2 * Copyright 1998 Massachusetts Institute of Technology
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
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
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
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
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
58 #include <sys/cdefs.h>
59 __FBSDID("$FreeBSD$");
61 #include <sys/param.h>
62 #include <sys/systm.h>
63 #include <sys/kernel.h>
64 #include <sys/limits.h>
66 #include <sys/malloc.h>
67 #include <sys/mutex.h>
68 #include <sys/bus.h> /* XXX debugging */
69 #include <machine/bus.h>
71 #include <sys/sysctl.h>
74 * We use a linked list rather than a bitmap because we need to be able to
75 * represent potentially huge objects (like all of a processor's physical
76 * address space). That is also why the indices are defined to have type
77 * `unsigned long' -- that being the largest integral type in ISO C (1990).
78 * The 1999 version of C allows `long long'; we may need to switch to that
79 * at some point in the future, particularly if we want to support 36-bit
80 * addresses on IA32 hardware.
84 TAILQ_ENTRY(resource_i) r_link;
85 LIST_ENTRY(resource_i) r_sharelink;
86 LIST_HEAD(, resource_i) *r_sharehead;
87 u_long r_start; /* index of the first entry in this resource */
88 u_long r_end; /* index of the last entry (inclusive) */
90 void *r_virtual; /* virtual address of this resource */
91 struct device *r_dev; /* device which has allocated this resource */
92 struct rman *r_rm; /* resource manager from whence this came */
93 int r_rid; /* optional rid for this resource. */
97 TUNABLE_INT("debug.rman_debug", &rman_debug);
98 SYSCTL_INT(_debug, OID_AUTO, rman_debug, CTLFLAG_RW,
99 &rman_debug, 0, "rman debug");
101 #define DPRINTF(params) if (rman_debug) printf params
103 static MALLOC_DEFINE(M_RMAN, "rman", "Resource manager");
105 struct rman_head rman_head;
106 static struct mtx rman_mtx; /* mutex to protect rman_head */
107 static int int_rman_activate_resource(struct rman *rm, struct resource_i *r,
108 struct resource_i **whohas);
109 static int int_rman_deactivate_resource(struct resource_i *r);
110 static int int_rman_release_resource(struct rman *rm, struct resource_i *r);
112 static __inline struct resource_i *
113 int_alloc_resource(int malloc_flag)
115 struct resource_i *r;
117 r = malloc(sizeof *r, M_RMAN, malloc_flag | M_ZERO);
125 rman_init(struct rman *rm)
131 TAILQ_INIT(&rman_head);
132 mtx_init(&rman_mtx, "rman head", NULL, MTX_DEF);
135 if (rm->rm_type == RMAN_UNINIT)
137 if (rm->rm_type == RMAN_GAUGE)
138 panic("implement RMAN_GAUGE");
140 TAILQ_INIT(&rm->rm_list);
141 rm->rm_mtx = malloc(sizeof *rm->rm_mtx, M_RMAN, M_NOWAIT | M_ZERO);
142 if (rm->rm_mtx == NULL)
144 mtx_init(rm->rm_mtx, "rman", NULL, MTX_DEF);
147 TAILQ_INSERT_TAIL(&rman_head, rm, rm_link);
148 mtx_unlock(&rman_mtx);
153 * NB: this interface is not robust against programming errors which
154 * add multiple copies of the same region.
157 rman_manage_region(struct rman *rm, u_long start, u_long end)
159 struct resource_i *r, *s, *t;
161 DPRINTF(("rman_manage_region: <%s> request: start %#lx, end %#lx\n",
162 rm->rm_descr, start, end));
163 r = int_alloc_resource(M_NOWAIT);
170 mtx_lock(rm->rm_mtx);
172 /* Skip entries before us. */
173 TAILQ_FOREACH(s, &rm->rm_list, r_link) {
174 if (s->r_end == ULONG_MAX)
176 if (s->r_end + 1 >= r->r_start)
180 /* If we ran off the end of the list, insert at the tail. */
182 TAILQ_INSERT_TAIL(&rm->rm_list, r, r_link);
184 /* Check for any overlap with the current region. */
185 if (r->r_start <= s->r_end && r->r_end >= s->r_start)
188 /* Check for any overlap with the next region. */
189 t = TAILQ_NEXT(s, r_link);
190 if (t && r->r_start <= t->r_end && r->r_end >= t->r_start)
194 * See if this region can be merged with the next region. If
195 * not, clear the pointer.
197 if (t && (r->r_end + 1 != t->r_start || t->r_flags != 0))
200 /* See if we can merge with the current region. */
201 if (s->r_end + 1 == r->r_start && s->r_flags == 0) {
202 /* Can we merge all 3 regions? */
205 TAILQ_REMOVE(&rm->rm_list, t, r_link);
213 /* Can we merge with just the next region? */
215 t->r_start = r->r_start;
218 TAILQ_INSERT_BEFORE(s, r, r_link);
222 mtx_unlock(rm->rm_mtx);
227 rman_init_from_resource(struct rman *rm, struct resource *r)
231 if ((rv = rman_init(rm)) != 0)
233 return (rman_manage_region(rm, r->__r_i->r_start, r->__r_i->r_end));
237 rman_fini(struct rman *rm)
239 struct resource_i *r;
241 mtx_lock(rm->rm_mtx);
242 TAILQ_FOREACH(r, &rm->rm_list, r_link) {
243 if (r->r_flags & RF_ALLOCATED) {
244 mtx_unlock(rm->rm_mtx);
250 * There really should only be one of these if we are in this
251 * state and the code is working properly, but it can't hurt.
253 while (!TAILQ_EMPTY(&rm->rm_list)) {
254 r = TAILQ_FIRST(&rm->rm_list);
255 TAILQ_REMOVE(&rm->rm_list, r, r_link);
258 mtx_unlock(rm->rm_mtx);
260 TAILQ_REMOVE(&rman_head, rm, rm_link);
261 mtx_unlock(&rman_mtx);
262 mtx_destroy(rm->rm_mtx);
263 free(rm->rm_mtx, M_RMAN);
269 rman_reserve_resource_bound(struct rman *rm, u_long start, u_long end,
270 u_long count, u_long bound, u_int flags,
274 struct resource_i *r, *s, *rv;
275 u_long rstart, rend, amask, bmask;
279 DPRINTF(("rman_reserve_resource_bound: <%s> request: [%#lx, %#lx], "
280 "length %#lx, flags %u, device %s\n", rm->rm_descr, start, end,
282 dev == NULL ? "<null>" : device_get_nameunit(dev)));
283 want_activate = (flags & RF_ACTIVE);
286 mtx_lock(rm->rm_mtx);
288 for (r = TAILQ_FIRST(&rm->rm_list);
289 r && r->r_end < start;
290 r = TAILQ_NEXT(r, r_link))
294 DPRINTF(("could not find a region\n"));
298 amask = (1ul << RF_ALIGNMENT(flags)) - 1;
299 /* If bound is 0, bmask will also be 0 */
300 bmask = ~(bound - 1);
302 * First try to find an acceptable totally-unshared region.
304 for (s = r; s; s = TAILQ_NEXT(s, r_link)) {
305 DPRINTF(("considering [%#lx, %#lx]\n", s->r_start, s->r_end));
306 if (s->r_start + count - 1 > end) {
307 DPRINTF(("s->r_start (%#lx) + count - 1> end (%#lx)\n",
311 if (s->r_flags & RF_ALLOCATED) {
312 DPRINTF(("region is allocated\n"));
315 rstart = ulmax(s->r_start, start);
317 * Try to find a region by adjusting to boundary and alignment
318 * until both conditions are satisfied. This is not an optimal
319 * algorithm, but in most cases it isn't really bad, either.
322 rstart = (rstart + amask) & ~amask;
323 if (((rstart ^ (rstart + count - 1)) & bmask) != 0)
324 rstart += bound - (rstart & ~bmask);
325 } while ((rstart & amask) != 0 && rstart < end &&
327 rend = ulmin(s->r_end, ulmax(rstart + count - 1, end));
329 DPRINTF(("adjusted start exceeds end\n"));
332 DPRINTF(("truncated region: [%#lx, %#lx]; size %#lx (requested %#lx)\n",
333 rstart, rend, (rend - rstart + 1), count));
335 if ((rend - rstart + 1) >= count) {
336 DPRINTF(("candidate region: [%#lx, %#lx], size %#lx\n",
337 rstart, rend, (rend - rstart + 1)));
338 if ((s->r_end - s->r_start + 1) == count) {
339 DPRINTF(("candidate region is entire chunk\n"));
341 rv->r_flags |= RF_ALLOCATED | flags;
347 * If s->r_start < rstart and
348 * s->r_end > rstart + count - 1, then
349 * we need to split the region into three pieces
350 * (the middle one will get returned to the user).
351 * Otherwise, we are allocating at either the
352 * beginning or the end of s, so we only need to
353 * split it in two. The first case requires
354 * two new allocations; the second requires but one.
356 rv = int_alloc_resource(M_NOWAIT);
359 rv->r_start = rstart;
360 rv->r_end = rstart + count - 1;
361 rv->r_flags = flags | RF_ALLOCATED;
365 if (s->r_start < rv->r_start && s->r_end > rv->r_end) {
366 DPRINTF(("splitting region in three parts: "
367 "[%#lx, %#lx]; [%#lx, %#lx]; [%#lx, %#lx]\n",
368 s->r_start, rv->r_start - 1,
369 rv->r_start, rv->r_end,
370 rv->r_end + 1, s->r_end));
372 * We are allocating in the middle.
374 r = int_alloc_resource(M_NOWAIT);
380 r->r_start = rv->r_end + 1;
382 r->r_flags = s->r_flags;
384 s->r_end = rv->r_start - 1;
385 TAILQ_INSERT_AFTER(&rm->rm_list, s, rv,
387 TAILQ_INSERT_AFTER(&rm->rm_list, rv, r,
389 } else if (s->r_start == rv->r_start) {
390 DPRINTF(("allocating from the beginning\n"));
392 * We are allocating at the beginning.
394 s->r_start = rv->r_end + 1;
395 TAILQ_INSERT_BEFORE(s, rv, r_link);
397 DPRINTF(("allocating at the end\n"));
399 * We are allocating at the end.
401 s->r_end = rv->r_start - 1;
402 TAILQ_INSERT_AFTER(&rm->rm_list, s, rv,
410 * Now find an acceptable shared region, if the client's requirements
411 * allow sharing. By our implementation restriction, a candidate
412 * region must match exactly by both size and sharing type in order
413 * to be considered compatible with the client's request. (The
414 * former restriction could probably be lifted without too much
415 * additional work, but this does not seem warranted.)
417 DPRINTF(("no unshared regions found\n"));
418 if ((flags & (RF_SHAREABLE | RF_TIMESHARE)) == 0)
421 for (s = r; s; s = TAILQ_NEXT(s, r_link)) {
422 if (s->r_start > end)
424 if ((s->r_flags & flags) != flags)
426 rstart = ulmax(s->r_start, start);
427 rend = ulmin(s->r_end, ulmax(start + count - 1, end));
428 if (s->r_start >= start && s->r_end <= end
429 && (s->r_end - s->r_start + 1) == count &&
430 (s->r_start & amask) == 0 &&
431 ((s->r_start ^ s->r_end) & bmask) == 0) {
432 rv = int_alloc_resource(M_NOWAIT);
435 rv->r_start = s->r_start;
436 rv->r_end = s->r_end;
437 rv->r_flags = s->r_flags &
438 (RF_ALLOCATED | RF_SHAREABLE | RF_TIMESHARE);
441 if (s->r_sharehead == NULL) {
442 s->r_sharehead = malloc(sizeof *s->r_sharehead,
443 M_RMAN, M_NOWAIT | M_ZERO);
444 if (s->r_sharehead == NULL) {
449 LIST_INIT(s->r_sharehead);
450 LIST_INSERT_HEAD(s->r_sharehead, s,
452 s->r_flags |= RF_FIRSTSHARE;
454 rv->r_sharehead = s->r_sharehead;
455 LIST_INSERT_HEAD(s->r_sharehead, rv, r_sharelink);
461 * We couldn't find anything.
465 * If the user specified RF_ACTIVE in the initial flags,
466 * which is reflected in `want_activate', we attempt to atomically
467 * activate the resource. If this fails, we release the resource
468 * and indicate overall failure. (This behavior probably doesn't
469 * make sense for RF_TIMESHARE-type resources.)
471 if (rv && want_activate) {
472 struct resource_i *whohas;
473 if (int_rman_activate_resource(rm, rv, &whohas)) {
474 int_rman_release_resource(rm, rv);
479 mtx_unlock(rm->rm_mtx);
480 return (rv == NULL ? NULL : &rv->r_r);
484 rman_reserve_resource(struct rman *rm, u_long start, u_long end, u_long count,
485 u_int flags, struct device *dev)
488 return (rman_reserve_resource_bound(rm, start, end, count, 0, flags,
493 int_rman_activate_resource(struct rman *rm, struct resource_i *r,
494 struct resource_i **whohas)
496 struct resource_i *s;
500 * If we are not timesharing, then there is nothing much to do.
501 * If we already have the resource, then there is nothing at all to do.
502 * If we are not on a sharing list with anybody else, then there is
505 if ((r->r_flags & RF_TIMESHARE) == 0
506 || (r->r_flags & RF_ACTIVE) != 0
507 || r->r_sharehead == NULL) {
508 r->r_flags |= RF_ACTIVE;
513 for (s = LIST_FIRST(r->r_sharehead); s && ok;
514 s = LIST_NEXT(s, r_sharelink)) {
515 if ((s->r_flags & RF_ACTIVE) != 0) {
521 r->r_flags |= RF_ACTIVE;
528 rman_activate_resource(struct resource *re)
531 struct resource_i *r, *whohas;
536 mtx_lock(rm->rm_mtx);
537 rv = int_rman_activate_resource(rm, r, &whohas);
538 mtx_unlock(rm->rm_mtx);
543 rman_await_resource(struct resource *re, int pri, int timo)
546 struct resource_i *r, *whohas;
551 mtx_lock(rm->rm_mtx);
553 rv = int_rman_activate_resource(rm, r, &whohas);
555 return (rv); /* returns with mutex held */
557 if (r->r_sharehead == NULL)
558 panic("rman_await_resource");
559 whohas->r_flags |= RF_WANTED;
560 rv = msleep(r->r_sharehead, rm->rm_mtx, pri, "rmwait", timo);
562 mtx_unlock(rm->rm_mtx);
569 int_rman_deactivate_resource(struct resource_i *r)
572 r->r_flags &= ~RF_ACTIVE;
573 if (r->r_flags & RF_WANTED) {
574 r->r_flags &= ~RF_WANTED;
575 wakeup(r->r_sharehead);
581 rman_deactivate_resource(struct resource *r)
586 mtx_lock(rm->rm_mtx);
587 int_rman_deactivate_resource(r->__r_i);
588 mtx_unlock(rm->rm_mtx);
593 int_rman_release_resource(struct rman *rm, struct resource_i *r)
595 struct resource_i *s, *t;
597 if (r->r_flags & RF_ACTIVE)
598 int_rman_deactivate_resource(r);
601 * Check for a sharing list first. If there is one, then we don't
602 * have to think as hard.
604 if (r->r_sharehead) {
606 * If a sharing list exists, then we know there are at
609 * If we are in the main circleq, appoint someone else.
611 LIST_REMOVE(r, r_sharelink);
612 s = LIST_FIRST(r->r_sharehead);
613 if (r->r_flags & RF_FIRSTSHARE) {
614 s->r_flags |= RF_FIRSTSHARE;
615 TAILQ_INSERT_BEFORE(r, s, r_link);
616 TAILQ_REMOVE(&rm->rm_list, r, r_link);
620 * Make sure that the sharing list goes away completely
621 * if the resource is no longer being shared at all.
623 if (LIST_NEXT(s, r_sharelink) == NULL) {
624 free(s->r_sharehead, M_RMAN);
625 s->r_sharehead = NULL;
626 s->r_flags &= ~RF_FIRSTSHARE;
632 * Look at the adjacent resources in the list and see if our
633 * segment can be merged with any of them. If either of the
634 * resources is allocated or is not exactly adjacent then they
635 * cannot be merged with our segment.
637 s = TAILQ_PREV(r, resource_head, r_link);
638 if (s != NULL && ((s->r_flags & RF_ALLOCATED) != 0 ||
639 s->r_end + 1 != r->r_start))
641 t = TAILQ_NEXT(r, r_link);
642 if (t != NULL && ((t->r_flags & RF_ALLOCATED) != 0 ||
643 r->r_end + 1 != t->r_start))
646 if (s != NULL && t != NULL) {
648 * Merge all three segments.
651 TAILQ_REMOVE(&rm->rm_list, r, r_link);
652 TAILQ_REMOVE(&rm->rm_list, t, r_link);
654 } else if (s != NULL) {
656 * Merge previous segment with ours.
659 TAILQ_REMOVE(&rm->rm_list, r, r_link);
660 } else if (t != NULL) {
662 * Merge next segment with ours.
664 t->r_start = r->r_start;
665 TAILQ_REMOVE(&rm->rm_list, r, r_link);
668 * At this point, we know there is nothing we
669 * can potentially merge with, because on each
670 * side, there is either nothing there or what is
671 * there is still allocated. In that case, we don't
672 * want to remove r from the list; we simply want to
673 * change it to an unallocated region and return
674 * without freeing anything.
676 r->r_flags &= ~RF_ALLOCATED;
686 rman_release_resource(struct resource *re)
689 struct resource_i *r;
694 mtx_lock(rm->rm_mtx);
695 rv = int_rman_release_resource(rm, r);
696 mtx_unlock(rm->rm_mtx);
701 rman_make_alignment_flags(uint32_t size)
706 * Find the hightest bit set, and add one if more than one bit
707 * set. We're effectively computing the ceil(log2(size)) here.
709 for (i = 31; i > 0; i--)
712 if (~(1 << i) & size)
715 return(RF_ALIGNMENT_LOG2(i));
719 rman_get_start(struct resource *r)
721 return (r->__r_i->r_start);
725 rman_get_end(struct resource *r)
727 return (r->__r_i->r_end);
731 rman_get_size(struct resource *r)
733 return (r->__r_i->r_end - r->__r_i->r_start + 1);
737 rman_get_flags(struct resource *r)
739 return (r->__r_i->r_flags);
743 rman_set_virtual(struct resource *r, void *v)
745 r->__r_i->r_virtual = v;
749 rman_get_virtual(struct resource *r)
751 return (r->__r_i->r_virtual);
755 rman_set_bustag(struct resource *r, bus_space_tag_t t)
761 rman_get_bustag(struct resource *r)
763 return (r->r_bustag);
767 rman_set_bushandle(struct resource *r, bus_space_handle_t h)
773 rman_get_bushandle(struct resource *r)
775 return (r->r_bushandle);
779 rman_set_rid(struct resource *r, int rid)
781 r->__r_i->r_rid = rid;
785 rman_set_start(struct resource *r, u_long start)
787 r->__r_i->r_start = start;
791 rman_set_end(struct resource *r, u_long end)
793 r->__r_i->r_end = end;
797 rman_get_rid(struct resource *r)
799 return (r->__r_i->r_rid);
803 rman_get_device(struct resource *r)
805 return (r->__r_i->r_dev);
809 rman_set_device(struct resource *r, struct device *dev)
811 r->__r_i->r_dev = dev;
815 rman_is_region_manager(struct resource *r, struct rman *rm)
818 return (r->__r_i->r_rm == rm);
822 * Sysctl interface for scanning the resource lists.
824 * We take two input parameters; the index into the list of resource
825 * managers, and the resource offset into the list.
828 sysctl_rman(SYSCTL_HANDLER_ARGS)
830 int *name = (int *)arg1;
831 u_int namelen = arg2;
832 int rman_idx, res_idx;
834 struct resource_i *res;
836 struct u_resource ures;
842 if (bus_data_generation_check(name[0]))
848 * Find the indexed resource manager
851 TAILQ_FOREACH(rm, &rman_head, rm_link) {
855 mtx_unlock(&rman_mtx);
860 * If the resource index is -1, we want details on the
864 bzero(&urm, sizeof(urm));
865 urm.rm_handle = (uintptr_t)rm;
866 strlcpy(urm.rm_descr, rm->rm_descr, RM_TEXTLEN);
867 urm.rm_start = rm->rm_start;
868 urm.rm_size = rm->rm_end - rm->rm_start + 1;
869 urm.rm_type = rm->rm_type;
871 error = SYSCTL_OUT(req, &urm, sizeof(urm));
876 * Find the indexed resource and return it.
878 mtx_lock(rm->rm_mtx);
879 TAILQ_FOREACH(res, &rm->rm_list, r_link) {
880 if (res_idx-- == 0) {
881 bzero(&ures, sizeof(ures));
882 ures.r_handle = (uintptr_t)res;
883 ures.r_parent = (uintptr_t)res->r_rm;
884 ures.r_device = (uintptr_t)res->r_dev;
885 if (res->r_dev != NULL) {
886 if (device_get_name(res->r_dev) != NULL) {
887 snprintf(ures.r_devname, RM_TEXTLEN,
889 device_get_name(res->r_dev),
890 device_get_unit(res->r_dev));
892 strlcpy(ures.r_devname, "nomatch",
896 ures.r_devname[0] = '\0';
898 ures.r_start = res->r_start;
899 ures.r_size = res->r_end - res->r_start + 1;
900 ures.r_flags = res->r_flags;
902 mtx_unlock(rm->rm_mtx);
903 error = SYSCTL_OUT(req, &ures, sizeof(ures));
907 mtx_unlock(rm->rm_mtx);
911 SYSCTL_NODE(_hw_bus, OID_AUTO, rman, CTLFLAG_RD, sysctl_rman,
912 "kernel resource manager");