1 /*	$OpenBSD: uvm_glue.c,v 1.39 2004/02/23 06:19:32 drahn 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
uvm_kernacc(addr,len,rw)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
uvm_chgkprot(addr,len,rw)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 	pmap_update(pmap_kernel());
176 }
177 #endif
178 
179 /*
180  * uvm_vslock: wire user memory for I/O
181  *
182  * - called from physio and sys___sysctl
183  * - XXXCDC: consider nuking this (or making it a macro?)
184  */
185 
186 int
uvm_vslock(p,addr,len,access_type)187 uvm_vslock(p, addr, len, access_type)
188 	struct proc *p;
189 	caddr_t	addr;
190 	size_t	len;
191 	vm_prot_t access_type;
192 {
193 	vm_map_t map;
194 	vaddr_t start, end;
195 	int rv;
196 
197 	map = &p->p_vmspace->vm_map;
198 	start = trunc_page((vaddr_t)addr);
199 	end = round_page((vaddr_t)addr + len);
200 	if (end <= start)
201 		return (EINVAL);
202 
203 	rv = uvm_fault_wire(map, start, end, access_type);
204 
205 	return (rv);
206 }
207 
208 /*
209  * uvm_vsunlock: unwire user memory wired by uvm_vslock()
210  *
211  * - called from physio and sys___sysctl
212  * - XXXCDC: consider nuking this (or making it a macro?)
213  */
214 
215 void
uvm_vsunlock(p,addr,len)216 uvm_vsunlock(p, addr, len)
217 	struct proc *p;
218 	caddr_t	addr;
219 	size_t	len;
220 {
221 	vaddr_t start, end;
222 
223 	start = trunc_page((vaddr_t)addr);
224 	end = round_page((vaddr_t)addr + len);
225 	if (end <= start)
226 		return;
227 
228 	uvm_fault_unwire(&p->p_vmspace->vm_map, start, end);
229 }
230 
231 /*
232  * uvm_fork: fork a virtual address space
233  *
234  * - the address space is copied as per parent map's inherit values
235  * - a new "user" structure is allocated for the child process
236  *	[filled in by MD layer...]
237  * - if specified, the child gets a new user stack described by
238  *	stack and stacksize
239  * - NOTE: the kernel stack may be at a different location in the child
240  *	process, and thus addresses of automatic variables may be invalid
241  *	after cpu_fork returns in the child process.  We do nothing here
242  *	after cpu_fork returns.
243  * - XXXCDC: we need a way for this to return a failure value rather
244  *   than just hang
245  */
246 void
uvm_fork(p1,p2,shared,stack,stacksize,func,arg)247 uvm_fork(p1, p2, shared, stack, stacksize, func, arg)
248 	struct proc *p1, *p2;
249 	boolean_t shared;
250 	void *stack;
251 	size_t stacksize;
252 	void (*func)(void *);
253 	void *arg;
254 {
255 	struct user *up = p2->p_addr;
256 	int rv;
257 
258 	if (shared == TRUE) {
259 		p2->p_vmspace = NULL;
260 		uvmspace_share(p1, p2);			/* share vmspace */
261 	} else
262 		p2->p_vmspace = uvmspace_fork(p1->p_vmspace); /* fork vmspace */
263 
264 	/*
265 	 * Wire down the U-area for the process, which contains the PCB
266 	 * and the kernel stack.  Wired state is stored in p->p_flag's
267 	 * P_INMEM bit rather than in the vm_map_entry's wired count
268 	 * to prevent kernel_map fragmentation.
269 	 *
270 	 * Note the kernel stack gets read/write accesses right off
271 	 * the bat.
272 	 */
273 	rv = uvm_fault_wire(kernel_map, (vaddr_t)up,
274 	    (vaddr_t)up + USPACE, VM_PROT_READ | VM_PROT_WRITE);
275 	if (rv != KERN_SUCCESS)
276 		panic("uvm_fork: uvm_fault_wire failed: %d", rv);
277 
278 	/*
279 	 * p_stats currently points at a field in the user struct.  Copy
280 	 * parts of p_stats, and zero out the rest.
281 	 */
282 	p2->p_stats = &up->u_stats;
283 	memset(&up->u_stats.pstat_startzero, 0,
284 	       ((caddr_t)&up->u_stats.pstat_endzero -
285 		(caddr_t)&up->u_stats.pstat_startzero));
286 	memcpy(&up->u_stats.pstat_startcopy, &p1->p_stats->pstat_startcopy,
287 	       ((caddr_t)&up->u_stats.pstat_endcopy -
288 		(caddr_t)&up->u_stats.pstat_startcopy));
289 
290 	/*
291 	 * cpu_fork() copy and update the pcb, and make the child ready
292 	 * to run.  If this is a normal user fork, the child will exit
293 	 * directly to user mode via child_return() on its first time
294 	 * slice and will not return here.  If this is a kernel thread,
295 	 * the specified entry point will be executed.
296 	 */
297 	cpu_fork(p1, p2, stack, stacksize, func, arg);
298 }
299 
300 /*
301  * uvm_exit: exit a virtual address space
302  *
303  * - the process passed to us is a dead (pre-zombie) process; we
304  *   are running on a different context now (the reaper).
305  * - we must run in a separate thread because freeing the vmspace
306  *   of the dead process may block.
307  */
308 void
uvm_exit(p)309 uvm_exit(p)
310 	struct proc *p;
311 {
312 	vaddr_t va = (vaddr_t)p->p_addr;
313 
314 	uvmspace_free(p->p_vmspace);
315 	p->p_flag &= ~P_INMEM;
316 	uvm_fault_unwire(kernel_map, va, va + USPACE);
317 	uvm_km_free(kernel_map, va, USPACE);
318 	p->p_addr = NULL;
319 }
320 
321 /*
322  * uvm_init_limit: init per-process VM limits
323  *
324  * - called for process 0 and then inherited by all others.
325  */
326 void
uvm_init_limits(p)327 uvm_init_limits(p)
328 	struct proc *p;
329 {
330 
331 	/*
332 	 * Set up the initial limits on process VM.  Set the maximum
333 	 * resident set size to be all of (reasonably) available memory.
334 	 * This causes any single, large process to start random page
335 	 * replacement once it fills memory.
336 	 */
337 
338 	p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ;
339 	p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ;
340 	p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ;
341 	p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ;
342 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free);
343 }
344 
345 #ifdef DEBUG
346 int	enableswap = 1;
347 int	swapdebug = 0;
348 #define	SDB_FOLLOW	1
349 #define SDB_SWAPIN	2
350 #define SDB_SWAPOUT	4
351 #endif
352 
353 /*
354  * uvm_swapin: swap in a process's u-area.
355  */
356 
357 void
uvm_swapin(p)358 uvm_swapin(p)
359 	struct proc *p;
360 {
361 	vaddr_t addr;
362 	int rv, s;
363 
364 	s = splstatclock();
365 	if (p->p_flag & P_SWAPIN) {
366 		splx(s);
367 		return;
368 	}
369 	p->p_flag |= P_SWAPIN;
370 	splx(s);
371 
372 	addr = (vaddr_t)p->p_addr;
373 	/* make P_INMEM true */
374 	if ((rv = uvm_fault_wire(kernel_map, addr, addr + USPACE,
375 	    VM_PROT_READ | VM_PROT_WRITE)) != KERN_SUCCESS)
376 		panic("uvm_swapin: uvm_fault_wire failed: %d", rv);
377 
378 	/*
379 	 * Some architectures need to be notified when the user area has
380 	 * moved to new physical page(s) (e.g.  see mips/mips/vm_machdep.c).
381 	 */
382 	cpu_swapin(p);
383 	s = splstatclock();
384 	if (p->p_stat == SRUN)
385 		setrunqueue(p);
386 	p->p_flag |= P_INMEM;
387 	p->p_flag &= ~P_SWAPIN;
388 	splx(s);
389 	p->p_swtime = 0;
390 	++uvmexp.swapins;
391 }
392 
393 /*
394  * uvm_scheduler: process zero main loop
395  *
396  * - attempt to swapin every swaped-out, runnable process in order of
397  *	priority.
398  * - if not enough memory, wake the pagedaemon and let it clear space.
399  */
400 
401 void
uvm_scheduler()402 uvm_scheduler()
403 {
404 	struct proc *p;
405 	int pri;
406 	struct proc *pp;
407 	int ppri;
408 
409 loop:
410 #ifdef DEBUG
411 	while (!enableswap)
412 		tsleep(&proc0, PVM, "noswap", 0);
413 #endif
414 	pp = NULL;		/* process to choose */
415 	ppri = INT_MIN;	/* its priority */
416 	LIST_FOREACH(p, &allproc, p_list) {
417 
418 		/* is it a runnable swapped out process? */
419 		if (p->p_stat == SRUN && (p->p_flag & P_INMEM) == 0) {
420 			pri = p->p_swtime + p->p_slptime -
421 			    (p->p_nice - NZERO) * 8;
422 			if (pri > ppri) {   /* higher priority?  remember it. */
423 				pp = p;
424 				ppri = pri;
425 			}
426 		}
427 	}
428 
429 #ifdef DEBUG
430 	if (swapdebug & SDB_FOLLOW)
431 		printf("scheduler: running, procp %p pri %d\n", pp, ppri);
432 #endif
433 	/*
434 	 * Nothing to do, back to sleep
435 	 */
436 	if ((p = pp) == NULL) {
437 		tsleep(&proc0, PVM, "scheduler", 0);
438 		goto loop;
439 	}
440 
441 	/*
442 	 * we have found swapped out process which we would like to bring
443 	 * back in.
444 	 *
445 	 * XXX: this part is really bogus because we could deadlock on memory
446 	 * despite our feeble check
447 	 */
448 	if (uvmexp.free > atop(USPACE)) {
449 #ifdef DEBUG
450 		if (swapdebug & SDB_SWAPIN)
451 			printf("swapin: pid %d(%s)@%p, pri %d free %d\n",
452 	     p->p_pid, p->p_comm, p->p_addr, ppri, uvmexp.free);
453 #endif
454 		uvm_swapin(p);
455 		goto loop;
456 	}
457 	/*
458 	 * not enough memory, jab the pageout daemon and wait til the coast
459 	 * is clear
460 	 */
461 #ifdef DEBUG
462 	if (swapdebug & SDB_FOLLOW)
463 		printf("scheduler: no room for pid %d(%s), free %d\n",
464 	   p->p_pid, p->p_comm, uvmexp.free);
465 #endif
466 	uvm_wait("schedpwait");
467 #ifdef DEBUG
468 	if (swapdebug & SDB_FOLLOW)
469 		printf("scheduler: room again, free %d\n", uvmexp.free);
470 #endif
471 	goto loop;
472 }
473 
474 /*
475  * swappable: is process "p" swappable?
476  */
477 
478 #define	swappable(p)							\
479 	(((p)->p_flag & (P_SYSTEM | P_INMEM | P_WEXIT)) == P_INMEM &&	\
480 	 (p)->p_holdcnt == 0)
481 
482 /*
483  * swapout_threads: find threads that can be swapped and unwire their
484  *	u-areas.
485  *
486  * - called by the pagedaemon
487  * - try and swap at least one processs
488  * - processes that are sleeping or stopped for maxslp or more seconds
489  *   are swapped... otherwise the longest-sleeping or stopped process
490  *   is swapped, otherwise the longest resident process...
491  */
492 void
uvm_swapout_threads()493 uvm_swapout_threads()
494 {
495 	struct proc *p;
496 	struct proc *outp, *outp2;
497 	int outpri, outpri2;
498 	int didswap = 0;
499 	extern int maxslp;
500 	/* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */
501 
502 #ifdef DEBUG
503 	if (!enableswap)
504 		return;
505 #endif
506 
507 	/*
508 	 * outp/outpri  : stop/sleep process with largest sleeptime < maxslp
509 	 * outp2/outpri2: the longest resident process (its swap time)
510 	 */
511 	outp = outp2 = NULL;
512 	outpri = outpri2 = 0;
513 	LIST_FOREACH(p, &allproc, p_list) {
514 		if (!swappable(p))
515 			continue;
516 		switch (p->p_stat) {
517 		case SRUN:
518 			if (p->p_swtime > outpri2) {
519 				outp2 = p;
520 				outpri2 = p->p_swtime;
521 			}
522 			continue;
523 
524 		case SSLEEP:
525 		case SSTOP:
526 			if (p->p_slptime >= maxslp) {
527 				uvm_swapout(p);
528 				didswap++;
529 			} else if (p->p_slptime > outpri) {
530 				outp = p;
531 				outpri = p->p_slptime;
532 			}
533 			continue;
534 		}
535 	}
536 
537 	/*
538 	 * If we didn't get rid of any real duds, toss out the next most
539 	 * likely sleeping/stopped or running candidate.  We only do this
540 	 * if we are real low on memory since we don't gain much by doing
541 	 * it (USPACE bytes).
542 	 */
543 	if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) {
544 		if ((p = outp) == NULL)
545 			p = outp2;
546 #ifdef DEBUG
547 		if (swapdebug & SDB_SWAPOUT)
548 			printf("swapout_threads: no duds, try procp %p\n", p);
549 #endif
550 		if (p)
551 			uvm_swapout(p);
552 	}
553 }
554 
555 /*
556  * uvm_swapout: swap out process "p"
557  *
558  * - currently "swapout" means "unwire U-area" and "pmap_collect()"
559  *   the pmap.
560  * - XXXCDC: should deactivate all process' private anonymous memory
561  */
562 
563 static void
uvm_swapout(p)564 uvm_swapout(p)
565 	struct proc *p;
566 {
567 	vaddr_t addr;
568 	int s;
569 
570 #ifdef DEBUG
571 	if (swapdebug & SDB_SWAPOUT)
572 		printf("swapout: pid %d(%s)@%p, stat %x pri %d free %d\n",
573 		    p->p_pid, p->p_comm, p->p_addr, p->p_stat,
574 		    p->p_slptime, uvmexp.free);
575 #endif
576 
577 	/*
578 	 * Mark it as (potentially) swapped out.
579 	 */
580 	s = splstatclock();
581 	if (!(p->p_flag & P_INMEM)) {
582 		splx(s);
583 		return;
584 	}
585 	p->p_flag &= ~P_INMEM;
586 	if (p->p_stat == SRUN)
587 		remrunqueue(p);
588 	splx(s);
589 	p->p_swtime = 0;
590 	++uvmexp.swapouts;
591 
592 	/*
593 	 * Do any machine-specific actions necessary before swapout.
594 	 * This can include saving floating point state, etc.
595 	 */
596 	cpu_swapout(p);
597 
598 	/*
599 	 * Unwire the to-be-swapped process's user struct and kernel stack.
600 	 */
601 	addr = (vaddr_t)p->p_addr;
602 	uvm_fault_unwire(kernel_map, addr, addr + USPACE); /* !P_INMEM */
603 	pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map));
604 }
605 
606