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 TAILQ_HEAD(mempool_callback_list, rte_tailq_entry);
47 static struct rte_tailq_elem callback_tailq = {
48 .name = "RTE_MEMPOOL_CALLBACK",
50 EAL_REGISTER_TAILQ(callback_tailq)
52 /* Invoke all registered mempool event callbacks. */
54 mempool_event_callback_invoke(enum rte_mempool_event event,
55 struct rte_mempool *mp);
57 #define CACHE_FLUSHTHRESH_MULTIPLIER 1.5
58 #define CALC_CACHE_FLUSHTHRESH(c) \
59 ((typeof(c))((c) * CACHE_FLUSHTHRESH_MULTIPLIER))
61 #if defined(RTE_ARCH_X86)
63 * return the greatest common divisor between a and b (fast algorithm)
66 static unsigned get_gcd(unsigned a, unsigned b)
91 * Depending on memory configuration on x86 arch, objects addresses are spread
92 * between channels and ranks in RAM: the pool allocator will add
93 * padding between objects. This function return the new size of the
97 arch_mem_object_align(unsigned int obj_size)
99 unsigned nrank, nchan;
100 unsigned new_obj_size;
102 /* get number of channels */
103 nchan = rte_memory_get_nchannel();
107 nrank = rte_memory_get_nrank();
111 /* process new object size */
112 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
113 while (get_gcd(new_obj_size, nrank * nchan) != 1)
115 return new_obj_size * RTE_MEMPOOL_ALIGN;
119 arch_mem_object_align(unsigned int obj_size)
125 struct pagesz_walk_arg {
131 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
133 struct pagesz_walk_arg *wa = arg;
137 * we need to only look at page sizes available for a particular socket
138 * ID. so, we either need an exact match on socket ID (can match both
139 * native and external memory), or, if SOCKET_ID_ANY was specified as a
140 * socket ID argument, we must only look at native memory and ignore any
141 * page sizes associated with external memory.
143 valid = msl->socket_id == wa->socket_id;
144 valid |= wa->socket_id == SOCKET_ID_ANY && msl->external == 0;
146 if (valid && msl->page_sz < wa->min)
147 wa->min = msl->page_sz;
153 get_min_page_size(int socket_id)
155 struct pagesz_walk_arg wa;
158 wa.socket_id = socket_id;
160 rte_memseg_list_walk(find_min_pagesz, &wa);
162 return wa.min == SIZE_MAX ? (size_t) rte_mem_page_size() : wa.min;
167 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
168 void *obj, rte_iova_t iova)
170 struct rte_mempool_objhdr *hdr;
171 struct rte_mempool_objtlr *tlr __rte_unused;
173 /* set mempool ptr in header */
174 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
177 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
178 mp->populated_size++;
180 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
181 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
182 tlr = rte_mempool_get_trailer(obj);
183 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
187 /* call obj_cb() for each mempool element */
189 rte_mempool_obj_iter(struct rte_mempool *mp,
190 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
192 struct rte_mempool_objhdr *hdr;
196 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
197 obj = (char *)hdr + sizeof(*hdr);
198 obj_cb(mp, obj_cb_arg, obj, n);
205 /* call mem_cb() for each mempool memory chunk */
207 rte_mempool_mem_iter(struct rte_mempool *mp,
208 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
210 struct rte_mempool_memhdr *hdr;
213 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
214 mem_cb(mp, mem_cb_arg, hdr, n);
221 /* get the header, trailer and total size of a mempool element. */
223 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
224 struct rte_mempool_objsz *sz)
226 struct rte_mempool_objsz lsz;
228 sz = (sz != NULL) ? sz : &lsz;
230 sz->header_size = sizeof(struct rte_mempool_objhdr);
231 if ((flags & RTE_MEMPOOL_F_NO_CACHE_ALIGN) == 0)
232 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
235 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
236 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
238 sz->trailer_size = 0;
241 /* element size is 8 bytes-aligned at least */
242 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
244 /* expand trailer to next cache line */
245 if ((flags & RTE_MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
246 sz->total_size = sz->header_size + sz->elt_size +
248 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
249 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
250 RTE_MEMPOOL_ALIGN_MASK);
254 * increase trailer to add padding between objects in order to
255 * spread them across memory channels/ranks
257 if ((flags & RTE_MEMPOOL_F_NO_SPREAD) == 0) {
259 new_size = arch_mem_object_align
260 (sz->header_size + sz->elt_size + sz->trailer_size);
261 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
264 /* this is the size of an object, including header and trailer */
265 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
267 return sz->total_size;
270 /* free a memchunk allocated with rte_memzone_reserve() */
272 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
275 const struct rte_memzone *mz = opaque;
276 rte_memzone_free(mz);
279 /* Free memory chunks used by a mempool. Objects must be in pool */
281 rte_mempool_free_memchunks(struct rte_mempool *mp)
283 struct rte_mempool_memhdr *memhdr;
286 while (!STAILQ_EMPTY(&mp->elt_list)) {
287 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
289 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
290 mp->populated_size--;
293 while (!STAILQ_EMPTY(&mp->mem_list)) {
294 memhdr = STAILQ_FIRST(&mp->mem_list);
295 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
296 if (memhdr->free_cb != NULL)
297 memhdr->free_cb(memhdr, memhdr->opaque);
304 mempool_ops_alloc_once(struct rte_mempool *mp)
308 /* create the internal ring if not already done */
309 if ((mp->flags & RTE_MEMPOOL_F_POOL_CREATED) == 0) {
310 ret = rte_mempool_ops_alloc(mp);
313 mp->flags |= RTE_MEMPOOL_F_POOL_CREATED;
318 /* Add objects in the pool, using a physically contiguous memory
319 * zone. Return the number of objects added, or a negative value
323 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
324 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
329 struct rte_mempool_memhdr *memhdr;
332 ret = mempool_ops_alloc_once(mp);
336 /* mempool is already populated */
337 if (mp->populated_size >= mp->size)
340 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
345 memhdr->addr = vaddr;
348 memhdr->free_cb = free_cb;
349 memhdr->opaque = opaque;
351 if (mp->flags & RTE_MEMPOOL_F_NO_CACHE_ALIGN)
352 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
354 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_MEMPOOL_ALIGN) - vaddr;
361 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
363 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
364 len - off, mempool_add_elem, NULL);
366 /* not enough room to store one object */
372 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
375 /* At least some objects in the pool can now be used for IO. */
376 if (iova != RTE_BAD_IOVA)
377 mp->flags &= ~RTE_MEMPOOL_F_NON_IO;
379 /* Report the mempool as ready only when fully populated. */
380 if (mp->populated_size >= mp->size)
381 mempool_event_callback_invoke(RTE_MEMPOOL_EVENT_READY, mp);
383 rte_mempool_trace_populate_iova(mp, vaddr, iova, len, free_cb, opaque);
394 struct rte_memseg *ms;
396 /* try registered memory first */
397 ms = rte_mem_virt2memseg(addr, NULL);
398 if (ms == NULL || ms->iova == RTE_BAD_IOVA)
399 /* fall back to actual physical address */
400 return rte_mem_virt2iova(addr);
401 return ms->iova + RTE_PTR_DIFF(addr, ms->addr);
404 /* Populate the mempool with a virtual area. Return the number of
405 * objects added, or a negative value on error.
408 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
409 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
413 size_t off, phys_len;
416 if (mp->flags & RTE_MEMPOOL_F_NO_IOVA_CONTIG)
417 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
418 len, free_cb, opaque);
420 for (off = 0; off < len &&
421 mp->populated_size < mp->size; off += phys_len) {
423 iova = get_iova(addr + off);
425 /* populate with the largest group of contiguous pages */
426 for (phys_len = RTE_MIN(
427 (size_t)(RTE_PTR_ALIGN_CEIL(addr + off + 1, pg_sz) -
430 off + phys_len < len;
431 phys_len = RTE_MIN(phys_len + pg_sz, len - off)) {
434 iova_tmp = get_iova(addr + off + phys_len);
436 if (iova_tmp == RTE_BAD_IOVA ||
437 iova_tmp != iova + phys_len)
441 ret = rte_mempool_populate_iova(mp, addr + off, iova,
442 phys_len, free_cb, opaque);
447 /* no need to call the free callback for next chunks */
452 rte_mempool_trace_populate_virt(mp, addr, len, pg_sz, free_cb, opaque);
456 rte_mempool_free_memchunks(mp);
460 /* Get the minimal page size used in a mempool before populating it. */
462 rte_mempool_get_page_size(struct rte_mempool *mp, size_t *pg_sz)
464 bool need_iova_contig_obj;
465 bool alloc_in_ext_mem;
468 /* check if we can retrieve a valid socket ID */
469 ret = rte_malloc_heap_socket_is_external(mp->socket_id);
472 alloc_in_ext_mem = (ret == 1);
473 need_iova_contig_obj = !(mp->flags & RTE_MEMPOOL_F_NO_IOVA_CONTIG);
475 if (!need_iova_contig_obj)
477 else if (rte_eal_has_hugepages() || alloc_in_ext_mem)
478 *pg_sz = get_min_page_size(mp->socket_id);
480 *pg_sz = rte_mem_page_size();
482 rte_mempool_trace_get_page_size(mp, *pg_sz);
486 /* Default function to populate the mempool: allocate memory in memzones,
487 * and populate them. Return the number of objects added, or a negative
491 rte_mempool_populate_default(struct rte_mempool *mp)
493 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
494 char mz_name[RTE_MEMZONE_NAMESIZE];
495 const struct rte_memzone *mz;
497 size_t align, pg_sz, pg_shift = 0;
501 bool need_iova_contig_obj;
502 size_t max_alloc_size = SIZE_MAX;
504 ret = mempool_ops_alloc_once(mp);
508 /* mempool must not be populated */
509 if (mp->nb_mem_chunks != 0)
513 * the following section calculates page shift and page size values.
515 * these values impact the result of calc_mem_size operation, which
516 * returns the amount of memory that should be allocated to store the
517 * desired number of objects. when not zero, it allocates more memory
518 * for the padding between objects, to ensure that an object does not
519 * cross a page boundary. in other words, page size/shift are to be set
520 * to zero if mempool elements won't care about page boundaries.
521 * there are several considerations for page size and page shift here.
523 * if we don't need our mempools to have physically contiguous objects,
524 * then just set page shift and page size to 0, because the user has
525 * indicated that there's no need to care about anything.
527 * if we do need contiguous objects (if a mempool driver has its
528 * own calc_size() method returning min_chunk_size = mem_size),
529 * there is also an option to reserve the entire mempool memory
530 * as one contiguous block of memory.
532 * if we require contiguous objects, but not necessarily the entire
533 * mempool reserved space to be contiguous, pg_sz will be != 0,
534 * and the default ops->populate() will take care of not placing
535 * objects across pages.
537 * if our IO addresses are physical, we may get memory from bigger
538 * pages, or we might get memory from smaller pages, and how much of it
539 * we require depends on whether we want bigger or smaller pages.
540 * However, requesting each and every memory size is too much work, so
541 * what we'll do instead is walk through the page sizes available, pick
542 * the smallest one and set up page shift to match that one. We will be
543 * wasting some space this way, but it's much nicer than looping around
544 * trying to reserve each and every page size.
546 * If we fail to get enough contiguous memory, then we'll go and
547 * reserve space in smaller chunks.
550 need_iova_contig_obj = !(mp->flags & RTE_MEMPOOL_F_NO_IOVA_CONTIG);
551 ret = rte_mempool_get_page_size(mp, &pg_sz);
556 pg_shift = rte_bsf32(pg_sz);
558 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
559 size_t min_chunk_size;
561 mem_size = rte_mempool_ops_calc_mem_size(
562 mp, n, pg_shift, &min_chunk_size, &align);
569 ret = snprintf(mz_name, sizeof(mz_name),
570 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
571 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
576 /* if we're trying to reserve contiguous memory, add appropriate
579 if (min_chunk_size == (size_t)mem_size)
580 mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
582 /* Allocate a memzone, retrying with a smaller area on ENOMEM */
584 mz = rte_memzone_reserve_aligned(mz_name,
585 RTE_MIN((size_t)mem_size, max_alloc_size),
586 mp->socket_id, mz_flags, align);
588 if (mz != NULL || rte_errno != ENOMEM)
591 max_alloc_size = RTE_MIN(max_alloc_size,
592 (size_t)mem_size) / 2;
593 } while (mz == NULL && max_alloc_size >= min_chunk_size);
600 if (need_iova_contig_obj)
605 if (pg_sz == 0 || (mz_flags & RTE_MEMZONE_IOVA_CONTIG))
606 ret = rte_mempool_populate_iova(mp, mz->addr,
608 rte_mempool_memchunk_mz_free,
609 (void *)(uintptr_t)mz);
611 ret = rte_mempool_populate_virt(mp, mz->addr,
613 rte_mempool_memchunk_mz_free,
614 (void *)(uintptr_t)mz);
615 if (ret == 0) /* should not happen */
618 rte_memzone_free(mz);
623 rte_mempool_trace_populate_default(mp);
627 rte_mempool_free_memchunks(mp);
631 /* return the memory size required for mempool objects in anonymous mem */
633 get_anon_size(const struct rte_mempool *mp)
636 size_t pg_sz, pg_shift;
637 size_t min_chunk_size;
640 pg_sz = rte_mem_page_size();
641 pg_shift = rte_bsf32(pg_sz);
642 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
643 &min_chunk_size, &align);
648 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
650 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
656 * Calculate size since memhdr->len has contiguous chunk length
657 * which may be smaller if anon map is split into many contiguous
658 * chunks. Result must be the same as we calculated on populate.
660 size = get_anon_size(memhdr->mp);
664 rte_mem_unmap(opaque, size);
667 /* populate the mempool with an anonymous mapping */
669 rte_mempool_populate_anon(struct rte_mempool *mp)
675 /* mempool is already populated, error */
676 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
681 ret = mempool_ops_alloc_once(mp);
687 size = get_anon_size(mp);
693 /* get chunk of virtually continuous memory */
694 addr = rte_mem_map(NULL, size, RTE_PROT_READ | RTE_PROT_WRITE,
695 RTE_MAP_SHARED | RTE_MAP_ANONYMOUS, -1, 0);
698 /* can't use MMAP_LOCKED, it does not exist on BSD */
699 if (rte_mem_lock(addr, size) < 0) {
700 rte_mem_unmap(addr, size);
704 ret = rte_mempool_populate_virt(mp, addr, size, rte_mem_page_size(),
705 rte_mempool_memchunk_anon_free, addr);
706 if (ret == 0) /* should not happen */
713 rte_mempool_trace_populate_anon(mp);
714 return mp->populated_size;
717 rte_mempool_free_memchunks(mp);
723 rte_mempool_free(struct rte_mempool *mp)
725 struct rte_mempool_list *mempool_list = NULL;
726 struct rte_tailq_entry *te;
731 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
732 rte_mcfg_tailq_write_lock();
733 /* find out tailq entry */
734 TAILQ_FOREACH(te, mempool_list, next) {
735 if (te->data == (void *)mp)
740 TAILQ_REMOVE(mempool_list, te, next);
743 rte_mcfg_tailq_write_unlock();
745 mempool_event_callback_invoke(RTE_MEMPOOL_EVENT_DESTROY, mp);
746 rte_mempool_trace_free(mp);
747 rte_mempool_free_memchunks(mp);
748 rte_mempool_ops_free(mp);
749 rte_memzone_free(mp->mz);
753 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
756 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
761 * Create and initialize a cache for objects that are retrieved from and
762 * returned to an underlying mempool. This structure is identical to the
763 * local_cache[lcore_id] pointed to by the mempool structure.
765 struct rte_mempool_cache *
766 rte_mempool_cache_create(uint32_t size, int socket_id)
768 struct rte_mempool_cache *cache;
770 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
775 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
776 RTE_CACHE_LINE_SIZE, socket_id);
778 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
783 mempool_cache_init(cache, size);
785 rte_mempool_trace_cache_create(size, socket_id, cache);
790 * Free a cache. It's the responsibility of the user to make sure that any
791 * remaining objects in the cache are flushed to the corresponding
795 rte_mempool_cache_free(struct rte_mempool_cache *cache)
797 rte_mempool_trace_cache_free(cache);
801 #define MEMPOOL_KNOWN_FLAGS (RTE_MEMPOOL_F_NO_SPREAD \
802 | RTE_MEMPOOL_F_NO_CACHE_ALIGN \
803 | RTE_MEMPOOL_F_SP_PUT \
804 | RTE_MEMPOOL_F_SC_GET \
805 | RTE_MEMPOOL_F_POOL_CREATED \
806 | RTE_MEMPOOL_F_NO_IOVA_CONTIG \
808 /* create an empty mempool */
810 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
811 unsigned cache_size, unsigned private_data_size,
812 int socket_id, unsigned flags)
814 char mz_name[RTE_MEMZONE_NAMESIZE];
815 struct rte_mempool_list *mempool_list;
816 struct rte_mempool *mp = NULL;
817 struct rte_tailq_entry *te = NULL;
818 const struct rte_memzone *mz = NULL;
820 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
821 struct rte_mempool_objsz objsz;
825 /* compilation-time checks */
826 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
827 RTE_CACHE_LINE_MASK) != 0);
828 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
829 RTE_CACHE_LINE_MASK) != 0);
830 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
831 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
832 RTE_CACHE_LINE_MASK) != 0);
833 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
834 RTE_CACHE_LINE_MASK) != 0);
837 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
839 /* asked for zero items */
845 /* asked cache too big */
846 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
847 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
852 /* enforce no unknown flag is passed by the application */
853 if ((flags & ~MEMPOOL_KNOWN_FLAGS) != 0) {
859 * No objects in the pool can be used for IO until it's populated
860 * with at least some objects with valid IOVA.
862 flags |= RTE_MEMPOOL_F_NON_IO;
864 /* "no cache align" imply "no spread" */
865 if (flags & RTE_MEMPOOL_F_NO_CACHE_ALIGN)
866 flags |= RTE_MEMPOOL_F_NO_SPREAD;
868 /* calculate mempool object sizes. */
869 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
874 rte_mcfg_mempool_write_lock();
877 * reserve a memory zone for this mempool: private data is
880 private_data_size = (private_data_size +
881 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
884 /* try to allocate tailq entry */
885 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
887 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
891 mempool_size = RTE_MEMPOOL_HEADER_SIZE(mp, cache_size);
892 mempool_size += private_data_size;
893 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
895 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
896 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
897 rte_errno = ENAMETOOLONG;
901 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
905 /* init the mempool structure */
907 memset(mp, 0, RTE_MEMPOOL_HEADER_SIZE(mp, cache_size));
908 ret = strlcpy(mp->name, name, sizeof(mp->name));
909 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
910 rte_errno = ENAMETOOLONG;
916 mp->socket_id = socket_id;
917 mp->elt_size = objsz.elt_size;
918 mp->header_size = objsz.header_size;
919 mp->trailer_size = objsz.trailer_size;
920 /* Size of default caches, zero means disabled. */
921 mp->cache_size = cache_size;
922 mp->private_data_size = private_data_size;
923 STAILQ_INIT(&mp->elt_list);
924 STAILQ_INIT(&mp->mem_list);
927 * local_cache pointer is set even if cache_size is zero.
928 * The local_cache points to just past the elt_pa[] array.
930 mp->local_cache = (struct rte_mempool_cache *)
931 RTE_PTR_ADD(mp, RTE_MEMPOOL_HEADER_SIZE(mp, 0));
933 /* Init all default caches. */
934 if (cache_size != 0) {
935 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
936 mempool_cache_init(&mp->local_cache[lcore_id],
942 rte_mcfg_tailq_write_lock();
943 TAILQ_INSERT_TAIL(mempool_list, te, next);
944 rte_mcfg_tailq_write_unlock();
945 rte_mcfg_mempool_write_unlock();
947 rte_mempool_trace_create_empty(name, n, elt_size, cache_size,
948 private_data_size, flags, mp);
952 rte_mcfg_mempool_write_unlock();
954 rte_mempool_free(mp);
958 /* create the mempool */
960 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
961 unsigned cache_size, unsigned private_data_size,
962 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
963 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
964 int socket_id, unsigned flags)
967 struct rte_mempool *mp;
969 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
970 private_data_size, socket_id, flags);
975 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
976 * set the correct index into the table of ops structs.
978 if ((flags & RTE_MEMPOOL_F_SP_PUT) && (flags & RTE_MEMPOOL_F_SC_GET))
979 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
980 else if (flags & RTE_MEMPOOL_F_SP_PUT)
981 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
982 else if (flags & RTE_MEMPOOL_F_SC_GET)
983 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
985 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
990 /* call the mempool priv initializer */
992 mp_init(mp, mp_init_arg);
994 if (rte_mempool_populate_default(mp) < 0)
997 /* call the object initializers */
999 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1001 rte_mempool_trace_create(name, n, elt_size, cache_size,
1002 private_data_size, mp_init, mp_init_arg, obj_init,
1003 obj_init_arg, flags, mp);
1007 rte_mempool_free(mp);
1011 /* Return the number of entries in the mempool */
1013 rte_mempool_avail_count(const struct rte_mempool *mp)
1018 count = rte_mempool_ops_get_count(mp);
1020 if (mp->cache_size == 0)
1023 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1024 count += mp->local_cache[lcore_id].len;
1027 * due to race condition (access to len is not locked), the
1028 * total can be greater than size... so fix the result
1030 if (count > mp->size)
1035 /* return the number of entries allocated from the mempool */
1037 rte_mempool_in_use_count(const struct rte_mempool *mp)
1039 return mp->size - rte_mempool_avail_count(mp);
1042 /* dump the cache status */
1044 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1048 unsigned cache_count;
1050 fprintf(f, " internal cache infos:\n");
1051 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1053 if (mp->cache_size == 0)
1056 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1057 cache_count = mp->local_cache[lcore_id].len;
1058 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1059 lcore_id, cache_count);
1060 count += cache_count;
1062 fprintf(f, " total_cache_count=%u\n", count);
1066 #ifndef __INTEL_COMPILER
1067 #pragma GCC diagnostic ignored "-Wcast-qual"
1070 /* check and update cookies or panic (internal) */
1071 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1072 void * const *obj_table_const, unsigned n, int free)
1074 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1075 struct rte_mempool_objhdr *hdr;
1076 struct rte_mempool_objtlr *tlr;
1082 /* Force to drop the "const" attribute. This is done only when
1083 * DEBUG is enabled */
1084 tmp = (void *) obj_table_const;
1090 if (rte_mempool_from_obj(obj) != mp)
1091 rte_panic("MEMPOOL: object is owned by another "
1094 hdr = rte_mempool_get_header(obj);
1095 cookie = hdr->cookie;
1098 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1099 RTE_LOG(CRIT, MEMPOOL,
1100 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1101 obj, (const void *) mp, cookie);
1102 rte_panic("MEMPOOL: bad header cookie (put)\n");
1104 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1105 } else if (free == 1) {
1106 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1107 RTE_LOG(CRIT, MEMPOOL,
1108 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1109 obj, (const void *) mp, cookie);
1110 rte_panic("MEMPOOL: bad header cookie (get)\n");
1112 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1113 } else if (free == 2) {
1114 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1115 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1116 RTE_LOG(CRIT, MEMPOOL,
1117 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1118 obj, (const void *) mp, cookie);
1119 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1122 tlr = rte_mempool_get_trailer(obj);
1123 cookie = tlr->cookie;
1124 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1125 RTE_LOG(CRIT, MEMPOOL,
1126 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1127 obj, (const void *) mp, cookie);
1128 rte_panic("MEMPOOL: bad trailer cookie\n");
1133 RTE_SET_USED(obj_table_const);
1140 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1141 void * const *first_obj_table_const, unsigned int n, int free)
1143 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1144 struct rte_mempool_info info;
1145 const size_t total_elt_sz =
1146 mp->header_size + mp->elt_size + mp->trailer_size;
1149 rte_mempool_ops_get_info(mp, &info);
1151 for (i = 0; i < n; ++i) {
1152 void *first_obj = first_obj_table_const[i];
1154 for (j = 0; j < info.contig_block_size; ++j) {
1157 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1158 rte_mempool_check_cookies(mp, &obj, 1, free);
1163 RTE_SET_USED(first_obj_table_const);
1169 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1171 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1172 void *obj, __rte_unused unsigned idx)
1174 RTE_MEMPOOL_CHECK_COOKIES(mp, &obj, 1, 2);
1178 mempool_audit_cookies(struct rte_mempool *mp)
1182 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1183 if (num != mp->size) {
1184 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1185 "iterated only over %u elements\n",
1190 #define mempool_audit_cookies(mp) do {} while(0)
1193 #ifndef __INTEL_COMPILER
1194 #pragma GCC diagnostic error "-Wcast-qual"
1197 /* check cookies before and after objects */
1199 mempool_audit_cache(const struct rte_mempool *mp)
1201 /* check cache size consistency */
1204 if (mp->cache_size == 0)
1207 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1208 const struct rte_mempool_cache *cache;
1209 cache = &mp->local_cache[lcore_id];
1210 if (cache->len > RTE_DIM(cache->objs)) {
1211 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1213 rte_panic("MEMPOOL: invalid cache len\n");
1218 /* check the consistency of mempool (size, cookies, ...) */
1220 rte_mempool_audit(struct rte_mempool *mp)
1222 mempool_audit_cache(mp);
1223 mempool_audit_cookies(mp);
1225 /* For case where mempool DEBUG is not set, and cache size is 0 */
1229 /* dump the status of the mempool on the console */
1231 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1233 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1234 struct rte_mempool_info info;
1235 struct rte_mempool_debug_stats sum;
1238 struct rte_mempool_memhdr *memhdr;
1239 struct rte_mempool_ops *ops;
1240 unsigned common_count;
1241 unsigned cache_count;
1244 RTE_ASSERT(f != NULL);
1245 RTE_ASSERT(mp != NULL);
1247 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1248 fprintf(f, " flags=%x\n", mp->flags);
1249 fprintf(f, " socket_id=%d\n", mp->socket_id);
1250 fprintf(f, " pool=%p\n", mp->pool_data);
1251 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1252 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1253 fprintf(f, " size=%"PRIu32"\n", mp->size);
1254 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1255 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1256 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1257 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1258 fprintf(f, " total_obj_size=%"PRIu32"\n",
1259 mp->header_size + mp->elt_size + mp->trailer_size);
1261 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1263 fprintf(f, " ops_index=%d\n", mp->ops_index);
1264 ops = rte_mempool_get_ops(mp->ops_index);
1265 fprintf(f, " ops_name: <%s>\n", (ops != NULL) ? ops->name : "NA");
1267 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1268 mem_len += memhdr->len;
1270 fprintf(f, " avg bytes/object=%#Lf\n",
1271 (long double)mem_len / mp->size);
1274 cache_count = rte_mempool_dump_cache(f, mp);
1275 common_count = rte_mempool_ops_get_count(mp);
1276 if ((cache_count + common_count) > mp->size)
1277 common_count = mp->size - cache_count;
1278 fprintf(f, " common_pool_count=%u\n", common_count);
1280 /* sum and dump statistics */
1281 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1282 rte_mempool_ops_get_info(mp, &info);
1283 memset(&sum, 0, sizeof(sum));
1284 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1285 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1286 sum.put_objs += mp->stats[lcore_id].put_objs;
1287 sum.put_common_pool_bulk += mp->stats[lcore_id].put_common_pool_bulk;
1288 sum.put_common_pool_objs += mp->stats[lcore_id].put_common_pool_objs;
1289 sum.get_common_pool_bulk += mp->stats[lcore_id].get_common_pool_bulk;
1290 sum.get_common_pool_objs += mp->stats[lcore_id].get_common_pool_objs;
1291 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1292 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1293 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1294 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1295 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1296 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1298 fprintf(f, " stats:\n");
1299 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1300 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1301 fprintf(f, " put_common_pool_bulk=%"PRIu64"\n", sum.put_common_pool_bulk);
1302 fprintf(f, " put_common_pool_objs=%"PRIu64"\n", sum.put_common_pool_objs);
1303 fprintf(f, " get_common_pool_bulk=%"PRIu64"\n", sum.get_common_pool_bulk);
1304 fprintf(f, " get_common_pool_objs=%"PRIu64"\n", sum.get_common_pool_objs);
1305 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1306 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1307 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1308 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1309 if (info.contig_block_size > 0) {
1310 fprintf(f, " get_success_blks=%"PRIu64"\n",
1311 sum.get_success_blks);
1312 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1315 fprintf(f, " no statistics available\n");
1318 rte_mempool_audit(mp);
1321 /* dump the status of all mempools on the console */
1323 rte_mempool_list_dump(FILE *f)
1325 struct rte_mempool *mp = NULL;
1326 struct rte_tailq_entry *te;
1327 struct rte_mempool_list *mempool_list;
1329 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1331 rte_mcfg_mempool_read_lock();
1333 TAILQ_FOREACH(te, mempool_list, next) {
1334 mp = (struct rte_mempool *) te->data;
1335 rte_mempool_dump(f, mp);
1338 rte_mcfg_mempool_read_unlock();
1341 /* search a mempool from its name */
1342 struct rte_mempool *
1343 rte_mempool_lookup(const char *name)
1345 struct rte_mempool *mp = NULL;
1346 struct rte_tailq_entry *te;
1347 struct rte_mempool_list *mempool_list;
1349 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1351 rte_mcfg_mempool_read_lock();
1353 TAILQ_FOREACH(te, mempool_list, next) {
1354 mp = (struct rte_mempool *) te->data;
1355 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1359 rte_mcfg_mempool_read_unlock();
1369 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1372 struct rte_tailq_entry *te = NULL;
1373 struct rte_mempool_list *mempool_list;
1376 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1378 rte_mcfg_mempool_read_lock();
1380 RTE_TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1381 (*func)((struct rte_mempool *) te->data, arg);
1384 rte_mcfg_mempool_read_unlock();
1387 struct mempool_callback_data {
1388 rte_mempool_event_callback *func;
1393 mempool_event_callback_invoke(enum rte_mempool_event event,
1394 struct rte_mempool *mp)
1396 struct mempool_callback_list *list;
1397 struct rte_tailq_entry *te;
1400 rte_mcfg_tailq_read_lock();
1401 list = RTE_TAILQ_CAST(callback_tailq.head, mempool_callback_list);
1402 RTE_TAILQ_FOREACH_SAFE(te, list, next, tmp_te) {
1403 struct mempool_callback_data *cb = te->data;
1404 rte_mcfg_tailq_read_unlock();
1405 cb->func(event, mp, cb->user_data);
1406 rte_mcfg_tailq_read_lock();
1408 rte_mcfg_tailq_read_unlock();
1412 rte_mempool_event_callback_register(rte_mempool_event_callback *func,
1415 struct mempool_callback_list *list;
1416 struct rte_tailq_entry *te = NULL;
1417 struct mempool_callback_data *cb;
1426 rte_mcfg_tailq_write_lock();
1427 list = RTE_TAILQ_CAST(callback_tailq.head, mempool_callback_list);
1428 RTE_TAILQ_FOREACH_SAFE(te, list, next, tmp_te) {
1430 if (cb->func == func && cb->user_data == user_data) {
1436 te = rte_zmalloc("mempool_cb_tail_entry", sizeof(*te), 0);
1438 RTE_LOG(ERR, MEMPOOL,
1439 "Cannot allocate event callback tailq entry!\n");
1444 cb = rte_malloc("mempool_cb_data", sizeof(*cb), 0);
1446 RTE_LOG(ERR, MEMPOOL,
1447 "Cannot allocate event callback!\n");
1454 cb->user_data = user_data;
1456 TAILQ_INSERT_TAIL(list, te, next);
1460 rte_mcfg_tailq_write_unlock();
1466 rte_mempool_event_callback_unregister(rte_mempool_event_callback *func,
1469 struct mempool_callback_list *list;
1470 struct rte_tailq_entry *te = NULL;
1471 struct mempool_callback_data *cb;
1474 rte_mcfg_tailq_write_lock();
1475 list = RTE_TAILQ_CAST(callback_tailq.head, mempool_callback_list);
1476 TAILQ_FOREACH(te, list, next) {
1478 if (cb->func == func && cb->user_data == user_data) {
1479 TAILQ_REMOVE(list, te, next);
1484 rte_mcfg_tailq_write_unlock();