sys/vm/vm_pageout.c

   1 /*-
   2  * Copyright (c) 1991 Regents of the University of California.
   3  * All rights reserved.
   4  * Copyright (c) 1994 John S. Dyson
   5  * All rights reserved.
   6  * Copyright (c) 1994 David Greenman
   7  * All rights reserved.
   8  * Copyright (c) 2005 Yahoo! Technologies Norway AS
   9  * All rights reserved.
  10  *
  11  * This code is derived from software contributed to Berkeley by
  12  * The Mach Operating System project at Carnegie-Mellon University.
  13  *
  14  * Redistribution and use in source and binary forms, with or without
  15  * modification, are permitted provided that the following conditions
  16  * are met:
  17  * 1. Redistributions of source code must retain the above copyright
  18  *    notice, this list of conditions and the following disclaimer.
  19  * 2. Redistributions in binary form must reproduce the above copyright
  20  *    notice, this list of conditions and the following disclaimer in the
  21  *    documentation and/or other materials provided with the distribution.
  22  * 3. All advertising materials mentioning features or use of this software
  23  *    must display the following acknowledgement:
  24  *      This product includes software developed by the University of
  25  *      California, Berkeley and its contributors.
  26  * 4. Neither the name of the University nor the names of its contributors
  27  *    may be used to endorse or promote products derived from this software
  28  *    without specific prior written permission.
  29  *
  30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  40  * SUCH DAMAGE.
  41  *
  42  *      from: @(#)vm_pageout.c  7.4 (Berkeley) 5/7/91
  43  *
  44  *
  45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  46  * All rights reserved.
  47  *
  48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  49  *
  50  * Permission to use, copy, modify and distribute this software and
  51  * its documentation is hereby granted, provided that both the copyright
  52  * notice and this permission notice appear in all copies of the
  53  * software, derivative works or modified versions, and any portions
  54  * thereof, and that both notices appear in supporting documentation.
  55  *
  56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  59  *
  60  * Carnegie Mellon requests users of this software to return to
  61  *
  62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  63  *  School of Computer Science
  64  *  Carnegie Mellon University
  65  *  Pittsburgh PA 15213-3890
  66  *
  67  * any improvements or extensions that they make and grant Carnegie the
  68  * rights to redistribute these changes.
  69  */
  70
  71 /*
  72  *      The proverbial page-out daemon.
  73  */
  74
  75 #include <sys/cdefs.h>
  76 __FBSDID("$FreeBSD$");
  77
  78 #include "opt_vm.h"
  79 #include "opt_kdtrace.h"
  80 #include <sys/param.h>
  81 #include <sys/systm.h>
  82 #include <sys/kernel.h>
  83 #include <sys/eventhandler.h>
  84 #include <sys/lock.h>
  85 #include <sys/mutex.h>
  86 #include <sys/proc.h>
  87 #include <sys/kthread.h>
  88 #include <sys/ktr.h>
  89 #include <sys/mount.h>
  90 #include <sys/racct.h>
  91 #include <sys/resourcevar.h>
  92 #include <sys/sched.h>
  93 #include <sys/sdt.h>
  94 #include <sys/signalvar.h>
  95 #include <sys/smp.h>
  96 #include <sys/time.h>
  97 #include <sys/vnode.h>
  98 #include <sys/vmmeter.h>
  99 #include <sys/rwlock.h>
 100 #include <sys/sx.h>
 101 #include <sys/sysctl.h>
 102
 103 #include <vm/vm.h>
 104 #include <vm/vm_param.h>
 105 #include <vm/vm_object.h>
 106 #include <vm/vm_page.h>
 107 #include <vm/vm_map.h>
 108 #include <vm/vm_pageout.h>
 109 #include <vm/vm_pager.h>
 110 #include <vm/vm_phys.h>
 111 #include <vm/swap_pager.h>
 112 #include <vm/vm_extern.h>
 113 #include <vm/uma.h>
 114
 115 /*
 116  * System initialization
 117  */
 118
 119 /* the kernel process "vm_pageout"*/
 120 static void vm_pageout(void);
 121 static void vm_pageout_init(void);
 122 static int vm_pageout_clean(vm_page_t);
 123 static void vm_pageout_scan(struct vm_domain *vmd, int pass);
 124 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
 125     int starting_page_shortage);
 126
 127 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
 128     NULL);
 129
 130 struct proc *pageproc;
 131
 132 static struct kproc_desc page_kp = {
 133         "pagedaemon",
 134         vm_pageout,
 135         &pageproc
 136 };
 137 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
 138     &page_kp);
 139
 140 SDT_PROVIDER_DEFINE(vm);
 141 SDT_PROBE_DEFINE(vm, , , vm__lowmem_cache);
 142 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
 143
 144 #if !defined(NO_SWAPPING)
 145 /* the kernel process "vm_daemon"*/
 146 static void vm_daemon(void);
 147 static struct   proc *vmproc;
 148
 149 static struct kproc_desc vm_kp = {
 150         "vmdaemon",
 151         vm_daemon,
 152         &vmproc
 153 };
 154 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
 155 #endif
 156
 157
 158 int vm_pages_needed;            /* Event on which pageout daemon sleeps */
 159 int vm_pageout_deficit;         /* Estimated number of pages deficit */
 160 int vm_pageout_pages_needed;    /* flag saying that the pageout daemon needs pages */
 161 int vm_pageout_wakeup_thresh;
 162 static int vm_pageout_oom_seq = 12;
 163
 164 #if !defined(NO_SWAPPING)
 165 static int vm_pageout_req_swapout;      /* XXX */
 166 static int vm_daemon_needed;
 167 static struct mtx vm_daemon_mtx;
 168 /* Allow for use by vm_pageout before vm_daemon is initialized. */
 169 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
 170 #endif
 171 static int vm_max_launder = 32;
 172 static int vm_pageout_update_period;
 173 static int defer_swap_pageouts;
 174 static int disable_swap_pageouts;
 175 static int lowmem_period = 10;
 176 static time_t lowmem_uptime;
 177
 178 #if defined(NO_SWAPPING)
 179 static int vm_swap_enabled = 0;
 180 static int vm_swap_idle_enabled = 0;
 181 #else
 182 static int vm_swap_enabled = 1;
 183 static int vm_swap_idle_enabled = 0;
 184 #endif
 185
 186 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
 187         CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0,
 188         "free page threshold for waking up the pageout daemon");
 189
 190 SYSCTL_INT(_vm, OID_AUTO, max_launder,
 191         CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
 192
 193 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
 194         CTLFLAG_RW, &vm_pageout_update_period, 0,
 195         "Maximum active LRU update period");
 196
 197 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0,
 198         "Low memory callback period");
 199
 200 #if defined(NO_SWAPPING)
 201 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
 202         CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
 203 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
 204         CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
 205 #else
 206 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
 207         CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
 208 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
 209         CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
 210 #endif
 211
 212 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
 213         CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
 214
 215 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
 216         CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
 217
 218 static int pageout_lock_miss;
 219 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
 220         CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
 221
 222 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
 223         CTLFLAG_RW, &vm_pageout_oom_seq, 0,
 224         "back-to-back calls to oom detector to start OOM");
 225
 226 #define VM_PAGEOUT_PAGE_COUNT 16
 227 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
 228
 229 int vm_page_max_wired;          /* XXX max # of wired pages system-wide */
 230 SYSCTL_INT(_vm, OID_AUTO, max_wired,
 231         CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
 232
 233 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
 234 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t,
 235     vm_paddr_t);
 236 #if !defined(NO_SWAPPING)
 237 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
 238 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
 239 static void vm_req_vmdaemon(int req);
 240 #endif
 241 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
 242
 243 /*
 244  * Initialize a dummy page for marking the caller's place in the specified
 245  * paging queue.  In principle, this function only needs to set the flag
 246  * PG_MARKER.  Nonetheless, it wirte busies and initializes the hold count
 247  * to one as safety precautions.
 248  */
 249 static void
 250 vm_pageout_init_marker(vm_page_t marker, u_short queue)
 251 {
 252
 253         bzero(marker, sizeof(*marker));
 254         marker->flags = PG_MARKER;
 255         marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
 256         marker->queue = queue;
 257         marker->hold_count = 1;
 258 }
 259
 260 /*
 261  * vm_pageout_fallback_object_lock:
 262  *
 263  * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
 264  * known to have failed and page queue must be either PQ_ACTIVE or
 265  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
 266  * while locking the vm object.  Use marker page to detect page queue
 267  * changes and maintain notion of next page on page queue.  Return
 268  * TRUE if no changes were detected, FALSE otherwise.  vm object is
 269  * locked on return.
 270  *
 271  * This function depends on both the lock portion of struct vm_object
 272  * and normal struct vm_page being type stable.
 273  */
 274 static boolean_t
 275 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
 276 {
 277         struct vm_page marker;
 278         struct vm_pagequeue *pq;
 279         boolean_t unchanged;
 280         u_short queue;
 281         vm_object_t object;
 282
 283         queue = m->queue;
 284         vm_pageout_init_marker(&marker, queue);
 285         pq = vm_page_pagequeue(m);
 286         object = m->object;
 287
 288         TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
 289         vm_pagequeue_unlock(pq);
 290         vm_page_unlock(m);
 291         VM_OBJECT_WLOCK(object);
 292         vm_page_lock(m);
 293         vm_pagequeue_lock(pq);
 294
 295         /*
 296          * The page's object might have changed, and/or the page might
 297          * have moved from its original position in the queue.  If the
 298          * page's object has changed, then the caller should abandon
 299          * processing the page because the wrong object lock was
 300          * acquired.  Use the marker's plinks.q, not the page's, to
 301          * determine if the page has been moved.  The state of the
 302          * page's plinks.q can be indeterminate; whereas, the marker's
 303          * plinks.q must be valid.
 304          */
 305         *next = TAILQ_NEXT(&marker, plinks.q);
 306         unchanged = m->object == object &&
 307             m == TAILQ_PREV(&marker, pglist, plinks.q);
 308         KASSERT(!unchanged || m->queue == queue,
 309             ("page %p queue %d %d", m, queue, m->queue));
 310         TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
 311         return (unchanged);
 312 }
 313
 314 /*
 315  * Lock the page while holding the page queue lock.  Use marker page
 316  * to detect page queue changes and maintain notion of next page on
 317  * page queue.  Return TRUE if no changes were detected, FALSE
 318  * otherwise.  The page is locked on return. The page queue lock might
 319  * be dropped and reacquired.
 320  *
 321  * This function depends on normal struct vm_page being type stable.
 322  */
 323 static boolean_t
 324 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
 325 {
 326         struct vm_page marker;
 327         struct vm_pagequeue *pq;
 328         boolean_t unchanged;
 329         u_short queue;
 330
 331         vm_page_lock_assert(m, MA_NOTOWNED);
 332         if (vm_page_trylock(m))
 333                 return (TRUE);
 334
 335         queue = m->queue;
 336         vm_pageout_init_marker(&marker, queue);
 337         pq = vm_page_pagequeue(m);
 338
 339         TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
 340         vm_pagequeue_unlock(pq);
 341         vm_page_lock(m);
 342         vm_pagequeue_lock(pq);
 343
 344         /* Page queue might have changed. */
 345         *next = TAILQ_NEXT(&marker, plinks.q);
 346         unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
 347         KASSERT(!unchanged || m->queue == queue,
 348             ("page %p queue %d %d", m, queue, m->queue));
 349         TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
 350         return (unchanged);
 351 }
 352
 353 /*
 354  * vm_pageout_clean:
 355  *
 356  * Clean the page and remove it from the laundry.
 357  *
 358  * We set the busy bit to cause potential page faults on this page to
 359  * block.  Note the careful timing, however, the busy bit isn't set till
 360  * late and we cannot do anything that will mess with the page.
 361  */
 362 static int
 363 vm_pageout_clean(vm_page_t m)
 364 {
 365         vm_object_t object;
 366         vm_page_t mc[2*vm_pageout_page_count], pb, ps;
 367         int pageout_count;
 368         int ib, is, page_base;
 369         vm_pindex_t pindex = m->pindex;
 370
 371         vm_page_lock_assert(m, MA_OWNED);
 372         object = m->object;
 373         VM_OBJECT_ASSERT_WLOCKED(object);
 374
 375         /*
 376          * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
 377          * with the new swapper, but we could have serious problems paging
 378          * out other object types if there is insufficient memory.
 379          *
 380          * Unfortunately, checking free memory here is far too late, so the
 381          * check has been moved up a procedural level.
 382          */
 383
 384         /*
 385          * Can't clean the page if it's busy or held.
 386          */
 387         vm_page_assert_unbusied(m);
 388         KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
 389         vm_page_unlock(m);
 390
 391         mc[vm_pageout_page_count] = pb = ps = m;
 392         pageout_count = 1;
 393         page_base = vm_pageout_page_count;
 394         ib = 1;
 395         is = 1;
 396
 397         /*
 398          * Scan object for clusterable pages.
 399          *
 400          * We can cluster ONLY if: ->> the page is NOT
 401          * clean, wired, busy, held, or mapped into a
 402          * buffer, and one of the following:
 403          * 1) The page is inactive, or a seldom used
 404          *    active page.
 405          * -or-
 406          * 2) we force the issue.
 407          *
 408          * During heavy mmap/modification loads the pageout
 409          * daemon can really fragment the underlying file
 410          * due to flushing pages out of order and not trying
 411          * align the clusters (which leave sporatic out-of-order
 412          * holes).  To solve this problem we do the reverse scan
 413          * first and attempt to align our cluster, then do a
 414          * forward scan if room remains.
 415          */
 416 more:
 417         while (ib && pageout_count < vm_pageout_page_count) {
 418                 vm_page_t p;
 419
 420                 if (ib > pindex) {
 421                         ib = 0;
 422                         break;
 423                 }
 424
 425                 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
 426                         ib = 0;
 427                         break;
 428                 }
 429                 vm_page_test_dirty(p);
 430                 if (p->dirty == 0) {
 431                         ib = 0;
 432                         break;
 433                 }
 434                 vm_page_lock(p);
 435                 if (p->queue != PQ_INACTIVE ||
 436                     p->hold_count != 0) {       /* may be undergoing I/O */
 437                         vm_page_unlock(p);
 438                         ib = 0;
 439                         break;
 440                 }
 441                 vm_page_unlock(p);
 442                 mc[--page_base] = pb = p;
 443                 ++pageout_count;
 444                 ++ib;
 445                 /*
 446                  * alignment boundry, stop here and switch directions.  Do
 447                  * not clear ib.
 448                  */
 449                 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
 450                         break;
 451         }
 452
 453         while (pageout_count < vm_pageout_page_count &&
 454             pindex + is < object->size) {
 455                 vm_page_t p;
 456
 457                 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
 458                         break;
 459                 vm_page_test_dirty(p);
 460                 if (p->dirty == 0)
 461                         break;
 462                 vm_page_lock(p);
 463                 if (p->queue != PQ_INACTIVE ||
 464                     p->hold_count != 0) {       /* may be undergoing I/O */
 465                         vm_page_unlock(p);
 466                         break;
 467                 }
 468                 vm_page_unlock(p);
 469                 mc[page_base + pageout_count] = ps = p;
 470                 ++pageout_count;
 471                 ++is;
 472         }
 473
 474         /*
 475          * If we exhausted our forward scan, continue with the reverse scan
 476          * when possible, even past a page boundry.  This catches boundry
 477          * conditions.
 478          */
 479         if (ib && pageout_count < vm_pageout_page_count)
 480                 goto more;
 481
 482         /*
 483          * we allow reads during pageouts...
 484          */
 485         return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
 486             NULL));
 487 }
 488
 489 /*
 490  * vm_pageout_flush() - launder the given pages
 491  *
 492  *      The given pages are laundered.  Note that we setup for the start of
 493  *      I/O ( i.e. busy the page ), mark it read-only, and bump the object
 494  *      reference count all in here rather then in the parent.  If we want
 495  *      the parent to do more sophisticated things we may have to change
 496  *      the ordering.
 497  *
 498  *      Returned runlen is the count of pages between mreq and first
 499  *      page after mreq with status VM_PAGER_AGAIN.
 500  *      *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
 501  *      for any page in runlen set.
 502  */
 503 int
 504 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
 505     boolean_t *eio)
 506 {
 507         vm_object_t object = mc[0]->object;
 508         int pageout_status[count];
 509         int numpagedout = 0;
 510         int i, runlen;
 511
 512         VM_OBJECT_ASSERT_WLOCKED(object);
 513
 514         /*
 515          * Initiate I/O.  Bump the vm_page_t->busy counter and
 516          * mark the pages read-only.
 517          *
 518          * We do not have to fixup the clean/dirty bits here... we can
 519          * allow the pager to do it after the I/O completes.
 520          *
 521          * NOTE! mc[i]->dirty may be partial or fragmented due to an
 522          * edge case with file fragments.
 523          */
 524         for (i = 0; i < count; i++) {
 525                 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
 526                     ("vm_pageout_flush: partially invalid page %p index %d/%d",
 527                         mc[i], i, count));
 528                 vm_page_sbusy(mc[i]);
 529                 pmap_remove_write(mc[i]);
 530         }
 531         vm_object_pip_add(object, count);
 532
 533         vm_pager_put_pages(object, mc, count, flags, pageout_status);
 534
 535         runlen = count - mreq;
 536         if (eio != NULL)
 537                 *eio = FALSE;
 538         for (i = 0; i < count; i++) {
 539                 vm_page_t mt = mc[i];
 540
 541                 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
 542                     !pmap_page_is_write_mapped(mt),
 543                     ("vm_pageout_flush: page %p is not write protected", mt));
 544                 switch (pageout_status[i]) {
 545                 case VM_PAGER_OK:
 546                 case VM_PAGER_PEND:
 547                         numpagedout++;
 548                         break;
 549                 case VM_PAGER_BAD:
 550                         /*
 551                          * Page outside of range of object. Right now we
 552                          * essentially lose the changes by pretending it
 553                          * worked.
 554                          */
 555                         vm_page_undirty(mt);
 556                         break;
 557                 case VM_PAGER_ERROR:
 558                 case VM_PAGER_FAIL:
 559                         /*
 560                          * If page couldn't be paged out, then reactivate the
 561                          * page so it doesn't clog the inactive list.  (We
 562                          * will try paging out it again later).
 563                          */
 564                         vm_page_lock(mt);
 565                         vm_page_activate(mt);
 566                         vm_page_unlock(mt);
 567                         if (eio != NULL && i >= mreq && i - mreq < runlen)
 568                                 *eio = TRUE;
 569                         break;
 570                 case VM_PAGER_AGAIN:
 571                         if (i >= mreq && i - mreq < runlen)
 572                                 runlen = i - mreq;
 573                         break;
 574                 }
 575
 576                 /*
 577                  * If the operation is still going, leave the page busy to
 578                  * block all other accesses. Also, leave the paging in
 579                  * progress indicator set so that we don't attempt an object
 580                  * collapse.
 581                  */
 582                 if (pageout_status[i] != VM_PAGER_PEND) {
 583                         vm_object_pip_wakeup(object);
 584                         vm_page_sunbusy(mt);
 585                         if (vm_page_count_severe()) {
 586                                 vm_page_lock(mt);
 587                                 vm_page_try_to_cache(mt);
 588                                 vm_page_unlock(mt);
 589                         }
 590                 }
 591         }
 592         if (prunlen != NULL)
 593                 *prunlen = runlen;
 594         return (numpagedout);
 595 }
 596
 597 static boolean_t
 598 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low,
 599     vm_paddr_t high)
 600 {
 601         struct mount *mp;
 602         struct vnode *vp;
 603         vm_object_t object;
 604         vm_paddr_t pa;
 605         vm_page_t m, m_tmp, next;
 606         int lockmode;
 607
 608         vm_pagequeue_lock(pq);
 609         TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) {
 610                 if ((m->flags & PG_MARKER) != 0)
 611                         continue;
 612                 pa = VM_PAGE_TO_PHYS(m);
 613                 if (pa < low || pa + PAGE_SIZE > high)
 614                         continue;
 615                 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
 616                         vm_page_unlock(m);
 617                         continue;
 618                 }
 619                 object = m->object;
 620                 if ((!VM_OBJECT_TRYWLOCK(object) &&
 621                     (!vm_pageout_fallback_object_lock(m, &next) ||
 622                     m->hold_count != 0)) || vm_page_busied(m)) {
 623                         vm_page_unlock(m);
 624                         VM_OBJECT_WUNLOCK(object);
 625                         continue;
 626                 }
 627                 vm_page_test_dirty(m);
 628                 if (m->dirty == 0 && object->ref_count != 0)
 629                         pmap_remove_all(m);
 630                 if (m->dirty != 0) {
 631                         vm_page_unlock(m);
 632                         if (tries == 0 || (object->flags & OBJ_DEAD) != 0) {
 633                                 VM_OBJECT_WUNLOCK(object);
 634                                 continue;
 635                         }
 636                         if (object->type == OBJT_VNODE) {
 637                                 vm_pagequeue_unlock(pq);
 638                                 vp = object->handle;
 639                                 vm_object_reference_locked(object);
 640                                 VM_OBJECT_WUNLOCK(object);
 641                                 (void)vn_start_write(vp, &mp, V_WAIT);
 642                                 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
 643                                     LK_SHARED : LK_EXCLUSIVE;
 644                                 vn_lock(vp, lockmode | LK_RETRY);
 645                                 VM_OBJECT_WLOCK(object);
 646                                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 647                                 VM_OBJECT_WUNLOCK(object);
 648                                 VOP_UNLOCK(vp, 0);
 649                                 vm_object_deallocate(object);
 650                                 vn_finished_write(mp);
 651                                 return (TRUE);
 652                         } else if (object->type == OBJT_SWAP ||
 653                             object->type == OBJT_DEFAULT) {
 654                                 vm_pagequeue_unlock(pq);
 655                                 m_tmp = m;
 656                                 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC,
 657                                     0, NULL, NULL);
 658                                 VM_OBJECT_WUNLOCK(object);
 659                                 return (TRUE);
 660                         }
 661                 } else {
 662                         /*
 663                          * Dequeue here to prevent lock recursion in
 664                          * vm_page_cache().
 665                          */
 666                         vm_page_dequeue_locked(m);
 667                         vm_page_cache(m);
 668                         vm_page_unlock(m);
 669                 }
 670                 VM_OBJECT_WUNLOCK(object);
 671         }
 672         vm_pagequeue_unlock(pq);
 673         return (FALSE);
 674 }
 675
 676 /*
 677  * Increase the number of cached pages.  The specified value, "tries",
 678  * determines which categories of pages are cached:
 679  *
 680  *  0: All clean, inactive pages within the specified physical address range
 681  *     are cached.  Will not sleep.
 682  *  1: The vm_lowmem handlers are called.  All inactive pages within
 683  *     the specified physical address range are cached.  May sleep.
 684  *  2: The vm_lowmem handlers are called.  All inactive and active pages
 685  *     within the specified physical address range are cached.  May sleep.
 686  */
 687 void
 688 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high)
 689 {
 690         int actl, actmax, inactl, inactmax, dom, initial_dom;
 691         static int start_dom = 0;
 692
 693         if (tries > 0) {
 694                 /*
 695                  * Decrease registered cache sizes.  The vm_lowmem handlers
 696                  * may acquire locks and/or sleep, so they can only be invoked
 697                  * when "tries" is greater than zero.
 698                  */
 699                 SDT_PROBE0(vm, , , vm__lowmem_cache);
 700                 EVENTHANDLER_INVOKE(vm_lowmem, 0);
 701
 702                 /*
 703                  * We do this explicitly after the caches have been drained
 704                  * above.
 705                  */
 706                 uma_reclaim();
 707         }
 708
 709         /*
 710          * Make the next scan start on the next domain.
 711          */
 712         initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains;
 713
 714         inactl = 0;
 715         inactmax = cnt.v_inactive_count;
 716         actl = 0;
 717         actmax = tries < 2 ? 0 : cnt.v_active_count;
 718         dom = initial_dom;
 719
 720         /*
 721          * Scan domains in round-robin order, first inactive queues,
 722          * then active.  Since domain usually owns large physically
 723          * contiguous chunk of memory, it makes sense to completely
 724          * exhaust one domain before switching to next, while growing
 725          * the pool of contiguous physical pages.
 726          *
 727          * Do not even start launder a domain which cannot contain
 728          * the specified address range, as indicated by segments
 729          * constituting the domain.
 730          */
 731 again_inact:
 732         if (inactl < inactmax) {
 733                 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
 734                     low, high) &&
 735                     vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE],
 736                     tries, low, high)) {
 737                         inactl++;
 738                         goto again_inact;
 739                 }
 740                 if (++dom == vm_ndomains)
 741                         dom = 0;
 742                 if (dom != initial_dom)
 743                         goto again_inact;
 744         }
 745 again_act:
 746         if (actl < actmax) {
 747                 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
 748                     low, high) &&
 749                     vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE],
 750                       tries, low, high)) {
 751                         actl++;
 752                         goto again_act;
 753                 }
 754                 if (++dom == vm_ndomains)
 755                         dom = 0;
 756                 if (dom != initial_dom)
 757                         goto again_act;
 758         }
 759 }
 760
 761 #if !defined(NO_SWAPPING)
 762 /*
 763  *      vm_pageout_object_deactivate_pages
 764  *
 765  *      Deactivate enough pages to satisfy the inactive target
 766  *      requirements.
 767  *
 768  *      The object and map must be locked.
 769  */
 770 static void
 771 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
 772     long desired)
 773 {
 774         vm_object_t backing_object, object;
 775         vm_page_t p;
 776         int act_delta, remove_mode;
 777
 778         VM_OBJECT_ASSERT_LOCKED(first_object);
 779         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 780                 return;
 781         for (object = first_object;; object = backing_object) {
 782                 if (pmap_resident_count(pmap) <= desired)
 783                         goto unlock_return;
 784                 VM_OBJECT_ASSERT_LOCKED(object);
 785                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 786                     object->paging_in_progress != 0)
 787                         goto unlock_return;
 788
 789                 remove_mode = 0;
 790                 if (object->shadow_count > 1)
 791                         remove_mode = 1;
 792                 /*
 793                  * Scan the object's entire memory queue.
 794                  */
 795                 TAILQ_FOREACH(p, &object->memq, listq) {
 796                         if (pmap_resident_count(pmap) <= desired)
 797                                 goto unlock_return;
 798                         if (vm_page_busied(p))
 799                                 continue;
 800                         PCPU_INC(cnt.v_pdpages);
 801                         vm_page_lock(p);
 802                         if (p->wire_count != 0 || p->hold_count != 0 ||
 803                             !pmap_page_exists_quick(pmap, p)) {
 804                                 vm_page_unlock(p);
 805                                 continue;
 806                         }
 807                         act_delta = pmap_ts_referenced(p);
 808                         if ((p->aflags & PGA_REFERENCED) != 0) {
 809                                 if (act_delta == 0)
 810                                         act_delta = 1;
 811                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 812                         }
 813                         if (p->queue != PQ_ACTIVE && act_delta != 0) {
 814                                 vm_page_activate(p);
 815                                 p->act_count += act_delta;
 816                         } else if (p->queue == PQ_ACTIVE) {
 817                                 if (act_delta == 0) {
 818                                         p->act_count -= min(p->act_count,
 819                                             ACT_DECLINE);
 820                                         if (!remove_mode && p->act_count == 0) {
 821                                                 pmap_remove_all(p);
 822                                                 vm_page_deactivate(p);
 823                                         } else
 824                                                 vm_page_requeue(p);
 825                                 } else {
 826                                         vm_page_activate(p);
 827                                         if (p->act_count < ACT_MAX -
 828                                             ACT_ADVANCE)
 829                                                 p->act_count += ACT_ADVANCE;
 830                                         vm_page_requeue(p);
 831                                 }
 832                         } else if (p->queue == PQ_INACTIVE)
 833                                 pmap_remove_all(p);
 834                         vm_page_unlock(p);
 835                 }
 836                 if ((backing_object = object->backing_object) == NULL)
 837                         goto unlock_return;
 838                 VM_OBJECT_RLOCK(backing_object);
 839                 if (object != first_object)
 840                         VM_OBJECT_RUNLOCK(object);
 841         }
 842 unlock_return:
 843         if (object != first_object)
 844                 VM_OBJECT_RUNLOCK(object);
 845 }
 846
 847 /*
 848  * deactivate some number of pages in a map, try to do it fairly, but
 849  * that is really hard to do.
 850  */
 851 static void
 852 vm_pageout_map_deactivate_pages(map, desired)
 853         vm_map_t map;
 854         long desired;
 855 {
 856         vm_map_entry_t tmpe;
 857         vm_object_t obj, bigobj;
 858         int nothingwired;
 859
 860         if (!vm_map_trylock(map))
 861                 return;
 862
 863         bigobj = NULL;
 864         nothingwired = TRUE;
 865
 866         /*
 867          * first, search out the biggest object, and try to free pages from
 868          * that.
 869          */
 870         tmpe = map->header.next;
 871         while (tmpe != &map->header) {
 872                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 873                         obj = tmpe->object.vm_object;
 874                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 875                                 if (obj->shadow_count <= 1 &&
 876                                     (bigobj == NULL ||
 877                                      bigobj->resident_page_count < obj->resident_page_count)) {
 878                                         if (bigobj != NULL)
 879                                                 VM_OBJECT_RUNLOCK(bigobj);
 880                                         bigobj = obj;
 881                                 } else
 882                                         VM_OBJECT_RUNLOCK(obj);
 883                         }
 884                 }
 885                 if (tmpe->wired_count > 0)
 886                         nothingwired = FALSE;
 887                 tmpe = tmpe->next;
 888         }
 889
 890         if (bigobj != NULL) {
 891                 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
 892                 VM_OBJECT_RUNLOCK(bigobj);
 893         }
 894         /*
 895          * Next, hunt around for other pages to deactivate.  We actually
 896          * do this search sort of wrong -- .text first is not the best idea.
 897          */
 898         tmpe = map->header.next;
 899         while (tmpe != &map->header) {
 900                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 901                         break;
 902                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 903                         obj = tmpe->object.vm_object;
 904                         if (obj != NULL) {
 905                                 VM_OBJECT_RLOCK(obj);
 906                                 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
 907                                 VM_OBJECT_RUNLOCK(obj);
 908                         }
 909                 }
 910                 tmpe = tmpe->next;
 911         }
 912
 913 #ifdef __ia64__
 914         /*
 915          * Remove all non-wired, managed mappings if a process is swapped out.
 916          * This will free page table pages.
 917          */
 918         if (desired == 0)
 919                 pmap_remove_pages(map->pmap);
 920 #else
 921         /*
 922          * Remove all mappings if a process is swapped out, this will free page
 923          * table pages.
 924          */
 925         if (desired == 0 && nothingwired) {
 926                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 927                     vm_map_max(map));
 928         }
 929 #endif
 930
 931         vm_map_unlock(map);
 932 }
 933 #endif          /* !defined(NO_SWAPPING) */
 934
 935 /*
 936  *      vm_pageout_scan does the dirty work for the pageout daemon.
 937  *
 938  *      pass 0 - Update active LRU/deactivate pages
 939  *      pass 1 - Move inactive to cache or free
 940  *      pass 2 - Launder dirty pages
 941  */
 942 static void
 943 vm_pageout_scan(struct vm_domain *vmd, int pass)
 944 {
 945         vm_page_t m, next;
 946         struct vm_pagequeue *pq;
 947         vm_object_t object;
 948         long min_scan;
 949         int act_delta, addl_page_shortage, deficit, maxscan, page_shortage;
 950         int vnodes_skipped = 0;
 951         int maxlaunder, scan_tick, scanned, starting_page_shortage;
 952         int lockmode;
 953         boolean_t queues_locked;
 954
 955         /*
 956          * If we need to reclaim memory ask kernel caches to return
 957          * some.  We rate limit to avoid thrashing.
 958          */
 959         if (vmd == &vm_dom[0] && pass > 0 &&
 960             (time_uptime - lowmem_uptime) >= lowmem_period) {
 961                 /*
 962                  * Decrease registered cache sizes.
 963                  */
 964                 SDT_PROBE0(vm, , , vm__lowmem_scan);
 965                 EVENTHANDLER_INVOKE(vm_lowmem, 0);
 966                 /*
 967                  * We do this explicitly after the caches have been
 968                  * drained above.
 969                  */
 970                 uma_reclaim();
 971                 lowmem_uptime = time_uptime;
 972         }
 973
 974         /*
 975          * The addl_page_shortage is the number of temporarily
 976          * stuck pages in the inactive queue.  In other words, the
 977          * number of pages from the inactive count that should be
 978          * discounted in setting the target for the active queue scan.
 979          */
 980         addl_page_shortage = 0;
 981
 982         /*
 983          * Calculate the number of pages we want to either free or move
 984          * to the cache.
 985          */
 986         if (pass > 0) {
 987                 deficit = atomic_readandclear_int(&vm_pageout_deficit);
 988                 page_shortage = vm_paging_target() + deficit;
 989         } else
 990                 page_shortage = deficit = 0;
 991         starting_page_shortage = page_shortage;
 992
 993         /*
 994          * maxlaunder limits the number of dirty pages we flush per scan.
 995          * For most systems a smaller value (16 or 32) is more robust under
 996          * extreme memory and disk pressure because any unnecessary writes
 997          * to disk can result in extreme performance degredation.  However,
 998          * systems with excessive dirty pages (especially when MAP_NOSYNC is
 999          * used) will die horribly with limited laundering.  If the pageout
1000          * daemon cannot clean enough pages in the first pass, we let it go
1001          * all out in succeeding passes.
1002          */
1003         if ((maxlaunder = vm_max_launder) <= 1)
1004                 maxlaunder = 1;
1005         if (pass > 1)
1006                 maxlaunder = 10000;
1007
1008         /*
1009          * Start scanning the inactive queue for pages we can move to the
1010          * cache or free.  The scan will stop when the target is reached or
1011          * we have scanned the entire inactive queue.  Note that m->act_count
1012          * is not used to form decisions for the inactive queue, only for the
1013          * active queue.
1014          */
1015         pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1016         maxscan = pq->pq_cnt;
1017         vm_pagequeue_lock(pq);
1018         queues_locked = TRUE;
1019         for (m = TAILQ_FIRST(&pq->pq_pl);
1020              m != NULL && maxscan-- > 0 && page_shortage > 0;
1021              m = next) {
1022                 vm_pagequeue_assert_locked(pq);
1023                 KASSERT(queues_locked, ("unlocked queues"));
1024                 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
1025
1026                 PCPU_INC(cnt.v_pdpages);
1027                 next = TAILQ_NEXT(m, plinks.q);
1028
1029                 /*
1030                  * skip marker pages
1031                  */
1032                 if (m->flags & PG_MARKER)
1033                         continue;
1034
1035                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1036                     ("Fictitious page %p cannot be in inactive queue", m));
1037                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1038                     ("Unmanaged page %p cannot be in inactive queue", m));
1039
1040                 /*
1041                  * The page or object lock acquisitions fail if the
1042                  * page was removed from the queue or moved to a
1043                  * different position within the queue.  In either
1044                  * case, addl_page_shortage should not be incremented.
1045                  */
1046                 if (!vm_pageout_page_lock(m, &next)) {
1047                         vm_page_unlock(m);
1048                         continue;
1049                 }
1050                 object = m->object;
1051                 if (!VM_OBJECT_TRYWLOCK(object) &&
1052                     !vm_pageout_fallback_object_lock(m, &next)) {
1053                         vm_page_unlock(m);
1054                         VM_OBJECT_WUNLOCK(object);
1055                         continue;
1056                 }
1057
1058                 /*
1059                  * Don't mess with busy pages, keep them at at the
1060                  * front of the queue, most likely they are being
1061                  * paged out.  Increment addl_page_shortage for busy
1062                  * pages, because they may leave the inactive queue
1063                  * shortly after page scan is finished.
1064                  */
1065                 if (vm_page_busied(m)) {
1066                         vm_page_unlock(m);
1067                         VM_OBJECT_WUNLOCK(object);
1068                         addl_page_shortage++;
1069                         continue;
1070                 }
1071
1072                 /*
1073                  * We unlock the inactive page queue, invalidating the
1074                  * 'next' pointer.  Use our marker to remember our
1075                  * place.
1076                  */
1077                 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1078                 vm_pagequeue_unlock(pq);
1079                 queues_locked = FALSE;
1080
1081                 /*
1082                  * We bump the activation count if the page has been
1083                  * referenced while in the inactive queue.  This makes
1084                  * it less likely that the page will be added back to the
1085                  * inactive queue prematurely again.  Here we check the
1086                  * page tables (or emulated bits, if any), given the upper
1087                  * level VM system not knowing anything about existing
1088                  * references.
1089                  */
1090                 act_delta = 0;
1091                 if ((m->aflags & PGA_REFERENCED) != 0) {
1092                         vm_page_aflag_clear(m, PGA_REFERENCED);
1093                         act_delta = 1;
1094                 }
1095                 if (object->ref_count != 0) {
1096                         act_delta += pmap_ts_referenced(m);
1097                 } else {
1098                         KASSERT(!pmap_page_is_mapped(m),
1099                             ("vm_pageout_scan: page %p is mapped", m));
1100                 }
1101
1102                 /*
1103                  * If the upper level VM system knows about any page
1104                  * references, we reactivate the page or requeue it.
1105                  */
1106                 if (act_delta != 0) {
1107                         if (object->ref_count) {
1108                                 vm_page_activate(m);
1109                                 m->act_count += act_delta + ACT_ADVANCE;
1110                         } else {
1111                                 vm_pagequeue_lock(pq);
1112                                 queues_locked = TRUE;
1113                                 vm_page_requeue_locked(m);
1114                         }
1115                         VM_OBJECT_WUNLOCK(object);
1116                         vm_page_unlock(m);
1117                         goto relock_queues;
1118                 }
1119
1120                 if (m->hold_count != 0) {
1121                         vm_page_unlock(m);
1122                         VM_OBJECT_WUNLOCK(object);
1123
1124                         /*
1125                          * Held pages are essentially stuck in the
1126                          * queue.  So, they ought to be discounted
1127                          * from the inactive count.  See the
1128                          * calculation of the page_shortage for the
1129                          * loop over the active queue below.
1130                          */
1131                         addl_page_shortage++;
1132                         goto relock_queues;
1133                 }
1134
1135                 /*
1136                  * If the page appears to be clean at the machine-independent
1137                  * layer, then remove all of its mappings from the pmap in
1138                  * anticipation of placing it onto the cache queue.  If,
1139                  * however, any of the page's mappings allow write access,
1140                  * then the page may still be modified until the last of those
1141                  * mappings are removed.
1142                  */
1143                 if (object->ref_count != 0) {
1144                         vm_page_test_dirty(m);
1145                         if (m->dirty == 0)
1146                                 pmap_remove_all(m);
1147                 }
1148
1149                 if (m->valid == 0) {
1150                         /*
1151                          * Invalid pages can be easily freed
1152                          */
1153                         vm_page_free(m);
1154                         PCPU_INC(cnt.v_dfree);
1155                         --page_shortage;
1156                 } else if (m->dirty == 0) {
1157                         /*
1158                          * Clean pages can be placed onto the cache queue.
1159                          * This effectively frees them.
1160                          */
1161                         vm_page_cache(m);
1162                         --page_shortage;
1163                 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) {
1164                         /*
1165                          * Dirty pages need to be paged out, but flushing
1166                          * a page is extremely expensive verses freeing
1167                          * a clean page.  Rather then artificially limiting
1168                          * the number of pages we can flush, we instead give
1169                          * dirty pages extra priority on the inactive queue
1170                          * by forcing them to be cycled through the queue
1171                          * twice before being flushed, after which the
1172                          * (now clean) page will cycle through once more
1173                          * before being freed.  This significantly extends
1174                          * the thrash point for a heavily loaded machine.
1175                          */
1176                         m->flags |= PG_WINATCFLS;
1177                         vm_pagequeue_lock(pq);
1178                         queues_locked = TRUE;
1179                         vm_page_requeue_locked(m);
1180                 } else if (maxlaunder > 0) {
1181                         /*
1182                          * We always want to try to flush some dirty pages if
1183                          * we encounter them, to keep the system stable.
1184                          * Normally this number is small, but under extreme
1185                          * pressure where there are insufficient clean pages
1186                          * on the inactive queue, we may have to go all out.
1187                          */
1188                         int swap_pageouts_ok;
1189                         struct vnode *vp = NULL;
1190                         struct mount *mp = NULL;
1191
1192                         if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
1193                                 swap_pageouts_ok = 1;
1194                         } else {
1195                                 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
1196                                 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
1197                                 vm_page_count_min());
1198
1199                         }
1200
1201                         /*
1202                          * We don't bother paging objects that are "dead".
1203                          * Those objects are in a "rundown" state.
1204                          */
1205                         if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
1206                                 vm_pagequeue_lock(pq);
1207                                 vm_page_unlock(m);
1208                                 VM_OBJECT_WUNLOCK(object);
1209                                 queues_locked = TRUE;
1210                                 vm_page_requeue_locked(m);
1211                                 goto relock_queues;
1212                         }
1213
1214                         /*
1215                          * The object is already known NOT to be dead.   It
1216                          * is possible for the vget() to block the whole
1217                          * pageout daemon, but the new low-memory handling
1218                          * code should prevent it.
1219                          *
1220                          * The previous code skipped locked vnodes and, worse,
1221                          * reordered pages in the queue.  This results in
1222                          * completely non-deterministic operation and, on a
1223                          * busy system, can lead to extremely non-optimal
1224                          * pageouts.  For example, it can cause clean pages
1225                          * to be freed and dirty pages to be moved to the end
1226                          * of the queue.  Since dirty pages are also moved to
1227                          * the end of the queue once-cleaned, this gives
1228                          * way too large a weighting to defering the freeing
1229                          * of dirty pages.
1230                          *
1231                          * We can't wait forever for the vnode lock, we might
1232                          * deadlock due to a vn_read() getting stuck in
1233                          * vm_wait while holding this vnode.  We skip the
1234                          * vnode if we can't get it in a reasonable amount
1235                          * of time.
1236                          */
1237                         if (object->type == OBJT_VNODE) {
1238                                 vm_page_unlock(m);
1239                                 vp = object->handle;
1240                                 if (vp->v_type == VREG &&
1241                                     vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1242                                         mp = NULL;
1243                                         ++pageout_lock_miss;
1244                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1245                                                 vnodes_skipped++;
1246                                         goto unlock_and_continue;
1247                                 }
1248                                 KASSERT(mp != NULL,
1249                                     ("vp %p with NULL v_mount", vp));
1250                                 vm_object_reference_locked(object);
1251                                 VM_OBJECT_WUNLOCK(object);
1252                                 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
1253                                     LK_SHARED : LK_EXCLUSIVE;
1254                                 if (vget(vp, lockmode | LK_TIMELOCK,
1255                                     curthread)) {
1256                                         VM_OBJECT_WLOCK(object);
1257                                         ++pageout_lock_miss;
1258                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1259                                                 vnodes_skipped++;
1260                                         vp = NULL;
1261                                         goto unlock_and_continue;
1262                                 }
1263                                 VM_OBJECT_WLOCK(object);
1264                                 vm_page_lock(m);
1265                                 vm_pagequeue_lock(pq);
1266                                 queues_locked = TRUE;
1267                                 /*
1268                                  * The page might have been moved to another
1269                                  * queue during potential blocking in vget()
1270                                  * above.  The page might have been freed and
1271                                  * reused for another vnode.
1272                                  */
1273                                 if (m->queue != PQ_INACTIVE ||
1274                                     m->object != object ||
1275                                     TAILQ_NEXT(m, plinks.q) != &vmd->vmd_marker) {
1276                                         vm_page_unlock(m);
1277                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1278                                                 vnodes_skipped++;
1279                                         goto unlock_and_continue;
1280                                 }
1281
1282                                 /*
1283                                  * The page may have been busied during the
1284                                  * blocking in vget().  We don't move the
1285                                  * page back onto the end of the queue so that
1286                                  * statistics are more correct if we don't.
1287                                  */
1288                                 if (vm_page_busied(m)) {
1289                                         vm_page_unlock(m);
1290                                         addl_page_shortage++;
1291                                         goto unlock_and_continue;
1292                                 }
1293
1294                                 /*
1295                                  * If the page has become held it might
1296                                  * be undergoing I/O, so skip it
1297                                  */
1298                                 if (m->hold_count != 0) {
1299                                         vm_page_unlock(m);
1300                                         addl_page_shortage++;
1301                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1302                                                 vnodes_skipped++;
1303                                         goto unlock_and_continue;
1304                                 }
1305                                 vm_pagequeue_unlock(pq);
1306                                 queues_locked = FALSE;
1307                         }
1308
1309                         /*
1310                          * If a page is dirty, then it is either being washed
1311                          * (but not yet cleaned) or it is still in the
1312                          * laundry.  If it is still in the laundry, then we
1313                          * start the cleaning operation.
1314                          *
1315                          * decrement page_shortage on success to account for
1316                          * the (future) cleaned page.  Otherwise we could wind
1317                          * up laundering or cleaning too many pages.
1318                          */
1319                         if (vm_pageout_clean(m) != 0) {
1320                                 --page_shortage;
1321                                 --maxlaunder;
1322                         }
1323 unlock_and_continue:
1324                         vm_page_lock_assert(m, MA_NOTOWNED);
1325                         VM_OBJECT_WUNLOCK(object);
1326                         if (mp != NULL) {
1327                                 if (queues_locked) {
1328                                         vm_pagequeue_unlock(pq);
1329                                         queues_locked = FALSE;
1330                                 }
1331                                 if (vp != NULL)
1332                                         vput(vp);
1333                                 vm_object_deallocate(object);
1334                                 vn_finished_write(mp);
1335                         }
1336                         vm_page_lock_assert(m, MA_NOTOWNED);
1337                         goto relock_queues;
1338                 }
1339                 vm_page_unlock(m);
1340                 VM_OBJECT_WUNLOCK(object);
1341 relock_queues:
1342                 if (!queues_locked) {
1343                         vm_pagequeue_lock(pq);
1344                         queues_locked = TRUE;
1345                 }
1346                 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1347                 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1348         }
1349         vm_pagequeue_unlock(pq);
1350
1351 #if !defined(NO_SWAPPING)
1352         /*
1353          * Wakeup the swapout daemon if we didn't cache or free the targeted
1354          * number of pages.
1355          */
1356         if (vm_swap_enabled && page_shortage > 0)
1357                 vm_req_vmdaemon(VM_SWAP_NORMAL);
1358 #endif
1359
1360         /*
1361          * Wakeup the sync daemon if we skipped a vnode in a writeable object
1362          * and we didn't cache or free enough pages.
1363          */
1364         if (vnodes_skipped > 0 && page_shortage > cnt.v_free_target -
1365             cnt.v_free_min)
1366                 (void)speedup_syncer();
1367
1368         /*
1369          * If the inactive queue scan fails repeatedly to meet its
1370          * target, kill the largest process.
1371          */
1372         vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1373
1374         /*
1375          * Compute the number of pages we want to try to move from the
1376          * active queue to the inactive queue.
1377          */
1378         page_shortage = cnt.v_inactive_target - cnt.v_inactive_count +
1379             vm_paging_target() + deficit + addl_page_shortage;
1380
1381         pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1382         vm_pagequeue_lock(pq);
1383         maxscan = pq->pq_cnt;
1384
1385         /*
1386          * If we're just idle polling attempt to visit every
1387          * active page within 'update_period' seconds.
1388          */
1389         scan_tick = ticks;
1390         if (vm_pageout_update_period != 0) {
1391                 min_scan = pq->pq_cnt;
1392                 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1393                 min_scan /= hz * vm_pageout_update_period;
1394         } else
1395                 min_scan = 0;
1396         if (min_scan > 0 || (page_shortage > 0 && maxscan > 0))
1397                 vmd->vmd_last_active_scan = scan_tick;
1398
1399         /*
1400          * Scan the active queue for pages that can be deactivated.  Update
1401          * the per-page activity counter and use it to identify deactivation
1402          * candidates.
1403          */
1404         for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1405             min_scan || (page_shortage > 0 && scanned < maxscan)); m = next,
1406             scanned++) {
1407
1408                 KASSERT(m->queue == PQ_ACTIVE,
1409                     ("vm_pageout_scan: page %p isn't active", m));
1410
1411                 next = TAILQ_NEXT(m, plinks.q);
1412                 if ((m->flags & PG_MARKER) != 0)
1413                         continue;
1414                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1415                     ("Fictitious page %p cannot be in active queue", m));
1416                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1417                     ("Unmanaged page %p cannot be in active queue", m));
1418                 if (!vm_pageout_page_lock(m, &next)) {
1419                         vm_page_unlock(m);
1420                         continue;
1421                 }
1422
1423                 /*
1424                  * The count for pagedaemon pages is done after checking the
1425                  * page for eligibility...
1426                  */
1427                 PCPU_INC(cnt.v_pdpages);
1428
1429                 /*
1430                  * Check to see "how much" the page has been used.
1431                  */
1432                 act_delta = 0;
1433                 if (m->aflags & PGA_REFERENCED) {
1434                         vm_page_aflag_clear(m, PGA_REFERENCED);
1435                         act_delta += 1;
1436                 }
1437                 /*
1438                  * Unlocked object ref count check.  Two races are possible.
1439                  * 1) The ref was transitioning to zero and we saw non-zero,
1440                  *    the pmap bits will be checked unnecessarily.
1441                  * 2) The ref was transitioning to one and we saw zero.
1442                  *    The page lock prevents a new reference to this page so
1443                  *    we need not check the reference bits.
1444                  */
1445                 if (m->object->ref_count != 0)
1446                         act_delta += pmap_ts_referenced(m);
1447
1448                 /*
1449                  * Advance or decay the act_count based on recent usage.
1450                  */
1451                 if (act_delta) {
1452                         m->act_count += ACT_ADVANCE + act_delta;
1453                         if (m->act_count > ACT_MAX)
1454                                 m->act_count = ACT_MAX;
1455                 } else {
1456                         m->act_count -= min(m->act_count, ACT_DECLINE);
1457                         act_delta = m->act_count;
1458                 }
1459
1460                 /*
1461                  * Move this page to the tail of the active or inactive
1462                  * queue depending on usage.
1463                  */
1464                 if (act_delta == 0) {
1465                         /* Dequeue to avoid later lock recursion. */
1466                         vm_page_dequeue_locked(m);
1467                         vm_page_deactivate(m);
1468                         page_shortage--;
1469                 } else
1470                         vm_page_requeue_locked(m);
1471                 vm_page_unlock(m);
1472         }
1473         vm_pagequeue_unlock(pq);
1474 #if !defined(NO_SWAPPING)
1475         /*
1476          * Idle process swapout -- run once per second.
1477          */
1478         if (vm_swap_idle_enabled) {
1479                 static long lsec;
1480                 if (time_second != lsec) {
1481                         vm_req_vmdaemon(VM_SWAP_IDLE);
1482                         lsec = time_second;
1483                 }
1484         }
1485 #endif
1486 }
1487
1488 static int vm_pageout_oom_vote;
1489
1490 /*
1491  * The pagedaemon threads randlomly select one to perform the
1492  * OOM.  Trying to kill processes before all pagedaemons
1493  * failed to reach free target is premature.
1494  */
1495 static void
1496 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1497     int starting_page_shortage)
1498 {
1499         int old_vote;
1500
1501         if (starting_page_shortage <= 0 || starting_page_shortage !=
1502             page_shortage)
1503                 vmd->vmd_oom_seq = 0;
1504         else
1505                 vmd->vmd_oom_seq++;
1506         if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1507                 if (vmd->vmd_oom) {
1508                         vmd->vmd_oom = FALSE;
1509                         atomic_subtract_int(&vm_pageout_oom_vote, 1);
1510                 }
1511                 return;
1512         }
1513
1514         /*
1515          * Do not follow the call sequence until OOM condition is
1516          * cleared.
1517          */
1518         vmd->vmd_oom_seq = 0;
1519
1520         if (vmd->vmd_oom)
1521                 return;
1522
1523         vmd->vmd_oom = TRUE;
1524         old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1525         if (old_vote != vm_ndomains - 1)
1526                 return;
1527
1528         /*
1529          * The current pagedaemon thread is the last in the quorum to
1530          * start OOM.  Initiate the selection and signaling of the
1531          * victim.
1532          */
1533         vm_pageout_oom(VM_OOM_MEM);
1534
1535         /*
1536          * After one round of OOM terror, recall our vote.  On the
1537          * next pass, current pagedaemon would vote again if the low
1538          * memory condition is still there, due to vmd_oom being
1539          * false.
1540          */
1541         vmd->vmd_oom = FALSE;
1542         atomic_subtract_int(&vm_pageout_oom_vote, 1);
1543 }
1544
1545 /*
1546  * The OOM killer is the page daemon's action of last resort when
1547  * memory allocation requests have been stalled for a prolonged period
1548  * of time because it cannot reclaim memory.  This function computes
1549  * the approximate number of physical pages that could be reclaimed if
1550  * the specified address space is destroyed.
1551  *
1552  * Private, anonymous memory owned by the address space is the
1553  * principal resource that we expect to recover after an OOM kill.
1554  * Since the physical pages mapped by the address space's COW entries
1555  * are typically shared pages, they are unlikely to be released and so
1556  * they are not counted.
1557  *
1558  * To get to the point where the page daemon runs the OOM killer, its
1559  * efforts to write-back vnode-backed pages may have stalled.  This
1560  * could be caused by a memory allocation deadlock in the write path
1561  * that might be resolved by an OOM kill.  Therefore, physical pages
1562  * belonging to vnode-backed objects are counted, because they might
1563  * be freed without being written out first if the address space holds
1564  * the last reference to an unlinked vnode.
1565  *
1566  * Similarly, physical pages belonging to OBJT_PHYS objects are
1567  * counted because the address space might hold the last reference to
1568  * the object.
1569  */
1570 static long
1571 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1572 {
1573         vm_map_t map;
1574         vm_map_entry_t entry;
1575         vm_object_t obj;
1576         long res;
1577
1578         map = &vmspace->vm_map;
1579         KASSERT(!map->system_map, ("system map"));
1580         sx_assert(&map->lock, SA_LOCKED);
1581         res = 0;
1582         for (entry = map->header.next; entry != &map->header;
1583             entry = entry->next) {
1584                 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1585                         continue;
1586                 obj = entry->object.vm_object;
1587                 if (obj == NULL)
1588                         continue;
1589                 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1590                     obj->ref_count != 1)
1591                         continue;
1592                 switch (obj->type) {
1593                 case OBJT_DEFAULT:
1594                 case OBJT_SWAP:
1595                 case OBJT_PHYS:
1596                 case OBJT_VNODE:
1597                         res += obj->resident_page_count;
1598                         break;
1599                 }
1600         }
1601         return (res);
1602 }
1603
1604 void
1605 vm_pageout_oom(int shortage)
1606 {
1607         struct proc *p, *bigproc;
1608         vm_offset_t size, bigsize;
1609         struct thread *td;
1610         struct vmspace *vm;
1611
1612         /*
1613          * We keep the process bigproc locked once we find it to keep anyone
1614          * from messing with it; however, there is a possibility of
1615          * deadlock if process B is bigproc and one of it's child processes
1616          * attempts to propagate a signal to B while we are waiting for A's
1617          * lock while walking this list.  To avoid this, we don't block on
1618          * the process lock but just skip a process if it is already locked.
1619          */
1620         bigproc = NULL;
1621         bigsize = 0;
1622         sx_slock(&allproc_lock);
1623         FOREACH_PROC_IN_SYSTEM(p) {
1624                 int breakout;
1625
1626                 PROC_LOCK(p);
1627
1628                 /*
1629                  * If this is a system, protected or killed process, skip it.
1630                  */
1631                 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1632                     P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1633                     p->p_pid == 1 || P_KILLED(p) ||
1634                     (p->p_pid < 48 && swap_pager_avail != 0)) {
1635                         PROC_UNLOCK(p);
1636                         continue;
1637                 }
1638                 /*
1639                  * If the process is in a non-running type state,
1640                  * don't touch it.  Check all the threads individually.
1641                  */
1642                 breakout = 0;
1643                 FOREACH_THREAD_IN_PROC(p, td) {
1644                         thread_lock(td);
1645                         if (!TD_ON_RUNQ(td) &&
1646                             !TD_IS_RUNNING(td) &&
1647                             !TD_IS_SLEEPING(td) &&
1648                             !TD_IS_SUSPENDED(td) &&
1649                             !TD_IS_SWAPPED(td)) {
1650                                 thread_unlock(td);
1651                                 breakout = 1;
1652                                 break;
1653                         }
1654                         thread_unlock(td);
1655                 }
1656                 if (breakout) {
1657                         PROC_UNLOCK(p);
1658                         continue;
1659                 }
1660                 /*
1661                  * get the process size
1662                  */
1663                 vm = vmspace_acquire_ref(p);
1664                 if (vm == NULL) {
1665                         PROC_UNLOCK(p);
1666                         continue;
1667                 }
1668                 _PHOLD(p);
1669                 if (!vm_map_trylock_read(&vm->vm_map)) {
1670                         _PRELE(p);
1671                         PROC_UNLOCK(p);
1672                         vmspace_free(vm);
1673                         continue;
1674                 }
1675                 PROC_UNLOCK(p);
1676                 size = vmspace_swap_count(vm);
1677                 if (shortage == VM_OOM_MEM)
1678                         size += vm_pageout_oom_pagecount(vm);
1679                 vm_map_unlock_read(&vm->vm_map);
1680                 vmspace_free(vm);
1681
1682                 /*
1683                  * If this process is bigger than the biggest one,
1684                  * remember it.
1685                  */
1686                 if (size > bigsize) {
1687                         if (bigproc != NULL)
1688                                 PRELE(bigproc);
1689                         bigproc = p;
1690                         bigsize = size;
1691                 } else {
1692                         PRELE(p);
1693                 }
1694         }
1695         sx_sunlock(&allproc_lock);
1696         if (bigproc != NULL) {
1697                 PROC_LOCK(bigproc);
1698                 killproc(bigproc, "out of swap space");
1699                 sched_nice(bigproc, PRIO_MIN);
1700                 _PRELE(bigproc);
1701                 PROC_UNLOCK(bigproc);
1702                 wakeup(&cnt.v_free_count);
1703         }
1704 }
1705
1706 static void
1707 vm_pageout_worker(void *arg)
1708 {
1709         struct vm_domain *domain;
1710         int domidx;
1711
1712         domidx = (uintptr_t)arg;
1713         domain = &vm_dom[domidx];
1714
1715         /*
1716          * XXXKIB It could be useful to bind pageout daemon threads to
1717          * the cores belonging to the domain, from which vm_page_array
1718          * is allocated.
1719          */
1720
1721         KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1722         domain->vmd_last_active_scan = ticks;
1723         vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1724
1725         /*
1726          * The pageout daemon worker is never done, so loop forever.
1727          */
1728         while (TRUE) {
1729                 /*
1730                  * If we have enough free memory, wakeup waiters.  Do
1731                  * not clear vm_pages_needed until we reach our target,
1732                  * otherwise we may be woken up over and over again and
1733                  * waste a lot of cpu.
1734                  */
1735                 mtx_lock(&vm_page_queue_free_mtx);
1736                 if (vm_pages_needed && !vm_page_count_min()) {
1737                         if (!vm_paging_needed())
1738                                 vm_pages_needed = 0;
1739                         wakeup(&cnt.v_free_count);
1740                 }
1741                 if (vm_pages_needed) {
1742                         /*
1743                          * We're still not done.  Either vm_pages_needed was
1744                          * set by another thread during the previous scan
1745                          * (typically, this happens during a level 0 scan) or
1746                          * vm_pages_needed was already set and the scan failed
1747                          * to free enough pages.  If we haven't yet performed
1748                          * a level >= 2 scan (unlimited dirty cleaning), then
1749                          * upgrade the level and scan again now.  Otherwise,
1750                          * sleep a bit and try again later.  While sleeping,
1751                          * vm_pages_needed can be cleared.
1752                          */
1753                         if (domain->vmd_pass > 1)
1754                                 msleep(&vm_pages_needed,
1755                                     &vm_page_queue_free_mtx, PVM, "psleep",
1756                                     hz / 2);
1757                 } else {
1758                         /*
1759                          * Good enough, sleep until required to refresh
1760                          * stats.
1761                          */
1762                         msleep(&vm_pages_needed, &vm_page_queue_free_mtx,
1763                             PVM, "psleep", hz);
1764                 }
1765                 if (vm_pages_needed) {
1766                         cnt.v_pdwakeups++;
1767                         domain->vmd_pass++;
1768                 } else
1769                         domain->vmd_pass = 0;
1770                 mtx_unlock(&vm_page_queue_free_mtx);
1771                 vm_pageout_scan(domain, domain->vmd_pass);
1772         }
1773 }
1774
1775 /*
1776  *      vm_pageout_init initialises basic pageout daemon settings.
1777  */
1778 static void
1779 vm_pageout_init(void)
1780 {
1781         /*
1782          * Initialize some paging parameters.
1783          */
1784         cnt.v_interrupt_free_min = 2;
1785         if (cnt.v_page_count < 2000)
1786                 vm_pageout_page_count = 8;
1787
1788         /*
1789          * v_free_reserved needs to include enough for the largest
1790          * swap pager structures plus enough for any pv_entry structs
1791          * when paging.
1792          */
1793         if (cnt.v_page_count > 1024)
1794                 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1795         else
1796                 cnt.v_free_min = 4;
1797         cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1798             cnt.v_interrupt_free_min;
1799         cnt.v_free_reserved = vm_pageout_page_count +
1800             cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1801         cnt.v_free_severe = cnt.v_free_min / 2;
1802         cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1803         cnt.v_free_min += cnt.v_free_reserved;
1804         cnt.v_free_severe += cnt.v_free_reserved;
1805         cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1806         if (cnt.v_inactive_target > cnt.v_free_count / 3)
1807                 cnt.v_inactive_target = cnt.v_free_count / 3;
1808
1809         /*
1810          * Set the default wakeup threshold to be 10% above the minimum
1811          * page limit.  This keeps the steady state out of shortfall.
1812          */
1813         vm_pageout_wakeup_thresh = (cnt.v_free_min / 10) * 11;
1814
1815         /*
1816          * Set interval in seconds for active scan.  We want to visit each
1817          * page at least once every ten minutes.  This is to prevent worst
1818          * case paging behaviors with stale active LRU.
1819          */
1820         if (vm_pageout_update_period == 0)
1821                 vm_pageout_update_period = 600;
1822
1823         /* XXX does not really belong here */
1824         if (vm_page_max_wired == 0)
1825                 vm_page_max_wired = cnt.v_free_count / 3;
1826 }
1827
1828 /*
1829  *     vm_pageout is the high level pageout daemon.
1830  */
1831 static void
1832 vm_pageout(void)
1833 {
1834         int error;
1835 #if MAXMEMDOM > 1
1836         int i;
1837 #endif
1838
1839         swap_pager_swap_init();
1840 #if MAXMEMDOM > 1
1841         for (i = 1; i < vm_ndomains; i++) {
1842                 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1843                     curproc, NULL, 0, 0, "dom%d", i);
1844                 if (error != 0) {
1845                         panic("starting pageout for domain %d, error %d\n",
1846                             i, error);
1847                 }
1848         }
1849 #endif
1850         error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
1851             0, 0, "uma");
1852         if (error != 0)
1853                 panic("starting uma_reclaim helper, error %d\n", error);
1854         vm_pageout_worker((void *)(uintptr_t)0);
1855 }
1856
1857 /*
1858  * Unless the free page queue lock is held by the caller, this function
1859  * should be regarded as advisory.  Specifically, the caller should
1860  * not msleep() on &cnt.v_free_count following this function unless
1861  * the free page queue lock is held until the msleep() is performed.
1862  */
1863 void
1864 pagedaemon_wakeup(void)
1865 {
1866
1867         if (!vm_pages_needed && curthread->td_proc != pageproc) {
1868                 vm_pages_needed = 1;
1869                 wakeup(&vm_pages_needed);
1870         }
1871 }
1872
1873 #if !defined(NO_SWAPPING)
1874 static void
1875 vm_req_vmdaemon(int req)
1876 {
1877         static int lastrun = 0;
1878
1879         mtx_lock(&vm_daemon_mtx);
1880         vm_pageout_req_swapout |= req;
1881         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1882                 wakeup(&vm_daemon_needed);
1883                 lastrun = ticks;
1884         }
1885         mtx_unlock(&vm_daemon_mtx);
1886 }
1887
1888 static void
1889 vm_daemon(void)
1890 {
1891         struct rlimit rsslim;
1892         struct proc *p;
1893         struct thread *td;
1894         struct vmspace *vm;
1895         int breakout, swapout_flags, tryagain, attempts;
1896 #ifdef RACCT
1897         uint64_t rsize, ravailable;
1898 #endif
1899
1900         while (TRUE) {
1901                 mtx_lock(&vm_daemon_mtx);
1902                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
1903 #ifdef RACCT
1904                     racct_enable ? hz : 0
1905 #else
1906                     0
1907 #endif
1908                 );
1909                 swapout_flags = vm_pageout_req_swapout;
1910                 vm_pageout_req_swapout = 0;
1911                 mtx_unlock(&vm_daemon_mtx);
1912                 if (swapout_flags)
1913                         swapout_procs(swapout_flags);
1914
1915                 /*
1916                  * scan the processes for exceeding their rlimits or if
1917                  * process is swapped out -- deactivate pages
1918                  */
1919                 tryagain = 0;
1920                 attempts = 0;
1921 again:
1922                 attempts++;
1923                 sx_slock(&allproc_lock);
1924                 FOREACH_PROC_IN_SYSTEM(p) {
1925                         vm_pindex_t limit, size;
1926
1927                         /*
1928                          * if this is a system process or if we have already
1929                          * looked at this process, skip it.
1930                          */
1931                         PROC_LOCK(p);
1932                         if (p->p_state != PRS_NORMAL ||
1933                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1934                                 PROC_UNLOCK(p);
1935                                 continue;
1936                         }
1937                         /*
1938                          * if the process is in a non-running type state,
1939                          * don't touch it.
1940                          */
1941                         breakout = 0;
1942                         FOREACH_THREAD_IN_PROC(p, td) {
1943                                 thread_lock(td);
1944                                 if (!TD_ON_RUNQ(td) &&
1945                                     !TD_IS_RUNNING(td) &&
1946                                     !TD_IS_SLEEPING(td) &&
1947                                     !TD_IS_SUSPENDED(td)) {
1948                                         thread_unlock(td);
1949                                         breakout = 1;
1950                                         break;
1951                                 }
1952                                 thread_unlock(td);
1953                         }
1954                         if (breakout) {
1955                                 PROC_UNLOCK(p);
1956                                 continue;
1957                         }
1958                         /*
1959                          * get a limit
1960                          */
1961                         lim_rlimit(p, RLIMIT_RSS, &rsslim);
1962                         limit = OFF_TO_IDX(
1963                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
1964
1965                         /*
1966                          * let processes that are swapped out really be
1967                          * swapped out set the limit to nothing (will force a
1968                          * swap-out.)
1969                          */
1970                         if ((p->p_flag & P_INMEM) == 0)
1971                                 limit = 0;      /* XXX */
1972                         vm = vmspace_acquire_ref(p);
1973                         PROC_UNLOCK(p);
1974                         if (vm == NULL)
1975                                 continue;
1976
1977                         size = vmspace_resident_count(vm);
1978                         if (size >= limit) {
1979                                 vm_pageout_map_deactivate_pages(
1980                                     &vm->vm_map, limit);
1981                         }
1982 #ifdef RACCT
1983                         if (racct_enable) {
1984                                 rsize = IDX_TO_OFF(size);
1985                                 PROC_LOCK(p);
1986                                 racct_set(p, RACCT_RSS, rsize);
1987                                 ravailable = racct_get_available(p, RACCT_RSS);
1988                                 PROC_UNLOCK(p);
1989                                 if (rsize > ravailable) {
1990                                         /*
1991                                          * Don't be overly aggressive; this
1992                                          * might be an innocent process,
1993                                          * and the limit could've been exceeded
1994                                          * by some memory hog.  Don't try
1995                                          * to deactivate more than 1/4th
1996                                          * of process' resident set size.
1997                                          */
1998                                         if (attempts <= 8) {
1999                                                 if (ravailable < rsize -
2000                                                     (rsize / 4)) {
2001                                                         ravailable = rsize -
2002                                                             (rsize / 4);
2003                                                 }
2004                                         }
2005                                         vm_pageout_map_deactivate_pages(
2006                                             &vm->vm_map,
2007                                             OFF_TO_IDX(ravailable));
2008                                         /* Update RSS usage after paging out. */
2009                                         size = vmspace_resident_count(vm);
2010                                         rsize = IDX_TO_OFF(size);
2011                                         PROC_LOCK(p);
2012                                         racct_set(p, RACCT_RSS, rsize);
2013                                         PROC_UNLOCK(p);
2014                                         if (rsize > ravailable)
2015                                                 tryagain = 1;
2016                                 }
2017                         }
2018 #endif
2019                         vmspace_free(vm);
2020                 }
2021                 sx_sunlock(&allproc_lock);
2022                 if (tryagain != 0 && attempts <= 10)
2023                         goto again;
2024         }
2025 }
2026 #endif                  /* !defined(NO_SWAPPING) */