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 */
200 TAILQ_HEAD(unrfr
,unr
) ppfree
; /* Items to be freed after mtx
205 #if defined(DIAGNOSTIC) || !defined(_KERNEL)
207 * Consistency check function.
209 * Checks the internal consistency as well as we can.
211 * Called at all boundaries of this API.
214 check_unrhdr(struct unrhdr
*uh
, int line
)
222 TAILQ_FOREACH(up
, &uh
->head
, list
) {
224 if (up
->ptr
!= uh
&& up
->ptr
!= NULL
) {
226 KASSERT (up
->len
<= NBITS
,
227 ("UNR inconsistency: len %u max %d (line %d)\n",
228 up
->len
, NBITS
, line
));
231 for (x
= 0; x
< up
->len
; x
++)
232 if (bit_test(ub
->map
, x
))
234 KASSERT (w
== ub
->busy
,
235 ("UNR inconsistency: busy %u found %u (line %d)\n",
238 } else if (up
->ptr
!= NULL
)
241 KASSERT (y
== uh
->busy
,
242 ("UNR inconsistency: items %u found %u (line %d)\n",
244 KASSERT (z
== uh
->alloc
,
245 ("UNR inconsistency: chunks %u found %u (line %d)\n",
246 uh
->alloc
, z
, line
));
252 check_unrhdr(struct unrhdr
*uh
, int line
)
261 * Userland memory management. Just use calloc and keep track of how
262 * many elements we have allocated for check_unrhdr().
265 static __inline
void *
266 new_unr(struct unrhdr
*uh
, void **p1
, void **p2
)
271 KASSERT(*p1
!= NULL
|| *p2
!= NULL
, ("Out of cached memory"));
284 delete_unr(struct unrhdr
*uh
, void *ptr
)
290 TAILQ_INSERT_TAIL(&uh
->ppfree
, up
, list
);
294 clean_unrhdrl(struct unrhdr
*uh
)
298 mtx_assert(uh
->mtx
, MA_OWNED
);
299 while ((up
= TAILQ_FIRST(&uh
->ppfree
)) != NULL
) {
300 TAILQ_REMOVE(&uh
->ppfree
, up
, list
);
309 clean_unrhdr(struct unrhdr
*uh
)
318 * Allocate a new unrheader set.
320 * Highest and lowest valid values given as paramters.
324 new_unrhdr(int low
, int high
, struct mtx
*mutex
)
329 ("UNR: use error: new_unrhdr(%u, %u)", low
, high
));
330 uh
= Malloc(sizeof *uh
);
335 TAILQ_INIT(&uh
->head
);
336 TAILQ_INIT(&uh
->ppfree
);
340 uh
->last
= 1 + (high
- low
);
341 check_unrhdr(uh
, __LINE__
);
346 delete_unrhdr(struct unrhdr
*uh
)
349 check_unrhdr(uh
, __LINE__
);
350 KASSERT(uh
->busy
== 0, ("unrhdr has %u allocations", uh
->busy
));
351 KASSERT(uh
->alloc
== 0, ("UNR memory leak in delete_unrhdr"));
352 KASSERT(TAILQ_FIRST(&uh
->ppfree
) == NULL
,
353 ("unrhdr has postponed item for free"));
358 is_bitmap(struct unrhdr
*uh
, struct unr
*up
)
360 return (up
->ptr
!= uh
&& up
->ptr
!= NULL
);
364 * Look for sequence of items which can be combined into a bitmap, if
365 * multiple are present, take the one which saves most memory.
367 * Return (1) if a sequence was found to indicate that another call
368 * might be able to do more. Return (0) if we found no suitable sequence.
370 * NB: called from alloc_unr(), no new memory allocation allowed.
373 optimize_unr(struct unrhdr
*uh
)
375 struct unr
*up
, *uf
, *us
;
376 struct unrb
*ub
, *ubf
;
380 * Look for the run of items (if any) which when collapsed into
381 * a bitmap would save most memory.
385 TAILQ_FOREACH(uf
, &uh
->head
, list
) {
386 if (uf
->len
>= NBITS
)
389 if (is_bitmap(uh
, uf
))
394 up
= TAILQ_NEXT(up
, list
);
397 if ((up
->len
+ l
) > NBITS
)
400 if (is_bitmap(uh
, up
))
413 * If the first element is not a bitmap, make it one.
414 * Trying to do so without allocating more memory complicates things
417 if (!is_bitmap(uh
, us
)) {
418 uf
= TAILQ_NEXT(us
, list
);
419 TAILQ_REMOVE(&uh
->head
, us
, list
);
421 l
= us
->ptr
== uh
? 1 : 0;
425 bit_nset(ub
->map
, 0, a
);
428 bit_nclear(ub
->map
, 0, a
);
430 if (!is_bitmap(uh
, uf
)) {
431 if (uf
->ptr
== NULL
) {
432 bit_nclear(ub
->map
, a
, a
+ uf
->len
- 1);
434 bit_nset(ub
->map
, a
, a
+ uf
->len
- 1);
442 for (l
= 0; l
< uf
->len
; l
++, a
++) {
443 if (bit_test(ubf
->map
, l
)) {
447 bit_clear(ub
->map
, a
);
451 delete_unr(uh
, uf
->ptr
);
458 uf
= TAILQ_NEXT(us
, list
);
461 if (uf
->len
+ us
->len
> NBITS
)
463 if (uf
->ptr
== NULL
) {
464 bit_nclear(ub
->map
, us
->len
, us
->len
+ uf
->len
- 1);
466 TAILQ_REMOVE(&uh
->head
, uf
, list
);
468 } else if (uf
->ptr
== uh
) {
469 bit_nset(ub
->map
, us
->len
, us
->len
+ uf
->len
- 1);
472 TAILQ_REMOVE(&uh
->head
, uf
, list
);
476 for (l
= 0; l
< uf
->len
; l
++, us
->len
++) {
477 if (bit_test(ubf
->map
, l
)) {
478 bit_set(ub
->map
, us
->len
);
481 bit_clear(ub
->map
, us
->len
);
484 TAILQ_REMOVE(&uh
->head
, uf
, list
);
492 * See if a given unr should be collapsed with a neighbor.
494 * NB: called from alloc_unr(), no new memory allocation allowed.
497 collapse_unr(struct unrhdr
*uh
, struct unr
*up
)
502 /* If bitmap is all set or clear, change it to runlength */
503 if (is_bitmap(uh
, up
)) {
505 if (ub
->busy
== up
->len
) {
506 delete_unr(uh
, up
->ptr
);
508 } else if (ub
->busy
== 0) {
509 delete_unr(uh
, up
->ptr
);
514 /* If nothing left in runlength, delete it */
516 upp
= TAILQ_PREV(up
, unrhd
, list
);
518 upp
= TAILQ_NEXT(up
, list
);
519 TAILQ_REMOVE(&uh
->head
, up
, list
);
524 /* If we have "hot-spot" still, merge with neighbor if possible */
526 upp
= TAILQ_PREV(up
, unrhd
, list
);
527 if (upp
!= NULL
&& up
->ptr
== upp
->ptr
) {
529 TAILQ_REMOVE(&uh
->head
, upp
, list
);
532 upp
= TAILQ_NEXT(up
, list
);
533 if (upp
!= NULL
&& up
->ptr
== upp
->ptr
) {
535 TAILQ_REMOVE(&uh
->head
, upp
, list
);
540 /* Merge into ->first if possible */
541 upp
= TAILQ_FIRST(&uh
->head
);
542 if (upp
!= NULL
&& upp
->ptr
== uh
) {
543 uh
->first
+= upp
->len
;
544 TAILQ_REMOVE(&uh
->head
, upp
, list
);
550 /* Merge into ->last if possible */
551 upp
= TAILQ_LAST(&uh
->head
, unrhd
);
552 if (upp
!= NULL
&& upp
->ptr
== NULL
) {
553 uh
->last
+= upp
->len
;
554 TAILQ_REMOVE(&uh
->head
, upp
, list
);
560 /* Try to make bitmaps */
561 while (optimize_unr(uh
))
566 * Allocate a free unr.
569 alloc_unrl(struct unrhdr
*uh
)
576 mtx_assert(uh
->mtx
, MA_OWNED
);
577 check_unrhdr(uh
, __LINE__
);
578 x
= uh
->low
+ uh
->first
;
580 up
= TAILQ_FIRST(&uh
->head
);
583 * If we have an ideal split, just adjust the first+last
585 if (up
== NULL
&& uh
->last
> 0) {
593 * We can always allocate from the first list element, so if we have
594 * nothing on the list, we must have run out of unit numbers.
599 KASSERT(up
->ptr
!= uh
, ("UNR first element is allocated"));
601 if (up
->ptr
== NULL
) { /* free run */
604 } else { /* bitmap */
606 KASSERT(ub
->busy
< up
->len
, ("UNR bitmap confusion"));
607 bit_ffc(ub
->map
, up
->len
, &y
);
608 KASSERT(y
!= -1, ("UNR corruption: No clear bit in bitmap."));
614 collapse_unr(uh
, up
);
619 alloc_unr(struct unrhdr
*uh
)
633 * If we can save unrs by using a bitmap, do so.
636 free_unrl(struct unrhdr
*uh
, u_int item
, void **p1
, void **p2
)
638 struct unr
*up
, *upp
, *upn
;
642 KASSERT(item
>= uh
->low
&& item
<= uh
->high
,
643 ("UNR: free_unr(%u) out of range [%u...%u]",
644 item
, uh
->low
, uh
->high
));
645 check_unrhdr(uh
, __LINE__
);
647 upp
= TAILQ_FIRST(&uh
->head
);
649 * Freeing in the ideal split case
651 if (item
+ 1 == uh
->first
&& upp
== NULL
) {
655 check_unrhdr(uh
, __LINE__
);
659 * Freeing in the ->first section. Create a run starting at the
660 * freed item. The code below will subdivide it.
662 if (item
< uh
->first
) {
663 up
= new_unr(uh
, p1
, p2
);
665 up
->len
= uh
->first
- item
;
666 TAILQ_INSERT_HEAD(&uh
->head
, up
, list
);
667 uh
->first
-= up
->len
;
672 /* Find the item which contains the unit we want to free */
673 TAILQ_FOREACH(up
, &uh
->head
, list
) {
679 /* Handle bitmap items */
680 if (is_bitmap(uh
, up
)) {
683 KASSERT(bit_test(ub
->map
, item
) != 0,
684 ("UNR: Freeing free item %d (bitmap)\n", item
));
685 bit_clear(ub
->map
, item
);
688 collapse_unr(uh
, up
);
692 KASSERT(up
->ptr
== uh
, ("UNR Freeing free item %d (run))\n", item
));
694 /* Just this one left, reap it */
698 collapse_unr(uh
, up
);
702 /* Check if we can shift the item into the previous 'free' run */
703 upp
= TAILQ_PREV(up
, unrhd
, list
);
704 if (item
== 0 && upp
!= NULL
&& upp
->ptr
== NULL
) {
708 collapse_unr(uh
, up
);
712 /* Check if we can shift the item to the next 'free' run */
713 upn
= TAILQ_NEXT(up
, list
);
714 if (item
== up
->len
- 1 && upn
!= NULL
&& upn
->ptr
== NULL
) {
718 collapse_unr(uh
, up
);
722 /* Split off the tail end, if any. */
723 pl
= up
->len
- (1 + item
);
725 upp
= new_unr(uh
, p1
, p2
);
728 TAILQ_INSERT_AFTER(&uh
->head
, up
, upp
, list
);
731 /* Split off head end, if any */
733 upp
= new_unr(uh
, p1
, p2
);
736 TAILQ_INSERT_BEFORE(up
, upp
, list
);
741 collapse_unr(uh
, up
);
745 free_unr(struct unrhdr
*uh
, u_int item
)
749 WITNESS_WARN(WARN_GIANTOK
| WARN_SLEEPOK
, NULL
, "free_unr");
750 p1
= Malloc(sizeof(struct unr
));
751 p2
= Malloc(sizeof(struct unr
));
753 free_unrl(uh
, item
, &p1
, &p2
);
762 #ifndef _KERNEL /* USERLAND test driver */
765 * Simple stochastic test driver for the above functions
769 print_unr(struct unrhdr
*uh
, struct unr
*up
)
774 printf(" %p len = %5u ", up
, up
->len
);
777 else if (up
->ptr
== uh
)
781 printf("bitmap(%d) [", ub
->busy
);
782 for (x
= 0; x
< up
->len
; x
++) {
783 if (bit_test(ub
->map
, x
))
793 print_unrhdr(struct unrhdr
*uh
)
799 "%p low = %u high = %u first = %u last = %u busy %u chunks = %u\n",
800 uh
, uh
->low
, uh
->high
, uh
->first
, uh
->last
, uh
->busy
, uh
->alloc
);
801 x
= uh
->low
+ uh
->first
;
802 TAILQ_FOREACH(up
, &uh
->head
, list
) {
803 printf(" from = %5u", x
);
805 if (up
->ptr
== NULL
|| up
->ptr
== uh
)
812 /* Number of unrs to test */
816 main(int argc __unused
, const char **argv __unused
)
822 setbuf(stdout
, NULL
);
823 uh
= new_unrhdr(0, NN
- 1, NULL
);
826 memset(a
, 0, sizeof a
);
828 fprintf(stderr
, "sizeof(struct unr) %d\n", sizeof (struct unr
));
829 fprintf(stderr
, "sizeof(struct unrb) %d\n", sizeof (struct unrb
));
830 fprintf(stderr
, "sizeof(struct unrhdr) %d\n", sizeof (struct unrhdr
));
831 fprintf(stderr
, "NBITS %d\n", NBITS
);
833 for (m
= 0; m
< NN
* 100; m
++) {
853 if (1) /* XXX: change this for detailed debug printout */
855 check_unrhdr(uh
, __LINE__
);
857 for (i
= 0; i
< NN
; i
++) {