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>
33 #include <rte_tailq.h>
35 #include "rte_mempool.h"
37 TAILQ_HEAD(rte_mempool_list, rte_tailq_entry);
39 static struct rte_tailq_elem rte_mempool_tailq = {
40 .name = "RTE_MEMPOOL",
42 EAL_REGISTER_TAILQ(rte_mempool_tailq)
44 #define CACHE_FLUSHTHRESH_MULTIPLIER 1.5
45 #define CALC_CACHE_FLUSHTHRESH(c) \
46 ((typeof(c))((c) * CACHE_FLUSHTHRESH_MULTIPLIER))
49 * return the greatest common divisor between a and b (fast algorithm)
52 static unsigned get_gcd(unsigned a, unsigned b)
77 * Depending on memory configuration, objects addresses are spread
78 * between channels and ranks in RAM: the pool allocator will add
79 * padding between objects. This function return the new size of the
82 static unsigned optimize_object_size(unsigned obj_size)
84 unsigned nrank, nchan;
85 unsigned new_obj_size;
87 /* get number of channels */
88 nchan = rte_memory_get_nchannel();
92 nrank = rte_memory_get_nrank();
96 /* process new object size */
97 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
98 while (get_gcd(new_obj_size, nrank * nchan) != 1)
100 return new_obj_size * RTE_MEMPOOL_ALIGN;
103 struct pagesz_walk_arg {
109 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
111 struct pagesz_walk_arg *wa = arg;
115 * we need to only look at page sizes available for a particular socket
116 * ID. so, we either need an exact match on socket ID (can match both
117 * native and external memory), or, if SOCKET_ID_ANY was specified as a
118 * socket ID argument, we must only look at native memory and ignore any
119 * page sizes associated with external memory.
121 valid = msl->socket_id == wa->socket_id;
122 valid |= wa->socket_id == SOCKET_ID_ANY && msl->external == 0;
124 if (valid && msl->page_sz < wa->min)
125 wa->min = msl->page_sz;
131 get_min_page_size(int socket_id)
133 struct pagesz_walk_arg wa;
136 wa.socket_id = socket_id;
138 rte_memseg_list_walk(find_min_pagesz, &wa);
140 return wa.min == SIZE_MAX ? (size_t) getpagesize() : wa.min;
145 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
146 void *obj, rte_iova_t iova)
148 struct rte_mempool_objhdr *hdr;
149 struct rte_mempool_objtlr *tlr __rte_unused;
151 /* set mempool ptr in header */
152 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
155 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
156 mp->populated_size++;
158 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
159 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
160 tlr = __mempool_get_trailer(obj);
161 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
165 /* call obj_cb() for each mempool element */
167 rte_mempool_obj_iter(struct rte_mempool *mp,
168 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
170 struct rte_mempool_objhdr *hdr;
174 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
175 obj = (char *)hdr + sizeof(*hdr);
176 obj_cb(mp, obj_cb_arg, obj, n);
183 /* call mem_cb() for each mempool memory chunk */
185 rte_mempool_mem_iter(struct rte_mempool *mp,
186 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
188 struct rte_mempool_memhdr *hdr;
191 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
192 mem_cb(mp, mem_cb_arg, hdr, n);
199 /* get the header, trailer and total size of a mempool element. */
201 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
202 struct rte_mempool_objsz *sz)
204 struct rte_mempool_objsz lsz;
206 sz = (sz != NULL) ? sz : &lsz;
208 sz->header_size = sizeof(struct rte_mempool_objhdr);
209 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
210 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
213 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
214 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
216 sz->trailer_size = 0;
219 /* element size is 8 bytes-aligned at least */
220 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
222 /* expand trailer to next cache line */
223 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
224 sz->total_size = sz->header_size + sz->elt_size +
226 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
227 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
228 RTE_MEMPOOL_ALIGN_MASK);
232 * increase trailer to add padding between objects in order to
233 * spread them across memory channels/ranks
235 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
237 new_size = optimize_object_size(sz->header_size + sz->elt_size +
239 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
242 /* this is the size of an object, including header and trailer */
243 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
245 return sz->total_size;
248 /* free a memchunk allocated with rte_memzone_reserve() */
250 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
253 const struct rte_memzone *mz = opaque;
254 rte_memzone_free(mz);
257 /* Free memory chunks used by a mempool. Objects must be in pool */
259 rte_mempool_free_memchunks(struct rte_mempool *mp)
261 struct rte_mempool_memhdr *memhdr;
264 while (!STAILQ_EMPTY(&mp->elt_list)) {
265 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
267 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
268 mp->populated_size--;
271 while (!STAILQ_EMPTY(&mp->mem_list)) {
272 memhdr = STAILQ_FIRST(&mp->mem_list);
273 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
274 if (memhdr->free_cb != NULL)
275 memhdr->free_cb(memhdr, memhdr->opaque);
282 mempool_ops_alloc_once(struct rte_mempool *mp)
286 /* create the internal ring if not already done */
287 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
288 ret = rte_mempool_ops_alloc(mp);
291 mp->flags |= MEMPOOL_F_POOL_CREATED;
296 /* Add objects in the pool, using a physically contiguous memory
297 * zone. Return the number of objects added, or a negative value
301 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
302 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
307 struct rte_mempool_memhdr *memhdr;
310 ret = mempool_ops_alloc_once(mp);
314 /* mempool is already populated */
315 if (mp->populated_size >= mp->size)
318 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
323 memhdr->addr = vaddr;
326 memhdr->free_cb = free_cb;
327 memhdr->opaque = opaque;
329 if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
330 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
332 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
339 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
341 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
342 len - off, mempool_add_elem, NULL);
344 /* not enough room to store one object */
350 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
359 /* Populate the mempool with a virtual area. Return the number of
360 * objects added, or a negative value on error.
363 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
364 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
368 size_t off, phys_len;
371 /* address and len must be page-aligned */
372 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
374 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
377 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
378 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
379 len, free_cb, opaque);
381 for (off = 0; off + pg_sz <= len &&
382 mp->populated_size < mp->size; off += phys_len) {
384 iova = rte_mem_virt2iova(addr + off);
386 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
391 /* populate with the largest group of contiguous pages */
392 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
395 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
397 if (iova_tmp != iova + phys_len)
401 ret = rte_mempool_populate_iova(mp, addr + off, iova,
402 phys_len, free_cb, opaque);
405 /* no need to call the free callback for next chunks */
413 rte_mempool_free_memchunks(mp);
417 /* Default function to populate the mempool: allocate memory in memzones,
418 * and populate them. Return the number of objects added, or a negative
422 rte_mempool_populate_default(struct rte_mempool *mp)
424 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
425 char mz_name[RTE_MEMZONE_NAMESIZE];
426 const struct rte_memzone *mz;
428 size_t align, pg_sz, pg_shift;
432 bool need_iova_contig_obj;
433 bool try_iova_contig_mempool;
434 bool alloc_in_ext_mem;
436 ret = mempool_ops_alloc_once(mp);
440 /* mempool must not be populated */
441 if (mp->nb_mem_chunks != 0)
445 * the following section calculates page shift and page size values.
447 * these values impact the result of calc_mem_size operation, which
448 * returns the amount of memory that should be allocated to store the
449 * desired number of objects. when not zero, it allocates more memory
450 * for the padding between objects, to ensure that an object does not
451 * cross a page boundary. in other words, page size/shift are to be set
452 * to zero if mempool elements won't care about page boundaries.
453 * there are several considerations for page size and page shift here.
455 * if we don't need our mempools to have physically contiguous objects,
456 * then just set page shift and page size to 0, because the user has
457 * indicated that there's no need to care about anything.
459 * if we do need contiguous objects, there is also an option to reserve
460 * the entire mempool memory as one contiguous block of memory, in
461 * which case the page shift and alignment wouldn't matter as well.
463 * if we require contiguous objects, but not necessarily the entire
464 * mempool reserved space to be contiguous, then there are two options.
466 * if our IO addresses are virtual, not actual physical (IOVA as VA
467 * case), then no page shift needed - our memory allocation will give us
468 * contiguous IO memory as far as the hardware is concerned, so
469 * act as if we're getting contiguous memory.
471 * if our IO addresses are physical, we may get memory from bigger
472 * pages, or we might get memory from smaller pages, and how much of it
473 * we require depends on whether we want bigger or smaller pages.
474 * However, requesting each and every memory size is too much work, so
475 * what we'll do instead is walk through the page sizes available, pick
476 * the smallest one and set up page shift to match that one. We will be
477 * wasting some space this way, but it's much nicer than looping around
478 * trying to reserve each and every page size.
480 * However, since size calculation will produce page-aligned sizes, it
481 * makes sense to first try and see if we can reserve the entire memzone
482 * in one contiguous chunk as well (otherwise we might end up wasting a
483 * 1G page on a 10MB memzone). If we fail to get enough contiguous
484 * memory, then we'll go and reserve space page-by-page.
486 * We also have to take into account the fact that memory that we're
487 * going to allocate from can belong to an externally allocated memory
488 * area, in which case the assumption of IOVA as VA mode being
489 * synonymous with IOVA contiguousness will not hold. We should also try
490 * to go for contiguous memory even if we're in no-huge mode, because
491 * external memory may in fact be IOVA-contiguous.
494 /* check if we can retrieve a valid socket ID */
495 ret = rte_malloc_heap_socket_is_external(mp->socket_id);
498 alloc_in_ext_mem = (ret == 1);
499 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
500 try_iova_contig_mempool = false;
502 if (!need_iova_contig_obj) {
505 } else if (!alloc_in_ext_mem && rte_eal_iova_mode() == RTE_IOVA_VA) {
508 } else if (rte_eal_has_hugepages() || alloc_in_ext_mem) {
509 try_iova_contig_mempool = true;
510 pg_sz = get_min_page_size(mp->socket_id);
511 pg_shift = rte_bsf32(pg_sz);
513 pg_sz = getpagesize();
514 pg_shift = rte_bsf32(pg_sz);
517 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
518 size_t min_chunk_size;
521 if (try_iova_contig_mempool || pg_sz == 0)
522 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
523 0, &min_chunk_size, &align);
525 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
526 pg_shift, &min_chunk_size, &align);
533 ret = snprintf(mz_name, sizeof(mz_name),
534 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
535 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
542 /* if we're trying to reserve contiguous memory, add appropriate
545 if (try_iova_contig_mempool)
546 flags |= RTE_MEMZONE_IOVA_CONTIG;
548 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
549 mp->socket_id, flags, align);
551 /* if we were trying to allocate contiguous memory, failed and
552 * minimum required contiguous chunk fits minimum page, adjust
553 * memzone size to the page size, and try again.
555 if (mz == NULL && try_iova_contig_mempool &&
556 min_chunk_size <= pg_sz) {
557 try_iova_contig_mempool = false;
558 flags &= ~RTE_MEMZONE_IOVA_CONTIG;
560 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
561 pg_shift, &min_chunk_size, &align);
567 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
568 mp->socket_id, flags, align);
570 /* don't try reserving with 0 size if we were asked to reserve
571 * IOVA-contiguous memory.
573 if (min_chunk_size < (size_t)mem_size && mz == NULL) {
574 /* not enough memory, retry with the biggest zone we
577 mz = rte_memzone_reserve_aligned(mz_name, 0,
578 mp->socket_id, flags,
579 RTE_MAX(pg_sz, align));
586 if (mz->len < min_chunk_size) {
587 rte_memzone_free(mz);
592 if (need_iova_contig_obj)
597 if (try_iova_contig_mempool || pg_sz == 0)
598 ret = rte_mempool_populate_iova(mp, mz->addr,
600 rte_mempool_memchunk_mz_free,
601 (void *)(uintptr_t)mz);
603 ret = rte_mempool_populate_virt(mp, mz->addr,
604 RTE_ALIGN_FLOOR(mz->len, pg_sz), pg_sz,
605 rte_mempool_memchunk_mz_free,
606 (void *)(uintptr_t)mz);
608 rte_memzone_free(mz);
616 rte_mempool_free_memchunks(mp);
620 /* return the memory size required for mempool objects in anonymous mem */
622 get_anon_size(const struct rte_mempool *mp)
625 size_t pg_sz, pg_shift;
626 size_t min_chunk_size;
629 pg_sz = getpagesize();
630 pg_shift = rte_bsf32(pg_sz);
631 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
632 &min_chunk_size, &align);
637 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
639 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
645 * Calculate size since memhdr->len has contiguous chunk length
646 * which may be smaller if anon map is split into many contiguous
647 * chunks. Result must be the same as we calculated on populate.
649 size = get_anon_size(memhdr->mp);
653 munmap(opaque, size);
656 /* populate the mempool with an anonymous mapping */
658 rte_mempool_populate_anon(struct rte_mempool *mp)
664 /* mempool is already populated, error */
665 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
670 ret = mempool_ops_alloc_once(mp);
674 size = get_anon_size(mp);
680 /* get chunk of virtually continuous memory */
681 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
682 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
683 if (addr == MAP_FAILED) {
687 /* can't use MMAP_LOCKED, it does not exist on BSD */
688 if (mlock(addr, size) < 0) {
694 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
695 rte_mempool_memchunk_anon_free, addr);
699 return mp->populated_size;
702 rte_mempool_free_memchunks(mp);
708 rte_mempool_free(struct rte_mempool *mp)
710 struct rte_mempool_list *mempool_list = NULL;
711 struct rte_tailq_entry *te;
716 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
717 rte_mcfg_tailq_write_lock();
718 /* find out tailq entry */
719 TAILQ_FOREACH(te, mempool_list, next) {
720 if (te->data == (void *)mp)
725 TAILQ_REMOVE(mempool_list, te, next);
728 rte_mcfg_tailq_write_unlock();
730 rte_mempool_free_memchunks(mp);
731 rte_mempool_ops_free(mp);
732 rte_memzone_free(mp->mz);
736 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
739 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
744 * Create and initialize a cache for objects that are retrieved from and
745 * returned to an underlying mempool. This structure is identical to the
746 * local_cache[lcore_id] pointed to by the mempool structure.
748 struct rte_mempool_cache *
749 rte_mempool_cache_create(uint32_t size, int socket_id)
751 struct rte_mempool_cache *cache;
753 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
758 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
759 RTE_CACHE_LINE_SIZE, socket_id);
761 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
766 mempool_cache_init(cache, size);
772 * Free a cache. It's the responsibility of the user to make sure that any
773 * remaining objects in the cache are flushed to the corresponding
777 rte_mempool_cache_free(struct rte_mempool_cache *cache)
782 /* create an empty mempool */
784 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
785 unsigned cache_size, unsigned private_data_size,
786 int socket_id, unsigned flags)
788 char mz_name[RTE_MEMZONE_NAMESIZE];
789 struct rte_mempool_list *mempool_list;
790 struct rte_mempool *mp = NULL;
791 struct rte_tailq_entry *te = NULL;
792 const struct rte_memzone *mz = NULL;
794 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
795 struct rte_mempool_objsz objsz;
799 /* compilation-time checks */
800 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
801 RTE_CACHE_LINE_MASK) != 0);
802 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
803 RTE_CACHE_LINE_MASK) != 0);
804 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
805 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
806 RTE_CACHE_LINE_MASK) != 0);
807 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
808 RTE_CACHE_LINE_MASK) != 0);
811 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
813 /* asked for zero items */
819 /* asked cache too big */
820 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
821 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
826 /* "no cache align" imply "no spread" */
827 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
828 flags |= MEMPOOL_F_NO_SPREAD;
830 /* calculate mempool object sizes. */
831 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
836 rte_mcfg_mempool_write_lock();
839 * reserve a memory zone for this mempool: private data is
842 private_data_size = (private_data_size +
843 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
846 /* try to allocate tailq entry */
847 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
849 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
853 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
854 mempool_size += private_data_size;
855 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
857 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
858 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
859 rte_errno = ENAMETOOLONG;
863 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
867 /* init the mempool structure */
869 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
870 ret = strlcpy(mp->name, name, sizeof(mp->name));
871 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
872 rte_errno = ENAMETOOLONG;
878 mp->socket_id = socket_id;
879 mp->elt_size = objsz.elt_size;
880 mp->header_size = objsz.header_size;
881 mp->trailer_size = objsz.trailer_size;
882 /* Size of default caches, zero means disabled. */
883 mp->cache_size = cache_size;
884 mp->private_data_size = private_data_size;
885 STAILQ_INIT(&mp->elt_list);
886 STAILQ_INIT(&mp->mem_list);
889 * local_cache pointer is set even if cache_size is zero.
890 * The local_cache points to just past the elt_pa[] array.
892 mp->local_cache = (struct rte_mempool_cache *)
893 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
895 /* Init all default caches. */
896 if (cache_size != 0) {
897 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
898 mempool_cache_init(&mp->local_cache[lcore_id],
904 rte_mcfg_tailq_write_lock();
905 TAILQ_INSERT_TAIL(mempool_list, te, next);
906 rte_mcfg_tailq_write_unlock();
907 rte_mcfg_mempool_write_unlock();
912 rte_mcfg_mempool_write_unlock();
914 rte_mempool_free(mp);
918 /* create the mempool */
920 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
921 unsigned cache_size, unsigned private_data_size,
922 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
923 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
924 int socket_id, unsigned flags)
927 struct rte_mempool *mp;
929 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
930 private_data_size, socket_id, flags);
935 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
936 * set the correct index into the table of ops structs.
938 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
939 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
940 else if (flags & MEMPOOL_F_SP_PUT)
941 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
942 else if (flags & MEMPOOL_F_SC_GET)
943 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
945 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
950 /* call the mempool priv initializer */
952 mp_init(mp, mp_init_arg);
954 if (rte_mempool_populate_default(mp) < 0)
957 /* call the object initializers */
959 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
964 rte_mempool_free(mp);
968 /* Return the number of entries in the mempool */
970 rte_mempool_avail_count(const struct rte_mempool *mp)
975 count = rte_mempool_ops_get_count(mp);
977 if (mp->cache_size == 0)
980 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
981 count += mp->local_cache[lcore_id].len;
984 * due to race condition (access to len is not locked), the
985 * total can be greater than size... so fix the result
987 if (count > mp->size)
992 /* return the number of entries allocated from the mempool */
994 rte_mempool_in_use_count(const struct rte_mempool *mp)
996 return mp->size - rte_mempool_avail_count(mp);
999 /* dump the cache status */
1001 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1005 unsigned cache_count;
1007 fprintf(f, " internal cache infos:\n");
1008 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1010 if (mp->cache_size == 0)
1013 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1014 cache_count = mp->local_cache[lcore_id].len;
1015 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1016 lcore_id, cache_count);
1017 count += cache_count;
1019 fprintf(f, " total_cache_count=%u\n", count);
1023 #ifndef __INTEL_COMPILER
1024 #pragma GCC diagnostic ignored "-Wcast-qual"
1027 /* check and update cookies or panic (internal) */
1028 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1029 void * const *obj_table_const, unsigned n, int free)
1031 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1032 struct rte_mempool_objhdr *hdr;
1033 struct rte_mempool_objtlr *tlr;
1039 /* Force to drop the "const" attribute. This is done only when
1040 * DEBUG is enabled */
1041 tmp = (void *) obj_table_const;
1047 if (rte_mempool_from_obj(obj) != mp)
1048 rte_panic("MEMPOOL: object is owned by another "
1051 hdr = __mempool_get_header(obj);
1052 cookie = hdr->cookie;
1055 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1056 RTE_LOG(CRIT, MEMPOOL,
1057 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1058 obj, (const void *) mp, cookie);
1059 rte_panic("MEMPOOL: bad header cookie (put)\n");
1061 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1062 } else if (free == 1) {
1063 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1064 RTE_LOG(CRIT, MEMPOOL,
1065 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1066 obj, (const void *) mp, cookie);
1067 rte_panic("MEMPOOL: bad header cookie (get)\n");
1069 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1070 } else if (free == 2) {
1071 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1072 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1073 RTE_LOG(CRIT, MEMPOOL,
1074 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1075 obj, (const void *) mp, cookie);
1076 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1079 tlr = __mempool_get_trailer(obj);
1080 cookie = tlr->cookie;
1081 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1082 RTE_LOG(CRIT, MEMPOOL,
1083 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1084 obj, (const void *) mp, cookie);
1085 rte_panic("MEMPOOL: bad trailer cookie\n");
1090 RTE_SET_USED(obj_table_const);
1097 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1098 void * const *first_obj_table_const, unsigned int n, int free)
1100 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1101 struct rte_mempool_info info;
1102 const size_t total_elt_sz =
1103 mp->header_size + mp->elt_size + mp->trailer_size;
1106 rte_mempool_ops_get_info(mp, &info);
1108 for (i = 0; i < n; ++i) {
1109 void *first_obj = first_obj_table_const[i];
1111 for (j = 0; j < info.contig_block_size; ++j) {
1114 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1115 rte_mempool_check_cookies(mp, &obj, 1, free);
1120 RTE_SET_USED(first_obj_table_const);
1126 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1128 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1129 void *obj, __rte_unused unsigned idx)
1131 __mempool_check_cookies(mp, &obj, 1, 2);
1135 mempool_audit_cookies(struct rte_mempool *mp)
1139 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1140 if (num != mp->size) {
1141 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1142 "iterated only over %u elements\n",
1147 #define mempool_audit_cookies(mp) do {} while(0)
1150 #ifndef __INTEL_COMPILER
1151 #pragma GCC diagnostic error "-Wcast-qual"
1154 /* check cookies before and after objects */
1156 mempool_audit_cache(const struct rte_mempool *mp)
1158 /* check cache size consistency */
1161 if (mp->cache_size == 0)
1164 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1165 const struct rte_mempool_cache *cache;
1166 cache = &mp->local_cache[lcore_id];
1167 if (cache->len > cache->flushthresh) {
1168 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1170 rte_panic("MEMPOOL: invalid cache len\n");
1175 /* check the consistency of mempool (size, cookies, ...) */
1177 rte_mempool_audit(struct rte_mempool *mp)
1179 mempool_audit_cache(mp);
1180 mempool_audit_cookies(mp);
1182 /* For case where mempool DEBUG is not set, and cache size is 0 */
1186 /* dump the status of the mempool on the console */
1188 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1190 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1191 struct rte_mempool_info info;
1192 struct rte_mempool_debug_stats sum;
1195 struct rte_mempool_memhdr *memhdr;
1196 unsigned common_count;
1197 unsigned cache_count;
1200 RTE_ASSERT(f != NULL);
1201 RTE_ASSERT(mp != NULL);
1203 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1204 fprintf(f, " flags=%x\n", mp->flags);
1205 fprintf(f, " pool=%p\n", mp->pool_data);
1206 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1207 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1208 fprintf(f, " size=%"PRIu32"\n", mp->size);
1209 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1210 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1211 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1212 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1213 fprintf(f, " total_obj_size=%"PRIu32"\n",
1214 mp->header_size + mp->elt_size + mp->trailer_size);
1216 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1218 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1219 mem_len += memhdr->len;
1221 fprintf(f, " avg bytes/object=%#Lf\n",
1222 (long double)mem_len / mp->size);
1225 cache_count = rte_mempool_dump_cache(f, mp);
1226 common_count = rte_mempool_ops_get_count(mp);
1227 if ((cache_count + common_count) > mp->size)
1228 common_count = mp->size - cache_count;
1229 fprintf(f, " common_pool_count=%u\n", common_count);
1231 /* sum and dump statistics */
1232 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1233 rte_mempool_ops_get_info(mp, &info);
1234 memset(&sum, 0, sizeof(sum));
1235 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1236 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1237 sum.put_objs += mp->stats[lcore_id].put_objs;
1238 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1239 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1240 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1241 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1242 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1243 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1245 fprintf(f, " stats:\n");
1246 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1247 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1248 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1249 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1250 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1251 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1252 if (info.contig_block_size > 0) {
1253 fprintf(f, " get_success_blks=%"PRIu64"\n",
1254 sum.get_success_blks);
1255 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1258 fprintf(f, " no statistics available\n");
1261 rte_mempool_audit(mp);
1264 /* dump the status of all mempools on the console */
1266 rte_mempool_list_dump(FILE *f)
1268 struct rte_mempool *mp = NULL;
1269 struct rte_tailq_entry *te;
1270 struct rte_mempool_list *mempool_list;
1272 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1274 rte_mcfg_mempool_read_lock();
1276 TAILQ_FOREACH(te, mempool_list, next) {
1277 mp = (struct rte_mempool *) te->data;
1278 rte_mempool_dump(f, mp);
1281 rte_mcfg_mempool_read_unlock();
1284 /* search a mempool from its name */
1285 struct rte_mempool *
1286 rte_mempool_lookup(const char *name)
1288 struct rte_mempool *mp = NULL;
1289 struct rte_tailq_entry *te;
1290 struct rte_mempool_list *mempool_list;
1292 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1294 rte_mcfg_mempool_read_lock();
1296 TAILQ_FOREACH(te, mempool_list, next) {
1297 mp = (struct rte_mempool *) te->data;
1298 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1302 rte_mcfg_mempool_read_unlock();
1312 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1315 struct rte_tailq_entry *te = NULL;
1316 struct rte_mempool_list *mempool_list;
1319 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1321 rte_mcfg_mempool_read_lock();
1323 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1324 (*func)((struct rte_mempool *) te->data, arg);
1327 rte_mcfg_mempool_read_unlock();