sys/vm/vm_pageout.c

   1 /*-
   2  * Copyright (c) 1991 Regents of the University of California.
   3  * All rights reserved.
   4  * Copyright (c) 1994 John S. Dyson
   5  * All rights reserved.
   6  * Copyright (c) 1994 David Greenman
   7  * All rights reserved.
   8  * Copyright (c) 2005 Yahoo! Technologies Norway AS
   9  * All rights reserved.
  10  *
  11  * This code is derived from software contributed to Berkeley by
  12  * The Mach Operating System project at Carnegie-Mellon University.
  13  *
  14  * Redistribution and use in source and binary forms, with or without
  15  * modification, are permitted provided that the following conditions
  16  * are met:
  17  * 1. Redistributions of source code must retain the above copyright
  18  *    notice, this list of conditions and the following disclaimer.
  19  * 2. Redistributions in binary form must reproduce the above copyright
  20  *    notice, this list of conditions and the following disclaimer in the
  21  *    documentation and/or other materials provided with the distribution.
  22  * 3. All advertising materials mentioning features or use of this software
  23  *    must display the following acknowledgement:
  24  *      This product includes software developed by the University of
  25  *      California, Berkeley and its contributors.
  26  * 4. Neither the name of the University nor the names of its contributors
  27  *    may be used to endorse or promote products derived from this software
  28  *    without specific prior written permission.
  29  *
  30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  40  * SUCH DAMAGE.
  41  *
  42  *      from: @(#)vm_pageout.c  7.4 (Berkeley) 5/7/91
  43  *
  44  *
  45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  46  * All rights reserved.
  47  *
  48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  49  *
  50  * Permission to use, copy, modify and distribute this software and
  51  * its documentation is hereby granted, provided that both the copyright
  52  * notice and this permission notice appear in all copies of the
  53  * software, derivative works or modified versions, and any portions
  54  * thereof, and that both notices appear in supporting documentation.
  55  *
  56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  59  *
  60  * Carnegie Mellon requests users of this software to return to
  61  *
  62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  63  *  School of Computer Science
  64  *  Carnegie Mellon University
  65  *  Pittsburgh PA 15213-3890
  66  *
  67  * any improvements or extensions that they make and grant Carnegie the
  68  * rights to redistribute these changes.
  69  */
  70
  71 /*
  72  *      The proverbial page-out daemon.
  73  */
  74
  75 #include <sys/cdefs.h>
  76 __FBSDID("$FreeBSD$");
  77
  78 #include "opt_vm.h"
  79 #include <sys/param.h>
  80 #include <sys/systm.h>
  81 #include <sys/kernel.h>
  82 #include <sys/eventhandler.h>
  83 #include <sys/lock.h>
  84 #include <sys/mutex.h>
  85 #include <sys/proc.h>
  86 #include <sys/kthread.h>
  87 #include <sys/ktr.h>
  88 #include <sys/mount.h>
  89 #include <sys/racct.h>
  90 #include <sys/resourcevar.h>
  91 #include <sys/sched.h>
  92 #include <sys/signalvar.h>
  93 #include <sys/vnode.h>
  94 #include <sys/vmmeter.h>
  95 #include <sys/sx.h>
  96 #include <sys/sysctl.h>
  97
  98 #include <vm/vm.h>
  99 #include <vm/vm_param.h>
 100 #include <vm/vm_object.h>
 101 #include <vm/vm_page.h>
 102 #include <vm/vm_map.h>
 103 #include <vm/vm_pageout.h>
 104 #include <vm/vm_pager.h>
 105 #include <vm/swap_pager.h>
 106 #include <vm/vm_extern.h>
 107 #include <vm/uma.h>
 108
 109 /*
 110  * System initialization
 111  */
 112
 113 /* the kernel process "vm_pageout"*/
 114 static void vm_pageout(void);
 115 static int vm_pageout_clean(vm_page_t);
 116 static void vm_pageout_scan(int pass);
 117
 118 struct proc *pageproc;
 119
 120 static struct kproc_desc page_kp = {
 121         "pagedaemon",
 122         vm_pageout,
 123         &pageproc
 124 };
 125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
 126     &page_kp);
 127
 128 #if !defined(NO_SWAPPING)
 129 /* the kernel process "vm_daemon"*/
 130 static void vm_daemon(void);
 131 static struct   proc *vmproc;
 132
 133 static struct kproc_desc vm_kp = {
 134         "vmdaemon",
 135         vm_daemon,
 136         &vmproc
 137 };
 138 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
 139 #endif
 140
 141
 142 int vm_pages_needed;            /* Event on which pageout daemon sleeps */
 143 int vm_pageout_deficit;         /* Estimated number of pages deficit */
 144 int vm_pageout_pages_needed;    /* flag saying that the pageout daemon needs pages */
 145
 146 #if !defined(NO_SWAPPING)
 147 static int vm_pageout_req_swapout;      /* XXX */
 148 static int vm_daemon_needed;
 149 static struct mtx vm_daemon_mtx;
 150 /* Allow for use by vm_pageout before vm_daemon is initialized. */
 151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
 152 #endif
 153 static int vm_max_launder = 32;
 154 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
 155 static int vm_pageout_full_stats_interval = 0;
 156 static int vm_pageout_algorithm=0;
 157 static int defer_swap_pageouts=0;
 158 static int disable_swap_pageouts=0;
 159
 160 #if defined(NO_SWAPPING)
 161 static int vm_swap_enabled=0;
 162 static int vm_swap_idle_enabled=0;
 163 #else
 164 static int vm_swap_enabled=1;
 165 static int vm_swap_idle_enabled=0;
 166 #endif
 167
 168 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
 169         CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
 170
 171 SYSCTL_INT(_vm, OID_AUTO, max_launder,
 172         CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
 173
 174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
 175         CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
 176
 177 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
 178         CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
 179
 180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
 181         CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
 182
 183 #if defined(NO_SWAPPING)
 184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
 185         CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
 186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
 187         CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
 188 #else
 189 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
 190         CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
 191 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
 192         CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
 193 #endif
 194
 195 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
 196         CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
 197
 198 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
 199         CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
 200
 201 static int pageout_lock_miss;
 202 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
 203         CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
 204
 205 #define VM_PAGEOUT_PAGE_COUNT 16
 206 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
 207
 208 int vm_page_max_wired;          /* XXX max # of wired pages system-wide */
 209 SYSCTL_INT(_vm, OID_AUTO, max_wired,
 210         CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
 211
 212 #if !defined(NO_SWAPPING)
 213 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
 214 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
 215 static void vm_req_vmdaemon(int req);
 216 #endif
 217 static void vm_pageout_page_stats(void);
 218
 219 /*
 220  * Initialize a dummy page for marking the caller's place in the specified
 221  * paging queue.  In principle, this function only needs to set the flag
 222  * PG_MARKER.  Nonetheless, it sets the flag VPO_BUSY and initializes the hold
 223  * count to one as safety precautions.
 224  */
 225 static void
 226 vm_pageout_init_marker(vm_page_t marker, u_short queue)
 227 {
 228
 229         bzero(marker, sizeof(*marker));
 230         marker->flags = PG_MARKER;
 231         marker->oflags = VPO_BUSY;
 232         marker->queue = queue;
 233         marker->hold_count = 1;
 234 }
 235
 236 /*
 237  * vm_pageout_fallback_object_lock:
 238  *
 239  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
 240  * known to have failed and page queue must be either PQ_ACTIVE or
 241  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
 242  * while locking the vm object.  Use marker page to detect page queue
 243  * changes and maintain notion of next page on page queue.  Return
 244  * TRUE if no changes were detected, FALSE otherwise.  vm object is
 245  * locked on return.
 246  *
 247  * This function depends on both the lock portion of struct vm_object
 248  * and normal struct vm_page being type stable.
 249  */
 250 boolean_t
 251 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
 252 {
 253         struct vm_page marker;
 254         boolean_t unchanged;
 255         u_short queue;
 256         vm_object_t object;
 257
 258         queue = m->queue;
 259         vm_pageout_init_marker(&marker, queue);
 260         object = m->object;
 261
 262         TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
 263                            m, &marker, pageq);
 264         vm_page_unlock_queues();
 265         vm_page_unlock(m);
 266         VM_OBJECT_LOCK(object);
 267         vm_page_lock(m);
 268         vm_page_lock_queues();
 269
 270         /* Page queue might have changed. */
 271         *next = TAILQ_NEXT(&marker, pageq);
 272         unchanged = (m->queue == queue &&
 273                      m->object == object &&
 274                      &marker == TAILQ_NEXT(m, pageq));
 275         TAILQ_REMOVE(&vm_page_queues[queue].pl,
 276                      &marker, pageq);
 277         return (unchanged);
 278 }
 279
 280 /*
 281  * Lock the page while holding the page queue lock.  Use marker page
 282  * to detect page queue changes and maintain notion of next page on
 283  * page queue.  Return TRUE if no changes were detected, FALSE
 284  * otherwise.  The page is locked on return. The page queue lock might
 285  * be dropped and reacquired.
 286  *
 287  * This function depends on normal struct vm_page being type stable.
 288  */
 289 boolean_t
 290 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
 291 {
 292         struct vm_page marker;
 293         boolean_t unchanged;
 294         u_short queue;
 295
 296         vm_page_lock_assert(m, MA_NOTOWNED);
 297         mtx_assert(&vm_page_queue_mtx, MA_OWNED);
 298
 299         if (vm_page_trylock(m))
 300                 return (TRUE);
 301
 302         queue = m->queue;
 303         vm_pageout_init_marker(&marker, queue);
 304
 305         TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq);
 306         vm_page_unlock_queues();
 307         vm_page_lock(m);
 308         vm_page_lock_queues();
 309
 310         /* Page queue might have changed. */
 311         *next = TAILQ_NEXT(&marker, pageq);
 312         unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq));
 313         TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq);
 314         return (unchanged);
 315 }
 316
 317 /*
 318  * vm_pageout_clean:
 319  *
 320  * Clean the page and remove it from the laundry.
 321  *
 322  * We set the busy bit to cause potential page faults on this page to
 323  * block.  Note the careful timing, however, the busy bit isn't set till
 324  * late and we cannot do anything that will mess with the page.
 325  */
 326 static int
 327 vm_pageout_clean(vm_page_t m)
 328 {
 329         vm_object_t object;
 330         vm_page_t mc[2*vm_pageout_page_count], pb, ps;
 331         int pageout_count;
 332         int ib, is, page_base;
 333         vm_pindex_t pindex = m->pindex;
 334
 335         vm_page_lock_assert(m, MA_OWNED);
 336         object = m->object;
 337         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 338
 339         /*
 340          * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
 341          * with the new swapper, but we could have serious problems paging
 342          * out other object types if there is insufficient memory.
 343          *
 344          * Unfortunately, checking free memory here is far too late, so the
 345          * check has been moved up a procedural level.
 346          */
 347
 348         /*
 349          * Can't clean the page if it's busy or held.
 350          */
 351         KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0,
 352             ("vm_pageout_clean: page %p is busy", m));
 353         KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
 354         vm_page_unlock(m);
 355
 356         mc[vm_pageout_page_count] = pb = ps = m;
 357         pageout_count = 1;
 358         page_base = vm_pageout_page_count;
 359         ib = 1;
 360         is = 1;
 361
 362         /*
 363          * Scan object for clusterable pages.
 364          *
 365          * We can cluster ONLY if: ->> the page is NOT
 366          * clean, wired, busy, held, or mapped into a
 367          * buffer, and one of the following:
 368          * 1) The page is inactive, or a seldom used
 369          *    active page.
 370          * -or-
 371          * 2) we force the issue.
 372          *
 373          * During heavy mmap/modification loads the pageout
 374          * daemon can really fragment the underlying file
 375          * due to flushing pages out of order and not trying
 376          * align the clusters (which leave sporatic out-of-order
 377          * holes).  To solve this problem we do the reverse scan
 378          * first and attempt to align our cluster, then do a
 379          * forward scan if room remains.
 380          */
 381 more:
 382         while (ib && pageout_count < vm_pageout_page_count) {
 383                 vm_page_t p;
 384
 385                 if (ib > pindex) {
 386                         ib = 0;
 387                         break;
 388                 }
 389
 390                 if ((p = vm_page_prev(pb)) == NULL ||
 391                     (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
 392                         ib = 0;
 393                         break;
 394                 }
 395                 vm_page_lock(p);
 396                 vm_page_test_dirty(p);
 397                 if (p->dirty == 0 ||
 398                     p->queue != PQ_INACTIVE ||
 399                     p->hold_count != 0) {       /* may be undergoing I/O */
 400                         vm_page_unlock(p);
 401                         ib = 0;
 402                         break;
 403                 }
 404                 vm_page_unlock(p);
 405                 mc[--page_base] = pb = p;
 406                 ++pageout_count;
 407                 ++ib;
 408                 /*
 409                  * alignment boundry, stop here and switch directions.  Do
 410                  * not clear ib.
 411                  */
 412                 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
 413                         break;
 414         }
 415
 416         while (pageout_count < vm_pageout_page_count &&
 417             pindex + is < object->size) {
 418                 vm_page_t p;
 419
 420                 if ((p = vm_page_next(ps)) == NULL ||
 421                     (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
 422                         break;
 423                 vm_page_lock(p);
 424                 vm_page_test_dirty(p);
 425                 if (p->dirty == 0 ||
 426                     p->queue != PQ_INACTIVE ||
 427                     p->hold_count != 0) {       /* may be undergoing I/O */
 428                         vm_page_unlock(p);
 429                         break;
 430                 }
 431                 vm_page_unlock(p);
 432                 mc[page_base + pageout_count] = ps = p;
 433                 ++pageout_count;
 434                 ++is;
 435         }
 436
 437         /*
 438          * If we exhausted our forward scan, continue with the reverse scan
 439          * when possible, even past a page boundry.  This catches boundry
 440          * conditions.
 441          */
 442         if (ib && pageout_count < vm_pageout_page_count)
 443                 goto more;
 444
 445         /*
 446          * we allow reads during pageouts...
 447          */
 448         return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
 449             NULL));
 450 }
 451
 452 /*
 453  * vm_pageout_flush() - launder the given pages
 454  *
 455  *      The given pages are laundered.  Note that we setup for the start of
 456  *      I/O ( i.e. busy the page ), mark it read-only, and bump the object
 457  *      reference count all in here rather then in the parent.  If we want
 458  *      the parent to do more sophisticated things we may have to change
 459  *      the ordering.
 460  *
 461  *      Returned runlen is the count of pages between mreq and first
 462  *      page after mreq with status VM_PAGER_AGAIN.
 463  *      *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
 464  *      for any page in runlen set.
 465  */
 466 int
 467 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
 468     boolean_t *eio)
 469 {
 470         vm_object_t object = mc[0]->object;
 471         int pageout_status[count];
 472         int numpagedout = 0;
 473         int i, runlen;
 474
 475         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 476         mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
 477
 478         /*
 479          * Initiate I/O.  Bump the vm_page_t->busy counter and
 480          * mark the pages read-only.
 481          *
 482          * We do not have to fixup the clean/dirty bits here... we can
 483          * allow the pager to do it after the I/O completes.
 484          *
 485          * NOTE! mc[i]->dirty may be partial or fragmented due to an
 486          * edge case with file fragments.
 487          */
 488         for (i = 0; i < count; i++) {
 489                 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
 490                     ("vm_pageout_flush: partially invalid page %p index %d/%d",
 491                         mc[i], i, count));
 492                 vm_page_io_start(mc[i]);
 493                 pmap_remove_write(mc[i]);
 494         }
 495         vm_object_pip_add(object, count);
 496
 497         vm_pager_put_pages(object, mc, count, flags, pageout_status);
 498
 499         runlen = count - mreq;
 500         if (eio != NULL)
 501                 *eio = FALSE;
 502         for (i = 0; i < count; i++) {
 503                 vm_page_t mt = mc[i];
 504
 505                 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
 506                     !pmap_page_is_write_mapped(mt),
 507                     ("vm_pageout_flush: page %p is not write protected", mt));
 508                 switch (pageout_status[i]) {
 509                 case VM_PAGER_OK:
 510                 case VM_PAGER_PEND:
 511                         numpagedout++;
 512                         break;
 513                 case VM_PAGER_BAD:
 514                         /*
 515                          * Page outside of range of object. Right now we
 516                          * essentially lose the changes by pretending it
 517                          * worked.
 518                          */
 519                         vm_page_undirty(mt);
 520                         break;
 521                 case VM_PAGER_ERROR:
 522                 case VM_PAGER_FAIL:
 523                         /*
 524                          * If page couldn't be paged out, then reactivate the
 525                          * page so it doesn't clog the inactive list.  (We
 526                          * will try paging out it again later).
 527                          */
 528                         vm_page_lock(mt);
 529                         vm_page_activate(mt);
 530                         vm_page_unlock(mt);
 531                         if (eio != NULL && i >= mreq && i - mreq < runlen)
 532                                 *eio = TRUE;
 533                         break;
 534                 case VM_PAGER_AGAIN:
 535                         if (i >= mreq && i - mreq < runlen)
 536                                 runlen = i - mreq;
 537                         break;
 538                 }
 539
 540                 /*
 541                  * If the operation is still going, leave the page busy to
 542                  * block all other accesses. Also, leave the paging in
 543                  * progress indicator set so that we don't attempt an object
 544                  * collapse.
 545                  */
 546                 if (pageout_status[i] != VM_PAGER_PEND) {
 547                         vm_object_pip_wakeup(object);
 548                         vm_page_io_finish(mt);
 549                         if (vm_page_count_severe()) {
 550                                 vm_page_lock(mt);
 551                                 vm_page_try_to_cache(mt);
 552                                 vm_page_unlock(mt);
 553                         }
 554                 }
 555         }
 556         if (prunlen != NULL)
 557                 *prunlen = runlen;
 558         return (numpagedout);
 559 }
 560
 561 #if !defined(NO_SWAPPING)
 562 /*
 563  *      vm_pageout_object_deactivate_pages
 564  *
 565  *      Deactivate enough pages to satisfy the inactive target
 566  *      requirements.
 567  *
 568  *      The object and map must be locked.
 569  */
 570 static void
 571 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
 572     long desired)
 573 {
 574         vm_object_t backing_object, object;
 575         vm_page_t p;
 576         int actcount, remove_mode;
 577
 578         VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
 579         if (first_object->type == OBJT_DEVICE ||
 580             first_object->type == OBJT_SG)
 581                 return;
 582         for (object = first_object;; object = backing_object) {
 583                 if (pmap_resident_count(pmap) <= desired)
 584                         goto unlock_return;
 585                 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 586                 if (object->type == OBJT_PHYS || object->paging_in_progress)
 587                         goto unlock_return;
 588
 589                 remove_mode = 0;
 590                 if (object->shadow_count > 1)
 591                         remove_mode = 1;
 592                 /*
 593                  * Scan the object's entire memory queue.
 594                  */
 595                 TAILQ_FOREACH(p, &object->memq, listq) {
 596                         if (pmap_resident_count(pmap) <= desired)
 597                                 goto unlock_return;
 598                         if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0)
 599                                 continue;
 600                         PCPU_INC(cnt.v_pdpages);
 601                         vm_page_lock(p);
 602                         if (p->wire_count != 0 || p->hold_count != 0 ||
 603                             !pmap_page_exists_quick(pmap, p)) {
 604                                 vm_page_unlock(p);
 605                                 continue;
 606                         }
 607                         actcount = pmap_ts_referenced(p);
 608                         if ((p->aflags & PGA_REFERENCED) != 0) {
 609                                 if (actcount == 0)
 610                                         actcount = 1;
 611                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 612                         }
 613                         if (p->queue != PQ_ACTIVE && actcount != 0) {
 614                                 vm_page_activate(p);
 615                                 p->act_count += actcount;
 616                         } else if (p->queue == PQ_ACTIVE) {
 617                                 if (actcount == 0) {
 618                                         p->act_count -= min(p->act_count,
 619                                             ACT_DECLINE);
 620                                         if (!remove_mode &&
 621                                             (vm_pageout_algorithm ||
 622                                             p->act_count == 0)) {
 623                                                 pmap_remove_all(p);
 624                                                 vm_page_deactivate(p);
 625                                         } else {
 626                                                 vm_page_lock_queues();
 627                                                 vm_page_requeue(p);
 628                                                 vm_page_unlock_queues();
 629                                         }
 630                                 } else {
 631                                         vm_page_activate(p);
 632                                         if (p->act_count < ACT_MAX -
 633                                             ACT_ADVANCE)
 634                                                 p->act_count += ACT_ADVANCE;
 635                                         vm_page_lock_queues();
 636                                         vm_page_requeue(p);
 637                                         vm_page_unlock_queues();
 638                                 }
 639                         } else if (p->queue == PQ_INACTIVE)
 640                                 pmap_remove_all(p);
 641                         vm_page_unlock(p);
 642                 }
 643                 if ((backing_object = object->backing_object) == NULL)
 644                         goto unlock_return;
 645                 VM_OBJECT_LOCK(backing_object);
 646                 if (object != first_object)
 647                         VM_OBJECT_UNLOCK(object);
 648         }
 649 unlock_return:
 650         if (object != first_object)
 651                 VM_OBJECT_UNLOCK(object);
 652 }
 653
 654 /*
 655  * deactivate some number of pages in a map, try to do it fairly, but
 656  * that is really hard to do.
 657  */
 658 static void
 659 vm_pageout_map_deactivate_pages(map, desired)
 660         vm_map_t map;
 661         long desired;
 662 {
 663         vm_map_entry_t tmpe;
 664         vm_object_t obj, bigobj;
 665         int nothingwired;
 666
 667         if (!vm_map_trylock(map))
 668                 return;
 669
 670         bigobj = NULL;
 671         nothingwired = TRUE;
 672
 673         /*
 674          * first, search out the biggest object, and try to free pages from
 675          * that.
 676          */
 677         tmpe = map->header.next;
 678         while (tmpe != &map->header) {
 679                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 680                         obj = tmpe->object.vm_object;
 681                         if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
 682                                 if (obj->shadow_count <= 1 &&
 683                                     (bigobj == NULL ||
 684                                      bigobj->resident_page_count < obj->resident_page_count)) {
 685                                         if (bigobj != NULL)
 686                                                 VM_OBJECT_UNLOCK(bigobj);
 687                                         bigobj = obj;
 688                                 } else
 689                                         VM_OBJECT_UNLOCK(obj);
 690                         }
 691                 }
 692                 if (tmpe->wired_count > 0)
 693                         nothingwired = FALSE;
 694                 tmpe = tmpe->next;
 695         }
 696
 697         if (bigobj != NULL) {
 698                 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
 699                 VM_OBJECT_UNLOCK(bigobj);
 700         }
 701         /*
 702          * Next, hunt around for other pages to deactivate.  We actually
 703          * do this search sort of wrong -- .text first is not the best idea.
 704          */
 705         tmpe = map->header.next;
 706         while (tmpe != &map->header) {
 707                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 708                         break;
 709                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 710                         obj = tmpe->object.vm_object;
 711                         if (obj != NULL) {
 712                                 VM_OBJECT_LOCK(obj);
 713                                 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
 714                                 VM_OBJECT_UNLOCK(obj);
 715                         }
 716                 }
 717                 tmpe = tmpe->next;
 718         }
 719
 720         /*
 721          * Remove all mappings if a process is swapped out, this will free page
 722          * table pages.
 723          */
 724         if (desired == 0 && nothingwired) {
 725                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 726                     vm_map_max(map));
 727         }
 728         vm_map_unlock(map);
 729 }
 730 #endif          /* !defined(NO_SWAPPING) */
 731
 732 /*
 733  *      vm_pageout_scan does the dirty work for the pageout daemon.
 734  */
 735 static void
 736 vm_pageout_scan(int pass)
 737 {
 738         vm_page_t m, next;
 739         struct vm_page marker;
 740         int page_shortage, maxscan, pcount;
 741         int addl_page_shortage;
 742         vm_object_t object;
 743         int actcount;
 744         int vnodes_skipped = 0;
 745         int maxlaunder;
 746         boolean_t queues_locked;
 747
 748         /*
 749          * Decrease registered cache sizes.
 750          */
 751         EVENTHANDLER_INVOKE(vm_lowmem, 0);
 752         /*
 753          * We do this explicitly after the caches have been drained above.
 754          */
 755         uma_reclaim();
 756
 757         /*
 758          * The addl_page_shortage is the number of temporarily
 759          * stuck pages in the inactive queue.  In other words, the
 760          * number of pages from cnt.v_inactive_count that should be
 761          * discounted in setting the target for the active queue scan.
 762          */
 763         addl_page_shortage = atomic_readandclear_int(&vm_pageout_deficit);
 764
 765         /*
 766          * Calculate the number of pages we want to either free or move
 767          * to the cache.
 768          */
 769         page_shortage = vm_paging_target() + addl_page_shortage;
 770
 771         vm_pageout_init_marker(&marker, PQ_INACTIVE);
 772
 773         /*
 774          * Start scanning the inactive queue for pages we can move to the
 775          * cache or free.  The scan will stop when the target is reached or
 776          * we have scanned the entire inactive queue.  Note that m->act_count
 777          * is not used to form decisions for the inactive queue, only for the
 778          * active queue.
 779          *
 780          * maxlaunder limits the number of dirty pages we flush per scan.
 781          * For most systems a smaller value (16 or 32) is more robust under
 782          * extreme memory and disk pressure because any unnecessary writes
 783          * to disk can result in extreme performance degredation.  However,
 784          * systems with excessive dirty pages (especially when MAP_NOSYNC is
 785          * used) will die horribly with limited laundering.  If the pageout
 786          * daemon cannot clean enough pages in the first pass, we let it go
 787          * all out in succeeding passes.
 788          */
 789         if ((maxlaunder = vm_max_launder) <= 1)
 790                 maxlaunder = 1;
 791         if (pass)
 792                 maxlaunder = 10000;
 793         vm_page_lock_queues();
 794         queues_locked = TRUE;
 795         maxscan = cnt.v_inactive_count;
 796
 797         for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
 798              m != NULL && maxscan-- > 0 && page_shortage > 0;
 799              m = next) {
 800                 KASSERT(queues_locked, ("unlocked queues"));
 801                 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
 802                 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
 803
 804                 cnt.v_pdpages++;
 805                 next = TAILQ_NEXT(m, pageq);
 806
 807                 /*
 808                  * skip marker pages
 809                  */
 810                 if (m->flags & PG_MARKER)
 811                         continue;
 812
 813                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
 814                     ("Fictitious page %p cannot be in inactive queue", m));
 815                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
 816                     ("Unmanaged page %p cannot be in inactive queue", m));
 817
 818                 /*
 819                  * The page or object lock acquisitions fail if the
 820                  * page was removed from the queue or moved to a
 821                  * different position within the queue.  In either
 822                  * case, addl_page_shortage should not be incremented.
 823                  */
 824                 if (!vm_pageout_page_lock(m, &next)) {
 825                         vm_page_unlock(m);
 826                         continue;
 827                 }
 828                 object = m->object;
 829                 if (!VM_OBJECT_TRYLOCK(object) &&
 830                     !vm_pageout_fallback_object_lock(m, &next)) {
 831                         vm_page_unlock(m);
 832                         VM_OBJECT_UNLOCK(object);
 833                         continue;
 834                 }
 835
 836                 /*
 837                  * Don't mess with busy pages, keep them at at the
 838                  * front of the queue, most likely they are being
 839                  * paged out.  Increment addl_page_shortage for busy
 840                  * pages, because they may leave the inactive queue
 841                  * shortly after page scan is finished.
 842                  */
 843                 if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0) {
 844                         vm_page_unlock(m);
 845                         VM_OBJECT_UNLOCK(object);
 846                         addl_page_shortage++;
 847                         continue;
 848                 }
 849
 850                 /*
 851                  * We unlock vm_page_queue_mtx, invalidating the
 852                  * 'next' pointer.  Use our marker to remember our
 853                  * place.
 854                  */
 855                 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
 856                     m, &marker, pageq);
 857                 vm_page_unlock_queues();
 858                 queues_locked = FALSE;
 859
 860                 /*
 861                  * If the object is not being used, we ignore previous
 862                  * references.
 863                  */
 864                 if (object->ref_count == 0) {
 865                         vm_page_aflag_clear(m, PGA_REFERENCED);
 866                         KASSERT(!pmap_page_is_mapped(m),
 867                             ("vm_pageout_scan: page %p is mapped", m));
 868
 869                 /*
 870                  * Otherwise, if the page has been referenced while in the
 871                  * inactive queue, we bump the "activation count" upwards,
 872                  * making it less likely that the page will be added back to
 873                  * the inactive queue prematurely again.  Here we check the
 874                  * page tables (or emulated bits, if any), given the upper
 875                  * level VM system not knowing anything about existing
 876                  * references.
 877                  */
 878                 } else if ((m->aflags & PGA_REFERENCED) == 0 &&
 879                     (actcount = pmap_ts_referenced(m)) != 0) {
 880                         vm_page_activate(m);
 881                         vm_page_unlock(m);
 882                         m->act_count += actcount + ACT_ADVANCE;
 883                         VM_OBJECT_UNLOCK(object);
 884                         goto relock_queues;
 885                 }
 886
 887                 /*
 888                  * If the upper level VM system knows about any page
 889                  * references, we activate the page.  We also set the
 890                  * "activation count" higher than normal so that we will less
 891                  * likely place pages back onto the inactive queue again.
 892                  */
 893                 if ((m->aflags & PGA_REFERENCED) != 0) {
 894                         vm_page_aflag_clear(m, PGA_REFERENCED);
 895                         actcount = pmap_ts_referenced(m);
 896                         vm_page_activate(m);
 897                         vm_page_unlock(m);
 898                         m->act_count += actcount + ACT_ADVANCE + 1;
 899                         VM_OBJECT_UNLOCK(object);
 900                         goto relock_queues;
 901                 }
 902
 903                 if (m->hold_count != 0) {
 904                         vm_page_unlock(m);
 905                         VM_OBJECT_UNLOCK(object);
 906
 907                         /*
 908                          * Held pages are essentially stuck in the
 909                          * queue.  So, they ought to be discounted
 910                          * from cnt.v_inactive_count.  See the
 911                          * calculation of the page_shortage for the
 912                          * loop over the active queue below.
 913                          */
 914                         addl_page_shortage++;
 915                         goto relock_queues;
 916                 }
 917
 918                 /*
 919                  * If the upper level VM system does not believe that the page
 920                  * is fully dirty, but it is mapped for write access, then we
 921                  * consult the pmap to see if the page's dirty status should
 922                  * be updated.
 923                  */
 924                 if (m->dirty != VM_PAGE_BITS_ALL &&
 925                     pmap_page_is_write_mapped(m)) {
 926                         /*
 927                          * Avoid a race condition: Unless write access is
 928                          * removed from the page, another processor could
 929                          * modify it before all access is removed by the call
 930                          * to vm_page_cache() below.  If vm_page_cache() finds
 931                          * that the page has been modified when it removes all
 932                          * access, it panics because it cannot cache dirty
 933                          * pages.  In principle, we could eliminate just write
 934                          * access here rather than all access.  In the expected
 935                          * case, when there are no last instant modifications
 936                          * to the page, removing all access will be cheaper
 937                          * overall.
 938                          */
 939                         if (pmap_is_modified(m))
 940                                 vm_page_dirty(m);
 941                         else if (m->dirty == 0)
 942                                 pmap_remove_all(m);
 943                 }
 944
 945                 if (m->valid == 0) {
 946                         /*
 947                          * Invalid pages can be easily freed
 948                          */
 949                         vm_page_free(m);
 950                         PCPU_INC(cnt.v_dfree);
 951                         --page_shortage;
 952                 } else if (m->dirty == 0) {
 953                         /*
 954                          * Clean pages can be placed onto the cache queue.
 955                          * This effectively frees them.
 956                          */
 957                         vm_page_cache(m);
 958                         --page_shortage;
 959                 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
 960                         /*
 961                          * Dirty pages need to be paged out, but flushing
 962                          * a page is extremely expensive verses freeing
 963                          * a clean page.  Rather then artificially limiting
 964                          * the number of pages we can flush, we instead give
 965                          * dirty pages extra priority on the inactive queue
 966                          * by forcing them to be cycled through the queue
 967                          * twice before being flushed, after which the
 968                          * (now clean) page will cycle through once more
 969                          * before being freed.  This significantly extends
 970                          * the thrash point for a heavily loaded machine.
 971                          */
 972                         m->flags |= PG_WINATCFLS;
 973                         vm_page_lock_queues();
 974                         queues_locked = TRUE;
 975                         vm_page_requeue(m);
 976                 } else if (maxlaunder > 0) {
 977                         /*
 978                          * We always want to try to flush some dirty pages if
 979                          * we encounter them, to keep the system stable.
 980                          * Normally this number is small, but under extreme
 981                          * pressure where there are insufficient clean pages
 982                          * on the inactive queue, we may have to go all out.
 983                          */
 984                         int swap_pageouts_ok, vfslocked = 0;
 985                         struct vnode *vp = NULL;
 986                         struct mount *mp = NULL;
 987
 988                         if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
 989                                 swap_pageouts_ok = 1;
 990                         } else {
 991                                 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
 992                                 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
 993                                 vm_page_count_min());
 994
 995                         }
 996
 997                         /*
 998                          * We don't bother paging objects that are "dead".
 999                          * Those objects are in a "rundown" state.
1000                          */
1001                         if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
1002                                 vm_page_lock_queues();
1003                                 vm_page_unlock(m);
1004                                 VM_OBJECT_UNLOCK(object);
1005                                 queues_locked = TRUE;
1006                                 vm_page_requeue(m);
1007                                 goto relock_queues;
1008                         }
1009
1010                         /*
1011                          * The object is already known NOT to be dead.   It
1012                          * is possible for the vget() to block the whole
1013                          * pageout daemon, but the new low-memory handling
1014                          * code should prevent it.
1015                          *
1016                          * The previous code skipped locked vnodes and, worse,
1017                          * reordered pages in the queue.  This results in
1018                          * completely non-deterministic operation and, on a
1019                          * busy system, can lead to extremely non-optimal
1020                          * pageouts.  For example, it can cause clean pages
1021                          * to be freed and dirty pages to be moved to the end
1022                          * of the queue.  Since dirty pages are also moved to
1023                          * the end of the queue once-cleaned, this gives
1024                          * way too large a weighting to defering the freeing
1025                          * of dirty pages.
1026                          *
1027                          * We can't wait forever for the vnode lock, we might
1028                          * deadlock due to a vn_read() getting stuck in
1029                          * vm_wait while holding this vnode.  We skip the
1030                          * vnode if we can't get it in a reasonable amount
1031                          * of time.
1032                          */
1033                         if (object->type == OBJT_VNODE) {
1034                                 vm_page_unlock(m);
1035                                 vp = object->handle;
1036                                 if (vp->v_type == VREG &&
1037                                     vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1038                                         mp = NULL;
1039                                         ++pageout_lock_miss;
1040                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1041                                                 vnodes_skipped++;
1042                                         goto unlock_and_continue;
1043                                 }
1044                                 KASSERT(mp != NULL,
1045                                     ("vp %p with NULL v_mount", vp));
1046                                 vm_object_reference_locked(object);
1047                                 VM_OBJECT_UNLOCK(object);
1048                                 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1049                                 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1050                                     curthread)) {
1051                                         VM_OBJECT_LOCK(object);
1052                                         ++pageout_lock_miss;
1053                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1054                                                 vnodes_skipped++;
1055                                         vp = NULL;
1056                                         goto unlock_and_continue;
1057                                 }
1058                                 VM_OBJECT_LOCK(object);
1059                                 vm_page_lock(m);
1060                                 vm_page_lock_queues();
1061                                 queues_locked = TRUE;
1062                                 /*
1063                                  * The page might have been moved to another
1064                                  * queue during potential blocking in vget()
1065                                  * above.  The page might have been freed and
1066                                  * reused for another vnode.
1067                                  */
1068                                 if (m->queue != PQ_INACTIVE ||
1069                                     m->object != object ||
1070                                     TAILQ_NEXT(m, pageq) != &marker) {
1071                                         vm_page_unlock(m);
1072                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1073                                                 vnodes_skipped++;
1074                                         goto unlock_and_continue;
1075                                 }
1076
1077                                 /*
1078                                  * The page may have been busied during the
1079                                  * blocking in vget().  We don't move the
1080                                  * page back onto the end of the queue so that
1081                                  * statistics are more correct if we don't.
1082                                  */
1083                                 if (m->busy || (m->oflags & VPO_BUSY)) {
1084                                         vm_page_unlock(m);
1085                                         goto unlock_and_continue;
1086                                 }
1087
1088                                 /*
1089                                  * If the page has become held it might
1090                                  * be undergoing I/O, so skip it
1091                                  */
1092                                 if (m->hold_count) {
1093                                         vm_page_unlock(m);
1094                                         vm_page_requeue(m);
1095                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1096                                                 vnodes_skipped++;
1097                                         goto unlock_and_continue;
1098                                 }
1099                                 vm_page_unlock_queues();
1100                                 queues_locked = FALSE;
1101                         }
1102
1103                         /*
1104                          * If a page is dirty, then it is either being washed
1105                          * (but not yet cleaned) or it is still in the
1106                          * laundry.  If it is still in the laundry, then we
1107                          * start the cleaning operation.
1108                          *
1109                          * decrement page_shortage on success to account for
1110                          * the (future) cleaned page.  Otherwise we could wind
1111                          * up laundering or cleaning too many pages.
1112                          */
1113                         if (vm_pageout_clean(m) != 0) {
1114                                 --page_shortage;
1115                                 --maxlaunder;
1116                         }
1117 unlock_and_continue:
1118                         vm_page_lock_assert(m, MA_NOTOWNED);
1119                         VM_OBJECT_UNLOCK(object);
1120                         if (mp != NULL) {
1121                                 if (queues_locked) {
1122                                         vm_page_unlock_queues();
1123                                         queues_locked = FALSE;
1124                                 }
1125                                 if (vp != NULL)
1126                                         vput(vp);
1127                                 VFS_UNLOCK_GIANT(vfslocked);
1128                                 vm_object_deallocate(object);
1129                                 vn_finished_write(mp);
1130                         }
1131                         vm_page_lock_assert(m, MA_NOTOWNED);
1132                         goto relock_queues;
1133                 }
1134                 vm_page_unlock(m);
1135                 VM_OBJECT_UNLOCK(object);
1136 relock_queues:
1137                 if (!queues_locked) {
1138                         vm_page_lock_queues();
1139                         queues_locked = TRUE;
1140                 }
1141                 next = TAILQ_NEXT(&marker, pageq);
1142                 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1143                     &marker, pageq);
1144         }
1145
1146         /*
1147          * Compute the number of pages we want to try to move from the
1148          * active queue to the inactive queue.
1149          */
1150         page_shortage = vm_paging_target() +
1151                 cnt.v_inactive_target - cnt.v_inactive_count;
1152         page_shortage += addl_page_shortage;
1153
1154         /*
1155          * Scan the active queue for things we can deactivate. We nominally
1156          * track the per-page activity counter and use it to locate
1157          * deactivation candidates.
1158          */
1159         pcount = cnt.v_active_count;
1160         m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1161         mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1162
1163         while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1164
1165                 KASSERT(m->queue == PQ_ACTIVE,
1166                     ("vm_pageout_scan: page %p isn't active", m));
1167
1168                 next = TAILQ_NEXT(m, pageq);
1169                 if ((m->flags & PG_MARKER) != 0) {
1170                         m = next;
1171                         continue;
1172                 }
1173                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1174                     ("Fictitious page %p cannot be in active queue", m));
1175                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1176                     ("Unmanaged page %p cannot be in active queue", m));
1177                 if (!vm_pageout_page_lock(m, &next)) {
1178                         vm_page_unlock(m);
1179                         m = next;
1180                         continue;
1181                 }
1182                 object = m->object;
1183                 if (!VM_OBJECT_TRYLOCK(object) &&
1184                     !vm_pageout_fallback_object_lock(m, &next)) {
1185                         VM_OBJECT_UNLOCK(object);
1186                         vm_page_unlock(m);
1187                         m = next;
1188                         continue;
1189                 }
1190
1191                 /*
1192                  * Don't deactivate pages that are busy.
1193                  */
1194                 if ((m->busy != 0) ||
1195                     (m->oflags & VPO_BUSY) ||
1196                     (m->hold_count != 0)) {
1197                         vm_page_unlock(m);
1198                         VM_OBJECT_UNLOCK(object);
1199                         vm_page_requeue(m);
1200                         m = next;
1201                         continue;
1202                 }
1203
1204                 /*
1205                  * The count for pagedaemon pages is done after checking the
1206                  * page for eligibility...
1207                  */
1208                 cnt.v_pdpages++;
1209
1210                 /*
1211                  * Check to see "how much" the page has been used.
1212                  */
1213                 actcount = 0;
1214                 if (object->ref_count != 0) {
1215                         if (m->aflags & PGA_REFERENCED) {
1216                                 actcount += 1;
1217                         }
1218                         actcount += pmap_ts_referenced(m);
1219                         if (actcount) {
1220                                 m->act_count += ACT_ADVANCE + actcount;
1221                                 if (m->act_count > ACT_MAX)
1222                                         m->act_count = ACT_MAX;
1223                         }
1224                 }
1225
1226                 /*
1227                  * Since we have "tested" this bit, we need to clear it now.
1228                  */
1229                 vm_page_aflag_clear(m, PGA_REFERENCED);
1230
1231                 /*
1232                  * Only if an object is currently being used, do we use the
1233                  * page activation count stats.
1234                  */
1235                 if (actcount && (object->ref_count != 0)) {
1236                         vm_page_requeue(m);
1237                 } else {
1238                         m->act_count -= min(m->act_count, ACT_DECLINE);
1239                         if (vm_pageout_algorithm ||
1240                             object->ref_count == 0 ||
1241                             m->act_count == 0) {
1242                                 page_shortage--;
1243                                 if (object->ref_count == 0) {
1244                                         KASSERT(!pmap_page_is_mapped(m),
1245                                     ("vm_pageout_scan: page %p is mapped", m));
1246                                         if (m->dirty == 0)
1247                                                 vm_page_cache(m);
1248                                         else
1249                                                 vm_page_deactivate(m);
1250                                 } else {
1251                                         vm_page_deactivate(m);
1252                                 }
1253                         } else {
1254                                 vm_page_requeue(m);
1255                         }
1256                 }
1257                 vm_page_unlock(m);
1258                 VM_OBJECT_UNLOCK(object);
1259                 m = next;
1260         }
1261         vm_page_unlock_queues();
1262 #if !defined(NO_SWAPPING)
1263         /*
1264          * Idle process swapout -- run once per second.
1265          */
1266         if (vm_swap_idle_enabled) {
1267                 static long lsec;
1268                 if (time_second != lsec) {
1269                         vm_req_vmdaemon(VM_SWAP_IDLE);
1270                         lsec = time_second;
1271                 }
1272         }
1273 #endif
1274
1275         /*
1276          * If we didn't get enough free pages, and we have skipped a vnode
1277          * in a writeable object, wakeup the sync daemon.  And kick swapout
1278          * if we did not get enough free pages.
1279          */
1280         if (vm_paging_target() > 0) {
1281                 if (vnodes_skipped && vm_page_count_min())
1282                         (void) speedup_syncer();
1283 #if !defined(NO_SWAPPING)
1284                 if (vm_swap_enabled && vm_page_count_target())
1285                         vm_req_vmdaemon(VM_SWAP_NORMAL);
1286 #endif
1287         }
1288
1289         /*
1290          * If we are critically low on one of RAM or swap and low on
1291          * the other, kill the largest process.  However, we avoid
1292          * doing this on the first pass in order to give ourselves a
1293          * chance to flush out dirty vnode-backed pages and to allow
1294          * active pages to be moved to the inactive queue and reclaimed.
1295          */
1296         if (pass != 0 &&
1297             ((swap_pager_avail < 64 && vm_page_count_min()) ||
1298              (swap_pager_full && vm_paging_target() > 0)))
1299                 vm_pageout_oom(VM_OOM_MEM);
1300 }
1301
1302
1303 void
1304 vm_pageout_oom(int shortage)
1305 {
1306         struct proc *p, *bigproc;
1307         vm_offset_t size, bigsize;
1308         struct thread *td;
1309         struct vmspace *vm;
1310
1311         /*
1312          * We keep the process bigproc locked once we find it to keep anyone
1313          * from messing with it; however, there is a possibility of
1314          * deadlock if process B is bigproc and one of it's child processes
1315          * attempts to propagate a signal to B while we are waiting for A's
1316          * lock while walking this list.  To avoid this, we don't block on
1317          * the process lock but just skip a process if it is already locked.
1318          */
1319         bigproc = NULL;
1320         bigsize = 0;
1321         sx_slock(&allproc_lock);
1322         FOREACH_PROC_IN_SYSTEM(p) {
1323                 int breakout;
1324
1325                 if (PROC_TRYLOCK(p) == 0)
1326                         continue;
1327                 /*
1328                  * If this is a system, protected or killed process, skip it.
1329                  */
1330                 if (p->p_state != PRS_NORMAL ||
1331                     (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1332                     (p->p_pid == 1) || P_KILLED(p) ||
1333                     ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1334                         PROC_UNLOCK(p);
1335                         continue;
1336                 }
1337                 /*
1338                  * If the process is in a non-running type state,
1339                  * don't touch it.  Check all the threads individually.
1340                  */
1341                 breakout = 0;
1342                 FOREACH_THREAD_IN_PROC(p, td) {
1343                         thread_lock(td);
1344                         if (!TD_ON_RUNQ(td) &&
1345                             !TD_IS_RUNNING(td) &&
1346                             !TD_IS_SLEEPING(td) &&
1347                             !TD_IS_SUSPENDED(td)) {
1348                                 thread_unlock(td);
1349                                 breakout = 1;
1350                                 break;
1351                         }
1352                         thread_unlock(td);
1353                 }
1354                 if (breakout) {
1355                         PROC_UNLOCK(p);
1356                         continue;
1357                 }
1358                 /*
1359                  * get the process size
1360                  */
1361                 vm = vmspace_acquire_ref(p);
1362                 if (vm == NULL) {
1363                         PROC_UNLOCK(p);
1364                         continue;
1365                 }
1366                 if (!vm_map_trylock_read(&vm->vm_map)) {
1367                         vmspace_free(vm);
1368                         PROC_UNLOCK(p);
1369                         continue;
1370                 }
1371                 size = vmspace_swap_count(vm);
1372                 vm_map_unlock_read(&vm->vm_map);
1373                 if (shortage == VM_OOM_MEM)
1374                         size += vmspace_resident_count(vm);
1375                 vmspace_free(vm);
1376                 /*
1377                  * if the this process is bigger than the biggest one
1378                  * remember it.
1379                  */
1380                 if (size > bigsize) {
1381                         if (bigproc != NULL)
1382                                 PROC_UNLOCK(bigproc);
1383                         bigproc = p;
1384                         bigsize = size;
1385                 } else
1386                         PROC_UNLOCK(p);
1387         }
1388         sx_sunlock(&allproc_lock);
1389         if (bigproc != NULL) {
1390                 killproc(bigproc, "out of swap space");
1391                 sched_nice(bigproc, PRIO_MIN);
1392                 PROC_UNLOCK(bigproc);
1393                 wakeup(&cnt.v_free_count);
1394         }
1395 }
1396
1397 /*
1398  * This routine tries to maintain the pseudo LRU active queue,
1399  * so that during long periods of time where there is no paging,
1400  * that some statistic accumulation still occurs.  This code
1401  * helps the situation where paging just starts to occur.
1402  */
1403 static void
1404 vm_pageout_page_stats()
1405 {
1406         vm_object_t object;
1407         vm_page_t m,next;
1408         int pcount,tpcount;             /* Number of pages to check */
1409         static int fullintervalcount = 0;
1410         int page_shortage;
1411
1412         page_shortage =
1413             (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1414             (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1415
1416         if (page_shortage <= 0)
1417                 return;
1418
1419         vm_page_lock_queues();
1420         pcount = cnt.v_active_count;
1421         fullintervalcount += vm_pageout_stats_interval;
1422         if (fullintervalcount < vm_pageout_full_stats_interval) {
1423                 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1424                     cnt.v_page_count;
1425                 if (pcount > tpcount)
1426                         pcount = tpcount;
1427         } else {
1428                 fullintervalcount = 0;
1429         }
1430
1431         m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1432         while ((m != NULL) && (pcount-- > 0)) {
1433                 int actcount;
1434
1435                 KASSERT(m->queue == PQ_ACTIVE,
1436                     ("vm_pageout_page_stats: page %p isn't active", m));
1437
1438                 next = TAILQ_NEXT(m, pageq);
1439                 if ((m->flags & PG_MARKER) != 0) {
1440                         m = next;
1441                         continue;
1442                 }
1443                 vm_page_lock_assert(m, MA_NOTOWNED);
1444                 if (!vm_pageout_page_lock(m, &next)) {
1445                         vm_page_unlock(m);
1446                         m = next;
1447                         continue;
1448                 }
1449                 object = m->object;
1450                 if (!VM_OBJECT_TRYLOCK(object) &&
1451                     !vm_pageout_fallback_object_lock(m, &next)) {
1452                         VM_OBJECT_UNLOCK(object);
1453                         vm_page_unlock(m);
1454                         m = next;
1455                         continue;
1456                 }
1457
1458                 /*
1459                  * Don't deactivate pages that are busy.
1460                  */
1461                 if ((m->busy != 0) ||
1462                     (m->oflags & VPO_BUSY) ||
1463                     (m->hold_count != 0)) {
1464                         vm_page_unlock(m);
1465                         VM_OBJECT_UNLOCK(object);
1466                         vm_page_requeue(m);
1467                         m = next;
1468                         continue;
1469                 }
1470
1471                 actcount = 0;
1472                 if (m->aflags & PGA_REFERENCED) {
1473                         vm_page_aflag_clear(m, PGA_REFERENCED);
1474                         actcount += 1;
1475                 }
1476
1477                 actcount += pmap_ts_referenced(m);
1478                 if (actcount) {
1479                         m->act_count += ACT_ADVANCE + actcount;
1480                         if (m->act_count > ACT_MAX)
1481                                 m->act_count = ACT_MAX;
1482                         vm_page_requeue(m);
1483                 } else {
1484                         if (m->act_count == 0) {
1485                                 /*
1486                                  * We turn off page access, so that we have
1487                                  * more accurate RSS stats.  We don't do this
1488                                  * in the normal page deactivation when the
1489                                  * system is loaded VM wise, because the
1490                                  * cost of the large number of page protect
1491                                  * operations would be higher than the value
1492                                  * of doing the operation.
1493                                  */
1494                                 pmap_remove_all(m);
1495                                 vm_page_deactivate(m);
1496                         } else {
1497                                 m->act_count -= min(m->act_count, ACT_DECLINE);
1498                                 vm_page_requeue(m);
1499                         }
1500                 }
1501                 vm_page_unlock(m);
1502                 VM_OBJECT_UNLOCK(object);
1503                 m = next;
1504         }
1505         vm_page_unlock_queues();
1506 }
1507
1508 /*
1509  *      vm_pageout is the high level pageout daemon.
1510  */
1511 static void
1512 vm_pageout()
1513 {
1514         int error, pass;
1515
1516         /*
1517          * Initialize some paging parameters.
1518          */
1519         cnt.v_interrupt_free_min = 2;
1520         if (cnt.v_page_count < 2000)
1521                 vm_pageout_page_count = 8;
1522
1523         /*
1524          * v_free_reserved needs to include enough for the largest
1525          * swap pager structures plus enough for any pv_entry structs
1526          * when paging.
1527          */
1528         if (cnt.v_page_count > 1024)
1529                 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1530         else
1531                 cnt.v_free_min = 4;
1532         cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1533             cnt.v_interrupt_free_min;
1534         cnt.v_free_reserved = vm_pageout_page_count +
1535             cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1536         cnt.v_free_severe = cnt.v_free_min / 2;
1537         cnt.v_free_min += cnt.v_free_reserved;
1538         cnt.v_free_severe += cnt.v_free_reserved;
1539
1540         /*
1541          * v_free_target and v_cache_min control pageout hysteresis.  Note
1542          * that these are more a measure of the VM cache queue hysteresis
1543          * then the VM free queue.  Specifically, v_free_target is the
1544          * high water mark (free+cache pages).
1545          *
1546          * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1547          * low water mark, while v_free_min is the stop.  v_cache_min must
1548          * be big enough to handle memory needs while the pageout daemon
1549          * is signalled and run to free more pages.
1550          */
1551         if (cnt.v_free_count > 6144)
1552                 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1553         else
1554                 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1555
1556         if (cnt.v_free_count > 2048) {
1557                 cnt.v_cache_min = cnt.v_free_target;
1558                 cnt.v_cache_max = 2 * cnt.v_cache_min;
1559                 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1560         } else {
1561                 cnt.v_cache_min = 0;
1562                 cnt.v_cache_max = 0;
1563                 cnt.v_inactive_target = cnt.v_free_count / 4;
1564         }
1565         if (cnt.v_inactive_target > cnt.v_free_count / 3)
1566                 cnt.v_inactive_target = cnt.v_free_count / 3;
1567
1568         /* XXX does not really belong here */
1569         if (vm_page_max_wired == 0)
1570                 vm_page_max_wired = cnt.v_free_count / 3;
1571
1572         if (vm_pageout_stats_max == 0)
1573                 vm_pageout_stats_max = cnt.v_free_target;
1574
1575         /*
1576          * Set interval in seconds for stats scan.
1577          */
1578         if (vm_pageout_stats_interval == 0)
1579                 vm_pageout_stats_interval = 5;
1580         if (vm_pageout_full_stats_interval == 0)
1581                 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1582
1583         swap_pager_swap_init();
1584         pass = 0;
1585         /*
1586          * The pageout daemon is never done, so loop forever.
1587          */
1588         while (TRUE) {
1589                 /*
1590                  * If we have enough free memory, wakeup waiters.  Do
1591                  * not clear vm_pages_needed until we reach our target,
1592                  * otherwise we may be woken up over and over again and
1593                  * waste a lot of cpu.
1594                  */
1595                 mtx_lock(&vm_page_queue_free_mtx);
1596                 if (vm_pages_needed && !vm_page_count_min()) {
1597                         if (!vm_paging_needed())
1598                                 vm_pages_needed = 0;
1599                         wakeup(&cnt.v_free_count);
1600                 }
1601                 if (vm_pages_needed) {
1602                         /*
1603                          * Still not done, take a second pass without waiting
1604                          * (unlimited dirty cleaning), otherwise sleep a bit
1605                          * and try again.
1606                          */
1607                         ++pass;
1608                         if (pass > 1)
1609                                 msleep(&vm_pages_needed,
1610                                     &vm_page_queue_free_mtx, PVM, "psleep",
1611                                     hz / 2);
1612                 } else {
1613                         /*
1614                          * Good enough, sleep & handle stats.  Prime the pass
1615                          * for the next run.
1616                          */
1617                         if (pass > 1)
1618                                 pass = 1;
1619                         else
1620                                 pass = 0;
1621                         error = msleep(&vm_pages_needed,
1622                             &vm_page_queue_free_mtx, PVM, "psleep",
1623                             vm_pageout_stats_interval * hz);
1624                         if (error && !vm_pages_needed) {
1625                                 mtx_unlock(&vm_page_queue_free_mtx);
1626                                 pass = 0;
1627                                 vm_pageout_page_stats();
1628                                 continue;
1629                         }
1630                 }
1631                 if (vm_pages_needed)
1632                         cnt.v_pdwakeups++;
1633                 mtx_unlock(&vm_page_queue_free_mtx);
1634                 vm_pageout_scan(pass);
1635         }
1636 }
1637
1638 /*
1639  * Unless the free page queue lock is held by the caller, this function
1640  * should be regarded as advisory.  Specifically, the caller should
1641  * not msleep() on &cnt.v_free_count following this function unless
1642  * the free page queue lock is held until the msleep() is performed.
1643  */
1644 void
1645 pagedaemon_wakeup()
1646 {
1647
1648         if (!vm_pages_needed && curthread->td_proc != pageproc) {
1649                 vm_pages_needed = 1;
1650                 wakeup(&vm_pages_needed);
1651         }
1652 }
1653
1654 #if !defined(NO_SWAPPING)
1655 static void
1656 vm_req_vmdaemon(int req)
1657 {
1658         static int lastrun = 0;
1659
1660         mtx_lock(&vm_daemon_mtx);
1661         vm_pageout_req_swapout |= req;
1662         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1663                 wakeup(&vm_daemon_needed);
1664                 lastrun = ticks;
1665         }
1666         mtx_unlock(&vm_daemon_mtx);
1667 }
1668
1669 static void
1670 vm_daemon()
1671 {
1672         struct rlimit rsslim;
1673         struct proc *p;
1674         struct thread *td;
1675         struct vmspace *vm;
1676         int breakout, swapout_flags, tryagain, attempts;
1677 #ifdef RACCT
1678         uint64_t rsize, ravailable;
1679 #endif
1680
1681         while (TRUE) {
1682                 mtx_lock(&vm_daemon_mtx);
1683 #ifdef RACCT
1684                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz);
1685 #else
1686                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1687 #endif
1688                 swapout_flags = vm_pageout_req_swapout;
1689                 vm_pageout_req_swapout = 0;
1690                 mtx_unlock(&vm_daemon_mtx);
1691                 if (swapout_flags)
1692                         swapout_procs(swapout_flags);
1693
1694                 /*
1695                  * scan the processes for exceeding their rlimits or if
1696                  * process is swapped out -- deactivate pages
1697                  */
1698                 tryagain = 0;
1699                 attempts = 0;
1700 again:
1701                 attempts++;
1702                 sx_slock(&allproc_lock);
1703                 FOREACH_PROC_IN_SYSTEM(p) {
1704                         vm_pindex_t limit, size;
1705
1706                         /*
1707                          * if this is a system process or if we have already
1708                          * looked at this process, skip it.
1709                          */
1710                         PROC_LOCK(p);
1711                         if (p->p_state != PRS_NORMAL ||
1712                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1713                                 PROC_UNLOCK(p);
1714                                 continue;
1715                         }
1716                         /*
1717                          * if the process is in a non-running type state,
1718                          * don't touch it.
1719                          */
1720                         breakout = 0;
1721                         FOREACH_THREAD_IN_PROC(p, td) {
1722                                 thread_lock(td);
1723                                 if (!TD_ON_RUNQ(td) &&
1724                                     !TD_IS_RUNNING(td) &&
1725                                     !TD_IS_SLEEPING(td) &&
1726                                     !TD_IS_SUSPENDED(td)) {
1727                                         thread_unlock(td);
1728                                         breakout = 1;
1729                                         break;
1730                                 }
1731                                 thread_unlock(td);
1732                         }
1733                         if (breakout) {
1734                                 PROC_UNLOCK(p);
1735                                 continue;
1736                         }
1737                         /*
1738                          * get a limit
1739                          */
1740                         lim_rlimit(p, RLIMIT_RSS, &rsslim);
1741                         limit = OFF_TO_IDX(
1742                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
1743
1744                         /*
1745                          * let processes that are swapped out really be
1746                          * swapped out set the limit to nothing (will force a
1747                          * swap-out.)
1748                          */
1749                         if ((p->p_flag & P_INMEM) == 0)
1750                                 limit = 0;      /* XXX */
1751                         vm = vmspace_acquire_ref(p);
1752                         PROC_UNLOCK(p);
1753                         if (vm == NULL)
1754                                 continue;
1755
1756                         size = vmspace_resident_count(vm);
1757                         if (limit >= 0 && size >= limit) {
1758                                 vm_pageout_map_deactivate_pages(
1759                                     &vm->vm_map, limit);
1760                         }
1761 #ifdef RACCT
1762                         rsize = IDX_TO_OFF(size);
1763                         PROC_LOCK(p);
1764                         racct_set(p, RACCT_RSS, rsize);
1765                         ravailable = racct_get_available(p, RACCT_RSS);
1766                         PROC_UNLOCK(p);
1767                         if (rsize > ravailable) {
1768                                 /*
1769                                  * Don't be overly aggressive; this might be
1770                                  * an innocent process, and the limit could've
1771                                  * been exceeded by some memory hog.  Don't
1772                                  * try to deactivate more than 1/4th of process'
1773                                  * resident set size.
1774                                  */
1775                                 if (attempts <= 8) {
1776                                         if (ravailable < rsize - (rsize / 4))
1777                                                 ravailable = rsize - (rsize / 4);
1778                                 }
1779                                 vm_pageout_map_deactivate_pages(
1780                                     &vm->vm_map, OFF_TO_IDX(ravailable));
1781                                 /* Update RSS usage after paging out. */
1782                                 size = vmspace_resident_count(vm);
1783                                 rsize = IDX_TO_OFF(size);
1784                                 PROC_LOCK(p);
1785                                 racct_set(p, RACCT_RSS, rsize);
1786                                 PROC_UNLOCK(p);
1787                                 if (rsize > ravailable)
1788                                         tryagain = 1;
1789                         }
1790 #endif
1791                         vmspace_free(vm);
1792                 }
1793                 sx_sunlock(&allproc_lock);
1794                 if (tryagain != 0 && attempts <= 10)
1795                         goto again;
1796         }
1797 }
1798 #endif                  /* !defined(NO_SWAPPING) */