1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 1986, 2010, Oracle and/or its affiliates. All rights reserved.
23 */
24
25 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
26 /* All Rights Reserved */
27
28 /*
29 * University Copyright- Copyright (c) 1982, 1986, 1988
30 * The Regents of the University of California
31 * All Rights Reserved
32 *
33 * University Acknowledgment- Portions of this document are derived from
34 * software developed by the University of California, Berkeley, and its
35 * contributors.
36 */
37
38 /*
39 * VM - shared or copy-on-write from a vnode/anonymous memory.
40 */
41
42 #include <sys/types.h>
43 #include <sys/param.h>
44 #include <sys/t_lock.h>
45 #include <sys/errno.h>
46 #include <sys/systm.h>
47 #include <sys/mman.h>
48 #include <sys/debug.h>
49 #include <sys/cred.h>
50 #include <sys/vmsystm.h>
51 #include <sys/tuneable.h>
52 #include <sys/bitmap.h>
53 #include <sys/swap.h>
54 #include <sys/kmem.h>
55 #include <sys/sysmacros.h>
56 #include <sys/vtrace.h>
57 #include <sys/cmn_err.h>
58 #include <sys/callb.h>
59 #include <sys/vm.h>
60 #include <sys/dumphdr.h>
61 #include <sys/lgrp.h>
62
63 #include <vm/hat.h>
64 #include <vm/as.h>
65 #include <vm/seg.h>
66 #include <vm/seg_vn.h>
67 #include <vm/pvn.h>
68 #include <vm/anon.h>
69 #include <vm/page.h>
70 #include <vm/vpage.h>
71 #include <sys/proc.h>
72 #include <sys/task.h>
73 #include <sys/project.h>
74 #include <sys/zone.h>
75 #include <sys/shm_impl.h>
76 /*
77 * Private seg op routines.
78 */
79 static int segvn_dup(struct seg *seg, struct seg *newseg);
80 static int segvn_unmap(struct seg *seg, caddr_t addr, size_t len);
81 static void segvn_free(struct seg *seg);
82 static faultcode_t segvn_fault(struct hat *hat, struct seg *seg,
83 caddr_t addr, size_t len, enum fault_type type,
84 enum seg_rw rw);
85 static faultcode_t segvn_faulta(struct seg *seg, caddr_t addr);
86 static int segvn_setprot(struct seg *seg, caddr_t addr,
87 size_t len, uint_t prot);
88 static int segvn_checkprot(struct seg *seg, caddr_t addr,
89 size_t len, uint_t prot);
90 static int segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta);
91 static int segvn_sync(struct seg *seg, caddr_t addr, size_t len,
92 int attr, uint_t flags);
93 static size_t segvn_incore(struct seg *seg, caddr_t addr, size_t len,
94 char *vec);
95 static int segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
96 int attr, int op, ulong_t *lockmap, size_t pos);
97 static int segvn_getprot(struct seg *seg, caddr_t addr, size_t len,
98 uint_t *protv);
99 static u_offset_t segvn_getoffset(struct seg *seg, caddr_t addr);
100 static int segvn_gettype(struct seg *seg, caddr_t addr);
101 static int segvn_getvp(struct seg *seg, caddr_t addr,
102 struct vnode **vpp);
103 static int segvn_advise(struct seg *seg, caddr_t addr, size_t len,
104 uint_t behav);
105 static void segvn_dump(struct seg *seg);
106 static int segvn_pagelock(struct seg *seg, caddr_t addr, size_t len,
107 struct page ***ppp, enum lock_type type, enum seg_rw rw);
108 static int segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len,
109 uint_t szc);
110 static int segvn_getmemid(struct seg *seg, caddr_t addr,
111 memid_t *memidp);
112 static lgrp_mem_policy_info_t *segvn_getpolicy(struct seg *, caddr_t);
113 static int segvn_capable(struct seg *seg, segcapability_t capable);
114
115 struct seg_ops segvn_ops = {
116 segvn_dup,
117 segvn_unmap,
118 segvn_free,
119 segvn_fault,
120 segvn_faulta,
121 segvn_setprot,
122 segvn_checkprot,
123 segvn_kluster,
124 segvn_sync,
125 segvn_incore,
126 segvn_lockop,
127 segvn_getprot,
128 segvn_getoffset,
129 segvn_gettype,
130 segvn_getvp,
131 segvn_advise,
132 segvn_dump,
133 segvn_pagelock,
134 segvn_setpagesize,
135 segvn_getmemid,
136 segvn_getpolicy,
137 segvn_capable,
138 };
139
140 /*
141 * Common zfod structures, provided as a shorthand for others to use.
142 */
143 static segvn_crargs_t zfod_segvn_crargs =
144 SEGVN_ZFOD_ARGS(PROT_ZFOD, PROT_ALL);
145 static segvn_crargs_t kzfod_segvn_crargs =
146 SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_USER,
147 PROT_ALL & ~PROT_USER);
148 static segvn_crargs_t stack_noexec_crargs =
149 SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_EXEC, PROT_ALL);
150
151 caddr_t zfod_argsp = (caddr_t)&zfod_segvn_crargs; /* user zfod argsp */
152 caddr_t kzfod_argsp = (caddr_t)&kzfod_segvn_crargs; /* kernel zfod argsp */
153 caddr_t stack_exec_argsp = (caddr_t)&zfod_segvn_crargs; /* executable stack */
154 caddr_t stack_noexec_argsp = (caddr_t)&stack_noexec_crargs; /* noexec stack */
155
156 #define vpgtob(n) ((n) * sizeof (struct vpage)) /* For brevity */
157
158 size_t segvn_comb_thrshld = UINT_MAX; /* patchable -- see 1196681 */
159
160 size_t segvn_pglock_comb_thrshld = (1UL << 16); /* 64K */
161 size_t segvn_pglock_comb_balign = (1UL << 16); /* 64K */
162 uint_t segvn_pglock_comb_bshift;
163 size_t segvn_pglock_comb_palign;
164
165 static int segvn_concat(struct seg *, struct seg *, int);
166 static int segvn_extend_prev(struct seg *, struct seg *,
167 struct segvn_crargs *, size_t);
168 static int segvn_extend_next(struct seg *, struct seg *,
169 struct segvn_crargs *, size_t);
170 static void segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
171 static void segvn_pagelist_rele(page_t **);
172 static void segvn_setvnode_mpss(vnode_t *);
173 static void segvn_relocate_pages(page_t **, page_t *);
174 static int segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
175 static int segvn_fill_vp_pages(struct segvn_data *, vnode_t *, u_offset_t,
176 uint_t, page_t **, page_t **, uint_t *, int *);
177 static faultcode_t segvn_fault_vnodepages(struct hat *, struct seg *, caddr_t,
178 caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
179 static faultcode_t segvn_fault_anonpages(struct hat *, struct seg *, caddr_t,
180 caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
181 static faultcode_t segvn_faultpage(struct hat *, struct seg *, caddr_t,
182 u_offset_t, struct vpage *, page_t **, uint_t,
183 enum fault_type, enum seg_rw, int);
184 static void segvn_vpage(struct seg *);
185 static size_t segvn_count_swap_by_vpages(struct seg *);
186
187 static void segvn_purge(struct seg *seg);
188 static int segvn_reclaim(void *, caddr_t, size_t, struct page **,
189 enum seg_rw, int);
190 static int shamp_reclaim(void *, caddr_t, size_t, struct page **,
191 enum seg_rw, int);
192
193 static int sameprot(struct seg *, caddr_t, size_t);
194
195 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
196 static int segvn_clrszc(struct seg *);
197 static struct seg *segvn_split_seg(struct seg *, caddr_t);
198 static int segvn_claim_pages(struct seg *, struct vpage *, u_offset_t,
199 ulong_t, uint_t);
200
201 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
202 size_t, void *, u_offset_t);
203
204 static struct kmem_cache *segvn_cache;
205 static struct kmem_cache **segvn_szc_cache;
206
207 #ifdef VM_STATS
208 static struct segvnvmstats_str {
209 ulong_t fill_vp_pages[31];
210 ulong_t fltvnpages[49];
211 ulong_t fullszcpages[10];
212 ulong_t relocatepages[3];
213 ulong_t fltanpages[17];
214 ulong_t pagelock[2];
215 ulong_t demoterange[3];
216 } segvnvmstats;
217 #endif /* VM_STATS */
218
219 #define SDR_RANGE 1 /* demote entire range */
220 #define SDR_END 2 /* demote non aligned ends only */
221
222 #define CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) { \
223 if ((len) != 0) { \
224 lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz); \
225 ASSERT(lpgaddr >= (seg)->s_base); \
226 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) + \
227 (len)), pgsz); \
228 ASSERT(lpgeaddr > lpgaddr); \
229 ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size); \
230 } else { \
231 lpgeaddr = lpgaddr = (addr); \
232 } \
233 }
234
235 /*ARGSUSED*/
236 static int
237 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
238 {
239 struct segvn_data *svd = buf;
240
241 rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
242 mutex_init(&svd->segfree_syncmtx, NULL, MUTEX_DEFAULT, NULL);
243 svd->svn_trnext = svd->svn_trprev = NULL;
244 return (0);
245 }
246
247 /*ARGSUSED1*/
248 static void
249 segvn_cache_destructor(void *buf, void *cdrarg)
250 {
251 struct segvn_data *svd = buf;
252
253 rw_destroy(&svd->lock);
254 mutex_destroy(&svd->segfree_syncmtx);
255 }
256
257 /*ARGSUSED*/
258 static int
259 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
260 {
261 bzero(buf, sizeof (svntr_t));
262 return (0);
263 }
264
265 /*
266 * Patching this variable to non-zero allows the system to run with
267 * stacks marked as "not executable". It's a bit of a kludge, but is
268 * provided as a tweakable for platforms that export those ABIs
269 * (e.g. sparc V8) that have executable stacks enabled by default.
270 * There are also some restrictions for platforms that don't actually
271 * implement 'noexec' protections.
272 *
273 * Once enabled, the system is (therefore) unable to provide a fully
274 * ABI-compliant execution environment, though practically speaking,
275 * most everything works. The exceptions are generally some interpreters
276 * and debuggers that create executable code on the stack and jump
277 * into it (without explicitly mprotecting the address range to include
278 * PROT_EXEC).
279 *
280 * One important class of applications that are disabled are those
281 * that have been transformed into malicious agents using one of the
282 * numerous "buffer overflow" attacks. See 4007890.
283 */
284 int noexec_user_stack = 0;
285 int noexec_user_stack_log = 1;
286
287 int segvn_lpg_disable = 0;
288 uint_t segvn_maxpgszc = 0;
289
290 ulong_t segvn_vmpss_clrszc_cnt;
291 ulong_t segvn_vmpss_clrszc_err;
292 ulong_t segvn_fltvnpages_clrszc_cnt;
293 ulong_t segvn_fltvnpages_clrszc_err;
294 ulong_t segvn_setpgsz_align_err;
295 ulong_t segvn_setpgsz_anon_align_err;
296 ulong_t segvn_setpgsz_getattr_err;
297 ulong_t segvn_setpgsz_eof_err;
298 ulong_t segvn_faultvnmpss_align_err1;
299 ulong_t segvn_faultvnmpss_align_err2;
300 ulong_t segvn_faultvnmpss_align_err3;
301 ulong_t segvn_faultvnmpss_align_err4;
302 ulong_t segvn_faultvnmpss_align_err5;
303 ulong_t segvn_vmpss_pageio_deadlk_err;
304
305 int segvn_use_regions = 1;
306
307 /*
308 * Segvn supports text replication optimization for NUMA platforms. Text
309 * replica's are represented by anon maps (amp). There's one amp per text file
310 * region per lgroup. A process chooses the amp for each of its text mappings
311 * based on the lgroup assignment of its main thread (t_tid = 1). All
312 * processes that want a replica on a particular lgroup for the same text file
313 * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
314 * with vp,off,size,szc used as a key. Text replication segments are read only
315 * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
316 * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
317 * pages. Replication amp is assigned to a segment when it gets its first
318 * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
319 * rechecks periodically if the process still maps an amp local to the main
320 * thread. If not async thread forces process to remap to an amp in the new
321 * home lgroup of the main thread. Current text replication implementation
322 * only provides the benefit to workloads that do most of their work in the
323 * main thread of a process or all the threads of a process run in the same
324 * lgroup. To extend text replication benefit to different types of
325 * multithreaded workloads further work would be needed in the hat layer to
326 * allow the same virtual address in the same hat to simultaneously map
327 * different physical addresses (i.e. page table replication would be needed
328 * for x86).
329 *
330 * amp pages are used instead of vnode pages as long as segment has a very
331 * simple life cycle. It's created via segvn_create(), handles S_EXEC
332 * (S_READ) pagefaults and is fully unmapped. If anything more complicated
333 * happens such as protection is changed, real COW fault happens, pagesize is
334 * changed, MC_LOCK is requested or segment is partially unmapped we turn off
335 * text replication by converting the segment back to vnode only segment
336 * (unmap segment's address range and set svd->amp to NULL).
337 *
338 * The original file can be changed after amp is inserted into
339 * svntr_hashtab. Processes that are launched after the file is already
340 * changed can't use the replica's created prior to the file change. To
341 * implement this functionality hash entries are timestamped. Replica's can
342 * only be used if current file modification time is the same as the timestamp
343 * saved when hash entry was created. However just timestamps alone are not
344 * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
345 * deal with file changes via MAP_SHARED mappings differently. When writable
346 * MAP_SHARED mappings are created to vnodes marked as executable we mark all
347 * existing replica's for this vnode as not usable for future text
348 * mappings. And we don't create new replica's for files that currently have
349 * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
350 * true).
351 */
352
353 #define SEGVN_TEXTREPL_MAXBYTES_FACTOR (20)
354 size_t segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
355
356 static ulong_t svntr_hashtab_sz = 512;
357 static svntr_bucket_t *svntr_hashtab = NULL;
358 static struct kmem_cache *svntr_cache;
359 static svntr_stats_t *segvn_textrepl_stats;
360 static ksema_t segvn_trasync_sem;
361
362 int segvn_disable_textrepl = 1;
363 size_t textrepl_size_thresh = (size_t)-1;
364 size_t segvn_textrepl_bytes = 0;
365 size_t segvn_textrepl_max_bytes = 0;
366 clock_t segvn_update_textrepl_interval = 0;
367 int segvn_update_tr_time = 10;
368 int segvn_disable_textrepl_update = 0;
369
370 static void segvn_textrepl(struct seg *);
371 static void segvn_textunrepl(struct seg *, int);
372 static void segvn_inval_trcache(vnode_t *);
373 static void segvn_trasync_thread(void);
374 static void segvn_trupdate_wakeup(void *);
375 static void segvn_trupdate(void);
376 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
377 ulong_t);
378
379 /*
380 * Initialize segvn data structures
381 */
382 void
383 segvn_init(void)
384 {
385 uint_t maxszc;
386 uint_t szc;
387 size_t pgsz;
388
389 segvn_cache = kmem_cache_create("segvn_cache",
390 sizeof (struct segvn_data), 0,
391 segvn_cache_constructor, segvn_cache_destructor, NULL,
392 NULL, NULL, 0);
393
394 if (segvn_lpg_disable == 0) {
395 szc = maxszc = page_num_pagesizes() - 1;
396 if (szc == 0) {
397 segvn_lpg_disable = 1;
398 }
399 if (page_get_pagesize(0) != PAGESIZE) {
400 panic("segvn_init: bad szc 0");
401 /*NOTREACHED*/
402 }
403 while (szc != 0) {
404 pgsz = page_get_pagesize(szc);
405 if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
406 panic("segvn_init: bad szc %d", szc);
407 /*NOTREACHED*/
408 }
409 szc--;
410 }
411 if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
412 segvn_maxpgszc = maxszc;
413 }
414
415 if (segvn_maxpgszc) {
416 segvn_szc_cache = (struct kmem_cache **)kmem_alloc(
417 (segvn_maxpgszc + 1) * sizeof (struct kmem_cache *),
418 KM_SLEEP);
419 }
420
421 for (szc = 1; szc <= segvn_maxpgszc; szc++) {
422 char str[32];
423
424 (void) sprintf(str, "segvn_szc_cache%d", szc);
425 segvn_szc_cache[szc] = kmem_cache_create(str,
426 page_get_pagecnt(szc) * sizeof (page_t *), 0,
427 NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
428 }
429
430
431 if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
432 segvn_use_regions = 0;
433
434 /*
435 * For now shared regions and text replication segvn support
436 * are mutually exclusive. This is acceptable because
437 * currently significant benefit from text replication was
438 * only observed on AMD64 NUMA platforms (due to relatively
439 * small L2$ size) and currently we don't support shared
440 * regions on x86.
441 */
442 if (segvn_use_regions && !segvn_disable_textrepl) {
443 segvn_disable_textrepl = 1;
444 }
445
446 #if defined(_LP64)
447 if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
448 !segvn_disable_textrepl) {
449 ulong_t i;
450 size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
451
452 svntr_cache = kmem_cache_create("svntr_cache",
453 sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
454 NULL, NULL, NULL, 0);
455 svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
456 for (i = 0; i < svntr_hashtab_sz; i++) {
457 mutex_init(&svntr_hashtab[i].tr_lock, NULL,
458 MUTEX_DEFAULT, NULL);
459 }
460 segvn_textrepl_max_bytes = ptob(physmem) /
461 segvn_textrepl_max_bytes_factor;
462 segvn_textrepl_stats = kmem_zalloc(NCPU *
463 sizeof (svntr_stats_t), KM_SLEEP);
464 sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
465 (void) thread_create(NULL, 0, segvn_trasync_thread,
466 NULL, 0, &p0, TS_RUN, minclsyspri);
467 }
468 #endif
469
470 if (!ISP2(segvn_pglock_comb_balign) ||
471 segvn_pglock_comb_balign < PAGESIZE) {
472 segvn_pglock_comb_balign = 1UL << 16; /* 64K */
473 }
474 segvn_pglock_comb_bshift = highbit(segvn_pglock_comb_balign) - 1;
475 segvn_pglock_comb_palign = btop(segvn_pglock_comb_balign);
476 }
477
478 #define SEGVN_PAGEIO ((void *)0x1)
479 #define SEGVN_NOPAGEIO ((void *)0x2)
480
481 static void
482 segvn_setvnode_mpss(vnode_t *vp)
483 {
484 int err;
485
486 ASSERT(vp->v_mpssdata == NULL ||
487 vp->v_mpssdata == SEGVN_PAGEIO ||
488 vp->v_mpssdata == SEGVN_NOPAGEIO);
489
490 if (vp->v_mpssdata == NULL) {
491 if (vn_vmpss_usepageio(vp)) {
492 err = VOP_PAGEIO(vp, (page_t *)NULL,
493 (u_offset_t)0, 0, 0, CRED(), NULL);
494 } else {
495 err = ENOSYS;
496 }
497 /*
498 * set v_mpssdata just once per vnode life
499 * so that it never changes.
500 */
501 mutex_enter(&vp->v_lock);
502 if (vp->v_mpssdata == NULL) {
503 if (err == EINVAL) {
504 vp->v_mpssdata = SEGVN_PAGEIO;
505 } else {
506 vp->v_mpssdata = SEGVN_NOPAGEIO;
507 }
508 }
509 mutex_exit(&vp->v_lock);
510 }
511 }
512
513 int
514 segvn_create(struct seg *seg, void *argsp)
515 {
516 struct segvn_crargs *a = (struct segvn_crargs *)argsp;
517 struct segvn_data *svd;
518 size_t swresv = 0;
519 struct cred *cred;
520 struct anon_map *amp;
521 int error = 0;
522 size_t pgsz;
523 lgrp_mem_policy_t mpolicy = LGRP_MEM_POLICY_DEFAULT;
524 int use_rgn = 0;
525 int trok = 0;
526
527 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
528
529 if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
530 panic("segvn_create type");
531 /*NOTREACHED*/
532 }
533
534 /*
535 * Check arguments. If a shared anon structure is given then
536 * it is illegal to also specify a vp.
537 */
538 if (a->amp != NULL && a->vp != NULL) {
539 panic("segvn_create anon_map");
540 /*NOTREACHED*/
541 }
542
543 if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
544 a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
545 segvn_use_regions) {
546 use_rgn = 1;
547 }
548
549 /* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
550 if (a->type == MAP_SHARED)
551 a->flags &= ~MAP_NORESERVE;
552
553 if (a->szc != 0) {
554 if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
555 (a->amp != NULL && a->type == MAP_PRIVATE) ||
556 (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
557 a->szc = 0;
558 } else {
559 if (a->szc > segvn_maxpgszc)
560 a->szc = segvn_maxpgszc;
561 pgsz = page_get_pagesize(a->szc);
562 if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
563 !IS_P2ALIGNED(seg->s_size, pgsz)) {
564 a->szc = 0;
565 } else if (a->vp != NULL) {
566 if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
567 /*
568 * paranoid check.
569 * hat_page_demote() is not supported
570 * on swapfs pages.
571 */
572 a->szc = 0;
573 } else if (map_addr_vacalign_check(seg->s_base,
574 a->offset & PAGEMASK)) {
575 a->szc = 0;
576 }
577 } else if (a->amp != NULL) {
578 pgcnt_t anum = btopr(a->offset);
579 pgcnt_t pgcnt = page_get_pagecnt(a->szc);
580 if (!IS_P2ALIGNED(anum, pgcnt)) {
581 a->szc = 0;
582 }
583 }
584 }
585 }
586
587 /*
588 * If segment may need private pages, reserve them now.
589 */
590 if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
591 (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
592 if (anon_resv_zone(seg->s_size,
593 seg->s_as->a_proc->p_zone) == 0)
594 return (EAGAIN);
595 swresv = seg->s_size;
596 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
597 seg, swresv, 1);
598 }
599
600 /*
601 * Reserve any mapping structures that may be required.
602 *
603 * Don't do it for segments that may use regions. It's currently a
604 * noop in the hat implementations anyway.
605 */
606 if (!use_rgn) {
607 hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
608 }
609
610 if (a->cred) {
611 cred = a->cred;
612 crhold(cred);
613 } else {
614 crhold(cred = CRED());
615 }
616
617 /* Inform the vnode of the new mapping */
618 if (a->vp != NULL) {
619 error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
620 seg->s_as, seg->s_base, seg->s_size, a->prot,
621 a->maxprot, a->type, cred, NULL);
622 if (error) {
623 if (swresv != 0) {
624 anon_unresv_zone(swresv,
625 seg->s_as->a_proc->p_zone);
626 TRACE_3(TR_FAC_VM, TR_ANON_PROC,
627 "anon proc:%p %lu %u", seg, swresv, 0);
628 }
629 crfree(cred);
630 if (!use_rgn) {
631 hat_unload(seg->s_as->a_hat, seg->s_base,
632 seg->s_size, HAT_UNLOAD_UNMAP);
633 }
634 return (error);
635 }
636 /*
637 * svntr_hashtab will be NULL if we support shared regions.
638 */
639 trok = ((a->flags & MAP_TEXT) &&
640 (seg->s_size > textrepl_size_thresh ||
641 (a->flags & _MAP_TEXTREPL)) &&
642 lgrp_optimizations() && svntr_hashtab != NULL &&
643 a->type == MAP_PRIVATE && swresv == 0 &&
644 !(a->flags & MAP_NORESERVE) &&
645 seg->s_as != &kas && a->vp->v_type == VREG);
646
647 ASSERT(!trok || !use_rgn);
648 }
649
650 /*
651 * MAP_NORESERVE mappings don't count towards the VSZ of a process
652 * until we fault the pages in.
653 */
654 if ((a->vp == NULL || a->vp->v_type != VREG) &&
655 a->flags & MAP_NORESERVE) {
656 seg->s_as->a_resvsize -= seg->s_size;
657 }
658
659 /*
660 * If more than one segment in the address space, and they're adjacent
661 * virtually, try to concatenate them. Don't concatenate if an
662 * explicit anon_map structure was supplied (e.g., SystemV shared
663 * memory) or if we'll use text replication for this segment.
664 */
665 if (a->amp == NULL && !use_rgn && !trok) {
666 struct seg *pseg, *nseg;
667 struct segvn_data *psvd, *nsvd;
668 lgrp_mem_policy_t ppolicy, npolicy;
669 uint_t lgrp_mem_policy_flags = 0;
670 extern lgrp_mem_policy_t lgrp_mem_default_policy;
671
672 /*
673 * Memory policy flags (lgrp_mem_policy_flags) is valid when
674 * extending stack/heap segments.
675 */
676 if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
677 !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
678 lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
679 } else {
680 /*
681 * Get policy when not extending it from another segment
682 */
683 mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
684 }
685
686 /*
687 * First, try to concatenate the previous and new segments
688 */
689 pseg = AS_SEGPREV(seg->s_as, seg);
690 if (pseg != NULL &&
691 pseg->s_base + pseg->s_size == seg->s_base &&
692 pseg->s_ops == &segvn_ops) {
693 /*
694 * Get memory allocation policy from previous segment.
695 * When extension is specified (e.g. for heap) apply
696 * this policy to the new segment regardless of the
697 * outcome of segment concatenation. Extension occurs
698 * for non-default policy otherwise default policy is
699 * used and is based on extended segment size.
700 */
701 psvd = (struct segvn_data *)pseg->s_data;
702 ppolicy = psvd->policy_info.mem_policy;
703 if (lgrp_mem_policy_flags ==
704 LGRP_MP_FLAG_EXTEND_UP) {
705 if (ppolicy != lgrp_mem_default_policy) {
706 mpolicy = ppolicy;
707 } else {
708 mpolicy = lgrp_mem_policy_default(
709 pseg->s_size + seg->s_size,
710 a->type);
711 }
712 }
713
714 if (mpolicy == ppolicy &&
715 (pseg->s_size + seg->s_size <=
716 segvn_comb_thrshld || psvd->amp == NULL) &&
717 segvn_extend_prev(pseg, seg, a, swresv) == 0) {
718 /*
719 * success! now try to concatenate
720 * with following seg
721 */
722 crfree(cred);
723 nseg = AS_SEGNEXT(pseg->s_as, pseg);
724 if (nseg != NULL &&
725 nseg != pseg &&
726 nseg->s_ops == &segvn_ops &&
727 pseg->s_base + pseg->s_size ==
728 nseg->s_base)
729 (void) segvn_concat(pseg, nseg, 0);
730 ASSERT(pseg->s_szc == 0 ||
731 (a->szc == pseg->s_szc &&
732 IS_P2ALIGNED(pseg->s_base, pgsz) &&
733 IS_P2ALIGNED(pseg->s_size, pgsz)));
734 return (0);
735 }
736 }
737
738 /*
739 * Failed, so try to concatenate with following seg
740 */
741 nseg = AS_SEGNEXT(seg->s_as, seg);
742 if (nseg != NULL &&
743 seg->s_base + seg->s_size == nseg->s_base &&
744 nseg->s_ops == &segvn_ops) {
745 /*
746 * Get memory allocation policy from next segment.
747 * When extension is specified (e.g. for stack) apply
748 * this policy to the new segment regardless of the
749 * outcome of segment concatenation. Extension occurs
750 * for non-default policy otherwise default policy is
751 * used and is based on extended segment size.
752 */
753 nsvd = (struct segvn_data *)nseg->s_data;
754 npolicy = nsvd->policy_info.mem_policy;
755 if (lgrp_mem_policy_flags ==
756 LGRP_MP_FLAG_EXTEND_DOWN) {
757 if (npolicy != lgrp_mem_default_policy) {
758 mpolicy = npolicy;
759 } else {
760 mpolicy = lgrp_mem_policy_default(
761 nseg->s_size + seg->s_size,
762 a->type);
763 }
764 }
765
766 if (mpolicy == npolicy &&
767 segvn_extend_next(seg, nseg, a, swresv) == 0) {
768 crfree(cred);
769 ASSERT(nseg->s_szc == 0 ||
770 (a->szc == nseg->s_szc &&
771 IS_P2ALIGNED(nseg->s_base, pgsz) &&
772 IS_P2ALIGNED(nseg->s_size, pgsz)));
773 return (0);
774 }
775 }
776 }
777
778 if (a->vp != NULL) {
779 VN_HOLD(a->vp);
780 if (a->type == MAP_SHARED)
781 lgrp_shm_policy_init(NULL, a->vp);
782 }
783 svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
784
785 seg->s_ops = &segvn_ops;
786 seg->s_data = (void *)svd;
787 seg->s_szc = a->szc;
788
789 svd->seg = seg;
790 svd->vp = a->vp;
791 /*
792 * Anonymous mappings have no backing file so the offset is meaningless.
793 */
794 svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
795 svd->prot = a->prot;
796 svd->maxprot = a->maxprot;
797 svd->pageprot = 0;
798 svd->type = a->type;
799 svd->vpage = NULL;
800 svd->cred = cred;
801 svd->advice = MADV_NORMAL;
802 svd->pageadvice = 0;
803 svd->flags = (ushort_t)a->flags;
804 svd->softlockcnt = 0;
805 svd->softlockcnt_sbase = 0;
806 svd->softlockcnt_send = 0;
807 svd->rcookie = HAT_INVALID_REGION_COOKIE;
808 svd->pageswap = 0;
809
810 if (a->szc != 0 && a->vp != NULL) {
811 segvn_setvnode_mpss(a->vp);
812 }
813 if (svd->type == MAP_SHARED && svd->vp != NULL &&
814 (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
815 ASSERT(vn_is_mapped(svd->vp, V_WRITE));
816 segvn_inval_trcache(svd->vp);
817 }
818
819 amp = a->amp;
820 if ((svd->amp = amp) == NULL) {
821 svd->anon_index = 0;
822 if (svd->type == MAP_SHARED) {
823 svd->swresv = 0;
824 /*
825 * Shared mappings to a vp need no other setup.
826 * If we have a shared mapping to an anon_map object
827 * which hasn't been allocated yet, allocate the
828 * struct now so that it will be properly shared
829 * by remembering the swap reservation there.
830 */
831 if (a->vp == NULL) {
832 svd->amp = anonmap_alloc(seg->s_size, swresv,
833 ANON_SLEEP);
834 svd->amp->a_szc = seg->s_szc;
835 }
836 } else {
837 /*
838 * Private mapping (with or without a vp).
839 * Allocate anon_map when needed.
840 */
841 svd->swresv = swresv;
842 }
843 } else {
844 pgcnt_t anon_num;
845
846 /*
847 * Mapping to an existing anon_map structure without a vp.
848 * For now we will insure that the segment size isn't larger
849 * than the size - offset gives us. Later on we may wish to
850 * have the anon array dynamically allocated itself so that
851 * we don't always have to allocate all the anon pointer slots.
852 * This of course involves adding extra code to check that we
853 * aren't trying to use an anon pointer slot beyond the end
854 * of the currently allocated anon array.
855 */
856 if ((amp->size - a->offset) < seg->s_size) {
857 panic("segvn_create anon_map size");
858 /*NOTREACHED*/
859 }
860
861 anon_num = btopr(a->offset);
862
863 if (a->type == MAP_SHARED) {
864 /*
865 * SHARED mapping to a given anon_map.
866 */
867 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
868 amp->refcnt++;
869 if (a->szc > amp->a_szc) {
870 amp->a_szc = a->szc;
871 }
872 ANON_LOCK_EXIT(&->a_rwlock);
873 svd->anon_index = anon_num;
874 svd->swresv = 0;
875 } else {
876 /*
877 * PRIVATE mapping to a given anon_map.
878 * Make sure that all the needed anon
879 * structures are created (so that we will
880 * share the underlying pages if nothing
881 * is written by this mapping) and then
882 * duplicate the anon array as is done
883 * when a privately mapped segment is dup'ed.
884 */
885 struct anon *ap;
886 caddr_t addr;
887 caddr_t eaddr;
888 ulong_t anon_idx;
889 int hat_flag = HAT_LOAD;
890
891 if (svd->flags & MAP_TEXT) {
892 hat_flag |= HAT_LOAD_TEXT;
893 }
894
895 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
896 svd->amp->a_szc = seg->s_szc;
897 svd->anon_index = 0;
898 svd->swresv = swresv;
899
900 /*
901 * Prevent 2 threads from allocating anon
902 * slots simultaneously.
903 */
904 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
905 eaddr = seg->s_base + seg->s_size;
906
907 for (anon_idx = anon_num, addr = seg->s_base;
908 addr < eaddr; addr += PAGESIZE, anon_idx++) {
909 page_t *pp;
910
911 if ((ap = anon_get_ptr(amp->ahp,
912 anon_idx)) != NULL)
913 continue;
914
915 /*
916 * Allocate the anon struct now.
917 * Might as well load up translation
918 * to the page while we're at it...
919 */
920 pp = anon_zero(seg, addr, &ap, cred);
921 if (ap == NULL || pp == NULL) {
922 panic("segvn_create anon_zero");
923 /*NOTREACHED*/
924 }
925
926 /*
927 * Re-acquire the anon_map lock and
928 * initialize the anon array entry.
929 */
930 ASSERT(anon_get_ptr(amp->ahp,
931 anon_idx) == NULL);
932 (void) anon_set_ptr(amp->ahp, anon_idx, ap,
933 ANON_SLEEP);
934
935 ASSERT(seg->s_szc == 0);
936 ASSERT(!IS_VMODSORT(pp->p_vnode));
937
938 ASSERT(use_rgn == 0);
939 hat_memload(seg->s_as->a_hat, addr, pp,
940 svd->prot & ~PROT_WRITE, hat_flag);
941
942 page_unlock(pp);
943 }
944 ASSERT(seg->s_szc == 0);
945 anon_dup(amp->ahp, anon_num, svd->amp->ahp,
946 0, seg->s_size);
947 ANON_LOCK_EXIT(&->a_rwlock);
948 }
949 }
950
951 /*
952 * Set default memory allocation policy for segment
953 *
954 * Always set policy for private memory at least for initialization
955 * even if this is a shared memory segment
956 */
957 (void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
958
959 if (svd->type == MAP_SHARED)
960 (void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
961 svd->vp, svd->offset, seg->s_size);
962
963 if (use_rgn) {
964 ASSERT(!trok);
965 ASSERT(svd->amp == NULL);
966 svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
967 seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
968 (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
969 HAT_REGION_TEXT);
970 }
971
972 ASSERT(!trok || !(svd->prot & PROT_WRITE));
973 svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
974
975 return (0);
976 }
977
978 /*
979 * Concatenate two existing segments, if possible.
980 * Return 0 on success, -1 if two segments are not compatible
981 * or -2 on memory allocation failure.
982 * If amp_cat == 1 then try and concat segments with anon maps
983 */
984 static int
985 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
986 {
987 struct segvn_data *svd1 = seg1->s_data;
988 struct segvn_data *svd2 = seg2->s_data;
989 struct anon_map *amp1 = svd1->amp;
990 struct anon_map *amp2 = svd2->amp;
991 struct vpage *vpage1 = svd1->vpage;
992 struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
993 size_t size, nvpsize;
994 pgcnt_t npages1, npages2;
995
996 ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
997 ASSERT(AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
998 ASSERT(seg1->s_ops == seg2->s_ops);
999
1000 if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
1001 HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1002 return (-1);
1003 }
1004
1005 /* both segments exist, try to merge them */
1006 #define incompat(x) (svd1->x != svd2->x)
1007 if (incompat(vp) || incompat(maxprot) ||
1008 (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
1009 (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
1010 incompat(type) || incompat(cred) || incompat(flags) ||
1011 seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
1012 (svd2->softlockcnt > 0) || svd1->softlockcnt_send > 0)
1013 return (-1);
1014 #undef incompat
1015
1016 /*
1017 * vp == NULL implies zfod, offset doesn't matter
1018 */
1019 if (svd1->vp != NULL &&
1020 svd1->offset + seg1->s_size != svd2->offset) {
1021 return (-1);
1022 }
1023
1024 /*
1025 * Don't concatenate if either segment uses text replication.
1026 */
1027 if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
1028 return (-1);
1029 }
1030
1031 /*
1032 * Fail early if we're not supposed to concatenate
1033 * segments with non NULL amp.
1034 */
1035 if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1036 return (-1);
1037 }
1038
1039 if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1040 if (amp1 != amp2) {
1041 return (-1);
1042 }
1043 if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1044 svd2->anon_index) {
1045 return (-1);
1046 }
1047 ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1048 }
1049
1050 /*
1051 * If either seg has vpages, create a new merged vpage array.
1052 */
1053 if (vpage1 != NULL || vpage2 != NULL) {
1054 struct vpage *vp, *evp;
1055
1056 npages1 = seg_pages(seg1);
1057 npages2 = seg_pages(seg2);
1058 nvpsize = vpgtob(npages1 + npages2);
1059
1060 if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1061 return (-2);
1062 }
1063
1064 if (vpage1 != NULL) {
1065 bcopy(vpage1, nvpage, vpgtob(npages1));
1066 } else {
1067 evp = nvpage + npages1;
1068 for (vp = nvpage; vp < evp; vp++) {
1069 VPP_SETPROT(vp, svd1->prot);
1070 VPP_SETADVICE(vp, svd1->advice);
1071 }
1072 }
1073
1074 if (vpage2 != NULL) {
1075 bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1076 } else {
1077 evp = nvpage + npages1 + npages2;
1078 for (vp = nvpage + npages1; vp < evp; vp++) {
1079 VPP_SETPROT(vp, svd2->prot);
1080 VPP_SETADVICE(vp, svd2->advice);
1081 }
1082 }
1083
1084 if (svd2->pageswap && (!svd1->pageswap && svd1->swresv)) {
1085 ASSERT(svd1->swresv == seg1->s_size);
1086 ASSERT(!(svd1->flags & MAP_NORESERVE));
1087 ASSERT(!(svd2->flags & MAP_NORESERVE));
1088 evp = nvpage + npages1;
1089 for (vp = nvpage; vp < evp; vp++) {
1090 VPP_SETSWAPRES(vp);
1091 }
1092 }
1093
1094 if (svd1->pageswap && (!svd2->pageswap && svd2->swresv)) {
1095 ASSERT(svd2->swresv == seg2->s_size);
1096 ASSERT(!(svd1->flags & MAP_NORESERVE));
1097 ASSERT(!(svd2->flags & MAP_NORESERVE));
1098 vp = nvpage + npages1;
1099 evp = vp + npages2;
1100 for (; vp < evp; vp++) {
1101 VPP_SETSWAPRES(vp);
1102 }
1103 }
1104 }
1105 ASSERT((vpage1 != NULL || vpage2 != NULL) ||
1106 (svd1->pageswap == 0 && svd2->pageswap == 0));
1107
1108 /*
1109 * If either segment has private pages, create a new merged anon
1110 * array. If mergeing shared anon segments just decrement anon map's
1111 * refcnt.
1112 */
1113 if (amp1 != NULL && svd1->type == MAP_SHARED) {
1114 ASSERT(amp1 == amp2 && svd1->vp == NULL);
1115 ANON_LOCK_ENTER(&1->a_rwlock, RW_WRITER);
1116 ASSERT(amp1->refcnt >= 2);
1117 amp1->refcnt--;
1118 ANON_LOCK_EXIT(&1->a_rwlock);
1119 svd2->amp = NULL;
1120 } else if (amp1 != NULL || amp2 != NULL) {
1121 struct anon_hdr *nahp;
1122 struct anon_map *namp = NULL;
1123 size_t asize;
1124
1125 ASSERT(svd1->type == MAP_PRIVATE);
1126
1127 asize = seg1->s_size + seg2->s_size;
1128 if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1129 if (nvpage != NULL) {
1130 kmem_free(nvpage, nvpsize);
1131 }
1132 return (-2);
1133 }
1134 if (amp1 != NULL) {
1135 /*
1136 * XXX anon rwlock is not really needed because
1137 * this is a private segment and we are writers.
1138 */
1139 ANON_LOCK_ENTER(&1->a_rwlock, RW_WRITER);
1140 ASSERT(amp1->refcnt == 1);
1141 if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1142 nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1143 anon_release(nahp, btop(asize));
1144 ANON_LOCK_EXIT(&1->a_rwlock);
1145 if (nvpage != NULL) {
1146 kmem_free(nvpage, nvpsize);
1147 }
1148 return (-2);
1149 }
1150 }
1151 if (amp2 != NULL) {
1152 ANON_LOCK_ENTER(&2->a_rwlock, RW_WRITER);
1153 ASSERT(amp2->refcnt == 1);
1154 if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1155 nahp, btop(seg1->s_size), btop(seg2->s_size),
1156 ANON_NOSLEEP)) {
1157 anon_release(nahp, btop(asize));
1158 ANON_LOCK_EXIT(&2->a_rwlock);
1159 if (amp1 != NULL) {
1160 ANON_LOCK_EXIT(&1->a_rwlock);
1161 }
1162 if (nvpage != NULL) {
1163 kmem_free(nvpage, nvpsize);
1164 }
1165 return (-2);
1166 }
1167 }
1168 if (amp1 != NULL) {
1169 namp = amp1;
1170 anon_release(amp1->ahp, btop(amp1->size));
1171 }
1172 if (amp2 != NULL) {
1173 if (namp == NULL) {
1174 ASSERT(amp1 == NULL);
1175 namp = amp2;
1176 anon_release(amp2->ahp, btop(amp2->size));
1177 } else {
1178 amp2->refcnt--;
1179 ANON_LOCK_EXIT(&2->a_rwlock);
1180 anonmap_free(amp2);
1181 }
1182 svd2->amp = NULL; /* needed for seg_free */
1183 }
1184 namp->ahp = nahp;
1185 namp->size = asize;
1186 svd1->amp = namp;
1187 svd1->anon_index = 0;
1188 ANON_LOCK_EXIT(&namp->a_rwlock);
1189 }
1190 /*
1191 * Now free the old vpage structures.
1192 */
1193 if (nvpage != NULL) {
1194 if (vpage1 != NULL) {
1195 kmem_free(vpage1, vpgtob(npages1));
1196 }
1197 if (vpage2 != NULL) {
1198 svd2->vpage = NULL;
1199 kmem_free(vpage2, vpgtob(npages2));
1200 }
1201 if (svd2->pageprot) {
1202 svd1->pageprot = 1;
1203 }
1204 if (svd2->pageadvice) {
1205 svd1->pageadvice = 1;
1206 }
1207 if (svd2->pageswap) {
1208 svd1->pageswap = 1;
1209 }
1210 svd1->vpage = nvpage;
1211 }
1212
1213 /* all looks ok, merge segments */
1214 svd1->swresv += svd2->swresv;
1215 svd2->swresv = 0; /* so seg_free doesn't release swap space */
1216 size = seg2->s_size;
1217 seg_free(seg2);
1218 seg1->s_size += size;
1219 return (0);
1220 }
1221
1222 /*
1223 * Extend the previous segment (seg1) to include the
1224 * new segment (seg2 + a), if possible.
1225 * Return 0 on success.
1226 */
1227 static int
1228 segvn_extend_prev(seg1, seg2, a, swresv)
1229 struct seg *seg1, *seg2;
1230 struct segvn_crargs *a;
1231 size_t swresv;
1232 {
1233 struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1234 size_t size;
1235 struct anon_map *amp1;
1236 struct vpage *new_vpage;
1237
1238 /*
1239 * We don't need any segment level locks for "segvn" data
1240 * since the address space is "write" locked.
1241 */
1242 ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
1243
1244 if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1245 return (-1);
1246 }
1247
1248 /* second segment is new, try to extend first */
1249 /* XXX - should also check cred */
1250 if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1251 (!svd1->pageprot && (svd1->prot != a->prot)) ||
1252 svd1->type != a->type || svd1->flags != a->flags ||
1253 seg1->s_szc != a->szc || svd1->softlockcnt_send > 0)
1254 return (-1);
1255
1256 /* vp == NULL implies zfod, offset doesn't matter */
1257 if (svd1->vp != NULL &&
1258 svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1259 return (-1);
1260
1261 if (svd1->tr_state != SEGVN_TR_OFF) {
1262 return (-1);
1263 }
1264
1265 amp1 = svd1->amp;
1266 if (amp1) {
1267 pgcnt_t newpgs;
1268
1269 /*
1270 * Segment has private pages, can data structures
1271 * be expanded?
1272 *
1273 * Acquire the anon_map lock to prevent it from changing,
1274 * if it is shared. This ensures that the anon_map
1275 * will not change while a thread which has a read/write
1276 * lock on an address space references it.
1277 * XXX - Don't need the anon_map lock at all if "refcnt"
1278 * is 1.
1279 *
1280 * Can't grow a MAP_SHARED segment with an anonmap because
1281 * there may be existing anon slots where we want to extend
1282 * the segment and we wouldn't know what to do with them
1283 * (e.g., for tmpfs right thing is to just leave them there,
1284 * for /dev/zero they should be cleared out).
1285 */
1286 if (svd1->type == MAP_SHARED)
1287 return (-1);
1288
1289 ANON_LOCK_ENTER(&1->a_rwlock, RW_WRITER);
1290 if (amp1->refcnt > 1) {
1291 ANON_LOCK_EXIT(&1->a_rwlock);
1292 return (-1);
1293 }
1294 newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1295 btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1296
1297 if (newpgs == 0) {
1298 ANON_LOCK_EXIT(&1->a_rwlock);
1299 return (-1);
1300 }
1301 amp1->size = ptob(newpgs);
1302 ANON_LOCK_EXIT(&1->a_rwlock);
1303 }
1304 if (svd1->vpage != NULL) {
1305 struct vpage *vp, *evp;
1306 new_vpage =
1307 kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1308 KM_NOSLEEP);
1309 if (new_vpage == NULL)
1310 return (-1);
1311 bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1312 kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1313 svd1->vpage = new_vpage;
1314
1315 vp = new_vpage + seg_pages(seg1);
1316 evp = vp + seg_pages(seg2);
1317 for (; vp < evp; vp++)
1318 VPP_SETPROT(vp, a->prot);
1319 if (svd1->pageswap && swresv) {
1320 ASSERT(!(svd1->flags & MAP_NORESERVE));
1321 ASSERT(swresv == seg2->s_size);
1322 vp = new_vpage + seg_pages(seg1);
1323 for (; vp < evp; vp++) {
1324 VPP_SETSWAPRES(vp);
1325 }
1326 }
1327 }
1328 ASSERT(svd1->vpage != NULL || svd1->pageswap == 0);
1329 size = seg2->s_size;
1330 seg_free(seg2);
1331 seg1->s_size += size;
1332 svd1->swresv += swresv;
1333 if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1334 svd1->type == MAP_SHARED && svd1->vp != NULL &&
1335 (svd1->vp->v_flag & VVMEXEC)) {
1336 ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1337 segvn_inval_trcache(svd1->vp);
1338 }
1339 return (0);
1340 }
1341
1342 /*
1343 * Extend the next segment (seg2) to include the
1344 * new segment (seg1 + a), if possible.
1345 * Return 0 on success.
1346 */
1347 static int
1348 segvn_extend_next(
1349 struct seg *seg1,
1350 struct seg *seg2,
1351 struct segvn_crargs *a,
1352 size_t swresv)
1353 {
1354 struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1355 size_t size;
1356 struct anon_map *amp2;
1357 struct vpage *new_vpage;
1358
1359 /*
1360 * We don't need any segment level locks for "segvn" data
1361 * since the address space is "write" locked.
1362 */
1363 ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as, &seg2->s_as->a_lock));
1364
1365 if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1366 return (-1);
1367 }
1368
1369 /* first segment is new, try to extend second */
1370 /* XXX - should also check cred */
1371 if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1372 (!svd2->pageprot && (svd2->prot != a->prot)) ||
1373 svd2->type != a->type || svd2->flags != a->flags ||
1374 seg2->s_szc != a->szc || svd2->softlockcnt_sbase > 0)
1375 return (-1);
1376 /* vp == NULL implies zfod, offset doesn't matter */
1377 if (svd2->vp != NULL &&
1378 (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1379 return (-1);
1380
1381 if (svd2->tr_state != SEGVN_TR_OFF) {
1382 return (-1);
1383 }
1384
1385 amp2 = svd2->amp;
1386 if (amp2) {
1387 pgcnt_t newpgs;
1388
1389 /*
1390 * Segment has private pages, can data structures
1391 * be expanded?
1392 *
1393 * Acquire the anon_map lock to prevent it from changing,
1394 * if it is shared. This ensures that the anon_map
1395 * will not change while a thread which has a read/write
1396 * lock on an address space references it.
1397 *
1398 * XXX - Don't need the anon_map lock at all if "refcnt"
1399 * is 1.
1400 */
1401 if (svd2->type == MAP_SHARED)
1402 return (-1);
1403
1404 ANON_LOCK_ENTER(&2->a_rwlock, RW_WRITER);
1405 if (amp2->refcnt > 1) {
1406 ANON_LOCK_EXIT(&2->a_rwlock);
1407 return (-1);
1408 }
1409 newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1410 btop(seg2->s_size), btop(seg1->s_size),
1411 ANON_NOSLEEP | ANON_GROWDOWN);
1412
1413 if (newpgs == 0) {
1414 ANON_LOCK_EXIT(&2->a_rwlock);
1415 return (-1);
1416 }
1417 amp2->size = ptob(newpgs);
1418 ANON_LOCK_EXIT(&2->a_rwlock);
1419 }
1420 if (svd2->vpage != NULL) {
1421 struct vpage *vp, *evp;
1422 new_vpage =
1423 kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1424 KM_NOSLEEP);
1425 if (new_vpage == NULL) {
1426 /* Not merging segments so adjust anon_index back */
1427 if (amp2)
1428 svd2->anon_index += seg_pages(seg1);
1429 return (-1);
1430 }
1431 bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1432 vpgtob(seg_pages(seg2)));
1433 kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1434 svd2->vpage = new_vpage;
1435
1436 vp = new_vpage;
1437 evp = vp + seg_pages(seg1);
1438 for (; vp < evp; vp++)
1439 VPP_SETPROT(vp, a->prot);
1440 if (svd2->pageswap && swresv) {
1441 ASSERT(!(svd2->flags & MAP_NORESERVE));
1442 ASSERT(swresv == seg1->s_size);
1443 vp = new_vpage;
1444 for (; vp < evp; vp++) {
1445 VPP_SETSWAPRES(vp);
1446 }
1447 }
1448 }
1449 ASSERT(svd2->vpage != NULL || svd2->pageswap == 0);
1450 size = seg1->s_size;
1451 seg_free(seg1);
1452 seg2->s_size += size;
1453 seg2->s_base -= size;
1454 svd2->offset -= size;
1455 svd2->swresv += swresv;
1456 if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1457 svd2->type == MAP_SHARED && svd2->vp != NULL &&
1458 (svd2->vp->v_flag & VVMEXEC)) {
1459 ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1460 segvn_inval_trcache(svd2->vp);
1461 }
1462 return (0);
1463 }
1464
1465 static int
1466 segvn_dup(struct seg *seg, struct seg *newseg)
1467 {
1468 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1469 struct segvn_data *newsvd;
1470 pgcnt_t npages = seg_pages(seg);
1471 int error = 0;
1472 uint_t prot;
1473 size_t len;
1474 struct anon_map *amp;
1475
1476 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1477 ASSERT(newseg->s_as->a_proc->p_parent == curproc);
1478
1479 /*
1480 * If segment has anon reserved, reserve more for the new seg.
1481 * For a MAP_NORESERVE segment swresv will be a count of all the
1482 * allocated anon slots; thus we reserve for the child as many slots
1483 * as the parent has allocated. This semantic prevents the child or
1484 * parent from dieing during a copy-on-write fault caused by trying
1485 * to write a shared pre-existing anon page.
1486 */
1487 if ((len = svd->swresv) != 0) {
1488 if (anon_resv(svd->swresv) == 0)
1489 return (ENOMEM);
1490
1491 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1492 seg, len, 0);
1493 }
1494
1495 newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1496
1497 newseg->s_ops = &segvn_ops;
1498 newseg->s_data = (void *)newsvd;
1499 newseg->s_szc = seg->s_szc;
1500
1501 newsvd->seg = newseg;
1502 if ((newsvd->vp = svd->vp) != NULL) {
1503 VN_HOLD(svd->vp);
1504 if (svd->type == MAP_SHARED)
1505 lgrp_shm_policy_init(NULL, svd->vp);
1506 }
1507 newsvd->offset = svd->offset;
1508 newsvd->prot = svd->prot;
1509 newsvd->maxprot = svd->maxprot;
1510 newsvd->pageprot = svd->pageprot;
1511 newsvd->type = svd->type;
1512 newsvd->cred = svd->cred;
1513 crhold(newsvd->cred);
1514 newsvd->advice = svd->advice;
1515 newsvd->pageadvice = svd->pageadvice;
1516 newsvd->swresv = svd->swresv;
1517 newsvd->pageswap = svd->pageswap;
1518 newsvd->flags = svd->flags;
1519 newsvd->softlockcnt = 0;
1520 newsvd->softlockcnt_sbase = 0;
1521 newsvd->softlockcnt_send = 0;
1522 newsvd->policy_info = svd->policy_info;
1523 newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1524
1525 if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1526 /*
1527 * Not attaching to a shared anon object.
1528 */
1529 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1530 svd->tr_state == SEGVN_TR_OFF);
1531 if (svd->tr_state == SEGVN_TR_ON) {
1532 ASSERT(newsvd->vp != NULL && amp != NULL);
1533 newsvd->tr_state = SEGVN_TR_INIT;
1534 } else {
1535 newsvd->tr_state = svd->tr_state;
1536 }
1537 newsvd->amp = NULL;
1538 newsvd->anon_index = 0;
1539 } else {
1540 /* regions for now are only used on pure vnode segments */
1541 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1542 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1543 newsvd->tr_state = SEGVN_TR_OFF;
1544 if (svd->type == MAP_SHARED) {
1545 newsvd->amp = amp;
1546 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
1547 amp->refcnt++;
1548 ANON_LOCK_EXIT(&->a_rwlock);
1549 newsvd->anon_index = svd->anon_index;
1550 } else {
1551 int reclaim = 1;
1552
1553 /*
1554 * Allocate and initialize new anon_map structure.
1555 */
1556 newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1557 ANON_SLEEP);
1558 newsvd->amp->a_szc = newseg->s_szc;
1559 newsvd->anon_index = 0;
1560
1561 /*
1562 * We don't have to acquire the anon_map lock
1563 * for the new segment (since it belongs to an
1564 * address space that is still not associated
1565 * with any process), or the segment in the old
1566 * address space (since all threads in it
1567 * are stopped while duplicating the address space).
1568 */
1569
1570 /*
1571 * The goal of the following code is to make sure that
1572 * softlocked pages do not end up as copy on write
1573 * pages. This would cause problems where one
1574 * thread writes to a page that is COW and a different
1575 * thread in the same process has softlocked it. The
1576 * softlock lock would move away from this process
1577 * because the write would cause this process to get
1578 * a copy (without the softlock).
1579 *
1580 * The strategy here is to just break the
1581 * sharing on pages that could possibly be
1582 * softlocked.
1583 */
1584 retry:
1585 if (svd->softlockcnt) {
1586 struct anon *ap, *newap;
1587 size_t i;
1588 uint_t vpprot;
1589 page_t *anon_pl[1+1], *pp;
1590 caddr_t addr;
1591 ulong_t old_idx = svd->anon_index;
1592 ulong_t new_idx = 0;
1593
1594 /*
1595 * The softlock count might be non zero
1596 * because some pages are still stuck in the
1597 * cache for lazy reclaim. Flush the cache
1598 * now. This should drop the count to zero.
1599 * [or there is really I/O going on to these
1600 * pages]. Note, we have the writers lock so
1601 * nothing gets inserted during the flush.
1602 */
1603 if (reclaim == 1) {
1604 segvn_purge(seg);
1605 reclaim = 0;
1606 goto retry;
1607 }
1608 i = btopr(seg->s_size);
1609 addr = seg->s_base;
1610 /*
1611 * XXX break cow sharing using PAGESIZE
1612 * pages. They will be relocated into larger
1613 * pages at fault time.
1614 */
1615 while (i-- > 0) {
1616 if (ap = anon_get_ptr(amp->ahp,
1617 old_idx)) {
1618 error = anon_getpage(&ap,
1619 &vpprot, anon_pl, PAGESIZE,
1620 seg, addr, S_READ,
1621 svd->cred);
1622 if (error) {
1623 newsvd->vpage = NULL;
1624 goto out;
1625 }
1626 /*
1627 * prot need not be computed
1628 * below 'cause anon_private is
1629 * going to ignore it anyway
1630 * as child doesn't inherit
1631 * pagelock from parent.
1632 */
1633 prot = svd->pageprot ?
1634 VPP_PROT(
1635 &svd->vpage[
1636 seg_page(seg, addr)])
1637 : svd->prot;
1638 pp = anon_private(&newap,
1639 newseg, addr, prot,
1640 anon_pl[0], 0,
1641 newsvd->cred);
1642 if (pp == NULL) {
1643 /* no mem abort */
1644 newsvd->vpage = NULL;
1645 error = ENOMEM;
1646 goto out;
1647 }
1648 (void) anon_set_ptr(
1649 newsvd->amp->ahp, new_idx,
1650 newap, ANON_SLEEP);
1651 page_unlock(pp);
1652 }
1653 addr += PAGESIZE;
1654 old_idx++;
1655 new_idx++;
1656 }
1657 } else { /* common case */
1658 if (seg->s_szc != 0) {
1659 /*
1660 * If at least one of anon slots of a
1661 * large page exists then make sure
1662 * all anon slots of a large page
1663 * exist to avoid partial cow sharing
1664 * of a large page in the future.
1665 */
1666 anon_dup_fill_holes(amp->ahp,
1667 svd->anon_index, newsvd->amp->ahp,
1668 0, seg->s_size, seg->s_szc,
1669 svd->vp != NULL);
1670 } else {
1671 anon_dup(amp->ahp, svd->anon_index,
1672 newsvd->amp->ahp, 0, seg->s_size);
1673 }
1674
1675 hat_clrattr(seg->s_as->a_hat, seg->s_base,
1676 seg->s_size, PROT_WRITE);
1677 }
1678 }
1679 }
1680 /*
1681 * If necessary, create a vpage structure for the new segment.
1682 * Do not copy any page lock indications.
1683 */
1684 if (svd->vpage != NULL) {
1685 uint_t i;
1686 struct vpage *ovp = svd->vpage;
1687 struct vpage *nvp;
1688
1689 nvp = newsvd->vpage =
1690 kmem_alloc(vpgtob(npages), KM_SLEEP);
1691 for (i = 0; i < npages; i++) {
1692 *nvp = *ovp++;
1693 VPP_CLRPPLOCK(nvp++);
1694 }
1695 } else
1696 newsvd->vpage = NULL;
1697
1698 /* Inform the vnode of the new mapping */
1699 if (newsvd->vp != NULL) {
1700 error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1701 newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1702 newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1703 }
1704 out:
1705 if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1706 ASSERT(newsvd->amp == NULL);
1707 ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1708 newsvd->rcookie = svd->rcookie;
1709 hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1710 }
1711 return (error);
1712 }
1713
1714
1715 /*
1716 * callback function to invoke free_vp_pages() for only those pages actually
1717 * processed by the HAT when a shared region is destroyed.
1718 */
1719 extern int free_pages;
1720
1721 static void
1722 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1723 size_t r_size, void *r_obj, u_offset_t r_objoff)
1724 {
1725 u_offset_t off;
1726 size_t len;
1727 vnode_t *vp = (vnode_t *)r_obj;
1728
1729 ASSERT(eaddr > saddr);
1730 ASSERT(saddr >= r_saddr);
1731 ASSERT(saddr < r_saddr + r_size);
1732 ASSERT(eaddr > r_saddr);
1733 ASSERT(eaddr <= r_saddr + r_size);
1734 ASSERT(vp != NULL);
1735
1736 if (!free_pages) {
1737 return;
1738 }
1739
1740 len = eaddr - saddr;
1741 off = (saddr - r_saddr) + r_objoff;
1742 free_vp_pages(vp, off, len);
1743 }
1744
1745 /*
1746 * callback function used by segvn_unmap to invoke free_vp_pages() for only
1747 * those pages actually processed by the HAT
1748 */
1749 static void
1750 segvn_hat_unload_callback(hat_callback_t *cb)
1751 {
1752 struct seg *seg = cb->hcb_data;
1753 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1754 size_t len;
1755 u_offset_t off;
1756
1757 ASSERT(svd->vp != NULL);
1758 ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1759 ASSERT(cb->hcb_start_addr >= seg->s_base);
1760
1761 len = cb->hcb_end_addr - cb->hcb_start_addr;
1762 off = cb->hcb_start_addr - seg->s_base;
1763 free_vp_pages(svd->vp, svd->offset + off, len);
1764 }
1765
1766 /*
1767 * This function determines the number of bytes of swap reserved by
1768 * a segment for which per-page accounting is present. It is used to
1769 * calculate the correct value of a segvn_data's swresv.
1770 */
1771 static size_t
1772 segvn_count_swap_by_vpages(struct seg *seg)
1773 {
1774 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1775 struct vpage *vp, *evp;
1776 size_t nswappages = 0;
1777
1778 ASSERT(svd->pageswap);
1779 ASSERT(svd->vpage != NULL);
1780
1781 evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
1782
1783 for (vp = svd->vpage; vp < evp; vp++) {
1784 if (VPP_ISSWAPRES(vp))
1785 nswappages++;
1786 }
1787
1788 return (nswappages << PAGESHIFT);
1789 }
1790
1791 static int
1792 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1793 {
1794 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1795 struct segvn_data *nsvd;
1796 struct seg *nseg;
1797 struct anon_map *amp;
1798 pgcnt_t opages; /* old segment size in pages */
1799 pgcnt_t npages; /* new segment size in pages */
1800 pgcnt_t dpages; /* pages being deleted (unmapped) */
1801 hat_callback_t callback; /* used for free_vp_pages() */
1802 hat_callback_t *cbp = NULL;
1803 caddr_t nbase;
1804 size_t nsize;
1805 size_t oswresv;
1806 int reclaim = 1;
1807
1808 /*
1809 * We don't need any segment level locks for "segvn" data
1810 * since the address space is "write" locked.
1811 */
1812 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1813
1814 /*
1815 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1816 * softlockcnt is protected from change by the as write lock.
1817 */
1818 retry:
1819 if (svd->softlockcnt > 0) {
1820 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1821
1822 /*
1823 * If this is shared segment non 0 softlockcnt
1824 * means locked pages are still in use.
1825 */
1826 if (svd->type == MAP_SHARED) {
1827 return (EAGAIN);
1828 }
1829
1830 /*
1831 * since we do have the writers lock nobody can fill
1832 * the cache during the purge. The flush either succeeds
1833 * or we still have pending I/Os.
1834 */
1835 if (reclaim == 1) {
1836 segvn_purge(seg);
1837 reclaim = 0;
1838 goto retry;
1839 }
1840 return (EAGAIN);
1841 }
1842
1843 /*
1844 * Check for bad sizes
1845 */
1846 if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1847 (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1848 panic("segvn_unmap");
1849 /*NOTREACHED*/
1850 }
1851
1852 if (seg->s_szc != 0) {
1853 size_t pgsz = page_get_pagesize(seg->s_szc);
1854 int err;
1855 if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1856 ASSERT(seg->s_base != addr || seg->s_size != len);
1857 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1858 ASSERT(svd->amp == NULL);
1859 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1860 hat_leave_region(seg->s_as->a_hat,
1861 svd->rcookie, HAT_REGION_TEXT);
1862 svd->rcookie = HAT_INVALID_REGION_COOKIE;
1863 /*
1864 * could pass a flag to segvn_demote_range()
1865 * below to tell it not to do any unloads but
1866 * this case is rare enough to not bother for
1867 * now.
1868 */
1869 } else if (svd->tr_state == SEGVN_TR_INIT) {
1870 svd->tr_state = SEGVN_TR_OFF;
1871 } else if (svd->tr_state == SEGVN_TR_ON) {
1872 ASSERT(svd->amp != NULL);
1873 segvn_textunrepl(seg, 1);
1874 ASSERT(svd->amp == NULL);
1875 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1876 }
1877 VM_STAT_ADD(segvnvmstats.demoterange[0]);
1878 err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1879 if (err == 0) {
1880 return (IE_RETRY);
1881 }
1882 return (err);
1883 }
1884 }
1885
1886 /* Inform the vnode of the unmapping. */
1887 if (svd->vp) {
1888 int error;
1889
1890 error = VOP_DELMAP(svd->vp,
1891 (offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1892 seg->s_as, addr, len, svd->prot, svd->maxprot,
1893 svd->type, svd->cred, NULL);
1894
1895 if (error == EAGAIN)
1896 return (error);
1897 }
1898
1899 /*
1900 * Remove any page locks set through this mapping.
1901 * If text replication is not off no page locks could have been
1902 * established via this mapping.
1903 */
1904 if (svd->tr_state == SEGVN_TR_OFF) {
1905 (void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1906 }
1907
1908 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1909 ASSERT(svd->amp == NULL);
1910 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1911 ASSERT(svd->type == MAP_PRIVATE);
1912 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1913 HAT_REGION_TEXT);
1914 svd->rcookie = HAT_INVALID_REGION_COOKIE;
1915 } else if (svd->tr_state == SEGVN_TR_ON) {
1916 ASSERT(svd->amp != NULL);
1917 ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1918 segvn_textunrepl(seg, 1);
1919 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1920 } else {
1921 if (svd->tr_state != SEGVN_TR_OFF) {
1922 ASSERT(svd->tr_state == SEGVN_TR_INIT);
1923 svd->tr_state = SEGVN_TR_OFF;
1924 }
1925 /*
1926 * Unload any hardware translations in the range to be taken
1927 * out. Use a callback to invoke free_vp_pages() effectively.
1928 */
1929 if (svd->vp != NULL && free_pages != 0) {
1930 callback.hcb_data = seg;
1931 callback.hcb_function = segvn_hat_unload_callback;
1932 cbp = &callback;
1933 }
1934 hat_unload_callback(seg->s_as->a_hat, addr, len,
1935 HAT_UNLOAD_UNMAP, cbp);
1936
1937 if (svd->type == MAP_SHARED && svd->vp != NULL &&
1938 (svd->vp->v_flag & VVMEXEC) &&
1939 ((svd->prot & PROT_WRITE) || svd->pageprot)) {
1940 segvn_inval_trcache(svd->vp);
1941 }
1942 }
1943
1944 /*
1945 * Check for entire segment
1946 */
1947 if (addr == seg->s_base && len == seg->s_size) {
1948 seg_free(seg);
1949 return (0);
1950 }
1951
1952 opages = seg_pages(seg);
1953 dpages = btop(len);
1954 npages = opages - dpages;
1955 amp = svd->amp;
1956 ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
1957
1958 /*
1959 * Check for beginning of segment
1960 */
1961 if (addr == seg->s_base) {
1962 if (svd->vpage != NULL) {
1963 size_t nbytes;
1964 struct vpage *ovpage;
1965
1966 ovpage = svd->vpage; /* keep pointer to vpage */
1967
1968 nbytes = vpgtob(npages);
1969 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1970 bcopy(&ovpage[dpages], svd->vpage, nbytes);
1971
1972 /* free up old vpage */
1973 kmem_free(ovpage, vpgtob(opages));
1974 }
1975 if (amp != NULL) {
1976 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
1977 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
1978 /*
1979 * Shared anon map is no longer in use. Before
1980 * freeing its pages purge all entries from
1981 * pcache that belong to this amp.
1982 */
1983 if (svd->type == MAP_SHARED) {
1984 ASSERT(amp->refcnt == 1);
1985 ASSERT(svd->softlockcnt == 0);
1986 anonmap_purge(amp);
1987 }
1988 /*
1989 * Free up now unused parts of anon_map array.
1990 */
1991 if (amp->a_szc == seg->s_szc) {
1992 if (seg->s_szc != 0) {
1993 anon_free_pages(amp->ahp,
1994 svd->anon_index, len,
1995 seg->s_szc);
1996 } else {
1997 anon_free(amp->ahp,
1998 svd->anon_index,
1999 len);
2000 }
2001 } else {
2002 ASSERT(svd->type == MAP_SHARED);
2003 ASSERT(amp->a_szc > seg->s_szc);
2004 anon_shmap_free_pages(amp,
2005 svd->anon_index, len);
2006 }
2007
2008 /*
2009 * Unreserve swap space for the
2010 * unmapped chunk of this segment in
2011 * case it's MAP_SHARED
2012 */
2013 if (svd->type == MAP_SHARED) {
2014 anon_unresv_zone(len,
2015 seg->s_as->a_proc->p_zone);
2016 amp->swresv -= len;
2017 }
2018 }
2019 ANON_LOCK_EXIT(&->a_rwlock);
2020 svd->anon_index += dpages;
2021 }
2022 if (svd->vp != NULL)
2023 svd->offset += len;
2024
2025 seg->s_base += len;
2026 seg->s_size -= len;
2027
2028 if (svd->swresv) {
2029 if (svd->flags & MAP_NORESERVE) {
2030 ASSERT(amp);
2031 oswresv = svd->swresv;
2032
2033 svd->swresv = ptob(anon_pages(amp->ahp,
2034 svd->anon_index, npages));
2035 anon_unresv_zone(oswresv - svd->swresv,
2036 seg->s_as->a_proc->p_zone);
2037 if (SEG_IS_PARTIAL_RESV(seg))
2038 seg->s_as->a_resvsize -= oswresv -
2039 svd->swresv;
2040 } else {
2041 size_t unlen;
2042
2043 if (svd->pageswap) {
2044 oswresv = svd->swresv;
2045 svd->swresv =
2046 segvn_count_swap_by_vpages(seg);
2047 ASSERT(oswresv >= svd->swresv);
2048 unlen = oswresv - svd->swresv;
2049 } else {
2050 svd->swresv -= len;
2051 ASSERT(svd->swresv == seg->s_size);
2052 unlen = len;
2053 }
2054 anon_unresv_zone(unlen,
2055 seg->s_as->a_proc->p_zone);
2056 }
2057 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2058 seg, len, 0);
2059 }
2060
2061 return (0);
2062 }
2063
2064 /*
2065 * Check for end of segment
2066 */
2067 if (addr + len == seg->s_base + seg->s_size) {
2068 if (svd->vpage != NULL) {
2069 size_t nbytes;
2070 struct vpage *ovpage;
2071
2072 ovpage = svd->vpage; /* keep pointer to vpage */
2073
2074 nbytes = vpgtob(npages);
2075 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2076 bcopy(ovpage, svd->vpage, nbytes);
2077
2078 /* free up old vpage */
2079 kmem_free(ovpage, vpgtob(opages));
2080
2081 }
2082 if (amp != NULL) {
2083 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
2084 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2085 /*
2086 * Free up now unused parts of anon_map array.
2087 */
2088 ulong_t an_idx = svd->anon_index + npages;
2089
2090 /*
2091 * Shared anon map is no longer in use. Before
2092 * freeing its pages purge all entries from
2093 * pcache that belong to this amp.
2094 */
2095 if (svd->type == MAP_SHARED) {
2096 ASSERT(amp->refcnt == 1);
2097 ASSERT(svd->softlockcnt == 0);
2098 anonmap_purge(amp);
2099 }
2100
2101 if (amp->a_szc == seg->s_szc) {
2102 if (seg->s_szc != 0) {
2103 anon_free_pages(amp->ahp,
2104 an_idx, len,
2105 seg->s_szc);
2106 } else {
2107 anon_free(amp->ahp, an_idx,
2108 len);
2109 }
2110 } else {
2111 ASSERT(svd->type == MAP_SHARED);
2112 ASSERT(amp->a_szc > seg->s_szc);
2113 anon_shmap_free_pages(amp,
2114 an_idx, len);
2115 }
2116
2117 /*
2118 * Unreserve swap space for the
2119 * unmapped chunk of this segment in
2120 * case it's MAP_SHARED
2121 */
2122 if (svd->type == MAP_SHARED) {
2123 anon_unresv_zone(len,
2124 seg->s_as->a_proc->p_zone);
2125 amp->swresv -= len;
2126 }
2127 }
2128 ANON_LOCK_EXIT(&->a_rwlock);
2129 }
2130
2131 seg->s_size -= len;
2132
2133 if (svd->swresv) {
2134 if (svd->flags & MAP_NORESERVE) {
2135 ASSERT(amp);
2136 oswresv = svd->swresv;
2137 svd->swresv = ptob(anon_pages(amp->ahp,
2138 svd->anon_index, npages));
2139 anon_unresv_zone(oswresv - svd->swresv,
2140 seg->s_as->a_proc->p_zone);
2141 if (SEG_IS_PARTIAL_RESV(seg))
2142 seg->s_as->a_resvsize -= oswresv -
2143 svd->swresv;
2144 } else {
2145 size_t unlen;
2146
2147 if (svd->pageswap) {
2148 oswresv = svd->swresv;
2149 svd->swresv =
2150 segvn_count_swap_by_vpages(seg);
2151 ASSERT(oswresv >= svd->swresv);
2152 unlen = oswresv - svd->swresv;
2153 } else {
2154 svd->swresv -= len;
2155 ASSERT(svd->swresv == seg->s_size);
2156 unlen = len;
2157 }
2158 anon_unresv_zone(unlen,
2159 seg->s_as->a_proc->p_zone);
2160 }
2161 TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2162 "anon proc:%p %lu %u", seg, len, 0);
2163 }
2164
2165 return (0);
2166 }
2167
2168 /*
2169 * The section to go is in the middle of the segment,
2170 * have to make it into two segments. nseg is made for
2171 * the high end while seg is cut down at the low end.
2172 */
2173 nbase = addr + len; /* new seg base */
2174 nsize = (seg->s_base + seg->s_size) - nbase; /* new seg size */
2175 seg->s_size = addr - seg->s_base; /* shrink old seg */
2176 nseg = seg_alloc(seg->s_as, nbase, nsize);
2177 if (nseg == NULL) {
2178 panic("segvn_unmap seg_alloc");
2179 /*NOTREACHED*/
2180 }
2181 nseg->s_ops = seg->s_ops;
2182 nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2183 nseg->s_data = (void *)nsvd;
2184 nseg->s_szc = seg->s_szc;
2185 *nsvd = *svd;
2186 nsvd->seg = nseg;
2187 nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2188 nsvd->swresv = 0;
2189 nsvd->softlockcnt = 0;
2190 nsvd->softlockcnt_sbase = 0;
2191 nsvd->softlockcnt_send = 0;
2192 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2193
2194 if (svd->vp != NULL) {
2195 VN_HOLD(nsvd->vp);
2196 if (nsvd->type == MAP_SHARED)
2197 lgrp_shm_policy_init(NULL, nsvd->vp);
2198 }
2199 crhold(svd->cred);
2200
2201 if (svd->vpage == NULL) {
2202 nsvd->vpage = NULL;
2203 } else {
2204 /* need to split vpage into two arrays */
2205 size_t nbytes;
2206 struct vpage *ovpage;
2207
2208 ovpage = svd->vpage; /* keep pointer to vpage */
2209
2210 npages = seg_pages(seg); /* seg has shrunk */
2211 nbytes = vpgtob(npages);
2212 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2213
2214 bcopy(ovpage, svd->vpage, nbytes);
2215
2216 npages = seg_pages(nseg);
2217 nbytes = vpgtob(npages);
2218 nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2219
2220 bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2221
2222 /* free up old vpage */
2223 kmem_free(ovpage, vpgtob(opages));
2224 }
2225
2226 if (amp == NULL) {
2227 nsvd->amp = NULL;
2228 nsvd->anon_index = 0;
2229 } else {
2230 /*
2231 * Need to create a new anon map for the new segment.
2232 * We'll also allocate a new smaller array for the old
2233 * smaller segment to save space.
2234 */
2235 opages = btop((uintptr_t)(addr - seg->s_base));
2236 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
2237 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2238 /*
2239 * Free up now unused parts of anon_map array.
2240 */
2241 ulong_t an_idx = svd->anon_index + opages;
2242
2243 /*
2244 * Shared anon map is no longer in use. Before
2245 * freeing its pages purge all entries from
2246 * pcache that belong to this amp.
2247 */
2248 if (svd->type == MAP_SHARED) {
2249 ASSERT(amp->refcnt == 1);
2250 ASSERT(svd->softlockcnt == 0);
2251 anonmap_purge(amp);
2252 }
2253
2254 if (amp->a_szc == seg->s_szc) {
2255 if (seg->s_szc != 0) {
2256 anon_free_pages(amp->ahp, an_idx, len,
2257 seg->s_szc);
2258 } else {
2259 anon_free(amp->ahp, an_idx,
2260 len);
2261 }
2262 } else {
2263 ASSERT(svd->type == MAP_SHARED);
2264 ASSERT(amp->a_szc > seg->s_szc);
2265 anon_shmap_free_pages(amp, an_idx, len);
2266 }
2267
2268 /*
2269 * Unreserve swap space for the
2270 * unmapped chunk of this segment in
2271 * case it's MAP_SHARED
2272 */
2273 if (svd->type == MAP_SHARED) {
2274 anon_unresv_zone(len,
2275 seg->s_as->a_proc->p_zone);
2276 amp->swresv -= len;
2277 }
2278 }
2279 nsvd->anon_index = svd->anon_index +
2280 btop((uintptr_t)(nseg->s_base - seg->s_base));
2281 if (svd->type == MAP_SHARED) {
2282 amp->refcnt++;
2283 nsvd->amp = amp;
2284 } else {
2285 struct anon_map *namp;
2286 struct anon_hdr *nahp;
2287
2288 ASSERT(svd->type == MAP_PRIVATE);
2289 nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2290 namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2291 namp->a_szc = seg->s_szc;
2292 (void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2293 0, btop(seg->s_size), ANON_SLEEP);
2294 (void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2295 namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2296 anon_release(amp->ahp, btop(amp->size));
2297 svd->anon_index = 0;
2298 nsvd->anon_index = 0;
2299 amp->ahp = nahp;
2300 amp->size = seg->s_size;
2301 nsvd->amp = namp;
2302 }
2303 ANON_LOCK_EXIT(&->a_rwlock);
2304 }
2305 if (svd->swresv) {
2306 if (svd->flags & MAP_NORESERVE) {
2307 ASSERT(amp);
2308 oswresv = svd->swresv;
2309 svd->swresv = ptob(anon_pages(amp->ahp,
2310 svd->anon_index, btop(seg->s_size)));
2311 nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2312 nsvd->anon_index, btop(nseg->s_size)));
2313 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2314 anon_unresv_zone(oswresv - (svd->swresv + nsvd->swresv),
2315 seg->s_as->a_proc->p_zone);
2316 if (SEG_IS_PARTIAL_RESV(seg))
2317 seg->s_as->a_resvsize -= oswresv -
2318 (svd->swresv + nsvd->swresv);
2319 } else {
2320 size_t unlen;
2321
2322 if (svd->pageswap) {
2323 oswresv = svd->swresv;
2324 svd->swresv = segvn_count_swap_by_vpages(seg);
2325 nsvd->swresv = segvn_count_swap_by_vpages(nseg);
2326 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2327 unlen = oswresv - (svd->swresv + nsvd->swresv);
2328 } else {
2329 if (seg->s_size + nseg->s_size + len !=
2330 svd->swresv) {
2331 panic("segvn_unmap: cannot split "
2332 "swap reservation");
2333 /*NOTREACHED*/
2334 }
2335 svd->swresv = seg->s_size;
2336 nsvd->swresv = nseg->s_size;
2337 unlen = len;
2338 }
2339 anon_unresv_zone(unlen,
2340 seg->s_as->a_proc->p_zone);
2341 }
2342 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2343 seg, len, 0);
2344 }
2345
2346 return (0); /* I'm glad that's all over with! */
2347 }
2348
2349 static void
2350 segvn_free(struct seg *seg)
2351 {
2352 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2353 pgcnt_t npages = seg_pages(seg);
2354 struct anon_map *amp;
2355 size_t len;
2356
2357 /*
2358 * We don't need any segment level locks for "segvn" data
2359 * since the address space is "write" locked.
2360 */
2361 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
2362 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2363
2364 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2365
2366 /*
2367 * Be sure to unlock pages. XXX Why do things get free'ed instead
2368 * of unmapped? XXX
2369 */
2370 (void) segvn_lockop(seg, seg->s_base, seg->s_size,
2371 0, MC_UNLOCK, NULL, 0);
2372
2373 /*
2374 * Deallocate the vpage and anon pointers if necessary and possible.
2375 */
2376 if (svd->vpage != NULL) {
2377 kmem_free(svd->vpage, vpgtob(npages));
2378 svd->vpage = NULL;
2379 }
2380 if ((amp = svd->amp) != NULL) {
2381 /*
2382 * If there are no more references to this anon_map
2383 * structure, then deallocate the structure after freeing
2384 * up all the anon slot pointers that we can.
2385 */
2386 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
2387 ASSERT(amp->a_szc >= seg->s_szc);
2388 if (--amp->refcnt == 0) {
2389 if (svd->type == MAP_PRIVATE) {
2390 /*
2391 * Private - we only need to anon_free
2392 * the part that this segment refers to.
2393 */
2394 if (seg->s_szc != 0) {
2395 anon_free_pages(amp->ahp,
2396 svd->anon_index, seg->s_size,
2397 seg->s_szc);
2398 } else {
2399 anon_free(amp->ahp, svd->anon_index,
2400 seg->s_size);
2401 }
2402 } else {
2403
2404 /*
2405 * Shared anon map is no longer in use. Before
2406 * freeing its pages purge all entries from
2407 * pcache that belong to this amp.
2408 */
2409 ASSERT(svd->softlockcnt == 0);
2410 anonmap_purge(amp);
2411
2412 /*
2413 * Shared - anon_free the entire
2414 * anon_map's worth of stuff and
2415 * release any swap reservation.
2416 */
2417 if (amp->a_szc != 0) {
2418 anon_shmap_free_pages(amp, 0,
2419 amp->size);
2420 } else {
2421 anon_free(amp->ahp, 0, amp->size);
2422 }
2423 if ((len = amp->swresv) != 0) {
2424 anon_unresv_zone(len,
2425 seg->s_as->a_proc->p_zone);
2426 TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2427 "anon proc:%p %lu %u", seg, len, 0);
2428 }
2429 }
2430 svd->amp = NULL;
2431 ANON_LOCK_EXIT(&->a_rwlock);
2432 anonmap_free(amp);
2433 } else if (svd->type == MAP_PRIVATE) {
2434 /*
2435 * We had a private mapping which still has
2436 * a held anon_map so just free up all the
2437 * anon slot pointers that we were using.
2438 */
2439 if (seg->s_szc != 0) {
2440 anon_free_pages(amp->ahp, svd->anon_index,
2441 seg->s_size, seg->s_szc);
2442 } else {
2443 anon_free(amp->ahp, svd->anon_index,
2444 seg->s_size);
2445 }
2446 ANON_LOCK_EXIT(&->a_rwlock);
2447 } else {
2448 ANON_LOCK_EXIT(&->a_rwlock);
2449 }
2450 }
2451
2452 /*
2453 * Release swap reservation.
2454 */
2455 if ((len = svd->swresv) != 0) {
2456 anon_unresv_zone(svd->swresv,
2457 seg->s_as->a_proc->p_zone);
2458 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2459 seg, len, 0);
2460 if (SEG_IS_PARTIAL_RESV(seg))
2461 seg->s_as->a_resvsize -= svd->swresv;
2462 svd->swresv = 0;
2463 }
2464 /*
2465 * Release claim on vnode, credentials, and finally free the
2466 * private data.
2467 */
2468 if (svd->vp != NULL) {
2469 if (svd->type == MAP_SHARED)
2470 lgrp_shm_policy_fini(NULL, svd->vp);
2471 VN_RELE(svd->vp);
2472 svd->vp = NULL;
2473 }
2474 crfree(svd->cred);
2475 svd->pageprot = 0;
2476 svd->pageadvice = 0;
2477 svd->pageswap = 0;
2478 svd->cred = NULL;
2479
2480 /*
2481 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2482 * still working with this segment without holding as lock (in case
2483 * it's called by pcache async thread).
2484 */
2485 ASSERT(svd->softlockcnt == 0);
2486 mutex_enter(&svd->segfree_syncmtx);
2487 mutex_exit(&svd->segfree_syncmtx);
2488
2489 seg->s_data = NULL;
2490 kmem_cache_free(segvn_cache, svd);
2491 }
2492
2493 /*
2494 * Do a F_SOFTUNLOCK call over the range requested. The range must have
2495 * already been F_SOFTLOCK'ed.
2496 * Caller must always match addr and len of a softunlock with a previous
2497 * softlock with exactly the same addr and len.
2498 */
2499 static void
2500 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2501 {
2502 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2503 page_t *pp;
2504 caddr_t adr;
2505 struct vnode *vp;
2506 u_offset_t offset;
2507 ulong_t anon_index;
2508 struct anon_map *amp;
2509 struct anon *ap = NULL;
2510
2511 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
2512 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2513
2514 if ((amp = svd->amp) != NULL)
2515 anon_index = svd->anon_index + seg_page(seg, addr);
2516
2517 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2518 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2519 hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2520 } else {
2521 hat_unlock(seg->s_as->a_hat, addr, len);
2522 }
2523 for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2524 if (amp != NULL) {
2525 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
2526 if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2527 != NULL) {
2528 swap_xlate(ap, &vp, &offset);
2529 } else {
2530 vp = svd->vp;
2531 offset = svd->offset +
2532 (uintptr_t)(adr - seg->s_base);
2533 }
2534 ANON_LOCK_EXIT(&->a_rwlock);
2535 } else {
2536 vp = svd->vp;
2537 offset = svd->offset +
2538 (uintptr_t)(adr - seg->s_base);
2539 }
2540
2541 /*
2542 * Use page_find() instead of page_lookup() to
2543 * find the page since we know that it is locked.
2544 */
2545 pp = page_find(vp, offset);
2546 if (pp == NULL) {
2547 panic(
2548 "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2549 (void *)adr, (void *)ap, (void *)vp, offset);
2550 /*NOTREACHED*/
2551 }
2552
2553 if (rw == S_WRITE) {
2554 hat_setrefmod(pp);
2555 if (seg->s_as->a_vbits)
2556 hat_setstat(seg->s_as, adr, PAGESIZE,
2557 P_REF | P_MOD);
2558 } else if (rw != S_OTHER) {
2559 hat_setref(pp);
2560 if (seg->s_as->a_vbits)
2561 hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2562 }
2563 TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2564 "segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2565 page_unlock(pp);
2566 }
2567 ASSERT(svd->softlockcnt >= btop(len));
2568 if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -btop(len))) {
2569 /*
2570 * All SOFTLOCKS are gone. Wakeup any waiting
2571 * unmappers so they can try again to unmap.
2572 * Check for waiters first without the mutex
2573 * held so we don't always grab the mutex on
2574 * softunlocks.
2575 */
2576 if (AS_ISUNMAPWAIT(seg->s_as)) {
2577 mutex_enter(&seg->s_as->a_contents);
2578 if (AS_ISUNMAPWAIT(seg->s_as)) {
2579 AS_CLRUNMAPWAIT(seg->s_as);
2580 cv_broadcast(&seg->s_as->a_cv);
2581 }
2582 mutex_exit(&seg->s_as->a_contents);
2583 }
2584 }
2585 }
2586
2587 #define PAGE_HANDLED ((page_t *)-1)
2588
2589 /*
2590 * Release all the pages in the NULL terminated ppp list
2591 * which haven't already been converted to PAGE_HANDLED.
2592 */
2593 static void
2594 segvn_pagelist_rele(page_t **ppp)
2595 {
2596 for (; *ppp != NULL; ppp++) {
2597 if (*ppp != PAGE_HANDLED)
2598 page_unlock(*ppp);
2599 }
2600 }
2601
2602 static int stealcow = 1;
2603
2604 /*
2605 * Workaround for viking chip bug. See bug id 1220902.
2606 * To fix this down in pagefault() would require importing so
2607 * much as and segvn code as to be unmaintainable.
2608 */
2609 int enable_mbit_wa = 0;
2610
2611 /*
2612 * Handles all the dirty work of getting the right
2613 * anonymous pages and loading up the translations.
2614 * This routine is called only from segvn_fault()
2615 * when looping over the range of addresses requested.
2616 *
2617 * The basic algorithm here is:
2618 * If this is an anon_zero case
2619 * Call anon_zero to allocate page
2620 * Load up translation
2621 * Return
2622 * endif
2623 * If this is an anon page
2624 * Use anon_getpage to get the page
2625 * else
2626 * Find page in pl[] list passed in
2627 * endif
2628 * If not a cow
2629 * Load up the translation to the page
2630 * return
2631 * endif
2632 * Call anon_private to handle cow
2633 * Load up (writable) translation to new page
2634 */
2635 static faultcode_t
2636 segvn_faultpage(
2637 struct hat *hat, /* the hat to use for mapping */
2638 struct seg *seg, /* seg_vn of interest */
2639 caddr_t addr, /* address in as */
2640 u_offset_t off, /* offset in vp */
2641 struct vpage *vpage, /* pointer to vpage for vp, off */
2642 page_t *pl[], /* object source page pointer */
2643 uint_t vpprot, /* access allowed to object pages */
2644 enum fault_type type, /* type of fault */
2645 enum seg_rw rw, /* type of access at fault */
2646 int brkcow) /* we may need to break cow */
2647 {
2648 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2649 page_t *pp, **ppp;
2650 uint_t pageflags = 0;
2651 page_t *anon_pl[1 + 1];
2652 page_t *opp = NULL; /* original page */
2653 uint_t prot;
2654 int err;
2655 int cow;
2656 int claim;
2657 int steal = 0;
2658 ulong_t anon_index;
2659 struct anon *ap, *oldap;
2660 struct anon_map *amp;
2661 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2662 int anon_lock = 0;
2663 anon_sync_obj_t cookie;
2664
2665 if (svd->flags & MAP_TEXT) {
2666 hat_flag |= HAT_LOAD_TEXT;
2667 }
2668
2669 ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2670 ASSERT(seg->s_szc == 0);
2671 ASSERT(svd->tr_state != SEGVN_TR_INIT);
2672
2673 /*
2674 * Initialize protection value for this page.
2675 * If we have per page protection values check it now.
2676 */
2677 if (svd->pageprot) {
2678 uint_t protchk;
2679
2680 switch (rw) {
2681 case S_READ:
2682 protchk = PROT_READ;
2683 break;
2684 case S_WRITE:
2685 protchk = PROT_WRITE;
2686 break;
2687 case S_EXEC:
2688 protchk = PROT_EXEC;
2689 break;
2690 case S_OTHER:
2691 default:
2692 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2693 break;
2694 }
2695
2696 prot = VPP_PROT(vpage);
2697 if ((prot & protchk) == 0)
2698 return (FC_PROT); /* illegal access type */
2699 } else {
2700 prot = svd->prot;
2701 }
2702
2703 if (type == F_SOFTLOCK) {
2704 atomic_add_long((ulong_t *)&svd->softlockcnt, 1);
2705 }
2706
2707 /*
2708 * Always acquire the anon array lock to prevent 2 threads from
2709 * allocating separate anon slots for the same "addr".
2710 */
2711
2712 if ((amp = svd->amp) != NULL) {
2713 ASSERT(RW_READ_HELD(&->a_rwlock));
2714 anon_index = svd->anon_index + seg_page(seg, addr);
2715 anon_array_enter(amp, anon_index, &cookie);
2716 anon_lock = 1;
2717 }
2718
2719 if (svd->vp == NULL && amp != NULL) {
2720 if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2721 /*
2722 * Allocate a (normally) writable anonymous page of
2723 * zeroes. If no advance reservations, reserve now.
2724 */
2725 if (svd->flags & MAP_NORESERVE) {
2726 if (anon_resv_zone(ptob(1),
2727 seg->s_as->a_proc->p_zone)) {
2728 atomic_add_long(&svd->swresv, ptob(1));
2729 atomic_add_long(&seg->s_as->a_resvsize,
2730 ptob(1));
2731 } else {
2732 err = ENOMEM;
2733 goto out;
2734 }
2735 }
2736 if ((pp = anon_zero(seg, addr, &ap,
2737 svd->cred)) == NULL) {
2738 err = ENOMEM;
2739 goto out; /* out of swap space */
2740 }
2741 /*
2742 * Re-acquire the anon_map lock and
2743 * initialize the anon array entry.
2744 */
2745 (void) anon_set_ptr(amp->ahp, anon_index, ap,
2746 ANON_SLEEP);
2747
2748 ASSERT(pp->p_szc == 0);
2749
2750 /*
2751 * Handle pages that have been marked for migration
2752 */
2753 if (lgrp_optimizations())
2754 page_migrate(seg, addr, &pp, 1);
2755
2756 if (enable_mbit_wa) {
2757 if (rw == S_WRITE)
2758 hat_setmod(pp);
2759 else if (!hat_ismod(pp))
2760 prot &= ~PROT_WRITE;
2761 }
2762 /*
2763 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2764 * with MC_LOCKAS, MCL_FUTURE) and this is a
2765 * MAP_NORESERVE segment, we may need to
2766 * permanently lock the page as it is being faulted
2767 * for the first time. The following text applies
2768 * only to MAP_NORESERVE segments:
2769 *
2770 * As per memcntl(2), if this segment was created
2771 * after MCL_FUTURE was applied (a "future"
2772 * segment), its pages must be locked. If this
2773 * segment existed at MCL_FUTURE application (a
2774 * "past" segment), the interface is unclear.
2775 *
2776 * We decide to lock only if vpage is present:
2777 *
2778 * - "future" segments will have a vpage array (see
2779 * as_map), and so will be locked as required
2780 *
2781 * - "past" segments may not have a vpage array,
2782 * depending on whether events (such as
2783 * mprotect) have occurred. Locking if vpage
2784 * exists will preserve legacy behavior. Not
2785 * locking if vpage is absent, will not break
2786 * the interface or legacy behavior. Note that
2787 * allocating vpage here if it's absent requires
2788 * upgrading the segvn reader lock, the cost of
2789 * which does not seem worthwhile.
2790 *
2791 * Usually testing and setting VPP_ISPPLOCK and
2792 * VPP_SETPPLOCK requires holding the segvn lock as
2793 * writer, but in this case all readers are
2794 * serializing on the anon array lock.
2795 */
2796 if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2797 (svd->flags & MAP_NORESERVE) &&
2798 !VPP_ISPPLOCK(vpage)) {
2799 proc_t *p = seg->s_as->a_proc;
2800 ASSERT(svd->type == MAP_PRIVATE);
2801 mutex_enter(&p->p_lock);
2802 if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2803 1) == 0) {
2804 claim = VPP_PROT(vpage) & PROT_WRITE;
2805 if (page_pp_lock(pp, claim, 0)) {
2806 VPP_SETPPLOCK(vpage);
2807 } else {
2808 rctl_decr_locked_mem(p, NULL,
2809 PAGESIZE, 1);
2810 }
2811 }
2812 mutex_exit(&p->p_lock);
2813 }
2814
2815 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2816 hat_memload(hat, addr, pp, prot, hat_flag);
2817
2818 if (!(hat_flag & HAT_LOAD_LOCK))
2819 page_unlock(pp);
2820
2821 anon_array_exit(&cookie);
2822 return (0);
2823 }
2824 }
2825
2826 /*
2827 * Obtain the page structure via anon_getpage() if it is
2828 * a private copy of an object (the result of a previous
2829 * copy-on-write).
2830 */
2831 if (amp != NULL) {
2832 if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2833 err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2834 seg, addr, rw, svd->cred);
2835 if (err)
2836 goto out;
2837
2838 if (svd->type == MAP_SHARED) {
2839 /*
2840 * If this is a shared mapping to an
2841 * anon_map, then ignore the write
2842 * permissions returned by anon_getpage().
2843 * They apply to the private mappings
2844 * of this anon_map.
2845 */
2846 vpprot |= PROT_WRITE;
2847 }
2848 opp = anon_pl[0];
2849 }
2850 }
2851
2852 /*
2853 * Search the pl[] list passed in if it is from the
2854 * original object (i.e., not a private copy).
2855 */
2856 if (opp == NULL) {
2857 /*
2858 * Find original page. We must be bringing it in
2859 * from the list in pl[].
2860 */
2861 for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2862 if (opp == PAGE_HANDLED)
2863 continue;
2864 ASSERT(opp->p_vnode == svd->vp); /* XXX */
2865 if (opp->p_offset == off)
2866 break;
2867 }
2868 if (opp == NULL) {
2869 panic("segvn_faultpage not found");
2870 /*NOTREACHED*/
2871 }
2872 *ppp = PAGE_HANDLED;
2873
2874 }
2875
2876 ASSERT(PAGE_LOCKED(opp));
2877
2878 TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2879 "segvn_fault:pp %p vp %p offset %llx", opp, NULL, 0);
2880
2881 /*
2882 * The fault is treated as a copy-on-write fault if a
2883 * write occurs on a private segment and the object
2884 * page (i.e., mapping) is write protected. We assume
2885 * that fatal protection checks have already been made.
2886 */
2887
2888 if (brkcow) {
2889 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2890 cow = !(vpprot & PROT_WRITE);
2891 } else if (svd->tr_state == SEGVN_TR_ON) {
2892 /*
2893 * If we are doing text replication COW on first touch.
2894 */
2895 ASSERT(amp != NULL);
2896 ASSERT(svd->vp != NULL);
2897 ASSERT(rw != S_WRITE);
2898 cow = (ap == NULL);
2899 } else {
2900 cow = 0;
2901 }
2902
2903 /*
2904 * If not a copy-on-write case load the translation
2905 * and return.
2906 */
2907 if (cow == 0) {
2908
2909 /*
2910 * Handle pages that have been marked for migration
2911 */
2912 if (lgrp_optimizations())
2913 page_migrate(seg, addr, &opp, 1);
2914
2915 if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2916 if (rw == S_WRITE)
2917 hat_setmod(opp);
2918 else if (rw != S_OTHER && !hat_ismod(opp))
2919 prot &= ~PROT_WRITE;
2920 }
2921
2922 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2923 (!svd->pageprot && svd->prot == (prot & vpprot)));
2924 ASSERT(amp == NULL ||
2925 svd->rcookie == HAT_INVALID_REGION_COOKIE);
2926 hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2927 svd->rcookie);
2928
2929 if (!(hat_flag & HAT_LOAD_LOCK))
2930 page_unlock(opp);
2931
2932 if (anon_lock) {
2933 anon_array_exit(&cookie);
2934 }
2935 return (0);
2936 }
2937
2938 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2939
2940 hat_setref(opp);
2941
2942 ASSERT(amp != NULL && anon_lock);
2943
2944 /*
2945 * Steal the page only if it isn't a private page
2946 * since stealing a private page is not worth the effort.
2947 */
2948 if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
2949 steal = 1;
2950
2951 /*
2952 * Steal the original page if the following conditions are true:
2953 *
2954 * We are low on memory, the page is not private, page is not large,
2955 * not shared, not modified, not `locked' or if we have it `locked'
2956 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2957 * that the page is not shared) and if it doesn't have any
2958 * translations. page_struct_lock isn't needed to look at p_cowcnt
2959 * and p_lckcnt because we first get exclusive lock on page.
2960 */
2961 (void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
2962
2963 if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
2964 page_tryupgrade(opp) && !hat_ismod(opp) &&
2965 ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
2966 (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
2967 vpage != NULL && VPP_ISPPLOCK(vpage)))) {
2968 /*
2969 * Check if this page has other translations
2970 * after unloading our translation.
2971 */
2972 if (hat_page_is_mapped(opp)) {
2973 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2974 hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
2975 HAT_UNLOAD);
2976 }
2977
2978 /*
2979 * hat_unload() might sync back someone else's recent
2980 * modification, so check again.
2981 */
2982 if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
2983 pageflags |= STEAL_PAGE;
2984 }
2985
2986 /*
2987 * If we have a vpage pointer, see if it indicates that we have
2988 * ``locked'' the page we map -- if so, tell anon_private to
2989 * transfer the locking resource to the new page.
2990 *
2991 * See Statement at the beginning of segvn_lockop regarding
2992 * the way lockcnts/cowcnts are handled during COW.
2993 *
2994 */
2995 if (vpage != NULL && VPP_ISPPLOCK(vpage))
2996 pageflags |= LOCK_PAGE;
2997
2998 /*
2999 * Allocate a private page and perform the copy.
3000 * For MAP_NORESERVE reserve swap space now, unless this
3001 * is a cow fault on an existing anon page in which case
3002 * MAP_NORESERVE will have made advance reservations.
3003 */
3004 if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
3005 if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
3006 atomic_add_long(&svd->swresv, ptob(1));
3007 atomic_add_long(&seg->s_as->a_resvsize, ptob(1));
3008 } else {
3009 page_unlock(opp);
3010 err = ENOMEM;
3011 goto out;
3012 }
3013 }
3014 oldap = ap;
3015 pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
3016 if (pp == NULL) {
3017 err = ENOMEM; /* out of swap space */
3018 goto out;
3019 }
3020
3021 /*
3022 * If we copied away from an anonymous page, then
3023 * we are one step closer to freeing up an anon slot.
3024 *
3025 * NOTE: The original anon slot must be released while
3026 * holding the "anon_map" lock. This is necessary to prevent
3027 * other threads from obtaining a pointer to the anon slot
3028 * which may be freed if its "refcnt" is 1.
3029 */
3030 if (oldap != NULL)
3031 anon_decref(oldap);
3032
3033 (void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3034
3035 /*
3036 * Handle pages that have been marked for migration
3037 */
3038 if (lgrp_optimizations())
3039 page_migrate(seg, addr, &pp, 1);
3040
3041 ASSERT(pp->p_szc == 0);
3042
3043 ASSERT(!IS_VMODSORT(pp->p_vnode));
3044 if (enable_mbit_wa) {
3045 if (rw == S_WRITE)
3046 hat_setmod(pp);
3047 else if (!hat_ismod(pp))
3048 prot &= ~PROT_WRITE;
3049 }
3050
3051 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3052 hat_memload(hat, addr, pp, prot, hat_flag);
3053
3054 if (!(hat_flag & HAT_LOAD_LOCK))
3055 page_unlock(pp);
3056
3057 ASSERT(anon_lock);
3058 anon_array_exit(&cookie);
3059 return (0);
3060 out:
3061 if (anon_lock)
3062 anon_array_exit(&cookie);
3063
3064 if (type == F_SOFTLOCK) {
3065 atomic_add_long((ulong_t *)&svd->softlockcnt, -1);
3066 }
3067 return (FC_MAKE_ERR(err));
3068 }
3069
3070 /*
3071 * relocate a bunch of smaller targ pages into one large repl page. all targ
3072 * pages must be complete pages smaller than replacement pages.
3073 * it's assumed that no page's szc can change since they are all PAGESIZE or
3074 * complete large pages locked SHARED.
3075 */
3076 static void
3077 segvn_relocate_pages(page_t **targ, page_t *replacement)
3078 {
3079 page_t *pp;
3080 pgcnt_t repl_npgs, curnpgs;
3081 pgcnt_t i;
3082 uint_t repl_szc = replacement->p_szc;
3083 page_t *first_repl = replacement;
3084 page_t *repl;
3085 spgcnt_t npgs;
3086
3087 VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3088
3089 ASSERT(repl_szc != 0);
3090 npgs = repl_npgs = page_get_pagecnt(repl_szc);
3091
3092 i = 0;
3093 while (repl_npgs) {
3094 spgcnt_t nreloc;
3095 int err;
3096 ASSERT(replacement != NULL);
3097 pp = targ[i];
3098 ASSERT(pp->p_szc < repl_szc);
3099 ASSERT(PAGE_EXCL(pp));
3100 ASSERT(!PP_ISFREE(pp));
3101 curnpgs = page_get_pagecnt(pp->p_szc);
3102 if (curnpgs == 1) {
3103 VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3104 repl = replacement;
3105 page_sub(&replacement, repl);
3106 ASSERT(PAGE_EXCL(repl));
3107 ASSERT(!PP_ISFREE(repl));
3108 ASSERT(repl->p_szc == repl_szc);
3109 } else {
3110 page_t *repl_savepp;
3111 int j;
3112 VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3113 repl_savepp = replacement;
3114 for (j = 0; j < curnpgs; j++) {
3115 repl = replacement;
3116 page_sub(&replacement, repl);
3117 ASSERT(PAGE_EXCL(repl));
3118 ASSERT(!PP_ISFREE(repl));
3119 ASSERT(repl->p_szc == repl_szc);
3120 ASSERT(page_pptonum(targ[i + j]) ==
3121 page_pptonum(targ[i]) + j);
3122 }
3123 repl = repl_savepp;
3124 ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3125 }
3126 err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3127 if (err || nreloc != curnpgs) {
3128 panic("segvn_relocate_pages: "
3129 "page_relocate failed err=%d curnpgs=%ld "
3130 "nreloc=%ld", err, curnpgs, nreloc);
3131 }
3132 ASSERT(curnpgs <= repl_npgs);
3133 repl_npgs -= curnpgs;
3134 i += curnpgs;
3135 }
3136 ASSERT(replacement == NULL);
3137
3138 repl = first_repl;
3139 repl_npgs = npgs;
3140 for (i = 0; i < repl_npgs; i++) {
3141 ASSERT(PAGE_EXCL(repl));
3142 ASSERT(!PP_ISFREE(repl));
3143 targ[i] = repl;
3144 page_downgrade(targ[i]);
3145 repl++;
3146 }
3147 }
3148
3149 /*
3150 * Check if all pages in ppa array are complete smaller than szc pages and
3151 * their roots will still be aligned relative to their current size if the
3152 * entire ppa array is relocated into one szc page. If these conditions are
3153 * not met return 0.
3154 *
3155 * If all pages are properly aligned attempt to upgrade their locks
3156 * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3157 * upgrdfail was set to 0 by caller.
3158 *
3159 * Return 1 if all pages are aligned and locked exclusively.
3160 *
3161 * If all pages in ppa array happen to be physically contiguous to make one
3162 * szc page and all exclusive locks are successfully obtained promote the page
3163 * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3164 */
3165 static int
3166 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3167 {
3168 page_t *pp;
3169 pfn_t pfn;
3170 pgcnt_t totnpgs = page_get_pagecnt(szc);
3171 pfn_t first_pfn;
3172 int contig = 1;
3173 pgcnt_t i;
3174 pgcnt_t j;
3175 uint_t curszc;
3176 pgcnt_t curnpgs;
3177 int root = 0;
3178
3179 ASSERT(szc > 0);
3180
3181 VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3182
3183 for (i = 0; i < totnpgs; i++) {
3184 pp = ppa[i];
3185 ASSERT(PAGE_SHARED(pp));
3186 ASSERT(!PP_ISFREE(pp));
3187 pfn = page_pptonum(pp);
3188 if (i == 0) {
3189 if (!IS_P2ALIGNED(pfn, totnpgs)) {
3190 contig = 0;
3191 } else {
3192 first_pfn = pfn;
3193 }
3194 } else if (contig && pfn != first_pfn + i) {
3195 contig = 0;
3196 }
3197 if (pp->p_szc == 0) {
3198 if (root) {
3199 VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3200 return (0);
3201 }
3202 } else if (!root) {
3203 if ((curszc = pp->p_szc) >= szc) {
3204 VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3205 return (0);
3206 }
3207 if (curszc == 0) {
3208 /*
3209 * p_szc changed means we don't have all pages
3210 * locked. return failure.
3211 */
3212 VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3213 return (0);
3214 }
3215 curnpgs = page_get_pagecnt(curszc);
3216 if (!IS_P2ALIGNED(pfn, curnpgs) ||
3217 !IS_P2ALIGNED(i, curnpgs)) {
3218 VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3219 return (0);
3220 }
3221 root = 1;
3222 } else {
3223 ASSERT(i > 0);
3224 VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3225 if (pp->p_szc != curszc) {
3226 VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3227 return (0);
3228 }
3229 if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3230 panic("segvn_full_szcpages: "
3231 "large page not physically contiguous");
3232 }
3233 if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3234 root = 0;
3235 }
3236 }
3237 }
3238
3239 for (i = 0; i < totnpgs; i++) {
3240 ASSERT(ppa[i]->p_szc < szc);
3241 if (!page_tryupgrade(ppa[i])) {
3242 for (j = 0; j < i; j++) {
3243 page_downgrade(ppa[j]);
3244 }
3245 *pszc = ppa[i]->p_szc;
3246 *upgrdfail = 1;
3247 VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3248 return (0);
3249 }
3250 }
3251
3252 /*
3253 * When a page is put a free cachelist its szc is set to 0. if file
3254 * system reclaimed pages from cachelist targ pages will be physically
3255 * contiguous with 0 p_szc. in this case just upgrade szc of targ
3256 * pages without any relocations.
3257 * To avoid any hat issues with previous small mappings
3258 * hat_pageunload() the target pages first.
3259 */
3260 if (contig) {
3261 VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3262 for (i = 0; i < totnpgs; i++) {
3263 (void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3264 }
3265 for (i = 0; i < totnpgs; i++) {
3266 ppa[i]->p_szc = szc;
3267 }
3268 for (i = 0; i < totnpgs; i++) {
3269 ASSERT(PAGE_EXCL(ppa[i]));
3270 page_downgrade(ppa[i]);
3271 }
3272 if (pszc != NULL) {
3273 *pszc = szc;
3274 }
3275 }
3276 VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3277 return (1);
3278 }
3279
3280 /*
3281 * Create physically contiguous pages for [vp, off] - [vp, off +
3282 * page_size(szc)) range and for private segment return them in ppa array.
3283 * Pages are created either via IO or relocations.
3284 *
3285 * Return 1 on success and 0 on failure.
3286 *
3287 * If physically contiguous pages already exist for this range return 1 without
3288 * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3289 * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3290 */
3291
3292 static int
3293 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, u_offset_t off,
3294 uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3295 int *downsize)
3296
3297 {
3298 page_t *pplist = *ppplist;
3299 size_t pgsz = page_get_pagesize(szc);
3300 pgcnt_t pages = btop(pgsz);
3301 ulong_t start_off = off;
3302 u_offset_t eoff = off + pgsz;
3303 spgcnt_t nreloc;
3304 u_offset_t io_off = off;
3305 size_t io_len;
3306 page_t *io_pplist = NULL;
3307 page_t *done_pplist = NULL;
3308 pgcnt_t pgidx = 0;
3309 page_t *pp;
3310 page_t *newpp;
3311 page_t *targpp;
3312 int io_err = 0;
3313 int i;
3314 pfn_t pfn;
3315 ulong_t ppages;
3316 page_t *targ_pplist = NULL;
3317 page_t *repl_pplist = NULL;
3318 page_t *tmp_pplist;
3319 int nios = 0;
3320 uint_t pszc;
3321 struct vattr va;
3322
3323 VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3324
3325 ASSERT(szc != 0);
3326 ASSERT(pplist->p_szc == szc);
3327
3328 /*
3329 * downsize will be set to 1 only if we fail to lock pages. this will
3330 * allow subsequent faults to try to relocate the page again. If we
3331 * fail due to misalignment don't downsize and let the caller map the
3332 * whole region with small mappings to avoid more faults into the area
3333 * where we can't get large pages anyway.
3334 */
3335 *downsize = 0;
3336
3337 while (off < eoff) {
3338 newpp = pplist;
3339 ASSERT(newpp != NULL);
3340 ASSERT(PAGE_EXCL(newpp));
3341 ASSERT(!PP_ISFREE(newpp));
3342 /*
3343 * we pass NULL for nrelocp to page_lookup_create()
3344 * so that it doesn't relocate. We relocate here
3345 * later only after we make sure we can lock all
3346 * pages in the range we handle and they are all
3347 * aligned.
3348 */
3349 pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3350 ASSERT(pp != NULL);
3351 ASSERT(!PP_ISFREE(pp));
3352 ASSERT(pp->p_vnode == vp);
3353 ASSERT(pp->p_offset == off);
3354 if (pp == newpp) {
3355 VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3356 page_sub(&pplist, pp);
3357 ASSERT(PAGE_EXCL(pp));
3358 ASSERT(page_iolock_assert(pp));
3359 page_list_concat(&io_pplist, &pp);
3360 off += PAGESIZE;
3361 continue;
3362 }
3363 VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3364 pfn = page_pptonum(pp);
3365 pszc = pp->p_szc;
3366 if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3367 IS_P2ALIGNED(pfn, pages)) {
3368 ASSERT(repl_pplist == NULL);
3369 ASSERT(done_pplist == NULL);
3370 ASSERT(pplist == *ppplist);
3371 page_unlock(pp);
3372 page_free_replacement_page(pplist);
3373 page_create_putback(pages);
3374 *ppplist = NULL;
3375 VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3376 return (1);
3377 }
3378 if (pszc >= szc) {
3379 page_unlock(pp);
3380 segvn_faultvnmpss_align_err1++;
3381 goto out;
3382 }
3383 ppages = page_get_pagecnt(pszc);
3384 if (!IS_P2ALIGNED(pfn, ppages)) {
3385 ASSERT(pszc > 0);
3386 /*
3387 * sizing down to pszc won't help.
3388 */
3389 page_unlock(pp);
3390 segvn_faultvnmpss_align_err2++;
3391 goto out;
3392 }
3393 pfn = page_pptonum(newpp);
3394 if (!IS_P2ALIGNED(pfn, ppages)) {
3395 ASSERT(pszc > 0);
3396 /*
3397 * sizing down to pszc won't help.
3398 */
3399 page_unlock(pp);
3400 segvn_faultvnmpss_align_err3++;
3401 goto out;
3402 }
3403 if (!PAGE_EXCL(pp)) {
3404 VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3405 page_unlock(pp);
3406 *downsize = 1;
3407 *ret_pszc = pp->p_szc;
3408 goto out;
3409 }
3410 targpp = pp;
3411 if (io_pplist != NULL) {
3412 VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3413 io_len = off - io_off;
3414 /*
3415 * Some file systems like NFS don't check EOF
3416 * conditions in VOP_PAGEIO(). Check it here
3417 * now that pages are locked SE_EXCL. Any file
3418 * truncation will wait until the pages are
3419 * unlocked so no need to worry that file will
3420 * be truncated after we check its size here.
3421 * XXX fix NFS to remove this check.
3422 */
3423 va.va_mask = AT_SIZE;
3424 if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3425 VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3426 page_unlock(targpp);
3427 goto out;
3428 }
3429 if (btopr(va.va_size) < btopr(io_off + io_len)) {
3430 VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3431 *downsize = 1;
3432 *ret_pszc = 0;
3433 page_unlock(targpp);
3434 goto out;
3435 }
3436 io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3437 B_READ, svd->cred, NULL);
3438 if (io_err) {
3439 VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3440 page_unlock(targpp);
3441 if (io_err == EDEADLK) {
3442 segvn_vmpss_pageio_deadlk_err++;
3443 }
3444 goto out;
3445 }
3446 nios++;
3447 VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3448 while (io_pplist != NULL) {
3449 pp = io_pplist;
3450 page_sub(&io_pplist, pp);
3451 ASSERT(page_iolock_assert(pp));
3452 page_io_unlock(pp);
3453 pgidx = (pp->p_offset - start_off) >>
3454 PAGESHIFT;
3455 ASSERT(pgidx < pages);
3456 ppa[pgidx] = pp;
3457 page_list_concat(&done_pplist, &pp);
3458 }
3459 }
3460 pp = targpp;
3461 ASSERT(PAGE_EXCL(pp));
3462 ASSERT(pp->p_szc <= pszc);
3463 if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3464 VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3465 page_unlock(pp);
3466 *downsize = 1;
3467 *ret_pszc = pp->p_szc;
3468 goto out;
3469 }
3470 VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3471 /*
3472 * page szc chould have changed before the entire group was
3473 * locked. reread page szc.
3474 */
3475 pszc = pp->p_szc;
3476 ppages = page_get_pagecnt(pszc);
3477
3478 /* link just the roots */
3479 page_list_concat(&targ_pplist, &pp);
3480 page_sub(&pplist, newpp);
3481 page_list_concat(&repl_pplist, &newpp);
3482 off += PAGESIZE;
3483 while (--ppages != 0) {
3484 newpp = pplist;
3485 page_sub(&pplist, newpp);
3486 off += PAGESIZE;
3487 }
3488 io_off = off;
3489 }
3490 if (io_pplist != NULL) {
3491 VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3492 io_len = eoff - io_off;
3493 va.va_mask = AT_SIZE;
3494 if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3495 VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3496 goto out;
3497 }
3498 if (btopr(va.va_size) < btopr(io_off + io_len)) {
3499 VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3500 *downsize = 1;
3501 *ret_pszc = 0;
3502 goto out;
3503 }
3504 io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3505 B_READ, svd->cred, NULL);
3506 if (io_err) {
3507 VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3508 if (io_err == EDEADLK) {
3509 segvn_vmpss_pageio_deadlk_err++;
3510 }
3511 goto out;
3512 }
3513 nios++;
3514 while (io_pplist != NULL) {
3515 pp = io_pplist;
3516 page_sub(&io_pplist, pp);
3517 ASSERT(page_iolock_assert(pp));
3518 page_io_unlock(pp);
3519 pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3520 ASSERT(pgidx < pages);
3521 ppa[pgidx] = pp;
3522 }
3523 }
3524 /*
3525 * we're now bound to succeed or panic.
3526 * remove pages from done_pplist. it's not needed anymore.
3527 */
3528 while (done_pplist != NULL) {
3529 pp = done_pplist;
3530 page_sub(&done_pplist, pp);
3531 }
3532 VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3533 ASSERT(pplist == NULL);
3534 *ppplist = NULL;
3535 while (targ_pplist != NULL) {
3536 int ret;
3537 VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3538 ASSERT(repl_pplist);
3539 pp = targ_pplist;
3540 page_sub(&targ_pplist, pp);
3541 pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3542 newpp = repl_pplist;
3543 page_sub(&repl_pplist, newpp);
3544 #ifdef DEBUG
3545 pfn = page_pptonum(pp);
3546 pszc = pp->p_szc;
3547 ppages = page_get_pagecnt(pszc);
3548 ASSERT(IS_P2ALIGNED(pfn, ppages));
3549 pfn = page_pptonum(newpp);
3550 ASSERT(IS_P2ALIGNED(pfn, ppages));
3551 ASSERT(P2PHASE(pfn, pages) == pgidx);
3552 #endif
3553 nreloc = 0;
3554 ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3555 if (ret != 0 || nreloc == 0) {
3556 panic("segvn_fill_vp_pages: "
3557 "page_relocate failed");
3558 }
3559 pp = newpp;
3560 while (nreloc-- != 0) {
3561 ASSERT(PAGE_EXCL(pp));
3562 ASSERT(pp->p_vnode == vp);
3563 ASSERT(pgidx ==
3564 ((pp->p_offset - start_off) >> PAGESHIFT));
3565 ppa[pgidx++] = pp;
3566 pp++;
3567 }
3568 }
3569
3570 if (svd->type == MAP_PRIVATE) {
3571 VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3572 for (i = 0; i < pages; i++) {
3573 ASSERT(ppa[i] != NULL);
3574 ASSERT(PAGE_EXCL(ppa[i]));
3575 ASSERT(ppa[i]->p_vnode == vp);
3576 ASSERT(ppa[i]->p_offset ==
3577 start_off + (i << PAGESHIFT));
3578 page_downgrade(ppa[i]);
3579 }
3580 ppa[pages] = NULL;
3581 } else {
3582 VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3583 /*
3584 * the caller will still call VOP_GETPAGE() for shared segments
3585 * to check FS write permissions. For private segments we map
3586 * file read only anyway. so no VOP_GETPAGE is needed.
3587 */
3588 for (i = 0; i < pages; i++) {
3589 ASSERT(ppa[i] != NULL);
3590 ASSERT(PAGE_EXCL(ppa[i]));
3591 ASSERT(ppa[i]->p_vnode == vp);
3592 ASSERT(ppa[i]->p_offset ==
3593 start_off + (i << PAGESHIFT));
3594 page_unlock(ppa[i]);
3595 }
3596 ppa[0] = NULL;
3597 }
3598
3599 return (1);
3600 out:
3601 /*
3602 * Do the cleanup. Unlock target pages we didn't relocate. They are
3603 * linked on targ_pplist by root pages. reassemble unused replacement
3604 * and io pages back to pplist.
3605 */
3606 if (io_pplist != NULL) {
3607 VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3608 pp = io_pplist;
3609 do {
3610 ASSERT(pp->p_vnode == vp);
3611 ASSERT(pp->p_offset == io_off);
3612 ASSERT(page_iolock_assert(pp));
3613 page_io_unlock(pp);
3614 page_hashout(pp, NULL);
3615 io_off += PAGESIZE;
3616 } while ((pp = pp->p_next) != io_pplist);
3617 page_list_concat(&io_pplist, &pplist);
3618 pplist = io_pplist;
3619 }
3620 tmp_pplist = NULL;
3621 while (targ_pplist != NULL) {
3622 VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3623 pp = targ_pplist;
3624 ASSERT(PAGE_EXCL(pp));
3625 page_sub(&targ_pplist, pp);
3626
3627 pszc = pp->p_szc;
3628 ppages = page_get_pagecnt(pszc);
3629 ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3630
3631 if (pszc != 0) {
3632 group_page_unlock(pp);
3633 }
3634 page_unlock(pp);
3635
3636 pp = repl_pplist;
3637 ASSERT(pp != NULL);
3638 ASSERT(PAGE_EXCL(pp));
3639 ASSERT(pp->p_szc == szc);
3640 page_sub(&repl_pplist, pp);
3641
3642 ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3643
3644 /* relink replacement page */
3645 page_list_concat(&tmp_pplist, &pp);
3646 while (--ppages != 0) {
3647 VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3648 pp++;
3649 ASSERT(PAGE_EXCL(pp));
3650 ASSERT(pp->p_szc == szc);
3651 page_list_concat(&tmp_pplist, &pp);
3652 }
3653 }
3654 if (tmp_pplist != NULL) {
3655 VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3656 page_list_concat(&tmp_pplist, &pplist);
3657 pplist = tmp_pplist;
3658 }
3659 /*
3660 * at this point all pages are either on done_pplist or
3661 * pplist. They can't be all on done_pplist otherwise
3662 * we'd've been done.
3663 */
3664 ASSERT(pplist != NULL);
3665 if (nios != 0) {
3666 VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3667 pp = pplist;
3668 do {
3669 VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3670 ASSERT(pp->p_szc == szc);
3671 ASSERT(PAGE_EXCL(pp));
3672 ASSERT(pp->p_vnode != vp);
3673 pp->p_szc = 0;
3674 } while ((pp = pp->p_next) != pplist);
3675
3676 pp = done_pplist;
3677 do {
3678 VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3679 ASSERT(pp->p_szc == szc);
3680 ASSERT(PAGE_EXCL(pp));
3681 ASSERT(pp->p_vnode == vp);
3682 pp->p_szc = 0;
3683 } while ((pp = pp->p_next) != done_pplist);
3684
3685 while (pplist != NULL) {
3686 VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3687 pp = pplist;
3688 page_sub(&pplist, pp);
3689 page_free(pp, 0);
3690 }
3691
3692 while (done_pplist != NULL) {
3693 VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3694 pp = done_pplist;
3695 page_sub(&done_pplist, pp);
3696 page_unlock(pp);
3697 }
3698 *ppplist = NULL;
3699 return (0);
3700 }
3701 ASSERT(pplist == *ppplist);
3702 if (io_err) {
3703 VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3704 /*
3705 * don't downsize on io error.
3706 * see if vop_getpage succeeds.
3707 * pplist may still be used in this case
3708 * for relocations.
3709 */
3710 return (0);
3711 }
3712 VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3713 page_free_replacement_page(pplist);
3714 page_create_putback(pages);
3715 *ppplist = NULL;
3716 return (0);
3717 }
3718
3719 int segvn_anypgsz = 0;
3720
3721 #define SEGVN_RESTORE_SOFTLOCK_VP(type, pages) \
3722 if ((type) == F_SOFTLOCK) { \
3723 atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3724 -(pages)); \
3725 }
3726
3727 #define SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot) \
3728 if (IS_VMODSORT((ppa)[0]->p_vnode)) { \
3729 if ((rw) == S_WRITE) { \
3730 for (i = 0; i < (pages); i++) { \
3731 ASSERT((ppa)[i]->p_vnode == \
3732 (ppa)[0]->p_vnode); \
3733 hat_setmod((ppa)[i]); \
3734 } \
3735 } else if ((rw) != S_OTHER && \
3736 ((prot) & (vpprot) & PROT_WRITE)) { \
3737 for (i = 0; i < (pages); i++) { \
3738 ASSERT((ppa)[i]->p_vnode == \
3739 (ppa)[0]->p_vnode); \
3740 if (!hat_ismod((ppa)[i])) { \
3741 prot &= ~PROT_WRITE; \
3742 break; \
3743 } \
3744 } \
3745 } \
3746 }
3747
3748 #ifdef VM_STATS
3749
3750 #define SEGVN_VMSTAT_FLTVNPAGES(idx) \
3751 VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3752
3753 #else /* VM_STATS */
3754
3755 #define SEGVN_VMSTAT_FLTVNPAGES(idx)
3756
3757 #endif
3758
3759 static faultcode_t
3760 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3761 caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3762 caddr_t eaddr, int brkcow)
3763 {
3764 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3765 struct anon_map *amp = svd->amp;
3766 uchar_t segtype = svd->type;
3767 uint_t szc = seg->s_szc;
3768 size_t pgsz = page_get_pagesize(szc);
3769 size_t maxpgsz = pgsz;
3770 pgcnt_t pages = btop(pgsz);
3771 pgcnt_t maxpages = pages;
3772 size_t ppasize = (pages + 1) * sizeof (page_t *);
3773 caddr_t a = lpgaddr;
3774 caddr_t maxlpgeaddr = lpgeaddr;
3775 u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3776 ulong_t aindx = svd->anon_index + seg_page(seg, a);
3777 struct vpage *vpage = (svd->vpage != NULL) ?
3778 &svd->vpage[seg_page(seg, a)] : NULL;
3779 vnode_t *vp = svd->vp;
3780 page_t **ppa;
3781 uint_t pszc;
3782 size_t ppgsz;
3783 pgcnt_t ppages;
3784 faultcode_t err = 0;
3785 int ierr;
3786 int vop_size_err = 0;
3787 uint_t protchk, prot, vpprot;
3788 ulong_t i;
3789 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3790 anon_sync_obj_t an_cookie;
3791 enum seg_rw arw;
3792 int alloc_failed = 0;
3793 int adjszc_chk;
3794 struct vattr va;
3795 int xhat = 0;
3796 page_t *pplist;
3797 pfn_t pfn;
3798 int physcontig;
3799 int upgrdfail;
3800 int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3801 int tron = (svd->tr_state == SEGVN_TR_ON);
3802
3803 ASSERT(szc != 0);
3804 ASSERT(vp != NULL);
3805 ASSERT(brkcow == 0 || amp != NULL);
3806 ASSERT(tron == 0 || amp != NULL);
3807 ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3808 ASSERT(!(svd->flags & MAP_NORESERVE));
3809 ASSERT(type != F_SOFTUNLOCK);
3810 ASSERT(IS_P2ALIGNED(a, maxpgsz));
3811 ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3812 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3813 ASSERT(seg->s_szc < NBBY * sizeof (int));
3814 ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3815 ASSERT(svd->tr_state != SEGVN_TR_INIT);
3816
3817 VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3818 VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3819
3820 if (svd->flags & MAP_TEXT) {
3821 hat_flag |= HAT_LOAD_TEXT;
3822 }
3823
3824 if (svd->pageprot) {
3825 switch (rw) {
3826 case S_READ:
3827 protchk = PROT_READ;
3828 break;
3829 case S_WRITE:
3830 protchk = PROT_WRITE;
3831 break;
3832 case S_EXEC:
3833 protchk = PROT_EXEC;
3834 break;
3835 case S_OTHER:
3836 default:
3837 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3838 break;
3839 }
3840 } else {
3841 prot = svd->prot;
3842 /* caller has already done segment level protection check. */
3843 }
3844
3845 if (seg->s_as->a_hat != hat) {
3846 xhat = 1;
3847 }
3848
3849 if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3850 SEGVN_VMSTAT_FLTVNPAGES(2);
3851 arw = S_READ;
3852 } else {
3853 arw = rw;
3854 }
3855
3856 ppa = kmem_alloc(ppasize, KM_SLEEP);
3857
3858 VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3859
3860 for (;;) {
3861 adjszc_chk = 0;
3862 for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3863 if (adjszc_chk) {
3864 while (szc < seg->s_szc) {
3865 uintptr_t e;
3866 uint_t tszc;
3867 tszc = segvn_anypgsz_vnode ? szc + 1 :
3868 seg->s_szc;
3869 ppgsz = page_get_pagesize(tszc);
3870 if (!IS_P2ALIGNED(a, ppgsz) ||
3871 ((alloc_failed >> tszc) & 0x1)) {
3872 break;
3873 }
3874 SEGVN_VMSTAT_FLTVNPAGES(4);
3875 szc = tszc;
3876 pgsz = ppgsz;
3877 pages = btop(pgsz);
3878 e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3879 lpgeaddr = (caddr_t)e;
3880 }
3881 }
3882
3883 again:
3884 if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3885 ASSERT(IS_P2ALIGNED(aindx, maxpages));
3886 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
3887 anon_array_enter(amp, aindx, &an_cookie);
3888 if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3889 SEGVN_VMSTAT_FLTVNPAGES(5);
3890 ASSERT(anon_pages(amp->ahp, aindx,
3891 maxpages) == maxpages);
3892 anon_array_exit(&an_cookie);
3893 ANON_LOCK_EXIT(&->a_rwlock);
3894 err = segvn_fault_anonpages(hat, seg,
3895 a, a + maxpgsz, type, rw,
3896 MAX(a, addr),
3897 MIN(a + maxpgsz, eaddr), brkcow);
3898 if (err != 0) {
3899 SEGVN_VMSTAT_FLTVNPAGES(6);
3900 goto out;
3901 }
3902 if (szc < seg->s_szc) {
3903 szc = seg->s_szc;
3904 pgsz = maxpgsz;
3905 pages = maxpages;
3906 lpgeaddr = maxlpgeaddr;
3907 }
3908 goto next;
3909 } else {
3910 ASSERT(anon_pages(amp->ahp, aindx,
3911 maxpages) == 0);
3912 SEGVN_VMSTAT_FLTVNPAGES(7);
3913 anon_array_exit(&an_cookie);
3914 ANON_LOCK_EXIT(&->a_rwlock);
3915 }
3916 }
3917 ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3918 ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3919
3920 if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3921 ASSERT(vpage != NULL);
3922 prot = VPP_PROT(vpage);
3923 ASSERT(sameprot(seg, a, maxpgsz));
3924 if ((prot & protchk) == 0) {
3925 SEGVN_VMSTAT_FLTVNPAGES(8);
3926 err = FC_PROT;
3927 goto out;
3928 }
3929 }
3930 if (type == F_SOFTLOCK) {
3931 atomic_add_long((ulong_t *)&svd->softlockcnt,
3932 pages);
3933 }
3934
3935 pplist = NULL;
3936 physcontig = 0;
3937 ppa[0] = NULL;
3938 if (!brkcow && !tron && szc &&
3939 !page_exists_physcontig(vp, off, szc,
3940 segtype == MAP_PRIVATE ? ppa : NULL)) {
3941 SEGVN_VMSTAT_FLTVNPAGES(9);
3942 if (page_alloc_pages(vp, seg, a, &pplist, NULL,
3943 szc, 0, 0) && type != F_SOFTLOCK) {
3944 SEGVN_VMSTAT_FLTVNPAGES(10);
3945 pszc = 0;
3946 ierr = -1;
3947 alloc_failed |= (1 << szc);
3948 break;
3949 }
3950 if (pplist != NULL &&
3951 vp->v_mpssdata == SEGVN_PAGEIO) {
3952 int downsize;
3953 SEGVN_VMSTAT_FLTVNPAGES(11);
3954 physcontig = segvn_fill_vp_pages(svd,
3955 vp, off, szc, ppa, &pplist,
3956 &pszc, &downsize);
3957 ASSERT(!physcontig || pplist == NULL);
3958 if (!physcontig && downsize &&
3959 type != F_SOFTLOCK) {
3960 ASSERT(pplist == NULL);
3961 SEGVN_VMSTAT_FLTVNPAGES(12);
3962 ierr = -1;
3963 break;
3964 }
3965 ASSERT(!physcontig ||
3966 segtype == MAP_PRIVATE ||
3967 ppa[0] == NULL);
3968 if (physcontig && ppa[0] == NULL) {
3969 physcontig = 0;
3970 }
3971 }
3972 } else if (!brkcow && !tron && szc && ppa[0] != NULL) {
3973 SEGVN_VMSTAT_FLTVNPAGES(13);
3974 ASSERT(segtype == MAP_PRIVATE);
3975 physcontig = 1;
3976 }
3977
3978 if (!physcontig) {
3979 SEGVN_VMSTAT_FLTVNPAGES(14);
3980 ppa[0] = NULL;
3981 ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
3982 &vpprot, ppa, pgsz, seg, a, arw,
3983 svd->cred, NULL);
3984 #ifdef DEBUG
3985 if (ierr == 0) {
3986 for (i = 0; i < pages; i++) {
3987 ASSERT(PAGE_LOCKED(ppa[i]));
3988 ASSERT(!PP_ISFREE(ppa[i]));
3989 ASSERT(ppa[i]->p_vnode == vp);
3990 ASSERT(ppa[i]->p_offset ==
3991 off + (i << PAGESHIFT));
3992 }
3993 }
3994 #endif /* DEBUG */
3995 if (segtype == MAP_PRIVATE) {
3996 SEGVN_VMSTAT_FLTVNPAGES(15);
3997 vpprot &= ~PROT_WRITE;
3998 }
3999 } else {
4000 ASSERT(segtype == MAP_PRIVATE);
4001 SEGVN_VMSTAT_FLTVNPAGES(16);
4002 vpprot = PROT_ALL & ~PROT_WRITE;
4003 ierr = 0;
4004 }
4005
4006 if (ierr != 0) {
4007 SEGVN_VMSTAT_FLTVNPAGES(17);
4008 if (pplist != NULL) {
4009 SEGVN_VMSTAT_FLTVNPAGES(18);
4010 page_free_replacement_page(pplist);
4011 page_create_putback(pages);
4012 }
4013 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4014 if (a + pgsz <= eaddr) {
4015 SEGVN_VMSTAT_FLTVNPAGES(19);
4016 err = FC_MAKE_ERR(ierr);
4017 goto out;
4018 }
4019 va.va_mask = AT_SIZE;
4020 if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL)) {
4021 SEGVN_VMSTAT_FLTVNPAGES(20);
4022 err = FC_MAKE_ERR(EIO);
4023 goto out;
4024 }
4025 if (btopr(va.va_size) >= btopr(off + pgsz)) {
4026 SEGVN_VMSTAT_FLTVNPAGES(21);
4027 err = FC_MAKE_ERR(ierr);
4028 goto out;
4029 }
4030 if (btopr(va.va_size) <
4031 btopr(off + (eaddr - a))) {
4032 SEGVN_VMSTAT_FLTVNPAGES(22);
4033 err = FC_MAKE_ERR(ierr);
4034 goto out;
4035 }
4036 if (brkcow || tron || type == F_SOFTLOCK) {
4037 /* can't reduce map area */
4038 SEGVN_VMSTAT_FLTVNPAGES(23);
4039 vop_size_err = 1;
4040 goto out;
4041 }
4042 SEGVN_VMSTAT_FLTVNPAGES(24);
4043 ASSERT(szc != 0);
4044 pszc = 0;
4045 ierr = -1;
4046 break;
4047 }
4048
4049 if (amp != NULL) {
4050 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
4051 anon_array_enter(amp, aindx, &an_cookie);
4052 }
4053 if (amp != NULL &&
4054 anon_get_ptr(amp->ahp, aindx) != NULL) {
4055 ulong_t taindx = P2ALIGN(aindx, maxpages);
4056
4057 SEGVN_VMSTAT_FLTVNPAGES(25);
4058 ASSERT(anon_pages(amp->ahp, taindx,
4059 maxpages) == maxpages);
4060 for (i = 0; i < pages; i++) {
4061 page_unlock(ppa[i]);
4062 }
4063 anon_array_exit(&an_cookie);
4064 ANON_LOCK_EXIT(&->a_rwlock);
4065 if (pplist != NULL) {
4066 page_free_replacement_page(pplist);
4067 page_create_putback(pages);
4068 }
4069 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4070 if (szc < seg->s_szc) {
4071 SEGVN_VMSTAT_FLTVNPAGES(26);
4072 /*
4073 * For private segments SOFTLOCK
4074 * either always breaks cow (any rw
4075 * type except S_READ_NOCOW) or
4076 * address space is locked as writer
4077 * (S_READ_NOCOW case) and anon slots
4078 * can't show up on second check.
4079 * Therefore if we are here for
4080 * SOFTLOCK case it must be a cow
4081 * break but cow break never reduces
4082 * szc. text replication (tron) in
4083 * this case works as cow break.
4084 * Thus the assert below.
4085 */
4086 ASSERT(!brkcow && !tron &&
4087 type != F_SOFTLOCK);
4088 pszc = seg->s_szc;
4089 ierr = -2;
4090 break;
4091 }
4092 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4093 goto again;
4094 }
4095 #ifdef DEBUG
4096 if (amp != NULL) {
4097 ulong_t taindx = P2ALIGN(aindx, maxpages);
4098 ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4099 }
4100 #endif /* DEBUG */
4101
4102 if (brkcow || tron) {
4103 ASSERT(amp != NULL);
4104 ASSERT(pplist == NULL);
4105 ASSERT(szc == seg->s_szc);
4106 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4107 ASSERT(IS_P2ALIGNED(aindx, maxpages));
4108 SEGVN_VMSTAT_FLTVNPAGES(27);
4109 ierr = anon_map_privatepages(amp, aindx, szc,
4110 seg, a, prot, ppa, vpage, segvn_anypgsz,
4111 tron ? PG_LOCAL : 0, svd->cred);
4112 if (ierr != 0) {
4113 SEGVN_VMSTAT_FLTVNPAGES(28);
4114 anon_array_exit(&an_cookie);
4115 ANON_LOCK_EXIT(&->a_rwlock);
4116 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4117 err = FC_MAKE_ERR(ierr);
4118 goto out;
4119 }
4120
4121 ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4122 /*
4123 * p_szc can't be changed for locked
4124 * swapfs pages.
4125 */
4126 ASSERT(svd->rcookie ==
4127 HAT_INVALID_REGION_COOKIE);
4128 hat_memload_array(hat, a, pgsz, ppa, prot,
4129 hat_flag);
4130
4131 if (!(hat_flag & HAT_LOAD_LOCK)) {
4132 SEGVN_VMSTAT_FLTVNPAGES(29);
4133 for (i = 0; i < pages; i++) {
4134 page_unlock(ppa[i]);
4135 }
4136 }
4137 anon_array_exit(&an_cookie);
4138 ANON_LOCK_EXIT(&->a_rwlock);
4139 goto next;
4140 }
4141
4142 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4143 (!svd->pageprot && svd->prot == (prot & vpprot)));
4144
4145 pfn = page_pptonum(ppa[0]);
4146 /*
4147 * hat_page_demote() needs an SE_EXCL lock on one of
4148 * constituent page_t's and it decreases root's p_szc
4149 * last. This means if root's p_szc is equal szc and
4150 * all its constituent pages are locked
4151 * hat_page_demote() that could have changed p_szc to
4152 * szc is already done and no new have page_demote()
4153 * can start for this large page.
4154 */
4155
4156 /*
4157 * we need to make sure same mapping size is used for
4158 * the same address range if there's a possibility the
4159 * adddress is already mapped because hat layer panics
4160 * when translation is loaded for the range already
4161 * mapped with a different page size. We achieve it
4162 * by always using largest page size possible subject
4163 * to the constraints of page size, segment page size
4164 * and page alignment. Since mappings are invalidated
4165 * when those constraints change and make it
4166 * impossible to use previously used mapping size no
4167 * mapping size conflicts should happen.
4168 */
4169
4170 chkszc:
4171 if ((pszc = ppa[0]->p_szc) == szc &&
4172 IS_P2ALIGNED(pfn, pages)) {
4173
4174 SEGVN_VMSTAT_FLTVNPAGES(30);
4175 #ifdef DEBUG
4176 for (i = 0; i < pages; i++) {
4177 ASSERT(PAGE_LOCKED(ppa[i]));
4178 ASSERT(!PP_ISFREE(ppa[i]));
4179 ASSERT(page_pptonum(ppa[i]) ==
4180 pfn + i);
4181 ASSERT(ppa[i]->p_szc == szc);
4182 ASSERT(ppa[i]->p_vnode == vp);
4183 ASSERT(ppa[i]->p_offset ==
4184 off + (i << PAGESHIFT));
4185 }
4186 #endif /* DEBUG */
4187 /*
4188 * All pages are of szc we need and they are
4189 * all locked so they can't change szc. load
4190 * translations.
4191 *
4192 * if page got promoted since last check
4193 * we don't need pplist.
4194 */
4195 if (pplist != NULL) {
4196 page_free_replacement_page(pplist);
4197 page_create_putback(pages);
4198 }
4199 if (PP_ISMIGRATE(ppa[0])) {
4200 page_migrate(seg, a, ppa, pages);
4201 }
4202 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4203 prot, vpprot);
4204 if (!xhat) {
4205 hat_memload_array_region(hat, a, pgsz,
4206 ppa, prot & vpprot, hat_flag,
4207 svd->rcookie);
4208 } else {
4209 /*
4210 * avoid large xhat mappings to FS
4211 * pages so that hat_page_demote()
4212 * doesn't need to check for xhat
4213 * large mappings.
4214 * Don't use regions with xhats.
4215 */
4216 for (i = 0; i < pages; i++) {
4217 hat_memload(hat,
4218 a + (i << PAGESHIFT),
4219 ppa[i], prot & vpprot,
4220 hat_flag);
4221 }
4222 }
4223
4224 if (!(hat_flag & HAT_LOAD_LOCK)) {
4225 for (i = 0; i < pages; i++) {
4226 page_unlock(ppa[i]);
4227 }
4228 }
4229 if (amp != NULL) {
4230 anon_array_exit(&an_cookie);
4231 ANON_LOCK_EXIT(&->a_rwlock);
4232 }
4233 goto next;
4234 }
4235
4236 /*
4237 * See if upsize is possible.
4238 */
4239 if (pszc > szc && szc < seg->s_szc &&
4240 (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4241 pgcnt_t aphase;
4242 uint_t pszc1 = MIN(pszc, seg->s_szc);
4243 ppgsz = page_get_pagesize(pszc1);
4244 ppages = btop(ppgsz);
4245 aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4246
4247 ASSERT(type != F_SOFTLOCK);
4248
4249 SEGVN_VMSTAT_FLTVNPAGES(31);
4250 if (aphase != P2PHASE(pfn, ppages)) {
4251 segvn_faultvnmpss_align_err4++;
4252 } else {
4253 SEGVN_VMSTAT_FLTVNPAGES(32);
4254 if (pplist != NULL) {
4255 page_t *pl = pplist;
4256 page_free_replacement_page(pl);
4257 page_create_putback(pages);
4258 }
4259 for (i = 0; i < pages; i++) {
4260 page_unlock(ppa[i]);
4261 }
4262 if (amp != NULL) {
4263 anon_array_exit(&an_cookie);
4264 ANON_LOCK_EXIT(&->a_rwlock);
4265 }
4266 pszc = pszc1;
4267 ierr = -2;
4268 break;
4269 }
4270 }
4271
4272 /*
4273 * check if we should use smallest mapping size.
4274 */
4275 upgrdfail = 0;
4276 if (szc == 0 || xhat ||
4277 (pszc >= szc &&
4278 !IS_P2ALIGNED(pfn, pages)) ||
4279 (pszc < szc &&
4280 !segvn_full_szcpages(ppa, szc, &upgrdfail,
4281 &pszc))) {
4282
4283 if (upgrdfail && type != F_SOFTLOCK) {
4284 /*
4285 * segvn_full_szcpages failed to lock
4286 * all pages EXCL. Size down.
4287 */
4288 ASSERT(pszc < szc);
4289
4290 SEGVN_VMSTAT_FLTVNPAGES(33);
4291
4292 if (pplist != NULL) {
4293 page_t *pl = pplist;
4294 page_free_replacement_page(pl);
4295 page_create_putback(pages);
4296 }
4297
4298 for (i = 0; i < pages; i++) {
4299 page_unlock(ppa[i]);
4300 }
4301 if (amp != NULL) {
4302 anon_array_exit(&an_cookie);
4303 ANON_LOCK_EXIT(&->a_rwlock);
4304 }
4305 ierr = -1;
4306 break;
4307 }
4308 if (szc != 0 && !xhat && !upgrdfail) {
4309 segvn_faultvnmpss_align_err5++;
4310 }
4311 SEGVN_VMSTAT_FLTVNPAGES(34);
4312 if (pplist != NULL) {
4313 page_free_replacement_page(pplist);
4314 page_create_putback(pages);
4315 }
4316 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4317 prot, vpprot);
4318 if (upgrdfail && segvn_anypgsz_vnode) {
4319 /* SOFTLOCK case */
4320 hat_memload_array_region(hat, a, pgsz,
4321 ppa, prot & vpprot, hat_flag,
4322 svd->rcookie);
4323 } else {
4324 for (i = 0; i < pages; i++) {
4325 hat_memload_region(hat,
4326 a + (i << PAGESHIFT),
4327 ppa[i], prot & vpprot,
4328 hat_flag, svd->rcookie);
4329 }
4330 }
4331 if (!(hat_flag & HAT_LOAD_LOCK)) {
4332 for (i = 0; i < pages; i++) {
4333 page_unlock(ppa[i]);
4334 }
4335 }
4336 if (amp != NULL) {
4337 anon_array_exit(&an_cookie);
4338 ANON_LOCK_EXIT(&->a_rwlock);
4339 }
4340 goto next;
4341 }
4342
4343 if (pszc == szc) {
4344 /*
4345 * segvn_full_szcpages() upgraded pages szc.
4346 */
4347 ASSERT(pszc == ppa[0]->p_szc);
4348 ASSERT(IS_P2ALIGNED(pfn, pages));
4349 goto chkszc;
4350 }
4351
4352 if (pszc > szc) {
4353 kmutex_t *szcmtx;
4354 SEGVN_VMSTAT_FLTVNPAGES(35);
4355 /*
4356 * p_szc of ppa[0] can change since we haven't
4357 * locked all constituent pages. Call
4358 * page_lock_szc() to prevent szc changes.
4359 * This should be a rare case that happens when
4360 * multiple segments use a different page size
4361 * to map the same file offsets.
4362 */
4363 szcmtx = page_szc_lock(ppa[0]);
4364 pszc = ppa[0]->p_szc;
4365 ASSERT(szcmtx != NULL || pszc == 0);
4366 ASSERT(ppa[0]->p_szc <= pszc);
4367 if (pszc <= szc) {
4368 SEGVN_VMSTAT_FLTVNPAGES(36);
4369 if (szcmtx != NULL) {
4370 mutex_exit(szcmtx);
4371 }
4372 goto chkszc;
4373 }
4374 if (pplist != NULL) {
4375 /*
4376 * page got promoted since last check.
4377 * we don't need preaalocated large
4378 * page.
4379 */
4380 SEGVN_VMSTAT_FLTVNPAGES(37);
4381 page_free_replacement_page(pplist);
4382 page_create_putback(pages);
4383 }
4384 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4385 prot, vpprot);
4386 hat_memload_array_region(hat, a, pgsz, ppa,
4387 prot & vpprot, hat_flag, svd->rcookie);
4388 mutex_exit(szcmtx);
4389 if (!(hat_flag & HAT_LOAD_LOCK)) {
4390 for (i = 0; i < pages; i++) {
4391 page_unlock(ppa[i]);
4392 }
4393 }
4394 if (amp != NULL) {
4395 anon_array_exit(&an_cookie);
4396 ANON_LOCK_EXIT(&->a_rwlock);
4397 }
4398 goto next;
4399 }
4400
4401 /*
4402 * if page got demoted since last check
4403 * we could have not allocated larger page.
4404 * allocate now.
4405 */
4406 if (pplist == NULL &&
4407 page_alloc_pages(vp, seg, a, &pplist, NULL,
4408 szc, 0, 0) && type != F_SOFTLOCK) {
4409 SEGVN_VMSTAT_FLTVNPAGES(38);
4410 for (i = 0; i < pages; i++) {
4411 page_unlock(ppa[i]);
4412 }
4413 if (amp != NULL) {
4414 anon_array_exit(&an_cookie);
4415 ANON_LOCK_EXIT(&->a_rwlock);
4416 }
4417 ierr = -1;
4418 alloc_failed |= (1 << szc);
4419 break;
4420 }
4421
4422 SEGVN_VMSTAT_FLTVNPAGES(39);
4423
4424 if (pplist != NULL) {
4425 segvn_relocate_pages(ppa, pplist);
4426 #ifdef DEBUG
4427 } else {
4428 ASSERT(type == F_SOFTLOCK);
4429 SEGVN_VMSTAT_FLTVNPAGES(40);
4430 #endif /* DEBUG */
4431 }
4432
4433 SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4434
4435 if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4436 ASSERT(type == F_SOFTLOCK);
4437 for (i = 0; i < pages; i++) {
4438 ASSERT(ppa[i]->p_szc < szc);
4439 hat_memload_region(hat,
4440 a + (i << PAGESHIFT),
4441 ppa[i], prot & vpprot, hat_flag,
4442 svd->rcookie);
4443 }
4444 } else {
4445 ASSERT(pplist != NULL || type == F_SOFTLOCK);
4446 hat_memload_array_region(hat, a, pgsz, ppa,
4447 prot & vpprot, hat_flag, svd->rcookie);
4448 }
4449 if (!(hat_flag & HAT_LOAD_LOCK)) {
4450 for (i = 0; i < pages; i++) {
4451 ASSERT(PAGE_SHARED(ppa[i]));
4452 page_unlock(ppa[i]);
4453 }
4454 }
4455 if (amp != NULL) {
4456 anon_array_exit(&an_cookie);
4457 ANON_LOCK_EXIT(&->a_rwlock);
4458 }
4459
4460 next:
4461 if (vpage != NULL) {
4462 vpage += pages;
4463 }
4464 adjszc_chk = 1;
4465 }
4466 if (a == lpgeaddr)
4467 break;
4468 ASSERT(a < lpgeaddr);
4469
4470 ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4471
4472 /*
4473 * ierr == -1 means we failed to map with a large page.
4474 * (either due to allocation/relocation failures or
4475 * misalignment with other mappings to this file.
4476 *
4477 * ierr == -2 means some other thread allocated a large page
4478 * after we gave up tp map with a large page. retry with
4479 * larger mapping.
4480 */
4481 ASSERT(ierr == -1 || ierr == -2);
4482 ASSERT(ierr == -2 || szc != 0);
4483 ASSERT(ierr == -1 || szc < seg->s_szc);
4484 if (ierr == -2) {
4485 SEGVN_VMSTAT_FLTVNPAGES(41);
4486 ASSERT(pszc > szc && pszc <= seg->s_szc);
4487 szc = pszc;
4488 } else if (segvn_anypgsz_vnode) {
4489 SEGVN_VMSTAT_FLTVNPAGES(42);
4490 szc--;
4491 } else {
4492 SEGVN_VMSTAT_FLTVNPAGES(43);
4493 ASSERT(pszc < szc);
4494 /*
4495 * other process created pszc large page.
4496 * but we still have to drop to 0 szc.
4497 */
4498 szc = 0;
4499 }
4500
4501 pgsz = page_get_pagesize(szc);
4502 pages = btop(pgsz);
4503 if (ierr == -2) {
4504 /*
4505 * Size up case. Note lpgaddr may only be needed for
4506 * softlock case so we don't adjust it here.
4507 */
4508 a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4509 ASSERT(a >= lpgaddr);
4510 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4511 off = svd->offset + (uintptr_t)(a - seg->s_base);
4512 aindx = svd->anon_index + seg_page(seg, a);
4513 vpage = (svd->vpage != NULL) ?
4514 &svd->vpage[seg_page(seg, a)] : NULL;
4515 } else {
4516 /*
4517 * Size down case. Note lpgaddr may only be needed for
4518 * softlock case so we don't adjust it here.
4519 */
4520 ASSERT(IS_P2ALIGNED(a, pgsz));
4521 ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4522 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4523 ASSERT(a < lpgeaddr);
4524 if (a < addr) {
4525 SEGVN_VMSTAT_FLTVNPAGES(44);
4526 /*
4527 * The beginning of the large page region can
4528 * be pulled to the right to make a smaller
4529 * region. We haven't yet faulted a single
4530 * page.
4531 */
4532 a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4533 ASSERT(a >= lpgaddr);
4534 off = svd->offset +
4535 (uintptr_t)(a - seg->s_base);
4536 aindx = svd->anon_index + seg_page(seg, a);
4537 vpage = (svd->vpage != NULL) ?
4538 &svd->vpage[seg_page(seg, a)] : NULL;
4539 }
4540 }
4541 }
4542 out:
4543 kmem_free(ppa, ppasize);
4544 if (!err && !vop_size_err) {
4545 SEGVN_VMSTAT_FLTVNPAGES(45);
4546 return (0);
4547 }
4548 if (type == F_SOFTLOCK && a > lpgaddr) {
4549 SEGVN_VMSTAT_FLTVNPAGES(46);
4550 segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4551 }
4552 if (!vop_size_err) {
4553 SEGVN_VMSTAT_FLTVNPAGES(47);
4554 return (err);
4555 }
4556 ASSERT(brkcow || tron || type == F_SOFTLOCK);
4557 /*
4558 * Large page end is mapped beyond the end of file and it's a cow
4559 * fault (can be a text replication induced cow) or softlock so we can't
4560 * reduce the map area. For now just demote the segment. This should
4561 * really only happen if the end of the file changed after the mapping
4562 * was established since when large page segments are created we make
4563 * sure they don't extend beyond the end of the file.
4564 */
4565 SEGVN_VMSTAT_FLTVNPAGES(48);
4566
4567 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4568 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4569 err = 0;
4570 if (seg->s_szc != 0) {
4571 segvn_fltvnpages_clrszc_cnt++;
4572 ASSERT(svd->softlockcnt == 0);
4573 err = segvn_clrszc(seg);
4574 if (err != 0) {
4575 segvn_fltvnpages_clrszc_err++;
4576 }
4577 }
4578 ASSERT(err || seg->s_szc == 0);
4579 SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4580 /* segvn_fault will do its job as if szc had been zero to begin with */
4581 return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4582 }
4583
4584 /*
4585 * This routine will attempt to fault in one large page.
4586 * it will use smaller pages if that fails.
4587 * It should only be called for pure anonymous segments.
4588 */
4589 static faultcode_t
4590 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4591 caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4592 caddr_t eaddr, int brkcow)
4593 {
4594 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4595 struct anon_map *amp = svd->amp;
4596 uchar_t segtype = svd->type;
4597 uint_t szc = seg->s_szc;
4598 size_t pgsz = page_get_pagesize(szc);
4599 size_t maxpgsz = pgsz;
4600 pgcnt_t pages = btop(pgsz);
4601 uint_t ppaszc = szc;
4602 caddr_t a = lpgaddr;
4603 ulong_t aindx = svd->anon_index + seg_page(seg, a);
4604 struct vpage *vpage = (svd->vpage != NULL) ?
4605 &svd->vpage[seg_page(seg, a)] : NULL;
4606 page_t **ppa;
4607 uint_t ppa_szc;
4608 faultcode_t err;
4609 int ierr;
4610 uint_t protchk, prot, vpprot;
4611 ulong_t i;
4612 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4613 anon_sync_obj_t cookie;
4614 int adjszc_chk;
4615 int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4616
4617 ASSERT(szc != 0);
4618 ASSERT(amp != NULL);
4619 ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4620 ASSERT(!(svd->flags & MAP_NORESERVE));
4621 ASSERT(type != F_SOFTUNLOCK);
4622 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4623 ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4624 ASSERT(svd->tr_state != SEGVN_TR_INIT);
4625
4626 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4627
4628 VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4629 VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4630
4631 if (svd->flags & MAP_TEXT) {
4632 hat_flag |= HAT_LOAD_TEXT;
4633 }
4634
4635 if (svd->pageprot) {
4636 switch (rw) {
4637 case S_READ:
4638 protchk = PROT_READ;
4639 break;
4640 case S_WRITE:
4641 protchk = PROT_WRITE;
4642 break;
4643 case S_EXEC:
4644 protchk = PROT_EXEC;
4645 break;
4646 case S_OTHER:
4647 default:
4648 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4649 break;
4650 }
4651 VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4652 } else {
4653 prot = svd->prot;
4654 /* caller has already done segment level protection check. */
4655 }
4656
4657 ppa = kmem_cache_alloc(segvn_szc_cache[ppaszc], KM_SLEEP);
4658 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
4659 for (;;) {
4660 adjszc_chk = 0;
4661 for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4662 if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4663 VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4664 ASSERT(vpage != NULL);
4665 prot = VPP_PROT(vpage);
4666 ASSERT(sameprot(seg, a, maxpgsz));
4667 if ((prot & protchk) == 0) {
4668 err = FC_PROT;
4669 goto error;
4670 }
4671 }
4672 if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4673 pgsz < maxpgsz) {
4674 ASSERT(a > lpgaddr);
4675 szc = seg->s_szc;
4676 pgsz = maxpgsz;
4677 pages = btop(pgsz);
4678 ASSERT(IS_P2ALIGNED(aindx, pages));
4679 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4680 pgsz);
4681 }
4682 if (type == F_SOFTLOCK) {
4683 atomic_add_long((ulong_t *)&svd->softlockcnt,
4684 pages);
4685 }
4686 anon_array_enter(amp, aindx, &cookie);
4687 ppa_szc = (uint_t)-1;
4688 ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4689 prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4690 segvn_anypgsz, pgflags, svd->cred);
4691 if (ierr != 0) {
4692 anon_array_exit(&cookie);
4693 VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4694 if (type == F_SOFTLOCK) {
4695 atomic_add_long(
4696 (ulong_t *)&svd->softlockcnt,
4697 -pages);
4698 }
4699 if (ierr > 0) {
4700 VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4701 err = FC_MAKE_ERR(ierr);
4702 goto error;
4703 }
4704 break;
4705 }
4706
4707 ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4708
4709 ASSERT(segtype == MAP_SHARED ||
4710 ppa[0]->p_szc <= szc);
4711 ASSERT(segtype == MAP_PRIVATE ||
4712 ppa[0]->p_szc >= szc);
4713
4714 /*
4715 * Handle pages that have been marked for migration
4716 */
4717 if (lgrp_optimizations())
4718 page_migrate(seg, a, ppa, pages);
4719
4720 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4721
4722 if (segtype == MAP_SHARED) {
4723 vpprot |= PROT_WRITE;
4724 }
4725
4726 hat_memload_array(hat, a, pgsz, ppa,
4727 prot & vpprot, hat_flag);
4728
4729 if (hat_flag & HAT_LOAD_LOCK) {
4730 VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4731 } else {
4732 VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4733 for (i = 0; i < pages; i++)
4734 page_unlock(ppa[i]);
4735 }
4736 if (vpage != NULL)
4737 vpage += pages;
4738
4739 anon_array_exit(&cookie);
4740 adjszc_chk = 1;
4741 }
4742 if (a == lpgeaddr)
4743 break;
4744 ASSERT(a < lpgeaddr);
4745 /*
4746 * ierr == -1 means we failed to allocate a large page.
4747 * so do a size down operation.
4748 *
4749 * ierr == -2 means some other process that privately shares
4750 * pages with this process has allocated a larger page and we
4751 * need to retry with larger pages. So do a size up
4752 * operation. This relies on the fact that large pages are
4753 * never partially shared i.e. if we share any constituent
4754 * page of a large page with another process we must share the
4755 * entire large page. Note this cannot happen for SOFTLOCK
4756 * case, unless current address (a) is at the beginning of the
4757 * next page size boundary because the other process couldn't
4758 * have relocated locked pages.
4759 */
4760 ASSERT(ierr == -1 || ierr == -2);
4761
4762 if (segvn_anypgsz) {
4763 ASSERT(ierr == -2 || szc != 0);
4764 ASSERT(ierr == -1 || szc < seg->s_szc);
4765 szc = (ierr == -1) ? szc - 1 : szc + 1;
4766 } else {
4767 /*
4768 * For non COW faults and segvn_anypgsz == 0
4769 * we need to be careful not to loop forever
4770 * if existing page is found with szc other
4771 * than 0 or seg->s_szc. This could be due
4772 * to page relocations on behalf of DR or
4773 * more likely large page creation. For this
4774 * case simply re-size to existing page's szc
4775 * if returned by anon_map_getpages().
4776 */
4777 if (ppa_szc == (uint_t)-1) {
4778 szc = (ierr == -1) ? 0 : seg->s_szc;
4779 } else {
4780 ASSERT(ppa_szc <= seg->s_szc);
4781 ASSERT(ierr == -2 || ppa_szc < szc);
4782 ASSERT(ierr == -1 || ppa_szc > szc);
4783 szc = ppa_szc;
4784 }
4785 }
4786
4787 pgsz = page_get_pagesize(szc);
4788 pages = btop(pgsz);
4789 ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4790 (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4791 if (type == F_SOFTLOCK) {
4792 /*
4793 * For softlocks we cannot reduce the fault area
4794 * (calculated based on the largest page size for this
4795 * segment) for size down and a is already next
4796 * page size aligned as assertted above for size
4797 * ups. Therefore just continue in case of softlock.
4798 */
4799 VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4800 continue; /* keep lint happy */
4801 } else if (ierr == -2) {
4802
4803 /*
4804 * Size up case. Note lpgaddr may only be needed for
4805 * softlock case so we don't adjust it here.
4806 */
4807 VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4808 a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4809 ASSERT(a >= lpgaddr);
4810 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4811 aindx = svd->anon_index + seg_page(seg, a);
4812 vpage = (svd->vpage != NULL) ?
4813 &svd->vpage[seg_page(seg, a)] : NULL;
4814 } else {
4815 /*
4816 * Size down case. Note lpgaddr may only be needed for
4817 * softlock case so we don't adjust it here.
4818 */
4819 VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4820 ASSERT(IS_P2ALIGNED(a, pgsz));
4821 ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4822 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4823 ASSERT(a < lpgeaddr);
4824 if (a < addr) {
4825 /*
4826 * The beginning of the large page region can
4827 * be pulled to the right to make a smaller
4828 * region. We haven't yet faulted a single
4829 * page.
4830 */
4831 VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4832 a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4833 ASSERT(a >= lpgaddr);
4834 aindx = svd->anon_index + seg_page(seg, a);
4835 vpage = (svd->vpage != NULL) ?
4836 &svd->vpage[seg_page(seg, a)] : NULL;
4837 }
4838 }
4839 }
4840 VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4841 ANON_LOCK_EXIT(&->a_rwlock);
4842 kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4843 return (0);
4844 error:
4845 VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4846 ANON_LOCK_EXIT(&->a_rwlock);
4847 kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4848 if (type == F_SOFTLOCK && a > lpgaddr) {
4849 VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4850 segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4851 }
4852 return (err);
4853 }
4854
4855 int fltadvice = 1; /* set to free behind pages for sequential access */
4856
4857 /*
4858 * This routine is called via a machine specific fault handling routine.
4859 * It is also called by software routines wishing to lock or unlock
4860 * a range of addresses.
4861 *
4862 * Here is the basic algorithm:
4863 * If unlocking
4864 * Call segvn_softunlock
4865 * Return
4866 * endif
4867 * Checking and set up work
4868 * If we will need some non-anonymous pages
4869 * Call VOP_GETPAGE over the range of non-anonymous pages
4870 * endif
4871 * Loop over all addresses requested
4872 * Call segvn_faultpage passing in page list
4873 * to load up translations and handle anonymous pages
4874 * endloop
4875 * Load up translation to any additional pages in page list not
4876 * already handled that fit into this segment
4877 */
4878 static faultcode_t
4879 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4880 enum fault_type type, enum seg_rw rw)
4881 {
4882 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4883 page_t **plp, **ppp, *pp;
4884 u_offset_t off;
4885 caddr_t a;
4886 struct vpage *vpage;
4887 uint_t vpprot, prot;
4888 int err;
4889 page_t *pl[PVN_GETPAGE_NUM + 1];
4890 size_t plsz, pl_alloc_sz;
4891 size_t page;
4892 ulong_t anon_index;
4893 struct anon_map *amp;
4894 int dogetpage = 0;
4895 caddr_t lpgaddr, lpgeaddr;
4896 size_t pgsz;
4897 anon_sync_obj_t cookie;
4898 int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4899
4900 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
4901 ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4902
4903 /*
4904 * First handle the easy stuff
4905 */
4906 if (type == F_SOFTUNLOCK) {
4907 if (rw == S_READ_NOCOW) {
4908 rw = S_READ;
4909 ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
4910 }
4911 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4912 pgsz = (seg->s_szc == 0) ? PAGESIZE :
4913 page_get_pagesize(seg->s_szc);
4914 VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4915 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4916 segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4917 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4918 return (0);
4919 }
4920
4921 ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4922 !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4923 if (brkcow == 0) {
4924 if (svd->tr_state == SEGVN_TR_INIT) {
4925 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4926 if (svd->tr_state == SEGVN_TR_INIT) {
4927 ASSERT(svd->vp != NULL && svd->amp == NULL);
4928 ASSERT(svd->flags & MAP_TEXT);
4929 ASSERT(svd->type == MAP_PRIVATE);
4930 segvn_textrepl(seg);
4931 ASSERT(svd->tr_state != SEGVN_TR_INIT);
4932 ASSERT(svd->tr_state != SEGVN_TR_ON ||
4933 svd->amp != NULL);
4934 }
4935 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4936 }
4937 } else if (svd->tr_state != SEGVN_TR_OFF) {
4938 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4939
4940 if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4941 ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4942 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4943 return (FC_PROT);
4944 }
4945
4946 if (svd->tr_state == SEGVN_TR_ON) {
4947 ASSERT(svd->vp != NULL && svd->amp != NULL);
4948 segvn_textunrepl(seg, 0);
4949 ASSERT(svd->amp == NULL &&
4950 svd->tr_state == SEGVN_TR_OFF);
4951 } else if (svd->tr_state != SEGVN_TR_OFF) {
4952 svd->tr_state = SEGVN_TR_OFF;
4953 }
4954 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4955 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4956 }
4957
4958 top:
4959 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4960
4961 /*
4962 * If we have the same protections for the entire segment,
4963 * insure that the access being attempted is legitimate.
4964 */
4965
4966 if (svd->pageprot == 0) {
4967 uint_t protchk;
4968
4969 switch (rw) {
4970 case S_READ:
4971 case S_READ_NOCOW:
4972 protchk = PROT_READ;
4973 break;
4974 case S_WRITE:
4975 protchk = PROT_WRITE;
4976 break;
4977 case S_EXEC:
4978 protchk = PROT_EXEC;
4979 break;
4980 case S_OTHER:
4981 default:
4982 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4983 break;
4984 }
4985
4986 if ((svd->prot & protchk) == 0) {
4987 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4988 return (FC_PROT); /* illegal access type */
4989 }
4990 }
4991
4992 if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
4993 /* this must be SOFTLOCK S_READ fault */
4994 ASSERT(svd->amp == NULL);
4995 ASSERT(svd->tr_state == SEGVN_TR_OFF);
4996 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4997 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4998 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
4999 /*
5000 * this must be the first ever non S_READ_NOCOW
5001 * softlock for this segment.
5002 */
5003 ASSERT(svd->softlockcnt == 0);
5004 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5005 HAT_REGION_TEXT);
5006 svd->rcookie = HAT_INVALID_REGION_COOKIE;
5007 }
5008 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5009 goto top;
5010 }
5011
5012 /*
5013 * We can't allow the long term use of softlocks for vmpss segments,
5014 * because in some file truncation cases we should be able to demote
5015 * the segment, which requires that there are no softlocks. The
5016 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5017 * segment is S_READ_NOCOW, where the caller holds the address space
5018 * locked as writer and calls softunlock before dropping the as lock.
5019 * S_READ_NOCOW is used by /proc to read memory from another user.
5020 *
5021 * Another deadlock between SOFTLOCK and file truncation can happen
5022 * because segvn_fault_vnodepages() calls the FS one pagesize at
5023 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5024 * can cause a deadlock because the first set of page_t's remain
5025 * locked SE_SHARED. To avoid this, we demote segments on a first
5026 * SOFTLOCK if they have a length greater than the segment's
5027 * page size.
5028 *
5029 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5030 * the access type is S_READ_NOCOW and the fault length is less than
5031 * or equal to the segment's page size. While this is quite restrictive,
5032 * it should be the most common case of SOFTLOCK against a vmpss
5033 * segment.
5034 *
5035 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5036 * caller makes sure no COW will be caused by another thread for a
5037 * softlocked page.
5038 */
5039 if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5040 int demote = 0;
5041
5042 if (rw != S_READ_NOCOW) {
5043 demote = 1;
5044 }
5045 if (!demote && len > PAGESIZE) {
5046 pgsz = page_get_pagesize(seg->s_szc);
5047 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5048 lpgeaddr);
5049 if (lpgeaddr - lpgaddr > pgsz) {
5050 demote = 1;
5051 }
5052 }
5053
5054 ASSERT(demote || AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5055
5056 if (demote) {
5057 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5058 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5059 if (seg->s_szc != 0) {
5060 segvn_vmpss_clrszc_cnt++;
5061 ASSERT(svd->softlockcnt == 0);
5062 err = segvn_clrszc(seg);
5063 if (err) {
5064 segvn_vmpss_clrszc_err++;
5065 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5066 return (FC_MAKE_ERR(err));
5067 }
5068 }
5069 ASSERT(seg->s_szc == 0);
5070 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5071 goto top;
5072 }
5073 }
5074
5075 /*
5076 * Check to see if we need to allocate an anon_map structure.
5077 */
5078 if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5079 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5080 /*
5081 * Drop the "read" lock on the segment and acquire
5082 * the "write" version since we have to allocate the
5083 * anon_map.
5084 */
5085 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5086 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5087
5088 if (svd->amp == NULL) {
5089 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5090 svd->amp->a_szc = seg->s_szc;
5091 }
5092 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5093
5094 /*
5095 * Start all over again since segment protections
5096 * may have changed after we dropped the "read" lock.
5097 */
5098 goto top;
5099 }
5100
5101 /*
5102 * S_READ_NOCOW vs S_READ distinction was
5103 * only needed for the code above. After
5104 * that we treat it as S_READ.
5105 */
5106 if (rw == S_READ_NOCOW) {
5107 ASSERT(type == F_SOFTLOCK);
5108 ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5109 rw = S_READ;
5110 }
5111
5112 amp = svd->amp;
5113
5114 /*
5115 * MADV_SEQUENTIAL work is ignored for large page segments.
5116 */
5117 if (seg->s_szc != 0) {
5118 pgsz = page_get_pagesize(seg->s_szc);
5119 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5120 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5121 if (svd->vp == NULL) {
5122 err = segvn_fault_anonpages(hat, seg, lpgaddr,
5123 lpgeaddr, type, rw, addr, addr + len, brkcow);
5124 } else {
5125 err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5126 lpgeaddr, type, rw, addr, addr + len, brkcow);
5127 if (err == IE_RETRY) {
5128 ASSERT(seg->s_szc == 0);
5129 ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5130 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5131 goto top;
5132 }
5133 }
5134 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5135 return (err);
5136 }
5137
5138 page = seg_page(seg, addr);
5139 if (amp != NULL) {
5140 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5141 anon_index = svd->anon_index + page;
5142
5143 if (type == F_PROT && rw == S_READ &&
5144 svd->tr_state == SEGVN_TR_OFF &&
5145 svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5146 size_t index = anon_index;
5147 struct anon *ap;
5148
5149 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5150 /*
5151 * The fast path could apply to S_WRITE also, except
5152 * that the protection fault could be caused by lazy
5153 * tlb flush when ro->rw. In this case, the pte is
5154 * RW already. But RO in the other cpu's tlb causes
5155 * the fault. Since hat_chgprot won't do anything if
5156 * pte doesn't change, we may end up faulting
5157 * indefinitely until the RO tlb entry gets replaced.
5158 */
5159 for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5160 anon_array_enter(amp, index, &cookie);
5161 ap = anon_get_ptr(amp->ahp, index);
5162 anon_array_exit(&cookie);
5163 if ((ap == NULL) || (ap->an_refcnt != 1)) {
5164 ANON_LOCK_EXIT(&->a_rwlock);
5165 goto slow;
5166 }
5167 }
5168 hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5169 ANON_LOCK_EXIT(&->a_rwlock);
5170 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5171 return (0);
5172 }
5173 }
5174 slow:
5175
5176 if (svd->vpage == NULL)
5177 vpage = NULL;
5178 else
5179 vpage = &svd->vpage[page];
5180
5181 off = svd->offset + (uintptr_t)(addr - seg->s_base);
5182
5183 /*
5184 * If MADV_SEQUENTIAL has been set for the particular page we
5185 * are faulting on, free behind all pages in the segment and put
5186 * them on the free list.
5187 */
5188
5189 if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5190 struct vpage *vpp;
5191 ulong_t fanon_index;
5192 size_t fpage;
5193 u_offset_t pgoff, fpgoff;
5194 struct vnode *fvp;
5195 struct anon *fap = NULL;
5196
5197 if (svd->advice == MADV_SEQUENTIAL ||
5198 (svd->pageadvice &&
5199 VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5200 pgoff = off - PAGESIZE;
5201 fpage = page - 1;
5202 if (vpage != NULL)
5203 vpp = &svd->vpage[fpage];
5204 if (amp != NULL)
5205 fanon_index = svd->anon_index + fpage;
5206
5207 while (pgoff > svd->offset) {
5208 if (svd->advice != MADV_SEQUENTIAL &&
5209 (!svd->pageadvice || (vpage &&
5210 VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5211 break;
5212
5213 /*
5214 * If this is an anon page, we must find the
5215 * correct <vp, offset> for it
5216 */
5217 fap = NULL;
5218 if (amp != NULL) {
5219 ANON_LOCK_ENTER(&->a_rwlock,
5220 RW_READER);
5221 anon_array_enter(amp, fanon_index,
5222 &cookie);
5223 fap = anon_get_ptr(amp->ahp,
5224 fanon_index);
5225 if (fap != NULL) {
5226 swap_xlate(fap, &fvp, &fpgoff);
5227 } else {
5228 fpgoff = pgoff;
5229 fvp = svd->vp;
5230 }
5231 anon_array_exit(&cookie);
5232 ANON_LOCK_EXIT(&->a_rwlock);
5233 } else {
5234 fpgoff = pgoff;
5235 fvp = svd->vp;
5236 }
5237 if (fvp == NULL)
5238 break; /* XXX */
5239 /*
5240 * Skip pages that are free or have an
5241 * "exclusive" lock.
5242 */
5243 pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5244 if (pp == NULL)
5245 break;
5246 /*
5247 * We don't need the page_struct_lock to test
5248 * as this is only advisory; even if we
5249 * acquire it someone might race in and lock
5250 * the page after we unlock and before the
5251 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5252 */
5253 if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5254 /*
5255 * Hold the vnode before releasing
5256 * the page lock to prevent it from
5257 * being freed and re-used by some
5258 * other thread.
5259 */
5260 VN_HOLD(fvp);
5261 page_unlock(pp);
5262 /*
5263 * We should build a page list
5264 * to kluster putpages XXX
5265 */
5266 (void) VOP_PUTPAGE(fvp,
5267 (offset_t)fpgoff, PAGESIZE,
5268 (B_DONTNEED|B_FREE|B_ASYNC),
5269 svd->cred, NULL);
5270 VN_RELE(fvp);
5271 } else {
5272 /*
5273 * XXX - Should the loop terminate if
5274 * the page is `locked'?
5275 */
5276 page_unlock(pp);
5277 }
5278 --vpp;
5279 --fanon_index;
5280 pgoff -= PAGESIZE;
5281 }
5282 }
5283 }
5284
5285 plp = pl;
5286 *plp = NULL;
5287 pl_alloc_sz = 0;
5288
5289 /*
5290 * See if we need to call VOP_GETPAGE for
5291 * *any* of the range being faulted on.
5292 * We can skip all of this work if there
5293 * was no original vnode.
5294 */
5295 if (svd->vp != NULL) {
5296 u_offset_t vp_off;
5297 size_t vp_len;
5298 struct anon *ap;
5299 vnode_t *vp;
5300
5301 vp_off = off;
5302 vp_len = len;
5303
5304 if (amp == NULL)
5305 dogetpage = 1;
5306 else {
5307 /*
5308 * Only acquire reader lock to prevent amp->ahp
5309 * from being changed. It's ok to miss pages,
5310 * hence we don't do anon_array_enter
5311 */
5312 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5313 ap = anon_get_ptr(amp->ahp, anon_index);
5314
5315 if (len <= PAGESIZE)
5316 /* inline non_anon() */
5317 dogetpage = (ap == NULL);
5318 else
5319 dogetpage = non_anon(amp->ahp, anon_index,
5320 &vp_off, &vp_len);
5321 ANON_LOCK_EXIT(&->a_rwlock);
5322 }
5323
5324 if (dogetpage) {
5325 enum seg_rw arw;
5326 struct as *as = seg->s_as;
5327
5328 if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5329 /*
5330 * Page list won't fit in local array,
5331 * allocate one of the needed size.
5332 */
5333 pl_alloc_sz =
5334 (btop(len) + 1) * sizeof (page_t *);
5335 plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5336 plp[0] = NULL;
5337 plsz = len;
5338 } else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5339 svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5340 (((size_t)(addr + PAGESIZE) <
5341 (size_t)(seg->s_base + seg->s_size)) &&
5342 hat_probe(as->a_hat, addr + PAGESIZE))) {
5343 /*
5344 * Ask VOP_GETPAGE to return the exact number
5345 * of pages if
5346 * (a) this is a COW fault, or
5347 * (b) this is a software fault, or
5348 * (c) next page is already mapped.
5349 */
5350 plsz = len;
5351 } else {
5352 /*
5353 * Ask VOP_GETPAGE to return adjacent pages
5354 * within the segment.
5355 */
5356 plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5357 ((seg->s_base + seg->s_size) - addr));
5358 ASSERT((addr + plsz) <=
5359 (seg->s_base + seg->s_size));
5360 }
5361
5362 /*
5363 * Need to get some non-anonymous pages.
5364 * We need to make only one call to GETPAGE to do
5365 * this to prevent certain deadlocking conditions
5366 * when we are doing locking. In this case
5367 * non_anon() should have picked up the smallest
5368 * range which includes all the non-anonymous
5369 * pages in the requested range. We have to
5370 * be careful regarding which rw flag to pass in
5371 * because on a private mapping, the underlying
5372 * object is never allowed to be written.
5373 */
5374 if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5375 arw = S_READ;
5376 } else {
5377 arw = rw;
5378 }
5379 vp = svd->vp;
5380 TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5381 "segvn_getpage:seg %p addr %p vp %p",
5382 seg, addr, vp);
5383 err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5384 &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5385 svd->cred, NULL);
5386 if (err) {
5387 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5388 segvn_pagelist_rele(plp);
5389 if (pl_alloc_sz)
5390 kmem_free(plp, pl_alloc_sz);
5391 return (FC_MAKE_ERR(err));
5392 }
5393 if (svd->type == MAP_PRIVATE)
5394 vpprot &= ~PROT_WRITE;
5395 }
5396 }
5397
5398 /*
5399 * N.B. at this time the plp array has all the needed non-anon
5400 * pages in addition to (possibly) having some adjacent pages.
5401 */
5402
5403 /*
5404 * Always acquire the anon_array_lock to prevent
5405 * 2 threads from allocating separate anon slots for
5406 * the same "addr".
5407 *
5408 * If this is a copy-on-write fault and we don't already
5409 * have the anon_array_lock, acquire it to prevent the
5410 * fault routine from handling multiple copy-on-write faults
5411 * on the same "addr" in the same address space.
5412 *
5413 * Only one thread should deal with the fault since after
5414 * it is handled, the other threads can acquire a translation
5415 * to the newly created private page. This prevents two or
5416 * more threads from creating different private pages for the
5417 * same fault.
5418 *
5419 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5420 * to prevent deadlock between this thread and another thread
5421 * which has soft-locked this page and wants to acquire serial_lock.
5422 * ( bug 4026339 )
5423 *
5424 * The fix for bug 4026339 becomes unnecessary when using the
5425 * locking scheme with per amp rwlock and a global set of hash
5426 * lock, anon_array_lock. If we steal a vnode page when low
5427 * on memory and upgrad the page lock through page_rename,
5428 * then the page is PAGE_HANDLED, nothing needs to be done
5429 * for this page after returning from segvn_faultpage.
5430 *
5431 * But really, the page lock should be downgraded after
5432 * the stolen page is page_rename'd.
5433 */
5434
5435 if (amp != NULL)
5436 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5437
5438 /*
5439 * Ok, now loop over the address range and handle faults
5440 */
5441 for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5442 err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5443 type, rw, brkcow);
5444 if (err) {
5445 if (amp != NULL)
5446 ANON_LOCK_EXIT(&->a_rwlock);
5447 if (type == F_SOFTLOCK && a > addr) {
5448 segvn_softunlock(seg, addr, (a - addr),
5449 S_OTHER);
5450 }
5451 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5452 segvn_pagelist_rele(plp);
5453 if (pl_alloc_sz)
5454 kmem_free(plp, pl_alloc_sz);
5455 return (err);
5456 }
5457 if (vpage) {
5458 vpage++;
5459 } else if (svd->vpage) {
5460 page = seg_page(seg, addr);
5461 vpage = &svd->vpage[++page];
5462 }
5463 }
5464
5465 /* Didn't get pages from the underlying fs so we're done */
5466 if (!dogetpage)
5467 goto done;
5468
5469 /*
5470 * Now handle any other pages in the list returned.
5471 * If the page can be used, load up the translations now.
5472 * Note that the for loop will only be entered if "plp"
5473 * is pointing to a non-NULL page pointer which means that
5474 * VOP_GETPAGE() was called and vpprot has been initialized.
5475 */
5476 if (svd->pageprot == 0)
5477 prot = svd->prot & vpprot;
5478
5479
5480 /*
5481 * Large Files: diff should be unsigned value because we started
5482 * supporting > 2GB segment sizes from 2.5.1 and when a
5483 * large file of size > 2GB gets mapped to address space
5484 * the diff value can be > 2GB.
5485 */
5486
5487 for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5488 size_t diff;
5489 struct anon *ap;
5490 int anon_index;
5491 anon_sync_obj_t cookie;
5492 int hat_flag = HAT_LOAD_ADV;
5493
5494 if (svd->flags & MAP_TEXT) {
5495 hat_flag |= HAT_LOAD_TEXT;
5496 }
5497
5498 if (pp == PAGE_HANDLED)
5499 continue;
5500
5501 if (svd->tr_state != SEGVN_TR_ON &&
5502 pp->p_offset >= svd->offset &&
5503 pp->p_offset < svd->offset + seg->s_size) {
5504
5505 diff = pp->p_offset - svd->offset;
5506
5507 /*
5508 * Large Files: Following is the assertion
5509 * validating the above cast.
5510 */
5511 ASSERT(svd->vp == pp->p_vnode);
5512
5513 page = btop(diff);
5514 if (svd->pageprot)
5515 prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5516
5517 /*
5518 * Prevent other threads in the address space from
5519 * creating private pages (i.e., allocating anon slots)
5520 * while we are in the process of loading translations
5521 * to additional pages returned by the underlying
5522 * object.
5523 */
5524 if (amp != NULL) {
5525 anon_index = svd->anon_index + page;
5526 anon_array_enter(amp, anon_index, &cookie);
5527 ap = anon_get_ptr(amp->ahp, anon_index);
5528 }
5529 if ((amp == NULL) || (ap == NULL)) {
5530 if (IS_VMODSORT(pp->p_vnode) ||
5531 enable_mbit_wa) {
5532 if (rw == S_WRITE)
5533 hat_setmod(pp);
5534 else if (rw != S_OTHER &&
5535 !hat_ismod(pp))
5536 prot &= ~PROT_WRITE;
5537 }
5538 /*
5539 * Skip mapping read ahead pages marked
5540 * for migration, so they will get migrated
5541 * properly on fault
5542 */
5543 ASSERT(amp == NULL ||
5544 svd->rcookie == HAT_INVALID_REGION_COOKIE);
5545 if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5546 hat_memload_region(hat,
5547 seg->s_base + diff,
5548 pp, prot, hat_flag,
5549 svd->rcookie);
5550 }
5551 }
5552 if (amp != NULL)
5553 anon_array_exit(&cookie);
5554 }
5555 page_unlock(pp);
5556 }
5557 done:
5558 if (amp != NULL)
5559 ANON_LOCK_EXIT(&->a_rwlock);
5560 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5561 if (pl_alloc_sz)
5562 kmem_free(plp, pl_alloc_sz);
5563 return (0);
5564 }
5565
5566 /*
5567 * This routine is used to start I/O on pages asynchronously. XXX it will
5568 * only create PAGESIZE pages. At fault time they will be relocated into
5569 * larger pages.
5570 */
5571 static faultcode_t
5572 segvn_faulta(struct seg *seg, caddr_t addr)
5573 {
5574 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5575 int err;
5576 struct anon_map *amp;
5577 vnode_t *vp;
5578
5579 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5580
5581 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5582 if ((amp = svd->amp) != NULL) {
5583 struct anon *ap;
5584
5585 /*
5586 * Reader lock to prevent amp->ahp from being changed.
5587 * This is advisory, it's ok to miss a page, so
5588 * we don't do anon_array_enter lock.
5589 */
5590 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5591 if ((ap = anon_get_ptr(amp->ahp,
5592 svd->anon_index + seg_page(seg, addr))) != NULL) {
5593
5594 err = anon_getpage(&ap, NULL, NULL,
5595 0, seg, addr, S_READ, svd->cred);
5596
5597 ANON_LOCK_EXIT(&->a_rwlock);
5598 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5599 if (err)
5600 return (FC_MAKE_ERR(err));
5601 return (0);
5602 }
5603 ANON_LOCK_EXIT(&->a_rwlock);
5604 }
5605
5606 if (svd->vp == NULL) {
5607 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5608 return (0); /* zfod page - do nothing now */
5609 }
5610
5611 vp = svd->vp;
5612 TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5613 "segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5614 err = VOP_GETPAGE(vp,
5615 (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5616 PAGESIZE, NULL, NULL, 0, seg, addr,
5617 S_OTHER, svd->cred, NULL);
5618
5619 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5620 if (err)
5621 return (FC_MAKE_ERR(err));
5622 return (0);
5623 }
5624
5625 static int
5626 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5627 {
5628 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5629 struct vpage *cvp, *svp, *evp;
5630 struct vnode *vp;
5631 size_t pgsz;
5632 pgcnt_t pgcnt;
5633 anon_sync_obj_t cookie;
5634 int unload_done = 0;
5635
5636 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5637
5638 if ((svd->maxprot & prot) != prot)
5639 return (EACCES); /* violated maxprot */
5640
5641 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5642
5643 /* return if prot is the same */
5644 if (!svd->pageprot && svd->prot == prot) {
5645 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5646 return (0);
5647 }
5648
5649 /*
5650 * Since we change protections we first have to flush the cache.
5651 * This makes sure all the pagelock calls have to recheck
5652 * protections.
5653 */
5654 if (svd->softlockcnt > 0) {
5655 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5656
5657 /*
5658 * If this is shared segment non 0 softlockcnt
5659 * means locked pages are still in use.
5660 */
5661 if (svd->type == MAP_SHARED) {
5662 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5663 return (EAGAIN);
5664 }
5665
5666 /*
5667 * Since we do have the segvn writers lock nobody can fill
5668 * the cache with entries belonging to this seg during
5669 * the purge. The flush either succeeds or we still have
5670 * pending I/Os.
5671 */
5672 segvn_purge(seg);
5673 if (svd->softlockcnt > 0) {
5674 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5675 return (EAGAIN);
5676 }
5677 }
5678
5679 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5680 ASSERT(svd->amp == NULL);
5681 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5682 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5683 HAT_REGION_TEXT);
5684 svd->rcookie = HAT_INVALID_REGION_COOKIE;
5685 unload_done = 1;
5686 } else if (svd->tr_state == SEGVN_TR_INIT) {
5687 svd->tr_state = SEGVN_TR_OFF;
5688 } else if (svd->tr_state == SEGVN_TR_ON) {
5689 ASSERT(svd->amp != NULL);
5690 segvn_textunrepl(seg, 0);
5691 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5692 unload_done = 1;
5693 }
5694
5695 if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5696 svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5697 ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5698 segvn_inval_trcache(svd->vp);
5699 }
5700 if (seg->s_szc != 0) {
5701 int err;
5702 pgsz = page_get_pagesize(seg->s_szc);
5703 pgcnt = pgsz >> PAGESHIFT;
5704 ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5705 if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5706 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5707 ASSERT(seg->s_base != addr || seg->s_size != len);
5708 /*
5709 * If we are holding the as lock as a reader then
5710 * we need to return IE_RETRY and let the as
5711 * layer drop and re-acquire the lock as a writer.
5712 */
5713 if (AS_READ_HELD(seg->s_as, &seg->s_as->a_lock))
5714 return (IE_RETRY);
5715 VM_STAT_ADD(segvnvmstats.demoterange[1]);
5716 if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5717 err = segvn_demote_range(seg, addr, len,
5718 SDR_END, 0);
5719 } else {
5720 uint_t szcvec = map_pgszcvec(seg->s_base,
5721 pgsz, (uintptr_t)seg->s_base,
5722 (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5723 err = segvn_demote_range(seg, addr, len,
5724 SDR_END, szcvec);
5725 }
5726 if (err == 0)
5727 return (IE_RETRY);
5728 if (err == ENOMEM)
5729 return (IE_NOMEM);
5730 return (err);
5731 }
5732 }
5733
5734
5735 /*
5736 * If it's a private mapping and we're making it writable then we
5737 * may have to reserve the additional swap space now. If we are
5738 * making writable only a part of the segment then we use its vpage
5739 * array to keep a record of the pages for which we have reserved
5740 * swap. In this case we set the pageswap field in the segment's
5741 * segvn structure to record this.
5742 *
5743 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5744 * removing write permission on the entire segment and we haven't
5745 * modified any pages, we can release the swap space.
5746 */
5747 if (svd->type == MAP_PRIVATE) {
5748 if (prot & PROT_WRITE) {
5749 if (!(svd->flags & MAP_NORESERVE) &&
5750 !(svd->swresv && svd->pageswap == 0)) {
5751 size_t sz = 0;
5752
5753 /*
5754 * Start by determining how much swap
5755 * space is required.
5756 */
5757 if (addr == seg->s_base &&
5758 len == seg->s_size &&
5759 svd->pageswap == 0) {
5760 /* The whole segment */
5761 sz = seg->s_size;
5762 } else {
5763 /*
5764 * Make sure that the vpage array
5765 * exists, and make a note of the
5766 * range of elements corresponding
5767 * to len.
5768 */
5769 segvn_vpage(seg);
5770 svp = &svd->vpage[seg_page(seg, addr)];
5771 evp = &svd->vpage[seg_page(seg,
5772 addr + len)];
5773
5774 if (svd->pageswap == 0) {
5775 /*
5776 * This is the first time we've
5777 * asked for a part of this
5778 * segment, so we need to
5779 * reserve everything we've
5780 * been asked for.
5781 */
5782 sz = len;
5783 } else {
5784 /*
5785 * We have to count the number
5786 * of pages required.
5787 */
5788 for (cvp = svp; cvp < evp;
5789 cvp++) {
5790 if (!VPP_ISSWAPRES(cvp))
5791 sz++;
5792 }
5793 sz <<= PAGESHIFT;
5794 }
5795 }
5796
5797 /* Try to reserve the necessary swap. */
5798 if (anon_resv_zone(sz,
5799 seg->s_as->a_proc->p_zone) == 0) {
5800 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5801 return (IE_NOMEM);
5802 }
5803
5804 /*
5805 * Make a note of how much swap space
5806 * we've reserved.
5807 */
5808 if (svd->pageswap == 0 && sz == seg->s_size) {
5809 svd->swresv = sz;
5810 } else {
5811 ASSERT(svd->vpage != NULL);
5812 svd->swresv += sz;
5813 svd->pageswap = 1;
5814 for (cvp = svp; cvp < evp; cvp++) {
5815 if (!VPP_ISSWAPRES(cvp))
5816 VPP_SETSWAPRES(cvp);
5817 }
5818 }
5819 }
5820 } else {
5821 /*
5822 * Swap space is released only if this segment
5823 * does not map anonymous memory, since read faults
5824 * on such segments still need an anon slot to read
5825 * in the data.
5826 */
5827 if (svd->swresv != 0 && svd->vp != NULL &&
5828 svd->amp == NULL && addr == seg->s_base &&
5829 len == seg->s_size && svd->pageprot == 0) {
5830 ASSERT(svd->pageswap == 0);
5831 anon_unresv_zone(svd->swresv,
5832 seg->s_as->a_proc->p_zone);
5833 svd->swresv = 0;
5834 TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5835 "anon proc:%p %lu %u", seg, 0, 0);
5836 }
5837 }
5838 }
5839
5840 if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5841 if (svd->prot == prot) {
5842 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5843 return (0); /* all done */
5844 }
5845 svd->prot = (uchar_t)prot;
5846 } else if (svd->type == MAP_PRIVATE) {
5847 struct anon *ap = NULL;
5848 page_t *pp;
5849 u_offset_t offset, off;
5850 struct anon_map *amp;
5851 ulong_t anon_idx = 0;
5852
5853 /*
5854 * A vpage structure exists or else the change does not
5855 * involve the entire segment. Establish a vpage structure
5856 * if none is there. Then, for each page in the range,
5857 * adjust its individual permissions. Note that write-
5858 * enabling a MAP_PRIVATE page can affect the claims for
5859 * locked down memory. Overcommitting memory terminates
5860 * the operation.
5861 */
5862 segvn_vpage(seg);
5863 svd->pageprot = 1;
5864 if ((amp = svd->amp) != NULL) {
5865 anon_idx = svd->anon_index + seg_page(seg, addr);
5866 ASSERT(seg->s_szc == 0 ||
5867 IS_P2ALIGNED(anon_idx, pgcnt));
5868 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5869 }
5870
5871 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5872 evp = &svd->vpage[seg_page(seg, addr + len)];
5873
5874 /*
5875 * See Statement at the beginning of segvn_lockop regarding
5876 * the way cowcnts and lckcnts are handled.
5877 */
5878 for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5879
5880 if (seg->s_szc != 0) {
5881 if (amp != NULL) {
5882 anon_array_enter(amp, anon_idx,
5883 &cookie);
5884 }
5885 if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5886 !segvn_claim_pages(seg, svp, offset,
5887 anon_idx, prot)) {
5888 if (amp != NULL) {
5889 anon_array_exit(&cookie);
5890 }
5891 break;
5892 }
5893 if (amp != NULL) {
5894 anon_array_exit(&cookie);
5895 }
5896 anon_idx++;
5897 } else {
5898 if (amp != NULL) {
5899 anon_array_enter(amp, anon_idx,
5900 &cookie);
5901 ap = anon_get_ptr(amp->ahp, anon_idx++);
5902 }
5903
5904 if (VPP_ISPPLOCK(svp) &&
5905 VPP_PROT(svp) != prot) {
5906
5907 if (amp == NULL || ap == NULL) {
5908 vp = svd->vp;
5909 off = offset;
5910 } else
5911 swap_xlate(ap, &vp, &off);
5912 if (amp != NULL)
5913 anon_array_exit(&cookie);
5914
5915 if ((pp = page_lookup(vp, off,
5916 SE_SHARED)) == NULL) {
5917 panic("segvn_setprot: no page");
5918 /*NOTREACHED*/
5919 }
5920 ASSERT(seg->s_szc == 0);
5921 if ((VPP_PROT(svp) ^ prot) &
5922 PROT_WRITE) {
5923 if (prot & PROT_WRITE) {
5924 if (!page_addclaim(
5925 pp)) {
5926 page_unlock(pp);
5927 break;
5928 }
5929 } else {
5930 if (!page_subclaim(
5931 pp)) {
5932 page_unlock(pp);
5933 break;
5934 }
5935 }
5936 }
5937 page_unlock(pp);
5938 } else if (amp != NULL)
5939 anon_array_exit(&cookie);
5940 }
5941 VPP_SETPROT(svp, prot);
5942 offset += PAGESIZE;
5943 }
5944 if (amp != NULL)
5945 ANON_LOCK_EXIT(&->a_rwlock);
5946
5947 /*
5948 * Did we terminate prematurely? If so, simply unload
5949 * the translations to the things we've updated so far.
5950 */
5951 if (svp != evp) {
5952 if (unload_done) {
5953 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5954 return (IE_NOMEM);
5955 }
5956 len = (svp - &svd->vpage[seg_page(seg, addr)]) *
5957 PAGESIZE;
5958 ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
5959 if (len != 0)
5960 hat_unload(seg->s_as->a_hat, addr,
5961 len, HAT_UNLOAD);
5962 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5963 return (IE_NOMEM);
5964 }
5965 } else {
5966 segvn_vpage(seg);
5967 svd->pageprot = 1;
5968 evp = &svd->vpage[seg_page(seg, addr + len)];
5969 for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5970 VPP_SETPROT(svp, prot);
5971 }
5972 }
5973
5974 if (unload_done) {
5975 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5976 return (0);
5977 }
5978
5979 if (((prot & PROT_WRITE) != 0 &&
5980 (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
5981 (prot & ~PROT_USER) == PROT_NONE) {
5982 /*
5983 * Either private or shared data with write access (in
5984 * which case we need to throw out all former translations
5985 * so that we get the right translations set up on fault
5986 * and we don't allow write access to any copy-on-write pages
5987 * that might be around or to prevent write access to pages
5988 * representing holes in a file), or we don't have permission
5989 * to access the memory at all (in which case we have to
5990 * unload any current translations that might exist).
5991 */
5992 hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
5993 } else {
5994 /*
5995 * A shared mapping or a private mapping in which write
5996 * protection is going to be denied - just change all the
5997 * protections over the range of addresses in question.
5998 * segvn does not support any other attributes other
5999 * than prot so we can use hat_chgattr.
6000 */
6001 hat_chgattr(seg->s_as->a_hat, addr, len, prot);
6002 }
6003
6004 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6005
6006 return (0);
6007 }
6008
6009 /*
6010 * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
6011 * to determine if the seg is capable of mapping the requested szc.
6012 */
6013 static int
6014 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
6015 {
6016 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6017 struct segvn_data *nsvd;
6018 struct anon_map *amp = svd->amp;
6019 struct seg *nseg;
6020 caddr_t eaddr = addr + len, a;
6021 size_t pgsz = page_get_pagesize(szc);
6022 pgcnt_t pgcnt = page_get_pagecnt(szc);
6023 int err;
6024 u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
6025
6026 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6027 ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6028
6029 if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6030 return (0);
6031 }
6032
6033 /*
6034 * addr should always be pgsz aligned but eaddr may be misaligned if
6035 * it's at the end of the segment.
6036 *
6037 * XXX we should assert this condition since as_setpagesize() logic
6038 * guarantees it.
6039 */
6040 if (!IS_P2ALIGNED(addr, pgsz) ||
6041 (!IS_P2ALIGNED(eaddr, pgsz) &&
6042 eaddr != seg->s_base + seg->s_size)) {
6043
6044 segvn_setpgsz_align_err++;
6045 return (EINVAL);
6046 }
6047
6048 if (amp != NULL && svd->type == MAP_SHARED) {
6049 ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6050 if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6051
6052 segvn_setpgsz_anon_align_err++;
6053 return (EINVAL);
6054 }
6055 }
6056
6057 if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6058 szc > segvn_maxpgszc) {
6059 return (EINVAL);
6060 }
6061
6062 /* paranoid check */
6063 if (svd->vp != NULL &&
6064 (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6065 return (EINVAL);
6066 }
6067
6068 if (seg->s_szc == 0 && svd->vp != NULL &&
6069 map_addr_vacalign_check(addr, off)) {
6070 return (EINVAL);
6071 }
6072
6073 /*
6074 * Check that protections are the same within new page
6075 * size boundaries.
6076 */
6077 if (svd->pageprot) {
6078 for (a = addr; a < eaddr; a += pgsz) {
6079 if ((a + pgsz) > eaddr) {
6080 if (!sameprot(seg, a, eaddr - a)) {
6081 return (EINVAL);
6082 }
6083 } else {
6084 if (!sameprot(seg, a, pgsz)) {
6085 return (EINVAL);
6086 }
6087 }
6088 }
6089 }
6090
6091 /*
6092 * Since we are changing page size we first have to flush
6093 * the cache. This makes sure all the pagelock calls have
6094 * to recheck protections.
6095 */
6096 if (svd->softlockcnt > 0) {
6097 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6098
6099 /*
6100 * If this is shared segment non 0 softlockcnt
6101 * means locked pages are still in use.
6102 */
6103 if (svd->type == MAP_SHARED) {
6104 return (EAGAIN);
6105 }
6106
6107 /*
6108 * Since we do have the segvn writers lock nobody can fill
6109 * the cache with entries belonging to this seg during
6110 * the purge. The flush either succeeds or we still have
6111 * pending I/Os.
6112 */
6113 segvn_purge(seg);
6114 if (svd->softlockcnt > 0) {
6115 return (EAGAIN);
6116 }
6117 }
6118
6119 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6120 ASSERT(svd->amp == NULL);
6121 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6122 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6123 HAT_REGION_TEXT);
6124 svd->rcookie = HAT_INVALID_REGION_COOKIE;
6125 } else if (svd->tr_state == SEGVN_TR_INIT) {
6126 svd->tr_state = SEGVN_TR_OFF;
6127 } else if (svd->tr_state == SEGVN_TR_ON) {
6128 ASSERT(svd->amp != NULL);
6129 segvn_textunrepl(seg, 1);
6130 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6131 amp = NULL;
6132 }
6133
6134 /*
6135 * Operation for sub range of existing segment.
6136 */
6137 if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6138 if (szc < seg->s_szc) {
6139 VM_STAT_ADD(segvnvmstats.demoterange[2]);
6140 err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6141 if (err == 0) {
6142 return (IE_RETRY);
6143 }
6144 if (err == ENOMEM) {
6145 return (IE_NOMEM);
6146 }
6147 return (err);
6148 }
6149 if (addr != seg->s_base) {
6150 nseg = segvn_split_seg(seg, addr);
6151 if (eaddr != (nseg->s_base + nseg->s_size)) {
6152 /* eaddr is szc aligned */
6153 (void) segvn_split_seg(nseg, eaddr);
6154 }
6155 return (IE_RETRY);
6156 }
6157 if (eaddr != (seg->s_base + seg->s_size)) {
6158 /* eaddr is szc aligned */
6159 (void) segvn_split_seg(seg, eaddr);
6160 }
6161 return (IE_RETRY);
6162 }
6163
6164 /*
6165 * Break any low level sharing and reset seg->s_szc to 0.
6166 */
6167 if ((err = segvn_clrszc(seg)) != 0) {
6168 if (err == ENOMEM) {
6169 err = IE_NOMEM;
6170 }
6171 return (err);
6172 }
6173 ASSERT(seg->s_szc == 0);
6174
6175 /*
6176 * If the end of the current segment is not pgsz aligned
6177 * then attempt to concatenate with the next segment.
6178 */
6179 if (!IS_P2ALIGNED(eaddr, pgsz)) {
6180 nseg = AS_SEGNEXT(seg->s_as, seg);
6181 if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6182 return (ENOMEM);
6183 }
6184 if (nseg->s_ops != &segvn_ops) {
6185 return (EINVAL);
6186 }
6187 nsvd = (struct segvn_data *)nseg->s_data;
6188 if (nsvd->softlockcnt > 0) {
6189 /*
6190 * If this is shared segment non 0 softlockcnt
6191 * means locked pages are still in use.
6192 */
6193 if (nsvd->type == MAP_SHARED) {
6194 return (EAGAIN);
6195 }
6196 segvn_purge(nseg);
6197 if (nsvd->softlockcnt > 0) {
6198 return (EAGAIN);
6199 }
6200 }
6201 err = segvn_clrszc(nseg);
6202 if (err == ENOMEM) {
6203 err = IE_NOMEM;
6204 }
6205 if (err != 0) {
6206 return (err);
6207 }
6208 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6209 err = segvn_concat(seg, nseg, 1);
6210 if (err == -1) {
6211 return (EINVAL);
6212 }
6213 if (err == -2) {
6214 return (IE_NOMEM);
6215 }
6216 return (IE_RETRY);
6217 }
6218
6219 /*
6220 * May need to re-align anon array to
6221 * new szc.
6222 */
6223 if (amp != NULL) {
6224 if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6225 struct anon_hdr *nahp;
6226
6227 ASSERT(svd->type == MAP_PRIVATE);
6228
6229 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
6230 ASSERT(amp->refcnt == 1);
6231 nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6232 if (nahp == NULL) {
6233 ANON_LOCK_EXIT(&->a_rwlock);
6234 return (IE_NOMEM);
6235 }
6236 if (anon_copy_ptr(amp->ahp, svd->anon_index,
6237 nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6238 anon_release(nahp, btop(amp->size));
6239 ANON_LOCK_EXIT(&->a_rwlock);
6240 return (IE_NOMEM);
6241 }
6242 anon_release(amp->ahp, btop(amp->size));
6243 amp->ahp = nahp;
6244 svd->anon_index = 0;
6245 ANON_LOCK_EXIT(&->a_rwlock);
6246 }
6247 }
6248 if (svd->vp != NULL && szc != 0) {
6249 struct vattr va;
6250 u_offset_t eoffpage = svd->offset;
6251 va.va_mask = AT_SIZE;
6252 eoffpage += seg->s_size;
6253 eoffpage = btopr(eoffpage);
6254 if (VOP_GETATTR(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6255 segvn_setpgsz_getattr_err++;
6256 return (EINVAL);
6257 }
6258 if (btopr(va.va_size) < eoffpage) {
6259 segvn_setpgsz_eof_err++;
6260 return (EINVAL);
6261 }
6262 if (amp != NULL) {
6263 /*
6264 * anon_fill_cow_holes() may call VOP_GETPAGE().
6265 * don't take anon map lock here to avoid holding it
6266 * across VOP_GETPAGE() calls that may call back into
6267 * segvn for klsutering checks. We don't really need
6268 * anon map lock here since it's a private segment and
6269 * we hold as level lock as writers.
6270 */
6271 if ((err = anon_fill_cow_holes(seg, seg->s_base,
6272 amp->ahp, svd->anon_index, svd->vp, svd->offset,
6273 seg->s_size, szc, svd->prot, svd->vpage,
6274 svd->cred)) != 0) {
6275 return (EINVAL);
6276 }
6277 }
6278 segvn_setvnode_mpss(svd->vp);
6279 }
6280
6281 if (amp != NULL) {
6282 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
6283 if (svd->type == MAP_PRIVATE) {
6284 amp->a_szc = szc;
6285 } else if (szc > amp->a_szc) {
6286 amp->a_szc = szc;
6287 }
6288 ANON_LOCK_EXIT(&->a_rwlock);
6289 }
6290
6291 seg->s_szc = szc;
6292
6293 return (0);
6294 }
6295
6296 static int
6297 segvn_clrszc(struct seg *seg)
6298 {
6299 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6300 struct anon_map *amp = svd->amp;
6301 size_t pgsz;
6302 pgcnt_t pages;
6303 int err = 0;
6304 caddr_t a = seg->s_base;
6305 caddr_t ea = a + seg->s_size;
6306 ulong_t an_idx = svd->anon_index;
6307 vnode_t *vp = svd->vp;
6308 struct vpage *vpage = svd->vpage;
6309 page_t *anon_pl[1 + 1], *pp;
6310 struct anon *ap, *oldap;
6311 uint_t prot = svd->prot, vpprot;
6312 int pageflag = 0;
6313
6314 ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6315 SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6316 ASSERT(svd->softlockcnt == 0);
6317
6318 if (vp == NULL && amp == NULL) {
6319 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6320 seg->s_szc = 0;
6321 return (0);
6322 }
6323
6324 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6325 ASSERT(svd->amp == NULL);
6326 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6327 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6328 HAT_REGION_TEXT);
6329 svd->rcookie = HAT_INVALID_REGION_COOKIE;
6330 } else if (svd->tr_state == SEGVN_TR_ON) {
6331 ASSERT(svd->amp != NULL);
6332 segvn_textunrepl(seg, 1);
6333 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6334 amp = NULL;
6335 } else {
6336 if (svd->tr_state != SEGVN_TR_OFF) {
6337 ASSERT(svd->tr_state == SEGVN_TR_INIT);
6338 svd->tr_state = SEGVN_TR_OFF;
6339 }
6340
6341 /*
6342 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6343 * unload argument is 0 when we are freeing the segment
6344 * and unload was already done.
6345 */
6346 hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6347 HAT_UNLOAD_UNMAP);
6348 }
6349
6350 if (amp == NULL || svd->type == MAP_SHARED) {
6351 seg->s_szc = 0;
6352 return (0);
6353 }
6354
6355 pgsz = page_get_pagesize(seg->s_szc);
6356 pages = btop(pgsz);
6357
6358 /*
6359 * XXX anon rwlock is not really needed because this is a
6360 * private segment and we are writers.
6361 */
6362 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
6363
6364 for (; a < ea; a += pgsz, an_idx += pages) {
6365 if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6366 ASSERT(vpage != NULL || svd->pageprot == 0);
6367 if (vpage != NULL) {
6368 ASSERT(sameprot(seg, a, pgsz));
6369 prot = VPP_PROT(vpage);
6370 pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6371 }
6372 if (seg->s_szc != 0) {
6373 ASSERT(vp == NULL || anon_pages(amp->ahp,
6374 an_idx, pages) == pages);
6375 if ((err = anon_map_demotepages(amp, an_idx,
6376 seg, a, prot, vpage, svd->cred)) != 0) {
6377 goto out;
6378 }
6379 } else {
6380 if (oldap->an_refcnt == 1) {
6381 continue;
6382 }
6383 if ((err = anon_getpage(&oldap, &vpprot,
6384 anon_pl, PAGESIZE, seg, a, S_READ,
6385 svd->cred))) {
6386 goto out;
6387 }
6388 if ((pp = anon_private(&ap, seg, a, prot,
6389 anon_pl[0], pageflag, svd->cred)) == NULL) {
6390 err = ENOMEM;
6391 goto out;
6392 }
6393 anon_decref(oldap);
6394 (void) anon_set_ptr(amp->ahp, an_idx, ap,
6395 ANON_SLEEP);
6396 page_unlock(pp);
6397 }
6398 }
6399 vpage = (vpage == NULL) ? NULL : vpage + pages;
6400 }
6401
6402 amp->a_szc = 0;
6403 seg->s_szc = 0;
6404 out:
6405 ANON_LOCK_EXIT(&->a_rwlock);
6406 return (err);
6407 }
6408
6409 static int
6410 segvn_claim_pages(
6411 struct seg *seg,
6412 struct vpage *svp,
6413 u_offset_t off,
6414 ulong_t anon_idx,
6415 uint_t prot)
6416 {
6417 pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6418 size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6419 page_t **ppa;
6420 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6421 struct anon_map *amp = svd->amp;
6422 struct vpage *evp = svp + pgcnt;
6423 caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6424 + seg->s_base;
6425 struct anon *ap;
6426 struct vnode *vp = svd->vp;
6427 page_t *pp;
6428 pgcnt_t pg_idx, i;
6429 int err = 0;
6430 anoff_t aoff;
6431 int anon = (amp != NULL) ? 1 : 0;
6432
6433 ASSERT(svd->type == MAP_PRIVATE);
6434 ASSERT(svd->vpage != NULL);
6435 ASSERT(seg->s_szc != 0);
6436 ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6437 ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6438 ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6439
6440 if (VPP_PROT(svp) == prot)
6441 return (1);
6442 if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6443 return (1);
6444
6445 ppa = kmem_alloc(ppasize, KM_SLEEP);
6446 if (anon && vp != NULL) {
6447 if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6448 anon = 0;
6449 ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6450 }
6451 ASSERT(!anon ||
6452 anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6453 }
6454
6455 for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6456 if (!VPP_ISPPLOCK(svp))
6457 continue;
6458 if (anon) {
6459 ap = anon_get_ptr(amp->ahp, anon_idx);
6460 if (ap == NULL) {
6461 panic("segvn_claim_pages: no anon slot");
6462 }
6463 swap_xlate(ap, &vp, &aoff);
6464 off = (u_offset_t)aoff;
6465 }
6466 ASSERT(vp != NULL);
6467 if ((pp = page_lookup(vp,
6468 (u_offset_t)off, SE_SHARED)) == NULL) {
6469 panic("segvn_claim_pages: no page");
6470 }
6471 ppa[pg_idx++] = pp;
6472 off += PAGESIZE;
6473 }
6474
6475 if (ppa[0] == NULL) {
6476 kmem_free(ppa, ppasize);
6477 return (1);
6478 }
6479
6480 ASSERT(pg_idx <= pgcnt);
6481 ppa[pg_idx] = NULL;
6482
6483
6484 /* Find each large page within ppa, and adjust its claim */
6485
6486 /* Does ppa cover a single large page? */
6487 if (ppa[0]->p_szc == seg->s_szc) {
6488 if (prot & PROT_WRITE)
6489 err = page_addclaim_pages(ppa);
6490 else
6491 err = page_subclaim_pages(ppa);
6492 } else {
6493 for (i = 0; ppa[i]; i += pgcnt) {
6494 ASSERT(IS_P2ALIGNED(page_pptonum(ppa[i]), pgcnt));
6495 if (prot & PROT_WRITE)
6496 err = page_addclaim_pages(&ppa[i]);
6497 else
6498 err = page_subclaim_pages(&ppa[i]);
6499 if (err == 0)
6500 break;
6501 }
6502 }
6503
6504 for (i = 0; i < pg_idx; i++) {
6505 ASSERT(ppa[i] != NULL);
6506 page_unlock(ppa[i]);
6507 }
6508
6509 kmem_free(ppa, ppasize);
6510 return (err);
6511 }
6512
6513 /*
6514 * Returns right (upper address) segment if split occurred.
6515 * If the address is equal to the beginning or end of its segment it returns
6516 * the current segment.
6517 */
6518 static struct seg *
6519 segvn_split_seg(struct seg *seg, caddr_t addr)
6520 {
6521 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6522 struct seg *nseg;
6523 size_t nsize;
6524 struct segvn_data *nsvd;
6525
6526 ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6527 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6528
6529 ASSERT(addr >= seg->s_base);
6530 ASSERT(addr <= seg->s_base + seg->s_size);
6531 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6532
6533 if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6534 return (seg);
6535
6536 nsize = seg->s_base + seg->s_size - addr;
6537 seg->s_size = addr - seg->s_base;
6538 nseg = seg_alloc(seg->s_as, addr, nsize);
6539 ASSERT(nseg != NULL);
6540 nseg->s_ops = seg->s_ops;
6541 nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6542 nseg->s_data = (void *)nsvd;
6543 nseg->s_szc = seg->s_szc;
6544 *nsvd = *svd;
6545 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6546 nsvd->seg = nseg;
6547 rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6548
6549 if (nsvd->vp != NULL) {
6550 VN_HOLD(nsvd->vp);
6551 nsvd->offset = svd->offset +
6552 (uintptr_t)(nseg->s_base - seg->s_base);
6553 if (nsvd->type == MAP_SHARED)
6554 lgrp_shm_policy_init(NULL, nsvd->vp);
6555 } else {
6556 /*
6557 * The offset for an anonymous segment has no signifigance in
6558 * terms of an offset into a file. If we were to use the above
6559 * calculation instead, the structures read out of
6560 * /proc/<pid>/xmap would be more difficult to decipher since
6561 * it would be unclear whether two seemingly contiguous
6562 * prxmap_t structures represented different segments or a
6563 * single segment that had been split up into multiple prxmap_t
6564 * structures (e.g. if some part of the segment had not yet
6565 * been faulted in).
6566 */
6567 nsvd->offset = 0;
6568 }
6569
6570 ASSERT(svd->softlockcnt == 0);
6571 ASSERT(svd->softlockcnt_sbase == 0);
6572 ASSERT(svd->softlockcnt_send == 0);
6573 crhold(svd->cred);
6574
6575 if (svd->vpage != NULL) {
6576 size_t bytes = vpgtob(seg_pages(seg));
6577 size_t nbytes = vpgtob(seg_pages(nseg));
6578 struct vpage *ovpage = svd->vpage;
6579
6580 svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6581 bcopy(ovpage, svd->vpage, bytes);
6582 nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6583 bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6584 kmem_free(ovpage, bytes + nbytes);
6585 }
6586 if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6587 struct anon_map *oamp = svd->amp, *namp;
6588 struct anon_hdr *nahp;
6589
6590 ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6591 ASSERT(oamp->refcnt == 1);
6592 nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6593 (void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6594 nahp, 0, btop(seg->s_size), ANON_SLEEP);
6595
6596 namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6597 namp->a_szc = nseg->s_szc;
6598 (void) anon_copy_ptr(oamp->ahp,
6599 svd->anon_index + btop(seg->s_size),
6600 namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6601 anon_release(oamp->ahp, btop(oamp->size));
6602 oamp->ahp = nahp;
6603 oamp->size = seg->s_size;
6604 svd->anon_index = 0;
6605 nsvd->amp = namp;
6606 nsvd->anon_index = 0;
6607 ANON_LOCK_EXIT(&oamp->a_rwlock);
6608 } else if (svd->amp != NULL) {
6609 pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6610 ASSERT(svd->amp == nsvd->amp);
6611 ASSERT(seg->s_szc <= svd->amp->a_szc);
6612 nsvd->anon_index = svd->anon_index + seg_pages(seg);
6613 ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6614 ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6615 svd->amp->refcnt++;
6616 ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6617 }
6618
6619 /*
6620 * Split the amount of swap reserved.
6621 */
6622 if (svd->swresv) {
6623 /*
6624 * For MAP_NORESERVE, only allocate swap reserve for pages
6625 * being used. Other segments get enough to cover whole
6626 * segment.
6627 */
6628 if (svd->flags & MAP_NORESERVE) {
6629 size_t oswresv;
6630
6631 ASSERT(svd->amp);
6632 oswresv = svd->swresv;
6633 svd->swresv = ptob(anon_pages(svd->amp->ahp,
6634 svd->anon_index, btop(seg->s_size)));
6635 nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6636 nsvd->anon_index, btop(nseg->s_size)));
6637 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6638 } else {
6639 if (svd->pageswap) {
6640 svd->swresv = segvn_count_swap_by_vpages(seg);
6641 ASSERT(nsvd->swresv >= svd->swresv);
6642 nsvd->swresv -= svd->swresv;
6643 } else {
6644 ASSERT(svd->swresv == seg->s_size +
6645 nseg->s_size);
6646 svd->swresv = seg->s_size;
6647 nsvd->swresv = nseg->s_size;
6648 }
6649 }
6650 }
6651
6652 return (nseg);
6653 }
6654
6655 /*
6656 * called on memory operations (unmap, setprot, setpagesize) for a subset
6657 * of a large page segment to either demote the memory range (SDR_RANGE)
6658 * or the ends (SDR_END) by addr/len.
6659 *
6660 * returns 0 on success. returns errno, including ENOMEM, on failure.
6661 */
6662 static int
6663 segvn_demote_range(
6664 struct seg *seg,
6665 caddr_t addr,
6666 size_t len,
6667 int flag,
6668 uint_t szcvec)
6669 {
6670 caddr_t eaddr = addr + len;
6671 caddr_t lpgaddr, lpgeaddr;
6672 struct seg *nseg;
6673 struct seg *badseg1 = NULL;
6674 struct seg *badseg2 = NULL;
6675 size_t pgsz;
6676 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6677 int err;
6678 uint_t szc = seg->s_szc;
6679 uint_t tszcvec;
6680
6681 ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6682 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6683 ASSERT(szc != 0);
6684 pgsz = page_get_pagesize(szc);
6685 ASSERT(seg->s_base != addr || seg->s_size != len);
6686 ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6687 ASSERT(svd->softlockcnt == 0);
6688 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6689 ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6690
6691 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6692 ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6693 if (flag == SDR_RANGE) {
6694 /* demote entire range */
6695 badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6696 (void) segvn_split_seg(nseg, lpgeaddr);
6697 ASSERT(badseg1->s_base == lpgaddr);
6698 ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6699 } else if (addr != lpgaddr) {
6700 ASSERT(flag == SDR_END);
6701 badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6702 if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6703 eaddr < lpgaddr + 2 * pgsz) {
6704 (void) segvn_split_seg(nseg, lpgeaddr);
6705 ASSERT(badseg1->s_base == lpgaddr);
6706 ASSERT(badseg1->s_size == 2 * pgsz);
6707 } else {
6708 nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6709 ASSERT(badseg1->s_base == lpgaddr);
6710 ASSERT(badseg1->s_size == pgsz);
6711 if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6712 ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6713 nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6714 badseg2 = nseg;
6715 (void) segvn_split_seg(nseg, lpgeaddr);
6716 ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6717 ASSERT(badseg2->s_size == pgsz);
6718 }
6719 }
6720 } else {
6721 ASSERT(flag == SDR_END);
6722 ASSERT(eaddr < lpgeaddr);
6723 badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6724 (void) segvn_split_seg(nseg, lpgeaddr);
6725 ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6726 ASSERT(badseg1->s_size == pgsz);
6727 }
6728
6729 ASSERT(badseg1 != NULL);
6730 ASSERT(badseg1->s_szc == szc);
6731 ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6732 badseg1->s_size == 2 * pgsz);
6733 ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6734 ASSERT(badseg1->s_size == pgsz ||
6735 sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6736 if (err = segvn_clrszc(badseg1)) {
6737 return (err);
6738 }
6739 ASSERT(badseg1->s_szc == 0);
6740
6741 if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6742 uint_t tszc = highbit(tszcvec) - 1;
6743 caddr_t ta = MAX(addr, badseg1->s_base);
6744 caddr_t te;
6745 size_t tpgsz = page_get_pagesize(tszc);
6746
6747 ASSERT(svd->type == MAP_SHARED);
6748 ASSERT(flag == SDR_END);
6749 ASSERT(tszc < szc && tszc > 0);
6750
6751 if (eaddr > badseg1->s_base + badseg1->s_size) {
6752 te = badseg1->s_base + badseg1->s_size;
6753 } else {
6754 te = eaddr;
6755 }
6756
6757 ASSERT(ta <= te);
6758 badseg1->s_szc = tszc;
6759 if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6760 if (badseg2 != NULL) {
6761 err = segvn_demote_range(badseg1, ta, te - ta,
6762 SDR_END, tszcvec);
6763 if (err != 0) {
6764 return (err);
6765 }
6766 } else {
6767 return (segvn_demote_range(badseg1, ta,
6768 te - ta, SDR_END, tszcvec));
6769 }
6770 }
6771 }
6772
6773 if (badseg2 == NULL)
6774 return (0);
6775 ASSERT(badseg2->s_szc == szc);
6776 ASSERT(badseg2->s_size == pgsz);
6777 ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6778 if (err = segvn_clrszc(badseg2)) {
6779 return (err);
6780 }
6781 ASSERT(badseg2->s_szc == 0);
6782
6783 if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6784 uint_t tszc = highbit(tszcvec) - 1;
6785 size_t tpgsz = page_get_pagesize(tszc);
6786
6787 ASSERT(svd->type == MAP_SHARED);
6788 ASSERT(flag == SDR_END);
6789 ASSERT(tszc < szc && tszc > 0);
6790 ASSERT(badseg2->s_base > addr);
6791 ASSERT(eaddr > badseg2->s_base);
6792 ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6793
6794 badseg2->s_szc = tszc;
6795 if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6796 return (segvn_demote_range(badseg2, badseg2->s_base,
6797 eaddr - badseg2->s_base, SDR_END, tszcvec));
6798 }
6799 }
6800
6801 return (0);
6802 }
6803
6804 static int
6805 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6806 {
6807 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6808 struct vpage *vp, *evp;
6809
6810 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6811
6812 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6813 /*
6814 * If segment protection can be used, simply check against them.
6815 */
6816 if (svd->pageprot == 0) {
6817 int err;
6818
6819 err = ((svd->prot & prot) != prot) ? EACCES : 0;
6820 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6821 return (err);
6822 }
6823
6824 /*
6825 * Have to check down to the vpage level.
6826 */
6827 evp = &svd->vpage[seg_page(seg, addr + len)];
6828 for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6829 if ((VPP_PROT(vp) & prot) != prot) {
6830 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6831 return (EACCES);
6832 }
6833 }
6834 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6835 return (0);
6836 }
6837
6838 static int
6839 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6840 {
6841 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6842 size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6843
6844 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6845
6846 if (pgno != 0) {
6847 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6848 if (svd->pageprot == 0) {
6849 do {
6850 protv[--pgno] = svd->prot;
6851 } while (pgno != 0);
6852 } else {
6853 size_t pgoff = seg_page(seg, addr);
6854
6855 do {
6856 pgno--;
6857 protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6858 } while (pgno != 0);
6859 }
6860 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6861 }
6862 return (0);
6863 }
6864
6865 static u_offset_t
6866 segvn_getoffset(struct seg *seg, caddr_t addr)
6867 {
6868 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6869
6870 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6871
6872 return (svd->offset + (uintptr_t)(addr - seg->s_base));
6873 }
6874
6875 /*ARGSUSED*/
6876 static int
6877 segvn_gettype(struct seg *seg, caddr_t addr)
6878 {
6879 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6880
6881 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6882
6883 return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6884 MAP_INITDATA)));
6885 }
6886
6887 /*ARGSUSED*/
6888 static int
6889 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6890 {
6891 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6892
6893 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6894
6895 *vpp = svd->vp;
6896 return (0);
6897 }
6898
6899 /*
6900 * Check to see if it makes sense to do kluster/read ahead to
6901 * addr + delta relative to the mapping at addr. We assume here
6902 * that delta is a signed PAGESIZE'd multiple (which can be negative).
6903 *
6904 * For segvn, we currently "approve" of the action if we are
6905 * still in the segment and it maps from the same vp/off,
6906 * or if the advice stored in segvn_data or vpages allows it.
6907 * Currently, klustering is not allowed only if MADV_RANDOM is set.
6908 */
6909 static int
6910 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6911 {
6912 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6913 struct anon *oap, *ap;
6914 ssize_t pd;
6915 size_t page;
6916 struct vnode *vp1, *vp2;
6917 u_offset_t off1, off2;
6918 struct anon_map *amp;
6919
6920 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6921 ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6922 SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6923
6924 if (addr + delta < seg->s_base ||
6925 addr + delta >= (seg->s_base + seg->s_size))
6926 return (-1); /* exceeded segment bounds */
6927
6928 pd = delta / (ssize_t)PAGESIZE; /* divide to preserve sign bit */
6929 page = seg_page(seg, addr);
6930
6931 /*
6932 * Check to see if either of the pages addr or addr + delta
6933 * have advice set that prevents klustering (if MADV_RANDOM advice
6934 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6935 * is negative).
6936 */
6937 if (svd->advice == MADV_RANDOM ||
6938 svd->advice == MADV_SEQUENTIAL && delta < 0)
6939 return (-1);
6940 else if (svd->pageadvice && svd->vpage) {
6941 struct vpage *bvpp, *evpp;
6942
6943 bvpp = &svd->vpage[page];
6944 evpp = &svd->vpage[page + pd];
6945 if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
6946 VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
6947 return (-1);
6948 if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
6949 VPP_ADVICE(evpp) == MADV_RANDOM)
6950 return (-1);
6951 }
6952
6953 if (svd->type == MAP_SHARED)
6954 return (0); /* shared mapping - all ok */
6955
6956 if ((amp = svd->amp) == NULL)
6957 return (0); /* off original vnode */
6958
6959 page += svd->anon_index;
6960
6961 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
6962
6963 oap = anon_get_ptr(amp->ahp, page);
6964 ap = anon_get_ptr(amp->ahp, page + pd);
6965
6966 ANON_LOCK_EXIT(&->a_rwlock);
6967
6968 if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
6969 return (-1); /* one with and one without an anon */
6970 }
6971
6972 if (oap == NULL) { /* implies that ap == NULL */
6973 return (0); /* off original vnode */
6974 }
6975
6976 /*
6977 * Now we know we have two anon pointers - check to
6978 * see if they happen to be properly allocated.
6979 */
6980
6981 /*
6982 * XXX We cheat here and don't lock the anon slots. We can't because
6983 * we may have been called from the anon layer which might already
6984 * have locked them. We are holding a refcnt on the slots so they
6985 * can't disappear. The worst that will happen is we'll get the wrong
6986 * names (vp, off) for the slots and make a poor klustering decision.
6987 */
6988 swap_xlate(ap, &vp1, &off1);
6989 swap_xlate(oap, &vp2, &off2);
6990
6991
6992 if (!VOP_CMP(vp1, vp2, NULL) || off1 - off2 != delta)
6993 return (-1);
6994 return (0);
6995 }
6996
6997 /*
6998 * Synchronize primary storage cache with real object in virtual memory.
6999 *
7000 * XXX - Anonymous pages should not be sync'ed out at all.
7001 */
7002 static int
7003 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7004 {
7005 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7006 struct vpage *vpp;
7007 page_t *pp;
7008 u_offset_t offset;
7009 struct vnode *vp;
7010 u_offset_t off;
7011 caddr_t eaddr;
7012 int bflags;
7013 int err = 0;
7014 int segtype;
7015 int pageprot;
7016 int prot;
7017 ulong_t anon_index;
7018 struct anon_map *amp;
7019 struct anon *ap;
7020 anon_sync_obj_t cookie;
7021
7022 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7023
7024 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7025
7026 if (svd->softlockcnt > 0) {
7027 /*
7028 * If this is shared segment non 0 softlockcnt
7029 * means locked pages are still in use.
7030 */
7031 if (svd->type == MAP_SHARED) {
7032 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7033 return (EAGAIN);
7034 }
7035
7036 /*
7037 * flush all pages from seg cache
7038 * otherwise we may deadlock in swap_putpage
7039 * for B_INVAL page (4175402).
7040 *
7041 * Even if we grab segvn WRITER's lock
7042 * here, there might be another thread which could've
7043 * successfully performed lookup/insert just before
7044 * we acquired the lock here. So, grabbing either
7045 * lock here is of not much use. Until we devise
7046 * a strategy at upper layers to solve the
7047 * synchronization issues completely, we expect
7048 * applications to handle this appropriately.
7049 */
7050 segvn_purge(seg);
7051 if (svd->softlockcnt > 0) {
7052 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7053 return (EAGAIN);
7054 }
7055 } else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7056 svd->amp->a_softlockcnt > 0) {
7057 /*
7058 * Try to purge this amp's entries from pcache. It will
7059 * succeed only if other segments that share the amp have no
7060 * outstanding softlock's.
7061 */
7062 segvn_purge(seg);
7063 if (svd->amp->a_softlockcnt > 0 || svd->softlockcnt > 0) {
7064 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7065 return (EAGAIN);
7066 }
7067 }
7068
7069 vpp = svd->vpage;
7070 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7071 bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7072 ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7073
7074 if (attr) {
7075 pageprot = attr & ~(SHARED|PRIVATE);
7076 segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7077
7078 /*
7079 * We are done if the segment types don't match
7080 * or if we have segment level protections and
7081 * they don't match.
7082 */
7083 if (svd->type != segtype) {
7084 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7085 return (0);
7086 }
7087 if (vpp == NULL) {
7088 if (svd->prot != pageprot) {
7089 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7090 return (0);
7091 }
7092 prot = svd->prot;
7093 } else
7094 vpp = &svd->vpage[seg_page(seg, addr)];
7095
7096 } else if (svd->vp && svd->amp == NULL &&
7097 (flags & MS_INVALIDATE) == 0) {
7098
7099 /*
7100 * No attributes, no anonymous pages and MS_INVALIDATE flag
7101 * is not on, just use one big request.
7102 */
7103 err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7104 bflags, svd->cred, NULL);
7105 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7106 return (err);
7107 }
7108
7109 if ((amp = svd->amp) != NULL)
7110 anon_index = svd->anon_index + seg_page(seg, addr);
7111
7112 for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7113 ap = NULL;
7114 if (amp != NULL) {
7115 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7116 anon_array_enter(amp, anon_index, &cookie);
7117 ap = anon_get_ptr(amp->ahp, anon_index++);
7118 if (ap != NULL) {
7119 swap_xlate(ap, &vp, &off);
7120 } else {
7121 vp = svd->vp;
7122 off = offset;
7123 }
7124 anon_array_exit(&cookie);
7125 ANON_LOCK_EXIT(&->a_rwlock);
7126 } else {
7127 vp = svd->vp;
7128 off = offset;
7129 }
7130 offset += PAGESIZE;
7131
7132 if (vp == NULL) /* untouched zfod page */
7133 continue;
7134
7135 if (attr) {
7136 if (vpp) {
7137 prot = VPP_PROT(vpp);
7138 vpp++;
7139 }
7140 if (prot != pageprot) {
7141 continue;
7142 }
7143 }
7144
7145 /*
7146 * See if any of these pages are locked -- if so, then we
7147 * will have to truncate an invalidate request at the first
7148 * locked one. We don't need the page_struct_lock to test
7149 * as this is only advisory; even if we acquire it someone
7150 * might race in and lock the page after we unlock and before
7151 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7152 */
7153 if (flags & MS_INVALIDATE) {
7154 if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7155 if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7156 page_unlock(pp);
7157 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7158 return (EBUSY);
7159 }
7160 if (ap != NULL && pp->p_szc != 0 &&
7161 page_tryupgrade(pp)) {
7162 if (pp->p_lckcnt == 0 &&
7163 pp->p_cowcnt == 0) {
7164 /*
7165 * swapfs VN_DISPOSE() won't
7166 * invalidate large pages.
7167 * Attempt to demote.
7168 * XXX can't help it if it
7169 * fails. But for swapfs
7170 * pages it is no big deal.
7171 */
7172 (void) page_try_demote_pages(
7173 pp);
7174 }
7175 }
7176 page_unlock(pp);
7177 }
7178 } else if (svd->type == MAP_SHARED && amp != NULL) {
7179 /*
7180 * Avoid writing out to disk ISM's large pages
7181 * because segspt_free_pages() relies on NULL an_pvp
7182 * of anon slots of such pages.
7183 */
7184
7185 ASSERT(svd->vp == NULL);
7186 /*
7187 * swapfs uses page_lookup_nowait if not freeing or
7188 * invalidating and skips a page if
7189 * page_lookup_nowait returns NULL.
7190 */
7191 pp = page_lookup_nowait(vp, off, SE_SHARED);
7192 if (pp == NULL) {
7193 continue;
7194 }
7195 if (pp->p_szc != 0) {
7196 page_unlock(pp);
7197 continue;
7198 }
7199
7200 /*
7201 * Note ISM pages are created large so (vp, off)'s
7202 * page cannot suddenly become large after we unlock
7203 * pp.
7204 */
7205 page_unlock(pp);
7206 }
7207 /*
7208 * XXX - Should ultimately try to kluster
7209 * calls to VOP_PUTPAGE() for performance.
7210 */
7211 VN_HOLD(vp);
7212 err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7213 (bflags | (IS_SWAPFSVP(vp) ? B_PAGE_NOWAIT : 0)),
7214 svd->cred, NULL);
7215
7216 VN_RELE(vp);
7217 if (err)
7218 break;
7219 }
7220 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7221 return (err);
7222 }
7223
7224 /*
7225 * Determine if we have data corresponding to pages in the
7226 * primary storage virtual memory cache (i.e., "in core").
7227 */
7228 static size_t
7229 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7230 {
7231 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7232 struct vnode *vp, *avp;
7233 u_offset_t offset, aoffset;
7234 size_t p, ep;
7235 int ret;
7236 struct vpage *vpp;
7237 page_t *pp;
7238 uint_t start;
7239 struct anon_map *amp; /* XXX - for locknest */
7240 struct anon *ap;
7241 uint_t attr;
7242 anon_sync_obj_t cookie;
7243
7244 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7245
7246 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7247 if (svd->amp == NULL && svd->vp == NULL) {
7248 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7249 bzero(vec, btopr(len));
7250 return (len); /* no anonymous pages created yet */
7251 }
7252
7253 p = seg_page(seg, addr);
7254 ep = seg_page(seg, addr + len);
7255 start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7256
7257 amp = svd->amp;
7258 for (; p < ep; p++, addr += PAGESIZE) {
7259 vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7260 ret = start;
7261 ap = NULL;
7262 avp = NULL;
7263 /* Grab the vnode/offset for the anon slot */
7264 if (amp != NULL) {
7265 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7266 anon_array_enter(amp, svd->anon_index + p, &cookie);
7267 ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7268 if (ap != NULL) {
7269 swap_xlate(ap, &avp, &aoffset);
7270 }
7271 anon_array_exit(&cookie);
7272 ANON_LOCK_EXIT(&->a_rwlock);
7273 }
7274 if ((avp != NULL) && page_exists(avp, aoffset)) {
7275 /* A page exists for the anon slot */
7276 ret |= SEG_PAGE_INCORE;
7277
7278 /*
7279 * If page is mapped and writable
7280 */
7281 attr = (uint_t)0;
7282 if ((hat_getattr(seg->s_as->a_hat, addr,
7283 &attr) != -1) && (attr & PROT_WRITE)) {
7284 ret |= SEG_PAGE_ANON;
7285 }
7286 /*
7287 * Don't get page_struct lock for lckcnt and cowcnt,
7288 * since this is purely advisory.
7289 */
7290 if ((pp = page_lookup_nowait(avp, aoffset,
7291 SE_SHARED)) != NULL) {
7292 if (pp->p_lckcnt)
7293 ret |= SEG_PAGE_SOFTLOCK;
7294 if (pp->p_cowcnt)
7295 ret |= SEG_PAGE_HASCOW;
7296 page_unlock(pp);
7297 }
7298 }
7299
7300 /* Gather vnode statistics */
7301 vp = svd->vp;
7302 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7303
7304 if (vp != NULL) {
7305 /*
7306 * Try to obtain a "shared" lock on the page
7307 * without blocking. If this fails, determine
7308 * if the page is in memory.
7309 */
7310 pp = page_lookup_nowait(vp, offset, SE_SHARED);
7311 if ((pp == NULL) && (page_exists(vp, offset))) {
7312 /* Page is incore, and is named */
7313 ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7314 }
7315 /*
7316 * Don't get page_struct lock for lckcnt and cowcnt,
7317 * since this is purely advisory.
7318 */
7319 if (pp != NULL) {
7320 ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7321 if (pp->p_lckcnt)
7322 ret |= SEG_PAGE_SOFTLOCK;
7323 if (pp->p_cowcnt)
7324 ret |= SEG_PAGE_HASCOW;
7325 page_unlock(pp);
7326 }
7327 }
7328
7329 /* Gather virtual page information */
7330 if (vpp) {
7331 if (VPP_ISPPLOCK(vpp))
7332 ret |= SEG_PAGE_LOCKED;
7333 vpp++;
7334 }
7335
7336 *vec++ = (char)ret;
7337 }
7338 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7339 return (len);
7340 }
7341
7342 /*
7343 * Statement for p_cowcnts/p_lckcnts.
7344 *
7345 * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7346 * irrespective of the following factors or anything else:
7347 *
7348 * (1) anon slots are populated or not
7349 * (2) cow is broken or not
7350 * (3) refcnt on ap is 1 or greater than 1
7351 *
7352 * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7353 * and munlock.
7354 *
7355 *
7356 * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7357 *
7358 * if vpage has PROT_WRITE
7359 * transfer cowcnt on the oldpage -> cowcnt on the newpage
7360 * else
7361 * transfer lckcnt on the oldpage -> lckcnt on the newpage
7362 *
7363 * During copy-on-write, decrement p_cowcnt on the oldpage and increment
7364 * p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7365 *
7366 * We may also break COW if softlocking on read access in the physio case.
7367 * In this case, vpage may not have PROT_WRITE. So, we need to decrement
7368 * p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7369 * vpage doesn't have PROT_WRITE.
7370 *
7371 *
7372 * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7373 *
7374 * If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7375 * increment p_lckcnt by calling page_subclaim() which takes care of
7376 * availrmem accounting and p_lckcnt overflow.
7377 *
7378 * If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7379 * increment p_cowcnt by calling page_addclaim() which takes care of
7380 * availrmem availability and p_cowcnt overflow.
7381 */
7382
7383 /*
7384 * Lock down (or unlock) pages mapped by this segment.
7385 *
7386 * XXX only creates PAGESIZE pages if anon slots are not initialized.
7387 * At fault time they will be relocated into larger pages.
7388 */
7389 static int
7390 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7391 int attr, int op, ulong_t *lockmap, size_t pos)
7392 {
7393 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7394 struct vpage *vpp;
7395 struct vpage *evp;
7396 page_t *pp;
7397 u_offset_t offset;
7398 u_offset_t off;
7399 int segtype;
7400 int pageprot;
7401 int claim;
7402 struct vnode *vp;
7403 ulong_t anon_index;
7404 struct anon_map *amp;
7405 struct anon *ap;
7406 struct vattr va;
7407 anon_sync_obj_t cookie;
7408 struct kshmid *sp = NULL;
7409 struct proc *p = curproc;
7410 kproject_t *proj = NULL;
7411 int chargeproc = 1;
7412 size_t locked_bytes = 0;
7413 size_t unlocked_bytes = 0;
7414 int err = 0;
7415
7416 /*
7417 * Hold write lock on address space because may split or concatenate
7418 * segments
7419 */
7420 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7421
7422 /*
7423 * If this is a shm, use shm's project and zone, else use
7424 * project and zone of calling process
7425 */
7426
7427 /* Determine if this segment backs a sysV shm */
7428 if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7429 ASSERT(svd->type == MAP_SHARED);
7430 ASSERT(svd->tr_state == SEGVN_TR_OFF);
7431 sp = svd->amp->a_sp;
7432 proj = sp->shm_perm.ipc_proj;
7433 chargeproc = 0;
7434 }
7435
7436 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7437 if (attr) {
7438 pageprot = attr & ~(SHARED|PRIVATE);
7439 segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7440
7441 /*
7442 * We are done if the segment types don't match
7443 * or if we have segment level protections and
7444 * they don't match.
7445 */
7446 if (svd->type != segtype) {
7447 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7448 return (0);
7449 }
7450 if (svd->pageprot == 0 && svd->prot != pageprot) {
7451 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7452 return (0);
7453 }
7454 }
7455
7456 if (op == MC_LOCK) {
7457 if (svd->tr_state == SEGVN_TR_INIT) {
7458 svd->tr_state = SEGVN_TR_OFF;
7459 } else if (svd->tr_state == SEGVN_TR_ON) {
7460 ASSERT(svd->amp != NULL);
7461 segvn_textunrepl(seg, 0);
7462 ASSERT(svd->amp == NULL &&
7463 svd->tr_state == SEGVN_TR_OFF);
7464 }
7465 }
7466
7467 /*
7468 * If we're locking, then we must create a vpage structure if
7469 * none exists. If we're unlocking, then check to see if there
7470 * is a vpage -- if not, then we could not have locked anything.
7471 */
7472
7473 if ((vpp = svd->vpage) == NULL) {
7474 if (op == MC_LOCK)
7475 segvn_vpage(seg);
7476 else {
7477 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7478 return (0);
7479 }
7480 }
7481
7482 /*
7483 * The anonymous data vector (i.e., previously
7484 * unreferenced mapping to swap space) can be allocated
7485 * by lazily testing for its existence.
7486 */
7487 if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7488 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7489 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7490 svd->amp->a_szc = seg->s_szc;
7491 }
7492
7493 if ((amp = svd->amp) != NULL) {
7494 anon_index = svd->anon_index + seg_page(seg, addr);
7495 }
7496
7497 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7498 evp = &svd->vpage[seg_page(seg, addr + len)];
7499
7500 if (sp != NULL)
7501 mutex_enter(&sp->shm_mlock);
7502
7503 /* determine number of unlocked bytes in range for lock operation */
7504 if (op == MC_LOCK) {
7505
7506 if (sp == NULL) {
7507 for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7508 vpp++) {
7509 if (!VPP_ISPPLOCK(vpp))
7510 unlocked_bytes += PAGESIZE;
7511 }
7512 } else {
7513 ulong_t i_idx, i_edx;
7514 anon_sync_obj_t i_cookie;
7515 struct anon *i_ap;
7516 struct vnode *i_vp;
7517 u_offset_t i_off;
7518
7519 /* Only count sysV pages once for locked memory */
7520 i_edx = svd->anon_index + seg_page(seg, addr + len);
7521 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7522 for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7523 anon_array_enter(amp, i_idx, &i_cookie);
7524 i_ap = anon_get_ptr(amp->ahp, i_idx);
7525 if (i_ap == NULL) {
7526 unlocked_bytes += PAGESIZE;
7527 anon_array_exit(&i_cookie);
7528 continue;
7529 }
7530 swap_xlate(i_ap, &i_vp, &i_off);
7531 anon_array_exit(&i_cookie);
7532 pp = page_lookup(i_vp, i_off, SE_SHARED);
7533 if (pp == NULL) {
7534 unlocked_bytes += PAGESIZE;
7535 continue;
7536 } else if (pp->p_lckcnt == 0)
7537 unlocked_bytes += PAGESIZE;
7538 page_unlock(pp);
7539 }
7540 ANON_LOCK_EXIT(&->a_rwlock);
7541 }
7542
7543 mutex_enter(&p->p_lock);
7544 err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7545 chargeproc);
7546 mutex_exit(&p->p_lock);
7547
7548 if (err) {
7549 if (sp != NULL)
7550 mutex_exit(&sp->shm_mlock);
7551 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7552 return (err);
7553 }
7554 }
7555 /*
7556 * Loop over all pages in the range. Process if we're locking and
7557 * page has not already been locked in this mapping; or if we're
7558 * unlocking and the page has been locked.
7559 */
7560 for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7561 vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7562 if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7563 ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7564 (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7565
7566 if (amp != NULL)
7567 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7568 /*
7569 * If this isn't a MAP_NORESERVE segment and
7570 * we're locking, allocate anon slots if they
7571 * don't exist. The page is brought in later on.
7572 */
7573 if (op == MC_LOCK && svd->vp == NULL &&
7574 ((svd->flags & MAP_NORESERVE) == 0) &&
7575 amp != NULL &&
7576 ((ap = anon_get_ptr(amp->ahp, anon_index))
7577 == NULL)) {
7578 anon_array_enter(amp, anon_index, &cookie);
7579
7580 if ((ap = anon_get_ptr(amp->ahp,
7581 anon_index)) == NULL) {
7582 pp = anon_zero(seg, addr, &ap,
7583 svd->cred);
7584 if (pp == NULL) {
7585 anon_array_exit(&cookie);
7586 ANON_LOCK_EXIT(&->a_rwlock);
7587 err = ENOMEM;
7588 goto out;
7589 }
7590 ASSERT(anon_get_ptr(amp->ahp,
7591 anon_index) == NULL);
7592 (void) anon_set_ptr(amp->ahp,
7593 anon_index, ap, ANON_SLEEP);
7594 page_unlock(pp);
7595 }
7596 anon_array_exit(&cookie);
7597 }
7598
7599 /*
7600 * Get name for page, accounting for
7601 * existence of private copy.
7602 */
7603 ap = NULL;
7604 if (amp != NULL) {
7605 anon_array_enter(amp, anon_index, &cookie);
7606 ap = anon_get_ptr(amp->ahp, anon_index);
7607 if (ap != NULL) {
7608 swap_xlate(ap, &vp, &off);
7609 } else {
7610 if (svd->vp == NULL &&
7611 (svd->flags & MAP_NORESERVE)) {
7612 anon_array_exit(&cookie);
7613 ANON_LOCK_EXIT(&->a_rwlock);
7614 continue;
7615 }
7616 vp = svd->vp;
7617 off = offset;
7618 }
7619 if (op != MC_LOCK || ap == NULL) {
7620 anon_array_exit(&cookie);
7621 ANON_LOCK_EXIT(&->a_rwlock);
7622 }
7623 } else {
7624 vp = svd->vp;
7625 off = offset;
7626 }
7627
7628 /*
7629 * Get page frame. It's ok if the page is
7630 * not available when we're unlocking, as this
7631 * may simply mean that a page we locked got
7632 * truncated out of existence after we locked it.
7633 *
7634 * Invoke VOP_GETPAGE() to obtain the page struct
7635 * since we may need to read it from disk if its
7636 * been paged out.
7637 */
7638 if (op != MC_LOCK)
7639 pp = page_lookup(vp, off, SE_SHARED);
7640 else {
7641 page_t *pl[1 + 1];
7642 int error;
7643
7644 ASSERT(vp != NULL);
7645
7646 error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7647 (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7648 S_OTHER, svd->cred, NULL);
7649
7650 if (error && ap != NULL) {
7651 anon_array_exit(&cookie);
7652 ANON_LOCK_EXIT(&->a_rwlock);
7653 }
7654
7655 /*
7656 * If the error is EDEADLK then we must bounce
7657 * up and drop all vm subsystem locks and then
7658 * retry the operation later
7659 * This behavior is a temporary measure because
7660 * ufs/sds logging is badly designed and will
7661 * deadlock if we don't allow this bounce to
7662 * happen. The real solution is to re-design
7663 * the logging code to work properly. See bug
7664 * 4125102 for details of the problem.
7665 */
7666 if (error == EDEADLK) {
7667 err = error;
7668 goto out;
7669 }
7670 /*
7671 * Quit if we fail to fault in the page. Treat
7672 * the failure as an error, unless the addr
7673 * is mapped beyond the end of a file.
7674 */
7675 if (error && svd->vp) {
7676 va.va_mask = AT_SIZE;
7677 if (VOP_GETATTR(svd->vp, &va, 0,
7678 svd->cred, NULL) != 0) {
7679 err = EIO;
7680 goto out;
7681 }
7682 if (btopr(va.va_size) >=
7683 btopr(off + 1)) {
7684 err = EIO;
7685 goto out;
7686 }
7687 goto out;
7688
7689 } else if (error) {
7690 err = EIO;
7691 goto out;
7692 }
7693 pp = pl[0];
7694 ASSERT(pp != NULL);
7695 }
7696
7697 /*
7698 * See Statement at the beginning of this routine.
7699 *
7700 * claim is always set if MAP_PRIVATE and PROT_WRITE
7701 * irrespective of following factors:
7702 *
7703 * (1) anon slots are populated or not
7704 * (2) cow is broken or not
7705 * (3) refcnt on ap is 1 or greater than 1
7706 *
7707 * See 4140683 for details
7708 */
7709 claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7710 (svd->type == MAP_PRIVATE));
7711
7712 /*
7713 * Perform page-level operation appropriate to
7714 * operation. If locking, undo the SOFTLOCK
7715 * performed to bring the page into memory
7716 * after setting the lock. If unlocking,
7717 * and no page was found, account for the claim
7718 * separately.
7719 */
7720 if (op == MC_LOCK) {
7721 int ret = 1; /* Assume success */
7722
7723 ASSERT(!VPP_ISPPLOCK(vpp));
7724
7725 ret = page_pp_lock(pp, claim, 0);
7726 if (ap != NULL) {
7727 if (ap->an_pvp != NULL) {
7728 anon_swap_free(ap, pp);
7729 }
7730 anon_array_exit(&cookie);
7731 ANON_LOCK_EXIT(&->a_rwlock);
7732 }
7733 if (ret == 0) {
7734 /* locking page failed */
7735 page_unlock(pp);
7736 err = EAGAIN;
7737 goto out;
7738 }
7739 VPP_SETPPLOCK(vpp);
7740 if (sp != NULL) {
7741 if (pp->p_lckcnt == 1)
7742 locked_bytes += PAGESIZE;
7743 } else
7744 locked_bytes += PAGESIZE;
7745
7746 if (lockmap != (ulong_t *)NULL)
7747 BT_SET(lockmap, pos);
7748
7749 page_unlock(pp);
7750 } else {
7751 ASSERT(VPP_ISPPLOCK(vpp));
7752 if (pp != NULL) {
7753 /* sysV pages should be locked */
7754 ASSERT(sp == NULL || pp->p_lckcnt > 0);
7755 page_pp_unlock(pp, claim, 0);
7756 if (sp != NULL) {
7757 if (pp->p_lckcnt == 0)
7758 unlocked_bytes
7759 += PAGESIZE;
7760 } else
7761 unlocked_bytes += PAGESIZE;
7762 page_unlock(pp);
7763 } else {
7764 ASSERT(sp == NULL);
7765 unlocked_bytes += PAGESIZE;
7766 }
7767 VPP_CLRPPLOCK(vpp);
7768 }
7769 }
7770 }
7771 out:
7772 if (op == MC_LOCK) {
7773 /* Credit back bytes that did not get locked */
7774 if ((unlocked_bytes - locked_bytes) > 0) {
7775 if (proj == NULL)
7776 mutex_enter(&p->p_lock);
7777 rctl_decr_locked_mem(p, proj,
7778 (unlocked_bytes - locked_bytes), chargeproc);
7779 if (proj == NULL)
7780 mutex_exit(&p->p_lock);
7781 }
7782
7783 } else {
7784 /* Account bytes that were unlocked */
7785 if (unlocked_bytes > 0) {
7786 if (proj == NULL)
7787 mutex_enter(&p->p_lock);
7788 rctl_decr_locked_mem(p, proj, unlocked_bytes,
7789 chargeproc);
7790 if (proj == NULL)
7791 mutex_exit(&p->p_lock);
7792 }
7793 }
7794 if (sp != NULL)
7795 mutex_exit(&sp->shm_mlock);
7796 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7797
7798 return (err);
7799 }
7800
7801 /*
7802 * Set advice from user for specified pages
7803 * There are 5 types of advice:
7804 * MADV_NORMAL - Normal (default) behavior (whatever that is)
7805 * MADV_RANDOM - Random page references
7806 * do not allow readahead or 'klustering'
7807 * MADV_SEQUENTIAL - Sequential page references
7808 * Pages previous to the one currently being
7809 * accessed (determined by fault) are 'not needed'
7810 * and are freed immediately
7811 * MADV_WILLNEED - Pages are likely to be used (fault ahead in mctl)
7812 * MADV_DONTNEED - Pages are not needed (synced out in mctl)
7813 * MADV_FREE - Contents can be discarded
7814 * MADV_ACCESS_DEFAULT- Default access
7815 * MADV_ACCESS_LWP - Next LWP will access heavily
7816 * MADV_ACCESS_MANY- Many LWPs or processes will access heavily
7817 */
7818 static int
7819 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
7820 {
7821 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7822 size_t page;
7823 int err = 0;
7824 int already_set;
7825 struct anon_map *amp;
7826 ulong_t anon_index;
7827 struct seg *next;
7828 lgrp_mem_policy_t policy;
7829 struct seg *prev;
7830 struct vnode *vp;
7831
7832 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7833
7834 /*
7835 * In case of MADV_FREE, we won't be modifying any segment private
7836 * data structures; so, we only need to grab READER's lock
7837 */
7838 if (behav != MADV_FREE) {
7839 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7840 if (svd->tr_state != SEGVN_TR_OFF) {
7841 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7842 return (0);
7843 }
7844 } else {
7845 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7846 }
7847
7848 /*
7849 * Large pages are assumed to be only turned on when accesses to the
7850 * segment's address range have spatial and temporal locality. That
7851 * justifies ignoring MADV_SEQUENTIAL for large page segments.
7852 * Also, ignore advice affecting lgroup memory allocation
7853 * if don't need to do lgroup optimizations on this system
7854 */
7855
7856 if ((behav == MADV_SEQUENTIAL &&
7857 (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
7858 (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
7859 behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
7860 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7861 return (0);
7862 }
7863
7864 if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
7865 behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
7866 /*
7867 * Since we are going to unload hat mappings
7868 * we first have to flush the cache. Otherwise
7869 * this might lead to system panic if another
7870 * thread is doing physio on the range whose
7871 * mappings are unloaded by madvise(3C).
7872 */
7873 if (svd->softlockcnt > 0) {
7874 /*
7875 * If this is shared segment non 0 softlockcnt
7876 * means locked pages are still in use.
7877 */
7878 if (svd->type == MAP_SHARED) {
7879 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7880 return (EAGAIN);
7881 }
7882 /*
7883 * Since we do have the segvn writers lock
7884 * nobody can fill the cache with entries
7885 * belonging to this seg during the purge.
7886 * The flush either succeeds or we still
7887 * have pending I/Os. In the later case,
7888 * madvise(3C) fails.
7889 */
7890 segvn_purge(seg);
7891 if (svd->softlockcnt > 0) {
7892 /*
7893 * Since madvise(3C) is advisory and
7894 * it's not part of UNIX98, madvise(3C)
7895 * failure here doesn't cause any hardship.
7896 * Note that we don't block in "as" layer.
7897 */
7898 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7899 return (EAGAIN);
7900 }
7901 } else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7902 svd->amp->a_softlockcnt > 0) {
7903 /*
7904 * Try to purge this amp's entries from pcache. It
7905 * will succeed only if other segments that share the
7906 * amp have no outstanding softlock's.
7907 */
7908 segvn_purge(seg);
7909 }
7910 }
7911
7912 amp = svd->amp;
7913 vp = svd->vp;
7914 if (behav == MADV_FREE) {
7915 /*
7916 * MADV_FREE is not supported for segments with
7917 * underlying object; if anonmap is NULL, anon slots
7918 * are not yet populated and there is nothing for
7919 * us to do. As MADV_FREE is advisory, we don't
7920 * return error in either case.
7921 */
7922 if (vp != NULL || amp == NULL) {
7923 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7924 return (0);
7925 }
7926
7927 segvn_purge(seg);
7928
7929 page = seg_page(seg, addr);
7930 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7931 anon_disclaim(amp, svd->anon_index + page, len);
7932 ANON_LOCK_EXIT(&->a_rwlock);
7933 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7934 return (0);
7935 }
7936
7937 /*
7938 * If advice is to be applied to entire segment,
7939 * use advice field in seg_data structure
7940 * otherwise use appropriate vpage entry.
7941 */
7942 if ((addr == seg->s_base) && (len == seg->s_size)) {
7943 switch (behav) {
7944 case MADV_ACCESS_LWP:
7945 case MADV_ACCESS_MANY:
7946 case MADV_ACCESS_DEFAULT:
7947 /*
7948 * Set memory allocation policy for this segment
7949 */
7950 policy = lgrp_madv_to_policy(behav, len, svd->type);
7951 if (svd->type == MAP_SHARED)
7952 already_set = lgrp_shm_policy_set(policy, amp,
7953 svd->anon_index, vp, svd->offset, len);
7954 else {
7955 /*
7956 * For private memory, need writers lock on
7957 * address space because the segment may be
7958 * split or concatenated when changing policy
7959 */
7960 if (AS_READ_HELD(seg->s_as,
7961 &seg->s_as->a_lock)) {
7962 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7963 return (IE_RETRY);
7964 }
7965
7966 already_set = lgrp_privm_policy_set(policy,
7967 &svd->policy_info, len);
7968 }
7969
7970 /*
7971 * If policy set already and it shouldn't be reapplied,
7972 * don't do anything.
7973 */
7974 if (already_set &&
7975 !LGRP_MEM_POLICY_REAPPLICABLE(policy))
7976 break;
7977
7978 /*
7979 * Mark any existing pages in given range for
7980 * migration
7981 */
7982 page_mark_migrate(seg, addr, len, amp, svd->anon_index,
7983 vp, svd->offset, 1);
7984
7985 /*
7986 * If same policy set already or this is a shared
7987 * memory segment, don't need to try to concatenate
7988 * segment with adjacent ones.
7989 */
7990 if (already_set || svd->type == MAP_SHARED)
7991 break;
7992
7993 /*
7994 * Try to concatenate this segment with previous
7995 * one and next one, since we changed policy for
7996 * this one and it may be compatible with adjacent
7997 * ones now.
7998 */
7999 prev = AS_SEGPREV(seg->s_as, seg);
8000 next = AS_SEGNEXT(seg->s_as, seg);
8001
8002 if (next && next->s_ops == &segvn_ops &&
8003 addr + len == next->s_base)
8004 (void) segvn_concat(seg, next, 1);
8005
8006 if (prev && prev->s_ops == &segvn_ops &&
8007 addr == prev->s_base + prev->s_size) {
8008 /*
8009 * Drop lock for private data of current
8010 * segment before concatenating (deleting) it
8011 * and return IE_REATTACH to tell as_ctl() that
8012 * current segment has changed
8013 */
8014 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8015 if (!segvn_concat(prev, seg, 1))
8016 err = IE_REATTACH;
8017
8018 return (err);
8019 }
8020 break;
8021
8022 case MADV_SEQUENTIAL:
8023 /*
8024 * unloading mapping guarantees
8025 * detection in segvn_fault
8026 */
8027 ASSERT(seg->s_szc == 0);
8028 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8029 hat_unload(seg->s_as->a_hat, addr, len,
8030 HAT_UNLOAD);
8031 /* FALLTHROUGH */
8032 case MADV_NORMAL:
8033 case MADV_RANDOM:
8034 svd->advice = (uchar_t)behav;
8035 svd->pageadvice = 0;
8036 break;
8037 case MADV_WILLNEED: /* handled in memcntl */
8038 case MADV_DONTNEED: /* handled in memcntl */
8039 case MADV_FREE: /* handled above */
8040 break;
8041 default:
8042 err = EINVAL;
8043 }
8044 } else {
8045 caddr_t eaddr;
8046 struct seg *new_seg;
8047 struct segvn_data *new_svd;
8048 u_offset_t off;
8049 caddr_t oldeaddr;
8050
8051 page = seg_page(seg, addr);
8052
8053 segvn_vpage(seg);
8054
8055 switch (behav) {
8056 struct vpage *bvpp, *evpp;
8057
8058 case MADV_ACCESS_LWP:
8059 case MADV_ACCESS_MANY:
8060 case MADV_ACCESS_DEFAULT:
8061 /*
8062 * Set memory allocation policy for portion of this
8063 * segment
8064 */
8065
8066 /*
8067 * Align address and length of advice to page
8068 * boundaries for large pages
8069 */
8070 if (seg->s_szc != 0) {
8071 size_t pgsz;
8072
8073 pgsz = page_get_pagesize(seg->s_szc);
8074 addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8075 len = P2ROUNDUP(len, pgsz);
8076 }
8077
8078 /*
8079 * Check to see whether policy is set already
8080 */
8081 policy = lgrp_madv_to_policy(behav, len, svd->type);
8082
8083 anon_index = svd->anon_index + page;
8084 off = svd->offset + (uintptr_t)(addr - seg->s_base);
8085
8086 if (svd->type == MAP_SHARED)
8087 already_set = lgrp_shm_policy_set(policy, amp,
8088 anon_index, vp, off, len);
8089 else
8090 already_set =
8091 (policy == svd->policy_info.mem_policy);
8092
8093 /*
8094 * If policy set already and it shouldn't be reapplied,
8095 * don't do anything.
8096 */
8097 if (already_set &&
8098 !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8099 break;
8100
8101 /*
8102 * For private memory, need writers lock on
8103 * address space because the segment may be
8104 * split or concatenated when changing policy
8105 */
8106 if (svd->type == MAP_PRIVATE &&
8107 AS_READ_HELD(seg->s_as, &seg->s_as->a_lock)) {
8108 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8109 return (IE_RETRY);
8110 }
8111
8112 /*
8113 * Mark any existing pages in given range for
8114 * migration
8115 */
8116 page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8117 vp, svd->offset, 1);
8118
8119 /*
8120 * Don't need to try to split or concatenate
8121 * segments, since policy is same or this is a shared
8122 * memory segment
8123 */
8124 if (already_set || svd->type == MAP_SHARED)
8125 break;
8126
8127 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8128 ASSERT(svd->amp == NULL);
8129 ASSERT(svd->tr_state == SEGVN_TR_OFF);
8130 ASSERT(svd->softlockcnt == 0);
8131 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8132 HAT_REGION_TEXT);
8133 svd->rcookie = HAT_INVALID_REGION_COOKIE;
8134 }
8135
8136 /*
8137 * Split off new segment if advice only applies to a
8138 * portion of existing segment starting in middle
8139 */
8140 new_seg = NULL;
8141 eaddr = addr + len;
8142 oldeaddr = seg->s_base + seg->s_size;
8143 if (addr > seg->s_base) {
8144 /*
8145 * Must flush I/O page cache
8146 * before splitting segment
8147 */
8148 if (svd->softlockcnt > 0)
8149 segvn_purge(seg);
8150
8151 /*
8152 * Split segment and return IE_REATTACH to tell
8153 * as_ctl() that current segment changed
8154 */
8155 new_seg = segvn_split_seg(seg, addr);
8156 new_svd = (struct segvn_data *)new_seg->s_data;
8157 err = IE_REATTACH;
8158
8159 /*
8160 * If new segment ends where old one
8161 * did, try to concatenate the new
8162 * segment with next one.
8163 */
8164 if (eaddr == oldeaddr) {
8165 /*
8166 * Set policy for new segment
8167 */
8168 (void) lgrp_privm_policy_set(policy,
8169 &new_svd->policy_info,
8170 new_seg->s_size);
8171
8172 next = AS_SEGNEXT(new_seg->s_as,
8173 new_seg);
8174
8175 if (next &&
8176 next->s_ops == &segvn_ops &&
8177 eaddr == next->s_base)
8178 (void) segvn_concat(new_seg,
8179 next, 1);
8180 }
8181 }
8182
8183 /*
8184 * Split off end of existing segment if advice only
8185 * applies to a portion of segment ending before
8186 * end of the existing segment
8187 */
8188 if (eaddr < oldeaddr) {
8189 /*
8190 * Must flush I/O page cache
8191 * before splitting segment
8192 */
8193 if (svd->softlockcnt > 0)
8194 segvn_purge(seg);
8195
8196 /*
8197 * If beginning of old segment was already
8198 * split off, use new segment to split end off
8199 * from.
8200 */
8201 if (new_seg != NULL && new_seg != seg) {
8202 /*
8203 * Split segment
8204 */
8205 (void) segvn_split_seg(new_seg, eaddr);
8206
8207 /*
8208 * Set policy for new segment
8209 */
8210 (void) lgrp_privm_policy_set(policy,
8211 &new_svd->policy_info,
8212 new_seg->s_size);
8213 } else {
8214 /*
8215 * Split segment and return IE_REATTACH
8216 * to tell as_ctl() that current
8217 * segment changed
8218 */
8219 (void) segvn_split_seg(seg, eaddr);
8220 err = IE_REATTACH;
8221
8222 (void) lgrp_privm_policy_set(policy,
8223 &svd->policy_info, seg->s_size);
8224
8225 /*
8226 * If new segment starts where old one
8227 * did, try to concatenate it with
8228 * previous segment.
8229 */
8230 if (addr == seg->s_base) {
8231 prev = AS_SEGPREV(seg->s_as,
8232 seg);
8233
8234 /*
8235 * Drop lock for private data
8236 * of current segment before
8237 * concatenating (deleting) it
8238 */
8239 if (prev &&
8240 prev->s_ops ==
8241 &segvn_ops &&
8242 addr == prev->s_base +
8243 prev->s_size) {
8244 SEGVN_LOCK_EXIT(
8245 seg->s_as,
8246 &svd->lock);
8247 (void) segvn_concat(
8248 prev, seg, 1);
8249 return (err);
8250 }
8251 }
8252 }
8253 }
8254 break;
8255 case MADV_SEQUENTIAL:
8256 ASSERT(seg->s_szc == 0);
8257 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8258 hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8259 /* FALLTHROUGH */
8260 case MADV_NORMAL:
8261 case MADV_RANDOM:
8262 bvpp = &svd->vpage[page];
8263 evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8264 for (; bvpp < evpp; bvpp++)
8265 VPP_SETADVICE(bvpp, behav);
8266 svd->advice = MADV_NORMAL;
8267 break;
8268 case MADV_WILLNEED: /* handled in memcntl */
8269 case MADV_DONTNEED: /* handled in memcntl */
8270 case MADV_FREE: /* handled above */
8271 break;
8272 default:
8273 err = EINVAL;
8274 }
8275 }
8276 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8277 return (err);
8278 }
8279
8280 /*
8281 * Create a vpage structure for this seg.
8282 */
8283 static void
8284 segvn_vpage(struct seg *seg)
8285 {
8286 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8287 struct vpage *vp, *evp;
8288
8289 ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8290
8291 /*
8292 * If no vpage structure exists, allocate one. Copy the protections
8293 * and the advice from the segment itself to the individual pages.
8294 */
8295 if (svd->vpage == NULL) {
8296 svd->pageadvice = 1;
8297 svd->vpage = kmem_zalloc(seg_pages(seg) * sizeof (struct vpage),
8298 KM_SLEEP);
8299 evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8300 for (vp = svd->vpage; vp < evp; vp++) {
8301 VPP_SETPROT(vp, svd->prot);
8302 VPP_SETADVICE(vp, svd->advice);
8303 }
8304 }
8305 }
8306
8307 /*
8308 * Dump the pages belonging to this segvn segment.
8309 */
8310 static void
8311 segvn_dump(struct seg *seg)
8312 {
8313 struct segvn_data *svd;
8314 page_t *pp;
8315 struct anon_map *amp;
8316 ulong_t anon_index;
8317 struct vnode *vp;
8318 u_offset_t off, offset;
8319 pfn_t pfn;
8320 pgcnt_t page, npages;
8321 caddr_t addr;
8322
8323 npages = seg_pages(seg);
8324 svd = (struct segvn_data *)seg->s_data;
8325 vp = svd->vp;
8326 off = offset = svd->offset;
8327 addr = seg->s_base;
8328
8329 if ((amp = svd->amp) != NULL) {
8330 anon_index = svd->anon_index;
8331 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
8332 }
8333
8334 for (page = 0; page < npages; page++, offset += PAGESIZE) {
8335 struct anon *ap;
8336 int we_own_it = 0;
8337
8338 if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8339 swap_xlate_nopanic(ap, &vp, &off);
8340 } else {
8341 vp = svd->vp;
8342 off = offset;
8343 }
8344
8345 /*
8346 * If pp == NULL, the page either does not exist
8347 * or is exclusively locked. So determine if it
8348 * exists before searching for it.
8349 */
8350
8351 if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8352 we_own_it = 1;
8353 else
8354 pp = page_exists(vp, off);
8355
8356 if (pp) {
8357 pfn = page_pptonum(pp);
8358 dump_addpage(seg->s_as, addr, pfn);
8359 if (we_own_it)
8360 page_unlock(pp);
8361 }
8362 addr += PAGESIZE;
8363 dump_timeleft = dump_timeout;
8364 }
8365
8366 if (amp != NULL)
8367 ANON_LOCK_EXIT(&->a_rwlock);
8368 }
8369
8370 #ifdef DEBUG
8371 static uint32_t segvn_pglock_mtbf = 0;
8372 #endif
8373
8374 #define PCACHE_SHWLIST ((page_t *)-2)
8375 #define NOPCACHE_SHWLIST ((page_t *)-1)
8376
8377 /*
8378 * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8379 * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8380 * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8381 * the same parts of the segment. Currently shadow list creation is only
8382 * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8383 * tagged with segment pointer, starting virtual address and length. This
8384 * approach for MAP_SHARED segments may add many pcache entries for the same
8385 * set of pages and lead to long hash chains that decrease pcache lookup
8386 * performance. To avoid this issue for shared segments shared anon map and
8387 * starting anon index are used for pcache entry tagging. This allows all
8388 * segments to share pcache entries for the same anon range and reduces pcache
8389 * chain's length as well as memory overhead from duplicate shadow lists and
8390 * pcache entries.
8391 *
8392 * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8393 * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8394 * part of softlockcnt accounting is done differently for private and shared
8395 * segments. In private segment case softlock is only incremented when a new
8396 * shadow list is created but not when an existing one is found via
8397 * seg_plookup(). pcache entries have reference count incremented/decremented
8398 * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8399 * reference count can be purged (and purging is needed before segment can be
8400 * freed). When a private segment pcache entry is purged segvn_reclaim() will
8401 * decrement softlockcnt. Since in private segment case each of its pcache
8402 * entries only belongs to this segment we can expect that when
8403 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8404 * segment purge will succeed and softlockcnt will drop to 0. In shared
8405 * segment case reference count in pcache entry counts active locks from many
8406 * different segments so we can't expect segment purging to succeed even when
8407 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8408 * segment. To be able to determine when there're no pending pagelocks in
8409 * shared segment case we don't rely on purging to make softlockcnt drop to 0
8410 * but instead softlockcnt is incremented and decremented for every
8411 * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8412 * list was created or an existing one was found. When softlockcnt drops to 0
8413 * this segment no longer has any claims for pcached shadow lists and the
8414 * segment can be freed even if there're still active pcache entries
8415 * shared by this segment anon map. Shared segment pcache entries belong to
8416 * anon map and are typically removed when anon map is freed after all
8417 * processes destroy the segments that use this anon map.
8418 */
8419 static int
8420 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8421 enum lock_type type, enum seg_rw rw)
8422 {
8423 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8424 size_t np;
8425 pgcnt_t adjustpages;
8426 pgcnt_t npages;
8427 ulong_t anon_index;
8428 uint_t protchk = (rw == S_READ) ? PROT_READ : PROT_WRITE;
8429 uint_t error;
8430 struct anon_map *amp;
8431 pgcnt_t anpgcnt;
8432 struct page **pplist, **pl, *pp;
8433 caddr_t a;
8434 size_t page;
8435 caddr_t lpgaddr, lpgeaddr;
8436 anon_sync_obj_t cookie;
8437 int anlock;
8438 struct anon_map *pamp;
8439 caddr_t paddr;
8440 seg_preclaim_cbfunc_t preclaim_callback;
8441 size_t pgsz;
8442 int use_pcache;
8443 size_t wlen;
8444 uint_t pflags = 0;
8445 int sftlck_sbase = 0;
8446 int sftlck_send = 0;
8447
8448 #ifdef DEBUG
8449 if (type == L_PAGELOCK && segvn_pglock_mtbf) {
8450 hrtime_t ts = gethrtime();
8451 if ((ts % segvn_pglock_mtbf) == 0) {
8452 return (ENOTSUP);
8453 }
8454 if ((ts % segvn_pglock_mtbf) == 1) {
8455 return (EFAULT);
8456 }
8457 }
8458 #endif
8459
8460 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8461 "segvn_pagelock: start seg %p addr %p", seg, addr);
8462
8463 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8464 ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
8465
8466 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8467
8468 /*
8469 * for now we only support pagelock to anon memory. We would have to
8470 * check protections for vnode objects and call into the vnode driver.
8471 * That's too much for a fast path. Let the fault entry point handle
8472 * it.
8473 */
8474 if (svd->vp != NULL) {
8475 if (type == L_PAGELOCK) {
8476 error = ENOTSUP;
8477 goto out;
8478 }
8479 panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8480 }
8481 if ((amp = svd->amp) == NULL) {
8482 if (type == L_PAGELOCK) {
8483 error = EFAULT;
8484 goto out;
8485 }
8486 panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8487 }
8488 if (rw != S_READ && rw != S_WRITE) {
8489 if (type == L_PAGELOCK) {
8490 error = ENOTSUP;
8491 goto out;
8492 }
8493 panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8494 }
8495
8496 if (seg->s_szc != 0) {
8497 /*
8498 * We are adjusting the pagelock region to the large page size
8499 * boundary because the unlocked part of a large page cannot
8500 * be freed anyway unless all constituent pages of a large
8501 * page are locked. Bigger regions reduce pcache chain length
8502 * and improve lookup performance. The tradeoff is that the
8503 * very first segvn_pagelock() call for a given page is more
8504 * expensive if only 1 page_t is needed for IO. This is only
8505 * an issue if pcache entry doesn't get reused by several
8506 * subsequent calls. We optimize here for the case when pcache
8507 * is heavily used by repeated IOs to the same address range.
8508 *
8509 * Note segment's page size cannot change while we are holding
8510 * as lock. And then it cannot change while softlockcnt is
8511 * not 0. This will allow us to correctly recalculate large
8512 * page size region for the matching pageunlock/reclaim call
8513 * since as_pageunlock() caller must always match
8514 * as_pagelock() call's addr and len.
8515 *
8516 * For pageunlock *ppp points to the pointer of page_t that
8517 * corresponds to the real unadjusted start address. Similar
8518 * for pagelock *ppp must point to the pointer of page_t that
8519 * corresponds to the real unadjusted start address.
8520 */
8521 pgsz = page_get_pagesize(seg->s_szc);
8522 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8523 adjustpages = btop((uintptr_t)(addr - lpgaddr));
8524 } else if (len < segvn_pglock_comb_thrshld) {
8525 lpgaddr = addr;
8526 lpgeaddr = addr + len;
8527 adjustpages = 0;
8528 pgsz = PAGESIZE;
8529 } else {
8530 /*
8531 * Align the address range of large enough requests to allow
8532 * combining of different shadow lists into 1 to reduce memory
8533 * overhead from potentially overlapping large shadow lists
8534 * (worst case is we have a 1MB IO into buffers with start
8535 * addresses separated by 4K). Alignment is only possible if
8536 * padded chunks have sufficient access permissions. Note
8537 * permissions won't change between L_PAGELOCK and
8538 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8539 * segvn_setprot() to wait until softlockcnt drops to 0. This
8540 * allows us to determine in L_PAGEUNLOCK the same range we
8541 * computed in L_PAGELOCK.
8542 *
8543 * If alignment is limited by segment ends set
8544 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8545 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8546 * per segment counters. In L_PAGEUNLOCK case decrease
8547 * softlockcnt_sbase/softlockcnt_send counters if
8548 * sftlck_sbase/sftlck_send flags are set. When
8549 * softlockcnt_sbase/softlockcnt_send are non 0
8550 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8551 * won't merge the segments. This restriction combined with
8552 * restriction on segment unmapping and splitting for segments
8553 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8554 * correctly determine the same range that was previously
8555 * locked by matching L_PAGELOCK.
8556 */
8557 pflags = SEGP_PSHIFT | (segvn_pglock_comb_bshift << 16);
8558 pgsz = PAGESIZE;
8559 if (svd->type == MAP_PRIVATE) {
8560 lpgaddr = (caddr_t)P2ALIGN((uintptr_t)addr,
8561 segvn_pglock_comb_balign);
8562 if (lpgaddr < seg->s_base) {
8563 lpgaddr = seg->s_base;
8564 sftlck_sbase = 1;
8565 }
8566 } else {
8567 ulong_t aix = svd->anon_index + seg_page(seg, addr);
8568 ulong_t aaix = P2ALIGN(aix, segvn_pglock_comb_palign);
8569 if (aaix < svd->anon_index) {
8570 lpgaddr = seg->s_base;
8571 sftlck_sbase = 1;
8572 } else {
8573 lpgaddr = addr - ptob(aix - aaix);
8574 ASSERT(lpgaddr >= seg->s_base);
8575 }
8576 }
8577 if (svd->pageprot && lpgaddr != addr) {
8578 struct vpage *vp = &svd->vpage[seg_page(seg, lpgaddr)];
8579 struct vpage *evp = &svd->vpage[seg_page(seg, addr)];
8580 while (vp < evp) {
8581 if ((VPP_PROT(vp) & protchk) == 0) {
8582 break;
8583 }
8584 vp++;
8585 }
8586 if (vp < evp) {
8587 lpgaddr = addr;
8588 pflags = 0;
8589 }
8590 }
8591 lpgeaddr = addr + len;
8592 if (pflags) {
8593 if (svd->type == MAP_PRIVATE) {
8594 lpgeaddr = (caddr_t)P2ROUNDUP(
8595 (uintptr_t)lpgeaddr,
8596 segvn_pglock_comb_balign);
8597 } else {
8598 ulong_t aix = svd->anon_index +
8599 seg_page(seg, lpgeaddr);
8600 ulong_t aaix = P2ROUNDUP(aix,
8601 segvn_pglock_comb_palign);
8602 if (aaix < aix) {
8603 lpgeaddr = 0;
8604 } else {
8605 lpgeaddr += ptob(aaix - aix);
8606 }
8607 }
8608 if (lpgeaddr == 0 ||
8609 lpgeaddr > seg->s_base + seg->s_size) {
8610 lpgeaddr = seg->s_base + seg->s_size;
8611 sftlck_send = 1;
8612 }
8613 }
8614 if (svd->pageprot && lpgeaddr != addr + len) {
8615 struct vpage *vp;
8616 struct vpage *evp;
8617
8618 vp = &svd->vpage[seg_page(seg, addr + len)];
8619 evp = &svd->vpage[seg_page(seg, lpgeaddr)];
8620
8621 while (vp < evp) {
8622 if ((VPP_PROT(vp) & protchk) == 0) {
8623 break;
8624 }
8625 vp++;
8626 }
8627 if (vp < evp) {
8628 lpgeaddr = addr + len;
8629 }
8630 }
8631 adjustpages = btop((uintptr_t)(addr - lpgaddr));
8632 }
8633
8634 /*
8635 * For MAP_SHARED segments we create pcache entries tagged by amp and
8636 * anon index so that we can share pcache entries with other segments
8637 * that map this amp. For private segments pcache entries are tagged
8638 * with segment and virtual address.
8639 */
8640 if (svd->type == MAP_SHARED) {
8641 pamp = amp;
8642 paddr = (caddr_t)((lpgaddr - seg->s_base) +
8643 ptob(svd->anon_index));
8644 preclaim_callback = shamp_reclaim;
8645 } else {
8646 pamp = NULL;
8647 paddr = lpgaddr;
8648 preclaim_callback = segvn_reclaim;
8649 }
8650
8651 if (type == L_PAGEUNLOCK) {
8652 VM_STAT_ADD(segvnvmstats.pagelock[0]);
8653
8654 /*
8655 * update hat ref bits for /proc. We need to make sure
8656 * that threads tracing the ref and mod bits of the
8657 * address space get the right data.
8658 * Note: page ref and mod bits are updated at reclaim time
8659 */
8660 if (seg->s_as->a_vbits) {
8661 for (a = addr; a < addr + len; a += PAGESIZE) {
8662 if (rw == S_WRITE) {
8663 hat_setstat(seg->s_as, a,
8664 PAGESIZE, P_REF | P_MOD);
8665 } else {
8666 hat_setstat(seg->s_as, a,
8667 PAGESIZE, P_REF);
8668 }
8669 }
8670 }
8671
8672 /*
8673 * Check the shadow list entry after the last page used in
8674 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
8675 * was not inserted into pcache and is not large page
8676 * adjusted. In this case call reclaim callback directly and
8677 * don't adjust the shadow list start and size for large
8678 * pages.
8679 */
8680 npages = btop(len);
8681 if ((*ppp)[npages] == NOPCACHE_SHWLIST) {
8682 void *ptag;
8683 if (pamp != NULL) {
8684 ASSERT(svd->type == MAP_SHARED);
8685 ptag = (void *)pamp;
8686 paddr = (caddr_t)((addr - seg->s_base) +
8687 ptob(svd->anon_index));
8688 } else {
8689 ptag = (void *)seg;
8690 paddr = addr;
8691 }
8692 (*preclaim_callback)(ptag, paddr, len, *ppp, rw, 0);
8693 } else {
8694 ASSERT((*ppp)[npages] == PCACHE_SHWLIST ||
8695 IS_SWAPFSVP((*ppp)[npages]->p_vnode));
8696 len = lpgeaddr - lpgaddr;
8697 npages = btop(len);
8698 seg_pinactive(seg, pamp, paddr, len,
8699 *ppp - adjustpages, rw, pflags, preclaim_callback);
8700 }
8701
8702 if (pamp != NULL) {
8703 ASSERT(svd->type == MAP_SHARED);
8704 ASSERT(svd->softlockcnt >= npages);
8705 atomic_add_long((ulong_t *)&svd->softlockcnt, -npages);
8706 }
8707
8708 if (sftlck_sbase) {
8709 ASSERT(svd->softlockcnt_sbase > 0);
8710 atomic_add_long((ulong_t *)&svd->softlockcnt_sbase, -1);
8711 }
8712 if (sftlck_send) {
8713 ASSERT(svd->softlockcnt_send > 0);
8714 atomic_add_long((ulong_t *)&svd->softlockcnt_send, -1);
8715 }
8716
8717 /*
8718 * If someone is blocked while unmapping, we purge
8719 * segment page cache and thus reclaim pplist synchronously
8720 * without waiting for seg_pasync_thread. This speeds up
8721 * unmapping in cases where munmap(2) is called, while
8722 * raw async i/o is still in progress or where a thread
8723 * exits on data fault in a multithreaded application.
8724 */
8725 if (AS_ISUNMAPWAIT(seg->s_as)) {
8726 if (svd->softlockcnt == 0) {
8727 mutex_enter(&seg->s_as->a_contents);
8728 if (AS_ISUNMAPWAIT(seg->s_as)) {
8729 AS_CLRUNMAPWAIT(seg->s_as);
8730 cv_broadcast(&seg->s_as->a_cv);
8731 }
8732 mutex_exit(&seg->s_as->a_contents);
8733 } else if (pamp == NULL) {
8734 /*
8735 * softlockcnt is not 0 and this is a
8736 * MAP_PRIVATE segment. Try to purge its
8737 * pcache entries to reduce softlockcnt.
8738 * If it drops to 0 segvn_reclaim()
8739 * will wake up a thread waiting on
8740 * unmapwait flag.
8741 *
8742 * We don't purge MAP_SHARED segments with non
8743 * 0 softlockcnt since IO is still in progress
8744 * for such segments.
8745 */
8746 ASSERT(svd->type == MAP_PRIVATE);
8747 segvn_purge(seg);
8748 }
8749 }
8750 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8751 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8752 "segvn_pagelock: unlock seg %p addr %p", seg, addr);
8753 return (0);
8754 }
8755
8756 /* The L_PAGELOCK case ... */
8757
8758 VM_STAT_ADD(segvnvmstats.pagelock[1]);
8759
8760 /*
8761 * For MAP_SHARED segments we have to check protections before
8762 * seg_plookup() since pcache entries may be shared by many segments
8763 * with potentially different page protections.
8764 */
8765 if (pamp != NULL) {
8766 ASSERT(svd->type == MAP_SHARED);
8767 if (svd->pageprot == 0) {
8768 if ((svd->prot & protchk) == 0) {
8769 error = EACCES;
8770 goto out;
8771 }
8772 } else {
8773 /*
8774 * check page protections
8775 */
8776 caddr_t ea;
8777
8778 if (seg->s_szc) {
8779 a = lpgaddr;
8780 ea = lpgeaddr;
8781 } else {
8782 a = addr;
8783 ea = addr + len;
8784 }
8785 for (; a < ea; a += pgsz) {
8786 struct vpage *vp;
8787
8788 ASSERT(seg->s_szc == 0 ||
8789 sameprot(seg, a, pgsz));
8790 vp = &svd->vpage[seg_page(seg, a)];
8791 if ((VPP_PROT(vp) & protchk) == 0) {
8792 error = EACCES;
8793 goto out;
8794 }
8795 }
8796 }
8797 }
8798
8799 /*
8800 * try to find pages in segment page cache
8801 */
8802 pplist = seg_plookup(seg, pamp, paddr, lpgeaddr - lpgaddr, rw, pflags);
8803 if (pplist != NULL) {
8804 if (pamp != NULL) {
8805 npages = btop((uintptr_t)(lpgeaddr - lpgaddr));
8806 ASSERT(svd->type == MAP_SHARED);
8807 atomic_add_long((ulong_t *)&svd->softlockcnt,
8808 npages);
8809 }
8810 if (sftlck_sbase) {
8811 atomic_add_long((ulong_t *)&svd->softlockcnt_sbase, 1);
8812 }
8813 if (sftlck_send) {
8814 atomic_add_long((ulong_t *)&svd->softlockcnt_send, 1);
8815 }
8816 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8817 *ppp = pplist + adjustpages;
8818 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
8819 "segvn_pagelock: cache hit seg %p addr %p", seg, addr);
8820 return (0);
8821 }
8822
8823 /*
8824 * For MAP_SHARED segments we already verified above that segment
8825 * protections allow this pagelock operation.
8826 */
8827 if (pamp == NULL) {
8828 ASSERT(svd->type == MAP_PRIVATE);
8829 if (svd->pageprot == 0) {
8830 if ((svd->prot & protchk) == 0) {
8831 error = EACCES;
8832 goto out;
8833 }
8834 if (svd->prot & PROT_WRITE) {
8835 wlen = lpgeaddr - lpgaddr;
8836 } else {
8837 wlen = 0;
8838 ASSERT(rw == S_READ);
8839 }
8840 } else {
8841 int wcont = 1;
8842 /*
8843 * check page protections
8844 */
8845 for (a = lpgaddr, wlen = 0; a < lpgeaddr; a += pgsz) {
8846 struct vpage *vp;
8847
8848 ASSERT(seg->s_szc == 0 ||
8849 sameprot(seg, a, pgsz));
8850 vp = &svd->vpage[seg_page(seg, a)];
8851 if ((VPP_PROT(vp) & protchk) == 0) {
8852 error = EACCES;
8853 goto out;
8854 }
8855 if (wcont && (VPP_PROT(vp) & PROT_WRITE)) {
8856 wlen += pgsz;
8857 } else {
8858 wcont = 0;
8859 ASSERT(rw == S_READ);
8860 }
8861 }
8862 }
8863 ASSERT(rw == S_READ || wlen == lpgeaddr - lpgaddr);
8864 ASSERT(rw == S_WRITE || wlen <= lpgeaddr - lpgaddr);
8865 }
8866
8867 /*
8868 * Only build large page adjusted shadow list if we expect to insert
8869 * it into pcache. For large enough pages it's a big overhead to
8870 * create a shadow list of the entire large page. But this overhead
8871 * should be amortized over repeated pcache hits on subsequent reuse
8872 * of this shadow list (IO into any range within this shadow list will
8873 * find it in pcache since we large page align the request for pcache
8874 * lookups). pcache performance is improved with bigger shadow lists
8875 * as it reduces the time to pcache the entire big segment and reduces
8876 * pcache chain length.
8877 */
8878 if (seg_pinsert_check(seg, pamp, paddr,
8879 lpgeaddr - lpgaddr, pflags) == SEGP_SUCCESS) {
8880 addr = lpgaddr;
8881 len = lpgeaddr - lpgaddr;
8882 use_pcache = 1;
8883 } else {
8884 use_pcache = 0;
8885 /*
8886 * Since this entry will not be inserted into the pcache, we
8887 * will not do any adjustments to the starting address or
8888 * size of the memory to be locked.
8889 */
8890 adjustpages = 0;
8891 }
8892 npages = btop(len);
8893
8894 pplist = kmem_alloc(sizeof (page_t *) * (npages + 1), KM_SLEEP);
8895 pl = pplist;
8896 *ppp = pplist + adjustpages;
8897 /*
8898 * If use_pcache is 0 this shadow list is not large page adjusted.
8899 * Record this info in the last entry of shadow array so that
8900 * L_PAGEUNLOCK can determine if it should large page adjust the
8901 * address range to find the real range that was locked.
8902 */
8903 pl[npages] = use_pcache ? PCACHE_SHWLIST : NOPCACHE_SHWLIST;
8904
8905 page = seg_page(seg, addr);
8906 anon_index = svd->anon_index + page;
8907
8908 anlock = 0;
8909 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
8910 ASSERT(amp->a_szc >= seg->s_szc);
8911 anpgcnt = page_get_pagecnt(amp->a_szc);
8912 for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
8913 struct anon *ap;
8914 struct vnode *vp;
8915 u_offset_t off;
8916
8917 /*
8918 * Lock and unlock anon array only once per large page.
8919 * anon_array_enter() locks the root anon slot according to
8920 * a_szc which can't change while anon map is locked. We lock
8921 * anon the first time through this loop and each time we
8922 * reach anon index that corresponds to a root of a large
8923 * page.
8924 */
8925 if (a == addr || P2PHASE(anon_index, anpgcnt) == 0) {
8926 ASSERT(anlock == 0);
8927 anon_array_enter(amp, anon_index, &cookie);
8928 anlock = 1;
8929 }
8930 ap = anon_get_ptr(amp->ahp, anon_index);
8931
8932 /*
8933 * We must never use seg_pcache for COW pages
8934 * because we might end up with original page still
8935 * lying in seg_pcache even after private page is
8936 * created. This leads to data corruption as
8937 * aio_write refers to the page still in cache
8938 * while all other accesses refer to the private
8939 * page.
8940 */
8941 if (ap == NULL || ap->an_refcnt != 1) {
8942 struct vpage *vpage;
8943
8944 if (seg->s_szc) {
8945 error = EFAULT;
8946 break;
8947 }
8948 if (svd->vpage != NULL) {
8949 vpage = &svd->vpage[seg_page(seg, a)];
8950 } else {
8951 vpage = NULL;
8952 }
8953 ASSERT(anlock);
8954 anon_array_exit(&cookie);
8955 anlock = 0;
8956 pp = NULL;
8957 error = segvn_faultpage(seg->s_as->a_hat, seg, a, 0,
8958 vpage, &pp, 0, F_INVAL, rw, 1);
8959 if (error) {
8960 error = fc_decode(error);
8961 break;
8962 }
8963 anon_array_enter(amp, anon_index, &cookie);
8964 anlock = 1;
8965 ap = anon_get_ptr(amp->ahp, anon_index);
8966 if (ap == NULL || ap->an_refcnt != 1) {
8967 error = EFAULT;
8968 break;
8969 }
8970 }
8971 swap_xlate(ap, &vp, &off);
8972 pp = page_lookup_nowait(vp, off, SE_SHARED);
8973 if (pp == NULL) {
8974 error = EFAULT;
8975 break;
8976 }
8977 if (ap->an_pvp != NULL) {
8978 anon_swap_free(ap, pp);
8979 }
8980 /*
8981 * Unlock anon if this is the last slot in a large page.
8982 */
8983 if (P2PHASE(anon_index, anpgcnt) == anpgcnt - 1) {
8984 ASSERT(anlock);
8985 anon_array_exit(&cookie);
8986 anlock = 0;
8987 }
8988 *pplist++ = pp;
8989 }
8990 if (anlock) { /* Ensure the lock is dropped */
8991 anon_array_exit(&cookie);
8992 }
8993 ANON_LOCK_EXIT(&->a_rwlock);
8994
8995 if (a >= addr + len) {
8996 atomic_add_long((ulong_t *)&svd->softlockcnt, npages);
8997 if (pamp != NULL) {
8998 ASSERT(svd->type == MAP_SHARED);
8999 atomic_add_long((ulong_t *)&pamp->a_softlockcnt,
9000 npages);
9001 wlen = len;
9002 }
9003 if (sftlck_sbase) {
9004 atomic_add_long((ulong_t *)&svd->softlockcnt_sbase, 1);
9005 }
9006 if (sftlck_send) {
9007 atomic_add_long((ulong_t *)&svd->softlockcnt_send, 1);
9008 }
9009 if (use_pcache) {
9010 (void) seg_pinsert(seg, pamp, paddr, len, wlen, pl,
9011 rw, pflags, preclaim_callback);
9012 }
9013 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9014 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
9015 "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
9016 return (0);
9017 }
9018
9019 pplist = pl;
9020 np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
9021 while (np > (uint_t)0) {
9022 ASSERT(PAGE_LOCKED(*pplist));
9023 page_unlock(*pplist);
9024 np--;
9025 pplist++;
9026 }
9027 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9028 out:
9029 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9030 *ppp = NULL;
9031 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
9032 "segvn_pagelock: cache miss seg %p addr %p", seg, addr);
9033 return (error);
9034 }
9035
9036 /*
9037 * purge any cached pages in the I/O page cache
9038 */
9039 static void
9040 segvn_purge(struct seg *seg)
9041 {
9042 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9043
9044 /*
9045 * pcache is only used by pure anon segments.
9046 */
9047 if (svd->amp == NULL || svd->vp != NULL) {
9048 return;
9049 }
9050
9051 /*
9052 * For MAP_SHARED segments non 0 segment's softlockcnt means
9053 * active IO is still in progress via this segment. So we only
9054 * purge MAP_SHARED segments when their softlockcnt is 0.
9055 */
9056 if (svd->type == MAP_PRIVATE) {
9057 if (svd->softlockcnt) {
9058 seg_ppurge(seg, NULL, 0);
9059 }
9060 } else if (svd->softlockcnt == 0 && svd->amp->a_softlockcnt != 0) {
9061 seg_ppurge(seg, svd->amp, 0);
9062 }
9063 }
9064
9065 /*
9066 * If async argument is not 0 we are called from pcache async thread and don't
9067 * hold AS lock.
9068 */
9069
9070 /*ARGSUSED*/
9071 static int
9072 segvn_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9073 enum seg_rw rw, int async)
9074 {
9075 struct seg *seg = (struct seg *)ptag;
9076 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9077 pgcnt_t np, npages;
9078 struct page **pl;
9079
9080 npages = np = btop(len);
9081 ASSERT(npages);
9082
9083 ASSERT(svd->vp == NULL && svd->amp != NULL);
9084 ASSERT(svd->softlockcnt >= npages);
9085 ASSERT(async || AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9086
9087 pl = pplist;
9088
9089 ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9090 ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9091
9092 while (np > (uint_t)0) {
9093 if (rw == S_WRITE) {
9094 hat_setrefmod(*pplist);
9095 } else {
9096 hat_setref(*pplist);
9097 }
9098 page_unlock(*pplist);
9099 np--;
9100 pplist++;
9101 }
9102
9103 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9104
9105 /*
9106 * If we are pcache async thread we don't hold AS lock. This means if
9107 * softlockcnt drops to 0 after the decrement below address space may
9108 * get freed. We can't allow it since after softlock derement to 0 we
9109 * still need to access as structure for possible wakeup of unmap
9110 * waiters. To prevent the disappearance of as we take this segment
9111 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9112 * make sure this routine completes before segment is freed.
9113 *
9114 * The second complication we have to deal with in async case is a
9115 * possibility of missed wake up of unmap wait thread. When we don't
9116 * hold as lock here we may take a_contents lock before unmap wait
9117 * thread that was first to see softlockcnt was still not 0. As a
9118 * result we'll fail to wake up an unmap wait thread. To avoid this
9119 * race we set nounmapwait flag in as structure if we drop softlockcnt
9120 * to 0 when we were called by pcache async thread. unmapwait thread
9121 * will not block if this flag is set.
9122 */
9123 if (async) {
9124 mutex_enter(&svd->segfree_syncmtx);
9125 }
9126
9127 if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -npages)) {
9128 if (async || AS_ISUNMAPWAIT(seg->s_as)) {
9129 mutex_enter(&seg->s_as->a_contents);
9130 if (async) {
9131 AS_SETNOUNMAPWAIT(seg->s_as);
9132 }
9133 if (AS_ISUNMAPWAIT(seg->s_as)) {
9134 AS_CLRUNMAPWAIT(seg->s_as);
9135 cv_broadcast(&seg->s_as->a_cv);
9136 }
9137 mutex_exit(&seg->s_as->a_contents);
9138 }
9139 }
9140
9141 if (async) {
9142 mutex_exit(&svd->segfree_syncmtx);
9143 }
9144 return (0);
9145 }
9146
9147 /*ARGSUSED*/
9148 static int
9149 shamp_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9150 enum seg_rw rw, int async)
9151 {
9152 amp_t *amp = (amp_t *)ptag;
9153 pgcnt_t np, npages;
9154 struct page **pl;
9155
9156 npages = np = btop(len);
9157 ASSERT(npages);
9158 ASSERT(amp->a_softlockcnt >= npages);
9159
9160 pl = pplist;
9161
9162 ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9163 ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9164
9165 while (np > (uint_t)0) {
9166 if (rw == S_WRITE) {
9167 hat_setrefmod(*pplist);
9168 } else {
9169 hat_setref(*pplist);
9170 }
9171 page_unlock(*pplist);
9172 np--;
9173 pplist++;
9174 }
9175
9176 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9177
9178 /*
9179 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9180 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9181 * and anonmap_purge() acquires a_purgemtx.
9182 */
9183 mutex_enter(&->a_purgemtx);
9184 if (!atomic_add_long_nv((ulong_t *)&->a_softlockcnt, -npages) &&
9185 amp->a_purgewait) {
9186 amp->a_purgewait = 0;
9187 cv_broadcast(&->a_purgecv);
9188 }
9189 mutex_exit(&->a_purgemtx);
9190 return (0);
9191 }
9192
9193 /*
9194 * get a memory ID for an addr in a given segment
9195 *
9196 * XXX only creates PAGESIZE pages if anon slots are not initialized.
9197 * At fault time they will be relocated into larger pages.
9198 */
9199 static int
9200 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
9201 {
9202 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9203 struct anon *ap = NULL;
9204 ulong_t anon_index;
9205 struct anon_map *amp;
9206 anon_sync_obj_t cookie;
9207
9208 if (svd->type == MAP_PRIVATE) {
9209 memidp->val[0] = (uintptr_t)seg->s_as;
9210 memidp->val[1] = (uintptr_t)addr;
9211 return (0);
9212 }
9213
9214 if (svd->type == MAP_SHARED) {
9215 if (svd->vp) {
9216 memidp->val[0] = (uintptr_t)svd->vp;
9217 memidp->val[1] = (u_longlong_t)svd->offset +
9218 (uintptr_t)(addr - seg->s_base);
9219 return (0);
9220 } else {
9221
9222 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
9223 if ((amp = svd->amp) != NULL) {
9224 anon_index = svd->anon_index +
9225 seg_page(seg, addr);
9226 }
9227 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9228
9229 ASSERT(amp != NULL);
9230
9231 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
9232 anon_array_enter(amp, anon_index, &cookie);
9233 ap = anon_get_ptr(amp->ahp, anon_index);
9234 if (ap == NULL) {
9235 page_t *pp;
9236
9237 pp = anon_zero(seg, addr, &ap, svd->cred);
9238 if (pp == NULL) {
9239 anon_array_exit(&cookie);
9240 ANON_LOCK_EXIT(&->a_rwlock);
9241 return (ENOMEM);
9242 }
9243 ASSERT(anon_get_ptr(amp->ahp, anon_index)
9244 == NULL);
9245 (void) anon_set_ptr(amp->ahp, anon_index,
9246 ap, ANON_SLEEP);
9247 page_unlock(pp);
9248 }
9249
9250 anon_array_exit(&cookie);
9251 ANON_LOCK_EXIT(&->a_rwlock);
9252
9253 memidp->val[0] = (uintptr_t)ap;
9254 memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
9255 return (0);
9256 }
9257 }
9258 return (EINVAL);
9259 }
9260
9261 static int
9262 sameprot(struct seg *seg, caddr_t a, size_t len)
9263 {
9264 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9265 struct vpage *vpage;
9266 spgcnt_t pages = btop(len);
9267 uint_t prot;
9268
9269 if (svd->pageprot == 0)
9270 return (1);
9271
9272 ASSERT(svd->vpage != NULL);
9273
9274 vpage = &svd->vpage[seg_page(seg, a)];
9275 prot = VPP_PROT(vpage);
9276 vpage++;
9277 pages--;
9278 while (pages-- > 0) {
9279 if (prot != VPP_PROT(vpage))
9280 return (0);
9281 vpage++;
9282 }
9283 return (1);
9284 }
9285
9286 /*
9287 * Get memory allocation policy info for specified address in given segment
9288 */
9289 static lgrp_mem_policy_info_t *
9290 segvn_getpolicy(struct seg *seg, caddr_t addr)
9291 {
9292 struct anon_map *amp;
9293 ulong_t anon_index;
9294 lgrp_mem_policy_info_t *policy_info;
9295 struct segvn_data *svn_data;
9296 u_offset_t vn_off;
9297 vnode_t *vp;
9298
9299 ASSERT(seg != NULL);
9300
9301 svn_data = (struct segvn_data *)seg->s_data;
9302 if (svn_data == NULL)
9303 return (NULL);
9304
9305 /*
9306 * Get policy info for private or shared memory
9307 */
9308 if (svn_data->type != MAP_SHARED) {
9309 if (svn_data->tr_state != SEGVN_TR_ON) {
9310 policy_info = &svn_data->policy_info;
9311 } else {
9312 policy_info = &svn_data->tr_policy_info;
9313 ASSERT(policy_info->mem_policy ==
9314 LGRP_MEM_POLICY_NEXT_SEG);
9315 }
9316 } else {
9317 amp = svn_data->amp;
9318 anon_index = svn_data->anon_index + seg_page(seg, addr);
9319 vp = svn_data->vp;
9320 vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
9321 policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
9322 }
9323
9324 return (policy_info);
9325 }
9326
9327 /*ARGSUSED*/
9328 static int
9329 segvn_capable(struct seg *seg, segcapability_t capability)
9330 {
9331 return (0);
9332 }
9333
9334 /*
9335 * Bind text vnode segment to an amp. If we bind successfully mappings will be
9336 * established to per vnode mapping per lgroup amp pages instead of to vnode
9337 * pages. There's one amp per vnode text mapping per lgroup. Many processes
9338 * may share the same text replication amp. If a suitable amp doesn't already
9339 * exist in svntr hash table create a new one. We may fail to bind to amp if
9340 * segment is not eligible for text replication. Code below first checks for
9341 * these conditions. If binding is successful segment tr_state is set to on
9342 * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9343 * svd->amp remains as NULL.
9344 */
9345 static void
9346 segvn_textrepl(struct seg *seg)
9347 {
9348 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9349 vnode_t *vp = svd->vp;
9350 u_offset_t off = svd->offset;
9351 size_t size = seg->s_size;
9352 u_offset_t eoff = off + size;
9353 uint_t szc = seg->s_szc;
9354 ulong_t hash = SVNTR_HASH_FUNC(vp);
9355 svntr_t *svntrp;
9356 struct vattr va;
9357 proc_t *p = seg->s_as->a_proc;
9358 lgrp_id_t lgrp_id;
9359 lgrp_id_t olid;
9360 int first;
9361 struct anon_map *amp;
9362
9363 ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9364 ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9365 ASSERT(p != NULL);
9366 ASSERT(svd->tr_state == SEGVN_TR_INIT);
9367 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9368 ASSERT(svd->flags & MAP_TEXT);
9369 ASSERT(svd->type == MAP_PRIVATE);
9370 ASSERT(vp != NULL && svd->amp == NULL);
9371 ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
9372 ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
9373 ASSERT(seg->s_as != &kas);
9374 ASSERT(off < eoff);
9375 ASSERT(svntr_hashtab != NULL);
9376
9377 /*
9378 * If numa optimizations are no longer desired bail out.
9379 */
9380 if (!lgrp_optimizations()) {
9381 svd->tr_state = SEGVN_TR_OFF;
9382 return;
9383 }
9384
9385 /*
9386 * Avoid creating anon maps with size bigger than the file size.
9387 * If VOP_GETATTR() call fails bail out.
9388 */
9389 va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9390 if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL) != 0) {
9391 svd->tr_state = SEGVN_TR_OFF;
9392 SEGVN_TR_ADDSTAT(gaerr);
9393 return;
9394 }
9395 if (btopr(va.va_size) < btopr(eoff)) {
9396 svd->tr_state = SEGVN_TR_OFF;
9397 SEGVN_TR_ADDSTAT(overmap);
9398 return;
9399 }
9400
9401 /*
9402 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9403 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9404 * mapping that checks if trcache for this vnode needs to be
9405 * invalidated can't miss us.
9406 */
9407 if (!(vp->v_flag & VVMEXEC)) {
9408 mutex_enter(&vp->v_lock);
9409 vp->v_flag |= VVMEXEC;
9410 mutex_exit(&vp->v_lock);
9411 }
9412 mutex_enter(&svntr_hashtab[hash].tr_lock);
9413 /*
9414 * Bail out if potentially MAP_SHARED writable mappings exist to this
9415 * vnode. We don't want to use old file contents from existing
9416 * replicas if this mapping was established after the original file
9417 * was changed.
9418 */
9419 if (vn_is_mapped(vp, V_WRITE)) {
9420 mutex_exit(&svntr_hashtab[hash].tr_lock);
9421 svd->tr_state = SEGVN_TR_OFF;
9422 SEGVN_TR_ADDSTAT(wrcnt);
9423 return;
9424 }
9425 svntrp = svntr_hashtab[hash].tr_head;
9426 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9427 ASSERT(svntrp->tr_refcnt != 0);
9428 if (svntrp->tr_vp != vp) {
9429 continue;
9430 }
9431
9432 /*
9433 * Bail out if the file or its attributes were changed after
9434 * this replication entry was created since we need to use the
9435 * latest file contents. Note that mtime test alone is not
9436 * sufficient because a user can explicitly change mtime via
9437 * utimes(2) interfaces back to the old value after modifiying
9438 * the file contents. To detect this case we also have to test
9439 * ctime which among other things records the time of the last
9440 * mtime change by utimes(2). ctime is not changed when the file
9441 * is only read or executed so we expect that typically existing
9442 * replication amp's can be used most of the time.
9443 */
9444 if (!svntrp->tr_valid ||
9445 svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9446 svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9447 svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9448 svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9449 mutex_exit(&svntr_hashtab[hash].tr_lock);
9450 svd->tr_state = SEGVN_TR_OFF;
9451 SEGVN_TR_ADDSTAT(stale);
9452 return;
9453 }
9454 /*
9455 * if off, eoff and szc match current segment we found the
9456 * existing entry we can use.
9457 */
9458 if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9459 svntrp->tr_szc == szc) {
9460 break;
9461 }
9462 /*
9463 * Don't create different but overlapping in file offsets
9464 * entries to avoid replication of the same file pages more
9465 * than once per lgroup.
9466 */
9467 if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9468 (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9469 mutex_exit(&svntr_hashtab[hash].tr_lock);
9470 svd->tr_state = SEGVN_TR_OFF;
9471 SEGVN_TR_ADDSTAT(overlap);
9472 return;
9473 }
9474 }
9475 /*
9476 * If we didn't find existing entry create a new one.
9477 */
9478 if (svntrp == NULL) {
9479 svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9480 if (svntrp == NULL) {
9481 mutex_exit(&svntr_hashtab[hash].tr_lock);
9482 svd->tr_state = SEGVN_TR_OFF;
9483 SEGVN_TR_ADDSTAT(nokmem);
9484 return;
9485 }
9486 #ifdef DEBUG
9487 {
9488 lgrp_id_t i;
9489 for (i = 0; i < NLGRPS_MAX; i++) {
9490 ASSERT(svntrp->tr_amp[i] == NULL);
9491 }
9492 }
9493 #endif /* DEBUG */
9494 svntrp->tr_vp = vp;
9495 svntrp->tr_off = off;
9496 svntrp->tr_eoff = eoff;
9497 svntrp->tr_szc = szc;
9498 svntrp->tr_valid = 1;
9499 svntrp->tr_mtime = va.va_mtime;
9500 svntrp->tr_ctime = va.va_ctime;
9501 svntrp->tr_refcnt = 0;
9502 svntrp->tr_next = svntr_hashtab[hash].tr_head;
9503 svntr_hashtab[hash].tr_head = svntrp;
9504 }
9505 first = 1;
9506 again:
9507 /*
9508 * We want to pick a replica with pages on main thread's (t_tid = 1,
9509 * aka T1) lgrp. Currently text replication is only optimized for
9510 * workloads that either have all threads of a process on the same
9511 * lgrp or execute their large text primarily on main thread.
9512 */
9513 lgrp_id = p->p_t1_lgrpid;
9514 if (lgrp_id == LGRP_NONE) {
9515 /*
9516 * In case exec() prefaults text on non main thread use
9517 * current thread lgrpid. It will become main thread anyway
9518 * soon.
9519 */
9520 lgrp_id = lgrp_home_id(curthread);
9521 }
9522 /*
9523 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet. Otherwise
9524 * just set it to NLGRPS_MAX if it's different from current process T1
9525 * home lgrp. p_tr_lgrpid is used to detect if process uses text
9526 * replication and T1 new home is different from lgrp used for text
9527 * replication. When this happens asyncronous segvn thread rechecks if
9528 * segments should change lgrps used for text replication. If we fail
9529 * to set p_tr_lgrpid with cas32 then set it to NLGRPS_MAX without cas
9530 * if it's not already NLGRPS_MAX and not equal lgrp_id we want to
9531 * use. We don't need to use cas in this case because another thread
9532 * that races in between our non atomic check and set may only change
9533 * p_tr_lgrpid to NLGRPS_MAX at this point.
9534 */
9535 ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9536 olid = p->p_tr_lgrpid;
9537 if (lgrp_id != olid && olid != NLGRPS_MAX) {
9538 lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9539 if (cas32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) != olid) {
9540 olid = p->p_tr_lgrpid;
9541 ASSERT(olid != LGRP_NONE);
9542 if (olid != lgrp_id && olid != NLGRPS_MAX) {
9543 p->p_tr_lgrpid = NLGRPS_MAX;
9544 }
9545 }
9546 ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9547 membar_producer();
9548 /*
9549 * lgrp_move_thread() won't schedule async recheck after
9550 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9551 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9552 * is not LGRP_NONE.
9553 */
9554 if (first && p->p_t1_lgrpid != LGRP_NONE &&
9555 p->p_t1_lgrpid != lgrp_id) {
9556 first = 0;
9557 goto again;
9558 }
9559 }
9560 /*
9561 * If no amp was created yet for lgrp_id create a new one as long as
9562 * we have enough memory to afford it.
9563 */
9564 if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9565 size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9566 if (trmem > segvn_textrepl_max_bytes) {
9567 SEGVN_TR_ADDSTAT(normem);
9568 goto fail;
9569 }
9570 if (anon_try_resv_zone(size, NULL) == 0) {
9571 SEGVN_TR_ADDSTAT(noanon);
9572 goto fail;
9573 }
9574 amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9575 if (amp == NULL) {
9576 anon_unresv_zone(size, NULL);
9577 SEGVN_TR_ADDSTAT(nokmem);
9578 goto fail;
9579 }
9580 ASSERT(amp->refcnt == 1);
9581 amp->a_szc = szc;
9582 svntrp->tr_amp[lgrp_id] = amp;
9583 SEGVN_TR_ADDSTAT(newamp);
9584 }
9585 svntrp->tr_refcnt++;
9586 ASSERT(svd->svn_trnext == NULL);
9587 ASSERT(svd->svn_trprev == NULL);
9588 svd->svn_trnext = svntrp->tr_svnhead;
9589 svd->svn_trprev = NULL;
9590 if (svntrp->tr_svnhead != NULL) {
9591 svntrp->tr_svnhead->svn_trprev = svd;
9592 }
9593 svntrp->tr_svnhead = svd;
9594 ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9595 ASSERT(amp->refcnt >= 1);
9596 svd->amp = amp;
9597 svd->anon_index = 0;
9598 svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9599 svd->tr_policy_info.mem_lgrpid = lgrp_id;
9600 svd->tr_state = SEGVN_TR_ON;
9601 mutex_exit(&svntr_hashtab[hash].tr_lock);
9602 SEGVN_TR_ADDSTAT(repl);
9603 return;
9604 fail:
9605 ASSERT(segvn_textrepl_bytes >= size);
9606 atomic_add_long(&segvn_textrepl_bytes, -size);
9607 ASSERT(svntrp != NULL);
9608 ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9609 if (svntrp->tr_refcnt == 0) {
9610 ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9611 svntr_hashtab[hash].tr_head = svntrp->tr_next;
9612 mutex_exit(&svntr_hashtab[hash].tr_lock);
9613 kmem_cache_free(svntr_cache, svntrp);
9614 } else {
9615 mutex_exit(&svntr_hashtab[hash].tr_lock);
9616 }
9617 svd->tr_state = SEGVN_TR_OFF;
9618 }
9619
9620 /*
9621 * Convert seg back to regular vnode mapping seg by unbinding it from its text
9622 * replication amp. This routine is most typically called when segment is
9623 * unmapped but can also be called when segment no longer qualifies for text
9624 * replication (e.g. due to protection changes). If unload_unmap is set use
9625 * HAT_UNLOAD_UNMAP flag in hat_unload_callback(). If we are the last user of
9626 * svntr free all its anon maps and remove it from the hash table.
9627 */
9628 static void
9629 segvn_textunrepl(struct seg *seg, int unload_unmap)
9630 {
9631 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9632 vnode_t *vp = svd->vp;
9633 u_offset_t off = svd->offset;
9634 size_t size = seg->s_size;
9635 u_offset_t eoff = off + size;
9636 uint_t szc = seg->s_szc;
9637 ulong_t hash = SVNTR_HASH_FUNC(vp);
9638 svntr_t *svntrp;
9639 svntr_t **prv_svntrp;
9640 lgrp_id_t lgrp_id = svd->tr_policy_info.mem_lgrpid;
9641 lgrp_id_t i;
9642
9643 ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9644 ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
9645 SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9646 ASSERT(svd->tr_state == SEGVN_TR_ON);
9647 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9648 ASSERT(svd->amp != NULL);
9649 ASSERT(svd->amp->refcnt >= 1);
9650 ASSERT(svd->anon_index == 0);
9651 ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9652 ASSERT(svntr_hashtab != NULL);
9653
9654 mutex_enter(&svntr_hashtab[hash].tr_lock);
9655 prv_svntrp = &svntr_hashtab[hash].tr_head;
9656 for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9657 ASSERT(svntrp->tr_refcnt != 0);
9658 if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9659 svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9660 break;
9661 }
9662 }
9663 if (svntrp == NULL) {
9664 panic("segvn_textunrepl: svntr record not found");
9665 }
9666 if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9667 panic("segvn_textunrepl: amp mismatch");
9668 }
9669 svd->tr_state = SEGVN_TR_OFF;
9670 svd->amp = NULL;
9671 if (svd->svn_trprev == NULL) {
9672 ASSERT(svntrp->tr_svnhead == svd);
9673 svntrp->tr_svnhead = svd->svn_trnext;
9674 if (svntrp->tr_svnhead != NULL) {
9675 svntrp->tr_svnhead->svn_trprev = NULL;
9676 }
9677 svd->svn_trnext = NULL;
9678 } else {
9679 svd->svn_trprev->svn_trnext = svd->svn_trnext;
9680 if (svd->svn_trnext != NULL) {
9681 svd->svn_trnext->svn_trprev = svd->svn_trprev;
9682 svd->svn_trnext = NULL;
9683 }
9684 svd->svn_trprev = NULL;
9685 }
9686 if (--svntrp->tr_refcnt) {
9687 mutex_exit(&svntr_hashtab[hash].tr_lock);
9688 goto done;
9689 }
9690 *prv_svntrp = svntrp->tr_next;
9691 mutex_exit(&svntr_hashtab[hash].tr_lock);
9692 for (i = 0; i < NLGRPS_MAX; i++) {
9693 struct anon_map *amp = svntrp->tr_amp[i];
9694 if (amp == NULL) {
9695 continue;
9696 }
9697 ASSERT(amp->refcnt == 1);
9698 ASSERT(amp->swresv == size);
9699 ASSERT(amp->size == size);
9700 ASSERT(amp->a_szc == szc);
9701 if (amp->a_szc != 0) {
9702 anon_free_pages(amp->ahp, 0, size, szc);
9703 } else {
9704 anon_free(amp->ahp, 0, size);
9705 }
9706 svntrp->tr_amp[i] = NULL;
9707 ASSERT(segvn_textrepl_bytes >= size);
9708 atomic_add_long(&segvn_textrepl_bytes, -size);
9709 anon_unresv_zone(amp->swresv, NULL);
9710 amp->refcnt = 0;
9711 anonmap_free(amp);
9712 }
9713 kmem_cache_free(svntr_cache, svntrp);
9714 done:
9715 hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9716 unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9717 }
9718
9719 /*
9720 * This is called when a MAP_SHARED writable mapping is created to a vnode
9721 * that is currently used for execution (VVMEXEC flag is set). In this case we
9722 * need to prevent further use of existing replicas.
9723 */
9724 static void
9725 segvn_inval_trcache(vnode_t *vp)
9726 {
9727 ulong_t hash = SVNTR_HASH_FUNC(vp);
9728 svntr_t *svntrp;
9729
9730 ASSERT(vp->v_flag & VVMEXEC);
9731
9732 if (svntr_hashtab == NULL) {
9733 return;
9734 }
9735
9736 mutex_enter(&svntr_hashtab[hash].tr_lock);
9737 svntrp = svntr_hashtab[hash].tr_head;
9738 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9739 ASSERT(svntrp->tr_refcnt != 0);
9740 if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9741 svntrp->tr_valid = 0;
9742 }
9743 }
9744 mutex_exit(&svntr_hashtab[hash].tr_lock);
9745 }
9746
9747 static void
9748 segvn_trasync_thread(void)
9749 {
9750 callb_cpr_t cpr_info;
9751 kmutex_t cpr_lock; /* just for CPR stuff */
9752
9753 mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
9754
9755 CALLB_CPR_INIT(&cpr_info, &cpr_lock,
9756 callb_generic_cpr, "segvn_async");
9757
9758 if (segvn_update_textrepl_interval == 0) {
9759 segvn_update_textrepl_interval = segvn_update_tr_time * hz;
9760 } else {
9761 segvn_update_textrepl_interval *= hz;
9762 }
9763 (void) timeout(segvn_trupdate_wakeup, NULL,
9764 segvn_update_textrepl_interval);
9765
9766 for (;;) {
9767 mutex_enter(&cpr_lock);
9768 CALLB_CPR_SAFE_BEGIN(&cpr_info);
9769 mutex_exit(&cpr_lock);
9770 sema_p(&segvn_trasync_sem);
9771 mutex_enter(&cpr_lock);
9772 CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
9773 mutex_exit(&cpr_lock);
9774 segvn_trupdate();
9775 }
9776 }
9777
9778 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
9779
9780 static void
9781 segvn_trupdate_wakeup(void *dummy)
9782 {
9783 uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
9784
9785 if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
9786 segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
9787 sema_v(&segvn_trasync_sem);
9788 }
9789
9790 if (!segvn_disable_textrepl_update &&
9791 segvn_update_textrepl_interval != 0) {
9792 (void) timeout(segvn_trupdate_wakeup, dummy,
9793 segvn_update_textrepl_interval);
9794 }
9795 }
9796
9797 static void
9798 segvn_trupdate(void)
9799 {
9800 ulong_t hash;
9801 svntr_t *svntrp;
9802 segvn_data_t *svd;
9803
9804 ASSERT(svntr_hashtab != NULL);
9805
9806 for (hash = 0; hash < svntr_hashtab_sz; hash++) {
9807 mutex_enter(&svntr_hashtab[hash].tr_lock);
9808 svntrp = svntr_hashtab[hash].tr_head;
9809 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9810 ASSERT(svntrp->tr_refcnt != 0);
9811 svd = svntrp->tr_svnhead;
9812 for (; svd != NULL; svd = svd->svn_trnext) {
9813 segvn_trupdate_seg(svd->seg, svd, svntrp,
9814 hash);
9815 }
9816 }
9817 mutex_exit(&svntr_hashtab[hash].tr_lock);
9818 }
9819 }
9820
9821 static void
9822 segvn_trupdate_seg(struct seg *seg,
9823 segvn_data_t *svd,
9824 svntr_t *svntrp,
9825 ulong_t hash)
9826 {
9827 proc_t *p;
9828 lgrp_id_t lgrp_id;
9829 struct as *as;
9830 size_t size;
9831 struct anon_map *amp;
9832
9833 ASSERT(svd->vp != NULL);
9834 ASSERT(svd->vp == svntrp->tr_vp);
9835 ASSERT(svd->offset == svntrp->tr_off);
9836 ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
9837 ASSERT(seg != NULL);
9838 ASSERT(svd->seg == seg);
9839 ASSERT(seg->s_data == (void *)svd);
9840 ASSERT(seg->s_szc == svntrp->tr_szc);
9841 ASSERT(svd->tr_state == SEGVN_TR_ON);
9842 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9843 ASSERT(svd->amp != NULL);
9844 ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9845 ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
9846 ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
9847 ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
9848 ASSERT(svntrp->tr_refcnt != 0);
9849 ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
9850
9851 as = seg->s_as;
9852 ASSERT(as != NULL && as != &kas);
9853 p = as->a_proc;
9854 ASSERT(p != NULL);
9855 ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9856 lgrp_id = p->p_t1_lgrpid;
9857 if (lgrp_id == LGRP_NONE) {
9858 return;
9859 }
9860 ASSERT(lgrp_id < NLGRPS_MAX);
9861 if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
9862 return;
9863 }
9864
9865 /*
9866 * Use tryenter locking since we are locking as/seg and svntr hash
9867 * lock in reverse from syncrounous thread order.
9868 */
9869 if (!AS_LOCK_TRYENTER(as, &as->a_lock, RW_READER)) {
9870 SEGVN_TR_ADDSTAT(nolock);
9871 if (segvn_lgrp_trthr_migrs_snpsht) {
9872 segvn_lgrp_trthr_migrs_snpsht = 0;
9873 }
9874 return;
9875 }
9876 if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
9877 AS_LOCK_EXIT(as, &as->a_lock);
9878 SEGVN_TR_ADDSTAT(nolock);
9879 if (segvn_lgrp_trthr_migrs_snpsht) {
9880 segvn_lgrp_trthr_migrs_snpsht = 0;
9881 }
9882 return;
9883 }
9884 size = seg->s_size;
9885 if (svntrp->tr_amp[lgrp_id] == NULL) {
9886 size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9887 if (trmem > segvn_textrepl_max_bytes) {
9888 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9889 AS_LOCK_EXIT(as, &as->a_lock);
9890 atomic_add_long(&segvn_textrepl_bytes, -size);
9891 SEGVN_TR_ADDSTAT(normem);
9892 return;
9893 }
9894 if (anon_try_resv_zone(size, NULL) == 0) {
9895 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9896 AS_LOCK_EXIT(as, &as->a_lock);
9897 atomic_add_long(&segvn_textrepl_bytes, -size);
9898 SEGVN_TR_ADDSTAT(noanon);
9899 return;
9900 }
9901 amp = anonmap_alloc(size, size, KM_NOSLEEP);
9902 if (amp == NULL) {
9903 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9904 AS_LOCK_EXIT(as, &as->a_lock);
9905 atomic_add_long(&segvn_textrepl_bytes, -size);
9906 anon_unresv_zone(size, NULL);
9907 SEGVN_TR_ADDSTAT(nokmem);
9908 return;
9909 }
9910 ASSERT(amp->refcnt == 1);
9911 amp->a_szc = seg->s_szc;
9912 svntrp->tr_amp[lgrp_id] = amp;
9913 }
9914 /*
9915 * We don't need to drop the bucket lock but here we give other
9916 * threads a chance. svntr and svd can't be unlinked as long as
9917 * segment lock is held as a writer and AS held as well. After we
9918 * retake bucket lock we'll continue from where we left. We'll be able
9919 * to reach the end of either list since new entries are always added
9920 * to the beginning of the lists.
9921 */
9922 mutex_exit(&svntr_hashtab[hash].tr_lock);
9923 hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
9924 mutex_enter(&svntr_hashtab[hash].tr_lock);
9925
9926 ASSERT(svd->tr_state == SEGVN_TR_ON);
9927 ASSERT(svd->amp != NULL);
9928 ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9929 ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
9930 ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
9931
9932 svd->tr_policy_info.mem_lgrpid = lgrp_id;
9933 svd->amp = svntrp->tr_amp[lgrp_id];
9934 p->p_tr_lgrpid = NLGRPS_MAX;
9935 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9936 AS_LOCK_EXIT(as, &as->a_lock);
9937
9938 ASSERT(svntrp->tr_refcnt != 0);
9939 ASSERT(svd->vp == svntrp->tr_vp);
9940 ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
9941 ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
9942 ASSERT(svd->seg == seg);
9943 ASSERT(svd->tr_state == SEGVN_TR_ON);
9944
9945 SEGVN_TR_ADDSTAT(asyncrepl);
9946 }