1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation.
3 * Copyright(c) 2016 6WIND S.A.
14 #include <sys/queue.h>
17 #include <rte_common.h>
19 #include <rte_debug.h>
20 #include <rte_memory.h>
21 #include <rte_memzone.h>
22 #include <rte_malloc.h>
23 #include <rte_atomic.h>
24 #include <rte_launch.h>
26 #include <rte_eal_memconfig.h>
27 #include <rte_per_lcore.h>
28 #include <rte_lcore.h>
29 #include <rte_branch_prediction.h>
30 #include <rte_errno.h>
31 #include <rte_string_fns.h>
32 #include <rte_spinlock.h>
34 #include "rte_mempool.h"
36 TAILQ_HEAD(rte_mempool_list, rte_tailq_entry);
38 static struct rte_tailq_elem rte_mempool_tailq = {
39 .name = "RTE_MEMPOOL",
41 EAL_REGISTER_TAILQ(rte_mempool_tailq)
43 #define CACHE_FLUSHTHRESH_MULTIPLIER 1.5
44 #define CALC_CACHE_FLUSHTHRESH(c) \
45 ((typeof(c))((c) * CACHE_FLUSHTHRESH_MULTIPLIER))
48 * return the greatest common divisor between a and b (fast algorithm)
51 static unsigned get_gcd(unsigned a, unsigned b)
76 * Depending on memory configuration, objects addresses are spread
77 * between channels and ranks in RAM: the pool allocator will add
78 * padding between objects. This function return the new size of the
81 static unsigned optimize_object_size(unsigned obj_size)
83 unsigned nrank, nchan;
84 unsigned new_obj_size;
86 /* get number of channels */
87 nchan = rte_memory_get_nchannel();
91 nrank = rte_memory_get_nrank();
95 /* process new object size */
96 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
97 while (get_gcd(new_obj_size, nrank * nchan) != 1)
99 return new_obj_size * RTE_MEMPOOL_ALIGN;
103 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
107 if (msl->page_sz < *min)
114 get_min_page_size(void)
116 size_t min_pagesz = SIZE_MAX;
118 rte_memseg_list_walk(find_min_pagesz, &min_pagesz);
120 return min_pagesz == SIZE_MAX ? (size_t) getpagesize() : min_pagesz;
125 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
126 void *obj, rte_iova_t iova)
128 struct rte_mempool_objhdr *hdr;
129 struct rte_mempool_objtlr *tlr __rte_unused;
131 /* set mempool ptr in header */
132 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
135 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
136 mp->populated_size++;
138 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
139 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
140 tlr = __mempool_get_trailer(obj);
141 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
145 /* call obj_cb() for each mempool element */
147 rte_mempool_obj_iter(struct rte_mempool *mp,
148 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
150 struct rte_mempool_objhdr *hdr;
154 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
155 obj = (char *)hdr + sizeof(*hdr);
156 obj_cb(mp, obj_cb_arg, obj, n);
163 /* call mem_cb() for each mempool memory chunk */
165 rte_mempool_mem_iter(struct rte_mempool *mp,
166 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
168 struct rte_mempool_memhdr *hdr;
171 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
172 mem_cb(mp, mem_cb_arg, hdr, n);
179 /* get the header, trailer and total size of a mempool element. */
181 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
182 struct rte_mempool_objsz *sz)
184 struct rte_mempool_objsz lsz;
186 sz = (sz != NULL) ? sz : &lsz;
188 sz->header_size = sizeof(struct rte_mempool_objhdr);
189 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
190 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
193 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
194 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
196 sz->trailer_size = 0;
199 /* element size is 8 bytes-aligned at least */
200 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
202 /* expand trailer to next cache line */
203 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
204 sz->total_size = sz->header_size + sz->elt_size +
206 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
207 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
208 RTE_MEMPOOL_ALIGN_MASK);
212 * increase trailer to add padding between objects in order to
213 * spread them across memory channels/ranks
215 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
217 new_size = optimize_object_size(sz->header_size + sz->elt_size +
219 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
222 /* this is the size of an object, including header and trailer */
223 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
225 return sz->total_size;
230 * Internal function to calculate required memory chunk size.
233 rte_mempool_calc_mem_size_helper(uint32_t elt_num, size_t total_elt_sz,
236 size_t obj_per_page, pg_num, pg_sz;
238 if (total_elt_sz == 0)
242 return total_elt_sz * elt_num;
244 pg_sz = (size_t)1 << pg_shift;
245 obj_per_page = pg_sz / total_elt_sz;
246 if (obj_per_page == 0)
247 return RTE_ALIGN_CEIL(total_elt_sz, pg_sz) * elt_num;
249 pg_num = (elt_num + obj_per_page - 1) / obj_per_page;
250 return pg_num << pg_shift;
253 /* free a memchunk allocated with rte_memzone_reserve() */
255 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
258 const struct rte_memzone *mz = opaque;
259 rte_memzone_free(mz);
262 /* Free memory chunks used by a mempool. Objects must be in pool */
264 rte_mempool_free_memchunks(struct rte_mempool *mp)
266 struct rte_mempool_memhdr *memhdr;
269 while (!STAILQ_EMPTY(&mp->elt_list)) {
270 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
272 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
273 mp->populated_size--;
276 while (!STAILQ_EMPTY(&mp->mem_list)) {
277 memhdr = STAILQ_FIRST(&mp->mem_list);
278 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
279 if (memhdr->free_cb != NULL)
280 memhdr->free_cb(memhdr, memhdr->opaque);
287 mempool_ops_alloc_once(struct rte_mempool *mp)
291 /* create the internal ring if not already done */
292 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
293 ret = rte_mempool_ops_alloc(mp);
296 mp->flags |= MEMPOOL_F_POOL_CREATED;
301 /* Add objects in the pool, using a physically contiguous memory
302 * zone. Return the number of objects added, or a negative value
306 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
307 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
312 struct rte_mempool_memhdr *memhdr;
315 ret = mempool_ops_alloc_once(mp);
319 /* mempool is already populated */
320 if (mp->populated_size >= mp->size)
323 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
328 memhdr->addr = vaddr;
331 memhdr->free_cb = free_cb;
332 memhdr->opaque = opaque;
334 if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
335 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
337 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
344 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
346 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
347 len - off, mempool_add_elem, NULL);
349 /* not enough room to store one object */
355 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
364 /* Populate the mempool with a virtual area. Return the number of
365 * objects added, or a negative value on error.
368 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
369 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
373 size_t off, phys_len;
376 /* address and len must be page-aligned */
377 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
379 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
382 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
383 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
384 len, free_cb, opaque);
386 for (off = 0; off + pg_sz <= len &&
387 mp->populated_size < mp->size; off += phys_len) {
389 iova = rte_mem_virt2iova(addr + off);
391 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
396 /* populate with the largest group of contiguous pages */
397 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
400 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
402 if (iova_tmp != iova + phys_len)
406 ret = rte_mempool_populate_iova(mp, addr + off, iova,
407 phys_len, free_cb, opaque);
410 /* no need to call the free callback for next chunks */
418 rte_mempool_free_memchunks(mp);
422 /* Default function to populate the mempool: allocate memory in memzones,
423 * and populate them. Return the number of objects added, or a negative
427 rte_mempool_populate_default(struct rte_mempool *mp)
429 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
430 char mz_name[RTE_MEMZONE_NAMESIZE];
431 const struct rte_memzone *mz;
433 size_t align, pg_sz, pg_shift;
437 bool no_contig, try_contig, no_pageshift;
439 ret = mempool_ops_alloc_once(mp);
443 /* mempool must not be populated */
444 if (mp->nb_mem_chunks != 0)
447 no_contig = mp->flags & MEMPOOL_F_NO_IOVA_CONTIG;
450 * the following section calculates page shift and page size values.
452 * these values impact the result of calc_mem_size operation, which
453 * returns the amount of memory that should be allocated to store the
454 * desired number of objects. when not zero, it allocates more memory
455 * for the padding between objects, to ensure that an object does not
456 * cross a page boundary. in other words, page size/shift are to be set
457 * to zero if mempool elements won't care about page boundaries.
458 * there are several considerations for page size and page shift here.
460 * if we don't need our mempools to have physically contiguous objects,
461 * then just set page shift and page size to 0, because the user has
462 * indicated that there's no need to care about anything.
464 * if we do need contiguous objects, there is also an option to reserve
465 * the entire mempool memory as one contiguous block of memory, in
466 * which case the page shift and alignment wouldn't matter as well.
468 * if we require contiguous objects, but not necessarily the entire
469 * mempool reserved space to be contiguous, then there are two options.
471 * if our IO addresses are virtual, not actual physical (IOVA as VA
472 * case), then no page shift needed - our memory allocation will give us
473 * contiguous IO memory as far as the hardware is concerned, so
474 * act as if we're getting contiguous memory.
476 * if our IO addresses are physical, we may get memory from bigger
477 * pages, or we might get memory from smaller pages, and how much of it
478 * we require depends on whether we want bigger or smaller pages.
479 * However, requesting each and every memory size is too much work, so
480 * what we'll do instead is walk through the page sizes available, pick
481 * the smallest one and set up page shift to match that one. We will be
482 * wasting some space this way, but it's much nicer than looping around
483 * trying to reserve each and every page size.
485 * However, since size calculation will produce page-aligned sizes, it
486 * makes sense to first try and see if we can reserve the entire memzone
487 * in one contiguous chunk as well (otherwise we might end up wasting a
488 * 1G page on a 10MB memzone). If we fail to get enough contiguous
489 * memory, then we'll go and reserve space page-by-page.
491 no_pageshift = no_contig || rte_eal_iova_mode() == RTE_IOVA_VA;
492 try_contig = !no_contig && !no_pageshift && rte_eal_has_hugepages();
497 } else if (try_contig) {
498 pg_sz = get_min_page_size();
499 pg_shift = rte_bsf32(pg_sz);
501 pg_sz = getpagesize();
502 pg_shift = rte_bsf32(pg_sz);
505 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
506 size_t min_chunk_size;
509 if (try_contig || no_pageshift)
510 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
511 0, &min_chunk_size, &align);
513 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
514 pg_shift, &min_chunk_size, &align);
521 ret = snprintf(mz_name, sizeof(mz_name),
522 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
523 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
530 /* if we're trying to reserve contiguous memory, add appropriate
534 flags |= RTE_MEMZONE_IOVA_CONTIG;
536 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
537 mp->socket_id, flags, align);
539 /* if we were trying to allocate contiguous memory, failed and
540 * minimum required contiguous chunk fits minimum page, adjust
541 * memzone size to the page size, and try again.
543 if (mz == NULL && try_contig && min_chunk_size <= pg_sz) {
545 flags &= ~RTE_MEMZONE_IOVA_CONTIG;
547 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
548 pg_shift, &min_chunk_size, &align);
554 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
555 mp->socket_id, flags, align);
557 /* don't try reserving with 0 size if we were asked to reserve
558 * IOVA-contiguous memory.
560 if (min_chunk_size < (size_t)mem_size && mz == NULL) {
561 /* not enough memory, retry with the biggest zone we
564 mz = rte_memzone_reserve_aligned(mz_name, 0,
565 mp->socket_id, flags,
566 RTE_MAX(pg_sz, align));
573 if (mz->len < min_chunk_size) {
574 rte_memzone_free(mz);
584 if (no_pageshift || try_contig)
585 ret = rte_mempool_populate_iova(mp, mz->addr,
587 rte_mempool_memchunk_mz_free,
588 (void *)(uintptr_t)mz);
590 ret = rte_mempool_populate_virt(mp, mz->addr,
591 RTE_ALIGN_FLOOR(mz->len, pg_sz), pg_sz,
592 rte_mempool_memchunk_mz_free,
593 (void *)(uintptr_t)mz);
595 rte_memzone_free(mz);
603 rte_mempool_free_memchunks(mp);
607 /* return the memory size required for mempool objects in anonymous mem */
609 get_anon_size(const struct rte_mempool *mp)
612 size_t pg_sz, pg_shift;
613 size_t min_chunk_size;
616 pg_sz = getpagesize();
617 pg_shift = rte_bsf32(pg_sz);
618 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
619 &min_chunk_size, &align);
624 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
626 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
632 * Calculate size since memhdr->len has contiguous chunk length
633 * which may be smaller if anon map is split into many contiguous
634 * chunks. Result must be the same as we calculated on populate.
636 size = get_anon_size(memhdr->mp);
640 munmap(opaque, size);
643 /* populate the mempool with an anonymous mapping */
645 rte_mempool_populate_anon(struct rte_mempool *mp)
651 /* mempool is already populated, error */
652 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
657 ret = mempool_ops_alloc_once(mp);
661 size = get_anon_size(mp);
667 /* get chunk of virtually continuous memory */
668 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
669 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
670 if (addr == MAP_FAILED) {
674 /* can't use MMAP_LOCKED, it does not exist on BSD */
675 if (mlock(addr, size) < 0) {
681 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
682 rte_mempool_memchunk_anon_free, addr);
686 return mp->populated_size;
689 rte_mempool_free_memchunks(mp);
695 rte_mempool_free(struct rte_mempool *mp)
697 struct rte_mempool_list *mempool_list = NULL;
698 struct rte_tailq_entry *te;
703 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
704 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
705 /* find out tailq entry */
706 TAILQ_FOREACH(te, mempool_list, next) {
707 if (te->data == (void *)mp)
712 TAILQ_REMOVE(mempool_list, te, next);
715 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
717 rte_mempool_free_memchunks(mp);
718 rte_mempool_ops_free(mp);
719 rte_memzone_free(mp->mz);
723 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
726 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
731 * Create and initialize a cache for objects that are retrieved from and
732 * returned to an underlying mempool. This structure is identical to the
733 * local_cache[lcore_id] pointed to by the mempool structure.
735 struct rte_mempool_cache *
736 rte_mempool_cache_create(uint32_t size, int socket_id)
738 struct rte_mempool_cache *cache;
740 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
745 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
746 RTE_CACHE_LINE_SIZE, socket_id);
748 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
753 mempool_cache_init(cache, size);
759 * Free a cache. It's the responsibility of the user to make sure that any
760 * remaining objects in the cache are flushed to the corresponding
764 rte_mempool_cache_free(struct rte_mempool_cache *cache)
769 /* create an empty mempool */
771 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
772 unsigned cache_size, unsigned private_data_size,
773 int socket_id, unsigned flags)
775 char mz_name[RTE_MEMZONE_NAMESIZE];
776 struct rte_mempool_list *mempool_list;
777 struct rte_mempool *mp = NULL;
778 struct rte_tailq_entry *te = NULL;
779 const struct rte_memzone *mz = NULL;
781 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
782 struct rte_mempool_objsz objsz;
786 /* compilation-time checks */
787 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
788 RTE_CACHE_LINE_MASK) != 0);
789 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
790 RTE_CACHE_LINE_MASK) != 0);
791 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
792 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
793 RTE_CACHE_LINE_MASK) != 0);
794 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
795 RTE_CACHE_LINE_MASK) != 0);
798 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
800 /* asked for zero items */
806 /* asked cache too big */
807 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
808 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
813 /* "no cache align" imply "no spread" */
814 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
815 flags |= MEMPOOL_F_NO_SPREAD;
817 /* calculate mempool object sizes. */
818 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
823 rte_rwlock_write_lock(RTE_EAL_MEMPOOL_RWLOCK);
826 * reserve a memory zone for this mempool: private data is
829 private_data_size = (private_data_size +
830 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
833 /* try to allocate tailq entry */
834 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
836 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
840 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
841 mempool_size += private_data_size;
842 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
844 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
845 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
846 rte_errno = ENAMETOOLONG;
850 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
854 /* init the mempool structure */
856 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
857 ret = snprintf(mp->name, sizeof(mp->name), "%s", name);
858 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
859 rte_errno = ENAMETOOLONG;
865 mp->socket_id = socket_id;
866 mp->elt_size = objsz.elt_size;
867 mp->header_size = objsz.header_size;
868 mp->trailer_size = objsz.trailer_size;
869 /* Size of default caches, zero means disabled. */
870 mp->cache_size = cache_size;
871 mp->private_data_size = private_data_size;
872 STAILQ_INIT(&mp->elt_list);
873 STAILQ_INIT(&mp->mem_list);
876 * local_cache pointer is set even if cache_size is zero.
877 * The local_cache points to just past the elt_pa[] array.
879 mp->local_cache = (struct rte_mempool_cache *)
880 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
882 /* Init all default caches. */
883 if (cache_size != 0) {
884 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
885 mempool_cache_init(&mp->local_cache[lcore_id],
891 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
892 TAILQ_INSERT_TAIL(mempool_list, te, next);
893 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
894 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
899 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
901 rte_mempool_free(mp);
905 /* create the mempool */
907 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
908 unsigned cache_size, unsigned private_data_size,
909 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
910 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
911 int socket_id, unsigned flags)
914 struct rte_mempool *mp;
916 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
917 private_data_size, socket_id, flags);
922 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
923 * set the correct index into the table of ops structs.
925 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
926 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
927 else if (flags & MEMPOOL_F_SP_PUT)
928 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
929 else if (flags & MEMPOOL_F_SC_GET)
930 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
932 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
937 /* call the mempool priv initializer */
939 mp_init(mp, mp_init_arg);
941 if (rte_mempool_populate_default(mp) < 0)
944 /* call the object initializers */
946 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
951 rte_mempool_free(mp);
955 /* Return the number of entries in the mempool */
957 rte_mempool_avail_count(const struct rte_mempool *mp)
962 count = rte_mempool_ops_get_count(mp);
964 if (mp->cache_size == 0)
967 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
968 count += mp->local_cache[lcore_id].len;
971 * due to race condition (access to len is not locked), the
972 * total can be greater than size... so fix the result
974 if (count > mp->size)
979 /* return the number of entries allocated from the mempool */
981 rte_mempool_in_use_count(const struct rte_mempool *mp)
983 return mp->size - rte_mempool_avail_count(mp);
986 /* dump the cache status */
988 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
992 unsigned cache_count;
994 fprintf(f, " internal cache infos:\n");
995 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
997 if (mp->cache_size == 0)
1000 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1001 cache_count = mp->local_cache[lcore_id].len;
1002 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1003 lcore_id, cache_count);
1004 count += cache_count;
1006 fprintf(f, " total_cache_count=%u\n", count);
1010 #ifndef __INTEL_COMPILER
1011 #pragma GCC diagnostic ignored "-Wcast-qual"
1014 /* check and update cookies or panic (internal) */
1015 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1016 void * const *obj_table_const, unsigned n, int free)
1018 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1019 struct rte_mempool_objhdr *hdr;
1020 struct rte_mempool_objtlr *tlr;
1026 /* Force to drop the "const" attribute. This is done only when
1027 * DEBUG is enabled */
1028 tmp = (void *) obj_table_const;
1034 if (rte_mempool_from_obj(obj) != mp)
1035 rte_panic("MEMPOOL: object is owned by another "
1038 hdr = __mempool_get_header(obj);
1039 cookie = hdr->cookie;
1042 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1043 RTE_LOG(CRIT, MEMPOOL,
1044 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1045 obj, (const void *) mp, cookie);
1046 rte_panic("MEMPOOL: bad header cookie (put)\n");
1048 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1049 } else if (free == 1) {
1050 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1051 RTE_LOG(CRIT, MEMPOOL,
1052 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1053 obj, (const void *) mp, cookie);
1054 rte_panic("MEMPOOL: bad header cookie (get)\n");
1056 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1057 } else if (free == 2) {
1058 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1059 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1060 RTE_LOG(CRIT, MEMPOOL,
1061 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1062 obj, (const void *) mp, cookie);
1063 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1066 tlr = __mempool_get_trailer(obj);
1067 cookie = tlr->cookie;
1068 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1069 RTE_LOG(CRIT, MEMPOOL,
1070 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1071 obj, (const void *) mp, cookie);
1072 rte_panic("MEMPOOL: bad trailer cookie\n");
1077 RTE_SET_USED(obj_table_const);
1084 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1085 void * const *first_obj_table_const, unsigned int n, int free)
1087 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1088 struct rte_mempool_info info;
1089 const size_t total_elt_sz =
1090 mp->header_size + mp->elt_size + mp->trailer_size;
1093 rte_mempool_ops_get_info(mp, &info);
1095 for (i = 0; i < n; ++i) {
1096 void *first_obj = first_obj_table_const[i];
1098 for (j = 0; j < info.contig_block_size; ++j) {
1101 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1102 rte_mempool_check_cookies(mp, &obj, 1, free);
1107 RTE_SET_USED(first_obj_table_const);
1113 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1115 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1116 void *obj, __rte_unused unsigned idx)
1118 __mempool_check_cookies(mp, &obj, 1, 2);
1122 mempool_audit_cookies(struct rte_mempool *mp)
1126 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1127 if (num != mp->size) {
1128 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1129 "iterated only over %u elements\n",
1134 #define mempool_audit_cookies(mp) do {} while(0)
1137 #ifndef __INTEL_COMPILER
1138 #pragma GCC diagnostic error "-Wcast-qual"
1141 /* check cookies before and after objects */
1143 mempool_audit_cache(const struct rte_mempool *mp)
1145 /* check cache size consistency */
1148 if (mp->cache_size == 0)
1151 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1152 const struct rte_mempool_cache *cache;
1153 cache = &mp->local_cache[lcore_id];
1154 if (cache->len > cache->flushthresh) {
1155 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1157 rte_panic("MEMPOOL: invalid cache len\n");
1162 /* check the consistency of mempool (size, cookies, ...) */
1164 rte_mempool_audit(struct rte_mempool *mp)
1166 mempool_audit_cache(mp);
1167 mempool_audit_cookies(mp);
1169 /* For case where mempool DEBUG is not set, and cache size is 0 */
1173 /* dump the status of the mempool on the console */
1175 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1177 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1178 struct rte_mempool_info info;
1179 struct rte_mempool_debug_stats sum;
1182 struct rte_mempool_memhdr *memhdr;
1183 unsigned common_count;
1184 unsigned cache_count;
1187 RTE_ASSERT(f != NULL);
1188 RTE_ASSERT(mp != NULL);
1190 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1191 fprintf(f, " flags=%x\n", mp->flags);
1192 fprintf(f, " pool=%p\n", mp->pool_data);
1193 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1194 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1195 fprintf(f, " size=%"PRIu32"\n", mp->size);
1196 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1197 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1198 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1199 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1200 fprintf(f, " total_obj_size=%"PRIu32"\n",
1201 mp->header_size + mp->elt_size + mp->trailer_size);
1203 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1205 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1206 mem_len += memhdr->len;
1208 fprintf(f, " avg bytes/object=%#Lf\n",
1209 (long double)mem_len / mp->size);
1212 cache_count = rte_mempool_dump_cache(f, mp);
1213 common_count = rte_mempool_ops_get_count(mp);
1214 if ((cache_count + common_count) > mp->size)
1215 common_count = mp->size - cache_count;
1216 fprintf(f, " common_pool_count=%u\n", common_count);
1218 /* sum and dump statistics */
1219 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1220 rte_mempool_ops_get_info(mp, &info);
1221 memset(&sum, 0, sizeof(sum));
1222 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1223 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1224 sum.put_objs += mp->stats[lcore_id].put_objs;
1225 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1226 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1227 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1228 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1229 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1230 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1232 fprintf(f, " stats:\n");
1233 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1234 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1235 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1236 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1237 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1238 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1239 if (info.contig_block_size > 0) {
1240 fprintf(f, " get_success_blks=%"PRIu64"\n",
1241 sum.get_success_blks);
1242 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1245 fprintf(f, " no statistics available\n");
1248 rte_mempool_audit(mp);
1251 /* dump the status of all mempools on the console */
1253 rte_mempool_list_dump(FILE *f)
1255 struct rte_mempool *mp = NULL;
1256 struct rte_tailq_entry *te;
1257 struct rte_mempool_list *mempool_list;
1259 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1261 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1263 TAILQ_FOREACH(te, mempool_list, next) {
1264 mp = (struct rte_mempool *) te->data;
1265 rte_mempool_dump(f, mp);
1268 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1271 /* search a mempool from its name */
1272 struct rte_mempool *
1273 rte_mempool_lookup(const char *name)
1275 struct rte_mempool *mp = NULL;
1276 struct rte_tailq_entry *te;
1277 struct rte_mempool_list *mempool_list;
1279 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1281 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1283 TAILQ_FOREACH(te, mempool_list, next) {
1284 mp = (struct rte_mempool *) te->data;
1285 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1289 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1299 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1302 struct rte_tailq_entry *te = NULL;
1303 struct rte_mempool_list *mempool_list;
1306 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1308 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1310 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1311 (*func)((struct rte_mempool *) te->data, arg);
1314 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);