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