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 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_glue.c 8.6 (Berkeley) 1/5/94 33 * 34 * 35 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 36 * All rights reserved. 37 * 38 * Permission to use, copy, modify and distribute this software and 39 * its documentation is hereby granted, provided that both the copyright 40 * notice and this permission notice appear in all copies of the 41 * software, derivative works or modified versions, and any portions 42 * thereof, and that both notices appear in supporting documentation. 43 * 44 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 45 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 46 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 47 * 48 * Carnegie Mellon requests users of this software to return to 49 * 50 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 51 * School of Computer Science 52 * Carnegie Mellon University 53 * Pittsburgh PA 15213-3890 54 * 55 * any improvements or extensions that they make and grant Carnegie the 56 * rights to redistribute these changes. 57 */ 58 59 #include <sys/cdefs.h> 60 __FBSDID("$FreeBSD$"); 61 62 #include "opt_vm.h" 63 #include "opt_kstack_pages.h" 64 #include "opt_kstack_max_pages.h" 65 66 #include <sys/param.h> 67 #include <sys/systm.h> 68 #include <sys/limits.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 <vm/vm.h> 83 #include <vm/vm_param.h> 84 #include <vm/pmap.h> 85 #include <vm/vm_map.h> 86 #include <vm/vm_page.h> 87 #include <vm/vm_pageout.h> 88 #include <vm/vm_object.h> 89 #include <vm/vm_kern.h> 90 #include <vm/vm_extern.h> 91 #include <vm/vm_pager.h> 92 #include <vm/swap_pager.h> 93 94 extern int maxslp; 95 96 /* 97 * System initialization 98 * 99 * Note: proc0 from proc.h 100 */ 101 static void vm_init_limits(void *); 102 SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0) 103 104 /* 105 * THIS MUST BE THE LAST INITIALIZATION ITEM!!! 106 * 107 * Note: run scheduling should be divorced from the vm system. 108 */ 109 static void scheduler(void *); 110 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_ANY, scheduler, NULL) 111 112 #ifndef NO_SWAPPING 113 static void swapout(struct proc *); 114 #endif 115 116 117 static volatile int proc0_rescan; 118 119 120 /* 121 * MPSAFE 122 * 123 * WARNING! This code calls vm_map_check_protection() which only checks 124 * the associated vm_map_entry range. It does not determine whether the 125 * contents of the memory is actually readable or writable. In most cases 126 * just checking the vm_map_entry is sufficient within the kernel's address 127 * space. 128 */ 129 int 130 kernacc(addr, len, rw) 131 void *addr; 132 int len, rw; 133 { 134 boolean_t rv; 135 vm_offset_t saddr, eaddr; 136 vm_prot_t prot; 137 138 KASSERT((rw & ~VM_PROT_ALL) == 0, 139 ("illegal ``rw'' argument to kernacc (%x)\n", rw)); 140 141 if ((vm_offset_t)addr + len > kernel_map->max_offset || 142 (vm_offset_t)addr + len < (vm_offset_t)addr) 143 return (FALSE); 144 145 prot = rw; 146 saddr = trunc_page((vm_offset_t)addr); 147 eaddr = round_page((vm_offset_t)addr + len); 148 vm_map_lock_read(kernel_map); 149 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot); 150 vm_map_unlock_read(kernel_map); 151 return (rv == TRUE); 152 } 153 154 /* 155 * MPSAFE 156 * 157 * WARNING! This code calls vm_map_check_protection() which only checks 158 * the associated vm_map_entry range. It does not determine whether the 159 * contents of the memory is actually readable or writable. vmapbuf(), 160 * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be 161 * used in conjuction with this call. 162 */ 163 int 164 useracc(addr, len, rw) 165 void *addr; 166 int len, rw; 167 { 168 boolean_t rv; 169 vm_prot_t prot; 170 vm_map_t map; 171 172 KASSERT((rw & ~VM_PROT_ALL) == 0, 173 ("illegal ``rw'' argument to useracc (%x)\n", rw)); 174 prot = rw; 175 map = &curproc->p_vmspace->vm_map; 176 if ((vm_offset_t)addr + len > vm_map_max(map) || 177 (vm_offset_t)addr + len < (vm_offset_t)addr) { 178 return (FALSE); 179 } 180 vm_map_lock_read(map); 181 rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr), 182 round_page((vm_offset_t)addr + len), prot); 183 vm_map_unlock_read(map); 184 return (rv == TRUE); 185 } 186 187 int 188 vslock(void *addr, size_t len) 189 { 190 vm_offset_t end, last, start; 191 vm_size_t npages; 192 int error; 193 194 last = (vm_offset_t)addr + len; 195 start = trunc_page((vm_offset_t)addr); 196 end = round_page(last); 197 if (last < (vm_offset_t)addr || end < (vm_offset_t)addr) 198 return (EINVAL); 199 npages = atop(end - start); 200 if (npages > vm_page_max_wired) 201 return (ENOMEM); 202 PROC_LOCK(curproc); 203 if (ptoa(npages + 204 pmap_wired_count(vm_map_pmap(&curproc->p_vmspace->vm_map))) > 205 lim_cur(curproc, RLIMIT_MEMLOCK)) { 206 PROC_UNLOCK(curproc); 207 return (ENOMEM); 208 } 209 PROC_UNLOCK(curproc); 210 #if 0 211 /* 212 * XXX - not yet 213 * 214 * The limit for transient usage of wired pages should be 215 * larger than for "permanent" wired pages (mlock()). 216 * 217 * Also, the sysctl code, which is the only present user 218 * of vslock(), does a hard loop on EAGAIN. 219 */ 220 if (npages + cnt.v_wire_count > vm_page_max_wired) 221 return (EAGAIN); 222 #endif 223 error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end, 224 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 225 /* 226 * Return EFAULT on error to match copy{in,out}() behaviour 227 * rather than returning ENOMEM like mlock() would. 228 */ 229 return (error == KERN_SUCCESS ? 0 : EFAULT); 230 } 231 232 void 233 vsunlock(void *addr, size_t len) 234 { 235 236 /* Rely on the parameter sanity checks performed by vslock(). */ 237 (void)vm_map_unwire(&curproc->p_vmspace->vm_map, 238 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), 239 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 240 } 241 242 #ifndef KSTACK_MAX_PAGES 243 #define KSTACK_MAX_PAGES 32 244 #endif 245 246 /* 247 * Create the kernel stack (including pcb for i386) for a new thread. 248 * This routine directly affects the fork perf for a process and 249 * create performance for a thread. 250 */ 251 void 252 vm_thread_new(struct thread *td, int pages) 253 { 254 vm_object_t ksobj; 255 vm_offset_t ks; 256 vm_page_t m, ma[KSTACK_MAX_PAGES]; 257 int i; 258 259 /* Bounds check */ 260 if (pages <= 1) 261 pages = KSTACK_PAGES; 262 else if (pages > KSTACK_MAX_PAGES) 263 pages = KSTACK_MAX_PAGES; 264 /* 265 * Allocate an object for the kstack. 266 */ 267 ksobj = vm_object_allocate(OBJT_DEFAULT, pages); 268 td->td_kstack_obj = ksobj; 269 /* 270 * Get a kernel virtual address for this thread's kstack. 271 */ 272 ks = kmem_alloc_nofault(kernel_map, 273 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 274 if (ks == 0) 275 panic("vm_thread_new: kstack allocation failed"); 276 if (KSTACK_GUARD_PAGES != 0) { 277 pmap_qremove(ks, KSTACK_GUARD_PAGES); 278 ks += KSTACK_GUARD_PAGES * PAGE_SIZE; 279 } 280 td->td_kstack = ks; 281 /* 282 * Knowing the number of pages allocated is useful when you 283 * want to deallocate them. 284 */ 285 td->td_kstack_pages = pages; 286 /* 287 * For the length of the stack, link in a real page of ram for each 288 * page of stack. 289 */ 290 VM_OBJECT_LOCK(ksobj); 291 for (i = 0; i < pages; i++) { 292 /* 293 * Get a kernel stack page. 294 */ 295 m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY | 296 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED); 297 ma[i] = m; 298 m->valid = VM_PAGE_BITS_ALL; 299 } 300 VM_OBJECT_UNLOCK(ksobj); 301 pmap_qenter(ks, ma, pages); 302 } 303 304 /* 305 * Dispose of a thread's kernel stack. 306 */ 307 void 308 vm_thread_dispose(struct thread *td) 309 { 310 vm_object_t ksobj; 311 vm_offset_t ks; 312 vm_page_t m; 313 int i, pages; 314 315 pages = td->td_kstack_pages; 316 ksobj = td->td_kstack_obj; 317 ks = td->td_kstack; 318 pmap_qremove(ks, pages); 319 VM_OBJECT_LOCK(ksobj); 320 for (i = 0; i < pages; i++) { 321 m = vm_page_lookup(ksobj, i); 322 if (m == NULL) 323 panic("vm_thread_dispose: kstack already missing?"); 324 vm_page_lock_queues(); 325 vm_page_unwire(m, 0); 326 vm_page_free(m); 327 vm_page_unlock_queues(); 328 } 329 VM_OBJECT_UNLOCK(ksobj); 330 vm_object_deallocate(ksobj); 331 kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE), 332 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 333 } 334 335 /* 336 * Allow a thread's kernel stack to be paged out. 337 */ 338 void 339 vm_thread_swapout(struct thread *td) 340 { 341 vm_object_t ksobj; 342 vm_page_t m; 343 int i, pages; 344 345 cpu_thread_swapout(td); 346 pages = td->td_kstack_pages; 347 ksobj = td->td_kstack_obj; 348 pmap_qremove(td->td_kstack, pages); 349 VM_OBJECT_LOCK(ksobj); 350 for (i = 0; i < pages; i++) { 351 m = vm_page_lookup(ksobj, i); 352 if (m == NULL) 353 panic("vm_thread_swapout: kstack already missing?"); 354 vm_page_lock_queues(); 355 vm_page_dirty(m); 356 vm_page_unwire(m, 0); 357 vm_page_unlock_queues(); 358 } 359 VM_OBJECT_UNLOCK(ksobj); 360 } 361 362 /* 363 * Bring the kernel stack for a specified thread back in. 364 */ 365 void 366 vm_thread_swapin(struct thread *td) 367 { 368 vm_object_t ksobj; 369 vm_page_t m, ma[KSTACK_MAX_PAGES]; 370 int i, pages, rv; 371 372 pages = td->td_kstack_pages; 373 ksobj = td->td_kstack_obj; 374 VM_OBJECT_LOCK(ksobj); 375 for (i = 0; i < pages; i++) { 376 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 377 if (m->valid != VM_PAGE_BITS_ALL) { 378 rv = vm_pager_get_pages(ksobj, &m, 1, 0); 379 if (rv != VM_PAGER_OK) 380 panic("vm_thread_swapin: cannot get kstack for proc: %d", td->td_proc->p_pid); 381 m = vm_page_lookup(ksobj, i); 382 m->valid = VM_PAGE_BITS_ALL; 383 } 384 ma[i] = m; 385 vm_page_lock_queues(); 386 vm_page_wire(m); 387 vm_page_wakeup(m); 388 vm_page_unlock_queues(); 389 } 390 VM_OBJECT_UNLOCK(ksobj); 391 pmap_qenter(td->td_kstack, ma, pages); 392 cpu_thread_swapin(td); 393 } 394 395 /* 396 * Set up a variable-sized alternate kstack. 397 */ 398 void 399 vm_thread_new_altkstack(struct thread *td, int pages) 400 { 401 402 td->td_altkstack = td->td_kstack; 403 td->td_altkstack_obj = td->td_kstack_obj; 404 td->td_altkstack_pages = td->td_kstack_pages; 405 406 vm_thread_new(td, pages); 407 } 408 409 /* 410 * Restore the original kstack. 411 */ 412 void 413 vm_thread_dispose_altkstack(struct thread *td) 414 { 415 416 vm_thread_dispose(td); 417 418 td->td_kstack = td->td_altkstack; 419 td->td_kstack_obj = td->td_altkstack_obj; 420 td->td_kstack_pages = td->td_altkstack_pages; 421 td->td_altkstack = 0; 422 td->td_altkstack_obj = NULL; 423 td->td_altkstack_pages = 0; 424 } 425 426 /* 427 * Implement fork's actions on an address space. 428 * Here we arrange for the address space to be copied or referenced, 429 * allocate a user struct (pcb and kernel stack), then call the 430 * machine-dependent layer to fill those in and make the new process 431 * ready to run. The new process is set up so that it returns directly 432 * to user mode to avoid stack copying and relocation problems. 433 */ 434 void 435 vm_forkproc(td, p2, td2, flags) 436 struct thread *td; 437 struct proc *p2; 438 struct thread *td2; 439 int flags; 440 { 441 struct proc *p1 = td->td_proc; 442 443 if ((flags & RFPROC) == 0) { 444 /* 445 * Divorce the memory, if it is shared, essentially 446 * this changes shared memory amongst threads, into 447 * COW locally. 448 */ 449 if ((flags & RFMEM) == 0) { 450 if (p1->p_vmspace->vm_refcnt > 1) { 451 vmspace_unshare(p1); 452 } 453 } 454 cpu_fork(td, p2, td2, flags); 455 return; 456 } 457 458 if (flags & RFMEM) { 459 p2->p_vmspace = p1->p_vmspace; 460 atomic_add_int(&p1->p_vmspace->vm_refcnt, 1); 461 } 462 463 while (vm_page_count_severe()) { 464 VM_WAIT; 465 } 466 467 if ((flags & RFMEM) == 0) { 468 p2->p_vmspace = vmspace_fork(p1->p_vmspace); 469 if (p1->p_vmspace->vm_shm) 470 shmfork(p1, p2); 471 } 472 473 /* 474 * cpu_fork will copy and update the pcb, set up the kernel stack, 475 * and make the child ready to run. 476 */ 477 cpu_fork(td, p2, td2, flags); 478 } 479 480 /* 481 * Called after process has been wait(2)'ed apon and is being reaped. 482 * The idea is to reclaim resources that we could not reclaim while 483 * the process was still executing. 484 */ 485 void 486 vm_waitproc(p) 487 struct proc *p; 488 { 489 490 vmspace_exitfree(p); /* and clean-out the vmspace */ 491 } 492 493 /* 494 * Set default limits for VM system. 495 * Called for proc 0, and then inherited by all others. 496 * 497 * XXX should probably act directly on proc0. 498 */ 499 static void 500 vm_init_limits(udata) 501 void *udata; 502 { 503 struct proc *p = udata; 504 struct plimit *limp; 505 int rss_limit; 506 507 /* 508 * Set up the initial limits on process VM. Set the maximum resident 509 * set size to be half of (reasonably) available memory. Since this 510 * is a soft limit, it comes into effect only when the system is out 511 * of memory - half of main memory helps to favor smaller processes, 512 * and reduces thrashing of the object cache. 513 */ 514 limp = p->p_limit; 515 limp->pl_rlimit[RLIMIT_STACK].rlim_cur = dflssiz; 516 limp->pl_rlimit[RLIMIT_STACK].rlim_max = maxssiz; 517 limp->pl_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz; 518 limp->pl_rlimit[RLIMIT_DATA].rlim_max = maxdsiz; 519 /* limit the limit to no less than 2MB */ 520 rss_limit = max(cnt.v_free_count, 512); 521 limp->pl_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit); 522 limp->pl_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY; 523 } 524 525 void 526 faultin(p) 527 struct proc *p; 528 { 529 #ifdef NO_SWAPPING 530 531 PROC_LOCK_ASSERT(p, MA_OWNED); 532 if ((p->p_sflag & PS_INMEM) == 0) 533 panic("faultin: proc swapped out with NO_SWAPPING!"); 534 #else /* !NO_SWAPPING */ 535 struct thread *td; 536 537 PROC_LOCK_ASSERT(p, MA_OWNED); 538 /* 539 * If another process is swapping in this process, 540 * just wait until it finishes. 541 */ 542 if (p->p_sflag & PS_SWAPPINGIN) 543 msleep(&p->p_sflag, &p->p_mtx, PVM, "faultin", 0); 544 else if ((p->p_sflag & PS_INMEM) == 0) { 545 /* 546 * Don't let another thread swap process p out while we are 547 * busy swapping it in. 548 */ 549 ++p->p_lock; 550 mtx_lock_spin(&sched_lock); 551 p->p_sflag |= PS_SWAPPINGIN; 552 mtx_unlock_spin(&sched_lock); 553 PROC_UNLOCK(p); 554 555 FOREACH_THREAD_IN_PROC(p, td) 556 vm_thread_swapin(td); 557 558 PROC_LOCK(p); 559 mtx_lock_spin(&sched_lock); 560 p->p_sflag &= ~PS_SWAPPINGIN; 561 p->p_sflag |= PS_INMEM; 562 FOREACH_THREAD_IN_PROC(p, td) { 563 TD_CLR_SWAPPED(td); 564 if (TD_CAN_RUN(td)) 565 setrunnable(td); 566 } 567 mtx_unlock_spin(&sched_lock); 568 569 wakeup(&p->p_sflag); 570 571 /* Allow other threads to swap p out now. */ 572 --p->p_lock; 573 } 574 #endif /* NO_SWAPPING */ 575 } 576 577 /* 578 * This swapin algorithm attempts to swap-in processes only if there 579 * is enough space for them. Of course, if a process waits for a long 580 * time, it will be swapped in anyway. 581 * 582 * XXXKSE - process with the thread with highest priority counts.. 583 * 584 * Giant is held on entry. 585 */ 586 /* ARGSUSED*/ 587 static void 588 scheduler(dummy) 589 void *dummy; 590 { 591 struct proc *p; 592 struct thread *td; 593 int pri; 594 struct proc *pp; 595 int ppri; 596 597 mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED); 598 mtx_unlock(&Giant); 599 600 loop: 601 if (vm_page_count_min()) { 602 VM_WAIT; 603 mtx_lock_spin(&sched_lock); 604 proc0_rescan = 0; 605 mtx_unlock_spin(&sched_lock); 606 goto loop; 607 } 608 609 pp = NULL; 610 ppri = INT_MIN; 611 sx_slock(&allproc_lock); 612 FOREACH_PROC_IN_SYSTEM(p) { 613 struct ksegrp *kg; 614 if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { 615 continue; 616 } 617 mtx_lock_spin(&sched_lock); 618 FOREACH_THREAD_IN_PROC(p, td) { 619 /* 620 * An otherwise runnable thread of a process 621 * swapped out has only the TDI_SWAPPED bit set. 622 * 623 */ 624 if (td->td_inhibitors == TDI_SWAPPED) { 625 kg = td->td_ksegrp; 626 pri = p->p_swtime + kg->kg_slptime; 627 if ((p->p_sflag & PS_SWAPINREQ) == 0) { 628 pri -= p->p_nice * 8; 629 } 630 631 /* 632 * if this ksegrp is higher priority 633 * and there is enough space, then select 634 * this process instead of the previous 635 * selection. 636 */ 637 if (pri > ppri) { 638 pp = p; 639 ppri = pri; 640 } 641 } 642 } 643 mtx_unlock_spin(&sched_lock); 644 } 645 sx_sunlock(&allproc_lock); 646 647 /* 648 * Nothing to do, back to sleep. 649 */ 650 if ((p = pp) == NULL) { 651 mtx_lock_spin(&sched_lock); 652 if (!proc0_rescan) { 653 TD_SET_IWAIT(&thread0); 654 mi_switch(SW_VOL, NULL); 655 } 656 proc0_rescan = 0; 657 mtx_unlock_spin(&sched_lock); 658 goto loop; 659 } 660 PROC_LOCK(p); 661 662 /* 663 * Another process may be bringing or may have already 664 * brought this process in while we traverse all threads. 665 * Or, this process may even be being swapped out again. 666 */ 667 if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { 668 PROC_UNLOCK(p); 669 mtx_lock_spin(&sched_lock); 670 proc0_rescan = 0; 671 mtx_unlock_spin(&sched_lock); 672 goto loop; 673 } 674 675 mtx_lock_spin(&sched_lock); 676 p->p_sflag &= ~PS_SWAPINREQ; 677 mtx_unlock_spin(&sched_lock); 678 679 /* 680 * We would like to bring someone in. (only if there is space). 681 * [What checks the space? ] 682 */ 683 faultin(p); 684 PROC_UNLOCK(p); 685 mtx_lock_spin(&sched_lock); 686 p->p_swtime = 0; 687 proc0_rescan = 0; 688 mtx_unlock_spin(&sched_lock); 689 goto loop; 690 } 691 692 void kick_proc0(void) 693 { 694 struct thread *td = &thread0; 695 696 697 if (TD_AWAITING_INTR(td)) { 698 CTR2(KTR_INTR, "%s: setrunqueue %d", __func__, 0); 699 TD_CLR_IWAIT(td); 700 setrunqueue(td, SRQ_INTR); 701 } else { 702 proc0_rescan = 1; 703 CTR2(KTR_INTR, "%s: state %d", 704 __func__, td->td_state); 705 } 706 707 } 708 709 710 #ifndef NO_SWAPPING 711 712 /* 713 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 714 */ 715 static int swap_idle_threshold1 = 2; 716 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, 717 &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); 718 719 /* 720 * Swap_idle_threshold2 is the time that a process can be idle before 721 * it will be swapped out, if idle swapping is enabled. 722 */ 723 static int swap_idle_threshold2 = 10; 724 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, 725 &swap_idle_threshold2, 0, "Time before a process will be swapped out"); 726 727 /* 728 * Swapout is driven by the pageout daemon. Very simple, we find eligible 729 * procs and unwire their u-areas. We try to always "swap" at least one 730 * process in case we need the room for a swapin. 731 * If any procs have been sleeping/stopped for at least maxslp seconds, 732 * they are swapped. Else, we swap the longest-sleeping or stopped process, 733 * if any, otherwise the longest-resident process. 734 */ 735 void 736 swapout_procs(action) 737 int action; 738 { 739 struct proc *p; 740 struct thread *td; 741 struct ksegrp *kg; 742 int didswap = 0; 743 744 retry: 745 sx_slock(&allproc_lock); 746 FOREACH_PROC_IN_SYSTEM(p) { 747 struct vmspace *vm; 748 int minslptime = 100000; 749 750 /* 751 * Watch out for a process in 752 * creation. It may have no 753 * address space or lock yet. 754 */ 755 mtx_lock_spin(&sched_lock); 756 if (p->p_state == PRS_NEW) { 757 mtx_unlock_spin(&sched_lock); 758 continue; 759 } 760 mtx_unlock_spin(&sched_lock); 761 762 /* 763 * An aio daemon switches its 764 * address space while running. 765 * Perform a quick check whether 766 * a process has P_SYSTEM. 767 */ 768 if ((p->p_flag & P_SYSTEM) != 0) 769 continue; 770 771 /* 772 * Do not swapout a process that 773 * is waiting for VM data 774 * structures as there is a possible 775 * deadlock. Test this first as 776 * this may block. 777 * 778 * Lock the map until swapout 779 * finishes, or a thread of this 780 * process may attempt to alter 781 * the map. 782 */ 783 PROC_LOCK(p); 784 vm = p->p_vmspace; 785 KASSERT(vm != NULL, 786 ("swapout_procs: a process has no address space")); 787 atomic_add_int(&vm->vm_refcnt, 1); 788 PROC_UNLOCK(p); 789 if (!vm_map_trylock(&vm->vm_map)) 790 goto nextproc1; 791 792 PROC_LOCK(p); 793 if (p->p_lock != 0 || 794 (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT) 795 ) != 0) { 796 goto nextproc2; 797 } 798 /* 799 * only aiod changes vmspace, however it will be 800 * skipped because of the if statement above checking 801 * for P_SYSTEM 802 */ 803 if ((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) != PS_INMEM) 804 goto nextproc2; 805 806 switch (p->p_state) { 807 default: 808 /* Don't swap out processes in any sort 809 * of 'special' state. */ 810 break; 811 812 case PRS_NORMAL: 813 mtx_lock_spin(&sched_lock); 814 /* 815 * do not swapout a realtime process 816 * Check all the thread groups.. 817 */ 818 FOREACH_KSEGRP_IN_PROC(p, kg) { 819 if (PRI_IS_REALTIME(kg->kg_pri_class)) 820 goto nextproc; 821 822 /* 823 * Guarantee swap_idle_threshold1 824 * time in memory. 825 */ 826 if (kg->kg_slptime < swap_idle_threshold1) 827 goto nextproc; 828 829 /* 830 * Do not swapout a process if it is 831 * waiting on a critical event of some 832 * kind or there is a thread whose 833 * pageable memory may be accessed. 834 * 835 * This could be refined to support 836 * swapping out a thread. 837 */ 838 FOREACH_THREAD_IN_GROUP(kg, td) { 839 if ((td->td_priority) < PSOCK || 840 !thread_safetoswapout(td)) 841 goto nextproc; 842 } 843 /* 844 * If the system is under memory stress, 845 * or if we are swapping 846 * idle processes >= swap_idle_threshold2, 847 * then swap the process out. 848 */ 849 if (((action & VM_SWAP_NORMAL) == 0) && 850 (((action & VM_SWAP_IDLE) == 0) || 851 (kg->kg_slptime < swap_idle_threshold2))) 852 goto nextproc; 853 854 if (minslptime > kg->kg_slptime) 855 minslptime = kg->kg_slptime; 856 } 857 858 /* 859 * If the pageout daemon didn't free enough pages, 860 * or if this process is idle and the system is 861 * configured to swap proactively, swap it out. 862 */ 863 if ((action & VM_SWAP_NORMAL) || 864 ((action & VM_SWAP_IDLE) && 865 (minslptime > swap_idle_threshold2))) { 866 swapout(p); 867 didswap++; 868 mtx_unlock_spin(&sched_lock); 869 PROC_UNLOCK(p); 870 vm_map_unlock(&vm->vm_map); 871 vmspace_free(vm); 872 sx_sunlock(&allproc_lock); 873 goto retry; 874 } 875 nextproc: 876 mtx_unlock_spin(&sched_lock); 877 } 878 nextproc2: 879 PROC_UNLOCK(p); 880 vm_map_unlock(&vm->vm_map); 881 nextproc1: 882 vmspace_free(vm); 883 continue; 884 } 885 sx_sunlock(&allproc_lock); 886 /* 887 * If we swapped something out, and another process needed memory, 888 * then wakeup the sched process. 889 */ 890 if (didswap) 891 wakeup(&proc0); 892 } 893 894 static void 895 swapout(p) 896 struct proc *p; 897 { 898 struct thread *td; 899 900 PROC_LOCK_ASSERT(p, MA_OWNED); 901 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 902 #if defined(SWAP_DEBUG) 903 printf("swapping out %d\n", p->p_pid); 904 #endif 905 906 /* 907 * The states of this process and its threads may have changed 908 * by now. Assuming that there is only one pageout daemon thread, 909 * this process should still be in memory. 910 */ 911 KASSERT((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) == PS_INMEM, 912 ("swapout: lost a swapout race?")); 913 914 #if defined(INVARIANTS) 915 /* 916 * Make sure that all threads are safe to be swapped out. 917 * 918 * Alternatively, we could swap out only safe threads. 919 */ 920 FOREACH_THREAD_IN_PROC(p, td) { 921 KASSERT(thread_safetoswapout(td), 922 ("swapout: there is a thread not safe for swapout")); 923 } 924 #endif /* INVARIANTS */ 925 926 ++p->p_stats->p_ru.ru_nswap; 927 /* 928 * remember the process resident count 929 */ 930 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 931 932 p->p_sflag &= ~PS_INMEM; 933 p->p_sflag |= PS_SWAPPINGOUT; 934 PROC_UNLOCK(p); 935 FOREACH_THREAD_IN_PROC(p, td) 936 TD_SET_SWAPPED(td); 937 mtx_unlock_spin(&sched_lock); 938 939 FOREACH_THREAD_IN_PROC(p, td) 940 vm_thread_swapout(td); 941 942 PROC_LOCK(p); 943 mtx_lock_spin(&sched_lock); 944 p->p_sflag &= ~PS_SWAPPINGOUT; 945 p->p_swtime = 0; 946 } 947 #endif /* !NO_SWAPPING */ 948