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 * Internal function to calculate required memory chunk size shared
231 * by default implementation of the corresponding callback and
232 * deprecated external function.
235 rte_mempool_calc_mem_size_helper(uint32_t elt_num, size_t total_elt_sz,
238 size_t obj_per_page, pg_num, pg_sz;
240 if (total_elt_sz == 0)
244 return total_elt_sz * elt_num;
246 pg_sz = (size_t)1 << pg_shift;
247 obj_per_page = pg_sz / total_elt_sz;
248 if (obj_per_page == 0)
249 return RTE_ALIGN_CEIL(total_elt_sz, pg_sz) * elt_num;
251 pg_num = (elt_num + obj_per_page - 1) / obj_per_page;
252 return pg_num << pg_shift;
256 * Calculate maximum amount of memory required to store given number of objects.
259 rte_mempool_xmem_size(uint32_t elt_num, size_t total_elt_sz, uint32_t pg_shift,
260 __rte_unused unsigned int flags)
262 return rte_mempool_calc_mem_size_helper(elt_num, total_elt_sz,
267 * Calculate how much memory would be actually required with the
268 * given memory footprint to store required number of elements.
271 rte_mempool_xmem_usage(__rte_unused void *vaddr, uint32_t elt_num,
272 size_t total_elt_sz, const rte_iova_t iova[], uint32_t pg_num,
273 uint32_t pg_shift, __rte_unused unsigned int flags)
275 uint32_t elt_cnt = 0;
276 rte_iova_t start, end;
278 size_t pg_sz = (size_t)1 << pg_shift;
280 /* if iova is NULL, assume contiguous memory */
283 end = pg_sz * pg_num;
287 end = iova[0] + pg_sz;
290 while (elt_cnt < elt_num) {
292 if (end - start >= total_elt_sz) {
293 /* enough contiguous memory, add an object */
294 start += total_elt_sz;
296 } else if (iova_idx < pg_num) {
297 /* no room to store one obj, add a page */
298 if (end == iova[iova_idx]) {
301 start = iova[iova_idx];
302 end = iova[iova_idx] + pg_sz;
307 /* no more page, return how many elements fit */
308 return -(size_t)elt_cnt;
312 return (size_t)iova_idx << pg_shift;
315 /* free a memchunk allocated with rte_memzone_reserve() */
317 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
320 const struct rte_memzone *mz = opaque;
321 rte_memzone_free(mz);
324 /* Free memory chunks used by a mempool. Objects must be in pool */
326 rte_mempool_free_memchunks(struct rte_mempool *mp)
328 struct rte_mempool_memhdr *memhdr;
331 while (!STAILQ_EMPTY(&mp->elt_list)) {
332 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
334 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
335 mp->populated_size--;
338 while (!STAILQ_EMPTY(&mp->mem_list)) {
339 memhdr = STAILQ_FIRST(&mp->mem_list);
340 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
341 if (memhdr->free_cb != NULL)
342 memhdr->free_cb(memhdr, memhdr->opaque);
349 mempool_ops_alloc_once(struct rte_mempool *mp)
353 /* create the internal ring if not already done */
354 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
355 ret = rte_mempool_ops_alloc(mp);
358 mp->flags |= MEMPOOL_F_POOL_CREATED;
363 /* Add objects in the pool, using a physically contiguous memory
364 * zone. Return the number of objects added, or a negative value
368 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
369 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
374 struct rte_mempool_memhdr *memhdr;
377 ret = mempool_ops_alloc_once(mp);
381 /* Notify memory area to mempool */
382 ret = rte_mempool_ops_register_memory_area(mp, vaddr, iova, len);
383 if (ret != -ENOTSUP && ret < 0)
386 /* mempool is already populated */
387 if (mp->populated_size >= mp->size)
390 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
395 memhdr->addr = vaddr;
398 memhdr->free_cb = free_cb;
399 memhdr->opaque = opaque;
401 if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
402 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
404 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
411 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
413 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
414 len - off, mempool_add_elem, NULL);
416 /* not enough room to store one object */
422 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
432 rte_mempool_populate_phys(struct rte_mempool *mp, char *vaddr,
433 phys_addr_t paddr, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
436 return rte_mempool_populate_iova(mp, vaddr, paddr, len, free_cb, opaque);
439 /* Add objects in the pool, using a table of physical pages. Return the
440 * number of objects added, or a negative value on error.
443 rte_mempool_populate_iova_tab(struct rte_mempool *mp, char *vaddr,
444 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift,
445 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
449 size_t pg_sz = (size_t)1 << pg_shift;
451 /* mempool must not be populated */
452 if (mp->nb_mem_chunks != 0)
455 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
456 return rte_mempool_populate_iova(mp, vaddr, RTE_BAD_IOVA,
457 pg_num * pg_sz, free_cb, opaque);
459 for (i = 0; i < pg_num && mp->populated_size < mp->size; i += n) {
461 /* populate with the largest group of contiguous pages */
462 for (n = 1; (i + n) < pg_num &&
463 iova[i + n - 1] + pg_sz == iova[i + n]; n++)
466 ret = rte_mempool_populate_iova(mp, vaddr + i * pg_sz,
467 iova[i], n * pg_sz, free_cb, opaque);
469 rte_mempool_free_memchunks(mp);
472 /* no need to call the free callback for next chunks */
480 rte_mempool_populate_phys_tab(struct rte_mempool *mp, char *vaddr,
481 const phys_addr_t paddr[], uint32_t pg_num, uint32_t pg_shift,
482 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
484 return rte_mempool_populate_iova_tab(mp, vaddr, paddr, pg_num, pg_shift,
488 /* Populate the mempool with a virtual area. Return the number of
489 * objects added, or a negative value on error.
492 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
493 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
497 size_t off, phys_len;
500 /* mempool must not be populated */
501 if (mp->nb_mem_chunks != 0)
503 /* address and len must be page-aligned */
504 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
506 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
509 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
510 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
511 len, free_cb, opaque);
513 for (off = 0; off + pg_sz <= len &&
514 mp->populated_size < mp->size; off += phys_len) {
516 iova = rte_mem_virt2iova(addr + off);
518 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
523 /* populate with the largest group of contiguous pages */
524 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
527 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
529 if (iova_tmp != iova + phys_len)
533 ret = rte_mempool_populate_iova(mp, addr + off, iova,
534 phys_len, free_cb, opaque);
537 /* no need to call the free callback for next chunks */
545 rte_mempool_free_memchunks(mp);
549 /* Default function to populate the mempool: allocate memory in memzones,
550 * and populate them. Return the number of objects added, or a negative
554 rte_mempool_populate_default(struct rte_mempool *mp)
556 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
557 char mz_name[RTE_MEMZONE_NAMESIZE];
558 const struct rte_memzone *mz;
560 size_t align, pg_sz, pg_shift;
564 bool no_contig, try_contig, no_pageshift;
566 ret = mempool_ops_alloc_once(mp);
570 /* mempool must not be populated */
571 if (mp->nb_mem_chunks != 0)
574 no_contig = mp->flags & MEMPOOL_F_NO_IOVA_CONTIG;
577 * the following section calculates page shift and page size values.
579 * these values impact the result of calc_mem_size operation, which
580 * returns the amount of memory that should be allocated to store the
581 * desired number of objects. when not zero, it allocates more memory
582 * for the padding between objects, to ensure that an object does not
583 * cross a page boundary. in other words, page size/shift are to be set
584 * to zero if mempool elements won't care about page boundaries.
585 * there are several considerations for page size and page shift here.
587 * if we don't need our mempools to have physically contiguous objects,
588 * then just set page shift and page size to 0, because the user has
589 * indicated that there's no need to care about anything.
591 * if we do need contiguous objects, there is also an option to reserve
592 * the entire mempool memory as one contiguous block of memory, in
593 * which case the page shift and alignment wouldn't matter as well.
595 * if we require contiguous objects, but not necessarily the entire
596 * mempool reserved space to be contiguous, then there are two options.
598 * if our IO addresses are virtual, not actual physical (IOVA as VA
599 * case), then no page shift needed - our memory allocation will give us
600 * contiguous IO memory as far as the hardware is concerned, so
601 * act as if we're getting contiguous memory.
603 * if our IO addresses are physical, we may get memory from bigger
604 * pages, or we might get memory from smaller pages, and how much of it
605 * we require depends on whether we want bigger or smaller pages.
606 * However, requesting each and every memory size is too much work, so
607 * what we'll do instead is walk through the page sizes available, pick
608 * the smallest one and set up page shift to match that one. We will be
609 * wasting some space this way, but it's much nicer than looping around
610 * trying to reserve each and every page size.
612 * However, since size calculation will produce page-aligned sizes, it
613 * makes sense to first try and see if we can reserve the entire memzone
614 * in one contiguous chunk as well (otherwise we might end up wasting a
615 * 1G page on a 10MB memzone). If we fail to get enough contiguous
616 * memory, then we'll go and reserve space page-by-page.
618 no_pageshift = no_contig || rte_eal_iova_mode() == RTE_IOVA_VA;
619 try_contig = !no_contig && !no_pageshift && rte_eal_has_hugepages();
624 } else if (try_contig) {
625 pg_sz = get_min_page_size();
626 pg_shift = rte_bsf32(pg_sz);
628 pg_sz = getpagesize();
629 pg_shift = rte_bsf32(pg_sz);
632 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
633 size_t min_chunk_size;
636 if (try_contig || no_pageshift)
637 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
638 0, &min_chunk_size, &align);
640 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
641 pg_shift, &min_chunk_size, &align);
648 ret = snprintf(mz_name, sizeof(mz_name),
649 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
650 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
657 /* if we're trying to reserve contiguous memory, add appropriate
661 flags |= RTE_MEMZONE_IOVA_CONTIG;
663 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
664 mp->socket_id, flags, align);
666 /* if we were trying to allocate contiguous memory, failed and
667 * minimum required contiguous chunk fits minimum page, adjust
668 * memzone size to the page size, and try again.
670 if (mz == NULL && try_contig && min_chunk_size <= pg_sz) {
672 flags &= ~RTE_MEMZONE_IOVA_CONTIG;
674 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
675 pg_shift, &min_chunk_size, &align);
681 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
682 mp->socket_id, flags, align);
684 /* don't try reserving with 0 size if we were asked to reserve
685 * IOVA-contiguous memory.
687 if (min_chunk_size < (size_t)mem_size && mz == NULL) {
688 /* not enough memory, retry with the biggest zone we
691 mz = rte_memzone_reserve_aligned(mz_name, 0,
692 mp->socket_id, flags, align);
699 if (mz->len < min_chunk_size) {
700 rte_memzone_free(mz);
710 if (no_pageshift || try_contig)
711 ret = rte_mempool_populate_iova(mp, mz->addr,
713 rte_mempool_memchunk_mz_free,
714 (void *)(uintptr_t)mz);
716 ret = rte_mempool_populate_virt(mp, mz->addr,
718 rte_mempool_memchunk_mz_free,
719 (void *)(uintptr_t)mz);
721 rte_memzone_free(mz);
729 rte_mempool_free_memchunks(mp);
733 /* return the memory size required for mempool objects in anonymous mem */
735 get_anon_size(const struct rte_mempool *mp)
738 size_t pg_sz, pg_shift;
739 size_t min_chunk_size;
742 pg_sz = getpagesize();
743 pg_shift = rte_bsf32(pg_sz);
744 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
745 &min_chunk_size, &align);
750 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
752 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
758 * Calculate size since memhdr->len has contiguous chunk length
759 * which may be smaller if anon map is split into many contiguous
760 * chunks. Result must be the same as we calculated on populate.
762 size = get_anon_size(memhdr->mp);
766 munmap(opaque, size);
769 /* populate the mempool with an anonymous mapping */
771 rte_mempool_populate_anon(struct rte_mempool *mp)
777 /* mempool is already populated, error */
778 if (!STAILQ_EMPTY(&mp->mem_list)) {
783 ret = mempool_ops_alloc_once(mp);
787 size = get_anon_size(mp);
793 /* get chunk of virtually continuous memory */
794 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
795 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
796 if (addr == MAP_FAILED) {
800 /* can't use MMAP_LOCKED, it does not exist on BSD */
801 if (mlock(addr, size) < 0) {
807 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
808 rte_mempool_memchunk_anon_free, addr);
812 return mp->populated_size;
815 rte_mempool_free_memchunks(mp);
821 rte_mempool_free(struct rte_mempool *mp)
823 struct rte_mempool_list *mempool_list = NULL;
824 struct rte_tailq_entry *te;
829 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
830 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
831 /* find out tailq entry */
832 TAILQ_FOREACH(te, mempool_list, next) {
833 if (te->data == (void *)mp)
838 TAILQ_REMOVE(mempool_list, te, next);
841 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
843 rte_mempool_free_memchunks(mp);
844 rte_mempool_ops_free(mp);
845 rte_memzone_free(mp->mz);
849 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
852 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
857 * Create and initialize a cache for objects that are retrieved from and
858 * returned to an underlying mempool. This structure is identical to the
859 * local_cache[lcore_id] pointed to by the mempool structure.
861 struct rte_mempool_cache *
862 rte_mempool_cache_create(uint32_t size, int socket_id)
864 struct rte_mempool_cache *cache;
866 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
871 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
872 RTE_CACHE_LINE_SIZE, socket_id);
874 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
879 mempool_cache_init(cache, size);
885 * Free a cache. It's the responsibility of the user to make sure that any
886 * remaining objects in the cache are flushed to the corresponding
890 rte_mempool_cache_free(struct rte_mempool_cache *cache)
895 /* create an empty mempool */
897 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
898 unsigned cache_size, unsigned private_data_size,
899 int socket_id, unsigned flags)
901 char mz_name[RTE_MEMZONE_NAMESIZE];
902 struct rte_mempool_list *mempool_list;
903 struct rte_mempool *mp = NULL;
904 struct rte_tailq_entry *te = NULL;
905 const struct rte_memzone *mz = NULL;
907 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
908 struct rte_mempool_objsz objsz;
912 /* compilation-time checks */
913 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
914 RTE_CACHE_LINE_MASK) != 0);
915 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
916 RTE_CACHE_LINE_MASK) != 0);
917 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
918 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
919 RTE_CACHE_LINE_MASK) != 0);
920 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
921 RTE_CACHE_LINE_MASK) != 0);
924 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
926 /* asked cache too big */
927 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
928 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
933 /* "no cache align" imply "no spread" */
934 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
935 flags |= MEMPOOL_F_NO_SPREAD;
937 /* calculate mempool object sizes. */
938 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
943 rte_rwlock_write_lock(RTE_EAL_MEMPOOL_RWLOCK);
946 * reserve a memory zone for this mempool: private data is
949 private_data_size = (private_data_size +
950 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
953 /* try to allocate tailq entry */
954 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
956 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
960 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
961 mempool_size += private_data_size;
962 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
964 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
965 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
966 rte_errno = ENAMETOOLONG;
970 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
974 /* init the mempool structure */
976 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
977 ret = snprintf(mp->name, sizeof(mp->name), "%s", name);
978 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
979 rte_errno = ENAMETOOLONG;
985 mp->socket_id = socket_id;
986 mp->elt_size = objsz.elt_size;
987 mp->header_size = objsz.header_size;
988 mp->trailer_size = objsz.trailer_size;
989 /* Size of default caches, zero means disabled. */
990 mp->cache_size = cache_size;
991 mp->private_data_size = private_data_size;
992 STAILQ_INIT(&mp->elt_list);
993 STAILQ_INIT(&mp->mem_list);
996 * local_cache pointer is set even if cache_size is zero.
997 * The local_cache points to just past the elt_pa[] array.
999 mp->local_cache = (struct rte_mempool_cache *)
1000 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
1002 /* Init all default caches. */
1003 if (cache_size != 0) {
1004 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1005 mempool_cache_init(&mp->local_cache[lcore_id],
1011 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
1012 TAILQ_INSERT_TAIL(mempool_list, te, next);
1013 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
1014 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1019 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1021 rte_mempool_free(mp);
1025 /* create the mempool */
1026 struct rte_mempool *
1027 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
1028 unsigned cache_size, unsigned private_data_size,
1029 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1030 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1031 int socket_id, unsigned flags)
1034 struct rte_mempool *mp;
1036 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1037 private_data_size, socket_id, flags);
1042 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
1043 * set the correct index into the table of ops structs.
1045 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
1046 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
1047 else if (flags & MEMPOOL_F_SP_PUT)
1048 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
1049 else if (flags & MEMPOOL_F_SC_GET)
1050 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
1052 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
1057 /* call the mempool priv initializer */
1059 mp_init(mp, mp_init_arg);
1061 if (rte_mempool_populate_default(mp) < 0)
1064 /* call the object initializers */
1066 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1071 rte_mempool_free(mp);
1076 * Create the mempool over already allocated chunk of memory.
1077 * That external memory buffer can consists of physically disjoint pages.
1078 * Setting vaddr to NULL, makes mempool to fallback to rte_mempool_create()
1081 struct rte_mempool *
1082 rte_mempool_xmem_create(const char *name, unsigned n, unsigned elt_size,
1083 unsigned cache_size, unsigned private_data_size,
1084 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1085 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1086 int socket_id, unsigned flags, void *vaddr,
1087 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift)
1089 struct rte_mempool *mp = NULL;
1092 /* no virtual address supplied, use rte_mempool_create() */
1094 return rte_mempool_create(name, n, elt_size, cache_size,
1095 private_data_size, mp_init, mp_init_arg,
1096 obj_init, obj_init_arg, socket_id, flags);
1098 /* check that we have both VA and PA */
1104 /* Check that pg_shift parameter is valid. */
1105 if (pg_shift > MEMPOOL_PG_SHIFT_MAX) {
1110 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1111 private_data_size, socket_id, flags);
1115 /* call the mempool priv initializer */
1117 mp_init(mp, mp_init_arg);
1119 ret = rte_mempool_populate_iova_tab(mp, vaddr, iova, pg_num, pg_shift,
1121 if (ret < 0 || ret != (int)mp->size)
1124 /* call the object initializers */
1126 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1131 rte_mempool_free(mp);
1135 /* Return the number of entries in the mempool */
1137 rte_mempool_avail_count(const struct rte_mempool *mp)
1142 count = rte_mempool_ops_get_count(mp);
1144 if (mp->cache_size == 0)
1147 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1148 count += mp->local_cache[lcore_id].len;
1151 * due to race condition (access to len is not locked), the
1152 * total can be greater than size... so fix the result
1154 if (count > mp->size)
1159 /* return the number of entries allocated from the mempool */
1161 rte_mempool_in_use_count(const struct rte_mempool *mp)
1163 return mp->size - rte_mempool_avail_count(mp);
1166 /* dump the cache status */
1168 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1172 unsigned cache_count;
1174 fprintf(f, " internal cache infos:\n");
1175 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1177 if (mp->cache_size == 0)
1180 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1181 cache_count = mp->local_cache[lcore_id].len;
1182 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1183 lcore_id, cache_count);
1184 count += cache_count;
1186 fprintf(f, " total_cache_count=%u\n", count);
1190 #ifndef __INTEL_COMPILER
1191 #pragma GCC diagnostic ignored "-Wcast-qual"
1194 /* check and update cookies or panic (internal) */
1195 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1196 void * const *obj_table_const, unsigned n, int free)
1198 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1199 struct rte_mempool_objhdr *hdr;
1200 struct rte_mempool_objtlr *tlr;
1206 /* Force to drop the "const" attribute. This is done only when
1207 * DEBUG is enabled */
1208 tmp = (void *) obj_table_const;
1214 if (rte_mempool_from_obj(obj) != mp)
1215 rte_panic("MEMPOOL: object is owned by another "
1218 hdr = __mempool_get_header(obj);
1219 cookie = hdr->cookie;
1222 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1223 RTE_LOG(CRIT, MEMPOOL,
1224 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1225 obj, (const void *) mp, cookie);
1226 rte_panic("MEMPOOL: bad header cookie (put)\n");
1228 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1229 } else if (free == 1) {
1230 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1231 RTE_LOG(CRIT, MEMPOOL,
1232 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1233 obj, (const void *) mp, cookie);
1234 rte_panic("MEMPOOL: bad header cookie (get)\n");
1236 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1237 } else if (free == 2) {
1238 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1239 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1240 RTE_LOG(CRIT, MEMPOOL,
1241 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1242 obj, (const void *) mp, cookie);
1243 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1246 tlr = __mempool_get_trailer(obj);
1247 cookie = tlr->cookie;
1248 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1249 RTE_LOG(CRIT, MEMPOOL,
1250 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1251 obj, (const void *) mp, cookie);
1252 rte_panic("MEMPOOL: bad trailer cookie\n");
1257 RTE_SET_USED(obj_table_const);
1263 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1265 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1266 void *obj, __rte_unused unsigned idx)
1268 __mempool_check_cookies(mp, &obj, 1, 2);
1272 mempool_audit_cookies(struct rte_mempool *mp)
1276 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1277 if (num != mp->size) {
1278 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1279 "iterated only over %u elements\n",
1284 #define mempool_audit_cookies(mp) do {} while(0)
1287 #ifndef __INTEL_COMPILER
1288 #pragma GCC diagnostic error "-Wcast-qual"
1291 /* check cookies before and after objects */
1293 mempool_audit_cache(const struct rte_mempool *mp)
1295 /* check cache size consistency */
1298 if (mp->cache_size == 0)
1301 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1302 const struct rte_mempool_cache *cache;
1303 cache = &mp->local_cache[lcore_id];
1304 if (cache->len > cache->flushthresh) {
1305 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1307 rte_panic("MEMPOOL: invalid cache len\n");
1312 /* check the consistency of mempool (size, cookies, ...) */
1314 rte_mempool_audit(struct rte_mempool *mp)
1316 mempool_audit_cache(mp);
1317 mempool_audit_cookies(mp);
1319 /* For case where mempool DEBUG is not set, and cache size is 0 */
1323 /* dump the status of the mempool on the console */
1325 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1327 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1328 struct rte_mempool_debug_stats sum;
1331 struct rte_mempool_memhdr *memhdr;
1332 unsigned common_count;
1333 unsigned cache_count;
1336 RTE_ASSERT(f != NULL);
1337 RTE_ASSERT(mp != NULL);
1339 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1340 fprintf(f, " flags=%x\n", mp->flags);
1341 fprintf(f, " pool=%p\n", mp->pool_data);
1342 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1343 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1344 fprintf(f, " size=%"PRIu32"\n", mp->size);
1345 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1346 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1347 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1348 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1349 fprintf(f, " total_obj_size=%"PRIu32"\n",
1350 mp->header_size + mp->elt_size + mp->trailer_size);
1352 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1354 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1355 mem_len += memhdr->len;
1357 fprintf(f, " avg bytes/object=%#Lf\n",
1358 (long double)mem_len / mp->size);
1361 cache_count = rte_mempool_dump_cache(f, mp);
1362 common_count = rte_mempool_ops_get_count(mp);
1363 if ((cache_count + common_count) > mp->size)
1364 common_count = mp->size - cache_count;
1365 fprintf(f, " common_pool_count=%u\n", common_count);
1367 /* sum and dump statistics */
1368 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1369 memset(&sum, 0, sizeof(sum));
1370 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1371 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1372 sum.put_objs += mp->stats[lcore_id].put_objs;
1373 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1374 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1375 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1376 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1378 fprintf(f, " stats:\n");
1379 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1380 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1381 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1382 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1383 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1384 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1386 fprintf(f, " no statistics available\n");
1389 rte_mempool_audit(mp);
1392 /* dump the status of all mempools on the console */
1394 rte_mempool_list_dump(FILE *f)
1396 struct rte_mempool *mp = NULL;
1397 struct rte_tailq_entry *te;
1398 struct rte_mempool_list *mempool_list;
1400 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1402 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1404 TAILQ_FOREACH(te, mempool_list, next) {
1405 mp = (struct rte_mempool *) te->data;
1406 rte_mempool_dump(f, mp);
1409 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1412 /* search a mempool from its name */
1413 struct rte_mempool *
1414 rte_mempool_lookup(const char *name)
1416 struct rte_mempool *mp = NULL;
1417 struct rte_tailq_entry *te;
1418 struct rte_mempool_list *mempool_list;
1420 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1422 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1424 TAILQ_FOREACH(te, mempool_list, next) {
1425 mp = (struct rte_mempool *) te->data;
1426 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1430 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1440 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1443 struct rte_tailq_entry *te = NULL;
1444 struct rte_mempool_list *mempool_list;
1447 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1449 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1451 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1452 (*func)((struct rte_mempool *) te->data, arg);
1455 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);