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