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