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