vm_glue.c revision 113918
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 * $FreeBSD: head/sys/vm/vm_glue.c 113918 2003-04-23 18:21:41Z jhb $ 63 */ 64 65#include "opt_vm.h" 66 67#include <sys/param.h> 68#include <sys/systm.h> 69#include <sys/lock.h> 70#include <sys/mutex.h> 71#include <sys/proc.h> 72#include <sys/resourcevar.h> 73#include <sys/shm.h> 74#include <sys/vmmeter.h> 75#include <sys/sx.h> 76#include <sys/sysctl.h> 77 78#include <sys/kernel.h> 79#include <sys/ktr.h> 80#include <sys/unistd.h> 81 82#include <machine/limits.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 if (upobj->resident_page_count != UAREA_PAGES) 269 panic("vm_proc_dispose: incorrect number of pages in upobj"); 270 vm_page_lock_queues(); 271 while ((m = TAILQ_FIRST(&upobj->memq)) != NULL) { 272 vm_page_busy(m); 273 vm_page_unwire(m, 0); 274 vm_page_free(m); 275 } 276 vm_page_unlock_queues(); 277 up = (vm_offset_t)p->p_uarea; 278 pmap_qremove(up, UAREA_PAGES); 279 kmem_free(kernel_map, up, UAREA_PAGES * PAGE_SIZE); 280 vm_object_deallocate(upobj); 281} 282 283#ifndef NO_SWAPPING 284/* 285 * Allow the U area for a process to be prejudicially paged out. 286 */ 287static void 288vm_proc_swapout(struct proc *p) 289{ 290 vm_object_t upobj; 291 vm_offset_t up; 292 vm_page_t m; 293 294 upobj = p->p_upages_obj; 295 if (upobj->resident_page_count != UAREA_PAGES) 296 panic("vm_proc_dispose: incorrect number of pages in upobj"); 297 vm_page_lock_queues(); 298 TAILQ_FOREACH(m, &upobj->memq, listq) { 299 vm_page_dirty(m); 300 vm_page_unwire(m, 0); 301 } 302 vm_page_unlock_queues(); 303 up = (vm_offset_t)p->p_uarea; 304 pmap_qremove(up, UAREA_PAGES); 305} 306 307/* 308 * Bring the U area for a specified process back in. 309 */ 310static void 311vm_proc_swapin(struct proc *p) 312{ 313 vm_page_t ma[UAREA_PAGES]; 314 vm_object_t upobj; 315 vm_offset_t up; 316 vm_page_t m; 317 int rv; 318 int i; 319 320 upobj = p->p_upages_obj; 321 for (i = 0; i < UAREA_PAGES; i++) { 322 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 323 if (m->valid != VM_PAGE_BITS_ALL) { 324 rv = vm_pager_get_pages(upobj, &m, 1, 0); 325 if (rv != VM_PAGER_OK) 326 panic("vm_proc_swapin: cannot get upage"); 327 } 328 ma[i] = m; 329 } 330 if (upobj->resident_page_count != UAREA_PAGES) 331 panic("vm_proc_swapin: lost pages from upobj"); 332 vm_page_lock_queues(); 333 TAILQ_FOREACH(m, &upobj->memq, listq) { 334 m->valid = VM_PAGE_BITS_ALL; 335 vm_page_wire(m); 336 vm_page_wakeup(m); 337 } 338 vm_page_unlock_queues(); 339 up = (vm_offset_t)p->p_uarea; 340 pmap_qenter(up, ma, UAREA_PAGES); 341} 342 343/* 344 * Swap in the UAREAs of all processes swapped out to the given device. 345 * The pages in the UAREA are marked dirty and their swap metadata is freed. 346 */ 347void 348vm_proc_swapin_all(int devidx) 349{ 350 struct proc *p; 351 vm_object_t object; 352 vm_page_t m; 353 354retry: 355 sx_slock(&allproc_lock); 356 FOREACH_PROC_IN_SYSTEM(p) { 357 PROC_LOCK(p); 358 359 object = p->p_upages_obj; 360 if (object != NULL && 361 swap_pager_isswapped(p->p_upages_obj, devidx)) { 362 sx_sunlock(&allproc_lock); 363 faultin(p); 364 PROC_UNLOCK(p); 365 vm_page_lock_queues(); 366 TAILQ_FOREACH(m, &object->memq, listq) 367 vm_page_dirty(m); 368 vm_page_unlock_queues(); 369 swap_pager_freespace(object, 0, 370 object->un_pager.swp.swp_bcount); 371 goto retry; 372 } 373 374 PROC_UNLOCK(p); 375 } 376 sx_sunlock(&allproc_lock); 377} 378#endif 379 380/* 381 * Implement fork's actions on an address space. 382 * Here we arrange for the address space to be copied or referenced, 383 * allocate a user struct (pcb and kernel stack), then call the 384 * machine-dependent layer to fill those in and make the new process 385 * ready to run. The new process is set up so that it returns directly 386 * to user mode to avoid stack copying and relocation problems. 387 */ 388void 389vm_forkproc(td, p2, td2, flags) 390 struct thread *td; 391 struct proc *p2; 392 struct thread *td2; 393 int flags; 394{ 395 struct proc *p1 = td->td_proc; 396 struct user *up; 397 398 GIANT_REQUIRED; 399 400 if ((flags & RFPROC) == 0) { 401 /* 402 * Divorce the memory, if it is shared, essentially 403 * this changes shared memory amongst threads, into 404 * COW locally. 405 */ 406 if ((flags & RFMEM) == 0) { 407 if (p1->p_vmspace->vm_refcnt > 1) { 408 vmspace_unshare(p1); 409 } 410 } 411 cpu_fork(td, p2, td2, flags); 412 return; 413 } 414 415 if (flags & RFMEM) { 416 p2->p_vmspace = p1->p_vmspace; 417 p1->p_vmspace->vm_refcnt++; 418 } 419 420 while (vm_page_count_severe()) { 421 VM_WAIT; 422 } 423 424 if ((flags & RFMEM) == 0) { 425 p2->p_vmspace = vmspace_fork(p1->p_vmspace); 426 427 pmap_pinit2(vmspace_pmap(p2->p_vmspace)); 428 429 if (p1->p_vmspace->vm_shm) 430 shmfork(p1, p2); 431 } 432 433 /* XXXKSE this is unsatisfactory but should be adequate */ 434 up = p2->p_uarea; 435 436 /* 437 * p_stats currently points at fields in the user struct 438 * but not at &u, instead at p_addr. Copy parts of 439 * p_stats; zero the rest of p_stats (statistics). 440 * 441 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need 442 * to share sigacts, so we use the up->u_sigacts. 443 */ 444 p2->p_stats = &up->u_stats; 445 if (p2->p_sigacts == NULL) { 446 if (p2->p_procsig->ps_refcnt != 1) 447 printf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid); 448 p2->p_sigacts = &up->u_sigacts; 449 up->u_sigacts = *p1->p_sigacts; 450 } 451 452 bzero(&up->u_stats.pstat_startzero, 453 (unsigned) ((caddr_t) &up->u_stats.pstat_endzero - 454 (caddr_t) &up->u_stats.pstat_startzero)); 455 bcopy(&p1->p_stats->pstat_startcopy, &up->u_stats.pstat_startcopy, 456 ((caddr_t) &up->u_stats.pstat_endcopy - 457 (caddr_t) &up->u_stats.pstat_startcopy)); 458 459 460 /* 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 cpu_wait(p); 479 vmspace_exitfree(p); /* and clean-out the vmspace */ 480} 481 482/* 483 * Set default limits for VM system. 484 * Called for proc 0, and then inherited by all others. 485 * 486 * XXX should probably act directly on proc0. 487 */ 488static void 489vm_init_limits(udata) 490 void *udata; 491{ 492 struct proc *p = udata; 493 int rss_limit; 494 495 /* 496 * Set up the initial limits on process VM. Set the maximum resident 497 * set size to be half of (reasonably) available memory. Since this 498 * is a soft limit, it comes into effect only when the system is out 499 * of memory - half of main memory helps to favor smaller processes, 500 * and reduces thrashing of the object cache. 501 */ 502 p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz; 503 p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz; 504 p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz; 505 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz; 506 /* limit the limit to no less than 2MB */ 507 rss_limit = max(cnt.v_free_count, 512); 508 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit); 509 p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY; 510} 511 512void 513faultin(p) 514 struct proc *p; 515{ 516#ifdef NO_SWAPPING 517 518 PROC_LOCK_ASSERT(p, MA_OWNED); 519 if ((p->p_sflag & PS_INMEM) == 0) 520 panic("faultin: proc swapped out with NO_SWAPPING!"); 521#else /* !NO_SWAPPING */ 522 struct thread *td; 523 524 GIANT_REQUIRED; 525 PROC_LOCK_ASSERT(p, MA_OWNED); 526 /* 527 * If another process is swapping in this process, 528 * just wait until it finishes. 529 */ 530 if (p->p_sflag & PS_SWAPPINGIN) 531 msleep(&p->p_sflag, &p->p_mtx, PVM, "faultin", 0); 532 else if ((p->p_sflag & PS_INMEM) == 0) { 533 /* 534 * Don't let another thread swap process p out while we are 535 * busy swapping it in. 536 */ 537 ++p->p_lock; 538 mtx_lock_spin(&sched_lock); 539 p->p_sflag |= PS_SWAPPINGIN; 540 mtx_unlock_spin(&sched_lock); 541 PROC_UNLOCK(p); 542 543 vm_proc_swapin(p); 544 FOREACH_THREAD_IN_PROC(p, td) 545 pmap_swapin_thread(td); 546 547 PROC_LOCK(p); 548 mtx_lock_spin(&sched_lock); 549 p->p_sflag &= ~PS_SWAPPINGIN; 550 p->p_sflag |= PS_INMEM; 551 FOREACH_THREAD_IN_PROC(p, td) { 552 TD_CLR_SWAPPED(td); 553 if (TD_CAN_RUN(td)) 554 setrunnable(td); 555 } 556 mtx_unlock_spin(&sched_lock); 557 558 wakeup(&p->p_sflag); 559 560 /* Allow other threads to swap p out now. */ 561 --p->p_lock; 562 } 563#endif /* NO_SWAPPING */ 564} 565 566/* 567 * This swapin algorithm attempts to swap-in processes only if there 568 * is enough space for them. Of course, if a process waits for a long 569 * time, it will be swapped in anyway. 570 * 571 * XXXKSE - process with the thread with highest priority counts.. 572 * 573 * Giant is still held at this point, to be released in tsleep. 574 */ 575/* ARGSUSED*/ 576static void 577scheduler(dummy) 578 void *dummy; 579{ 580 struct proc *p; 581 struct thread *td; 582 int pri; 583 struct proc *pp; 584 int ppri; 585 586 mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED); 587 /* GIANT_REQUIRED */ 588 589loop: 590 if (vm_page_count_min()) { 591 VM_WAIT; 592 goto loop; 593 } 594 595 pp = NULL; 596 ppri = INT_MIN; 597 sx_slock(&allproc_lock); 598 FOREACH_PROC_IN_SYSTEM(p) { 599 struct ksegrp *kg; 600 if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { 601 continue; 602 } 603 mtx_lock_spin(&sched_lock); 604 FOREACH_THREAD_IN_PROC(p, td) { 605 /* 606 * An otherwise runnable thread of a process 607 * swapped out has only the TDI_SWAPPED bit set. 608 * 609 */ 610 if (td->td_inhibitors == TDI_SWAPPED) { 611 kg = td->td_ksegrp; 612 pri = p->p_swtime + kg->kg_slptime; 613 if ((p->p_sflag & PS_SWAPINREQ) == 0) { 614 pri -= kg->kg_nice * 8; 615 } 616 617 /* 618 * if this ksegrp is higher priority 619 * and there is enough space, then select 620 * this process instead of the previous 621 * selection. 622 */ 623 if (pri > ppri) { 624 pp = p; 625 ppri = pri; 626 } 627 } 628 } 629 mtx_unlock_spin(&sched_lock); 630 } 631 sx_sunlock(&allproc_lock); 632 633 /* 634 * Nothing to do, back to sleep. 635 */ 636 if ((p = pp) == NULL) { 637 tsleep(&proc0, PVM, "sched", maxslp * hz / 2); 638 goto loop; 639 } 640 PROC_LOCK(p); 641 642 /* 643 * Another process may be bringing or may have already 644 * brought this process in while we traverse all threads. 645 * Or, this process may even be being swapped out again. 646 */ 647 if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { 648 PROC_UNLOCK(p); 649 goto loop; 650 } 651 652 mtx_lock_spin(&sched_lock); 653 p->p_sflag &= ~PS_SWAPINREQ; 654 mtx_unlock_spin(&sched_lock); 655 656 /* 657 * We would like to bring someone in. (only if there is space). 658 * [What checks the space? ] 659 */ 660 faultin(p); 661 PROC_UNLOCK(p); 662 mtx_lock_spin(&sched_lock); 663 p->p_swtime = 0; 664 mtx_unlock_spin(&sched_lock); 665 goto loop; 666} 667 668#ifndef NO_SWAPPING 669 670/* 671 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 672 */ 673static int swap_idle_threshold1 = 2; 674SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, 675 &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); 676 677/* 678 * Swap_idle_threshold2 is the time that a process can be idle before 679 * it will be swapped out, if idle swapping is enabled. 680 */ 681static int swap_idle_threshold2 = 10; 682SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, 683 &swap_idle_threshold2, 0, "Time before a process will be swapped out"); 684 685/* 686 * Swapout is driven by the pageout daemon. Very simple, we find eligible 687 * procs and unwire their u-areas. We try to always "swap" at least one 688 * process in case we need the room for a swapin. 689 * If any procs have been sleeping/stopped for at least maxslp seconds, 690 * they are swapped. Else, we swap the longest-sleeping or stopped process, 691 * if any, otherwise the longest-resident process. 692 */ 693void 694swapout_procs(action) 695int action; 696{ 697 struct proc *p; 698 struct thread *td; 699 struct ksegrp *kg; 700 struct proc *outp, *outp2; 701 int outpri, outpri2; 702 int didswap = 0; 703 704 GIANT_REQUIRED; 705 706 outp = outp2 = NULL; 707 outpri = outpri2 = INT_MIN; 708retry: 709 sx_slock(&allproc_lock); 710 FOREACH_PROC_IN_SYSTEM(p) { 711 struct vmspace *vm; 712 int minslptime = 100000; 713 714 /* 715 * Watch out for a process in 716 * creation. It may have no 717 * address space or lock yet. 718 */ 719 mtx_lock_spin(&sched_lock); 720 if (p->p_state == PRS_NEW) { 721 mtx_unlock_spin(&sched_lock); 722 continue; 723 } 724 mtx_unlock_spin(&sched_lock); 725 726 /* 727 * An aio daemon switches its 728 * address space while running. 729 * Perform a quick check whether 730 * a process has P_SYSTEM. 731 */ 732 PROC_LOCK(p); 733 if ((p->p_flag & P_SYSTEM) != 0) { 734 PROC_UNLOCK(p); 735 continue; 736 } 737 738 /* 739 * Do not swapout a process that 740 * is waiting for VM data 741 * structures as there is a possible 742 * deadlock. Test this first as 743 * this may block. 744 * 745 * Lock the map until swapout 746 * finishes, or a thread of this 747 * process may attempt to alter 748 * the map. 749 */ 750 vm = p->p_vmspace; 751 KASSERT(vm != NULL, 752 ("swapout_procs: a process has no address space")); 753 ++vm->vm_refcnt; 754 PROC_UNLOCK(p); 755 if (!vm_map_trylock(&vm->vm_map)) 756 goto nextproc1; 757 758 PROC_LOCK(p); 759 if (p->p_lock != 0 || 760 (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT) 761 ) != 0) { 762 goto nextproc2; 763 } 764 /* 765 * only aiod changes vmspace, however it will be 766 * skipped because of the if statement above checking 767 * for P_SYSTEM 768 */ 769 if ((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) != PS_INMEM) 770 goto nextproc2; 771 772 mtx_lock_spin(&sched_lock); 773 switch (p->p_state) { 774 default: 775 /* Don't swap out processes in any sort 776 * of 'special' state. */ 777 goto nextproc; 778 779 case PRS_NORMAL: 780 /* 781 * do not swapout a realtime process 782 * Check all the thread groups.. 783 */ 784 FOREACH_KSEGRP_IN_PROC(p, kg) { 785 if (PRI_IS_REALTIME(kg->kg_pri_class)) 786 goto nextproc; 787 788 /* 789 * Guarantee swap_idle_threshold1 790 * time in memory. 791 */ 792 if (kg->kg_slptime < swap_idle_threshold1) 793 goto nextproc; 794 795 /* 796 * Do not swapout a process if it is 797 * waiting on a critical event of some 798 * kind or there is a thread whose 799 * pageable memory may be accessed. 800 * 801 * This could be refined to support 802 * swapping out a thread. 803 */ 804 FOREACH_THREAD_IN_GROUP(kg, td) { 805 if ((td->td_priority) < PSOCK || 806 !thread_safetoswapout(td)) 807 goto nextproc; 808 } 809 /* 810 * If the system is under memory stress, 811 * or if we are swapping 812 * idle processes >= swap_idle_threshold2, 813 * then swap the process out. 814 */ 815 if (((action & VM_SWAP_NORMAL) == 0) && 816 (((action & VM_SWAP_IDLE) == 0) || 817 (kg->kg_slptime < swap_idle_threshold2))) 818 goto nextproc; 819 820 if (minslptime > kg->kg_slptime) 821 minslptime = kg->kg_slptime; 822 } 823 824 /* 825 * If the process has been asleep for awhile and had 826 * most of its pages taken away already, swap it out. 827 */ 828 if ((action & VM_SWAP_NORMAL) || 829 ((action & VM_SWAP_IDLE) && 830 (minslptime > swap_idle_threshold2))) { 831 swapout(p); 832 didswap++; 833 mtx_unlock_spin(&sched_lock); 834 PROC_UNLOCK(p); 835 vm_map_unlock(&vm->vm_map); 836 vmspace_free(vm); 837 sx_sunlock(&allproc_lock); 838 goto retry; 839 } 840 } 841nextproc: 842 mtx_unlock_spin(&sched_lock); 843nextproc2: 844 PROC_UNLOCK(p); 845 vm_map_unlock(&vm->vm_map); 846nextproc1: 847 vmspace_free(vm); 848 continue; 849 } 850 sx_sunlock(&allproc_lock); 851 /* 852 * If we swapped something out, and another process needed memory, 853 * then wakeup the sched process. 854 */ 855 if (didswap) 856 wakeup(&proc0); 857} 858 859static void 860swapout(p) 861 struct proc *p; 862{ 863 struct thread *td; 864 865 PROC_LOCK_ASSERT(p, MA_OWNED); 866 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 867#if defined(SWAP_DEBUG) 868 printf("swapping out %d\n", p->p_pid); 869#endif 870 871 /* 872 * The states of this process and its threads may have changed 873 * by now. Assuming that there is only one pageout daemon thread, 874 * this process should still be in memory. 875 */ 876 KASSERT((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) == PS_INMEM, 877 ("swapout: lost a swapout race?")); 878 879#if defined(INVARIANTS) 880 /* 881 * Make sure that all threads are safe to be swapped out. 882 * 883 * Alternatively, we could swap out only safe threads. 884 */ 885 FOREACH_THREAD_IN_PROC(p, td) { 886 KASSERT(thread_safetoswapout(td), 887 ("swapout: there is a thread not safe for swapout")); 888 } 889#endif /* INVARIANTS */ 890 891 ++p->p_stats->p_ru.ru_nswap; 892 /* 893 * remember the process resident count 894 */ 895 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 896 897 p->p_sflag &= ~PS_INMEM; 898 p->p_sflag |= PS_SWAPPINGOUT; 899 PROC_UNLOCK(p); 900 FOREACH_THREAD_IN_PROC(p, td) 901 TD_SET_SWAPPED(td); 902 mtx_unlock_spin(&sched_lock); 903 904 vm_proc_swapout(p); 905 FOREACH_THREAD_IN_PROC(p, td) 906 pmap_swapout_thread(td); 907 908 PROC_LOCK(p); 909 mtx_lock_spin(&sched_lock); 910 p->p_sflag &= ~PS_SWAPPINGOUT; 911 p->p_swtime = 0; 912} 913#endif /* !NO_SWAPPING */ 914