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