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 void swapclear(struct proc *);
 160 static int swapout(struct proc *);
 161 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
 162 static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long);
 163 static void swapout_procs(int action);
 164 static void vm_req_vmdaemon(int req);
 165 static void vm_thread_swapin(struct thread *td);
 166 static void vm_thread_swapout(struct thread *td);
 167
 168 /*
 169  *      vm_swapout_object_deactivate_pages
 170  *
 171  *      Deactivate enough pages to satisfy the inactive target
 172  *      requirements.
 173  *
 174  *      The object and map must be locked.
 175  */
 176 static void
 177 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
 178     long desired)
 179 {
 180         vm_object_t backing_object, object;
 181         vm_page_t p;
 182         int act_delta, remove_mode;
 183
 184         VM_OBJECT_ASSERT_LOCKED(first_object);
 185         if ((first_object->flags & OBJ_FICTITIOUS) != 0)
 186                 return;
 187         for (object = first_object;; object = backing_object) {
 188                 if (pmap_resident_count(pmap) <= desired)
 189                         goto unlock_return;
 190                 VM_OBJECT_ASSERT_LOCKED(object);
 191                 if ((object->flags & OBJ_UNMANAGED) != 0 ||
 192                     object->paging_in_progress != 0)
 193                         goto unlock_return;
 194
 195                 remove_mode = 0;
 196                 if (object->shadow_count > 1)
 197                         remove_mode = 1;
 198                 /*
 199                  * Scan the object's entire memory queue.
 200                  */
 201                 TAILQ_FOREACH(p, &object->memq, listq) {
 202                         if (pmap_resident_count(pmap) <= desired)
 203                                 goto unlock_return;
 204                         if (vm_page_busied(p))
 205                                 continue;
 206                         VM_CNT_INC(v_pdpages);
 207                         vm_page_lock(p);
 208                         if (p->wire_count != 0 || p->hold_count != 0 ||
 209                             !pmap_page_exists_quick(pmap, p)) {
 210                                 vm_page_unlock(p);
 211                                 continue;
 212                         }
 213                         act_delta = pmap_ts_referenced(p);
 214                         if ((p->aflags & PGA_REFERENCED) != 0) {
 215                                 if (act_delta == 0)
 216                                         act_delta = 1;
 217                                 vm_page_aflag_clear(p, PGA_REFERENCED);
 218                         }
 219                         if (!vm_page_active(p) && act_delta != 0) {
 220                                 vm_page_activate(p);
 221                                 p->act_count += act_delta;
 222                         } else if (vm_page_active(p)) {
 223                                 if (act_delta == 0) {
 224                                         p->act_count -= min(p->act_count,
 225                                             ACT_DECLINE);
 226                                         if (!remove_mode && p->act_count == 0) {
 227                                                 pmap_remove_all(p);
 228                                                 vm_page_deactivate(p);
 229                                         } else
 230                                                 vm_page_requeue(p);
 231                                 } else {
 232                                         vm_page_activate(p);
 233                                         if (p->act_count < ACT_MAX -
 234                                             ACT_ADVANCE)
 235                                                 p->act_count += ACT_ADVANCE;
 236                                         vm_page_requeue(p);
 237                                 }
 238                         } else if (vm_page_inactive(p))
 239                                 pmap_remove_all(p);
 240                         vm_page_unlock(p);
 241                 }
 242                 if ((backing_object = object->backing_object) == NULL)
 243                         goto unlock_return;
 244                 VM_OBJECT_RLOCK(backing_object);
 245                 if (object != first_object)
 246                         VM_OBJECT_RUNLOCK(object);
 247         }
 248 unlock_return:
 249         if (object != first_object)
 250                 VM_OBJECT_RUNLOCK(object);
 251 }
 252
 253 /*
 254  * deactivate some number of pages in a map, try to do it fairly, but
 255  * that is really hard to do.
 256  */
 257 static void
 258 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
 259 {
 260         vm_map_entry_t tmpe;
 261         vm_object_t obj, bigobj;
 262         int nothingwired;
 263
 264         if (!vm_map_trylock(map))
 265                 return;
 266
 267         bigobj = NULL;
 268         nothingwired = TRUE;
 269
 270         /*
 271          * first, search out the biggest object, and try to free pages from
 272          * that.
 273          */
 274         tmpe = map->header.next;
 275         while (tmpe != &map->header) {
 276                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 277                         obj = tmpe->object.vm_object;
 278                         if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
 279                                 if (obj->shadow_count <= 1 &&
 280                                     (bigobj == NULL ||
 281                                      bigobj->resident_page_count <
 282                                      obj->resident_page_count)) {
 283                                         if (bigobj != NULL)
 284                                                 VM_OBJECT_RUNLOCK(bigobj);
 285                                         bigobj = obj;
 286                                 } else
 287                                         VM_OBJECT_RUNLOCK(obj);
 288                         }
 289                 }
 290                 if (tmpe->wired_count > 0)
 291                         nothingwired = FALSE;
 292                 tmpe = tmpe->next;
 293         }
 294
 295         if (bigobj != NULL) {
 296                 vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
 297                 VM_OBJECT_RUNLOCK(bigobj);
 298         }
 299         /*
 300          * Next, hunt around for other pages to deactivate.  We actually
 301          * do this search sort of wrong -- .text first is not the best idea.
 302          */
 303         tmpe = map->header.next;
 304         while (tmpe != &map->header) {
 305                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
 306                         break;
 307                 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
 308                         obj = tmpe->object.vm_object;
 309                         if (obj != NULL) {
 310                                 VM_OBJECT_RLOCK(obj);
 311                                 vm_swapout_object_deactivate_pages(map->pmap,
 312                                     obj, desired);
 313                                 VM_OBJECT_RUNLOCK(obj);
 314                         }
 315                 }
 316                 tmpe = tmpe->next;
 317         }
 318
 319         /*
 320          * Remove all mappings if a process is swapped out, this will free page
 321          * table pages.
 322          */
 323         if (desired == 0 && nothingwired) {
 324                 pmap_remove(vm_map_pmap(map), vm_map_min(map),
 325                     vm_map_max(map));
 326         }
 327
 328         vm_map_unlock(map);
 329 }
 330
 331 /*
 332  * Swap out requests
 333  */
 334 #define VM_SWAP_NORMAL 1
 335 #define VM_SWAP_IDLE 2
 336
 337 void
 338 vm_swapout_run(void)
 339 {
 340
 341         if (vm_swap_enabled)
 342                 vm_req_vmdaemon(VM_SWAP_NORMAL);
 343 }
 344
 345 /*
 346  * Idle process swapout -- run once per second when pagedaemons are
 347  * reclaiming pages.
 348  */
 349 void
 350 vm_swapout_run_idle(void)
 351 {
 352         static long lsec;
 353
 354         if (!vm_swap_idle_enabled || time_second == lsec)
 355                 return;
 356         vm_req_vmdaemon(VM_SWAP_IDLE);
 357         lsec = time_second;
 358 }
 359
 360 static void
 361 vm_req_vmdaemon(int req)
 362 {
 363         static int lastrun = 0;
 364
 365         mtx_lock(&vm_daemon_mtx);
 366         vm_pageout_req_swapout |= req;
 367         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
 368                 wakeup(&vm_daemon_needed);
 369                 lastrun = ticks;
 370         }
 371         mtx_unlock(&vm_daemon_mtx);
 372 }
 373
 374 static void
 375 vm_daemon(void)
 376 {
 377         struct rlimit rsslim;
 378         struct proc *p;
 379         struct thread *td;
 380         struct vmspace *vm;
 381         int breakout, swapout_flags, tryagain, attempts;
 382 #ifdef RACCT
 383         uint64_t rsize, ravailable;
 384 #endif
 385
 386         while (TRUE) {
 387                 mtx_lock(&vm_daemon_mtx);
 388                 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
 389 #ifdef RACCT
 390                     racct_enable ? hz : 0
 391 #else
 392                     0
 393 #endif
 394                 );
 395                 swapout_flags = vm_pageout_req_swapout;
 396                 vm_pageout_req_swapout = 0;
 397                 mtx_unlock(&vm_daemon_mtx);
 398                 if (swapout_flags)
 399                         swapout_procs(swapout_flags);
 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                         goto again;
 519         }
 520 }
 521
 522 /*
 523  * Allow a thread's kernel stack to be paged out.
 524  */
 525 static void
 526 vm_thread_swapout(struct thread *td)
 527 {
 528         vm_object_t ksobj;
 529         vm_page_t m;
 530         int i, pages;
 531
 532         cpu_thread_swapout(td);
 533         pages = td->td_kstack_pages;
 534         ksobj = td->td_kstack_obj;
 535         pmap_qremove(td->td_kstack, pages);
 536         VM_OBJECT_WLOCK(ksobj);
 537         for (i = 0; i < pages; i++) {
 538                 m = vm_page_lookup(ksobj, i);
 539                 if (m == NULL)
 540                         panic("vm_thread_swapout: kstack already missing?");
 541                 vm_page_dirty(m);
 542                 vm_page_lock(m);
 543                 vm_page_unwire(m, PQ_INACTIVE);
 544                 vm_page_unlock(m);
 545         }
 546         VM_OBJECT_WUNLOCK(ksobj);
 547 }
 548
 549 /*
 550  * Bring the kernel stack for a specified thread back in.
 551  */
 552 static void
 553 vm_thread_swapin(struct thread *td)
 554 {
 555         vm_object_t ksobj;
 556         vm_page_t ma[KSTACK_MAX_PAGES];
 557         int pages;
 558
 559         pages = td->td_kstack_pages;
 560         ksobj = td->td_kstack_obj;
 561         VM_OBJECT_WLOCK(ksobj);
 562         (void)vm_page_grab_pages(ksobj, 0, VM_ALLOC_NORMAL | VM_ALLOC_WIRED, ma,
 563             pages);
 564         for (int i = 0; i < pages;) {
 565                 int j, a, count, rv;
 566
 567                 vm_page_assert_xbusied(ma[i]);
 568                 if (ma[i]->valid == VM_PAGE_BITS_ALL) {
 569                         vm_page_xunbusy(ma[i]);
 570                         i++;
 571                         continue;
 572                 }
 573                 vm_object_pip_add(ksobj, 1);
 574                 for (j = i + 1; j < pages; j++)
 575                         if (ma[j]->valid == VM_PAGE_BITS_ALL)
 576                                 break;
 577                 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
 578                 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
 579                 count = min(a + 1, j - i);
 580                 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
 581                 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
 582                     __func__, td->td_proc->p_pid));
 583                 vm_object_pip_wakeup(ksobj);
 584                 for (j = i; j < i + count; j++)
 585                         vm_page_xunbusy(ma[j]);
 586                 i += count;
 587         }
 588         VM_OBJECT_WUNLOCK(ksobj);
 589         pmap_qenter(td->td_kstack, ma, pages);
 590         cpu_thread_swapin(td);
 591 }
 592
 593 void
 594 faultin(struct proc *p)
 595 {
 596         struct thread *td;
 597
 598         PROC_LOCK_ASSERT(p, MA_OWNED);
 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         if ((p->p_flag & P_INMEM) == 0) {
 609                 /*
 610                  * Don't let another thread swap process p out while we are
 611                  * busy swapping it in.
 612                  */
 613                 ++p->p_lock;
 614                 p->p_flag |= P_SWAPPINGIN;
 615                 PROC_UNLOCK(p);
 616
 617                 /*
 618                  * We hold no lock here because the list of threads
 619                  * can not change while all threads in the process are
 620                  * swapped out.
 621                  */
 622                 FOREACH_THREAD_IN_PROC(p, td)
 623                         vm_thread_swapin(td);
 624                 PROC_LOCK(p);
 625                 swapclear(p);
 626                 p->p_swtick = ticks;
 627
 628                 wakeup(&p->p_flag);
 629
 630                 /* Allow other threads to swap p out now. */
 631                 --p->p_lock;
 632         }
 633 }
 634
 635 /*
 636  * This swapin algorithm attempts to swap-in processes only if there
 637  * is enough space for them.  Of course, if a process waits for a long
 638  * time, it will be swapped in anyway.
 639  */
 640 void
 641 swapper(void)
 642 {
 643         struct proc *p;
 644         struct thread *td;
 645         struct proc *pp;
 646         int slptime;
 647         int swtime;
 648         int ppri;
 649         int pri;
 650
 651 loop:
 652         if (vm_page_count_min()) {
 653                 VM_WAIT;
 654                 goto loop;
 655         }
 656
 657         pp = NULL;
 658         ppri = INT_MIN;
 659         sx_slock(&allproc_lock);
 660         FOREACH_PROC_IN_SYSTEM(p) {
 661                 PROC_LOCK(p);
 662                 if (p->p_state == PRS_NEW ||
 663                     p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
 664                         PROC_UNLOCK(p);
 665                         continue;
 666                 }
 667                 swtime = (ticks - p->p_swtick) / hz;
 668                 FOREACH_THREAD_IN_PROC(p, td) {
 669                         /*
 670                          * An otherwise runnable thread of a process
 671                          * swapped out has only the TDI_SWAPPED bit set.
 672                          */
 673                         thread_lock(td);
 674                         if (td->td_inhibitors == TDI_SWAPPED) {
 675                                 slptime = (ticks - td->td_slptick) / hz;
 676                                 pri = swtime + slptime;
 677                                 if ((td->td_flags & TDF_SWAPINREQ) == 0)
 678                                         pri -= p->p_nice * 8;
 679                                 /*
 680                                  * if this thread is higher priority
 681                                  * and there is enough space, then select
 682                                  * this process instead of the previous
 683                                  * selection.
 684                                  */
 685                                 if (pri > ppri) {
 686                                         pp = p;
 687                                         ppri = pri;
 688                                 }
 689                         }
 690                         thread_unlock(td);
 691                 }
 692                 PROC_UNLOCK(p);
 693         }
 694         sx_sunlock(&allproc_lock);
 695
 696         /*
 697          * Nothing to do, back to sleep.
 698          */
 699         if ((p = pp) == NULL) {
 700                 tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2);
 701                 goto loop;
 702         }
 703         PROC_LOCK(p);
 704
 705         /*
 706          * Another process may be bringing or may have already
 707          * brought this process in while we traverse all threads.
 708          * Or, this process may even be being swapped out again.
 709          */
 710         if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
 711                 PROC_UNLOCK(p);
 712                 goto loop;
 713         }
 714
 715         /*
 716          * We would like to bring someone in. (only if there is space).
 717          * [What checks the space? ]
 718          */
 719         faultin(p);
 720         PROC_UNLOCK(p);
 721         goto loop;
 722 }
 723
 724 /*
 725  * First, if any processes have been sleeping or stopped for at least
 726  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
 727  * no such processes exist, then the longest-sleeping or stopped
 728  * process is swapped out.  Finally, and only as a last resort, if
 729  * there are no sleeping or stopped processes, the longest-resident
 730  * process is swapped out.
 731  */
 732 static void
 733 swapout_procs(int action)
 734 {
 735         struct proc *p;
 736         struct thread *td;
 737         int didswap = 0;
 738
 739 retry:
 740         sx_slock(&allproc_lock);
 741         FOREACH_PROC_IN_SYSTEM(p) {
 742                 struct vmspace *vm;
 743                 int minslptime = 100000;
 744                 int slptime;
 745
 746                 PROC_LOCK(p);
 747                 /*
 748                  * Watch out for a process in
 749                  * creation.  It may have no
 750                  * address space or lock yet.
 751                  */
 752                 if (p->p_state == PRS_NEW) {
 753                         PROC_UNLOCK(p);
 754                         continue;
 755                 }
 756                 /*
 757                  * An aio daemon switches its
 758                  * address space while running.
 759                  * Perform a quick check whether
 760                  * a process has P_SYSTEM.
 761                  * Filter out exiting processes.
 762                  */
 763                 if ((p->p_flag & (P_SYSTEM | P_WEXIT)) != 0) {
 764                         PROC_UNLOCK(p);
 765                         continue;
 766                 }
 767                 _PHOLD_LITE(p);
 768                 PROC_UNLOCK(p);
 769                 sx_sunlock(&allproc_lock);
 770
 771                 /*
 772                  * Do not swapout a process that
 773                  * is waiting for VM data
 774                  * structures as there is a possible
 775                  * deadlock.  Test this first as
 776                  * this may block.
 777                  *
 778                  * Lock the map until swapout
 779                  * finishes, or a thread of this
 780                  * process may attempt to alter
 781                  * the map.
 782                  */
 783                 vm = vmspace_acquire_ref(p);
 784                 if (vm == NULL)
 785                         goto nextproc2;
 786                 if (!vm_map_trylock(&vm->vm_map))
 787                         goto nextproc1;
 788
 789                 PROC_LOCK(p);
 790                 if (p->p_lock != 1 || (p->p_flag & (P_STOPPED_SINGLE |
 791                     P_TRACED | P_SYSTEM)) != 0)
 792                         goto nextproc;
 793
 794                 /*
 795                  * only aiod changes vmspace, however it will be
 796                  * skipped because of the if statement above checking
 797                  * for P_SYSTEM
 798                  */
 799                 if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM)
 800                         goto nextproc;
 801
 802                 switch (p->p_state) {
 803                 default:
 804                         /* Don't swap out processes in any sort
 805                          * of 'special' state. */
 806                         break;
 807
 808                 case PRS_NORMAL:
 809                         /*
 810                          * do not swapout a realtime process
 811                          * Check all the thread groups..
 812                          */
 813                         FOREACH_THREAD_IN_PROC(p, td) {
 814                                 thread_lock(td);
 815                                 if (PRI_IS_REALTIME(td->td_pri_class)) {
 816                                         thread_unlock(td);
 817                                         goto nextproc;
 818                                 }
 819                                 slptime = (ticks - td->td_slptick) / hz;
 820                                 /*
 821                                  * Guarantee swap_idle_threshold1
 822                                  * time in memory.
 823                                  */
 824                                 if (slptime < swap_idle_threshold1) {
 825                                         thread_unlock(td);
 826                                         goto nextproc;
 827                                 }
 828
 829                                 /*
 830                                  * Do not swapout a process if it is
 831                                  * waiting on a critical event of some
 832                                  * kind or there is a thread whose
 833                                  * pageable memory may be accessed.
 834                                  *
 835                                  * This could be refined to support
 836                                  * swapping out a thread.
 837                                  */
 838                                 if (!thread_safetoswapout(td)) {
 839                                         thread_unlock(td);
 840                                         goto nextproc;
 841                                 }
 842                                 /*
 843                                  * If the system is under memory stress,
 844                                  * or if we are swapping
 845                                  * idle processes >= swap_idle_threshold2,
 846                                  * then swap the process out.
 847                                  */
 848                                 if (((action & VM_SWAP_NORMAL) == 0) &&
 849                                     (((action & VM_SWAP_IDLE) == 0) ||
 850                                     (slptime < swap_idle_threshold2))) {
 851                                         thread_unlock(td);
 852                                         goto nextproc;
 853                                 }
 854
 855                                 if (minslptime > slptime)
 856                                         minslptime = slptime;
 857                                 thread_unlock(td);
 858                         }
 859
 860                         /*
 861                          * If the pageout daemon didn't free enough pages,
 862                          * or if this process is idle and the system is
 863                          * configured to swap proactively, swap it out.
 864                          */
 865                         if ((action & VM_SWAP_NORMAL) ||
 866                                 ((action & VM_SWAP_IDLE) &&
 867                                  (minslptime > swap_idle_threshold2))) {
 868                                 _PRELE(p);
 869                                 if (swapout(p) == 0)
 870                                         didswap++;
 871                                 PROC_UNLOCK(p);
 872                                 vm_map_unlock(&vm->vm_map);
 873                                 vmspace_free(vm);
 874                                 goto retry;
 875                         }
 876                 }
 877 nextproc:
 878                 PROC_UNLOCK(p);
 879                 vm_map_unlock(&vm->vm_map);
 880 nextproc1:
 881                 vmspace_free(vm);
 882 nextproc2:
 883                 sx_slock(&allproc_lock);
 884                 PRELE(p);
 885         }
 886         sx_sunlock(&allproc_lock);
 887         /*
 888          * If we swapped something out, and another process needed memory,
 889          * then wakeup the sched process.
 890          */
 891         if (didswap)
 892                 wakeup(&proc0);
 893 }
 894
 895 static void
 896 swapclear(struct proc *p)
 897 {
 898         struct thread *td;
 899
 900         PROC_LOCK_ASSERT(p, MA_OWNED);
 901
 902         FOREACH_THREAD_IN_PROC(p, td) {
 903                 thread_lock(td);
 904                 td->td_flags |= TDF_INMEM;
 905                 td->td_flags &= ~TDF_SWAPINREQ;
 906                 TD_CLR_SWAPPED(td);
 907                 if (TD_CAN_RUN(td))
 908                         if (setrunnable(td)) {
 909 #ifdef INVARIANTS
 910                                 /*
 911                                  * XXX: We just cleared TDI_SWAPPED
 912                                  * above and set TDF_INMEM, so this
 913                                  * should never happen.
 914                                  */
 915                                 panic("not waking up swapper");
 916 #endif
 917                         }
 918                 thread_unlock(td);
 919         }
 920         p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
 921         p->p_flag |= P_INMEM;
 922 }
 923
 924 static int
 925 swapout(struct proc *p)
 926 {
 927         struct thread *td;
 928
 929         PROC_LOCK_ASSERT(p, MA_OWNED);
 930
 931         /*
 932          * The states of this process and its threads may have changed
 933          * by now.  Assuming that there is only one pageout daemon thread,
 934          * this process should still be in memory.
 935          */
 936         KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
 937             P_INMEM, ("swapout: lost a swapout race?"));
 938
 939         /*
 940          * remember the process resident count
 941          */
 942         p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
 943         /*
 944          * Check and mark all threads before we proceed.
 945          */
 946         p->p_flag &= ~P_INMEM;
 947         p->p_flag |= P_SWAPPINGOUT;
 948         FOREACH_THREAD_IN_PROC(p, td) {
 949                 thread_lock(td);
 950                 if (!thread_safetoswapout(td)) {
 951                         thread_unlock(td);
 952                         swapclear(p);
 953                         return (EBUSY);
 954                 }
 955                 td->td_flags &= ~TDF_INMEM;
 956                 TD_SET_SWAPPED(td);
 957                 thread_unlock(td);
 958         }
 959         td = FIRST_THREAD_IN_PROC(p);
 960         ++td->td_ru.ru_nswap;
 961         PROC_UNLOCK(p);
 962
 963         /*
 964          * This list is stable because all threads are now prevented from
 965          * running.  The list is only modified in the context of a running
 966          * thread in this process.
 967          */
 968         FOREACH_THREAD_IN_PROC(p, td)
 969                 vm_thread_swapout(td);
 970
 971         PROC_LOCK(p);
 972         p->p_flag &= ~P_SWAPPINGOUT;
 973         p->p_swtick = ticks;
 974         return (0);
 975 }