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 2010 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */ 27 /* All Rights Reserved */ 28 29 /* 30 * University Copyright- Copyright (c) 1982, 1986, 1988 31 * The Regents of the University of California 32 * All Rights Reserved 33 * 34 * University Acknowledgment- Portions of this document are derived from 35 * software developed by the University of California, Berkeley, and its 36 * contributors. 37 */ 38 39 /* 40 * VM - address spaces. 41 */ 42 43 #include <sys/types.h> 44 #include <sys/t_lock.h> 45 #include <sys/param.h> 46 #include <sys/errno.h> 47 #include <sys/systm.h> 48 #include <sys/mman.h> 49 #include <sys/sysmacros.h> 50 #include <sys/cpuvar.h> 51 #include <sys/sysinfo.h> 52 #include <sys/kmem.h> 53 #include <sys/vnode.h> 54 #include <sys/vmsystm.h> 55 #include <sys/cmn_err.h> 56 #include <sys/debug.h> 57 #include <sys/tnf_probe.h> 58 #include <sys/vtrace.h> 59 60 #include <vm/hat.h> 61 #include <vm/xhat.h> 62 #include <vm/as.h> 63 #include <vm/seg.h> 64 #include <vm/seg_vn.h> 65 #include <vm/seg_dev.h> 66 #include <vm/seg_kmem.h> 67 #include <vm/seg_map.h> 68 #include <vm/seg_spt.h> 69 #include <vm/page.h> 70 71 clock_t deadlk_wait = 1; /* number of ticks to wait before retrying */ 72 73 static struct kmem_cache *as_cache; 74 75 static void as_setwatchprot(struct as *, caddr_t, size_t, uint_t); 76 static void as_clearwatchprot(struct as *, caddr_t, size_t); 77 int as_map_locked(struct as *, caddr_t, size_t, int ((*)()), void *); 78 79 80 /* 81 * Verifying the segment lists is very time-consuming; it may not be 82 * desirable always to define VERIFY_SEGLIST when DEBUG is set. 83 */ 84 #ifdef DEBUG 85 #define VERIFY_SEGLIST 86 int do_as_verify = 0; 87 #endif 88 89 /* 90 * Allocate a new callback data structure entry and fill in the events of 91 * interest, the address range of interest, and the callback argument. 92 * Link the entry on the as->a_callbacks list. A callback entry for the 93 * entire address space may be specified with vaddr = 0 and size = -1. 94 * 95 * CALLERS RESPONSIBILITY: If not calling from within the process context for 96 * the specified as, the caller must guarantee persistence of the specified as 97 * for the duration of this function (eg. pages being locked within the as 98 * will guarantee persistence). 99 */ 100 int 101 as_add_callback(struct as *as, void (*cb_func)(), void *arg, uint_t events, 102 caddr_t vaddr, size_t size, int sleepflag) 103 { 104 struct as_callback *current_head, *cb; 105 caddr_t saddr; 106 size_t rsize; 107 108 /* callback function and an event are mandatory */ 109 if ((cb_func == NULL) || ((events & AS_ALL_EVENT) == 0)) 110 return (EINVAL); 111 112 /* Adding a callback after as_free has been called is not allowed */ 113 if (as == &kas) 114 return (ENOMEM); 115 116 /* 117 * vaddr = 0 and size = -1 is used to indicate that the callback range 118 * is the entire address space so no rounding is done in that case. 119 */ 120 if (size != -1) { 121 saddr = (caddr_t)((uintptr_t)vaddr & (uintptr_t)PAGEMASK); 122 rsize = (((size_t)(vaddr + size) + PAGEOFFSET) & PAGEMASK) - 123 (size_t)saddr; 124 /* check for wraparound */ 125 if (saddr + rsize < saddr) 126 return (ENOMEM); 127 } else { 128 if (vaddr != 0) 129 return (EINVAL); 130 saddr = vaddr; 131 rsize = size; 132 } 133 134 /* Allocate and initialize a callback entry */ 135 cb = kmem_zalloc(sizeof (struct as_callback), sleepflag); 136 if (cb == NULL) 137 return (EAGAIN); 138 139 cb->ascb_func = cb_func; 140 cb->ascb_arg = arg; 141 cb->ascb_events = events; 142 cb->ascb_saddr = saddr; 143 cb->ascb_len = rsize; 144 145 /* Add the entry to the list */ 146 mutex_enter(&as->a_contents); 147 current_head = as->a_callbacks; 148 as->a_callbacks = cb; 149 cb->ascb_next = current_head; 150 151 /* 152 * The call to this function may lose in a race with 153 * a pertinent event - eg. a thread does long term memory locking 154 * but before the callback is added another thread executes as_unmap. 155 * A broadcast here resolves that. 156 */ 157 if ((cb->ascb_events & AS_UNMAPWAIT_EVENT) && AS_ISUNMAPWAIT(as)) { 158 AS_CLRUNMAPWAIT(as); 159 cv_broadcast(&as->a_cv); 160 } 161 162 mutex_exit(&as->a_contents); 163 return (0); 164 } 165 166 /* 167 * Search the callback list for an entry which pertains to arg. 168 * 169 * This is called from within the client upon completion of the callback. 170 * RETURN VALUES: 171 * AS_CALLBACK_DELETED (callback entry found and deleted) 172 * AS_CALLBACK_NOTFOUND (no callback entry found - this is ok) 173 * AS_CALLBACK_DELETE_DEFERRED (callback is in process, delete of this 174 * entry will be made in as_do_callbacks) 175 * 176 * If as_delete_callback encounters a matching entry with AS_CALLBACK_CALLED 177 * set, it indicates that as_do_callbacks is processing this entry. The 178 * AS_ALL_EVENT events are cleared in the entry, and a broadcast is made 179 * to unblock as_do_callbacks, in case it is blocked. 180 * 181 * CALLERS RESPONSIBILITY: If not calling from within the process context for 182 * the specified as, the caller must guarantee persistence of the specified as 183 * for the duration of this function (eg. pages being locked within the as 184 * will guarantee persistence). 185 */ 186 uint_t 187 as_delete_callback(struct as *as, void *arg) 188 { 189 struct as_callback **prevcb = &as->a_callbacks; 190 struct as_callback *cb; 191 uint_t rc = AS_CALLBACK_NOTFOUND; 192 193 mutex_enter(&as->a_contents); 194 for (cb = as->a_callbacks; cb; prevcb = &cb->ascb_next, cb = *prevcb) { 195 if (cb->ascb_arg != arg) 196 continue; 197 198 /* 199 * If the events indicate AS_CALLBACK_CALLED, just clear 200 * AS_ALL_EVENT in the events field and wakeup the thread 201 * that may be waiting in as_do_callbacks. as_do_callbacks 202 * will take care of removing this entry from the list. In 203 * that case, return AS_CALLBACK_DELETE_DEFERRED. Otherwise 204 * (AS_CALLBACK_CALLED not set), just remove it from the 205 * list, return the memory and return AS_CALLBACK_DELETED. 206 */ 207 if ((cb->ascb_events & AS_CALLBACK_CALLED) != 0) { 208 /* leave AS_CALLBACK_CALLED */ 209 cb->ascb_events &= ~AS_ALL_EVENT; 210 rc = AS_CALLBACK_DELETE_DEFERRED; 211 cv_broadcast(&as->a_cv); 212 } else { 213 *prevcb = cb->ascb_next; 214 kmem_free(cb, sizeof (struct as_callback)); 215 rc = AS_CALLBACK_DELETED; 216 } 217 break; 218 } 219 mutex_exit(&as->a_contents); 220 return (rc); 221 } 222 223 /* 224 * Searches the as callback list for a matching entry. 225 * Returns a pointer to the first matching callback, or NULL if 226 * nothing is found. 227 * This function never sleeps so it is ok to call it with more 228 * locks held but the (required) a_contents mutex. 229 * 230 * See also comment on as_do_callbacks below. 231 */ 232 static struct as_callback * 233 as_find_callback(struct as *as, uint_t events, caddr_t event_addr, 234 size_t event_len) 235 { 236 struct as_callback *cb; 237 238 ASSERT(MUTEX_HELD(&as->a_contents)); 239 for (cb = as->a_callbacks; cb != NULL; cb = cb->ascb_next) { 240 /* 241 * If the callback has not already been called, then 242 * check if events or address range pertains. An event_len 243 * of zero means do an unconditional callback. 244 */ 245 if (((cb->ascb_events & AS_CALLBACK_CALLED) != 0) || 246 ((event_len != 0) && (((cb->ascb_events & events) == 0) || 247 (event_addr + event_len < cb->ascb_saddr) || 248 (event_addr > (cb->ascb_saddr + cb->ascb_len))))) { 249 continue; 250 } 251 break; 252 } 253 return (cb); 254 } 255 256 /* 257 * Executes a given callback and removes it from the callback list for 258 * this address space. 259 * This function may sleep so the caller must drop all locks except 260 * a_contents before calling this func. 261 * 262 * See also comments on as_do_callbacks below. 263 */ 264 static void 265 as_execute_callback(struct as *as, struct as_callback *cb, 266 uint_t events) 267 { 268 struct as_callback **prevcb; 269 void *cb_arg; 270 271 ASSERT(MUTEX_HELD(&as->a_contents) && (cb->ascb_events & events)); 272 cb->ascb_events |= AS_CALLBACK_CALLED; 273 mutex_exit(&as->a_contents); 274 (*cb->ascb_func)(as, cb->ascb_arg, events); 275 mutex_enter(&as->a_contents); 276 /* 277 * the callback function is required to delete the callback 278 * when the callback function determines it is OK for 279 * this thread to continue. as_delete_callback will clear 280 * the AS_ALL_EVENT in the events field when it is deleted. 281 * If the callback function called as_delete_callback, 282 * events will already be cleared and there will be no blocking. 283 */ 284 while ((cb->ascb_events & events) != 0) { 285 cv_wait(&as->a_cv, &as->a_contents); 286 } 287 /* 288 * This entry needs to be taken off the list. Normally, the 289 * callback func itself does that, but unfortunately the list 290 * may have changed while the callback was running because the 291 * a_contents mutex was dropped and someone else other than the 292 * callback func itself could have called as_delete_callback, 293 * so we have to search to find this entry again. The entry 294 * must have AS_CALLBACK_CALLED, and have the same 'arg'. 295 */ 296 cb_arg = cb->ascb_arg; 297 prevcb = &as->a_callbacks; 298 for (cb = as->a_callbacks; cb != NULL; 299 prevcb = &cb->ascb_next, cb = *prevcb) { 300 if (((cb->ascb_events & AS_CALLBACK_CALLED) == 0) || 301 (cb_arg != cb->ascb_arg)) { 302 continue; 303 } 304 *prevcb = cb->ascb_next; 305 kmem_free(cb, sizeof (struct as_callback)); 306 break; 307 } 308 } 309 310 /* 311 * Check the callback list for a matching event and intersection of 312 * address range. If there is a match invoke the callback. Skip an entry if: 313 * - a callback is already in progress for this entry (AS_CALLBACK_CALLED) 314 * - not event of interest 315 * - not address range of interest 316 * 317 * An event_len of zero indicates a request for an unconditional callback 318 * (regardless of event), only the AS_CALLBACK_CALLED is checked. The 319 * a_contents lock must be dropped before a callback, so only one callback 320 * can be done before returning. Return -1 (true) if a callback was 321 * executed and removed from the list, else return 0 (false). 322 * 323 * The logically separate parts, i.e. finding a matching callback and 324 * executing a given callback have been separated into two functions 325 * so that they can be called with different sets of locks held beyond 326 * the always-required a_contents. as_find_callback does not sleep so 327 * it is ok to call it if more locks than a_contents (i.e. the a_lock 328 * rwlock) are held. as_execute_callback on the other hand may sleep 329 * so all locks beyond a_contents must be dropped by the caller if one 330 * does not want to end comatose. 331 */ 332 static int 333 as_do_callbacks(struct as *as, uint_t events, caddr_t event_addr, 334 size_t event_len) 335 { 336 struct as_callback *cb; 337 338 if ((cb = as_find_callback(as, events, event_addr, event_len))) { 339 as_execute_callback(as, cb, events); 340 return (-1); 341 } 342 return (0); 343 } 344 345 /* 346 * Search for the segment containing addr. If a segment containing addr 347 * exists, that segment is returned. If no such segment exists, and 348 * the list spans addresses greater than addr, then the first segment 349 * whose base is greater than addr is returned; otherwise, NULL is 350 * returned unless tail is true, in which case the last element of the 351 * list is returned. 352 * 353 * a_seglast is used to cache the last found segment for repeated 354 * searches to the same addr (which happens frequently). 355 */ 356 struct seg * 357 as_findseg(struct as *as, caddr_t addr, int tail) 358 { 359 struct seg *seg = as->a_seglast; 360 avl_index_t where; 361 362 ASSERT(AS_LOCK_HELD(as, &as->a_lock)); 363 364 if (seg != NULL && 365 seg->s_base <= addr && 366 addr < seg->s_base + seg->s_size) 367 return (seg); 368 369 seg = avl_find(&as->a_segtree, &addr, &where); 370 if (seg != NULL) 371 return (as->a_seglast = seg); 372 373 seg = avl_nearest(&as->a_segtree, where, AVL_AFTER); 374 if (seg == NULL && tail) 375 seg = avl_last(&as->a_segtree); 376 return (as->a_seglast = seg); 377 } 378 379 #ifdef VERIFY_SEGLIST 380 /* 381 * verify that the linked list is coherent 382 */ 383 static void 384 as_verify(struct as *as) 385 { 386 struct seg *seg, *seglast, *p, *n; 387 uint_t nsegs = 0; 388 389 if (do_as_verify == 0) 390 return; 391 392 seglast = as->a_seglast; 393 394 for (seg = AS_SEGFIRST(as); seg != NULL; seg = AS_SEGNEXT(as, seg)) { 395 ASSERT(seg->s_as == as); 396 p = AS_SEGPREV(as, seg); 397 n = AS_SEGNEXT(as, seg); 398 ASSERT(p == NULL || p->s_as == as); 399 ASSERT(p == NULL || p->s_base < seg->s_base); 400 ASSERT(n == NULL || n->s_base > seg->s_base); 401 ASSERT(n != NULL || seg == avl_last(&as->a_segtree)); 402 if (seg == seglast) 403 seglast = NULL; 404 nsegs++; 405 } 406 ASSERT(seglast == NULL); 407 ASSERT(avl_numnodes(&as->a_segtree) == nsegs); 408 } 409 #endif /* VERIFY_SEGLIST */ 410 411 /* 412 * Add a new segment to the address space. The avl_find() 413 * may be expensive so we attempt to use last segment accessed 414 * in as_gap() as an insertion point. 415 */ 416 int 417 as_addseg(struct as *as, struct seg *newseg) 418 { 419 struct seg *seg; 420 caddr_t addr; 421 caddr_t eaddr; 422 avl_index_t where; 423 424 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 425 426 as->a_updatedir = 1; /* inform /proc */ 427 gethrestime(&as->a_updatetime); 428 429 if (as->a_lastgaphl != NULL) { 430 struct seg *hseg = NULL; 431 struct seg *lseg = NULL; 432 433 if (as->a_lastgaphl->s_base > newseg->s_base) { 434 hseg = as->a_lastgaphl; 435 lseg = AVL_PREV(&as->a_segtree, hseg); 436 } else { 437 lseg = as->a_lastgaphl; 438 hseg = AVL_NEXT(&as->a_segtree, lseg); 439 } 440 441 if (hseg && lseg && lseg->s_base < newseg->s_base && 442 hseg->s_base > newseg->s_base) { 443 avl_insert_here(&as->a_segtree, newseg, lseg, 444 AVL_AFTER); 445 as->a_lastgaphl = NULL; 446 as->a_seglast = newseg; 447 return (0); 448 } 449 as->a_lastgaphl = NULL; 450 } 451 452 addr = newseg->s_base; 453 eaddr = addr + newseg->s_size; 454 again: 455 456 seg = avl_find(&as->a_segtree, &addr, &where); 457 458 if (seg == NULL) 459 seg = avl_nearest(&as->a_segtree, where, AVL_AFTER); 460 461 if (seg == NULL) 462 seg = avl_last(&as->a_segtree); 463 464 if (seg != NULL) { 465 caddr_t base = seg->s_base; 466 467 /* 468 * If top of seg is below the requested address, then 469 * the insertion point is at the end of the linked list, 470 * and seg points to the tail of the list. Otherwise, 471 * the insertion point is immediately before seg. 472 */ 473 if (base + seg->s_size > addr) { 474 if (addr >= base || eaddr > base) { 475 #ifdef __sparc 476 extern struct seg_ops segnf_ops; 477 478 /* 479 * no-fault segs must disappear if overlaid. 480 * XXX need new segment type so 481 * we don't have to check s_ops 482 */ 483 if (seg->s_ops == &segnf_ops) { 484 seg_unmap(seg); 485 goto again; 486 } 487 #endif 488 return (-1); /* overlapping segment */ 489 } 490 } 491 } 492 as->a_seglast = newseg; 493 avl_insert(&as->a_segtree, newseg, where); 494 495 #ifdef VERIFY_SEGLIST 496 as_verify(as); 497 #endif 498 return (0); 499 } 500 501 struct seg * 502 as_removeseg(struct as *as, struct seg *seg) 503 { 504 avl_tree_t *t; 505 506 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 507 508 as->a_updatedir = 1; /* inform /proc */ 509 gethrestime(&as->a_updatetime); 510 511 if (seg == NULL) 512 return (NULL); 513 514 t = &as->a_segtree; 515 if (as->a_seglast == seg) 516 as->a_seglast = NULL; 517 as->a_lastgaphl = NULL; 518 519 /* 520 * if this segment is at an address higher than 521 * a_lastgap, set a_lastgap to the next segment (NULL if last segment) 522 */ 523 if (as->a_lastgap && 524 (seg == as->a_lastgap || seg->s_base > as->a_lastgap->s_base)) 525 as->a_lastgap = AVL_NEXT(t, seg); 526 527 /* 528 * remove the segment from the seg tree 529 */ 530 avl_remove(t, seg); 531 532 #ifdef VERIFY_SEGLIST 533 as_verify(as); 534 #endif 535 return (seg); 536 } 537 538 /* 539 * Find a segment containing addr. 540 */ 541 struct seg * 542 as_segat(struct as *as, caddr_t addr) 543 { 544 struct seg *seg = as->a_seglast; 545 546 ASSERT(AS_LOCK_HELD(as, &as->a_lock)); 547 548 if (seg != NULL && seg->s_base <= addr && 549 addr < seg->s_base + seg->s_size) 550 return (seg); 551 552 seg = avl_find(&as->a_segtree, &addr, NULL); 553 return (seg); 554 } 555 556 /* 557 * Serialize all searches for holes in an address space to 558 * prevent two or more threads from allocating the same virtual 559 * address range. The address space must not be "read/write" 560 * locked by the caller since we may block. 561 */ 562 void 563 as_rangelock(struct as *as) 564 { 565 mutex_enter(&as->a_contents); 566 while (AS_ISCLAIMGAP(as)) 567 cv_wait(&as->a_cv, &as->a_contents); 568 AS_SETCLAIMGAP(as); 569 mutex_exit(&as->a_contents); 570 } 571 572 /* 573 * Release hold on a_state & AS_CLAIMGAP and signal any other blocked threads. 574 */ 575 void 576 as_rangeunlock(struct as *as) 577 { 578 mutex_enter(&as->a_contents); 579 AS_CLRCLAIMGAP(as); 580 cv_signal(&as->a_cv); 581 mutex_exit(&as->a_contents); 582 } 583 584 /* 585 * compar segments (or just an address) by segment address range 586 */ 587 static int 588 as_segcompar(const void *x, const void *y) 589 { 590 struct seg *a = (struct seg *)x; 591 struct seg *b = (struct seg *)y; 592 593 if (a->s_base < b->s_base) 594 return (-1); 595 if (a->s_base >= b->s_base + b->s_size) 596 return (1); 597 return (0); 598 } 599 600 601 void 602 as_avlinit(struct as *as) 603 { 604 avl_create(&as->a_segtree, as_segcompar, sizeof (struct seg), 605 offsetof(struct seg, s_tree)); 606 avl_create(&as->a_wpage, wp_compare, sizeof (struct watched_page), 607 offsetof(struct watched_page, wp_link)); 608 } 609 610 /*ARGSUSED*/ 611 static int 612 as_constructor(void *buf, void *cdrarg, int kmflags) 613 { 614 struct as *as = buf; 615 616 mutex_init(&as->a_contents, NULL, MUTEX_DEFAULT, NULL); 617 cv_init(&as->a_cv, NULL, CV_DEFAULT, NULL); 618 rw_init(&as->a_lock, NULL, RW_DEFAULT, NULL); 619 as_avlinit(as); 620 return (0); 621 } 622 623 /*ARGSUSED1*/ 624 static void 625 as_destructor(void *buf, void *cdrarg) 626 { 627 struct as *as = buf; 628 629 avl_destroy(&as->a_segtree); 630 mutex_destroy(&as->a_contents); 631 cv_destroy(&as->a_cv); 632 rw_destroy(&as->a_lock); 633 } 634 635 void 636 as_init(void) 637 { 638 as_cache = kmem_cache_create("as_cache", sizeof (struct as), 0, 639 as_constructor, as_destructor, NULL, NULL, NULL, 0); 640 } 641 642 /* 643 * Allocate and initialize an address space data structure. 644 * We call hat_alloc to allow any machine dependent 645 * information in the hat structure to be initialized. 646 */ 647 struct as * 648 as_alloc(void) 649 { 650 struct as *as; 651 652 as = kmem_cache_alloc(as_cache, KM_SLEEP); 653 654 as->a_flags = 0; 655 as->a_vbits = 0; 656 as->a_hrm = NULL; 657 as->a_seglast = NULL; 658 as->a_size = 0; 659 as->a_resvsize = 0; 660 as->a_updatedir = 0; 661 gethrestime(&as->a_updatetime); 662 as->a_objectdir = NULL; 663 as->a_sizedir = 0; 664 as->a_userlimit = (caddr_t)USERLIMIT; 665 as->a_lastgap = NULL; 666 as->a_lastgaphl = NULL; 667 as->a_callbacks = NULL; 668 669 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 670 as->a_hat = hat_alloc(as); /* create hat for default system mmu */ 671 AS_LOCK_EXIT(as, &as->a_lock); 672 673 as->a_xhat = NULL; 674 675 return (as); 676 } 677 678 /* 679 * Free an address space data structure. 680 * Need to free the hat first and then 681 * all the segments on this as and finally 682 * the space for the as struct itself. 683 */ 684 void 685 as_free(struct as *as) 686 { 687 struct hat *hat = as->a_hat; 688 struct seg *seg, *next; 689 int called = 0; 690 691 top: 692 /* 693 * Invoke ALL callbacks. as_do_callbacks will do one callback 694 * per call, and not return (-1) until the callback has completed. 695 * When as_do_callbacks returns zero, all callbacks have completed. 696 */ 697 mutex_enter(&as->a_contents); 698 while (as->a_callbacks && as_do_callbacks(as, AS_ALL_EVENT, 0, 0)) 699 ; 700 701 /* This will prevent new XHATs from attaching to as */ 702 if (!called) 703 AS_SETBUSY(as); 704 mutex_exit(&as->a_contents); 705 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 706 707 if (!called) { 708 called = 1; 709 hat_free_start(hat); 710 if (as->a_xhat != NULL) 711 xhat_free_start_all(as); 712 } 713 for (seg = AS_SEGFIRST(as); seg != NULL; seg = next) { 714 int err; 715 716 next = AS_SEGNEXT(as, seg); 717 retry: 718 err = SEGOP_UNMAP(seg, seg->s_base, seg->s_size); 719 if (err == EAGAIN) { 720 mutex_enter(&as->a_contents); 721 if (as->a_callbacks) { 722 AS_LOCK_EXIT(as, &as->a_lock); 723 } else if (!AS_ISNOUNMAPWAIT(as)) { 724 /* 725 * Memory is currently locked. Wait for a 726 * cv_signal that it has been unlocked, then 727 * try the operation again. 728 */ 729 if (AS_ISUNMAPWAIT(as) == 0) 730 cv_broadcast(&as->a_cv); 731 AS_SETUNMAPWAIT(as); 732 AS_LOCK_EXIT(as, &as->a_lock); 733 while (AS_ISUNMAPWAIT(as)) 734 cv_wait(&as->a_cv, &as->a_contents); 735 } else { 736 /* 737 * We may have raced with 738 * segvn_reclaim()/segspt_reclaim(). In this 739 * case clean nounmapwait flag and retry since 740 * softlockcnt in this segment may be already 741 * 0. We don't drop as writer lock so our 742 * number of retries without sleeping should 743 * be very small. See segvn_reclaim() for 744 * more comments. 745 */ 746 AS_CLRNOUNMAPWAIT(as); 747 mutex_exit(&as->a_contents); 748 goto retry; 749 } 750 mutex_exit(&as->a_contents); 751 goto top; 752 } else { 753 /* 754 * We do not expect any other error return at this 755 * time. This is similar to an ASSERT in seg_unmap() 756 */ 757 ASSERT(err == 0); 758 } 759 } 760 hat_free_end(hat); 761 if (as->a_xhat != NULL) 762 xhat_free_end_all(as); 763 AS_LOCK_EXIT(as, &as->a_lock); 764 765 /* /proc stuff */ 766 ASSERT(avl_numnodes(&as->a_wpage) == 0); 767 if (as->a_objectdir) { 768 kmem_free(as->a_objectdir, as->a_sizedir * sizeof (vnode_t *)); 769 as->a_objectdir = NULL; 770 as->a_sizedir = 0; 771 } 772 773 /* 774 * Free the struct as back to kmem. Assert it has no segments. 775 */ 776 ASSERT(avl_numnodes(&as->a_segtree) == 0); 777 kmem_cache_free(as_cache, as); 778 } 779 780 int 781 as_dup(struct as *as, struct proc *forkedproc) 782 { 783 struct as *newas; 784 struct seg *seg, *newseg; 785 size_t purgesize = 0; 786 int error; 787 788 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 789 as_clearwatch(as); 790 newas = as_alloc(); 791 newas->a_userlimit = as->a_userlimit; 792 newas->a_proc = forkedproc; 793 794 AS_LOCK_ENTER(newas, &newas->a_lock, RW_WRITER); 795 796 /* This will prevent new XHATs from attaching */ 797 mutex_enter(&as->a_contents); 798 AS_SETBUSY(as); 799 mutex_exit(&as->a_contents); 800 mutex_enter(&newas->a_contents); 801 AS_SETBUSY(newas); 802 mutex_exit(&newas->a_contents); 803 804 (void) hat_dup(as->a_hat, newas->a_hat, NULL, 0, HAT_DUP_SRD); 805 806 for (seg = AS_SEGFIRST(as); seg != NULL; seg = AS_SEGNEXT(as, seg)) { 807 808 if (seg->s_flags & S_PURGE) { 809 purgesize += seg->s_size; 810 continue; 811 } 812 813 newseg = seg_alloc(newas, seg->s_base, seg->s_size); 814 if (newseg == NULL) { 815 AS_LOCK_EXIT(newas, &newas->a_lock); 816 as_setwatch(as); 817 mutex_enter(&as->a_contents); 818 AS_CLRBUSY(as); 819 mutex_exit(&as->a_contents); 820 AS_LOCK_EXIT(as, &as->a_lock); 821 as_free(newas); 822 return (-1); 823 } 824 if ((error = SEGOP_DUP(seg, newseg)) != 0) { 825 /* 826 * We call seg_free() on the new seg 827 * because the segment is not set up 828 * completely; i.e. it has no ops. 829 */ 830 as_setwatch(as); 831 mutex_enter(&as->a_contents); 832 AS_CLRBUSY(as); 833 mutex_exit(&as->a_contents); 834 AS_LOCK_EXIT(as, &as->a_lock); 835 seg_free(newseg); 836 AS_LOCK_EXIT(newas, &newas->a_lock); 837 as_free(newas); 838 return (error); 839 } 840 newas->a_size += seg->s_size; 841 } 842 newas->a_resvsize = as->a_resvsize - purgesize; 843 844 error = hat_dup(as->a_hat, newas->a_hat, NULL, 0, HAT_DUP_ALL); 845 if (as->a_xhat != NULL) 846 error |= xhat_dup_all(as, newas, NULL, 0, HAT_DUP_ALL); 847 848 mutex_enter(&newas->a_contents); 849 AS_CLRBUSY(newas); 850 mutex_exit(&newas->a_contents); 851 AS_LOCK_EXIT(newas, &newas->a_lock); 852 853 as_setwatch(as); 854 mutex_enter(&as->a_contents); 855 AS_CLRBUSY(as); 856 mutex_exit(&as->a_contents); 857 AS_LOCK_EXIT(as, &as->a_lock); 858 if (error != 0) { 859 as_free(newas); 860 return (error); 861 } 862 forkedproc->p_as = newas; 863 return (0); 864 } 865 866 /* 867 * Handle a ``fault'' at addr for size bytes. 868 */ 869 faultcode_t 870 as_fault(struct hat *hat, struct as *as, caddr_t addr, size_t size, 871 enum fault_type type, enum seg_rw rw) 872 { 873 struct seg *seg; 874 caddr_t raddr; /* rounded down addr */ 875 size_t rsize; /* rounded up size */ 876 size_t ssize; 877 faultcode_t res = 0; 878 caddr_t addrsav; 879 struct seg *segsav; 880 int as_lock_held; 881 klwp_t *lwp = ttolwp(curthread); 882 int is_xhat = 0; 883 int holding_wpage = 0; 884 extern struct seg_ops segdev_ops; 885 886 887 888 if (as->a_hat != hat) { 889 /* This must be an XHAT then */ 890 is_xhat = 1; 891 892 if ((type != F_INVAL) || (as == &kas)) 893 return (FC_NOSUPPORT); 894 } 895 896 retry: 897 if (!is_xhat) { 898 /* 899 * Indicate that the lwp is not to be stopped while waiting 900 * for a pagefault. This is to avoid deadlock while debugging 901 * a process via /proc over NFS (in particular). 902 */ 903 if (lwp != NULL) 904 lwp->lwp_nostop++; 905 906 /* 907 * same length must be used when we softlock and softunlock. 908 * We don't support softunlocking lengths less than 909 * the original length when there is largepage support. 910 * See seg_dev.c for more comments. 911 */ 912 switch (type) { 913 914 case F_SOFTLOCK: 915 CPU_STATS_ADD_K(vm, softlock, 1); 916 break; 917 918 case F_SOFTUNLOCK: 919 break; 920 921 case F_PROT: 922 CPU_STATS_ADD_K(vm, prot_fault, 1); 923 break; 924 925 case F_INVAL: 926 CPU_STATS_ENTER_K(); 927 CPU_STATS_ADDQ(CPU, vm, as_fault, 1); 928 if (as == &kas) 929 CPU_STATS_ADDQ(CPU, vm, kernel_asflt, 1); 930 CPU_STATS_EXIT_K(); 931 break; 932 } 933 } 934 935 /* Kernel probe */ 936 TNF_PROBE_3(address_fault, "vm pagefault", /* CSTYLED */, 937 tnf_opaque, address, addr, 938 tnf_fault_type, fault_type, type, 939 tnf_seg_access, access, rw); 940 941 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 942 rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) - 943 (size_t)raddr; 944 945 /* 946 * XXX -- Don't grab the as lock for segkmap. We should grab it for 947 * correctness, but then we could be stuck holding this lock for 948 * a LONG time if the fault needs to be resolved on a slow 949 * filesystem, and then no-one will be able to exec new commands, 950 * as exec'ing requires the write lock on the as. 951 */ 952 if (as == &kas && segkmap && segkmap->s_base <= raddr && 953 raddr + size < segkmap->s_base + segkmap->s_size) { 954 /* 955 * if (as==&kas), this can't be XHAT: we've already returned 956 * FC_NOSUPPORT. 957 */ 958 seg = segkmap; 959 as_lock_held = 0; 960 } else { 961 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 962 if (is_xhat && avl_numnodes(&as->a_wpage) != 0) { 963 /* 964 * Grab and hold the writers' lock on the as 965 * if the fault is to a watched page. 966 * This will keep CPUs from "peeking" at the 967 * address range while we're temporarily boosting 968 * the permissions for the XHAT device to 969 * resolve the fault in the segment layer. 970 * 971 * We could check whether faulted address 972 * is within a watched page and only then grab 973 * the writer lock, but this is simpler. 974 */ 975 AS_LOCK_EXIT(as, &as->a_lock); 976 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 977 } 978 979 seg = as_segat(as, raddr); 980 if (seg == NULL) { 981 AS_LOCK_EXIT(as, &as->a_lock); 982 if ((lwp != NULL) && (!is_xhat)) 983 lwp->lwp_nostop--; 984 return (FC_NOMAP); 985 } 986 987 as_lock_held = 1; 988 } 989 990 addrsav = raddr; 991 segsav = seg; 992 993 for (; rsize != 0; rsize -= ssize, raddr += ssize) { 994 if (raddr >= seg->s_base + seg->s_size) { 995 seg = AS_SEGNEXT(as, seg); 996 if (seg == NULL || raddr != seg->s_base) { 997 res = FC_NOMAP; 998 break; 999 } 1000 } 1001 if (raddr + rsize > seg->s_base + seg->s_size) 1002 ssize = seg->s_base + seg->s_size - raddr; 1003 else 1004 ssize = rsize; 1005 1006 if (!is_xhat || (seg->s_ops != &segdev_ops)) { 1007 1008 if (is_xhat && avl_numnodes(&as->a_wpage) != 0 && 1009 pr_is_watchpage_as(raddr, rw, as)) { 1010 /* 1011 * Handle watch pages. If we're faulting on a 1012 * watched page from an X-hat, we have to 1013 * restore the original permissions while we 1014 * handle the fault. 1015 */ 1016 as_clearwatch(as); 1017 holding_wpage = 1; 1018 } 1019 1020 res = SEGOP_FAULT(hat, seg, raddr, ssize, type, rw); 1021 1022 /* Restore watchpoints */ 1023 if (holding_wpage) { 1024 as_setwatch(as); 1025 holding_wpage = 0; 1026 } 1027 1028 if (res != 0) 1029 break; 1030 } else { 1031 /* XHAT does not support seg_dev */ 1032 res = FC_NOSUPPORT; 1033 break; 1034 } 1035 } 1036 1037 /* 1038 * If we were SOFTLOCKing and encountered a failure, 1039 * we must SOFTUNLOCK the range we already did. (Maybe we 1040 * should just panic if we are SOFTLOCKing or even SOFTUNLOCKing 1041 * right here...) 1042 */ 1043 if (res != 0 && type == F_SOFTLOCK) { 1044 for (seg = segsav; addrsav < raddr; addrsav += ssize) { 1045 if (addrsav >= seg->s_base + seg->s_size) 1046 seg = AS_SEGNEXT(as, seg); 1047 ASSERT(seg != NULL); 1048 /* 1049 * Now call the fault routine again to perform the 1050 * unlock using S_OTHER instead of the rw variable 1051 * since we never got a chance to touch the pages. 1052 */ 1053 if (raddr > seg->s_base + seg->s_size) 1054 ssize = seg->s_base + seg->s_size - addrsav; 1055 else 1056 ssize = raddr - addrsav; 1057 (void) SEGOP_FAULT(hat, seg, addrsav, ssize, 1058 F_SOFTUNLOCK, S_OTHER); 1059 } 1060 } 1061 if (as_lock_held) 1062 AS_LOCK_EXIT(as, &as->a_lock); 1063 if ((lwp != NULL) && (!is_xhat)) 1064 lwp->lwp_nostop--; 1065 1066 /* 1067 * If the lower levels returned EDEADLK for a fault, 1068 * It means that we should retry the fault. Let's wait 1069 * a bit also to let the deadlock causing condition clear. 1070 * This is part of a gross hack to work around a design flaw 1071 * in the ufs/sds logging code and should go away when the 1072 * logging code is re-designed to fix the problem. See bug 1073 * 4125102 for details of the problem. 1074 */ 1075 if (FC_ERRNO(res) == EDEADLK) { 1076 delay(deadlk_wait); 1077 res = 0; 1078 goto retry; 1079 } 1080 return (res); 1081 } 1082 1083 1084 1085 /* 1086 * Asynchronous ``fault'' at addr for size bytes. 1087 */ 1088 faultcode_t 1089 as_faulta(struct as *as, caddr_t addr, size_t size) 1090 { 1091 struct seg *seg; 1092 caddr_t raddr; /* rounded down addr */ 1093 size_t rsize; /* rounded up size */ 1094 faultcode_t res = 0; 1095 klwp_t *lwp = ttolwp(curthread); 1096 1097 retry: 1098 /* 1099 * Indicate that the lwp is not to be stopped while waiting 1100 * for a pagefault. This is to avoid deadlock while debugging 1101 * a process via /proc over NFS (in particular). 1102 */ 1103 if (lwp != NULL) 1104 lwp->lwp_nostop++; 1105 1106 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 1107 rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) - 1108 (size_t)raddr; 1109 1110 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 1111 seg = as_segat(as, raddr); 1112 if (seg == NULL) { 1113 AS_LOCK_EXIT(as, &as->a_lock); 1114 if (lwp != NULL) 1115 lwp->lwp_nostop--; 1116 return (FC_NOMAP); 1117 } 1118 1119 for (; rsize != 0; rsize -= PAGESIZE, raddr += PAGESIZE) { 1120 if (raddr >= seg->s_base + seg->s_size) { 1121 seg = AS_SEGNEXT(as, seg); 1122 if (seg == NULL || raddr != seg->s_base) { 1123 res = FC_NOMAP; 1124 break; 1125 } 1126 } 1127 res = SEGOP_FAULTA(seg, raddr); 1128 if (res != 0) 1129 break; 1130 } 1131 AS_LOCK_EXIT(as, &as->a_lock); 1132 if (lwp != NULL) 1133 lwp->lwp_nostop--; 1134 /* 1135 * If the lower levels returned EDEADLK for a fault, 1136 * It means that we should retry the fault. Let's wait 1137 * a bit also to let the deadlock causing condition clear. 1138 * This is part of a gross hack to work around a design flaw 1139 * in the ufs/sds logging code and should go away when the 1140 * logging code is re-designed to fix the problem. See bug 1141 * 4125102 for details of the problem. 1142 */ 1143 if (FC_ERRNO(res) == EDEADLK) { 1144 delay(deadlk_wait); 1145 res = 0; 1146 goto retry; 1147 } 1148 return (res); 1149 } 1150 1151 /* 1152 * Set the virtual mapping for the interval from [addr : addr + size) 1153 * in address space `as' to have the specified protection. 1154 * It is ok for the range to cross over several segments, 1155 * as long as they are contiguous. 1156 */ 1157 int 1158 as_setprot(struct as *as, caddr_t addr, size_t size, uint_t prot) 1159 { 1160 struct seg *seg; 1161 struct as_callback *cb; 1162 size_t ssize; 1163 caddr_t raddr; /* rounded down addr */ 1164 size_t rsize; /* rounded up size */ 1165 int error = 0, writer = 0; 1166 caddr_t saveraddr; 1167 size_t saversize; 1168 1169 setprot_top: 1170 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 1171 rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) - 1172 (size_t)raddr; 1173 1174 if (raddr + rsize < raddr) /* check for wraparound */ 1175 return (ENOMEM); 1176 1177 saveraddr = raddr; 1178 saversize = rsize; 1179 1180 /* 1181 * Normally we only lock the as as a reader. But 1182 * if due to setprot the segment driver needs to split 1183 * a segment it will return IE_RETRY. Therefore we re-acquire 1184 * the as lock as a writer so the segment driver can change 1185 * the seg list. Also the segment driver will return IE_RETRY 1186 * after it has changed the segment list so we therefore keep 1187 * locking as a writer. Since these opeartions should be rare 1188 * want to only lock as a writer when necessary. 1189 */ 1190 if (writer || avl_numnodes(&as->a_wpage) != 0) { 1191 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 1192 } else { 1193 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 1194 } 1195 1196 as_clearwatchprot(as, raddr, rsize); 1197 seg = as_segat(as, raddr); 1198 if (seg == NULL) { 1199 as_setwatch(as); 1200 AS_LOCK_EXIT(as, &as->a_lock); 1201 return (ENOMEM); 1202 } 1203 1204 for (; rsize != 0; rsize -= ssize, raddr += ssize) { 1205 if (raddr >= seg->s_base + seg->s_size) { 1206 seg = AS_SEGNEXT(as, seg); 1207 if (seg == NULL || raddr != seg->s_base) { 1208 error = ENOMEM; 1209 break; 1210 } 1211 } 1212 if ((raddr + rsize) > (seg->s_base + seg->s_size)) 1213 ssize = seg->s_base + seg->s_size - raddr; 1214 else 1215 ssize = rsize; 1216 retry: 1217 error = SEGOP_SETPROT(seg, raddr, ssize, prot); 1218 1219 if (error == IE_NOMEM) { 1220 error = EAGAIN; 1221 break; 1222 } 1223 1224 if (error == IE_RETRY) { 1225 AS_LOCK_EXIT(as, &as->a_lock); 1226 writer = 1; 1227 goto setprot_top; 1228 } 1229 1230 if (error == EAGAIN) { 1231 /* 1232 * Make sure we have a_lock as writer. 1233 */ 1234 if (writer == 0) { 1235 AS_LOCK_EXIT(as, &as->a_lock); 1236 writer = 1; 1237 goto setprot_top; 1238 } 1239 1240 /* 1241 * Memory is currently locked. It must be unlocked 1242 * before this operation can succeed through a retry. 1243 * The possible reasons for locked memory and 1244 * corresponding strategies for unlocking are: 1245 * (1) Normal I/O 1246 * wait for a signal that the I/O operation 1247 * has completed and the memory is unlocked. 1248 * (2) Asynchronous I/O 1249 * The aio subsystem does not unlock pages when 1250 * the I/O is completed. Those pages are unlocked 1251 * when the application calls aiowait/aioerror. 1252 * So, to prevent blocking forever, cv_broadcast() 1253 * is done to wake up aio_cleanup_thread. 1254 * Subsequently, segvn_reclaim will be called, and 1255 * that will do AS_CLRUNMAPWAIT() and wake us up. 1256 * (3) Long term page locking: 1257 * Drivers intending to have pages locked for a 1258 * period considerably longer than for normal I/O 1259 * (essentially forever) may have registered for a 1260 * callback so they may unlock these pages on 1261 * request. This is needed to allow this operation 1262 * to succeed. Each entry on the callback list is 1263 * examined. If the event or address range pertains 1264 * the callback is invoked (unless it already is in 1265 * progress). The a_contents lock must be dropped 1266 * before the callback, so only one callback can 1267 * be done at a time. Go to the top and do more 1268 * until zero is returned. If zero is returned, 1269 * either there were no callbacks for this event 1270 * or they were already in progress. 1271 */ 1272 mutex_enter(&as->a_contents); 1273 if (as->a_callbacks && 1274 (cb = as_find_callback(as, AS_SETPROT_EVENT, 1275 seg->s_base, seg->s_size))) { 1276 AS_LOCK_EXIT(as, &as->a_lock); 1277 as_execute_callback(as, cb, AS_SETPROT_EVENT); 1278 } else if (!AS_ISNOUNMAPWAIT(as)) { 1279 if (AS_ISUNMAPWAIT(as) == 0) 1280 cv_broadcast(&as->a_cv); 1281 AS_SETUNMAPWAIT(as); 1282 AS_LOCK_EXIT(as, &as->a_lock); 1283 while (AS_ISUNMAPWAIT(as)) 1284 cv_wait(&as->a_cv, &as->a_contents); 1285 } else { 1286 /* 1287 * We may have raced with 1288 * segvn_reclaim()/segspt_reclaim(). In this 1289 * case clean nounmapwait flag and retry since 1290 * softlockcnt in this segment may be already 1291 * 0. We don't drop as writer lock so our 1292 * number of retries without sleeping should 1293 * be very small. See segvn_reclaim() for 1294 * more comments. 1295 */ 1296 AS_CLRNOUNMAPWAIT(as); 1297 mutex_exit(&as->a_contents); 1298 goto retry; 1299 } 1300 mutex_exit(&as->a_contents); 1301 goto setprot_top; 1302 } else if (error != 0) 1303 break; 1304 } 1305 if (error != 0) { 1306 as_setwatch(as); 1307 } else { 1308 as_setwatchprot(as, saveraddr, saversize, prot); 1309 } 1310 AS_LOCK_EXIT(as, &as->a_lock); 1311 return (error); 1312 } 1313 1314 /* 1315 * Check to make sure that the interval [addr, addr + size) 1316 * in address space `as' has at least the specified protection. 1317 * It is ok for the range to cross over several segments, as long 1318 * as they are contiguous. 1319 */ 1320 int 1321 as_checkprot(struct as *as, caddr_t addr, size_t size, uint_t prot) 1322 { 1323 struct seg *seg; 1324 size_t ssize; 1325 caddr_t raddr; /* rounded down addr */ 1326 size_t rsize; /* rounded up size */ 1327 int error = 0; 1328 1329 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 1330 rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) - 1331 (size_t)raddr; 1332 1333 if (raddr + rsize < raddr) /* check for wraparound */ 1334 return (ENOMEM); 1335 1336 /* 1337 * This is ugly as sin... 1338 * Normally, we only acquire the address space readers lock. 1339 * However, if the address space has watchpoints present, 1340 * we must acquire the writer lock on the address space for 1341 * the benefit of as_clearwatchprot() and as_setwatchprot(). 1342 */ 1343 if (avl_numnodes(&as->a_wpage) != 0) 1344 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 1345 else 1346 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 1347 as_clearwatchprot(as, raddr, rsize); 1348 seg = as_segat(as, raddr); 1349 if (seg == NULL) { 1350 as_setwatch(as); 1351 AS_LOCK_EXIT(as, &as->a_lock); 1352 return (ENOMEM); 1353 } 1354 1355 for (; rsize != 0; rsize -= ssize, raddr += ssize) { 1356 if (raddr >= seg->s_base + seg->s_size) { 1357 seg = AS_SEGNEXT(as, seg); 1358 if (seg == NULL || raddr != seg->s_base) { 1359 error = ENOMEM; 1360 break; 1361 } 1362 } 1363 if ((raddr + rsize) > (seg->s_base + seg->s_size)) 1364 ssize = seg->s_base + seg->s_size - raddr; 1365 else 1366 ssize = rsize; 1367 1368 error = SEGOP_CHECKPROT(seg, raddr, ssize, prot); 1369 if (error != 0) 1370 break; 1371 } 1372 as_setwatch(as); 1373 AS_LOCK_EXIT(as, &as->a_lock); 1374 return (error); 1375 } 1376 1377 int 1378 as_unmap(struct as *as, caddr_t addr, size_t size) 1379 { 1380 struct seg *seg, *seg_next; 1381 struct as_callback *cb; 1382 caddr_t raddr, eaddr; 1383 size_t ssize, rsize = 0; 1384 int err; 1385 1386 top: 1387 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 1388 eaddr = (caddr_t)(((uintptr_t)(addr + size) + PAGEOFFSET) & 1389 (uintptr_t)PAGEMASK); 1390 1391 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 1392 1393 as->a_updatedir = 1; /* inform /proc */ 1394 gethrestime(&as->a_updatetime); 1395 1396 /* 1397 * Use as_findseg to find the first segment in the range, then 1398 * step through the segments in order, following s_next. 1399 */ 1400 as_clearwatchprot(as, raddr, eaddr - raddr); 1401 1402 for (seg = as_findseg(as, raddr, 0); seg != NULL; seg = seg_next) { 1403 if (eaddr <= seg->s_base) 1404 break; /* eaddr was in a gap; all done */ 1405 1406 /* this is implied by the test above */ 1407 ASSERT(raddr < eaddr); 1408 1409 if (raddr < seg->s_base) 1410 raddr = seg->s_base; /* raddr was in a gap */ 1411 1412 if (eaddr > (seg->s_base + seg->s_size)) 1413 ssize = seg->s_base + seg->s_size - raddr; 1414 else 1415 ssize = eaddr - raddr; 1416 1417 /* 1418 * Save next segment pointer since seg can be 1419 * destroyed during the segment unmap operation. 1420 */ 1421 seg_next = AS_SEGNEXT(as, seg); 1422 1423 /* 1424 * We didn't count /dev/null mappings, so ignore them here. 1425 * We'll handle MAP_NORESERVE cases in segvn_unmap(). (Again, 1426 * we have to do this check here while we have seg.) 1427 */ 1428 rsize = 0; 1429 if (!SEG_IS_DEVNULL_MAPPING(seg) && 1430 !SEG_IS_PARTIAL_RESV(seg)) 1431 rsize = ssize; 1432 1433 retry: 1434 err = SEGOP_UNMAP(seg, raddr, ssize); 1435 if (err == EAGAIN) { 1436 /* 1437 * Memory is currently locked. It must be unlocked 1438 * before this operation can succeed through a retry. 1439 * The possible reasons for locked memory and 1440 * corresponding strategies for unlocking are: 1441 * (1) Normal I/O 1442 * wait for a signal that the I/O operation 1443 * has completed and the memory is unlocked. 1444 * (2) Asynchronous I/O 1445 * The aio subsystem does not unlock pages when 1446 * the I/O is completed. Those pages are unlocked 1447 * when the application calls aiowait/aioerror. 1448 * So, to prevent blocking forever, cv_broadcast() 1449 * is done to wake up aio_cleanup_thread. 1450 * Subsequently, segvn_reclaim will be called, and 1451 * that will do AS_CLRUNMAPWAIT() and wake us up. 1452 * (3) Long term page locking: 1453 * Drivers intending to have pages locked for a 1454 * period considerably longer than for normal I/O 1455 * (essentially forever) may have registered for a 1456 * callback so they may unlock these pages on 1457 * request. This is needed to allow this operation 1458 * to succeed. Each entry on the callback list is 1459 * examined. If the event or address range pertains 1460 * the callback is invoked (unless it already is in 1461 * progress). The a_contents lock must be dropped 1462 * before the callback, so only one callback can 1463 * be done at a time. Go to the top and do more 1464 * until zero is returned. If zero is returned, 1465 * either there were no callbacks for this event 1466 * or they were already in progress. 1467 */ 1468 mutex_enter(&as->a_contents); 1469 if (as->a_callbacks && 1470 (cb = as_find_callback(as, AS_UNMAP_EVENT, 1471 seg->s_base, seg->s_size))) { 1472 AS_LOCK_EXIT(as, &as->a_lock); 1473 as_execute_callback(as, cb, AS_UNMAP_EVENT); 1474 } else if (!AS_ISNOUNMAPWAIT(as)) { 1475 if (AS_ISUNMAPWAIT(as) == 0) 1476 cv_broadcast(&as->a_cv); 1477 AS_SETUNMAPWAIT(as); 1478 AS_LOCK_EXIT(as, &as->a_lock); 1479 while (AS_ISUNMAPWAIT(as)) 1480 cv_wait(&as->a_cv, &as->a_contents); 1481 } else { 1482 /* 1483 * We may have raced with 1484 * segvn_reclaim()/segspt_reclaim(). In this 1485 * case clean nounmapwait flag and retry since 1486 * softlockcnt in this segment may be already 1487 * 0. We don't drop as writer lock so our 1488 * number of retries without sleeping should 1489 * be very small. See segvn_reclaim() for 1490 * more comments. 1491 */ 1492 AS_CLRNOUNMAPWAIT(as); 1493 mutex_exit(&as->a_contents); 1494 goto retry; 1495 } 1496 mutex_exit(&as->a_contents); 1497 goto top; 1498 } else if (err == IE_RETRY) { 1499 AS_LOCK_EXIT(as, &as->a_lock); 1500 goto top; 1501 } else if (err) { 1502 as_setwatch(as); 1503 AS_LOCK_EXIT(as, &as->a_lock); 1504 return (-1); 1505 } 1506 1507 as->a_size -= ssize; 1508 if (rsize) 1509 as->a_resvsize -= rsize; 1510 raddr += ssize; 1511 } 1512 AS_LOCK_EXIT(as, &as->a_lock); 1513 return (0); 1514 } 1515 1516 static int 1517 as_map_segvn_segs(struct as *as, caddr_t addr, size_t size, uint_t szcvec, 1518 int (*crfp)(), struct segvn_crargs *vn_a, int *segcreated) 1519 { 1520 uint_t szc; 1521 uint_t nszc; 1522 int error; 1523 caddr_t a; 1524 caddr_t eaddr; 1525 size_t segsize; 1526 struct seg *seg; 1527 size_t pgsz; 1528 int do_off = (vn_a->vp != NULL || vn_a->amp != NULL); 1529 uint_t save_szcvec; 1530 1531 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 1532 ASSERT(IS_P2ALIGNED(addr, PAGESIZE)); 1533 ASSERT(IS_P2ALIGNED(size, PAGESIZE)); 1534 ASSERT(vn_a->vp == NULL || vn_a->amp == NULL); 1535 if (!do_off) { 1536 vn_a->offset = 0; 1537 } 1538 1539 if (szcvec <= 1) { 1540 seg = seg_alloc(as, addr, size); 1541 if (seg == NULL) { 1542 return (ENOMEM); 1543 } 1544 vn_a->szc = 0; 1545 error = (*crfp)(seg, vn_a); 1546 if (error != 0) { 1547 seg_free(seg); 1548 } else { 1549 as->a_size += size; 1550 as->a_resvsize += size; 1551 } 1552 return (error); 1553 } 1554 1555 eaddr = addr + size; 1556 save_szcvec = szcvec; 1557 szcvec >>= 1; 1558 szc = 0; 1559 nszc = 0; 1560 while (szcvec) { 1561 if ((szcvec & 0x1) == 0) { 1562 nszc++; 1563 szcvec >>= 1; 1564 continue; 1565 } 1566 nszc++; 1567 pgsz = page_get_pagesize(nszc); 1568 a = (caddr_t)P2ROUNDUP((uintptr_t)addr, pgsz); 1569 if (a != addr) { 1570 ASSERT(a < eaddr); 1571 segsize = a - addr; 1572 seg = seg_alloc(as, addr, segsize); 1573 if (seg == NULL) { 1574 return (ENOMEM); 1575 } 1576 vn_a->szc = szc; 1577 error = (*crfp)(seg, vn_a); 1578 if (error != 0) { 1579 seg_free(seg); 1580 return (error); 1581 } 1582 as->a_size += segsize; 1583 as->a_resvsize += segsize; 1584 *segcreated = 1; 1585 if (do_off) { 1586 vn_a->offset += segsize; 1587 } 1588 addr = a; 1589 } 1590 szc = nszc; 1591 szcvec >>= 1; 1592 } 1593 1594 ASSERT(addr < eaddr); 1595 szcvec = save_szcvec | 1; /* add 8K pages */ 1596 while (szcvec) { 1597 a = (caddr_t)P2ALIGN((uintptr_t)eaddr, pgsz); 1598 ASSERT(a >= addr); 1599 if (a != addr) { 1600 segsize = a - addr; 1601 seg = seg_alloc(as, addr, segsize); 1602 if (seg == NULL) { 1603 return (ENOMEM); 1604 } 1605 vn_a->szc = szc; 1606 error = (*crfp)(seg, vn_a); 1607 if (error != 0) { 1608 seg_free(seg); 1609 return (error); 1610 } 1611 as->a_size += segsize; 1612 as->a_resvsize += segsize; 1613 *segcreated = 1; 1614 if (do_off) { 1615 vn_a->offset += segsize; 1616 } 1617 addr = a; 1618 } 1619 szcvec &= ~(1 << szc); 1620 if (szcvec) { 1621 szc = highbit(szcvec) - 1; 1622 pgsz = page_get_pagesize(szc); 1623 } 1624 } 1625 ASSERT(addr == eaddr); 1626 1627 return (0); 1628 } 1629 1630 static int 1631 as_map_vnsegs(struct as *as, caddr_t addr, size_t size, 1632 int (*crfp)(), struct segvn_crargs *vn_a, int *segcreated) 1633 { 1634 uint_t mapflags = vn_a->flags & (MAP_TEXT | MAP_INITDATA); 1635 int type = (vn_a->type == MAP_SHARED) ? MAPPGSZC_SHM : MAPPGSZC_PRIVM; 1636 uint_t szcvec = map_pgszcvec(addr, size, (uintptr_t)addr, mapflags, 1637 type, 0); 1638 int error; 1639 struct seg *seg; 1640 struct vattr va; 1641 u_offset_t eoff; 1642 size_t save_size = 0; 1643 extern size_t textrepl_size_thresh; 1644 1645 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 1646 ASSERT(IS_P2ALIGNED(addr, PAGESIZE)); 1647 ASSERT(IS_P2ALIGNED(size, PAGESIZE)); 1648 ASSERT(vn_a->vp != NULL); 1649 ASSERT(vn_a->amp == NULL); 1650 1651 again: 1652 if (szcvec <= 1) { 1653 seg = seg_alloc(as, addr, size); 1654 if (seg == NULL) { 1655 return (ENOMEM); 1656 } 1657 vn_a->szc = 0; 1658 error = (*crfp)(seg, vn_a); 1659 if (error != 0) { 1660 seg_free(seg); 1661 } else { 1662 as->a_size += size; 1663 as->a_resvsize += size; 1664 } 1665 return (error); 1666 } 1667 1668 va.va_mask = AT_SIZE; 1669 if (VOP_GETATTR(vn_a->vp, &va, ATTR_HINT, vn_a->cred, NULL) != 0) { 1670 szcvec = 0; 1671 goto again; 1672 } 1673 eoff = vn_a->offset & PAGEMASK; 1674 if (eoff >= va.va_size) { 1675 szcvec = 0; 1676 goto again; 1677 } 1678 eoff += size; 1679 if (btopr(va.va_size) < btopr(eoff)) { 1680 save_size = size; 1681 size = va.va_size - (vn_a->offset & PAGEMASK); 1682 size = P2ROUNDUP_TYPED(size, PAGESIZE, size_t); 1683 szcvec = map_pgszcvec(addr, size, (uintptr_t)addr, mapflags, 1684 type, 0); 1685 if (szcvec <= 1) { 1686 size = save_size; 1687 goto again; 1688 } 1689 } 1690 1691 if (size > textrepl_size_thresh) { 1692 vn_a->flags |= _MAP_TEXTREPL; 1693 } 1694 error = as_map_segvn_segs(as, addr, size, szcvec, crfp, vn_a, 1695 segcreated); 1696 if (error != 0) { 1697 return (error); 1698 } 1699 if (save_size) { 1700 addr += size; 1701 size = save_size - size; 1702 szcvec = 0; 1703 goto again; 1704 } 1705 return (0); 1706 } 1707 1708 /* 1709 * as_map_ansegs: shared or private anonymous memory. Note that the flags 1710 * passed to map_pgszvec cannot be MAP_INITDATA, for anon. 1711 */ 1712 static int 1713 as_map_ansegs(struct as *as, caddr_t addr, size_t size, 1714 int (*crfp)(), struct segvn_crargs *vn_a, int *segcreated) 1715 { 1716 uint_t szcvec; 1717 uchar_t type; 1718 1719 ASSERT(vn_a->type == MAP_SHARED || vn_a->type == MAP_PRIVATE); 1720 if (vn_a->type == MAP_SHARED) { 1721 type = MAPPGSZC_SHM; 1722 } else if (vn_a->type == MAP_PRIVATE) { 1723 if (vn_a->szc == AS_MAP_HEAP) { 1724 type = MAPPGSZC_HEAP; 1725 } else if (vn_a->szc == AS_MAP_STACK) { 1726 type = MAPPGSZC_STACK; 1727 } else { 1728 type = MAPPGSZC_PRIVM; 1729 } 1730 } 1731 szcvec = map_pgszcvec(addr, size, vn_a->amp == NULL ? 1732 (uintptr_t)addr : (uintptr_t)P2ROUNDUP(vn_a->offset, PAGESIZE), 1733 (vn_a->flags & MAP_TEXT), type, 0); 1734 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 1735 ASSERT(IS_P2ALIGNED(addr, PAGESIZE)); 1736 ASSERT(IS_P2ALIGNED(size, PAGESIZE)); 1737 ASSERT(vn_a->vp == NULL); 1738 1739 return (as_map_segvn_segs(as, addr, size, szcvec, 1740 crfp, vn_a, segcreated)); 1741 } 1742 1743 int 1744 as_map(struct as *as, caddr_t addr, size_t size, int (*crfp)(), void *argsp) 1745 { 1746 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 1747 return (as_map_locked(as, addr, size, crfp, argsp)); 1748 } 1749 1750 int 1751 as_map_locked(struct as *as, caddr_t addr, size_t size, int (*crfp)(), 1752 void *argsp) 1753 { 1754 struct seg *seg = NULL; 1755 caddr_t raddr; /* rounded down addr */ 1756 size_t rsize; /* rounded up size */ 1757 int error; 1758 int unmap = 0; 1759 struct proc *p = curproc; 1760 struct segvn_crargs crargs; 1761 1762 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 1763 rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) - 1764 (size_t)raddr; 1765 1766 /* 1767 * check for wrap around 1768 */ 1769 if ((raddr + rsize < raddr) || (as->a_size > (ULONG_MAX - size))) { 1770 AS_LOCK_EXIT(as, &as->a_lock); 1771 return (ENOMEM); 1772 } 1773 1774 as->a_updatedir = 1; /* inform /proc */ 1775 gethrestime(&as->a_updatetime); 1776 1777 if (as != &kas && as->a_size + rsize > (size_t)p->p_vmem_ctl) { 1778 AS_LOCK_EXIT(as, &as->a_lock); 1779 1780 (void) rctl_action(rctlproc_legacy[RLIMIT_VMEM], p->p_rctls, p, 1781 RCA_UNSAFE_ALL); 1782 1783 return (ENOMEM); 1784 } 1785 1786 if (AS_MAP_CHECK_VNODE_LPOOB(crfp, argsp)) { 1787 crargs = *(struct segvn_crargs *)argsp; 1788 error = as_map_vnsegs(as, raddr, rsize, crfp, &crargs, &unmap); 1789 if (error != 0) { 1790 AS_LOCK_EXIT(as, &as->a_lock); 1791 if (unmap) { 1792 (void) as_unmap(as, addr, size); 1793 } 1794 return (error); 1795 } 1796 } else if (AS_MAP_CHECK_ANON_LPOOB(crfp, argsp)) { 1797 crargs = *(struct segvn_crargs *)argsp; 1798 error = as_map_ansegs(as, raddr, rsize, crfp, &crargs, &unmap); 1799 if (error != 0) { 1800 AS_LOCK_EXIT(as, &as->a_lock); 1801 if (unmap) { 1802 (void) as_unmap(as, addr, size); 1803 } 1804 return (error); 1805 } 1806 } else { 1807 seg = seg_alloc(as, addr, size); 1808 if (seg == NULL) { 1809 AS_LOCK_EXIT(as, &as->a_lock); 1810 return (ENOMEM); 1811 } 1812 1813 error = (*crfp)(seg, argsp); 1814 if (error != 0) { 1815 seg_free(seg); 1816 AS_LOCK_EXIT(as, &as->a_lock); 1817 return (error); 1818 } 1819 /* 1820 * Add size now so as_unmap will work if as_ctl fails. 1821 */ 1822 as->a_size += rsize; 1823 as->a_resvsize += rsize; 1824 } 1825 1826 as_setwatch(as); 1827 1828 /* 1829 * If the address space is locked, 1830 * establish memory locks for the new segment. 1831 */ 1832 mutex_enter(&as->a_contents); 1833 if (AS_ISPGLCK(as)) { 1834 mutex_exit(&as->a_contents); 1835 AS_LOCK_EXIT(as, &as->a_lock); 1836 error = as_ctl(as, addr, size, MC_LOCK, 0, 0, NULL, 0); 1837 if (error != 0) 1838 (void) as_unmap(as, addr, size); 1839 } else { 1840 mutex_exit(&as->a_contents); 1841 AS_LOCK_EXIT(as, &as->a_lock); 1842 } 1843 return (error); 1844 } 1845 1846 1847 /* 1848 * Delete all segments in the address space marked with S_PURGE. 1849 * This is currently used for Sparc V9 nofault ASI segments (seg_nf.c). 1850 * These segments are deleted as a first step before calls to as_gap(), so 1851 * that they don't affect mmap() or shmat(). 1852 */ 1853 void 1854 as_purge(struct as *as) 1855 { 1856 struct seg *seg; 1857 struct seg *next_seg; 1858 1859 /* 1860 * the setting of NEEDSPURGE is protect by as_rangelock(), so 1861 * no need to grab a_contents mutex for this check 1862 */ 1863 if ((as->a_flags & AS_NEEDSPURGE) == 0) 1864 return; 1865 1866 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 1867 next_seg = NULL; 1868 seg = AS_SEGFIRST(as); 1869 while (seg != NULL) { 1870 next_seg = AS_SEGNEXT(as, seg); 1871 if (seg->s_flags & S_PURGE) 1872 SEGOP_UNMAP(seg, seg->s_base, seg->s_size); 1873 seg = next_seg; 1874 } 1875 AS_LOCK_EXIT(as, &as->a_lock); 1876 1877 mutex_enter(&as->a_contents); 1878 as->a_flags &= ~AS_NEEDSPURGE; 1879 mutex_exit(&as->a_contents); 1880 } 1881 1882 /* 1883 * Find a hole within [*basep, *basep + *lenp), which contains a mappable 1884 * range of addresses at least "minlen" long, where the base of the range is 1885 * at "off" phase from an "align" boundary and there is space for a 1886 * "redzone"-sized redzone on eithe rside of the range. Thus, 1887 * if align was 4M and off was 16k, the user wants a hole which will start 1888 * 16k into a 4M page. 1889 * 1890 * If flags specifies AH_HI, the hole will have the highest possible address 1891 * in the range. We use the as->a_lastgap field to figure out where to 1892 * start looking for a gap. 1893 * 1894 * Otherwise, the gap will have the lowest possible address. 1895 * 1896 * If flags specifies AH_CONTAIN, the hole will contain the address addr. 1897 * 1898 * If an adequate hole is found, *basep and *lenp are set to reflect the part of 1899 * the hole that is within range, and 0 is returned. On failure, -1 is returned. 1900 * 1901 * NOTE: This routine is not correct when base+len overflows caddr_t. 1902 */ 1903 int 1904 as_gap_aligned(struct as *as, size_t minlen, caddr_t *basep, size_t *lenp, 1905 uint_t flags, caddr_t addr, size_t align, size_t redzone, size_t off) 1906 { 1907 caddr_t lobound = *basep; 1908 caddr_t hibound = lobound + *lenp; 1909 struct seg *lseg, *hseg; 1910 caddr_t lo, hi; 1911 int forward; 1912 caddr_t save_base; 1913 size_t save_len; 1914 size_t save_minlen; 1915 size_t save_redzone; 1916 int fast_path = 1; 1917 1918 save_base = *basep; 1919 save_len = *lenp; 1920 save_minlen = minlen; 1921 save_redzone = redzone; 1922 1923 /* 1924 * For the first pass/fast_path, just add align and redzone into 1925 * minlen since if we get an allocation, we can guarantee that it 1926 * will fit the alignment and redzone requested. 1927 * This increases the chance that hibound will be adjusted to 1928 * a_lastgap->s_base which will likely allow us to find an 1929 * acceptable hole in the address space quicker. 1930 * If we can't find a hole with this fast_path, then we look for 1931 * smaller holes in which the alignment and offset may allow 1932 * the allocation to fit. 1933 */ 1934 minlen += align; 1935 minlen += 2 * redzone; 1936 redzone = 0; 1937 1938 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 1939 if (AS_SEGFIRST(as) == NULL) { 1940 if (valid_va_range_aligned(basep, lenp, minlen, flags & AH_DIR, 1941 align, redzone, off)) { 1942 AS_LOCK_EXIT(as, &as->a_lock); 1943 return (0); 1944 } else { 1945 AS_LOCK_EXIT(as, &as->a_lock); 1946 *basep = save_base; 1947 *lenp = save_len; 1948 return (-1); 1949 } 1950 } 1951 1952 retry: 1953 /* 1954 * Set up to iterate over all the inter-segment holes in the given 1955 * direction. lseg is NULL for the lowest-addressed hole and hseg is 1956 * NULL for the highest-addressed hole. If moving backwards, we reset 1957 * sseg to denote the highest-addressed segment. 1958 */ 1959 forward = (flags & AH_DIR) == AH_LO; 1960 if (forward) { 1961 hseg = as_findseg(as, lobound, 1); 1962 lseg = AS_SEGPREV(as, hseg); 1963 } else { 1964 1965 /* 1966 * If allocating at least as much as the last allocation, 1967 * use a_lastgap's base as a better estimate of hibound. 1968 */ 1969 if (as->a_lastgap && 1970 minlen >= as->a_lastgap->s_size && 1971 hibound >= as->a_lastgap->s_base) 1972 hibound = as->a_lastgap->s_base; 1973 1974 hseg = as_findseg(as, hibound, 1); 1975 if (hseg->s_base + hseg->s_size < hibound) { 1976 lseg = hseg; 1977 hseg = NULL; 1978 } else { 1979 lseg = AS_SEGPREV(as, hseg); 1980 } 1981 } 1982 1983 for (;;) { 1984 /* 1985 * Set lo and hi to the hole's boundaries. (We should really 1986 * use MAXADDR in place of hibound in the expression below, 1987 * but can't express it easily; using hibound in its place is 1988 * harmless.) 1989 */ 1990 lo = (lseg == NULL) ? 0 : lseg->s_base + lseg->s_size; 1991 hi = (hseg == NULL) ? hibound : hseg->s_base; 1992 /* 1993 * If the iteration has moved past the interval from lobound 1994 * to hibound it's pointless to continue. 1995 */ 1996 if ((forward && lo > hibound) || (!forward && hi < lobound)) 1997 break; 1998 else if (lo > hibound || hi < lobound) 1999 goto cont; 2000 /* 2001 * Candidate hole lies at least partially within the allowable 2002 * range. Restrict it to fall completely within that range, 2003 * i.e., to [max(lo, lobound), min(hi, hibound)]. 2004 */ 2005 if (lo < lobound) 2006 lo = lobound; 2007 if (hi > hibound) 2008 hi = hibound; 2009 /* 2010 * Verify that the candidate hole is big enough and meets 2011 * hardware constraints. If the hole is too small, no need 2012 * to do the further checks since they will fail. 2013 */ 2014 *basep = lo; 2015 *lenp = hi - lo; 2016 if (*lenp >= minlen && valid_va_range_aligned(basep, lenp, 2017 minlen, forward ? AH_LO : AH_HI, align, redzone, off) && 2018 ((flags & AH_CONTAIN) == 0 || 2019 (*basep <= addr && *basep + *lenp > addr))) { 2020 if (!forward) 2021 as->a_lastgap = hseg; 2022 if (hseg != NULL) 2023 as->a_lastgaphl = hseg; 2024 else 2025 as->a_lastgaphl = lseg; 2026 AS_LOCK_EXIT(as, &as->a_lock); 2027 return (0); 2028 } 2029 cont: 2030 /* 2031 * Move to the next hole. 2032 */ 2033 if (forward) { 2034 lseg = hseg; 2035 if (lseg == NULL) 2036 break; 2037 hseg = AS_SEGNEXT(as, hseg); 2038 } else { 2039 hseg = lseg; 2040 if (hseg == NULL) 2041 break; 2042 lseg = AS_SEGPREV(as, lseg); 2043 } 2044 } 2045 if (fast_path && (align != 0 || save_redzone != 0)) { 2046 fast_path = 0; 2047 minlen = save_minlen; 2048 redzone = save_redzone; 2049 goto retry; 2050 } 2051 *basep = save_base; 2052 *lenp = save_len; 2053 AS_LOCK_EXIT(as, &as->a_lock); 2054 return (-1); 2055 } 2056 2057 /* 2058 * Find a hole of at least size minlen within [*basep, *basep + *lenp). 2059 * 2060 * If flags specifies AH_HI, the hole will have the highest possible address 2061 * in the range. We use the as->a_lastgap field to figure out where to 2062 * start looking for a gap. 2063 * 2064 * Otherwise, the gap will have the lowest possible address. 2065 * 2066 * If flags specifies AH_CONTAIN, the hole will contain the address addr. 2067 * 2068 * If an adequate hole is found, base and len are set to reflect the part of 2069 * the hole that is within range, and 0 is returned, otherwise, 2070 * -1 is returned. 2071 * 2072 * NOTE: This routine is not correct when base+len overflows caddr_t. 2073 */ 2074 int 2075 as_gap(struct as *as, size_t minlen, caddr_t *basep, size_t *lenp, uint_t flags, 2076 caddr_t addr) 2077 { 2078 2079 return (as_gap_aligned(as, minlen, basep, lenp, flags, addr, 0, 0, 0)); 2080 } 2081 2082 /* 2083 * Return the next range within [base, base + len) that is backed 2084 * with "real memory". Skip holes and non-seg_vn segments. 2085 * We're lazy and only return one segment at a time. 2086 */ 2087 int 2088 as_memory(struct as *as, caddr_t *basep, size_t *lenp) 2089 { 2090 extern struct seg_ops segspt_shmops; /* needs a header file */ 2091 struct seg *seg; 2092 caddr_t addr, eaddr; 2093 caddr_t segend; 2094 2095 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 2096 2097 addr = *basep; 2098 eaddr = addr + *lenp; 2099 2100 seg = as_findseg(as, addr, 0); 2101 if (seg != NULL) 2102 addr = MAX(seg->s_base, addr); 2103 2104 for (;;) { 2105 if (seg == NULL || addr >= eaddr || eaddr <= seg->s_base) { 2106 AS_LOCK_EXIT(as, &as->a_lock); 2107 return (EINVAL); 2108 } 2109 2110 if (seg->s_ops == &segvn_ops) { 2111 segend = seg->s_base + seg->s_size; 2112 break; 2113 } 2114 2115 /* 2116 * We do ISM by looking into the private data 2117 * to determine the real size of the segment. 2118 */ 2119 if (seg->s_ops == &segspt_shmops) { 2120 segend = seg->s_base + spt_realsize(seg); 2121 if (addr < segend) 2122 break; 2123 } 2124 2125 seg = AS_SEGNEXT(as, seg); 2126 2127 if (seg != NULL) 2128 addr = seg->s_base; 2129 } 2130 2131 *basep = addr; 2132 2133 if (segend > eaddr) 2134 *lenp = eaddr - addr; 2135 else 2136 *lenp = segend - addr; 2137 2138 AS_LOCK_EXIT(as, &as->a_lock); 2139 return (0); 2140 } 2141 2142 /* 2143 * Determine whether data from the mappings in interval [addr, addr + size) 2144 * are in the primary memory (core) cache. 2145 */ 2146 int 2147 as_incore(struct as *as, caddr_t addr, 2148 size_t size, char *vec, size_t *sizep) 2149 { 2150 struct seg *seg; 2151 size_t ssize; 2152 caddr_t raddr; /* rounded down addr */ 2153 size_t rsize; /* rounded up size */ 2154 size_t isize; /* iteration size */ 2155 int error = 0; /* result, assume success */ 2156 2157 *sizep = 0; 2158 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 2159 rsize = ((((size_t)addr + size) + PAGEOFFSET) & PAGEMASK) - 2160 (size_t)raddr; 2161 2162 if (raddr + rsize < raddr) /* check for wraparound */ 2163 return (ENOMEM); 2164 2165 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 2166 seg = as_segat(as, raddr); 2167 if (seg == NULL) { 2168 AS_LOCK_EXIT(as, &as->a_lock); 2169 return (-1); 2170 } 2171 2172 for (; rsize != 0; rsize -= ssize, raddr += ssize) { 2173 if (raddr >= seg->s_base + seg->s_size) { 2174 seg = AS_SEGNEXT(as, seg); 2175 if (seg == NULL || raddr != seg->s_base) { 2176 error = -1; 2177 break; 2178 } 2179 } 2180 if ((raddr + rsize) > (seg->s_base + seg->s_size)) 2181 ssize = seg->s_base + seg->s_size - raddr; 2182 else 2183 ssize = rsize; 2184 *sizep += isize = SEGOP_INCORE(seg, raddr, ssize, vec); 2185 if (isize != ssize) { 2186 error = -1; 2187 break; 2188 } 2189 vec += btopr(ssize); 2190 } 2191 AS_LOCK_EXIT(as, &as->a_lock); 2192 return (error); 2193 } 2194 2195 static void 2196 as_segunlock(struct seg *seg, caddr_t addr, int attr, 2197 ulong_t *bitmap, size_t position, size_t npages) 2198 { 2199 caddr_t range_start; 2200 size_t pos1 = position; 2201 size_t pos2; 2202 size_t size; 2203 size_t end_pos = npages + position; 2204 2205 while (bt_range(bitmap, &pos1, &pos2, end_pos)) { 2206 size = ptob((pos2 - pos1)); 2207 range_start = (caddr_t)((uintptr_t)addr + 2208 ptob(pos1 - position)); 2209 2210 (void) SEGOP_LOCKOP(seg, range_start, size, attr, MC_UNLOCK, 2211 (ulong_t *)NULL, (size_t)NULL); 2212 pos1 = pos2; 2213 } 2214 } 2215 2216 static void 2217 as_unlockerr(struct as *as, int attr, ulong_t *mlock_map, 2218 caddr_t raddr, size_t rsize) 2219 { 2220 struct seg *seg = as_segat(as, raddr); 2221 size_t ssize; 2222 2223 while (rsize != 0) { 2224 if (raddr >= seg->s_base + seg->s_size) 2225 seg = AS_SEGNEXT(as, seg); 2226 2227 if ((raddr + rsize) > (seg->s_base + seg->s_size)) 2228 ssize = seg->s_base + seg->s_size - raddr; 2229 else 2230 ssize = rsize; 2231 2232 as_segunlock(seg, raddr, attr, mlock_map, 0, btopr(ssize)); 2233 2234 rsize -= ssize; 2235 raddr += ssize; 2236 } 2237 } 2238 2239 /* 2240 * Cache control operations over the interval [addr, addr + size) in 2241 * address space "as". 2242 */ 2243 /*ARGSUSED*/ 2244 int 2245 as_ctl(struct as *as, caddr_t addr, size_t size, int func, int attr, 2246 uintptr_t arg, ulong_t *lock_map, size_t pos) 2247 { 2248 struct seg *seg; /* working segment */ 2249 caddr_t raddr; /* rounded down addr */ 2250 caddr_t initraddr; /* saved initial rounded down addr */ 2251 size_t rsize; /* rounded up size */ 2252 size_t initrsize; /* saved initial rounded up size */ 2253 size_t ssize; /* size of seg */ 2254 int error = 0; /* result */ 2255 size_t mlock_size; /* size of bitmap */ 2256 ulong_t *mlock_map; /* pointer to bitmap used */ 2257 /* to represent the locked */ 2258 /* pages. */ 2259 retry: 2260 if (error == IE_RETRY) 2261 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 2262 else 2263 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 2264 2265 /* 2266 * If these are address space lock/unlock operations, loop over 2267 * all segments in the address space, as appropriate. 2268 */ 2269 if (func == MC_LOCKAS) { 2270 size_t npages, idx; 2271 size_t rlen = 0; /* rounded as length */ 2272 2273 idx = pos; 2274 2275 if (arg & MCL_FUTURE) { 2276 mutex_enter(&as->a_contents); 2277 AS_SETPGLCK(as); 2278 mutex_exit(&as->a_contents); 2279 } 2280 if ((arg & MCL_CURRENT) == 0) { 2281 AS_LOCK_EXIT(as, &as->a_lock); 2282 return (0); 2283 } 2284 2285 seg = AS_SEGFIRST(as); 2286 if (seg == NULL) { 2287 AS_LOCK_EXIT(as, &as->a_lock); 2288 return (0); 2289 } 2290 2291 do { 2292 raddr = (caddr_t)((uintptr_t)seg->s_base & 2293 (uintptr_t)PAGEMASK); 2294 rlen += (((uintptr_t)(seg->s_base + seg->s_size) + 2295 PAGEOFFSET) & PAGEMASK) - (uintptr_t)raddr; 2296 } while ((seg = AS_SEGNEXT(as, seg)) != NULL); 2297 2298 mlock_size = BT_BITOUL(btopr(rlen)); 2299 if ((mlock_map = (ulong_t *)kmem_zalloc(mlock_size * 2300 sizeof (ulong_t), KM_NOSLEEP)) == NULL) { 2301 AS_LOCK_EXIT(as, &as->a_lock); 2302 return (EAGAIN); 2303 } 2304 2305 for (seg = AS_SEGFIRST(as); seg; seg = AS_SEGNEXT(as, seg)) { 2306 error = SEGOP_LOCKOP(seg, seg->s_base, 2307 seg->s_size, attr, MC_LOCK, mlock_map, pos); 2308 if (error != 0) 2309 break; 2310 pos += seg_pages(seg); 2311 } 2312 2313 if (error) { 2314 for (seg = AS_SEGFIRST(as); seg != NULL; 2315 seg = AS_SEGNEXT(as, seg)) { 2316 2317 raddr = (caddr_t)((uintptr_t)seg->s_base & 2318 (uintptr_t)PAGEMASK); 2319 npages = seg_pages(seg); 2320 as_segunlock(seg, raddr, attr, mlock_map, 2321 idx, npages); 2322 idx += npages; 2323 } 2324 } 2325 2326 kmem_free(mlock_map, mlock_size * sizeof (ulong_t)); 2327 AS_LOCK_EXIT(as, &as->a_lock); 2328 goto lockerr; 2329 } else if (func == MC_UNLOCKAS) { 2330 mutex_enter(&as->a_contents); 2331 AS_CLRPGLCK(as); 2332 mutex_exit(&as->a_contents); 2333 2334 for (seg = AS_SEGFIRST(as); seg; seg = AS_SEGNEXT(as, seg)) { 2335 error = SEGOP_LOCKOP(seg, seg->s_base, 2336 seg->s_size, attr, MC_UNLOCK, NULL, 0); 2337 if (error != 0) 2338 break; 2339 } 2340 2341 AS_LOCK_EXIT(as, &as->a_lock); 2342 goto lockerr; 2343 } 2344 2345 /* 2346 * Normalize addresses and sizes. 2347 */ 2348 initraddr = raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 2349 initrsize = rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) - 2350 (size_t)raddr; 2351 2352 if (raddr + rsize < raddr) { /* check for wraparound */ 2353 AS_LOCK_EXIT(as, &as->a_lock); 2354 return (ENOMEM); 2355 } 2356 2357 /* 2358 * Get initial segment. 2359 */ 2360 if ((seg = as_segat(as, raddr)) == NULL) { 2361 AS_LOCK_EXIT(as, &as->a_lock); 2362 return (ENOMEM); 2363 } 2364 2365 if (func == MC_LOCK) { 2366 mlock_size = BT_BITOUL(btopr(rsize)); 2367 if ((mlock_map = (ulong_t *)kmem_zalloc(mlock_size * 2368 sizeof (ulong_t), KM_NOSLEEP)) == NULL) { 2369 AS_LOCK_EXIT(as, &as->a_lock); 2370 return (EAGAIN); 2371 } 2372 } 2373 2374 /* 2375 * Loop over all segments. If a hole in the address range is 2376 * discovered, then fail. For each segment, perform the appropriate 2377 * control operation. 2378 */ 2379 while (rsize != 0) { 2380 2381 /* 2382 * Make sure there's no hole, calculate the portion 2383 * of the next segment to be operated over. 2384 */ 2385 if (raddr >= seg->s_base + seg->s_size) { 2386 seg = AS_SEGNEXT(as, seg); 2387 if (seg == NULL || raddr != seg->s_base) { 2388 if (func == MC_LOCK) { 2389 as_unlockerr(as, attr, mlock_map, 2390 initraddr, initrsize - rsize); 2391 kmem_free(mlock_map, 2392 mlock_size * sizeof (ulong_t)); 2393 } 2394 AS_LOCK_EXIT(as, &as->a_lock); 2395 return (ENOMEM); 2396 } 2397 } 2398 if ((raddr + rsize) > (seg->s_base + seg->s_size)) 2399 ssize = seg->s_base + seg->s_size - raddr; 2400 else 2401 ssize = rsize; 2402 2403 /* 2404 * Dispatch on specific function. 2405 */ 2406 switch (func) { 2407 2408 /* 2409 * Synchronize cached data from mappings with backing 2410 * objects. 2411 */ 2412 case MC_SYNC: 2413 if (error = SEGOP_SYNC(seg, raddr, ssize, 2414 attr, (uint_t)arg)) { 2415 AS_LOCK_EXIT(as, &as->a_lock); 2416 return (error); 2417 } 2418 break; 2419 2420 /* 2421 * Lock pages in memory. 2422 */ 2423 case MC_LOCK: 2424 if (error = SEGOP_LOCKOP(seg, raddr, ssize, 2425 attr, func, mlock_map, pos)) { 2426 as_unlockerr(as, attr, mlock_map, initraddr, 2427 initrsize - rsize + ssize); 2428 kmem_free(mlock_map, mlock_size * 2429 sizeof (ulong_t)); 2430 AS_LOCK_EXIT(as, &as->a_lock); 2431 goto lockerr; 2432 } 2433 break; 2434 2435 /* 2436 * Unlock mapped pages. 2437 */ 2438 case MC_UNLOCK: 2439 (void) SEGOP_LOCKOP(seg, raddr, ssize, attr, func, 2440 (ulong_t *)NULL, (size_t)NULL); 2441 break; 2442 2443 /* 2444 * Store VM advise for mapped pages in segment layer. 2445 */ 2446 case MC_ADVISE: 2447 error = SEGOP_ADVISE(seg, raddr, ssize, (uint_t)arg); 2448 2449 /* 2450 * Check for regular errors and special retry error 2451 */ 2452 if (error) { 2453 if (error == IE_RETRY) { 2454 /* 2455 * Need to acquire writers lock, so 2456 * have to drop readers lock and start 2457 * all over again 2458 */ 2459 AS_LOCK_EXIT(as, &as->a_lock); 2460 goto retry; 2461 } else if (error == IE_REATTACH) { 2462 /* 2463 * Find segment for current address 2464 * because current segment just got 2465 * split or concatenated 2466 */ 2467 seg = as_segat(as, raddr); 2468 if (seg == NULL) { 2469 AS_LOCK_EXIT(as, &as->a_lock); 2470 return (ENOMEM); 2471 } 2472 } else { 2473 /* 2474 * Regular error 2475 */ 2476 AS_LOCK_EXIT(as, &as->a_lock); 2477 return (error); 2478 } 2479 } 2480 break; 2481 2482 /* 2483 * Can't happen. 2484 */ 2485 default: 2486 panic("as_ctl: bad operation %d", func); 2487 /*NOTREACHED*/ 2488 } 2489 2490 rsize -= ssize; 2491 raddr += ssize; 2492 } 2493 2494 if (func == MC_LOCK) 2495 kmem_free(mlock_map, mlock_size * sizeof (ulong_t)); 2496 AS_LOCK_EXIT(as, &as->a_lock); 2497 return (0); 2498 lockerr: 2499 2500 /* 2501 * If the lower levels returned EDEADLK for a segment lockop, 2502 * it means that we should retry the operation. Let's wait 2503 * a bit also to let the deadlock causing condition clear. 2504 * This is part of a gross hack to work around a design flaw 2505 * in the ufs/sds logging code and should go away when the 2506 * logging code is re-designed to fix the problem. See bug 2507 * 4125102 for details of the problem. 2508 */ 2509 if (error == EDEADLK) { 2510 delay(deadlk_wait); 2511 error = 0; 2512 goto retry; 2513 } 2514 return (error); 2515 } 2516 2517 int 2518 fc_decode(faultcode_t fault_err) 2519 { 2520 int error = 0; 2521 2522 switch (FC_CODE(fault_err)) { 2523 case FC_OBJERR: 2524 error = FC_ERRNO(fault_err); 2525 break; 2526 case FC_PROT: 2527 error = EACCES; 2528 break; 2529 default: 2530 error = EFAULT; 2531 break; 2532 } 2533 return (error); 2534 } 2535 2536 /* 2537 * Pagelock pages from a range that spans more than 1 segment. Obtain shadow 2538 * lists from each segment and copy them to one contiguous shadow list (plist) 2539 * as expected by the caller. Save pointers to per segment shadow lists at 2540 * the tail of plist so that they can be used during as_pageunlock(). 2541 */ 2542 static int 2543 as_pagelock_segs(struct as *as, struct seg *seg, struct page ***ppp, 2544 caddr_t addr, size_t size, enum seg_rw rw) 2545 { 2546 caddr_t sv_addr = addr; 2547 size_t sv_size = size; 2548 struct seg *sv_seg = seg; 2549 ulong_t segcnt = 1; 2550 ulong_t cnt; 2551 size_t ssize; 2552 pgcnt_t npages = btop(size); 2553 page_t **plist; 2554 page_t **pl; 2555 int error; 2556 caddr_t eaddr; 2557 faultcode_t fault_err = 0; 2558 pgcnt_t pl_off; 2559 extern struct seg_ops segspt_shmops; 2560 2561 ASSERT(AS_LOCK_HELD(as, &as->a_lock)); 2562 ASSERT(seg != NULL); 2563 ASSERT(addr >= seg->s_base && addr < seg->s_base + seg->s_size); 2564 ASSERT(addr + size > seg->s_base + seg->s_size); 2565 ASSERT(IS_P2ALIGNED(size, PAGESIZE)); 2566 ASSERT(IS_P2ALIGNED(addr, PAGESIZE)); 2567 2568 /* 2569 * Count the number of segments covered by the range we are about to 2570 * lock. The segment count is used to size the shadow list we return 2571 * back to the caller. 2572 */ 2573 for (; size != 0; size -= ssize, addr += ssize) { 2574 if (addr >= seg->s_base + seg->s_size) { 2575 2576 seg = AS_SEGNEXT(as, seg); 2577 if (seg == NULL || addr != seg->s_base) { 2578 AS_LOCK_EXIT(as, &as->a_lock); 2579 return (EFAULT); 2580 } 2581 /* 2582 * Do a quick check if subsequent segments 2583 * will most likely support pagelock. 2584 */ 2585 if (seg->s_ops == &segvn_ops) { 2586 vnode_t *vp; 2587 2588 if (SEGOP_GETVP(seg, addr, &vp) != 0 || 2589 vp != NULL) { 2590 AS_LOCK_EXIT(as, &as->a_lock); 2591 goto slow; 2592 } 2593 } else if (seg->s_ops != &segspt_shmops) { 2594 AS_LOCK_EXIT(as, &as->a_lock); 2595 goto slow; 2596 } 2597 segcnt++; 2598 } 2599 if (addr + size > seg->s_base + seg->s_size) { 2600 ssize = seg->s_base + seg->s_size - addr; 2601 } else { 2602 ssize = size; 2603 } 2604 } 2605 ASSERT(segcnt > 1); 2606 2607 plist = kmem_zalloc((npages + segcnt) * sizeof (page_t *), KM_SLEEP); 2608 2609 addr = sv_addr; 2610 size = sv_size; 2611 seg = sv_seg; 2612 2613 for (cnt = 0, pl_off = 0; size != 0; size -= ssize, addr += ssize) { 2614 if (addr >= seg->s_base + seg->s_size) { 2615 seg = AS_SEGNEXT(as, seg); 2616 ASSERT(seg != NULL && addr == seg->s_base); 2617 cnt++; 2618 ASSERT(cnt < segcnt); 2619 } 2620 if (addr + size > seg->s_base + seg->s_size) { 2621 ssize = seg->s_base + seg->s_size - addr; 2622 } else { 2623 ssize = size; 2624 } 2625 pl = &plist[npages + cnt]; 2626 error = SEGOP_PAGELOCK(seg, addr, ssize, (page_t ***)pl, 2627 L_PAGELOCK, rw); 2628 if (error) { 2629 break; 2630 } 2631 ASSERT(plist[npages + cnt] != NULL); 2632 ASSERT(pl_off + btop(ssize) <= npages); 2633 bcopy(plist[npages + cnt], &plist[pl_off], 2634 btop(ssize) * sizeof (page_t *)); 2635 pl_off += btop(ssize); 2636 } 2637 2638 if (size == 0) { 2639 AS_LOCK_EXIT(as, &as->a_lock); 2640 ASSERT(cnt == segcnt - 1); 2641 *ppp = plist; 2642 return (0); 2643 } 2644 2645 /* 2646 * one of pagelock calls failed. The error type is in error variable. 2647 * Unlock what we've locked so far and retry with F_SOFTLOCK if error 2648 * type is either EFAULT or ENOTSUP. Otherwise just return the error 2649 * back to the caller. 2650 */ 2651 2652 eaddr = addr; 2653 seg = sv_seg; 2654 2655 for (cnt = 0, addr = sv_addr; addr < eaddr; addr += ssize) { 2656 if (addr >= seg->s_base + seg->s_size) { 2657 seg = AS_SEGNEXT(as, seg); 2658 ASSERT(seg != NULL && addr == seg->s_base); 2659 cnt++; 2660 ASSERT(cnt < segcnt); 2661 } 2662 if (eaddr > seg->s_base + seg->s_size) { 2663 ssize = seg->s_base + seg->s_size - addr; 2664 } else { 2665 ssize = eaddr - addr; 2666 } 2667 pl = &plist[npages + cnt]; 2668 ASSERT(*pl != NULL); 2669 (void) SEGOP_PAGELOCK(seg, addr, ssize, (page_t ***)pl, 2670 L_PAGEUNLOCK, rw); 2671 } 2672 2673 AS_LOCK_EXIT(as, &as->a_lock); 2674 2675 kmem_free(plist, (npages + segcnt) * sizeof (page_t *)); 2676 2677 if (error != ENOTSUP && error != EFAULT) { 2678 return (error); 2679 } 2680 2681 slow: 2682 /* 2683 * If we are here because pagelock failed due to the need to cow fault 2684 * in the pages we want to lock F_SOFTLOCK will do this job and in 2685 * next as_pagelock() call for this address range pagelock will 2686 * hopefully succeed. 2687 */ 2688 fault_err = as_fault(as->a_hat, as, sv_addr, sv_size, F_SOFTLOCK, rw); 2689 if (fault_err != 0) { 2690 return (fc_decode(fault_err)); 2691 } 2692 *ppp = NULL; 2693 2694 return (0); 2695 } 2696 2697 /* 2698 * lock pages in a given address space. Return shadow list. If 2699 * the list is NULL, the MMU mapping is also locked. 2700 */ 2701 int 2702 as_pagelock(struct as *as, struct page ***ppp, caddr_t addr, 2703 size_t size, enum seg_rw rw) 2704 { 2705 size_t rsize; 2706 caddr_t raddr; 2707 faultcode_t fault_err; 2708 struct seg *seg; 2709 int err; 2710 2711 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_AS_LOCK_START, 2712 "as_pagelock_start: addr %p size %ld", addr, size); 2713 2714 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 2715 rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) - 2716 (size_t)raddr; 2717 2718 /* 2719 * if the request crosses two segments let 2720 * as_fault handle it. 2721 */ 2722 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 2723 2724 seg = as_segat(as, raddr); 2725 if (seg == NULL) { 2726 AS_LOCK_EXIT(as, &as->a_lock); 2727 return (EFAULT); 2728 } 2729 ASSERT(raddr >= seg->s_base && raddr < seg->s_base + seg->s_size); 2730 if (raddr + rsize > seg->s_base + seg->s_size) { 2731 return (as_pagelock_segs(as, seg, ppp, raddr, rsize, rw)); 2732 } 2733 if (raddr + rsize <= raddr) { 2734 AS_LOCK_EXIT(as, &as->a_lock); 2735 return (EFAULT); 2736 } 2737 2738 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEG_LOCK_START, 2739 "seg_lock_1_start: raddr %p rsize %ld", raddr, rsize); 2740 2741 /* 2742 * try to lock pages and pass back shadow list 2743 */ 2744 err = SEGOP_PAGELOCK(seg, raddr, rsize, ppp, L_PAGELOCK, rw); 2745 2746 TRACE_0(TR_FAC_PHYSIO, TR_PHYSIO_SEG_LOCK_END, "seg_lock_1_end"); 2747 2748 AS_LOCK_EXIT(as, &as->a_lock); 2749 2750 if (err == 0 || (err != ENOTSUP && err != EFAULT)) { 2751 return (err); 2752 } 2753 2754 /* 2755 * Use F_SOFTLOCK to lock the pages because pagelock failed either due 2756 * to no pagelock support for this segment or pages need to be cow 2757 * faulted in. If fault is needed F_SOFTLOCK will do this job for 2758 * this as_pagelock() call and in the next as_pagelock() call for the 2759 * same address range pagelock call will hopefull succeed. 2760 */ 2761 fault_err = as_fault(as->a_hat, as, addr, size, F_SOFTLOCK, rw); 2762 if (fault_err != 0) { 2763 return (fc_decode(fault_err)); 2764 } 2765 *ppp = NULL; 2766 2767 TRACE_0(TR_FAC_PHYSIO, TR_PHYSIO_AS_LOCK_END, "as_pagelock_end"); 2768 return (0); 2769 } 2770 2771 /* 2772 * unlock pages locked by as_pagelock_segs(). Retrieve per segment shadow 2773 * lists from the end of plist and call pageunlock interface for each segment. 2774 * Drop as lock and free plist. 2775 */ 2776 static void 2777 as_pageunlock_segs(struct as *as, struct seg *seg, caddr_t addr, size_t size, 2778 struct page **plist, enum seg_rw rw) 2779 { 2780 ulong_t cnt; 2781 caddr_t eaddr = addr + size; 2782 pgcnt_t npages = btop(size); 2783 size_t ssize; 2784 page_t **pl; 2785 2786 ASSERT(AS_LOCK_HELD(as, &as->a_lock)); 2787 ASSERT(seg != NULL); 2788 ASSERT(addr >= seg->s_base && addr < seg->s_base + seg->s_size); 2789 ASSERT(addr + size > seg->s_base + seg->s_size); 2790 ASSERT(IS_P2ALIGNED(size, PAGESIZE)); 2791 ASSERT(IS_P2ALIGNED(addr, PAGESIZE)); 2792 ASSERT(plist != NULL); 2793 2794 for (cnt = 0; addr < eaddr; addr += ssize) { 2795 if (addr >= seg->s_base + seg->s_size) { 2796 seg = AS_SEGNEXT(as, seg); 2797 ASSERT(seg != NULL && addr == seg->s_base); 2798 cnt++; 2799 } 2800 if (eaddr > seg->s_base + seg->s_size) { 2801 ssize = seg->s_base + seg->s_size - addr; 2802 } else { 2803 ssize = eaddr - addr; 2804 } 2805 pl = &plist[npages + cnt]; 2806 ASSERT(*pl != NULL); 2807 (void) SEGOP_PAGELOCK(seg, addr, ssize, (page_t ***)pl, 2808 L_PAGEUNLOCK, rw); 2809 } 2810 ASSERT(cnt > 0); 2811 AS_LOCK_EXIT(as, &as->a_lock); 2812 2813 cnt++; 2814 kmem_free(plist, (npages + cnt) * sizeof (page_t *)); 2815 } 2816 2817 /* 2818 * unlock pages in a given address range 2819 */ 2820 void 2821 as_pageunlock(struct as *as, struct page **pp, caddr_t addr, size_t size, 2822 enum seg_rw rw) 2823 { 2824 struct seg *seg; 2825 size_t rsize; 2826 caddr_t raddr; 2827 2828 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_AS_UNLOCK_START, 2829 "as_pageunlock_start: addr %p size %ld", addr, size); 2830 2831 /* 2832 * if the shadow list is NULL, as_pagelock was 2833 * falling back to as_fault 2834 */ 2835 if (pp == NULL) { 2836 (void) as_fault(as->a_hat, as, addr, size, F_SOFTUNLOCK, rw); 2837 return; 2838 } 2839 2840 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 2841 rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) - 2842 (size_t)raddr; 2843 2844 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 2845 seg = as_segat(as, raddr); 2846 ASSERT(seg != NULL); 2847 2848 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEG_UNLOCK_START, 2849 "seg_unlock_start: raddr %p rsize %ld", raddr, rsize); 2850 2851 ASSERT(raddr >= seg->s_base && raddr < seg->s_base + seg->s_size); 2852 if (raddr + rsize <= seg->s_base + seg->s_size) { 2853 SEGOP_PAGELOCK(seg, raddr, rsize, &pp, L_PAGEUNLOCK, rw); 2854 } else { 2855 as_pageunlock_segs(as, seg, raddr, rsize, pp, rw); 2856 return; 2857 } 2858 AS_LOCK_EXIT(as, &as->a_lock); 2859 TRACE_0(TR_FAC_PHYSIO, TR_PHYSIO_AS_UNLOCK_END, "as_pageunlock_end"); 2860 } 2861 2862 int 2863 as_setpagesize(struct as *as, caddr_t addr, size_t size, uint_t szc, 2864 boolean_t wait) 2865 { 2866 struct seg *seg; 2867 size_t ssize; 2868 caddr_t raddr; /* rounded down addr */ 2869 size_t rsize; /* rounded up size */ 2870 int error = 0; 2871 size_t pgsz = page_get_pagesize(szc); 2872 2873 setpgsz_top: 2874 if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(size, pgsz)) { 2875 return (EINVAL); 2876 } 2877 2878 raddr = addr; 2879 rsize = size; 2880 2881 if (raddr + rsize < raddr) /* check for wraparound */ 2882 return (ENOMEM); 2883 2884 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 2885 as_clearwatchprot(as, raddr, rsize); 2886 seg = as_segat(as, raddr); 2887 if (seg == NULL) { 2888 as_setwatch(as); 2889 AS_LOCK_EXIT(as, &as->a_lock); 2890 return (ENOMEM); 2891 } 2892 2893 for (; rsize != 0; rsize -= ssize, raddr += ssize) { 2894 if (raddr >= seg->s_base + seg->s_size) { 2895 seg = AS_SEGNEXT(as, seg); 2896 if (seg == NULL || raddr != seg->s_base) { 2897 error = ENOMEM; 2898 break; 2899 } 2900 } 2901 if ((raddr + rsize) > (seg->s_base + seg->s_size)) { 2902 ssize = seg->s_base + seg->s_size - raddr; 2903 } else { 2904 ssize = rsize; 2905 } 2906 2907 retry: 2908 error = SEGOP_SETPAGESIZE(seg, raddr, ssize, szc); 2909 2910 if (error == IE_NOMEM) { 2911 error = EAGAIN; 2912 break; 2913 } 2914 2915 if (error == IE_RETRY) { 2916 AS_LOCK_EXIT(as, &as->a_lock); 2917 goto setpgsz_top; 2918 } 2919 2920 if (error == ENOTSUP) { 2921 error = EINVAL; 2922 break; 2923 } 2924 2925 if (wait && (error == EAGAIN)) { 2926 /* 2927 * Memory is currently locked. It must be unlocked 2928 * before this operation can succeed through a retry. 2929 * The possible reasons for locked memory and 2930 * corresponding strategies for unlocking are: 2931 * (1) Normal I/O 2932 * wait for a signal that the I/O operation 2933 * has completed and the memory is unlocked. 2934 * (2) Asynchronous I/O 2935 * The aio subsystem does not unlock pages when 2936 * the I/O is completed. Those pages are unlocked 2937 * when the application calls aiowait/aioerror. 2938 * So, to prevent blocking forever, cv_broadcast() 2939 * is done to wake up aio_cleanup_thread. 2940 * Subsequently, segvn_reclaim will be called, and 2941 * that will do AS_CLRUNMAPWAIT() and wake us up. 2942 * (3) Long term page locking: 2943 * This is not relevant for as_setpagesize() 2944 * because we cannot change the page size for 2945 * driver memory. The attempt to do so will 2946 * fail with a different error than EAGAIN so 2947 * there's no need to trigger as callbacks like 2948 * as_unmap, as_setprot or as_free would do. 2949 */ 2950 mutex_enter(&as->a_contents); 2951 if (!AS_ISNOUNMAPWAIT(as)) { 2952 if (AS_ISUNMAPWAIT(as) == 0) { 2953 cv_broadcast(&as->a_cv); 2954 } 2955 AS_SETUNMAPWAIT(as); 2956 AS_LOCK_EXIT(as, &as->a_lock); 2957 while (AS_ISUNMAPWAIT(as)) { 2958 cv_wait(&as->a_cv, &as->a_contents); 2959 } 2960 } else { 2961 /* 2962 * We may have raced with 2963 * segvn_reclaim()/segspt_reclaim(). In this 2964 * case clean nounmapwait flag and retry since 2965 * softlockcnt in this segment may be already 2966 * 0. We don't drop as writer lock so our 2967 * number of retries without sleeping should 2968 * be very small. See segvn_reclaim() for 2969 * more comments. 2970 */ 2971 AS_CLRNOUNMAPWAIT(as); 2972 mutex_exit(&as->a_contents); 2973 goto retry; 2974 } 2975 mutex_exit(&as->a_contents); 2976 goto setpgsz_top; 2977 } else if (error != 0) { 2978 break; 2979 } 2980 } 2981 as_setwatch(as); 2982 AS_LOCK_EXIT(as, &as->a_lock); 2983 return (error); 2984 } 2985 2986 /* 2987 * as_iset3_default_lpsize() just calls SEGOP_SETPAGESIZE() on all segments 2988 * in its chunk where s_szc is less than the szc we want to set. 2989 */ 2990 static int 2991 as_iset3_default_lpsize(struct as *as, caddr_t raddr, size_t rsize, uint_t szc, 2992 int *retry) 2993 { 2994 struct seg *seg; 2995 size_t ssize; 2996 int error; 2997 2998 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 2999 3000 seg = as_segat(as, raddr); 3001 if (seg == NULL) { 3002 panic("as_iset3_default_lpsize: no seg"); 3003 } 3004 3005 for (; rsize != 0; rsize -= ssize, raddr += ssize) { 3006 if (raddr >= seg->s_base + seg->s_size) { 3007 seg = AS_SEGNEXT(as, seg); 3008 if (seg == NULL || raddr != seg->s_base) { 3009 panic("as_iset3_default_lpsize: as changed"); 3010 } 3011 } 3012 if ((raddr + rsize) > (seg->s_base + seg->s_size)) { 3013 ssize = seg->s_base + seg->s_size - raddr; 3014 } else { 3015 ssize = rsize; 3016 } 3017 3018 if (szc > seg->s_szc) { 3019 error = SEGOP_SETPAGESIZE(seg, raddr, ssize, szc); 3020 /* Only retry on EINVAL segments that have no vnode. */ 3021 if (error == EINVAL) { 3022 vnode_t *vp = NULL; 3023 if ((SEGOP_GETTYPE(seg, raddr) & MAP_SHARED) && 3024 (SEGOP_GETVP(seg, raddr, &vp) != 0 || 3025 vp == NULL)) { 3026 *retry = 1; 3027 } else { 3028 *retry = 0; 3029 } 3030 } 3031 if (error) { 3032 return (error); 3033 } 3034 } 3035 } 3036 return (0); 3037 } 3038 3039 /* 3040 * as_iset2_default_lpsize() calls as_iset3_default_lpsize() to set the 3041 * pagesize on each segment in its range, but if any fails with EINVAL, 3042 * then it reduces the pagesizes to the next size in the bitmap and 3043 * retries as_iset3_default_lpsize(). The reason why the code retries 3044 * smaller allowed sizes on EINVAL is because (a) the anon offset may not 3045 * match the bigger sizes, and (b) it's hard to get this offset (to begin 3046 * with) to pass to map_pgszcvec(). 3047 */ 3048 static int 3049 as_iset2_default_lpsize(struct as *as, caddr_t addr, size_t size, uint_t szc, 3050 uint_t szcvec) 3051 { 3052 int error; 3053 int retry; 3054 3055 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 3056 3057 for (;;) { 3058 error = as_iset3_default_lpsize(as, addr, size, szc, &retry); 3059 if (error == EINVAL && retry) { 3060 szcvec &= ~(1 << szc); 3061 if (szcvec <= 1) { 3062 return (EINVAL); 3063 } 3064 szc = highbit(szcvec) - 1; 3065 } else { 3066 return (error); 3067 } 3068 } 3069 } 3070 3071 /* 3072 * as_iset1_default_lpsize() breaks its chunk into areas where existing 3073 * segments have a smaller szc than we want to set. For each such area, 3074 * it calls as_iset2_default_lpsize() 3075 */ 3076 static int 3077 as_iset1_default_lpsize(struct as *as, caddr_t raddr, size_t rsize, uint_t szc, 3078 uint_t szcvec) 3079 { 3080 struct seg *seg; 3081 size_t ssize; 3082 caddr_t setaddr = raddr; 3083 size_t setsize = 0; 3084 int set; 3085 int error; 3086 3087 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 3088 3089 seg = as_segat(as, raddr); 3090 if (seg == NULL) { 3091 panic("as_iset1_default_lpsize: no seg"); 3092 } 3093 if (seg->s_szc < szc) { 3094 set = 1; 3095 } else { 3096 set = 0; 3097 } 3098 3099 for (; rsize != 0; rsize -= ssize, raddr += ssize, setsize += ssize) { 3100 if (raddr >= seg->s_base + seg->s_size) { 3101 seg = AS_SEGNEXT(as, seg); 3102 if (seg == NULL || raddr != seg->s_base) { 3103 panic("as_iset1_default_lpsize: as changed"); 3104 } 3105 if (seg->s_szc >= szc && set) { 3106 ASSERT(setsize != 0); 3107 error = as_iset2_default_lpsize(as, 3108 setaddr, setsize, szc, szcvec); 3109 if (error) { 3110 return (error); 3111 } 3112 set = 0; 3113 } else if (seg->s_szc < szc && !set) { 3114 setaddr = raddr; 3115 setsize = 0; 3116 set = 1; 3117 } 3118 } 3119 if ((raddr + rsize) > (seg->s_base + seg->s_size)) { 3120 ssize = seg->s_base + seg->s_size - raddr; 3121 } else { 3122 ssize = rsize; 3123 } 3124 } 3125 error = 0; 3126 if (set) { 3127 ASSERT(setsize != 0); 3128 error = as_iset2_default_lpsize(as, setaddr, setsize, 3129 szc, szcvec); 3130 } 3131 return (error); 3132 } 3133 3134 /* 3135 * as_iset_default_lpsize() breaks its chunk according to the size code bitmap 3136 * returned by map_pgszcvec() (similar to as_map_segvn_segs()), and passes each 3137 * chunk to as_iset1_default_lpsize(). 3138 */ 3139 static int 3140 as_iset_default_lpsize(struct as *as, caddr_t addr, size_t size, int flags, 3141 int type) 3142 { 3143 int rtype = (type & MAP_SHARED) ? MAPPGSZC_SHM : MAPPGSZC_PRIVM; 3144 uint_t szcvec = map_pgszcvec(addr, size, (uintptr_t)addr, 3145 flags, rtype, 1); 3146 uint_t szc; 3147 uint_t nszc; 3148 int error; 3149 caddr_t a; 3150 caddr_t eaddr; 3151 size_t segsize; 3152 size_t pgsz; 3153 uint_t save_szcvec; 3154 3155 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 3156 ASSERT(IS_P2ALIGNED(addr, PAGESIZE)); 3157 ASSERT(IS_P2ALIGNED(size, PAGESIZE)); 3158 3159 szcvec &= ~1; 3160 if (szcvec <= 1) { /* skip if base page size */ 3161 return (0); 3162 } 3163 3164 /* Get the pagesize of the first larger page size. */ 3165 szc = lowbit(szcvec) - 1; 3166 pgsz = page_get_pagesize(szc); 3167 eaddr = addr + size; 3168 addr = (caddr_t)P2ROUNDUP((uintptr_t)addr, pgsz); 3169 eaddr = (caddr_t)P2ALIGN((uintptr_t)eaddr, pgsz); 3170 3171 save_szcvec = szcvec; 3172 szcvec >>= (szc + 1); 3173 nszc = szc; 3174 while (szcvec) { 3175 if ((szcvec & 0x1) == 0) { 3176 nszc++; 3177 szcvec >>= 1; 3178 continue; 3179 } 3180 nszc++; 3181 pgsz = page_get_pagesize(nszc); 3182 a = (caddr_t)P2ROUNDUP((uintptr_t)addr, pgsz); 3183 if (a != addr) { 3184 ASSERT(szc > 0); 3185 ASSERT(a < eaddr); 3186 segsize = a - addr; 3187 error = as_iset1_default_lpsize(as, addr, segsize, szc, 3188 save_szcvec); 3189 if (error) { 3190 return (error); 3191 } 3192 addr = a; 3193 } 3194 szc = nszc; 3195 szcvec >>= 1; 3196 } 3197 3198 ASSERT(addr < eaddr); 3199 szcvec = save_szcvec; 3200 while (szcvec) { 3201 a = (caddr_t)P2ALIGN((uintptr_t)eaddr, pgsz); 3202 ASSERT(a >= addr); 3203 if (a != addr) { 3204 ASSERT(szc > 0); 3205 segsize = a - addr; 3206 error = as_iset1_default_lpsize(as, addr, segsize, szc, 3207 save_szcvec); 3208 if (error) { 3209 return (error); 3210 } 3211 addr = a; 3212 } 3213 szcvec &= ~(1 << szc); 3214 if (szcvec) { 3215 szc = highbit(szcvec) - 1; 3216 pgsz = page_get_pagesize(szc); 3217 } 3218 } 3219 ASSERT(addr == eaddr); 3220 3221 return (0); 3222 } 3223 3224 /* 3225 * Set the default large page size for the range. Called via memcntl with 3226 * page size set to 0. as_set_default_lpsize breaks the range down into 3227 * chunks with the same type/flags, ignores-non segvn segments, and passes 3228 * each chunk to as_iset_default_lpsize(). 3229 */ 3230 int 3231 as_set_default_lpsize(struct as *as, caddr_t addr, size_t size) 3232 { 3233 struct seg *seg; 3234 caddr_t raddr; 3235 size_t rsize; 3236 size_t ssize; 3237 int rtype, rflags; 3238 int stype, sflags; 3239 int error; 3240 caddr_t setaddr; 3241 size_t setsize; 3242 int segvn; 3243 3244 if (size == 0) 3245 return (0); 3246 3247 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 3248 again: 3249 error = 0; 3250 3251 raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 3252 rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) - 3253 (size_t)raddr; 3254 3255 if (raddr + rsize < raddr) { /* check for wraparound */ 3256 AS_LOCK_EXIT(as, &as->a_lock); 3257 return (ENOMEM); 3258 } 3259 as_clearwatchprot(as, raddr, rsize); 3260 seg = as_segat(as, raddr); 3261 if (seg == NULL) { 3262 as_setwatch(as); 3263 AS_LOCK_EXIT(as, &as->a_lock); 3264 return (ENOMEM); 3265 } 3266 if (seg->s_ops == &segvn_ops) { 3267 rtype = SEGOP_GETTYPE(seg, addr); 3268 rflags = rtype & (MAP_TEXT | MAP_INITDATA); 3269 rtype = rtype & (MAP_SHARED | MAP_PRIVATE); 3270 segvn = 1; 3271 } else { 3272 segvn = 0; 3273 } 3274 setaddr = raddr; 3275 setsize = 0; 3276 3277 for (; rsize != 0; rsize -= ssize, raddr += ssize, setsize += ssize) { 3278 if (raddr >= (seg->s_base + seg->s_size)) { 3279 seg = AS_SEGNEXT(as, seg); 3280 if (seg == NULL || raddr != seg->s_base) { 3281 error = ENOMEM; 3282 break; 3283 } 3284 if (seg->s_ops == &segvn_ops) { 3285 stype = SEGOP_GETTYPE(seg, raddr); 3286 sflags = stype & (MAP_TEXT | MAP_INITDATA); 3287 stype &= (MAP_SHARED | MAP_PRIVATE); 3288 if (segvn && (rflags != sflags || 3289 rtype != stype)) { 3290 /* 3291 * The next segment is also segvn but 3292 * has different flags and/or type. 3293 */ 3294 ASSERT(setsize != 0); 3295 error = as_iset_default_lpsize(as, 3296 setaddr, setsize, rflags, rtype); 3297 if (error) { 3298 break; 3299 } 3300 rflags = sflags; 3301 rtype = stype; 3302 setaddr = raddr; 3303 setsize = 0; 3304 } else if (!segvn) { 3305 rflags = sflags; 3306 rtype = stype; 3307 setaddr = raddr; 3308 setsize = 0; 3309 segvn = 1; 3310 } 3311 } else if (segvn) { 3312 /* The next segment is not segvn. */ 3313 ASSERT(setsize != 0); 3314 error = as_iset_default_lpsize(as, 3315 setaddr, setsize, rflags, rtype); 3316 if (error) { 3317 break; 3318 } 3319 segvn = 0; 3320 } 3321 } 3322 if ((raddr + rsize) > (seg->s_base + seg->s_size)) { 3323 ssize = seg->s_base + seg->s_size - raddr; 3324 } else { 3325 ssize = rsize; 3326 } 3327 } 3328 if (error == 0 && segvn) { 3329 /* The last chunk when rsize == 0. */ 3330 ASSERT(setsize != 0); 3331 error = as_iset_default_lpsize(as, setaddr, setsize, 3332 rflags, rtype); 3333 } 3334 3335 if (error == IE_RETRY) { 3336 goto again; 3337 } else if (error == IE_NOMEM) { 3338 error = EAGAIN; 3339 } else if (error == ENOTSUP) { 3340 error = EINVAL; 3341 } else if (error == EAGAIN) { 3342 mutex_enter(&as->a_contents); 3343 if (!AS_ISNOUNMAPWAIT(as)) { 3344 if (AS_ISUNMAPWAIT(as) == 0) { 3345 cv_broadcast(&as->a_cv); 3346 } 3347 AS_SETUNMAPWAIT(as); 3348 AS_LOCK_EXIT(as, &as->a_lock); 3349 while (AS_ISUNMAPWAIT(as)) { 3350 cv_wait(&as->a_cv, &as->a_contents); 3351 } 3352 mutex_exit(&as->a_contents); 3353 AS_LOCK_ENTER(as, &as->a_lock, RW_WRITER); 3354 } else { 3355 /* 3356 * We may have raced with 3357 * segvn_reclaim()/segspt_reclaim(). In this case 3358 * clean nounmapwait flag and retry since softlockcnt 3359 * in this segment may be already 0. We don't drop as 3360 * writer lock so our number of retries without 3361 * sleeping should be very small. See segvn_reclaim() 3362 * for more comments. 3363 */ 3364 AS_CLRNOUNMAPWAIT(as); 3365 mutex_exit(&as->a_contents); 3366 } 3367 goto again; 3368 } 3369 3370 as_setwatch(as); 3371 AS_LOCK_EXIT(as, &as->a_lock); 3372 return (error); 3373 } 3374 3375 /* 3376 * Setup all of the uninitialized watched pages that we can. 3377 */ 3378 void 3379 as_setwatch(struct as *as) 3380 { 3381 struct watched_page *pwp; 3382 struct seg *seg; 3383 caddr_t vaddr; 3384 uint_t prot; 3385 int err, retrycnt; 3386 3387 if (avl_numnodes(&as->a_wpage) == 0) 3388 return; 3389 3390 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 3391 3392 for (pwp = avl_first(&as->a_wpage); pwp != NULL; 3393 pwp = AVL_NEXT(&as->a_wpage, pwp)) { 3394 retrycnt = 0; 3395 retry: 3396 vaddr = pwp->wp_vaddr; 3397 if (pwp->wp_oprot != 0 || /* already set up */ 3398 (seg = as_segat(as, vaddr)) == NULL || 3399 SEGOP_GETPROT(seg, vaddr, 0, &prot) != 0) 3400 continue; 3401 3402 pwp->wp_oprot = prot; 3403 if (pwp->wp_read) 3404 prot &= ~(PROT_READ|PROT_WRITE|PROT_EXEC); 3405 if (pwp->wp_write) 3406 prot &= ~PROT_WRITE; 3407 if (pwp->wp_exec) 3408 prot &= ~(PROT_READ|PROT_WRITE|PROT_EXEC); 3409 if (!(pwp->wp_flags & WP_NOWATCH) && prot != pwp->wp_oprot) { 3410 err = SEGOP_SETPROT(seg, vaddr, PAGESIZE, prot); 3411 if (err == IE_RETRY) { 3412 pwp->wp_oprot = 0; 3413 ASSERT(retrycnt == 0); 3414 retrycnt++; 3415 goto retry; 3416 } 3417 } 3418 pwp->wp_prot = prot; 3419 } 3420 } 3421 3422 /* 3423 * Clear all of the watched pages in the address space. 3424 */ 3425 void 3426 as_clearwatch(struct as *as) 3427 { 3428 struct watched_page *pwp; 3429 struct seg *seg; 3430 caddr_t vaddr; 3431 uint_t prot; 3432 int err, retrycnt; 3433 3434 if (avl_numnodes(&as->a_wpage) == 0) 3435 return; 3436 3437 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 3438 3439 for (pwp = avl_first(&as->a_wpage); pwp != NULL; 3440 pwp = AVL_NEXT(&as->a_wpage, pwp)) { 3441 retrycnt = 0; 3442 retry: 3443 vaddr = pwp->wp_vaddr; 3444 if (pwp->wp_oprot == 0 || /* not set up */ 3445 (seg = as_segat(as, vaddr)) == NULL) 3446 continue; 3447 3448 if ((prot = pwp->wp_oprot) != pwp->wp_prot) { 3449 err = SEGOP_SETPROT(seg, vaddr, PAGESIZE, prot); 3450 if (err == IE_RETRY) { 3451 ASSERT(retrycnt == 0); 3452 retrycnt++; 3453 goto retry; 3454 } 3455 } 3456 pwp->wp_oprot = 0; 3457 pwp->wp_prot = 0; 3458 } 3459 } 3460 3461 /* 3462 * Force a new setup for all the watched pages in the range. 3463 */ 3464 static void 3465 as_setwatchprot(struct as *as, caddr_t addr, size_t size, uint_t prot) 3466 { 3467 struct watched_page *pwp; 3468 struct watched_page tpw; 3469 caddr_t eaddr = addr + size; 3470 caddr_t vaddr; 3471 struct seg *seg; 3472 int err, retrycnt; 3473 uint_t wprot; 3474 avl_index_t where; 3475 3476 if (avl_numnodes(&as->a_wpage) == 0) 3477 return; 3478 3479 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 3480 3481 tpw.wp_vaddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 3482 if ((pwp = avl_find(&as->a_wpage, &tpw, &where)) == NULL) 3483 pwp = avl_nearest(&as->a_wpage, where, AVL_AFTER); 3484 3485 while (pwp != NULL && pwp->wp_vaddr < eaddr) { 3486 retrycnt = 0; 3487 vaddr = pwp->wp_vaddr; 3488 3489 wprot = prot; 3490 if (pwp->wp_read) 3491 wprot &= ~(PROT_READ|PROT_WRITE|PROT_EXEC); 3492 if (pwp->wp_write) 3493 wprot &= ~PROT_WRITE; 3494 if (pwp->wp_exec) 3495 wprot &= ~(PROT_READ|PROT_WRITE|PROT_EXEC); 3496 if (!(pwp->wp_flags & WP_NOWATCH) && wprot != pwp->wp_oprot) { 3497 retry: 3498 seg = as_segat(as, vaddr); 3499 if (seg == NULL) { 3500 panic("as_setwatchprot: no seg"); 3501 /*NOTREACHED*/ 3502 } 3503 err = SEGOP_SETPROT(seg, vaddr, PAGESIZE, wprot); 3504 if (err == IE_RETRY) { 3505 ASSERT(retrycnt == 0); 3506 retrycnt++; 3507 goto retry; 3508 } 3509 } 3510 pwp->wp_oprot = prot; 3511 pwp->wp_prot = wprot; 3512 3513 pwp = AVL_NEXT(&as->a_wpage, pwp); 3514 } 3515 } 3516 3517 /* 3518 * Clear all of the watched pages in the range. 3519 */ 3520 static void 3521 as_clearwatchprot(struct as *as, caddr_t addr, size_t size) 3522 { 3523 caddr_t eaddr = addr + size; 3524 struct watched_page *pwp; 3525 struct watched_page tpw; 3526 uint_t prot; 3527 struct seg *seg; 3528 int err, retrycnt; 3529 avl_index_t where; 3530 3531 if (avl_numnodes(&as->a_wpage) == 0) 3532 return; 3533 3534 tpw.wp_vaddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK); 3535 if ((pwp = avl_find(&as->a_wpage, &tpw, &where)) == NULL) 3536 pwp = avl_nearest(&as->a_wpage, where, AVL_AFTER); 3537 3538 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 3539 3540 while (pwp != NULL && pwp->wp_vaddr < eaddr) { 3541 3542 if ((prot = pwp->wp_oprot) != 0) { 3543 retrycnt = 0; 3544 3545 if (prot != pwp->wp_prot) { 3546 retry: 3547 seg = as_segat(as, pwp->wp_vaddr); 3548 if (seg == NULL) 3549 continue; 3550 err = SEGOP_SETPROT(seg, pwp->wp_vaddr, 3551 PAGESIZE, prot); 3552 if (err == IE_RETRY) { 3553 ASSERT(retrycnt == 0); 3554 retrycnt++; 3555 goto retry; 3556 3557 } 3558 } 3559 pwp->wp_oprot = 0; 3560 pwp->wp_prot = 0; 3561 } 3562 3563 pwp = AVL_NEXT(&as->a_wpage, pwp); 3564 } 3565 } 3566 3567 void 3568 as_signal_proc(struct as *as, k_siginfo_t *siginfo) 3569 { 3570 struct proc *p; 3571 3572 mutex_enter(&pidlock); 3573 for (p = practive; p; p = p->p_next) { 3574 if (p->p_as == as) { 3575 mutex_enter(&p->p_lock); 3576 if (p->p_as == as) 3577 sigaddq(p, NULL, siginfo, KM_NOSLEEP); 3578 mutex_exit(&p->p_lock); 3579 } 3580 } 3581 mutex_exit(&pidlock); 3582 } 3583 3584 /* 3585 * return memory object ID 3586 */ 3587 int 3588 as_getmemid(struct as *as, caddr_t addr, memid_t *memidp) 3589 { 3590 struct seg *seg; 3591 int sts; 3592 3593 AS_LOCK_ENTER(as, &as->a_lock, RW_READER); 3594 seg = as_segat(as, addr); 3595 if (seg == NULL) { 3596 AS_LOCK_EXIT(as, &as->a_lock); 3597 return (EFAULT); 3598 } 3599 /* 3600 * catch old drivers which may not support getmemid 3601 */ 3602 if (seg->s_ops->getmemid == NULL) { 3603 AS_LOCK_EXIT(as, &as->a_lock); 3604 return (ENODEV); 3605 } 3606 3607 sts = SEGOP_GETMEMID(seg, addr, memidp); 3608 3609 AS_LOCK_EXIT(as, &as->a_lock); 3610 return (sts); 3611 }