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_MEMPOOL_ALIGN) - 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);
360 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
361 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
366 ret = __rte_mempool_populate_iova(mp, vaddr, iova, len, free_cb,
377 struct rte_memseg *ms;
379 /* try registered memory first */
380 ms = rte_mem_virt2memseg(addr, NULL);
381 if (ms == NULL || ms->iova == RTE_BAD_IOVA)
382 /* fall back to actual physical address */
383 return rte_mem_virt2iova(addr);
384 return ms->iova + RTE_PTR_DIFF(addr, ms->addr);
387 /* Populate the mempool with a virtual area. Return the number of
388 * objects added, or a negative value on error.
391 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
392 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
396 size_t off, phys_len;
399 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
400 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
401 len, free_cb, opaque);
403 for (off = 0; off < len &&
404 mp->populated_size < mp->size; off += phys_len) {
406 iova = get_iova(addr + off);
408 /* populate with the largest group of contiguous pages */
409 for (phys_len = RTE_MIN(
410 (size_t)(RTE_PTR_ALIGN_CEIL(addr + off + 1, pg_sz) -
413 off + phys_len < len;
414 phys_len = RTE_MIN(phys_len + pg_sz, len - off)) {
417 iova_tmp = get_iova(addr + off + phys_len);
419 if (iova_tmp == RTE_BAD_IOVA ||
420 iova_tmp != iova + phys_len)
424 ret = __rte_mempool_populate_iova(mp, addr + off, iova,
425 phys_len, free_cb, opaque);
430 /* no need to call the free callback for next chunks */
441 rte_mempool_free_memchunks(mp);
445 /* Get the minimal page size used in a mempool before populating it. */
447 rte_mempool_get_page_size(struct rte_mempool *mp, size_t *pg_sz)
449 bool need_iova_contig_obj;
450 bool alloc_in_ext_mem;
453 /* check if we can retrieve a valid socket ID */
454 ret = rte_malloc_heap_socket_is_external(mp->socket_id);
457 alloc_in_ext_mem = (ret == 1);
458 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
460 if (!need_iova_contig_obj)
462 else if (rte_eal_has_hugepages() || alloc_in_ext_mem)
463 *pg_sz = get_min_page_size(mp->socket_id);
465 *pg_sz = getpagesize();
470 /* Default function to populate the mempool: allocate memory in memzones,
471 * and populate them. Return the number of objects added, or a negative
475 rte_mempool_populate_default(struct rte_mempool *mp)
477 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
478 char mz_name[RTE_MEMZONE_NAMESIZE];
479 const struct rte_memzone *mz;
481 size_t align, pg_sz, pg_shift = 0;
485 bool need_iova_contig_obj;
486 size_t max_alloc_size = SIZE_MAX;
488 ret = mempool_ops_alloc_once(mp);
492 /* mempool must not be populated */
493 if (mp->nb_mem_chunks != 0)
497 * the following section calculates page shift and page size values.
499 * these values impact the result of calc_mem_size operation, which
500 * returns the amount of memory that should be allocated to store the
501 * desired number of objects. when not zero, it allocates more memory
502 * for the padding between objects, to ensure that an object does not
503 * cross a page boundary. in other words, page size/shift are to be set
504 * to zero if mempool elements won't care about page boundaries.
505 * there are several considerations for page size and page shift here.
507 * if we don't need our mempools to have physically contiguous objects,
508 * then just set page shift and page size to 0, because the user has
509 * indicated that there's no need to care about anything.
511 * if we do need contiguous objects (if a mempool driver has its
512 * own calc_size() method returning min_chunk_size = mem_size),
513 * there is also an option to reserve the entire mempool memory
514 * as one contiguous block of memory.
516 * if we require contiguous objects, but not necessarily the entire
517 * mempool reserved space to be contiguous, pg_sz will be != 0,
518 * and the default ops->populate() will take care of not placing
519 * objects across pages.
521 * if our IO addresses are physical, we may get memory from bigger
522 * pages, or we might get memory from smaller pages, and how much of it
523 * we require depends on whether we want bigger or smaller pages.
524 * However, requesting each and every memory size is too much work, so
525 * what we'll do instead is walk through the page sizes available, pick
526 * the smallest one and set up page shift to match that one. We will be
527 * wasting some space this way, but it's much nicer than looping around
528 * trying to reserve each and every page size.
530 * If we fail to get enough contiguous memory, then we'll go and
531 * reserve space in smaller chunks.
534 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
535 ret = rte_mempool_get_page_size(mp, &pg_sz);
540 pg_shift = rte_bsf32(pg_sz);
542 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
543 size_t min_chunk_size;
545 mem_size = rte_mempool_ops_calc_mem_size(
546 mp, n, pg_shift, &min_chunk_size, &align);
553 ret = snprintf(mz_name, sizeof(mz_name),
554 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
555 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
560 /* if we're trying to reserve contiguous memory, add appropriate
563 if (min_chunk_size == (size_t)mem_size)
564 mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
566 /* Allocate a memzone, retrying with a smaller area on ENOMEM */
568 mz = rte_memzone_reserve_aligned(mz_name,
569 RTE_MIN((size_t)mem_size, max_alloc_size),
570 mp->socket_id, mz_flags, align);
572 if (mz == NULL && rte_errno != ENOMEM)
575 max_alloc_size = RTE_MIN(max_alloc_size,
576 (size_t)mem_size) / 2;
577 } while (mz == NULL && max_alloc_size >= min_chunk_size);
584 if (need_iova_contig_obj)
589 if (pg_sz == 0 || (mz_flags & RTE_MEMZONE_IOVA_CONTIG))
590 ret = rte_mempool_populate_iova(mp, mz->addr,
592 rte_mempool_memchunk_mz_free,
593 (void *)(uintptr_t)mz);
595 ret = rte_mempool_populate_virt(mp, mz->addr,
597 rte_mempool_memchunk_mz_free,
598 (void *)(uintptr_t)mz);
600 rte_memzone_free(mz);
608 rte_mempool_free_memchunks(mp);
612 /* return the memory size required for mempool objects in anonymous mem */
614 get_anon_size(const struct rte_mempool *mp)
617 size_t pg_sz, pg_shift;
618 size_t min_chunk_size;
621 pg_sz = getpagesize();
622 pg_shift = rte_bsf32(pg_sz);
623 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
624 &min_chunk_size, &align);
629 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
631 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
637 * Calculate size since memhdr->len has contiguous chunk length
638 * which may be smaller if anon map is split into many contiguous
639 * chunks. Result must be the same as we calculated on populate.
641 size = get_anon_size(memhdr->mp);
645 munmap(opaque, size);
648 /* populate the mempool with an anonymous mapping */
650 rte_mempool_populate_anon(struct rte_mempool *mp)
656 /* mempool is already populated, error */
657 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
662 ret = mempool_ops_alloc_once(mp);
668 size = get_anon_size(mp);
674 /* get chunk of virtually continuous memory */
675 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
676 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
677 if (addr == MAP_FAILED) {
681 /* can't use MMAP_LOCKED, it does not exist on BSD */
682 if (mlock(addr, size) < 0) {
688 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
689 rte_mempool_memchunk_anon_free, addr);
695 return mp->populated_size;
698 rte_mempool_free_memchunks(mp);
704 rte_mempool_free(struct rte_mempool *mp)
706 struct rte_mempool_list *mempool_list = NULL;
707 struct rte_tailq_entry *te;
712 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
713 rte_mcfg_tailq_write_lock();
714 /* find out tailq entry */
715 TAILQ_FOREACH(te, mempool_list, next) {
716 if (te->data == (void *)mp)
721 TAILQ_REMOVE(mempool_list, te, next);
724 rte_mcfg_tailq_write_unlock();
726 rte_mempool_free_memchunks(mp);
727 rte_mempool_ops_free(mp);
728 rte_memzone_free(mp->mz);
732 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
735 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
740 * Create and initialize a cache for objects that are retrieved from and
741 * returned to an underlying mempool. This structure is identical to the
742 * local_cache[lcore_id] pointed to by the mempool structure.
744 struct rte_mempool_cache *
745 rte_mempool_cache_create(uint32_t size, int socket_id)
747 struct rte_mempool_cache *cache;
749 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
754 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
755 RTE_CACHE_LINE_SIZE, socket_id);
757 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
762 mempool_cache_init(cache, size);
768 * Free a cache. It's the responsibility of the user to make sure that any
769 * remaining objects in the cache are flushed to the corresponding
773 rte_mempool_cache_free(struct rte_mempool_cache *cache)
778 /* create an empty mempool */
780 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
781 unsigned cache_size, unsigned private_data_size,
782 int socket_id, unsigned flags)
784 char mz_name[RTE_MEMZONE_NAMESIZE];
785 struct rte_mempool_list *mempool_list;
786 struct rte_mempool *mp = NULL;
787 struct rte_tailq_entry *te = NULL;
788 const struct rte_memzone *mz = NULL;
790 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
791 struct rte_mempool_objsz objsz;
795 /* compilation-time checks */
796 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
797 RTE_CACHE_LINE_MASK) != 0);
798 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
799 RTE_CACHE_LINE_MASK) != 0);
800 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
801 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
802 RTE_CACHE_LINE_MASK) != 0);
803 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
804 RTE_CACHE_LINE_MASK) != 0);
807 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
809 /* asked for zero items */
815 /* asked cache too big */
816 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
817 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
822 /* "no cache align" imply "no spread" */
823 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
824 flags |= MEMPOOL_F_NO_SPREAD;
826 /* calculate mempool object sizes. */
827 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
832 rte_mcfg_mempool_write_lock();
835 * reserve a memory zone for this mempool: private data is
838 private_data_size = (private_data_size +
839 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
842 /* try to allocate tailq entry */
843 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
845 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
849 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
850 mempool_size += private_data_size;
851 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
853 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
854 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
855 rte_errno = ENAMETOOLONG;
859 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
863 /* init the mempool structure */
865 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
866 ret = strlcpy(mp->name, name, sizeof(mp->name));
867 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
868 rte_errno = ENAMETOOLONG;
874 mp->socket_id = socket_id;
875 mp->elt_size = objsz.elt_size;
876 mp->header_size = objsz.header_size;
877 mp->trailer_size = objsz.trailer_size;
878 /* Size of default caches, zero means disabled. */
879 mp->cache_size = cache_size;
880 mp->private_data_size = private_data_size;
881 STAILQ_INIT(&mp->elt_list);
882 STAILQ_INIT(&mp->mem_list);
885 * local_cache pointer is set even if cache_size is zero.
886 * The local_cache points to just past the elt_pa[] array.
888 mp->local_cache = (struct rte_mempool_cache *)
889 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
891 /* Init all default caches. */
892 if (cache_size != 0) {
893 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
894 mempool_cache_init(&mp->local_cache[lcore_id],
900 rte_mcfg_tailq_write_lock();
901 TAILQ_INSERT_TAIL(mempool_list, te, next);
902 rte_mcfg_tailq_write_unlock();
903 rte_mcfg_mempool_write_unlock();
908 rte_mcfg_mempool_write_unlock();
910 rte_mempool_free(mp);
914 /* create the mempool */
916 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
917 unsigned cache_size, unsigned private_data_size,
918 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
919 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
920 int socket_id, unsigned flags)
923 struct rte_mempool *mp;
925 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
926 private_data_size, socket_id, flags);
931 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
932 * set the correct index into the table of ops structs.
934 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
935 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
936 else if (flags & MEMPOOL_F_SP_PUT)
937 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
938 else if (flags & MEMPOOL_F_SC_GET)
939 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
941 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
946 /* call the mempool priv initializer */
948 mp_init(mp, mp_init_arg);
950 if (rte_mempool_populate_default(mp) < 0)
953 /* call the object initializers */
955 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
960 rte_mempool_free(mp);
964 /* Return the number of entries in the mempool */
966 rte_mempool_avail_count(const struct rte_mempool *mp)
971 count = rte_mempool_ops_get_count(mp);
973 if (mp->cache_size == 0)
976 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
977 count += mp->local_cache[lcore_id].len;
980 * due to race condition (access to len is not locked), the
981 * total can be greater than size... so fix the result
983 if (count > mp->size)
988 /* return the number of entries allocated from the mempool */
990 rte_mempool_in_use_count(const struct rte_mempool *mp)
992 return mp->size - rte_mempool_avail_count(mp);
995 /* dump the cache status */
997 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1001 unsigned cache_count;
1003 fprintf(f, " internal cache infos:\n");
1004 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1006 if (mp->cache_size == 0)
1009 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1010 cache_count = mp->local_cache[lcore_id].len;
1011 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1012 lcore_id, cache_count);
1013 count += cache_count;
1015 fprintf(f, " total_cache_count=%u\n", count);
1019 #ifndef __INTEL_COMPILER
1020 #pragma GCC diagnostic ignored "-Wcast-qual"
1023 /* check and update cookies or panic (internal) */
1024 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1025 void * const *obj_table_const, unsigned n, int free)
1027 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1028 struct rte_mempool_objhdr *hdr;
1029 struct rte_mempool_objtlr *tlr;
1035 /* Force to drop the "const" attribute. This is done only when
1036 * DEBUG is enabled */
1037 tmp = (void *) obj_table_const;
1043 if (rte_mempool_from_obj(obj) != mp)
1044 rte_panic("MEMPOOL: object is owned by another "
1047 hdr = __mempool_get_header(obj);
1048 cookie = hdr->cookie;
1051 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1052 RTE_LOG(CRIT, MEMPOOL,
1053 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1054 obj, (const void *) mp, cookie);
1055 rte_panic("MEMPOOL: bad header cookie (put)\n");
1057 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1058 } else if (free == 1) {
1059 if (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 (get)\n");
1065 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1066 } else if (free == 2) {
1067 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1068 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1069 RTE_LOG(CRIT, MEMPOOL,
1070 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1071 obj, (const void *) mp, cookie);
1072 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1075 tlr = __mempool_get_trailer(obj);
1076 cookie = tlr->cookie;
1077 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1078 RTE_LOG(CRIT, MEMPOOL,
1079 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1080 obj, (const void *) mp, cookie);
1081 rte_panic("MEMPOOL: bad trailer cookie\n");
1086 RTE_SET_USED(obj_table_const);
1093 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1094 void * const *first_obj_table_const, unsigned int n, int free)
1096 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1097 struct rte_mempool_info info;
1098 const size_t total_elt_sz =
1099 mp->header_size + mp->elt_size + mp->trailer_size;
1102 rte_mempool_ops_get_info(mp, &info);
1104 for (i = 0; i < n; ++i) {
1105 void *first_obj = first_obj_table_const[i];
1107 for (j = 0; j < info.contig_block_size; ++j) {
1110 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1111 rte_mempool_check_cookies(mp, &obj, 1, free);
1116 RTE_SET_USED(first_obj_table_const);
1122 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1124 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1125 void *obj, __rte_unused unsigned idx)
1127 __mempool_check_cookies(mp, &obj, 1, 2);
1131 mempool_audit_cookies(struct rte_mempool *mp)
1135 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1136 if (num != mp->size) {
1137 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1138 "iterated only over %u elements\n",
1143 #define mempool_audit_cookies(mp) do {} while(0)
1146 #ifndef __INTEL_COMPILER
1147 #pragma GCC diagnostic error "-Wcast-qual"
1150 /* check cookies before and after objects */
1152 mempool_audit_cache(const struct rte_mempool *mp)
1154 /* check cache size consistency */
1157 if (mp->cache_size == 0)
1160 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1161 const struct rte_mempool_cache *cache;
1162 cache = &mp->local_cache[lcore_id];
1163 if (cache->len > cache->flushthresh) {
1164 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1166 rte_panic("MEMPOOL: invalid cache len\n");
1171 /* check the consistency of mempool (size, cookies, ...) */
1173 rte_mempool_audit(struct rte_mempool *mp)
1175 mempool_audit_cache(mp);
1176 mempool_audit_cookies(mp);
1178 /* For case where mempool DEBUG is not set, and cache size is 0 */
1182 /* dump the status of the mempool on the console */
1184 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1186 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1187 struct rte_mempool_info info;
1188 struct rte_mempool_debug_stats sum;
1191 struct rte_mempool_memhdr *memhdr;
1192 unsigned common_count;
1193 unsigned cache_count;
1196 RTE_ASSERT(f != NULL);
1197 RTE_ASSERT(mp != NULL);
1199 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1200 fprintf(f, " flags=%x\n", mp->flags);
1201 fprintf(f, " pool=%p\n", mp->pool_data);
1202 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1203 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1204 fprintf(f, " size=%"PRIu32"\n", mp->size);
1205 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1206 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1207 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1208 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1209 fprintf(f, " total_obj_size=%"PRIu32"\n",
1210 mp->header_size + mp->elt_size + mp->trailer_size);
1212 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1214 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1215 mem_len += memhdr->len;
1217 fprintf(f, " avg bytes/object=%#Lf\n",
1218 (long double)mem_len / mp->size);
1221 cache_count = rte_mempool_dump_cache(f, mp);
1222 common_count = rte_mempool_ops_get_count(mp);
1223 if ((cache_count + common_count) > mp->size)
1224 common_count = mp->size - cache_count;
1225 fprintf(f, " common_pool_count=%u\n", common_count);
1227 /* sum and dump statistics */
1228 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1229 rte_mempool_ops_get_info(mp, &info);
1230 memset(&sum, 0, sizeof(sum));
1231 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1232 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1233 sum.put_objs += mp->stats[lcore_id].put_objs;
1234 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1235 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1236 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1237 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1238 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1239 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1241 fprintf(f, " stats:\n");
1242 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1243 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1244 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1245 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1246 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1247 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1248 if (info.contig_block_size > 0) {
1249 fprintf(f, " get_success_blks=%"PRIu64"\n",
1250 sum.get_success_blks);
1251 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1254 fprintf(f, " no statistics available\n");
1257 rte_mempool_audit(mp);
1260 /* dump the status of all mempools on the console */
1262 rte_mempool_list_dump(FILE *f)
1264 struct rte_mempool *mp = NULL;
1265 struct rte_tailq_entry *te;
1266 struct rte_mempool_list *mempool_list;
1268 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1270 rte_mcfg_mempool_read_lock();
1272 TAILQ_FOREACH(te, mempool_list, next) {
1273 mp = (struct rte_mempool *) te->data;
1274 rte_mempool_dump(f, mp);
1277 rte_mcfg_mempool_read_unlock();
1280 /* search a mempool from its name */
1281 struct rte_mempool *
1282 rte_mempool_lookup(const char *name)
1284 struct rte_mempool *mp = NULL;
1285 struct rte_tailq_entry *te;
1286 struct rte_mempool_list *mempool_list;
1288 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1290 rte_mcfg_mempool_read_lock();
1292 TAILQ_FOREACH(te, mempool_list, next) {
1293 mp = (struct rte_mempool *) te->data;
1294 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1298 rte_mcfg_mempool_read_unlock();
1308 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1311 struct rte_tailq_entry *te = NULL;
1312 struct rte_mempool_list *mempool_list;
1315 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1317 rte_mcfg_mempool_read_lock();
1319 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1320 (*func)((struct rte_mempool *) te->data, arg);
1323 rte_mcfg_mempool_read_unlock();