vm_pageout.c revision 216899
1284677Sdim/*- 2284677Sdim * Copyright (c) 1991 Regents of the University of California. 3284677Sdim * All rights reserved. 4284677Sdim * Copyright (c) 1994 John S. Dyson 5284677Sdim * All rights reserved. 6284677Sdim * Copyright (c) 1994 David Greenman 7284677Sdim * All rights reserved. 8284677Sdim * Copyright (c) 2005 Yahoo! Technologies Norway AS 9284677Sdim * All rights reserved. 10284677Sdim * 11341825Sdim * This code is derived from software contributed to Berkeley by 12284677Sdim * The Mach Operating System project at Carnegie-Mellon University. 13284677Sdim * 14284677Sdim * Redistribution and use in source and binary forms, with or without 15284677Sdim * modification, are permitted provided that the following conditions 16284677Sdim * are met: 17284677Sdim * 1. Redistributions of source code must retain the above copyright 18284677Sdim * notice, this list of conditions and the following disclaimer. 19314564Sdim * 2. Redistributions in binary form must reproduce the above copyright 20284677Sdim * notice, this list of conditions and the following disclaimer in the 21284677Sdim * documentation and/or other materials provided with the distribution. 22341825Sdim * 3. All advertising materials mentioning features or use of this software 23314564Sdim * must display the following acknowledgement: 24314564Sdim * This product includes software developed by the University of 25314564Sdim * California, Berkeley and its contributors. 26314564Sdim * 4. Neither the name of the University nor the names of its contributors 27314564Sdim * may be used to endorse or promote products derived from this software 28314564Sdim * without specific prior written permission. 29284677Sdim * 30314564Sdim * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31314564Sdim * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32314564Sdim * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33314564Sdim * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34314564Sdim * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35284677Sdim * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36314564Sdim * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37284677Sdim * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38284677Sdim * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39314564Sdim * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40314564Sdim * SUCH DAMAGE. 41314564Sdim * 42314564Sdim * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 43314564Sdim * 44341825Sdim * 45314564Sdim * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46314564Sdim * All rights reserved. 47314564Sdim * 48314564Sdim * Authors: Avadis Tevanian, Jr., Michael Wayne Young 49314564Sdim * 50284677Sdim * Permission to use, copy, modify and distribute this software and 51284677Sdim * its documentation is hereby granted, provided that both the copyright 52284677Sdim * notice and this permission notice appear in all copies of the 53341825Sdim * software, derivative works or modified versions, and any portions 54341825Sdim * thereof, and that both notices appear in supporting documentation. 55309124Sdim * 56309124Sdim * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 57341825Sdim * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 58341825Sdim * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 59341825Sdim * 60341825Sdim * Carnegie Mellon requests users of this software to return to 61341825Sdim * 62341825Sdim * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 63341825Sdim * School of Computer Science 64284677Sdim * Carnegie Mellon University 65341825Sdim * Pittsburgh PA 15213-3890 66284677Sdim * 67309124Sdim * any improvements or extensions that they make and grant Carnegie the 68309124Sdim * rights to redistribute these changes. 69309124Sdim */ 70309124Sdim 71309124Sdim/* 72309124Sdim * The proverbial page-out daemon. 73309124Sdim */ 74309124Sdim 75309124Sdim#include <sys/cdefs.h> 76309124Sdim__FBSDID("$FreeBSD: head/sys/vm/vm_pageout.c 216899 2011-01-03 00:41:56Z alc $"); 77309124Sdim 78309124Sdim#include "opt_vm.h" 79309124Sdim#include <sys/param.h> 80309124Sdim#include <sys/systm.h> 81309124Sdim#include <sys/kernel.h> 82309124Sdim#include <sys/eventhandler.h> 83309124Sdim#include <sys/lock.h> 84309124Sdim#include <sys/mutex.h> 85309124Sdim#include <sys/proc.h> 86309124Sdim#include <sys/kthread.h> 87309124Sdim#include <sys/ktr.h> 88309124Sdim#include <sys/mount.h> 89309124Sdim#include <sys/resourcevar.h> 90309124Sdim#include <sys/sched.h> 91309124Sdim#include <sys/signalvar.h> 92309124Sdim#include <sys/vnode.h> 93309124Sdim#include <sys/vmmeter.h> 94309124Sdim#include <sys/sx.h> 95309124Sdim#include <sys/sysctl.h> 96309124Sdim 97309124Sdim#include <vm/vm.h> 98309124Sdim#include <vm/vm_param.h> 99309124Sdim#include <vm/vm_object.h> 100309124Sdim#include <vm/vm_page.h> 101309124Sdim#include <vm/vm_map.h> 102309124Sdim#include <vm/vm_pageout.h> 103309124Sdim#include <vm/vm_pager.h> 104314564Sdim#include <vm/swap_pager.h> 105314564Sdim#include <vm/vm_extern.h> 106314564Sdim#include <vm/uma.h> 107314564Sdim 108314564Sdim/* 109314564Sdim * System initialization 110314564Sdim */ 111314564Sdim 112314564Sdim/* the kernel process "vm_pageout"*/ 113314564Sdimstatic void vm_pageout(void); 114314564Sdimstatic int vm_pageout_clean(vm_page_t); 115314564Sdimstatic void vm_pageout_scan(int pass); 116309124Sdim 117309124Sdimstruct proc *pageproc; 118309124Sdim 119309124Sdimstatic struct kproc_desc page_kp = { 120309124Sdim "pagedaemon", 121284677Sdim vm_pageout, 122284677Sdim &pageproc 123284677Sdim}; 124284677SdimSYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, 125284677Sdim &page_kp); 126284677Sdim 127284677Sdim#if !defined(NO_SWAPPING) 128284677Sdim/* the kernel process "vm_daemon"*/ 129284677Sdimstatic void vm_daemon(void); 130284677Sdimstatic struct proc *vmproc; 131284677Sdim 132284677Sdimstatic struct kproc_desc vm_kp = { 133284677Sdim "vmdaemon", 134284677Sdim vm_daemon, 135284677Sdim &vmproc 136284677Sdim}; 137284677SdimSYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 138284677Sdim#endif 139284677Sdim 140284677Sdim 141284677Sdimint vm_pages_needed; /* Event on which pageout daemon sleeps */ 142284677Sdimint vm_pageout_deficit; /* Estimated number of pages deficit */ 143284677Sdimint vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 144284677Sdim 145284677Sdim#if !defined(NO_SWAPPING) 146284677Sdimstatic int vm_pageout_req_swapout; /* XXX */ 147284677Sdimstatic int vm_daemon_needed; 148284677Sdimstatic struct mtx vm_daemon_mtx; 149284677Sdim/* Allow for use by vm_pageout before vm_daemon is initialized. */ 150284677SdimMTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 151284677Sdim#endif 152341825Sdimstatic int vm_max_launder = 32; 153341825Sdimstatic int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 154341825Sdimstatic int vm_pageout_full_stats_interval = 0; 155341825Sdimstatic int vm_pageout_algorithm=0; 156341825Sdimstatic int defer_swap_pageouts=0; 157284677Sdimstatic int disable_swap_pageouts=0; 158284677Sdim 159284677Sdim#if defined(NO_SWAPPING) 160284677Sdimstatic int vm_swap_enabled=0; 161284677Sdimstatic int vm_swap_idle_enabled=0; 162284677Sdim#else 163309124Sdimstatic int vm_swap_enabled=1; 164284677Sdimstatic int vm_swap_idle_enabled=0; 165284677Sdim#endif 166284677Sdim 167284677SdimSYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 168284677Sdim CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 169284677Sdim 170284677SdimSYSCTL_INT(_vm, OID_AUTO, max_launder, 171284677Sdim CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 172284677Sdim 173284677SdimSYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 174284677Sdim CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 175284677Sdim 176284677SdimSYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 177284677Sdim CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 178284677Sdim 179284677SdimSYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 180284677Sdim CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 181284677Sdim 182284677Sdim#if defined(NO_SWAPPING) 183284677SdimSYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 184284677Sdim CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 185284677SdimSYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 186284677Sdim CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 187284677Sdim#else 188284677SdimSYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 189284677Sdim CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 190284677SdimSYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 191284677Sdim CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 192284677Sdim#endif 193284677Sdim 194284677SdimSYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 195284677Sdim CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 196284677Sdim 197284677SdimSYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 198284677Sdim CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 199284677Sdim 200284677Sdimstatic int pageout_lock_miss; 201284677SdimSYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 202284677Sdim CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 203284677Sdim 204284677Sdim#define VM_PAGEOUT_PAGE_COUNT 16 205284677Sdimint vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 206284677Sdim 207284677Sdimint vm_page_max_wired; /* XXX max # of wired pages system-wide */ 208284677SdimSYSCTL_INT(_vm, OID_AUTO, max_wired, 209284677Sdim CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 210284677Sdim 211284677Sdim#if !defined(NO_SWAPPING) 212284677Sdimstatic void vm_pageout_map_deactivate_pages(vm_map_t, long); 213284677Sdimstatic void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 214284677Sdimstatic void vm_req_vmdaemon(int req); 215284677Sdim#endif 216284677Sdimstatic void vm_pageout_page_stats(void); 217284677Sdim 218284677Sdimstatic void 219284677Sdimvm_pageout_init_marker(vm_page_t marker, u_short queue) 220284677Sdim{ 221327952Sdim 222327952Sdim bzero(marker, sizeof(*marker)); 223327952Sdim marker->flags = PG_FICTITIOUS | PG_MARKER; 224327952Sdim marker->oflags = VPO_BUSY; 225327952Sdim marker->queue = queue; 226341825Sdim marker->wire_count = 1; 227341825Sdim} 228341825Sdim 229341825Sdim/* 230341825Sdim * vm_pageout_fallback_object_lock: 231341825Sdim * 232341825Sdim * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 233341825Sdim * known to have failed and page queue must be either PQ_ACTIVE or 234341825Sdim * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 235341825Sdim * while locking the vm object. Use marker page to detect page queue 236341825Sdim * changes and maintain notion of next page on page queue. Return 237341825Sdim * TRUE if no changes were detected, FALSE otherwise. vm object is 238341825Sdim * locked on return. 239341825Sdim * 240341825Sdim * This function depends on both the lock portion of struct vm_object 241341825Sdim * and normal struct vm_page being type stable. 242341825Sdim */ 243341825Sdimboolean_t 244341825Sdimvm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 245341825Sdim{ 246341825Sdim struct vm_page marker; 247341825Sdim boolean_t unchanged; 248341825Sdim u_short queue; 249341825Sdim vm_object_t object; 250341825Sdim 251341825Sdim queue = m->queue; 252341825Sdim vm_pageout_init_marker(&marker, queue); 253341825Sdim object = m->object; 254284677Sdim 255284677Sdim TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 256284677Sdim m, &marker, pageq); 257284677Sdim vm_page_unlock_queues(); 258284677Sdim vm_page_unlock(m); 259284677Sdim VM_OBJECT_LOCK(object); 260284677Sdim vm_page_lock(m); 261284677Sdim vm_page_lock_queues(); 262284677Sdim 263284677Sdim /* Page queue might have changed. */ 264321369Sdim *next = TAILQ_NEXT(&marker, pageq); 265321369Sdim unchanged = (m->queue == queue && 266321369Sdim m->object == object && 267321369Sdim &marker == TAILQ_NEXT(m, pageq)); 268321369Sdim TAILQ_REMOVE(&vm_page_queues[queue].pl, 269321369Sdim &marker, pageq); 270321369Sdim return (unchanged); 271321369Sdim} 272321369Sdim 273284677Sdim/* 274309124Sdim * Lock the page while holding the page queue lock. Use marker page 275309124Sdim * to detect page queue changes and maintain notion of next page on 276309124Sdim * page queue. Return TRUE if no changes were detected, FALSE 277309124Sdim * otherwise. The page is locked on return. The page queue lock might 278309124Sdim * be dropped and reacquired. 279309124Sdim * 280314564Sdim * This function depends on normal struct vm_page being type stable. 281284677Sdim */ 282284677Sdimboolean_t 283296417Sdimvm_pageout_page_lock(vm_page_t m, vm_page_t *next) 284314564Sdim{ 285314564Sdim struct vm_page marker; 286341825Sdim boolean_t unchanged; 287296417Sdim u_short queue; 288296417Sdim 289309124Sdim vm_page_lock_assert(m, MA_NOTOWNED); 290309124Sdim mtx_assert(&vm_page_queue_mtx, MA_OWNED); 291309124Sdim 292314564Sdim if (vm_page_trylock(m)) 293284677Sdim return (TRUE); 294309124Sdim 295341825Sdim queue = m->queue; 296284677Sdim vm_pageout_init_marker(&marker, queue); 297309124Sdim 298284677Sdim TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq); 299284677Sdim vm_page_unlock_queues(); 300284677Sdim vm_page_lock(m); 301309124Sdim vm_page_lock_queues(); 302341825Sdim 303284677Sdim /* Page queue might have changed. */ 304284677Sdim *next = TAILQ_NEXT(&marker, pageq); 305284677Sdim unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq)); 306284677Sdim TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq); 307309124Sdim return (unchanged); 308341825Sdim} 309284677Sdim 310284677Sdim/* 311284677Sdim * vm_pageout_clean: 312341825Sdim * 313309124Sdim * Clean the page and remove it from the laundry. 314321369Sdim * 315284677Sdim * We set the busy bit to cause potential page faults on this page to 316284677Sdim * block. Note the careful timing, however, the busy bit isn't set till 317284677Sdim * late and we cannot do anything that will mess with the page. 318284677Sdim */ 319284677Sdimstatic int 320284677Sdimvm_pageout_clean(vm_page_t m) 321341825Sdim{ 322309124Sdim vm_object_t object; 323341825Sdim vm_page_t mc[2*vm_pageout_page_count], pb, ps; 324309124Sdim int pageout_count; 325309124Sdim int ib, is, page_base; 326284677Sdim vm_pindex_t pindex = m->pindex; 327284677Sdim 328284677Sdim vm_page_lock_assert(m, MA_OWNED); 329341825Sdim object = m->object; 330309124Sdim VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 331341825Sdim 332309124Sdim /* 333309124Sdim * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 334284677Sdim * with the new swapper, but we could have serious problems paging 335284677Sdim * out other object types if there is insufficient memory. 336284677Sdim * 337341825Sdim * Unfortunately, checking free memory here is far too late, so the 338309124Sdim * check has been moved up a procedural level. 339284677Sdim */ 340284677Sdim 341309124Sdim /* 342284677Sdim * Can't clean the page if it's busy or held. 343284677Sdim */ 344284677Sdim KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0, 345341825Sdim ("vm_pageout_clean: page %p is busy", m)); 346309124Sdim KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m)); 347309124Sdim vm_page_unlock(m); 348309124Sdim 349309124Sdim mc[vm_pageout_page_count] = pb = ps = m; 350309124Sdim pageout_count = 1; 351284677Sdim page_base = vm_pageout_page_count; 352314564Sdim ib = 1; 353341825Sdim is = 1; 354309124Sdim 355309124Sdim /* 356284677Sdim * Scan object for clusterable pages. 357314564Sdim * 358341825Sdim * We can cluster ONLY if: ->> the page is NOT 359309124Sdim * clean, wired, busy, held, or mapped into a 360309124Sdim * buffer, and one of the following: 361341825Sdim * 1) The page is inactive, or a seldom used 362341825Sdim * active page. 363309124Sdim * -or- 364309124Sdim * 2) we force the issue. 365309124Sdim * 366309124Sdim * During heavy mmap/modification loads the pageout 367314564Sdim * daemon can really fragment the underlying file 368341825Sdim * due to flushing pages out of order and not trying 369309124Sdim * align the clusters (which leave sporatic out-of-order 370341825Sdim * holes). To solve this problem we do the reverse scan 371341825Sdim * first and attempt to align our cluster, then do a 372309124Sdim * forward scan if room remains. 373284677Sdim */ 374284677Sdimmore: 375284677Sdim while (ib && pageout_count < vm_pageout_page_count) { 376341825Sdim vm_page_t p; 377341825Sdim 378341825Sdim if (ib > pindex) { 379309124Sdim ib = 0; 380321369Sdim break; 381321369Sdim } 382309124Sdim 383284677Sdim if ((p = vm_page_prev(pb)) == NULL || 384314564Sdim (p->oflags & VPO_BUSY) != 0 || p->busy != 0) { 385341825Sdim ib = 0; 386309124Sdim break; 387321369Sdim } 388321369Sdim vm_page_lock(p); 389321369Sdim vm_page_test_dirty(p); 390309124Sdim if (p->dirty == 0 || 391296417Sdim p->queue != PQ_INACTIVE || 392284677Sdim p->hold_count != 0) { /* may be undergoing I/O */ 393341825Sdim vm_page_unlock(p); 394341825Sdim ib = 0; 395321369Sdim break; 396309124Sdim } 397284677Sdim vm_page_unlock(p); 398341825Sdim mc[--page_base] = pb = p; 399284677Sdim ++pageout_count; 400341825Sdim ++ib; 401341825Sdim /* 402321369Sdim * alignment boundry, stop here and switch directions. Do 403341825Sdim * not clear ib. 404309124Sdim */ 405309124Sdim if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 406341825Sdim break; 407321369Sdim } 408341825Sdim 409284677Sdim while (pageout_count < vm_pageout_page_count && 410284677Sdim pindex + is < object->size) { 411284677Sdim vm_page_t p; 412341825Sdim 413284677Sdim if ((p = vm_page_next(ps)) == NULL || 414341825Sdim (p->oflags & VPO_BUSY) != 0 || p->busy != 0) 415341825Sdim break; 416321369Sdim vm_page_lock(p); 417341825Sdim vm_page_test_dirty(p); 418284677Sdim if (p->dirty == 0 || 419309124Sdim p->queue != PQ_INACTIVE || 420341825Sdim p->hold_count != 0) { /* may be undergoing I/O */ 421321369Sdim vm_page_unlock(p); 422341825Sdim break; 423284677Sdim } 424284677Sdim vm_page_unlock(p); 425284677Sdim mc[page_base + pageout_count] = ps = p; 426341825Sdim ++pageout_count; 427341825Sdim ++is; 428284677Sdim } 429284677Sdim 430309124Sdim /* 431284677Sdim * If we exhausted our forward scan, continue with the reverse scan 432284677Sdim * when possible, even past a page boundry. This catches boundry 433284677Sdim * conditions. 434341825Sdim */ 435341825Sdim if (ib && pageout_count < vm_pageout_page_count) 436284677Sdim goto more; 437284677Sdim 438284677Sdim /* 439284677Sdim * we allow reads during pageouts... 440284677Sdim */ 441284677Sdim return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL)); 442284677Sdim} 443284677Sdim 444296417Sdim/* 445284677Sdim * vm_pageout_flush() - launder the given pages 446284677Sdim * 447341825Sdim * The given pages are laundered. Note that we setup for the start of 448309124Sdim * I/O ( i.e. busy the page ), mark it read-only, and bump the object 449321369Sdim * reference count all in here rather then in the parent. If we want 450321369Sdim * the parent to do more sophisticated things we may have to change 451321369Sdim * the ordering. 452321369Sdim * 453321369Sdim * Returned runlen is the count of pages between mreq and first 454321369Sdim * page after mreq with status VM_PAGER_AGAIN. 455321369Sdim */ 456309124Sdimint 457309124Sdimvm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen) 458284677Sdim{ 459284677Sdim vm_object_t object = mc[0]->object; 460341825Sdim int pageout_status[count]; 461284677Sdim int numpagedout = 0; 462284677Sdim int i, runlen; 463284677Sdim 464284677Sdim VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 465309124Sdim mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 466284677Sdim 467284677Sdim /* 468284677Sdim * Initiate I/O. Bump the vm_page_t->busy counter and 469284677Sdim * mark the pages read-only. 470284677Sdim * 471284677Sdim * We do not have to fixup the clean/dirty bits here... we can 472284677Sdim * allow the pager to do it after the I/O completes. 473284677Sdim * 474284677Sdim * NOTE! mc[i]->dirty may be partial or fragmented due to an 475284677Sdim * edge case with file fragments. 476284677Sdim */ 477284677Sdim for (i = 0; i < count; i++) { 478284677Sdim KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 479284677Sdim ("vm_pageout_flush: partially invalid page %p index %d/%d", 480284677Sdim mc[i], i, count)); 481284677Sdim vm_page_io_start(mc[i]); 482284677Sdim pmap_remove_write(mc[i]); 483284677Sdim } 484284677Sdim vm_object_pip_add(object, count); 485284677Sdim 486284677Sdim vm_pager_put_pages(object, mc, count, flags, pageout_status); 487284677Sdim 488284677Sdim runlen = count - mreq; 489284677Sdim for (i = 0; i < count; i++) { 490284677Sdim vm_page_t mt = mc[i]; 491284677Sdim 492284677Sdim KASSERT(pageout_status[i] == VM_PAGER_PEND || 493284677Sdim (mt->flags & PG_WRITEABLE) == 0, 494284677Sdim ("vm_pageout_flush: page %p is not write protected", mt)); 495284677Sdim switch (pageout_status[i]) { 496284677Sdim case VM_PAGER_OK: 497284677Sdim case VM_PAGER_PEND: 498284677Sdim numpagedout++; 499309124Sdim break; 500284677Sdim case VM_PAGER_BAD: 501284677Sdim /* 502284677Sdim * Page outside of range of object. Right now we 503284677Sdim * essentially lose the changes by pretending it 504284677Sdim * worked. 505341825Sdim */ 506284677Sdim vm_page_undirty(mt); 507284677Sdim break; 508284677Sdim case VM_PAGER_ERROR: 509284677Sdim case VM_PAGER_FAIL: 510284677Sdim /* 511284677Sdim * If page couldn't be paged out, then reactivate the 512284677Sdim * page so it doesn't clog the inactive list. (We 513284677Sdim * will try paging out it again later). 514284677Sdim */ 515284677Sdim vm_page_lock(mt); 516284677Sdim vm_page_activate(mt); 517314564Sdim vm_page_unlock(mt); 518284677Sdim break; 519284677Sdim case VM_PAGER_AGAIN: 520284677Sdim if (i >= mreq && i - mreq < runlen) 521284677Sdim runlen = i - mreq; 522284677Sdim break; 523284677Sdim } 524284677Sdim 525284677Sdim /* 526284677Sdim * If the operation is still going, leave the page busy to 527284677Sdim * block all other accesses. Also, leave the paging in 528284677Sdim * progress indicator set so that we don't attempt an object 529284677Sdim * collapse. 530284677Sdim */ 531321369Sdim if (pageout_status[i] != VM_PAGER_PEND) { 532341825Sdim vm_object_pip_wakeup(object); 533341825Sdim vm_page_io_finish(mt); 534284677Sdim if (vm_page_count_severe()) { 535284677Sdim vm_page_lock(mt); 536341825Sdim vm_page_try_to_cache(mt); 537284677Sdim vm_page_unlock(mt); 538284677Sdim } 539341825Sdim } 540284677Sdim } 541284677Sdim if (prunlen != NULL) 542284677Sdim *prunlen = runlen; 543314564Sdim return (numpagedout); 544284677Sdim} 545284677Sdim 546284677Sdim#if !defined(NO_SWAPPING) 547284677Sdim/* 548284677Sdim * vm_pageout_object_deactivate_pages 549284677Sdim * 550284677Sdim * Deactivate enough pages to satisfy the inactive target 551284677Sdim * requirements. 552284677Sdim * 553284677Sdim * The object and map must be locked. 554284677Sdim */ 555284677Sdimstatic void 556284677Sdimvm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 557284677Sdim long desired) 558284677Sdim{ 559284677Sdim vm_object_t backing_object, object; 560284677Sdim vm_page_t p; 561284677Sdim int actcount, remove_mode; 562284677Sdim 563284677Sdim VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 564284677Sdim if (first_object->type == OBJT_DEVICE || 565284677Sdim first_object->type == OBJT_SG) 566284677Sdim return; 567284677Sdim for (object = first_object;; object = backing_object) { 568284677Sdim if (pmap_resident_count(pmap) <= desired) 569341825Sdim goto unlock_return; 570284677Sdim VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 571284677Sdim if (object->type == OBJT_PHYS || object->paging_in_progress) 572284677Sdim goto unlock_return; 573284677Sdim 574284677Sdim remove_mode = 0; 575284677Sdim if (object->shadow_count > 1) 576284677Sdim remove_mode = 1; 577284677Sdim /* 578284677Sdim * Scan the object's entire memory queue. 579284677Sdim */ 580284677Sdim TAILQ_FOREACH(p, &object->memq, listq) { 581284677Sdim if (pmap_resident_count(pmap) <= desired) 582284677Sdim goto unlock_return; 583284677Sdim if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0) 584284677Sdim continue; 585284677Sdim PCPU_INC(cnt.v_pdpages); 586284677Sdim vm_page_lock(p); 587284677Sdim if (p->wire_count != 0 || p->hold_count != 0 || 588284677Sdim !pmap_page_exists_quick(pmap, p)) { 589284677Sdim vm_page_unlock(p); 590284677Sdim continue; 591309124Sdim } 592321369Sdim actcount = pmap_ts_referenced(p); 593341825Sdim if ((p->flags & PG_REFERENCED) != 0) { 594309124Sdim if (actcount == 0) 595309124Sdim actcount = 1; 596284677Sdim vm_page_lock_queues(); 597284677Sdim vm_page_flag_clear(p, PG_REFERENCED); 598284677Sdim vm_page_unlock_queues(); 599284677Sdim } 600284677Sdim if (p->queue != PQ_ACTIVE && actcount != 0) { 601284677Sdim vm_page_activate(p); 602284677Sdim p->act_count += actcount; 603284677Sdim } else if (p->queue == PQ_ACTIVE) { 604284677Sdim if (actcount == 0) { 605284677Sdim p->act_count -= min(p->act_count, 606284677Sdim ACT_DECLINE); 607284677Sdim if (!remove_mode && 608284677Sdim (vm_pageout_algorithm || 609284677Sdim p->act_count == 0)) { 610284677Sdim pmap_remove_all(p); 611284677Sdim vm_page_deactivate(p); 612284677Sdim } else { 613284677Sdim vm_page_lock_queues(); 614284677Sdim vm_page_requeue(p); 615284677Sdim vm_page_unlock_queues(); 616341825Sdim } 617341825Sdim } else { 618284677Sdim vm_page_activate(p); 619341825Sdim if (p->act_count < ACT_MAX - 620341825Sdim ACT_ADVANCE) 621284677Sdim p->act_count += ACT_ADVANCE; 622341825Sdim vm_page_lock_queues(); 623341825Sdim vm_page_requeue(p); 624284677Sdim vm_page_unlock_queues(); 625341825Sdim } 626341825Sdim } else if (p->queue == PQ_INACTIVE) 627284677Sdim pmap_remove_all(p); 628341825Sdim vm_page_unlock(p); 629341825Sdim } 630284677Sdim if ((backing_object = object->backing_object) == NULL) 631341825Sdim goto unlock_return; 632341825Sdim VM_OBJECT_LOCK(backing_object); 633284677Sdim if (object != first_object) 634309124Sdim VM_OBJECT_UNLOCK(object); 635309124Sdim } 636309124Sdimunlock_return: 637309124Sdim if (object != first_object) 638309124Sdim VM_OBJECT_UNLOCK(object); 639321369Sdim} 640321369Sdim 641284677Sdim/* 642309124Sdim * deactivate some number of pages in a map, try to do it fairly, but 643284677Sdim * that is really hard to do. 644284677Sdim */ 645284677Sdimstatic void 646284677Sdimvm_pageout_map_deactivate_pages(map, desired) 647284677Sdim vm_map_t map; 648284677Sdim long desired; 649284677Sdim{ 650284677Sdim vm_map_entry_t tmpe; 651284677Sdim vm_object_t obj, bigobj; 652284677Sdim int nothingwired; 653284677Sdim 654284677Sdim if (!vm_map_trylock(map)) 655284677Sdim return; 656284677Sdim 657284677Sdim bigobj = NULL; 658284677Sdim nothingwired = TRUE; 659309124Sdim 660284677Sdim /* 661284677Sdim * first, search out the biggest object, and try to free pages from 662314564Sdim * that. 663314564Sdim */ 664314564Sdim tmpe = map->header.next; 665314564Sdim while (tmpe != &map->header) { 666284677Sdim if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 667309124Sdim obj = tmpe->object.vm_object; 668309124Sdim if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 669309124Sdim if (obj->shadow_count <= 1 && 670309124Sdim (bigobj == NULL || 671309124Sdim bigobj->resident_page_count < obj->resident_page_count)) { 672284677Sdim if (bigobj != NULL) 673284677Sdim VM_OBJECT_UNLOCK(bigobj); 674284677Sdim bigobj = obj; 675284677Sdim } else 676284677Sdim VM_OBJECT_UNLOCK(obj); 677284677Sdim } 678284677Sdim } 679284677Sdim if (tmpe->wired_count > 0) 680284677Sdim nothingwired = FALSE; 681284677Sdim tmpe = tmpe->next; 682284677Sdim } 683284677Sdim 684284677Sdim if (bigobj != NULL) { 685284677Sdim vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 686284677Sdim VM_OBJECT_UNLOCK(bigobj); 687284677Sdim } 688284677Sdim /* 689284677Sdim * Next, hunt around for other pages to deactivate. We actually 690284677Sdim * do this search sort of wrong -- .text first is not the best idea. 691284677Sdim */ 692284677Sdim tmpe = map->header.next; 693284677Sdim while (tmpe != &map->header) { 694284677Sdim if (pmap_resident_count(vm_map_pmap(map)) <= desired) 695284677Sdim break; 696284677Sdim if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 697284677Sdim obj = tmpe->object.vm_object; 698284677Sdim if (obj != NULL) { 699314564Sdim VM_OBJECT_LOCK(obj); 700286684Sdim vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 701286684Sdim VM_OBJECT_UNLOCK(obj); 702286684Sdim } 703284677Sdim } 704284677Sdim tmpe = tmpe->next; 705284677Sdim } 706284677Sdim 707284677Sdim /* 708284677Sdim * Remove all mappings if a process is swapped out, this will free page 709284677Sdim * table pages. 710284677Sdim */ 711284677Sdim if (desired == 0 && nothingwired) { 712284677Sdim tmpe = map->header.next; 713284677Sdim while (tmpe != &map->header) { 714284677Sdim pmap_remove(vm_map_pmap(map), tmpe->start, tmpe->end); 715284677Sdim tmpe = tmpe->next; 716284677Sdim } 717284677Sdim } 718284677Sdim vm_map_unlock(map); 719284677Sdim} 720284677Sdim#endif /* !defined(NO_SWAPPING) */ 721284677Sdim 722284677Sdim/* 723284677Sdim * vm_pageout_scan does the dirty work for the pageout daemon. 724284677Sdim */ 725284677Sdimstatic void 726284677Sdimvm_pageout_scan(int pass) 727284677Sdim{ 728284677Sdim vm_page_t m, next; 729284677Sdim struct vm_page marker; 730284677Sdim int page_shortage, maxscan, pcount; 731284677Sdim int addl_page_shortage, addl_page_shortage_init; 732284677Sdim vm_object_t object; 733284677Sdim int actcount; 734284677Sdim int vnodes_skipped = 0; 735284677Sdim int maxlaunder; 736284677Sdim 737284677Sdim /* 738284677Sdim * Decrease registered cache sizes. 739284677Sdim */ 740309124Sdim EVENTHANDLER_INVOKE(vm_lowmem, 0); 741309124Sdim /* 742309124Sdim * We do this explicitly after the caches have been drained above. 743309124Sdim */ 744321369Sdim uma_reclaim(); 745309124Sdim 746309124Sdim addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 747309124Sdim 748309124Sdim /* 749321369Sdim * Calculate the number of pages we want to either free or move 750309124Sdim * to the cache. 751309124Sdim */ 752309124Sdim page_shortage = vm_paging_target() + addl_page_shortage_init; 753309124Sdim 754309124Sdim vm_pageout_init_marker(&marker, PQ_INACTIVE); 755284677Sdim 756284677Sdim /* 757284677Sdim * Start scanning the inactive queue for pages we can move to the 758284677Sdim * cache or free. The scan will stop when the target is reached or 759284677Sdim * we have scanned the entire inactive queue. Note that m->act_count 760284677Sdim * is not used to form decisions for the inactive queue, only for the 761296417Sdim * active queue. 762296417Sdim * 763284677Sdim * maxlaunder limits the number of dirty pages we flush per scan. 764284677Sdim * For most systems a smaller value (16 or 32) is more robust under 765284677Sdim * extreme memory and disk pressure because any unnecessary writes 766284677Sdim * to disk can result in extreme performance degredation. However, 767284677Sdim * systems with excessive dirty pages (especially when MAP_NOSYNC is 768284677Sdim * used) will die horribly with limited laundering. If the pageout 769284677Sdim * daemon cannot clean enough pages in the first pass, we let it go 770284677Sdim * all out in succeeding passes. 771284677Sdim */ 772284677Sdim if ((maxlaunder = vm_max_launder) <= 1) 773284677Sdim maxlaunder = 1; 774284677Sdim if (pass) 775284677Sdim maxlaunder = 10000; 776284677Sdim vm_page_lock_queues(); 777284677Sdimrescan0: 778284677Sdim addl_page_shortage = addl_page_shortage_init; 779284677Sdim maxscan = cnt.v_inactive_count; 780284677Sdim 781284677Sdim for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 782341825Sdim m != NULL && maxscan-- > 0 && page_shortage > 0; 783284677Sdim m = next) { 784284677Sdim 785284677Sdim cnt.v_pdpages++; 786284677Sdim 787284677Sdim if (m->queue != PQ_INACTIVE) 788284677Sdim goto rescan0; 789284677Sdim 790284677Sdim next = TAILQ_NEXT(m, pageq); 791284677Sdim 792284677Sdim /* 793284677Sdim * skip marker pages 794284677Sdim */ 795284677Sdim if (m->flags & PG_MARKER) 796284677Sdim continue; 797284677Sdim 798284677Sdim /* 799284677Sdim * Lock the page. 800284677Sdim */ 801284677Sdim if (!vm_pageout_page_lock(m, &next)) { 802284677Sdim vm_page_unlock(m); 803284677Sdim addl_page_shortage++; 804284677Sdim continue; 805284677Sdim } 806284677Sdim 807284677Sdim /* 808284677Sdim * A held page may be undergoing I/O, so skip it. 809284677Sdim */ 810284677Sdim if (m->hold_count) { 811284677Sdim vm_page_unlock(m); 812284677Sdim vm_page_requeue(m); 813284677Sdim addl_page_shortage++; 814284677Sdim continue; 815284677Sdim } 816284677Sdim 817284677Sdim /* 818284677Sdim * Don't mess with busy pages, keep in the front of the 819284677Sdim * queue, most likely are being paged out. 820284677Sdim */ 821284677Sdim object = m->object; 822284677Sdim if (!VM_OBJECT_TRYLOCK(object) && 823284677Sdim (!vm_pageout_fallback_object_lock(m, &next) || 824284677Sdim m->hold_count != 0)) { 825284677Sdim VM_OBJECT_UNLOCK(object); 826284677Sdim vm_page_unlock(m); 827284677Sdim addl_page_shortage++; 828284677Sdim continue; 829284677Sdim } 830284677Sdim if (m->busy || (m->oflags & VPO_BUSY)) { 831284677Sdim vm_page_unlock(m); 832284677Sdim VM_OBJECT_UNLOCK(object); 833284677Sdim addl_page_shortage++; 834284677Sdim continue; 835284677Sdim } 836284677Sdim 837284677Sdim /* 838284677Sdim * If the object is not being used, we ignore previous 839284677Sdim * references. 840284677Sdim */ 841284677Sdim if (object->ref_count == 0) { 842284677Sdim vm_page_flag_clear(m, PG_REFERENCED); 843284677Sdim KASSERT(!pmap_page_is_mapped(m), 844284677Sdim ("vm_pageout_scan: page %p is mapped", m)); 845284677Sdim 846284677Sdim /* 847284677Sdim * Otherwise, if the page has been referenced while in the 848284677Sdim * inactive queue, we bump the "activation count" upwards, 849284677Sdim * making it less likely that the page will be added back to 850284677Sdim * the inactive queue prematurely again. Here we check the 851284677Sdim * page tables (or emulated bits, if any), given the upper 852284677Sdim * level VM system not knowing anything about existing 853284677Sdim * references. 854284677Sdim */ 855284677Sdim } else if (((m->flags & PG_REFERENCED) == 0) && 856284677Sdim (actcount = pmap_ts_referenced(m))) { 857284677Sdim vm_page_activate(m); 858284677Sdim VM_OBJECT_UNLOCK(object); 859284677Sdim m->act_count += (actcount + ACT_ADVANCE); 860284677Sdim vm_page_unlock(m); 861284677Sdim continue; 862284677Sdim } 863284677Sdim 864284677Sdim /* 865284677Sdim * If the upper level VM system knows about any page 866284677Sdim * references, we activate the page. We also set the 867284677Sdim * "activation count" higher than normal so that we will less 868284677Sdim * likely place pages back onto the inactive queue again. 869284677Sdim */ 870284677Sdim if ((m->flags & PG_REFERENCED) != 0) { 871284677Sdim vm_page_flag_clear(m, PG_REFERENCED); 872284677Sdim actcount = pmap_ts_referenced(m); 873284677Sdim vm_page_activate(m); 874284677Sdim VM_OBJECT_UNLOCK(object); 875284677Sdim m->act_count += (actcount + ACT_ADVANCE + 1); 876284677Sdim vm_page_unlock(m); 877284677Sdim continue; 878284677Sdim } 879284677Sdim 880284677Sdim /* 881309124Sdim * If the upper level VM system does not believe that the page 882284677Sdim * is fully dirty, but it is mapped for write access, then we 883284677Sdim * consult the pmap to see if the page's dirty status should 884284677Sdim * be updated. 885284677Sdim */ 886284677Sdim if (m->dirty != VM_PAGE_BITS_ALL && 887309124Sdim (m->flags & PG_WRITEABLE) != 0) { 888309124Sdim /* 889284677Sdim * Avoid a race condition: Unless write access is 890284677Sdim * removed from the page, another processor could 891309124Sdim * modify it before all access is removed by the call 892309124Sdim * to vm_page_cache() below. If vm_page_cache() finds 893309124Sdim * that the page has been modified when it removes all 894309124Sdim * access, it panics because it cannot cache dirty 895309124Sdim * pages. In principle, we could eliminate just write 896309124Sdim * access here rather than all access. In the expected 897309124Sdim * case, when there are no last instant modifications 898309124Sdim * to the page, removing all access will be cheaper 899309124Sdim * overall. 900309124Sdim */ 901284677Sdim if (pmap_is_modified(m)) 902309124Sdim vm_page_dirty(m); 903284677Sdim else if (m->dirty == 0) 904284677Sdim pmap_remove_all(m); 905284677Sdim } 906341825Sdim 907284677Sdim if (m->valid == 0) { 908284677Sdim /* 909284677Sdim * Invalid pages can be easily freed 910284677Sdim */ 911284677Sdim vm_page_free(m); 912284677Sdim cnt.v_dfree++; 913309124Sdim --page_shortage; 914284677Sdim } else if (m->dirty == 0) { 915284677Sdim /* 916284677Sdim * Clean pages can be placed onto the cache queue. 917284677Sdim * This effectively frees them. 918284677Sdim */ 919284677Sdim vm_page_cache(m); 920284677Sdim --page_shortage; 921284677Sdim } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 922284677Sdim /* 923284677Sdim * Dirty pages need to be paged out, but flushing 924284677Sdim * a page is extremely expensive verses freeing 925284677Sdim * a clean page. Rather then artificially limiting 926309124Sdim * the number of pages we can flush, we instead give 927309124Sdim * dirty pages extra priority on the inactive queue 928309124Sdim * by forcing them to be cycled through the queue 929309124Sdim * twice before being flushed, after which the 930309124Sdim * (now clean) page will cycle through once more 931309124Sdim * before being freed. This significantly extends 932309124Sdim * the thrash point for a heavily loaded machine. 933309124Sdim */ 934309124Sdim vm_page_flag_set(m, PG_WINATCFLS); 935309124Sdim vm_page_requeue(m); 936309124Sdim } else if (maxlaunder > 0) { 937309124Sdim /* 938309124Sdim * We always want to try to flush some dirty pages if 939309124Sdim * we encounter them, to keep the system stable. 940284677Sdim * Normally this number is small, but under extreme 941284677Sdim * pressure where there are insufficient clean pages 942284677Sdim * on the inactive queue, we may have to go all out. 943284677Sdim */ 944284677Sdim int swap_pageouts_ok, vfslocked = 0; 945284677Sdim struct vnode *vp = NULL; 946284677Sdim struct mount *mp = NULL; 947284677Sdim 948284677Sdim if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 949284677Sdim swap_pageouts_ok = 1; 950284677Sdim } else { 951284677Sdim swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 952284677Sdim swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 953321369Sdim vm_page_count_min()); 954284677Sdim 955284677Sdim } 956284677Sdim 957284677Sdim /* 958284677Sdim * We don't bother paging objects that are "dead". 959284677Sdim * Those objects are in a "rundown" state. 960284677Sdim */ 961284677Sdim if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 962284677Sdim vm_page_unlock(m); 963284677Sdim VM_OBJECT_UNLOCK(object); 964284677Sdim vm_page_requeue(m); 965284677Sdim continue; 966284677Sdim } 967284677Sdim 968284677Sdim /* 969284677Sdim * Following operations may unlock 970284677Sdim * vm_page_queue_mtx, invalidating the 'next' 971284677Sdim * pointer. To prevent an inordinate number 972284677Sdim * of restarts we use our marker to remember 973284677Sdim * our place. 974284677Sdim * 975284677Sdim */ 976284677Sdim TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 977284677Sdim m, &marker, pageq); 978284677Sdim /* 979284677Sdim * The object is already known NOT to be dead. It 980284677Sdim * is possible for the vget() to block the whole 981284677Sdim * pageout daemon, but the new low-memory handling 982284677Sdim * code should prevent it. 983284677Sdim * 984284677Sdim * The previous code skipped locked vnodes and, worse, 985284677Sdim * reordered pages in the queue. This results in 986284677Sdim * completely non-deterministic operation and, on a 987284677Sdim * busy system, can lead to extremely non-optimal 988284677Sdim * pageouts. For example, it can cause clean pages 989284677Sdim * to be freed and dirty pages to be moved to the end 990284677Sdim * of the queue. Since dirty pages are also moved to 991284677Sdim * the end of the queue once-cleaned, this gives 992284677Sdim * way too large a weighting to defering the freeing 993284677Sdim * of dirty pages. 994284677Sdim * 995284677Sdim * We can't wait forever for the vnode lock, we might 996284677Sdim * deadlock due to a vn_read() getting stuck in 997284677Sdim * vm_wait while holding this vnode. We skip the 998284677Sdim * vnode if we can't get it in a reasonable amount 999284677Sdim * of time. 1000284677Sdim */ 1001284677Sdim if (object->type == OBJT_VNODE) { 1002284677Sdim vm_page_unlock_queues(); 1003284677Sdim vm_page_unlock(m); 1004284677Sdim vp = object->handle; 1005284677Sdim if (vp->v_type == VREG && 1006284677Sdim vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1007284677Sdim mp = NULL; 1008284677Sdim ++pageout_lock_miss; 1009284677Sdim if (object->flags & OBJ_MIGHTBEDIRTY) 1010284677Sdim vnodes_skipped++; 1011284677Sdim vm_page_lock_queues(); 1012284677Sdim goto unlock_and_continue; 1013284677Sdim } 1014284677Sdim KASSERT(mp != NULL, 1015284677Sdim ("vp %p with NULL v_mount", vp)); 1016284677Sdim vm_object_reference_locked(object); 1017284677Sdim VM_OBJECT_UNLOCK(object); 1018284677Sdim vfslocked = VFS_LOCK_GIANT(vp->v_mount); 1019284677Sdim if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 1020284677Sdim curthread)) { 1021284677Sdim VM_OBJECT_LOCK(object); 1022284677Sdim vm_page_lock_queues(); 1023284677Sdim ++pageout_lock_miss; 1024284677Sdim if (object->flags & OBJ_MIGHTBEDIRTY) 1025284677Sdim vnodes_skipped++; 1026284677Sdim vp = NULL; 1027284677Sdim goto unlock_and_continue; 1028284677Sdim } 1029284677Sdim VM_OBJECT_LOCK(object); 1030284677Sdim vm_page_lock(m); 1031284677Sdim vm_page_lock_queues(); 1032284677Sdim /* 1033284677Sdim * The page might have been moved to another 1034284677Sdim * queue during potential blocking in vget() 1035284677Sdim * above. The page might have been freed and 1036284677Sdim * reused for another vnode. 1037284677Sdim */ 1038284677Sdim if (m->queue != PQ_INACTIVE || 1039284677Sdim m->object != object || 1040284677Sdim TAILQ_NEXT(m, pageq) != &marker) { 1041284677Sdim vm_page_unlock(m); 1042284677Sdim if (object->flags & OBJ_MIGHTBEDIRTY) 1043284677Sdim vnodes_skipped++; 1044284677Sdim goto unlock_and_continue; 1045284677Sdim } 1046284677Sdim 1047284677Sdim /* 1048284677Sdim * The page may have been busied during the 1049284677Sdim * blocking in vget(). We don't move the 1050284677Sdim * page back onto the end of the queue so that 1051284677Sdim * statistics are more correct if we don't. 1052284677Sdim */ 1053284677Sdim if (m->busy || (m->oflags & VPO_BUSY)) { 1054284677Sdim vm_page_unlock(m); 1055284677Sdim goto unlock_and_continue; 1056284677Sdim } 1057284677Sdim 1058284677Sdim /* 1059284677Sdim * If the page has become held it might 1060284677Sdim * be undergoing I/O, so skip it 1061284677Sdim */ 1062284677Sdim if (m->hold_count) { 1063284677Sdim vm_page_unlock(m); 1064284677Sdim vm_page_requeue(m); 1065284677Sdim if (object->flags & OBJ_MIGHTBEDIRTY) 1066284677Sdim vnodes_skipped++; 1067284677Sdim goto unlock_and_continue; 1068284677Sdim } 1069284677Sdim } 1070284677Sdim 1071284677Sdim /* 1072284677Sdim * If a page is dirty, then it is either being washed 1073284677Sdim * (but not yet cleaned) or it is still in the 1074284677Sdim * laundry. If it is still in the laundry, then we 1075284677Sdim * start the cleaning operation. 1076284677Sdim * 1077284677Sdim * decrement page_shortage on success to account for 1078284677Sdim * the (future) cleaned page. Otherwise we could wind 1079284677Sdim * up laundering or cleaning too many pages. 1080284677Sdim */ 1081284677Sdim vm_page_unlock_queues(); 1082284677Sdim if (vm_pageout_clean(m) != 0) { 1083284677Sdim --page_shortage; 1084284677Sdim --maxlaunder; 1085284677Sdim } 1086284677Sdim vm_page_lock_queues(); 1087284677Sdimunlock_and_continue: 1088284677Sdim vm_page_lock_assert(m, MA_NOTOWNED); 1089284677Sdim VM_OBJECT_UNLOCK(object); 1090284677Sdim if (mp != NULL) { 1091284677Sdim vm_page_unlock_queues(); 1092284677Sdim if (vp != NULL) 1093284677Sdim vput(vp); 1094284677Sdim VFS_UNLOCK_GIANT(vfslocked); 1095284677Sdim vm_object_deallocate(object); 1096284677Sdim vn_finished_write(mp); 1097284677Sdim vm_page_lock_queues(); 1098284677Sdim } 1099284677Sdim next = TAILQ_NEXT(&marker, pageq); 1100284677Sdim TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1101284677Sdim &marker, pageq); 1102284677Sdim vm_page_lock_assert(m, MA_NOTOWNED); 1103284677Sdim continue; 1104284677Sdim } 1105284677Sdim vm_page_unlock(m); 1106284677Sdim VM_OBJECT_UNLOCK(object); 1107284677Sdim } 1108284677Sdim 1109284677Sdim /* 1110284677Sdim * Compute the number of pages we want to try to move from the 1111284677Sdim * active queue to the inactive queue. 1112284677Sdim */ 1113284677Sdim page_shortage = vm_paging_target() + 1114284677Sdim cnt.v_inactive_target - cnt.v_inactive_count; 1115284677Sdim page_shortage += addl_page_shortage; 1116284677Sdim 1117284677Sdim /* 1118284677Sdim * Scan the active queue for things we can deactivate. We nominally 1119284677Sdim * track the per-page activity counter and use it to locate 1120284677Sdim * deactivation candidates. 1121284677Sdim */ 1122284677Sdim pcount = cnt.v_active_count; 1123284677Sdim m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1124284677Sdim mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1125284677Sdim 1126284677Sdim while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1127284677Sdim 1128284677Sdim KASSERT(m->queue == PQ_ACTIVE, 1129284677Sdim ("vm_pageout_scan: page %p isn't active", m)); 1130284677Sdim 1131284677Sdim next = TAILQ_NEXT(m, pageq); 1132284677Sdim if ((m->flags & PG_MARKER) != 0) { 1133284677Sdim m = next; 1134284677Sdim continue; 1135284677Sdim } 1136284677Sdim if (!vm_pageout_page_lock(m, &next)) { 1137284677Sdim vm_page_unlock(m); 1138309124Sdim m = next; 1139309124Sdim continue; 1140309124Sdim } 1141314564Sdim object = m->object; 1142314564Sdim if (!VM_OBJECT_TRYLOCK(object) && 1143314564Sdim !vm_pageout_fallback_object_lock(m, &next)) { 1144321369Sdim VM_OBJECT_UNLOCK(object); 1145309124Sdim vm_page_unlock(m); 1146314564Sdim m = next; 1147309124Sdim continue; 1148314564Sdim } 1149314564Sdim 1150309124Sdim /* 1151314564Sdim * Don't deactivate pages that are busy. 1152321369Sdim */ 1153314564Sdim if ((m->busy != 0) || 1154314564Sdim (m->oflags & VPO_BUSY) || 1155309124Sdim (m->hold_count != 0)) { 1156309124Sdim vm_page_unlock(m); 1157314564Sdim VM_OBJECT_UNLOCK(object); 1158314564Sdim vm_page_requeue(m); 1159314564Sdim m = next; 1160314564Sdim continue; 1161309124Sdim } 1162314564Sdim 1163309124Sdim /* 1164309124Sdim * The count for pagedaemon pages is done after checking the 1165314564Sdim * page for eligibility... 1166314564Sdim */ 1167314564Sdim cnt.v_pdpages++; 1168314564Sdim 1169309124Sdim /* 1170314564Sdim * Check to see "how much" the page has been used. 1171314564Sdim */ 1172314564Sdim actcount = 0; 1173314564Sdim if (object->ref_count != 0) { 1174314564Sdim if (m->flags & PG_REFERENCED) { 1175314564Sdim actcount += 1; 1176314564Sdim } 1177327952Sdim actcount += pmap_ts_referenced(m); 1178327952Sdim if (actcount) { 1179327952Sdim m->act_count += ACT_ADVANCE + actcount; 1180314564Sdim if (m->act_count > ACT_MAX) 1181314564Sdim m->act_count = ACT_MAX; 1182314564Sdim } 1183314564Sdim } 1184314564Sdim 1185309124Sdim /* 1186309124Sdim * Since we have "tested" this bit, we need to clear it now. 1187314564Sdim */ 1188309124Sdim vm_page_flag_clear(m, PG_REFERENCED); 1189309124Sdim 1190309124Sdim /* 1191314564Sdim * Only if an object is currently being used, do we use the 1192314564Sdim * page activation count stats. 1193309124Sdim */ 1194309124Sdim if (actcount && (object->ref_count != 0)) { 1195309124Sdim vm_page_requeue(m); 1196314564Sdim } else { 1197309124Sdim m->act_count -= min(m->act_count, ACT_DECLINE); 1198309124Sdim if (vm_pageout_algorithm || 1199314564Sdim object->ref_count == 0 || 1200309124Sdim m->act_count == 0) { 1201309124Sdim page_shortage--; 1202309124Sdim if (object->ref_count == 0) { 1203314564Sdim KASSERT(!pmap_page_is_mapped(m), 1204314564Sdim ("vm_pageout_scan: page %p is mapped", m)); 1205314564Sdim if (m->dirty == 0) 1206314564Sdim vm_page_cache(m); 1207314564Sdim else 1208314564Sdim vm_page_deactivate(m); 1209314564Sdim } else { 1210314564Sdim vm_page_deactivate(m); 1211309124Sdim } 1212309124Sdim } else { 1213314564Sdim vm_page_requeue(m); 1214309124Sdim } 1215314564Sdim } 1216314564Sdim vm_page_unlock(m); 1217314564Sdim VM_OBJECT_UNLOCK(object); 1218327952Sdim m = next; 1219314564Sdim } 1220314564Sdim vm_page_unlock_queues(); 1221314564Sdim#if !defined(NO_SWAPPING) 1222314564Sdim /* 1223314564Sdim * Idle process swapout -- run once per second. 1224314564Sdim */ 1225314564Sdim if (vm_swap_idle_enabled) { 1226314564Sdim static long lsec; 1227309124Sdim if (time_second != lsec) { 1228309124Sdim vm_req_vmdaemon(VM_SWAP_IDLE); 1229284677Sdim lsec = time_second; 1230284677Sdim } 1231309124Sdim } 1232314564Sdim#endif 1233314564Sdim 1234314564Sdim /* 1235309124Sdim * If we didn't get enough free pages, and we have skipped a vnode 1236284677Sdim * in a writeable object, wakeup the sync daemon. And kick swapout 1237314564Sdim * if we did not get enough free pages. 1238314564Sdim */ 1239314564Sdim if (vm_paging_target() > 0) { 1240314564Sdim if (vnodes_skipped && vm_page_count_min()) 1241314564Sdim (void) speedup_syncer(); 1242314564Sdim#if !defined(NO_SWAPPING) 1243314564Sdim if (vm_swap_enabled && vm_page_count_target()) 1244321369Sdim vm_req_vmdaemon(VM_SWAP_NORMAL); 1245314564Sdim#endif 1246321369Sdim } 1247321369Sdim 1248314564Sdim /* 1249314564Sdim * If we are critically low on one of RAM or swap and low on 1250314564Sdim * the other, kill the largest process. However, we avoid 1251314564Sdim * doing this on the first pass in order to give ourselves a 1252314564Sdim * chance to flush out dirty vnode-backed pages and to allow 1253314564Sdim * active pages to be moved to the inactive queue and reclaimed. 1254314564Sdim */ 1255314564Sdim if (pass != 0 && 1256314564Sdim ((swap_pager_avail < 64 && vm_page_count_min()) || 1257314564Sdim (swap_pager_full && vm_paging_target() > 0))) 1258314564Sdim vm_pageout_oom(VM_OOM_MEM); 1259284677Sdim} 1260284677Sdim 1261314564Sdim 1262314564Sdimvoid 1263314564Sdimvm_pageout_oom(int shortage) 1264314564Sdim{ 1265314564Sdim struct proc *p, *bigproc; 1266309124Sdim vm_offset_t size, bigsize; 1267314564Sdim struct thread *td; 1268314564Sdim struct vmspace *vm; 1269314564Sdim 1270321369Sdim /* 1271314564Sdim * We keep the process bigproc locked once we find it to keep anyone 1272314564Sdim * from messing with it; however, there is a possibility of 1273284677Sdim * deadlock if process B is bigproc and one of it's child processes 1274284677Sdim * attempts to propagate a signal to B while we are waiting for A's 1275284677Sdim * lock while walking this list. To avoid this, we don't block on 1276284677Sdim * the process lock but just skip a process if it is already locked. 1277284677Sdim */ 1278284677Sdim bigproc = NULL; 1279284677Sdim bigsize = 0; 1280314564Sdim sx_slock(&allproc_lock); 1281284677Sdim FOREACH_PROC_IN_SYSTEM(p) { 1282284677Sdim int breakout; 1283314564Sdim 1284314564Sdim if (PROC_TRYLOCK(p) == 0) 1285314564Sdim continue; 1286314564Sdim /* 1287314564Sdim * If this is a system, protected or killed process, skip it. 1288314564Sdim */ 1289314564Sdim if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1290314564Sdim (p->p_pid == 1) || P_KILLED(p) || 1291314564Sdim ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1292314564Sdim PROC_UNLOCK(p); 1293284677Sdim continue; 1294314564Sdim } 1295314564Sdim /* 1296284677Sdim * If the process is in a non-running type state, 1297284677Sdim * don't touch it. Check all the threads individually. 1298284677Sdim */ 1299284677Sdim breakout = 0; 1300284677Sdim FOREACH_THREAD_IN_PROC(p, td) { 1301284677Sdim thread_lock(td); 1302284677Sdim if (!TD_ON_RUNQ(td) && 1303284677Sdim !TD_IS_RUNNING(td) && 1304309124Sdim !TD_IS_SLEEPING(td)) { 1305284677Sdim thread_unlock(td); 1306284677Sdim breakout = 1; 1307284677Sdim break; 1308284677Sdim } 1309314564Sdim thread_unlock(td); 1310284677Sdim } 1311314564Sdim if (breakout) { 1312284677Sdim PROC_UNLOCK(p); 1313284677Sdim continue; 1314284677Sdim } 1315284677Sdim /* 1316284677Sdim * get the process size 1317284677Sdim */ 1318284677Sdim vm = vmspace_acquire_ref(p); 1319284677Sdim if (vm == NULL) { 1320284677Sdim PROC_UNLOCK(p); 1321284677Sdim continue; 1322314564Sdim } 1323321369Sdim if (!vm_map_trylock_read(&vm->vm_map)) { 1324284677Sdim vmspace_free(vm); 1325284677Sdim PROC_UNLOCK(p); 1326309124Sdim continue; 1327309124Sdim } 1328309124Sdim size = vmspace_swap_count(vm); 1329314564Sdim vm_map_unlock_read(&vm->vm_map); 1330314564Sdim if (shortage == VM_OOM_MEM) 1331314564Sdim size += vmspace_resident_count(vm); 1332314564Sdim vmspace_free(vm); 1333314564Sdim /* 1334284677Sdim * if the this process is bigger than the biggest one 1335284677Sdim * remember it. 1336314564Sdim */ 1337314564Sdim if (size > bigsize) { 1338284677Sdim if (bigproc != NULL) 1339284677Sdim PROC_UNLOCK(bigproc); 1340284677Sdim bigproc = p; 1341284677Sdim bigsize = size; 1342327952Sdim } else 1343284677Sdim PROC_UNLOCK(p); 1344327952Sdim } 1345284677Sdim sx_sunlock(&allproc_lock); 1346327952Sdim if (bigproc != NULL) { 1347284677Sdim killproc(bigproc, "out of swap space"); 1348327952Sdim sched_nice(bigproc, PRIO_MIN); 1349284677Sdim PROC_UNLOCK(bigproc); 1350327952Sdim wakeup(&cnt.v_free_count); 1351284677Sdim } 1352327952Sdim} 1353284677Sdim 1354327952Sdim/* 1355284677Sdim * This routine tries to maintain the pseudo LRU active queue, 1356327952Sdim * so that during long periods of time where there is no paging, 1357284677Sdim * that some statistic accumulation still occurs. This code 1358327952Sdim * helps the situation where paging just starts to occur. 1359284677Sdim */ 1360327952Sdimstatic void 1361284677Sdimvm_pageout_page_stats() 1362327952Sdim{ 1363284677Sdim vm_object_t object; 1364327952Sdim vm_page_t m,next; 1365284677Sdim int pcount,tpcount; /* Number of pages to check */ 1366327952Sdim static int fullintervalcount = 0; 1367284677Sdim int page_shortage; 1368327952Sdim 1369284677Sdim page_shortage = 1370327952Sdim (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1371284677Sdim (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1372327952Sdim 1373284677Sdim if (page_shortage <= 0) 1374327952Sdim return; 1375284677Sdim 1376284677Sdim vm_page_lock_queues(); 1377284677Sdim pcount = cnt.v_active_count; 1378284677Sdim fullintervalcount += vm_pageout_stats_interval; 1379284677Sdim if (fullintervalcount < vm_pageout_full_stats_interval) { 1380284677Sdim tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1381309124Sdim cnt.v_page_count; 1382309124Sdim if (pcount > tpcount) 1383284677Sdim pcount = tpcount; 1384309124Sdim } else { 1385309124Sdim fullintervalcount = 0; 1386309124Sdim } 1387314564Sdim 1388284677Sdim m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1389314564Sdim while ((m != NULL) && (pcount-- > 0)) { 1390314564Sdim int actcount; 1391314564Sdim 1392284677Sdim KASSERT(m->queue == PQ_ACTIVE, 1393314564Sdim ("vm_pageout_page_stats: page %p isn't active", m)); 1394314564Sdim 1395314564Sdim next = TAILQ_NEXT(m, pageq); 1396314564Sdim if ((m->flags & PG_MARKER) != 0) { 1397284677Sdim m = next; 1398314564Sdim continue; 1399314564Sdim } 1400314564Sdim vm_page_lock_assert(m, MA_NOTOWNED); 1401314564Sdim if (!vm_pageout_page_lock(m, &next)) { 1402314564Sdim vm_page_unlock(m); 1403314564Sdim m = next; 1404314564Sdim continue; 1405327952Sdim } 1406327952Sdim object = m->object; 1407327952Sdim if (!VM_OBJECT_TRYLOCK(object) && 1408314564Sdim !vm_pageout_fallback_object_lock(m, &next)) { 1409309124Sdim VM_OBJECT_UNLOCK(object); 1410309124Sdim vm_page_unlock(m); 1411314564Sdim m = next; 1412309124Sdim continue; 1413309124Sdim } 1414309124Sdim 1415309124Sdim /* 1416309124Sdim * Don't deactivate pages that are busy. 1417309124Sdim */ 1418314564Sdim if ((m->busy != 0) || 1419309124Sdim (m->oflags & VPO_BUSY) || 1420309124Sdim (m->hold_count != 0)) { 1421309124Sdim vm_page_unlock(m); 1422309124Sdim VM_OBJECT_UNLOCK(object); 1423314564Sdim vm_page_requeue(m); 1424309124Sdim m = next; 1425314564Sdim continue; 1426314564Sdim } 1427314564Sdim 1428284677Sdim actcount = 0; 1429284677Sdim if (m->flags & PG_REFERENCED) { 1430309124Sdim vm_page_flag_clear(m, PG_REFERENCED); 1431314564Sdim actcount += 1; 1432314564Sdim } 1433309124Sdim 1434309124Sdim actcount += pmap_ts_referenced(m); 1435309124Sdim if (actcount) { 1436309124Sdim m->act_count += ACT_ADVANCE + actcount; 1437309124Sdim if (m->act_count > ACT_MAX) 1438309124Sdim m->act_count = ACT_MAX; 1439309124Sdim vm_page_requeue(m); 1440309124Sdim } else { 1441309124Sdim if (m->act_count == 0) { 1442309124Sdim /* 1443309124Sdim * We turn off page access, so that we have 1444321369Sdim * more accurate RSS stats. We don't do this 1445309124Sdim * in the normal page deactivation when the 1446309124Sdim * system is loaded VM wise, because the 1447309124Sdim * cost of the large number of page protect 1448309124Sdim * operations would be higher than the value 1449309124Sdim * of doing the operation. 1450309124Sdim */ 1451309124Sdim pmap_remove_all(m); 1452309124Sdim vm_page_deactivate(m); 1453309124Sdim } else { 1454321369Sdim m->act_count -= min(m->act_count, ACT_DECLINE); 1455321369Sdim vm_page_requeue(m); 1456314564Sdim } 1457314564Sdim } 1458284677Sdim vm_page_unlock(m); 1459284677Sdim VM_OBJECT_UNLOCK(object); 1460341825Sdim m = next; 1461327952Sdim } 1462284677Sdim vm_page_unlock_queues(); 1463284677Sdim} 1464284677Sdim 1465284677Sdim/* 1466341825Sdim * vm_pageout is the high level pageout daemon. 1467284677Sdim */ 1468341825Sdimstatic void 1469284677Sdimvm_pageout() 1470341825Sdim{ 1471284677Sdim int error, pass; 1472284677Sdim 1473284677Sdim /* 1474284677Sdim * Initialize some paging parameters. 1475284677Sdim */ 1476321369Sdim cnt.v_interrupt_free_min = 2; 1477284677Sdim if (cnt.v_page_count < 2000) 1478284677Sdim vm_pageout_page_count = 8; 1479284677Sdim 1480284677Sdim /* 1481284677Sdim * v_free_reserved needs to include enough for the largest 1482284677Sdim * swap pager structures plus enough for any pv_entry structs 1483284677Sdim * when paging. 1484284677Sdim */ 1485284677Sdim if (cnt.v_page_count > 1024) 1486284677Sdim cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1487284677Sdim else 1488284677Sdim cnt.v_free_min = 4; 1489284677Sdim cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1490284677Sdim cnt.v_interrupt_free_min; 1491284677Sdim cnt.v_free_reserved = vm_pageout_page_count + 1492284677Sdim cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1493284677Sdim cnt.v_free_severe = cnt.v_free_min / 2; 1494284677Sdim cnt.v_free_min += cnt.v_free_reserved; 1495284677Sdim cnt.v_free_severe += cnt.v_free_reserved; 1496284677Sdim 1497284677Sdim /* 1498284677Sdim * v_free_target and v_cache_min control pageout hysteresis. Note 1499284677Sdim * that these are more a measure of the VM cache queue hysteresis 1500284677Sdim * then the VM free queue. Specifically, v_free_target is the 1501284677Sdim * high water mark (free+cache pages). 1502309124Sdim * 1503309124Sdim * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1504309124Sdim * low water mark, while v_free_min is the stop. v_cache_min must 1505284677Sdim * be big enough to handle memory needs while the pageout daemon 1506284677Sdim * is signalled and run to free more pages. 1507284677Sdim */ 1508284677Sdim if (cnt.v_free_count > 6144) 1509284677Sdim cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1510284677Sdim else 1511284677Sdim cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1512321369Sdim 1513284677Sdim if (cnt.v_free_count > 2048) { 1514284677Sdim cnt.v_cache_min = cnt.v_free_target; 1515284677Sdim cnt.v_cache_max = 2 * cnt.v_cache_min; 1516314564Sdim cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1517314564Sdim } else { 1518314564Sdim cnt.v_cache_min = 0; 1519314564Sdim cnt.v_cache_max = 0; 1520314564Sdim cnt.v_inactive_target = cnt.v_free_count / 4; 1521314564Sdim } 1522284677Sdim if (cnt.v_inactive_target > cnt.v_free_count / 3) 1523314564Sdim cnt.v_inactive_target = cnt.v_free_count / 3; 1524284677Sdim 1525284677Sdim /* XXX does not really belong here */ 1526284677Sdim if (vm_page_max_wired == 0) 1527284677Sdim vm_page_max_wired = cnt.v_free_count / 3; 1528284677Sdim 1529284677Sdim if (vm_pageout_stats_max == 0) 1530284677Sdim vm_pageout_stats_max = cnt.v_free_target; 1531284677Sdim 1532284677Sdim /* 1533284677Sdim * Set interval in seconds for stats scan. 1534284677Sdim */ 1535309124Sdim if (vm_pageout_stats_interval == 0) 1536309124Sdim vm_pageout_stats_interval = 5; 1537309124Sdim if (vm_pageout_full_stats_interval == 0) 1538341825Sdim vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1539309124Sdim 1540309124Sdim swap_pager_swap_init(); 1541309124Sdim pass = 0; 1542309124Sdim /* 1543309124Sdim * The pageout daemon is never done, so loop forever. 1544309124Sdim */ 1545309124Sdim while (TRUE) { 1546309124Sdim /* 1547309124Sdim * If we have enough free memory, wakeup waiters. Do 1548309124Sdim * not clear vm_pages_needed until we reach our target, 1549309124Sdim * otherwise we may be woken up over and over again and 1550341825Sdim * waste a lot of cpu. 1551341825Sdim */ 1552341825Sdim mtx_lock(&vm_page_queue_free_mtx); 1553341825Sdim if (vm_pages_needed && !vm_page_count_min()) { 1554341825Sdim if (!vm_paging_needed()) 1555341825Sdim vm_pages_needed = 0; 1556341825Sdim wakeup(&cnt.v_free_count); 1557341825Sdim } 1558341825Sdim if (vm_pages_needed) { 1559341825Sdim /* 1560341825Sdim * Still not done, take a second pass without waiting 1561341825Sdim * (unlimited dirty cleaning), otherwise sleep a bit 1562341825Sdim * and try again. 1563341825Sdim */ 1564341825Sdim ++pass; 1565341825Sdim if (pass > 1) 1566341825Sdim msleep(&vm_pages_needed, 1567341825Sdim &vm_page_queue_free_mtx, PVM, "psleep", 1568341825Sdim hz / 2); 1569341825Sdim } else { 1570341825Sdim /* 1571341825Sdim * Good enough, sleep & handle stats. Prime the pass 1572284677Sdim * for the next run. 1573284677Sdim */ 1574284677Sdim if (pass > 1) 1575284677Sdim pass = 1; 1576309124Sdim else 1577309124Sdim pass = 0; 1578309124Sdim error = msleep(&vm_pages_needed, 1579284677Sdim &vm_page_queue_free_mtx, PVM, "psleep", 1580284677Sdim vm_pageout_stats_interval * hz); 1581284677Sdim if (error && !vm_pages_needed) { 1582284677Sdim mtx_unlock(&vm_page_queue_free_mtx); 1583284677Sdim pass = 0; 1584284677Sdim vm_pageout_page_stats(); 1585314564Sdim continue; 1586321369Sdim } 1587314564Sdim } 1588314564Sdim if (vm_pages_needed) 1589314564Sdim cnt.v_pdwakeups++; 1590284677Sdim mtx_unlock(&vm_page_queue_free_mtx); 1591284677Sdim vm_pageout_scan(pass); 1592284677Sdim } 1593284677Sdim} 1594284677Sdim 1595284677Sdim/* 1596284677Sdim * Unless the free page queue lock is held by the caller, this function 1597284677Sdim * should be regarded as advisory. Specifically, the caller should 1598284677Sdim * not msleep() on &cnt.v_free_count following this function unless 1599284677Sdim * the free page queue lock is held until the msleep() is performed. 1600284677Sdim */ 1601309124Sdimvoid 1602341825Sdimpagedaemon_wakeup() 1603284677Sdim{ 1604284677Sdim 1605284677Sdim if (!vm_pages_needed && curthread->td_proc != pageproc) { 1606284677Sdim vm_pages_needed = 1; 1607284677Sdim wakeup(&vm_pages_needed); 1608284677Sdim } 1609341825Sdim} 1610284677Sdim 1611284677Sdim#if !defined(NO_SWAPPING) 1612284677Sdimstatic void 1613284677Sdimvm_req_vmdaemon(int req) 1614284677Sdim{ 1615284677Sdim static int lastrun = 0; 1616284677Sdim 1617284677Sdim mtx_lock(&vm_daemon_mtx); 1618284677Sdim vm_pageout_req_swapout |= req; 1619284677Sdim if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1620284677Sdim wakeup(&vm_daemon_needed); 1621284677Sdim lastrun = ticks; 1622284677Sdim } 1623284677Sdim mtx_unlock(&vm_daemon_mtx); 1624284677Sdim} 1625284677Sdim 1626284677Sdimstatic void 1627284677Sdimvm_daemon() 1628284677Sdim{ 1629284677Sdim struct rlimit rsslim; 1630284677Sdim struct proc *p; 1631284677Sdim struct thread *td; 1632284677Sdim struct vmspace *vm; 1633341825Sdim int breakout, swapout_flags; 1634284677Sdim 1635284677Sdim while (TRUE) { 1636309124Sdim mtx_lock(&vm_daemon_mtx); 1637309124Sdim msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1638341825Sdim swapout_flags = vm_pageout_req_swapout; 1639341825Sdim vm_pageout_req_swapout = 0; 1640341825Sdim mtx_unlock(&vm_daemon_mtx); 1641314564Sdim if (swapout_flags) 1642314564Sdim swapout_procs(swapout_flags); 1643284677Sdim 1644284677Sdim /* 1645284677Sdim * scan the processes for exceeding their rlimits or if 1646284677Sdim * process is swapped out -- deactivate pages 1647284677Sdim */ 1648284677Sdim sx_slock(&allproc_lock); 1649286684Sdim FOREACH_PROC_IN_SYSTEM(p) { 1650286684Sdim vm_pindex_t limit, size; 1651284677Sdim 1652284677Sdim /* 1653284677Sdim * if this is a system process or if we have already 1654284677Sdim * looked at this process, skip it. 1655284677Sdim */ 1656321369Sdim PROC_LOCK(p); 1657321369Sdim if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1658321369Sdim PROC_UNLOCK(p); 1659321369Sdim continue; 1660321369Sdim } 1661321369Sdim /* 1662321369Sdim * if the process is in a non-running type state, 1663321369Sdim * don't touch it. 1664321369Sdim */ 1665309124Sdim breakout = 0; 1666309124Sdim FOREACH_THREAD_IN_PROC(p, td) { 1667309124Sdim thread_lock(td); 1668309124Sdim if (!TD_ON_RUNQ(td) && 1669309124Sdim !TD_IS_RUNNING(td) && 1670309124Sdim !TD_IS_SLEEPING(td)) { 1671309124Sdim thread_unlock(td); 1672309124Sdim breakout = 1; 1673309124Sdim break; 1674309124Sdim } 1675309124Sdim thread_unlock(td); 1676309124Sdim } 1677309124Sdim if (breakout) { 1678309124Sdim PROC_UNLOCK(p); 1679309124Sdim continue; 1680309124Sdim } 1681309124Sdim /* 1682309124Sdim * get a limit 1683309124Sdim */ 1684309124Sdim lim_rlimit(p, RLIMIT_RSS, &rsslim); 1685284677Sdim limit = OFF_TO_IDX( 1686284677Sdim qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1687284677Sdim 1688284677Sdim /* 1689284677Sdim * let processes that are swapped out really be 1690284677Sdim * swapped out set the limit to nothing (will force a 1691284677Sdim * swap-out.) 1692284677Sdim */ 1693284677Sdim if ((p->p_flag & P_INMEM) == 0) 1694284677Sdim limit = 0; /* XXX */ 1695284677Sdim vm = vmspace_acquire_ref(p); 1696284677Sdim PROC_UNLOCK(p); 1697284677Sdim if (vm == NULL) 1698309124Sdim continue; 1699284677Sdim 1700284677Sdim size = vmspace_resident_count(vm); 1701284677Sdim if (limit >= 0 && size >= limit) { 1702284677Sdim vm_pageout_map_deactivate_pages( 1703284677Sdim &vm->vm_map, limit); 1704284677Sdim } 1705284677Sdim vmspace_free(vm); 1706284677Sdim } 1707284677Sdim sx_sunlock(&allproc_lock); 1708284677Sdim } 1709284677Sdim} 1710284677Sdim#endif /* !defined(NO_SWAPPING) */ 1711284677Sdim