1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2017-2018 Intel Corporation
14 #include <rte_common.h>
16 #include <rte_errno.h>
17 #include <rte_spinlock.h>
18 #include <rte_tailq.h>
20 #include "eal_filesystem.h"
21 #include "eal_private.h"
23 #include "rte_fbarray.h"
25 #define MASK_SHIFT 6ULL
26 #define MASK_ALIGN (1ULL << MASK_SHIFT)
27 #define MASK_LEN_TO_IDX(x) ((x) >> MASK_SHIFT)
28 #define MASK_LEN_TO_MOD(x) ((x) - RTE_ALIGN_FLOOR(x, MASK_ALIGN))
29 #define MASK_GET_IDX(idx, mod) ((idx << MASK_SHIFT) + mod)
32 * We use this to keep track of created/attached memory areas to prevent user
33 * errors in API usage.
36 TAILQ_ENTRY(mem_area
) next
;
41 TAILQ_HEAD(mem_area_head
, mem_area
);
42 /* local per-process tailq */
43 static struct mem_area_head mem_area_tailq
=
44 TAILQ_HEAD_INITIALIZER(mem_area_tailq
);
45 static rte_spinlock_t mem_area_lock
= RTE_SPINLOCK_INITIALIZER
;
48 * This is a mask that is always stored at the end of array, to provide fast
49 * way of finding free/used spots without looping through each element.
58 calc_mask_size(unsigned int len
)
60 /* mask must be multiple of MASK_ALIGN, even though length of array
61 * itself may not be aligned on that boundary.
63 len
= RTE_ALIGN_CEIL(len
, MASK_ALIGN
);
64 return sizeof(struct used_mask
) +
65 sizeof(uint64_t) * MASK_LEN_TO_IDX(len
);
69 calc_data_size(size_t page_sz
, unsigned int elt_sz
, unsigned int len
)
71 size_t data_sz
= elt_sz
* len
;
72 size_t msk_sz
= calc_mask_size(len
);
73 return RTE_ALIGN_CEIL(data_sz
+ msk_sz
, page_sz
);
76 static struct used_mask
*
77 get_used_mask(void *data
, unsigned int elt_sz
, unsigned int len
)
79 return (struct used_mask
*) RTE_PTR_ADD(data
, elt_sz
* len
);
83 resize_and_map(int fd
, void *addr
, size_t len
)
88 if (ftruncate(fd
, len
)) {
89 RTE_LOG(ERR
, EAL
, "Cannot truncate %s\n", path
);
90 /* pass errno up the chain */
95 map_addr
= mmap(addr
, len
, PROT_READ
| PROT_WRITE
,
96 MAP_SHARED
| MAP_FIXED
, fd
, 0);
97 if (map_addr
!= addr
) {
98 RTE_LOG(ERR
, EAL
, "mmap() failed: %s\n", strerror(errno
));
99 /* pass errno up the chain */
107 overlap(const struct mem_area
*ma
, const void *start
, size_t len
)
109 const void *end
= RTE_PTR_ADD(start
, len
);
110 const void *ma_start
= ma
->addr
;
111 const void *ma_end
= RTE_PTR_ADD(ma
->addr
, ma
->len
);
114 if (start
>= ma_start
&& start
< ma_end
)
117 if (end
>= ma_start
&& end
< ma_end
)
123 find_next_n(const struct rte_fbarray
*arr
, unsigned int start
, unsigned int n
,
126 const struct used_mask
*msk
= get_used_mask(arr
->data
, arr
->elt_sz
,
128 unsigned int msk_idx
, lookahead_idx
, first
, first_mod
;
129 unsigned int last
, last_mod
;
130 uint64_t last_msk
, ignore_msk
;
133 * mask only has granularity of MASK_ALIGN, but start may not be aligned
134 * on that boundary, so construct a special mask to exclude anything we
135 * don't want to see to avoid confusing ctz.
137 first
= MASK_LEN_TO_IDX(start
);
138 first_mod
= MASK_LEN_TO_MOD(start
);
139 ignore_msk
= ~((1ULL << first_mod
) - 1);
141 /* array length may not be aligned, so calculate ignore mask for last
144 last
= MASK_LEN_TO_IDX(arr
->len
);
145 last_mod
= MASK_LEN_TO_MOD(arr
->len
);
146 last_msk
= ~(-1ULL << last_mod
);
148 for (msk_idx
= first
; msk_idx
< msk
->n_masks
; msk_idx
++) {
149 uint64_t cur_msk
, lookahead_msk
;
150 unsigned int run_start
, clz
, left
;
153 * The process of getting n consecutive bits for arbitrary n is
154 * a bit involved, but here it is in a nutshell:
156 * 1. let n be the number of consecutive bits we're looking for
157 * 2. check if n can fit in one mask, and if so, do n-1
158 * rshift-ands to see if there is an appropriate run inside
160 * 2a. if we found a run, bail out early
161 * 2b. if we didn't find a run, proceed
162 * 3. invert the mask and count leading zeroes (that is, count
163 * how many consecutive set bits we had starting from the
164 * end of current mask) as k
165 * 3a. if k is 0, continue to next mask
166 * 3b. if k is not 0, we have a potential run
167 * 4. to satisfy our requirements, next mask must have n-k
168 * consecutive set bits right at the start, so we will do
169 * (n-k-1) rshift-ands and check if first bit is set.
171 * Step 4 will need to be repeated if (n-k) > MASK_ALIGN until
172 * we either run out of masks, lose the run, or find what we
175 cur_msk
= msk
->data
[msk_idx
];
178 /* if we're looking for free spaces, invert the mask */
182 /* combine current ignore mask with last index ignore mask */
184 ignore_msk
|= last_msk
;
186 /* if we have an ignore mask, ignore once */
188 cur_msk
&= ignore_msk
;
192 /* if n can fit in within a single mask, do a search */
193 if (n
<= MASK_ALIGN
) {
194 uint64_t tmp_msk
= cur_msk
;
196 for (s_idx
= 0; s_idx
< n
- 1; s_idx
++)
197 tmp_msk
&= tmp_msk
>> 1ULL;
198 /* we found what we were looking for */
200 run_start
= __builtin_ctzll(tmp_msk
);
201 return MASK_GET_IDX(msk_idx
, run_start
);
206 * we didn't find our run within the mask, or n > MASK_ALIGN,
207 * so we're going for plan B.
210 /* count leading zeroes on inverted mask */
212 clz
= sizeof(cur_msk
) * 8;
214 clz
= __builtin_clzll(~cur_msk
);
216 /* if there aren't any runs at the end either, just continue */
220 /* we have a partial run at the end, so try looking ahead */
221 run_start
= MASK_ALIGN
- clz
;
224 for (lookahead_idx
= msk_idx
+ 1; lookahead_idx
< msk
->n_masks
;
226 unsigned int s_idx
, need
;
227 lookahead_msk
= msk
->data
[lookahead_idx
];
229 /* if we're looking for free space, invert the mask */
231 lookahead_msk
= ~lookahead_msk
;
233 /* figure out how many consecutive bits we need here */
234 need
= RTE_MIN(left
, MASK_ALIGN
);
236 for (s_idx
= 0; s_idx
< need
- 1; s_idx
++)
237 lookahead_msk
&= lookahead_msk
>> 1ULL;
239 /* if first bit is not set, we've lost the run */
240 if ((lookahead_msk
& 1) == 0) {
242 * we've scanned this far, so we know there are
243 * no runs in the space we've lookahead-scanned
244 * as well, so skip that on next iteration.
246 ignore_msk
= ~((1ULL << need
) - 1);
247 msk_idx
= lookahead_idx
;
253 /* check if we've found what we were looking for */
260 /* we didn't find anything, so continue */
264 return MASK_GET_IDX(msk_idx
, run_start
);
266 /* we didn't find anything */
267 rte_errno
= used
? ENOENT
: ENOSPC
;
272 find_next(const struct rte_fbarray
*arr
, unsigned int start
, bool used
)
274 const struct used_mask
*msk
= get_used_mask(arr
->data
, arr
->elt_sz
,
276 unsigned int idx
, first
, first_mod
;
277 unsigned int last
, last_mod
;
278 uint64_t last_msk
, ignore_msk
;
281 * mask only has granularity of MASK_ALIGN, but start may not be aligned
282 * on that boundary, so construct a special mask to exclude anything we
283 * don't want to see to avoid confusing ctz.
285 first
= MASK_LEN_TO_IDX(start
);
286 first_mod
= MASK_LEN_TO_MOD(start
);
287 ignore_msk
= ~((1ULL << first_mod
) - 1ULL);
289 /* array length may not be aligned, so calculate ignore mask for last
292 last
= MASK_LEN_TO_IDX(arr
->len
);
293 last_mod
= MASK_LEN_TO_MOD(arr
->len
);
294 last_msk
= ~(-(1ULL) << last_mod
);
296 for (idx
= first
; idx
< msk
->n_masks
; idx
++) {
297 uint64_t cur
= msk
->data
[idx
];
300 /* if we're looking for free entries, invert mask */
307 /* ignore everything before start on first iteration */
311 /* check if we have any entries */
316 * find first set bit - that will correspond to whatever it is
317 * that we're looking for.
319 found
= __builtin_ctzll(cur
);
320 return MASK_GET_IDX(idx
, found
);
322 /* we didn't find anything */
323 rte_errno
= used
? ENOENT
: ENOSPC
;
328 find_contig(const struct rte_fbarray
*arr
, unsigned int start
, bool used
)
330 const struct used_mask
*msk
= get_used_mask(arr
->data
, arr
->elt_sz
,
332 unsigned int idx
, first
, first_mod
;
333 unsigned int last
, last_mod
;
335 unsigned int need_len
, result
= 0;
337 /* array length may not be aligned, so calculate ignore mask for last
340 last
= MASK_LEN_TO_IDX(arr
->len
);
341 last_mod
= MASK_LEN_TO_MOD(arr
->len
);
342 last_msk
= ~(-(1ULL) << last_mod
);
344 first
= MASK_LEN_TO_IDX(start
);
345 first_mod
= MASK_LEN_TO_MOD(start
);
346 for (idx
= first
; idx
< msk
->n_masks
; idx
++, result
+= need_len
) {
347 uint64_t cur
= msk
->data
[idx
];
348 unsigned int run_len
;
350 need_len
= MASK_ALIGN
;
352 /* if we're looking for free entries, invert mask */
356 /* if this is last mask, ignore everything after last bit */
360 /* ignore everything before start on first iteration */
363 /* at the start, we don't need the full mask len */
364 need_len
-= first_mod
;
367 /* we will be looking for zeroes, so invert the mask */
370 /* if mask is zero, we have a complete run */
375 * see if current run ends before mask end.
377 run_len
= __builtin_ctzll(cur
);
379 /* add however many zeroes we've had in the last run and quit */
380 if (run_len
< need_len
) {
389 find_prev_n(const struct rte_fbarray
*arr
, unsigned int start
, unsigned int n
,
392 const struct used_mask
*msk
= get_used_mask(arr
->data
, arr
->elt_sz
,
394 unsigned int msk_idx
, lookbehind_idx
, first
, first_mod
;
398 * mask only has granularity of MASK_ALIGN, but start may not be aligned
399 * on that boundary, so construct a special mask to exclude anything we
400 * don't want to see to avoid confusing ctz.
402 first
= MASK_LEN_TO_IDX(start
);
403 first_mod
= MASK_LEN_TO_MOD(start
);
404 /* we're going backwards, so mask must start from the top */
405 ignore_msk
= first_mod
== MASK_ALIGN
- 1 ?
406 -1ULL : /* prevent overflow */
407 ~(-1ULL << (first_mod
+ 1));
409 /* go backwards, include zero */
412 uint64_t cur_msk
, lookbehind_msk
;
413 unsigned int run_start
, run_end
, ctz
, left
;
416 * The process of getting n consecutive bits from the top for
417 * arbitrary n is a bit involved, but here it is in a nutshell:
419 * 1. let n be the number of consecutive bits we're looking for
420 * 2. check if n can fit in one mask, and if so, do n-1
421 * lshift-ands to see if there is an appropriate run inside
423 * 2a. if we found a run, bail out early
424 * 2b. if we didn't find a run, proceed
425 * 3. invert the mask and count trailing zeroes (that is, count
426 * how many consecutive set bits we had starting from the
427 * start of current mask) as k
428 * 3a. if k is 0, continue to next mask
429 * 3b. if k is not 0, we have a potential run
430 * 4. to satisfy our requirements, next mask must have n-k
431 * consecutive set bits at the end, so we will do (n-k-1)
432 * lshift-ands and check if last bit is set.
434 * Step 4 will need to be repeated if (n-k) > MASK_ALIGN until
435 * we either run out of masks, lose the run, or find what we
438 cur_msk
= msk
->data
[msk_idx
];
441 /* if we're looking for free spaces, invert the mask */
445 /* if we have an ignore mask, ignore once */
447 cur_msk
&= ignore_msk
;
451 /* if n can fit in within a single mask, do a search */
452 if (n
<= MASK_ALIGN
) {
453 uint64_t tmp_msk
= cur_msk
;
455 for (s_idx
= 0; s_idx
< n
- 1; s_idx
++)
456 tmp_msk
&= tmp_msk
<< 1ULL;
457 /* we found what we were looking for */
459 /* clz will give us offset from end of mask, and
460 * we only get the end of our run, not start,
461 * so adjust result to point to where start
464 run_start
= MASK_ALIGN
-
465 __builtin_clzll(tmp_msk
) - n
;
466 return MASK_GET_IDX(msk_idx
, run_start
);
471 * we didn't find our run within the mask, or n > MASK_ALIGN,
472 * so we're going for plan B.
475 /* count trailing zeroes on inverted mask */
477 ctz
= sizeof(cur_msk
) * 8;
479 ctz
= __builtin_ctzll(~cur_msk
);
481 /* if there aren't any runs at the start either, just
487 /* we have a partial run at the start, so try looking behind */
488 run_end
= MASK_GET_IDX(msk_idx
, ctz
);
491 /* go backwards, include zero */
492 lookbehind_idx
= msk_idx
- 1;
494 /* we can't lookbehind as we've run out of masks, so stop */
499 const uint64_t last_bit
= 1ULL << (MASK_ALIGN
- 1);
500 unsigned int s_idx
, need
;
502 lookbehind_msk
= msk
->data
[lookbehind_idx
];
504 /* if we're looking for free space, invert the mask */
506 lookbehind_msk
= ~lookbehind_msk
;
508 /* figure out how many consecutive bits we need here */
509 need
= RTE_MIN(left
, MASK_ALIGN
);
511 for (s_idx
= 0; s_idx
< need
- 1; s_idx
++)
512 lookbehind_msk
&= lookbehind_msk
<< 1ULL;
514 /* if last bit is not set, we've lost the run */
515 if ((lookbehind_msk
& last_bit
) == 0) {
517 * we've scanned this far, so we know there are
518 * no runs in the space we've lookbehind-scanned
519 * as well, so skip that on next iteration.
521 ignore_msk
= -1ULL << need
;
522 msk_idx
= lookbehind_idx
;
528 /* check if we've found what we were looking for */
533 } while ((lookbehind_idx
--) != 0); /* decrement after check to
537 /* we didn't find anything, so continue */
541 /* we've found what we were looking for, but we only know where
542 * the run ended, so calculate start position.
545 } while (msk_idx
-- != 0); /* decrement after check to include zero */
546 /* we didn't find anything */
547 rte_errno
= used
? ENOENT
: ENOSPC
;
552 find_prev(const struct rte_fbarray
*arr
, unsigned int start
, bool used
)
554 const struct used_mask
*msk
= get_used_mask(arr
->data
, arr
->elt_sz
,
556 unsigned int idx
, first
, first_mod
;
560 * mask only has granularity of MASK_ALIGN, but start may not be aligned
561 * on that boundary, so construct a special mask to exclude anything we
562 * don't want to see to avoid confusing clz.
564 first
= MASK_LEN_TO_IDX(start
);
565 first_mod
= MASK_LEN_TO_MOD(start
);
566 /* we're going backwards, so mask must start from the top */
567 ignore_msk
= first_mod
== MASK_ALIGN
- 1 ?
568 -1ULL : /* prevent overflow */
569 ~(-1ULL << (first_mod
+ 1));
571 /* go backwards, include zero */
574 uint64_t cur
= msk
->data
[idx
];
577 /* if we're looking for free entries, invert mask */
581 /* ignore everything before start on first iteration */
585 /* check if we have any entries */
590 * find last set bit - that will correspond to whatever it is
591 * that we're looking for. we're counting trailing zeroes, thus
592 * the value we get is counted from end of mask, so calculate
593 * position from start of mask.
595 found
= MASK_ALIGN
- __builtin_clzll(cur
) - 1;
597 return MASK_GET_IDX(idx
, found
);
598 } while (idx
-- != 0); /* decrement after check to include zero*/
600 /* we didn't find anything */
601 rte_errno
= used
? ENOENT
: ENOSPC
;
606 find_rev_contig(const struct rte_fbarray
*arr
, unsigned int start
, bool used
)
608 const struct used_mask
*msk
= get_used_mask(arr
->data
, arr
->elt_sz
,
610 unsigned int idx
, first
, first_mod
;
611 unsigned int need_len
, result
= 0;
613 first
= MASK_LEN_TO_IDX(start
);
614 first_mod
= MASK_LEN_TO_MOD(start
);
616 /* go backwards, include zero */
619 uint64_t cur
= msk
->data
[idx
];
620 unsigned int run_len
;
622 need_len
= MASK_ALIGN
;
624 /* if we're looking for free entries, invert mask */
628 /* ignore everything after start on first iteration */
630 unsigned int end_len
= MASK_ALIGN
- first_mod
- 1;
632 /* at the start, we don't need the full mask len */
636 /* we will be looking for zeroes, so invert the mask */
639 /* if mask is zero, we have a complete run */
644 * see where run ends, starting from the end.
646 run_len
= __builtin_clzll(cur
);
648 /* add however many zeroes we've had in the last run and quit */
649 if (run_len
< need_len
) {
655 } while (idx
-- != 0); /* decrement after check to include zero */
660 set_used(struct rte_fbarray
*arr
, unsigned int idx
, bool used
)
662 struct used_mask
*msk
;
663 uint64_t msk_bit
= 1ULL << MASK_LEN_TO_MOD(idx
);
664 unsigned int msk_idx
= MASK_LEN_TO_IDX(idx
);
668 if (arr
== NULL
|| idx
>= arr
->len
) {
672 msk
= get_used_mask(arr
->data
, arr
->elt_sz
, arr
->len
);
675 /* prevent array from changing under us */
676 rte_rwlock_write_lock(&arr
->rwlock
);
678 already_used
= (msk
->data
[msk_idx
] & msk_bit
) != 0;
680 /* nothing to be done */
681 if (used
== already_used
)
685 msk
->data
[msk_idx
] |= msk_bit
;
688 msk
->data
[msk_idx
] &= ~msk_bit
;
692 rte_rwlock_write_unlock(&arr
->rwlock
);
698 fully_validate(const char *name
, unsigned int elt_sz
, unsigned int len
)
700 if (name
== NULL
|| elt_sz
== 0 || len
== 0 || len
> INT_MAX
) {
705 if (strnlen(name
, RTE_FBARRAY_NAME_LEN
) == RTE_FBARRAY_NAME_LEN
) {
706 rte_errno
= ENAMETOOLONG
;
712 int __rte_experimental
713 rte_fbarray_init(struct rte_fbarray
*arr
, const char *name
, unsigned int len
,
716 size_t page_sz
, mmap_len
;
718 struct used_mask
*msk
;
719 struct mem_area
*ma
= NULL
;
728 if (fully_validate(name
, elt_sz
, len
))
731 /* allocate mem area before doing anything */
732 ma
= malloc(sizeof(*ma
));
738 page_sz
= sysconf(_SC_PAGESIZE
);
739 if (page_sz
== (size_t)-1) {
744 /* calculate our memory limits */
745 mmap_len
= calc_data_size(page_sz
, elt_sz
, len
);
747 data
= eal_get_virtual_area(NULL
, &mmap_len
, page_sz
, 0, 0);
753 rte_spinlock_lock(&mem_area_lock
);
757 if (internal_config
.no_shconf
) {
758 /* remap virtual area as writable */
759 void *new_data
= mmap(data
, mmap_len
, PROT_READ
| PROT_WRITE
,
760 MAP_FIXED
| MAP_PRIVATE
| MAP_ANONYMOUS
, fd
, 0);
761 if (new_data
== MAP_FAILED
) {
762 RTE_LOG(DEBUG
, EAL
, "%s(): couldn't remap anonymous memory: %s\n",
763 __func__
, strerror(errno
));
767 eal_get_fbarray_path(path
, sizeof(path
), name
);
770 * Each fbarray is unique to process namespace, i.e. the
771 * filename depends on process prefix. Try to take out a lock
772 * and see if we succeed. If we don't, someone else is using it
775 fd
= open(path
, O_CREAT
| O_RDWR
, 0600);
777 RTE_LOG(DEBUG
, EAL
, "%s(): couldn't open %s: %s\n",
778 __func__
, path
, strerror(errno
));
781 } else if (flock(fd
, LOCK_EX
| LOCK_NB
)) {
782 RTE_LOG(DEBUG
, EAL
, "%s(): couldn't lock %s: %s\n",
783 __func__
, path
, strerror(errno
));
788 /* take out a non-exclusive lock, so that other processes could
789 * still attach to it, but no other process could reinitialize
792 if (flock(fd
, LOCK_SH
| LOCK_NB
)) {
797 if (resize_and_map(fd
, data
, mmap_len
))
804 /* do not close fd - keep it until detach/destroy */
805 TAILQ_INSERT_TAIL(&mem_area_tailq
, ma
, next
);
807 /* initialize the data */
808 memset(data
, 0, mmap_len
);
810 /* populate data structure */
811 strlcpy(arr
->name
, name
, sizeof(arr
->name
));
814 arr
->elt_sz
= elt_sz
;
817 msk
= get_used_mask(data
, elt_sz
, len
);
818 msk
->n_masks
= MASK_LEN_TO_IDX(RTE_ALIGN_CEIL(len
, MASK_ALIGN
));
820 rte_rwlock_init(&arr
->rwlock
);
822 rte_spinlock_unlock(&mem_area_lock
);
827 munmap(data
, mmap_len
);
832 rte_spinlock_unlock(&mem_area_lock
);
836 int __rte_experimental
837 rte_fbarray_attach(struct rte_fbarray
*arr
)
839 struct mem_area
*ma
= NULL
, *tmp
= NULL
;
840 size_t page_sz
, mmap_len
;
851 * we don't need to synchronize attach as two values we need (element
852 * size and array length) are constant for the duration of life of
853 * the array, so the parts we care about will not race.
856 if (fully_validate(arr
->name
, arr
->elt_sz
, arr
->len
))
859 ma
= malloc(sizeof(*ma
));
865 page_sz
= sysconf(_SC_PAGESIZE
);
866 if (page_sz
== (size_t)-1) {
871 mmap_len
= calc_data_size(page_sz
, arr
->elt_sz
, arr
->len
);
873 /* check the tailq - maybe user has already mapped this address space */
874 rte_spinlock_lock(&mem_area_lock
);
876 TAILQ_FOREACH(tmp
, &mem_area_tailq
, next
) {
877 if (overlap(tmp
, arr
->data
, mmap_len
)) {
883 /* we know this memory area is unique, so proceed */
885 data
= eal_get_virtual_area(arr
->data
, &mmap_len
, page_sz
, 0, 0);
889 eal_get_fbarray_path(path
, sizeof(path
), arr
->name
);
891 fd
= open(path
, O_RDWR
);
897 /* lock the file, to let others know we're using it */
898 if (flock(fd
, LOCK_SH
| LOCK_NB
)) {
903 if (resize_and_map(fd
, data
, mmap_len
))
906 /* store our new memory area */
908 ma
->fd
= fd
; /* keep fd until detach/destroy */
911 TAILQ_INSERT_TAIL(&mem_area_tailq
, ma
, next
);
915 rte_spinlock_unlock(&mem_area_lock
);
919 munmap(data
, mmap_len
);
923 rte_spinlock_unlock(&mem_area_lock
);
927 int __rte_experimental
928 rte_fbarray_detach(struct rte_fbarray
*arr
)
930 struct mem_area
*tmp
= NULL
;
940 * we don't need to synchronize detach as two values we need (element
941 * size and total capacity) are constant for the duration of life of
942 * the array, so the parts we care about will not race. if the user is
943 * detaching while doing something else in the same process, we can't
944 * really do anything about it, things will blow up either way.
947 size_t page_sz
= sysconf(_SC_PAGESIZE
);
949 if (page_sz
== (size_t)-1)
952 mmap_len
= calc_data_size(page_sz
, arr
->elt_sz
, arr
->len
);
954 /* does this area exist? */
955 rte_spinlock_lock(&mem_area_lock
);
957 TAILQ_FOREACH(tmp
, &mem_area_tailq
, next
) {
958 if (tmp
->addr
== arr
->data
&& tmp
->len
== mmap_len
)
967 munmap(arr
->data
, mmap_len
);
969 /* area is unmapped, close fd and remove the tailq entry */
972 TAILQ_REMOVE(&mem_area_tailq
, tmp
, next
);
977 rte_spinlock_unlock(&mem_area_lock
);
981 int __rte_experimental
982 rte_fbarray_destroy(struct rte_fbarray
*arr
)
984 struct mem_area
*tmp
= NULL
;
995 * we don't need to synchronize detach as two values we need (element
996 * size and total capacity) are constant for the duration of life of
997 * the array, so the parts we care about will not race. if the user is
998 * detaching while doing something else in the same process, we can't
999 * really do anything about it, things will blow up either way.
1002 size_t page_sz
= sysconf(_SC_PAGESIZE
);
1004 if (page_sz
== (size_t)-1)
1007 mmap_len
= calc_data_size(page_sz
, arr
->elt_sz
, arr
->len
);
1009 /* does this area exist? */
1010 rte_spinlock_lock(&mem_area_lock
);
1012 TAILQ_FOREACH(tmp
, &mem_area_tailq
, next
) {
1013 if (tmp
->addr
== arr
->data
&& tmp
->len
== mmap_len
)
1021 /* with no shconf, there were never any files to begin with */
1022 if (!internal_config
.no_shconf
) {
1024 * attempt to get an exclusive lock on the file, to ensure it
1025 * has been detached by all other processes
1028 if (flock(fd
, LOCK_EX
| LOCK_NB
)) {
1029 RTE_LOG(DEBUG
, EAL
, "Cannot destroy fbarray - another process is using it\n");
1035 /* we're OK to destroy the file */
1036 eal_get_fbarray_path(path
, sizeof(path
), arr
->name
);
1038 RTE_LOG(DEBUG
, EAL
, "Cannot unlink fbarray: %s\n",
1042 * we're still holding an exclusive lock, so drop it to
1045 flock(fd
, LOCK_SH
| LOCK_NB
);
1052 munmap(arr
->data
, mmap_len
);
1054 /* area is unmapped, remove the tailq entry */
1055 TAILQ_REMOVE(&mem_area_tailq
, tmp
, next
);
1059 rte_spinlock_unlock(&mem_area_lock
);
1063 void * __rte_experimental
1064 rte_fbarray_get(const struct rte_fbarray
*arr
, unsigned int idx
)
1072 if (idx
>= arr
->len
) {
1077 ret
= RTE_PTR_ADD(arr
->data
, idx
* arr
->elt_sz
);
1082 int __rte_experimental
1083 rte_fbarray_set_used(struct rte_fbarray
*arr
, unsigned int idx
)
1085 return set_used(arr
, idx
, true);
1088 int __rte_experimental
1089 rte_fbarray_set_free(struct rte_fbarray
*arr
, unsigned int idx
)
1091 return set_used(arr
, idx
, false);
1094 int __rte_experimental
1095 rte_fbarray_is_used(struct rte_fbarray
*arr
, unsigned int idx
)
1097 struct used_mask
*msk
;
1102 if (arr
== NULL
|| idx
>= arr
->len
) {
1107 /* prevent array from changing under us */
1108 rte_rwlock_read_lock(&arr
->rwlock
);
1110 msk
= get_used_mask(arr
->data
, arr
->elt_sz
, arr
->len
);
1111 msk_idx
= MASK_LEN_TO_IDX(idx
);
1112 msk_bit
= 1ULL << MASK_LEN_TO_MOD(idx
);
1114 ret
= (msk
->data
[msk_idx
] & msk_bit
) != 0;
1116 rte_rwlock_read_unlock(&arr
->rwlock
);
1122 fbarray_find(struct rte_fbarray
*arr
, unsigned int start
, bool next
, bool used
)
1126 if (arr
== NULL
|| start
>= arr
->len
) {
1131 /* prevent array from changing under us */
1132 rte_rwlock_read_lock(&arr
->rwlock
);
1134 /* cheap checks to prevent doing useless work */
1136 if (arr
->len
== arr
->count
) {
1140 if (arr
->count
== 0) {
1145 if (arr
->count
== 0) {
1149 if (arr
->len
== arr
->count
) {
1155 ret
= find_next(arr
, start
, used
);
1157 ret
= find_prev(arr
, start
, used
);
1159 rte_rwlock_read_unlock(&arr
->rwlock
);
1163 int __rte_experimental
1164 rte_fbarray_find_next_free(struct rte_fbarray
*arr
, unsigned int start
)
1166 return fbarray_find(arr
, start
, true, false);
1169 int __rte_experimental
1170 rte_fbarray_find_next_used(struct rte_fbarray
*arr
, unsigned int start
)
1172 return fbarray_find(arr
, start
, true, true);
1175 int __rte_experimental
1176 rte_fbarray_find_prev_free(struct rte_fbarray
*arr
, unsigned int start
)
1178 return fbarray_find(arr
, start
, false, false);
1181 int __rte_experimental
1182 rte_fbarray_find_prev_used(struct rte_fbarray
*arr
, unsigned int start
)
1184 return fbarray_find(arr
, start
, false, true);
1188 fbarray_find_n(struct rte_fbarray
*arr
, unsigned int start
, unsigned int n
,
1189 bool next
, bool used
)
1193 if (arr
== NULL
|| start
>= arr
->len
|| n
> arr
->len
|| n
== 0) {
1197 if (next
&& (arr
->len
- start
) < n
) {
1198 rte_errno
= used
? ENOENT
: ENOSPC
;
1201 if (!next
&& start
< (n
- 1)) {
1202 rte_errno
= used
? ENOENT
: ENOSPC
;
1206 /* prevent array from changing under us */
1207 rte_rwlock_read_lock(&arr
->rwlock
);
1209 /* cheap checks to prevent doing useless work */
1211 if (arr
->len
== arr
->count
|| arr
->len
- arr
->count
< n
) {
1215 if (arr
->count
== 0) {
1216 ret
= next
? start
: start
- n
+ 1;
1220 if (arr
->count
< n
) {
1224 if (arr
->count
== arr
->len
) {
1225 ret
= next
? start
: start
- n
+ 1;
1231 ret
= find_next_n(arr
, start
, n
, used
);
1233 ret
= find_prev_n(arr
, start
, n
, used
);
1235 rte_rwlock_read_unlock(&arr
->rwlock
);
1239 int __rte_experimental
1240 rte_fbarray_find_next_n_free(struct rte_fbarray
*arr
, unsigned int start
,
1243 return fbarray_find_n(arr
, start
, n
, true, false);
1246 int __rte_experimental
1247 rte_fbarray_find_next_n_used(struct rte_fbarray
*arr
, unsigned int start
,
1250 return fbarray_find_n(arr
, start
, n
, true, true);
1253 int __rte_experimental
1254 rte_fbarray_find_prev_n_free(struct rte_fbarray
*arr
, unsigned int start
,
1257 return fbarray_find_n(arr
, start
, n
, false, false);
1260 int __rte_experimental
1261 rte_fbarray_find_prev_n_used(struct rte_fbarray
*arr
, unsigned int start
,
1264 return fbarray_find_n(arr
, start
, n
, false, true);
1268 fbarray_find_contig(struct rte_fbarray
*arr
, unsigned int start
, bool next
,
1273 if (arr
== NULL
|| start
>= arr
->len
) {
1278 /* prevent array from changing under us */
1279 rte_rwlock_read_lock(&arr
->rwlock
);
1281 /* cheap checks to prevent doing useless work */
1283 if (arr
->count
== 0) {
1287 if (next
&& arr
->count
== arr
->len
) {
1288 ret
= arr
->len
- start
;
1291 if (!next
&& arr
->count
== arr
->len
) {
1296 if (arr
->len
== arr
->count
) {
1300 if (next
&& arr
->count
== 0) {
1301 ret
= arr
->len
- start
;
1304 if (!next
&& arr
->count
== 0) {
1311 ret
= find_contig(arr
, start
, used
);
1313 ret
= find_rev_contig(arr
, start
, used
);
1315 rte_rwlock_read_unlock(&arr
->rwlock
);
1320 fbarray_find_biggest(struct rte_fbarray
*arr
, unsigned int start
, bool used
,
1323 int cur_idx
, next_idx
, cur_len
, biggest_idx
, biggest_len
;
1324 /* don't stack if conditions, use function pointers instead */
1325 int (*find_func
)(struct rte_fbarray
*, unsigned int);
1326 int (*find_contig_func
)(struct rte_fbarray
*, unsigned int);
1328 if (arr
== NULL
|| start
>= arr
->len
) {
1332 /* the other API calls already do their fair share of cheap checks, so
1333 * no need to do them here.
1336 /* the API's called are thread-safe, but something may still happen
1337 * inbetween the API calls, so lock the fbarray. all other API's are
1338 * read-locking the fbarray, so read lock here is OK.
1340 rte_rwlock_read_lock(&arr
->rwlock
);
1342 /* pick out appropriate functions */
1345 find_func
= rte_fbarray_find_prev_used
;
1346 find_contig_func
= rte_fbarray_find_rev_contig_used
;
1348 find_func
= rte_fbarray_find_next_used
;
1349 find_contig_func
= rte_fbarray_find_contig_used
;
1353 find_func
= rte_fbarray_find_prev_free
;
1354 find_contig_func
= rte_fbarray_find_rev_contig_free
;
1356 find_func
= rte_fbarray_find_next_free
;
1357 find_contig_func
= rte_fbarray_find_contig_free
;
1362 biggest_idx
= -1; /* default is error */
1365 cur_idx
= find_func(arr
, cur_idx
);
1367 /* block found, check its length */
1369 cur_len
= find_contig_func(arr
, cur_idx
);
1370 /* decide where we go next */
1371 next_idx
= rev
? cur_idx
- cur_len
: cur_idx
+ cur_len
;
1372 /* move current index to start of chunk */
1373 cur_idx
= rev
? next_idx
+ 1 : cur_idx
;
1375 if (cur_len
> biggest_len
) {
1376 biggest_idx
= cur_idx
;
1377 biggest_len
= cur_len
;
1380 /* in reverse mode, next_idx may be -1 if chunk started
1381 * at array beginning. this means there's no more work
1387 /* nothing more to find, stop. however, a failed API
1388 * call has set rte_errno, which we want to ignore, as
1389 * reaching the end of fbarray is not an error.
1395 /* if we didn't find anything at all, set rte_errno */
1396 if (biggest_idx
< 0)
1397 rte_errno
= used
? ENOENT
: ENOSPC
;
1399 rte_rwlock_read_unlock(&arr
->rwlock
);
1403 int __rte_experimental
1404 rte_fbarray_find_biggest_free(struct rte_fbarray
*arr
, unsigned int start
)
1406 return fbarray_find_biggest(arr
, start
, false, false);
1409 int __rte_experimental
1410 rte_fbarray_find_biggest_used(struct rte_fbarray
*arr
, unsigned int start
)
1412 return fbarray_find_biggest(arr
, start
, true, false);
1415 int __rte_experimental
1416 rte_fbarray_find_rev_biggest_free(struct rte_fbarray
*arr
, unsigned int start
)
1418 return fbarray_find_biggest(arr
, start
, false, true);
1421 int __rte_experimental
1422 rte_fbarray_find_rev_biggest_used(struct rte_fbarray
*arr
, unsigned int start
)
1424 return fbarray_find_biggest(arr
, start
, true, true);
1428 int __rte_experimental
1429 rte_fbarray_find_contig_free(struct rte_fbarray
*arr
, unsigned int start
)
1431 return fbarray_find_contig(arr
, start
, true, false);
1434 int __rte_experimental
1435 rte_fbarray_find_contig_used(struct rte_fbarray
*arr
, unsigned int start
)
1437 return fbarray_find_contig(arr
, start
, true, true);
1440 int __rte_experimental
1441 rte_fbarray_find_rev_contig_free(struct rte_fbarray
*arr
, unsigned int start
)
1443 return fbarray_find_contig(arr
, start
, false, false);
1446 int __rte_experimental
1447 rte_fbarray_find_rev_contig_used(struct rte_fbarray
*arr
, unsigned int start
)
1449 return fbarray_find_contig(arr
, start
, false, true);
1452 int __rte_experimental
1453 rte_fbarray_find_idx(const struct rte_fbarray
*arr
, const void *elt
)
1459 * no need to synchronize as it doesn't matter if underlying data
1460 * changes - we're doing pointer arithmetic here.
1463 if (arr
== NULL
|| elt
== NULL
) {
1467 end
= RTE_PTR_ADD(arr
->data
, arr
->elt_sz
* arr
->len
);
1468 if (elt
< arr
->data
|| elt
>= end
) {
1473 ret
= RTE_PTR_DIFF(elt
, arr
->data
) / arr
->elt_sz
;
1478 void __rte_experimental
1479 rte_fbarray_dump_metadata(struct rte_fbarray
*arr
, FILE *f
)
1481 struct used_mask
*msk
;
1484 if (arr
== NULL
|| f
== NULL
) {
1489 if (fully_validate(arr
->name
, arr
->elt_sz
, arr
->len
)) {
1490 fprintf(f
, "Invalid file-backed array\n");
1494 /* prevent array from changing under us */
1495 rte_rwlock_read_lock(&arr
->rwlock
);
1497 fprintf(f
, "File-backed array: %s\n", arr
->name
);
1498 fprintf(f
, "size: %i occupied: %i elt_sz: %i\n",
1499 arr
->len
, arr
->count
, arr
->elt_sz
);
1501 msk
= get_used_mask(arr
->data
, arr
->elt_sz
, arr
->len
);
1503 for (i
= 0; i
< msk
->n_masks
; i
++)
1504 fprintf(f
, "msk idx %i: 0x%016" PRIx64
"\n", i
, msk
->data
[i
]);
1506 rte_rwlock_read_unlock(&arr
->rwlock
);