subr_rman.c revision 128172
1336815Sdim/*
2336815Sdim * Copyright 1998 Massachusetts Institute of Technology
3353358Sdim *
4353358Sdim * Permission to use, copy, modify, and distribute this software and
5353358Sdim * its documentation for any purpose and without fee is hereby
6336815Sdim * granted, provided that both the above copyright notice and this
7336815Sdim * permission notice appear in all copies, that both the above
8336815Sdim * copyright notice and this permission notice appear in all
9336815Sdim * supporting documentation, and that the name of M.I.T. not be used
10336815Sdim * in advertising or publicity pertaining to distribution of the
11336815Sdim * software without specific, written prior permission.  M.I.T. makes
12336815Sdim * no representations about the suitability of this software for any
13336815Sdim * purpose.  It is provided "as is" without express or implied
14336815Sdim * warranty.
15336815Sdim *
16336815Sdim * THIS SOFTWARE IS PROVIDED BY M.I.T. ``AS IS''.  M.I.T. DISCLAIMS
17336815Sdim * ALL EXPRESS OR IMPLIED WARRANTIES WITH REGARD TO THIS SOFTWARE,
18336815Sdim * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
19336815Sdim * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT
20336815Sdim * SHALL M.I.T. BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
21336815Sdim * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
22336815Sdim * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF
23336815Sdim * USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
24336815Sdim * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
25336815Sdim * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
26336815Sdim * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27336815Sdim * SUCH DAMAGE.
28336815Sdim */
29336815Sdim
30336815Sdim/*
31336815Sdim * The kernel resource manager.  This code is responsible for keeping track
32336815Sdim * of hardware resources which are apportioned out to various drivers.
33336815Sdim * It does not actually assign those resources, and it is not expected
34336815Sdim * that end-device drivers will call into this code directly.  Rather,
35336815Sdim * the code which implements the buses that those devices are attached to,
36336815Sdim * and the code which manages CPU resources, will call this code, and the
37336815Sdim * end-device drivers will make upcalls to that code to actually perform
38 * the allocation.
39 *
40 * There are two sorts of resources managed by this code.  The first is
41 * the more familiar array (RMAN_ARRAY) type; resources in this class
42 * consist of a sequence of individually-allocatable objects which have
43 * been numbered in some well-defined order.  Most of the resources
44 * are of this type, as it is the most familiar.  The second type is
45 * called a gauge (RMAN_GAUGE), and models fungible resources (i.e.,
46 * resources in which each instance is indistinguishable from every
47 * other instance).  The principal anticipated application of gauges
48 * is in the context of power consumption, where a bus may have a specific
49 * power budget which all attached devices share.  RMAN_GAUGE is not
50 * implemented yet.
51 *
52 * For array resources, we make one simplifying assumption: two clients
53 * sharing the same resource must use the same range of indices.  That
54 * is to say, sharing of overlapping-but-not-identical regions is not
55 * permitted.
56 */
57
58#include <sys/cdefs.h>
59__FBSDID("$FreeBSD: head/sys/kern/subr_rman.c 128172 2004-04-12 23:02:21Z imp $");
60
61#include <sys/param.h>
62#include <sys/systm.h>
63#include <sys/kernel.h>
64#include <sys/lock.h>
65#include <sys/malloc.h>
66#include <sys/mutex.h>
67#include <sys/bus.h>		/* XXX debugging */
68#include <machine/bus.h>
69#include <sys/rman.h>
70#include <sys/sysctl.h>
71
72int     rman_debug = 0;
73TUNABLE_INT("debug.rman_debug", &rman_debug);
74SYSCTL_INT(_debug, OID_AUTO, rman_debug, CTLFLAG_RW,
75    &rman_debug, 0, "rman debug");
76
77#define DPRINTF(params) if (rman_debug) printf params
78
79static MALLOC_DEFINE(M_RMAN, "rman", "Resource manager");
80
81struct	rman_head rman_head;
82static	struct mtx rman_mtx; /* mutex to protect rman_head */
83static	int int_rman_activate_resource(struct rman *rm, struct resource *r,
84				       struct resource **whohas);
85static	int int_rman_deactivate_resource(struct resource *r);
86static	int int_rman_release_resource(struct rman *rm, struct resource *r);
87
88int
89rman_init(struct rman *rm)
90{
91	static int once;
92
93	if (once == 0) {
94		once = 1;
95		TAILQ_INIT(&rman_head);
96		mtx_init(&rman_mtx, "rman head", NULL, MTX_DEF);
97	}
98
99	if (rm->rm_type == RMAN_UNINIT)
100		panic("rman_init");
101	if (rm->rm_type == RMAN_GAUGE)
102		panic("implement RMAN_GAUGE");
103
104	TAILQ_INIT(&rm->rm_list);
105	rm->rm_mtx = malloc(sizeof *rm->rm_mtx, M_RMAN, M_NOWAIT | M_ZERO);
106	if (rm->rm_mtx == 0)
107		return ENOMEM;
108	mtx_init(rm->rm_mtx, "rman", NULL, MTX_DEF);
109
110	mtx_lock(&rman_mtx);
111	TAILQ_INSERT_TAIL(&rman_head, rm, rm_link);
112	mtx_unlock(&rman_mtx);
113	return 0;
114}
115
116/*
117 * NB: this interface is not robust against programming errors which
118 * add multiple copies of the same region.
119 */
120int
121rman_manage_region(struct rman *rm, u_long start, u_long end)
122{
123	struct resource *r, *s;
124
125	r = malloc(sizeof *r, M_RMAN, M_NOWAIT | M_ZERO);
126	if (r == 0)
127		return ENOMEM;
128	r->r_start = start;
129	r->r_end = end;
130	r->r_rm = rm;
131
132	mtx_lock(rm->rm_mtx);
133	for (s = TAILQ_FIRST(&rm->rm_list);
134	     s && s->r_end < r->r_start;
135	     s = TAILQ_NEXT(s, r_link))
136		;
137
138	if (s == NULL) {
139		TAILQ_INSERT_TAIL(&rm->rm_list, r, r_link);
140	} else {
141		TAILQ_INSERT_BEFORE(s, r, r_link);
142	}
143
144	mtx_unlock(rm->rm_mtx);
145	return 0;
146}
147
148int
149rman_fini(struct rman *rm)
150{
151	struct resource *r;
152
153	mtx_lock(rm->rm_mtx);
154	TAILQ_FOREACH(r, &rm->rm_list, r_link) {
155		if (r->r_flags & RF_ALLOCATED) {
156			mtx_unlock(rm->rm_mtx);
157			return EBUSY;
158		}
159	}
160
161	/*
162	 * There really should only be one of these if we are in this
163	 * state and the code is working properly, but it can't hurt.
164	 */
165	while (!TAILQ_EMPTY(&rm->rm_list)) {
166		r = TAILQ_FIRST(&rm->rm_list);
167		TAILQ_REMOVE(&rm->rm_list, r, r_link);
168		free(r, M_RMAN);
169	}
170	mtx_unlock(rm->rm_mtx);
171	mtx_lock(&rman_mtx);
172	TAILQ_REMOVE(&rman_head, rm, rm_link);
173	mtx_unlock(&rman_mtx);
174	mtx_destroy(rm->rm_mtx);
175	free(rm->rm_mtx, M_RMAN);
176
177	return 0;
178}
179
180struct resource *
181rman_reserve_resource_bound(struct rman *rm, u_long start, u_long end,
182		      u_long count, u_long bound,  u_int flags,
183		      struct device *dev)
184{
185	u_int	want_activate;
186	struct	resource *r, *s, *rv;
187	u_long	rstart, rend, amask, bmask;
188
189	rv = 0;
190
191	DPRINTF(("rman_reserve_resource: <%s> request: [%#lx, %#lx], length "
192	       "%#lx, flags %u, device %s\n", rm->rm_descr, start, end, count,
193	       flags, dev == NULL ? "<null>" : device_get_nameunit(dev)));
194	want_activate = (flags & RF_ACTIVE);
195	flags &= ~RF_ACTIVE;
196
197	mtx_lock(rm->rm_mtx);
198
199	for (r = TAILQ_FIRST(&rm->rm_list);
200	     r && r->r_end < start;
201	     r = TAILQ_NEXT(r, r_link))
202		;
203
204	if (r == NULL) {
205		DPRINTF(("could not find a region\n"));
206		goto out;
207	}
208
209	amask = (1ul << RF_ALIGNMENT(flags)) - 1;
210	/* If bound is 0, bmask will also be 0 */
211	bmask = ~(bound - 1);
212	/*
213	 * First try to find an acceptable totally-unshared region.
214	 */
215	for (s = r; s; s = TAILQ_NEXT(s, r_link)) {
216		DPRINTF(("considering [%#lx, %#lx]\n", s->r_start, s->r_end));
217		if (s->r_start > end) {
218			DPRINTF(("s->r_start (%#lx) > end (%#lx)\n", s->r_start, end));
219			break;
220		}
221		if (s->r_flags & RF_ALLOCATED) {
222			DPRINTF(("region is allocated\n"));
223			continue;
224		}
225		rstart = ulmax(s->r_start, start);
226		/*
227		 * Try to find a region by adjusting to boundary and alignment
228		 * until both conditions are satisfied. This is not an optimal
229		 * algorithm, but in most cases it isn't really bad, either.
230		 */
231		do {
232			rstart = (rstart + amask) & ~amask;
233			if (((rstart ^ (rstart + count - 1)) & bmask) != 0)
234				rstart += bound - (rstart & ~bmask);
235		} while ((rstart & amask) != 0 && rstart < end &&
236		    rstart < s->r_end);
237		rend = ulmin(s->r_end, ulmax(rstart + count - 1, end));
238		if (rstart > rend) {
239			DPRINTF(("adjusted start exceeds end\n"));
240			continue;
241		}
242		DPRINTF(("truncated region: [%#lx, %#lx]; size %#lx (requested %#lx)\n",
243		       rstart, rend, (rend - rstart + 1), count));
244
245		if ((rend - rstart + 1) >= count) {
246			DPRINTF(("candidate region: [%#lx, %#lx], size %#lx\n",
247			       rend, rstart, (rend - rstart + 1)));
248			if ((s->r_end - s->r_start + 1) == count) {
249				DPRINTF(("candidate region is entire chunk\n"));
250				rv = s;
251				rv->r_flags |= RF_ALLOCATED | flags;
252				rv->r_dev = dev;
253				goto out;
254			}
255
256			/*
257			 * If s->r_start < rstart and
258			 *    s->r_end > rstart + count - 1, then
259			 * we need to split the region into three pieces
260			 * (the middle one will get returned to the user).
261			 * Otherwise, we are allocating at either the
262			 * beginning or the end of s, so we only need to
263			 * split it in two.  The first case requires
264			 * two new allocations; the second requires but one.
265			 */
266			rv = malloc(sizeof *rv, M_RMAN, M_NOWAIT | M_ZERO);
267			if (rv == 0)
268				goto out;
269			rv->r_start = rstart;
270			rv->r_end = rstart + count - 1;
271			rv->r_flags = flags | RF_ALLOCATED;
272			rv->r_dev = dev;
273			rv->r_rm = rm;
274
275			if (s->r_start < rv->r_start && s->r_end > rv->r_end) {
276				DPRINTF(("splitting region in three parts: "
277				       "[%#lx, %#lx]; [%#lx, %#lx]; [%#lx, %#lx]\n",
278				       s->r_start, rv->r_start - 1,
279				       rv->r_start, rv->r_end,
280				       rv->r_end + 1, s->r_end));
281				/*
282				 * We are allocating in the middle.
283				 */
284				r = malloc(sizeof *r, M_RMAN, M_NOWAIT|M_ZERO);
285				if (r == 0) {
286					free(rv, M_RMAN);
287					rv = 0;
288					goto out;
289				}
290				r->r_start = rv->r_end + 1;
291				r->r_end = s->r_end;
292				r->r_flags = s->r_flags;
293				r->r_rm = rm;
294				s->r_end = rv->r_start - 1;
295				TAILQ_INSERT_AFTER(&rm->rm_list, s, rv,
296						     r_link);
297				TAILQ_INSERT_AFTER(&rm->rm_list, rv, r,
298						     r_link);
299			} else if (s->r_start == rv->r_start) {
300				DPRINTF(("allocating from the beginning\n"));
301				/*
302				 * We are allocating at the beginning.
303				 */
304				s->r_start = rv->r_end + 1;
305				TAILQ_INSERT_BEFORE(s, rv, r_link);
306			} else {
307				DPRINTF(("allocating at the end\n"));
308				/*
309				 * We are allocating at the end.
310				 */
311				s->r_end = rv->r_start - 1;
312				TAILQ_INSERT_AFTER(&rm->rm_list, s, rv,
313						     r_link);
314			}
315			goto out;
316		}
317	}
318
319	/*
320	 * Now find an acceptable shared region, if the client's requirements
321	 * allow sharing.  By our implementation restriction, a candidate
322	 * region must match exactly by both size and sharing type in order
323	 * to be considered compatible with the client's request.  (The
324	 * former restriction could probably be lifted without too much
325	 * additional work, but this does not seem warranted.)
326	 */
327	DPRINTF(("no unshared regions found\n"));
328	if ((flags & (RF_SHAREABLE | RF_TIMESHARE)) == 0)
329		goto out;
330
331	for (s = r; s; s = TAILQ_NEXT(s, r_link)) {
332		if (s->r_start > end)
333			break;
334		if ((s->r_flags & flags) != flags)
335			continue;
336		rstart = ulmax(s->r_start, start);
337		rend = ulmin(s->r_end, ulmax(start + count - 1, end));
338		if (s->r_start >= start && s->r_end <= end
339		    && (s->r_end - s->r_start + 1) == count &&
340		    (s->r_start & amask) == 0 &&
341		    ((s->r_start ^ s->r_end) & bmask) == 0) {
342			rv = malloc(sizeof *rv, M_RMAN, M_NOWAIT | M_ZERO);
343			if (rv == 0)
344				goto out;
345			rv->r_start = s->r_start;
346			rv->r_end = s->r_end;
347			rv->r_flags = s->r_flags &
348				(RF_ALLOCATED | RF_SHAREABLE | RF_TIMESHARE);
349			rv->r_dev = dev;
350			rv->r_rm = rm;
351			if (s->r_sharehead == 0) {
352				s->r_sharehead = malloc(sizeof *s->r_sharehead,
353						M_RMAN, M_NOWAIT | M_ZERO);
354				if (s->r_sharehead == 0) {
355					free(rv, M_RMAN);
356					rv = 0;
357					goto out;
358				}
359				LIST_INIT(s->r_sharehead);
360				LIST_INSERT_HEAD(s->r_sharehead, s,
361						 r_sharelink);
362				s->r_flags |= RF_FIRSTSHARE;
363			}
364			rv->r_sharehead = s->r_sharehead;
365			LIST_INSERT_HEAD(s->r_sharehead, rv, r_sharelink);
366			goto out;
367		}
368	}
369
370	/*
371	 * We couldn't find anything.
372	 */
373out:
374	/*
375	 * If the user specified RF_ACTIVE in the initial flags,
376	 * which is reflected in `want_activate', we attempt to atomically
377	 * activate the resource.  If this fails, we release the resource
378	 * and indicate overall failure.  (This behavior probably doesn't
379	 * make sense for RF_TIMESHARE-type resources.)
380	 */
381	if (rv && want_activate) {
382		struct resource *whohas;
383		if (int_rman_activate_resource(rm, rv, &whohas)) {
384			int_rman_release_resource(rm, rv);
385			rv = 0;
386		}
387	}
388
389	mtx_unlock(rm->rm_mtx);
390	return (rv);
391}
392
393struct resource *
394rman_reserve_resource(struct rman *rm, u_long start, u_long end, u_long count,
395		      u_int flags, struct device *dev)
396{
397
398	return (rman_reserve_resource_bound(rm, start, end, count, 0, flags,
399	    dev));
400}
401
402static int
403int_rman_activate_resource(struct rman *rm, struct resource *r,
404			   struct resource **whohas)
405{
406	struct resource *s;
407	int ok;
408
409	/*
410	 * If we are not timesharing, then there is nothing much to do.
411	 * If we already have the resource, then there is nothing at all to do.
412	 * If we are not on a sharing list with anybody else, then there is
413	 * little to do.
414	 */
415	if ((r->r_flags & RF_TIMESHARE) == 0
416	    || (r->r_flags & RF_ACTIVE) != 0
417	    || r->r_sharehead == 0) {
418		r->r_flags |= RF_ACTIVE;
419		return 0;
420	}
421
422	ok = 1;
423	for (s = LIST_FIRST(r->r_sharehead); s && ok;
424	     s = LIST_NEXT(s, r_sharelink)) {
425		if ((s->r_flags & RF_ACTIVE) != 0) {
426			ok = 0;
427			*whohas = s;
428		}
429	}
430	if (ok) {
431		r->r_flags |= RF_ACTIVE;
432		return 0;
433	}
434	return EBUSY;
435}
436
437int
438rman_activate_resource(struct resource *r)
439{
440	int rv;
441	struct resource *whohas;
442	struct rman *rm;
443
444	rm = r->r_rm;
445	mtx_lock(rm->rm_mtx);
446	rv = int_rman_activate_resource(rm, r, &whohas);
447	mtx_unlock(rm->rm_mtx);
448	return rv;
449}
450
451int
452rman_await_resource(struct resource *r, int pri, int timo)
453{
454	int	rv;
455	struct	resource *whohas;
456	struct	rman *rm;
457
458	rm = r->r_rm;
459	mtx_lock(rm->rm_mtx);
460	for (;;) {
461		rv = int_rman_activate_resource(rm, r, &whohas);
462		if (rv != EBUSY)
463			return (rv);	/* returns with mutex held */
464
465		if (r->r_sharehead == 0)
466			panic("rman_await_resource");
467		whohas->r_flags |= RF_WANTED;
468		rv = msleep(r->r_sharehead, rm->rm_mtx, pri, "rmwait", timo);
469		if (rv) {
470			mtx_unlock(rm->rm_mtx);
471			return (rv);
472		}
473	}
474}
475
476static int
477int_rman_deactivate_resource(struct resource *r)
478{
479
480	r->r_flags &= ~RF_ACTIVE;
481	if (r->r_flags & RF_WANTED) {
482		r->r_flags &= ~RF_WANTED;
483		wakeup(r->r_sharehead);
484	}
485	return 0;
486}
487
488int
489rman_deactivate_resource(struct resource *r)
490{
491	struct	rman *rm;
492
493	rm = r->r_rm;
494	mtx_lock(rm->rm_mtx);
495	int_rman_deactivate_resource(r);
496	mtx_unlock(rm->rm_mtx);
497	return 0;
498}
499
500static int
501int_rman_release_resource(struct rman *rm, struct resource *r)
502{
503	struct	resource *s, *t;
504
505	if (r->r_flags & RF_ACTIVE)
506		int_rman_deactivate_resource(r);
507
508	/*
509	 * Check for a sharing list first.  If there is one, then we don't
510	 * have to think as hard.
511	 */
512	if (r->r_sharehead) {
513		/*
514		 * If a sharing list exists, then we know there are at
515		 * least two sharers.
516		 *
517		 * If we are in the main circleq, appoint someone else.
518		 */
519		LIST_REMOVE(r, r_sharelink);
520		s = LIST_FIRST(r->r_sharehead);
521		if (r->r_flags & RF_FIRSTSHARE) {
522			s->r_flags |= RF_FIRSTSHARE;
523			TAILQ_INSERT_BEFORE(r, s, r_link);
524			TAILQ_REMOVE(&rm->rm_list, r, r_link);
525		}
526
527		/*
528		 * Make sure that the sharing list goes away completely
529		 * if the resource is no longer being shared at all.
530		 */
531		if (LIST_NEXT(s, r_sharelink) == 0) {
532			free(s->r_sharehead, M_RMAN);
533			s->r_sharehead = 0;
534			s->r_flags &= ~RF_FIRSTSHARE;
535		}
536		goto out;
537	}
538
539	/*
540	 * Look at the adjacent resources in the list and see if our
541	 * segment can be merged with any of them.
542	 */
543	s = TAILQ_PREV(r, resource_head, r_link);
544	t = TAILQ_NEXT(r, r_link);
545
546	if (s != NULL && (s->r_flags & RF_ALLOCATED) == 0
547	    && t != NULL && (t->r_flags & RF_ALLOCATED) == 0) {
548		/*
549		 * Merge all three segments.
550		 */
551		s->r_end = t->r_end;
552		TAILQ_REMOVE(&rm->rm_list, r, r_link);
553		TAILQ_REMOVE(&rm->rm_list, t, r_link);
554		free(t, M_RMAN);
555	} else if (s != NULL && (s->r_flags & RF_ALLOCATED) == 0) {
556		/*
557		 * Merge previous segment with ours.
558		 */
559		s->r_end = r->r_end;
560		TAILQ_REMOVE(&rm->rm_list, r, r_link);
561	} else if (t != NULL && (t->r_flags & RF_ALLOCATED) == 0) {
562		/*
563		 * Merge next segment with ours.
564		 */
565		t->r_start = r->r_start;
566		TAILQ_REMOVE(&rm->rm_list, r, r_link);
567	} else {
568		/*
569		 * At this point, we know there is nothing we
570		 * can potentially merge with, because on each
571		 * side, there is either nothing there or what is
572		 * there is still allocated.  In that case, we don't
573		 * want to remove r from the list; we simply want to
574		 * change it to an unallocated region and return
575		 * without freeing anything.
576		 */
577		r->r_flags &= ~RF_ALLOCATED;
578		return 0;
579	}
580
581out:
582	free(r, M_RMAN);
583	return 0;
584}
585
586int
587rman_release_resource(struct resource *r)
588{
589	int	rv;
590	struct	rman *rm = r->r_rm;
591
592	mtx_lock(rm->rm_mtx);
593	rv = int_rman_release_resource(rm, r);
594	mtx_unlock(rm->rm_mtx);
595	return (rv);
596}
597
598uint32_t
599rman_make_alignment_flags(uint32_t size)
600{
601	int	i;
602
603	/*
604	 * Find the hightest bit set, and add one if more than one bit
605	 * set.  We're effectively computing the ceil(log2(size)) here.
606	 */
607	for (i = 31; i > 0; i--)
608		if ((1 << i) & size)
609			break;
610	if (~(1 << i) & size)
611		i++;
612
613	return(RF_ALIGNMENT_LOG2(i));
614}
615
616u_long
617rman_get_start(struct resource *r)
618{
619	return (r->r_start);
620}
621
622u_long
623rman_get_end(struct resource *r)
624{
625	return (r->r_end);
626}
627
628u_long
629rman_get_size(struct resource *r)
630{
631	return (r->r_end - r->r_start + 1);
632}
633
634u_int
635rman_get_flags(struct resource *r)
636{
637	return (r->r_flags);
638}
639
640void
641rman_set_virtual(struct resource *r, void *v)
642{
643	r->r_virtual = v;
644}
645
646void *
647rman_get_virtual(struct resource *r)
648{
649	return (r->r_virtual);
650}
651
652void
653rman_set_bustag(struct resource *r, bus_space_tag_t t)
654{
655	r->r_bustag = t;
656}
657
658bus_space_tag_t
659rman_get_bustag(struct resource *r)
660{
661	return (r->r_bustag);
662}
663
664void
665rman_set_bushandle(struct resource *r, bus_space_handle_t h)
666{
667	r->r_bushandle = h;
668}
669
670bus_space_handle_t
671rman_get_bushandle(struct resource *r)
672{
673	return (r->r_bushandle);
674}
675
676void
677rman_set_rid(struct resource *r, int rid)
678{
679	r->r_rid = rid;
680}
681
682int
683rman_get_rid(struct resource *r)
684{
685	return (r->r_rid);
686}
687
688struct device *
689rman_get_device(struct resource *r)
690{
691	return (r->r_dev);
692}
693