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/kthread.h>
  89 #include <sys/ktr.h>
  90 #include <sys/mount.h>
  91 #include <sys/racct.h>
  92 #include <sys/resourcevar.h>
  93 #include <sys/refcount.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                     REFCOUNT_COUNT(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
 212                         /*
 213                          * The page may acquire a wiring after this check.
 214                          * The page daemon handles wired pages, so there is
 215                          * no harm done if a wiring appears while we are
 216                          * attempting to deactivate the page.
 217                          */
 218                         if (vm_page_busied(p) || vm_page_wired(p))
 219                                 continue;
 220                         VM_CNT_INC(v_pdpages);
 221                         if (!pmap_page_exists_quick(pmap, p))
 222                                 continue;
 223                         act_delta = pmap_ts_referenced(p);
 224                         vm_page_lock(p);
 225                         if ((p->aflags & PGA_REFERENCED) != 0) {
 226                                 if (act_delta == 0)
 227                                         act_delta = 1;
 228                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 229                         }
 230                         if (!vm_page_active(p) && act_delta != 0) {
 231                                 vm_page_activate(p);
 232                                 p->act_count += act_delta;
 233                         } else if (vm_page_active(p)) {
 234                                 /*
 235                                  * The page daemon does not requeue pages
 236                                  * after modifying their activation count.
 237                                  */
 238                                 if (act_delta == 0) {
 239                                         p->act_count -= min(p->act_count,
 240                                             ACT_DECLINE);
 241                                         if (!remove_mode && p->act_count == 0) {
 242                                                 (void)vm_page_try_remove_all(p);
 243                                                 vm_page_deactivate(p);
 244                                         }
 245                                 } else {
 246                                         vm_page_activate(p);
 247                                         if (p->act_count < ACT_MAX -
 248                                             ACT_ADVANCE)
 249                                                 p->act_count += ACT_ADVANCE;
 250                                 }
 251                         } else if (vm_page_inactive(p))
 252                                 (void)vm_page_try_remove_all(p);
 253                         vm_page_unlock(p);
 254                 }
 255                 if ((backing_object = object->backing_object) == NULL)
 256                         goto unlock_return;
 257                 VM_OBJECT_RLOCK(backing_object);
 258                 if (object != first_object)
 259                         VM_OBJECT_RUNLOCK(object);
 260         }
 261 unlock_return:
 262         if (object != first_object)
 263                 VM_OBJECT_RUNLOCK(object);
 264 }
 265
 266 /*
 267  * deactivate some number of pages in a map, try to do it fairly, but
 268  * that is really hard to do.
 269  */
 270 static void
 271 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
 272 {
 273         vm_map_entry_t tmpe;
 274         vm_object_t obj, bigobj;
 275         int nothingwired;
 276
 277         if (!vm_map_trylock_read(map))
 278                 return;
 279
 280         bigobj = NULL;
 281         nothingwired = TRUE;
 282
 283         /*
 284          * first, search out the biggest object, and try to free pages from
 285          * that.
 286          */
 287         tmpe = map->header.next;
 288         while (tmpe != &map->header) {
 289                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 290                         obj = tmpe->object.vm_object;
 291                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 292                                 if (obj->shadow_count <= 1 &&
 293                                     (bigobj == NULL ||
 294                                      bigobj->resident_page_count <
 295                                      obj->resident_page_count)) {
 296                                         if (bigobj != NULL)
 297                                                 VM_OBJECT_RUNLOCK(bigobj);
 298                                         bigobj = obj;
 299                                 } else
 300                                         VM_OBJECT_RUNLOCK(obj);
 301                         }
 302                 }
 303                 if (tmpe->wired_count > 0)
 304                         nothingwired = FALSE;
 305                 tmpe = tmpe->next;
 306         }
 307
 308         if (bigobj != NULL) {
 309                 vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
 310                 VM_OBJECT_RUNLOCK(bigobj);
 311         }
 312         /*
 313          * Next, hunt around for other pages to deactivate.  We actually
 314          * do this search sort of wrong -- .text first is not the best idea.
 315          */
 316         tmpe = map->header.next;
 317         while (tmpe != &map->header) {
 318                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 319                         break;
 320                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 321                         obj = tmpe->object.vm_object;
 322                         if (obj != NULL) {
 323                                 VM_OBJECT_RLOCK(obj);
 324                                 vm_swapout_object_deactivate_pages(map->pmap,
 325                                     obj, desired);
 326                                 VM_OBJECT_RUNLOCK(obj);
 327                         }
 328                 }
 329                 tmpe = tmpe->next;
 330         }
 331
 332         /*
 333          * Remove all mappings if a process is swapped out, this will free page
 334          * table pages.
 335          */
 336         if (desired == 0 && nothingwired) {
 337                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 338                     vm_map_max(map));
 339         }
 340
 341         vm_map_unlock_read(map);
 342 }
 343
 344 /*
 345  * Swap out requests
 346  */
 347 #define VM_SWAP_NORMAL 1
 348 #define VM_SWAP_IDLE 2
 349
 350 void
 351 vm_swapout_run(void)
 352 {
 353
 354         if (vm_swap_enabled)
 355                 vm_req_vmdaemon(VM_SWAP_NORMAL);
 356 }
 357
 358 /*
 359  * Idle process swapout -- run once per second when pagedaemons are
 360  * reclaiming pages.
 361  */
 362 void
 363 vm_swapout_run_idle(void)
 364 {
 365         static long lsec;
 366
 367         if (!vm_swap_idle_enabled || time_second == lsec)
 368                 return;
 369         vm_req_vmdaemon(VM_SWAP_IDLE);
 370         lsec = time_second;
 371 }
 372
 373 static void
 374 vm_req_vmdaemon(int req)
 375 {
 376         static int lastrun = 0;
 377
 378         mtx_lock(&vm_daemon_mtx);
 379         vm_pageout_req_swapout |= req;
 380         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
 381                 wakeup(&vm_daemon_needed);
 382                 lastrun = ticks;
 383         }
 384         mtx_unlock(&vm_daemon_mtx);
 385 }
 386
 387 static void
 388 vm_daemon(void)
 389 {
 390         struct rlimit rsslim;
 391         struct proc *p;
 392         struct thread *td;
 393         struct vmspace *vm;
 394         int breakout, swapout_flags, tryagain, attempts;
 395 #ifdef RACCT
 396         uint64_t rsize, ravailable;
 397 #endif
 398
 399         while (TRUE) {
 400                 mtx_lock(&vm_daemon_mtx);
 401                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
 402 #ifdef RACCT
 403                     racct_enable ? hz : 0
 404 #else
 405                     0
 406 #endif
 407                 );
 408                 swapout_flags = vm_pageout_req_swapout;
 409                 vm_pageout_req_swapout = 0;
 410                 mtx_unlock(&vm_daemon_mtx);
 411                 if (swapout_flags != 0) {
 412                         /*
 413                          * Drain the per-CPU page queue batches as a deadlock
 414                          * avoidance measure.
 415                          */
 416                         if ((swapout_flags & VM_SWAP_NORMAL) != 0)
 417                                 vm_page_pqbatch_drain();
 418                         swapout_procs(swapout_flags);
 419                 }
 420
 421                 /*
 422                  * scan the processes for exceeding their rlimits or if
 423                  * process is swapped out -- deactivate pages
 424                  */
 425                 tryagain = 0;
 426                 attempts = 0;
 427 again:
 428                 attempts++;
 429                 sx_slock(&allproc_lock);
 430                 FOREACH_PROC_IN_SYSTEM(p) {
 431                         vm_pindex_t limit, size;
 432
 433                         /*
 434                          * if this is a system process or if we have already
 435                          * looked at this process, skip it.
 436                          */
 437                         PROC_LOCK(p);
 438                         if (p->p_state != PRS_NORMAL ||
 439                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
 440                                 PROC_UNLOCK(p);
 441                                 continue;
 442                         }
 443                         /*
 444                          * if the process is in a non-running type state,
 445                          * don't touch it.
 446                          */
 447                         breakout = 0;
 448                         FOREACH_THREAD_IN_PROC(p, td) {
 449                                 thread_lock(td);
 450                                 if (!TD_ON_RUNQ(td) &&
 451                                     !TD_IS_RUNNING(td) &&
 452                                     !TD_IS_SLEEPING(td) &&
 453                                     !TD_IS_SUSPENDED(td)) {
 454                                         thread_unlock(td);
 455                                         breakout = 1;
 456                                         break;
 457                                 }
 458                                 thread_unlock(td);
 459                         }
 460                         if (breakout) {
 461                                 PROC_UNLOCK(p);
 462                                 continue;
 463                         }
 464                         /*
 465                          * get a limit
 466                          */
 467                         lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
 468                         limit = OFF_TO_IDX(
 469                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
 470
 471                         /*
 472                          * let processes that are swapped out really be
 473                          * swapped out set the limit to nothing (will force a
 474                          * swap-out.)
 475                          */
 476                         if ((p->p_flag & P_INMEM) == 0)
 477                                 limit = 0;      /* XXX */
 478                         vm = vmspace_acquire_ref(p);
 479                         _PHOLD_LITE(p);
 480                         PROC_UNLOCK(p);
 481                         if (vm == NULL) {
 482                                 PRELE(p);
 483                                 continue;
 484                         }
 485                         sx_sunlock(&allproc_lock);
 486
 487                         size = vmspace_resident_count(vm);
 488                         if (size >= limit) {
 489                                 vm_swapout_map_deactivate_pages(
 490                                     &vm->vm_map, limit);
 491                                 size = vmspace_resident_count(vm);
 492                         }
 493 #ifdef RACCT
 494                         if (racct_enable) {
 495                                 rsize = IDX_TO_OFF(size);
 496                                 PROC_LOCK(p);
 497                                 if (p->p_state == PRS_NORMAL)
 498                                         racct_set(p, RACCT_RSS, rsize);
 499                                 ravailable = racct_get_available(p, RACCT_RSS);
 500                                 PROC_UNLOCK(p);
 501                                 if (rsize > ravailable) {
 502                                         /*
 503                                          * Don't be overly aggressive; this
 504                                          * might be an innocent process,
 505                                          * and the limit could've been exceeded
 506                                          * by some memory hog.  Don't try
 507                                          * to deactivate more than 1/4th
 508                                          * of process' resident set size.
 509                                          */
 510                                         if (attempts <= 8) {
 511                                                 if (ravailable < rsize -
 512                                                     (rsize / 4)) {
 513                                                         ravailable = rsize -
 514                                                             (rsize / 4);
 515                                                 }
 516                                         }
 517                                         vm_swapout_map_deactivate_pages(
 518                                             &vm->vm_map,
 519                                             OFF_TO_IDX(ravailable));
 520                                         /* Update RSS usage after paging out. */
 521                                         size = vmspace_resident_count(vm);
 522                                         rsize = IDX_TO_OFF(size);
 523                                         PROC_LOCK(p);
 524                                         if (p->p_state == PRS_NORMAL)
 525                                                 racct_set(p, RACCT_RSS, rsize);
 526                                         PROC_UNLOCK(p);
 527                                         if (rsize > ravailable)
 528                                                 tryagain = 1;
 529                                 }
 530                         }
 531 #endif
 532                         vmspace_free(vm);
 533                         sx_slock(&allproc_lock);
 534                         PRELE(p);
 535                 }
 536                 sx_sunlock(&allproc_lock);
 537                 if (tryagain != 0 && attempts <= 10) {
 538                         maybe_yield();
 539                         goto again;
 540                 }
 541         }
 542 }
 543
 544 /*
 545  * Allow a thread's kernel stack to be paged out.
 546  */
 547 static void
 548 vm_thread_swapout(struct thread *td)
 549 {
 550         vm_object_t ksobj;
 551         vm_page_t m;
 552         int i, pages;
 553
 554         cpu_thread_swapout(td);
 555         pages = td->td_kstack_pages;
 556         ksobj = td->td_kstack_obj;
 557         pmap_qremove(td->td_kstack, pages);
 558         VM_OBJECT_WLOCK(ksobj);
 559         for (i = 0; i < pages; i++) {
 560                 m = vm_page_lookup(ksobj, i);
 561                 if (m == NULL)
 562                         panic("vm_thread_swapout: kstack already missing?");
 563                 vm_page_dirty(m);
 564                 vm_page_unwire(m, PQ_LAUNDRY);
 565         }
 566         VM_OBJECT_WUNLOCK(ksobj);
 567 }
 568
 569 /*
 570  * Bring the kernel stack for a specified thread back in.
 571  */
 572 static void
 573 vm_thread_swapin(struct thread *td, int oom_alloc)
 574 {
 575         vm_object_t ksobj;
 576         vm_page_t ma[KSTACK_MAX_PAGES];
 577         int a, count, i, j, pages, rv;
 578
 579         pages = td->td_kstack_pages;
 580         ksobj = td->td_kstack_obj;
 581         VM_OBJECT_WLOCK(ksobj);
 582         (void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma,
 583             pages);
 584         for (i = 0; i < pages;) {
 585                 vm_page_assert_xbusied(ma[i]);
 586                 if (ma[i]->valid == VM_PAGE_BITS_ALL) {
 587                         vm_page_xunbusy(ma[i]);
 588                         i++;
 589                         continue;
 590                 }
 591                 vm_object_pip_add(ksobj, 1);
 592                 for (j = i + 1; j < pages; j++)
 593                         if (ma[j]->valid == VM_PAGE_BITS_ALL)
 594                                 break;
 595                 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
 596                 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
 597                 count = min(a + 1, j - i);
 598                 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
 599                 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
 600                     __func__, td->td_proc->p_pid));
 601                 vm_object_pip_wakeup(ksobj);
 602                 for (j = i; j < i + count; j++)
 603                         vm_page_xunbusy(ma[j]);
 604                 i += count;
 605         }
 606         VM_OBJECT_WUNLOCK(ksobj);
 607         pmap_qenter(td->td_kstack, ma, pages);
 608         cpu_thread_swapin(td);
 609 }
 610
 611 void
 612 faultin(struct proc *p)
 613 {
 614         struct thread *td;
 615         int oom_alloc;
 616
 617         PROC_LOCK_ASSERT(p, MA_OWNED);
 618
 619         /*
 620          * If another process is swapping in this process,
 621          * just wait until it finishes.
 622          */
 623         if (p->p_flag & P_SWAPPINGIN) {
 624                 while (p->p_flag & P_SWAPPINGIN)
 625                         msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
 626                 return;
 627         }
 628
 629         if ((p->p_flag & P_INMEM) == 0) {
 630                 oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
 631                     VM_ALLOC_NORMAL;
 632
 633                 /*
 634                  * Don't let another thread swap process p out while we are
 635                  * busy swapping it in.
 636                  */
 637                 ++p->p_lock;
 638                 p->p_flag |= P_SWAPPINGIN;
 639                 PROC_UNLOCK(p);
 640                 sx_xlock(&allproc_lock);
 641                 MPASS(swapped_cnt > 0);
 642                 swapped_cnt--;
 643                 if (curthread != &thread0)
 644                         swap_inprogress++;
 645                 sx_xunlock(&allproc_lock);
 646
 647                 /*
 648                  * We hold no lock here because the list of threads
 649                  * can not change while all threads in the process are
 650                  * swapped out.
 651                  */
 652                 FOREACH_THREAD_IN_PROC(p, td)
 653                         vm_thread_swapin(td, oom_alloc);
 654
 655                 if (curthread != &thread0) {
 656                         sx_xlock(&allproc_lock);
 657                         MPASS(swap_inprogress > 0);
 658                         swap_inprogress--;
 659                         last_swapin = ticks;
 660                         sx_xunlock(&allproc_lock);
 661                 }
 662                 PROC_LOCK(p);
 663                 swapclear(p);
 664                 p->p_swtick = ticks;
 665
 666                 /* Allow other threads to swap p out now. */
 667                 wakeup(&p->p_flag);
 668                 --p->p_lock;
 669         }
 670 }
 671
 672 /*
 673  * This swapin algorithm attempts to swap-in processes only if there
 674  * is enough space for them.  Of course, if a process waits for a long
 675  * time, it will be swapped in anyway.
 676  */
 677
 678 static struct proc *
 679 swapper_selector(bool wkilled_only)
 680 {
 681         struct proc *p, *res;
 682         struct thread *td;
 683         int ppri, pri, slptime, swtime;
 684
 685         sx_assert(&allproc_lock, SA_SLOCKED);
 686         if (swapped_cnt == 0)
 687                 return (NULL);
 688         res = NULL;
 689         ppri = INT_MIN;
 690         FOREACH_PROC_IN_SYSTEM(p) {
 691                 PROC_LOCK(p);
 692                 if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
 693                     P_SWAPPINGIN | P_INMEM)) != 0) {
 694                         PROC_UNLOCK(p);
 695                         continue;
 696                 }
 697                 if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
 698                         /*
 699                          * A swapped-out process might have mapped a
 700                          * large portion of the system's pages as
 701                          * anonymous memory.  There is no other way to
 702                          * release the memory other than to kill the
 703                          * process, for which we need to swap it in.
 704                          */
 705                         return (p);
 706                 }
 707                 if (wkilled_only) {
 708                         PROC_UNLOCK(p);
 709                         continue;
 710                 }
 711                 swtime = (ticks - p->p_swtick) / hz;
 712                 FOREACH_THREAD_IN_PROC(p, td) {
 713                         /*
 714                          * An otherwise runnable thread of a process
 715                          * swapped out has only the TDI_SWAPPED bit set.
 716                          */
 717                         thread_lock(td);
 718                         if (td->td_inhibitors == TDI_SWAPPED) {
 719                                 slptime = (ticks - td->td_slptick) / hz;
 720                                 pri = swtime + slptime;
 721                                 if ((td->td_flags & TDF_SWAPINREQ) == 0)
 722                                         pri -= p->p_nice * 8;
 723                                 /*
 724                                  * if this thread is higher priority
 725                                  * and there is enough space, then select
 726                                  * this process instead of the previous
 727                                  * selection.
 728                                  */
 729                                 if (pri > ppri) {
 730                                         res = p;
 731                                         ppri = pri;
 732                                 }
 733                         }
 734                         thread_unlock(td);
 735                 }
 736                 PROC_UNLOCK(p);
 737         }
 738
 739         if (res != NULL)
 740                 PROC_LOCK(res);
 741         return (res);
 742 }
 743
 744 #define SWAPIN_INTERVAL (MAXSLP * hz / 2)
 745
 746 /*
 747  * Limit swapper to swap in one non-WKILLED process in MAXSLP/2
 748  * interval, assuming that there is:
 749  * - at least one domain that is not suffering from a shortage of free memory;
 750  * - no parallel swap-ins;
 751  * - no other swap-ins in the current SWAPIN_INTERVAL.
 752  */
 753 static bool
 754 swapper_wkilled_only(void)
 755 {
 756
 757         return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 ||
 758             (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL);
 759 }
 760
 761 void
 762 swapper(void)
 763 {
 764         struct proc *p;
 765
 766         for (;;) {
 767                 sx_slock(&allproc_lock);
 768                 p = swapper_selector(swapper_wkilled_only());
 769                 sx_sunlock(&allproc_lock);
 770
 771                 if (p == NULL) {
 772                         tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL);
 773                 } else {
 774                         PROC_LOCK_ASSERT(p, MA_OWNED);
 775
 776                         /*
 777                          * Another process may be bringing or may have
 778                          * already brought this process in while we
 779                          * traverse all threads.  Or, this process may
 780                          * have exited or even being swapped out
 781                          * again.
 782                          */
 783                         if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
 784                             P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
 785                                 faultin(p);
 786                         }
 787                         PROC_UNLOCK(p);
 788                 }
 789         }
 790 }
 791
 792 /*
 793  * First, if any processes have been sleeping or stopped for at least
 794  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
 795  * no such processes exist, then the longest-sleeping or stopped
 796  * process is swapped out.  Finally, and only as a last resort, if
 797  * there are no sleeping or stopped processes, the longest-resident
 798  * process is swapped out.
 799  */
 800 static void
 801 swapout_procs(int action)
 802 {
 803         struct proc *p;
 804         struct thread *td;
 805         int slptime;
 806         bool didswap, doswap;
 807
 808         MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
 809
 810         didswap = false;
 811         sx_slock(&allproc_lock);
 812         FOREACH_PROC_IN_SYSTEM(p) {
 813                 /*
 814                  * Filter out not yet fully constructed processes.  Do
 815                  * not swap out held processes.  Avoid processes which
 816                  * are system, exiting, execing, traced, already swapped
 817                  * out or are in the process of being swapped in or out.
 818                  */
 819                 PROC_LOCK(p);
 820                 if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
 821                     (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
 822                     P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
 823                     P_INMEM) {
 824                         PROC_UNLOCK(p);
 825                         continue;
 826                 }
 827
 828                 /*
 829                  * Further consideration of this process for swap out
 830                  * requires iterating over its threads.  We release
 831                  * allproc_lock here so that process creation and
 832                  * destruction are not blocked while we iterate.
 833                  *
 834                  * To later reacquire allproc_lock and resume
 835                  * iteration over the allproc list, we will first have
 836                  * to release the lock on the process.  We place a
 837                  * hold on the process so that it remains in the
 838                  * allproc list while it is unlocked.
 839                  */
 840                 _PHOLD_LITE(p);
 841                 sx_sunlock(&allproc_lock);
 842
 843                 /*
 844                  * Do not swapout a realtime process.
 845                  * Guarantee swap_idle_threshold1 time in memory.
 846                  * If the system is under memory stress, or if we are
 847                  * swapping idle processes >= swap_idle_threshold2,
 848                  * then swap the process out.
 849                  */
 850                 doswap = true;
 851                 FOREACH_THREAD_IN_PROC(p, td) {
 852                         thread_lock(td);
 853                         slptime = (ticks - td->td_slptick) / hz;
 854                         if (PRI_IS_REALTIME(td->td_pri_class) ||
 855                             slptime < swap_idle_threshold1 ||
 856                             !thread_safetoswapout(td) ||
 857                             ((action & VM_SWAP_NORMAL) == 0 &&
 858                             slptime < swap_idle_threshold2))
 859                                 doswap = false;
 860                         thread_unlock(td);
 861                         if (!doswap)
 862                                 break;
 863                 }
 864                 if (doswap && swapout(p) == 0)
 865                         didswap = true;
 866
 867                 PROC_UNLOCK(p);
 868                 if (didswap) {
 869                         sx_xlock(&allproc_lock);
 870                         swapped_cnt++;
 871                         sx_downgrade(&allproc_lock);
 872                 } else
 873                         sx_slock(&allproc_lock);
 874                 PRELE(p);
 875         }
 876         sx_sunlock(&allproc_lock);
 877
 878         /*
 879          * If we swapped something out, and another process needed memory,
 880          * then wakeup the sched process.
 881          */
 882         if (didswap)
 883                 wakeup(&proc0);
 884 }
 885
 886 static void
 887 swapclear(struct proc *p)
 888 {
 889         struct thread *td;
 890
 891         PROC_LOCK_ASSERT(p, MA_OWNED);
 892
 893         FOREACH_THREAD_IN_PROC(p, td) {
 894                 thread_lock(td);
 895                 td->td_flags |= TDF_INMEM;
 896                 td->td_flags &= ~TDF_SWAPINREQ;
 897                 TD_CLR_SWAPPED(td);
 898                 if (TD_CAN_RUN(td))
 899                         if (setrunnable(td)) {
 900 #ifdef INVARIANTS
 901                                 /*
 902                                  * XXX: We just cleared TDI_SWAPPED
 903                                  * above and set TDF_INMEM, so this
 904                                  * should never happen.
 905                                  */
 906                                 panic("not waking up swapper");
 907 #endif
 908                         }
 909                 thread_unlock(td);
 910         }
 911         p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
 912         p->p_flag |= P_INMEM;
 913 }
 914
 915 static int
 916 swapout(struct proc *p)
 917 {
 918         struct thread *td;
 919
 920         PROC_LOCK_ASSERT(p, MA_OWNED);
 921
 922         /*
 923          * The states of this process and its threads may have changed
 924          * by now.  Assuming that there is only one pageout daemon thread,
 925          * this process should still be in memory.
 926          */
 927         KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
 928             P_INMEM, ("swapout: lost a swapout race?"));
 929
 930         /*
 931          * Remember the resident count.
 932          */
 933         p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
 934
 935         /*
 936          * Check and mark all threads before we proceed.
 937          */
 938         p->p_flag &= ~P_INMEM;
 939         p->p_flag |= P_SWAPPINGOUT;
 940         FOREACH_THREAD_IN_PROC(p, td) {
 941                 thread_lock(td);
 942                 if (!thread_safetoswapout(td)) {
 943                         thread_unlock(td);
 944                         swapclear(p);
 945                         return (EBUSY);
 946                 }
 947                 td->td_flags &= ~TDF_INMEM;
 948                 TD_SET_SWAPPED(td);
 949                 thread_unlock(td);
 950         }
 951         td = FIRST_THREAD_IN_PROC(p);
 952         ++td->td_ru.ru_nswap;
 953         PROC_UNLOCK(p);
 954
 955         /*
 956          * This list is stable because all threads are now prevented from
 957          * running.  The list is only modified in the context of a running
 958          * thread in this process.
 959          */
 960         FOREACH_THREAD_IN_PROC(p, td)
 961                 vm_thread_swapout(td);
 962
 963         PROC_LOCK(p);
 964         p->p_flag &= ~P_SWAPPINGOUT;
 965         p->p_swtick = ticks;
 966         return (0);
 967 }