uvm_loan.c revision 1.90
1/* $NetBSD: uvm_loan.c,v 1.90 2019/12/14 15:08:45 ad 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 1.90 2019/12/14 15:08:45 ad 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 PG_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 pg->interlock 81 * must be locked. to modify pg->loan_count, both the owner of the page 82 * and pg->interlock 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 pg->interlock 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->flags & PG_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 if (pg->wire_count > 0) { 416 UVMHIST_LOG(loanhist, "->K wired %#jx", (uintptr_t)pg, 0, 0, 0); 417 KASSERT(pg->uobject == NULL); 418 uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, NULL); 419 return (-1); 420 } 421 if (pg->loan_count == 0) { 422 pmap_page_protect(pg, VM_PROT_READ); 423 } 424 mutex_enter(&pg->interlock); 425 pg->loan_count++; 426 KASSERT(pg->loan_count > 0); /* detect wrap-around */ 427 mutex_exit(&pg->interlock); 428 uvm_pageactivate(pg); 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 if (pg->wire_count > 0) { 466 UVMHIST_LOG(loanhist, "wired %#jx", (uintptr_t)pg, 467 0, 0, 0); 468 error = EBUSY; 469 break; 470 } 471 if (pg->loan_count == 0) { 472 pmap_page_protect(pg, VM_PROT_READ); 473 } 474 mutex_enter(&pg->interlock); 475 pg->loan_count++; 476 KASSERT(pg->loan_count > 0); /* detect wrap-around */ 477 mutex_exit(&pg->interlock); 478 uvm_pageactivate(pg); 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 uvm_page_unbusy(&pg, 1); 578 mutex_exit(slock); 579 } 580 goto reget; 581 } 582 583 npendloan++; 584 pgpp++; 585 KASSERT(origpgpp + ndone + npendloan == pgpp); 586 } 587 KASSERT(slock != NULL); 588 KASSERT(npendloan > 0); 589 error = uvm_loanpage(pgpp - npendloan, npendloan); 590 mutex_exit(slock); 591 if (error) 592 goto fail; 593 ndone += npendloan; 594 KASSERT(origpgpp + ndone == pgpp); 595 } 596 597 return 0; 598 599fail: 600 uvm_unloan(origpgpp, ndone, UVM_LOAN_TOPAGE); 601 602 return error; 603} 604 605/* 606 * uvm_loanuobj: loan a page from a uobj out 607 * 608 * => called with map, amap, uobj locked 609 * => return value: 610 * -1 = fatal error, everything is unlocked, abort. 611 * 0 = lookup in ufi went stale, everything unlocked, relookup and 612 * try again 613 * 1 = got it, everything still locked 614 */ 615 616static int 617uvm_loanuobj(struct uvm_faultinfo *ufi, void ***output, int flags, vaddr_t va) 618{ 619 struct vm_amap *amap = ufi->entry->aref.ar_amap; 620 struct uvm_object *uobj = ufi->entry->object.uvm_obj; 621 struct vm_page *pg; 622 int error, npages; 623 bool locked; 624 625 UVMHIST_FUNC(__func__); UVMHIST_CALLED(loanhist); 626 627 /* 628 * first we must make sure the page is resident. 629 * 630 * XXXCDC: duplicate code with uvm_fault(). 631 */ 632 633 /* locked: maps(read), amap(if there) */ 634 mutex_enter(uobj->vmobjlock); 635 /* locked: maps(read), amap(if there), uobj */ 636 637 if (uobj->pgops->pgo_get) { /* try locked pgo_get */ 638 npages = 1; 639 pg = NULL; 640 error = (*uobj->pgops->pgo_get)(uobj, 641 va - ufi->entry->start + ufi->entry->offset, 642 &pg, &npages, 0, VM_PROT_READ, MADV_NORMAL, PGO_LOCKED); 643 } else { 644 error = EIO; /* must have pgo_get op */ 645 } 646 647 /* 648 * check the result of the locked pgo_get. if there is a problem, 649 * then we fail the loan. 650 */ 651 652 if (error && error != EBUSY) { 653 uvmfault_unlockall(ufi, amap, uobj); 654 return (-1); 655 } 656 657 /* 658 * if we need to unlock for I/O, do so now. 659 */ 660 661 if (error == EBUSY) { 662 uvmfault_unlockall(ufi, amap, NULL); 663 664 /* locked: uobj */ 665 npages = 1; 666 error = (*uobj->pgops->pgo_get)(uobj, 667 va - ufi->entry->start + ufi->entry->offset, 668 &pg, &npages, 0, VM_PROT_READ, MADV_NORMAL, PGO_SYNCIO); 669 /* locked: <nothing> */ 670 671 if (error) { 672 if (error == EAGAIN) { 673 kpause("fltagain2", false, hz/2, NULL); 674 return (0); 675 } 676 return (-1); 677 } 678 679 /* 680 * pgo_get was a success. attempt to relock everything. 681 */ 682 683 locked = uvmfault_relock(ufi); 684 if (locked && amap) 685 amap_lock(amap); 686 uobj = pg->uobject; 687 mutex_enter(uobj->vmobjlock); 688 689 /* 690 * verify that the page has not be released and re-verify 691 * that amap slot is still free. if there is a problem we 692 * drop our lock (thus force a lookup refresh/retry). 693 */ 694 695 if ((pg->flags & PG_RELEASED) != 0 || 696 (locked && amap && amap_lookup(&ufi->entry->aref, 697 ufi->orig_rvaddr - ufi->entry->start))) { 698 if (locked) 699 uvmfault_unlockall(ufi, amap, NULL); 700 locked = false; 701 } 702 703 /* 704 * didn't get the lock? release the page and retry. 705 */ 706 707 if (locked == false) { 708 if (pg->flags & PG_WANTED) { 709 wakeup(pg); 710 } 711 if (pg->flags & PG_RELEASED) { 712 uvm_pagefree(pg); 713 mutex_exit(uobj->vmobjlock); 714 return (0); 715 } 716 uvm_pageactivate(pg); 717 pg->flags &= ~(PG_BUSY|PG_WANTED); 718 UVM_PAGE_OWN(pg, NULL); 719 mutex_exit(uobj->vmobjlock); 720 return (0); 721 } 722 } 723 724 KASSERT(uobj == pg->uobject); 725 726 /* 727 * at this point we have the page we want ("pg") marked PG_BUSY for us 728 * and we have all data structures locked. do the loanout. page can 729 * not be PG_RELEASED (we caught this above). 730 */ 731 732 if ((flags & UVM_LOAN_TOANON) == 0) { 733 if (uvm_loanpage(&pg, 1)) { 734 uvmfault_unlockall(ufi, amap, uobj); 735 return (-1); 736 } 737 mutex_exit(uobj->vmobjlock); 738 **output = pg; 739 (*output)++; 740 return (1); 741 } 742 743#ifdef notdef 744 /* 745 * must be a loan to an anon. check to see if there is already 746 * an anon associated with this page. if so, then just return 747 * a reference to this object. the page should already be 748 * mapped read-only because it is already on loan. 749 */ 750 751 if (pg->uanon) { 752 /* XXX: locking */ 753 anon = pg->uanon; 754 anon->an_ref++; 755 if (pg->flags & PG_WANTED) { 756 wakeup(pg); 757 } 758 pg->flags &= ~(PG_WANTED|PG_BUSY); 759 UVM_PAGE_OWN(pg, NULL); 760 mutex_exit(uobj->vmobjlock); 761 **output = anon; 762 (*output)++; 763 return (1); 764 } 765 766 /* 767 * need to allocate a new anon 768 */ 769 770 anon = uvm_analloc(); 771 if (anon == NULL) { 772 goto fail; 773 } 774 if (pg->wire_count > 0) { 775 UVMHIST_LOG(loanhist, "wired %#jx", (uintptr_t)pg, 0, 0, 0); 776 goto fail; 777 } 778 if (pg->loan_count == 0) { 779 pmap_page_protect(pg, VM_PROT_READ); 780 } 781 mutex_enter(&pg->interlock); 782 pg->loan_count++; 783 KASSERT(pg->loan_count > 0); /* detect wrap-around */ 784 pg->uanon = anon; 785 anon->an_page = pg; 786 anon->an_lock = /* TODO: share amap lock */ 787 mutex_exit(&pg->interlock); 788 uvm_pageactivate(pg); 789 if (pg->flags & PG_WANTED) { 790 wakeup(pg); 791 } 792 pg->flags &= ~(PG_WANTED|PG_BUSY); 793 UVM_PAGE_OWN(pg, NULL); 794 mutex_exit(uobj->vmobjlock); 795 mutex_exit(&anon->an_lock); 796 **output = anon; 797 (*output)++; 798 return (1); 799 800fail: 801 UVMHIST_LOG(loanhist, "fail", 0,0,0,0); 802 /* 803 * unlock everything and bail out. 804 */ 805 if (pg->flags & PG_WANTED) { 806 wakeup(pg); 807 } 808 pg->flags &= ~(PG_WANTED|PG_BUSY); 809 UVM_PAGE_OWN(pg, NULL); 810 uvmfault_unlockall(ufi, amap, uobj, NULL); 811 if (anon) { 812 anon->an_ref--; 813 uvm_anon_free(anon); 814 } 815#endif /* notdef */ 816 return (-1); 817} 818 819/* 820 * uvm_loanzero: loan a zero-fill page out 821 * 822 * => called with map, amap, uobj locked 823 * => return value: 824 * -1 = fatal error, everything is unlocked, abort. 825 * 0 = lookup in ufi went stale, everything unlocked, relookup and 826 * try again 827 * 1 = got it, everything still locked 828 */ 829 830static struct uvm_object uvm_loanzero_object; 831static kmutex_t uvm_loanzero_lock __cacheline_aligned; 832 833static int 834uvm_loanzero(struct uvm_faultinfo *ufi, void ***output, int flags) 835{ 836 struct vm_page *pg; 837 struct vm_amap *amap = ufi->entry->aref.ar_amap; 838 839 UVMHIST_FUNC(__func__); UVMHIST_CALLED(loanhist); 840again: 841 mutex_enter(uvm_loanzero_object.vmobjlock); 842 843 /* 844 * first, get ahold of our single zero page. 845 */ 846 847 if (__predict_false((pg = 848 TAILQ_FIRST(&uvm_loanzero_object.memq)) == NULL)) { 849 while ((pg = uvm_pagealloc(&uvm_loanzero_object, 0, NULL, 850 UVM_PGA_ZERO)) == NULL) { 851 mutex_exit(uvm_loanzero_object.vmobjlock); 852 uvmfault_unlockall(ufi, amap, NULL); 853 uvm_wait("loanzero"); 854 if (!uvmfault_relock(ufi)) { 855 return (0); 856 } 857 if (amap) { 858 amap_lock(amap); 859 } 860 goto again; 861 } 862 863 /* got a zero'd page. */ 864 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE); 865 pg->flags |= PG_RDONLY; 866 uvm_pageactivate(pg); 867 UVM_PAGE_OWN(pg, NULL); 868 } 869 870 if ((flags & UVM_LOAN_TOANON) == 0) { /* loaning to kernel-page */ 871 mutex_enter(&pg->interlock); 872 pg->loan_count++; 873 KASSERT(pg->loan_count > 0); /* detect wrap-around */ 874 mutex_exit(&pg->interlock); 875 mutex_exit(uvm_loanzero_object.vmobjlock); 876 **output = pg; 877 (*output)++; 878 return (1); 879 } 880 881#ifdef notdef 882 /* 883 * loaning to an anon. check to see if there is already an anon 884 * associated with this page. if so, then just return a reference 885 * to this object. 886 */ 887 888 if (pg->uanon) { 889 anon = pg->uanon; 890 mutex_enter(&anon->an_lock); 891 anon->an_ref++; 892 mutex_exit(&anon->an_lock); 893 mutex_exit(uvm_loanzero_object.vmobjlock); 894 **output = anon; 895 (*output)++; 896 return (1); 897 } 898 899 /* 900 * need to allocate a new anon 901 */ 902 903 anon = uvm_analloc(); 904 if (anon == NULL) { 905 /* out of swap causes us to fail */ 906 mutex_exit(uvm_loanzero_object.vmobjlock); 907 uvmfault_unlockall(ufi, amap, NULL, NULL); 908 return (-1); 909 } 910 anon->an_page = pg; 911 pg->uanon = anon; 912 mutex_enter(&pg->interlock); 913 pg->loan_count++; 914 KASSERT(pg->loan_count > 0); /* detect wrap-around */ 915 mutex_exit(&pg->interlock); 916 uvm_pageactivate(pg); 917 mutex_exit(&anon->an_lock); 918 mutex_exit(uvm_loanzero_object.vmobjlock); 919 **output = anon; 920 (*output)++; 921 return (1); 922#else 923 return (-1); 924#endif 925} 926 927 928/* 929 * uvm_unloananon: kill loans on anons (basically a normal ref drop) 930 * 931 * => we expect all our resources to be unlocked 932 */ 933 934static void 935uvm_unloananon(struct vm_anon **aloans, int nanons) 936{ 937#ifdef notdef 938 struct vm_anon *anon, *to_free = NULL; 939 940 /* TODO: locking */ 941 amap_lock(amap); 942 while (nanons-- > 0) { 943 anon = *aloans++; 944 if (--anon->an_ref == 0) { 945 anon->an_link = to_free; 946 to_free = anon; 947 } 948 } 949 uvm_anon_freelst(amap, to_free); 950#endif /* notdef */ 951} 952 953/* 954 * uvm_unloanpage: kill loans on pages loaned out to the kernel 955 * 956 * => we expect all our resources to be unlocked 957 */ 958 959static void 960uvm_unloanpage(struct vm_page **ploans, int npages) 961{ 962 struct vm_page *pg; 963 kmutex_t *slock; 964 965 while (npages-- > 0) { 966 pg = *ploans++; 967 968 /* 969 * do a little dance to acquire the object or anon lock 970 * as appropriate. we are locking in the wrong order, 971 * so we have to do a try-lock here. 972 */ 973 974 mutex_enter(&pg->interlock); 975 slock = NULL; 976 while (pg->uobject != NULL || pg->uanon != NULL) { 977 if (pg->uobject != NULL) { 978 slock = pg->uobject->vmobjlock; 979 } else { 980 slock = pg->uanon->an_lock; 981 } 982 if (mutex_tryenter(slock)) { 983 break; 984 } 985 /* XXX Better than yielding but inadequate. */ 986 kpause("livelock", false, 1, &pg->interlock); 987 slock = NULL; 988 } 989 990 /* 991 * drop our loan. if page is owned by an anon but 992 * PG_ANON is not set, the page was loaned to the anon 993 * from an object which dropped ownership, so resolve 994 * this by turning the anon's loan into real ownership 995 * (ie. decrement loan_count again and set PG_ANON). 996 * after all this, if there are no loans left, put the 997 * page back a paging queue (if the page is owned by 998 * an anon) or free it (if the page is now unowned). 999 */ 1000 1001 KASSERT(pg->loan_count > 0); 1002 pg->loan_count--; 1003 if (pg->uobject == NULL && pg->uanon != NULL && 1004 (pg->flags & PG_ANON) == 0) { 1005 KASSERT(pg->loan_count > 0); 1006 pg->loan_count--; 1007 pg->flags |= PG_ANON; 1008 } 1009 mutex_exit(&pg->interlock); 1010 if (pg->loan_count == 0 && pg->uobject == NULL && 1011 pg->uanon == NULL) { 1012 KASSERT((pg->flags & PG_BUSY) == 0); 1013 uvm_pagefree(pg); 1014 } 1015 if (slock != NULL) { 1016 mutex_exit(slock); 1017 } 1018 } 1019} 1020 1021/* 1022 * uvm_unloan: kill loans on pages or anons. 1023 */ 1024 1025void 1026uvm_unloan(void *v, int npages, int flags) 1027{ 1028 if (flags & UVM_LOAN_TOANON) { 1029 uvm_unloananon(v, npages); 1030 } else { 1031 uvm_unloanpage(v, npages); 1032 } 1033} 1034 1035/* 1036 * Minimal pager for uvm_loanzero_object. We need to provide a "put" 1037 * method, because the page can end up on a paging queue, and the 1038 * page daemon will want to call pgo_put when it encounters the page 1039 * on the inactive list. 1040 */ 1041 1042static int 1043ulz_put(struct uvm_object *uobj, voff_t start, voff_t stop, int flags) 1044{ 1045 struct vm_page *pg; 1046 1047 KDASSERT(uobj == &uvm_loanzero_object); 1048 1049 /* 1050 * Don't need to do any work here if we're not freeing pages. 1051 */ 1052 1053 if ((flags & PGO_FREE) == 0) { 1054 mutex_exit(uobj->vmobjlock); 1055 return 0; 1056 } 1057 1058 /* 1059 * we don't actually want to ever free the uvm_loanzero_page, so 1060 * just reactivate or dequeue it. 1061 */ 1062 1063 pg = TAILQ_FIRST(&uobj->memq); 1064 KASSERT(pg != NULL); 1065 KASSERT(TAILQ_NEXT(pg, listq.queue) == NULL); 1066 1067 if (pg->uanon) 1068 uvm_pageactivate(pg); 1069 else { 1070 uvm_pagedequeue(pg); 1071 } 1072 1073 mutex_exit(uobj->vmobjlock); 1074 return 0; 1075} 1076 1077static const struct uvm_pagerops ulz_pager = { 1078 .pgo_put = ulz_put, 1079}; 1080 1081/* 1082 * uvm_loan_init(): initialize the uvm_loan() facility. 1083 */ 1084 1085void 1086uvm_loan_init(void) 1087{ 1088 1089 mutex_init(&uvm_loanzero_lock, MUTEX_DEFAULT, IPL_NONE); 1090 uvm_obj_init(&uvm_loanzero_object, &ulz_pager, false, 0); 1091 uvm_obj_setlock(&uvm_loanzero_object, &uvm_loanzero_lock); 1092 1093 UVMHIST_INIT(loanhist, 300); 1094} 1095 1096/* 1097 * uvm_loanbreak: break loan on a uobj page 1098 * 1099 * => called with uobj locked 1100 * => the page should be busy 1101 * => return value: 1102 * newly allocated page if succeeded 1103 */ 1104struct vm_page * 1105uvm_loanbreak(struct vm_page *uobjpage) 1106{ 1107 struct vm_page *pg; 1108 struct uvm_object *uobj __diagused = uobjpage->uobject; 1109 1110 KASSERT(uobj != NULL); 1111 KASSERT(mutex_owned(uobj->vmobjlock)); 1112 KASSERT(uobjpage->flags & PG_BUSY); 1113 1114 /* alloc new un-owned page */ 1115 pg = uvm_pagealloc(NULL, 0, NULL, 0); 1116 if (pg == NULL) 1117 return NULL; 1118 1119 /* 1120 * copy the data from the old page to the new 1121 * one and clear the fake flags on the new page (keep it busy). 1122 * force a reload of the old page by clearing it from all 1123 * pmaps. 1124 * transfer dirtiness of the old page to the new page. 1125 * then lock pg->interlock to rename the pages. 1126 */ 1127 1128 uvm_pagecopy(uobjpage, pg); /* old -> new */ 1129 pg->flags &= ~PG_FAKE; 1130 pmap_page_protect(uobjpage, VM_PROT_NONE); 1131 if ((uobjpage->flags & PG_CLEAN) != 0 && !pmap_clear_modify(uobjpage)) { 1132 pmap_clear_modify(pg); 1133 pg->flags |= PG_CLEAN; 1134 } else { 1135 /* uvm_pagecopy marked it dirty */ 1136 KASSERT((pg->flags & PG_CLEAN) == 0); 1137 /* a object with a dirty page should be dirty. */ 1138 KASSERT(!UVM_OBJ_IS_CLEAN(uobj)); 1139 } 1140 if (uobjpage->flags & PG_WANTED) 1141 wakeup(uobjpage); 1142 /* uobj still locked */ 1143 uobjpage->flags &= ~(PG_WANTED|PG_BUSY); 1144 UVM_PAGE_OWN(uobjpage, NULL); 1145 1146 /* 1147 * replace uobjpage with new page. 1148 */ 1149 1150 mutex_enter(&uobjpage->interlock); 1151 uvm_pagereplace(uobjpage, pg); 1152 mutex_exit(&uobjpage->interlock); 1153 1154 /* 1155 * if the page is no longer referenced by 1156 * an anon (i.e. we are breaking an O->K 1157 * loan), then remove it from any pageq's. 1158 */ 1159 if (uobjpage->uanon == NULL) 1160 uvm_pagedequeue(uobjpage); 1161 1162 /* 1163 * at this point we have absolutely no 1164 * control over uobjpage 1165 */ 1166 1167 uvm_pageactivate(pg); 1168 1169 /* 1170 * done! loan is broken and "pg" is 1171 * PG_BUSY. it can now replace uobjpage. 1172 */ 1173 1174 return pg; 1175} 1176 1177int 1178uvm_loanbreak_anon(struct vm_anon *anon, struct uvm_object *uobj) 1179{ 1180 struct vm_page *pg, *dequeuepg; 1181 1182 KASSERT(mutex_owned(anon->an_lock)); 1183 KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock)); 1184 1185 /* get new un-owned replacement page */ 1186 pg = uvm_pagealloc(NULL, 0, NULL, 0); 1187 if (pg == NULL) { 1188 return ENOMEM; 1189 } 1190 1191 /* copy old -> new */ 1192 uvm_pagecopy(anon->an_page, pg); 1193 1194 /* force reload */ 1195 pmap_page_protect(anon->an_page, VM_PROT_NONE); 1196 if (pg < anon->an_page) { 1197 mutex_enter(&pg->interlock); 1198 mutex_enter(&anon->an_page->interlock); 1199 } else { 1200 mutex_enter(&anon->an_page->interlock); 1201 mutex_enter(&pg->interlock); 1202 } 1203 anon->an_page->uanon = NULL; 1204 /* in case we owned */ 1205 anon->an_page->flags &= ~PG_ANON; 1206 1207 if (uobj) { 1208 /* if we were receiver of loan */ 1209 anon->an_page->loan_count--; 1210 dequeuepg = NULL; 1211 } else { 1212 /* 1213 * we were the lender (A->K); need to remove the page from 1214 * pageq's. 1215 */ 1216 dequeuepg = anon->an_page; 1217 } 1218 1219 /* install new page in anon */ 1220 anon->an_page = pg; 1221 pg->uanon = anon; 1222 pg->flags |= PG_ANON; 1223 1224 mutex_exit(&pg->interlock); 1225 mutex_exit(&anon->an_page->interlock); 1226 uvm_pageactivate(pg); 1227 if (dequeuepg != NULL) { 1228 uvm_pagedequeue(anon->an_page); 1229 } 1230 1231 pg->flags &= ~(PG_BUSY|PG_FAKE); 1232 UVM_PAGE_OWN(pg, NULL); 1233 1234 if (uobj) { 1235 mutex_exit(uobj->vmobjlock); 1236 } 1237 1238 /* done! */ 1239 1240 return 0; 1241} 1242