1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation.
3 * Copyright(c) 2016 6WIND S.A.
13 #include <sys/queue.h>
15 #include <rte_common.h>
17 #include <rte_debug.h>
18 #include <rte_memory.h>
19 #include <rte_memzone.h>
20 #include <rte_malloc.h>
22 #include <rte_eal_memconfig.h>
23 #include <rte_errno.h>
24 #include <rte_string_fns.h>
25 #include <rte_tailq.h>
26 #include <rte_eal_paging.h>
27 #include <rte_telemetry.h>
29 #include "rte_mempool.h"
30 #include "rte_mempool_trace.h"
32 TAILQ_HEAD(rte_mempool_list, rte_tailq_entry);
34 static struct rte_tailq_elem rte_mempool_tailq = {
35 .name = "RTE_MEMPOOL",
37 EAL_REGISTER_TAILQ(rte_mempool_tailq)
39 TAILQ_HEAD(mempool_callback_list, rte_tailq_entry);
41 static struct rte_tailq_elem callback_tailq = {
42 .name = "RTE_MEMPOOL_CALLBACK",
44 EAL_REGISTER_TAILQ(callback_tailq)
46 /* Invoke all registered mempool event callbacks. */
48 mempool_event_callback_invoke(enum rte_mempool_event event,
49 struct rte_mempool *mp);
51 #define CACHE_FLUSHTHRESH_MULTIPLIER 1.5
52 #define CALC_CACHE_FLUSHTHRESH(c) \
53 ((typeof(c))((c) * CACHE_FLUSHTHRESH_MULTIPLIER))
55 #if defined(RTE_ARCH_X86)
57 * return the greatest common divisor between a and b (fast algorithm)
60 static unsigned get_gcd(unsigned a, unsigned b)
85 * Depending on memory configuration on x86 arch, objects addresses are spread
86 * between channels and ranks in RAM: the pool allocator will add
87 * padding between objects. This function return the new size of the
91 arch_mem_object_align(unsigned int obj_size)
93 unsigned nrank, nchan;
94 unsigned new_obj_size;
96 /* get number of channels */
97 nchan = rte_memory_get_nchannel();
101 nrank = rte_memory_get_nrank();
105 /* process new object size */
106 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
107 while (get_gcd(new_obj_size, nrank * nchan) != 1)
109 return new_obj_size * RTE_MEMPOOL_ALIGN;
113 arch_mem_object_align(unsigned int obj_size)
119 struct pagesz_walk_arg {
125 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
127 struct pagesz_walk_arg *wa = arg;
131 * we need to only look at page sizes available for a particular socket
132 * ID. so, we either need an exact match on socket ID (can match both
133 * native and external memory), or, if SOCKET_ID_ANY was specified as a
134 * socket ID argument, we must only look at native memory and ignore any
135 * page sizes associated with external memory.
137 valid = msl->socket_id == wa->socket_id;
138 valid |= wa->socket_id == SOCKET_ID_ANY && msl->external == 0;
140 if (valid && msl->page_sz < wa->min)
141 wa->min = msl->page_sz;
147 get_min_page_size(int socket_id)
149 struct pagesz_walk_arg wa;
152 wa.socket_id = socket_id;
154 rte_memseg_list_walk(find_min_pagesz, &wa);
156 return wa.min == SIZE_MAX ? (size_t) rte_mem_page_size() : wa.min;
161 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
162 void *obj, rte_iova_t iova)
164 struct rte_mempool_objhdr *hdr;
165 struct rte_mempool_objtlr *tlr __rte_unused;
167 /* set mempool ptr in header */
168 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
171 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
172 mp->populated_size++;
174 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
175 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
176 tlr = rte_mempool_get_trailer(obj);
177 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
181 /* call obj_cb() for each mempool element */
183 rte_mempool_obj_iter(struct rte_mempool *mp,
184 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
186 struct rte_mempool_objhdr *hdr;
190 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
191 obj = (char *)hdr + sizeof(*hdr);
192 obj_cb(mp, obj_cb_arg, obj, n);
199 /* call mem_cb() for each mempool memory chunk */
201 rte_mempool_mem_iter(struct rte_mempool *mp,
202 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
204 struct rte_mempool_memhdr *hdr;
207 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
208 mem_cb(mp, mem_cb_arg, hdr, n);
215 /* get the header, trailer and total size of a mempool element. */
217 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
218 struct rte_mempool_objsz *sz)
220 struct rte_mempool_objsz lsz;
222 sz = (sz != NULL) ? sz : &lsz;
224 sz->header_size = sizeof(struct rte_mempool_objhdr);
225 if ((flags & RTE_MEMPOOL_F_NO_CACHE_ALIGN) == 0)
226 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
229 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
230 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
232 sz->trailer_size = 0;
235 /* element size is 8 bytes-aligned at least */
236 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
238 /* expand trailer to next cache line */
239 if ((flags & RTE_MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
240 sz->total_size = sz->header_size + sz->elt_size +
242 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
243 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
244 RTE_MEMPOOL_ALIGN_MASK);
248 * increase trailer to add padding between objects in order to
249 * spread them across memory channels/ranks
251 if ((flags & RTE_MEMPOOL_F_NO_SPREAD) == 0) {
253 new_size = arch_mem_object_align
254 (sz->header_size + sz->elt_size + sz->trailer_size);
255 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
258 /* this is the size of an object, including header and trailer */
259 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
261 return sz->total_size;
264 /* free a memchunk allocated with rte_memzone_reserve() */
266 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
269 const struct rte_memzone *mz = opaque;
270 rte_memzone_free(mz);
273 /* Free memory chunks used by a mempool. Objects must be in pool */
275 rte_mempool_free_memchunks(struct rte_mempool *mp)
277 struct rte_mempool_memhdr *memhdr;
280 while (!STAILQ_EMPTY(&mp->elt_list)) {
281 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
283 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
284 mp->populated_size--;
287 while (!STAILQ_EMPTY(&mp->mem_list)) {
288 memhdr = STAILQ_FIRST(&mp->mem_list);
289 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
290 if (memhdr->free_cb != NULL)
291 memhdr->free_cb(memhdr, memhdr->opaque);
298 mempool_ops_alloc_once(struct rte_mempool *mp)
302 /* create the internal ring if not already done */
303 if ((mp->flags & RTE_MEMPOOL_F_POOL_CREATED) == 0) {
304 ret = rte_mempool_ops_alloc(mp);
307 mp->flags |= RTE_MEMPOOL_F_POOL_CREATED;
312 /* Add objects in the pool, using a physically contiguous memory
313 * zone. Return the number of objects added, or a negative value
317 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
318 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
323 struct rte_mempool_memhdr *memhdr;
326 ret = mempool_ops_alloc_once(mp);
330 /* mempool is already populated */
331 if (mp->populated_size >= mp->size)
334 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
339 memhdr->addr = vaddr;
342 memhdr->free_cb = free_cb;
343 memhdr->opaque = opaque;
345 if (mp->flags & RTE_MEMPOOL_F_NO_CACHE_ALIGN)
346 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
348 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_MEMPOOL_ALIGN) - vaddr;
355 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
357 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
358 len - off, mempool_add_elem, NULL);
360 /* not enough room to store one object */
366 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
369 /* Check if at least some objects in the pool are now usable for IO. */
370 if (!(mp->flags & RTE_MEMPOOL_F_NO_IOVA_CONTIG) && iova != RTE_BAD_IOVA)
371 mp->flags &= ~RTE_MEMPOOL_F_NON_IO;
373 /* Report the mempool as ready only when fully populated. */
374 if (mp->populated_size >= mp->size)
375 mempool_event_callback_invoke(RTE_MEMPOOL_EVENT_READY, mp);
377 rte_mempool_trace_populate_iova(mp, vaddr, iova, len, free_cb, opaque);
388 struct rte_memseg *ms;
390 /* try registered memory first */
391 ms = rte_mem_virt2memseg(addr, NULL);
392 if (ms == NULL || ms->iova == RTE_BAD_IOVA)
393 /* fall back to actual physical address */
394 return rte_mem_virt2iova(addr);
395 return ms->iova + RTE_PTR_DIFF(addr, ms->addr);
398 /* Populate the mempool with a virtual area. Return the number of
399 * objects added, or a negative value on error.
402 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
403 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
407 size_t off, phys_len;
410 if (mp->flags & RTE_MEMPOOL_F_NO_IOVA_CONTIG)
411 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
412 len, free_cb, opaque);
414 for (off = 0; off < len &&
415 mp->populated_size < mp->size; off += phys_len) {
417 iova = get_iova(addr + off);
419 /* populate with the largest group of contiguous pages */
420 for (phys_len = RTE_MIN(
421 (size_t)(RTE_PTR_ALIGN_CEIL(addr + off + 1, pg_sz) -
424 off + phys_len < len;
425 phys_len = RTE_MIN(phys_len + pg_sz, len - off)) {
428 iova_tmp = get_iova(addr + off + phys_len);
430 if (iova_tmp == RTE_BAD_IOVA ||
431 iova_tmp != iova + phys_len)
435 ret = rte_mempool_populate_iova(mp, addr + off, iova,
436 phys_len, free_cb, opaque);
441 /* no need to call the free callback for next chunks */
446 rte_mempool_trace_populate_virt(mp, addr, len, pg_sz, free_cb, opaque);
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 & RTE_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 = rte_mem_page_size();
476 rte_mempool_trace_get_page_size(mp, *pg_sz);
480 /* Default function to populate the mempool: allocate memory in memzones,
481 * and populate them. Return the number of objects added, or a negative
485 rte_mempool_populate_default(struct rte_mempool *mp)
487 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
488 char mz_name[RTE_MEMZONE_NAMESIZE];
489 const struct rte_memzone *mz;
491 size_t align, pg_sz, pg_shift = 0;
495 bool need_iova_contig_obj;
496 size_t max_alloc_size = SIZE_MAX;
498 ret = mempool_ops_alloc_once(mp);
502 /* mempool must not be populated */
503 if (mp->nb_mem_chunks != 0)
507 * the following section calculates page shift and page size values.
509 * these values impact the result of calc_mem_size operation, which
510 * returns the amount of memory that should be allocated to store the
511 * desired number of objects. when not zero, it allocates more memory
512 * for the padding between objects, to ensure that an object does not
513 * cross a page boundary. in other words, page size/shift are to be set
514 * to zero if mempool elements won't care about page boundaries.
515 * there are several considerations for page size and page shift here.
517 * if we don't need our mempools to have physically contiguous objects,
518 * then just set page shift and page size to 0, because the user has
519 * indicated that there's no need to care about anything.
521 * if we do need contiguous objects (if a mempool driver has its
522 * own calc_size() method returning min_chunk_size = mem_size),
523 * there is also an option to reserve the entire mempool memory
524 * as one contiguous block of memory.
526 * if we require contiguous objects, but not necessarily the entire
527 * mempool reserved space to be contiguous, pg_sz will be != 0,
528 * and the default ops->populate() will take care of not placing
529 * objects across pages.
531 * if our IO addresses are physical, we may get memory from bigger
532 * pages, or we might get memory from smaller pages, and how much of it
533 * we require depends on whether we want bigger or smaller pages.
534 * However, requesting each and every memory size is too much work, so
535 * what we'll do instead is walk through the page sizes available, pick
536 * the smallest one and set up page shift to match that one. We will be
537 * wasting some space this way, but it's much nicer than looping around
538 * trying to reserve each and every page size.
540 * If we fail to get enough contiguous memory, then we'll go and
541 * reserve space in smaller chunks.
544 need_iova_contig_obj = !(mp->flags & RTE_MEMPOOL_F_NO_IOVA_CONTIG);
545 ret = rte_mempool_get_page_size(mp, &pg_sz);
550 pg_shift = rte_bsf32(pg_sz);
552 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
553 size_t min_chunk_size;
555 mem_size = rte_mempool_ops_calc_mem_size(
556 mp, n, pg_shift, &min_chunk_size, &align);
563 ret = snprintf(mz_name, sizeof(mz_name),
564 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
565 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
570 /* if we're trying to reserve contiguous memory, add appropriate
573 if (min_chunk_size == (size_t)mem_size)
574 mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
576 /* Allocate a memzone, retrying with a smaller area on ENOMEM */
578 mz = rte_memzone_reserve_aligned(mz_name,
579 RTE_MIN((size_t)mem_size, max_alloc_size),
580 mp->socket_id, mz_flags, align);
582 if (mz != NULL || rte_errno != ENOMEM)
585 max_alloc_size = RTE_MIN(max_alloc_size,
586 (size_t)mem_size) / 2;
587 } while (mz == NULL && max_alloc_size >= min_chunk_size);
594 if (need_iova_contig_obj)
599 if (pg_sz == 0 || (mz_flags & RTE_MEMZONE_IOVA_CONTIG))
600 ret = rte_mempool_populate_iova(mp, mz->addr,
602 rte_mempool_memchunk_mz_free,
603 (void *)(uintptr_t)mz);
605 ret = rte_mempool_populate_virt(mp, mz->addr,
607 rte_mempool_memchunk_mz_free,
608 (void *)(uintptr_t)mz);
609 if (ret == 0) /* should not happen */
612 rte_memzone_free(mz);
617 rte_mempool_trace_populate_default(mp);
621 rte_mempool_free_memchunks(mp);
625 /* return the memory size required for mempool objects in anonymous mem */
627 get_anon_size(const struct rte_mempool *mp)
630 size_t pg_sz, pg_shift;
631 size_t min_chunk_size;
634 pg_sz = rte_mem_page_size();
635 pg_shift = rte_bsf32(pg_sz);
636 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
637 &min_chunk_size, &align);
642 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
644 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
650 * Calculate size since memhdr->len has contiguous chunk length
651 * which may be smaller if anon map is split into many contiguous
652 * chunks. Result must be the same as we calculated on populate.
654 size = get_anon_size(memhdr->mp);
658 rte_mem_unmap(opaque, size);
661 /* populate the mempool with an anonymous mapping */
663 rte_mempool_populate_anon(struct rte_mempool *mp)
669 /* mempool is already populated, error */
670 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
675 ret = mempool_ops_alloc_once(mp);
681 size = get_anon_size(mp);
687 /* get chunk of virtually continuous memory */
688 addr = rte_mem_map(NULL, size, RTE_PROT_READ | RTE_PROT_WRITE,
689 RTE_MAP_SHARED | RTE_MAP_ANONYMOUS, -1, 0);
692 /* can't use MMAP_LOCKED, it does not exist on BSD */
693 if (rte_mem_lock(addr, size) < 0) {
694 rte_mem_unmap(addr, size);
698 ret = rte_mempool_populate_virt(mp, addr, size, rte_mem_page_size(),
699 rte_mempool_memchunk_anon_free, addr);
700 if (ret == 0) /* should not happen */
707 rte_mempool_trace_populate_anon(mp);
708 return mp->populated_size;
711 rte_mempool_free_memchunks(mp);
717 rte_mempool_free(struct rte_mempool *mp)
719 struct rte_mempool_list *mempool_list = NULL;
720 struct rte_tailq_entry *te;
725 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
726 rte_mcfg_tailq_write_lock();
727 /* find out tailq entry */
728 TAILQ_FOREACH(te, mempool_list, next) {
729 if (te->data == (void *)mp)
734 TAILQ_REMOVE(mempool_list, te, next);
737 rte_mcfg_tailq_write_unlock();
739 mempool_event_callback_invoke(RTE_MEMPOOL_EVENT_DESTROY, mp);
740 rte_mempool_trace_free(mp);
741 rte_mempool_free_memchunks(mp);
742 rte_mempool_ops_free(mp);
743 rte_memzone_free(mp->mz);
747 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
750 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
755 * Create and initialize a cache for objects that are retrieved from and
756 * returned to an underlying mempool. This structure is identical to the
757 * local_cache[lcore_id] pointed to by the mempool structure.
759 struct rte_mempool_cache *
760 rte_mempool_cache_create(uint32_t size, int socket_id)
762 struct rte_mempool_cache *cache;
764 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
769 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
770 RTE_CACHE_LINE_SIZE, socket_id);
772 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
777 mempool_cache_init(cache, size);
779 rte_mempool_trace_cache_create(size, socket_id, cache);
784 * Free a cache. It's the responsibility of the user to make sure that any
785 * remaining objects in the cache are flushed to the corresponding
789 rte_mempool_cache_free(struct rte_mempool_cache *cache)
791 rte_mempool_trace_cache_free(cache);
795 /* create an empty mempool */
797 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
798 unsigned cache_size, unsigned private_data_size,
799 int socket_id, unsigned flags)
801 char mz_name[RTE_MEMZONE_NAMESIZE];
802 struct rte_mempool_list *mempool_list;
803 struct rte_mempool *mp = NULL;
804 struct rte_tailq_entry *te = NULL;
805 const struct rte_memzone *mz = NULL;
807 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
808 struct rte_mempool_objsz objsz;
812 /* compilation-time checks */
813 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
814 RTE_CACHE_LINE_MASK) != 0);
815 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
816 RTE_CACHE_LINE_MASK) != 0);
817 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
818 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
819 RTE_CACHE_LINE_MASK) != 0);
820 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
821 RTE_CACHE_LINE_MASK) != 0);
824 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
826 /* asked for zero items */
832 /* asked cache too big */
833 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
834 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
839 /* enforce only user flags are passed by the application */
840 if ((flags & ~RTE_MEMPOOL_VALID_USER_FLAGS) != 0) {
846 * No objects in the pool can be used for IO until it's populated
847 * with at least some objects with valid IOVA.
849 flags |= RTE_MEMPOOL_F_NON_IO;
851 /* "no cache align" imply "no spread" */
852 if (flags & RTE_MEMPOOL_F_NO_CACHE_ALIGN)
853 flags |= RTE_MEMPOOL_F_NO_SPREAD;
855 /* calculate mempool object sizes. */
856 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
861 rte_mcfg_mempool_write_lock();
864 * reserve a memory zone for this mempool: private data is
867 private_data_size = (private_data_size +
868 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
871 /* try to allocate tailq entry */
872 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
874 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
878 mempool_size = RTE_MEMPOOL_HEADER_SIZE(mp, cache_size);
879 mempool_size += private_data_size;
880 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
882 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
883 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
884 rte_errno = ENAMETOOLONG;
888 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
892 /* init the mempool structure */
894 memset(mp, 0, RTE_MEMPOOL_HEADER_SIZE(mp, cache_size));
895 ret = strlcpy(mp->name, name, sizeof(mp->name));
896 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
897 rte_errno = ENAMETOOLONG;
903 mp->socket_id = socket_id;
904 mp->elt_size = objsz.elt_size;
905 mp->header_size = objsz.header_size;
906 mp->trailer_size = objsz.trailer_size;
907 /* Size of default caches, zero means disabled. */
908 mp->cache_size = cache_size;
909 mp->private_data_size = private_data_size;
910 STAILQ_INIT(&mp->elt_list);
911 STAILQ_INIT(&mp->mem_list);
914 * local_cache pointer is set even if cache_size is zero.
915 * The local_cache points to just past the elt_pa[] array.
917 mp->local_cache = (struct rte_mempool_cache *)
918 RTE_PTR_ADD(mp, RTE_MEMPOOL_HEADER_SIZE(mp, 0));
920 /* Init all default caches. */
921 if (cache_size != 0) {
922 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
923 mempool_cache_init(&mp->local_cache[lcore_id],
929 rte_mcfg_tailq_write_lock();
930 TAILQ_INSERT_TAIL(mempool_list, te, next);
931 rte_mcfg_tailq_write_unlock();
932 rte_mcfg_mempool_write_unlock();
934 rte_mempool_trace_create_empty(name, n, elt_size, cache_size,
935 private_data_size, flags, mp);
939 rte_mcfg_mempool_write_unlock();
941 rte_mempool_free(mp);
945 /* create the mempool */
947 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
948 unsigned cache_size, unsigned private_data_size,
949 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
950 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
951 int socket_id, unsigned flags)
954 struct rte_mempool *mp;
956 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
957 private_data_size, socket_id, flags);
962 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
963 * set the correct index into the table of ops structs.
965 if ((flags & RTE_MEMPOOL_F_SP_PUT) && (flags & RTE_MEMPOOL_F_SC_GET))
966 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
967 else if (flags & RTE_MEMPOOL_F_SP_PUT)
968 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
969 else if (flags & RTE_MEMPOOL_F_SC_GET)
970 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
972 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
977 /* call the mempool priv initializer */
979 mp_init(mp, mp_init_arg);
981 if (rte_mempool_populate_default(mp) < 0)
984 /* call the object initializers */
986 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
988 rte_mempool_trace_create(name, n, elt_size, cache_size,
989 private_data_size, mp_init, mp_init_arg, obj_init,
990 obj_init_arg, flags, mp);
994 rte_mempool_free(mp);
998 /* Return the number of entries in the mempool */
1000 rte_mempool_avail_count(const struct rte_mempool *mp)
1005 count = rte_mempool_ops_get_count(mp);
1007 if (mp->cache_size == 0)
1010 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1011 count += mp->local_cache[lcore_id].len;
1014 * due to race condition (access to len is not locked), the
1015 * total can be greater than size... so fix the result
1017 if (count > mp->size)
1022 /* return the number of entries allocated from the mempool */
1024 rte_mempool_in_use_count(const struct rte_mempool *mp)
1026 return mp->size - rte_mempool_avail_count(mp);
1029 /* dump the cache status */
1031 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1035 unsigned cache_count;
1037 fprintf(f, " internal cache infos:\n");
1038 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1040 if (mp->cache_size == 0)
1043 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1044 cache_count = mp->local_cache[lcore_id].len;
1045 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1046 lcore_id, cache_count);
1047 count += cache_count;
1049 fprintf(f, " total_cache_count=%u\n", count);
1053 #ifndef __INTEL_COMPILER
1054 #pragma GCC diagnostic ignored "-Wcast-qual"
1057 /* check and update cookies or panic (internal) */
1058 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1059 void * const *obj_table_const, unsigned n, int free)
1061 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1062 struct rte_mempool_objhdr *hdr;
1063 struct rte_mempool_objtlr *tlr;
1069 /* Force to drop the "const" attribute. This is done only when
1070 * DEBUG is enabled */
1071 tmp = (void *) obj_table_const;
1077 if (rte_mempool_from_obj(obj) != mp)
1078 rte_panic("MEMPOOL: object is owned by another "
1081 hdr = rte_mempool_get_header(obj);
1082 cookie = hdr->cookie;
1085 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1086 RTE_LOG(CRIT, MEMPOOL,
1087 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1088 obj, (const void *) mp, cookie);
1089 rte_panic("MEMPOOL: bad header cookie (put)\n");
1091 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1092 } else if (free == 1) {
1093 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1094 RTE_LOG(CRIT, MEMPOOL,
1095 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1096 obj, (const void *) mp, cookie);
1097 rte_panic("MEMPOOL: bad header cookie (get)\n");
1099 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1100 } else if (free == 2) {
1101 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1102 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1103 RTE_LOG(CRIT, MEMPOOL,
1104 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1105 obj, (const void *) mp, cookie);
1106 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1109 tlr = rte_mempool_get_trailer(obj);
1110 cookie = tlr->cookie;
1111 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1112 RTE_LOG(CRIT, MEMPOOL,
1113 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1114 obj, (const void *) mp, cookie);
1115 rte_panic("MEMPOOL: bad trailer cookie\n");
1120 RTE_SET_USED(obj_table_const);
1127 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1128 void * const *first_obj_table_const, unsigned int n, int free)
1130 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1131 struct rte_mempool_info info;
1132 const size_t total_elt_sz =
1133 mp->header_size + mp->elt_size + mp->trailer_size;
1136 rte_mempool_ops_get_info(mp, &info);
1138 for (i = 0; i < n; ++i) {
1139 void *first_obj = first_obj_table_const[i];
1141 for (j = 0; j < info.contig_block_size; ++j) {
1144 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1145 rte_mempool_check_cookies(mp, &obj, 1, free);
1150 RTE_SET_USED(first_obj_table_const);
1156 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1158 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1159 void *obj, __rte_unused unsigned idx)
1161 RTE_MEMPOOL_CHECK_COOKIES(mp, &obj, 1, 2);
1165 mempool_audit_cookies(struct rte_mempool *mp)
1169 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1170 if (num != mp->size) {
1171 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1172 "iterated only over %u elements\n",
1177 #define mempool_audit_cookies(mp) do {} while(0)
1180 #ifndef __INTEL_COMPILER
1181 #pragma GCC diagnostic error "-Wcast-qual"
1184 /* check cookies before and after objects */
1186 mempool_audit_cache(const struct rte_mempool *mp)
1188 /* check cache size consistency */
1191 if (mp->cache_size == 0)
1194 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1195 const struct rte_mempool_cache *cache;
1196 cache = &mp->local_cache[lcore_id];
1197 if (cache->len > RTE_DIM(cache->objs)) {
1198 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1200 rte_panic("MEMPOOL: invalid cache len\n");
1205 /* check the consistency of mempool (size, cookies, ...) */
1207 rte_mempool_audit(struct rte_mempool *mp)
1209 mempool_audit_cache(mp);
1210 mempool_audit_cookies(mp);
1212 /* For case where mempool DEBUG is not set, and cache size is 0 */
1216 /* dump the status of the mempool on the console */
1218 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1220 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1221 struct rte_mempool_info info;
1222 struct rte_mempool_debug_stats sum;
1225 struct rte_mempool_memhdr *memhdr;
1226 struct rte_mempool_ops *ops;
1227 unsigned common_count;
1228 unsigned cache_count;
1231 RTE_ASSERT(f != NULL);
1232 RTE_ASSERT(mp != NULL);
1234 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1235 fprintf(f, " flags=%x\n", mp->flags);
1236 fprintf(f, " socket_id=%d\n", mp->socket_id);
1237 fprintf(f, " pool=%p\n", mp->pool_data);
1238 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1239 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1240 fprintf(f, " size=%"PRIu32"\n", mp->size);
1241 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1242 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1243 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1244 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1245 fprintf(f, " total_obj_size=%"PRIu32"\n",
1246 mp->header_size + mp->elt_size + mp->trailer_size);
1248 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1250 fprintf(f, " ops_index=%d\n", mp->ops_index);
1251 ops = rte_mempool_get_ops(mp->ops_index);
1252 fprintf(f, " ops_name: <%s>\n", (ops != NULL) ? ops->name : "NA");
1254 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1255 mem_len += memhdr->len;
1257 fprintf(f, " avg bytes/object=%#Lf\n",
1258 (long double)mem_len / mp->size);
1261 cache_count = rte_mempool_dump_cache(f, mp);
1262 common_count = rte_mempool_ops_get_count(mp);
1263 if ((cache_count + common_count) > mp->size)
1264 common_count = mp->size - cache_count;
1265 fprintf(f, " common_pool_count=%u\n", common_count);
1267 /* sum and dump statistics */
1268 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1269 rte_mempool_ops_get_info(mp, &info);
1270 memset(&sum, 0, sizeof(sum));
1271 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1272 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1273 sum.put_objs += mp->stats[lcore_id].put_objs;
1274 sum.put_common_pool_bulk += mp->stats[lcore_id].put_common_pool_bulk;
1275 sum.put_common_pool_objs += mp->stats[lcore_id].put_common_pool_objs;
1276 sum.get_common_pool_bulk += mp->stats[lcore_id].get_common_pool_bulk;
1277 sum.get_common_pool_objs += mp->stats[lcore_id].get_common_pool_objs;
1278 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1279 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1280 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1281 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1282 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1283 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1285 fprintf(f, " stats:\n");
1286 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1287 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1288 fprintf(f, " put_common_pool_bulk=%"PRIu64"\n", sum.put_common_pool_bulk);
1289 fprintf(f, " put_common_pool_objs=%"PRIu64"\n", sum.put_common_pool_objs);
1290 fprintf(f, " get_common_pool_bulk=%"PRIu64"\n", sum.get_common_pool_bulk);
1291 fprintf(f, " get_common_pool_objs=%"PRIu64"\n", sum.get_common_pool_objs);
1292 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1293 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1294 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1295 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1296 if (info.contig_block_size > 0) {
1297 fprintf(f, " get_success_blks=%"PRIu64"\n",
1298 sum.get_success_blks);
1299 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1302 fprintf(f, " no statistics available\n");
1305 rte_mempool_audit(mp);
1308 /* dump the status of all mempools on the console */
1310 rte_mempool_list_dump(FILE *f)
1312 struct rte_mempool *mp = NULL;
1313 struct rte_tailq_entry *te;
1314 struct rte_mempool_list *mempool_list;
1316 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1318 rte_mcfg_mempool_read_lock();
1320 TAILQ_FOREACH(te, mempool_list, next) {
1321 mp = (struct rte_mempool *) te->data;
1322 rte_mempool_dump(f, mp);
1325 rte_mcfg_mempool_read_unlock();
1328 /* search a mempool from its name */
1329 struct rte_mempool *
1330 rte_mempool_lookup(const char *name)
1332 struct rte_mempool *mp = NULL;
1333 struct rte_tailq_entry *te;
1334 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(te, mempool_list, next) {
1341 mp = (struct rte_mempool *) te->data;
1342 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1346 rte_mcfg_mempool_read_unlock();
1356 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1359 struct rte_tailq_entry *te = NULL;
1360 struct rte_mempool_list *mempool_list;
1363 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1365 rte_mcfg_mempool_read_lock();
1367 RTE_TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1368 (*func)((struct rte_mempool *) te->data, arg);
1371 rte_mcfg_mempool_read_unlock();
1374 struct mempool_callback_data {
1375 rte_mempool_event_callback *func;
1380 mempool_event_callback_invoke(enum rte_mempool_event event,
1381 struct rte_mempool *mp)
1383 struct mempool_callback_list *list;
1384 struct rte_tailq_entry *te;
1387 rte_mcfg_tailq_read_lock();
1388 list = RTE_TAILQ_CAST(callback_tailq.head, mempool_callback_list);
1389 RTE_TAILQ_FOREACH_SAFE(te, list, next, tmp_te) {
1390 struct mempool_callback_data *cb = te->data;
1391 rte_mcfg_tailq_read_unlock();
1392 cb->func(event, mp, cb->user_data);
1393 rte_mcfg_tailq_read_lock();
1395 rte_mcfg_tailq_read_unlock();
1399 rte_mempool_event_callback_register(rte_mempool_event_callback *func,
1402 struct mempool_callback_list *list;
1403 struct rte_tailq_entry *te = NULL;
1404 struct mempool_callback_data *cb;
1413 rte_mcfg_tailq_write_lock();
1414 list = RTE_TAILQ_CAST(callback_tailq.head, mempool_callback_list);
1415 RTE_TAILQ_FOREACH_SAFE(te, list, next, tmp_te) {
1417 if (cb->func == func && cb->user_data == user_data) {
1423 te = rte_zmalloc("mempool_cb_tail_entry", sizeof(*te), 0);
1425 RTE_LOG(ERR, MEMPOOL,
1426 "Cannot allocate event callback tailq entry!\n");
1431 cb = rte_malloc("mempool_cb_data", sizeof(*cb), 0);
1433 RTE_LOG(ERR, MEMPOOL,
1434 "Cannot allocate event callback!\n");
1441 cb->user_data = user_data;
1443 TAILQ_INSERT_TAIL(list, te, next);
1447 rte_mcfg_tailq_write_unlock();
1453 rte_mempool_event_callback_unregister(rte_mempool_event_callback *func,
1456 struct mempool_callback_list *list;
1457 struct rte_tailq_entry *te = NULL;
1458 struct mempool_callback_data *cb;
1461 rte_mcfg_tailq_write_lock();
1462 list = RTE_TAILQ_CAST(callback_tailq.head, mempool_callback_list);
1463 TAILQ_FOREACH(te, list, next) {
1465 if (cb->func == func && cb->user_data == user_data) {
1466 TAILQ_REMOVE(list, te, next);
1471 rte_mcfg_tailq_write_unlock();
1482 mempool_list_cb(struct rte_mempool *mp, void *arg)
1484 struct rte_tel_data *d = (struct rte_tel_data *)arg;
1486 rte_tel_data_add_array_string(d, mp->name);
1490 mempool_handle_list(const char *cmd __rte_unused,
1491 const char *params __rte_unused, struct rte_tel_data *d)
1493 rte_tel_data_start_array(d, RTE_TEL_STRING_VAL);
1494 rte_mempool_walk(mempool_list_cb, d);
1498 struct mempool_info_cb_arg {
1500 struct rte_tel_data *d;
1504 mempool_info_cb(struct rte_mempool *mp, void *arg)
1506 struct mempool_info_cb_arg *info = (struct mempool_info_cb_arg *)arg;
1507 const struct rte_memzone *mz;
1509 if (strncmp(mp->name, info->pool_name, RTE_MEMZONE_NAMESIZE))
1512 rte_tel_data_add_dict_string(info->d, "name", mp->name);
1513 rte_tel_data_add_dict_int(info->d, "pool_id", mp->pool_id);
1514 rte_tel_data_add_dict_int(info->d, "flags", mp->flags);
1515 rte_tel_data_add_dict_int(info->d, "socket_id", mp->socket_id);
1516 rte_tel_data_add_dict_int(info->d, "size", mp->size);
1517 rte_tel_data_add_dict_int(info->d, "cache_size", mp->cache_size);
1518 rte_tel_data_add_dict_int(info->d, "elt_size", mp->elt_size);
1519 rte_tel_data_add_dict_int(info->d, "header_size", mp->header_size);
1520 rte_tel_data_add_dict_int(info->d, "trailer_size", mp->trailer_size);
1521 rte_tel_data_add_dict_int(info->d, "private_data_size",
1522 mp->private_data_size);
1523 rte_tel_data_add_dict_int(info->d, "ops_index", mp->ops_index);
1524 rte_tel_data_add_dict_int(info->d, "populated_size",
1525 mp->populated_size);
1528 rte_tel_data_add_dict_string(info->d, "mz_name", mz->name);
1529 rte_tel_data_add_dict_int(info->d, "mz_len", mz->len);
1530 rte_tel_data_add_dict_int(info->d, "mz_hugepage_sz",
1532 rte_tel_data_add_dict_int(info->d, "mz_socket_id", mz->socket_id);
1533 rte_tel_data_add_dict_int(info->d, "mz_flags", mz->flags);
1537 mempool_handle_info(const char *cmd __rte_unused, const char *params,
1538 struct rte_tel_data *d)
1540 struct mempool_info_cb_arg mp_arg;
1541 char name[RTE_MEMZONE_NAMESIZE];
1543 if (!params || strlen(params) == 0)
1546 rte_strlcpy(name, params, RTE_MEMZONE_NAMESIZE);
1548 rte_tel_data_start_dict(d);
1549 mp_arg.pool_name = name;
1551 rte_mempool_walk(mempool_info_cb, &mp_arg);
1556 RTE_INIT(mempool_init_telemetry)
1558 rte_telemetry_register_cmd("/mempool/list", mempool_handle_list,
1559 "Returns list of available mempool. Takes no parameters");
1560 rte_telemetry_register_cmd("/mempool/info", mempool_handle_info,
1561 "Returns mempool info. Parameters: pool_name");