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 <sys/param.h>
  80 #include <sys/systm.h>
  81 #include <sys/kernel.h>
  82 #include <sys/eventhandler.h>
  83 #include <sys/lock.h>
  84 #include <sys/mutex.h>
  85 #include <sys/proc.h>
  86 #include <sys/kthread.h>
  87 #include <sys/ktr.h>
  88 #include <sys/mount.h>
  89 #include <sys/racct.h>
  90 #include <sys/resourcevar.h>
  91 #include <sys/sched.h>
  92 #include <sys/signalvar.h>
  93 #include <sys/vnode.h>
  94 #include <sys/vmmeter.h>
  95 #include <sys/sx.h>
  96 #include <sys/sysctl.h>
  97
  98 #include <vm/vm.h>
  99 #include <vm/vm_param.h>
 100 #include <vm/vm_object.h>
 101 #include <vm/vm_page.h>
 102 #include <vm/vm_map.h>
 103 #include <vm/vm_pageout.h>
 104 #include <vm/vm_pager.h>
 105 #include <vm/swap_pager.h>
 106 #include <vm/vm_extern.h>
 107 #include <vm/uma.h>
 108
 109 /*
 110  * System initialization
 111  */
 112
 113 /* the kernel process "vm_pageout"*/
 114 static void vm_pageout(void);
 115 static int vm_pageout_clean(vm_page_t);
 116 static void vm_pageout_scan(int pass);
 117
 118 struct proc *pageproc;
 119
 120 static struct kproc_desc page_kp = {
 121         "pagedaemon",
 122         vm_pageout,
 123         &pageproc
 124 };
 125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
 126     &page_kp);
 127
 128 #if !defined(NO_SWAPPING)
 129 /* the kernel process "vm_daemon"*/
 130 static void vm_daemon(void);
 131 static struct   proc *vmproc;
 132
 133 static struct kproc_desc vm_kp = {
 134         "vmdaemon",
 135         vm_daemon,
 136         &vmproc
 137 };
 138 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
 139 #endif
 140
 141
 142 int vm_pages_needed;            /* Event on which pageout daemon sleeps */
 143 int vm_pageout_deficit;         /* Estimated number of pages deficit */
 144 int vm_pageout_pages_needed;    /* flag saying that the pageout daemon needs pages */
 145
 146 #if !defined(NO_SWAPPING)
 147 static int vm_pageout_req_swapout;      /* XXX */
 148 static int vm_daemon_needed;
 149 static struct mtx vm_daemon_mtx;
 150 /* Allow for use by vm_pageout before vm_daemon is initialized. */
 151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
 152 #endif
 153 static int vm_max_launder = 32;
 154 static int vm_pageout_stats_max;
 155 static int vm_pageout_stats;
 156 static int vm_pageout_stats_interval;
 157 static int vm_pageout_full_stats;
 158 static int vm_pageout_full_stats_interval;
 159 static int vm_pageout_algorithm;
 160 static int defer_swap_pageouts;
 161 static int disable_swap_pageouts;
 162
 163 #if defined(NO_SWAPPING)
 164 static int vm_swap_enabled = 0;
 165 static int vm_swap_idle_enabled = 0;
 166 #else
 167 static int vm_swap_enabled = 1;
 168 static int vm_swap_idle_enabled = 0;
 169 #endif
 170
 171 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
 172         CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
 173
 174 SYSCTL_INT(_vm, OID_AUTO, max_launder,
 175         CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
 176
 177 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
 178         CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
 179
 180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats,
 181         CTLFLAG_RD, &vm_pageout_stats, 0, "Number of partial stats scans");
 182
 183 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
 184         CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
 185
 186 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats,
 187         CTLFLAG_RD, &vm_pageout_full_stats, 0, "Number of full stats scans");
 188
 189 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
 190         CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
 191
 192 #if defined(NO_SWAPPING)
 193 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
 194         CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
 195 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
 196         CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
 197 #else
 198 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
 199         CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
 200 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
 201         CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
 202 #endif
 203
 204 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
 205         CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
 206
 207 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
 208         CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
 209
 210 static int pageout_lock_miss;
 211 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
 212         CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
 213
 214 #define VM_PAGEOUT_PAGE_COUNT 16
 215 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
 216
 217 int vm_page_max_wired;          /* XXX max # of wired pages system-wide */
 218 SYSCTL_INT(_vm, OID_AUTO, max_wired,
 219         CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
 220
 221 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
 222 static boolean_t vm_pageout_launder(int, int, vm_paddr_t, vm_paddr_t);
 223 #if !defined(NO_SWAPPING)
 224 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
 225 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
 226 static void vm_req_vmdaemon(int req);
 227 #endif
 228 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
 229 static void vm_pageout_page_stats(void);
 230
 231 /*
 232  * Initialize a dummy page for marking the caller's place in the specified
 233  * paging queue.  In principle, this function only needs to set the flag
 234  * PG_MARKER.  Nonetheless, it sets the flag VPO_BUSY and initializes the hold
 235  * count to one as safety precautions.
 236  */
 237 static void
 238 vm_pageout_init_marker(vm_page_t marker, u_short queue)
 239 {
 240
 241         bzero(marker, sizeof(*marker));
 242         marker->flags = PG_MARKER;
 243         marker->oflags = VPO_BUSY;
 244         marker->queue = queue;
 245         marker->hold_count = 1;
 246 }
 247
 248 /*
 249  * vm_pageout_fallback_object_lock:
 250  *
 251  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
 252  * known to have failed and page queue must be either PQ_ACTIVE or
 253  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
 254  * while locking the vm object.  Use marker page to detect page queue
 255  * changes and maintain notion of next page on page queue.  Return
 256  * TRUE if no changes were detected, FALSE otherwise.  vm object is
 257  * locked on return.
 258  *
 259  * This function depends on both the lock portion of struct vm_object
 260  * and normal struct vm_page being type stable.
 261  */
 262 static boolean_t
 263 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
 264 {
 265         struct vm_page marker;
 266         struct vm_pagequeue *pq;
 267         boolean_t unchanged;
 268         u_short queue;
 269         vm_object_t object;
 270
 271         queue = m->queue;
 272         vm_pageout_init_marker(&marker, queue);
 273         pq = &vm_pagequeues[queue];
 274         object = m->object;
 275
 276         TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, pageq);
 277         vm_pagequeue_unlock(pq);
 278         vm_page_unlock(m);
 279         VM_OBJECT_LOCK(object);
 280         vm_page_lock(m);
 281         vm_pagequeue_lock(pq);
 282
 283         /* Page queue might have changed. */
 284         *next = TAILQ_NEXT(&marker, pageq);
 285         unchanged = (m->queue == queue &&
 286                      m->object == object &&
 287                      &marker == TAILQ_NEXT(m, pageq));
 288         TAILQ_REMOVE(&pq->pq_pl, &marker, pageq);
 289         return (unchanged);
 290 }
 291
 292 /*
 293  * Lock the page while holding the page queue lock.  Use marker page
 294  * to detect page queue changes and maintain notion of next page on
 295  * page queue.  Return TRUE if no changes were detected, FALSE
 296  * otherwise.  The page is locked on return. The page queue lock might
 297  * be dropped and reacquired.
 298  *
 299  * This function depends on normal struct vm_page being type stable.
 300  */
 301 static boolean_t
 302 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
 303 {
 304         struct vm_page marker;
 305         struct vm_pagequeue *pq;
 306         boolean_t unchanged;
 307         u_short queue;
 308
 309         vm_page_lock_assert(m, MA_NOTOWNED);
 310         if (vm_page_trylock(m))
 311                 return (TRUE);
 312
 313         queue = m->queue;
 314         vm_pageout_init_marker(&marker, queue);
 315         pq = &vm_pagequeues[queue];
 316
 317         TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, pageq);
 318         vm_pagequeue_unlock(pq);
 319         vm_page_lock(m);
 320         vm_pagequeue_lock(pq);
 321
 322         /* Page queue might have changed. */
 323         *next = TAILQ_NEXT(&marker, pageq);
 324         unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq));
 325         TAILQ_REMOVE(&pq->pq_pl, &marker, pageq);
 326         return (unchanged);
 327 }
 328
 329 /*
 330  * vm_pageout_clean:
 331  *
 332  * Clean the page and remove it from the laundry.
 333  *
 334  * We set the busy bit to cause potential page faults on this page to
 335  * block.  Note the careful timing, however, the busy bit isn't set till
 336  * late and we cannot do anything that will mess with the page.
 337  */
 338 static int
 339 vm_pageout_clean(vm_page_t m)
 340 {
 341         vm_object_t object;
 342         vm_page_t mc[2*vm_pageout_page_count], pb, ps;
 343         int pageout_count;
 344         int ib, is, page_base;
 345         vm_pindex_t pindex = m->pindex;
 346
 347         vm_page_lock_assert(m, MA_OWNED);
 348         object = m->object;
 349         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 350
 351         /*
 352          * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
 353          * with the new swapper, but we could have serious problems paging
 354          * out other object types if there is insufficient memory.
 355          *
 356          * Unfortunately, checking free memory here is far too late, so the
 357          * check has been moved up a procedural level.
 358          */
 359
 360         /*
 361          * Can't clean the page if it's busy or held.
 362          */
 363         KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0,
 364             ("vm_pageout_clean: page %p is busy", m));
 365         KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
 366         vm_page_unlock(m);
 367
 368         mc[vm_pageout_page_count] = pb = ps = m;
 369         pageout_count = 1;
 370         page_base = vm_pageout_page_count;
 371         ib = 1;
 372         is = 1;
 373
 374         /*
 375          * Scan object for clusterable pages.
 376          *
 377          * We can cluster ONLY if: ->> the page is NOT
 378          * clean, wired, busy, held, or mapped into a
 379          * buffer, and one of the following:
 380          * 1) The page is inactive, or a seldom used
 381          *    active page.
 382          * -or-
 383          * 2) we force the issue.
 384          *
 385          * During heavy mmap/modification loads the pageout
 386          * daemon can really fragment the underlying file
 387          * due to flushing pages out of order and not trying
 388          * align the clusters (which leave sporatic out-of-order
 389          * holes).  To solve this problem we do the reverse scan
 390          * first and attempt to align our cluster, then do a
 391          * forward scan if room remains.
 392          */
 393 more:
 394         while (ib && pageout_count < vm_pageout_page_count) {
 395                 vm_page_t p;
 396
 397                 if (ib > pindex) {
 398                         ib = 0;
 399                         break;
 400                 }
 401
 402                 if ((p = vm_page_prev(pb)) == NULL ||
 403                     (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
 404                         ib = 0;
 405                         break;
 406                 }
 407                 vm_page_lock(p);
 408                 vm_page_test_dirty(p);
 409                 if (p->dirty == 0 ||
 410                     p->queue != PQ_INACTIVE ||
 411                     p->hold_count != 0) {       /* may be undergoing I/O */
 412                         vm_page_unlock(p);
 413                         ib = 0;
 414                         break;
 415                 }
 416                 vm_page_unlock(p);
 417                 mc[--page_base] = pb = p;
 418                 ++pageout_count;
 419                 ++ib;
 420                 /*
 421                  * alignment boundry, stop here and switch directions.  Do
 422                  * not clear ib.
 423                  */
 424                 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
 425                         break;
 426         }
 427
 428         while (pageout_count < vm_pageout_page_count &&
 429             pindex + is < object->size) {
 430                 vm_page_t p;
 431
 432                 if ((p = vm_page_next(ps)) == NULL ||
 433                     (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
 434                         break;
 435                 vm_page_lock(p);
 436                 vm_page_test_dirty(p);
 437                 if (p->dirty == 0 ||
 438                     p->queue != PQ_INACTIVE ||
 439                     p->hold_count != 0) {       /* may be undergoing I/O */
 440                         vm_page_unlock(p);
 441                         break;
 442                 }
 443                 vm_page_unlock(p);
 444                 mc[page_base + pageout_count] = ps = p;
 445                 ++pageout_count;
 446                 ++is;
 447         }
 448
 449         /*
 450          * If we exhausted our forward scan, continue with the reverse scan
 451          * when possible, even past a page boundry.  This catches boundry
 452          * conditions.
 453          */
 454         if (ib && pageout_count < vm_pageout_page_count)
 455                 goto more;
 456
 457         /*
 458          * we allow reads during pageouts...
 459          */
 460         return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
 461             NULL));
 462 }
 463
 464 /*
 465  * vm_pageout_flush() - launder the given pages
 466  *
 467  *      The given pages are laundered.  Note that we setup for the start of
 468  *      I/O ( i.e. busy the page ), mark it read-only, and bump the object
 469  *      reference count all in here rather then in the parent.  If we want
 470  *      the parent to do more sophisticated things we may have to change
 471  *      the ordering.
 472  *
 473  *      Returned runlen is the count of pages between mreq and first
 474  *      page after mreq with status VM_PAGER_AGAIN.
 475  *      *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
 476  *      for any page in runlen set.
 477  */
 478 int
 479 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
 480     boolean_t *eio)
 481 {
 482         vm_object_t object = mc[0]->object;
 483         int pageout_status[count];
 484         int numpagedout = 0;
 485         int i, runlen;
 486
 487         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 488
 489         /*
 490          * Initiate I/O.  Bump the vm_page_t->busy counter and
 491          * mark the pages read-only.
 492          *
 493          * We do not have to fixup the clean/dirty bits here... we can
 494          * allow the pager to do it after the I/O completes.
 495          *
 496          * NOTE! mc[i]->dirty may be partial or fragmented due to an
 497          * edge case with file fragments.
 498          */
 499         for (i = 0; i < count; i++) {
 500                 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
 501                     ("vm_pageout_flush: partially invalid page %p index %d/%d",
 502                         mc[i], i, count));
 503                 vm_page_io_start(mc[i]);
 504                 pmap_remove_write(mc[i]);
 505         }
 506         vm_object_pip_add(object, count);
 507
 508         vm_pager_put_pages(object, mc, count, flags, pageout_status);
 509
 510         runlen = count - mreq;
 511         if (eio != NULL)
 512                 *eio = FALSE;
 513         for (i = 0; i < count; i++) {
 514                 vm_page_t mt = mc[i];
 515
 516                 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
 517                     !pmap_page_is_write_mapped(mt),
 518                     ("vm_pageout_flush: page %p is not write protected", mt));
 519                 switch (pageout_status[i]) {
 520                 case VM_PAGER_OK:
 521                 case VM_PAGER_PEND:
 522                         numpagedout++;
 523                         break;
 524                 case VM_PAGER_BAD:
 525                         /*
 526                          * Page outside of range of object. Right now we
 527                          * essentially lose the changes by pretending it
 528                          * worked.
 529                          */
 530                         vm_page_undirty(mt);
 531                         break;
 532                 case VM_PAGER_ERROR:
 533                 case VM_PAGER_FAIL:
 534                         /*
 535                          * If page couldn't be paged out, then reactivate the
 536                          * page so it doesn't clog the inactive list.  (We
 537                          * will try paging out it again later).
 538                          */
 539                         vm_page_lock(mt);
 540                         vm_page_activate(mt);
 541                         vm_page_unlock(mt);
 542                         if (eio != NULL && i >= mreq && i - mreq < runlen)
 543                                 *eio = TRUE;
 544                         break;
 545                 case VM_PAGER_AGAIN:
 546                         if (i >= mreq && i - mreq < runlen)
 547                                 runlen = i - mreq;
 548                         break;
 549                 }
 550
 551                 /*
 552                  * If the operation is still going, leave the page busy to
 553                  * block all other accesses. Also, leave the paging in
 554                  * progress indicator set so that we don't attempt an object
 555                  * collapse.
 556                  */
 557                 if (pageout_status[i] != VM_PAGER_PEND) {
 558                         vm_object_pip_wakeup(object);
 559                         vm_page_io_finish(mt);
 560                         if (vm_page_count_severe()) {
 561                                 vm_page_lock(mt);
 562                                 vm_page_try_to_cache(mt);
 563                                 vm_page_unlock(mt);
 564                         }
 565                 }
 566         }
 567         if (prunlen != NULL)
 568                 *prunlen = runlen;
 569         return (numpagedout);
 570 }
 571
 572 static boolean_t
 573 vm_pageout_launder(int queue, int tries, vm_paddr_t low, vm_paddr_t high)
 574 {
 575         struct mount *mp;
 576         struct vm_pagequeue *pq;
 577         struct vnode *vp;
 578         vm_object_t object;
 579         vm_paddr_t pa;
 580         vm_page_t m, m_tmp, next;
 581
 582         pq = &vm_pagequeues[queue];
 583         vm_pagequeue_lock(pq);
 584         TAILQ_FOREACH_SAFE(m, &pq->pq_pl, pageq, next) {
 585                 KASSERT(m->queue == queue,
 586                     ("vm_pageout_launder: page %p's queue is not %d", m,
 587                     queue));
 588                 if ((m->flags & PG_MARKER) != 0)
 589                         continue;
 590                 pa = VM_PAGE_TO_PHYS(m);
 591                 if (pa < low || pa + PAGE_SIZE > high)
 592                         continue;
 593                 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
 594                         vm_page_unlock(m);
 595                         continue;
 596                 }
 597                 object = m->object;
 598                 if ((!VM_OBJECT_TRYLOCK(object) &&
 599                     (!vm_pageout_fallback_object_lock(m, &next) ||
 600                     m->hold_count != 0)) || (m->oflags & VPO_BUSY) != 0 ||
 601                     m->busy != 0) {
 602                         vm_page_unlock(m);
 603                         VM_OBJECT_UNLOCK(object);
 604                         continue;
 605                 }
 606                 vm_page_test_dirty(m);
 607                 if (m->dirty == 0 && object->ref_count != 0)
 608                         pmap_remove_all(m);
 609                 if (m->dirty != 0) {
 610                         vm_page_unlock(m);
 611                         if (tries == 0 || (object->flags & OBJ_DEAD) != 0) {
 612                                 VM_OBJECT_UNLOCK(object);
 613                                 continue;
 614                         }
 615                         if (object->type == OBJT_VNODE) {
 616                                 vm_pagequeue_unlock(pq);
 617                                 vp = object->handle;
 618                                 vm_object_reference_locked(object);
 619                                 VM_OBJECT_UNLOCK(object);
 620                                 (void)vn_start_write(vp, &mp, V_WAIT);
 621                                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 622                                 VM_OBJECT_LOCK(object);
 623                                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 624                                 VM_OBJECT_UNLOCK(object);
 625                                 VOP_UNLOCK(vp, 0);
 626                                 vm_object_deallocate(object);
 627                                 vn_finished_write(mp);
 628                                 return (TRUE);
 629                         } else if (object->type == OBJT_SWAP ||
 630                             object->type == OBJT_DEFAULT) {
 631                                 vm_pagequeue_unlock(pq);
 632                                 m_tmp = m;
 633                                 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC,
 634                                     0, NULL, NULL);
 635                                 VM_OBJECT_UNLOCK(object);
 636                                 return (TRUE);
 637                         }
 638                 } else {
 639                         /*
 640                          * Dequeue here to prevent lock recursion in
 641                          * vm_page_cache().
 642                          */
 643                         vm_page_dequeue_locked(m);
 644                         vm_page_cache(m);
 645                         vm_page_unlock(m);
 646                 }
 647                 VM_OBJECT_UNLOCK(object);
 648         }
 649         vm_pagequeue_unlock(pq);
 650         return (FALSE);
 651 }
 652
 653 /*
 654  * Increase the number of cached pages.  The specified value, "tries",
 655  * determines which categories of pages are cached:
 656  *
 657  *  0: All clean, inactive pages within the specified physical address range
 658  *     are cached.  Will not sleep.
 659  *  1: The vm_lowmem handlers are called.  All inactive pages within
 660  *     the specified physical address range are cached.  May sleep.
 661  *  2: The vm_lowmem handlers are called.  All inactive and active pages
 662  *     within the specified physical address range are cached.  May sleep.
 663  */
 664 void
 665 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high)
 666 {
 667         int actl, actmax, inactl, inactmax;
 668
 669         if (tries > 0) {
 670                 /*
 671                  * Decrease registered cache sizes.  The vm_lowmem handlers
 672                  * may acquire locks and/or sleep, so they can only be invoked
 673                  * when "tries" is greater than zero.
 674                  */
 675                 EVENTHANDLER_INVOKE(vm_lowmem, 0);
 676
 677                 /*
 678                  * We do this explicitly after the caches have been drained
 679                  * above.
 680                  */
 681                 uma_reclaim();
 682         }
 683         inactl = 0;
 684         inactmax = cnt.v_inactive_count;
 685         actl = 0;
 686         actmax = tries < 2 ? 0 : cnt.v_active_count;
 687 again:
 688         if (inactl < inactmax && vm_pageout_launder(PQ_INACTIVE, tries, low,
 689             high)) {
 690                 inactl++;
 691                 goto again;
 692         }
 693         if (actl < actmax && vm_pageout_launder(PQ_ACTIVE, tries, low, high)) {
 694                 actl++;
 695                 goto again;
 696         }
 697 }
 698
 699 #if !defined(NO_SWAPPING)
 700 /*
 701  *      vm_pageout_object_deactivate_pages
 702  *
 703  *      Deactivate enough pages to satisfy the inactive target
 704  *      requirements.
 705  *
 706  *      The object and map must be locked.
 707  */
 708 static void
 709 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
 710     long desired)
 711 {
 712         vm_object_t backing_object, object;
 713         vm_page_t p;
 714         int actcount, remove_mode;
 715
 716         VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
 717         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 718                 return;
 719         for (object = first_object;; object = backing_object) {
 720                 if (pmap_resident_count(pmap) <= desired)
 721                         goto unlock_return;
 722                 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 723                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 724                     object->paging_in_progress != 0)
 725                         goto unlock_return;
 726
 727                 remove_mode = 0;
 728                 if (object->shadow_count > 1)
 729                         remove_mode = 1;
 730                 /*
 731                  * Scan the object's entire memory queue.
 732                  */
 733                 TAILQ_FOREACH(p, &object->memq, listq) {
 734                         if (pmap_resident_count(pmap) <= desired)
 735                                 goto unlock_return;
 736                         if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0)
 737                                 continue;
 738                         PCPU_INC(cnt.v_pdpages);
 739                         vm_page_lock(p);
 740                         if (p->wire_count != 0 || p->hold_count != 0 ||
 741                             !pmap_page_exists_quick(pmap, p)) {
 742                                 vm_page_unlock(p);
 743                                 continue;
 744                         }
 745                         actcount = pmap_ts_referenced(p);
 746                         if ((p->aflags & PGA_REFERENCED) != 0) {
 747                                 if (actcount == 0)
 748                                         actcount = 1;
 749                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 750                         }
 751                         if (p->queue != PQ_ACTIVE && actcount != 0) {
 752                                 vm_page_activate(p);
 753                                 p->act_count += actcount;
 754                         } else if (p->queue == PQ_ACTIVE) {
 755                                 if (actcount == 0) {
 756                                         p->act_count -= min(p->act_count,
 757                                             ACT_DECLINE);
 758                                         if (!remove_mode &&
 759                                             (vm_pageout_algorithm ||
 760                                             p->act_count == 0)) {
 761                                                 pmap_remove_all(p);
 762                                                 vm_page_deactivate(p);
 763                                         } else
 764                                                 vm_page_requeue(p);
 765                                 } else {
 766                                         vm_page_activate(p);
 767                                         if (p->act_count < ACT_MAX -
 768                                             ACT_ADVANCE)
 769                                                 p->act_count += ACT_ADVANCE;
 770                                         vm_page_requeue(p);
 771                                 }
 772                         } else if (p->queue == PQ_INACTIVE)
 773                                 pmap_remove_all(p);
 774                         vm_page_unlock(p);
 775                 }
 776                 if ((backing_object = object->backing_object) == NULL)
 777                         goto unlock_return;
 778                 VM_OBJECT_LOCK(backing_object);
 779                 if (object != first_object)
 780                         VM_OBJECT_UNLOCK(object);
 781         }
 782 unlock_return:
 783         if (object != first_object)
 784                 VM_OBJECT_UNLOCK(object);
 785 }
 786
 787 /*
 788  * deactivate some number of pages in a map, try to do it fairly, but
 789  * that is really hard to do.
 790  */
 791 static void
 792 vm_pageout_map_deactivate_pages(map, desired)
 793         vm_map_t map;
 794         long desired;
 795 {
 796         vm_map_entry_t tmpe;
 797         vm_object_t obj, bigobj;
 798         int nothingwired;
 799
 800         if (!vm_map_trylock(map))
 801                 return;
 802
 803         bigobj = NULL;
 804         nothingwired = TRUE;
 805
 806         /*
 807          * first, search out the biggest object, and try to free pages from
 808          * that.
 809          */
 810         tmpe = map->header.next;
 811         while (tmpe != &map->header) {
 812                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 813                         obj = tmpe->object.vm_object;
 814                         if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
 815                                 if (obj->shadow_count <= 1 &&
 816                                     (bigobj == NULL ||
 817                                      bigobj->resident_page_count < obj->resident_page_count)) {
 818                                         if (bigobj != NULL)
 819                                                 VM_OBJECT_UNLOCK(bigobj);
 820                                         bigobj = obj;
 821                                 } else
 822                                         VM_OBJECT_UNLOCK(obj);
 823                         }
 824                 }
 825                 if (tmpe->wired_count > 0)
 826                         nothingwired = FALSE;
 827                 tmpe = tmpe->next;
 828         }
 829
 830         if (bigobj != NULL) {
 831                 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
 832                 VM_OBJECT_UNLOCK(bigobj);
 833         }
 834         /*
 835          * Next, hunt around for other pages to deactivate.  We actually
 836          * do this search sort of wrong -- .text first is not the best idea.
 837          */
 838         tmpe = map->header.next;
 839         while (tmpe != &map->header) {
 840                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 841                         break;
 842                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 843                         obj = tmpe->object.vm_object;
 844                         if (obj != NULL) {
 845                                 VM_OBJECT_LOCK(obj);
 846                                 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
 847                                 VM_OBJECT_UNLOCK(obj);
 848                         }
 849                 }
 850                 tmpe = tmpe->next;
 851         }
 852
 853         /*
 854          * Remove all mappings if a process is swapped out, this will free page
 855          * table pages.
 856          */
 857         if (desired == 0 && nothingwired) {
 858                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 859                     vm_map_max(map));
 860         }
 861         vm_map_unlock(map);
 862 }
 863 #endif          /* !defined(NO_SWAPPING) */
 864
 865 /*
 866  *      vm_pageout_scan does the dirty work for the pageout daemon.
 867  */
 868 static void
 869 vm_pageout_scan(int pass)
 870 {
 871         vm_page_t m, next;
 872         struct vm_page marker;
 873         struct vm_pagequeue *pq;
 874         int page_shortage, maxscan, pcount;
 875         int addl_page_shortage;
 876         vm_object_t object;
 877         int actcount;
 878         int vnodes_skipped = 0;
 879         int maxlaunder;
 880         boolean_t queues_locked;
 881
 882         vm_pageout_init_marker(&marker, PQ_INACTIVE);
 883
 884         /*
 885          * Decrease registered cache sizes.
 886          */
 887         EVENTHANDLER_INVOKE(vm_lowmem, 0);
 888         /*
 889          * We do this explicitly after the caches have been drained above.
 890          */
 891         uma_reclaim();
 892
 893         /*
 894          * The addl_page_shortage is the number of temporarily
 895          * stuck pages in the inactive queue.  In other words, the
 896          * number of pages from cnt.v_inactive_count that should be
 897          * discounted in setting the target for the active queue scan.
 898          */
 899         addl_page_shortage = atomic_readandclear_int(&vm_pageout_deficit);
 900
 901         /*
 902          * Calculate the number of pages we want to either free or move
 903          * to the cache.
 904          */
 905         page_shortage = vm_paging_target() + addl_page_shortage;
 906
 907         /*
 908          * maxlaunder limits the number of dirty pages we flush per scan.
 909          * For most systems a smaller value (16 or 32) is more robust under
 910          * extreme memory and disk pressure because any unnecessary writes
 911          * to disk can result in extreme performance degredation.  However,
 912          * systems with excessive dirty pages (especially when MAP_NOSYNC is
 913          * used) will die horribly with limited laundering.  If the pageout
 914          * daemon cannot clean enough pages in the first pass, we let it go
 915          * all out in succeeding passes.
 916          */
 917         if ((maxlaunder = vm_max_launder) <= 1)
 918                 maxlaunder = 1;
 919         if (pass)
 920                 maxlaunder = 10000;
 921
 922         maxscan = cnt.v_inactive_count;
 923
 924         /*
 925          * Start scanning the inactive queue for pages we can move to the
 926          * cache or free.  The scan will stop when the target is reached or
 927          * we have scanned the entire inactive queue.  Note that m->act_count
 928          * is not used to form decisions for the inactive queue, only for the
 929          * active queue.
 930          */
 931         pq = &vm_pagequeues[PQ_INACTIVE];
 932         vm_pagequeue_lock(pq);
 933         queues_locked = TRUE;
 934         for (m = TAILQ_FIRST(&pq->pq_pl);
 935              m != NULL && maxscan-- > 0 && page_shortage > 0;
 936              m = next) {
 937                 vm_pagequeue_assert_locked(pq);
 938                 KASSERT(queues_locked, ("unlocked queues"));
 939                 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
 940
 941                 PCPU_INC(cnt.v_pdpages);
 942                 next = TAILQ_NEXT(m, pageq);
 943
 944                 /*
 945                  * skip marker pages
 946                  */
 947                 if (m->flags & PG_MARKER)
 948                         continue;
 949
 950                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
 951                     ("Fictitious page %p cannot be in inactive queue", m));
 952                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
 953                     ("Unmanaged page %p cannot be in inactive queue", m));
 954
 955                 /*
 956                  * The page or object lock acquisitions fail if the
 957                  * page was removed from the queue or moved to a
 958                  * different position within the queue.  In either
 959                  * case, addl_page_shortage should not be incremented.
 960                  */
 961                 if (!vm_pageout_page_lock(m, &next)) {
 962                         vm_page_unlock(m);
 963                         continue;
 964                 }
 965                 object = m->object;
 966                 if (!VM_OBJECT_TRYLOCK(object) &&
 967                     !vm_pageout_fallback_object_lock(m, &next)) {
 968                         vm_page_unlock(m);
 969                         VM_OBJECT_UNLOCK(object);
 970                         continue;
 971                 }
 972
 973                 /*
 974                  * Don't mess with busy pages, keep them at at the
 975                  * front of the queue, most likely they are being
 976                  * paged out.  Increment addl_page_shortage for busy
 977                  * pages, because they may leave the inactive queue
 978                  * shortly after page scan is finished.
 979                  */
 980                 if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0) {
 981                         vm_page_unlock(m);
 982                         VM_OBJECT_UNLOCK(object);
 983                         addl_page_shortage++;
 984                         continue;
 985                 }
 986
 987                 /*
 988                  * We unlock the inactive page queue, invalidating the
 989                  * 'next' pointer.  Use our marker to remember our
 990                  * place.
 991                  */
 992                 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, pageq);
 993                 vm_pagequeue_unlock(pq);
 994                 queues_locked = FALSE;
 995
 996                 /*
 997                  * If the object is not being used, we ignore previous
 998                  * references.
 999                  */
1000                 if (object->ref_count == 0) {
1001                         vm_page_aflag_clear(m, PGA_REFERENCED);
1002                         KASSERT(!pmap_page_is_mapped(m),
1003                             ("vm_pageout_scan: page %p is mapped", m));
1004
1005                 /*
1006                  * Otherwise, if the page has been referenced while in the
1007                  * inactive queue, we bump the "activation count" upwards,
1008                  * making it less likely that the page will be added back to
1009                  * the inactive queue prematurely again.  Here we check the
1010                  * page tables (or emulated bits, if any), given the upper
1011                  * level VM system not knowing anything about existing
1012                  * references.
1013                  */
1014                 } else if ((m->aflags & PGA_REFERENCED) == 0 &&
1015                     (actcount = pmap_ts_referenced(m)) != 0) {
1016                         vm_page_activate(m);
1017                         vm_page_unlock(m);
1018                         m->act_count += actcount + ACT_ADVANCE;
1019                         VM_OBJECT_UNLOCK(object);
1020                         goto relock_queues;
1021                 }
1022
1023                 /*
1024                  * If the upper level VM system knows about any page
1025                  * references, we activate the page.  We also set the
1026                  * "activation count" higher than normal so that we will less
1027                  * likely place pages back onto the inactive queue again.
1028                  */
1029                 if ((m->aflags & PGA_REFERENCED) != 0) {
1030                         vm_page_aflag_clear(m, PGA_REFERENCED);
1031                         actcount = pmap_ts_referenced(m);
1032                         vm_page_activate(m);
1033                         vm_page_unlock(m);
1034                         m->act_count += actcount + ACT_ADVANCE + 1;
1035                         VM_OBJECT_UNLOCK(object);
1036                         goto relock_queues;
1037                 }
1038
1039                 if (m->hold_count != 0) {
1040                         vm_page_unlock(m);
1041                         VM_OBJECT_UNLOCK(object);
1042
1043                         /*
1044                          * Held pages are essentially stuck in the
1045                          * queue.  So, they ought to be discounted
1046                          * from cnt.v_inactive_count.  See the
1047                          * calculation of the page_shortage for the
1048                          * loop over the active queue below.
1049                          */
1050                         addl_page_shortage++;
1051                         goto relock_queues;
1052                 }
1053
1054                 /*
1055                  * If the page appears to be clean at the machine-independent
1056                  * layer, then remove all of its mappings from the pmap in
1057                  * anticipation of placing it onto the cache queue.  If,
1058                  * however, any of the page's mappings allow write access,
1059                  * then the page may still be modified until the last of those
1060                  * mappings are removed.
1061                  */
1062                 vm_page_test_dirty(m);
1063                 if (m->dirty == 0 && object->ref_count != 0)
1064                         pmap_remove_all(m);
1065
1066                 if (m->valid == 0) {
1067                         /*
1068                          * Invalid pages can be easily freed
1069                          */
1070                         vm_page_free(m);
1071                         PCPU_INC(cnt.v_dfree);
1072                         --page_shortage;
1073                 } else if (m->dirty == 0) {
1074                         /*
1075                          * Clean pages can be placed onto the cache queue.
1076                          * This effectively frees them.
1077                          */
1078                         vm_page_cache(m);
1079                         --page_shortage;
1080                 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
1081                         /*
1082                          * Dirty pages need to be paged out, but flushing
1083                          * a page is extremely expensive verses freeing
1084                          * a clean page.  Rather then artificially limiting
1085                          * the number of pages we can flush, we instead give
1086                          * dirty pages extra priority on the inactive queue
1087                          * by forcing them to be cycled through the queue
1088                          * twice before being flushed, after which the
1089                          * (now clean) page will cycle through once more
1090                          * before being freed.  This significantly extends
1091                          * the thrash point for a heavily loaded machine.
1092                          */
1093                         m->flags |= PG_WINATCFLS;
1094                         vm_pagequeue_lock(pq);
1095                         queues_locked = TRUE;
1096                         vm_page_requeue_locked(m);
1097                 } else if (maxlaunder > 0) {
1098                         /*
1099                          * We always want to try to flush some dirty pages if
1100                          * we encounter them, to keep the system stable.
1101                          * Normally this number is small, but under extreme
1102                          * pressure where there are insufficient clean pages
1103                          * on the inactive queue, we may have to go all out.
1104                          */
1105                         int swap_pageouts_ok;
1106                         struct vnode *vp = NULL;
1107                         struct mount *mp = NULL;
1108
1109                         if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
1110                                 swap_pageouts_ok = 1;
1111                         } else {
1112                                 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
1113                                 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
1114                                 vm_page_count_min());
1115
1116                         }
1117
1118                         /*
1119                          * We don't bother paging objects that are "dead".
1120                          * Those objects are in a "rundown" state.
1121                          */
1122                         if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
1123                                 vm_pagequeue_lock(pq);
1124                                 vm_page_unlock(m);
1125                                 VM_OBJECT_UNLOCK(object);
1126                                 queues_locked = TRUE;
1127                                 vm_page_requeue_locked(m);
1128                                 goto relock_queues;
1129                         }
1130
1131                         /*
1132                          * The object is already known NOT to be dead.   It
1133                          * is possible for the vget() to block the whole
1134                          * pageout daemon, but the new low-memory handling
1135                          * code should prevent it.
1136                          *
1137                          * The previous code skipped locked vnodes and, worse,
1138                          * reordered pages in the queue.  This results in
1139                          * completely non-deterministic operation and, on a
1140                          * busy system, can lead to extremely non-optimal
1141                          * pageouts.  For example, it can cause clean pages
1142                          * to be freed and dirty pages to be moved to the end
1143                          * of the queue.  Since dirty pages are also moved to
1144                          * the end of the queue once-cleaned, this gives
1145                          * way too large a weighting to defering the freeing
1146                          * of dirty pages.
1147                          *
1148                          * We can't wait forever for the vnode lock, we might
1149                          * deadlock due to a vn_read() getting stuck in
1150                          * vm_wait while holding this vnode.  We skip the
1151                          * vnode if we can't get it in a reasonable amount
1152                          * of time.
1153                          */
1154                         if (object->type == OBJT_VNODE) {
1155                                 vm_page_unlock(m);
1156                                 vp = object->handle;
1157                                 if (vp->v_type == VREG &&
1158                                     vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1159                                         mp = NULL;
1160                                         ++pageout_lock_miss;
1161                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1162                                                 vnodes_skipped++;
1163                                         goto unlock_and_continue;
1164                                 }
1165                                 KASSERT(mp != NULL,
1166                                     ("vp %p with NULL v_mount", vp));
1167                                 vm_object_reference_locked(object);
1168                                 VM_OBJECT_UNLOCK(object);
1169                                 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1170                                     curthread)) {
1171                                         VM_OBJECT_LOCK(object);
1172                                         ++pageout_lock_miss;
1173                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1174                                                 vnodes_skipped++;
1175                                         vp = NULL;
1176                                         goto unlock_and_continue;
1177                                 }
1178                                 VM_OBJECT_LOCK(object);
1179                                 vm_page_lock(m);
1180                                 vm_pagequeue_lock(pq);
1181                                 queues_locked = TRUE;
1182                                 /*
1183                                  * The page might have been moved to another
1184                                  * queue during potential blocking in vget()
1185                                  * above.  The page might have been freed and
1186                                  * reused for another vnode.
1187                                  */
1188                                 if (m->queue != PQ_INACTIVE ||
1189                                     m->object != object ||
1190                                     TAILQ_NEXT(m, pageq) != &marker) {
1191                                         vm_page_unlock(m);
1192                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1193                                                 vnodes_skipped++;
1194                                         goto unlock_and_continue;
1195                                 }
1196
1197                                 /*
1198                                  * The page may have been busied during the
1199                                  * blocking in vget().  We don't move the
1200                                  * page back onto the end of the queue so that
1201                                  * statistics are more correct if we don't.
1202                                  */
1203                                 if (m->busy || (m->oflags & VPO_BUSY)) {
1204                                         vm_page_unlock(m);
1205                                         goto unlock_and_continue;
1206                                 }
1207
1208                                 /*
1209                                  * If the page has become held it might
1210                                  * be undergoing I/O, so skip it
1211                                  */
1212                                 if (m->hold_count) {
1213                                         vm_page_unlock(m);
1214                                         vm_page_requeue_locked(m);
1215                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1216                                                 vnodes_skipped++;
1217                                         goto unlock_and_continue;
1218                                 }
1219                                 vm_pagequeue_unlock(pq);
1220                                 queues_locked = FALSE;
1221                         }
1222
1223                         /*
1224                          * If a page is dirty, then it is either being washed
1225                          * (but not yet cleaned) or it is still in the
1226                          * laundry.  If it is still in the laundry, then we
1227                          * start the cleaning operation.
1228                          *
1229                          * decrement page_shortage on success to account for
1230                          * the (future) cleaned page.  Otherwise we could wind
1231                          * up laundering or cleaning too many pages.
1232                          */
1233                         if (vm_pageout_clean(m) != 0) {
1234                                 --page_shortage;
1235                                 --maxlaunder;
1236                         }
1237 unlock_and_continue:
1238                         vm_page_lock_assert(m, MA_NOTOWNED);
1239                         VM_OBJECT_UNLOCK(object);
1240                         if (mp != NULL) {
1241                                 if (queues_locked) {
1242                                         vm_pagequeue_unlock(pq);
1243                                         queues_locked = FALSE;
1244                                 }
1245                                 if (vp != NULL)
1246                                         vput(vp);
1247                                 vm_object_deallocate(object);
1248                                 vn_finished_write(mp);
1249                         }
1250                         vm_page_lock_assert(m, MA_NOTOWNED);
1251                         goto relock_queues;
1252                 }
1253                 vm_page_unlock(m);
1254                 VM_OBJECT_UNLOCK(object);
1255 relock_queues:
1256                 if (!queues_locked) {
1257                         vm_pagequeue_lock(pq);
1258                         queues_locked = TRUE;
1259                 }
1260                 next = TAILQ_NEXT(&marker, pageq);
1261                 TAILQ_REMOVE(&pq->pq_pl, &marker, pageq);
1262         }
1263         vm_pagequeue_unlock(pq);
1264
1265         /*
1266          * Compute the number of pages we want to try to move from the
1267          * active queue to the inactive queue.
1268          */
1269         page_shortage = vm_paging_target() +
1270                 cnt.v_inactive_target - cnt.v_inactive_count;
1271         page_shortage += addl_page_shortage;
1272
1273         /*
1274          * Scan the active queue for things we can deactivate. We nominally
1275          * track the per-page activity counter and use it to locate
1276          * deactivation candidates.
1277          */
1278         pcount = cnt.v_active_count;
1279         pq = &vm_pagequeues[PQ_ACTIVE];
1280         vm_pagequeue_lock(pq);
1281         m = TAILQ_FIRST(&pq->pq_pl);
1282         while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1283
1284                 KASSERT(m->queue == PQ_ACTIVE,
1285                     ("vm_pageout_scan: page %p isn't active", m));
1286
1287                 next = TAILQ_NEXT(m, pageq);
1288                 if ((m->flags & PG_MARKER) != 0) {
1289                         m = next;
1290                         continue;
1291                 }
1292                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1293                     ("Fictitious page %p cannot be in active queue", m));
1294                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1295                     ("Unmanaged page %p cannot be in active queue", m));
1296                 if (!vm_pageout_page_lock(m, &next)) {
1297                         vm_page_unlock(m);
1298                         m = next;
1299                         continue;
1300                 }
1301                 object = m->object;
1302                 if (!VM_OBJECT_TRYLOCK(object) &&
1303                     !vm_pageout_fallback_object_lock(m, &next)) {
1304                         VM_OBJECT_UNLOCK(object);
1305                         vm_page_unlock(m);
1306                         m = next;
1307                         continue;
1308                 }
1309
1310                 /*
1311                  * Don't deactivate pages that are busy.
1312                  */
1313                 if ((m->busy != 0) ||
1314                     (m->oflags & VPO_BUSY) ||
1315                     (m->hold_count != 0)) {
1316                         vm_page_unlock(m);
1317                         VM_OBJECT_UNLOCK(object);
1318                         vm_page_requeue_locked(m);
1319                         m = next;
1320                         continue;
1321                 }
1322
1323                 /*
1324                  * The count for pagedaemon pages is done after checking the
1325                  * page for eligibility...
1326                  */
1327                 PCPU_INC(cnt.v_pdpages);
1328
1329                 /*
1330                  * Check to see "how much" the page has been used.
1331                  */
1332                 actcount = 0;
1333                 if (object->ref_count != 0) {
1334                         if (m->aflags & PGA_REFERENCED) {
1335                                 actcount += 1;
1336                         }
1337                         actcount += pmap_ts_referenced(m);
1338                         if (actcount) {
1339                                 m->act_count += ACT_ADVANCE + actcount;
1340                                 if (m->act_count > ACT_MAX)
1341                                         m->act_count = ACT_MAX;
1342                         }
1343                 }
1344
1345                 /*
1346                  * Since we have "tested" this bit, we need to clear it now.
1347                  */
1348                 vm_page_aflag_clear(m, PGA_REFERENCED);
1349
1350                 /*
1351                  * Only if an object is currently being used, do we use the
1352                  * page activation count stats.
1353                  */
1354                 if (actcount != 0 && object->ref_count != 0)
1355                         vm_page_requeue_locked(m);
1356                 else {
1357                         m->act_count -= min(m->act_count, ACT_DECLINE);
1358                         if (vm_pageout_algorithm ||
1359                             object->ref_count == 0 ||
1360                             m->act_count == 0) {
1361                                 page_shortage--;
1362                                 /* Dequeue to avoid later lock recursion. */
1363                                 vm_page_dequeue_locked(m);
1364                                 if (object->ref_count == 0) {
1365                                         KASSERT(!pmap_page_is_mapped(m),
1366                                     ("vm_pageout_scan: page %p is mapped", m));
1367                                         if (m->dirty == 0)
1368                                                 vm_page_cache(m);
1369                                         else
1370                                                 vm_page_deactivate(m);
1371                                 } else {
1372                                         vm_page_deactivate(m);
1373                                 }
1374                         } else
1375                                 vm_page_requeue_locked(m);
1376                 }
1377                 vm_page_unlock(m);
1378                 VM_OBJECT_UNLOCK(object);
1379                 m = next;
1380         }
1381         vm_pagequeue_unlock(pq);
1382 #if !defined(NO_SWAPPING)
1383         /*
1384          * Idle process swapout -- run once per second.
1385          */
1386         if (vm_swap_idle_enabled) {
1387                 static long lsec;
1388                 if (time_second != lsec) {
1389                         vm_req_vmdaemon(VM_SWAP_IDLE);
1390                         lsec = time_second;
1391                 }
1392         }
1393 #endif
1394
1395         /*
1396          * If we didn't get enough free pages, and we have skipped a vnode
1397          * in a writeable object, wakeup the sync daemon.  And kick swapout
1398          * if we did not get enough free pages.
1399          */
1400         if (vm_paging_target() > 0) {
1401                 if (vnodes_skipped && vm_page_count_min())
1402                         (void) speedup_syncer();
1403 #if !defined(NO_SWAPPING)
1404                 if (vm_swap_enabled && vm_page_count_target())
1405                         vm_req_vmdaemon(VM_SWAP_NORMAL);
1406 #endif
1407         }
1408
1409         /*
1410          * If we are critically low on one of RAM or swap and low on
1411          * the other, kill the largest process.  However, we avoid
1412          * doing this on the first pass in order to give ourselves a
1413          * chance to flush out dirty vnode-backed pages and to allow
1414          * active pages to be moved to the inactive queue and reclaimed.
1415          */
1416         if (pass != 0 &&
1417             ((swap_pager_avail < 64 && vm_page_count_min()) ||
1418              (swap_pager_full && vm_paging_target() > 0)))
1419                 vm_pageout_oom(VM_OOM_MEM);
1420 }
1421
1422
1423 void
1424 vm_pageout_oom(int shortage)
1425 {
1426         struct proc *p, *bigproc;
1427         vm_offset_t size, bigsize;
1428         struct thread *td;
1429         struct vmspace *vm;
1430
1431         /*
1432          * We keep the process bigproc locked once we find it to keep anyone
1433          * from messing with it; however, there is a possibility of
1434          * deadlock if process B is bigproc and one of it's child processes
1435          * attempts to propagate a signal to B while we are waiting for A's
1436          * lock while walking this list.  To avoid this, we don't block on
1437          * the process lock but just skip a process if it is already locked.
1438          */
1439         bigproc = NULL;
1440         bigsize = 0;
1441         sx_slock(&allproc_lock);
1442         FOREACH_PROC_IN_SYSTEM(p) {
1443                 int breakout;
1444
1445                 if (PROC_TRYLOCK(p) == 0)
1446                         continue;
1447                 /*
1448                  * If this is a system, protected or killed process, skip it.
1449                  */
1450                 if (p->p_state != PRS_NORMAL ||
1451                     (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1452                     (p->p_pid == 1) || P_KILLED(p) ||
1453                     ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1454                         PROC_UNLOCK(p);
1455                         continue;
1456                 }
1457                 /*
1458                  * If the process is in a non-running type state,
1459                  * don't touch it.  Check all the threads individually.
1460                  */
1461                 breakout = 0;
1462                 FOREACH_THREAD_IN_PROC(p, td) {
1463                         thread_lock(td);
1464                         if (!TD_ON_RUNQ(td) &&
1465                             !TD_IS_RUNNING(td) &&
1466                             !TD_IS_SLEEPING(td) &&
1467                             !TD_IS_SUSPENDED(td)) {
1468                                 thread_unlock(td);
1469                                 breakout = 1;
1470                                 break;
1471                         }
1472                         thread_unlock(td);
1473                 }
1474                 if (breakout) {
1475                         PROC_UNLOCK(p);
1476                         continue;
1477                 }
1478                 /*
1479                  * get the process size
1480                  */
1481                 vm = vmspace_acquire_ref(p);
1482                 if (vm == NULL) {
1483                         PROC_UNLOCK(p);
1484                         continue;
1485                 }
1486                 if (!vm_map_trylock_read(&vm->vm_map)) {
1487                         vmspace_free(vm);
1488                         PROC_UNLOCK(p);
1489                         continue;
1490                 }
1491                 size = vmspace_swap_count(vm);
1492                 vm_map_unlock_read(&vm->vm_map);
1493                 if (shortage == VM_OOM_MEM)
1494                         size += vmspace_resident_count(vm);
1495                 vmspace_free(vm);
1496                 /*
1497                  * if the this process is bigger than the biggest one
1498                  * remember it.
1499                  */
1500                 if (size > bigsize) {
1501                         if (bigproc != NULL)
1502                                 PROC_UNLOCK(bigproc);
1503                         bigproc = p;
1504                         bigsize = size;
1505                 } else
1506                         PROC_UNLOCK(p);
1507         }
1508         sx_sunlock(&allproc_lock);
1509         if (bigproc != NULL) {
1510                 killproc(bigproc, "out of swap space");
1511                 sched_nice(bigproc, PRIO_MIN);
1512                 PROC_UNLOCK(bigproc);
1513                 wakeup(&cnt.v_free_count);
1514         }
1515 }
1516
1517 /*
1518  * This routine tries to maintain the pseudo LRU active queue,
1519  * so that during long periods of time where there is no paging,
1520  * that some statistic accumulation still occurs.  This code
1521  * helps the situation where paging just starts to occur.
1522  */
1523 static void
1524 vm_pageout_page_stats(void)
1525 {
1526         struct vm_pagequeue *pq;
1527         vm_object_t object;
1528         vm_page_t m, next;
1529         int pcount, tpcount;            /* Number of pages to check */
1530         static int fullintervalcount = 0;
1531         int page_shortage;
1532
1533         page_shortage =
1534             (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1535             (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1536
1537         if (page_shortage <= 0)
1538                 return;
1539
1540         pcount = cnt.v_active_count;
1541         fullintervalcount += vm_pageout_stats_interval;
1542         if (fullintervalcount < vm_pageout_full_stats_interval) {
1543                 vm_pageout_stats++;
1544                 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1545                     cnt.v_page_count;
1546                 if (pcount > tpcount)
1547                         pcount = tpcount;
1548         } else {
1549                 vm_pageout_full_stats++;
1550                 fullintervalcount = 0;
1551         }
1552
1553         pq = &vm_pagequeues[PQ_ACTIVE];
1554         vm_pagequeue_lock(pq);
1555         m = TAILQ_FIRST(&pq->pq_pl);
1556         while ((m != NULL) && (pcount-- > 0)) {
1557                 int actcount;
1558
1559                 KASSERT(m->queue == PQ_ACTIVE,
1560                     ("vm_pageout_page_stats: page %p isn't active", m));
1561
1562                 next = TAILQ_NEXT(m, pageq);
1563                 if ((m->flags & PG_MARKER) != 0) {
1564                         m = next;
1565                         continue;
1566                 }
1567                 vm_page_lock_assert(m, MA_NOTOWNED);
1568                 if (!vm_pageout_page_lock(m, &next)) {
1569                         vm_page_unlock(m);
1570                         m = next;
1571                         continue;
1572                 }
1573                 object = m->object;
1574                 if (!VM_OBJECT_TRYLOCK(object) &&
1575                     !vm_pageout_fallback_object_lock(m, &next)) {
1576                         VM_OBJECT_UNLOCK(object);
1577                         vm_page_unlock(m);
1578                         m = next;
1579                         continue;
1580                 }
1581
1582                 /*
1583                  * Don't deactivate pages that are busy.
1584                  */
1585                 if ((m->busy != 0) ||
1586                     (m->oflags & VPO_BUSY) ||
1587                     (m->hold_count != 0)) {
1588                         vm_page_unlock(m);
1589                         VM_OBJECT_UNLOCK(object);
1590                         vm_page_requeue_locked(m);
1591                         m = next;
1592                         continue;
1593                 }
1594
1595                 actcount = 0;
1596                 if (m->aflags & PGA_REFERENCED) {
1597                         vm_page_aflag_clear(m, PGA_REFERENCED);
1598                         actcount += 1;
1599                 }
1600
1601                 actcount += pmap_ts_referenced(m);
1602                 if (actcount) {
1603                         m->act_count += ACT_ADVANCE + actcount;
1604                         if (m->act_count > ACT_MAX)
1605                                 m->act_count = ACT_MAX;
1606                         vm_page_requeue_locked(m);
1607                 } else {
1608                         if (m->act_count == 0) {
1609                                 /*
1610                                  * We turn off page access, so that we have
1611                                  * more accurate RSS stats.  We don't do this
1612                                  * in the normal page deactivation when the
1613                                  * system is loaded VM wise, because the
1614                                  * cost of the large number of page protect
1615                                  * operations would be higher than the value
1616                                  * of doing the operation.
1617                                  */
1618                                 pmap_remove_all(m);
1619                                 /* Dequeue to avoid later lock recursion. */
1620                                 vm_page_dequeue_locked(m);
1621                                 vm_page_deactivate(m);
1622                         } else {
1623                                 m->act_count -= min(m->act_count, ACT_DECLINE);
1624                                 vm_page_requeue_locked(m);
1625                         }
1626                 }
1627                 vm_page_unlock(m);
1628                 VM_OBJECT_UNLOCK(object);
1629                 m = next;
1630         }
1631         vm_pagequeue_unlock(pq);
1632 }
1633
1634 /*
1635  *      vm_pageout is the high level pageout daemon.
1636  */
1637 static void
1638 vm_pageout(void)
1639 {
1640         int error, pass;
1641
1642         /*
1643          * Initialize some paging parameters.
1644          */
1645         cnt.v_interrupt_free_min = 2;
1646         if (cnt.v_page_count < 2000)
1647                 vm_pageout_page_count = 8;
1648
1649         /*
1650          * v_free_reserved needs to include enough for the largest
1651          * swap pager structures plus enough for any pv_entry structs
1652          * when paging.
1653          */
1654         if (cnt.v_page_count > 1024)
1655                 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1656         else
1657                 cnt.v_free_min = 4;
1658         cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1659             cnt.v_interrupt_free_min;
1660         cnt.v_free_reserved = vm_pageout_page_count +
1661             cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1662         cnt.v_free_severe = cnt.v_free_min / 2;
1663         cnt.v_free_min += cnt.v_free_reserved;
1664         cnt.v_free_severe += cnt.v_free_reserved;
1665
1666         /*
1667          * v_free_target and v_cache_min control pageout hysteresis.  Note
1668          * that these are more a measure of the VM cache queue hysteresis
1669          * then the VM free queue.  Specifically, v_free_target is the
1670          * high water mark (free+cache pages).
1671          *
1672          * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1673          * low water mark, while v_free_min is the stop.  v_cache_min must
1674          * be big enough to handle memory needs while the pageout daemon
1675          * is signalled and run to free more pages.
1676          */
1677         if (cnt.v_free_count > 6144)
1678                 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1679         else
1680                 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1681
1682         if (cnt.v_free_count > 2048) {
1683                 cnt.v_cache_min = cnt.v_free_target;
1684                 cnt.v_cache_max = 2 * cnt.v_cache_min;
1685                 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1686         } else {
1687                 cnt.v_cache_min = 0;
1688                 cnt.v_cache_max = 0;
1689                 cnt.v_inactive_target = cnt.v_free_count / 4;
1690         }
1691         if (cnt.v_inactive_target > cnt.v_free_count / 3)
1692                 cnt.v_inactive_target = cnt.v_free_count / 3;
1693
1694         /* XXX does not really belong here */
1695         if (vm_page_max_wired == 0)
1696                 vm_page_max_wired = cnt.v_free_count / 3;
1697
1698         if (vm_pageout_stats_max == 0)
1699                 vm_pageout_stats_max = cnt.v_free_target;
1700
1701         /*
1702          * Set interval in seconds for stats scan.
1703          */
1704         if (vm_pageout_stats_interval == 0)
1705                 vm_pageout_stats_interval = 5;
1706         if (vm_pageout_full_stats_interval == 0)
1707                 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1708
1709         swap_pager_swap_init();
1710         pass = 0;
1711         /*
1712          * The pageout daemon is never done, so loop forever.
1713          */
1714         while (TRUE) {
1715                 /*
1716                  * If we have enough free memory, wakeup waiters.  Do
1717                  * not clear vm_pages_needed until we reach our target,
1718                  * otherwise we may be woken up over and over again and
1719                  * waste a lot of cpu.
1720                  */
1721                 mtx_lock(&vm_page_queue_free_mtx);
1722                 if (vm_pages_needed && !vm_page_count_min()) {
1723                         if (!vm_paging_needed())
1724                                 vm_pages_needed = 0;
1725                         wakeup(&cnt.v_free_count);
1726                 }
1727                 if (vm_pages_needed) {
1728                         /*
1729                          * Still not done, take a second pass without waiting
1730                          * (unlimited dirty cleaning), otherwise sleep a bit
1731                          * and try again.
1732                          */
1733                         ++pass;
1734                         if (pass > 1)
1735                                 msleep(&vm_pages_needed,
1736                                     &vm_page_queue_free_mtx, PVM, "psleep",
1737                                     hz / 2);
1738                 } else {
1739                         /*
1740                          * Good enough, sleep & handle stats.  Prime the pass
1741                          * for the next run.
1742                          */
1743                         if (pass > 1)
1744                                 pass = 1;
1745                         else
1746                                 pass = 0;
1747                         error = msleep(&vm_pages_needed,
1748                             &vm_page_queue_free_mtx, PVM, "psleep",
1749                             vm_pageout_stats_interval * hz);
1750                         if (error && !vm_pages_needed) {
1751                                 mtx_unlock(&vm_page_queue_free_mtx);
1752                                 pass = 0;
1753                                 vm_pageout_page_stats();
1754                                 continue;
1755                         }
1756                 }
1757                 if (vm_pages_needed)
1758                         cnt.v_pdwakeups++;
1759                 mtx_unlock(&vm_page_queue_free_mtx);
1760                 vm_pageout_scan(pass);
1761         }
1762 }
1763
1764 /*
1765  * Unless the free page queue lock is held by the caller, this function
1766  * should be regarded as advisory.  Specifically, the caller should
1767  * not msleep() on &cnt.v_free_count following this function unless
1768  * the free page queue lock is held until the msleep() is performed.
1769  */
1770 void
1771 pagedaemon_wakeup(void)
1772 {
1773
1774         if (!vm_pages_needed && curthread->td_proc != pageproc) {
1775                 vm_pages_needed = 1;
1776                 wakeup(&vm_pages_needed);
1777         }
1778 }
1779
1780 #if !defined(NO_SWAPPING)
1781 static void
1782 vm_req_vmdaemon(int req)
1783 {
1784         static int lastrun = 0;
1785
1786         mtx_lock(&vm_daemon_mtx);
1787         vm_pageout_req_swapout |= req;
1788         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1789                 wakeup(&vm_daemon_needed);
1790                 lastrun = ticks;
1791         }
1792         mtx_unlock(&vm_daemon_mtx);
1793 }
1794
1795 static void
1796 vm_daemon(void)
1797 {
1798         struct rlimit rsslim;
1799         struct proc *p;
1800         struct thread *td;
1801         struct vmspace *vm;
1802         int breakout, swapout_flags, tryagain, attempts;
1803 #ifdef RACCT
1804         uint64_t rsize, ravailable;
1805 #endif
1806
1807         while (TRUE) {
1808                 mtx_lock(&vm_daemon_mtx);
1809 #ifdef RACCT
1810                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz);
1811 #else
1812                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1813 #endif
1814                 swapout_flags = vm_pageout_req_swapout;
1815                 vm_pageout_req_swapout = 0;
1816                 mtx_unlock(&vm_daemon_mtx);
1817                 if (swapout_flags)
1818                         swapout_procs(swapout_flags);
1819
1820                 /*
1821                  * scan the processes for exceeding their rlimits or if
1822                  * process is swapped out -- deactivate pages
1823                  */
1824                 tryagain = 0;
1825                 attempts = 0;
1826 again:
1827                 attempts++;
1828                 sx_slock(&allproc_lock);
1829                 FOREACH_PROC_IN_SYSTEM(p) {
1830                         vm_pindex_t limit, size;
1831
1832                         /*
1833                          * if this is a system process or if we have already
1834                          * looked at this process, skip it.
1835                          */
1836                         PROC_LOCK(p);
1837                         if (p->p_state != PRS_NORMAL ||
1838                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1839                                 PROC_UNLOCK(p);
1840                                 continue;
1841                         }
1842                         /*
1843                          * if the process is in a non-running type state,
1844                          * don't touch it.
1845                          */
1846                         breakout = 0;
1847                         FOREACH_THREAD_IN_PROC(p, td) {
1848                                 thread_lock(td);
1849                                 if (!TD_ON_RUNQ(td) &&
1850                                     !TD_IS_RUNNING(td) &&
1851                                     !TD_IS_SLEEPING(td) &&
1852                                     !TD_IS_SUSPENDED(td)) {
1853                                         thread_unlock(td);
1854                                         breakout = 1;
1855                                         break;
1856                                 }
1857                                 thread_unlock(td);
1858                         }
1859                         if (breakout) {
1860                                 PROC_UNLOCK(p);
1861                                 continue;
1862                         }
1863                         /*
1864                          * get a limit
1865                          */
1866                         lim_rlimit(p, RLIMIT_RSS, &rsslim);
1867                         limit = OFF_TO_IDX(
1868                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
1869
1870                         /*
1871                          * let processes that are swapped out really be
1872                          * swapped out set the limit to nothing (will force a
1873                          * swap-out.)
1874                          */
1875                         if ((p->p_flag & P_INMEM) == 0)
1876                                 limit = 0;      /* XXX */
1877                         vm = vmspace_acquire_ref(p);
1878                         PROC_UNLOCK(p);
1879                         if (vm == NULL)
1880                                 continue;
1881
1882                         size = vmspace_resident_count(vm);
1883                         if (size >= limit) {
1884                                 vm_pageout_map_deactivate_pages(
1885                                     &vm->vm_map, limit);
1886                         }
1887 #ifdef RACCT
1888                         rsize = IDX_TO_OFF(size);
1889                         PROC_LOCK(p);
1890                         racct_set(p, RACCT_RSS, rsize);
1891                         ravailable = racct_get_available(p, RACCT_RSS);
1892                         PROC_UNLOCK(p);
1893                         if (rsize > ravailable) {
1894                                 /*
1895                                  * Don't be overly aggressive; this might be
1896                                  * an innocent process, and the limit could've
1897                                  * been exceeded by some memory hog.  Don't
1898                                  * try to deactivate more than 1/4th of process'
1899                                  * resident set size.
1900                                  */
1901                                 if (attempts <= 8) {
1902                                         if (ravailable < rsize - (rsize / 4))
1903                                                 ravailable = rsize - (rsize / 4);
1904                                 }
1905                                 vm_pageout_map_deactivate_pages(
1906                                     &vm->vm_map, OFF_TO_IDX(ravailable));
1907                                 /* Update RSS usage after paging out. */
1908                                 size = vmspace_resident_count(vm);
1909                                 rsize = IDX_TO_OFF(size);
1910                                 PROC_LOCK(p);
1911                                 racct_set(p, RACCT_RSS, rsize);
1912                                 PROC_UNLOCK(p);
1913                                 if (rsize > ravailable)
1914                                         tryagain = 1;
1915                         }
1916 #endif
1917                         vmspace_free(vm);
1918                 }
1919                 sx_sunlock(&allproc_lock);
1920                 if (tryagain != 0 && attempts <= 10)
1921                         goto again;
1922         }
1923 }
1924 #endif                  /* !defined(NO_SWAPPING) */