1/* $OpenBSD: uvm_fault.c,v 1.135 2023/09/05 05:08:26 guenther Exp $ */ 2/* $NetBSD: uvm_fault.c,v 1.51 2000/08/06 00:22:53 thorpej Exp $ */ 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 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 * 28 * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp 29 */ 30 31/* 32 * uvm_fault.c: fault handler 33 */ 34 35#include <sys/param.h> 36#include <sys/systm.h> 37#include <sys/kernel.h> 38#include <sys/percpu.h> 39#include <sys/proc.h> 40#include <sys/malloc.h> 41#include <sys/mman.h> 42#include <sys/tracepoint.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) 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 */ 146struct uvm_advice { 147 int nback; 148 int nforw; 149}; 150 151/* 152 * page range array: set up in uvmfault_init(). 153 */ 154static struct uvm_advice uvmadvice[MADV_MASK + 1]; 155 156#define UVM_MAXRANGE 16 /* must be max() of nback+nforw+1 */ 157 158/* 159 * private prototypes 160 */ 161static void uvmfault_amapcopy(struct uvm_faultinfo *); 162static inline void uvmfault_anonflush(struct vm_anon **, int); 163void uvmfault_unlockmaps(struct uvm_faultinfo *, boolean_t); 164void uvmfault_update_stats(struct uvm_faultinfo *); 165 166/* 167 * inline functions 168 */ 169/* 170 * uvmfault_anonflush: try and deactivate pages in specified anons 171 * 172 * => does not have to deactivate page if it is busy 173 */ 174static inline void 175uvmfault_anonflush(struct vm_anon **anons, int n) 176{ 177 int lcv; 178 struct vm_page *pg; 179 180 for (lcv = 0; lcv < n; lcv++) { 181 if (anons[lcv] == NULL) 182 continue; 183 KASSERT(rw_lock_held(anons[lcv]->an_lock)); 184 pg = anons[lcv]->an_page; 185 if (pg && (pg->pg_flags & PG_BUSY) == 0) { 186 uvm_lock_pageq(); 187 if (pg->wire_count == 0) { 188 pmap_page_protect(pg, PROT_NONE); 189 uvm_pagedeactivate(pg); 190 } 191 uvm_unlock_pageq(); 192 } 193 } 194} 195 196/* 197 * normal functions 198 */ 199/* 200 * uvmfault_init: compute proper values for the uvmadvice[] array. 201 */ 202void 203uvmfault_init(void) 204{ 205 int npages; 206 207 npages = atop(16384); 208 if (npages > 0) { 209 KASSERT(npages <= UVM_MAXRANGE / 2); 210 uvmadvice[MADV_NORMAL].nforw = npages; 211 uvmadvice[MADV_NORMAL].nback = npages - 1; 212 } 213 214 npages = atop(32768); 215 if (npages > 0) { 216 KASSERT(npages <= UVM_MAXRANGE / 2); 217 uvmadvice[MADV_SEQUENTIAL].nforw = npages - 1; 218 uvmadvice[MADV_SEQUENTIAL].nback = npages; 219 } 220} 221 222/* 223 * uvmfault_amapcopy: clear "needs_copy" in a map. 224 * 225 * => called with VM data structures unlocked (usually, see below) 226 * => we get a write lock on the maps and clear needs_copy for a VA 227 * => if we are out of RAM we sleep (waiting for more) 228 */ 229static void 230uvmfault_amapcopy(struct uvm_faultinfo *ufi) 231{ 232 for (;;) { 233 /* 234 * no mapping? give up. 235 */ 236 if (uvmfault_lookup(ufi, TRUE) == FALSE) 237 return; 238 239 /* 240 * copy if needed. 241 */ 242 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) 243 amap_copy(ufi->map, ufi->entry, M_NOWAIT, 244 UVM_ET_ISSTACK(ufi->entry) ? FALSE : TRUE, 245 ufi->orig_rvaddr, ufi->orig_rvaddr + 1); 246 247 /* 248 * didn't work? must be out of RAM. unlock and sleep. 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 uvmfault_unlockmaps(ufi, TRUE); 260 return; 261 } 262 /*NOTREACHED*/ 263} 264 265/* 266 * uvmfault_anonget: get data in an anon into a non-busy, non-released 267 * page in that anon. 268 * 269 * => Map, amap and thus anon should be locked by caller. 270 * => If we fail, we unlock everything and error is returned. 271 * => If we are successful, return with everything still locked. 272 * => We do not move the page on the queues [gets moved later]. If we 273 * allocate a new page [we_own], it gets put on the queues. Either way, 274 * the result is that the page is on the queues at return time 275 */ 276int 277uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap, 278 struct vm_anon *anon) 279{ 280 struct vm_page *pg; 281 int error; 282 283 KASSERT(rw_lock_held(anon->an_lock)); 284 KASSERT(anon->an_lock == amap->am_lock); 285 286 /* Increment the counters.*/ 287 counters_inc(uvmexp_counters, flt_anget); 288 if (anon->an_page) { 289 curproc->p_ru.ru_minflt++; 290 } else { 291 curproc->p_ru.ru_majflt++; 292 } 293 error = 0; 294 295 /* 296 * Loop until we get the anon data, or fail. 297 */ 298 for (;;) { 299 boolean_t we_own, locked; 300 /* 301 * Note: 'we_own' will become true if we set PG_BUSY on a page. 302 */ 303 we_own = FALSE; 304 pg = anon->an_page; 305 306 /* 307 * Is page resident? Make sure it is not busy/released. 308 */ 309 if (pg) { 310 KASSERT(pg->pg_flags & PQ_ANON); 311 KASSERT(pg->uanon == anon); 312 313 /* 314 * if the page is busy, we drop all the locks and 315 * try again. 316 */ 317 if ((pg->pg_flags & (PG_BUSY|PG_RELEASED)) == 0) 318 return (VM_PAGER_OK); 319 atomic_setbits_int(&pg->pg_flags, PG_WANTED); 320 counters_inc(uvmexp_counters, flt_pgwait); 321 322 /* 323 * The last unlock must be an atomic unlock and wait 324 * on the owner of page. 325 */ 326 if (pg->uobject) { 327 /* Owner of page is UVM object. */ 328 uvmfault_unlockall(ufi, amap, NULL); 329 rwsleep_nsec(pg, pg->uobject->vmobjlock, 330 PVM | PNORELOCK, "anonget1", INFSLP); 331 } else { 332 /* Owner of page is anon. */ 333 uvmfault_unlockall(ufi, NULL, NULL); 334 rwsleep_nsec(pg, anon->an_lock, PVM | PNORELOCK, 335 "anonget2", INFSLP); 336 } 337 } else { 338 /* 339 * No page, therefore allocate one. 340 */ 341 pg = uvm_pagealloc(NULL, 0, anon, 0); 342 if (pg == NULL) { 343 /* Out of memory. Wait a little. */ 344 uvmfault_unlockall(ufi, amap, NULL); 345 counters_inc(uvmexp_counters, flt_noram); 346 uvm_wait("flt_noram1"); 347 } else { 348 /* PG_BUSY bit is set. */ 349 we_own = TRUE; 350 uvmfault_unlockall(ufi, amap, NULL); 351 352 /* 353 * Pass a PG_BUSY+PG_FAKE+PG_CLEAN page into 354 * the uvm_swap_get() function with all data 355 * structures unlocked. Note that it is OK 356 * to read an_swslot here, because we hold 357 * PG_BUSY on the page. 358 */ 359 counters_inc(uvmexp_counters, pageins); 360 error = uvm_swap_get(pg, anon->an_swslot, 361 PGO_SYNCIO); 362 363 /* 364 * We clean up after the I/O below in the 365 * 'we_own' case. 366 */ 367 } 368 } 369 370 /* 371 * Re-lock the map and anon. 372 */ 373 locked = uvmfault_relock(ufi); 374 if (locked || we_own) { 375 rw_enter(anon->an_lock, RW_WRITE); 376 } 377 378 /* 379 * If we own the page (i.e. we set PG_BUSY), then we need 380 * to clean up after the I/O. There are three cases to 381 * consider: 382 * 383 * 1) Page was released during I/O: free anon and ReFault. 384 * 2) I/O not OK. Free the page and cause the fault to fail. 385 * 3) I/O OK! Activate the page and sync with the non-we_own 386 * case (i.e. drop anon lock if not locked). 387 */ 388 if (we_own) { 389 if (pg->pg_flags & PG_WANTED) { 390 wakeup(pg); 391 } 392 393 /* 394 * if we were RELEASED during I/O, then our anon is 395 * no longer part of an amap. we need to free the 396 * anon and try again. 397 */ 398 if (pg->pg_flags & PG_RELEASED) { 399 KASSERT(anon->an_ref == 0); 400 /* 401 * Released while we had unlocked amap. 402 */ 403 if (locked) 404 uvmfault_unlockall(ufi, NULL, NULL); 405 uvm_anon_release(anon); /* frees page for us */ 406 counters_inc(uvmexp_counters, flt_pgrele); 407 return (VM_PAGER_REFAULT); /* refault! */ 408 } 409 410 if (error != VM_PAGER_OK) { 411 KASSERT(error != VM_PAGER_PEND); 412 413 /* remove page from anon */ 414 anon->an_page = NULL; 415 416 /* 417 * Remove the swap slot from the anon and 418 * mark the anon as having no real slot. 419 * Do not free the swap slot, thus preventing 420 * it from being used again. 421 */ 422 uvm_swap_markbad(anon->an_swslot, 1); 423 anon->an_swslot = SWSLOT_BAD; 424 425 /* 426 * Note: page was never !PG_BUSY, so it 427 * cannot be mapped and thus no need to 428 * pmap_page_protect() it. 429 */ 430 uvm_lock_pageq(); 431 uvm_pagefree(pg); 432 uvm_unlock_pageq(); 433 434 if (locked) { 435 uvmfault_unlockall(ufi, NULL, NULL); 436 } 437 rw_exit(anon->an_lock); 438 return (VM_PAGER_ERROR); 439 } 440 441 /* 442 * We have successfully read the page, activate it. 443 */ 444 pmap_clear_modify(pg); 445 uvm_lock_pageq(); 446 uvm_pageactivate(pg); 447 uvm_unlock_pageq(); 448 atomic_clearbits_int(&pg->pg_flags, 449 PG_WANTED|PG_BUSY|PG_FAKE); 450 UVM_PAGE_OWN(pg, NULL); 451 } 452 453 /* 454 * We were not able to re-lock the map - restart the fault. 455 */ 456 if (!locked) { 457 if (we_own) { 458 rw_exit(anon->an_lock); 459 } 460 return (VM_PAGER_REFAULT); 461 } 462 463 /* 464 * Verify that no one has touched the amap and moved 465 * the anon on us. 466 */ 467 if (ufi != NULL && amap_lookup(&ufi->entry->aref, 468 ufi->orig_rvaddr - ufi->entry->start) != anon) { 469 470 uvmfault_unlockall(ufi, amap, NULL); 471 return (VM_PAGER_REFAULT); 472 } 473 474 /* 475 * Retry.. 476 */ 477 counters_inc(uvmexp_counters, flt_anretry); 478 continue; 479 480 } 481 /*NOTREACHED*/ 482} 483 484/* 485 * Update statistics after fault resolution. 486 * - maxrss 487 */ 488void 489uvmfault_update_stats(struct uvm_faultinfo *ufi) 490{ 491 struct vm_map *map; 492 struct proc *p; 493 vsize_t res; 494 495 map = ufi->orig_map; 496 497 /* 498 * If this is a nested pmap (eg, a virtual machine pmap managed 499 * by vmm(4) on amd64/i386), don't do any updating, just return. 500 * 501 * pmap_nested() on other archs is #defined to 0, so this is a 502 * no-op. 503 */ 504 if (pmap_nested(map->pmap)) 505 return; 506 507 /* Update the maxrss for the process. */ 508 if (map->flags & VM_MAP_ISVMSPACE) { 509 p = curproc; 510 KASSERT(p != NULL && &p->p_vmspace->vm_map == map); 511 512 res = pmap_resident_count(map->pmap); 513 /* Convert res from pages to kilobytes. */ 514 res <<= (PAGE_SHIFT - 10); 515 516 if (p->p_ru.ru_maxrss < res) 517 p->p_ru.ru_maxrss = res; 518 } 519} 520 521/* 522 * F A U L T - m a i n e n t r y p o i n t 523 */ 524 525/* 526 * uvm_fault: page fault handler 527 * 528 * => called from MD code to resolve a page fault 529 * => VM data structures usually should be unlocked. however, it is 530 * possible to call here with the main map locked if the caller 531 * gets a write lock, sets it recursive, and then calls us (c.f. 532 * uvm_map_pageable). this should be avoided because it keeps 533 * the map locked off during I/O. 534 * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT 535 */ 536#define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 537 ~PROT_WRITE : PROT_MASK) 538struct uvm_faultctx { 539 /* 540 * the following members are set up by uvm_fault_check() and 541 * read-only after that. 542 */ 543 vm_prot_t enter_prot; 544 vm_prot_t access_type; 545 vaddr_t startva; 546 int npages; 547 int centeridx; 548 boolean_t narrow; 549 boolean_t wired; 550 paddr_t pa_flags; 551}; 552 553int uvm_fault_check( 554 struct uvm_faultinfo *, struct uvm_faultctx *, 555 struct vm_anon ***); 556 557int uvm_fault_upper( 558 struct uvm_faultinfo *, struct uvm_faultctx *, 559 struct vm_anon **, vm_fault_t); 560boolean_t uvm_fault_upper_lookup( 561 struct uvm_faultinfo *, const struct uvm_faultctx *, 562 struct vm_anon **, struct vm_page **); 563 564int uvm_fault_lower( 565 struct uvm_faultinfo *, struct uvm_faultctx *, 566 struct vm_page **, vm_fault_t); 567 568int 569uvm_fault(vm_map_t orig_map, vaddr_t vaddr, vm_fault_t fault_type, 570 vm_prot_t access_type) 571{ 572 struct uvm_faultinfo ufi; 573 struct uvm_faultctx flt; 574 boolean_t shadowed; 575 struct vm_anon *anons_store[UVM_MAXRANGE], **anons; 576 struct vm_page *pages[UVM_MAXRANGE]; 577 int error; 578 579 counters_inc(uvmexp_counters, faults); 580 TRACEPOINT(uvm, fault, vaddr, fault_type, access_type, NULL); 581 582 /* 583 * init the IN parameters in the ufi 584 */ 585 ufi.orig_map = orig_map; 586 ufi.orig_rvaddr = trunc_page(vaddr); 587 ufi.orig_size = PAGE_SIZE; /* can't get any smaller than this */ 588 if (fault_type == VM_FAULT_WIRE) 589 flt.narrow = TRUE; /* don't look for neighborhood 590 * pages on wire */ 591 else 592 flt.narrow = FALSE; /* normal fault */ 593 flt.access_type = access_type; 594 595 596 error = ERESTART; 597 while (error == ERESTART) { /* ReFault: */ 598 anons = anons_store; 599 600 error = uvm_fault_check(&ufi, &flt, &anons); 601 if (error != 0) 602 continue; 603 604 /* True if there is an anon at the faulting address */ 605 shadowed = uvm_fault_upper_lookup(&ufi, &flt, anons, pages); 606 if (shadowed == TRUE) { 607 /* case 1: fault on an anon in our amap */ 608 error = uvm_fault_upper(&ufi, &flt, anons, fault_type); 609 } else { 610 struct uvm_object *uobj = ufi.entry->object.uvm_obj; 611 612 /* 613 * if the desired page is not shadowed by the amap and 614 * we have a backing object, then we check to see if 615 * the backing object would prefer to handle the fault 616 * itself (rather than letting us do it with the usual 617 * pgo_get hook). the backing object signals this by 618 * providing a pgo_fault routine. 619 */ 620 if (uobj != NULL && uobj->pgops->pgo_fault != NULL) { 621 KERNEL_LOCK(); 622 rw_enter(uobj->vmobjlock, RW_WRITE); 623 error = uobj->pgops->pgo_fault(&ufi, 624 flt.startva, pages, flt.npages, 625 flt.centeridx, fault_type, flt.access_type, 626 PGO_LOCKED); 627 KERNEL_UNLOCK(); 628 629 if (error == VM_PAGER_OK) 630 error = 0; 631 else if (error == VM_PAGER_REFAULT) 632 error = ERESTART; 633 else 634 error = EACCES; 635 } else { 636 /* case 2: fault on backing obj or zero fill */ 637 error = uvm_fault_lower(&ufi, &flt, pages, 638 fault_type); 639 } 640 } 641 } 642 643 return error; 644} 645 646/* 647 * uvm_fault_check: check prot, handle needs-copy, etc. 648 * 649 * 1. lookup entry. 650 * 2. check protection. 651 * 3. adjust fault condition (mainly for simulated fault). 652 * 4. handle needs-copy (lazy amap copy). 653 * 5. establish range of interest for neighbor fault (aka pre-fault). 654 * 6. look up anons (if amap exists). 655 * 7. flush pages (if MADV_SEQUENTIAL) 656 * 657 * => called with nothing locked. 658 * => if we fail (result != 0) we unlock everything. 659 * => initialize/adjust many members of flt. 660 */ 661int 662uvm_fault_check(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 663 struct vm_anon ***ranons) 664{ 665 struct vm_amap *amap; 666 struct uvm_object *uobj; 667 int nback, nforw; 668 669 /* 670 * lookup and lock the maps 671 */ 672 if (uvmfault_lookup(ufi, FALSE) == FALSE) { 673 return EFAULT; 674 } 675 /* locked: maps(read) */ 676 677#ifdef DIAGNOSTIC 678 if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) 679 panic("uvm_fault: fault on non-pageable map (%p, 0x%lx)", 680 ufi->map, ufi->orig_rvaddr); 681#endif 682 683 /* 684 * check protection 685 */ 686 if ((ufi->entry->protection & flt->access_type) != flt->access_type) { 687 uvmfault_unlockmaps(ufi, FALSE); 688 return EACCES; 689 } 690 691 /* 692 * "enter_prot" is the protection we want to enter the page in at. 693 * for certain pages (e.g. copy-on-write pages) this protection can 694 * be more strict than ufi->entry->protection. "wired" means either 695 * the entry is wired or we are fault-wiring the pg. 696 */ 697 698 flt->enter_prot = ufi->entry->protection; 699 flt->pa_flags = UVM_ET_ISWC(ufi->entry) ? PMAP_WC : 0; 700 flt->wired = VM_MAPENT_ISWIRED(ufi->entry) || (flt->narrow == TRUE); 701 if (flt->wired) 702 flt->access_type = flt->enter_prot; /* full access for wired */ 703 704 /* handle "needs_copy" case. */ 705 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) { 706 if ((flt->access_type & PROT_WRITE) || 707 (ufi->entry->object.uvm_obj == NULL)) { 708 /* need to clear */ 709 uvmfault_unlockmaps(ufi, FALSE); 710 uvmfault_amapcopy(ufi); 711 counters_inc(uvmexp_counters, flt_amcopy); 712 return ERESTART; 713 } else { 714 /* 715 * ensure that we pmap_enter page R/O since 716 * needs_copy is still true 717 */ 718 flt->enter_prot &= ~PROT_WRITE; 719 } 720 } 721 722 /* 723 * identify the players 724 */ 725 amap = ufi->entry->aref.ar_amap; /* upper layer */ 726 uobj = ufi->entry->object.uvm_obj; /* lower layer */ 727 728 /* 729 * check for a case 0 fault. if nothing backing the entry then 730 * error now. 731 */ 732 if (amap == NULL && uobj == NULL) { 733 uvmfault_unlockmaps(ufi, FALSE); 734 return EFAULT; 735 } 736 737 /* 738 * for a case 2B fault waste no time on adjacent pages because 739 * they are likely already entered. 740 */ 741 if (uobj != NULL && amap != NULL && 742 (flt->access_type & PROT_WRITE) != 0) { 743 /* wide fault (!narrow) */ 744 flt->narrow = TRUE; 745 } 746 747 /* 748 * establish range of interest based on advice from mapper 749 * and then clip to fit map entry. note that we only want 750 * to do this the first time through the fault. if we 751 * ReFault we will disable this by setting "narrow" to true. 752 */ 753 if (flt->narrow == FALSE) { 754 755 /* wide fault (!narrow) */ 756 nback = min(uvmadvice[ufi->entry->advice].nback, 757 (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT); 758 flt->startva = ufi->orig_rvaddr - ((vsize_t)nback << PAGE_SHIFT); 759 nforw = min(uvmadvice[ufi->entry->advice].nforw, 760 ((ufi->entry->end - ufi->orig_rvaddr) >> PAGE_SHIFT) - 1); 761 /* 762 * note: "-1" because we don't want to count the 763 * faulting page as forw 764 */ 765 flt->npages = nback + nforw + 1; 766 flt->centeridx = nback; 767 768 flt->narrow = TRUE; /* ensure only once per-fault */ 769 } else { 770 /* narrow fault! */ 771 nback = nforw = 0; 772 flt->startva = ufi->orig_rvaddr; 773 flt->npages = 1; 774 flt->centeridx = 0; 775 } 776 777 /* 778 * if we've got an amap then lock it and extract current anons. 779 */ 780 if (amap) { 781 amap_lock(amap); 782 amap_lookups(&ufi->entry->aref, 783 flt->startva - ufi->entry->start, *ranons, flt->npages); 784 } else { 785 *ranons = NULL; /* to be safe */ 786 } 787 788 /* 789 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages 790 * now and then forget about them (for the rest of the fault). 791 */ 792 if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) { 793 /* flush back-page anons? */ 794 if (amap) 795 uvmfault_anonflush(*ranons, nback); 796 797 /* 798 * flush object? 799 */ 800 if (uobj) { 801 voff_t uoff; 802 803 uoff = (flt->startva - ufi->entry->start) + ufi->entry->offset; 804 rw_enter(uobj->vmobjlock, RW_WRITE); 805 (void) uobj->pgops->pgo_flush(uobj, uoff, uoff + 806 ((vsize_t)nback << PAGE_SHIFT), PGO_DEACTIVATE); 807 rw_exit(uobj->vmobjlock); 808 } 809 810 /* now forget about the backpages */ 811 if (amap) 812 *ranons += nback; 813 flt->startva += ((vsize_t)nback << PAGE_SHIFT); 814 flt->npages -= nback; 815 flt->centeridx = 0; 816 } 817 818 return 0; 819} 820 821/* 822 * uvm_fault_upper_lookup: look up existing h/w mapping and amap. 823 * 824 * iterate range of interest: 825 * 1. check if h/w mapping exists. if yes, we don't care 826 * 2. check if anon exists. if not, page is lower. 827 * 3. if anon exists, enter h/w mapping for neighbors. 828 * 829 * => called with amap locked (if exists). 830 */ 831boolean_t 832uvm_fault_upper_lookup(struct uvm_faultinfo *ufi, 833 const struct uvm_faultctx *flt, struct vm_anon **anons, 834 struct vm_page **pages) 835{ 836 struct vm_amap *amap = ufi->entry->aref.ar_amap; 837 struct vm_anon *anon; 838 boolean_t shadowed; 839 vaddr_t currva; 840 paddr_t pa; 841 int lcv; 842 843 /* locked: maps(read), amap(if there) */ 844 KASSERT(amap == NULL || 845 rw_write_held(amap->am_lock)); 846 847 /* 848 * map in the backpages and frontpages we found in the amap in hopes 849 * of preventing future faults. we also init the pages[] array as 850 * we go. 851 */ 852 currva = flt->startva; 853 shadowed = FALSE; 854 for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) { 855 /* 856 * dont play with VAs that are already mapped 857 * except for center) 858 */ 859 if (lcv != flt->centeridx && 860 pmap_extract(ufi->orig_map->pmap, currva, &pa)) { 861 pages[lcv] = PGO_DONTCARE; 862 continue; 863 } 864 865 /* 866 * unmapped or center page. check if any anon at this level. 867 */ 868 if (amap == NULL || anons[lcv] == NULL) { 869 pages[lcv] = NULL; 870 continue; 871 } 872 873 /* 874 * check for present page and map if possible. 875 */ 876 pages[lcv] = PGO_DONTCARE; 877 if (lcv == flt->centeridx) { /* save center for later! */ 878 shadowed = TRUE; 879 continue; 880 } 881 anon = anons[lcv]; 882 KASSERT(anon->an_lock == amap->am_lock); 883 if (anon->an_page && 884 (anon->an_page->pg_flags & (PG_RELEASED|PG_BUSY)) == 0) { 885 uvm_lock_pageq(); 886 uvm_pageactivate(anon->an_page); /* reactivate */ 887 uvm_unlock_pageq(); 888 counters_inc(uvmexp_counters, flt_namap); 889 890 /* 891 * Since this isn't the page that's actually faulting, 892 * ignore pmap_enter() failures; it's not critical 893 * that we enter these right now. 894 */ 895 (void) pmap_enter(ufi->orig_map->pmap, currva, 896 VM_PAGE_TO_PHYS(anon->an_page) | flt->pa_flags, 897 (anon->an_ref > 1) ? 898 (flt->enter_prot & ~PROT_WRITE) : flt->enter_prot, 899 PMAP_CANFAIL | 900 (VM_MAPENT_ISWIRED(ufi->entry) ? PMAP_WIRED : 0)); 901 } 902 } 903 if (flt->npages > 1) 904 pmap_update(ufi->orig_map->pmap); 905 906 return shadowed; 907} 908 909/* 910 * uvm_fault_upper: handle upper fault. 911 * 912 * 1. acquire anon lock. 913 * 2. get anon. let uvmfault_anonget do the dirty work. 914 * 3. if COW, promote data to new anon 915 * 4. enter h/w mapping 916 */ 917int 918uvm_fault_upper(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 919 struct vm_anon **anons, vm_fault_t fault_type) 920{ 921 struct vm_amap *amap = ufi->entry->aref.ar_amap; 922 struct vm_anon *oanon, *anon = anons[flt->centeridx]; 923 struct vm_page *pg = NULL; 924 int error, ret; 925 926 /* locked: maps(read), amap, anon */ 927 KASSERT(rw_write_held(amap->am_lock)); 928 KASSERT(anon->an_lock == amap->am_lock); 929 930 /* 931 * no matter if we have case 1A or case 1B we are going to need to 932 * have the anon's memory resident. ensure that now. 933 */ 934 /* 935 * let uvmfault_anonget do the dirty work. 936 * if it fails (!OK) it will unlock everything for us. 937 * if it succeeds, locks are still valid and locked. 938 * also, if it is OK, then the anon's page is on the queues. 939 * if the page is on loan from a uvm_object, then anonget will 940 * lock that object for us if it does not fail. 941 */ 942 error = uvmfault_anonget(ufi, amap, anon); 943 switch (error) { 944 case VM_PAGER_OK: 945 break; 946 947 case VM_PAGER_REFAULT: 948 return ERESTART; 949 950 case VM_PAGER_ERROR: 951 /* 952 * An error occurred while trying to bring in the 953 * page -- this is the only error we return right 954 * now. 955 */ 956 return EACCES; /* XXX */ 957 default: 958#ifdef DIAGNOSTIC 959 panic("uvm_fault: uvmfault_anonget -> %d", error); 960#else 961 return EACCES; 962#endif 963 } 964 965 KASSERT(rw_write_held(amap->am_lock)); 966 KASSERT(anon->an_lock == amap->am_lock); 967 968 /* 969 * if we are case 1B then we will need to allocate a new blank 970 * anon to transfer the data into. note that we have a lock 971 * on anon, so no one can busy or release the page until we are done. 972 * also note that the ref count can't drop to zero here because 973 * it is > 1 and we are only dropping one ref. 974 * 975 * in the (hopefully very rare) case that we are out of RAM we 976 * will unlock, wait for more RAM, and refault. 977 * 978 * if we are out of anon VM we wait for RAM to become available. 979 */ 980 981 if ((flt->access_type & PROT_WRITE) != 0 && anon->an_ref > 1) { 982 counters_inc(uvmexp_counters, flt_acow); 983 oanon = anon; /* oanon = old */ 984 anon = uvm_analloc(); 985 if (anon) { 986 anon->an_lock = amap->am_lock; 987 pg = uvm_pagealloc(NULL, 0, anon, 0); 988 } 989 990 /* check for out of RAM */ 991 if (anon == NULL || pg == NULL) { 992 uvmfault_unlockall(ufi, amap, NULL); 993 if (anon == NULL) 994 counters_inc(uvmexp_counters, flt_noanon); 995 else { 996 anon->an_lock = NULL; 997 anon->an_ref--; 998 uvm_anfree(anon); 999 counters_inc(uvmexp_counters, flt_noram); 1000 } 1001 1002 if (uvm_swapisfull()) 1003 return ENOMEM; 1004 1005 /* out of RAM, wait for more */ 1006 if (anon == NULL) 1007 uvm_anwait(); 1008 else 1009 uvm_wait("flt_noram3"); 1010 return ERESTART; 1011 } 1012 1013 /* got all resources, replace anon with nanon */ 1014 uvm_pagecopy(oanon->an_page, pg); /* pg now !PG_CLEAN */ 1015 /* un-busy! new page */ 1016 atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE); 1017 UVM_PAGE_OWN(pg, NULL); 1018 ret = amap_add(&ufi->entry->aref, 1019 ufi->orig_rvaddr - ufi->entry->start, anon, 1); 1020 KASSERT(ret == 0); 1021 1022 /* deref: can not drop to zero here by defn! */ 1023 oanon->an_ref--; 1024 1025#if defined(MULTIPROCESSOR) && !defined(__HAVE_PMAP_MPSAFE_ENTER_COW) 1026 /* 1027 * If there are multiple threads, either uvm or the 1028 * pmap has to make sure no threads see the old RO 1029 * mapping once any have seen the new RW mapping. 1030 * uvm does it by inserting the new mapping RO and 1031 * letting it fault again. 1032 * This is only a problem on MP systems. 1033 */ 1034 if (P_HASSIBLING(curproc)) { 1035 flt->enter_prot &= ~PROT_WRITE; 1036 flt->access_type &= ~PROT_WRITE; 1037 } 1038#endif 1039 1040 /* 1041 * note: anon is _not_ locked, but we have the sole references 1042 * to in from amap. 1043 * thus, no one can get at it until we are done with it. 1044 */ 1045 } else { 1046 counters_inc(uvmexp_counters, flt_anon); 1047 oanon = anon; 1048 pg = anon->an_page; 1049 if (anon->an_ref > 1) /* disallow writes to ref > 1 anons */ 1050 flt->enter_prot = flt->enter_prot & ~PROT_WRITE; 1051 } 1052 1053 /* 1054 * now map the page in . 1055 */ 1056 if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr, 1057 VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot, 1058 flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) { 1059 /* 1060 * No need to undo what we did; we can simply think of 1061 * this as the pmap throwing away the mapping information. 1062 * 1063 * We do, however, have to go through the ReFault path, 1064 * as the map may change while we're asleep. 1065 */ 1066 uvmfault_unlockall(ufi, amap, NULL); 1067 if (uvm_swapisfull()) { 1068 /* XXX instrumentation */ 1069 return ENOMEM; 1070 } 1071 /* XXX instrumentation */ 1072 uvm_wait("flt_pmfail1"); 1073 return ERESTART; 1074 } 1075 1076 /* 1077 * ... update the page queues. 1078 */ 1079 uvm_lock_pageq(); 1080 1081 if (fault_type == VM_FAULT_WIRE) { 1082 uvm_pagewire(pg); 1083 /* 1084 * since the now-wired page cannot be paged out, 1085 * release its swap resources for others to use. 1086 * since an anon with no swap cannot be PG_CLEAN, 1087 * clear its clean flag now. 1088 */ 1089 atomic_clearbits_int(&pg->pg_flags, PG_CLEAN); 1090 uvm_anon_dropswap(anon); 1091 } else { 1092 /* activate it */ 1093 uvm_pageactivate(pg); 1094 } 1095 1096 uvm_unlock_pageq(); 1097 1098 /* 1099 * done case 1! finish up by unlocking everything and returning success 1100 */ 1101 uvmfault_unlockall(ufi, amap, NULL); 1102 pmap_update(ufi->orig_map->pmap); 1103 return 0; 1104} 1105 1106/* 1107 * uvm_fault_lower_lookup: look up on-memory uobj pages. 1108 * 1109 * 1. get on-memory pages. 1110 * 2. if failed, give up (get only center page later). 1111 * 3. if succeeded, enter h/w mapping of neighbor pages. 1112 */ 1113 1114struct vm_page * 1115uvm_fault_lower_lookup( 1116 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 1117 struct vm_page **pages) 1118{ 1119 struct uvm_object *uobj = ufi->entry->object.uvm_obj; 1120 struct vm_page *uobjpage = NULL; 1121 int lcv, gotpages; 1122 vaddr_t currva; 1123 1124 rw_enter(uobj->vmobjlock, RW_WRITE); 1125 1126 counters_inc(uvmexp_counters, flt_lget); 1127 gotpages = flt->npages; 1128 (void) uobj->pgops->pgo_get(uobj, 1129 ufi->entry->offset + (flt->startva - ufi->entry->start), 1130 pages, &gotpages, flt->centeridx, 1131 flt->access_type & MASK(ufi->entry), ufi->entry->advice, 1132 PGO_LOCKED); 1133 1134 /* 1135 * check for pages to map, if we got any 1136 */ 1137 if (gotpages == 0) { 1138 return NULL; 1139 } 1140 1141 currva = flt->startva; 1142 for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) { 1143 if (pages[lcv] == NULL || 1144 pages[lcv] == PGO_DONTCARE) 1145 continue; 1146 1147 KASSERT((pages[lcv]->pg_flags & PG_RELEASED) == 0); 1148 1149 /* 1150 * if center page is resident and not 1151 * PG_BUSY, then pgo_get made it PG_BUSY 1152 * for us and gave us a handle to it. 1153 * remember this page as "uobjpage." 1154 * (for later use). 1155 */ 1156 if (lcv == flt->centeridx) { 1157 uobjpage = pages[lcv]; 1158 continue; 1159 } 1160 1161 /* 1162 * note: calling pgo_get with locked data 1163 * structures returns us pages which are 1164 * neither busy nor released, so we don't 1165 * need to check for this. we can just 1166 * directly enter the page (after moving it 1167 * to the head of the active queue [useful?]). 1168 */ 1169 1170 uvm_lock_pageq(); 1171 uvm_pageactivate(pages[lcv]); /* reactivate */ 1172 uvm_unlock_pageq(); 1173 counters_inc(uvmexp_counters, flt_nomap); 1174 1175 /* 1176 * Since this page isn't the page that's 1177 * actually faulting, ignore pmap_enter() 1178 * failures; it's not critical that we 1179 * enter these right now. 1180 */ 1181 (void) pmap_enter(ufi->orig_map->pmap, currva, 1182 VM_PAGE_TO_PHYS(pages[lcv]) | flt->pa_flags, 1183 flt->enter_prot & MASK(ufi->entry), 1184 PMAP_CANFAIL | 1185 (flt->wired ? PMAP_WIRED : 0)); 1186 1187 /* 1188 * NOTE: page can't be PG_WANTED because 1189 * we've held the lock the whole time 1190 * we've had the handle. 1191 */ 1192 atomic_clearbits_int(&pages[lcv]->pg_flags, PG_BUSY); 1193 UVM_PAGE_OWN(pages[lcv], NULL); 1194 } 1195 pmap_update(ufi->orig_map->pmap); 1196 1197 return uobjpage; 1198} 1199 1200/* 1201 * uvm_fault_lower: handle lower fault. 1202 * 1203 */ 1204int 1205uvm_fault_lower(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 1206 struct vm_page **pages, vm_fault_t fault_type) 1207{ 1208 struct vm_amap *amap = ufi->entry->aref.ar_amap; 1209 struct uvm_object *uobj = ufi->entry->object.uvm_obj; 1210 boolean_t promote, locked; 1211 int result; 1212 struct vm_page *uobjpage, *pg = NULL; 1213 struct vm_anon *anon = NULL; 1214 voff_t uoff; 1215 1216 /* 1217 * now, if the desired page is not shadowed by the amap and we have 1218 * a backing object that does not have a special fault routine, then 1219 * we ask (with pgo_get) the object for resident pages that we care 1220 * about and attempt to map them in. we do not let pgo_get block 1221 * (PGO_LOCKED). 1222 */ 1223 if (uobj == NULL) { 1224 /* zero fill; don't care neighbor pages */ 1225 uobjpage = NULL; 1226 } else { 1227 uobjpage = uvm_fault_lower_lookup(ufi, flt, pages); 1228 } 1229 1230 /* 1231 * note that at this point we are done with any front or back pages. 1232 * we are now going to focus on the center page (i.e. the one we've 1233 * faulted on). if we have faulted on the bottom (uobj) 1234 * layer [i.e. case 2] and the page was both present and available, 1235 * then we've got a pointer to it as "uobjpage" and we've already 1236 * made it BUSY. 1237 */ 1238 1239 /* 1240 * locked: 1241 */ 1242 KASSERT(amap == NULL || 1243 rw_write_held(amap->am_lock)); 1244 KASSERT(uobj == NULL || 1245 rw_write_held(uobj->vmobjlock)); 1246 1247 /* 1248 * note that uobjpage can not be PGO_DONTCARE at this point. we now 1249 * set uobjpage to PGO_DONTCARE if we are doing a zero fill. if we 1250 * have a backing object, check and see if we are going to promote 1251 * the data up to an anon during the fault. 1252 */ 1253 if (uobj == NULL) { 1254 uobjpage = PGO_DONTCARE; 1255 promote = TRUE; /* always need anon here */ 1256 } else { 1257 KASSERT(uobjpage != PGO_DONTCARE); 1258 promote = (flt->access_type & PROT_WRITE) && 1259 UVM_ET_ISCOPYONWRITE(ufi->entry); 1260 } 1261 1262 /* 1263 * if uobjpage is not null then we do not need to do I/O to get the 1264 * uobjpage. 1265 * 1266 * if uobjpage is null, then we need to ask the pager to 1267 * get the data for us. once we have the data, we need to reverify 1268 * the state the world. we are currently not holding any resources. 1269 */ 1270 if (uobjpage) { 1271 /* update rusage counters */ 1272 curproc->p_ru.ru_minflt++; 1273 } else { 1274 int gotpages; 1275 1276 /* update rusage counters */ 1277 curproc->p_ru.ru_majflt++; 1278 1279 uvmfault_unlockall(ufi, amap, NULL); 1280 1281 counters_inc(uvmexp_counters, flt_get); 1282 gotpages = 1; 1283 uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset; 1284 result = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages, 1285 0, flt->access_type & MASK(ufi->entry), ufi->entry->advice, 1286 PGO_SYNCIO); 1287 1288 /* 1289 * recover from I/O 1290 */ 1291 if (result != VM_PAGER_OK) { 1292 KASSERT(result != VM_PAGER_PEND); 1293 1294 if (result == VM_PAGER_AGAIN) { 1295 tsleep_nsec(&nowake, PVM, "fltagain2", 1296 MSEC_TO_NSEC(5)); 1297 return ERESTART; 1298 } 1299 1300 if (!UVM_ET_ISNOFAULT(ufi->entry)) 1301 return (EIO); 1302 1303 uobjpage = PGO_DONTCARE; 1304 uobj = NULL; 1305 promote = TRUE; 1306 } 1307 1308 /* re-verify the state of the world. */ 1309 locked = uvmfault_relock(ufi); 1310 if (locked && amap != NULL) 1311 amap_lock(amap); 1312 1313 /* might be changed */ 1314 if (uobjpage != PGO_DONTCARE) { 1315 uobj = uobjpage->uobject; 1316 rw_enter(uobj->vmobjlock, RW_WRITE); 1317 } 1318 1319 /* 1320 * Re-verify that amap slot is still free. if there is 1321 * a problem, we clean up. 1322 */ 1323 if (locked && amap && amap_lookup(&ufi->entry->aref, 1324 ufi->orig_rvaddr - ufi->entry->start)) { 1325 if (locked) 1326 uvmfault_unlockall(ufi, amap, NULL); 1327 locked = FALSE; 1328 } 1329 1330 /* didn't get the lock? release the page and retry. */ 1331 if (locked == FALSE && uobjpage != PGO_DONTCARE) { 1332 uvm_lock_pageq(); 1333 /* make sure it is in queues */ 1334 uvm_pageactivate(uobjpage); 1335 uvm_unlock_pageq(); 1336 1337 if (uobjpage->pg_flags & PG_WANTED) 1338 /* still holding object lock */ 1339 wakeup(uobjpage); 1340 atomic_clearbits_int(&uobjpage->pg_flags, 1341 PG_BUSY|PG_WANTED); 1342 UVM_PAGE_OWN(uobjpage, NULL); 1343 } 1344 1345 if (locked == FALSE) { 1346 if (uobjpage != PGO_DONTCARE) 1347 rw_exit(uobj->vmobjlock); 1348 return ERESTART; 1349 } 1350 1351 /* 1352 * we have the data in uobjpage which is PG_BUSY 1353 */ 1354 } 1355 1356 /* 1357 * notes: 1358 * - at this point uobjpage can not be NULL 1359 * - at this point uobjpage could be PG_WANTED (handle later) 1360 */ 1361 if (promote == FALSE) { 1362 /* 1363 * we are not promoting. if the mapping is COW ensure that we 1364 * don't give more access than we should (e.g. when doing a read 1365 * fault on a COPYONWRITE mapping we want to map the COW page in 1366 * R/O even though the entry protection could be R/W). 1367 * 1368 * set "pg" to the page we want to map in (uobjpage, usually) 1369 */ 1370 counters_inc(uvmexp_counters, flt_obj); 1371 if (UVM_ET_ISCOPYONWRITE(ufi->entry)) 1372 flt->enter_prot &= ~PROT_WRITE; 1373 pg = uobjpage; /* map in the actual object */ 1374 1375 /* assert(uobjpage != PGO_DONTCARE) */ 1376 1377 /* 1378 * we are faulting directly on the page. 1379 */ 1380 } else { 1381 /* 1382 * if we are going to promote the data to an anon we 1383 * allocate a blank anon here and plug it into our amap. 1384 */ 1385#ifdef DIAGNOSTIC 1386 if (amap == NULL) 1387 panic("uvm_fault: want to promote data, but no anon"); 1388#endif 1389 1390 anon = uvm_analloc(); 1391 if (anon) { 1392 /* 1393 * In `Fill in data...' below, if 1394 * uobjpage == PGO_DONTCARE, we want 1395 * a zero'd, dirty page, so have 1396 * uvm_pagealloc() do that for us. 1397 */ 1398 anon->an_lock = amap->am_lock; 1399 pg = uvm_pagealloc(NULL, 0, anon, 1400 (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0); 1401 } 1402 1403 /* 1404 * out of memory resources? 1405 */ 1406 if (anon == NULL || pg == NULL) { 1407 /* 1408 * arg! must unbusy our page and fail or sleep. 1409 */ 1410 if (uobjpage != PGO_DONTCARE) { 1411 uvm_lock_pageq(); 1412 uvm_pageactivate(uobjpage); 1413 uvm_unlock_pageq(); 1414 1415 if (uobjpage->pg_flags & PG_WANTED) 1416 wakeup(uobjpage); 1417 atomic_clearbits_int(&uobjpage->pg_flags, 1418 PG_BUSY|PG_WANTED); 1419 UVM_PAGE_OWN(uobjpage, NULL); 1420 } 1421 1422 /* unlock and fail ... */ 1423 uvmfault_unlockall(ufi, amap, uobj); 1424 if (anon == NULL) 1425 counters_inc(uvmexp_counters, flt_noanon); 1426 else { 1427 anon->an_lock = NULL; 1428 anon->an_ref--; 1429 uvm_anfree(anon); 1430 counters_inc(uvmexp_counters, flt_noram); 1431 } 1432 1433 if (uvm_swapisfull()) 1434 return (ENOMEM); 1435 1436 /* out of RAM, wait for more */ 1437 if (anon == NULL) 1438 uvm_anwait(); 1439 else 1440 uvm_wait("flt_noram5"); 1441 return ERESTART; 1442 } 1443 1444 /* 1445 * fill in the data 1446 */ 1447 if (uobjpage != PGO_DONTCARE) { 1448 counters_inc(uvmexp_counters, flt_prcopy); 1449 /* copy page [pg now dirty] */ 1450 uvm_pagecopy(uobjpage, pg); 1451 1452 /* 1453 * promote to shared amap? make sure all sharing 1454 * procs see it 1455 */ 1456 if ((amap_flags(amap) & AMAP_SHARED) != 0) { 1457 pmap_page_protect(uobjpage, PROT_NONE); 1458 } 1459#if defined(MULTIPROCESSOR) && !defined(__HAVE_PMAP_MPSAFE_ENTER_COW) 1460 /* 1461 * Otherwise: 1462 * If there are multiple threads, either uvm or the 1463 * pmap has to make sure no threads see the old RO 1464 * mapping once any have seen the new RW mapping. 1465 * uvm does it here by forcing it to PROT_NONE before 1466 * inserting the new mapping. 1467 */ 1468 else if (P_HASSIBLING(curproc)) { 1469 pmap_page_protect(uobjpage, PROT_NONE); 1470 } 1471#endif 1472 1473 /* dispose of uobjpage. drop handle to uobj as well. */ 1474 if (uobjpage->pg_flags & PG_WANTED) 1475 wakeup(uobjpage); 1476 atomic_clearbits_int(&uobjpage->pg_flags, 1477 PG_BUSY|PG_WANTED); 1478 UVM_PAGE_OWN(uobjpage, NULL); 1479 uvm_lock_pageq(); 1480 uvm_pageactivate(uobjpage); 1481 uvm_unlock_pageq(); 1482 rw_exit(uobj->vmobjlock); 1483 uobj = NULL; 1484 } else { 1485 counters_inc(uvmexp_counters, flt_przero); 1486 /* 1487 * Page is zero'd and marked dirty by uvm_pagealloc() 1488 * above. 1489 */ 1490 } 1491 1492 if (amap_add(&ufi->entry->aref, 1493 ufi->orig_rvaddr - ufi->entry->start, anon, 0)) { 1494 uvmfault_unlockall(ufi, amap, uobj); 1495 uvm_anfree(anon); 1496 counters_inc(uvmexp_counters, flt_noamap); 1497 1498 if (uvm_swapisfull()) 1499 return (ENOMEM); 1500 1501 amap_populate(&ufi->entry->aref, 1502 ufi->orig_rvaddr - ufi->entry->start); 1503 return ERESTART; 1504 } 1505 } 1506 1507 /* note: pg is either the uobjpage or the new page in the new anon */ 1508 /* 1509 * all resources are present. we can now map it in and free our 1510 * resources. 1511 */ 1512 if (amap == NULL) 1513 KASSERT(anon == NULL); 1514 else { 1515 KASSERT(rw_write_held(amap->am_lock)); 1516 KASSERT(anon == NULL || anon->an_lock == amap->am_lock); 1517 } 1518 if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr, 1519 VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot, 1520 flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) { 1521 /* 1522 * No need to undo what we did; we can simply think of 1523 * this as the pmap throwing away the mapping information. 1524 * 1525 * We do, however, have to go through the ReFault path, 1526 * as the map may change while we're asleep. 1527 */ 1528 if (pg->pg_flags & PG_WANTED) 1529 wakeup(pg); 1530 1531 atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED); 1532 UVM_PAGE_OWN(pg, NULL); 1533 uvmfault_unlockall(ufi, amap, uobj); 1534 if (uvm_swapisfull()) { 1535 /* XXX instrumentation */ 1536 return (ENOMEM); 1537 } 1538 /* XXX instrumentation */ 1539 uvm_wait("flt_pmfail2"); 1540 return ERESTART; 1541 } 1542 1543 if (fault_type == VM_FAULT_WIRE) { 1544 uvm_lock_pageq(); 1545 uvm_pagewire(pg); 1546 uvm_unlock_pageq(); 1547 if (pg->pg_flags & PQ_AOBJ) { 1548 /* 1549 * since the now-wired page cannot be paged out, 1550 * release its swap resources for others to use. 1551 * since an aobj page with no swap cannot be clean, 1552 * mark it dirty now. 1553 * 1554 * use pg->uobject here. if the page is from a 1555 * tmpfs vnode, the pages are backed by its UAO and 1556 * not the vnode. 1557 */ 1558 KASSERT(uobj != NULL); 1559 KASSERT(uobj->vmobjlock == pg->uobject->vmobjlock); 1560 atomic_clearbits_int(&pg->pg_flags, PG_CLEAN); 1561 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 1562 } 1563 } else { 1564 /* activate it */ 1565 uvm_lock_pageq(); 1566 uvm_pageactivate(pg); 1567 uvm_unlock_pageq(); 1568 } 1569 1570 if (pg->pg_flags & PG_WANTED) 1571 wakeup(pg); 1572 1573 atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED); 1574 UVM_PAGE_OWN(pg, NULL); 1575 uvmfault_unlockall(ufi, amap, uobj); 1576 pmap_update(ufi->orig_map->pmap); 1577 1578 return (0); 1579} 1580 1581 1582/* 1583 * uvm_fault_wire: wire down a range of virtual addresses in a map. 1584 * 1585 * => map may be read-locked by caller, but MUST NOT be write-locked. 1586 * => if map is read-locked, any operations which may cause map to 1587 * be write-locked in uvm_fault() must be taken care of by 1588 * the caller. See uvm_map_pageable(). 1589 */ 1590int 1591uvm_fault_wire(vm_map_t map, vaddr_t start, vaddr_t end, vm_prot_t access_type) 1592{ 1593 vaddr_t va; 1594 int rv; 1595 1596 /* 1597 * now fault it in a page at a time. if the fault fails then we have 1598 * to undo what we have done. note that in uvm_fault PROT_NONE 1599 * is replaced with the max protection if fault_type is VM_FAULT_WIRE. 1600 */ 1601 for (va = start ; va < end ; va += PAGE_SIZE) { 1602 rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type); 1603 if (rv) { 1604 if (va != start) { 1605 uvm_fault_unwire(map, start, va); 1606 } 1607 return (rv); 1608 } 1609 } 1610 1611 return (0); 1612} 1613 1614/* 1615 * uvm_fault_unwire(): unwire range of virtual space. 1616 */ 1617void 1618uvm_fault_unwire(vm_map_t map, vaddr_t start, vaddr_t end) 1619{ 1620 1621 vm_map_lock_read(map); 1622 uvm_fault_unwire_locked(map, start, end); 1623 vm_map_unlock_read(map); 1624} 1625 1626/* 1627 * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire(). 1628 * 1629 * => map must be at least read-locked. 1630 */ 1631void 1632uvm_fault_unwire_locked(vm_map_t map, vaddr_t start, vaddr_t end) 1633{ 1634 vm_map_entry_t entry, oentry = NULL, next; 1635 pmap_t pmap = vm_map_pmap(map); 1636 vaddr_t va; 1637 paddr_t pa; 1638 struct vm_page *pg; 1639 1640 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 1641 vm_map_assert_anylock(map); 1642 1643 /* 1644 * we assume that the area we are unwiring has actually been wired 1645 * in the first place. this means that we should be able to extract 1646 * the PAs from the pmap. 1647 */ 1648 1649 /* 1650 * find the beginning map entry for the region. 1651 */ 1652 KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map)); 1653 if (uvm_map_lookup_entry(map, start, &entry) == FALSE) 1654 panic("uvm_fault_unwire_locked: address not in map"); 1655 1656 for (va = start; va < end ; va += PAGE_SIZE) { 1657 if (pmap_extract(pmap, va, &pa) == FALSE) 1658 continue; 1659 1660 /* 1661 * find the map entry for the current address. 1662 */ 1663 KASSERT(va >= entry->start); 1664 while (entry && va >= entry->end) { 1665 next = RBT_NEXT(uvm_map_addr, entry); 1666 entry = next; 1667 } 1668 1669 if (entry == NULL) 1670 return; 1671 if (va < entry->start) 1672 continue; 1673 1674 /* 1675 * lock it. 1676 */ 1677 if (entry != oentry) { 1678 if (oentry != NULL) { 1679 uvm_map_unlock_entry(oentry); 1680 } 1681 uvm_map_lock_entry(entry); 1682 oentry = entry; 1683 } 1684 1685 /* 1686 * if the entry is no longer wired, tell the pmap. 1687 */ 1688 if (VM_MAPENT_ISWIRED(entry) == 0) 1689 pmap_unwire(pmap, va); 1690 1691 pg = PHYS_TO_VM_PAGE(pa); 1692 if (pg) { 1693 uvm_lock_pageq(); 1694 uvm_pageunwire(pg); 1695 uvm_unlock_pageq(); 1696 } 1697 } 1698 1699 if (oentry != NULL) { 1700 uvm_map_unlock_entry(oentry); 1701 } 1702} 1703 1704/* 1705 * uvmfault_unlockmaps: unlock the maps 1706 */ 1707void 1708uvmfault_unlockmaps(struct uvm_faultinfo *ufi, boolean_t write_locked) 1709{ 1710 /* 1711 * ufi can be NULL when this isn't really a fault, 1712 * but merely paging in anon data. 1713 */ 1714 if (ufi == NULL) { 1715 return; 1716 } 1717 1718 uvmfault_update_stats(ufi); 1719 if (write_locked) { 1720 vm_map_unlock(ufi->map); 1721 } else { 1722 vm_map_unlock_read(ufi->map); 1723 } 1724} 1725 1726/* 1727 * uvmfault_unlockall: unlock everything passed in. 1728 * 1729 * => maps must be read-locked (not write-locked). 1730 */ 1731void 1732uvmfault_unlockall(struct uvm_faultinfo *ufi, struct vm_amap *amap, 1733 struct uvm_object *uobj) 1734{ 1735 if (uobj) 1736 rw_exit(uobj->vmobjlock); 1737 if (amap != NULL) 1738 amap_unlock(amap); 1739 uvmfault_unlockmaps(ufi, FALSE); 1740} 1741 1742/* 1743 * uvmfault_lookup: lookup a virtual address in a map 1744 * 1745 * => caller must provide a uvm_faultinfo structure with the IN 1746 * params properly filled in 1747 * => we will lookup the map entry (handling submaps) as we go 1748 * => if the lookup is a success we will return with the maps locked 1749 * => if "write_lock" is TRUE, we write_lock the map, otherwise we only 1750 * get a read lock. 1751 * => note that submaps can only appear in the kernel and they are 1752 * required to use the same virtual addresses as the map they 1753 * are referenced by (thus address translation between the main 1754 * map and the submap is unnecessary). 1755 */ 1756 1757boolean_t 1758uvmfault_lookup(struct uvm_faultinfo *ufi, boolean_t write_lock) 1759{ 1760 vm_map_t tmpmap; 1761 1762 /* 1763 * init ufi values for lookup. 1764 */ 1765 ufi->map = ufi->orig_map; 1766 ufi->size = ufi->orig_size; 1767 1768 /* 1769 * keep going down levels until we are done. note that there can 1770 * only be two levels so we won't loop very long. 1771 */ 1772 while (1) { 1773 if (ufi->orig_rvaddr < ufi->map->min_offset || 1774 ufi->orig_rvaddr >= ufi->map->max_offset) 1775 return FALSE; 1776 1777 /* lock map */ 1778 if (write_lock) { 1779 vm_map_lock(ufi->map); 1780 } else { 1781 vm_map_lock_read(ufi->map); 1782 } 1783 1784 /* lookup */ 1785 if (!uvm_map_lookup_entry(ufi->map, ufi->orig_rvaddr, 1786 &ufi->entry)) { 1787 uvmfault_unlockmaps(ufi, write_lock); 1788 return FALSE; 1789 } 1790 1791 /* reduce size if necessary */ 1792 if (ufi->entry->end - ufi->orig_rvaddr < ufi->size) 1793 ufi->size = ufi->entry->end - ufi->orig_rvaddr; 1794 1795 /* 1796 * submap? replace map with the submap and lookup again. 1797 * note: VAs in submaps must match VAs in main map. 1798 */ 1799 if (UVM_ET_ISSUBMAP(ufi->entry)) { 1800 tmpmap = ufi->entry->object.sub_map; 1801 uvmfault_unlockmaps(ufi, write_lock); 1802 ufi->map = tmpmap; 1803 continue; 1804 } 1805 1806 /* 1807 * got it! 1808 */ 1809 ufi->mapv = ufi->map->timestamp; 1810 return TRUE; 1811 1812 } /* while loop */ 1813 1814 /*NOTREACHED*/ 1815} 1816 1817/* 1818 * uvmfault_relock: attempt to relock the same version of the map 1819 * 1820 * => fault data structures should be unlocked before calling. 1821 * => if a success (TRUE) maps will be locked after call. 1822 */ 1823boolean_t 1824uvmfault_relock(struct uvm_faultinfo *ufi) 1825{ 1826 /* 1827 * ufi can be NULL when this isn't really a fault, 1828 * but merely paging in anon data. 1829 */ 1830 if (ufi == NULL) { 1831 return TRUE; 1832 } 1833 1834 counters_inc(uvmexp_counters, flt_relck); 1835 1836 /* 1837 * relock map. fail if version mismatch (in which case nothing 1838 * gets locked). 1839 */ 1840 vm_map_lock_read(ufi->map); 1841 if (ufi->mapv != ufi->map->timestamp) { 1842 vm_map_unlock_read(ufi->map); 1843 return FALSE; 1844 } 1845 1846 counters_inc(uvmexp_counters, flt_relckok); 1847 return TRUE; /* got it! */ 1848} 1849