xref: /netbsd-src/sys/uvm/uvm_glue.c (revision 27578b9aac214cc7796ead81dcc5427e79d5f2a0)
1 /*	$NetBSD: uvm_glue.c,v 1.52 2001/09/15 20:36:45 chs 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@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_kgdb.h"
70 #include "opt_sysv.h"
71 #include "opt_uvmhist.h"
72 
73 /*
74  * uvm_glue.c: glue functions
75  */
76 
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/proc.h>
80 #include <sys/resourcevar.h>
81 #include <sys/buf.h>
82 #include <sys/user.h>
83 #ifdef SYSVSHM
84 #include <sys/shm.h>
85 #endif
86 
87 #include <uvm/uvm.h>
88 
89 #include <machine/cpu.h>
90 
91 /*
92  * local prototypes
93  */
94 
95 static void uvm_swapout __P((struct proc *));
96 
97 /*
98  * XXXCDC: do these really belong here?
99  */
100 
101 int readbuffers = 0;		/* allow KGDB to read kern buffer pool */
102 				/* XXX: see uvm_kernacc */
103 
104 
105 /*
106  * uvm_kernacc: can the kernel access a region of memory
107  *
108  * - called from malloc [DIAGNOSTIC], and /dev/kmem driver (mem.c)
109  */
110 
111 boolean_t
112 uvm_kernacc(addr, len, rw)
113 	caddr_t addr;
114 	size_t len;
115 	int rw;
116 {
117 	boolean_t rv;
118 	vaddr_t saddr, eaddr;
119 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
120 
121 	saddr = trunc_page((vaddr_t)addr);
122 	eaddr = round_page((vaddr_t)addr + len);
123 	vm_map_lock_read(kernel_map);
124 	rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot);
125 	vm_map_unlock_read(kernel_map);
126 
127 	/*
128 	 * XXX there are still some things (e.g. the buffer cache) that
129 	 * are managed behind the VM system's back so even though an
130 	 * address is accessible in the mind of the VM system, there may
131 	 * not be physical pages where the VM thinks there is.  This can
132 	 * lead to bogus allocation of pages in the kernel address space
133 	 * or worse, inconsistencies at the pmap level.  We only worry
134 	 * about the buffer cache for now.
135 	 */
136 	if (!readbuffers && rv && (eaddr > (vaddr_t)buffers &&
137 			     saddr < (vaddr_t)buffers + MAXBSIZE * nbuf))
138 		rv = FALSE;
139 	return(rv);
140 }
141 
142 /*
143  * uvm_useracc: can the user access it?
144  *
145  * - called from physio() and sys___sysctl().
146  */
147 
148 boolean_t
149 uvm_useracc(addr, len, rw)
150 	caddr_t addr;
151 	size_t len;
152 	int rw;
153 {
154 	struct vm_map *map;
155 	boolean_t rv;
156 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
157 
158 	/* XXX curproc */
159 	map = &curproc->p_vmspace->vm_map;
160 
161 	vm_map_lock_read(map);
162 	rv = uvm_map_checkprot(map, trunc_page((vaddr_t)addr),
163 	    round_page((vaddr_t)addr + len), prot);
164 	vm_map_unlock_read(map);
165 
166 	return(rv);
167 }
168 
169 #ifdef KGDB
170 /*
171  * Change protections on kernel pages from addr to addr+len
172  * (presumably so debugger can plant a breakpoint).
173  *
174  * We force the protection change at the pmap level.  If we were
175  * to use vm_map_protect a change to allow writing would be lazily-
176  * applied meaning we would still take a protection fault, something
177  * we really don't want to do.  It would also fragment the kernel
178  * map unnecessarily.  We cannot use pmap_protect since it also won't
179  * enforce a write-enable request.  Using pmap_enter is the only way
180  * we can ensure the change takes place properly.
181  */
182 void
183 uvm_chgkprot(addr, len, rw)
184 	caddr_t addr;
185 	size_t len;
186 	int rw;
187 {
188 	vm_prot_t prot;
189 	paddr_t pa;
190 	vaddr_t sva, eva;
191 
192 	prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE;
193 	eva = round_page((vaddr_t)addr + len);
194 	for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) {
195 		/*
196 		 * Extract physical address for the page.
197 		 */
198 		if (pmap_extract(pmap_kernel(), sva, &pa) == FALSE)
199 			panic("chgkprot: invalid page");
200 		pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED);
201 	}
202 	pmap_update(pmap_kernel());
203 }
204 #endif
205 
206 /*
207  * uvm_vslock: wire user memory for I/O
208  *
209  * - called from physio and sys___sysctl
210  * - XXXCDC: consider nuking this (or making it a macro?)
211  */
212 
213 int
214 uvm_vslock(p, addr, len, access_type)
215 	struct proc *p;
216 	caddr_t	addr;
217 	size_t	len;
218 	vm_prot_t access_type;
219 {
220 	struct vm_map *map;
221 	vaddr_t start, end;
222 	int error;
223 
224 	map = &p->p_vmspace->vm_map;
225 	start = trunc_page((vaddr_t)addr);
226 	end = round_page((vaddr_t)addr + len);
227 	error = uvm_fault_wire(map, start, end, access_type);
228 	return error;
229 }
230 
231 /*
232  * uvm_vsunlock: unwire user memory wired by uvm_vslock()
233  *
234  * - called from physio and sys___sysctl
235  * - XXXCDC: consider nuking this (or making it a macro?)
236  */
237 
238 void
239 uvm_vsunlock(p, addr, len)
240 	struct proc *p;
241 	caddr_t	addr;
242 	size_t	len;
243 {
244 	uvm_fault_unwire(&p->p_vmspace->vm_map, trunc_page((vaddr_t)addr),
245 		round_page((vaddr_t)addr + len));
246 }
247 
248 /*
249  * uvm_fork: fork a virtual address space
250  *
251  * - the address space is copied as per parent map's inherit values
252  * - a new "user" structure is allocated for the child process
253  *	[filled in by MD layer...]
254  * - if specified, the child gets a new user stack described by
255  *	stack and stacksize
256  * - NOTE: the kernel stack may be at a different location in the child
257  *	process, and thus addresses of automatic variables may be invalid
258  *	after cpu_fork returns in the child process.  We do nothing here
259  *	after cpu_fork returns.
260  * - XXXCDC: we need a way for this to return a failure value rather
261  *   than just hang
262  */
263 void
264 uvm_fork(p1, p2, shared, stack, stacksize, func, arg)
265 	struct proc *p1, *p2;
266 	boolean_t shared;
267 	void *stack;
268 	size_t stacksize;
269 	void (*func) __P((void *));
270 	void *arg;
271 {
272 	struct user *up = p2->p_addr;
273 	int error;
274 
275 	if (shared == TRUE) {
276 		p2->p_vmspace = NULL;
277 		uvmspace_share(p1, p2);
278 	} else
279 		p2->p_vmspace = uvmspace_fork(p1->p_vmspace);
280 
281 	/*
282 	 * Wire down the U-area for the process, which contains the PCB
283 	 * and the kernel stack.  Wired state is stored in p->p_flag's
284 	 * P_INMEM bit rather than in the vm_map_entry's wired count
285 	 * to prevent kernel_map fragmentation.
286 	 *
287 	 * Note the kernel stack gets read/write accesses right off
288 	 * the bat.
289 	 */
290 	error = uvm_fault_wire(kernel_map, (vaddr_t)up,
291 	    (vaddr_t)up + USPACE, VM_PROT_READ | VM_PROT_WRITE);
292 	if (error)
293 		panic("uvm_fork: uvm_fault_wire failed: %d", error);
294 
295 	/*
296 	 * p_stats currently points at a field in the user struct.  Copy
297 	 * parts of p_stats, and zero out the rest.
298 	 */
299 	p2->p_stats = &up->u_stats;
300 	memset(&up->u_stats.pstat_startzero, 0,
301 	       ((caddr_t)&up->u_stats.pstat_endzero -
302 		(caddr_t)&up->u_stats.pstat_startzero));
303 	memcpy(&up->u_stats.pstat_startcopy, &p1->p_stats->pstat_startcopy,
304 	       ((caddr_t)&up->u_stats.pstat_endcopy -
305 		(caddr_t)&up->u_stats.pstat_startcopy));
306 
307 	/*
308 	 * cpu_fork() copy and update the pcb, and make the child ready
309 	 * to run.  If this is a normal user fork, the child will exit
310 	 * directly to user mode via child_return() on its first time
311 	 * slice and will not return here.  If this is a kernel thread,
312 	 * the specified entry point will be executed.
313 	 */
314 	cpu_fork(p1, p2, stack, stacksize, func, arg);
315 }
316 
317 /*
318  * uvm_exit: exit a virtual address space
319  *
320  * - the process passed to us is a dead (pre-zombie) process; we
321  *   are running on a different context now (the reaper).
322  * - we must run in a separate thread because freeing the vmspace
323  *   of the dead process may block.
324  */
325 void
326 uvm_exit(p)
327 	struct proc *p;
328 {
329 	vaddr_t va = (vaddr_t)p->p_addr;
330 
331 	uvmspace_free(p->p_vmspace);
332 	p->p_flag &= ~P_INMEM;
333 	uvm_fault_unwire(kernel_map, va, va + USPACE);
334 	uvm_km_free(kernel_map, va, USPACE);
335 	p->p_addr = NULL;
336 }
337 
338 /*
339  * uvm_init_limit: init per-process VM limits
340  *
341  * - called for process 0 and then inherited by all others.
342  */
343 void
344 uvm_init_limits(p)
345 	struct proc *p;
346 {
347 
348 	/*
349 	 * Set up the initial limits on process VM.  Set the maximum
350 	 * resident set size to be all of (reasonably) available memory.
351 	 * This causes any single, large process to start random page
352 	 * replacement once it fills memory.
353 	 */
354 
355 	p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ;
356 	p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ;
357 	p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ;
358 	p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ;
359 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free);
360 }
361 
362 #ifdef DEBUG
363 int	enableswap = 1;
364 int	swapdebug = 0;
365 #define	SDB_FOLLOW	1
366 #define SDB_SWAPIN	2
367 #define SDB_SWAPOUT	4
368 #endif
369 
370 /*
371  * uvm_swapin: swap in a process's u-area.
372  */
373 
374 void
375 uvm_swapin(p)
376 	struct proc *p;
377 {
378 	vaddr_t addr;
379 	int s, error;
380 
381 	addr = (vaddr_t)p->p_addr;
382 	/* make P_INMEM true */
383 	error = uvm_fault_wire(kernel_map, addr, addr + USPACE,
384 	    VM_PROT_READ | VM_PROT_WRITE);
385 	if (error) {
386 		panic("uvm_swapin: rewiring stack failed: %d", error);
387 	}
388 
389 	/*
390 	 * Some architectures need to be notified when the user area has
391 	 * moved to new physical page(s) (e.g.  see mips/mips/vm_machdep.c).
392 	 */
393 	cpu_swapin(p);
394 	SCHED_LOCK(s);
395 	if (p->p_stat == SRUN)
396 		setrunqueue(p);
397 	p->p_flag |= P_INMEM;
398 	SCHED_UNLOCK(s);
399 	p->p_swtime = 0;
400 	++uvmexp.swapins;
401 }
402 
403 /*
404  * uvm_scheduler: process zero main loop
405  *
406  * - attempt to swapin every swaped-out, runnable process in order of
407  *	priority.
408  * - if not enough memory, wake the pagedaemon and let it clear space.
409  */
410 
411 void
412 uvm_scheduler()
413 {
414 	struct proc *p;
415 	int pri;
416 	struct proc *pp;
417 	int ppri;
418 
419 loop:
420 #ifdef DEBUG
421 	while (!enableswap)
422 		tsleep(&proc0, PVM, "noswap", 0);
423 #endif
424 	pp = NULL;		/* process to choose */
425 	ppri = INT_MIN;	/* its priority */
426 	proclist_lock_read();
427 	LIST_FOREACH(p, &allproc, p_list) {
428 
429 		/* is it a runnable swapped out process? */
430 		if (p->p_stat == SRUN && (p->p_flag & P_INMEM) == 0) {
431 			pri = p->p_swtime + p->p_slptime -
432 			    (p->p_nice - NZERO) * 8;
433 			if (pri > ppri) {   /* higher priority?  remember it. */
434 				pp = p;
435 				ppri = pri;
436 			}
437 		}
438 	}
439 	/*
440 	 * XXXSMP: possible unlock/sleep race between here and the
441 	 * "scheduler" tsleep below..
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(&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 	uvm_wait("schedpwait");
483 #ifdef DEBUG
484 	if (swapdebug & SDB_FOLLOW)
485 		printf("scheduler: room again, free %d\n", uvmexp.free);
486 #endif
487 	goto loop;
488 }
489 
490 /*
491  * swappable: is process "p" swappable?
492  */
493 
494 #define	swappable(p)							\
495 	(((p)->p_flag & (P_SYSTEM | P_INMEM | P_WEXIT)) == P_INMEM &&	\
496 	 (p)->p_holdcnt == 0)
497 
498 /*
499  * swapout_threads: find threads that can be swapped and unwire their
500  *	u-areas.
501  *
502  * - called by the pagedaemon
503  * - try and swap at least one processs
504  * - processes that are sleeping or stopped for maxslp or more seconds
505  *   are swapped... otherwise the longest-sleeping or stopped process
506  *   is swapped, otherwise the longest resident process...
507  */
508 void
509 uvm_swapout_threads()
510 {
511 	struct proc *p;
512 	struct proc *outp, *outp2;
513 	int outpri, outpri2;
514 	int didswap = 0;
515 	extern int maxslp;
516 	/* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */
517 
518 #ifdef DEBUG
519 	if (!enableswap)
520 		return;
521 #endif
522 
523 	/*
524 	 * outp/outpri  : stop/sleep process with largest sleeptime < maxslp
525 	 * outp2/outpri2: the longest resident process (its swap time)
526 	 */
527 	outp = outp2 = NULL;
528 	outpri = outpri2 = 0;
529 	proclist_lock_read();
530 	LIST_FOREACH(p, &allproc, p_list) {
531 		if (!swappable(p))
532 			continue;
533 		switch (p->p_stat) {
534 		case SRUN:
535 		case SONPROC:
536 			if (p->p_swtime > outpri2) {
537 				outp2 = p;
538 				outpri2 = p->p_swtime;
539 			}
540 			continue;
541 
542 		case SSLEEP:
543 		case SSTOP:
544 			if (p->p_slptime >= maxslp) {
545 				uvm_swapout(p);
546 				didswap++;
547 			} else if (p->p_slptime > outpri) {
548 				outp = p;
549 				outpri = p->p_slptime;
550 			}
551 			continue;
552 		}
553 	}
554 	proclist_unlock_read();
555 
556 	/*
557 	 * If we didn't get rid of any real duds, toss out the next most
558 	 * likely sleeping/stopped or running candidate.  We only do this
559 	 * if we are real low on memory since we don't gain much by doing
560 	 * it (USPACE bytes).
561 	 */
562 	if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) {
563 		if ((p = outp) == NULL)
564 			p = outp2;
565 #ifdef DEBUG
566 		if (swapdebug & SDB_SWAPOUT)
567 			printf("swapout_threads: no duds, try procp %p\n", p);
568 #endif
569 		if (p)
570 			uvm_swapout(p);
571 	}
572 }
573 
574 /*
575  * uvm_swapout: swap out process "p"
576  *
577  * - currently "swapout" means "unwire U-area" and "pmap_collect()"
578  *   the pmap.
579  * - XXXCDC: should deactivate all process' private anonymous memory
580  */
581 
582 static void
583 uvm_swapout(p)
584 	struct proc *p;
585 {
586 	vaddr_t addr;
587 	int s;
588 
589 #ifdef DEBUG
590 	if (swapdebug & SDB_SWAPOUT)
591 		printf("swapout: pid %d(%s)@%p, stat %x pri %d free %d\n",
592 	   p->p_pid, p->p_comm, p->p_addr, p->p_stat,
593 	   p->p_slptime, uvmexp.free);
594 #endif
595 
596 	/*
597 	 * Do any machine-specific actions necessary before swapout.
598 	 * This can include saving floating point state, etc.
599 	 */
600 	cpu_swapout(p);
601 
602 	/*
603 	 * Mark it as (potentially) swapped out.
604 	 */
605 	SCHED_LOCK(s);
606 	p->p_flag &= ~P_INMEM;
607 	if (p->p_stat == SRUN)
608 		remrunqueue(p);
609 	SCHED_UNLOCK(s);
610 	p->p_swtime = 0;
611 	++uvmexp.swapouts;
612 
613 	/*
614 	 * Unwire the to-be-swapped process's user struct and kernel stack.
615 	 */
616 	addr = (vaddr_t)p->p_addr;
617 	uvm_fault_unwire(kernel_map, addr, addr + USPACE); /* !P_INMEM */
618 	pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map));
619 }
620 
621