This is a major revamping of the pageout and low-memory handling code.
[dragonfly.git] / sys / vm / vm_glue.c
CommitLineData
984263bc
MD
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 $
4ecf7cc9 63 * $DragonFly: src/sys/vm/vm_glue.c,v 1.56 2008/07/01 02:02:56 dillon Exp $
984263bc
MD
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>
12e4aaff 93#include <vm/vm_page2.h>
cdd46d2e 94#include <sys/thread2.h>
e3161323 95#include <sys/sysref2.h>
984263bc
MD
96
97/*
984263bc
MD
98 * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
99 *
100 * Note: run scheduling should be divorced from the vm system.
101 */
1388df65 102static void scheduler (void *);
984263bc
MD
103SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL)
104
344ad853
MD
105#ifdef INVARIANTS
106
107static int swap_debug = 0;
108SYSCTL_INT(_vm, OID_AUTO, swap_debug,
109 CTLFLAG_RW, &swap_debug, 0, "");
110
111#endif
112
113static int scheduler_notify;
984263bc 114
1388df65 115static void swapout (struct proc *);
984263bc
MD
116
117int
6cebf0fc 118kernacc(c_caddr_t addr, int len, int rw)
984263bc
MD
119{
120 boolean_t rv;
121 vm_offset_t saddr, eaddr;
122 vm_prot_t prot;
123
124 KASSERT((rw & (~VM_PROT_ALL)) == 0,
125 ("illegal ``rw'' argument to kernacc (%x)\n", rw));
ed468fd2 126
03debc4a
MD
127 /*
128 * The globaldata space is not part of the kernel_map proper,
129 * check access separately.
130 */
131 if (is_globaldata_space((vm_offset_t)addr, (vm_offset_t)(addr + len)))
132 return (TRUE);
133
134 /*
135 * Nominal kernel memory access - check access via kernel_map.
136 */
e4846942 137 if ((vm_offset_t)addr + len > kernel_map.max_offset ||
ed468fd2
MD
138 (vm_offset_t)addr + len < (vm_offset_t)addr) {
139 return (FALSE);
140 }
984263bc
MD
141 prot = rw;
142 saddr = trunc_page((vm_offset_t)addr);
143 eaddr = round_page((vm_offset_t)addr + len);
e4846942
MD
144 vm_map_lock_read(&kernel_map);
145 rv = vm_map_check_protection(&kernel_map, saddr, eaddr, prot);
146 vm_map_unlock_read(&kernel_map);
984263bc
MD
147 return (rv == TRUE);
148}
149
150int
6cebf0fc 151useracc(c_caddr_t addr, int len, int rw)
984263bc
MD
152{
153 boolean_t rv;
154 vm_prot_t prot;
155 vm_map_t map;
156 vm_map_entry_t save_hint;
157
158 KASSERT((rw & (~VM_PROT_ALL)) == 0,
159 ("illegal ``rw'' argument to useracc (%x)\n", rw));
160 prot = rw;
161 /*
162 * XXX - check separately to disallow access to user area and user
163 * page tables - they are in the map.
164 *
88181b08 165 * XXX - VM_MAX_USER_ADDRESS is an end address, not a max. It was once
984263bc
MD
166 * only used (as an end address) in trap.c. Use it as an end address
167 * here too. This bogusness has spread. I just fixed where it was
168 * used as a max in vm_mmap.c.
169 */
88181b08 170 if ((vm_offset_t) addr + len > /* XXX */ VM_MAX_USER_ADDRESS
984263bc
MD
171 || (vm_offset_t) addr + len < (vm_offset_t) addr) {
172 return (FALSE);
173 }
174 map = &curproc->p_vmspace->vm_map;
175 vm_map_lock_read(map);
176 /*
177 * We save the map hint, and restore it. Useracc appears to distort
178 * the map hint unnecessarily.
179 */
180 save_hint = map->hint;
181 rv = vm_map_check_protection(map,
182 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), prot);
183 map->hint = save_hint;
184 vm_map_unlock_read(map);
185
186 return (rv == TRUE);
187}
188
189void
57e43348 190vslock(caddr_t addr, u_int len)
984263bc 191{
fc00aa2b
MD
192 if (len) {
193 vm_map_wire(&curproc->p_vmspace->vm_map,
194 trunc_page((vm_offset_t)addr),
195 round_page((vm_offset_t)addr + len), 0);
196 }
984263bc
MD
197}
198
199void
57e43348 200vsunlock(caddr_t addr, u_int len)
984263bc 201{
fc00aa2b
MD
202 if (len) {
203 vm_map_wire(&curproc->p_vmspace->vm_map,
204 trunc_page((vm_offset_t)addr),
205 round_page((vm_offset_t)addr + len),
206 KM_PAGEABLE);
207 }
984263bc
MD
208}
209
210/*
211 * Implement fork's actions on an address space.
212 * Here we arrange for the address space to be copied or referenced,
213 * allocate a user struct (pcb and kernel stack), then call the
214 * machine-dependent layer to fill those in and make the new process
215 * ready to run. The new process is set up so that it returns directly
216 * to user mode to avoid stack copying and relocation problems.
217 */
218void
13d13d89 219vm_fork(struct proc *p1, struct proc *p2, int flags)
984263bc 220{
984263bc
MD
221 if ((flags & RFPROC) == 0) {
222 /*
223 * Divorce the memory, if it is shared, essentially
224 * this changes shared memory amongst threads, into
225 * COW locally.
226 */
227 if ((flags & RFMEM) == 0) {
e3161323 228 if (p1->p_vmspace->vm_sysref.refcnt > 1) {
984263bc
MD
229 vmspace_unshare(p1);
230 }
231 }
13d13d89 232 cpu_fork(ONLY_LWP_IN_PROC(p1), NULL, flags);
984263bc
MD
233 return;
234 }
235
236 if (flags & RFMEM) {
237 p2->p_vmspace = p1->p_vmspace;
e3161323 238 sysref_get(&p1->p_vmspace->vm_sysref);
984263bc
MD
239 }
240
241 while (vm_page_count_severe()) {
4ecf7cc9 242 vm_wait(0);
984263bc
MD
243 }
244
245 if ((flags & RFMEM) == 0) {
246 p2->p_vmspace = vmspace_fork(p1->p_vmspace);
247
248 pmap_pinit2(vmspace_pmap(p2->p_vmspace));
249
250 if (p1->p_vmspace->vm_shm)
251 shmfork(p1, p2);
252 }
253
13d13d89 254 pmap_init_proc(p2);
984263bc
MD
255}
256
257/*
258 * Called after process has been wait(2)'ed apon and is being reaped.
259 * The idea is to reclaim resources that we could not reclaim while
260 * the process was still executing.
261 */
262void
263vm_waitproc(struct proc *p)
264{
99df837e 265 cpu_proc_wait(p);
984263bc
MD
266 vmspace_exitfree(p); /* and clean-out the vmspace */
267}
268
269/*
ba39e2e0 270 * Set default limits for VM system. Call during proc0's initialization.
984263bc 271 */
ba39e2e0
MD
272void
273vm_init_limits(struct proc *p)
984263bc 274{
984263bc
MD
275 int rss_limit;
276
277 /*
278 * Set up the initial limits on process VM. Set the maximum resident
279 * set size to be half of (reasonably) available memory. Since this
280 * is a soft limit, it comes into effect only when the system is out
281 * of memory - half of main memory helps to favor smaller processes,
282 * and reduces thrashing of the object cache.
283 */
284 p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz;
285 p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz;
286 p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz;
287 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz;
288 /* limit the limit to no less than 2MB */
12e4aaff 289 rss_limit = max(vmstats.v_free_count, 512);
984263bc
MD
290 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
291 p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
292}
293
344ad853
MD
294/*
295 * Faultin the specified process. Note that the process can be in any
296 * state. Just clear P_SWAPPEDOUT and call wakeup in case the process is
297 * sleeping.
298 */
984263bc 299void
57e43348 300faultin(struct proc *p)
984263bc 301{
344ad853 302 if (p->p_flag & P_SWAPPEDOUT) {
8ec60c3f 303 /*
344ad853
MD
304 * The process is waiting in the kernel to return to user
305 * mode but cannot until P_SWAPPEDOUT gets cleared.
8ec60c3f 306 */
344ad853
MD
307 crit_enter();
308 p->p_flag &= ~(P_SWAPPEDOUT | P_SWAPWAIT);
309#ifdef INVARIANTS
310 if (swap_debug)
086c1d7e 311 kprintf("swapping in %d (%s)\n", p->p_pid, p->p_comm);
344ad853
MD
312#endif
313 wakeup(p);
984263bc 314
cdd46d2e 315 crit_exit();
984263bc
MD
316 }
317}
318
319/*
f8c3996b
MD
320 * Kernel initialization eventually falls through to this function,
321 * which is process 0.
322 *
984263bc
MD
323 * This swapin algorithm attempts to swap-in processes only if there
324 * is enough space for them. Of course, if a process waits for a long
325 * time, it will be swapped in anyway.
326 */
344ad853 327
8fa76237
MD
328struct scheduler_info {
329 struct proc *pp;
330 int ppri;
331};
332
333static int scheduler_callback(struct proc *p, void *data);
334
984263bc 335static void
57e43348 336scheduler(void *dummy)
984263bc 337{
8fa76237 338 struct scheduler_info info;
5f910b2f 339 struct proc *p;
984263bc 340
f8c3996b 341 KKASSERT(!IN_CRITICAL_SECT(curthread));
984263bc 342loop:
344ad853
MD
343 scheduler_notify = 0;
344 /*
345 * Don't try to swap anything in if we are low on memory.
346 */
20479584 347 if (vm_page_count_severe()) {
4ecf7cc9 348 vm_wait(0);
984263bc
MD
349 goto loop;
350 }
351
344ad853
MD
352 /*
353 * Look for a good candidate to wake up
354 */
8fa76237
MD
355 info.pp = NULL;
356 info.ppri = INT_MIN;
357 allproc_scan(scheduler_callback, &info);
984263bc
MD
358
359 /*
344ad853
MD
360 * Nothing to do, back to sleep for at least 1/10 of a second. If
361 * we are woken up, immediately process the next request. If
362 * multiple requests have built up the first is processed
363 * immediately and the rest are staggered.
984263bc 364 */
8fa76237 365 if ((p = info.pp) == NULL) {
344ad853
MD
366 tsleep(&proc0, 0, "nowork", hz / 10);
367 if (scheduler_notify == 0)
368 tsleep(&scheduler_notify, 0, "nowork", 0);
984263bc
MD
369 goto loop;
370 }
984263bc
MD
371
372 /*
344ad853
MD
373 * Fault the selected process in, then wait for a short period of
374 * time and loop up.
375 *
376 * XXX we need a heuristic to get a measure of system stress and
377 * then adjust our stagger wakeup delay accordingly.
984263bc
MD
378 */
379 faultin(p);
380 p->p_swtime = 0;
8fa76237 381 PRELE(p);
344ad853 382 tsleep(&proc0, 0, "swapin", hz / 10);
984263bc
MD
383 goto loop;
384}
385
8fa76237
MD
386static int
387scheduler_callback(struct proc *p, void *data)
388{
389 struct scheduler_info *info = data;
08f2f1bb 390 struct lwp *lp;
8fa76237
MD
391 segsz_t pgs;
392 int pri;
393
394 if (p->p_flag & P_SWAPWAIT) {
c7e98b2f
SS
395 pri = 0;
396 FOREACH_LWP_IN_PROC(lp, p) {
397 /* XXX lwp might need a different metric */
398 pri += lp->lwp_slptime;
399 }
400 pri += p->p_swtime - p->p_nice * 8;
8fa76237
MD
401
402 /*
403 * The more pages paged out while we were swapped,
404 * the more work we have to do to get up and running
405 * again and the lower our wakeup priority.
406 *
407 * Each second of sleep time is worth ~1MB
408 */
409 pgs = vmspace_resident_count(p->p_vmspace);
410 if (pgs < p->p_vmspace->vm_swrss) {
411 pri -= (p->p_vmspace->vm_swrss - pgs) /
412 (1024 * 1024 / PAGE_SIZE);
413 }
414
415 /*
416 * If this process is higher priority and there is
417 * enough space, then select this process instead of
418 * the previous selection.
419 */
420 if (pri > info->ppri) {
421 if (info->pp)
422 PRELE(info->pp);
423 PHOLD(p);
424 info->pp = p;
425 info->ppri = pri;
426 }
427 }
428 return(0);
429}
430
344ad853
MD
431void
432swapin_request(void)
433{
434 if (scheduler_notify == 0) {
435 scheduler_notify = 1;
436 wakeup(&scheduler_notify);
437 }
438}
439
984263bc
MD
440#ifndef NO_SWAPPING
441
442#define swappable(p) \
443 (((p)->p_lock == 0) && \
344ad853 444 ((p)->p_flag & (P_TRACED|P_SYSTEM|P_SWAPPEDOUT|P_WEXIT)) == 0)
984263bc
MD
445
446
447/*
448 * Swap_idle_threshold1 is the guaranteed swapped in time for a process
449 */
46311ac2 450static int swap_idle_threshold1 = 15;
984263bc
MD
451SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1,
452 CTLFLAG_RW, &swap_idle_threshold1, 0, "");
453
454/*
455 * Swap_idle_threshold2 is the time that a process can be idle before
344ad853
MD
456 * it will be swapped out, if idle swapping is enabled. Default is
457 * one minute.
984263bc 458 */
344ad853 459static int swap_idle_threshold2 = 60;
984263bc
MD
460SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2,
461 CTLFLAG_RW, &swap_idle_threshold2, 0, "");
462
463/*
464 * Swapout is driven by the pageout daemon. Very simple, we find eligible
344ad853
MD
465 * procs and mark them as being swapped out. This will cause the kernel
466 * to prefer to pageout those proc's pages first and the procs in question
467 * will not return to user mode until the swapper tells them they can.
468 *
984263bc
MD
469 * If any procs have been sleeping/stopped for at least maxslp seconds,
470 * they are swapped. Else, we swap the longest-sleeping or stopped process,
471 * if any, otherwise the longest-resident process.
472 */
8fa76237
MD
473
474static int swapout_procs_callback(struct proc *p, void *data);
475
984263bc 476void
57e43348 477swapout_procs(int action)
984263bc 478{
8fa76237
MD
479 allproc_scan(swapout_procs_callback, &action);
480}
984263bc 481
8fa76237
MD
482static int
483swapout_procs_callback(struct proc *p, void *data)
484{
485 struct vmspace *vm;
08f2f1bb 486 struct lwp *lp;
8fa76237 487 int action = *(int *)data;
c7e98b2f 488 int minslp = -1;
8fa76237
MD
489
490 if (!swappable(p))
491 return(0);
492
493 vm = p->p_vmspace;
494
c7e98b2f
SS
495 /*
496 * We only consider active processes.
497 */
498 if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
499 return(0);
164b8401 500
c7e98b2f 501 FOREACH_LWP_IN_PROC(lp, p) {
8fa76237
MD
502 /*
503 * do not swap out a realtime process
504 */
08f2f1bb 505 if (RTP_PRIO_IS_REALTIME(lp->lwp_rtprio.type))
8fa76237 506 return(0);
984263bc 507
8fa76237
MD
508 /*
509 * Guarentee swap_idle_threshold time in memory
510 */
08f2f1bb 511 if (lp->lwp_slptime < swap_idle_threshold1)
8fa76237
MD
512 return(0);
513
514 /*
515 * If the system is under memory stress, or if we
516 * are swapping idle processes >= swap_idle_threshold2,
517 * then swap the process out.
518 */
519 if (((action & VM_SWAP_NORMAL) == 0) &&
520 (((action & VM_SWAP_IDLE) == 0) ||
08f2f1bb 521 (lp->lwp_slptime < swap_idle_threshold2))) {
8fa76237 522 return(0);
984263bc 523 }
8fa76237 524
c7e98b2f
SS
525 if (minslp == -1 || lp->lwp_slptime < minslp)
526 minslp = lp->lwp_slptime;
527 }
8fa76237 528
e3161323 529 sysref_get(&vm->vm_sysref);
8fa76237 530
c7e98b2f
SS
531 /*
532 * If the process has been asleep for awhile, swap
533 * it out.
534 */
535 if ((action & VM_SWAP_NORMAL) ||
536 ((action & VM_SWAP_IDLE) &&
537 (minslp > swap_idle_threshold2))) {
538 swapout(p);
984263bc 539 }
c7e98b2f
SS
540
541 /*
542 * cleanup our reference
543 */
e3161323 544 sysref_put(&vm->vm_sysref);
c7e98b2f 545
8fa76237 546 return(0);
984263bc
MD
547}
548
549static void
57e43348 550swapout(struct proc *p)
984263bc 551{
344ad853
MD
552#ifdef INVARIANTS
553 if (swap_debug)
086c1d7e 554 kprintf("swapping out %d (%s)\n", p->p_pid, p->p_comm);
984263bc 555#endif
fde7ac71 556 ++p->p_ru.ru_nswap;
984263bc
MD
557 /*
558 * remember the process resident count
559 */
560 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
344ad853 561 p->p_flag |= P_SWAPPEDOUT;
984263bc
MD
562 p->p_swtime = 0;
563}
344ad853 564
984263bc 565#endif /* !NO_SWAPPING */
344ad853 566