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 __FBSDID("$FreeBSD$");
  75
  76 #include "opt_kstack_pages.h"
  77 #include "opt_kstack_max_pages.h"
  78 #include "opt_vm.h"
  79
  80 #include <sys/param.h>
  81 #include <sys/systm.h>
  82 #include <sys/limits.h>
  83 #include <sys/kernel.h>
  84 #include <sys/eventhandler.h>
  85 #include <sys/lock.h>
  86 #include <sys/mutex.h>
  87 #include <sys/proc.h>
  88 #include <sys/_kstack_cache.h>
  89 #include <sys/kthread.h>
  90 #include <sys/ktr.h>
  91 #include <sys/mount.h>
  92 #include <sys/racct.h>
  93 #include <sys/resourcevar.h>
  94 #include <sys/sched.h>
  95 #include <sys/sdt.h>
  96 #include <sys/signalvar.h>
  97 #include <sys/smp.h>
  98 #include <sys/time.h>
  99 #include <sys/vnode.h>
 100 #include <sys/vmmeter.h>
 101 #include <sys/rwlock.h>
 102 #include <sys/sx.h>
 103 #include <sys/sysctl.h>
 104
 105 #include <vm/vm.h>
 106 #include <vm/vm_param.h>
 107 #include <vm/vm_object.h>
 108 #include <vm/vm_page.h>
 109 #include <vm/vm_map.h>
 110 #include <vm/vm_pageout.h>
 111 #include <vm/vm_pager.h>
 112 #include <vm/vm_phys.h>
 113 #include <vm/swap_pager.h>
 114 #include <vm/vm_extern.h>
 115 #include <vm/uma.h>
 116
 117 /* the kernel process "vm_daemon" */
 118 static void vm_daemon(void);
 119 static struct proc *vmproc;
 120
 121 static struct kproc_desc vm_kp = {
 122         "vmdaemon",
 123         vm_daemon,
 124         &vmproc
 125 };
 126 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
 127
 128 static int vm_swap_enabled = 1;
 129 static int vm_swap_idle_enabled = 0;
 130
 131 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW,
 132     &vm_swap_enabled, 0,
 133     "Enable entire process swapout");
 134 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW,
 135     &vm_swap_idle_enabled, 0,
 136     "Allow swapout on idle criteria");
 137
 138 /*
 139  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
 140  */
 141 static int swap_idle_threshold1 = 2;
 142 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
 143     &swap_idle_threshold1, 0,
 144     "Guaranteed swapped in time for a process");
 145
 146 /*
 147  * Swap_idle_threshold2 is the time that a process can be idle before
 148  * it will be swapped out, if idle swapping is enabled.
 149  */
 150 static int swap_idle_threshold2 = 10;
 151 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
 152     &swap_idle_threshold2, 0,
 153     "Time before a process will be swapped out");
 154
 155 static int vm_pageout_req_swapout;      /* XXX */
 156 static int vm_daemon_needed;
 157 static struct mtx vm_daemon_mtx;
 158 /* Allow for use by vm_pageout before vm_daemon is initialized. */
 159 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
 160
 161 static int swapped_cnt;
 162
 163 static void swapclear(struct proc *);
 164 static int swapout(struct proc *);
 165 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
 166 static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long);
 167 static void swapout_procs(int action);
 168 static void vm_req_vmdaemon(int req);
 169 static void vm_thread_swapout(struct thread *td);
 170
 171 /*
 172  *      vm_swapout_object_deactivate_pages
 173  *
 174  *      Deactivate enough pages to satisfy the inactive target
 175  *      requirements.
 176  *
 177  *      The object and map must be locked.
 178  */
 179 static void
 180 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
 181     long desired)
 182 {
 183         vm_object_t backing_object, object;
 184         vm_page_t p;
 185         int act_delta, remove_mode;
 186
 187         VM_OBJECT_ASSERT_LOCKED(first_object);
 188         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 189                 return;
 190         for (object = first_object;; object = backing_object) {
 191                 if (pmap_resident_count(pmap) <= desired)
 192                         goto unlock_return;
 193                 VM_OBJECT_ASSERT_LOCKED(object);
 194                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 195                     object->paging_in_progress != 0)
 196                         goto unlock_return;
 197
 198                 remove_mode = 0;
 199                 if (object->shadow_count > 1)
 200                         remove_mode = 1;
 201                 /*
 202                  * Scan the object's entire memory queue.
 203                  */
 204                 TAILQ_FOREACH(p, &object->memq, listq) {
 205                         if (pmap_resident_count(pmap) <= desired)
 206                                 goto unlock_return;
 207                         if (should_yield())
 208                                 goto unlock_return;
 209                         if (vm_page_busied(p))
 210                                 continue;
 211                         VM_CNT_INC(v_pdpages);
 212                         vm_page_lock(p);
 213                         if (vm_page_held(p) ||
 214                             !pmap_page_exists_quick(pmap, p)) {
 215                                 vm_page_unlock(p);
 216                                 continue;
 217                         }
 218                         act_delta = pmap_ts_referenced(p);
 219                         if ((p->aflags & PGA_REFERENCED) != 0) {
 220                                 if (act_delta == 0)
 221                                         act_delta = 1;
 222                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 223                         }
 224                         if (!vm_page_active(p) && act_delta != 0) {
 225                                 vm_page_activate(p);
 226                                 p->act_count += act_delta;
 227                         } else if (vm_page_active(p)) {
 228                                 if (act_delta == 0) {
 229                                         p->act_count -= min(p->act_count,
 230                                             ACT_DECLINE);
 231                                         if (!remove_mode && p->act_count == 0) {
 232                                                 pmap_remove_all(p);
 233                                                 vm_page_deactivate(p);
 234                                         } else
 235                                                 vm_page_requeue(p);
 236                                 } else {
 237                                         vm_page_activate(p);
 238                                         if (p->act_count < ACT_MAX -
 239                                             ACT_ADVANCE)
 240                                                 p->act_count += ACT_ADVANCE;
 241                                         vm_page_requeue(p);
 242                                 }
 243                         } else if (vm_page_inactive(p))
 244                                 pmap_remove_all(p);
 245                         vm_page_unlock(p);
 246                 }
 247                 if ((backing_object = object->backing_object) == NULL)
 248                         goto unlock_return;
 249                 VM_OBJECT_RLOCK(backing_object);
 250                 if (object != first_object)
 251                         VM_OBJECT_RUNLOCK(object);
 252         }
 253 unlock_return:
 254         if (object != first_object)
 255                 VM_OBJECT_RUNLOCK(object);
 256 }
 257
 258 /*
 259  * deactivate some number of pages in a map, try to do it fairly, but
 260  * that is really hard to do.
 261  */
 262 static void
 263 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
 264 {
 265         vm_map_entry_t tmpe;
 266         vm_object_t obj, bigobj;
 267         int nothingwired;
 268
 269         if (!vm_map_trylock_read(map))
 270                 return;
 271
 272         bigobj = NULL;
 273         nothingwired = TRUE;
 274
 275         /*
 276          * first, search out the biggest object, and try to free pages from
 277          * that.
 278          */
 279         tmpe = map->header.next;
 280         while (tmpe != &map->header) {
 281                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 282                         obj = tmpe->object.vm_object;
 283                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 284                                 if (obj->shadow_count <= 1 &&
 285                                     (bigobj == NULL ||
 286                                      bigobj->resident_page_count <
 287                                      obj->resident_page_count)) {
 288                                         if (bigobj != NULL)
 289                                                 VM_OBJECT_RUNLOCK(bigobj);
 290                                         bigobj = obj;
 291                                 } else
 292                                         VM_OBJECT_RUNLOCK(obj);
 293                         }
 294                 }
 295                 if (tmpe->wired_count > 0)
 296                         nothingwired = FALSE;
 297                 tmpe = tmpe->next;
 298         }
 299
 300         if (bigobj != NULL) {
 301                 vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
 302                 VM_OBJECT_RUNLOCK(bigobj);
 303         }
 304         /*
 305          * Next, hunt around for other pages to deactivate.  We actually
 306          * do this search sort of wrong -- .text first is not the best idea.
 307          */
 308         tmpe = map->header.next;
 309         while (tmpe != &map->header) {
 310                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 311                         break;
 312                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 313                         obj = tmpe->object.vm_object;
 314                         if (obj != NULL) {
 315                                 VM_OBJECT_RLOCK(obj);
 316                                 vm_swapout_object_deactivate_pages(map->pmap,
 317                                     obj, desired);
 318                                 VM_OBJECT_RUNLOCK(obj);
 319                         }
 320                 }
 321                 tmpe = tmpe->next;
 322         }
 323
 324         /*
 325          * Remove all mappings if a process is swapped out, this will free page
 326          * table pages.
 327          */
 328         if (desired == 0 && nothingwired) {
 329                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 330                     vm_map_max(map));
 331         }
 332
 333         vm_map_unlock_read(map);
 334 }
 335
 336 /*
 337  * Swap out requests
 338  */
 339 #define VM_SWAP_NORMAL 1
 340 #define VM_SWAP_IDLE 2
 341
 342 void
 343 vm_swapout_run(void)
 344 {
 345
 346         if (vm_swap_enabled)
 347                 vm_req_vmdaemon(VM_SWAP_NORMAL);
 348 }
 349
 350 /*
 351  * Idle process swapout -- run once per second when pagedaemons are
 352  * reclaiming pages.
 353  */
 354 void
 355 vm_swapout_run_idle(void)
 356 {
 357         static long lsec;
 358
 359         if (!vm_swap_idle_enabled || time_second == lsec)
 360                 return;
 361         vm_req_vmdaemon(VM_SWAP_IDLE);
 362         lsec = time_second;
 363 }
 364
 365 static void
 366 vm_req_vmdaemon(int req)
 367 {
 368         static int lastrun = 0;
 369
 370         mtx_lock(&vm_daemon_mtx);
 371         vm_pageout_req_swapout |= req;
 372         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
 373                 wakeup(&vm_daemon_needed);
 374                 lastrun = ticks;
 375         }
 376         mtx_unlock(&vm_daemon_mtx);
 377 }
 378
 379 static void
 380 vm_daemon(void)
 381 {
 382         struct rlimit rsslim;
 383         struct proc *p;
 384         struct thread *td;
 385         struct vmspace *vm;
 386         int breakout, swapout_flags, tryagain, attempts;
 387 #ifdef RACCT
 388         uint64_t rsize, ravailable;
 389 #endif
 390
 391         while (TRUE) {
 392                 mtx_lock(&vm_daemon_mtx);
 393                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
 394 #ifdef RACCT
 395                     racct_enable ? hz : 0
 396 #else
 397                     0
 398 #endif
 399                 );
 400                 swapout_flags = vm_pageout_req_swapout;
 401                 vm_pageout_req_swapout = 0;
 402                 mtx_unlock(&vm_daemon_mtx);
 403                 if (swapout_flags != 0) {
 404                         /*
 405                          * Drain the per-CPU page queue batches as a deadlock
 406                          * avoidance measure.
 407                          */
 408                         if ((swapout_flags & VM_SWAP_NORMAL) != 0)
 409                                 vm_page_drain_pqbatch();
 410                         swapout_procs(swapout_flags);
 411                 }
 412
 413                 /*
 414                  * scan the processes for exceeding their rlimits or if
 415                  * process is swapped out -- deactivate pages
 416                  */
 417                 tryagain = 0;
 418                 attempts = 0;
 419 again:
 420                 attempts++;
 421                 sx_slock(&allproc_lock);
 422                 FOREACH_PROC_IN_SYSTEM(p) {
 423                         vm_pindex_t limit, size;
 424
 425                         /*
 426                          * if this is a system process or if we have already
 427                          * looked at this process, skip it.
 428                          */
 429                         PROC_LOCK(p);
 430                         if (p->p_state != PRS_NORMAL ||
 431                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
 432                                 PROC_UNLOCK(p);
 433                                 continue;
 434                         }
 435                         /*
 436                          * if the process is in a non-running type state,
 437                          * don't touch it.
 438                          */
 439                         breakout = 0;
 440                         FOREACH_THREAD_IN_PROC(p, td) {
 441                                 thread_lock(td);
 442                                 if (!TD_ON_RUNQ(td) &&
 443                                     !TD_IS_RUNNING(td) &&
 444                                     !TD_IS_SLEEPING(td) &&
 445                                     !TD_IS_SUSPENDED(td)) {
 446                                         thread_unlock(td);
 447                                         breakout = 1;
 448                                         break;
 449                                 }
 450                                 thread_unlock(td);
 451                         }
 452                         if (breakout) {
 453                                 PROC_UNLOCK(p);
 454                                 continue;
 455                         }
 456                         /*
 457                          * get a limit
 458                          */
 459                         lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
 460                         limit = OFF_TO_IDX(
 461                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
 462
 463                         /*
 464                          * let processes that are swapped out really be
 465                          * swapped out set the limit to nothing (will force a
 466                          * swap-out.)
 467                          */
 468                         if ((p->p_flag & P_INMEM) == 0)
 469                                 limit = 0;      /* XXX */
 470                         vm = vmspace_acquire_ref(p);
 471                         _PHOLD_LITE(p);
 472                         PROC_UNLOCK(p);
 473                         if (vm == NULL) {
 474                                 PRELE(p);
 475                                 continue;
 476                         }
 477                         sx_sunlock(&allproc_lock);
 478
 479                         size = vmspace_resident_count(vm);
 480                         if (size >= limit) {
 481                                 vm_swapout_map_deactivate_pages(
 482                                     &vm->vm_map, limit);
 483                                 size = vmspace_resident_count(vm);
 484                         }
 485 #ifdef RACCT
 486                         if (racct_enable) {
 487                                 rsize = IDX_TO_OFF(size);
 488                                 PROC_LOCK(p);
 489                                 if (p->p_state == PRS_NORMAL)
 490                                         racct_set(p, RACCT_RSS, rsize);
 491                                 ravailable = racct_get_available(p, RACCT_RSS);
 492                                 PROC_UNLOCK(p);
 493                                 if (rsize > ravailable) {
 494                                         /*
 495                                          * Don't be overly aggressive; this
 496                                          * might be an innocent process,
 497                                          * and the limit could've been exceeded
 498                                          * by some memory hog.  Don't try
 499                                          * to deactivate more than 1/4th
 500                                          * of process' resident set size.
 501                                          */
 502                                         if (attempts <= 8) {
 503                                                 if (ravailable < rsize -
 504                                                     (rsize / 4)) {
 505                                                         ravailable = rsize -
 506                                                             (rsize / 4);
 507                                                 }
 508                                         }
 509                                         vm_swapout_map_deactivate_pages(
 510                                             &vm->vm_map,
 511                                             OFF_TO_IDX(ravailable));
 512                                         /* Update RSS usage after paging out. */
 513                                         size = vmspace_resident_count(vm);
 514                                         rsize = IDX_TO_OFF(size);
 515                                         PROC_LOCK(p);
 516                                         if (p->p_state == PRS_NORMAL)
 517                                                 racct_set(p, RACCT_RSS, rsize);
 518                                         PROC_UNLOCK(p);
 519                                         if (rsize > ravailable)
 520                                                 tryagain = 1;
 521                                 }
 522                         }
 523 #endif
 524                         vmspace_free(vm);
 525                         sx_slock(&allproc_lock);
 526                         PRELE(p);
 527                 }
 528                 sx_sunlock(&allproc_lock);
 529                 if (tryagain != 0 && attempts <= 10) {
 530                         maybe_yield();
 531                         goto again;
 532                 }
 533         }
 534 }
 535
 536 /*
 537  * Allow a thread's kernel stack to be paged out.
 538  */
 539 static void
 540 vm_thread_swapout(struct thread *td)
 541 {
 542         vm_object_t ksobj;
 543         vm_page_t m;
 544         int i, pages;
 545
 546         cpu_thread_swapout(td);
 547         pages = td->td_kstack_pages;
 548         ksobj = td->td_kstack_obj;
 549         pmap_qremove(td->td_kstack, pages);
 550         VM_OBJECT_WLOCK(ksobj);
 551         for (i = 0; i < pages; i++) {
 552                 m = vm_page_lookup(ksobj, i);
 553                 if (m == NULL)
 554                         panic("vm_thread_swapout: kstack already missing?");
 555                 vm_page_dirty(m);
 556                 vm_page_lock(m);
 557                 vm_page_unwire(m, PQ_LAUNDRY);
 558                 vm_page_unlock(m);
 559         }
 560         VM_OBJECT_WUNLOCK(ksobj);
 561 }
 562
 563 /*
 564  * Bring the kernel stack for a specified thread back in.
 565  */
 566 static void
 567 vm_thread_swapin(struct thread *td, int oom_alloc)
 568 {
 569         vm_object_t ksobj;
 570         vm_page_t ma[KSTACK_MAX_PAGES];
 571         int a, count, i, j, pages, rv;
 572
 573         pages = td->td_kstack_pages;
 574         ksobj = td->td_kstack_obj;
 575         VM_OBJECT_WLOCK(ksobj);
 576         (void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma,
 577             pages);
 578         for (i = 0; i < pages;) {
 579                 vm_page_assert_xbusied(ma[i]);
 580                 if (ma[i]->valid == VM_PAGE_BITS_ALL) {
 581                         vm_page_xunbusy(ma[i]);
 582                         i++;
 583                         continue;
 584                 }
 585                 vm_object_pip_add(ksobj, 1);
 586                 for (j = i + 1; j < pages; j++)
 587                         if (ma[j]->valid == VM_PAGE_BITS_ALL)
 588                                 break;
 589                 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
 590                 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
 591                 count = min(a + 1, j - i);
 592                 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
 593                 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
 594                     __func__, td->td_proc->p_pid));
 595                 vm_object_pip_wakeup(ksobj);
 596                 for (j = i; j < i + count; j++)
 597                         vm_page_xunbusy(ma[j]);
 598                 i += count;
 599         }
 600         VM_OBJECT_WUNLOCK(ksobj);
 601         pmap_qenter(td->td_kstack, ma, pages);
 602         cpu_thread_swapin(td);
 603 }
 604
 605 void
 606 faultin(struct proc *p)
 607 {
 608         struct thread *td;
 609         int oom_alloc;
 610
 611         PROC_LOCK_ASSERT(p, MA_OWNED);
 612
 613         /*
 614          * If another process is swapping in this process,
 615          * just wait until it finishes.
 616          */
 617         if (p->p_flag & P_SWAPPINGIN) {
 618                 while (p->p_flag & P_SWAPPINGIN)
 619                         msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
 620                 return;
 621         }
 622
 623         if ((p->p_flag & P_INMEM) == 0) {
 624                 oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
 625                     VM_ALLOC_NORMAL;
 626
 627                 /*
 628                  * Don't let another thread swap process p out while we are
 629                  * busy swapping it in.
 630                  */
 631                 ++p->p_lock;
 632                 p->p_flag |= P_SWAPPINGIN;
 633                 PROC_UNLOCK(p);
 634                 sx_xlock(&allproc_lock);
 635                 MPASS(swapped_cnt > 0);
 636                 swapped_cnt--;
 637                 sx_xunlock(&allproc_lock);
 638
 639                 /*
 640                  * We hold no lock here because the list of threads
 641                  * can not change while all threads in the process are
 642                  * swapped out.
 643                  */
 644                 FOREACH_THREAD_IN_PROC(p, td)
 645                         vm_thread_swapin(td, oom_alloc);
 646
 647                 PROC_LOCK(p);
 648                 swapclear(p);
 649                 p->p_swtick = ticks;
 650
 651                 /* Allow other threads to swap p out now. */
 652                 wakeup(&p->p_flag);
 653                 --p->p_lock;
 654         }
 655 }
 656
 657 /*
 658  * This swapin algorithm attempts to swap-in processes only if there
 659  * is enough space for them.  Of course, if a process waits for a long
 660  * time, it will be swapped in anyway.
 661  */
 662
 663 static struct proc *
 664 swapper_selector(void)
 665 {
 666         struct proc *p, *res;
 667         struct thread *td;
 668         int min_flag, ppri, pri, slptime, swtime;
 669
 670         sx_assert(&allproc_lock, SA_SLOCKED);
 671         if (swapped_cnt == 0)
 672                 return (NULL);
 673         res = NULL;
 674         ppri = INT_MIN;
 675         min_flag = vm_page_count_min();
 676         FOREACH_PROC_IN_SYSTEM(p) {
 677                 PROC_LOCK(p);
 678                 if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
 679                     P_SWAPPINGIN | P_INMEM)) != 0) {
 680                         PROC_UNLOCK(p);
 681                         continue;
 682                 }
 683                 if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
 684                         /*
 685                          * A swapped-out process might have mapped a
 686                          * large portion of the system's pages as
 687                          * anonymous memory.  There is no other way to
 688                          * release the memory other than to kill the
 689                          * process, for which we need to swap it in.
 690                          */
 691                         return (p);
 692                 }
 693                 if (min_flag) {
 694                         PROC_UNLOCK(p);
 695                         continue;
 696                 }
 697                 swtime = (ticks - p->p_swtick) / hz;
 698                 FOREACH_THREAD_IN_PROC(p, td) {
 699                         /*
 700                          * An otherwise runnable thread of a process
 701                          * swapped out has only the TDI_SWAPPED bit set.
 702                          */
 703                         thread_lock(td);
 704                         if (td->td_inhibitors == TDI_SWAPPED) {
 705                                 slptime = (ticks - td->td_slptick) / hz;
 706                                 pri = swtime + slptime;
 707                                 if ((td->td_flags & TDF_SWAPINREQ) == 0)
 708                                         pri -= p->p_nice * 8;
 709                                 /*
 710                                  * if this thread is higher priority
 711                                  * and there is enough space, then select
 712                                  * this process instead of the previous
 713                                  * selection.
 714                                  */
 715                                 if (pri > ppri) {
 716                                         res = p;
 717                                         ppri = pri;
 718                                 }
 719                         }
 720                         thread_unlock(td);
 721                 }
 722                 PROC_UNLOCK(p);
 723         }
 724         if (res != NULL)
 725                 PROC_LOCK(res);
 726         return (res);
 727 }
 728
 729 void
 730 swapper(void)
 731 {
 732         struct proc *p;
 733
 734         for (;;) {
 735                 sx_slock(&allproc_lock);
 736                 p = swapper_selector();
 737                 sx_sunlock(&allproc_lock);
 738
 739                 if (p == NULL) {
 740                         tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2);
 741                 } else {
 742                         PROC_LOCK_ASSERT(p, MA_OWNED);
 743
 744                         /*
 745                          * Another process may be bringing or may have
 746                          * already brought this process in while we
 747                          * traverse all threads.  Or, this process may
 748                          * have exited or even being swapped out
 749                          * again.
 750                          */
 751                         if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
 752                             P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
 753                                 faultin(p);
 754                         }
 755                         PROC_UNLOCK(p);
 756                 }
 757         }
 758 }
 759
 760 /*
 761  * First, if any processes have been sleeping or stopped for at least
 762  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
 763  * no such processes exist, then the longest-sleeping or stopped
 764  * process is swapped out.  Finally, and only as a last resort, if
 765  * there are no sleeping or stopped processes, the longest-resident
 766  * process is swapped out.
 767  */
 768 static void
 769 swapout_procs(int action)
 770 {
 771         struct proc *p;
 772         struct thread *td;
 773         int slptime;
 774         bool didswap, doswap;
 775
 776         MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
 777
 778         didswap = false;
 779         sx_slock(&allproc_lock);
 780         FOREACH_PROC_IN_SYSTEM(p) {
 781                 /*
 782                  * Filter out not yet fully constructed processes.  Do
 783                  * not swap out held processes.  Avoid processes which
 784                  * are system, exiting, execing, traced, already swapped
 785                  * out or are in the process of being swapped in or out.
 786                  */
 787                 PROC_LOCK(p);
 788                 if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
 789                     (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
 790                     P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
 791                     P_INMEM) {
 792                         PROC_UNLOCK(p);
 793                         continue;
 794                 }
 795
 796                 /*
 797                  * Further consideration of this process for swap out
 798                  * requires iterating over its threads.  We release
 799                  * allproc_lock here so that process creation and
 800                  * destruction are not blocked while we iterate.
 801                  *
 802                  * To later reacquire allproc_lock and resume
 803                  * iteration over the allproc list, we will first have
 804                  * to release the lock on the process.  We place a
 805                  * hold on the process so that it remains in the
 806                  * allproc list while it is unlocked.
 807                  */
 808                 _PHOLD_LITE(p);
 809                 sx_sunlock(&allproc_lock);
 810
 811                 /*
 812                  * Do not swapout a realtime process.
 813                  * Guarantee swap_idle_threshold1 time in memory.
 814                  * If the system is under memory stress, or if we are
 815                  * swapping idle processes >= swap_idle_threshold2,
 816                  * then swap the process out.
 817                  */
 818                 doswap = true;
 819                 FOREACH_THREAD_IN_PROC(p, td) {
 820                         thread_lock(td);
 821                         slptime = (ticks - td->td_slptick) / hz;
 822                         if (PRI_IS_REALTIME(td->td_pri_class) ||
 823                             slptime < swap_idle_threshold1 ||
 824                             !thread_safetoswapout(td) ||
 825                             ((action & VM_SWAP_NORMAL) == 0 &&
 826                             slptime < swap_idle_threshold2))
 827                                 doswap = false;
 828                         thread_unlock(td);
 829                         if (!doswap)
 830                                 break;
 831                 }
 832                 if (doswap && swapout(p) == 0)
 833                         didswap = true;
 834
 835                 PROC_UNLOCK(p);
 836                 if (didswap) {
 837                         sx_xlock(&allproc_lock);
 838                         swapped_cnt++;
 839                         sx_downgrade(&allproc_lock);
 840                 } else
 841                         sx_slock(&allproc_lock);
 842                 PRELE(p);
 843         }
 844         sx_sunlock(&allproc_lock);
 845
 846         /*
 847          * If we swapped something out, and another process needed memory,
 848          * then wakeup the sched process.
 849          */
 850         if (didswap)
 851                 wakeup(&proc0);
 852 }
 853
 854 static void
 855 swapclear(struct proc *p)
 856 {
 857         struct thread *td;
 858
 859         PROC_LOCK_ASSERT(p, MA_OWNED);
 860
 861         FOREACH_THREAD_IN_PROC(p, td) {
 862                 thread_lock(td);
 863                 td->td_flags |= TDF_INMEM;
 864                 td->td_flags &= ~TDF_SWAPINREQ;
 865                 TD_CLR_SWAPPED(td);
 866                 if (TD_CAN_RUN(td))
 867                         if (setrunnable(td)) {
 868 #ifdef INVARIANTS
 869                                 /*
 870                                  * XXX: We just cleared TDI_SWAPPED
 871                                  * above and set TDF_INMEM, so this
 872                                  * should never happen.
 873                                  */
 874                                 panic("not waking up swapper");
 875 #endif
 876                         }
 877                 thread_unlock(td);
 878         }
 879         p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
 880         p->p_flag |= P_INMEM;
 881 }
 882
 883 static int
 884 swapout(struct proc *p)
 885 {
 886         struct thread *td;
 887
 888         PROC_LOCK_ASSERT(p, MA_OWNED);
 889
 890         /*
 891          * The states of this process and its threads may have changed
 892          * by now.  Assuming that there is only one pageout daemon thread,
 893          * this process should still be in memory.
 894          */
 895         KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
 896             P_INMEM, ("swapout: lost a swapout race?"));
 897
 898         /*
 899          * Remember the resident count.
 900          */
 901         p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
 902
 903         /*
 904          * Check and mark all threads before we proceed.
 905          */
 906         p->p_flag &= ~P_INMEM;
 907         p->p_flag |= P_SWAPPINGOUT;
 908         FOREACH_THREAD_IN_PROC(p, td) {
 909                 thread_lock(td);
 910                 if (!thread_safetoswapout(td)) {
 911                         thread_unlock(td);
 912                         swapclear(p);
 913                         return (EBUSY);
 914                 }
 915                 td->td_flags &= ~TDF_INMEM;
 916                 TD_SET_SWAPPED(td);
 917                 thread_unlock(td);
 918         }
 919         td = FIRST_THREAD_IN_PROC(p);
 920         ++td->td_ru.ru_nswap;
 921         PROC_UNLOCK(p);
 922
 923         /*
 924          * This list is stable because all threads are now prevented from
 925          * running.  The list is only modified in the context of a running
 926          * thread in this process.
 927          */
 928         FOREACH_THREAD_IN_PROC(p, td)
 929                 vm_thread_swapout(td);
 930
 931         PROC_LOCK(p);
 932         p->p_flag &= ~P_SWAPPINGOUT;
 933         p->p_swtick = ticks;
 934         return (0);
 935 }