uvm_glue.c revision 1.36 1 /* $NetBSD: uvm_glue.c,v 1.36 2000/06/18 05:20:27 simonb Exp $ */
2
3 /*
4 * Copyright (c) 1997 Charles D. Cranor and Washington University.
5 * Copyright (c) 1991, 1993, The Regents of the University of California.
6 *
7 * All rights reserved.
8 *
9 * This code is derived from software contributed to Berkeley by
10 * The Mach Operating System project at Carnegie-Mellon University.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by Charles D. Cranor,
23 * Washington University, the University of California, Berkeley and
24 * its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * SUCH DAMAGE.
40 *
41 * @(#)vm_glue.c 8.6 (Berkeley) 1/5/94
42 * from: Id: uvm_glue.c,v 1.1.2.8 1998/02/07 01:16:54 chs Exp
43 *
44 *
45 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46 * All rights reserved.
47 *
48 * Permission to use, copy, modify and distribute this software and
49 * its documentation is hereby granted, provided that both the copyright
50 * notice and this permission notice appear in all copies of the
51 * software, derivative works or modified versions, and any portions
52 * thereof, and that both notices appear in supporting documentation.
53 *
54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57 *
58 * Carnegie Mellon requests users of this software to return to
59 *
60 * Software Distribution Coordinator or Software.Distribution (at) CS.CMU.EDU
61 * School of Computer Science
62 * Carnegie Mellon University
63 * Pittsburgh PA 15213-3890
64 *
65 * any improvements or extensions that they make and grant Carnegie the
66 * rights to redistribute these changes.
67 */
68
69 #include "opt_uvmhist.h"
70 #include "opt_sysv.h"
71
72 /*
73 * uvm_glue.c: glue functions
74 */
75
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/proc.h>
79 #include <sys/resourcevar.h>
80 #include <sys/buf.h>
81 #include <sys/user.h>
82 #ifdef SYSVSHM
83 #include <sys/shm.h>
84 #endif
85
86 #include <vm/vm.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_kern.h>
89
90 #include <uvm/uvm.h>
91
92 #include <machine/cpu.h>
93
94 /*
95 * local prototypes
96 */
97
98 static void uvm_swapout __P((struct proc *));
99
100 /*
101 * XXXCDC: do these really belong here?
102 */
103
104 unsigned maxdmap = MAXDSIZ; /* kern_resource.c: RLIMIT_DATA max */
105 unsigned maxsmap = MAXSSIZ; /* kern_resource.c: RLIMIT_STACK max */
106
107 int readbuffers = 0; /* allow KGDB to read kern buffer pool */
108 /* XXX: see uvm_kernacc */
109
110
111 /*
112 * uvm_kernacc: can the kernel access a region of memory
113 *
114 * - called from malloc [DIAGNOSTIC], and /dev/kmem driver (mem.c)
115 */
116
117 boolean_t
118 uvm_kernacc(addr, len, rw)
119 caddr_t addr;
120 size_t len;
121 int rw;
122 {
123 boolean_t rv;
124 vaddr_t saddr, eaddr;
125 vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
126
127 saddr = trunc_page((vaddr_t)addr);
128 eaddr = round_page((vaddr_t)addr+len);
129 vm_map_lock_read(kernel_map);
130 rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot);
131 vm_map_unlock_read(kernel_map);
132
133 /*
134 * XXX there are still some things (e.g. the buffer cache) that
135 * are managed behind the VM system's back so even though an
136 * address is accessible in the mind of the VM system, there may
137 * not be physical pages where the VM thinks there is. This can
138 * lead to bogus allocation of pages in the kernel address space
139 * or worse, inconsistencies at the pmap level. We only worry
140 * about the buffer cache for now.
141 */
142 if (!readbuffers && rv && (eaddr > (vaddr_t)buffers &&
143 saddr < (vaddr_t)buffers + MAXBSIZE * nbuf))
144 rv = FALSE;
145 return(rv);
146 }
147
148 /*
149 * uvm_useracc: can the user access it?
150 *
151 * - called from physio() and sys___sysctl().
152 */
153
154 boolean_t
155 uvm_useracc(addr, len, rw)
156 caddr_t addr;
157 size_t len;
158 int rw;
159 {
160 vm_map_t map;
161 boolean_t rv;
162 vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
163
164 /* XXX curproc */
165 map = &curproc->p_vmspace->vm_map;
166
167 vm_map_lock_read(map);
168 rv = uvm_map_checkprot(map, trunc_page((vaddr_t)addr),
169 round_page((vaddr_t)addr+len), prot);
170 vm_map_unlock_read(map);
171
172 return(rv);
173 }
174
175 #ifdef KGDB
176 /*
177 * Change protections on kernel pages from addr to addr+len
178 * (presumably so debugger can plant a breakpoint).
179 *
180 * We force the protection change at the pmap level. If we were
181 * to use vm_map_protect a change to allow writing would be lazily-
182 * applied meaning we would still take a protection fault, something
183 * we really don't want to do. It would also fragment the kernel
184 * map unnecessarily. We cannot use pmap_protect since it also won't
185 * enforce a write-enable request. Using pmap_enter is the only way
186 * we can ensure the change takes place properly.
187 */
188 void
189 uvm_chgkprot(addr, len, rw)
190 caddr_t addr;
191 size_t len;
192 int rw;
193 {
194 vm_prot_t prot;
195 paddr_t pa;
196 vaddr_t sva, eva;
197
198 prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE;
199 eva = round_page((vaddr_t)addr + len);
200 for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) {
201 /*
202 * Extract physical address for the page.
203 * We use a cheezy hack to differentiate physical
204 * page 0 from an invalid mapping, not that it
205 * really matters...
206 */
207 if (pmap_extract(pmap_kernel(), sva, &pa) == FALSE)
208 panic("chgkprot: invalid page");
209 pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED);
210 }
211 }
212 #endif
213
214 /*
215 * vslock: wire user memory for I/O
216 *
217 * - called from physio and sys___sysctl
218 * - XXXCDC: consider nuking this (or making it a macro?)
219 */
220
221 int
222 uvm_vslock(p, addr, len, access_type)
223 struct proc *p;
224 caddr_t addr;
225 size_t len;
226 vm_prot_t access_type;
227 {
228 vm_map_t map;
229 vaddr_t start, end;
230 int rv;
231
232 map = &p->p_vmspace->vm_map;
233 start = trunc_page((vaddr_t)addr);
234 end = round_page((vaddr_t)addr + len);
235
236 rv = uvm_fault_wire(map, start, end, access_type);
237
238 return (rv);
239 }
240
241 /*
242 * vslock: wire user memory for I/O
243 *
244 * - called from physio and sys___sysctl
245 * - XXXCDC: consider nuking this (or making it a macro?)
246 */
247
248 void
249 uvm_vsunlock(p, addr, len)
250 struct proc *p;
251 caddr_t addr;
252 size_t len;
253 {
254 uvm_fault_unwire(&p->p_vmspace->vm_map, trunc_page((vaddr_t)addr),
255 round_page((vaddr_t)addr+len));
256 }
257
258 /*
259 * uvm_fork: fork a virtual address space
260 *
261 * - the address space is copied as per parent map's inherit values
262 * - a new "user" structure is allocated for the child process
263 * [filled in by MD layer...]
264 * - if specified, the child gets a new user stack described by
265 * stack and stacksize
266 * - NOTE: the kernel stack may be at a different location in the child
267 * process, and thus addresses of automatic variables may be invalid
268 * after cpu_fork returns in the child process. We do nothing here
269 * after cpu_fork returns.
270 * - XXXCDC: we need a way for this to return a failure value rather
271 * than just hang
272 */
273 void
274 uvm_fork(p1, p2, shared, stack, stacksize, func, arg)
275 struct proc *p1, *p2;
276 boolean_t shared;
277 void *stack;
278 size_t stacksize;
279 void (*func) __P((void *));
280 void *arg;
281 {
282 struct user *up = p2->p_addr;
283 int rv;
284
285 if (shared == TRUE)
286 uvmspace_share(p1, p2); /* share vmspace */
287 else
288 p2->p_vmspace = uvmspace_fork(p1->p_vmspace); /* fork vmspace */
289
290 /*
291 * Wire down the U-area for the process, which contains the PCB
292 * and the kernel stack. Wired state is stored in p->p_flag's
293 * P_INMEM bit rather than in the vm_map_entry's wired count
294 * to prevent kernel_map fragmentation.
295 *
296 * Note the kernel stack gets read/write accesses right off
297 * the bat.
298 */
299 rv = uvm_fault_wire(kernel_map, (vaddr_t)up,
300 (vaddr_t)up + USPACE, VM_PROT_READ | VM_PROT_WRITE);
301 if (rv != KERN_SUCCESS)
302 panic("uvm_fork: uvm_fault_wire failed: %d", rv);
303
304 /*
305 * p_stats currently points at a field in the user struct. Copy
306 * parts of p_stats, and zero out the rest.
307 */
308 p2->p_stats = &up->u_stats;
309 memset(&up->u_stats.pstat_startzero, 0,
310 (unsigned) ((caddr_t)&up->u_stats.pstat_endzero -
311 (caddr_t)&up->u_stats.pstat_startzero));
312 memcpy(&up->u_stats.pstat_startcopy, &p1->p_stats->pstat_startcopy,
313 ((caddr_t)&up->u_stats.pstat_endcopy -
314 (caddr_t)&up->u_stats.pstat_startcopy));
315
316 /*
317 * cpu_fork() copy and update the pcb, and make the child ready
318 * to run. If this is a normal user fork, the child will exit
319 * directly to user mode via child_return() on its first time
320 * slice and will not return here. If this is a kernel thread,
321 * the specified entry point will be executed.
322 */
323 cpu_fork(p1, p2, stack, stacksize, func, arg);
324 }
325
326 /*
327 * uvm_exit: exit a virtual address space
328 *
329 * - the process passed to us is a dead (pre-zombie) process; we
330 * are running on a different context now (the reaper).
331 * - we must run in a separate thread because freeing the vmspace
332 * of the dead process may block.
333 */
334 void
335 uvm_exit(p)
336 struct proc *p;
337 {
338
339 uvmspace_free(p->p_vmspace);
340 uvm_km_free(kernel_map, (vaddr_t)p->p_addr, USPACE);
341 p->p_addr = NULL;
342 }
343
344 /*
345 * uvm_init_limit: init per-process VM limits
346 *
347 * - called for process 0 and then inherited by all others.
348 */
349 void
350 uvm_init_limits(p)
351 struct proc *p;
352 {
353
354 /*
355 * Set up the initial limits on process VM. Set the maximum
356 * resident set size to be all of (reasonably) available memory.
357 * This causes any single, large process to start random page
358 * replacement once it fills memory.
359 */
360
361 p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ;
362 p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ;
363 p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ;
364 p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ;
365 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free);
366 }
367
368 #ifdef DEBUG
369 int enableswap = 1;
370 int swapdebug = 0;
371 #define SDB_FOLLOW 1
372 #define SDB_SWAPIN 2
373 #define SDB_SWAPOUT 4
374 #endif
375
376 /*
377 * uvm_swapin: swap in a process's u-area.
378 */
379
380 void
381 uvm_swapin(p)
382 struct proc *p;
383 {
384 vaddr_t addr;
385 int s;
386
387 addr = (vaddr_t)p->p_addr;
388 /* make P_INMEM true */
389 uvm_fault_wire(kernel_map, addr, addr + USPACE,
390 VM_PROT_READ | VM_PROT_WRITE);
391
392 /*
393 * Some architectures need to be notified when the user area has
394 * moved to new physical page(s) (e.g. see mips/mips/vm_machdep.c).
395 */
396 cpu_swapin(p);
397 s = splstatclock();
398 if (p->p_stat == SRUN)
399 setrunqueue(p);
400 p->p_flag |= P_INMEM;
401 splx(s);
402 p->p_swtime = 0;
403 ++uvmexp.swapins;
404 }
405
406 /*
407 * uvm_scheduler: process zero main loop
408 *
409 * - attempt to swapin every swaped-out, runnable process in order of
410 * priority.
411 * - if not enough memory, wake the pagedaemon and let it clear space.
412 */
413
414 void
415 uvm_scheduler()
416 {
417 struct proc *p;
418 int pri;
419 struct proc *pp;
420 int ppri;
421 UVMHIST_FUNC("uvm_scheduler"); UVMHIST_CALLED(maphist);
422
423 loop:
424 #ifdef DEBUG
425 while (!enableswap)
426 tsleep((caddr_t)&proc0, PVM, "noswap", 0);
427 #endif
428 pp = NULL; /* process to choose */
429 ppri = INT_MIN; /* its priority */
430 proclist_lock_read();
431 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
432
433 /* is it a runnable swapped out process? */
434 if (p->p_stat == SRUN && (p->p_flag & P_INMEM) == 0) {
435 pri = p->p_swtime + p->p_slptime -
436 (p->p_nice - NZERO) * 8;
437 if (pri > ppri) { /* higher priority? remember it. */
438 pp = p;
439 ppri = pri;
440 }
441 }
442 }
443 proclist_unlock_read();
444
445 #ifdef DEBUG
446 if (swapdebug & SDB_FOLLOW)
447 printf("scheduler: running, procp %p pri %d\n", pp, ppri);
448 #endif
449 /*
450 * Nothing to do, back to sleep
451 */
452 if ((p = pp) == NULL) {
453 tsleep((caddr_t)&proc0, PVM, "scheduler", 0);
454 goto loop;
455 }
456
457 /*
458 * we have found swapped out process which we would like to bring
459 * back in.
460 *
461 * XXX: this part is really bogus cuz we could deadlock on memory
462 * despite our feeble check
463 */
464 if (uvmexp.free > atop(USPACE)) {
465 #ifdef DEBUG
466 if (swapdebug & SDB_SWAPIN)
467 printf("swapin: pid %d(%s)@%p, pri %d free %d\n",
468 p->p_pid, p->p_comm, p->p_addr, ppri, uvmexp.free);
469 #endif
470 uvm_swapin(p);
471 goto loop;
472 }
473 /*
474 * not enough memory, jab the pageout daemon and wait til the coast
475 * is clear
476 */
477 #ifdef DEBUG
478 if (swapdebug & SDB_FOLLOW)
479 printf("scheduler: no room for pid %d(%s), free %d\n",
480 p->p_pid, p->p_comm, uvmexp.free);
481 #endif
482 (void) splhigh();
483 uvm_wait("schedpwait");
484 (void) spl0();
485 #ifdef DEBUG
486 if (swapdebug & SDB_FOLLOW)
487 printf("scheduler: room again, free %d\n", uvmexp.free);
488 #endif
489 goto loop;
490 }
491
492 /*
493 * swappable: is process "p" swappable?
494 */
495
496 #define swappable(p) \
497 (((p)->p_flag & (P_SYSTEM | P_INMEM | P_WEXIT)) == P_INMEM && \
498 (p)->p_holdcnt == 0)
499
500 /*
501 * swapout_threads: find threads that can be swapped and unwire their
502 * u-areas.
503 *
504 * - called by the pagedaemon
505 * - try and swap at least one processs
506 * - processes that are sleeping or stopped for maxslp or more seconds
507 * are swapped... otherwise the longest-sleeping or stopped process
508 * is swapped, otherwise the longest resident process...
509 */
510 void
511 uvm_swapout_threads()
512 {
513 struct proc *p;
514 struct proc *outp, *outp2;
515 int outpri, outpri2;
516 int didswap = 0;
517 extern int maxslp;
518 /* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */
519
520 #ifdef DEBUG
521 if (!enableswap)
522 return;
523 #endif
524
525 /*
526 * outp/outpri : stop/sleep process with largest sleeptime < maxslp
527 * outp2/outpri2: the longest resident process (its swap time)
528 */
529 outp = outp2 = NULL;
530 outpri = outpri2 = 0;
531 proclist_lock_read();
532 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
533 if (!swappable(p))
534 continue;
535 switch (p->p_stat) {
536 case SRUN:
537 case SONPROC:
538 if (p->p_swtime > outpri2) {
539 outp2 = p;
540 outpri2 = p->p_swtime;
541 }
542 continue;
543
544 case SSLEEP:
545 case SSTOP:
546 if (p->p_slptime >= maxslp) {
547 uvm_swapout(p); /* zap! */
548 didswap++;
549 } else if (p->p_slptime > outpri) {
550 outp = p;
551 outpri = p->p_slptime;
552 }
553 continue;
554 }
555 }
556 proclist_unlock_read();
557
558 /*
559 * If we didn't get rid of any real duds, toss out the next most
560 * likely sleeping/stopped or running candidate. We only do this
561 * if we are real low on memory since we don't gain much by doing
562 * it (USPACE bytes).
563 */
564 if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) {
565 if ((p = outp) == NULL)
566 p = outp2;
567 #ifdef DEBUG
568 if (swapdebug & SDB_SWAPOUT)
569 printf("swapout_threads: no duds, try procp %p\n", p);
570 #endif
571 if (p)
572 uvm_swapout(p);
573 }
574 }
575
576 /*
577 * uvm_swapout: swap out process "p"
578 *
579 * - currently "swapout" means "unwire U-area" and "pmap_collect()"
580 * the pmap.
581 * - XXXCDC: should deactivate all process' private anonymous memory
582 */
583
584 static void
585 uvm_swapout(p)
586 struct proc *p;
587 {
588 vaddr_t addr;
589 int s;
590
591 #ifdef DEBUG
592 if (swapdebug & SDB_SWAPOUT)
593 printf("swapout: pid %d(%s)@%p, stat %x pri %d free %d\n",
594 p->p_pid, p->p_comm, p->p_addr, p->p_stat,
595 p->p_slptime, uvmexp.free);
596 #endif
597
598 /*
599 * Do any machine-specific actions necessary before swapout.
600 * This can include saving floating point state, etc.
601 */
602 cpu_swapout(p);
603
604 /*
605 * Unwire the to-be-swapped process's user struct and kernel stack.
606 */
607 addr = (vaddr_t)p->p_addr;
608 uvm_fault_unwire(kernel_map, addr, addr + USPACE); /* !P_INMEM */
609 pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map));
610
611 /*
612 * Mark it as (potentially) swapped out.
613 */
614 s = splstatclock();
615 p->p_flag &= ~P_INMEM;
616 if (p->p_stat == SRUN)
617 remrunqueue(p);
618 splx(s);
619 p->p_swtime = 0;
620 ++uvmexp.swapouts;
621 }
622
623