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))
48 #if defined(RTE_ARCH_X86)
50 * return the greatest common divisor between a and b (fast algorithm)
53 static unsigned get_gcd(unsigned a, unsigned b)
78 * Depending on memory configuration on x86 arch, objects addresses are spread
79 * between channels and ranks in RAM: the pool allocator will add
80 * padding between objects. This function return the new size of the
84 arch_mem_object_align(unsigned int obj_size)
86 unsigned nrank, nchan;
87 unsigned new_obj_size;
89 /* get number of channels */
90 nchan = rte_memory_get_nchannel();
94 nrank = rte_memory_get_nrank();
98 /* process new object size */
99 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
100 while (get_gcd(new_obj_size, nrank * nchan) != 1)
102 return new_obj_size * RTE_MEMPOOL_ALIGN;
106 arch_mem_object_align(unsigned int obj_size)
112 struct pagesz_walk_arg {
118 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
120 struct pagesz_walk_arg *wa = arg;
124 * we need to only look at page sizes available for a particular socket
125 * ID. so, we either need an exact match on socket ID (can match both
126 * native and external memory), or, if SOCKET_ID_ANY was specified as a
127 * socket ID argument, we must only look at native memory and ignore any
128 * page sizes associated with external memory.
130 valid = msl->socket_id == wa->socket_id;
131 valid |= wa->socket_id == SOCKET_ID_ANY && msl->external == 0;
133 if (valid && msl->page_sz < wa->min)
134 wa->min = msl->page_sz;
140 get_min_page_size(int socket_id)
142 struct pagesz_walk_arg wa;
145 wa.socket_id = socket_id;
147 rte_memseg_list_walk(find_min_pagesz, &wa);
149 return wa.min == SIZE_MAX ? (size_t) getpagesize() : wa.min;
154 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
155 void *obj, rte_iova_t iova)
157 struct rte_mempool_objhdr *hdr;
158 struct rte_mempool_objtlr *tlr __rte_unused;
160 /* set mempool ptr in header */
161 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
164 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
165 mp->populated_size++;
167 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
168 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
169 tlr = __mempool_get_trailer(obj);
170 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
174 /* call obj_cb() for each mempool element */
176 rte_mempool_obj_iter(struct rte_mempool *mp,
177 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
179 struct rte_mempool_objhdr *hdr;
183 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
184 obj = (char *)hdr + sizeof(*hdr);
185 obj_cb(mp, obj_cb_arg, obj, n);
192 /* call mem_cb() for each mempool memory chunk */
194 rte_mempool_mem_iter(struct rte_mempool *mp,
195 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
197 struct rte_mempool_memhdr *hdr;
200 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
201 mem_cb(mp, mem_cb_arg, hdr, n);
208 /* get the header, trailer and total size of a mempool element. */
210 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
211 struct rte_mempool_objsz *sz)
213 struct rte_mempool_objsz lsz;
215 sz = (sz != NULL) ? sz : &lsz;
217 sz->header_size = sizeof(struct rte_mempool_objhdr);
218 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
219 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
222 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
223 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
225 sz->trailer_size = 0;
228 /* element size is 8 bytes-aligned at least */
229 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
231 /* expand trailer to next cache line */
232 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
233 sz->total_size = sz->header_size + sz->elt_size +
235 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
236 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
237 RTE_MEMPOOL_ALIGN_MASK);
241 * increase trailer to add padding between objects in order to
242 * spread them across memory channels/ranks
244 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
246 new_size = arch_mem_object_align
247 (sz->header_size + sz->elt_size + sz->trailer_size);
248 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
251 /* this is the size of an object, including header and trailer */
252 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
254 return sz->total_size;
257 /* free a memchunk allocated with rte_memzone_reserve() */
259 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
262 const struct rte_memzone *mz = opaque;
263 rte_memzone_free(mz);
266 /* Free memory chunks used by a mempool. Objects must be in pool */
268 rte_mempool_free_memchunks(struct rte_mempool *mp)
270 struct rte_mempool_memhdr *memhdr;
273 while (!STAILQ_EMPTY(&mp->elt_list)) {
274 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
276 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
277 mp->populated_size--;
280 while (!STAILQ_EMPTY(&mp->mem_list)) {
281 memhdr = STAILQ_FIRST(&mp->mem_list);
282 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
283 if (memhdr->free_cb != NULL)
284 memhdr->free_cb(memhdr, memhdr->opaque);
291 mempool_ops_alloc_once(struct rte_mempool *mp)
295 /* create the internal ring if not already done */
296 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
297 ret = rte_mempool_ops_alloc(mp);
300 mp->flags |= MEMPOOL_F_POOL_CREATED;
305 /* Add objects in the pool, using a physically contiguous memory
306 * zone. Return the number of objects added, or a negative value
310 __rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
311 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
316 struct rte_mempool_memhdr *memhdr;
319 ret = mempool_ops_alloc_once(mp);
323 /* mempool is already populated */
324 if (mp->populated_size >= mp->size)
327 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
332 memhdr->addr = vaddr;
335 memhdr->free_cb = free_cb;
336 memhdr->opaque = opaque;
338 if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
339 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
341 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_MEMPOOL_ALIGN) - vaddr;
348 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
350 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
351 len - off, mempool_add_elem, NULL);
353 /* not enough room to store one object */
359 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
369 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
370 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
375 ret = __rte_mempool_populate_iova(mp, vaddr, iova, len, free_cb,
386 struct rte_memseg *ms;
388 /* try registered memory first */
389 ms = rte_mem_virt2memseg(addr, NULL);
390 if (ms == NULL || ms->iova == RTE_BAD_IOVA)
391 /* fall back to actual physical address */
392 return rte_mem_virt2iova(addr);
393 return ms->iova + RTE_PTR_DIFF(addr, ms->addr);
396 /* Populate the mempool with a virtual area. Return the number of
397 * objects added, or a negative value on error.
400 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
401 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
405 size_t off, phys_len;
408 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
409 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
410 len, free_cb, opaque);
412 for (off = 0; off < len &&
413 mp->populated_size < mp->size; off += phys_len) {
415 iova = get_iova(addr + off);
417 /* populate with the largest group of contiguous pages */
418 for (phys_len = RTE_MIN(
419 (size_t)(RTE_PTR_ALIGN_CEIL(addr + off + 1, pg_sz) -
422 off + phys_len < len;
423 phys_len = RTE_MIN(phys_len + pg_sz, len - off)) {
426 iova_tmp = get_iova(addr + off + phys_len);
428 if (iova_tmp == RTE_BAD_IOVA ||
429 iova_tmp != iova + phys_len)
433 ret = __rte_mempool_populate_iova(mp, addr + off, iova,
434 phys_len, free_cb, opaque);
439 /* no need to call the free callback for next chunks */
450 rte_mempool_free_memchunks(mp);
454 /* Get the minimal page size used in a mempool before populating it. */
456 rte_mempool_get_page_size(struct rte_mempool *mp, size_t *pg_sz)
458 bool need_iova_contig_obj;
459 bool alloc_in_ext_mem;
462 /* check if we can retrieve a valid socket ID */
463 ret = rte_malloc_heap_socket_is_external(mp->socket_id);
466 alloc_in_ext_mem = (ret == 1);
467 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
469 if (!need_iova_contig_obj)
471 else if (rte_eal_has_hugepages() || alloc_in_ext_mem)
472 *pg_sz = get_min_page_size(mp->socket_id);
474 *pg_sz = getpagesize();
479 /* Default function to populate the mempool: allocate memory in memzones,
480 * and populate them. Return the number of objects added, or a negative
484 rte_mempool_populate_default(struct rte_mempool *mp)
486 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
487 char mz_name[RTE_MEMZONE_NAMESIZE];
488 const struct rte_memzone *mz;
490 size_t align, pg_sz, pg_shift = 0;
494 bool need_iova_contig_obj;
495 size_t max_alloc_size = SIZE_MAX;
497 ret = mempool_ops_alloc_once(mp);
501 /* mempool must not be populated */
502 if (mp->nb_mem_chunks != 0)
506 * the following section calculates page shift and page size values.
508 * these values impact the result of calc_mem_size operation, which
509 * returns the amount of memory that should be allocated to store the
510 * desired number of objects. when not zero, it allocates more memory
511 * for the padding between objects, to ensure that an object does not
512 * cross a page boundary. in other words, page size/shift are to be set
513 * to zero if mempool elements won't care about page boundaries.
514 * there are several considerations for page size and page shift here.
516 * if we don't need our mempools to have physically contiguous objects,
517 * then just set page shift and page size to 0, because the user has
518 * indicated that there's no need to care about anything.
520 * if we do need contiguous objects (if a mempool driver has its
521 * own calc_size() method returning min_chunk_size = mem_size),
522 * there is also an option to reserve the entire mempool memory
523 * as one contiguous block of memory.
525 * if we require contiguous objects, but not necessarily the entire
526 * mempool reserved space to be contiguous, pg_sz will be != 0,
527 * and the default ops->populate() will take care of not placing
528 * objects across pages.
530 * if our IO addresses are physical, we may get memory from bigger
531 * pages, or we might get memory from smaller pages, and how much of it
532 * we require depends on whether we want bigger or smaller pages.
533 * However, requesting each and every memory size is too much work, so
534 * what we'll do instead is walk through the page sizes available, pick
535 * the smallest one and set up page shift to match that one. We will be
536 * wasting some space this way, but it's much nicer than looping around
537 * trying to reserve each and every page size.
539 * If we fail to get enough contiguous memory, then we'll go and
540 * reserve space in smaller chunks.
543 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
544 ret = rte_mempool_get_page_size(mp, &pg_sz);
549 pg_shift = rte_bsf32(pg_sz);
551 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
552 size_t min_chunk_size;
554 mem_size = rte_mempool_ops_calc_mem_size(
555 mp, n, pg_shift, &min_chunk_size, &align);
562 ret = snprintf(mz_name, sizeof(mz_name),
563 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
564 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
569 /* if we're trying to reserve contiguous memory, add appropriate
572 if (min_chunk_size == (size_t)mem_size)
573 mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
575 /* Allocate a memzone, retrying with a smaller area on ENOMEM */
577 mz = rte_memzone_reserve_aligned(mz_name,
578 RTE_MIN((size_t)mem_size, max_alloc_size),
579 mp->socket_id, mz_flags, align);
581 if (mz == NULL && rte_errno != ENOMEM)
584 max_alloc_size = RTE_MIN(max_alloc_size,
585 (size_t)mem_size) / 2;
586 } while (mz == NULL && max_alloc_size >= min_chunk_size);
593 if (need_iova_contig_obj)
598 if (pg_sz == 0 || (mz_flags & RTE_MEMZONE_IOVA_CONTIG))
599 ret = rte_mempool_populate_iova(mp, mz->addr,
601 rte_mempool_memchunk_mz_free,
602 (void *)(uintptr_t)mz);
604 ret = rte_mempool_populate_virt(mp, mz->addr,
606 rte_mempool_memchunk_mz_free,
607 (void *)(uintptr_t)mz);
609 rte_memzone_free(mz);
617 rte_mempool_free_memchunks(mp);
621 /* return the memory size required for mempool objects in anonymous mem */
623 get_anon_size(const struct rte_mempool *mp)
626 size_t pg_sz, pg_shift;
627 size_t min_chunk_size;
630 pg_sz = getpagesize();
631 pg_shift = rte_bsf32(pg_sz);
632 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
633 &min_chunk_size, &align);
638 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
640 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
646 * Calculate size since memhdr->len has contiguous chunk length
647 * which may be smaller if anon map is split into many contiguous
648 * chunks. Result must be the same as we calculated on populate.
650 size = get_anon_size(memhdr->mp);
654 munmap(opaque, size);
657 /* populate the mempool with an anonymous mapping */
659 rte_mempool_populate_anon(struct rte_mempool *mp)
665 /* mempool is already populated, error */
666 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
671 ret = mempool_ops_alloc_once(mp);
677 size = get_anon_size(mp);
683 /* get chunk of virtually continuous memory */
684 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
685 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
686 if (addr == MAP_FAILED) {
690 /* can't use MMAP_LOCKED, it does not exist on BSD */
691 if (mlock(addr, size) < 0) {
697 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
698 rte_mempool_memchunk_anon_free, addr);
704 return mp->populated_size;
707 rte_mempool_free_memchunks(mp);
713 rte_mempool_free(struct rte_mempool *mp)
715 struct rte_mempool_list *mempool_list = NULL;
716 struct rte_tailq_entry *te;
721 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
722 rte_mcfg_tailq_write_lock();
723 /* find out tailq entry */
724 TAILQ_FOREACH(te, mempool_list, next) {
725 if (te->data == (void *)mp)
730 TAILQ_REMOVE(mempool_list, te, next);
733 rte_mcfg_tailq_write_unlock();
735 rte_mempool_free_memchunks(mp);
736 rte_mempool_ops_free(mp);
737 rte_memzone_free(mp->mz);
741 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
744 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
749 * Create and initialize a cache for objects that are retrieved from and
750 * returned to an underlying mempool. This structure is identical to the
751 * local_cache[lcore_id] pointed to by the mempool structure.
753 struct rte_mempool_cache *
754 rte_mempool_cache_create(uint32_t size, int socket_id)
756 struct rte_mempool_cache *cache;
758 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
763 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
764 RTE_CACHE_LINE_SIZE, socket_id);
766 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
771 mempool_cache_init(cache, size);
777 * Free a cache. It's the responsibility of the user to make sure that any
778 * remaining objects in the cache are flushed to the corresponding
782 rte_mempool_cache_free(struct rte_mempool_cache *cache)
787 /* create an empty mempool */
789 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
790 unsigned cache_size, unsigned private_data_size,
791 int socket_id, unsigned flags)
793 char mz_name[RTE_MEMZONE_NAMESIZE];
794 struct rte_mempool_list *mempool_list;
795 struct rte_mempool *mp = NULL;
796 struct rte_tailq_entry *te = NULL;
797 const struct rte_memzone *mz = NULL;
799 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
800 struct rte_mempool_objsz objsz;
804 /* compilation-time checks */
805 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
806 RTE_CACHE_LINE_MASK) != 0);
807 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
808 RTE_CACHE_LINE_MASK) != 0);
809 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
810 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
811 RTE_CACHE_LINE_MASK) != 0);
812 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
813 RTE_CACHE_LINE_MASK) != 0);
816 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
818 /* asked for zero items */
824 /* asked cache too big */
825 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
826 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
831 /* "no cache align" imply "no spread" */
832 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
833 flags |= MEMPOOL_F_NO_SPREAD;
835 /* calculate mempool object sizes. */
836 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
841 rte_mcfg_mempool_write_lock();
844 * reserve a memory zone for this mempool: private data is
847 private_data_size = (private_data_size +
848 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
851 /* try to allocate tailq entry */
852 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
854 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
858 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
859 mempool_size += private_data_size;
860 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
862 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
863 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
864 rte_errno = ENAMETOOLONG;
868 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
872 /* init the mempool structure */
874 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
875 ret = strlcpy(mp->name, name, sizeof(mp->name));
876 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
877 rte_errno = ENAMETOOLONG;
883 mp->socket_id = socket_id;
884 mp->elt_size = objsz.elt_size;
885 mp->header_size = objsz.header_size;
886 mp->trailer_size = objsz.trailer_size;
887 /* Size of default caches, zero means disabled. */
888 mp->cache_size = cache_size;
889 mp->private_data_size = private_data_size;
890 STAILQ_INIT(&mp->elt_list);
891 STAILQ_INIT(&mp->mem_list);
894 * local_cache pointer is set even if cache_size is zero.
895 * The local_cache points to just past the elt_pa[] array.
897 mp->local_cache = (struct rte_mempool_cache *)
898 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
900 /* Init all default caches. */
901 if (cache_size != 0) {
902 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
903 mempool_cache_init(&mp->local_cache[lcore_id],
909 rte_mcfg_tailq_write_lock();
910 TAILQ_INSERT_TAIL(mempool_list, te, next);
911 rte_mcfg_tailq_write_unlock();
912 rte_mcfg_mempool_write_unlock();
917 rte_mcfg_mempool_write_unlock();
919 rte_mempool_free(mp);
923 /* create the mempool */
925 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
926 unsigned cache_size, unsigned private_data_size,
927 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
928 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
929 int socket_id, unsigned flags)
932 struct rte_mempool *mp;
934 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
935 private_data_size, socket_id, flags);
940 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
941 * set the correct index into the table of ops structs.
943 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
944 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
945 else if (flags & MEMPOOL_F_SP_PUT)
946 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
947 else if (flags & MEMPOOL_F_SC_GET)
948 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
950 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
955 /* call the mempool priv initializer */
957 mp_init(mp, mp_init_arg);
959 if (rte_mempool_populate_default(mp) < 0)
962 /* call the object initializers */
964 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
969 rte_mempool_free(mp);
973 /* Return the number of entries in the mempool */
975 rte_mempool_avail_count(const struct rte_mempool *mp)
980 count = rte_mempool_ops_get_count(mp);
982 if (mp->cache_size == 0)
985 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
986 count += mp->local_cache[lcore_id].len;
989 * due to race condition (access to len is not locked), the
990 * total can be greater than size... so fix the result
992 if (count > mp->size)
997 /* return the number of entries allocated from the mempool */
999 rte_mempool_in_use_count(const struct rte_mempool *mp)
1001 return mp->size - rte_mempool_avail_count(mp);
1004 /* dump the cache status */
1006 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1010 unsigned cache_count;
1012 fprintf(f, " internal cache infos:\n");
1013 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1015 if (mp->cache_size == 0)
1018 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1019 cache_count = mp->local_cache[lcore_id].len;
1020 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1021 lcore_id, cache_count);
1022 count += cache_count;
1024 fprintf(f, " total_cache_count=%u\n", count);
1028 #ifndef __INTEL_COMPILER
1029 #pragma GCC diagnostic ignored "-Wcast-qual"
1032 /* check and update cookies or panic (internal) */
1033 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1034 void * const *obj_table_const, unsigned n, int free)
1036 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1037 struct rte_mempool_objhdr *hdr;
1038 struct rte_mempool_objtlr *tlr;
1044 /* Force to drop the "const" attribute. This is done only when
1045 * DEBUG is enabled */
1046 tmp = (void *) obj_table_const;
1052 if (rte_mempool_from_obj(obj) != mp)
1053 rte_panic("MEMPOOL: object is owned by another "
1056 hdr = __mempool_get_header(obj);
1057 cookie = hdr->cookie;
1060 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1061 RTE_LOG(CRIT, MEMPOOL,
1062 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1063 obj, (const void *) mp, cookie);
1064 rte_panic("MEMPOOL: bad header cookie (put)\n");
1066 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1067 } else if (free == 1) {
1068 if (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 (get)\n");
1074 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1075 } else if (free == 2) {
1076 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1077 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1078 RTE_LOG(CRIT, MEMPOOL,
1079 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1080 obj, (const void *) mp, cookie);
1081 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1084 tlr = __mempool_get_trailer(obj);
1085 cookie = tlr->cookie;
1086 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1087 RTE_LOG(CRIT, MEMPOOL,
1088 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1089 obj, (const void *) mp, cookie);
1090 rte_panic("MEMPOOL: bad trailer cookie\n");
1095 RTE_SET_USED(obj_table_const);
1102 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1103 void * const *first_obj_table_const, unsigned int n, int free)
1105 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1106 struct rte_mempool_info info;
1107 const size_t total_elt_sz =
1108 mp->header_size + mp->elt_size + mp->trailer_size;
1111 rte_mempool_ops_get_info(mp, &info);
1113 for (i = 0; i < n; ++i) {
1114 void *first_obj = first_obj_table_const[i];
1116 for (j = 0; j < info.contig_block_size; ++j) {
1119 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1120 rte_mempool_check_cookies(mp, &obj, 1, free);
1125 RTE_SET_USED(first_obj_table_const);
1131 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1133 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1134 void *obj, __rte_unused unsigned idx)
1136 __mempool_check_cookies(mp, &obj, 1, 2);
1140 mempool_audit_cookies(struct rte_mempool *mp)
1144 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1145 if (num != mp->size) {
1146 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1147 "iterated only over %u elements\n",
1152 #define mempool_audit_cookies(mp) do {} while(0)
1155 #ifndef __INTEL_COMPILER
1156 #pragma GCC diagnostic error "-Wcast-qual"
1159 /* check cookies before and after objects */
1161 mempool_audit_cache(const struct rte_mempool *mp)
1163 /* check cache size consistency */
1166 if (mp->cache_size == 0)
1169 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1170 const struct rte_mempool_cache *cache;
1171 cache = &mp->local_cache[lcore_id];
1172 if (cache->len > cache->flushthresh) {
1173 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1175 rte_panic("MEMPOOL: invalid cache len\n");
1180 /* check the consistency of mempool (size, cookies, ...) */
1182 rte_mempool_audit(struct rte_mempool *mp)
1184 mempool_audit_cache(mp);
1185 mempool_audit_cookies(mp);
1187 /* For case where mempool DEBUG is not set, and cache size is 0 */
1191 /* dump the status of the mempool on the console */
1193 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1195 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1196 struct rte_mempool_info info;
1197 struct rte_mempool_debug_stats sum;
1200 struct rte_mempool_memhdr *memhdr;
1201 unsigned common_count;
1202 unsigned cache_count;
1205 RTE_ASSERT(f != NULL);
1206 RTE_ASSERT(mp != NULL);
1208 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1209 fprintf(f, " flags=%x\n", mp->flags);
1210 fprintf(f, " pool=%p\n", mp->pool_data);
1211 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1212 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1213 fprintf(f, " size=%"PRIu32"\n", mp->size);
1214 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1215 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1216 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1217 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1218 fprintf(f, " total_obj_size=%"PRIu32"\n",
1219 mp->header_size + mp->elt_size + mp->trailer_size);
1221 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1223 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1224 mem_len += memhdr->len;
1226 fprintf(f, " avg bytes/object=%#Lf\n",
1227 (long double)mem_len / mp->size);
1230 cache_count = rte_mempool_dump_cache(f, mp);
1231 common_count = rte_mempool_ops_get_count(mp);
1232 if ((cache_count + common_count) > mp->size)
1233 common_count = mp->size - cache_count;
1234 fprintf(f, " common_pool_count=%u\n", common_count);
1236 /* sum and dump statistics */
1237 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1238 rte_mempool_ops_get_info(mp, &info);
1239 memset(&sum, 0, sizeof(sum));
1240 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1241 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1242 sum.put_objs += mp->stats[lcore_id].put_objs;
1243 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1244 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1245 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1246 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1247 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1248 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1250 fprintf(f, " stats:\n");
1251 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1252 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1253 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1254 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1255 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1256 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1257 if (info.contig_block_size > 0) {
1258 fprintf(f, " get_success_blks=%"PRIu64"\n",
1259 sum.get_success_blks);
1260 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1263 fprintf(f, " no statistics available\n");
1266 rte_mempool_audit(mp);
1269 /* dump the status of all mempools on the console */
1271 rte_mempool_list_dump(FILE *f)
1273 struct rte_mempool *mp = NULL;
1274 struct rte_tailq_entry *te;
1275 struct rte_mempool_list *mempool_list;
1277 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1279 rte_mcfg_mempool_read_lock();
1281 TAILQ_FOREACH(te, mempool_list, next) {
1282 mp = (struct rte_mempool *) te->data;
1283 rte_mempool_dump(f, mp);
1286 rte_mcfg_mempool_read_unlock();
1289 /* search a mempool from its name */
1290 struct rte_mempool *
1291 rte_mempool_lookup(const char *name)
1293 struct rte_mempool *mp = NULL;
1294 struct rte_tailq_entry *te;
1295 struct rte_mempool_list *mempool_list;
1297 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1299 rte_mcfg_mempool_read_lock();
1301 TAILQ_FOREACH(te, mempool_list, next) {
1302 mp = (struct rte_mempool *) te->data;
1303 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1307 rte_mcfg_mempool_read_unlock();
1317 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1320 struct rte_tailq_entry *te = NULL;
1321 struct rte_mempool_list *mempool_list;
1324 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1326 rte_mcfg_mempool_read_lock();
1328 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1329 (*func)((struct rte_mempool *) te->data, arg);
1332 rte_mcfg_mempool_read_unlock();