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, void *obj, rte_iova_t iova)
127 struct rte_mempool_objhdr *hdr;
128 struct rte_mempool_objtlr *tlr __rte_unused;
130 /* set mempool ptr in header */
131 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
134 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
135 mp->populated_size++;
137 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
138 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
139 tlr = __mempool_get_trailer(obj);
140 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
143 /* enqueue in ring */
144 rte_mempool_ops_enqueue_bulk(mp, &obj, 1);
147 /* call obj_cb() for each mempool element */
149 rte_mempool_obj_iter(struct rte_mempool *mp,
150 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
152 struct rte_mempool_objhdr *hdr;
156 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
157 obj = (char *)hdr + sizeof(*hdr);
158 obj_cb(mp, obj_cb_arg, obj, n);
165 /* call mem_cb() for each mempool memory chunk */
167 rte_mempool_mem_iter(struct rte_mempool *mp,
168 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
170 struct rte_mempool_memhdr *hdr;
173 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
174 mem_cb(mp, mem_cb_arg, hdr, n);
181 /* get the header, trailer and total size of a mempool element. */
183 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
184 struct rte_mempool_objsz *sz)
186 struct rte_mempool_objsz lsz;
188 sz = (sz != NULL) ? sz : &lsz;
190 sz->header_size = sizeof(struct rte_mempool_objhdr);
191 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
192 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
195 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
196 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
198 sz->trailer_size = 0;
201 /* element size is 8 bytes-aligned at least */
202 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
204 /* expand trailer to next cache line */
205 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
206 sz->total_size = sz->header_size + sz->elt_size +
208 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
209 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
210 RTE_MEMPOOL_ALIGN_MASK);
214 * increase trailer to add padding between objects in order to
215 * spread them across memory channels/ranks
217 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
219 new_size = optimize_object_size(sz->header_size + sz->elt_size +
221 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
224 /* this is the size of an object, including header and trailer */
225 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
227 return sz->total_size;
232 * Calculate maximum amount of memory required to store given number of objects.
235 rte_mempool_xmem_size(uint32_t elt_num, size_t total_elt_sz, uint32_t pg_shift,
238 size_t obj_per_page, pg_num, pg_sz;
241 mask = MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS | MEMPOOL_F_CAPA_PHYS_CONTIG;
242 if ((flags & mask) == mask)
243 /* alignment need one additional object */
246 if (total_elt_sz == 0)
250 return total_elt_sz * elt_num;
252 pg_sz = (size_t)1 << pg_shift;
253 obj_per_page = pg_sz / total_elt_sz;
254 if (obj_per_page == 0)
255 return RTE_ALIGN_CEIL(total_elt_sz, pg_sz) * elt_num;
257 pg_num = (elt_num + obj_per_page - 1) / obj_per_page;
258 return pg_num << pg_shift;
262 * Calculate how much memory would be actually required with the
263 * given memory footprint to store required number of elements.
266 rte_mempool_xmem_usage(__rte_unused void *vaddr, uint32_t elt_num,
267 size_t total_elt_sz, const rte_iova_t iova[], uint32_t pg_num,
268 uint32_t pg_shift, unsigned int flags)
270 uint32_t elt_cnt = 0;
271 rte_iova_t start, end;
273 size_t pg_sz = (size_t)1 << pg_shift;
276 mask = MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS | MEMPOOL_F_CAPA_PHYS_CONTIG;
277 if ((flags & mask) == mask)
278 /* alignment need one additional object */
281 /* if iova is NULL, assume contiguous memory */
284 end = pg_sz * pg_num;
288 end = iova[0] + pg_sz;
291 while (elt_cnt < elt_num) {
293 if (end - start >= total_elt_sz) {
294 /* enough contiguous memory, add an object */
295 start += total_elt_sz;
297 } else if (iova_idx < pg_num) {
298 /* no room to store one obj, add a page */
299 if (end == iova[iova_idx]) {
302 start = iova[iova_idx];
303 end = iova[iova_idx] + pg_sz;
308 /* no more page, return how many elements fit */
309 return -(size_t)elt_cnt;
313 return (size_t)iova_idx << pg_shift;
316 /* free a memchunk allocated with rte_memzone_reserve() */
318 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
321 const struct rte_memzone *mz = opaque;
322 rte_memzone_free(mz);
325 /* Free memory chunks used by a mempool. Objects must be in pool */
327 rte_mempool_free_memchunks(struct rte_mempool *mp)
329 struct rte_mempool_memhdr *memhdr;
332 while (!STAILQ_EMPTY(&mp->elt_list)) {
333 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
335 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
336 mp->populated_size--;
339 while (!STAILQ_EMPTY(&mp->mem_list)) {
340 memhdr = STAILQ_FIRST(&mp->mem_list);
341 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
342 if (memhdr->free_cb != NULL)
343 memhdr->free_cb(memhdr, memhdr->opaque);
350 mempool_ops_alloc_once(struct rte_mempool *mp)
354 /* create the internal ring if not already done */
355 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
356 ret = rte_mempool_ops_alloc(mp);
359 mp->flags |= MEMPOOL_F_POOL_CREATED;
364 /* Add objects in the pool, using a physically contiguous memory
365 * zone. Return the number of objects added, or a negative value
369 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
370 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
373 unsigned total_elt_sz;
374 unsigned int mp_capa_flags;
377 struct rte_mempool_memhdr *memhdr;
380 ret = mempool_ops_alloc_once(mp);
384 /* Notify memory area to mempool */
385 ret = rte_mempool_ops_register_memory_area(mp, vaddr, iova, len);
386 if (ret != -ENOTSUP && ret < 0)
389 /* mempool is already populated */
390 if (mp->populated_size >= mp->size)
393 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
395 /* Get mempool capabilities */
397 ret = rte_mempool_ops_get_capabilities(mp, &mp_capa_flags);
398 if ((ret < 0) && (ret != -ENOTSUP))
401 /* update mempool capabilities */
402 mp->flags |= mp_capa_flags;
404 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
409 memhdr->addr = vaddr;
412 memhdr->free_cb = free_cb;
413 memhdr->opaque = opaque;
415 if (mp_capa_flags & MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS)
416 /* align object start address to a multiple of total_elt_sz */
417 off = total_elt_sz - ((uintptr_t)vaddr % total_elt_sz);
418 else if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
419 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
421 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
423 while (off + total_elt_sz <= len && mp->populated_size < mp->size) {
424 off += mp->header_size;
425 if (iova == RTE_BAD_IOVA)
426 mempool_add_elem(mp, (char *)vaddr + off,
429 mempool_add_elem(mp, (char *)vaddr + off, iova + off);
430 off += mp->elt_size + mp->trailer_size;
434 /* not enough room to store one object */
440 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
450 rte_mempool_populate_phys(struct rte_mempool *mp, char *vaddr,
451 phys_addr_t paddr, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
454 return rte_mempool_populate_iova(mp, vaddr, paddr, len, free_cb, opaque);
457 /* Add objects in the pool, using a table of physical pages. Return the
458 * number of objects added, or a negative value on error.
461 rte_mempool_populate_iova_tab(struct rte_mempool *mp, char *vaddr,
462 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift,
463 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
467 size_t pg_sz = (size_t)1 << pg_shift;
469 /* mempool must not be populated */
470 if (mp->nb_mem_chunks != 0)
473 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
474 return rte_mempool_populate_iova(mp, vaddr, RTE_BAD_IOVA,
475 pg_num * pg_sz, free_cb, opaque);
477 for (i = 0; i < pg_num && mp->populated_size < mp->size; i += n) {
479 /* populate with the largest group of contiguous pages */
480 for (n = 1; (i + n) < pg_num &&
481 iova[i + n - 1] + pg_sz == iova[i + n]; n++)
484 ret = rte_mempool_populate_iova(mp, vaddr + i * pg_sz,
485 iova[i], n * pg_sz, free_cb, opaque);
487 rte_mempool_free_memchunks(mp);
490 /* no need to call the free callback for next chunks */
498 rte_mempool_populate_phys_tab(struct rte_mempool *mp, char *vaddr,
499 const phys_addr_t paddr[], uint32_t pg_num, uint32_t pg_shift,
500 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
502 return rte_mempool_populate_iova_tab(mp, vaddr, paddr, pg_num, pg_shift,
506 /* Populate the mempool with a virtual area. Return the number of
507 * objects added, or a negative value on error.
510 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
511 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
515 size_t off, phys_len;
518 /* mempool must not be populated */
519 if (mp->nb_mem_chunks != 0)
521 /* address and len must be page-aligned */
522 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
524 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
527 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
528 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
529 len, free_cb, opaque);
531 for (off = 0; off + pg_sz <= len &&
532 mp->populated_size < mp->size; off += phys_len) {
534 iova = rte_mem_virt2iova(addr + off);
536 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
541 /* populate with the largest group of contiguous pages */
542 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
545 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
547 if (iova_tmp != iova + phys_len)
551 ret = rte_mempool_populate_iova(mp, addr + off, iova,
552 phys_len, free_cb, opaque);
555 /* no need to call the free callback for next chunks */
563 rte_mempool_free_memchunks(mp);
567 /* Default function to populate the mempool: allocate memory in memzones,
568 * and populate them. Return the number of objects added, or a negative
572 rte_mempool_populate_default(struct rte_mempool *mp)
574 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
575 char mz_name[RTE_MEMZONE_NAMESIZE];
576 const struct rte_memzone *mz;
578 size_t align, pg_sz, pg_shift;
582 bool no_contig, try_contig, no_pageshift;
584 ret = mempool_ops_alloc_once(mp);
588 /* mempool must not be populated */
589 if (mp->nb_mem_chunks != 0)
592 no_contig = mp->flags & MEMPOOL_F_NO_IOVA_CONTIG;
595 * the following section calculates page shift and page size values.
597 * these values impact the result of calc_mem_size operation, which
598 * returns the amount of memory that should be allocated to store the
599 * desired number of objects. when not zero, it allocates more memory
600 * for the padding between objects, to ensure that an object does not
601 * cross a page boundary. in other words, page size/shift are to be set
602 * to zero if mempool elements won't care about page boundaries.
603 * there are several considerations for page size and page shift here.
605 * if we don't need our mempools to have physically contiguous objects,
606 * then just set page shift and page size to 0, because the user has
607 * indicated that there's no need to care about anything.
609 * if we do need contiguous objects, there is also an option to reserve
610 * the entire mempool memory as one contiguous block of memory, in
611 * which case the page shift and alignment wouldn't matter as well.
613 * if we require contiguous objects, but not necessarily the entire
614 * mempool reserved space to be contiguous, then there are two options.
616 * if our IO addresses are virtual, not actual physical (IOVA as VA
617 * case), then no page shift needed - our memory allocation will give us
618 * contiguous IO memory as far as the hardware is concerned, so
619 * act as if we're getting contiguous memory.
621 * if our IO addresses are physical, we may get memory from bigger
622 * pages, or we might get memory from smaller pages, and how much of it
623 * we require depends on whether we want bigger or smaller pages.
624 * However, requesting each and every memory size is too much work, so
625 * what we'll do instead is walk through the page sizes available, pick
626 * the smallest one and set up page shift to match that one. We will be
627 * wasting some space this way, but it's much nicer than looping around
628 * trying to reserve each and every page size.
630 * However, since size calculation will produce page-aligned sizes, it
631 * makes sense to first try and see if we can reserve the entire memzone
632 * in one contiguous chunk as well (otherwise we might end up wasting a
633 * 1G page on a 10MB memzone). If we fail to get enough contiguous
634 * memory, then we'll go and reserve space page-by-page.
636 no_pageshift = no_contig || rte_eal_iova_mode() == RTE_IOVA_VA;
637 try_contig = !no_contig && !no_pageshift && rte_eal_has_hugepages();
642 } else if (try_contig) {
643 pg_sz = get_min_page_size();
644 pg_shift = rte_bsf32(pg_sz);
646 pg_sz = getpagesize();
647 pg_shift = rte_bsf32(pg_sz);
650 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
651 size_t min_chunk_size;
654 if (try_contig || no_pageshift)
655 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
656 0, &min_chunk_size, &align);
658 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
659 pg_shift, &min_chunk_size, &align);
666 ret = snprintf(mz_name, sizeof(mz_name),
667 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
668 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
675 /* if we're trying to reserve contiguous memory, add appropriate
679 flags |= RTE_MEMZONE_IOVA_CONTIG;
681 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
682 mp->socket_id, flags, align);
684 /* if we were trying to allocate contiguous memory, failed and
685 * minimum required contiguous chunk fits minimum page, adjust
686 * memzone size to the page size, and try again.
688 if (mz == NULL && try_contig && min_chunk_size <= pg_sz) {
690 flags &= ~RTE_MEMZONE_IOVA_CONTIG;
692 mem_size = rte_mempool_ops_calc_mem_size(mp, n,
693 pg_shift, &min_chunk_size, &align);
699 mz = rte_memzone_reserve_aligned(mz_name, mem_size,
700 mp->socket_id, flags, align);
702 /* don't try reserving with 0 size if we were asked to reserve
703 * IOVA-contiguous memory.
705 if (min_chunk_size < (size_t)mem_size && mz == NULL) {
706 /* not enough memory, retry with the biggest zone we
709 mz = rte_memzone_reserve_aligned(mz_name, 0,
710 mp->socket_id, flags, align);
717 if (mz->len < min_chunk_size) {
718 rte_memzone_free(mz);
728 if (no_pageshift || try_contig)
729 ret = rte_mempool_populate_iova(mp, mz->addr,
731 rte_mempool_memchunk_mz_free,
732 (void *)(uintptr_t)mz);
734 ret = rte_mempool_populate_virt(mp, mz->addr,
736 rte_mempool_memchunk_mz_free,
737 (void *)(uintptr_t)mz);
739 rte_memzone_free(mz);
747 rte_mempool_free_memchunks(mp);
751 /* return the memory size required for mempool objects in anonymous mem */
753 get_anon_size(const struct rte_mempool *mp)
756 size_t pg_sz, pg_shift;
757 size_t min_chunk_size;
760 pg_sz = getpagesize();
761 pg_shift = rte_bsf32(pg_sz);
762 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
763 &min_chunk_size, &align);
768 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
770 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
776 * Calculate size since memhdr->len has contiguous chunk length
777 * which may be smaller if anon map is split into many contiguous
778 * chunks. Result must be the same as we calculated on populate.
780 size = get_anon_size(memhdr->mp);
784 munmap(opaque, size);
787 /* populate the mempool with an anonymous mapping */
789 rte_mempool_populate_anon(struct rte_mempool *mp)
795 /* mempool is already populated, error */
796 if (!STAILQ_EMPTY(&mp->mem_list)) {
801 ret = mempool_ops_alloc_once(mp);
805 size = get_anon_size(mp);
811 /* get chunk of virtually continuous memory */
812 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
813 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
814 if (addr == MAP_FAILED) {
818 /* can't use MMAP_LOCKED, it does not exist on BSD */
819 if (mlock(addr, size) < 0) {
825 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
826 rte_mempool_memchunk_anon_free, addr);
830 return mp->populated_size;
833 rte_mempool_free_memchunks(mp);
839 rte_mempool_free(struct rte_mempool *mp)
841 struct rte_mempool_list *mempool_list = NULL;
842 struct rte_tailq_entry *te;
847 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
848 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
849 /* find out tailq entry */
850 TAILQ_FOREACH(te, mempool_list, next) {
851 if (te->data == (void *)mp)
856 TAILQ_REMOVE(mempool_list, te, next);
859 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
861 rte_mempool_free_memchunks(mp);
862 rte_mempool_ops_free(mp);
863 rte_memzone_free(mp->mz);
867 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
870 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
875 * Create and initialize a cache for objects that are retrieved from and
876 * returned to an underlying mempool. This structure is identical to the
877 * local_cache[lcore_id] pointed to by the mempool structure.
879 struct rte_mempool_cache *
880 rte_mempool_cache_create(uint32_t size, int socket_id)
882 struct rte_mempool_cache *cache;
884 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
889 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
890 RTE_CACHE_LINE_SIZE, socket_id);
892 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
897 mempool_cache_init(cache, size);
903 * Free a cache. It's the responsibility of the user to make sure that any
904 * remaining objects in the cache are flushed to the corresponding
908 rte_mempool_cache_free(struct rte_mempool_cache *cache)
913 /* create an empty mempool */
915 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
916 unsigned cache_size, unsigned private_data_size,
917 int socket_id, unsigned flags)
919 char mz_name[RTE_MEMZONE_NAMESIZE];
920 struct rte_mempool_list *mempool_list;
921 struct rte_mempool *mp = NULL;
922 struct rte_tailq_entry *te = NULL;
923 const struct rte_memzone *mz = NULL;
925 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
926 struct rte_mempool_objsz objsz;
930 /* compilation-time checks */
931 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
932 RTE_CACHE_LINE_MASK) != 0);
933 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
934 RTE_CACHE_LINE_MASK) != 0);
935 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
936 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
937 RTE_CACHE_LINE_MASK) != 0);
938 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
939 RTE_CACHE_LINE_MASK) != 0);
942 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
944 /* asked cache too big */
945 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
946 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
951 /* "no cache align" imply "no spread" */
952 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
953 flags |= MEMPOOL_F_NO_SPREAD;
955 /* calculate mempool object sizes. */
956 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
961 rte_rwlock_write_lock(RTE_EAL_MEMPOOL_RWLOCK);
964 * reserve a memory zone for this mempool: private data is
967 private_data_size = (private_data_size +
968 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
971 /* try to allocate tailq entry */
972 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
974 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
978 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
979 mempool_size += private_data_size;
980 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
982 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
983 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
984 rte_errno = ENAMETOOLONG;
988 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
992 /* init the mempool structure */
994 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
995 ret = snprintf(mp->name, sizeof(mp->name), "%s", name);
996 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
997 rte_errno = ENAMETOOLONG;
1003 mp->socket_id = socket_id;
1004 mp->elt_size = objsz.elt_size;
1005 mp->header_size = objsz.header_size;
1006 mp->trailer_size = objsz.trailer_size;
1007 /* Size of default caches, zero means disabled. */
1008 mp->cache_size = cache_size;
1009 mp->private_data_size = private_data_size;
1010 STAILQ_INIT(&mp->elt_list);
1011 STAILQ_INIT(&mp->mem_list);
1014 * local_cache pointer is set even if cache_size is zero.
1015 * The local_cache points to just past the elt_pa[] array.
1017 mp->local_cache = (struct rte_mempool_cache *)
1018 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
1020 /* Init all default caches. */
1021 if (cache_size != 0) {
1022 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1023 mempool_cache_init(&mp->local_cache[lcore_id],
1029 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
1030 TAILQ_INSERT_TAIL(mempool_list, te, next);
1031 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
1032 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1037 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1039 rte_mempool_free(mp);
1043 /* create the mempool */
1044 struct rte_mempool *
1045 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
1046 unsigned cache_size, unsigned private_data_size,
1047 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1048 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1049 int socket_id, unsigned flags)
1052 struct rte_mempool *mp;
1054 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1055 private_data_size, socket_id, flags);
1060 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
1061 * set the correct index into the table of ops structs.
1063 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
1064 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
1065 else if (flags & MEMPOOL_F_SP_PUT)
1066 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
1067 else if (flags & MEMPOOL_F_SC_GET)
1068 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
1070 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
1075 /* call the mempool priv initializer */
1077 mp_init(mp, mp_init_arg);
1079 if (rte_mempool_populate_default(mp) < 0)
1082 /* call the object initializers */
1084 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1089 rte_mempool_free(mp);
1094 * Create the mempool over already allocated chunk of memory.
1095 * That external memory buffer can consists of physically disjoint pages.
1096 * Setting vaddr to NULL, makes mempool to fallback to rte_mempool_create()
1099 struct rte_mempool *
1100 rte_mempool_xmem_create(const char *name, unsigned n, unsigned elt_size,
1101 unsigned cache_size, unsigned private_data_size,
1102 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1103 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1104 int socket_id, unsigned flags, void *vaddr,
1105 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift)
1107 struct rte_mempool *mp = NULL;
1110 /* no virtual address supplied, use rte_mempool_create() */
1112 return rte_mempool_create(name, n, elt_size, cache_size,
1113 private_data_size, mp_init, mp_init_arg,
1114 obj_init, obj_init_arg, socket_id, flags);
1116 /* check that we have both VA and PA */
1122 /* Check that pg_shift parameter is valid. */
1123 if (pg_shift > MEMPOOL_PG_SHIFT_MAX) {
1128 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1129 private_data_size, socket_id, flags);
1133 /* call the mempool priv initializer */
1135 mp_init(mp, mp_init_arg);
1137 ret = rte_mempool_populate_iova_tab(mp, vaddr, iova, pg_num, pg_shift,
1139 if (ret < 0 || ret != (int)mp->size)
1142 /* call the object initializers */
1144 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1149 rte_mempool_free(mp);
1153 /* Return the number of entries in the mempool */
1155 rte_mempool_avail_count(const struct rte_mempool *mp)
1160 count = rte_mempool_ops_get_count(mp);
1162 if (mp->cache_size == 0)
1165 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1166 count += mp->local_cache[lcore_id].len;
1169 * due to race condition (access to len is not locked), the
1170 * total can be greater than size... so fix the result
1172 if (count > mp->size)
1177 /* return the number of entries allocated from the mempool */
1179 rte_mempool_in_use_count(const struct rte_mempool *mp)
1181 return mp->size - rte_mempool_avail_count(mp);
1184 /* dump the cache status */
1186 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1190 unsigned cache_count;
1192 fprintf(f, " internal cache infos:\n");
1193 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1195 if (mp->cache_size == 0)
1198 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1199 cache_count = mp->local_cache[lcore_id].len;
1200 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1201 lcore_id, cache_count);
1202 count += cache_count;
1204 fprintf(f, " total_cache_count=%u\n", count);
1208 #ifndef __INTEL_COMPILER
1209 #pragma GCC diagnostic ignored "-Wcast-qual"
1212 /* check and update cookies or panic (internal) */
1213 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1214 void * const *obj_table_const, unsigned n, int free)
1216 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1217 struct rte_mempool_objhdr *hdr;
1218 struct rte_mempool_objtlr *tlr;
1224 /* Force to drop the "const" attribute. This is done only when
1225 * DEBUG is enabled */
1226 tmp = (void *) obj_table_const;
1232 if (rte_mempool_from_obj(obj) != mp)
1233 rte_panic("MEMPOOL: object is owned by another "
1236 hdr = __mempool_get_header(obj);
1237 cookie = hdr->cookie;
1240 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1241 RTE_LOG(CRIT, MEMPOOL,
1242 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1243 obj, (const void *) mp, cookie);
1244 rte_panic("MEMPOOL: bad header cookie (put)\n");
1246 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1247 } else if (free == 1) {
1248 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1249 RTE_LOG(CRIT, MEMPOOL,
1250 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1251 obj, (const void *) mp, cookie);
1252 rte_panic("MEMPOOL: bad header cookie (get)\n");
1254 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1255 } else if (free == 2) {
1256 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1257 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1258 RTE_LOG(CRIT, MEMPOOL,
1259 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1260 obj, (const void *) mp, cookie);
1261 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1264 tlr = __mempool_get_trailer(obj);
1265 cookie = tlr->cookie;
1266 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1267 RTE_LOG(CRIT, MEMPOOL,
1268 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1269 obj, (const void *) mp, cookie);
1270 rte_panic("MEMPOOL: bad trailer cookie\n");
1275 RTE_SET_USED(obj_table_const);
1281 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1283 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1284 void *obj, __rte_unused unsigned idx)
1286 __mempool_check_cookies(mp, &obj, 1, 2);
1290 mempool_audit_cookies(struct rte_mempool *mp)
1294 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1295 if (num != mp->size) {
1296 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1297 "iterated only over %u elements\n",
1302 #define mempool_audit_cookies(mp) do {} while(0)
1305 #ifndef __INTEL_COMPILER
1306 #pragma GCC diagnostic error "-Wcast-qual"
1309 /* check cookies before and after objects */
1311 mempool_audit_cache(const struct rte_mempool *mp)
1313 /* check cache size consistency */
1316 if (mp->cache_size == 0)
1319 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1320 const struct rte_mempool_cache *cache;
1321 cache = &mp->local_cache[lcore_id];
1322 if (cache->len > cache->flushthresh) {
1323 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1325 rte_panic("MEMPOOL: invalid cache len\n");
1330 /* check the consistency of mempool (size, cookies, ...) */
1332 rte_mempool_audit(struct rte_mempool *mp)
1334 mempool_audit_cache(mp);
1335 mempool_audit_cookies(mp);
1337 /* For case where mempool DEBUG is not set, and cache size is 0 */
1341 /* dump the status of the mempool on the console */
1343 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1345 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1346 struct rte_mempool_debug_stats sum;
1349 struct rte_mempool_memhdr *memhdr;
1350 unsigned common_count;
1351 unsigned cache_count;
1354 RTE_ASSERT(f != NULL);
1355 RTE_ASSERT(mp != NULL);
1357 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1358 fprintf(f, " flags=%x\n", mp->flags);
1359 fprintf(f, " pool=%p\n", mp->pool_data);
1360 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1361 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1362 fprintf(f, " size=%"PRIu32"\n", mp->size);
1363 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1364 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1365 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1366 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1367 fprintf(f, " total_obj_size=%"PRIu32"\n",
1368 mp->header_size + mp->elt_size + mp->trailer_size);
1370 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1372 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1373 mem_len += memhdr->len;
1375 fprintf(f, " avg bytes/object=%#Lf\n",
1376 (long double)mem_len / mp->size);
1379 cache_count = rte_mempool_dump_cache(f, mp);
1380 common_count = rte_mempool_ops_get_count(mp);
1381 if ((cache_count + common_count) > mp->size)
1382 common_count = mp->size - cache_count;
1383 fprintf(f, " common_pool_count=%u\n", common_count);
1385 /* sum and dump statistics */
1386 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1387 memset(&sum, 0, sizeof(sum));
1388 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1389 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1390 sum.put_objs += mp->stats[lcore_id].put_objs;
1391 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1392 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1393 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1394 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1396 fprintf(f, " stats:\n");
1397 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1398 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1399 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1400 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1401 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1402 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1404 fprintf(f, " no statistics available\n");
1407 rte_mempool_audit(mp);
1410 /* dump the status of all mempools on the console */
1412 rte_mempool_list_dump(FILE *f)
1414 struct rte_mempool *mp = NULL;
1415 struct rte_tailq_entry *te;
1416 struct rte_mempool_list *mempool_list;
1418 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1420 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1422 TAILQ_FOREACH(te, mempool_list, next) {
1423 mp = (struct rte_mempool *) te->data;
1424 rte_mempool_dump(f, mp);
1427 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1430 /* search a mempool from its name */
1431 struct rte_mempool *
1432 rte_mempool_lookup(const char *name)
1434 struct rte_mempool *mp = NULL;
1435 struct rte_tailq_entry *te;
1436 struct rte_mempool_list *mempool_list;
1438 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1440 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1442 TAILQ_FOREACH(te, mempool_list, next) {
1443 mp = (struct rte_mempool *) te->data;
1444 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1448 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1458 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1461 struct rte_tailq_entry *te = NULL;
1462 struct rte_mempool_list *mempool_list;
1465 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1467 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1469 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1470 (*func)((struct rte_mempool *) te->data, arg);
1473 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);