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>
16 #include <rte_common.h>
18 #include <rte_debug.h>
19 #include <rte_memory.h>
20 #include <rte_memzone.h>
21 #include <rte_malloc.h>
22 #include <rte_atomic.h>
23 #include <rte_launch.h>
25 #include <rte_eal_memconfig.h>
26 #include <rte_per_lcore.h>
27 #include <rte_lcore.h>
28 #include <rte_branch_prediction.h>
29 #include <rte_errno.h>
30 #include <rte_string_fns.h>
31 #include <rte_spinlock.h>
32 #include <rte_tailq.h>
33 #include <rte_eal_paging.h>
35 #include "rte_mempool.h"
36 #include "rte_mempool_trace.h"
38 TAILQ_HEAD(rte_mempool_list, rte_tailq_entry);
40 static struct rte_tailq_elem rte_mempool_tailq = {
41 .name = "RTE_MEMPOOL",
43 EAL_REGISTER_TAILQ(rte_mempool_tailq)
45 #define CACHE_FLUSHTHRESH_MULTIPLIER 1.5
46 #define CALC_CACHE_FLUSHTHRESH(c) \
47 ((typeof(c))((c) * CACHE_FLUSHTHRESH_MULTIPLIER))
49 #if defined(RTE_ARCH_X86)
51 * return the greatest common divisor between a and b (fast algorithm)
54 static unsigned get_gcd(unsigned a, unsigned b)
79 * Depending on memory configuration on x86 arch, objects addresses are spread
80 * between channels and ranks in RAM: the pool allocator will add
81 * padding between objects. This function return the new size of the
85 arch_mem_object_align(unsigned int obj_size)
87 unsigned nrank, nchan;
88 unsigned new_obj_size;
90 /* get number of channels */
91 nchan = rte_memory_get_nchannel();
95 nrank = rte_memory_get_nrank();
99 /* process new object size */
100 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
101 while (get_gcd(new_obj_size, nrank * nchan) != 1)
103 return new_obj_size * RTE_MEMPOOL_ALIGN;
107 arch_mem_object_align(unsigned int obj_size)
113 struct pagesz_walk_arg {
119 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
121 struct pagesz_walk_arg *wa = arg;
125 * we need to only look at page sizes available for a particular socket
126 * ID. so, we either need an exact match on socket ID (can match both
127 * native and external memory), or, if SOCKET_ID_ANY was specified as a
128 * socket ID argument, we must only look at native memory and ignore any
129 * page sizes associated with external memory.
131 valid = msl->socket_id == wa->socket_id;
132 valid |= wa->socket_id == SOCKET_ID_ANY && msl->external == 0;
134 if (valid && msl->page_sz < wa->min)
135 wa->min = msl->page_sz;
141 get_min_page_size(int socket_id)
143 struct pagesz_walk_arg wa;
146 wa.socket_id = socket_id;
148 rte_memseg_list_walk(find_min_pagesz, &wa);
150 return wa.min == SIZE_MAX ? (size_t) rte_mem_page_size() : wa.min;
155 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
156 void *obj, rte_iova_t iova)
158 struct rte_mempool_objhdr *hdr;
159 struct rte_mempool_objtlr *tlr __rte_unused;
161 /* set mempool ptr in header */
162 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
165 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
166 mp->populated_size++;
168 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
169 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
170 tlr = __mempool_get_trailer(obj);
171 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
175 /* call obj_cb() for each mempool element */
177 rte_mempool_obj_iter(struct rte_mempool *mp,
178 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
180 struct rte_mempool_objhdr *hdr;
184 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
185 obj = (char *)hdr + sizeof(*hdr);
186 obj_cb(mp, obj_cb_arg, obj, n);
193 /* call mem_cb() for each mempool memory chunk */
195 rte_mempool_mem_iter(struct rte_mempool *mp,
196 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
198 struct rte_mempool_memhdr *hdr;
201 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
202 mem_cb(mp, mem_cb_arg, hdr, n);
209 /* get the header, trailer and total size of a mempool element. */
211 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
212 struct rte_mempool_objsz *sz)
214 struct rte_mempool_objsz lsz;
216 sz = (sz != NULL) ? sz : &lsz;
218 sz->header_size = sizeof(struct rte_mempool_objhdr);
219 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
220 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
223 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
224 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
226 sz->trailer_size = 0;
229 /* element size is 8 bytes-aligned at least */
230 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
232 /* expand trailer to next cache line */
233 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
234 sz->total_size = sz->header_size + sz->elt_size +
236 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
237 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
238 RTE_MEMPOOL_ALIGN_MASK);
242 * increase trailer to add padding between objects in order to
243 * spread them across memory channels/ranks
245 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
247 new_size = arch_mem_object_align
248 (sz->header_size + sz->elt_size + sz->trailer_size);
249 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
252 /* this is the size of an object, including header and trailer */
253 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
255 return sz->total_size;
258 /* free a memchunk allocated with rte_memzone_reserve() */
260 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
263 const struct rte_memzone *mz = opaque;
264 rte_memzone_free(mz);
267 /* Free memory chunks used by a mempool. Objects must be in pool */
269 rte_mempool_free_memchunks(struct rte_mempool *mp)
271 struct rte_mempool_memhdr *memhdr;
274 while (!STAILQ_EMPTY(&mp->elt_list)) {
275 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
277 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
278 mp->populated_size--;
281 while (!STAILQ_EMPTY(&mp->mem_list)) {
282 memhdr = STAILQ_FIRST(&mp->mem_list);
283 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
284 if (memhdr->free_cb != NULL)
285 memhdr->free_cb(memhdr, memhdr->opaque);
292 mempool_ops_alloc_once(struct rte_mempool *mp)
296 /* create the internal ring if not already done */
297 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
298 ret = rte_mempool_ops_alloc(mp);
301 mp->flags |= MEMPOOL_F_POOL_CREATED;
306 /* Add objects in the pool, using a physically contiguous memory
307 * zone. Return the number of objects added, or a negative value
311 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
312 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
317 struct rte_mempool_memhdr *memhdr;
320 ret = mempool_ops_alloc_once(mp);
324 /* mempool is already populated */
325 if (mp->populated_size >= mp->size)
328 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
333 memhdr->addr = vaddr;
336 memhdr->free_cb = free_cb;
337 memhdr->opaque = opaque;
339 if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
340 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
342 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_MEMPOOL_ALIGN) - vaddr;
349 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
351 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
352 len - off, mempool_add_elem, NULL);
354 /* not enough room to store one object */
360 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
363 rte_mempool_trace_populate_iova(mp, vaddr, iova, len, free_cb, opaque);
374 struct rte_memseg *ms;
376 /* try registered memory first */
377 ms = rte_mem_virt2memseg(addr, NULL);
378 if (ms == NULL || ms->iova == RTE_BAD_IOVA)
379 /* fall back to actual physical address */
380 return rte_mem_virt2iova(addr);
381 return ms->iova + RTE_PTR_DIFF(addr, ms->addr);
384 /* Populate the mempool with a virtual area. Return the number of
385 * objects added, or a negative value on error.
388 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
389 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
393 size_t off, phys_len;
396 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
397 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
398 len, free_cb, opaque);
400 for (off = 0; off < len &&
401 mp->populated_size < mp->size; off += phys_len) {
403 iova = get_iova(addr + off);
405 /* populate with the largest group of contiguous pages */
406 for (phys_len = RTE_MIN(
407 (size_t)(RTE_PTR_ALIGN_CEIL(addr + off + 1, pg_sz) -
410 off + phys_len < len;
411 phys_len = RTE_MIN(phys_len + pg_sz, len - off)) {
414 iova_tmp = get_iova(addr + off + phys_len);
416 if (iova_tmp == RTE_BAD_IOVA ||
417 iova_tmp != iova + phys_len)
421 ret = rte_mempool_populate_iova(mp, addr + off, iova,
422 phys_len, free_cb, opaque);
427 /* no need to call the free callback for next chunks */
432 rte_mempool_trace_populate_virt(mp, addr, len, pg_sz, free_cb, opaque);
436 rte_mempool_free_memchunks(mp);
440 /* Get the minimal page size used in a mempool before populating it. */
442 rte_mempool_get_page_size(struct rte_mempool *mp, size_t *pg_sz)
444 bool need_iova_contig_obj;
445 bool alloc_in_ext_mem;
448 /* check if we can retrieve a valid socket ID */
449 ret = rte_malloc_heap_socket_is_external(mp->socket_id);
452 alloc_in_ext_mem = (ret == 1);
453 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
455 if (!need_iova_contig_obj)
457 else if (rte_eal_has_hugepages() || alloc_in_ext_mem)
458 *pg_sz = get_min_page_size(mp->socket_id);
460 *pg_sz = rte_mem_page_size();
462 rte_mempool_trace_get_page_size(mp, *pg_sz);
466 /* Default function to populate the mempool: allocate memory in memzones,
467 * and populate them. Return the number of objects added, or a negative
471 rte_mempool_populate_default(struct rte_mempool *mp)
473 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
474 char mz_name[RTE_MEMZONE_NAMESIZE];
475 const struct rte_memzone *mz;
477 size_t align, pg_sz, pg_shift = 0;
481 bool need_iova_contig_obj;
482 size_t max_alloc_size = SIZE_MAX;
484 ret = mempool_ops_alloc_once(mp);
488 /* mempool must not be populated */
489 if (mp->nb_mem_chunks != 0)
493 * the following section calculates page shift and page size values.
495 * these values impact the result of calc_mem_size operation, which
496 * returns the amount of memory that should be allocated to store the
497 * desired number of objects. when not zero, it allocates more memory
498 * for the padding between objects, to ensure that an object does not
499 * cross a page boundary. in other words, page size/shift are to be set
500 * to zero if mempool elements won't care about page boundaries.
501 * there are several considerations for page size and page shift here.
503 * if we don't need our mempools to have physically contiguous objects,
504 * then just set page shift and page size to 0, because the user has
505 * indicated that there's no need to care about anything.
507 * if we do need contiguous objects (if a mempool driver has its
508 * own calc_size() method returning min_chunk_size = mem_size),
509 * there is also an option to reserve the entire mempool memory
510 * as one contiguous block of memory.
512 * if we require contiguous objects, but not necessarily the entire
513 * mempool reserved space to be contiguous, pg_sz will be != 0,
514 * and the default ops->populate() will take care of not placing
515 * objects across pages.
517 * if our IO addresses are physical, we may get memory from bigger
518 * pages, or we might get memory from smaller pages, and how much of it
519 * we require depends on whether we want bigger or smaller pages.
520 * However, requesting each and every memory size is too much work, so
521 * what we'll do instead is walk through the page sizes available, pick
522 * the smallest one and set up page shift to match that one. We will be
523 * wasting some space this way, but it's much nicer than looping around
524 * trying to reserve each and every page size.
526 * If we fail to get enough contiguous memory, then we'll go and
527 * reserve space in smaller chunks.
530 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
531 ret = rte_mempool_get_page_size(mp, &pg_sz);
536 pg_shift = rte_bsf32(pg_sz);
538 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
539 size_t min_chunk_size;
541 mem_size = rte_mempool_ops_calc_mem_size(
542 mp, n, pg_shift, &min_chunk_size, &align);
549 ret = snprintf(mz_name, sizeof(mz_name),
550 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
551 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
556 /* if we're trying to reserve contiguous memory, add appropriate
559 if (min_chunk_size == (size_t)mem_size)
560 mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
562 /* Allocate a memzone, retrying with a smaller area on ENOMEM */
564 mz = rte_memzone_reserve_aligned(mz_name,
565 RTE_MIN((size_t)mem_size, max_alloc_size),
566 mp->socket_id, mz_flags, align);
568 if (mz != NULL || rte_errno != ENOMEM)
571 max_alloc_size = RTE_MIN(max_alloc_size,
572 (size_t)mem_size) / 2;
573 } while (mz == NULL && max_alloc_size >= min_chunk_size);
580 if (need_iova_contig_obj)
585 if (pg_sz == 0 || (mz_flags & RTE_MEMZONE_IOVA_CONTIG))
586 ret = rte_mempool_populate_iova(mp, mz->addr,
588 rte_mempool_memchunk_mz_free,
589 (void *)(uintptr_t)mz);
591 ret = rte_mempool_populate_virt(mp, mz->addr,
593 rte_mempool_memchunk_mz_free,
594 (void *)(uintptr_t)mz);
595 if (ret == 0) /* should not happen */
598 rte_memzone_free(mz);
603 rte_mempool_trace_populate_default(mp);
607 rte_mempool_free_memchunks(mp);
611 /* return the memory size required for mempool objects in anonymous mem */
613 get_anon_size(const struct rte_mempool *mp)
616 size_t pg_sz, pg_shift;
617 size_t min_chunk_size;
620 pg_sz = rte_mem_page_size();
621 pg_shift = rte_bsf32(pg_sz);
622 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
623 &min_chunk_size, &align);
628 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
630 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
636 * Calculate size since memhdr->len has contiguous chunk length
637 * which may be smaller if anon map is split into many contiguous
638 * chunks. Result must be the same as we calculated on populate.
640 size = get_anon_size(memhdr->mp);
644 rte_mem_unmap(opaque, size);
647 /* populate the mempool with an anonymous mapping */
649 rte_mempool_populate_anon(struct rte_mempool *mp)
655 /* mempool is already populated, error */
656 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
661 ret = mempool_ops_alloc_once(mp);
667 size = get_anon_size(mp);
673 /* get chunk of virtually continuous memory */
674 addr = rte_mem_map(NULL, size, RTE_PROT_READ | RTE_PROT_WRITE,
675 RTE_MAP_SHARED | RTE_MAP_ANONYMOUS, -1, 0);
678 /* can't use MMAP_LOCKED, it does not exist on BSD */
679 if (rte_mem_lock(addr, size) < 0) {
680 rte_mem_unmap(addr, size);
684 ret = rte_mempool_populate_virt(mp, addr, size, rte_mem_page_size(),
685 rte_mempool_memchunk_anon_free, addr);
686 if (ret == 0) /* should not happen */
693 rte_mempool_trace_populate_anon(mp);
694 return mp->populated_size;
697 rte_mempool_free_memchunks(mp);
703 rte_mempool_free(struct rte_mempool *mp)
705 struct rte_mempool_list *mempool_list = NULL;
706 struct rte_tailq_entry *te;
711 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
712 rte_mcfg_tailq_write_lock();
713 /* find out tailq entry */
714 TAILQ_FOREACH(te, mempool_list, next) {
715 if (te->data == (void *)mp)
720 TAILQ_REMOVE(mempool_list, te, next);
723 rte_mcfg_tailq_write_unlock();
725 rte_mempool_trace_free(mp);
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);
764 rte_mempool_trace_cache_create(size, socket_id, cache);
769 * Free a cache. It's the responsibility of the user to make sure that any
770 * remaining objects in the cache are flushed to the corresponding
774 rte_mempool_cache_free(struct rte_mempool_cache *cache)
776 rte_mempool_trace_cache_free(cache);
780 /* create an empty mempool */
782 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
783 unsigned cache_size, unsigned private_data_size,
784 int socket_id, unsigned flags)
786 char mz_name[RTE_MEMZONE_NAMESIZE];
787 struct rte_mempool_list *mempool_list;
788 struct rte_mempool *mp = NULL;
789 struct rte_tailq_entry *te = NULL;
790 const struct rte_memzone *mz = NULL;
792 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
793 struct rte_mempool_objsz objsz;
797 /* compilation-time checks */
798 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
799 RTE_CACHE_LINE_MASK) != 0);
800 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
801 RTE_CACHE_LINE_MASK) != 0);
802 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
803 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
804 RTE_CACHE_LINE_MASK) != 0);
805 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
806 RTE_CACHE_LINE_MASK) != 0);
809 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
811 /* asked for zero items */
817 /* asked cache too big */
818 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
819 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
824 /* "no cache align" imply "no spread" */
825 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
826 flags |= MEMPOOL_F_NO_SPREAD;
828 /* calculate mempool object sizes. */
829 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
834 rte_mcfg_mempool_write_lock();
837 * reserve a memory zone for this mempool: private data is
840 private_data_size = (private_data_size +
841 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
844 /* try to allocate tailq entry */
845 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
847 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
851 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
852 mempool_size += private_data_size;
853 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
855 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
856 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
857 rte_errno = ENAMETOOLONG;
861 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
865 /* init the mempool structure */
867 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
868 ret = strlcpy(mp->name, name, sizeof(mp->name));
869 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
870 rte_errno = ENAMETOOLONG;
876 mp->socket_id = socket_id;
877 mp->elt_size = objsz.elt_size;
878 mp->header_size = objsz.header_size;
879 mp->trailer_size = objsz.trailer_size;
880 /* Size of default caches, zero means disabled. */
881 mp->cache_size = cache_size;
882 mp->private_data_size = private_data_size;
883 STAILQ_INIT(&mp->elt_list);
884 STAILQ_INIT(&mp->mem_list);
887 * local_cache pointer is set even if cache_size is zero.
888 * The local_cache points to just past the elt_pa[] array.
890 mp->local_cache = (struct rte_mempool_cache *)
891 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
893 /* Init all default caches. */
894 if (cache_size != 0) {
895 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
896 mempool_cache_init(&mp->local_cache[lcore_id],
902 rte_mcfg_tailq_write_lock();
903 TAILQ_INSERT_TAIL(mempool_list, te, next);
904 rte_mcfg_tailq_write_unlock();
905 rte_mcfg_mempool_write_unlock();
907 rte_mempool_trace_create_empty(name, n, elt_size, cache_size,
908 private_data_size, flags, mp);
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);
961 rte_mempool_trace_create(name, n, elt_size, cache_size,
962 private_data_size, mp_init, mp_init_arg, obj_init,
963 obj_init_arg, flags, mp);
967 rte_mempool_free(mp);
971 /* Return the number of entries in the mempool */
973 rte_mempool_avail_count(const struct rte_mempool *mp)
978 count = rte_mempool_ops_get_count(mp);
980 if (mp->cache_size == 0)
983 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
984 count += mp->local_cache[lcore_id].len;
987 * due to race condition (access to len is not locked), the
988 * total can be greater than size... so fix the result
990 if (count > mp->size)
995 /* return the number of entries allocated from the mempool */
997 rte_mempool_in_use_count(const struct rte_mempool *mp)
999 return mp->size - rte_mempool_avail_count(mp);
1002 /* dump the cache status */
1004 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1008 unsigned cache_count;
1010 fprintf(f, " internal cache infos:\n");
1011 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1013 if (mp->cache_size == 0)
1016 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1017 cache_count = mp->local_cache[lcore_id].len;
1018 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1019 lcore_id, cache_count);
1020 count += cache_count;
1022 fprintf(f, " total_cache_count=%u\n", count);
1026 #ifndef __INTEL_COMPILER
1027 #pragma GCC diagnostic ignored "-Wcast-qual"
1030 /* check and update cookies or panic (internal) */
1031 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1032 void * const *obj_table_const, unsigned n, int free)
1034 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1035 struct rte_mempool_objhdr *hdr;
1036 struct rte_mempool_objtlr *tlr;
1042 /* Force to drop the "const" attribute. This is done only when
1043 * DEBUG is enabled */
1044 tmp = (void *) obj_table_const;
1050 if (rte_mempool_from_obj(obj) != mp)
1051 rte_panic("MEMPOOL: object is owned by another "
1054 hdr = __mempool_get_header(obj);
1055 cookie = hdr->cookie;
1058 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1059 RTE_LOG(CRIT, MEMPOOL,
1060 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1061 obj, (const void *) mp, cookie);
1062 rte_panic("MEMPOOL: bad header cookie (put)\n");
1064 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1065 } else if (free == 1) {
1066 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1067 RTE_LOG(CRIT, MEMPOOL,
1068 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1069 obj, (const void *) mp, cookie);
1070 rte_panic("MEMPOOL: bad header cookie (get)\n");
1072 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1073 } else if (free == 2) {
1074 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1075 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1076 RTE_LOG(CRIT, MEMPOOL,
1077 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1078 obj, (const void *) mp, cookie);
1079 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1082 tlr = __mempool_get_trailer(obj);
1083 cookie = tlr->cookie;
1084 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1085 RTE_LOG(CRIT, MEMPOOL,
1086 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1087 obj, (const void *) mp, cookie);
1088 rte_panic("MEMPOOL: bad trailer cookie\n");
1093 RTE_SET_USED(obj_table_const);
1100 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1101 void * const *first_obj_table_const, unsigned int n, int free)
1103 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1104 struct rte_mempool_info info;
1105 const size_t total_elt_sz =
1106 mp->header_size + mp->elt_size + mp->trailer_size;
1109 rte_mempool_ops_get_info(mp, &info);
1111 for (i = 0; i < n; ++i) {
1112 void *first_obj = first_obj_table_const[i];
1114 for (j = 0; j < info.contig_block_size; ++j) {
1117 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1118 rte_mempool_check_cookies(mp, &obj, 1, free);
1123 RTE_SET_USED(first_obj_table_const);
1129 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1131 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1132 void *obj, __rte_unused unsigned idx)
1134 __mempool_check_cookies(mp, &obj, 1, 2);
1138 mempool_audit_cookies(struct rte_mempool *mp)
1142 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1143 if (num != mp->size) {
1144 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1145 "iterated only over %u elements\n",
1150 #define mempool_audit_cookies(mp) do {} while(0)
1153 #ifndef __INTEL_COMPILER
1154 #pragma GCC diagnostic error "-Wcast-qual"
1157 /* check cookies before and after objects */
1159 mempool_audit_cache(const struct rte_mempool *mp)
1161 /* check cache size consistency */
1164 if (mp->cache_size == 0)
1167 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1168 const struct rte_mempool_cache *cache;
1169 cache = &mp->local_cache[lcore_id];
1170 if (cache->len > RTE_DIM(cache->objs)) {
1171 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1173 rte_panic("MEMPOOL: invalid cache len\n");
1178 /* check the consistency of mempool (size, cookies, ...) */
1180 rte_mempool_audit(struct rte_mempool *mp)
1182 mempool_audit_cache(mp);
1183 mempool_audit_cookies(mp);
1185 /* For case where mempool DEBUG is not set, and cache size is 0 */
1189 /* dump the status of the mempool on the console */
1191 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1193 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1194 struct rte_mempool_info info;
1195 struct rte_mempool_debug_stats sum;
1198 struct rte_mempool_memhdr *memhdr;
1199 struct rte_mempool_ops *ops;
1200 unsigned common_count;
1201 unsigned cache_count;
1204 RTE_ASSERT(f != NULL);
1205 RTE_ASSERT(mp != NULL);
1207 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1208 fprintf(f, " flags=%x\n", mp->flags);
1209 fprintf(f, " socket_id=%d\n", mp->socket_id);
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 fprintf(f, " ops_index=%d\n", mp->ops_index);
1224 ops = rte_mempool_get_ops(mp->ops_index);
1225 fprintf(f, " ops_name: <%s>\n", (ops != NULL) ? ops->name : "NA");
1227 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1228 mem_len += memhdr->len;
1230 fprintf(f, " avg bytes/object=%#Lf\n",
1231 (long double)mem_len / mp->size);
1234 cache_count = rte_mempool_dump_cache(f, mp);
1235 common_count = rte_mempool_ops_get_count(mp);
1236 if ((cache_count + common_count) > mp->size)
1237 common_count = mp->size - cache_count;
1238 fprintf(f, " common_pool_count=%u\n", common_count);
1240 /* sum and dump statistics */
1241 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1242 rte_mempool_ops_get_info(mp, &info);
1243 memset(&sum, 0, sizeof(sum));
1244 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1245 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1246 sum.put_objs += mp->stats[lcore_id].put_objs;
1247 sum.put_common_pool_bulk += mp->stats[lcore_id].put_common_pool_bulk;
1248 sum.put_common_pool_objs += mp->stats[lcore_id].put_common_pool_objs;
1249 sum.get_common_pool_bulk += mp->stats[lcore_id].get_common_pool_bulk;
1250 sum.get_common_pool_objs += mp->stats[lcore_id].get_common_pool_objs;
1251 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1252 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1253 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1254 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1255 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1256 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1258 fprintf(f, " stats:\n");
1259 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1260 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1261 fprintf(f, " put_common_pool_bulk=%"PRIu64"\n", sum.put_common_pool_bulk);
1262 fprintf(f, " put_common_pool_objs=%"PRIu64"\n", sum.put_common_pool_objs);
1263 fprintf(f, " get_common_pool_bulk=%"PRIu64"\n", sum.get_common_pool_bulk);
1264 fprintf(f, " get_common_pool_objs=%"PRIu64"\n", sum.get_common_pool_objs);
1265 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1266 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1267 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1268 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1269 if (info.contig_block_size > 0) {
1270 fprintf(f, " get_success_blks=%"PRIu64"\n",
1271 sum.get_success_blks);
1272 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1275 fprintf(f, " no statistics available\n");
1278 rte_mempool_audit(mp);
1281 /* dump the status of all mempools on the console */
1283 rte_mempool_list_dump(FILE *f)
1285 struct rte_mempool *mp = NULL;
1286 struct rte_tailq_entry *te;
1287 struct rte_mempool_list *mempool_list;
1289 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1291 rte_mcfg_mempool_read_lock();
1293 TAILQ_FOREACH(te, mempool_list, next) {
1294 mp = (struct rte_mempool *) te->data;
1295 rte_mempool_dump(f, mp);
1298 rte_mcfg_mempool_read_unlock();
1301 /* search a mempool from its name */
1302 struct rte_mempool *
1303 rte_mempool_lookup(const char *name)
1305 struct rte_mempool *mp = NULL;
1306 struct rte_tailq_entry *te;
1307 struct rte_mempool_list *mempool_list;
1309 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1311 rte_mcfg_mempool_read_lock();
1313 TAILQ_FOREACH(te, mempool_list, next) {
1314 mp = (struct rte_mempool *) te->data;
1315 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1319 rte_mcfg_mempool_read_unlock();
1329 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1332 struct rte_tailq_entry *te = NULL;
1333 struct rte_mempool_list *mempool_list;
1336 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1338 rte_mcfg_mempool_read_lock();
1340 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1341 (*func)((struct rte_mempool *) te->data, arg);
1344 rte_mcfg_mempool_read_unlock();