xref: /openbsd-src/sys/uvm/uvm_glue.c (revision b2ea75c1b17e1a9a339660e7ed45cd24946b230e)
1 /*	$OpenBSD: uvm_glue.c,v 1.19 2001/08/11 10:57:22 art Exp $	*/
2 /*	$NetBSD: uvm_glue.c,v 1.36 2000/06/18 05:20:27 simonb 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 <vm/vm.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_kern.h>
87 
88 #include <uvm/uvm.h>
89 
90 #include <machine/cpu.h>
91 
92 /*
93  * local prototypes
94  */
95 
96 static void uvm_swapout __P((struct proc *));
97 
98 /*
99  * XXXCDC: do these really belong here?
100  */
101 
102 unsigned maxdmap = MAXDSIZ;	/* kern_resource.c: RLIMIT_DATA max */
103 unsigned maxsmap = MAXSSIZ;	/* kern_resource.c: RLIMIT_STACK max */
104 
105 int readbuffers = 0;		/* allow KGDB to read kern buffer pool */
106 				/* XXX: see uvm_kernacc */
107 
108 
109 /*
110  * uvm_kernacc: can the kernel access a region of memory
111  *
112  * - called from malloc [DIAGNOSTIC], and /dev/kmem driver (mem.c)
113  */
114 
115 boolean_t
116 uvm_kernacc(addr, len, rw)
117 	caddr_t addr;
118 	size_t len;
119 	int rw;
120 {
121 	boolean_t rv;
122 	vaddr_t saddr, eaddr;
123 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
124 
125 	saddr = trunc_page((vaddr_t)addr);
126 	eaddr = round_page((vaddr_t)addr+len);
127 	vm_map_lock_read(kernel_map);
128 	rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot);
129 	vm_map_unlock_read(kernel_map);
130 
131 	/*
132 	 * XXX there are still some things (e.g. the buffer cache) that
133 	 * are managed behind the VM system's back so even though an
134 	 * address is accessible in the mind of the VM system, there may
135 	 * not be physical pages where the VM thinks there is.  This can
136 	 * lead to bogus allocation of pages in the kernel address space
137 	 * or worse, inconsistencies at the pmap level.  We only worry
138 	 * about the buffer cache for now.
139 	 */
140 	if (!readbuffers && rv && (eaddr > (vaddr_t)buffers &&
141 			     saddr < (vaddr_t)buffers + MAXBSIZE * nbuf))
142 		rv = FALSE;
143 	return(rv);
144 }
145 
146 /*
147  * uvm_useracc: can the user access it?
148  *
149  * - called from physio() and sys___sysctl().
150  */
151 
152 boolean_t
153 uvm_useracc(addr, len, rw)
154 	caddr_t addr;
155 	size_t len;
156 	int rw;
157 {
158 	vm_map_t map;
159 	boolean_t rv;
160 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
161 
162 	/* XXX curproc */
163 	map = &curproc->p_vmspace->vm_map;
164 
165 	vm_map_lock_read(map);
166 	rv = uvm_map_checkprot(map, trunc_page((vaddr_t)addr),
167 	    round_page((vaddr_t)addr+len), prot);
168 	vm_map_unlock_read(map);
169 
170 	return(rv);
171 }
172 
173 #ifdef KGDB
174 /*
175  * Change protections on kernel pages from addr to addr+len
176  * (presumably so debugger can plant a breakpoint).
177  *
178  * We force the protection change at the pmap level.  If we were
179  * to use vm_map_protect a change to allow writing would be lazily-
180  * applied meaning we would still take a protection fault, something
181  * we really don't want to do.  It would also fragment the kernel
182  * map unnecessarily.  We cannot use pmap_protect since it also won't
183  * enforce a write-enable request.  Using pmap_enter is the only way
184  * we can ensure the change takes place properly.
185  */
186 void
187 uvm_chgkprot(addr, len, rw)
188 	caddr_t addr;
189 	size_t len;
190 	int rw;
191 {
192 	vm_prot_t prot;
193 	paddr_t pa;
194 	vaddr_t sva, eva;
195 
196 	prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE;
197 	eva = round_page((vaddr_t)addr + len);
198 	for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) {
199 		/*
200 		 * Extract physical address for the page.
201 		 * We use a cheezy hack to differentiate physical
202 		 * page 0 from an invalid mapping, not that it
203 		 * really matters...
204 		 */
205 		if (pmap_extract(pmap_kernel(), sva, &pa) == FALSE)
206 			panic("chgkprot: invalid page");
207 		pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED);
208 	}
209 }
210 #endif
211 
212 /*
213  * vslock: wire user memory for I/O
214  *
215  * - called from physio and sys___sysctl
216  * - XXXCDC: consider nuking this (or making it a macro?)
217  */
218 
219 int
220 uvm_vslock(p, addr, len, access_type)
221 	struct proc *p;
222 	caddr_t	addr;
223 	size_t	len;
224 	vm_prot_t access_type;
225 {
226 	vm_map_t map;
227 	vaddr_t start, end;
228 	int rv;
229 
230 	map = &p->p_vmspace->vm_map;
231 	start = trunc_page((vaddr_t)addr);
232 	end = round_page((vaddr_t)addr + len);
233 
234 	rv = uvm_fault_wire(map, start, end, access_type);
235 
236 	return (rv);
237 }
238 
239 /*
240  * vslock: wire user memory for I/O
241  *
242  * - called from physio and sys___sysctl
243  * - XXXCDC: consider nuking this (or making it a macro?)
244  */
245 
246 void
247 uvm_vsunlock(p, addr, len)
248 	struct proc *p;
249 	caddr_t	addr;
250 	size_t	len;
251 {
252 	uvm_fault_unwire(&p->p_vmspace->vm_map, trunc_page((vaddr_t)addr),
253 		round_page((vaddr_t)addr+len));
254 }
255 
256 /*
257  * uvm_fork: fork a virtual address space
258  *
259  * - the address space is copied as per parent map's inherit values
260  * - a new "user" structure is allocated for the child process
261  *	[filled in by MD layer...]
262  * - if specified, the child gets a new user stack described by
263  *	stack and stacksize
264  * - NOTE: the kernel stack may be at a different location in the child
265  *	process, and thus addresses of automatic variables may be invalid
266  *	after cpu_fork returns in the child process.  We do nothing here
267  *	after cpu_fork returns.
268  * - XXXCDC: we need a way for this to return a failure value rather
269  *   than just hang
270  */
271 void
272 uvm_fork(p1, p2, shared, stack, stacksize)
273 	struct proc *p1, *p2;
274 	boolean_t shared;
275 	void *stack;
276 	size_t stacksize;
277 {
278 	struct user *up = p2->p_addr;
279 	int rv;
280 
281 	if (shared == TRUE)
282 		uvmspace_share(p1, p2);			/* share vmspace */
283 	else
284 		p2->p_vmspace = uvmspace_fork(p1->p_vmspace); /* fork vmspace */
285 
286 	/*
287 	 * Wire down the U-area for the process, which contains the PCB
288 	 * and the kernel stack.  Wired state is stored in p->p_flag's
289 	 * P_INMEM bit rather than in the vm_map_entry's wired count
290 	 * to prevent kernel_map fragmentation.
291 	 *
292 	 * Note the kernel stack gets read/write accesses right off
293 	 * the bat.
294 	 */
295 	rv = uvm_fault_wire(kernel_map, (vaddr_t)up,
296 	    (vaddr_t)up + USPACE, VM_PROT_READ | VM_PROT_WRITE);
297 	if (rv != KERN_SUCCESS)
298 		panic("uvm_fork: uvm_fault_wire failed: %d", rv);
299 
300 	/*
301 	 * p_stats currently points at a field in the user struct.  Copy
302 	 * parts of p_stats, and zero out the rest.
303 	 */
304 	p2->p_stats = &up->u_stats;
305 	memset(&up->u_stats.pstat_startzero, 0,
306 	(unsigned) ((caddr_t)&up->u_stats.pstat_endzero -
307 		    (caddr_t)&up->u_stats.pstat_startzero));
308 	memcpy(&up->u_stats.pstat_startcopy, &p1->p_stats->pstat_startcopy,
309 	((caddr_t)&up->u_stats.pstat_endcopy -
310 	 (caddr_t)&up->u_stats.pstat_startcopy));
311 
312 	/*
313 	 * cpu_fork will copy and update the kernel stack and pcb, and make
314 	 * the child ready to run.  The child will exit directly to user
315 	 * mode on its first time slice, and will not return here.
316 	 */
317 	cpu_fork(p1, p2, stack, stacksize);
318 }
319 
320 /*
321  * uvm_exit: exit a virtual address space
322  *
323  * - the process passed to us is a dead (pre-zombie) process; we
324  *   are running on a different context now (the reaper).
325  * - we must run in a separate thread because freeing the vmspace
326  *   of the dead process may block.
327  */
328 void
329 uvm_exit(p)
330 	struct proc *p;
331 {
332 
333 	uvmspace_free(p->p_vmspace);
334 	uvm_km_free(kernel_map, (vaddr_t)p->p_addr, 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;
380 
381 	addr = (vaddr_t)p->p_addr;
382 	/* make P_INMEM true */
383 	uvm_fault_wire(kernel_map, addr, addr + USPACE,
384 	    VM_PROT_READ | VM_PROT_WRITE);
385 
386 	/*
387 	 * Some architectures need to be notified when the user area has
388 	 * moved to new physical page(s) (e.g.  see mips/mips/vm_machdep.c).
389 	 */
390 	cpu_swapin(p);
391 	s = splstatclock();
392 	if (p->p_stat == SRUN)
393 		setrunqueue(p);
394 	p->p_flag |= P_INMEM;
395 	splx(s);
396 	p->p_swtime = 0;
397 	++uvmexp.swapins;
398 }
399 
400 /*
401  * uvm_scheduler: process zero main loop
402  *
403  * - attempt to swapin every swaped-out, runnable process in order of
404  *	priority.
405  * - if not enough memory, wake the pagedaemon and let it clear space.
406  */
407 
408 void
409 uvm_scheduler()
410 {
411 	struct proc *p;
412 	int pri;
413 	struct proc *pp;
414 	int ppri;
415 	UVMHIST_FUNC("uvm_scheduler"); UVMHIST_CALLED(maphist);
416 
417 loop:
418 #ifdef DEBUG
419 	while (!enableswap)
420 		tsleep((caddr_t)&proc0, PVM, "noswap", 0);
421 #endif
422 	pp = NULL;		/* process to choose */
423 	ppri = INT_MIN;	/* its priority */
424 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
425 
426 		/* is it a runnable swapped out process? */
427 		if (p->p_stat == SRUN && (p->p_flag & P_INMEM) == 0) {
428 			pri = p->p_swtime + p->p_slptime -
429 			    (p->p_nice - NZERO) * 8;
430 			if (pri > ppri) {   /* higher priority?  remember it. */
431 				pp = p;
432 				ppri = pri;
433 			}
434 		}
435 	}
436 
437 #ifdef DEBUG
438 	if (swapdebug & SDB_FOLLOW)
439 		printf("scheduler: running, procp %p pri %d\n", pp, ppri);
440 #endif
441 	/*
442 	 * Nothing to do, back to sleep
443 	 */
444 	if ((p = pp) == NULL) {
445 		tsleep((caddr_t)&proc0, PVM, "scheduler", 0);
446 		goto loop;
447 	}
448 
449 	/*
450 	 * we have found swapped out process which we would like to bring
451 	 * back in.
452 	 *
453 	 * XXX: this part is really bogus cuz we could deadlock on memory
454 	 * despite our feeble check
455 	 */
456 	if (uvmexp.free > atop(USPACE)) {
457 #ifdef DEBUG
458 		if (swapdebug & SDB_SWAPIN)
459 			printf("swapin: pid %d(%s)@%p, pri %d free %d\n",
460 	     p->p_pid, p->p_comm, p->p_addr, ppri, uvmexp.free);
461 #endif
462 		uvm_swapin(p);
463 		goto loop;
464 	}
465 	/*
466 	 * not enough memory, jab the pageout daemon and wait til the coast
467 	 * is clear
468 	 */
469 #ifdef DEBUG
470 	if (swapdebug & SDB_FOLLOW)
471 		printf("scheduler: no room for pid %d(%s), free %d\n",
472 	   p->p_pid, p->p_comm, uvmexp.free);
473 #endif
474 	uvm_wait("schedpwait");
475 #ifdef DEBUG
476 	if (swapdebug & SDB_FOLLOW)
477 		printf("scheduler: room again, free %d\n", uvmexp.free);
478 #endif
479 	goto loop;
480 }
481 
482 /*
483  * swappable: is process "p" swappable?
484  */
485 
486 #define	swappable(p)							\
487 	(((p)->p_flag & (P_SYSTEM | P_INMEM | P_WEXIT)) == P_INMEM &&	\
488 	 (p)->p_holdcnt == 0)
489 
490 /*
491  * swapout_threads: find threads that can be swapped and unwire their
492  *	u-areas.
493  *
494  * - called by the pagedaemon
495  * - try and swap at least one processs
496  * - processes that are sleeping or stopped for maxslp or more seconds
497  *   are swapped... otherwise the longest-sleeping or stopped process
498  *   is swapped, otherwise the longest resident process...
499  */
500 void
501 uvm_swapout_threads()
502 {
503 	struct proc *p;
504 	struct proc *outp, *outp2;
505 	int outpri, outpri2;
506 	int didswap = 0;
507 	extern int maxslp;
508 	/* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */
509 
510 #ifdef DEBUG
511 	if (!enableswap)
512 		return;
513 #endif
514 
515 	/*
516 	 * outp/outpri  : stop/sleep process with largest sleeptime < maxslp
517 	 * outp2/outpri2: the longest resident process (its swap time)
518 	 */
519 	outp = outp2 = NULL;
520 	outpri = outpri2 = 0;
521 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
522 		if (!swappable(p))
523 			continue;
524 		switch (p->p_stat) {
525 		case SRUN:
526 			if (p->p_swtime > outpri2) {
527 				outp2 = p;
528 				outpri2 = p->p_swtime;
529 			}
530 			continue;
531 
532 		case SSLEEP:
533 		case SSTOP:
534 			if (p->p_slptime >= maxslp) {
535 				uvm_swapout(p);			/* zap! */
536 				didswap++;
537 			} else if (p->p_slptime > outpri) {
538 				outp = p;
539 				outpri = p->p_slptime;
540 			}
541 			continue;
542 		}
543 	}
544 
545 	/*
546 	 * If we didn't get rid of any real duds, toss out the next most
547 	 * likely sleeping/stopped or running candidate.  We only do this
548 	 * if we are real low on memory since we don't gain much by doing
549 	 * it (USPACE bytes).
550 	 */
551 	if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) {
552 		if ((p = outp) == NULL)
553 			p = outp2;
554 #ifdef DEBUG
555 		if (swapdebug & SDB_SWAPOUT)
556 			printf("swapout_threads: no duds, try procp %p\n", p);
557 #endif
558 		if (p)
559 			uvm_swapout(p);
560 	}
561 }
562 
563 /*
564  * uvm_swapout: swap out process "p"
565  *
566  * - currently "swapout" means "unwire U-area" and "pmap_collect()"
567  *   the pmap.
568  * - XXXCDC: should deactivate all process' private anonymous memory
569  */
570 
571 static void
572 uvm_swapout(p)
573 	struct proc *p;
574 {
575 	vaddr_t addr;
576 	int s;
577 
578 #ifdef DEBUG
579 	if (swapdebug & SDB_SWAPOUT)
580 		printf("swapout: pid %d(%s)@%p, stat %x pri %d free %d\n",
581 	   p->p_pid, p->p_comm, p->p_addr, p->p_stat,
582 	   p->p_slptime, uvmexp.free);
583 #endif
584 
585 	/*
586 	 * Do any machine-specific actions necessary before swapout.
587 	 * This can include saving floating point state, etc.
588 	 */
589 	cpu_swapout(p);
590 
591 	/*
592 	 * Unwire the to-be-swapped process's user struct and kernel stack.
593 	 */
594 	addr = (vaddr_t)p->p_addr;
595 	uvm_fault_unwire(kernel_map, addr, addr + USPACE); /* !P_INMEM */
596 	pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map));
597 
598 	/*
599 	 * Mark it as (potentially) swapped out.
600 	 */
601 	s = splstatclock();
602 	p->p_flag &= ~P_INMEM;
603 	if (p->p_stat == SRUN)
604 		remrunqueue(p);
605 	splx(s);
606 	p->p_swtime = 0;
607 	++uvmexp.swapouts;
608 }
609 
610