1 /* $NetBSD: uvm_fault.c,v 1.90 2005/02/07 11:57:38 yamt Exp $ */ 2 3 /* 4 * 5 * Copyright (c) 1997 Charles D. Cranor and Washington University. 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by Charles D. Cranor and 19 * Washington University. 20 * 4. The name of the author may not be used to endorse or promote products 21 * derived from this software without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 24 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 25 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 26 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 27 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 28 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 29 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 30 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 32 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 33 * 34 * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp 35 */ 36 37 /* 38 * uvm_fault.c: fault handler 39 */ 40 41 #include <sys/cdefs.h> 42 __KERNEL_RCSID(0, "$NetBSD: uvm_fault.c,v 1.90 2005/02/07 11:57:38 yamt Exp $"); 43 44 #include "opt_uvmhist.h" 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/kernel.h> 49 #include <sys/proc.h> 50 #include <sys/malloc.h> 51 #include <sys/mman.h> 52 #include <sys/user.h> 53 54 #include <uvm/uvm.h> 55 56 /* 57 * 58 * a word on page faults: 59 * 60 * types of page faults we handle: 61 * 62 * CASE 1: upper layer faults CASE 2: lower layer faults 63 * 64 * CASE 1A CASE 1B CASE 2A CASE 2B 65 * read/write1 write>1 read/write +-cow_write/zero 66 * | | | | 67 * +--|--+ +--|--+ +-----+ + | + | +-----+ 68 * amap | V | | ----------->new| | | | ^ | 69 * +-----+ +-----+ +-----+ + | + | +--|--+ 70 * | | | 71 * +-----+ +-----+ +--|--+ | +--|--+ 72 * uobj | d/c | | d/c | | V | +----| | 73 * +-----+ +-----+ +-----+ +-----+ 74 * 75 * d/c = don't care 76 * 77 * case [0]: layerless fault 78 * no amap or uobj is present. this is an error. 79 * 80 * case [1]: upper layer fault [anon active] 81 * 1A: [read] or [write with anon->an_ref == 1] 82 * I/O takes place in top level anon and uobj is not touched. 83 * 1B: [write with anon->an_ref > 1] 84 * new anon is alloc'd and data is copied off ["COW"] 85 * 86 * case [2]: lower layer fault [uobj] 87 * 2A: [read on non-NULL uobj] or [write to non-copy_on_write area] 88 * I/O takes place directly in object. 89 * 2B: [write to copy_on_write] or [read on NULL uobj] 90 * data is "promoted" from uobj to a new anon. 91 * if uobj is null, then we zero fill. 92 * 93 * we follow the standard UVM locking protocol ordering: 94 * 95 * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ) 96 * we hold a PG_BUSY page if we unlock for I/O 97 * 98 * 99 * the code is structured as follows: 100 * 101 * - init the "IN" params in the ufi structure 102 * ReFault: 103 * - do lookups [locks maps], check protection, handle needs_copy 104 * - check for case 0 fault (error) 105 * - establish "range" of fault 106 * - if we have an amap lock it and extract the anons 107 * - if sequential advice deactivate pages behind us 108 * - at the same time check pmap for unmapped areas and anon for pages 109 * that we could map in (and do map it if found) 110 * - check object for resident pages that we could map in 111 * - if (case 2) goto Case2 112 * - >>> handle case 1 113 * - ensure source anon is resident in RAM 114 * - if case 1B alloc new anon and copy from source 115 * - map the correct page in 116 * Case2: 117 * - >>> handle case 2 118 * - ensure source page is resident (if uobj) 119 * - if case 2B alloc new anon and copy from source (could be zero 120 * fill if uobj == NULL) 121 * - map the correct page in 122 * - done! 123 * 124 * note on paging: 125 * if we have to do I/O we place a PG_BUSY page in the correct object, 126 * unlock everything, and do the I/O. when I/O is done we must reverify 127 * the state of the world before assuming that our data structures are 128 * valid. [because mappings could change while the map is unlocked] 129 * 130 * alternative 1: unbusy the page in question and restart the page fault 131 * from the top (ReFault). this is easy but does not take advantage 132 * of the information that we already have from our previous lookup, 133 * although it is possible that the "hints" in the vm_map will help here. 134 * 135 * alternative 2: the system already keeps track of a "version" number of 136 * a map. [i.e. every time you write-lock a map (e.g. to change a 137 * mapping) you bump the version number up by one...] so, we can save 138 * the version number of the map before we release the lock and start I/O. 139 * then when I/O is done we can relock and check the version numbers 140 * to see if anything changed. this might save us some over 1 because 141 * we don't have to unbusy the page and may be less compares(?). 142 * 143 * alternative 3: put in backpointers or a way to "hold" part of a map 144 * in place while I/O is in progress. this could be complex to 145 * implement (especially with structures like amap that can be referenced 146 * by multiple map entries, and figuring out what should wait could be 147 * complex as well...). 148 * 149 * given that we are not currently multiprocessor or multithreaded we might 150 * as well choose alternative 2 now. maybe alternative 3 would be useful 151 * in the future. XXX keep in mind for future consideration//rechecking. 152 */ 153 154 /* 155 * local data structures 156 */ 157 158 struct uvm_advice { 159 int advice; 160 int nback; 161 int nforw; 162 }; 163 164 /* 165 * page range array: 166 * note: index in array must match "advice" value 167 * XXX: borrowed numbers from freebsd. do they work well for us? 168 */ 169 170 static struct uvm_advice uvmadvice[] = { 171 { MADV_NORMAL, 3, 4 }, 172 { MADV_RANDOM, 0, 0 }, 173 { MADV_SEQUENTIAL, 8, 7}, 174 }; 175 176 #define UVM_MAXRANGE 16 /* must be MAX() of nback+nforw+1 */ 177 178 /* 179 * private prototypes 180 */ 181 182 static void uvmfault_amapcopy(struct uvm_faultinfo *); 183 static __inline void uvmfault_anonflush(struct vm_anon **, int); 184 185 /* 186 * inline functions 187 */ 188 189 /* 190 * uvmfault_anonflush: try and deactivate pages in specified anons 191 * 192 * => does not have to deactivate page if it is busy 193 */ 194 195 static __inline void 196 uvmfault_anonflush(anons, n) 197 struct vm_anon **anons; 198 int n; 199 { 200 int lcv; 201 struct vm_page *pg; 202 203 for (lcv = 0 ; lcv < n ; lcv++) { 204 if (anons[lcv] == NULL) 205 continue; 206 simple_lock(&anons[lcv]->an_lock); 207 pg = anons[lcv]->u.an_page; 208 if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) { 209 uvm_lock_pageq(); 210 if (pg->wire_count == 0) { 211 pmap_clear_reference(pg); 212 uvm_pagedeactivate(pg); 213 } 214 uvm_unlock_pageq(); 215 } 216 simple_unlock(&anons[lcv]->an_lock); 217 } 218 } 219 220 /* 221 * normal functions 222 */ 223 224 /* 225 * uvmfault_amapcopy: clear "needs_copy" in a map. 226 * 227 * => called with VM data structures unlocked (usually, see below) 228 * => we get a write lock on the maps and clear needs_copy for a VA 229 * => if we are out of RAM we sleep (waiting for more) 230 */ 231 232 static void 233 uvmfault_amapcopy(ufi) 234 struct uvm_faultinfo *ufi; 235 { 236 for (;;) { 237 238 /* 239 * no mapping? give up. 240 */ 241 242 if (uvmfault_lookup(ufi, TRUE) == FALSE) 243 return; 244 245 /* 246 * copy if needed. 247 */ 248 249 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) 250 amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE, 251 ufi->orig_rvaddr, ufi->orig_rvaddr + 1); 252 253 /* 254 * didn't work? must be out of RAM. unlock and sleep. 255 */ 256 257 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) { 258 uvmfault_unlockmaps(ufi, TRUE); 259 uvm_wait("fltamapcopy"); 260 continue; 261 } 262 263 /* 264 * got it! unlock and return. 265 */ 266 267 uvmfault_unlockmaps(ufi, TRUE); 268 return; 269 } 270 /*NOTREACHED*/ 271 } 272 273 /* 274 * uvmfault_anonget: get data in an anon into a non-busy, non-released 275 * page in that anon. 276 * 277 * => maps, amap, and anon locked by caller. 278 * => if we fail (result != 0) we unlock everything. 279 * => if we are successful, we return with everything still locked. 280 * => we don't move the page on the queues [gets moved later] 281 * => if we allocate a new page [we_own], it gets put on the queues. 282 * either way, the result is that the page is on the queues at return time 283 * => for pages which are on loan from a uvm_object (and thus are not 284 * owned by the anon): if successful, we return with the owning object 285 * locked. the caller must unlock this object when it unlocks everything 286 * else. 287 */ 288 289 int 290 uvmfault_anonget(ufi, amap, anon) 291 struct uvm_faultinfo *ufi; 292 struct vm_amap *amap; 293 struct vm_anon *anon; 294 { 295 boolean_t we_own; /* we own anon's page? */ 296 boolean_t locked; /* did we relock? */ 297 struct vm_page *pg; 298 int error; 299 UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist); 300 301 LOCK_ASSERT(simple_lock_held(&anon->an_lock)); 302 303 error = 0; 304 uvmexp.fltanget++; 305 /* bump rusage counters */ 306 if (anon->u.an_page) 307 curproc->p_stats->p_ru.ru_minflt++; 308 else 309 curproc->p_stats->p_ru.ru_majflt++; 310 311 /* 312 * loop until we get it, or fail. 313 */ 314 315 for (;;) { 316 we_own = FALSE; /* TRUE if we set PG_BUSY on a page */ 317 pg = anon->u.an_page; 318 319 /* 320 * if there is a resident page and it is loaned, then anon 321 * may not own it. call out to uvm_anon_lockpage() to ensure 322 * the real owner of the page has been identified and locked. 323 */ 324 325 if (pg && pg->loan_count) 326 pg = uvm_anon_lockloanpg(anon); 327 328 /* 329 * page there? make sure it is not busy/released. 330 */ 331 332 if (pg) { 333 334 /* 335 * at this point, if the page has a uobject [meaning 336 * we have it on loan], then that uobject is locked 337 * by us! if the page is busy, we drop all the 338 * locks (including uobject) and try again. 339 */ 340 341 if ((pg->flags & PG_BUSY) == 0) { 342 UVMHIST_LOG(maphist, "<- OK",0,0,0,0); 343 return (0); 344 } 345 pg->flags |= PG_WANTED; 346 uvmexp.fltpgwait++; 347 348 /* 349 * the last unlock must be an atomic unlock+wait on 350 * the owner of page 351 */ 352 353 if (pg->uobject) { /* owner is uobject ? */ 354 uvmfault_unlockall(ufi, amap, NULL, anon); 355 UVMHIST_LOG(maphist, " unlock+wait on uobj",0, 356 0,0,0); 357 UVM_UNLOCK_AND_WAIT(pg, 358 &pg->uobject->vmobjlock, 359 FALSE, "anonget1",0); 360 } else { 361 /* anon owns page */ 362 uvmfault_unlockall(ufi, amap, NULL, NULL); 363 UVMHIST_LOG(maphist, " unlock+wait on anon",0, 364 0,0,0); 365 UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0, 366 "anonget2",0); 367 } 368 } else { 369 370 /* 371 * no page, we must try and bring it in. 372 */ 373 374 pg = uvm_pagealloc(NULL, 0, anon, 0); 375 if (pg == NULL) { /* out of RAM. */ 376 uvmfault_unlockall(ufi, amap, NULL, anon); 377 uvmexp.fltnoram++; 378 UVMHIST_LOG(maphist, " noram -- UVM_WAIT",0, 379 0,0,0); 380 uvm_wait("flt_noram1"); 381 } else { 382 /* we set the PG_BUSY bit */ 383 we_own = TRUE; 384 uvmfault_unlockall(ufi, amap, NULL, anon); 385 386 /* 387 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN 388 * page into the uvm_swap_get function with 389 * all data structures unlocked. note that 390 * it is ok to read an_swslot here because 391 * we hold PG_BUSY on the page. 392 */ 393 uvmexp.pageins++; 394 error = uvm_swap_get(pg, anon->an_swslot, 395 PGO_SYNCIO); 396 397 /* 398 * we clean up after the i/o below in the 399 * "we_own" case 400 */ 401 } 402 } 403 404 /* 405 * now relock and try again 406 */ 407 408 locked = uvmfault_relock(ufi); 409 if (locked && amap != NULL) { 410 amap_lock(amap); 411 } 412 if (locked || we_own) 413 simple_lock(&anon->an_lock); 414 415 /* 416 * if we own the page (i.e. we set PG_BUSY), then we need 417 * to clean up after the I/O. there are three cases to 418 * consider: 419 * [1] page released during I/O: free anon and ReFault. 420 * [2] I/O not OK. free the page and cause the fault 421 * to fail. 422 * [3] I/O OK! activate the page and sync with the 423 * non-we_own case (i.e. drop anon lock if not locked). 424 */ 425 426 if (we_own) { 427 if (pg->flags & PG_WANTED) { 428 wakeup(pg); 429 } 430 if (error) { 431 432 /* 433 * remove the swap slot from the anon 434 * and mark the anon as having no real slot. 435 * don't free the swap slot, thus preventing 436 * it from being used again. 437 */ 438 439 if (anon->an_swslot > 0) 440 uvm_swap_markbad(anon->an_swslot, 1); 441 anon->an_swslot = SWSLOT_BAD; 442 443 if ((pg->flags & PG_RELEASED) != 0) 444 goto released; 445 446 /* 447 * note: page was never !PG_BUSY, so it 448 * can't be mapped and thus no need to 449 * pmap_page_protect it... 450 */ 451 452 uvm_lock_pageq(); 453 uvm_pagefree(pg); 454 uvm_unlock_pageq(); 455 456 if (locked) 457 uvmfault_unlockall(ufi, amap, NULL, 458 anon); 459 else 460 simple_unlock(&anon->an_lock); 461 UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0); 462 return error; 463 } 464 465 if ((pg->flags & PG_RELEASED) != 0) { 466 released: 467 KASSERT(anon->an_ref == 0); 468 469 /* 470 * released while we unlocked amap. 471 */ 472 473 if (locked) 474 uvmfault_unlockall(ufi, amap, NULL, 475 NULL); 476 477 uvm_anon_release(anon); 478 479 if (error) { 480 UVMHIST_LOG(maphist, 481 "<- ERROR/RELEASED", 0,0,0,0); 482 return error; 483 } 484 485 UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0); 486 return ERESTART; 487 } 488 489 /* 490 * we've successfully read the page, activate it. 491 */ 492 493 uvm_lock_pageq(); 494 uvm_pageactivate(pg); 495 uvm_unlock_pageq(); 496 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE); 497 UVM_PAGE_OWN(pg, NULL); 498 if (!locked) 499 simple_unlock(&anon->an_lock); 500 } 501 502 /* 503 * we were not able to relock. restart fault. 504 */ 505 506 if (!locked) { 507 UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0); 508 return (ERESTART); 509 } 510 511 /* 512 * verify no one has touched the amap and moved the anon on us. 513 */ 514 515 if (ufi != NULL && 516 amap_lookup(&ufi->entry->aref, 517 ufi->orig_rvaddr - ufi->entry->start) != anon) { 518 519 uvmfault_unlockall(ufi, amap, NULL, anon); 520 UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0); 521 return (ERESTART); 522 } 523 524 /* 525 * try it again! 526 */ 527 528 uvmexp.fltanretry++; 529 continue; 530 } 531 /*NOTREACHED*/ 532 } 533 534 /* 535 * F A U L T - m a i n e n t r y p o i n t 536 */ 537 538 /* 539 * uvm_fault: page fault handler 540 * 541 * => called from MD code to resolve a page fault 542 * => VM data structures usually should be unlocked. however, it is 543 * possible to call here with the main map locked if the caller 544 * gets a write lock, sets it recusive, and then calls us (c.f. 545 * uvm_map_pageable). this should be avoided because it keeps 546 * the map locked off during I/O. 547 * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT 548 */ 549 550 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 551 ~VM_PROT_WRITE : VM_PROT_ALL) 552 553 int 554 uvm_fault(orig_map, vaddr, fault_type, access_type) 555 struct vm_map *orig_map; 556 vaddr_t vaddr; 557 vm_fault_t fault_type; 558 vm_prot_t access_type; 559 { 560 struct uvm_faultinfo ufi; 561 vm_prot_t enter_prot, check_prot; 562 boolean_t wired, narrow, promote, locked, shadowed, wire_fault, cow_now; 563 int npages, nback, nforw, centeridx, error, lcv, gotpages; 564 vaddr_t startva, currva; 565 voff_t uoff; 566 struct vm_amap *amap; 567 struct uvm_object *uobj; 568 struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon; 569 struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage; 570 UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist); 571 572 UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)", 573 orig_map, vaddr, fault_type, access_type); 574 575 anon = NULL; 576 pg = NULL; 577 578 uvmexp.faults++; /* XXX: locking? */ 579 580 /* 581 * init the IN parameters in the ufi 582 */ 583 584 ufi.orig_map = orig_map; 585 ufi.orig_rvaddr = trunc_page(vaddr); 586 ufi.orig_size = PAGE_SIZE; /* can't get any smaller than this */ 587 wire_fault = fault_type == VM_FAULT_WIRE || 588 fault_type == VM_FAULT_WIREMAX; 589 if (wire_fault) 590 narrow = TRUE; /* don't look for neighborhood 591 * pages on wire */ 592 else 593 narrow = FALSE; /* normal fault */ 594 595 /* 596 * "goto ReFault" means restart the page fault from ground zero. 597 */ 598 ReFault: 599 600 /* 601 * lookup and lock the maps 602 */ 603 604 if (uvmfault_lookup(&ufi, FALSE) == FALSE) { 605 UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0); 606 return (EFAULT); 607 } 608 /* locked: maps(read) */ 609 610 #ifdef DIAGNOSTIC 611 if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) { 612 printf("Page fault on non-pageable map:\n"); 613 printf("ufi.map = %p\n", ufi.map); 614 printf("ufi.orig_map = %p\n", ufi.orig_map); 615 printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr); 616 panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0"); 617 } 618 #endif 619 620 /* 621 * check protection 622 */ 623 624 check_prot = fault_type == VM_FAULT_WIREMAX ? 625 ufi.entry->max_protection : ufi.entry->protection; 626 if ((check_prot & access_type) != access_type) { 627 UVMHIST_LOG(maphist, 628 "<- protection failure (prot=0x%x, access=0x%x)", 629 ufi.entry->protection, access_type, 0, 0); 630 uvmfault_unlockmaps(&ufi, FALSE); 631 return EACCES; 632 } 633 634 /* 635 * "enter_prot" is the protection we want to enter the page in at. 636 * for certain pages (e.g. copy-on-write pages) this protection can 637 * be more strict than ufi.entry->protection. "wired" means either 638 * the entry is wired or we are fault-wiring the pg. 639 */ 640 641 enter_prot = ufi.entry->protection; 642 wired = VM_MAPENT_ISWIRED(ufi.entry) || wire_fault; 643 if (wired) { 644 access_type = enter_prot; /* full access for wired */ 645 cow_now = (check_prot & VM_PROT_WRITE) != 0; 646 } else { 647 cow_now = (access_type & VM_PROT_WRITE) != 0; 648 } 649 650 /* 651 * handle "needs_copy" case. if we need to copy the amap we will 652 * have to drop our readlock and relock it with a write lock. (we 653 * need a write lock to change anything in a map entry [e.g. 654 * needs_copy]). 655 */ 656 657 if (UVM_ET_ISNEEDSCOPY(ufi.entry)) { 658 KASSERT(fault_type != VM_FAULT_WIREMAX); 659 if (cow_now || (ufi.entry->object.uvm_obj == NULL)) { 660 /* need to clear */ 661 UVMHIST_LOG(maphist, 662 " need to clear needs_copy and refault",0,0,0,0); 663 uvmfault_unlockmaps(&ufi, FALSE); 664 uvmfault_amapcopy(&ufi); 665 uvmexp.fltamcopy++; 666 goto ReFault; 667 668 } else { 669 670 /* 671 * ensure that we pmap_enter page R/O since 672 * needs_copy is still true 673 */ 674 675 enter_prot &= ~VM_PROT_WRITE; 676 } 677 } 678 679 /* 680 * identify the players 681 */ 682 683 amap = ufi.entry->aref.ar_amap; /* top layer */ 684 uobj = ufi.entry->object.uvm_obj; /* bottom layer */ 685 686 /* 687 * check for a case 0 fault. if nothing backing the entry then 688 * error now. 689 */ 690 691 if (amap == NULL && uobj == NULL) { 692 uvmfault_unlockmaps(&ufi, FALSE); 693 UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0); 694 return (EFAULT); 695 } 696 697 /* 698 * establish range of interest based on advice from mapper 699 * and then clip to fit map entry. note that we only want 700 * to do this the first time through the fault. if we 701 * ReFault we will disable this by setting "narrow" to true. 702 */ 703 704 if (narrow == FALSE) { 705 706 /* wide fault (!narrow) */ 707 KASSERT(uvmadvice[ufi.entry->advice].advice == 708 ufi.entry->advice); 709 nback = MIN(uvmadvice[ufi.entry->advice].nback, 710 (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT); 711 startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT); 712 nforw = MIN(uvmadvice[ufi.entry->advice].nforw, 713 ((ufi.entry->end - ufi.orig_rvaddr) >> 714 PAGE_SHIFT) - 1); 715 /* 716 * note: "-1" because we don't want to count the 717 * faulting page as forw 718 */ 719 npages = nback + nforw + 1; 720 centeridx = nback; 721 722 narrow = TRUE; /* ensure only once per-fault */ 723 724 } else { 725 726 /* narrow fault! */ 727 nback = nforw = 0; 728 startva = ufi.orig_rvaddr; 729 npages = 1; 730 centeridx = 0; 731 732 } 733 734 /* locked: maps(read) */ 735 UVMHIST_LOG(maphist, " narrow=%d, back=%d, forw=%d, startva=0x%x", 736 narrow, nback, nforw, startva); 737 UVMHIST_LOG(maphist, " entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry, 738 amap, uobj, 0); 739 740 /* 741 * if we've got an amap, lock it and extract current anons. 742 */ 743 744 if (amap) { 745 amap_lock(amap); 746 anons = anons_store; 747 amap_lookups(&ufi.entry->aref, startva - ufi.entry->start, 748 anons, npages); 749 } else { 750 anons = NULL; /* to be safe */ 751 } 752 753 /* locked: maps(read), amap(if there) */ 754 755 /* 756 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages 757 * now and then forget about them (for the rest of the fault). 758 */ 759 760 if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) { 761 762 UVMHIST_LOG(maphist, " MADV_SEQUENTIAL: flushing backpages", 763 0,0,0,0); 764 /* flush back-page anons? */ 765 if (amap) 766 uvmfault_anonflush(anons, nback); 767 768 /* flush object? */ 769 if (uobj) { 770 uoff = (startva - ufi.entry->start) + ufi.entry->offset; 771 simple_lock(&uobj->vmobjlock); 772 (void) (uobj->pgops->pgo_put)(uobj, uoff, uoff + 773 (nback << PAGE_SHIFT), PGO_DEACTIVATE); 774 } 775 776 /* now forget about the backpages */ 777 if (amap) 778 anons += nback; 779 startva += (nback << PAGE_SHIFT); 780 npages -= nback; 781 nback = centeridx = 0; 782 } 783 784 /* locked: maps(read), amap(if there) */ 785 786 /* 787 * map in the backpages and frontpages we found in the amap in hopes 788 * of preventing future faults. we also init the pages[] array as 789 * we go. 790 */ 791 792 currva = startva; 793 shadowed = FALSE; 794 for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) { 795 796 /* 797 * dont play with VAs that are already mapped 798 * except for center) 799 */ 800 if (lcv != centeridx && 801 pmap_extract(ufi.orig_map->pmap, currva, NULL)) { 802 pages[lcv] = PGO_DONTCARE; 803 continue; 804 } 805 806 /* 807 * unmapped or center page. check if any anon at this level. 808 */ 809 if (amap == NULL || anons[lcv] == NULL) { 810 pages[lcv] = NULL; 811 continue; 812 } 813 814 /* 815 * check for present page and map if possible. re-activate it. 816 */ 817 818 pages[lcv] = PGO_DONTCARE; 819 if (lcv == centeridx) { /* save center for later! */ 820 shadowed = TRUE; 821 continue; 822 } 823 anon = anons[lcv]; 824 simple_lock(&anon->an_lock); 825 /* ignore loaned pages */ 826 if (anon->u.an_page && anon->u.an_page->loan_count == 0 && 827 (anon->u.an_page->flags & PG_BUSY) == 0) { 828 uvm_lock_pageq(); 829 uvm_pageactivate(anon->u.an_page); 830 uvm_unlock_pageq(); 831 UVMHIST_LOG(maphist, 832 " MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x", 833 ufi.orig_map->pmap, currva, anon->u.an_page, 0); 834 uvmexp.fltnamap++; 835 836 /* 837 * Since this isn't the page that's actually faulting, 838 * ignore pmap_enter() failures; it's not critical 839 * that we enter these right now. 840 */ 841 842 (void) pmap_enter(ufi.orig_map->pmap, currva, 843 VM_PAGE_TO_PHYS(anon->u.an_page), 844 (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) : 845 enter_prot, 846 PMAP_CANFAIL | 847 (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0)); 848 } 849 simple_unlock(&anon->an_lock); 850 pmap_update(ufi.orig_map->pmap); 851 } 852 853 /* locked: maps(read), amap(if there) */ 854 /* (shadowed == TRUE) if there is an anon at the faulting address */ 855 UVMHIST_LOG(maphist, " shadowed=%d, will_get=%d", shadowed, 856 (uobj && shadowed == FALSE),0,0); 857 858 /* 859 * note that if we are really short of RAM we could sleep in the above 860 * call to pmap_enter with everything locked. bad? 861 * 862 * XXX Actually, that is bad; pmap_enter() should just fail in that 863 * XXX case. --thorpej 864 */ 865 866 /* 867 * if the desired page is not shadowed by the amap and we have a 868 * backing object, then we check to see if the backing object would 869 * prefer to handle the fault itself (rather than letting us do it 870 * with the usual pgo_get hook). the backing object signals this by 871 * providing a pgo_fault routine. 872 */ 873 874 if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) { 875 simple_lock(&uobj->vmobjlock); 876 877 /* locked: maps(read), amap (if there), uobj */ 878 error = uobj->pgops->pgo_fault(&ufi, startva, pages, npages, 879 centeridx, fault_type, access_type, PGO_LOCKED|PGO_SYNCIO); 880 881 /* locked: nothing, pgo_fault has unlocked everything */ 882 883 if (error == ERESTART) 884 goto ReFault; /* try again! */ 885 /* 886 * object fault routine responsible for pmap_update(). 887 */ 888 return error; 889 } 890 891 /* 892 * now, if the desired page is not shadowed by the amap and we have 893 * a backing object that does not have a special fault routine, then 894 * we ask (with pgo_get) the object for resident pages that we care 895 * about and attempt to map them in. we do not let pgo_get block 896 * (PGO_LOCKED). 897 */ 898 899 if (uobj && shadowed == FALSE) { 900 simple_lock(&uobj->vmobjlock); 901 902 /* locked (!shadowed): maps(read), amap (if there), uobj */ 903 /* 904 * the following call to pgo_get does _not_ change locking state 905 */ 906 907 uvmexp.fltlget++; 908 gotpages = npages; 909 (void) uobj->pgops->pgo_get(uobj, ufi.entry->offset + 910 (startva - ufi.entry->start), 911 pages, &gotpages, centeridx, 912 access_type & MASK(ufi.entry), 913 ufi.entry->advice, PGO_LOCKED); 914 915 /* 916 * check for pages to map, if we got any 917 */ 918 919 uobjpage = NULL; 920 921 if (gotpages) { 922 currva = startva; 923 for (lcv = 0; lcv < npages; 924 lcv++, currva += PAGE_SIZE) { 925 struct vm_page *curpg; 926 boolean_t readonly; 927 928 curpg = pages[lcv]; 929 if (curpg == NULL || curpg == PGO_DONTCARE) { 930 continue; 931 } 932 933 /* 934 * if center page is resident and not 935 * PG_BUSY|PG_RELEASED then pgo_get 936 * made it PG_BUSY for us and gave 937 * us a handle to it. remember this 938 * page as "uobjpage." (for later use). 939 */ 940 941 if (lcv == centeridx) { 942 uobjpage = curpg; 943 UVMHIST_LOG(maphist, " got uobjpage " 944 "(0x%x) with locked get", 945 uobjpage, 0,0,0); 946 continue; 947 } 948 949 /* 950 * calling pgo_get with PGO_LOCKED returns us 951 * pages which are neither busy nor released, 952 * so we don't need to check for this. 953 * we can just directly enter the pages. 954 */ 955 956 uvm_lock_pageq(); 957 uvm_pageactivate(curpg); 958 uvm_unlock_pageq(); 959 UVMHIST_LOG(maphist, 960 " MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x", 961 ufi.orig_map->pmap, currva, curpg, 0); 962 uvmexp.fltnomap++; 963 964 /* 965 * Since this page isn't the page that's 966 * actually faulting, ignore pmap_enter() 967 * failures; it's not critical that we 968 * enter these right now. 969 */ 970 KASSERT((curpg->flags & PG_PAGEOUT) == 0); 971 KASSERT((curpg->flags & PG_RELEASED) == 0); 972 readonly = (curpg->flags & PG_RDONLY) 973 || (curpg->loan_count > 0); 974 975 (void) pmap_enter(ufi.orig_map->pmap, currva, 976 VM_PAGE_TO_PHYS(curpg), 977 readonly ? 978 enter_prot & ~VM_PROT_WRITE : 979 enter_prot & MASK(ufi.entry), 980 PMAP_CANFAIL | 981 (wired ? PMAP_WIRED : 0)); 982 983 /* 984 * NOTE: page can't be PG_WANTED or PG_RELEASED 985 * because we've held the lock the whole time 986 * we've had the handle. 987 */ 988 989 curpg->flags &= ~(PG_BUSY); 990 UVM_PAGE_OWN(curpg, NULL); 991 } 992 pmap_update(ufi.orig_map->pmap); 993 } 994 } else { 995 uobjpage = NULL; 996 } 997 998 /* locked (shadowed): maps(read), amap */ 999 /* locked (!shadowed): maps(read), amap(if there), 1000 uobj(if !null), uobjpage(if !null) */ 1001 1002 /* 1003 * note that at this point we are done with any front or back pages. 1004 * we are now going to focus on the center page (i.e. the one we've 1005 * faulted on). if we have faulted on the top (anon) layer 1006 * [i.e. case 1], then the anon we want is anons[centeridx] (we have 1007 * not touched it yet). if we have faulted on the bottom (uobj) 1008 * layer [i.e. case 2] and the page was both present and available, 1009 * then we've got a pointer to it as "uobjpage" and we've already 1010 * made it BUSY. 1011 */ 1012 1013 /* 1014 * there are four possible cases we must address: 1A, 1B, 2A, and 2B 1015 */ 1016 1017 /* 1018 * redirect case 2: if we are not shadowed, go to case 2. 1019 */ 1020 1021 if (shadowed == FALSE) 1022 goto Case2; 1023 1024 /* locked: maps(read), amap */ 1025 1026 /* 1027 * handle case 1: fault on an anon in our amap 1028 */ 1029 1030 anon = anons[centeridx]; 1031 UVMHIST_LOG(maphist, " case 1 fault: anon=0x%x", anon, 0,0,0); 1032 simple_lock(&anon->an_lock); 1033 1034 /* locked: maps(read), amap, anon */ 1035 1036 /* 1037 * no matter if we have case 1A or case 1B we are going to need to 1038 * have the anon's memory resident. ensure that now. 1039 */ 1040 1041 /* 1042 * let uvmfault_anonget do the dirty work. 1043 * if it fails (!OK) it will unlock everything for us. 1044 * if it succeeds, locks are still valid and locked. 1045 * also, if it is OK, then the anon's page is on the queues. 1046 * if the page is on loan from a uvm_object, then anonget will 1047 * lock that object for us if it does not fail. 1048 */ 1049 1050 error = uvmfault_anonget(&ufi, amap, anon); 1051 switch (error) { 1052 case 0: 1053 break; 1054 1055 case ERESTART: 1056 goto ReFault; 1057 1058 case EAGAIN: 1059 tsleep(&lbolt, PVM, "fltagain1", 0); 1060 goto ReFault; 1061 1062 default: 1063 return error; 1064 } 1065 1066 /* 1067 * uobj is non null if the page is on loan from an object (i.e. uobj) 1068 */ 1069 1070 uobj = anon->u.an_page->uobject; /* locked by anonget if !NULL */ 1071 1072 /* locked: maps(read), amap, anon, uobj(if one) */ 1073 1074 /* 1075 * special handling for loaned pages 1076 */ 1077 1078 if (anon->u.an_page->loan_count) { 1079 1080 if (!cow_now) { 1081 1082 /* 1083 * for read faults on loaned pages we just cap the 1084 * protection at read-only. 1085 */ 1086 1087 enter_prot = enter_prot & ~VM_PROT_WRITE; 1088 1089 } else { 1090 /* 1091 * note that we can't allow writes into a loaned page! 1092 * 1093 * if we have a write fault on a loaned page in an 1094 * anon then we need to look at the anon's ref count. 1095 * if it is greater than one then we are going to do 1096 * a normal copy-on-write fault into a new anon (this 1097 * is not a problem). however, if the reference count 1098 * is one (a case where we would normally allow a 1099 * write directly to the page) then we need to kill 1100 * the loan before we continue. 1101 */ 1102 1103 /* >1 case is already ok */ 1104 if (anon->an_ref == 1) { 1105 1106 /* get new un-owned replacement page */ 1107 pg = uvm_pagealloc(NULL, 0, NULL, 0); 1108 if (pg == NULL) { 1109 uvmfault_unlockall(&ufi, amap, uobj, 1110 anon); 1111 uvm_wait("flt_noram2"); 1112 goto ReFault; 1113 } 1114 1115 /* 1116 * copy data, kill loan, and drop uobj lock 1117 * (if any) 1118 */ 1119 /* copy old -> new */ 1120 uvm_pagecopy(anon->u.an_page, pg); 1121 1122 /* force reload */ 1123 pmap_page_protect(anon->u.an_page, 1124 VM_PROT_NONE); 1125 uvm_lock_pageq(); /* KILL loan */ 1126 1127 anon->u.an_page->uanon = NULL; 1128 /* in case we owned */ 1129 anon->u.an_page->pqflags &= ~PQ_ANON; 1130 1131 if (uobj) { 1132 /* if we were receiver of loan */ 1133 anon->u.an_page->loan_count--; 1134 } else { 1135 /* 1136 * we were the lender (A->K); need 1137 * to remove the page from pageq's. 1138 */ 1139 uvm_pagedequeue(anon->u.an_page); 1140 } 1141 1142 uvm_pageactivate(pg); 1143 uvm_unlock_pageq(); 1144 if (uobj) { 1145 simple_unlock(&uobj->vmobjlock); 1146 uobj = NULL; 1147 } 1148 1149 /* install new page in anon */ 1150 anon->u.an_page = pg; 1151 pg->uanon = anon; 1152 pg->pqflags |= PQ_ANON; 1153 pg->flags &= ~(PG_BUSY|PG_FAKE); 1154 UVM_PAGE_OWN(pg, NULL); 1155 1156 /* done! */ 1157 } /* ref == 1 */ 1158 } /* write fault */ 1159 } /* loan count */ 1160 1161 /* 1162 * if we are case 1B then we will need to allocate a new blank 1163 * anon to transfer the data into. note that we have a lock 1164 * on anon, so no one can busy or release the page until we are done. 1165 * also note that the ref count can't drop to zero here because 1166 * it is > 1 and we are only dropping one ref. 1167 * 1168 * in the (hopefully very rare) case that we are out of RAM we 1169 * will unlock, wait for more RAM, and refault. 1170 * 1171 * if we are out of anon VM we kill the process (XXX: could wait?). 1172 */ 1173 1174 if (cow_now && anon->an_ref > 1) { 1175 1176 UVMHIST_LOG(maphist, " case 1B: COW fault",0,0,0,0); 1177 uvmexp.flt_acow++; 1178 oanon = anon; /* oanon = old, locked anon */ 1179 anon = uvm_analloc(); 1180 if (anon) { 1181 /* new anon is locked! */ 1182 pg = uvm_pagealloc(NULL, 0, anon, 0); 1183 } 1184 1185 /* check for out of RAM */ 1186 if (anon == NULL || pg == NULL) { 1187 if (anon) { 1188 anon->an_ref--; 1189 simple_unlock(&anon->an_lock); 1190 uvm_anfree(anon); 1191 } 1192 uvmfault_unlockall(&ufi, amap, uobj, oanon); 1193 if (anon == NULL || uvm_swapisfull()) { 1194 UVMHIST_LOG(maphist, 1195 "<- failed. out of VM",0,0,0,0); 1196 uvmexp.fltnoanon++; 1197 return ENOMEM; 1198 } 1199 1200 uvmexp.fltnoram++; 1201 uvm_wait("flt_noram3"); /* out of RAM, wait for more */ 1202 goto ReFault; 1203 } 1204 1205 /* got all resources, replace anon with nanon */ 1206 uvm_pagecopy(oanon->u.an_page, pg); 1207 uvm_lock_pageq(); 1208 uvm_pageactivate(pg); 1209 pg->flags &= ~(PG_BUSY|PG_FAKE); 1210 uvm_unlock_pageq(); 1211 UVM_PAGE_OWN(pg, NULL); 1212 amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start, 1213 anon, 1); 1214 1215 /* deref: can not drop to zero here by defn! */ 1216 oanon->an_ref--; 1217 1218 /* 1219 * note: oanon is still locked, as is the new anon. we 1220 * need to check for this later when we unlock oanon; if 1221 * oanon != anon, we'll have to unlock anon, too. 1222 */ 1223 1224 } else { 1225 1226 uvmexp.flt_anon++; 1227 oanon = anon; /* old, locked anon is same as anon */ 1228 pg = anon->u.an_page; 1229 if (anon->an_ref > 1) /* disallow writes to ref > 1 anons */ 1230 enter_prot = enter_prot & ~VM_PROT_WRITE; 1231 1232 } 1233 1234 /* locked: maps(read), amap, oanon, anon (if different from oanon) */ 1235 1236 /* 1237 * now map the page in. 1238 */ 1239 1240 UVMHIST_LOG(maphist, " MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x", 1241 ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0); 1242 if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg), 1243 enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) 1244 != 0) { 1245 1246 /* 1247 * No need to undo what we did; we can simply think of 1248 * this as the pmap throwing away the mapping information. 1249 * 1250 * We do, however, have to go through the ReFault path, 1251 * as the map may change while we're asleep. 1252 */ 1253 1254 if (anon != oanon) 1255 simple_unlock(&anon->an_lock); 1256 uvmfault_unlockall(&ufi, amap, uobj, oanon); 1257 if (uvm_swapisfull()) { 1258 UVMHIST_LOG(maphist, 1259 "<- failed. out of VM",0,0,0,0); 1260 /* XXX instrumentation */ 1261 return ENOMEM; 1262 } 1263 /* XXX instrumentation */ 1264 uvm_wait("flt_pmfail1"); 1265 goto ReFault; 1266 } 1267 1268 /* 1269 * ... update the page queues. 1270 */ 1271 1272 uvm_lock_pageq(); 1273 if (wire_fault) { 1274 uvm_pagewire(pg); 1275 1276 /* 1277 * since the now-wired page cannot be paged out, 1278 * release its swap resources for others to use. 1279 * since an anon with no swap cannot be PG_CLEAN, 1280 * clear its clean flag now. 1281 */ 1282 1283 pg->flags &= ~(PG_CLEAN); 1284 uvm_anon_dropswap(anon); 1285 } else { 1286 uvm_pageactivate(pg); 1287 } 1288 uvm_unlock_pageq(); 1289 1290 /* 1291 * done case 1! finish up by unlocking everything and returning success 1292 */ 1293 1294 if (anon != oanon) 1295 simple_unlock(&anon->an_lock); 1296 uvmfault_unlockall(&ufi, amap, uobj, oanon); 1297 pmap_update(ufi.orig_map->pmap); 1298 return 0; 1299 1300 Case2: 1301 /* 1302 * handle case 2: faulting on backing object or zero fill 1303 */ 1304 1305 /* 1306 * locked: 1307 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null) 1308 */ 1309 1310 /* 1311 * note that uobjpage can not be PGO_DONTCARE at this point. we now 1312 * set uobjpage to PGO_DONTCARE if we are doing a zero fill. if we 1313 * have a backing object, check and see if we are going to promote 1314 * the data up to an anon during the fault. 1315 */ 1316 1317 if (uobj == NULL) { 1318 uobjpage = PGO_DONTCARE; 1319 promote = TRUE; /* always need anon here */ 1320 } else { 1321 KASSERT(uobjpage != PGO_DONTCARE); 1322 promote = cow_now && UVM_ET_ISCOPYONWRITE(ufi.entry); 1323 } 1324 UVMHIST_LOG(maphist, " case 2 fault: promote=%d, zfill=%d", 1325 promote, (uobj == NULL), 0,0); 1326 1327 /* 1328 * if uobjpage is not null then we do not need to do I/O to get the 1329 * uobjpage. 1330 * 1331 * if uobjpage is null, then we need to unlock and ask the pager to 1332 * get the data for us. once we have the data, we need to reverify 1333 * the state the world. we are currently not holding any resources. 1334 */ 1335 1336 if (uobjpage) { 1337 /* update rusage counters */ 1338 curproc->p_stats->p_ru.ru_minflt++; 1339 } else { 1340 /* update rusage counters */ 1341 curproc->p_stats->p_ru.ru_majflt++; 1342 1343 /* locked: maps(read), amap(if there), uobj */ 1344 uvmfault_unlockall(&ufi, amap, NULL, NULL); 1345 /* locked: uobj */ 1346 1347 uvmexp.fltget++; 1348 gotpages = 1; 1349 uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset; 1350 error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages, 1351 0, access_type & MASK(ufi.entry), ufi.entry->advice, 1352 PGO_SYNCIO); 1353 /* locked: uobjpage(if no error) */ 1354 1355 /* 1356 * recover from I/O 1357 */ 1358 1359 if (error) { 1360 if (error == EAGAIN) { 1361 UVMHIST_LOG(maphist, 1362 " pgo_get says TRY AGAIN!",0,0,0,0); 1363 tsleep(&lbolt, PVM, "fltagain2", 0); 1364 goto ReFault; 1365 } 1366 1367 UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)", 1368 error, 0,0,0); 1369 return error; 1370 } 1371 1372 /* locked: uobjpage */ 1373 1374 uvm_lock_pageq(); 1375 uvm_pageactivate(uobjpage); 1376 uvm_unlock_pageq(); 1377 1378 /* 1379 * re-verify the state of the world by first trying to relock 1380 * the maps. always relock the object. 1381 */ 1382 1383 locked = uvmfault_relock(&ufi); 1384 if (locked && amap) 1385 amap_lock(amap); 1386 simple_lock(&uobj->vmobjlock); 1387 1388 /* locked(locked): maps(read), amap(if !null), uobj, uobjpage */ 1389 /* locked(!locked): uobj, uobjpage */ 1390 1391 /* 1392 * verify that the page has not be released and re-verify 1393 * that amap slot is still free. if there is a problem, 1394 * we unlock and clean up. 1395 */ 1396 1397 if ((uobjpage->flags & PG_RELEASED) != 0 || 1398 (locked && amap && 1399 amap_lookup(&ufi.entry->aref, 1400 ufi.orig_rvaddr - ufi.entry->start))) { 1401 if (locked) 1402 uvmfault_unlockall(&ufi, amap, NULL, NULL); 1403 locked = FALSE; 1404 } 1405 1406 /* 1407 * didn't get the lock? release the page and retry. 1408 */ 1409 1410 if (locked == FALSE) { 1411 UVMHIST_LOG(maphist, 1412 " wasn't able to relock after fault: retry", 1413 0,0,0,0); 1414 if (uobjpage->flags & PG_WANTED) 1415 wakeup(uobjpage); 1416 if (uobjpage->flags & PG_RELEASED) { 1417 uvmexp.fltpgrele++; 1418 uvm_pagefree(uobjpage); 1419 goto ReFault; 1420 } 1421 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1422 UVM_PAGE_OWN(uobjpage, NULL); 1423 simple_unlock(&uobj->vmobjlock); 1424 goto ReFault; 1425 } 1426 1427 /* 1428 * we have the data in uobjpage which is busy and 1429 * not released. we are holding object lock (so the page 1430 * can't be released on us). 1431 */ 1432 1433 /* locked: maps(read), amap(if !null), uobj, uobjpage */ 1434 } 1435 1436 /* 1437 * locked: 1438 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj) 1439 */ 1440 1441 /* 1442 * notes: 1443 * - at this point uobjpage can not be NULL 1444 * - at this point uobjpage can not be PG_RELEASED (since we checked 1445 * for it above) 1446 * - at this point uobjpage could be PG_WANTED (handle later) 1447 */ 1448 1449 if (promote == FALSE) { 1450 1451 /* 1452 * we are not promoting. if the mapping is COW ensure that we 1453 * don't give more access than we should (e.g. when doing a read 1454 * fault on a COPYONWRITE mapping we want to map the COW page in 1455 * R/O even though the entry protection could be R/W). 1456 * 1457 * set "pg" to the page we want to map in (uobjpage, usually) 1458 */ 1459 1460 /* no anon in this case. */ 1461 anon = NULL; 1462 1463 uvmexp.flt_obj++; 1464 if (UVM_ET_ISCOPYONWRITE(ufi.entry)) 1465 enter_prot &= ~VM_PROT_WRITE; 1466 pg = uobjpage; /* map in the actual object */ 1467 1468 /* assert(uobjpage != PGO_DONTCARE) */ 1469 1470 /* 1471 * we are faulting directly on the page. be careful 1472 * about writing to loaned pages... 1473 */ 1474 1475 if (uobjpage->loan_count) { 1476 if (!cow_now) { 1477 /* read fault: cap the protection at readonly */ 1478 /* cap! */ 1479 enter_prot = enter_prot & ~VM_PROT_WRITE; 1480 } else { 1481 /* write fault: must break the loan here */ 1482 1483 pg = uvm_loanbreak(uobjpage); 1484 if (pg == NULL) { 1485 1486 /* 1487 * drop ownership of page, it can't 1488 * be released 1489 */ 1490 1491 if (uobjpage->flags & PG_WANTED) 1492 wakeup(uobjpage); 1493 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1494 UVM_PAGE_OWN(uobjpage, NULL); 1495 1496 uvmfault_unlockall(&ufi, amap, uobj, 1497 NULL); 1498 UVMHIST_LOG(maphist, 1499 " out of RAM breaking loan, waiting", 1500 0,0,0,0); 1501 uvmexp.fltnoram++; 1502 uvm_wait("flt_noram4"); 1503 goto ReFault; 1504 } 1505 uobjpage = pg; 1506 } 1507 } 1508 } else { 1509 1510 /* 1511 * if we are going to promote the data to an anon we 1512 * allocate a blank anon here and plug it into our amap. 1513 */ 1514 #if DIAGNOSTIC 1515 if (amap == NULL) 1516 panic("uvm_fault: want to promote data, but no anon"); 1517 #endif 1518 1519 anon = uvm_analloc(); 1520 if (anon) { 1521 1522 /* 1523 * The new anon is locked. 1524 * 1525 * In `Fill in data...' below, if 1526 * uobjpage == PGO_DONTCARE, we want 1527 * a zero'd, dirty page, so have 1528 * uvm_pagealloc() do that for us. 1529 */ 1530 1531 pg = uvm_pagealloc(NULL, 0, anon, 1532 (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0); 1533 } 1534 1535 /* 1536 * out of memory resources? 1537 */ 1538 1539 if (anon == NULL || pg == NULL) { 1540 if (anon != NULL) { 1541 anon->an_ref--; 1542 simple_unlock(&anon->an_lock); 1543 uvm_anfree(anon); 1544 } 1545 1546 /* 1547 * arg! must unbusy our page and fail or sleep. 1548 */ 1549 1550 if (uobjpage != PGO_DONTCARE) { 1551 if (uobjpage->flags & PG_WANTED) 1552 /* still holding object lock */ 1553 wakeup(uobjpage); 1554 1555 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1556 UVM_PAGE_OWN(uobjpage, NULL); 1557 } 1558 1559 /* unlock and fail ... */ 1560 uvmfault_unlockall(&ufi, amap, uobj, NULL); 1561 if (anon == NULL || uvm_swapisfull()) { 1562 UVMHIST_LOG(maphist, " promote: out of VM", 1563 0,0,0,0); 1564 uvmexp.fltnoanon++; 1565 return ENOMEM; 1566 } 1567 1568 UVMHIST_LOG(maphist, " out of RAM, waiting for more", 1569 0,0,0,0); 1570 uvmexp.fltnoram++; 1571 uvm_wait("flt_noram5"); 1572 goto ReFault; 1573 } 1574 1575 /* 1576 * fill in the data 1577 */ 1578 1579 if (uobjpage != PGO_DONTCARE) { 1580 uvmexp.flt_prcopy++; 1581 /* copy page [pg now dirty] */ 1582 uvm_pagecopy(uobjpage, pg); 1583 1584 /* 1585 * promote to shared amap? make sure all sharing 1586 * procs see it 1587 */ 1588 1589 if ((amap_flags(amap) & AMAP_SHARED) != 0) { 1590 pmap_page_protect(uobjpage, VM_PROT_NONE); 1591 /* 1592 * XXX: PAGE MIGHT BE WIRED! 1593 */ 1594 } 1595 1596 /* 1597 * dispose of uobjpage. it can't be PG_RELEASED 1598 * since we still hold the object lock. 1599 * drop handle to uobj as well. 1600 */ 1601 1602 if (uobjpage->flags & PG_WANTED) 1603 /* still have the obj lock */ 1604 wakeup(uobjpage); 1605 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1606 UVM_PAGE_OWN(uobjpage, NULL); 1607 simple_unlock(&uobj->vmobjlock); 1608 uobj = NULL; 1609 1610 UVMHIST_LOG(maphist, 1611 " promote uobjpage 0x%x to anon/page 0x%x/0x%x", 1612 uobjpage, anon, pg, 0); 1613 1614 } else { 1615 uvmexp.flt_przero++; 1616 1617 /* 1618 * Page is zero'd and marked dirty by uvm_pagealloc() 1619 * above. 1620 */ 1621 1622 UVMHIST_LOG(maphist," zero fill anon/page 0x%x/0%x", 1623 anon, pg, 0, 0); 1624 } 1625 amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start, 1626 anon, 0); 1627 } 1628 1629 /* 1630 * locked: 1631 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj), 1632 * anon(if !null), pg(if anon) 1633 * 1634 * note: pg is either the uobjpage or the new page in the new anon 1635 */ 1636 1637 /* 1638 * all resources are present. we can now map it in and free our 1639 * resources. 1640 */ 1641 1642 UVMHIST_LOG(maphist, 1643 " MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d", 1644 ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote); 1645 KASSERT((access_type & VM_PROT_WRITE) == 0 || 1646 (pg->flags & PG_RDONLY) == 0); 1647 if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg), 1648 pg->flags & PG_RDONLY ? enter_prot & ~VM_PROT_WRITE : enter_prot, 1649 access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) { 1650 1651 /* 1652 * No need to undo what we did; we can simply think of 1653 * this as the pmap throwing away the mapping information. 1654 * 1655 * We do, however, have to go through the ReFault path, 1656 * as the map may change while we're asleep. 1657 */ 1658 1659 if (pg->flags & PG_WANTED) 1660 wakeup(pg); 1661 1662 /* 1663 * note that pg can't be PG_RELEASED since we did not drop 1664 * the object lock since the last time we checked. 1665 */ 1666 1667 pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED); 1668 UVM_PAGE_OWN(pg, NULL); 1669 uvmfault_unlockall(&ufi, amap, uobj, anon); 1670 if (uvm_swapisfull()) { 1671 UVMHIST_LOG(maphist, 1672 "<- failed. out of VM",0,0,0,0); 1673 /* XXX instrumentation */ 1674 return ENOMEM; 1675 } 1676 /* XXX instrumentation */ 1677 uvm_wait("flt_pmfail2"); 1678 goto ReFault; 1679 } 1680 1681 uvm_lock_pageq(); 1682 if (wire_fault) { 1683 uvm_pagewire(pg); 1684 if (pg->pqflags & PQ_AOBJ) { 1685 1686 /* 1687 * since the now-wired page cannot be paged out, 1688 * release its swap resources for others to use. 1689 * since an aobj page with no swap cannot be PG_CLEAN, 1690 * clear its clean flag now. 1691 */ 1692 1693 pg->flags &= ~(PG_CLEAN); 1694 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 1695 } 1696 } else { 1697 uvm_pageactivate(pg); 1698 } 1699 uvm_unlock_pageq(); 1700 if (pg->flags & PG_WANTED) 1701 wakeup(pg); 1702 1703 /* 1704 * note that pg can't be PG_RELEASED since we did not drop the object 1705 * lock since the last time we checked. 1706 */ 1707 1708 pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED); 1709 UVM_PAGE_OWN(pg, NULL); 1710 uvmfault_unlockall(&ufi, amap, uobj, anon); 1711 pmap_update(ufi.orig_map->pmap); 1712 UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0); 1713 return 0; 1714 } 1715 1716 /* 1717 * uvm_fault_wire: wire down a range of virtual addresses in a map. 1718 * 1719 * => map may be read-locked by caller, but MUST NOT be write-locked. 1720 * => if map is read-locked, any operations which may cause map to 1721 * be write-locked in uvm_fault() must be taken care of by 1722 * the caller. See uvm_map_pageable(). 1723 */ 1724 1725 int 1726 uvm_fault_wire(map, start, end, fault_type, access_type) 1727 struct vm_map *map; 1728 vaddr_t start, end; 1729 vm_fault_t fault_type; 1730 vm_prot_t access_type; 1731 { 1732 vaddr_t va; 1733 int error; 1734 1735 /* 1736 * now fault it in a page at a time. if the fault fails then we have 1737 * to undo what we have done. note that in uvm_fault VM_PROT_NONE 1738 * is replaced with the max protection if fault_type is VM_FAULT_WIRE. 1739 */ 1740 1741 /* 1742 * XXX work around overflowing a vaddr_t. this prevents us from 1743 * wiring the last page in the address space, though. 1744 */ 1745 if (start > end) { 1746 return EFAULT; 1747 } 1748 1749 for (va = start ; va < end ; va += PAGE_SIZE) { 1750 error = uvm_fault(map, va, fault_type, access_type); 1751 if (error) { 1752 if (va != start) { 1753 uvm_fault_unwire(map, start, va); 1754 } 1755 return error; 1756 } 1757 } 1758 return 0; 1759 } 1760 1761 /* 1762 * uvm_fault_unwire(): unwire range of virtual space. 1763 */ 1764 1765 void 1766 uvm_fault_unwire(map, start, end) 1767 struct vm_map *map; 1768 vaddr_t start, end; 1769 { 1770 vm_map_lock_read(map); 1771 uvm_fault_unwire_locked(map, start, end); 1772 vm_map_unlock_read(map); 1773 } 1774 1775 /* 1776 * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire(). 1777 * 1778 * => map must be at least read-locked. 1779 */ 1780 1781 void 1782 uvm_fault_unwire_locked(map, start, end) 1783 struct vm_map *map; 1784 vaddr_t start, end; 1785 { 1786 struct vm_map_entry *entry; 1787 pmap_t pmap = vm_map_pmap(map); 1788 vaddr_t va; 1789 paddr_t pa; 1790 struct vm_page *pg; 1791 1792 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 1793 1794 /* 1795 * we assume that the area we are unwiring has actually been wired 1796 * in the first place. this means that we should be able to extract 1797 * the PAs from the pmap. we also lock out the page daemon so that 1798 * we can call uvm_pageunwire. 1799 */ 1800 1801 uvm_lock_pageq(); 1802 1803 /* 1804 * find the beginning map entry for the region. 1805 */ 1806 1807 KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map)); 1808 if (uvm_map_lookup_entry(map, start, &entry) == FALSE) 1809 panic("uvm_fault_unwire_locked: address not in map"); 1810 1811 for (va = start; va < end; va += PAGE_SIZE) { 1812 if (pmap_extract(pmap, va, &pa) == FALSE) 1813 continue; 1814 1815 /* 1816 * find the map entry for the current address. 1817 */ 1818 1819 KASSERT(va >= entry->start); 1820 while (va >= entry->end) { 1821 KASSERT(entry->next != &map->header && 1822 entry->next->start <= entry->end); 1823 entry = entry->next; 1824 } 1825 1826 /* 1827 * if the entry is no longer wired, tell the pmap. 1828 */ 1829 1830 if (VM_MAPENT_ISWIRED(entry) == 0) 1831 pmap_unwire(pmap, va); 1832 1833 pg = PHYS_TO_VM_PAGE(pa); 1834 if (pg) 1835 uvm_pageunwire(pg); 1836 } 1837 1838 uvm_unlock_pageq(); 1839 } 1840