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 2006 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25 
  26 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */
  27 /* All Rights Reserved */
  28 
  29 /*
  30  * Portions of this source code were derived from Berkeley 4.3 BSD
  31  * under license from the Regents of the University of California.
  32  */
  33 
  34 /*
  35  * VM - segment for non-faulting loads.
  36  */
  37 
  38 #include <sys/types.h>
  39 #include <sys/t_lock.h>
  40 #include <sys/param.h>
  41 #include <sys/mman.h>
  42 #include <sys/errno.h>
  43 #include <sys/kmem.h>
  44 #include <sys/cmn_err.h>
  45 #include <sys/vnode.h>
  46 #include <sys/proc.h>
  47 #include <sys/conf.h>
  48 #include <sys/debug.h>
  49 #include <sys/archsystm.h>
  50 #include <sys/lgrp.h>
  51 
  52 #include <vm/page.h>
  53 #include <vm/hat.h>
  54 #include <vm/as.h>
  55 #include <vm/seg.h>
  56 #include <vm/vpage.h>
  57 
  58 /*
  59  * Private seg op routines.
  60  */
  61 static int      segnf_dup(struct seg *seg, struct seg *newseg);
  62 static int      segnf_unmap(struct seg *seg, caddr_t addr, size_t len);
  63 static void     segnf_free(struct seg *seg);
  64 static faultcode_t segnf_nomap(void);
  65 static int      segnf_setprot(struct seg *seg, caddr_t addr,
  66                     size_t len, uint_t prot);
  67 static int      segnf_checkprot(struct seg *seg, caddr_t addr,
  68                     size_t len, uint_t prot);
  69 static void     segnf_badop(void);
  70 static int      segnf_nop(void);
  71 static int      segnf_getprot(struct seg *seg, caddr_t addr,
  72                     size_t len, uint_t *protv);
  73 static u_offset_t segnf_getoffset(struct seg *seg, caddr_t addr);
  74 static int      segnf_gettype(struct seg *seg, caddr_t addr);
  75 static int      segnf_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp);
  76 static void     segnf_dump(struct seg *seg);
  77 static int      segnf_pagelock(struct seg *seg, caddr_t addr, size_t len,
  78                     struct page ***ppp, enum lock_type type, enum seg_rw rw);
  79 static int      segnf_setpagesize(struct seg *seg, caddr_t addr, size_t len,
  80                     uint_t szc);
  81 static int      segnf_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp);
  82 static lgrp_mem_policy_info_t   *segnf_getpolicy(struct seg *seg,
  83     caddr_t addr);
  84 
  85 
  86 struct seg_ops segnf_ops = {
  87         .dup            = segnf_dup,
  88         .unmap          = segnf_unmap,
  89         .free           = segnf_free,
  90         .fault          = (faultcode_t (*)(struct hat *, struct seg *, caddr_t,
  91             size_t, enum fault_type, enum seg_rw))segnf_nomap,
  92         .faulta         = (faultcode_t (*)(struct seg *, caddr_t)) segnf_nomap,
  93         .setprot        = segnf_setprot,
  94         .checkprot      = segnf_checkprot,
  95         .kluster        = (int (*)())segnf_badop,
  96         .sync           = (int (*)(struct seg *, caddr_t, size_t, int, uint_t))
  97                 segnf_nop,
  98         .incore         = (size_t (*)(struct seg *, caddr_t, size_t, char *))
  99                 segnf_nop,
 100         .lockop         = (int (*)(struct seg *, caddr_t, size_t, int, int,
 101             ulong_t *, size_t))segnf_nop,
 102         .getprot        = segnf_getprot,
 103         .getoffset      = segnf_getoffset,
 104         .gettype        = segnf_gettype,
 105         .getvp          = segnf_getvp,
 106         .advise         = (int (*)(struct seg *, caddr_t, size_t, uint_t))
 107                 segnf_nop,
 108         .dump           = segnf_dump,
 109         .pagelock       = segnf_pagelock,
 110         .setpagesize    = segnf_setpagesize,
 111         .getmemid       = segnf_getmemid,
 112         .getpolicy      = segnf_getpolicy,
 113 };
 114 
 115 /*
 116  * vnode and page for the page of zeros we use for the nf mappings.
 117  */
 118 static kmutex_t segnf_lock;
 119 static struct vnode nfvp;
 120 static struct page **nfpp;
 121 
 122 #define addr_to_vcolor(addr)                                            \
 123         (shm_alignment) ?                                               \
 124         ((int)(((uintptr_t)(addr) & (shm_alignment - 1)) >> PAGESHIFT)) : 0
 125 
 126 /*
 127  * We try to limit the number of Non-fault segments created.
 128  * Non fault segments are created to optimize sparc V9 code which uses
 129  * the sparc nonfaulting load ASI (ASI_PRIMARY_NOFAULT).
 130  *
 131  * There are several reasons why creating too many non-fault segments
 132  * could cause problems.
 133  *
 134  *      First, excessive allocation of kernel resources for the seg
 135  *      structures and the HAT data to map the zero pages.
 136  *
 137  *      Secondly, creating nofault segments actually uses up user virtual
 138  *      address space. This makes it unavailable for subsequent mmap(0, ...)
 139  *      calls which use as_gap() to find empty va regions.  Creation of too
 140  *      many nofault segments could thus interfere with the ability of the
 141  *      runtime linker to load a shared object.
 142  */
 143 #define MAXSEGFORNF     (10000)
 144 #define MAXNFSEARCH     (5)
 145 
 146 
 147 /*
 148  * Must be called from startup()
 149  */
 150 void
 151 segnf_init()
 152 {
 153         mutex_init(&segnf_lock, NULL, MUTEX_DEFAULT, NULL);
 154 }
 155 
 156 
 157 /*
 158  * Create a no-fault segment.
 159  *
 160  * The no-fault segment is not technically necessary, as the code in
 161  * nfload() in trap.c will emulate the SPARC instruction and load
 162  * a value of zero in the destination register.
 163  *
 164  * However, this code tries to put a page of zero's at the nofault address
 165  * so that subsequent non-faulting loads to the same page will not
 166  * trap with a tlb miss.
 167  *
 168  * In order to help limit the number of segments we merge adjacent nofault
 169  * segments into a single segment.  If we get a large number of segments
 170  * we'll also try to delete a random other nf segment.
 171  */
 172 /* ARGSUSED */
 173 int
 174 segnf_create(struct seg *seg, void *argsp)
 175 {
 176         uint_t prot;
 177         pgcnt_t vacpgs;
 178         u_offset_t off = 0;
 179         caddr_t vaddr = NULL;
 180         int i, color;
 181         struct seg *s1;
 182         struct seg *s2;
 183         size_t size;
 184         struct as *as = seg->s_as;
 185 
 186         ASSERT(as && AS_WRITE_HELD(as, &as->a_lock));
 187 
 188         /*
 189          * Need a page per virtual color or just 1 if no vac.
 190          */
 191         mutex_enter(&segnf_lock);
 192         if (nfpp == NULL) {
 193                 struct seg kseg;
 194 
 195                 vacpgs = 1;
 196                 if (shm_alignment > PAGESIZE) {
 197                         vacpgs = shm_alignment >> PAGESHIFT;
 198                 }
 199 
 200                 nfpp = kmem_alloc(sizeof (*nfpp) * vacpgs, KM_SLEEP);
 201 
 202                 kseg.s_as = &kas;
 203                 for (i = 0; i < vacpgs; i++, off += PAGESIZE,
 204                     vaddr += PAGESIZE) {
 205                         nfpp[i] = page_create_va(&nfvp, off, PAGESIZE,
 206                             PG_WAIT | PG_NORELOC, &kseg, vaddr);
 207                         page_io_unlock(nfpp[i]);
 208                         page_downgrade(nfpp[i]);
 209                         pagezero(nfpp[i], 0, PAGESIZE);
 210                 }
 211         }
 212         mutex_exit(&segnf_lock);
 213 
 214         hat_map(as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
 215 
 216         /*
 217          * s_data can't be NULL because of ASSERTS in the common vm code.
 218          */
 219         seg->s_ops = &segnf_ops;
 220         seg->s_data = seg;
 221         seg->s_flags |= S_PURGE;
 222 
 223         mutex_enter(&as->a_contents);
 224         as->a_flags |= AS_NEEDSPURGE;
 225         mutex_exit(&as->a_contents);
 226 
 227         prot = PROT_READ;
 228         color = addr_to_vcolor(seg->s_base);
 229         if (as != &kas)
 230                 prot |= PROT_USER;
 231         hat_memload(as->a_hat, seg->s_base, nfpp[color],
 232             prot | HAT_NOFAULT, HAT_LOAD);
 233 
 234         /*
 235          * At this point see if we can concatenate a segment to
 236          * a non-fault segment immediately before and/or after it.
 237          */
 238         if ((s1 = AS_SEGPREV(as, seg)) != NULL &&
 239             s1->s_ops == &segnf_ops &&
 240             s1->s_base + s1->s_size == seg->s_base) {
 241                 size = s1->s_size;
 242                 seg_free(s1);
 243                 seg->s_base -= size;
 244                 seg->s_size += size;
 245         }
 246 
 247         if ((s2 = AS_SEGNEXT(as, seg)) != NULL &&
 248             s2->s_ops == &segnf_ops &&
 249             seg->s_base + seg->s_size == s2->s_base) {
 250                 size = s2->s_size;
 251                 seg_free(s2);
 252                 seg->s_size += size;
 253         }
 254 
 255         /*
 256          * if we already have a lot of segments, try to delete some other
 257          * nofault segment to reduce the probability of uncontrolled segment
 258          * creation.
 259          *
 260          * the code looks around quickly (no more than MAXNFSEARCH segments
 261          * each way) for another NF segment and then deletes it.
 262          */
 263         if (avl_numnodes(&as->a_segtree) > MAXSEGFORNF) {
 264                 size = 0;
 265                 s2 = NULL;
 266                 s1 = AS_SEGPREV(as, seg);
 267                 while (size++ < MAXNFSEARCH && s1 != NULL) {
 268                         if (s1->s_ops == &segnf_ops)
 269                                 s2 = s1;
 270                         s1 = AS_SEGPREV(s1->s_as, seg);
 271                 }
 272                 if (s2 == NULL) {
 273                         s1 = AS_SEGNEXT(as, seg);
 274                         while (size-- > 0 && s1 != NULL) {
 275                                 if (s1->s_ops == &segnf_ops)
 276                                         s2 = s1;
 277                                 s1 = AS_SEGNEXT(as, seg);
 278                         }
 279                 }
 280                 if (s2 != NULL)
 281                         seg_unmap(s2);
 282         }
 283 
 284         return (0);
 285 }
 286 
 287 /*
 288  * Never really need "No fault" segments, so they aren't dup'd.
 289  */
 290 /* ARGSUSED */
 291 static int
 292 segnf_dup(struct seg *seg, struct seg *newseg)
 293 {
 294         panic("segnf_dup");
 295         return (0);
 296 }
 297 
 298 /*
 299  * Split a segment at addr for length len.
 300  */
 301 static int
 302 segnf_unmap(struct seg *seg, caddr_t addr, size_t len)
 303 {
 304         ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
 305 
 306         /*
 307          * Check for bad sizes.
 308          */
 309         if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
 310             (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
 311                 cmn_err(CE_PANIC, "segnf_unmap: bad unmap size");
 312         }
 313 
 314         /*
 315          * Unload any hardware translations in the range to be taken out.
 316          */
 317         hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD_UNMAP);
 318 
 319         if (addr == seg->s_base && len == seg->s_size) {
 320                 /*
 321                  * Freeing entire segment.
 322                  */
 323                 seg_free(seg);
 324         } else if (addr == seg->s_base) {
 325                 /*
 326                  * Freeing the beginning of the segment.
 327                  */
 328                 seg->s_base += len;
 329                 seg->s_size -= len;
 330         } else if (addr + len == seg->s_base + seg->s_size) {
 331                 /*
 332                  * Freeing the end of the segment.
 333                  */
 334                 seg->s_size -= len;
 335         } else {
 336                 /*
 337                  * The section to go is in the middle of the segment, so we
 338                  * have to cut it into two segments.  We shrink the existing
 339                  * "seg" at the low end, and create "nseg" for the high end.
 340                  */
 341                 caddr_t nbase = addr + len;
 342                 size_t nsize = (seg->s_base + seg->s_size) - nbase;
 343                 struct seg *nseg;
 344 
 345                 /*
 346                  * Trim down "seg" before trying to stick "nseg" into the as.
 347                  */
 348                 seg->s_size = addr - seg->s_base;
 349                 nseg = seg_alloc(seg->s_as, nbase, nsize);
 350                 if (nseg == NULL)
 351                         cmn_err(CE_PANIC, "segnf_unmap: seg_alloc failed");
 352 
 353                 /*
 354                  * s_data can't be NULL because of ASSERTs in common VM code.
 355                  */
 356                 nseg->s_ops = seg->s_ops;
 357                 nseg->s_data = nseg;
 358                 nseg->s_flags |= S_PURGE;
 359                 mutex_enter(&seg->s_as->a_contents);
 360                 seg->s_as->a_flags |= AS_NEEDSPURGE;
 361                 mutex_exit(&seg->s_as->a_contents);
 362         }
 363 
 364         return (0);
 365 }
 366 
 367 /*
 368  * Free a segment.
 369  */
 370 static void
 371 segnf_free(struct seg *seg)
 372 {
 373         ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
 374 }
 375 
 376 /*
 377  * No faults allowed on segnf.
 378  */
 379 static faultcode_t
 380 segnf_nomap(void)
 381 {
 382         return (FC_NOMAP);
 383 }
 384 
 385 /* ARGSUSED */
 386 static int
 387 segnf_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
 388 {
 389         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 390         return (EACCES);
 391 }
 392 
 393 /* ARGSUSED */
 394 static int
 395 segnf_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
 396 {
 397         uint_t sprot;
 398         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 399 
 400         sprot = seg->s_as == &kas ?  PROT_READ : PROT_READ|PROT_USER;
 401         return ((prot & sprot) == prot ? 0 : EACCES);
 402 }
 403 
 404 static void
 405 segnf_badop(void)
 406 {
 407         panic("segnf_badop");
 408         /*NOTREACHED*/
 409 }
 410 
 411 static int
 412 segnf_nop(void)
 413 {
 414         return (0);
 415 }
 416 
 417 static int
 418 segnf_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
 419 {
 420         size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
 421         size_t p;
 422         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 423 
 424         for (p = 0; p < pgno; ++p)
 425                 protv[p] = PROT_READ;
 426         return (0);
 427 }
 428 
 429 /* ARGSUSED */
 430 static u_offset_t
 431 segnf_getoffset(struct seg *seg, caddr_t addr)
 432 {
 433         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 434 
 435         return ((u_offset_t)0);
 436 }
 437 
 438 /* ARGSUSED */
 439 static int
 440 segnf_gettype(struct seg *seg, caddr_t addr)
 441 {
 442         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 443 
 444         return (MAP_SHARED);
 445 }
 446 
 447 /* ARGSUSED */
 448 static int
 449 segnf_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
 450 {
 451         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 452 
 453         *vpp = &nfvp;
 454         return (0);
 455 }
 456 
 457 /*
 458  * segnf pages are not dumped, so we just return
 459  */
 460 /* ARGSUSED */
 461 static void
 462 segnf_dump(struct seg *seg)
 463 {}
 464 
 465 /*ARGSUSED*/
 466 static int
 467 segnf_pagelock(struct seg *seg, caddr_t addr, size_t len,
 468     struct page ***ppp, enum lock_type type, enum seg_rw rw)
 469 {
 470         return (ENOTSUP);
 471 }
 472 
 473 /*ARGSUSED*/
 474 static int
 475 segnf_setpagesize(struct seg *seg, caddr_t addr, size_t len,
 476     uint_t szc)
 477 {
 478         return (ENOTSUP);
 479 }
 480 
 481 /*ARGSUSED*/
 482 static int
 483 segnf_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
 484 {
 485         return (ENODEV);
 486 }
 487 
 488 /*ARGSUSED*/
 489 static lgrp_mem_policy_info_t *
 490 segnf_getpolicy(struct seg *seg, caddr_t addr)
 491 {
 492         return (NULL);
 493 }