sys/vm/vm_swapout.c

   1 /*-
   2  * Copyright (c) 1991 Regents of the University of California.
   3  * All rights reserved.
   4  * Copyright (c) 1994 John S. Dyson
   5  * All rights reserved.
   6  * Copyright (c) 1994 David Greenman
   7  * All rights reserved.
   8  * Copyright (c) 2005 Yahoo! Technologies Norway AS
   9  * All rights reserved.
  10  *
  11  * This code is derived from software contributed to Berkeley by
  12  * The Mach Operating System project at Carnegie-Mellon University.
  13  *
  14  * Redistribution and use in source and binary forms, with or without
  15  * modification, are permitted provided that the following conditions
  16  * are met:
  17  * 1. Redistributions of source code must retain the above copyright
  18  *    notice, this list of conditions and the following disclaimer.
  19  * 2. Redistributions in binary form must reproduce the above copyright
  20  *    notice, this list of conditions and the following disclaimer in the
  21  *    documentation and/or other materials provided with the distribution.
  22  * 3. All advertising materials mentioning features or use of this software
  23  *    must display the following acknowledgement:
  24  *      This product includes software developed by the University of
  25  *      California, Berkeley and its contributors.
  26  * 4. Neither the name of the University nor the names of its contributors
  27  *    may be used to endorse or promote products derived from this software
  28  *    without specific prior written permission.
  29  *
  30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  40  * SUCH DAMAGE.
  41  *
  42  *      from: @(#)vm_pageout.c  7.4 (Berkeley) 5/7/91
  43  *
  44  *
  45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  46  * All rights reserved.
  47  *
  48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  49  *
  50  * Permission to use, copy, modify and distribute this software and
  51  * its documentation is hereby granted, provided that both the copyright
  52  * notice and this permission notice appear in all copies of the
  53  * software, derivative works or modified versions, and any portions
  54  * thereof, and that both notices appear in supporting documentation.
  55  *
  56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  59  *
  60  * Carnegie Mellon requests users of this software to return to
  61  *
  62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  63  *  School of Computer Science
  64  *  Carnegie Mellon University
  65  *  Pittsburgh PA 15213-3890
  66  *
  67  * any improvements or extensions that they make and grant Carnegie the
  68  * rights to redistribute these changes.
  69  */
  70
  71 #include <sys/cdefs.h>
  72 __FBSDID("$FreeBSD$");
  73
  74 #include "opt_kstack_pages.h"
  75 #include "opt_kstack_max_pages.h"
  76 #include "opt_vm.h"
  77
  78 #include <sys/param.h>
  79 #include <sys/systm.h>
  80 #include <sys/limits.h>
  81 #include <sys/kernel.h>
  82 #include <sys/eventhandler.h>
  83 #include <sys/lock.h>
  84 #include <sys/mutex.h>
  85 #include <sys/proc.h>
  86 #include <sys/_kstack_cache.h>
  87 #include <sys/kthread.h>
  88 #include <sys/ktr.h>
  89 #include <sys/mount.h>
  90 #include <sys/racct.h>
  91 #include <sys/resourcevar.h>
  92 #include <sys/sched.h>
  93 #include <sys/sdt.h>
  94 #include <sys/signalvar.h>
  95 #include <sys/smp.h>
  96 #include <sys/time.h>
  97 #include <sys/vnode.h>
  98 #include <sys/vmmeter.h>
  99 #include <sys/rwlock.h>
 100 #include <sys/sx.h>
 101 #include <sys/sysctl.h>
 102
 103 #include <vm/vm.h>
 104 #include <vm/vm_param.h>
 105 #include <vm/vm_object.h>
 106 #include <vm/vm_page.h>
 107 #include <vm/vm_map.h>
 108 #include <vm/vm_pageout.h>
 109 #include <vm/vm_pager.h>
 110 #include <vm/vm_phys.h>
 111 #include <vm/swap_pager.h>
 112 #include <vm/vm_extern.h>
 113 #include <vm/uma.h>
 114
 115 /* the kernel process "vm_daemon" */
 116 static void vm_daemon(void);
 117 static struct proc *vmproc;
 118
 119 static struct kproc_desc vm_kp = {
 120         "vmdaemon",
 121         vm_daemon,
 122         &vmproc
 123 };
 124 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
 125
 126 static int vm_swap_enabled = 1;
 127 static int vm_swap_idle_enabled = 0;
 128
 129 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW,
 130     &vm_swap_enabled, 0,
 131     "Enable entire process swapout");
 132 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW,
 133     &vm_swap_idle_enabled, 0,
 134     "Allow swapout on idle criteria");
 135
 136 /*
 137  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
 138  */
 139 static int swap_idle_threshold1 = 2;
 140 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
 141     &swap_idle_threshold1, 0,
 142     "Guaranteed swapped in time for a process");
 143
 144 /*
 145  * Swap_idle_threshold2 is the time that a process can be idle before
 146  * it will be swapped out, if idle swapping is enabled.
 147  */
 148 static int swap_idle_threshold2 = 10;
 149 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
 150     &swap_idle_threshold2, 0,
 151     "Time before a process will be swapped out");
 152
 153 static int vm_pageout_req_swapout;      /* XXX */
 154 static int vm_daemon_needed;
 155 static struct mtx vm_daemon_mtx;
 156 /* Allow for use by vm_pageout before vm_daemon is initialized. */
 157 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
 158
 159 static int swapped_cnt;
 160 static int swap_inprogress;     /* Pending swap-ins done outside swapper. */
 161 static int last_swapin;
 162
 163 static void swapclear(struct proc *);
 164 static int swapout(struct proc *);
 165 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
 166 static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long);
 167 static void swapout_procs(int action);
 168 static void vm_req_vmdaemon(int req);
 169 static void vm_thread_swapout(struct thread *td);
 170
 171 /*
 172  *      vm_swapout_object_deactivate_pages
 173  *
 174  *      Deactivate enough pages to satisfy the inactive target
 175  *      requirements.
 176  *
 177  *      The object and map must be locked.
 178  */
 179 static void
 180 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
 181     long desired)
 182 {
 183         vm_object_t backing_object, object;
 184         vm_page_t p;
 185         int act_delta, remove_mode;
 186
 187         VM_OBJECT_ASSERT_LOCKED(first_object);
 188         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 189                 return;
 190         for (object = first_object;; object = backing_object) {
 191                 if (pmap_resident_count(pmap) <= desired)
 192                         goto unlock_return;
 193                 VM_OBJECT_ASSERT_LOCKED(object);
 194                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 195                     object->paging_in_progress != 0)
 196                         goto unlock_return;
 197
 198                 remove_mode = 0;
 199                 if (object->shadow_count > 1)
 200                         remove_mode = 1;
 201                 /*
 202                  * Scan the object's entire memory queue.
 203                  */
 204                 TAILQ_FOREACH(p, &object->memq, listq) {
 205                         if (pmap_resident_count(pmap) <= desired)
 206                                 goto unlock_return;
 207                         if (should_yield())
 208                                 goto unlock_return;
 209                         if (vm_page_busied(p))
 210                                 continue;
 211                         PCPU_INC(cnt.v_pdpages);
 212                         vm_page_lock(p);
 213                         if (p->wire_count != 0 || p->hold_count != 0 ||
 214                             !pmap_page_exists_quick(pmap, p)) {
 215                                 vm_page_unlock(p);
 216                                 continue;
 217                         }
 218                         act_delta = pmap_ts_referenced(p);
 219                         if ((p->aflags & PGA_REFERENCED) != 0) {
 220                                 if (act_delta == 0)
 221                                         act_delta = 1;
 222                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 223                         }
 224                         if (!vm_page_active(p) && act_delta != 0) {
 225                                 vm_page_activate(p);
 226                                 p->act_count += act_delta;
 227                         } else if (vm_page_active(p)) {
 228                                 if (act_delta == 0) {
 229                                         p->act_count -= min(p->act_count,
 230                                             ACT_DECLINE);
 231                                         if (!remove_mode && p->act_count == 0) {
 232                                                 pmap_remove_all(p);
 233                                                 vm_page_deactivate(p);
 234                                         } else
 235                                                 vm_page_requeue(p);
 236                                 } else {
 237                                         vm_page_activate(p);
 238                                         if (p->act_count < ACT_MAX -
 239                                             ACT_ADVANCE)
 240                                                 p->act_count += ACT_ADVANCE;
 241                                         vm_page_requeue(p);
 242                                 }
 243                         } else if (vm_page_inactive(p))
 244                                 pmap_remove_all(p);
 245                         vm_page_unlock(p);
 246                 }
 247                 if ((backing_object = object->backing_object) == NULL)
 248                         goto unlock_return;
 249                 VM_OBJECT_RLOCK(backing_object);
 250                 if (object != first_object)
 251                         VM_OBJECT_RUNLOCK(object);
 252         }
 253 unlock_return:
 254         if (object != first_object)
 255                 VM_OBJECT_RUNLOCK(object);
 256 }
 257
 258 /*
 259  * deactivate some number of pages in a map, try to do it fairly, but
 260  * that is really hard to do.
 261  */
 262 static void
 263 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
 264 {
 265         vm_map_entry_t tmpe;
 266         vm_object_t obj, bigobj;
 267         int nothingwired;
 268
 269         if (!vm_map_trylock_read(map))
 270                 return;
 271
 272         bigobj = NULL;
 273         nothingwired = TRUE;
 274
 275         /*
 276          * first, search out the biggest object, and try to free pages from
 277          * that.
 278          */
 279         tmpe = map->header.next;
 280         while (tmpe != &map->header) {
 281                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 282                         obj = tmpe->object.vm_object;
 283                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 284                                 if (obj->shadow_count <= 1 &&
 285                                     (bigobj == NULL ||
 286                                      bigobj->resident_page_count <
 287                                      obj->resident_page_count)) {
 288                                         if (bigobj != NULL)
 289                                                 VM_OBJECT_RUNLOCK(bigobj);
 290                                         bigobj = obj;
 291                                 } else
 292                                         VM_OBJECT_RUNLOCK(obj);
 293                         }
 294                 }
 295                 if (tmpe->wired_count > 0)
 296                         nothingwired = FALSE;
 297                 tmpe = tmpe->next;
 298         }
 299
 300         if (bigobj != NULL) {
 301                 vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
 302                 VM_OBJECT_RUNLOCK(bigobj);
 303         }
 304         /*
 305          * Next, hunt around for other pages to deactivate.  We actually
 306          * do this search sort of wrong -- .text first is not the best idea.
 307          */
 308         tmpe = map->header.next;
 309         while (tmpe != &map->header) {
 310                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 311                         break;
 312                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 313                         obj = tmpe->object.vm_object;
 314                         if (obj != NULL) {
 315                                 VM_OBJECT_RLOCK(obj);
 316                                 vm_swapout_object_deactivate_pages(map->pmap,
 317                                     obj, desired);
 318                                 VM_OBJECT_RUNLOCK(obj);
 319                         }
 320                 }
 321                 tmpe = tmpe->next;
 322         }
 323
 324         /*
 325          * Remove all mappings if a process is swapped out, this will free page
 326          * table pages.
 327          */
 328         if (desired == 0 && nothingwired) {
 329                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 330                     vm_map_max(map));
 331         }
 332
 333         vm_map_unlock_read(map);
 334 }
 335
 336 /*
 337  * Swap out requests
 338  */
 339 #define VM_SWAP_NORMAL 1
 340 #define VM_SWAP_IDLE 2
 341
 342 void
 343 vm_swapout_run(void)
 344 {
 345
 346         if (vm_swap_enabled)
 347                 vm_req_vmdaemon(VM_SWAP_NORMAL);
 348 }
 349
 350 /*
 351  * Idle process swapout -- run once per second when pagedaemons are
 352  * reclaiming pages.
 353  */
 354 void
 355 vm_swapout_run_idle(void)
 356 {
 357         static long lsec;
 358
 359         if (!vm_swap_idle_enabled || time_second == lsec)
 360                 return;
 361         vm_req_vmdaemon(VM_SWAP_IDLE);
 362         lsec = time_second;
 363 }
 364
 365 static void
 366 vm_req_vmdaemon(int req)
 367 {
 368         static int lastrun = 0;
 369
 370         mtx_lock(&vm_daemon_mtx);
 371         vm_pageout_req_swapout |= req;
 372         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
 373                 wakeup(&vm_daemon_needed);
 374                 lastrun = ticks;
 375         }
 376         mtx_unlock(&vm_daemon_mtx);
 377 }
 378
 379 static void
 380 vm_daemon(void)
 381 {
 382         struct rlimit rsslim;
 383         struct proc *p;
 384         struct thread *td;
 385         struct vmspace *vm;
 386         int breakout, swapout_flags, tryagain, attempts;
 387 #ifdef RACCT
 388         uint64_t rsize, ravailable;
 389 #endif
 390
 391         while (TRUE) {
 392                 mtx_lock(&vm_daemon_mtx);
 393                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
 394 #ifdef RACCT
 395                     racct_enable ? hz : 0
 396 #else
 397                     0
 398 #endif
 399                 );
 400                 swapout_flags = vm_pageout_req_swapout;
 401                 vm_pageout_req_swapout = 0;
 402                 mtx_unlock(&vm_daemon_mtx);
 403                 if (swapout_flags)
 404                         swapout_procs(swapout_flags);
 405
 406                 /*
 407                  * scan the processes for exceeding their rlimits or if
 408                  * process is swapped out -- deactivate pages
 409                  */
 410                 tryagain = 0;
 411                 attempts = 0;
 412 again:
 413                 attempts++;
 414                 sx_slock(&allproc_lock);
 415                 FOREACH_PROC_IN_SYSTEM(p) {
 416                         vm_pindex_t limit, size;
 417
 418                         /*
 419                          * if this is a system process or if we have already
 420                          * looked at this process, skip it.
 421                          */
 422                         PROC_LOCK(p);
 423                         if (p->p_state != PRS_NORMAL ||
 424                             p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
 425                                 PROC_UNLOCK(p);
 426                                 continue;
 427                         }
 428                         /*
 429                          * if the process is in a non-running type state,
 430                          * don't touch it.
 431                          */
 432                         breakout = 0;
 433                         FOREACH_THREAD_IN_PROC(p, td) {
 434                                 thread_lock(td);
 435                                 if (!TD_ON_RUNQ(td) &&
 436                                     !TD_IS_RUNNING(td) &&
 437                                     !TD_IS_SLEEPING(td) &&
 438                                     !TD_IS_SUSPENDED(td)) {
 439                                         thread_unlock(td);
 440                                         breakout = 1;
 441                                         break;
 442                                 }
 443                                 thread_unlock(td);
 444                         }
 445                         if (breakout) {
 446                                 PROC_UNLOCK(p);
 447                                 continue;
 448                         }
 449                         /*
 450                          * get a limit
 451                          */
 452                         lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
 453                         limit = OFF_TO_IDX(
 454                             qmin(rsslim.rlim_cur, rsslim.rlim_max));
 455
 456                         /*
 457                          * let processes that are swapped out really be
 458                          * swapped out set the limit to nothing (will force a
 459                          * swap-out.)
 460                          */
 461                         if ((p->p_flag & P_INMEM) == 0)
 462                                 limit = 0;      /* XXX */
 463                         vm = vmspace_acquire_ref(p);
 464                         _PHOLD_LITE(p);
 465                         PROC_UNLOCK(p);
 466                         if (vm == NULL) {
 467                                 PRELE(p);
 468                                 continue;
 469                         }
 470                         sx_sunlock(&allproc_lock);
 471
 472                         size = vmspace_resident_count(vm);
 473                         if (size >= limit) {
 474                                 vm_swapout_map_deactivate_pages(
 475                                     &vm->vm_map, limit);
 476                                 size = vmspace_resident_count(vm);
 477                         }
 478 #ifdef RACCT
 479                         if (racct_enable) {
 480                                 rsize = IDX_TO_OFF(size);
 481                                 PROC_LOCK(p);
 482                                 if (p->p_state == PRS_NORMAL)
 483                                         racct_set(p, RACCT_RSS, rsize);
 484                                 ravailable = racct_get_available(p, RACCT_RSS);
 485                                 PROC_UNLOCK(p);
 486                                 if (rsize > ravailable) {
 487                                         /*
 488                                          * Don't be overly aggressive; this
 489                                          * might be an innocent process,
 490                                          * and the limit could've been exceeded
 491                                          * by some memory hog.  Don't try
 492                                          * to deactivate more than 1/4th
 493                                          * of process' resident set size.
 494                                          */
 495                                         if (attempts <= 8) {
 496                                                 if (ravailable < rsize -
 497                                                     (rsize / 4)) {
 498                                                         ravailable = rsize -
 499                                                             (rsize / 4);
 500                                                 }
 501                                         }
 502                                         vm_swapout_map_deactivate_pages(
 503                                             &vm->vm_map,
 504                                             OFF_TO_IDX(ravailable));
 505                                         /* Update RSS usage after paging out. */
 506                                         size = vmspace_resident_count(vm);
 507                                         rsize = IDX_TO_OFF(size);
 508                                         PROC_LOCK(p);
 509                                         if (p->p_state == PRS_NORMAL)
 510                                                 racct_set(p, RACCT_RSS, rsize);
 511                                         PROC_UNLOCK(p);
 512                                         if (rsize > ravailable)
 513                                                 tryagain = 1;
 514                                 }
 515                         }
 516 #endif
 517                         vmspace_free(vm);
 518                         sx_slock(&allproc_lock);
 519                         PRELE(p);
 520                 }
 521                 sx_sunlock(&allproc_lock);
 522                 if (tryagain != 0 && attempts <= 10) {
 523                         maybe_yield();
 524                         goto again;
 525                 }
 526         }
 527 }
 528
 529 /*
 530  * Allow a thread's kernel stack to be paged out.
 531  */
 532 static void
 533 vm_thread_swapout(struct thread *td)
 534 {
 535         vm_object_t ksobj;
 536         vm_page_t m;
 537         int i, pages;
 538
 539         cpu_thread_swapout(td);
 540         pages = td->td_kstack_pages;
 541         ksobj = td->td_kstack_obj;
 542         pmap_qremove(td->td_kstack, pages);
 543         VM_OBJECT_WLOCK(ksobj);
 544         for (i = 0; i < pages; i++) {
 545                 m = vm_page_lookup(ksobj, i);
 546                 if (m == NULL)
 547                         panic("vm_thread_swapout: kstack already missing?");
 548                 vm_page_dirty(m);
 549                 vm_page_lock(m);
 550                 vm_page_unwire(m, PQ_INACTIVE);
 551                 vm_page_unlock(m);
 552         }
 553         VM_OBJECT_WUNLOCK(ksobj);
 554 }
 555
 556 /*
 557  * Bring the kernel stack for a specified thread back in.
 558  */
 559 static void
 560 vm_thread_swapin(struct thread *td, int oom_alloc)
 561 {
 562         vm_object_t ksobj;
 563         vm_page_t ma[KSTACK_MAX_PAGES];
 564         int a, count, i, j, pages, rv;
 565
 566         pages = td->td_kstack_pages;
 567         ksobj = td->td_kstack_obj;
 568         VM_OBJECT_WLOCK(ksobj);
 569         (void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma,
 570             pages);
 571         for (i = 0; i < pages;) {
 572                 vm_page_assert_xbusied(ma[i]);
 573                 if (ma[i]->valid == VM_PAGE_BITS_ALL) {
 574                         vm_page_xunbusy(ma[i]);
 575                         i++;
 576                         continue;
 577                 }
 578                 vm_object_pip_add(ksobj, 1);
 579                 for (j = i + 1; j < pages; j++)
 580                         if (ma[j]->valid == VM_PAGE_BITS_ALL)
 581                                 break;
 582                 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
 583                 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
 584                 count = min(a + 1, j - i);
 585                 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
 586                 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
 587                     __func__, td->td_proc->p_pid));
 588                 vm_object_pip_wakeup(ksobj);
 589                 for (j = i; j < i + count; j++)
 590                         vm_page_xunbusy(ma[j]);
 591                 i += count;
 592         }
 593         VM_OBJECT_WUNLOCK(ksobj);
 594         pmap_qenter(td->td_kstack, ma, pages);
 595         cpu_thread_swapin(td);
 596 }
 597
 598 void
 599 faultin(struct proc *p)
 600 {
 601         struct thread *td;
 602         int oom_alloc;
 603
 604         PROC_LOCK_ASSERT(p, MA_OWNED);
 605
 606         /*
 607          * If another process is swapping in this process,
 608          * just wait until it finishes.
 609          */
 610         if (p->p_flag & P_SWAPPINGIN) {
 611                 while (p->p_flag & P_SWAPPINGIN)
 612                         msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
 613                 return;
 614         }
 615
 616         if ((p->p_flag & P_INMEM) == 0) {
 617                 oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
 618                     VM_ALLOC_NORMAL;
 619
 620                 /*
 621                  * Don't let another thread swap process p out while we are
 622                  * busy swapping it in.
 623                  */
 624                 ++p->p_lock;
 625                 p->p_flag |= P_SWAPPINGIN;
 626                 PROC_UNLOCK(p);
 627                 sx_xlock(&allproc_lock);
 628                 MPASS(swapped_cnt > 0);
 629                 swapped_cnt--;
 630                 if (curthread != &thread0)
 631                         swap_inprogress++;
 632                 sx_xunlock(&allproc_lock);
 633
 634                 /*
 635                  * We hold no lock here because the list of threads
 636                  * can not change while all threads in the process are
 637                  * swapped out.
 638                  */
 639                 FOREACH_THREAD_IN_PROC(p, td)
 640                         vm_thread_swapin(td, oom_alloc);
 641
 642                 if (curthread != &thread0) {
 643                         sx_xlock(&allproc_lock);
 644                         MPASS(swap_inprogress > 0);
 645                         swap_inprogress--;
 646                         last_swapin = ticks;
 647                         sx_xunlock(&allproc_lock);
 648                 }
 649                 PROC_LOCK(p);
 650                 swapclear(p);
 651                 p->p_swtick = ticks;
 652
 653                 /* Allow other threads to swap p out now. */
 654                 wakeup(&p->p_flag);
 655                 --p->p_lock;
 656         }
 657 }
 658
 659 /*
 660  * This swapin algorithm attempts to swap-in processes only if there
 661  * is enough space for them.  Of course, if a process waits for a long
 662  * time, it will be swapped in anyway.
 663  */
 664
 665 static struct proc *
 666 swapper_selector(bool wkilled_only)
 667 {
 668         struct proc *p, *res;
 669         struct thread *td;
 670         int ppri, pri, slptime, swtime;
 671
 672         sx_assert(&allproc_lock, SA_SLOCKED);
 673         if (swapped_cnt == 0)
 674                 return (NULL);
 675         res = NULL;
 676         ppri = INT_MIN;
 677         FOREACH_PROC_IN_SYSTEM(p) {
 678                 PROC_LOCK(p);
 679                 if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
 680                     P_SWAPPINGIN | P_INMEM)) != 0) {
 681                         PROC_UNLOCK(p);
 682                         continue;
 683                 }
 684                 if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
 685                         /*
 686                          * A swapped-out process might have mapped a
 687                          * large portion of the system's pages as
 688                          * anonymous memory.  There is no other way to
 689                          * release the memory other than to kill the
 690                          * process, for which we need to swap it in.
 691                          */
 692                         return (p);
 693                 }
 694                 if (wkilled_only) {
 695                         PROC_UNLOCK(p);
 696                         continue;
 697                 }
 698                 swtime = (ticks - p->p_swtick) / hz;
 699                 FOREACH_THREAD_IN_PROC(p, td) {
 700                         /*
 701                          * An otherwise runnable thread of a process
 702                          * swapped out has only the TDI_SWAPPED bit set.
 703                          */
 704                         thread_lock(td);
 705                         if (td->td_inhibitors == TDI_SWAPPED) {
 706                                 slptime = (ticks - td->td_slptick) / hz;
 707                                 pri = swtime + slptime;
 708                                 if ((td->td_flags & TDF_SWAPINREQ) == 0)
 709                                         pri -= p->p_nice * 8;
 710                                 /*
 711                                  * if this thread is higher priority
 712                                  * and there is enough space, then select
 713                                  * this process instead of the previous
 714                                  * selection.
 715                                  */
 716                                 if (pri > ppri) {
 717                                         res = p;
 718                                         ppri = pri;
 719                                 }
 720                         }
 721                         thread_unlock(td);
 722                 }
 723                 PROC_UNLOCK(p);
 724         }
 725
 726         if (res != NULL)
 727                 PROC_LOCK(res);
 728         return (res);
 729 }
 730
 731 #define SWAPIN_INTERVAL (MAXSLP * hz / 2)
 732
 733 /*
 734  * Limit swapper to swap in one non-WKILLED process in MAXSLP/2
 735  * interval, assuming that there is:
 736  * - no memory shortage;
 737  * - no parallel swap-ins;
 738  * - no other swap-ins in the current SWAPIN_INTERVAL.
 739  */
 740 static bool
 741 swapper_wkilled_only(void)
 742 {
 743
 744         return (vm_page_count_min() || swap_inprogress > 0 ||
 745             (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL);
 746 }
 747
 748 void
 749 swapper(void)
 750 {
 751         struct proc *p;
 752
 753         for (;;) {
 754                 sx_slock(&allproc_lock);
 755                 p = swapper_selector(swapper_wkilled_only());
 756                 sx_sunlock(&allproc_lock);
 757
 758                 if (p == NULL) {
 759                         tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL);
 760                 } else {
 761                         PROC_LOCK_ASSERT(p, MA_OWNED);
 762
 763                         /*
 764                          * Another process may be bringing or may have
 765                          * already brought this process in while we
 766                          * traverse all threads.  Or, this process may
 767                          * have exited or even being swapped out
 768                          * again.
 769                          */
 770                         if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
 771                             P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
 772                                 faultin(p);
 773                         }
 774                         PROC_UNLOCK(p);
 775                 }
 776         }
 777 }
 778
 779 /*
 780  * First, if any processes have been sleeping or stopped for at least
 781  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
 782  * no such processes exist, then the longest-sleeping or stopped
 783  * process is swapped out.  Finally, and only as a last resort, if
 784  * there are no sleeping or stopped processes, the longest-resident
 785  * process is swapped out.
 786  */
 787 static void
 788 swapout_procs(int action)
 789 {
 790         struct proc *p;
 791         struct thread *td;
 792         int slptime;
 793         bool didswap, doswap;
 794
 795         MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
 796
 797         didswap = false;
 798         sx_slock(&allproc_lock);
 799         FOREACH_PROC_IN_SYSTEM(p) {
 800                 /*
 801                  * Filter out not yet fully constructed processes.  Do
 802                  * not swap out held processes.  Avoid processes which
 803                  * are system, exiting, execing, traced, already swapped
 804                  * out or are in the process of being swapped in or out.
 805                  */
 806                 PROC_LOCK(p);
 807                 if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
 808                     (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
 809                     P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
 810                     P_INMEM) {
 811                         PROC_UNLOCK(p);
 812                         continue;
 813                 }
 814
 815                 /*
 816                  * Further consideration of this process for swap out
 817                  * requires iterating over its threads.  We release
 818                  * allproc_lock here so that process creation and
 819                  * destruction are not blocked while we iterate.
 820                  *
 821                  * To later reacquire allproc_lock and resume
 822                  * iteration over the allproc list, we will first have
 823                  * to release the lock on the process.  We place a
 824                  * hold on the process so that it remains in the
 825                  * allproc list while it is unlocked.
 826                  */
 827                 _PHOLD_LITE(p);
 828                 sx_sunlock(&allproc_lock);
 829
 830                 /*
 831                  * Do not swapout a realtime process.
 832                  * Guarantee swap_idle_threshold1 time in memory.
 833                  * If the system is under memory stress, or if we are
 834                  * swapping idle processes >= swap_idle_threshold2,
 835                  * then swap the process out.
 836                  */
 837                 doswap = true;
 838                 FOREACH_THREAD_IN_PROC(p, td) {
 839                         thread_lock(td);
 840                         slptime = (ticks - td->td_slptick) / hz;
 841                         if (PRI_IS_REALTIME(td->td_pri_class) ||
 842                             slptime < swap_idle_threshold1 ||
 843                             !thread_safetoswapout(td) ||
 844                             ((action & VM_SWAP_NORMAL) == 0 &&
 845                             slptime < swap_idle_threshold2))
 846                                 doswap = false;
 847                         thread_unlock(td);
 848                         if (!doswap)
 849                                 break;
 850                 }
 851                 if (doswap && swapout(p) == 0)
 852                         didswap = true;
 853
 854                 PROC_UNLOCK(p);
 855                 if (didswap) {
 856                         sx_xlock(&allproc_lock);
 857                         swapped_cnt++;
 858                         sx_downgrade(&allproc_lock);
 859                 } else
 860                         sx_slock(&allproc_lock);
 861                 PRELE(p);
 862         }
 863         sx_sunlock(&allproc_lock);
 864
 865         /*
 866          * If we swapped something out, and another process needed memory,
 867          * then wakeup the sched process.
 868          */
 869         if (didswap)
 870                 wakeup(&proc0);
 871 }
 872
 873 static void
 874 swapclear(struct proc *p)
 875 {
 876         struct thread *td;
 877
 878         PROC_LOCK_ASSERT(p, MA_OWNED);
 879
 880         FOREACH_THREAD_IN_PROC(p, td) {
 881                 thread_lock(td);
 882                 td->td_flags |= TDF_INMEM;
 883                 td->td_flags &= ~TDF_SWAPINREQ;
 884                 TD_CLR_SWAPPED(td);
 885                 if (TD_CAN_RUN(td))
 886                         if (setrunnable(td)) {
 887 #ifdef INVARIANTS
 888                                 /*
 889                                  * XXX: We just cleared TDI_SWAPPED
 890                                  * above and set TDF_INMEM, so this
 891                                  * should never happen.
 892                                  */
 893                                 panic("not waking up swapper");
 894 #endif
 895                         }
 896                 thread_unlock(td);
 897         }
 898         p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
 899         p->p_flag |= P_INMEM;
 900 }
 901
 902 static int
 903 swapout(struct proc *p)
 904 {
 905         struct thread *td;
 906
 907         PROC_LOCK_ASSERT(p, MA_OWNED);
 908
 909         /*
 910          * The states of this process and its threads may have changed
 911          * by now.  Assuming that there is only one pageout daemon thread,
 912          * this process should still be in memory.
 913          */
 914         KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
 915             P_INMEM, ("swapout: lost a swapout race?"));
 916
 917         /*
 918          * Remember the resident count.
 919          */
 920         p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
 921
 922         /*
 923          * Check and mark all threads before we proceed.
 924          */
 925         p->p_flag &= ~P_INMEM;
 926         p->p_flag |= P_SWAPPINGOUT;
 927         FOREACH_THREAD_IN_PROC(p, td) {
 928                 thread_lock(td);
 929                 if (!thread_safetoswapout(td)) {
 930                         thread_unlock(td);
 931                         swapclear(p);
 932                         return (EBUSY);
 933                 }
 934                 td->td_flags &= ~TDF_INMEM;
 935                 TD_SET_SWAPPED(td);
 936                 thread_unlock(td);
 937         }
 938         td = FIRST_THREAD_IN_PROC(p);
 939         ++td->td_ru.ru_nswap;
 940         PROC_UNLOCK(p);
 941
 942         /*
 943          * This list is stable because all threads are now prevented from
 944          * running.  The list is only modified in the context of a running
 945          * thread in this process.
 946          */
 947         FOREACH_THREAD_IN_PROC(p, td)
 948                 vm_thread_swapout(td);
 949
 950         PROC_LOCK(p);
 951         p->p_flag &= ~P_SWAPPINGOUT;
 952         p->p_swtick = ticks;
 953         return (0);
 954 }