vm_glue.c revision 116328
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 116328 2003-06-14 06:20:25Z 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 * Set up a variable-sized alternate kstack. 392 */ 393void 394vm_thread_new_altkstack(struct thread *td, int pages) 395{ 396 397 td->td_altkstack = td->td_kstack; 398 td->td_altkstack_obj = td->td_kstack_obj; 399 td->td_altkstack_pages = td->td_kstack_pages; 400 401 pmap_new_thread(td, pages); 402} 403 404/* 405 * Restore the original kstack. 406 */ 407void 408vm_thread_dispose_altkstack(struct thread *td) 409{ 410 411 pmap_dispose_thread(td); 412 413 td->td_kstack = td->td_altkstack; 414 td->td_kstack_obj = td->td_altkstack_obj; 415 td->td_kstack_pages = td->td_altkstack_pages; 416 td->td_altkstack = 0; 417 td->td_altkstack_obj = NULL; 418 td->td_altkstack_pages = 0; 419} 420 421/* 422 * Implement fork's actions on an address space. 423 * Here we arrange for the address space to be copied or referenced, 424 * allocate a user struct (pcb and kernel stack), then call the 425 * machine-dependent layer to fill those in and make the new process 426 * ready to run. The new process is set up so that it returns directly 427 * to user mode to avoid stack copying and relocation problems. 428 */ 429void 430vm_forkproc(td, p2, td2, flags) 431 struct thread *td; 432 struct proc *p2; 433 struct thread *td2; 434 int flags; 435{ 436 struct proc *p1 = td->td_proc; 437 struct user *up; 438 439 GIANT_REQUIRED; 440 441 if ((flags & RFPROC) == 0) { 442 /* 443 * Divorce the memory, if it is shared, essentially 444 * this changes shared memory amongst threads, into 445 * COW locally. 446 */ 447 if ((flags & RFMEM) == 0) { 448 if (p1->p_vmspace->vm_refcnt > 1) { 449 vmspace_unshare(p1); 450 } 451 } 452 cpu_fork(td, p2, td2, flags); 453 return; 454 } 455 456 if (flags & RFMEM) { 457 p2->p_vmspace = p1->p_vmspace; 458 p1->p_vmspace->vm_refcnt++; 459 } 460 461 while (vm_page_count_severe()) { 462 VM_WAIT; 463 } 464 465 if ((flags & RFMEM) == 0) { 466 p2->p_vmspace = vmspace_fork(p1->p_vmspace); 467 468 pmap_pinit2(vmspace_pmap(p2->p_vmspace)); 469 470 if (p1->p_vmspace->vm_shm) 471 shmfork(p1, p2); 472 } 473 474 /* XXXKSE this is unsatisfactory but should be adequate */ 475 up = p2->p_uarea; 476 MPASS(p2->p_sigacts != NULL); 477 478 /* 479 * p_stats currently points at fields in the user struct 480 * but not at &u, instead at p_addr. Copy parts of 481 * p_stats; zero the rest of p_stats (statistics). 482 */ 483 p2->p_stats = &up->u_stats; 484 bzero(&up->u_stats.pstat_startzero, 485 (unsigned) ((caddr_t) &up->u_stats.pstat_endzero - 486 (caddr_t) &up->u_stats.pstat_startzero)); 487 bcopy(&p1->p_stats->pstat_startcopy, &up->u_stats.pstat_startcopy, 488 ((caddr_t) &up->u_stats.pstat_endcopy - 489 (caddr_t) &up->u_stats.pstat_startcopy)); 490 491 /* 492 * cpu_fork will copy and update the pcb, set up the kernel stack, 493 * and make the child ready to run. 494 */ 495 cpu_fork(td, p2, td2, flags); 496} 497 498/* 499 * Called after process has been wait(2)'ed apon and is being reaped. 500 * The idea is to reclaim resources that we could not reclaim while 501 * the process was still executing. 502 */ 503void 504vm_waitproc(p) 505 struct proc *p; 506{ 507 508 GIANT_REQUIRED; 509 vmspace_exitfree(p); /* and clean-out the vmspace */ 510} 511 512/* 513 * Set default limits for VM system. 514 * Called for proc 0, and then inherited by all others. 515 * 516 * XXX should probably act directly on proc0. 517 */ 518static void 519vm_init_limits(udata) 520 void *udata; 521{ 522 struct proc *p = udata; 523 int rss_limit; 524 525 /* 526 * Set up the initial limits on process VM. Set the maximum resident 527 * set size to be half of (reasonably) available memory. Since this 528 * is a soft limit, it comes into effect only when the system is out 529 * of memory - half of main memory helps to favor smaller processes, 530 * and reduces thrashing of the object cache. 531 */ 532 p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz; 533 p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz; 534 p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz; 535 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz; 536 /* limit the limit to no less than 2MB */ 537 rss_limit = max(cnt.v_free_count, 512); 538 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit); 539 p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY; 540} 541 542void 543faultin(p) 544 struct proc *p; 545{ 546#ifdef NO_SWAPPING 547 548 PROC_LOCK_ASSERT(p, MA_OWNED); 549 if ((p->p_sflag & PS_INMEM) == 0) 550 panic("faultin: proc swapped out with NO_SWAPPING!"); 551#else /* !NO_SWAPPING */ 552 struct thread *td; 553 554 GIANT_REQUIRED; 555 PROC_LOCK_ASSERT(p, MA_OWNED); 556 /* 557 * If another process is swapping in this process, 558 * just wait until it finishes. 559 */ 560 if (p->p_sflag & PS_SWAPPINGIN) 561 msleep(&p->p_sflag, &p->p_mtx, PVM, "faultin", 0); 562 else if ((p->p_sflag & PS_INMEM) == 0) { 563 /* 564 * Don't let another thread swap process p out while we are 565 * busy swapping it in. 566 */ 567 ++p->p_lock; 568 mtx_lock_spin(&sched_lock); 569 p->p_sflag |= PS_SWAPPINGIN; 570 mtx_unlock_spin(&sched_lock); 571 PROC_UNLOCK(p); 572 573 vm_proc_swapin(p); 574 FOREACH_THREAD_IN_PROC(p, td) 575 pmap_swapin_thread(td); 576 577 PROC_LOCK(p); 578 mtx_lock_spin(&sched_lock); 579 p->p_sflag &= ~PS_SWAPPINGIN; 580 p->p_sflag |= PS_INMEM; 581 FOREACH_THREAD_IN_PROC(p, td) { 582 TD_CLR_SWAPPED(td); 583 if (TD_CAN_RUN(td)) 584 setrunnable(td); 585 } 586 mtx_unlock_spin(&sched_lock); 587 588 wakeup(&p->p_sflag); 589 590 /* Allow other threads to swap p out now. */ 591 --p->p_lock; 592 } 593#endif /* NO_SWAPPING */ 594} 595 596/* 597 * This swapin algorithm attempts to swap-in processes only if there 598 * is enough space for them. Of course, if a process waits for a long 599 * time, it will be swapped in anyway. 600 * 601 * XXXKSE - process with the thread with highest priority counts.. 602 * 603 * Giant is still held at this point, to be released in tsleep. 604 */ 605/* ARGSUSED*/ 606static void 607scheduler(dummy) 608 void *dummy; 609{ 610 struct proc *p; 611 struct thread *td; 612 int pri; 613 struct proc *pp; 614 int ppri; 615 616 mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED); 617 /* GIANT_REQUIRED */ 618 619loop: 620 if (vm_page_count_min()) { 621 VM_WAIT; 622 goto loop; 623 } 624 625 pp = NULL; 626 ppri = INT_MIN; 627 sx_slock(&allproc_lock); 628 FOREACH_PROC_IN_SYSTEM(p) { 629 struct ksegrp *kg; 630 if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { 631 continue; 632 } 633 mtx_lock_spin(&sched_lock); 634 FOREACH_THREAD_IN_PROC(p, td) { 635 /* 636 * An otherwise runnable thread of a process 637 * swapped out has only the TDI_SWAPPED bit set. 638 * 639 */ 640 if (td->td_inhibitors == TDI_SWAPPED) { 641 kg = td->td_ksegrp; 642 pri = p->p_swtime + kg->kg_slptime; 643 if ((p->p_sflag & PS_SWAPINREQ) == 0) { 644 pri -= kg->kg_nice * 8; 645 } 646 647 /* 648 * if this ksegrp is higher priority 649 * and there is enough space, then select 650 * this process instead of the previous 651 * selection. 652 */ 653 if (pri > ppri) { 654 pp = p; 655 ppri = pri; 656 } 657 } 658 } 659 mtx_unlock_spin(&sched_lock); 660 } 661 sx_sunlock(&allproc_lock); 662 663 /* 664 * Nothing to do, back to sleep. 665 */ 666 if ((p = pp) == NULL) { 667 tsleep(&proc0, PVM, "sched", maxslp * hz / 2); 668 goto loop; 669 } 670 PROC_LOCK(p); 671 672 /* 673 * Another process may be bringing or may have already 674 * brought this process in while we traverse all threads. 675 * Or, this process may even be being swapped out again. 676 */ 677 if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { 678 PROC_UNLOCK(p); 679 goto loop; 680 } 681 682 mtx_lock_spin(&sched_lock); 683 p->p_sflag &= ~PS_SWAPINREQ; 684 mtx_unlock_spin(&sched_lock); 685 686 /* 687 * We would like to bring someone in. (only if there is space). 688 * [What checks the space? ] 689 */ 690 faultin(p); 691 PROC_UNLOCK(p); 692 mtx_lock_spin(&sched_lock); 693 p->p_swtime = 0; 694 mtx_unlock_spin(&sched_lock); 695 goto loop; 696} 697 698#ifndef NO_SWAPPING 699 700/* 701 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 702 */ 703static int swap_idle_threshold1 = 2; 704SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, 705 &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); 706 707/* 708 * Swap_idle_threshold2 is the time that a process can be idle before 709 * it will be swapped out, if idle swapping is enabled. 710 */ 711static int swap_idle_threshold2 = 10; 712SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, 713 &swap_idle_threshold2, 0, "Time before a process will be swapped out"); 714 715/* 716 * Swapout is driven by the pageout daemon. Very simple, we find eligible 717 * procs and unwire their u-areas. We try to always "swap" at least one 718 * process in case we need the room for a swapin. 719 * If any procs have been sleeping/stopped for at least maxslp seconds, 720 * they are swapped. Else, we swap the longest-sleeping or stopped process, 721 * if any, otherwise the longest-resident process. 722 */ 723void 724swapout_procs(action) 725int action; 726{ 727 struct proc *p; 728 struct thread *td; 729 struct ksegrp *kg; 730 int didswap = 0; 731 732 GIANT_REQUIRED; 733 734retry: 735 sx_slock(&allproc_lock); 736 FOREACH_PROC_IN_SYSTEM(p) { 737 struct vmspace *vm; 738 int minslptime = 100000; 739 740 /* 741 * Watch out for a process in 742 * creation. It may have no 743 * address space or lock yet. 744 */ 745 mtx_lock_spin(&sched_lock); 746 if (p->p_state == PRS_NEW) { 747 mtx_unlock_spin(&sched_lock); 748 continue; 749 } 750 mtx_unlock_spin(&sched_lock); 751 752 /* 753 * An aio daemon switches its 754 * address space while running. 755 * Perform a quick check whether 756 * a process has P_SYSTEM. 757 */ 758 if ((p->p_flag & P_SYSTEM) != 0) 759 continue; 760 761 /* 762 * Do not swapout a process that 763 * is waiting for VM data 764 * structures as there is a possible 765 * deadlock. Test this first as 766 * this may block. 767 * 768 * Lock the map until swapout 769 * finishes, or a thread of this 770 * process may attempt to alter 771 * the map. 772 */ 773 PROC_LOCK(p); 774 vm = p->p_vmspace; 775 KASSERT(vm != NULL, 776 ("swapout_procs: a process has no address space")); 777 ++vm->vm_refcnt; 778 PROC_UNLOCK(p); 779 if (!vm_map_trylock(&vm->vm_map)) 780 goto nextproc1; 781 782 PROC_LOCK(p); 783 if (p->p_lock != 0 || 784 (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT) 785 ) != 0) { 786 goto nextproc2; 787 } 788 /* 789 * only aiod changes vmspace, however it will be 790 * skipped because of the if statement above checking 791 * for P_SYSTEM 792 */ 793 if ((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) != PS_INMEM) 794 goto nextproc2; 795 796 switch (p->p_state) { 797 default: 798 /* Don't swap out processes in any sort 799 * of 'special' state. */ 800 break; 801 802 case PRS_NORMAL: 803 mtx_lock_spin(&sched_lock); 804 /* 805 * do not swapout a realtime process 806 * Check all the thread groups.. 807 */ 808 FOREACH_KSEGRP_IN_PROC(p, kg) { 809 if (PRI_IS_REALTIME(kg->kg_pri_class)) 810 goto nextproc; 811 812 /* 813 * Guarantee swap_idle_threshold1 814 * time in memory. 815 */ 816 if (kg->kg_slptime < swap_idle_threshold1) 817 goto nextproc; 818 819 /* 820 * Do not swapout a process if it is 821 * waiting on a critical event of some 822 * kind or there is a thread whose 823 * pageable memory may be accessed. 824 * 825 * This could be refined to support 826 * swapping out a thread. 827 */ 828 FOREACH_THREAD_IN_GROUP(kg, td) { 829 if ((td->td_priority) < PSOCK || 830 !thread_safetoswapout(td)) 831 goto nextproc; 832 } 833 /* 834 * If the system is under memory stress, 835 * or if we are swapping 836 * idle processes >= swap_idle_threshold2, 837 * then swap the process out. 838 */ 839 if (((action & VM_SWAP_NORMAL) == 0) && 840 (((action & VM_SWAP_IDLE) == 0) || 841 (kg->kg_slptime < swap_idle_threshold2))) 842 goto nextproc; 843 844 if (minslptime > kg->kg_slptime) 845 minslptime = kg->kg_slptime; 846 } 847 848 /* 849 * If the process has been asleep for awhile and had 850 * most of its pages taken away already, swap it out. 851 */ 852 if ((action & VM_SWAP_NORMAL) || 853 ((action & VM_SWAP_IDLE) && 854 (minslptime > swap_idle_threshold2))) { 855 swapout(p); 856 didswap++; 857 mtx_unlock_spin(&sched_lock); 858 PROC_UNLOCK(p); 859 vm_map_unlock(&vm->vm_map); 860 vmspace_free(vm); 861 sx_sunlock(&allproc_lock); 862 goto retry; 863 } 864nextproc: 865 mtx_unlock_spin(&sched_lock); 866 } 867nextproc2: 868 PROC_UNLOCK(p); 869 vm_map_unlock(&vm->vm_map); 870nextproc1: 871 vmspace_free(vm); 872 continue; 873 } 874 sx_sunlock(&allproc_lock); 875 /* 876 * If we swapped something out, and another process needed memory, 877 * then wakeup the sched process. 878 */ 879 if (didswap) 880 wakeup(&proc0); 881} 882 883static void 884swapout(p) 885 struct proc *p; 886{ 887 struct thread *td; 888 889 PROC_LOCK_ASSERT(p, MA_OWNED); 890 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 891#if defined(SWAP_DEBUG) 892 printf("swapping out %d\n", p->p_pid); 893#endif 894 895 /* 896 * The states of this process and its threads may have changed 897 * by now. Assuming that there is only one pageout daemon thread, 898 * this process should still be in memory. 899 */ 900 KASSERT((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) == PS_INMEM, 901 ("swapout: lost a swapout race?")); 902 903#if defined(INVARIANTS) 904 /* 905 * Make sure that all threads are safe to be swapped out. 906 * 907 * Alternatively, we could swap out only safe threads. 908 */ 909 FOREACH_THREAD_IN_PROC(p, td) { 910 KASSERT(thread_safetoswapout(td), 911 ("swapout: there is a thread not safe for swapout")); 912 } 913#endif /* INVARIANTS */ 914 915 ++p->p_stats->p_ru.ru_nswap; 916 /* 917 * remember the process resident count 918 */ 919 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 920 921 p->p_sflag &= ~PS_INMEM; 922 p->p_sflag |= PS_SWAPPINGOUT; 923 PROC_UNLOCK(p); 924 FOREACH_THREAD_IN_PROC(p, td) 925 TD_SET_SWAPPED(td); 926 mtx_unlock_spin(&sched_lock); 927 928 vm_proc_swapout(p); 929 FOREACH_THREAD_IN_PROC(p, td) 930 pmap_swapout_thread(td); 931 932 PROC_LOCK(p); 933 mtx_lock_spin(&sched_lock); 934 p->p_sflag &= ~PS_SWAPPINGOUT; 935 p->p_swtime = 0; 936} 937#endif /* !NO_SWAPPING */ 938