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. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_glue.c 8.6 (Berkeley) 1/5/94 37 * 38 * 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 * 62 * $FreeBSD: src/sys/vm/vm_glue.c,v 1.94.2.4 2003/01/13 22:51:17 dillon Exp $ 63 * $DragonFly: src/sys/vm/vm_glue.c,v 1.41 2006/05/25 07:36:37 dillon Exp $ 64 */ 65 66 #include "opt_vm.h" 67 68 #include <sys/param.h> 69 #include <sys/systm.h> 70 #include <sys/proc.h> 71 #include <sys/resourcevar.h> 72 #include <sys/buf.h> 73 #include <sys/shm.h> 74 #include <sys/vmmeter.h> 75 #include <sys/sysctl.h> 76 77 #include <sys/kernel.h> 78 #include <sys/unistd.h> 79 80 #include <machine/limits.h> 81 82 #include <vm/vm.h> 83 #include <vm/vm_param.h> 84 #include <sys/lock.h> 85 #include <vm/pmap.h> 86 #include <vm/vm_map.h> 87 #include <vm/vm_page.h> 88 #include <vm/vm_pageout.h> 89 #include <vm/vm_kern.h> 90 #include <vm/vm_extern.h> 91 92 #include <sys/user.h> 93 #include <vm/vm_page2.h> 94 #include <sys/thread2.h> 95 96 /* 97 * System initialization 98 * 99 * Note: proc0 from proc.h 100 */ 101 102 static void vm_init_limits (void *); 103 SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0) 104 105 /* 106 * THIS MUST BE THE LAST INITIALIZATION ITEM!!! 107 * 108 * Note: run scheduling should be divorced from the vm system. 109 */ 110 static void scheduler (void *); 111 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL) 112 113 #ifdef INVARIANTS 114 115 static int swap_debug = 0; 116 SYSCTL_INT(_vm, OID_AUTO, swap_debug, 117 CTLFLAG_RW, &swap_debug, 0, ""); 118 119 #endif 120 121 static int scheduler_notify; 122 123 static void swapout (struct proc *); 124 125 int 126 kernacc(c_caddr_t addr, int len, int rw) 127 { 128 boolean_t rv; 129 vm_offset_t saddr, eaddr; 130 vm_prot_t prot; 131 132 KASSERT((rw & (~VM_PROT_ALL)) == 0, 133 ("illegal ``rw'' argument to kernacc (%x)\n", rw)); 134 135 if ((vm_offset_t)addr + len > kernel_map->max_offset || 136 (vm_offset_t)addr + len < (vm_offset_t)addr) { 137 return (FALSE); 138 } 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 if (rv == FALSE && is_globaldata_space(saddr, eaddr)) 147 rv = TRUE; 148 return (rv == TRUE); 149 } 150 151 int 152 useracc(c_caddr_t addr, int len, int rw) 153 { 154 boolean_t rv; 155 vm_prot_t prot; 156 vm_map_t map; 157 vm_map_entry_t save_hint; 158 159 KASSERT((rw & (~VM_PROT_ALL)) == 0, 160 ("illegal ``rw'' argument to useracc (%x)\n", rw)); 161 prot = rw; 162 /* 163 * XXX - check separately to disallow access to user area and user 164 * page tables - they are in the map. 165 * 166 * XXX - VM_MAXUSER_ADDRESS is an end address, not a max. It was once 167 * only used (as an end address) in trap.c. Use it as an end address 168 * here too. This bogusness has spread. I just fixed where it was 169 * used as a max in vm_mmap.c. 170 */ 171 if ((vm_offset_t) addr + len > /* XXX */ VM_MAXUSER_ADDRESS 172 || (vm_offset_t) addr + len < (vm_offset_t) addr) { 173 return (FALSE); 174 } 175 map = &curproc->p_vmspace->vm_map; 176 vm_map_lock_read(map); 177 /* 178 * We save the map hint, and restore it. Useracc appears to distort 179 * the map hint unnecessarily. 180 */ 181 save_hint = map->hint; 182 rv = vm_map_check_protection(map, 183 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), prot); 184 map->hint = save_hint; 185 vm_map_unlock_read(map); 186 187 return (rv == TRUE); 188 } 189 190 void 191 vslock(caddr_t addr, u_int len) 192 { 193 vm_map_wire(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr), 194 round_page((vm_offset_t)addr + len), 0); 195 } 196 197 void 198 vsunlock(caddr_t addr, u_int len) 199 { 200 vm_map_wire(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr), 201 round_page((vm_offset_t)addr + len), KM_PAGEABLE); 202 } 203 204 /* 205 * Implement fork's actions on an address space. 206 * Here we arrange for the address space to be copied or referenced, 207 * allocate a user struct (pcb and kernel stack), then call the 208 * machine-dependent layer to fill those in and make the new process 209 * ready to run. The new process is set up so that it returns directly 210 * to user mode to avoid stack copying and relocation problems. 211 */ 212 void 213 vm_fork(struct proc *p1, struct proc *p2, int flags) 214 { 215 struct user *up; 216 struct thread *td2; 217 218 if ((flags & RFPROC) == 0) { 219 /* 220 * Divorce the memory, if it is shared, essentially 221 * this changes shared memory amongst threads, into 222 * COW locally. 223 */ 224 if ((flags & RFMEM) == 0) { 225 if (p1->p_vmspace->vm_refcnt > 1) { 226 vmspace_unshare(p1); 227 } 228 } 229 cpu_fork(p1, p2, flags); 230 return; 231 } 232 233 if (flags & RFMEM) { 234 p2->p_vmspace = p1->p_vmspace; 235 p1->p_vmspace->vm_refcnt++; 236 } 237 238 while (vm_page_count_severe()) { 239 vm_wait(); 240 } 241 242 if ((flags & RFMEM) == 0) { 243 p2->p_vmspace = vmspace_fork(p1->p_vmspace); 244 245 pmap_pinit2(vmspace_pmap(p2->p_vmspace)); 246 247 if (p1->p_vmspace->vm_shm) 248 shmfork(p1, p2); 249 } 250 251 td2 = lwkt_alloc_thread(NULL, LWKT_THREAD_STACK, -1, 0); 252 pmap_init_proc(p2, td2); 253 lwkt_setpri(td2, TDPRI_KERN_USER); 254 lwkt_set_comm(td2, "%s", p1->p_comm); 255 256 up = p2->p_addr; 257 258 /* 259 * p_stats currently points at fields in the user struct 260 * but not at &u, instead at p_addr. Copy parts of 261 * p_stats; zero the rest of p_stats (statistics). 262 * 263 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need 264 * to share sigacts, so we use the up->u_sigacts. 265 */ 266 p2->p_stats = &up->u_stats; 267 if (p2->p_sigacts == NULL) { 268 if (p2->p_procsig->ps_refcnt != 1) 269 printf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid); 270 p2->p_sigacts = &up->u_sigacts; 271 up->u_sigacts = *p1->p_sigacts; 272 } 273 274 bzero(&up->u_stats, sizeof(struct pstats)); 275 276 /* 277 * cpu_fork will copy and update the pcb, set up the kernel stack, 278 * and make the child ready to run. 279 */ 280 cpu_fork(p1, p2, flags); 281 } 282 283 /* 284 * Called after process has been wait(2)'ed apon and is being reaped. 285 * The idea is to reclaim resources that we could not reclaim while 286 * the process was still executing. 287 */ 288 void 289 vm_waitproc(struct proc *p) 290 { 291 p->p_stats = NULL; 292 cpu_proc_wait(p); 293 vmspace_exitfree(p); /* and clean-out the vmspace */ 294 } 295 296 /* 297 * Set default limits for VM system. 298 * Called for proc 0, and then inherited by all others. 299 * 300 * XXX should probably act directly on proc0. 301 */ 302 static void 303 vm_init_limits(void *udata) 304 { 305 struct proc *p = udata; 306 int rss_limit; 307 308 /* 309 * Set up the initial limits on process VM. Set the maximum resident 310 * set size to be half of (reasonably) available memory. Since this 311 * is a soft limit, it comes into effect only when the system is out 312 * of memory - half of main memory helps to favor smaller processes, 313 * and reduces thrashing of the object cache. 314 */ 315 p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz; 316 p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz; 317 p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz; 318 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz; 319 /* limit the limit to no less than 2MB */ 320 rss_limit = max(vmstats.v_free_count, 512); 321 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit); 322 p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY; 323 } 324 325 /* 326 * Faultin the specified process. Note that the process can be in any 327 * state. Just clear P_SWAPPEDOUT and call wakeup in case the process is 328 * sleeping. 329 */ 330 void 331 faultin(struct proc *p) 332 { 333 if (p->p_flag & P_SWAPPEDOUT) { 334 /* 335 * The process is waiting in the kernel to return to user 336 * mode but cannot until P_SWAPPEDOUT gets cleared. 337 */ 338 crit_enter(); 339 p->p_flag &= ~(P_SWAPPEDOUT | P_SWAPWAIT); 340 #ifdef INVARIANTS 341 if (swap_debug) 342 printf("swapping in %d (%s)\n", p->p_pid, p->p_comm); 343 #endif 344 wakeup(p); 345 346 crit_exit(); 347 } 348 } 349 350 /* 351 * Kernel initialization eventually falls through to this function, 352 * which is process 0. 353 * 354 * This swapin algorithm attempts to swap-in processes only if there 355 * is enough space for them. Of course, if a process waits for a long 356 * time, it will be swapped in anyway. 357 */ 358 359 struct scheduler_info { 360 struct proc *pp; 361 int ppri; 362 }; 363 364 static int scheduler_callback(struct proc *p, void *data); 365 366 static void 367 scheduler(void *dummy) 368 { 369 struct scheduler_info info; 370 struct proc *p; 371 372 KKASSERT(!IN_CRITICAL_SECT(curthread)); 373 loop: 374 scheduler_notify = 0; 375 /* 376 * Don't try to swap anything in if we are low on memory. 377 */ 378 if (vm_page_count_min()) { 379 vm_wait(); 380 goto loop; 381 } 382 383 /* 384 * Look for a good candidate to wake up 385 */ 386 info.pp = NULL; 387 info.ppri = INT_MIN; 388 allproc_scan(scheduler_callback, &info); 389 390 /* 391 * Nothing to do, back to sleep for at least 1/10 of a second. If 392 * we are woken up, immediately process the next request. If 393 * multiple requests have built up the first is processed 394 * immediately and the rest are staggered. 395 */ 396 if ((p = info.pp) == NULL) { 397 tsleep(&proc0, 0, "nowork", hz / 10); 398 if (scheduler_notify == 0) 399 tsleep(&scheduler_notify, 0, "nowork", 0); 400 goto loop; 401 } 402 403 /* 404 * Fault the selected process in, then wait for a short period of 405 * time and loop up. 406 * 407 * XXX we need a heuristic to get a measure of system stress and 408 * then adjust our stagger wakeup delay accordingly. 409 */ 410 faultin(p); 411 p->p_swtime = 0; 412 PRELE(p); 413 tsleep(&proc0, 0, "swapin", hz / 10); 414 goto loop; 415 } 416 417 static int 418 scheduler_callback(struct proc *p, void *data) 419 { 420 struct scheduler_info *info = data; 421 segsz_t pgs; 422 int pri; 423 424 if (p->p_flag & P_SWAPWAIT) { 425 pri = p->p_swtime + p->p_slptime - p->p_nice * 8; 426 427 /* 428 * The more pages paged out while we were swapped, 429 * the more work we have to do to get up and running 430 * again and the lower our wakeup priority. 431 * 432 * Each second of sleep time is worth ~1MB 433 */ 434 pgs = vmspace_resident_count(p->p_vmspace); 435 if (pgs < p->p_vmspace->vm_swrss) { 436 pri -= (p->p_vmspace->vm_swrss - pgs) / 437 (1024 * 1024 / PAGE_SIZE); 438 } 439 440 /* 441 * If this process is higher priority and there is 442 * enough space, then select this process instead of 443 * the previous selection. 444 */ 445 if (pri > info->ppri) { 446 if (info->pp) 447 PRELE(info->pp); 448 PHOLD(p); 449 info->pp = p; 450 info->ppri = pri; 451 } 452 } 453 return(0); 454 } 455 456 void 457 swapin_request(void) 458 { 459 if (scheduler_notify == 0) { 460 scheduler_notify = 1; 461 wakeup(&scheduler_notify); 462 } 463 } 464 465 #ifndef NO_SWAPPING 466 467 #define swappable(p) \ 468 (((p)->p_lock == 0) && \ 469 ((p)->p_flag & (P_TRACED|P_SYSTEM|P_SWAPPEDOUT|P_WEXIT)) == 0) 470 471 472 /* 473 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 474 */ 475 static int swap_idle_threshold1 = 15; 476 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, 477 CTLFLAG_RW, &swap_idle_threshold1, 0, ""); 478 479 /* 480 * Swap_idle_threshold2 is the time that a process can be idle before 481 * it will be swapped out, if idle swapping is enabled. Default is 482 * one minute. 483 */ 484 static int swap_idle_threshold2 = 60; 485 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, 486 CTLFLAG_RW, &swap_idle_threshold2, 0, ""); 487 488 /* 489 * Swapout is driven by the pageout daemon. Very simple, we find eligible 490 * procs and mark them as being swapped out. This will cause the kernel 491 * to prefer to pageout those proc's pages first and the procs in question 492 * will not return to user mode until the swapper tells them they can. 493 * 494 * If any procs have been sleeping/stopped for at least maxslp seconds, 495 * they are swapped. Else, we swap the longest-sleeping or stopped process, 496 * if any, otherwise the longest-resident process. 497 */ 498 499 static int swapout_procs_callback(struct proc *p, void *data); 500 501 void 502 swapout_procs(int action) 503 { 504 allproc_scan(swapout_procs_callback, &action); 505 } 506 507 static int 508 swapout_procs_callback(struct proc *p, void *data) 509 { 510 struct vmspace *vm; 511 int action = *(int *)data; 512 513 if (!swappable(p)) 514 return(0); 515 516 vm = p->p_vmspace; 517 518 if (p->p_stat == SSLEEP || p->p_stat == SRUN) { 519 /* 520 * do not swap out a realtime process 521 */ 522 if (RTP_PRIO_IS_REALTIME(p->p_lwp.lwp_rtprio.type)) 523 return(0); 524 525 /* 526 * Guarentee swap_idle_threshold time in memory 527 */ 528 if (p->p_slptime < swap_idle_threshold1) 529 return(0); 530 531 /* 532 * If the system is under memory stress, or if we 533 * are swapping idle processes >= swap_idle_threshold2, 534 * then swap the process out. 535 */ 536 if (((action & VM_SWAP_NORMAL) == 0) && 537 (((action & VM_SWAP_IDLE) == 0) || 538 (p->p_slptime < swap_idle_threshold2))) { 539 return(0); 540 } 541 542 ++vm->vm_refcnt; 543 544 /* 545 * If the process has been asleep for awhile, swap 546 * it out. 547 */ 548 if ((action & VM_SWAP_NORMAL) || 549 ((action & VM_SWAP_IDLE) && 550 (p->p_slptime > swap_idle_threshold2))) { 551 swapout(p); 552 } 553 554 /* 555 * cleanup our reference 556 */ 557 vmspace_free(vm); 558 } 559 return(0); 560 } 561 562 static void 563 swapout(struct proc *p) 564 { 565 #ifdef INVARIANTS 566 if (swap_debug) 567 printf("swapping out %d (%s)\n", p->p_pid, p->p_comm); 568 #endif 569 ++p->p_stats->p_ru.ru_nswap; 570 /* 571 * remember the process resident count 572 */ 573 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 574 p->p_flag |= P_SWAPPEDOUT; 575 p->p_swtime = 0; 576 } 577 578 #endif /* !NO_SWAPPING */ 579 580