subr_rman.c revision 131414
1/*
2 * Copyright 1998 Massachusetts Institute of Technology
3 *
4 * Permission to use, copy, modify, and distribute this software and
5 * its documentation for any purpose and without fee is hereby
6 * granted, provided that both the above copyright notice and this
7 * permission notice appear in all copies, that both the above
8 * copyright notice and this permission notice appear in all
9 * supporting documentation, and that the name of M.I.T. not be used
10 * in advertising or publicity pertaining to distribution of the
11 * software without specific, written prior permission.  M.I.T. makes
12 * no representations about the suitability of this software for any
13 * purpose.  It is provided "as is" without express or implied
14 * warranty.
15 *
16 * THIS SOFTWARE IS PROVIDED BY M.I.T. ``AS IS''.  M.I.T. DISCLAIMS
17 * ALL EXPRESS OR IMPLIED WARRANTIES WITH REGARD TO THIS SOFTWARE,
18 * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
19 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT
20 * SHALL M.I.T. BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF
23 * USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
24 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
25 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
26 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30/*
31 * The kernel resource manager.  This code is responsible for keeping track
32 * of hardware resources which are apportioned out to various drivers.
33 * It does not actually assign those resources, and it is not expected
34 * that end-device drivers will call into this code directly.  Rather,
35 * the code which implements the buses that those devices are attached to,
36 * and the code which manages CPU resources, will call this code, and the
37 * 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 131414 2004-07-01 16:22:10Z imp $");
60
61#define __RMAN_RESOURCE_VISIBLE
62#include <sys/param.h>
63#include <sys/systm.h>
64#include <sys/kernel.h>
65#include <sys/lock.h>
66#include <sys/malloc.h>
67#include <sys/mutex.h>
68#include <sys/bus.h>		/* XXX debugging */
69#include <machine/bus.h>
70#include <sys/rman.h>
71#include <sys/sysctl.h>
72
73int     rman_debug = 0;
74TUNABLE_INT("debug.rman_debug", &rman_debug);
75SYSCTL_INT(_debug, OID_AUTO, rman_debug, CTLFLAG_RW,
76    &rman_debug, 0, "rman debug");
77
78#define DPRINTF(params) if (rman_debug) printf params
79
80static MALLOC_DEFINE(M_RMAN, "rman", "Resource manager");
81
82struct	rman_head rman_head;
83static	struct mtx rman_mtx; /* mutex to protect rman_head */
84static	int int_rman_activate_resource(struct rman *rm, struct resource *r,
85				       struct resource **whohas);
86static	int int_rman_deactivate_resource(struct resource *r);
87static	int int_rman_release_resource(struct rman *rm, struct resource *r);
88
89int
90rman_init(struct rman *rm)
91{
92	static int once;
93
94	if (once == 0) {
95		once = 1;
96		TAILQ_INIT(&rman_head);
97		mtx_init(&rman_mtx, "rman head", NULL, MTX_DEF);
98	}
99
100	if (rm->rm_type == RMAN_UNINIT)
101		panic("rman_init");
102	if (rm->rm_type == RMAN_GAUGE)
103		panic("implement RMAN_GAUGE");
104
105	TAILQ_INIT(&rm->rm_list);
106	rm->rm_mtx = malloc(sizeof *rm->rm_mtx, M_RMAN, M_NOWAIT | M_ZERO);
107	if (rm->rm_mtx == 0)
108		return ENOMEM;
109	mtx_init(rm->rm_mtx, "rman", NULL, MTX_DEF);
110
111	mtx_lock(&rman_mtx);
112	TAILQ_INSERT_TAIL(&rman_head, rm, rm_link);
113	mtx_unlock(&rman_mtx);
114	return 0;
115}
116
117/*
118 * NB: this interface is not robust against programming errors which
119 * add multiple copies of the same region.
120 */
121int
122rman_manage_region(struct rman *rm, u_long start, u_long end)
123{
124	struct resource *r, *s;
125
126	r = malloc(sizeof *r, M_RMAN, M_NOWAIT | M_ZERO);
127	if (r == 0)
128		return ENOMEM;
129	r->r_start = start;
130	r->r_end = end;
131	r->r_rm = rm;
132
133	mtx_lock(rm->rm_mtx);
134	for (s = TAILQ_FIRST(&rm->rm_list);
135	     s && s->r_end < r->r_start;
136	     s = TAILQ_NEXT(s, r_link))
137		;
138
139	if (s == NULL) {
140		TAILQ_INSERT_TAIL(&rm->rm_list, r, r_link);
141	} else {
142		TAILQ_INSERT_BEFORE(s, r, r_link);
143	}
144
145	mtx_unlock(rm->rm_mtx);
146	return 0;
147}
148
149int
150rman_fini(struct rman *rm)
151{
152	struct resource *r;
153
154	mtx_lock(rm->rm_mtx);
155	TAILQ_FOREACH(r, &rm->rm_list, r_link) {
156		if (r->r_flags & RF_ALLOCATED) {
157			mtx_unlock(rm->rm_mtx);
158			return EBUSY;
159		}
160	}
161
162	/*
163	 * There really should only be one of these if we are in this
164	 * state and the code is working properly, but it can't hurt.
165	 */
166	while (!TAILQ_EMPTY(&rm->rm_list)) {
167		r = TAILQ_FIRST(&rm->rm_list);
168		TAILQ_REMOVE(&rm->rm_list, r, r_link);
169		free(r, M_RMAN);
170	}
171	mtx_unlock(rm->rm_mtx);
172	mtx_lock(&rman_mtx);
173	TAILQ_REMOVE(&rman_head, rm, rm_link);
174	mtx_unlock(&rman_mtx);
175	mtx_destroy(rm->rm_mtx);
176	free(rm->rm_mtx, M_RMAN);
177
178	return 0;
179}
180
181struct resource *
182rman_reserve_resource_bound(struct rman *rm, u_long start, u_long end,
183		      u_long count, u_long bound,  u_int flags,
184		      struct device *dev)
185{
186	u_int	want_activate;
187	struct	resource *r, *s, *rv;
188	u_long	rstart, rend, amask, bmask;
189
190	rv = 0;
191
192	DPRINTF(("rman_reserve_resource: <%s> request: [%#lx, %#lx], length "
193	       "%#lx, flags %u, device %s\n", rm->rm_descr, start, end, count,
194	       flags, dev == NULL ? "<null>" : device_get_nameunit(dev)));
195	want_activate = (flags & RF_ACTIVE);
196	flags &= ~RF_ACTIVE;
197
198	mtx_lock(rm->rm_mtx);
199
200	for (r = TAILQ_FIRST(&rm->rm_list);
201	     r && r->r_end < start;
202	     r = TAILQ_NEXT(r, r_link))
203		;
204
205	if (r == NULL) {
206		DPRINTF(("could not find a region\n"));
207		goto out;
208	}
209
210	amask = (1ul << RF_ALIGNMENT(flags)) - 1;
211	/* If bound is 0, bmask will also be 0 */
212	bmask = ~(bound - 1);
213	/*
214	 * First try to find an acceptable totally-unshared region.
215	 */
216	for (s = r; s; s = TAILQ_NEXT(s, r_link)) {
217		DPRINTF(("considering [%#lx, %#lx]\n", s->r_start, s->r_end));
218		if (s->r_start > end) {
219			DPRINTF(("s->r_start (%#lx) > end (%#lx)\n", s->r_start, end));
220			break;
221		}
222		if (s->r_flags & RF_ALLOCATED) {
223			DPRINTF(("region is allocated\n"));
224			continue;
225		}
226		rstart = ulmax(s->r_start, start);
227		/*
228		 * Try to find a region by adjusting to boundary and alignment
229		 * until both conditions are satisfied. This is not an optimal
230		 * algorithm, but in most cases it isn't really bad, either.
231		 */
232		do {
233			rstart = (rstart + amask) & ~amask;
234			if (((rstart ^ (rstart + count - 1)) & bmask) != 0)
235				rstart += bound - (rstart & ~bmask);
236		} while ((rstart & amask) != 0 && rstart < end &&
237		    rstart < s->r_end);
238		rend = ulmin(s->r_end, ulmax(rstart + count - 1, end));
239		if (rstart > rend) {
240			DPRINTF(("adjusted start exceeds end\n"));
241			continue;
242		}
243		DPRINTF(("truncated region: [%#lx, %#lx]; size %#lx (requested %#lx)\n",
244		       rstart, rend, (rend - rstart + 1), count));
245
246		if ((rend - rstart + 1) >= count) {
247			DPRINTF(("candidate region: [%#lx, %#lx], size %#lx\n",
248			       rend, rstart, (rend - rstart + 1)));
249			if ((s->r_end - s->r_start + 1) == count) {
250				DPRINTF(("candidate region is entire chunk\n"));
251				rv = s;
252				rv->r_flags |= RF_ALLOCATED | flags;
253				rv->r_dev = dev;
254				goto out;
255			}
256
257			/*
258			 * If s->r_start < rstart and
259			 *    s->r_end > rstart + count - 1, then
260			 * we need to split the region into three pieces
261			 * (the middle one will get returned to the user).
262			 * Otherwise, we are allocating at either the
263			 * beginning or the end of s, so we only need to
264			 * split it in two.  The first case requires
265			 * two new allocations; the second requires but one.
266			 */
267			rv = malloc(sizeof *rv, M_RMAN, M_NOWAIT | M_ZERO);
268			if (rv == 0)
269				goto out;
270			rv->r_start = rstart;
271			rv->r_end = rstart + count - 1;
272			rv->r_flags = flags | RF_ALLOCATED;
273			rv->r_dev = dev;
274			rv->r_rm = rm;
275
276			if (s->r_start < rv->r_start && s->r_end > rv->r_end) {
277				DPRINTF(("splitting region in three parts: "
278				       "[%#lx, %#lx]; [%#lx, %#lx]; [%#lx, %#lx]\n",
279				       s->r_start, rv->r_start - 1,
280				       rv->r_start, rv->r_end,
281				       rv->r_end + 1, s->r_end));
282				/*
283				 * We are allocating in the middle.
284				 */
285				r = malloc(sizeof *r, M_RMAN, M_NOWAIT|M_ZERO);
286				if (r == 0) {
287					free(rv, M_RMAN);
288					rv = 0;
289					goto out;
290				}
291				r->r_start = rv->r_end + 1;
292				r->r_end = s->r_end;
293				r->r_flags = s->r_flags;
294				r->r_rm = rm;
295				s->r_end = rv->r_start - 1;
296				TAILQ_INSERT_AFTER(&rm->rm_list, s, rv,
297						     r_link);
298				TAILQ_INSERT_AFTER(&rm->rm_list, rv, r,
299						     r_link);
300			} else if (s->r_start == rv->r_start) {
301				DPRINTF(("allocating from the beginning\n"));
302				/*
303				 * We are allocating at the beginning.
304				 */
305				s->r_start = rv->r_end + 1;
306				TAILQ_INSERT_BEFORE(s, rv, r_link);
307			} else {
308				DPRINTF(("allocating at the end\n"));
309				/*
310				 * We are allocating at the end.
311				 */
312				s->r_end = rv->r_start - 1;
313				TAILQ_INSERT_AFTER(&rm->rm_list, s, rv,
314						     r_link);
315			}
316			goto out;
317		}
318	}
319
320	/*
321	 * Now find an acceptable shared region, if the client's requirements
322	 * allow sharing.  By our implementation restriction, a candidate
323	 * region must match exactly by both size and sharing type in order
324	 * to be considered compatible with the client's request.  (The
325	 * former restriction could probably be lifted without too much
326	 * additional work, but this does not seem warranted.)
327	 */
328	DPRINTF(("no unshared regions found\n"));
329	if ((flags & (RF_SHAREABLE | RF_TIMESHARE)) == 0)
330		goto out;
331
332	for (s = r; s; s = TAILQ_NEXT(s, r_link)) {
333		if (s->r_start > end)
334			break;
335		if ((s->r_flags & flags) != flags)
336			continue;
337		rstart = ulmax(s->r_start, start);
338		rend = ulmin(s->r_end, ulmax(start + count - 1, end));
339		if (s->r_start >= start && s->r_end <= end
340		    && (s->r_end - s->r_start + 1) == count &&
341		    (s->r_start & amask) == 0 &&
342		    ((s->r_start ^ s->r_end) & bmask) == 0) {
343			rv = malloc(sizeof *rv, M_RMAN, M_NOWAIT | M_ZERO);
344			if (rv == 0)
345				goto out;
346			rv->r_start = s->r_start;
347			rv->r_end = s->r_end;
348			rv->r_flags = s->r_flags &
349				(RF_ALLOCATED | RF_SHAREABLE | RF_TIMESHARE);
350			rv->r_dev = dev;
351			rv->r_rm = rm;
352			if (s->r_sharehead == 0) {
353				s->r_sharehead = malloc(sizeof *s->r_sharehead,
354						M_RMAN, M_NOWAIT | M_ZERO);
355				if (s->r_sharehead == 0) {
356					free(rv, M_RMAN);
357					rv = 0;
358					goto out;
359				}
360				LIST_INIT(s->r_sharehead);
361				LIST_INSERT_HEAD(s->r_sharehead, s,
362						 r_sharelink);
363				s->r_flags |= RF_FIRSTSHARE;
364			}
365			rv->r_sharehead = s->r_sharehead;
366			LIST_INSERT_HEAD(s->r_sharehead, rv, r_sharelink);
367			goto out;
368		}
369	}
370
371	/*
372	 * We couldn't find anything.
373	 */
374out:
375	/*
376	 * If the user specified RF_ACTIVE in the initial flags,
377	 * which is reflected in `want_activate', we attempt to atomically
378	 * activate the resource.  If this fails, we release the resource
379	 * and indicate overall failure.  (This behavior probably doesn't
380	 * make sense for RF_TIMESHARE-type resources.)
381	 */
382	if (rv && want_activate) {
383		struct resource *whohas;
384		if (int_rman_activate_resource(rm, rv, &whohas)) {
385			int_rman_release_resource(rm, rv);
386			rv = 0;
387		}
388	}
389
390	mtx_unlock(rm->rm_mtx);
391	return (rv);
392}
393
394struct resource *
395rman_reserve_resource(struct rman *rm, u_long start, u_long end, u_long count,
396		      u_int flags, struct device *dev)
397{
398
399	return (rman_reserve_resource_bound(rm, start, end, count, 0, flags,
400	    dev));
401}
402
403static int
404int_rman_activate_resource(struct rman *rm, struct resource *r,
405			   struct resource **whohas)
406{
407	struct resource *s;
408	int ok;
409
410	/*
411	 * If we are not timesharing, then there is nothing much to do.
412	 * If we already have the resource, then there is nothing at all to do.
413	 * If we are not on a sharing list with anybody else, then there is
414	 * little to do.
415	 */
416	if ((r->r_flags & RF_TIMESHARE) == 0
417	    || (r->r_flags & RF_ACTIVE) != 0
418	    || r->r_sharehead == 0) {
419		r->r_flags |= RF_ACTIVE;
420		return 0;
421	}
422
423	ok = 1;
424	for (s = LIST_FIRST(r->r_sharehead); s && ok;
425	     s = LIST_NEXT(s, r_sharelink)) {
426		if ((s->r_flags & RF_ACTIVE) != 0) {
427			ok = 0;
428			*whohas = s;
429		}
430	}
431	if (ok) {
432		r->r_flags |= RF_ACTIVE;
433		return 0;
434	}
435	return EBUSY;
436}
437
438int
439rman_activate_resource(struct resource *r)
440{
441	int rv;
442	struct resource *whohas;
443	struct rman *rm;
444
445	rm = r->r_rm;
446	mtx_lock(rm->rm_mtx);
447	rv = int_rman_activate_resource(rm, r, &whohas);
448	mtx_unlock(rm->rm_mtx);
449	return rv;
450}
451
452int
453rman_await_resource(struct resource *r, int pri, int timo)
454{
455	int	rv;
456	struct	resource *whohas;
457	struct	rman *rm;
458
459	rm = r->r_rm;
460	mtx_lock(rm->rm_mtx);
461	for (;;) {
462		rv = int_rman_activate_resource(rm, r, &whohas);
463		if (rv != EBUSY)
464			return (rv);	/* returns with mutex held */
465
466		if (r->r_sharehead == 0)
467			panic("rman_await_resource");
468		whohas->r_flags |= RF_WANTED;
469		rv = msleep(r->r_sharehead, rm->rm_mtx, pri, "rmwait", timo);
470		if (rv) {
471			mtx_unlock(rm->rm_mtx);
472			return (rv);
473		}
474	}
475}
476
477static int
478int_rman_deactivate_resource(struct resource *r)
479{
480
481	r->r_flags &= ~RF_ACTIVE;
482	if (r->r_flags & RF_WANTED) {
483		r->r_flags &= ~RF_WANTED;
484		wakeup(r->r_sharehead);
485	}
486	return 0;
487}
488
489int
490rman_deactivate_resource(struct resource *r)
491{
492	struct	rman *rm;
493
494	rm = r->r_rm;
495	mtx_lock(rm->rm_mtx);
496	int_rman_deactivate_resource(r);
497	mtx_unlock(rm->rm_mtx);
498	return 0;
499}
500
501static int
502int_rman_release_resource(struct rman *rm, struct resource *r)
503{
504	struct	resource *s, *t;
505
506	if (r->r_flags & RF_ACTIVE)
507		int_rman_deactivate_resource(r);
508
509	/*
510	 * Check for a sharing list first.  If there is one, then we don't
511	 * have to think as hard.
512	 */
513	if (r->r_sharehead) {
514		/*
515		 * If a sharing list exists, then we know there are at
516		 * least two sharers.
517		 *
518		 * If we are in the main circleq, appoint someone else.
519		 */
520		LIST_REMOVE(r, r_sharelink);
521		s = LIST_FIRST(r->r_sharehead);
522		if (r->r_flags & RF_FIRSTSHARE) {
523			s->r_flags |= RF_FIRSTSHARE;
524			TAILQ_INSERT_BEFORE(r, s, r_link);
525			TAILQ_REMOVE(&rm->rm_list, r, r_link);
526		}
527
528		/*
529		 * Make sure that the sharing list goes away completely
530		 * if the resource is no longer being shared at all.
531		 */
532		if (LIST_NEXT(s, r_sharelink) == 0) {
533			free(s->r_sharehead, M_RMAN);
534			s->r_sharehead = 0;
535			s->r_flags &= ~RF_FIRSTSHARE;
536		}
537		goto out;
538	}
539
540	/*
541	 * Look at the adjacent resources in the list and see if our
542	 * segment can be merged with any of them.
543	 */
544	s = TAILQ_PREV(r, resource_head, r_link);
545	t = TAILQ_NEXT(r, r_link);
546
547	if (s != NULL && (s->r_flags & RF_ALLOCATED) == 0
548	    && t != NULL && (t->r_flags & RF_ALLOCATED) == 0) {
549		/*
550		 * Merge all three segments.
551		 */
552		s->r_end = t->r_end;
553		TAILQ_REMOVE(&rm->rm_list, r, r_link);
554		TAILQ_REMOVE(&rm->rm_list, t, r_link);
555		free(t, M_RMAN);
556	} else if (s != NULL && (s->r_flags & RF_ALLOCATED) == 0) {
557		/*
558		 * Merge previous segment with ours.
559		 */
560		s->r_end = r->r_end;
561		TAILQ_REMOVE(&rm->rm_list, r, r_link);
562	} else if (t != NULL && (t->r_flags & RF_ALLOCATED) == 0) {
563		/*
564		 * Merge next segment with ours.
565		 */
566		t->r_start = r->r_start;
567		TAILQ_REMOVE(&rm->rm_list, r, r_link);
568	} else {
569		/*
570		 * At this point, we know there is nothing we
571		 * can potentially merge with, because on each
572		 * side, there is either nothing there or what is
573		 * there is still allocated.  In that case, we don't
574		 * want to remove r from the list; we simply want to
575		 * change it to an unallocated region and return
576		 * without freeing anything.
577		 */
578		r->r_flags &= ~RF_ALLOCATED;
579		return 0;
580	}
581
582out:
583	free(r, M_RMAN);
584	return 0;
585}
586
587int
588rman_release_resource(struct resource *r)
589{
590	int	rv;
591	struct	rman *rm = r->r_rm;
592
593	mtx_lock(rm->rm_mtx);
594	rv = int_rman_release_resource(rm, r);
595	mtx_unlock(rm->rm_mtx);
596	return (rv);
597}
598
599uint32_t
600rman_make_alignment_flags(uint32_t size)
601{
602	int	i;
603
604	/*
605	 * Find the hightest bit set, and add one if more than one bit
606	 * set.  We're effectively computing the ceil(log2(size)) here.
607	 */
608	for (i = 31; i > 0; i--)
609		if ((1 << i) & size)
610			break;
611	if (~(1 << i) & size)
612		i++;
613
614	return(RF_ALIGNMENT_LOG2(i));
615}
616
617u_long
618rman_get_start(struct resource *r)
619{
620	return (r->r_start);
621}
622
623u_long
624rman_get_end(struct resource *r)
625{
626	return (r->r_end);
627}
628
629u_long
630rman_get_size(struct resource *r)
631{
632	return (r->r_end - r->r_start + 1);
633}
634
635u_int
636rman_get_flags(struct resource *r)
637{
638	return (r->r_flags);
639}
640
641void
642rman_set_virtual(struct resource *r, void *v)
643{
644	r->r_virtual = v;
645}
646
647void *
648rman_get_virtual(struct resource *r)
649{
650	return (r->r_virtual);
651}
652
653void
654rman_set_bustag(struct resource *r, bus_space_tag_t t)
655{
656	r->r_bustag = t;
657}
658
659bus_space_tag_t
660rman_get_bustag(struct resource *r)
661{
662	return (r->r_bustag);
663}
664
665void
666rman_set_bushandle(struct resource *r, bus_space_handle_t h)
667{
668	r->r_bushandle = h;
669}
670
671bus_space_handle_t
672rman_get_bushandle(struct resource *r)
673{
674	return (r->r_bushandle);
675}
676
677void
678rman_set_rid(struct resource *r, int rid)
679{
680	r->r_rid = rid;
681}
682
683void
684rman_set_start(struct resource *r, u_long start)
685{
686	r->r_start = start;
687}
688
689void
690rman_set_end(struct resource *r, u_long end)
691{
692	r->r_end = end;
693}
694
695int
696rman_get_rid(struct resource *r)
697{
698	return (r->r_rid);
699}
700
701struct device *
702rman_get_device(struct resource *r)
703{
704	return (r->r_dev);
705}
706