uvm_loan.c revision 1.85
1/* $NetBSD: uvm_loan.c,v 1.85 2017/10/28 00:37:13 pgoyette Exp $ */ 2 3/* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 19 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 21 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 22 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 23 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 24 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 25 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 * 27 * from: Id: uvm_loan.c,v 1.1.6.4 1998/02/06 05:08:43 chs Exp 28 */ 29 30/* 31 * uvm_loan.c: page loanout handler 32 */ 33 34#include <sys/cdefs.h> 35__KERNEL_RCSID(0, "$NetBSD: uvm_loan.c,v 1.85 2017/10/28 00:37:13 pgoyette Exp $"); 36 37#include <sys/param.h> 38#include <sys/systm.h> 39#include <sys/kernel.h> 40#include <sys/mman.h> 41 42#include <uvm/uvm.h> 43 44#ifdef UVMHIST 45UVMHIST_DEFINE(loanhist); 46#endif 47 48/* 49 * "loaned" pages are pages which are (read-only, copy-on-write) loaned 50 * from the VM system to other parts of the kernel. this allows page 51 * copying to be avoided (e.g. you can loan pages from objs/anons to 52 * the mbuf system). 53 * 54 * there are 3 types of loans possible: 55 * O->K uvm_object page to wired kernel page (e.g. mbuf data area) 56 * A->K anon page to wired kernel page (e.g. mbuf data area) 57 * O->A uvm_object to anon loan (e.g. vnode page to an anon) 58 * note that it possible to have an O page loaned to both an A and K 59 * at the same time. 60 * 61 * loans are tracked by pg->loan_count. an O->A page will have both 62 * a uvm_object and a vm_anon, but PQ_ANON will not be set. this sort 63 * of page is considered "owned" by the uvm_object (not the anon). 64 * 65 * each loan of a page to the kernel bumps the pg->wire_count. the 66 * kernel mappings for these pages will be read-only and wired. since 67 * the page will also be wired, it will not be a candidate for pageout, 68 * and thus will never be pmap_page_protect()'d with VM_PROT_NONE. a 69 * write fault in the kernel to one of these pages will not cause 70 * copy-on-write. instead, the page fault is considered fatal. this 71 * is because the kernel mapping will have no way to look up the 72 * object/anon which the page is owned by. this is a good side-effect, 73 * since a kernel write to a loaned page is an error. 74 * 75 * owners that want to free their pages and discover that they are 76 * loaned out simply "disown" them (the page becomes an orphan). these 77 * pages should be freed when the last loan is dropped. in some cases 78 * an anon may "adopt" an orphaned page. 79 * 80 * locking: to read pg->loan_count either the owner or the page queues 81 * must be locked. to modify pg->loan_count, both the owner of the page 82 * and the PQs must be locked. pg->flags is (as always) locked by 83 * the owner of the page. 84 * 85 * note that locking from the "loaned" side is tricky since the object 86 * getting the loaned page has no reference to the page's owner and thus 87 * the owner could "die" at any time. in order to prevent the owner 88 * from dying the page queues should be locked. this forces us to sometimes 89 * use "try" locking. 90 * 91 * loans are typically broken by the following events: 92 * 1. user-level xwrite fault to a loaned page 93 * 2. pageout of clean+inactive O->A loaned page 94 * 3. owner frees page (e.g. pager flush) 95 * 96 * note that loaning a page causes all mappings of the page to become 97 * read-only (via pmap_page_protect). this could have an unexpected 98 * effect on normal "wired" pages if one is not careful (XXX). 99 */ 100 101/* 102 * local prototypes 103 */ 104 105static int uvm_loananon(struct uvm_faultinfo *, void ***, 106 int, struct vm_anon *); 107static int uvm_loanuobj(struct uvm_faultinfo *, void ***, 108 int, vaddr_t); 109static int uvm_loanzero(struct uvm_faultinfo *, void ***, int); 110static void uvm_unloananon(struct vm_anon **, int); 111static void uvm_unloanpage(struct vm_page **, int); 112static int uvm_loanpage(struct vm_page **, int); 113 114 115/* 116 * inlines 117 */ 118 119/* 120 * uvm_loanentry: loan out pages in a map entry (helper fn for uvm_loan()) 121 * 122 * => "ufi" is the result of a successful map lookup (meaning that 123 * on entry the map is locked by the caller) 124 * => we may unlock and then relock the map if needed (for I/O) 125 * => we put our output result in "output" 126 * => we always return with the map unlocked 127 * => possible return values: 128 * -1 == error, map is unlocked 129 * 0 == map relock error (try again!), map is unlocked 130 * >0 == number of pages we loaned, map is unlocked 131 * 132 * NOTE: We can live with this being an inline, because it is only called 133 * from one place. 134 */ 135 136static inline int 137uvm_loanentry(struct uvm_faultinfo *ufi, void ***output, int flags) 138{ 139 vaddr_t curaddr = ufi->orig_rvaddr; 140 vsize_t togo = ufi->size; 141 struct vm_aref *aref = &ufi->entry->aref; 142 struct uvm_object *uobj = ufi->entry->object.uvm_obj; 143 struct vm_anon *anon; 144 int rv, result = 0; 145 146 UVMHIST_FUNC(__func__); UVMHIST_CALLED(loanhist); 147 148 /* 149 * lock us the rest of the way down (we unlock before return) 150 */ 151 if (aref->ar_amap) { 152 amap_lock(aref->ar_amap); 153 } 154 155 /* 156 * loop until done 157 */ 158 while (togo) { 159 160 /* 161 * find the page we want. check the anon layer first. 162 */ 163 164 if (aref->ar_amap) { 165 anon = amap_lookup(aref, curaddr - ufi->entry->start); 166 } else { 167 anon = NULL; 168 } 169 170 /* locked: map, amap, uobj */ 171 if (anon) { 172 rv = uvm_loananon(ufi, output, flags, anon); 173 } else if (uobj) { 174 rv = uvm_loanuobj(ufi, output, flags, curaddr); 175 } else if (UVM_ET_ISCOPYONWRITE(ufi->entry)) { 176 rv = uvm_loanzero(ufi, output, flags); 177 } else { 178 uvmfault_unlockall(ufi, aref->ar_amap, uobj); 179 rv = -1; 180 } 181 /* locked: if (rv > 0) => map, amap, uobj [o.w. unlocked] */ 182 KASSERT(rv > 0 || aref->ar_amap == NULL || 183 !mutex_owned(aref->ar_amap->am_lock)); 184 KASSERT(rv > 0 || uobj == NULL || 185 !mutex_owned(uobj->vmobjlock)); 186 187 /* total failure */ 188 if (rv < 0) { 189 UVMHIST_LOG(loanhist, "failure %jd", rv, 0,0,0); 190 return (-1); 191 } 192 193 /* relock failed, need to do another lookup */ 194 if (rv == 0) { 195 UVMHIST_LOG(loanhist, "relock failure %jd", result 196 ,0,0,0); 197 return (result); 198 } 199 200 /* 201 * got it... advance to next page 202 */ 203 204 result++; 205 togo -= PAGE_SIZE; 206 curaddr += PAGE_SIZE; 207 } 208 209 /* 210 * unlock what we locked, unlock the maps and return 211 */ 212 213 if (aref->ar_amap) { 214 amap_unlock(aref->ar_amap); 215 } 216 uvmfault_unlockmaps(ufi, false); 217 UVMHIST_LOG(loanhist, "done %jd", result, 0,0,0); 218 return (result); 219} 220 221/* 222 * normal functions 223 */ 224 225/* 226 * uvm_loan: loan pages in a map out to anons or to the kernel 227 * 228 * => map should be unlocked 229 * => start and len should be multiples of PAGE_SIZE 230 * => result is either an array of anon's or vm_pages (depending on flags) 231 * => flag values: UVM_LOAN_TOANON - loan to anons 232 * UVM_LOAN_TOPAGE - loan to wired kernel page 233 * one and only one of these flags must be set! 234 * => returns 0 (success), or an appropriate error number 235 */ 236 237int 238uvm_loan(struct vm_map *map, vaddr_t start, vsize_t len, void *v, int flags) 239{ 240 struct uvm_faultinfo ufi; 241 void **result, **output; 242 int rv, error; 243 244 UVMHIST_FUNC(__func__); UVMHIST_CALLED(loanhist); 245 246 /* 247 * ensure that one and only one of the flags is set 248 */ 249 250 KASSERT(((flags & UVM_LOAN_TOANON) == 0) ^ 251 ((flags & UVM_LOAN_TOPAGE) == 0)); 252 253 /* 254 * "output" is a pointer to the current place to put the loaned page. 255 */ 256 257 result = v; 258 output = &result[0]; /* start at the beginning ... */ 259 260 /* 261 * while we've got pages to do 262 */ 263 264 while (len > 0) { 265 266 /* 267 * fill in params for a call to uvmfault_lookup 268 */ 269 270 ufi.orig_map = map; 271 ufi.orig_rvaddr = start; 272 ufi.orig_size = len; 273 274 /* 275 * do the lookup, the only time this will fail is if we hit on 276 * an unmapped region (an error) 277 */ 278 279 if (!uvmfault_lookup(&ufi, false)) { 280 error = ENOENT; 281 goto fail; 282 } 283 284 /* 285 * map now locked. now do the loanout... 286 */ 287 288 rv = uvm_loanentry(&ufi, &output, flags); 289 if (rv < 0) { 290 /* all unlocked due to error */ 291 error = EINVAL; 292 goto fail; 293 } 294 295 /* 296 * done! the map is unlocked. advance, if possible. 297 * 298 * XXXCDC: could be recoded to hold the map lock with 299 * smarter code (but it only happens on map entry 300 * boundaries, so it isn't that bad). 301 */ 302 303 if (rv) { 304 rv <<= PAGE_SHIFT; 305 len -= rv; 306 start += rv; 307 } 308 } 309 UVMHIST_LOG(loanhist, "success", 0,0,0,0); 310 return 0; 311 312fail: 313 /* 314 * failed to complete loans. drop any loans and return failure code. 315 * map is already unlocked. 316 */ 317 318 if (output - result) { 319 if (flags & UVM_LOAN_TOANON) { 320 uvm_unloananon((struct vm_anon **)result, 321 output - result); 322 } else { 323 uvm_unloanpage((struct vm_page **)result, 324 output - result); 325 } 326 } 327 UVMHIST_LOG(loanhist, "error %jd", error,0,0,0); 328 return (error); 329} 330 331/* 332 * uvm_loananon: loan a page from an anon out 333 * 334 * => called with map, amap, uobj locked 335 * => return value: 336 * -1 = fatal error, everything is unlocked, abort. 337 * 0 = lookup in ufi went stale, everything unlocked, relookup and 338 * try again 339 * 1 = got it, everything still locked 340 */ 341 342int 343uvm_loananon(struct uvm_faultinfo *ufi, void ***output, int flags, 344 struct vm_anon *anon) 345{ 346 struct vm_page *pg; 347 int error; 348 349 UVMHIST_FUNC(__func__); UVMHIST_CALLED(loanhist); 350 351 /* 352 * if we are loaning to "another" anon then it is easy, we just 353 * bump the reference count on the current anon and return a 354 * pointer to it (it becomes copy-on-write shared). 355 */ 356 357 if (flags & UVM_LOAN_TOANON) { 358 KASSERT(mutex_owned(anon->an_lock)); 359 pg = anon->an_page; 360 if (pg && (pg->pqflags & PQ_ANON) != 0 && anon->an_ref == 1) { 361 if (pg->wire_count > 0) { 362 UVMHIST_LOG(loanhist, "->A wired %#jx", 363 (uintptr_t)pg, 0, 0, 0); 364 uvmfault_unlockall(ufi, 365 ufi->entry->aref.ar_amap, 366 ufi->entry->object.uvm_obj); 367 return (-1); 368 } 369 pmap_page_protect(pg, VM_PROT_READ); 370 } 371 anon->an_ref++; 372 **output = anon; 373 (*output)++; 374 UVMHIST_LOG(loanhist, "->A done", 0,0,0,0); 375 return (1); 376 } 377 378 /* 379 * we are loaning to a kernel-page. we need to get the page 380 * resident so we can wire it. uvmfault_anonget will handle 381 * this for us. 382 */ 383 384 KASSERT(mutex_owned(anon->an_lock)); 385 error = uvmfault_anonget(ufi, ufi->entry->aref.ar_amap, anon); 386 387 /* 388 * if we were unable to get the anon, then uvmfault_anonget has 389 * unlocked everything and returned an error code. 390 */ 391 392 if (error) { 393 UVMHIST_LOG(loanhist, "error %jd", error,0,0,0); 394 395 /* need to refault (i.e. refresh our lookup) ? */ 396 if (error == ERESTART) { 397 return (0); 398 } 399 400 /* "try again"? sleep a bit and retry ... */ 401 if (error == EAGAIN) { 402 kpause("loanagain", false, hz/2, NULL); 403 return (0); 404 } 405 406 /* otherwise flag it as an error */ 407 return (-1); 408 } 409 410 /* 411 * we have the page and its owner locked: do the loan now. 412 */ 413 414 pg = anon->an_page; 415 mutex_enter(&uvm_pageqlock); 416 if (pg->wire_count > 0) { 417 mutex_exit(&uvm_pageqlock); 418 UVMHIST_LOG(loanhist, "->K wired %#jx", (uintptr_t)pg, 0, 0, 0); 419 KASSERT(pg->uobject == NULL); 420 uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, NULL); 421 return (-1); 422 } 423 if (pg->loan_count == 0) { 424 pmap_page_protect(pg, VM_PROT_READ); 425 } 426 pg->loan_count++; 427 uvm_pageactivate(pg); 428 mutex_exit(&uvm_pageqlock); 429 **output = pg; 430 (*output)++; 431 432 /* unlock and return success */ 433 if (pg->uobject) 434 mutex_exit(pg->uobject->vmobjlock); 435 UVMHIST_LOG(loanhist, "->K done", 0,0,0,0); 436 return (1); 437} 438 439/* 440 * uvm_loanpage: loan out pages to kernel (->K) 441 * 442 * => pages should be object-owned and the object should be locked. 443 * => in the case of error, the object might be unlocked and relocked. 444 * => caller should busy the pages beforehand. 445 * => pages will be unbusied. 446 * => fail with EBUSY if meet a wired page. 447 */ 448static int 449uvm_loanpage(struct vm_page **pgpp, int npages) 450{ 451 int i; 452 int error = 0; 453 454 UVMHIST_FUNC(__func__); UVMHIST_CALLED(loanhist); 455 456 for (i = 0; i < npages; i++) { 457 struct vm_page *pg = pgpp[i]; 458 459 KASSERT(pg->uobject != NULL); 460 KASSERT(pg->uobject == pgpp[0]->uobject); 461 KASSERT(!(pg->flags & (PG_RELEASED|PG_PAGEOUT))); 462 KASSERT(mutex_owned(pg->uobject->vmobjlock)); 463 KASSERT(pg->flags & PG_BUSY); 464 465 mutex_enter(&uvm_pageqlock); 466 if (pg->wire_count > 0) { 467 mutex_exit(&uvm_pageqlock); 468 UVMHIST_LOG(loanhist, "wired %#jx", (uintptr_t)pg, 469 0, 0, 0); 470 error = EBUSY; 471 break; 472 } 473 if (pg->loan_count == 0) { 474 pmap_page_protect(pg, VM_PROT_READ); 475 } 476 pg->loan_count++; 477 uvm_pageactivate(pg); 478 mutex_exit(&uvm_pageqlock); 479 } 480 481 uvm_page_unbusy(pgpp, npages); 482 483 if (error) { 484 /* 485 * backout what we've done 486 */ 487 kmutex_t *slock = pgpp[0]->uobject->vmobjlock; 488 489 mutex_exit(slock); 490 uvm_unloan(pgpp, i, UVM_LOAN_TOPAGE); 491 mutex_enter(slock); 492 } 493 494 UVMHIST_LOG(loanhist, "done %jd", error, 0, 0, 0); 495 return error; 496} 497 498/* 499 * XXX UBC temp limit 500 * number of pages to get at once. 501 * should be <= MAX_READ_AHEAD in genfs_vnops.c 502 */ 503#define UVM_LOAN_GET_CHUNK 16 504 505/* 506 * uvm_loanuobjpages: loan pages from a uobj out (O->K) 507 * 508 * => uobj shouldn't be locked. (we'll lock it) 509 * => fail with EBUSY if we meet a wired page. 510 */ 511int 512uvm_loanuobjpages(struct uvm_object *uobj, voff_t pgoff, int orignpages, 513 struct vm_page **origpgpp) 514{ 515 int ndone; /* # of pages loaned out */ 516 struct vm_page **pgpp; 517 int error; 518 int i; 519 kmutex_t *slock; 520 521 pgpp = origpgpp; 522 for (ndone = 0; ndone < orignpages; ) { 523 int npages; 524 /* npendloan: # of pages busied but not loand out yet. */ 525 int npendloan = 0xdead; /* XXX gcc */ 526reget: 527 npages = MIN(UVM_LOAN_GET_CHUNK, orignpages - ndone); 528 mutex_enter(uobj->vmobjlock); 529 error = (*uobj->pgops->pgo_get)(uobj, 530 pgoff + (ndone << PAGE_SHIFT), pgpp, &npages, 0, 531 VM_PROT_READ, 0, PGO_SYNCIO); 532 if (error == EAGAIN) { 533 kpause("loanuopg", false, hz/2, NULL); 534 continue; 535 } 536 if (error) 537 goto fail; 538 539 KASSERT(npages > 0); 540 541 /* loan and unbusy pages */ 542 slock = NULL; 543 for (i = 0; i < npages; i++) { 544 kmutex_t *nextslock; /* slock for next page */ 545 struct vm_page *pg = *pgpp; 546 547 /* XXX assuming that the page is owned by uobj */ 548 KASSERT(pg->uobject != NULL); 549 nextslock = pg->uobject->vmobjlock; 550 551 if (slock != nextslock) { 552 if (slock) { 553 KASSERT(npendloan > 0); 554 error = uvm_loanpage(pgpp - npendloan, 555 npendloan); 556 mutex_exit(slock); 557 if (error) 558 goto fail; 559 ndone += npendloan; 560 KASSERT(origpgpp + ndone == pgpp); 561 } 562 slock = nextslock; 563 npendloan = 0; 564 mutex_enter(slock); 565 } 566 567 if ((pg->flags & PG_RELEASED) != 0) { 568 /* 569 * release pages and try again. 570 */ 571 mutex_exit(slock); 572 for (; i < npages; i++) { 573 pg = pgpp[i]; 574 slock = pg->uobject->vmobjlock; 575 576 mutex_enter(slock); 577 mutex_enter(&uvm_pageqlock); 578 uvm_page_unbusy(&pg, 1); 579 mutex_exit(&uvm_pageqlock); 580 mutex_exit(slock); 581 } 582 goto reget; 583 } 584 585 npendloan++; 586 pgpp++; 587 KASSERT(origpgpp + ndone + npendloan == pgpp); 588 } 589 KASSERT(slock != NULL); 590 KASSERT(npendloan > 0); 591 error = uvm_loanpage(pgpp - npendloan, npendloan); 592 mutex_exit(slock); 593 if (error) 594 goto fail; 595 ndone += npendloan; 596 KASSERT(origpgpp + ndone == pgpp); 597 } 598 599 return 0; 600 601fail: 602 uvm_unloan(origpgpp, ndone, UVM_LOAN_TOPAGE); 603 604 return error; 605} 606 607/* 608 * uvm_loanuobj: loan a page from a uobj out 609 * 610 * => called with map, amap, uobj locked 611 * => return value: 612 * -1 = fatal error, everything is unlocked, abort. 613 * 0 = lookup in ufi went stale, everything unlocked, relookup and 614 * try again 615 * 1 = got it, everything still locked 616 */ 617 618static int 619uvm_loanuobj(struct uvm_faultinfo *ufi, void ***output, int flags, vaddr_t va) 620{ 621 struct vm_amap *amap = ufi->entry->aref.ar_amap; 622 struct uvm_object *uobj = ufi->entry->object.uvm_obj; 623 struct vm_page *pg; 624 int error, npages; 625 bool locked; 626 627 UVMHIST_FUNC(__func__); UVMHIST_CALLED(loanhist); 628 629 /* 630 * first we must make sure the page is resident. 631 * 632 * XXXCDC: duplicate code with uvm_fault(). 633 */ 634 635 /* locked: maps(read), amap(if there) */ 636 mutex_enter(uobj->vmobjlock); 637 /* locked: maps(read), amap(if there), uobj */ 638 639 if (uobj->pgops->pgo_get) { /* try locked pgo_get */ 640 npages = 1; 641 pg = NULL; 642 error = (*uobj->pgops->pgo_get)(uobj, 643 va - ufi->entry->start + ufi->entry->offset, 644 &pg, &npages, 0, VM_PROT_READ, MADV_NORMAL, PGO_LOCKED); 645 } else { 646 error = EIO; /* must have pgo_get op */ 647 } 648 649 /* 650 * check the result of the locked pgo_get. if there is a problem, 651 * then we fail the loan. 652 */ 653 654 if (error && error != EBUSY) { 655 uvmfault_unlockall(ufi, amap, uobj); 656 return (-1); 657 } 658 659 /* 660 * if we need to unlock for I/O, do so now. 661 */ 662 663 if (error == EBUSY) { 664 uvmfault_unlockall(ufi, amap, NULL); 665 666 /* locked: uobj */ 667 npages = 1; 668 error = (*uobj->pgops->pgo_get)(uobj, 669 va - ufi->entry->start + ufi->entry->offset, 670 &pg, &npages, 0, VM_PROT_READ, MADV_NORMAL, PGO_SYNCIO); 671 /* locked: <nothing> */ 672 673 if (error) { 674 if (error == EAGAIN) { 675 kpause("fltagain2", false, hz/2, NULL); 676 return (0); 677 } 678 return (-1); 679 } 680 681 /* 682 * pgo_get was a success. attempt to relock everything. 683 */ 684 685 locked = uvmfault_relock(ufi); 686 if (locked && amap) 687 amap_lock(amap); 688 uobj = pg->uobject; 689 mutex_enter(uobj->vmobjlock); 690 691 /* 692 * verify that the page has not be released and re-verify 693 * that amap slot is still free. if there is a problem we 694 * drop our lock (thus force a lookup refresh/retry). 695 */ 696 697 if ((pg->flags & PG_RELEASED) != 0 || 698 (locked && amap && amap_lookup(&ufi->entry->aref, 699 ufi->orig_rvaddr - ufi->entry->start))) { 700 if (locked) 701 uvmfault_unlockall(ufi, amap, NULL); 702 locked = false; 703 } 704 705 /* 706 * didn't get the lock? release the page and retry. 707 */ 708 709 if (locked == false) { 710 if (pg->flags & PG_WANTED) { 711 wakeup(pg); 712 } 713 if (pg->flags & PG_RELEASED) { 714 mutex_enter(&uvm_pageqlock); 715 uvm_pagefree(pg); 716 mutex_exit(&uvm_pageqlock); 717 mutex_exit(uobj->vmobjlock); 718 return (0); 719 } 720 mutex_enter(&uvm_pageqlock); 721 uvm_pageactivate(pg); 722 mutex_exit(&uvm_pageqlock); 723 pg->flags &= ~(PG_BUSY|PG_WANTED); 724 UVM_PAGE_OWN(pg, NULL); 725 mutex_exit(uobj->vmobjlock); 726 return (0); 727 } 728 } 729 730 KASSERT(uobj == pg->uobject); 731 732 /* 733 * at this point we have the page we want ("pg") marked PG_BUSY for us 734 * and we have all data structures locked. do the loanout. page can 735 * not be PG_RELEASED (we caught this above). 736 */ 737 738 if ((flags & UVM_LOAN_TOANON) == 0) { 739 if (uvm_loanpage(&pg, 1)) { 740 uvmfault_unlockall(ufi, amap, uobj); 741 return (-1); 742 } 743 mutex_exit(uobj->vmobjlock); 744 **output = pg; 745 (*output)++; 746 return (1); 747 } 748 749#ifdef notdef 750 /* 751 * must be a loan to an anon. check to see if there is already 752 * an anon associated with this page. if so, then just return 753 * a reference to this object. the page should already be 754 * mapped read-only because it is already on loan. 755 */ 756 757 if (pg->uanon) { 758 /* XXX: locking */ 759 anon = pg->uanon; 760 anon->an_ref++; 761 if (pg->flags & PG_WANTED) { 762 wakeup(pg); 763 } 764 pg->flags &= ~(PG_WANTED|PG_BUSY); 765 UVM_PAGE_OWN(pg, NULL); 766 mutex_exit(uobj->vmobjlock); 767 **output = anon; 768 (*output)++; 769 return (1); 770 } 771 772 /* 773 * need to allocate a new anon 774 */ 775 776 anon = uvm_analloc(); 777 if (anon == NULL) { 778 goto fail; 779 } 780 mutex_enter(&uvm_pageqlock); 781 if (pg->wire_count > 0) { 782 mutex_exit(&uvm_pageqlock); 783 UVMHIST_LOG(loanhist, "wired %#jx", (uintptr_t)pg, 0, 0, 0); 784 goto fail; 785 } 786 if (pg->loan_count == 0) { 787 pmap_page_protect(pg, VM_PROT_READ); 788 } 789 pg->loan_count++; 790 pg->uanon = anon; 791 anon->an_page = pg; 792 anon->an_lock = /* TODO: share amap lock */ 793 uvm_pageactivate(pg); 794 mutex_exit(&uvm_pageqlock); 795 if (pg->flags & PG_WANTED) { 796 wakeup(pg); 797 } 798 pg->flags &= ~(PG_WANTED|PG_BUSY); 799 UVM_PAGE_OWN(pg, NULL); 800 mutex_exit(uobj->vmobjlock); 801 mutex_exit(&anon->an_lock); 802 **output = anon; 803 (*output)++; 804 return (1); 805 806fail: 807 UVMHIST_LOG(loanhist, "fail", 0,0,0,0); 808 /* 809 * unlock everything and bail out. 810 */ 811 if (pg->flags & PG_WANTED) { 812 wakeup(pg); 813 } 814 pg->flags &= ~(PG_WANTED|PG_BUSY); 815 UVM_PAGE_OWN(pg, NULL); 816 uvmfault_unlockall(ufi, amap, uobj, NULL); 817 if (anon) { 818 anon->an_ref--; 819 uvm_anon_free(anon); 820 } 821#endif /* notdef */ 822 return (-1); 823} 824 825/* 826 * uvm_loanzero: loan a zero-fill page out 827 * 828 * => called with map, amap, uobj locked 829 * => return value: 830 * -1 = fatal error, everything is unlocked, abort. 831 * 0 = lookup in ufi went stale, everything unlocked, relookup and 832 * try again 833 * 1 = got it, everything still locked 834 */ 835 836static struct uvm_object uvm_loanzero_object; 837static kmutex_t uvm_loanzero_lock; 838 839static int 840uvm_loanzero(struct uvm_faultinfo *ufi, void ***output, int flags) 841{ 842 struct vm_page *pg; 843 struct vm_amap *amap = ufi->entry->aref.ar_amap; 844 845 UVMHIST_FUNC(__func__); UVMHIST_CALLED(loanhist); 846again: 847 mutex_enter(uvm_loanzero_object.vmobjlock); 848 849 /* 850 * first, get ahold of our single zero page. 851 */ 852 853 if (__predict_false((pg = 854 TAILQ_FIRST(&uvm_loanzero_object.memq)) == NULL)) { 855 while ((pg = uvm_pagealloc(&uvm_loanzero_object, 0, NULL, 856 UVM_PGA_ZERO)) == NULL) { 857 mutex_exit(uvm_loanzero_object.vmobjlock); 858 uvmfault_unlockall(ufi, amap, NULL); 859 uvm_wait("loanzero"); 860 if (!uvmfault_relock(ufi)) { 861 return (0); 862 } 863 if (amap) { 864 amap_lock(amap); 865 } 866 goto again; 867 } 868 869 /* got a zero'd page. */ 870 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE); 871 pg->flags |= PG_RDONLY; 872 mutex_enter(&uvm_pageqlock); 873 uvm_pageactivate(pg); 874 mutex_exit(&uvm_pageqlock); 875 UVM_PAGE_OWN(pg, NULL); 876 } 877 878 if ((flags & UVM_LOAN_TOANON) == 0) { /* loaning to kernel-page */ 879 mutex_enter(&uvm_pageqlock); 880 pg->loan_count++; 881 mutex_exit(&uvm_pageqlock); 882 mutex_exit(uvm_loanzero_object.vmobjlock); 883 **output = pg; 884 (*output)++; 885 return (1); 886 } 887 888#ifdef notdef 889 /* 890 * loaning to an anon. check to see if there is already an anon 891 * associated with this page. if so, then just return a reference 892 * to this object. 893 */ 894 895 if (pg->uanon) { 896 anon = pg->uanon; 897 mutex_enter(&anon->an_lock); 898 anon->an_ref++; 899 mutex_exit(&anon->an_lock); 900 mutex_exit(uvm_loanzero_object.vmobjlock); 901 **output = anon; 902 (*output)++; 903 return (1); 904 } 905 906 /* 907 * need to allocate a new anon 908 */ 909 910 anon = uvm_analloc(); 911 if (anon == NULL) { 912 /* out of swap causes us to fail */ 913 mutex_exit(uvm_loanzero_object.vmobjlock); 914 uvmfault_unlockall(ufi, amap, NULL, NULL); 915 return (-1); 916 } 917 anon->an_page = pg; 918 pg->uanon = anon; 919 mutex_enter(&uvm_pageqlock); 920 pg->loan_count++; 921 uvm_pageactivate(pg); 922 mutex_exit(&uvm_pageqlock); 923 mutex_exit(&anon->an_lock); 924 mutex_exit(uvm_loanzero_object.vmobjlock); 925 **output = anon; 926 (*output)++; 927 return (1); 928#else 929 return (-1); 930#endif 931} 932 933 934/* 935 * uvm_unloananon: kill loans on anons (basically a normal ref drop) 936 * 937 * => we expect all our resources to be unlocked 938 */ 939 940static void 941uvm_unloananon(struct vm_anon **aloans, int nanons) 942{ 943#ifdef notdef 944 struct vm_anon *anon, *to_free = NULL; 945 946 /* TODO: locking */ 947 amap_lock(amap); 948 while (nanons-- > 0) { 949 anon = *aloans++; 950 if (--anon->an_ref == 0) { 951 anon->an_link = to_free; 952 to_free = anon; 953 } 954 } 955 uvm_anon_freelst(amap, to_free); 956#endif /* notdef */ 957} 958 959/* 960 * uvm_unloanpage: kill loans on pages loaned out to the kernel 961 * 962 * => we expect all our resources to be unlocked 963 */ 964 965static void 966uvm_unloanpage(struct vm_page **ploans, int npages) 967{ 968 struct vm_page *pg; 969 kmutex_t *slock; 970 971 mutex_enter(&uvm_pageqlock); 972 while (npages-- > 0) { 973 pg = *ploans++; 974 975 /* 976 * do a little dance to acquire the object or anon lock 977 * as appropriate. we are locking in the wrong order, 978 * so we have to do a try-lock here. 979 */ 980 981 slock = NULL; 982 while (pg->uobject != NULL || pg->uanon != NULL) { 983 if (pg->uobject != NULL) { 984 slock = pg->uobject->vmobjlock; 985 } else { 986 slock = pg->uanon->an_lock; 987 } 988 if (mutex_tryenter(slock)) { 989 break; 990 } 991 /* XXX Better than yielding but inadequate. */ 992 kpause("livelock", false, 1, &uvm_pageqlock); 993 slock = NULL; 994 } 995 996 /* 997 * drop our loan. if page is owned by an anon but 998 * PQ_ANON is not set, the page was loaned to the anon 999 * from an object which dropped ownership, so resolve 1000 * this by turning the anon's loan into real ownership 1001 * (ie. decrement loan_count again and set PQ_ANON). 1002 * after all this, if there are no loans left, put the 1003 * page back a paging queue (if the page is owned by 1004 * an anon) or free it (if the page is now unowned). 1005 */ 1006 1007 KASSERT(pg->loan_count > 0); 1008 pg->loan_count--; 1009 if (pg->uobject == NULL && pg->uanon != NULL && 1010 (pg->pqflags & PQ_ANON) == 0) { 1011 KASSERT(pg->loan_count > 0); 1012 pg->loan_count--; 1013 pg->pqflags |= PQ_ANON; 1014 } 1015 if (pg->loan_count == 0 && pg->uobject == NULL && 1016 pg->uanon == NULL) { 1017 KASSERT((pg->flags & PG_BUSY) == 0); 1018 uvm_pagefree(pg); 1019 } 1020 if (slock != NULL) { 1021 mutex_exit(slock); 1022 } 1023 } 1024 mutex_exit(&uvm_pageqlock); 1025} 1026 1027/* 1028 * uvm_unloan: kill loans on pages or anons. 1029 */ 1030 1031void 1032uvm_unloan(void *v, int npages, int flags) 1033{ 1034 if (flags & UVM_LOAN_TOANON) { 1035 uvm_unloananon(v, npages); 1036 } else { 1037 uvm_unloanpage(v, npages); 1038 } 1039} 1040 1041/* 1042 * Minimal pager for uvm_loanzero_object. We need to provide a "put" 1043 * method, because the page can end up on a paging queue, and the 1044 * page daemon will want to call pgo_put when it encounters the page 1045 * on the inactive list. 1046 */ 1047 1048static int 1049ulz_put(struct uvm_object *uobj, voff_t start, voff_t stop, int flags) 1050{ 1051 struct vm_page *pg; 1052 1053 KDASSERT(uobj == &uvm_loanzero_object); 1054 1055 /* 1056 * Don't need to do any work here if we're not freeing pages. 1057 */ 1058 1059 if ((flags & PGO_FREE) == 0) { 1060 mutex_exit(uobj->vmobjlock); 1061 return 0; 1062 } 1063 1064 /* 1065 * we don't actually want to ever free the uvm_loanzero_page, so 1066 * just reactivate or dequeue it. 1067 */ 1068 1069 pg = TAILQ_FIRST(&uobj->memq); 1070 KASSERT(pg != NULL); 1071 KASSERT(TAILQ_NEXT(pg, listq.queue) == NULL); 1072 1073 mutex_enter(&uvm_pageqlock); 1074 if (pg->uanon) 1075 uvm_pageactivate(pg); 1076 else 1077 uvm_pagedequeue(pg); 1078 mutex_exit(&uvm_pageqlock); 1079 1080 mutex_exit(uobj->vmobjlock); 1081 return 0; 1082} 1083 1084static const struct uvm_pagerops ulz_pager = { 1085 .pgo_put = ulz_put, 1086}; 1087 1088/* 1089 * uvm_loan_init(): initialize the uvm_loan() facility. 1090 */ 1091 1092void 1093uvm_loan_init(void) 1094{ 1095 1096 mutex_init(&uvm_loanzero_lock, MUTEX_DEFAULT, IPL_NONE); 1097 uvm_obj_init(&uvm_loanzero_object, &ulz_pager, false, 0); 1098 uvm_obj_setlock(&uvm_loanzero_object, &uvm_loanzero_lock); 1099 1100 UVMHIST_INIT(loanhist, 300); 1101} 1102 1103/* 1104 * uvm_loanbreak: break loan on a uobj page 1105 * 1106 * => called with uobj locked 1107 * => the page should be busy 1108 * => return value: 1109 * newly allocated page if succeeded 1110 */ 1111struct vm_page * 1112uvm_loanbreak(struct vm_page *uobjpage) 1113{ 1114 struct vm_page *pg; 1115 struct uvm_object *uobj __diagused = uobjpage->uobject; 1116 1117 KASSERT(uobj != NULL); 1118 KASSERT(mutex_owned(uobj->vmobjlock)); 1119 KASSERT(uobjpage->flags & PG_BUSY); 1120 1121 /* alloc new un-owned page */ 1122 pg = uvm_pagealloc(NULL, 0, NULL, 0); 1123 if (pg == NULL) 1124 return NULL; 1125 1126 /* 1127 * copy the data from the old page to the new 1128 * one and clear the fake flags on the new page (keep it busy). 1129 * force a reload of the old page by clearing it from all 1130 * pmaps. 1131 * transfer dirtiness of the old page to the new page. 1132 * then lock the page queues to rename the pages. 1133 */ 1134 1135 uvm_pagecopy(uobjpage, pg); /* old -> new */ 1136 pg->flags &= ~PG_FAKE; 1137 pmap_page_protect(uobjpage, VM_PROT_NONE); 1138 if ((uobjpage->flags & PG_CLEAN) != 0 && !pmap_clear_modify(uobjpage)) { 1139 pmap_clear_modify(pg); 1140 pg->flags |= PG_CLEAN; 1141 } else { 1142 /* uvm_pagecopy marked it dirty */ 1143 KASSERT((pg->flags & PG_CLEAN) == 0); 1144 /* a object with a dirty page should be dirty. */ 1145 KASSERT(!UVM_OBJ_IS_CLEAN(uobj)); 1146 } 1147 if (uobjpage->flags & PG_WANTED) 1148 wakeup(uobjpage); 1149 /* uobj still locked */ 1150 uobjpage->flags &= ~(PG_WANTED|PG_BUSY); 1151 UVM_PAGE_OWN(uobjpage, NULL); 1152 1153 mutex_enter(&uvm_pageqlock); 1154 1155 /* 1156 * replace uobjpage with new page. 1157 */ 1158 1159 uvm_pagereplace(uobjpage, pg); 1160 1161 /* 1162 * if the page is no longer referenced by 1163 * an anon (i.e. we are breaking an O->K 1164 * loan), then remove it from any pageq's. 1165 */ 1166 if (uobjpage->uanon == NULL) 1167 uvm_pagedequeue(uobjpage); 1168 1169 /* 1170 * at this point we have absolutely no 1171 * control over uobjpage 1172 */ 1173 1174 /* install new page */ 1175 uvm_pageactivate(pg); 1176 mutex_exit(&uvm_pageqlock); 1177 1178 /* 1179 * done! loan is broken and "pg" is 1180 * PG_BUSY. it can now replace uobjpage. 1181 */ 1182 1183 return pg; 1184} 1185 1186int 1187uvm_loanbreak_anon(struct vm_anon *anon, struct uvm_object *uobj) 1188{ 1189 struct vm_page *pg; 1190 1191 KASSERT(mutex_owned(anon->an_lock)); 1192 KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock)); 1193 1194 /* get new un-owned replacement page */ 1195 pg = uvm_pagealloc(NULL, 0, NULL, 0); 1196 if (pg == NULL) { 1197 return ENOMEM; 1198 } 1199 1200 /* copy old -> new */ 1201 uvm_pagecopy(anon->an_page, pg); 1202 1203 /* force reload */ 1204 pmap_page_protect(anon->an_page, VM_PROT_NONE); 1205 mutex_enter(&uvm_pageqlock); /* KILL loan */ 1206 1207 anon->an_page->uanon = NULL; 1208 /* in case we owned */ 1209 anon->an_page->pqflags &= ~PQ_ANON; 1210 1211 if (uobj) { 1212 /* if we were receiver of loan */ 1213 anon->an_page->loan_count--; 1214 } else { 1215 /* 1216 * we were the lender (A->K); need to remove the page from 1217 * pageq's. 1218 */ 1219 uvm_pagedequeue(anon->an_page); 1220 } 1221 1222 if (uobj) { 1223 mutex_exit(uobj->vmobjlock); 1224 } 1225 1226 /* install new page in anon */ 1227 anon->an_page = pg; 1228 pg->uanon = anon; 1229 pg->pqflags |= PQ_ANON; 1230 1231 uvm_pageactivate(pg); 1232 mutex_exit(&uvm_pageqlock); 1233 1234 pg->flags &= ~(PG_BUSY|PG_FAKE); 1235 UVM_PAGE_OWN(pg, NULL); 1236 1237 /* done! */ 1238 1239 return 0; 1240} 1241