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/sched.h>
  94 #include <sys/sdt.h>
  95 #include <sys/signalvar.h>
  96 #include <sys/smp.h>
  97 #include <sys/time.h>
  98 #include <sys/vnode.h>
  99 #include <sys/vmmeter.h>
 100 #include <sys/rwlock.h>
 101 #include <sys/sx.h>
 102 #include <sys/sysctl.h>
 103
 104 #include <vm/vm.h>
 105 #include <vm/vm_param.h>
 106 #include <vm/vm_object.h>
 107 #include <vm/vm_page.h>
 108 #include <vm/vm_map.h>
 109 #include <vm/vm_pageout.h>
 110 #include <vm/vm_pager.h>
 111 #include <vm/vm_phys.h>
 112 #include <vm/swap_pager.h>
 113 #include <vm/vm_extern.h>
 114 #include <vm/uma.h>
 115
 116 /* the kernel process "vm_daemon" */
 117 static void vm_daemon(void);
 118 static struct proc *vmproc;
 119
 120 static struct kproc_desc vm_kp = {
 121         "vmdaemon",
 122         vm_daemon,
 123         &vmproc
 124 };
 125 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
 126
 127 static int vm_swap_enabled = 1;
 128 static int vm_swap_idle_enabled = 0;
 129
 130 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW,
 131     &vm_swap_enabled, 0,
 132     "Enable entire process swapout");
 133 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW,
 134     &vm_swap_idle_enabled, 0,
 135     "Allow swapout on idle criteria");
 136
 137 /*
 138  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
 139  */
 140 static int swap_idle_threshold1 = 2;
 141 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
 142     &swap_idle_threshold1, 0,
 143     "Guaranteed swapped in time for a process");
 144
 145 /*
 146  * Swap_idle_threshold2 is the time that a process can be idle before
 147  * it will be swapped out, if idle swapping is enabled.
 148  */
 149 static int swap_idle_threshold2 = 10;
 150 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
 151     &swap_idle_threshold2, 0,
 152     "Time before a process will be swapped out");
 153
 154 static int vm_pageout_req_swapout;      /* XXX */
 155 static int vm_daemon_needed;
 156 static struct mtx vm_daemon_mtx;
 157 /* Allow for use by vm_pageout before vm_daemon is initialized. */
 158 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
 159
 160 static int swapped_cnt;
 161 static int swap_inprogress;     /* Pending swap-ins done outside swapper. */
 162 static int last_swapin;
 163
 164 static void swapclear(struct proc *);
 165 static int swapout(struct proc *);
 166 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
 167 static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long);
 168 static void swapout_procs(int action);
 169 static void vm_req_vmdaemon(int req);
 170 static void vm_thread_swapout(struct thread *td);
 171
 172 /*
 173  *      vm_swapout_object_deactivate_pages
 174  *
 175  *      Deactivate enough pages to satisfy the inactive target
 176  *      requirements.
 177  *
 178  *      The object and map must be locked.
 179  */
 180 static void
 181 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
 182     long desired)
 183 {
 184         vm_object_t backing_object, object;
 185         vm_page_t p;
 186         int act_delta, remove_mode;
 187
 188         VM_OBJECT_ASSERT_LOCKED(first_object);
 189         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 190                 return;
 191         for (object = first_object;; object = backing_object) {
 192                 if (pmap_resident_count(pmap) <= desired)
 193                         goto unlock_return;
 194                 VM_OBJECT_ASSERT_LOCKED(object);
 195                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 196                     object->paging_in_progress != 0)
 197                         goto unlock_return;
 198
 199                 remove_mode = 0;
 200                 if (object->shadow_count > 1)
 201                         remove_mode = 1;
 202                 /*
 203                  * Scan the object's entire memory queue.
 204                  */
 205                 TAILQ_FOREACH(p, &object->memq, listq) {
 206                         if (pmap_resident_count(pmap) <= desired)
 207                                 goto unlock_return;
 208                         if (should_yield())
 209                                 goto unlock_return;
 210                         if (vm_page_busied(p))
 211                                 continue;
 212                         VM_CNT_INC(v_pdpages);
 213                         vm_page_lock(p);
 214                         if (vm_page_wired(p) ||
 215                             !pmap_page_exists_quick(pmap, p)) {
 216                                 vm_page_unlock(p);
 217                                 continue;
 218                         }
 219                         act_delta = pmap_ts_referenced(p);
 220                         if ((p->aflags & PGA_REFERENCED) != 0) {
 221                                 if (act_delta == 0)
 222                                         act_delta = 1;
 223                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 224                         }
 225                         if (!vm_page_active(p) && act_delta != 0) {
 226                                 vm_page_activate(p);
 227                                 p->act_count += act_delta;
 228                         } else if (vm_page_active(p)) {
 229                                 /*
 230                                  * The page daemon does not requeue pages
 231                                  * after modifying their activation count.
 232                                  */
 233                                 if (act_delta == 0) {
 234                                         p->act_count -= min(p->act_count,
 235                                             ACT_DECLINE);
 236                                         if (!remove_mode && p->act_count == 0) {
 237                                                 pmap_remove_all(p);
 238                                                 vm_page_deactivate(p);
 239                                         }
 240                                 } else {
 241                                         vm_page_activate(p);
 242                                         if (p->act_count < ACT_MAX -
 243                                             ACT_ADVANCE)
 244                                                 p->act_count += ACT_ADVANCE;
 245                                 }
 246                         } else if (vm_page_inactive(p))
 247                                 pmap_remove_all(p);
 248                         vm_page_unlock(p);
 249                 }
 250                 if ((backing_object = object->backing_object) == NULL)
 251                         goto unlock_return;
 252                 VM_OBJECT_RLOCK(backing_object);
 253                 if (object != first_object)
 254                         VM_OBJECT_RUNLOCK(object);
 255         }
 256 unlock_return:
 257         if (object != first_object)
 258                 VM_OBJECT_RUNLOCK(object);
 259 }
 260
 261 /*
 262  * deactivate some number of pages in a map, try to do it fairly, but
 263  * that is really hard to do.
 264  */
 265 static void
 266 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
 267 {
 268         vm_map_entry_t tmpe;
 269         vm_object_t obj, bigobj;
 270         int nothingwired;
 271
 272         if (!vm_map_trylock_read(map))
 273                 return;
 274
 275         bigobj = NULL;
 276         nothingwired = TRUE;
 277
 278         /*
 279          * first, search out the biggest object, and try to free pages from
 280          * that.
 281          */
 282         tmpe = map->header.next;
 283         while (tmpe != &map->header) {
 284                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 285                         obj = tmpe->object.vm_object;
 286                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 287                                 if (obj->shadow_count <= 1 &&
 288                                     (bigobj == NULL ||
 289                                      bigobj->resident_page_count <
 290                                      obj->resident_page_count)) {
 291                                         if (bigobj != NULL)
 292                                                 VM_OBJECT_RUNLOCK(bigobj);
 293                                         bigobj = obj;
 294                                 } else
 295                                         VM_OBJECT_RUNLOCK(obj);
 296                         }
 297                 }
 298                 if (tmpe->wired_count > 0)
 299                         nothingwired = FALSE;
 300                 tmpe = tmpe->next;
 301         }
 302
 303         if (bigobj != NULL) {
 304                 vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
 305                 VM_OBJECT_RUNLOCK(bigobj);
 306         }
 307         /*
 308          * Next, hunt around for other pages to deactivate.  We actually
 309          * do this search sort of wrong -- .text first is not the best idea.
 310          */
 311         tmpe = map->header.next;
 312         while (tmpe != &map->header) {
 313                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 314                         break;
 315                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 316                         obj = tmpe->object.vm_object;
 317                         if (obj != NULL) {
 318                                 VM_OBJECT_RLOCK(obj);
 319                                 vm_swapout_object_deactivate_pages(map->pmap,
 320                                     obj, desired);
 321                                 VM_OBJECT_RUNLOCK(obj);
 322                         }
 323                 }
 324                 tmpe = tmpe->next;
 325         }
 326
 327         /*
 328          * Remove all mappings if a process is swapped out, this will free page
 329          * table pages.
 330          */
 331         if (desired == 0 && nothingwired) {
 332                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 333                     vm_map_max(map));
 334         }
 335
 336         vm_map_unlock_read(map);
 337 }
 338
 339 /*
 340  * Swap out requests
 341  */
 342 #define VM_SWAP_NORMAL 1
 343 #define VM_SWAP_IDLE 2
 344
 345 void
 346 vm_swapout_run(void)
 347 {
 348
 349         if (vm_swap_enabled)
 350                 vm_req_vmdaemon(VM_SWAP_NORMAL);
 351 }
 352
 353 /*
 354  * Idle process swapout -- run once per second when pagedaemons are
 355  * reclaiming pages.
 356  */
 357 void
 358 vm_swapout_run_idle(void)
 359 {
 360         static long lsec;
 361
 362         if (!vm_swap_idle_enabled || time_second == lsec)
 363                 return;
 364         vm_req_vmdaemon(VM_SWAP_IDLE);
 365         lsec = time_second;
 366 }
 367
 368 static void
 369 vm_req_vmdaemon(int req)
 370 {
 371         static int lastrun = 0;
 372
 373         mtx_lock(&vm_daemon_mtx);
 374         vm_pageout_req_swapout |= req;
 375         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
 376                 wakeup(&vm_daemon_needed);
 377                 lastrun = ticks;
 378         }
 379         mtx_unlock(&vm_daemon_mtx);
 380 }
 381
 382 static void
 383 vm_daemon(void)
 384 {
 385         struct rlimit rsslim;
 386         struct proc *p;
 387         struct thread *td;
 388         struct vmspace *vm;
 389         int breakout, swapout_flags, tryagain, attempts;
 390 #ifdef RACCT
 391         uint64_t rsize, ravailable;
 392 #endif
 393
 394         while (TRUE) {
 395                 mtx_lock(&vm_daemon_mtx);
 396                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
 397 #ifdef RACCT
 398                     racct_enable ? hz : 0
 399 #else
 400                     0
 401 #endif
 402                 );
 403                 swapout_flags = vm_pageout_req_swapout;
 404                 vm_pageout_req_swapout = 0;
 405                 mtx_unlock(&vm_daemon_mtx);
 406                 if (swapout_flags != 0) {
 407                         /*
 408                          * Drain the per-CPU page queue batches as a deadlock
 409                          * avoidance measure.
 410                          */
 411                         if ((swapout_flags & VM_SWAP_NORMAL) != 0)
 412                                 vm_page_drain_pqbatch();
 413                         swapout_procs(swapout_flags);
 414                 }
 415
 416                 /*
 417                  * scan the processes for exceeding their rlimits or if
 418                  * process is swapped out -- deactivate pages
 419                  */
 420                 tryagain = 0;
 421                 attempts = 0;
 422 again:
 423                 attempts++;
 424                 sx_slock(&allproc_lock);
 425                 FOREACH_PROC_IN_SYSTEM(p) {
 426                         vm_pindex_t limit, size;
 427
 428                         /*
 429                          * if this is a system process or if we have already
 430                          * looked at this process, skip it.
 431                          */
 432                         PROC_LOCK(p);
 433                         if (p->p_state != PRS_NORMAL ||
 434                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
 435                                 PROC_UNLOCK(p);
 436                                 continue;
 437                         }
 438                         /*
 439                          * if the process is in a non-running type state,
 440                          * don't touch it.
 441                          */
 442                         breakout = 0;
 443                         FOREACH_THREAD_IN_PROC(p, td) {
 444                                 thread_lock(td);
 445                                 if (!TD_ON_RUNQ(td) &&
 446                                     !TD_IS_RUNNING(td) &&
 447                                     !TD_IS_SLEEPING(td) &&
 448                                     !TD_IS_SUSPENDED(td)) {
 449                                         thread_unlock(td);
 450                                         breakout = 1;
 451                                         break;
 452                                 }
 453                                 thread_unlock(td);
 454                         }
 455                         if (breakout) {
 456                                 PROC_UNLOCK(p);
 457                                 continue;
 458                         }
 459                         /*
 460                          * get a limit
 461                          */
 462                         lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
 463                         limit = OFF_TO_IDX(
 464                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
 465
 466                         /*
 467                          * let processes that are swapped out really be
 468                          * swapped out set the limit to nothing (will force a
 469                          * swap-out.)
 470                          */
 471                         if ((p->p_flag & P_INMEM) == 0)
 472                                 limit = 0;      /* XXX */
 473                         vm = vmspace_acquire_ref(p);
 474                         _PHOLD_LITE(p);
 475                         PROC_UNLOCK(p);
 476                         if (vm == NULL) {
 477                                 PRELE(p);
 478                                 continue;
 479                         }
 480                         sx_sunlock(&allproc_lock);
 481
 482                         size = vmspace_resident_count(vm);
 483                         if (size >= limit) {
 484                                 vm_swapout_map_deactivate_pages(
 485                                     &vm->vm_map, limit);
 486                                 size = vmspace_resident_count(vm);
 487                         }
 488 #ifdef RACCT
 489                         if (racct_enable) {
 490                                 rsize = IDX_TO_OFF(size);
 491                                 PROC_LOCK(p);
 492                                 if (p->p_state == PRS_NORMAL)
 493                                         racct_set(p, RACCT_RSS, rsize);
 494                                 ravailable = racct_get_available(p, RACCT_RSS);
 495                                 PROC_UNLOCK(p);
 496                                 if (rsize > ravailable) {
 497                                         /*
 498                                          * Don't be overly aggressive; this
 499                                          * might be an innocent process,
 500                                          * and the limit could've been exceeded
 501                                          * by some memory hog.  Don't try
 502                                          * to deactivate more than 1/4th
 503                                          * of process' resident set size.
 504                                          */
 505                                         if (attempts <= 8) {
 506                                                 if (ravailable < rsize -
 507                                                     (rsize / 4)) {
 508                                                         ravailable = rsize -
 509                                                             (rsize / 4);
 510                                                 }
 511                                         }
 512                                         vm_swapout_map_deactivate_pages(
 513                                             &vm->vm_map,
 514                                             OFF_TO_IDX(ravailable));
 515                                         /* Update RSS usage after paging out. */
 516                                         size = vmspace_resident_count(vm);
 517                                         rsize = IDX_TO_OFF(size);
 518                                         PROC_LOCK(p);
 519                                         if (p->p_state == PRS_NORMAL)
 520                                                 racct_set(p, RACCT_RSS, rsize);
 521                                         PROC_UNLOCK(p);
 522                                         if (rsize > ravailable)
 523                                                 tryagain = 1;
 524                                 }
 525                         }
 526 #endif
 527                         vmspace_free(vm);
 528                         sx_slock(&allproc_lock);
 529                         PRELE(p);
 530                 }
 531                 sx_sunlock(&allproc_lock);
 532                 if (tryagain != 0 && attempts <= 10) {
 533                         maybe_yield();
 534                         goto again;
 535                 }
 536         }
 537 }
 538
 539 /*
 540  * Allow a thread's kernel stack to be paged out.
 541  */
 542 static void
 543 vm_thread_swapout(struct thread *td)
 544 {
 545         vm_object_t ksobj;
 546         vm_page_t m;
 547         int i, pages;
 548
 549         cpu_thread_swapout(td);
 550         pages = td->td_kstack_pages;
 551         ksobj = td->td_kstack_obj;
 552         pmap_qremove(td->td_kstack, pages);
 553         VM_OBJECT_WLOCK(ksobj);
 554         for (i = 0; i < pages; i++) {
 555                 m = vm_page_lookup(ksobj, i);
 556                 if (m == NULL)
 557                         panic("vm_thread_swapout: kstack already missing?");
 558                 vm_page_dirty(m);
 559                 vm_page_lock(m);
 560                 vm_page_unwire(m, PQ_LAUNDRY);
 561                 vm_page_unlock(m);
 562         }
 563         VM_OBJECT_WUNLOCK(ksobj);
 564 }
 565
 566 /*
 567  * Bring the kernel stack for a specified thread back in.
 568  */
 569 static void
 570 vm_thread_swapin(struct thread *td, int oom_alloc)
 571 {
 572         vm_object_t ksobj;
 573         vm_page_t ma[KSTACK_MAX_PAGES];
 574         int a, count, i, j, pages, rv;
 575
 576         pages = td->td_kstack_pages;
 577         ksobj = td->td_kstack_obj;
 578         VM_OBJECT_WLOCK(ksobj);
 579         (void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma,
 580             pages);
 581         for (i = 0; i < pages;) {
 582                 vm_page_assert_xbusied(ma[i]);
 583                 if (ma[i]->valid == VM_PAGE_BITS_ALL) {
 584                         vm_page_xunbusy(ma[i]);
 585                         i++;
 586                         continue;
 587                 }
 588                 vm_object_pip_add(ksobj, 1);
 589                 for (j = i + 1; j < pages; j++)
 590                         if (ma[j]->valid == VM_PAGE_BITS_ALL)
 591                                 break;
 592                 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
 593                 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
 594                 count = min(a + 1, j - i);
 595                 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
 596                 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
 597                     __func__, td->td_proc->p_pid));
 598                 vm_object_pip_wakeup(ksobj);
 599                 for (j = i; j < i + count; j++)
 600                         vm_page_xunbusy(ma[j]);
 601                 i += count;
 602         }
 603         VM_OBJECT_WUNLOCK(ksobj);
 604         pmap_qenter(td->td_kstack, ma, pages);
 605         cpu_thread_swapin(td);
 606 }
 607
 608 void
 609 faultin(struct proc *p)
 610 {
 611         struct thread *td;
 612         int oom_alloc;
 613
 614         PROC_LOCK_ASSERT(p, MA_OWNED);
 615
 616         /*
 617          * If another process is swapping in this process,
 618          * just wait until it finishes.
 619          */
 620         if (p->p_flag & P_SWAPPINGIN) {
 621                 while (p->p_flag & P_SWAPPINGIN)
 622                         msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
 623                 return;
 624         }
 625
 626         if ((p->p_flag & P_INMEM) == 0) {
 627                 oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
 628                     VM_ALLOC_NORMAL;
 629
 630                 /*
 631                  * Don't let another thread swap process p out while we are
 632                  * busy swapping it in.
 633                  */
 634                 ++p->p_lock;
 635                 p->p_flag |= P_SWAPPINGIN;
 636                 PROC_UNLOCK(p);
 637                 sx_xlock(&allproc_lock);
 638                 MPASS(swapped_cnt > 0);
 639                 swapped_cnt--;
 640                 if (curthread != &thread0)
 641                         swap_inprogress++;
 642                 sx_xunlock(&allproc_lock);
 643
 644                 /*
 645                  * We hold no lock here because the list of threads
 646                  * can not change while all threads in the process are
 647                  * swapped out.
 648                  */
 649                 FOREACH_THREAD_IN_PROC(p, td)
 650                         vm_thread_swapin(td, oom_alloc);
 651
 652                 if (curthread != &thread0) {
 653                         sx_xlock(&allproc_lock);
 654                         MPASS(swap_inprogress > 0);
 655                         swap_inprogress--;
 656                         last_swapin = ticks;
 657                         sx_xunlock(&allproc_lock);
 658                 }
 659                 PROC_LOCK(p);
 660                 swapclear(p);
 661                 p->p_swtick = ticks;
 662
 663                 /* Allow other threads to swap p out now. */
 664                 wakeup(&p->p_flag);
 665                 --p->p_lock;
 666         }
 667 }
 668
 669 /*
 670  * This swapin algorithm attempts to swap-in processes only if there
 671  * is enough space for them.  Of course, if a process waits for a long
 672  * time, it will be swapped in anyway.
 673  */
 674
 675 static struct proc *
 676 swapper_selector(bool wkilled_only)
 677 {
 678         struct proc *p, *res;
 679         struct thread *td;
 680         int ppri, pri, slptime, swtime;
 681
 682         sx_assert(&allproc_lock, SA_SLOCKED);
 683         if (swapped_cnt == 0)
 684                 return (NULL);
 685         res = NULL;
 686         ppri = INT_MIN;
 687         FOREACH_PROC_IN_SYSTEM(p) {
 688                 PROC_LOCK(p);
 689                 if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
 690                     P_SWAPPINGIN | P_INMEM)) != 0) {
 691                         PROC_UNLOCK(p);
 692                         continue;
 693                 }
 694                 if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
 695                         /*
 696                          * A swapped-out process might have mapped a
 697                          * large portion of the system's pages as
 698                          * anonymous memory.  There is no other way to
 699                          * release the memory other than to kill the
 700                          * process, for which we need to swap it in.
 701                          */
 702                         return (p);
 703                 }
 704                 if (wkilled_only) {
 705                         PROC_UNLOCK(p);
 706                         continue;
 707                 }
 708                 swtime = (ticks - p->p_swtick) / hz;
 709                 FOREACH_THREAD_IN_PROC(p, td) {
 710                         /*
 711                          * An otherwise runnable thread of a process
 712                          * swapped out has only the TDI_SWAPPED bit set.
 713                          */
 714                         thread_lock(td);
 715                         if (td->td_inhibitors == TDI_SWAPPED) {
 716                                 slptime = (ticks - td->td_slptick) / hz;
 717                                 pri = swtime + slptime;
 718                                 if ((td->td_flags & TDF_SWAPINREQ) == 0)
 719                                         pri -= p->p_nice * 8;
 720                                 /*
 721                                  * if this thread is higher priority
 722                                  * and there is enough space, then select
 723                                  * this process instead of the previous
 724                                  * selection.
 725                                  */
 726                                 if (pri > ppri) {
 727                                         res = p;
 728                                         ppri = pri;
 729                                 }
 730                         }
 731                         thread_unlock(td);
 732                 }
 733                 PROC_UNLOCK(p);
 734         }
 735
 736         if (res != NULL)
 737                 PROC_LOCK(res);
 738         return (res);
 739 }
 740
 741 #define SWAPIN_INTERVAL (MAXSLP * hz / 2)
 742
 743 /*
 744  * Limit swapper to swap in one non-WKILLED process in MAXSLP/2
 745  * interval, assuming that there is:
 746  * - at least one domain that is not suffering from a shortage of free memory;
 747  * - no parallel swap-ins;
 748  * - no other swap-ins in the current SWAPIN_INTERVAL.
 749  */
 750 static bool
 751 swapper_wkilled_only(void)
 752 {
 753
 754         return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 ||
 755             (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL);
 756 }
 757
 758 void
 759 swapper(void)
 760 {
 761         struct proc *p;
 762
 763         for (;;) {
 764                 sx_slock(&allproc_lock);
 765                 p = swapper_selector(swapper_wkilled_only());
 766                 sx_sunlock(&allproc_lock);
 767
 768                 if (p == NULL) {
 769                         tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL);
 770                 } else {
 771                         PROC_LOCK_ASSERT(p, MA_OWNED);
 772
 773                         /*
 774                          * Another process may be bringing or may have
 775                          * already brought this process in while we
 776                          * traverse all threads.  Or, this process may
 777                          * have exited or even being swapped out
 778                          * again.
 779                          */
 780                         if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
 781                             P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
 782                                 faultin(p);
 783                         }
 784                         PROC_UNLOCK(p);
 785                 }
 786         }
 787 }
 788
 789 /*
 790  * First, if any processes have been sleeping or stopped for at least
 791  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
 792  * no such processes exist, then the longest-sleeping or stopped
 793  * process is swapped out.  Finally, and only as a last resort, if
 794  * there are no sleeping or stopped processes, the longest-resident
 795  * process is swapped out.
 796  */
 797 static void
 798 swapout_procs(int action)
 799 {
 800         struct proc *p;
 801         struct thread *td;
 802         int slptime;
 803         bool didswap, doswap;
 804
 805         MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
 806
 807         didswap = false;
 808         sx_slock(&allproc_lock);
 809         FOREACH_PROC_IN_SYSTEM(p) {
 810                 /*
 811                  * Filter out not yet fully constructed processes.  Do
 812                  * not swap out held processes.  Avoid processes which
 813                  * are system, exiting, execing, traced, already swapped
 814                  * out or are in the process of being swapped in or out.
 815                  */
 816                 PROC_LOCK(p);
 817                 if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
 818                     (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
 819                     P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
 820                     P_INMEM) {
 821                         PROC_UNLOCK(p);
 822                         continue;
 823                 }
 824
 825                 /*
 826                  * Further consideration of this process for swap out
 827                  * requires iterating over its threads.  We release
 828                  * allproc_lock here so that process creation and
 829                  * destruction are not blocked while we iterate.
 830                  *
 831                  * To later reacquire allproc_lock and resume
 832                  * iteration over the allproc list, we will first have
 833                  * to release the lock on the process.  We place a
 834                  * hold on the process so that it remains in the
 835                  * allproc list while it is unlocked.
 836                  */
 837                 _PHOLD_LITE(p);
 838                 sx_sunlock(&allproc_lock);
 839
 840                 /*
 841                  * Do not swapout a realtime process.
 842                  * Guarantee swap_idle_threshold1 time in memory.
 843                  * If the system is under memory stress, or if we are
 844                  * swapping idle processes >= swap_idle_threshold2,
 845                  * then swap the process out.
 846                  */
 847                 doswap = true;
 848                 FOREACH_THREAD_IN_PROC(p, td) {
 849                         thread_lock(td);
 850                         slptime = (ticks - td->td_slptick) / hz;
 851                         if (PRI_IS_REALTIME(td->td_pri_class) ||
 852                             slptime < swap_idle_threshold1 ||
 853                             !thread_safetoswapout(td) ||
 854                             ((action & VM_SWAP_NORMAL) == 0 &&
 855                             slptime < swap_idle_threshold2))
 856                                 doswap = false;
 857                         thread_unlock(td);
 858                         if (!doswap)
 859                                 break;
 860                 }
 861                 if (doswap && swapout(p) == 0)
 862                         didswap = true;
 863
 864                 PROC_UNLOCK(p);
 865                 if (didswap) {
 866                         sx_xlock(&allproc_lock);
 867                         swapped_cnt++;
 868                         sx_downgrade(&allproc_lock);
 869                 } else
 870                         sx_slock(&allproc_lock);
 871                 PRELE(p);
 872         }
 873         sx_sunlock(&allproc_lock);
 874
 875         /*
 876          * If we swapped something out, and another process needed memory,
 877          * then wakeup the sched process.
 878          */
 879         if (didswap)
 880                 wakeup(&proc0);
 881 }
 882
 883 static void
 884 swapclear(struct proc *p)
 885 {
 886         struct thread *td;
 887
 888         PROC_LOCK_ASSERT(p, MA_OWNED);
 889
 890         FOREACH_THREAD_IN_PROC(p, td) {
 891                 thread_lock(td);
 892                 td->td_flags |= TDF_INMEM;
 893                 td->td_flags &= ~TDF_SWAPINREQ;
 894                 TD_CLR_SWAPPED(td);
 895                 if (TD_CAN_RUN(td))
 896                         if (setrunnable(td)) {
 897 #ifdef INVARIANTS
 898                                 /*
 899                                  * XXX: We just cleared TDI_SWAPPED
 900                                  * above and set TDF_INMEM, so this
 901                                  * should never happen.
 902                                  */
 903                                 panic("not waking up swapper");
 904 #endif
 905                         }
 906                 thread_unlock(td);
 907         }
 908         p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
 909         p->p_flag |= P_INMEM;
 910 }
 911
 912 static int
 913 swapout(struct proc *p)
 914 {
 915         struct thread *td;
 916
 917         PROC_LOCK_ASSERT(p, MA_OWNED);
 918
 919         /*
 920          * The states of this process and its threads may have changed
 921          * by now.  Assuming that there is only one pageout daemon thread,
 922          * this process should still be in memory.
 923          */
 924         KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
 925             P_INMEM, ("swapout: lost a swapout race?"));
 926
 927         /*
 928          * Remember the resident count.
 929          */
 930         p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
 931
 932         /*
 933          * Check and mark all threads before we proceed.
 934          */
 935         p->p_flag &= ~P_INMEM;
 936         p->p_flag |= P_SWAPPINGOUT;
 937         FOREACH_THREAD_IN_PROC(p, td) {
 938                 thread_lock(td);
 939                 if (!thread_safetoswapout(td)) {
 940                         thread_unlock(td);
 941                         swapclear(p);
 942                         return (EBUSY);
 943                 }
 944                 td->td_flags &= ~TDF_INMEM;
 945                 TD_SET_SWAPPED(td);
 946                 thread_unlock(td);
 947         }
 948         td = FIRST_THREAD_IN_PROC(p);
 949         ++td->td_ru.ru_nswap;
 950         PROC_UNLOCK(p);
 951
 952         /*
 953          * This list is stable because all threads are now prevented from
 954          * running.  The list is only modified in the context of a running
 955          * thread in this process.
 956          */
 957         FOREACH_THREAD_IN_PROC(p, td)
 958                 vm_thread_swapout(td);
 959
 960         PROC_LOCK(p);
 961         p->p_flag &= ~P_SWAPPINGOUT;
 962         p->p_swtick = ticks;
 963         return (0);
 964 }