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 1.42 2006/06/27 16:38:42 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 /* 136 * The globaldata space is not part of the kernel_map proper, 137 * check access separately. 138 */ 139 if (is_globaldata_space((vm_offset_t)addr, (vm_offset_t)(addr + len))) 140 return (TRUE); 141 142 /* 143 * Nominal kernel memory access - check access via kernel_map. 144 */ 145 if ((vm_offset_t)addr + len > kernel_map->max_offset || 146 (vm_offset_t)addr + len < (vm_offset_t)addr) { 147 return (FALSE); 148 } 149 prot = rw; 150 saddr = trunc_page((vm_offset_t)addr); 151 eaddr = round_page((vm_offset_t)addr + len); 152 vm_map_lock_read(kernel_map); 153 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot); 154 vm_map_unlock_read(kernel_map); 155 return (rv == TRUE); 156 } 157 158 int 159 useracc(c_caddr_t addr, int len, int rw) 160 { 161 boolean_t rv; 162 vm_prot_t prot; 163 vm_map_t map; 164 vm_map_entry_t save_hint; 165 166 KASSERT((rw & (~VM_PROT_ALL)) == 0, 167 ("illegal ``rw'' argument to useracc (%x)\n", rw)); 168 prot = rw; 169 /* 170 * XXX - check separately to disallow access to user area and user 171 * page tables - they are in the map. 172 * 173 * XXX - VM_MAXUSER_ADDRESS is an end address, not a max. It was once 174 * only used (as an end address) in trap.c. Use it as an end address 175 * here too. This bogusness has spread. I just fixed where it was 176 * used as a max in vm_mmap.c. 177 */ 178 if ((vm_offset_t) addr + len > /* XXX */ VM_MAXUSER_ADDRESS 179 || (vm_offset_t) addr + len < (vm_offset_t) addr) { 180 return (FALSE); 181 } 182 map = &curproc->p_vmspace->vm_map; 183 vm_map_lock_read(map); 184 /* 185 * We save the map hint, and restore it. Useracc appears to distort 186 * the map hint unnecessarily. 187 */ 188 save_hint = map->hint; 189 rv = vm_map_check_protection(map, 190 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), prot); 191 map->hint = save_hint; 192 vm_map_unlock_read(map); 193 194 return (rv == TRUE); 195 } 196 197 void 198 vslock(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), 0); 202 } 203 204 void 205 vsunlock(caddr_t addr, u_int len) 206 { 207 vm_map_wire(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr), 208 round_page((vm_offset_t)addr + len), KM_PAGEABLE); 209 } 210 211 /* 212 * Implement fork's actions on an address space. 213 * Here we arrange for the address space to be copied or referenced, 214 * allocate a user struct (pcb and kernel stack), then call the 215 * machine-dependent layer to fill those in and make the new process 216 * ready to run. The new process is set up so that it returns directly 217 * to user mode to avoid stack copying and relocation problems. 218 */ 219 void 220 vm_fork(struct proc *p1, struct proc *p2, int flags) 221 { 222 struct user *up; 223 struct thread *td2; 224 225 if ((flags & RFPROC) == 0) { 226 /* 227 * Divorce the memory, if it is shared, essentially 228 * this changes shared memory amongst threads, into 229 * COW locally. 230 */ 231 if ((flags & RFMEM) == 0) { 232 if (p1->p_vmspace->vm_refcnt > 1) { 233 vmspace_unshare(p1); 234 } 235 } 236 cpu_fork(p1, p2, flags); 237 return; 238 } 239 240 if (flags & RFMEM) { 241 p2->p_vmspace = p1->p_vmspace; 242 p1->p_vmspace->vm_refcnt++; 243 } 244 245 while (vm_page_count_severe()) { 246 vm_wait(); 247 } 248 249 if ((flags & RFMEM) == 0) { 250 p2->p_vmspace = vmspace_fork(p1->p_vmspace); 251 252 pmap_pinit2(vmspace_pmap(p2->p_vmspace)); 253 254 if (p1->p_vmspace->vm_shm) 255 shmfork(p1, p2); 256 } 257 258 td2 = lwkt_alloc_thread(NULL, LWKT_THREAD_STACK, -1, 0); 259 pmap_init_proc(p2, td2); 260 lwkt_setpri(td2, TDPRI_KERN_USER); 261 lwkt_set_comm(td2, "%s", p1->p_comm); 262 263 up = p2->p_addr; 264 265 /* 266 * p_stats currently points at fields in the user struct 267 * but not at &u, instead at p_addr. Copy parts of 268 * p_stats; zero the rest of p_stats (statistics). 269 * 270 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need 271 * to share sigacts, so we use the up->u_sigacts. 272 */ 273 p2->p_stats = &up->u_stats; 274 if (p2->p_sigacts == NULL) { 275 if (p2->p_procsig->ps_refcnt != 1) 276 printf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid); 277 p2->p_sigacts = &up->u_sigacts; 278 up->u_sigacts = *p1->p_sigacts; 279 } 280 281 bzero(&up->u_stats, sizeof(struct pstats)); 282 283 /* 284 * cpu_fork will copy and update the pcb, set up the kernel stack, 285 * and make the child ready to run. 286 */ 287 cpu_fork(p1, p2, flags); 288 } 289 290 /* 291 * Called after process has been wait(2)'ed apon and is being reaped. 292 * The idea is to reclaim resources that we could not reclaim while 293 * the process was still executing. 294 */ 295 void 296 vm_waitproc(struct proc *p) 297 { 298 p->p_stats = NULL; 299 cpu_proc_wait(p); 300 vmspace_exitfree(p); /* and clean-out the vmspace */ 301 } 302 303 /* 304 * Set default limits for VM system. 305 * Called for proc 0, and then inherited by all others. 306 * 307 * XXX should probably act directly on proc0. 308 */ 309 static void 310 vm_init_limits(void *udata) 311 { 312 struct proc *p = udata; 313 int rss_limit; 314 315 /* 316 * Set up the initial limits on process VM. Set the maximum resident 317 * set size to be half of (reasonably) available memory. Since this 318 * is a soft limit, it comes into effect only when the system is out 319 * of memory - half of main memory helps to favor smaller processes, 320 * and reduces thrashing of the object cache. 321 */ 322 p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz; 323 p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz; 324 p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz; 325 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz; 326 /* limit the limit to no less than 2MB */ 327 rss_limit = max(vmstats.v_free_count, 512); 328 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit); 329 p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY; 330 } 331 332 /* 333 * Faultin the specified process. Note that the process can be in any 334 * state. Just clear P_SWAPPEDOUT and call wakeup in case the process is 335 * sleeping. 336 */ 337 void 338 faultin(struct proc *p) 339 { 340 if (p->p_flag & P_SWAPPEDOUT) { 341 /* 342 * The process is waiting in the kernel to return to user 343 * mode but cannot until P_SWAPPEDOUT gets cleared. 344 */ 345 crit_enter(); 346 p->p_flag &= ~(P_SWAPPEDOUT | P_SWAPWAIT); 347 #ifdef INVARIANTS 348 if (swap_debug) 349 printf("swapping in %d (%s)\n", p->p_pid, p->p_comm); 350 #endif 351 wakeup(p); 352 353 crit_exit(); 354 } 355 } 356 357 /* 358 * Kernel initialization eventually falls through to this function, 359 * which is process 0. 360 * 361 * This swapin algorithm attempts to swap-in processes only if there 362 * is enough space for them. Of course, if a process waits for a long 363 * time, it will be swapped in anyway. 364 */ 365 366 struct scheduler_info { 367 struct proc *pp; 368 int ppri; 369 }; 370 371 static int scheduler_callback(struct proc *p, void *data); 372 373 static void 374 scheduler(void *dummy) 375 { 376 struct scheduler_info info; 377 struct proc *p; 378 379 KKASSERT(!IN_CRITICAL_SECT(curthread)); 380 loop: 381 scheduler_notify = 0; 382 /* 383 * Don't try to swap anything in if we are low on memory. 384 */ 385 if (vm_page_count_min()) { 386 vm_wait(); 387 goto loop; 388 } 389 390 /* 391 * Look for a good candidate to wake up 392 */ 393 info.pp = NULL; 394 info.ppri = INT_MIN; 395 allproc_scan(scheduler_callback, &info); 396 397 /* 398 * Nothing to do, back to sleep for at least 1/10 of a second. If 399 * we are woken up, immediately process the next request. If 400 * multiple requests have built up the first is processed 401 * immediately and the rest are staggered. 402 */ 403 if ((p = info.pp) == NULL) { 404 tsleep(&proc0, 0, "nowork", hz / 10); 405 if (scheduler_notify == 0) 406 tsleep(&scheduler_notify, 0, "nowork", 0); 407 goto loop; 408 } 409 410 /* 411 * Fault the selected process in, then wait for a short period of 412 * time and loop up. 413 * 414 * XXX we need a heuristic to get a measure of system stress and 415 * then adjust our stagger wakeup delay accordingly. 416 */ 417 faultin(p); 418 p->p_swtime = 0; 419 PRELE(p); 420 tsleep(&proc0, 0, "swapin", hz / 10); 421 goto loop; 422 } 423 424 static int 425 scheduler_callback(struct proc *p, void *data) 426 { 427 struct scheduler_info *info = data; 428 segsz_t pgs; 429 int pri; 430 431 if (p->p_flag & P_SWAPWAIT) { 432 pri = p->p_swtime + p->p_slptime - p->p_nice * 8; 433 434 /* 435 * The more pages paged out while we were swapped, 436 * the more work we have to do to get up and running 437 * again and the lower our wakeup priority. 438 * 439 * Each second of sleep time is worth ~1MB 440 */ 441 pgs = vmspace_resident_count(p->p_vmspace); 442 if (pgs < p->p_vmspace->vm_swrss) { 443 pri -= (p->p_vmspace->vm_swrss - pgs) / 444 (1024 * 1024 / PAGE_SIZE); 445 } 446 447 /* 448 * If this process is higher priority and there is 449 * enough space, then select this process instead of 450 * the previous selection. 451 */ 452 if (pri > info->ppri) { 453 if (info->pp) 454 PRELE(info->pp); 455 PHOLD(p); 456 info->pp = p; 457 info->ppri = pri; 458 } 459 } 460 return(0); 461 } 462 463 void 464 swapin_request(void) 465 { 466 if (scheduler_notify == 0) { 467 scheduler_notify = 1; 468 wakeup(&scheduler_notify); 469 } 470 } 471 472 #ifndef NO_SWAPPING 473 474 #define swappable(p) \ 475 (((p)->p_lock == 0) && \ 476 ((p)->p_flag & (P_TRACED|P_SYSTEM|P_SWAPPEDOUT|P_WEXIT)) == 0) 477 478 479 /* 480 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 481 */ 482 static int swap_idle_threshold1 = 15; 483 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, 484 CTLFLAG_RW, &swap_idle_threshold1, 0, ""); 485 486 /* 487 * Swap_idle_threshold2 is the time that a process can be idle before 488 * it will be swapped out, if idle swapping is enabled. Default is 489 * one minute. 490 */ 491 static int swap_idle_threshold2 = 60; 492 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, 493 CTLFLAG_RW, &swap_idle_threshold2, 0, ""); 494 495 /* 496 * Swapout is driven by the pageout daemon. Very simple, we find eligible 497 * procs and mark them as being swapped out. This will cause the kernel 498 * to prefer to pageout those proc's pages first and the procs in question 499 * will not return to user mode until the swapper tells them they can. 500 * 501 * If any procs have been sleeping/stopped for at least maxslp seconds, 502 * they are swapped. Else, we swap the longest-sleeping or stopped process, 503 * if any, otherwise the longest-resident process. 504 */ 505 506 static int swapout_procs_callback(struct proc *p, void *data); 507 508 void 509 swapout_procs(int action) 510 { 511 allproc_scan(swapout_procs_callback, &action); 512 } 513 514 static int 515 swapout_procs_callback(struct proc *p, void *data) 516 { 517 struct vmspace *vm; 518 int action = *(int *)data; 519 520 if (!swappable(p)) 521 return(0); 522 523 vm = p->p_vmspace; 524 525 if (p->p_stat == SSLEEP || p->p_stat == SRUN) { 526 /* 527 * do not swap out a realtime process 528 */ 529 if (RTP_PRIO_IS_REALTIME(p->p_lwp.lwp_rtprio.type)) 530 return(0); 531 532 /* 533 * Guarentee swap_idle_threshold time in memory 534 */ 535 if (p->p_slptime < swap_idle_threshold1) 536 return(0); 537 538 /* 539 * If the system is under memory stress, or if we 540 * are swapping idle processes >= swap_idle_threshold2, 541 * then swap the process out. 542 */ 543 if (((action & VM_SWAP_NORMAL) == 0) && 544 (((action & VM_SWAP_IDLE) == 0) || 545 (p->p_slptime < swap_idle_threshold2))) { 546 return(0); 547 } 548 549 ++vm->vm_refcnt; 550 551 /* 552 * If the process has been asleep for awhile, swap 553 * it out. 554 */ 555 if ((action & VM_SWAP_NORMAL) || 556 ((action & VM_SWAP_IDLE) && 557 (p->p_slptime > swap_idle_threshold2))) { 558 swapout(p); 559 } 560 561 /* 562 * cleanup our reference 563 */ 564 vmspace_free(vm); 565 } 566 return(0); 567 } 568 569 static void 570 swapout(struct proc *p) 571 { 572 #ifdef INVARIANTS 573 if (swap_debug) 574 printf("swapping out %d (%s)\n", p->p_pid, p->p_comm); 575 #endif 576 ++p->p_stats->p_ru.ru_nswap; 577 /* 578 * remember the process resident count 579 */ 580 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 581 p->p_flag |= P_SWAPPEDOUT; 582 p->p_swtime = 0; 583 } 584 585 #endif /* !NO_SWAPPING */ 586 587