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/racct.h> 73#include <sys/resourcevar.h> 74#include <sys/sched.h> 75#include <sys/sf_buf.h> 76#include <sys/shm.h> 77#include <sys/vmmeter.h> 78#include <sys/sx.h> 79#include <sys/sysctl.h> 80#include <sys/_kstack_cache.h> 81#include <sys/eventhandler.h> 82#include <sys/kernel.h> 83#include <sys/ktr.h> 84#include <sys/unistd.h> 85 86#include <vm/vm.h> 87#include <vm/vm_param.h> 88#include <vm/pmap.h> 89#include <vm/vm_map.h> 90#include <vm/vm_page.h> 91#include <vm/vm_pageout.h> 92#include <vm/vm_object.h> 93#include <vm/vm_kern.h> 94#include <vm/vm_extern.h> 95#include <vm/vm_pager.h> 96#include <vm/swap_pager.h> 97 98#ifndef NO_SWAPPING 99static int swapout(struct proc *); 100static void swapclear(struct proc *); 101static void vm_thread_swapin(struct thread *td); 102static void vm_thread_swapout(struct thread *td); 103#endif 104 105/* 106 * MPSAFE 107 * 108 * WARNING! This code calls vm_map_check_protection() which only checks 109 * the associated vm_map_entry range. It does not determine whether the 110 * contents of the memory is actually readable or writable. In most cases 111 * just checking the vm_map_entry is sufficient within the kernel's address 112 * space. 113 */ 114int 115kernacc(addr, len, rw) 116 void *addr; 117 int len, rw; 118{ 119 boolean_t rv; 120 vm_offset_t saddr, eaddr; 121 vm_prot_t prot; 122 123 KASSERT((rw & ~VM_PROT_ALL) == 0, 124 ("illegal ``rw'' argument to kernacc (%x)\n", rw)); 125 126 if ((vm_offset_t)addr + len > kernel_map->max_offset || 127 (vm_offset_t)addr + len < (vm_offset_t)addr) 128 return (FALSE); 129 130 prot = rw; 131 saddr = trunc_page((vm_offset_t)addr); 132 eaddr = round_page((vm_offset_t)addr + len); 133 vm_map_lock_read(kernel_map); 134 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot); 135 vm_map_unlock_read(kernel_map); 136 return (rv == TRUE); 137} 138 139/* 140 * MPSAFE 141 * 142 * WARNING! This code calls vm_map_check_protection() which only checks 143 * the associated vm_map_entry range. It does not determine whether the 144 * contents of the memory is actually readable or writable. vmapbuf(), 145 * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be 146 * used in conjuction with this call. 147 */ 148int 149useracc(addr, len, rw) 150 void *addr; 151 int len, rw; 152{ 153 boolean_t rv; 154 vm_prot_t prot; 155 vm_map_t map; 156 157 KASSERT((rw & ~VM_PROT_ALL) == 0, 158 ("illegal ``rw'' argument to useracc (%x)\n", rw)); 159 prot = rw; 160 map = &curproc->p_vmspace->vm_map; 161 if ((vm_offset_t)addr + len > vm_map_max(map) || 162 (vm_offset_t)addr + len < (vm_offset_t)addr) { 163 return (FALSE); 164 } 165 vm_map_lock_read(map); 166 rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr), 167 round_page((vm_offset_t)addr + len), prot); 168 vm_map_unlock_read(map); 169 return (rv == TRUE); 170} 171 172int 173vslock(void *addr, size_t len) 174{ 175 vm_offset_t end, last, start; 176 vm_size_t npages; 177 int error; 178 179 last = (vm_offset_t)addr + len; 180 start = trunc_page((vm_offset_t)addr); 181 end = round_page(last); 182 if (last < (vm_offset_t)addr || end < (vm_offset_t)addr) 183 return (EINVAL); 184 npages = atop(end - start); 185 if (npages > vm_page_max_wired) 186 return (ENOMEM); 187#if 0 188 /* 189 * XXX - not yet 190 * 191 * The limit for transient usage of wired pages should be 192 * larger than for "permanent" wired pages (mlock()). 193 * 194 * Also, the sysctl code, which is the only present user 195 * of vslock(), does a hard loop on EAGAIN. 196 */ 197 if (npages + cnt.v_wire_count > vm_page_max_wired) 198 return (EAGAIN); 199#endif 200 error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end, 201 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 202 /* 203 * Return EFAULT on error to match copy{in,out}() behaviour 204 * rather than returning ENOMEM like mlock() would. 205 */ 206 return (error == KERN_SUCCESS ? 0 : EFAULT); 207} 208 209void 210vsunlock(void *addr, size_t len) 211{ 212 213 /* Rely on the parameter sanity checks performed by vslock(). */ 214 (void)vm_map_unwire(&curproc->p_vmspace->vm_map, 215 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), 216 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 217} 218 219/* 220 * Pin the page contained within the given object at the given offset. If the 221 * page is not resident, allocate and load it using the given object's pager. 222 * Return the pinned page if successful; otherwise, return NULL. 223 */ 224static vm_page_t 225vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset) 226{ 227 vm_page_t m, ma[1]; 228 vm_pindex_t pindex; 229 int rv; 230 231 VM_OBJECT_LOCK(object); 232 pindex = OFF_TO_IDX(offset); 233 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 234 if (m->valid != VM_PAGE_BITS_ALL) { 235 ma[0] = m; 236 rv = vm_pager_get_pages(object, ma, 1, 0); 237 m = vm_page_lookup(object, pindex); 238 if (m == NULL) 239 goto out; 240 if (rv != VM_PAGER_OK) { 241 vm_page_lock(m); 242 vm_page_free(m); 243 vm_page_unlock(m); 244 m = NULL; 245 goto out; 246 } 247 } 248 vm_page_lock(m); 249 vm_page_hold(m); 250 vm_page_unlock(m); 251 vm_page_wakeup(m); 252out: 253 VM_OBJECT_UNLOCK(object); 254 return (m); 255} 256 257/* 258 * Return a CPU private mapping to the page at the given offset within the 259 * given object. The page is pinned before it is mapped. 260 */ 261struct sf_buf * 262vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset) 263{ 264 vm_page_t m; 265 266 m = vm_imgact_hold_page(object, offset); 267 if (m == NULL) 268 return (NULL); 269 sched_pin(); 270 return (sf_buf_alloc(m, SFB_CPUPRIVATE)); 271} 272 273/* 274 * Destroy the given CPU private mapping and unpin the page that it mapped. 275 */ 276void 277vm_imgact_unmap_page(struct sf_buf *sf) 278{ 279 vm_page_t m; 280 281 m = sf_buf_page(sf); 282 sf_buf_free(sf); 283 sched_unpin(); 284 vm_page_lock(m); 285 vm_page_unhold(m); 286 vm_page_unlock(m); 287} 288 289void 290vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz) 291{ 292 293 pmap_sync_icache(map->pmap, va, sz); 294} 295 296struct kstack_cache_entry *kstack_cache; 297static int kstack_cache_size = 128; 298static int kstacks; 299static struct mtx kstack_cache_mtx; 300SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0, 301 ""); 302SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0, 303 ""); 304 305#ifndef KSTACK_MAX_PAGES 306#define KSTACK_MAX_PAGES 32 307#endif 308 309/* 310 * Create the kernel stack (including pcb for i386) for a new thread. 311 * This routine directly affects the fork perf for a process and 312 * create performance for a thread. 313 */ 314int 315vm_thread_new(struct thread *td, int pages) 316{ 317 vm_object_t ksobj; 318 vm_offset_t ks; 319 vm_page_t m, ma[KSTACK_MAX_PAGES]; 320 struct kstack_cache_entry *ks_ce; 321 int i; 322 323 /* Bounds check */ 324 if (pages <= 1) 325 pages = KSTACK_PAGES; 326 else if (pages > KSTACK_MAX_PAGES) 327 pages = KSTACK_MAX_PAGES; 328 329 if (pages == KSTACK_PAGES) { 330 mtx_lock(&kstack_cache_mtx); 331 if (kstack_cache != NULL) { 332 ks_ce = kstack_cache; 333 kstack_cache = ks_ce->next_ks_entry; 334 mtx_unlock(&kstack_cache_mtx); 335 336 td->td_kstack_obj = ks_ce->ksobj; 337 td->td_kstack = (vm_offset_t)ks_ce; 338 td->td_kstack_pages = KSTACK_PAGES; 339 return (1); 340 } 341 mtx_unlock(&kstack_cache_mtx); 342 } 343 344 /* 345 * Allocate an object for the kstack. 346 */ 347 ksobj = vm_object_allocate(OBJT_DEFAULT, pages); 348 349 /* 350 * Get a kernel virtual address for this thread's kstack. 351 */ 352#if defined(__mips__) 353 /* 354 * We need to align the kstack's mapped address to fit within 355 * a single TLB entry. 356 */ 357 ks = kmem_alloc_nofault_space(kernel_map, 358 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE, VMFS_TLB_ALIGNED_SPACE); 359#else 360 ks = kmem_alloc_nofault(kernel_map, 361 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 362#endif 363 if (ks == 0) { 364 printf("vm_thread_new: kstack allocation failed\n"); 365 vm_object_deallocate(ksobj); 366 return (0); 367 } 368 369 atomic_add_int(&kstacks, 1); 370 if (KSTACK_GUARD_PAGES != 0) { 371 pmap_qremove(ks, KSTACK_GUARD_PAGES); 372 ks += KSTACK_GUARD_PAGES * PAGE_SIZE; 373 } 374 td->td_kstack_obj = ksobj; 375 td->td_kstack = ks; 376 /* 377 * Knowing the number of pages allocated is useful when you 378 * want to deallocate them. 379 */ 380 td->td_kstack_pages = pages; 381 /* 382 * For the length of the stack, link in a real page of ram for each 383 * page of stack. 384 */ 385 VM_OBJECT_LOCK(ksobj); 386 for (i = 0; i < pages; i++) { 387 /* 388 * Get a kernel stack page. 389 */ 390 m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY | 391 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED); 392 ma[i] = m; 393 m->valid = VM_PAGE_BITS_ALL; 394 } 395 VM_OBJECT_UNLOCK(ksobj); 396 pmap_qenter(ks, ma, pages); 397 return (1); 398} 399 400static void 401vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages) 402{ 403 vm_page_t m; 404 int i; 405 406 atomic_add_int(&kstacks, -1); 407 pmap_qremove(ks, pages); 408 VM_OBJECT_LOCK(ksobj); 409 for (i = 0; i < pages; i++) { 410 m = vm_page_lookup(ksobj, i); 411 if (m == NULL) 412 panic("vm_thread_dispose: kstack already missing?"); 413 vm_page_lock(m); 414 vm_page_unwire(m, 0); 415 vm_page_free(m); 416 vm_page_unlock(m); 417 } 418 VM_OBJECT_UNLOCK(ksobj); 419 vm_object_deallocate(ksobj); 420 kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE), 421 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 422} 423 424/* 425 * Dispose of a thread's kernel stack. 426 */ 427void 428vm_thread_dispose(struct thread *td) 429{ 430 vm_object_t ksobj; 431 vm_offset_t ks; 432 struct kstack_cache_entry *ks_ce; 433 int pages; 434 435 pages = td->td_kstack_pages; 436 ksobj = td->td_kstack_obj; 437 ks = td->td_kstack; 438 td->td_kstack = 0; 439 td->td_kstack_pages = 0; 440 if (pages == KSTACK_PAGES && kstacks <= kstack_cache_size) { 441 ks_ce = (struct kstack_cache_entry *)ks; 442 ks_ce->ksobj = ksobj; 443 mtx_lock(&kstack_cache_mtx); 444 ks_ce->next_ks_entry = kstack_cache; 445 kstack_cache = ks_ce; 446 mtx_unlock(&kstack_cache_mtx); 447 return; 448 } 449 vm_thread_stack_dispose(ksobj, ks, pages); 450} 451 452static void 453vm_thread_stack_lowmem(void *nulll) 454{ 455 struct kstack_cache_entry *ks_ce, *ks_ce1; 456 457 mtx_lock(&kstack_cache_mtx); 458 ks_ce = kstack_cache; 459 kstack_cache = NULL; 460 mtx_unlock(&kstack_cache_mtx); 461 462 while (ks_ce != NULL) { 463 ks_ce1 = ks_ce; 464 ks_ce = ks_ce->next_ks_entry; 465 466 vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1, 467 KSTACK_PAGES); 468 } 469} 470 471static void 472kstack_cache_init(void *nulll) 473{ 474 475 EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL, 476 EVENTHANDLER_PRI_ANY); 477} 478 479MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF); 480SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL); 481 482#ifndef NO_SWAPPING 483/* 484 * Allow a thread's kernel stack to be paged out. 485 */ 486static void 487vm_thread_swapout(struct thread *td) 488{ 489 vm_object_t ksobj; 490 vm_page_t m; 491 int i, pages; 492 493 cpu_thread_swapout(td); 494 pages = td->td_kstack_pages; 495 ksobj = td->td_kstack_obj; 496 pmap_qremove(td->td_kstack, pages); 497 VM_OBJECT_LOCK(ksobj); 498 for (i = 0; i < pages; i++) { 499 m = vm_page_lookup(ksobj, i); 500 if (m == NULL) 501 panic("vm_thread_swapout: kstack already missing?"); 502 vm_page_dirty(m); 503 vm_page_lock(m); 504 vm_page_unwire(m, 0); 505 vm_page_unlock(m); 506 } 507 VM_OBJECT_UNLOCK(ksobj); 508} 509 510/* 511 * Bring the kernel stack for a specified thread back in. 512 */ 513static void 514vm_thread_swapin(struct thread *td) 515{ 516 vm_object_t ksobj; 517 vm_page_t ma[KSTACK_MAX_PAGES]; 518 int i, j, k, pages, rv; 519 520 pages = td->td_kstack_pages; 521 ksobj = td->td_kstack_obj; 522 VM_OBJECT_LOCK(ksobj); 523 for (i = 0; i < pages; i++) 524 ma[i] = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY | 525 VM_ALLOC_WIRED); 526 for (i = 0; i < pages; i++) { 527 if (ma[i]->valid != VM_PAGE_BITS_ALL) { 528 KASSERT(ma[i]->oflags & VPO_BUSY, 529 ("lost busy 1")); 530 vm_object_pip_add(ksobj, 1); 531 for (j = i + 1; j < pages; j++) { 532 KASSERT(ma[j]->valid == VM_PAGE_BITS_ALL || 533 (ma[j]->oflags & VPO_BUSY), 534 ("lost busy 2")); 535 if (ma[j]->valid == VM_PAGE_BITS_ALL) 536 break; 537 } 538 rv = vm_pager_get_pages(ksobj, ma + i, j - i, 0); 539 if (rv != VM_PAGER_OK) 540 panic("vm_thread_swapin: cannot get kstack for proc: %d", 541 td->td_proc->p_pid); 542 vm_object_pip_wakeup(ksobj); 543 for (k = i; k < j; k++) 544 ma[k] = vm_page_lookup(ksobj, k); 545 vm_page_wakeup(ma[i]); 546 } else if (ma[i]->oflags & VPO_BUSY) 547 vm_page_wakeup(ma[i]); 548 } 549 VM_OBJECT_UNLOCK(ksobj); 550 pmap_qenter(td->td_kstack, ma, pages); 551 cpu_thread_swapin(td); 552} 553#endif /* !NO_SWAPPING */ 554 555/* 556 * Implement fork's actions on an address space. 557 * Here we arrange for the address space to be copied or referenced, 558 * allocate a user struct (pcb and kernel stack), then call the 559 * machine-dependent layer to fill those in and make the new process 560 * ready to run. The new process is set up so that it returns directly 561 * to user mode to avoid stack copying and relocation problems. 562 */ 563int 564vm_forkproc(td, p2, td2, vm2, flags) 565 struct thread *td; 566 struct proc *p2; 567 struct thread *td2; 568 struct vmspace *vm2; 569 int flags; 570{ 571 struct proc *p1 = td->td_proc; 572 int error; 573 574 if ((flags & RFPROC) == 0) { 575 /* 576 * Divorce the memory, if it is shared, essentially 577 * this changes shared memory amongst threads, into 578 * COW locally. 579 */ 580 if ((flags & RFMEM) == 0) { 581 if (p1->p_vmspace->vm_refcnt > 1) { 582 error = vmspace_unshare(p1); 583 if (error) 584 return (error); 585 } 586 } 587 cpu_fork(td, p2, td2, flags); 588 return (0); 589 } 590 591 if (flags & RFMEM) { 592 p2->p_vmspace = p1->p_vmspace; 593 atomic_add_int(&p1->p_vmspace->vm_refcnt, 1); 594 } 595 596 while (vm_page_count_severe()) { 597 VM_WAIT; 598 } 599 600 if ((flags & RFMEM) == 0) { 601 p2->p_vmspace = vm2; 602 if (p1->p_vmspace->vm_shm) 603 shmfork(p1, p2); 604 } 605 606 /* 607 * cpu_fork will copy and update the pcb, set up the kernel stack, 608 * and make the child ready to run. 609 */ 610 cpu_fork(td, p2, td2, flags); 611 return (0); 612} 613 614/* 615 * Called after process has been wait(2)'ed apon and is being reaped. 616 * The idea is to reclaim resources that we could not reclaim while 617 * the process was still executing. 618 */ 619void 620vm_waitproc(p) 621 struct proc *p; 622{ 623 624 vmspace_exitfree(p); /* and clean-out the vmspace */ 625} 626 627void 628faultin(p) 629 struct proc *p; 630{ 631#ifdef NO_SWAPPING 632 633 PROC_LOCK_ASSERT(p, MA_OWNED); 634 if ((p->p_flag & P_INMEM) == 0) 635 panic("faultin: proc swapped out with NO_SWAPPING!"); 636#else /* !NO_SWAPPING */ 637 struct thread *td; 638 639 PROC_LOCK_ASSERT(p, MA_OWNED); 640 /* 641 * If another process is swapping in this process, 642 * just wait until it finishes. 643 */ 644 if (p->p_flag & P_SWAPPINGIN) { 645 while (p->p_flag & P_SWAPPINGIN) 646 msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0); 647 return; 648 } 649 if ((p->p_flag & P_INMEM) == 0) { 650 /* 651 * Don't let another thread swap process p out while we are 652 * busy swapping it in. 653 */ 654 ++p->p_lock; 655 p->p_flag |= P_SWAPPINGIN; 656 PROC_UNLOCK(p); 657 658 /* 659 * We hold no lock here because the list of threads 660 * can not change while all threads in the process are 661 * swapped out. 662 */ 663 FOREACH_THREAD_IN_PROC(p, td) 664 vm_thread_swapin(td); 665 PROC_LOCK(p); 666 swapclear(p); 667 p->p_swtick = ticks; 668 669 wakeup(&p->p_flag); 670 671 /* Allow other threads to swap p out now. */ 672 --p->p_lock; 673 } 674#endif /* NO_SWAPPING */ 675} 676 677/* 678 * This swapin algorithm attempts to swap-in processes only if there 679 * is enough space for them. Of course, if a process waits for a long 680 * time, it will be swapped in anyway. 681 * 682 * Giant is held on entry. 683 */ 684void 685swapper(void) 686{ 687 struct proc *p; 688 struct thread *td; 689 struct proc *pp; 690 int slptime; 691 int swtime; 692 int ppri; 693 int pri; 694 695loop: 696 if (vm_page_count_min()) { 697 VM_WAIT; 698 goto loop; 699 } 700 701 pp = NULL; 702 ppri = INT_MIN; 703 sx_slock(&allproc_lock); 704 FOREACH_PROC_IN_SYSTEM(p) { 705 PROC_LOCK(p); 706 if (p->p_state == PRS_NEW || 707 p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) { 708 PROC_UNLOCK(p); 709 continue; 710 } 711 swtime = (ticks - p->p_swtick) / hz; 712 FOREACH_THREAD_IN_PROC(p, td) { 713 /* 714 * An otherwise runnable thread of a process 715 * swapped out has only the TDI_SWAPPED bit set. 716 * 717 */ 718 thread_lock(td); 719 if (td->td_inhibitors == TDI_SWAPPED) { 720 slptime = (ticks - td->td_slptick) / hz; 721 pri = swtime + slptime; 722 if ((td->td_flags & TDF_SWAPINREQ) == 0) 723 pri -= p->p_nice * 8; 724 /* 725 * if this thread is higher priority 726 * and there is enough space, then select 727 * this process instead of the previous 728 * selection. 729 */ 730 if (pri > ppri) { 731 pp = p; 732 ppri = pri; 733 } 734 } 735 thread_unlock(td); 736 } 737 PROC_UNLOCK(p); 738 } 739 sx_sunlock(&allproc_lock); 740 741 /* 742 * Nothing to do, back to sleep. 743 */ 744 if ((p = pp) == NULL) { 745 tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2); 746 goto loop; 747 } 748 PROC_LOCK(p); 749 750 /* 751 * Another process may be bringing or may have already 752 * brought this process in while we traverse all threads. 753 * Or, this process may even be being swapped out again. 754 */ 755 if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) { 756 PROC_UNLOCK(p); 757 goto loop; 758 } 759 760 /* 761 * We would like to bring someone in. (only if there is space). 762 * [What checks the space? ] 763 */ 764 faultin(p); 765 PROC_UNLOCK(p); 766 goto loop; 767} 768 769void 770kick_proc0(void) 771{ 772 773 wakeup(&proc0); 774} 775 776#ifndef NO_SWAPPING 777 778/* 779 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 780 */ 781static int swap_idle_threshold1 = 2; 782SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, 783 &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); 784 785/* 786 * Swap_idle_threshold2 is the time that a process can be idle before 787 * it will be swapped out, if idle swapping is enabled. 788 */ 789static int swap_idle_threshold2 = 10; 790SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, 791 &swap_idle_threshold2, 0, "Time before a process will be swapped out"); 792 793/* 794 * First, if any processes have been sleeping or stopped for at least 795 * "swap_idle_threshold1" seconds, they are swapped out. If, however, 796 * no such processes exist, then the longest-sleeping or stopped 797 * process is swapped out. Finally, and only as a last resort, if 798 * there are no sleeping or stopped processes, the longest-resident 799 * process is swapped out. 800 */ 801void 802swapout_procs(action) 803int action; 804{ 805 struct proc *p; 806 struct thread *td; 807 int didswap = 0; 808 809retry: 810 sx_slock(&allproc_lock); 811 FOREACH_PROC_IN_SYSTEM(p) { 812 struct vmspace *vm; 813 int minslptime = 100000; 814 int slptime; 815 816 /* 817 * Watch out for a process in 818 * creation. It may have no 819 * address space or lock yet. 820 */ 821 if (p->p_state == PRS_NEW) 822 continue; 823 /* 824 * An aio daemon switches its 825 * address space while running. 826 * Perform a quick check whether 827 * a process has P_SYSTEM. 828 */ 829 if ((p->p_flag & P_SYSTEM) != 0) 830 continue; 831 /* 832 * Do not swapout a process that 833 * is waiting for VM data 834 * structures as there is a possible 835 * deadlock. Test this first as 836 * this may block. 837 * 838 * Lock the map until swapout 839 * finishes, or a thread of this 840 * process may attempt to alter 841 * the map. 842 */ 843 vm = vmspace_acquire_ref(p); 844 if (vm == NULL) 845 continue; 846 if (!vm_map_trylock(&vm->vm_map)) 847 goto nextproc1; 848 849 PROC_LOCK(p); 850 if (p->p_lock != 0 || 851 (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT) 852 ) != 0) { 853 goto nextproc; 854 } 855 /* 856 * only aiod changes vmspace, however it will be 857 * skipped because of the if statement above checking 858 * for P_SYSTEM 859 */ 860 if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM) 861 goto nextproc; 862 863 switch (p->p_state) { 864 default: 865 /* Don't swap out processes in any sort 866 * of 'special' state. */ 867 break; 868 869 case PRS_NORMAL: 870 /* 871 * do not swapout a realtime process 872 * Check all the thread groups.. 873 */ 874 FOREACH_THREAD_IN_PROC(p, td) { 875 thread_lock(td); 876 if (PRI_IS_REALTIME(td->td_pri_class)) { 877 thread_unlock(td); 878 goto nextproc; 879 } 880 slptime = (ticks - td->td_slptick) / hz; 881 /* 882 * Guarantee swap_idle_threshold1 883 * time in memory. 884 */ 885 if (slptime < swap_idle_threshold1) { 886 thread_unlock(td); 887 goto nextproc; 888 } 889 890 /* 891 * Do not swapout a process if it is 892 * waiting on a critical event of some 893 * kind or there is a thread whose 894 * pageable memory may be accessed. 895 * 896 * This could be refined to support 897 * swapping out a thread. 898 */ 899 if (!thread_safetoswapout(td)) { 900 thread_unlock(td); 901 goto nextproc; 902 } 903 /* 904 * If the system is under memory stress, 905 * or if we are swapping 906 * idle processes >= swap_idle_threshold2, 907 * then swap the process out. 908 */ 909 if (((action & VM_SWAP_NORMAL) == 0) && 910 (((action & VM_SWAP_IDLE) == 0) || 911 (slptime < swap_idle_threshold2))) { 912 thread_unlock(td); 913 goto nextproc; 914 } 915 916 if (minslptime > slptime) 917 minslptime = slptime; 918 thread_unlock(td); 919 } 920 921 /* 922 * If the pageout daemon didn't free enough pages, 923 * or if this process is idle and the system is 924 * configured to swap proactively, swap it out. 925 */ 926 if ((action & VM_SWAP_NORMAL) || 927 ((action & VM_SWAP_IDLE) && 928 (minslptime > swap_idle_threshold2))) { 929 if (swapout(p) == 0) 930 didswap++; 931 PROC_UNLOCK(p); 932 vm_map_unlock(&vm->vm_map); 933 vmspace_free(vm); 934 sx_sunlock(&allproc_lock); 935 goto retry; 936 } 937 } 938nextproc: 939 PROC_UNLOCK(p); 940 vm_map_unlock(&vm->vm_map); 941nextproc1: 942 vmspace_free(vm); 943 continue; 944 } 945 sx_sunlock(&allproc_lock); 946 /* 947 * If we swapped something out, and another process needed memory, 948 * then wakeup the sched process. 949 */ 950 if (didswap) 951 wakeup(&proc0); 952} 953 954static void 955swapclear(p) 956 struct proc *p; 957{ 958 struct thread *td; 959 960 PROC_LOCK_ASSERT(p, MA_OWNED); 961 962 FOREACH_THREAD_IN_PROC(p, td) { 963 thread_lock(td); 964 td->td_flags |= TDF_INMEM; 965 td->td_flags &= ~TDF_SWAPINREQ; 966 TD_CLR_SWAPPED(td); 967 if (TD_CAN_RUN(td)) 968 if (setrunnable(td)) { 969#ifdef INVARIANTS 970 /* 971 * XXX: We just cleared TDI_SWAPPED 972 * above and set TDF_INMEM, so this 973 * should never happen. 974 */ 975 panic("not waking up swapper"); 976#endif 977 } 978 thread_unlock(td); 979 } 980 p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT); 981 p->p_flag |= P_INMEM; 982} 983 984static int 985swapout(p) 986 struct proc *p; 987{ 988 struct thread *td; 989 990 PROC_LOCK_ASSERT(p, MA_OWNED); 991#if defined(SWAP_DEBUG) 992 printf("swapping out %d\n", p->p_pid); 993#endif 994 995 /* 996 * The states of this process and its threads may have changed 997 * by now. Assuming that there is only one pageout daemon thread, 998 * this process should still be in memory. 999 */ 1000 KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM, 1001 ("swapout: lost a swapout race?")); 1002 1003 /* 1004 * remember the process resident count 1005 */ 1006 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 1007 /* 1008 * Check and mark all threads before we proceed. 1009 */ 1010 p->p_flag &= ~P_INMEM; 1011 p->p_flag |= P_SWAPPINGOUT; 1012 FOREACH_THREAD_IN_PROC(p, td) { 1013 thread_lock(td); 1014 if (!thread_safetoswapout(td)) { 1015 thread_unlock(td); 1016 swapclear(p); 1017 return (EBUSY); 1018 } 1019 td->td_flags &= ~TDF_INMEM; 1020 TD_SET_SWAPPED(td); 1021 thread_unlock(td); 1022 } 1023 td = FIRST_THREAD_IN_PROC(p); 1024 ++td->td_ru.ru_nswap; 1025 PROC_UNLOCK(p); 1026 1027 /* 1028 * This list is stable because all threads are now prevented from 1029 * running. The list is only modified in the context of a running 1030 * thread in this process. 1031 */ 1032 FOREACH_THREAD_IN_PROC(p, td) 1033 vm_thread_swapout(td); 1034 1035 PROC_LOCK(p); 1036 p->p_flag &= ~P_SWAPPINGOUT; 1037 p->p_swtick = ticks; 1038 return (0); 1039} 1040#endif /* !NO_SWAPPING */ 1041