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/_kstack_cache.h> 89 #include <sys/kthread.h> 90 #include <sys/ktr.h> 91 #include <sys/mount.h> 92 #include <sys/racct.h> 93 #include <sys/resourcevar.h> 94 #include <sys/sched.h> 95 #include <sys/sdt.h> 96 #include <sys/signalvar.h> 97 #include <sys/smp.h> 98 #include <sys/time.h> 99 #include <sys/vnode.h> 100 #include <sys/vmmeter.h> 101 #include <sys/rwlock.h> 102 #include <sys/sx.h> 103 #include <sys/sysctl.h> 104 105 #include <vm/vm.h> 106 #include <vm/vm_param.h> 107 #include <vm/vm_object.h> 108 #include <vm/vm_page.h> 109 #include <vm/vm_map.h> 110 #include <vm/vm_pageout.h> 111 #include <vm/vm_pager.h> 112 #include <vm/vm_phys.h> 113 #include <vm/swap_pager.h> 114 #include <vm/vm_extern.h> 115 #include <vm/uma.h> 116 117 /* the kernel process "vm_daemon" */ 118 static void vm_daemon(void); 119 static struct proc *vmproc; 120 121 static struct kproc_desc vm_kp = { 122 "vmdaemon", 123 vm_daemon, 124 &vmproc 125 }; 126 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 127 128 static int vm_swap_enabled = 1; 129 static int vm_swap_idle_enabled = 0; 130 131 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW, 132 &vm_swap_enabled, 0, 133 "Enable entire process swapout"); 134 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW, 135 &vm_swap_idle_enabled, 0, 136 "Allow swapout on idle criteria"); 137 138 /* 139 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 140 */ 141 static int swap_idle_threshold1 = 2; 142 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, 143 &swap_idle_threshold1, 0, 144 "Guaranteed swapped in time for a process"); 145 146 /* 147 * Swap_idle_threshold2 is the time that a process can be idle before 148 * it will be swapped out, if idle swapping is enabled. 149 */ 150 static int swap_idle_threshold2 = 10; 151 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, 152 &swap_idle_threshold2, 0, 153 "Time before a process will be swapped out"); 154 155 static int vm_pageout_req_swapout; /* XXX */ 156 static int vm_daemon_needed; 157 static struct mtx vm_daemon_mtx; 158 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 159 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 160 161 static void swapclear(struct proc *); 162 static int swapout(struct proc *); 163 static void vm_swapout_map_deactivate_pages(vm_map_t, long); 164 static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long); 165 static void swapout_procs(int action); 166 static void vm_req_vmdaemon(int req); 167 static void vm_thread_swapin(struct thread *td); 168 static void vm_thread_swapout(struct thread *td); 169 170 /* 171 * vm_swapout_object_deactivate_pages 172 * 173 * Deactivate enough pages to satisfy the inactive target 174 * requirements. 175 * 176 * The object and map must be locked. 177 */ 178 static void 179 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 180 long desired) 181 { 182 vm_object_t backing_object, object; 183 vm_page_t p; 184 int act_delta, remove_mode; 185 186 VM_OBJECT_ASSERT_LOCKED(first_object); 187 if ((first_object->flags & OBJ_FICTITIOUS) != 0) 188 return; 189 for (object = first_object;; object = backing_object) { 190 if (pmap_resident_count(pmap) <= desired) 191 goto unlock_return; 192 VM_OBJECT_ASSERT_LOCKED(object); 193 if ((object->flags & OBJ_UNMANAGED) != 0 || 194 object->paging_in_progress != 0) 195 goto unlock_return; 196 197 remove_mode = 0; 198 if (object->shadow_count > 1) 199 remove_mode = 1; 200 /* 201 * Scan the object's entire memory queue. 202 */ 203 TAILQ_FOREACH(p, &object->memq, listq) { 204 if (pmap_resident_count(pmap) <= desired) 205 goto unlock_return; 206 if (should_yield()) 207 goto unlock_return; 208 if (vm_page_busied(p)) 209 continue; 210 VM_CNT_INC(v_pdpages); 211 vm_page_lock(p); 212 if (vm_page_held(p) || 213 !pmap_page_exists_quick(pmap, p)) { 214 vm_page_unlock(p); 215 continue; 216 } 217 act_delta = pmap_ts_referenced(p); 218 if ((p->aflags & PGA_REFERENCED) != 0) { 219 if (act_delta == 0) 220 act_delta = 1; 221 vm_page_aflag_clear(p, PGA_REFERENCED); 222 } 223 if (!vm_page_active(p) && act_delta != 0) { 224 vm_page_activate(p); 225 p->act_count += act_delta; 226 } else if (vm_page_active(p)) { 227 if (act_delta == 0) { 228 p->act_count -= min(p->act_count, 229 ACT_DECLINE); 230 if (!remove_mode && p->act_count == 0) { 231 pmap_remove_all(p); 232 vm_page_deactivate(p); 233 } else 234 vm_page_requeue(p); 235 } else { 236 vm_page_activate(p); 237 if (p->act_count < ACT_MAX - 238 ACT_ADVANCE) 239 p->act_count += ACT_ADVANCE; 240 vm_page_requeue(p); 241 } 242 } else if (vm_page_inactive(p)) 243 pmap_remove_all(p); 244 vm_page_unlock(p); 245 } 246 if ((backing_object = object->backing_object) == NULL) 247 goto unlock_return; 248 VM_OBJECT_RLOCK(backing_object); 249 if (object != first_object) 250 VM_OBJECT_RUNLOCK(object); 251 } 252 unlock_return: 253 if (object != first_object) 254 VM_OBJECT_RUNLOCK(object); 255 } 256 257 /* 258 * deactivate some number of pages in a map, try to do it fairly, but 259 * that is really hard to do. 260 */ 261 static void 262 vm_swapout_map_deactivate_pages(vm_map_t map, long desired) 263 { 264 vm_map_entry_t tmpe; 265 vm_object_t obj, bigobj; 266 int nothingwired; 267 268 if (!vm_map_trylock_read(map)) 269 return; 270 271 bigobj = NULL; 272 nothingwired = TRUE; 273 274 /* 275 * first, search out the biggest object, and try to free pages from 276 * that. 277 */ 278 tmpe = map->header.next; 279 while (tmpe != &map->header) { 280 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 281 obj = tmpe->object.vm_object; 282 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { 283 if (obj->shadow_count <= 1 && 284 (bigobj == NULL || 285 bigobj->resident_page_count < 286 obj->resident_page_count)) { 287 if (bigobj != NULL) 288 VM_OBJECT_RUNLOCK(bigobj); 289 bigobj = obj; 290 } else 291 VM_OBJECT_RUNLOCK(obj); 292 } 293 } 294 if (tmpe->wired_count > 0) 295 nothingwired = FALSE; 296 tmpe = tmpe->next; 297 } 298 299 if (bigobj != NULL) { 300 vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired); 301 VM_OBJECT_RUNLOCK(bigobj); 302 } 303 /* 304 * Next, hunt around for other pages to deactivate. We actually 305 * do this search sort of wrong -- .text first is not the best idea. 306 */ 307 tmpe = map->header.next; 308 while (tmpe != &map->header) { 309 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 310 break; 311 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 312 obj = tmpe->object.vm_object; 313 if (obj != NULL) { 314 VM_OBJECT_RLOCK(obj); 315 vm_swapout_object_deactivate_pages(map->pmap, 316 obj, desired); 317 VM_OBJECT_RUNLOCK(obj); 318 } 319 } 320 tmpe = tmpe->next; 321 } 322 323 /* 324 * Remove all mappings if a process is swapped out, this will free page 325 * table pages. 326 */ 327 if (desired == 0 && nothingwired) { 328 pmap_remove(vm_map_pmap(map), vm_map_min(map), 329 vm_map_max(map)); 330 } 331 332 vm_map_unlock_read(map); 333 } 334 335 /* 336 * Swap out requests 337 */ 338 #define VM_SWAP_NORMAL 1 339 #define VM_SWAP_IDLE 2 340 341 void 342 vm_swapout_run(void) 343 { 344 345 if (vm_swap_enabled) 346 vm_req_vmdaemon(VM_SWAP_NORMAL); 347 } 348 349 /* 350 * Idle process swapout -- run once per second when pagedaemons are 351 * reclaiming pages. 352 */ 353 void 354 vm_swapout_run_idle(void) 355 { 356 static long lsec; 357 358 if (!vm_swap_idle_enabled || time_second == lsec) 359 return; 360 vm_req_vmdaemon(VM_SWAP_IDLE); 361 lsec = time_second; 362 } 363 364 static void 365 vm_req_vmdaemon(int req) 366 { 367 static int lastrun = 0; 368 369 mtx_lock(&vm_daemon_mtx); 370 vm_pageout_req_swapout |= req; 371 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 372 wakeup(&vm_daemon_needed); 373 lastrun = ticks; 374 } 375 mtx_unlock(&vm_daemon_mtx); 376 } 377 378 static void 379 vm_daemon(void) 380 { 381 struct rlimit rsslim; 382 struct proc *p; 383 struct thread *td; 384 struct vmspace *vm; 385 int breakout, swapout_flags, tryagain, attempts; 386 #ifdef RACCT 387 uint64_t rsize, ravailable; 388 #endif 389 390 while (TRUE) { 391 mtx_lock(&vm_daemon_mtx); 392 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 393 #ifdef RACCT 394 racct_enable ? hz : 0 395 #else 396 0 397 #endif 398 ); 399 swapout_flags = vm_pageout_req_swapout; 400 vm_pageout_req_swapout = 0; 401 mtx_unlock(&vm_daemon_mtx); 402 if (swapout_flags != 0) { 403 /* 404 * Drain the per-CPU page queue batches as a deadlock 405 * avoidance measure. 406 */ 407 if ((swapout_flags & VM_SWAP_NORMAL) != 0) 408 vm_page_drain_pqbatch(); 409 swapout_procs(swapout_flags); 410 } 411 412 /* 413 * scan the processes for exceeding their rlimits or if 414 * process is swapped out -- deactivate pages 415 */ 416 tryagain = 0; 417 attempts = 0; 418 again: 419 attempts++; 420 sx_slock(&allproc_lock); 421 FOREACH_PROC_IN_SYSTEM(p) { 422 vm_pindex_t limit, size; 423 424 /* 425 * if this is a system process or if we have already 426 * looked at this process, skip it. 427 */ 428 PROC_LOCK(p); 429 if (p->p_state != PRS_NORMAL || 430 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 431 PROC_UNLOCK(p); 432 continue; 433 } 434 /* 435 * if the process is in a non-running type state, 436 * don't touch it. 437 */ 438 breakout = 0; 439 FOREACH_THREAD_IN_PROC(p, td) { 440 thread_lock(td); 441 if (!TD_ON_RUNQ(td) && 442 !TD_IS_RUNNING(td) && 443 !TD_IS_SLEEPING(td) && 444 !TD_IS_SUSPENDED(td)) { 445 thread_unlock(td); 446 breakout = 1; 447 break; 448 } 449 thread_unlock(td); 450 } 451 if (breakout) { 452 PROC_UNLOCK(p); 453 continue; 454 } 455 /* 456 * get a limit 457 */ 458 lim_rlimit_proc(p, RLIMIT_RSS, &rsslim); 459 limit = OFF_TO_IDX( 460 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 461 462 /* 463 * let processes that are swapped out really be 464 * swapped out set the limit to nothing (will force a 465 * swap-out.) 466 */ 467 if ((p->p_flag & P_INMEM) == 0) 468 limit = 0; /* XXX */ 469 vm = vmspace_acquire_ref(p); 470 _PHOLD_LITE(p); 471 PROC_UNLOCK(p); 472 if (vm == NULL) { 473 PRELE(p); 474 continue; 475 } 476 sx_sunlock(&allproc_lock); 477 478 size = vmspace_resident_count(vm); 479 if (size >= limit) { 480 vm_swapout_map_deactivate_pages( 481 &vm->vm_map, limit); 482 size = vmspace_resident_count(vm); 483 } 484 #ifdef RACCT 485 if (racct_enable) { 486 rsize = IDX_TO_OFF(size); 487 PROC_LOCK(p); 488 if (p->p_state == PRS_NORMAL) 489 racct_set(p, RACCT_RSS, rsize); 490 ravailable = racct_get_available(p, RACCT_RSS); 491 PROC_UNLOCK(p); 492 if (rsize > ravailable) { 493 /* 494 * Don't be overly aggressive; this 495 * might be an innocent process, 496 * and the limit could've been exceeded 497 * by some memory hog. Don't try 498 * to deactivate more than 1/4th 499 * of process' resident set size. 500 */ 501 if (attempts <= 8) { 502 if (ravailable < rsize - 503 (rsize / 4)) { 504 ravailable = rsize - 505 (rsize / 4); 506 } 507 } 508 vm_swapout_map_deactivate_pages( 509 &vm->vm_map, 510 OFF_TO_IDX(ravailable)); 511 /* Update RSS usage after paging out. */ 512 size = vmspace_resident_count(vm); 513 rsize = IDX_TO_OFF(size); 514 PROC_LOCK(p); 515 if (p->p_state == PRS_NORMAL) 516 racct_set(p, RACCT_RSS, rsize); 517 PROC_UNLOCK(p); 518 if (rsize > ravailable) 519 tryagain = 1; 520 } 521 } 522 #endif 523 vmspace_free(vm); 524 sx_slock(&allproc_lock); 525 PRELE(p); 526 } 527 sx_sunlock(&allproc_lock); 528 if (tryagain != 0 && attempts <= 10) { 529 maybe_yield(); 530 goto again; 531 } 532 } 533 } 534 535 /* 536 * Allow a thread's kernel stack to be paged out. 537 */ 538 static void 539 vm_thread_swapout(struct thread *td) 540 { 541 vm_object_t ksobj; 542 vm_page_t m; 543 int i, pages; 544 545 cpu_thread_swapout(td); 546 pages = td->td_kstack_pages; 547 ksobj = td->td_kstack_obj; 548 pmap_qremove(td->td_kstack, pages); 549 VM_OBJECT_WLOCK(ksobj); 550 for (i = 0; i < pages; i++) { 551 m = vm_page_lookup(ksobj, i); 552 if (m == NULL) 553 panic("vm_thread_swapout: kstack already missing?"); 554 vm_page_dirty(m); 555 vm_page_lock(m); 556 vm_page_unwire(m, PQ_LAUNDRY); 557 vm_page_unlock(m); 558 } 559 VM_OBJECT_WUNLOCK(ksobj); 560 } 561 562 /* 563 * Bring the kernel stack for a specified thread back in. 564 */ 565 static void 566 vm_thread_swapin(struct thread *td) 567 { 568 vm_object_t ksobj; 569 vm_page_t ma[KSTACK_MAX_PAGES]; 570 int a, count, i, j, pages, rv; 571 572 pages = td->td_kstack_pages; 573 ksobj = td->td_kstack_obj; 574 VM_OBJECT_WLOCK(ksobj); 575 (void)vm_page_grab_pages(ksobj, 0, VM_ALLOC_NORMAL | VM_ALLOC_WIRED, ma, 576 pages); 577 for (i = 0; i < pages;) { 578 vm_page_assert_xbusied(ma[i]); 579 if (ma[i]->valid == VM_PAGE_BITS_ALL) { 580 vm_page_xunbusy(ma[i]); 581 i++; 582 continue; 583 } 584 vm_object_pip_add(ksobj, 1); 585 for (j = i + 1; j < pages; j++) 586 if (ma[j]->valid == VM_PAGE_BITS_ALL) 587 break; 588 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a); 589 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i])); 590 count = min(a + 1, j - i); 591 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL); 592 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d", 593 __func__, td->td_proc->p_pid)); 594 vm_object_pip_wakeup(ksobj); 595 for (j = i; j < i + count; j++) 596 vm_page_xunbusy(ma[j]); 597 i += count; 598 } 599 VM_OBJECT_WUNLOCK(ksobj); 600 pmap_qenter(td->td_kstack, ma, pages); 601 cpu_thread_swapin(td); 602 } 603 604 void 605 faultin(struct proc *p) 606 { 607 struct thread *td; 608 609 PROC_LOCK_ASSERT(p, MA_OWNED); 610 /* 611 * If another process is swapping in this process, 612 * just wait until it finishes. 613 */ 614 if (p->p_flag & P_SWAPPINGIN) { 615 while (p->p_flag & P_SWAPPINGIN) 616 msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0); 617 return; 618 } 619 if ((p->p_flag & P_INMEM) == 0) { 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); 627 628 /* 629 * We hold no lock here because the list of threads 630 * can not change while all threads in the process are 631 * swapped out. 632 */ 633 FOREACH_THREAD_IN_PROC(p, td) 634 vm_thread_swapin(td); 635 PROC_LOCK(p); 636 swapclear(p); 637 p->p_swtick = ticks; 638 639 wakeup(&p->p_flag); 640 641 /* Allow other threads to swap p out now. */ 642 --p->p_lock; 643 } 644 } 645 646 /* 647 * This swapin algorithm attempts to swap-in processes only if there 648 * is enough space for them. Of course, if a process waits for a long 649 * time, it will be swapped in anyway. 650 */ 651 void 652 swapper(void) 653 { 654 struct proc *p, *pp; 655 struct thread *td; 656 int ppri, pri, slptime, swtime; 657 658 loop: 659 if (vm_page_count_min()) { 660 vm_wait_min(); 661 goto loop; 662 } 663 664 pp = NULL; 665 ppri = INT_MIN; 666 sx_slock(&allproc_lock); 667 FOREACH_PROC_IN_SYSTEM(p) { 668 PROC_LOCK(p); 669 if (p->p_state == PRS_NEW || 670 p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) { 671 PROC_UNLOCK(p); 672 continue; 673 } 674 swtime = (ticks - p->p_swtick) / hz; 675 FOREACH_THREAD_IN_PROC(p, td) { 676 /* 677 * An otherwise runnable thread of a process 678 * swapped out has only the TDI_SWAPPED bit set. 679 */ 680 thread_lock(td); 681 if (td->td_inhibitors == TDI_SWAPPED) { 682 slptime = (ticks - td->td_slptick) / hz; 683 pri = swtime + slptime; 684 if ((td->td_flags & TDF_SWAPINREQ) == 0) 685 pri -= p->p_nice * 8; 686 /* 687 * if this thread is higher priority 688 * and there is enough space, then select 689 * this process instead of the previous 690 * selection. 691 */ 692 if (pri > ppri) { 693 pp = p; 694 ppri = pri; 695 } 696 } 697 thread_unlock(td); 698 } 699 PROC_UNLOCK(p); 700 } 701 sx_sunlock(&allproc_lock); 702 703 /* 704 * Nothing to do, back to sleep. 705 */ 706 if ((p = pp) == NULL) { 707 tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2); 708 goto loop; 709 } 710 PROC_LOCK(p); 711 712 /* 713 * Another process may be bringing or may have already 714 * brought this process in while we traverse all threads. 715 * Or, this process may even be being swapped out again. 716 */ 717 if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) { 718 PROC_UNLOCK(p); 719 goto loop; 720 } 721 722 /* 723 * We would like to bring someone in. 724 */ 725 faultin(p); 726 PROC_UNLOCK(p); 727 goto loop; 728 } 729 730 /* 731 * First, if any processes have been sleeping or stopped for at least 732 * "swap_idle_threshold1" seconds, they are swapped out. If, however, 733 * no such processes exist, then the longest-sleeping or stopped 734 * process is swapped out. Finally, and only as a last resort, if 735 * there are no sleeping or stopped processes, the longest-resident 736 * process is swapped out. 737 */ 738 static void 739 swapout_procs(int action) 740 { 741 struct proc *p; 742 struct thread *td; 743 int slptime; 744 bool didswap, doswap; 745 746 MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0); 747 748 didswap = false; 749 sx_slock(&allproc_lock); 750 FOREACH_PROC_IN_SYSTEM(p) { 751 /* 752 * Filter out not yet fully constructed processes. Do 753 * not swap out held processes. Avoid processes which 754 * are system, exiting, execing, traced, already swapped 755 * out or are in the process of being swapped in or out. 756 */ 757 PROC_LOCK(p); 758 if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag & 759 (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE | 760 P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) != 761 P_INMEM) { 762 PROC_UNLOCK(p); 763 continue; 764 } 765 766 /* 767 * Further consideration of this process for swap out 768 * requires iterating over its threads. We release 769 * allproc_lock here so that process creation and 770 * destruction are not blocked while we iterate. 771 * 772 * To later reacquire allproc_lock and resume 773 * iteration over the allproc list, we will first have 774 * to release the lock on the process. We place a 775 * hold on the process so that it remains in the 776 * allproc list while it is unlocked. 777 */ 778 _PHOLD_LITE(p); 779 sx_sunlock(&allproc_lock); 780 781 /* 782 * Do not swapout a realtime process. 783 * Guarantee swap_idle_threshold1 time in memory. 784 * If the system is under memory stress, or if we are 785 * swapping idle processes >= swap_idle_threshold2, 786 * then swap the process out. 787 */ 788 doswap = true; 789 FOREACH_THREAD_IN_PROC(p, td) { 790 thread_lock(td); 791 slptime = (ticks - td->td_slptick) / hz; 792 if (PRI_IS_REALTIME(td->td_pri_class) || 793 slptime < swap_idle_threshold1 || 794 !thread_safetoswapout(td) || 795 ((action & VM_SWAP_NORMAL) == 0 && 796 slptime < swap_idle_threshold2)) 797 doswap = false; 798 thread_unlock(td); 799 if (!doswap) 800 break; 801 } 802 if (doswap && swapout(p) == 0) 803 didswap = true; 804 805 PROC_UNLOCK(p); 806 sx_slock(&allproc_lock); 807 PRELE(p); 808 } 809 sx_sunlock(&allproc_lock); 810 811 /* 812 * If we swapped something out, and another process needed memory, 813 * then wakeup the sched process. 814 */ 815 if (didswap) 816 wakeup(&proc0); 817 } 818 819 static void 820 swapclear(struct proc *p) 821 { 822 struct thread *td; 823 824 PROC_LOCK_ASSERT(p, MA_OWNED); 825 826 FOREACH_THREAD_IN_PROC(p, td) { 827 thread_lock(td); 828 td->td_flags |= TDF_INMEM; 829 td->td_flags &= ~TDF_SWAPINREQ; 830 TD_CLR_SWAPPED(td); 831 if (TD_CAN_RUN(td)) 832 if (setrunnable(td)) { 833 #ifdef INVARIANTS 834 /* 835 * XXX: We just cleared TDI_SWAPPED 836 * above and set TDF_INMEM, so this 837 * should never happen. 838 */ 839 panic("not waking up swapper"); 840 #endif 841 } 842 thread_unlock(td); 843 } 844 p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT); 845 p->p_flag |= P_INMEM; 846 } 847 848 static int 849 swapout(struct proc *p) 850 { 851 struct thread *td; 852 853 PROC_LOCK_ASSERT(p, MA_OWNED); 854 855 /* 856 * The states of this process and its threads may have changed 857 * by now. Assuming that there is only one pageout daemon thread, 858 * this process should still be in memory. 859 */ 860 KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) == 861 P_INMEM, ("swapout: lost a swapout race?")); 862 863 /* 864 * Remember the resident count. 865 */ 866 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 867 868 /* 869 * Check and mark all threads before we proceed. 870 */ 871 p->p_flag &= ~P_INMEM; 872 p->p_flag |= P_SWAPPINGOUT; 873 FOREACH_THREAD_IN_PROC(p, td) { 874 thread_lock(td); 875 if (!thread_safetoswapout(td)) { 876 thread_unlock(td); 877 swapclear(p); 878 return (EBUSY); 879 } 880 td->td_flags &= ~TDF_INMEM; 881 TD_SET_SWAPPED(td); 882 thread_unlock(td); 883 } 884 td = FIRST_THREAD_IN_PROC(p); 885 ++td->td_ru.ru_nswap; 886 PROC_UNLOCK(p); 887 888 /* 889 * This list is stable because all threads are now prevented from 890 * running. The list is only modified in the context of a running 891 * thread in this process. 892 */ 893 FOREACH_THREAD_IN_PROC(p, td) 894 vm_thread_swapout(td); 895 896 PROC_LOCK(p); 897 p->p_flag &= ~P_SWAPPINGOUT; 898 p->p_swtick = ticks; 899 return (0); 900 } 901