xref: /netbsd-src/sys/uvm/uvm_glue.c (revision 4472dbe5e3bd91ef2540bada7a7ca7384627ff9b)
1 /*	$NetBSD: uvm_glue.c,v 1.34 2000/05/28 05:49:06 thorpej 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_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_sleep: atomic unlock and sleep for UVM_UNLOCK_AND_WAIT().
113  */
114 
115 void
116 uvm_sleep(event, slock, canintr, msg, timo)
117 	void *event;
118 	struct simplelock *slock;
119 	boolean_t canintr;
120 	const char *msg;
121 	int timo;
122 {
123 	int s, pri;
124 
125 	pri = PVM;
126 	if (canintr)
127 		pri |= PCATCH;
128 
129 	s = splhigh();
130 	if (slock != NULL)
131 		simple_unlock(slock);
132 	(void) tsleep(event, pri, msg, timo);
133 	splx(s);
134 }
135 
136 /*
137  * uvm_kernacc: can the kernel access a region of memory
138  *
139  * - called from malloc [DIAGNOSTIC], and /dev/kmem driver (mem.c)
140  */
141 
142 boolean_t
143 uvm_kernacc(addr, len, rw)
144 	caddr_t addr;
145 	size_t len;
146 	int rw;
147 {
148 	boolean_t rv;
149 	vaddr_t saddr, eaddr;
150 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
151 
152 	saddr = trunc_page((vaddr_t)addr);
153 	eaddr = round_page((vaddr_t)addr+len);
154 	vm_map_lock_read(kernel_map);
155 	rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot);
156 	vm_map_unlock_read(kernel_map);
157 
158 	/*
159 	 * XXX there are still some things (e.g. the buffer cache) that
160 	 * are managed behind the VM system's back so even though an
161 	 * address is accessible in the mind of the VM system, there may
162 	 * not be physical pages where the VM thinks there is.  This can
163 	 * lead to bogus allocation of pages in the kernel address space
164 	 * or worse, inconsistencies at the pmap level.  We only worry
165 	 * about the buffer cache for now.
166 	 */
167 	if (!readbuffers && rv && (eaddr > (vaddr_t)buffers &&
168 			     saddr < (vaddr_t)buffers + MAXBSIZE * nbuf))
169 		rv = FALSE;
170 	return(rv);
171 }
172 
173 /*
174  * uvm_useracc: can the user access it?
175  *
176  * - called from physio() and sys___sysctl().
177  */
178 
179 boolean_t
180 uvm_useracc(addr, len, rw)
181 	caddr_t addr;
182 	size_t len;
183 	int rw;
184 {
185 	vm_map_t map;
186 	boolean_t rv;
187 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
188 
189 	/* XXX curproc */
190 	map = &curproc->p_vmspace->vm_map;
191 
192 	vm_map_lock_read(map);
193 	rv = uvm_map_checkprot(map, trunc_page((vaddr_t)addr),
194 	    round_page((vaddr_t)addr+len), prot);
195 	vm_map_unlock_read(map);
196 
197 	return(rv);
198 }
199 
200 #ifdef KGDB
201 /*
202  * Change protections on kernel pages from addr to addr+len
203  * (presumably so debugger can plant a breakpoint).
204  *
205  * We force the protection change at the pmap level.  If we were
206  * to use vm_map_protect a change to allow writing would be lazily-
207  * applied meaning we would still take a protection fault, something
208  * we really don't want to do.  It would also fragment the kernel
209  * map unnecessarily.  We cannot use pmap_protect since it also won't
210  * enforce a write-enable request.  Using pmap_enter is the only way
211  * we can ensure the change takes place properly.
212  */
213 void
214 uvm_chgkprot(addr, len, rw)
215 	caddr_t addr;
216 	size_t len;
217 	int rw;
218 {
219 	vm_prot_t prot;
220 	paddr_t pa;
221 	vaddr_t sva, eva;
222 
223 	prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE;
224 	eva = round_page((vaddr_t)addr + len);
225 	for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) {
226 		/*
227 		 * Extract physical address for the page.
228 		 * We use a cheezy hack to differentiate physical
229 		 * page 0 from an invalid mapping, not that it
230 		 * really matters...
231 		 */
232 		if (pmap_extract(pmap_kernel(), sva, &pa) == FALSE)
233 			panic("chgkprot: invalid page");
234 		pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED);
235 	}
236 }
237 #endif
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 int
247 uvm_vslock(p, addr, len, access_type)
248 	struct proc *p;
249 	caddr_t	addr;
250 	size_t	len;
251 	vm_prot_t access_type;
252 {
253 	vm_map_t map;
254 	vaddr_t start, end;
255 	int rv;
256 
257 	map = &p->p_vmspace->vm_map;
258 	start = trunc_page((vaddr_t)addr);
259 	end = round_page((vaddr_t)addr + len);
260 
261 	rv = uvm_fault_wire(map, start, end, access_type);
262 
263 	return (rv);
264 }
265 
266 /*
267  * vslock: wire user memory for I/O
268  *
269  * - called from physio and sys___sysctl
270  * - XXXCDC: consider nuking this (or making it a macro?)
271  */
272 
273 void
274 uvm_vsunlock(p, addr, len)
275 	struct proc *p;
276 	caddr_t	addr;
277 	size_t	len;
278 {
279 	uvm_fault_unwire(&p->p_vmspace->vm_map, trunc_page((vaddr_t)addr),
280 		round_page((vaddr_t)addr+len));
281 }
282 
283 /*
284  * uvm_fork: fork a virtual address space
285  *
286  * - the address space is copied as per parent map's inherit values
287  * - a new "user" structure is allocated for the child process
288  *	[filled in by MD layer...]
289  * - if specified, the child gets a new user stack described by
290  *	stack and stacksize
291  * - NOTE: the kernel stack may be at a different location in the child
292  *	process, and thus addresses of automatic variables may be invalid
293  *	after cpu_fork returns in the child process.  We do nothing here
294  *	after cpu_fork returns.
295  * - XXXCDC: we need a way for this to return a failure value rather
296  *   than just hang
297  */
298 void
299 uvm_fork(p1, p2, shared, stack, stacksize, func, arg)
300 	struct proc *p1, *p2;
301 	boolean_t shared;
302 	void *stack;
303 	size_t stacksize;
304 	void (*func) __P((void *));
305 	void *arg;
306 {
307 	struct user *up = p2->p_addr;
308 	int rv;
309 
310 	if (shared == TRUE)
311 		uvmspace_share(p1, p2);			/* share vmspace */
312 	else
313 		p2->p_vmspace = uvmspace_fork(p1->p_vmspace); /* fork vmspace */
314 
315 	/*
316 	 * Wire down the U-area for the process, which contains the PCB
317 	 * and the kernel stack.  Wired state is stored in p->p_flag's
318 	 * P_INMEM bit rather than in the vm_map_entry's wired count
319 	 * to prevent kernel_map fragmentation.
320 	 *
321 	 * Note the kernel stack gets read/write accesses right off
322 	 * the bat.
323 	 */
324 	rv = uvm_fault_wire(kernel_map, (vaddr_t)up,
325 	    (vaddr_t)up + USPACE, VM_PROT_READ | VM_PROT_WRITE);
326 	if (rv != KERN_SUCCESS)
327 		panic("uvm_fork: uvm_fault_wire failed: %d", rv);
328 
329 	/*
330 	 * p_stats currently points at a field in the user struct.  Copy
331 	 * parts of p_stats, and zero out the rest.
332 	 */
333 	p2->p_stats = &up->u_stats;
334 	memset(&up->u_stats.pstat_startzero, 0,
335 	(unsigned) ((caddr_t)&up->u_stats.pstat_endzero -
336 		    (caddr_t)&up->u_stats.pstat_startzero));
337 	memcpy(&up->u_stats.pstat_startcopy, &p1->p_stats->pstat_startcopy,
338 	((caddr_t)&up->u_stats.pstat_endcopy -
339 	 (caddr_t)&up->u_stats.pstat_startcopy));
340 
341 	/*
342 	 * cpu_fork() copy and update the pcb, and make the child ready
343 	 * to run.  If this is a normal user fork, the child will exit
344 	 * directly to user mode via child_return() on its first time
345 	 * slice and will not return here.  If this is a kernel thread,
346 	 * the specified entry point will be executed.
347 	 */
348 	cpu_fork(p1, p2, stack, stacksize, func, arg);
349 }
350 
351 /*
352  * uvm_exit: exit a virtual address space
353  *
354  * - the process passed to us is a dead (pre-zombie) process; we
355  *   are running on a different context now (the reaper).
356  * - we must run in a separate thread because freeing the vmspace
357  *   of the dead process may block.
358  */
359 void
360 uvm_exit(p)
361 	struct proc *p;
362 {
363 
364 	uvmspace_free(p->p_vmspace);
365 	uvm_km_free(kernel_map, (vaddr_t)p->p_addr, USPACE);
366 }
367 
368 /*
369  * uvm_init_limit: init per-process VM limits
370  *
371  * - called for process 0 and then inherited by all others.
372  */
373 void
374 uvm_init_limits(p)
375 	struct proc *p;
376 {
377 
378 	/*
379 	 * Set up the initial limits on process VM.  Set the maximum
380 	 * resident set size to be all of (reasonably) available memory.
381 	 * This causes any single, large process to start random page
382 	 * replacement once it fills memory.
383 	 */
384 
385 	p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ;
386 	p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ;
387 	p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ;
388 	p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ;
389 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free);
390 }
391 
392 #ifdef DEBUG
393 int	enableswap = 1;
394 int	swapdebug = 0;
395 #define	SDB_FOLLOW	1
396 #define SDB_SWAPIN	2
397 #define SDB_SWAPOUT	4
398 #endif
399 
400 /*
401  * uvm_swapin: swap in a process's u-area.
402  */
403 
404 void
405 uvm_swapin(p)
406 	struct proc *p;
407 {
408 	vaddr_t addr;
409 	int s;
410 
411 	addr = (vaddr_t)p->p_addr;
412 	/* make P_INMEM true */
413 	uvm_fault_wire(kernel_map, addr, addr + USPACE,
414 	    VM_PROT_READ | VM_PROT_WRITE);
415 
416 	/*
417 	 * Some architectures need to be notified when the user area has
418 	 * moved to new physical page(s) (e.g.  see mips/mips/vm_machdep.c).
419 	 */
420 	cpu_swapin(p);
421 	s = splstatclock();
422 	if (p->p_stat == SRUN)
423 		setrunqueue(p);
424 	p->p_flag |= P_INMEM;
425 	splx(s);
426 	p->p_swtime = 0;
427 	++uvmexp.swapins;
428 }
429 
430 /*
431  * uvm_scheduler: process zero main loop
432  *
433  * - attempt to swapin every swaped-out, runnable process in order of
434  *	priority.
435  * - if not enough memory, wake the pagedaemon and let it clear space.
436  */
437 
438 void
439 uvm_scheduler()
440 {
441 	struct proc *p;
442 	int pri;
443 	struct proc *pp;
444 	int ppri;
445 	UVMHIST_FUNC("uvm_scheduler"); UVMHIST_CALLED(maphist);
446 
447 loop:
448 #ifdef DEBUG
449 	while (!enableswap)
450 		tsleep((caddr_t)&proc0, PVM, "noswap", 0);
451 #endif
452 	pp = NULL;		/* process to choose */
453 	ppri = INT_MIN;	/* its priority */
454 	proclist_lock_read();
455 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
456 
457 		/* is it a runnable swapped out process? */
458 		if (p->p_stat == SRUN && (p->p_flag & P_INMEM) == 0) {
459 			pri = p->p_swtime + p->p_slptime -
460 			    (p->p_nice - NZERO) * 8;
461 			if (pri > ppri) {   /* higher priority?  remember it. */
462 				pp = p;
463 				ppri = pri;
464 			}
465 		}
466 	}
467 	proclist_unlock_read();
468 
469 #ifdef DEBUG
470 	if (swapdebug & SDB_FOLLOW)
471 		printf("scheduler: running, procp %p pri %d\n", pp, ppri);
472 #endif
473 	/*
474 	 * Nothing to do, back to sleep
475 	 */
476 	if ((p = pp) == NULL) {
477 		tsleep((caddr_t)&proc0, PVM, "scheduler", 0);
478 		goto loop;
479 	}
480 
481 	/*
482 	 * we have found swapped out process which we would like to bring
483 	 * back in.
484 	 *
485 	 * XXX: this part is really bogus cuz we could deadlock on memory
486 	 * despite our feeble check
487 	 */
488 	if (uvmexp.free > atop(USPACE)) {
489 #ifdef DEBUG
490 		if (swapdebug & SDB_SWAPIN)
491 			printf("swapin: pid %d(%s)@%p, pri %d free %d\n",
492 	     p->p_pid, p->p_comm, p->p_addr, ppri, uvmexp.free);
493 #endif
494 		uvm_swapin(p);
495 		goto loop;
496 	}
497 	/*
498 	 * not enough memory, jab the pageout daemon and wait til the coast
499 	 * is clear
500 	 */
501 #ifdef DEBUG
502 	if (swapdebug & SDB_FOLLOW)
503 		printf("scheduler: no room for pid %d(%s), free %d\n",
504 	   p->p_pid, p->p_comm, uvmexp.free);
505 #endif
506 	(void) splhigh();
507 	uvm_wait("schedpwait");
508 	(void) spl0();
509 #ifdef DEBUG
510 	if (swapdebug & SDB_FOLLOW)
511 		printf("scheduler: room again, free %d\n", uvmexp.free);
512 #endif
513 	goto loop;
514 }
515 
516 /*
517  * swappable: is process "p" swappable?
518  */
519 
520 #define	swappable(p)							\
521 	(((p)->p_flag & (P_SYSTEM | P_INMEM | P_WEXIT)) == P_INMEM &&	\
522 	 (p)->p_holdcnt == 0)
523 
524 /*
525  * swapout_threads: find threads that can be swapped and unwire their
526  *	u-areas.
527  *
528  * - called by the pagedaemon
529  * - try and swap at least one processs
530  * - processes that are sleeping or stopped for maxslp or more seconds
531  *   are swapped... otherwise the longest-sleeping or stopped process
532  *   is swapped, otherwise the longest resident process...
533  */
534 void
535 uvm_swapout_threads()
536 {
537 	struct proc *p;
538 	struct proc *outp, *outp2;
539 	int outpri, outpri2;
540 	int didswap = 0;
541 	extern int maxslp;
542 	/* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */
543 
544 #ifdef DEBUG
545 	if (!enableswap)
546 		return;
547 #endif
548 
549 	/*
550 	 * outp/outpri  : stop/sleep process with largest sleeptime < maxslp
551 	 * outp2/outpri2: the longest resident process (its swap time)
552 	 */
553 	outp = outp2 = NULL;
554 	outpri = outpri2 = 0;
555 	proclist_lock_read();
556 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
557 		if (!swappable(p))
558 			continue;
559 		switch (p->p_stat) {
560 		case SRUN:
561 		case SONPROC:
562 			if (p->p_swtime > outpri2) {
563 				outp2 = p;
564 				outpri2 = p->p_swtime;
565 			}
566 			continue;
567 
568 		case SSLEEP:
569 		case SSTOP:
570 			if (p->p_slptime >= maxslp) {
571 				uvm_swapout(p);			/* zap! */
572 				didswap++;
573 			} else if (p->p_slptime > outpri) {
574 				outp = p;
575 				outpri = p->p_slptime;
576 			}
577 			continue;
578 		}
579 	}
580 	proclist_unlock_read();
581 
582 	/*
583 	 * If we didn't get rid of any real duds, toss out the next most
584 	 * likely sleeping/stopped or running candidate.  We only do this
585 	 * if we are real low on memory since we don't gain much by doing
586 	 * it (USPACE bytes).
587 	 */
588 	if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) {
589 		if ((p = outp) == NULL)
590 			p = outp2;
591 #ifdef DEBUG
592 		if (swapdebug & SDB_SWAPOUT)
593 			printf("swapout_threads: no duds, try procp %p\n", p);
594 #endif
595 		if (p)
596 			uvm_swapout(p);
597 	}
598 }
599 
600 /*
601  * uvm_swapout: swap out process "p"
602  *
603  * - currently "swapout" means "unwire U-area" and "pmap_collect()"
604  *   the pmap.
605  * - XXXCDC: should deactivate all process' private anonymous memory
606  */
607 
608 static void
609 uvm_swapout(p)
610 	struct proc *p;
611 {
612 	vaddr_t addr;
613 	int s;
614 
615 #ifdef DEBUG
616 	if (swapdebug & SDB_SWAPOUT)
617 		printf("swapout: pid %d(%s)@%p, stat %x pri %d free %d\n",
618 	   p->p_pid, p->p_comm, p->p_addr, p->p_stat,
619 	   p->p_slptime, uvmexp.free);
620 #endif
621 
622 	/*
623 	 * Do any machine-specific actions necessary before swapout.
624 	 * This can include saving floating point state, etc.
625 	 */
626 	cpu_swapout(p);
627 
628 	/*
629 	 * Unwire the to-be-swapped process's user struct and kernel stack.
630 	 */
631 	addr = (vaddr_t)p->p_addr;
632 	uvm_fault_unwire(kernel_map, addr, addr + USPACE); /* !P_INMEM */
633 	pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map));
634 
635 	/*
636 	 * Mark it as (potentially) swapped out.
637 	 */
638 	s = splstatclock();
639 	p->p_flag &= ~P_INMEM;
640 	if (p->p_stat == SRUN)
641 		remrunqueue(p);
642 	splx(s);
643 	p->p_swtime = 0;
644 	++uvmexp.swapouts;
645 }
646 
647