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;
102 struct pagesz_walk_arg {
108 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
110 struct pagesz_walk_arg *wa = arg;
114 * we need to only look at page sizes available for a particular socket
115 * ID. so, we either need an exact match on socket ID (can match both
116 * native and external memory), or, if SOCKET_ID_ANY was specified as a
117 * socket ID argument, we must only look at native memory and ignore any
118 * page sizes associated with external memory.
120 valid = msl->socket_id == wa->socket_id;
121 valid |= wa->socket_id == SOCKET_ID_ANY && msl->external == 0;
123 if (valid && msl->page_sz < wa->min)
124 wa->min = msl->page_sz;
130 get_min_page_size(int socket_id)
132 struct pagesz_walk_arg wa;
135 wa.socket_id = socket_id;
137 rte_memseg_list_walk(find_min_pagesz, &wa);
139 return wa.min == SIZE_MAX ? (size_t) getpagesize() : wa.min;
144 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
145 void *obj, rte_iova_t iova)
147 struct rte_mempool_objhdr *hdr;
148 struct rte_mempool_objtlr *tlr __rte_unused;
150 /* set mempool ptr in header */
151 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
154 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
155 mp->populated_size++;
157 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
158 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
159 tlr = __mempool_get_trailer(obj);
160 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
164 /* call obj_cb() for each mempool element */
166 rte_mempool_obj_iter(struct rte_mempool *mp,
167 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
169 struct rte_mempool_objhdr *hdr;
173 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
174 obj = (char *)hdr + sizeof(*hdr);
175 obj_cb(mp, obj_cb_arg, obj, n);
182 /* call mem_cb() for each mempool memory chunk */
184 rte_mempool_mem_iter(struct rte_mempool *mp,
185 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
187 struct rte_mempool_memhdr *hdr;
190 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
191 mem_cb(mp, mem_cb_arg, hdr, n);
198 /* get the header, trailer and total size of a mempool element. */
200 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
201 struct rte_mempool_objsz *sz)
203 struct rte_mempool_objsz lsz;
205 sz = (sz != NULL) ? sz : &lsz;
207 sz->header_size = sizeof(struct rte_mempool_objhdr);
208 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
209 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
212 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
213 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
215 sz->trailer_size = 0;
218 /* element size is 8 bytes-aligned at least */
219 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
221 /* expand trailer to next cache line */
222 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
223 sz->total_size = sz->header_size + sz->elt_size +
225 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
226 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
227 RTE_MEMPOOL_ALIGN_MASK);
231 * increase trailer to add padding between objects in order to
232 * spread them across memory channels/ranks
234 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
236 new_size = optimize_object_size(sz->header_size + sz->elt_size +
238 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
241 /* this is the size of an object, including header and trailer */
242 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
244 return sz->total_size;
247 /* free a memchunk allocated with rte_memzone_reserve() */
249 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
252 const struct rte_memzone *mz = opaque;
253 rte_memzone_free(mz);
256 /* Free memory chunks used by a mempool. Objects must be in pool */
258 rte_mempool_free_memchunks(struct rte_mempool *mp)
260 struct rte_mempool_memhdr *memhdr;
263 while (!STAILQ_EMPTY(&mp->elt_list)) {
264 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
266 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
267 mp->populated_size--;
270 while (!STAILQ_EMPTY(&mp->mem_list)) {
271 memhdr = STAILQ_FIRST(&mp->mem_list);
272 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
273 if (memhdr->free_cb != NULL)
274 memhdr->free_cb(memhdr, memhdr->opaque);
281 mempool_ops_alloc_once(struct rte_mempool *mp)
285 /* create the internal ring if not already done */
286 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
287 ret = rte_mempool_ops_alloc(mp);
290 mp->flags |= MEMPOOL_F_POOL_CREATED;
295 /* Add objects in the pool, using a physically contiguous memory
296 * zone. Return the number of objects added, or a negative value
300 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
301 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
306 struct rte_mempool_memhdr *memhdr;
309 ret = mempool_ops_alloc_once(mp);
313 /* mempool is already populated */
314 if (mp->populated_size >= mp->size)
317 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
322 memhdr->addr = vaddr;
325 memhdr->free_cb = free_cb;
326 memhdr->opaque = opaque;
328 if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
329 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
331 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
338 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
340 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
341 len - off, mempool_add_elem, NULL);
343 /* not enough room to store one object */
349 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
358 /* Populate the mempool with a virtual area. Return the number of
359 * objects added, or a negative value on error.
362 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
363 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
367 size_t off, phys_len;
370 /* address and len must be page-aligned */
371 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
373 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
376 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
377 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
378 len, free_cb, opaque);
380 for (off = 0; off + pg_sz <= len &&
381 mp->populated_size < mp->size; off += phys_len) {
383 iova = rte_mem_virt2iova(addr + off);
385 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
390 /* populate with the largest group of contiguous pages */
391 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
394 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
396 if (iova_tmp != iova + phys_len)
400 ret = rte_mempool_populate_iova(mp, addr + off, iova,
401 phys_len, free_cb, opaque);
404 /* no need to call the free callback for next chunks */
412 rte_mempool_free_memchunks(mp);
416 /* Default function to populate the mempool: allocate memory in memzones,
417 * and populate them. Return the number of objects added, or a negative
421 rte_mempool_populate_default(struct rte_mempool *mp)
423 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
424 char mz_name[RTE_MEMZONE_NAMESIZE];
425 const struct rte_memzone *mz;
427 size_t align, pg_sz, pg_shift;
431 bool no_contig, try_contig, no_pageshift;
433 ret = mempool_ops_alloc_once(mp);
437 /* mempool must not be populated */
438 if (mp->nb_mem_chunks != 0)
441 no_contig = mp->flags & MEMPOOL_F_NO_IOVA_CONTIG;
444 * the following section calculates page shift and page size values.
446 * these values impact the result of calc_mem_size operation, which
447 * returns the amount of memory that should be allocated to store the
448 * desired number of objects. when not zero, it allocates more memory
449 * for the padding between objects, to ensure that an object does not
450 * cross a page boundary. in other words, page size/shift are to be set
451 * to zero if mempool elements won't care about page boundaries.
452 * there are several considerations for page size and page shift here.
454 * if we don't need our mempools to have physically contiguous objects,
455 * then just set page shift and page size to 0, because the user has
456 * indicated that there's no need to care about anything.
458 * if we do need contiguous objects, there is also an option to reserve
459 * the entire mempool memory as one contiguous block of memory, in
460 * which case the page shift and alignment wouldn't matter as well.
462 * if we require contiguous objects, but not necessarily the entire
463 * mempool reserved space to be contiguous, then there are two options.
465 * if our IO addresses are virtual, not actual physical (IOVA as VA
466 * case), then no page shift needed - our memory allocation will give us
467 * contiguous IO memory as far as the hardware is concerned, so
468 * act as if we're getting contiguous memory.
470 * if our IO addresses are physical, we may get memory from bigger
471 * pages, or we might get memory from smaller pages, and how much of it
472 * we require depends on whether we want bigger or smaller pages.
473 * However, requesting each and every memory size is too much work, so
474 * what we'll do instead is walk through the page sizes available, pick
475 * the smallest one and set up page shift to match that one. We will be
476 * wasting some space this way, but it's much nicer than looping around
477 * trying to reserve each and every page size.
479 * However, since size calculation will produce page-aligned sizes, it
480 * makes sense to first try and see if we can reserve the entire memzone
481 * in one contiguous chunk as well (otherwise we might end up wasting a
482 * 1G page on a 10MB memzone). If we fail to get enough contiguous
483 * memory, then we'll go and reserve space page-by-page.
485 no_pageshift = no_contig || rte_eal_iova_mode() == RTE_IOVA_VA;
486 try_contig = !no_contig && !no_pageshift && rte_eal_has_hugepages();
491 } else if (try_contig) {
492 pg_sz = get_min_page_size(mp->socket_id);
493 pg_shift = rte_bsf32(pg_sz);
495 pg_sz = getpagesize();
496 pg_shift = rte_bsf32(pg_sz);
499 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
500 size_t min_chunk_size;
503 if (try_contig || no_pageshift)
504 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
505 0, &min_chunk_size, &align);
507 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
508 pg_shift, &min_chunk_size, &align);
515 ret = snprintf(mz_name, sizeof(mz_name),
516 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
517 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
524 /* if we're trying to reserve contiguous memory, add appropriate
528 flags |= RTE_MEMZONE_IOVA_CONTIG;
530 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
531 mp->socket_id, flags, align);
533 /* if we were trying to allocate contiguous memory, failed and
534 * minimum required contiguous chunk fits minimum page, adjust
535 * memzone size to the page size, and try again.
537 if (mz == NULL && try_contig && min_chunk_size <= pg_sz) {
539 flags &= ~RTE_MEMZONE_IOVA_CONTIG;
541 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
542 pg_shift, &min_chunk_size, &align);
548 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
549 mp->socket_id, flags, align);
551 /* don't try reserving with 0 size if we were asked to reserve
552 * IOVA-contiguous memory.
554 if (min_chunk_size < (size_t)mem_size && mz == NULL) {
555 /* not enough memory, retry with the biggest zone we
558 mz = rte_memzone_reserve_aligned(mz_name, 0,
559 mp->socket_id, flags,
560 RTE_MAX(pg_sz, align));
567 if (mz->len < min_chunk_size) {
568 rte_memzone_free(mz);
578 if (no_pageshift || try_contig)
579 ret = rte_mempool_populate_iova(mp, mz->addr,
581 rte_mempool_memchunk_mz_free,
582 (void *)(uintptr_t)mz);
584 ret = rte_mempool_populate_virt(mp, mz->addr,
585 RTE_ALIGN_FLOOR(mz->len, pg_sz), pg_sz,
586 rte_mempool_memchunk_mz_free,
587 (void *)(uintptr_t)mz);
589 rte_memzone_free(mz);
597 rte_mempool_free_memchunks(mp);
601 /* return the memory size required for mempool objects in anonymous mem */
603 get_anon_size(const struct rte_mempool *mp)
606 size_t pg_sz, pg_shift;
607 size_t min_chunk_size;
610 pg_sz = getpagesize();
611 pg_shift = rte_bsf32(pg_sz);
612 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
613 &min_chunk_size, &align);
618 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
620 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
626 * Calculate size since memhdr->len has contiguous chunk length
627 * which may be smaller if anon map is split into many contiguous
628 * chunks. Result must be the same as we calculated on populate.
630 size = get_anon_size(memhdr->mp);
634 munmap(opaque, size);
637 /* populate the mempool with an anonymous mapping */
639 rte_mempool_populate_anon(struct rte_mempool *mp)
645 /* mempool is already populated, error */
646 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
651 ret = mempool_ops_alloc_once(mp);
655 size = get_anon_size(mp);
661 /* get chunk of virtually continuous memory */
662 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
663 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
664 if (addr == MAP_FAILED) {
668 /* can't use MMAP_LOCKED, it does not exist on BSD */
669 if (mlock(addr, size) < 0) {
675 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
676 rte_mempool_memchunk_anon_free, addr);
680 return mp->populated_size;
683 rte_mempool_free_memchunks(mp);
689 rte_mempool_free(struct rte_mempool *mp)
691 struct rte_mempool_list *mempool_list = NULL;
692 struct rte_tailq_entry *te;
697 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
698 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
699 /* find out tailq entry */
700 TAILQ_FOREACH(te, mempool_list, next) {
701 if (te->data == (void *)mp)
706 TAILQ_REMOVE(mempool_list, te, next);
709 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
711 rte_mempool_free_memchunks(mp);
712 rte_mempool_ops_free(mp);
713 rte_memzone_free(mp->mz);
717 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
720 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
725 * Create and initialize a cache for objects that are retrieved from and
726 * returned to an underlying mempool. This structure is identical to the
727 * local_cache[lcore_id] pointed to by the mempool structure.
729 struct rte_mempool_cache *
730 rte_mempool_cache_create(uint32_t size, int socket_id)
732 struct rte_mempool_cache *cache;
734 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
739 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
740 RTE_CACHE_LINE_SIZE, socket_id);
742 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
747 mempool_cache_init(cache, size);
753 * Free a cache. It's the responsibility of the user to make sure that any
754 * remaining objects in the cache are flushed to the corresponding
758 rte_mempool_cache_free(struct rte_mempool_cache *cache)
763 /* create an empty mempool */
765 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
766 unsigned cache_size, unsigned private_data_size,
767 int socket_id, unsigned flags)
769 char mz_name[RTE_MEMZONE_NAMESIZE];
770 struct rte_mempool_list *mempool_list;
771 struct rte_mempool *mp = NULL;
772 struct rte_tailq_entry *te = NULL;
773 const struct rte_memzone *mz = NULL;
775 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
776 struct rte_mempool_objsz objsz;
780 /* compilation-time checks */
781 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
782 RTE_CACHE_LINE_MASK) != 0);
783 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
784 RTE_CACHE_LINE_MASK) != 0);
785 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
786 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
787 RTE_CACHE_LINE_MASK) != 0);
788 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
789 RTE_CACHE_LINE_MASK) != 0);
792 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
794 /* asked for zero items */
800 /* asked cache too big */
801 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
802 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
807 /* "no cache align" imply "no spread" */
808 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
809 flags |= MEMPOOL_F_NO_SPREAD;
811 /* calculate mempool object sizes. */
812 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
817 rte_rwlock_write_lock(RTE_EAL_MEMPOOL_RWLOCK);
820 * reserve a memory zone for this mempool: private data is
823 private_data_size = (private_data_size +
824 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
827 /* try to allocate tailq entry */
828 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
830 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
834 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
835 mempool_size += private_data_size;
836 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
838 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
839 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
840 rte_errno = ENAMETOOLONG;
844 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
848 /* init the mempool structure */
850 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
851 ret = snprintf(mp->name, sizeof(mp->name), "%s", name);
852 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
853 rte_errno = ENAMETOOLONG;
859 mp->socket_id = socket_id;
860 mp->elt_size = objsz.elt_size;
861 mp->header_size = objsz.header_size;
862 mp->trailer_size = objsz.trailer_size;
863 /* Size of default caches, zero means disabled. */
864 mp->cache_size = cache_size;
865 mp->private_data_size = private_data_size;
866 STAILQ_INIT(&mp->elt_list);
867 STAILQ_INIT(&mp->mem_list);
870 * local_cache pointer is set even if cache_size is zero.
871 * The local_cache points to just past the elt_pa[] array.
873 mp->local_cache = (struct rte_mempool_cache *)
874 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
876 /* Init all default caches. */
877 if (cache_size != 0) {
878 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
879 mempool_cache_init(&mp->local_cache[lcore_id],
885 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
886 TAILQ_INSERT_TAIL(mempool_list, te, next);
887 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
888 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
893 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
895 rte_mempool_free(mp);
899 /* create the mempool */
901 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
902 unsigned cache_size, unsigned private_data_size,
903 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
904 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
905 int socket_id, unsigned flags)
908 struct rte_mempool *mp;
910 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
911 private_data_size, socket_id, flags);
916 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
917 * set the correct index into the table of ops structs.
919 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
920 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
921 else if (flags & MEMPOOL_F_SP_PUT)
922 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
923 else if (flags & MEMPOOL_F_SC_GET)
924 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
926 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
931 /* call the mempool priv initializer */
933 mp_init(mp, mp_init_arg);
935 if (rte_mempool_populate_default(mp) < 0)
938 /* call the object initializers */
940 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
945 rte_mempool_free(mp);
949 /* Return the number of entries in the mempool */
951 rte_mempool_avail_count(const struct rte_mempool *mp)
956 count = rte_mempool_ops_get_count(mp);
958 if (mp->cache_size == 0)
961 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
962 count += mp->local_cache[lcore_id].len;
965 * due to race condition (access to len is not locked), the
966 * total can be greater than size... so fix the result
968 if (count > mp->size)
973 /* return the number of entries allocated from the mempool */
975 rte_mempool_in_use_count(const struct rte_mempool *mp)
977 return mp->size - rte_mempool_avail_count(mp);
980 /* dump the cache status */
982 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
986 unsigned cache_count;
988 fprintf(f, " internal cache infos:\n");
989 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
991 if (mp->cache_size == 0)
994 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
995 cache_count = mp->local_cache[lcore_id].len;
996 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
997 lcore_id, cache_count);
998 count += cache_count;
1000 fprintf(f, " total_cache_count=%u\n", count);
1004 #ifndef __INTEL_COMPILER
1005 #pragma GCC diagnostic ignored "-Wcast-qual"
1008 /* check and update cookies or panic (internal) */
1009 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1010 void * const *obj_table_const, unsigned n, int free)
1012 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1013 struct rte_mempool_objhdr *hdr;
1014 struct rte_mempool_objtlr *tlr;
1020 /* Force to drop the "const" attribute. This is done only when
1021 * DEBUG is enabled */
1022 tmp = (void *) obj_table_const;
1028 if (rte_mempool_from_obj(obj) != mp)
1029 rte_panic("MEMPOOL: object is owned by another "
1032 hdr = __mempool_get_header(obj);
1033 cookie = hdr->cookie;
1036 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1037 RTE_LOG(CRIT, MEMPOOL,
1038 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1039 obj, (const void *) mp, cookie);
1040 rte_panic("MEMPOOL: bad header cookie (put)\n");
1042 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1043 } else if (free == 1) {
1044 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1045 RTE_LOG(CRIT, MEMPOOL,
1046 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1047 obj, (const void *) mp, cookie);
1048 rte_panic("MEMPOOL: bad header cookie (get)\n");
1050 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1051 } else if (free == 2) {
1052 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1053 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1054 RTE_LOG(CRIT, MEMPOOL,
1055 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1056 obj, (const void *) mp, cookie);
1057 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1060 tlr = __mempool_get_trailer(obj);
1061 cookie = tlr->cookie;
1062 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1063 RTE_LOG(CRIT, MEMPOOL,
1064 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1065 obj, (const void *) mp, cookie);
1066 rte_panic("MEMPOOL: bad trailer cookie\n");
1071 RTE_SET_USED(obj_table_const);
1078 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1079 void * const *first_obj_table_const, unsigned int n, int free)
1081 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1082 struct rte_mempool_info info;
1083 const size_t total_elt_sz =
1084 mp->header_size + mp->elt_size + mp->trailer_size;
1087 rte_mempool_ops_get_info(mp, &info);
1089 for (i = 0; i < n; ++i) {
1090 void *first_obj = first_obj_table_const[i];
1092 for (j = 0; j < info.contig_block_size; ++j) {
1095 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1096 rte_mempool_check_cookies(mp, &obj, 1, free);
1101 RTE_SET_USED(first_obj_table_const);
1107 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1109 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1110 void *obj, __rte_unused unsigned idx)
1112 __mempool_check_cookies(mp, &obj, 1, 2);
1116 mempool_audit_cookies(struct rte_mempool *mp)
1120 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1121 if (num != mp->size) {
1122 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1123 "iterated only over %u elements\n",
1128 #define mempool_audit_cookies(mp) do {} while(0)
1131 #ifndef __INTEL_COMPILER
1132 #pragma GCC diagnostic error "-Wcast-qual"
1135 /* check cookies before and after objects */
1137 mempool_audit_cache(const struct rte_mempool *mp)
1139 /* check cache size consistency */
1142 if (mp->cache_size == 0)
1145 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1146 const struct rte_mempool_cache *cache;
1147 cache = &mp->local_cache[lcore_id];
1148 if (cache->len > cache->flushthresh) {
1149 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1151 rte_panic("MEMPOOL: invalid cache len\n");
1156 /* check the consistency of mempool (size, cookies, ...) */
1158 rte_mempool_audit(struct rte_mempool *mp)
1160 mempool_audit_cache(mp);
1161 mempool_audit_cookies(mp);
1163 /* For case where mempool DEBUG is not set, and cache size is 0 */
1167 /* dump the status of the mempool on the console */
1169 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1171 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1172 struct rte_mempool_info info;
1173 struct rte_mempool_debug_stats sum;
1176 struct rte_mempool_memhdr *memhdr;
1177 unsigned common_count;
1178 unsigned cache_count;
1181 RTE_ASSERT(f != NULL);
1182 RTE_ASSERT(mp != NULL);
1184 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1185 fprintf(f, " flags=%x\n", mp->flags);
1186 fprintf(f, " pool=%p\n", mp->pool_data);
1187 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1188 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1189 fprintf(f, " size=%"PRIu32"\n", mp->size);
1190 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1191 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1192 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1193 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1194 fprintf(f, " total_obj_size=%"PRIu32"\n",
1195 mp->header_size + mp->elt_size + mp->trailer_size);
1197 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1199 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1200 mem_len += memhdr->len;
1202 fprintf(f, " avg bytes/object=%#Lf\n",
1203 (long double)mem_len / mp->size);
1206 cache_count = rte_mempool_dump_cache(f, mp);
1207 common_count = rte_mempool_ops_get_count(mp);
1208 if ((cache_count + common_count) > mp->size)
1209 common_count = mp->size - cache_count;
1210 fprintf(f, " common_pool_count=%u\n", common_count);
1212 /* sum and dump statistics */
1213 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1214 rte_mempool_ops_get_info(mp, &info);
1215 memset(&sum, 0, sizeof(sum));
1216 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1217 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1218 sum.put_objs += mp->stats[lcore_id].put_objs;
1219 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1220 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1221 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1222 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1223 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1224 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1226 fprintf(f, " stats:\n");
1227 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1228 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1229 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1230 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1231 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1232 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1233 if (info.contig_block_size > 0) {
1234 fprintf(f, " get_success_blks=%"PRIu64"\n",
1235 sum.get_success_blks);
1236 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1239 fprintf(f, " no statistics available\n");
1242 rte_mempool_audit(mp);
1245 /* dump the status of all mempools on the console */
1247 rte_mempool_list_dump(FILE *f)
1249 struct rte_mempool *mp = NULL;
1250 struct rte_tailq_entry *te;
1251 struct rte_mempool_list *mempool_list;
1253 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1255 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1257 TAILQ_FOREACH(te, mempool_list, next) {
1258 mp = (struct rte_mempool *) te->data;
1259 rte_mempool_dump(f, mp);
1262 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1265 /* search a mempool from its name */
1266 struct rte_mempool *
1267 rte_mempool_lookup(const char *name)
1269 struct rte_mempool *mp = NULL;
1270 struct rte_tailq_entry *te;
1271 struct rte_mempool_list *mempool_list;
1273 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1275 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1277 TAILQ_FOREACH(te, mempool_list, next) {
1278 mp = (struct rte_mempool *) te->data;
1279 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1283 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1293 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1296 struct rte_tailq_entry *te = NULL;
1297 struct rte_mempool_list *mempool_list;
1300 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1302 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1304 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1305 (*func)((struct rte_mempool *) te->data, arg);
1308 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);