2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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
36 * from: @(#)vm_glue.c 8.6 (Berkeley) 1/5/94
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
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.
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.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
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 1.38 2005/11/22 08:41:06 dillon Exp $
68 #include <sys/param.h>
69 #include <sys/systm.h>
71 #include <sys/resourcevar.h>
74 #include <sys/vmmeter.h>
75 #include <sys/sysctl.h>
77 #include <sys/kernel.h>
78 #include <sys/unistd.h>
80 #include <machine/limits.h>
83 #include <vm/vm_param.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>
93 #include <vm/vm_page2.h>
94 #include <sys/thread2.h>
97 * System initialization
99 * Note: proc0 from proc.h
102 static void vm_init_limits (void *);
103 SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0)
106 * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
108 * Note: run scheduling should be divorced from the vm system.
110 static void scheduler (void *);
111 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL)
115 static int swap_debug = 0;
116 SYSCTL_INT(_vm, OID_AUTO, swap_debug,
117 CTLFLAG_RW, &swap_debug, 0, "");
121 static int scheduler_notify;
123 static void swapout (struct proc *);
126 kernacc(c_caddr_t addr, int len, int rw)
129 vm_offset_t saddr, eaddr;
132 KASSERT((rw & (~VM_PROT_ALL)) == 0,
133 ("illegal ``rw'' argument to kernacc (%x)\n", rw));
135 saddr = trunc_page((vm_offset_t)addr);
136 eaddr = round_page((vm_offset_t)addr + len);
137 vm_map_lock_read(kernel_map);
138 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
139 vm_map_unlock_read(kernel_map);
140 if (rv == FALSE && is_globaldata_space(saddr, eaddr))
146 useracc(c_caddr_t addr, int len, int rw)
151 vm_map_entry_t save_hint;
153 KASSERT((rw & (~VM_PROT_ALL)) == 0,
154 ("illegal ``rw'' argument to useracc (%x)\n", rw));
157 * XXX - check separately to disallow access to user area and user
158 * page tables - they are in the map.
160 * XXX - VM_MAXUSER_ADDRESS is an end address, not a max. It was once
161 * only used (as an end address) in trap.c. Use it as an end address
162 * here too. This bogusness has spread. I just fixed where it was
163 * used as a max in vm_mmap.c.
165 if ((vm_offset_t) addr + len > /* XXX */ VM_MAXUSER_ADDRESS
166 || (vm_offset_t) addr + len < (vm_offset_t) addr) {
169 map = &curproc->p_vmspace->vm_map;
170 vm_map_lock_read(map);
172 * We save the map hint, and restore it. Useracc appears to distort
173 * the map hint unnecessarily.
175 save_hint = map->hint;
176 rv = vm_map_check_protection(map,
177 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), prot);
178 map->hint = save_hint;
179 vm_map_unlock_read(map);
185 vslock(caddr_t addr, u_int len)
187 vm_map_wire(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr),
188 round_page((vm_offset_t)addr + len), 0);
192 vsunlock(caddr_t addr, u_int len)
194 vm_map_wire(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr),
195 round_page((vm_offset_t)addr + len), KM_PAGEABLE);
199 * Implement fork's actions on an address space.
200 * Here we arrange for the address space to be copied or referenced,
201 * allocate a user struct (pcb and kernel stack), then call the
202 * machine-dependent layer to fill those in and make the new process
203 * ready to run. The new process is set up so that it returns directly
204 * to user mode to avoid stack copying and relocation problems.
207 vm_fork(struct proc *p1, struct proc *p2, int flags)
212 if ((flags & RFPROC) == 0) {
214 * Divorce the memory, if it is shared, essentially
215 * this changes shared memory amongst threads, into
218 if ((flags & RFMEM) == 0) {
219 if (p1->p_vmspace->vm_refcnt > 1) {
223 cpu_fork(p1, p2, flags);
228 p2->p_vmspace = p1->p_vmspace;
229 p1->p_vmspace->vm_refcnt++;
232 while (vm_page_count_severe()) {
236 if ((flags & RFMEM) == 0) {
237 p2->p_vmspace = vmspace_fork(p1->p_vmspace);
239 pmap_pinit2(vmspace_pmap(p2->p_vmspace));
241 if (p1->p_vmspace->vm_shm)
245 td2 = lwkt_alloc_thread(NULL, LWKT_THREAD_STACK, -1, 0);
246 pmap_init_proc(p2, td2);
247 lwkt_setpri(td2, TDPRI_KERN_USER);
248 lwkt_set_comm(td2, "%s", p1->p_comm);
253 * p_stats currently points at fields in the user struct
254 * but not at &u, instead at p_addr. Copy parts of
255 * p_stats; zero the rest of p_stats (statistics).
257 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need
258 * to share sigacts, so we use the up->u_sigacts.
260 p2->p_stats = &up->u_stats;
261 if (p2->p_sigacts == NULL) {
262 if (p2->p_procsig->ps_refcnt != 1)
263 printf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid);
264 p2->p_sigacts = &up->u_sigacts;
265 up->u_sigacts = *p1->p_sigacts;
268 bzero(&up->u_stats, sizeof(struct pstats));
271 * cpu_fork will copy and update the pcb, set up the kernel stack,
272 * and make the child ready to run.
274 cpu_fork(p1, p2, flags);
278 * Called after process has been wait(2)'ed apon and is being reaped.
279 * The idea is to reclaim resources that we could not reclaim while
280 * the process was still executing.
283 vm_waitproc(struct proc *p)
287 vmspace_exitfree(p); /* and clean-out the vmspace */
291 * Set default limits for VM system.
292 * Called for proc 0, and then inherited by all others.
294 * XXX should probably act directly on proc0.
297 vm_init_limits(void *udata)
299 struct proc *p = udata;
303 * Set up the initial limits on process VM. Set the maximum resident
304 * set size to be half of (reasonably) available memory. Since this
305 * is a soft limit, it comes into effect only when the system is out
306 * of memory - half of main memory helps to favor smaller processes,
307 * and reduces thrashing of the object cache.
309 p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz;
310 p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz;
311 p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz;
312 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz;
313 /* limit the limit to no less than 2MB */
314 rss_limit = max(vmstats.v_free_count, 512);
315 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
316 p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
320 * Faultin the specified process. Note that the process can be in any
321 * state. Just clear P_SWAPPEDOUT and call wakeup in case the process is
325 faultin(struct proc *p)
327 if (p->p_flag & P_SWAPPEDOUT) {
329 * The process is waiting in the kernel to return to user
330 * mode but cannot until P_SWAPPEDOUT gets cleared.
333 p->p_flag &= ~(P_SWAPPEDOUT | P_SWAPWAIT);
336 printf("swapping in %d (%s)\n", p->p_pid, p->p_comm);
345 * Kernel initialization eventually falls through to this function,
346 * which is process 0.
348 * This swapin algorithm attempts to swap-in processes only if there
349 * is enough space for them. Of course, if a process waits for a long
350 * time, it will be swapped in anyway.
355 scheduler(void *dummy)
363 KKASSERT(!IN_CRITICAL_SECT(curthread));
365 scheduler_notify = 0;
367 * Don't try to swap anything in if we are low on memory.
369 if (vm_page_count_min()) {
375 * Look for a good candidate to wake up
379 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
380 if (p->p_flag & P_SWAPWAIT) {
381 pri = p->p_swtime + p->p_slptime - p->p_nice * 8;
384 * The more pages paged out while we were swapped,
385 * the more work we have to do to get up and running
386 * again and the lower our wakeup priority.
388 * Each second of sleep time is worth ~1MB
390 pgs = vmspace_resident_count(p->p_vmspace);
391 if (pgs < p->p_vmspace->vm_swrss) {
392 pri -= (p->p_vmspace->vm_swrss - pgs) /
393 (1024 * 1024 / PAGE_SIZE);
397 * if this process is higher priority and there is
398 * enough space, then select this process instead of
399 * the previous selection.
409 * Nothing to do, back to sleep for at least 1/10 of a second. If
410 * we are woken up, immediately process the next request. If
411 * multiple requests have built up the first is processed
412 * immediately and the rest are staggered.
414 if ((p = pp) == NULL) {
415 tsleep(&proc0, 0, "nowork", hz / 10);
416 if (scheduler_notify == 0)
417 tsleep(&scheduler_notify, 0, "nowork", 0);
422 * Fault the selected process in, then wait for a short period of
425 * XXX we need a heuristic to get a measure of system stress and
426 * then adjust our stagger wakeup delay accordingly.
430 tsleep(&proc0, 0, "swapin", hz / 10);
437 if (scheduler_notify == 0) {
438 scheduler_notify = 1;
439 wakeup(&scheduler_notify);
445 #define swappable(p) \
446 (((p)->p_lock == 0) && \
447 ((p)->p_flag & (P_TRACED|P_SYSTEM|P_SWAPPEDOUT|P_WEXIT)) == 0)
451 * Swap_idle_threshold1 is the guaranteed swapped in time for a process
453 static int swap_idle_threshold1 = 2;
454 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1,
455 CTLFLAG_RW, &swap_idle_threshold1, 0, "");
458 * Swap_idle_threshold2 is the time that a process can be idle before
459 * it will be swapped out, if idle swapping is enabled. Default is
462 static int swap_idle_threshold2 = 60;
463 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2,
464 CTLFLAG_RW, &swap_idle_threshold2, 0, "");
467 * Swapout is driven by the pageout daemon. Very simple, we find eligible
468 * procs and mark them as being swapped out. This will cause the kernel
469 * to prefer to pageout those proc's pages first and the procs in question
470 * will not return to user mode until the swapper tells them they can.
472 * If any procs have been sleeping/stopped for at least maxslp seconds,
473 * they are swapped. Else, we swap the longest-sleeping or stopped process,
474 * if any, otherwise the longest-resident process.
477 swapout_procs(int action)
480 struct proc *outp, *outp2;
484 outpri = outpri2 = INT_MIN;
486 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
493 if (p->p_stat == SSLEEP || p->p_stat == SRUN) {
495 * do not swap out a realtime process
497 if (RTP_PRIO_IS_REALTIME(p->p_lwp.lwp_rtprio.type))
501 * Guarentee swap_idle_threshold time in memory
503 if (p->p_slptime < swap_idle_threshold1)
507 * If the system is under memory stress, or if we
508 * are swapping idle processes >= swap_idle_threshold2,
509 * then swap the process out.
511 if (((action & VM_SWAP_NORMAL) == 0) &&
512 (((action & VM_SWAP_IDLE) == 0) ||
513 (p->p_slptime < swap_idle_threshold2))) {
520 * If the process has been asleep for awhile, swap
523 if ((action & VM_SWAP_NORMAL) ||
524 ((action & VM_SWAP_IDLE) &&
525 (p->p_slptime > swap_idle_threshold2))) {
532 * cleanup our reference
540 swapout(struct proc *p)
544 printf("swapping out %d (%s)\n", p->p_pid, p->p_comm);
546 ++p->p_stats->p_ru.ru_nswap;
548 * remember the process resident count
550 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
551 p->p_flag |= P_SWAPPEDOUT;
555 #endif /* !NO_SWAPPING */