1 /* $NetBSD: uvm_fault.c,v 1.196 2014/08/10 16:44:37 tls Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 19 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 21 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 22 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 23 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 24 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 25 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 * 27 * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp 28 */ 29 30 /* 31 * uvm_fault.c: fault handler 32 */ 33 34 #include <sys/cdefs.h> 35 __KERNEL_RCSID(0, "$NetBSD: uvm_fault.c,v 1.196 2014/08/10 16:44:37 tls Exp $"); 36 37 #include "opt_uvmhist.h" 38 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/kernel.h> 42 #include <sys/mman.h> 43 44 #include <uvm/uvm.h> 45 46 /* 47 * 48 * a word on page faults: 49 * 50 * types of page faults we handle: 51 * 52 * CASE 1: upper layer faults CASE 2: lower layer faults 53 * 54 * CASE 1A CASE 1B CASE 2A CASE 2B 55 * read/write1 write>1 read/write +-cow_write/zero 56 * | | | | 57 * +--|--+ +--|--+ +-----+ + | + | +-----+ 58 * amap | V | | ---------> new | | | | ^ | 59 * +-----+ +-----+ +-----+ + | + | +--|--+ 60 * | | | 61 * +-----+ +-----+ +--|--+ | +--|--+ 62 * uobj | d/c | | d/c | | V | +----+ | 63 * +-----+ +-----+ +-----+ +-----+ 64 * 65 * d/c = don't care 66 * 67 * case [0]: layerless fault 68 * no amap or uobj is present. this is an error. 69 * 70 * case [1]: upper layer fault [anon active] 71 * 1A: [read] or [write with anon->an_ref == 1] 72 * I/O takes place in upper level anon and uobj is not touched. 73 * 1B: [write with anon->an_ref > 1] 74 * new anon is alloc'd and data is copied off ["COW"] 75 * 76 * case [2]: lower layer fault [uobj] 77 * 2A: [read on non-NULL uobj] or [write to non-copy_on_write area] 78 * I/O takes place directly in object. 79 * 2B: [write to copy_on_write] or [read on NULL uobj] 80 * data is "promoted" from uobj to a new anon. 81 * if uobj is null, then we zero fill. 82 * 83 * we follow the standard UVM locking protocol ordering: 84 * 85 * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ) 86 * we hold a PG_BUSY page if we unlock for I/O 87 * 88 * 89 * the code is structured as follows: 90 * 91 * - init the "IN" params in the ufi structure 92 * ReFault: (ERESTART returned to the loop in uvm_fault_internal) 93 * - do lookups [locks maps], check protection, handle needs_copy 94 * - check for case 0 fault (error) 95 * - establish "range" of fault 96 * - if we have an amap lock it and extract the anons 97 * - if sequential advice deactivate pages behind us 98 * - at the same time check pmap for unmapped areas and anon for pages 99 * that we could map in (and do map it if found) 100 * - check object for resident pages that we could map in 101 * - if (case 2) goto Case2 102 * - >>> handle case 1 103 * - ensure source anon is resident in RAM 104 * - if case 1B alloc new anon and copy from source 105 * - map the correct page in 106 * Case2: 107 * - >>> handle case 2 108 * - ensure source page is resident (if uobj) 109 * - if case 2B alloc new anon and copy from source (could be zero 110 * fill if uobj == NULL) 111 * - map the correct page in 112 * - done! 113 * 114 * note on paging: 115 * if we have to do I/O we place a PG_BUSY page in the correct object, 116 * unlock everything, and do the I/O. when I/O is done we must reverify 117 * the state of the world before assuming that our data structures are 118 * valid. [because mappings could change while the map is unlocked] 119 * 120 * alternative 1: unbusy the page in question and restart the page fault 121 * from the top (ReFault). this is easy but does not take advantage 122 * of the information that we already have from our previous lookup, 123 * although it is possible that the "hints" in the vm_map will help here. 124 * 125 * alternative 2: the system already keeps track of a "version" number of 126 * a map. [i.e. every time you write-lock a map (e.g. to change a 127 * mapping) you bump the version number up by one...] so, we can save 128 * the version number of the map before we release the lock and start I/O. 129 * then when I/O is done we can relock and check the version numbers 130 * to see if anything changed. this might save us some over 1 because 131 * we don't have to unbusy the page and may be less compares(?). 132 * 133 * alternative 3: put in backpointers or a way to "hold" part of a map 134 * in place while I/O is in progress. this could be complex to 135 * implement (especially with structures like amap that can be referenced 136 * by multiple map entries, and figuring out what should wait could be 137 * complex as well...). 138 * 139 * we use alternative 2. given that we are multi-threaded now we may want 140 * to reconsider the choice. 141 */ 142 143 /* 144 * local data structures 145 */ 146 147 struct uvm_advice { 148 int advice; 149 int nback; 150 int nforw; 151 }; 152 153 /* 154 * page range array: 155 * note: index in array must match "advice" value 156 * XXX: borrowed numbers from freebsd. do they work well for us? 157 */ 158 159 static const struct uvm_advice uvmadvice[] = { 160 { UVM_ADV_NORMAL, 3, 4 }, 161 { UVM_ADV_RANDOM, 0, 0 }, 162 { UVM_ADV_SEQUENTIAL, 8, 7}, 163 }; 164 165 #define UVM_MAXRANGE 16 /* must be MAX() of nback+nforw+1 */ 166 167 /* 168 * private prototypes 169 */ 170 171 /* 172 * externs from other modules 173 */ 174 175 extern int start_init_exec; /* Is init_main() done / init running? */ 176 177 /* 178 * inline functions 179 */ 180 181 /* 182 * uvmfault_anonflush: try and deactivate pages in specified anons 183 * 184 * => does not have to deactivate page if it is busy 185 */ 186 187 static inline void 188 uvmfault_anonflush(struct vm_anon **anons, int n) 189 { 190 int lcv; 191 struct vm_page *pg; 192 193 for (lcv = 0; lcv < n; lcv++) { 194 if (anons[lcv] == NULL) 195 continue; 196 KASSERT(mutex_owned(anons[lcv]->an_lock)); 197 pg = anons[lcv]->an_page; 198 if (pg && (pg->flags & PG_BUSY) == 0) { 199 mutex_enter(&uvm_pageqlock); 200 if (pg->wire_count == 0) { 201 uvm_pagedeactivate(pg); 202 } 203 mutex_exit(&uvm_pageqlock); 204 } 205 } 206 } 207 208 /* 209 * normal functions 210 */ 211 212 /* 213 * uvmfault_amapcopy: clear "needs_copy" in a map. 214 * 215 * => called with VM data structures unlocked (usually, see below) 216 * => we get a write lock on the maps and clear needs_copy for a VA 217 * => if we are out of RAM we sleep (waiting for more) 218 */ 219 220 static void 221 uvmfault_amapcopy(struct uvm_faultinfo *ufi) 222 { 223 for (;;) { 224 225 /* 226 * no mapping? give up. 227 */ 228 229 if (uvmfault_lookup(ufi, true) == false) 230 return; 231 232 /* 233 * copy if needed. 234 */ 235 236 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) 237 amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT, 238 ufi->orig_rvaddr, ufi->orig_rvaddr + 1); 239 240 /* 241 * didn't work? must be out of RAM. unlock and sleep. 242 */ 243 244 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) { 245 uvmfault_unlockmaps(ufi, true); 246 uvm_wait("fltamapcopy"); 247 continue; 248 } 249 250 /* 251 * got it! unlock and return. 252 */ 253 254 uvmfault_unlockmaps(ufi, true); 255 return; 256 } 257 /*NOTREACHED*/ 258 } 259 260 /* 261 * uvmfault_anonget: get data in an anon into a non-busy, non-released 262 * page in that anon. 263 * 264 * => Map, amap and thus anon should be locked by caller. 265 * => If we fail, we unlock everything and error is returned. 266 * => If we are successful, return with everything still locked. 267 * => We do not move the page on the queues [gets moved later]. If we 268 * allocate a new page [we_own], it gets put on the queues. Either way, 269 * the result is that the page is on the queues at return time 270 * => For pages which are on loan from a uvm_object (and thus are not owned 271 * by the anon): if successful, return with the owning object locked. 272 * The caller must unlock this object when it unlocks everything else. 273 */ 274 275 int 276 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap, 277 struct vm_anon *anon) 278 { 279 struct vm_page *pg; 280 int error; 281 282 UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist); 283 KASSERT(mutex_owned(anon->an_lock)); 284 KASSERT(anon->an_lock == amap->am_lock); 285 286 /* Increment the counters.*/ 287 uvmexp.fltanget++; 288 if (anon->an_page) { 289 curlwp->l_ru.ru_minflt++; 290 } else { 291 curlwp->l_ru.ru_majflt++; 292 } 293 error = 0; 294 295 /* 296 * Loop until we get the anon data, or fail. 297 */ 298 299 for (;;) { 300 bool we_own, locked; 301 /* 302 * Note: 'we_own' will become true if we set PG_BUSY on a page. 303 */ 304 we_own = false; 305 pg = anon->an_page; 306 307 /* 308 * If there is a resident page and it is loaned, then anon 309 * may not own it. Call out to uvm_anon_lockloanpg() to 310 * identify and lock the real owner of the page. 311 */ 312 313 if (pg && pg->loan_count) 314 pg = uvm_anon_lockloanpg(anon); 315 316 /* 317 * Is page resident? Make sure it is not busy/released. 318 */ 319 320 if (pg) { 321 322 /* 323 * at this point, if the page has a uobject [meaning 324 * we have it on loan], then that uobject is locked 325 * by us! if the page is busy, we drop all the 326 * locks (including uobject) and try again. 327 */ 328 329 if ((pg->flags & PG_BUSY) == 0) { 330 UVMHIST_LOG(maphist, "<- OK",0,0,0,0); 331 return 0; 332 } 333 pg->flags |= PG_WANTED; 334 uvmexp.fltpgwait++; 335 336 /* 337 * The last unlock must be an atomic unlock and wait 338 * on the owner of page. 339 */ 340 341 if (pg->uobject) { 342 /* Owner of page is UVM object. */ 343 uvmfault_unlockall(ufi, amap, NULL); 344 UVMHIST_LOG(maphist, " unlock+wait on uobj",0, 345 0,0,0); 346 UVM_UNLOCK_AND_WAIT(pg, 347 pg->uobject->vmobjlock, 348 false, "anonget1", 0); 349 } else { 350 /* Owner of page is anon. */ 351 uvmfault_unlockall(ufi, NULL, NULL); 352 UVMHIST_LOG(maphist, " unlock+wait on anon",0, 353 0,0,0); 354 UVM_UNLOCK_AND_WAIT(pg, anon->an_lock, 355 false, "anonget2", 0); 356 } 357 } else { 358 #if defined(VMSWAP) 359 /* 360 * No page, therefore allocate one. 361 */ 362 363 pg = uvm_pagealloc(NULL, 364 ufi != NULL ? ufi->orig_rvaddr : 0, 365 anon, ufi != NULL ? UVM_FLAG_COLORMATCH : 0); 366 if (pg == NULL) { 367 /* Out of memory. Wait a little. */ 368 uvmfault_unlockall(ufi, amap, NULL); 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 /* PG_BUSY bit is set. */ 378 we_own = true; 379 uvmfault_unlockall(ufi, amap, NULL); 380 381 /* 382 * Pass a PG_BUSY+PG_FAKE+PG_CLEAN page into 383 * the uvm_swap_get() function with all data 384 * structures unlocked. Note that it is OK 385 * to read an_swslot here, because we hold 386 * 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 398 panic("%s: no page", __func__); 399 #endif /* defined(VMSWAP) */ 400 } 401 402 /* 403 * Re-lock the map and anon. 404 */ 405 406 locked = uvmfault_relock(ufi); 407 if (locked || we_own) { 408 mutex_enter(anon->an_lock); 409 } 410 411 /* 412 * If we own the page (i.e. we set PG_BUSY), then we need 413 * to clean up after the I/O. There are three cases to 414 * consider: 415 * 416 * 1) Page was released during I/O: free anon and ReFault. 417 * 2) I/O not OK. Free the page and cause the fault to fail. 418 * 3) I/O OK! Activate the page and sync with the non-we_own 419 * case (i.e. drop anon lock if not locked). 420 */ 421 422 if (we_own) { 423 #if defined(VMSWAP) 424 if (pg->flags & PG_WANTED) { 425 wakeup(pg); 426 } 427 if (error) { 428 429 /* 430 * Remove the swap slot from the anon and 431 * mark the anon as having no real slot. 432 * Do not free the swap slot, thus preventing 433 * it from being used again. 434 */ 435 436 if (anon->an_swslot > 0) { 437 uvm_swap_markbad(anon->an_swslot, 1); 438 } 439 anon->an_swslot = SWSLOT_BAD; 440 441 if ((pg->flags & PG_RELEASED) != 0) { 442 goto released; 443 } 444 445 /* 446 * Note: page was never !PG_BUSY, so it 447 * cannot 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, NULL, NULL); 457 } 458 mutex_exit(anon->an_lock); 459 UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0); 460 return error; 461 } 462 463 if ((pg->flags & PG_RELEASED) != 0) { 464 released: 465 KASSERT(anon->an_ref == 0); 466 467 /* 468 * Released while we had unlocked amap. 469 */ 470 471 if (locked) { 472 uvmfault_unlockall(ufi, NULL, NULL); 473 } 474 uvm_anon_release(anon); 475 476 if (error) { 477 UVMHIST_LOG(maphist, 478 "<- ERROR/RELEASED", 0,0,0,0); 479 return error; 480 } 481 482 UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0); 483 return ERESTART; 484 } 485 486 /* 487 * We have successfully read the page, activate it. 488 */ 489 490 mutex_enter(&uvm_pageqlock); 491 uvm_pageactivate(pg); 492 mutex_exit(&uvm_pageqlock); 493 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE); 494 UVM_PAGE_OWN(pg, NULL); 495 #else 496 panic("%s: we_own", __func__); 497 #endif /* defined(VMSWAP) */ 498 } 499 500 /* 501 * We were not able to re-lock the map - restart the fault. 502 */ 503 504 if (!locked) { 505 if (we_own) { 506 mutex_exit(anon->an_lock); 507 } 508 UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0); 509 return ERESTART; 510 } 511 512 /* 513 * Verify that no one has touched the amap and moved 514 * the anon on us. 515 */ 516 517 if (ufi != NULL && amap_lookup(&ufi->entry->aref, 518 ufi->orig_rvaddr - ufi->entry->start) != anon) { 519 520 uvmfault_unlockall(ufi, amap, NULL); 521 UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0); 522 return ERESTART; 523 } 524 525 /* 526 * Retry.. 527 */ 528 529 uvmexp.fltanretry++; 530 continue; 531 } 532 /*NOTREACHED*/ 533 } 534 535 /* 536 * uvmfault_promote: promote data to a new anon. used for 1B and 2B. 537 * 538 * 1. allocate an anon and a page. 539 * 2. fill its contents. 540 * 3. put it into amap. 541 * 542 * => if we fail (result != 0) we unlock everything. 543 * => on success, return a new locked anon via 'nanon'. 544 * (*nanon)->an_page will be a resident, locked, dirty page. 545 * => it's caller's responsibility to put the promoted nanon->an_page to the 546 * page queue. 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 KASSERT(mutex_owned(amap->am_lock)); 586 KASSERT(oanon == NULL || amap->am_lock == oanon->an_lock); 587 KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock)); 588 589 if (*spare != NULL) { 590 anon = *spare; 591 *spare = NULL; 592 } else { 593 anon = uvm_analloc(); 594 } 595 if (anon) { 596 597 /* 598 * The new anon is locked. 599 * 600 * if opg == NULL, we want a zero'd, dirty page, 601 * so have uvm_pagealloc() do that for us. 602 */ 603 604 KASSERT(anon->an_lock == NULL); 605 anon->an_lock = amap->am_lock; 606 pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon, 607 UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0)); 608 if (pg == NULL) { 609 anon->an_lock = NULL; 610 } 611 } else { 612 pg = NULL; 613 } 614 615 /* 616 * out of memory resources? 617 */ 618 619 if (pg == NULL) { 620 /* save anon for the next try. */ 621 if (anon != NULL) { 622 *spare = anon; 623 } 624 625 /* unlock and fail ... */ 626 uvm_page_unbusy(&uobjpage, 1); 627 uvmfault_unlockall(ufi, amap, uobj); 628 if (!uvm_reclaimable()) { 629 UVMHIST_LOG(maphist, "out of VM", 0,0,0,0); 630 uvmexp.fltnoanon++; 631 error = ENOMEM; 632 goto done; 633 } 634 635 UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0); 636 uvmexp.fltnoram++; 637 uvm_wait("flt_noram5"); 638 error = ERESTART; 639 goto done; 640 } 641 642 /* copy page [pg now dirty] */ 643 if (opg) { 644 uvm_pagecopy(opg, pg); 645 } 646 647 amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon, 648 oanon != NULL); 649 650 *nanon = anon; 651 error = 0; 652 done: 653 return error; 654 } 655 656 657 /* 658 * F A U L T - m a i n e n t r y p o i n t 659 */ 660 661 /* 662 * uvm_fault: page fault handler 663 * 664 * => called from MD code to resolve a page fault 665 * => VM data structures usually should be unlocked. however, it is 666 * possible to call here with the main map locked if the caller 667 * gets a write lock, sets it recusive, and then calls us (c.f. 668 * uvm_map_pageable). this should be avoided because it keeps 669 * the map locked off during I/O. 670 * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT 671 */ 672 673 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 674 ~VM_PROT_WRITE : VM_PROT_ALL) 675 676 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */ 677 #define UVM_FAULT_WIRE (1 << 0) 678 #define UVM_FAULT_MAXPROT (1 << 1) 679 680 struct uvm_faultctx { 681 682 /* 683 * the following members are set up by uvm_fault_check() and 684 * read-only after that. 685 * 686 * note that narrow is used by uvm_fault_check() to change 687 * the behaviour after ERESTART. 688 * 689 * most of them might change after RESTART if the underlying 690 * map entry has been changed behind us. an exception is 691 * wire_paging, which does never change. 692 */ 693 vm_prot_t access_type; 694 vaddr_t startva; 695 int npages; 696 int centeridx; 697 bool narrow; /* work on a single requested page only */ 698 bool wire_mapping; /* request a PMAP_WIRED mapping 699 (UVM_FAULT_WIRE or VM_MAPENT_ISWIRED) */ 700 bool wire_paging; /* request uvm_pagewire 701 (true for UVM_FAULT_WIRE) */ 702 bool cow_now; /* VM_PROT_WRITE is actually requested 703 (ie. should break COW and page loaning) */ 704 705 /* 706 * enter_prot is set up by uvm_fault_check() and clamped 707 * (ie. drop the VM_PROT_WRITE bit) in various places in case 708 * of !cow_now. 709 */ 710 vm_prot_t enter_prot; /* prot at which we want to enter pages in */ 711 712 /* 713 * the following member is for uvmfault_promote() and ERESTART. 714 */ 715 struct vm_anon *anon_spare; 716 717 /* 718 * the folloing is actually a uvm_fault_lower() internal. 719 * it's here merely for debugging. 720 * (or due to the mechanical separation of the function?) 721 */ 722 bool promote; 723 }; 724 725 static inline int uvm_fault_check( 726 struct uvm_faultinfo *, struct uvm_faultctx *, 727 struct vm_anon ***, bool); 728 729 static int uvm_fault_upper( 730 struct uvm_faultinfo *, struct uvm_faultctx *, 731 struct vm_anon **); 732 static inline int uvm_fault_upper_lookup( 733 struct uvm_faultinfo *, const struct uvm_faultctx *, 734 struct vm_anon **, struct vm_page **); 735 static inline void uvm_fault_upper_neighbor( 736 struct uvm_faultinfo *, const struct uvm_faultctx *, 737 vaddr_t, struct vm_page *, bool); 738 static inline int uvm_fault_upper_loan( 739 struct uvm_faultinfo *, struct uvm_faultctx *, 740 struct vm_anon *, struct uvm_object **); 741 static inline int uvm_fault_upper_promote( 742 struct uvm_faultinfo *, struct uvm_faultctx *, 743 struct uvm_object *, struct vm_anon *); 744 static inline int uvm_fault_upper_direct( 745 struct uvm_faultinfo *, struct uvm_faultctx *, 746 struct uvm_object *, struct vm_anon *); 747 static int uvm_fault_upper_enter( 748 struct uvm_faultinfo *, const struct uvm_faultctx *, 749 struct uvm_object *, struct vm_anon *, 750 struct vm_page *, struct vm_anon *); 751 static inline void uvm_fault_upper_done( 752 struct uvm_faultinfo *, const struct uvm_faultctx *, 753 struct vm_anon *, struct vm_page *); 754 755 static int uvm_fault_lower( 756 struct uvm_faultinfo *, struct uvm_faultctx *, 757 struct vm_page **); 758 static inline void uvm_fault_lower_lookup( 759 struct uvm_faultinfo *, const struct uvm_faultctx *, 760 struct vm_page **); 761 static inline void uvm_fault_lower_neighbor( 762 struct uvm_faultinfo *, const struct uvm_faultctx *, 763 vaddr_t, struct vm_page *, bool); 764 static inline int uvm_fault_lower_io( 765 struct uvm_faultinfo *, const struct uvm_faultctx *, 766 struct uvm_object **, struct vm_page **); 767 static inline int uvm_fault_lower_direct( 768 struct uvm_faultinfo *, struct uvm_faultctx *, 769 struct uvm_object *, struct vm_page *); 770 static inline int uvm_fault_lower_direct_loan( 771 struct uvm_faultinfo *, struct uvm_faultctx *, 772 struct uvm_object *, struct vm_page **, 773 struct vm_page **); 774 static inline int uvm_fault_lower_promote( 775 struct uvm_faultinfo *, struct uvm_faultctx *, 776 struct uvm_object *, struct vm_page *); 777 static int uvm_fault_lower_enter( 778 struct uvm_faultinfo *, const struct uvm_faultctx *, 779 struct uvm_object *, 780 struct vm_anon *, struct vm_page *); 781 static inline void uvm_fault_lower_done( 782 struct uvm_faultinfo *, const struct uvm_faultctx *, 783 struct uvm_object *, struct vm_page *); 784 785 int 786 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr, 787 vm_prot_t access_type, int fault_flag) 788 { 789 struct cpu_data *cd; 790 struct uvm_cpu *ucpu; 791 struct uvm_faultinfo ufi; 792 struct uvm_faultctx flt = { 793 .access_type = access_type, 794 795 /* don't look for neighborhood * pages on "wire" fault */ 796 .narrow = (fault_flag & UVM_FAULT_WIRE) != 0, 797 798 /* "wire" fault causes wiring of both mapping and paging */ 799 .wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0, 800 .wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0, 801 }; 802 const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0; 803 struct vm_anon *anons_store[UVM_MAXRANGE], **anons; 804 struct vm_page *pages_store[UVM_MAXRANGE], **pages; 805 int error; 806 807 UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist); 808 809 UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, at=%d, ff=%d)", 810 orig_map, vaddr, access_type, fault_flag); 811 812 cd = &(curcpu()->ci_data); 813 cd->cpu_nfault++; 814 ucpu = cd->cpu_uvm; 815 816 /* Don't flood RNG subsystem with samples. */ 817 if (cd->cpu_nfault % 503) 818 goto norng; 819 820 /* Don't count anything until user interaction is possible */ 821 if (__predict_true(start_init_exec)) { 822 kpreempt_disable(); 823 rnd_add_uint32(&ucpu->rs, 824 sizeof(vaddr_t) == sizeof(uint32_t) ? 825 (uint32_t)vaddr : sizeof(vaddr_t) == 826 sizeof(uint64_t) ? 827 (uint32_t)(vaddr & 0x00000000ffffffff) : 828 (uint32_t)(cd->cpu_nfault & 0x00000000ffffffff)); 829 kpreempt_enable(); 830 } 831 norng: 832 /* 833 * init the IN parameters in the ufi 834 */ 835 836 ufi.orig_map = orig_map; 837 ufi.orig_rvaddr = trunc_page(vaddr); 838 ufi.orig_size = PAGE_SIZE; /* can't get any smaller than this */ 839 840 error = ERESTART; 841 while (error == ERESTART) { /* ReFault: */ 842 anons = anons_store; 843 pages = pages_store; 844 845 error = uvm_fault_check(&ufi, &flt, &anons, maxprot); 846 if (error != 0) 847 continue; 848 849 error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages); 850 if (error != 0) 851 continue; 852 853 if (pages[flt.centeridx] == PGO_DONTCARE) 854 error = uvm_fault_upper(&ufi, &flt, anons); 855 else { 856 struct uvm_object * const uobj = 857 ufi.entry->object.uvm_obj; 858 859 if (uobj && uobj->pgops->pgo_fault != NULL) { 860 /* 861 * invoke "special" fault routine. 862 */ 863 mutex_enter(uobj->vmobjlock); 864 /* locked: maps(read), amap(if there), uobj */ 865 error = uobj->pgops->pgo_fault(&ufi, 866 flt.startva, pages, flt.npages, 867 flt.centeridx, flt.access_type, 868 PGO_LOCKED|PGO_SYNCIO); 869 870 /* 871 * locked: nothing, pgo_fault has unlocked 872 * everything 873 */ 874 875 /* 876 * object fault routine responsible for 877 * pmap_update(). 878 */ 879 } else { 880 error = uvm_fault_lower(&ufi, &flt, pages); 881 } 882 } 883 } 884 885 if (flt.anon_spare != NULL) { 886 flt.anon_spare->an_ref--; 887 KASSERT(flt.anon_spare->an_ref == 0); 888 KASSERT(flt.anon_spare->an_lock == NULL); 889 uvm_anon_free(flt.anon_spare); 890 } 891 return error; 892 } 893 894 /* 895 * uvm_fault_check: check prot, handle needs-copy, etc. 896 * 897 * 1. lookup entry. 898 * 2. check protection. 899 * 3. adjust fault condition (mainly for simulated fault). 900 * 4. handle needs-copy (lazy amap copy). 901 * 5. establish range of interest for neighbor fault (aka pre-fault). 902 * 6. look up anons (if amap exists). 903 * 7. flush pages (if MADV_SEQUENTIAL) 904 * 905 * => called with nothing locked. 906 * => if we fail (result != 0) we unlock everything. 907 * => initialize/adjust many members of flt. 908 */ 909 910 static int 911 uvm_fault_check( 912 struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 913 struct vm_anon ***ranons, bool maxprot) 914 { 915 struct vm_amap *amap; 916 struct uvm_object *uobj; 917 vm_prot_t check_prot; 918 int nback, nforw; 919 UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist); 920 921 /* 922 * lookup and lock the maps 923 */ 924 925 if (uvmfault_lookup(ufi, false) == false) { 926 UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", ufi->orig_rvaddr, 927 0,0,0); 928 return EFAULT; 929 } 930 /* locked: maps(read) */ 931 932 #ifdef DIAGNOSTIC 933 if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) { 934 printf("Page fault on non-pageable map:\n"); 935 printf("ufi->map = %p\n", ufi->map); 936 printf("ufi->orig_map = %p\n", ufi->orig_map); 937 printf("ufi->orig_rvaddr = 0x%lx\n", (u_long) ufi->orig_rvaddr); 938 panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0"); 939 } 940 #endif 941 942 /* 943 * check protection 944 */ 945 946 check_prot = maxprot ? 947 ufi->entry->max_protection : ufi->entry->protection; 948 if ((check_prot & flt->access_type) != flt->access_type) { 949 UVMHIST_LOG(maphist, 950 "<- protection failure (prot=0x%x, access=0x%x)", 951 ufi->entry->protection, flt->access_type, 0, 0); 952 uvmfault_unlockmaps(ufi, false); 953 return EACCES; 954 } 955 956 /* 957 * "enter_prot" is the protection we want to enter the page in at. 958 * for certain pages (e.g. copy-on-write pages) this protection can 959 * be more strict than ufi->entry->protection. "wired" means either 960 * the entry is wired or we are fault-wiring the pg. 961 */ 962 963 flt->enter_prot = ufi->entry->protection; 964 if (VM_MAPENT_ISWIRED(ufi->entry)) 965 flt->wire_mapping = true; 966 967 if (flt->wire_mapping) { 968 flt->access_type = flt->enter_prot; /* full access for wired */ 969 flt->cow_now = (check_prot & VM_PROT_WRITE) != 0; 970 } else { 971 flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0; 972 } 973 974 flt->promote = false; 975 976 /* 977 * handle "needs_copy" case. if we need to copy the amap we will 978 * have to drop our readlock and relock it with a write lock. (we 979 * need a write lock to change anything in a map entry [e.g. 980 * needs_copy]). 981 */ 982 983 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) { 984 if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) { 985 KASSERT(!maxprot); 986 /* need to clear */ 987 UVMHIST_LOG(maphist, 988 " need to clear needs_copy and refault",0,0,0,0); 989 uvmfault_unlockmaps(ufi, false); 990 uvmfault_amapcopy(ufi); 991 uvmexp.fltamcopy++; 992 return ERESTART; 993 994 } else { 995 996 /* 997 * ensure that we pmap_enter page R/O since 998 * needs_copy is still true 999 */ 1000 1001 flt->enter_prot &= ~VM_PROT_WRITE; 1002 } 1003 } 1004 1005 /* 1006 * identify the players 1007 */ 1008 1009 amap = ufi->entry->aref.ar_amap; /* upper layer */ 1010 uobj = ufi->entry->object.uvm_obj; /* lower layer */ 1011 1012 /* 1013 * check for a case 0 fault. if nothing backing the entry then 1014 * error now. 1015 */ 1016 1017 if (amap == NULL && uobj == NULL) { 1018 uvmfault_unlockmaps(ufi, false); 1019 UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0); 1020 return EFAULT; 1021 } 1022 1023 /* 1024 * establish range of interest based on advice from mapper 1025 * and then clip to fit map entry. note that we only want 1026 * to do this the first time through the fault. if we 1027 * ReFault we will disable this by setting "narrow" to true. 1028 */ 1029 1030 if (flt->narrow == false) { 1031 1032 /* wide fault (!narrow) */ 1033 KASSERT(uvmadvice[ufi->entry->advice].advice == 1034 ufi->entry->advice); 1035 nback = MIN(uvmadvice[ufi->entry->advice].nback, 1036 (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT); 1037 flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT); 1038 /* 1039 * note: "-1" because we don't want to count the 1040 * faulting page as forw 1041 */ 1042 nforw = MIN(uvmadvice[ufi->entry->advice].nforw, 1043 ((ufi->entry->end - ufi->orig_rvaddr) >> 1044 PAGE_SHIFT) - 1); 1045 flt->npages = nback + nforw + 1; 1046 flt->centeridx = nback; 1047 1048 flt->narrow = true; /* ensure only once per-fault */ 1049 1050 } else { 1051 1052 /* narrow fault! */ 1053 nback = nforw = 0; 1054 flt->startva = ufi->orig_rvaddr; 1055 flt->npages = 1; 1056 flt->centeridx = 0; 1057 1058 } 1059 /* offset from entry's start to pgs' start */ 1060 const voff_t eoff = flt->startva - ufi->entry->start; 1061 1062 /* locked: maps(read) */ 1063 UVMHIST_LOG(maphist, " narrow=%d, back=%d, forw=%d, startva=0x%x", 1064 flt->narrow, nback, nforw, flt->startva); 1065 UVMHIST_LOG(maphist, " entry=0x%x, amap=0x%x, obj=0x%x", ufi->entry, 1066 amap, uobj, 0); 1067 1068 /* 1069 * if we've got an amap, lock it and extract current anons. 1070 */ 1071 1072 if (amap) { 1073 amap_lock(amap); 1074 amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages); 1075 } else { 1076 *ranons = NULL; /* to be safe */ 1077 } 1078 1079 /* locked: maps(read), amap(if there) */ 1080 KASSERT(amap == NULL || mutex_owned(amap->am_lock)); 1081 1082 /* 1083 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages 1084 * now and then forget about them (for the rest of the fault). 1085 */ 1086 1087 if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) { 1088 1089 UVMHIST_LOG(maphist, " MADV_SEQUENTIAL: flushing backpages", 1090 0,0,0,0); 1091 /* flush back-page anons? */ 1092 if (amap) 1093 uvmfault_anonflush(*ranons, nback); 1094 1095 /* flush object? */ 1096 if (uobj) { 1097 voff_t uoff; 1098 1099 uoff = ufi->entry->offset + eoff; 1100 mutex_enter(uobj->vmobjlock); 1101 (void) (uobj->pgops->pgo_put)(uobj, uoff, uoff + 1102 (nback << PAGE_SHIFT), PGO_DEACTIVATE); 1103 } 1104 1105 /* now forget about the backpages */ 1106 if (amap) 1107 *ranons += nback; 1108 flt->startva += (nback << PAGE_SHIFT); 1109 flt->npages -= nback; 1110 flt->centeridx = 0; 1111 } 1112 /* 1113 * => startva is fixed 1114 * => npages is fixed 1115 */ 1116 KASSERT(flt->startva <= ufi->orig_rvaddr); 1117 KASSERT(ufi->orig_rvaddr + ufi->orig_size <= 1118 flt->startva + (flt->npages << PAGE_SHIFT)); 1119 return 0; 1120 } 1121 1122 /* 1123 * uvm_fault_upper_lookup: look up existing h/w mapping and amap. 1124 * 1125 * iterate range of interest: 1126 * 1. check if h/w mapping exists. if yes, we don't care 1127 * 2. check if anon exists. if not, page is lower. 1128 * 3. if anon exists, enter h/w mapping for neighbors. 1129 * 1130 * => called with amap locked (if exists). 1131 */ 1132 1133 static int 1134 uvm_fault_upper_lookup( 1135 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 1136 struct vm_anon **anons, struct vm_page **pages) 1137 { 1138 struct vm_amap *amap = ufi->entry->aref.ar_amap; 1139 int lcv; 1140 vaddr_t currva; 1141 bool shadowed __unused; 1142 UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist); 1143 1144 /* locked: maps(read), amap(if there) */ 1145 KASSERT(amap == NULL || mutex_owned(amap->am_lock)); 1146 1147 /* 1148 * map in the backpages and frontpages we found in the amap in hopes 1149 * of preventing future faults. we also init the pages[] array as 1150 * we go. 1151 */ 1152 1153 currva = flt->startva; 1154 shadowed = false; 1155 for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) { 1156 /* 1157 * don't play with VAs that are already mapped 1158 * (except for center) 1159 */ 1160 if (lcv != flt->centeridx && 1161 pmap_extract(ufi->orig_map->pmap, currva, NULL)) { 1162 pages[lcv] = PGO_DONTCARE; 1163 continue; 1164 } 1165 1166 /* 1167 * unmapped or center page. check if any anon at this level. 1168 */ 1169 if (amap == NULL || anons[lcv] == NULL) { 1170 pages[lcv] = NULL; 1171 continue; 1172 } 1173 1174 /* 1175 * check for present page and map if possible. re-activate it. 1176 */ 1177 1178 pages[lcv] = PGO_DONTCARE; 1179 if (lcv == flt->centeridx) { /* save center for later! */ 1180 shadowed = true; 1181 continue; 1182 } 1183 1184 struct vm_anon *anon = anons[lcv]; 1185 struct vm_page *pg = anon->an_page; 1186 1187 KASSERT(anon->an_lock == amap->am_lock); 1188 1189 /* Ignore loaned and busy pages. */ 1190 if (pg && pg->loan_count == 0 && (pg->flags & PG_BUSY) == 0) { 1191 uvm_fault_upper_neighbor(ufi, flt, currva, 1192 pg, anon->an_ref > 1); 1193 } 1194 } 1195 1196 /* locked: maps(read), amap(if there) */ 1197 KASSERT(amap == NULL || mutex_owned(amap->am_lock)); 1198 /* (shadowed == true) if there is an anon at the faulting address */ 1199 UVMHIST_LOG(maphist, " shadowed=%d, will_get=%d", shadowed, 1200 (ufi->entry->object.uvm_obj && shadowed != false),0,0); 1201 1202 /* 1203 * note that if we are really short of RAM we could sleep in the above 1204 * call to pmap_enter with everything locked. bad? 1205 * 1206 * XXX Actually, that is bad; pmap_enter() should just fail in that 1207 * XXX case. --thorpej 1208 */ 1209 1210 return 0; 1211 } 1212 1213 /* 1214 * uvm_fault_upper_neighbor: enter single lower neighbor page. 1215 * 1216 * => called with amap and anon locked. 1217 */ 1218 1219 static void 1220 uvm_fault_upper_neighbor( 1221 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 1222 vaddr_t currva, struct vm_page *pg, bool readonly) 1223 { 1224 UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist); 1225 1226 /* locked: amap, anon */ 1227 1228 mutex_enter(&uvm_pageqlock); 1229 uvm_pageenqueue(pg); 1230 mutex_exit(&uvm_pageqlock); 1231 UVMHIST_LOG(maphist, 1232 " MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x", 1233 ufi->orig_map->pmap, currva, pg, 0); 1234 uvmexp.fltnamap++; 1235 1236 /* 1237 * Since this page isn't the page that's actually faulting, 1238 * ignore pmap_enter() failures; it's not critical that we 1239 * enter these right now. 1240 */ 1241 1242 (void) pmap_enter(ufi->orig_map->pmap, currva, 1243 VM_PAGE_TO_PHYS(pg), 1244 readonly ? (flt->enter_prot & ~VM_PROT_WRITE) : 1245 flt->enter_prot, 1246 PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0)); 1247 1248 pmap_update(ufi->orig_map->pmap); 1249 } 1250 1251 /* 1252 * uvm_fault_upper: handle upper fault. 1253 * 1254 * 1. acquire anon lock. 1255 * 2. get anon. let uvmfault_anonget do the dirty work. 1256 * 3. handle loan. 1257 * 4. dispatch direct or promote handlers. 1258 */ 1259 1260 static int 1261 uvm_fault_upper( 1262 struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 1263 struct vm_anon **anons) 1264 { 1265 struct vm_amap * const amap = ufi->entry->aref.ar_amap; 1266 struct vm_anon * const anon = anons[flt->centeridx]; 1267 struct uvm_object *uobj; 1268 int error; 1269 UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist); 1270 1271 /* locked: maps(read), amap, anon */ 1272 KASSERT(mutex_owned(amap->am_lock)); 1273 KASSERT(anon->an_lock == amap->am_lock); 1274 1275 /* 1276 * handle case 1: fault on an anon in our amap 1277 */ 1278 1279 UVMHIST_LOG(maphist, " case 1 fault: anon=0x%x", anon, 0,0,0); 1280 1281 /* 1282 * no matter if we have case 1A or case 1B we are going to need to 1283 * have the anon's memory resident. ensure that now. 1284 */ 1285 1286 /* 1287 * let uvmfault_anonget do the dirty work. 1288 * if it fails (!OK) it will unlock everything for us. 1289 * if it succeeds, locks are still valid and locked. 1290 * also, if it is OK, then the anon's page is on the queues. 1291 * if the page is on loan from a uvm_object, then anonget will 1292 * lock that object for us if it does not fail. 1293 */ 1294 1295 error = uvmfault_anonget(ufi, amap, anon); 1296 switch (error) { 1297 case 0: 1298 break; 1299 1300 case ERESTART: 1301 return ERESTART; 1302 1303 case EAGAIN: 1304 kpause("fltagain1", false, hz/2, NULL); 1305 return ERESTART; 1306 1307 default: 1308 return error; 1309 } 1310 1311 /* 1312 * uobj is non null if the page is on loan from an object (i.e. uobj) 1313 */ 1314 1315 uobj = anon->an_page->uobject; /* locked by anonget if !NULL */ 1316 1317 /* locked: maps(read), amap, anon, uobj(if one) */ 1318 KASSERT(mutex_owned(amap->am_lock)); 1319 KASSERT(anon->an_lock == amap->am_lock); 1320 KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock)); 1321 1322 /* 1323 * special handling for loaned pages 1324 */ 1325 1326 if (anon->an_page->loan_count) { 1327 error = uvm_fault_upper_loan(ufi, flt, anon, &uobj); 1328 if (error != 0) 1329 return error; 1330 } 1331 1332 /* 1333 * if we are case 1B then we will need to allocate a new blank 1334 * anon to transfer the data into. note that we have a lock 1335 * on anon, so no one can busy or release the page until we are done. 1336 * also note that the ref count can't drop to zero here because 1337 * it is > 1 and we are only dropping one ref. 1338 * 1339 * in the (hopefully very rare) case that we are out of RAM we 1340 * will unlock, wait for more RAM, and refault. 1341 * 1342 * if we are out of anon VM we kill the process (XXX: could wait?). 1343 */ 1344 1345 if (flt->cow_now && anon->an_ref > 1) { 1346 flt->promote = true; 1347 error = uvm_fault_upper_promote(ufi, flt, uobj, anon); 1348 } else { 1349 error = uvm_fault_upper_direct(ufi, flt, uobj, anon); 1350 } 1351 return error; 1352 } 1353 1354 /* 1355 * uvm_fault_upper_loan: handle loaned upper page. 1356 * 1357 * 1. if not cow'ing now, simply adjust flt->enter_prot. 1358 * 2. if cow'ing now, and if ref count is 1, break loan. 1359 */ 1360 1361 static int 1362 uvm_fault_upper_loan( 1363 struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 1364 struct vm_anon *anon, struct uvm_object **ruobj) 1365 { 1366 struct vm_amap * const amap = ufi->entry->aref.ar_amap; 1367 int error = 0; 1368 UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist); 1369 1370 if (!flt->cow_now) { 1371 1372 /* 1373 * for read faults on loaned pages we just cap the 1374 * protection at read-only. 1375 */ 1376 1377 flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE; 1378 1379 } else { 1380 /* 1381 * note that we can't allow writes into a loaned page! 1382 * 1383 * if we have a write fault on a loaned page in an 1384 * anon then we need to look at the anon's ref count. 1385 * if it is greater than one then we are going to do 1386 * a normal copy-on-write fault into a new anon (this 1387 * is not a problem). however, if the reference count 1388 * is one (a case where we would normally allow a 1389 * write directly to the page) then we need to kill 1390 * the loan before we continue. 1391 */ 1392 1393 /* >1 case is already ok */ 1394 if (anon->an_ref == 1) { 1395 error = uvm_loanbreak_anon(anon, *ruobj); 1396 if (error != 0) { 1397 uvmfault_unlockall(ufi, amap, *ruobj); 1398 uvm_wait("flt_noram2"); 1399 return ERESTART; 1400 } 1401 /* if we were a loan reciever uobj is gone */ 1402 if (*ruobj) 1403 *ruobj = NULL; 1404 } 1405 } 1406 return error; 1407 } 1408 1409 /* 1410 * uvm_fault_upper_promote: promote upper page. 1411 * 1412 * 1. call uvmfault_promote. 1413 * 2. enqueue page. 1414 * 3. deref. 1415 * 4. pass page to uvm_fault_upper_enter. 1416 */ 1417 1418 static int 1419 uvm_fault_upper_promote( 1420 struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 1421 struct uvm_object *uobj, struct vm_anon *anon) 1422 { 1423 struct vm_anon * const oanon = anon; 1424 struct vm_page *pg; 1425 int error; 1426 UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist); 1427 1428 UVMHIST_LOG(maphist, " case 1B: COW fault",0,0,0,0); 1429 uvmexp.flt_acow++; 1430 1431 error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon, 1432 &flt->anon_spare); 1433 switch (error) { 1434 case 0: 1435 break; 1436 case ERESTART: 1437 return ERESTART; 1438 default: 1439 return error; 1440 } 1441 1442 KASSERT(anon == NULL || anon->an_lock == oanon->an_lock); 1443 1444 pg = anon->an_page; 1445 mutex_enter(&uvm_pageqlock); 1446 uvm_pageenqueue(pg); /* uvm_fault_upper_done will activate the page */ 1447 mutex_exit(&uvm_pageqlock); 1448 pg->flags &= ~(PG_BUSY|PG_FAKE); 1449 UVM_PAGE_OWN(pg, NULL); 1450 1451 /* deref: can not drop to zero here by defn! */ 1452 KASSERT(oanon->an_ref > 1); 1453 oanon->an_ref--; 1454 1455 /* 1456 * note: oanon is still locked, as is the new anon. we 1457 * need to check for this later when we unlock oanon; if 1458 * oanon != anon, we'll have to unlock anon, too. 1459 */ 1460 1461 return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon); 1462 } 1463 1464 /* 1465 * uvm_fault_upper_direct: handle direct fault. 1466 */ 1467 1468 static int 1469 uvm_fault_upper_direct( 1470 struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 1471 struct uvm_object *uobj, struct vm_anon *anon) 1472 { 1473 struct vm_anon * const oanon = anon; 1474 struct vm_page *pg; 1475 UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist); 1476 1477 uvmexp.flt_anon++; 1478 pg = anon->an_page; 1479 if (anon->an_ref > 1) /* disallow writes to ref > 1 anons */ 1480 flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE; 1481 1482 return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon); 1483 } 1484 1485 /* 1486 * uvm_fault_upper_enter: enter h/w mapping of upper page. 1487 */ 1488 1489 static int 1490 uvm_fault_upper_enter( 1491 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 1492 struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg, 1493 struct vm_anon *oanon) 1494 { 1495 struct vm_amap * const amap = ufi->entry->aref.ar_amap; 1496 UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist); 1497 1498 /* locked: maps(read), amap, oanon, anon(if different from oanon) */ 1499 KASSERT(mutex_owned(amap->am_lock)); 1500 KASSERT(anon->an_lock == amap->am_lock); 1501 KASSERT(oanon->an_lock == amap->am_lock); 1502 KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock)); 1503 1504 /* 1505 * now map the page in. 1506 */ 1507 1508 UVMHIST_LOG(maphist, 1509 " MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x, promote=%d", 1510 ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote); 1511 if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr, 1512 VM_PAGE_TO_PHYS(pg), 1513 flt->enter_prot, flt->access_type | PMAP_CANFAIL | 1514 (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) { 1515 1516 /* 1517 * No need to undo what we did; we can simply think of 1518 * this as the pmap throwing away the mapping information. 1519 * 1520 * We do, however, have to go through the ReFault path, 1521 * as the map may change while we're asleep. 1522 */ 1523 1524 uvmfault_unlockall(ufi, amap, uobj); 1525 if (!uvm_reclaimable()) { 1526 UVMHIST_LOG(maphist, 1527 "<- failed. out of VM",0,0,0,0); 1528 /* XXX instrumentation */ 1529 return ENOMEM; 1530 } 1531 /* XXX instrumentation */ 1532 uvm_wait("flt_pmfail1"); 1533 return ERESTART; 1534 } 1535 1536 uvm_fault_upper_done(ufi, flt, anon, pg); 1537 1538 /* 1539 * done case 1! finish up by unlocking everything and returning success 1540 */ 1541 1542 pmap_update(ufi->orig_map->pmap); 1543 uvmfault_unlockall(ufi, amap, uobj); 1544 return 0; 1545 } 1546 1547 /* 1548 * uvm_fault_upper_done: queue upper center page. 1549 */ 1550 1551 static void 1552 uvm_fault_upper_done( 1553 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 1554 struct vm_anon *anon, struct vm_page *pg) 1555 { 1556 const bool wire_paging = flt->wire_paging; 1557 1558 UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist); 1559 1560 /* 1561 * ... update the page queues. 1562 */ 1563 1564 mutex_enter(&uvm_pageqlock); 1565 if (wire_paging) { 1566 uvm_pagewire(pg); 1567 1568 /* 1569 * since the now-wired page cannot be paged out, 1570 * release its swap resources for others to use. 1571 * since an anon with no swap cannot be PG_CLEAN, 1572 * clear its clean flag now. 1573 */ 1574 1575 pg->flags &= ~(PG_CLEAN); 1576 1577 } else { 1578 uvm_pageactivate(pg); 1579 } 1580 mutex_exit(&uvm_pageqlock); 1581 1582 if (wire_paging) { 1583 uvm_anon_dropswap(anon); 1584 } 1585 } 1586 1587 /* 1588 * uvm_fault_lower: handle lower fault. 1589 * 1590 * 1. check uobj 1591 * 1.1. if null, ZFOD. 1592 * 1.2. if not null, look up unnmapped neighbor pages. 1593 * 2. for center page, check if promote. 1594 * 2.1. ZFOD always needs promotion. 1595 * 2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode). 1596 * 3. if uobj is not ZFOD and page is not found, do i/o. 1597 * 4. dispatch either direct / promote fault. 1598 */ 1599 1600 static int 1601 uvm_fault_lower( 1602 struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 1603 struct vm_page **pages) 1604 { 1605 #ifdef DIAGNOSTIC 1606 struct vm_amap *amap = ufi->entry->aref.ar_amap; 1607 #endif 1608 struct uvm_object *uobj = ufi->entry->object.uvm_obj; 1609 struct vm_page *uobjpage; 1610 int error; 1611 UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist); 1612 1613 /* 1614 * now, if the desired page is not shadowed by the amap and we have 1615 * a backing object that does not have a special fault routine, then 1616 * we ask (with pgo_get) the object for resident pages that we care 1617 * about and attempt to map them in. we do not let pgo_get block 1618 * (PGO_LOCKED). 1619 */ 1620 1621 if (uobj == NULL) { 1622 /* zero fill; don't care neighbor pages */ 1623 uobjpage = NULL; 1624 } else { 1625 uvm_fault_lower_lookup(ufi, flt, pages); 1626 uobjpage = pages[flt->centeridx]; 1627 } 1628 1629 /* 1630 * note that at this point we are done with any front or back pages. 1631 * we are now going to focus on the center page (i.e. the one we've 1632 * faulted on). if we have faulted on the upper (anon) layer 1633 * [i.e. case 1], then the anon we want is anons[centeridx] (we have 1634 * not touched it yet). if we have faulted on the bottom (uobj) 1635 * layer [i.e. case 2] and the page was both present and available, 1636 * then we've got a pointer to it as "uobjpage" and we've already 1637 * made it BUSY. 1638 */ 1639 1640 /* 1641 * locked: 1642 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null) 1643 */ 1644 KASSERT(amap == NULL || mutex_owned(amap->am_lock)); 1645 KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock)); 1646 KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0); 1647 1648 /* 1649 * note that uobjpage can not be PGO_DONTCARE at this point. we now 1650 * set uobjpage to PGO_DONTCARE if we are doing a zero fill. if we 1651 * have a backing object, check and see if we are going to promote 1652 * the data up to an anon during the fault. 1653 */ 1654 1655 if (uobj == NULL) { 1656 uobjpage = PGO_DONTCARE; 1657 flt->promote = true; /* always need anon here */ 1658 } else { 1659 KASSERT(uobjpage != PGO_DONTCARE); 1660 flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry); 1661 } 1662 UVMHIST_LOG(maphist, " case 2 fault: promote=%d, zfill=%d", 1663 flt->promote, (uobj == NULL), 0,0); 1664 1665 /* 1666 * if uobjpage is not null then we do not need to do I/O to get the 1667 * uobjpage. 1668 * 1669 * if uobjpage is null, then we need to unlock and ask the pager to 1670 * get the data for us. once we have the data, we need to reverify 1671 * the state the world. we are currently not holding any resources. 1672 */ 1673 1674 if (uobjpage) { 1675 /* update rusage counters */ 1676 curlwp->l_ru.ru_minflt++; 1677 } else { 1678 error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage); 1679 if (error != 0) 1680 return error; 1681 } 1682 1683 /* 1684 * locked: 1685 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj) 1686 */ 1687 KASSERT(amap == NULL || mutex_owned(amap->am_lock)); 1688 KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock)); 1689 KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0); 1690 1691 /* 1692 * notes: 1693 * - at this point uobjpage can not be NULL 1694 * - at this point uobjpage can not be PG_RELEASED (since we checked 1695 * for it above) 1696 * - at this point uobjpage could be PG_WANTED (handle later) 1697 */ 1698 1699 KASSERT(uobjpage != NULL); 1700 KASSERT(uobj == NULL || uobj == uobjpage->uobject); 1701 KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) || 1702 (uobjpage->flags & PG_CLEAN) != 0); 1703 1704 if (!flt->promote) { 1705 error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage); 1706 } else { 1707 error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage); 1708 } 1709 return error; 1710 } 1711 1712 /* 1713 * uvm_fault_lower_lookup: look up on-memory uobj pages. 1714 * 1715 * 1. get on-memory pages. 1716 * 2. if failed, give up (get only center page later). 1717 * 3. if succeeded, enter h/w mapping of neighbor pages. 1718 */ 1719 1720 static void 1721 uvm_fault_lower_lookup( 1722 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 1723 struct vm_page **pages) 1724 { 1725 struct uvm_object *uobj = ufi->entry->object.uvm_obj; 1726 int lcv, gotpages; 1727 vaddr_t currva; 1728 UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist); 1729 1730 mutex_enter(uobj->vmobjlock); 1731 /* Locked: maps(read), amap(if there), uobj */ 1732 1733 uvmexp.fltlget++; 1734 gotpages = flt->npages; 1735 (void) uobj->pgops->pgo_get(uobj, 1736 ufi->entry->offset + flt->startva - ufi->entry->start, 1737 pages, &gotpages, flt->centeridx, 1738 flt->access_type & MASK(ufi->entry), ufi->entry->advice, PGO_LOCKED); 1739 1740 KASSERT(mutex_owned(uobj->vmobjlock)); 1741 1742 /* 1743 * check for pages to map, if we got any 1744 */ 1745 1746 if (gotpages == 0) { 1747 pages[flt->centeridx] = NULL; 1748 return; 1749 } 1750 1751 currva = flt->startva; 1752 for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) { 1753 struct vm_page *curpg; 1754 1755 curpg = pages[lcv]; 1756 if (curpg == NULL || curpg == PGO_DONTCARE) { 1757 continue; 1758 } 1759 KASSERT(curpg->uobject == uobj); 1760 1761 /* 1762 * if center page is resident and not PG_BUSY|PG_RELEASED 1763 * then pgo_get made it PG_BUSY for us and gave us a handle 1764 * to it. 1765 */ 1766 1767 if (lcv == flt->centeridx) { 1768 UVMHIST_LOG(maphist, " got uobjpage " 1769 "(0x%x) with locked get", 1770 curpg, 0,0,0); 1771 } else { 1772 bool readonly = (curpg->flags & PG_RDONLY) 1773 || (curpg->loan_count > 0) 1774 || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject); 1775 1776 uvm_fault_lower_neighbor(ufi, flt, 1777 currva, curpg, readonly); 1778 } 1779 } 1780 pmap_update(ufi->orig_map->pmap); 1781 } 1782 1783 /* 1784 * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page. 1785 */ 1786 1787 static void 1788 uvm_fault_lower_neighbor( 1789 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 1790 vaddr_t currva, struct vm_page *pg, bool readonly) 1791 { 1792 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 1793 1794 /* locked: maps(read), amap(if there), uobj */ 1795 1796 /* 1797 * calling pgo_get with PGO_LOCKED returns us pages which 1798 * are neither busy nor released, so we don't need to check 1799 * for this. we can just directly enter the pages. 1800 */ 1801 1802 mutex_enter(&uvm_pageqlock); 1803 uvm_pageenqueue(pg); 1804 mutex_exit(&uvm_pageqlock); 1805 UVMHIST_LOG(maphist, 1806 " MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x", 1807 ufi->orig_map->pmap, currva, pg, 0); 1808 uvmexp.fltnomap++; 1809 1810 /* 1811 * Since this page isn't the page that's actually faulting, 1812 * ignore pmap_enter() failures; it's not critical that we 1813 * enter these right now. 1814 * NOTE: page can't be PG_WANTED or PG_RELEASED because we've 1815 * held the lock the whole time we've had the handle. 1816 */ 1817 KASSERT((pg->flags & PG_PAGEOUT) == 0); 1818 KASSERT((pg->flags & PG_RELEASED) == 0); 1819 KASSERT((pg->flags & PG_WANTED) == 0); 1820 KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) || (pg->flags & PG_CLEAN) != 0); 1821 pg->flags &= ~(PG_BUSY); 1822 UVM_PAGE_OWN(pg, NULL); 1823 1824 KASSERT(mutex_owned(pg->uobject->vmobjlock)); 1825 (void) pmap_enter(ufi->orig_map->pmap, currva, 1826 VM_PAGE_TO_PHYS(pg), 1827 readonly ? (flt->enter_prot & ~VM_PROT_WRITE) : 1828 flt->enter_prot & MASK(ufi->entry), 1829 PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0)); 1830 } 1831 1832 /* 1833 * uvm_fault_lower_io: get lower page from backing store. 1834 * 1835 * 1. unlock everything, because i/o will block. 1836 * 2. call pgo_get. 1837 * 3. if failed, recover. 1838 * 4. if succeeded, relock everything and verify things. 1839 */ 1840 1841 static int 1842 uvm_fault_lower_io( 1843 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 1844 struct uvm_object **ruobj, struct vm_page **ruobjpage) 1845 { 1846 struct vm_amap * const amap = ufi->entry->aref.ar_amap; 1847 struct uvm_object *uobj = *ruobj; 1848 struct vm_page *pg; 1849 bool locked; 1850 int gotpages; 1851 int error; 1852 voff_t uoff; 1853 UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist); 1854 1855 /* update rusage counters */ 1856 curlwp->l_ru.ru_majflt++; 1857 1858 /* Locked: maps(read), amap(if there), uobj */ 1859 uvmfault_unlockall(ufi, amap, NULL); 1860 1861 /* Locked: uobj */ 1862 KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock)); 1863 1864 uvmexp.fltget++; 1865 gotpages = 1; 1866 pg = NULL; 1867 uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset; 1868 error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages, 1869 0, flt->access_type & MASK(ufi->entry), ufi->entry->advice, 1870 PGO_SYNCIO); 1871 /* locked: pg(if no error) */ 1872 1873 /* 1874 * recover from I/O 1875 */ 1876 1877 if (error) { 1878 if (error == EAGAIN) { 1879 UVMHIST_LOG(maphist, 1880 " pgo_get says TRY AGAIN!",0,0,0,0); 1881 kpause("fltagain2", false, hz/2, NULL); 1882 return ERESTART; 1883 } 1884 1885 #if 0 1886 KASSERT(error != ERESTART); 1887 #else 1888 /* XXXUEBS don't re-fault? */ 1889 if (error == ERESTART) 1890 error = EIO; 1891 #endif 1892 1893 UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)", 1894 error, 0,0,0); 1895 return error; 1896 } 1897 1898 /* 1899 * re-verify the state of the world by first trying to relock 1900 * the maps. always relock the object. 1901 */ 1902 1903 locked = uvmfault_relock(ufi); 1904 if (locked && amap) 1905 amap_lock(amap); 1906 1907 /* might be changed */ 1908 uobj = pg->uobject; 1909 1910 mutex_enter(uobj->vmobjlock); 1911 KASSERT((pg->flags & PG_BUSY) != 0); 1912 1913 mutex_enter(&uvm_pageqlock); 1914 uvm_pageactivate(pg); 1915 mutex_exit(&uvm_pageqlock); 1916 1917 /* locked(locked): maps(read), amap(if !null), uobj, pg */ 1918 /* locked(!locked): uobj, pg */ 1919 1920 /* 1921 * verify that the page has not be released and re-verify 1922 * that amap slot is still free. if there is a problem, 1923 * we unlock and clean up. 1924 */ 1925 1926 if ((pg->flags & PG_RELEASED) != 0 || 1927 (locked && amap && amap_lookup(&ufi->entry->aref, 1928 ufi->orig_rvaddr - ufi->entry->start))) { 1929 if (locked) 1930 uvmfault_unlockall(ufi, amap, NULL); 1931 locked = false; 1932 } 1933 1934 /* 1935 * didn't get the lock? release the page and retry. 1936 */ 1937 1938 if (locked == false) { 1939 UVMHIST_LOG(maphist, 1940 " wasn't able to relock after fault: retry", 1941 0,0,0,0); 1942 if (pg->flags & PG_WANTED) { 1943 wakeup(pg); 1944 } 1945 if ((pg->flags & PG_RELEASED) == 0) { 1946 pg->flags &= ~(PG_BUSY | PG_WANTED); 1947 UVM_PAGE_OWN(pg, NULL); 1948 } else { 1949 uvmexp.fltpgrele++; 1950 uvm_pagefree(pg); 1951 } 1952 mutex_exit(uobj->vmobjlock); 1953 return ERESTART; 1954 } 1955 1956 /* 1957 * we have the data in pg which is busy and 1958 * not released. we are holding object lock (so the page 1959 * can't be released on us). 1960 */ 1961 1962 /* locked: maps(read), amap(if !null), uobj, pg */ 1963 1964 *ruobj = uobj; 1965 *ruobjpage = pg; 1966 return 0; 1967 } 1968 1969 /* 1970 * uvm_fault_lower_direct: fault lower center page 1971 * 1972 * 1. adjust flt->enter_prot. 1973 * 2. if page is loaned, resolve. 1974 */ 1975 1976 int 1977 uvm_fault_lower_direct( 1978 struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 1979 struct uvm_object *uobj, struct vm_page *uobjpage) 1980 { 1981 struct vm_page *pg; 1982 UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist); 1983 1984 /* 1985 * we are not promoting. if the mapping is COW ensure that we 1986 * don't give more access than we should (e.g. when doing a read 1987 * fault on a COPYONWRITE mapping we want to map the COW page in 1988 * R/O even though the entry protection could be R/W). 1989 * 1990 * set "pg" to the page we want to map in (uobjpage, usually) 1991 */ 1992 1993 uvmexp.flt_obj++; 1994 if (UVM_ET_ISCOPYONWRITE(ufi->entry) || 1995 UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject)) 1996 flt->enter_prot &= ~VM_PROT_WRITE; 1997 pg = uobjpage; /* map in the actual object */ 1998 1999 KASSERT(uobjpage != PGO_DONTCARE); 2000 2001 /* 2002 * we are faulting directly on the page. be careful 2003 * about writing to loaned pages... 2004 */ 2005 2006 if (uobjpage->loan_count) { 2007 uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage); 2008 } 2009 KASSERT(pg == uobjpage); 2010 2011 KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0); 2012 return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg); 2013 } 2014 2015 /* 2016 * uvm_fault_lower_direct_loan: resolve loaned page. 2017 * 2018 * 1. if not cow'ing, adjust flt->enter_prot. 2019 * 2. if cow'ing, break loan. 2020 */ 2021 2022 static int 2023 uvm_fault_lower_direct_loan( 2024 struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 2025 struct uvm_object *uobj, struct vm_page **rpg, 2026 struct vm_page **ruobjpage) 2027 { 2028 struct vm_amap * const amap = ufi->entry->aref.ar_amap; 2029 struct vm_page *pg; 2030 struct vm_page *uobjpage = *ruobjpage; 2031 UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist); 2032 2033 if (!flt->cow_now) { 2034 /* read fault: cap the protection at readonly */ 2035 /* cap! */ 2036 flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE; 2037 } else { 2038 /* write fault: must break the loan here */ 2039 2040 pg = uvm_loanbreak(uobjpage); 2041 if (pg == NULL) { 2042 2043 /* 2044 * drop ownership of page, it can't be released 2045 */ 2046 2047 if (uobjpage->flags & PG_WANTED) 2048 wakeup(uobjpage); 2049 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 2050 UVM_PAGE_OWN(uobjpage, NULL); 2051 2052 uvmfault_unlockall(ufi, amap, uobj); 2053 UVMHIST_LOG(maphist, 2054 " out of RAM breaking loan, waiting", 2055 0,0,0,0); 2056 uvmexp.fltnoram++; 2057 uvm_wait("flt_noram4"); 2058 return ERESTART; 2059 } 2060 *rpg = pg; 2061 *ruobjpage = pg; 2062 } 2063 return 0; 2064 } 2065 2066 /* 2067 * uvm_fault_lower_promote: promote lower page. 2068 * 2069 * 1. call uvmfault_promote. 2070 * 2. fill in data. 2071 * 3. if not ZFOD, dispose old page. 2072 */ 2073 2074 int 2075 uvm_fault_lower_promote( 2076 struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 2077 struct uvm_object *uobj, struct vm_page *uobjpage) 2078 { 2079 struct vm_amap * const amap = ufi->entry->aref.ar_amap; 2080 struct vm_anon *anon; 2081 struct vm_page *pg; 2082 int error; 2083 UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist); 2084 2085 KASSERT(amap != NULL); 2086 2087 /* 2088 * If we are going to promote the data to an anon we 2089 * allocate a blank anon here and plug it into our amap. 2090 */ 2091 error = uvmfault_promote(ufi, NULL, uobjpage, 2092 &anon, &flt->anon_spare); 2093 switch (error) { 2094 case 0: 2095 break; 2096 case ERESTART: 2097 return ERESTART; 2098 default: 2099 return error; 2100 } 2101 2102 pg = anon->an_page; 2103 2104 /* 2105 * Fill in the data. 2106 */ 2107 KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0); 2108 2109 if (uobjpage != PGO_DONTCARE) { 2110 uvmexp.flt_prcopy++; 2111 2112 /* 2113 * promote to shared amap? make sure all sharing 2114 * procs see it 2115 */ 2116 2117 if ((amap_flags(amap) & AMAP_SHARED) != 0) { 2118 pmap_page_protect(uobjpage, VM_PROT_NONE); 2119 /* 2120 * XXX: PAGE MIGHT BE WIRED! 2121 */ 2122 } 2123 2124 /* 2125 * dispose of uobjpage. it can't be PG_RELEASED 2126 * since we still hold the object lock. 2127 */ 2128 2129 if (uobjpage->flags & PG_WANTED) { 2130 /* still have the obj lock */ 2131 wakeup(uobjpage); 2132 } 2133 uobjpage->flags &= ~(PG_BUSY|PG_WANTED); 2134 UVM_PAGE_OWN(uobjpage, NULL); 2135 2136 UVMHIST_LOG(maphist, 2137 " promote uobjpage 0x%x to anon/page 0x%x/0x%x", 2138 uobjpage, anon, pg, 0); 2139 2140 } else { 2141 uvmexp.flt_przero++; 2142 2143 /* 2144 * Page is zero'd and marked dirty by 2145 * uvmfault_promote(). 2146 */ 2147 2148 UVMHIST_LOG(maphist," zero fill anon/page 0x%x/0%x", 2149 anon, pg, 0, 0); 2150 } 2151 2152 return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg); 2153 } 2154 2155 /* 2156 * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted 2157 * from the lower page. 2158 */ 2159 2160 int 2161 uvm_fault_lower_enter( 2162 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 2163 struct uvm_object *uobj, 2164 struct vm_anon *anon, struct vm_page *pg) 2165 { 2166 struct vm_amap * const amap = ufi->entry->aref.ar_amap; 2167 int error; 2168 UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist); 2169 2170 /* 2171 * Locked: 2172 * 2173 * maps(read), amap(if !null), uobj(if !null), 2174 * anon(if !null), pg(if anon), unlock_uobj(if !null) 2175 * 2176 * Note: pg is either the uobjpage or the new page in the new anon. 2177 */ 2178 KASSERT(amap == NULL || mutex_owned(amap->am_lock)); 2179 KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock)); 2180 KASSERT(anon == NULL || anon->an_lock == amap->am_lock); 2181 KASSERT((pg->flags & PG_BUSY) != 0); 2182 2183 /* 2184 * all resources are present. we can now map it in and free our 2185 * resources. 2186 */ 2187 2188 UVMHIST_LOG(maphist, 2189 " MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d", 2190 ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote); 2191 KASSERT((flt->access_type & VM_PROT_WRITE) == 0 || 2192 (pg->flags & PG_RDONLY) == 0); 2193 if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr, 2194 VM_PAGE_TO_PHYS(pg), 2195 (pg->flags & PG_RDONLY) != 0 ? 2196 flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot, 2197 flt->access_type | PMAP_CANFAIL | 2198 (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) { 2199 2200 /* 2201 * No need to undo what we did; we can simply think of 2202 * this as the pmap throwing away the mapping information. 2203 * 2204 * We do, however, have to go through the ReFault path, 2205 * as the map may change while we're asleep. 2206 */ 2207 2208 /* 2209 * ensure that the page is queued in the case that 2210 * we just promoted the page. 2211 */ 2212 2213 mutex_enter(&uvm_pageqlock); 2214 uvm_pageenqueue(pg); 2215 mutex_exit(&uvm_pageqlock); 2216 2217 if (pg->flags & PG_WANTED) 2218 wakeup(pg); 2219 2220 /* 2221 * note that pg can't be PG_RELEASED since we did not drop 2222 * the object lock since the last time we checked. 2223 */ 2224 KASSERT((pg->flags & PG_RELEASED) == 0); 2225 2226 pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED); 2227 UVM_PAGE_OWN(pg, NULL); 2228 2229 uvmfault_unlockall(ufi, amap, uobj); 2230 if (!uvm_reclaimable()) { 2231 UVMHIST_LOG(maphist, 2232 "<- failed. out of VM",0,0,0,0); 2233 /* XXX instrumentation */ 2234 error = ENOMEM; 2235 return error; 2236 } 2237 /* XXX instrumentation */ 2238 uvm_wait("flt_pmfail2"); 2239 return ERESTART; 2240 } 2241 2242 uvm_fault_lower_done(ufi, flt, uobj, pg); 2243 2244 /* 2245 * note that pg can't be PG_RELEASED since we did not drop the object 2246 * lock since the last time we checked. 2247 */ 2248 KASSERT((pg->flags & PG_RELEASED) == 0); 2249 if (pg->flags & PG_WANTED) 2250 wakeup(pg); 2251 pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED); 2252 UVM_PAGE_OWN(pg, NULL); 2253 2254 pmap_update(ufi->orig_map->pmap); 2255 uvmfault_unlockall(ufi, amap, uobj); 2256 2257 UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0); 2258 return 0; 2259 } 2260 2261 /* 2262 * uvm_fault_lower_done: queue lower center page. 2263 */ 2264 2265 void 2266 uvm_fault_lower_done( 2267 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 2268 struct uvm_object *uobj, struct vm_page *pg) 2269 { 2270 bool dropswap = false; 2271 2272 UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist); 2273 2274 mutex_enter(&uvm_pageqlock); 2275 if (flt->wire_paging) { 2276 uvm_pagewire(pg); 2277 if (pg->pqflags & PQ_AOBJ) { 2278 2279 /* 2280 * since the now-wired page cannot be paged out, 2281 * release its swap resources for others to use. 2282 * since an aobj page with no swap cannot be PG_CLEAN, 2283 * clear its clean flag now. 2284 */ 2285 2286 KASSERT(uobj != NULL); 2287 pg->flags &= ~(PG_CLEAN); 2288 dropswap = true; 2289 } 2290 } else { 2291 uvm_pageactivate(pg); 2292 } 2293 mutex_exit(&uvm_pageqlock); 2294 2295 if (dropswap) { 2296 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 2297 } 2298 } 2299 2300 2301 /* 2302 * uvm_fault_wire: wire down a range of virtual addresses in a map. 2303 * 2304 * => map may be read-locked by caller, but MUST NOT be write-locked. 2305 * => if map is read-locked, any operations which may cause map to 2306 * be write-locked in uvm_fault() must be taken care of by 2307 * the caller. See uvm_map_pageable(). 2308 */ 2309 2310 int 2311 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end, 2312 vm_prot_t access_type, int maxprot) 2313 { 2314 vaddr_t va; 2315 int error; 2316 2317 /* 2318 * now fault it in a page at a time. if the fault fails then we have 2319 * to undo what we have done. note that in uvm_fault VM_PROT_NONE 2320 * is replaced with the max protection if fault_type is VM_FAULT_WIRE. 2321 */ 2322 2323 /* 2324 * XXX work around overflowing a vaddr_t. this prevents us from 2325 * wiring the last page in the address space, though. 2326 */ 2327 if (start > end) { 2328 return EFAULT; 2329 } 2330 2331 for (va = start; va < end; va += PAGE_SIZE) { 2332 error = uvm_fault_internal(map, va, access_type, 2333 (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE); 2334 if (error) { 2335 if (va != start) { 2336 uvm_fault_unwire(map, start, va); 2337 } 2338 return error; 2339 } 2340 } 2341 return 0; 2342 } 2343 2344 /* 2345 * uvm_fault_unwire(): unwire range of virtual space. 2346 */ 2347 2348 void 2349 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end) 2350 { 2351 vm_map_lock_read(map); 2352 uvm_fault_unwire_locked(map, start, end); 2353 vm_map_unlock_read(map); 2354 } 2355 2356 /* 2357 * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire(). 2358 * 2359 * => map must be at least read-locked. 2360 */ 2361 2362 void 2363 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end) 2364 { 2365 struct vm_map_entry *entry, *oentry; 2366 pmap_t pmap = vm_map_pmap(map); 2367 vaddr_t va; 2368 paddr_t pa; 2369 struct vm_page *pg; 2370 2371 /* 2372 * we assume that the area we are unwiring has actually been wired 2373 * in the first place. this means that we should be able to extract 2374 * the PAs from the pmap. we also lock out the page daemon so that 2375 * we can call uvm_pageunwire. 2376 */ 2377 2378 /* 2379 * find the beginning map entry for the region. 2380 */ 2381 2382 KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map)); 2383 if (uvm_map_lookup_entry(map, start, &entry) == false) 2384 panic("uvm_fault_unwire_locked: address not in map"); 2385 2386 oentry = NULL; 2387 for (va = start; va < end; va += PAGE_SIZE) { 2388 if (pmap_extract(pmap, va, &pa) == false) 2389 continue; 2390 2391 /* 2392 * find the map entry for the current address. 2393 */ 2394 2395 KASSERT(va >= entry->start); 2396 while (va >= entry->end) { 2397 KASSERT(entry->next != &map->header && 2398 entry->next->start <= entry->end); 2399 entry = entry->next; 2400 } 2401 2402 /* 2403 * lock it. 2404 */ 2405 2406 if (entry != oentry) { 2407 if (oentry != NULL) { 2408 mutex_exit(&uvm_pageqlock); 2409 uvm_map_unlock_entry(oentry); 2410 } 2411 uvm_map_lock_entry(entry); 2412 mutex_enter(&uvm_pageqlock); 2413 oentry = entry; 2414 } 2415 2416 /* 2417 * if the entry is no longer wired, tell the pmap. 2418 */ 2419 2420 if (VM_MAPENT_ISWIRED(entry) == 0) 2421 pmap_unwire(pmap, va); 2422 2423 pg = PHYS_TO_VM_PAGE(pa); 2424 if (pg) 2425 uvm_pageunwire(pg); 2426 } 2427 2428 if (oentry != NULL) { 2429 mutex_exit(&uvm_pageqlock); 2430 uvm_map_unlock_entry(entry); 2431 } 2432 } 2433