1 /* $NetBSD: uvm_fault.c,v 1.114 2006/10/12 10:14:20 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 1.114 2006/10/12 10:14:20 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 const 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 /* 183 * inline functions 184 */ 185 186 /* 187 * uvmfault_anonflush: try and deactivate pages in specified anons 188 * 189 * => does not have to deactivate page if it is busy 190 */ 191 192 static inline void 193 uvmfault_anonflush(struct vm_anon **anons, int n) 194 { 195 int lcv; 196 struct vm_page *pg; 197 198 for (lcv = 0 ; lcv < n ; lcv++) { 199 if (anons[lcv] == NULL) 200 continue; 201 simple_lock(&anons[lcv]->an_lock); 202 pg = anons[lcv]->an_page; 203 if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) { 204 uvm_lock_pageq(); 205 if (pg->wire_count == 0) { 206 pmap_clear_reference(pg); 207 uvm_pagedeactivate(pg); 208 } 209 uvm_unlock_pageq(); 210 } 211 simple_unlock(&anons[lcv]->an_lock); 212 } 213 } 214 215 /* 216 * normal functions 217 */ 218 219 /* 220 * uvmfault_amapcopy: clear "needs_copy" in a map. 221 * 222 * => called with VM data structures unlocked (usually, see below) 223 * => we get a write lock on the maps and clear needs_copy for a VA 224 * => if we are out of RAM we sleep (waiting for more) 225 */ 226 227 static void 228 uvmfault_amapcopy(struct uvm_faultinfo *ufi) 229 { 230 for (;;) { 231 232 /* 233 * no mapping? give up. 234 */ 235 236 if (uvmfault_lookup(ufi, TRUE) == FALSE) 237 return; 238 239 /* 240 * copy if needed. 241 */ 242 243 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) 244 amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT, 245 ufi->orig_rvaddr, ufi->orig_rvaddr + 1); 246 247 /* 248 * didn't work? must be out of RAM. unlock and sleep. 249 */ 250 251 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) { 252 uvmfault_unlockmaps(ufi, TRUE); 253 uvm_wait("fltamapcopy"); 254 continue; 255 } 256 257 /* 258 * got it! unlock and return. 259 */ 260 261 uvmfault_unlockmaps(ufi, TRUE); 262 return; 263 } 264 /*NOTREACHED*/ 265 } 266 267 /* 268 * uvmfault_anonget: get data in an anon into a non-busy, non-released 269 * page in that anon. 270 * 271 * => maps, amap, and anon locked by caller. 272 * => if we fail (result != 0) we unlock everything. 273 * => if we are successful, we return with everything still locked. 274 * => we don't move the page on the queues [gets moved later] 275 * => if we allocate a new page [we_own], it gets put on the queues. 276 * either way, the result is that the page is on the queues at return time 277 * => for pages which are on loan from a uvm_object (and thus are not 278 * owned by the anon): if successful, we return with the owning object 279 * locked. the caller must unlock this object when it unlocks everything 280 * else. 281 */ 282 283 int 284 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap, 285 struct vm_anon *anon) 286 { 287 boolean_t we_own; /* we own anon's page? */ 288 boolean_t locked; /* did we relock? */ 289 struct vm_page *pg; 290 int error; 291 UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist); 292 293 LOCK_ASSERT(simple_lock_held(&anon->an_lock)); 294 295 error = 0; 296 uvmexp.fltanget++; 297 /* bump rusage counters */ 298 if (anon->an_page) 299 curproc->p_stats->p_ru.ru_minflt++; 300 else 301 curproc->p_stats->p_ru.ru_majflt++; 302 303 /* 304 * loop until we get it, or fail. 305 */ 306 307 for (;;) { 308 we_own = FALSE; /* TRUE if we set PG_BUSY on a page */ 309 pg = anon->an_page; 310 311 /* 312 * if there is a resident page and it is loaned, then anon 313 * may not own it. call out to uvm_anon_lockpage() to ensure 314 * the real owner of the page has been identified and locked. 315 */ 316 317 if (pg && pg->loan_count) 318 pg = uvm_anon_lockloanpg(anon); 319 320 /* 321 * page there? make sure it is not busy/released. 322 */ 323 324 if (pg) { 325 326 /* 327 * at this point, if the page has a uobject [meaning 328 * we have it on loan], then that uobject is locked 329 * by us! if the page is busy, we drop all the 330 * locks (including uobject) and try again. 331 */ 332 333 if ((pg->flags & PG_BUSY) == 0) { 334 UVMHIST_LOG(maphist, "<- OK",0,0,0,0); 335 return (0); 336 } 337 pg->flags |= PG_WANTED; 338 uvmexp.fltpgwait++; 339 340 /* 341 * the last unlock must be an atomic unlock+wait on 342 * the owner of page 343 */ 344 345 if (pg->uobject) { /* owner is uobject ? */ 346 uvmfault_unlockall(ufi, amap, NULL, anon); 347 UVMHIST_LOG(maphist, " unlock+wait on uobj",0, 348 0,0,0); 349 UVM_UNLOCK_AND_WAIT(pg, 350 &pg->uobject->vmobjlock, 351 FALSE, "anonget1",0); 352 } else { 353 /* anon owns page */ 354 uvmfault_unlockall(ufi, amap, NULL, NULL); 355 UVMHIST_LOG(maphist, " unlock+wait on anon",0, 356 0,0,0); 357 UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0, 358 "anonget2",0); 359 } 360 } else { 361 #if defined(VMSWAP) 362 363 /* 364 * no page, we must try and bring it in. 365 */ 366 367 pg = uvm_pagealloc(NULL, 0, anon, 0); 368 if (pg == NULL) { /* out of RAM. */ 369 uvmfault_unlockall(ufi, amap, NULL, anon); 370 uvmexp.fltnoram++; 371 UVMHIST_LOG(maphist, " noram -- UVM_WAIT",0, 372 0,0,0); 373 if (!uvm_reclaimable()) { 374 return ENOMEM; 375 } 376 uvm_wait("flt_noram1"); 377 } else { 378 /* we set the PG_BUSY bit */ 379 we_own = TRUE; 380 uvmfault_unlockall(ufi, amap, NULL, anon); 381 382 /* 383 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN 384 * page into the uvm_swap_get function with 385 * all data structures unlocked. note that 386 * it is ok to read an_swslot here because 387 * we hold PG_BUSY on the page. 388 */ 389 uvmexp.pageins++; 390 error = uvm_swap_get(pg, anon->an_swslot, 391 PGO_SYNCIO); 392 393 /* 394 * we clean up after the i/o below in the 395 * "we_own" case 396 */ 397 } 398 #else /* defined(VMSWAP) */ 399 panic("%s: no page", __func__); 400 #endif /* defined(VMSWAP) */ 401 } 402 403 /* 404 * now relock and try again 405 */ 406 407 locked = uvmfault_relock(ufi); 408 if (locked && amap != NULL) { 409 amap_lock(amap); 410 } 411 if (locked || we_own) 412 simple_lock(&anon->an_lock); 413 414 /* 415 * if we own the page (i.e. we set PG_BUSY), then we need 416 * to clean up after the I/O. there are three cases to 417 * consider: 418 * [1] page released during I/O: free anon and ReFault. 419 * [2] I/O not OK. free the page and cause the fault 420 * to fail. 421 * [3] I/O OK! activate the page and sync with the 422 * non-we_own case (i.e. drop anon lock if not locked). 423 */ 424 425 if (we_own) { 426 #if defined(VMSWAP) 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 #else /* defined(VMSWAP) */ 501 panic("%s: we_own", __func__); 502 #endif /* defined(VMSWAP) */ 503 } 504 505 /* 506 * we were not able to relock. restart fault. 507 */ 508 509 if (!locked) { 510 UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0); 511 return (ERESTART); 512 } 513 514 /* 515 * verify no one has touched the amap and moved the anon on us. 516 */ 517 518 if (ufi != NULL && 519 amap_lookup(&ufi->entry->aref, 520 ufi->orig_rvaddr - ufi->entry->start) != anon) { 521 522 uvmfault_unlockall(ufi, amap, NULL, anon); 523 UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0); 524 return (ERESTART); 525 } 526 527 /* 528 * try it again! 529 */ 530 531 uvmexp.fltanretry++; 532 continue; 533 } 534 /*NOTREACHED*/ 535 } 536 537 /* 538 * uvmfault_promote: promote data to a new anon. used for 1B and 2B. 539 * 540 * 1. allocate an anon and a page. 541 * 2. fill its contents. 542 * 3. put it into amap. 543 * 544 * => if we fail (result != 0) we unlock everything. 545 * => on success, return a new locked anon via 'nanon'. 546 * (*nanon)->an_page will be a resident, locked, dirty page. 547 */ 548 549 static int 550 uvmfault_promote(struct uvm_faultinfo *ufi, 551 struct vm_anon *oanon, 552 struct vm_page *uobjpage, 553 struct vm_anon **nanon, /* OUT: allocated anon */ 554 struct vm_anon **spare) 555 { 556 struct vm_amap *amap = ufi->entry->aref.ar_amap; 557 struct uvm_object *uobj; 558 struct vm_anon *anon; 559 struct vm_page *pg; 560 struct vm_page *opg; 561 int error; 562 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 563 564 if (oanon) { 565 /* anon COW */ 566 opg = oanon->an_page; 567 KASSERT(opg != NULL); 568 KASSERT(opg->uobject == NULL || opg->loan_count > 0); 569 } else if (uobjpage != PGO_DONTCARE) { 570 /* object-backed COW */ 571 opg = uobjpage; 572 } else { 573 /* ZFOD */ 574 opg = NULL; 575 } 576 if (opg != NULL) { 577 uobj = opg->uobject; 578 } else { 579 uobj = NULL; 580 } 581 582 KASSERT(amap != NULL); 583 KASSERT(uobjpage != NULL); 584 KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0); 585 LOCK_ASSERT(simple_lock_held(&amap->am_l)); 586 LOCK_ASSERT(oanon == NULL || simple_lock_held(&oanon->an_lock)); 587 LOCK_ASSERT(uobj == NULL || simple_lock_held(&uobj->vmobjlock)); 588 LOCK_ASSERT(*spare == NULL || !simple_lock_held(&(*spare)->an_lock)); 589 590 if (*spare != NULL) { 591 anon = *spare; 592 *spare = NULL; 593 simple_lock(&anon->an_lock); 594 } else if (ufi->map != kernel_map) { 595 anon = uvm_analloc(); 596 } else { 597 UVMHIST_LOG(maphist, "kernel_map, unlock and retry", 0,0,0,0); 598 599 /* 600 * we can't allocate anons with kernel_map locked. 601 */ 602 603 uvm_page_unbusy(&uobjpage, 1); 604 uvmfault_unlockall(ufi, amap, uobj, oanon); 605 606 *spare = uvm_analloc(); 607 if (*spare == NULL) { 608 goto nomem; 609 } 610 simple_unlock(&(*spare)->an_lock); 611 error = ERESTART; 612 goto done; 613 } 614 if (anon) { 615 616 /* 617 * The new anon is locked. 618 * 619 * if opg == NULL, we want a zero'd, dirty page, 620 * so have uvm_pagealloc() do that for us. 621 */ 622 623 pg = uvm_pagealloc(NULL, 0, anon, 624 (opg == NULL) ? UVM_PGA_ZERO : 0); 625 } else { 626 pg = NULL; 627 } 628 629 /* 630 * out of memory resources? 631 */ 632 633 if (pg == NULL) { 634 /* save anon for the next try. */ 635 if (anon != NULL) { 636 simple_unlock(&anon->an_lock); 637 *spare = anon; 638 } 639 640 /* unlock and fail ... */ 641 uvm_page_unbusy(&uobjpage, 1); 642 uvmfault_unlockall(ufi, amap, uobj, oanon); 643 nomem: 644 if (!uvm_reclaimable()) { 645 UVMHIST_LOG(maphist, "out of VM", 0,0,0,0); 646 uvmexp.fltnoanon++; 647 error = ENOMEM; 648 goto done; 649 } 650 651 UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0); 652 uvmexp.fltnoram++; 653 uvm_wait("flt_noram5"); 654 error = ERESTART; 655 goto done; 656 } 657 658 /* copy page [pg now dirty] */ 659 if (opg) { 660 uvm_pagecopy(opg, pg); 661 } 662 663 amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon, 664 oanon != NULL); 665 666 *nanon = anon; 667 error = 0; 668 done: 669 return error; 670 } 671 672 673 /* 674 * F A U L T - m a i n e n t r y p o i n t 675 */ 676 677 /* 678 * uvm_fault: page fault handler 679 * 680 * => called from MD code to resolve a page fault 681 * => VM data structures usually should be unlocked. however, it is 682 * possible to call here with the main map locked if the caller 683 * gets a write lock, sets it recusive, and then calls us (c.f. 684 * uvm_map_pageable). this should be avoided because it keeps 685 * the map locked off during I/O. 686 * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT 687 */ 688 689 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 690 ~VM_PROT_WRITE : VM_PROT_ALL) 691 692 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */ 693 #define UVM_FAULT_WIRE 1 694 #define UVM_FAULT_WIREMAX 2 695 696 int 697 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr, 698 vm_prot_t access_type, int fault_flag) 699 { 700 struct uvm_faultinfo ufi; 701 vm_prot_t enter_prot, check_prot; 702 boolean_t wired, narrow, promote, locked, shadowed, wire_fault, cow_now; 703 int npages, nback, nforw, centeridx, error, lcv, gotpages; 704 vaddr_t startva, currva; 705 voff_t uoff; 706 struct vm_amap *amap; 707 struct uvm_object *uobj; 708 struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon; 709 struct vm_anon *anon_spare; 710 struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage; 711 UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist); 712 713 UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, at=%d, ff=%d)", 714 orig_map, vaddr, access_type, fault_flag); 715 716 anon = anon_spare = NULL; 717 pg = NULL; 718 719 uvmexp.faults++; /* XXX: locking? */ 720 721 /* 722 * init the IN parameters in the ufi 723 */ 724 725 ufi.orig_map = orig_map; 726 ufi.orig_rvaddr = trunc_page(vaddr); 727 ufi.orig_size = PAGE_SIZE; /* can't get any smaller than this */ 728 wire_fault = (fault_flag > 0); 729 if (wire_fault) 730 narrow = TRUE; /* don't look for neighborhood 731 * pages on wire */ 732 else 733 narrow = FALSE; /* normal fault */ 734 735 /* 736 * "goto ReFault" means restart the page fault from ground zero. 737 */ 738 ReFault: 739 740 /* 741 * lookup and lock the maps 742 */ 743 744 if (uvmfault_lookup(&ufi, FALSE) == FALSE) { 745 UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0); 746 error = EFAULT; 747 goto done; 748 } 749 /* locked: maps(read) */ 750 751 #ifdef DIAGNOSTIC 752 if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) { 753 printf("Page fault on non-pageable map:\n"); 754 printf("ufi.map = %p\n", ufi.map); 755 printf("ufi.orig_map = %p\n", ufi.orig_map); 756 printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr); 757 panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0"); 758 } 759 #endif 760 761 /* 762 * check protection 763 */ 764 765 check_prot = fault_flag == UVM_FAULT_WIREMAX ? 766 ufi.entry->max_protection : ufi.entry->protection; 767 if ((check_prot & access_type) != access_type) { 768 UVMHIST_LOG(maphist, 769 "<- protection failure (prot=0x%x, access=0x%x)", 770 ufi.entry->protection, access_type, 0, 0); 771 uvmfault_unlockmaps(&ufi, FALSE); 772 error = EACCES; 773 goto done; 774 } 775 776 /* 777 * "enter_prot" is the protection we want to enter the page in at. 778 * for certain pages (e.g. copy-on-write pages) this protection can 779 * be more strict than ufi.entry->protection. "wired" means either 780 * the entry is wired or we are fault-wiring the pg. 781 */ 782 783 enter_prot = ufi.entry->protection; 784 wired = VM_MAPENT_ISWIRED(ufi.entry) || wire_fault; 785 if (wired) { 786 access_type = enter_prot; /* full access for wired */ 787 cow_now = (check_prot & VM_PROT_WRITE) != 0; 788 } else { 789 cow_now = (access_type & VM_PROT_WRITE) != 0; 790 } 791 792 /* 793 * handle "needs_copy" case. if we need to copy the amap we will 794 * have to drop our readlock and relock it with a write lock. (we 795 * need a write lock to change anything in a map entry [e.g. 796 * needs_copy]). 797 */ 798 799 if (UVM_ET_ISNEEDSCOPY(ufi.entry)) { 800 KASSERT(fault_flag != UVM_FAULT_WIREMAX); 801 if (cow_now || (ufi.entry->object.uvm_obj == NULL)) { 802 /* need to clear */ 803 UVMHIST_LOG(maphist, 804 " need to clear needs_copy and refault",0,0,0,0); 805 uvmfault_unlockmaps(&ufi, FALSE); 806 uvmfault_amapcopy(&ufi); 807 uvmexp.fltamcopy++; 808 goto ReFault; 809 810 } else { 811 812 /* 813 * ensure that we pmap_enter page R/O since 814 * needs_copy is still true 815 */ 816 817 enter_prot &= ~VM_PROT_WRITE; 818 } 819 } 820 821 /* 822 * identify the players 823 */ 824 825 amap = ufi.entry->aref.ar_amap; /* top layer */ 826 uobj = ufi.entry->object.uvm_obj; /* bottom layer */ 827 828 /* 829 * check for a case 0 fault. if nothing backing the entry then 830 * error now. 831 */ 832 833 if (amap == NULL && uobj == NULL) { 834 uvmfault_unlockmaps(&ufi, FALSE); 835 UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0); 836 error = EFAULT; 837 goto done; 838 } 839 840 /* 841 * establish range of interest based on advice from mapper 842 * and then clip to fit map entry. note that we only want 843 * to do this the first time through the fault. if we 844 * ReFault we will disable this by setting "narrow" to true. 845 */ 846 847 if (narrow == FALSE) { 848 849 /* wide fault (!narrow) */ 850 KASSERT(uvmadvice[ufi.entry->advice].advice == 851 ufi.entry->advice); 852 nback = MIN(uvmadvice[ufi.entry->advice].nback, 853 (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT); 854 startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT); 855 nforw = MIN(uvmadvice[ufi.entry->advice].nforw, 856 ((ufi.entry->end - ufi.orig_rvaddr) >> 857 PAGE_SHIFT) - 1); 858 /* 859 * note: "-1" because we don't want to count the 860 * faulting page as forw 861 */ 862 npages = nback + nforw + 1; 863 centeridx = nback; 864 865 narrow = TRUE; /* ensure only once per-fault */ 866 867 } else { 868 869 /* narrow fault! */ 870 nback = nforw = 0; 871 startva = ufi.orig_rvaddr; 872 npages = 1; 873 centeridx = 0; 874 875 } 876 877 /* locked: maps(read) */ 878 UVMHIST_LOG(maphist, " narrow=%d, back=%d, forw=%d, startva=0x%x", 879 narrow, nback, nforw, startva); 880 UVMHIST_LOG(maphist, " entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry, 881 amap, uobj, 0); 882 883 /* 884 * if we've got an amap, lock it and extract current anons. 885 */ 886 887 if (amap) { 888 amap_lock(amap); 889 anons = anons_store; 890 amap_lookups(&ufi.entry->aref, startva - ufi.entry->start, 891 anons, npages); 892 } else { 893 anons = NULL; /* to be safe */ 894 } 895 896 /* locked: maps(read), amap(if there) */ 897 LOCK_ASSERT(amap == NULL || simple_lock_held(&amap->am_l)); 898 899 /* 900 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages 901 * now and then forget about them (for the rest of the fault). 902 */ 903 904 if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) { 905 906 UVMHIST_LOG(maphist, " MADV_SEQUENTIAL: flushing backpages", 907 0,0,0,0); 908 /* flush back-page anons? */ 909 if (amap) 910 uvmfault_anonflush(anons, nback); 911 912 /* flush object? */ 913 if (uobj) { 914 uoff = (startva - ufi.entry->start) + ufi.entry->offset; 915 simple_lock(&uobj->vmobjlock); 916 (void) (uobj->pgops->pgo_put)(uobj, uoff, uoff + 917 (nback << PAGE_SHIFT), PGO_DEACTIVATE); 918 } 919 920 /* now forget about the backpages */ 921 if (amap) 922 anons += nback; 923 startva += (nback << PAGE_SHIFT); 924 npages -= nback; 925 nback = centeridx = 0; 926 } 927 928 /* locked: maps(read), amap(if there) */ 929 LOCK_ASSERT(amap == NULL || simple_lock_held(&amap->am_l)); 930 931 /* 932 * map in the backpages and frontpages we found in the amap in hopes 933 * of preventing future faults. we also init the pages[] array as 934 * we go. 935 */ 936 937 currva = startva; 938 shadowed = FALSE; 939 for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) { 940 941 /* 942 * dont play with VAs that are already mapped 943 * except for center) 944 */ 945 if (lcv != centeridx && 946 pmap_extract(ufi.orig_map->pmap, currva, NULL)) { 947 pages[lcv] = PGO_DONTCARE; 948 continue; 949 } 950 951 /* 952 * unmapped or center page. check if any anon at this level. 953 */ 954 if (amap == NULL || anons[lcv] == NULL) { 955 pages[lcv] = NULL; 956 continue; 957 } 958 959 /* 960 * check for present page and map if possible. re-activate it. 961 */ 962 963 pages[lcv] = PGO_DONTCARE; 964 if (lcv == centeridx) { /* save center for later! */ 965 shadowed = TRUE; 966 continue; 967 } 968 anon = anons[lcv]; 969 simple_lock(&anon->an_lock); 970 /* ignore loaned pages */ 971 if (anon->an_page && anon->an_page->loan_count == 0 && 972 (anon->an_page->flags & PG_BUSY) == 0) { 973 uvm_lock_pageq(); 974 uvm_pageenqueue(anon->an_page); 975 uvm_unlock_pageq(); 976 UVMHIST_LOG(maphist, 977 " MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x", 978 ufi.orig_map->pmap, currva, anon->an_page, 0); 979 uvmexp.fltnamap++; 980 981 /* 982 * Since this isn't the page that's actually faulting, 983 * ignore pmap_enter() failures; it's not critical 984 * that we enter these right now. 985 */ 986 987 (void) pmap_enter(ufi.orig_map->pmap, currva, 988 VM_PAGE_TO_PHYS(anon->an_page), 989 (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) : 990 enter_prot, 991 PMAP_CANFAIL | 992 (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0)); 993 } 994 simple_unlock(&anon->an_lock); 995 pmap_update(ufi.orig_map->pmap); 996 } 997 998 /* locked: maps(read), amap(if there) */ 999 LOCK_ASSERT(amap == NULL || simple_lock_held(&amap->am_l)); 1000 /* (shadowed == TRUE) if there is an anon at the faulting address */ 1001 UVMHIST_LOG(maphist, " shadowed=%d, will_get=%d", shadowed, 1002 (uobj && shadowed == FALSE),0,0); 1003 1004 /* 1005 * note that if we are really short of RAM we could sleep in the above 1006 * call to pmap_enter with everything locked. bad? 1007 * 1008 * XXX Actually, that is bad; pmap_enter() should just fail in that 1009 * XXX case. --thorpej 1010 */ 1011 1012 /* 1013 * if the desired page is not shadowed by the amap and we have a 1014 * backing object, then we check to see if the backing object would 1015 * prefer to handle the fault itself (rather than letting us do it 1016 * with the usual pgo_get hook). the backing object signals this by 1017 * providing a pgo_fault routine. 1018 */ 1019 1020 if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) { 1021 simple_lock(&uobj->vmobjlock); 1022 1023 /* locked: maps(read), amap (if there), uobj */ 1024 error = uobj->pgops->pgo_fault(&ufi, startva, pages, npages, 1025 centeridx, access_type, PGO_LOCKED|PGO_SYNCIO); 1026 1027 /* locked: nothing, pgo_fault has unlocked everything */ 1028 1029 if (error == ERESTART) 1030 goto ReFault; /* try again! */ 1031 /* 1032 * object fault routine responsible for pmap_update(). 1033 */ 1034 goto done; 1035 } 1036 1037 /* 1038 * now, if the desired page is not shadowed by the amap and we have 1039 * a backing object that does not have a special fault routine, then 1040 * we ask (with pgo_get) the object for resident pages that we care 1041 * about and attempt to map them in. we do not let pgo_get block 1042 * (PGO_LOCKED). 1043 */ 1044 1045 if (uobj && shadowed == FALSE) { 1046 simple_lock(&uobj->vmobjlock); 1047 1048 /* locked (!shadowed): maps(read), amap (if there), uobj */ 1049 /* 1050 * the following call to pgo_get does _not_ change locking state 1051 */ 1052 1053 uvmexp.fltlget++; 1054 gotpages = npages; 1055 (void) uobj->pgops->pgo_get(uobj, ufi.entry->offset + 1056 (startva - ufi.entry->start), 1057 pages, &gotpages, centeridx, 1058 access_type & MASK(ufi.entry), 1059 ufi.entry->advice, PGO_LOCKED); 1060 1061 /* 1062 * check for pages to map, if we got any 1063 */ 1064 1065 uobjpage = NULL; 1066 1067 if (gotpages) { 1068 currva = startva; 1069 for (lcv = 0; lcv < npages; 1070 lcv++, currva += PAGE_SIZE) { 1071 struct vm_page *curpg; 1072 boolean_t readonly; 1073 1074 curpg = pages[lcv]; 1075 if (curpg == NULL || curpg == PGO_DONTCARE) { 1076 continue; 1077 } 1078 KASSERT(curpg->uobject == uobj); 1079 1080 /* 1081 * if center page is resident and not 1082 * PG_BUSY|PG_RELEASED then pgo_get 1083 * made it PG_BUSY for us and gave 1084 * us a handle to it. remember this 1085 * page as "uobjpage." (for later use). 1086 */ 1087 1088 if (lcv == centeridx) { 1089 uobjpage = curpg; 1090 UVMHIST_LOG(maphist, " got uobjpage " 1091 "(0x%x) with locked get", 1092 uobjpage, 0,0,0); 1093 continue; 1094 } 1095 1096 /* 1097 * calling pgo_get with PGO_LOCKED returns us 1098 * pages which are neither busy nor released, 1099 * so we don't need to check for this. 1100 * we can just directly enter the pages. 1101 */ 1102 1103 uvm_lock_pageq(); 1104 uvm_pageenqueue(curpg); 1105 uvm_unlock_pageq(); 1106 UVMHIST_LOG(maphist, 1107 " MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x", 1108 ufi.orig_map->pmap, currva, curpg, 0); 1109 uvmexp.fltnomap++; 1110 1111 /* 1112 * Since this page isn't the page that's 1113 * actually faulting, ignore pmap_enter() 1114 * failures; it's not critical that we 1115 * enter these right now. 1116 */ 1117 KASSERT((curpg->flags & PG_PAGEOUT) == 0); 1118 KASSERT((curpg->flags & PG_RELEASED) == 0); 1119 KASSERT(!UVM_OBJ_IS_CLEAN(curpg->uobject) || 1120 (curpg->flags & PG_CLEAN) != 0); 1121 readonly = (curpg->flags & PG_RDONLY) 1122 || (curpg->loan_count > 0) 1123 || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject); 1124 1125 (void) pmap_enter(ufi.orig_map->pmap, currva, 1126 VM_PAGE_TO_PHYS(curpg), 1127 readonly ? 1128 enter_prot & ~VM_PROT_WRITE : 1129 enter_prot & MASK(ufi.entry), 1130 PMAP_CANFAIL | 1131 (wired ? PMAP_WIRED : 0)); 1132 1133 /* 1134 * NOTE: page can't be PG_WANTED or PG_RELEASED 1135 * because we've held the lock the whole time 1136 * we've had the handle. 1137 */ 1138 KASSERT((curpg->flags & PG_WANTED) == 0); 1139 KASSERT((curpg->flags & PG_RELEASED) == 0); 1140 1141 curpg->flags &= ~(PG_BUSY); 1142 UVM_PAGE_OWN(curpg, NULL); 1143 } 1144 pmap_update(ufi.orig_map->pmap); 1145 } 1146 } else { 1147 uobjpage = NULL; 1148 } 1149 1150 /* locked (shadowed): maps(read), amap */ 1151 /* locked (!shadowed): maps(read), amap(if there), 1152 uobj(if !null), uobjpage(if !null) */ 1153 if (shadowed) { 1154 LOCK_ASSERT(simple_lock_held(&amap->am_l)); 1155 } else { 1156 LOCK_ASSERT(amap == NULL || simple_lock_held(&amap->am_l)); 1157 LOCK_ASSERT(uobj == NULL || simple_lock_held(&uobj->vmobjlock)); 1158 KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0); 1159 } 1160 1161 /* 1162 * note that at this point we are done with any front or back pages. 1163 * we are now going to focus on the center page (i.e. the one we've 1164 * faulted on). if we have faulted on the top (anon) layer 1165 * [i.e. case 1], then the anon we want is anons[centeridx] (we have 1166 * not touched it yet). if we have faulted on the bottom (uobj) 1167 * layer [i.e. case 2] and the page was both present and available, 1168 * then we've got a pointer to it as "uobjpage" and we've already 1169 * made it BUSY. 1170 */ 1171 1172 /* 1173 * there are four possible cases we must address: 1A, 1B, 2A, and 2B 1174 */ 1175 1176 /* 1177 * redirect case 2: if we are not shadowed, go to case 2. 1178 */ 1179 1180 if (shadowed == FALSE) 1181 goto Case2; 1182 1183 /* locked: maps(read), amap */ 1184 1185 /* 1186 * handle case 1: fault on an anon in our amap 1187 */ 1188 1189 anon = anons[centeridx]; 1190 UVMHIST_LOG(maphist, " case 1 fault: anon=0x%x", anon, 0,0,0); 1191 simple_lock(&anon->an_lock); 1192 1193 /* locked: maps(read), amap, anon */ 1194 LOCK_ASSERT(simple_lock_held(&amap->am_l)); 1195 LOCK_ASSERT(simple_lock_held(&anon->an_lock)); 1196 1197 /* 1198 * no matter if we have case 1A or case 1B we are going to need to 1199 * have the anon's memory resident. ensure that now. 1200 */ 1201 1202 /* 1203 * let uvmfault_anonget do the dirty work. 1204 * if it fails (!OK) it will unlock everything for us. 1205 * if it succeeds, locks are still valid and locked. 1206 * also, if it is OK, then the anon's page is on the queues. 1207 * if the page is on loan from a uvm_object, then anonget will 1208 * lock that object for us if it does not fail. 1209 */ 1210 1211 error = uvmfault_anonget(&ufi, amap, anon); 1212 switch (error) { 1213 case 0: 1214 break; 1215 1216 case ERESTART: 1217 goto ReFault; 1218 1219 case EAGAIN: 1220 tsleep(&lbolt, PVM, "fltagain1", 0); 1221 goto ReFault; 1222 1223 default: 1224 goto done; 1225 } 1226 1227 /* 1228 * uobj is non null if the page is on loan from an object (i.e. uobj) 1229 */ 1230 1231 uobj = anon->an_page->uobject; /* locked by anonget if !NULL */ 1232 1233 /* locked: maps(read), amap, anon, uobj(if one) */ 1234 LOCK_ASSERT(simple_lock_held(&amap->am_l)); 1235 LOCK_ASSERT(simple_lock_held(&anon->an_lock)); 1236 LOCK_ASSERT(uobj == NULL || simple_lock_held(&uobj->vmobjlock)); 1237 1238 /* 1239 * special handling for loaned pages 1240 */ 1241 1242 if (anon->an_page->loan_count) { 1243 1244 if (!cow_now) { 1245 1246 /* 1247 * for read faults on loaned pages we just cap the 1248 * protection at read-only. 1249 */ 1250 1251 enter_prot = enter_prot & ~VM_PROT_WRITE; 1252 1253 } else { 1254 /* 1255 * note that we can't allow writes into a loaned page! 1256 * 1257 * if we have a write fault on a loaned page in an 1258 * anon then we need to look at the anon's ref count. 1259 * if it is greater than one then we are going to do 1260 * a normal copy-on-write fault into a new anon (this 1261 * is not a problem). however, if the reference count 1262 * is one (a case where we would normally allow a 1263 * write directly to the page) then we need to kill 1264 * the loan before we continue. 1265 */ 1266 1267 /* >1 case is already ok */ 1268 if (anon->an_ref == 1) { 1269 1270 /* get new un-owned replacement page */ 1271 pg = uvm_pagealloc(NULL, 0, NULL, 0); 1272 if (pg == NULL) { 1273 uvmfault_unlockall(&ufi, amap, uobj, 1274 anon); 1275 uvm_wait("flt_noram2"); 1276 goto ReFault; 1277 } 1278 1279 /* 1280 * copy data, kill loan, and drop uobj lock 1281 * (if any) 1282 */ 1283 /* copy old -> new */ 1284 uvm_pagecopy(anon->an_page, pg); 1285 1286 /* force reload */ 1287 pmap_page_protect(anon->an_page, 1288 VM_PROT_NONE); 1289 uvm_lock_pageq(); /* KILL loan */ 1290 1291 anon->an_page->uanon = NULL; 1292 /* in case we owned */ 1293 anon->an_page->pqflags &= ~PQ_ANON; 1294 1295 if (uobj) { 1296 /* if we were receiver of loan */ 1297 anon->an_page->loan_count--; 1298 } else { 1299 /* 1300 * we were the lender (A->K); need 1301 * to remove the page from pageq's. 1302 */ 1303 uvm_pagedequeue(anon->an_page); 1304 } 1305 1306 if (uobj) { 1307 simple_unlock(&uobj->vmobjlock); 1308 uobj = NULL; 1309 } 1310 1311 /* install new page in anon */ 1312 anon->an_page = pg; 1313 pg->uanon = anon; 1314 pg->pqflags |= PQ_ANON; 1315 1316 uvm_pageactivate(pg); 1317 uvm_unlock_pageq(); 1318 1319 pg->flags &= ~(PG_BUSY|PG_FAKE); 1320 UVM_PAGE_OWN(pg, NULL); 1321 1322 /* done! */ 1323 } /* ref == 1 */ 1324 } /* write fault */ 1325 } /* loan count */ 1326 1327 /* 1328 * if we are case 1B then we will need to allocate a new blank 1329 * anon to transfer the data into. note that we have a lock 1330 * on anon, so no one can busy or release the page until we are done. 1331 * also note that the ref count can't drop to zero here because 1332 * it is > 1 and we are only dropping one ref. 1333 * 1334 * in the (hopefully very rare) case that we are out of RAM we 1335 * will unlock, wait for more RAM, and refault. 1336 * 1337 * if we are out of anon VM we kill the process (XXX: could wait?). 1338 */ 1339 1340 if (cow_now && anon->an_ref > 1) { 1341 1342 UVMHIST_LOG(maphist, " case 1B: COW fault",0,0,0,0); 1343 uvmexp.flt_acow++; 1344 oanon = anon; /* oanon = old, locked anon */ 1345 1346 error = uvmfault_promote(&ufi, oanon, PGO_DONTCARE, 1347 &anon, &anon_spare); 1348 switch (error) { 1349 case 0: 1350 break; 1351 case ERESTART: 1352 goto ReFault; 1353 default: 1354 goto done; 1355 } 1356 1357 pg = anon->an_page; 1358 uvm_lock_pageq(); 1359 uvm_pageactivate(pg); 1360 uvm_unlock_pageq(); 1361 pg->flags &= ~(PG_BUSY|PG_FAKE); 1362 UVM_PAGE_OWN(pg, NULL); 1363 1364 /* deref: can not drop to zero here by defn! */ 1365 oanon->an_ref--; 1366 1367 /* 1368 * note: oanon is still locked, as is the new anon. we 1369 * need to check for this later when we unlock oanon; if 1370 * oanon != anon, we'll have to unlock anon, too. 1371 */ 1372 1373 } else { 1374 1375 uvmexp.flt_anon++; 1376 oanon = anon; /* old, locked anon is same as anon */ 1377 pg = anon->an_page; 1378 if (anon->an_ref > 1) /* disallow writes to ref > 1 anons */ 1379 enter_prot = enter_prot & ~VM_PROT_WRITE; 1380 1381 } 1382 1383 /* locked: maps(read), amap, oanon, anon (if different from oanon) */ 1384 LOCK_ASSERT(simple_lock_held(&amap->am_l)); 1385 LOCK_ASSERT(simple_lock_held(&anon->an_lock)); 1386 LOCK_ASSERT(simple_lock_held(&oanon->an_lock)); 1387 1388 /* 1389 * now map the page in. 1390 */ 1391 1392 UVMHIST_LOG(maphist, " MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x", 1393 ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0); 1394 if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg), 1395 enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) 1396 != 0) { 1397 1398 /* 1399 * No need to undo what we did; we can simply think of 1400 * this as the pmap throwing away the mapping information. 1401 * 1402 * We do, however, have to go through the ReFault path, 1403 * as the map may change while we're asleep. 1404 */ 1405 1406 if (anon != oanon) 1407 simple_unlock(&anon->an_lock); 1408 uvmfault_unlockall(&ufi, amap, uobj, oanon); 1409 if (!uvm_reclaimable()) { 1410 UVMHIST_LOG(maphist, 1411 "<- failed. out of VM",0,0,0,0); 1412 /* XXX instrumentation */ 1413 error = ENOMEM; 1414 goto done; 1415 } 1416 /* XXX instrumentation */ 1417 uvm_wait("flt_pmfail1"); 1418 goto ReFault; 1419 } 1420 1421 /* 1422 * ... update the page queues. 1423 */ 1424 1425 uvm_lock_pageq(); 1426 if (wire_fault) { 1427 uvm_pagewire(pg); 1428 1429 /* 1430 * since the now-wired page cannot be paged out, 1431 * release its swap resources for others to use. 1432 * since an anon with no swap cannot be PG_CLEAN, 1433 * clear its clean flag now. 1434 */ 1435 1436 pg->flags &= ~(PG_CLEAN); 1437 uvm_anon_dropswap(anon); 1438 } else { 1439 uvm_pageactivate(pg); 1440 } 1441 uvm_unlock_pageq(); 1442 1443 /* 1444 * done case 1! finish up by unlocking everything and returning success 1445 */ 1446 1447 if (anon != oanon) 1448 simple_unlock(&anon->an_lock); 1449 uvmfault_unlockall(&ufi, amap, uobj, oanon); 1450 pmap_update(ufi.orig_map->pmap); 1451 error = 0; 1452 goto done; 1453 1454 Case2: 1455 /* 1456 * handle case 2: faulting on backing object or zero fill 1457 */ 1458 1459 /* 1460 * locked: 1461 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null) 1462 */ 1463 LOCK_ASSERT(amap == NULL || simple_lock_held(&amap->am_l)); 1464 LOCK_ASSERT(uobj == NULL || simple_lock_held(&uobj->vmobjlock)); 1465 LOCK_ASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0); 1466 1467 /* 1468 * note that uobjpage can not be PGO_DONTCARE at this point. we now 1469 * set uobjpage to PGO_DONTCARE if we are doing a zero fill. if we 1470 * have a backing object, check and see if we are going to promote 1471 * the data up to an anon during the fault. 1472 */ 1473 1474 if (uobj == NULL) { 1475 uobjpage = PGO_DONTCARE; 1476 promote = TRUE; /* always need anon here */ 1477 } else { 1478 KASSERT(uobjpage != PGO_DONTCARE); 1479 promote = cow_now && UVM_ET_ISCOPYONWRITE(ufi.entry); 1480 } 1481 UVMHIST_LOG(maphist, " case 2 fault: promote=%d, zfill=%d", 1482 promote, (uobj == NULL), 0,0); 1483 1484 /* 1485 * if uobjpage is not null then we do not need to do I/O to get the 1486 * uobjpage. 1487 * 1488 * if uobjpage is null, then we need to unlock and ask the pager to 1489 * get the data for us. once we have the data, we need to reverify 1490 * the state the world. we are currently not holding any resources. 1491 */ 1492 1493 if (uobjpage) { 1494 /* update rusage counters */ 1495 curproc->p_stats->p_ru.ru_minflt++; 1496 } else { 1497 /* update rusage counters */ 1498 curproc->p_stats->p_ru.ru_majflt++; 1499 1500 /* locked: maps(read), amap(if there), uobj */ 1501 uvmfault_unlockall(&ufi, amap, NULL, NULL); 1502 /* locked: uobj */ 1503 1504 uvmexp.fltget++; 1505 gotpages = 1; 1506 uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset; 1507 error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages, 1508 0, access_type & MASK(ufi.entry), ufi.entry->advice, 1509 PGO_SYNCIO); 1510 /* locked: uobjpage(if no error) */ 1511 LOCK_ASSERT(error != 0 || (uobjpage->flags & PG_BUSY) != 0); 1512 1513 /* 1514 * recover from I/O 1515 */ 1516 1517 if (error) { 1518 if (error == EAGAIN) { 1519 UVMHIST_LOG(maphist, 1520 " pgo_get says TRY AGAIN!",0,0,0,0); 1521 tsleep(&lbolt, PVM, "fltagain2", 0); 1522 goto ReFault; 1523 } 1524 1525 UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)", 1526 error, 0,0,0); 1527 goto done; 1528 } 1529 1530 /* locked: uobjpage */ 1531 1532 uvm_lock_pageq(); 1533 uvm_pageactivate(uobjpage); 1534 uvm_unlock_pageq(); 1535 1536 /* 1537 * re-verify the state of the world by first trying to relock 1538 * the maps. always relock the object. 1539 */ 1540 1541 locked = uvmfault_relock(&ufi); 1542 if (locked && amap) 1543 amap_lock(amap); 1544 uobj = uobjpage->uobject; 1545 simple_lock(&uobj->vmobjlock); 1546 1547 /* locked(locked): maps(read), amap(if !null), uobj, uobjpage */ 1548 /* locked(!locked): uobj, uobjpage */ 1549 1550 /* 1551 * verify that the page has not be released and re-verify 1552 * that amap slot is still free. if there is a problem, 1553 * we unlock and clean up. 1554 */ 1555 1556 if ((uobjpage->flags & PG_RELEASED) != 0 || 1557 (locked && amap && 1558 amap_lookup(&ufi.entry->aref, 1559 ufi.orig_rvaddr - ufi.entry->start))) { 1560 if (locked) 1561 uvmfault_unlockall(&ufi, amap, NULL, NULL); 1562 locked = FALSE; 1563 } 1564 1565 /* 1566 * didn't get the lock? release the page and retry. 1567 */ 1568 1569 if (locked == FALSE) { 1570 UVMHIST_LOG(maphist, 1571 " wasn't able to relock after fault: retry", 1572 0,0,0,0); 1573 if (uobjpage->flags & PG_WANTED) 1574 wakeup(uobjpage); 1575 if (uobjpage->flags & PG_RELEASED) { 1576 uvmexp.fltpgrele++; 1577 uvm_pagefree(uobjpage); 1578 goto ReFault; 1579 } 1580 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1581 UVM_PAGE_OWN(uobjpage, NULL); 1582 simple_unlock(&uobj->vmobjlock); 1583 goto ReFault; 1584 } 1585 1586 /* 1587 * we have the data in uobjpage which is busy and 1588 * not released. we are holding object lock (so the page 1589 * can't be released on us). 1590 */ 1591 1592 /* locked: maps(read), amap(if !null), uobj, uobjpage */ 1593 } 1594 1595 /* 1596 * locked: 1597 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj) 1598 */ 1599 LOCK_ASSERT(amap == NULL || simple_lock_held(&amap->am_l)); 1600 LOCK_ASSERT(uobj == NULL || simple_lock_held(&uobj->vmobjlock)); 1601 LOCK_ASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0); 1602 1603 /* 1604 * notes: 1605 * - at this point uobjpage can not be NULL 1606 * - at this point uobjpage can not be PG_RELEASED (since we checked 1607 * for it above) 1608 * - at this point uobjpage could be PG_WANTED (handle later) 1609 */ 1610 1611 KASSERT(uobj == NULL || uobj == uobjpage->uobject); 1612 KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) || 1613 (uobjpage->flags & PG_CLEAN) != 0); 1614 if (promote == FALSE) { 1615 1616 /* 1617 * we are not promoting. if the mapping is COW ensure that we 1618 * don't give more access than we should (e.g. when doing a read 1619 * fault on a COPYONWRITE mapping we want to map the COW page in 1620 * R/O even though the entry protection could be R/W). 1621 * 1622 * set "pg" to the page we want to map in (uobjpage, usually) 1623 */ 1624 1625 /* no anon in this case. */ 1626 anon = NULL; 1627 1628 uvmexp.flt_obj++; 1629 if (UVM_ET_ISCOPYONWRITE(ufi.entry) || 1630 UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject)) 1631 enter_prot &= ~VM_PROT_WRITE; 1632 pg = uobjpage; /* map in the actual object */ 1633 1634 KASSERT(uobjpage != PGO_DONTCARE); 1635 1636 /* 1637 * we are faulting directly on the page. be careful 1638 * about writing to loaned pages... 1639 */ 1640 1641 if (uobjpage->loan_count) { 1642 if (!cow_now) { 1643 /* read fault: cap the protection at readonly */ 1644 /* cap! */ 1645 enter_prot = enter_prot & ~VM_PROT_WRITE; 1646 } else { 1647 /* write fault: must break the loan here */ 1648 1649 pg = uvm_loanbreak(uobjpage); 1650 if (pg == NULL) { 1651 1652 /* 1653 * drop ownership of page, it can't 1654 * be released 1655 */ 1656 1657 if (uobjpage->flags & PG_WANTED) 1658 wakeup(uobjpage); 1659 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1660 UVM_PAGE_OWN(uobjpage, NULL); 1661 1662 uvmfault_unlockall(&ufi, amap, uobj, 1663 NULL); 1664 UVMHIST_LOG(maphist, 1665 " out of RAM breaking loan, waiting", 1666 0,0,0,0); 1667 uvmexp.fltnoram++; 1668 uvm_wait("flt_noram4"); 1669 goto ReFault; 1670 } 1671 uobjpage = pg; 1672 } 1673 } 1674 } else { 1675 1676 /* 1677 * if we are going to promote the data to an anon we 1678 * allocate a blank anon here and plug it into our amap. 1679 */ 1680 #if DIAGNOSTIC 1681 if (amap == NULL) 1682 panic("uvm_fault: want to promote data, but no anon"); 1683 #endif 1684 error = uvmfault_promote(&ufi, NULL, uobjpage, 1685 &anon, &anon_spare); 1686 switch (error) { 1687 case 0: 1688 break; 1689 case ERESTART: 1690 goto ReFault; 1691 default: 1692 goto done; 1693 } 1694 1695 pg = anon->an_page; 1696 1697 /* 1698 * fill in the data 1699 */ 1700 1701 if (uobjpage != PGO_DONTCARE) { 1702 uvmexp.flt_prcopy++; 1703 1704 /* 1705 * promote to shared amap? make sure all sharing 1706 * procs see it 1707 */ 1708 1709 if ((amap_flags(amap) & AMAP_SHARED) != 0) { 1710 pmap_page_protect(uobjpage, VM_PROT_NONE); 1711 /* 1712 * XXX: PAGE MIGHT BE WIRED! 1713 */ 1714 } 1715 1716 /* 1717 * dispose of uobjpage. it can't be PG_RELEASED 1718 * since we still hold the object lock. 1719 * drop handle to uobj as well. 1720 */ 1721 1722 if (uobjpage->flags & PG_WANTED) 1723 /* still have the obj lock */ 1724 wakeup(uobjpage); 1725 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1726 UVM_PAGE_OWN(uobjpage, NULL); 1727 simple_unlock(&uobj->vmobjlock); 1728 uobj = NULL; 1729 1730 UVMHIST_LOG(maphist, 1731 " promote uobjpage 0x%x to anon/page 0x%x/0x%x", 1732 uobjpage, anon, pg, 0); 1733 1734 } else { 1735 uvmexp.flt_przero++; 1736 1737 /* 1738 * Page is zero'd and marked dirty by 1739 * uvmfault_promote(). 1740 */ 1741 1742 UVMHIST_LOG(maphist," zero fill anon/page 0x%x/0%x", 1743 anon, pg, 0, 0); 1744 } 1745 } 1746 1747 /* 1748 * locked: 1749 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj), 1750 * anon(if !null), pg(if anon) 1751 * 1752 * note: pg is either the uobjpage or the new page in the new anon 1753 */ 1754 LOCK_ASSERT(amap == NULL || simple_lock_held(&amap->am_l)); 1755 LOCK_ASSERT(uobj == NULL || simple_lock_held(&uobj->vmobjlock)); 1756 LOCK_ASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0); 1757 LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock)); 1758 LOCK_ASSERT((pg->flags & PG_BUSY) != 0); 1759 1760 /* 1761 * all resources are present. we can now map it in and free our 1762 * resources. 1763 */ 1764 1765 UVMHIST_LOG(maphist, 1766 " MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d", 1767 ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote); 1768 KASSERT((access_type & VM_PROT_WRITE) == 0 || 1769 (pg->flags & PG_RDONLY) == 0); 1770 if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg), 1771 pg->flags & PG_RDONLY ? enter_prot & ~VM_PROT_WRITE : enter_prot, 1772 access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) { 1773 1774 /* 1775 * No need to undo what we did; we can simply think of 1776 * this as the pmap throwing away the mapping information. 1777 * 1778 * We do, however, have to go through the ReFault path, 1779 * as the map may change while we're asleep. 1780 */ 1781 1782 if (pg->flags & PG_WANTED) 1783 wakeup(pg); 1784 1785 /* 1786 * note that pg can't be PG_RELEASED since we did not drop 1787 * the object lock since the last time we checked. 1788 */ 1789 KASSERT((pg->flags & PG_RELEASED) == 0); 1790 1791 pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED); 1792 UVM_PAGE_OWN(pg, NULL); 1793 uvmfault_unlockall(&ufi, amap, uobj, anon); 1794 if (!uvm_reclaimable()) { 1795 UVMHIST_LOG(maphist, 1796 "<- failed. out of VM",0,0,0,0); 1797 /* XXX instrumentation */ 1798 error = ENOMEM; 1799 goto done; 1800 } 1801 /* XXX instrumentation */ 1802 uvm_wait("flt_pmfail2"); 1803 goto ReFault; 1804 } 1805 1806 uvm_lock_pageq(); 1807 if (wire_fault) { 1808 uvm_pagewire(pg); 1809 if (pg->pqflags & PQ_AOBJ) { 1810 1811 /* 1812 * since the now-wired page cannot be paged out, 1813 * release its swap resources for others to use. 1814 * since an aobj page with no swap cannot be PG_CLEAN, 1815 * clear its clean flag now. 1816 */ 1817 1818 KASSERT(uobj != NULL); 1819 pg->flags &= ~(PG_CLEAN); 1820 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 1821 } 1822 } else { 1823 uvm_pageactivate(pg); 1824 } 1825 uvm_unlock_pageq(); 1826 if (pg->flags & PG_WANTED) 1827 wakeup(pg); 1828 1829 /* 1830 * note that pg can't be PG_RELEASED since we did not drop the object 1831 * lock since the last time we checked. 1832 */ 1833 KASSERT((pg->flags & PG_RELEASED) == 0); 1834 1835 pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED); 1836 UVM_PAGE_OWN(pg, NULL); 1837 uvmfault_unlockall(&ufi, amap, uobj, anon); 1838 pmap_update(ufi.orig_map->pmap); 1839 UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0); 1840 error = 0; 1841 done: 1842 if (anon_spare != NULL) { 1843 anon_spare->an_ref--; 1844 uvm_anfree(anon_spare); 1845 } 1846 return error; 1847 } 1848 1849 1850 /* 1851 * uvm_fault_wire: wire down a range of virtual addresses in a map. 1852 * 1853 * => map may be read-locked by caller, but MUST NOT be write-locked. 1854 * => if map is read-locked, any operations which may cause map to 1855 * be write-locked in uvm_fault() must be taken care of by 1856 * the caller. See uvm_map_pageable(). 1857 */ 1858 1859 int 1860 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end, 1861 vm_prot_t access_type, int wiremax) 1862 { 1863 vaddr_t va; 1864 int error; 1865 1866 /* 1867 * now fault it in a page at a time. if the fault fails then we have 1868 * to undo what we have done. note that in uvm_fault VM_PROT_NONE 1869 * is replaced with the max protection if fault_type is VM_FAULT_WIRE. 1870 */ 1871 1872 /* 1873 * XXX work around overflowing a vaddr_t. this prevents us from 1874 * wiring the last page in the address space, though. 1875 */ 1876 if (start > end) { 1877 return EFAULT; 1878 } 1879 1880 for (va = start ; va < end ; va += PAGE_SIZE) { 1881 error = uvm_fault_internal(map, va, access_type, 1882 wiremax ? UVM_FAULT_WIREMAX : UVM_FAULT_WIRE); 1883 if (error) { 1884 if (va != start) { 1885 uvm_fault_unwire(map, start, va); 1886 } 1887 return error; 1888 } 1889 } 1890 return 0; 1891 } 1892 1893 /* 1894 * uvm_fault_unwire(): unwire range of virtual space. 1895 */ 1896 1897 void 1898 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end) 1899 { 1900 vm_map_lock_read(map); 1901 uvm_fault_unwire_locked(map, start, end); 1902 vm_map_unlock_read(map); 1903 } 1904 1905 /* 1906 * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire(). 1907 * 1908 * => map must be at least read-locked. 1909 */ 1910 1911 void 1912 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end) 1913 { 1914 struct vm_map_entry *entry; 1915 pmap_t pmap = vm_map_pmap(map); 1916 vaddr_t va; 1917 paddr_t pa; 1918 struct vm_page *pg; 1919 1920 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 1921 1922 /* 1923 * we assume that the area we are unwiring has actually been wired 1924 * in the first place. this means that we should be able to extract 1925 * the PAs from the pmap. we also lock out the page daemon so that 1926 * we can call uvm_pageunwire. 1927 */ 1928 1929 uvm_lock_pageq(); 1930 1931 /* 1932 * find the beginning map entry for the region. 1933 */ 1934 1935 KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map)); 1936 if (uvm_map_lookup_entry(map, start, &entry) == FALSE) 1937 panic("uvm_fault_unwire_locked: address not in map"); 1938 1939 for (va = start; va < end; va += PAGE_SIZE) { 1940 if (pmap_extract(pmap, va, &pa) == FALSE) 1941 continue; 1942 1943 /* 1944 * find the map entry for the current address. 1945 */ 1946 1947 KASSERT(va >= entry->start); 1948 while (va >= entry->end) { 1949 KASSERT(entry->next != &map->header && 1950 entry->next->start <= entry->end); 1951 entry = entry->next; 1952 } 1953 1954 /* 1955 * if the entry is no longer wired, tell the pmap. 1956 */ 1957 1958 if (VM_MAPENT_ISWIRED(entry) == 0) 1959 pmap_unwire(pmap, va); 1960 1961 pg = PHYS_TO_VM_PAGE(pa); 1962 if (pg) 1963 uvm_pageunwire(pg); 1964 } 1965 1966 uvm_unlock_pageq(); 1967 } 1968