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);
349 /* Add objects in the pool, using a physically contiguous memory
350 * zone. Return the number of objects added, or a negative value
354 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
355 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
358 unsigned total_elt_sz;
359 unsigned int mp_capa_flags;
362 struct rte_mempool_memhdr *memhdr;
365 /* create the internal ring if not already done */
366 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
367 ret = rte_mempool_ops_alloc(mp);
370 mp->flags |= MEMPOOL_F_POOL_CREATED;
373 /* Notify memory area to mempool */
374 ret = rte_mempool_ops_register_memory_area(mp, vaddr, iova, len);
375 if (ret != -ENOTSUP && ret < 0)
378 /* mempool is already populated */
379 if (mp->populated_size >= mp->size)
382 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
384 /* Get mempool capabilities */
386 ret = rte_mempool_ops_get_capabilities(mp, &mp_capa_flags);
387 if ((ret < 0) && (ret != -ENOTSUP))
390 /* update mempool capabilities */
391 mp->flags |= mp_capa_flags;
393 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
398 memhdr->addr = vaddr;
401 memhdr->free_cb = free_cb;
402 memhdr->opaque = opaque;
404 if (mp_capa_flags & MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS)
405 /* align object start address to a multiple of total_elt_sz */
406 off = total_elt_sz - ((uintptr_t)vaddr % total_elt_sz);
407 else if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
408 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
410 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
412 while (off + total_elt_sz <= len && mp->populated_size < mp->size) {
413 off += mp->header_size;
414 if (iova == RTE_BAD_IOVA)
415 mempool_add_elem(mp, (char *)vaddr + off,
418 mempool_add_elem(mp, (char *)vaddr + off, iova + off);
419 off += mp->elt_size + mp->trailer_size;
423 /* not enough room to store one object */
429 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
439 rte_mempool_populate_phys(struct rte_mempool *mp, char *vaddr,
440 phys_addr_t paddr, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
443 return rte_mempool_populate_iova(mp, vaddr, paddr, len, free_cb, opaque);
446 /* Add objects in the pool, using a table of physical pages. Return the
447 * number of objects added, or a negative value on error.
450 rte_mempool_populate_iova_tab(struct rte_mempool *mp, char *vaddr,
451 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift,
452 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
456 size_t pg_sz = (size_t)1 << pg_shift;
458 /* mempool must not be populated */
459 if (mp->nb_mem_chunks != 0)
462 if (mp->flags & MEMPOOL_F_NO_PHYS_CONTIG)
463 return rte_mempool_populate_iova(mp, vaddr, RTE_BAD_IOVA,
464 pg_num * pg_sz, free_cb, opaque);
466 for (i = 0; i < pg_num && mp->populated_size < mp->size; i += n) {
468 /* populate with the largest group of contiguous pages */
469 for (n = 1; (i + n) < pg_num &&
470 iova[i + n - 1] + pg_sz == iova[i + n]; n++)
473 ret = rte_mempool_populate_iova(mp, vaddr + i * pg_sz,
474 iova[i], n * pg_sz, free_cb, opaque);
476 rte_mempool_free_memchunks(mp);
479 /* no need to call the free callback for next chunks */
487 rte_mempool_populate_phys_tab(struct rte_mempool *mp, char *vaddr,
488 const phys_addr_t paddr[], uint32_t pg_num, uint32_t pg_shift,
489 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
491 return rte_mempool_populate_iova_tab(mp, vaddr, paddr, pg_num, pg_shift,
495 /* Populate the mempool with a virtual area. Return the number of
496 * objects added, or a negative value on error.
499 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
500 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
504 size_t off, phys_len;
507 /* mempool must not be populated */
508 if (mp->nb_mem_chunks != 0)
510 /* address and len must be page-aligned */
511 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
513 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
516 if (mp->flags & MEMPOOL_F_NO_PHYS_CONTIG)
517 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
518 len, free_cb, opaque);
520 for (off = 0; off + pg_sz <= len &&
521 mp->populated_size < mp->size; off += phys_len) {
523 iova = rte_mem_virt2iova(addr + off);
525 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
530 /* populate with the largest group of contiguous pages */
531 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
534 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
536 if (iova_tmp != iova + phys_len)
540 ret = rte_mempool_populate_iova(mp, addr + off, iova,
541 phys_len, free_cb, opaque);
544 /* no need to call the free callback for next chunks */
552 rte_mempool_free_memchunks(mp);
556 /* Default function to populate the mempool: allocate memory in memzones,
557 * and populate them. Return the number of objects added, or a negative
561 rte_mempool_populate_default(struct rte_mempool *mp)
563 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
564 char mz_name[RTE_MEMZONE_NAMESIZE];
565 const struct rte_memzone *mz;
566 size_t size, total_elt_sz, align, pg_sz, pg_shift;
569 unsigned int mp_flags;
571 bool force_contig, no_contig, try_contig, no_pageshift;
573 /* mempool must not be populated */
574 if (mp->nb_mem_chunks != 0)
577 /* Get mempool capabilities */
579 ret = rte_mempool_ops_get_capabilities(mp, &mp_flags);
580 if ((ret < 0) && (ret != -ENOTSUP))
583 /* update mempool capabilities */
584 mp->flags |= mp_flags;
586 no_contig = mp->flags & MEMPOOL_F_NO_PHYS_CONTIG;
587 force_contig = mp->flags & MEMPOOL_F_CAPA_PHYS_CONTIG;
590 * the following section calculates page shift and page size values.
592 * these values impact the result of rte_mempool_xmem_size(), which
593 * returns the amount of memory that should be allocated to store the
594 * desired number of objects. when not zero, it allocates more memory
595 * for the padding between objects, to ensure that an object does not
596 * cross a page boundary. in other words, page size/shift are to be set
597 * to zero if mempool elements won't care about page boundaries.
598 * there are several considerations for page size and page shift here.
600 * if we don't need our mempools to have physically contiguous objects,
601 * then just set page shift and page size to 0, because the user has
602 * indicated that there's no need to care about anything.
604 * if we do need contiguous objects, there is also an option to reserve
605 * the entire mempool memory as one contiguous block of memory, in
606 * which case the page shift and alignment wouldn't matter as well.
608 * if we require contiguous objects, but not necessarily the entire
609 * mempool reserved space to be contiguous, then there are two options.
611 * if our IO addresses are virtual, not actual physical (IOVA as VA
612 * case), then no page shift needed - our memory allocation will give us
613 * contiguous physical memory as far as the hardware is concerned, so
614 * act as if we're getting contiguous memory.
616 * if our IO addresses are physical, we may get memory from bigger
617 * pages, or we might get memory from smaller pages, and how much of it
618 * we require depends on whether we want bigger or smaller pages.
619 * However, requesting each and every memory size is too much work, so
620 * what we'll do instead is walk through the page sizes available, pick
621 * the smallest one and set up page shift to match that one. We will be
622 * wasting some space this way, but it's much nicer than looping around
623 * trying to reserve each and every page size.
625 * However, since size calculation will produce page-aligned sizes, it
626 * makes sense to first try and see if we can reserve the entire memzone
627 * in one contiguous chunk as well (otherwise we might end up wasting a
628 * 1G page on a 10MB memzone). If we fail to get enough contiguous
629 * memory, then we'll go and reserve space page-by-page.
631 no_pageshift = no_contig || force_contig ||
632 rte_eal_iova_mode() == RTE_IOVA_VA;
633 try_contig = !no_contig && !no_pageshift && rte_eal_has_hugepages();
635 mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
640 align = RTE_CACHE_LINE_SIZE;
641 } else if (try_contig) {
642 pg_sz = get_min_page_size();
643 pg_shift = rte_bsf32(pg_sz);
644 /* we're trying to reserve contiguous memzone first, so try
645 * align to cache line; if we fail to reserve a contiguous
646 * memzone, we'll adjust alignment to equal pagesize later.
648 align = RTE_CACHE_LINE_SIZE;
650 pg_sz = getpagesize();
651 pg_shift = rte_bsf32(pg_sz);
655 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
656 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
658 if (try_contig || no_pageshift)
659 size = rte_mempool_xmem_size(n, total_elt_sz, 0,
662 size = rte_mempool_xmem_size(n, total_elt_sz, pg_shift,
665 ret = snprintf(mz_name, sizeof(mz_name),
666 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
667 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
674 /* if we're trying to reserve contiguous memory, add appropriate
678 flags |= RTE_MEMZONE_IOVA_CONTIG;
680 mz = rte_memzone_reserve_aligned(mz_name, size, mp->socket_id,
683 /* if we were trying to allocate contiguous memory, adjust
684 * memzone size and page size to fit smaller page sizes, and
687 if (mz == NULL && try_contig) {
689 flags &= ~RTE_MEMZONE_IOVA_CONTIG;
691 size = rte_mempool_xmem_size(n, total_elt_sz,
692 pg_shift, mp->flags);
694 mz = rte_memzone_reserve_aligned(mz_name, size,
695 mp->socket_id, flags, align);
697 /* don't try reserving with 0 size if we were asked to reserve
698 * IOVA-contiguous memory.
700 if (!force_contig && mz == NULL) {
701 /* not enough memory, retry with the biggest zone we
704 mz = rte_memzone_reserve_aligned(mz_name, 0,
705 mp->socket_id, flags, align);
717 if (no_pageshift || try_contig)
718 ret = rte_mempool_populate_iova(mp, mz->addr,
720 rte_mempool_memchunk_mz_free,
721 (void *)(uintptr_t)mz);
723 ret = rte_mempool_populate_virt(mp, mz->addr,
725 rte_mempool_memchunk_mz_free,
726 (void *)(uintptr_t)mz);
728 rte_memzone_free(mz);
736 rte_mempool_free_memchunks(mp);
740 /* return the memory size required for mempool objects in anonymous mem */
742 get_anon_size(const struct rte_mempool *mp)
744 size_t size, total_elt_sz, pg_sz, pg_shift;
746 pg_sz = getpagesize();
747 pg_shift = rte_bsf32(pg_sz);
748 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
749 size = rte_mempool_xmem_size(mp->size, total_elt_sz, pg_shift,
755 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
757 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
760 munmap(opaque, get_anon_size(memhdr->mp));
763 /* populate the mempool with an anonymous mapping */
765 rte_mempool_populate_anon(struct rte_mempool *mp)
771 /* mempool is already populated, error */
772 if (!STAILQ_EMPTY(&mp->mem_list)) {
777 /* get chunk of virtually continuous memory */
778 size = get_anon_size(mp);
779 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
780 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
781 if (addr == MAP_FAILED) {
785 /* can't use MMAP_LOCKED, it does not exist on BSD */
786 if (mlock(addr, size) < 0) {
792 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
793 rte_mempool_memchunk_anon_free, addr);
797 return mp->populated_size;
800 rte_mempool_free_memchunks(mp);
806 rte_mempool_free(struct rte_mempool *mp)
808 struct rte_mempool_list *mempool_list = NULL;
809 struct rte_tailq_entry *te;
814 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
815 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
816 /* find out tailq entry */
817 TAILQ_FOREACH(te, mempool_list, next) {
818 if (te->data == (void *)mp)
823 TAILQ_REMOVE(mempool_list, te, next);
826 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
828 rte_mempool_free_memchunks(mp);
829 rte_mempool_ops_free(mp);
830 rte_memzone_free(mp->mz);
834 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
837 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
842 * Create and initialize a cache for objects that are retrieved from and
843 * returned to an underlying mempool. This structure is identical to the
844 * local_cache[lcore_id] pointed to by the mempool structure.
846 struct rte_mempool_cache *
847 rte_mempool_cache_create(uint32_t size, int socket_id)
849 struct rte_mempool_cache *cache;
851 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
856 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
857 RTE_CACHE_LINE_SIZE, socket_id);
859 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
864 mempool_cache_init(cache, size);
870 * Free a cache. It's the responsibility of the user to make sure that any
871 * remaining objects in the cache are flushed to the corresponding
875 rte_mempool_cache_free(struct rte_mempool_cache *cache)
880 /* create an empty mempool */
882 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
883 unsigned cache_size, unsigned private_data_size,
884 int socket_id, unsigned flags)
886 char mz_name[RTE_MEMZONE_NAMESIZE];
887 struct rte_mempool_list *mempool_list;
888 struct rte_mempool *mp = NULL;
889 struct rte_tailq_entry *te = NULL;
890 const struct rte_memzone *mz = NULL;
892 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
893 struct rte_mempool_objsz objsz;
897 /* compilation-time checks */
898 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
899 RTE_CACHE_LINE_MASK) != 0);
900 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
901 RTE_CACHE_LINE_MASK) != 0);
902 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
903 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
904 RTE_CACHE_LINE_MASK) != 0);
905 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
906 RTE_CACHE_LINE_MASK) != 0);
909 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
911 /* asked cache too big */
912 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
913 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
918 /* "no cache align" imply "no spread" */
919 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
920 flags |= MEMPOOL_F_NO_SPREAD;
922 /* calculate mempool object sizes. */
923 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
928 rte_rwlock_write_lock(RTE_EAL_MEMPOOL_RWLOCK);
931 * reserve a memory zone for this mempool: private data is
934 private_data_size = (private_data_size +
935 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
938 /* try to allocate tailq entry */
939 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
941 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
945 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
946 mempool_size += private_data_size;
947 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
949 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
950 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
951 rte_errno = ENAMETOOLONG;
955 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
959 /* init the mempool structure */
961 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
962 ret = snprintf(mp->name, sizeof(mp->name), "%s", name);
963 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
964 rte_errno = ENAMETOOLONG;
970 mp->socket_id = socket_id;
971 mp->elt_size = objsz.elt_size;
972 mp->header_size = objsz.header_size;
973 mp->trailer_size = objsz.trailer_size;
974 /* Size of default caches, zero means disabled. */
975 mp->cache_size = cache_size;
976 mp->private_data_size = private_data_size;
977 STAILQ_INIT(&mp->elt_list);
978 STAILQ_INIT(&mp->mem_list);
981 * local_cache pointer is set even if cache_size is zero.
982 * The local_cache points to just past the elt_pa[] array.
984 mp->local_cache = (struct rte_mempool_cache *)
985 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
987 /* Init all default caches. */
988 if (cache_size != 0) {
989 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
990 mempool_cache_init(&mp->local_cache[lcore_id],
996 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
997 TAILQ_INSERT_TAIL(mempool_list, te, next);
998 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
999 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1004 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1006 rte_mempool_free(mp);
1010 /* create the mempool */
1011 struct rte_mempool *
1012 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
1013 unsigned cache_size, unsigned private_data_size,
1014 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1015 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1016 int socket_id, unsigned flags)
1019 struct rte_mempool *mp;
1021 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1022 private_data_size, socket_id, flags);
1027 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
1028 * set the correct index into the table of ops structs.
1030 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
1031 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
1032 else if (flags & MEMPOOL_F_SP_PUT)
1033 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
1034 else if (flags & MEMPOOL_F_SC_GET)
1035 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
1037 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
1042 /* call the mempool priv initializer */
1044 mp_init(mp, mp_init_arg);
1046 if (rte_mempool_populate_default(mp) < 0)
1049 /* call the object initializers */
1051 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1056 rte_mempool_free(mp);
1061 * Create the mempool over already allocated chunk of memory.
1062 * That external memory buffer can consists of physically disjoint pages.
1063 * Setting vaddr to NULL, makes mempool to fallback to rte_mempool_create()
1066 struct rte_mempool *
1067 rte_mempool_xmem_create(const char *name, unsigned n, unsigned elt_size,
1068 unsigned cache_size, unsigned private_data_size,
1069 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1070 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1071 int socket_id, unsigned flags, void *vaddr,
1072 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift)
1074 struct rte_mempool *mp = NULL;
1077 /* no virtual address supplied, use rte_mempool_create() */
1079 return rte_mempool_create(name, n, elt_size, cache_size,
1080 private_data_size, mp_init, mp_init_arg,
1081 obj_init, obj_init_arg, socket_id, flags);
1083 /* check that we have both VA and PA */
1089 /* Check that pg_shift parameter is valid. */
1090 if (pg_shift > MEMPOOL_PG_SHIFT_MAX) {
1095 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1096 private_data_size, socket_id, flags);
1100 /* call the mempool priv initializer */
1102 mp_init(mp, mp_init_arg);
1104 ret = rte_mempool_populate_iova_tab(mp, vaddr, iova, pg_num, pg_shift,
1106 if (ret < 0 || ret != (int)mp->size)
1109 /* call the object initializers */
1111 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1116 rte_mempool_free(mp);
1120 /* Return the number of entries in the mempool */
1122 rte_mempool_avail_count(const struct rte_mempool *mp)
1127 count = rte_mempool_ops_get_count(mp);
1129 if (mp->cache_size == 0)
1132 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1133 count += mp->local_cache[lcore_id].len;
1136 * due to race condition (access to len is not locked), the
1137 * total can be greater than size... so fix the result
1139 if (count > mp->size)
1144 /* return the number of entries allocated from the mempool */
1146 rte_mempool_in_use_count(const struct rte_mempool *mp)
1148 return mp->size - rte_mempool_avail_count(mp);
1151 /* dump the cache status */
1153 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1157 unsigned cache_count;
1159 fprintf(f, " internal cache infos:\n");
1160 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1162 if (mp->cache_size == 0)
1165 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1166 cache_count = mp->local_cache[lcore_id].len;
1167 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1168 lcore_id, cache_count);
1169 count += cache_count;
1171 fprintf(f, " total_cache_count=%u\n", count);
1175 #ifndef __INTEL_COMPILER
1176 #pragma GCC diagnostic ignored "-Wcast-qual"
1179 /* check and update cookies or panic (internal) */
1180 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1181 void * const *obj_table_const, unsigned n, int free)
1183 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1184 struct rte_mempool_objhdr *hdr;
1185 struct rte_mempool_objtlr *tlr;
1191 /* Force to drop the "const" attribute. This is done only when
1192 * DEBUG is enabled */
1193 tmp = (void *) obj_table_const;
1199 if (rte_mempool_from_obj(obj) != mp)
1200 rte_panic("MEMPOOL: object is owned by another "
1203 hdr = __mempool_get_header(obj);
1204 cookie = hdr->cookie;
1207 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1208 RTE_LOG(CRIT, MEMPOOL,
1209 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1210 obj, (const void *) mp, cookie);
1211 rte_panic("MEMPOOL: bad header cookie (put)\n");
1213 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1214 } else if (free == 1) {
1215 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1216 RTE_LOG(CRIT, MEMPOOL,
1217 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1218 obj, (const void *) mp, cookie);
1219 rte_panic("MEMPOOL: bad header cookie (get)\n");
1221 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1222 } else if (free == 2) {
1223 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1224 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1225 RTE_LOG(CRIT, MEMPOOL,
1226 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1227 obj, (const void *) mp, cookie);
1228 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1231 tlr = __mempool_get_trailer(obj);
1232 cookie = tlr->cookie;
1233 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1234 RTE_LOG(CRIT, MEMPOOL,
1235 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1236 obj, (const void *) mp, cookie);
1237 rte_panic("MEMPOOL: bad trailer cookie\n");
1242 RTE_SET_USED(obj_table_const);
1248 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1250 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1251 void *obj, __rte_unused unsigned idx)
1253 __mempool_check_cookies(mp, &obj, 1, 2);
1257 mempool_audit_cookies(struct rte_mempool *mp)
1261 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1262 if (num != mp->size) {
1263 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1264 "iterated only over %u elements\n",
1269 #define mempool_audit_cookies(mp) do {} while(0)
1272 #ifndef __INTEL_COMPILER
1273 #pragma GCC diagnostic error "-Wcast-qual"
1276 /* check cookies before and after objects */
1278 mempool_audit_cache(const struct rte_mempool *mp)
1280 /* check cache size consistency */
1283 if (mp->cache_size == 0)
1286 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1287 const struct rte_mempool_cache *cache;
1288 cache = &mp->local_cache[lcore_id];
1289 if (cache->len > cache->flushthresh) {
1290 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1292 rte_panic("MEMPOOL: invalid cache len\n");
1297 /* check the consistency of mempool (size, cookies, ...) */
1299 rte_mempool_audit(struct rte_mempool *mp)
1301 mempool_audit_cache(mp);
1302 mempool_audit_cookies(mp);
1304 /* For case where mempool DEBUG is not set, and cache size is 0 */
1308 /* dump the status of the mempool on the console */
1310 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1312 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1313 struct rte_mempool_debug_stats sum;
1316 struct rte_mempool_memhdr *memhdr;
1317 unsigned common_count;
1318 unsigned cache_count;
1321 RTE_ASSERT(f != NULL);
1322 RTE_ASSERT(mp != NULL);
1324 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1325 fprintf(f, " flags=%x\n", mp->flags);
1326 fprintf(f, " pool=%p\n", mp->pool_data);
1327 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1328 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1329 fprintf(f, " size=%"PRIu32"\n", mp->size);
1330 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1331 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1332 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1333 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1334 fprintf(f, " total_obj_size=%"PRIu32"\n",
1335 mp->header_size + mp->elt_size + mp->trailer_size);
1337 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1339 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1340 mem_len += memhdr->len;
1342 fprintf(f, " avg bytes/object=%#Lf\n",
1343 (long double)mem_len / mp->size);
1346 cache_count = rte_mempool_dump_cache(f, mp);
1347 common_count = rte_mempool_ops_get_count(mp);
1348 if ((cache_count + common_count) > mp->size)
1349 common_count = mp->size - cache_count;
1350 fprintf(f, " common_pool_count=%u\n", common_count);
1352 /* sum and dump statistics */
1353 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1354 memset(&sum, 0, sizeof(sum));
1355 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1356 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1357 sum.put_objs += mp->stats[lcore_id].put_objs;
1358 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1359 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1360 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1361 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1363 fprintf(f, " stats:\n");
1364 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1365 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1366 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1367 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1368 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1369 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1371 fprintf(f, " no statistics available\n");
1374 rte_mempool_audit(mp);
1377 /* dump the status of all mempools on the console */
1379 rte_mempool_list_dump(FILE *f)
1381 struct rte_mempool *mp = NULL;
1382 struct rte_tailq_entry *te;
1383 struct rte_mempool_list *mempool_list;
1385 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1387 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1389 TAILQ_FOREACH(te, mempool_list, next) {
1390 mp = (struct rte_mempool *) te->data;
1391 rte_mempool_dump(f, mp);
1394 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1397 /* search a mempool from its name */
1398 struct rte_mempool *
1399 rte_mempool_lookup(const char *name)
1401 struct rte_mempool *mp = NULL;
1402 struct rte_tailq_entry *te;
1403 struct rte_mempool_list *mempool_list;
1405 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1407 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1409 TAILQ_FOREACH(te, mempool_list, next) {
1410 mp = (struct rte_mempool *) te->data;
1411 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1415 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1425 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1428 struct rte_tailq_entry *te = NULL;
1429 struct rte_mempool_list *mempool_list;
1432 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1434 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1436 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1437 (*func)((struct rte_mempool *) te->data, arg);
1440 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);