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
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 i386.
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 i386 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 i386 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>
82 #include <sys/mutex.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) malloc(foo, M_UNIT, M_WAITOK | M_ZERO)
93 #define Free(foo) free(foo, M_UNIT)
95 static struct mtx unitmtx;
97 MTX_SYSINIT(unit, &unitmtx, "unit# allocation", MTX_DEF);
99 #else /* ...USERLAND */
105 #define KASSERT(cond, arg) \
114 #define Malloc(foo) _Malloc(foo, __LINE__)
116 _Malloc(size_t foo, int line)
119 KASSERT(no_alloc == 0, ("malloc in wrong place() line %d", line));
120 return (calloc(foo, 1));
122 #define Free(foo) free(foo)
132 mtx_lock(struct mtx *mp)
134 KASSERT(mp->state == 0, ("mutex already locked"));
139 mtx_unlock(struct mtx *mp)
141 KASSERT(mp->state == 1, ("mutex not locked"));
148 mtx_assert(struct mtx *mp, int flag)
150 if (flag == MA_OWNED) {
151 KASSERT(mp->state == 1, ("mtx_assert(MA_OWNED) not true"));
155 #define CTASSERT(foo)
157 #endif /* USERLAND */
160 * This is our basic building block.
162 * It can be used in three different ways depending on the value of the ptr
164 * If ptr is NULL, it represents a run of free items.
165 * If ptr points to the unrhdr it represents a run of allocated items.
166 * Otherwise it points to an bitstring of allocated items.
168 * For runs the len field is the length of the run.
169 * For bitmaps the len field represents the number of allocated items.
171 * The bitmap is the same size as struct unr to optimize memory management.
174 TAILQ_ENTRY(unr) list;
181 bitstr_t map[sizeof(struct unr) - 1];
184 CTASSERT(sizeof(struct unr) == sizeof(struct unrb));
186 /* Number of bits in the bitmap */
187 #define NBITS ((int)sizeof(((struct unrb *)NULL)->map) * 8)
189 /* Header element for a unr number space. */
192 TAILQ_HEAD(unrhd,unr) head;
193 u_int low; /* Lowest item */
194 u_int high; /* Highest item */
195 u_int busy; /* Count of allocated items */
196 u_int alloc; /* Count of memory allocations */
197 u_int first; /* items in allocated from start */
198 u_int last; /* items free at end */
203 #if defined(DIAGNOSTIC) || !defined(_KERNEL)
205 * Consistency check function.
207 * Checks the internal consistency as well as we can.
209 * Called at all boundaries of this API.
212 check_unrhdr(struct unrhdr *uh, int line)
220 TAILQ_FOREACH(up, &uh->head, list) {
222 if (up->ptr != uh && up->ptr != NULL) {
224 KASSERT (up->len <= NBITS,
225 ("UNR inconsistency: len %u max %d (line %d)\n",
226 up->len, NBITS, line));
229 for (x = 0; x < up->len; x++)
230 if (bit_test(ub->map, x))
232 KASSERT (w == ub->busy,
233 ("UNR inconsistency: busy %u found %u (line %d)\n",
236 } else if (up->ptr != NULL)
239 KASSERT (y == uh->busy,
240 ("UNR inconsistency: items %u found %u (line %d)\n",
242 KASSERT (z == uh->alloc,
243 ("UNR inconsistency: chunks %u found %u (line %d)\n",
244 uh->alloc, z, line));
250 check_unrhdr(struct unrhdr *uh, int line)
259 * Userland memory management. Just use calloc and keep track of how
260 * many elements we have allocated for check_unrhdr().
263 static __inline void *
264 new_unr(struct unrhdr *uh, void **p1, void **p2)
269 KASSERT(*p1 != NULL || *p2 != NULL, ("Out of cached memory"));
282 delete_unr(struct unrhdr *uh, void *ptr)
290 * Allocate a new unrheader set.
292 * Highest and lowest valid values given as paramters.
296 new_unrhdr(int low, int high, struct mtx *mutex)
301 ("UNR: use error: new_unrhdr(%u, %u)", low, high));
302 uh = Malloc(sizeof *uh);
307 TAILQ_INIT(&uh->head);
311 uh->last = 1 + (high - low);
312 check_unrhdr(uh, __LINE__);
317 delete_unrhdr(struct unrhdr *uh)
320 check_unrhdr(uh, __LINE__);
321 KASSERT(uh->busy == 0, ("unrhdr has %u allocations", uh->busy));
322 KASSERT(uh->alloc == 0, ("UNR memory leak in delete_unrhdr"));
327 is_bitmap(struct unrhdr *uh, struct unr *up)
329 return (up->ptr != uh && up->ptr != NULL);
333 * Look for sequence of items which can be combined into a bitmap, if
334 * multiple are present, take the one which saves most memory.
336 * Return (1) if a sequence was found to indicate that another call
337 * might be able to do more. Return (0) if we found no suitable sequence.
339 * NB: called from alloc_unr(), no new memory allocation allowed.
342 optimize_unr(struct unrhdr *uh)
344 struct unr *up, *uf, *us;
345 struct unrb *ub, *ubf;
349 * Look for the run of items (if any) which when collapsed into
350 * a bitmap would save most memory.
354 TAILQ_FOREACH(uf, &uh->head, list) {
355 if (uf->len >= NBITS)
358 if (is_bitmap(uh, uf))
363 up = TAILQ_NEXT(up, list);
366 if ((up->len + l) > NBITS)
369 if (is_bitmap(uh, up))
382 * If the first element is not a bitmap, make it one.
383 * Trying to do so without allocating more memory complicates things
386 if (!is_bitmap(uh, us)) {
387 uf = TAILQ_NEXT(us, list);
388 TAILQ_REMOVE(&uh->head, us, list);
390 l = us->ptr == uh ? 1 : 0;
394 bit_nset(ub->map, 0, a);
397 bit_nclear(ub->map, 0, a);
399 if (!is_bitmap(uh, uf)) {
400 if (uf->ptr == NULL) {
401 bit_nclear(ub->map, a, a + uf->len - 1);
403 bit_nset(ub->map, a, a + uf->len - 1);
411 for (l = 0; l < uf->len; l++, a++) {
412 if (bit_test(ubf->map, l)) {
416 bit_clear(ub->map, a);
420 delete_unr(uh, uf->ptr);
427 uf = TAILQ_NEXT(us, list);
430 if (uf->len + us->len > NBITS)
432 if (uf->ptr == NULL) {
433 bit_nclear(ub->map, us->len, us->len + uf->len - 1);
435 TAILQ_REMOVE(&uh->head, uf, list);
437 } else if (uf->ptr == uh) {
438 bit_nset(ub->map, us->len, us->len + uf->len - 1);
441 TAILQ_REMOVE(&uh->head, uf, list);
445 for (l = 0; l < uf->len; l++, us->len++) {
446 if (bit_test(ubf->map, l)) {
447 bit_set(ub->map, us->len);
450 bit_clear(ub->map, us->len);
453 TAILQ_REMOVE(&uh->head, uf, list);
461 * See if a given unr should be collapsed with a neighbor.
463 * NB: called from alloc_unr(), no new memory allocation allowed.
466 collapse_unr(struct unrhdr *uh, struct unr *up)
471 /* If bitmap is all set or clear, change it to runlength */
472 if (is_bitmap(uh, up)) {
474 if (ub->busy == up->len) {
475 delete_unr(uh, up->ptr);
477 } else if (ub->busy == 0) {
478 delete_unr(uh, up->ptr);
483 /* If nothing left in runlength, delete it */
485 upp = TAILQ_PREV(up, unrhd, list);
487 upp = TAILQ_NEXT(up, list);
488 TAILQ_REMOVE(&uh->head, up, list);
493 /* If we have "hot-spot" still, merge with neighbor if possible */
495 upp = TAILQ_PREV(up, unrhd, list);
496 if (upp != NULL && up->ptr == upp->ptr) {
498 TAILQ_REMOVE(&uh->head, upp, list);
501 upp = TAILQ_NEXT(up, list);
502 if (upp != NULL && up->ptr == upp->ptr) {
504 TAILQ_REMOVE(&uh->head, upp, list);
509 /* Merge into ->first if possible */
510 upp = TAILQ_FIRST(&uh->head);
511 if (upp != NULL && upp->ptr == uh) {
512 uh->first += upp->len;
513 TAILQ_REMOVE(&uh->head, upp, list);
519 /* Merge into ->last if possible */
520 upp = TAILQ_LAST(&uh->head, unrhd);
521 if (upp != NULL && upp->ptr == NULL) {
522 uh->last += upp->len;
523 TAILQ_REMOVE(&uh->head, upp, list);
529 /* Try to make bitmaps */
530 while (optimize_unr(uh))
535 * Allocate a free unr.
538 alloc_unrl(struct unrhdr *uh)
545 mtx_assert(uh->mtx, MA_OWNED);
546 check_unrhdr(uh, __LINE__);
547 x = uh->low + uh->first;
549 up = TAILQ_FIRST(&uh->head);
552 * If we have an ideal split, just adjust the first+last
554 if (up == NULL && uh->last > 0) {
562 * We can always allocate from the first list element, so if we have
563 * nothing on the list, we must have run out of unit numbers.
568 KASSERT(up->ptr != uh, ("UNR first element is allocated"));
570 if (up->ptr == NULL) { /* free run */
573 } else { /* bitmap */
575 KASSERT(ub->busy < up->len, ("UNR bitmap confusion"));
576 bit_ffc(ub->map, up->len, &y);
577 KASSERT(y != -1, ("UNR corruption: No clear bit in bitmap."));
583 collapse_unr(uh, up);
588 alloc_unr(struct unrhdr *uh)
601 * If we can save unrs by using a bitmap, do so.
604 free_unrl(struct unrhdr *uh, u_int item, void **p1, void **p2)
606 struct unr *up, *upp, *upn;
610 KASSERT(item >= uh->low && item <= uh->high,
611 ("UNR: free_unr(%u) out of range [%u...%u]",
612 item, uh->low, uh->high));
613 check_unrhdr(uh, __LINE__);
615 upp = TAILQ_FIRST(&uh->head);
617 * Freeing in the ideal split case
619 if (item + 1 == uh->first && upp == NULL) {
623 check_unrhdr(uh, __LINE__);
627 * Freeing in the ->first section. Create a run starting at the
628 * freed item. The code below will subdivide it.
630 if (item < uh->first) {
631 up = new_unr(uh, p1, p2);
633 up->len = uh->first - item;
634 TAILQ_INSERT_HEAD(&uh->head, up, list);
635 uh->first -= up->len;
640 /* Find the item which contains the unit we want to free */
641 TAILQ_FOREACH(up, &uh->head, list) {
647 /* Handle bitmap items */
648 if (is_bitmap(uh, up)) {
651 KASSERT(bit_test(ub->map, item) != 0,
652 ("UNR: Freeing free item %d (bitmap)\n", item));
653 bit_clear(ub->map, item);
656 collapse_unr(uh, up);
660 KASSERT(up->ptr == uh, ("UNR Freeing free item %d (run))\n", item));
662 /* Just this one left, reap it */
666 collapse_unr(uh, up);
670 /* Check if we can shift the item into the previous 'free' run */
671 upp = TAILQ_PREV(up, unrhd, list);
672 if (item == 0 && upp != NULL && upp->ptr == NULL) {
676 collapse_unr(uh, up);
680 /* Check if we can shift the item to the next 'free' run */
681 upn = TAILQ_NEXT(up, list);
682 if (item == up->len - 1 && upn != NULL && upn->ptr == NULL) {
686 collapse_unr(uh, up);
690 /* Split off the tail end, if any. */
691 pl = up->len - (1 + item);
693 upp = new_unr(uh, p1, p2);
696 TAILQ_INSERT_AFTER(&uh->head, up, upp, list);
699 /* Split off head end, if any */
701 upp = new_unr(uh, p1, p2);
704 TAILQ_INSERT_BEFORE(up, upp, list);
709 collapse_unr(uh, up);
713 free_unr(struct unrhdr *uh, u_int item)
717 p1 = Malloc(sizeof(struct unr));
718 p2 = Malloc(sizeof(struct unr));
720 free_unrl(uh, item, &p1, &p2);
728 #ifndef _KERNEL /* USERLAND test driver */
731 * Simple stochastic test driver for the above functions
735 print_unr(struct unrhdr *uh, struct unr *up)
740 printf(" %p len = %5u ", up, up->len);
743 else if (up->ptr == uh)
747 printf("bitmap(%d) [", ub->busy);
748 for (x = 0; x < up->len; x++) {
749 if (bit_test(ub->map, x))
759 print_unrhdr(struct unrhdr *uh)
765 "%p low = %u high = %u first = %u last = %u busy %u chunks = %u\n",
766 uh, uh->low, uh->high, uh->first, uh->last, uh->busy, uh->alloc);
767 x = uh->low + uh->first;
768 TAILQ_FOREACH(up, &uh->head, list) {
769 printf(" from = %5u", x);
771 if (up->ptr == NULL || up->ptr == uh)
778 /* Number of unrs to test */
782 main(int argc __unused, const char **argv __unused)
788 setbuf(stdout, NULL);
789 uh = new_unrhdr(0, NN - 1, NULL);
792 memset(a, 0, sizeof a);
794 fprintf(stderr, "sizeof(struct unr) %d\n", sizeof (struct unr));
795 fprintf(stderr, "sizeof(struct unrb) %d\n", sizeof (struct unrb));
796 fprintf(stderr, "sizeof(struct unrhdr) %d\n", sizeof (struct unrhdr));
797 fprintf(stderr, "NBITS %d\n", NBITS);
799 for (m = 0; m < NN * 100; m++) {
819 if (1) /* XXX: change this for detailed debug printout */
821 check_unrhdr(uh, __LINE__);
823 for (i = 0; i < NN; i++) {