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 1.48 2007/01/01 22:51:18 corecode 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_MAX_USER_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_MAX_USER_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 if (len) { 201 vm_map_wire(&curproc->p_vmspace->vm_map, 202 trunc_page((vm_offset_t)addr), 203 round_page((vm_offset_t)addr + len), 0); 204 } 205 } 206 207 void 208 vsunlock(caddr_t addr, u_int len) 209 { 210 if (len) { 211 vm_map_wire(&curproc->p_vmspace->vm_map, 212 trunc_page((vm_offset_t)addr), 213 round_page((vm_offset_t)addr + len), 214 KM_PAGEABLE); 215 } 216 } 217 218 /* 219 * Implement fork's actions on an address space. 220 * Here we arrange for the address space to be copied or referenced, 221 * allocate a user struct (pcb and kernel stack), then call the 222 * machine-dependent layer to fill those in and make the new process 223 * ready to run. The new process is set up so that it returns directly 224 * to user mode to avoid stack copying and relocation problems. 225 */ 226 void 227 vm_fork(struct lwp *lp1, struct proc *p2, int flags) 228 { 229 struct user *up; 230 struct proc *p1 = lp1->lwp_proc; 231 struct thread *td2; 232 233 if ((flags & RFPROC) == 0) { 234 /* 235 * Divorce the memory, if it is shared, essentially 236 * this changes shared memory amongst threads, into 237 * COW locally. 238 */ 239 if ((flags & RFMEM) == 0) { 240 if (p1->p_vmspace->vm_refcnt > 1) { 241 vmspace_unshare(p1); 242 } 243 } 244 cpu_fork(lp1, NULL, flags); 245 return; 246 } 247 248 if (flags & RFMEM) { 249 p2->p_vmspace = p1->p_vmspace; 250 p1->p_vmspace->vm_refcnt++; 251 } 252 253 while (vm_page_count_severe()) { 254 vm_wait(); 255 } 256 257 if ((flags & RFMEM) == 0) { 258 p2->p_vmspace = vmspace_fork(p1->p_vmspace); 259 260 pmap_pinit2(vmspace_pmap(p2->p_vmspace)); 261 262 if (p1->p_vmspace->vm_shm) 263 shmfork(p1, p2); 264 } 265 266 td2 = lwkt_alloc_thread(NULL, LWKT_THREAD_STACK, -1, 0); 267 pmap_init_proc(p2, td2); 268 lwkt_setpri(td2, TDPRI_KERN_USER); 269 lwkt_set_comm(td2, "%s", p1->p_comm); 270 271 up = p2->p_addr; 272 273 /* 274 * p_stats currently points at fields in the user struct 275 * but not at &u, instead at p_addr. Copy parts of 276 * p_stats; zero the rest of p_stats (statistics). 277 * 278 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need 279 * to share sigacts, so we use the up->u_sigacts. 280 */ 281 if (p2->p_sigacts == NULL) { 282 if (p2->p_procsig->ps_refcnt != 1) 283 kprintf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid); 284 p2->p_sigacts = &up->u_sigacts; 285 up->u_sigacts = *p1->p_sigacts; 286 } 287 288 /* 289 * cpu_fork will copy and update the pcb, set up the kernel stack, 290 * and make the child ready to run. 291 */ 292 cpu_fork(lp1, td2->td_lwp, flags); 293 } 294 295 /* 296 * Called after process has been wait(2)'ed apon and is being reaped. 297 * The idea is to reclaim resources that we could not reclaim while 298 * the process was still executing. 299 */ 300 void 301 vm_waitproc(struct proc *p) 302 { 303 cpu_proc_wait(p); 304 vmspace_exitfree(p); /* and clean-out the vmspace */ 305 } 306 307 /* 308 * Set default limits for VM system. 309 * Called for proc 0, and then inherited by all others. 310 * 311 * XXX should probably act directly on proc0. 312 */ 313 static void 314 vm_init_limits(void *udata) 315 { 316 struct proc *p = udata; 317 int rss_limit; 318 319 /* 320 * Set up the initial limits on process VM. Set the maximum resident 321 * set size to be half of (reasonably) available memory. Since this 322 * is a soft limit, it comes into effect only when the system is out 323 * of memory - half of main memory helps to favor smaller processes, 324 * and reduces thrashing of the object cache. 325 */ 326 p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz; 327 p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz; 328 p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz; 329 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz; 330 /* limit the limit to no less than 2MB */ 331 rss_limit = max(vmstats.v_free_count, 512); 332 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit); 333 p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY; 334 } 335 336 /* 337 * Faultin the specified process. Note that the process can be in any 338 * state. Just clear P_SWAPPEDOUT and call wakeup in case the process is 339 * sleeping. 340 */ 341 void 342 faultin(struct proc *p) 343 { 344 if (p->p_flag & P_SWAPPEDOUT) { 345 /* 346 * The process is waiting in the kernel to return to user 347 * mode but cannot until P_SWAPPEDOUT gets cleared. 348 */ 349 crit_enter(); 350 p->p_flag &= ~(P_SWAPPEDOUT | P_SWAPWAIT); 351 #ifdef INVARIANTS 352 if (swap_debug) 353 kprintf("swapping in %d (%s)\n", p->p_pid, p->p_comm); 354 #endif 355 wakeup(p); 356 357 crit_exit(); 358 } 359 } 360 361 /* 362 * Kernel initialization eventually falls through to this function, 363 * which is process 0. 364 * 365 * This swapin algorithm attempts to swap-in processes only if there 366 * is enough space for them. Of course, if a process waits for a long 367 * time, it will be swapped in anyway. 368 */ 369 370 struct scheduler_info { 371 struct proc *pp; 372 int ppri; 373 }; 374 375 static int scheduler_callback(struct proc *p, void *data); 376 377 static void 378 scheduler(void *dummy) 379 { 380 struct scheduler_info info; 381 struct proc *p; 382 383 KKASSERT(!IN_CRITICAL_SECT(curthread)); 384 loop: 385 scheduler_notify = 0; 386 /* 387 * Don't try to swap anything in if we are low on memory. 388 */ 389 if (vm_page_count_min()) { 390 vm_wait(); 391 goto loop; 392 } 393 394 /* 395 * Look for a good candidate to wake up 396 */ 397 info.pp = NULL; 398 info.ppri = INT_MIN; 399 allproc_scan(scheduler_callback, &info); 400 401 /* 402 * Nothing to do, back to sleep for at least 1/10 of a second. If 403 * we are woken up, immediately process the next request. If 404 * multiple requests have built up the first is processed 405 * immediately and the rest are staggered. 406 */ 407 if ((p = info.pp) == NULL) { 408 tsleep(&proc0, 0, "nowork", hz / 10); 409 if (scheduler_notify == 0) 410 tsleep(&scheduler_notify, 0, "nowork", 0); 411 goto loop; 412 } 413 414 /* 415 * Fault the selected process in, then wait for a short period of 416 * time and loop up. 417 * 418 * XXX we need a heuristic to get a measure of system stress and 419 * then adjust our stagger wakeup delay accordingly. 420 */ 421 faultin(p); 422 p->p_swtime = 0; 423 PRELE(p); 424 tsleep(&proc0, 0, "swapin", hz / 10); 425 goto loop; 426 } 427 428 static int 429 scheduler_callback(struct proc *p, void *data) 430 { 431 struct scheduler_info *info = data; 432 segsz_t pgs; 433 int pri; 434 435 if (p->p_flag & P_SWAPWAIT) { 436 pri = p->p_swtime + p->p_slptime - p->p_nice * 8; 437 438 /* 439 * The more pages paged out while we were swapped, 440 * the more work we have to do to get up and running 441 * again and the lower our wakeup priority. 442 * 443 * Each second of sleep time is worth ~1MB 444 */ 445 pgs = vmspace_resident_count(p->p_vmspace); 446 if (pgs < p->p_vmspace->vm_swrss) { 447 pri -= (p->p_vmspace->vm_swrss - pgs) / 448 (1024 * 1024 / PAGE_SIZE); 449 } 450 451 /* 452 * If this process is higher priority and there is 453 * enough space, then select this process instead of 454 * the previous selection. 455 */ 456 if (pri > info->ppri) { 457 if (info->pp) 458 PRELE(info->pp); 459 PHOLD(p); 460 info->pp = p; 461 info->ppri = pri; 462 } 463 } 464 return(0); 465 } 466 467 void 468 swapin_request(void) 469 { 470 if (scheduler_notify == 0) { 471 scheduler_notify = 1; 472 wakeup(&scheduler_notify); 473 } 474 } 475 476 #ifndef NO_SWAPPING 477 478 #define swappable(p) \ 479 (((p)->p_lock == 0) && \ 480 ((p)->p_flag & (P_TRACED|P_SYSTEM|P_SWAPPEDOUT|P_WEXIT)) == 0) 481 482 483 /* 484 * Swap_idle_threshold1 is the guaranteed swapped in time for a process 485 */ 486 static int swap_idle_threshold1 = 15; 487 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, 488 CTLFLAG_RW, &swap_idle_threshold1, 0, ""); 489 490 /* 491 * Swap_idle_threshold2 is the time that a process can be idle before 492 * it will be swapped out, if idle swapping is enabled. Default is 493 * one minute. 494 */ 495 static int swap_idle_threshold2 = 60; 496 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, 497 CTLFLAG_RW, &swap_idle_threshold2, 0, ""); 498 499 /* 500 * Swapout is driven by the pageout daemon. Very simple, we find eligible 501 * procs and mark them as being swapped out. This will cause the kernel 502 * to prefer to pageout those proc's pages first and the procs in question 503 * will not return to user mode until the swapper tells them they can. 504 * 505 * If any procs have been sleeping/stopped for at least maxslp seconds, 506 * they are swapped. Else, we swap the longest-sleeping or stopped process, 507 * if any, otherwise the longest-resident process. 508 */ 509 510 static int swapout_procs_callback(struct proc *p, void *data); 511 512 void 513 swapout_procs(int action) 514 { 515 allproc_scan(swapout_procs_callback, &action); 516 } 517 518 static int 519 swapout_procs_callback(struct proc *p, void *data) 520 { 521 struct vmspace *vm; 522 int action = *(int *)data; 523 524 if (!swappable(p)) 525 return(0); 526 527 vm = p->p_vmspace; 528 529 if (p->p_stat == SSLEEP || p->p_stat == SRUN) { 530 /* 531 * do not swap out a realtime process 532 */ 533 if (RTP_PRIO_IS_REALTIME(p->p_lwp.lwp_rtprio.type)) 534 return(0); 535 536 /* 537 * Guarentee swap_idle_threshold time in memory 538 */ 539 if (p->p_slptime < swap_idle_threshold1) 540 return(0); 541 542 /* 543 * If the system is under memory stress, or if we 544 * are swapping idle processes >= swap_idle_threshold2, 545 * then swap the process out. 546 */ 547 if (((action & VM_SWAP_NORMAL) == 0) && 548 (((action & VM_SWAP_IDLE) == 0) || 549 (p->p_slptime < swap_idle_threshold2))) { 550 return(0); 551 } 552 553 ++vm->vm_refcnt; 554 555 /* 556 * If the process has been asleep for awhile, swap 557 * it out. 558 */ 559 if ((action & VM_SWAP_NORMAL) || 560 ((action & VM_SWAP_IDLE) && 561 (p->p_slptime > swap_idle_threshold2))) { 562 swapout(p); 563 } 564 565 /* 566 * cleanup our reference 567 */ 568 vmspace_free(vm); 569 } 570 return(0); 571 } 572 573 static void 574 swapout(struct proc *p) 575 { 576 #ifdef INVARIANTS 577 if (swap_debug) 578 kprintf("swapping out %d (%s)\n", p->p_pid, p->p_comm); 579 #endif 580 ++p->p_ru.ru_nswap; 581 /* 582 * remember the process resident count 583 */ 584 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); 585 p->p_flag |= P_SWAPPEDOUT; 586 p->p_swtime = 0; 587 } 588 589 #endif /* !NO_SWAPPING */ 590 591