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 static int swap_inprogress;     /* Pending swap-ins done outside swapper. */
 163 static int last_swapin;
 164
 165 static void swapclear(struct proc *);
 166 static int swapout(struct proc *);
 167 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
 168 static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long);
 169 static void swapout_procs(int action);
 170 static void vm_req_vmdaemon(int req);
 171 static void vm_thread_swapout(struct thread *td);
 172
 173 /*
 174  *      vm_swapout_object_deactivate_pages
 175  *
 176  *      Deactivate enough pages to satisfy the inactive target
 177  *      requirements.
 178  *
 179  *      The object and map must be locked.
 180  */
 181 static void
 182 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
 183     long desired)
 184 {
 185         vm_object_t backing_object, object;
 186         vm_page_t p;
 187         int act_delta, remove_mode;
 188
 189         VM_OBJECT_ASSERT_LOCKED(first_object);
 190         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 191                 return;
 192         for (object = first_object;; object = backing_object) {
 193                 if (pmap_resident_count(pmap) <= desired)
 194                         goto unlock_return;
 195                 VM_OBJECT_ASSERT_LOCKED(object);
 196                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 197                     object->paging_in_progress != 0)
 198                         goto unlock_return;
 199
 200                 remove_mode = 0;
 201                 if (object->shadow_count > 1)
 202                         remove_mode = 1;
 203                 /*
 204                  * Scan the object's entire memory queue.
 205                  */
 206                 TAILQ_FOREACH(p, &object->memq, listq) {
 207                         if (pmap_resident_count(pmap) <= desired)
 208                                 goto unlock_return;
 209                         if (should_yield())
 210                                 goto unlock_return;
 211                         if (vm_page_busied(p))
 212                                 continue;
 213                         VM_CNT_INC(v_pdpages);
 214                         vm_page_lock(p);
 215                         if (vm_page_held(p) ||
 216                             !pmap_page_exists_quick(pmap, p)) {
 217                                 vm_page_unlock(p);
 218                                 continue;
 219                         }
 220                         act_delta = pmap_ts_referenced(p);
 221                         if ((p->aflags & PGA_REFERENCED) != 0) {
 222                                 if (act_delta == 0)
 223                                         act_delta = 1;
 224                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 225                         }
 226                         if (!vm_page_active(p) && act_delta != 0) {
 227                                 vm_page_activate(p);
 228                                 p->act_count += act_delta;
 229                         } else if (vm_page_active(p)) {
 230                                 /*
 231                                  * The page daemon does not requeue pages
 232                                  * after modifying their activation count.
 233                                  */
 234                                 if (act_delta == 0) {
 235                                         p->act_count -= min(p->act_count,
 236                                             ACT_DECLINE);
 237                                         if (!remove_mode && p->act_count == 0) {
 238                                                 pmap_remove_all(p);
 239                                                 vm_page_deactivate(p);
 240                                         }
 241                                 } else {
 242                                         vm_page_activate(p);
 243                                         if (p->act_count < ACT_MAX -
 244                                             ACT_ADVANCE)
 245                                                 p->act_count += ACT_ADVANCE;
 246                                 }
 247                         } else if (vm_page_inactive(p))
 248                                 pmap_remove_all(p);
 249                         vm_page_unlock(p);
 250                 }
 251                 if ((backing_object = object->backing_object) == NULL)
 252                         goto unlock_return;
 253                 VM_OBJECT_RLOCK(backing_object);
 254                 if (object != first_object)
 255                         VM_OBJECT_RUNLOCK(object);
 256         }
 257 unlock_return:
 258         if (object != first_object)
 259                 VM_OBJECT_RUNLOCK(object);
 260 }
 261
 262 /*
 263  * deactivate some number of pages in a map, try to do it fairly, but
 264  * that is really hard to do.
 265  */
 266 static void
 267 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
 268 {
 269         vm_map_entry_t tmpe;
 270         vm_object_t obj, bigobj;
 271         int nothingwired;
 272
 273         if (!vm_map_trylock_read(map))
 274                 return;
 275
 276         bigobj = NULL;
 277         nothingwired = TRUE;
 278
 279         /*
 280          * first, search out the biggest object, and try to free pages from
 281          * that.
 282          */
 283         tmpe = map->header.next;
 284         while (tmpe != &map->header) {
 285                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 286                         obj = tmpe->object.vm_object;
 287                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 288                                 if (obj->shadow_count <= 1 &&
 289                                     (bigobj == NULL ||
 290                                      bigobj->resident_page_count <
 291                                      obj->resident_page_count)) {
 292                                         if (bigobj != NULL)
 293                                                 VM_OBJECT_RUNLOCK(bigobj);
 294                                         bigobj = obj;
 295                                 } else
 296                                         VM_OBJECT_RUNLOCK(obj);
 297                         }
 298                 }
 299                 if (tmpe->wired_count > 0)
 300                         nothingwired = FALSE;
 301                 tmpe = tmpe->next;
 302         }
 303
 304         if (bigobj != NULL) {
 305                 vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
 306                 VM_OBJECT_RUNLOCK(bigobj);
 307         }
 308         /*
 309          * Next, hunt around for other pages to deactivate.  We actually
 310          * do this search sort of wrong -- .text first is not the best idea.
 311          */
 312         tmpe = map->header.next;
 313         while (tmpe != &map->header) {
 314                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 315                         break;
 316                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 317                         obj = tmpe->object.vm_object;
 318                         if (obj != NULL) {
 319                                 VM_OBJECT_RLOCK(obj);
 320                                 vm_swapout_object_deactivate_pages(map->pmap,
 321                                     obj, desired);
 322                                 VM_OBJECT_RUNLOCK(obj);
 323                         }
 324                 }
 325                 tmpe = tmpe->next;
 326         }
 327
 328         /*
 329          * Remove all mappings if a process is swapped out, this will free page
 330          * table pages.
 331          */
 332         if (desired == 0 && nothingwired) {
 333                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 334                     vm_map_max(map));
 335         }
 336
 337         vm_map_unlock_read(map);
 338 }
 339
 340 /*
 341  * Swap out requests
 342  */
 343 #define VM_SWAP_NORMAL 1
 344 #define VM_SWAP_IDLE 2
 345
 346 void
 347 vm_swapout_run(void)
 348 {
 349
 350         if (vm_swap_enabled)
 351                 vm_req_vmdaemon(VM_SWAP_NORMAL);
 352 }
 353
 354 /*
 355  * Idle process swapout -- run once per second when pagedaemons are
 356  * reclaiming pages.
 357  */
 358 void
 359 vm_swapout_run_idle(void)
 360 {
 361         static long lsec;
 362
 363         if (!vm_swap_idle_enabled || time_second == lsec)
 364                 return;
 365         vm_req_vmdaemon(VM_SWAP_IDLE);
 366         lsec = time_second;
 367 }
 368
 369 static void
 370 vm_req_vmdaemon(int req)
 371 {
 372         static int lastrun = 0;
 373
 374         mtx_lock(&vm_daemon_mtx);
 375         vm_pageout_req_swapout |= req;
 376         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
 377                 wakeup(&vm_daemon_needed);
 378                 lastrun = ticks;
 379         }
 380         mtx_unlock(&vm_daemon_mtx);
 381 }
 382
 383 static void
 384 vm_daemon(void)
 385 {
 386         struct rlimit rsslim;
 387         struct proc *p;
 388         struct thread *td;
 389         struct vmspace *vm;
 390         int breakout, swapout_flags, tryagain, attempts;
 391 #ifdef RACCT
 392         uint64_t rsize, ravailable;
 393 #endif
 394
 395         while (TRUE) {
 396                 mtx_lock(&vm_daemon_mtx);
 397                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
 398 #ifdef RACCT
 399                     racct_enable ? hz : 0
 400 #else
 401                     0
 402 #endif
 403                 );
 404                 swapout_flags = vm_pageout_req_swapout;
 405                 vm_pageout_req_swapout = 0;
 406                 mtx_unlock(&vm_daemon_mtx);
 407                 if (swapout_flags != 0) {
 408                         /*
 409                          * Drain the per-CPU page queue batches as a deadlock
 410                          * avoidance measure.
 411                          */
 412                         if ((swapout_flags & VM_SWAP_NORMAL) != 0)
 413                                 vm_page_drain_pqbatch();
 414                         swapout_procs(swapout_flags);
 415                 }
 416
 417                 /*
 418                  * scan the processes for exceeding their rlimits or if
 419                  * process is swapped out -- deactivate pages
 420                  */
 421                 tryagain = 0;
 422                 attempts = 0;
 423 again:
 424                 attempts++;
 425                 sx_slock(&allproc_lock);
 426                 FOREACH_PROC_IN_SYSTEM(p) {
 427                         vm_pindex_t limit, size;
 428
 429                         /*
 430                          * if this is a system process or if we have already
 431                          * looked at this process, skip it.
 432                          */
 433                         PROC_LOCK(p);
 434                         if (p->p_state != PRS_NORMAL ||
 435                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
 436                                 PROC_UNLOCK(p);
 437                                 continue;
 438                         }
 439                         /*
 440                          * if the process is in a non-running type state,
 441                          * don't touch it.
 442                          */
 443                         breakout = 0;
 444                         FOREACH_THREAD_IN_PROC(p, td) {
 445                                 thread_lock(td);
 446                                 if (!TD_ON_RUNQ(td) &&
 447                                     !TD_IS_RUNNING(td) &&
 448                                     !TD_IS_SLEEPING(td) &&
 449                                     !TD_IS_SUSPENDED(td)) {
 450                                         thread_unlock(td);
 451                                         breakout = 1;
 452                                         break;
 453                                 }
 454                                 thread_unlock(td);
 455                         }
 456                         if (breakout) {
 457                                 PROC_UNLOCK(p);
 458                                 continue;
 459                         }
 460                         /*
 461                          * get a limit
 462                          */
 463                         lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
 464                         limit = OFF_TO_IDX(
 465                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
 466
 467                         /*
 468                          * let processes that are swapped out really be
 469                          * swapped out set the limit to nothing (will force a
 470                          * swap-out.)
 471                          */
 472                         if ((p->p_flag & P_INMEM) == 0)
 473                                 limit = 0;      /* XXX */
 474                         vm = vmspace_acquire_ref(p);
 475                         _PHOLD_LITE(p);
 476                         PROC_UNLOCK(p);
 477                         if (vm == NULL) {
 478                                 PRELE(p);
 479                                 continue;
 480                         }
 481                         sx_sunlock(&allproc_lock);
 482
 483                         size = vmspace_resident_count(vm);
 484                         if (size >= limit) {
 485                                 vm_swapout_map_deactivate_pages(
 486                                     &vm->vm_map, limit);
 487                                 size = vmspace_resident_count(vm);
 488                         }
 489 #ifdef RACCT
 490                         if (racct_enable) {
 491                                 rsize = IDX_TO_OFF(size);
 492                                 PROC_LOCK(p);
 493                                 if (p->p_state == PRS_NORMAL)
 494                                         racct_set(p, RACCT_RSS, rsize);
 495                                 ravailable = racct_get_available(p, RACCT_RSS);
 496                                 PROC_UNLOCK(p);
 497                                 if (rsize > ravailable) {
 498                                         /*
 499                                          * Don't be overly aggressive; this
 500                                          * might be an innocent process,
 501                                          * and the limit could've been exceeded
 502                                          * by some memory hog.  Don't try
 503                                          * to deactivate more than 1/4th
 504                                          * of process' resident set size.
 505                                          */
 506                                         if (attempts <= 8) {
 507                                                 if (ravailable < rsize -
 508                                                     (rsize / 4)) {
 509                                                         ravailable = rsize -
 510                                                             (rsize / 4);
 511                                                 }
 512                                         }
 513                                         vm_swapout_map_deactivate_pages(
 514                                             &vm->vm_map,
 515                                             OFF_TO_IDX(ravailable));
 516                                         /* Update RSS usage after paging out. */
 517                                         size = vmspace_resident_count(vm);
 518                                         rsize = IDX_TO_OFF(size);
 519                                         PROC_LOCK(p);
 520                                         if (p->p_state == PRS_NORMAL)
 521                                                 racct_set(p, RACCT_RSS, rsize);
 522                                         PROC_UNLOCK(p);
 523                                         if (rsize > ravailable)
 524                                                 tryagain = 1;
 525                                 }
 526                         }
 527 #endif
 528                         vmspace_free(vm);
 529                         sx_slock(&allproc_lock);
 530                         PRELE(p);
 531                 }
 532                 sx_sunlock(&allproc_lock);
 533                 if (tryagain != 0 && attempts <= 10) {
 534                         maybe_yield();
 535                         goto again;
 536                 }
 537         }
 538 }
 539
 540 /*
 541  * Allow a thread's kernel stack to be paged out.
 542  */
 543 static void
 544 vm_thread_swapout(struct thread *td)
 545 {
 546         vm_object_t ksobj;
 547         vm_page_t m;
 548         int i, pages;
 549
 550         cpu_thread_swapout(td);
 551         pages = td->td_kstack_pages;
 552         ksobj = td->td_kstack_obj;
 553         pmap_qremove(td->td_kstack, pages);
 554         VM_OBJECT_WLOCK(ksobj);
 555         for (i = 0; i < pages; i++) {
 556                 m = vm_page_lookup(ksobj, i);
 557                 if (m == NULL)
 558                         panic("vm_thread_swapout: kstack already missing?");
 559                 vm_page_dirty(m);
 560                 vm_page_lock(m);
 561                 vm_page_unwire(m, PQ_LAUNDRY);
 562                 vm_page_unlock(m);
 563         }
 564         VM_OBJECT_WUNLOCK(ksobj);
 565 }
 566
 567 /*
 568  * Bring the kernel stack for a specified thread back in.
 569  */
 570 static void
 571 vm_thread_swapin(struct thread *td, int oom_alloc)
 572 {
 573         vm_object_t ksobj;
 574         vm_page_t ma[KSTACK_MAX_PAGES];
 575         int a, count, i, j, pages, rv;
 576
 577         pages = td->td_kstack_pages;
 578         ksobj = td->td_kstack_obj;
 579         VM_OBJECT_WLOCK(ksobj);
 580         (void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma,
 581             pages);
 582         for (i = 0; i < pages;) {
 583                 vm_page_assert_xbusied(ma[i]);
 584                 if (ma[i]->valid == VM_PAGE_BITS_ALL) {
 585                         vm_page_xunbusy(ma[i]);
 586                         i++;
 587                         continue;
 588                 }
 589                 vm_object_pip_add(ksobj, 1);
 590                 for (j = i + 1; j < pages; j++)
 591                         if (ma[j]->valid == VM_PAGE_BITS_ALL)
 592                                 break;
 593                 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
 594                 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
 595                 count = min(a + 1, j - i);
 596                 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
 597                 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
 598                     __func__, td->td_proc->p_pid));
 599                 vm_object_pip_wakeup(ksobj);
 600                 for (j = i; j < i + count; j++)
 601                         vm_page_xunbusy(ma[j]);
 602                 i += count;
 603         }
 604         VM_OBJECT_WUNLOCK(ksobj);
 605         pmap_qenter(td->td_kstack, ma, pages);
 606         cpu_thread_swapin(td);
 607 }
 608
 609 void
 610 faultin(struct proc *p)
 611 {
 612         struct thread *td;
 613         int oom_alloc;
 614
 615         PROC_LOCK_ASSERT(p, MA_OWNED);
 616
 617         /*
 618          * If another process is swapping in this process,
 619          * just wait until it finishes.
 620          */
 621         if (p->p_flag & P_SWAPPINGIN) {
 622                 while (p->p_flag & P_SWAPPINGIN)
 623                         msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
 624                 return;
 625         }
 626
 627         if ((p->p_flag & P_INMEM) == 0) {
 628                 oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
 629                     VM_ALLOC_NORMAL;
 630
 631                 /*
 632                  * Don't let another thread swap process p out while we are
 633                  * busy swapping it in.
 634                  */
 635                 ++p->p_lock;
 636                 p->p_flag |= P_SWAPPINGIN;
 637                 PROC_UNLOCK(p);
 638                 sx_xlock(&allproc_lock);
 639                 MPASS(swapped_cnt > 0);
 640                 swapped_cnt--;
 641                 if (curthread != &thread0)
 642                         swap_inprogress++;
 643                 sx_xunlock(&allproc_lock);
 644
 645                 /*
 646                  * We hold no lock here because the list of threads
 647                  * can not change while all threads in the process are
 648                  * swapped out.
 649                  */
 650                 FOREACH_THREAD_IN_PROC(p, td)
 651                         vm_thread_swapin(td, oom_alloc);
 652
 653                 if (curthread != &thread0) {
 654                         sx_xlock(&allproc_lock);
 655                         MPASS(swap_inprogress > 0);
 656                         swap_inprogress--;
 657                         last_swapin = ticks;
 658                         sx_xunlock(&allproc_lock);
 659                 }
 660                 PROC_LOCK(p);
 661                 swapclear(p);
 662                 p->p_swtick = ticks;
 663
 664                 /* Allow other threads to swap p out now. */
 665                 wakeup(&p->p_flag);
 666                 --p->p_lock;
 667         }
 668 }
 669
 670 /*
 671  * This swapin algorithm attempts to swap-in processes only if there
 672  * is enough space for them.  Of course, if a process waits for a long
 673  * time, it will be swapped in anyway.
 674  */
 675
 676 static struct proc *
 677 swapper_selector(bool wkilled_only)
 678 {
 679         struct proc *p, *res;
 680         struct thread *td;
 681         int ppri, pri, slptime, swtime;
 682
 683         sx_assert(&allproc_lock, SA_SLOCKED);
 684         if (swapped_cnt == 0)
 685                 return (NULL);
 686         res = NULL;
 687         ppri = INT_MIN;
 688         FOREACH_PROC_IN_SYSTEM(p) {
 689                 PROC_LOCK(p);
 690                 if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
 691                     P_SWAPPINGIN | P_INMEM)) != 0) {
 692                         PROC_UNLOCK(p);
 693                         continue;
 694                 }
 695                 if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
 696                         /*
 697                          * A swapped-out process might have mapped a
 698                          * large portion of the system's pages as
 699                          * anonymous memory.  There is no other way to
 700                          * release the memory other than to kill the
 701                          * process, for which we need to swap it in.
 702                          */
 703                         return (p);
 704                 }
 705                 if (wkilled_only) {
 706                         PROC_UNLOCK(p);
 707                         continue;
 708                 }
 709                 swtime = (ticks - p->p_swtick) / hz;
 710                 FOREACH_THREAD_IN_PROC(p, td) {
 711                         /*
 712                          * An otherwise runnable thread of a process
 713                          * swapped out has only the TDI_SWAPPED bit set.
 714                          */
 715                         thread_lock(td);
 716                         if (td->td_inhibitors == TDI_SWAPPED) {
 717                                 slptime = (ticks - td->td_slptick) / hz;
 718                                 pri = swtime + slptime;
 719                                 if ((td->td_flags & TDF_SWAPINREQ) == 0)
 720                                         pri -= p->p_nice * 8;
 721                                 /*
 722                                  * if this thread is higher priority
 723                                  * and there is enough space, then select
 724                                  * this process instead of the previous
 725                                  * selection.
 726                                  */
 727                                 if (pri > ppri) {
 728                                         res = p;
 729                                         ppri = pri;
 730                                 }
 731                         }
 732                         thread_unlock(td);
 733                 }
 734                 PROC_UNLOCK(p);
 735         }
 736
 737         if (res != NULL)
 738                 PROC_LOCK(res);
 739         return (res);
 740 }
 741
 742 #define SWAPIN_INTERVAL (MAXSLP * hz / 2)
 743
 744 /*
 745  * Limit swapper to swap in one non-WKILLED process in MAXSLP/2
 746  * interval, assuming that there is:
 747  * - there exists at least one domain that is not suffering from a shortage of
 748  *   free memory;
 749  * - no parallel swap-ins;
 750  * - no other swap-ins in the current SWAPIN_INTERVAL.
 751  */
 752 static bool
 753 swapper_wkilled_only(void)
 754 {
 755
 756         return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 ||
 757             (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL);
 758 }
 759
 760 void
 761 swapper(void)
 762 {
 763         struct proc *p;
 764
 765         for (;;) {
 766                 sx_slock(&allproc_lock);
 767                 p = swapper_selector(swapper_wkilled_only());
 768                 sx_sunlock(&allproc_lock);
 769
 770                 if (p == NULL) {
 771                         tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL);
 772                 } else {
 773                         PROC_LOCK_ASSERT(p, MA_OWNED);
 774
 775                         /*
 776                          * Another process may be bringing or may have
 777                          * already brought this process in while we
 778                          * traverse all threads.  Or, this process may
 779                          * have exited or even being swapped out
 780                          * again.
 781                          */
 782                         if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
 783                             P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
 784                                 faultin(p);
 785                         }
 786                         PROC_UNLOCK(p);
 787                 }
 788         }
 789 }
 790
 791 /*
 792  * First, if any processes have been sleeping or stopped for at least
 793  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
 794  * no such processes exist, then the longest-sleeping or stopped
 795  * process is swapped out.  Finally, and only as a last resort, if
 796  * there are no sleeping or stopped processes, the longest-resident
 797  * process is swapped out.
 798  */
 799 static void
 800 swapout_procs(int action)
 801 {
 802         struct proc *p;
 803         struct thread *td;
 804         int slptime;
 805         bool didswap, doswap;
 806
 807         MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
 808
 809         didswap = false;
 810         sx_slock(&allproc_lock);
 811         FOREACH_PROC_IN_SYSTEM(p) {
 812                 /*
 813                  * Filter out not yet fully constructed processes.  Do
 814                  * not swap out held processes.  Avoid processes which
 815                  * are system, exiting, execing, traced, already swapped
 816                  * out or are in the process of being swapped in or out.
 817                  */
 818                 PROC_LOCK(p);
 819                 if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
 820                     (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
 821                     P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
 822                     P_INMEM) {
 823                         PROC_UNLOCK(p);
 824                         continue;
 825                 }
 826
 827                 /*
 828                  * Further consideration of this process for swap out
 829                  * requires iterating over its threads.  We release
 830                  * allproc_lock here so that process creation and
 831                  * destruction are not blocked while we iterate.
 832                  *
 833                  * To later reacquire allproc_lock and resume
 834                  * iteration over the allproc list, we will first have
 835                  * to release the lock on the process.  We place a
 836                  * hold on the process so that it remains in the
 837                  * allproc list while it is unlocked.
 838                  */
 839                 _PHOLD_LITE(p);
 840                 sx_sunlock(&allproc_lock);
 841
 842                 /*
 843                  * Do not swapout a realtime process.
 844                  * Guarantee swap_idle_threshold1 time in memory.
 845                  * If the system is under memory stress, or if we are
 846                  * swapping idle processes >= swap_idle_threshold2,
 847                  * then swap the process out.
 848                  */
 849                 doswap = true;
 850                 FOREACH_THREAD_IN_PROC(p, td) {
 851                         thread_lock(td);
 852                         slptime = (ticks - td->td_slptick) / hz;
 853                         if (PRI_IS_REALTIME(td->td_pri_class) ||
 854                             slptime < swap_idle_threshold1 ||
 855                             !thread_safetoswapout(td) ||
 856                             ((action & VM_SWAP_NORMAL) == 0 &&
 857                             slptime < swap_idle_threshold2))
 858                                 doswap = false;
 859                         thread_unlock(td);
 860                         if (!doswap)
 861                                 break;
 862                 }
 863                 if (doswap && swapout(p) == 0)
 864                         didswap = true;
 865
 866                 PROC_UNLOCK(p);
 867                 if (didswap) {
 868                         sx_xlock(&allproc_lock);
 869                         swapped_cnt++;
 870                         sx_downgrade(&allproc_lock);
 871                 } else
 872                         sx_slock(&allproc_lock);
 873                 PRELE(p);
 874         }
 875         sx_sunlock(&allproc_lock);
 876
 877         /*
 878          * If we swapped something out, and another process needed memory,
 879          * then wakeup the sched process.
 880          */
 881         if (didswap)
 882                 wakeup(&proc0);
 883 }
 884
 885 static void
 886 swapclear(struct proc *p)
 887 {
 888         struct thread *td;
 889
 890         PROC_LOCK_ASSERT(p, MA_OWNED);
 891
 892         FOREACH_THREAD_IN_PROC(p, td) {
 893                 thread_lock(td);
 894                 td->td_flags |= TDF_INMEM;
 895                 td->td_flags &= ~TDF_SWAPINREQ;
 896                 TD_CLR_SWAPPED(td);
 897                 if (TD_CAN_RUN(td))
 898                         if (setrunnable(td)) {
 899 #ifdef INVARIANTS
 900                                 /*
 901                                  * XXX: We just cleared TDI_SWAPPED
 902                                  * above and set TDF_INMEM, so this
 903                                  * should never happen.
 904                                  */
 905                                 panic("not waking up swapper");
 906 #endif
 907                         }
 908                 thread_unlock(td);
 909         }
 910         p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
 911         p->p_flag |= P_INMEM;
 912 }
 913
 914 static int
 915 swapout(struct proc *p)
 916 {
 917         struct thread *td;
 918
 919         PROC_LOCK_ASSERT(p, MA_OWNED);
 920
 921         /*
 922          * The states of this process and its threads may have changed
 923          * by now.  Assuming that there is only one pageout daemon thread,
 924          * this process should still be in memory.
 925          */
 926         KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
 927             P_INMEM, ("swapout: lost a swapout race?"));
 928
 929         /*
 930          * Remember the resident count.
 931          */
 932         p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
 933
 934         /*
 935          * Check and mark all threads before we proceed.
 936          */
 937         p->p_flag &= ~P_INMEM;
 938         p->p_flag |= P_SWAPPINGOUT;
 939         FOREACH_THREAD_IN_PROC(p, td) {
 940                 thread_lock(td);
 941                 if (!thread_safetoswapout(td)) {
 942                         thread_unlock(td);
 943                         swapclear(p);
 944                         return (EBUSY);
 945                 }
 946                 td->td_flags &= ~TDF_INMEM;
 947                 TD_SET_SWAPPED(td);
 948                 thread_unlock(td);
 949         }
 950         td = FIRST_THREAD_IN_PROC(p);
 951         ++td->td_ru.ru_nswap;
 952         PROC_UNLOCK(p);
 953
 954         /*
 955          * This list is stable because all threads are now prevented from
 956          * running.  The list is only modified in the context of a running
 957          * thread in this process.
 958          */
 959         FOREACH_THREAD_IN_PROC(p, td)
 960                 vm_thread_swapout(td);
 961
 962         PROC_LOCK(p);
 963         p->p_flag &= ~P_SWAPPINGOUT;
 964         p->p_swtick = ticks;
 965         return (0);
 966 }