1 /* $NetBSD: uvm_fault.c,v 1.124 2008/03/27 19:06:52 ad 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 1.124 2008/03/27 19:06:52 ad 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 mutex_enter(&anons[lcv]->an_lock); 202 pg = anons[lcv]->an_page; 203 if (pg && (pg->flags & PG_BUSY) == 0) { 204 mutex_enter(&uvm_pageqlock); 205 if (pg->wire_count == 0) { 206 uvm_pagedeactivate(pg); 207 } 208 mutex_exit(&uvm_pageqlock); 209 } 210 mutex_exit(&anons[lcv]->an_lock); 211 } 212 } 213 214 /* 215 * normal functions 216 */ 217 218 /* 219 * uvmfault_amapcopy: clear "needs_copy" in a map. 220 * 221 * => called with VM data structures unlocked (usually, see below) 222 * => we get a write lock on the maps and clear needs_copy for a VA 223 * => if we are out of RAM we sleep (waiting for more) 224 */ 225 226 static void 227 uvmfault_amapcopy(struct uvm_faultinfo *ufi) 228 { 229 for (;;) { 230 231 /* 232 * no mapping? give up. 233 */ 234 235 if (uvmfault_lookup(ufi, true) == false) 236 return; 237 238 /* 239 * copy if needed. 240 */ 241 242 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) 243 amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT, 244 ufi->orig_rvaddr, ufi->orig_rvaddr + 1); 245 246 /* 247 * didn't work? must be out of RAM. unlock and sleep. 248 */ 249 250 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) { 251 uvmfault_unlockmaps(ufi, true); 252 uvm_wait("fltamapcopy"); 253 continue; 254 } 255 256 /* 257 * got it! unlock and return. 258 */ 259 260 uvmfault_unlockmaps(ufi, true); 261 return; 262 } 263 /*NOTREACHED*/ 264 } 265 266 /* 267 * uvmfault_anonget: get data in an anon into a non-busy, non-released 268 * page in that anon. 269 * 270 * => maps, amap, and anon locked by caller. 271 * => if we fail (result != 0) we unlock everything. 272 * => if we are successful, we return with everything still locked. 273 * => we don't move the page on the queues [gets moved later] 274 * => if we allocate a new page [we_own], it gets put on the queues. 275 * either way, the result is that the page is on the queues at return time 276 * => for pages which are on loan from a uvm_object (and thus are not 277 * owned by the anon): if successful, we return with the owning object 278 * locked. the caller must unlock this object when it unlocks everything 279 * else. 280 */ 281 282 int 283 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap, 284 struct vm_anon *anon) 285 { 286 bool we_own; /* we own anon's page? */ 287 bool locked; /* did we relock? */ 288 struct vm_page *pg; 289 int error; 290 UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist); 291 292 KASSERT(mutex_owned(&anon->an_lock)); 293 294 error = 0; 295 uvmexp.fltanget++; 296 /* bump rusage counters */ 297 if (anon->an_page) 298 curlwp->l_ru.ru_minflt++; 299 else 300 curlwp->l_ru.ru_majflt++; 301 302 /* 303 * loop until we get it, or fail. 304 */ 305 306 for (;;) { 307 we_own = false; /* true if we set PG_BUSY on a page */ 308 pg = anon->an_page; 309 310 /* 311 * if there is a resident page and it is loaned, then anon 312 * may not own it. call out to uvm_anon_lockpage() to ensure 313 * the real owner of the page has been identified and locked. 314 */ 315 316 if (pg && pg->loan_count) 317 pg = uvm_anon_lockloanpg(anon); 318 319 /* 320 * page there? make sure it is not busy/released. 321 */ 322 323 if (pg) { 324 325 /* 326 * at this point, if the page has a uobject [meaning 327 * we have it on loan], then that uobject is locked 328 * by us! if the page is busy, we drop all the 329 * locks (including uobject) and try again. 330 */ 331 332 if ((pg->flags & PG_BUSY) == 0) { 333 UVMHIST_LOG(maphist, "<- OK",0,0,0,0); 334 return (0); 335 } 336 pg->flags |= PG_WANTED; 337 uvmexp.fltpgwait++; 338 339 /* 340 * the last unlock must be an atomic unlock+wait on 341 * the owner of page 342 */ 343 344 if (pg->uobject) { /* owner is uobject ? */ 345 uvmfault_unlockall(ufi, amap, NULL, anon); 346 UVMHIST_LOG(maphist, " unlock+wait on uobj",0, 347 0,0,0); 348 UVM_UNLOCK_AND_WAIT(pg, 349 &pg->uobject->vmobjlock, 350 false, "anonget1",0); 351 } else { 352 /* anon owns page */ 353 uvmfault_unlockall(ufi, amap, NULL, NULL); 354 UVMHIST_LOG(maphist, " unlock+wait on anon",0, 355 0,0,0); 356 UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0, 357 "anonget2",0); 358 } 359 } else { 360 #if defined(VMSWAP) 361 362 /* 363 * no page, we must try and bring it in. 364 */ 365 366 pg = uvm_pagealloc(NULL, 0, anon, 0); 367 if (pg == NULL) { /* out of RAM. */ 368 uvmfault_unlockall(ufi, amap, NULL, anon); 369 uvmexp.fltnoram++; 370 UVMHIST_LOG(maphist, " noram -- UVM_WAIT",0, 371 0,0,0); 372 if (!uvm_reclaimable()) { 373 return ENOMEM; 374 } 375 uvm_wait("flt_noram1"); 376 } else { 377 /* we set the PG_BUSY bit */ 378 we_own = true; 379 uvmfault_unlockall(ufi, amap, NULL, anon); 380 381 /* 382 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN 383 * page into the uvm_swap_get function with 384 * all data structures unlocked. note that 385 * it is ok to read an_swslot here because 386 * we hold PG_BUSY on the page. 387 */ 388 uvmexp.pageins++; 389 error = uvm_swap_get(pg, anon->an_swslot, 390 PGO_SYNCIO); 391 392 /* 393 * we clean up after the i/o below in the 394 * "we_own" case 395 */ 396 } 397 #else /* defined(VMSWAP) */ 398 panic("%s: no page", __func__); 399 #endif /* defined(VMSWAP) */ 400 } 401 402 /* 403 * now relock and try again 404 */ 405 406 locked = uvmfault_relock(ufi); 407 if (locked && amap != NULL) { 408 amap_lock(amap); 409 } 410 if (locked || we_own) 411 mutex_enter(&anon->an_lock); 412 413 /* 414 * if we own the page (i.e. we set PG_BUSY), then we need 415 * to clean up after the I/O. there are three cases to 416 * consider: 417 * [1] page released during I/O: free anon and ReFault. 418 * [2] I/O not OK. free the page and cause the fault 419 * to fail. 420 * [3] I/O OK! activate the page and sync with the 421 * non-we_own case (i.e. drop anon lock if not locked). 422 */ 423 424 if (we_own) { 425 #if defined(VMSWAP) 426 if (pg->flags & PG_WANTED) { 427 wakeup(pg); 428 } 429 if (error) { 430 431 /* 432 * remove the swap slot from the anon 433 * and mark the anon as having no real slot. 434 * don't free the swap slot, thus preventing 435 * it from being used again. 436 */ 437 438 if (anon->an_swslot > 0) 439 uvm_swap_markbad(anon->an_swslot, 1); 440 anon->an_swslot = SWSLOT_BAD; 441 442 if ((pg->flags & PG_RELEASED) != 0) 443 goto released; 444 445 /* 446 * note: page was never !PG_BUSY, so it 447 * can't be mapped and thus no need to 448 * pmap_page_protect it... 449 */ 450 451 mutex_enter(&uvm_pageqlock); 452 uvm_pagefree(pg); 453 mutex_exit(&uvm_pageqlock); 454 455 if (locked) 456 uvmfault_unlockall(ufi, amap, NULL, 457 anon); 458 else 459 mutex_exit(&anon->an_lock); 460 UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0); 461 return error; 462 } 463 464 if ((pg->flags & PG_RELEASED) != 0) { 465 released: 466 KASSERT(anon->an_ref == 0); 467 468 /* 469 * released while we unlocked amap. 470 */ 471 472 if (locked) 473 uvmfault_unlockall(ufi, amap, NULL, 474 NULL); 475 476 uvm_anon_release(anon); 477 478 if (error) { 479 UVMHIST_LOG(maphist, 480 "<- ERROR/RELEASED", 0,0,0,0); 481 return error; 482 } 483 484 UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0); 485 return ERESTART; 486 } 487 488 /* 489 * we've successfully read the page, activate it. 490 */ 491 492 mutex_enter(&uvm_pageqlock); 493 uvm_pageactivate(pg); 494 mutex_exit(&uvm_pageqlock); 495 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE); 496 UVM_PAGE_OWN(pg, NULL); 497 if (!locked) 498 mutex_exit(&anon->an_lock); 499 #else /* defined(VMSWAP) */ 500 panic("%s: we_own", __func__); 501 #endif /* defined(VMSWAP) */ 502 } 503 504 /* 505 * we were not able to relock. restart fault. 506 */ 507 508 if (!locked) { 509 UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0); 510 return (ERESTART); 511 } 512 513 /* 514 * verify no one has touched the amap and moved the anon on us. 515 */ 516 517 if (ufi != NULL && 518 amap_lookup(&ufi->entry->aref, 519 ufi->orig_rvaddr - ufi->entry->start) != anon) { 520 521 uvmfault_unlockall(ufi, amap, NULL, anon); 522 UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0); 523 return (ERESTART); 524 } 525 526 /* 527 * try it again! 528 */ 529 530 uvmexp.fltanretry++; 531 continue; 532 } 533 /*NOTREACHED*/ 534 } 535 536 /* 537 * uvmfault_promote: promote data to a new anon. used for 1B and 2B. 538 * 539 * 1. allocate an anon and a page. 540 * 2. fill its contents. 541 * 3. put it into amap. 542 * 543 * => if we fail (result != 0) we unlock everything. 544 * => on success, return a new locked anon via 'nanon'. 545 * (*nanon)->an_page will be a resident, locked, dirty page. 546 */ 547 548 static int 549 uvmfault_promote(struct uvm_faultinfo *ufi, 550 struct vm_anon *oanon, 551 struct vm_page *uobjpage, 552 struct vm_anon **nanon, /* OUT: allocated anon */ 553 struct vm_anon **spare) 554 { 555 struct vm_amap *amap = ufi->entry->aref.ar_amap; 556 struct uvm_object *uobj; 557 struct vm_anon *anon; 558 struct vm_page *pg; 559 struct vm_page *opg; 560 int error; 561 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 562 563 if (oanon) { 564 /* anon COW */ 565 opg = oanon->an_page; 566 KASSERT(opg != NULL); 567 KASSERT(opg->uobject == NULL || opg->loan_count > 0); 568 } else if (uobjpage != PGO_DONTCARE) { 569 /* object-backed COW */ 570 opg = uobjpage; 571 } else { 572 /* ZFOD */ 573 opg = NULL; 574 } 575 if (opg != NULL) { 576 uobj = opg->uobject; 577 } else { 578 uobj = NULL; 579 } 580 581 KASSERT(amap != NULL); 582 KASSERT(uobjpage != NULL); 583 KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0); 584 KASSERT(mutex_owned(&amap->am_l)); 585 KASSERT(oanon == NULL || mutex_owned(&oanon->an_lock)); 586 KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock)); 587 #if 0 588 KASSERT(*spare == NULL || !mutex_owned(&(*spare)->an_lock)); 589 #endif 590 591 if (*spare != NULL) { 592 anon = *spare; 593 *spare = NULL; 594 mutex_enter(&anon->an_lock); 595 } else if (ufi->map != kernel_map) { 596 anon = uvm_analloc(); 597 } else { 598 UVMHIST_LOG(maphist, "kernel_map, unlock and retry", 0,0,0,0); 599 600 /* 601 * we can't allocate anons with kernel_map locked. 602 */ 603 604 uvm_page_unbusy(&uobjpage, 1); 605 uvmfault_unlockall(ufi, amap, uobj, oanon); 606 607 *spare = uvm_analloc(); 608 if (*spare == NULL) { 609 goto nomem; 610 } 611 mutex_exit(&(*spare)->an_lock); 612 error = ERESTART; 613 goto done; 614 } 615 if (anon) { 616 617 /* 618 * The new anon is locked. 619 * 620 * if opg == NULL, we want a zero'd, dirty page, 621 * so have uvm_pagealloc() do that for us. 622 */ 623 624 pg = uvm_pagealloc(NULL, 0, anon, 625 (opg == NULL) ? UVM_PGA_ZERO : 0); 626 } else { 627 pg = NULL; 628 } 629 630 /* 631 * out of memory resources? 632 */ 633 634 if (pg == NULL) { 635 /* save anon for the next try. */ 636 if (anon != NULL) { 637 mutex_exit(&anon->an_lock); 638 *spare = anon; 639 } 640 641 /* unlock and fail ... */ 642 uvm_page_unbusy(&uobjpage, 1); 643 uvmfault_unlockall(ufi, amap, uobj, oanon); 644 nomem: 645 if (!uvm_reclaimable()) { 646 UVMHIST_LOG(maphist, "out of VM", 0,0,0,0); 647 uvmexp.fltnoanon++; 648 error = ENOMEM; 649 goto done; 650 } 651 652 UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0); 653 uvmexp.fltnoram++; 654 uvm_wait("flt_noram5"); 655 error = ERESTART; 656 goto done; 657 } 658 659 /* copy page [pg now dirty] */ 660 if (opg) { 661 uvm_pagecopy(opg, pg); 662 } 663 664 amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon, 665 oanon != NULL); 666 667 *nanon = anon; 668 error = 0; 669 done: 670 return error; 671 } 672 673 674 /* 675 * F A U L T - m a i n e n t r y p o i n t 676 */ 677 678 /* 679 * uvm_fault: page fault handler 680 * 681 * => called from MD code to resolve a page fault 682 * => VM data structures usually should be unlocked. however, it is 683 * possible to call here with the main map locked if the caller 684 * gets a write lock, sets it recusive, and then calls us (c.f. 685 * uvm_map_pageable). this should be avoided because it keeps 686 * the map locked off during I/O. 687 * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT 688 */ 689 690 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 691 ~VM_PROT_WRITE : VM_PROT_ALL) 692 693 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */ 694 #define UVM_FAULT_WIRE 1 695 #define UVM_FAULT_WIREMAX 2 696 697 int 698 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr, 699 vm_prot_t access_type, int fault_flag) 700 { 701 struct uvm_faultinfo ufi; 702 vm_prot_t enter_prot, check_prot; 703 bool wired, narrow, promote, locked, shadowed, wire_fault, cow_now; 704 int npages, nback, nforw, centeridx, error, lcv, gotpages; 705 vaddr_t startva, currva; 706 voff_t uoff; 707 struct vm_amap *amap; 708 struct uvm_object *uobj; 709 struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon; 710 struct vm_anon *anon_spare; 711 struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage; 712 UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist); 713 714 UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, at=%d, ff=%d)", 715 orig_map, vaddr, access_type, fault_flag); 716 717 anon = anon_spare = NULL; 718 pg = NULL; 719 720 uvmexp.faults++; /* XXX: locking? */ 721 722 /* 723 * init the IN parameters in the ufi 724 */ 725 726 ufi.orig_map = orig_map; 727 ufi.orig_rvaddr = trunc_page(vaddr); 728 ufi.orig_size = PAGE_SIZE; /* can't get any smaller than this */ 729 wire_fault = (fault_flag > 0); 730 if (wire_fault) 731 narrow = true; /* don't look for neighborhood 732 * pages on wire */ 733 else 734 narrow = false; /* normal fault */ 735 736 /* 737 * "goto ReFault" means restart the page fault from ground zero. 738 */ 739 ReFault: 740 741 /* 742 * lookup and lock the maps 743 */ 744 745 if (uvmfault_lookup(&ufi, false) == false) { 746 UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0); 747 error = EFAULT; 748 goto done; 749 } 750 /* locked: maps(read) */ 751 752 #ifdef DIAGNOSTIC 753 if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) { 754 printf("Page fault on non-pageable map:\n"); 755 printf("ufi.map = %p\n", ufi.map); 756 printf("ufi.orig_map = %p\n", ufi.orig_map); 757 printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr); 758 panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0"); 759 } 760 #endif 761 762 /* 763 * check protection 764 */ 765 766 check_prot = fault_flag == UVM_FAULT_WIREMAX ? 767 ufi.entry->max_protection : ufi.entry->protection; 768 if ((check_prot & access_type) != access_type) { 769 UVMHIST_LOG(maphist, 770 "<- protection failure (prot=0x%x, access=0x%x)", 771 ufi.entry->protection, access_type, 0, 0); 772 uvmfault_unlockmaps(&ufi, false); 773 error = EACCES; 774 goto done; 775 } 776 777 /* 778 * "enter_prot" is the protection we want to enter the page in at. 779 * for certain pages (e.g. copy-on-write pages) this protection can 780 * be more strict than ufi.entry->protection. "wired" means either 781 * the entry is wired or we are fault-wiring the pg. 782 */ 783 784 enter_prot = ufi.entry->protection; 785 wired = VM_MAPENT_ISWIRED(ufi.entry) || wire_fault; 786 if (wired) { 787 access_type = enter_prot; /* full access for wired */ 788 cow_now = (check_prot & VM_PROT_WRITE) != 0; 789 } else { 790 cow_now = (access_type & VM_PROT_WRITE) != 0; 791 } 792 793 /* 794 * handle "needs_copy" case. if we need to copy the amap we will 795 * have to drop our readlock and relock it with a write lock. (we 796 * need a write lock to change anything in a map entry [e.g. 797 * needs_copy]). 798 */ 799 800 if (UVM_ET_ISNEEDSCOPY(ufi.entry)) { 801 if (cow_now || (ufi.entry->object.uvm_obj == NULL)) { 802 KASSERT(fault_flag != UVM_FAULT_WIREMAX); 803 /* need to clear */ 804 UVMHIST_LOG(maphist, 805 " need to clear needs_copy and refault",0,0,0,0); 806 uvmfault_unlockmaps(&ufi, false); 807 uvmfault_amapcopy(&ufi); 808 uvmexp.fltamcopy++; 809 goto ReFault; 810 811 } else { 812 813 /* 814 * ensure that we pmap_enter page R/O since 815 * needs_copy is still true 816 */ 817 818 enter_prot &= ~VM_PROT_WRITE; 819 } 820 } 821 822 /* 823 * identify the players 824 */ 825 826 amap = ufi.entry->aref.ar_amap; /* top layer */ 827 uobj = ufi.entry->object.uvm_obj; /* bottom layer */ 828 829 /* 830 * check for a case 0 fault. if nothing backing the entry then 831 * error now. 832 */ 833 834 if (amap == NULL && uobj == NULL) { 835 uvmfault_unlockmaps(&ufi, false); 836 UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0); 837 error = EFAULT; 838 goto done; 839 } 840 841 /* 842 * establish range of interest based on advice from mapper 843 * and then clip to fit map entry. note that we only want 844 * to do this the first time through the fault. if we 845 * ReFault we will disable this by setting "narrow" to true. 846 */ 847 848 if (narrow == false) { 849 850 /* wide fault (!narrow) */ 851 KASSERT(uvmadvice[ufi.entry->advice].advice == 852 ufi.entry->advice); 853 nback = MIN(uvmadvice[ufi.entry->advice].nback, 854 (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT); 855 startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT); 856 nforw = MIN(uvmadvice[ufi.entry->advice].nforw, 857 ((ufi.entry->end - ufi.orig_rvaddr) >> 858 PAGE_SHIFT) - 1); 859 /* 860 * note: "-1" because we don't want to count the 861 * faulting page as forw 862 */ 863 npages = nback + nforw + 1; 864 centeridx = nback; 865 866 narrow = true; /* ensure only once per-fault */ 867 868 } else { 869 870 /* narrow fault! */ 871 nback = nforw = 0; 872 startva = ufi.orig_rvaddr; 873 npages = 1; 874 centeridx = 0; 875 876 } 877 878 /* locked: maps(read) */ 879 UVMHIST_LOG(maphist, " narrow=%d, back=%d, forw=%d, startva=0x%x", 880 narrow, nback, nforw, startva); 881 UVMHIST_LOG(maphist, " entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry, 882 amap, uobj, 0); 883 884 /* 885 * if we've got an amap, lock it and extract current anons. 886 */ 887 888 if (amap) { 889 amap_lock(amap); 890 anons = anons_store; 891 amap_lookups(&ufi.entry->aref, startva - ufi.entry->start, 892 anons, npages); 893 } else { 894 anons = NULL; /* to be safe */ 895 } 896 897 /* locked: maps(read), amap(if there) */ 898 KASSERT(amap == NULL || mutex_owned(&amap->am_l)); 899 900 /* 901 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages 902 * now and then forget about them (for the rest of the fault). 903 */ 904 905 if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) { 906 907 UVMHIST_LOG(maphist, " MADV_SEQUENTIAL: flushing backpages", 908 0,0,0,0); 909 /* flush back-page anons? */ 910 if (amap) 911 uvmfault_anonflush(anons, nback); 912 913 /* flush object? */ 914 if (uobj) { 915 uoff = (startva - ufi.entry->start) + ufi.entry->offset; 916 mutex_enter(&uobj->vmobjlock); 917 (void) (uobj->pgops->pgo_put)(uobj, uoff, uoff + 918 (nback << PAGE_SHIFT), PGO_DEACTIVATE); 919 } 920 921 /* now forget about the backpages */ 922 if (amap) 923 anons += nback; 924 startva += (nback << PAGE_SHIFT); 925 npages -= nback; 926 nback = centeridx = 0; 927 } 928 929 /* locked: maps(read), amap(if there) */ 930 KASSERT(amap == NULL || mutex_owned(&amap->am_l)); 931 932 /* 933 * map in the backpages and frontpages we found in the amap in hopes 934 * of preventing future faults. we also init the pages[] array as 935 * we go. 936 */ 937 938 currva = startva; 939 shadowed = false; 940 for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) { 941 942 /* 943 * dont play with VAs that are already mapped 944 * except for center) 945 */ 946 if (lcv != centeridx && 947 pmap_extract(ufi.orig_map->pmap, currva, NULL)) { 948 pages[lcv] = PGO_DONTCARE; 949 continue; 950 } 951 952 /* 953 * unmapped or center page. check if any anon at this level. 954 */ 955 if (amap == NULL || anons[lcv] == NULL) { 956 pages[lcv] = NULL; 957 continue; 958 } 959 960 /* 961 * check for present page and map if possible. re-activate it. 962 */ 963 964 pages[lcv] = PGO_DONTCARE; 965 if (lcv == centeridx) { /* save center for later! */ 966 shadowed = true; 967 continue; 968 } 969 anon = anons[lcv]; 970 mutex_enter(&anon->an_lock); 971 /* ignore loaned pages */ 972 if (anon->an_page && anon->an_page->loan_count == 0 && 973 (anon->an_page->flags & PG_BUSY) == 0) { 974 mutex_enter(&uvm_pageqlock); 975 uvm_pageenqueue(anon->an_page); 976 mutex_exit(&uvm_pageqlock); 977 UVMHIST_LOG(maphist, 978 " MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x", 979 ufi.orig_map->pmap, currva, anon->an_page, 0); 980 uvmexp.fltnamap++; 981 982 /* 983 * Since this isn't the page that's actually faulting, 984 * ignore pmap_enter() failures; it's not critical 985 * that we enter these right now. 986 */ 987 988 (void) pmap_enter(ufi.orig_map->pmap, currva, 989 VM_PAGE_TO_PHYS(anon->an_page), 990 (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) : 991 enter_prot, 992 PMAP_CANFAIL | 993 (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0)); 994 } 995 mutex_exit(&anon->an_lock); 996 pmap_update(ufi.orig_map->pmap); 997 } 998 999 /* locked: maps(read), amap(if there) */ 1000 KASSERT(amap == NULL || mutex_owned(&amap->am_l)); 1001 /* (shadowed == true) if there is an anon at the faulting address */ 1002 UVMHIST_LOG(maphist, " shadowed=%d, will_get=%d", shadowed, 1003 (uobj && shadowed == false),0,0); 1004 1005 /* 1006 * note that if we are really short of RAM we could sleep in the above 1007 * call to pmap_enter with everything locked. bad? 1008 * 1009 * XXX Actually, that is bad; pmap_enter() should just fail in that 1010 * XXX case. --thorpej 1011 */ 1012 1013 /* 1014 * if the desired page is not shadowed by the amap and we have a 1015 * backing object, then we check to see if the backing object would 1016 * prefer to handle the fault itself (rather than letting us do it 1017 * with the usual pgo_get hook). the backing object signals this by 1018 * providing a pgo_fault routine. 1019 */ 1020 1021 if (uobj && shadowed == false && uobj->pgops->pgo_fault != NULL) { 1022 mutex_enter(&uobj->vmobjlock); 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 mutex_enter(&uobj->vmobjlock); 1047 /* locked (!shadowed): maps(read), amap (if there), uobj */ 1048 /* 1049 * the following call to pgo_get does _not_ change locking state 1050 */ 1051 1052 uvmexp.fltlget++; 1053 gotpages = npages; 1054 (void) uobj->pgops->pgo_get(uobj, ufi.entry->offset + 1055 (startva - ufi.entry->start), 1056 pages, &gotpages, centeridx, 1057 access_type & MASK(ufi.entry), 1058 ufi.entry->advice, PGO_LOCKED); 1059 1060 /* 1061 * check for pages to map, if we got any 1062 */ 1063 1064 uobjpage = NULL; 1065 1066 if (gotpages) { 1067 currva = startva; 1068 for (lcv = 0; lcv < npages; 1069 lcv++, currva += PAGE_SIZE) { 1070 struct vm_page *curpg; 1071 bool readonly; 1072 1073 curpg = pages[lcv]; 1074 if (curpg == NULL || curpg == PGO_DONTCARE) { 1075 continue; 1076 } 1077 KASSERT(curpg->uobject == uobj); 1078 1079 /* 1080 * if center page is resident and not 1081 * PG_BUSY|PG_RELEASED then pgo_get 1082 * made it PG_BUSY for us and gave 1083 * us a handle to it. remember this 1084 * page as "uobjpage." (for later use). 1085 */ 1086 1087 if (lcv == centeridx) { 1088 uobjpage = curpg; 1089 UVMHIST_LOG(maphist, " got uobjpage " 1090 "(0x%x) with locked get", 1091 uobjpage, 0,0,0); 1092 continue; 1093 } 1094 1095 /* 1096 * calling pgo_get with PGO_LOCKED returns us 1097 * pages which are neither busy nor released, 1098 * so we don't need to check for this. 1099 * we can just directly enter the pages. 1100 */ 1101 1102 mutex_enter(&uvm_pageqlock); 1103 uvm_pageenqueue(curpg); 1104 mutex_exit(&uvm_pageqlock); 1105 UVMHIST_LOG(maphist, 1106 " MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x", 1107 ufi.orig_map->pmap, currva, curpg, 0); 1108 uvmexp.fltnomap++; 1109 1110 /* 1111 * Since this page isn't the page that's 1112 * actually faulting, ignore pmap_enter() 1113 * failures; it's not critical that we 1114 * enter these right now. 1115 */ 1116 KASSERT((curpg->flags & PG_PAGEOUT) == 0); 1117 KASSERT((curpg->flags & PG_RELEASED) == 0); 1118 KASSERT(!UVM_OBJ_IS_CLEAN(curpg->uobject) || 1119 (curpg->flags & PG_CLEAN) != 0); 1120 readonly = (curpg->flags & PG_RDONLY) 1121 || (curpg->loan_count > 0) 1122 || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject); 1123 1124 (void) pmap_enter(ufi.orig_map->pmap, currva, 1125 VM_PAGE_TO_PHYS(curpg), 1126 readonly ? 1127 enter_prot & ~VM_PROT_WRITE : 1128 enter_prot & MASK(ufi.entry), 1129 PMAP_CANFAIL | 1130 (wired ? PMAP_WIRED : 0)); 1131 1132 /* 1133 * NOTE: page can't be PG_WANTED or PG_RELEASED 1134 * because we've held the lock the whole time 1135 * we've had the handle. 1136 */ 1137 KASSERT((curpg->flags & PG_WANTED) == 0); 1138 KASSERT((curpg->flags & PG_RELEASED) == 0); 1139 1140 curpg->flags &= ~(PG_BUSY); 1141 UVM_PAGE_OWN(curpg, NULL); 1142 } 1143 pmap_update(ufi.orig_map->pmap); 1144 } 1145 } else { 1146 uobjpage = NULL; 1147 } 1148 1149 /* locked (shadowed): maps(read), amap */ 1150 /* locked (!shadowed): maps(read), amap(if there), 1151 uobj(if !null), uobjpage(if !null) */ 1152 if (shadowed) { 1153 KASSERT(mutex_owned(&amap->am_l)); 1154 } else { 1155 KASSERT(amap == NULL || mutex_owned(&amap->am_l)); 1156 KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock)); 1157 KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0); 1158 } 1159 1160 /* 1161 * note that at this point we are done with any front or back pages. 1162 * we are now going to focus on the center page (i.e. the one we've 1163 * faulted on). if we have faulted on the top (anon) layer 1164 * [i.e. case 1], then the anon we want is anons[centeridx] (we have 1165 * not touched it yet). if we have faulted on the bottom (uobj) 1166 * layer [i.e. case 2] and the page was both present and available, 1167 * then we've got a pointer to it as "uobjpage" and we've already 1168 * made it BUSY. 1169 */ 1170 1171 /* 1172 * there are four possible cases we must address: 1A, 1B, 2A, and 2B 1173 */ 1174 1175 /* 1176 * redirect case 2: if we are not shadowed, go to case 2. 1177 */ 1178 1179 if (shadowed == false) 1180 goto Case2; 1181 1182 /* locked: maps(read), amap */ 1183 1184 /* 1185 * handle case 1: fault on an anon in our amap 1186 */ 1187 1188 anon = anons[centeridx]; 1189 UVMHIST_LOG(maphist, " case 1 fault: anon=0x%x", anon, 0,0,0); 1190 mutex_enter(&anon->an_lock); 1191 1192 /* locked: maps(read), amap, anon */ 1193 KASSERT(mutex_owned(&amap->am_l)); 1194 KASSERT(mutex_owned(&anon->an_lock)); 1195 1196 /* 1197 * no matter if we have case 1A or case 1B we are going to need to 1198 * have the anon's memory resident. ensure that now. 1199 */ 1200 1201 /* 1202 * let uvmfault_anonget do the dirty work. 1203 * if it fails (!OK) it will unlock everything for us. 1204 * if it succeeds, locks are still valid and locked. 1205 * also, if it is OK, then the anon's page is on the queues. 1206 * if the page is on loan from a uvm_object, then anonget will 1207 * lock that object for us if it does not fail. 1208 */ 1209 1210 error = uvmfault_anonget(&ufi, amap, anon); 1211 switch (error) { 1212 case 0: 1213 break; 1214 1215 case ERESTART: 1216 goto ReFault; 1217 1218 case EAGAIN: 1219 tsleep(&lbolt, PVM, "fltagain1", 0); 1220 goto ReFault; 1221 1222 default: 1223 goto done; 1224 } 1225 1226 /* 1227 * uobj is non null if the page is on loan from an object (i.e. uobj) 1228 */ 1229 1230 uobj = anon->an_page->uobject; /* locked by anonget if !NULL */ 1231 1232 /* locked: maps(read), amap, anon, uobj(if one) */ 1233 KASSERT(mutex_owned(&amap->am_l)); 1234 KASSERT(mutex_owned(&anon->an_lock)); 1235 KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock)); 1236 1237 /* 1238 * special handling for loaned pages 1239 */ 1240 1241 if (anon->an_page->loan_count) { 1242 1243 if (!cow_now) { 1244 1245 /* 1246 * for read faults on loaned pages we just cap the 1247 * protection at read-only. 1248 */ 1249 1250 enter_prot = enter_prot & ~VM_PROT_WRITE; 1251 1252 } else { 1253 /* 1254 * note that we can't allow writes into a loaned page! 1255 * 1256 * if we have a write fault on a loaned page in an 1257 * anon then we need to look at the anon's ref count. 1258 * if it is greater than one then we are going to do 1259 * a normal copy-on-write fault into a new anon (this 1260 * is not a problem). however, if the reference count 1261 * is one (a case where we would normally allow a 1262 * write directly to the page) then we need to kill 1263 * the loan before we continue. 1264 */ 1265 1266 /* >1 case is already ok */ 1267 if (anon->an_ref == 1) { 1268 1269 /* get new un-owned replacement page */ 1270 pg = uvm_pagealloc(NULL, 0, NULL, 0); 1271 if (pg == NULL) { 1272 uvmfault_unlockall(&ufi, amap, uobj, 1273 anon); 1274 uvm_wait("flt_noram2"); 1275 goto ReFault; 1276 } 1277 1278 /* 1279 * copy data, kill loan, and drop uobj lock 1280 * (if any) 1281 */ 1282 /* copy old -> new */ 1283 uvm_pagecopy(anon->an_page, pg); 1284 1285 /* force reload */ 1286 pmap_page_protect(anon->an_page, VM_PROT_NONE); 1287 mutex_enter(&uvm_pageqlock); /* KILL loan */ 1288 1289 anon->an_page->uanon = NULL; 1290 /* in case we owned */ 1291 anon->an_page->pqflags &= ~PQ_ANON; 1292 1293 if (uobj) { 1294 /* if we were receiver of loan */ 1295 anon->an_page->loan_count--; 1296 } else { 1297 /* 1298 * we were the lender (A->K); need 1299 * to remove the page from pageq's. 1300 */ 1301 uvm_pagedequeue(anon->an_page); 1302 } 1303 1304 if (uobj) { 1305 mutex_exit(&uobj->vmobjlock); 1306 uobj = NULL; 1307 } 1308 1309 /* install new page in anon */ 1310 anon->an_page = pg; 1311 pg->uanon = anon; 1312 pg->pqflags |= PQ_ANON; 1313 1314 uvm_pageactivate(pg); 1315 mutex_exit(&uvm_pageqlock); 1316 1317 pg->flags &= ~(PG_BUSY|PG_FAKE); 1318 UVM_PAGE_OWN(pg, NULL); 1319 1320 /* done! */ 1321 } /* ref == 1 */ 1322 } /* write fault */ 1323 } /* loan count */ 1324 1325 /* 1326 * if we are case 1B then we will need to allocate a new blank 1327 * anon to transfer the data into. note that we have a lock 1328 * on anon, so no one can busy or release the page until we are done. 1329 * also note that the ref count can't drop to zero here because 1330 * it is > 1 and we are only dropping one ref. 1331 * 1332 * in the (hopefully very rare) case that we are out of RAM we 1333 * will unlock, wait for more RAM, and refault. 1334 * 1335 * if we are out of anon VM we kill the process (XXX: could wait?). 1336 */ 1337 1338 if (cow_now && anon->an_ref > 1) { 1339 1340 UVMHIST_LOG(maphist, " case 1B: COW fault",0,0,0,0); 1341 uvmexp.flt_acow++; 1342 oanon = anon; /* oanon = old, locked anon */ 1343 1344 error = uvmfault_promote(&ufi, oanon, PGO_DONTCARE, 1345 &anon, &anon_spare); 1346 switch (error) { 1347 case 0: 1348 break; 1349 case ERESTART: 1350 goto ReFault; 1351 default: 1352 goto done; 1353 } 1354 1355 pg = anon->an_page; 1356 mutex_enter(&uvm_pageqlock); 1357 uvm_pageactivate(pg); 1358 mutex_exit(&uvm_pageqlock); 1359 pg->flags &= ~(PG_BUSY|PG_FAKE); 1360 UVM_PAGE_OWN(pg, NULL); 1361 1362 /* deref: can not drop to zero here by defn! */ 1363 oanon->an_ref--; 1364 1365 /* 1366 * note: oanon is still locked, as is the new anon. we 1367 * need to check for this later when we unlock oanon; if 1368 * oanon != anon, we'll have to unlock anon, too. 1369 */ 1370 1371 } else { 1372 1373 uvmexp.flt_anon++; 1374 oanon = anon; /* old, locked anon is same as anon */ 1375 pg = anon->an_page; 1376 if (anon->an_ref > 1) /* disallow writes to ref > 1 anons */ 1377 enter_prot = enter_prot & ~VM_PROT_WRITE; 1378 1379 } 1380 1381 /* locked: maps(read), amap, oanon, anon (if different from oanon) */ 1382 KASSERT(mutex_owned(&amap->am_l)); 1383 KASSERT(mutex_owned(&anon->an_lock)); 1384 KASSERT(mutex_owned(&oanon->an_lock)); 1385 1386 /* 1387 * now map the page in. 1388 */ 1389 1390 UVMHIST_LOG(maphist, " MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x", 1391 ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0); 1392 if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg), 1393 enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) 1394 != 0) { 1395 1396 /* 1397 * No need to undo what we did; we can simply think of 1398 * this as the pmap throwing away the mapping information. 1399 * 1400 * We do, however, have to go through the ReFault path, 1401 * as the map may change while we're asleep. 1402 */ 1403 1404 if (anon != oanon) 1405 mutex_exit(&anon->an_lock); 1406 uvmfault_unlockall(&ufi, amap, uobj, oanon); 1407 if (!uvm_reclaimable()) { 1408 UVMHIST_LOG(maphist, 1409 "<- failed. out of VM",0,0,0,0); 1410 /* XXX instrumentation */ 1411 error = ENOMEM; 1412 goto done; 1413 } 1414 /* XXX instrumentation */ 1415 uvm_wait("flt_pmfail1"); 1416 goto ReFault; 1417 } 1418 1419 /* 1420 * ... update the page queues. 1421 */ 1422 1423 mutex_enter(&uvm_pageqlock); 1424 if (wire_fault) { 1425 uvm_pagewire(pg); 1426 1427 /* 1428 * since the now-wired page cannot be paged out, 1429 * release its swap resources for others to use. 1430 * since an anon with no swap cannot be PG_CLEAN, 1431 * clear its clean flag now. 1432 */ 1433 1434 pg->flags &= ~(PG_CLEAN); 1435 uvm_anon_dropswap(anon); 1436 } else { 1437 uvm_pageactivate(pg); 1438 } 1439 mutex_exit(&uvm_pageqlock); 1440 1441 /* 1442 * done case 1! finish up by unlocking everything and returning success 1443 */ 1444 1445 if (anon != oanon) 1446 mutex_exit(&anon->an_lock); 1447 uvmfault_unlockall(&ufi, amap, uobj, oanon); 1448 pmap_update(ufi.orig_map->pmap); 1449 error = 0; 1450 goto done; 1451 1452 Case2: 1453 /* 1454 * handle case 2: faulting on backing object or zero fill 1455 */ 1456 1457 /* 1458 * locked: 1459 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null) 1460 */ 1461 KASSERT(amap == NULL || mutex_owned(&amap->am_l)); 1462 KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock)); 1463 KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0); 1464 1465 /* 1466 * note that uobjpage can not be PGO_DONTCARE at this point. we now 1467 * set uobjpage to PGO_DONTCARE if we are doing a zero fill. if we 1468 * have a backing object, check and see if we are going to promote 1469 * the data up to an anon during the fault. 1470 */ 1471 1472 if (uobj == NULL) { 1473 uobjpage = PGO_DONTCARE; 1474 promote = true; /* always need anon here */ 1475 } else { 1476 KASSERT(uobjpage != PGO_DONTCARE); 1477 promote = cow_now && UVM_ET_ISCOPYONWRITE(ufi.entry); 1478 } 1479 UVMHIST_LOG(maphist, " case 2 fault: promote=%d, zfill=%d", 1480 promote, (uobj == NULL), 0,0); 1481 1482 /* 1483 * if uobjpage is not null then we do not need to do I/O to get the 1484 * uobjpage. 1485 * 1486 * if uobjpage is null, then we need to unlock and ask the pager to 1487 * get the data for us. once we have the data, we need to reverify 1488 * the state the world. we are currently not holding any resources. 1489 */ 1490 1491 if (uobjpage) { 1492 /* update rusage counters */ 1493 curlwp->l_ru.ru_minflt++; 1494 } else { 1495 /* update rusage counters */ 1496 curlwp->l_ru.ru_majflt++; 1497 1498 /* locked: maps(read), amap(if there), uobj */ 1499 uvmfault_unlockall(&ufi, amap, NULL, NULL); 1500 /* locked: uobj */ 1501 1502 uvmexp.fltget++; 1503 gotpages = 1; 1504 uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset; 1505 error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages, 1506 0, access_type & MASK(ufi.entry), ufi.entry->advice, 1507 PGO_SYNCIO); 1508 /* locked: uobjpage(if no error) */ 1509 KASSERT(error != 0 || (uobjpage->flags & PG_BUSY) != 0); 1510 1511 /* 1512 * recover from I/O 1513 */ 1514 1515 if (error) { 1516 if (error == EAGAIN) { 1517 UVMHIST_LOG(maphist, 1518 " pgo_get says TRY AGAIN!",0,0,0,0); 1519 tsleep(&lbolt, PVM, "fltagain2", 0); 1520 goto ReFault; 1521 } 1522 1523 UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)", 1524 error, 0,0,0); 1525 goto done; 1526 } 1527 1528 /* locked: uobjpage */ 1529 1530 mutex_enter(&uvm_pageqlock); 1531 uvm_pageactivate(uobjpage); 1532 mutex_exit(&uvm_pageqlock); 1533 1534 /* 1535 * re-verify the state of the world by first trying to relock 1536 * the maps. always relock the object. 1537 */ 1538 1539 locked = uvmfault_relock(&ufi); 1540 if (locked && amap) 1541 amap_lock(amap); 1542 uobj = uobjpage->uobject; 1543 mutex_enter(&uobj->vmobjlock); 1544 1545 /* locked(locked): maps(read), amap(if !null), uobj, uobjpage */ 1546 /* locked(!locked): uobj, uobjpage */ 1547 1548 /* 1549 * verify that the page has not be released and re-verify 1550 * that amap slot is still free. if there is a problem, 1551 * we unlock and clean up. 1552 */ 1553 1554 if ((uobjpage->flags & PG_RELEASED) != 0 || 1555 (locked && amap && 1556 amap_lookup(&ufi.entry->aref, 1557 ufi.orig_rvaddr - ufi.entry->start))) { 1558 if (locked) 1559 uvmfault_unlockall(&ufi, amap, NULL, NULL); 1560 locked = false; 1561 } 1562 1563 /* 1564 * didn't get the lock? release the page and retry. 1565 */ 1566 1567 if (locked == false) { 1568 UVMHIST_LOG(maphist, 1569 " wasn't able to relock after fault: retry", 1570 0,0,0,0); 1571 if (uobjpage->flags & PG_WANTED) 1572 wakeup(uobjpage); 1573 if (uobjpage->flags & PG_RELEASED) { 1574 uvmexp.fltpgrele++; 1575 uvm_pagefree(uobjpage); 1576 goto ReFault; 1577 } 1578 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1579 UVM_PAGE_OWN(uobjpage, NULL); 1580 mutex_exit(&uobj->vmobjlock); 1581 goto ReFault; 1582 } 1583 1584 /* 1585 * we have the data in uobjpage which is busy and 1586 * not released. we are holding object lock (so the page 1587 * can't be released on us). 1588 */ 1589 1590 /* locked: maps(read), amap(if !null), uobj, uobjpage */ 1591 } 1592 1593 /* 1594 * locked: 1595 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj) 1596 */ 1597 KASSERT(amap == NULL || mutex_owned(&amap->am_l)); 1598 KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock)); 1599 KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0); 1600 1601 /* 1602 * notes: 1603 * - at this point uobjpage can not be NULL 1604 * - at this point uobjpage can not be PG_RELEASED (since we checked 1605 * for it above) 1606 * - at this point uobjpage could be PG_WANTED (handle later) 1607 */ 1608 1609 KASSERT(uobj == NULL || uobj == uobjpage->uobject); 1610 KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) || 1611 (uobjpage->flags & PG_CLEAN) != 0); 1612 if (promote == false) { 1613 1614 /* 1615 * we are not promoting. if the mapping is COW ensure that we 1616 * don't give more access than we should (e.g. when doing a read 1617 * fault on a COPYONWRITE mapping we want to map the COW page in 1618 * R/O even though the entry protection could be R/W). 1619 * 1620 * set "pg" to the page we want to map in (uobjpage, usually) 1621 */ 1622 1623 /* no anon in this case. */ 1624 anon = NULL; 1625 1626 uvmexp.flt_obj++; 1627 if (UVM_ET_ISCOPYONWRITE(ufi.entry) || 1628 UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject)) 1629 enter_prot &= ~VM_PROT_WRITE; 1630 pg = uobjpage; /* map in the actual object */ 1631 1632 KASSERT(uobjpage != PGO_DONTCARE); 1633 1634 /* 1635 * we are faulting directly on the page. be careful 1636 * about writing to loaned pages... 1637 */ 1638 1639 if (uobjpage->loan_count) { 1640 if (!cow_now) { 1641 /* read fault: cap the protection at readonly */ 1642 /* cap! */ 1643 enter_prot = enter_prot & ~VM_PROT_WRITE; 1644 } else { 1645 /* write fault: must break the loan here */ 1646 1647 pg = uvm_loanbreak(uobjpage); 1648 if (pg == NULL) { 1649 1650 /* 1651 * drop ownership of page, it can't 1652 * be released 1653 */ 1654 1655 if (uobjpage->flags & PG_WANTED) 1656 wakeup(uobjpage); 1657 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1658 UVM_PAGE_OWN(uobjpage, NULL); 1659 1660 uvmfault_unlockall(&ufi, amap, uobj, 1661 NULL); 1662 UVMHIST_LOG(maphist, 1663 " out of RAM breaking loan, waiting", 1664 0,0,0,0); 1665 uvmexp.fltnoram++; 1666 uvm_wait("flt_noram4"); 1667 goto ReFault; 1668 } 1669 uobjpage = pg; 1670 } 1671 } 1672 } else { 1673 1674 /* 1675 * if we are going to promote the data to an anon we 1676 * allocate a blank anon here and plug it into our amap. 1677 */ 1678 #if DIAGNOSTIC 1679 if (amap == NULL) 1680 panic("uvm_fault: want to promote data, but no anon"); 1681 #endif 1682 error = uvmfault_promote(&ufi, NULL, uobjpage, 1683 &anon, &anon_spare); 1684 switch (error) { 1685 case 0: 1686 break; 1687 case ERESTART: 1688 goto ReFault; 1689 default: 1690 goto done; 1691 } 1692 1693 pg = anon->an_page; 1694 1695 /* 1696 * fill in the data 1697 */ 1698 1699 if (uobjpage != PGO_DONTCARE) { 1700 uvmexp.flt_prcopy++; 1701 1702 /* 1703 * promote to shared amap? make sure all sharing 1704 * procs see it 1705 */ 1706 1707 if ((amap_flags(amap) & AMAP_SHARED) != 0) { 1708 pmap_page_protect(uobjpage, VM_PROT_NONE); 1709 /* 1710 * XXX: PAGE MIGHT BE WIRED! 1711 */ 1712 } 1713 1714 /* 1715 * dispose of uobjpage. it can't be PG_RELEASED 1716 * since we still hold the object lock. 1717 * drop handle to uobj as well. 1718 */ 1719 1720 if (uobjpage->flags & PG_WANTED) 1721 /* still have the obj lock */ 1722 wakeup(uobjpage); 1723 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 1724 UVM_PAGE_OWN(uobjpage, NULL); 1725 mutex_exit(&uobj->vmobjlock); 1726 uobj = NULL; 1727 1728 UVMHIST_LOG(maphist, 1729 " promote uobjpage 0x%x to anon/page 0x%x/0x%x", 1730 uobjpage, anon, pg, 0); 1731 1732 } else { 1733 uvmexp.flt_przero++; 1734 1735 /* 1736 * Page is zero'd and marked dirty by 1737 * uvmfault_promote(). 1738 */ 1739 1740 UVMHIST_LOG(maphist," zero fill anon/page 0x%x/0%x", 1741 anon, pg, 0, 0); 1742 } 1743 } 1744 1745 /* 1746 * locked: 1747 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj), 1748 * anon(if !null), pg(if anon) 1749 * 1750 * note: pg is either the uobjpage or the new page in the new anon 1751 */ 1752 KASSERT(amap == NULL || mutex_owned(&amap->am_l)); 1753 KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock)); 1754 KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0); 1755 KASSERT(anon == NULL || mutex_owned(&anon->an_lock)); 1756 KASSERT((pg->flags & PG_BUSY) != 0); 1757 1758 /* 1759 * all resources are present. we can now map it in and free our 1760 * resources. 1761 */ 1762 1763 UVMHIST_LOG(maphist, 1764 " MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d", 1765 ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote); 1766 KASSERT((access_type & VM_PROT_WRITE) == 0 || 1767 (pg->flags & PG_RDONLY) == 0); 1768 if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg), 1769 pg->flags & PG_RDONLY ? enter_prot & ~VM_PROT_WRITE : enter_prot, 1770 access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) { 1771 1772 /* 1773 * No need to undo what we did; we can simply think of 1774 * this as the pmap throwing away the mapping information. 1775 * 1776 * We do, however, have to go through the ReFault path, 1777 * as the map may change while we're asleep. 1778 */ 1779 1780 if (pg->flags & PG_WANTED) 1781 wakeup(pg); 1782 1783 /* 1784 * note that pg can't be PG_RELEASED since we did not drop 1785 * the object lock since the last time we checked. 1786 */ 1787 KASSERT((pg->flags & PG_RELEASED) == 0); 1788 1789 pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED); 1790 UVM_PAGE_OWN(pg, NULL); 1791 uvmfault_unlockall(&ufi, amap, uobj, anon); 1792 if (!uvm_reclaimable()) { 1793 UVMHIST_LOG(maphist, 1794 "<- failed. out of VM",0,0,0,0); 1795 /* XXX instrumentation */ 1796 error = ENOMEM; 1797 goto done; 1798 } 1799 /* XXX instrumentation */ 1800 uvm_wait("flt_pmfail2"); 1801 goto ReFault; 1802 } 1803 1804 mutex_enter(&uvm_pageqlock); 1805 if (wire_fault) { 1806 uvm_pagewire(pg); 1807 if (pg->pqflags & PQ_AOBJ) { 1808 1809 /* 1810 * since the now-wired page cannot be paged out, 1811 * release its swap resources for others to use. 1812 * since an aobj page with no swap cannot be PG_CLEAN, 1813 * clear its clean flag now. 1814 */ 1815 1816 KASSERT(uobj != NULL); 1817 pg->flags &= ~(PG_CLEAN); 1818 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 1819 } 1820 } else { 1821 uvm_pageactivate(pg); 1822 } 1823 mutex_exit(&uvm_pageqlock); 1824 if (pg->flags & PG_WANTED) 1825 wakeup(pg); 1826 1827 /* 1828 * note that pg can't be PG_RELEASED since we did not drop the object 1829 * lock since the last time we checked. 1830 */ 1831 KASSERT((pg->flags & PG_RELEASED) == 0); 1832 1833 pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED); 1834 UVM_PAGE_OWN(pg, NULL); 1835 uvmfault_unlockall(&ufi, amap, uobj, anon); 1836 pmap_update(ufi.orig_map->pmap); 1837 UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0); 1838 error = 0; 1839 done: 1840 if (anon_spare != NULL) { 1841 anon_spare->an_ref--; 1842 uvm_anfree(anon_spare); 1843 } 1844 return error; 1845 } 1846 1847 1848 /* 1849 * uvm_fault_wire: wire down a range of virtual addresses in a map. 1850 * 1851 * => map may be read-locked by caller, but MUST NOT be write-locked. 1852 * => if map is read-locked, any operations which may cause map to 1853 * be write-locked in uvm_fault() must be taken care of by 1854 * the caller. See uvm_map_pageable(). 1855 */ 1856 1857 int 1858 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end, 1859 vm_prot_t access_type, int wiremax) 1860 { 1861 vaddr_t va; 1862 int error; 1863 1864 /* 1865 * now fault it in a page at a time. if the fault fails then we have 1866 * to undo what we have done. note that in uvm_fault VM_PROT_NONE 1867 * is replaced with the max protection if fault_type is VM_FAULT_WIRE. 1868 */ 1869 1870 /* 1871 * XXX work around overflowing a vaddr_t. this prevents us from 1872 * wiring the last page in the address space, though. 1873 */ 1874 if (start > end) { 1875 return EFAULT; 1876 } 1877 1878 for (va = start ; va < end ; va += PAGE_SIZE) { 1879 error = uvm_fault_internal(map, va, access_type, 1880 wiremax ? UVM_FAULT_WIREMAX : UVM_FAULT_WIRE); 1881 if (error) { 1882 if (va != start) { 1883 uvm_fault_unwire(map, start, va); 1884 } 1885 return error; 1886 } 1887 } 1888 return 0; 1889 } 1890 1891 /* 1892 * uvm_fault_unwire(): unwire range of virtual space. 1893 */ 1894 1895 void 1896 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end) 1897 { 1898 vm_map_lock_read(map); 1899 uvm_fault_unwire_locked(map, start, end); 1900 vm_map_unlock_read(map); 1901 } 1902 1903 /* 1904 * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire(). 1905 * 1906 * => map must be at least read-locked. 1907 */ 1908 1909 void 1910 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end) 1911 { 1912 struct vm_map_entry *entry; 1913 pmap_t pmap = vm_map_pmap(map); 1914 vaddr_t va; 1915 paddr_t pa; 1916 struct vm_page *pg; 1917 1918 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 1919 1920 /* 1921 * we assume that the area we are unwiring has actually been wired 1922 * in the first place. this means that we should be able to extract 1923 * the PAs from the pmap. we also lock out the page daemon so that 1924 * we can call uvm_pageunwire. 1925 */ 1926 1927 mutex_enter(&uvm_pageqlock); 1928 1929 /* 1930 * find the beginning map entry for the region. 1931 */ 1932 1933 KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map)); 1934 if (uvm_map_lookup_entry(map, start, &entry) == false) 1935 panic("uvm_fault_unwire_locked: address not in map"); 1936 1937 for (va = start; va < end; va += PAGE_SIZE) { 1938 if (pmap_extract(pmap, va, &pa) == false) 1939 continue; 1940 1941 /* 1942 * find the map entry for the current address. 1943 */ 1944 1945 KASSERT(va >= entry->start); 1946 while (va >= entry->end) { 1947 KASSERT(entry->next != &map->header && 1948 entry->next->start <= entry->end); 1949 entry = entry->next; 1950 } 1951 1952 /* 1953 * if the entry is no longer wired, tell the pmap. 1954 */ 1955 1956 if (VM_MAPENT_ISWIRED(entry) == 0) 1957 pmap_unwire(pmap, va); 1958 1959 pg = PHYS_TO_VM_PAGE(pa); 1960 if (pg) 1961 uvm_pageunwire(pg); 1962 } 1963 1964 mutex_exit(&uvm_pageqlock); 1965 } 1966