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>
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>
33 #include "rte_mempool.h"
35 TAILQ_HEAD(rte_mempool_list, rte_tailq_entry);
37 static struct rte_tailq_elem rte_mempool_tailq = {
38 .name = "RTE_MEMPOOL",
40 EAL_REGISTER_TAILQ(rte_mempool_tailq)
42 #define CACHE_FLUSHTHRESH_MULTIPLIER 1.5
43 #define CALC_CACHE_FLUSHTHRESH(c) \
44 ((typeof(c))((c) * CACHE_FLUSHTHRESH_MULTIPLIER))
47 * return the greatest common divisor between a and b (fast algorithm)
50 static unsigned get_gcd(unsigned a, unsigned b)
75 * Depending on memory configuration, objects addresses are spread
76 * between channels and ranks in RAM: the pool allocator will add
77 * padding between objects. This function return the new size of the
80 static unsigned optimize_object_size(unsigned obj_size)
82 unsigned nrank, nchan;
83 unsigned new_obj_size;
85 /* get number of channels */
86 nchan = rte_memory_get_nchannel();
90 nrank = rte_memory_get_nrank();
94 /* process new object size */
95 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
96 while (get_gcd(new_obj_size, nrank * nchan) != 1)
98 return new_obj_size * RTE_MEMPOOL_ALIGN;
102 mempool_add_elem(struct rte_mempool *mp, void *obj, rte_iova_t iova)
104 struct rte_mempool_objhdr *hdr;
105 struct rte_mempool_objtlr *tlr __rte_unused;
107 /* set mempool ptr in header */
108 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
111 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
112 mp->populated_size++;
114 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
115 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
116 tlr = __mempool_get_trailer(obj);
117 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
120 /* enqueue in ring */
121 rte_mempool_ops_enqueue_bulk(mp, &obj, 1);
124 /* call obj_cb() for each mempool element */
126 rte_mempool_obj_iter(struct rte_mempool *mp,
127 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
129 struct rte_mempool_objhdr *hdr;
133 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
134 obj = (char *)hdr + sizeof(*hdr);
135 obj_cb(mp, obj_cb_arg, obj, n);
142 /* call mem_cb() for each mempool memory chunk */
144 rte_mempool_mem_iter(struct rte_mempool *mp,
145 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
147 struct rte_mempool_memhdr *hdr;
150 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
151 mem_cb(mp, mem_cb_arg, hdr, n);
158 /* get the header, trailer and total size of a mempool element. */
160 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
161 struct rte_mempool_objsz *sz)
163 struct rte_mempool_objsz lsz;
165 sz = (sz != NULL) ? sz : &lsz;
167 sz->header_size = sizeof(struct rte_mempool_objhdr);
168 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
169 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
172 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
173 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
175 sz->trailer_size = 0;
178 /* element size is 8 bytes-aligned at least */
179 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
181 /* expand trailer to next cache line */
182 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
183 sz->total_size = sz->header_size + sz->elt_size +
185 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
186 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
187 RTE_MEMPOOL_ALIGN_MASK);
191 * increase trailer to add padding between objects in order to
192 * spread them across memory channels/ranks
194 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
196 new_size = optimize_object_size(sz->header_size + sz->elt_size +
198 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
201 /* this is the size of an object, including header and trailer */
202 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
204 return sz->total_size;
209 * Calculate maximum amount of memory required to store given number of objects.
212 rte_mempool_xmem_size(uint32_t elt_num, size_t total_elt_sz, uint32_t pg_shift,
215 size_t obj_per_page, pg_num, pg_sz;
218 mask = MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS | MEMPOOL_F_CAPA_PHYS_CONTIG;
219 if ((flags & mask) == mask)
220 /* alignment need one additional object */
223 if (total_elt_sz == 0)
227 return total_elt_sz * elt_num;
229 pg_sz = (size_t)1 << pg_shift;
230 obj_per_page = pg_sz / total_elt_sz;
231 if (obj_per_page == 0)
232 return RTE_ALIGN_CEIL(total_elt_sz, pg_sz) * elt_num;
234 pg_num = (elt_num + obj_per_page - 1) / obj_per_page;
235 return pg_num << pg_shift;
239 * Calculate how much memory would be actually required with the
240 * given memory footprint to store required number of elements.
243 rte_mempool_xmem_usage(__rte_unused void *vaddr, uint32_t elt_num,
244 size_t total_elt_sz, const rte_iova_t iova[], uint32_t pg_num,
245 uint32_t pg_shift, unsigned int flags)
247 uint32_t elt_cnt = 0;
248 rte_iova_t start, end;
250 size_t pg_sz = (size_t)1 << pg_shift;
253 mask = MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS | MEMPOOL_F_CAPA_PHYS_CONTIG;
254 if ((flags & mask) == mask)
255 /* alignment need one additional object */
258 /* if iova is NULL, assume contiguous memory */
261 end = pg_sz * pg_num;
265 end = iova[0] + pg_sz;
268 while (elt_cnt < elt_num) {
270 if (end - start >= total_elt_sz) {
271 /* enough contiguous memory, add an object */
272 start += total_elt_sz;
274 } else if (iova_idx < pg_num) {
275 /* no room to store one obj, add a page */
276 if (end == iova[iova_idx]) {
279 start = iova[iova_idx];
280 end = iova[iova_idx] + pg_sz;
285 /* no more page, return how many elements fit */
286 return -(size_t)elt_cnt;
290 return (size_t)iova_idx << pg_shift;
293 /* free a memchunk allocated with rte_memzone_reserve() */
295 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
298 const struct rte_memzone *mz = opaque;
299 rte_memzone_free(mz);
302 /* Free memory chunks used by a mempool. Objects must be in pool */
304 rte_mempool_free_memchunks(struct rte_mempool *mp)
306 struct rte_mempool_memhdr *memhdr;
309 while (!STAILQ_EMPTY(&mp->elt_list)) {
310 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
312 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
313 mp->populated_size--;
316 while (!STAILQ_EMPTY(&mp->mem_list)) {
317 memhdr = STAILQ_FIRST(&mp->mem_list);
318 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
319 if (memhdr->free_cb != NULL)
320 memhdr->free_cb(memhdr, memhdr->opaque);
326 /* Add objects in the pool, using a physically contiguous memory
327 * zone. Return the number of objects added, or a negative value
331 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
332 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
335 unsigned total_elt_sz;
336 unsigned int mp_capa_flags;
339 struct rte_mempool_memhdr *memhdr;
342 /* create the internal ring if not already done */
343 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
344 ret = rte_mempool_ops_alloc(mp);
347 mp->flags |= MEMPOOL_F_POOL_CREATED;
350 /* Notify memory area to mempool */
351 ret = rte_mempool_ops_register_memory_area(mp, vaddr, iova, len);
352 if (ret != -ENOTSUP && ret < 0)
355 /* mempool is already populated */
356 if (mp->populated_size >= mp->size)
359 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
361 /* Get mempool capabilities */
363 ret = rte_mempool_ops_get_capabilities(mp, &mp_capa_flags);
364 if ((ret < 0) && (ret != -ENOTSUP))
367 /* update mempool capabilities */
368 mp->flags |= mp_capa_flags;
370 /* Detect pool area has sufficient space for elements */
371 if (mp_capa_flags & MEMPOOL_F_CAPA_PHYS_CONTIG) {
372 if (len < total_elt_sz * mp->size) {
373 RTE_LOG(ERR, MEMPOOL,
374 "pool area %" PRIx64 " not enough\n",
380 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
385 memhdr->addr = vaddr;
388 memhdr->free_cb = free_cb;
389 memhdr->opaque = opaque;
391 if (mp_capa_flags & MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS)
392 /* align object start address to a multiple of total_elt_sz */
393 off = total_elt_sz - ((uintptr_t)vaddr % total_elt_sz);
394 else if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
395 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
397 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
399 while (off + total_elt_sz <= len && mp->populated_size < mp->size) {
400 off += mp->header_size;
401 if (iova == RTE_BAD_IOVA)
402 mempool_add_elem(mp, (char *)vaddr + off,
405 mempool_add_elem(mp, (char *)vaddr + off, iova + off);
406 off += mp->elt_size + mp->trailer_size;
410 /* not enough room to store one object */
414 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
420 rte_mempool_populate_phys(struct rte_mempool *mp, char *vaddr,
421 phys_addr_t paddr, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
424 return rte_mempool_populate_iova(mp, vaddr, paddr, len, free_cb, opaque);
427 /* Add objects in the pool, using a table of physical pages. Return the
428 * number of objects added, or a negative value on error.
431 rte_mempool_populate_iova_tab(struct rte_mempool *mp, char *vaddr,
432 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift,
433 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
437 size_t pg_sz = (size_t)1 << pg_shift;
439 /* mempool must not be populated */
440 if (mp->nb_mem_chunks != 0)
443 if (mp->flags & MEMPOOL_F_NO_PHYS_CONTIG)
444 return rte_mempool_populate_iova(mp, vaddr, RTE_BAD_IOVA,
445 pg_num * pg_sz, free_cb, opaque);
447 for (i = 0; i < pg_num && mp->populated_size < mp->size; i += n) {
449 /* populate with the largest group of contiguous pages */
450 for (n = 1; (i + n) < pg_num &&
451 iova[i + n - 1] + pg_sz == iova[i + n]; n++)
454 ret = rte_mempool_populate_iova(mp, vaddr + i * pg_sz,
455 iova[i], n * pg_sz, free_cb, opaque);
457 rte_mempool_free_memchunks(mp);
460 /* no need to call the free callback for next chunks */
468 rte_mempool_populate_phys_tab(struct rte_mempool *mp, char *vaddr,
469 const phys_addr_t paddr[], uint32_t pg_num, uint32_t pg_shift,
470 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
472 return rte_mempool_populate_iova_tab(mp, vaddr, paddr, pg_num, pg_shift,
476 /* Populate the mempool with a virtual area. Return the number of
477 * objects added, or a negative value on error.
480 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
481 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
485 size_t off, phys_len;
488 /* mempool must not be populated */
489 if (mp->nb_mem_chunks != 0)
491 /* address and len must be page-aligned */
492 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
494 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
497 if (mp->flags & MEMPOOL_F_NO_PHYS_CONTIG)
498 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
499 len, free_cb, opaque);
501 for (off = 0; off + pg_sz <= len &&
502 mp->populated_size < mp->size; off += phys_len) {
504 iova = rte_mem_virt2iova(addr + off);
506 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
511 /* populate with the largest group of contiguous pages */
512 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
515 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
517 if (iova_tmp != iova + phys_len)
521 ret = rte_mempool_populate_iova(mp, addr + off, iova,
522 phys_len, free_cb, opaque);
525 /* no need to call the free callback for next chunks */
533 rte_mempool_free_memchunks(mp);
537 /* Default function to populate the mempool: allocate memory in memzones,
538 * and populate them. Return the number of objects added, or a negative
542 rte_mempool_populate_default(struct rte_mempool *mp)
544 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
545 char mz_name[RTE_MEMZONE_NAMESIZE];
546 const struct rte_memzone *mz;
547 size_t size, total_elt_sz, align, pg_sz, pg_shift;
550 unsigned int mp_flags;
553 /* mempool must not be populated */
554 if (mp->nb_mem_chunks != 0)
557 /* Get mempool capabilities */
559 ret = rte_mempool_ops_get_capabilities(mp, &mp_flags);
560 if ((ret < 0) && (ret != -ENOTSUP))
563 /* update mempool capabilities */
564 mp->flags |= mp_flags;
566 if (rte_eal_has_hugepages()) {
567 pg_shift = 0; /* not needed, zone is physically contiguous */
569 align = RTE_CACHE_LINE_SIZE;
571 pg_sz = getpagesize();
572 pg_shift = rte_bsf32(pg_sz);
576 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
577 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
578 size = rte_mempool_xmem_size(n, total_elt_sz, pg_shift,
581 ret = snprintf(mz_name, sizeof(mz_name),
582 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
583 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
588 mz = rte_memzone_reserve_aligned(mz_name, size,
589 mp->socket_id, mz_flags, align);
590 /* not enough memory, retry with the biggest zone we have */
592 mz = rte_memzone_reserve_aligned(mz_name, 0,
593 mp->socket_id, mz_flags, align);
599 if (mp->flags & MEMPOOL_F_NO_PHYS_CONTIG)
604 if (rte_eal_has_hugepages())
605 ret = rte_mempool_populate_iova(mp, mz->addr,
607 rte_mempool_memchunk_mz_free,
608 (void *)(uintptr_t)mz);
610 ret = rte_mempool_populate_virt(mp, mz->addr,
612 rte_mempool_memchunk_mz_free,
613 (void *)(uintptr_t)mz);
615 rte_memzone_free(mz);
623 rte_mempool_free_memchunks(mp);
627 /* return the memory size required for mempool objects in anonymous mem */
629 get_anon_size(const struct rte_mempool *mp)
631 size_t size, total_elt_sz, pg_sz, pg_shift;
633 pg_sz = getpagesize();
634 pg_shift = rte_bsf32(pg_sz);
635 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
636 size = rte_mempool_xmem_size(mp->size, total_elt_sz, pg_shift,
642 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
644 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
647 munmap(opaque, get_anon_size(memhdr->mp));
650 /* populate the mempool with an anonymous mapping */
652 rte_mempool_populate_anon(struct rte_mempool *mp)
658 /* mempool is already populated, error */
659 if (!STAILQ_EMPTY(&mp->mem_list)) {
664 /* get chunk of virtually continuous memory */
665 size = get_anon_size(mp);
666 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
667 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
668 if (addr == MAP_FAILED) {
672 /* can't use MMAP_LOCKED, it does not exist on BSD */
673 if (mlock(addr, size) < 0) {
679 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
680 rte_mempool_memchunk_anon_free, addr);
684 return mp->populated_size;
687 rte_mempool_free_memchunks(mp);
693 rte_mempool_free(struct rte_mempool *mp)
695 struct rte_mempool_list *mempool_list = NULL;
696 struct rte_tailq_entry *te;
701 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
702 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
703 /* find out tailq entry */
704 TAILQ_FOREACH(te, mempool_list, next) {
705 if (te->data == (void *)mp)
710 TAILQ_REMOVE(mempool_list, te, next);
713 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
715 rte_mempool_free_memchunks(mp);
716 rte_mempool_ops_free(mp);
717 rte_memzone_free(mp->mz);
721 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
724 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
729 * Create and initialize a cache for objects that are retrieved from and
730 * returned to an underlying mempool. This structure is identical to the
731 * local_cache[lcore_id] pointed to by the mempool structure.
733 struct rte_mempool_cache *
734 rte_mempool_cache_create(uint32_t size, int socket_id)
736 struct rte_mempool_cache *cache;
738 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
743 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
744 RTE_CACHE_LINE_SIZE, socket_id);
746 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
751 mempool_cache_init(cache, size);
757 * Free a cache. It's the responsibility of the user to make sure that any
758 * remaining objects in the cache are flushed to the corresponding
762 rte_mempool_cache_free(struct rte_mempool_cache *cache)
767 /* create an empty mempool */
769 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
770 unsigned cache_size, unsigned private_data_size,
771 int socket_id, unsigned flags)
773 char mz_name[RTE_MEMZONE_NAMESIZE];
774 struct rte_mempool_list *mempool_list;
775 struct rte_mempool *mp = NULL;
776 struct rte_tailq_entry *te = NULL;
777 const struct rte_memzone *mz = NULL;
779 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
780 struct rte_mempool_objsz objsz;
784 /* compilation-time checks */
785 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
786 RTE_CACHE_LINE_MASK) != 0);
787 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
788 RTE_CACHE_LINE_MASK) != 0);
789 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
790 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
791 RTE_CACHE_LINE_MASK) != 0);
792 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
793 RTE_CACHE_LINE_MASK) != 0);
796 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
798 /* asked cache too big */
799 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
800 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
805 /* "no cache align" imply "no spread" */
806 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
807 flags |= MEMPOOL_F_NO_SPREAD;
809 /* calculate mempool object sizes. */
810 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
815 rte_rwlock_write_lock(RTE_EAL_MEMPOOL_RWLOCK);
818 * reserve a memory zone for this mempool: private data is
821 private_data_size = (private_data_size +
822 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
825 /* try to allocate tailq entry */
826 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
828 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
832 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
833 mempool_size += private_data_size;
834 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
836 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
837 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
838 rte_errno = ENAMETOOLONG;
842 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
846 /* init the mempool structure */
848 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
849 ret = snprintf(mp->name, sizeof(mp->name), "%s", name);
850 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
851 rte_errno = ENAMETOOLONG;
857 mp->socket_id = socket_id;
858 mp->elt_size = objsz.elt_size;
859 mp->header_size = objsz.header_size;
860 mp->trailer_size = objsz.trailer_size;
861 /* Size of default caches, zero means disabled. */
862 mp->cache_size = cache_size;
863 mp->private_data_size = private_data_size;
864 STAILQ_INIT(&mp->elt_list);
865 STAILQ_INIT(&mp->mem_list);
868 * local_cache pointer is set even if cache_size is zero.
869 * The local_cache points to just past the elt_pa[] array.
871 mp->local_cache = (struct rte_mempool_cache *)
872 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
874 /* Init all default caches. */
875 if (cache_size != 0) {
876 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
877 mempool_cache_init(&mp->local_cache[lcore_id],
883 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
884 TAILQ_INSERT_TAIL(mempool_list, te, next);
885 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
886 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
891 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
893 rte_mempool_free(mp);
897 /* create the mempool */
899 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
900 unsigned cache_size, unsigned private_data_size,
901 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
902 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
903 int socket_id, unsigned flags)
906 struct rte_mempool *mp;
908 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
909 private_data_size, socket_id, flags);
914 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
915 * set the correct index into the table of ops structs.
917 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
918 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
919 else if (flags & MEMPOOL_F_SP_PUT)
920 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
921 else if (flags & MEMPOOL_F_SC_GET)
922 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
924 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
929 /* call the mempool priv initializer */
931 mp_init(mp, mp_init_arg);
933 if (rte_mempool_populate_default(mp) < 0)
936 /* call the object initializers */
938 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
943 rte_mempool_free(mp);
948 * Create the mempool over already allocated chunk of memory.
949 * That external memory buffer can consists of physically disjoint pages.
950 * Setting vaddr to NULL, makes mempool to fallback to rte_mempool_create()
954 rte_mempool_xmem_create(const char *name, unsigned n, unsigned elt_size,
955 unsigned cache_size, unsigned private_data_size,
956 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
957 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
958 int socket_id, unsigned flags, void *vaddr,
959 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift)
961 struct rte_mempool *mp = NULL;
964 /* no virtual address supplied, use rte_mempool_create() */
966 return rte_mempool_create(name, n, elt_size, cache_size,
967 private_data_size, mp_init, mp_init_arg,
968 obj_init, obj_init_arg, socket_id, flags);
970 /* check that we have both VA and PA */
976 /* Check that pg_shift parameter is valid. */
977 if (pg_shift > MEMPOOL_PG_SHIFT_MAX) {
982 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
983 private_data_size, socket_id, flags);
987 /* call the mempool priv initializer */
989 mp_init(mp, mp_init_arg);
991 ret = rte_mempool_populate_iova_tab(mp, vaddr, iova, pg_num, pg_shift,
993 if (ret < 0 || ret != (int)mp->size)
996 /* call the object initializers */
998 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1003 rte_mempool_free(mp);
1007 /* Return the number of entries in the mempool */
1009 rte_mempool_avail_count(const struct rte_mempool *mp)
1014 count = rte_mempool_ops_get_count(mp);
1016 if (mp->cache_size == 0)
1019 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1020 count += mp->local_cache[lcore_id].len;
1023 * due to race condition (access to len is not locked), the
1024 * total can be greater than size... so fix the result
1026 if (count > mp->size)
1031 /* return the number of entries allocated from the mempool */
1033 rte_mempool_in_use_count(const struct rte_mempool *mp)
1035 return mp->size - rte_mempool_avail_count(mp);
1038 /* dump the cache status */
1040 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1044 unsigned cache_count;
1046 fprintf(f, " internal cache infos:\n");
1047 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1049 if (mp->cache_size == 0)
1052 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1053 cache_count = mp->local_cache[lcore_id].len;
1054 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1055 lcore_id, cache_count);
1056 count += cache_count;
1058 fprintf(f, " total_cache_count=%u\n", count);
1062 #ifndef __INTEL_COMPILER
1063 #pragma GCC diagnostic ignored "-Wcast-qual"
1066 /* check and update cookies or panic (internal) */
1067 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1068 void * const *obj_table_const, unsigned n, int free)
1070 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1071 struct rte_mempool_objhdr *hdr;
1072 struct rte_mempool_objtlr *tlr;
1078 /* Force to drop the "const" attribute. This is done only when
1079 * DEBUG is enabled */
1080 tmp = (void *) obj_table_const;
1086 if (rte_mempool_from_obj(obj) != mp)
1087 rte_panic("MEMPOOL: object is owned by another "
1090 hdr = __mempool_get_header(obj);
1091 cookie = hdr->cookie;
1094 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1095 RTE_LOG(CRIT, MEMPOOL,
1096 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1097 obj, (const void *) mp, cookie);
1098 rte_panic("MEMPOOL: bad header cookie (put)\n");
1100 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1101 } else if (free == 1) {
1102 if (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 (get)\n");
1108 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1109 } else if (free == 2) {
1110 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1111 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1112 RTE_LOG(CRIT, MEMPOOL,
1113 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1114 obj, (const void *) mp, cookie);
1115 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1118 tlr = __mempool_get_trailer(obj);
1119 cookie = tlr->cookie;
1120 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1121 RTE_LOG(CRIT, MEMPOOL,
1122 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1123 obj, (const void *) mp, cookie);
1124 rte_panic("MEMPOOL: bad trailer cookie\n");
1129 RTE_SET_USED(obj_table_const);
1135 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1137 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1138 void *obj, __rte_unused unsigned idx)
1140 __mempool_check_cookies(mp, &obj, 1, 2);
1144 mempool_audit_cookies(struct rte_mempool *mp)
1148 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1149 if (num != mp->size) {
1150 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1151 "iterated only over %u elements\n",
1156 #define mempool_audit_cookies(mp) do {} while(0)
1159 #ifndef __INTEL_COMPILER
1160 #pragma GCC diagnostic error "-Wcast-qual"
1163 /* check cookies before and after objects */
1165 mempool_audit_cache(const struct rte_mempool *mp)
1167 /* check cache size consistency */
1170 if (mp->cache_size == 0)
1173 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1174 const struct rte_mempool_cache *cache;
1175 cache = &mp->local_cache[lcore_id];
1176 if (cache->len > cache->flushthresh) {
1177 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1179 rte_panic("MEMPOOL: invalid cache len\n");
1184 /* check the consistency of mempool (size, cookies, ...) */
1186 rte_mempool_audit(struct rte_mempool *mp)
1188 mempool_audit_cache(mp);
1189 mempool_audit_cookies(mp);
1191 /* For case where mempool DEBUG is not set, and cache size is 0 */
1195 /* dump the status of the mempool on the console */
1197 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1199 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1200 struct rte_mempool_debug_stats sum;
1203 struct rte_mempool_memhdr *memhdr;
1204 unsigned common_count;
1205 unsigned cache_count;
1208 RTE_ASSERT(f != NULL);
1209 RTE_ASSERT(mp != NULL);
1211 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1212 fprintf(f, " flags=%x\n", mp->flags);
1213 fprintf(f, " pool=%p\n", mp->pool_data);
1214 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1215 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1216 fprintf(f, " size=%"PRIu32"\n", mp->size);
1217 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1218 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1219 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1220 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1221 fprintf(f, " total_obj_size=%"PRIu32"\n",
1222 mp->header_size + mp->elt_size + mp->trailer_size);
1224 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1226 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1227 mem_len += memhdr->len;
1229 fprintf(f, " avg bytes/object=%#Lf\n",
1230 (long double)mem_len / mp->size);
1233 cache_count = rte_mempool_dump_cache(f, mp);
1234 common_count = rte_mempool_ops_get_count(mp);
1235 if ((cache_count + common_count) > mp->size)
1236 common_count = mp->size - cache_count;
1237 fprintf(f, " common_pool_count=%u\n", common_count);
1239 /* sum and dump statistics */
1240 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1241 memset(&sum, 0, sizeof(sum));
1242 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1243 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1244 sum.put_objs += mp->stats[lcore_id].put_objs;
1245 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1246 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1247 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1248 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1250 fprintf(f, " stats:\n");
1251 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1252 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1253 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1254 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1255 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1256 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1258 fprintf(f, " no statistics available\n");
1261 rte_mempool_audit(mp);
1264 /* dump the status of all mempools on the console */
1266 rte_mempool_list_dump(FILE *f)
1268 struct rte_mempool *mp = NULL;
1269 struct rte_tailq_entry *te;
1270 struct rte_mempool_list *mempool_list;
1272 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1274 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1276 TAILQ_FOREACH(te, mempool_list, next) {
1277 mp = (struct rte_mempool *) te->data;
1278 rte_mempool_dump(f, mp);
1281 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1284 /* search a mempool from its name */
1285 struct rte_mempool *
1286 rte_mempool_lookup(const char *name)
1288 struct rte_mempool *mp = NULL;
1289 struct rte_tailq_entry *te;
1290 struct rte_mempool_list *mempool_list;
1292 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1294 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1296 TAILQ_FOREACH(te, mempool_list, next) {
1297 mp = (struct rte_mempool *) te->data;
1298 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1302 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1312 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1315 struct rte_tailq_entry *te = NULL;
1316 struct rte_mempool_list *mempool_list;
1319 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1321 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1323 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1324 (*func)((struct rte_mempool *) te->data, arg);
1327 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);