sys/vm/vm_pageout.c

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