1 /* $OpenBSD: uvm_fault.c,v 1.129 2022/04/04 09:27:05 kettenis 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 */ 146 struct uvm_advice { 147 int nback; 148 int nforw; 149 }; 150 151 /* 152 * page range array: set up in uvmfault_init(). 153 */ 154 static 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 */ 161 static void uvmfault_amapcopy(struct uvm_faultinfo *); 162 static inline void uvmfault_anonflush(struct vm_anon **, int); 163 void uvmfault_unlockmaps(struct uvm_faultinfo *, boolean_t); 164 void 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 */ 174 static inline void 175 uvmfault_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 */ 202 void 203 uvmfault_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 */ 229 static void 230 uvmfault_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 */ 276 int 277 uvmfault_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 /* un-busy! */ 393 atomic_clearbits_int(&pg->pg_flags, 394 PG_WANTED|PG_BUSY|PG_FAKE); 395 UVM_PAGE_OWN(pg, NULL); 396 397 /* 398 * if we were RELEASED during I/O, then our anon is 399 * no longer part of an amap. we need to free the 400 * anon and try again. 401 */ 402 if (pg->pg_flags & PG_RELEASED) { 403 pmap_page_protect(pg, PROT_NONE); 404 KASSERT(anon->an_ref == 0); 405 /* 406 * Released while we had unlocked amap. 407 */ 408 if (locked) 409 uvmfault_unlockall(ufi, NULL, NULL); 410 uvm_anon_release(anon); /* frees page for us */ 411 counters_inc(uvmexp_counters, flt_pgrele); 412 return (VM_PAGER_REFAULT); /* refault! */ 413 } 414 415 if (error != VM_PAGER_OK) { 416 KASSERT(error != VM_PAGER_PEND); 417 418 /* remove page from anon */ 419 anon->an_page = NULL; 420 421 /* 422 * Remove the swap slot from the anon and 423 * mark the anon as having no real slot. 424 * Do not free the swap slot, thus preventing 425 * it from being used again. 426 */ 427 uvm_swap_markbad(anon->an_swslot, 1); 428 anon->an_swslot = SWSLOT_BAD; 429 430 /* 431 * Note: page was never !PG_BUSY, so it 432 * cannot be mapped and thus no need to 433 * pmap_page_protect() it. 434 */ 435 uvm_lock_pageq(); 436 uvm_pagefree(pg); 437 uvm_unlock_pageq(); 438 439 if (locked) { 440 uvmfault_unlockall(ufi, NULL, NULL); 441 } 442 rw_exit(anon->an_lock); 443 return (VM_PAGER_ERROR); 444 } 445 446 /* 447 * We have successfully read the page, activate it. 448 */ 449 pmap_clear_modify(pg); 450 uvm_lock_pageq(); 451 uvm_pageactivate(pg); 452 uvm_unlock_pageq(); 453 } 454 455 /* 456 * We were not able to re-lock the map - restart the fault. 457 */ 458 if (!locked) { 459 if (we_own) { 460 rw_exit(anon->an_lock); 461 } 462 return (VM_PAGER_REFAULT); 463 } 464 465 /* 466 * Verify that no one has touched the amap and moved 467 * the anon on us. 468 */ 469 if (ufi != NULL && amap_lookup(&ufi->entry->aref, 470 ufi->orig_rvaddr - ufi->entry->start) != anon) { 471 472 uvmfault_unlockall(ufi, amap, NULL); 473 return (VM_PAGER_REFAULT); 474 } 475 476 /* 477 * Retry.. 478 */ 479 counters_inc(uvmexp_counters, flt_anretry); 480 continue; 481 482 } 483 /*NOTREACHED*/ 484 } 485 486 /* 487 * Update statistics after fault resolution. 488 * - maxrss 489 */ 490 void 491 uvmfault_update_stats(struct uvm_faultinfo *ufi) 492 { 493 struct vm_map *map; 494 struct proc *p; 495 vsize_t res; 496 497 map = ufi->orig_map; 498 499 /* 500 * If this is a nested pmap (eg, a virtual machine pmap managed 501 * by vmm(4) on amd64/i386), don't do any updating, just return. 502 * 503 * pmap_nested() on other archs is #defined to 0, so this is a 504 * no-op. 505 */ 506 if (pmap_nested(map->pmap)) 507 return; 508 509 /* Update the maxrss for the process. */ 510 if (map->flags & VM_MAP_ISVMSPACE) { 511 p = curproc; 512 KASSERT(p != NULL && &p->p_vmspace->vm_map == map); 513 514 res = pmap_resident_count(map->pmap); 515 /* Convert res from pages to kilobytes. */ 516 res <<= (PAGE_SHIFT - 10); 517 518 if (p->p_ru.ru_maxrss < res) 519 p->p_ru.ru_maxrss = res; 520 } 521 } 522 523 /* 524 * F A U L T - m a i n e n t r y p o i n t 525 */ 526 527 /* 528 * uvm_fault: page fault handler 529 * 530 * => called from MD code to resolve a page fault 531 * => VM data structures usually should be unlocked. however, it is 532 * possible to call here with the main map locked if the caller 533 * gets a write lock, sets it recursive, and then calls us (c.f. 534 * uvm_map_pageable). this should be avoided because it keeps 535 * the map locked off during I/O. 536 * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT 537 */ 538 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 539 ~PROT_WRITE : PROT_MASK) 540 struct uvm_faultctx { 541 /* 542 * the following members are set up by uvm_fault_check() and 543 * read-only after that. 544 */ 545 vm_prot_t enter_prot; 546 vm_prot_t access_type; 547 vaddr_t startva; 548 int npages; 549 int centeridx; 550 boolean_t narrow; 551 boolean_t wired; 552 paddr_t pa_flags; 553 }; 554 555 int uvm_fault_check( 556 struct uvm_faultinfo *, struct uvm_faultctx *, 557 struct vm_anon ***); 558 559 int uvm_fault_upper( 560 struct uvm_faultinfo *, struct uvm_faultctx *, 561 struct vm_anon **, vm_fault_t); 562 boolean_t uvm_fault_upper_lookup( 563 struct uvm_faultinfo *, const struct uvm_faultctx *, 564 struct vm_anon **, struct vm_page **); 565 566 int uvm_fault_lower( 567 struct uvm_faultinfo *, struct uvm_faultctx *, 568 struct vm_page **, vm_fault_t); 569 570 int 571 uvm_fault(vm_map_t orig_map, vaddr_t vaddr, vm_fault_t fault_type, 572 vm_prot_t access_type) 573 { 574 struct uvm_faultinfo ufi; 575 struct uvm_faultctx flt; 576 boolean_t shadowed; 577 struct vm_anon *anons_store[UVM_MAXRANGE], **anons; 578 struct vm_page *pages[UVM_MAXRANGE]; 579 int error; 580 581 counters_inc(uvmexp_counters, faults); 582 TRACEPOINT(uvm, fault, vaddr, fault_type, access_type, NULL); 583 584 /* 585 * init the IN parameters in the ufi 586 */ 587 ufi.orig_map = orig_map; 588 ufi.orig_rvaddr = trunc_page(vaddr); 589 ufi.orig_size = PAGE_SIZE; /* can't get any smaller than this */ 590 if (fault_type == VM_FAULT_WIRE) 591 flt.narrow = TRUE; /* don't look for neighborhood 592 * pages on wire */ 593 else 594 flt.narrow = FALSE; /* normal fault */ 595 flt.access_type = access_type; 596 597 598 error = ERESTART; 599 while (error == ERESTART) { /* ReFault: */ 600 anons = anons_store; 601 602 error = uvm_fault_check(&ufi, &flt, &anons); 603 if (error != 0) 604 continue; 605 606 /* True if there is an anon at the faulting address */ 607 shadowed = uvm_fault_upper_lookup(&ufi, &flt, anons, pages); 608 if (shadowed == TRUE) { 609 /* case 1: fault on an anon in our amap */ 610 error = uvm_fault_upper(&ufi, &flt, anons, fault_type); 611 } else { 612 struct uvm_object *uobj = ufi.entry->object.uvm_obj; 613 614 /* 615 * if the desired page is not shadowed by the amap and 616 * we have a backing object, then we check to see if 617 * the backing object would prefer to handle the fault 618 * itself (rather than letting us do it with the usual 619 * pgo_get hook). the backing object signals this by 620 * providing a pgo_fault routine. 621 */ 622 if (uobj != NULL && uobj->pgops->pgo_fault != NULL) { 623 KERNEL_LOCK(); 624 rw_enter(uobj->vmobjlock, RW_WRITE); 625 error = uobj->pgops->pgo_fault(&ufi, 626 flt.startva, pages, flt.npages, 627 flt.centeridx, fault_type, flt.access_type, 628 PGO_LOCKED); 629 KERNEL_UNLOCK(); 630 631 if (error == VM_PAGER_OK) 632 error = 0; 633 else if (error == VM_PAGER_REFAULT) 634 error = ERESTART; 635 else 636 error = EACCES; 637 } else { 638 /* case 2: fault on backing obj or zero fill */ 639 error = uvm_fault_lower(&ufi, &flt, pages, 640 fault_type); 641 } 642 } 643 } 644 645 return error; 646 } 647 648 /* 649 * uvm_fault_check: check prot, handle needs-copy, etc. 650 * 651 * 1. lookup entry. 652 * 2. check protection. 653 * 3. adjust fault condition (mainly for simulated fault). 654 * 4. handle needs-copy (lazy amap copy). 655 * 5. establish range of interest for neighbor fault (aka pre-fault). 656 * 6. look up anons (if amap exists). 657 * 7. flush pages (if MADV_SEQUENTIAL) 658 * 659 * => called with nothing locked. 660 * => if we fail (result != 0) we unlock everything. 661 * => initialize/adjust many members of flt. 662 */ 663 int 664 uvm_fault_check(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 665 struct vm_anon ***ranons) 666 { 667 struct vm_amap *amap; 668 struct uvm_object *uobj; 669 int nback, nforw; 670 671 /* 672 * lookup and lock the maps 673 */ 674 if (uvmfault_lookup(ufi, FALSE) == FALSE) { 675 return EFAULT; 676 } 677 /* locked: maps(read) */ 678 679 #ifdef DIAGNOSTIC 680 if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) 681 panic("uvm_fault: fault on non-pageable map (%p, 0x%lx)", 682 ufi->map, ufi->orig_rvaddr); 683 #endif 684 685 /* 686 * check protection 687 */ 688 if ((ufi->entry->protection & flt->access_type) != flt->access_type) { 689 uvmfault_unlockmaps(ufi, FALSE); 690 return EACCES; 691 } 692 693 /* 694 * "enter_prot" is the protection we want to enter the page in at. 695 * for certain pages (e.g. copy-on-write pages) this protection can 696 * be more strict than ufi->entry->protection. "wired" means either 697 * the entry is wired or we are fault-wiring the pg. 698 */ 699 700 flt->enter_prot = ufi->entry->protection; 701 flt->pa_flags = UVM_ET_ISWC(ufi->entry) ? PMAP_WC : 0; 702 flt->wired = VM_MAPENT_ISWIRED(ufi->entry) || (flt->narrow == TRUE); 703 if (flt->wired) 704 flt->access_type = flt->enter_prot; /* full access for wired */ 705 706 /* handle "needs_copy" case. */ 707 if (UVM_ET_ISNEEDSCOPY(ufi->entry)) { 708 if ((flt->access_type & PROT_WRITE) || 709 (ufi->entry->object.uvm_obj == NULL)) { 710 /* need to clear */ 711 uvmfault_unlockmaps(ufi, FALSE); 712 uvmfault_amapcopy(ufi); 713 counters_inc(uvmexp_counters, flt_amcopy); 714 return ERESTART; 715 } else { 716 /* 717 * ensure that we pmap_enter page R/O since 718 * needs_copy is still true 719 */ 720 flt->enter_prot &= ~PROT_WRITE; 721 } 722 } 723 724 /* 725 * identify the players 726 */ 727 amap = ufi->entry->aref.ar_amap; /* upper layer */ 728 uobj = ufi->entry->object.uvm_obj; /* lower layer */ 729 730 /* 731 * check for a case 0 fault. if nothing backing the entry then 732 * error now. 733 */ 734 if (amap == NULL && uobj == NULL) { 735 uvmfault_unlockmaps(ufi, FALSE); 736 return EFAULT; 737 } 738 739 /* 740 * establish range of interest based on advice from mapper 741 * and then clip to fit map entry. note that we only want 742 * to do this the first time through the fault. if we 743 * ReFault we will disable this by setting "narrow" to true. 744 */ 745 if (flt->narrow == FALSE) { 746 747 /* wide fault (!narrow) */ 748 nback = min(uvmadvice[ufi->entry->advice].nback, 749 (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT); 750 flt->startva = ufi->orig_rvaddr - ((vsize_t)nback << PAGE_SHIFT); 751 nforw = min(uvmadvice[ufi->entry->advice].nforw, 752 ((ufi->entry->end - ufi->orig_rvaddr) >> PAGE_SHIFT) - 1); 753 /* 754 * note: "-1" because we don't want to count the 755 * faulting page as forw 756 */ 757 flt->npages = nback + nforw + 1; 758 flt->centeridx = nback; 759 760 flt->narrow = TRUE; /* ensure only once per-fault */ 761 } else { 762 /* narrow fault! */ 763 nback = nforw = 0; 764 flt->startva = ufi->orig_rvaddr; 765 flt->npages = 1; 766 flt->centeridx = 0; 767 } 768 769 /* 770 * if we've got an amap then lock it and extract current anons. 771 */ 772 if (amap) { 773 amap_lock(amap); 774 amap_lookups(&ufi->entry->aref, 775 flt->startva - ufi->entry->start, *ranons, flt->npages); 776 } else { 777 *ranons = NULL; /* to be safe */ 778 } 779 780 /* 781 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages 782 * now and then forget about them (for the rest of the fault). 783 */ 784 if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) { 785 /* flush back-page anons? */ 786 if (amap) 787 uvmfault_anonflush(*ranons, nback); 788 789 /* 790 * flush object? 791 */ 792 if (uobj) { 793 voff_t uoff; 794 795 uoff = (flt->startva - ufi->entry->start) + ufi->entry->offset; 796 rw_enter(uobj->vmobjlock, RW_WRITE); 797 (void) uobj->pgops->pgo_flush(uobj, uoff, uoff + 798 ((vsize_t)nback << PAGE_SHIFT), PGO_DEACTIVATE); 799 rw_exit(uobj->vmobjlock); 800 } 801 802 /* now forget about the backpages */ 803 if (amap) 804 *ranons += nback; 805 flt->startva += ((vsize_t)nback << PAGE_SHIFT); 806 flt->npages -= nback; 807 flt->centeridx = 0; 808 } 809 810 return 0; 811 } 812 813 /* 814 * uvm_fault_upper_lookup: look up existing h/w mapping and amap. 815 * 816 * iterate range of interest: 817 * 1. check if h/w mapping exists. if yes, we don't care 818 * 2. check if anon exists. if not, page is lower. 819 * 3. if anon exists, enter h/w mapping for neighbors. 820 * 821 * => called with amap locked (if exists). 822 */ 823 boolean_t 824 uvm_fault_upper_lookup(struct uvm_faultinfo *ufi, 825 const struct uvm_faultctx *flt, struct vm_anon **anons, 826 struct vm_page **pages) 827 { 828 struct vm_amap *amap = ufi->entry->aref.ar_amap; 829 struct vm_anon *anon; 830 boolean_t shadowed; 831 vaddr_t currva; 832 paddr_t pa; 833 int lcv; 834 835 /* locked: maps(read), amap(if there) */ 836 KASSERT(amap == NULL || 837 rw_write_held(amap->am_lock)); 838 839 /* 840 * map in the backpages and frontpages we found in the amap in hopes 841 * of preventing future faults. we also init the pages[] array as 842 * we go. 843 */ 844 currva = flt->startva; 845 shadowed = FALSE; 846 for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) { 847 /* 848 * dont play with VAs that are already mapped 849 * except for center) 850 */ 851 if (lcv != flt->centeridx && 852 pmap_extract(ufi->orig_map->pmap, currva, &pa)) { 853 pages[lcv] = PGO_DONTCARE; 854 continue; 855 } 856 857 /* 858 * unmapped or center page. check if any anon at this level. 859 */ 860 if (amap == NULL || anons[lcv] == NULL) { 861 pages[lcv] = NULL; 862 continue; 863 } 864 865 /* 866 * check for present page and map if possible. 867 */ 868 pages[lcv] = PGO_DONTCARE; 869 if (lcv == flt->centeridx) { /* save center for later! */ 870 shadowed = TRUE; 871 continue; 872 } 873 anon = anons[lcv]; 874 KASSERT(anon->an_lock == amap->am_lock); 875 if (anon->an_page && 876 (anon->an_page->pg_flags & (PG_RELEASED|PG_BUSY)) == 0) { 877 uvm_lock_pageq(); 878 uvm_pageactivate(anon->an_page); /* reactivate */ 879 uvm_unlock_pageq(); 880 counters_inc(uvmexp_counters, flt_namap); 881 882 /* 883 * Since this isn't the page that's actually faulting, 884 * ignore pmap_enter() failures; it's not critical 885 * that we enter these right now. 886 */ 887 (void) pmap_enter(ufi->orig_map->pmap, currva, 888 VM_PAGE_TO_PHYS(anon->an_page) | flt->pa_flags, 889 (anon->an_ref > 1) ? 890 (flt->enter_prot & ~PROT_WRITE) : flt->enter_prot, 891 PMAP_CANFAIL | 892 (VM_MAPENT_ISWIRED(ufi->entry) ? PMAP_WIRED : 0)); 893 } 894 } 895 if (flt->npages > 1) 896 pmap_update(ufi->orig_map->pmap); 897 898 return shadowed; 899 } 900 901 /* 902 * uvm_fault_upper: handle upper fault. 903 * 904 * 1. acquire anon lock. 905 * 2. get anon. let uvmfault_anonget do the dirty work. 906 * 3. if COW, promote data to new anon 907 * 4. enter h/w mapping 908 */ 909 int 910 uvm_fault_upper(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 911 struct vm_anon **anons, vm_fault_t fault_type) 912 { 913 struct vm_amap *amap = ufi->entry->aref.ar_amap; 914 struct vm_anon *oanon, *anon = anons[flt->centeridx]; 915 struct vm_page *pg = NULL; 916 int error, ret; 917 918 /* locked: maps(read), amap, anon */ 919 KASSERT(rw_write_held(amap->am_lock)); 920 KASSERT(anon->an_lock == amap->am_lock); 921 922 /* 923 * no matter if we have case 1A or case 1B we are going to need to 924 * have the anon's memory resident. ensure that now. 925 */ 926 /* 927 * let uvmfault_anonget do the dirty work. 928 * if it fails (!OK) it will unlock everything for us. 929 * if it succeeds, locks are still valid and locked. 930 * also, if it is OK, then the anon's page is on the queues. 931 * if the page is on loan from a uvm_object, then anonget will 932 * lock that object for us if it does not fail. 933 */ 934 error = uvmfault_anonget(ufi, amap, anon); 935 switch (error) { 936 case VM_PAGER_OK: 937 break; 938 939 case VM_PAGER_REFAULT: 940 return ERESTART; 941 942 case VM_PAGER_ERROR: 943 /* 944 * An error occurred while trying to bring in the 945 * page -- this is the only error we return right 946 * now. 947 */ 948 return EACCES; /* XXX */ 949 default: 950 #ifdef DIAGNOSTIC 951 panic("uvm_fault: uvmfault_anonget -> %d", error); 952 #else 953 return EACCES; 954 #endif 955 } 956 957 KASSERT(rw_write_held(amap->am_lock)); 958 KASSERT(anon->an_lock == amap->am_lock); 959 960 /* 961 * if we are case 1B then we will need to allocate a new blank 962 * anon to transfer the data into. note that we have a lock 963 * on anon, so no one can busy or release the page until we are done. 964 * also note that the ref count can't drop to zero here because 965 * it is > 1 and we are only dropping one ref. 966 * 967 * in the (hopefully very rare) case that we are out of RAM we 968 * will unlock, wait for more RAM, and refault. 969 * 970 * if we are out of anon VM we wait for RAM to become available. 971 */ 972 973 if ((flt->access_type & PROT_WRITE) != 0 && anon->an_ref > 1) { 974 counters_inc(uvmexp_counters, flt_acow); 975 oanon = anon; /* oanon = old */ 976 anon = uvm_analloc(); 977 if (anon) { 978 anon->an_lock = amap->am_lock; 979 pg = uvm_pagealloc(NULL, 0, anon, 0); 980 } 981 982 /* check for out of RAM */ 983 if (anon == NULL || pg == NULL) { 984 uvmfault_unlockall(ufi, amap, NULL); 985 if (anon == NULL) 986 counters_inc(uvmexp_counters, flt_noanon); 987 else { 988 anon->an_lock = NULL; 989 anon->an_ref--; 990 uvm_anfree(anon); 991 counters_inc(uvmexp_counters, flt_noram); 992 } 993 994 if (uvm_swapisfull()) 995 return ENOMEM; 996 997 /* out of RAM, wait for more */ 998 if (anon == NULL) 999 uvm_anwait(); 1000 else 1001 uvm_wait("flt_noram3"); 1002 return ERESTART; 1003 } 1004 1005 /* got all resources, replace anon with nanon */ 1006 uvm_pagecopy(oanon->an_page, pg); /* pg now !PG_CLEAN */ 1007 /* un-busy! new page */ 1008 atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE); 1009 UVM_PAGE_OWN(pg, NULL); 1010 ret = amap_add(&ufi->entry->aref, 1011 ufi->orig_rvaddr - ufi->entry->start, anon, 1); 1012 KASSERT(ret == 0); 1013 1014 /* deref: can not drop to zero here by defn! */ 1015 oanon->an_ref--; 1016 1017 #ifndef __HAVE_PMAP_MPSAFE_ENTER_COW 1018 /* 1019 * If there are multiple threads, either uvm or the 1020 * pmap has to make sure no threads see the old RO 1021 * mapping once any have seen the new RW mapping. 1022 * uvm does it by inserting the new mapping RO and 1023 * letting it fault again. 1024 */ 1025 if (P_HASSIBLING(curproc)) { 1026 flt->enter_prot &= ~PROT_WRITE; 1027 flt->access_type &= ~PROT_WRITE; 1028 } 1029 #endif 1030 1031 /* 1032 * note: anon is _not_ locked, but we have the sole references 1033 * to in from amap. 1034 * thus, no one can get at it until we are done with it. 1035 */ 1036 } else { 1037 counters_inc(uvmexp_counters, flt_anon); 1038 oanon = anon; 1039 pg = anon->an_page; 1040 if (anon->an_ref > 1) /* disallow writes to ref > 1 anons */ 1041 flt->enter_prot = flt->enter_prot & ~PROT_WRITE; 1042 } 1043 1044 /* 1045 * now map the page in . 1046 */ 1047 if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr, 1048 VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot, 1049 flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) { 1050 /* 1051 * No need to undo what we did; we can simply think of 1052 * this as the pmap throwing away the mapping information. 1053 * 1054 * We do, however, have to go through the ReFault path, 1055 * as the map may change while we're asleep. 1056 */ 1057 uvmfault_unlockall(ufi, amap, NULL); 1058 if (uvm_swapisfull()) { 1059 /* XXX instrumentation */ 1060 return ENOMEM; 1061 } 1062 /* XXX instrumentation */ 1063 uvm_wait("flt_pmfail1"); 1064 return ERESTART; 1065 } 1066 1067 /* 1068 * ... update the page queues. 1069 */ 1070 uvm_lock_pageq(); 1071 1072 if (fault_type == VM_FAULT_WIRE) { 1073 uvm_pagewire(pg); 1074 /* 1075 * since the now-wired page cannot be paged out, 1076 * release its swap resources for others to use. 1077 * since an anon with no swap cannot be PG_CLEAN, 1078 * clear its clean flag now. 1079 */ 1080 atomic_clearbits_int(&pg->pg_flags, PG_CLEAN); 1081 uvm_anon_dropswap(anon); 1082 } else { 1083 /* activate it */ 1084 uvm_pageactivate(pg); 1085 } 1086 1087 uvm_unlock_pageq(); 1088 1089 /* 1090 * done case 1! finish up by unlocking everything and returning success 1091 */ 1092 uvmfault_unlockall(ufi, amap, NULL); 1093 pmap_update(ufi->orig_map->pmap); 1094 return 0; 1095 } 1096 1097 /* 1098 * uvm_fault_lower_lookup: look up on-memory uobj pages. 1099 * 1100 * 1. get on-memory pages. 1101 * 2. if failed, give up (get only center page later). 1102 * 3. if succeeded, enter h/w mapping of neighbor pages. 1103 */ 1104 1105 struct vm_page * 1106 uvm_fault_lower_lookup( 1107 struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt, 1108 struct vm_page **pages) 1109 { 1110 struct uvm_object *uobj = ufi->entry->object.uvm_obj; 1111 struct vm_page *uobjpage = NULL; 1112 int lcv, gotpages; 1113 vaddr_t currva; 1114 1115 rw_enter(uobj->vmobjlock, RW_WRITE); 1116 1117 counters_inc(uvmexp_counters, flt_lget); 1118 gotpages = flt->npages; 1119 (void) uobj->pgops->pgo_get(uobj, 1120 ufi->entry->offset + (flt->startva - ufi->entry->start), 1121 pages, &gotpages, flt->centeridx, 1122 flt->access_type & MASK(ufi->entry), ufi->entry->advice, 1123 PGO_LOCKED); 1124 1125 /* 1126 * check for pages to map, if we got any 1127 */ 1128 if (gotpages == 0) { 1129 return NULL; 1130 } 1131 1132 currva = flt->startva; 1133 for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) { 1134 if (pages[lcv] == NULL || 1135 pages[lcv] == PGO_DONTCARE) 1136 continue; 1137 1138 KASSERT((pages[lcv]->pg_flags & PG_RELEASED) == 0); 1139 1140 /* 1141 * if center page is resident and not 1142 * PG_BUSY, then pgo_get made it PG_BUSY 1143 * for us and gave us a handle to it. 1144 * remember this page as "uobjpage." 1145 * (for later use). 1146 */ 1147 if (lcv == flt->centeridx) { 1148 uobjpage = pages[lcv]; 1149 continue; 1150 } 1151 1152 /* 1153 * note: calling pgo_get with locked data 1154 * structures returns us pages which are 1155 * neither busy nor released, so we don't 1156 * need to check for this. we can just 1157 * directly enter the page (after moving it 1158 * to the head of the active queue [useful?]). 1159 */ 1160 1161 uvm_lock_pageq(); 1162 uvm_pageactivate(pages[lcv]); /* reactivate */ 1163 uvm_unlock_pageq(); 1164 counters_inc(uvmexp_counters, flt_nomap); 1165 1166 /* 1167 * Since this page isn't the page that's 1168 * actually faulting, ignore pmap_enter() 1169 * failures; it's not critical that we 1170 * enter these right now. 1171 */ 1172 (void) pmap_enter(ufi->orig_map->pmap, currva, 1173 VM_PAGE_TO_PHYS(pages[lcv]) | flt->pa_flags, 1174 flt->enter_prot & MASK(ufi->entry), 1175 PMAP_CANFAIL | 1176 (flt->wired ? PMAP_WIRED : 0)); 1177 1178 /* 1179 * NOTE: page can't be PG_WANTED because 1180 * we've held the lock the whole time 1181 * we've had the handle. 1182 */ 1183 atomic_clearbits_int(&pages[lcv]->pg_flags, PG_BUSY); 1184 UVM_PAGE_OWN(pages[lcv], NULL); 1185 } 1186 pmap_update(ufi->orig_map->pmap); 1187 1188 return uobjpage; 1189 } 1190 1191 /* 1192 * uvm_fault_lower: handle lower fault. 1193 * 1194 */ 1195 int 1196 uvm_fault_lower(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt, 1197 struct vm_page **pages, vm_fault_t fault_type) 1198 { 1199 struct vm_amap *amap = ufi->entry->aref.ar_amap; 1200 struct uvm_object *uobj = ufi->entry->object.uvm_obj; 1201 boolean_t promote, locked; 1202 int result; 1203 struct vm_page *uobjpage, *pg = NULL; 1204 struct vm_anon *anon = NULL; 1205 voff_t uoff; 1206 1207 /* 1208 * now, if the desired page is not shadowed by the amap and we have 1209 * a backing object that does not have a special fault routine, then 1210 * we ask (with pgo_get) the object for resident pages that we care 1211 * about and attempt to map them in. we do not let pgo_get block 1212 * (PGO_LOCKED). 1213 */ 1214 if (uobj == NULL) { 1215 /* zero fill; don't care neighbor pages */ 1216 uobjpage = NULL; 1217 } else { 1218 uobjpage = uvm_fault_lower_lookup(ufi, flt, pages); 1219 } 1220 1221 /* 1222 * note that at this point we are done with any front or back pages. 1223 * we are now going to focus on the center page (i.e. the one we've 1224 * faulted on). if we have faulted on the bottom (uobj) 1225 * layer [i.e. case 2] and the page was both present and available, 1226 * then we've got a pointer to it as "uobjpage" and we've already 1227 * made it BUSY. 1228 */ 1229 1230 /* 1231 * locked: 1232 */ 1233 KASSERT(amap == NULL || 1234 rw_write_held(amap->am_lock)); 1235 KASSERT(uobj == NULL || 1236 rw_write_held(uobj->vmobjlock)); 1237 1238 /* 1239 * note that uobjpage can not be PGO_DONTCARE at this point. we now 1240 * set uobjpage to PGO_DONTCARE if we are doing a zero fill. if we 1241 * have a backing object, check and see if we are going to promote 1242 * the data up to an anon during the fault. 1243 */ 1244 if (uobj == NULL) { 1245 uobjpage = PGO_DONTCARE; 1246 promote = TRUE; /* always need anon here */ 1247 } else { 1248 KASSERT(uobjpage != PGO_DONTCARE); 1249 promote = (flt->access_type & PROT_WRITE) && 1250 UVM_ET_ISCOPYONWRITE(ufi->entry); 1251 } 1252 1253 /* 1254 * if uobjpage is not null then we do not need to do I/O to get the 1255 * uobjpage. 1256 * 1257 * if uobjpage is null, then we need to ask the pager to 1258 * get the data for us. once we have the data, we need to reverify 1259 * the state the world. we are currently not holding any resources. 1260 */ 1261 if (uobjpage) { 1262 /* update rusage counters */ 1263 curproc->p_ru.ru_minflt++; 1264 } else { 1265 int gotpages; 1266 1267 /* update rusage counters */ 1268 curproc->p_ru.ru_majflt++; 1269 1270 uvmfault_unlockall(ufi, amap, NULL); 1271 1272 counters_inc(uvmexp_counters, flt_get); 1273 gotpages = 1; 1274 uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset; 1275 result = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages, 1276 0, flt->access_type & MASK(ufi->entry), ufi->entry->advice, 1277 PGO_SYNCIO); 1278 1279 /* 1280 * recover from I/O 1281 */ 1282 if (result != VM_PAGER_OK) { 1283 KASSERT(result != VM_PAGER_PEND); 1284 1285 if (result == VM_PAGER_AGAIN) { 1286 tsleep_nsec(&nowake, PVM, "fltagain2", 1287 MSEC_TO_NSEC(5)); 1288 return ERESTART; 1289 } 1290 1291 if (!UVM_ET_ISNOFAULT(ufi->entry)) 1292 return (EIO); 1293 1294 uobjpage = PGO_DONTCARE; 1295 uobj = NULL; 1296 promote = TRUE; 1297 } 1298 1299 /* re-verify the state of the world. */ 1300 locked = uvmfault_relock(ufi); 1301 if (locked && amap != NULL) 1302 amap_lock(amap); 1303 1304 /* might be changed */ 1305 if (uobjpage != PGO_DONTCARE) { 1306 uobj = uobjpage->uobject; 1307 rw_enter(uobj->vmobjlock, RW_WRITE); 1308 } 1309 1310 /* 1311 * Re-verify that amap slot is still free. if there is 1312 * a problem, we clean up. 1313 */ 1314 if (locked && amap && amap_lookup(&ufi->entry->aref, 1315 ufi->orig_rvaddr - ufi->entry->start)) { 1316 if (locked) 1317 uvmfault_unlockall(ufi, amap, NULL); 1318 locked = FALSE; 1319 } 1320 1321 /* didn't get the lock? release the page and retry. */ 1322 if (locked == FALSE && uobjpage != PGO_DONTCARE) { 1323 uvm_lock_pageq(); 1324 /* make sure it is in queues */ 1325 uvm_pageactivate(uobjpage); 1326 uvm_unlock_pageq(); 1327 1328 if (uobjpage->pg_flags & PG_WANTED) 1329 /* still holding object lock */ 1330 wakeup(uobjpage); 1331 atomic_clearbits_int(&uobjpage->pg_flags, 1332 PG_BUSY|PG_WANTED); 1333 UVM_PAGE_OWN(uobjpage, NULL); 1334 } 1335 1336 if (locked == FALSE) { 1337 if (uobjpage != PGO_DONTCARE) 1338 rw_exit(uobj->vmobjlock); 1339 return ERESTART; 1340 } 1341 1342 /* 1343 * we have the data in uobjpage which is PG_BUSY 1344 */ 1345 } 1346 1347 /* 1348 * notes: 1349 * - at this point uobjpage can not be NULL 1350 * - at this point uobjpage could be PG_WANTED (handle later) 1351 */ 1352 if (promote == FALSE) { 1353 /* 1354 * we are not promoting. if the mapping is COW ensure that we 1355 * don't give more access than we should (e.g. when doing a read 1356 * fault on a COPYONWRITE mapping we want to map the COW page in 1357 * R/O even though the entry protection could be R/W). 1358 * 1359 * set "pg" to the page we want to map in (uobjpage, usually) 1360 */ 1361 counters_inc(uvmexp_counters, flt_obj); 1362 if (UVM_ET_ISCOPYONWRITE(ufi->entry)) 1363 flt->enter_prot &= ~PROT_WRITE; 1364 pg = uobjpage; /* map in the actual object */ 1365 1366 /* assert(uobjpage != PGO_DONTCARE) */ 1367 1368 /* 1369 * we are faulting directly on the page. 1370 */ 1371 } else { 1372 /* 1373 * if we are going to promote the data to an anon we 1374 * allocate a blank anon here and plug it into our amap. 1375 */ 1376 #ifdef DIAGNOSTIC 1377 if (amap == NULL) 1378 panic("uvm_fault: want to promote data, but no anon"); 1379 #endif 1380 1381 anon = uvm_analloc(); 1382 if (anon) { 1383 /* 1384 * In `Fill in data...' below, if 1385 * uobjpage == PGO_DONTCARE, we want 1386 * a zero'd, dirty page, so have 1387 * uvm_pagealloc() do that for us. 1388 */ 1389 anon->an_lock = amap->am_lock; 1390 pg = uvm_pagealloc(NULL, 0, anon, 1391 (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0); 1392 } 1393 1394 /* 1395 * out of memory resources? 1396 */ 1397 if (anon == NULL || pg == NULL) { 1398 /* 1399 * arg! must unbusy our page and fail or sleep. 1400 */ 1401 if (uobjpage != PGO_DONTCARE) { 1402 uvm_lock_pageq(); 1403 uvm_pageactivate(uobjpage); 1404 uvm_unlock_pageq(); 1405 1406 if (uobjpage->pg_flags & PG_WANTED) 1407 wakeup(uobjpage); 1408 atomic_clearbits_int(&uobjpage->pg_flags, 1409 PG_BUSY|PG_WANTED); 1410 UVM_PAGE_OWN(uobjpage, NULL); 1411 } 1412 1413 /* unlock and fail ... */ 1414 uvmfault_unlockall(ufi, amap, uobj); 1415 if (anon == NULL) 1416 counters_inc(uvmexp_counters, flt_noanon); 1417 else { 1418 anon->an_lock = NULL; 1419 anon->an_ref--; 1420 uvm_anfree(anon); 1421 counters_inc(uvmexp_counters, flt_noram); 1422 } 1423 1424 if (uvm_swapisfull()) 1425 return (ENOMEM); 1426 1427 /* out of RAM, wait for more */ 1428 if (anon == NULL) 1429 uvm_anwait(); 1430 else 1431 uvm_wait("flt_noram5"); 1432 return ERESTART; 1433 } 1434 1435 /* 1436 * fill in the data 1437 */ 1438 if (uobjpage != PGO_DONTCARE) { 1439 counters_inc(uvmexp_counters, flt_prcopy); 1440 /* copy page [pg now dirty] */ 1441 uvm_pagecopy(uobjpage, pg); 1442 1443 /* 1444 * promote to shared amap? make sure all sharing 1445 * procs see it 1446 */ 1447 if ((amap_flags(amap) & AMAP_SHARED) != 0) { 1448 pmap_page_protect(uobjpage, PROT_NONE); 1449 } 1450 1451 /* dispose of uobjpage. drop handle to uobj as well. */ 1452 if (uobjpage->pg_flags & PG_WANTED) 1453 wakeup(uobjpage); 1454 atomic_clearbits_int(&uobjpage->pg_flags, 1455 PG_BUSY|PG_WANTED); 1456 UVM_PAGE_OWN(uobjpage, NULL); 1457 uvm_lock_pageq(); 1458 uvm_pageactivate(uobjpage); 1459 uvm_unlock_pageq(); 1460 rw_exit(uobj->vmobjlock); 1461 uobj = NULL; 1462 } else { 1463 counters_inc(uvmexp_counters, flt_przero); 1464 /* 1465 * Page is zero'd and marked dirty by uvm_pagealloc() 1466 * above. 1467 */ 1468 } 1469 1470 if (amap_add(&ufi->entry->aref, 1471 ufi->orig_rvaddr - ufi->entry->start, anon, 0)) { 1472 uvmfault_unlockall(ufi, amap, uobj); 1473 uvm_anfree(anon); 1474 counters_inc(uvmexp_counters, flt_noamap); 1475 1476 if (uvm_swapisfull()) 1477 return (ENOMEM); 1478 1479 amap_populate(&ufi->entry->aref, 1480 ufi->orig_rvaddr - ufi->entry->start); 1481 return ERESTART; 1482 } 1483 } 1484 1485 /* note: pg is either the uobjpage or the new page in the new anon */ 1486 /* 1487 * all resources are present. we can now map it in and free our 1488 * resources. 1489 */ 1490 if (amap == NULL) 1491 KASSERT(anon == NULL); 1492 else { 1493 KASSERT(rw_write_held(amap->am_lock)); 1494 KASSERT(anon == NULL || anon->an_lock == amap->am_lock); 1495 } 1496 if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr, 1497 VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot, 1498 flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) { 1499 /* 1500 * No need to undo what we did; we can simply think of 1501 * this as the pmap throwing away the mapping information. 1502 * 1503 * We do, however, have to go through the ReFault path, 1504 * as the map may change while we're asleep. 1505 */ 1506 if (pg->pg_flags & PG_WANTED) 1507 wakeup(pg); 1508 1509 atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED); 1510 UVM_PAGE_OWN(pg, NULL); 1511 uvmfault_unlockall(ufi, amap, uobj); 1512 if (uvm_swapisfull()) { 1513 /* XXX instrumentation */ 1514 return (ENOMEM); 1515 } 1516 /* XXX instrumentation */ 1517 uvm_wait("flt_pmfail2"); 1518 return ERESTART; 1519 } 1520 1521 if (fault_type == VM_FAULT_WIRE) { 1522 uvm_lock_pageq(); 1523 uvm_pagewire(pg); 1524 uvm_unlock_pageq(); 1525 if (pg->pg_flags & PQ_AOBJ) { 1526 /* 1527 * since the now-wired page cannot be paged out, 1528 * release its swap resources for others to use. 1529 * since an aobj page with no swap cannot be clean, 1530 * mark it dirty now. 1531 * 1532 * use pg->uobject here. if the page is from a 1533 * tmpfs vnode, the pages are backed by its UAO and 1534 * not the vnode. 1535 */ 1536 KASSERT(uobj != NULL); 1537 KASSERT(uobj->vmobjlock == pg->uobject->vmobjlock); 1538 atomic_clearbits_int(&pg->pg_flags, PG_CLEAN); 1539 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 1540 } 1541 } else { 1542 /* activate it */ 1543 uvm_lock_pageq(); 1544 uvm_pageactivate(pg); 1545 uvm_unlock_pageq(); 1546 } 1547 1548 if (pg->pg_flags & PG_WANTED) 1549 wakeup(pg); 1550 1551 atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED); 1552 UVM_PAGE_OWN(pg, NULL); 1553 uvmfault_unlockall(ufi, amap, uobj); 1554 pmap_update(ufi->orig_map->pmap); 1555 1556 return (0); 1557 } 1558 1559 1560 /* 1561 * uvm_fault_wire: wire down a range of virtual addresses in a map. 1562 * 1563 * => map may be read-locked by caller, but MUST NOT be write-locked. 1564 * => if map is read-locked, any operations which may cause map to 1565 * be write-locked in uvm_fault() must be taken care of by 1566 * the caller. See uvm_map_pageable(). 1567 */ 1568 int 1569 uvm_fault_wire(vm_map_t map, vaddr_t start, vaddr_t end, vm_prot_t access_type) 1570 { 1571 vaddr_t va; 1572 int rv; 1573 1574 /* 1575 * now fault it in a page at a time. if the fault fails then we have 1576 * to undo what we have done. note that in uvm_fault PROT_NONE 1577 * is replaced with the max protection if fault_type is VM_FAULT_WIRE. 1578 */ 1579 for (va = start ; va < end ; va += PAGE_SIZE) { 1580 rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type); 1581 if (rv) { 1582 if (va != start) { 1583 uvm_fault_unwire(map, start, va); 1584 } 1585 return (rv); 1586 } 1587 } 1588 1589 return (0); 1590 } 1591 1592 /* 1593 * uvm_fault_unwire(): unwire range of virtual space. 1594 */ 1595 void 1596 uvm_fault_unwire(vm_map_t map, vaddr_t start, vaddr_t end) 1597 { 1598 1599 vm_map_lock_read(map); 1600 uvm_fault_unwire_locked(map, start, end); 1601 vm_map_unlock_read(map); 1602 } 1603 1604 /* 1605 * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire(). 1606 * 1607 * => map must be at least read-locked. 1608 */ 1609 void 1610 uvm_fault_unwire_locked(vm_map_t map, vaddr_t start, vaddr_t end) 1611 { 1612 vm_map_entry_t entry, oentry = NULL, next; 1613 pmap_t pmap = vm_map_pmap(map); 1614 vaddr_t va; 1615 paddr_t pa; 1616 struct vm_page *pg; 1617 1618 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 1619 1620 /* 1621 * we assume that the area we are unwiring has actually been wired 1622 * in the first place. this means that we should be able to extract 1623 * the PAs from the pmap. 1624 */ 1625 1626 /* 1627 * find the beginning map entry for the region. 1628 */ 1629 KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map)); 1630 if (uvm_map_lookup_entry(map, start, &entry) == FALSE) 1631 panic("uvm_fault_unwire_locked: address not in map"); 1632 1633 for (va = start; va < end ; va += PAGE_SIZE) { 1634 if (pmap_extract(pmap, va, &pa) == FALSE) 1635 continue; 1636 1637 /* 1638 * find the map entry for the current address. 1639 */ 1640 KASSERT(va >= entry->start); 1641 while (entry && va >= entry->end) { 1642 next = RBT_NEXT(uvm_map_addr, entry); 1643 entry = next; 1644 } 1645 1646 if (entry == NULL) 1647 return; 1648 if (va < entry->start) 1649 continue; 1650 1651 /* 1652 * lock it. 1653 */ 1654 if (entry != oentry) { 1655 if (oentry != NULL) { 1656 uvm_map_unlock_entry(oentry); 1657 } 1658 uvm_map_lock_entry(entry); 1659 oentry = entry; 1660 } 1661 1662 /* 1663 * if the entry is no longer wired, tell the pmap. 1664 */ 1665 if (VM_MAPENT_ISWIRED(entry) == 0) 1666 pmap_unwire(pmap, va); 1667 1668 pg = PHYS_TO_VM_PAGE(pa); 1669 if (pg) { 1670 uvm_lock_pageq(); 1671 uvm_pageunwire(pg); 1672 uvm_unlock_pageq(); 1673 } 1674 } 1675 1676 if (oentry != NULL) { 1677 uvm_map_unlock_entry(oentry); 1678 } 1679 } 1680 1681 /* 1682 * uvmfault_unlockmaps: unlock the maps 1683 */ 1684 void 1685 uvmfault_unlockmaps(struct uvm_faultinfo *ufi, boolean_t write_locked) 1686 { 1687 /* 1688 * ufi can be NULL when this isn't really a fault, 1689 * but merely paging in anon data. 1690 */ 1691 if (ufi == NULL) { 1692 return; 1693 } 1694 1695 uvmfault_update_stats(ufi); 1696 if (write_locked) { 1697 vm_map_unlock(ufi->map); 1698 } else { 1699 vm_map_unlock_read(ufi->map); 1700 } 1701 } 1702 1703 /* 1704 * uvmfault_unlockall: unlock everything passed in. 1705 * 1706 * => maps must be read-locked (not write-locked). 1707 */ 1708 void 1709 uvmfault_unlockall(struct uvm_faultinfo *ufi, struct vm_amap *amap, 1710 struct uvm_object *uobj) 1711 { 1712 if (uobj) 1713 rw_exit(uobj->vmobjlock); 1714 if (amap != NULL) 1715 amap_unlock(amap); 1716 uvmfault_unlockmaps(ufi, FALSE); 1717 } 1718 1719 /* 1720 * uvmfault_lookup: lookup a virtual address in a map 1721 * 1722 * => caller must provide a uvm_faultinfo structure with the IN 1723 * params properly filled in 1724 * => we will lookup the map entry (handling submaps) as we go 1725 * => if the lookup is a success we will return with the maps locked 1726 * => if "write_lock" is TRUE, we write_lock the map, otherwise we only 1727 * get a read lock. 1728 * => note that submaps can only appear in the kernel and they are 1729 * required to use the same virtual addresses as the map they 1730 * are referenced by (thus address translation between the main 1731 * map and the submap is unnecessary). 1732 */ 1733 1734 boolean_t 1735 uvmfault_lookup(struct uvm_faultinfo *ufi, boolean_t write_lock) 1736 { 1737 vm_map_t tmpmap; 1738 1739 /* 1740 * init ufi values for lookup. 1741 */ 1742 ufi->map = ufi->orig_map; 1743 ufi->size = ufi->orig_size; 1744 1745 /* 1746 * keep going down levels until we are done. note that there can 1747 * only be two levels so we won't loop very long. 1748 */ 1749 while (1) { 1750 if (ufi->orig_rvaddr < ufi->map->min_offset || 1751 ufi->orig_rvaddr >= ufi->map->max_offset) 1752 return FALSE; 1753 1754 /* lock map */ 1755 if (write_lock) { 1756 vm_map_lock(ufi->map); 1757 } else { 1758 vm_map_lock_read(ufi->map); 1759 } 1760 1761 /* lookup */ 1762 if (!uvm_map_lookup_entry(ufi->map, ufi->orig_rvaddr, 1763 &ufi->entry)) { 1764 uvmfault_unlockmaps(ufi, write_lock); 1765 return FALSE; 1766 } 1767 1768 /* reduce size if necessary */ 1769 if (ufi->entry->end - ufi->orig_rvaddr < ufi->size) 1770 ufi->size = ufi->entry->end - ufi->orig_rvaddr; 1771 1772 /* 1773 * submap? replace map with the submap and lookup again. 1774 * note: VAs in submaps must match VAs in main map. 1775 */ 1776 if (UVM_ET_ISSUBMAP(ufi->entry)) { 1777 tmpmap = ufi->entry->object.sub_map; 1778 uvmfault_unlockmaps(ufi, write_lock); 1779 ufi->map = tmpmap; 1780 continue; 1781 } 1782 1783 /* 1784 * got it! 1785 */ 1786 ufi->mapv = ufi->map->timestamp; 1787 return TRUE; 1788 1789 } /* while loop */ 1790 1791 /*NOTREACHED*/ 1792 } 1793 1794 /* 1795 * uvmfault_relock: attempt to relock the same version of the map 1796 * 1797 * => fault data structures should be unlocked before calling. 1798 * => if a success (TRUE) maps will be locked after call. 1799 */ 1800 boolean_t 1801 uvmfault_relock(struct uvm_faultinfo *ufi) 1802 { 1803 /* 1804 * ufi can be NULL when this isn't really a fault, 1805 * but merely paging in anon data. 1806 */ 1807 if (ufi == NULL) { 1808 return TRUE; 1809 } 1810 1811 counters_inc(uvmexp_counters, flt_relck); 1812 1813 /* 1814 * relock map. fail if version mismatch (in which case nothing 1815 * gets locked). 1816 */ 1817 vm_map_lock_read(ufi->map); 1818 if (ufi->mapv != ufi->map->timestamp) { 1819 vm_map_unlock_read(ufi->map); 1820 return FALSE; 1821 } 1822 1823 counters_inc(uvmexp_counters, flt_relckok); 1824 return TRUE; /* got it! */ 1825 } 1826