1 /*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * Copyright (c) 2005 Yahoo! Technologies Norway AS 9 * All rights reserved. 10 * 11 * This code is derived from software contributed to Berkeley by 12 * The Mach Operating System project at Carnegie-Mellon University. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and 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 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 43 * 44 * 45 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46 * All rights reserved. 47 * 48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 49 * 50 * Permission to use, copy, modify and distribute this software and 51 * its documentation is hereby granted, provided that both the copyright 52 * notice and this permission notice appear in all copies of the 53 * software, derivative works or modified versions, and any portions 54 * thereof, and that both notices appear in supporting documentation. 55 * 56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 59 * 60 * Carnegie Mellon requests users of this software to return to 61 * 62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 63 * School of Computer Science 64 * Carnegie Mellon University 65 * Pittsburgh PA 15213-3890 66 * 67 * any improvements or extensions that they make and grant Carnegie the 68 * rights to redistribute these changes. 69 */ 70 71 /* 72 * The proverbial page-out daemon. 73 */ 74 75 #include <sys/cdefs.h> 76 __FBSDID("$FreeBSD$"); 77 78 #include "opt_vm.h" 79 #include <sys/param.h> 80 #include <sys/systm.h> 81 #include <sys/kernel.h> 82 #include <sys/eventhandler.h> 83 #include <sys/lock.h> 84 #include <sys/mutex.h> 85 #include <sys/proc.h> 86 #include <sys/kthread.h> 87 #include <sys/ktr.h> 88 #include <sys/mount.h> 89 #include <sys/racct.h> 90 #include <sys/resourcevar.h> 91 #include <sys/sched.h> 92 #include <sys/signalvar.h> 93 #include <sys/vnode.h> 94 #include <sys/vmmeter.h> 95 #include <sys/sx.h> 96 #include <sys/sysctl.h> 97 98 #include <vm/vm.h> 99 #include <vm/vm_param.h> 100 #include <vm/vm_object.h> 101 #include <vm/vm_page.h> 102 #include <vm/vm_map.h> 103 #include <vm/vm_pageout.h> 104 #include <vm/vm_pager.h> 105 #include <vm/swap_pager.h> 106 #include <vm/vm_extern.h> 107 #include <vm/uma.h> 108 109 /* 110 * System initialization 111 */ 112 113 /* the kernel process "vm_pageout"*/ 114 static void vm_pageout(void); 115 static int vm_pageout_clean(vm_page_t); 116 static void vm_pageout_scan(int pass); 117 118 struct proc *pageproc; 119 120 static struct kproc_desc page_kp = { 121 "pagedaemon", 122 vm_pageout, 123 &pageproc 124 }; 125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, 126 &page_kp); 127 128 #if !defined(NO_SWAPPING) 129 /* the kernel process "vm_daemon"*/ 130 static void vm_daemon(void); 131 static struct proc *vmproc; 132 133 static struct kproc_desc vm_kp = { 134 "vmdaemon", 135 vm_daemon, 136 &vmproc 137 }; 138 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 139 #endif 140 141 142 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 143 int vm_pageout_deficit; /* Estimated number of pages deficit */ 144 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 145 146 #if !defined(NO_SWAPPING) 147 static int vm_pageout_req_swapout; /* XXX */ 148 static int vm_daemon_needed; 149 static struct mtx vm_daemon_mtx; 150 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 152 #endif 153 static int vm_max_launder = 32; 154 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 155 static int vm_pageout_full_stats_interval = 0; 156 static int vm_pageout_algorithm=0; 157 static int defer_swap_pageouts=0; 158 static int disable_swap_pageouts=0; 159 160 #if defined(NO_SWAPPING) 161 static int vm_swap_enabled=0; 162 static int vm_swap_idle_enabled=0; 163 #else 164 static int vm_swap_enabled=1; 165 static int vm_swap_idle_enabled=0; 166 #endif 167 168 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 169 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 170 171 SYSCTL_INT(_vm, OID_AUTO, max_launder, 172 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 173 174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 175 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 176 177 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 178 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 179 180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 181 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 182 183 #if defined(NO_SWAPPING) 184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 185 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 187 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 188 #else 189 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 190 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 191 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 192 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 193 #endif 194 195 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 196 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 197 198 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 199 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 200 201 static int pageout_lock_miss; 202 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 203 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 204 205 #define VM_PAGEOUT_PAGE_COUNT 16 206 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 207 208 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 209 SYSCTL_INT(_vm, OID_AUTO, max_wired, 210 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 211 212 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *); 213 static boolean_t vm_pageout_launder(int, int, vm_paddr_t, vm_paddr_t); 214 #if !defined(NO_SWAPPING) 215 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 216 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 217 static void vm_req_vmdaemon(int req); 218 #endif 219 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *); 220 static void vm_pageout_page_stats(void); 221 static void vm_pageout_requeue(vm_page_t m); 222 223 /* 224 * Initialize a dummy page for marking the caller's place in the specified 225 * paging queue. In principle, this function only needs to set the flag 226 * PG_MARKER. Nonetheless, it sets the flag VPO_BUSY and initializes the hold 227 * count to one as safety precautions. 228 */ 229 static void 230 vm_pageout_init_marker(vm_page_t marker, u_short queue) 231 { 232 233 bzero(marker, sizeof(*marker)); 234 marker->flags = PG_MARKER; 235 marker->oflags = VPO_BUSY; 236 marker->queue = queue; 237 marker->hold_count = 1; 238 } 239 240 /* 241 * vm_pageout_fallback_object_lock: 242 * 243 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 244 * known to have failed and page queue must be either PQ_ACTIVE or 245 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 246 * while locking the vm object. Use marker page to detect page queue 247 * changes and maintain notion of next page on page queue. Return 248 * TRUE if no changes were detected, FALSE otherwise. vm object is 249 * locked on return. 250 * 251 * This function depends on both the lock portion of struct vm_object 252 * and normal struct vm_page being type stable. 253 */ 254 static boolean_t 255 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 256 { 257 struct vm_page marker; 258 boolean_t unchanged; 259 u_short queue; 260 vm_object_t object; 261 262 queue = m->queue; 263 vm_pageout_init_marker(&marker, queue); 264 object = m->object; 265 266 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 267 m, &marker, pageq); 268 vm_page_unlock_queues(); 269 vm_page_unlock(m); 270 VM_OBJECT_LOCK(object); 271 vm_page_lock(m); 272 vm_page_lock_queues(); 273 274 /* Page queue might have changed. */ 275 *next = TAILQ_NEXT(&marker, pageq); 276 unchanged = (m->queue == queue && 277 m->object == object && 278 &marker == TAILQ_NEXT(m, pageq)); 279 TAILQ_REMOVE(&vm_page_queues[queue].pl, 280 &marker, pageq); 281 return (unchanged); 282 } 283 284 /* 285 * Lock the page while holding the page queue lock. Use marker page 286 * to detect page queue changes and maintain notion of next page on 287 * page queue. Return TRUE if no changes were detected, FALSE 288 * otherwise. The page is locked on return. The page queue lock might 289 * be dropped and reacquired. 290 * 291 * This function depends on normal struct vm_page being type stable. 292 */ 293 static boolean_t 294 vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 295 { 296 struct vm_page marker; 297 boolean_t unchanged; 298 u_short queue; 299 300 vm_page_lock_assert(m, MA_NOTOWNED); 301 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 302 303 if (vm_page_trylock(m)) 304 return (TRUE); 305 306 queue = m->queue; 307 vm_pageout_init_marker(&marker, queue); 308 309 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq); 310 vm_page_unlock_queues(); 311 vm_page_lock(m); 312 vm_page_lock_queues(); 313 314 /* Page queue might have changed. */ 315 *next = TAILQ_NEXT(&marker, pageq); 316 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq)); 317 TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq); 318 return (unchanged); 319 } 320 321 /* 322 * vm_pageout_clean: 323 * 324 * Clean the page and remove it from the laundry. 325 * 326 * We set the busy bit to cause potential page faults on this page to 327 * block. Note the careful timing, however, the busy bit isn't set till 328 * late and we cannot do anything that will mess with the page. 329 */ 330 static int 331 vm_pageout_clean(vm_page_t m) 332 { 333 vm_object_t object; 334 vm_page_t mc[2*vm_pageout_page_count], pb, ps; 335 int pageout_count; 336 int ib, is, page_base; 337 vm_pindex_t pindex = m->pindex; 338 339 vm_page_lock_assert(m, MA_OWNED); 340 object = m->object; 341 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 342 343 /* 344 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 345 * with the new swapper, but we could have serious problems paging 346 * out other object types if there is insufficient memory. 347 * 348 * Unfortunately, checking free memory here is far too late, so the 349 * check has been moved up a procedural level. 350 */ 351 352 /* 353 * Can't clean the page if it's busy or held. 354 */ 355 KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0, 356 ("vm_pageout_clean: page %p is busy", m)); 357 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m)); 358 vm_page_unlock(m); 359 360 mc[vm_pageout_page_count] = pb = ps = m; 361 pageout_count = 1; 362 page_base = vm_pageout_page_count; 363 ib = 1; 364 is = 1; 365 366 /* 367 * Scan object for clusterable pages. 368 * 369 * We can cluster ONLY if: ->> the page is NOT 370 * clean, wired, busy, held, or mapped into a 371 * buffer, and one of the following: 372 * 1) The page is inactive, or a seldom used 373 * active page. 374 * -or- 375 * 2) we force the issue. 376 * 377 * During heavy mmap/modification loads the pageout 378 * daemon can really fragment the underlying file 379 * due to flushing pages out of order and not trying 380 * align the clusters (which leave sporatic out-of-order 381 * holes). To solve this problem we do the reverse scan 382 * first and attempt to align our cluster, then do a 383 * forward scan if room remains. 384 */ 385 more: 386 while (ib && pageout_count < vm_pageout_page_count) { 387 vm_page_t p; 388 389 if (ib > pindex) { 390 ib = 0; 391 break; 392 } 393 394 if ((p = vm_page_prev(pb)) == NULL || 395 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) { 396 ib = 0; 397 break; 398 } 399 vm_page_lock(p); 400 vm_page_test_dirty(p); 401 if (p->dirty == 0 || 402 p->queue != PQ_INACTIVE || 403 p->hold_count != 0) { /* may be undergoing I/O */ 404 vm_page_unlock(p); 405 ib = 0; 406 break; 407 } 408 vm_page_unlock(p); 409 mc[--page_base] = pb = p; 410 ++pageout_count; 411 ++ib; 412 /* 413 * alignment boundry, stop here and switch directions. Do 414 * not clear ib. 415 */ 416 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 417 break; 418 } 419 420 while (pageout_count < vm_pageout_page_count && 421 pindex + is < object->size) { 422 vm_page_t p; 423 424 if ((p = vm_page_next(ps)) == NULL || 425 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) 426 break; 427 vm_page_lock(p); 428 vm_page_test_dirty(p); 429 if (p->dirty == 0 || 430 p->queue != PQ_INACTIVE || 431 p->hold_count != 0) { /* may be undergoing I/O */ 432 vm_page_unlock(p); 433 break; 434 } 435 vm_page_unlock(p); 436 mc[page_base + pageout_count] = ps = p; 437 ++pageout_count; 438 ++is; 439 } 440 441 /* 442 * If we exhausted our forward scan, continue with the reverse scan 443 * when possible, even past a page boundry. This catches boundry 444 * conditions. 445 */ 446 if (ib && pageout_count < vm_pageout_page_count) 447 goto more; 448 449 /* 450 * we allow reads during pageouts... 451 */ 452 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL, 453 NULL)); 454 } 455 456 /* 457 * vm_pageout_flush() - launder the given pages 458 * 459 * The given pages are laundered. Note that we setup for the start of 460 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 461 * reference count all in here rather then in the parent. If we want 462 * the parent to do more sophisticated things we may have to change 463 * the ordering. 464 * 465 * Returned runlen is the count of pages between mreq and first 466 * page after mreq with status VM_PAGER_AGAIN. 467 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 468 * for any page in runlen set. 469 */ 470 int 471 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 472 boolean_t *eio) 473 { 474 vm_object_t object = mc[0]->object; 475 int pageout_status[count]; 476 int numpagedout = 0; 477 int i, runlen; 478 479 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 480 481 /* 482 * Initiate I/O. Bump the vm_page_t->busy counter and 483 * mark the pages read-only. 484 * 485 * We do not have to fixup the clean/dirty bits here... we can 486 * allow the pager to do it after the I/O completes. 487 * 488 * NOTE! mc[i]->dirty may be partial or fragmented due to an 489 * edge case with file fragments. 490 */ 491 for (i = 0; i < count; i++) { 492 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 493 ("vm_pageout_flush: partially invalid page %p index %d/%d", 494 mc[i], i, count)); 495 vm_page_io_start(mc[i]); 496 pmap_remove_write(mc[i]); 497 } 498 vm_object_pip_add(object, count); 499 500 vm_pager_put_pages(object, mc, count, flags, pageout_status); 501 502 runlen = count - mreq; 503 if (eio != NULL) 504 *eio = FALSE; 505 for (i = 0; i < count; i++) { 506 vm_page_t mt = mc[i]; 507 508 KASSERT(pageout_status[i] == VM_PAGER_PEND || 509 !pmap_page_is_write_mapped(mt), 510 ("vm_pageout_flush: page %p is not write protected", mt)); 511 switch (pageout_status[i]) { 512 case VM_PAGER_OK: 513 case VM_PAGER_PEND: 514 numpagedout++; 515 break; 516 case VM_PAGER_BAD: 517 /* 518 * Page outside of range of object. Right now we 519 * essentially lose the changes by pretending it 520 * worked. 521 */ 522 vm_page_undirty(mt); 523 break; 524 case VM_PAGER_ERROR: 525 case VM_PAGER_FAIL: 526 /* 527 * If page couldn't be paged out, then reactivate the 528 * page so it doesn't clog the inactive list. (We 529 * will try paging out it again later). 530 */ 531 vm_page_lock(mt); 532 vm_page_activate(mt); 533 vm_page_unlock(mt); 534 if (eio != NULL && i >= mreq && i - mreq < runlen) 535 *eio = TRUE; 536 break; 537 case VM_PAGER_AGAIN: 538 if (i >= mreq && i - mreq < runlen) 539 runlen = i - mreq; 540 break; 541 } 542 543 /* 544 * If the operation is still going, leave the page busy to 545 * block all other accesses. Also, leave the paging in 546 * progress indicator set so that we don't attempt an object 547 * collapse. 548 */ 549 if (pageout_status[i] != VM_PAGER_PEND) { 550 vm_object_pip_wakeup(object); 551 vm_page_io_finish(mt); 552 if (vm_page_count_severe()) { 553 vm_page_lock(mt); 554 vm_page_try_to_cache(mt); 555 vm_page_unlock(mt); 556 } 557 } 558 } 559 if (prunlen != NULL) 560 *prunlen = runlen; 561 return (numpagedout); 562 } 563 564 static boolean_t 565 vm_pageout_launder(int queue, int tries, vm_paddr_t low, vm_paddr_t high) 566 { 567 struct mount *mp; 568 struct vnode *vp; 569 vm_object_t object; 570 vm_paddr_t pa; 571 vm_page_t m, m_tmp, next; 572 int vfslocked; 573 574 vm_page_lock_queues(); 575 TAILQ_FOREACH_SAFE(m, &vm_page_queues[queue].pl, pageq, next) { 576 KASSERT(m->queue == queue, 577 ("vm_pageout_launder: page %p's queue is not %d", m, 578 queue)); 579 if ((m->flags & PG_MARKER) != 0) 580 continue; 581 pa = VM_PAGE_TO_PHYS(m); 582 if (pa < low || pa + PAGE_SIZE > high) 583 continue; 584 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) { 585 vm_page_unlock(m); 586 continue; 587 } 588 object = m->object; 589 if ((!VM_OBJECT_TRYLOCK(object) && 590 (!vm_pageout_fallback_object_lock(m, &next) || 591 m->hold_count != 0)) || (m->oflags & VPO_BUSY) != 0 || 592 m->busy != 0) { 593 vm_page_unlock(m); 594 VM_OBJECT_UNLOCK(object); 595 continue; 596 } 597 vm_page_test_dirty(m); 598 if (m->dirty == 0) 599 pmap_remove_all(m); 600 if (m->dirty != 0) { 601 vm_page_unlock(m); 602 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) { 603 VM_OBJECT_UNLOCK(object); 604 continue; 605 } 606 if (object->type == OBJT_VNODE) { 607 vm_page_unlock_queues(); 608 vp = object->handle; 609 vm_object_reference_locked(object); 610 VM_OBJECT_UNLOCK(object); 611 (void)vn_start_write(vp, &mp, V_WAIT); 612 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 613 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 614 VM_OBJECT_LOCK(object); 615 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 616 VM_OBJECT_UNLOCK(object); 617 VOP_UNLOCK(vp, 0); 618 VFS_UNLOCK_GIANT(vfslocked); 619 vm_object_deallocate(object); 620 vn_finished_write(mp); 621 return (TRUE); 622 } else if (object->type == OBJT_SWAP || 623 object->type == OBJT_DEFAULT) { 624 vm_page_unlock_queues(); 625 m_tmp = m; 626 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC, 627 0, NULL, NULL); 628 VM_OBJECT_UNLOCK(object); 629 return (TRUE); 630 } 631 } else { 632 vm_page_cache(m); 633 vm_page_unlock(m); 634 } 635 VM_OBJECT_UNLOCK(object); 636 } 637 vm_page_unlock_queues(); 638 return (FALSE); 639 } 640 641 /* 642 * Increase the number of cached pages. The specified value, "tries", 643 * determines which categories of pages are cached: 644 * 645 * 0: All clean, inactive pages within the specified physical address range 646 * are cached. Will not sleep. 647 * 1: The vm_lowmem handlers are called. All inactive pages within 648 * the specified physical address range are cached. May sleep. 649 * 2: The vm_lowmem handlers are called. All inactive and active pages 650 * within the specified physical address range are cached. May sleep. 651 */ 652 void 653 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high) 654 { 655 int actl, actmax, inactl, inactmax; 656 657 if (tries > 0) { 658 /* 659 * Decrease registered cache sizes. The vm_lowmem handlers 660 * may acquire locks and/or sleep, so they can only be invoked 661 * when "tries" is greater than zero. 662 */ 663 EVENTHANDLER_INVOKE(vm_lowmem, 0); 664 665 /* 666 * We do this explicitly after the caches have been drained 667 * above. 668 */ 669 uma_reclaim(); 670 } 671 inactl = 0; 672 inactmax = cnt.v_inactive_count; 673 actl = 0; 674 actmax = tries < 2 ? 0 : cnt.v_active_count; 675 again: 676 if (inactl < inactmax && vm_pageout_launder(PQ_INACTIVE, tries, low, 677 high)) { 678 inactl++; 679 goto again; 680 } 681 if (actl < actmax && vm_pageout_launder(PQ_ACTIVE, tries, low, high)) { 682 actl++; 683 goto again; 684 } 685 } 686 687 #if !defined(NO_SWAPPING) 688 /* 689 * vm_pageout_object_deactivate_pages 690 * 691 * Deactivate enough pages to satisfy the inactive target 692 * requirements. 693 * 694 * The object and map must be locked. 695 */ 696 static void 697 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 698 long desired) 699 { 700 vm_object_t backing_object, object; 701 vm_page_t p; 702 int actcount, remove_mode; 703 704 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 705 if (first_object->type == OBJT_DEVICE || 706 first_object->type == OBJT_SG) 707 return; 708 for (object = first_object;; object = backing_object) { 709 if (pmap_resident_count(pmap) <= desired) 710 goto unlock_return; 711 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 712 if (object->type == OBJT_PHYS || object->paging_in_progress) 713 goto unlock_return; 714 715 remove_mode = 0; 716 if (object->shadow_count > 1) 717 remove_mode = 1; 718 /* 719 * Scan the object's entire memory queue. 720 */ 721 TAILQ_FOREACH(p, &object->memq, listq) { 722 if (pmap_resident_count(pmap) <= desired) 723 goto unlock_return; 724 if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0) 725 continue; 726 PCPU_INC(cnt.v_pdpages); 727 vm_page_lock(p); 728 if (p->wire_count != 0 || p->hold_count != 0 || 729 !pmap_page_exists_quick(pmap, p)) { 730 vm_page_unlock(p); 731 continue; 732 } 733 actcount = pmap_ts_referenced(p); 734 if ((p->aflags & PGA_REFERENCED) != 0) { 735 if (actcount == 0) 736 actcount = 1; 737 vm_page_aflag_clear(p, PGA_REFERENCED); 738 } 739 if (p->queue != PQ_ACTIVE && actcount != 0) { 740 vm_page_activate(p); 741 p->act_count += actcount; 742 } else if (p->queue == PQ_ACTIVE) { 743 if (actcount == 0) { 744 p->act_count -= min(p->act_count, 745 ACT_DECLINE); 746 if (!remove_mode && 747 (vm_pageout_algorithm || 748 p->act_count == 0)) { 749 pmap_remove_all(p); 750 vm_page_deactivate(p); 751 } else { 752 vm_page_lock_queues(); 753 vm_pageout_requeue(p); 754 vm_page_unlock_queues(); 755 } 756 } else { 757 vm_page_activate(p); 758 if (p->act_count < ACT_MAX - 759 ACT_ADVANCE) 760 p->act_count += ACT_ADVANCE; 761 vm_page_lock_queues(); 762 vm_pageout_requeue(p); 763 vm_page_unlock_queues(); 764 } 765 } else if (p->queue == PQ_INACTIVE) 766 pmap_remove_all(p); 767 vm_page_unlock(p); 768 } 769 if ((backing_object = object->backing_object) == NULL) 770 goto unlock_return; 771 VM_OBJECT_LOCK(backing_object); 772 if (object != first_object) 773 VM_OBJECT_UNLOCK(object); 774 } 775 unlock_return: 776 if (object != first_object) 777 VM_OBJECT_UNLOCK(object); 778 } 779 780 /* 781 * deactivate some number of pages in a map, try to do it fairly, but 782 * that is really hard to do. 783 */ 784 static void 785 vm_pageout_map_deactivate_pages(map, desired) 786 vm_map_t map; 787 long desired; 788 { 789 vm_map_entry_t tmpe; 790 vm_object_t obj, bigobj; 791 int nothingwired; 792 793 if (!vm_map_trylock(map)) 794 return; 795 796 bigobj = NULL; 797 nothingwired = TRUE; 798 799 /* 800 * first, search out the biggest object, and try to free pages from 801 * that. 802 */ 803 tmpe = map->header.next; 804 while (tmpe != &map->header) { 805 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 806 obj = tmpe->object.vm_object; 807 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 808 if (obj->shadow_count <= 1 && 809 (bigobj == NULL || 810 bigobj->resident_page_count < obj->resident_page_count)) { 811 if (bigobj != NULL) 812 VM_OBJECT_UNLOCK(bigobj); 813 bigobj = obj; 814 } else 815 VM_OBJECT_UNLOCK(obj); 816 } 817 } 818 if (tmpe->wired_count > 0) 819 nothingwired = FALSE; 820 tmpe = tmpe->next; 821 } 822 823 if (bigobj != NULL) { 824 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 825 VM_OBJECT_UNLOCK(bigobj); 826 } 827 /* 828 * Next, hunt around for other pages to deactivate. We actually 829 * do this search sort of wrong -- .text first is not the best idea. 830 */ 831 tmpe = map->header.next; 832 while (tmpe != &map->header) { 833 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 834 break; 835 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 836 obj = tmpe->object.vm_object; 837 if (obj != NULL) { 838 VM_OBJECT_LOCK(obj); 839 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 840 VM_OBJECT_UNLOCK(obj); 841 } 842 } 843 tmpe = tmpe->next; 844 } 845 846 /* 847 * Remove all mappings if a process is swapped out, this will free page 848 * table pages. 849 */ 850 if (desired == 0 && nothingwired) { 851 pmap_remove(vm_map_pmap(map), vm_map_min(map), 852 vm_map_max(map)); 853 } 854 vm_map_unlock(map); 855 } 856 #endif /* !defined(NO_SWAPPING) */ 857 858 /* 859 * vm_pageout_requeue: 860 * 861 * Move the specified page to the tail of its present page queue. 862 * 863 * The page queues must be locked. 864 */ 865 static void 866 vm_pageout_requeue(vm_page_t m) 867 { 868 struct vpgqueues *vpq; 869 870 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 871 KASSERT(m->queue != PQ_NONE, 872 ("vm_pageout_requeue: page %p is not queued", m)); 873 vpq = &vm_page_queues[m->queue]; 874 TAILQ_REMOVE(&vpq->pl, m, pageq); 875 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 876 } 877 878 /* 879 * vm_pageout_scan does the dirty work for the pageout daemon. 880 */ 881 static void 882 vm_pageout_scan(int pass) 883 { 884 vm_page_t m, next; 885 struct vm_page marker; 886 int page_shortage, maxscan, pcount; 887 int addl_page_shortage; 888 vm_object_t object; 889 int actcount; 890 int vnodes_skipped = 0; 891 int maxlaunder; 892 boolean_t queues_locked; 893 894 /* 895 * Decrease registered cache sizes. 896 */ 897 EVENTHANDLER_INVOKE(vm_lowmem, 0); 898 /* 899 * We do this explicitly after the caches have been drained above. 900 */ 901 uma_reclaim(); 902 903 /* 904 * The addl_page_shortage is the number of temporarily 905 * stuck pages in the inactive queue. In other words, the 906 * number of pages from cnt.v_inactive_count that should be 907 * discounted in setting the target for the active queue scan. 908 */ 909 addl_page_shortage = atomic_readandclear_int(&vm_pageout_deficit); 910 911 /* 912 * Calculate the number of pages we want to either free or move 913 * to the cache. 914 */ 915 page_shortage = vm_paging_target() + addl_page_shortage; 916 917 vm_pageout_init_marker(&marker, PQ_INACTIVE); 918 919 /* 920 * Start scanning the inactive queue for pages we can move to the 921 * cache or free. The scan will stop when the target is reached or 922 * we have scanned the entire inactive queue. Note that m->act_count 923 * is not used to form decisions for the inactive queue, only for the 924 * active queue. 925 * 926 * maxlaunder limits the number of dirty pages we flush per scan. 927 * For most systems a smaller value (16 or 32) is more robust under 928 * extreme memory and disk pressure because any unnecessary writes 929 * to disk can result in extreme performance degredation. However, 930 * systems with excessive dirty pages (especially when MAP_NOSYNC is 931 * used) will die horribly with limited laundering. If the pageout 932 * daemon cannot clean enough pages in the first pass, we let it go 933 * all out in succeeding passes. 934 */ 935 if ((maxlaunder = vm_max_launder) <= 1) 936 maxlaunder = 1; 937 if (pass) 938 maxlaunder = 10000; 939 vm_page_lock_queues(); 940 queues_locked = TRUE; 941 maxscan = cnt.v_inactive_count; 942 943 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 944 m != NULL && maxscan-- > 0 && page_shortage > 0; 945 m = next) { 946 KASSERT(queues_locked, ("unlocked queues")); 947 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 948 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m)); 949 950 cnt.v_pdpages++; 951 next = TAILQ_NEXT(m, pageq); 952 953 /* 954 * skip marker pages 955 */ 956 if (m->flags & PG_MARKER) 957 continue; 958 959 KASSERT((m->flags & PG_FICTITIOUS) == 0, 960 ("Fictitious page %p cannot be in inactive queue", m)); 961 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 962 ("Unmanaged page %p cannot be in inactive queue", m)); 963 964 /* 965 * The page or object lock acquisitions fail if the 966 * page was removed from the queue or moved to a 967 * different position within the queue. In either 968 * case, addl_page_shortage should not be incremented. 969 */ 970 if (!vm_pageout_page_lock(m, &next)) { 971 vm_page_unlock(m); 972 continue; 973 } 974 object = m->object; 975 if (!VM_OBJECT_TRYLOCK(object) && 976 !vm_pageout_fallback_object_lock(m, &next)) { 977 vm_page_unlock(m); 978 VM_OBJECT_UNLOCK(object); 979 continue; 980 } 981 982 /* 983 * Don't mess with busy pages, keep them at at the 984 * front of the queue, most likely they are being 985 * paged out. Increment addl_page_shortage for busy 986 * pages, because they may leave the inactive queue 987 * shortly after page scan is finished. 988 */ 989 if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0) { 990 vm_page_unlock(m); 991 VM_OBJECT_UNLOCK(object); 992 addl_page_shortage++; 993 continue; 994 } 995 996 /* 997 * We unlock vm_page_queue_mtx, invalidating the 998 * 'next' pointer. Use our marker to remember our 999 * place. 1000 */ 1001 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 1002 m, &marker, pageq); 1003 vm_page_unlock_queues(); 1004 queues_locked = FALSE; 1005 1006 /* 1007 * If the object is not being used, we ignore previous 1008 * references. 1009 */ 1010 if (object->ref_count == 0) { 1011 vm_page_aflag_clear(m, PGA_REFERENCED); 1012 KASSERT(!pmap_page_is_mapped(m), 1013 ("vm_pageout_scan: page %p is mapped", m)); 1014 1015 /* 1016 * Otherwise, if the page has been referenced while in the 1017 * inactive queue, we bump the "activation count" upwards, 1018 * making it less likely that the page will be added back to 1019 * the inactive queue prematurely again. Here we check the 1020 * page tables (or emulated bits, if any), given the upper 1021 * level VM system not knowing anything about existing 1022 * references. 1023 */ 1024 } else if ((m->aflags & PGA_REFERENCED) == 0 && 1025 (actcount = pmap_ts_referenced(m)) != 0) { 1026 vm_page_activate(m); 1027 vm_page_unlock(m); 1028 m->act_count += actcount + ACT_ADVANCE; 1029 VM_OBJECT_UNLOCK(object); 1030 goto relock_queues; 1031 } 1032 1033 /* 1034 * If the upper level VM system knows about any page 1035 * references, we activate the page. We also set the 1036 * "activation count" higher than normal so that we will less 1037 * likely place pages back onto the inactive queue again. 1038 */ 1039 if ((m->aflags & PGA_REFERENCED) != 0) { 1040 vm_page_aflag_clear(m, PGA_REFERENCED); 1041 actcount = pmap_ts_referenced(m); 1042 vm_page_activate(m); 1043 vm_page_unlock(m); 1044 m->act_count += actcount + ACT_ADVANCE + 1; 1045 VM_OBJECT_UNLOCK(object); 1046 goto relock_queues; 1047 } 1048 1049 if (m->hold_count != 0) { 1050 vm_page_unlock(m); 1051 VM_OBJECT_UNLOCK(object); 1052 1053 /* 1054 * Held pages are essentially stuck in the 1055 * queue. So, they ought to be discounted 1056 * from cnt.v_inactive_count. See the 1057 * calculation of the page_shortage for the 1058 * loop over the active queue below. 1059 */ 1060 addl_page_shortage++; 1061 goto relock_queues; 1062 } 1063 1064 /* 1065 * If the upper level VM system does not believe that the page 1066 * is fully dirty, but it is mapped for write access, then we 1067 * consult the pmap to see if the page's dirty status should 1068 * be updated. 1069 */ 1070 if (m->dirty != VM_PAGE_BITS_ALL && 1071 pmap_page_is_write_mapped(m)) { 1072 /* 1073 * Avoid a race condition: Unless write access is 1074 * removed from the page, another processor could 1075 * modify it before all access is removed by the call 1076 * to vm_page_cache() below. If vm_page_cache() finds 1077 * that the page has been modified when it removes all 1078 * access, it panics because it cannot cache dirty 1079 * pages. In principle, we could eliminate just write 1080 * access here rather than all access. In the expected 1081 * case, when there are no last instant modifications 1082 * to the page, removing all access will be cheaper 1083 * overall. 1084 */ 1085 if (pmap_is_modified(m)) 1086 vm_page_dirty(m); 1087 else if (m->dirty == 0) 1088 pmap_remove_all(m); 1089 } 1090 1091 if (m->valid == 0) { 1092 /* 1093 * Invalid pages can be easily freed 1094 */ 1095 vm_page_free(m); 1096 PCPU_INC(cnt.v_dfree); 1097 --page_shortage; 1098 } else if (m->dirty == 0) { 1099 /* 1100 * Clean pages can be placed onto the cache queue. 1101 * This effectively frees them. 1102 */ 1103 vm_page_cache(m); 1104 --page_shortage; 1105 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 1106 /* 1107 * Dirty pages need to be paged out, but flushing 1108 * a page is extremely expensive verses freeing 1109 * a clean page. Rather then artificially limiting 1110 * the number of pages we can flush, we instead give 1111 * dirty pages extra priority on the inactive queue 1112 * by forcing them to be cycled through the queue 1113 * twice before being flushed, after which the 1114 * (now clean) page will cycle through once more 1115 * before being freed. This significantly extends 1116 * the thrash point for a heavily loaded machine. 1117 */ 1118 m->flags |= PG_WINATCFLS; 1119 vm_page_lock_queues(); 1120 queues_locked = TRUE; 1121 vm_pageout_requeue(m); 1122 } else if (maxlaunder > 0) { 1123 /* 1124 * We always want to try to flush some dirty pages if 1125 * we encounter them, to keep the system stable. 1126 * Normally this number is small, but under extreme 1127 * pressure where there are insufficient clean pages 1128 * on the inactive queue, we may have to go all out. 1129 */ 1130 int swap_pageouts_ok, vfslocked = 0; 1131 struct vnode *vp = NULL; 1132 struct mount *mp = NULL; 1133 1134 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 1135 swap_pageouts_ok = 1; 1136 } else { 1137 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 1138 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 1139 vm_page_count_min()); 1140 1141 } 1142 1143 /* 1144 * We don't bother paging objects that are "dead". 1145 * Those objects are in a "rundown" state. 1146 */ 1147 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 1148 vm_page_lock_queues(); 1149 vm_page_unlock(m); 1150 VM_OBJECT_UNLOCK(object); 1151 queues_locked = TRUE; 1152 vm_pageout_requeue(m); 1153 goto relock_queues; 1154 } 1155 1156 /* 1157 * The object is already known NOT to be dead. It 1158 * is possible for the vget() to block the whole 1159 * pageout daemon, but the new low-memory handling 1160 * code should prevent it. 1161 * 1162 * The previous code skipped locked vnodes and, worse, 1163 * reordered pages in the queue. This results in 1164 * completely non-deterministic operation and, on a 1165 * busy system, can lead to extremely non-optimal 1166 * pageouts. For example, it can cause clean pages 1167 * to be freed and dirty pages to be moved to the end 1168 * of the queue. Since dirty pages are also moved to 1169 * the end of the queue once-cleaned, this gives 1170 * way too large a weighting to defering the freeing 1171 * of dirty pages. 1172 * 1173 * We can't wait forever for the vnode lock, we might 1174 * deadlock due to a vn_read() getting stuck in 1175 * vm_wait while holding this vnode. We skip the 1176 * vnode if we can't get it in a reasonable amount 1177 * of time. 1178 */ 1179 if (object->type == OBJT_VNODE) { 1180 vm_page_unlock(m); 1181 vp = object->handle; 1182 if (vp->v_type == VREG && 1183 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1184 mp = NULL; 1185 ++pageout_lock_miss; 1186 if (object->flags & OBJ_MIGHTBEDIRTY) 1187 vnodes_skipped++; 1188 goto unlock_and_continue; 1189 } 1190 KASSERT(mp != NULL, 1191 ("vp %p with NULL v_mount", vp)); 1192 vm_object_reference_locked(object); 1193 VM_OBJECT_UNLOCK(object); 1194 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 1195 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 1196 curthread)) { 1197 VM_OBJECT_LOCK(object); 1198 ++pageout_lock_miss; 1199 if (object->flags & OBJ_MIGHTBEDIRTY) 1200 vnodes_skipped++; 1201 vp = NULL; 1202 goto unlock_and_continue; 1203 } 1204 VM_OBJECT_LOCK(object); 1205 vm_page_lock(m); 1206 vm_page_lock_queues(); 1207 queues_locked = TRUE; 1208 /* 1209 * The page might have been moved to another 1210 * queue during potential blocking in vget() 1211 * above. The page might have been freed and 1212 * reused for another vnode. 1213 */ 1214 if (m->queue != PQ_INACTIVE || 1215 m->object != object || 1216 TAILQ_NEXT(m, pageq) != &marker) { 1217 vm_page_unlock(m); 1218 if (object->flags & OBJ_MIGHTBEDIRTY) 1219 vnodes_skipped++; 1220 goto unlock_and_continue; 1221 } 1222 1223 /* 1224 * The page may have been busied during the 1225 * blocking in vget(). We don't move the 1226 * page back onto the end of the queue so that 1227 * statistics are more correct if we don't. 1228 */ 1229 if (m->busy || (m->oflags & VPO_BUSY)) { 1230 vm_page_unlock(m); 1231 goto unlock_and_continue; 1232 } 1233 1234 /* 1235 * If the page has become held it might 1236 * be undergoing I/O, so skip it 1237 */ 1238 if (m->hold_count) { 1239 vm_page_unlock(m); 1240 vm_pageout_requeue(m); 1241 if (object->flags & OBJ_MIGHTBEDIRTY) 1242 vnodes_skipped++; 1243 goto unlock_and_continue; 1244 } 1245 vm_page_unlock_queues(); 1246 queues_locked = FALSE; 1247 } 1248 1249 /* 1250 * If a page is dirty, then it is either being washed 1251 * (but not yet cleaned) or it is still in the 1252 * laundry. If it is still in the laundry, then we 1253 * start the cleaning operation. 1254 * 1255 * decrement page_shortage on success to account for 1256 * the (future) cleaned page. Otherwise we could wind 1257 * up laundering or cleaning too many pages. 1258 */ 1259 if (vm_pageout_clean(m) != 0) { 1260 --page_shortage; 1261 --maxlaunder; 1262 } 1263 unlock_and_continue: 1264 vm_page_lock_assert(m, MA_NOTOWNED); 1265 VM_OBJECT_UNLOCK(object); 1266 if (mp != NULL) { 1267 if (queues_locked) { 1268 vm_page_unlock_queues(); 1269 queues_locked = FALSE; 1270 } 1271 if (vp != NULL) 1272 vput(vp); 1273 VFS_UNLOCK_GIANT(vfslocked); 1274 vm_object_deallocate(object); 1275 vn_finished_write(mp); 1276 } 1277 vm_page_lock_assert(m, MA_NOTOWNED); 1278 goto relock_queues; 1279 } 1280 vm_page_unlock(m); 1281 VM_OBJECT_UNLOCK(object); 1282 relock_queues: 1283 if (!queues_locked) { 1284 vm_page_lock_queues(); 1285 queues_locked = TRUE; 1286 } 1287 next = TAILQ_NEXT(&marker, pageq); 1288 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1289 &marker, pageq); 1290 } 1291 1292 /* 1293 * Compute the number of pages we want to try to move from the 1294 * active queue to the inactive queue. 1295 */ 1296 page_shortage = vm_paging_target() + 1297 cnt.v_inactive_target - cnt.v_inactive_count; 1298 page_shortage += addl_page_shortage; 1299 1300 /* 1301 * Scan the active queue for things we can deactivate. We nominally 1302 * track the per-page activity counter and use it to locate 1303 * deactivation candidates. 1304 */ 1305 pcount = cnt.v_active_count; 1306 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1307 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1308 1309 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1310 1311 KASSERT(m->queue == PQ_ACTIVE, 1312 ("vm_pageout_scan: page %p isn't active", m)); 1313 1314 next = TAILQ_NEXT(m, pageq); 1315 if ((m->flags & PG_MARKER) != 0) { 1316 m = next; 1317 continue; 1318 } 1319 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1320 ("Fictitious page %p cannot be in active queue", m)); 1321 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1322 ("Unmanaged page %p cannot be in active queue", m)); 1323 if (!vm_pageout_page_lock(m, &next)) { 1324 vm_page_unlock(m); 1325 m = next; 1326 continue; 1327 } 1328 object = m->object; 1329 if (!VM_OBJECT_TRYLOCK(object) && 1330 !vm_pageout_fallback_object_lock(m, &next)) { 1331 VM_OBJECT_UNLOCK(object); 1332 vm_page_unlock(m); 1333 m = next; 1334 continue; 1335 } 1336 1337 /* 1338 * Don't deactivate pages that are busy. 1339 */ 1340 if ((m->busy != 0) || 1341 (m->oflags & VPO_BUSY) || 1342 (m->hold_count != 0)) { 1343 vm_page_unlock(m); 1344 VM_OBJECT_UNLOCK(object); 1345 vm_pageout_requeue(m); 1346 m = next; 1347 continue; 1348 } 1349 1350 /* 1351 * The count for pagedaemon pages is done after checking the 1352 * page for eligibility... 1353 */ 1354 cnt.v_pdpages++; 1355 1356 /* 1357 * Check to see "how much" the page has been used. 1358 */ 1359 actcount = 0; 1360 if (object->ref_count != 0) { 1361 if (m->aflags & PGA_REFERENCED) { 1362 actcount += 1; 1363 } 1364 actcount += pmap_ts_referenced(m); 1365 if (actcount) { 1366 m->act_count += ACT_ADVANCE + actcount; 1367 if (m->act_count > ACT_MAX) 1368 m->act_count = ACT_MAX; 1369 } 1370 } 1371 1372 /* 1373 * Since we have "tested" this bit, we need to clear it now. 1374 */ 1375 vm_page_aflag_clear(m, PGA_REFERENCED); 1376 1377 /* 1378 * Only if an object is currently being used, do we use the 1379 * page activation count stats. 1380 */ 1381 if (actcount && (object->ref_count != 0)) { 1382 vm_pageout_requeue(m); 1383 } else { 1384 m->act_count -= min(m->act_count, ACT_DECLINE); 1385 if (vm_pageout_algorithm || 1386 object->ref_count == 0 || 1387 m->act_count == 0) { 1388 page_shortage--; 1389 if (object->ref_count == 0) { 1390 KASSERT(!pmap_page_is_mapped(m), 1391 ("vm_pageout_scan: page %p is mapped", m)); 1392 if (m->dirty == 0) 1393 vm_page_cache(m); 1394 else 1395 vm_page_deactivate(m); 1396 } else { 1397 vm_page_deactivate(m); 1398 } 1399 } else { 1400 vm_pageout_requeue(m); 1401 } 1402 } 1403 vm_page_unlock(m); 1404 VM_OBJECT_UNLOCK(object); 1405 m = next; 1406 } 1407 vm_page_unlock_queues(); 1408 #if !defined(NO_SWAPPING) 1409 /* 1410 * Idle process swapout -- run once per second. 1411 */ 1412 if (vm_swap_idle_enabled) { 1413 static long lsec; 1414 if (time_second != lsec) { 1415 vm_req_vmdaemon(VM_SWAP_IDLE); 1416 lsec = time_second; 1417 } 1418 } 1419 #endif 1420 1421 /* 1422 * If we didn't get enough free pages, and we have skipped a vnode 1423 * in a writeable object, wakeup the sync daemon. And kick swapout 1424 * if we did not get enough free pages. 1425 */ 1426 if (vm_paging_target() > 0) { 1427 if (vnodes_skipped && vm_page_count_min()) 1428 (void) speedup_syncer(); 1429 #if !defined(NO_SWAPPING) 1430 if (vm_swap_enabled && vm_page_count_target()) 1431 vm_req_vmdaemon(VM_SWAP_NORMAL); 1432 #endif 1433 } 1434 1435 /* 1436 * If we are critically low on one of RAM or swap and low on 1437 * the other, kill the largest process. However, we avoid 1438 * doing this on the first pass in order to give ourselves a 1439 * chance to flush out dirty vnode-backed pages and to allow 1440 * active pages to be moved to the inactive queue and reclaimed. 1441 */ 1442 if (pass != 0 && 1443 ((swap_pager_avail < 64 && vm_page_count_min()) || 1444 (swap_pager_full && vm_paging_target() > 0))) 1445 vm_pageout_oom(VM_OOM_MEM); 1446 } 1447 1448 1449 void 1450 vm_pageout_oom(int shortage) 1451 { 1452 struct proc *p, *bigproc; 1453 vm_offset_t size, bigsize; 1454 struct thread *td; 1455 struct vmspace *vm; 1456 1457 /* 1458 * We keep the process bigproc locked once we find it to keep anyone 1459 * from messing with it; however, there is a possibility of 1460 * deadlock if process B is bigproc and one of it's child processes 1461 * attempts to propagate a signal to B while we are waiting for A's 1462 * lock while walking this list. To avoid this, we don't block on 1463 * the process lock but just skip a process if it is already locked. 1464 */ 1465 bigproc = NULL; 1466 bigsize = 0; 1467 sx_slock(&allproc_lock); 1468 FOREACH_PROC_IN_SYSTEM(p) { 1469 int breakout; 1470 1471 if (PROC_TRYLOCK(p) == 0) 1472 continue; 1473 /* 1474 * If this is a system, protected or killed process, skip it. 1475 */ 1476 if (p->p_state != PRS_NORMAL || 1477 (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1478 (p->p_pid == 1) || P_KILLED(p) || 1479 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1480 PROC_UNLOCK(p); 1481 continue; 1482 } 1483 /* 1484 * If the process is in a non-running type state, 1485 * don't touch it. Check all the threads individually. 1486 */ 1487 breakout = 0; 1488 FOREACH_THREAD_IN_PROC(p, td) { 1489 thread_lock(td); 1490 if (!TD_ON_RUNQ(td) && 1491 !TD_IS_RUNNING(td) && 1492 !TD_IS_SLEEPING(td) && 1493 !TD_IS_SUSPENDED(td)) { 1494 thread_unlock(td); 1495 breakout = 1; 1496 break; 1497 } 1498 thread_unlock(td); 1499 } 1500 if (breakout) { 1501 PROC_UNLOCK(p); 1502 continue; 1503 } 1504 /* 1505 * get the process size 1506 */ 1507 vm = vmspace_acquire_ref(p); 1508 if (vm == NULL) { 1509 PROC_UNLOCK(p); 1510 continue; 1511 } 1512 if (!vm_map_trylock_read(&vm->vm_map)) { 1513 vmspace_free(vm); 1514 PROC_UNLOCK(p); 1515 continue; 1516 } 1517 size = vmspace_swap_count(vm); 1518 vm_map_unlock_read(&vm->vm_map); 1519 if (shortage == VM_OOM_MEM) 1520 size += vmspace_resident_count(vm); 1521 vmspace_free(vm); 1522 /* 1523 * if the this process is bigger than the biggest one 1524 * remember it. 1525 */ 1526 if (size > bigsize) { 1527 if (bigproc != NULL) 1528 PROC_UNLOCK(bigproc); 1529 bigproc = p; 1530 bigsize = size; 1531 } else 1532 PROC_UNLOCK(p); 1533 } 1534 sx_sunlock(&allproc_lock); 1535 if (bigproc != NULL) { 1536 killproc(bigproc, "out of swap space"); 1537 sched_nice(bigproc, PRIO_MIN); 1538 PROC_UNLOCK(bigproc); 1539 wakeup(&cnt.v_free_count); 1540 } 1541 } 1542 1543 /* 1544 * This routine tries to maintain the pseudo LRU active queue, 1545 * so that during long periods of time where there is no paging, 1546 * that some statistic accumulation still occurs. This code 1547 * helps the situation where paging just starts to occur. 1548 */ 1549 static void 1550 vm_pageout_page_stats() 1551 { 1552 vm_object_t object; 1553 vm_page_t m,next; 1554 int pcount,tpcount; /* Number of pages to check */ 1555 static int fullintervalcount = 0; 1556 int page_shortage; 1557 1558 page_shortage = 1559 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1560 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1561 1562 if (page_shortage <= 0) 1563 return; 1564 1565 vm_page_lock_queues(); 1566 pcount = cnt.v_active_count; 1567 fullintervalcount += vm_pageout_stats_interval; 1568 if (fullintervalcount < vm_pageout_full_stats_interval) { 1569 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1570 cnt.v_page_count; 1571 if (pcount > tpcount) 1572 pcount = tpcount; 1573 } else { 1574 fullintervalcount = 0; 1575 } 1576 1577 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1578 while ((m != NULL) && (pcount-- > 0)) { 1579 int actcount; 1580 1581 KASSERT(m->queue == PQ_ACTIVE, 1582 ("vm_pageout_page_stats: page %p isn't active", m)); 1583 1584 next = TAILQ_NEXT(m, pageq); 1585 if ((m->flags & PG_MARKER) != 0) { 1586 m = next; 1587 continue; 1588 } 1589 vm_page_lock_assert(m, MA_NOTOWNED); 1590 if (!vm_pageout_page_lock(m, &next)) { 1591 vm_page_unlock(m); 1592 m = next; 1593 continue; 1594 } 1595 object = m->object; 1596 if (!VM_OBJECT_TRYLOCK(object) && 1597 !vm_pageout_fallback_object_lock(m, &next)) { 1598 VM_OBJECT_UNLOCK(object); 1599 vm_page_unlock(m); 1600 m = next; 1601 continue; 1602 } 1603 1604 /* 1605 * Don't deactivate pages that are busy. 1606 */ 1607 if ((m->busy != 0) || 1608 (m->oflags & VPO_BUSY) || 1609 (m->hold_count != 0)) { 1610 vm_page_unlock(m); 1611 VM_OBJECT_UNLOCK(object); 1612 vm_pageout_requeue(m); 1613 m = next; 1614 continue; 1615 } 1616 1617 actcount = 0; 1618 if (m->aflags & PGA_REFERENCED) { 1619 vm_page_aflag_clear(m, PGA_REFERENCED); 1620 actcount += 1; 1621 } 1622 1623 actcount += pmap_ts_referenced(m); 1624 if (actcount) { 1625 m->act_count += ACT_ADVANCE + actcount; 1626 if (m->act_count > ACT_MAX) 1627 m->act_count = ACT_MAX; 1628 vm_pageout_requeue(m); 1629 } else { 1630 if (m->act_count == 0) { 1631 /* 1632 * We turn off page access, so that we have 1633 * more accurate RSS stats. We don't do this 1634 * in the normal page deactivation when the 1635 * system is loaded VM wise, because the 1636 * cost of the large number of page protect 1637 * operations would be higher than the value 1638 * of doing the operation. 1639 */ 1640 pmap_remove_all(m); 1641 vm_page_deactivate(m); 1642 } else { 1643 m->act_count -= min(m->act_count, ACT_DECLINE); 1644 vm_pageout_requeue(m); 1645 } 1646 } 1647 vm_page_unlock(m); 1648 VM_OBJECT_UNLOCK(object); 1649 m = next; 1650 } 1651 vm_page_unlock_queues(); 1652 } 1653 1654 /* 1655 * vm_pageout is the high level pageout daemon. 1656 */ 1657 static void 1658 vm_pageout() 1659 { 1660 int error, pass; 1661 1662 /* 1663 * Initialize some paging parameters. 1664 */ 1665 cnt.v_interrupt_free_min = 2; 1666 if (cnt.v_page_count < 2000) 1667 vm_pageout_page_count = 8; 1668 1669 /* 1670 * v_free_reserved needs to include enough for the largest 1671 * swap pager structures plus enough for any pv_entry structs 1672 * when paging. 1673 */ 1674 if (cnt.v_page_count > 1024) 1675 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1676 else 1677 cnt.v_free_min = 4; 1678 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1679 cnt.v_interrupt_free_min; 1680 cnt.v_free_reserved = vm_pageout_page_count + 1681 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1682 cnt.v_free_severe = cnt.v_free_min / 2; 1683 cnt.v_free_min += cnt.v_free_reserved; 1684 cnt.v_free_severe += cnt.v_free_reserved; 1685 1686 /* 1687 * v_free_target and v_cache_min control pageout hysteresis. Note 1688 * that these are more a measure of the VM cache queue hysteresis 1689 * then the VM free queue. Specifically, v_free_target is the 1690 * high water mark (free+cache pages). 1691 * 1692 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1693 * low water mark, while v_free_min is the stop. v_cache_min must 1694 * be big enough to handle memory needs while the pageout daemon 1695 * is signalled and run to free more pages. 1696 */ 1697 if (cnt.v_free_count > 6144) 1698 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1699 else 1700 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1701 1702 if (cnt.v_free_count > 2048) { 1703 cnt.v_cache_min = cnt.v_free_target; 1704 cnt.v_cache_max = 2 * cnt.v_cache_min; 1705 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1706 } else { 1707 cnt.v_cache_min = 0; 1708 cnt.v_cache_max = 0; 1709 cnt.v_inactive_target = cnt.v_free_count / 4; 1710 } 1711 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1712 cnt.v_inactive_target = cnt.v_free_count / 3; 1713 1714 /* XXX does not really belong here */ 1715 if (vm_page_max_wired == 0) 1716 vm_page_max_wired = cnt.v_free_count / 3; 1717 1718 if (vm_pageout_stats_max == 0) 1719 vm_pageout_stats_max = cnt.v_free_target; 1720 1721 /* 1722 * Set interval in seconds for stats scan. 1723 */ 1724 if (vm_pageout_stats_interval == 0) 1725 vm_pageout_stats_interval = 5; 1726 if (vm_pageout_full_stats_interval == 0) 1727 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1728 1729 swap_pager_swap_init(); 1730 pass = 0; 1731 /* 1732 * The pageout daemon is never done, so loop forever. 1733 */ 1734 while (TRUE) { 1735 /* 1736 * If we have enough free memory, wakeup waiters. Do 1737 * not clear vm_pages_needed until we reach our target, 1738 * otherwise we may be woken up over and over again and 1739 * waste a lot of cpu. 1740 */ 1741 mtx_lock(&vm_page_queue_free_mtx); 1742 if (vm_pages_needed && !vm_page_count_min()) { 1743 if (!vm_paging_needed()) 1744 vm_pages_needed = 0; 1745 wakeup(&cnt.v_free_count); 1746 } 1747 if (vm_pages_needed) { 1748 /* 1749 * Still not done, take a second pass without waiting 1750 * (unlimited dirty cleaning), otherwise sleep a bit 1751 * and try again. 1752 */ 1753 ++pass; 1754 if (pass > 1) 1755 msleep(&vm_pages_needed, 1756 &vm_page_queue_free_mtx, PVM, "psleep", 1757 hz / 2); 1758 } else { 1759 /* 1760 * Good enough, sleep & handle stats. Prime the pass 1761 * for the next run. 1762 */ 1763 if (pass > 1) 1764 pass = 1; 1765 else 1766 pass = 0; 1767 error = msleep(&vm_pages_needed, 1768 &vm_page_queue_free_mtx, PVM, "psleep", 1769 vm_pageout_stats_interval * hz); 1770 if (error && !vm_pages_needed) { 1771 mtx_unlock(&vm_page_queue_free_mtx); 1772 pass = 0; 1773 vm_pageout_page_stats(); 1774 continue; 1775 } 1776 } 1777 if (vm_pages_needed) 1778 cnt.v_pdwakeups++; 1779 mtx_unlock(&vm_page_queue_free_mtx); 1780 vm_pageout_scan(pass); 1781 } 1782 } 1783 1784 /* 1785 * Unless the free page queue lock is held by the caller, this function 1786 * should be regarded as advisory. Specifically, the caller should 1787 * not msleep() on &cnt.v_free_count following this function unless 1788 * the free page queue lock is held until the msleep() is performed. 1789 */ 1790 void 1791 pagedaemon_wakeup() 1792 { 1793 1794 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1795 vm_pages_needed = 1; 1796 wakeup(&vm_pages_needed); 1797 } 1798 } 1799 1800 #if !defined(NO_SWAPPING) 1801 static void 1802 vm_req_vmdaemon(int req) 1803 { 1804 static int lastrun = 0; 1805 1806 mtx_lock(&vm_daemon_mtx); 1807 vm_pageout_req_swapout |= req; 1808 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1809 wakeup(&vm_daemon_needed); 1810 lastrun = ticks; 1811 } 1812 mtx_unlock(&vm_daemon_mtx); 1813 } 1814 1815 static void 1816 vm_daemon() 1817 { 1818 struct rlimit rsslim; 1819 struct proc *p; 1820 struct thread *td; 1821 struct vmspace *vm; 1822 int breakout, swapout_flags, tryagain, attempts; 1823 #ifdef RACCT 1824 uint64_t rsize, ravailable; 1825 #endif 1826 1827 while (TRUE) { 1828 mtx_lock(&vm_daemon_mtx); 1829 #ifdef RACCT 1830 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz); 1831 #else 1832 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1833 #endif 1834 swapout_flags = vm_pageout_req_swapout; 1835 vm_pageout_req_swapout = 0; 1836 mtx_unlock(&vm_daemon_mtx); 1837 if (swapout_flags) 1838 swapout_procs(swapout_flags); 1839 1840 /* 1841 * scan the processes for exceeding their rlimits or if 1842 * process is swapped out -- deactivate pages 1843 */ 1844 tryagain = 0; 1845 attempts = 0; 1846 again: 1847 attempts++; 1848 sx_slock(&allproc_lock); 1849 FOREACH_PROC_IN_SYSTEM(p) { 1850 vm_pindex_t limit, size; 1851 1852 /* 1853 * if this is a system process or if we have already 1854 * looked at this process, skip it. 1855 */ 1856 PROC_LOCK(p); 1857 if (p->p_state != PRS_NORMAL || 1858 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1859 PROC_UNLOCK(p); 1860 continue; 1861 } 1862 /* 1863 * if the process is in a non-running type state, 1864 * don't touch it. 1865 */ 1866 breakout = 0; 1867 FOREACH_THREAD_IN_PROC(p, td) { 1868 thread_lock(td); 1869 if (!TD_ON_RUNQ(td) && 1870 !TD_IS_RUNNING(td) && 1871 !TD_IS_SLEEPING(td) && 1872 !TD_IS_SUSPENDED(td)) { 1873 thread_unlock(td); 1874 breakout = 1; 1875 break; 1876 } 1877 thread_unlock(td); 1878 } 1879 if (breakout) { 1880 PROC_UNLOCK(p); 1881 continue; 1882 } 1883 /* 1884 * get a limit 1885 */ 1886 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1887 limit = OFF_TO_IDX( 1888 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1889 1890 /* 1891 * let processes that are swapped out really be 1892 * swapped out set the limit to nothing (will force a 1893 * swap-out.) 1894 */ 1895 if ((p->p_flag & P_INMEM) == 0) 1896 limit = 0; /* XXX */ 1897 vm = vmspace_acquire_ref(p); 1898 PROC_UNLOCK(p); 1899 if (vm == NULL) 1900 continue; 1901 1902 size = vmspace_resident_count(vm); 1903 if (limit >= 0 && size >= limit) { 1904 vm_pageout_map_deactivate_pages( 1905 &vm->vm_map, limit); 1906 } 1907 #ifdef RACCT 1908 rsize = IDX_TO_OFF(size); 1909 PROC_LOCK(p); 1910 racct_set(p, RACCT_RSS, rsize); 1911 ravailable = racct_get_available(p, RACCT_RSS); 1912 PROC_UNLOCK(p); 1913 if (rsize > ravailable) { 1914 /* 1915 * Don't be overly aggressive; this might be 1916 * an innocent process, and the limit could've 1917 * been exceeded by some memory hog. Don't 1918 * try to deactivate more than 1/4th of process' 1919 * resident set size. 1920 */ 1921 if (attempts <= 8) { 1922 if (ravailable < rsize - (rsize / 4)) 1923 ravailable = rsize - (rsize / 4); 1924 } 1925 vm_pageout_map_deactivate_pages( 1926 &vm->vm_map, OFF_TO_IDX(ravailable)); 1927 /* Update RSS usage after paging out. */ 1928 size = vmspace_resident_count(vm); 1929 rsize = IDX_TO_OFF(size); 1930 PROC_LOCK(p); 1931 racct_set(p, RACCT_RSS, rsize); 1932 PROC_UNLOCK(p); 1933 if (rsize > ravailable) 1934 tryagain = 1; 1935 } 1936 #endif 1937 vmspace_free(vm); 1938 } 1939 sx_sunlock(&allproc_lock); 1940 if (tryagain != 0 && attempts <= 10) 1941 goto again; 1942 } 1943 } 1944 #endif /* !defined(NO_SWAPPING) */ 1945