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 80 #include <sys/param.h> 81 #include <sys/systm.h> 82 #include <sys/kernel.h> 83 #include <sys/eventhandler.h> 84 #include <sys/lock.h> 85 #include <sys/mutex.h> 86 #include <sys/proc.h> 87 #include <sys/kthread.h> 88 #include <sys/ktr.h> 89 #include <sys/mount.h> 90 #include <sys/racct.h> 91 #include <sys/resourcevar.h> 92 #include <sys/sched.h> 93 #include <sys/sdt.h> 94 #include <sys/signalvar.h> 95 #include <sys/smp.h> 96 #include <sys/time.h> 97 #include <sys/vnode.h> 98 #include <sys/vmmeter.h> 99 #include <sys/rwlock.h> 100 #include <sys/sx.h> 101 #include <sys/sysctl.h> 102 103 #include <vm/vm.h> 104 #include <vm/vm_param.h> 105 #include <vm/vm_object.h> 106 #include <vm/vm_page.h> 107 #include <vm/vm_map.h> 108 #include <vm/vm_pageout.h> 109 #include <vm/vm_pager.h> 110 #include <vm/vm_phys.h> 111 #include <vm/swap_pager.h> 112 #include <vm/vm_extern.h> 113 #include <vm/uma.h> 114 115 /* 116 * System initialization 117 */ 118 119 /* the kernel process "vm_pageout"*/ 120 static void vm_pageout(void); 121 static void vm_pageout_init(void); 122 static int vm_pageout_clean(vm_page_t m); 123 static int vm_pageout_cluster(vm_page_t m); 124 static void vm_pageout_scan(struct vm_domain *vmd, int pass); 125 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, 126 int starting_page_shortage); 127 128 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, 129 NULL); 130 131 struct proc *pageproc; 132 133 static struct kproc_desc page_kp = { 134 "pagedaemon", 135 vm_pageout, 136 &pageproc 137 }; 138 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, 139 &page_kp); 140 141 SDT_PROVIDER_DEFINE(vm); 142 SDT_PROBE_DEFINE(vm, , , vm__lowmem_cache); 143 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); 144 145 #if !defined(NO_SWAPPING) 146 /* the kernel process "vm_daemon"*/ 147 static void vm_daemon(void); 148 static struct proc *vmproc; 149 150 static struct kproc_desc vm_kp = { 151 "vmdaemon", 152 vm_daemon, 153 &vmproc 154 }; 155 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 156 #endif 157 158 159 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 160 int vm_pageout_deficit; /* Estimated number of pages deficit */ 161 int vm_pageout_wakeup_thresh; 162 static int vm_pageout_oom_seq = 12; 163 164 #if !defined(NO_SWAPPING) 165 static int vm_pageout_req_swapout; /* XXX */ 166 static int vm_daemon_needed; 167 static struct mtx vm_daemon_mtx; 168 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 169 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 170 #endif 171 static int vm_max_launder = 32; 172 static int vm_pageout_update_period; 173 static int defer_swap_pageouts; 174 static int disable_swap_pageouts; 175 static int lowmem_period = 10; 176 static time_t lowmem_uptime; 177 178 #if defined(NO_SWAPPING) 179 static int vm_swap_enabled = 0; 180 static int vm_swap_idle_enabled = 0; 181 #else 182 static int vm_swap_enabled = 1; 183 static int vm_swap_idle_enabled = 0; 184 #endif 185 186 static int vm_panic_on_oom = 0; 187 188 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, 189 CTLFLAG_RWTUN, &vm_panic_on_oom, 0, 190 "panic on out of memory instead of killing the largest process"); 191 192 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh, 193 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0, 194 "free page threshold for waking up the pageout daemon"); 195 196 SYSCTL_INT(_vm, OID_AUTO, max_launder, 197 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 198 199 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, 200 CTLFLAG_RW, &vm_pageout_update_period, 0, 201 "Maximum active LRU update period"); 202 203 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0, 204 "Low memory callback period"); 205 206 #if defined(NO_SWAPPING) 207 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 208 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 209 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 210 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 211 #else 212 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 213 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 214 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 215 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 216 #endif 217 218 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 219 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 220 221 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 222 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 223 224 static int pageout_lock_miss; 225 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 226 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 227 228 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq, 229 CTLFLAG_RW, &vm_pageout_oom_seq, 0, 230 "back-to-back calls to oom detector to start OOM"); 231 232 #define VM_PAGEOUT_PAGE_COUNT 16 233 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 234 235 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 236 SYSCTL_INT(_vm, OID_AUTO, max_wired, 237 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 238 239 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *); 240 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t, 241 vm_paddr_t); 242 #if !defined(NO_SWAPPING) 243 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 244 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 245 static void vm_req_vmdaemon(int req); 246 #endif 247 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *); 248 249 /* 250 * Initialize a dummy page for marking the caller's place in the specified 251 * paging queue. In principle, this function only needs to set the flag 252 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count 253 * to one as safety precautions. 254 */ 255 static void 256 vm_pageout_init_marker(vm_page_t marker, u_short queue) 257 { 258 259 bzero(marker, sizeof(*marker)); 260 marker->flags = PG_MARKER; 261 marker->busy_lock = VPB_SINGLE_EXCLUSIVER; 262 marker->queue = queue; 263 marker->hold_count = 1; 264 } 265 266 /* 267 * vm_pageout_fallback_object_lock: 268 * 269 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is 270 * known to have failed and page queue must be either PQ_ACTIVE or 271 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 272 * while locking the vm object. Use marker page to detect page queue 273 * changes and maintain notion of next page on page queue. Return 274 * TRUE if no changes were detected, FALSE otherwise. vm object is 275 * locked on return. 276 * 277 * This function depends on both the lock portion of struct vm_object 278 * and normal struct vm_page being type stable. 279 */ 280 static boolean_t 281 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 282 { 283 struct vm_page marker; 284 struct vm_pagequeue *pq; 285 boolean_t unchanged; 286 u_short queue; 287 vm_object_t object; 288 289 queue = m->queue; 290 vm_pageout_init_marker(&marker, queue); 291 pq = vm_page_pagequeue(m); 292 object = m->object; 293 294 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 295 vm_pagequeue_unlock(pq); 296 vm_page_unlock(m); 297 VM_OBJECT_WLOCK(object); 298 vm_page_lock(m); 299 vm_pagequeue_lock(pq); 300 301 /* 302 * The page's object might have changed, and/or the page might 303 * have moved from its original position in the queue. If the 304 * page's object has changed, then the caller should abandon 305 * processing the page because the wrong object lock was 306 * acquired. Use the marker's plinks.q, not the page's, to 307 * determine if the page has been moved. The state of the 308 * page's plinks.q can be indeterminate; whereas, the marker's 309 * plinks.q must be valid. 310 */ 311 *next = TAILQ_NEXT(&marker, plinks.q); 312 unchanged = m->object == object && 313 m == TAILQ_PREV(&marker, pglist, plinks.q); 314 KASSERT(!unchanged || m->queue == queue, 315 ("page %p queue %d %d", m, queue, m->queue)); 316 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 317 return (unchanged); 318 } 319 320 /* 321 * Lock the page while holding the page queue lock. Use marker page 322 * to detect page queue changes and maintain notion of next page on 323 * page queue. Return TRUE if no changes were detected, FALSE 324 * otherwise. The page is locked on return. The page queue lock might 325 * be dropped and reacquired. 326 * 327 * This function depends on normal struct vm_page being type stable. 328 */ 329 static boolean_t 330 vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 331 { 332 struct vm_page marker; 333 struct vm_pagequeue *pq; 334 boolean_t unchanged; 335 u_short queue; 336 337 vm_page_lock_assert(m, MA_NOTOWNED); 338 if (vm_page_trylock(m)) 339 return (TRUE); 340 341 queue = m->queue; 342 vm_pageout_init_marker(&marker, queue); 343 pq = vm_page_pagequeue(m); 344 345 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 346 vm_pagequeue_unlock(pq); 347 vm_page_lock(m); 348 vm_pagequeue_lock(pq); 349 350 /* Page queue might have changed. */ 351 *next = TAILQ_NEXT(&marker, plinks.q); 352 unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q); 353 KASSERT(!unchanged || m->queue == queue, 354 ("page %p queue %d %d", m, queue, m->queue)); 355 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 356 return (unchanged); 357 } 358 359 /* 360 * vm_pageout_clean: 361 * 362 * Clean the page and remove it from the laundry. 363 * 364 * We set the busy bit to cause potential page faults on this page to 365 * block. Note the careful timing, however, the busy bit isn't set till 366 * late and we cannot do anything that will mess with the page. 367 */ 368 static int 369 vm_pageout_cluster(vm_page_t m) 370 { 371 vm_object_t object; 372 vm_page_t mc[2*vm_pageout_page_count], pb, ps; 373 int pageout_count; 374 int ib, is, page_base; 375 vm_pindex_t pindex = m->pindex; 376 377 vm_page_lock_assert(m, MA_OWNED); 378 object = m->object; 379 VM_OBJECT_ASSERT_WLOCKED(object); 380 381 /* 382 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 383 * with the new swapper, but we could have serious problems paging 384 * out other object types if there is insufficient memory. 385 * 386 * Unfortunately, checking free memory here is far too late, so the 387 * check has been moved up a procedural level. 388 */ 389 390 /* 391 * Can't clean the page if it's busy or held. 392 */ 393 vm_page_assert_unbusied(m); 394 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m)); 395 vm_page_unlock(m); 396 397 mc[vm_pageout_page_count] = pb = ps = m; 398 pageout_count = 1; 399 page_base = vm_pageout_page_count; 400 ib = 1; 401 is = 1; 402 403 /* 404 * Scan object for clusterable pages. 405 * 406 * We can cluster ONLY if: ->> the page is NOT 407 * clean, wired, busy, held, or mapped into a 408 * buffer, and one of the following: 409 * 1) The page is inactive, or a seldom used 410 * active page. 411 * -or- 412 * 2) we force the issue. 413 * 414 * During heavy mmap/modification loads the pageout 415 * daemon can really fragment the underlying file 416 * due to flushing pages out of order and not trying 417 * align the clusters (which leave sporatic out-of-order 418 * holes). To solve this problem we do the reverse scan 419 * first and attempt to align our cluster, then do a 420 * forward scan if room remains. 421 */ 422 more: 423 while (ib && pageout_count < vm_pageout_page_count) { 424 vm_page_t p; 425 426 if (ib > pindex) { 427 ib = 0; 428 break; 429 } 430 431 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) { 432 ib = 0; 433 break; 434 } 435 vm_page_test_dirty(p); 436 if (p->dirty == 0) { 437 ib = 0; 438 break; 439 } 440 vm_page_lock(p); 441 if (p->queue != PQ_INACTIVE || 442 p->hold_count != 0) { /* may be undergoing I/O */ 443 vm_page_unlock(p); 444 ib = 0; 445 break; 446 } 447 vm_page_unlock(p); 448 mc[--page_base] = pb = p; 449 ++pageout_count; 450 ++ib; 451 /* 452 * alignment boundry, stop here and switch directions. Do 453 * not clear ib. 454 */ 455 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 456 break; 457 } 458 459 while (pageout_count < vm_pageout_page_count && 460 pindex + is < object->size) { 461 vm_page_t p; 462 463 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p)) 464 break; 465 vm_page_test_dirty(p); 466 if (p->dirty == 0) 467 break; 468 vm_page_lock(p); 469 if (p->queue != PQ_INACTIVE || 470 p->hold_count != 0) { /* may be undergoing I/O */ 471 vm_page_unlock(p); 472 break; 473 } 474 vm_page_unlock(p); 475 mc[page_base + pageout_count] = ps = p; 476 ++pageout_count; 477 ++is; 478 } 479 480 /* 481 * If we exhausted our forward scan, continue with the reverse scan 482 * when possible, even past a page boundry. This catches boundry 483 * conditions. 484 */ 485 if (ib && pageout_count < vm_pageout_page_count) 486 goto more; 487 488 /* 489 * we allow reads during pageouts... 490 */ 491 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL, 492 NULL)); 493 } 494 495 /* 496 * vm_pageout_flush() - launder the given pages 497 * 498 * The given pages are laundered. Note that we setup for the start of 499 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 500 * reference count all in here rather then in the parent. If we want 501 * the parent to do more sophisticated things we may have to change 502 * the ordering. 503 * 504 * Returned runlen is the count of pages between mreq and first 505 * page after mreq with status VM_PAGER_AGAIN. 506 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 507 * for any page in runlen set. 508 */ 509 int 510 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 511 boolean_t *eio) 512 { 513 vm_object_t object = mc[0]->object; 514 int pageout_status[count]; 515 int numpagedout = 0; 516 int i, runlen; 517 518 VM_OBJECT_ASSERT_WLOCKED(object); 519 520 /* 521 * Initiate I/O. Bump the vm_page_t->busy counter and 522 * mark the pages read-only. 523 * 524 * We do not have to fixup the clean/dirty bits here... we can 525 * allow the pager to do it after the I/O completes. 526 * 527 * NOTE! mc[i]->dirty may be partial or fragmented due to an 528 * edge case with file fragments. 529 */ 530 for (i = 0; i < count; i++) { 531 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 532 ("vm_pageout_flush: partially invalid page %p index %d/%d", 533 mc[i], i, count)); 534 vm_page_sbusy(mc[i]); 535 pmap_remove_write(mc[i]); 536 } 537 vm_object_pip_add(object, count); 538 539 vm_pager_put_pages(object, mc, count, flags, pageout_status); 540 541 runlen = count - mreq; 542 if (eio != NULL) 543 *eio = FALSE; 544 for (i = 0; i < count; i++) { 545 vm_page_t mt = mc[i]; 546 547 KASSERT(pageout_status[i] == VM_PAGER_PEND || 548 !pmap_page_is_write_mapped(mt), 549 ("vm_pageout_flush: page %p is not write protected", mt)); 550 switch (pageout_status[i]) { 551 case VM_PAGER_OK: 552 case VM_PAGER_PEND: 553 numpagedout++; 554 break; 555 case VM_PAGER_BAD: 556 /* 557 * Page outside of range of object. Right now we 558 * essentially lose the changes by pretending it 559 * worked. 560 */ 561 vm_page_undirty(mt); 562 break; 563 case VM_PAGER_ERROR: 564 case VM_PAGER_FAIL: 565 /* 566 * If page couldn't be paged out, then reactivate the 567 * page so it doesn't clog the inactive list. (We 568 * will try paging out it again later). 569 */ 570 vm_page_lock(mt); 571 vm_page_activate(mt); 572 vm_page_unlock(mt); 573 if (eio != NULL && i >= mreq && i - mreq < runlen) 574 *eio = TRUE; 575 break; 576 case VM_PAGER_AGAIN: 577 if (i >= mreq && i - mreq < runlen) 578 runlen = i - mreq; 579 break; 580 } 581 582 /* 583 * If the operation is still going, leave the page busy to 584 * block all other accesses. Also, leave the paging in 585 * progress indicator set so that we don't attempt an object 586 * collapse. 587 */ 588 if (pageout_status[i] != VM_PAGER_PEND) { 589 vm_object_pip_wakeup(object); 590 vm_page_sunbusy(mt); 591 } 592 } 593 if (prunlen != NULL) 594 *prunlen = runlen; 595 return (numpagedout); 596 } 597 598 static boolean_t 599 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low, 600 vm_paddr_t high) 601 { 602 struct mount *mp; 603 struct vnode *vp; 604 vm_object_t object; 605 vm_paddr_t pa; 606 vm_page_t m, m_tmp, next; 607 int lockmode; 608 609 vm_pagequeue_lock(pq); 610 TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) { 611 if ((m->flags & PG_MARKER) != 0) 612 continue; 613 pa = VM_PAGE_TO_PHYS(m); 614 if (pa < low || pa + PAGE_SIZE > high) 615 continue; 616 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) { 617 vm_page_unlock(m); 618 continue; 619 } 620 object = m->object; 621 if ((!VM_OBJECT_TRYWLOCK(object) && 622 (!vm_pageout_fallback_object_lock(m, &next) || 623 m->hold_count != 0)) || vm_page_busied(m)) { 624 vm_page_unlock(m); 625 VM_OBJECT_WUNLOCK(object); 626 continue; 627 } 628 vm_page_test_dirty(m); 629 if (m->dirty == 0 && object->ref_count != 0) 630 pmap_remove_all(m); 631 if (m->dirty != 0) { 632 vm_page_unlock(m); 633 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) { 634 VM_OBJECT_WUNLOCK(object); 635 continue; 636 } 637 if (object->type == OBJT_VNODE) { 638 vm_pagequeue_unlock(pq); 639 vp = object->handle; 640 vm_object_reference_locked(object); 641 VM_OBJECT_WUNLOCK(object); 642 (void)vn_start_write(vp, &mp, V_WAIT); 643 lockmode = MNT_SHARED_WRITES(vp->v_mount) ? 644 LK_SHARED : LK_EXCLUSIVE; 645 vn_lock(vp, lockmode | LK_RETRY); 646 VM_OBJECT_WLOCK(object); 647 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 648 VM_OBJECT_WUNLOCK(object); 649 VOP_UNLOCK(vp, 0); 650 vm_object_deallocate(object); 651 vn_finished_write(mp); 652 return (TRUE); 653 } else if (object->type == OBJT_SWAP || 654 object->type == OBJT_DEFAULT) { 655 vm_pagequeue_unlock(pq); 656 m_tmp = m; 657 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC, 658 0, NULL, NULL); 659 VM_OBJECT_WUNLOCK(object); 660 return (TRUE); 661 } 662 } else { 663 /* 664 * Dequeue here to prevent lock recursion in 665 * vm_page_cache(). 666 */ 667 vm_page_dequeue_locked(m); 668 vm_page_cache(m); 669 vm_page_unlock(m); 670 } 671 VM_OBJECT_WUNLOCK(object); 672 } 673 vm_pagequeue_unlock(pq); 674 return (FALSE); 675 } 676 677 /* 678 * Increase the number of cached pages. The specified value, "tries", 679 * determines which categories of pages are cached: 680 * 681 * 0: All clean, inactive pages within the specified physical address range 682 * are cached. Will not sleep. 683 * 1: The vm_lowmem handlers are called. All inactive pages within 684 * the specified physical address range are cached. May sleep. 685 * 2: The vm_lowmem handlers are called. All inactive and active pages 686 * within the specified physical address range are cached. May sleep. 687 */ 688 void 689 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high) 690 { 691 int actl, actmax, inactl, inactmax, dom, initial_dom; 692 static int start_dom = 0; 693 694 if (tries > 0) { 695 /* 696 * Decrease registered cache sizes. The vm_lowmem handlers 697 * may acquire locks and/or sleep, so they can only be invoked 698 * when "tries" is greater than zero. 699 */ 700 SDT_PROBE0(vm, , , vm__lowmem_cache); 701 EVENTHANDLER_INVOKE(vm_lowmem, 0); 702 703 /* 704 * We do this explicitly after the caches have been drained 705 * above. 706 */ 707 uma_reclaim(); 708 } 709 710 /* 711 * Make the next scan start on the next domain. 712 */ 713 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains; 714 715 inactl = 0; 716 inactmax = vm_cnt.v_inactive_count; 717 actl = 0; 718 actmax = tries < 2 ? 0 : vm_cnt.v_active_count; 719 dom = initial_dom; 720 721 /* 722 * Scan domains in round-robin order, first inactive queues, 723 * then active. Since domain usually owns large physically 724 * contiguous chunk of memory, it makes sense to completely 725 * exhaust one domain before switching to next, while growing 726 * the pool of contiguous physical pages. 727 * 728 * Do not even start launder a domain which cannot contain 729 * the specified address range, as indicated by segments 730 * constituting the domain. 731 */ 732 again_inact: 733 if (inactl < inactmax) { 734 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 735 low, high) && 736 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE], 737 tries, low, high)) { 738 inactl++; 739 goto again_inact; 740 } 741 if (++dom == vm_ndomains) 742 dom = 0; 743 if (dom != initial_dom) 744 goto again_inact; 745 } 746 again_act: 747 if (actl < actmax) { 748 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 749 low, high) && 750 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE], 751 tries, low, high)) { 752 actl++; 753 goto again_act; 754 } 755 if (++dom == vm_ndomains) 756 dom = 0; 757 if (dom != initial_dom) 758 goto again_act; 759 } 760 } 761 762 #if !defined(NO_SWAPPING) 763 /* 764 * vm_pageout_object_deactivate_pages 765 * 766 * Deactivate enough pages to satisfy the inactive target 767 * requirements. 768 * 769 * The object and map must be locked. 770 */ 771 static void 772 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 773 long desired) 774 { 775 vm_object_t backing_object, object; 776 vm_page_t p; 777 int act_delta, remove_mode; 778 779 VM_OBJECT_ASSERT_LOCKED(first_object); 780 if ((first_object->flags & OBJ_FICTITIOUS) != 0) 781 return; 782 for (object = first_object;; object = backing_object) { 783 if (pmap_resident_count(pmap) <= desired) 784 goto unlock_return; 785 VM_OBJECT_ASSERT_LOCKED(object); 786 if ((object->flags & OBJ_UNMANAGED) != 0 || 787 object->paging_in_progress != 0) 788 goto unlock_return; 789 790 remove_mode = 0; 791 if (object->shadow_count > 1) 792 remove_mode = 1; 793 /* 794 * Scan the object's entire memory queue. 795 */ 796 TAILQ_FOREACH(p, &object->memq, listq) { 797 if (pmap_resident_count(pmap) <= desired) 798 goto unlock_return; 799 if (vm_page_busied(p)) 800 continue; 801 PCPU_INC(cnt.v_pdpages); 802 vm_page_lock(p); 803 if (p->wire_count != 0 || p->hold_count != 0 || 804 !pmap_page_exists_quick(pmap, p)) { 805 vm_page_unlock(p); 806 continue; 807 } 808 act_delta = pmap_ts_referenced(p); 809 if ((p->aflags & PGA_REFERENCED) != 0) { 810 if (act_delta == 0) 811 act_delta = 1; 812 vm_page_aflag_clear(p, PGA_REFERENCED); 813 } 814 if (p->queue != PQ_ACTIVE && act_delta != 0) { 815 vm_page_activate(p); 816 p->act_count += act_delta; 817 } else if (p->queue == PQ_ACTIVE) { 818 if (act_delta == 0) { 819 p->act_count -= min(p->act_count, 820 ACT_DECLINE); 821 if (!remove_mode && p->act_count == 0) { 822 pmap_remove_all(p); 823 vm_page_deactivate(p); 824 } else 825 vm_page_requeue(p); 826 } else { 827 vm_page_activate(p); 828 if (p->act_count < ACT_MAX - 829 ACT_ADVANCE) 830 p->act_count += ACT_ADVANCE; 831 vm_page_requeue(p); 832 } 833 } else if (p->queue == PQ_INACTIVE) 834 pmap_remove_all(p); 835 vm_page_unlock(p); 836 } 837 if ((backing_object = object->backing_object) == NULL) 838 goto unlock_return; 839 VM_OBJECT_RLOCK(backing_object); 840 if (object != first_object) 841 VM_OBJECT_RUNLOCK(object); 842 } 843 unlock_return: 844 if (object != first_object) 845 VM_OBJECT_RUNLOCK(object); 846 } 847 848 /* 849 * deactivate some number of pages in a map, try to do it fairly, but 850 * that is really hard to do. 851 */ 852 static void 853 vm_pageout_map_deactivate_pages(map, desired) 854 vm_map_t map; 855 long desired; 856 { 857 vm_map_entry_t tmpe; 858 vm_object_t obj, bigobj; 859 int nothingwired; 860 861 if (!vm_map_trylock(map)) 862 return; 863 864 bigobj = NULL; 865 nothingwired = TRUE; 866 867 /* 868 * first, search out the biggest object, and try to free pages from 869 * that. 870 */ 871 tmpe = map->header.next; 872 while (tmpe != &map->header) { 873 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 874 obj = tmpe->object.vm_object; 875 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { 876 if (obj->shadow_count <= 1 && 877 (bigobj == NULL || 878 bigobj->resident_page_count < obj->resident_page_count)) { 879 if (bigobj != NULL) 880 VM_OBJECT_RUNLOCK(bigobj); 881 bigobj = obj; 882 } else 883 VM_OBJECT_RUNLOCK(obj); 884 } 885 } 886 if (tmpe->wired_count > 0) 887 nothingwired = FALSE; 888 tmpe = tmpe->next; 889 } 890 891 if (bigobj != NULL) { 892 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 893 VM_OBJECT_RUNLOCK(bigobj); 894 } 895 /* 896 * Next, hunt around for other pages to deactivate. We actually 897 * do this search sort of wrong -- .text first is not the best idea. 898 */ 899 tmpe = map->header.next; 900 while (tmpe != &map->header) { 901 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 902 break; 903 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 904 obj = tmpe->object.vm_object; 905 if (obj != NULL) { 906 VM_OBJECT_RLOCK(obj); 907 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 908 VM_OBJECT_RUNLOCK(obj); 909 } 910 } 911 tmpe = tmpe->next; 912 } 913 914 /* 915 * Remove all mappings if a process is swapped out, this will free page 916 * table pages. 917 */ 918 if (desired == 0 && nothingwired) { 919 pmap_remove(vm_map_pmap(map), vm_map_min(map), 920 vm_map_max(map)); 921 } 922 923 vm_map_unlock(map); 924 } 925 #endif /* !defined(NO_SWAPPING) */ 926 927 /* 928 * Attempt to acquire all of the necessary locks to launder a page and 929 * then call through the clustering layer to PUTPAGES. Wait a short 930 * time for a vnode lock. 931 * 932 * Requires the page and object lock on entry, releases both before return. 933 * Returns 0 on success and an errno otherwise. 934 */ 935 static int 936 vm_pageout_clean(vm_page_t m) 937 { 938 struct vnode *vp; 939 struct mount *mp; 940 vm_object_t object; 941 vm_pindex_t pindex; 942 int error, lockmode; 943 944 vm_page_assert_locked(m); 945 object = m->object; 946 VM_OBJECT_ASSERT_WLOCKED(object); 947 error = 0; 948 vp = NULL; 949 mp = NULL; 950 951 /* 952 * The object is already known NOT to be dead. It 953 * is possible for the vget() to block the whole 954 * pageout daemon, but the new low-memory handling 955 * code should prevent it. 956 * 957 * We can't wait forever for the vnode lock, we might 958 * deadlock due to a vn_read() getting stuck in 959 * vm_wait while holding this vnode. We skip the 960 * vnode if we can't get it in a reasonable amount 961 * of time. 962 */ 963 if (object->type == OBJT_VNODE) { 964 vm_page_unlock(m); 965 vp = object->handle; 966 if (vp->v_type == VREG && 967 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 968 mp = NULL; 969 error = EDEADLK; 970 goto unlock_all; 971 } 972 KASSERT(mp != NULL, 973 ("vp %p with NULL v_mount", vp)); 974 vm_object_reference_locked(object); 975 pindex = m->pindex; 976 VM_OBJECT_WUNLOCK(object); 977 lockmode = MNT_SHARED_WRITES(vp->v_mount) ? 978 LK_SHARED : LK_EXCLUSIVE; 979 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) { 980 vp = NULL; 981 error = EDEADLK; 982 goto unlock_mp; 983 } 984 VM_OBJECT_WLOCK(object); 985 vm_page_lock(m); 986 /* 987 * While the object and page were unlocked, the page 988 * may have been: 989 * (1) moved to a different queue, 990 * (2) reallocated to a different object, 991 * (3) reallocated to a different offset, or 992 * (4) cleaned. 993 */ 994 if (m->queue != PQ_INACTIVE || m->object != object || 995 m->pindex != pindex || m->dirty == 0) { 996 vm_page_unlock(m); 997 error = ENXIO; 998 goto unlock_all; 999 } 1000 1001 /* 1002 * The page may have been busied or held while the object 1003 * and page locks were released. 1004 */ 1005 if (vm_page_busied(m) || m->hold_count != 0) { 1006 vm_page_unlock(m); 1007 error = EBUSY; 1008 goto unlock_all; 1009 } 1010 } 1011 1012 /* 1013 * If a page is dirty, then it is either being washed 1014 * (but not yet cleaned) or it is still in the 1015 * laundry. If it is still in the laundry, then we 1016 * start the cleaning operation. 1017 */ 1018 if (vm_pageout_cluster(m) == 0) 1019 error = EIO; 1020 1021 unlock_all: 1022 VM_OBJECT_WUNLOCK(object); 1023 1024 unlock_mp: 1025 vm_page_lock_assert(m, MA_NOTOWNED); 1026 if (mp != NULL) { 1027 if (vp != NULL) 1028 vput(vp); 1029 vm_object_deallocate(object); 1030 vn_finished_write(mp); 1031 } 1032 1033 return (error); 1034 } 1035 1036 /* 1037 * vm_pageout_scan does the dirty work for the pageout daemon. 1038 * 1039 * pass 0 - Update active LRU/deactivate pages 1040 * pass 1 - Move inactive to cache or free 1041 * pass 2 - Launder dirty pages 1042 */ 1043 static void 1044 vm_pageout_scan(struct vm_domain *vmd, int pass) 1045 { 1046 vm_page_t m, next; 1047 struct vm_pagequeue *pq; 1048 vm_object_t object; 1049 long min_scan; 1050 int act_delta, addl_page_shortage, deficit, error, maxlaunder, maxscan; 1051 int page_shortage, scan_tick, scanned, starting_page_shortage; 1052 int vnodes_skipped; 1053 boolean_t pageout_ok, queues_locked; 1054 1055 /* 1056 * If we need to reclaim memory ask kernel caches to return 1057 * some. We rate limit to avoid thrashing. 1058 */ 1059 if (vmd == &vm_dom[0] && pass > 0 && 1060 (time_uptime - lowmem_uptime) >= lowmem_period) { 1061 /* 1062 * Decrease registered cache sizes. 1063 */ 1064 SDT_PROBE0(vm, , , vm__lowmem_scan); 1065 EVENTHANDLER_INVOKE(vm_lowmem, 0); 1066 /* 1067 * We do this explicitly after the caches have been 1068 * drained above. 1069 */ 1070 uma_reclaim(); 1071 lowmem_uptime = time_uptime; 1072 } 1073 1074 /* 1075 * The addl_page_shortage is the number of temporarily 1076 * stuck pages in the inactive queue. In other words, the 1077 * number of pages from the inactive count that should be 1078 * discounted in setting the target for the active queue scan. 1079 */ 1080 addl_page_shortage = 0; 1081 1082 /* 1083 * Calculate the number of pages we want to either free or move 1084 * to the cache. 1085 */ 1086 if (pass > 0) { 1087 deficit = atomic_readandclear_int(&vm_pageout_deficit); 1088 page_shortage = vm_paging_target() + deficit; 1089 } else 1090 page_shortage = deficit = 0; 1091 starting_page_shortage = page_shortage; 1092 1093 /* 1094 * maxlaunder limits the number of dirty pages we flush per scan. 1095 * For most systems a smaller value (16 or 32) is more robust under 1096 * extreme memory and disk pressure because any unnecessary writes 1097 * to disk can result in extreme performance degredation. However, 1098 * systems with excessive dirty pages (especially when MAP_NOSYNC is 1099 * used) will die horribly with limited laundering. If the pageout 1100 * daemon cannot clean enough pages in the first pass, we let it go 1101 * all out in succeeding passes. 1102 */ 1103 if ((maxlaunder = vm_max_launder) <= 1) 1104 maxlaunder = 1; 1105 if (pass > 1) 1106 maxlaunder = 10000; 1107 1108 vnodes_skipped = 0; 1109 1110 /* 1111 * Start scanning the inactive queue for pages we can move to the 1112 * cache or free. The scan will stop when the target is reached or 1113 * we have scanned the entire inactive queue. Note that m->act_count 1114 * is not used to form decisions for the inactive queue, only for the 1115 * active queue. 1116 */ 1117 pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; 1118 maxscan = pq->pq_cnt; 1119 vm_pagequeue_lock(pq); 1120 queues_locked = TRUE; 1121 for (m = TAILQ_FIRST(&pq->pq_pl); 1122 m != NULL && maxscan-- > 0 && page_shortage > 0; 1123 m = next) { 1124 vm_pagequeue_assert_locked(pq); 1125 KASSERT(queues_locked, ("unlocked queues")); 1126 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m)); 1127 1128 PCPU_INC(cnt.v_pdpages); 1129 next = TAILQ_NEXT(m, plinks.q); 1130 1131 /* 1132 * skip marker pages 1133 */ 1134 if (m->flags & PG_MARKER) 1135 continue; 1136 1137 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1138 ("Fictitious page %p cannot be in inactive queue", m)); 1139 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1140 ("Unmanaged page %p cannot be in inactive queue", m)); 1141 1142 /* 1143 * The page or object lock acquisitions fail if the 1144 * page was removed from the queue or moved to a 1145 * different position within the queue. In either 1146 * case, addl_page_shortage should not be incremented. 1147 */ 1148 if (!vm_pageout_page_lock(m, &next)) 1149 goto unlock_page; 1150 else if (m->hold_count != 0) { 1151 /* 1152 * Held pages are essentially stuck in the 1153 * queue. So, they ought to be discounted 1154 * from the inactive count. See the 1155 * calculation of the page_shortage for the 1156 * loop over the active queue below. 1157 */ 1158 addl_page_shortage++; 1159 goto unlock_page; 1160 } 1161 object = m->object; 1162 if (!VM_OBJECT_TRYWLOCK(object)) { 1163 if (!vm_pageout_fallback_object_lock(m, &next)) 1164 goto unlock_object; 1165 else if (m->hold_count != 0) { 1166 addl_page_shortage++; 1167 goto unlock_object; 1168 } 1169 } 1170 if (vm_page_busied(m)) { 1171 /* 1172 * Don't mess with busy pages. Leave them at 1173 * the front of the queue. Most likely, they 1174 * are being paged out and will leave the 1175 * queue shortly after the scan finishes. So, 1176 * they ought to be discounted from the 1177 * inactive count. 1178 */ 1179 addl_page_shortage++; 1180 unlock_object: 1181 VM_OBJECT_WUNLOCK(object); 1182 unlock_page: 1183 vm_page_unlock(m); 1184 continue; 1185 } 1186 KASSERT(m->hold_count == 0, ("Held page %p", m)); 1187 1188 /* 1189 * We unlock the inactive page queue, invalidating the 1190 * 'next' pointer. Use our marker to remember our 1191 * place. 1192 */ 1193 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q); 1194 vm_pagequeue_unlock(pq); 1195 queues_locked = FALSE; 1196 1197 /* 1198 * Invalid pages can be easily freed. They cannot be 1199 * mapped, vm_page_free() asserts this. 1200 */ 1201 if (m->valid == 0) 1202 goto free_page; 1203 1204 /* 1205 * If the page has been referenced and the object is not dead, 1206 * reactivate or requeue the page depending on whether the 1207 * object is mapped. 1208 */ 1209 if ((m->aflags & PGA_REFERENCED) != 0) { 1210 vm_page_aflag_clear(m, PGA_REFERENCED); 1211 act_delta = 1; 1212 } else 1213 act_delta = 0; 1214 if (object->ref_count != 0) { 1215 act_delta += pmap_ts_referenced(m); 1216 } else { 1217 KASSERT(!pmap_page_is_mapped(m), 1218 ("vm_pageout_scan: page %p is mapped", m)); 1219 } 1220 if (act_delta != 0) { 1221 if (object->ref_count != 0) { 1222 vm_page_activate(m); 1223 1224 /* 1225 * Increase the activation count if the page 1226 * was referenced while in the inactive queue. 1227 * This makes it less likely that the page will 1228 * be returned prematurely to the inactive 1229 * queue. 1230 */ 1231 m->act_count += act_delta + ACT_ADVANCE; 1232 goto drop_page; 1233 } else if ((object->flags & OBJ_DEAD) == 0) 1234 goto requeue_page; 1235 } 1236 1237 /* 1238 * If the page appears to be clean at the machine-independent 1239 * layer, then remove all of its mappings from the pmap in 1240 * anticipation of placing it onto the cache queue. If, 1241 * however, any of the page's mappings allow write access, 1242 * then the page may still be modified until the last of those 1243 * mappings are removed. 1244 */ 1245 if (object->ref_count != 0) { 1246 vm_page_test_dirty(m); 1247 if (m->dirty == 0) 1248 pmap_remove_all(m); 1249 } 1250 1251 if (m->dirty == 0) { 1252 /* 1253 * Clean pages can be freed. 1254 */ 1255 free_page: 1256 vm_page_free(m); 1257 PCPU_INC(cnt.v_dfree); 1258 --page_shortage; 1259 } else if ((object->flags & OBJ_DEAD) != 0) { 1260 /* 1261 * Leave dirty pages from dead objects at the front of 1262 * the queue. They are being paged out and freed by 1263 * the thread that destroyed the object. They will 1264 * leave the queue shortly after the scan finishes, so 1265 * they should be discounted from the inactive count. 1266 */ 1267 addl_page_shortage++; 1268 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) { 1269 /* 1270 * Dirty pages need to be paged out, but flushing 1271 * a page is extremely expensive versus freeing 1272 * a clean page. Rather then artificially limiting 1273 * the number of pages we can flush, we instead give 1274 * dirty pages extra priority on the inactive queue 1275 * by forcing them to be cycled through the queue 1276 * twice before being flushed, after which the 1277 * (now clean) page will cycle through once more 1278 * before being freed. This significantly extends 1279 * the thrash point for a heavily loaded machine. 1280 */ 1281 m->flags |= PG_WINATCFLS; 1282 requeue_page: 1283 vm_pagequeue_lock(pq); 1284 queues_locked = TRUE; 1285 vm_page_requeue_locked(m); 1286 } else if (maxlaunder > 0) { 1287 /* 1288 * We always want to try to flush some dirty pages if 1289 * we encounter them, to keep the system stable. 1290 * Normally this number is small, but under extreme 1291 * pressure where there are insufficient clean pages 1292 * on the inactive queue, we may have to go all out. 1293 */ 1294 1295 if (object->type != OBJT_SWAP && 1296 object->type != OBJT_DEFAULT) 1297 pageout_ok = TRUE; 1298 else if (disable_swap_pageouts) 1299 pageout_ok = FALSE; 1300 else if (defer_swap_pageouts) 1301 pageout_ok = vm_page_count_min(); 1302 else 1303 pageout_ok = TRUE; 1304 if (!pageout_ok) 1305 goto requeue_page; 1306 error = vm_pageout_clean(m); 1307 /* 1308 * Decrement page_shortage on success to account for 1309 * the (future) cleaned page. Otherwise we could wind 1310 * up laundering or cleaning too many pages. 1311 */ 1312 if (error == 0) { 1313 page_shortage--; 1314 maxlaunder--; 1315 } else if (error == EDEADLK) { 1316 pageout_lock_miss++; 1317 vnodes_skipped++; 1318 } else if (error == EBUSY) { 1319 addl_page_shortage++; 1320 } 1321 vm_page_lock_assert(m, MA_NOTOWNED); 1322 goto relock_queues; 1323 } 1324 drop_page: 1325 vm_page_unlock(m); 1326 VM_OBJECT_WUNLOCK(object); 1327 relock_queues: 1328 if (!queues_locked) { 1329 vm_pagequeue_lock(pq); 1330 queues_locked = TRUE; 1331 } 1332 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q); 1333 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q); 1334 } 1335 vm_pagequeue_unlock(pq); 1336 1337 #if !defined(NO_SWAPPING) 1338 /* 1339 * Wakeup the swapout daemon if we didn't cache or free the targeted 1340 * number of pages. 1341 */ 1342 if (vm_swap_enabled && page_shortage > 0) 1343 vm_req_vmdaemon(VM_SWAP_NORMAL); 1344 #endif 1345 1346 /* 1347 * Wakeup the sync daemon if we skipped a vnode in a writeable object 1348 * and we didn't cache or free enough pages. 1349 */ 1350 if (vnodes_skipped > 0 && page_shortage > vm_cnt.v_free_target - 1351 vm_cnt.v_free_min) 1352 (void)speedup_syncer(); 1353 1354 /* 1355 * If the inactive queue scan fails repeatedly to meet its 1356 * target, kill the largest process. 1357 */ 1358 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage); 1359 1360 /* 1361 * Compute the number of pages we want to try to move from the 1362 * active queue to the inactive queue. 1363 */ 1364 page_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count + 1365 vm_paging_target() + deficit + addl_page_shortage; 1366 1367 pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; 1368 vm_pagequeue_lock(pq); 1369 maxscan = pq->pq_cnt; 1370 1371 /* 1372 * If we're just idle polling attempt to visit every 1373 * active page within 'update_period' seconds. 1374 */ 1375 scan_tick = ticks; 1376 if (vm_pageout_update_period != 0) { 1377 min_scan = pq->pq_cnt; 1378 min_scan *= scan_tick - vmd->vmd_last_active_scan; 1379 min_scan /= hz * vm_pageout_update_period; 1380 } else 1381 min_scan = 0; 1382 if (min_scan > 0 || (page_shortage > 0 && maxscan > 0)) 1383 vmd->vmd_last_active_scan = scan_tick; 1384 1385 /* 1386 * Scan the active queue for pages that can be deactivated. Update 1387 * the per-page activity counter and use it to identify deactivation 1388 * candidates. 1389 */ 1390 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned < 1391 min_scan || (page_shortage > 0 && scanned < maxscan)); m = next, 1392 scanned++) { 1393 1394 KASSERT(m->queue == PQ_ACTIVE, 1395 ("vm_pageout_scan: page %p isn't active", m)); 1396 1397 next = TAILQ_NEXT(m, plinks.q); 1398 if ((m->flags & PG_MARKER) != 0) 1399 continue; 1400 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1401 ("Fictitious page %p cannot be in active queue", m)); 1402 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1403 ("Unmanaged page %p cannot be in active queue", m)); 1404 if (!vm_pageout_page_lock(m, &next)) { 1405 vm_page_unlock(m); 1406 continue; 1407 } 1408 1409 /* 1410 * The count for pagedaemon pages is done after checking the 1411 * page for eligibility... 1412 */ 1413 PCPU_INC(cnt.v_pdpages); 1414 1415 /* 1416 * Check to see "how much" the page has been used. 1417 */ 1418 if ((m->aflags & PGA_REFERENCED) != 0) { 1419 vm_page_aflag_clear(m, PGA_REFERENCED); 1420 act_delta = 1; 1421 } else 1422 act_delta = 0; 1423 1424 /* 1425 * Unlocked object ref count check. Two races are possible. 1426 * 1) The ref was transitioning to zero and we saw non-zero, 1427 * the pmap bits will be checked unnecessarily. 1428 * 2) The ref was transitioning to one and we saw zero. 1429 * The page lock prevents a new reference to this page so 1430 * we need not check the reference bits. 1431 */ 1432 if (m->object->ref_count != 0) 1433 act_delta += pmap_ts_referenced(m); 1434 1435 /* 1436 * Advance or decay the act_count based on recent usage. 1437 */ 1438 if (act_delta != 0) { 1439 m->act_count += ACT_ADVANCE + act_delta; 1440 if (m->act_count > ACT_MAX) 1441 m->act_count = ACT_MAX; 1442 } else 1443 m->act_count -= min(m->act_count, ACT_DECLINE); 1444 1445 /* 1446 * Move this page to the tail of the active or inactive 1447 * queue depending on usage. 1448 */ 1449 if (m->act_count == 0) { 1450 /* Dequeue to avoid later lock recursion. */ 1451 vm_page_dequeue_locked(m); 1452 vm_page_deactivate(m); 1453 page_shortage--; 1454 } else 1455 vm_page_requeue_locked(m); 1456 vm_page_unlock(m); 1457 } 1458 vm_pagequeue_unlock(pq); 1459 #if !defined(NO_SWAPPING) 1460 /* 1461 * Idle process swapout -- run once per second. 1462 */ 1463 if (vm_swap_idle_enabled) { 1464 static long lsec; 1465 if (time_second != lsec) { 1466 vm_req_vmdaemon(VM_SWAP_IDLE); 1467 lsec = time_second; 1468 } 1469 } 1470 #endif 1471 } 1472 1473 static int vm_pageout_oom_vote; 1474 1475 /* 1476 * The pagedaemon threads randlomly select one to perform the 1477 * OOM. Trying to kill processes before all pagedaemons 1478 * failed to reach free target is premature. 1479 */ 1480 static void 1481 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, 1482 int starting_page_shortage) 1483 { 1484 int old_vote; 1485 1486 if (starting_page_shortage <= 0 || starting_page_shortage != 1487 page_shortage) 1488 vmd->vmd_oom_seq = 0; 1489 else 1490 vmd->vmd_oom_seq++; 1491 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) { 1492 if (vmd->vmd_oom) { 1493 vmd->vmd_oom = FALSE; 1494 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1495 } 1496 return; 1497 } 1498 1499 /* 1500 * Do not follow the call sequence until OOM condition is 1501 * cleared. 1502 */ 1503 vmd->vmd_oom_seq = 0; 1504 1505 if (vmd->vmd_oom) 1506 return; 1507 1508 vmd->vmd_oom = TRUE; 1509 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); 1510 if (old_vote != vm_ndomains - 1) 1511 return; 1512 1513 /* 1514 * The current pagedaemon thread is the last in the quorum to 1515 * start OOM. Initiate the selection and signaling of the 1516 * victim. 1517 */ 1518 vm_pageout_oom(VM_OOM_MEM); 1519 1520 /* 1521 * After one round of OOM terror, recall our vote. On the 1522 * next pass, current pagedaemon would vote again if the low 1523 * memory condition is still there, due to vmd_oom being 1524 * false. 1525 */ 1526 vmd->vmd_oom = FALSE; 1527 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1528 } 1529 1530 /* 1531 * The OOM killer is the page daemon's action of last resort when 1532 * memory allocation requests have been stalled for a prolonged period 1533 * of time because it cannot reclaim memory. This function computes 1534 * the approximate number of physical pages that could be reclaimed if 1535 * the specified address space is destroyed. 1536 * 1537 * Private, anonymous memory owned by the address space is the 1538 * principal resource that we expect to recover after an OOM kill. 1539 * Since the physical pages mapped by the address space's COW entries 1540 * are typically shared pages, they are unlikely to be released and so 1541 * they are not counted. 1542 * 1543 * To get to the point where the page daemon runs the OOM killer, its 1544 * efforts to write-back vnode-backed pages may have stalled. This 1545 * could be caused by a memory allocation deadlock in the write path 1546 * that might be resolved by an OOM kill. Therefore, physical pages 1547 * belonging to vnode-backed objects are counted, because they might 1548 * be freed without being written out first if the address space holds 1549 * the last reference to an unlinked vnode. 1550 * 1551 * Similarly, physical pages belonging to OBJT_PHYS objects are 1552 * counted because the address space might hold the last reference to 1553 * the object. 1554 */ 1555 static long 1556 vm_pageout_oom_pagecount(struct vmspace *vmspace) 1557 { 1558 vm_map_t map; 1559 vm_map_entry_t entry; 1560 vm_object_t obj; 1561 long res; 1562 1563 map = &vmspace->vm_map; 1564 KASSERT(!map->system_map, ("system map")); 1565 sx_assert(&map->lock, SA_LOCKED); 1566 res = 0; 1567 for (entry = map->header.next; entry != &map->header; 1568 entry = entry->next) { 1569 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 1570 continue; 1571 obj = entry->object.vm_object; 1572 if (obj == NULL) 1573 continue; 1574 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 && 1575 obj->ref_count != 1) 1576 continue; 1577 switch (obj->type) { 1578 case OBJT_DEFAULT: 1579 case OBJT_SWAP: 1580 case OBJT_PHYS: 1581 case OBJT_VNODE: 1582 res += obj->resident_page_count; 1583 break; 1584 } 1585 } 1586 return (res); 1587 } 1588 1589 void 1590 vm_pageout_oom(int shortage) 1591 { 1592 struct proc *p, *bigproc; 1593 vm_offset_t size, bigsize; 1594 struct thread *td; 1595 struct vmspace *vm; 1596 1597 /* 1598 * We keep the process bigproc locked once we find it to keep anyone 1599 * from messing with it; however, there is a possibility of 1600 * deadlock if process B is bigproc and one of it's child processes 1601 * attempts to propagate a signal to B while we are waiting for A's 1602 * lock while walking this list. To avoid this, we don't block on 1603 * the process lock but just skip a process if it is already locked. 1604 */ 1605 bigproc = NULL; 1606 bigsize = 0; 1607 sx_slock(&allproc_lock); 1608 FOREACH_PROC_IN_SYSTEM(p) { 1609 int breakout; 1610 1611 PROC_LOCK(p); 1612 1613 /* 1614 * If this is a system, protected or killed process, skip it. 1615 */ 1616 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | 1617 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || 1618 p->p_pid == 1 || P_KILLED(p) || 1619 (p->p_pid < 48 && swap_pager_avail != 0)) { 1620 PROC_UNLOCK(p); 1621 continue; 1622 } 1623 /* 1624 * If the process is in a non-running type state, 1625 * don't touch it. Check all the threads individually. 1626 */ 1627 breakout = 0; 1628 FOREACH_THREAD_IN_PROC(p, td) { 1629 thread_lock(td); 1630 if (!TD_ON_RUNQ(td) && 1631 !TD_IS_RUNNING(td) && 1632 !TD_IS_SLEEPING(td) && 1633 !TD_IS_SUSPENDED(td) && 1634 !TD_IS_SWAPPED(td)) { 1635 thread_unlock(td); 1636 breakout = 1; 1637 break; 1638 } 1639 thread_unlock(td); 1640 } 1641 if (breakout) { 1642 PROC_UNLOCK(p); 1643 continue; 1644 } 1645 /* 1646 * get the process size 1647 */ 1648 vm = vmspace_acquire_ref(p); 1649 if (vm == NULL) { 1650 PROC_UNLOCK(p); 1651 continue; 1652 } 1653 _PHOLD(p); 1654 if (!vm_map_trylock_read(&vm->vm_map)) { 1655 _PRELE(p); 1656 PROC_UNLOCK(p); 1657 vmspace_free(vm); 1658 continue; 1659 } 1660 PROC_UNLOCK(p); 1661 size = vmspace_swap_count(vm); 1662 if (shortage == VM_OOM_MEM) 1663 size += vm_pageout_oom_pagecount(vm); 1664 vm_map_unlock_read(&vm->vm_map); 1665 vmspace_free(vm); 1666 1667 /* 1668 * If this process is bigger than the biggest one, 1669 * remember it. 1670 */ 1671 if (size > bigsize) { 1672 if (bigproc != NULL) 1673 PRELE(bigproc); 1674 bigproc = p; 1675 bigsize = size; 1676 } else { 1677 PRELE(p); 1678 } 1679 } 1680 sx_sunlock(&allproc_lock); 1681 if (bigproc != NULL) { 1682 if (vm_panic_on_oom != 0) 1683 panic("out of swap space"); 1684 PROC_LOCK(bigproc); 1685 killproc(bigproc, "out of swap space"); 1686 sched_nice(bigproc, PRIO_MIN); 1687 _PRELE(bigproc); 1688 PROC_UNLOCK(bigproc); 1689 wakeup(&vm_cnt.v_free_count); 1690 } 1691 } 1692 1693 static void 1694 vm_pageout_worker(void *arg) 1695 { 1696 struct vm_domain *domain; 1697 int domidx; 1698 1699 domidx = (uintptr_t)arg; 1700 domain = &vm_dom[domidx]; 1701 1702 /* 1703 * XXXKIB It could be useful to bind pageout daemon threads to 1704 * the cores belonging to the domain, from which vm_page_array 1705 * is allocated. 1706 */ 1707 1708 KASSERT(domain->vmd_segs != 0, ("domain without segments")); 1709 domain->vmd_last_active_scan = ticks; 1710 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE); 1711 vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE); 1712 TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl, 1713 &domain->vmd_inacthead, plinks.q); 1714 1715 /* 1716 * The pageout daemon worker is never done, so loop forever. 1717 */ 1718 while (TRUE) { 1719 /* 1720 * If we have enough free memory, wakeup waiters. Do 1721 * not clear vm_pages_needed until we reach our target, 1722 * otherwise we may be woken up over and over again and 1723 * waste a lot of cpu. 1724 */ 1725 mtx_lock(&vm_page_queue_free_mtx); 1726 if (vm_pages_needed && !vm_page_count_min()) { 1727 if (!vm_paging_needed()) 1728 vm_pages_needed = 0; 1729 wakeup(&vm_cnt.v_free_count); 1730 } 1731 if (vm_pages_needed) { 1732 /* 1733 * We're still not done. Either vm_pages_needed was 1734 * set by another thread during the previous scan 1735 * (typically, this happens during a level 0 scan) or 1736 * vm_pages_needed was already set and the scan failed 1737 * to free enough pages. If we haven't yet performed 1738 * a level >= 2 scan (unlimited dirty cleaning), then 1739 * upgrade the level and scan again now. Otherwise, 1740 * sleep a bit and try again later. While sleeping, 1741 * vm_pages_needed can be cleared. 1742 */ 1743 if (domain->vmd_pass > 1) 1744 msleep(&vm_pages_needed, 1745 &vm_page_queue_free_mtx, PVM, "psleep", 1746 hz / 2); 1747 } else { 1748 /* 1749 * Good enough, sleep until required to refresh 1750 * stats. 1751 */ 1752 msleep(&vm_pages_needed, &vm_page_queue_free_mtx, 1753 PVM, "psleep", hz); 1754 } 1755 if (vm_pages_needed) { 1756 vm_cnt.v_pdwakeups++; 1757 domain->vmd_pass++; 1758 } else 1759 domain->vmd_pass = 0; 1760 mtx_unlock(&vm_page_queue_free_mtx); 1761 vm_pageout_scan(domain, domain->vmd_pass); 1762 } 1763 } 1764 1765 /* 1766 * vm_pageout_init initialises basic pageout daemon settings. 1767 */ 1768 static void 1769 vm_pageout_init(void) 1770 { 1771 /* 1772 * Initialize some paging parameters. 1773 */ 1774 vm_cnt.v_interrupt_free_min = 2; 1775 if (vm_cnt.v_page_count < 2000) 1776 vm_pageout_page_count = 8; 1777 1778 /* 1779 * v_free_reserved needs to include enough for the largest 1780 * swap pager structures plus enough for any pv_entry structs 1781 * when paging. 1782 */ 1783 if (vm_cnt.v_page_count > 1024) 1784 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200; 1785 else 1786 vm_cnt.v_free_min = 4; 1787 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1788 vm_cnt.v_interrupt_free_min; 1789 vm_cnt.v_free_reserved = vm_pageout_page_count + 1790 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768); 1791 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2; 1792 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved; 1793 vm_cnt.v_free_min += vm_cnt.v_free_reserved; 1794 vm_cnt.v_free_severe += vm_cnt.v_free_reserved; 1795 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2; 1796 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3) 1797 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3; 1798 1799 /* 1800 * Set the default wakeup threshold to be 10% above the minimum 1801 * page limit. This keeps the steady state out of shortfall. 1802 */ 1803 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11; 1804 1805 /* 1806 * Set interval in seconds for active scan. We want to visit each 1807 * page at least once every ten minutes. This is to prevent worst 1808 * case paging behaviors with stale active LRU. 1809 */ 1810 if (vm_pageout_update_period == 0) 1811 vm_pageout_update_period = 600; 1812 1813 /* XXX does not really belong here */ 1814 if (vm_page_max_wired == 0) 1815 vm_page_max_wired = vm_cnt.v_free_count / 3; 1816 } 1817 1818 /* 1819 * vm_pageout is the high level pageout daemon. 1820 */ 1821 static void 1822 vm_pageout(void) 1823 { 1824 int error; 1825 #if MAXMEMDOM > 1 1826 int i; 1827 #endif 1828 1829 swap_pager_swap_init(); 1830 #if MAXMEMDOM > 1 1831 for (i = 1; i < vm_ndomains; i++) { 1832 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, 1833 curproc, NULL, 0, 0, "dom%d", i); 1834 if (error != 0) { 1835 panic("starting pageout for domain %d, error %d\n", 1836 i, error); 1837 } 1838 } 1839 #endif 1840 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL, 1841 0, 0, "uma"); 1842 if (error != 0) 1843 panic("starting uma_reclaim helper, error %d\n", error); 1844 vm_pageout_worker((void *)(uintptr_t)0); 1845 } 1846 1847 /* 1848 * Unless the free page queue lock is held by the caller, this function 1849 * should be regarded as advisory. Specifically, the caller should 1850 * not msleep() on &vm_cnt.v_free_count following this function unless 1851 * the free page queue lock is held until the msleep() is performed. 1852 */ 1853 void 1854 pagedaemon_wakeup(void) 1855 { 1856 1857 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1858 vm_pages_needed = 1; 1859 wakeup(&vm_pages_needed); 1860 } 1861 } 1862 1863 #if !defined(NO_SWAPPING) 1864 static void 1865 vm_req_vmdaemon(int req) 1866 { 1867 static int lastrun = 0; 1868 1869 mtx_lock(&vm_daemon_mtx); 1870 vm_pageout_req_swapout |= req; 1871 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1872 wakeup(&vm_daemon_needed); 1873 lastrun = ticks; 1874 } 1875 mtx_unlock(&vm_daemon_mtx); 1876 } 1877 1878 static void 1879 vm_daemon(void) 1880 { 1881 struct rlimit rsslim; 1882 struct proc *p; 1883 struct thread *td; 1884 struct vmspace *vm; 1885 int breakout, swapout_flags, tryagain, attempts; 1886 #ifdef RACCT 1887 uint64_t rsize, ravailable; 1888 #endif 1889 1890 while (TRUE) { 1891 mtx_lock(&vm_daemon_mtx); 1892 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 1893 #ifdef RACCT 1894 racct_enable ? hz : 0 1895 #else 1896 0 1897 #endif 1898 ); 1899 swapout_flags = vm_pageout_req_swapout; 1900 vm_pageout_req_swapout = 0; 1901 mtx_unlock(&vm_daemon_mtx); 1902 if (swapout_flags) 1903 swapout_procs(swapout_flags); 1904 1905 /* 1906 * scan the processes for exceeding their rlimits or if 1907 * process is swapped out -- deactivate pages 1908 */ 1909 tryagain = 0; 1910 attempts = 0; 1911 again: 1912 attempts++; 1913 sx_slock(&allproc_lock); 1914 FOREACH_PROC_IN_SYSTEM(p) { 1915 vm_pindex_t limit, size; 1916 1917 /* 1918 * if this is a system process or if we have already 1919 * looked at this process, skip it. 1920 */ 1921 PROC_LOCK(p); 1922 if (p->p_state != PRS_NORMAL || 1923 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1924 PROC_UNLOCK(p); 1925 continue; 1926 } 1927 /* 1928 * if the process is in a non-running type state, 1929 * don't touch it. 1930 */ 1931 breakout = 0; 1932 FOREACH_THREAD_IN_PROC(p, td) { 1933 thread_lock(td); 1934 if (!TD_ON_RUNQ(td) && 1935 !TD_IS_RUNNING(td) && 1936 !TD_IS_SLEEPING(td) && 1937 !TD_IS_SUSPENDED(td)) { 1938 thread_unlock(td); 1939 breakout = 1; 1940 break; 1941 } 1942 thread_unlock(td); 1943 } 1944 if (breakout) { 1945 PROC_UNLOCK(p); 1946 continue; 1947 } 1948 /* 1949 * get a limit 1950 */ 1951 lim_rlimit_proc(p, RLIMIT_RSS, &rsslim); 1952 limit = OFF_TO_IDX( 1953 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1954 1955 /* 1956 * let processes that are swapped out really be 1957 * swapped out set the limit to nothing (will force a 1958 * swap-out.) 1959 */ 1960 if ((p->p_flag & P_INMEM) == 0) 1961 limit = 0; /* XXX */ 1962 vm = vmspace_acquire_ref(p); 1963 PROC_UNLOCK(p); 1964 if (vm == NULL) 1965 continue; 1966 1967 size = vmspace_resident_count(vm); 1968 if (size >= limit) { 1969 vm_pageout_map_deactivate_pages( 1970 &vm->vm_map, limit); 1971 } 1972 #ifdef RACCT 1973 if (racct_enable) { 1974 rsize = IDX_TO_OFF(size); 1975 PROC_LOCK(p); 1976 racct_set(p, RACCT_RSS, rsize); 1977 ravailable = racct_get_available(p, RACCT_RSS); 1978 PROC_UNLOCK(p); 1979 if (rsize > ravailable) { 1980 /* 1981 * Don't be overly aggressive; this 1982 * might be an innocent process, 1983 * and the limit could've been exceeded 1984 * by some memory hog. Don't try 1985 * to deactivate more than 1/4th 1986 * of process' resident set size. 1987 */ 1988 if (attempts <= 8) { 1989 if (ravailable < rsize - 1990 (rsize / 4)) { 1991 ravailable = rsize - 1992 (rsize / 4); 1993 } 1994 } 1995 vm_pageout_map_deactivate_pages( 1996 &vm->vm_map, 1997 OFF_TO_IDX(ravailable)); 1998 /* Update RSS usage after paging out. */ 1999 size = vmspace_resident_count(vm); 2000 rsize = IDX_TO_OFF(size); 2001 PROC_LOCK(p); 2002 racct_set(p, RACCT_RSS, rsize); 2003 PROC_UNLOCK(p); 2004 if (rsize > ravailable) 2005 tryagain = 1; 2006 } 2007 } 2008 #endif 2009 vmspace_free(vm); 2010 } 2011 sx_sunlock(&allproc_lock); 2012 if (tryagain != 0 && attempts <= 10) 2013 goto again; 2014 } 2015 } 2016 #endif /* !defined(NO_SWAPPING) */ 2017