sys/vm/vm_swapout.c

   1 /*-
   2  * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
   3  *
   4  * Copyright (c) 1991 Regents of the University of California.
   5  * All rights reserved.
   6  * Copyright (c) 1994 John S. Dyson
   7  * All rights reserved.
   8  * Copyright (c) 1994 David Greenman
   9  * All rights reserved.
  10  * Copyright (c) 2005 Yahoo! Technologies Norway AS
  11  * All rights reserved.
  12  *
  13  * This code is derived from software contributed to Berkeley by
  14  * The Mach Operating System project at Carnegie-Mellon University.
  15  *
  16  * Redistribution and use in source and binary forms, with or without
  17  * modification, are permitted provided that the following conditions
  18  * are met:
  19  * 1. Redistributions of source code must retain the above copyright
  20  *    notice, this list of conditions and the following disclaimer.
  21  * 2. Redistributions in binary form must reproduce the above copyright
  22  *    notice, this list of conditions and the following disclaimer in the
  23  *    documentation and/or other materials provided with the distribution.
  24  * 3. All advertising materials mentioning features or use of this software
  25  *    must display the following acknowledgement:
  26  *      This product includes software developed by the University of
  27  *      California, Berkeley and its contributors.
  28  * 4. Neither the name of the University nor the names of its contributors
  29  *    may be used to endorse or promote products derived from this software
  30  *    without specific prior written permission.
  31  *
  32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  42  * SUCH DAMAGE.
  43  *
  44  *      from: @(#)vm_pageout.c  7.4 (Berkeley) 5/7/91
  45  *
  46  *
  47  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  48  * All rights reserved.
  49  *
  50  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  51  *
  52  * Permission to use, copy, modify and distribute this software and
  53  * its documentation is hereby granted, provided that both the copyright
  54  * notice and this permission notice appear in all copies of the
  55  * software, derivative works or modified versions, and any portions
  56  * thereof, and that both notices appear in supporting documentation.
  57  *
  58  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  59  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  60  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  61  *
  62  * Carnegie Mellon requests users of this software to return to
  63  *
  64  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  65  *  School of Computer Science
  66  *  Carnegie Mellon University
  67  *  Pittsburgh PA 15213-3890
  68  *
  69  * any improvements or extensions that they make and grant Carnegie the
  70  * rights to redistribute these changes.
  71  */
  72
  73 #include <sys/cdefs.h>
  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/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/refcount.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_kern.h>
 106 #include <vm/vm_object.h>
 107 #include <vm/vm_page.h>
 108 #include <vm/vm_map.h>
 109 #include <vm/vm_pageout.h>
 110 #include <vm/vm_pager.h>
 111 #include <vm/vm_phys.h>
 112 #include <vm/swap_pager.h>
 113 #include <vm/vm_extern.h>
 114 #include <vm/uma.h>
 115
 116 /* the kernel process "vm_daemon" */
 117 static void vm_daemon(void);
 118 static struct proc *vmproc;
 119
 120 static struct kproc_desc vm_kp = {
 121         "vmdaemon",
 122         vm_daemon,
 123         &vmproc
 124 };
 125 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
 126
 127 static int vm_swap_enabled = 1;
 128 static int vm_swap_idle_enabled = 0;
 129
 130 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW,
 131     &vm_swap_enabled, 0,
 132     "Enable entire process swapout");
 133 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW,
 134     &vm_swap_idle_enabled, 0,
 135     "Allow swapout on idle criteria");
 136
 137 /*
 138  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
 139  */
 140 static int swap_idle_threshold1 = 2;
 141 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
 142     &swap_idle_threshold1, 0,
 143     "Guaranteed swapped in time for a process");
 144
 145 /*
 146  * Swap_idle_threshold2 is the time that a process can be idle before
 147  * it will be swapped out, if idle swapping is enabled.
 148  */
 149 static int swap_idle_threshold2 = 10;
 150 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
 151     &swap_idle_threshold2, 0,
 152     "Time before a process will be swapped out");
 153
 154 static int vm_daemon_timeout = 0;
 155 SYSCTL_INT(_vm, OID_AUTO, vmdaemon_timeout, CTLFLAG_RW,
 156     &vm_daemon_timeout, 0,
 157     "Time between vmdaemon runs");
 158
 159 static int vm_pageout_req_swapout;      /* XXX */
 160 static int vm_daemon_needed;
 161 static struct mtx vm_daemon_mtx;
 162 /* Allow for use by vm_pageout before vm_daemon is initialized. */
 163 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
 164
 165 static int swapped_cnt;
 166 static int swap_inprogress;     /* Pending swap-ins done outside swapper. */
 167 static int last_swapin;
 168
 169 static void swapclear(struct proc *);
 170 static int swapout(struct proc *);
 171 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
 172 static void vm_swapout_object_deactivate(pmap_t, vm_object_t, long);
 173 static void swapout_procs(int action);
 174 static void vm_req_vmdaemon(int req);
 175 static void vm_thread_swapout(struct thread *td);
 176
 177 static void
 178 vm_swapout_object_deactivate_page(pmap_t pmap, vm_page_t m, bool unmap)
 179 {
 180
 181         /*
 182          * Ignore unreclaimable wired pages.  Repeat the check after busying
 183          * since a busy holder may wire the page.
 184          */
 185         if (vm_page_wired(m) || !vm_page_tryxbusy(m))
 186                 return;
 187
 188         if (vm_page_wired(m) || !pmap_page_exists_quick(pmap, m)) {
 189                 vm_page_xunbusy(m);
 190                 return;
 191         }
 192         if (!pmap_is_referenced(m)) {
 193                 if (!vm_page_active(m))
 194                         (void)vm_page_try_remove_all(m);
 195                 else if (unmap && vm_page_try_remove_all(m))
 196                         vm_page_deactivate(m);
 197         }
 198         vm_page_xunbusy(m);
 199 }
 200
 201 /*
 202  *      vm_swapout_object_deactivate
 203  *
 204  *      Deactivate enough pages to satisfy the inactive target
 205  *      requirements.
 206  *
 207  *      The object and map must be locked.
 208  */
 209 static void
 210 vm_swapout_object_deactivate(pmap_t pmap, vm_object_t first_object,
 211     long desired)
 212 {
 213         vm_object_t backing_object, object;
 214         vm_page_t m;
 215         bool unmap;
 216
 217         VM_OBJECT_ASSERT_LOCKED(first_object);
 218         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 219                 return;
 220         for (object = first_object;; object = backing_object) {
 221                 if (pmap_resident_count(pmap) <= desired)
 222                         goto unlock_return;
 223                 VM_OBJECT_ASSERT_LOCKED(object);
 224                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 225                     blockcount_read(&object->paging_in_progress) > 0)
 226                         goto unlock_return;
 227
 228                 unmap = true;
 229                 if (object->shadow_count > 1)
 230                         unmap = false;
 231
 232                 /*
 233                  * Scan the object's entire memory queue.
 234                  */
 235                 TAILQ_FOREACH(m, &object->memq, listq) {
 236                         if (pmap_resident_count(pmap) <= desired)
 237                                 goto unlock_return;
 238                         if (should_yield())
 239                                 goto unlock_return;
 240                         vm_swapout_object_deactivate_page(pmap, m, unmap);
 241                 }
 242                 if ((backing_object = object->backing_object) == NULL)
 243                         goto unlock_return;
 244                 VM_OBJECT_RLOCK(backing_object);
 245                 if (object != first_object)
 246                         VM_OBJECT_RUNLOCK(object);
 247         }
 248 unlock_return:
 249         if (object != first_object)
 250                 VM_OBJECT_RUNLOCK(object);
 251 }
 252
 253 /*
 254  * deactivate some number of pages in a map, try to do it fairly, but
 255  * that is really hard to do.
 256  */
 257 static void
 258 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
 259 {
 260         vm_map_entry_t tmpe;
 261         vm_object_t obj, bigobj;
 262         int nothingwired;
 263
 264         if (!vm_map_trylock_read(map))
 265                 return;
 266
 267         bigobj = NULL;
 268         nothingwired = TRUE;
 269
 270         /*
 271          * first, search out the biggest object, and try to free pages from
 272          * that.
 273          */
 274         VM_MAP_ENTRY_FOREACH(tmpe, map) {
 275                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 276                         obj = tmpe->object.vm_object;
 277                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 278                                 if (obj->shadow_count <= 1 &&
 279                                     (bigobj == NULL ||
 280                                      bigobj->resident_page_count <
 281                                      obj->resident_page_count)) {
 282                                         if (bigobj != NULL)
 283                                                 VM_OBJECT_RUNLOCK(bigobj);
 284                                         bigobj = obj;
 285                                 } else
 286                                         VM_OBJECT_RUNLOCK(obj);
 287                         }
 288                 }
 289                 if (tmpe->wired_count > 0)
 290                         nothingwired = FALSE;
 291         }
 292
 293         if (bigobj != NULL) {
 294                 vm_swapout_object_deactivate(map->pmap, bigobj, desired);
 295                 VM_OBJECT_RUNLOCK(bigobj);
 296         }
 297         /*
 298          * Next, hunt around for other pages to deactivate.  We actually
 299          * do this search sort of wrong -- .text first is not the best idea.
 300          */
 301         VM_MAP_ENTRY_FOREACH(tmpe, map) {
 302                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 303                         break;
 304                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 305                         obj = tmpe->object.vm_object;
 306                         if (obj != NULL) {
 307                                 VM_OBJECT_RLOCK(obj);
 308                                 vm_swapout_object_deactivate(map->pmap, obj,
 309                                     desired);
 310                                 VM_OBJECT_RUNLOCK(obj);
 311                         }
 312                 }
 313         }
 314
 315         /*
 316          * Remove all mappings if a process is swapped out, this will free page
 317          * table pages.
 318          */
 319         if (desired == 0 && nothingwired) {
 320                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 321                     vm_map_max(map));
 322         }
 323
 324         vm_map_unlock_read(map);
 325 }
 326
 327 /*
 328  * Swap out requests
 329  */
 330 #define VM_SWAP_NORMAL 1
 331 #define VM_SWAP_IDLE 2
 332
 333 void
 334 vm_swapout_run(void)
 335 {
 336
 337         if (vm_swap_enabled)
 338                 vm_req_vmdaemon(VM_SWAP_NORMAL);
 339 }
 340
 341 /*
 342  * Idle process swapout -- run once per second when pagedaemons are
 343  * reclaiming pages.
 344  */
 345 void
 346 vm_swapout_run_idle(void)
 347 {
 348         static long lsec;
 349
 350         if (!vm_swap_idle_enabled || time_second == lsec)
 351                 return;
 352         vm_req_vmdaemon(VM_SWAP_IDLE);
 353         lsec = time_second;
 354 }
 355
 356 static void
 357 vm_req_vmdaemon(int req)
 358 {
 359         static int lastrun = 0;
 360
 361         mtx_lock(&vm_daemon_mtx);
 362         vm_pageout_req_swapout |= req;
 363         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
 364                 wakeup(&vm_daemon_needed);
 365                 lastrun = ticks;
 366         }
 367         mtx_unlock(&vm_daemon_mtx);
 368 }
 369
 370 static void
 371 vm_daemon(void)
 372 {
 373         struct rlimit rsslim;
 374         struct proc *p;
 375         struct thread *td;
 376         struct vmspace *vm;
 377         int breakout, swapout_flags, tryagain, attempts;
 378 #ifdef RACCT
 379         uint64_t rsize, ravailable;
 380
 381         if (racct_enable && vm_daemon_timeout == 0)
 382                 vm_daemon_timeout = hz;
 383 #endif
 384
 385         while (TRUE) {
 386                 mtx_lock(&vm_daemon_mtx);
 387                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
 388                     vm_daemon_timeout);
 389                 swapout_flags = vm_pageout_req_swapout;
 390                 vm_pageout_req_swapout = 0;
 391                 mtx_unlock(&vm_daemon_mtx);
 392                 if (swapout_flags != 0) {
 393                         /*
 394                          * Drain the per-CPU page queue batches as a deadlock
 395                          * avoidance measure.
 396                          */
 397                         if ((swapout_flags & VM_SWAP_NORMAL) != 0)
 398                                 vm_page_pqbatch_drain();
 399                         swapout_procs(swapout_flags);
 400                 }
 401
 402                 /*
 403                  * scan the processes for exceeding their rlimits or if
 404                  * process is swapped out -- deactivate pages
 405                  */
 406                 tryagain = 0;
 407                 attempts = 0;
 408 again:
 409                 attempts++;
 410                 sx_slock(&allproc_lock);
 411                 FOREACH_PROC_IN_SYSTEM(p) {
 412                         vm_pindex_t limit, size;
 413
 414                         /*
 415                          * if this is a system process or if we have already
 416                          * looked at this process, skip it.
 417                          */
 418                         PROC_LOCK(p);
 419                         if (p->p_state != PRS_NORMAL ||
 420                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
 421                                 PROC_UNLOCK(p);
 422                                 continue;
 423                         }
 424                         /*
 425                          * if the process is in a non-running type state,
 426                          * don't touch it.
 427                          */
 428                         breakout = 0;
 429                         FOREACH_THREAD_IN_PROC(p, td) {
 430                                 thread_lock(td);
 431                                 if (!TD_ON_RUNQ(td) &&
 432                                     !TD_IS_RUNNING(td) &&
 433                                     !TD_IS_SLEEPING(td) &&
 434                                     !TD_IS_SUSPENDED(td)) {
 435                                         thread_unlock(td);
 436                                         breakout = 1;
 437                                         break;
 438                                 }
 439                                 thread_unlock(td);
 440                         }
 441                         if (breakout) {
 442                                 PROC_UNLOCK(p);
 443                                 continue;
 444                         }
 445                         /*
 446                          * get a limit
 447                          */
 448                         lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
 449                         limit = OFF_TO_IDX(
 450                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
 451
 452                         /*
 453                          * let processes that are swapped out really be
 454                          * swapped out set the limit to nothing (will force a
 455                          * swap-out.)
 456                          */
 457                         if ((p->p_flag & P_INMEM) == 0)
 458                                 limit = 0;      /* XXX */
 459                         vm = vmspace_acquire_ref(p);
 460                         _PHOLD_LITE(p);
 461                         PROC_UNLOCK(p);
 462                         if (vm == NULL) {
 463                                 PRELE(p);
 464                                 continue;
 465                         }
 466                         sx_sunlock(&allproc_lock);
 467
 468                         size = vmspace_resident_count(vm);
 469                         if (size >= limit) {
 470                                 vm_swapout_map_deactivate_pages(
 471                                     &vm->vm_map, limit);
 472                                 size = vmspace_resident_count(vm);
 473                         }
 474 #ifdef RACCT
 475                         if (racct_enable) {
 476                                 rsize = IDX_TO_OFF(size);
 477                                 PROC_LOCK(p);
 478                                 if (p->p_state == PRS_NORMAL)
 479                                         racct_set(p, RACCT_RSS, rsize);
 480                                 ravailable = racct_get_available(p, RACCT_RSS);
 481                                 PROC_UNLOCK(p);
 482                                 if (rsize > ravailable) {
 483                                         /*
 484                                          * Don't be overly aggressive; this
 485                                          * might be an innocent process,
 486                                          * and the limit could've been exceeded
 487                                          * by some memory hog.  Don't try
 488                                          * to deactivate more than 1/4th
 489                                          * of process' resident set size.
 490                                          */
 491                                         if (attempts <= 8) {
 492                                                 if (ravailable < rsize -
 493                                                     (rsize / 4)) {
 494                                                         ravailable = rsize -
 495                                                             (rsize / 4);
 496                                                 }
 497                                         }
 498                                         vm_swapout_map_deactivate_pages(
 499                                             &vm->vm_map,
 500                                             OFF_TO_IDX(ravailable));
 501                                         /* Update RSS usage after paging out. */
 502                                         size = vmspace_resident_count(vm);
 503                                         rsize = IDX_TO_OFF(size);
 504                                         PROC_LOCK(p);
 505                                         if (p->p_state == PRS_NORMAL)
 506                                                 racct_set(p, RACCT_RSS, rsize);
 507                                         PROC_UNLOCK(p);
 508                                         if (rsize > ravailable)
 509                                                 tryagain = 1;
 510                                 }
 511                         }
 512 #endif
 513                         vmspace_free(vm);
 514                         sx_slock(&allproc_lock);
 515                         PRELE(p);
 516                 }
 517                 sx_sunlock(&allproc_lock);
 518                 if (tryagain != 0 && attempts <= 10) {
 519                         maybe_yield();
 520                         goto again;
 521                 }
 522         }
 523 }
 524
 525 /*
 526  * Allow a thread's kernel stack to be paged out.
 527  */
 528 static void
 529 vm_thread_swapout(struct thread *td)
 530 {
 531         vm_page_t m;
 532         vm_offset_t kaddr;
 533         vm_pindex_t pindex;
 534         int i, pages;
 535
 536         cpu_thread_swapout(td);
 537         kaddr = td->td_kstack;
 538         pages = td->td_kstack_pages;
 539         pindex = atop(kaddr - VM_MIN_KERNEL_ADDRESS);
 540         pmap_qremove(kaddr, pages);
 541         VM_OBJECT_WLOCK(kstack_object);
 542         for (i = 0; i < pages; i++) {
 543                 m = vm_page_lookup(kstack_object, pindex + i);
 544                 if (m == NULL)
 545                         panic("vm_thread_swapout: kstack already missing?");
 546                 vm_page_dirty(m);
 547                 vm_page_xunbusy_unchecked(m);
 548                 vm_page_unwire(m, PQ_LAUNDRY);
 549         }
 550         VM_OBJECT_WUNLOCK(kstack_object);
 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, int oom_alloc)
 558 {
 559         vm_page_t ma[KSTACK_MAX_PAGES];
 560         vm_offset_t kaddr;
 561         int a, count, i, j, pages, rv __diagused;
 562
 563         kaddr = td->td_kstack;
 564         pages = td->td_kstack_pages;
 565         vm_thread_stack_back(td->td_domain.dr_policy, kaddr, ma, pages,
 566             oom_alloc);
 567         for (i = 0; i < pages;) {
 568                 vm_page_assert_xbusied(ma[i]);
 569                 if (vm_page_all_valid(ma[i])) {
 570                         i++;
 571                         continue;
 572                 }
 573                 vm_object_pip_add(kstack_object, 1);
 574                 for (j = i + 1; j < pages; j++)
 575                         if (vm_page_all_valid(ma[j]))
 576                                 break;
 577                 VM_OBJECT_WLOCK(kstack_object);
 578                 rv = vm_pager_has_page(kstack_object, ma[i]->pindex, NULL, &a);
 579                 VM_OBJECT_WUNLOCK(kstack_object);
 580                 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
 581                 count = min(a + 1, j - i);
 582                 rv = vm_pager_get_pages(kstack_object, 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(kstack_object);
 586                 i += count;
 587         }
 588         pmap_qenter(kaddr, ma, pages);
 589         cpu_thread_swapin(td);
 590 }
 591
 592 void
 593 faultin(struct proc *p)
 594 {
 595         struct thread *td;
 596         int oom_alloc;
 597
 598         PROC_LOCK_ASSERT(p, MA_OWNED);
 599
 600         /*
 601          * If another process is swapping in this process,
 602          * just wait until it finishes.
 603          */
 604         if (p->p_flag & P_SWAPPINGIN) {
 605                 while (p->p_flag & P_SWAPPINGIN)
 606                         msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
 607                 return;
 608         }
 609
 610         if ((p->p_flag & P_INMEM) == 0) {
 611                 oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
 612                     VM_ALLOC_NORMAL;
 613
 614                 /*
 615                  * Don't let another thread swap process p out while we are
 616                  * busy swapping it in.
 617                  */
 618                 ++p->p_lock;
 619                 p->p_flag |= P_SWAPPINGIN;
 620                 PROC_UNLOCK(p);
 621                 sx_xlock(&allproc_lock);
 622                 MPASS(swapped_cnt > 0);
 623                 swapped_cnt--;
 624                 if (curthread != &thread0)
 625                         swap_inprogress++;
 626                 sx_xunlock(&allproc_lock);
 627
 628                 /*
 629                  * We hold no lock here because the list of threads
 630                  * can not change while all threads in the process are
 631                  * swapped out.
 632                  */
 633                 FOREACH_THREAD_IN_PROC(p, td)
 634                         vm_thread_swapin(td, oom_alloc);
 635
 636                 if (curthread != &thread0) {
 637                         sx_xlock(&allproc_lock);
 638                         MPASS(swap_inprogress > 0);
 639                         swap_inprogress--;
 640                         last_swapin = ticks;
 641                         sx_xunlock(&allproc_lock);
 642                 }
 643                 PROC_LOCK(p);
 644                 swapclear(p);
 645                 p->p_swtick = ticks;
 646
 647                 /* Allow other threads to swap p out now. */
 648                 wakeup(&p->p_flag);
 649                 --p->p_lock;
 650         }
 651 }
 652
 653 /*
 654  * This swapin algorithm attempts to swap-in processes only if there
 655  * is enough space for them.  Of course, if a process waits for a long
 656  * time, it will be swapped in anyway.
 657  */
 658
 659 static struct proc *
 660 swapper_selector(bool wkilled_only)
 661 {
 662         struct proc *p, *res;
 663         struct thread *td;
 664         int ppri, pri, slptime, swtime;
 665
 666         sx_assert(&allproc_lock, SA_SLOCKED);
 667         if (swapped_cnt == 0)
 668                 return (NULL);
 669         res = NULL;
 670         ppri = INT_MIN;
 671         FOREACH_PROC_IN_SYSTEM(p) {
 672                 PROC_LOCK(p);
 673                 if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
 674                     P_SWAPPINGIN | P_INMEM)) != 0) {
 675                         PROC_UNLOCK(p);
 676                         continue;
 677                 }
 678                 if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
 679                         /*
 680                          * A swapped-out process might have mapped a
 681                          * large portion of the system's pages as
 682                          * anonymous memory.  There is no other way to
 683                          * release the memory other than to kill the
 684                          * process, for which we need to swap it in.
 685                          */
 686                         return (p);
 687                 }
 688                 if (wkilled_only) {
 689                         PROC_UNLOCK(p);
 690                         continue;
 691                 }
 692                 swtime = (ticks - p->p_swtick) / hz;
 693                 FOREACH_THREAD_IN_PROC(p, td) {
 694                         /*
 695                          * An otherwise runnable thread of a process
 696                          * swapped out has only the TDI_SWAPPED bit set.
 697                          */
 698                         thread_lock(td);
 699                         if (td->td_inhibitors == TDI_SWAPPED) {
 700                                 slptime = (ticks - td->td_slptick) / hz;
 701                                 pri = swtime + slptime;
 702                                 if ((td->td_flags & TDF_SWAPINREQ) == 0)
 703                                         pri -= p->p_nice * 8;
 704                                 /*
 705                                  * if this thread is higher priority
 706                                  * and there is enough space, then select
 707                                  * this process instead of the previous
 708                                  * selection.
 709                                  */
 710                                 if (pri > ppri) {
 711                                         res = p;
 712                                         ppri = pri;
 713                                 }
 714                         }
 715                         thread_unlock(td);
 716                 }
 717                 PROC_UNLOCK(p);
 718         }
 719
 720         if (res != NULL)
 721                 PROC_LOCK(res);
 722         return (res);
 723 }
 724
 725 #define SWAPIN_INTERVAL (MAXSLP * hz / 2)
 726
 727 /*
 728  * Limit swapper to swap in one non-WKILLED process in MAXSLP/2
 729  * interval, assuming that there is:
 730  * - at least one domain that is not suffering from a shortage of free memory;
 731  * - no parallel swap-ins;
 732  * - no other swap-ins in the current SWAPIN_INTERVAL.
 733  */
 734 static bool
 735 swapper_wkilled_only(void)
 736 {
 737
 738         return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 ||
 739             (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL);
 740 }
 741
 742 void
 743 swapper(void)
 744 {
 745         struct proc *p;
 746
 747         for (;;) {
 748                 sx_slock(&allproc_lock);
 749                 p = swapper_selector(swapper_wkilled_only());
 750                 sx_sunlock(&allproc_lock);
 751
 752                 if (p == NULL) {
 753                         tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL);
 754                 } else {
 755                         PROC_LOCK_ASSERT(p, MA_OWNED);
 756
 757                         /*
 758                          * Another process may be bringing or may have
 759                          * already brought this process in while we
 760                          * traverse all threads.  Or, this process may
 761                          * have exited or even being swapped out
 762                          * again.
 763                          */
 764                         if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
 765                             P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
 766                                 faultin(p);
 767                         }
 768                         PROC_UNLOCK(p);
 769                 }
 770         }
 771 }
 772
 773 /*
 774  * First, if any processes have been sleeping or stopped for at least
 775  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
 776  * no such processes exist, then the longest-sleeping or stopped
 777  * process is swapped out.  Finally, and only as a last resort, if
 778  * there are no sleeping or stopped processes, the longest-resident
 779  * process is swapped out.
 780  */
 781 static void
 782 swapout_procs(int action)
 783 {
 784         struct proc *p;
 785         struct thread *td;
 786         int slptime;
 787         bool didswap, doswap;
 788
 789         MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
 790
 791         didswap = false;
 792         sx_slock(&allproc_lock);
 793         FOREACH_PROC_IN_SYSTEM(p) {
 794                 /*
 795                  * Filter out not yet fully constructed processes.  Do
 796                  * not swap out held processes.  Avoid processes which
 797                  * are system, exiting, execing, traced, already swapped
 798                  * out or are in the process of being swapped in or out.
 799                  */
 800                 PROC_LOCK(p);
 801                 if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
 802                     (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
 803                     P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
 804                     P_INMEM) {
 805                         PROC_UNLOCK(p);
 806                         continue;
 807                 }
 808
 809                 /*
 810                  * Further consideration of this process for swap out
 811                  * requires iterating over its threads.  We release
 812                  * allproc_lock here so that process creation and
 813                  * destruction are not blocked while we iterate.
 814                  *
 815                  * To later reacquire allproc_lock and resume
 816                  * iteration over the allproc list, we will first have
 817                  * to release the lock on the process.  We place a
 818                  * hold on the process so that it remains in the
 819                  * allproc list while it is unlocked.
 820                  */
 821                 _PHOLD_LITE(p);
 822                 sx_sunlock(&allproc_lock);
 823
 824                 /*
 825                  * Do not swapout a realtime process.
 826                  * Guarantee swap_idle_threshold1 time in memory.
 827                  * If the system is under memory stress, or if we are
 828                  * swapping idle processes >= swap_idle_threshold2,
 829                  * then swap the process out.
 830                  */
 831                 doswap = true;
 832                 FOREACH_THREAD_IN_PROC(p, td) {
 833                         thread_lock(td);
 834                         slptime = (ticks - td->td_slptick) / hz;
 835                         if (PRI_IS_REALTIME(td->td_pri_class) ||
 836                             slptime < swap_idle_threshold1 ||
 837                             !thread_safetoswapout(td) ||
 838                             ((action & VM_SWAP_NORMAL) == 0 &&
 839                             slptime < swap_idle_threshold2))
 840                                 doswap = false;
 841                         thread_unlock(td);
 842                         if (!doswap)
 843                                 break;
 844                 }
 845                 if (doswap && swapout(p) == 0)
 846                         didswap = true;
 847
 848                 PROC_UNLOCK(p);
 849                 if (didswap) {
 850                         sx_xlock(&allproc_lock);
 851                         swapped_cnt++;
 852                         sx_downgrade(&allproc_lock);
 853                 } else
 854                         sx_slock(&allproc_lock);
 855                 PRELE(p);
 856         }
 857         sx_sunlock(&allproc_lock);
 858
 859         /*
 860          * If we swapped something out, and another process needed memory,
 861          * then wakeup the sched process.
 862          */
 863         if (didswap)
 864                 wakeup(&proc0);
 865 }
 866
 867 static void
 868 swapclear(struct proc *p)
 869 {
 870         struct thread *td;
 871
 872         PROC_LOCK_ASSERT(p, MA_OWNED);
 873
 874         FOREACH_THREAD_IN_PROC(p, td) {
 875                 thread_lock(td);
 876                 td->td_flags |= TDF_INMEM;
 877                 td->td_flags &= ~TDF_SWAPINREQ;
 878                 TD_CLR_SWAPPED(td);
 879                 if (TD_CAN_RUN(td)) {
 880                         if (setrunnable(td, 0)) {
 881 #ifdef INVARIANTS
 882                                 /*
 883                                  * XXX: We just cleared TDI_SWAPPED
 884                                  * above and set TDF_INMEM, so this
 885                                  * should never happen.
 886                                  */
 887                                 panic("not waking up swapper");
 888 #endif
 889                         }
 890                 } else
 891                         thread_unlock(td);
 892         }
 893         p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
 894         p->p_flag |= P_INMEM;
 895 }
 896
 897 static int
 898 swapout(struct proc *p)
 899 {
 900         struct thread *td;
 901
 902         PROC_LOCK_ASSERT(p, MA_OWNED);
 903
 904         /*
 905          * The states of this process and its threads may have changed
 906          * by now.  Assuming that there is only one pageout daemon thread,
 907          * this process should still be in memory.
 908          */
 909         KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
 910             P_INMEM, ("swapout: lost a swapout race?"));
 911
 912         /*
 913          * Remember the resident count.
 914          */
 915         p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
 916
 917         /*
 918          * Check and mark all threads before we proceed.
 919          */
 920         p->p_flag &= ~P_INMEM;
 921         p->p_flag |= P_SWAPPINGOUT;
 922         FOREACH_THREAD_IN_PROC(p, td) {
 923                 thread_lock(td);
 924                 if (!thread_safetoswapout(td)) {
 925                         thread_unlock(td);
 926                         swapclear(p);
 927                         return (EBUSY);
 928                 }
 929                 td->td_flags &= ~TDF_INMEM;
 930                 TD_SET_SWAPPED(td);
 931                 thread_unlock(td);
 932         }
 933         td = FIRST_THREAD_IN_PROC(p);
 934         ++td->td_ru.ru_nswap;
 935         PROC_UNLOCK(p);
 936
 937         /*
 938          * This list is stable because all threads are now prevented from
 939          * running.  The list is only modified in the context of a running
 940          * thread in this process.
 941          */
 942         FOREACH_THREAD_IN_PROC(p, td)
 943                 vm_thread_swapout(td);
 944
 945         PROC_LOCK(p);
 946         p->p_flag &= ~P_SWAPPINGOUT;
 947         p->p_swtick = ticks;
 948         return (0);
 949 }