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
  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 m, int *numpagedout);
 123 static int vm_pageout_cluster(vm_page_t m);
 124 static bool vm_pageout_scan(struct vm_domain *vmd, int pass);
 125 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
 126     int starting_page_shortage);
 127
 128 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
 129     NULL);
 130
 131 struct proc *pageproc;
 132
 133 static struct kproc_desc page_kp = {
 134         "pagedaemon",
 135         vm_pageout,
 136         &pageproc
 137 };
 138 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
 139     &page_kp);
 140
 141 SDT_PROVIDER_DEFINE(vm);
 142 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
 143
 144 /* Pagedaemon activity rates, in subdivisions of one second. */
 145 #define VM_LAUNDER_RATE         10
 146 #define VM_INACT_SCAN_RATE      2
 147
 148 int vm_pageout_deficit;         /* Estimated number of pages deficit */
 149 u_int vm_pageout_wakeup_thresh;
 150 static int vm_pageout_oom_seq = 12;
 151 bool vm_pageout_wanted;         /* Event on which pageout daemon sleeps */
 152 bool vm_pages_needed;           /* Are threads waiting for free pages? */
 153
 154 /* Pending request for dirty page laundering. */
 155 static enum {
 156         VM_LAUNDRY_IDLE,
 157         VM_LAUNDRY_BACKGROUND,
 158         VM_LAUNDRY_SHORTFALL
 159 } vm_laundry_request = VM_LAUNDRY_IDLE;
 160
 161 static int vm_pageout_update_period;
 162 static int disable_swap_pageouts;
 163 static int lowmem_period = 10;
 164 static time_t lowmem_uptime;
 165 static int swapdev_enabled;
 166
 167 static int vm_panic_on_oom = 0;
 168
 169 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
 170         CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
 171         "panic on out of memory instead of killing the largest process");
 172
 173 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
 174         CTLFLAG_RWTUN, &vm_pageout_wakeup_thresh, 0,
 175         "free page threshold for waking up the pageout daemon");
 176
 177 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
 178         CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
 179         "Maximum active LRU update period");
 180
 181 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
 182         "Low memory callback period");
 183
 184 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
 185         CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
 186
 187 static int pageout_lock_miss;
 188 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
 189         CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
 190
 191 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
 192         CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
 193         "back-to-back calls to oom detector to start OOM");
 194
 195 static int act_scan_laundry_weight = 3;
 196 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
 197     &act_scan_laundry_weight, 0,
 198     "weight given to clean vs. dirty pages in active queue scans");
 199
 200 static u_int vm_background_launder_target;
 201 SYSCTL_UINT(_vm, OID_AUTO, background_launder_target, CTLFLAG_RWTUN,
 202     &vm_background_launder_target, 0,
 203     "background laundering target, in pages");
 204
 205 static u_int vm_background_launder_rate = 4096;
 206 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
 207     &vm_background_launder_rate, 0,
 208     "background laundering rate, in kilobytes per second");
 209
 210 static u_int vm_background_launder_max = 20 * 1024;
 211 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
 212     &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
 213
 214 int vm_pageout_page_count = 32;
 215
 216 int vm_page_max_wired;          /* XXX max # of wired pages system-wide */
 217 SYSCTL_INT(_vm, OID_AUTO, max_wired,
 218         CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
 219
 220 static u_int isqrt(u_int num);
 221 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
 222 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
 223     bool in_shortfall);
 224 static void vm_pageout_laundry_worker(void *arg);
 225 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
 226
 227 /*
 228  * Initialize a dummy page for marking the caller's place in the specified
 229  * paging queue.  In principle, this function only needs to set the flag
 230  * PG_MARKER.  Nonetheless, it write busies and initializes the hold count
 231  * to one as safety precautions.
 232  */
 233 static void
 234 vm_pageout_init_marker(vm_page_t marker, u_short queue)
 235 {
 236
 237         bzero(marker, sizeof(*marker));
 238         marker->flags = PG_MARKER;
 239         marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
 240         marker->queue = queue;
 241         marker->hold_count = 1;
 242 }
 243
 244 /*
 245  * vm_pageout_fallback_object_lock:
 246  *
 247  * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
 248  * known to have failed and page queue must be either PQ_ACTIVE or
 249  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queue
 250  * while locking the vm object.  Use marker page to detect page queue
 251  * changes and maintain notion of next page on page queue.  Return
 252  * TRUE if no changes were detected, FALSE otherwise.  vm object is
 253  * locked on return.
 254  *
 255  * This function depends on both the lock portion of struct vm_object
 256  * and normal struct vm_page being type stable.
 257  */
 258 static boolean_t
 259 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
 260 {
 261         struct vm_page marker;
 262         struct vm_pagequeue *pq;
 263         boolean_t unchanged;
 264         u_short queue;
 265         vm_object_t object;
 266
 267         queue = m->queue;
 268         vm_pageout_init_marker(&marker, queue);
 269         pq = vm_page_pagequeue(m);
 270         object = m->object;
 271
 272         TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
 273         vm_pagequeue_unlock(pq);
 274         vm_page_unlock(m);
 275         VM_OBJECT_WLOCK(object);
 276         vm_page_lock(m);
 277         vm_pagequeue_lock(pq);
 278
 279         /*
 280          * The page's object might have changed, and/or the page might
 281          * have moved from its original position in the queue.  If the
 282          * page's object has changed, then the caller should abandon
 283          * processing the page because the wrong object lock was
 284          * acquired.  Use the marker's plinks.q, not the page's, to
 285          * determine if the page has been moved.  The state of the
 286          * page's plinks.q can be indeterminate; whereas, the marker's
 287          * plinks.q must be valid.
 288          */
 289         *next = TAILQ_NEXT(&marker, plinks.q);
 290         unchanged = m->object == object &&
 291             m == TAILQ_PREV(&marker, pglist, plinks.q);
 292         KASSERT(!unchanged || m->queue == queue,
 293             ("page %p queue %d %d", m, queue, m->queue));
 294         TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
 295         return (unchanged);
 296 }
 297
 298 /*
 299  * Lock the page while holding the page queue lock.  Use marker page
 300  * to detect page queue changes and maintain notion of next page on
 301  * page queue.  Return TRUE if no changes were detected, FALSE
 302  * otherwise.  The page is locked on return. The page queue lock might
 303  * be dropped and reacquired.
 304  *
 305  * This function depends on normal struct vm_page being type stable.
 306  */
 307 static boolean_t
 308 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
 309 {
 310         struct vm_page marker;
 311         struct vm_pagequeue *pq;
 312         boolean_t unchanged;
 313         u_short queue;
 314
 315         vm_page_lock_assert(m, MA_NOTOWNED);
 316         if (vm_page_trylock(m))
 317                 return (TRUE);
 318
 319         queue = m->queue;
 320         vm_pageout_init_marker(&marker, queue);
 321         pq = vm_page_pagequeue(m);
 322
 323         TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
 324         vm_pagequeue_unlock(pq);
 325         vm_page_lock(m);
 326         vm_pagequeue_lock(pq);
 327
 328         /* Page queue might have changed. */
 329         *next = TAILQ_NEXT(&marker, plinks.q);
 330         unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
 331         KASSERT(!unchanged || m->queue == queue,
 332             ("page %p queue %d %d", m, queue, m->queue));
 333         TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
 334         return (unchanged);
 335 }
 336
 337 /*
 338  * Scan for pages at adjacent offsets within the given page's object that are
 339  * eligible for laundering, form a cluster of these pages and the given page,
 340  * and launder that cluster.
 341  */
 342 static int
 343 vm_pageout_cluster(vm_page_t m)
 344 {
 345         vm_object_t object;
 346         vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
 347         vm_pindex_t pindex;
 348         int ib, is, page_base, pageout_count;
 349
 350         vm_page_assert_locked(m);
 351         object = m->object;
 352         VM_OBJECT_ASSERT_WLOCKED(object);
 353         pindex = m->pindex;
 354
 355         /*
 356          * We can't clean the page if it is busy or held.
 357          */
 358         vm_page_assert_unbusied(m);
 359         KASSERT(m->hold_count == 0, ("page %p is held", m));
 360
 361         pmap_remove_write(m);
 362         vm_page_unlock(m);
 363
 364         mc[vm_pageout_page_count] = pb = ps = m;
 365         pageout_count = 1;
 366         page_base = vm_pageout_page_count;
 367         ib = 1;
 368         is = 1;
 369
 370         /*
 371          * We can cluster only if the page is not clean, busy, or held, and
 372          * the page is in the laundry queue.
 373          *
 374          * During heavy mmap/modification loads the pageout
 375          * daemon can really fragment the underlying file
 376          * due to flushing pages out of order and not trying to
 377          * align the clusters (which leaves sporadic out-of-order
 378          * holes).  To solve this problem we do the reverse scan
 379          * first and attempt to align our cluster, then do a
 380          * forward scan if room remains.
 381          */
 382 more:
 383         while (ib != 0 && pageout_count < vm_pageout_page_count) {
 384                 if (ib > pindex) {
 385                         ib = 0;
 386                         break;
 387                 }
 388                 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
 389                         ib = 0;
 390                         break;
 391                 }
 392                 vm_page_test_dirty(p);
 393                 if (p->dirty == 0) {
 394                         ib = 0;
 395                         break;
 396                 }
 397                 vm_page_lock(p);
 398                 if (!vm_page_in_laundry(p) ||
 399                     p->hold_count != 0) {       /* may be undergoing I/O */
 400                         vm_page_unlock(p);
 401                         ib = 0;
 402                         break;
 403                 }
 404                 pmap_remove_write(p);
 405                 vm_page_unlock(p);
 406                 mc[--page_base] = pb = p;
 407                 ++pageout_count;
 408                 ++ib;
 409
 410                 /*
 411                  * We are at an alignment boundary.  Stop here, and switch
 412                  * directions.  Do not clear ib.
 413                  */
 414                 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
 415                         break;
 416         }
 417         while (pageout_count < vm_pageout_page_count &&
 418             pindex + is < object->size) {
 419                 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
 420                         break;
 421                 vm_page_test_dirty(p);
 422                 if (p->dirty == 0)
 423                         break;
 424                 vm_page_lock(p);
 425                 if (!vm_page_in_laundry(p) ||
 426                     p->hold_count != 0) {       /* may be undergoing I/O */
 427                         vm_page_unlock(p);
 428                         break;
 429                 }
 430                 pmap_remove_write(p);
 431                 vm_page_unlock(p);
 432                 mc[page_base + pageout_count] = ps = p;
 433                 ++pageout_count;
 434                 ++is;
 435         }
 436
 437         /*
 438          * If we exhausted our forward scan, continue with the reverse scan
 439          * when possible, even past an alignment boundary.  This catches
 440          * boundary conditions.
 441          */
 442         if (ib != 0 && pageout_count < vm_pageout_page_count)
 443                 goto more;
 444
 445         return (vm_pageout_flush(&mc[page_base], pageout_count,
 446             VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
 447 }
 448
 449 /*
 450  * vm_pageout_flush() - launder the given pages
 451  *
 452  *      The given pages are laundered.  Note that we setup for the start of
 453  *      I/O ( i.e. busy the page ), mark it read-only, and bump the object
 454  *      reference count all in here rather then in the parent.  If we want
 455  *      the parent to do more sophisticated things we may have to change
 456  *      the ordering.
 457  *
 458  *      Returned runlen is the count of pages between mreq and first
 459  *      page after mreq with status VM_PAGER_AGAIN.
 460  *      *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
 461  *      for any page in runlen set.
 462  */
 463 int
 464 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
 465     boolean_t *eio)
 466 {
 467         vm_object_t object = mc[0]->object;
 468         int pageout_status[count];
 469         int numpagedout = 0;
 470         int i, runlen;
 471
 472         VM_OBJECT_ASSERT_WLOCKED(object);
 473
 474         /*
 475          * Initiate I/O.  Mark the pages busy and verify that they're valid
 476          * and read-only.
 477          *
 478          * We do not have to fixup the clean/dirty bits here... we can
 479          * allow the pager to do it after the I/O completes.
 480          *
 481          * NOTE! mc[i]->dirty may be partial or fragmented due to an
 482          * edge case with file fragments.
 483          */
 484         for (i = 0; i < count; i++) {
 485                 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
 486                     ("vm_pageout_flush: partially invalid page %p index %d/%d",
 487                         mc[i], i, count));
 488                 KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0,
 489                     ("vm_pageout_flush: writeable page %p", mc[i]));
 490                 vm_page_sbusy(mc[i]);
 491         }
 492         vm_object_pip_add(object, count);
 493
 494         vm_pager_put_pages(object, mc, count, flags, pageout_status);
 495
 496         runlen = count - mreq;
 497         if (eio != NULL)
 498                 *eio = FALSE;
 499         for (i = 0; i < count; i++) {
 500                 vm_page_t mt = mc[i];
 501
 502                 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
 503                     !pmap_page_is_write_mapped(mt),
 504                     ("vm_pageout_flush: page %p is not write protected", mt));
 505                 switch (pageout_status[i]) {
 506                 case VM_PAGER_OK:
 507                         vm_page_lock(mt);
 508                         if (vm_page_in_laundry(mt))
 509                                 vm_page_deactivate_noreuse(mt);
 510                         vm_page_unlock(mt);
 511                         /* FALLTHROUGH */
 512                 case VM_PAGER_PEND:
 513                         numpagedout++;
 514                         break;
 515                 case VM_PAGER_BAD:
 516                         /*
 517                          * The page is outside the object's range.  We pretend
 518                          * that the page out worked and clean the page, so the
 519                          * changes will be lost if the page is reclaimed by
 520                          * the page daemon.
 521                          */
 522                         vm_page_undirty(mt);
 523                         vm_page_lock(mt);
 524                         if (vm_page_in_laundry(mt))
 525                                 vm_page_deactivate_noreuse(mt);
 526                         vm_page_unlock(mt);
 527                         break;
 528                 case VM_PAGER_ERROR:
 529                 case VM_PAGER_FAIL:
 530                         /*
 531                          * If the page couldn't be paged out to swap because the
 532                          * pager wasn't able to find space, place the page in
 533                          * the PQ_UNSWAPPABLE holding queue.  This is an
 534                          * optimization that prevents the page daemon from
 535                          * wasting CPU cycles on pages that cannot be reclaimed
 536                          * becase no swap device is configured.
 537                          *
 538                          * Otherwise, reactivate the page so that it doesn't
 539                          * clog the laundry and inactive queues.  (We will try
 540                          * paging it out again later.)
 541                          */
 542                         vm_page_lock(mt);
 543                         if (object->type == OBJT_SWAP &&
 544                             pageout_status[i] == VM_PAGER_FAIL) {
 545                                 vm_page_unswappable(mt);
 546                                 numpagedout++;
 547                         } else
 548                                 vm_page_activate(mt);
 549                         vm_page_unlock(mt);
 550                         if (eio != NULL && i >= mreq && i - mreq < runlen)
 551                                 *eio = TRUE;
 552                         break;
 553                 case VM_PAGER_AGAIN:
 554                         if (i >= mreq && i - mreq < runlen)
 555                                 runlen = i - mreq;
 556                         break;
 557                 }
 558
 559                 /*
 560                  * If the operation is still going, leave the page busy to
 561                  * block all other accesses. Also, leave the paging in
 562                  * progress indicator set so that we don't attempt an object
 563                  * collapse.
 564                  */
 565                 if (pageout_status[i] != VM_PAGER_PEND) {
 566                         vm_object_pip_wakeup(object);
 567                         vm_page_sunbusy(mt);
 568                 }
 569         }
 570         if (prunlen != NULL)
 571                 *prunlen = runlen;
 572         return (numpagedout);
 573 }
 574
 575 static void
 576 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
 577 {
 578
 579         atomic_store_rel_int(&swapdev_enabled, 1);
 580 }
 581
 582 static void
 583 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
 584 {
 585
 586         if (swap_pager_nswapdev() == 1)
 587                 atomic_store_rel_int(&swapdev_enabled, 0);
 588 }
 589
 590 /*
 591  * Attempt to acquire all of the necessary locks to launder a page and
 592  * then call through the clustering layer to PUTPAGES.  Wait a short
 593  * time for a vnode lock.
 594  *
 595  * Requires the page and object lock on entry, releases both before return.
 596  * Returns 0 on success and an errno otherwise.
 597  */
 598 static int
 599 vm_pageout_clean(vm_page_t m, int *numpagedout)
 600 {
 601         struct vnode *vp;
 602         struct mount *mp;
 603         vm_object_t object;
 604         vm_pindex_t pindex;
 605         int error, lockmode;
 606
 607         vm_page_assert_locked(m);
 608         object = m->object;
 609         VM_OBJECT_ASSERT_WLOCKED(object);
 610         error = 0;
 611         vp = NULL;
 612         mp = NULL;
 613
 614         /*
 615          * The object is already known NOT to be dead.   It
 616          * is possible for the vget() to block the whole
 617          * pageout daemon, but the new low-memory handling
 618          * code should prevent it.
 619          *
 620          * We can't wait forever for the vnode lock, we might
 621          * deadlock due to a vn_read() getting stuck in
 622          * vm_wait while holding this vnode.  We skip the
 623          * vnode if we can't get it in a reasonable amount
 624          * of time.
 625          */
 626         if (object->type == OBJT_VNODE) {
 627                 vm_page_unlock(m);
 628                 vp = object->handle;
 629                 if (vp->v_type == VREG &&
 630                     vn_start_write(vp, &mp, V_NOWAIT) != 0) {
 631                         mp = NULL;
 632                         error = EDEADLK;
 633                         goto unlock_all;
 634                 }
 635                 KASSERT(mp != NULL,
 636                     ("vp %p with NULL v_mount", vp));
 637                 vm_object_reference_locked(object);
 638                 pindex = m->pindex;
 639                 VM_OBJECT_WUNLOCK(object);
 640                 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
 641                     LK_SHARED : LK_EXCLUSIVE;
 642                 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
 643                         vp = NULL;
 644                         error = EDEADLK;
 645                         goto unlock_mp;
 646                 }
 647                 VM_OBJECT_WLOCK(object);
 648                 vm_page_lock(m);
 649                 /*
 650                  * While the object and page were unlocked, the page
 651                  * may have been:
 652                  * (1) moved to a different queue,
 653                  * (2) reallocated to a different object,
 654                  * (3) reallocated to a different offset, or
 655                  * (4) cleaned.
 656                  */
 657                 if (!vm_page_in_laundry(m) || m->object != object ||
 658                     m->pindex != pindex || m->dirty == 0) {
 659                         vm_page_unlock(m);
 660                         error = ENXIO;
 661                         goto unlock_all;
 662                 }
 663
 664                 /*
 665                  * The page may have been busied or held while the object
 666                  * and page locks were released.
 667                  */
 668                 if (vm_page_busied(m) || m->hold_count != 0) {
 669                         vm_page_unlock(m);
 670                         error = EBUSY;
 671                         goto unlock_all;
 672                 }
 673         }
 674
 675         /*
 676          * If a page is dirty, then it is either being washed
 677          * (but not yet cleaned) or it is still in the
 678          * laundry.  If it is still in the laundry, then we
 679          * start the cleaning operation.
 680          */
 681         if ((*numpagedout = vm_pageout_cluster(m)) == 0)
 682                 error = EIO;
 683
 684 unlock_all:
 685         VM_OBJECT_WUNLOCK(object);
 686
 687 unlock_mp:
 688         vm_page_lock_assert(m, MA_NOTOWNED);
 689         if (mp != NULL) {
 690                 if (vp != NULL)
 691                         vput(vp);
 692                 vm_object_deallocate(object);
 693                 vn_finished_write(mp);
 694         }
 695
 696         return (error);
 697 }
 698
 699 /*
 700  * Attempt to launder the specified number of pages.
 701  *
 702  * Returns the number of pages successfully laundered.
 703  */
 704 static int
 705 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
 706 {
 707         struct vm_pagequeue *pq;
 708         vm_object_t object;
 709         vm_page_t m, next;
 710         int act_delta, error, maxscan, numpagedout, starting_target;
 711         int vnodes_skipped;
 712         bool pageout_ok, queue_locked;
 713
 714         starting_target = launder;
 715         vnodes_skipped = 0;
 716
 717         /*
 718          * Scan the laundry queues for pages eligible to be laundered.  We stop
 719          * once the target number of dirty pages have been laundered, or once
 720          * we've reached the end of the queue.  A single iteration of this loop
 721          * may cause more than one page to be laundered because of clustering.
 722          *
 723          * maxscan ensures that we don't re-examine requeued pages.  Any
 724          * additional pages written as part of a cluster are subtracted from
 725          * maxscan since they must be taken from the laundry queue.
 726          *
 727          * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
 728          * swap devices are configured.
 729          */
 730         if (atomic_load_acq_int(&swapdev_enabled))
 731                 pq = &vmd->vmd_pagequeues[PQ_UNSWAPPABLE];
 732         else
 733                 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
 734
 735 scan:
 736         vm_pagequeue_lock(pq);
 737         maxscan = pq->pq_cnt;
 738         queue_locked = true;
 739         for (m = TAILQ_FIRST(&pq->pq_pl);
 740             m != NULL && maxscan-- > 0 && launder > 0;
 741             m = next) {
 742                 vm_pagequeue_assert_locked(pq);
 743                 KASSERT(queue_locked, ("unlocked laundry queue"));
 744                 KASSERT(vm_page_in_laundry(m),
 745                     ("page %p has an inconsistent queue", m));
 746                 next = TAILQ_NEXT(m, plinks.q);
 747                 if ((m->flags & PG_MARKER) != 0)
 748                         continue;
 749                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
 750                     ("PG_FICTITIOUS page %p cannot be in laundry queue", m));
 751                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
 752                     ("VPO_UNMANAGED page %p cannot be in laundry queue", m));
 753                 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
 754                         vm_page_unlock(m);
 755                         continue;
 756                 }
 757                 object = m->object;
 758                 if ((!VM_OBJECT_TRYWLOCK(object) &&
 759                     (!vm_pageout_fallback_object_lock(m, &next) ||
 760                     m->hold_count != 0)) || vm_page_busied(m)) {
 761                         VM_OBJECT_WUNLOCK(object);
 762                         vm_page_unlock(m);
 763                         continue;
 764                 }
 765
 766                 /*
 767                  * Unlock the laundry queue, invalidating the 'next' pointer.
 768                  * Use a marker to remember our place in the laundry queue.
 769                  */
 770                 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_laundry_marker,
 771                     plinks.q);
 772                 vm_pagequeue_unlock(pq);
 773                 queue_locked = false;
 774
 775                 /*
 776                  * Invalid pages can be easily freed.  They cannot be
 777                  * mapped; vm_page_free() asserts this.
 778                  */
 779                 if (m->valid == 0)
 780                         goto free_page;
 781
 782                 /*
 783                  * If the page has been referenced and the object is not dead,
 784                  * reactivate or requeue the page depending on whether the
 785                  * object is mapped.
 786                  */
 787                 if ((m->aflags & PGA_REFERENCED) != 0) {
 788                         vm_page_aflag_clear(m, PGA_REFERENCED);
 789                         act_delta = 1;
 790                 } else
 791                         act_delta = 0;
 792                 if (object->ref_count != 0)
 793                         act_delta += pmap_ts_referenced(m);
 794                 else {
 795                         KASSERT(!pmap_page_is_mapped(m),
 796                             ("page %p is mapped", m));
 797                 }
 798                 if (act_delta != 0) {
 799                         if (object->ref_count != 0) {
 800                                 VM_CNT_INC(v_reactivated);
 801                                 vm_page_activate(m);
 802
 803                                 /*
 804                                  * Increase the activation count if the page
 805                                  * was referenced while in the laundry queue.
 806                                  * This makes it less likely that the page will
 807                                  * be returned prematurely to the inactive
 808                                  * queue.
 809                                  */
 810                                 m->act_count += act_delta + ACT_ADVANCE;
 811
 812                                 /*
 813                                  * If this was a background laundering, count
 814                                  * activated pages towards our target.  The
 815                                  * purpose of background laundering is to ensure
 816                                  * that pages are eventually cycled through the
 817                                  * laundry queue, and an activation is a valid
 818                                  * way out.
 819                                  */
 820                                 if (!in_shortfall)
 821                                         launder--;
 822                                 goto drop_page;
 823                         } else if ((object->flags & OBJ_DEAD) == 0)
 824                                 goto requeue_page;
 825                 }
 826
 827                 /*
 828                  * If the page appears to be clean at the machine-independent
 829                  * layer, then remove all of its mappings from the pmap in
 830                  * anticipation of freeing it.  If, however, any of the page's
 831                  * mappings allow write access, then the page may still be
 832                  * modified until the last of those mappings are removed.
 833                  */
 834                 if (object->ref_count != 0) {
 835                         vm_page_test_dirty(m);
 836                         if (m->dirty == 0)
 837                                 pmap_remove_all(m);
 838                 }
 839
 840                 /*
 841                  * Clean pages are freed, and dirty pages are paged out unless
 842                  * they belong to a dead object.  Requeueing dirty pages from
 843                  * dead objects is pointless, as they are being paged out and
 844                  * freed by the thread that destroyed the object.
 845                  */
 846                 if (m->dirty == 0) {
 847 free_page:
 848                         vm_page_free(m);
 849                         VM_CNT_INC(v_dfree);
 850                 } else if ((object->flags & OBJ_DEAD) == 0) {
 851                         if (object->type != OBJT_SWAP &&
 852                             object->type != OBJT_DEFAULT)
 853                                 pageout_ok = true;
 854                         else if (disable_swap_pageouts)
 855                                 pageout_ok = false;
 856                         else
 857                                 pageout_ok = true;
 858                         if (!pageout_ok) {
 859 requeue_page:
 860                                 vm_pagequeue_lock(pq);
 861                                 queue_locked = true;
 862                                 vm_page_requeue_locked(m);
 863                                 goto drop_page;
 864                         }
 865
 866                         /*
 867                          * Form a cluster with adjacent, dirty pages from the
 868                          * same object, and page out that entire cluster.
 869                          *
 870                          * The adjacent, dirty pages must also be in the
 871                          * laundry.  However, their mappings are not checked
 872                          * for new references.  Consequently, a recently
 873                          * referenced page may be paged out.  However, that
 874                          * page will not be prematurely reclaimed.  After page
 875                          * out, the page will be placed in the inactive queue,
 876                          * where any new references will be detected and the
 877                          * page reactivated.
 878                          */
 879                         error = vm_pageout_clean(m, &numpagedout);
 880                         if (error == 0) {
 881                                 launder -= numpagedout;
 882                                 maxscan -= numpagedout - 1;
 883                         } else if (error == EDEADLK) {
 884                                 pageout_lock_miss++;
 885                                 vnodes_skipped++;
 886                         }
 887                         goto relock_queue;
 888                 }
 889 drop_page:
 890                 vm_page_unlock(m);
 891                 VM_OBJECT_WUNLOCK(object);
 892 relock_queue:
 893                 if (!queue_locked) {
 894                         vm_pagequeue_lock(pq);
 895                         queue_locked = true;
 896                 }
 897                 next = TAILQ_NEXT(&vmd->vmd_laundry_marker, plinks.q);
 898                 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_laundry_marker, plinks.q);
 899         }
 900         vm_pagequeue_unlock(pq);
 901
 902         if (launder > 0 && pq == &vmd->vmd_pagequeues[PQ_UNSWAPPABLE]) {
 903                 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
 904                 goto scan;
 905         }
 906
 907         /*
 908          * Wakeup the sync daemon if we skipped a vnode in a writeable object
 909          * and we didn't launder enough pages.
 910          */
 911         if (vnodes_skipped > 0 && launder > 0)
 912                 (void)speedup_syncer();
 913
 914         return (starting_target - launder);
 915 }
 916
 917 /*
 918  * Compute the integer square root.
 919  */
 920 static u_int
 921 isqrt(u_int num)
 922 {
 923         u_int bit, root, tmp;
 924
 925         bit = 1u << ((NBBY * sizeof(u_int)) - 2);
 926         while (bit > num)
 927                 bit >>= 2;
 928         root = 0;
 929         while (bit != 0) {
 930                 tmp = root + bit;
 931                 root >>= 1;
 932                 if (num >= tmp) {
 933                         num -= tmp;
 934                         root += bit;
 935                 }
 936                 bit >>= 2;
 937         }
 938         return (root);
 939 }
 940
 941 /*
 942  * Perform the work of the laundry thread: periodically wake up and determine
 943  * whether any pages need to be laundered.  If so, determine the number of pages
 944  * that need to be laundered, and launder them.
 945  */
 946 static void
 947 vm_pageout_laundry_worker(void *arg)
 948 {
 949         struct vm_domain *domain;
 950         struct vm_pagequeue *pq;
 951         uint64_t nclean, ndirty;
 952         u_int last_launder, wakeups;
 953         int domidx, last_target, launder, shortfall, shortfall_cycle, target;
 954         bool in_shortfall;
 955
 956         domidx = (uintptr_t)arg;
 957         domain = &vm_dom[domidx];
 958         pq = &domain->vmd_pagequeues[PQ_LAUNDRY];
 959         KASSERT(domain->vmd_segs != 0, ("domain without segments"));
 960         vm_pageout_init_marker(&domain->vmd_laundry_marker, PQ_LAUNDRY);
 961
 962         shortfall = 0;
 963         in_shortfall = false;
 964         shortfall_cycle = 0;
 965         target = 0;
 966         last_launder = 0;
 967
 968         /*
 969          * Calls to these handlers are serialized by the swap syscall lock.
 970          */
 971         (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, domain,
 972             EVENTHANDLER_PRI_ANY);
 973         (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, domain,
 974             EVENTHANDLER_PRI_ANY);
 975
 976         /*
 977          * The pageout laundry worker is never done, so loop forever.
 978          */
 979         for (;;) {
 980                 KASSERT(target >= 0, ("negative target %d", target));
 981                 KASSERT(shortfall_cycle >= 0,
 982                     ("negative cycle %d", shortfall_cycle));
 983                 launder = 0;
 984                 wakeups = VM_CNT_FETCH(v_pdwakeups);
 985
 986                 /*
 987                  * First determine whether we need to launder pages to meet a
 988                  * shortage of free pages.
 989                  */
 990                 if (shortfall > 0) {
 991                         in_shortfall = true;
 992                         shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
 993                         target = shortfall;
 994                 } else if (!in_shortfall)
 995                         goto trybackground;
 996                 else if (shortfall_cycle == 0 || vm_laundry_target() <= 0) {
 997                         /*
 998                          * We recently entered shortfall and began laundering
 999                          * pages.  If we have completed that laundering run
1000                          * (and we are no longer in shortfall) or we have met
1001                          * our laundry target through other activity, then we
1002                          * can stop laundering pages.
1003                          */
1004                         in_shortfall = false;
1005                         target = 0;
1006                         goto trybackground;
1007                 }
1008                 last_launder = wakeups;
1009                 launder = target / shortfall_cycle--;
1010                 goto dolaundry;
1011
1012                 /*
1013                  * There's no immediate need to launder any pages; see if we
1014                  * meet the conditions to perform background laundering:
1015                  *
1016                  * 1. The ratio of dirty to clean inactive pages exceeds the
1017                  *    background laundering threshold and the pagedaemon has
1018                  *    been woken up to reclaim pages since our last
1019                  *    laundering, or
1020                  * 2. we haven't yet reached the target of the current
1021                  *    background laundering run.
1022                  *
1023                  * The background laundering threshold is not a constant.
1024                  * Instead, it is a slowly growing function of the number of
1025                  * page daemon wakeups since the last laundering.  Thus, as the
1026                  * ratio of dirty to clean inactive pages grows, the amount of
1027                  * memory pressure required to trigger laundering decreases.
1028                  */
1029 trybackground:
1030                 nclean = vm_cnt.v_inactive_count + vm_cnt.v_free_count;
1031                 ndirty = vm_cnt.v_laundry_count;
1032                 if (target == 0 && wakeups != last_launder &&
1033                     ndirty * isqrt(wakeups - last_launder) >= nclean) {
1034                         target = vm_background_launder_target;
1035                 }
1036
1037                 /*
1038                  * We have a non-zero background laundering target.  If we've
1039                  * laundered up to our maximum without observing a page daemon
1040                  * wakeup, just stop.  This is a safety belt that ensures we
1041                  * don't launder an excessive amount if memory pressure is low
1042                  * and the ratio of dirty to clean pages is large.  Otherwise,
1043                  * proceed at the background laundering rate.
1044                  */
1045                 if (target > 0) {
1046                         if (wakeups != last_launder) {
1047                                 last_launder = wakeups;
1048                                 last_target = target;
1049                         } else if (last_target - target >=
1050                             vm_background_launder_max * PAGE_SIZE / 1024) {
1051                                 target = 0;
1052                         }
1053                         launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1054                         launder /= VM_LAUNDER_RATE;
1055                         if (launder > target)
1056                                 launder = target;
1057                 }
1058
1059 dolaundry:
1060                 if (launder > 0) {
1061                         /*
1062                          * Because of I/O clustering, the number of laundered
1063                          * pages could exceed "target" by the maximum size of
1064                          * a cluster minus one.
1065                          */
1066                         target -= min(vm_pageout_launder(domain, launder,
1067                             in_shortfall), target);
1068                         pause("laundp", hz / VM_LAUNDER_RATE);
1069                 }
1070
1071                 /*
1072                  * If we're not currently laundering pages and the page daemon
1073                  * hasn't posted a new request, sleep until the page daemon
1074                  * kicks us.
1075                  */
1076                 vm_pagequeue_lock(pq);
1077                 if (target == 0 && vm_laundry_request == VM_LAUNDRY_IDLE)
1078                         (void)mtx_sleep(&vm_laundry_request,
1079                             vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1080
1081                 /*
1082                  * If the pagedaemon has indicated that it's in shortfall, start
1083                  * a shortfall laundering unless we're already in the middle of
1084                  * one.  This may preempt a background laundering.
1085                  */
1086                 if (vm_laundry_request == VM_LAUNDRY_SHORTFALL &&
1087                     (!in_shortfall || shortfall_cycle == 0)) {
1088                         shortfall = vm_laundry_target() + vm_pageout_deficit;
1089                         target = 0;
1090                 } else
1091                         shortfall = 0;
1092
1093                 if (target == 0)
1094                         vm_laundry_request = VM_LAUNDRY_IDLE;
1095                 vm_pagequeue_unlock(pq);
1096         }
1097 }
1098
1099 /*
1100  *      vm_pageout_scan does the dirty work for the pageout daemon.
1101  *
1102  *      pass == 0: Update active LRU/deactivate pages
1103  *      pass >= 1: Free inactive pages
1104  *
1105  * Returns true if pass was zero or enough pages were freed by the inactive
1106  * queue scan to meet the target.
1107  */
1108 static bool
1109 vm_pageout_scan(struct vm_domain *vmd, int pass)
1110 {
1111         vm_page_t m, next;
1112         struct vm_pagequeue *pq;
1113         vm_object_t object;
1114         long min_scan;
1115         int act_delta, addl_page_shortage, deficit, inactq_shortage, maxscan;
1116         int page_shortage, scan_tick, scanned, starting_page_shortage;
1117         boolean_t queue_locked;
1118
1119         /*
1120          * If we need to reclaim memory ask kernel caches to return
1121          * some.  We rate limit to avoid thrashing.
1122          */
1123         if (vmd == &vm_dom[0] && pass > 0 &&
1124             (time_uptime - lowmem_uptime) >= lowmem_period) {
1125                 /*
1126                  * Decrease registered cache sizes.
1127                  */
1128                 SDT_PROBE0(vm, , , vm__lowmem_scan);
1129                 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1130                 /*
1131                  * We do this explicitly after the caches have been
1132                  * drained above.
1133                  */
1134                 uma_reclaim();
1135                 lowmem_uptime = time_uptime;
1136         }
1137
1138         /*
1139          * The addl_page_shortage is the number of temporarily
1140          * stuck pages in the inactive queue.  In other words, the
1141          * number of pages from the inactive count that should be
1142          * discounted in setting the target for the active queue scan.
1143          */
1144         addl_page_shortage = 0;
1145
1146         /*
1147          * Calculate the number of pages that we want to free.  This number
1148          * can be negative if many pages are freed between the wakeup call to
1149          * the page daemon and this calculation.
1150          */
1151         if (pass > 0) {
1152                 deficit = atomic_readandclear_int(&vm_pageout_deficit);
1153                 page_shortage = vm_paging_target() + deficit;
1154         } else
1155                 page_shortage = deficit = 0;
1156         starting_page_shortage = page_shortage;
1157
1158         /*
1159          * Start scanning the inactive queue for pages that we can free.  The
1160          * scan will stop when we reach the target or we have scanned the
1161          * entire queue.  (Note that m->act_count is not used to make
1162          * decisions for the inactive queue, only for the active queue.)
1163          */
1164         pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1165         maxscan = pq->pq_cnt;
1166         vm_pagequeue_lock(pq);
1167         queue_locked = TRUE;
1168         for (m = TAILQ_FIRST(&pq->pq_pl);
1169              m != NULL && maxscan-- > 0 && page_shortage > 0;
1170              m = next) {
1171                 vm_pagequeue_assert_locked(pq);
1172                 KASSERT(queue_locked, ("unlocked inactive queue"));
1173                 KASSERT(vm_page_inactive(m), ("Inactive queue %p", m));
1174
1175                 VM_CNT_INC(v_pdpages);
1176                 next = TAILQ_NEXT(m, plinks.q);
1177
1178                 /*
1179                  * skip marker pages
1180                  */
1181                 if (m->flags & PG_MARKER)
1182                         continue;
1183
1184                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1185                     ("Fictitious page %p cannot be in inactive queue", m));
1186                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1187                     ("Unmanaged page %p cannot be in inactive queue", m));
1188
1189                 /*
1190                  * The page or object lock acquisitions fail if the
1191                  * page was removed from the queue or moved to a
1192                  * different position within the queue.  In either
1193                  * case, addl_page_shortage should not be incremented.
1194                  */
1195                 if (!vm_pageout_page_lock(m, &next))
1196                         goto unlock_page;
1197                 else if (m->hold_count != 0) {
1198                         /*
1199                          * Held pages are essentially stuck in the
1200                          * queue.  So, they ought to be discounted
1201                          * from the inactive count.  See the
1202                          * calculation of inactq_shortage before the
1203                          * loop over the active queue below.
1204                          */
1205                         addl_page_shortage++;
1206                         goto unlock_page;
1207                 }
1208                 object = m->object;
1209                 if (!VM_OBJECT_TRYWLOCK(object)) {
1210                         if (!vm_pageout_fallback_object_lock(m, &next))
1211                                 goto unlock_object;
1212                         else if (m->hold_count != 0) {
1213                                 addl_page_shortage++;
1214                                 goto unlock_object;
1215                         }
1216                 }
1217                 if (vm_page_busied(m)) {
1218                         /*
1219                          * Don't mess with busy pages.  Leave them at
1220                          * the front of the queue.  Most likely, they
1221                          * are being paged out and will leave the
1222                          * queue shortly after the scan finishes.  So,
1223                          * they ought to be discounted from the
1224                          * inactive count.
1225                          */
1226                         addl_page_shortage++;
1227 unlock_object:
1228                         VM_OBJECT_WUNLOCK(object);
1229 unlock_page:
1230                         vm_page_unlock(m);
1231                         continue;
1232                 }
1233                 KASSERT(m->hold_count == 0, ("Held page %p", m));
1234
1235                 /*
1236                  * Dequeue the inactive page and unlock the inactive page
1237                  * queue, invalidating the 'next' pointer.  Dequeueing the
1238                  * page here avoids a later reacquisition (and release) of
1239                  * the inactive page queue lock when vm_page_activate(),
1240                  * vm_page_free(), or vm_page_launder() is called.  Use a
1241                  * marker to remember our place in the inactive queue.
1242                  */
1243                 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1244                 vm_page_dequeue_locked(m);
1245                 vm_pagequeue_unlock(pq);
1246                 queue_locked = FALSE;
1247
1248                 /*
1249                  * Invalid pages can be easily freed. They cannot be
1250                  * mapped, vm_page_free() asserts this.
1251                  */
1252                 if (m->valid == 0)
1253                         goto free_page;
1254
1255                 /*
1256                  * If the page has been referenced and the object is not dead,
1257                  * reactivate or requeue the page depending on whether the
1258                  * object is mapped.
1259                  */
1260                 if ((m->aflags & PGA_REFERENCED) != 0) {
1261                         vm_page_aflag_clear(m, PGA_REFERENCED);
1262                         act_delta = 1;
1263                 } else
1264                         act_delta = 0;
1265                 if (object->ref_count != 0) {
1266                         act_delta += pmap_ts_referenced(m);
1267                 } else {
1268                         KASSERT(!pmap_page_is_mapped(m),
1269                             ("vm_pageout_scan: page %p is mapped", m));
1270                 }
1271                 if (act_delta != 0) {
1272                         if (object->ref_count != 0) {
1273                                 VM_CNT_INC(v_reactivated);
1274                                 vm_page_activate(m);
1275
1276                                 /*
1277                                  * Increase the activation count if the page
1278                                  * was referenced while in the inactive queue.
1279                                  * This makes it less likely that the page will
1280                                  * be returned prematurely to the inactive
1281                                  * queue.
1282                                  */
1283                                 m->act_count += act_delta + ACT_ADVANCE;
1284                                 goto drop_page;
1285                         } else if ((object->flags & OBJ_DEAD) == 0) {
1286                                 vm_pagequeue_lock(pq);
1287                                 queue_locked = TRUE;
1288                                 m->queue = PQ_INACTIVE;
1289                                 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
1290                                 vm_pagequeue_cnt_inc(pq);
1291                                 goto drop_page;
1292                         }
1293                 }
1294
1295                 /*
1296                  * If the page appears to be clean at the machine-independent
1297                  * layer, then remove all of its mappings from the pmap in
1298                  * anticipation of freeing it.  If, however, any of the page's
1299                  * mappings allow write access, then the page may still be
1300                  * modified until the last of those mappings are removed.
1301                  */
1302                 if (object->ref_count != 0) {
1303                         vm_page_test_dirty(m);
1304                         if (m->dirty == 0)
1305                                 pmap_remove_all(m);
1306                 }
1307
1308                 /*
1309                  * Clean pages can be freed, but dirty pages must be sent back
1310                  * to the laundry, unless they belong to a dead object.
1311                  * Requeueing dirty pages from dead objects is pointless, as
1312                  * they are being paged out and freed by the thread that
1313                  * destroyed the object.
1314                  */
1315                 if (m->dirty == 0) {
1316 free_page:
1317                         vm_page_free(m);
1318                         VM_CNT_INC(v_dfree);
1319                         --page_shortage;
1320                 } else if ((object->flags & OBJ_DEAD) == 0)
1321                         vm_page_launder(m);
1322 drop_page:
1323                 vm_page_unlock(m);
1324                 VM_OBJECT_WUNLOCK(object);
1325                 if (!queue_locked) {
1326                         vm_pagequeue_lock(pq);
1327                         queue_locked = TRUE;
1328                 }
1329                 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1330                 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1331         }
1332         vm_pagequeue_unlock(pq);
1333
1334         /*
1335          * Wake up the laundry thread so that it can perform any needed
1336          * laundering.  If we didn't meet our target, we're in shortfall and
1337          * need to launder more aggressively.  If PQ_LAUNDRY is empty and no
1338          * swap devices are configured, the laundry thread has no work to do, so
1339          * don't bother waking it up.
1340          */
1341         if (vm_laundry_request == VM_LAUNDRY_IDLE &&
1342             starting_page_shortage > 0) {
1343                 pq = &vm_dom[0].vmd_pagequeues[PQ_LAUNDRY];
1344                 vm_pagequeue_lock(pq);
1345                 if (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled)) {
1346                         if (page_shortage > 0) {
1347                                 vm_laundry_request = VM_LAUNDRY_SHORTFALL;
1348                                 VM_CNT_INC(v_pdshortfalls);
1349                         } else if (vm_laundry_request != VM_LAUNDRY_SHORTFALL)
1350                                 vm_laundry_request = VM_LAUNDRY_BACKGROUND;
1351                         wakeup(&vm_laundry_request);
1352                 }
1353                 vm_pagequeue_unlock(pq);
1354         }
1355
1356         /*
1357          * Wakeup the swapout daemon if we didn't free the targeted number of
1358          * pages.
1359          */
1360         if (page_shortage > 0)
1361                 vm_swapout_run();
1362
1363         /*
1364          * If the inactive queue scan fails repeatedly to meet its
1365          * target, kill the largest process.
1366          */
1367         vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1368
1369         /*
1370          * Compute the number of pages we want to try to move from the
1371          * active queue to either the inactive or laundry queue.
1372          *
1373          * When scanning active pages, we make clean pages count more heavily
1374          * towards the page shortage than dirty pages.  This is because dirty
1375          * pages must be laundered before they can be reused and thus have less
1376          * utility when attempting to quickly alleviate a shortage.  However,
1377          * this weighting also causes the scan to deactivate dirty pages more
1378          * more aggressively, improving the effectiveness of clustering and
1379          * ensuring that they can eventually be reused.
1380          */
1381         inactq_shortage = vm_cnt.v_inactive_target - (vm_cnt.v_inactive_count +
1382             vm_cnt.v_laundry_count / act_scan_laundry_weight) +
1383             vm_paging_target() + deficit + addl_page_shortage;
1384         page_shortage *= act_scan_laundry_weight;
1385
1386         pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1387         vm_pagequeue_lock(pq);
1388         maxscan = pq->pq_cnt;
1389
1390         /*
1391          * If we're just idle polling attempt to visit every
1392          * active page within 'update_period' seconds.
1393          */
1394         scan_tick = ticks;
1395         if (vm_pageout_update_period != 0) {
1396                 min_scan = pq->pq_cnt;
1397                 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1398                 min_scan /= hz * vm_pageout_update_period;
1399         } else
1400                 min_scan = 0;
1401         if (min_scan > 0 || (inactq_shortage > 0 && maxscan > 0))
1402                 vmd->vmd_last_active_scan = scan_tick;
1403
1404         /*
1405          * Scan the active queue for pages that can be deactivated.  Update
1406          * the per-page activity counter and use it to identify deactivation
1407          * candidates.  Held pages may be deactivated.
1408          */
1409         for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1410             min_scan || (inactq_shortage > 0 && scanned < maxscan)); m = next,
1411             scanned++) {
1412                 KASSERT(m->queue == PQ_ACTIVE,
1413                     ("vm_pageout_scan: page %p isn't active", m));
1414                 next = TAILQ_NEXT(m, plinks.q);
1415                 if ((m->flags & PG_MARKER) != 0)
1416                         continue;
1417                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1418                     ("Fictitious page %p cannot be in active queue", m));
1419                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1420                     ("Unmanaged page %p cannot be in active queue", m));
1421                 if (!vm_pageout_page_lock(m, &next)) {
1422                         vm_page_unlock(m);
1423                         continue;
1424                 }
1425
1426                 /*
1427                  * The count for page daemon pages is updated after checking
1428                  * the page for eligibility.
1429                  */
1430                 VM_CNT_INC(v_pdpages);
1431
1432                 /*
1433                  * Check to see "how much" the page has been used.
1434                  */
1435                 if ((m->aflags & PGA_REFERENCED) != 0) {
1436                         vm_page_aflag_clear(m, PGA_REFERENCED);
1437                         act_delta = 1;
1438                 } else
1439                         act_delta = 0;
1440
1441                 /*
1442                  * Perform an unsynchronized object ref count check.  While
1443                  * the page lock ensures that the page is not reallocated to
1444                  * another object, in particular, one with unmanaged mappings
1445                  * that cannot support pmap_ts_referenced(), two races are,
1446                  * nonetheless, possible:
1447                  * 1) The count was transitioning to zero, but we saw a non-
1448                  *    zero value.  pmap_ts_referenced() will return zero
1449                  *    because the page is not mapped.
1450                  * 2) The count was transitioning to one, but we saw zero.
1451                  *    This race delays the detection of a new reference.  At
1452                  *    worst, we will deactivate and reactivate the page.
1453                  */
1454                 if (m->object->ref_count != 0)
1455                         act_delta += pmap_ts_referenced(m);
1456
1457                 /*
1458                  * Advance or decay the act_count based on recent usage.
1459                  */
1460                 if (act_delta != 0) {
1461                         m->act_count += ACT_ADVANCE + act_delta;
1462                         if (m->act_count > ACT_MAX)
1463                                 m->act_count = ACT_MAX;
1464                 } else
1465                         m->act_count -= min(m->act_count, ACT_DECLINE);
1466
1467                 /*
1468                  * Move this page to the tail of the active, inactive or laundry
1469                  * queue depending on usage.
1470                  */
1471                 if (m->act_count == 0) {
1472                         /* Dequeue to avoid later lock recursion. */
1473                         vm_page_dequeue_locked(m);
1474
1475                         /*
1476                          * When not short for inactive pages, let dirty pages go
1477                          * through the inactive queue before moving to the
1478                          * laundry queues.  This gives them some extra time to
1479                          * be reactivated, potentially avoiding an expensive
1480                          * pageout.  During a page shortage, the inactive queue
1481                          * is necessarily small, so we may move dirty pages
1482                          * directly to the laundry queue.
1483                          */
1484                         if (inactq_shortage <= 0)
1485                                 vm_page_deactivate(m);
1486                         else {
1487                                 /*
1488                                  * Calling vm_page_test_dirty() here would
1489                                  * require acquisition of the object's write
1490                                  * lock.  However, during a page shortage,
1491                                  * directing dirty pages into the laundry
1492                                  * queue is only an optimization and not a
1493                                  * requirement.  Therefore, we simply rely on
1494                                  * the opportunistic updates to the page's
1495                                  * dirty field by the pmap.
1496                                  */
1497                                 if (m->dirty == 0) {
1498                                         vm_page_deactivate(m);
1499                                         inactq_shortage -=
1500                                             act_scan_laundry_weight;
1501                                 } else {
1502                                         vm_page_launder(m);
1503                                         inactq_shortage--;
1504                                 }
1505                         }
1506                 } else
1507                         vm_page_requeue_locked(m);
1508                 vm_page_unlock(m);
1509         }
1510         vm_pagequeue_unlock(pq);
1511         if (pass > 0)
1512                 vm_swapout_run_idle();
1513         return (page_shortage <= 0);
1514 }
1515
1516 static int vm_pageout_oom_vote;
1517
1518 /*
1519  * The pagedaemon threads randlomly select one to perform the
1520  * OOM.  Trying to kill processes before all pagedaemons
1521  * failed to reach free target is premature.
1522  */
1523 static void
1524 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1525     int starting_page_shortage)
1526 {
1527         int old_vote;
1528
1529         if (starting_page_shortage <= 0 || starting_page_shortage !=
1530             page_shortage)
1531                 vmd->vmd_oom_seq = 0;
1532         else
1533                 vmd->vmd_oom_seq++;
1534         if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1535                 if (vmd->vmd_oom) {
1536                         vmd->vmd_oom = FALSE;
1537                         atomic_subtract_int(&vm_pageout_oom_vote, 1);
1538                 }
1539                 return;
1540         }
1541
1542         /*
1543          * Do not follow the call sequence until OOM condition is
1544          * cleared.
1545          */
1546         vmd->vmd_oom_seq = 0;
1547
1548         if (vmd->vmd_oom)
1549                 return;
1550
1551         vmd->vmd_oom = TRUE;
1552         old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1553         if (old_vote != vm_ndomains - 1)
1554                 return;
1555
1556         /*
1557          * The current pagedaemon thread is the last in the quorum to
1558          * start OOM.  Initiate the selection and signaling of the
1559          * victim.
1560          */
1561         vm_pageout_oom(VM_OOM_MEM);
1562
1563         /*
1564          * After one round of OOM terror, recall our vote.  On the
1565          * next pass, current pagedaemon would vote again if the low
1566          * memory condition is still there, due to vmd_oom being
1567          * false.
1568          */
1569         vmd->vmd_oom = FALSE;
1570         atomic_subtract_int(&vm_pageout_oom_vote, 1);
1571 }
1572
1573 /*
1574  * The OOM killer is the page daemon's action of last resort when
1575  * memory allocation requests have been stalled for a prolonged period
1576  * of time because it cannot reclaim memory.  This function computes
1577  * the approximate number of physical pages that could be reclaimed if
1578  * the specified address space is destroyed.
1579  *
1580  * Private, anonymous memory owned by the address space is the
1581  * principal resource that we expect to recover after an OOM kill.
1582  * Since the physical pages mapped by the address space's COW entries
1583  * are typically shared pages, they are unlikely to be released and so
1584  * they are not counted.
1585  *
1586  * To get to the point where the page daemon runs the OOM killer, its
1587  * efforts to write-back vnode-backed pages may have stalled.  This
1588  * could be caused by a memory allocation deadlock in the write path
1589  * that might be resolved by an OOM kill.  Therefore, physical pages
1590  * belonging to vnode-backed objects are counted, because they might
1591  * be freed without being written out first if the address space holds
1592  * the last reference to an unlinked vnode.
1593  *
1594  * Similarly, physical pages belonging to OBJT_PHYS objects are
1595  * counted because the address space might hold the last reference to
1596  * the object.
1597  */
1598 static long
1599 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1600 {
1601         vm_map_t map;
1602         vm_map_entry_t entry;
1603         vm_object_t obj;
1604         long res;
1605
1606         map = &vmspace->vm_map;
1607         KASSERT(!map->system_map, ("system map"));
1608         sx_assert(&map->lock, SA_LOCKED);
1609         res = 0;
1610         for (entry = map->header.next; entry != &map->header;
1611             entry = entry->next) {
1612                 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1613                         continue;
1614                 obj = entry->object.vm_object;
1615                 if (obj == NULL)
1616                         continue;
1617                 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1618                     obj->ref_count != 1)
1619                         continue;
1620                 switch (obj->type) {
1621                 case OBJT_DEFAULT:
1622                 case OBJT_SWAP:
1623                 case OBJT_PHYS:
1624                 case OBJT_VNODE:
1625                         res += obj->resident_page_count;
1626                         break;
1627                 }
1628         }
1629         return (res);
1630 }
1631
1632 void
1633 vm_pageout_oom(int shortage)
1634 {
1635         struct proc *p, *bigproc;
1636         vm_offset_t size, bigsize;
1637         struct thread *td;
1638         struct vmspace *vm;
1639         bool breakout;
1640
1641         /*
1642          * We keep the process bigproc locked once we find it to keep anyone
1643          * from messing with it; however, there is a possibility of
1644          * deadlock if process B is bigproc and one of its child processes
1645          * attempts to propagate a signal to B while we are waiting for A's
1646          * lock while walking this list.  To avoid this, we don't block on
1647          * the process lock but just skip a process if it is already locked.
1648          */
1649         bigproc = NULL;
1650         bigsize = 0;
1651         sx_slock(&allproc_lock);
1652         FOREACH_PROC_IN_SYSTEM(p) {
1653                 PROC_LOCK(p);
1654
1655                 /*
1656                  * If this is a system, protected or killed process, skip it.
1657                  */
1658                 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1659                     P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1660                     p->p_pid == 1 || P_KILLED(p) ||
1661                     (p->p_pid < 48 && swap_pager_avail != 0)) {
1662                         PROC_UNLOCK(p);
1663                         continue;
1664                 }
1665                 /*
1666                  * If the process is in a non-running type state,
1667                  * don't touch it.  Check all the threads individually.
1668                  */
1669                 breakout = false;
1670                 FOREACH_THREAD_IN_PROC(p, td) {
1671                         thread_lock(td);
1672                         if (!TD_ON_RUNQ(td) &&
1673                             !TD_IS_RUNNING(td) &&
1674                             !TD_IS_SLEEPING(td) &&
1675                             !TD_IS_SUSPENDED(td) &&
1676                             !TD_IS_SWAPPED(td)) {
1677                                 thread_unlock(td);
1678                                 breakout = true;
1679                                 break;
1680                         }
1681                         thread_unlock(td);
1682                 }
1683                 if (breakout) {
1684                         PROC_UNLOCK(p);
1685                         continue;
1686                 }
1687                 /*
1688                  * get the process size
1689                  */
1690                 vm = vmspace_acquire_ref(p);
1691                 if (vm == NULL) {
1692                         PROC_UNLOCK(p);
1693                         continue;
1694                 }
1695                 _PHOLD_LITE(p);
1696                 PROC_UNLOCK(p);
1697                 sx_sunlock(&allproc_lock);
1698                 if (!vm_map_trylock_read(&vm->vm_map)) {
1699                         vmspace_free(vm);
1700                         sx_slock(&allproc_lock);
1701                         PRELE(p);
1702                         continue;
1703                 }
1704                 size = vmspace_swap_count(vm);
1705                 if (shortage == VM_OOM_MEM)
1706                         size += vm_pageout_oom_pagecount(vm);
1707                 vm_map_unlock_read(&vm->vm_map);
1708                 vmspace_free(vm);
1709                 sx_slock(&allproc_lock);
1710
1711                 /*
1712                  * If this process is bigger than the biggest one,
1713                  * remember it.
1714                  */
1715                 if (size > bigsize) {
1716                         if (bigproc != NULL)
1717                                 PRELE(bigproc);
1718                         bigproc = p;
1719                         bigsize = size;
1720                 } else {
1721                         PRELE(p);
1722                 }
1723         }
1724         sx_sunlock(&allproc_lock);
1725         if (bigproc != NULL) {
1726                 if (vm_panic_on_oom != 0)
1727                         panic("out of swap space");
1728                 PROC_LOCK(bigproc);
1729                 killproc(bigproc, "out of swap space");
1730                 sched_nice(bigproc, PRIO_MIN);
1731                 _PRELE(bigproc);
1732                 PROC_UNLOCK(bigproc);
1733                 wakeup(&vm_cnt.v_free_count);
1734         }
1735 }
1736
1737 static void
1738 vm_pageout_worker(void *arg)
1739 {
1740         struct vm_domain *domain;
1741         int domidx, pass;
1742         bool target_met;
1743
1744         domidx = (uintptr_t)arg;
1745         domain = &vm_dom[domidx];
1746         pass = 0;
1747         target_met = true;
1748
1749         /*
1750          * XXXKIB It could be useful to bind pageout daemon threads to
1751          * the cores belonging to the domain, from which vm_page_array
1752          * is allocated.
1753          */
1754
1755         KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1756         domain->vmd_last_active_scan = ticks;
1757         vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1758         vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE);
1759         TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl,
1760             &domain->vmd_inacthead, plinks.q);
1761
1762         /*
1763          * The pageout daemon worker is never done, so loop forever.
1764          */
1765         while (TRUE) {
1766                 mtx_lock(&vm_page_queue_free_mtx);
1767
1768                 /*
1769                  * Generally, after a level >= 1 scan, if there are enough
1770                  * free pages to wakeup the waiters, then they are already
1771                  * awake.  A call to vm_page_free() during the scan awakened
1772                  * them.  However, in the following case, this wakeup serves
1773                  * to bound the amount of time that a thread might wait.
1774                  * Suppose a thread's call to vm_page_alloc() fails, but
1775                  * before that thread calls VM_WAIT, enough pages are freed by
1776                  * other threads to alleviate the free page shortage.  The
1777                  * thread will, nonetheless, wait until another page is freed
1778                  * or this wakeup is performed.
1779                  */
1780                 if (vm_pages_needed && !vm_page_count_min()) {
1781                         vm_pages_needed = false;
1782                         wakeup(&vm_cnt.v_free_count);
1783                 }
1784
1785                 /*
1786                  * Do not clear vm_pageout_wanted until we reach our free page
1787                  * target.  Otherwise, we may be awakened over and over again,
1788                  * wasting CPU time.
1789                  */
1790                 if (vm_pageout_wanted && target_met)
1791                         vm_pageout_wanted = false;
1792
1793                 /*
1794                  * Might the page daemon receive a wakeup call?
1795                  */
1796                 if (vm_pageout_wanted) {
1797                         /*
1798                          * No.  Either vm_pageout_wanted was set by another
1799                          * thread during the previous scan, which must have
1800                          * been a level 0 scan, or vm_pageout_wanted was
1801                          * already set and the scan failed to free enough
1802                          * pages.  If we haven't yet performed a level >= 1
1803                          * (page reclamation) scan, then increase the level
1804                          * and scan again now.  Otherwise, sleep a bit and
1805                          * try again later.
1806                          */
1807                         mtx_unlock(&vm_page_queue_free_mtx);
1808                         if (pass >= 1)
1809                                 pause("psleep", hz / VM_INACT_SCAN_RATE);
1810                         pass++;
1811                 } else {
1812                         /*
1813                          * Yes.  Sleep until pages need to be reclaimed or
1814                          * have their reference stats updated.
1815                          */
1816                         if (mtx_sleep(&vm_pageout_wanted,
1817                             &vm_page_queue_free_mtx, PDROP | PVM, "psleep",
1818                             hz) == 0) {
1819                                 VM_CNT_INC(v_pdwakeups);
1820                                 pass = 1;
1821                         } else
1822                                 pass = 0;
1823                 }
1824
1825                 target_met = vm_pageout_scan(domain, pass);
1826         }
1827 }
1828
1829 /*
1830  *      vm_pageout_init initialises basic pageout daemon settings.
1831  */
1832 static void
1833 vm_pageout_init(void)
1834 {
1835         /*
1836          * Initialize some paging parameters.
1837          */
1838         vm_cnt.v_interrupt_free_min = 2;
1839         if (vm_cnt.v_page_count < 2000)
1840                 vm_pageout_page_count = 8;
1841
1842         /*
1843          * v_free_reserved needs to include enough for the largest
1844          * swap pager structures plus enough for any pv_entry structs
1845          * when paging.
1846          */
1847         if (vm_cnt.v_page_count > 1024)
1848                 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
1849         else
1850                 vm_cnt.v_free_min = 4;
1851         vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1852             vm_cnt.v_interrupt_free_min;
1853         vm_cnt.v_free_reserved = vm_pageout_page_count +
1854             vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
1855         vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
1856         vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
1857         vm_cnt.v_free_min += vm_cnt.v_free_reserved;
1858         vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
1859         vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
1860         if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
1861                 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
1862
1863         /*
1864          * Set the default wakeup threshold to be 10% above the minimum
1865          * page limit.  This keeps the steady state out of shortfall.
1866          */
1867         vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
1868
1869         /*
1870          * Set interval in seconds for active scan.  We want to visit each
1871          * page at least once every ten minutes.  This is to prevent worst
1872          * case paging behaviors with stale active LRU.
1873          */
1874         if (vm_pageout_update_period == 0)
1875                 vm_pageout_update_period = 600;
1876
1877         /* XXX does not really belong here */
1878         if (vm_page_max_wired == 0)
1879                 vm_page_max_wired = vm_cnt.v_free_count / 3;
1880
1881         /*
1882          * Target amount of memory to move out of the laundry queue during a
1883          * background laundering.  This is proportional to the amount of system
1884          * memory.
1885          */
1886         vm_background_launder_target = (vm_cnt.v_free_target -
1887             vm_cnt.v_free_min) / 10;
1888 }
1889
1890 /*
1891  *     vm_pageout is the high level pageout daemon.
1892  */
1893 static void
1894 vm_pageout(void)
1895 {
1896         int error;
1897 #ifdef VM_NUMA_ALLOC
1898         int i;
1899 #endif
1900
1901         swap_pager_swap_init();
1902         error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL,
1903             0, 0, "laundry: dom0");
1904         if (error != 0)
1905                 panic("starting laundry for domain 0, error %d", error);
1906 #ifdef VM_NUMA_ALLOC
1907         for (i = 1; i < vm_ndomains; i++) {
1908                 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1909                     curproc, NULL, 0, 0, "dom%d", i);
1910                 if (error != 0) {
1911                         panic("starting pageout for domain %d, error %d\n",
1912                             i, error);
1913                 }
1914         }
1915 #endif
1916         error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
1917             0, 0, "uma");
1918         if (error != 0)
1919                 panic("starting uma_reclaim helper, error %d\n", error);
1920         vm_pageout_worker((void *)(uintptr_t)0);
1921 }
1922
1923 /*
1924  * Unless the free page queue lock is held by the caller, this function
1925  * should be regarded as advisory.  Specifically, the caller should
1926  * not msleep() on &vm_cnt.v_free_count following this function unless
1927  * the free page queue lock is held until the msleep() is performed.
1928  */
1929 void
1930 pagedaemon_wakeup(void)
1931 {
1932
1933         if (!vm_pageout_wanted && curthread->td_proc != pageproc) {
1934                 vm_pageout_wanted = true;
1935                 wakeup(&vm_pageout_wanted);
1936         }
1937 }