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 get_min_page_size(void)
105 const struct rte_mem_config *mcfg =
106 rte_eal_get_configuration()->mem_config;
108 size_t min_pagesz = SIZE_MAX;
110 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
111 const struct rte_memseg *ms = &mcfg->memseg[i];
113 if (ms->addr == NULL)
116 if (ms->hugepage_sz < min_pagesz)
117 min_pagesz = ms->hugepage_sz;
120 return min_pagesz == SIZE_MAX ? (size_t) getpagesize() : min_pagesz;
124 mempool_add_elem(struct rte_mempool *mp, void *obj, rte_iova_t iova)
126 struct rte_mempool_objhdr *hdr;
127 struct rte_mempool_objtlr *tlr __rte_unused;
129 /* set mempool ptr in header */
130 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
133 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
134 mp->populated_size++;
136 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
137 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
138 tlr = __mempool_get_trailer(obj);
139 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
142 /* enqueue in ring */
143 rte_mempool_ops_enqueue_bulk(mp, &obj, 1);
146 /* call obj_cb() for each mempool element */
148 rte_mempool_obj_iter(struct rte_mempool *mp,
149 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
151 struct rte_mempool_objhdr *hdr;
155 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
156 obj = (char *)hdr + sizeof(*hdr);
157 obj_cb(mp, obj_cb_arg, obj, n);
164 /* call mem_cb() for each mempool memory chunk */
166 rte_mempool_mem_iter(struct rte_mempool *mp,
167 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
169 struct rte_mempool_memhdr *hdr;
172 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
173 mem_cb(mp, mem_cb_arg, hdr, n);
180 /* get the header, trailer and total size of a mempool element. */
182 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
183 struct rte_mempool_objsz *sz)
185 struct rte_mempool_objsz lsz;
187 sz = (sz != NULL) ? sz : &lsz;
189 sz->header_size = sizeof(struct rte_mempool_objhdr);
190 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
191 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
194 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
195 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
197 sz->trailer_size = 0;
200 /* element size is 8 bytes-aligned at least */
201 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
203 /* expand trailer to next cache line */
204 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
205 sz->total_size = sz->header_size + sz->elt_size +
207 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
208 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
209 RTE_MEMPOOL_ALIGN_MASK);
213 * increase trailer to add padding between objects in order to
214 * spread them across memory channels/ranks
216 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
218 new_size = optimize_object_size(sz->header_size + sz->elt_size +
220 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
223 /* this is the size of an object, including header and trailer */
224 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
226 return sz->total_size;
231 * Calculate maximum amount of memory required to store given number of objects.
234 rte_mempool_xmem_size(uint32_t elt_num, size_t total_elt_sz, uint32_t pg_shift,
237 size_t obj_per_page, pg_num, pg_sz;
240 mask = MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS | MEMPOOL_F_CAPA_PHYS_CONTIG;
241 if ((flags & mask) == mask)
242 /* alignment need one additional object */
245 if (total_elt_sz == 0)
249 return total_elt_sz * elt_num;
251 pg_sz = (size_t)1 << pg_shift;
252 obj_per_page = pg_sz / total_elt_sz;
253 if (obj_per_page == 0)
254 return RTE_ALIGN_CEIL(total_elt_sz, pg_sz) * elt_num;
256 pg_num = (elt_num + obj_per_page - 1) / obj_per_page;
257 return pg_num << pg_shift;
261 * Calculate how much memory would be actually required with the
262 * given memory footprint to store required number of elements.
265 rte_mempool_xmem_usage(__rte_unused void *vaddr, uint32_t elt_num,
266 size_t total_elt_sz, const rte_iova_t iova[], uint32_t pg_num,
267 uint32_t pg_shift, unsigned int flags)
269 uint32_t elt_cnt = 0;
270 rte_iova_t start, end;
272 size_t pg_sz = (size_t)1 << pg_shift;
275 mask = MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS | MEMPOOL_F_CAPA_PHYS_CONTIG;
276 if ((flags & mask) == mask)
277 /* alignment need one additional object */
280 /* if iova is NULL, assume contiguous memory */
283 end = pg_sz * pg_num;
287 end = iova[0] + pg_sz;
290 while (elt_cnt < elt_num) {
292 if (end - start >= total_elt_sz) {
293 /* enough contiguous memory, add an object */
294 start += total_elt_sz;
296 } else if (iova_idx < pg_num) {
297 /* no room to store one obj, add a page */
298 if (end == iova[iova_idx]) {
301 start = iova[iova_idx];
302 end = iova[iova_idx] + pg_sz;
307 /* no more page, return how many elements fit */
308 return -(size_t)elt_cnt;
312 return (size_t)iova_idx << pg_shift;
315 /* free a memchunk allocated with rte_memzone_reserve() */
317 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
320 const struct rte_memzone *mz = opaque;
321 rte_memzone_free(mz);
324 /* Free memory chunks used by a mempool. Objects must be in pool */
326 rte_mempool_free_memchunks(struct rte_mempool *mp)
328 struct rte_mempool_memhdr *memhdr;
331 while (!STAILQ_EMPTY(&mp->elt_list)) {
332 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
334 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
335 mp->populated_size--;
338 while (!STAILQ_EMPTY(&mp->mem_list)) {
339 memhdr = STAILQ_FIRST(&mp->mem_list);
340 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
341 if (memhdr->free_cb != NULL)
342 memhdr->free_cb(memhdr, memhdr->opaque);
348 /* Add objects in the pool, using a physically contiguous memory
349 * zone. Return the number of objects added, or a negative value
353 rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
354 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
357 unsigned total_elt_sz;
358 unsigned int mp_capa_flags;
361 struct rte_mempool_memhdr *memhdr;
364 /* create the internal ring if not already done */
365 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
366 ret = rte_mempool_ops_alloc(mp);
369 mp->flags |= MEMPOOL_F_POOL_CREATED;
372 /* Notify memory area to mempool */
373 ret = rte_mempool_ops_register_memory_area(mp, vaddr, iova, len);
374 if (ret != -ENOTSUP && ret < 0)
377 /* mempool is already populated */
378 if (mp->populated_size >= mp->size)
381 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
383 /* Get mempool capabilities */
385 ret = rte_mempool_ops_get_capabilities(mp, &mp_capa_flags);
386 if ((ret < 0) && (ret != -ENOTSUP))
389 /* update mempool capabilities */
390 mp->flags |= mp_capa_flags;
392 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
397 memhdr->addr = vaddr;
400 memhdr->free_cb = free_cb;
401 memhdr->opaque = opaque;
403 if (mp_capa_flags & MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS)
404 /* align object start address to a multiple of total_elt_sz */
405 off = total_elt_sz - ((uintptr_t)vaddr % total_elt_sz);
406 else if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
407 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
409 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) - vaddr;
411 while (off + total_elt_sz <= len && mp->populated_size < mp->size) {
412 off += mp->header_size;
413 if (iova == RTE_BAD_IOVA)
414 mempool_add_elem(mp, (char *)vaddr + off,
417 mempool_add_elem(mp, (char *)vaddr + off, iova + off);
418 off += mp->elt_size + mp->trailer_size;
422 /* not enough room to store one object */
426 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
432 rte_mempool_populate_phys(struct rte_mempool *mp, char *vaddr,
433 phys_addr_t paddr, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
436 return rte_mempool_populate_iova(mp, vaddr, paddr, len, free_cb, opaque);
439 /* Add objects in the pool, using a table of physical pages. Return the
440 * number of objects added, or a negative value on error.
443 rte_mempool_populate_iova_tab(struct rte_mempool *mp, char *vaddr,
444 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift,
445 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
449 size_t pg_sz = (size_t)1 << pg_shift;
451 /* mempool must not be populated */
452 if (mp->nb_mem_chunks != 0)
455 if (mp->flags & MEMPOOL_F_NO_PHYS_CONTIG)
456 return rte_mempool_populate_iova(mp, vaddr, RTE_BAD_IOVA,
457 pg_num * pg_sz, free_cb, opaque);
459 for (i = 0; i < pg_num && mp->populated_size < mp->size; i += n) {
461 /* populate with the largest group of contiguous pages */
462 for (n = 1; (i + n) < pg_num &&
463 iova[i + n - 1] + pg_sz == iova[i + n]; n++)
466 ret = rte_mempool_populate_iova(mp, vaddr + i * pg_sz,
467 iova[i], n * pg_sz, free_cb, opaque);
469 rte_mempool_free_memchunks(mp);
472 /* no need to call the free callback for next chunks */
480 rte_mempool_populate_phys_tab(struct rte_mempool *mp, char *vaddr,
481 const phys_addr_t paddr[], uint32_t pg_num, uint32_t pg_shift,
482 rte_mempool_memchunk_free_cb_t *free_cb, void *opaque)
484 return rte_mempool_populate_iova_tab(mp, vaddr, paddr, pg_num, pg_shift,
488 /* Populate the mempool with a virtual area. Return the number of
489 * objects added, or a negative value on error.
492 rte_mempool_populate_virt(struct rte_mempool *mp, char *addr,
493 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
497 size_t off, phys_len;
500 /* mempool must not be populated */
501 if (mp->nb_mem_chunks != 0)
503 /* address and len must be page-aligned */
504 if (RTE_PTR_ALIGN_CEIL(addr, pg_sz) != addr)
506 if (RTE_ALIGN_CEIL(len, pg_sz) != len)
509 if (mp->flags & MEMPOOL_F_NO_PHYS_CONTIG)
510 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
511 len, free_cb, opaque);
513 for (off = 0; off + pg_sz <= len &&
514 mp->populated_size < mp->size; off += phys_len) {
516 iova = rte_mem_virt2iova(addr + off);
518 if (iova == RTE_BAD_IOVA && rte_eal_has_hugepages()) {
523 /* populate with the largest group of contiguous pages */
524 for (phys_len = pg_sz; off + phys_len < len; phys_len += pg_sz) {
527 iova_tmp = rte_mem_virt2iova(addr + off + phys_len);
529 if (iova_tmp != iova + phys_len)
533 ret = rte_mempool_populate_iova(mp, addr + off, iova,
534 phys_len, free_cb, opaque);
537 /* no need to call the free callback for next chunks */
545 rte_mempool_free_memchunks(mp);
549 /* Default function to populate the mempool: allocate memory in memzones,
550 * and populate them. Return the number of objects added, or a negative
554 rte_mempool_populate_default(struct rte_mempool *mp)
556 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
557 char mz_name[RTE_MEMZONE_NAMESIZE];
558 const struct rte_memzone *mz;
559 size_t size, total_elt_sz, align, pg_sz, pg_shift;
562 unsigned int mp_flags;
564 bool force_contig, no_contig, try_contig, no_pageshift;
566 /* mempool must not be populated */
567 if (mp->nb_mem_chunks != 0)
570 /* Get mempool capabilities */
572 ret = rte_mempool_ops_get_capabilities(mp, &mp_flags);
573 if ((ret < 0) && (ret != -ENOTSUP))
576 /* update mempool capabilities */
577 mp->flags |= mp_flags;
579 no_contig = mp->flags & MEMPOOL_F_NO_PHYS_CONTIG;
580 force_contig = mp->flags & MEMPOOL_F_CAPA_PHYS_CONTIG;
583 * the following section calculates page shift and page size values.
585 * these values impact the result of rte_mempool_xmem_size(), which
586 * returns the amount of memory that should be allocated to store the
587 * desired number of objects. when not zero, it allocates more memory
588 * for the padding between objects, to ensure that an object does not
589 * cross a page boundary. in other words, page size/shift are to be set
590 * to zero if mempool elements won't care about page boundaries.
591 * there are several considerations for page size and page shift here.
593 * if we don't need our mempools to have physically contiguous objects,
594 * then just set page shift and page size to 0, because the user has
595 * indicated that there's no need to care about anything.
597 * if we do need contiguous objects, there is also an option to reserve
598 * the entire mempool memory as one contiguous block of memory, in
599 * which case the page shift and alignment wouldn't matter as well.
601 * if we require contiguous objects, but not necessarily the entire
602 * mempool reserved space to be contiguous, then there are two options.
604 * if our IO addresses are virtual, not actual physical (IOVA as VA
605 * case), then no page shift needed - our memory allocation will give us
606 * contiguous physical memory as far as the hardware is concerned, so
607 * act as if we're getting contiguous memory.
609 * if our IO addresses are physical, we may get memory from bigger
610 * pages, or we might get memory from smaller pages, and how much of it
611 * we require depends on whether we want bigger or smaller pages.
612 * However, requesting each and every memory size is too much work, so
613 * what we'll do instead is walk through the page sizes available, pick
614 * the smallest one and set up page shift to match that one. We will be
615 * wasting some space this way, but it's much nicer than looping around
616 * trying to reserve each and every page size.
618 * However, since size calculation will produce page-aligned sizes, it
619 * makes sense to first try and see if we can reserve the entire memzone
620 * in one contiguous chunk as well (otherwise we might end up wasting a
621 * 1G page on a 10MB memzone). If we fail to get enough contiguous
622 * memory, then we'll go and reserve space page-by-page.
624 no_pageshift = no_contig || force_contig ||
625 rte_eal_iova_mode() == RTE_IOVA_VA;
626 try_contig = !no_contig && !no_pageshift && rte_eal_has_hugepages();
628 mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
633 align = RTE_CACHE_LINE_SIZE;
634 } else if (try_contig) {
635 pg_sz = get_min_page_size();
636 pg_shift = rte_bsf32(pg_sz);
637 /* we're trying to reserve contiguous memzone first, so try
638 * align to cache line; if we fail to reserve a contiguous
639 * memzone, we'll adjust alignment to equal pagesize later.
641 align = RTE_CACHE_LINE_SIZE;
643 pg_sz = getpagesize();
644 pg_shift = rte_bsf32(pg_sz);
648 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
649 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
651 if (try_contig || no_pageshift)
652 size = rte_mempool_xmem_size(n, total_elt_sz, 0,
655 size = rte_mempool_xmem_size(n, total_elt_sz, pg_shift,
658 ret = snprintf(mz_name, sizeof(mz_name),
659 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
660 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
667 /* if we're trying to reserve contiguous memory, add appropriate
671 flags |= RTE_MEMZONE_IOVA_CONTIG;
673 mz = rte_memzone_reserve_aligned(mz_name, size, mp->socket_id,
676 /* if we were trying to allocate contiguous memory, adjust
677 * memzone size and page size to fit smaller page sizes, and
680 if (mz == NULL && try_contig) {
682 flags &= ~RTE_MEMZONE_IOVA_CONTIG;
684 size = rte_mempool_xmem_size(n, total_elt_sz,
685 pg_shift, mp->flags);
687 mz = rte_memzone_reserve_aligned(mz_name, size,
688 mp->socket_id, flags, align);
690 /* don't try reserving with 0 size if we were asked to reserve
691 * IOVA-contiguous memory.
693 if (!force_contig && mz == NULL) {
694 /* not enough memory, retry with the biggest zone we
697 mz = rte_memzone_reserve_aligned(mz_name, 0,
698 mp->socket_id, flags, align);
710 if (no_pageshift || try_contig)
711 ret = rte_mempool_populate_iova(mp, mz->addr,
713 rte_mempool_memchunk_mz_free,
714 (void *)(uintptr_t)mz);
716 ret = rte_mempool_populate_virt(mp, mz->addr,
718 rte_mempool_memchunk_mz_free,
719 (void *)(uintptr_t)mz);
721 rte_memzone_free(mz);
729 rte_mempool_free_memchunks(mp);
733 /* return the memory size required for mempool objects in anonymous mem */
735 get_anon_size(const struct rte_mempool *mp)
737 size_t size, total_elt_sz, pg_sz, pg_shift;
739 pg_sz = getpagesize();
740 pg_shift = rte_bsf32(pg_sz);
741 total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
742 size = rte_mempool_xmem_size(mp->size, total_elt_sz, pg_shift,
748 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
750 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
753 munmap(opaque, get_anon_size(memhdr->mp));
756 /* populate the mempool with an anonymous mapping */
758 rte_mempool_populate_anon(struct rte_mempool *mp)
764 /* mempool is already populated, error */
765 if (!STAILQ_EMPTY(&mp->mem_list)) {
770 /* get chunk of virtually continuous memory */
771 size = get_anon_size(mp);
772 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
773 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
774 if (addr == MAP_FAILED) {
778 /* can't use MMAP_LOCKED, it does not exist on BSD */
779 if (mlock(addr, size) < 0) {
785 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
786 rte_mempool_memchunk_anon_free, addr);
790 return mp->populated_size;
793 rte_mempool_free_memchunks(mp);
799 rte_mempool_free(struct rte_mempool *mp)
801 struct rte_mempool_list *mempool_list = NULL;
802 struct rte_tailq_entry *te;
807 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
808 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
809 /* find out tailq entry */
810 TAILQ_FOREACH(te, mempool_list, next) {
811 if (te->data == (void *)mp)
816 TAILQ_REMOVE(mempool_list, te, next);
819 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
821 rte_mempool_free_memchunks(mp);
822 rte_mempool_ops_free(mp);
823 rte_memzone_free(mp->mz);
827 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
830 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
835 * Create and initialize a cache for objects that are retrieved from and
836 * returned to an underlying mempool. This structure is identical to the
837 * local_cache[lcore_id] pointed to by the mempool structure.
839 struct rte_mempool_cache *
840 rte_mempool_cache_create(uint32_t size, int socket_id)
842 struct rte_mempool_cache *cache;
844 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
849 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
850 RTE_CACHE_LINE_SIZE, socket_id);
852 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
857 mempool_cache_init(cache, size);
863 * Free a cache. It's the responsibility of the user to make sure that any
864 * remaining objects in the cache are flushed to the corresponding
868 rte_mempool_cache_free(struct rte_mempool_cache *cache)
873 /* create an empty mempool */
875 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
876 unsigned cache_size, unsigned private_data_size,
877 int socket_id, unsigned flags)
879 char mz_name[RTE_MEMZONE_NAMESIZE];
880 struct rte_mempool_list *mempool_list;
881 struct rte_mempool *mp = NULL;
882 struct rte_tailq_entry *te = NULL;
883 const struct rte_memzone *mz = NULL;
885 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
886 struct rte_mempool_objsz objsz;
890 /* compilation-time checks */
891 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
892 RTE_CACHE_LINE_MASK) != 0);
893 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
894 RTE_CACHE_LINE_MASK) != 0);
895 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
896 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
897 RTE_CACHE_LINE_MASK) != 0);
898 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
899 RTE_CACHE_LINE_MASK) != 0);
902 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
904 /* asked cache too big */
905 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
906 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
911 /* "no cache align" imply "no spread" */
912 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
913 flags |= MEMPOOL_F_NO_SPREAD;
915 /* calculate mempool object sizes. */
916 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
921 rte_rwlock_write_lock(RTE_EAL_MEMPOOL_RWLOCK);
924 * reserve a memory zone for this mempool: private data is
927 private_data_size = (private_data_size +
928 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
931 /* try to allocate tailq entry */
932 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
934 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
938 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
939 mempool_size += private_data_size;
940 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
942 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
943 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
944 rte_errno = ENAMETOOLONG;
948 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
952 /* init the mempool structure */
954 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
955 ret = snprintf(mp->name, sizeof(mp->name), "%s", name);
956 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
957 rte_errno = ENAMETOOLONG;
963 mp->socket_id = socket_id;
964 mp->elt_size = objsz.elt_size;
965 mp->header_size = objsz.header_size;
966 mp->trailer_size = objsz.trailer_size;
967 /* Size of default caches, zero means disabled. */
968 mp->cache_size = cache_size;
969 mp->private_data_size = private_data_size;
970 STAILQ_INIT(&mp->elt_list);
971 STAILQ_INIT(&mp->mem_list);
974 * local_cache pointer is set even if cache_size is zero.
975 * The local_cache points to just past the elt_pa[] array.
977 mp->local_cache = (struct rte_mempool_cache *)
978 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
980 /* Init all default caches. */
981 if (cache_size != 0) {
982 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
983 mempool_cache_init(&mp->local_cache[lcore_id],
989 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
990 TAILQ_INSERT_TAIL(mempool_list, te, next);
991 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
992 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
997 rte_rwlock_write_unlock(RTE_EAL_MEMPOOL_RWLOCK);
999 rte_mempool_free(mp);
1003 /* create the mempool */
1004 struct rte_mempool *
1005 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
1006 unsigned cache_size, unsigned private_data_size,
1007 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1008 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1009 int socket_id, unsigned flags)
1012 struct rte_mempool *mp;
1014 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1015 private_data_size, socket_id, flags);
1020 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
1021 * set the correct index into the table of ops structs.
1023 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
1024 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
1025 else if (flags & MEMPOOL_F_SP_PUT)
1026 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
1027 else if (flags & MEMPOOL_F_SC_GET)
1028 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
1030 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
1035 /* call the mempool priv initializer */
1037 mp_init(mp, mp_init_arg);
1039 if (rte_mempool_populate_default(mp) < 0)
1042 /* call the object initializers */
1044 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1049 rte_mempool_free(mp);
1054 * Create the mempool over already allocated chunk of memory.
1055 * That external memory buffer can consists of physically disjoint pages.
1056 * Setting vaddr to NULL, makes mempool to fallback to rte_mempool_create()
1059 struct rte_mempool *
1060 rte_mempool_xmem_create(const char *name, unsigned n, unsigned elt_size,
1061 unsigned cache_size, unsigned private_data_size,
1062 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
1063 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
1064 int socket_id, unsigned flags, void *vaddr,
1065 const rte_iova_t iova[], uint32_t pg_num, uint32_t pg_shift)
1067 struct rte_mempool *mp = NULL;
1070 /* no virtual address supplied, use rte_mempool_create() */
1072 return rte_mempool_create(name, n, elt_size, cache_size,
1073 private_data_size, mp_init, mp_init_arg,
1074 obj_init, obj_init_arg, socket_id, flags);
1076 /* check that we have both VA and PA */
1082 /* Check that pg_shift parameter is valid. */
1083 if (pg_shift > MEMPOOL_PG_SHIFT_MAX) {
1088 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1089 private_data_size, socket_id, flags);
1093 /* call the mempool priv initializer */
1095 mp_init(mp, mp_init_arg);
1097 ret = rte_mempool_populate_iova_tab(mp, vaddr, iova, pg_num, pg_shift,
1099 if (ret < 0 || ret != (int)mp->size)
1102 /* call the object initializers */
1104 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1109 rte_mempool_free(mp);
1113 /* Return the number of entries in the mempool */
1115 rte_mempool_avail_count(const struct rte_mempool *mp)
1120 count = rte_mempool_ops_get_count(mp);
1122 if (mp->cache_size == 0)
1125 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1126 count += mp->local_cache[lcore_id].len;
1129 * due to race condition (access to len is not locked), the
1130 * total can be greater than size... so fix the result
1132 if (count > mp->size)
1137 /* return the number of entries allocated from the mempool */
1139 rte_mempool_in_use_count(const struct rte_mempool *mp)
1141 return mp->size - rte_mempool_avail_count(mp);
1144 /* dump the cache status */
1146 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1150 unsigned cache_count;
1152 fprintf(f, " internal cache infos:\n");
1153 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1155 if (mp->cache_size == 0)
1158 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1159 cache_count = mp->local_cache[lcore_id].len;
1160 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1161 lcore_id, cache_count);
1162 count += cache_count;
1164 fprintf(f, " total_cache_count=%u\n", count);
1168 #ifndef __INTEL_COMPILER
1169 #pragma GCC diagnostic ignored "-Wcast-qual"
1172 /* check and update cookies or panic (internal) */
1173 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1174 void * const *obj_table_const, unsigned n, int free)
1176 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1177 struct rte_mempool_objhdr *hdr;
1178 struct rte_mempool_objtlr *tlr;
1184 /* Force to drop the "const" attribute. This is done only when
1185 * DEBUG is enabled */
1186 tmp = (void *) obj_table_const;
1192 if (rte_mempool_from_obj(obj) != mp)
1193 rte_panic("MEMPOOL: object is owned by another "
1196 hdr = __mempool_get_header(obj);
1197 cookie = hdr->cookie;
1200 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1201 RTE_LOG(CRIT, MEMPOOL,
1202 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1203 obj, (const void *) mp, cookie);
1204 rte_panic("MEMPOOL: bad header cookie (put)\n");
1206 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1207 } else if (free == 1) {
1208 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1209 RTE_LOG(CRIT, MEMPOOL,
1210 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1211 obj, (const void *) mp, cookie);
1212 rte_panic("MEMPOOL: bad header cookie (get)\n");
1214 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1215 } else if (free == 2) {
1216 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1217 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1218 RTE_LOG(CRIT, MEMPOOL,
1219 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1220 obj, (const void *) mp, cookie);
1221 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1224 tlr = __mempool_get_trailer(obj);
1225 cookie = tlr->cookie;
1226 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1227 RTE_LOG(CRIT, MEMPOOL,
1228 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1229 obj, (const void *) mp, cookie);
1230 rte_panic("MEMPOOL: bad trailer cookie\n");
1235 RTE_SET_USED(obj_table_const);
1241 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1243 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1244 void *obj, __rte_unused unsigned idx)
1246 __mempool_check_cookies(mp, &obj, 1, 2);
1250 mempool_audit_cookies(struct rte_mempool *mp)
1254 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1255 if (num != mp->size) {
1256 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1257 "iterated only over %u elements\n",
1262 #define mempool_audit_cookies(mp) do {} while(0)
1265 #ifndef __INTEL_COMPILER
1266 #pragma GCC diagnostic error "-Wcast-qual"
1269 /* check cookies before and after objects */
1271 mempool_audit_cache(const struct rte_mempool *mp)
1273 /* check cache size consistency */
1276 if (mp->cache_size == 0)
1279 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1280 const struct rte_mempool_cache *cache;
1281 cache = &mp->local_cache[lcore_id];
1282 if (cache->len > cache->flushthresh) {
1283 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1285 rte_panic("MEMPOOL: invalid cache len\n");
1290 /* check the consistency of mempool (size, cookies, ...) */
1292 rte_mempool_audit(struct rte_mempool *mp)
1294 mempool_audit_cache(mp);
1295 mempool_audit_cookies(mp);
1297 /* For case where mempool DEBUG is not set, and cache size is 0 */
1301 /* dump the status of the mempool on the console */
1303 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1305 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1306 struct rte_mempool_debug_stats sum;
1309 struct rte_mempool_memhdr *memhdr;
1310 unsigned common_count;
1311 unsigned cache_count;
1314 RTE_ASSERT(f != NULL);
1315 RTE_ASSERT(mp != NULL);
1317 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1318 fprintf(f, " flags=%x\n", mp->flags);
1319 fprintf(f, " pool=%p\n", mp->pool_data);
1320 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1321 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1322 fprintf(f, " size=%"PRIu32"\n", mp->size);
1323 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1324 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1325 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1326 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1327 fprintf(f, " total_obj_size=%"PRIu32"\n",
1328 mp->header_size + mp->elt_size + mp->trailer_size);
1330 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1332 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1333 mem_len += memhdr->len;
1335 fprintf(f, " avg bytes/object=%#Lf\n",
1336 (long double)mem_len / mp->size);
1339 cache_count = rte_mempool_dump_cache(f, mp);
1340 common_count = rte_mempool_ops_get_count(mp);
1341 if ((cache_count + common_count) > mp->size)
1342 common_count = mp->size - cache_count;
1343 fprintf(f, " common_pool_count=%u\n", common_count);
1345 /* sum and dump statistics */
1346 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1347 memset(&sum, 0, sizeof(sum));
1348 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1349 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1350 sum.put_objs += mp->stats[lcore_id].put_objs;
1351 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1352 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1353 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1354 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1356 fprintf(f, " stats:\n");
1357 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1358 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1359 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1360 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1361 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1362 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1364 fprintf(f, " no statistics available\n");
1367 rte_mempool_audit(mp);
1370 /* dump the status of all mempools on the console */
1372 rte_mempool_list_dump(FILE *f)
1374 struct rte_mempool *mp = NULL;
1375 struct rte_tailq_entry *te;
1376 struct rte_mempool_list *mempool_list;
1378 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1380 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1382 TAILQ_FOREACH(te, mempool_list, next) {
1383 mp = (struct rte_mempool *) te->data;
1384 rte_mempool_dump(f, mp);
1387 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1390 /* search a mempool from its name */
1391 struct rte_mempool *
1392 rte_mempool_lookup(const char *name)
1394 struct rte_mempool *mp = NULL;
1395 struct rte_tailq_entry *te;
1396 struct rte_mempool_list *mempool_list;
1398 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1400 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1402 TAILQ_FOREACH(te, mempool_list, next) {
1403 mp = (struct rte_mempool *) te->data;
1404 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1408 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);
1418 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1421 struct rte_tailq_entry *te = NULL;
1422 struct rte_mempool_list *mempool_list;
1425 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1427 rte_rwlock_read_lock(RTE_EAL_MEMPOOL_RWLOCK);
1429 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1430 (*func)((struct rte_mempool *) te->data, arg);
1433 rte_rwlock_read_unlock(RTE_EAL_MEMPOOL_RWLOCK);