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/smp.h> 94 #include <sys/vnode.h> 95 #include <sys/vmmeter.h> 96 #include <sys/rwlock.h> 97 #include <sys/sx.h> 98 #include <sys/sysctl.h> 99 100 #include <vm/vm.h> 101 #include <vm/vm_param.h> 102 #include <vm/vm_object.h> 103 #include <vm/vm_page.h> 104 #include <vm/vm_map.h> 105 #include <vm/vm_pageout.h> 106 #include <vm/vm_pager.h> 107 #include <vm/vm_phys.h> 108 #include <vm/swap_pager.h> 109 #include <vm/vm_extern.h> 110 #include <vm/uma.h> 111 112 /* 113 * System initialization 114 */ 115 116 /* the kernel process "vm_pageout"*/ 117 static void vm_pageout(void); 118 static int vm_pageout_clean(vm_page_t); 119 static void vm_pageout_scan(struct vm_domain *vmd, int pass); 120 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass); 121 122 struct proc *pageproc; 123 124 static struct kproc_desc page_kp = { 125 "pagedaemon", 126 vm_pageout, 127 &pageproc 128 }; 129 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, 130 &page_kp); 131 132 #if !defined(NO_SWAPPING) 133 /* the kernel process "vm_daemon"*/ 134 static void vm_daemon(void); 135 static struct proc *vmproc; 136 137 static struct kproc_desc vm_kp = { 138 "vmdaemon", 139 vm_daemon, 140 &vmproc 141 }; 142 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 143 #endif 144 145 146 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 147 int vm_pageout_deficit; /* Estimated number of pages deficit */ 148 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 149 150 #if !defined(NO_SWAPPING) 151 static int vm_pageout_req_swapout; /* XXX */ 152 static int vm_daemon_needed; 153 static struct mtx vm_daemon_mtx; 154 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 155 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 156 #endif 157 static int vm_max_launder = 32; 158 static int vm_pageout_stats_max; 159 static int vm_pageout_stats; 160 static int vm_pageout_stats_interval; 161 static int vm_pageout_full_stats; 162 static int vm_pageout_full_stats_interval; 163 static int defer_swap_pageouts; 164 static int disable_swap_pageouts; 165 166 #if defined(NO_SWAPPING) 167 static int vm_swap_enabled = 0; 168 static int vm_swap_idle_enabled = 0; 169 #else 170 static int vm_swap_enabled = 1; 171 static int vm_swap_idle_enabled = 0; 172 #endif 173 174 SYSCTL_INT(_vm, OID_AUTO, max_launder, 175 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 176 177 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 178 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 179 180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats, 181 CTLFLAG_RD, &vm_pageout_stats, 0, "Number of partial stats scans"); 182 183 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 184 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 185 186 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats, 187 CTLFLAG_RD, &vm_pageout_full_stats, 0, "Number of full stats scans"); 188 189 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 190 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 191 192 #if defined(NO_SWAPPING) 193 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 194 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 195 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 196 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 197 #else 198 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 199 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 200 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 201 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 202 #endif 203 204 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 205 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 206 207 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 208 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 209 210 static int pageout_lock_miss; 211 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 212 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 213 214 #define VM_PAGEOUT_PAGE_COUNT 16 215 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 216 217 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 218 SYSCTL_INT(_vm, OID_AUTO, max_wired, 219 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 220 221 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *); 222 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t, 223 vm_paddr_t); 224 #if !defined(NO_SWAPPING) 225 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 226 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 227 static void vm_req_vmdaemon(int req); 228 #endif 229 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *); 230 static void vm_pageout_page_stats(struct vm_domain *vmd); 231 232 /* 233 * Initialize a dummy page for marking the caller's place in the specified 234 * paging queue. In principle, this function only needs to set the flag 235 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count 236 * to one as safety precautions. 237 */ 238 static void 239 vm_pageout_init_marker(vm_page_t marker, u_short queue) 240 { 241 242 bzero(marker, sizeof(*marker)); 243 marker->flags = PG_MARKER; 244 marker->busy_lock = VPB_SINGLE_EXCLUSIVER; 245 marker->queue = queue; 246 marker->hold_count = 1; 247 } 248 249 /* 250 * vm_pageout_fallback_object_lock: 251 * 252 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is 253 * known to have failed and page queue must be either PQ_ACTIVE or 254 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 255 * while locking the vm object. Use marker page to detect page queue 256 * changes and maintain notion of next page on page queue. Return 257 * TRUE if no changes were detected, FALSE otherwise. vm object is 258 * locked on return. 259 * 260 * This function depends on both the lock portion of struct vm_object 261 * and normal struct vm_page being type stable. 262 */ 263 static boolean_t 264 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 265 { 266 struct vm_page marker; 267 struct vm_pagequeue *pq; 268 boolean_t unchanged; 269 u_short queue; 270 vm_object_t object; 271 272 queue = m->queue; 273 vm_pageout_init_marker(&marker, queue); 274 pq = vm_page_pagequeue(m); 275 object = m->object; 276 277 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 278 vm_pagequeue_unlock(pq); 279 vm_page_unlock(m); 280 VM_OBJECT_WLOCK(object); 281 vm_page_lock(m); 282 vm_pagequeue_lock(pq); 283 284 /* Page queue might have changed. */ 285 *next = TAILQ_NEXT(&marker, plinks.q); 286 unchanged = (m->queue == queue && 287 m->object == object && 288 &marker == TAILQ_NEXT(m, plinks.q)); 289 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 290 return (unchanged); 291 } 292 293 /* 294 * Lock the page while holding the page queue lock. Use marker page 295 * to detect page queue changes and maintain notion of next page on 296 * page queue. Return TRUE if no changes were detected, FALSE 297 * otherwise. The page is locked on return. The page queue lock might 298 * be dropped and reacquired. 299 * 300 * This function depends on normal struct vm_page being type stable. 301 */ 302 static boolean_t 303 vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 304 { 305 struct vm_page marker; 306 struct vm_pagequeue *pq; 307 boolean_t unchanged; 308 u_short queue; 309 310 vm_page_lock_assert(m, MA_NOTOWNED); 311 if (vm_page_trylock(m)) 312 return (TRUE); 313 314 queue = m->queue; 315 vm_pageout_init_marker(&marker, queue); 316 pq = vm_page_pagequeue(m); 317 318 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 319 vm_pagequeue_unlock(pq); 320 vm_page_lock(m); 321 vm_pagequeue_lock(pq); 322 323 /* Page queue might have changed. */ 324 *next = TAILQ_NEXT(&marker, plinks.q); 325 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, plinks.q)); 326 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 327 return (unchanged); 328 } 329 330 /* 331 * vm_pageout_clean: 332 * 333 * Clean the page and remove it from the laundry. 334 * 335 * We set the busy bit to cause potential page faults on this page to 336 * block. Note the careful timing, however, the busy bit isn't set till 337 * late and we cannot do anything that will mess with the page. 338 */ 339 static int 340 vm_pageout_clean(vm_page_t m) 341 { 342 vm_object_t object; 343 vm_page_t mc[2*vm_pageout_page_count], pb, ps; 344 int pageout_count; 345 int ib, is, page_base; 346 vm_pindex_t pindex = m->pindex; 347 348 vm_page_lock_assert(m, MA_OWNED); 349 object = m->object; 350 VM_OBJECT_ASSERT_WLOCKED(object); 351 352 /* 353 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 354 * with the new swapper, but we could have serious problems paging 355 * out other object types if there is insufficient memory. 356 * 357 * Unfortunately, checking free memory here is far too late, so the 358 * check has been moved up a procedural level. 359 */ 360 361 /* 362 * Can't clean the page if it's busy or held. 363 */ 364 vm_page_assert_unbusied(m); 365 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m)); 366 vm_page_unlock(m); 367 368 mc[vm_pageout_page_count] = pb = ps = m; 369 pageout_count = 1; 370 page_base = vm_pageout_page_count; 371 ib = 1; 372 is = 1; 373 374 /* 375 * Scan object for clusterable pages. 376 * 377 * We can cluster ONLY if: ->> the page is NOT 378 * clean, wired, busy, held, or mapped into a 379 * buffer, and one of the following: 380 * 1) The page is inactive, or a seldom used 381 * active page. 382 * -or- 383 * 2) we force the issue. 384 * 385 * During heavy mmap/modification loads the pageout 386 * daemon can really fragment the underlying file 387 * due to flushing pages out of order and not trying 388 * align the clusters (which leave sporatic out-of-order 389 * holes). To solve this problem we do the reverse scan 390 * first and attempt to align our cluster, then do a 391 * forward scan if room remains. 392 */ 393 more: 394 while (ib && pageout_count < vm_pageout_page_count) { 395 vm_page_t p; 396 397 if (ib > pindex) { 398 ib = 0; 399 break; 400 } 401 402 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) { 403 ib = 0; 404 break; 405 } 406 vm_page_lock(p); 407 vm_page_test_dirty(p); 408 if (p->dirty == 0 || 409 p->queue != PQ_INACTIVE || 410 p->hold_count != 0) { /* may be undergoing I/O */ 411 vm_page_unlock(p); 412 ib = 0; 413 break; 414 } 415 vm_page_unlock(p); 416 mc[--page_base] = pb = p; 417 ++pageout_count; 418 ++ib; 419 /* 420 * alignment boundry, stop here and switch directions. Do 421 * not clear ib. 422 */ 423 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 424 break; 425 } 426 427 while (pageout_count < vm_pageout_page_count && 428 pindex + is < object->size) { 429 vm_page_t p; 430 431 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p)) 432 break; 433 vm_page_lock(p); 434 vm_page_test_dirty(p); 435 if (p->dirty == 0 || 436 p->queue != PQ_INACTIVE || 437 p->hold_count != 0) { /* may be undergoing I/O */ 438 vm_page_unlock(p); 439 break; 440 } 441 vm_page_unlock(p); 442 mc[page_base + pageout_count] = ps = p; 443 ++pageout_count; 444 ++is; 445 } 446 447 /* 448 * If we exhausted our forward scan, continue with the reverse scan 449 * when possible, even past a page boundry. This catches boundry 450 * conditions. 451 */ 452 if (ib && pageout_count < vm_pageout_page_count) 453 goto more; 454 455 /* 456 * we allow reads during pageouts... 457 */ 458 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL, 459 NULL)); 460 } 461 462 /* 463 * vm_pageout_flush() - launder the given pages 464 * 465 * The given pages are laundered. Note that we setup for the start of 466 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 467 * reference count all in here rather then in the parent. If we want 468 * the parent to do more sophisticated things we may have to change 469 * the ordering. 470 * 471 * Returned runlen is the count of pages between mreq and first 472 * page after mreq with status VM_PAGER_AGAIN. 473 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 474 * for any page in runlen set. 475 */ 476 int 477 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 478 boolean_t *eio) 479 { 480 vm_object_t object = mc[0]->object; 481 int pageout_status[count]; 482 int numpagedout = 0; 483 int i, runlen; 484 485 VM_OBJECT_ASSERT_WLOCKED(object); 486 487 /* 488 * Initiate I/O. Bump the vm_page_t->busy counter and 489 * mark the pages read-only. 490 * 491 * We do not have to fixup the clean/dirty bits here... we can 492 * allow the pager to do it after the I/O completes. 493 * 494 * NOTE! mc[i]->dirty may be partial or fragmented due to an 495 * edge case with file fragments. 496 */ 497 for (i = 0; i < count; i++) { 498 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 499 ("vm_pageout_flush: partially invalid page %p index %d/%d", 500 mc[i], i, count)); 501 vm_page_sbusy(mc[i]); 502 pmap_remove_write(mc[i]); 503 } 504 vm_object_pip_add(object, count); 505 506 vm_pager_put_pages(object, mc, count, flags, pageout_status); 507 508 runlen = count - mreq; 509 if (eio != NULL) 510 *eio = FALSE; 511 for (i = 0; i < count; i++) { 512 vm_page_t mt = mc[i]; 513 514 KASSERT(pageout_status[i] == VM_PAGER_PEND || 515 !pmap_page_is_write_mapped(mt), 516 ("vm_pageout_flush: page %p is not write protected", mt)); 517 switch (pageout_status[i]) { 518 case VM_PAGER_OK: 519 case VM_PAGER_PEND: 520 numpagedout++; 521 break; 522 case VM_PAGER_BAD: 523 /* 524 * Page outside of range of object. Right now we 525 * essentially lose the changes by pretending it 526 * worked. 527 */ 528 vm_page_undirty(mt); 529 break; 530 case VM_PAGER_ERROR: 531 case VM_PAGER_FAIL: 532 /* 533 * If page couldn't be paged out, then reactivate the 534 * page so it doesn't clog the inactive list. (We 535 * will try paging out it again later). 536 */ 537 vm_page_lock(mt); 538 vm_page_activate(mt); 539 vm_page_unlock(mt); 540 if (eio != NULL && i >= mreq && i - mreq < runlen) 541 *eio = TRUE; 542 break; 543 case VM_PAGER_AGAIN: 544 if (i >= mreq && i - mreq < runlen) 545 runlen = i - mreq; 546 break; 547 } 548 549 /* 550 * If the operation is still going, leave the page busy to 551 * block all other accesses. Also, leave the paging in 552 * progress indicator set so that we don't attempt an object 553 * collapse. 554 */ 555 if (pageout_status[i] != VM_PAGER_PEND) { 556 vm_object_pip_wakeup(object); 557 vm_page_sunbusy(mt); 558 if (vm_page_count_severe()) { 559 vm_page_lock(mt); 560 vm_page_try_to_cache(mt); 561 vm_page_unlock(mt); 562 } 563 } 564 } 565 if (prunlen != NULL) 566 *prunlen = runlen; 567 return (numpagedout); 568 } 569 570 static boolean_t 571 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low, 572 vm_paddr_t high) 573 { 574 struct mount *mp; 575 struct vnode *vp; 576 vm_object_t object; 577 vm_paddr_t pa; 578 vm_page_t m, m_tmp, next; 579 580 vm_pagequeue_lock(pq); 581 TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) { 582 if ((m->flags & PG_MARKER) != 0) 583 continue; 584 pa = VM_PAGE_TO_PHYS(m); 585 if (pa < low || pa + PAGE_SIZE > high) 586 continue; 587 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) { 588 vm_page_unlock(m); 589 continue; 590 } 591 object = m->object; 592 if ((!VM_OBJECT_TRYWLOCK(object) && 593 (!vm_pageout_fallback_object_lock(m, &next) || 594 m->hold_count != 0)) || vm_page_busied(m)) { 595 vm_page_unlock(m); 596 VM_OBJECT_WUNLOCK(object); 597 continue; 598 } 599 vm_page_test_dirty(m); 600 if (m->dirty == 0 && object->ref_count != 0) 601 pmap_remove_all(m); 602 if (m->dirty != 0) { 603 vm_page_unlock(m); 604 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) { 605 VM_OBJECT_WUNLOCK(object); 606 continue; 607 } 608 if (object->type == OBJT_VNODE) { 609 vm_pagequeue_unlock(pq); 610 vp = object->handle; 611 vm_object_reference_locked(object); 612 VM_OBJECT_WUNLOCK(object); 613 (void)vn_start_write(vp, &mp, V_WAIT); 614 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 615 VM_OBJECT_WLOCK(object); 616 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 617 VM_OBJECT_WUNLOCK(object); 618 VOP_UNLOCK(vp, 0); 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_pagequeue_unlock(pq); 625 m_tmp = m; 626 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC, 627 0, NULL, NULL); 628 VM_OBJECT_WUNLOCK(object); 629 return (TRUE); 630 } 631 } else { 632 /* 633 * Dequeue here to prevent lock recursion in 634 * vm_page_cache(). 635 */ 636 vm_page_dequeue_locked(m); 637 vm_page_cache(m); 638 vm_page_unlock(m); 639 } 640 VM_OBJECT_WUNLOCK(object); 641 } 642 vm_pagequeue_unlock(pq); 643 return (FALSE); 644 } 645 646 /* 647 * Increase the number of cached pages. The specified value, "tries", 648 * determines which categories of pages are cached: 649 * 650 * 0: All clean, inactive pages within the specified physical address range 651 * are cached. Will not sleep. 652 * 1: The vm_lowmem handlers are called. All inactive pages within 653 * the specified physical address range are cached. May sleep. 654 * 2: The vm_lowmem handlers are called. All inactive and active pages 655 * within the specified physical address range are cached. May sleep. 656 */ 657 void 658 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high) 659 { 660 int actl, actmax, inactl, inactmax, dom, initial_dom; 661 static int start_dom = 0; 662 663 if (tries > 0) { 664 /* 665 * Decrease registered cache sizes. The vm_lowmem handlers 666 * may acquire locks and/or sleep, so they can only be invoked 667 * when "tries" is greater than zero. 668 */ 669 EVENTHANDLER_INVOKE(vm_lowmem, 0); 670 671 /* 672 * We do this explicitly after the caches have been drained 673 * above. 674 */ 675 uma_reclaim(); 676 } 677 678 /* 679 * Make the next scan start on the next domain. 680 */ 681 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains; 682 683 inactl = 0; 684 inactmax = cnt.v_inactive_count; 685 actl = 0; 686 actmax = tries < 2 ? 0 : cnt.v_active_count; 687 dom = initial_dom; 688 689 /* 690 * Scan domains in round-robin order, first inactive queues, 691 * then active. Since domain usually owns large physically 692 * contiguous chunk of memory, it makes sense to completely 693 * exhaust one domain before switching to next, while growing 694 * the pool of contiguous physical pages. 695 * 696 * Do not even start launder a domain which cannot contain 697 * the specified address range, as indicated by segments 698 * constituting the domain. 699 */ 700 again: 701 if (inactl < inactmax) { 702 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 703 low, high) && 704 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE], 705 tries, low, high)) { 706 inactl++; 707 goto again; 708 } 709 if (++dom == vm_ndomains) 710 dom = 0; 711 if (dom != initial_dom) 712 goto again; 713 } 714 if (actl < actmax) { 715 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 716 low, high) && 717 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE], 718 tries, low, high)) { 719 actl++; 720 goto again; 721 } 722 if (++dom == vm_ndomains) 723 dom = 0; 724 if (dom != initial_dom) 725 goto again; 726 } 727 } 728 729 #if !defined(NO_SWAPPING) 730 /* 731 * vm_pageout_object_deactivate_pages 732 * 733 * Deactivate enough pages to satisfy the inactive target 734 * requirements. 735 * 736 * The object and map must be locked. 737 */ 738 static void 739 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 740 long desired) 741 { 742 vm_object_t backing_object, object; 743 vm_page_t p; 744 int act_delta, remove_mode; 745 746 VM_OBJECT_ASSERT_LOCKED(first_object); 747 if ((first_object->flags & OBJ_FICTITIOUS) != 0) 748 return; 749 for (object = first_object;; object = backing_object) { 750 if (pmap_resident_count(pmap) <= desired) 751 goto unlock_return; 752 VM_OBJECT_ASSERT_LOCKED(object); 753 if ((object->flags & OBJ_UNMANAGED) != 0 || 754 object->paging_in_progress != 0) 755 goto unlock_return; 756 757 remove_mode = 0; 758 if (object->shadow_count > 1) 759 remove_mode = 1; 760 /* 761 * Scan the object's entire memory queue. 762 */ 763 TAILQ_FOREACH(p, &object->memq, listq) { 764 if (pmap_resident_count(pmap) <= desired) 765 goto unlock_return; 766 if (vm_page_busied(p)) 767 continue; 768 PCPU_INC(cnt.v_pdpages); 769 vm_page_lock(p); 770 if (p->wire_count != 0 || p->hold_count != 0 || 771 !pmap_page_exists_quick(pmap, p)) { 772 vm_page_unlock(p); 773 continue; 774 } 775 act_delta = pmap_ts_referenced(p); 776 if ((p->aflags & PGA_REFERENCED) != 0) { 777 if (act_delta == 0) 778 act_delta = 1; 779 vm_page_aflag_clear(p, PGA_REFERENCED); 780 } 781 if (p->queue != PQ_ACTIVE && act_delta != 0) { 782 vm_page_activate(p); 783 p->act_count += act_delta; 784 } else if (p->queue == PQ_ACTIVE) { 785 if (act_delta == 0) { 786 p->act_count -= min(p->act_count, 787 ACT_DECLINE); 788 if (!remove_mode && p->act_count == 0) { 789 pmap_remove_all(p); 790 vm_page_deactivate(p); 791 } else 792 vm_page_requeue(p); 793 } else { 794 vm_page_activate(p); 795 if (p->act_count < ACT_MAX - 796 ACT_ADVANCE) 797 p->act_count += ACT_ADVANCE; 798 vm_page_requeue(p); 799 } 800 } else if (p->queue == PQ_INACTIVE) 801 pmap_remove_all(p); 802 vm_page_unlock(p); 803 } 804 if ((backing_object = object->backing_object) == NULL) 805 goto unlock_return; 806 VM_OBJECT_RLOCK(backing_object); 807 if (object != first_object) 808 VM_OBJECT_RUNLOCK(object); 809 } 810 unlock_return: 811 if (object != first_object) 812 VM_OBJECT_RUNLOCK(object); 813 } 814 815 /* 816 * deactivate some number of pages in a map, try to do it fairly, but 817 * that is really hard to do. 818 */ 819 static void 820 vm_pageout_map_deactivate_pages(map, desired) 821 vm_map_t map; 822 long desired; 823 { 824 vm_map_entry_t tmpe; 825 vm_object_t obj, bigobj; 826 int nothingwired; 827 828 if (!vm_map_trylock(map)) 829 return; 830 831 bigobj = NULL; 832 nothingwired = TRUE; 833 834 /* 835 * first, search out the biggest object, and try to free pages from 836 * that. 837 */ 838 tmpe = map->header.next; 839 while (tmpe != &map->header) { 840 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 841 obj = tmpe->object.vm_object; 842 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { 843 if (obj->shadow_count <= 1 && 844 (bigobj == NULL || 845 bigobj->resident_page_count < obj->resident_page_count)) { 846 if (bigobj != NULL) 847 VM_OBJECT_RUNLOCK(bigobj); 848 bigobj = obj; 849 } else 850 VM_OBJECT_RUNLOCK(obj); 851 } 852 } 853 if (tmpe->wired_count > 0) 854 nothingwired = FALSE; 855 tmpe = tmpe->next; 856 } 857 858 if (bigobj != NULL) { 859 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 860 VM_OBJECT_RUNLOCK(bigobj); 861 } 862 /* 863 * Next, hunt around for other pages to deactivate. We actually 864 * do this search sort of wrong -- .text first is not the best idea. 865 */ 866 tmpe = map->header.next; 867 while (tmpe != &map->header) { 868 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 869 break; 870 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 871 obj = tmpe->object.vm_object; 872 if (obj != NULL) { 873 VM_OBJECT_RLOCK(obj); 874 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 875 VM_OBJECT_RUNLOCK(obj); 876 } 877 } 878 tmpe = tmpe->next; 879 } 880 881 /* 882 * Remove all mappings if a process is swapped out, this will free page 883 * table pages. 884 */ 885 if (desired == 0 && nothingwired) { 886 pmap_remove(vm_map_pmap(map), vm_map_min(map), 887 vm_map_max(map)); 888 } 889 vm_map_unlock(map); 890 } 891 #endif /* !defined(NO_SWAPPING) */ 892 893 /* 894 * vm_pageout_scan does the dirty work for the pageout daemon. 895 */ 896 static void 897 vm_pageout_scan(struct vm_domain *vmd, int pass) 898 { 899 vm_page_t m, next; 900 struct vm_pagequeue *pq; 901 int page_shortage, maxscan, pcount; 902 int addl_page_shortage; 903 vm_object_t object; 904 int act_delta; 905 int vnodes_skipped = 0; 906 int maxlaunder; 907 boolean_t queues_locked; 908 909 /* 910 * Decrease registered cache sizes. 911 */ 912 EVENTHANDLER_INVOKE(vm_lowmem, 0); 913 /* 914 * We do this explicitly after the caches have been drained above. 915 */ 916 uma_reclaim(); 917 918 /* 919 * The addl_page_shortage is the number of temporarily 920 * stuck pages in the inactive queue. In other words, the 921 * number of pages from the inactive count that should be 922 * discounted in setting the target for the active queue scan. 923 */ 924 addl_page_shortage = atomic_readandclear_int(&vm_pageout_deficit); 925 926 /* 927 * Calculate the number of pages we want to either free or move 928 * to the cache. 929 */ 930 page_shortage = vm_paging_target() + addl_page_shortage; 931 932 /* 933 * maxlaunder limits the number of dirty pages we flush per scan. 934 * For most systems a smaller value (16 or 32) is more robust under 935 * extreme memory and disk pressure because any unnecessary writes 936 * to disk can result in extreme performance degredation. However, 937 * systems with excessive dirty pages (especially when MAP_NOSYNC is 938 * used) will die horribly with limited laundering. If the pageout 939 * daemon cannot clean enough pages in the first pass, we let it go 940 * all out in succeeding passes. 941 */ 942 if ((maxlaunder = vm_max_launder) <= 1) 943 maxlaunder = 1; 944 if (pass) 945 maxlaunder = 10000; 946 947 /* 948 * Start scanning the inactive queue for pages we can move to the 949 * cache or free. The scan will stop when the target is reached or 950 * we have scanned the entire inactive queue. Note that m->act_count 951 * is not used to form decisions for the inactive queue, only for the 952 * active queue. 953 */ 954 pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; 955 maxscan = pq->pq_cnt; 956 vm_pagequeue_lock(pq); 957 queues_locked = TRUE; 958 for (m = TAILQ_FIRST(&pq->pq_pl); 959 m != NULL && maxscan-- > 0 && page_shortage > 0; 960 m = next) { 961 vm_pagequeue_assert_locked(pq); 962 KASSERT(queues_locked, ("unlocked queues")); 963 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m)); 964 965 PCPU_INC(cnt.v_pdpages); 966 next = TAILQ_NEXT(m, plinks.q); 967 968 /* 969 * skip marker pages 970 */ 971 if (m->flags & PG_MARKER) 972 continue; 973 974 KASSERT((m->flags & PG_FICTITIOUS) == 0, 975 ("Fictitious page %p cannot be in inactive queue", m)); 976 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 977 ("Unmanaged page %p cannot be in inactive queue", m)); 978 979 /* 980 * The page or object lock acquisitions fail if the 981 * page was removed from the queue or moved to a 982 * different position within the queue. In either 983 * case, addl_page_shortage should not be incremented. 984 */ 985 if (!vm_pageout_page_lock(m, &next)) { 986 vm_page_unlock(m); 987 continue; 988 } 989 object = m->object; 990 if (!VM_OBJECT_TRYWLOCK(object) && 991 !vm_pageout_fallback_object_lock(m, &next)) { 992 vm_page_unlock(m); 993 VM_OBJECT_WUNLOCK(object); 994 continue; 995 } 996 997 /* 998 * Don't mess with busy pages, keep them at at the 999 * front of the queue, most likely they are being 1000 * paged out. Increment addl_page_shortage for busy 1001 * pages, because they may leave the inactive queue 1002 * shortly after page scan is finished. 1003 */ 1004 if (vm_page_busied(m)) { 1005 vm_page_unlock(m); 1006 VM_OBJECT_WUNLOCK(object); 1007 addl_page_shortage++; 1008 continue; 1009 } 1010 1011 /* 1012 * We unlock the inactive page queue, invalidating the 1013 * 'next' pointer. Use our marker to remember our 1014 * place. 1015 */ 1016 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q); 1017 vm_pagequeue_unlock(pq); 1018 queues_locked = FALSE; 1019 1020 /* 1021 * We bump the activation count if the page has been 1022 * referenced while in the inactive queue. This makes 1023 * it less likely that the page will be added back to the 1024 * inactive queue prematurely again. Here we check the 1025 * page tables (or emulated bits, if any), given the upper 1026 * level VM system not knowing anything about existing 1027 * references. 1028 */ 1029 act_delta = 0; 1030 if ((m->aflags & PGA_REFERENCED) != 0) { 1031 vm_page_aflag_clear(m, PGA_REFERENCED); 1032 act_delta = 1; 1033 } 1034 if (object->ref_count != 0) { 1035 act_delta += pmap_ts_referenced(m); 1036 } else { 1037 KASSERT(!pmap_page_is_mapped(m), 1038 ("vm_pageout_scan: page %p is mapped", m)); 1039 } 1040 1041 /* 1042 * If the upper level VM system knows about any page 1043 * references, we reactivate the page or requeue it. 1044 */ 1045 if (act_delta != 0) { 1046 if (object->ref_count) { 1047 vm_page_activate(m); 1048 m->act_count += act_delta + ACT_ADVANCE; 1049 } else { 1050 vm_pagequeue_lock(pq); 1051 queues_locked = TRUE; 1052 vm_page_requeue_locked(m); 1053 } 1054 VM_OBJECT_WUNLOCK(object); 1055 vm_page_unlock(m); 1056 goto relock_queues; 1057 } 1058 1059 if (m->hold_count != 0) { 1060 vm_page_unlock(m); 1061 VM_OBJECT_WUNLOCK(object); 1062 1063 /* 1064 * Held pages are essentially stuck in the 1065 * queue. So, they ought to be discounted 1066 * from the inactive count. See the 1067 * calculation of the page_shortage for the 1068 * loop over the active queue below. 1069 */ 1070 addl_page_shortage++; 1071 goto relock_queues; 1072 } 1073 1074 /* 1075 * If the page appears to be clean at the machine-independent 1076 * layer, then remove all of its mappings from the pmap in 1077 * anticipation of placing it onto the cache queue. If, 1078 * however, any of the page's mappings allow write access, 1079 * then the page may still be modified until the last of those 1080 * mappings are removed. 1081 */ 1082 vm_page_test_dirty(m); 1083 if (m->dirty == 0 && object->ref_count != 0) 1084 pmap_remove_all(m); 1085 1086 if (m->valid == 0) { 1087 /* 1088 * Invalid pages can be easily freed 1089 */ 1090 vm_page_free(m); 1091 PCPU_INC(cnt.v_dfree); 1092 --page_shortage; 1093 } else if (m->dirty == 0) { 1094 /* 1095 * Clean pages can be placed onto the cache queue. 1096 * This effectively frees them. 1097 */ 1098 vm_page_cache(m); 1099 --page_shortage; 1100 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 1101 /* 1102 * Dirty pages need to be paged out, but flushing 1103 * a page is extremely expensive verses freeing 1104 * a clean page. Rather then artificially limiting 1105 * the number of pages we can flush, we instead give 1106 * dirty pages extra priority on the inactive queue 1107 * by forcing them to be cycled through the queue 1108 * twice before being flushed, after which the 1109 * (now clean) page will cycle through once more 1110 * before being freed. This significantly extends 1111 * the thrash point for a heavily loaded machine. 1112 */ 1113 m->flags |= PG_WINATCFLS; 1114 vm_pagequeue_lock(pq); 1115 queues_locked = TRUE; 1116 vm_page_requeue_locked(m); 1117 } else if (maxlaunder > 0) { 1118 /* 1119 * We always want to try to flush some dirty pages if 1120 * we encounter them, to keep the system stable. 1121 * Normally this number is small, but under extreme 1122 * pressure where there are insufficient clean pages 1123 * on the inactive queue, we may have to go all out. 1124 */ 1125 int swap_pageouts_ok; 1126 struct vnode *vp = NULL; 1127 struct mount *mp = NULL; 1128 1129 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 1130 swap_pageouts_ok = 1; 1131 } else { 1132 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 1133 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 1134 vm_page_count_min()); 1135 1136 } 1137 1138 /* 1139 * We don't bother paging objects that are "dead". 1140 * Those objects are in a "rundown" state. 1141 */ 1142 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 1143 vm_pagequeue_lock(pq); 1144 vm_page_unlock(m); 1145 VM_OBJECT_WUNLOCK(object); 1146 queues_locked = TRUE; 1147 vm_page_requeue_locked(m); 1148 goto relock_queues; 1149 } 1150 1151 /* 1152 * The object is already known NOT to be dead. It 1153 * is possible for the vget() to block the whole 1154 * pageout daemon, but the new low-memory handling 1155 * code should prevent it. 1156 * 1157 * The previous code skipped locked vnodes and, worse, 1158 * reordered pages in the queue. This results in 1159 * completely non-deterministic operation and, on a 1160 * busy system, can lead to extremely non-optimal 1161 * pageouts. For example, it can cause clean pages 1162 * to be freed and dirty pages to be moved to the end 1163 * of the queue. Since dirty pages are also moved to 1164 * the end of the queue once-cleaned, this gives 1165 * way too large a weighting to defering the freeing 1166 * of dirty pages. 1167 * 1168 * We can't wait forever for the vnode lock, we might 1169 * deadlock due to a vn_read() getting stuck in 1170 * vm_wait while holding this vnode. We skip the 1171 * vnode if we can't get it in a reasonable amount 1172 * of time. 1173 */ 1174 if (object->type == OBJT_VNODE) { 1175 vm_page_unlock(m); 1176 vp = object->handle; 1177 if (vp->v_type == VREG && 1178 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1179 mp = NULL; 1180 ++pageout_lock_miss; 1181 if (object->flags & OBJ_MIGHTBEDIRTY) 1182 vnodes_skipped++; 1183 goto unlock_and_continue; 1184 } 1185 KASSERT(mp != NULL, 1186 ("vp %p with NULL v_mount", vp)); 1187 vm_object_reference_locked(object); 1188 VM_OBJECT_WUNLOCK(object); 1189 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 1190 curthread)) { 1191 VM_OBJECT_WLOCK(object); 1192 ++pageout_lock_miss; 1193 if (object->flags & OBJ_MIGHTBEDIRTY) 1194 vnodes_skipped++; 1195 vp = NULL; 1196 goto unlock_and_continue; 1197 } 1198 VM_OBJECT_WLOCK(object); 1199 vm_page_lock(m); 1200 vm_pagequeue_lock(pq); 1201 queues_locked = TRUE; 1202 /* 1203 * The page might have been moved to another 1204 * queue during potential blocking in vget() 1205 * above. The page might have been freed and 1206 * reused for another vnode. 1207 */ 1208 if (m->queue != PQ_INACTIVE || 1209 m->object != object || 1210 TAILQ_NEXT(m, plinks.q) != &vmd->vmd_marker) { 1211 vm_page_unlock(m); 1212 if (object->flags & OBJ_MIGHTBEDIRTY) 1213 vnodes_skipped++; 1214 goto unlock_and_continue; 1215 } 1216 1217 /* 1218 * The page may have been busied during the 1219 * blocking in vget(). We don't move the 1220 * page back onto the end of the queue so that 1221 * statistics are more correct if we don't. 1222 */ 1223 if (vm_page_busied(m)) { 1224 vm_page_unlock(m); 1225 goto unlock_and_continue; 1226 } 1227 1228 /* 1229 * If the page has become held it might 1230 * be undergoing I/O, so skip it 1231 */ 1232 if (m->hold_count) { 1233 vm_page_unlock(m); 1234 vm_page_requeue_locked(m); 1235 if (object->flags & OBJ_MIGHTBEDIRTY) 1236 vnodes_skipped++; 1237 goto unlock_and_continue; 1238 } 1239 vm_pagequeue_unlock(pq); 1240 queues_locked = FALSE; 1241 } 1242 1243 /* 1244 * If a page is dirty, then it is either being washed 1245 * (but not yet cleaned) or it is still in the 1246 * laundry. If it is still in the laundry, then we 1247 * start the cleaning operation. 1248 * 1249 * decrement page_shortage on success to account for 1250 * the (future) cleaned page. Otherwise we could wind 1251 * up laundering or cleaning too many pages. 1252 */ 1253 if (vm_pageout_clean(m) != 0) { 1254 --page_shortage; 1255 --maxlaunder; 1256 } 1257 unlock_and_continue: 1258 vm_page_lock_assert(m, MA_NOTOWNED); 1259 VM_OBJECT_WUNLOCK(object); 1260 if (mp != NULL) { 1261 if (queues_locked) { 1262 vm_pagequeue_unlock(pq); 1263 queues_locked = FALSE; 1264 } 1265 if (vp != NULL) 1266 vput(vp); 1267 vm_object_deallocate(object); 1268 vn_finished_write(mp); 1269 } 1270 vm_page_lock_assert(m, MA_NOTOWNED); 1271 goto relock_queues; 1272 } 1273 vm_page_unlock(m); 1274 VM_OBJECT_WUNLOCK(object); 1275 relock_queues: 1276 if (!queues_locked) { 1277 vm_pagequeue_lock(pq); 1278 queues_locked = TRUE; 1279 } 1280 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q); 1281 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q); 1282 } 1283 vm_pagequeue_unlock(pq); 1284 1285 /* 1286 * Compute the number of pages we want to try to move from the 1287 * active queue to the inactive queue. 1288 */ 1289 page_shortage = vm_paging_target() + 1290 cnt.v_inactive_target - cnt.v_inactive_count; 1291 page_shortage += addl_page_shortage; 1292 1293 /* 1294 * Scan the active queue for things we can deactivate. We nominally 1295 * track the per-page activity counter and use it to locate 1296 * deactivation candidates. 1297 */ 1298 pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; 1299 pcount = pq->pq_cnt; 1300 vm_pagequeue_lock(pq); 1301 m = TAILQ_FIRST(&pq->pq_pl); 1302 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1303 1304 KASSERT(m->queue == PQ_ACTIVE, 1305 ("vm_pageout_scan: page %p isn't active", m)); 1306 1307 next = TAILQ_NEXT(m, plinks.q); 1308 if ((m->flags & PG_MARKER) != 0) { 1309 m = next; 1310 continue; 1311 } 1312 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1313 ("Fictitious page %p cannot be in active queue", m)); 1314 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1315 ("Unmanaged page %p cannot be in active queue", m)); 1316 if (!vm_pageout_page_lock(m, &next)) { 1317 vm_page_unlock(m); 1318 m = next; 1319 continue; 1320 } 1321 object = m->object; 1322 if (!VM_OBJECT_TRYWLOCK(object) && 1323 !vm_pageout_fallback_object_lock(m, &next)) { 1324 VM_OBJECT_WUNLOCK(object); 1325 vm_page_unlock(m); 1326 m = next; 1327 continue; 1328 } 1329 1330 /* 1331 * Don't deactivate pages that are busy. 1332 */ 1333 if (vm_page_busied(m) || m->hold_count != 0) { 1334 vm_page_unlock(m); 1335 VM_OBJECT_WUNLOCK(object); 1336 vm_page_requeue_locked(m); 1337 m = next; 1338 continue; 1339 } 1340 1341 /* 1342 * The count for pagedaemon pages is done after checking the 1343 * page for eligibility... 1344 */ 1345 PCPU_INC(cnt.v_pdpages); 1346 1347 /* 1348 * Check to see "how much" the page has been used. 1349 */ 1350 act_delta = 0; 1351 if (m->aflags & PGA_REFERENCED) { 1352 vm_page_aflag_clear(m, PGA_REFERENCED); 1353 act_delta += 1; 1354 } 1355 if (object->ref_count != 0) 1356 act_delta += pmap_ts_referenced(m); 1357 1358 /* 1359 * Advance or decay the act_count based on recent usage. 1360 */ 1361 if (act_delta) { 1362 m->act_count += ACT_ADVANCE + act_delta; 1363 if (m->act_count > ACT_MAX) 1364 m->act_count = ACT_MAX; 1365 } else { 1366 m->act_count -= min(m->act_count, ACT_DECLINE); 1367 act_delta = m->act_count; 1368 } 1369 1370 /* 1371 * Move this page to the tail of the active or inactive 1372 * queue depending on usage. 1373 */ 1374 if (act_delta == 0) { 1375 KASSERT(object->ref_count != 0 || 1376 !pmap_page_is_mapped(m), 1377 ("vm_pageout_scan: page %p is mapped", m)); 1378 /* Dequeue to avoid later lock recursion. */ 1379 vm_page_dequeue_locked(m); 1380 vm_page_deactivate(m); 1381 page_shortage--; 1382 } else 1383 vm_page_requeue_locked(m); 1384 vm_page_unlock(m); 1385 VM_OBJECT_WUNLOCK(object); 1386 m = next; 1387 } 1388 vm_pagequeue_unlock(pq); 1389 #if !defined(NO_SWAPPING) 1390 /* 1391 * Idle process swapout -- run once per second. 1392 */ 1393 if (vm_swap_idle_enabled) { 1394 static long lsec; 1395 if (time_second != lsec) { 1396 vm_req_vmdaemon(VM_SWAP_IDLE); 1397 lsec = time_second; 1398 } 1399 } 1400 #endif 1401 1402 /* 1403 * If we didn't get enough free pages, and we have skipped a vnode 1404 * in a writeable object, wakeup the sync daemon. And kick swapout 1405 * if we did not get enough free pages. 1406 */ 1407 if (vm_paging_target() > 0) { 1408 if (vnodes_skipped && vm_page_count_min()) 1409 (void) speedup_syncer(); 1410 #if !defined(NO_SWAPPING) 1411 if (vm_swap_enabled && vm_page_count_target()) 1412 vm_req_vmdaemon(VM_SWAP_NORMAL); 1413 #endif 1414 } 1415 1416 /* 1417 * If we are critically low on one of RAM or swap and low on 1418 * the other, kill the largest process. However, we avoid 1419 * doing this on the first pass in order to give ourselves a 1420 * chance to flush out dirty vnode-backed pages and to allow 1421 * active pages to be moved to the inactive queue and reclaimed. 1422 */ 1423 vm_pageout_mightbe_oom(vmd, pass); 1424 } 1425 1426 static int vm_pageout_oom_vote; 1427 1428 /* 1429 * The pagedaemon threads randlomly select one to perform the 1430 * OOM. Trying to kill processes before all pagedaemons 1431 * failed to reach free target is premature. 1432 */ 1433 static void 1434 vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass) 1435 { 1436 int old_vote; 1437 1438 if (pass == 0 || !((swap_pager_avail < 64 && vm_page_count_min()) || 1439 (swap_pager_full && vm_paging_target() > 0))) { 1440 if (vmd->vmd_oom) { 1441 vmd->vmd_oom = FALSE; 1442 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1443 } 1444 return; 1445 } 1446 1447 if (vmd->vmd_oom) 1448 return; 1449 1450 vmd->vmd_oom = TRUE; 1451 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); 1452 if (old_vote != vm_ndomains - 1) 1453 return; 1454 1455 /* 1456 * The current pagedaemon thread is the last in the quorum to 1457 * start OOM. Initiate the selection and signaling of the 1458 * victim. 1459 */ 1460 vm_pageout_oom(VM_OOM_MEM); 1461 1462 /* 1463 * After one round of OOM terror, recall our vote. On the 1464 * next pass, current pagedaemon would vote again if the low 1465 * memory condition is still there, due to vmd_oom being 1466 * false. 1467 */ 1468 vmd->vmd_oom = FALSE; 1469 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1470 } 1471 1472 void 1473 vm_pageout_oom(int shortage) 1474 { 1475 struct proc *p, *bigproc; 1476 vm_offset_t size, bigsize; 1477 struct thread *td; 1478 struct vmspace *vm; 1479 1480 /* 1481 * We keep the process bigproc locked once we find it to keep anyone 1482 * from messing with it; however, there is a possibility of 1483 * deadlock if process B is bigproc and one of it's child processes 1484 * attempts to propagate a signal to B while we are waiting for A's 1485 * lock while walking this list. To avoid this, we don't block on 1486 * the process lock but just skip a process if it is already locked. 1487 */ 1488 bigproc = NULL; 1489 bigsize = 0; 1490 sx_slock(&allproc_lock); 1491 FOREACH_PROC_IN_SYSTEM(p) { 1492 int breakout; 1493 1494 if (PROC_TRYLOCK(p) == 0) 1495 continue; 1496 /* 1497 * If this is a system, protected or killed process, skip it. 1498 */ 1499 if (p->p_state != PRS_NORMAL || 1500 (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1501 (p->p_pid == 1) || P_KILLED(p) || 1502 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1503 PROC_UNLOCK(p); 1504 continue; 1505 } 1506 /* 1507 * If the process is in a non-running type state, 1508 * don't touch it. Check all the threads individually. 1509 */ 1510 breakout = 0; 1511 FOREACH_THREAD_IN_PROC(p, td) { 1512 thread_lock(td); 1513 if (!TD_ON_RUNQ(td) && 1514 !TD_IS_RUNNING(td) && 1515 !TD_IS_SLEEPING(td) && 1516 !TD_IS_SUSPENDED(td)) { 1517 thread_unlock(td); 1518 breakout = 1; 1519 break; 1520 } 1521 thread_unlock(td); 1522 } 1523 if (breakout) { 1524 PROC_UNLOCK(p); 1525 continue; 1526 } 1527 /* 1528 * get the process size 1529 */ 1530 vm = vmspace_acquire_ref(p); 1531 if (vm == NULL) { 1532 PROC_UNLOCK(p); 1533 continue; 1534 } 1535 if (!vm_map_trylock_read(&vm->vm_map)) { 1536 vmspace_free(vm); 1537 PROC_UNLOCK(p); 1538 continue; 1539 } 1540 size = vmspace_swap_count(vm); 1541 vm_map_unlock_read(&vm->vm_map); 1542 if (shortage == VM_OOM_MEM) 1543 size += vmspace_resident_count(vm); 1544 vmspace_free(vm); 1545 /* 1546 * if the this process is bigger than the biggest one 1547 * remember it. 1548 */ 1549 if (size > bigsize) { 1550 if (bigproc != NULL) 1551 PROC_UNLOCK(bigproc); 1552 bigproc = p; 1553 bigsize = size; 1554 } else 1555 PROC_UNLOCK(p); 1556 } 1557 sx_sunlock(&allproc_lock); 1558 if (bigproc != NULL) { 1559 killproc(bigproc, "out of swap space"); 1560 sched_nice(bigproc, PRIO_MIN); 1561 PROC_UNLOCK(bigproc); 1562 wakeup(&cnt.v_free_count); 1563 } 1564 } 1565 1566 /* 1567 * This routine tries to maintain the pseudo LRU active queue, 1568 * so that during long periods of time where there is no paging, 1569 * that some statistic accumulation still occurs. This code 1570 * helps the situation where paging just starts to occur. 1571 */ 1572 static void 1573 vm_pageout_page_stats(struct vm_domain *vmd) 1574 { 1575 struct vm_pagequeue *pq; 1576 vm_object_t object; 1577 vm_page_t m, next; 1578 int pcount, tpcount; /* Number of pages to check */ 1579 int actcount, page_shortage; 1580 1581 page_shortage = 1582 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1583 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1584 1585 if (page_shortage <= 0) 1586 return; 1587 1588 pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; 1589 1590 /* 1591 * pcount limits the depth of the queue scan. In particular, 1592 * for the full scan, it prevents the iteration from looking 1593 * into the requeued pages. The limit is not exact since the 1594 * page queue lock is dropped during the iteration. 1595 */ 1596 pcount = pq->pq_cnt; 1597 vmd->vmd_fullintervalcount += vm_pageout_stats_interval; 1598 if (vmd->vmd_fullintervalcount < vm_pageout_full_stats_interval) { 1599 atomic_add_int(&vm_pageout_stats, 1); 1600 tpcount = (int64_t)vm_pageout_stats_max * pcount / 1601 vmd->vmd_page_count; 1602 if (pcount > tpcount) 1603 pcount = tpcount; 1604 } else { 1605 atomic_add_int(&vm_pageout_full_stats, 1); 1606 vmd->vmd_fullintervalcount = 0; 1607 } 1608 1609 vm_pagequeue_lock(pq); 1610 m = TAILQ_FIRST(&pq->pq_pl); 1611 while (m != NULL && pcount-- > 0) { 1612 KASSERT(m->queue == PQ_ACTIVE, 1613 ("vm_pageout_page_stats: page %p isn't active", m)); 1614 1615 next = TAILQ_NEXT(m, plinks.q); 1616 if ((m->flags & PG_MARKER) != 0) { 1617 m = next; 1618 continue; 1619 } 1620 vm_page_lock_assert(m, MA_NOTOWNED); 1621 if (!vm_pageout_page_lock(m, &next)) { 1622 vm_page_unlock(m); 1623 m = next; 1624 continue; 1625 } 1626 object = m->object; 1627 if (!VM_OBJECT_TRYWLOCK(object) && 1628 !vm_pageout_fallback_object_lock(m, &next)) { 1629 VM_OBJECT_WUNLOCK(object); 1630 vm_page_unlock(m); 1631 m = next; 1632 continue; 1633 } 1634 1635 /* 1636 * Don't deactivate pages that are busy or held. 1637 */ 1638 if (vm_page_busied(m) || m->hold_count != 0) { 1639 vm_page_unlock(m); 1640 VM_OBJECT_WUNLOCK(object); 1641 vm_page_requeue_locked(m); 1642 m = next; 1643 continue; 1644 } 1645 1646 actcount = 0; 1647 if (m->aflags & PGA_REFERENCED) { 1648 vm_page_aflag_clear(m, PGA_REFERENCED); 1649 actcount += 1; 1650 } 1651 1652 actcount += pmap_ts_referenced(m); 1653 if (actcount != 0) { 1654 m->act_count += ACT_ADVANCE + actcount; 1655 if (m->act_count > ACT_MAX) 1656 m->act_count = ACT_MAX; 1657 vm_page_requeue_locked(m); 1658 } else { 1659 if (m->act_count == 0) { 1660 /* 1661 * We turn off page access, so that we have 1662 * more accurate RSS stats. We don't do this 1663 * in the normal page deactivation when the 1664 * system is loaded VM wise, because the 1665 * cost of the large number of page protect 1666 * operations would be higher than the value 1667 * of doing the operation. 1668 */ 1669 pmap_remove_all(m); 1670 /* Dequeue to avoid later lock recursion. */ 1671 vm_page_dequeue_locked(m); 1672 vm_page_deactivate(m); 1673 } else { 1674 m->act_count -= min(m->act_count, ACT_DECLINE); 1675 vm_page_requeue_locked(m); 1676 } 1677 } 1678 vm_page_unlock(m); 1679 VM_OBJECT_WUNLOCK(object); 1680 m = next; 1681 } 1682 vm_pagequeue_unlock(pq); 1683 } 1684 1685 static void 1686 vm_pageout_worker(void *arg) 1687 { 1688 struct vm_domain *domain; 1689 struct pcpu *pc; 1690 int cpu, error, domidx; 1691 1692 domidx = (uintptr_t)arg; 1693 domain = &vm_dom[domidx]; 1694 1695 /* 1696 * XXXKIB The bind is rather arbitrary. With some minor 1697 * complications, we could assign the cpuset consisting of all 1698 * CPUs in the same domain. In fact, it even does not matter 1699 * if the CPU we bind to is in the affinity domain of this 1700 * page queue, we only need to establish the fair distribution 1701 * of pagedaemon threads among CPUs. 1702 * 1703 * XXXKIB It would be useful to allocate vm_pages for the 1704 * domain from the domain, and put pcpu area into the page 1705 * owned by the domain. 1706 */ 1707 if (mem_affinity != NULL) { 1708 CPU_FOREACH(cpu) { 1709 pc = pcpu_find(cpu); 1710 if (pc->pc_domain == domidx) { 1711 thread_lock(curthread); 1712 sched_bind(curthread, cpu); 1713 thread_unlock(curthread); 1714 break; 1715 } 1716 } 1717 } 1718 1719 KASSERT(domain->vmd_segs != 0, ("domain without segments")); 1720 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE); 1721 1722 /* 1723 * The pageout daemon worker is never done, so loop forever. 1724 */ 1725 while (TRUE) { 1726 /* 1727 * If we have enough free memory, wakeup waiters. Do 1728 * not clear vm_pages_needed until we reach our target, 1729 * otherwise we may be woken up over and over again and 1730 * waste a lot of cpu. 1731 */ 1732 mtx_lock(&vm_page_queue_free_mtx); 1733 if (vm_pages_needed && !vm_page_count_min()) { 1734 if (!vm_paging_needed()) 1735 vm_pages_needed = 0; 1736 wakeup(&cnt.v_free_count); 1737 } 1738 if (vm_pages_needed) { 1739 /* 1740 * Still not done, take a second pass without waiting 1741 * (unlimited dirty cleaning), otherwise sleep a bit 1742 * and try again. 1743 */ 1744 ++(domain->vmd_pass); 1745 if (domain->vmd_pass > 1) 1746 msleep(&vm_pages_needed, 1747 &vm_page_queue_free_mtx, PVM, "psleep", 1748 hz / 2); 1749 } else { 1750 /* 1751 * Good enough, sleep & handle stats. Prime the pass 1752 * for the next run. 1753 */ 1754 if (domain->vmd_pass > 1) 1755 domain->vmd_pass = 1; 1756 else 1757 domain->vmd_pass = 0; 1758 error = msleep(&vm_pages_needed, 1759 &vm_page_queue_free_mtx, PVM, "psleep", 1760 vm_pageout_stats_interval * hz); 1761 if (error && !vm_pages_needed) { 1762 mtx_unlock(&vm_page_queue_free_mtx); 1763 domain->vmd_pass = 0; 1764 vm_pageout_page_stats(domain); 1765 continue; 1766 } 1767 } 1768 if (vm_pages_needed) 1769 cnt.v_pdwakeups++; 1770 mtx_unlock(&vm_page_queue_free_mtx); 1771 vm_pageout_scan(domain, domain->vmd_pass); 1772 } 1773 } 1774 1775 /* 1776 * vm_pageout is the high level pageout daemon. 1777 */ 1778 static void 1779 vm_pageout(void) 1780 { 1781 #if MAXMEMDOM > 1 1782 int error, i; 1783 #endif 1784 1785 /* 1786 * Initialize some paging parameters. 1787 */ 1788 cnt.v_interrupt_free_min = 2; 1789 if (cnt.v_page_count < 2000) 1790 vm_pageout_page_count = 8; 1791 1792 /* 1793 * v_free_reserved needs to include enough for the largest 1794 * swap pager structures plus enough for any pv_entry structs 1795 * when paging. 1796 */ 1797 if (cnt.v_page_count > 1024) 1798 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1799 else 1800 cnt.v_free_min = 4; 1801 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1802 cnt.v_interrupt_free_min; 1803 cnt.v_free_reserved = vm_pageout_page_count + 1804 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1805 cnt.v_free_severe = cnt.v_free_min / 2; 1806 cnt.v_free_min += cnt.v_free_reserved; 1807 cnt.v_free_severe += cnt.v_free_reserved; 1808 1809 /* 1810 * v_free_target and v_cache_min control pageout hysteresis. Note 1811 * that these are more a measure of the VM cache queue hysteresis 1812 * then the VM free queue. Specifically, v_free_target is the 1813 * high water mark (free+cache pages). 1814 * 1815 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1816 * low water mark, while v_free_min is the stop. v_cache_min must 1817 * be big enough to handle memory needs while the pageout daemon 1818 * is signalled and run to free more pages. 1819 */ 1820 if (cnt.v_free_count > 6144) 1821 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1822 else 1823 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1824 1825 if (cnt.v_free_count > 2048) { 1826 cnt.v_cache_min = cnt.v_free_target; 1827 cnt.v_cache_max = 2 * cnt.v_cache_min; 1828 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1829 } else { 1830 cnt.v_cache_min = 0; 1831 cnt.v_cache_max = 0; 1832 cnt.v_inactive_target = cnt.v_free_count / 4; 1833 } 1834 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1835 cnt.v_inactive_target = cnt.v_free_count / 3; 1836 1837 /* XXX does not really belong here */ 1838 if (vm_page_max_wired == 0) 1839 vm_page_max_wired = cnt.v_free_count / 3; 1840 1841 if (vm_pageout_stats_max == 0) 1842 vm_pageout_stats_max = cnt.v_free_target; 1843 1844 /* 1845 * Set interval in seconds for stats scan. 1846 */ 1847 if (vm_pageout_stats_interval == 0) 1848 vm_pageout_stats_interval = 5; 1849 if (vm_pageout_full_stats_interval == 0) 1850 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1851 1852 swap_pager_swap_init(); 1853 #if MAXMEMDOM > 1 1854 for (i = 1; i < vm_ndomains; i++) { 1855 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, 1856 curproc, NULL, 0, 0, "dom%d", i); 1857 if (error != 0) { 1858 panic("starting pageout for domain %d, error %d\n", 1859 i, error); 1860 } 1861 } 1862 #endif 1863 vm_pageout_worker((uintptr_t)0); 1864 } 1865 1866 /* 1867 * Unless the free page queue lock is held by the caller, this function 1868 * should be regarded as advisory. Specifically, the caller should 1869 * not msleep() on &cnt.v_free_count following this function unless 1870 * the free page queue lock is held until the msleep() is performed. 1871 */ 1872 void 1873 pagedaemon_wakeup(void) 1874 { 1875 1876 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1877 vm_pages_needed = 1; 1878 wakeup(&vm_pages_needed); 1879 } 1880 } 1881 1882 #if !defined(NO_SWAPPING) 1883 static void 1884 vm_req_vmdaemon(int req) 1885 { 1886 static int lastrun = 0; 1887 1888 mtx_lock(&vm_daemon_mtx); 1889 vm_pageout_req_swapout |= req; 1890 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1891 wakeup(&vm_daemon_needed); 1892 lastrun = ticks; 1893 } 1894 mtx_unlock(&vm_daemon_mtx); 1895 } 1896 1897 static void 1898 vm_daemon(void) 1899 { 1900 struct rlimit rsslim; 1901 struct proc *p; 1902 struct thread *td; 1903 struct vmspace *vm; 1904 int breakout, swapout_flags, tryagain, attempts; 1905 #ifdef RACCT 1906 uint64_t rsize, ravailable; 1907 #endif 1908 1909 while (TRUE) { 1910 mtx_lock(&vm_daemon_mtx); 1911 #ifdef RACCT 1912 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz); 1913 #else 1914 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1915 #endif 1916 swapout_flags = vm_pageout_req_swapout; 1917 vm_pageout_req_swapout = 0; 1918 mtx_unlock(&vm_daemon_mtx); 1919 if (swapout_flags) 1920 swapout_procs(swapout_flags); 1921 1922 /* 1923 * scan the processes for exceeding their rlimits or if 1924 * process is swapped out -- deactivate pages 1925 */ 1926 tryagain = 0; 1927 attempts = 0; 1928 again: 1929 attempts++; 1930 sx_slock(&allproc_lock); 1931 FOREACH_PROC_IN_SYSTEM(p) { 1932 vm_pindex_t limit, size; 1933 1934 /* 1935 * if this is a system process or if we have already 1936 * looked at this process, skip it. 1937 */ 1938 PROC_LOCK(p); 1939 if (p->p_state != PRS_NORMAL || 1940 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1941 PROC_UNLOCK(p); 1942 continue; 1943 } 1944 /* 1945 * if the process is in a non-running type state, 1946 * don't touch it. 1947 */ 1948 breakout = 0; 1949 FOREACH_THREAD_IN_PROC(p, td) { 1950 thread_lock(td); 1951 if (!TD_ON_RUNQ(td) && 1952 !TD_IS_RUNNING(td) && 1953 !TD_IS_SLEEPING(td) && 1954 !TD_IS_SUSPENDED(td)) { 1955 thread_unlock(td); 1956 breakout = 1; 1957 break; 1958 } 1959 thread_unlock(td); 1960 } 1961 if (breakout) { 1962 PROC_UNLOCK(p); 1963 continue; 1964 } 1965 /* 1966 * get a limit 1967 */ 1968 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1969 limit = OFF_TO_IDX( 1970 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1971 1972 /* 1973 * let processes that are swapped out really be 1974 * swapped out set the limit to nothing (will force a 1975 * swap-out.) 1976 */ 1977 if ((p->p_flag & P_INMEM) == 0) 1978 limit = 0; /* XXX */ 1979 vm = vmspace_acquire_ref(p); 1980 PROC_UNLOCK(p); 1981 if (vm == NULL) 1982 continue; 1983 1984 size = vmspace_resident_count(vm); 1985 if (size >= limit) { 1986 vm_pageout_map_deactivate_pages( 1987 &vm->vm_map, limit); 1988 } 1989 #ifdef RACCT 1990 rsize = IDX_TO_OFF(size); 1991 PROC_LOCK(p); 1992 racct_set(p, RACCT_RSS, rsize); 1993 ravailable = racct_get_available(p, RACCT_RSS); 1994 PROC_UNLOCK(p); 1995 if (rsize > ravailable) { 1996 /* 1997 * Don't be overly aggressive; this might be 1998 * an innocent process, and the limit could've 1999 * been exceeded by some memory hog. Don't 2000 * try to deactivate more than 1/4th of process' 2001 * resident set size. 2002 */ 2003 if (attempts <= 8) { 2004 if (ravailable < rsize - (rsize / 4)) 2005 ravailable = rsize - (rsize / 4); 2006 } 2007 vm_pageout_map_deactivate_pages( 2008 &vm->vm_map, OFF_TO_IDX(ravailable)); 2009 /* Update RSS usage after paging out. */ 2010 size = vmspace_resident_count(vm); 2011 rsize = IDX_TO_OFF(size); 2012 PROC_LOCK(p); 2013 racct_set(p, RACCT_RSS, rsize); 2014 PROC_UNLOCK(p); 2015 if (rsize > ravailable) 2016 tryagain = 1; 2017 } 2018 #endif 2019 vmspace_free(vm); 2020 } 2021 sx_sunlock(&allproc_lock); 2022 if (tryagain != 0 && attempts <= 10) 2023 goto again; 2024 } 2025 } 2026 #endif /* !defined(NO_SWAPPING) */ 2027