1 /* $NetBSD: uvm_pdaemon.c,v 1.102 2011/02/02 15:25:27 chuck Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * Copyright (c) 1991, 1993, The Regents of the University of California. 6 * 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)vm_pageout.c 8.5 (Berkeley) 2/14/94 37 * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp 38 * 39 * 40 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 41 * All rights reserved. 42 * 43 * Permission to use, copy, modify and distribute this software and 44 * its documentation is hereby granted, provided that both the copyright 45 * notice and this permission notice appear in all copies of the 46 * software, derivative works or modified versions, and any portions 47 * thereof, and that both notices appear in supporting documentation. 48 * 49 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 50 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 51 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 52 * 53 * Carnegie Mellon requests users of this software to return to 54 * 55 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 56 * School of Computer Science 57 * Carnegie Mellon University 58 * Pittsburgh PA 15213-3890 59 * 60 * any improvements or extensions that they make and grant Carnegie the 61 * rights to redistribute these changes. 62 */ 63 64 /* 65 * uvm_pdaemon.c: the page daemon 66 */ 67 68 #include <sys/cdefs.h> 69 __KERNEL_RCSID(0, "$NetBSD: uvm_pdaemon.c,v 1.102 2011/02/02 15:25:27 chuck Exp $"); 70 71 #include "opt_uvmhist.h" 72 #include "opt_readahead.h" 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/module.h> 81 #include <sys/atomic.h> 82 83 #include <uvm/uvm.h> 84 #include <uvm/uvm_pdpolicy.h> 85 86 /* 87 * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate 88 * in a pass thru the inactive list when swap is full. the value should be 89 * "small"... if it's too large we'll cycle the active pages thru the inactive 90 * queue too quickly to for them to be referenced and avoid being freed. 91 */ 92 93 #define UVMPD_NUMDIRTYREACTS 16 94 95 #define UVMPD_NUMTRYLOCKOWNER 16 96 97 /* 98 * local prototypes 99 */ 100 101 static void uvmpd_scan(void); 102 static void uvmpd_scan_queue(void); 103 static void uvmpd_tune(void); 104 105 static unsigned int uvm_pagedaemon_waiters; 106 107 /* 108 * XXX hack to avoid hangs when large processes fork. 109 */ 110 u_int uvm_extrapages; 111 112 static kmutex_t uvm_reclaim_lock; 113 114 SLIST_HEAD(uvm_reclaim_hooks, uvm_reclaim_hook) uvm_reclaim_list; 115 116 /* 117 * uvm_wait: wait (sleep) for the page daemon to free some pages 118 * 119 * => should be called with all locks released 120 * => should _not_ be called by the page daemon (to avoid deadlock) 121 */ 122 123 void 124 uvm_wait(const char *wmsg) 125 { 126 int timo = 0; 127 128 mutex_spin_enter(&uvm_fpageqlock); 129 130 /* 131 * check for page daemon going to sleep (waiting for itself) 132 */ 133 134 if (curlwp == uvm.pagedaemon_lwp && uvmexp.paging == 0) { 135 /* 136 * now we have a problem: the pagedaemon wants to go to 137 * sleep until it frees more memory. but how can it 138 * free more memory if it is asleep? that is a deadlock. 139 * we have two options: 140 * [1] panic now 141 * [2] put a timeout on the sleep, thus causing the 142 * pagedaemon to only pause (rather than sleep forever) 143 * 144 * note that option [2] will only help us if we get lucky 145 * and some other process on the system breaks the deadlock 146 * by exiting or freeing memory (thus allowing the pagedaemon 147 * to continue). for now we panic if DEBUG is defined, 148 * otherwise we hope for the best with option [2] (better 149 * yet, this should never happen in the first place!). 150 */ 151 152 printf("pagedaemon: deadlock detected!\n"); 153 timo = hz >> 3; /* set timeout */ 154 #if defined(DEBUG) 155 /* DEBUG: panic so we can debug it */ 156 panic("pagedaemon deadlock"); 157 #endif 158 } 159 160 uvm_pagedaemon_waiters++; 161 wakeup(&uvm.pagedaemon); /* wake the daemon! */ 162 UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm_fpageqlock, false, wmsg, timo); 163 } 164 165 /* 166 * uvm_kick_pdaemon: perform checks to determine if we need to 167 * give the pagedaemon a nudge, and do so if necessary. 168 * 169 * => called with uvm_fpageqlock held. 170 */ 171 172 void 173 uvm_kick_pdaemon(void) 174 { 175 176 KASSERT(mutex_owned(&uvm_fpageqlock)); 177 178 if (uvmexp.free + uvmexp.paging < uvmexp.freemin || 179 (uvmexp.free + uvmexp.paging < uvmexp.freetarg && 180 uvmpdpol_needsscan_p())) { 181 wakeup(&uvm.pagedaemon); 182 } 183 } 184 185 /* 186 * uvmpd_tune: tune paging parameters 187 * 188 * => called when ever memory is added (or removed?) to the system 189 * => caller must call with page queues locked 190 */ 191 192 static void 193 uvmpd_tune(void) 194 { 195 int val; 196 197 UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist); 198 199 /* 200 * try to keep 0.5% of available RAM free, but limit to between 201 * 128k and 1024k per-CPU. XXX: what are these values good for? 202 */ 203 val = uvmexp.npages / 200; 204 val = MAX(val, (128*1024) >> PAGE_SHIFT); 205 val = MIN(val, (1024*1024) >> PAGE_SHIFT); 206 val *= ncpu; 207 208 /* Make sure there's always a user page free. */ 209 if (val < uvmexp.reserve_kernel + 1) 210 val = uvmexp.reserve_kernel + 1; 211 uvmexp.freemin = val; 212 213 /* Calculate free target. */ 214 val = (uvmexp.freemin * 4) / 3; 215 if (val <= uvmexp.freemin) 216 val = uvmexp.freemin + 1; 217 uvmexp.freetarg = val + atomic_swap_uint(&uvm_extrapages, 0); 218 219 uvmexp.wiredmax = uvmexp.npages / 3; 220 UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d", 221 uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0); 222 } 223 224 /* 225 * uvm_pageout: the main loop for the pagedaemon 226 */ 227 228 void 229 uvm_pageout(void *arg) 230 { 231 int bufcnt, npages = 0; 232 int extrapages = 0; 233 struct pool *pp; 234 uint64_t where; 235 struct uvm_reclaim_hook *hook; 236 237 UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist); 238 239 UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0); 240 241 /* 242 * ensure correct priority and set paging parameters... 243 */ 244 245 uvm.pagedaemon_lwp = curlwp; 246 mutex_enter(&uvm_pageqlock); 247 npages = uvmexp.npages; 248 uvmpd_tune(); 249 mutex_exit(&uvm_pageqlock); 250 251 /* 252 * main loop 253 */ 254 255 for (;;) { 256 bool needsscan, needsfree; 257 258 mutex_spin_enter(&uvm_fpageqlock); 259 if (uvm_pagedaemon_waiters == 0 || uvmexp.paging > 0) { 260 UVMHIST_LOG(pdhist," <<SLEEPING>>",0,0,0,0); 261 UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon, 262 &uvm_fpageqlock, false, "pgdaemon", 0); 263 uvmexp.pdwoke++; 264 UVMHIST_LOG(pdhist," <<WOKE UP>>",0,0,0,0); 265 } else { 266 mutex_spin_exit(&uvm_fpageqlock); 267 } 268 269 /* 270 * now lock page queues and recompute inactive count 271 */ 272 273 mutex_enter(&uvm_pageqlock); 274 if (npages != uvmexp.npages || extrapages != uvm_extrapages) { 275 npages = uvmexp.npages; 276 extrapages = uvm_extrapages; 277 mutex_spin_enter(&uvm_fpageqlock); 278 uvmpd_tune(); 279 mutex_spin_exit(&uvm_fpageqlock); 280 } 281 282 uvmpdpol_tune(); 283 284 /* 285 * Estimate a hint. Note that bufmem are returned to 286 * system only when entire pool page is empty. 287 */ 288 mutex_spin_enter(&uvm_fpageqlock); 289 bufcnt = uvmexp.freetarg - uvmexp.free; 290 if (bufcnt < 0) 291 bufcnt = 0; 292 293 UVMHIST_LOG(pdhist," free/ftarg=%d/%d", 294 uvmexp.free, uvmexp.freetarg, 0,0); 295 296 needsfree = uvmexp.free + uvmexp.paging < uvmexp.freetarg; 297 needsscan = needsfree || uvmpdpol_needsscan_p(); 298 299 /* 300 * scan if needed 301 */ 302 if (needsscan) { 303 mutex_spin_exit(&uvm_fpageqlock); 304 uvmpd_scan(); 305 mutex_spin_enter(&uvm_fpageqlock); 306 } 307 308 /* 309 * if there's any free memory to be had, 310 * wake up any waiters. 311 */ 312 if (uvmexp.free > uvmexp.reserve_kernel || 313 uvmexp.paging == 0) { 314 wakeup(&uvmexp.free); 315 uvm_pagedaemon_waiters = 0; 316 } 317 mutex_spin_exit(&uvm_fpageqlock); 318 319 /* 320 * scan done. unlock page queues (the only lock we are holding) 321 */ 322 mutex_exit(&uvm_pageqlock); 323 324 /* 325 * if we don't need free memory, we're done. 326 */ 327 328 if (!needsfree) 329 continue; 330 331 /* 332 * start draining pool resources now that we're not 333 * holding any locks. 334 */ 335 pool_drain_start(&pp, &where); 336 337 /* 338 * kill unused metadata buffers. 339 */ 340 mutex_enter(&bufcache_lock); 341 buf_drain(bufcnt << PAGE_SHIFT); 342 mutex_exit(&bufcache_lock); 343 344 mutex_enter(&uvm_reclaim_lock); 345 SLIST_FOREACH(hook, &uvm_reclaim_list, uvm_reclaim_next) { 346 (*hook->uvm_reclaim_hook)(); 347 } 348 mutex_exit(&uvm_reclaim_lock); 349 350 /* 351 * complete draining the pools. 352 */ 353 pool_drain_end(pp, where); 354 } 355 /*NOTREACHED*/ 356 } 357 358 359 /* 360 * uvm_aiodone_worker: a workqueue callback for the aiodone daemon. 361 */ 362 363 void 364 uvm_aiodone_worker(struct work *wk, void *dummy) 365 { 366 struct buf *bp = (void *)wk; 367 368 KASSERT(&bp->b_work == wk); 369 370 /* 371 * process an i/o that's done. 372 */ 373 374 (*bp->b_iodone)(bp); 375 } 376 377 void 378 uvm_pageout_start(int npages) 379 { 380 381 mutex_spin_enter(&uvm_fpageqlock); 382 uvmexp.paging += npages; 383 mutex_spin_exit(&uvm_fpageqlock); 384 } 385 386 void 387 uvm_pageout_done(int npages) 388 { 389 390 mutex_spin_enter(&uvm_fpageqlock); 391 KASSERT(uvmexp.paging >= npages); 392 uvmexp.paging -= npages; 393 394 /* 395 * wake up either of pagedaemon or LWPs waiting for it. 396 */ 397 398 if (uvmexp.free <= uvmexp.reserve_kernel) { 399 wakeup(&uvm.pagedaemon); 400 } else { 401 wakeup(&uvmexp.free); 402 uvm_pagedaemon_waiters = 0; 403 } 404 mutex_spin_exit(&uvm_fpageqlock); 405 } 406 407 /* 408 * uvmpd_trylockowner: trylock the page's owner. 409 * 410 * => called with pageq locked. 411 * => resolve orphaned O->A loaned page. 412 * => return the locked mutex on success. otherwise, return NULL. 413 */ 414 415 kmutex_t * 416 uvmpd_trylockowner(struct vm_page *pg) 417 { 418 struct uvm_object *uobj = pg->uobject; 419 kmutex_t *slock; 420 421 KASSERT(mutex_owned(&uvm_pageqlock)); 422 423 if (uobj != NULL) { 424 slock = &uobj->vmobjlock; 425 } else { 426 struct vm_anon *anon = pg->uanon; 427 428 KASSERT(anon != NULL); 429 slock = &anon->an_lock; 430 } 431 432 if (!mutex_tryenter(slock)) { 433 return NULL; 434 } 435 436 if (uobj == NULL) { 437 438 /* 439 * set PQ_ANON if it isn't set already. 440 */ 441 442 if ((pg->pqflags & PQ_ANON) == 0) { 443 KASSERT(pg->loan_count > 0); 444 pg->loan_count--; 445 pg->pqflags |= PQ_ANON; 446 /* anon now owns it */ 447 } 448 } 449 450 return slock; 451 } 452 453 #if defined(VMSWAP) 454 struct swapcluster { 455 int swc_slot; 456 int swc_nallocated; 457 int swc_nused; 458 struct vm_page *swc_pages[howmany(MAXPHYS, MIN_PAGE_SIZE)]; 459 }; 460 461 static void 462 swapcluster_init(struct swapcluster *swc) 463 { 464 465 swc->swc_slot = 0; 466 swc->swc_nused = 0; 467 } 468 469 static int 470 swapcluster_allocslots(struct swapcluster *swc) 471 { 472 int slot; 473 int npages; 474 475 if (swc->swc_slot != 0) { 476 return 0; 477 } 478 479 /* Even with strange MAXPHYS, the shift 480 implicitly rounds down to a page. */ 481 npages = MAXPHYS >> PAGE_SHIFT; 482 slot = uvm_swap_alloc(&npages, true); 483 if (slot == 0) { 484 return ENOMEM; 485 } 486 swc->swc_slot = slot; 487 swc->swc_nallocated = npages; 488 swc->swc_nused = 0; 489 490 return 0; 491 } 492 493 static int 494 swapcluster_add(struct swapcluster *swc, struct vm_page *pg) 495 { 496 int slot; 497 struct uvm_object *uobj; 498 499 KASSERT(swc->swc_slot != 0); 500 KASSERT(swc->swc_nused < swc->swc_nallocated); 501 KASSERT((pg->pqflags & PQ_SWAPBACKED) != 0); 502 503 slot = swc->swc_slot + swc->swc_nused; 504 uobj = pg->uobject; 505 if (uobj == NULL) { 506 KASSERT(mutex_owned(&pg->uanon->an_lock)); 507 pg->uanon->an_swslot = slot; 508 } else { 509 int result; 510 511 KASSERT(mutex_owned(&uobj->vmobjlock)); 512 result = uao_set_swslot(uobj, pg->offset >> PAGE_SHIFT, slot); 513 if (result == -1) { 514 return ENOMEM; 515 } 516 } 517 swc->swc_pages[swc->swc_nused] = pg; 518 swc->swc_nused++; 519 520 return 0; 521 } 522 523 static void 524 swapcluster_flush(struct swapcluster *swc, bool now) 525 { 526 int slot; 527 int nused; 528 int nallocated; 529 int error; 530 531 if (swc->swc_slot == 0) { 532 return; 533 } 534 KASSERT(swc->swc_nused <= swc->swc_nallocated); 535 536 slot = swc->swc_slot; 537 nused = swc->swc_nused; 538 nallocated = swc->swc_nallocated; 539 540 /* 541 * if this is the final pageout we could have a few 542 * unused swap blocks. if so, free them now. 543 */ 544 545 if (nused < nallocated) { 546 if (!now) { 547 return; 548 } 549 uvm_swap_free(slot + nused, nallocated - nused); 550 } 551 552 /* 553 * now start the pageout. 554 */ 555 556 if (nused > 0) { 557 uvmexp.pdpageouts++; 558 uvm_pageout_start(nused); 559 error = uvm_swap_put(slot, swc->swc_pages, nused, 0); 560 KASSERT(error == 0 || error == ENOMEM); 561 } 562 563 /* 564 * zero swslot to indicate that we are 565 * no longer building a swap-backed cluster. 566 */ 567 568 swc->swc_slot = 0; 569 swc->swc_nused = 0; 570 } 571 572 static int 573 swapcluster_nused(struct swapcluster *swc) 574 { 575 576 return swc->swc_nused; 577 } 578 579 /* 580 * uvmpd_dropswap: free any swap allocated to this page. 581 * 582 * => called with owner locked. 583 * => return true if a page had an associated slot. 584 */ 585 586 static bool 587 uvmpd_dropswap(struct vm_page *pg) 588 { 589 bool result = false; 590 struct vm_anon *anon = pg->uanon; 591 592 if ((pg->pqflags & PQ_ANON) && anon->an_swslot) { 593 uvm_swap_free(anon->an_swslot, 1); 594 anon->an_swslot = 0; 595 pg->flags &= ~PG_CLEAN; 596 result = true; 597 } else if (pg->pqflags & PQ_AOBJ) { 598 int slot = uao_set_swslot(pg->uobject, 599 pg->offset >> PAGE_SHIFT, 0); 600 if (slot) { 601 uvm_swap_free(slot, 1); 602 pg->flags &= ~PG_CLEAN; 603 result = true; 604 } 605 } 606 607 return result; 608 } 609 610 /* 611 * uvmpd_trydropswap: try to free any swap allocated to this page. 612 * 613 * => return true if a slot is successfully freed. 614 */ 615 616 bool 617 uvmpd_trydropswap(struct vm_page *pg) 618 { 619 kmutex_t *slock; 620 bool result; 621 622 if ((pg->flags & PG_BUSY) != 0) { 623 return false; 624 } 625 626 /* 627 * lock the page's owner. 628 */ 629 630 slock = uvmpd_trylockowner(pg); 631 if (slock == NULL) { 632 return false; 633 } 634 635 /* 636 * skip this page if it's busy. 637 */ 638 639 if ((pg->flags & PG_BUSY) != 0) { 640 mutex_exit(slock); 641 return false; 642 } 643 644 result = uvmpd_dropswap(pg); 645 646 mutex_exit(slock); 647 648 return result; 649 } 650 651 #endif /* defined(VMSWAP) */ 652 653 /* 654 * uvmpd_scan_queue: scan an replace candidate list for pages 655 * to clean or free. 656 * 657 * => called with page queues locked 658 * => we work on meeting our free target by converting inactive pages 659 * into free pages. 660 * => we handle the building of swap-backed clusters 661 */ 662 663 static void 664 uvmpd_scan_queue(void) 665 { 666 struct vm_page *p; 667 struct uvm_object *uobj; 668 struct vm_anon *anon; 669 #if defined(VMSWAP) 670 struct swapcluster swc; 671 #endif /* defined(VMSWAP) */ 672 int dirtyreacts; 673 int lockownerfail; 674 kmutex_t *slock; 675 UVMHIST_FUNC("uvmpd_scan_queue"); UVMHIST_CALLED(pdhist); 676 677 /* 678 * swslot is non-zero if we are building a swap cluster. we want 679 * to stay in the loop while we have a page to scan or we have 680 * a swap-cluster to build. 681 */ 682 683 #if defined(VMSWAP) 684 swapcluster_init(&swc); 685 #endif /* defined(VMSWAP) */ 686 687 dirtyreacts = 0; 688 lockownerfail = 0; 689 uvmpdpol_scaninit(); 690 691 while (/* CONSTCOND */ 1) { 692 693 /* 694 * see if we've met the free target. 695 */ 696 697 if (uvmexp.free + uvmexp.paging 698 #if defined(VMSWAP) 699 + swapcluster_nused(&swc) 700 #endif /* defined(VMSWAP) */ 701 >= uvmexp.freetarg << 2 || 702 dirtyreacts == UVMPD_NUMDIRTYREACTS) { 703 UVMHIST_LOG(pdhist," met free target: " 704 "exit loop", 0, 0, 0, 0); 705 break; 706 } 707 708 p = uvmpdpol_selectvictim(); 709 if (p == NULL) { 710 break; 711 } 712 KASSERT(uvmpdpol_pageisqueued_p(p)); 713 KASSERT(p->wire_count == 0); 714 715 /* 716 * we are below target and have a new page to consider. 717 */ 718 719 anon = p->uanon; 720 uobj = p->uobject; 721 722 /* 723 * first we attempt to lock the object that this page 724 * belongs to. if our attempt fails we skip on to 725 * the next page (no harm done). it is important to 726 * "try" locking the object as we are locking in the 727 * wrong order (pageq -> object) and we don't want to 728 * deadlock. 729 * 730 * the only time we expect to see an ownerless page 731 * (i.e. a page with no uobject and !PQ_ANON) is if an 732 * anon has loaned a page from a uvm_object and the 733 * uvm_object has dropped the ownership. in that 734 * case, the anon can "take over" the loaned page 735 * and make it its own. 736 */ 737 738 slock = uvmpd_trylockowner(p); 739 if (slock == NULL) { 740 /* 741 * yield cpu to make a chance for an LWP holding 742 * the lock run. otherwise we can busy-loop too long 743 * if the page queue is filled with a lot of pages 744 * from few objects. 745 */ 746 lockownerfail++; 747 if (lockownerfail > UVMPD_NUMTRYLOCKOWNER) { 748 mutex_exit(&uvm_pageqlock); 749 /* XXX Better than yielding but inadequate. */ 750 kpause("livelock", false, 1, NULL); 751 mutex_enter(&uvm_pageqlock); 752 lockownerfail = 0; 753 } 754 continue; 755 } 756 if (p->flags & PG_BUSY) { 757 mutex_exit(slock); 758 uvmexp.pdbusy++; 759 continue; 760 } 761 762 /* does the page belong to an object? */ 763 if (uobj != NULL) { 764 uvmexp.pdobscan++; 765 } else { 766 #if defined(VMSWAP) 767 KASSERT(anon != NULL); 768 uvmexp.pdanscan++; 769 #else /* defined(VMSWAP) */ 770 panic("%s: anon", __func__); 771 #endif /* defined(VMSWAP) */ 772 } 773 774 775 /* 776 * we now have the object and the page queues locked. 777 * if the page is not swap-backed, call the object's 778 * pager to flush and free the page. 779 */ 780 781 #if defined(READAHEAD_STATS) 782 if ((p->pqflags & PQ_READAHEAD) != 0) { 783 p->pqflags &= ~PQ_READAHEAD; 784 uvm_ra_miss.ev_count++; 785 } 786 #endif /* defined(READAHEAD_STATS) */ 787 788 if ((p->pqflags & PQ_SWAPBACKED) == 0) { 789 KASSERT(uobj != NULL); 790 mutex_exit(&uvm_pageqlock); 791 (void) (uobj->pgops->pgo_put)(uobj, p->offset, 792 p->offset + PAGE_SIZE, PGO_CLEANIT|PGO_FREE); 793 mutex_enter(&uvm_pageqlock); 794 continue; 795 } 796 797 /* 798 * the page is swap-backed. remove all the permissions 799 * from the page so we can sync the modified info 800 * without any race conditions. if the page is clean 801 * we can free it now and continue. 802 */ 803 804 pmap_page_protect(p, VM_PROT_NONE); 805 if ((p->flags & PG_CLEAN) && pmap_clear_modify(p)) { 806 p->flags &= ~(PG_CLEAN); 807 } 808 if (p->flags & PG_CLEAN) { 809 int slot; 810 int pageidx; 811 812 pageidx = p->offset >> PAGE_SHIFT; 813 uvm_pagefree(p); 814 uvmexp.pdfreed++; 815 816 /* 817 * for anons, we need to remove the page 818 * from the anon ourselves. for aobjs, 819 * pagefree did that for us. 820 */ 821 822 if (anon) { 823 KASSERT(anon->an_swslot != 0); 824 anon->an_page = NULL; 825 slot = anon->an_swslot; 826 } else { 827 slot = uao_find_swslot(uobj, pageidx); 828 } 829 mutex_exit(slock); 830 831 if (slot > 0) { 832 /* this page is now only in swap. */ 833 mutex_enter(&uvm_swap_data_lock); 834 KASSERT(uvmexp.swpgonly < uvmexp.swpginuse); 835 uvmexp.swpgonly++; 836 mutex_exit(&uvm_swap_data_lock); 837 } 838 continue; 839 } 840 841 #if defined(VMSWAP) 842 /* 843 * this page is dirty, skip it if we'll have met our 844 * free target when all the current pageouts complete. 845 */ 846 847 if (uvmexp.free + uvmexp.paging > uvmexp.freetarg << 2) { 848 mutex_exit(slock); 849 continue; 850 } 851 852 /* 853 * free any swap space allocated to the page since 854 * we'll have to write it again with its new data. 855 */ 856 857 uvmpd_dropswap(p); 858 859 /* 860 * start new swap pageout cluster (if necessary). 861 * 862 * if swap is full reactivate this page so that 863 * we eventually cycle all pages through the 864 * inactive queue. 865 */ 866 867 if (swapcluster_allocslots(&swc)) { 868 dirtyreacts++; 869 uvm_pageactivate(p); 870 mutex_exit(slock); 871 continue; 872 } 873 874 /* 875 * at this point, we're definitely going reuse this 876 * page. mark the page busy and delayed-free. 877 * we should remove the page from the page queues 878 * so we don't ever look at it again. 879 * adjust counters and such. 880 */ 881 882 p->flags |= PG_BUSY; 883 UVM_PAGE_OWN(p, "scan_queue"); 884 885 p->flags |= PG_PAGEOUT; 886 uvm_pagedequeue(p); 887 888 uvmexp.pgswapout++; 889 mutex_exit(&uvm_pageqlock); 890 891 /* 892 * add the new page to the cluster. 893 */ 894 895 if (swapcluster_add(&swc, p)) { 896 p->flags &= ~(PG_BUSY|PG_PAGEOUT); 897 UVM_PAGE_OWN(p, NULL); 898 mutex_enter(&uvm_pageqlock); 899 dirtyreacts++; 900 uvm_pageactivate(p); 901 mutex_exit(slock); 902 continue; 903 } 904 mutex_exit(slock); 905 906 swapcluster_flush(&swc, false); 907 mutex_enter(&uvm_pageqlock); 908 909 /* 910 * the pageout is in progress. bump counters and set up 911 * for the next loop. 912 */ 913 914 uvmexp.pdpending++; 915 916 #else /* defined(VMSWAP) */ 917 uvm_pageactivate(p); 918 mutex_exit(slock); 919 #endif /* defined(VMSWAP) */ 920 } 921 922 #if defined(VMSWAP) 923 mutex_exit(&uvm_pageqlock); 924 swapcluster_flush(&swc, true); 925 mutex_enter(&uvm_pageqlock); 926 #endif /* defined(VMSWAP) */ 927 } 928 929 /* 930 * uvmpd_scan: scan the page queues and attempt to meet our targets. 931 * 932 * => called with pageq's locked 933 */ 934 935 static void 936 uvmpd_scan(void) 937 { 938 int swap_shortage, pages_freed; 939 UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist); 940 941 uvmexp.pdrevs++; 942 943 /* 944 * work on meeting our targets. first we work on our free target 945 * by converting inactive pages into free pages. then we work on 946 * meeting our inactive target by converting active pages to 947 * inactive ones. 948 */ 949 950 UVMHIST_LOG(pdhist, " starting 'free' loop",0,0,0,0); 951 952 pages_freed = uvmexp.pdfreed; 953 uvmpd_scan_queue(); 954 pages_freed = uvmexp.pdfreed - pages_freed; 955 956 /* 957 * detect if we're not going to be able to page anything out 958 * until we free some swap resources from active pages. 959 */ 960 961 swap_shortage = 0; 962 if (uvmexp.free < uvmexp.freetarg && 963 uvmexp.swpginuse >= uvmexp.swpgavail && 964 !uvm_swapisfull() && 965 pages_freed == 0) { 966 swap_shortage = uvmexp.freetarg - uvmexp.free; 967 } 968 969 uvmpdpol_balancequeue(swap_shortage); 970 971 /* 972 * if still below the minimum target, try unloading kernel 973 * modules. 974 */ 975 976 if (uvmexp.free < uvmexp.freemin) { 977 module_thread_kick(); 978 } 979 } 980 981 /* 982 * uvm_reclaimable: decide whether to wait for pagedaemon. 983 * 984 * => return true if it seems to be worth to do uvm_wait. 985 * 986 * XXX should be tunable. 987 * XXX should consider pools, etc? 988 */ 989 990 bool 991 uvm_reclaimable(void) 992 { 993 int filepages; 994 int active, inactive; 995 996 /* 997 * if swap is not full, no problem. 998 */ 999 1000 if (!uvm_swapisfull()) { 1001 return true; 1002 } 1003 1004 /* 1005 * file-backed pages can be reclaimed even when swap is full. 1006 * if we have more than 1/16 of pageable memory or 5MB, try to reclaim. 1007 * 1008 * XXX assume the worst case, ie. all wired pages are file-backed. 1009 * 1010 * XXX should consider about other reclaimable memory. 1011 * XXX ie. pools, traditional buffer cache. 1012 */ 1013 1014 filepages = uvmexp.filepages + uvmexp.execpages - uvmexp.wired; 1015 uvm_estimatepageable(&active, &inactive); 1016 if (filepages >= MIN((active + inactive) >> 4, 1017 5 * 1024 * 1024 >> PAGE_SHIFT)) { 1018 return true; 1019 } 1020 1021 /* 1022 * kill the process, fail allocation, etc.. 1023 */ 1024 1025 return false; 1026 } 1027 1028 void 1029 uvm_estimatepageable(int *active, int *inactive) 1030 { 1031 1032 uvmpdpol_estimatepageable(active, inactive); 1033 } 1034 1035 void 1036 uvm_reclaim_init(void) 1037 { 1038 1039 /* Initialize UVM reclaim hooks. */ 1040 mutex_init(&uvm_reclaim_lock, MUTEX_DEFAULT, IPL_NONE); 1041 SLIST_INIT(&uvm_reclaim_list); 1042 } 1043 1044 void 1045 uvm_reclaim_hook_add(struct uvm_reclaim_hook *hook) 1046 { 1047 1048 KASSERT(hook != NULL); 1049 1050 mutex_enter(&uvm_reclaim_lock); 1051 SLIST_INSERT_HEAD(&uvm_reclaim_list, hook, uvm_reclaim_next); 1052 mutex_exit(&uvm_reclaim_lock); 1053 } 1054 1055 void 1056 uvm_reclaim_hook_del(struct uvm_reclaim_hook *hook_entry) 1057 { 1058 struct uvm_reclaim_hook *hook; 1059 1060 KASSERT(hook_entry != NULL); 1061 1062 mutex_enter(&uvm_reclaim_lock); 1063 SLIST_FOREACH(hook, &uvm_reclaim_list, uvm_reclaim_next) { 1064 if (hook != hook_entry) { 1065 continue; 1066 } 1067 1068 SLIST_REMOVE(&uvm_reclaim_list, hook, uvm_reclaim_hook, 1069 uvm_reclaim_next); 1070 break; 1071 } 1072 1073 mutex_exit(&uvm_reclaim_lock); 1074 } 1075