2 * Copyright (c) 2004 Poul-Henning Kamp
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * $FreeBSD: src/sys/kern/subr_unit.c,v 1.7 2005/03/14 06:51:29 phk Exp $
29 * Unit number allocation functions.
31 * These functions implement a mixed run-length/bitmap management of unit
32 * number spaces in a very memory efficient manner.
34 * Allocation policy is always lowest free number first.
36 * A return value of -1 signals that no more unit numbers are available.
38 * There is no cost associated with the range of unitnumbers, so unless
39 * the resource really is finite, specify INT_MAX to new_unrhdr() and
40 * forget about checking the return value.
42 * If a mutex is not provided when the unit number space is created, a
43 * default global mutex is used. The advantage to passing a mutex in, is
44 * that the the alloc_unrl() function can be called with the mutex already
45 * held (it will not be released by alloc_unrl()).
47 * The allocation function alloc_unr{l}() never sleeps (but it may block on
48 * the mutex of course).
50 * Freeing a unit number may require allocating memory, and can therefore
51 * sleep so the free_unr() function does not come in a pre-locked variant.
53 * A userland test program is included.
55 * Memory usage is a very complex function of the the exact allocation
56 * pattern, but always very compact:
57 * * For the very typical case where a single unbroken run of unit
58 * numbers are allocated 44 bytes are used on x86.
59 * * For a unit number space of 1000 units and the random pattern
60 * in the usermode test program included, the worst case usage
61 * was 252 bytes on x86 for 500 allocated and 500 free units.
62 * * For a unit number space of 10000 units and the random pattern
63 * in the usermode test program included, the worst case usage
64 * was 798 bytes on x86 for 5000 allocated and 5000 free units.
65 * * The worst case is where every other unit number is allocated and
66 * the the rest are free. In that case 44 + N/4 bytes are used where
67 * N is the number of the highest unit allocated.
70 #include <sys/types.h>
71 #include <sys/queue.h>
72 #include <sys/bitstring.h>
76 #include <sys/param.h>
77 #include <sys/malloc.h>
78 #include <sys/kernel.h>
79 #include <sys/systm.h>
80 #include <sys/limits.h>
85 * In theory it would be smarter to allocate the individual blocks
86 * with the zone allocator, but at this time the expectation is that
87 * there will typically not even be enough allocations to fill a single
88 * page, so we stick with malloc for now.
90 static MALLOC_DEFINE(M_UNIT, "Unitno", "Unit number allocation");
92 #define Malloc(foo) kmalloc(foo, M_UNIT, M_WAITOK | M_ZERO)
93 #define Free(foo) kfree(foo, M_UNIT)
95 static struct lock unit_lock;
97 LOCK_SYSINIT(unit, &unit_lock, "unit# allocation", LK_CANRECURSE);
99 #else /* ...USERLAND */
101 /* No unit allocation on DragonFly's userland */
103 #endif /* USERLAND */
106 * This is our basic building block.
108 * It can be used in three different ways depending on the value of the ptr
110 * If ptr is NULL, it represents a run of free items.
111 * If ptr points to the unrhdr it represents a run of allocated items.
112 * Otherwise it points to an bitstring of allocated items.
114 * For runs the len field is the length of the run.
115 * For bitmaps the len field represents the number of allocated items.
117 * The bitmap is the same size as struct unr to optimize memory management.
120 TAILQ_ENTRY(unr) list;
127 bitstr_t map[sizeof(struct unr) - 1];
130 CTASSERT(sizeof(struct unr) == sizeof(struct unrb));
132 /* Number of bits in the bitmap */
133 #define NBITS ((int)sizeof(((struct unrb *)NULL)->map) * 8)
135 /* Header element for a unr number space. */
138 TAILQ_HEAD(unrhd,unr) head;
139 u_int low; /* Lowest item */
140 u_int high; /* Highest item */
141 u_int busy; /* Count of allocated items */
142 u_int alloc; /* Count of memory allocations */
143 u_int first; /* items in allocated from start */
144 u_int last; /* items free at end */
149 #if defined(DIAGNOSTIC) || !defined(_KERNEL)
151 * Consistency check function.
153 * Checks the internal consistency as well as we can.
155 * Called at all boundaries of this API.
158 check_unrhdr(struct unrhdr *uh, int line)
166 TAILQ_FOREACH(up, &uh->head, list) {
168 if (up->ptr != uh && up->ptr != NULL) {
170 KASSERT (up->len <= NBITS,
171 ("UNR inconsistency: len %u max %d (line %d)\n",
172 up->len, NBITS, line));
175 for (x = 0; x < up->len; x++)
176 if (bit_test(ub->map, x))
178 KASSERT (w == ub->busy,
179 ("UNR inconsistency: busy %u found %u (line %d)\n",
182 } else if (up->ptr != NULL)
185 KASSERT (y == uh->busy,
186 ("UNR inconsistency: items %u found %u (line %d)\n",
188 KASSERT (z == uh->alloc,
189 ("UNR inconsistency: chunks %u found %u (line %d)\n",
190 uh->alloc, z, line));
196 check_unrhdr(struct unrhdr *uh, int line)
205 * Userland memory management. Just use calloc and keep track of how
206 * many elements we have allocated for check_unrhdr().
209 static __inline void *
210 new_unr(struct unrhdr *uh, void **p1, void **p2)
215 KASSERT(*p1 != NULL || *p2 != NULL, ("Out of cached memory"));
228 delete_unr(struct unrhdr *uh, void *ptr)
236 * Allocate a new unrheader set.
238 * Highest and lowest valid values given as paramters.
242 new_unrhdr(int low, int high, struct lock *lock)
247 ("UNR: use error: new_unrhdr(%u, %u)", low, high));
248 uh = Malloc(sizeof *uh);
252 uh->lock = &unit_lock;
253 TAILQ_INIT(&uh->head);
257 uh->last = 1 + (high - low);
258 check_unrhdr(uh, __LINE__);
263 delete_unrhdr(struct unrhdr *uh)
266 check_unrhdr(uh, __LINE__);
267 KASSERT(uh->busy == 0, ("unrhdr has %u allocations", uh->busy));
268 KASSERT(uh->alloc == 0, ("UNR memory leak in delete_unrhdr"));
273 is_bitmap(struct unrhdr *uh, struct unr *up)
275 return (up->ptr != uh && up->ptr != NULL);
279 * Look for sequence of items which can be combined into a bitmap, if
280 * multiple are present, take the one which saves most memory.
282 * Return (1) if a sequence was found to indicate that another call
283 * might be able to do more. Return (0) if we found no suitable sequence.
285 * NB: called from alloc_unr(), no new memory allocation allowed.
288 optimize_unr(struct unrhdr *uh)
290 struct unr *up, *uf, *us;
291 struct unrb *ub, *ubf;
295 * Look for the run of items (if any) which when collapsed into
296 * a bitmap would save most memory.
300 TAILQ_FOREACH(uf, &uh->head, list) {
301 if (uf->len >= NBITS)
304 if (is_bitmap(uh, uf))
309 up = TAILQ_NEXT(up, list);
312 if ((up->len + l) > NBITS)
315 if (is_bitmap(uh, up))
328 * If the first element is not a bitmap, make it one.
329 * Trying to do so without allocating more memory complicates things
332 if (!is_bitmap(uh, us)) {
333 uf = TAILQ_NEXT(us, list);
334 TAILQ_REMOVE(&uh->head, us, list);
336 l = us->ptr == uh ? 1 : 0;
340 bit_nset(ub->map, 0, a);
343 bit_nclear(ub->map, 0, a);
345 if (!is_bitmap(uh, uf)) {
346 if (uf->ptr == NULL) {
347 bit_nclear(ub->map, a, a + uf->len - 1);
349 bit_nset(ub->map, a, a + uf->len - 1);
357 for (l = 0; l < uf->len; l++, a++) {
358 if (bit_test(ubf->map, l)) {
362 bit_clear(ub->map, a);
366 delete_unr(uh, uf->ptr);
373 uf = TAILQ_NEXT(us, list);
376 if (uf->len + us->len > NBITS)
378 if (uf->ptr == NULL) {
379 bit_nclear(ub->map, us->len, us->len + uf->len - 1);
381 TAILQ_REMOVE(&uh->head, uf, list);
383 } else if (uf->ptr == uh) {
384 bit_nset(ub->map, us->len, us->len + uf->len - 1);
387 TAILQ_REMOVE(&uh->head, uf, list);
391 for (l = 0; l < uf->len; l++, us->len++) {
392 if (bit_test(ubf->map, l)) {
393 bit_set(ub->map, us->len);
396 bit_clear(ub->map, us->len);
399 TAILQ_REMOVE(&uh->head, uf, list);
407 * See if a given unr should be collapsed with a neighbor.
409 * NB: called from alloc_unr(), no new memory allocation allowed.
412 collapse_unr(struct unrhdr *uh, struct unr *up)
417 /* If bitmap is all set or clear, change it to runlength */
418 if (is_bitmap(uh, up)) {
420 if (ub->busy == up->len) {
421 delete_unr(uh, up->ptr);
423 } else if (ub->busy == 0) {
424 delete_unr(uh, up->ptr);
429 /* If nothing left in runlength, delete it */
431 upp = TAILQ_PREV(up, unrhd, list);
433 upp = TAILQ_NEXT(up, list);
434 TAILQ_REMOVE(&uh->head, up, list);
439 /* If we have "hot-spot" still, merge with neighbor if possible */
441 upp = TAILQ_PREV(up, unrhd, list);
442 if (upp != NULL && up->ptr == upp->ptr) {
444 TAILQ_REMOVE(&uh->head, upp, list);
447 upp = TAILQ_NEXT(up, list);
448 if (upp != NULL && up->ptr == upp->ptr) {
450 TAILQ_REMOVE(&uh->head, upp, list);
455 /* Merge into ->first if possible */
456 upp = TAILQ_FIRST(&uh->head);
457 if (upp != NULL && upp->ptr == uh) {
458 uh->first += upp->len;
459 TAILQ_REMOVE(&uh->head, upp, list);
465 /* Merge into ->last if possible */
466 upp = TAILQ_LAST(&uh->head, unrhd);
467 if (upp != NULL && upp->ptr == NULL) {
468 uh->last += upp->len;
469 TAILQ_REMOVE(&uh->head, upp, list);
475 /* Try to make bitmaps */
476 while (optimize_unr(uh))
481 * Allocate a free unr.
484 alloc_unrl(struct unrhdr *uh)
490 struct lock *ml __debugvar = uh->lock;
491 struct thread *td __debugvar = curthread;
493 KKASSERT(lockstatus(ml, td) != 0);
494 check_unrhdr(uh, __LINE__);
495 x = uh->low + uh->first;
497 up = TAILQ_FIRST(&uh->head);
500 * If we have an ideal split, just adjust the first+last
502 if (up == NULL && uh->last > 0) {
510 * We can always allocate from the first list element, so if we have
511 * nothing on the list, we must have run out of unit numbers.
516 KASSERT(up->ptr != uh, ("UNR first element is allocated"));
518 if (up->ptr == NULL) { /* free run */
521 } else { /* bitmap */
523 KASSERT(ub->busy < up->len, ("UNR bitmap confusion"));
524 bit_ffc(ub->map, up->len, &y);
525 KASSERT(y != -1, ("UNR corruption: No clear bit in bitmap."));
531 collapse_unr(uh, up);
536 alloc_unr(struct unrhdr *uh)
540 lockmgr(uh->lock, LK_EXCLUSIVE);
542 lockmgr(uh->lock, LK_RELEASE);
549 * If we can save unrs by using a bitmap, do so.
552 free_unrl(struct unrhdr *uh, u_int item, void **p1, void **p2)
554 struct unr *up, *upp, *upn;
558 KASSERT(item >= uh->low && item <= uh->high,
559 ("UNR: free_unr(%u) out of range [%u...%u]",
560 item, uh->low, uh->high));
561 check_unrhdr(uh, __LINE__);
563 upp = TAILQ_FIRST(&uh->head);
565 * Freeing in the ideal split case
567 if (item + 1 == uh->first && upp == NULL) {
571 check_unrhdr(uh, __LINE__);
575 * Freeing in the ->first section. Create a run starting at the
576 * freed item. The code below will subdivide it.
578 if (item < uh->first) {
579 up = new_unr(uh, p1, p2);
581 up->len = uh->first - item;
582 TAILQ_INSERT_HEAD(&uh->head, up, list);
583 uh->first -= up->len;
588 /* Find the item which contains the unit we want to free */
589 TAILQ_FOREACH(up, &uh->head, list) {
595 /* Handle bitmap items */
596 if (is_bitmap(uh, up)) {
599 KASSERT(bit_test(ub->map, item) != 0,
600 ("UNR: Freeing free item %d (bitmap)\n", item));
601 bit_clear(ub->map, item);
604 collapse_unr(uh, up);
608 KASSERT(up->ptr == uh, ("UNR Freeing free item %d (run))\n", item));
610 /* Just this one left, reap it */
614 collapse_unr(uh, up);
618 /* Check if we can shift the item into the previous 'free' run */
619 upp = TAILQ_PREV(up, unrhd, list);
620 if (item == 0 && upp != NULL && upp->ptr == NULL) {
624 collapse_unr(uh, up);
628 /* Check if we can shift the item to the next 'free' run */
629 upn = TAILQ_NEXT(up, list);
630 if (item == up->len - 1 && upn != NULL && upn->ptr == NULL) {
634 collapse_unr(uh, up);
638 /* Split off the tail end, if any. */
639 pl = up->len - (1 + item);
641 upp = new_unr(uh, p1, p2);
644 TAILQ_INSERT_AFTER(&uh->head, up, upp, list);
647 /* Split off head end, if any */
649 upp = new_unr(uh, p1, p2);
652 TAILQ_INSERT_BEFORE(up, upp, list);
657 collapse_unr(uh, up);
661 free_unr(struct unrhdr *uh, u_int item)
665 p1 = Malloc(sizeof(struct unr));
666 p2 = Malloc(sizeof(struct unr));
667 lockmgr(uh->lock, LK_EXCLUSIVE);
668 free_unrl(uh, item, &p1, &p2);
669 lockmgr(uh->lock, LK_RELEASE);
676 #ifndef _KERNEL /* USERLAND test driver */
679 * Simple stochastic test driver for the above functions
683 print_unr(struct unrhdr *uh, struct unr *up)
688 printf(" %p len = %5u ", up, up->len);
691 else if (up->ptr == uh)
695 printf("bitmap(%d) [", ub->busy);
696 for (x = 0; x < up->len; x++) {
697 if (bit_test(ub->map, x))
707 print_unrhdr(struct unrhdr *uh)
713 "%p low = %u high = %u first = %u last = %u busy %u chunks = %u\n",
714 uh, uh->low, uh->high, uh->first, uh->last, uh->busy, uh->alloc);
715 x = uh->low + uh->first;
716 TAILQ_FOREACH(up, &uh->head, list) {
717 printf(" from = %5u", x);
719 if (up->ptr == NULL || up->ptr == uh)
726 /* Number of unrs to test */
730 main(int argc __unused, const char **argv __unused)
736 setbuf(stdout, NULL);
737 uh = new_unrhdr(0, NN - 1, NULL);
740 memset(a, 0, sizeof a);
742 fprintf(stderr, "sizeof(struct unr) %d\n", sizeof (struct unr));
743 fprintf(stderr, "sizeof(struct unrb) %d\n", sizeof (struct unrb));
744 fprintf(stderr, "sizeof(struct unrhdr) %d\n", sizeof (struct unrhdr));
745 fprintf(stderr, "NBITS %d\n", NBITS);
747 for (m = 0; m < NN * 100; m++) {
767 if (1) /* XXX: change this for detailed debug printout */
769 check_unrhdr(uh, __LINE__);
771 for (i = 0; i < NN; i++) {