1 /* $OpenBSD: uvm_pdaemon.c,v 1.30 2006/07/31 11:51:29 mickey Exp $ */ 2 /* $NetBSD: uvm_pdaemon.c,v 1.23 2000/08/20 10:24:14 bjh21 Exp $ */ 3 4 /* 5 * Copyright (c) 1997 Charles D. Cranor and Washington University. 6 * Copyright (c) 1991, 1993, The Regents of the University of California. 7 * 8 * All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * The Mach Operating System project at Carnegie-Mellon University. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by Charles D. Cranor, 24 * Washington University, the University of California, Berkeley and 25 * its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * @(#)vm_pageout.c 8.5 (Berkeley) 2/14/94 43 * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp 44 * 45 * 46 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 47 * All rights reserved. 48 * 49 * Permission to use, copy, modify and distribute this software and 50 * its documentation is hereby granted, provided that both the copyright 51 * notice and this permission notice appear in all copies of the 52 * software, derivative works or modified versions, and any portions 53 * thereof, and that both notices appear in supporting documentation. 54 * 55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 58 * 59 * Carnegie Mellon requests users of this software to return to 60 * 61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 62 * School of Computer Science 63 * Carnegie Mellon University 64 * Pittsburgh PA 15213-3890 65 * 66 * any improvements or extensions that they make and grant Carnegie the 67 * rights to redistribute these changes. 68 */ 69 70 /* 71 * uvm_pdaemon.c: the page daemon 72 */ 73 74 #include <sys/param.h> 75 #include <sys/proc.h> 76 #include <sys/systm.h> 77 #include <sys/kernel.h> 78 #include <sys/pool.h> 79 #include <sys/buf.h> 80 #include <sys/vnode.h> 81 82 #include <uvm/uvm.h> 83 84 /* 85 * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate 86 * in a pass thru the inactive list when swap is full. the value should be 87 * "small"... if it's too large we'll cycle the active pages thru the inactive 88 * queue too quickly to for them to be referenced and avoid being freed. 89 */ 90 91 #define UVMPD_NUMDIRTYREACTS 16 92 93 94 /* 95 * local prototypes 96 */ 97 98 static void uvmpd_scan(void); 99 static boolean_t uvmpd_scan_inactive(struct pglist *); 100 static void uvmpd_tune(void); 101 102 /* 103 * uvm_wait: wait (sleep) for the page daemon to free some pages 104 * 105 * => should be called with all locks released 106 * => should _not_ be called by the page daemon (to avoid deadlock) 107 */ 108 109 void 110 uvm_wait(wmsg) 111 const char *wmsg; 112 { 113 int timo = 0; 114 int s = splbio(); 115 116 /* 117 * check for page daemon going to sleep (waiting for itself) 118 */ 119 120 if (curproc == uvm.pagedaemon_proc) { 121 /* 122 * now we have a problem: the pagedaemon wants to go to 123 * sleep until it frees more memory. but how can it 124 * free more memory if it is asleep? that is a deadlock. 125 * we have two options: 126 * [1] panic now 127 * [2] put a timeout on the sleep, thus causing the 128 * pagedaemon to only pause (rather than sleep forever) 129 * 130 * note that option [2] will only help us if we get lucky 131 * and some other process on the system breaks the deadlock 132 * by exiting or freeing memory (thus allowing the pagedaemon 133 * to continue). for now we panic if DEBUG is defined, 134 * otherwise we hope for the best with option [2] (better 135 * yet, this should never happen in the first place!). 136 */ 137 138 printf("pagedaemon: deadlock detected!\n"); 139 timo = hz >> 3; /* set timeout */ 140 #if defined(DEBUG) 141 /* DEBUG: panic so we can debug it */ 142 panic("pagedaemon deadlock"); 143 #endif 144 } 145 146 simple_lock(&uvm.pagedaemon_lock); 147 wakeup(&uvm.pagedaemon); /* wake the daemon! */ 148 UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm.pagedaemon_lock, FALSE, wmsg, 149 timo); 150 151 splx(s); 152 } 153 154 155 /* 156 * uvmpd_tune: tune paging parameters 157 * 158 * => called when ever memory is added (or removed?) to the system 159 * => caller must call with page queues locked 160 */ 161 162 static void 163 uvmpd_tune() 164 { 165 UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist); 166 167 uvmexp.freemin = uvmexp.npages / 30; 168 169 /* between 16k and 512k */ 170 /* XXX: what are these values good for? */ 171 uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT); 172 uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT); 173 174 /* Make sure there's always a user page free. */ 175 if (uvmexp.freemin < uvmexp.reserve_kernel + 1) 176 uvmexp.freemin = uvmexp.reserve_kernel + 1; 177 178 uvmexp.freetarg = (uvmexp.freemin * 4) / 3; 179 if (uvmexp.freetarg <= uvmexp.freemin) 180 uvmexp.freetarg = uvmexp.freemin + 1; 181 182 /* uvmexp.inactarg: computed in main daemon loop */ 183 184 uvmexp.wiredmax = uvmexp.npages / 3; 185 UVMHIST_LOG(pdhist, "<- done, freemin=%ld, freetarg=%ld, wiredmax=%ld", 186 uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0); 187 } 188 189 /* 190 * uvm_pageout: the main loop for the pagedaemon 191 */ 192 193 void 194 uvm_pageout(void *arg) 195 { 196 int npages = 0; 197 UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist); 198 199 UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0); 200 201 /* 202 * ensure correct priority and set paging parameters... 203 */ 204 205 uvm.pagedaemon_proc = curproc; 206 (void) spl0(); 207 uvm_lock_pageq(); 208 npages = uvmexp.npages; 209 uvmpd_tune(); 210 uvm_unlock_pageq(); 211 212 /* 213 * main loop 214 */ 215 216 for (;;) { 217 simple_lock(&uvm.pagedaemon_lock); 218 219 UVMHIST_LOG(pdhist," <<SLEEPING>>",0,0,0,0); 220 UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon, 221 &uvm.pagedaemon_lock, FALSE, "pgdaemon", 0); 222 uvmexp.pdwoke++; 223 UVMHIST_LOG(pdhist," <<WOKE UP>>",0,0,0,0); 224 225 /* 226 * now lock page queues and recompute inactive count 227 */ 228 229 uvm_lock_pageq(); 230 if (npages != uvmexp.npages) { /* check for new pages? */ 231 npages = uvmexp.npages; 232 uvmpd_tune(); 233 } 234 235 uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3; 236 if (uvmexp.inactarg <= uvmexp.freetarg) { 237 uvmexp.inactarg = uvmexp.freetarg + 1; 238 } 239 240 UVMHIST_LOG(pdhist," free/ftarg=%ld/%ld, inact/itarg=%ld/%ld", 241 uvmexp.free, uvmexp.freetarg, uvmexp.inactive, 242 uvmexp.inactarg); 243 244 /* 245 * scan if needed 246 */ 247 248 #ifdef UBC 249 if (uvmexp.free + uvmexp.paging < uvmexp.freetarg || 250 uvmexp.inactive < uvmexp.inactarg || 251 uvm_pgcnt_vnode > 252 (uvmexp.active + uvmexp.inactive + uvmexp.wired + 253 uvmexp.free) * 13 / 16) { 254 #else 255 if (uvmexp.free < uvmexp.freetarg || 256 uvmexp.inactive < uvmexp.inactarg) { 257 #endif 258 uvmpd_scan(); 259 } 260 261 /* 262 * if there's any free memory to be had, 263 * wake up any waiters. 264 */ 265 266 if (uvmexp.free > uvmexp.reserve_kernel || 267 uvmexp.paging == 0) { 268 wakeup(&uvmexp.free); 269 } 270 271 /* 272 * scan done. unlock page queues (the only lock we are holding) 273 */ 274 275 uvm_unlock_pageq(); 276 } 277 /*NOTREACHED*/ 278 } 279 280 281 /* 282 * uvm_aiodone_daemon: main loop for the aiodone daemon. 283 */ 284 285 void 286 uvm_aiodone_daemon(void *arg) 287 { 288 int s, free; 289 struct buf *bp, *nbp; 290 UVMHIST_FUNC("uvm_aiodoned"); UVMHIST_CALLED(pdhist); 291 292 for (;;) { 293 294 /* 295 * carefully attempt to go to sleep (without losing "wakeups"!). 296 * we need splbio because we want to make sure the aio_done list 297 * is totally empty before we go to sleep. 298 */ 299 300 s = splbio(); 301 simple_lock(&uvm.aiodoned_lock); 302 if (TAILQ_FIRST(&uvm.aio_done) == NULL) { 303 UVMHIST_LOG(pdhist," <<SLEEPING>>",0,0,0,0); 304 UVM_UNLOCK_AND_WAIT(&uvm.aiodoned, 305 &uvm.aiodoned_lock, FALSE, "aiodoned", 0); 306 UVMHIST_LOG(pdhist," <<WOKE UP>>",0,0,0,0); 307 308 /* relock aiodoned_lock, still at splbio */ 309 simple_lock(&uvm.aiodoned_lock); 310 } 311 312 /* 313 * check for done aio structures 314 */ 315 316 bp = TAILQ_FIRST(&uvm.aio_done); 317 if (bp) { 318 TAILQ_INIT(&uvm.aio_done); 319 } 320 321 simple_unlock(&uvm.aiodoned_lock); 322 splx(s); 323 324 /* 325 * process each i/o that's done. 326 */ 327 328 free = uvmexp.free; 329 while (bp != NULL) { 330 if (bp->b_flags & B_PDAEMON) { 331 uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT; 332 } 333 nbp = TAILQ_NEXT(bp, b_freelist); 334 s = splbio(); /* b_iodone must by called at splbio */ 335 (*bp->b_iodone)(bp); 336 splx(s); 337 bp = nbp; 338 } 339 if (free <= uvmexp.reserve_kernel) { 340 s = uvm_lock_fpageq(); 341 wakeup(&uvm.pagedaemon); 342 uvm_unlock_fpageq(s); 343 } else { 344 simple_lock(&uvm.pagedaemon_lock); 345 wakeup(&uvmexp.free); 346 simple_unlock(&uvm.pagedaemon_lock); 347 } 348 } 349 } 350 351 352 353 /* 354 * uvmpd_scan_inactive: scan an inactive list for pages to clean or free. 355 * 356 * => called with page queues locked 357 * => we work on meeting our free target by converting inactive pages 358 * into free pages. 359 * => we handle the building of swap-backed clusters 360 * => we return TRUE if we are exiting because we met our target 361 */ 362 363 static boolean_t 364 uvmpd_scan_inactive(pglst) 365 struct pglist *pglst; 366 { 367 boolean_t retval = FALSE; /* assume we haven't hit target */ 368 int s, free, result; 369 struct vm_page *p, *nextpg; 370 struct uvm_object *uobj; 371 struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp; 372 int npages; 373 struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; /* XXX: see below */ 374 int swnpages, swcpages; /* XXX: see below */ 375 int swslot; 376 struct vm_anon *anon; 377 boolean_t swap_backed; 378 vaddr_t start; 379 int dirtyreacts; 380 UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist); 381 382 /* 383 * note: we currently keep swap-backed pages on a separate inactive 384 * list from object-backed pages. however, merging the two lists 385 * back together again hasn't been ruled out. thus, we keep our 386 * swap cluster in "swpps" rather than in pps (allows us to mix 387 * clustering types in the event of a mixed inactive queue). 388 */ 389 390 /* 391 * swslot is non-zero if we are building a swap cluster. we want 392 * to stay in the loop while we have a page to scan or we have 393 * a swap-cluster to build. 394 */ 395 396 swslot = 0; 397 swnpages = swcpages = 0; 398 free = 0; 399 dirtyreacts = 0; 400 401 for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) { 402 403 /* 404 * note that p can be NULL iff we have traversed the whole 405 * list and need to do one final swap-backed clustered pageout. 406 */ 407 408 uobj = NULL; 409 anon = NULL; 410 411 if (p) { 412 413 /* 414 * update our copy of "free" and see if we've met 415 * our target 416 */ 417 418 s = uvm_lock_fpageq(); 419 free = uvmexp.free; 420 uvm_unlock_fpageq(s); 421 422 if (free + uvmexp.paging >= uvmexp.freetarg << 2 || 423 dirtyreacts == UVMPD_NUMDIRTYREACTS) { 424 UVMHIST_LOG(pdhist," met free target: " 425 "exit loop", 0, 0, 0, 0); 426 retval = TRUE; 427 428 if (swslot == 0) { 429 /* exit now if no swap-i/o pending */ 430 break; 431 } 432 433 /* set p to null to signal final swap i/o */ 434 p = NULL; 435 } 436 } 437 438 if (p) { /* if (we have a new page to consider) */ 439 440 /* 441 * we are below target and have a new page to consider. 442 */ 443 uvmexp.pdscans++; 444 nextpg = TAILQ_NEXT(p, pageq); 445 446 /* 447 * move referenced pages back to active queue and 448 * skip to next page (unlikely to happen since 449 * inactive pages shouldn't have any valid mappings 450 * and we cleared reference before deactivating). 451 */ 452 453 if (pmap_is_referenced(p)) { 454 uvm_pageactivate(p); 455 uvmexp.pdreact++; 456 continue; 457 } 458 459 /* 460 * first we attempt to lock the object that this page 461 * belongs to. if our attempt fails we skip on to 462 * the next page (no harm done). it is important to 463 * "try" locking the object as we are locking in the 464 * wrong order (pageq -> object) and we don't want to 465 * deadlock. 466 * 467 * the only time we expect to see an ownerless page 468 * (i.e. a page with no uobject and !PQ_ANON) is if an 469 * anon has loaned a page from a uvm_object and the 470 * uvm_object has dropped the ownership. in that 471 * case, the anon can "take over" the loaned page 472 * and make it its own. 473 */ 474 475 /* is page part of an anon or ownerless ? */ 476 if ((p->pqflags & PQ_ANON) || p->uobject == NULL) { 477 anon = p->uanon; 478 KASSERT(anon != NULL); 479 if (!simple_lock_try(&anon->an_lock)) { 480 /* lock failed, skip this page */ 481 continue; 482 } 483 484 /* 485 * if the page is ownerless, claim it in the 486 * name of "anon"! 487 */ 488 489 if ((p->pqflags & PQ_ANON) == 0) { 490 KASSERT(p->loan_count > 0); 491 p->loan_count--; 492 p->pqflags |= PQ_ANON; 493 /* anon now owns it */ 494 } 495 if (p->flags & PG_BUSY) { 496 simple_unlock(&anon->an_lock); 497 uvmexp.pdbusy++; 498 /* someone else owns page, skip it */ 499 continue; 500 } 501 uvmexp.pdanscan++; 502 } else { 503 uobj = p->uobject; 504 KASSERT(uobj != NULL); 505 if (!simple_lock_try(&uobj->vmobjlock)) { 506 /* lock failed, skip this page */ 507 continue; 508 } 509 if (p->flags & PG_BUSY) { 510 simple_unlock(&uobj->vmobjlock); 511 uvmexp.pdbusy++; 512 /* someone else owns page, skip it */ 513 continue; 514 } 515 uvmexp.pdobscan++; 516 } 517 518 /* 519 * we now have the object and the page queues locked. 520 * the page is not busy. if the page is clean we 521 * can free it now and continue. 522 */ 523 524 if (p->flags & PG_CLEAN) { 525 if (p->pqflags & PQ_SWAPBACKED) { 526 /* this page now lives only in swap */ 527 simple_lock(&uvm.swap_data_lock); 528 uvmexp.swpgonly++; 529 simple_unlock(&uvm.swap_data_lock); 530 } 531 532 /* zap all mappings with pmap_page_protect... */ 533 pmap_page_protect(p, VM_PROT_NONE); 534 uvm_pagefree(p); 535 uvmexp.pdfreed++; 536 537 if (anon) { 538 539 /* 540 * an anonymous page can only be clean 541 * if it has backing store assigned. 542 */ 543 544 KASSERT(anon->an_swslot != 0); 545 546 /* remove from object */ 547 anon->u.an_page = NULL; 548 simple_unlock(&anon->an_lock); 549 } else { 550 /* pagefree has already removed the 551 * page from the object */ 552 simple_unlock(&uobj->vmobjlock); 553 } 554 continue; 555 } 556 557 /* 558 * this page is dirty, skip it if we'll have met our 559 * free target when all the current pageouts complete. 560 */ 561 562 if (free + uvmexp.paging > uvmexp.freetarg << 2) { 563 if (anon) { 564 simple_unlock(&anon->an_lock); 565 } else { 566 simple_unlock(&uobj->vmobjlock); 567 } 568 continue; 569 } 570 571 /* 572 * this page is dirty, but we can't page it out 573 * since all pages in swap are only in swap. 574 * reactivate it so that we eventually cycle 575 * all pages thru the inactive queue. 576 */ 577 578 KASSERT(uvmexp.swpgonly <= uvmexp.swpages); 579 if ((p->pqflags & PQ_SWAPBACKED) && 580 uvmexp.swpgonly == uvmexp.swpages) { 581 dirtyreacts++; 582 uvm_pageactivate(p); 583 if (anon) { 584 simple_unlock(&anon->an_lock); 585 } else { 586 simple_unlock(&uobj->vmobjlock); 587 } 588 continue; 589 } 590 591 /* 592 * if the page is swap-backed and dirty and swap space 593 * is full, free any swap allocated to the page 594 * so that other pages can be paged out. 595 */ 596 597 KASSERT(uvmexp.swpginuse <= uvmexp.swpages); 598 if ((p->pqflags & PQ_SWAPBACKED) && 599 uvmexp.swpginuse == uvmexp.swpages) { 600 601 if ((p->pqflags & PQ_ANON) && 602 p->uanon->an_swslot) { 603 uvm_swap_free(p->uanon->an_swslot, 1); 604 p->uanon->an_swslot = 0; 605 } 606 if (p->pqflags & PQ_AOBJ) { 607 uao_dropswap(p->uobject, 608 p->offset >> PAGE_SHIFT); 609 } 610 } 611 612 /* 613 * the page we are looking at is dirty. we must 614 * clean it before it can be freed. to do this we 615 * first mark the page busy so that no one else will 616 * touch the page. we write protect all the mappings 617 * of the page so that no one touches it while it is 618 * in I/O. 619 */ 620 621 swap_backed = ((p->pqflags & PQ_SWAPBACKED) != 0); 622 p->flags |= PG_BUSY; /* now we own it */ 623 UVM_PAGE_OWN(p, "scan_inactive"); 624 pmap_page_protect(p, VM_PROT_READ); 625 uvmexp.pgswapout++; 626 627 /* 628 * for swap-backed pages we need to (re)allocate 629 * swap space. 630 */ 631 632 if (swap_backed) { 633 634 /* 635 * free old swap slot (if any) 636 */ 637 638 if (anon) { 639 if (anon->an_swslot) { 640 uvm_swap_free(anon->an_swslot, 641 1); 642 anon->an_swslot = 0; 643 } 644 } else { 645 uao_dropswap(uobj, 646 p->offset >> PAGE_SHIFT); 647 } 648 649 /* 650 * start new cluster (if necessary) 651 */ 652 653 if (swslot == 0) { 654 swnpages = MAXBSIZE >> PAGE_SHIFT; 655 swslot = uvm_swap_alloc(&swnpages, 656 TRUE); 657 if (swslot == 0) { 658 /* no swap? give up! */ 659 p->flags &= ~PG_BUSY; 660 UVM_PAGE_OWN(p, NULL); 661 if (anon) 662 simple_unlock( 663 &anon->an_lock); 664 else 665 simple_unlock( 666 &uobj->vmobjlock); 667 continue; 668 } 669 swcpages = 0; /* cluster is empty */ 670 } 671 672 /* 673 * add block to cluster 674 */ 675 676 swpps[swcpages] = p; 677 if (anon) 678 anon->an_swslot = swslot + swcpages; 679 else 680 uao_set_swslot(uobj, 681 p->offset >> PAGE_SHIFT, 682 swslot + swcpages); 683 swcpages++; 684 } 685 } else { 686 687 /* if p == NULL we must be doing a last swap i/o */ 688 swap_backed = TRUE; 689 } 690 691 /* 692 * now consider doing the pageout. 693 * 694 * for swap-backed pages, we do the pageout if we have either 695 * filled the cluster (in which case (swnpages == swcpages) or 696 * run out of pages (p == NULL). 697 * 698 * for object pages, we always do the pageout. 699 */ 700 701 if (swap_backed) { 702 if (p) { /* if we just added a page to cluster */ 703 if (anon) 704 simple_unlock(&anon->an_lock); 705 else 706 simple_unlock(&uobj->vmobjlock); 707 708 /* cluster not full yet? */ 709 if (swcpages < swnpages) 710 continue; 711 } 712 713 /* starting I/O now... set up for it */ 714 npages = swcpages; 715 ppsp = swpps; 716 /* for swap-backed pages only */ 717 start = (vaddr_t) swslot; 718 719 /* if this is final pageout we could have a few 720 * extra swap blocks */ 721 if (swcpages < swnpages) { 722 uvm_swap_free(swslot + swcpages, 723 (swnpages - swcpages)); 724 } 725 } else { 726 /* normal object pageout */ 727 ppsp = pps; 728 npages = sizeof(pps) / sizeof(struct vm_page *); 729 /* not looked at because PGO_ALLPAGES is set */ 730 start = 0; 731 } 732 733 /* 734 * now do the pageout. 735 * 736 * for swap_backed pages we have already built the cluster. 737 * for !swap_backed pages, uvm_pager_put will call the object's 738 * "make put cluster" function to build a cluster on our behalf. 739 * 740 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct 741 * it to free the cluster pages for us on a successful I/O (it 742 * always does this for un-successful I/O requests). this 743 * allows us to do clustered pageout without having to deal 744 * with cluster pages at this level. 745 * 746 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST: 747 * IN: locked: uobj (if !swap_backed), page queues 748 * OUT: locked: uobj (if !swap_backed && result !=VM_PAGER_PEND) 749 * !locked: pageqs, uobj (if swap_backed || VM_PAGER_PEND) 750 * 751 * [the bit about VM_PAGER_PEND saves us one lock-unlock pair] 752 */ 753 754 /* locked: uobj (if !swap_backed), page queues */ 755 uvmexp.pdpageouts++; 756 result = uvm_pager_put(swap_backed ? NULL : uobj, p, 757 &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0); 758 /* locked: uobj (if !swap_backed && result != PEND) */ 759 /* unlocked: pageqs, object (if swap_backed ||result == PEND) */ 760 761 /* 762 * if we did i/o to swap, zero swslot to indicate that we are 763 * no longer building a swap-backed cluster. 764 */ 765 766 if (swap_backed) 767 swslot = 0; /* done with this cluster */ 768 769 /* 770 * first, we check for VM_PAGER_PEND which means that the 771 * async I/O is in progress and the async I/O done routine 772 * will clean up after us. in this case we move on to the 773 * next page. 774 * 775 * there is a very remote chance that the pending async i/o can 776 * finish _before_ we get here. if that happens, our page "p" 777 * may no longer be on the inactive queue. so we verify this 778 * when determining the next page (starting over at the head if 779 * we've lost our inactive page). 780 */ 781 782 if (result == VM_PAGER_PEND) { 783 uvmexp.paging += npages; 784 uvm_lock_pageq(); 785 uvmexp.pdpending++; 786 if (p) { 787 if (p->pqflags & PQ_INACTIVE) 788 nextpg = TAILQ_NEXT(p, pageq); 789 else 790 nextpg = TAILQ_FIRST(pglst); 791 } else { 792 nextpg = NULL; 793 } 794 continue; 795 } 796 797 #ifdef UBC 798 if (result == VM_PAGER_ERROR && 799 curproc == uvm.pagedaemon_proc) { 800 uvm_lock_pageq(); 801 nextpg = TAILQ_NEXT(p, pageq); 802 uvm_pageactivate(p); 803 continue; 804 } 805 #endif 806 807 /* 808 * clean up "p" if we have one 809 */ 810 811 if (p) { 812 /* 813 * the I/O request to "p" is done and uvm_pager_put 814 * has freed any cluster pages it may have allocated 815 * during I/O. all that is left for us to do is 816 * clean up page "p" (which is still PG_BUSY). 817 * 818 * our result could be one of the following: 819 * VM_PAGER_OK: successful pageout 820 * 821 * VM_PAGER_AGAIN: tmp resource shortage, we skip 822 * to next page 823 * VM_PAGER_{FAIL,ERROR,BAD}: an error. we 824 * "reactivate" page to get it out of the way (it 825 * will eventually drift back into the inactive 826 * queue for a retry). 827 * VM_PAGER_UNLOCK: should never see this as it is 828 * only valid for "get" operations 829 */ 830 831 /* relock p's object: page queues not lock yet, so 832 * no need for "try" */ 833 834 /* !swap_backed case: already locked... */ 835 if (swap_backed) { 836 if (anon) 837 simple_lock(&anon->an_lock); 838 else 839 simple_lock(&uobj->vmobjlock); 840 } 841 842 #ifdef DIAGNOSTIC 843 if (result == VM_PAGER_UNLOCK) 844 panic("pagedaemon: pageout returned " 845 "invalid 'unlock' code"); 846 #endif 847 848 /* handle PG_WANTED now */ 849 if (p->flags & PG_WANTED) 850 /* still holding object lock */ 851 wakeup(p); 852 853 p->flags &= ~(PG_BUSY|PG_WANTED); 854 UVM_PAGE_OWN(p, NULL); 855 856 /* released during I/O? */ 857 if (p->flags & PG_RELEASED) { 858 if (anon) { 859 /* remove page so we can get nextpg */ 860 anon->u.an_page = NULL; 861 862 simple_unlock(&anon->an_lock); 863 uvm_anfree(anon); /* kills anon */ 864 pmap_page_protect(p, VM_PROT_NONE); 865 anon = NULL; 866 uvm_lock_pageq(); 867 nextpg = TAILQ_NEXT(p, pageq); 868 /* free released page */ 869 uvm_pagefree(p); 870 871 } else { 872 873 /* 874 * pgo_releasepg nukes the page and 875 * gets "nextpg" for us. it returns 876 * with the page queues locked (when 877 * given nextpg ptr). 878 */ 879 880 if (!uobj->pgops->pgo_releasepg(p, 881 &nextpg)) 882 /* uobj died after release */ 883 uobj = NULL; 884 885 /* 886 * lock page queues here so that they're 887 * always locked at the end of the loop. 888 */ 889 890 uvm_lock_pageq(); 891 } 892 } else { /* page was not released during I/O */ 893 uvm_lock_pageq(); 894 nextpg = TAILQ_NEXT(p, pageq); 895 if (result != VM_PAGER_OK) { 896 /* pageout was a failure... */ 897 if (result != VM_PAGER_AGAIN) 898 uvm_pageactivate(p); 899 pmap_clear_reference(p); 900 /* XXXCDC: if (swap_backed) FREE p's 901 * swap block? */ 902 } else { 903 /* pageout was a success... */ 904 pmap_clear_reference(p); 905 pmap_clear_modify(p); 906 p->flags |= PG_CLEAN; 907 } 908 } 909 910 /* 911 * drop object lock (if there is an object left). do 912 * a safety check of nextpg to make sure it is on the 913 * inactive queue (it should be since PG_BUSY pages on 914 * the inactive queue can't be re-queued [note: not 915 * true for active queue]). 916 */ 917 918 if (anon) 919 simple_unlock(&anon->an_lock); 920 else if (uobj) 921 simple_unlock(&uobj->vmobjlock); 922 923 } else { 924 925 /* 926 * if p is null in this loop, make sure it stays null 927 * in the next loop. 928 */ 929 930 nextpg = NULL; 931 932 /* 933 * lock page queues here just so they're always locked 934 * at the end of the loop. 935 */ 936 937 uvm_lock_pageq(); 938 } 939 940 if (nextpg && (nextpg->pqflags & PQ_INACTIVE) == 0) { 941 nextpg = TAILQ_FIRST(pglst); /* reload! */ 942 } 943 } 944 return (retval); 945 } 946 947 /* 948 * uvmpd_scan: scan the page queues and attempt to meet our targets. 949 * 950 * => called with pageq's locked 951 */ 952 953 void 954 uvmpd_scan() 955 { 956 int s, free, inactive_shortage, swap_shortage, pages_freed; 957 struct vm_page *p, *nextpg; 958 struct uvm_object *uobj; 959 boolean_t got_it; 960 UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist); 961 962 uvmexp.pdrevs++; /* counter */ 963 uobj = NULL; 964 965 /* 966 * get current "free" page count 967 */ 968 s = uvm_lock_fpageq(); 969 free = uvmexp.free; 970 uvm_unlock_fpageq(s); 971 972 #ifndef __SWAP_BROKEN 973 /* 974 * swap out some processes if we are below our free target. 975 * we need to unlock the page queues for this. 976 */ 977 if (free < uvmexp.freetarg) { 978 uvmexp.pdswout++; 979 UVMHIST_LOG(pdhist," free %ld < target %ld: swapout", free, 980 uvmexp.freetarg, 0, 0); 981 uvm_unlock_pageq(); 982 uvm_swapout_threads(); 983 uvm_lock_pageq(); 984 985 } 986 #endif 987 988 /* 989 * now we want to work on meeting our targets. first we work on our 990 * free target by converting inactive pages into free pages. then 991 * we work on meeting our inactive target by converting active pages 992 * to inactive ones. 993 */ 994 995 UVMHIST_LOG(pdhist, " starting 'free' loop",0,0,0,0); 996 997 /* 998 * alternate starting queue between swap and object based on the 999 * low bit of uvmexp.pdrevs (which we bump by one each call). 1000 */ 1001 1002 got_it = FALSE; 1003 pages_freed = uvmexp.pdfreed; 1004 if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0) 1005 got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp); 1006 if (!got_it) 1007 got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj); 1008 if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0) 1009 (void) uvmpd_scan_inactive(&uvm.page_inactive_swp); 1010 pages_freed = uvmexp.pdfreed - pages_freed; 1011 1012 /* 1013 * we have done the scan to get free pages. now we work on meeting 1014 * our inactive target. 1015 */ 1016 1017 inactive_shortage = uvmexp.inactarg - uvmexp.inactive; 1018 1019 /* 1020 * detect if we're not going to be able to page anything out 1021 * until we free some swap resources from active pages. 1022 */ 1023 1024 swap_shortage = 0; 1025 if (uvmexp.free < uvmexp.freetarg && 1026 uvmexp.swpginuse == uvmexp.swpages && 1027 uvmexp.swpgonly < uvmexp.swpages && 1028 pages_freed == 0) { 1029 swap_shortage = uvmexp.freetarg - uvmexp.free; 1030 } 1031 1032 UVMHIST_LOG(pdhist, " loop 2: inactive_shortage=%ld swap_shortage=%ld", 1033 inactive_shortage, swap_shortage,0,0); 1034 for (p = TAILQ_FIRST(&uvm.page_active); 1035 p != NULL && (inactive_shortage > 0 || swap_shortage > 0); 1036 p = nextpg) { 1037 nextpg = TAILQ_NEXT(p, pageq); 1038 if (p->flags & PG_BUSY) 1039 continue; /* quick check before trying to lock */ 1040 1041 /* 1042 * lock the page's owner. 1043 */ 1044 /* is page anon owned or ownerless? */ 1045 if ((p->pqflags & PQ_ANON) || p->uobject == NULL) { 1046 KASSERT(p->uanon != NULL); 1047 if (!simple_lock_try(&p->uanon->an_lock)) 1048 continue; 1049 1050 /* take over the page? */ 1051 if ((p->pqflags & PQ_ANON) == 0) { 1052 KASSERT(p->loan_count > 0); 1053 p->loan_count--; 1054 p->pqflags |= PQ_ANON; 1055 } 1056 } else { 1057 if (!simple_lock_try(&p->uobject->vmobjlock)) 1058 continue; 1059 } 1060 1061 /* 1062 * skip this page if it's busy. 1063 */ 1064 1065 if ((p->flags & PG_BUSY) != 0) { 1066 if (p->pqflags & PQ_ANON) 1067 simple_unlock(&p->uanon->an_lock); 1068 else 1069 simple_unlock(&p->uobject->vmobjlock); 1070 continue; 1071 } 1072 1073 /* 1074 * if there's a shortage of swap, free any swap allocated 1075 * to this page so that other pages can be paged out. 1076 */ 1077 1078 if (swap_shortage > 0) { 1079 if ((p->pqflags & PQ_ANON) && p->uanon->an_swslot) { 1080 uvm_swap_free(p->uanon->an_swslot, 1); 1081 p->uanon->an_swslot = 0; 1082 p->flags &= ~PG_CLEAN; 1083 swap_shortage--; 1084 } 1085 if (p->pqflags & PQ_AOBJ) { 1086 int slot = uao_set_swslot(p->uobject, 1087 p->offset >> PAGE_SHIFT, 0); 1088 if (slot) { 1089 uvm_swap_free(slot, 1); 1090 p->flags &= ~PG_CLEAN; 1091 swap_shortage--; 1092 } 1093 } 1094 } 1095 1096 /* 1097 * deactivate this page if there's a shortage of 1098 * inactive pages. 1099 */ 1100 1101 if (inactive_shortage > 0) { 1102 pmap_page_protect(p, VM_PROT_NONE); 1103 /* no need to check wire_count as pg is "active" */ 1104 uvm_pagedeactivate(p); 1105 uvmexp.pddeact++; 1106 inactive_shortage--; 1107 } 1108 if (p->pqflags & PQ_ANON) 1109 simple_unlock(&p->uanon->an_lock); 1110 else 1111 simple_unlock(&p->uobject->vmobjlock); 1112 } 1113 } 1114