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 "opt_kdtrace.h" 80 #include <sys/param.h> 81 #include <sys/systm.h> 82 #include <sys/kernel.h> 83 #include <sys/eventhandler.h> 84 #include <sys/lock.h> 85 #include <sys/mutex.h> 86 #include <sys/proc.h> 87 #include <sys/kthread.h> 88 #include <sys/ktr.h> 89 #include <sys/mount.h> 90 #include <sys/racct.h> 91 #include <sys/resourcevar.h> 92 #include <sys/sched.h> 93 #include <sys/sdt.h> 94 #include <sys/signalvar.h> 95 #include <sys/smp.h> 96 #include <sys/vnode.h> 97 #include <sys/vmmeter.h> 98 #include <sys/rwlock.h> 99 #include <sys/sx.h> 100 #include <sys/sysctl.h> 101 102 #include <vm/vm.h> 103 #include <vm/vm_param.h> 104 #include <vm/vm_object.h> 105 #include <vm/vm_page.h> 106 #include <vm/vm_map.h> 107 #include <vm/vm_pageout.h> 108 #include <vm/vm_pager.h> 109 #include <vm/vm_phys.h> 110 #include <vm/swap_pager.h> 111 #include <vm/vm_extern.h> 112 #include <vm/uma.h> 113 114 /* 115 * System initialization 116 */ 117 118 /* the kernel process "vm_pageout"*/ 119 static void vm_pageout(void); 120 static void vm_pageout_init(void); 121 static int vm_pageout_clean(vm_page_t m); 122 static int vm_pageout_cluster(vm_page_t m); 123 static void vm_pageout_scan(struct vm_domain *vmd, int pass); 124 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass); 125 126 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, 127 NULL); 128 129 struct proc *pageproc; 130 131 static struct kproc_desc page_kp = { 132 "pagedaemon", 133 vm_pageout, 134 &pageproc 135 }; 136 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, 137 &page_kp); 138 139 SDT_PROVIDER_DEFINE(vm); 140 SDT_PROBE_DEFINE(vm, , , vm__lowmem_cache); 141 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); 142 143 #if !defined(NO_SWAPPING) 144 /* the kernel process "vm_daemon"*/ 145 static void vm_daemon(void); 146 static struct proc *vmproc; 147 148 static struct kproc_desc vm_kp = { 149 "vmdaemon", 150 vm_daemon, 151 &vmproc 152 }; 153 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 154 #endif 155 156 157 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 158 int vm_pageout_deficit; /* Estimated number of pages deficit */ 159 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 160 int vm_pageout_wakeup_thresh; 161 162 #if !defined(NO_SWAPPING) 163 static int vm_pageout_req_swapout; /* XXX */ 164 static int vm_daemon_needed; 165 static struct mtx vm_daemon_mtx; 166 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 167 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 168 #endif 169 static int vm_max_launder = 32; 170 static int vm_pageout_update_period; 171 static int defer_swap_pageouts; 172 static int disable_swap_pageouts; 173 static int lowmem_period = 10; 174 static int lowmem_ticks; 175 176 #if defined(NO_SWAPPING) 177 static int vm_swap_enabled = 0; 178 static int vm_swap_idle_enabled = 0; 179 #else 180 static int vm_swap_enabled = 1; 181 static int vm_swap_idle_enabled = 0; 182 #endif 183 184 static int vm_panic_on_oom = 0; 185 186 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, 187 CTLFLAG_RWTUN, &vm_panic_on_oom, 0, 188 "panic on out of memory instead of killing the largest process"); 189 190 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh, 191 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0, 192 "free page threshold for waking up the pageout daemon"); 193 194 SYSCTL_INT(_vm, OID_AUTO, max_launder, 195 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 196 197 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, 198 CTLFLAG_RW, &vm_pageout_update_period, 0, 199 "Maximum active LRU update period"); 200 201 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0, 202 "Low memory callback period"); 203 204 #if defined(NO_SWAPPING) 205 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 206 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 207 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 208 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 209 #else 210 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 211 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 212 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 213 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 214 #endif 215 216 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 217 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 218 219 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 220 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 221 222 static int pageout_lock_miss; 223 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 224 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 225 226 #define VM_PAGEOUT_PAGE_COUNT 16 227 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 228 229 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 230 SYSCTL_INT(_vm, OID_AUTO, max_wired, 231 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 232 233 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *); 234 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t, 235 vm_paddr_t); 236 #if !defined(NO_SWAPPING) 237 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 238 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 239 static void vm_req_vmdaemon(int req); 240 #endif 241 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *); 242 243 /* 244 * Initialize a dummy page for marking the caller's place in the specified 245 * paging queue. In principle, this function only needs to set the flag 246 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count 247 * to one as safety precautions. 248 */ 249 static void 250 vm_pageout_init_marker(vm_page_t marker, u_short queue) 251 { 252 253 bzero(marker, sizeof(*marker)); 254 marker->flags = PG_MARKER; 255 marker->busy_lock = VPB_SINGLE_EXCLUSIVER; 256 marker->queue = queue; 257 marker->hold_count = 1; 258 } 259 260 /* 261 * vm_pageout_fallback_object_lock: 262 * 263 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is 264 * known to have failed and page queue must be either PQ_ACTIVE or 265 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 266 * while locking the vm object. Use marker page to detect page queue 267 * changes and maintain notion of next page on page queue. Return 268 * TRUE if no changes were detected, FALSE otherwise. vm object is 269 * locked on return. 270 * 271 * This function depends on both the lock portion of struct vm_object 272 * and normal struct vm_page being type stable. 273 */ 274 static boolean_t 275 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 276 { 277 struct vm_page marker; 278 struct vm_pagequeue *pq; 279 boolean_t unchanged; 280 u_short queue; 281 vm_object_t object; 282 283 queue = m->queue; 284 vm_pageout_init_marker(&marker, queue); 285 pq = vm_page_pagequeue(m); 286 object = m->object; 287 288 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 289 vm_pagequeue_unlock(pq); 290 vm_page_unlock(m); 291 VM_OBJECT_WLOCK(object); 292 vm_page_lock(m); 293 vm_pagequeue_lock(pq); 294 295 /* Page queue might have changed. */ 296 *next = TAILQ_NEXT(&marker, plinks.q); 297 unchanged = (m->queue == queue && 298 m->object == object && 299 &marker == TAILQ_NEXT(m, plinks.q)); 300 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 301 return (unchanged); 302 } 303 304 /* 305 * Lock the page while holding the page queue lock. Use marker page 306 * to detect page queue changes and maintain notion of next page on 307 * page queue. Return TRUE if no changes were detected, FALSE 308 * otherwise. The page is locked on return. The page queue lock might 309 * be dropped and reacquired. 310 * 311 * This function depends on normal struct vm_page being type stable. 312 */ 313 static boolean_t 314 vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 315 { 316 struct vm_page marker; 317 struct vm_pagequeue *pq; 318 boolean_t unchanged; 319 u_short queue; 320 321 vm_page_lock_assert(m, MA_NOTOWNED); 322 if (vm_page_trylock(m)) 323 return (TRUE); 324 325 queue = m->queue; 326 vm_pageout_init_marker(&marker, queue); 327 pq = vm_page_pagequeue(m); 328 329 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 330 vm_pagequeue_unlock(pq); 331 vm_page_lock(m); 332 vm_pagequeue_lock(pq); 333 334 /* Page queue might have changed. */ 335 *next = TAILQ_NEXT(&marker, plinks.q); 336 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, plinks.q)); 337 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 338 return (unchanged); 339 } 340 341 /* 342 * vm_pageout_clean: 343 * 344 * Clean the page and remove it from the laundry. 345 * 346 * We set the busy bit to cause potential page faults on this page to 347 * block. Note the careful timing, however, the busy bit isn't set till 348 * late and we cannot do anything that will mess with the page. 349 */ 350 static int 351 vm_pageout_cluster(vm_page_t m) 352 { 353 vm_object_t object; 354 vm_page_t mc[2*vm_pageout_page_count], pb, ps; 355 int pageout_count; 356 int ib, is, page_base; 357 vm_pindex_t pindex = m->pindex; 358 359 vm_page_lock_assert(m, MA_OWNED); 360 object = m->object; 361 VM_OBJECT_ASSERT_WLOCKED(object); 362 363 /* 364 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 365 * with the new swapper, but we could have serious problems paging 366 * out other object types if there is insufficient memory. 367 * 368 * Unfortunately, checking free memory here is far too late, so the 369 * check has been moved up a procedural level. 370 */ 371 372 /* 373 * Can't clean the page if it's busy or held. 374 */ 375 vm_page_assert_unbusied(m); 376 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m)); 377 vm_page_unlock(m); 378 379 mc[vm_pageout_page_count] = pb = ps = m; 380 pageout_count = 1; 381 page_base = vm_pageout_page_count; 382 ib = 1; 383 is = 1; 384 385 /* 386 * Scan object for clusterable pages. 387 * 388 * We can cluster ONLY if: ->> the page is NOT 389 * clean, wired, busy, held, or mapped into a 390 * buffer, and one of the following: 391 * 1) The page is inactive, or a seldom used 392 * active page. 393 * -or- 394 * 2) we force the issue. 395 * 396 * During heavy mmap/modification loads the pageout 397 * daemon can really fragment the underlying file 398 * due to flushing pages out of order and not trying 399 * align the clusters (which leave sporatic out-of-order 400 * holes). To solve this problem we do the reverse scan 401 * first and attempt to align our cluster, then do a 402 * forward scan if room remains. 403 */ 404 more: 405 while (ib && pageout_count < vm_pageout_page_count) { 406 vm_page_t p; 407 408 if (ib > pindex) { 409 ib = 0; 410 break; 411 } 412 413 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) { 414 ib = 0; 415 break; 416 } 417 vm_page_lock(p); 418 vm_page_test_dirty(p); 419 if (p->dirty == 0 || 420 p->queue != PQ_INACTIVE || 421 p->hold_count != 0) { /* may be undergoing I/O */ 422 vm_page_unlock(p); 423 ib = 0; 424 break; 425 } 426 vm_page_unlock(p); 427 mc[--page_base] = pb = p; 428 ++pageout_count; 429 ++ib; 430 /* 431 * alignment boundry, stop here and switch directions. Do 432 * not clear ib. 433 */ 434 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 435 break; 436 } 437 438 while (pageout_count < vm_pageout_page_count && 439 pindex + is < object->size) { 440 vm_page_t p; 441 442 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p)) 443 break; 444 vm_page_lock(p); 445 vm_page_test_dirty(p); 446 if (p->dirty == 0 || 447 p->queue != PQ_INACTIVE || 448 p->hold_count != 0) { /* may be undergoing I/O */ 449 vm_page_unlock(p); 450 break; 451 } 452 vm_page_unlock(p); 453 mc[page_base + pageout_count] = ps = p; 454 ++pageout_count; 455 ++is; 456 } 457 458 /* 459 * If we exhausted our forward scan, continue with the reverse scan 460 * when possible, even past a page boundry. This catches boundry 461 * conditions. 462 */ 463 if (ib && pageout_count < vm_pageout_page_count) 464 goto more; 465 466 /* 467 * we allow reads during pageouts... 468 */ 469 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL, 470 NULL)); 471 } 472 473 /* 474 * vm_pageout_flush() - launder the given pages 475 * 476 * The given pages are laundered. Note that we setup for the start of 477 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 478 * reference count all in here rather then in the parent. If we want 479 * the parent to do more sophisticated things we may have to change 480 * the ordering. 481 * 482 * Returned runlen is the count of pages between mreq and first 483 * page after mreq with status VM_PAGER_AGAIN. 484 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 485 * for any page in runlen set. 486 */ 487 int 488 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 489 boolean_t *eio) 490 { 491 vm_object_t object = mc[0]->object; 492 int pageout_status[count]; 493 int numpagedout = 0; 494 int i, runlen; 495 496 VM_OBJECT_ASSERT_WLOCKED(object); 497 498 /* 499 * Initiate I/O. Bump the vm_page_t->busy counter and 500 * mark the pages read-only. 501 * 502 * We do not have to fixup the clean/dirty bits here... we can 503 * allow the pager to do it after the I/O completes. 504 * 505 * NOTE! mc[i]->dirty may be partial or fragmented due to an 506 * edge case with file fragments. 507 */ 508 for (i = 0; i < count; i++) { 509 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 510 ("vm_pageout_flush: partially invalid page %p index %d/%d", 511 mc[i], i, count)); 512 vm_page_sbusy(mc[i]); 513 pmap_remove_write(mc[i]); 514 } 515 vm_object_pip_add(object, count); 516 517 vm_pager_put_pages(object, mc, count, flags, pageout_status); 518 519 runlen = count - mreq; 520 if (eio != NULL) 521 *eio = FALSE; 522 for (i = 0; i < count; i++) { 523 vm_page_t mt = mc[i]; 524 525 KASSERT(pageout_status[i] == VM_PAGER_PEND || 526 !pmap_page_is_write_mapped(mt), 527 ("vm_pageout_flush: page %p is not write protected", mt)); 528 switch (pageout_status[i]) { 529 case VM_PAGER_OK: 530 case VM_PAGER_PEND: 531 numpagedout++; 532 break; 533 case VM_PAGER_BAD: 534 /* 535 * Page outside of range of object. Right now we 536 * essentially lose the changes by pretending it 537 * worked. 538 */ 539 vm_page_undirty(mt); 540 break; 541 case VM_PAGER_ERROR: 542 case VM_PAGER_FAIL: 543 /* 544 * If page couldn't be paged out, then reactivate the 545 * page so it doesn't clog the inactive list. (We 546 * will try paging out it again later). 547 */ 548 vm_page_lock(mt); 549 vm_page_activate(mt); 550 vm_page_unlock(mt); 551 if (eio != NULL && i >= mreq && i - mreq < runlen) 552 *eio = TRUE; 553 break; 554 case VM_PAGER_AGAIN: 555 if (i >= mreq && i - mreq < runlen) 556 runlen = i - mreq; 557 break; 558 } 559 560 /* 561 * If the operation is still going, leave the page busy to 562 * block all other accesses. Also, leave the paging in 563 * progress indicator set so that we don't attempt an object 564 * collapse. 565 */ 566 if (pageout_status[i] != VM_PAGER_PEND) { 567 vm_object_pip_wakeup(object); 568 vm_page_sunbusy(mt); 569 } 570 } 571 if (prunlen != NULL) 572 *prunlen = runlen; 573 return (numpagedout); 574 } 575 576 static boolean_t 577 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low, 578 vm_paddr_t high) 579 { 580 struct mount *mp; 581 struct vnode *vp; 582 vm_object_t object; 583 vm_paddr_t pa; 584 vm_page_t m, m_tmp, next; 585 int lockmode; 586 587 vm_pagequeue_lock(pq); 588 TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) { 589 if ((m->flags & PG_MARKER) != 0) 590 continue; 591 pa = VM_PAGE_TO_PHYS(m); 592 if (pa < low || pa + PAGE_SIZE > high) 593 continue; 594 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) { 595 vm_page_unlock(m); 596 continue; 597 } 598 object = m->object; 599 if ((!VM_OBJECT_TRYWLOCK(object) && 600 (!vm_pageout_fallback_object_lock(m, &next) || 601 m->hold_count != 0)) || vm_page_busied(m)) { 602 vm_page_unlock(m); 603 VM_OBJECT_WUNLOCK(object); 604 continue; 605 } 606 vm_page_test_dirty(m); 607 if (m->dirty == 0 && object->ref_count != 0) 608 pmap_remove_all(m); 609 if (m->dirty != 0) { 610 vm_page_unlock(m); 611 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) { 612 VM_OBJECT_WUNLOCK(object); 613 continue; 614 } 615 if (object->type == OBJT_VNODE) { 616 vm_pagequeue_unlock(pq); 617 vp = object->handle; 618 vm_object_reference_locked(object); 619 VM_OBJECT_WUNLOCK(object); 620 (void)vn_start_write(vp, &mp, V_WAIT); 621 lockmode = MNT_SHARED_WRITES(vp->v_mount) ? 622 LK_SHARED : LK_EXCLUSIVE; 623 vn_lock(vp, lockmode | LK_RETRY); 624 VM_OBJECT_WLOCK(object); 625 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 626 VM_OBJECT_WUNLOCK(object); 627 VOP_UNLOCK(vp, 0); 628 vm_object_deallocate(object); 629 vn_finished_write(mp); 630 return (TRUE); 631 } else if (object->type == OBJT_SWAP || 632 object->type == OBJT_DEFAULT) { 633 vm_pagequeue_unlock(pq); 634 m_tmp = m; 635 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC, 636 0, NULL, NULL); 637 VM_OBJECT_WUNLOCK(object); 638 return (TRUE); 639 } 640 } else { 641 /* 642 * Dequeue here to prevent lock recursion in 643 * vm_page_cache(). 644 */ 645 vm_page_dequeue_locked(m); 646 vm_page_cache(m); 647 vm_page_unlock(m); 648 } 649 VM_OBJECT_WUNLOCK(object); 650 } 651 vm_pagequeue_unlock(pq); 652 return (FALSE); 653 } 654 655 /* 656 * Increase the number of cached pages. The specified value, "tries", 657 * determines which categories of pages are cached: 658 * 659 * 0: All clean, inactive pages within the specified physical address range 660 * are cached. Will not sleep. 661 * 1: The vm_lowmem handlers are called. All inactive pages within 662 * the specified physical address range are cached. May sleep. 663 * 2: The vm_lowmem handlers are called. All inactive and active pages 664 * within the specified physical address range are cached. May sleep. 665 */ 666 void 667 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high) 668 { 669 int actl, actmax, inactl, inactmax, dom, initial_dom; 670 static int start_dom = 0; 671 672 if (tries > 0) { 673 /* 674 * Decrease registered cache sizes. The vm_lowmem handlers 675 * may acquire locks and/or sleep, so they can only be invoked 676 * when "tries" is greater than zero. 677 */ 678 SDT_PROBE0(vm, , , vm__lowmem_cache); 679 EVENTHANDLER_INVOKE(vm_lowmem, 0); 680 681 /* 682 * We do this explicitly after the caches have been drained 683 * above. 684 */ 685 uma_reclaim(); 686 } 687 688 /* 689 * Make the next scan start on the next domain. 690 */ 691 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains; 692 693 inactl = 0; 694 inactmax = vm_cnt.v_inactive_count; 695 actl = 0; 696 actmax = tries < 2 ? 0 : vm_cnt.v_active_count; 697 dom = initial_dom; 698 699 /* 700 * Scan domains in round-robin order, first inactive queues, 701 * then active. Since domain usually owns large physically 702 * contiguous chunk of memory, it makes sense to completely 703 * exhaust one domain before switching to next, while growing 704 * the pool of contiguous physical pages. 705 * 706 * Do not even start launder a domain which cannot contain 707 * the specified address range, as indicated by segments 708 * constituting the domain. 709 */ 710 again: 711 if (inactl < inactmax) { 712 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 713 low, high) && 714 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE], 715 tries, low, high)) { 716 inactl++; 717 goto again; 718 } 719 if (++dom == vm_ndomains) 720 dom = 0; 721 if (dom != initial_dom) 722 goto again; 723 } 724 if (actl < actmax) { 725 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 726 low, high) && 727 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE], 728 tries, low, high)) { 729 actl++; 730 goto again; 731 } 732 if (++dom == vm_ndomains) 733 dom = 0; 734 if (dom != initial_dom) 735 goto again; 736 } 737 } 738 739 #if !defined(NO_SWAPPING) 740 /* 741 * vm_pageout_object_deactivate_pages 742 * 743 * Deactivate enough pages to satisfy the inactive target 744 * requirements. 745 * 746 * The object and map must be locked. 747 */ 748 static void 749 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 750 long desired) 751 { 752 vm_object_t backing_object, object; 753 vm_page_t p; 754 int act_delta, remove_mode; 755 756 VM_OBJECT_ASSERT_LOCKED(first_object); 757 if ((first_object->flags & OBJ_FICTITIOUS) != 0) 758 return; 759 for (object = first_object;; object = backing_object) { 760 if (pmap_resident_count(pmap) <= desired) 761 goto unlock_return; 762 VM_OBJECT_ASSERT_LOCKED(object); 763 if ((object->flags & OBJ_UNMANAGED) != 0 || 764 object->paging_in_progress != 0) 765 goto unlock_return; 766 767 remove_mode = 0; 768 if (object->shadow_count > 1) 769 remove_mode = 1; 770 /* 771 * Scan the object's entire memory queue. 772 */ 773 TAILQ_FOREACH(p, &object->memq, listq) { 774 if (pmap_resident_count(pmap) <= desired) 775 goto unlock_return; 776 if (vm_page_busied(p)) 777 continue; 778 PCPU_INC(cnt.v_pdpages); 779 vm_page_lock(p); 780 if (p->wire_count != 0 || p->hold_count != 0 || 781 !pmap_page_exists_quick(pmap, p)) { 782 vm_page_unlock(p); 783 continue; 784 } 785 act_delta = pmap_ts_referenced(p); 786 if ((p->aflags & PGA_REFERENCED) != 0) { 787 if (act_delta == 0) 788 act_delta = 1; 789 vm_page_aflag_clear(p, PGA_REFERENCED); 790 } 791 if (p->queue != PQ_ACTIVE && act_delta != 0) { 792 vm_page_activate(p); 793 p->act_count += act_delta; 794 } else if (p->queue == PQ_ACTIVE) { 795 if (act_delta == 0) { 796 p->act_count -= min(p->act_count, 797 ACT_DECLINE); 798 if (!remove_mode && p->act_count == 0) { 799 pmap_remove_all(p); 800 vm_page_deactivate(p); 801 } else 802 vm_page_requeue(p); 803 } else { 804 vm_page_activate(p); 805 if (p->act_count < ACT_MAX - 806 ACT_ADVANCE) 807 p->act_count += ACT_ADVANCE; 808 vm_page_requeue(p); 809 } 810 } else if (p->queue == PQ_INACTIVE) 811 pmap_remove_all(p); 812 vm_page_unlock(p); 813 } 814 if ((backing_object = object->backing_object) == NULL) 815 goto unlock_return; 816 VM_OBJECT_RLOCK(backing_object); 817 if (object != first_object) 818 VM_OBJECT_RUNLOCK(object); 819 } 820 unlock_return: 821 if (object != first_object) 822 VM_OBJECT_RUNLOCK(object); 823 } 824 825 /* 826 * deactivate some number of pages in a map, try to do it fairly, but 827 * that is really hard to do. 828 */ 829 static void 830 vm_pageout_map_deactivate_pages(map, desired) 831 vm_map_t map; 832 long desired; 833 { 834 vm_map_entry_t tmpe; 835 vm_object_t obj, bigobj; 836 int nothingwired; 837 838 if (!vm_map_trylock(map)) 839 return; 840 841 bigobj = NULL; 842 nothingwired = TRUE; 843 844 /* 845 * first, search out the biggest object, and try to free pages from 846 * that. 847 */ 848 tmpe = map->header.next; 849 while (tmpe != &map->header) { 850 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 851 obj = tmpe->object.vm_object; 852 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { 853 if (obj->shadow_count <= 1 && 854 (bigobj == NULL || 855 bigobj->resident_page_count < obj->resident_page_count)) { 856 if (bigobj != NULL) 857 VM_OBJECT_RUNLOCK(bigobj); 858 bigobj = obj; 859 } else 860 VM_OBJECT_RUNLOCK(obj); 861 } 862 } 863 if (tmpe->wired_count > 0) 864 nothingwired = FALSE; 865 tmpe = tmpe->next; 866 } 867 868 if (bigobj != NULL) { 869 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 870 VM_OBJECT_RUNLOCK(bigobj); 871 } 872 /* 873 * Next, hunt around for other pages to deactivate. We actually 874 * do this search sort of wrong -- .text first is not the best idea. 875 */ 876 tmpe = map->header.next; 877 while (tmpe != &map->header) { 878 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 879 break; 880 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 881 obj = tmpe->object.vm_object; 882 if (obj != NULL) { 883 VM_OBJECT_RLOCK(obj); 884 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 885 VM_OBJECT_RUNLOCK(obj); 886 } 887 } 888 tmpe = tmpe->next; 889 } 890 891 /* 892 * Remove all mappings if a process is swapped out, this will free page 893 * table pages. 894 */ 895 if (desired == 0 && nothingwired) { 896 pmap_remove(vm_map_pmap(map), vm_map_min(map), 897 vm_map_max(map)); 898 } 899 900 vm_map_unlock(map); 901 } 902 #endif /* !defined(NO_SWAPPING) */ 903 904 /* 905 * Attempt to acquire all of the necessary locks to launder a page and 906 * then call through the clustering layer to PUTPAGES. Wait a short 907 * time for a vnode lock. 908 * 909 * Requires the page and object lock on entry, releases both before return. 910 * Returns 0 on success and an errno otherwise. 911 */ 912 static int 913 vm_pageout_clean(vm_page_t m) 914 { 915 struct vnode *vp; 916 struct mount *mp; 917 vm_object_t object; 918 vm_pindex_t pindex; 919 int error, lockmode; 920 921 vm_page_assert_locked(m); 922 object = m->object; 923 VM_OBJECT_ASSERT_WLOCKED(object); 924 error = 0; 925 vp = NULL; 926 mp = NULL; 927 928 /* 929 * The object is already known NOT to be dead. It 930 * is possible for the vget() to block the whole 931 * pageout daemon, but the new low-memory handling 932 * code should prevent it. 933 * 934 * We can't wait forever for the vnode lock, we might 935 * deadlock due to a vn_read() getting stuck in 936 * vm_wait while holding this vnode. We skip the 937 * vnode if we can't get it in a reasonable amount 938 * of time. 939 */ 940 if (object->type == OBJT_VNODE) { 941 vm_page_unlock(m); 942 vp = object->handle; 943 if (vp->v_type == VREG && 944 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 945 mp = NULL; 946 error = EDEADLK; 947 goto unlock_all; 948 } 949 KASSERT(mp != NULL, 950 ("vp %p with NULL v_mount", vp)); 951 vm_object_reference_locked(object); 952 pindex = m->pindex; 953 VM_OBJECT_WUNLOCK(object); 954 lockmode = MNT_SHARED_WRITES(vp->v_mount) ? 955 LK_SHARED : LK_EXCLUSIVE; 956 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) { 957 vp = NULL; 958 error = EDEADLK; 959 goto unlock_mp; 960 } 961 VM_OBJECT_WLOCK(object); 962 vm_page_lock(m); 963 /* 964 * While the object and page were unlocked, the page 965 * may have been: 966 * (1) moved to a different queue, 967 * (2) reallocated to a different object, 968 * (3) reallocated to a different offset, or 969 * (4) cleaned. 970 */ 971 if (m->queue != PQ_INACTIVE || m->object != object || 972 m->pindex != pindex || m->dirty == 0) { 973 vm_page_unlock(m); 974 error = ENXIO; 975 goto unlock_all; 976 } 977 978 /* 979 * The page may have been busied or held while the object 980 * and page locks were released. 981 */ 982 if (vm_page_busied(m) || m->hold_count != 0) { 983 vm_page_unlock(m); 984 error = EBUSY; 985 goto unlock_all; 986 } 987 } 988 989 /* 990 * If a page is dirty, then it is either being washed 991 * (but not yet cleaned) or it is still in the 992 * laundry. If it is still in the laundry, then we 993 * start the cleaning operation. 994 */ 995 if (vm_pageout_cluster(m) == 0) 996 error = EIO; 997 998 unlock_all: 999 VM_OBJECT_WUNLOCK(object); 1000 1001 unlock_mp: 1002 vm_page_lock_assert(m, MA_NOTOWNED); 1003 if (mp != NULL) { 1004 if (vp != NULL) 1005 vput(vp); 1006 vm_object_deallocate(object); 1007 vn_finished_write(mp); 1008 } 1009 1010 return (error); 1011 } 1012 1013 /* 1014 * vm_pageout_scan does the dirty work for the pageout daemon. 1015 * 1016 * pass 0 - Update active LRU/deactivate pages 1017 * pass 1 - Move inactive to cache or free 1018 * pass 2 - Launder dirty pages 1019 */ 1020 static void 1021 vm_pageout_scan(struct vm_domain *vmd, int pass) 1022 { 1023 vm_page_t m, next; 1024 struct vm_pagequeue *pq; 1025 vm_object_t object; 1026 long min_scan; 1027 int act_delta, addl_page_shortage, deficit, maxscan, page_shortage; 1028 int vnodes_skipped = 0; 1029 int maxlaunder, scan_tick, scanned; 1030 boolean_t queues_locked; 1031 1032 /* 1033 * If we need to reclaim memory ask kernel caches to return 1034 * some. We rate limit to avoid thrashing. 1035 */ 1036 if (vmd == &vm_dom[0] && pass > 0 && 1037 (ticks - lowmem_ticks) / hz >= lowmem_period) { 1038 /* 1039 * Decrease registered cache sizes. 1040 */ 1041 SDT_PROBE0(vm, , , vm__lowmem_scan); 1042 EVENTHANDLER_INVOKE(vm_lowmem, 0); 1043 /* 1044 * We do this explicitly after the caches have been 1045 * drained above. 1046 */ 1047 uma_reclaim(); 1048 lowmem_ticks = ticks; 1049 } 1050 1051 /* 1052 * The addl_page_shortage is the number of temporarily 1053 * stuck pages in the inactive queue. In other words, the 1054 * number of pages from the inactive count that should be 1055 * discounted in setting the target for the active queue scan. 1056 */ 1057 addl_page_shortage = 0; 1058 1059 /* 1060 * Calculate the number of pages we want to either free or move 1061 * to the cache. 1062 */ 1063 if (pass > 0) { 1064 deficit = atomic_readandclear_int(&vm_pageout_deficit); 1065 page_shortage = vm_paging_target() + deficit; 1066 } else 1067 page_shortage = deficit = 0; 1068 1069 /* 1070 * maxlaunder limits the number of dirty pages we flush per scan. 1071 * For most systems a smaller value (16 or 32) is more robust under 1072 * extreme memory and disk pressure because any unnecessary writes 1073 * to disk can result in extreme performance degredation. However, 1074 * systems with excessive dirty pages (especially when MAP_NOSYNC is 1075 * used) will die horribly with limited laundering. If the pageout 1076 * daemon cannot clean enough pages in the first pass, we let it go 1077 * all out in succeeding passes. 1078 */ 1079 if ((maxlaunder = vm_max_launder) <= 1) 1080 maxlaunder = 1; 1081 if (pass > 1) 1082 maxlaunder = 10000; 1083 1084 /* 1085 * Start scanning the inactive queue for pages we can move to the 1086 * cache or free. The scan will stop when the target is reached or 1087 * we have scanned the entire inactive queue. Note that m->act_count 1088 * is not used to form decisions for the inactive queue, only for the 1089 * active queue. 1090 */ 1091 pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; 1092 maxscan = pq->pq_cnt; 1093 vm_pagequeue_lock(pq); 1094 queues_locked = TRUE; 1095 for (m = TAILQ_FIRST(&pq->pq_pl); 1096 m != NULL && maxscan-- > 0 && page_shortage > 0; 1097 m = next) { 1098 vm_pagequeue_assert_locked(pq); 1099 KASSERT(queues_locked, ("unlocked queues")); 1100 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m)); 1101 1102 PCPU_INC(cnt.v_pdpages); 1103 next = TAILQ_NEXT(m, plinks.q); 1104 1105 /* 1106 * skip marker pages 1107 */ 1108 if (m->flags & PG_MARKER) 1109 continue; 1110 1111 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1112 ("Fictitious page %p cannot be in inactive queue", m)); 1113 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1114 ("Unmanaged page %p cannot be in inactive queue", m)); 1115 1116 /* 1117 * The page or object lock acquisitions fail if the 1118 * page was removed from the queue or moved to a 1119 * different position within the queue. In either 1120 * case, addl_page_shortage should not be incremented. 1121 */ 1122 if (!vm_pageout_page_lock(m, &next)) { 1123 vm_page_unlock(m); 1124 continue; 1125 } 1126 object = m->object; 1127 if (!VM_OBJECT_TRYWLOCK(object) && 1128 !vm_pageout_fallback_object_lock(m, &next)) { 1129 vm_page_unlock(m); 1130 VM_OBJECT_WUNLOCK(object); 1131 continue; 1132 } 1133 1134 /* 1135 * Don't mess with busy pages, keep them at at the 1136 * front of the queue, most likely they are being 1137 * paged out. Increment addl_page_shortage for busy 1138 * pages, because they may leave the inactive queue 1139 * shortly after page scan is finished. 1140 */ 1141 if (vm_page_busied(m)) { 1142 vm_page_unlock(m); 1143 VM_OBJECT_WUNLOCK(object); 1144 addl_page_shortage++; 1145 continue; 1146 } 1147 1148 /* 1149 * We unlock the inactive page queue, invalidating the 1150 * 'next' pointer. Use our marker to remember our 1151 * place. 1152 */ 1153 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q); 1154 vm_pagequeue_unlock(pq); 1155 queues_locked = FALSE; 1156 1157 /* 1158 * Invalid pages can be easily freed. They cannot be 1159 * mapped, vm_page_free() asserts this. 1160 */ 1161 if (m->valid == 0 && m->hold_count == 0) { 1162 vm_page_free(m); 1163 PCPU_INC(cnt.v_dfree); 1164 --page_shortage; 1165 goto drop_page; 1166 } 1167 1168 /* 1169 * We bump the activation count if the page has been 1170 * referenced while in the inactive queue. This makes 1171 * it less likely that the page will be added back to the 1172 * inactive queue prematurely again. Here we check the 1173 * page tables (or emulated bits, if any), given the upper 1174 * level VM system not knowing anything about existing 1175 * references. 1176 */ 1177 if ((m->aflags & PGA_REFERENCED) != 0) { 1178 vm_page_aflag_clear(m, PGA_REFERENCED); 1179 act_delta = 1; 1180 } else 1181 act_delta = 0; 1182 if (object->ref_count != 0) { 1183 act_delta += pmap_ts_referenced(m); 1184 } else { 1185 KASSERT(!pmap_page_is_mapped(m), 1186 ("vm_pageout_scan: page %p is mapped", m)); 1187 } 1188 1189 /* 1190 * If the upper level VM system knows about any page 1191 * references, we reactivate the page or requeue it. 1192 */ 1193 if (act_delta != 0) { 1194 if (object->ref_count != 0) { 1195 vm_page_activate(m); 1196 m->act_count += act_delta + ACT_ADVANCE; 1197 } else { 1198 vm_pagequeue_lock(pq); 1199 queues_locked = TRUE; 1200 vm_page_requeue_locked(m); 1201 } 1202 goto drop_page; 1203 } 1204 1205 if (m->hold_count != 0) { 1206 /* 1207 * Held pages are essentially stuck in the 1208 * queue. So, they ought to be discounted 1209 * from the inactive count. See the 1210 * calculation of the page_shortage for the 1211 * loop over the active queue below. 1212 */ 1213 addl_page_shortage++; 1214 goto drop_page; 1215 } 1216 1217 /* 1218 * If the page appears to be clean at the machine-independent 1219 * layer, then remove all of its mappings from the pmap in 1220 * anticipation of placing it onto the cache queue. If, 1221 * however, any of the page's mappings allow write access, 1222 * then the page may still be modified until the last of those 1223 * mappings are removed. 1224 */ 1225 if (object->ref_count != 0) { 1226 vm_page_test_dirty(m); 1227 if (m->dirty == 0) 1228 pmap_remove_all(m); 1229 } 1230 1231 if (m->dirty == 0) { 1232 /* 1233 * Clean pages can be freed. 1234 */ 1235 vm_page_free(m); 1236 PCPU_INC(cnt.v_dfree); 1237 --page_shortage; 1238 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) { 1239 /* 1240 * Dirty pages need to be paged out, but flushing 1241 * a page is extremely expensive versus freeing 1242 * a clean page. Rather then artificially limiting 1243 * the number of pages we can flush, we instead give 1244 * dirty pages extra priority on the inactive queue 1245 * by forcing them to be cycled through the queue 1246 * twice before being flushed, after which the 1247 * (now clean) page will cycle through once more 1248 * before being freed. This significantly extends 1249 * the thrash point for a heavily loaded machine. 1250 */ 1251 m->flags |= PG_WINATCFLS; 1252 vm_pagequeue_lock(pq); 1253 queues_locked = TRUE; 1254 vm_page_requeue_locked(m); 1255 } else if (maxlaunder > 0) { 1256 /* 1257 * We always want to try to flush some dirty pages if 1258 * we encounter them, to keep the system stable. 1259 * Normally this number is small, but under extreme 1260 * pressure where there are insufficient clean pages 1261 * on the inactive queue, we may have to go all out. 1262 */ 1263 int swap_pageouts_ok; 1264 int error; 1265 1266 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 1267 swap_pageouts_ok = 1; 1268 } else { 1269 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 1270 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 1271 vm_page_count_min()); 1272 1273 } 1274 1275 /* 1276 * We don't bother paging objects that are "dead". 1277 * Those objects are in a "rundown" state. 1278 */ 1279 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 1280 vm_pagequeue_lock(pq); 1281 vm_page_unlock(m); 1282 VM_OBJECT_WUNLOCK(object); 1283 queues_locked = TRUE; 1284 vm_page_requeue_locked(m); 1285 goto relock_queues; 1286 } 1287 error = vm_pageout_clean(m); 1288 /* 1289 * Decrement page_shortage on success to account for 1290 * the (future) cleaned page. Otherwise we could wind 1291 * up laundering or cleaning too many pages. 1292 */ 1293 if (error == 0) { 1294 page_shortage--; 1295 maxlaunder--; 1296 } else if (error == EDEADLK) { 1297 pageout_lock_miss++; 1298 vnodes_skipped++; 1299 } else if (error == EBUSY) { 1300 addl_page_shortage++; 1301 } 1302 vm_page_lock_assert(m, MA_NOTOWNED); 1303 goto relock_queues; 1304 } 1305 drop_page: 1306 vm_page_unlock(m); 1307 VM_OBJECT_WUNLOCK(object); 1308 relock_queues: 1309 if (!queues_locked) { 1310 vm_pagequeue_lock(pq); 1311 queues_locked = TRUE; 1312 } 1313 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q); 1314 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q); 1315 } 1316 vm_pagequeue_unlock(pq); 1317 1318 #if !defined(NO_SWAPPING) 1319 /* 1320 * Wakeup the swapout daemon if we didn't cache or free the targeted 1321 * number of pages. 1322 */ 1323 if (vm_swap_enabled && page_shortage > 0) 1324 vm_req_vmdaemon(VM_SWAP_NORMAL); 1325 #endif 1326 1327 /* 1328 * Wakeup the sync daemon if we skipped a vnode in a writeable object 1329 * and we didn't cache or free enough pages. 1330 */ 1331 if (vnodes_skipped > 0 && page_shortage > vm_cnt.v_free_target - 1332 vm_cnt.v_free_min) 1333 (void)speedup_syncer(); 1334 1335 /* 1336 * Compute the number of pages we want to try to move from the 1337 * active queue to the inactive queue. 1338 */ 1339 page_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count + 1340 vm_paging_target() + deficit + addl_page_shortage; 1341 1342 pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; 1343 vm_pagequeue_lock(pq); 1344 maxscan = pq->pq_cnt; 1345 1346 /* 1347 * If we're just idle polling attempt to visit every 1348 * active page within 'update_period' seconds. 1349 */ 1350 scan_tick = ticks; 1351 if (vm_pageout_update_period != 0) { 1352 min_scan = pq->pq_cnt; 1353 min_scan *= scan_tick - vmd->vmd_last_active_scan; 1354 min_scan /= hz * vm_pageout_update_period; 1355 } else 1356 min_scan = 0; 1357 if (min_scan > 0 || (page_shortage > 0 && maxscan > 0)) 1358 vmd->vmd_last_active_scan = scan_tick; 1359 1360 /* 1361 * Scan the active queue for pages that can be deactivated. Update 1362 * the per-page activity counter and use it to identify deactivation 1363 * candidates. 1364 */ 1365 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned < 1366 min_scan || (page_shortage > 0 && scanned < maxscan)); m = next, 1367 scanned++) { 1368 1369 KASSERT(m->queue == PQ_ACTIVE, 1370 ("vm_pageout_scan: page %p isn't active", m)); 1371 1372 next = TAILQ_NEXT(m, plinks.q); 1373 if ((m->flags & PG_MARKER) != 0) 1374 continue; 1375 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1376 ("Fictitious page %p cannot be in active queue", m)); 1377 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1378 ("Unmanaged page %p cannot be in active queue", m)); 1379 if (!vm_pageout_page_lock(m, &next)) { 1380 vm_page_unlock(m); 1381 continue; 1382 } 1383 1384 /* 1385 * The count for pagedaemon pages is done after checking the 1386 * page for eligibility... 1387 */ 1388 PCPU_INC(cnt.v_pdpages); 1389 1390 /* 1391 * Check to see "how much" the page has been used. 1392 */ 1393 if ((m->aflags & PGA_REFERENCED) != 0) { 1394 vm_page_aflag_clear(m, PGA_REFERENCED); 1395 act_delta = 1; 1396 } else 1397 act_delta = 0; 1398 1399 /* 1400 * Unlocked object ref count check. Two races are possible. 1401 * 1) The ref was transitioning to zero and we saw non-zero, 1402 * the pmap bits will be checked unnecessarily. 1403 * 2) The ref was transitioning to one and we saw zero. 1404 * The page lock prevents a new reference to this page so 1405 * we need not check the reference bits. 1406 */ 1407 if (m->object->ref_count != 0) 1408 act_delta += pmap_ts_referenced(m); 1409 1410 /* 1411 * Advance or decay the act_count based on recent usage. 1412 */ 1413 if (act_delta != 0) { 1414 m->act_count += ACT_ADVANCE + act_delta; 1415 if (m->act_count > ACT_MAX) 1416 m->act_count = ACT_MAX; 1417 } else 1418 m->act_count -= min(m->act_count, ACT_DECLINE); 1419 1420 /* 1421 * Move this page to the tail of the active or inactive 1422 * queue depending on usage. 1423 */ 1424 if (m->act_count == 0) { 1425 /* Dequeue to avoid later lock recursion. */ 1426 vm_page_dequeue_locked(m); 1427 vm_page_deactivate(m); 1428 page_shortage--; 1429 } else 1430 vm_page_requeue_locked(m); 1431 vm_page_unlock(m); 1432 } 1433 vm_pagequeue_unlock(pq); 1434 #if !defined(NO_SWAPPING) 1435 /* 1436 * Idle process swapout -- run once per second. 1437 */ 1438 if (vm_swap_idle_enabled) { 1439 static long lsec; 1440 if (time_second != lsec) { 1441 vm_req_vmdaemon(VM_SWAP_IDLE); 1442 lsec = time_second; 1443 } 1444 } 1445 #endif 1446 1447 /* 1448 * If we are critically low on one of RAM or swap and low on 1449 * the other, kill the largest process. However, we avoid 1450 * doing this on the first pass in order to give ourselves a 1451 * chance to flush out dirty vnode-backed pages and to allow 1452 * active pages to be moved to the inactive queue and reclaimed. 1453 */ 1454 vm_pageout_mightbe_oom(vmd, pass); 1455 } 1456 1457 static int vm_pageout_oom_vote; 1458 1459 /* 1460 * The pagedaemon threads randlomly select one to perform the 1461 * OOM. Trying to kill processes before all pagedaemons 1462 * failed to reach free target is premature. 1463 */ 1464 static void 1465 vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass) 1466 { 1467 int old_vote; 1468 1469 if (pass <= 1 || !((swap_pager_avail < 64 && vm_page_count_min()) || 1470 (swap_pager_full && vm_paging_target() > 0))) { 1471 if (vmd->vmd_oom) { 1472 vmd->vmd_oom = FALSE; 1473 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1474 } 1475 return; 1476 } 1477 1478 if (vmd->vmd_oom) 1479 return; 1480 1481 vmd->vmd_oom = TRUE; 1482 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); 1483 if (old_vote != vm_ndomains - 1) 1484 return; 1485 1486 /* 1487 * The current pagedaemon thread is the last in the quorum to 1488 * start OOM. Initiate the selection and signaling of the 1489 * victim. 1490 */ 1491 vm_pageout_oom(VM_OOM_MEM); 1492 1493 /* 1494 * After one round of OOM terror, recall our vote. On the 1495 * next pass, current pagedaemon would vote again if the low 1496 * memory condition is still there, due to vmd_oom being 1497 * false. 1498 */ 1499 vmd->vmd_oom = FALSE; 1500 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1501 } 1502 1503 void 1504 vm_pageout_oom(int shortage) 1505 { 1506 struct proc *p, *bigproc; 1507 vm_offset_t size, bigsize; 1508 struct thread *td; 1509 struct vmspace *vm; 1510 1511 /* 1512 * We keep the process bigproc locked once we find it to keep anyone 1513 * from messing with it; however, there is a possibility of 1514 * deadlock if process B is bigproc and one of it's child processes 1515 * attempts to propagate a signal to B while we are waiting for A's 1516 * lock while walking this list. To avoid this, we don't block on 1517 * the process lock but just skip a process if it is already locked. 1518 */ 1519 bigproc = NULL; 1520 bigsize = 0; 1521 sx_slock(&allproc_lock); 1522 FOREACH_PROC_IN_SYSTEM(p) { 1523 int breakout; 1524 1525 PROC_LOCK(p); 1526 1527 /* 1528 * If this is a system, protected or killed process, skip it. 1529 */ 1530 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | 1531 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || 1532 p->p_pid == 1 || P_KILLED(p) || 1533 (p->p_pid < 48 && swap_pager_avail != 0)) { 1534 PROC_UNLOCK(p); 1535 continue; 1536 } 1537 /* 1538 * If the process is in a non-running type state, 1539 * don't touch it. Check all the threads individually. 1540 */ 1541 breakout = 0; 1542 FOREACH_THREAD_IN_PROC(p, td) { 1543 thread_lock(td); 1544 if (!TD_ON_RUNQ(td) && 1545 !TD_IS_RUNNING(td) && 1546 !TD_IS_SLEEPING(td) && 1547 !TD_IS_SUSPENDED(td)) { 1548 thread_unlock(td); 1549 breakout = 1; 1550 break; 1551 } 1552 thread_unlock(td); 1553 } 1554 if (breakout) { 1555 PROC_UNLOCK(p); 1556 continue; 1557 } 1558 /* 1559 * get the process size 1560 */ 1561 vm = vmspace_acquire_ref(p); 1562 if (vm == NULL) { 1563 PROC_UNLOCK(p); 1564 continue; 1565 } 1566 _PHOLD(p); 1567 if (!vm_map_trylock_read(&vm->vm_map)) { 1568 _PRELE(p); 1569 PROC_UNLOCK(p); 1570 vmspace_free(vm); 1571 continue; 1572 } 1573 PROC_UNLOCK(p); 1574 size = vmspace_swap_count(vm); 1575 vm_map_unlock_read(&vm->vm_map); 1576 if (shortage == VM_OOM_MEM) 1577 size += vmspace_resident_count(vm); 1578 vmspace_free(vm); 1579 /* 1580 * if the this process is bigger than the biggest one 1581 * remember it. 1582 */ 1583 if (size > bigsize) { 1584 if (bigproc != NULL) 1585 PRELE(bigproc); 1586 bigproc = p; 1587 bigsize = size; 1588 } else { 1589 PRELE(p); 1590 } 1591 } 1592 sx_sunlock(&allproc_lock); 1593 if (bigproc != NULL) { 1594 if (vm_panic_on_oom != 0) 1595 panic("out of swap space"); 1596 PROC_LOCK(bigproc); 1597 killproc(bigproc, "out of swap space"); 1598 sched_nice(bigproc, PRIO_MIN); 1599 _PRELE(bigproc); 1600 PROC_UNLOCK(bigproc); 1601 wakeup(&vm_cnt.v_free_count); 1602 } 1603 } 1604 1605 static void 1606 vm_pageout_worker(void *arg) 1607 { 1608 struct vm_domain *domain; 1609 int domidx; 1610 1611 domidx = (uintptr_t)arg; 1612 domain = &vm_dom[domidx]; 1613 1614 /* 1615 * XXXKIB It could be useful to bind pageout daemon threads to 1616 * the cores belonging to the domain, from which vm_page_array 1617 * is allocated. 1618 */ 1619 1620 KASSERT(domain->vmd_segs != 0, ("domain without segments")); 1621 domain->vmd_last_active_scan = ticks; 1622 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE); 1623 1624 /* 1625 * The pageout daemon worker is never done, so loop forever. 1626 */ 1627 while (TRUE) { 1628 /* 1629 * If we have enough free memory, wakeup waiters. Do 1630 * not clear vm_pages_needed until we reach our target, 1631 * otherwise we may be woken up over and over again and 1632 * waste a lot of cpu. 1633 */ 1634 mtx_lock(&vm_page_queue_free_mtx); 1635 if (vm_pages_needed && !vm_page_count_min()) { 1636 if (!vm_paging_needed()) 1637 vm_pages_needed = 0; 1638 wakeup(&vm_cnt.v_free_count); 1639 } 1640 if (vm_pages_needed) { 1641 /* 1642 * Still not done, take a second pass without waiting 1643 * (unlimited dirty cleaning), otherwise sleep a bit 1644 * and try again. 1645 */ 1646 if (domain->vmd_pass > 1) 1647 msleep(&vm_pages_needed, 1648 &vm_page_queue_free_mtx, PVM, "psleep", 1649 hz / 2); 1650 } else { 1651 /* 1652 * Good enough, sleep until required to refresh 1653 * stats. 1654 */ 1655 domain->vmd_pass = 0; 1656 msleep(&vm_pages_needed, &vm_page_queue_free_mtx, 1657 PVM, "psleep", hz); 1658 1659 } 1660 if (vm_pages_needed) { 1661 vm_cnt.v_pdwakeups++; 1662 domain->vmd_pass++; 1663 } 1664 mtx_unlock(&vm_page_queue_free_mtx); 1665 vm_pageout_scan(domain, domain->vmd_pass); 1666 } 1667 } 1668 1669 /* 1670 * vm_pageout_init initialises basic pageout daemon settings. 1671 */ 1672 static void 1673 vm_pageout_init(void) 1674 { 1675 /* 1676 * Initialize some paging parameters. 1677 */ 1678 vm_cnt.v_interrupt_free_min = 2; 1679 if (vm_cnt.v_page_count < 2000) 1680 vm_pageout_page_count = 8; 1681 1682 /* 1683 * v_free_reserved needs to include enough for the largest 1684 * swap pager structures plus enough for any pv_entry structs 1685 * when paging. 1686 */ 1687 if (vm_cnt.v_page_count > 1024) 1688 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200; 1689 else 1690 vm_cnt.v_free_min = 4; 1691 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1692 vm_cnt.v_interrupt_free_min; 1693 vm_cnt.v_free_reserved = vm_pageout_page_count + 1694 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768); 1695 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2; 1696 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved; 1697 vm_cnt.v_free_min += vm_cnt.v_free_reserved; 1698 vm_cnt.v_free_severe += vm_cnt.v_free_reserved; 1699 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2; 1700 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3) 1701 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3; 1702 1703 /* 1704 * Set the default wakeup threshold to be 10% above the minimum 1705 * page limit. This keeps the steady state out of shortfall. 1706 */ 1707 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11; 1708 1709 /* 1710 * Set interval in seconds for active scan. We want to visit each 1711 * page at least once every ten minutes. This is to prevent worst 1712 * case paging behaviors with stale active LRU. 1713 */ 1714 if (vm_pageout_update_period == 0) 1715 vm_pageout_update_period = 600; 1716 1717 /* XXX does not really belong here */ 1718 if (vm_page_max_wired == 0) 1719 vm_page_max_wired = vm_cnt.v_free_count / 3; 1720 } 1721 1722 /* 1723 * vm_pageout is the high level pageout daemon. 1724 */ 1725 static void 1726 vm_pageout(void) 1727 { 1728 int error; 1729 #if MAXMEMDOM > 1 1730 int i; 1731 #endif 1732 1733 swap_pager_swap_init(); 1734 #if MAXMEMDOM > 1 1735 for (i = 1; i < vm_ndomains; i++) { 1736 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, 1737 curproc, NULL, 0, 0, "dom%d", i); 1738 if (error != 0) { 1739 panic("starting pageout for domain %d, error %d\n", 1740 i, error); 1741 } 1742 } 1743 #endif 1744 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL, 1745 0, 0, "uma"); 1746 if (error != 0) 1747 panic("starting uma_reclaim helper, error %d\n", error); 1748 vm_pageout_worker((void *)(uintptr_t)0); 1749 } 1750 1751 /* 1752 * Unless the free page queue lock is held by the caller, this function 1753 * should be regarded as advisory. Specifically, the caller should 1754 * not msleep() on &vm_cnt.v_free_count following this function unless 1755 * the free page queue lock is held until the msleep() is performed. 1756 */ 1757 void 1758 pagedaemon_wakeup(void) 1759 { 1760 1761 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1762 vm_pages_needed = 1; 1763 wakeup(&vm_pages_needed); 1764 } 1765 } 1766 1767 #if !defined(NO_SWAPPING) 1768 static void 1769 vm_req_vmdaemon(int req) 1770 { 1771 static int lastrun = 0; 1772 1773 mtx_lock(&vm_daemon_mtx); 1774 vm_pageout_req_swapout |= req; 1775 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1776 wakeup(&vm_daemon_needed); 1777 lastrun = ticks; 1778 } 1779 mtx_unlock(&vm_daemon_mtx); 1780 } 1781 1782 static void 1783 vm_daemon(void) 1784 { 1785 struct rlimit rsslim; 1786 struct proc *p; 1787 struct thread *td; 1788 struct vmspace *vm; 1789 int breakout, swapout_flags, tryagain, attempts; 1790 #ifdef RACCT 1791 uint64_t rsize, ravailable; 1792 #endif 1793 1794 while (TRUE) { 1795 mtx_lock(&vm_daemon_mtx); 1796 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 1797 #ifdef RACCT 1798 racct_enable ? hz : 0 1799 #else 1800 0 1801 #endif 1802 ); 1803 swapout_flags = vm_pageout_req_swapout; 1804 vm_pageout_req_swapout = 0; 1805 mtx_unlock(&vm_daemon_mtx); 1806 if (swapout_flags) 1807 swapout_procs(swapout_flags); 1808 1809 /* 1810 * scan the processes for exceeding their rlimits or if 1811 * process is swapped out -- deactivate pages 1812 */ 1813 tryagain = 0; 1814 attempts = 0; 1815 again: 1816 attempts++; 1817 sx_slock(&allproc_lock); 1818 FOREACH_PROC_IN_SYSTEM(p) { 1819 vm_pindex_t limit, size; 1820 1821 /* 1822 * if this is a system process or if we have already 1823 * looked at this process, skip it. 1824 */ 1825 PROC_LOCK(p); 1826 if (p->p_state != PRS_NORMAL || 1827 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1828 PROC_UNLOCK(p); 1829 continue; 1830 } 1831 /* 1832 * if the process is in a non-running type state, 1833 * don't touch it. 1834 */ 1835 breakout = 0; 1836 FOREACH_THREAD_IN_PROC(p, td) { 1837 thread_lock(td); 1838 if (!TD_ON_RUNQ(td) && 1839 !TD_IS_RUNNING(td) && 1840 !TD_IS_SLEEPING(td) && 1841 !TD_IS_SUSPENDED(td)) { 1842 thread_unlock(td); 1843 breakout = 1; 1844 break; 1845 } 1846 thread_unlock(td); 1847 } 1848 if (breakout) { 1849 PROC_UNLOCK(p); 1850 continue; 1851 } 1852 /* 1853 * get a limit 1854 */ 1855 lim_rlimit_proc(p, RLIMIT_RSS, &rsslim); 1856 limit = OFF_TO_IDX( 1857 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1858 1859 /* 1860 * let processes that are swapped out really be 1861 * swapped out set the limit to nothing (will force a 1862 * swap-out.) 1863 */ 1864 if ((p->p_flag & P_INMEM) == 0) 1865 limit = 0; /* XXX */ 1866 vm = vmspace_acquire_ref(p); 1867 PROC_UNLOCK(p); 1868 if (vm == NULL) 1869 continue; 1870 1871 size = vmspace_resident_count(vm); 1872 if (size >= limit) { 1873 vm_pageout_map_deactivate_pages( 1874 &vm->vm_map, limit); 1875 } 1876 #ifdef RACCT 1877 if (racct_enable) { 1878 rsize = IDX_TO_OFF(size); 1879 PROC_LOCK(p); 1880 racct_set(p, RACCT_RSS, rsize); 1881 ravailable = racct_get_available(p, RACCT_RSS); 1882 PROC_UNLOCK(p); 1883 if (rsize > ravailable) { 1884 /* 1885 * Don't be overly aggressive; this 1886 * might be an innocent process, 1887 * and the limit could've been exceeded 1888 * by some memory hog. Don't try 1889 * to deactivate more than 1/4th 1890 * of process' resident set size. 1891 */ 1892 if (attempts <= 8) { 1893 if (ravailable < rsize - 1894 (rsize / 4)) { 1895 ravailable = rsize - 1896 (rsize / 4); 1897 } 1898 } 1899 vm_pageout_map_deactivate_pages( 1900 &vm->vm_map, 1901 OFF_TO_IDX(ravailable)); 1902 /* Update RSS usage after paging out. */ 1903 size = vmspace_resident_count(vm); 1904 rsize = IDX_TO_OFF(size); 1905 PROC_LOCK(p); 1906 racct_set(p, RACCT_RSS, rsize); 1907 PROC_UNLOCK(p); 1908 if (rsize > ravailable) 1909 tryagain = 1; 1910 } 1911 } 1912 #endif 1913 vmspace_free(vm); 1914 } 1915 sx_sunlock(&allproc_lock); 1916 if (tryagain != 0 && attempts <= 10) 1917 goto again; 1918 } 1919 } 1920 #endif /* !defined(NO_SWAPPING) */ 1921