1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2017-2018 Intel Corporation
13 #include <rte_common.h>
15 #include <rte_errno.h>
16 #include <rte_spinlock.h>
17 #include <rte_tailq.h>
19 #include "eal_filesystem.h"
20 #include "eal_private.h"
22 #include "rte_fbarray.h"
24 #define MASK_SHIFT 6ULL
25 #define MASK_ALIGN (1ULL << MASK_SHIFT)
26 #define MASK_LEN_TO_IDX(x) ((x) >> MASK_SHIFT)
27 #define MASK_LEN_TO_MOD(x) ((x) - RTE_ALIGN_FLOOR(x, MASK_ALIGN))
28 #define MASK_GET_IDX(idx, mod) ((idx << MASK_SHIFT) + mod)
31 * This is a mask that is always stored at the end of array, to provide fast
32 * way of finding free/used spots without looping through each element.
41 calc_mask_size(unsigned int len)
43 /* mask must be multiple of MASK_ALIGN, even though length of array
44 * itself may not be aligned on that boundary.
46 len = RTE_ALIGN_CEIL(len, MASK_ALIGN);
47 return sizeof(struct used_mask) +
48 sizeof(uint64_t) * MASK_LEN_TO_IDX(len);
52 calc_data_size(size_t page_sz, unsigned int elt_sz, unsigned int len)
54 size_t data_sz = elt_sz * len;
55 size_t msk_sz = calc_mask_size(len);
56 return RTE_ALIGN_CEIL(data_sz + msk_sz, page_sz);
59 static struct used_mask *
60 get_used_mask(void *data, unsigned int elt_sz, unsigned int len)
62 return (struct used_mask *) RTE_PTR_ADD(data, elt_sz * len);
66 resize_and_map(int fd, void *addr, size_t len)
71 if (ftruncate(fd, len)) {
72 RTE_LOG(ERR, EAL, "Cannot truncate %s\n", path);
73 /* pass errno up the chain */
78 map_addr = mmap(addr, len, PROT_READ | PROT_WRITE,
79 MAP_SHARED | MAP_FIXED, fd, 0);
80 if (map_addr != addr) {
81 RTE_LOG(ERR, EAL, "mmap() failed: %s\n", strerror(errno));
82 /* pass errno up the chain */
90 find_next_n(const struct rte_fbarray *arr, unsigned int start, unsigned int n,
93 const struct used_mask *msk = get_used_mask(arr->data, arr->elt_sz,
95 unsigned int msk_idx, lookahead_idx, first, first_mod;
96 unsigned int last, last_mod;
97 uint64_t last_msk, ignore_msk;
100 * mask only has granularity of MASK_ALIGN, but start may not be aligned
101 * on that boundary, so construct a special mask to exclude anything we
102 * don't want to see to avoid confusing ctz.
104 first = MASK_LEN_TO_IDX(start);
105 first_mod = MASK_LEN_TO_MOD(start);
106 ignore_msk = ~((1ULL << first_mod) - 1);
108 /* array length may not be aligned, so calculate ignore mask for last
111 last = MASK_LEN_TO_IDX(arr->len);
112 last_mod = MASK_LEN_TO_MOD(arr->len);
113 last_msk = ~(-1ULL << last_mod);
115 for (msk_idx = first; msk_idx < msk->n_masks; msk_idx++) {
116 uint64_t cur_msk, lookahead_msk;
117 unsigned int run_start, clz, left;
120 * The process of getting n consecutive bits for arbitrary n is
121 * a bit involved, but here it is in a nutshell:
123 * 1. let n be the number of consecutive bits we're looking for
124 * 2. check if n can fit in one mask, and if so, do n-1
125 * rshift-ands to see if there is an appropriate run inside
127 * 2a. if we found a run, bail out early
128 * 2b. if we didn't find a run, proceed
129 * 3. invert the mask and count leading zeroes (that is, count
130 * how many consecutive set bits we had starting from the
131 * end of current mask) as k
132 * 3a. if k is 0, continue to next mask
133 * 3b. if k is not 0, we have a potential run
134 * 4. to satisfy our requirements, next mask must have n-k
135 * consecutive set bits right at the start, so we will do
136 * (n-k-1) rshift-ands and check if first bit is set.
138 * Step 4 will need to be repeated if (n-k) > MASK_ALIGN until
139 * we either run out of masks, lose the run, or find what we
142 cur_msk = msk->data[msk_idx];
145 /* if we're looking for free spaces, invert the mask */
149 /* combine current ignore mask with last index ignore mask */
151 ignore_msk |= last_msk;
153 /* if we have an ignore mask, ignore once */
155 cur_msk &= ignore_msk;
159 /* if n can fit in within a single mask, do a search */
160 if (n <= MASK_ALIGN) {
161 uint64_t tmp_msk = cur_msk;
163 for (s_idx = 0; s_idx < n - 1; s_idx++)
164 tmp_msk &= tmp_msk >> 1ULL;
165 /* we found what we were looking for */
167 run_start = __builtin_ctzll(tmp_msk);
168 return MASK_GET_IDX(msk_idx, run_start);
173 * we didn't find our run within the mask, or n > MASK_ALIGN,
174 * so we're going for plan B.
177 /* count leading zeroes on inverted mask */
179 clz = sizeof(cur_msk) * 8;
181 clz = __builtin_clzll(~cur_msk);
183 /* if there aren't any runs at the end either, just continue */
187 /* we have a partial run at the end, so try looking ahead */
188 run_start = MASK_ALIGN - clz;
191 for (lookahead_idx = msk_idx + 1; lookahead_idx < msk->n_masks;
193 unsigned int s_idx, need;
194 lookahead_msk = msk->data[lookahead_idx];
196 /* if we're looking for free space, invert the mask */
198 lookahead_msk = ~lookahead_msk;
200 /* figure out how many consecutive bits we need here */
201 need = RTE_MIN(left, MASK_ALIGN);
203 for (s_idx = 0; s_idx < need - 1; s_idx++)
204 lookahead_msk &= lookahead_msk >> 1ULL;
206 /* if first bit is not set, we've lost the run */
207 if ((lookahead_msk & 1) == 0) {
209 * we've scanned this far, so we know there are
210 * no runs in the space we've lookahead-scanned
211 * as well, so skip that on next iteration.
213 ignore_msk = ~((1ULL << need) - 1);
214 msk_idx = lookahead_idx;
220 /* check if we've found what we were looking for */
227 /* we didn't find anything, so continue */
231 return MASK_GET_IDX(msk_idx, run_start);
233 /* we didn't find anything */
234 rte_errno = used ? -ENOENT : -ENOSPC;
239 find_next(const struct rte_fbarray *arr, unsigned int start, bool used)
241 const struct used_mask *msk = get_used_mask(arr->data, arr->elt_sz,
243 unsigned int idx, first, first_mod;
244 unsigned int last, last_mod;
245 uint64_t last_msk, ignore_msk;
248 * mask only has granularity of MASK_ALIGN, but start may not be aligned
249 * on that boundary, so construct a special mask to exclude anything we
250 * don't want to see to avoid confusing ctz.
252 first = MASK_LEN_TO_IDX(start);
253 first_mod = MASK_LEN_TO_MOD(start);
254 ignore_msk = ~((1ULL << first_mod) - 1ULL);
256 /* array length may not be aligned, so calculate ignore mask for last
259 last = MASK_LEN_TO_IDX(arr->len);
260 last_mod = MASK_LEN_TO_MOD(arr->len);
261 last_msk = ~(-(1ULL) << last_mod);
263 for (idx = first; idx < msk->n_masks; idx++) {
264 uint64_t cur = msk->data[idx];
267 /* if we're looking for free entries, invert mask */
274 /* ignore everything before start on first iteration */
278 /* check if we have any entries */
283 * find first set bit - that will correspond to whatever it is
284 * that we're looking for.
286 found = __builtin_ctzll(cur);
287 return MASK_GET_IDX(idx, found);
289 /* we didn't find anything */
290 rte_errno = used ? -ENOENT : -ENOSPC;
295 find_contig(const struct rte_fbarray *arr, unsigned int start, bool used)
297 const struct used_mask *msk = get_used_mask(arr->data, arr->elt_sz,
299 unsigned int idx, first, first_mod;
300 unsigned int last, last_mod;
302 unsigned int need_len, result = 0;
304 /* array length may not be aligned, so calculate ignore mask for last
307 last = MASK_LEN_TO_IDX(arr->len);
308 last_mod = MASK_LEN_TO_MOD(arr->len);
309 last_msk = ~(-(1ULL) << last_mod);
311 first = MASK_LEN_TO_IDX(start);
312 first_mod = MASK_LEN_TO_MOD(start);
313 for (idx = first; idx < msk->n_masks; idx++, result += need_len) {
314 uint64_t cur = msk->data[idx];
315 unsigned int run_len;
317 need_len = MASK_ALIGN;
319 /* if we're looking for free entries, invert mask */
323 /* if this is last mask, ignore everything after last bit */
327 /* ignore everything before start on first iteration */
330 /* at the start, we don't need the full mask len */
331 need_len -= first_mod;
334 /* we will be looking for zeroes, so invert the mask */
337 /* if mask is zero, we have a complete run */
342 * see if current run ends before mask end.
344 run_len = __builtin_ctzll(cur);
346 /* add however many zeroes we've had in the last run and quit */
347 if (run_len < need_len) {
356 set_used(struct rte_fbarray *arr, unsigned int idx, bool used)
358 struct used_mask *msk = get_used_mask(arr->data, arr->elt_sz, arr->len);
359 uint64_t msk_bit = 1ULL << MASK_LEN_TO_MOD(idx);
360 unsigned int msk_idx = MASK_LEN_TO_IDX(idx);
364 if (arr == NULL || idx >= arr->len) {
370 /* prevent array from changing under us */
371 rte_rwlock_write_lock(&arr->rwlock);
373 already_used = (msk->data[msk_idx] & msk_bit) != 0;
375 /* nothing to be done */
376 if (used == already_used)
380 msk->data[msk_idx] |= msk_bit;
383 msk->data[msk_idx] &= ~msk_bit;
387 rte_rwlock_write_unlock(&arr->rwlock);
393 fully_validate(const char *name, unsigned int elt_sz, unsigned int len)
395 if (name == NULL || elt_sz == 0 || len == 0 || len > INT_MAX) {
400 if (strnlen(name, RTE_FBARRAY_NAME_LEN) == RTE_FBARRAY_NAME_LEN) {
401 rte_errno = ENAMETOOLONG;
407 int __rte_experimental
408 rte_fbarray_init(struct rte_fbarray *arr, const char *name, unsigned int len,
411 size_t page_sz, mmap_len;
413 struct used_mask *msk;
422 if (fully_validate(name, elt_sz, len))
425 page_sz = sysconf(_SC_PAGESIZE);
426 if (page_sz == (size_t)-1)
429 /* calculate our memory limits */
430 mmap_len = calc_data_size(page_sz, elt_sz, len);
432 data = eal_get_virtual_area(NULL, &mmap_len, page_sz, 0, 0);
436 eal_get_fbarray_path(path, sizeof(path), name);
439 * Each fbarray is unique to process namespace, i.e. the filename
440 * depends on process prefix. Try to take out a lock and see if we
441 * succeed. If we don't, someone else is using it already.
443 fd = open(path, O_CREAT | O_RDWR, 0600);
445 RTE_LOG(DEBUG, EAL, "%s(): couldn't open %s: %s\n", __func__,
446 path, strerror(errno));
449 } else if (flock(fd, LOCK_EX | LOCK_NB)) {
450 RTE_LOG(DEBUG, EAL, "%s(): couldn't lock %s: %s\n", __func__,
451 path, strerror(errno));
456 /* take out a non-exclusive lock, so that other processes could still
457 * attach to it, but no other process could reinitialize it.
459 if (flock(fd, LOCK_SH | LOCK_NB)) {
464 if (resize_and_map(fd, data, mmap_len))
467 /* we've mmap'ed the file, we can now close the fd */
470 /* initialize the data */
471 memset(data, 0, mmap_len);
473 /* populate data structure */
474 snprintf(arr->name, sizeof(arr->name), "%s", name);
477 arr->elt_sz = elt_sz;
480 msk = get_used_mask(data, elt_sz, len);
481 msk->n_masks = MASK_LEN_TO_IDX(RTE_ALIGN_CEIL(len, MASK_ALIGN));
483 rte_rwlock_init(&arr->rwlock);
488 munmap(data, mmap_len);
494 int __rte_experimental
495 rte_fbarray_attach(struct rte_fbarray *arr)
497 size_t page_sz, mmap_len;
508 * we don't need to synchronize attach as two values we need (element
509 * size and array length) are constant for the duration of life of
510 * the array, so the parts we care about will not race.
513 if (fully_validate(arr->name, arr->elt_sz, arr->len))
516 page_sz = sysconf(_SC_PAGESIZE);
517 if (page_sz == (size_t)-1)
520 mmap_len = calc_data_size(page_sz, arr->elt_sz, arr->len);
522 data = eal_get_virtual_area(arr->data, &mmap_len, page_sz, 0, 0);
526 eal_get_fbarray_path(path, sizeof(path), arr->name);
528 fd = open(path, O_RDWR);
534 /* lock the file, to let others know we're using it */
535 if (flock(fd, LOCK_SH | LOCK_NB)) {
540 if (resize_and_map(fd, data, mmap_len))
550 munmap(data, mmap_len);
556 int __rte_experimental
557 rte_fbarray_detach(struct rte_fbarray *arr)
565 * we don't need to synchronize detach as two values we need (element
566 * size and total capacity) are constant for the duration of life of
567 * the array, so the parts we care about will not race. if the user is
568 * detaching while doing something else in the same process, we can't
569 * really do anything about it, things will blow up either way.
572 size_t page_sz = sysconf(_SC_PAGESIZE);
574 /* this may already be unmapped (e.g. repeated call from previously
575 * failed destroy(), but this is on user, we can't (easily) know if this
578 munmap(arr->data, calc_data_size(page_sz, arr->elt_sz, arr->len));
583 int __rte_experimental
584 rte_fbarray_destroy(struct rte_fbarray *arr)
589 ret = rte_fbarray_detach(arr);
593 /* try deleting the file */
594 eal_get_fbarray_path(path, sizeof(path), arr->name);
596 fd = open(path, O_RDONLY);
597 if (flock(fd, LOCK_EX | LOCK_NB)) {
598 RTE_LOG(DEBUG, EAL, "Cannot destroy fbarray - another process is using it\n");
604 memset(arr, 0, sizeof(*arr));
611 void * __rte_experimental
612 rte_fbarray_get(const struct rte_fbarray *arr, unsigned int idx)
620 if (idx >= arr->len) {
625 ret = RTE_PTR_ADD(arr->data, idx * arr->elt_sz);
630 int __rte_experimental
631 rte_fbarray_set_used(struct rte_fbarray *arr, unsigned int idx)
633 return set_used(arr, idx, true);
636 int __rte_experimental
637 rte_fbarray_set_free(struct rte_fbarray *arr, unsigned int idx)
639 return set_used(arr, idx, false);
642 int __rte_experimental
643 rte_fbarray_is_used(struct rte_fbarray *arr, unsigned int idx)
645 struct used_mask *msk;
650 if (arr == NULL || idx >= arr->len) {
655 /* prevent array from changing under us */
656 rte_rwlock_read_lock(&arr->rwlock);
658 msk = get_used_mask(arr->data, arr->elt_sz, arr->len);
659 msk_idx = MASK_LEN_TO_IDX(idx);
660 msk_bit = 1ULL << MASK_LEN_TO_MOD(idx);
662 ret = (msk->data[msk_idx] & msk_bit) != 0;
664 rte_rwlock_read_unlock(&arr->rwlock);
669 int __rte_experimental
670 rte_fbarray_find_next_free(struct rte_fbarray *arr, unsigned int start)
674 if (arr == NULL || start >= arr->len) {
679 /* prevent array from changing under us */
680 rte_rwlock_read_lock(&arr->rwlock);
682 if (arr->len == arr->count) {
687 ret = find_next(arr, start, false);
689 rte_rwlock_read_unlock(&arr->rwlock);
693 int __rte_experimental
694 rte_fbarray_find_next_used(struct rte_fbarray *arr, unsigned int start)
698 if (arr == NULL || start >= arr->len) {
703 /* prevent array from changing under us */
704 rte_rwlock_read_lock(&arr->rwlock);
706 if (arr->count == 0) {
711 ret = find_next(arr, start, true);
713 rte_rwlock_read_unlock(&arr->rwlock);
717 int __rte_experimental
718 rte_fbarray_find_next_n_free(struct rte_fbarray *arr, unsigned int start,
723 if (arr == NULL || start >= arr->len || n > arr->len) {
728 /* prevent array from changing under us */
729 rte_rwlock_read_lock(&arr->rwlock);
731 if (arr->len == arr->count || arr->len - arr->count < n) {
736 ret = find_next_n(arr, start, n, false);
738 rte_rwlock_read_unlock(&arr->rwlock);
742 int __rte_experimental
743 rte_fbarray_find_next_n_used(struct rte_fbarray *arr, unsigned int start,
748 if (arr == NULL || start >= arr->len || n > arr->len) {
753 /* prevent array from changing under us */
754 rte_rwlock_read_lock(&arr->rwlock);
756 if (arr->count < n) {
761 ret = find_next_n(arr, start, n, true);
763 rte_rwlock_read_unlock(&arr->rwlock);
767 int __rte_experimental
768 rte_fbarray_find_contig_free(struct rte_fbarray *arr, unsigned int start)
772 if (arr == NULL || start >= arr->len) {
777 /* prevent array from changing under us */
778 rte_rwlock_read_lock(&arr->rwlock);
780 if (arr->len == arr->count) {
785 if (arr->count == 0) {
786 ret = arr->len - start;
790 ret = find_contig(arr, start, false);
792 rte_rwlock_read_unlock(&arr->rwlock);
796 int __rte_experimental
797 rte_fbarray_find_contig_used(struct rte_fbarray *arr, unsigned int start)
801 if (arr == NULL || start >= arr->len) {
806 /* prevent array from changing under us */
807 rte_rwlock_read_lock(&arr->rwlock);
809 ret = find_contig(arr, start, true);
811 rte_rwlock_read_unlock(&arr->rwlock);
815 int __rte_experimental
816 rte_fbarray_find_idx(const struct rte_fbarray *arr, const void *elt)
822 * no need to synchronize as it doesn't matter if underlying data
823 * changes - we're doing pointer arithmetic here.
826 if (arr == NULL || elt == NULL) {
830 end = RTE_PTR_ADD(arr->data, arr->elt_sz * arr->len);
831 if (elt < arr->data || elt >= end) {
836 ret = RTE_PTR_DIFF(elt, arr->data) / arr->elt_sz;
841 void __rte_experimental
842 rte_fbarray_dump_metadata(struct rte_fbarray *arr, FILE *f)
844 struct used_mask *msk;
847 if (arr == NULL || f == NULL) {
852 if (fully_validate(arr->name, arr->elt_sz, arr->len)) {
853 fprintf(f, "Invalid file-backed array\n");
857 /* prevent array from changing under us */
858 rte_rwlock_read_lock(&arr->rwlock);
860 fprintf(f, "File-backed array: %s\n", arr->name);
861 fprintf(f, "size: %i occupied: %i elt_sz: %i\n",
862 arr->len, arr->count, arr->elt_sz);
864 msk = get_used_mask(arr->data, arr->elt_sz, arr->len);
866 for (i = 0; i < msk->n_masks; i++)
867 fprintf(f, "msk idx %i: 0x%016" PRIx64 "\n", i, msk->data[i]);
869 rte_rwlock_read_unlock(&arr->rwlock);