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 int vm_pageout_wakeup_thresh; 150 151 #if !defined(NO_SWAPPING) 152 static int vm_pageout_req_swapout; /* XXX */ 153 static int vm_daemon_needed; 154 static struct mtx vm_daemon_mtx; 155 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 156 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 157 #endif 158 static int vm_max_launder = 32; 159 static int vm_pageout_update_period; 160 static int defer_swap_pageouts; 161 static int disable_swap_pageouts; 162 static int lowmem_period = 10; 163 static int lowmem_ticks; 164 165 #if defined(NO_SWAPPING) 166 static int vm_swap_enabled = 0; 167 static int vm_swap_idle_enabled = 0; 168 #else 169 static int vm_swap_enabled = 1; 170 static int vm_swap_idle_enabled = 0; 171 #endif 172 173 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh, 174 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0, 175 "free page threshold for waking up the pageout daemon"); 176 177 SYSCTL_INT(_vm, OID_AUTO, max_launder, 178 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 179 180 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, 181 CTLFLAG_RW, &vm_pageout_update_period, 0, 182 "Maximum active LRU update period"); 183 184 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0, 185 "Low memory callback period"); 186 187 #if defined(NO_SWAPPING) 188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 189 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 191 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 192 #else 193 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 194 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 195 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 196 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 197 #endif 198 199 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 200 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 201 202 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 203 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 204 205 static int pageout_lock_miss; 206 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 207 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 208 209 #define VM_PAGEOUT_PAGE_COUNT 16 210 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 211 212 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 213 SYSCTL_INT(_vm, OID_AUTO, max_wired, 214 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 215 216 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *); 217 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t, 218 vm_paddr_t); 219 #if !defined(NO_SWAPPING) 220 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 221 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 222 static void vm_req_vmdaemon(int req); 223 #endif 224 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *); 225 226 /* 227 * Initialize a dummy page for marking the caller's place in the specified 228 * paging queue. In principle, this function only needs to set the flag 229 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count 230 * to one as safety precautions. 231 */ 232 static void 233 vm_pageout_init_marker(vm_page_t marker, u_short queue) 234 { 235 236 bzero(marker, sizeof(*marker)); 237 marker->flags = PG_MARKER; 238 marker->busy_lock = VPB_SINGLE_EXCLUSIVER; 239 marker->queue = queue; 240 marker->hold_count = 1; 241 } 242 243 /* 244 * vm_pageout_fallback_object_lock: 245 * 246 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is 247 * known to have failed and page queue must be either PQ_ACTIVE or 248 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 249 * while locking the vm object. Use marker page to detect page queue 250 * changes and maintain notion of next page on page queue. Return 251 * TRUE if no changes were detected, FALSE otherwise. vm object is 252 * locked on return. 253 * 254 * This function depends on both the lock portion of struct vm_object 255 * and normal struct vm_page being type stable. 256 */ 257 static boolean_t 258 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 259 { 260 struct vm_page marker; 261 struct vm_pagequeue *pq; 262 boolean_t unchanged; 263 u_short queue; 264 vm_object_t object; 265 266 queue = m->queue; 267 vm_pageout_init_marker(&marker, queue); 268 pq = vm_page_pagequeue(m); 269 object = m->object; 270 271 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 272 vm_pagequeue_unlock(pq); 273 vm_page_unlock(m); 274 VM_OBJECT_WLOCK(object); 275 vm_page_lock(m); 276 vm_pagequeue_lock(pq); 277 278 /* Page queue might have changed. */ 279 *next = TAILQ_NEXT(&marker, plinks.q); 280 unchanged = (m->queue == queue && 281 m->object == object && 282 &marker == TAILQ_NEXT(m, plinks.q)); 283 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 284 return (unchanged); 285 } 286 287 /* 288 * Lock the page while holding the page queue lock. Use marker page 289 * to detect page queue changes and maintain notion of next page on 290 * page queue. Return TRUE if no changes were detected, FALSE 291 * otherwise. The page is locked on return. The page queue lock might 292 * be dropped and reacquired. 293 * 294 * This function depends on normal struct vm_page being type stable. 295 */ 296 static boolean_t 297 vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 298 { 299 struct vm_page marker; 300 struct vm_pagequeue *pq; 301 boolean_t unchanged; 302 u_short queue; 303 304 vm_page_lock_assert(m, MA_NOTOWNED); 305 if (vm_page_trylock(m)) 306 return (TRUE); 307 308 queue = m->queue; 309 vm_pageout_init_marker(&marker, queue); 310 pq = vm_page_pagequeue(m); 311 312 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 313 vm_pagequeue_unlock(pq); 314 vm_page_lock(m); 315 vm_pagequeue_lock(pq); 316 317 /* Page queue might have changed. */ 318 *next = TAILQ_NEXT(&marker, plinks.q); 319 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, plinks.q)); 320 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 321 return (unchanged); 322 } 323 324 /* 325 * vm_pageout_clean: 326 * 327 * Clean the page and remove it from the laundry. 328 * 329 * We set the busy bit to cause potential page faults on this page to 330 * block. Note the careful timing, however, the busy bit isn't set till 331 * late and we cannot do anything that will mess with the page. 332 */ 333 static int 334 vm_pageout_clean(vm_page_t m) 335 { 336 vm_object_t object; 337 vm_page_t mc[2*vm_pageout_page_count], pb, ps; 338 int pageout_count; 339 int ib, is, page_base; 340 vm_pindex_t pindex = m->pindex; 341 342 vm_page_lock_assert(m, MA_OWNED); 343 object = m->object; 344 VM_OBJECT_ASSERT_WLOCKED(object); 345 346 /* 347 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 348 * with the new swapper, but we could have serious problems paging 349 * out other object types if there is insufficient memory. 350 * 351 * Unfortunately, checking free memory here is far too late, so the 352 * check has been moved up a procedural level. 353 */ 354 355 /* 356 * Can't clean the page if it's busy or held. 357 */ 358 vm_page_assert_unbusied(m); 359 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m)); 360 vm_page_unlock(m); 361 362 mc[vm_pageout_page_count] = pb = ps = m; 363 pageout_count = 1; 364 page_base = vm_pageout_page_count; 365 ib = 1; 366 is = 1; 367 368 /* 369 * Scan object for clusterable pages. 370 * 371 * We can cluster ONLY if: ->> the page is NOT 372 * clean, wired, busy, held, or mapped into a 373 * buffer, and one of the following: 374 * 1) The page is inactive, or a seldom used 375 * active page. 376 * -or- 377 * 2) we force the issue. 378 * 379 * During heavy mmap/modification loads the pageout 380 * daemon can really fragment the underlying file 381 * due to flushing pages out of order and not trying 382 * align the clusters (which leave sporatic out-of-order 383 * holes). To solve this problem we do the reverse scan 384 * first and attempt to align our cluster, then do a 385 * forward scan if room remains. 386 */ 387 more: 388 while (ib && pageout_count < vm_pageout_page_count) { 389 vm_page_t p; 390 391 if (ib > pindex) { 392 ib = 0; 393 break; 394 } 395 396 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) { 397 ib = 0; 398 break; 399 } 400 vm_page_lock(p); 401 vm_page_test_dirty(p); 402 if (p->dirty == 0 || 403 p->queue != PQ_INACTIVE || 404 p->hold_count != 0) { /* may be undergoing I/O */ 405 vm_page_unlock(p); 406 ib = 0; 407 break; 408 } 409 vm_page_unlock(p); 410 mc[--page_base] = pb = p; 411 ++pageout_count; 412 ++ib; 413 /* 414 * alignment boundry, stop here and switch directions. Do 415 * not clear ib. 416 */ 417 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 418 break; 419 } 420 421 while (pageout_count < vm_pageout_page_count && 422 pindex + is < object->size) { 423 vm_page_t p; 424 425 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p)) 426 break; 427 vm_page_lock(p); 428 vm_page_test_dirty(p); 429 if (p->dirty == 0 || 430 p->queue != PQ_INACTIVE || 431 p->hold_count != 0) { /* may be undergoing I/O */ 432 vm_page_unlock(p); 433 break; 434 } 435 vm_page_unlock(p); 436 mc[page_base + pageout_count] = ps = p; 437 ++pageout_count; 438 ++is; 439 } 440 441 /* 442 * If we exhausted our forward scan, continue with the reverse scan 443 * when possible, even past a page boundry. This catches boundry 444 * conditions. 445 */ 446 if (ib && pageout_count < vm_pageout_page_count) 447 goto more; 448 449 /* 450 * we allow reads during pageouts... 451 */ 452 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL, 453 NULL)); 454 } 455 456 /* 457 * vm_pageout_flush() - launder the given pages 458 * 459 * The given pages are laundered. Note that we setup for the start of 460 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 461 * reference count all in here rather then in the parent. If we want 462 * the parent to do more sophisticated things we may have to change 463 * the ordering. 464 * 465 * Returned runlen is the count of pages between mreq and first 466 * page after mreq with status VM_PAGER_AGAIN. 467 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 468 * for any page in runlen set. 469 */ 470 int 471 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 472 boolean_t *eio) 473 { 474 vm_object_t object = mc[0]->object; 475 int pageout_status[count]; 476 int numpagedout = 0; 477 int i, runlen; 478 479 VM_OBJECT_ASSERT_WLOCKED(object); 480 481 /* 482 * Initiate I/O. Bump the vm_page_t->busy counter and 483 * mark the pages read-only. 484 * 485 * We do not have to fixup the clean/dirty bits here... we can 486 * allow the pager to do it after the I/O completes. 487 * 488 * NOTE! mc[i]->dirty may be partial or fragmented due to an 489 * edge case with file fragments. 490 */ 491 for (i = 0; i < count; i++) { 492 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 493 ("vm_pageout_flush: partially invalid page %p index %d/%d", 494 mc[i], i, count)); 495 vm_page_sbusy(mc[i]); 496 pmap_remove_write(mc[i]); 497 } 498 vm_object_pip_add(object, count); 499 500 vm_pager_put_pages(object, mc, count, flags, pageout_status); 501 502 runlen = count - mreq; 503 if (eio != NULL) 504 *eio = FALSE; 505 for (i = 0; i < count; i++) { 506 vm_page_t mt = mc[i]; 507 508 KASSERT(pageout_status[i] == VM_PAGER_PEND || 509 !pmap_page_is_write_mapped(mt), 510 ("vm_pageout_flush: page %p is not write protected", mt)); 511 switch (pageout_status[i]) { 512 case VM_PAGER_OK: 513 case VM_PAGER_PEND: 514 numpagedout++; 515 break; 516 case VM_PAGER_BAD: 517 /* 518 * Page outside of range of object. Right now we 519 * essentially lose the changes by pretending it 520 * worked. 521 */ 522 vm_page_undirty(mt); 523 break; 524 case VM_PAGER_ERROR: 525 case VM_PAGER_FAIL: 526 /* 527 * If page couldn't be paged out, then reactivate the 528 * page so it doesn't clog the inactive list. (We 529 * will try paging out it again later). 530 */ 531 vm_page_lock(mt); 532 vm_page_activate(mt); 533 vm_page_unlock(mt); 534 if (eio != NULL && i >= mreq && i - mreq < runlen) 535 *eio = TRUE; 536 break; 537 case VM_PAGER_AGAIN: 538 if (i >= mreq && i - mreq < runlen) 539 runlen = i - mreq; 540 break; 541 } 542 543 /* 544 * If the operation is still going, leave the page busy to 545 * block all other accesses. Also, leave the paging in 546 * progress indicator set so that we don't attempt an object 547 * collapse. 548 */ 549 if (pageout_status[i] != VM_PAGER_PEND) { 550 vm_object_pip_wakeup(object); 551 vm_page_sunbusy(mt); 552 if (vm_page_count_severe()) { 553 vm_page_lock(mt); 554 vm_page_try_to_cache(mt); 555 vm_page_unlock(mt); 556 } 557 } 558 } 559 if (prunlen != NULL) 560 *prunlen = runlen; 561 return (numpagedout); 562 } 563 564 static boolean_t 565 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low, 566 vm_paddr_t high) 567 { 568 struct mount *mp; 569 struct vnode *vp; 570 vm_object_t object; 571 vm_paddr_t pa; 572 vm_page_t m, m_tmp, next; 573 574 vm_pagequeue_lock(pq); 575 TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) { 576 if ((m->flags & PG_MARKER) != 0) 577 continue; 578 pa = VM_PAGE_TO_PHYS(m); 579 if (pa < low || pa + PAGE_SIZE > high) 580 continue; 581 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) { 582 vm_page_unlock(m); 583 continue; 584 } 585 object = m->object; 586 if ((!VM_OBJECT_TRYWLOCK(object) && 587 (!vm_pageout_fallback_object_lock(m, &next) || 588 m->hold_count != 0)) || vm_page_busied(m)) { 589 vm_page_unlock(m); 590 VM_OBJECT_WUNLOCK(object); 591 continue; 592 } 593 vm_page_test_dirty(m); 594 if (m->dirty == 0 && object->ref_count != 0) 595 pmap_remove_all(m); 596 if (m->dirty != 0) { 597 vm_page_unlock(m); 598 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) { 599 VM_OBJECT_WUNLOCK(object); 600 continue; 601 } 602 if (object->type == OBJT_VNODE) { 603 vm_pagequeue_unlock(pq); 604 vp = object->handle; 605 vm_object_reference_locked(object); 606 VM_OBJECT_WUNLOCK(object); 607 (void)vn_start_write(vp, &mp, V_WAIT); 608 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 609 VM_OBJECT_WLOCK(object); 610 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 611 VM_OBJECT_WUNLOCK(object); 612 VOP_UNLOCK(vp, 0); 613 vm_object_deallocate(object); 614 vn_finished_write(mp); 615 return (TRUE); 616 } else if (object->type == OBJT_SWAP || 617 object->type == OBJT_DEFAULT) { 618 vm_pagequeue_unlock(pq); 619 m_tmp = m; 620 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC, 621 0, NULL, NULL); 622 VM_OBJECT_WUNLOCK(object); 623 return (TRUE); 624 } 625 } else { 626 /* 627 * Dequeue here to prevent lock recursion in 628 * vm_page_cache(). 629 */ 630 vm_page_dequeue_locked(m); 631 vm_page_cache(m); 632 vm_page_unlock(m); 633 } 634 VM_OBJECT_WUNLOCK(object); 635 } 636 vm_pagequeue_unlock(pq); 637 return (FALSE); 638 } 639 640 /* 641 * Increase the number of cached pages. The specified value, "tries", 642 * determines which categories of pages are cached: 643 * 644 * 0: All clean, inactive pages within the specified physical address range 645 * are cached. Will not sleep. 646 * 1: The vm_lowmem handlers are called. All inactive pages within 647 * the specified physical address range are cached. May sleep. 648 * 2: The vm_lowmem handlers are called. All inactive and active pages 649 * within the specified physical address range are cached. May sleep. 650 */ 651 void 652 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high) 653 { 654 int actl, actmax, inactl, inactmax, dom, initial_dom; 655 static int start_dom = 0; 656 657 if (tries > 0) { 658 /* 659 * Decrease registered cache sizes. The vm_lowmem handlers 660 * may acquire locks and/or sleep, so they can only be invoked 661 * when "tries" is greater than zero. 662 */ 663 EVENTHANDLER_INVOKE(vm_lowmem, 0); 664 665 /* 666 * We do this explicitly after the caches have been drained 667 * above. 668 */ 669 uma_reclaim(); 670 } 671 672 /* 673 * Make the next scan start on the next domain. 674 */ 675 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains; 676 677 inactl = 0; 678 inactmax = cnt.v_inactive_count; 679 actl = 0; 680 actmax = tries < 2 ? 0 : cnt.v_active_count; 681 dom = initial_dom; 682 683 /* 684 * Scan domains in round-robin order, first inactive queues, 685 * then active. Since domain usually owns large physically 686 * contiguous chunk of memory, it makes sense to completely 687 * exhaust one domain before switching to next, while growing 688 * the pool of contiguous physical pages. 689 * 690 * Do not even start launder a domain which cannot contain 691 * the specified address range, as indicated by segments 692 * constituting the domain. 693 */ 694 again: 695 if (inactl < inactmax) { 696 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 697 low, high) && 698 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE], 699 tries, low, high)) { 700 inactl++; 701 goto again; 702 } 703 if (++dom == vm_ndomains) 704 dom = 0; 705 if (dom != initial_dom) 706 goto again; 707 } 708 if (actl < actmax) { 709 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 710 low, high) && 711 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE], 712 tries, low, high)) { 713 actl++; 714 goto again; 715 } 716 if (++dom == vm_ndomains) 717 dom = 0; 718 if (dom != initial_dom) 719 goto again; 720 } 721 } 722 723 #if !defined(NO_SWAPPING) 724 /* 725 * vm_pageout_object_deactivate_pages 726 * 727 * Deactivate enough pages to satisfy the inactive target 728 * requirements. 729 * 730 * The object and map must be locked. 731 */ 732 static void 733 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 734 long desired) 735 { 736 vm_object_t backing_object, object; 737 vm_page_t p; 738 int act_delta, remove_mode; 739 740 VM_OBJECT_ASSERT_LOCKED(first_object); 741 if ((first_object->flags & OBJ_FICTITIOUS) != 0) 742 return; 743 for (object = first_object;; object = backing_object) { 744 if (pmap_resident_count(pmap) <= desired) 745 goto unlock_return; 746 VM_OBJECT_ASSERT_LOCKED(object); 747 if ((object->flags & OBJ_UNMANAGED) != 0 || 748 object->paging_in_progress != 0) 749 goto unlock_return; 750 751 remove_mode = 0; 752 if (object->shadow_count > 1) 753 remove_mode = 1; 754 /* 755 * Scan the object's entire memory queue. 756 */ 757 TAILQ_FOREACH(p, &object->memq, listq) { 758 if (pmap_resident_count(pmap) <= desired) 759 goto unlock_return; 760 if (vm_page_busied(p)) 761 continue; 762 PCPU_INC(cnt.v_pdpages); 763 vm_page_lock(p); 764 if (p->wire_count != 0 || p->hold_count != 0 || 765 !pmap_page_exists_quick(pmap, p)) { 766 vm_page_unlock(p); 767 continue; 768 } 769 act_delta = pmap_ts_referenced(p); 770 if ((p->aflags & PGA_REFERENCED) != 0) { 771 if (act_delta == 0) 772 act_delta = 1; 773 vm_page_aflag_clear(p, PGA_REFERENCED); 774 } 775 if (p->queue != PQ_ACTIVE && act_delta != 0) { 776 vm_page_activate(p); 777 p->act_count += act_delta; 778 } else if (p->queue == PQ_ACTIVE) { 779 if (act_delta == 0) { 780 p->act_count -= min(p->act_count, 781 ACT_DECLINE); 782 if (!remove_mode && p->act_count == 0) { 783 pmap_remove_all(p); 784 vm_page_deactivate(p); 785 } else 786 vm_page_requeue(p); 787 } else { 788 vm_page_activate(p); 789 if (p->act_count < ACT_MAX - 790 ACT_ADVANCE) 791 p->act_count += ACT_ADVANCE; 792 vm_page_requeue(p); 793 } 794 } else if (p->queue == PQ_INACTIVE) 795 pmap_remove_all(p); 796 vm_page_unlock(p); 797 } 798 if ((backing_object = object->backing_object) == NULL) 799 goto unlock_return; 800 VM_OBJECT_RLOCK(backing_object); 801 if (object != first_object) 802 VM_OBJECT_RUNLOCK(object); 803 } 804 unlock_return: 805 if (object != first_object) 806 VM_OBJECT_RUNLOCK(object); 807 } 808 809 /* 810 * deactivate some number of pages in a map, try to do it fairly, but 811 * that is really hard to do. 812 */ 813 static void 814 vm_pageout_map_deactivate_pages(map, desired) 815 vm_map_t map; 816 long desired; 817 { 818 vm_map_entry_t tmpe; 819 vm_object_t obj, bigobj; 820 int nothingwired; 821 822 if (!vm_map_trylock(map)) 823 return; 824 825 bigobj = NULL; 826 nothingwired = TRUE; 827 828 /* 829 * first, search out the biggest object, and try to free pages from 830 * that. 831 */ 832 tmpe = map->header.next; 833 while (tmpe != &map->header) { 834 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 835 obj = tmpe->object.vm_object; 836 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { 837 if (obj->shadow_count <= 1 && 838 (bigobj == NULL || 839 bigobj->resident_page_count < obj->resident_page_count)) { 840 if (bigobj != NULL) 841 VM_OBJECT_RUNLOCK(bigobj); 842 bigobj = obj; 843 } else 844 VM_OBJECT_RUNLOCK(obj); 845 } 846 } 847 if (tmpe->wired_count > 0) 848 nothingwired = FALSE; 849 tmpe = tmpe->next; 850 } 851 852 if (bigobj != NULL) { 853 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 854 VM_OBJECT_RUNLOCK(bigobj); 855 } 856 /* 857 * Next, hunt around for other pages to deactivate. We actually 858 * do this search sort of wrong -- .text first is not the best idea. 859 */ 860 tmpe = map->header.next; 861 while (tmpe != &map->header) { 862 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 863 break; 864 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 865 obj = tmpe->object.vm_object; 866 if (obj != NULL) { 867 VM_OBJECT_RLOCK(obj); 868 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 869 VM_OBJECT_RUNLOCK(obj); 870 } 871 } 872 tmpe = tmpe->next; 873 } 874 875 /* 876 * Remove all mappings if a process is swapped out, this will free page 877 * table pages. 878 */ 879 if (desired == 0 && nothingwired) { 880 pmap_remove(vm_map_pmap(map), vm_map_min(map), 881 vm_map_max(map)); 882 } 883 vm_map_unlock(map); 884 } 885 #endif /* !defined(NO_SWAPPING) */ 886 887 /* 888 * vm_pageout_scan does the dirty work for the pageout daemon. 889 * 890 * pass 0 - Update active LRU/deactivate pages 891 * pass 1 - Move inactive to cache or free 892 * pass 2 - Launder dirty pages 893 */ 894 static void 895 vm_pageout_scan(struct vm_domain *vmd, int pass) 896 { 897 vm_page_t m, next; 898 struct vm_pagequeue *pq; 899 int page_shortage, maxscan, pcount; 900 int addl_page_shortage; 901 vm_object_t object; 902 int act_delta; 903 int vnodes_skipped = 0; 904 int maxlaunder; 905 boolean_t queues_locked; 906 907 /* 908 * If we need to reclaim memory ask kernel caches to return 909 * some. We rate limit to avoid thrashing. 910 */ 911 if (vmd == &vm_dom[0] && pass > 0 && 912 lowmem_ticks + (lowmem_period * hz) < ticks) { 913 /* 914 * Decrease registered cache sizes. 915 */ 916 EVENTHANDLER_INVOKE(vm_lowmem, 0); 917 /* 918 * We do this explicitly after the caches have been 919 * drained above. 920 */ 921 uma_reclaim(); 922 lowmem_ticks = ticks; 923 } 924 925 /* 926 * The addl_page_shortage is the number of temporarily 927 * stuck pages in the inactive queue. In other words, the 928 * number of pages from the inactive count that should be 929 * discounted in setting the target for the active queue scan. 930 */ 931 addl_page_shortage = atomic_readandclear_int(&vm_pageout_deficit); 932 933 /* 934 * Calculate the number of pages we want to either free or move 935 * to the cache. 936 */ 937 page_shortage = vm_paging_target() + addl_page_shortage; 938 939 /* 940 * maxlaunder limits the number of dirty pages we flush per scan. 941 * For most systems a smaller value (16 or 32) is more robust under 942 * extreme memory and disk pressure because any unnecessary writes 943 * to disk can result in extreme performance degredation. However, 944 * systems with excessive dirty pages (especially when MAP_NOSYNC is 945 * used) will die horribly with limited laundering. If the pageout 946 * daemon cannot clean enough pages in the first pass, we let it go 947 * all out in succeeding passes. 948 */ 949 if ((maxlaunder = vm_max_launder) <= 1) 950 maxlaunder = 1; 951 if (pass > 1) 952 maxlaunder = 10000; 953 954 /* 955 * Start scanning the inactive queue for pages we can move to the 956 * cache or free. The scan will stop when the target is reached or 957 * we have scanned the entire inactive queue. Note that m->act_count 958 * is not used to form decisions for the inactive queue, only for the 959 * active queue. 960 */ 961 pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; 962 maxscan = pq->pq_cnt; 963 vm_pagequeue_lock(pq); 964 queues_locked = TRUE; 965 for (m = TAILQ_FIRST(&pq->pq_pl); 966 m != NULL && maxscan-- > 0 && page_shortage > 0; 967 m = next) { 968 vm_pagequeue_assert_locked(pq); 969 KASSERT(queues_locked, ("unlocked queues")); 970 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m)); 971 972 PCPU_INC(cnt.v_pdpages); 973 next = TAILQ_NEXT(m, plinks.q); 974 975 /* 976 * skip marker pages 977 */ 978 if (m->flags & PG_MARKER) 979 continue; 980 981 KASSERT((m->flags & PG_FICTITIOUS) == 0, 982 ("Fictitious page %p cannot be in inactive queue", m)); 983 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 984 ("Unmanaged page %p cannot be in inactive queue", m)); 985 986 /* 987 * The page or object lock acquisitions fail if the 988 * page was removed from the queue or moved to a 989 * different position within the queue. In either 990 * case, addl_page_shortage should not be incremented. 991 */ 992 if (!vm_pageout_page_lock(m, &next)) { 993 vm_page_unlock(m); 994 continue; 995 } 996 object = m->object; 997 if (!VM_OBJECT_TRYWLOCK(object) && 998 !vm_pageout_fallback_object_lock(m, &next)) { 999 vm_page_unlock(m); 1000 VM_OBJECT_WUNLOCK(object); 1001 continue; 1002 } 1003 1004 /* 1005 * Don't mess with busy pages, keep them at at the 1006 * front of the queue, most likely they are being 1007 * paged out. Increment addl_page_shortage for busy 1008 * pages, because they may leave the inactive queue 1009 * shortly after page scan is finished. 1010 */ 1011 if (vm_page_busied(m)) { 1012 vm_page_unlock(m); 1013 VM_OBJECT_WUNLOCK(object); 1014 addl_page_shortage++; 1015 continue; 1016 } 1017 1018 /* 1019 * We unlock the inactive page queue, invalidating the 1020 * 'next' pointer. Use our marker to remember our 1021 * place. 1022 */ 1023 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q); 1024 vm_pagequeue_unlock(pq); 1025 queues_locked = FALSE; 1026 1027 /* 1028 * We bump the activation count if the page has been 1029 * referenced while in the inactive queue. This makes 1030 * it less likely that the page will be added back to the 1031 * inactive queue prematurely again. Here we check the 1032 * page tables (or emulated bits, if any), given the upper 1033 * level VM system not knowing anything about existing 1034 * references. 1035 */ 1036 act_delta = 0; 1037 if ((m->aflags & PGA_REFERENCED) != 0) { 1038 vm_page_aflag_clear(m, PGA_REFERENCED); 1039 act_delta = 1; 1040 } 1041 if (object->ref_count != 0) { 1042 act_delta += pmap_ts_referenced(m); 1043 } else { 1044 KASSERT(!pmap_page_is_mapped(m), 1045 ("vm_pageout_scan: page %p is mapped", m)); 1046 } 1047 1048 /* 1049 * If the upper level VM system knows about any page 1050 * references, we reactivate the page or requeue it. 1051 */ 1052 if (act_delta != 0) { 1053 if (object->ref_count) { 1054 vm_page_activate(m); 1055 m->act_count += act_delta + ACT_ADVANCE; 1056 } else { 1057 vm_pagequeue_lock(pq); 1058 queues_locked = TRUE; 1059 vm_page_requeue_locked(m); 1060 } 1061 VM_OBJECT_WUNLOCK(object); 1062 vm_page_unlock(m); 1063 goto relock_queues; 1064 } 1065 1066 if (m->hold_count != 0) { 1067 vm_page_unlock(m); 1068 VM_OBJECT_WUNLOCK(object); 1069 1070 /* 1071 * Held pages are essentially stuck in the 1072 * queue. So, they ought to be discounted 1073 * from the inactive count. See the 1074 * calculation of the page_shortage for the 1075 * loop over the active queue below. 1076 */ 1077 addl_page_shortage++; 1078 goto relock_queues; 1079 } 1080 1081 /* 1082 * If the page appears to be clean at the machine-independent 1083 * layer, then remove all of its mappings from the pmap in 1084 * anticipation of placing it onto the cache queue. If, 1085 * however, any of the page's mappings allow write access, 1086 * then the page may still be modified until the last of those 1087 * mappings are removed. 1088 */ 1089 vm_page_test_dirty(m); 1090 if (m->dirty == 0 && object->ref_count != 0) 1091 pmap_remove_all(m); 1092 1093 if (m->valid == 0) { 1094 /* 1095 * Invalid pages can be easily freed 1096 */ 1097 vm_page_free(m); 1098 PCPU_INC(cnt.v_dfree); 1099 --page_shortage; 1100 } else if (m->dirty == 0) { 1101 /* 1102 * Clean pages can be placed onto the cache queue. 1103 * This effectively frees them. 1104 */ 1105 vm_page_cache(m); 1106 --page_shortage; 1107 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) { 1108 /* 1109 * Dirty pages need to be paged out, but flushing 1110 * a page is extremely expensive verses freeing 1111 * a clean page. Rather then artificially limiting 1112 * the number of pages we can flush, we instead give 1113 * dirty pages extra priority on the inactive queue 1114 * by forcing them to be cycled through the queue 1115 * twice before being flushed, after which the 1116 * (now clean) page will cycle through once more 1117 * before being freed. This significantly extends 1118 * the thrash point for a heavily loaded machine. 1119 */ 1120 m->flags |= PG_WINATCFLS; 1121 vm_pagequeue_lock(pq); 1122 queues_locked = TRUE; 1123 vm_page_requeue_locked(m); 1124 } else if (maxlaunder > 0) { 1125 /* 1126 * We always want to try to flush some dirty pages if 1127 * we encounter them, to keep the system stable. 1128 * Normally this number is small, but under extreme 1129 * pressure where there are insufficient clean pages 1130 * on the inactive queue, we may have to go all out. 1131 */ 1132 int swap_pageouts_ok; 1133 struct vnode *vp = NULL; 1134 struct mount *mp = NULL; 1135 1136 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 1137 swap_pageouts_ok = 1; 1138 } else { 1139 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 1140 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 1141 vm_page_count_min()); 1142 1143 } 1144 1145 /* 1146 * We don't bother paging objects that are "dead". 1147 * Those objects are in a "rundown" state. 1148 */ 1149 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 1150 vm_pagequeue_lock(pq); 1151 vm_page_unlock(m); 1152 VM_OBJECT_WUNLOCK(object); 1153 queues_locked = TRUE; 1154 vm_page_requeue_locked(m); 1155 goto relock_queues; 1156 } 1157 1158 /* 1159 * The object is already known NOT to be dead. It 1160 * is possible for the vget() to block the whole 1161 * pageout daemon, but the new low-memory handling 1162 * code should prevent it. 1163 * 1164 * The previous code skipped locked vnodes and, worse, 1165 * reordered pages in the queue. This results in 1166 * completely non-deterministic operation and, on a 1167 * busy system, can lead to extremely non-optimal 1168 * pageouts. For example, it can cause clean pages 1169 * to be freed and dirty pages to be moved to the end 1170 * of the queue. Since dirty pages are also moved to 1171 * the end of the queue once-cleaned, this gives 1172 * way too large a weighting to defering the freeing 1173 * of dirty pages. 1174 * 1175 * We can't wait forever for the vnode lock, we might 1176 * deadlock due to a vn_read() getting stuck in 1177 * vm_wait while holding this vnode. We skip the 1178 * vnode if we can't get it in a reasonable amount 1179 * of time. 1180 */ 1181 if (object->type == OBJT_VNODE) { 1182 vm_page_unlock(m); 1183 vp = object->handle; 1184 if (vp->v_type == VREG && 1185 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1186 mp = NULL; 1187 ++pageout_lock_miss; 1188 if (object->flags & OBJ_MIGHTBEDIRTY) 1189 vnodes_skipped++; 1190 goto unlock_and_continue; 1191 } 1192 KASSERT(mp != NULL, 1193 ("vp %p with NULL v_mount", vp)); 1194 vm_object_reference_locked(object); 1195 VM_OBJECT_WUNLOCK(object); 1196 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 1197 curthread)) { 1198 VM_OBJECT_WLOCK(object); 1199 ++pageout_lock_miss; 1200 if (object->flags & OBJ_MIGHTBEDIRTY) 1201 vnodes_skipped++; 1202 vp = NULL; 1203 goto unlock_and_continue; 1204 } 1205 VM_OBJECT_WLOCK(object); 1206 vm_page_lock(m); 1207 vm_pagequeue_lock(pq); 1208 queues_locked = TRUE; 1209 /* 1210 * The page might have been moved to another 1211 * queue during potential blocking in vget() 1212 * above. The page might have been freed and 1213 * reused for another vnode. 1214 */ 1215 if (m->queue != PQ_INACTIVE || 1216 m->object != object || 1217 TAILQ_NEXT(m, plinks.q) != &vmd->vmd_marker) { 1218 vm_page_unlock(m); 1219 if (object->flags & OBJ_MIGHTBEDIRTY) 1220 vnodes_skipped++; 1221 goto unlock_and_continue; 1222 } 1223 1224 /* 1225 * The page may have been busied during the 1226 * blocking in vget(). We don't move the 1227 * page back onto the end of the queue so that 1228 * statistics are more correct if we don't. 1229 */ 1230 if (vm_page_busied(m)) { 1231 vm_page_unlock(m); 1232 goto unlock_and_continue; 1233 } 1234 1235 /* 1236 * If the page has become held it might 1237 * be undergoing I/O, so skip it 1238 */ 1239 if (m->hold_count) { 1240 vm_page_unlock(m); 1241 vm_page_requeue_locked(m); 1242 if (object->flags & OBJ_MIGHTBEDIRTY) 1243 vnodes_skipped++; 1244 goto unlock_and_continue; 1245 } 1246 vm_pagequeue_unlock(pq); 1247 queues_locked = FALSE; 1248 } 1249 1250 /* 1251 * If a page is dirty, then it is either being washed 1252 * (but not yet cleaned) or it is still in the 1253 * laundry. If it is still in the laundry, then we 1254 * start the cleaning operation. 1255 * 1256 * decrement page_shortage on success to account for 1257 * the (future) cleaned page. Otherwise we could wind 1258 * up laundering or cleaning too many pages. 1259 */ 1260 if (vm_pageout_clean(m) != 0) { 1261 --page_shortage; 1262 --maxlaunder; 1263 } 1264 unlock_and_continue: 1265 vm_page_lock_assert(m, MA_NOTOWNED); 1266 VM_OBJECT_WUNLOCK(object); 1267 if (mp != NULL) { 1268 if (queues_locked) { 1269 vm_pagequeue_unlock(pq); 1270 queues_locked = FALSE; 1271 } 1272 if (vp != NULL) 1273 vput(vp); 1274 vm_object_deallocate(object); 1275 vn_finished_write(mp); 1276 } 1277 vm_page_lock_assert(m, MA_NOTOWNED); 1278 goto relock_queues; 1279 } 1280 vm_page_unlock(m); 1281 VM_OBJECT_WUNLOCK(object); 1282 relock_queues: 1283 if (!queues_locked) { 1284 vm_pagequeue_lock(pq); 1285 queues_locked = TRUE; 1286 } 1287 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q); 1288 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q); 1289 } 1290 vm_pagequeue_unlock(pq); 1291 1292 /* 1293 * Compute the number of pages we want to try to move from the 1294 * active queue to the inactive queue. 1295 */ 1296 pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; 1297 vm_pagequeue_lock(pq); 1298 pcount = pq->pq_cnt; 1299 page_shortage = vm_paging_target() + 1300 cnt.v_inactive_target - cnt.v_inactive_count; 1301 page_shortage += addl_page_shortage; 1302 /* 1303 * If we're just idle polling attempt to visit every 1304 * active page within 'update_period' seconds. 1305 */ 1306 if (pass == 0 && vm_pageout_update_period != 0) { 1307 pcount /= vm_pageout_update_period; 1308 page_shortage = pcount; 1309 } 1310 1311 /* 1312 * Scan the active queue for things we can deactivate. We nominally 1313 * track the per-page activity counter and use it to locate 1314 * deactivation candidates. 1315 */ 1316 m = TAILQ_FIRST(&pq->pq_pl); 1317 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1318 1319 KASSERT(m->queue == PQ_ACTIVE, 1320 ("vm_pageout_scan: page %p isn't active", m)); 1321 1322 next = TAILQ_NEXT(m, plinks.q); 1323 if ((m->flags & PG_MARKER) != 0) { 1324 m = next; 1325 continue; 1326 } 1327 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1328 ("Fictitious page %p cannot be in active queue", m)); 1329 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1330 ("Unmanaged page %p cannot be in active queue", m)); 1331 if (!vm_pageout_page_lock(m, &next)) { 1332 vm_page_unlock(m); 1333 m = next; 1334 continue; 1335 } 1336 1337 /* 1338 * The count for pagedaemon pages is done after checking the 1339 * page for eligibility... 1340 */ 1341 PCPU_INC(cnt.v_pdpages); 1342 1343 /* 1344 * Check to see "how much" the page has been used. 1345 */ 1346 act_delta = 0; 1347 if (m->aflags & PGA_REFERENCED) { 1348 vm_page_aflag_clear(m, PGA_REFERENCED); 1349 act_delta += 1; 1350 } 1351 /* 1352 * Unlocked object ref count check. Two races are possible. 1353 * 1) The ref was transitioning to zero and we saw non-zero, 1354 * the pmap bits will be checked unnecessarily. 1355 * 2) The ref was transitioning to one and we saw zero. 1356 * The page lock prevents a new reference to this page so 1357 * we need not check the reference bits. 1358 */ 1359 if (m->object->ref_count != 0) 1360 act_delta += pmap_ts_referenced(m); 1361 1362 /* 1363 * Advance or decay the act_count based on recent usage. 1364 */ 1365 if (act_delta) { 1366 m->act_count += ACT_ADVANCE + act_delta; 1367 if (m->act_count > ACT_MAX) 1368 m->act_count = ACT_MAX; 1369 } else { 1370 m->act_count -= min(m->act_count, ACT_DECLINE); 1371 act_delta = m->act_count; 1372 } 1373 1374 /* 1375 * Move this page to the tail of the active or inactive 1376 * queue depending on usage. 1377 */ 1378 if (act_delta == 0) { 1379 /* Dequeue to avoid later lock recursion. */ 1380 vm_page_dequeue_locked(m); 1381 vm_page_deactivate(m); 1382 page_shortage--; 1383 } else 1384 vm_page_requeue_locked(m); 1385 vm_page_unlock(m); 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 <= 1 || !((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 static void 1567 vm_pageout_worker(void *arg) 1568 { 1569 struct vm_domain *domain; 1570 int domidx; 1571 1572 domidx = (uintptr_t)arg; 1573 domain = &vm_dom[domidx]; 1574 1575 /* 1576 * XXXKIB It could be useful to bind pageout daemon threads to 1577 * the cores belonging to the domain, from which vm_page_array 1578 * is allocated. 1579 */ 1580 1581 KASSERT(domain->vmd_segs != 0, ("domain without segments")); 1582 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE); 1583 1584 /* 1585 * The pageout daemon worker is never done, so loop forever. 1586 */ 1587 while (TRUE) { 1588 /* 1589 * If we have enough free memory, wakeup waiters. Do 1590 * not clear vm_pages_needed until we reach our target, 1591 * otherwise we may be woken up over and over again and 1592 * waste a lot of cpu. 1593 */ 1594 mtx_lock(&vm_page_queue_free_mtx); 1595 if (vm_pages_needed && !vm_page_count_min()) { 1596 if (!vm_paging_needed()) 1597 vm_pages_needed = 0; 1598 wakeup(&cnt.v_free_count); 1599 } 1600 if (vm_pages_needed) { 1601 /* 1602 * Still not done, take a second pass without waiting 1603 * (unlimited dirty cleaning), otherwise sleep a bit 1604 * and try again. 1605 */ 1606 if (domain->vmd_pass > 1) 1607 msleep(&vm_pages_needed, 1608 &vm_page_queue_free_mtx, PVM, "psleep", 1609 hz / 2); 1610 } else { 1611 /* 1612 * Good enough, sleep until required to refresh 1613 * stats. 1614 */ 1615 domain->vmd_pass = 0; 1616 msleep(&vm_pages_needed, &vm_page_queue_free_mtx, 1617 PVM, "psleep", hz); 1618 1619 } 1620 if (vm_pages_needed) { 1621 cnt.v_pdwakeups++; 1622 domain->vmd_pass++; 1623 } 1624 mtx_unlock(&vm_page_queue_free_mtx); 1625 vm_pageout_scan(domain, domain->vmd_pass); 1626 } 1627 } 1628 1629 /* 1630 * vm_pageout is the high level pageout daemon. 1631 */ 1632 static void 1633 vm_pageout(void) 1634 { 1635 #if MAXMEMDOM > 1 1636 int error, i; 1637 #endif 1638 1639 /* 1640 * Initialize some paging parameters. 1641 */ 1642 cnt.v_interrupt_free_min = 2; 1643 if (cnt.v_page_count < 2000) 1644 vm_pageout_page_count = 8; 1645 1646 /* 1647 * v_free_reserved needs to include enough for the largest 1648 * swap pager structures plus enough for any pv_entry structs 1649 * when paging. 1650 */ 1651 if (cnt.v_page_count > 1024) 1652 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1653 else 1654 cnt.v_free_min = 4; 1655 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1656 cnt.v_interrupt_free_min; 1657 cnt.v_free_reserved = vm_pageout_page_count + 1658 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1659 cnt.v_free_severe = cnt.v_free_min / 2; 1660 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1661 cnt.v_free_min += cnt.v_free_reserved; 1662 cnt.v_free_severe += cnt.v_free_reserved; 1663 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1664 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1665 cnt.v_inactive_target = cnt.v_free_count / 3; 1666 1667 /* 1668 * Set the default wakeup threshold to be 10% above the minimum 1669 * page limit. This keeps the steady state out of shortfall. 1670 */ 1671 vm_pageout_wakeup_thresh = (cnt.v_free_min / 10) * 11; 1672 1673 /* 1674 * Set interval in seconds for active scan. We want to visit each 1675 * page at least once every ten minutes. This is to prevent worst 1676 * case paging behaviors with stale active LRU. 1677 */ 1678 if (vm_pageout_update_period == 0) 1679 vm_pageout_update_period = 600; 1680 1681 /* XXX does not really belong here */ 1682 if (vm_page_max_wired == 0) 1683 vm_page_max_wired = cnt.v_free_count / 3; 1684 1685 swap_pager_swap_init(); 1686 #if MAXMEMDOM > 1 1687 for (i = 1; i < vm_ndomains; i++) { 1688 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, 1689 curproc, NULL, 0, 0, "dom%d", i); 1690 if (error != 0) { 1691 panic("starting pageout for domain %d, error %d\n", 1692 i, error); 1693 } 1694 } 1695 #endif 1696 vm_pageout_worker((uintptr_t)0); 1697 } 1698 1699 /* 1700 * Unless the free page queue lock is held by the caller, this function 1701 * should be regarded as advisory. Specifically, the caller should 1702 * not msleep() on &cnt.v_free_count following this function unless 1703 * the free page queue lock is held until the msleep() is performed. 1704 */ 1705 void 1706 pagedaemon_wakeup(void) 1707 { 1708 1709 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1710 vm_pages_needed = 1; 1711 wakeup(&vm_pages_needed); 1712 } 1713 } 1714 1715 #if !defined(NO_SWAPPING) 1716 static void 1717 vm_req_vmdaemon(int req) 1718 { 1719 static int lastrun = 0; 1720 1721 mtx_lock(&vm_daemon_mtx); 1722 vm_pageout_req_swapout |= req; 1723 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1724 wakeup(&vm_daemon_needed); 1725 lastrun = ticks; 1726 } 1727 mtx_unlock(&vm_daemon_mtx); 1728 } 1729 1730 static void 1731 vm_daemon(void) 1732 { 1733 struct rlimit rsslim; 1734 struct proc *p; 1735 struct thread *td; 1736 struct vmspace *vm; 1737 int breakout, swapout_flags, tryagain, attempts; 1738 #ifdef RACCT 1739 uint64_t rsize, ravailable; 1740 #endif 1741 1742 while (TRUE) { 1743 mtx_lock(&vm_daemon_mtx); 1744 #ifdef RACCT 1745 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz); 1746 #else 1747 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1748 #endif 1749 swapout_flags = vm_pageout_req_swapout; 1750 vm_pageout_req_swapout = 0; 1751 mtx_unlock(&vm_daemon_mtx); 1752 if (swapout_flags) 1753 swapout_procs(swapout_flags); 1754 1755 /* 1756 * scan the processes for exceeding their rlimits or if 1757 * process is swapped out -- deactivate pages 1758 */ 1759 tryagain = 0; 1760 attempts = 0; 1761 again: 1762 attempts++; 1763 sx_slock(&allproc_lock); 1764 FOREACH_PROC_IN_SYSTEM(p) { 1765 vm_pindex_t limit, size; 1766 1767 /* 1768 * if this is a system process or if we have already 1769 * looked at this process, skip it. 1770 */ 1771 PROC_LOCK(p); 1772 if (p->p_state != PRS_NORMAL || 1773 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1774 PROC_UNLOCK(p); 1775 continue; 1776 } 1777 /* 1778 * if the process is in a non-running type state, 1779 * don't touch it. 1780 */ 1781 breakout = 0; 1782 FOREACH_THREAD_IN_PROC(p, td) { 1783 thread_lock(td); 1784 if (!TD_ON_RUNQ(td) && 1785 !TD_IS_RUNNING(td) && 1786 !TD_IS_SLEEPING(td) && 1787 !TD_IS_SUSPENDED(td)) { 1788 thread_unlock(td); 1789 breakout = 1; 1790 break; 1791 } 1792 thread_unlock(td); 1793 } 1794 if (breakout) { 1795 PROC_UNLOCK(p); 1796 continue; 1797 } 1798 /* 1799 * get a limit 1800 */ 1801 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1802 limit = OFF_TO_IDX( 1803 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1804 1805 /* 1806 * let processes that are swapped out really be 1807 * swapped out set the limit to nothing (will force a 1808 * swap-out.) 1809 */ 1810 if ((p->p_flag & P_INMEM) == 0) 1811 limit = 0; /* XXX */ 1812 vm = vmspace_acquire_ref(p); 1813 PROC_UNLOCK(p); 1814 if (vm == NULL) 1815 continue; 1816 1817 size = vmspace_resident_count(vm); 1818 if (size >= limit) { 1819 vm_pageout_map_deactivate_pages( 1820 &vm->vm_map, limit); 1821 } 1822 #ifdef RACCT 1823 rsize = IDX_TO_OFF(size); 1824 PROC_LOCK(p); 1825 racct_set(p, RACCT_RSS, rsize); 1826 ravailable = racct_get_available(p, RACCT_RSS); 1827 PROC_UNLOCK(p); 1828 if (rsize > ravailable) { 1829 /* 1830 * Don't be overly aggressive; this might be 1831 * an innocent process, and the limit could've 1832 * been exceeded by some memory hog. Don't 1833 * try to deactivate more than 1/4th of process' 1834 * resident set size. 1835 */ 1836 if (attempts <= 8) { 1837 if (ravailable < rsize - (rsize / 4)) 1838 ravailable = rsize - (rsize / 4); 1839 } 1840 vm_pageout_map_deactivate_pages( 1841 &vm->vm_map, OFF_TO_IDX(ravailable)); 1842 /* Update RSS usage after paging out. */ 1843 size = vmspace_resident_count(vm); 1844 rsize = IDX_TO_OFF(size); 1845 PROC_LOCK(p); 1846 racct_set(p, RACCT_RSS, rsize); 1847 PROC_UNLOCK(p); 1848 if (rsize > ravailable) 1849 tryagain = 1; 1850 } 1851 #endif 1852 vmspace_free(vm); 1853 } 1854 sx_sunlock(&allproc_lock); 1855 if (tryagain != 0 && attempts <= 10) 1856 goto again; 1857 } 1858 } 1859 #endif /* !defined(NO_SWAPPING) */ 1860