1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation.
3 * Copyright(c) 2016 6WIND S.A.
14 #include <sys/queue.h>
17 #include <rte_common.h>
19 #include <rte_debug.h>
20 #include <rte_memory.h>
21 #include <rte_memzone.h>
22 #include <rte_malloc.h>
23 #include <rte_atomic.h>
24 #include <rte_launch.h>
26 #include <rte_eal_memconfig.h>
27 #include <rte_per_lcore.h>
28 #include <rte_lcore.h>
29 #include <rte_branch_prediction.h>
30 #include <rte_errno.h>
31 #include <rte_string_fns.h>
32 #include <rte_spinlock.h>
34 #include "rte_mempool.h"
36 TAILQ_HEAD(rte_mempool_list, rte_tailq_entry);
38 static struct rte_tailq_elem rte_mempool_tailq = {
39 .name = "RTE_MEMPOOL",
41 EAL_REGISTER_TAILQ(rte_mempool_tailq)
43 #define CACHE_FLUSHTHRESH_MULTIPLIER 1.5
44 #define CALC_CACHE_FLUSHTHRESH(c) \
45 ((typeof(c))((c) * CACHE_FLUSHTHRESH_MULTIPLIER))
48 * return the greatest common divisor between a and b (fast algorithm)
51 static unsigned get_gcd(unsigned a, unsigned b)
76 * Depending on memory configuration, objects addresses are spread
77 * between channels and ranks in RAM: the pool allocator will add
78 * padding between objects. This function return the new size of the
81 static unsigned optimize_object_size(unsigned obj_size)
83 unsigned nrank, nchan;
84 unsigned new_obj_size;
86 /* get number of channels */
87 nchan = rte_memory_get_nchannel();
91 nrank = rte_memory_get_nrank();
95 /* process new object size */
96 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
97 while (get_gcd(new_obj_size, nrank * nchan) != 1)
99 return new_obj_size * RTE_MEMPOOL_ALIGN;
103 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
107 if (msl->page_sz < *min)
114 get_min_page_size(void)
116 size_t min_pagesz = SIZE_MAX;
118 rte_memseg_list_walk(find_min_pagesz, &min_pagesz);
120 return min_pagesz == SIZE_MAX ? (size_t) getpagesize() : min_pagesz;
125 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
126 void *obj, rte_iova_t iova)
128 struct rte_mempool_objhdr *hdr;
129 struct rte_mempool_objtlr *tlr __rte_unused;
131 /* set mempool ptr in header */
132 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
135 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
136 mp->populated_size++;
138 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
139 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
140 tlr = __mempool_get_trailer(obj);
141 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
145 /* call obj_cb() for each mempool element */
147 rte_mempool_obj_iter(struct rte_mempool *mp,
148 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
150 struct rte_mempool_objhdr *hdr;
154 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
155 obj = (char *)hdr + sizeof(*hdr);
156 obj_cb(mp, obj_cb_arg, obj, n);
163 /* call mem_cb() for each mempool memory chunk */
165 rte_mempool_mem_iter(struct rte_mempool *mp,
166 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
168 struct rte_mempool_memhdr *hdr;
171 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
172 mem_cb(mp, mem_cb_arg, hdr, n);
179 /* get the header, trailer and total size of a mempool element. */
181 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
182 struct rte_mempool_objsz *sz)
184 struct rte_mempool_objsz lsz;
186 sz = (sz != NULL) ? sz : &lsz;
188 sz->header_size = sizeof(struct rte_mempool_objhdr);
189 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
190 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
193 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
194 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
196 sz->trailer_size = 0;
199 /* element size is 8 bytes-aligned at least */
200 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
202 /* expand trailer to next cache line */
203 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
204 sz->total_size = sz->header_size + sz->elt_size +
206 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
207 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
208 RTE_MEMPOOL_ALIGN_MASK);
212 * increase trailer to add padding between objects in order to
213 * spread them across memory channels/ranks
215 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
217 new_size = optimize_object_size(sz->header_size + sz->elt_size +
219 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
222 /* this is the size of an object, including header and trailer */
223 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
225 return sz->total_size;
230 * Calculate maximum amount of memory required to store given number of objects.
233 rte_mempool_xmem_size(uint32_t elt_num, size_t total_elt_sz, uint32_t pg_shift,
234 __rte_unused unsigned int flags)
236 size_t obj_per_page, pg_num, pg_sz;
238 if (total_elt_sz == 0)
242 return total_elt_sz * elt_num;
244 pg_sz = (size_t)1 << pg_shift;
245 obj_per_page = pg_sz / total_elt_sz;
246 if (obj_per_page == 0)
247 return RTE_ALIGN_CEIL(total_elt_sz, pg_sz) * elt_num;
249 pg_num = (elt_num + obj_per_page - 1) / obj_per_page;
250 return pg_num << pg_shift;
254 * Calculate how much memory would be actually required with the
255 * given memory footprint to store required number of elements.
258 rte_mempool_xmem_usage(__rte_unused void *vaddr, uint32_t elt_num,
259 size_t total_elt_sz, const rte_iova_t iova[], uint32_t pg_num,
260 uint32_t pg_shift, __rte_unused unsigned int flags)
262 uint32_t elt_cnt = 0;
263 rte_iova_t start, end;
265 size_t pg_sz = (size_t)1 << pg_shift;
267 /* if iova is NULL, assume contiguous memory */
270 end = pg_sz * pg_num;
274 end = iova[0] + pg_sz;
277 while (elt_cnt < elt_num) {
279 if (end - start >= total_elt_sz) {
280 /* enough contiguous memory, add an object */
281 start += total_elt_sz;
283 } else if (iova_idx < pg_num) {
284 /* no room to store one obj, add a page */
285 if (end == iova[iova_idx]) {
288 start = iova[iova_idx];
289 end = iova[iova_idx] + pg_sz;
294 /* no more page, return how many elements fit */
295 return -(size_t)elt_cnt;
299 return (size_t)iova_idx << pg_shift;
302 /* free a memchunk allocated with rte_memzone_reserve() */
304 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
307 const struct rte_memzone *mz = opaque;
308 rte_memzone_free(mz);
311 /* Free memory chunks used by a mempool. Objects must be in pool */
313 rte_mempool_free_memchunks(struct rte_mempool *mp)
315 struct rte_mempool_memhdr *memhdr;
318 while (!STAILQ_EMPTY(&mp->elt_list)) {
319 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
321 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
322 mp->populated_size--;
325 while (!STAILQ_EMPTY(&mp->mem_list)) {
326 memhdr = STAILQ_FIRST(&mp->mem_list);
327 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
328 if (memhdr->free_cb != NULL)
329 memhdr->free_cb(memhdr, memhdr->opaque);
336 mempool_ops_alloc_once(struct rte_mempool *mp)
340 /* create the internal ring if not already done */
341 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
342 ret = rte_mempool_ops_alloc(mp);
345 mp->flags |= MEMPOOL_F_POOL_CREATED;
350 /* Add objects in the pool, using a physically contiguous memory
351 * zone. Return the number of objects added, or a negative value
355 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
356 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
361 struct rte_mempool_memhdr *memhdr;
364 ret = mempool_ops_alloc_once(mp);
368 /* Notify memory area to mempool */
369 ret = rte_mempool_ops_register_memory_area(mp, vaddr, iova, len);
370 if (ret != -ENOTSUP && ret < 0)
373 /* mempool is already populated */
374 if (mp->populated_size >= mp->size)
377 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
382 memhdr->addr = vaddr;
385 memhdr->free_cb = free_cb;
386 memhdr->opaque = opaque;
388 if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
389 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
391 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
398 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
400 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
401 len - off, mempool_add_elem, NULL);
403 /* not enough room to store one object */
409 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
419 rte_mempool_populate_phys(struct rte_mempool *mp, char *vaddr,
420 phys_addr_t paddr, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
423 return rte_mempool_populate_iova(mp, vaddr, paddr, len, free_cb, opaque);
426 /* Add objects in the pool, using a table of physical pages. Return the
427 * number of objects added, or a negative value on error.
430 rte_mempool_populate_iova_tab(struct rte_mempool *mp, char *vaddr,
431 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift,
432 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
436 size_t pg_sz = (size_t)1 << pg_shift;
438 /* mempool must not be populated */
439 if (mp->nb_mem_chunks != 0)
442 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
443 return rte_mempool_populate_iova(mp, vaddr, RTE_BAD_IOVA,
444 pg_num * pg_sz, free_cb, opaque);
446 for (i = 0; i < pg_num && mp->populated_size < mp->size; i += n) {
448 /* populate with the largest group of contiguous pages */
449 for (n = 1; (i + n) < pg_num &&
450 iova[i + n - 1] + pg_sz == iova[i + n]; n++)
453 ret = rte_mempool_populate_iova(mp, vaddr + i * pg_sz,
454 iova[i], n * pg_sz, free_cb, opaque);
456 rte_mempool_free_memchunks(mp);
459 /* no need to call the free callback for next chunks */
467 rte_mempool_populate_phys_tab(struct rte_mempool *mp, char *vaddr,
468 const phys_addr_t paddr[], uint32_t pg_num, uint32_t pg_shift,
469 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
471 return rte_mempool_populate_iova_tab(mp, vaddr, paddr, pg_num, pg_shift,
475 /* Populate the mempool with a virtual area. Return the number of
476 * objects added, or a negative value on error.
479 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
480 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
484 size_t off, phys_len;
487 /* mempool must not be populated */
488 if (mp->nb_mem_chunks != 0)
490 /* address and len must be page-aligned */
491 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
493 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
496 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
497 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
498 len, free_cb, opaque);
500 for (off = 0; off + pg_sz <= len &&
501 mp->populated_size < mp->size; off += phys_len) {
503 iova = rte_mem_virt2iova(addr + off);
505 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
510 /* populate with the largest group of contiguous pages */
511 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
514 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
516 if (iova_tmp != iova + phys_len)
520 ret = rte_mempool_populate_iova(mp, addr + off, iova,
521 phys_len, free_cb, opaque);
524 /* no need to call the free callback for next chunks */
532 rte_mempool_free_memchunks(mp);
536 /* Default function to populate the mempool: allocate memory in memzones,
537 * and populate them. Return the number of objects added, or a negative
541 rte_mempool_populate_default(struct rte_mempool *mp)
543 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
544 char mz_name[RTE_MEMZONE_NAMESIZE];
545 const struct rte_memzone *mz;
547 size_t align, pg_sz, pg_shift;
551 bool no_contig, try_contig, no_pageshift;
553 ret = mempool_ops_alloc_once(mp);
557 /* mempool must not be populated */
558 if (mp->nb_mem_chunks != 0)
561 no_contig = mp->flags & MEMPOOL_F_NO_IOVA_CONTIG;
564 * the following section calculates page shift and page size values.
566 * these values impact the result of calc_mem_size operation, which
567 * returns the amount of memory that should be allocated to store the
568 * desired number of objects. when not zero, it allocates more memory
569 * for the padding between objects, to ensure that an object does not
570 * cross a page boundary. in other words, page size/shift are to be set
571 * to zero if mempool elements won't care about page boundaries.
572 * there are several considerations for page size and page shift here.
574 * if we don't need our mempools to have physically contiguous objects,
575 * then just set page shift and page size to 0, because the user has
576 * indicated that there's no need to care about anything.
578 * if we do need contiguous objects, there is also an option to reserve
579 * the entire mempool memory as one contiguous block of memory, in
580 * which case the page shift and alignment wouldn't matter as well.
582 * if we require contiguous objects, but not necessarily the entire
583 * mempool reserved space to be contiguous, then there are two options.
585 * if our IO addresses are virtual, not actual physical (IOVA as VA
586 * case), then no page shift needed - our memory allocation will give us
587 * contiguous IO memory as far as the hardware is concerned, so
588 * act as if we're getting contiguous memory.
590 * if our IO addresses are physical, we may get memory from bigger
591 * pages, or we might get memory from smaller pages, and how much of it
592 * we require depends on whether we want bigger or smaller pages.
593 * However, requesting each and every memory size is too much work, so
594 * what we'll do instead is walk through the page sizes available, pick
595 * the smallest one and set up page shift to match that one. We will be
596 * wasting some space this way, but it's much nicer than looping around
597 * trying to reserve each and every page size.
599 * However, since size calculation will produce page-aligned sizes, it
600 * makes sense to first try and see if we can reserve the entire memzone
601 * in one contiguous chunk as well (otherwise we might end up wasting a
602 * 1G page on a 10MB memzone). If we fail to get enough contiguous
603 * memory, then we'll go and reserve space page-by-page.
605 no_pageshift = no_contig || rte_eal_iova_mode() == RTE_IOVA_VA;
606 try_contig = !no_contig && !no_pageshift && rte_eal_has_hugepages();
611 } else if (try_contig) {
612 pg_sz = get_min_page_size();
613 pg_shift = rte_bsf32(pg_sz);
615 pg_sz = getpagesize();
616 pg_shift = rte_bsf32(pg_sz);
619 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
620 size_t min_chunk_size;
623 if (try_contig || no_pageshift)
624 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
625 0, &min_chunk_size, &align);
627 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
628 pg_shift, &min_chunk_size, &align);
635 ret = snprintf(mz_name, sizeof(mz_name),
636 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
637 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
644 /* if we're trying to reserve contiguous memory, add appropriate
648 flags |= RTE_MEMZONE_IOVA_CONTIG;
650 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
651 mp->socket_id, flags, align);
653 /* if we were trying to allocate contiguous memory, failed and
654 * minimum required contiguous chunk fits minimum page, adjust
655 * memzone size to the page size, and try again.
657 if (mz == NULL && try_contig && min_chunk_size <= pg_sz) {
659 flags &= ~RTE_MEMZONE_IOVA_CONTIG;
661 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
662 pg_shift, &min_chunk_size, &align);
668 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
669 mp->socket_id, flags, align);
671 /* don't try reserving with 0 size if we were asked to reserve
672 * IOVA-contiguous memory.
674 if (min_chunk_size < (size_t)mem_size && mz == NULL) {
675 /* not enough memory, retry with the biggest zone we
678 mz = rte_memzone_reserve_aligned(mz_name, 0,
679 mp->socket_id, flags, align);
686 if (mz->len < min_chunk_size) {
687 rte_memzone_free(mz);
697 if (no_pageshift || try_contig)
698 ret = rte_mempool_populate_iova(mp, mz->addr,
700 rte_mempool_memchunk_mz_free,
701 (void *)(uintptr_t)mz);
703 ret = rte_mempool_populate_virt(mp, mz->addr,
705 rte_mempool_memchunk_mz_free,
706 (void *)(uintptr_t)mz);
708 rte_memzone_free(mz);
716 rte_mempool_free_memchunks(mp);
720 /* return the memory size required for mempool objects in anonymous mem */
722 get_anon_size(const struct rte_mempool *mp)
725 size_t pg_sz, pg_shift;
726 size_t min_chunk_size;
729 pg_sz = getpagesize();
730 pg_shift = rte_bsf32(pg_sz);
731 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
732 &min_chunk_size, &align);
737 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
739 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
745 * Calculate size since memhdr->len has contiguous chunk length
746 * which may be smaller if anon map is split into many contiguous
747 * chunks. Result must be the same as we calculated on populate.
749 size = get_anon_size(memhdr->mp);
753 munmap(opaque, size);
756 /* populate the mempool with an anonymous mapping */
758 rte_mempool_populate_anon(struct rte_mempool *mp)
764 /* mempool is already populated, error */
765 if (!STAILQ_EMPTY(&mp->mem_list)) {
770 ret = mempool_ops_alloc_once(mp);
774 size = get_anon_size(mp);
780 /* get chunk of virtually continuous memory */
781 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
782 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
783 if (addr == MAP_FAILED) {
787 /* can't use MMAP_LOCKED, it does not exist on BSD */
788 if (mlock(addr, size) < 0) {
794 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
795 rte_mempool_memchunk_anon_free, addr);
799 return mp->populated_size;
802 rte_mempool_free_memchunks(mp);
808 rte_mempool_free(struct rte_mempool *mp)
810 struct rte_mempool_list *mempool_list = NULL;
811 struct rte_tailq_entry *te;
816 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
817 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
818 /* find out tailq entry */
819 TAILQ_FOREACH(te, mempool_list, next) {
820 if (te->data == (void *)mp)
825 TAILQ_REMOVE(mempool_list, te, next);
828 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
830 rte_mempool_free_memchunks(mp);
831 rte_mempool_ops_free(mp);
832 rte_memzone_free(mp->mz);
836 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
839 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
844 * Create and initialize a cache for objects that are retrieved from and
845 * returned to an underlying mempool. This structure is identical to the
846 * local_cache[lcore_id] pointed to by the mempool structure.
848 struct rte_mempool_cache *
849 rte_mempool_cache_create(uint32_t size, int socket_id)
851 struct rte_mempool_cache *cache;
853 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
858 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
859 RTE_CACHE_LINE_SIZE, socket_id);
861 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
866 mempool_cache_init(cache, size);
872 * Free a cache. It's the responsibility of the user to make sure that any
873 * remaining objects in the cache are flushed to the corresponding
877 rte_mempool_cache_free(struct rte_mempool_cache *cache)
882 /* create an empty mempool */
884 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
885 unsigned cache_size, unsigned private_data_size,
886 int socket_id, unsigned flags)
888 char mz_name[RTE_MEMZONE_NAMESIZE];
889 struct rte_mempool_list *mempool_list;
890 struct rte_mempool *mp = NULL;
891 struct rte_tailq_entry *te = NULL;
892 const struct rte_memzone *mz = NULL;
894 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
895 struct rte_mempool_objsz objsz;
899 /* compilation-time checks */
900 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
901 RTE_CACHE_LINE_MASK) != 0);
902 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
903 RTE_CACHE_LINE_MASK) != 0);
904 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
905 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
906 RTE_CACHE_LINE_MASK) != 0);
907 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
908 RTE_CACHE_LINE_MASK) != 0);
911 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
913 /* asked cache too big */
914 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
915 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
920 /* "no cache align" imply "no spread" */
921 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
922 flags |= MEMPOOL_F_NO_SPREAD;
924 /* calculate mempool object sizes. */
925 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
930 rte_rwlock_write_lock(RTE_EAL_MEMPOOL_RWLOCK);
933 * reserve a memory zone for this mempool: private data is
936 private_data_size = (private_data_size +
937 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
940 /* try to allocate tailq entry */
941 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
943 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
947 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
948 mempool_size += private_data_size;
949 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
951 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
952 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
953 rte_errno = ENAMETOOLONG;
957 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
961 /* init the mempool structure */
963 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
964 ret = snprintf(mp->name, sizeof(mp->name), "%s", name);
965 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
966 rte_errno = ENAMETOOLONG;
972 mp->socket_id = socket_id;
973 mp->elt_size = objsz.elt_size;
974 mp->header_size = objsz.header_size;
975 mp->trailer_size = objsz.trailer_size;
976 /* Size of default caches, zero means disabled. */
977 mp->cache_size = cache_size;
978 mp->private_data_size = private_data_size;
979 STAILQ_INIT(&mp->elt_list);
980 STAILQ_INIT(&mp->mem_list);
983 * local_cache pointer is set even if cache_size is zero.
984 * The local_cache points to just past the elt_pa[] array.
986 mp->local_cache = (struct rte_mempool_cache *)
987 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
989 /* Init all default caches. */
990 if (cache_size != 0) {
991 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
992 mempool_cache_init(&mp->local_cache[lcore_id],
998 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
999 TAILQ_INSERT_TAIL(mempool_list, te, next);
1000 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
1001 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1006 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1008 rte_mempool_free(mp);
1012 /* create the mempool */
1013 struct rte_mempool *
1014 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
1015 unsigned cache_size, unsigned private_data_size,
1016 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1017 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1018 int socket_id, unsigned flags)
1021 struct rte_mempool *mp;
1023 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1024 private_data_size, socket_id, flags);
1029 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
1030 * set the correct index into the table of ops structs.
1032 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
1033 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
1034 else if (flags & MEMPOOL_F_SP_PUT)
1035 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
1036 else if (flags & MEMPOOL_F_SC_GET)
1037 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
1039 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
1044 /* call the mempool priv initializer */
1046 mp_init(mp, mp_init_arg);
1048 if (rte_mempool_populate_default(mp) < 0)
1051 /* call the object initializers */
1053 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1058 rte_mempool_free(mp);
1063 * Create the mempool over already allocated chunk of memory.
1064 * That external memory buffer can consists of physically disjoint pages.
1065 * Setting vaddr to NULL, makes mempool to fallback to rte_mempool_create()
1068 struct rte_mempool *
1069 rte_mempool_xmem_create(const char *name, unsigned n, unsigned elt_size,
1070 unsigned cache_size, unsigned private_data_size,
1071 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1072 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1073 int socket_id, unsigned flags, void *vaddr,
1074 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift)
1076 struct rte_mempool *mp = NULL;
1079 /* no virtual address supplied, use rte_mempool_create() */
1081 return rte_mempool_create(name, n, elt_size, cache_size,
1082 private_data_size, mp_init, mp_init_arg,
1083 obj_init, obj_init_arg, socket_id, flags);
1085 /* check that we have both VA and PA */
1091 /* Check that pg_shift parameter is valid. */
1092 if (pg_shift > MEMPOOL_PG_SHIFT_MAX) {
1097 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1098 private_data_size, socket_id, flags);
1102 /* call the mempool priv initializer */
1104 mp_init(mp, mp_init_arg);
1106 ret = rte_mempool_populate_iova_tab(mp, vaddr, iova, pg_num, pg_shift,
1108 if (ret < 0 || ret != (int)mp->size)
1111 /* call the object initializers */
1113 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1118 rte_mempool_free(mp);
1122 /* Return the number of entries in the mempool */
1124 rte_mempool_avail_count(const struct rte_mempool *mp)
1129 count = rte_mempool_ops_get_count(mp);
1131 if (mp->cache_size == 0)
1134 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1135 count += mp->local_cache[lcore_id].len;
1138 * due to race condition (access to len is not locked), the
1139 * total can be greater than size... so fix the result
1141 if (count > mp->size)
1146 /* return the number of entries allocated from the mempool */
1148 rte_mempool_in_use_count(const struct rte_mempool *mp)
1150 return mp->size - rte_mempool_avail_count(mp);
1153 /* dump the cache status */
1155 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1159 unsigned cache_count;
1161 fprintf(f, " internal cache infos:\n");
1162 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1164 if (mp->cache_size == 0)
1167 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1168 cache_count = mp->local_cache[lcore_id].len;
1169 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1170 lcore_id, cache_count);
1171 count += cache_count;
1173 fprintf(f, " total_cache_count=%u\n", count);
1177 #ifndef __INTEL_COMPILER
1178 #pragma GCC diagnostic ignored "-Wcast-qual"
1181 /* check and update cookies or panic (internal) */
1182 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1183 void * const *obj_table_const, unsigned n, int free)
1185 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1186 struct rte_mempool_objhdr *hdr;
1187 struct rte_mempool_objtlr *tlr;
1193 /* Force to drop the "const" attribute. This is done only when
1194 * DEBUG is enabled */
1195 tmp = (void *) obj_table_const;
1201 if (rte_mempool_from_obj(obj) != mp)
1202 rte_panic("MEMPOOL: object is owned by another "
1205 hdr = __mempool_get_header(obj);
1206 cookie = hdr->cookie;
1209 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1210 RTE_LOG(CRIT, MEMPOOL,
1211 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1212 obj, (const void *) mp, cookie);
1213 rte_panic("MEMPOOL: bad header cookie (put)\n");
1215 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1216 } else if (free == 1) {
1217 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1218 RTE_LOG(CRIT, MEMPOOL,
1219 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1220 obj, (const void *) mp, cookie);
1221 rte_panic("MEMPOOL: bad header cookie (get)\n");
1223 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1224 } else if (free == 2) {
1225 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1226 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1227 RTE_LOG(CRIT, MEMPOOL,
1228 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1229 obj, (const void *) mp, cookie);
1230 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1233 tlr = __mempool_get_trailer(obj);
1234 cookie = tlr->cookie;
1235 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1236 RTE_LOG(CRIT, MEMPOOL,
1237 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1238 obj, (const void *) mp, cookie);
1239 rte_panic("MEMPOOL: bad trailer cookie\n");
1244 RTE_SET_USED(obj_table_const);
1250 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1252 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1253 void *obj, __rte_unused unsigned idx)
1255 __mempool_check_cookies(mp, &obj, 1, 2);
1259 mempool_audit_cookies(struct rte_mempool *mp)
1263 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1264 if (num != mp->size) {
1265 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1266 "iterated only over %u elements\n",
1271 #define mempool_audit_cookies(mp) do {} while(0)
1274 #ifndef __INTEL_COMPILER
1275 #pragma GCC diagnostic error "-Wcast-qual"
1278 /* check cookies before and after objects */
1280 mempool_audit_cache(const struct rte_mempool *mp)
1282 /* check cache size consistency */
1285 if (mp->cache_size == 0)
1288 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1289 const struct rte_mempool_cache *cache;
1290 cache = &mp->local_cache[lcore_id];
1291 if (cache->len > cache->flushthresh) {
1292 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1294 rte_panic("MEMPOOL: invalid cache len\n");
1299 /* check the consistency of mempool (size, cookies, ...) */
1301 rte_mempool_audit(struct rte_mempool *mp)
1303 mempool_audit_cache(mp);
1304 mempool_audit_cookies(mp);
1306 /* For case where mempool DEBUG is not set, and cache size is 0 */
1310 /* dump the status of the mempool on the console */
1312 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1314 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1315 struct rte_mempool_debug_stats sum;
1318 struct rte_mempool_memhdr *memhdr;
1319 unsigned common_count;
1320 unsigned cache_count;
1323 RTE_ASSERT(f != NULL);
1324 RTE_ASSERT(mp != NULL);
1326 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1327 fprintf(f, " flags=%x\n", mp->flags);
1328 fprintf(f, " pool=%p\n", mp->pool_data);
1329 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1330 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1331 fprintf(f, " size=%"PRIu32"\n", mp->size);
1332 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1333 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1334 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1335 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1336 fprintf(f, " total_obj_size=%"PRIu32"\n",
1337 mp->header_size + mp->elt_size + mp->trailer_size);
1339 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1341 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1342 mem_len += memhdr->len;
1344 fprintf(f, " avg bytes/object=%#Lf\n",
1345 (long double)mem_len / mp->size);
1348 cache_count = rte_mempool_dump_cache(f, mp);
1349 common_count = rte_mempool_ops_get_count(mp);
1350 if ((cache_count + common_count) > mp->size)
1351 common_count = mp->size - cache_count;
1352 fprintf(f, " common_pool_count=%u\n", common_count);
1354 /* sum and dump statistics */
1355 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1356 memset(&sum, 0, sizeof(sum));
1357 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1358 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1359 sum.put_objs += mp->stats[lcore_id].put_objs;
1360 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1361 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1362 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1363 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1365 fprintf(f, " stats:\n");
1366 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1367 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1368 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1369 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1370 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1371 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1373 fprintf(f, " no statistics available\n");
1376 rte_mempool_audit(mp);
1379 /* dump the status of all mempools on the console */
1381 rte_mempool_list_dump(FILE *f)
1383 struct rte_mempool *mp = NULL;
1384 struct rte_tailq_entry *te;
1385 struct rte_mempool_list *mempool_list;
1387 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1389 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1391 TAILQ_FOREACH(te, mempool_list, next) {
1392 mp = (struct rte_mempool *) te->data;
1393 rte_mempool_dump(f, mp);
1396 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1399 /* search a mempool from its name */
1400 struct rte_mempool *
1401 rte_mempool_lookup(const char *name)
1403 struct rte_mempool *mp = NULL;
1404 struct rte_tailq_entry *te;
1405 struct rte_mempool_list *mempool_list;
1407 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1409 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1411 TAILQ_FOREACH(te, mempool_list, next) {
1412 mp = (struct rte_mempool *) te->data;
1413 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1417 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1427 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1430 struct rte_tailq_entry *te = NULL;
1431 struct rte_mempool_list *mempool_list;
1434 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1436 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1438 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1439 (*func)((struct rte_mempool *) te->data, arg);
1442 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);