xref: /openbsd-src/sys/uvm/uvm_glue.c (revision 8500990981f885cbe5e6a4958549cacc238b5ae6)
1 /*	$OpenBSD: uvm_glue.c,v 1.38 2003/11/20 17:54:55 grange Exp $	*/
2 /*	$NetBSD: uvm_glue.c,v 1.44 2001/02/06 19:54:44 eeh Exp $	*/
3 
4 /*
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * Copyright (c) 1991, 1993, The Regents of the University of California.
7  *
8  * All rights reserved.
9  *
10  * This code is derived from software contributed to Berkeley by
11  * The Mach Operating System project at Carnegie-Mellon University.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. All advertising materials mentioning features or use of this software
22  *    must display the following acknowledgement:
23  *	This product includes software developed by Charles D. Cranor,
24  *      Washington University, the University of California, Berkeley and
25  *      its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vm_glue.c	8.6 (Berkeley) 1/5/94
43  * from: Id: uvm_glue.c,v 1.1.2.8 1998/02/07 01:16:54 chs Exp
44  *
45  *
46  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
47  * All rights reserved.
48  *
49  * Permission to use, copy, modify and distribute this software and
50  * its documentation is hereby granted, provided that both the copyright
51  * notice and this permission notice appear in all copies of the
52  * software, derivative works or modified versions, and any portions
53  * thereof, and that both notices appear in supporting documentation.
54  *
55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58  *
59  * Carnegie Mellon requests users of this software to return to
60  *
61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
62  *  School of Computer Science
63  *  Carnegie Mellon University
64  *  Pittsburgh PA 15213-3890
65  *
66  * any improvements or extensions that they make and grant Carnegie the
67  * rights to redistribute these changes.
68  */
69 
70 /*
71  * uvm_glue.c: glue functions
72  */
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/proc.h>
77 #include <sys/resourcevar.h>
78 #include <sys/buf.h>
79 #include <sys/user.h>
80 #ifdef SYSVSHM
81 #include <sys/shm.h>
82 #endif
83 
84 #include <uvm/uvm.h>
85 
86 #include <machine/cpu.h>
87 
88 /*
89  * local prototypes
90  */
91 
92 static void uvm_swapout(struct proc *);
93 
94 /*
95  * XXXCDC: do these really belong here?
96  */
97 
98 int readbuffers = 0;		/* allow KGDB to read kern buffer pool */
99 				/* XXX: see uvm_kernacc */
100 
101 
102 /*
103  * uvm_kernacc: can the kernel access a region of memory
104  *
105  * - called from malloc [DIAGNOSTIC], and /dev/kmem driver (mem.c)
106  */
107 
108 boolean_t
109 uvm_kernacc(addr, len, rw)
110 	caddr_t addr;
111 	size_t len;
112 	int rw;
113 {
114 	boolean_t rv;
115 	vaddr_t saddr, eaddr;
116 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
117 
118 	saddr = trunc_page((vaddr_t)addr);
119 	eaddr = round_page((vaddr_t)addr + len);
120 	vm_map_lock_read(kernel_map);
121 	rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot);
122 	vm_map_unlock_read(kernel_map);
123 
124 	/*
125 	 * XXX there are still some things (e.g. the buffer cache) that
126 	 * are managed behind the VM system's back so even though an
127 	 * address is accessible in the mind of the VM system, there may
128 	 * not be physical pages where the VM thinks there is.  This can
129 	 * lead to bogus allocation of pages in the kernel address space
130 	 * or worse, inconsistencies at the pmap level.  We only worry
131 	 * about the buffer cache for now.
132 	 */
133 	if (!readbuffers && rv && (eaddr > (vaddr_t)buffers &&
134 			     saddr < (vaddr_t)buffers + MAXBSIZE * nbuf))
135 		rv = FALSE;
136 	return(rv);
137 }
138 
139 #ifdef KGDB
140 /*
141  * Change protections on kernel pages from addr to addr+len
142  * (presumably so debugger can plant a breakpoint).
143  *
144  * We force the protection change at the pmap level.  If we were
145  * to use vm_map_protect a change to allow writing would be lazily-
146  * applied meaning we would still take a protection fault, something
147  * we really don't want to do.  It would also fragment the kernel
148  * map unnecessarily.  We cannot use pmap_protect since it also won't
149  * enforce a write-enable request.  Using pmap_enter is the only way
150  * we can ensure the change takes place properly.
151  */
152 void
153 uvm_chgkprot(addr, len, rw)
154 	caddr_t addr;
155 	size_t len;
156 	int rw;
157 {
158 	vm_prot_t prot;
159 	paddr_t pa;
160 	vaddr_t sva, eva;
161 
162 	prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE;
163 	eva = round_page((vaddr_t)addr + len);
164 	for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) {
165 		/*
166 		 * Extract physical address for the page.
167 		 * We use a cheezy hack to differentiate physical
168 		 * page 0 from an invalid mapping, not that it
169 		 * really matters...
170 		 */
171 		if (pmap_extract(pmap_kernel(), sva, &pa) == FALSE)
172 			panic("chgkprot: invalid page");
173 		pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED);
174 	}
175 }
176 #endif
177 
178 /*
179  * uvm_vslock: wire user memory for I/O
180  *
181  * - called from physio and sys___sysctl
182  * - XXXCDC: consider nuking this (or making it a macro?)
183  */
184 
185 int
186 uvm_vslock(p, addr, len, access_type)
187 	struct proc *p;
188 	caddr_t	addr;
189 	size_t	len;
190 	vm_prot_t access_type;
191 {
192 	vm_map_t map;
193 	vaddr_t start, end;
194 	int rv;
195 
196 	map = &p->p_vmspace->vm_map;
197 	start = trunc_page((vaddr_t)addr);
198 	end = round_page((vaddr_t)addr + len);
199 	if (end <= start)
200 		return (EINVAL);
201 
202 	rv = uvm_fault_wire(map, start, end, access_type);
203 
204 	return (rv);
205 }
206 
207 /*
208  * uvm_vsunlock: unwire user memory wired by uvm_vslock()
209  *
210  * - called from physio and sys___sysctl
211  * - XXXCDC: consider nuking this (or making it a macro?)
212  */
213 
214 void
215 uvm_vsunlock(p, addr, len)
216 	struct proc *p;
217 	caddr_t	addr;
218 	size_t	len;
219 {
220 	vaddr_t start, end;
221 
222 	start = trunc_page((vaddr_t)addr);
223 	end = round_page((vaddr_t)addr + len);
224 	if (end <= start)
225 		return;
226 
227 	uvm_fault_unwire(&p->p_vmspace->vm_map, start, end);
228 }
229 
230 /*
231  * uvm_fork: fork a virtual address space
232  *
233  * - the address space is copied as per parent map's inherit values
234  * - a new "user" structure is allocated for the child process
235  *	[filled in by MD layer...]
236  * - if specified, the child gets a new user stack described by
237  *	stack and stacksize
238  * - NOTE: the kernel stack may be at a different location in the child
239  *	process, and thus addresses of automatic variables may be invalid
240  *	after cpu_fork returns in the child process.  We do nothing here
241  *	after cpu_fork returns.
242  * - XXXCDC: we need a way for this to return a failure value rather
243  *   than just hang
244  */
245 void
246 uvm_fork(p1, p2, shared, stack, stacksize, func, arg)
247 	struct proc *p1, *p2;
248 	boolean_t shared;
249 	void *stack;
250 	size_t stacksize;
251 	void (*func)(void *);
252 	void *arg;
253 {
254 	struct user *up = p2->p_addr;
255 	int rv;
256 
257 	if (shared == TRUE) {
258 		p2->p_vmspace = NULL;
259 		uvmspace_share(p1, p2);			/* share vmspace */
260 	} else
261 		p2->p_vmspace = uvmspace_fork(p1->p_vmspace); /* fork vmspace */
262 
263 	/*
264 	 * Wire down the U-area for the process, which contains the PCB
265 	 * and the kernel stack.  Wired state is stored in p->p_flag's
266 	 * P_INMEM bit rather than in the vm_map_entry's wired count
267 	 * to prevent kernel_map fragmentation.
268 	 *
269 	 * Note the kernel stack gets read/write accesses right off
270 	 * the bat.
271 	 */
272 	rv = uvm_fault_wire(kernel_map, (vaddr_t)up,
273 	    (vaddr_t)up + USPACE, VM_PROT_READ | VM_PROT_WRITE);
274 	if (rv != KERN_SUCCESS)
275 		panic("uvm_fork: uvm_fault_wire failed: %d", rv);
276 
277 	/*
278 	 * p_stats currently points at a field in the user struct.  Copy
279 	 * parts of p_stats, and zero out the rest.
280 	 */
281 	p2->p_stats = &up->u_stats;
282 	memset(&up->u_stats.pstat_startzero, 0,
283 	       ((caddr_t)&up->u_stats.pstat_endzero -
284 		(caddr_t)&up->u_stats.pstat_startzero));
285 	memcpy(&up->u_stats.pstat_startcopy, &p1->p_stats->pstat_startcopy,
286 	       ((caddr_t)&up->u_stats.pstat_endcopy -
287 		(caddr_t)&up->u_stats.pstat_startcopy));
288 
289 	/*
290 	 * cpu_fork() copy and update the pcb, and make the child ready
291 	 * to run.  If this is a normal user fork, the child will exit
292 	 * directly to user mode via child_return() on its first time
293 	 * slice and will not return here.  If this is a kernel thread,
294 	 * the specified entry point will be executed.
295 	 */
296 	cpu_fork(p1, p2, stack, stacksize, func, arg);
297 }
298 
299 /*
300  * uvm_exit: exit a virtual address space
301  *
302  * - the process passed to us is a dead (pre-zombie) process; we
303  *   are running on a different context now (the reaper).
304  * - we must run in a separate thread because freeing the vmspace
305  *   of the dead process may block.
306  */
307 void
308 uvm_exit(p)
309 	struct proc *p;
310 {
311 	vaddr_t va = (vaddr_t)p->p_addr;
312 
313 	uvmspace_free(p->p_vmspace);
314 	p->p_flag &= ~P_INMEM;
315 	uvm_fault_unwire(kernel_map, va, va + USPACE);
316 	uvm_km_free(kernel_map, va, USPACE);
317 	p->p_addr = NULL;
318 }
319 
320 /*
321  * uvm_init_limit: init per-process VM limits
322  *
323  * - called for process 0 and then inherited by all others.
324  */
325 void
326 uvm_init_limits(p)
327 	struct proc *p;
328 {
329 
330 	/*
331 	 * Set up the initial limits on process VM.  Set the maximum
332 	 * resident set size to be all of (reasonably) available memory.
333 	 * This causes any single, large process to start random page
334 	 * replacement once it fills memory.
335 	 */
336 
337 	p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ;
338 	p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ;
339 	p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ;
340 	p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ;
341 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free);
342 }
343 
344 #ifdef DEBUG
345 int	enableswap = 1;
346 int	swapdebug = 0;
347 #define	SDB_FOLLOW	1
348 #define SDB_SWAPIN	2
349 #define SDB_SWAPOUT	4
350 #endif
351 
352 /*
353  * uvm_swapin: swap in a process's u-area.
354  */
355 
356 void
357 uvm_swapin(p)
358 	struct proc *p;
359 {
360 	vaddr_t addr;
361 	int rv, s;
362 
363 	s = splstatclock();
364 	if (p->p_flag & P_SWAPIN) {
365 		splx(s);
366 		return;
367 	}
368 	p->p_flag |= P_SWAPIN;
369 	splx(s);
370 
371 	addr = (vaddr_t)p->p_addr;
372 	/* make P_INMEM true */
373 	if ((rv = uvm_fault_wire(kernel_map, addr, addr + USPACE,
374 	    VM_PROT_READ | VM_PROT_WRITE)) != KERN_SUCCESS)
375 		panic("uvm_swapin: uvm_fault_wire failed: %d", rv);
376 
377 	/*
378 	 * Some architectures need to be notified when the user area has
379 	 * moved to new physical page(s) (e.g.  see mips/mips/vm_machdep.c).
380 	 */
381 	cpu_swapin(p);
382 	s = splstatclock();
383 	if (p->p_stat == SRUN)
384 		setrunqueue(p);
385 	p->p_flag |= P_INMEM;
386 	p->p_flag &= ~P_SWAPIN;
387 	splx(s);
388 	p->p_swtime = 0;
389 	++uvmexp.swapins;
390 }
391 
392 /*
393  * uvm_scheduler: process zero main loop
394  *
395  * - attempt to swapin every swaped-out, runnable process in order of
396  *	priority.
397  * - if not enough memory, wake the pagedaemon and let it clear space.
398  */
399 
400 void
401 uvm_scheduler()
402 {
403 	struct proc *p;
404 	int pri;
405 	struct proc *pp;
406 	int ppri;
407 
408 loop:
409 #ifdef DEBUG
410 	while (!enableswap)
411 		tsleep(&proc0, PVM, "noswap", 0);
412 #endif
413 	pp = NULL;		/* process to choose */
414 	ppri = INT_MIN;	/* its priority */
415 	LIST_FOREACH(p, &allproc, p_list) {
416 
417 		/* is it a runnable swapped out process? */
418 		if (p->p_stat == SRUN && (p->p_flag & P_INMEM) == 0) {
419 			pri = p->p_swtime + p->p_slptime -
420 			    (p->p_nice - NZERO) * 8;
421 			if (pri > ppri) {   /* higher priority?  remember it. */
422 				pp = p;
423 				ppri = pri;
424 			}
425 		}
426 	}
427 
428 #ifdef DEBUG
429 	if (swapdebug & SDB_FOLLOW)
430 		printf("scheduler: running, procp %p pri %d\n", pp, ppri);
431 #endif
432 	/*
433 	 * Nothing to do, back to sleep
434 	 */
435 	if ((p = pp) == NULL) {
436 		tsleep(&proc0, PVM, "scheduler", 0);
437 		goto loop;
438 	}
439 
440 	/*
441 	 * we have found swapped out process which we would like to bring
442 	 * back in.
443 	 *
444 	 * XXX: this part is really bogus because we could deadlock on memory
445 	 * despite our feeble check
446 	 */
447 	if (uvmexp.free > atop(USPACE)) {
448 #ifdef DEBUG
449 		if (swapdebug & SDB_SWAPIN)
450 			printf("swapin: pid %d(%s)@%p, pri %d free %d\n",
451 	     p->p_pid, p->p_comm, p->p_addr, ppri, uvmexp.free);
452 #endif
453 		uvm_swapin(p);
454 		goto loop;
455 	}
456 	/*
457 	 * not enough memory, jab the pageout daemon and wait til the coast
458 	 * is clear
459 	 */
460 #ifdef DEBUG
461 	if (swapdebug & SDB_FOLLOW)
462 		printf("scheduler: no room for pid %d(%s), free %d\n",
463 	   p->p_pid, p->p_comm, uvmexp.free);
464 #endif
465 	uvm_wait("schedpwait");
466 #ifdef DEBUG
467 	if (swapdebug & SDB_FOLLOW)
468 		printf("scheduler: room again, free %d\n", uvmexp.free);
469 #endif
470 	goto loop;
471 }
472 
473 /*
474  * swappable: is process "p" swappable?
475  */
476 
477 #define	swappable(p)							\
478 	(((p)->p_flag & (P_SYSTEM | P_INMEM | P_WEXIT)) == P_INMEM &&	\
479 	 (p)->p_holdcnt == 0)
480 
481 /*
482  * swapout_threads: find threads that can be swapped and unwire their
483  *	u-areas.
484  *
485  * - called by the pagedaemon
486  * - try and swap at least one processs
487  * - processes that are sleeping or stopped for maxslp or more seconds
488  *   are swapped... otherwise the longest-sleeping or stopped process
489  *   is swapped, otherwise the longest resident process...
490  */
491 void
492 uvm_swapout_threads()
493 {
494 	struct proc *p;
495 	struct proc *outp, *outp2;
496 	int outpri, outpri2;
497 	int didswap = 0;
498 	extern int maxslp;
499 	/* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */
500 
501 #ifdef DEBUG
502 	if (!enableswap)
503 		return;
504 #endif
505 
506 	/*
507 	 * outp/outpri  : stop/sleep process with largest sleeptime < maxslp
508 	 * outp2/outpri2: the longest resident process (its swap time)
509 	 */
510 	outp = outp2 = NULL;
511 	outpri = outpri2 = 0;
512 	LIST_FOREACH(p, &allproc, p_list) {
513 		if (!swappable(p))
514 			continue;
515 		switch (p->p_stat) {
516 		case SRUN:
517 			if (p->p_swtime > outpri2) {
518 				outp2 = p;
519 				outpri2 = p->p_swtime;
520 			}
521 			continue;
522 
523 		case SSLEEP:
524 		case SSTOP:
525 			if (p->p_slptime >= maxslp) {
526 				uvm_swapout(p);
527 				didswap++;
528 			} else if (p->p_slptime > outpri) {
529 				outp = p;
530 				outpri = p->p_slptime;
531 			}
532 			continue;
533 		}
534 	}
535 
536 	/*
537 	 * If we didn't get rid of any real duds, toss out the next most
538 	 * likely sleeping/stopped or running candidate.  We only do this
539 	 * if we are real low on memory since we don't gain much by doing
540 	 * it (USPACE bytes).
541 	 */
542 	if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) {
543 		if ((p = outp) == NULL)
544 			p = outp2;
545 #ifdef DEBUG
546 		if (swapdebug & SDB_SWAPOUT)
547 			printf("swapout_threads: no duds, try procp %p\n", p);
548 #endif
549 		if (p)
550 			uvm_swapout(p);
551 	}
552 }
553 
554 /*
555  * uvm_swapout: swap out process "p"
556  *
557  * - currently "swapout" means "unwire U-area" and "pmap_collect()"
558  *   the pmap.
559  * - XXXCDC: should deactivate all process' private anonymous memory
560  */
561 
562 static void
563 uvm_swapout(p)
564 	struct proc *p;
565 {
566 	vaddr_t addr;
567 	int s;
568 
569 #ifdef DEBUG
570 	if (swapdebug & SDB_SWAPOUT)
571 		printf("swapout: pid %d(%s)@%p, stat %x pri %d free %d\n",
572 		    p->p_pid, p->p_comm, p->p_addr, p->p_stat,
573 		    p->p_slptime, uvmexp.free);
574 #endif
575 
576 	/*
577 	 * Mark it as (potentially) swapped out.
578 	 */
579 	s = splstatclock();
580 	if (!(p->p_flag & P_INMEM)) {
581 		splx(s);
582 		return;
583 	}
584 	p->p_flag &= ~P_INMEM;
585 	if (p->p_stat == SRUN)
586 		remrunqueue(p);
587 	splx(s);
588 	p->p_swtime = 0;
589 	++uvmexp.swapouts;
590 
591 	/*
592 	 * Do any machine-specific actions necessary before swapout.
593 	 * This can include saving floating point state, etc.
594 	 */
595 	cpu_swapout(p);
596 
597 	/*
598 	 * Unwire the to-be-swapped process's user struct and kernel stack.
599 	 */
600 	addr = (vaddr_t)p->p_addr;
601 	uvm_fault_unwire(kernel_map, addr, addr + USPACE); /* !P_INMEM */
602 	pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map));
603 }
604 
605