sys/vm/vm_swapout.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 #include <sys/cdefs.h>
  72 __FBSDID("$FreeBSD$");
  73
  74 #include "opt_kstack_pages.h"
  75 #include "opt_kstack_max_pages.h"
  76 #include "opt_vm.h"
  77
  78 #include <sys/param.h>
  79 #include <sys/systm.h>
  80 #include <sys/limits.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/_kstack_cache.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 /* the kernel process "vm_daemon" */
 116 static void vm_daemon(void);
 117 static struct proc *vmproc;
 118
 119 static struct kproc_desc vm_kp = {
 120         "vmdaemon",
 121         vm_daemon,
 122         &vmproc
 123 };
 124 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
 125
 126 static int vm_swap_enabled = 1;
 127 static int vm_swap_idle_enabled = 0;
 128
 129 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW,
 130     &vm_swap_enabled, 0,
 131     "Enable entire process swapout");
 132 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW,
 133     &vm_swap_idle_enabled, 0,
 134     "Allow swapout on idle criteria");
 135
 136 /*
 137  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
 138  */
 139 static int swap_idle_threshold1 = 2;
 140 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
 141     &swap_idle_threshold1, 0,
 142     "Guaranteed swapped in time for a process");
 143
 144 /*
 145  * Swap_idle_threshold2 is the time that a process can be idle before
 146  * it will be swapped out, if idle swapping is enabled.
 147  */
 148 static int swap_idle_threshold2 = 10;
 149 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
 150     &swap_idle_threshold2, 0,
 151     "Time before a process will be swapped out");
 152
 153 static int vm_pageout_req_swapout;      /* XXX */
 154 static int vm_daemon_needed;
 155 static struct mtx vm_daemon_mtx;
 156 /* Allow for use by vm_pageout before vm_daemon is initialized. */
 157 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
 158
 159 static void swapclear(struct proc *);
 160 static int swapout(struct proc *);
 161 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
 162 static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long);
 163 static void swapout_procs(int action);
 164 static void vm_req_vmdaemon(int req);
 165 static void vm_thread_swapin(struct thread *td);
 166 static void vm_thread_swapout(struct thread *td);
 167
 168 /*
 169  *      vm_swapout_object_deactivate_pages
 170  *
 171  *      Deactivate enough pages to satisfy the inactive target
 172  *      requirements.
 173  *
 174  *      The object and map must be locked.
 175  */
 176 static void
 177 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
 178     long desired)
 179 {
 180         vm_object_t backing_object, object;
 181         vm_page_t p;
 182         int act_delta, remove_mode;
 183
 184         VM_OBJECT_ASSERT_LOCKED(first_object);
 185         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 186                 return;
 187         for (object = first_object;; object = backing_object) {
 188                 if (pmap_resident_count(pmap) <= desired)
 189                         goto unlock_return;
 190                 VM_OBJECT_ASSERT_LOCKED(object);
 191                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 192                     object->paging_in_progress != 0)
 193                         goto unlock_return;
 194
 195                 remove_mode = 0;
 196                 if (object->shadow_count > 1)
 197                         remove_mode = 1;
 198                 /*
 199                  * Scan the object's entire memory queue.
 200                  */
 201                 TAILQ_FOREACH(p, &object->memq, listq) {
 202                         if (pmap_resident_count(pmap) <= desired)
 203                                 goto unlock_return;
 204                         if (should_yield())
 205                                 goto unlock_return;
 206                         if (vm_page_busied(p))
 207                                 continue;
 208                         PCPU_INC(cnt.v_pdpages);
 209                         vm_page_lock(p);
 210                         if (p->wire_count != 0 || p->hold_count != 0 ||
 211                             !pmap_page_exists_quick(pmap, p)) {
 212                                 vm_page_unlock(p);
 213                                 continue;
 214                         }
 215                         act_delta = pmap_ts_referenced(p);
 216                         if ((p->aflags & PGA_REFERENCED) != 0) {
 217                                 if (act_delta == 0)
 218                                         act_delta = 1;
 219                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 220                         }
 221                         if (!vm_page_active(p) && act_delta != 0) {
 222                                 vm_page_activate(p);
 223                                 p->act_count += act_delta;
 224                         } else if (vm_page_active(p)) {
 225                                 if (act_delta == 0) {
 226                                         p->act_count -= min(p->act_count,
 227                                             ACT_DECLINE);
 228                                         if (!remove_mode && p->act_count == 0) {
 229                                                 pmap_remove_all(p);
 230                                                 vm_page_deactivate(p);
 231                                         } else
 232                                                 vm_page_requeue(p);
 233                                 } else {
 234                                         vm_page_activate(p);
 235                                         if (p->act_count < ACT_MAX -
 236                                             ACT_ADVANCE)
 237                                                 p->act_count += ACT_ADVANCE;
 238                                         vm_page_requeue(p);
 239                                 }
 240                         } else if (vm_page_inactive(p))
 241                                 pmap_remove_all(p);
 242                         vm_page_unlock(p);
 243                 }
 244                 if ((backing_object = object->backing_object) == NULL)
 245                         goto unlock_return;
 246                 VM_OBJECT_RLOCK(backing_object);
 247                 if (object != first_object)
 248                         VM_OBJECT_RUNLOCK(object);
 249         }
 250 unlock_return:
 251         if (object != first_object)
 252                 VM_OBJECT_RUNLOCK(object);
 253 }
 254
 255 /*
 256  * deactivate some number of pages in a map, try to do it fairly, but
 257  * that is really hard to do.
 258  */
 259 static void
 260 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
 261 {
 262         vm_map_entry_t tmpe;
 263         vm_object_t obj, bigobj;
 264         int nothingwired;
 265
 266         if (!vm_map_trylock_read(map))
 267                 return;
 268
 269         bigobj = NULL;
 270         nothingwired = TRUE;
 271
 272         /*
 273          * first, search out the biggest object, and try to free pages from
 274          * that.
 275          */
 276         tmpe = map->header.next;
 277         while (tmpe != &map->header) {
 278                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 279                         obj = tmpe->object.vm_object;
 280                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 281                                 if (obj->shadow_count <= 1 &&
 282                                     (bigobj == NULL ||
 283                                      bigobj->resident_page_count <
 284                                      obj->resident_page_count)) {
 285                                         if (bigobj != NULL)
 286                                                 VM_OBJECT_RUNLOCK(bigobj);
 287                                         bigobj = obj;
 288                                 } else
 289                                         VM_OBJECT_RUNLOCK(obj);
 290                         }
 291                 }
 292                 if (tmpe->wired_count > 0)
 293                         nothingwired = FALSE;
 294                 tmpe = tmpe->next;
 295         }
 296
 297         if (bigobj != NULL) {
 298                 vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
 299                 VM_OBJECT_RUNLOCK(bigobj);
 300         }
 301         /*
 302          * Next, hunt around for other pages to deactivate.  We actually
 303          * do this search sort of wrong -- .text first is not the best idea.
 304          */
 305         tmpe = map->header.next;
 306         while (tmpe != &map->header) {
 307                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 308                         break;
 309                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 310                         obj = tmpe->object.vm_object;
 311                         if (obj != NULL) {
 312                                 VM_OBJECT_RLOCK(obj);
 313                                 vm_swapout_object_deactivate_pages(map->pmap,
 314                                     obj, desired);
 315                                 VM_OBJECT_RUNLOCK(obj);
 316                         }
 317                 }
 318                 tmpe = tmpe->next;
 319         }
 320
 321         /*
 322          * Remove all mappings if a process is swapped out, this will free page
 323          * table pages.
 324          */
 325         if (desired == 0 && nothingwired) {
 326                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 327                     vm_map_max(map));
 328         }
 329
 330         vm_map_unlock_read(map);
 331 }
 332
 333 /*
 334  * Swap out requests
 335  */
 336 #define VM_SWAP_NORMAL 1
 337 #define VM_SWAP_IDLE 2
 338
 339 void
 340 vm_swapout_run(void)
 341 {
 342
 343         if (vm_swap_enabled)
 344                 vm_req_vmdaemon(VM_SWAP_NORMAL);
 345 }
 346
 347 /*
 348  * Idle process swapout -- run once per second when pagedaemons are
 349  * reclaiming pages.
 350  */
 351 void
 352 vm_swapout_run_idle(void)
 353 {
 354         static long lsec;
 355
 356         if (!vm_swap_idle_enabled || time_second == lsec)
 357                 return;
 358         vm_req_vmdaemon(VM_SWAP_IDLE);
 359         lsec = time_second;
 360 }
 361
 362 static void
 363 vm_req_vmdaemon(int req)
 364 {
 365         static int lastrun = 0;
 366
 367         mtx_lock(&vm_daemon_mtx);
 368         vm_pageout_req_swapout |= req;
 369         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
 370                 wakeup(&vm_daemon_needed);
 371                 lastrun = ticks;
 372         }
 373         mtx_unlock(&vm_daemon_mtx);
 374 }
 375
 376 static void
 377 vm_daemon(void)
 378 {
 379         struct rlimit rsslim;
 380         struct proc *p;
 381         struct thread *td;
 382         struct vmspace *vm;
 383         int breakout, swapout_flags, tryagain, attempts;
 384 #ifdef RACCT
 385         uint64_t rsize, ravailable;
 386 #endif
 387
 388         while (TRUE) {
 389                 mtx_lock(&vm_daemon_mtx);
 390                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
 391 #ifdef RACCT
 392                     racct_enable ? hz : 0
 393 #else
 394                     0
 395 #endif
 396                 );
 397                 swapout_flags = vm_pageout_req_swapout;
 398                 vm_pageout_req_swapout = 0;
 399                 mtx_unlock(&vm_daemon_mtx);
 400                 if (swapout_flags)
 401                         swapout_procs(swapout_flags);
 402
 403                 /*
 404                  * scan the processes for exceeding their rlimits or if
 405                  * process is swapped out -- deactivate pages
 406                  */
 407                 tryagain = 0;
 408                 attempts = 0;
 409 again:
 410                 attempts++;
 411                 sx_slock(&allproc_lock);
 412                 FOREACH_PROC_IN_SYSTEM(p) {
 413                         vm_pindex_t limit, size;
 414
 415                         /*
 416                          * if this is a system process or if we have already
 417                          * looked at this process, skip it.
 418                          */
 419                         PROC_LOCK(p);
 420                         if (p->p_state != PRS_NORMAL ||
 421                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
 422                                 PROC_UNLOCK(p);
 423                                 continue;
 424                         }
 425                         /*
 426                          * if the process is in a non-running type state,
 427                          * don't touch it.
 428                          */
 429                         breakout = 0;
 430                         FOREACH_THREAD_IN_PROC(p, td) {
 431                                 thread_lock(td);
 432                                 if (!TD_ON_RUNQ(td) &&
 433                                     !TD_IS_RUNNING(td) &&
 434                                     !TD_IS_SLEEPING(td) &&
 435                                     !TD_IS_SUSPENDED(td)) {
 436                                         thread_unlock(td);
 437                                         breakout = 1;
 438                                         break;
 439                                 }
 440                                 thread_unlock(td);
 441                         }
 442                         if (breakout) {
 443                                 PROC_UNLOCK(p);
 444                                 continue;
 445                         }
 446                         /*
 447                          * get a limit
 448                          */
 449                         lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
 450                         limit = OFF_TO_IDX(
 451                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
 452
 453                         /*
 454                          * let processes that are swapped out really be
 455                          * swapped out set the limit to nothing (will force a
 456                          * swap-out.)
 457                          */
 458                         if ((p->p_flag & P_INMEM) == 0)
 459                                 limit = 0;      /* XXX */
 460                         vm = vmspace_acquire_ref(p);
 461                         _PHOLD_LITE(p);
 462                         PROC_UNLOCK(p);
 463                         if (vm == NULL) {
 464                                 PRELE(p);
 465                                 continue;
 466                         }
 467                         sx_sunlock(&allproc_lock);
 468
 469                         size = vmspace_resident_count(vm);
 470                         if (size >= limit) {
 471                                 vm_swapout_map_deactivate_pages(
 472                                     &vm->vm_map, limit);
 473                                 size = vmspace_resident_count(vm);
 474                         }
 475 #ifdef RACCT
 476                         if (racct_enable) {
 477                                 rsize = IDX_TO_OFF(size);
 478                                 PROC_LOCK(p);
 479                                 if (p->p_state == PRS_NORMAL)
 480                                         racct_set(p, RACCT_RSS, rsize);
 481                                 ravailable = racct_get_available(p, RACCT_RSS);
 482                                 PROC_UNLOCK(p);
 483                                 if (rsize > ravailable) {
 484                                         /*
 485                                          * Don't be overly aggressive; this
 486                                          * might be an innocent process,
 487                                          * and the limit could've been exceeded
 488                                          * by some memory hog.  Don't try
 489                                          * to deactivate more than 1/4th
 490                                          * of process' resident set size.
 491                                          */
 492                                         if (attempts <= 8) {
 493                                                 if (ravailable < rsize -
 494                                                     (rsize / 4)) {
 495                                                         ravailable = rsize -
 496                                                             (rsize / 4);
 497                                                 }
 498                                         }
 499                                         vm_swapout_map_deactivate_pages(
 500                                             &vm->vm_map,
 501                                             OFF_TO_IDX(ravailable));
 502                                         /* Update RSS usage after paging out. */
 503                                         size = vmspace_resident_count(vm);
 504                                         rsize = IDX_TO_OFF(size);
 505                                         PROC_LOCK(p);
 506                                         if (p->p_state == PRS_NORMAL)
 507                                                 racct_set(p, RACCT_RSS, rsize);
 508                                         PROC_UNLOCK(p);
 509                                         if (rsize > ravailable)
 510                                                 tryagain = 1;
 511                                 }
 512                         }
 513 #endif
 514                         vmspace_free(vm);
 515                         sx_slock(&allproc_lock);
 516                         PRELE(p);
 517                 }
 518                 sx_sunlock(&allproc_lock);
 519                 if (tryagain != 0 && attempts <= 10) {
 520                         maybe_yield();
 521                         goto again;
 522                 }
 523         }
 524 }
 525
 526 /*
 527  * Allow a thread's kernel stack to be paged out.
 528  */
 529 static void
 530 vm_thread_swapout(struct thread *td)
 531 {
 532         vm_object_t ksobj;
 533         vm_page_t m;
 534         int i, pages;
 535
 536         cpu_thread_swapout(td);
 537         pages = td->td_kstack_pages;
 538         ksobj = td->td_kstack_obj;
 539         pmap_qremove(td->td_kstack, pages);
 540         VM_OBJECT_WLOCK(ksobj);
 541         for (i = 0; i < pages; i++) {
 542                 m = vm_page_lookup(ksobj, i);
 543                 if (m == NULL)
 544                         panic("vm_thread_swapout: kstack already missing?");
 545                 vm_page_dirty(m);
 546                 vm_page_lock(m);
 547                 vm_page_unwire(m, PQ_INACTIVE);
 548                 vm_page_unlock(m);
 549         }
 550         VM_OBJECT_WUNLOCK(ksobj);
 551 }
 552
 553 /*
 554  * Bring the kernel stack for a specified thread back in.
 555  */
 556 static void
 557 vm_thread_swapin(struct thread *td)
 558 {
 559         vm_object_t ksobj;
 560         vm_page_t ma[KSTACK_MAX_PAGES];
 561         int a, count, i, j, pages, rv;
 562
 563         pages = td->td_kstack_pages;
 564         ksobj = td->td_kstack_obj;
 565         VM_OBJECT_WLOCK(ksobj);
 566         (void)vm_page_grab_pages(ksobj, 0, VM_ALLOC_NORMAL | VM_ALLOC_WIRED, ma,
 567             pages);
 568         for (i = 0; i < pages;) {
 569                 vm_page_assert_xbusied(ma[i]);
 570                 if (ma[i]->valid == VM_PAGE_BITS_ALL) {
 571                         vm_page_xunbusy(ma[i]);
 572                         i++;
 573                         continue;
 574                 }
 575                 vm_object_pip_add(ksobj, 1);
 576                 for (j = i + 1; j < pages; j++)
 577                         if (ma[j]->valid == VM_PAGE_BITS_ALL)
 578                                 break;
 579                 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
 580                 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
 581                 count = min(a + 1, j - i);
 582                 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
 583                 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
 584                     __func__, td->td_proc->p_pid));
 585                 vm_object_pip_wakeup(ksobj);
 586                 for (j = i; j < i + count; j++)
 587                         vm_page_xunbusy(ma[j]);
 588                 i += count;
 589         }
 590         VM_OBJECT_WUNLOCK(ksobj);
 591         pmap_qenter(td->td_kstack, ma, pages);
 592         cpu_thread_swapin(td);
 593 }
 594
 595 void
 596 faultin(struct proc *p)
 597 {
 598         struct thread *td;
 599
 600         PROC_LOCK_ASSERT(p, MA_OWNED);
 601         /*
 602          * If another process is swapping in this process,
 603          * just wait until it finishes.
 604          */
 605         if (p->p_flag & P_SWAPPINGIN) {
 606                 while (p->p_flag & P_SWAPPINGIN)
 607                         msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
 608                 return;
 609         }
 610         if ((p->p_flag & P_INMEM) == 0) {
 611                 /*
 612                  * Don't let another thread swap process p out while we are
 613                  * busy swapping it in.
 614                  */
 615                 ++p->p_lock;
 616                 p->p_flag |= P_SWAPPINGIN;
 617                 PROC_UNLOCK(p);
 618
 619                 /*
 620                  * We hold no lock here because the list of threads
 621                  * can not change while all threads in the process are
 622                  * swapped out.
 623                  */
 624                 FOREACH_THREAD_IN_PROC(p, td)
 625                         vm_thread_swapin(td);
 626                 PROC_LOCK(p);
 627                 swapclear(p);
 628                 p->p_swtick = ticks;
 629
 630                 wakeup(&p->p_flag);
 631
 632                 /* Allow other threads to swap p out now. */
 633                 --p->p_lock;
 634         }
 635 }
 636
 637 /*
 638  * This swapin algorithm attempts to swap-in processes only if there
 639  * is enough space for them.  Of course, if a process waits for a long
 640  * time, it will be swapped in anyway.
 641  */
 642 void
 643 swapper(void)
 644 {
 645         struct proc *p, *pp;
 646         struct thread *td;
 647         int ppri, pri, slptime, swtime;
 648
 649 loop:
 650         if (vm_page_count_min()) {
 651                 VM_WAIT;
 652                 goto loop;
 653         }
 654
 655         pp = NULL;
 656         ppri = INT_MIN;
 657         sx_slock(&allproc_lock);
 658         FOREACH_PROC_IN_SYSTEM(p) {
 659                 PROC_LOCK(p);
 660                 if (p->p_state == PRS_NEW ||
 661                     p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
 662                         PROC_UNLOCK(p);
 663                         continue;
 664                 }
 665                 swtime = (ticks - p->p_swtick) / hz;
 666                 FOREACH_THREAD_IN_PROC(p, td) {
 667                         /*
 668                          * An otherwise runnable thread of a process
 669                          * swapped out has only the TDI_SWAPPED bit set.
 670                          */
 671                         thread_lock(td);
 672                         if (td->td_inhibitors == TDI_SWAPPED) {
 673                                 slptime = (ticks - td->td_slptick) / hz;
 674                                 pri = swtime + slptime;
 675                                 if ((td->td_flags & TDF_SWAPINREQ) == 0)
 676                                         pri -= p->p_nice * 8;
 677                                 /*
 678                                  * if this thread is higher priority
 679                                  * and there is enough space, then select
 680                                  * this process instead of the previous
 681                                  * selection.
 682                                  */
 683                                 if (pri > ppri) {
 684                                         pp = p;
 685                                         ppri = pri;
 686                                 }
 687                         }
 688                         thread_unlock(td);
 689                 }
 690                 PROC_UNLOCK(p);
 691         }
 692         sx_sunlock(&allproc_lock);
 693
 694         /*
 695          * Nothing to do, back to sleep.
 696          */
 697         if ((p = pp) == NULL) {
 698                 tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2);
 699                 goto loop;
 700         }
 701         PROC_LOCK(p);
 702
 703         /*
 704          * Another process may be bringing or may have already
 705          * brought this process in while we traverse all threads.
 706          * Or, this process may even be being swapped out again.
 707          */
 708         if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
 709                 PROC_UNLOCK(p);
 710                 goto loop;
 711         }
 712
 713         /*
 714          * We would like to bring someone in.
 715          */
 716         faultin(p);
 717         PROC_UNLOCK(p);
 718         goto loop;
 719 }
 720
 721 /*
 722  * First, if any processes have been sleeping or stopped for at least
 723  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
 724  * no such processes exist, then the longest-sleeping or stopped
 725  * process is swapped out.  Finally, and only as a last resort, if
 726  * there are no sleeping or stopped processes, the longest-resident
 727  * process is swapped out.
 728  */
 729 static void
 730 swapout_procs(int action)
 731 {
 732         struct proc *p;
 733         struct thread *td;
 734         int minslptime, slptime;
 735         bool didswap;
 736
 737         minslptime = 100000;
 738         didswap = false;
 739 retry:
 740         sx_slock(&allproc_lock);
 741         FOREACH_PROC_IN_SYSTEM(p) {
 742                 PROC_LOCK(p);
 743                 /*
 744                  * Watch out for a process in
 745                  * creation.  It may have no
 746                  * address space or lock yet.
 747                  */
 748                 if (p->p_state == PRS_NEW) {
 749                         PROC_UNLOCK(p);
 750                         continue;
 751                 }
 752                 /*
 753                  * An aio daemon switches its
 754                  * address space while running.
 755                  * Perform a quick check whether
 756                  * a process has P_SYSTEM.
 757                  * Filter out exiting processes.
 758                  */
 759                 if ((p->p_flag & (P_SYSTEM | P_WEXIT)) != 0) {
 760                         PROC_UNLOCK(p);
 761                         continue;
 762                 }
 763                 _PHOLD_LITE(p);
 764                 PROC_UNLOCK(p);
 765                 sx_sunlock(&allproc_lock);
 766
 767                 PROC_LOCK(p);
 768                 if (p->p_lock != 1 || (p->p_flag & (P_STOPPED_SINGLE |
 769                     P_TRACED | P_SYSTEM)) != 0)
 770                         goto nextproc;
 771
 772                 /*
 773                  * Only aiod changes vmspace.  However, it will be
 774                  * skipped because of the if statement above checking
 775                  * for P_SYSTEM.
 776                  */
 777                 if ((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) !=
 778                     P_INMEM)
 779                         goto nextproc;
 780
 781                 switch (p->p_state) {
 782                 default:
 783                         /*
 784                          * Don't swap out processes in any sort
 785                          * of 'special' state.
 786                          */
 787                         break;
 788
 789                 case PRS_NORMAL:
 790                         /*
 791                          * do not swapout a realtime process
 792                          * Check all the thread groups..
 793                          */
 794                         FOREACH_THREAD_IN_PROC(p, td) {
 795                                 thread_lock(td);
 796                                 if (PRI_IS_REALTIME(td->td_pri_class)) {
 797                                         thread_unlock(td);
 798                                         goto nextproc;
 799                                 }
 800                                 slptime = (ticks - td->td_slptick) / hz;
 801                                 /*
 802                                  * Guarantee swap_idle_threshold1
 803                                  * time in memory.
 804                                  */
 805                                 if (slptime < swap_idle_threshold1) {
 806                                         thread_unlock(td);
 807                                         goto nextproc;
 808                                 }
 809
 810                                 /*
 811                                  * Do not swapout a process if it is
 812                                  * waiting on a critical event of some
 813                                  * kind or there is a thread whose
 814                                  * pageable memory may be accessed.
 815                                  *
 816                                  * This could be refined to support
 817                                  * swapping out a thread.
 818                                  */
 819                                 if (!thread_safetoswapout(td)) {
 820                                         thread_unlock(td);
 821                                         goto nextproc;
 822                                 }
 823                                 /*
 824                                  * If the system is under memory stress,
 825                                  * or if we are swapping
 826                                  * idle processes >= swap_idle_threshold2,
 827                                  * then swap the process out.
 828                                  */
 829                                 if ((action & VM_SWAP_NORMAL) == 0 &&
 830                                     ((action & VM_SWAP_IDLE) == 0 ||
 831                                     slptime < swap_idle_threshold2)) {
 832                                         thread_unlock(td);
 833                                         goto nextproc;
 834                                 }
 835
 836                                 if (minslptime > slptime)
 837                                         minslptime = slptime;
 838                                 thread_unlock(td);
 839                         }
 840
 841                         /*
 842                          * If the pageout daemon didn't free enough pages,
 843                          * or if this process is idle and the system is
 844                          * configured to swap proactively, swap it out.
 845                          */
 846                         if ((action & VM_SWAP_NORMAL) != 0 ||
 847                             ((action & VM_SWAP_IDLE) != 0 &&
 848                             minslptime > swap_idle_threshold2)) {
 849                                 _PRELE(p);
 850                                 if (swapout(p) == 0)
 851                                         didswap = true;
 852                                 PROC_UNLOCK(p);
 853                                 goto retry;
 854                         }
 855                 }
 856 nextproc:
 857                 PROC_UNLOCK(p);
 858                 sx_slock(&allproc_lock);
 859                 PRELE(p);
 860         }
 861         sx_sunlock(&allproc_lock);
 862         /*
 863          * If we swapped something out, and another process needed memory,
 864          * then wakeup the sched process.
 865          */
 866         if (didswap)
 867                 wakeup(&proc0);
 868 }
 869
 870 static void
 871 swapclear(struct proc *p)
 872 {
 873         struct thread *td;
 874
 875         PROC_LOCK_ASSERT(p, MA_OWNED);
 876
 877         FOREACH_THREAD_IN_PROC(p, td) {
 878                 thread_lock(td);
 879                 td->td_flags |= TDF_INMEM;
 880                 td->td_flags &= ~TDF_SWAPINREQ;
 881                 TD_CLR_SWAPPED(td);
 882                 if (TD_CAN_RUN(td))
 883                         if (setrunnable(td)) {
 884 #ifdef INVARIANTS
 885                                 /*
 886                                  * XXX: We just cleared TDI_SWAPPED
 887                                  * above and set TDF_INMEM, so this
 888                                  * should never happen.
 889                                  */
 890                                 panic("not waking up swapper");
 891 #endif
 892                         }
 893                 thread_unlock(td);
 894         }
 895         p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
 896         p->p_flag |= P_INMEM;
 897 }
 898
 899 static int
 900 swapout(struct proc *p)
 901 {
 902         struct thread *td;
 903
 904         PROC_LOCK_ASSERT(p, MA_OWNED);
 905
 906         /*
 907          * The states of this process and its threads may have changed
 908          * by now.  Assuming that there is only one pageout daemon thread,
 909          * this process should still be in memory.
 910          */
 911         KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
 912             P_INMEM, ("swapout: lost a swapout race?"));
 913
 914         /*
 915          * Remember the resident count.
 916          */
 917         p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
 918
 919         /*
 920          * Check and mark all threads before we proceed.
 921          */
 922         p->p_flag &= ~P_INMEM;
 923         p->p_flag |= P_SWAPPINGOUT;
 924         FOREACH_THREAD_IN_PROC(p, td) {
 925                 thread_lock(td);
 926                 if (!thread_safetoswapout(td)) {
 927                         thread_unlock(td);
 928                         swapclear(p);
 929                         return (EBUSY);
 930                 }
 931                 td->td_flags &= ~TDF_INMEM;
 932                 TD_SET_SWAPPED(td);
 933                 thread_unlock(td);
 934         }
 935         td = FIRST_THREAD_IN_PROC(p);
 936         ++td->td_ru.ru_nswap;
 937         PROC_UNLOCK(p);
 938
 939         /*
 940          * This list is stable because all threads are now prevented from
 941          * running.  The list is only modified in the context of a running
 942          * thread in this process.
 943          */
 944         FOREACH_THREAD_IN_PROC(p, td)
 945                 vm_thread_swapout(td);
 946
 947         PROC_LOCK(p);
 948         p->p_flag &= ~P_SWAPPINGOUT;
 949         p->p_swtick = ticks;
 950         return (0);
 951 }