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>
16 #include <rte_common.h>
18 #include <rte_debug.h>
19 #include <rte_memory.h>
20 #include <rte_memzone.h>
21 #include <rte_malloc.h>
22 #include <rte_atomic.h>
23 #include <rte_launch.h>
25 #include <rte_eal_memconfig.h>
26 #include <rte_per_lcore.h>
27 #include <rte_lcore.h>
28 #include <rte_branch_prediction.h>
29 #include <rte_errno.h>
30 #include <rte_string_fns.h>
31 #include <rte_spinlock.h>
32 #include <rte_tailq.h>
33 #include <rte_function_versioning.h>
34 #include <rte_eal_paging.h>
37 #include "rte_mempool.h"
38 #include "rte_mempool_trace.h"
40 TAILQ_HEAD(rte_mempool_list, rte_tailq_entry);
42 static struct rte_tailq_elem rte_mempool_tailq = {
43 .name = "RTE_MEMPOOL",
45 EAL_REGISTER_TAILQ(rte_mempool_tailq)
47 #define CACHE_FLUSHTHRESH_MULTIPLIER 1.5
48 #define CALC_CACHE_FLUSHTHRESH(c) \
49 ((typeof(c))((c) * CACHE_FLUSHTHRESH_MULTIPLIER))
51 #if defined(RTE_ARCH_X86)
53 * return the greatest common divisor between a and b (fast algorithm)
56 static unsigned get_gcd(unsigned a, unsigned b)
81 * Depending on memory configuration on x86 arch, objects addresses are spread
82 * between channels and ranks in RAM: the pool allocator will add
83 * padding between objects. This function return the new size of the
87 arch_mem_object_align(unsigned int obj_size)
89 unsigned nrank, nchan;
90 unsigned new_obj_size;
92 /* get number of channels */
93 nchan = rte_memory_get_nchannel();
97 nrank = rte_memory_get_nrank();
101 /* process new object size */
102 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
103 while (get_gcd(new_obj_size, nrank * nchan) != 1)
105 return new_obj_size * RTE_MEMPOOL_ALIGN;
109 arch_mem_object_align(unsigned int obj_size)
115 struct pagesz_walk_arg {
121 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
123 struct pagesz_walk_arg *wa = arg;
127 * we need to only look at page sizes available for a particular socket
128 * ID. so, we either need an exact match on socket ID (can match both
129 * native and external memory), or, if SOCKET_ID_ANY was specified as a
130 * socket ID argument, we must only look at native memory and ignore any
131 * page sizes associated with external memory.
133 valid = msl->socket_id == wa->socket_id;
134 valid |= wa->socket_id == SOCKET_ID_ANY && msl->external == 0;
136 if (valid && msl->page_sz < wa->min)
137 wa->min = msl->page_sz;
143 get_min_page_size(int socket_id)
145 struct pagesz_walk_arg wa;
148 wa.socket_id = socket_id;
150 rte_memseg_list_walk(find_min_pagesz, &wa);
152 return wa.min == SIZE_MAX ? (size_t) rte_mem_page_size() : wa.min;
157 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
158 void *obj, rte_iova_t iova)
160 struct rte_mempool_objhdr *hdr;
161 struct rte_mempool_objtlr *tlr __rte_unused;
163 /* set mempool ptr in header */
164 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
167 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
168 mp->populated_size++;
170 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
171 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
172 tlr = __mempool_get_trailer(obj);
173 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
177 /* call obj_cb() for each mempool element */
179 rte_mempool_obj_iter(struct rte_mempool *mp,
180 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
182 struct rte_mempool_objhdr *hdr;
186 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
187 obj = (char *)hdr + sizeof(*hdr);
188 obj_cb(mp, obj_cb_arg, obj, n);
195 /* call mem_cb() for each mempool memory chunk */
197 rte_mempool_mem_iter(struct rte_mempool *mp,
198 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
200 struct rte_mempool_memhdr *hdr;
203 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
204 mem_cb(mp, mem_cb_arg, hdr, n);
211 /* get the header, trailer and total size of a mempool element. */
213 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
214 struct rte_mempool_objsz *sz)
216 struct rte_mempool_objsz lsz;
218 sz = (sz != NULL) ? sz : &lsz;
220 sz->header_size = sizeof(struct rte_mempool_objhdr);
221 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
222 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
225 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
226 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
228 sz->trailer_size = 0;
231 /* element size is 8 bytes-aligned at least */
232 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
234 /* expand trailer to next cache line */
235 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
236 sz->total_size = sz->header_size + sz->elt_size +
238 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
239 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
240 RTE_MEMPOOL_ALIGN_MASK);
244 * increase trailer to add padding between objects in order to
245 * spread them across memory channels/ranks
247 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
249 new_size = arch_mem_object_align
250 (sz->header_size + sz->elt_size + sz->trailer_size);
251 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
254 /* this is the size of an object, including header and trailer */
255 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
257 return sz->total_size;
260 /* free a memchunk allocated with rte_memzone_reserve() */
262 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
265 const struct rte_memzone *mz = opaque;
266 rte_memzone_free(mz);
269 /* Free memory chunks used by a mempool. Objects must be in pool */
271 rte_mempool_free_memchunks(struct rte_mempool *mp)
273 struct rte_mempool_memhdr *memhdr;
276 while (!STAILQ_EMPTY(&mp->elt_list)) {
277 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
279 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
280 mp->populated_size--;
283 while (!STAILQ_EMPTY(&mp->mem_list)) {
284 memhdr = STAILQ_FIRST(&mp->mem_list);
285 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
286 if (memhdr->free_cb != NULL)
287 memhdr->free_cb(memhdr, memhdr->opaque);
294 mempool_ops_alloc_once(struct rte_mempool *mp)
298 /* create the internal ring if not already done */
299 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
300 ret = rte_mempool_ops_alloc(mp);
303 mp->flags |= MEMPOOL_F_POOL_CREATED;
309 rte_mempool_populate_iova_v21(struct rte_mempool *mp, char *vaddr,
310 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
313 /* Add objects in the pool, using a physically contiguous memory
314 * zone. Return the number of objects added, or a negative value
318 rte_mempool_populate_iova_v21(struct rte_mempool *mp, char *vaddr,
319 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
324 struct rte_mempool_memhdr *memhdr;
327 ret = mempool_ops_alloc_once(mp);
331 /* mempool is already populated */
332 if (mp->populated_size >= mp->size)
335 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
340 memhdr->addr = vaddr;
343 memhdr->free_cb = free_cb;
344 memhdr->opaque = opaque;
346 if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
347 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
349 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_MEMPOOL_ALIGN) - vaddr;
356 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
358 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
359 len - off, mempool_add_elem, NULL);
361 /* not enough room to store one object */
367 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
370 rte_mempool_trace_populate_iova(mp, vaddr, iova, len, free_cb, opaque);
378 BIND_DEFAULT_SYMBOL(rte_mempool_populate_iova, _v21, 21);
380 int rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
381 rte_iova_t iova, size_t len,
382 rte_mempool_memchunk_free_cb_t *free_cb,
384 rte_mempool_populate_iova_v21);
387 rte_mempool_populate_iova_v20(struct rte_mempool *mp, char *vaddr,
388 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
392 rte_mempool_populate_iova_v20(struct rte_mempool *mp, char *vaddr,
393 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
398 ret = rte_mempool_populate_iova_v21(mp, vaddr, iova, len, free_cb,
405 VERSION_SYMBOL(rte_mempool_populate_iova, _v20, 20.0);
410 struct rte_memseg *ms;
412 /* try registered memory first */
413 ms = rte_mem_virt2memseg(addr, NULL);
414 if (ms == NULL || ms->iova == RTE_BAD_IOVA)
415 /* fall back to actual physical address */
416 return rte_mem_virt2iova(addr);
417 return ms->iova + RTE_PTR_DIFF(addr, ms->addr);
421 rte_mempool_populate_virt_v21(struct rte_mempool *mp, char *addr,
422 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
425 /* Populate the mempool with a virtual area. Return the number of
426 * objects added, or a negative value on error.
429 rte_mempool_populate_virt_v21(struct rte_mempool *mp, char *addr,
430 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
434 size_t off, phys_len;
437 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
438 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
439 len, free_cb, opaque);
441 for (off = 0; off < len &&
442 mp->populated_size < mp->size; off += phys_len) {
444 iova = get_iova(addr + off);
446 /* populate with the largest group of contiguous pages */
447 for (phys_len = RTE_MIN(
448 (size_t)(RTE_PTR_ALIGN_CEIL(addr + off + 1, pg_sz) -
451 off + phys_len < len;
452 phys_len = RTE_MIN(phys_len + pg_sz, len - off)) {
455 iova_tmp = get_iova(addr + off + phys_len);
457 if (iova_tmp == RTE_BAD_IOVA ||
458 iova_tmp != iova + phys_len)
462 ret = rte_mempool_populate_iova_v21(mp, addr + off, iova,
463 phys_len, free_cb, opaque);
468 /* no need to call the free callback for next chunks */
473 rte_mempool_trace_populate_virt(mp, addr, len, pg_sz, free_cb, opaque);
477 rte_mempool_free_memchunks(mp);
480 BIND_DEFAULT_SYMBOL(rte_mempool_populate_virt, _v21, 21);
482 int rte_mempool_populate_virt(struct rte_mempool *mp,
483 char *addr, size_t len, size_t pg_sz,
484 rte_mempool_memchunk_free_cb_t *free_cb,
486 rte_mempool_populate_virt_v21);
489 rte_mempool_populate_virt_v20(struct rte_mempool *mp, char *addr,
490 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
494 rte_mempool_populate_virt_v20(struct rte_mempool *mp, char *addr,
495 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
500 ret = rte_mempool_populate_virt_v21(mp, addr, len, pg_sz,
508 VERSION_SYMBOL(rte_mempool_populate_virt, _v20, 20.0);
510 /* Get the minimal page size used in a mempool before populating it. */
512 rte_mempool_get_page_size(struct rte_mempool *mp, size_t *pg_sz)
514 bool need_iova_contig_obj;
515 bool alloc_in_ext_mem;
518 /* check if we can retrieve a valid socket ID */
519 ret = rte_malloc_heap_socket_is_external(mp->socket_id);
522 alloc_in_ext_mem = (ret == 1);
523 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
525 if (!need_iova_contig_obj)
527 else if (rte_eal_has_hugepages() || alloc_in_ext_mem)
528 *pg_sz = get_min_page_size(mp->socket_id);
530 *pg_sz = rte_mem_page_size();
532 rte_mempool_trace_get_page_size(mp, *pg_sz);
536 /* Default function to populate the mempool: allocate memory in memzones,
537 * and populate them. Return the number of objects added, or a negative
541 rte_mempool_populate_default(struct rte_mempool *mp)
543 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
544 char mz_name[RTE_MEMZONE_NAMESIZE];
545 const struct rte_memzone *mz;
547 size_t align, pg_sz, pg_shift = 0;
551 bool need_iova_contig_obj;
552 size_t max_alloc_size = SIZE_MAX;
554 ret = mempool_ops_alloc_once(mp);
558 /* mempool must not be populated */
559 if (mp->nb_mem_chunks != 0)
563 * the following section calculates page shift and page size values.
565 * these values impact the result of calc_mem_size operation, which
566 * returns the amount of memory that should be allocated to store the
567 * desired number of objects. when not zero, it allocates more memory
568 * for the padding between objects, to ensure that an object does not
569 * cross a page boundary. in other words, page size/shift are to be set
570 * to zero if mempool elements won't care about page boundaries.
571 * there are several considerations for page size and page shift here.
573 * if we don't need our mempools to have physically contiguous objects,
574 * then just set page shift and page size to 0, because the user has
575 * indicated that there's no need to care about anything.
577 * if we do need contiguous objects (if a mempool driver has its
578 * own calc_size() method returning min_chunk_size = mem_size),
579 * there is also an option to reserve the entire mempool memory
580 * as one contiguous block of memory.
582 * if we require contiguous objects, but not necessarily the entire
583 * mempool reserved space to be contiguous, pg_sz will be != 0,
584 * and the default ops->populate() will take care of not placing
585 * objects across pages.
587 * if our IO addresses are physical, we may get memory from bigger
588 * pages, or we might get memory from smaller pages, and how much of it
589 * we require depends on whether we want bigger or smaller pages.
590 * However, requesting each and every memory size is too much work, so
591 * what we'll do instead is walk through the page sizes available, pick
592 * the smallest one and set up page shift to match that one. We will be
593 * wasting some space this way, but it's much nicer than looping around
594 * trying to reserve each and every page size.
596 * If we fail to get enough contiguous memory, then we'll go and
597 * reserve space in smaller chunks.
600 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
601 ret = rte_mempool_get_page_size(mp, &pg_sz);
606 pg_shift = rte_bsf32(pg_sz);
608 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
609 size_t min_chunk_size;
611 mem_size = rte_mempool_ops_calc_mem_size(
612 mp, n, pg_shift, &min_chunk_size, &align);
619 ret = snprintf(mz_name, sizeof(mz_name),
620 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
621 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
626 /* if we're trying to reserve contiguous memory, add appropriate
629 if (min_chunk_size == (size_t)mem_size)
630 mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
632 /* Allocate a memzone, retrying with a smaller area on ENOMEM */
634 mz = rte_memzone_reserve_aligned(mz_name,
635 RTE_MIN((size_t)mem_size, max_alloc_size),
636 mp->socket_id, mz_flags, align);
638 if (mz == NULL && rte_errno != ENOMEM)
641 max_alloc_size = RTE_MIN(max_alloc_size,
642 (size_t)mem_size) / 2;
643 } while (mz == NULL && max_alloc_size >= min_chunk_size);
650 if (need_iova_contig_obj)
655 if (pg_sz == 0 || (mz_flags & RTE_MEMZONE_IOVA_CONTIG))
656 ret = rte_mempool_populate_iova(mp, mz->addr,
658 rte_mempool_memchunk_mz_free,
659 (void *)(uintptr_t)mz);
661 ret = rte_mempool_populate_virt(mp, mz->addr,
663 rte_mempool_memchunk_mz_free,
664 (void *)(uintptr_t)mz);
665 if (ret == 0) /* should not happen */
668 rte_memzone_free(mz);
673 rte_mempool_trace_populate_default(mp);
677 rte_mempool_free_memchunks(mp);
681 /* return the memory size required for mempool objects in anonymous mem */
683 get_anon_size(const struct rte_mempool *mp)
686 size_t pg_sz, pg_shift;
687 size_t min_chunk_size;
690 pg_sz = rte_mem_page_size();
691 pg_shift = rte_bsf32(pg_sz);
692 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
693 &min_chunk_size, &align);
698 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
700 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
706 * Calculate size since memhdr->len has contiguous chunk length
707 * which may be smaller if anon map is split into many contiguous
708 * chunks. Result must be the same as we calculated on populate.
710 size = get_anon_size(memhdr->mp);
714 rte_mem_unmap(opaque, size);
717 /* populate the mempool with an anonymous mapping */
719 rte_mempool_populate_anon(struct rte_mempool *mp)
725 /* mempool is already populated, error */
726 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
731 ret = mempool_ops_alloc_once(mp);
737 size = get_anon_size(mp);
743 /* get chunk of virtually continuous memory */
744 addr = rte_mem_map(NULL, size, RTE_PROT_READ | RTE_PROT_WRITE,
745 RTE_MAP_SHARED | RTE_MAP_ANONYMOUS, -1, 0);
748 /* can't use MMAP_LOCKED, it does not exist on BSD */
749 if (rte_mem_lock(addr, size) < 0) {
750 rte_mem_unmap(addr, size);
754 ret = rte_mempool_populate_virt(mp, addr, size, rte_mem_page_size(),
755 rte_mempool_memchunk_anon_free, addr);
756 if (ret == 0) /* should not happen */
763 rte_mempool_trace_populate_anon(mp);
764 return mp->populated_size;
767 rte_mempool_free_memchunks(mp);
773 rte_mempool_free(struct rte_mempool *mp)
775 struct rte_mempool_list *mempool_list = NULL;
776 struct rte_tailq_entry *te;
781 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
782 rte_mcfg_tailq_write_lock();
783 /* find out tailq entry */
784 TAILQ_FOREACH(te, mempool_list, next) {
785 if (te->data == (void *)mp)
790 TAILQ_REMOVE(mempool_list, te, next);
793 rte_mcfg_tailq_write_unlock();
795 rte_mempool_trace_free(mp);
796 rte_mempool_free_memchunks(mp);
797 rte_mempool_ops_free(mp);
798 rte_memzone_free(mp->mz);
802 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
805 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
810 * Create and initialize a cache for objects that are retrieved from and
811 * returned to an underlying mempool. This structure is identical to the
812 * local_cache[lcore_id] pointed to by the mempool structure.
814 struct rte_mempool_cache *
815 rte_mempool_cache_create(uint32_t size, int socket_id)
817 struct rte_mempool_cache *cache;
819 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
824 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
825 RTE_CACHE_LINE_SIZE, socket_id);
827 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
832 mempool_cache_init(cache, size);
834 rte_mempool_trace_cache_create(size, socket_id, cache);
839 * Free a cache. It's the responsibility of the user to make sure that any
840 * remaining objects in the cache are flushed to the corresponding
844 rte_mempool_cache_free(struct rte_mempool_cache *cache)
846 rte_mempool_trace_cache_free(cache);
850 /* create an empty mempool */
852 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
853 unsigned cache_size, unsigned private_data_size,
854 int socket_id, unsigned flags)
856 char mz_name[RTE_MEMZONE_NAMESIZE];
857 struct rte_mempool_list *mempool_list;
858 struct rte_mempool *mp = NULL;
859 struct rte_tailq_entry *te = NULL;
860 const struct rte_memzone *mz = NULL;
862 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
863 struct rte_mempool_objsz objsz;
867 /* compilation-time checks */
868 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
869 RTE_CACHE_LINE_MASK) != 0);
870 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
871 RTE_CACHE_LINE_MASK) != 0);
872 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
873 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
874 RTE_CACHE_LINE_MASK) != 0);
875 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
876 RTE_CACHE_LINE_MASK) != 0);
879 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
881 /* asked for zero items */
887 /* asked cache too big */
888 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
889 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
894 /* "no cache align" imply "no spread" */
895 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
896 flags |= MEMPOOL_F_NO_SPREAD;
898 /* calculate mempool object sizes. */
899 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
904 rte_mcfg_mempool_write_lock();
907 * reserve a memory zone for this mempool: private data is
910 private_data_size = (private_data_size +
911 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
914 /* try to allocate tailq entry */
915 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
917 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
921 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
922 mempool_size += private_data_size;
923 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
925 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
926 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
927 rte_errno = ENAMETOOLONG;
931 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
935 /* init the mempool structure */
937 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
938 ret = strlcpy(mp->name, name, sizeof(mp->name));
939 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
940 rte_errno = ENAMETOOLONG;
946 mp->socket_id = socket_id;
947 mp->elt_size = objsz.elt_size;
948 mp->header_size = objsz.header_size;
949 mp->trailer_size = objsz.trailer_size;
950 /* Size of default caches, zero means disabled. */
951 mp->cache_size = cache_size;
952 mp->private_data_size = private_data_size;
953 STAILQ_INIT(&mp->elt_list);
954 STAILQ_INIT(&mp->mem_list);
957 * local_cache pointer is set even if cache_size is zero.
958 * The local_cache points to just past the elt_pa[] array.
960 mp->local_cache = (struct rte_mempool_cache *)
961 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
963 /* Init all default caches. */
964 if (cache_size != 0) {
965 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
966 mempool_cache_init(&mp->local_cache[lcore_id],
972 rte_mcfg_tailq_write_lock();
973 TAILQ_INSERT_TAIL(mempool_list, te, next);
974 rte_mcfg_tailq_write_unlock();
975 rte_mcfg_mempool_write_unlock();
977 rte_mempool_trace_create_empty(name, n, elt_size, cache_size,
978 private_data_size, flags, mp);
982 rte_mcfg_mempool_write_unlock();
984 rte_mempool_free(mp);
988 /* create the mempool */
990 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
991 unsigned cache_size, unsigned private_data_size,
992 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
993 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
994 int socket_id, unsigned flags)
997 struct rte_mempool *mp;
999 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1000 private_data_size, socket_id, flags);
1005 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
1006 * set the correct index into the table of ops structs.
1008 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
1009 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
1010 else if (flags & MEMPOOL_F_SP_PUT)
1011 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
1012 else if (flags & MEMPOOL_F_SC_GET)
1013 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
1015 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
1020 /* call the mempool priv initializer */
1022 mp_init(mp, mp_init_arg);
1024 if (rte_mempool_populate_default(mp) < 0)
1027 /* call the object initializers */
1029 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1031 rte_mempool_trace_create(name, n, elt_size, cache_size,
1032 private_data_size, mp_init, mp_init_arg, obj_init,
1033 obj_init_arg, flags, mp);
1037 rte_mempool_free(mp);
1041 /* Return the number of entries in the mempool */
1043 rte_mempool_avail_count(const struct rte_mempool *mp)
1048 count = rte_mempool_ops_get_count(mp);
1050 if (mp->cache_size == 0)
1053 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1054 count += mp->local_cache[lcore_id].len;
1057 * due to race condition (access to len is not locked), the
1058 * total can be greater than size... so fix the result
1060 if (count > mp->size)
1065 /* return the number of entries allocated from the mempool */
1067 rte_mempool_in_use_count(const struct rte_mempool *mp)
1069 return mp->size - rte_mempool_avail_count(mp);
1072 /* dump the cache status */
1074 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1078 unsigned cache_count;
1080 fprintf(f, " internal cache infos:\n");
1081 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1083 if (mp->cache_size == 0)
1086 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1087 cache_count = mp->local_cache[lcore_id].len;
1088 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1089 lcore_id, cache_count);
1090 count += cache_count;
1092 fprintf(f, " total_cache_count=%u\n", count);
1096 #ifndef __INTEL_COMPILER
1097 #pragma GCC diagnostic ignored "-Wcast-qual"
1100 /* check and update cookies or panic (internal) */
1101 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1102 void * const *obj_table_const, unsigned n, int free)
1104 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1105 struct rte_mempool_objhdr *hdr;
1106 struct rte_mempool_objtlr *tlr;
1112 /* Force to drop the "const" attribute. This is done only when
1113 * DEBUG is enabled */
1114 tmp = (void *) obj_table_const;
1120 if (rte_mempool_from_obj(obj) != mp)
1121 rte_panic("MEMPOOL: object is owned by another "
1124 hdr = __mempool_get_header(obj);
1125 cookie = hdr->cookie;
1128 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1129 RTE_LOG(CRIT, MEMPOOL,
1130 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1131 obj, (const void *) mp, cookie);
1132 rte_panic("MEMPOOL: bad header cookie (put)\n");
1134 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1135 } else if (free == 1) {
1136 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1137 RTE_LOG(CRIT, MEMPOOL,
1138 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1139 obj, (const void *) mp, cookie);
1140 rte_panic("MEMPOOL: bad header cookie (get)\n");
1142 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1143 } else if (free == 2) {
1144 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1145 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1146 RTE_LOG(CRIT, MEMPOOL,
1147 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1148 obj, (const void *) mp, cookie);
1149 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1152 tlr = __mempool_get_trailer(obj);
1153 cookie = tlr->cookie;
1154 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1155 RTE_LOG(CRIT, MEMPOOL,
1156 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1157 obj, (const void *) mp, cookie);
1158 rte_panic("MEMPOOL: bad trailer cookie\n");
1163 RTE_SET_USED(obj_table_const);
1170 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1171 void * const *first_obj_table_const, unsigned int n, int free)
1173 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1174 struct rte_mempool_info info;
1175 const size_t total_elt_sz =
1176 mp->header_size + mp->elt_size + mp->trailer_size;
1179 rte_mempool_ops_get_info(mp, &info);
1181 for (i = 0; i < n; ++i) {
1182 void *first_obj = first_obj_table_const[i];
1184 for (j = 0; j < info.contig_block_size; ++j) {
1187 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1188 rte_mempool_check_cookies(mp, &obj, 1, free);
1193 RTE_SET_USED(first_obj_table_const);
1199 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1201 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1202 void *obj, __rte_unused unsigned idx)
1204 __mempool_check_cookies(mp, &obj, 1, 2);
1208 mempool_audit_cookies(struct rte_mempool *mp)
1212 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1213 if (num != mp->size) {
1214 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1215 "iterated only over %u elements\n",
1220 #define mempool_audit_cookies(mp) do {} while(0)
1223 #ifndef __INTEL_COMPILER
1224 #pragma GCC diagnostic error "-Wcast-qual"
1227 /* check cookies before and after objects */
1229 mempool_audit_cache(const struct rte_mempool *mp)
1231 /* check cache size consistency */
1234 if (mp->cache_size == 0)
1237 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1238 const struct rte_mempool_cache *cache;
1239 cache = &mp->local_cache[lcore_id];
1240 if (cache->len > cache->flushthresh) {
1241 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1243 rte_panic("MEMPOOL: invalid cache len\n");
1248 /* check the consistency of mempool (size, cookies, ...) */
1250 rte_mempool_audit(struct rte_mempool *mp)
1252 mempool_audit_cache(mp);
1253 mempool_audit_cookies(mp);
1255 /* For case where mempool DEBUG is not set, and cache size is 0 */
1259 /* dump the status of the mempool on the console */
1261 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1263 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1264 struct rte_mempool_info info;
1265 struct rte_mempool_debug_stats sum;
1268 struct rte_mempool_memhdr *memhdr;
1269 unsigned common_count;
1270 unsigned cache_count;
1273 RTE_ASSERT(f != NULL);
1274 RTE_ASSERT(mp != NULL);
1276 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1277 fprintf(f, " flags=%x\n", mp->flags);
1278 fprintf(f, " pool=%p\n", mp->pool_data);
1279 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1280 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1281 fprintf(f, " size=%"PRIu32"\n", mp->size);
1282 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1283 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1284 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1285 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1286 fprintf(f, " total_obj_size=%"PRIu32"\n",
1287 mp->header_size + mp->elt_size + mp->trailer_size);
1289 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1291 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1292 mem_len += memhdr->len;
1294 fprintf(f, " avg bytes/object=%#Lf\n",
1295 (long double)mem_len / mp->size);
1298 cache_count = rte_mempool_dump_cache(f, mp);
1299 common_count = rte_mempool_ops_get_count(mp);
1300 if ((cache_count + common_count) > mp->size)
1301 common_count = mp->size - cache_count;
1302 fprintf(f, " common_pool_count=%u\n", common_count);
1304 /* sum and dump statistics */
1305 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1306 rte_mempool_ops_get_info(mp, &info);
1307 memset(&sum, 0, sizeof(sum));
1308 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1309 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1310 sum.put_objs += mp->stats[lcore_id].put_objs;
1311 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1312 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1313 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1314 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1315 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1316 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1318 fprintf(f, " stats:\n");
1319 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1320 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1321 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1322 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1323 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1324 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1325 if (info.contig_block_size > 0) {
1326 fprintf(f, " get_success_blks=%"PRIu64"\n",
1327 sum.get_success_blks);
1328 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1331 fprintf(f, " no statistics available\n");
1334 rte_mempool_audit(mp);
1337 /* dump the status of all mempools on the console */
1339 rte_mempool_list_dump(FILE *f)
1341 struct rte_mempool *mp = NULL;
1342 struct rte_tailq_entry *te;
1343 struct rte_mempool_list *mempool_list;
1345 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1347 rte_mcfg_mempool_read_lock();
1349 TAILQ_FOREACH(te, mempool_list, next) {
1350 mp = (struct rte_mempool *) te->data;
1351 rte_mempool_dump(f, mp);
1354 rte_mcfg_mempool_read_unlock();
1357 /* search a mempool from its name */
1358 struct rte_mempool *
1359 rte_mempool_lookup(const char *name)
1361 struct rte_mempool *mp = NULL;
1362 struct rte_tailq_entry *te;
1363 struct rte_mempool_list *mempool_list;
1365 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1367 rte_mcfg_mempool_read_lock();
1369 TAILQ_FOREACH(te, mempool_list, next) {
1370 mp = (struct rte_mempool *) te->data;
1371 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1375 rte_mcfg_mempool_read_unlock();
1385 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1388 struct rte_tailq_entry *te = NULL;
1389 struct rte_mempool_list *mempool_list;
1392 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1394 rte_mcfg_mempool_read_lock();
1396 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1397 (*func)((struct rte_mempool *) te->data, arg);
1400 rte_mcfg_mempool_read_unlock();