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_kern.h>
 108 #include <vm/vm_object.h>
 109 #include <vm/vm_page.h>
 110 #include <vm/vm_map.h>
 111 #include <vm/vm_pageout.h>
 112 #include <vm/vm_pager.h>
 113 #include <vm/vm_phys.h>
 114 #include <vm/swap_pager.h>
 115 #include <vm/vm_extern.h>
 116 #include <vm/uma.h>
 117
 118 /* the kernel process "vm_daemon" */
 119 static void vm_daemon(void);
 120 static struct proc *vmproc;
 121
 122 static struct kproc_desc vm_kp = {
 123         "vmdaemon",
 124         vm_daemon,
 125         &vmproc
 126 };
 127 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
 128
 129 static int vm_swap_enabled = 1;
 130 static int vm_swap_idle_enabled = 0;
 131
 132 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW,
 133     &vm_swap_enabled, 0,
 134     "Enable entire process swapout");
 135 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW,
 136     &vm_swap_idle_enabled, 0,
 137     "Allow swapout on idle criteria");
 138
 139 /*
 140  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
 141  */
 142 static int swap_idle_threshold1 = 2;
 143 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
 144     &swap_idle_threshold1, 0,
 145     "Guaranteed swapped in time for a process");
 146
 147 /*
 148  * Swap_idle_threshold2 is the time that a process can be idle before
 149  * it will be swapped out, if idle swapping is enabled.
 150  */
 151 static int swap_idle_threshold2 = 10;
 152 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
 153     &swap_idle_threshold2, 0,
 154     "Time before a process will be swapped out");
 155
 156 static int vm_pageout_req_swapout;      /* XXX */
 157 static int vm_daemon_needed;
 158 static struct mtx vm_daemon_mtx;
 159 /* Allow for use by vm_pageout before vm_daemon is initialized. */
 160 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
 161
 162 static int swapped_cnt;
 163 static int swap_inprogress;     /* Pending swap-ins done outside swapper. */
 164 static int last_swapin;
 165
 166 static void swapclear(struct proc *);
 167 static int swapout(struct proc *);
 168 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
 169 static void vm_swapout_object_deactivate(pmap_t, vm_object_t, long);
 170 static void swapout_procs(int action);
 171 static void vm_req_vmdaemon(int req);
 172 static void vm_thread_swapout(struct thread *td);
 173
 174 static void
 175 vm_swapout_object_deactivate_page(pmap_t pmap, vm_page_t m, bool unmap)
 176 {
 177
 178         /*
 179          * Ignore unreclaimable wired pages.  Repeat the check after busying
 180          * since a busy holder may wire the page.
 181          */
 182         if (vm_page_wired(m) || !vm_page_tryxbusy(m))
 183                 return;
 184
 185         if (vm_page_wired(m) || !pmap_page_exists_quick(pmap, m)) {
 186                 vm_page_xunbusy(m);
 187                 return;
 188         }
 189         if (!pmap_is_referenced(m)) {
 190                 if (!vm_page_active(m))
 191                         (void)vm_page_try_remove_all(m);
 192                 else if (unmap && vm_page_try_remove_all(m))
 193                         vm_page_deactivate(m);
 194         }
 195         vm_page_xunbusy(m);
 196 }
 197
 198 /*
 199  *      vm_swapout_object_deactivate
 200  *
 201  *      Deactivate enough pages to satisfy the inactive target
 202  *      requirements.
 203  *
 204  *      The object and map must be locked.
 205  */
 206 static void
 207 vm_swapout_object_deactivate(pmap_t pmap, vm_object_t first_object,
 208     long desired)
 209 {
 210         vm_object_t backing_object, object;
 211         vm_page_t m;
 212         bool unmap;
 213
 214         VM_OBJECT_ASSERT_LOCKED(first_object);
 215         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 216                 return;
 217         for (object = first_object;; object = backing_object) {
 218                 if (pmap_resident_count(pmap) <= desired)
 219                         goto unlock_return;
 220                 VM_OBJECT_ASSERT_LOCKED(object);
 221                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 222                     blockcount_read(&object->paging_in_progress) > 0)
 223                         goto unlock_return;
 224
 225                 unmap = true;
 226                 if (object->shadow_count > 1)
 227                         unmap = false;
 228
 229                 /*
 230                  * Scan the object's entire memory queue.
 231                  */
 232                 TAILQ_FOREACH(m, &object->memq, listq) {
 233                         if (pmap_resident_count(pmap) <= desired)
 234                                 goto unlock_return;
 235                         if (should_yield())
 236                                 goto unlock_return;
 237                         vm_swapout_object_deactivate_page(pmap, m, unmap);
 238                 }
 239                 if ((backing_object = object->backing_object) == NULL)
 240                         goto unlock_return;
 241                 VM_OBJECT_RLOCK(backing_object);
 242                 if (object != first_object)
 243                         VM_OBJECT_RUNLOCK(object);
 244         }
 245 unlock_return:
 246         if (object != first_object)
 247                 VM_OBJECT_RUNLOCK(object);
 248 }
 249
 250 /*
 251  * deactivate some number of pages in a map, try to do it fairly, but
 252  * that is really hard to do.
 253  */
 254 static void
 255 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
 256 {
 257         vm_map_entry_t tmpe;
 258         vm_object_t obj, bigobj;
 259         int nothingwired;
 260
 261         if (!vm_map_trylock_read(map))
 262                 return;
 263
 264         bigobj = NULL;
 265         nothingwired = TRUE;
 266
 267         /*
 268          * first, search out the biggest object, and try to free pages from
 269          * that.
 270          */
 271         VM_MAP_ENTRY_FOREACH(tmpe, map) {
 272                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 273                         obj = tmpe->object.vm_object;
 274                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 275                                 if (obj->shadow_count <= 1 &&
 276                                     (bigobj == NULL ||
 277                                      bigobj->resident_page_count <
 278                                      obj->resident_page_count)) {
 279                                         if (bigobj != NULL)
 280                                                 VM_OBJECT_RUNLOCK(bigobj);
 281                                         bigobj = obj;
 282                                 } else
 283                                         VM_OBJECT_RUNLOCK(obj);
 284                         }
 285                 }
 286                 if (tmpe->wired_count > 0)
 287                         nothingwired = FALSE;
 288         }
 289
 290         if (bigobj != NULL) {
 291                 vm_swapout_object_deactivate(map->pmap, bigobj, desired);
 292                 VM_OBJECT_RUNLOCK(bigobj);
 293         }
 294         /*
 295          * Next, hunt around for other pages to deactivate.  We actually
 296          * do this search sort of wrong -- .text first is not the best idea.
 297          */
 298         VM_MAP_ENTRY_FOREACH(tmpe, map) {
 299                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 300                         break;
 301                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 302                         obj = tmpe->object.vm_object;
 303                         if (obj != NULL) {
 304                                 VM_OBJECT_RLOCK(obj);
 305                                 vm_swapout_object_deactivate(map->pmap, obj,
 306                                     desired);
 307                                 VM_OBJECT_RUNLOCK(obj);
 308                         }
 309                 }
 310         }
 311
 312         /*
 313          * Remove all mappings if a process is swapped out, this will free page
 314          * table pages.
 315          */
 316         if (desired == 0 && nothingwired) {
 317                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 318                     vm_map_max(map));
 319         }
 320
 321         vm_map_unlock_read(map);
 322 }
 323
 324 /*
 325  * Swap out requests
 326  */
 327 #define VM_SWAP_NORMAL 1
 328 #define VM_SWAP_IDLE 2
 329
 330 void
 331 vm_swapout_run(void)
 332 {
 333
 334         if (vm_swap_enabled)
 335                 vm_req_vmdaemon(VM_SWAP_NORMAL);
 336 }
 337
 338 /*
 339  * Idle process swapout -- run once per second when pagedaemons are
 340  * reclaiming pages.
 341  */
 342 void
 343 vm_swapout_run_idle(void)
 344 {
 345         static long lsec;
 346
 347         if (!vm_swap_idle_enabled || time_second == lsec)
 348                 return;
 349         vm_req_vmdaemon(VM_SWAP_IDLE);
 350         lsec = time_second;
 351 }
 352
 353 static void
 354 vm_req_vmdaemon(int req)
 355 {
 356         static int lastrun = 0;
 357
 358         mtx_lock(&vm_daemon_mtx);
 359         vm_pageout_req_swapout |= req;
 360         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
 361                 wakeup(&vm_daemon_needed);
 362                 lastrun = ticks;
 363         }
 364         mtx_unlock(&vm_daemon_mtx);
 365 }
 366
 367 static void
 368 vm_daemon(void)
 369 {
 370         struct rlimit rsslim;
 371         struct proc *p;
 372         struct thread *td;
 373         struct vmspace *vm;
 374         int breakout, swapout_flags, tryagain, attempts;
 375 #ifdef RACCT
 376         uint64_t rsize, ravailable;
 377 #endif
 378
 379         while (TRUE) {
 380                 mtx_lock(&vm_daemon_mtx);
 381                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
 382 #ifdef RACCT
 383                     racct_enable ? hz : 0
 384 #else
 385                     0
 386 #endif
 387                 );
 388                 swapout_flags = vm_pageout_req_swapout;
 389                 vm_pageout_req_swapout = 0;
 390                 mtx_unlock(&vm_daemon_mtx);
 391                 if (swapout_flags != 0) {
 392                         /*
 393                          * Drain the per-CPU page queue batches as a deadlock
 394                          * avoidance measure.
 395                          */
 396                         if ((swapout_flags & VM_SWAP_NORMAL) != 0)
 397                                 vm_page_pqbatch_drain();
 398                         swapout_procs(swapout_flags);
 399                 }
 400
 401                 /*
 402                  * scan the processes for exceeding their rlimits or if
 403                  * process is swapped out -- deactivate pages
 404                  */
 405                 tryagain = 0;
 406                 attempts = 0;
 407 again:
 408                 attempts++;
 409                 sx_slock(&allproc_lock);
 410                 FOREACH_PROC_IN_SYSTEM(p) {
 411                         vm_pindex_t limit, size;
 412
 413                         /*
 414                          * if this is a system process or if we have already
 415                          * looked at this process, skip it.
 416                          */
 417                         PROC_LOCK(p);
 418                         if (p->p_state != PRS_NORMAL ||
 419                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
 420                                 PROC_UNLOCK(p);
 421                                 continue;
 422                         }
 423                         /*
 424                          * if the process is in a non-running type state,
 425                          * don't touch it.
 426                          */
 427                         breakout = 0;
 428                         FOREACH_THREAD_IN_PROC(p, td) {
 429                                 thread_lock(td);
 430                                 if (!TD_ON_RUNQ(td) &&
 431                                     !TD_IS_RUNNING(td) &&
 432                                     !TD_IS_SLEEPING(td) &&
 433                                     !TD_IS_SUSPENDED(td)) {
 434                                         thread_unlock(td);
 435                                         breakout = 1;
 436                                         break;
 437                                 }
 438                                 thread_unlock(td);
 439                         }
 440                         if (breakout) {
 441                                 PROC_UNLOCK(p);
 442                                 continue;
 443                         }
 444                         /*
 445                          * get a limit
 446                          */
 447                         lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
 448                         limit = OFF_TO_IDX(
 449                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
 450
 451                         /*
 452                          * let processes that are swapped out really be
 453                          * swapped out set the limit to nothing (will force a
 454                          * swap-out.)
 455                          */
 456                         if ((p->p_flag & P_INMEM) == 0)
 457                                 limit = 0;      /* XXX */
 458                         vm = vmspace_acquire_ref(p);
 459                         _PHOLD_LITE(p);
 460                         PROC_UNLOCK(p);
 461                         if (vm == NULL) {
 462                                 PRELE(p);
 463                                 continue;
 464                         }
 465                         sx_sunlock(&allproc_lock);
 466
 467                         size = vmspace_resident_count(vm);
 468                         if (size >= limit) {
 469                                 vm_swapout_map_deactivate_pages(
 470                                     &vm->vm_map, limit);
 471                                 size = vmspace_resident_count(vm);
 472                         }
 473 #ifdef RACCT
 474                         if (racct_enable) {
 475                                 rsize = IDX_TO_OFF(size);
 476                                 PROC_LOCK(p);
 477                                 if (p->p_state == PRS_NORMAL)
 478                                         racct_set(p, RACCT_RSS, rsize);
 479                                 ravailable = racct_get_available(p, RACCT_RSS);
 480                                 PROC_UNLOCK(p);
 481                                 if (rsize > ravailable) {
 482                                         /*
 483                                          * Don't be overly aggressive; this
 484                                          * might be an innocent process,
 485                                          * and the limit could've been exceeded
 486                                          * by some memory hog.  Don't try
 487                                          * to deactivate more than 1/4th
 488                                          * of process' resident set size.
 489                                          */
 490                                         if (attempts <= 8) {
 491                                                 if (ravailable < rsize -
 492                                                     (rsize / 4)) {
 493                                                         ravailable = rsize -
 494                                                             (rsize / 4);
 495                                                 }
 496                                         }
 497                                         vm_swapout_map_deactivate_pages(
 498                                             &vm->vm_map,
 499                                             OFF_TO_IDX(ravailable));
 500                                         /* Update RSS usage after paging out. */
 501                                         size = vmspace_resident_count(vm);
 502                                         rsize = IDX_TO_OFF(size);
 503                                         PROC_LOCK(p);
 504                                         if (p->p_state == PRS_NORMAL)
 505                                                 racct_set(p, RACCT_RSS, rsize);
 506                                         PROC_UNLOCK(p);
 507                                         if (rsize > ravailable)
 508                                                 tryagain = 1;
 509                                 }
 510                         }
 511 #endif
 512                         vmspace_free(vm);
 513                         sx_slock(&allproc_lock);
 514                         PRELE(p);
 515                 }
 516                 sx_sunlock(&allproc_lock);
 517                 if (tryagain != 0 && attempts <= 10) {
 518                         maybe_yield();
 519                         goto again;
 520                 }
 521         }
 522 }
 523
 524 /*
 525  * Allow a thread's kernel stack to be paged out.
 526  */
 527 static void
 528 vm_thread_swapout(struct thread *td)
 529 {
 530         vm_page_t m;
 531         vm_offset_t kaddr;
 532         vm_pindex_t pindex;
 533         int i, pages;
 534
 535         cpu_thread_swapout(td);
 536         kaddr = td->td_kstack;
 537         pages = td->td_kstack_pages;
 538         pindex = atop(kaddr - VM_MIN_KERNEL_ADDRESS);
 539         pmap_qremove(kaddr, pages);
 540         VM_OBJECT_WLOCK(kstack_object);
 541         for (i = 0; i < pages; i++) {
 542                 m = vm_page_lookup(kstack_object, pindex + i);
 543                 if (m == NULL)
 544                         panic("vm_thread_swapout: kstack already missing?");
 545                 vm_page_dirty(m);
 546                 vm_page_xunbusy_unchecked(m);
 547                 vm_page_unwire(m, PQ_LAUNDRY);
 548         }
 549         VM_OBJECT_WUNLOCK(kstack_object);
 550 }
 551
 552 /*
 553  * Bring the kernel stack for a specified thread back in.
 554  */
 555 static void
 556 vm_thread_swapin(struct thread *td, int oom_alloc)
 557 {
 558         vm_page_t ma[KSTACK_MAX_PAGES];
 559         vm_offset_t kaddr;
 560         int a, count, i, j, pages, rv;
 561
 562         kaddr = td->td_kstack;
 563         pages = td->td_kstack_pages;
 564         vm_thread_stack_back(td->td_domain.dr_policy, kaddr, ma, pages,
 565             oom_alloc);
 566         for (i = 0; i < pages;) {
 567                 vm_page_assert_xbusied(ma[i]);
 568                 if (vm_page_all_valid(ma[i])) {
 569                         i++;
 570                         continue;
 571                 }
 572                 vm_object_pip_add(kstack_object, 1);
 573                 for (j = i + 1; j < pages; j++)
 574                         if (vm_page_all_valid(ma[j]))
 575                                 break;
 576                 VM_OBJECT_WLOCK(kstack_object);
 577                 rv = vm_pager_has_page(kstack_object, ma[i]->pindex, NULL, &a);
 578                 VM_OBJECT_WUNLOCK(kstack_object);
 579                 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
 580                 count = min(a + 1, j - i);
 581                 rv = vm_pager_get_pages(kstack_object, ma + i, count, NULL, NULL);
 582                 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
 583                     __func__, td->td_proc->p_pid));
 584                 vm_object_pip_wakeup(kstack_object);
 585                 i += count;
 586         }
 587         pmap_qenter(kaddr, ma, pages);
 588         cpu_thread_swapin(td);
 589 }
 590
 591 void
 592 faultin(struct proc *p)
 593 {
 594         struct thread *td;
 595         int oom_alloc;
 596
 597         PROC_LOCK_ASSERT(p, MA_OWNED);
 598
 599         /*
 600          * If another process is swapping in this process,
 601          * just wait until it finishes.
 602          */
 603         if (p->p_flag & P_SWAPPINGIN) {
 604                 while (p->p_flag & P_SWAPPINGIN)
 605                         msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
 606                 return;
 607         }
 608
 609         if ((p->p_flag & P_INMEM) == 0) {
 610                 oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
 611                     VM_ALLOC_NORMAL;
 612
 613                 /*
 614                  * Don't let another thread swap process p out while we are
 615                  * busy swapping it in.
 616                  */
 617                 ++p->p_lock;
 618                 p->p_flag |= P_SWAPPINGIN;
 619                 PROC_UNLOCK(p);
 620                 sx_xlock(&allproc_lock);
 621                 MPASS(swapped_cnt > 0);
 622                 swapped_cnt--;
 623                 if (curthread != &thread0)
 624                         swap_inprogress++;
 625                 sx_xunlock(&allproc_lock);
 626
 627                 /*
 628                  * We hold no lock here because the list of threads
 629                  * can not change while all threads in the process are
 630                  * swapped out.
 631                  */
 632                 FOREACH_THREAD_IN_PROC(p, td)
 633                         vm_thread_swapin(td, oom_alloc);
 634
 635                 if (curthread != &thread0) {
 636                         sx_xlock(&allproc_lock);
 637                         MPASS(swap_inprogress > 0);
 638                         swap_inprogress--;
 639                         last_swapin = ticks;
 640                         sx_xunlock(&allproc_lock);
 641                 }
 642                 PROC_LOCK(p);
 643                 swapclear(p);
 644                 p->p_swtick = ticks;
 645
 646                 /* Allow other threads to swap p out now. */
 647                 wakeup(&p->p_flag);
 648                 --p->p_lock;
 649         }
 650 }
 651
 652 /*
 653  * This swapin algorithm attempts to swap-in processes only if there
 654  * is enough space for them.  Of course, if a process waits for a long
 655  * time, it will be swapped in anyway.
 656  */
 657
 658 static struct proc *
 659 swapper_selector(bool wkilled_only)
 660 {
 661         struct proc *p, *res;
 662         struct thread *td;
 663         int ppri, pri, slptime, swtime;
 664
 665         sx_assert(&allproc_lock, SA_SLOCKED);
 666         if (swapped_cnt == 0)
 667                 return (NULL);
 668         res = NULL;
 669         ppri = INT_MIN;
 670         FOREACH_PROC_IN_SYSTEM(p) {
 671                 PROC_LOCK(p);
 672                 if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
 673                     P_SWAPPINGIN | P_INMEM)) != 0) {
 674                         PROC_UNLOCK(p);
 675                         continue;
 676                 }
 677                 if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
 678                         /*
 679                          * A swapped-out process might have mapped a
 680                          * large portion of the system's pages as
 681                          * anonymous memory.  There is no other way to
 682                          * release the memory other than to kill the
 683                          * process, for which we need to swap it in.
 684                          */
 685                         return (p);
 686                 }
 687                 if (wkilled_only) {
 688                         PROC_UNLOCK(p);
 689                         continue;
 690                 }
 691                 swtime = (ticks - p->p_swtick) / hz;
 692                 FOREACH_THREAD_IN_PROC(p, td) {
 693                         /*
 694                          * An otherwise runnable thread of a process
 695                          * swapped out has only the TDI_SWAPPED bit set.
 696                          */
 697                         thread_lock(td);
 698                         if (td->td_inhibitors == TDI_SWAPPED) {
 699                                 slptime = (ticks - td->td_slptick) / hz;
 700                                 pri = swtime + slptime;
 701                                 if ((td->td_flags & TDF_SWAPINREQ) == 0)
 702                                         pri -= p->p_nice * 8;
 703                                 /*
 704                                  * if this thread is higher priority
 705                                  * and there is enough space, then select
 706                                  * this process instead of the previous
 707                                  * selection.
 708                                  */
 709                                 if (pri > ppri) {
 710                                         res = p;
 711                                         ppri = pri;
 712                                 }
 713                         }
 714                         thread_unlock(td);
 715                 }
 716                 PROC_UNLOCK(p);
 717         }
 718
 719         if (res != NULL)
 720                 PROC_LOCK(res);
 721         return (res);
 722 }
 723
 724 #define SWAPIN_INTERVAL (MAXSLP * hz / 2)
 725
 726 /*
 727  * Limit swapper to swap in one non-WKILLED process in MAXSLP/2
 728  * interval, assuming that there is:
 729  * - at least one domain that is not suffering from a shortage of free memory;
 730  * - no parallel swap-ins;
 731  * - no other swap-ins in the current SWAPIN_INTERVAL.
 732  */
 733 static bool
 734 swapper_wkilled_only(void)
 735 {
 736
 737         return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 ||
 738             (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL);
 739 }
 740
 741 void
 742 swapper(void)
 743 {
 744         struct proc *p;
 745
 746         for (;;) {
 747                 sx_slock(&allproc_lock);
 748                 p = swapper_selector(swapper_wkilled_only());
 749                 sx_sunlock(&allproc_lock);
 750
 751                 if (p == NULL) {
 752                         tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL);
 753                 } else {
 754                         PROC_LOCK_ASSERT(p, MA_OWNED);
 755
 756                         /*
 757                          * Another process may be bringing or may have
 758                          * already brought this process in while we
 759                          * traverse all threads.  Or, this process may
 760                          * have exited or even being swapped out
 761                          * again.
 762                          */
 763                         if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
 764                             P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
 765                                 faultin(p);
 766                         }
 767                         PROC_UNLOCK(p);
 768                 }
 769         }
 770 }
 771
 772 /*
 773  * First, if any processes have been sleeping or stopped for at least
 774  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
 775  * no such processes exist, then the longest-sleeping or stopped
 776  * process is swapped out.  Finally, and only as a last resort, if
 777  * there are no sleeping or stopped processes, the longest-resident
 778  * process is swapped out.
 779  */
 780 static void
 781 swapout_procs(int action)
 782 {
 783         struct proc *p;
 784         struct thread *td;
 785         int slptime;
 786         bool didswap, doswap;
 787
 788         MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
 789
 790         didswap = false;
 791         sx_slock(&allproc_lock);
 792         FOREACH_PROC_IN_SYSTEM(p) {
 793                 /*
 794                  * Filter out not yet fully constructed processes.  Do
 795                  * not swap out held processes.  Avoid processes which
 796                  * are system, exiting, execing, traced, already swapped
 797                  * out or are in the process of being swapped in or out.
 798                  */
 799                 PROC_LOCK(p);
 800                 if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
 801                     (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
 802                     P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
 803                     P_INMEM) {
 804                         PROC_UNLOCK(p);
 805                         continue;
 806                 }
 807
 808                 /*
 809                  * Further consideration of this process for swap out
 810                  * requires iterating over its threads.  We release
 811                  * allproc_lock here so that process creation and
 812                  * destruction are not blocked while we iterate.
 813                  *
 814                  * To later reacquire allproc_lock and resume
 815                  * iteration over the allproc list, we will first have
 816                  * to release the lock on the process.  We place a
 817                  * hold on the process so that it remains in the
 818                  * allproc list while it is unlocked.
 819                  */
 820                 _PHOLD_LITE(p);
 821                 sx_sunlock(&allproc_lock);
 822
 823                 /*
 824                  * Do not swapout a realtime process.
 825                  * Guarantee swap_idle_threshold1 time in memory.
 826                  * If the system is under memory stress, or if we are
 827                  * swapping idle processes >= swap_idle_threshold2,
 828                  * then swap the process out.
 829                  */
 830                 doswap = true;
 831                 FOREACH_THREAD_IN_PROC(p, td) {
 832                         thread_lock(td);
 833                         slptime = (ticks - td->td_slptick) / hz;
 834                         if (PRI_IS_REALTIME(td->td_pri_class) ||
 835                             slptime < swap_idle_threshold1 ||
 836                             !thread_safetoswapout(td) ||
 837                             ((action & VM_SWAP_NORMAL) == 0 &&
 838                             slptime < swap_idle_threshold2))
 839                                 doswap = false;
 840                         thread_unlock(td);
 841                         if (!doswap)
 842                                 break;
 843                 }
 844                 if (doswap && swapout(p) == 0)
 845                         didswap = true;
 846
 847                 PROC_UNLOCK(p);
 848                 if (didswap) {
 849                         sx_xlock(&allproc_lock);
 850                         swapped_cnt++;
 851                         sx_downgrade(&allproc_lock);
 852                 } else
 853                         sx_slock(&allproc_lock);
 854                 PRELE(p);
 855         }
 856         sx_sunlock(&allproc_lock);
 857
 858         /*
 859          * If we swapped something out, and another process needed memory,
 860          * then wakeup the sched process.
 861          */
 862         if (didswap)
 863                 wakeup(&proc0);
 864 }
 865
 866 static void
 867 swapclear(struct proc *p)
 868 {
 869         struct thread *td;
 870
 871         PROC_LOCK_ASSERT(p, MA_OWNED);
 872
 873         FOREACH_THREAD_IN_PROC(p, td) {
 874                 thread_lock(td);
 875                 td->td_flags |= TDF_INMEM;
 876                 td->td_flags &= ~TDF_SWAPINREQ;
 877                 TD_CLR_SWAPPED(td);
 878                 if (TD_CAN_RUN(td)) {
 879                         if (setrunnable(td, 0)) {
 880 #ifdef INVARIANTS
 881                                 /*
 882                                  * XXX: We just cleared TDI_SWAPPED
 883                                  * above and set TDF_INMEM, so this
 884                                  * should never happen.
 885                                  */
 886                                 panic("not waking up swapper");
 887 #endif
 888                         }
 889                 } else
 890                         thread_unlock(td);
 891         }
 892         p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
 893         p->p_flag |= P_INMEM;
 894 }
 895
 896 static int
 897 swapout(struct proc *p)
 898 {
 899         struct thread *td;
 900
 901         PROC_LOCK_ASSERT(p, MA_OWNED);
 902
 903         /*
 904          * The states of this process and its threads may have changed
 905          * by now.  Assuming that there is only one pageout daemon thread,
 906          * this process should still be in memory.
 907          */
 908         KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
 909             P_INMEM, ("swapout: lost a swapout race?"));
 910
 911         /*
 912          * Remember the resident count.
 913          */
 914         p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
 915
 916         /*
 917          * Check and mark all threads before we proceed.
 918          */
 919         p->p_flag &= ~P_INMEM;
 920         p->p_flag |= P_SWAPPINGOUT;
 921         FOREACH_THREAD_IN_PROC(p, td) {
 922                 thread_lock(td);
 923                 if (!thread_safetoswapout(td)) {
 924                         thread_unlock(td);
 925                         swapclear(p);
 926                         return (EBUSY);
 927                 }
 928                 td->td_flags &= ~TDF_INMEM;
 929                 TD_SET_SWAPPED(td);
 930                 thread_unlock(td);
 931         }
 932         td = FIRST_THREAD_IN_PROC(p);
 933         ++td->td_ru.ru_nswap;
 934         PROC_UNLOCK(p);
 935
 936         /*
 937          * This list is stable because all threads are now prevented from
 938          * running.  The list is only modified in the context of a running
 939          * thread in this process.
 940          */
 941         FOREACH_THREAD_IN_PROC(p, td)
 942                 vm_thread_swapout(td);
 943
 944         PROC_LOCK(p);
 945         p->p_flag &= ~P_SWAPPINGOUT;
 946         p->p_swtick = ticks;
 947         return (0);
 948 }