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,
236 size_t obj_per_page, pg_num, pg_sz;
239 mask = MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS | MEMPOOL_F_CAPA_PHYS_CONTIG;
240 if ((flags & mask) == mask)
241 /* alignment need one additional object */
244 if (total_elt_sz == 0)
248 return total_elt_sz * elt_num;
250 pg_sz = (size_t)1 << pg_shift;
251 obj_per_page = pg_sz / total_elt_sz;
252 if (obj_per_page == 0)
253 return RTE_ALIGN_CEIL(total_elt_sz, pg_sz) * elt_num;
255 pg_num = (elt_num + obj_per_page - 1) / obj_per_page;
256 return pg_num << pg_shift;
260 * Calculate how much memory would be actually required with the
261 * given memory footprint to store required number of elements.
264 rte_mempool_xmem_usage(__rte_unused void *vaddr, uint32_t elt_num,
265 size_t total_elt_sz, const rte_iova_t iova[], uint32_t pg_num,
266 uint32_t pg_shift, unsigned int flags)
268 uint32_t elt_cnt = 0;
269 rte_iova_t start, end;
271 size_t pg_sz = (size_t)1 << pg_shift;
274 mask = MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS | MEMPOOL_F_CAPA_PHYS_CONTIG;
275 if ((flags & mask) == mask)
276 /* alignment need one additional object */
279 /* if iova is NULL, assume contiguous memory */
282 end = pg_sz * pg_num;
286 end = iova[0] + pg_sz;
289 while (elt_cnt < elt_num) {
291 if (end - start >= total_elt_sz) {
292 /* enough contiguous memory, add an object */
293 start += total_elt_sz;
295 } else if (iova_idx < pg_num) {
296 /* no room to store one obj, add a page */
297 if (end == iova[iova_idx]) {
300 start = iova[iova_idx];
301 end = iova[iova_idx] + pg_sz;
306 /* no more page, return how many elements fit */
307 return -(size_t)elt_cnt;
311 return (size_t)iova_idx << pg_shift;
314 /* free a memchunk allocated with rte_memzone_reserve() */
316 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
319 const struct rte_memzone *mz = opaque;
320 rte_memzone_free(mz);
323 /* Free memory chunks used by a mempool. Objects must be in pool */
325 rte_mempool_free_memchunks(struct rte_mempool *mp)
327 struct rte_mempool_memhdr *memhdr;
330 while (!STAILQ_EMPTY(&mp->elt_list)) {
331 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
333 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
334 mp->populated_size--;
337 while (!STAILQ_EMPTY(&mp->mem_list)) {
338 memhdr = STAILQ_FIRST(&mp->mem_list);
339 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
340 if (memhdr->free_cb != NULL)
341 memhdr->free_cb(memhdr, memhdr->opaque);
348 mempool_ops_alloc_once(struct rte_mempool *mp)
352 /* create the internal ring if not already done */
353 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
354 ret = rte_mempool_ops_alloc(mp);
357 mp->flags |= MEMPOOL_F_POOL_CREATED;
362 /* Add objects in the pool, using a physically contiguous memory
363 * zone. Return the number of objects added, or a negative value
367 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
368 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
371 unsigned total_elt_sz;
372 unsigned int mp_capa_flags;
375 struct rte_mempool_memhdr *memhdr;
378 ret = mempool_ops_alloc_once(mp);
382 /* Notify memory area to mempool */
383 ret = rte_mempool_ops_register_memory_area(mp, vaddr, iova, len);
384 if (ret != -ENOTSUP && ret < 0)
387 /* mempool is already populated */
388 if (mp->populated_size >= mp->size)
391 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
393 /* Get mempool capabilities */
395 ret = rte_mempool_ops_get_capabilities(mp, &mp_capa_flags);
396 if ((ret < 0) && (ret != -ENOTSUP))
399 /* update mempool capabilities */
400 mp->flags |= mp_capa_flags;
402 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
407 memhdr->addr = vaddr;
410 memhdr->free_cb = free_cb;
411 memhdr->opaque = opaque;
413 if (mp_capa_flags & MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS)
414 /* align object start address to a multiple of total_elt_sz */
415 off = total_elt_sz - ((uintptr_t)vaddr % total_elt_sz);
416 else if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
417 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
419 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
426 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
428 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
429 len - off, mempool_add_elem, NULL);
431 /* not enough room to store one object */
437 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
447 rte_mempool_populate_phys(struct rte_mempool *mp, char *vaddr,
448 phys_addr_t paddr, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
451 return rte_mempool_populate_iova(mp, vaddr, paddr, len, free_cb, opaque);
454 /* Add objects in the pool, using a table of physical pages. Return the
455 * number of objects added, or a negative value on error.
458 rte_mempool_populate_iova_tab(struct rte_mempool *mp, char *vaddr,
459 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift,
460 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
464 size_t pg_sz = (size_t)1 << pg_shift;
466 /* mempool must not be populated */
467 if (mp->nb_mem_chunks != 0)
470 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
471 return rte_mempool_populate_iova(mp, vaddr, RTE_BAD_IOVA,
472 pg_num * pg_sz, free_cb, opaque);
474 for (i = 0; i < pg_num && mp->populated_size < mp->size; i += n) {
476 /* populate with the largest group of contiguous pages */
477 for (n = 1; (i + n) < pg_num &&
478 iova[i + n - 1] + pg_sz == iova[i + n]; n++)
481 ret = rte_mempool_populate_iova(mp, vaddr + i * pg_sz,
482 iova[i], n * pg_sz, free_cb, opaque);
484 rte_mempool_free_memchunks(mp);
487 /* no need to call the free callback for next chunks */
495 rte_mempool_populate_phys_tab(struct rte_mempool *mp, char *vaddr,
496 const phys_addr_t paddr[], uint32_t pg_num, uint32_t pg_shift,
497 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
499 return rte_mempool_populate_iova_tab(mp, vaddr, paddr, pg_num, pg_shift,
503 /* Populate the mempool with a virtual area. Return the number of
504 * objects added, or a negative value on error.
507 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
508 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
512 size_t off, phys_len;
515 /* mempool must not be populated */
516 if (mp->nb_mem_chunks != 0)
518 /* address and len must be page-aligned */
519 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
521 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
524 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
525 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
526 len, free_cb, opaque);
528 for (off = 0; off + pg_sz <= len &&
529 mp->populated_size < mp->size; off += phys_len) {
531 iova = rte_mem_virt2iova(addr + off);
533 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
538 /* populate with the largest group of contiguous pages */
539 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
542 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
544 if (iova_tmp != iova + phys_len)
548 ret = rte_mempool_populate_iova(mp, addr + off, iova,
549 phys_len, free_cb, opaque);
552 /* no need to call the free callback for next chunks */
560 rte_mempool_free_memchunks(mp);
564 /* Default function to populate the mempool: allocate memory in memzones,
565 * and populate them. Return the number of objects added, or a negative
569 rte_mempool_populate_default(struct rte_mempool *mp)
571 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
572 char mz_name[RTE_MEMZONE_NAMESIZE];
573 const struct rte_memzone *mz;
575 size_t align, pg_sz, pg_shift;
579 bool no_contig, try_contig, no_pageshift;
581 ret = mempool_ops_alloc_once(mp);
585 /* mempool must not be populated */
586 if (mp->nb_mem_chunks != 0)
589 no_contig = mp->flags & MEMPOOL_F_NO_IOVA_CONTIG;
592 * the following section calculates page shift and page size values.
594 * these values impact the result of calc_mem_size operation, which
595 * returns the amount of memory that should be allocated to store the
596 * desired number of objects. when not zero, it allocates more memory
597 * for the padding between objects, to ensure that an object does not
598 * cross a page boundary. in other words, page size/shift are to be set
599 * to zero if mempool elements won't care about page boundaries.
600 * there are several considerations for page size and page shift here.
602 * if we don't need our mempools to have physically contiguous objects,
603 * then just set page shift and page size to 0, because the user has
604 * indicated that there's no need to care about anything.
606 * if we do need contiguous objects, there is also an option to reserve
607 * the entire mempool memory as one contiguous block of memory, in
608 * which case the page shift and alignment wouldn't matter as well.
610 * if we require contiguous objects, but not necessarily the entire
611 * mempool reserved space to be contiguous, then there are two options.
613 * if our IO addresses are virtual, not actual physical (IOVA as VA
614 * case), then no page shift needed - our memory allocation will give us
615 * contiguous IO memory as far as the hardware is concerned, so
616 * act as if we're getting contiguous memory.
618 * if our IO addresses are physical, we may get memory from bigger
619 * pages, or we might get memory from smaller pages, and how much of it
620 * we require depends on whether we want bigger or smaller pages.
621 * However, requesting each and every memory size is too much work, so
622 * what we'll do instead is walk through the page sizes available, pick
623 * the smallest one and set up page shift to match that one. We will be
624 * wasting some space this way, but it's much nicer than looping around
625 * trying to reserve each and every page size.
627 * However, since size calculation will produce page-aligned sizes, it
628 * makes sense to first try and see if we can reserve the entire memzone
629 * in one contiguous chunk as well (otherwise we might end up wasting a
630 * 1G page on a 10MB memzone). If we fail to get enough contiguous
631 * memory, then we'll go and reserve space page-by-page.
633 no_pageshift = no_contig || rte_eal_iova_mode() == RTE_IOVA_VA;
634 try_contig = !no_contig && !no_pageshift && rte_eal_has_hugepages();
639 } else if (try_contig) {
640 pg_sz = get_min_page_size();
641 pg_shift = rte_bsf32(pg_sz);
643 pg_sz = getpagesize();
644 pg_shift = rte_bsf32(pg_sz);
647 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
648 size_t min_chunk_size;
651 if (try_contig || no_pageshift)
652 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
653 0, &min_chunk_size, &align);
655 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
656 pg_shift, &min_chunk_size, &align);
663 ret = snprintf(mz_name, sizeof(mz_name),
664 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
665 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
672 /* if we're trying to reserve contiguous memory, add appropriate
676 flags |= RTE_MEMZONE_IOVA_CONTIG;
678 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
679 mp->socket_id, flags, align);
681 /* if we were trying to allocate contiguous memory, failed and
682 * minimum required contiguous chunk fits minimum page, adjust
683 * memzone size to the page size, and try again.
685 if (mz == NULL && try_contig && min_chunk_size <= pg_sz) {
687 flags &= ~RTE_MEMZONE_IOVA_CONTIG;
689 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
690 pg_shift, &min_chunk_size, &align);
696 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
697 mp->socket_id, flags, align);
699 /* don't try reserving with 0 size if we were asked to reserve
700 * IOVA-contiguous memory.
702 if (min_chunk_size < (size_t)mem_size && mz == NULL) {
703 /* not enough memory, retry with the biggest zone we
706 mz = rte_memzone_reserve_aligned(mz_name, 0,
707 mp->socket_id, flags, align);
714 if (mz->len < min_chunk_size) {
715 rte_memzone_free(mz);
725 if (no_pageshift || try_contig)
726 ret = rte_mempool_populate_iova(mp, mz->addr,
728 rte_mempool_memchunk_mz_free,
729 (void *)(uintptr_t)mz);
731 ret = rte_mempool_populate_virt(mp, mz->addr,
733 rte_mempool_memchunk_mz_free,
734 (void *)(uintptr_t)mz);
736 rte_memzone_free(mz);
744 rte_mempool_free_memchunks(mp);
748 /* return the memory size required for mempool objects in anonymous mem */
750 get_anon_size(const struct rte_mempool *mp)
753 size_t pg_sz, pg_shift;
754 size_t min_chunk_size;
757 pg_sz = getpagesize();
758 pg_shift = rte_bsf32(pg_sz);
759 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
760 &min_chunk_size, &align);
765 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
767 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
773 * Calculate size since memhdr->len has contiguous chunk length
774 * which may be smaller if anon map is split into many contiguous
775 * chunks. Result must be the same as we calculated on populate.
777 size = get_anon_size(memhdr->mp);
781 munmap(opaque, size);
784 /* populate the mempool with an anonymous mapping */
786 rte_mempool_populate_anon(struct rte_mempool *mp)
792 /* mempool is already populated, error */
793 if (!STAILQ_EMPTY(&mp->mem_list)) {
798 ret = mempool_ops_alloc_once(mp);
802 size = get_anon_size(mp);
808 /* get chunk of virtually continuous memory */
809 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
810 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
811 if (addr == MAP_FAILED) {
815 /* can't use MMAP_LOCKED, it does not exist on BSD */
816 if (mlock(addr, size) < 0) {
822 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
823 rte_mempool_memchunk_anon_free, addr);
827 return mp->populated_size;
830 rte_mempool_free_memchunks(mp);
836 rte_mempool_free(struct rte_mempool *mp)
838 struct rte_mempool_list *mempool_list = NULL;
839 struct rte_tailq_entry *te;
844 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
845 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
846 /* find out tailq entry */
847 TAILQ_FOREACH(te, mempool_list, next) {
848 if (te->data == (void *)mp)
853 TAILQ_REMOVE(mempool_list, te, next);
856 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
858 rte_mempool_free_memchunks(mp);
859 rte_mempool_ops_free(mp);
860 rte_memzone_free(mp->mz);
864 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
867 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
872 * Create and initialize a cache for objects that are retrieved from and
873 * returned to an underlying mempool. This structure is identical to the
874 * local_cache[lcore_id] pointed to by the mempool structure.
876 struct rte_mempool_cache *
877 rte_mempool_cache_create(uint32_t size, int socket_id)
879 struct rte_mempool_cache *cache;
881 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
886 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
887 RTE_CACHE_LINE_SIZE, socket_id);
889 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
894 mempool_cache_init(cache, size);
900 * Free a cache. It's the responsibility of the user to make sure that any
901 * remaining objects in the cache are flushed to the corresponding
905 rte_mempool_cache_free(struct rte_mempool_cache *cache)
910 /* create an empty mempool */
912 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
913 unsigned cache_size, unsigned private_data_size,
914 int socket_id, unsigned flags)
916 char mz_name[RTE_MEMZONE_NAMESIZE];
917 struct rte_mempool_list *mempool_list;
918 struct rte_mempool *mp = NULL;
919 struct rte_tailq_entry *te = NULL;
920 const struct rte_memzone *mz = NULL;
922 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
923 struct rte_mempool_objsz objsz;
927 /* compilation-time checks */
928 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
929 RTE_CACHE_LINE_MASK) != 0);
930 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
931 RTE_CACHE_LINE_MASK) != 0);
932 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
933 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
934 RTE_CACHE_LINE_MASK) != 0);
935 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
936 RTE_CACHE_LINE_MASK) != 0);
939 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
941 /* asked cache too big */
942 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
943 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
948 /* "no cache align" imply "no spread" */
949 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
950 flags |= MEMPOOL_F_NO_SPREAD;
952 /* calculate mempool object sizes. */
953 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
958 rte_rwlock_write_lock(RTE_EAL_MEMPOOL_RWLOCK);
961 * reserve a memory zone for this mempool: private data is
964 private_data_size = (private_data_size +
965 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
968 /* try to allocate tailq entry */
969 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
971 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
975 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
976 mempool_size += private_data_size;
977 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
979 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
980 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
981 rte_errno = ENAMETOOLONG;
985 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
989 /* init the mempool structure */
991 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
992 ret = snprintf(mp->name, sizeof(mp->name), "%s", name);
993 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
994 rte_errno = ENAMETOOLONG;
1000 mp->socket_id = socket_id;
1001 mp->elt_size = objsz.elt_size;
1002 mp->header_size = objsz.header_size;
1003 mp->trailer_size = objsz.trailer_size;
1004 /* Size of default caches, zero means disabled. */
1005 mp->cache_size = cache_size;
1006 mp->private_data_size = private_data_size;
1007 STAILQ_INIT(&mp->elt_list);
1008 STAILQ_INIT(&mp->mem_list);
1011 * local_cache pointer is set even if cache_size is zero.
1012 * The local_cache points to just past the elt_pa[] array.
1014 mp->local_cache = (struct rte_mempool_cache *)
1015 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
1017 /* Init all default caches. */
1018 if (cache_size != 0) {
1019 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1020 mempool_cache_init(&mp->local_cache[lcore_id],
1026 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
1027 TAILQ_INSERT_TAIL(mempool_list, te, next);
1028 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
1029 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1034 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1036 rte_mempool_free(mp);
1040 /* create the mempool */
1041 struct rte_mempool *
1042 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
1043 unsigned cache_size, unsigned private_data_size,
1044 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1045 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1046 int socket_id, unsigned flags)
1049 struct rte_mempool *mp;
1051 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1052 private_data_size, socket_id, flags);
1057 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
1058 * set the correct index into the table of ops structs.
1060 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
1061 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
1062 else if (flags & MEMPOOL_F_SP_PUT)
1063 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
1064 else if (flags & MEMPOOL_F_SC_GET)
1065 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
1067 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
1072 /* call the mempool priv initializer */
1074 mp_init(mp, mp_init_arg);
1076 if (rte_mempool_populate_default(mp) < 0)
1079 /* call the object initializers */
1081 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1086 rte_mempool_free(mp);
1091 * Create the mempool over already allocated chunk of memory.
1092 * That external memory buffer can consists of physically disjoint pages.
1093 * Setting vaddr to NULL, makes mempool to fallback to rte_mempool_create()
1096 struct rte_mempool *
1097 rte_mempool_xmem_create(const char *name, unsigned n, unsigned elt_size,
1098 unsigned cache_size, unsigned private_data_size,
1099 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1100 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1101 int socket_id, unsigned flags, void *vaddr,
1102 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift)
1104 struct rte_mempool *mp = NULL;
1107 /* no virtual address supplied, use rte_mempool_create() */
1109 return rte_mempool_create(name, n, elt_size, cache_size,
1110 private_data_size, mp_init, mp_init_arg,
1111 obj_init, obj_init_arg, socket_id, flags);
1113 /* check that we have both VA and PA */
1119 /* Check that pg_shift parameter is valid. */
1120 if (pg_shift > MEMPOOL_PG_SHIFT_MAX) {
1125 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1126 private_data_size, socket_id, flags);
1130 /* call the mempool priv initializer */
1132 mp_init(mp, mp_init_arg);
1134 ret = rte_mempool_populate_iova_tab(mp, vaddr, iova, pg_num, pg_shift,
1136 if (ret < 0 || ret != (int)mp->size)
1139 /* call the object initializers */
1141 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1146 rte_mempool_free(mp);
1150 /* Return the number of entries in the mempool */
1152 rte_mempool_avail_count(const struct rte_mempool *mp)
1157 count = rte_mempool_ops_get_count(mp);
1159 if (mp->cache_size == 0)
1162 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1163 count += mp->local_cache[lcore_id].len;
1166 * due to race condition (access to len is not locked), the
1167 * total can be greater than size... so fix the result
1169 if (count > mp->size)
1174 /* return the number of entries allocated from the mempool */
1176 rte_mempool_in_use_count(const struct rte_mempool *mp)
1178 return mp->size - rte_mempool_avail_count(mp);
1181 /* dump the cache status */
1183 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1187 unsigned cache_count;
1189 fprintf(f, " internal cache infos:\n");
1190 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1192 if (mp->cache_size == 0)
1195 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1196 cache_count = mp->local_cache[lcore_id].len;
1197 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1198 lcore_id, cache_count);
1199 count += cache_count;
1201 fprintf(f, " total_cache_count=%u\n", count);
1205 #ifndef __INTEL_COMPILER
1206 #pragma GCC diagnostic ignored "-Wcast-qual"
1209 /* check and update cookies or panic (internal) */
1210 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1211 void * const *obj_table_const, unsigned n, int free)
1213 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1214 struct rte_mempool_objhdr *hdr;
1215 struct rte_mempool_objtlr *tlr;
1221 /* Force to drop the "const" attribute. This is done only when
1222 * DEBUG is enabled */
1223 tmp = (void *) obj_table_const;
1229 if (rte_mempool_from_obj(obj) != mp)
1230 rte_panic("MEMPOOL: object is owned by another "
1233 hdr = __mempool_get_header(obj);
1234 cookie = hdr->cookie;
1237 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1238 RTE_LOG(CRIT, MEMPOOL,
1239 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1240 obj, (const void *) mp, cookie);
1241 rte_panic("MEMPOOL: bad header cookie (put)\n");
1243 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1244 } else if (free == 1) {
1245 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1246 RTE_LOG(CRIT, MEMPOOL,
1247 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1248 obj, (const void *) mp, cookie);
1249 rte_panic("MEMPOOL: bad header cookie (get)\n");
1251 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1252 } else if (free == 2) {
1253 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1254 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1255 RTE_LOG(CRIT, MEMPOOL,
1256 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1257 obj, (const void *) mp, cookie);
1258 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1261 tlr = __mempool_get_trailer(obj);
1262 cookie = tlr->cookie;
1263 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1264 RTE_LOG(CRIT, MEMPOOL,
1265 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1266 obj, (const void *) mp, cookie);
1267 rte_panic("MEMPOOL: bad trailer cookie\n");
1272 RTE_SET_USED(obj_table_const);
1278 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1280 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1281 void *obj, __rte_unused unsigned idx)
1283 __mempool_check_cookies(mp, &obj, 1, 2);
1287 mempool_audit_cookies(struct rte_mempool *mp)
1291 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1292 if (num != mp->size) {
1293 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1294 "iterated only over %u elements\n",
1299 #define mempool_audit_cookies(mp) do {} while(0)
1302 #ifndef __INTEL_COMPILER
1303 #pragma GCC diagnostic error "-Wcast-qual"
1306 /* check cookies before and after objects */
1308 mempool_audit_cache(const struct rte_mempool *mp)
1310 /* check cache size consistency */
1313 if (mp->cache_size == 0)
1316 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1317 const struct rte_mempool_cache *cache;
1318 cache = &mp->local_cache[lcore_id];
1319 if (cache->len > cache->flushthresh) {
1320 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1322 rte_panic("MEMPOOL: invalid cache len\n");
1327 /* check the consistency of mempool (size, cookies, ...) */
1329 rte_mempool_audit(struct rte_mempool *mp)
1331 mempool_audit_cache(mp);
1332 mempool_audit_cookies(mp);
1334 /* For case where mempool DEBUG is not set, and cache size is 0 */
1338 /* dump the status of the mempool on the console */
1340 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1342 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1343 struct rte_mempool_debug_stats sum;
1346 struct rte_mempool_memhdr *memhdr;
1347 unsigned common_count;
1348 unsigned cache_count;
1351 RTE_ASSERT(f != NULL);
1352 RTE_ASSERT(mp != NULL);
1354 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1355 fprintf(f, " flags=%x\n", mp->flags);
1356 fprintf(f, " pool=%p\n", mp->pool_data);
1357 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1358 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1359 fprintf(f, " size=%"PRIu32"\n", mp->size);
1360 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1361 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1362 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1363 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1364 fprintf(f, " total_obj_size=%"PRIu32"\n",
1365 mp->header_size + mp->elt_size + mp->trailer_size);
1367 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1369 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1370 mem_len += memhdr->len;
1372 fprintf(f, " avg bytes/object=%#Lf\n",
1373 (long double)mem_len / mp->size);
1376 cache_count = rte_mempool_dump_cache(f, mp);
1377 common_count = rte_mempool_ops_get_count(mp);
1378 if ((cache_count + common_count) > mp->size)
1379 common_count = mp->size - cache_count;
1380 fprintf(f, " common_pool_count=%u\n", common_count);
1382 /* sum and dump statistics */
1383 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1384 memset(&sum, 0, sizeof(sum));
1385 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1386 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1387 sum.put_objs += mp->stats[lcore_id].put_objs;
1388 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1389 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1390 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1391 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1393 fprintf(f, " stats:\n");
1394 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1395 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1396 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1397 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1398 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1399 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1401 fprintf(f, " no statistics available\n");
1404 rte_mempool_audit(mp);
1407 /* dump the status of all mempools on the console */
1409 rte_mempool_list_dump(FILE *f)
1411 struct rte_mempool *mp = NULL;
1412 struct rte_tailq_entry *te;
1413 struct rte_mempool_list *mempool_list;
1415 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1417 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1419 TAILQ_FOREACH(te, mempool_list, next) {
1420 mp = (struct rte_mempool *) te->data;
1421 rte_mempool_dump(f, mp);
1424 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1427 /* search a mempool from its name */
1428 struct rte_mempool *
1429 rte_mempool_lookup(const char *name)
1431 struct rte_mempool *mp = NULL;
1432 struct rte_tailq_entry *te;
1433 struct rte_mempool_list *mempool_list;
1435 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1437 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1439 TAILQ_FOREACH(te, mempool_list, next) {
1440 mp = (struct rte_mempool *) te->data;
1441 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1445 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1455 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1458 struct rte_tailq_entry *te = NULL;
1459 struct rte_mempool_list *mempool_list;
1462 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1464 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1466 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1467 (*func)((struct rte_mempool *) te->data, arg);
1470 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);