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/sched.h> 74 #include <sys/sf_buf.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 extern int maxslp; 97 98 /* 99 * System initialization 100 * 101 * Note: proc0 from proc.h 102 */ 103 static void vm_init_limits(void *); 104 SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0) 105 106 /* 107 * THIS MUST BE THE LAST INITIALIZATION ITEM!!! 108 * 109 * Note: run scheduling should be divorced from the vm system. 110 */ 111 static void scheduler(void *); 112 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_ANY, scheduler, NULL) 113 114 #ifndef NO_SWAPPING 115 static int swapout(struct proc *); 116 static void swapclear(struct proc *); 117 #endif 118 119 120 static volatile int proc0_rescan; 121 122 123 /* 124 * MPSAFE 125 * 126 * WARNING! This code calls vm_map_check_protection() which only checks 127 * the associated vm_map_entry range. It does not determine whether the 128 * contents of the memory is actually readable or writable. In most cases 129 * just checking the vm_map_entry is sufficient within the kernel's address 130 * space. 131 */ 132 int 133 kernacc(addr, len, rw) 134 void *addr; 135 int len, rw; 136 { 137 boolean_t rv; 138 vm_offset_t saddr, eaddr; 139 vm_prot_t prot; 140 141 KASSERT((rw & ~VM_PROT_ALL) == 0, 142 ("illegal ``rw'' argument to kernacc (%x)\n", rw)); 143 144 if ((vm_offset_t)addr + len > kernel_map->max_offset || 145 (vm_offset_t)addr + len < (vm_offset_t)addr) 146 return (FALSE); 147 148 prot = rw; 149 saddr = trunc_page((vm_offset_t)addr); 150 eaddr = round_page((vm_offset_t)addr + len); 151 vm_map_lock_read(kernel_map); 152 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot); 153 vm_map_unlock_read(kernel_map); 154 return (rv == TRUE); 155 } 156 157 /* 158 * MPSAFE 159 * 160 * WARNING! This code calls vm_map_check_protection() which only checks 161 * the associated vm_map_entry range. It does not determine whether the 162 * contents of the memory is actually readable or writable. vmapbuf(), 163 * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be 164 * used in conjuction with this call. 165 */ 166 int 167 useracc(addr, len, rw) 168 void *addr; 169 int len, rw; 170 { 171 boolean_t rv; 172 vm_prot_t prot; 173 vm_map_t map; 174 175 KASSERT((rw & ~VM_PROT_ALL) == 0, 176 ("illegal ``rw'' argument to useracc (%x)\n", rw)); 177 prot = rw; 178 map = &curproc->p_vmspace->vm_map; 179 if ((vm_offset_t)addr + len > vm_map_max(map) || 180 (vm_offset_t)addr + len < (vm_offset_t)addr) { 181 return (FALSE); 182 } 183 vm_map_lock_read(map); 184 rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr), 185 round_page((vm_offset_t)addr + len), prot); 186 vm_map_unlock_read(map); 187 return (rv == TRUE); 188 } 189 190 int 191 vslock(void *addr, size_t len) 192 { 193 vm_offset_t end, last, start; 194 vm_size_t npages; 195 int error; 196 197 last = (vm_offset_t)addr + len; 198 start = trunc_page((vm_offset_t)addr); 199 end = round_page(last); 200 if (last < (vm_offset_t)addr || end < (vm_offset_t)addr) 201 return (EINVAL); 202 npages = atop(end - start); 203 if (npages > vm_page_max_wired) 204 return (ENOMEM); 205 PROC_LOCK(curproc); 206 if (ptoa(npages + 207 pmap_wired_count(vm_map_pmap(&curproc->p_vmspace->vm_map))) > 208 lim_cur(curproc, RLIMIT_MEMLOCK)) { 209 PROC_UNLOCK(curproc); 210 return (ENOMEM); 211 } 212 PROC_UNLOCK(curproc); 213 #if 0 214 /* 215 * XXX - not yet 216 * 217 * The limit for transient usage of wired pages should be 218 * larger than for "permanent" wired pages (mlock()). 219 * 220 * Also, the sysctl code, which is the only present user 221 * of vslock(), does a hard loop on EAGAIN. 222 */ 223 if (npages + cnt.v_wire_count > vm_page_max_wired) 224 return (EAGAIN); 225 #endif 226 error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end, 227 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 228 /* 229 * Return EFAULT on error to match copy{in,out}() behaviour 230 * rather than returning ENOMEM like mlock() would. 231 */ 232 return (error == KERN_SUCCESS ? 0 : EFAULT); 233 } 234 235 void 236 vsunlock(void *addr, size_t len) 237 { 238 239 /* Rely on the parameter sanity checks performed by vslock(). */ 240 (void)vm_map_unwire(&curproc->p_vmspace->vm_map, 241 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), 242 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 243 } 244 245 /* 246 * Pin the page contained within the given object at the given offset. If the 247 * page is not resident, allocate and load it using the given object's pager. 248 * Return the pinned page if successful; otherwise, return NULL. 249 */ 250 static vm_page_t 251 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset) 252 { 253 vm_page_t m, ma[1]; 254 vm_pindex_t pindex; 255 int rv; 256 257 VM_OBJECT_LOCK(object); 258 pindex = OFF_TO_IDX(offset); 259 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 260 if ((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) { 261 ma[0] = m; 262 rv = vm_pager_get_pages(object, ma, 1, 0); 263 m = vm_page_lookup(object, pindex); 264 if (m == NULL) 265 goto out; 266 if (m->valid == 0 || rv != VM_PAGER_OK) { 267 vm_page_lock_queues(); 268 vm_page_free(m); 269 vm_page_unlock_queues(); 270 m = NULL; 271 goto out; 272 } 273 } 274 vm_page_lock_queues(); 275 vm_page_hold(m); 276 vm_page_unlock_queues(); 277 vm_page_wakeup(m); 278 out: 279 VM_OBJECT_UNLOCK(object); 280 return (m); 281 } 282 283 /* 284 * Return a CPU private mapping to the page at the given offset within the 285 * given object. The page is pinned before it is mapped. 286 */ 287 struct sf_buf * 288 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset) 289 { 290 vm_page_t m; 291 292 m = vm_imgact_hold_page(object, offset); 293 if (m == NULL) 294 return (NULL); 295 sched_pin(); 296 return (sf_buf_alloc(m, SFB_CPUPRIVATE)); 297 } 298 299 /* 300 * Destroy the given CPU private mapping and unpin the page that it mapped. 301 */ 302 void 303 vm_imgact_unmap_page(struct sf_buf *sf) 304 { 305 vm_page_t m; 306 307 m = sf_buf_page(sf); 308 sf_buf_free(sf); 309 sched_unpin(); 310 vm_page_lock_queues(); 311 vm_page_unhold(m); 312 vm_page_unlock_queues(); 313 } 314 315 #ifndef KSTACK_MAX_PAGES 316 #define KSTACK_MAX_PAGES 32 317 #endif 318 319 /* 320 * Create the kernel stack (including pcb for i386) for a new thread. 321 * This routine directly affects the fork perf for a process and 322 * create performance for a thread. 323 */ 324 void 325 vm_thread_new(struct thread *td, int pages) 326 { 327 vm_object_t ksobj; 328 vm_offset_t ks; 329 vm_page_t m, ma[KSTACK_MAX_PAGES]; 330 int i; 331 332 /* Bounds check */ 333 if (pages <= 1) 334 pages = KSTACK_PAGES; 335 else if (pages > KSTACK_MAX_PAGES) 336 pages = KSTACK_MAX_PAGES; 337 /* 338 * Allocate an object for the kstack. 339 */ 340 ksobj = vm_object_allocate(OBJT_DEFAULT, pages); 341 td->td_kstack_obj = ksobj; 342 /* 343 * Get a kernel virtual address for this thread's kstack. 344 */ 345 ks = kmem_alloc_nofault(kernel_map, 346 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 347 if (ks == 0) 348 panic("vm_thread_new: kstack allocation failed"); 349 if (KSTACK_GUARD_PAGES != 0) { 350 pmap_qremove(ks, KSTACK_GUARD_PAGES); 351 ks += KSTACK_GUARD_PAGES * PAGE_SIZE; 352 } 353 td->td_kstack = ks; 354 /* 355 * Knowing the number of pages allocated is useful when you 356 * want to deallocate them. 357 */ 358 td->td_kstack_pages = pages; 359 /* 360 * For the length of the stack, link in a real page of ram for each 361 * page of stack. 362 */ 363 VM_OBJECT_LOCK(ksobj); 364 for (i = 0; i < pages; i++) { 365 /* 366 * Get a kernel stack page. 367 */ 368 m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY | 369 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED); 370 ma[i] = m; 371 m->valid = VM_PAGE_BITS_ALL; 372 } 373 VM_OBJECT_UNLOCK(ksobj); 374 pmap_qenter(ks, ma, pages); 375 } 376 377 /* 378 * Dispose of a thread's kernel stack. 379 */ 380 void 381 vm_thread_dispose(struct thread *td) 382 { 383 vm_object_t ksobj; 384 vm_offset_t ks; 385 vm_page_t m; 386 int i, pages; 387 388 pages = td->td_kstack_pages; 389 ksobj = td->td_kstack_obj; 390 ks = td->td_kstack; 391 pmap_qremove(ks, pages); 392 VM_OBJECT_LOCK(ksobj); 393 for (i = 0; i < pages; i++) { 394 m = vm_page_lookup(ksobj, i); 395 if (m == NULL) 396 panic("vm_thread_dispose: kstack already missing?"); 397 vm_page_lock_queues(); 398 vm_page_unwire(m, 0); 399 vm_page_free(m); 400 vm_page_unlock_queues(); 401 } 402 VM_OBJECT_UNLOCK(ksobj); 403 vm_object_deallocate(ksobj); 404 kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE), 405 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 406 } 407 408 /* 409 * Allow a thread's kernel stack to be paged out. 410 */ 411 void 412 vm_thread_swapout(struct thread *td) 413 { 414 vm_object_t ksobj; 415 vm_page_t m; 416 int i, pages; 417 418 cpu_thread_swapout(td); 419 pages = td->td_kstack_pages; 420 ksobj = td->td_kstack_obj; 421 pmap_qremove(td->td_kstack, pages); 422 VM_OBJECT_LOCK(ksobj); 423 for (i = 0; i < pages; i++) { 424 m = vm_page_lookup(ksobj, i); 425 if (m == NULL) 426 panic("vm_thread_swapout: kstack already missing?"); 427 vm_page_lock_queues(); 428 vm_page_dirty(m); 429 vm_page_unwire(m, 0); 430 vm_page_unlock_queues(); 431 } 432 VM_OBJECT_UNLOCK(ksobj); 433 } 434 435 /* 436 * Bring the kernel stack for a specified thread back in. 437 */ 438 void 439 vm_thread_swapin(struct thread *td) 440 { 441 vm_object_t ksobj; 442 vm_page_t m, ma[KSTACK_MAX_PAGES]; 443 int i, pages, rv; 444 445 pages = td->td_kstack_pages; 446 ksobj = td->td_kstack_obj; 447 VM_OBJECT_LOCK(ksobj); 448 for (i = 0; i < pages; i++) { 449 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 450 if (m->valid != VM_PAGE_BITS_ALL) { 451 rv = vm_pager_get_pages(ksobj, &m, 1, 0); 452 if (rv != VM_PAGER_OK) 453 panic("vm_thread_swapin: cannot get kstack for proc: %d", td->td_proc->p_pid); 454 m = vm_page_lookup(ksobj, i); 455 m->valid = VM_PAGE_BITS_ALL; 456 } 457 ma[i] = m; 458 vm_page_lock_queues(); 459 vm_page_wire(m); 460 vm_page_unlock_queues(); 461 vm_page_wakeup(m); 462 } 463 VM_OBJECT_UNLOCK(ksobj); 464 pmap_qenter(td->td_kstack, ma, pages); 465 cpu_thread_swapin(td); 466 } 467 468 /* 469 * Set up a variable-sized alternate kstack. 470 */ 471 void 472 vm_thread_new_altkstack(struct thread *td, int pages) 473 { 474 475 td->td_altkstack = td->td_kstack; 476 td->td_altkstack_obj = td->td_kstack_obj; 477 td->td_altkstack_pages = td->td_kstack_pages; 478 479 vm_thread_new(td, pages); 480 } 481 482 /* 483 * Restore the original kstack. 484 */ 485 void 486 vm_thread_dispose_altkstack(struct thread *td) 487 { 488 489 vm_thread_dispose(td); 490 491 td->td_kstack = td->td_altkstack; 492 td->td_kstack_obj = td->td_altkstack_obj; 493 td->td_kstack_pages = td->td_altkstack_pages; 494 td->td_altkstack = 0; 495 td->td_altkstack_obj = NULL; 496 td->td_altkstack_pages = 0; 497 } 498 499 /* 500 * Implement fork's actions on an address space. 501 * Here we arrange for the address space to be copied or referenced, 502 * allocate a user struct (pcb and kernel stack), then call the 503 * machine-dependent layer to fill those in and make the new process 504 * ready to run. The new process is set up so that it returns directly 505 * to user mode to avoid stack copying and relocation problems. 506 */ 507 void 508 vm_forkproc(td, p2, td2, flags) 509 struct thread *td; 510 struct proc *p2; 511 struct thread *td2; 512 int flags; 513 { 514 struct proc *p1 = td->td_proc; 515 516 if ((flags & RFPROC) == 0) { 517 /* 518 * Divorce the memory, if it is shared, essentially 519 * this changes shared memory amongst threads, into 520 * COW locally. 521 */ 522 if ((flags & RFMEM) == 0) { 523 if (p1->p_vmspace->vm_refcnt > 1) { 524 vmspace_unshare(p1); 525 } 526 } 527 cpu_fork(td, p2, td2, flags); 528 return; 529 } 530 531 if (flags & RFMEM) { 532 p2->p_vmspace = p1->p_vmspace; 533 atomic_add_int(&p1->p_vmspace->vm_refcnt, 1); 534 } 535 536 while (vm_page_count_severe()) { 537 VM_WAIT; 538 } 539 540 if ((flags & RFMEM) == 0) { 541 p2->p_vmspace = vmspace_fork(p1->p_vmspace); 542 if (p1->p_vmspace->vm_shm) 543 shmfork(p1, p2); 544 } 545 546 /* 547 * cpu_fork will copy and update the pcb, set up the kernel stack, 548 * and make the child ready to run. 549 */ 550 cpu_fork(td, p2, td2, flags); 551 } 552 553 /* 554 * Called after process has been wait(2)'ed apon and is being reaped. 555 * The idea is to reclaim resources that we could not reclaim while 556 * the process was still executing. 557 */ 558 void 559 vm_waitproc(p) 560 struct proc *p; 561 { 562 563 vmspace_exitfree(p); /* and clean-out the vmspace */ 564 } 565 566 /* 567 * Set default limits for VM system. 568 * Called for proc 0, and then inherited by all others. 569 * 570 * XXX should probably act directly on proc0. 571 */ 572 static void 573 vm_init_limits(udata) 574 void *udata; 575 { 576 struct proc *p = udata; 577 struct plimit *limp; 578 int rss_limit; 579 580 /* 581 * Set up the initial limits on process VM. Set the maximum resident 582 * set size to be half of (reasonably) available memory. Since this 583 * is a soft limit, it comes into effect only when the system is out 584 * of memory - half of main memory helps to favor smaller processes, 585 * and reduces thrashing of the object cache. 586 */ 587 limp = p->p_limit; 588 limp->pl_rlimit[RLIMIT_STACK].rlim_cur = dflssiz; 589 limp->pl_rlimit[RLIMIT_STACK].rlim_max = maxssiz; 590 limp->pl_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz; 591 limp->pl_rlimit[RLIMIT_DATA].rlim_max = maxdsiz; 592 /* limit the limit to no less than 2MB */ 593 rss_limit = max(cnt.v_free_count, 512); 594 limp->pl_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit); 595 limp->pl_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY; 596 } 597 598 void 599 faultin(p) 600 struct proc *p; 601 { 602 #ifdef NO_SWAPPING 603 604 PROC_LOCK_ASSERT(p, MA_OWNED); 605 if ((p->p_flag & P_INMEM) == 0) 606 panic("faultin: proc swapped out with NO_SWAPPING!"); 607 #else /* !NO_SWAPPING */ 608 struct thread *td; 609 610 PROC_LOCK_ASSERT(p, MA_OWNED); 611 /* 612 * If another process is swapping in this process, 613 * just wait until it finishes. 614 */ 615 if (p->p_flag & P_SWAPPINGIN) { 616 while (p->p_flag & P_SWAPPINGIN) 617 msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0); 618 return; 619 } 620 if ((p->p_flag & P_INMEM) == 0) { 621 /* 622 * Don't let another thread swap process p out while we are 623 * busy swapping it in. 624 */ 625 ++p->p_lock; 626 p->p_flag |= P_SWAPPINGIN; 627 PROC_UNLOCK(p); 628 629 /* 630 * We hold no lock here because the list of threads 631 * can not change while all threads in the process are 632 * swapped out. 633 */ 634 FOREACH_THREAD_IN_PROC(p, td) 635 vm_thread_swapin(td); 636 PROC_LOCK(p); 637 PROC_SLOCK(p); 638 swapclear(p); 639 p->p_swtick = ticks; 640 PROC_SUNLOCK(p); 641 642 wakeup(&p->p_flag); 643 644 /* Allow other threads to swap p out now. */ 645 --p->p_lock; 646 } 647 #endif /* NO_SWAPPING */ 648 } 649 650 /* 651 * This swapin algorithm attempts to swap-in processes only if there 652 * is enough space for them. Of course, if a process waits for a long 653 * time, it will be swapped in anyway. 654 * 655 * XXXKSE - process with the thread with highest priority counts.. 656 * 657 * Giant is held on entry. 658 */ 659 /* ARGSUSED*/ 660 static void 661 scheduler(dummy) 662 void *dummy; 663 { 664 struct proc *p; 665 struct thread *td; 666 struct proc *pp; 667 int slptime; 668 int swtime; 669 int ppri; 670 int pri; 671 672 mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED); 673 mtx_unlock(&Giant); 674 675 loop: 676 if (vm_page_count_min()) { 677 VM_WAIT; 678 thread_lock(&thread0); 679 proc0_rescan = 0; 680 thread_unlock(&thread0); 681 goto loop; 682 } 683 684 pp = NULL; 685 ppri = INT_MIN; 686 sx_slock(&allproc_lock); 687 FOREACH_PROC_IN_SYSTEM(p) { 688 PROC_LOCK(p); 689 if (p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) { 690 PROC_UNLOCK(p); 691 continue; 692 } 693 swtime = (ticks - p->p_swtick) / hz; 694 PROC_SLOCK(p); 695 FOREACH_THREAD_IN_PROC(p, td) { 696 /* 697 * An otherwise runnable thread of a process 698 * swapped out has only the TDI_SWAPPED bit set. 699 * 700 */ 701 thread_lock(td); 702 if (td->td_inhibitors == TDI_SWAPPED) { 703 slptime = (ticks - td->td_slptick) / hz; 704 pri = swtime + slptime; 705 if ((td->td_flags & TDF_SWAPINREQ) == 0) 706 pri -= p->p_nice * 8; 707 /* 708 * if this thread is higher priority 709 * and there is enough space, then select 710 * this process instead of the previous 711 * selection. 712 */ 713 if (pri > ppri) { 714 pp = p; 715 ppri = pri; 716 } 717 } 718 thread_unlock(td); 719 } 720 PROC_SUNLOCK(p); 721 PROC_UNLOCK(p); 722 } 723 sx_sunlock(&allproc_lock); 724 725 /* 726 * Nothing to do, back to sleep. 727 */ 728 if ((p = pp) == NULL) { 729 thread_lock(&thread0); 730 if (!proc0_rescan) { 731 TD_SET_IWAIT(&thread0); 732 mi_switch(SW_VOL, NULL); 733 } 734 proc0_rescan = 0; 735 thread_unlock(&thread0); 736 goto loop; 737 } 738 PROC_LOCK(p); 739 740 /* 741 * Another process may be bringing or may have already 742 * brought this process in while we traverse all threads. 743 * Or, this process may even be being swapped out again. 744 */ 745 if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) { 746 PROC_UNLOCK(p); 747 thread_lock(&thread0); 748 proc0_rescan = 0; 749 thread_unlock(&thread0); 750 goto loop; 751 } 752 753 /* 754 * We would like to bring someone in. (only if there is space). 755 * [What checks the space? ] 756 */ 757 faultin(p); 758 PROC_UNLOCK(p); 759 thread_lock(&thread0); 760 proc0_rescan = 0; 761 thread_unlock(&thread0); 762 goto loop; 763 } 764 765 void kick_proc0(void) 766 { 767 struct thread *td = &thread0; 768 769 /* XXX This will probably cause a LOR in some cases */ 770 thread_lock(td); 771 if (TD_AWAITING_INTR(td)) { 772 CTR2(KTR_INTR, "%s: sched_add %d", __func__, 0); 773 TD_CLR_IWAIT(td); 774 sched_add(td, SRQ_INTR); 775 } else { 776 proc0_rescan = 1; 777 CTR2(KTR_INTR, "%s: state %d", 778 __func__, td->td_state); 779 } 780 thread_unlock(td); 781 782 } 783 784 785 #ifndef NO_SWAPPING 786 787 /* 788 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 789 */ 790 static int swap_idle_threshold1 = 2; 791 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, 792 &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); 793 794 /* 795 * Swap_idle_threshold2 is the time that a process can be idle before 796 * it will be swapped out, if idle swapping is enabled. 797 */ 798 static int swap_idle_threshold2 = 10; 799 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, 800 &swap_idle_threshold2, 0, "Time before a process will be swapped out"); 801 802 /* 803 * Swapout is driven by the pageout daemon. Very simple, we find eligible 804 * procs and swap out their stacks. We try to always "swap" at least one 805 * process in case we need the room for a swapin. 806 * If any procs have been sleeping/stopped for at least maxslp seconds, 807 * they are swapped. Else, we swap the longest-sleeping or stopped process, 808 * if any, otherwise the longest-resident process. 809 */ 810 void 811 swapout_procs(action) 812 int action; 813 { 814 struct proc *p; 815 struct thread *td; 816 int didswap = 0; 817 818 retry: 819 sx_slock(&allproc_lock); 820 FOREACH_PROC_IN_SYSTEM(p) { 821 struct vmspace *vm; 822 int minslptime = 100000; 823 int slptime; 824 825 /* 826 * Watch out for a process in 827 * creation. It may have no 828 * address space or lock yet. 829 */ 830 if (p->p_state == PRS_NEW) 831 continue; 832 /* 833 * An aio daemon switches its 834 * address space while running. 835 * Perform a quick check whether 836 * a process has P_SYSTEM. 837 */ 838 if ((p->p_flag & P_SYSTEM) != 0) 839 continue; 840 /* 841 * Do not swapout a process that 842 * is waiting for VM data 843 * structures as there is a possible 844 * deadlock. Test this first as 845 * this may block. 846 * 847 * Lock the map until swapout 848 * finishes, or a thread of this 849 * process may attempt to alter 850 * the map. 851 */ 852 vm = vmspace_acquire_ref(p); 853 if (vm == NULL) 854 continue; 855 if (!vm_map_trylock(&vm->vm_map)) 856 goto nextproc1; 857 858 PROC_LOCK(p); 859 if (p->p_lock != 0 || 860 (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT) 861 ) != 0) { 862 goto nextproc2; 863 } 864 /* 865 * only aiod changes vmspace, however it will be 866 * skipped because of the if statement above checking 867 * for P_SYSTEM 868 */ 869 if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM) 870 goto nextproc2; 871 872 switch (p->p_state) { 873 default: 874 /* Don't swap out processes in any sort 875 * of 'special' state. */ 876 break; 877 878 case PRS_NORMAL: 879 PROC_SLOCK(p); 880 /* 881 * do not swapout a realtime process 882 * Check all the thread groups.. 883 */ 884 FOREACH_THREAD_IN_PROC(p, td) { 885 thread_lock(td); 886 if (PRI_IS_REALTIME(td->td_pri_class)) { 887 thread_unlock(td); 888 goto nextproc; 889 } 890 slptime = (ticks - td->td_slptick) / hz; 891 /* 892 * Guarantee swap_idle_threshold1 893 * time in memory. 894 */ 895 if (slptime < swap_idle_threshold1) { 896 thread_unlock(td); 897 goto nextproc; 898 } 899 900 /* 901 * Do not swapout a process if it is 902 * waiting on a critical event of some 903 * kind or there is a thread whose 904 * pageable memory may be accessed. 905 * 906 * This could be refined to support 907 * swapping out a thread. 908 */ 909 if ((td->td_priority) < PSOCK || 910 !thread_safetoswapout(td)) { 911 thread_unlock(td); 912 goto nextproc; 913 } 914 /* 915 * If the system is under memory stress, 916 * or if we are swapping 917 * idle processes >= swap_idle_threshold2, 918 * then swap the process out. 919 */ 920 if (((action & VM_SWAP_NORMAL) == 0) && 921 (((action & VM_SWAP_IDLE) == 0) || 922 (slptime < swap_idle_threshold2))) { 923 thread_unlock(td); 924 goto nextproc; 925 } 926 927 if (minslptime > slptime) 928 minslptime = slptime; 929 thread_unlock(td); 930 } 931 932 /* 933 * If the pageout daemon didn't free enough pages, 934 * or if this process is idle and the system is 935 * configured to swap proactively, swap it out. 936 */ 937 if ((action & VM_SWAP_NORMAL) || 938 ((action & VM_SWAP_IDLE) && 939 (minslptime > swap_idle_threshold2))) { 940 if (swapout(p) == 0) 941 didswap++; 942 PROC_SUNLOCK(p); 943 PROC_UNLOCK(p); 944 vm_map_unlock(&vm->vm_map); 945 vmspace_free(vm); 946 sx_sunlock(&allproc_lock); 947 goto retry; 948 } 949 nextproc: 950 PROC_SUNLOCK(p); 951 } 952 nextproc2: 953 PROC_UNLOCK(p); 954 vm_map_unlock(&vm->vm_map); 955 nextproc1: 956 vmspace_free(vm); 957 continue; 958 } 959 sx_sunlock(&allproc_lock); 960 /* 961 * If we swapped something out, and another process needed memory, 962 * then wakeup the sched process. 963 */ 964 if (didswap) 965 wakeup(&proc0); 966 } 967 968 static void 969 swapclear(p) 970 struct proc *p; 971 { 972 struct thread *td; 973 974 PROC_LOCK_ASSERT(p, MA_OWNED); 975 PROC_SLOCK_ASSERT(p, MA_OWNED); 976 977 FOREACH_THREAD_IN_PROC(p, td) { 978 thread_lock(td); 979 td->td_flags |= TDF_INMEM; 980 td->td_flags &= ~TDF_SWAPINREQ; 981 TD_CLR_SWAPPED(td); 982 if (TD_CAN_RUN(td)) 983 setrunnable(td); 984 thread_unlock(td); 985 } 986 p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT); 987 p->p_flag |= P_INMEM; 988 } 989 990 static int 991 swapout(p) 992 struct proc *p; 993 { 994 struct thread *td; 995 996 PROC_LOCK_ASSERT(p, MA_OWNED); 997 PROC_SLOCK_ASSERT(p, MA_OWNED | MA_NOTRECURSED); 998 #if defined(SWAP_DEBUG) 999 printf("swapping out %d\n", p->p_pid); 1000 #endif 1001 1002 /* 1003 * The states of this process and its threads may have changed 1004 * by now. Assuming that there is only one pageout daemon thread, 1005 * this process should still be in memory. 1006 */ 1007 KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM, 1008 ("swapout: lost a swapout race?")); 1009 1010 /* 1011 * remember the process resident count 1012 */ 1013 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 1014 /* 1015 * Check and mark all threads before we proceed. 1016 */ 1017 p->p_flag &= ~P_INMEM; 1018 p->p_flag |= P_SWAPPINGOUT; 1019 FOREACH_THREAD_IN_PROC(p, td) { 1020 thread_lock(td); 1021 if (!thread_safetoswapout(td)) { 1022 thread_unlock(td); 1023 swapclear(p); 1024 return (EBUSY); 1025 } 1026 td->td_flags &= ~TDF_INMEM; 1027 TD_SET_SWAPPED(td); 1028 thread_unlock(td); 1029 } 1030 td = FIRST_THREAD_IN_PROC(p); 1031 ++td->td_ru.ru_nswap; 1032 PROC_SUNLOCK(p); 1033 PROC_UNLOCK(p); 1034 1035 /* 1036 * This list is stable because all threads are now prevented from 1037 * running. The list is only modified in the context of a running 1038 * thread in this process. 1039 */ 1040 FOREACH_THREAD_IN_PROC(p, td) 1041 vm_thread_swapout(td); 1042 1043 PROC_LOCK(p); 1044 p->p_flag &= ~P_SWAPPINGOUT; 1045 PROC_SLOCK(p); 1046 p->p_swtick = ticks; 1047 return (0); 1048 } 1049 #endif /* !NO_SWAPPING */ 1050