subr_rman.c revision 143664
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 143664 2005-03-15 20:15:15Z 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	DPRINTF(("rman_manage_region: <%s> request: start %#lx, end %#lx\n",
127	    rm->rm_descr, start, end));
128	r = malloc(sizeof *r, M_RMAN, M_NOWAIT | M_ZERO);
129	if (r == 0)
130		return ENOMEM;
131	r->r_start = start;
132	r->r_end = end;
133	r->r_rm = rm;
134
135	mtx_lock(rm->rm_mtx);
136	for (s = TAILQ_FIRST(&rm->rm_list);
137	     s && s->r_end < r->r_start;
138	     s = TAILQ_NEXT(s, r_link))
139		;
140
141	if (s == NULL) {
142		TAILQ_INSERT_TAIL(&rm->rm_list, r, r_link);
143	} else {
144		TAILQ_INSERT_BEFORE(s, r, r_link);
145	}
146
147	mtx_unlock(rm->rm_mtx);
148	return 0;
149}
150
151int
152rman_fini(struct rman *rm)
153{
154	struct resource *r;
155
156	mtx_lock(rm->rm_mtx);
157	TAILQ_FOREACH(r, &rm->rm_list, r_link) {
158		if (r->r_flags & RF_ALLOCATED) {
159			mtx_unlock(rm->rm_mtx);
160			return EBUSY;
161		}
162	}
163
164	/*
165	 * There really should only be one of these if we are in this
166	 * state and the code is working properly, but it can't hurt.
167	 */
168	while (!TAILQ_EMPTY(&rm->rm_list)) {
169		r = TAILQ_FIRST(&rm->rm_list);
170		TAILQ_REMOVE(&rm->rm_list, r, r_link);
171		free(r, M_RMAN);
172	}
173	mtx_unlock(rm->rm_mtx);
174	mtx_lock(&rman_mtx);
175	TAILQ_REMOVE(&rman_head, rm, rm_link);
176	mtx_unlock(&rman_mtx);
177	mtx_destroy(rm->rm_mtx);
178	free(rm->rm_mtx, M_RMAN);
179
180	return 0;
181}
182
183struct resource *
184rman_reserve_resource_bound(struct rman *rm, u_long start, u_long end,
185		      u_long count, u_long bound,  u_int flags,
186		      struct device *dev)
187{
188	u_int	want_activate;
189	struct	resource *r, *s, *rv;
190	u_long	rstart, rend, amask, bmask;
191
192	rv = 0;
193
194	DPRINTF(("rman_reserve_resource: <%s> request: [%#lx, %#lx], length "
195	       "%#lx, flags %u, device %s\n", rm->rm_descr, start, end, count,
196	       flags, dev == NULL ? "<null>" : device_get_nameunit(dev)));
197	want_activate = (flags & RF_ACTIVE);
198	flags &= ~RF_ACTIVE;
199
200	mtx_lock(rm->rm_mtx);
201
202	for (r = TAILQ_FIRST(&rm->rm_list);
203	     r && r->r_end < start;
204	     r = TAILQ_NEXT(r, r_link))
205		;
206
207	if (r == NULL) {
208		DPRINTF(("could not find a region\n"));
209		goto out;
210	}
211
212	amask = (1ul << RF_ALIGNMENT(flags)) - 1;
213	/* If bound is 0, bmask will also be 0 */
214	bmask = ~(bound - 1);
215	/*
216	 * First try to find an acceptable totally-unshared region.
217	 */
218	for (s = r; s; s = TAILQ_NEXT(s, r_link)) {
219		DPRINTF(("considering [%#lx, %#lx]\n", s->r_start, s->r_end));
220		if (s->r_start > end) {
221			DPRINTF(("s->r_start (%#lx) > end (%#lx)\n", s->r_start, end));
222			break;
223		}
224		if (s->r_flags & RF_ALLOCATED) {
225			DPRINTF(("region is allocated\n"));
226			continue;
227		}
228		rstart = ulmax(s->r_start, start);
229		/*
230		 * Try to find a region by adjusting to boundary and alignment
231		 * until both conditions are satisfied. This is not an optimal
232		 * algorithm, but in most cases it isn't really bad, either.
233		 */
234		do {
235			rstart = (rstart + amask) & ~amask;
236			if (((rstart ^ (rstart + count - 1)) & bmask) != 0)
237				rstart += bound - (rstart & ~bmask);
238		} while ((rstart & amask) != 0 && rstart < end &&
239		    rstart < s->r_end);
240		rend = ulmin(s->r_end, ulmax(rstart + count - 1, end));
241		if (rstart > rend) {
242			DPRINTF(("adjusted start exceeds end\n"));
243			continue;
244		}
245		DPRINTF(("truncated region: [%#lx, %#lx]; size %#lx (requested %#lx)\n",
246		       rstart, rend, (rend - rstart + 1), count));
247
248		if ((rend - rstart + 1) >= count) {
249			DPRINTF(("candidate region: [%#lx, %#lx], size %#lx\n",
250			       rstart, rend, (rend - rstart + 1)));
251			if ((s->r_end - s->r_start + 1) == count) {
252				DPRINTF(("candidate region is entire chunk\n"));
253				rv = s;
254				rv->r_flags |= RF_ALLOCATED | flags;
255				rv->r_dev = dev;
256				goto out;
257			}
258
259			/*
260			 * If s->r_start < rstart and
261			 *    s->r_end > rstart + count - 1, then
262			 * we need to split the region into three pieces
263			 * (the middle one will get returned to the user).
264			 * Otherwise, we are allocating at either the
265			 * beginning or the end of s, so we only need to
266			 * split it in two.  The first case requires
267			 * two new allocations; the second requires but one.
268			 */
269			rv = malloc(sizeof *rv, M_RMAN, M_NOWAIT | M_ZERO);
270			if (rv == 0)
271				goto out;
272			rv->r_start = rstart;
273			rv->r_end = rstart + count - 1;
274			rv->r_flags = flags | RF_ALLOCATED;
275			rv->r_dev = dev;
276			rv->r_rm = rm;
277
278			if (s->r_start < rv->r_start && s->r_end > rv->r_end) {
279				DPRINTF(("splitting region in three parts: "
280				       "[%#lx, %#lx]; [%#lx, %#lx]; [%#lx, %#lx]\n",
281				       s->r_start, rv->r_start - 1,
282				       rv->r_start, rv->r_end,
283				       rv->r_end + 1, s->r_end));
284				/*
285				 * We are allocating in the middle.
286				 */
287				r = malloc(sizeof *r, M_RMAN, M_NOWAIT|M_ZERO);
288				if (r == 0) {
289					free(rv, M_RMAN);
290					rv = 0;
291					goto out;
292				}
293				r->r_start = rv->r_end + 1;
294				r->r_end = s->r_end;
295				r->r_flags = s->r_flags;
296				r->r_rm = rm;
297				s->r_end = rv->r_start - 1;
298				TAILQ_INSERT_AFTER(&rm->rm_list, s, rv,
299						     r_link);
300				TAILQ_INSERT_AFTER(&rm->rm_list, rv, r,
301						     r_link);
302			} else if (s->r_start == rv->r_start) {
303				DPRINTF(("allocating from the beginning\n"));
304				/*
305				 * We are allocating at the beginning.
306				 */
307				s->r_start = rv->r_end + 1;
308				TAILQ_INSERT_BEFORE(s, rv, r_link);
309			} else {
310				DPRINTF(("allocating at the end\n"));
311				/*
312				 * We are allocating at the end.
313				 */
314				s->r_end = rv->r_start - 1;
315				TAILQ_INSERT_AFTER(&rm->rm_list, s, rv,
316						     r_link);
317			}
318			goto out;
319		}
320	}
321
322	/*
323	 * Now find an acceptable shared region, if the client's requirements
324	 * allow sharing.  By our implementation restriction, a candidate
325	 * region must match exactly by both size and sharing type in order
326	 * to be considered compatible with the client's request.  (The
327	 * former restriction could probably be lifted without too much
328	 * additional work, but this does not seem warranted.)
329	 */
330	DPRINTF(("no unshared regions found\n"));
331	if ((flags & (RF_SHAREABLE | RF_TIMESHARE)) == 0)
332		goto out;
333
334	for (s = r; s; s = TAILQ_NEXT(s, r_link)) {
335		if (s->r_start > end)
336			break;
337		if ((s->r_flags & flags) != flags)
338			continue;
339		rstart = ulmax(s->r_start, start);
340		rend = ulmin(s->r_end, ulmax(start + count - 1, end));
341		if (s->r_start >= start && s->r_end <= end
342		    && (s->r_end - s->r_start + 1) == count &&
343		    (s->r_start & amask) == 0 &&
344		    ((s->r_start ^ s->r_end) & bmask) == 0) {
345			rv = malloc(sizeof *rv, M_RMAN, M_NOWAIT | M_ZERO);
346			if (rv == 0)
347				goto out;
348			rv->r_start = s->r_start;
349			rv->r_end = s->r_end;
350			rv->r_flags = s->r_flags &
351				(RF_ALLOCATED | RF_SHAREABLE | RF_TIMESHARE);
352			rv->r_dev = dev;
353			rv->r_rm = rm;
354			if (s->r_sharehead == 0) {
355				s->r_sharehead = malloc(sizeof *s->r_sharehead,
356						M_RMAN, M_NOWAIT | M_ZERO);
357				if (s->r_sharehead == 0) {
358					free(rv, M_RMAN);
359					rv = 0;
360					goto out;
361				}
362				LIST_INIT(s->r_sharehead);
363				LIST_INSERT_HEAD(s->r_sharehead, s,
364						 r_sharelink);
365				s->r_flags |= RF_FIRSTSHARE;
366			}
367			rv->r_sharehead = s->r_sharehead;
368			LIST_INSERT_HEAD(s->r_sharehead, rv, r_sharelink);
369			goto out;
370		}
371	}
372
373	/*
374	 * We couldn't find anything.
375	 */
376out:
377	/*
378	 * If the user specified RF_ACTIVE in the initial flags,
379	 * which is reflected in `want_activate', we attempt to atomically
380	 * activate the resource.  If this fails, we release the resource
381	 * and indicate overall failure.  (This behavior probably doesn't
382	 * make sense for RF_TIMESHARE-type resources.)
383	 */
384	if (rv && want_activate) {
385		struct resource *whohas;
386		if (int_rman_activate_resource(rm, rv, &whohas)) {
387			int_rman_release_resource(rm, rv);
388			rv = 0;
389		}
390	}
391
392	mtx_unlock(rm->rm_mtx);
393	return (rv);
394}
395
396struct resource *
397rman_reserve_resource(struct rman *rm, u_long start, u_long end, u_long count,
398		      u_int flags, struct device *dev)
399{
400
401	return (rman_reserve_resource_bound(rm, start, end, count, 0, flags,
402	    dev));
403}
404
405static int
406int_rman_activate_resource(struct rman *rm, struct resource *r,
407			   struct resource **whohas)
408{
409	struct resource *s;
410	int ok;
411
412	/*
413	 * If we are not timesharing, then there is nothing much to do.
414	 * If we already have the resource, then there is nothing at all to do.
415	 * If we are not on a sharing list with anybody else, then there is
416	 * little to do.
417	 */
418	if ((r->r_flags & RF_TIMESHARE) == 0
419	    || (r->r_flags & RF_ACTIVE) != 0
420	    || r->r_sharehead == 0) {
421		r->r_flags |= RF_ACTIVE;
422		return 0;
423	}
424
425	ok = 1;
426	for (s = LIST_FIRST(r->r_sharehead); s && ok;
427	     s = LIST_NEXT(s, r_sharelink)) {
428		if ((s->r_flags & RF_ACTIVE) != 0) {
429			ok = 0;
430			*whohas = s;
431		}
432	}
433	if (ok) {
434		r->r_flags |= RF_ACTIVE;
435		return 0;
436	}
437	return EBUSY;
438}
439
440int
441rman_activate_resource(struct resource *r)
442{
443	int rv;
444	struct resource *whohas;
445	struct rman *rm;
446
447	rm = r->r_rm;
448	mtx_lock(rm->rm_mtx);
449	rv = int_rman_activate_resource(rm, r, &whohas);
450	mtx_unlock(rm->rm_mtx);
451	return rv;
452}
453
454int
455rman_await_resource(struct resource *r, int pri, int timo)
456{
457	int	rv;
458	struct	resource *whohas;
459	struct	rman *rm;
460
461	rm = r->r_rm;
462	mtx_lock(rm->rm_mtx);
463	for (;;) {
464		rv = int_rman_activate_resource(rm, r, &whohas);
465		if (rv != EBUSY)
466			return (rv);	/* returns with mutex held */
467
468		if (r->r_sharehead == 0)
469			panic("rman_await_resource");
470		whohas->r_flags |= RF_WANTED;
471		rv = msleep(r->r_sharehead, rm->rm_mtx, pri, "rmwait", timo);
472		if (rv) {
473			mtx_unlock(rm->rm_mtx);
474			return (rv);
475		}
476	}
477}
478
479static int
480int_rman_deactivate_resource(struct resource *r)
481{
482
483	r->r_flags &= ~RF_ACTIVE;
484	if (r->r_flags & RF_WANTED) {
485		r->r_flags &= ~RF_WANTED;
486		wakeup(r->r_sharehead);
487	}
488	return 0;
489}
490
491int
492rman_deactivate_resource(struct resource *r)
493{
494	struct	rman *rm;
495
496	rm = r->r_rm;
497	mtx_lock(rm->rm_mtx);
498	int_rman_deactivate_resource(r);
499	mtx_unlock(rm->rm_mtx);
500	return 0;
501}
502
503static int
504int_rman_release_resource(struct rman *rm, struct resource *r)
505{
506	struct	resource *s, *t;
507
508	if (r->r_flags & RF_ACTIVE)
509		int_rman_deactivate_resource(r);
510
511	/*
512	 * Check for a sharing list first.  If there is one, then we don't
513	 * have to think as hard.
514	 */
515	if (r->r_sharehead) {
516		/*
517		 * If a sharing list exists, then we know there are at
518		 * least two sharers.
519		 *
520		 * If we are in the main circleq, appoint someone else.
521		 */
522		LIST_REMOVE(r, r_sharelink);
523		s = LIST_FIRST(r->r_sharehead);
524		if (r->r_flags & RF_FIRSTSHARE) {
525			s->r_flags |= RF_FIRSTSHARE;
526			TAILQ_INSERT_BEFORE(r, s, r_link);
527			TAILQ_REMOVE(&rm->rm_list, r, r_link);
528		}
529
530		/*
531		 * Make sure that the sharing list goes away completely
532		 * if the resource is no longer being shared at all.
533		 */
534		if (LIST_NEXT(s, r_sharelink) == 0) {
535			free(s->r_sharehead, M_RMAN);
536			s->r_sharehead = 0;
537			s->r_flags &= ~RF_FIRSTSHARE;
538		}
539		goto out;
540	}
541
542	/*
543	 * Look at the adjacent resources in the list and see if our
544	 * segment can be merged with any of them.  If either of the
545	 * resources is allocated or is not exactly adjacent then they
546	 * cannot be merged with our segment.
547	 */
548	s = TAILQ_PREV(r, resource_head, r_link);
549	if (s != NULL && ((s->r_flags & RF_ALLOCATED) != 0 ||
550	    s->r_end + 1 != r->r_start))
551		s = NULL;
552	t = TAILQ_NEXT(r, r_link);
553	if (t != NULL && ((t->r_flags & RF_ALLOCATED) != 0 ||
554	    r->r_end + 1 != t->r_start))
555		t = NULL;
556
557	if (s != NULL && t != NULL) {
558		/*
559		 * Merge all three segments.
560		 */
561		s->r_end = t->r_end;
562		TAILQ_REMOVE(&rm->rm_list, r, r_link);
563		TAILQ_REMOVE(&rm->rm_list, t, r_link);
564		free(t, M_RMAN);
565	} else if (s != NULL) {
566		/*
567		 * Merge previous segment with ours.
568		 */
569		s->r_end = r->r_end;
570		TAILQ_REMOVE(&rm->rm_list, r, r_link);
571	} else if (t != NULL) {
572		/*
573		 * Merge next segment with ours.
574		 */
575		t->r_start = r->r_start;
576		TAILQ_REMOVE(&rm->rm_list, r, r_link);
577	} else {
578		/*
579		 * At this point, we know there is nothing we
580		 * can potentially merge with, because on each
581		 * side, there is either nothing there or what is
582		 * there is still allocated.  In that case, we don't
583		 * want to remove r from the list; we simply want to
584		 * change it to an unallocated region and return
585		 * without freeing anything.
586		 */
587		r->r_flags &= ~RF_ALLOCATED;
588		return 0;
589	}
590
591out:
592	free(r, M_RMAN);
593	return 0;
594}
595
596int
597rman_release_resource(struct resource *r)
598{
599	int	rv;
600	struct	rman *rm = r->r_rm;
601
602	mtx_lock(rm->rm_mtx);
603	rv = int_rman_release_resource(rm, r);
604	mtx_unlock(rm->rm_mtx);
605	return (rv);
606}
607
608uint32_t
609rman_make_alignment_flags(uint32_t size)
610{
611	int	i;
612
613	/*
614	 * Find the hightest bit set, and add one if more than one bit
615	 * set.  We're effectively computing the ceil(log2(size)) here.
616	 */
617	for (i = 31; i > 0; i--)
618		if ((1 << i) & size)
619			break;
620	if (~(1 << i) & size)
621		i++;
622
623	return(RF_ALIGNMENT_LOG2(i));
624}
625
626u_long
627rman_get_start(struct resource *r)
628{
629	return (r->r_start);
630}
631
632u_long
633rman_get_end(struct resource *r)
634{
635	return (r->r_end);
636}
637
638u_long
639rman_get_size(struct resource *r)
640{
641	return (r->r_end - r->r_start + 1);
642}
643
644u_int
645rman_get_flags(struct resource *r)
646{
647	return (r->r_flags);
648}
649
650void
651rman_set_virtual(struct resource *r, void *v)
652{
653	r->r_virtual = v;
654}
655
656void *
657rman_get_virtual(struct resource *r)
658{
659	return (r->r_virtual);
660}
661
662void
663rman_set_bustag(struct resource *r, bus_space_tag_t t)
664{
665	r->r_bustag = t;
666}
667
668bus_space_tag_t
669rman_get_bustag(struct resource *r)
670{
671	return (r->r_bustag);
672}
673
674void
675rman_set_bushandle(struct resource *r, bus_space_handle_t h)
676{
677	r->r_bushandle = h;
678}
679
680bus_space_handle_t
681rman_get_bushandle(struct resource *r)
682{
683	return (r->r_bushandle);
684}
685
686void
687rman_set_rid(struct resource *r, int rid)
688{
689	r->r_rid = rid;
690}
691
692void
693rman_set_start(struct resource *r, u_long start)
694{
695	r->r_start = start;
696}
697
698void
699rman_set_end(struct resource *r, u_long end)
700{
701	r->r_end = end;
702}
703
704int
705rman_get_rid(struct resource *r)
706{
707	return (r->r_rid);
708}
709
710struct device *
711rman_get_device(struct resource *r)
712{
713	return (r->r_dev);
714}
715