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 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 { 335 vm_object_t ksobj; 336 struct kstack_cache_entry *next_ks_entry; 337 }; 338 339 static struct kstack_cache_entry *kstack_cache; 340 static int kstack_cache_size = 128; 341 static int kstacks; 342 static struct mtx kstack_cache_mtx; 343 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0, 344 ""); 345 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0, 346 ""); 347 348 #ifndef KSTACK_MAX_PAGES 349 #define KSTACK_MAX_PAGES 32 350 #endif 351 352 /* 353 * Create the kernel stack (including pcb for i386) for a new thread. 354 * This routine directly affects the fork perf for a process and 355 * create performance for a thread. 356 */ 357 int 358 vm_thread_new(struct thread *td, int pages) 359 { 360 vm_object_t ksobj; 361 vm_offset_t ks; 362 vm_page_t m, ma[KSTACK_MAX_PAGES]; 363 struct kstack_cache_entry *ks_ce; 364 int i; 365 366 /* Bounds check */ 367 if (pages <= 1) 368 pages = KSTACK_PAGES; 369 else if (pages > KSTACK_MAX_PAGES) 370 pages = KSTACK_MAX_PAGES; 371 372 if (pages == KSTACK_PAGES) { 373 mtx_lock(&kstack_cache_mtx); 374 if (kstack_cache != NULL) { 375 ks_ce = kstack_cache; 376 kstack_cache = ks_ce->next_ks_entry; 377 mtx_unlock(&kstack_cache_mtx); 378 379 td->td_kstack_obj = ks_ce->ksobj; 380 td->td_kstack = (vm_offset_t)ks_ce; 381 td->td_kstack_pages = KSTACK_PAGES; 382 return (1); 383 } 384 mtx_unlock(&kstack_cache_mtx); 385 } 386 387 /* 388 * Allocate an object for the kstack. 389 */ 390 ksobj = vm_object_allocate(OBJT_DEFAULT, pages); 391 392 /* 393 * Get a kernel virtual address for this thread's kstack. 394 */ 395 #if defined(__mips__) 396 /* 397 * We need to align the kstack's mapped address to fit within 398 * a single TLB entry. 399 */ 400 ks = kmem_alloc_nofault_space(kernel_map, 401 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE, VMFS_TLB_ALIGNED_SPACE); 402 #else 403 ks = kmem_alloc_nofault(kernel_map, 404 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 405 #endif 406 if (ks == 0) { 407 printf("vm_thread_new: kstack allocation failed\n"); 408 vm_object_deallocate(ksobj); 409 return (0); 410 } 411 412 atomic_add_int(&kstacks, 1); 413 if (KSTACK_GUARD_PAGES != 0) { 414 pmap_qremove(ks, KSTACK_GUARD_PAGES); 415 ks += KSTACK_GUARD_PAGES * PAGE_SIZE; 416 } 417 td->td_kstack_obj = ksobj; 418 td->td_kstack = ks; 419 /* 420 * Knowing the number of pages allocated is useful when you 421 * want to deallocate them. 422 */ 423 td->td_kstack_pages = pages; 424 /* 425 * For the length of the stack, link in a real page of ram for each 426 * page of stack. 427 */ 428 VM_OBJECT_LOCK(ksobj); 429 for (i = 0; i < pages; i++) { 430 /* 431 * Get a kernel stack page. 432 */ 433 m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY | 434 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED); 435 ma[i] = m; 436 m->valid = VM_PAGE_BITS_ALL; 437 } 438 VM_OBJECT_UNLOCK(ksobj); 439 pmap_qenter(ks, ma, pages); 440 return (1); 441 } 442 443 static void 444 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages) 445 { 446 vm_page_t m; 447 int i; 448 449 atomic_add_int(&kstacks, -1); 450 pmap_qremove(ks, pages); 451 VM_OBJECT_LOCK(ksobj); 452 for (i = 0; i < pages; i++) { 453 m = vm_page_lookup(ksobj, i); 454 if (m == NULL) 455 panic("vm_thread_dispose: kstack already missing?"); 456 vm_page_lock(m); 457 vm_page_unwire(m, 0); 458 vm_page_free(m); 459 vm_page_unlock(m); 460 } 461 VM_OBJECT_UNLOCK(ksobj); 462 vm_object_deallocate(ksobj); 463 kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE), 464 (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); 465 } 466 467 /* 468 * Dispose of a thread's kernel stack. 469 */ 470 void 471 vm_thread_dispose(struct thread *td) 472 { 473 vm_object_t ksobj; 474 vm_offset_t ks; 475 struct kstack_cache_entry *ks_ce; 476 int pages; 477 478 pages = td->td_kstack_pages; 479 ksobj = td->td_kstack_obj; 480 ks = td->td_kstack; 481 td->td_kstack = 0; 482 td->td_kstack_pages = 0; 483 if (pages == KSTACK_PAGES && kstacks <= kstack_cache_size) { 484 ks_ce = (struct kstack_cache_entry *)ks; 485 ks_ce->ksobj = ksobj; 486 mtx_lock(&kstack_cache_mtx); 487 ks_ce->next_ks_entry = kstack_cache; 488 kstack_cache = ks_ce; 489 mtx_unlock(&kstack_cache_mtx); 490 return; 491 } 492 vm_thread_stack_dispose(ksobj, ks, pages); 493 } 494 495 static void 496 vm_thread_stack_lowmem(void *nulll) 497 { 498 struct kstack_cache_entry *ks_ce, *ks_ce1; 499 500 mtx_lock(&kstack_cache_mtx); 501 ks_ce = kstack_cache; 502 kstack_cache = NULL; 503 mtx_unlock(&kstack_cache_mtx); 504 505 while (ks_ce != NULL) { 506 ks_ce1 = ks_ce; 507 ks_ce = ks_ce->next_ks_entry; 508 509 vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1, 510 KSTACK_PAGES); 511 } 512 } 513 514 static void 515 kstack_cache_init(void *nulll) 516 { 517 518 EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL, 519 EVENTHANDLER_PRI_ANY); 520 } 521 522 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF); 523 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL); 524 525 #ifndef NO_SWAPPING 526 /* 527 * Allow a thread's kernel stack to be paged out. 528 */ 529 static void 530 vm_thread_swapout(struct thread *td) 531 { 532 vm_object_t ksobj; 533 vm_page_t m; 534 int i, pages; 535 536 cpu_thread_swapout(td); 537 pages = td->td_kstack_pages; 538 ksobj = td->td_kstack_obj; 539 pmap_qremove(td->td_kstack, pages); 540 VM_OBJECT_LOCK(ksobj); 541 for (i = 0; i < pages; i++) { 542 m = vm_page_lookup(ksobj, i); 543 if (m == NULL) 544 panic("vm_thread_swapout: kstack already missing?"); 545 vm_page_dirty(m); 546 vm_page_lock(m); 547 vm_page_unwire(m, 0); 548 vm_page_unlock(m); 549 } 550 VM_OBJECT_UNLOCK(ksobj); 551 } 552 553 /* 554 * Bring the kernel stack for a specified thread back in. 555 */ 556 static void 557 vm_thread_swapin(struct thread *td) 558 { 559 vm_object_t ksobj; 560 vm_page_t ma[KSTACK_MAX_PAGES]; 561 int i, j, k, pages, rv; 562 563 pages = td->td_kstack_pages; 564 ksobj = td->td_kstack_obj; 565 VM_OBJECT_LOCK(ksobj); 566 for (i = 0; i < pages; i++) 567 ma[i] = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY | 568 VM_ALLOC_WIRED); 569 for (i = 0; i < pages; i++) { 570 if (ma[i]->valid != VM_PAGE_BITS_ALL) { 571 KASSERT(ma[i]->oflags & VPO_BUSY, 572 ("lost busy 1")); 573 vm_object_pip_add(ksobj, 1); 574 for (j = i + 1; j < pages; j++) { 575 KASSERT(ma[j]->valid == VM_PAGE_BITS_ALL || 576 (ma[j]->oflags & VPO_BUSY), 577 ("lost busy 2")); 578 if (ma[j]->valid == VM_PAGE_BITS_ALL) 579 break; 580 } 581 rv = vm_pager_get_pages(ksobj, ma + i, j - i, 0); 582 if (rv != VM_PAGER_OK) 583 panic("vm_thread_swapin: cannot get kstack for proc: %d", 584 td->td_proc->p_pid); 585 vm_object_pip_wakeup(ksobj); 586 for (k = i; k < j; k++) 587 ma[k] = vm_page_lookup(ksobj, k); 588 vm_page_wakeup(ma[i]); 589 } else if (ma[i]->oflags & VPO_BUSY) 590 vm_page_wakeup(ma[i]); 591 } 592 VM_OBJECT_UNLOCK(ksobj); 593 pmap_qenter(td->td_kstack, ma, pages); 594 cpu_thread_swapin(td); 595 } 596 #endif /* !NO_SWAPPING */ 597 598 /* 599 * Implement fork's actions on an address space. 600 * Here we arrange for the address space to be copied or referenced, 601 * allocate a user struct (pcb and kernel stack), then call the 602 * machine-dependent layer to fill those in and make the new process 603 * ready to run. The new process is set up so that it returns directly 604 * to user mode to avoid stack copying and relocation problems. 605 */ 606 int 607 vm_forkproc(td, p2, td2, vm2, flags) 608 struct thread *td; 609 struct proc *p2; 610 struct thread *td2; 611 struct vmspace *vm2; 612 int flags; 613 { 614 struct proc *p1 = td->td_proc; 615 int error; 616 617 if ((flags & RFPROC) == 0) { 618 /* 619 * Divorce the memory, if it is shared, essentially 620 * this changes shared memory amongst threads, into 621 * COW locally. 622 */ 623 if ((flags & RFMEM) == 0) { 624 if (p1->p_vmspace->vm_refcnt > 1) { 625 error = vmspace_unshare(p1); 626 if (error) 627 return (error); 628 } 629 } 630 cpu_fork(td, p2, td2, flags); 631 return (0); 632 } 633 634 if (flags & RFMEM) { 635 p2->p_vmspace = p1->p_vmspace; 636 atomic_add_int(&p1->p_vmspace->vm_refcnt, 1); 637 } 638 639 while (vm_page_count_severe()) { 640 VM_WAIT; 641 } 642 643 if ((flags & RFMEM) == 0) { 644 p2->p_vmspace = vm2; 645 if (p1->p_vmspace->vm_shm) 646 shmfork(p1, p2); 647 } 648 649 /* 650 * cpu_fork will copy and update the pcb, set up the kernel stack, 651 * and make the child ready to run. 652 */ 653 cpu_fork(td, p2, td2, flags); 654 return (0); 655 } 656 657 /* 658 * Called after process has been wait(2)'ed apon and is being reaped. 659 * The idea is to reclaim resources that we could not reclaim while 660 * the process was still executing. 661 */ 662 void 663 vm_waitproc(p) 664 struct proc *p; 665 { 666 667 vmspace_exitfree(p); /* and clean-out the vmspace */ 668 } 669 670 void 671 faultin(p) 672 struct proc *p; 673 { 674 #ifdef NO_SWAPPING 675 676 PROC_LOCK_ASSERT(p, MA_OWNED); 677 if ((p->p_flag & P_INMEM) == 0) 678 panic("faultin: proc swapped out with NO_SWAPPING!"); 679 #else /* !NO_SWAPPING */ 680 struct thread *td; 681 682 PROC_LOCK_ASSERT(p, MA_OWNED); 683 /* 684 * If another process is swapping in this process, 685 * just wait until it finishes. 686 */ 687 if (p->p_flag & P_SWAPPINGIN) { 688 while (p->p_flag & P_SWAPPINGIN) 689 msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0); 690 return; 691 } 692 if ((p->p_flag & P_INMEM) == 0) { 693 /* 694 * Don't let another thread swap process p out while we are 695 * busy swapping it in. 696 */ 697 ++p->p_lock; 698 p->p_flag |= P_SWAPPINGIN; 699 PROC_UNLOCK(p); 700 701 /* 702 * We hold no lock here because the list of threads 703 * can not change while all threads in the process are 704 * swapped out. 705 */ 706 FOREACH_THREAD_IN_PROC(p, td) 707 vm_thread_swapin(td); 708 PROC_LOCK(p); 709 swapclear(p); 710 p->p_swtick = ticks; 711 712 wakeup(&p->p_flag); 713 714 /* Allow other threads to swap p out now. */ 715 --p->p_lock; 716 } 717 #endif /* NO_SWAPPING */ 718 } 719 720 /* 721 * This swapin algorithm attempts to swap-in processes only if there 722 * is enough space for them. Of course, if a process waits for a long 723 * time, it will be swapped in anyway. 724 * 725 * Giant is held on entry. 726 */ 727 /* ARGSUSED*/ 728 static void 729 scheduler(dummy) 730 void *dummy; 731 { 732 struct proc *p; 733 struct thread *td; 734 struct proc *pp; 735 int slptime; 736 int swtime; 737 int ppri; 738 int pri; 739 740 mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED); 741 mtx_unlock(&Giant); 742 743 loop: 744 if (vm_page_count_min()) { 745 VM_WAIT; 746 goto loop; 747 } 748 749 pp = NULL; 750 ppri = INT_MIN; 751 sx_slock(&allproc_lock); 752 FOREACH_PROC_IN_SYSTEM(p) { 753 PROC_LOCK(p); 754 if (p->p_state == PRS_NEW || 755 p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) { 756 PROC_UNLOCK(p); 757 continue; 758 } 759 swtime = (ticks - p->p_swtick) / hz; 760 FOREACH_THREAD_IN_PROC(p, td) { 761 /* 762 * An otherwise runnable thread of a process 763 * swapped out has only the TDI_SWAPPED bit set. 764 * 765 */ 766 thread_lock(td); 767 if (td->td_inhibitors == TDI_SWAPPED) { 768 slptime = (ticks - td->td_slptick) / hz; 769 pri = swtime + slptime; 770 if ((td->td_flags & TDF_SWAPINREQ) == 0) 771 pri -= p->p_nice * 8; 772 /* 773 * if this thread is higher priority 774 * and there is enough space, then select 775 * this process instead of the previous 776 * selection. 777 */ 778 if (pri > ppri) { 779 pp = p; 780 ppri = pri; 781 } 782 } 783 thread_unlock(td); 784 } 785 PROC_UNLOCK(p); 786 } 787 sx_sunlock(&allproc_lock); 788 789 /* 790 * Nothing to do, back to sleep. 791 */ 792 if ((p = pp) == NULL) { 793 tsleep(&proc0, PVM, "sched", MAXSLP * hz / 2); 794 goto loop; 795 } 796 PROC_LOCK(p); 797 798 /* 799 * Another process may be bringing or may have already 800 * brought this process in while we traverse all threads. 801 * Or, this process may even be being swapped out again. 802 */ 803 if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) { 804 PROC_UNLOCK(p); 805 goto loop; 806 } 807 808 /* 809 * We would like to bring someone in. (only if there is space). 810 * [What checks the space? ] 811 */ 812 faultin(p); 813 PROC_UNLOCK(p); 814 goto loop; 815 } 816 817 void 818 kick_proc0(void) 819 { 820 821 wakeup(&proc0); 822 } 823 824 #ifndef NO_SWAPPING 825 826 /* 827 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 828 */ 829 static int swap_idle_threshold1 = 2; 830 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, 831 &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); 832 833 /* 834 * Swap_idle_threshold2 is the time that a process can be idle before 835 * it will be swapped out, if idle swapping is enabled. 836 */ 837 static int swap_idle_threshold2 = 10; 838 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, 839 &swap_idle_threshold2, 0, "Time before a process will be swapped out"); 840 841 /* 842 * First, if any processes have been sleeping or stopped for at least 843 * "swap_idle_threshold1" seconds, they are swapped out. If, however, 844 * no such processes exist, then the longest-sleeping or stopped 845 * process is swapped out. Finally, and only as a last resort, if 846 * there are no sleeping or stopped processes, the longest-resident 847 * process is swapped out. 848 */ 849 void 850 swapout_procs(action) 851 int action; 852 { 853 struct proc *p; 854 struct thread *td; 855 int didswap = 0; 856 857 retry: 858 sx_slock(&allproc_lock); 859 FOREACH_PROC_IN_SYSTEM(p) { 860 struct vmspace *vm; 861 int minslptime = 100000; 862 int slptime; 863 864 /* 865 * Watch out for a process in 866 * creation. It may have no 867 * address space or lock yet. 868 */ 869 if (p->p_state == PRS_NEW) 870 continue; 871 /* 872 * An aio daemon switches its 873 * address space while running. 874 * Perform a quick check whether 875 * a process has P_SYSTEM. 876 */ 877 if ((p->p_flag & P_SYSTEM) != 0) 878 continue; 879 /* 880 * Do not swapout a process that 881 * is waiting for VM data 882 * structures as there is a possible 883 * deadlock. Test this first as 884 * this may block. 885 * 886 * Lock the map until swapout 887 * finishes, or a thread of this 888 * process may attempt to alter 889 * the map. 890 */ 891 vm = vmspace_acquire_ref(p); 892 if (vm == NULL) 893 continue; 894 if (!vm_map_trylock(&vm->vm_map)) 895 goto nextproc1; 896 897 PROC_LOCK(p); 898 if (p->p_lock != 0 || 899 (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT) 900 ) != 0) { 901 goto nextproc; 902 } 903 /* 904 * only aiod changes vmspace, however it will be 905 * skipped because of the if statement above checking 906 * for P_SYSTEM 907 */ 908 if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM) 909 goto nextproc; 910 911 switch (p->p_state) { 912 default: 913 /* Don't swap out processes in any sort 914 * of 'special' state. */ 915 break; 916 917 case PRS_NORMAL: 918 /* 919 * do not swapout a realtime process 920 * Check all the thread groups.. 921 */ 922 FOREACH_THREAD_IN_PROC(p, td) { 923 thread_lock(td); 924 if (PRI_IS_REALTIME(td->td_pri_class)) { 925 thread_unlock(td); 926 goto nextproc; 927 } 928 slptime = (ticks - td->td_slptick) / hz; 929 /* 930 * Guarantee swap_idle_threshold1 931 * time in memory. 932 */ 933 if (slptime < swap_idle_threshold1) { 934 thread_unlock(td); 935 goto nextproc; 936 } 937 938 /* 939 * Do not swapout a process if it is 940 * waiting on a critical event of some 941 * kind or there is a thread whose 942 * pageable memory may be accessed. 943 * 944 * This could be refined to support 945 * swapping out a thread. 946 */ 947 if (!thread_safetoswapout(td)) { 948 thread_unlock(td); 949 goto nextproc; 950 } 951 /* 952 * If the system is under memory stress, 953 * or if we are swapping 954 * idle processes >= swap_idle_threshold2, 955 * then swap the process out. 956 */ 957 if (((action & VM_SWAP_NORMAL) == 0) && 958 (((action & VM_SWAP_IDLE) == 0) || 959 (slptime < swap_idle_threshold2))) { 960 thread_unlock(td); 961 goto nextproc; 962 } 963 964 if (minslptime > slptime) 965 minslptime = slptime; 966 thread_unlock(td); 967 } 968 969 /* 970 * If the pageout daemon didn't free enough pages, 971 * or if this process is idle and the system is 972 * configured to swap proactively, swap it out. 973 */ 974 if ((action & VM_SWAP_NORMAL) || 975 ((action & VM_SWAP_IDLE) && 976 (minslptime > swap_idle_threshold2))) { 977 if (swapout(p) == 0) 978 didswap++; 979 PROC_UNLOCK(p); 980 vm_map_unlock(&vm->vm_map); 981 vmspace_free(vm); 982 sx_sunlock(&allproc_lock); 983 goto retry; 984 } 985 } 986 nextproc: 987 PROC_UNLOCK(p); 988 vm_map_unlock(&vm->vm_map); 989 nextproc1: 990 vmspace_free(vm); 991 continue; 992 } 993 sx_sunlock(&allproc_lock); 994 /* 995 * If we swapped something out, and another process needed memory, 996 * then wakeup the sched process. 997 */ 998 if (didswap) 999 wakeup(&proc0); 1000 } 1001 1002 static void 1003 swapclear(p) 1004 struct proc *p; 1005 { 1006 struct thread *td; 1007 1008 PROC_LOCK_ASSERT(p, MA_OWNED); 1009 1010 FOREACH_THREAD_IN_PROC(p, td) { 1011 thread_lock(td); 1012 td->td_flags |= TDF_INMEM; 1013 td->td_flags &= ~TDF_SWAPINREQ; 1014 TD_CLR_SWAPPED(td); 1015 if (TD_CAN_RUN(td)) 1016 if (setrunnable(td)) { 1017 #ifdef INVARIANTS 1018 /* 1019 * XXX: We just cleared TDI_SWAPPED 1020 * above and set TDF_INMEM, so this 1021 * should never happen. 1022 */ 1023 panic("not waking up swapper"); 1024 #endif 1025 } 1026 thread_unlock(td); 1027 } 1028 p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT); 1029 p->p_flag |= P_INMEM; 1030 } 1031 1032 static int 1033 swapout(p) 1034 struct proc *p; 1035 { 1036 struct thread *td; 1037 1038 PROC_LOCK_ASSERT(p, MA_OWNED); 1039 #if defined(SWAP_DEBUG) 1040 printf("swapping out %d\n", p->p_pid); 1041 #endif 1042 1043 /* 1044 * The states of this process and its threads may have changed 1045 * by now. Assuming that there is only one pageout daemon thread, 1046 * this process should still be in memory. 1047 */ 1048 KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM, 1049 ("swapout: lost a swapout race?")); 1050 1051 /* 1052 * remember the process resident count 1053 */ 1054 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 1055 /* 1056 * Check and mark all threads before we proceed. 1057 */ 1058 p->p_flag &= ~P_INMEM; 1059 p->p_flag |= P_SWAPPINGOUT; 1060 FOREACH_THREAD_IN_PROC(p, td) { 1061 thread_lock(td); 1062 if (!thread_safetoswapout(td)) { 1063 thread_unlock(td); 1064 swapclear(p); 1065 return (EBUSY); 1066 } 1067 td->td_flags &= ~TDF_INMEM; 1068 TD_SET_SWAPPED(td); 1069 thread_unlock(td); 1070 } 1071 td = FIRST_THREAD_IN_PROC(p); 1072 ++td->td_ru.ru_nswap; 1073 PROC_UNLOCK(p); 1074 1075 /* 1076 * This list is stable because all threads are now prevented from 1077 * running. The list is only modified in the context of a running 1078 * thread in this process. 1079 */ 1080 FOREACH_THREAD_IN_PROC(p, td) 1081 vm_thread_swapout(td); 1082 1083 PROC_LOCK(p); 1084 p->p_flag &= ~P_SWAPPINGOUT; 1085 p->p_swtick = ticks; 1086 return (0); 1087 } 1088 #endif /* !NO_SWAPPING */ 1089