vm_glue.c revision 116279
1/* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_glue.c 8.6 (Berkeley) 1/5/94 37 * 38 * 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63#include <sys/cdefs.h> 64__FBSDID("$FreeBSD: head/sys/vm/vm_glue.c 116279 2003-06-13 03:02:28Z alc $"); 65 66#include "opt_vm.h" 67 68#include <sys/param.h> 69#include <sys/systm.h> 70#include <sys/limits.h> 71#include <sys/lock.h> 72#include <sys/mutex.h> 73#include <sys/proc.h> 74#include <sys/resourcevar.h> 75#include <sys/shm.h> 76#include <sys/vmmeter.h> 77#include <sys/sx.h> 78#include <sys/sysctl.h> 79 80#include <sys/kernel.h> 81#include <sys/ktr.h> 82#include <sys/unistd.h> 83 84#include <vm/vm.h> 85#include <vm/vm_param.h> 86#include <vm/pmap.h> 87#include <vm/vm_map.h> 88#include <vm/vm_page.h> 89#include <vm/vm_pageout.h> 90#include <vm/vm_object.h> 91#include <vm/vm_kern.h> 92#include <vm/vm_extern.h> 93#include <vm/vm_pager.h> 94#include <vm/swap_pager.h> 95 96#include <sys/user.h> 97 98extern int maxslp; 99 100/* 101 * System initialization 102 * 103 * Note: proc0 from proc.h 104 */ 105static void vm_init_limits(void *); 106SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0) 107 108/* 109 * THIS MUST BE THE LAST INITIALIZATION ITEM!!! 110 * 111 * Note: run scheduling should be divorced from the vm system. 112 */ 113static void scheduler(void *); 114SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL) 115 116#ifndef NO_SWAPPING 117static void swapout(struct proc *); 118static void vm_proc_swapin(struct proc *p); 119static void vm_proc_swapout(struct proc *p); 120#endif 121 122/* 123 * MPSAFE 124 * 125 * WARNING! This code calls vm_map_check_protection() which only checks 126 * the associated vm_map_entry range. It does not determine whether the 127 * contents of the memory is actually readable or writable. In most cases 128 * just checking the vm_map_entry is sufficient within the kernel's address 129 * space. 130 */ 131int 132kernacc(addr, len, rw) 133 void *addr; 134 int len, rw; 135{ 136 boolean_t rv; 137 vm_offset_t saddr, eaddr; 138 vm_prot_t prot; 139 140 KASSERT((rw & ~VM_PROT_ALL) == 0, 141 ("illegal ``rw'' argument to kernacc (%x)\n", rw)); 142 prot = rw; 143 saddr = trunc_page((vm_offset_t)addr); 144 eaddr = round_page((vm_offset_t)addr + len); 145 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot); 146 return (rv == TRUE); 147} 148 149/* 150 * MPSAFE 151 * 152 * WARNING! This code calls vm_map_check_protection() which only checks 153 * the associated vm_map_entry range. It does not determine whether the 154 * contents of the memory is actually readable or writable. vmapbuf(), 155 * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be 156 * used in conjuction with this call. 157 */ 158int 159useracc(addr, len, rw) 160 void *addr; 161 int len, rw; 162{ 163 boolean_t rv; 164 vm_prot_t prot; 165 vm_map_t map; 166 167 KASSERT((rw & ~VM_PROT_ALL) == 0, 168 ("illegal ``rw'' argument to useracc (%x)\n", rw)); 169 prot = rw; 170 map = &curproc->p_vmspace->vm_map; 171 if ((vm_offset_t)addr + len > vm_map_max(map) || 172 (vm_offset_t)addr + len < (vm_offset_t)addr) { 173 return (FALSE); 174 } 175 rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr), 176 round_page((vm_offset_t)addr + len), prot); 177 return (rv == TRUE); 178} 179 180/* 181 * MPSAFE 182 */ 183void 184vslock(addr, len) 185 void *addr; 186 u_int len; 187{ 188 189 vm_map_wire(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr), 190 round_page((vm_offset_t)addr + len), FALSE); 191} 192 193/* 194 * MPSAFE 195 */ 196void 197vsunlock(addr, len) 198 void *addr; 199 u_int len; 200{ 201 202 vm_map_unwire(&curproc->p_vmspace->vm_map, 203 trunc_page((vm_offset_t)addr), 204 round_page((vm_offset_t)addr + len), FALSE); 205} 206 207/* 208 * Create the U area for a new process. 209 * This routine directly affects the fork perf for a process. 210 */ 211void 212vm_proc_new(struct proc *p) 213{ 214 vm_page_t ma[UAREA_PAGES]; 215 vm_object_t upobj; 216 vm_offset_t up; 217 vm_page_t m; 218 u_int i; 219 220 /* 221 * Allocate object for the upage. 222 */ 223 upobj = vm_object_allocate(OBJT_DEFAULT, UAREA_PAGES); 224 p->p_upages_obj = upobj; 225 226 /* 227 * Get a kernel virtual address for the U area for this process. 228 */ 229 up = kmem_alloc_nofault(kernel_map, UAREA_PAGES * PAGE_SIZE); 230 if (up == 0) 231 panic("vm_proc_new: upage allocation failed"); 232 p->p_uarea = (struct user *)up; 233 234 for (i = 0; i < UAREA_PAGES; i++) { 235 /* 236 * Get a uarea page. 237 */ 238 m = vm_page_grab(upobj, i, 239 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED); 240 ma[i] = m; 241 242 vm_page_lock_queues(); 243 vm_page_wakeup(m); 244 vm_page_flag_clear(m, PG_ZERO); 245 m->valid = VM_PAGE_BITS_ALL; 246 vm_page_unlock_queues(); 247 } 248 249 /* 250 * Enter the pages into the kernel address space. 251 */ 252 pmap_qenter(up, ma, UAREA_PAGES); 253} 254 255/* 256 * Dispose the U area for a process that has exited. 257 * This routine directly impacts the exit perf of a process. 258 * XXX proc_zone is marked UMA_ZONE_NOFREE, so this should never be called. 259 */ 260void 261vm_proc_dispose(struct proc *p) 262{ 263 vm_object_t upobj; 264 vm_offset_t up; 265 vm_page_t m; 266 267 upobj = p->p_upages_obj; 268 VM_OBJECT_LOCK(upobj); 269 if (upobj->resident_page_count != UAREA_PAGES) 270 panic("vm_proc_dispose: incorrect number of pages in upobj"); 271 vm_page_lock_queues(); 272 while ((m = TAILQ_FIRST(&upobj->memq)) != NULL) { 273 vm_page_busy(m); 274 vm_page_unwire(m, 0); 275 vm_page_free(m); 276 } 277 vm_page_unlock_queues(); 278 VM_OBJECT_UNLOCK(upobj); 279 up = (vm_offset_t)p->p_uarea; 280 pmap_qremove(up, UAREA_PAGES); 281 kmem_free(kernel_map, up, UAREA_PAGES * PAGE_SIZE); 282 vm_object_deallocate(upobj); 283} 284 285#ifndef NO_SWAPPING 286/* 287 * Allow the U area for a process to be prejudicially paged out. 288 */ 289static void 290vm_proc_swapout(struct proc *p) 291{ 292 vm_object_t upobj; 293 vm_offset_t up; 294 vm_page_t m; 295 296 upobj = p->p_upages_obj; 297 VM_OBJECT_LOCK(upobj); 298 if (upobj->resident_page_count != UAREA_PAGES) 299 panic("vm_proc_dispose: incorrect number of pages in upobj"); 300 vm_page_lock_queues(); 301 TAILQ_FOREACH(m, &upobj->memq, listq) { 302 vm_page_dirty(m); 303 vm_page_unwire(m, 0); 304 } 305 vm_page_unlock_queues(); 306 VM_OBJECT_UNLOCK(upobj); 307 up = (vm_offset_t)p->p_uarea; 308 pmap_qremove(up, UAREA_PAGES); 309} 310 311/* 312 * Bring the U area for a specified process back in. 313 */ 314static void 315vm_proc_swapin(struct proc *p) 316{ 317 vm_page_t ma[UAREA_PAGES]; 318 vm_object_t upobj; 319 vm_offset_t up; 320 vm_page_t m; 321 int rv; 322 int i; 323 324 upobj = p->p_upages_obj; 325 VM_OBJECT_LOCK(upobj); 326 for (i = 0; i < UAREA_PAGES; i++) { 327 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 328 if (m->valid != VM_PAGE_BITS_ALL) { 329 rv = vm_pager_get_pages(upobj, &m, 1, 0); 330 if (rv != VM_PAGER_OK) 331 panic("vm_proc_swapin: cannot get upage"); 332 } 333 ma[i] = m; 334 } 335 if (upobj->resident_page_count != UAREA_PAGES) 336 panic("vm_proc_swapin: lost pages from upobj"); 337 vm_page_lock_queues(); 338 TAILQ_FOREACH(m, &upobj->memq, listq) { 339 m->valid = VM_PAGE_BITS_ALL; 340 vm_page_wire(m); 341 vm_page_wakeup(m); 342 } 343 vm_page_unlock_queues(); 344 VM_OBJECT_UNLOCK(upobj); 345 up = (vm_offset_t)p->p_uarea; 346 pmap_qenter(up, ma, UAREA_PAGES); 347} 348 349/* 350 * Swap in the UAREAs of all processes swapped out to the given device. 351 * The pages in the UAREA are marked dirty and their swap metadata is freed. 352 */ 353void 354vm_proc_swapin_all(int devidx) 355{ 356 struct proc *p; 357 vm_object_t object; 358 vm_page_t m; 359 360retry: 361 sx_slock(&allproc_lock); 362 FOREACH_PROC_IN_SYSTEM(p) { 363 PROC_LOCK(p); 364 object = p->p_upages_obj; 365 if (object != NULL) { 366 VM_OBJECT_LOCK(object); 367 if (swap_pager_isswapped(object, devidx)) { 368 VM_OBJECT_UNLOCK(object); 369 sx_sunlock(&allproc_lock); 370 faultin(p); 371 PROC_UNLOCK(p); 372 VM_OBJECT_LOCK(object); 373 vm_page_lock_queues(); 374 TAILQ_FOREACH(m, &object->memq, listq) 375 vm_page_dirty(m); 376 vm_page_unlock_queues(); 377 swap_pager_freespace(object, 0, 378 object->un_pager.swp.swp_bcount); 379 VM_OBJECT_UNLOCK(object); 380 goto retry; 381 } 382 VM_OBJECT_UNLOCK(object); 383 } 384 PROC_UNLOCK(p); 385 } 386 sx_sunlock(&allproc_lock); 387} 388#endif 389 390/* 391 * Implement fork's actions on an address space. 392 * Here we arrange for the address space to be copied or referenced, 393 * allocate a user struct (pcb and kernel stack), then call the 394 * machine-dependent layer to fill those in and make the new process 395 * ready to run. The new process is set up so that it returns directly 396 * to user mode to avoid stack copying and relocation problems. 397 */ 398void 399vm_forkproc(td, p2, td2, flags) 400 struct thread *td; 401 struct proc *p2; 402 struct thread *td2; 403 int flags; 404{ 405 struct proc *p1 = td->td_proc; 406 struct user *up; 407 408 GIANT_REQUIRED; 409 410 if ((flags & RFPROC) == 0) { 411 /* 412 * Divorce the memory, if it is shared, essentially 413 * this changes shared memory amongst threads, into 414 * COW locally. 415 */ 416 if ((flags & RFMEM) == 0) { 417 if (p1->p_vmspace->vm_refcnt > 1) { 418 vmspace_unshare(p1); 419 } 420 } 421 cpu_fork(td, p2, td2, flags); 422 return; 423 } 424 425 if (flags & RFMEM) { 426 p2->p_vmspace = p1->p_vmspace; 427 p1->p_vmspace->vm_refcnt++; 428 } 429 430 while (vm_page_count_severe()) { 431 VM_WAIT; 432 } 433 434 if ((flags & RFMEM) == 0) { 435 p2->p_vmspace = vmspace_fork(p1->p_vmspace); 436 437 pmap_pinit2(vmspace_pmap(p2->p_vmspace)); 438 439 if (p1->p_vmspace->vm_shm) 440 shmfork(p1, p2); 441 } 442 443 /* XXXKSE this is unsatisfactory but should be adequate */ 444 up = p2->p_uarea; 445 MPASS(p2->p_sigacts != NULL); 446 447 /* 448 * p_stats currently points at fields in the user struct 449 * but not at &u, instead at p_addr. Copy parts of 450 * p_stats; zero the rest of p_stats (statistics). 451 */ 452 p2->p_stats = &up->u_stats; 453 bzero(&up->u_stats.pstat_startzero, 454 (unsigned) ((caddr_t) &up->u_stats.pstat_endzero - 455 (caddr_t) &up->u_stats.pstat_startzero)); 456 bcopy(&p1->p_stats->pstat_startcopy, &up->u_stats.pstat_startcopy, 457 ((caddr_t) &up->u_stats.pstat_endcopy - 458 (caddr_t) &up->u_stats.pstat_startcopy)); 459 460 /* 461 * cpu_fork will copy and update the pcb, set up the kernel stack, 462 * and make the child ready to run. 463 */ 464 cpu_fork(td, p2, td2, flags); 465} 466 467/* 468 * Called after process has been wait(2)'ed apon and is being reaped. 469 * The idea is to reclaim resources that we could not reclaim while 470 * the process was still executing. 471 */ 472void 473vm_waitproc(p) 474 struct proc *p; 475{ 476 477 GIANT_REQUIRED; 478 vmspace_exitfree(p); /* and clean-out the vmspace */ 479} 480 481/* 482 * Set default limits for VM system. 483 * Called for proc 0, and then inherited by all others. 484 * 485 * XXX should probably act directly on proc0. 486 */ 487static void 488vm_init_limits(udata) 489 void *udata; 490{ 491 struct proc *p = udata; 492 int rss_limit; 493 494 /* 495 * Set up the initial limits on process VM. Set the maximum resident 496 * set size to be half of (reasonably) available memory. Since this 497 * is a soft limit, it comes into effect only when the system is out 498 * of memory - half of main memory helps to favor smaller processes, 499 * and reduces thrashing of the object cache. 500 */ 501 p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz; 502 p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz; 503 p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz; 504 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz; 505 /* limit the limit to no less than 2MB */ 506 rss_limit = max(cnt.v_free_count, 512); 507 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit); 508 p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY; 509} 510 511void 512faultin(p) 513 struct proc *p; 514{ 515#ifdef NO_SWAPPING 516 517 PROC_LOCK_ASSERT(p, MA_OWNED); 518 if ((p->p_sflag & PS_INMEM) == 0) 519 panic("faultin: proc swapped out with NO_SWAPPING!"); 520#else /* !NO_SWAPPING */ 521 struct thread *td; 522 523 GIANT_REQUIRED; 524 PROC_LOCK_ASSERT(p, MA_OWNED); 525 /* 526 * If another process is swapping in this process, 527 * just wait until it finishes. 528 */ 529 if (p->p_sflag & PS_SWAPPINGIN) 530 msleep(&p->p_sflag, &p->p_mtx, PVM, "faultin", 0); 531 else if ((p->p_sflag & PS_INMEM) == 0) { 532 /* 533 * Don't let another thread swap process p out while we are 534 * busy swapping it in. 535 */ 536 ++p->p_lock; 537 mtx_lock_spin(&sched_lock); 538 p->p_sflag |= PS_SWAPPINGIN; 539 mtx_unlock_spin(&sched_lock); 540 PROC_UNLOCK(p); 541 542 vm_proc_swapin(p); 543 FOREACH_THREAD_IN_PROC(p, td) 544 pmap_swapin_thread(td); 545 546 PROC_LOCK(p); 547 mtx_lock_spin(&sched_lock); 548 p->p_sflag &= ~PS_SWAPPINGIN; 549 p->p_sflag |= PS_INMEM; 550 FOREACH_THREAD_IN_PROC(p, td) { 551 TD_CLR_SWAPPED(td); 552 if (TD_CAN_RUN(td)) 553 setrunnable(td); 554 } 555 mtx_unlock_spin(&sched_lock); 556 557 wakeup(&p->p_sflag); 558 559 /* Allow other threads to swap p out now. */ 560 --p->p_lock; 561 } 562#endif /* NO_SWAPPING */ 563} 564 565/* 566 * This swapin algorithm attempts to swap-in processes only if there 567 * is enough space for them. Of course, if a process waits for a long 568 * time, it will be swapped in anyway. 569 * 570 * XXXKSE - process with the thread with highest priority counts.. 571 * 572 * Giant is still held at this point, to be released in tsleep. 573 */ 574/* ARGSUSED*/ 575static void 576scheduler(dummy) 577 void *dummy; 578{ 579 struct proc *p; 580 struct thread *td; 581 int pri; 582 struct proc *pp; 583 int ppri; 584 585 mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED); 586 /* GIANT_REQUIRED */ 587 588loop: 589 if (vm_page_count_min()) { 590 VM_WAIT; 591 goto loop; 592 } 593 594 pp = NULL; 595 ppri = INT_MIN; 596 sx_slock(&allproc_lock); 597 FOREACH_PROC_IN_SYSTEM(p) { 598 struct ksegrp *kg; 599 if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { 600 continue; 601 } 602 mtx_lock_spin(&sched_lock); 603 FOREACH_THREAD_IN_PROC(p, td) { 604 /* 605 * An otherwise runnable thread of a process 606 * swapped out has only the TDI_SWAPPED bit set. 607 * 608 */ 609 if (td->td_inhibitors == TDI_SWAPPED) { 610 kg = td->td_ksegrp; 611 pri = p->p_swtime + kg->kg_slptime; 612 if ((p->p_sflag & PS_SWAPINREQ) == 0) { 613 pri -= kg->kg_nice * 8; 614 } 615 616 /* 617 * if this ksegrp is higher priority 618 * and there is enough space, then select 619 * this process instead of the previous 620 * selection. 621 */ 622 if (pri > ppri) { 623 pp = p; 624 ppri = pri; 625 } 626 } 627 } 628 mtx_unlock_spin(&sched_lock); 629 } 630 sx_sunlock(&allproc_lock); 631 632 /* 633 * Nothing to do, back to sleep. 634 */ 635 if ((p = pp) == NULL) { 636 tsleep(&proc0, PVM, "sched", maxslp * hz / 2); 637 goto loop; 638 } 639 PROC_LOCK(p); 640 641 /* 642 * Another process may be bringing or may have already 643 * brought this process in while we traverse all threads. 644 * Or, this process may even be being swapped out again. 645 */ 646 if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { 647 PROC_UNLOCK(p); 648 goto loop; 649 } 650 651 mtx_lock_spin(&sched_lock); 652 p->p_sflag &= ~PS_SWAPINREQ; 653 mtx_unlock_spin(&sched_lock); 654 655 /* 656 * We would like to bring someone in. (only if there is space). 657 * [What checks the space? ] 658 */ 659 faultin(p); 660 PROC_UNLOCK(p); 661 mtx_lock_spin(&sched_lock); 662 p->p_swtime = 0; 663 mtx_unlock_spin(&sched_lock); 664 goto loop; 665} 666 667#ifndef NO_SWAPPING 668 669/* 670 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 671 */ 672static int swap_idle_threshold1 = 2; 673SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, 674 &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); 675 676/* 677 * Swap_idle_threshold2 is the time that a process can be idle before 678 * it will be swapped out, if idle swapping is enabled. 679 */ 680static int swap_idle_threshold2 = 10; 681SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, 682 &swap_idle_threshold2, 0, "Time before a process will be swapped out"); 683 684/* 685 * Swapout is driven by the pageout daemon. Very simple, we find eligible 686 * procs and unwire their u-areas. We try to always "swap" at least one 687 * process in case we need the room for a swapin. 688 * If any procs have been sleeping/stopped for at least maxslp seconds, 689 * they are swapped. Else, we swap the longest-sleeping or stopped process, 690 * if any, otherwise the longest-resident process. 691 */ 692void 693swapout_procs(action) 694int action; 695{ 696 struct proc *p; 697 struct thread *td; 698 struct ksegrp *kg; 699 int didswap = 0; 700 701 GIANT_REQUIRED; 702 703retry: 704 sx_slock(&allproc_lock); 705 FOREACH_PROC_IN_SYSTEM(p) { 706 struct vmspace *vm; 707 int minslptime = 100000; 708 709 /* 710 * Watch out for a process in 711 * creation. It may have no 712 * address space or lock yet. 713 */ 714 mtx_lock_spin(&sched_lock); 715 if (p->p_state == PRS_NEW) { 716 mtx_unlock_spin(&sched_lock); 717 continue; 718 } 719 mtx_unlock_spin(&sched_lock); 720 721 /* 722 * An aio daemon switches its 723 * address space while running. 724 * Perform a quick check whether 725 * a process has P_SYSTEM. 726 */ 727 if ((p->p_flag & P_SYSTEM) != 0) 728 continue; 729 730 /* 731 * Do not swapout a process that 732 * is waiting for VM data 733 * structures as there is a possible 734 * deadlock. Test this first as 735 * this may block. 736 * 737 * Lock the map until swapout 738 * finishes, or a thread of this 739 * process may attempt to alter 740 * the map. 741 */ 742 PROC_LOCK(p); 743 vm = p->p_vmspace; 744 KASSERT(vm != NULL, 745 ("swapout_procs: a process has no address space")); 746 ++vm->vm_refcnt; 747 PROC_UNLOCK(p); 748 if (!vm_map_trylock(&vm->vm_map)) 749 goto nextproc1; 750 751 PROC_LOCK(p); 752 if (p->p_lock != 0 || 753 (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT) 754 ) != 0) { 755 goto nextproc2; 756 } 757 /* 758 * only aiod changes vmspace, however it will be 759 * skipped because of the if statement above checking 760 * for P_SYSTEM 761 */ 762 if ((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) != PS_INMEM) 763 goto nextproc2; 764 765 switch (p->p_state) { 766 default: 767 /* Don't swap out processes in any sort 768 * of 'special' state. */ 769 break; 770 771 case PRS_NORMAL: 772 mtx_lock_spin(&sched_lock); 773 /* 774 * do not swapout a realtime process 775 * Check all the thread groups.. 776 */ 777 FOREACH_KSEGRP_IN_PROC(p, kg) { 778 if (PRI_IS_REALTIME(kg->kg_pri_class)) 779 goto nextproc; 780 781 /* 782 * Guarantee swap_idle_threshold1 783 * time in memory. 784 */ 785 if (kg->kg_slptime < swap_idle_threshold1) 786 goto nextproc; 787 788 /* 789 * Do not swapout a process if it is 790 * waiting on a critical event of some 791 * kind or there is a thread whose 792 * pageable memory may be accessed. 793 * 794 * This could be refined to support 795 * swapping out a thread. 796 */ 797 FOREACH_THREAD_IN_GROUP(kg, td) { 798 if ((td->td_priority) < PSOCK || 799 !thread_safetoswapout(td)) 800 goto nextproc; 801 } 802 /* 803 * If the system is under memory stress, 804 * or if we are swapping 805 * idle processes >= swap_idle_threshold2, 806 * then swap the process out. 807 */ 808 if (((action & VM_SWAP_NORMAL) == 0) && 809 (((action & VM_SWAP_IDLE) == 0) || 810 (kg->kg_slptime < swap_idle_threshold2))) 811 goto nextproc; 812 813 if (minslptime > kg->kg_slptime) 814 minslptime = kg->kg_slptime; 815 } 816 817 /* 818 * If the process has been asleep for awhile and had 819 * most of its pages taken away already, swap it out. 820 */ 821 if ((action & VM_SWAP_NORMAL) || 822 ((action & VM_SWAP_IDLE) && 823 (minslptime > swap_idle_threshold2))) { 824 swapout(p); 825 didswap++; 826 mtx_unlock_spin(&sched_lock); 827 PROC_UNLOCK(p); 828 vm_map_unlock(&vm->vm_map); 829 vmspace_free(vm); 830 sx_sunlock(&allproc_lock); 831 goto retry; 832 } 833nextproc: 834 mtx_unlock_spin(&sched_lock); 835 } 836nextproc2: 837 PROC_UNLOCK(p); 838 vm_map_unlock(&vm->vm_map); 839nextproc1: 840 vmspace_free(vm); 841 continue; 842 } 843 sx_sunlock(&allproc_lock); 844 /* 845 * If we swapped something out, and another process needed memory, 846 * then wakeup the sched process. 847 */ 848 if (didswap) 849 wakeup(&proc0); 850} 851 852static void 853swapout(p) 854 struct proc *p; 855{ 856 struct thread *td; 857 858 PROC_LOCK_ASSERT(p, MA_OWNED); 859 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 860#if defined(SWAP_DEBUG) 861 printf("swapping out %d\n", p->p_pid); 862#endif 863 864 /* 865 * The states of this process and its threads may have changed 866 * by now. Assuming that there is only one pageout daemon thread, 867 * this process should still be in memory. 868 */ 869 KASSERT((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) == PS_INMEM, 870 ("swapout: lost a swapout race?")); 871 872#if defined(INVARIANTS) 873 /* 874 * Make sure that all threads are safe to be swapped out. 875 * 876 * Alternatively, we could swap out only safe threads. 877 */ 878 FOREACH_THREAD_IN_PROC(p, td) { 879 KASSERT(thread_safetoswapout(td), 880 ("swapout: there is a thread not safe for swapout")); 881 } 882#endif /* INVARIANTS */ 883 884 ++p->p_stats->p_ru.ru_nswap; 885 /* 886 * remember the process resident count 887 */ 888 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 889 890 p->p_sflag &= ~PS_INMEM; 891 p->p_sflag |= PS_SWAPPINGOUT; 892 PROC_UNLOCK(p); 893 FOREACH_THREAD_IN_PROC(p, td) 894 TD_SET_SWAPPED(td); 895 mtx_unlock_spin(&sched_lock); 896 897 vm_proc_swapout(p); 898 FOREACH_THREAD_IN_PROC(p, td) 899 pmap_swapout_thread(td); 900 901 PROC_LOCK(p); 902 mtx_lock_spin(&sched_lock); 903 p->p_sflag &= ~PS_SWAPPINGOUT; 904 p->p_swtime = 0; 905} 906#endif /* !NO_SWAPPING */ 907