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. 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 #include "opt_kstack_usage_prof.h" 66 67 #include <sys/param.h> 68 #include <sys/systm.h> 69 #include <sys/limits.h> 70 #include <sys/lock.h> 71 #include <sys/malloc.h> 72 #include <sys/mutex.h> 73 #include <sys/proc.h> 74 #include <sys/racct.h> 75 #include <sys/resourcevar.h> 76 #include <sys/rwlock.h> 77 #include <sys/sched.h> 78 #include <sys/sf_buf.h> 79 #include <sys/shm.h> 80 #include <sys/vmmeter.h> 81 #include <sys/vmem.h> 82 #include <sys/sx.h> 83 #include <sys/sysctl.h> 84 #include <sys/_kstack_cache.h> 85 #include <sys/eventhandler.h> 86 #include <sys/kernel.h> 87 #include <sys/ktr.h> 88 #include <sys/unistd.h> 89 90 #include <vm/vm.h> 91 #include <vm/vm_param.h> 92 #include <vm/pmap.h> 93 #include <vm/vm_map.h> 94 #include <vm/vm_page.h> 95 #include <vm/vm_pageout.h> 96 #include <vm/vm_object.h> 97 #include <vm/vm_kern.h> 98 #include <vm/vm_extern.h> 99 #include <vm/vm_pager.h> 100 #include <vm/swap_pager.h> 101 102 #include <machine/cpu.h> 103 104 #ifndef NO_SWAPPING 105 static int swapout(struct proc *); 106 static void swapclear(struct proc *); 107 static void vm_thread_swapin(struct thread *td); 108 static void vm_thread_swapout(struct thread *td); 109 #endif 110 111 /* 112 * MPSAFE 113 * 114 * WARNING! This code calls vm_map_check_protection() which only checks 115 * the associated vm_map_entry range. It does not determine whether the 116 * contents of the memory is actually readable or writable. In most cases 117 * just checking the vm_map_entry is sufficient within the kernel's address 118 * space. 119 */ 120 int 121 kernacc(addr, len, rw) 122 void *addr; 123 int len, rw; 124 { 125 boolean_t rv; 126 vm_offset_t saddr, eaddr; 127 vm_prot_t prot; 128 129 KASSERT((rw & ~VM_PROT_ALL) == 0, 130 ("illegal ``rw'' argument to kernacc (%x)\n", rw)); 131 132 if ((vm_offset_t)addr + len > kernel_map->max_offset || 133 (vm_offset_t)addr + len < (vm_offset_t)addr) 134 return (FALSE); 135 136 prot = rw; 137 saddr = trunc_page((vm_offset_t)addr); 138 eaddr = round_page((vm_offset_t)addr + len); 139 vm_map_lock_read(kernel_map); 140 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot); 141 vm_map_unlock_read(kernel_map); 142 return (rv == TRUE); 143 } 144 145 /* 146 * MPSAFE 147 * 148 * WARNING! This code calls vm_map_check_protection() which only checks 149 * the associated vm_map_entry range. It does not determine whether the 150 * contents of the memory is actually readable or writable. vmapbuf(), 151 * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be 152 * used in conjunction with this call. 153 */ 154 int 155 useracc(addr, len, rw) 156 void *addr; 157 int len, rw; 158 { 159 boolean_t rv; 160 vm_prot_t prot; 161 vm_map_t map; 162 163 KASSERT((rw & ~VM_PROT_ALL) == 0, 164 ("illegal ``rw'' argument to useracc (%x)\n", rw)); 165 prot = rw; 166 map = &curproc->p_vmspace->vm_map; 167 if ((vm_offset_t)addr + len > vm_map_max(map) || 168 (vm_offset_t)addr + len < (vm_offset_t)addr) { 169 return (FALSE); 170 } 171 vm_map_lock_read(map); 172 rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr), 173 round_page((vm_offset_t)addr + len), prot); 174 vm_map_unlock_read(map); 175 return (rv == TRUE); 176 } 177 178 int 179 vslock(void *addr, size_t len) 180 { 181 vm_offset_t end, last, start; 182 vm_size_t npages; 183 int error; 184 185 last = (vm_offset_t)addr + len; 186 start = trunc_page((vm_offset_t)addr); 187 end = round_page(last); 188 if (last < (vm_offset_t)addr || end < (vm_offset_t)addr) 189 return (EINVAL); 190 npages = atop(end - start); 191 if (npages > vm_page_max_wired) 192 return (ENOMEM); 193 #if 0 194 /* 195 * XXX - not yet 196 * 197 * The limit for transient usage of wired pages should be 198 * larger than for "permanent" wired pages (mlock()). 199 * 200 * Also, the sysctl code, which is the only present user 201 * of vslock(), does a hard loop on EAGAIN. 202 */ 203 if (npages + vm_cnt.v_wire_count > vm_page_max_wired) 204 return (EAGAIN); 205 #endif 206 error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end, 207 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 208 /* 209 * Return EFAULT on error to match copy{in,out}() behaviour 210 * rather than returning ENOMEM like mlock() would. 211 */ 212 return (error == KERN_SUCCESS ? 0 : EFAULT); 213 } 214 215 void 216 vsunlock(void *addr, size_t len) 217 { 218 219 /* Rely on the parameter sanity checks performed by vslock(). */ 220 (void)vm_map_unwire(&curproc->p_vmspace->vm_map, 221 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), 222 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 223 } 224 225 /* 226 * Pin the page contained within the given object at the given offset. If the 227 * page is not resident, allocate and load it using the given object's pager. 228 * Return the pinned page if successful; otherwise, return NULL. 229 */ 230 static vm_page_t 231 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset) 232 { 233 vm_page_t m; 234 vm_pindex_t pindex; 235 int rv; 236 237 VM_OBJECT_WLOCK(object); 238 pindex = OFF_TO_IDX(offset); 239 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 240 if (m->valid != VM_PAGE_BITS_ALL) { 241 vm_page_xbusy(m); 242 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 243 if (rv != VM_PAGER_OK) { 244 vm_page_lock(m); 245 vm_page_free(m); 246 vm_page_unlock(m); 247 m = NULL; 248 goto out; 249 } 250 vm_page_xunbusy(m); 251 } 252 vm_page_lock(m); 253 vm_page_hold(m); 254 vm_page_activate(m); 255 vm_page_unlock(m); 256 out: 257 VM_OBJECT_WUNLOCK(object); 258 return (m); 259 } 260 261 /* 262 * Return a CPU private mapping to the page at the given offset within the 263 * given object. The page is pinned before it is mapped. 264 */ 265 struct sf_buf * 266 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset) 267 { 268 vm_page_t m; 269 270 m = vm_imgact_hold_page(object, offset); 271 if (m == NULL) 272 return (NULL); 273 sched_pin(); 274 return (sf_buf_alloc(m, SFB_CPUPRIVATE)); 275 } 276 277 /* 278 * Destroy the given CPU private mapping and unpin the page that it mapped. 279 */ 280 void 281 vm_imgact_unmap_page(struct sf_buf *sf) 282 { 283 vm_page_t m; 284 285 m = sf_buf_page(sf); 286 sf_buf_free(sf); 287 sched_unpin(); 288 vm_page_lock(m); 289 vm_page_unhold(m); 290 vm_page_unlock(m); 291 } 292 293 void 294 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz) 295 { 296 297 pmap_sync_icache(map->pmap, va, sz); 298 } 299 300 struct kstack_cache_entry *kstack_cache; 301 static int kstack_cache_size = 128; 302 static int kstacks; 303 static struct mtx kstack_cache_mtx; 304 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF); 305 306 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0, 307 ""); 308 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0, 309 ""); 310 311 #ifndef KSTACK_MAX_PAGES 312 #define KSTACK_MAX_PAGES 32 313 #endif 314 315 /* 316 * Create the kernel stack (including pcb for i386) for a new thread. 317 * This routine directly affects the fork perf for a process and 318 * create performance for a thread. 319 */ 320 int 321 vm_thread_new(struct thread *td, int pages) 322 { 323 vm_object_t ksobj; 324 vm_offset_t ks; 325 vm_page_t ma[KSTACK_MAX_PAGES]; 326 struct kstack_cache_entry *ks_ce; 327 int i; 328 329 /* Bounds check */ 330 if (pages <= 1) 331 pages = kstack_pages; 332 else if (pages > KSTACK_MAX_PAGES) 333 pages = KSTACK_MAX_PAGES; 334 335 if (pages == kstack_pages) { 336 mtx_lock(&kstack_cache_mtx); 337 if (kstack_cache != NULL) { 338 ks_ce = kstack_cache; 339 kstack_cache = ks_ce->next_ks_entry; 340 mtx_unlock(&kstack_cache_mtx); 341 342 td->td_kstack_obj = ks_ce->ksobj; 343 td->td_kstack = (vm_offset_t)ks_ce; 344 td->td_kstack_pages = kstack_pages; 345 return (1); 346 } 347 mtx_unlock(&kstack_cache_mtx); 348 } 349 350 /* 351 * Allocate an object for the kstack. 352 */ 353 ksobj = vm_object_allocate(OBJT_DEFAULT, pages); 354 355 /* 356 * Get a kernel virtual address for this thread's kstack. 357 */ 358 #if defined(__mips__) 359 /* 360 * We need to align the kstack's mapped address to fit within 361 * a single TLB entry. 362 */ 363 if (vmem_xalloc(kernel_arena, (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE, 364 PAGE_SIZE * 2, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, 365 M_BESTFIT | M_NOWAIT, &ks)) { 366 ks = 0; 367 } 368 #else 369 ks = kva_alloc((pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 370 #endif 371 if (ks == 0) { 372 printf("vm_thread_new: kstack allocation failed\n"); 373 vm_object_deallocate(ksobj); 374 return (0); 375 } 376 377 atomic_add_int(&kstacks, 1); 378 if (KSTACK_GUARD_PAGES != 0) { 379 pmap_qremove(ks, KSTACK_GUARD_PAGES); 380 ks += KSTACK_GUARD_PAGES * PAGE_SIZE; 381 } 382 td->td_kstack_obj = ksobj; 383 td->td_kstack = ks; 384 /* 385 * Knowing the number of pages allocated is useful when you 386 * want to deallocate them. 387 */ 388 td->td_kstack_pages = pages; 389 /* 390 * For the length of the stack, link in a real page of ram for each 391 * page of stack. 392 */ 393 VM_OBJECT_WLOCK(ksobj); 394 (void)vm_page_grab_pages(ksobj, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY | 395 VM_ALLOC_WIRED, ma, pages); 396 for (i = 0; i < pages; i++) 397 ma[i]->valid = VM_PAGE_BITS_ALL; 398 VM_OBJECT_WUNLOCK(ksobj); 399 pmap_qenter(ks, ma, pages); 400 return (1); 401 } 402 403 static void 404 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages) 405 { 406 vm_page_t m; 407 int i; 408 409 atomic_add_int(&kstacks, -1); 410 pmap_qremove(ks, pages); 411 VM_OBJECT_WLOCK(ksobj); 412 for (i = 0; i < pages; i++) { 413 m = vm_page_lookup(ksobj, i); 414 if (m == NULL) 415 panic("vm_thread_dispose: kstack already missing?"); 416 vm_page_lock(m); 417 vm_page_unwire(m, PQ_NONE); 418 vm_page_free(m); 419 vm_page_unlock(m); 420 } 421 VM_OBJECT_WUNLOCK(ksobj); 422 vm_object_deallocate(ksobj); 423 kva_free(ks - (KSTACK_GUARD_PAGES * PAGE_SIZE), 424 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 425 } 426 427 /* 428 * Dispose of a thread's kernel stack. 429 */ 430 void 431 vm_thread_dispose(struct thread *td) 432 { 433 vm_object_t ksobj; 434 vm_offset_t ks; 435 struct kstack_cache_entry *ks_ce; 436 int pages; 437 438 pages = td->td_kstack_pages; 439 ksobj = td->td_kstack_obj; 440 ks = td->td_kstack; 441 td->td_kstack = 0; 442 td->td_kstack_pages = 0; 443 if (pages == kstack_pages && kstacks <= kstack_cache_size) { 444 ks_ce = (struct kstack_cache_entry *)ks; 445 ks_ce->ksobj = ksobj; 446 mtx_lock(&kstack_cache_mtx); 447 ks_ce->next_ks_entry = kstack_cache; 448 kstack_cache = ks_ce; 449 mtx_unlock(&kstack_cache_mtx); 450 return; 451 } 452 vm_thread_stack_dispose(ksobj, ks, pages); 453 } 454 455 static void 456 vm_thread_stack_lowmem(void *nulll) 457 { 458 struct kstack_cache_entry *ks_ce, *ks_ce1; 459 460 mtx_lock(&kstack_cache_mtx); 461 ks_ce = kstack_cache; 462 kstack_cache = NULL; 463 mtx_unlock(&kstack_cache_mtx); 464 465 while (ks_ce != NULL) { 466 ks_ce1 = ks_ce; 467 ks_ce = ks_ce->next_ks_entry; 468 469 vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1, 470 kstack_pages); 471 } 472 } 473 474 static void 475 kstack_cache_init(void *nulll) 476 { 477 478 EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL, 479 EVENTHANDLER_PRI_ANY); 480 } 481 482 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL); 483 484 #ifdef KSTACK_USAGE_PROF 485 /* 486 * Track maximum stack used by a thread in kernel. 487 */ 488 static int max_kstack_used; 489 490 SYSCTL_INT(_debug, OID_AUTO, max_kstack_used, CTLFLAG_RD, 491 &max_kstack_used, 0, 492 "Maxiumum stack depth used by a thread in kernel"); 493 494 void 495 intr_prof_stack_use(struct thread *td, struct trapframe *frame) 496 { 497 vm_offset_t stack_top; 498 vm_offset_t current; 499 int used, prev_used; 500 501 /* 502 * Testing for interrupted kernel mode isn't strictly 503 * needed. It optimizes the execution, since interrupts from 504 * usermode will have only the trap frame on the stack. 505 */ 506 if (TRAPF_USERMODE(frame)) 507 return; 508 509 stack_top = td->td_kstack + td->td_kstack_pages * PAGE_SIZE; 510 current = (vm_offset_t)(uintptr_t)&stack_top; 511 512 /* 513 * Try to detect if interrupt is using kernel thread stack. 514 * Hardware could use a dedicated stack for interrupt handling. 515 */ 516 if (stack_top <= current || current < td->td_kstack) 517 return; 518 519 used = stack_top - current; 520 for (;;) { 521 prev_used = max_kstack_used; 522 if (prev_used >= used) 523 break; 524 if (atomic_cmpset_int(&max_kstack_used, prev_used, used)) 525 break; 526 } 527 } 528 #endif /* KSTACK_USAGE_PROF */ 529 530 #ifndef NO_SWAPPING 531 /* 532 * Allow a thread's kernel stack to be paged out. 533 */ 534 static void 535 vm_thread_swapout(struct thread *td) 536 { 537 vm_object_t ksobj; 538 vm_page_t m; 539 int i, pages; 540 541 cpu_thread_swapout(td); 542 pages = td->td_kstack_pages; 543 ksobj = td->td_kstack_obj; 544 pmap_qremove(td->td_kstack, pages); 545 VM_OBJECT_WLOCK(ksobj); 546 for (i = 0; i < pages; i++) { 547 m = vm_page_lookup(ksobj, i); 548 if (m == NULL) 549 panic("vm_thread_swapout: kstack already missing?"); 550 vm_page_dirty(m); 551 vm_page_lock(m); 552 vm_page_unwire(m, PQ_INACTIVE); 553 vm_page_unlock(m); 554 } 555 VM_OBJECT_WUNLOCK(ksobj); 556 } 557 558 /* 559 * Bring the kernel stack for a specified thread back in. 560 */ 561 static void 562 vm_thread_swapin(struct thread *td) 563 { 564 vm_object_t ksobj; 565 vm_page_t ma[KSTACK_MAX_PAGES]; 566 int pages; 567 568 pages = td->td_kstack_pages; 569 ksobj = td->td_kstack_obj; 570 VM_OBJECT_WLOCK(ksobj); 571 (void)vm_page_grab_pages(ksobj, 0, VM_ALLOC_NORMAL | VM_ALLOC_WIRED, ma, 572 pages); 573 for (int i = 0; i < pages;) { 574 int j, a, count, rv; 575 576 vm_page_assert_xbusied(ma[i]); 577 if (ma[i]->valid == VM_PAGE_BITS_ALL) { 578 vm_page_xunbusy(ma[i]); 579 i++; 580 continue; 581 } 582 vm_object_pip_add(ksobj, 1); 583 for (j = i + 1; j < pages; j++) 584 if (ma[j]->valid == VM_PAGE_BITS_ALL) 585 break; 586 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a); 587 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i])); 588 count = min(a + 1, j - i); 589 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL); 590 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d", 591 __func__, td->td_proc->p_pid)); 592 vm_object_pip_wakeup(ksobj); 593 for (j = i; j < i + count; j++) 594 vm_page_xunbusy(ma[j]); 595 i += count; 596 } 597 VM_OBJECT_WUNLOCK(ksobj); 598 pmap_qenter(td->td_kstack, ma, pages); 599 cpu_thread_swapin(td); 600 } 601 #endif /* !NO_SWAPPING */ 602 603 /* 604 * Implement fork's actions on an address space. 605 * Here we arrange for the address space to be copied or referenced, 606 * allocate a user struct (pcb and kernel stack), then call the 607 * machine-dependent layer to fill those in and make the new process 608 * ready to run. The new process is set up so that it returns directly 609 * to user mode to avoid stack copying and relocation problems. 610 */ 611 int 612 vm_forkproc(td, p2, td2, vm2, flags) 613 struct thread *td; 614 struct proc *p2; 615 struct thread *td2; 616 struct vmspace *vm2; 617 int flags; 618 { 619 struct proc *p1 = td->td_proc; 620 int error; 621 622 if ((flags & RFPROC) == 0) { 623 /* 624 * Divorce the memory, if it is shared, essentially 625 * this changes shared memory amongst threads, into 626 * COW locally. 627 */ 628 if ((flags & RFMEM) == 0) { 629 if (p1->p_vmspace->vm_refcnt > 1) { 630 error = vmspace_unshare(p1); 631 if (error) 632 return (error); 633 } 634 } 635 cpu_fork(td, p2, td2, flags); 636 return (0); 637 } 638 639 if (flags & RFMEM) { 640 p2->p_vmspace = p1->p_vmspace; 641 atomic_add_int(&p1->p_vmspace->vm_refcnt, 1); 642 } 643 644 while (vm_page_count_severe()) { 645 VM_WAIT; 646 } 647 648 if ((flags & RFMEM) == 0) { 649 p2->p_vmspace = vm2; 650 if (p1->p_vmspace->vm_shm) 651 shmfork(p1, p2); 652 } 653 654 /* 655 * cpu_fork will copy and update the pcb, set up the kernel stack, 656 * and make the child ready to run. 657 */ 658 cpu_fork(td, p2, td2, flags); 659 return (0); 660 } 661 662 /* 663 * Called after process has been wait(2)'ed upon and is being reaped. 664 * The idea is to reclaim resources that we could not reclaim while 665 * the process was still executing. 666 */ 667 void 668 vm_waitproc(p) 669 struct proc *p; 670 { 671 672 vmspace_exitfree(p); /* and clean-out the vmspace */ 673 } 674 675 void 676 faultin(p) 677 struct proc *p; 678 { 679 #ifdef NO_SWAPPING 680 681 PROC_LOCK_ASSERT(p, MA_OWNED); 682 if ((p->p_flag & P_INMEM) == 0) 683 panic("faultin: proc swapped out with NO_SWAPPING!"); 684 #else /* !NO_SWAPPING */ 685 struct thread *td; 686 687 PROC_LOCK_ASSERT(p, MA_OWNED); 688 /* 689 * If another process is swapping in this process, 690 * just wait until it finishes. 691 */ 692 if (p->p_flag & P_SWAPPINGIN) { 693 while (p->p_flag & P_SWAPPINGIN) 694 msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0); 695 return; 696 } 697 if ((p->p_flag & P_INMEM) == 0) { 698 /* 699 * Don't let another thread swap process p out while we are 700 * busy swapping it in. 701 */ 702 ++p->p_lock; 703 p->p_flag |= P_SWAPPINGIN; 704 PROC_UNLOCK(p); 705 706 /* 707 * We hold no lock here because the list of threads 708 * can not change while all threads in the process are 709 * swapped out. 710 */ 711 FOREACH_THREAD_IN_PROC(p, td) 712 vm_thread_swapin(td); 713 PROC_LOCK(p); 714 swapclear(p); 715 p->p_swtick = ticks; 716 717 wakeup(&p->p_flag); 718 719 /* Allow other threads to swap p out now. */ 720 --p->p_lock; 721 } 722 #endif /* NO_SWAPPING */ 723 } 724 725 /* 726 * This swapin algorithm attempts to swap-in processes only if there 727 * is enough space for them. Of course, if a process waits for a long 728 * time, it will be swapped in anyway. 729 */ 730 void 731 swapper(void) 732 { 733 struct proc *p; 734 struct thread *td; 735 struct proc *pp; 736 int slptime; 737 int swtime; 738 int ppri; 739 int pri; 740 741 loop: 742 if (vm_page_count_min()) { 743 VM_WAIT; 744 goto loop; 745 } 746 747 pp = NULL; 748 ppri = INT_MIN; 749 sx_slock(&allproc_lock); 750 FOREACH_PROC_IN_SYSTEM(p) { 751 PROC_LOCK(p); 752 if (p->p_state == PRS_NEW || 753 p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) { 754 PROC_UNLOCK(p); 755 continue; 756 } 757 swtime = (ticks - p->p_swtick) / hz; 758 FOREACH_THREAD_IN_PROC(p, td) { 759 /* 760 * An otherwise runnable thread of a process 761 * swapped out has only the TDI_SWAPPED bit set. 762 * 763 */ 764 thread_lock(td); 765 if (td->td_inhibitors == TDI_SWAPPED) { 766 slptime = (ticks - td->td_slptick) / hz; 767 pri = swtime + slptime; 768 if ((td->td_flags & TDF_SWAPINREQ) == 0) 769 pri -= p->p_nice * 8; 770 /* 771 * if this thread is higher priority 772 * and there is enough space, then select 773 * this process instead of the previous 774 * selection. 775 */ 776 if (pri > ppri) { 777 pp = p; 778 ppri = pri; 779 } 780 } 781 thread_unlock(td); 782 } 783 PROC_UNLOCK(p); 784 } 785 sx_sunlock(&allproc_lock); 786 787 /* 788 * Nothing to do, back to sleep. 789 */ 790 if ((p = pp) == NULL) { 791 tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2); 792 goto loop; 793 } 794 PROC_LOCK(p); 795 796 /* 797 * Another process may be bringing or may have already 798 * brought this process in while we traverse all threads. 799 * Or, this process may even be being swapped out again. 800 */ 801 if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) { 802 PROC_UNLOCK(p); 803 goto loop; 804 } 805 806 /* 807 * We would like to bring someone in. (only if there is space). 808 * [What checks the space? ] 809 */ 810 faultin(p); 811 PROC_UNLOCK(p); 812 goto loop; 813 } 814 815 void 816 kick_proc0(void) 817 { 818 819 wakeup(&proc0); 820 } 821 822 #ifndef NO_SWAPPING 823 824 /* 825 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 826 */ 827 static int swap_idle_threshold1 = 2; 828 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, 829 &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); 830 831 /* 832 * Swap_idle_threshold2 is the time that a process can be idle before 833 * it will be swapped out, if idle swapping is enabled. 834 */ 835 static int swap_idle_threshold2 = 10; 836 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, 837 &swap_idle_threshold2, 0, "Time before a process will be swapped out"); 838 839 /* 840 * First, if any processes have been sleeping or stopped for at least 841 * "swap_idle_threshold1" seconds, they are swapped out. If, however, 842 * no such processes exist, then the longest-sleeping or stopped 843 * process is swapped out. Finally, and only as a last resort, if 844 * there are no sleeping or stopped processes, the longest-resident 845 * process is swapped out. 846 */ 847 void 848 swapout_procs(action) 849 int action; 850 { 851 struct proc *p; 852 struct thread *td; 853 int didswap = 0; 854 855 retry: 856 sx_slock(&allproc_lock); 857 FOREACH_PROC_IN_SYSTEM(p) { 858 struct vmspace *vm; 859 int minslptime = 100000; 860 int slptime; 861 862 PROC_LOCK(p); 863 /* 864 * Watch out for a process in 865 * creation. It may have no 866 * address space or lock yet. 867 */ 868 if (p->p_state == PRS_NEW) { 869 PROC_UNLOCK(p); 870 continue; 871 } 872 /* 873 * An aio daemon switches its 874 * address space while running. 875 * Perform a quick check whether 876 * a process has P_SYSTEM. 877 * Filter out exiting processes. 878 */ 879 if ((p->p_flag & (P_SYSTEM | P_WEXIT)) != 0) { 880 PROC_UNLOCK(p); 881 continue; 882 } 883 _PHOLD_LITE(p); 884 PROC_UNLOCK(p); 885 sx_sunlock(&allproc_lock); 886 887 /* 888 * Do not swapout a process that 889 * is waiting for VM data 890 * structures as there is a possible 891 * deadlock. Test this first as 892 * this may block. 893 * 894 * Lock the map until swapout 895 * finishes, or a thread of this 896 * process may attempt to alter 897 * the map. 898 */ 899 vm = vmspace_acquire_ref(p); 900 if (vm == NULL) 901 goto nextproc2; 902 if (!vm_map_trylock(&vm->vm_map)) 903 goto nextproc1; 904 905 PROC_LOCK(p); 906 if (p->p_lock != 1 || (p->p_flag & (P_STOPPED_SINGLE | 907 P_TRACED | P_SYSTEM)) != 0) 908 goto nextproc; 909 910 /* 911 * only aiod changes vmspace, however it will be 912 * skipped because of the if statement above checking 913 * for P_SYSTEM 914 */ 915 if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM) 916 goto nextproc; 917 918 switch (p->p_state) { 919 default: 920 /* Don't swap out processes in any sort 921 * of 'special' state. */ 922 break; 923 924 case PRS_NORMAL: 925 /* 926 * do not swapout a realtime process 927 * Check all the thread groups.. 928 */ 929 FOREACH_THREAD_IN_PROC(p, td) { 930 thread_lock(td); 931 if (PRI_IS_REALTIME(td->td_pri_class)) { 932 thread_unlock(td); 933 goto nextproc; 934 } 935 slptime = (ticks - td->td_slptick) / hz; 936 /* 937 * Guarantee swap_idle_threshold1 938 * time in memory. 939 */ 940 if (slptime < swap_idle_threshold1) { 941 thread_unlock(td); 942 goto nextproc; 943 } 944 945 /* 946 * Do not swapout a process if it is 947 * waiting on a critical event of some 948 * kind or there is a thread whose 949 * pageable memory may be accessed. 950 * 951 * This could be refined to support 952 * swapping out a thread. 953 */ 954 if (!thread_safetoswapout(td)) { 955 thread_unlock(td); 956 goto nextproc; 957 } 958 /* 959 * If the system is under memory stress, 960 * or if we are swapping 961 * idle processes >= swap_idle_threshold2, 962 * then swap the process out. 963 */ 964 if (((action & VM_SWAP_NORMAL) == 0) && 965 (((action & VM_SWAP_IDLE) == 0) || 966 (slptime < swap_idle_threshold2))) { 967 thread_unlock(td); 968 goto nextproc; 969 } 970 971 if (minslptime > slptime) 972 minslptime = slptime; 973 thread_unlock(td); 974 } 975 976 /* 977 * If the pageout daemon didn't free enough pages, 978 * or if this process is idle and the system is 979 * configured to swap proactively, swap it out. 980 */ 981 if ((action & VM_SWAP_NORMAL) || 982 ((action & VM_SWAP_IDLE) && 983 (minslptime > swap_idle_threshold2))) { 984 _PRELE(p); 985 if (swapout(p) == 0) 986 didswap++; 987 PROC_UNLOCK(p); 988 vm_map_unlock(&vm->vm_map); 989 vmspace_free(vm); 990 goto retry; 991 } 992 } 993 nextproc: 994 PROC_UNLOCK(p); 995 vm_map_unlock(&vm->vm_map); 996 nextproc1: 997 vmspace_free(vm); 998 nextproc2: 999 sx_slock(&allproc_lock); 1000 PRELE(p); 1001 } 1002 sx_sunlock(&allproc_lock); 1003 /* 1004 * If we swapped something out, and another process needed memory, 1005 * then wakeup the sched process. 1006 */ 1007 if (didswap) 1008 wakeup(&proc0); 1009 } 1010 1011 static void 1012 swapclear(p) 1013 struct proc *p; 1014 { 1015 struct thread *td; 1016 1017 PROC_LOCK_ASSERT(p, MA_OWNED); 1018 1019 FOREACH_THREAD_IN_PROC(p, td) { 1020 thread_lock(td); 1021 td->td_flags |= TDF_INMEM; 1022 td->td_flags &= ~TDF_SWAPINREQ; 1023 TD_CLR_SWAPPED(td); 1024 if (TD_CAN_RUN(td)) 1025 if (setrunnable(td)) { 1026 #ifdef INVARIANTS 1027 /* 1028 * XXX: We just cleared TDI_SWAPPED 1029 * above and set TDF_INMEM, so this 1030 * should never happen. 1031 */ 1032 panic("not waking up swapper"); 1033 #endif 1034 } 1035 thread_unlock(td); 1036 } 1037 p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT); 1038 p->p_flag |= P_INMEM; 1039 } 1040 1041 static int 1042 swapout(p) 1043 struct proc *p; 1044 { 1045 struct thread *td; 1046 1047 PROC_LOCK_ASSERT(p, MA_OWNED); 1048 #if defined(SWAP_DEBUG) 1049 printf("swapping out %d\n", p->p_pid); 1050 #endif 1051 1052 /* 1053 * The states of this process and its threads may have changed 1054 * by now. Assuming that there is only one pageout daemon thread, 1055 * this process should still be in memory. 1056 */ 1057 KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM, 1058 ("swapout: lost a swapout race?")); 1059 1060 /* 1061 * remember the process resident count 1062 */ 1063 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 1064 /* 1065 * Check and mark all threads before we proceed. 1066 */ 1067 p->p_flag &= ~P_INMEM; 1068 p->p_flag |= P_SWAPPINGOUT; 1069 FOREACH_THREAD_IN_PROC(p, td) { 1070 thread_lock(td); 1071 if (!thread_safetoswapout(td)) { 1072 thread_unlock(td); 1073 swapclear(p); 1074 return (EBUSY); 1075 } 1076 td->td_flags &= ~TDF_INMEM; 1077 TD_SET_SWAPPED(td); 1078 thread_unlock(td); 1079 } 1080 td = FIRST_THREAD_IN_PROC(p); 1081 ++td->td_ru.ru_nswap; 1082 PROC_UNLOCK(p); 1083 1084 /* 1085 * This list is stable because all threads are now prevented from 1086 * running. The list is only modified in the context of a running 1087 * thread in this process. 1088 */ 1089 FOREACH_THREAD_IN_PROC(p, td) 1090 vm_thread_swapout(td); 1091 1092 PROC_LOCK(p); 1093 p->p_flag &= ~P_SWAPPINGOUT; 1094 p->p_swtick = ticks; 1095 return (0); 1096 } 1097 #endif /* !NO_SWAPPING */ 1098