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>
33 #include <rte_tailq.h>
34 #include <rte_function_versioning.h>
36 #include "rte_mempool.h"
37 #include "rte_mempool_trace.h"
39 TAILQ_HEAD(rte_mempool_list, rte_tailq_entry);
41 static struct rte_tailq_elem rte_mempool_tailq = {
42 .name = "RTE_MEMPOOL",
44 EAL_REGISTER_TAILQ(rte_mempool_tailq)
46 #define CACHE_FLUSHTHRESH_MULTIPLIER 1.5
47 #define CALC_CACHE_FLUSHTHRESH(c) \
48 ((typeof(c))((c) * CACHE_FLUSHTHRESH_MULTIPLIER))
50 #if defined(RTE_ARCH_X86)
52 * return the greatest common divisor between a and b (fast algorithm)
55 static unsigned get_gcd(unsigned a, unsigned b)
80 * Depending on memory configuration on x86 arch, objects addresses are spread
81 * between channels and ranks in RAM: the pool allocator will add
82 * padding between objects. This function return the new size of the
86 arch_mem_object_align(unsigned int obj_size)
88 unsigned nrank, nchan;
89 unsigned new_obj_size;
91 /* get number of channels */
92 nchan = rte_memory_get_nchannel();
96 nrank = rte_memory_get_nrank();
100 /* process new object size */
101 new_obj_size = (obj_size + RTE_MEMPOOL_ALIGN_MASK) / RTE_MEMPOOL_ALIGN;
102 while (get_gcd(new_obj_size, nrank * nchan) != 1)
104 return new_obj_size * RTE_MEMPOOL_ALIGN;
108 arch_mem_object_align(unsigned int obj_size)
114 struct pagesz_walk_arg {
120 find_min_pagesz(const struct rte_memseg_list *msl, void *arg)
122 struct pagesz_walk_arg *wa = arg;
126 * we need to only look at page sizes available for a particular socket
127 * ID. so, we either need an exact match on socket ID (can match both
128 * native and external memory), or, if SOCKET_ID_ANY was specified as a
129 * socket ID argument, we must only look at native memory and ignore any
130 * page sizes associated with external memory.
132 valid = msl->socket_id == wa->socket_id;
133 valid |= wa->socket_id == SOCKET_ID_ANY && msl->external == 0;
135 if (valid && msl->page_sz < wa->min)
136 wa->min = msl->page_sz;
142 get_min_page_size(int socket_id)
144 struct pagesz_walk_arg wa;
147 wa.socket_id = socket_id;
149 rte_memseg_list_walk(find_min_pagesz, &wa);
151 return wa.min == SIZE_MAX ? (size_t) getpagesize() : wa.min;
156 mempool_add_elem(struct rte_mempool *mp, __rte_unused void *opaque,
157 void *obj, rte_iova_t iova)
159 struct rte_mempool_objhdr *hdr;
160 struct rte_mempool_objtlr *tlr __rte_unused;
162 /* set mempool ptr in header */
163 hdr = RTE_PTR_SUB(obj, sizeof(*hdr));
166 STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
167 mp->populated_size++;
169 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
170 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
171 tlr = __mempool_get_trailer(obj);
172 tlr->cookie = RTE_MEMPOOL_TRAILER_COOKIE;
176 /* call obj_cb() for each mempool element */
178 rte_mempool_obj_iter(struct rte_mempool *mp,
179 rte_mempool_obj_cb_t *obj_cb, void *obj_cb_arg)
181 struct rte_mempool_objhdr *hdr;
185 STAILQ_FOREACH(hdr, &mp->elt_list, next) {
186 obj = (char *)hdr + sizeof(*hdr);
187 obj_cb(mp, obj_cb_arg, obj, n);
194 /* call mem_cb() for each mempool memory chunk */
196 rte_mempool_mem_iter(struct rte_mempool *mp,
197 rte_mempool_mem_cb_t *mem_cb, void *mem_cb_arg)
199 struct rte_mempool_memhdr *hdr;
202 STAILQ_FOREACH(hdr, &mp->mem_list, next) {
203 mem_cb(mp, mem_cb_arg, hdr, n);
210 /* get the header, trailer and total size of a mempool element. */
212 rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
213 struct rte_mempool_objsz *sz)
215 struct rte_mempool_objsz lsz;
217 sz = (sz != NULL) ? sz : &lsz;
219 sz->header_size = sizeof(struct rte_mempool_objhdr);
220 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0)
221 sz->header_size = RTE_ALIGN_CEIL(sz->header_size,
224 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
225 sz->trailer_size = sizeof(struct rte_mempool_objtlr);
227 sz->trailer_size = 0;
230 /* element size is 8 bytes-aligned at least */
231 sz->elt_size = RTE_ALIGN_CEIL(elt_size, sizeof(uint64_t));
233 /* expand trailer to next cache line */
234 if ((flags & MEMPOOL_F_NO_CACHE_ALIGN) == 0) {
235 sz->total_size = sz->header_size + sz->elt_size +
237 sz->trailer_size += ((RTE_MEMPOOL_ALIGN -
238 (sz->total_size & RTE_MEMPOOL_ALIGN_MASK)) &
239 RTE_MEMPOOL_ALIGN_MASK);
243 * increase trailer to add padding between objects in order to
244 * spread them across memory channels/ranks
246 if ((flags & MEMPOOL_F_NO_SPREAD) == 0) {
248 new_size = arch_mem_object_align
249 (sz->header_size + sz->elt_size + sz->trailer_size);
250 sz->trailer_size = new_size - sz->header_size - sz->elt_size;
253 /* this is the size of an object, including header and trailer */
254 sz->total_size = sz->header_size + sz->elt_size + sz->trailer_size;
256 return sz->total_size;
259 /* free a memchunk allocated with rte_memzone_reserve() */
261 rte_mempool_memchunk_mz_free(__rte_unused struct rte_mempool_memhdr *memhdr,
264 const struct rte_memzone *mz = opaque;
265 rte_memzone_free(mz);
268 /* Free memory chunks used by a mempool. Objects must be in pool */
270 rte_mempool_free_memchunks(struct rte_mempool *mp)
272 struct rte_mempool_memhdr *memhdr;
275 while (!STAILQ_EMPTY(&mp->elt_list)) {
276 rte_mempool_ops_dequeue_bulk(mp, &elt, 1);
278 STAILQ_REMOVE_HEAD(&mp->elt_list, next);
279 mp->populated_size--;
282 while (!STAILQ_EMPTY(&mp->mem_list)) {
283 memhdr = STAILQ_FIRST(&mp->mem_list);
284 STAILQ_REMOVE_HEAD(&mp->mem_list, next);
285 if (memhdr->free_cb != NULL)
286 memhdr->free_cb(memhdr, memhdr->opaque);
293 mempool_ops_alloc_once(struct rte_mempool *mp)
297 /* create the internal ring if not already done */
298 if ((mp->flags & MEMPOOL_F_POOL_CREATED) == 0) {
299 ret = rte_mempool_ops_alloc(mp);
302 mp->flags |= MEMPOOL_F_POOL_CREATED;
308 rte_mempool_populate_iova_v21(struct rte_mempool *mp, char *vaddr,
309 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
312 /* Add objects in the pool, using a physically contiguous memory
313 * zone. Return the number of objects added, or a negative value
317 rte_mempool_populate_iova_v21(struct rte_mempool *mp, char *vaddr,
318 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
323 struct rte_mempool_memhdr *memhdr;
326 ret = mempool_ops_alloc_once(mp);
330 /* mempool is already populated */
331 if (mp->populated_size >= mp->size)
334 memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
339 memhdr->addr = vaddr;
342 memhdr->free_cb = free_cb;
343 memhdr->opaque = opaque;
345 if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
346 off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
348 off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_MEMPOOL_ALIGN) - vaddr;
355 i = rte_mempool_ops_populate(mp, mp->size - mp->populated_size,
357 (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off),
358 len - off, mempool_add_elem, NULL);
360 /* not enough room to store one object */
366 STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
369 rte_mempool_trace_populate_iova(mp, vaddr, iova, len, free_cb, opaque);
377 BIND_DEFAULT_SYMBOL(rte_mempool_populate_iova, _v21, 21);
379 int rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
380 rte_iova_t iova, size_t len,
381 rte_mempool_memchunk_free_cb_t *free_cb,
383 rte_mempool_populate_iova_v21);
386 rte_mempool_populate_iova_v20(struct rte_mempool *mp, char *vaddr,
387 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
391 rte_mempool_populate_iova_v20(struct rte_mempool *mp, char *vaddr,
392 rte_iova_t iova, size_t len, rte_mempool_memchunk_free_cb_t *free_cb,
397 ret = rte_mempool_populate_iova_v21(mp, vaddr, iova, len, free_cb,
404 VERSION_SYMBOL(rte_mempool_populate_iova, _v20, 20.0);
409 struct rte_memseg *ms;
411 /* try registered memory first */
412 ms = rte_mem_virt2memseg(addr, NULL);
413 if (ms == NULL || ms->iova == RTE_BAD_IOVA)
414 /* fall back to actual physical address */
415 return rte_mem_virt2iova(addr);
416 return ms->iova + RTE_PTR_DIFF(addr, ms->addr);
420 rte_mempool_populate_virt_v21(struct rte_mempool *mp, char *addr,
421 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
424 /* Populate the mempool with a virtual area. Return the number of
425 * objects added, or a negative value on error.
428 rte_mempool_populate_virt_v21(struct rte_mempool *mp, char *addr,
429 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
433 size_t off, phys_len;
436 if (mp->flags & MEMPOOL_F_NO_IOVA_CONTIG)
437 return rte_mempool_populate_iova(mp, addr, RTE_BAD_IOVA,
438 len, free_cb, opaque);
440 for (off = 0; off < len &&
441 mp->populated_size < mp->size; off += phys_len) {
443 iova = get_iova(addr + off);
445 /* populate with the largest group of contiguous pages */
446 for (phys_len = RTE_MIN(
447 (size_t)(RTE_PTR_ALIGN_CEIL(addr + off + 1, pg_sz) -
450 off + phys_len < len;
451 phys_len = RTE_MIN(phys_len + pg_sz, len - off)) {
454 iova_tmp = get_iova(addr + off + phys_len);
456 if (iova_tmp == RTE_BAD_IOVA ||
457 iova_tmp != iova + phys_len)
461 ret = rte_mempool_populate_iova_v21(mp, addr + off, iova,
462 phys_len, free_cb, opaque);
467 /* no need to call the free callback for next chunks */
472 rte_mempool_trace_populate_virt(mp, addr, len, pg_sz, free_cb, opaque);
476 rte_mempool_free_memchunks(mp);
479 BIND_DEFAULT_SYMBOL(rte_mempool_populate_virt, _v21, 21);
481 int rte_mempool_populate_virt(struct rte_mempool *mp,
482 char *addr, size_t len, size_t pg_sz,
483 rte_mempool_memchunk_free_cb_t *free_cb,
485 rte_mempool_populate_virt_v21);
488 rte_mempool_populate_virt_v20(struct rte_mempool *mp, char *addr,
489 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
493 rte_mempool_populate_virt_v20(struct rte_mempool *mp, char *addr,
494 size_t len, size_t pg_sz, rte_mempool_memchunk_free_cb_t *free_cb,
499 ret = rte_mempool_populate_virt_v21(mp, addr, len, pg_sz,
507 VERSION_SYMBOL(rte_mempool_populate_virt, _v20, 20.0);
509 /* Get the minimal page size used in a mempool before populating it. */
511 rte_mempool_get_page_size(struct rte_mempool *mp, size_t *pg_sz)
513 bool need_iova_contig_obj;
514 bool alloc_in_ext_mem;
517 /* check if we can retrieve a valid socket ID */
518 ret = rte_malloc_heap_socket_is_external(mp->socket_id);
521 alloc_in_ext_mem = (ret == 1);
522 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
524 if (!need_iova_contig_obj)
526 else if (rte_eal_has_hugepages() || alloc_in_ext_mem)
527 *pg_sz = get_min_page_size(mp->socket_id);
529 *pg_sz = getpagesize();
531 rte_mempool_trace_get_page_size(mp, *pg_sz);
535 /* Default function to populate the mempool: allocate memory in memzones,
536 * and populate them. Return the number of objects added, or a negative
540 rte_mempool_populate_default(struct rte_mempool *mp)
542 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
543 char mz_name[RTE_MEMZONE_NAMESIZE];
544 const struct rte_memzone *mz;
546 size_t align, pg_sz, pg_shift = 0;
550 bool need_iova_contig_obj;
551 size_t max_alloc_size = SIZE_MAX;
553 ret = mempool_ops_alloc_once(mp);
557 /* mempool must not be populated */
558 if (mp->nb_mem_chunks != 0)
562 * the following section calculates page shift and page size values.
564 * these values impact the result of calc_mem_size operation, which
565 * returns the amount of memory that should be allocated to store the
566 * desired number of objects. when not zero, it allocates more memory
567 * for the padding between objects, to ensure that an object does not
568 * cross a page boundary. in other words, page size/shift are to be set
569 * to zero if mempool elements won't care about page boundaries.
570 * there are several considerations for page size and page shift here.
572 * if we don't need our mempools to have physically contiguous objects,
573 * then just set page shift and page size to 0, because the user has
574 * indicated that there's no need to care about anything.
576 * if we do need contiguous objects (if a mempool driver has its
577 * own calc_size() method returning min_chunk_size = mem_size),
578 * there is also an option to reserve the entire mempool memory
579 * as one contiguous block of memory.
581 * if we require contiguous objects, but not necessarily the entire
582 * mempool reserved space to be contiguous, pg_sz will be != 0,
583 * and the default ops->populate() will take care of not placing
584 * objects across pages.
586 * if our IO addresses are physical, we may get memory from bigger
587 * pages, or we might get memory from smaller pages, and how much of it
588 * we require depends on whether we want bigger or smaller pages.
589 * However, requesting each and every memory size is too much work, so
590 * what we'll do instead is walk through the page sizes available, pick
591 * the smallest one and set up page shift to match that one. We will be
592 * wasting some space this way, but it's much nicer than looping around
593 * trying to reserve each and every page size.
595 * If we fail to get enough contiguous memory, then we'll go and
596 * reserve space in smaller chunks.
599 need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
600 ret = rte_mempool_get_page_size(mp, &pg_sz);
605 pg_shift = rte_bsf32(pg_sz);
607 for (mz_id = 0, n = mp->size; n > 0; mz_id++, n -= ret) {
608 size_t min_chunk_size;
610 mem_size = rte_mempool_ops_calc_mem_size(
611 mp, n, pg_shift, &min_chunk_size, &align);
618 ret = snprintf(mz_name, sizeof(mz_name),
619 RTE_MEMPOOL_MZ_FORMAT "_%d", mp->name, mz_id);
620 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
625 /* if we're trying to reserve contiguous memory, add appropriate
628 if (min_chunk_size == (size_t)mem_size)
629 mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
631 /* Allocate a memzone, retrying with a smaller area on ENOMEM */
633 mz = rte_memzone_reserve_aligned(mz_name,
634 RTE_MIN((size_t)mem_size, max_alloc_size),
635 mp->socket_id, mz_flags, align);
637 if (mz == NULL && rte_errno != ENOMEM)
640 max_alloc_size = RTE_MIN(max_alloc_size,
641 (size_t)mem_size) / 2;
642 } while (mz == NULL && max_alloc_size >= min_chunk_size);
649 if (need_iova_contig_obj)
654 if (pg_sz == 0 || (mz_flags & RTE_MEMZONE_IOVA_CONTIG))
655 ret = rte_mempool_populate_iova(mp, mz->addr,
657 rte_mempool_memchunk_mz_free,
658 (void *)(uintptr_t)mz);
660 ret = rte_mempool_populate_virt(mp, mz->addr,
662 rte_mempool_memchunk_mz_free,
663 (void *)(uintptr_t)mz);
664 if (ret == 0) /* should not happen */
667 rte_memzone_free(mz);
672 rte_mempool_trace_populate_default(mp);
676 rte_mempool_free_memchunks(mp);
680 /* return the memory size required for mempool objects in anonymous mem */
682 get_anon_size(const struct rte_mempool *mp)
685 size_t pg_sz, pg_shift;
686 size_t min_chunk_size;
689 pg_sz = getpagesize();
690 pg_shift = rte_bsf32(pg_sz);
691 size = rte_mempool_ops_calc_mem_size(mp, mp->size, pg_shift,
692 &min_chunk_size, &align);
697 /* unmap a memory zone mapped by rte_mempool_populate_anon() */
699 rte_mempool_memchunk_anon_free(struct rte_mempool_memhdr *memhdr,
705 * Calculate size since memhdr->len has contiguous chunk length
706 * which may be smaller if anon map is split into many contiguous
707 * chunks. Result must be the same as we calculated on populate.
709 size = get_anon_size(memhdr->mp);
713 munmap(opaque, size);
716 /* populate the mempool with an anonymous mapping */
718 rte_mempool_populate_anon(struct rte_mempool *mp)
724 /* mempool is already populated, error */
725 if ((!STAILQ_EMPTY(&mp->mem_list)) || mp->nb_mem_chunks != 0) {
730 ret = mempool_ops_alloc_once(mp);
736 size = get_anon_size(mp);
742 /* get chunk of virtually continuous memory */
743 addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
744 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
745 if (addr == MAP_FAILED) {
749 /* can't use MMAP_LOCKED, it does not exist on BSD */
750 if (mlock(addr, size) < 0) {
756 ret = rte_mempool_populate_virt(mp, addr, size, getpagesize(),
757 rte_mempool_memchunk_anon_free, addr);
758 if (ret == 0) /* should not happen */
765 rte_mempool_trace_populate_anon(mp);
766 return mp->populated_size;
769 rte_mempool_free_memchunks(mp);
775 rte_mempool_free(struct rte_mempool *mp)
777 struct rte_mempool_list *mempool_list = NULL;
778 struct rte_tailq_entry *te;
783 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
784 rte_mcfg_tailq_write_lock();
785 /* find out tailq entry */
786 TAILQ_FOREACH(te, mempool_list, next) {
787 if (te->data == (void *)mp)
792 TAILQ_REMOVE(mempool_list, te, next);
795 rte_mcfg_tailq_write_unlock();
797 rte_mempool_trace_free(mp);
798 rte_mempool_free_memchunks(mp);
799 rte_mempool_ops_free(mp);
800 rte_memzone_free(mp->mz);
804 mempool_cache_init(struct rte_mempool_cache *cache, uint32_t size)
807 cache->flushthresh = CALC_CACHE_FLUSHTHRESH(size);
812 * Create and initialize a cache for objects that are retrieved from and
813 * returned to an underlying mempool. This structure is identical to the
814 * local_cache[lcore_id] pointed to by the mempool structure.
816 struct rte_mempool_cache *
817 rte_mempool_cache_create(uint32_t size, int socket_id)
819 struct rte_mempool_cache *cache;
821 if (size == 0 || size > RTE_MEMPOOL_CACHE_MAX_SIZE) {
826 cache = rte_zmalloc_socket("MEMPOOL_CACHE", sizeof(*cache),
827 RTE_CACHE_LINE_SIZE, socket_id);
829 RTE_LOG(ERR, MEMPOOL, "Cannot allocate mempool cache.\n");
834 mempool_cache_init(cache, size);
836 rte_mempool_trace_cache_create(size, socket_id, cache);
841 * Free a cache. It's the responsibility of the user to make sure that any
842 * remaining objects in the cache are flushed to the corresponding
846 rte_mempool_cache_free(struct rte_mempool_cache *cache)
848 rte_mempool_trace_cache_free(cache);
852 /* create an empty mempool */
854 rte_mempool_create_empty(const char *name, unsigned n, unsigned elt_size,
855 unsigned cache_size, unsigned private_data_size,
856 int socket_id, unsigned flags)
858 char mz_name[RTE_MEMZONE_NAMESIZE];
859 struct rte_mempool_list *mempool_list;
860 struct rte_mempool *mp = NULL;
861 struct rte_tailq_entry *te = NULL;
862 const struct rte_memzone *mz = NULL;
864 unsigned int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
865 struct rte_mempool_objsz objsz;
869 /* compilation-time checks */
870 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool) &
871 RTE_CACHE_LINE_MASK) != 0);
872 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_cache) &
873 RTE_CACHE_LINE_MASK) != 0);
874 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
875 RTE_BUILD_BUG_ON((sizeof(struct rte_mempool_debug_stats) &
876 RTE_CACHE_LINE_MASK) != 0);
877 RTE_BUILD_BUG_ON((offsetof(struct rte_mempool, stats) &
878 RTE_CACHE_LINE_MASK) != 0);
881 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
883 /* asked for zero items */
889 /* asked cache too big */
890 if (cache_size > RTE_MEMPOOL_CACHE_MAX_SIZE ||
891 CALC_CACHE_FLUSHTHRESH(cache_size) > n) {
896 /* "no cache align" imply "no spread" */
897 if (flags & MEMPOOL_F_NO_CACHE_ALIGN)
898 flags |= MEMPOOL_F_NO_SPREAD;
900 /* calculate mempool object sizes. */
901 if (!rte_mempool_calc_obj_size(elt_size, flags, &objsz)) {
906 rte_mcfg_mempool_write_lock();
909 * reserve a memory zone for this mempool: private data is
912 private_data_size = (private_data_size +
913 RTE_MEMPOOL_ALIGN_MASK) & (~RTE_MEMPOOL_ALIGN_MASK);
916 /* try to allocate tailq entry */
917 te = rte_zmalloc("MEMPOOL_TAILQ_ENTRY", sizeof(*te), 0);
919 RTE_LOG(ERR, MEMPOOL, "Cannot allocate tailq entry!\n");
923 mempool_size = MEMPOOL_HEADER_SIZE(mp, cache_size);
924 mempool_size += private_data_size;
925 mempool_size = RTE_ALIGN_CEIL(mempool_size, RTE_MEMPOOL_ALIGN);
927 ret = snprintf(mz_name, sizeof(mz_name), RTE_MEMPOOL_MZ_FORMAT, name);
928 if (ret < 0 || ret >= (int)sizeof(mz_name)) {
929 rte_errno = ENAMETOOLONG;
933 mz = rte_memzone_reserve(mz_name, mempool_size, socket_id, mz_flags);
937 /* init the mempool structure */
939 memset(mp, 0, MEMPOOL_HEADER_SIZE(mp, cache_size));
940 ret = strlcpy(mp->name, name, sizeof(mp->name));
941 if (ret < 0 || ret >= (int)sizeof(mp->name)) {
942 rte_errno = ENAMETOOLONG;
948 mp->socket_id = socket_id;
949 mp->elt_size = objsz.elt_size;
950 mp->header_size = objsz.header_size;
951 mp->trailer_size = objsz.trailer_size;
952 /* Size of default caches, zero means disabled. */
953 mp->cache_size = cache_size;
954 mp->private_data_size = private_data_size;
955 STAILQ_INIT(&mp->elt_list);
956 STAILQ_INIT(&mp->mem_list);
959 * local_cache pointer is set even if cache_size is zero.
960 * The local_cache points to just past the elt_pa[] array.
962 mp->local_cache = (struct rte_mempool_cache *)
963 RTE_PTR_ADD(mp, MEMPOOL_HEADER_SIZE(mp, 0));
965 /* Init all default caches. */
966 if (cache_size != 0) {
967 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
968 mempool_cache_init(&mp->local_cache[lcore_id],
974 rte_mcfg_tailq_write_lock();
975 TAILQ_INSERT_TAIL(mempool_list, te, next);
976 rte_mcfg_tailq_write_unlock();
977 rte_mcfg_mempool_write_unlock();
979 rte_mempool_trace_create_empty(name, n, elt_size, cache_size,
980 private_data_size, flags, mp);
984 rte_mcfg_mempool_write_unlock();
986 rte_mempool_free(mp);
990 /* create the mempool */
992 rte_mempool_create(const char *name, unsigned n, unsigned elt_size,
993 unsigned cache_size, unsigned private_data_size,
994 rte_mempool_ctor_t *mp_init, void *mp_init_arg,
995 rte_mempool_obj_cb_t *obj_init, void *obj_init_arg,
996 int socket_id, unsigned flags)
999 struct rte_mempool *mp;
1001 mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
1002 private_data_size, socket_id, flags);
1007 * Since we have 4 combinations of the SP/SC/MP/MC examine the flags to
1008 * set the correct index into the table of ops structs.
1010 if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
1011 ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
1012 else if (flags & MEMPOOL_F_SP_PUT)
1013 ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
1014 else if (flags & MEMPOOL_F_SC_GET)
1015 ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
1017 ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
1022 /* call the mempool priv initializer */
1024 mp_init(mp, mp_init_arg);
1026 if (rte_mempool_populate_default(mp) < 0)
1029 /* call the object initializers */
1031 rte_mempool_obj_iter(mp, obj_init, obj_init_arg);
1033 rte_mempool_trace_create(name, n, elt_size, cache_size,
1034 private_data_size, mp_init, mp_init_arg, obj_init,
1035 obj_init_arg, flags, mp);
1039 rte_mempool_free(mp);
1043 /* Return the number of entries in the mempool */
1045 rte_mempool_avail_count(const struct rte_mempool *mp)
1050 count = rte_mempool_ops_get_count(mp);
1052 if (mp->cache_size == 0)
1055 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
1056 count += mp->local_cache[lcore_id].len;
1059 * due to race condition (access to len is not locked), the
1060 * total can be greater than size... so fix the result
1062 if (count > mp->size)
1067 /* return the number of entries allocated from the mempool */
1069 rte_mempool_in_use_count(const struct rte_mempool *mp)
1071 return mp->size - rte_mempool_avail_count(mp);
1074 /* dump the cache status */
1076 rte_mempool_dump_cache(FILE *f, const struct rte_mempool *mp)
1080 unsigned cache_count;
1082 fprintf(f, " internal cache infos:\n");
1083 fprintf(f, " cache_size=%"PRIu32"\n", mp->cache_size);
1085 if (mp->cache_size == 0)
1088 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1089 cache_count = mp->local_cache[lcore_id].len;
1090 fprintf(f, " cache_count[%u]=%"PRIu32"\n",
1091 lcore_id, cache_count);
1092 count += cache_count;
1094 fprintf(f, " total_cache_count=%u\n", count);
1098 #ifndef __INTEL_COMPILER
1099 #pragma GCC diagnostic ignored "-Wcast-qual"
1102 /* check and update cookies or panic (internal) */
1103 void rte_mempool_check_cookies(const struct rte_mempool *mp,
1104 void * const *obj_table_const, unsigned n, int free)
1106 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1107 struct rte_mempool_objhdr *hdr;
1108 struct rte_mempool_objtlr *tlr;
1114 /* Force to drop the "const" attribute. This is done only when
1115 * DEBUG is enabled */
1116 tmp = (void *) obj_table_const;
1122 if (rte_mempool_from_obj(obj) != mp)
1123 rte_panic("MEMPOOL: object is owned by another "
1126 hdr = __mempool_get_header(obj);
1127 cookie = hdr->cookie;
1130 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1) {
1131 RTE_LOG(CRIT, MEMPOOL,
1132 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1133 obj, (const void *) mp, cookie);
1134 rte_panic("MEMPOOL: bad header cookie (put)\n");
1136 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE2;
1137 } else if (free == 1) {
1138 if (cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1139 RTE_LOG(CRIT, MEMPOOL,
1140 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1141 obj, (const void *) mp, cookie);
1142 rte_panic("MEMPOOL: bad header cookie (get)\n");
1144 hdr->cookie = RTE_MEMPOOL_HEADER_COOKIE1;
1145 } else if (free == 2) {
1146 if (cookie != RTE_MEMPOOL_HEADER_COOKIE1 &&
1147 cookie != RTE_MEMPOOL_HEADER_COOKIE2) {
1148 RTE_LOG(CRIT, MEMPOOL,
1149 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1150 obj, (const void *) mp, cookie);
1151 rte_panic("MEMPOOL: bad header cookie (audit)\n");
1154 tlr = __mempool_get_trailer(obj);
1155 cookie = tlr->cookie;
1156 if (cookie != RTE_MEMPOOL_TRAILER_COOKIE) {
1157 RTE_LOG(CRIT, MEMPOOL,
1158 "obj=%p, mempool=%p, cookie=%" PRIx64 "\n",
1159 obj, (const void *) mp, cookie);
1160 rte_panic("MEMPOOL: bad trailer cookie\n");
1165 RTE_SET_USED(obj_table_const);
1172 rte_mempool_contig_blocks_check_cookies(const struct rte_mempool *mp,
1173 void * const *first_obj_table_const, unsigned int n, int free)
1175 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1176 struct rte_mempool_info info;
1177 const size_t total_elt_sz =
1178 mp->header_size + mp->elt_size + mp->trailer_size;
1181 rte_mempool_ops_get_info(mp, &info);
1183 for (i = 0; i < n; ++i) {
1184 void *first_obj = first_obj_table_const[i];
1186 for (j = 0; j < info.contig_block_size; ++j) {
1189 obj = (void *)((uintptr_t)first_obj + j * total_elt_sz);
1190 rte_mempool_check_cookies(mp, &obj, 1, free);
1195 RTE_SET_USED(first_obj_table_const);
1201 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1203 mempool_obj_audit(struct rte_mempool *mp, __rte_unused void *opaque,
1204 void *obj, __rte_unused unsigned idx)
1206 __mempool_check_cookies(mp, &obj, 1, 2);
1210 mempool_audit_cookies(struct rte_mempool *mp)
1214 num = rte_mempool_obj_iter(mp, mempool_obj_audit, NULL);
1215 if (num != mp->size) {
1216 rte_panic("rte_mempool_obj_iter(mempool=%p, size=%u) "
1217 "iterated only over %u elements\n",
1222 #define mempool_audit_cookies(mp) do {} while(0)
1225 #ifndef __INTEL_COMPILER
1226 #pragma GCC diagnostic error "-Wcast-qual"
1229 /* check cookies before and after objects */
1231 mempool_audit_cache(const struct rte_mempool *mp)
1233 /* check cache size consistency */
1236 if (mp->cache_size == 0)
1239 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1240 const struct rte_mempool_cache *cache;
1241 cache = &mp->local_cache[lcore_id];
1242 if (cache->len > cache->flushthresh) {
1243 RTE_LOG(CRIT, MEMPOOL, "badness on cache[%u]\n",
1245 rte_panic("MEMPOOL: invalid cache len\n");
1250 /* check the consistency of mempool (size, cookies, ...) */
1252 rte_mempool_audit(struct rte_mempool *mp)
1254 mempool_audit_cache(mp);
1255 mempool_audit_cookies(mp);
1257 /* For case where mempool DEBUG is not set, and cache size is 0 */
1261 /* dump the status of the mempool on the console */
1263 rte_mempool_dump(FILE *f, struct rte_mempool *mp)
1265 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1266 struct rte_mempool_info info;
1267 struct rte_mempool_debug_stats sum;
1270 struct rte_mempool_memhdr *memhdr;
1271 unsigned common_count;
1272 unsigned cache_count;
1275 RTE_ASSERT(f != NULL);
1276 RTE_ASSERT(mp != NULL);
1278 fprintf(f, "mempool <%s>@%p\n", mp->name, mp);
1279 fprintf(f, " flags=%x\n", mp->flags);
1280 fprintf(f, " pool=%p\n", mp->pool_data);
1281 fprintf(f, " iova=0x%" PRIx64 "\n", mp->mz->iova);
1282 fprintf(f, " nb_mem_chunks=%u\n", mp->nb_mem_chunks);
1283 fprintf(f, " size=%"PRIu32"\n", mp->size);
1284 fprintf(f, " populated_size=%"PRIu32"\n", mp->populated_size);
1285 fprintf(f, " header_size=%"PRIu32"\n", mp->header_size);
1286 fprintf(f, " elt_size=%"PRIu32"\n", mp->elt_size);
1287 fprintf(f, " trailer_size=%"PRIu32"\n", mp->trailer_size);
1288 fprintf(f, " total_obj_size=%"PRIu32"\n",
1289 mp->header_size + mp->elt_size + mp->trailer_size);
1291 fprintf(f, " private_data_size=%"PRIu32"\n", mp->private_data_size);
1293 STAILQ_FOREACH(memhdr, &mp->mem_list, next)
1294 mem_len += memhdr->len;
1296 fprintf(f, " avg bytes/object=%#Lf\n",
1297 (long double)mem_len / mp->size);
1300 cache_count = rte_mempool_dump_cache(f, mp);
1301 common_count = rte_mempool_ops_get_count(mp);
1302 if ((cache_count + common_count) > mp->size)
1303 common_count = mp->size - cache_count;
1304 fprintf(f, " common_pool_count=%u\n", common_count);
1306 /* sum and dump statistics */
1307 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1308 rte_mempool_ops_get_info(mp, &info);
1309 memset(&sum, 0, sizeof(sum));
1310 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1311 sum.put_bulk += mp->stats[lcore_id].put_bulk;
1312 sum.put_objs += mp->stats[lcore_id].put_objs;
1313 sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
1314 sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
1315 sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
1316 sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
1317 sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
1318 sum.get_fail_blks += mp->stats[lcore_id].get_fail_blks;
1320 fprintf(f, " stats:\n");
1321 fprintf(f, " put_bulk=%"PRIu64"\n", sum.put_bulk);
1322 fprintf(f, " put_objs=%"PRIu64"\n", sum.put_objs);
1323 fprintf(f, " get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
1324 fprintf(f, " get_success_objs=%"PRIu64"\n", sum.get_success_objs);
1325 fprintf(f, " get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
1326 fprintf(f, " get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
1327 if (info.contig_block_size > 0) {
1328 fprintf(f, " get_success_blks=%"PRIu64"\n",
1329 sum.get_success_blks);
1330 fprintf(f, " get_fail_blks=%"PRIu64"\n", sum.get_fail_blks);
1333 fprintf(f, " no statistics available\n");
1336 rte_mempool_audit(mp);
1339 /* dump the status of all mempools on the console */
1341 rte_mempool_list_dump(FILE *f)
1343 struct rte_mempool *mp = NULL;
1344 struct rte_tailq_entry *te;
1345 struct rte_mempool_list *mempool_list;
1347 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1349 rte_mcfg_mempool_read_lock();
1351 TAILQ_FOREACH(te, mempool_list, next) {
1352 mp = (struct rte_mempool *) te->data;
1353 rte_mempool_dump(f, mp);
1356 rte_mcfg_mempool_read_unlock();
1359 /* search a mempool from its name */
1360 struct rte_mempool *
1361 rte_mempool_lookup(const char *name)
1363 struct rte_mempool *mp = NULL;
1364 struct rte_tailq_entry *te;
1365 struct rte_mempool_list *mempool_list;
1367 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1369 rte_mcfg_mempool_read_lock();
1371 TAILQ_FOREACH(te, mempool_list, next) {
1372 mp = (struct rte_mempool *) te->data;
1373 if (strncmp(name, mp->name, RTE_MEMPOOL_NAMESIZE) == 0)
1377 rte_mcfg_mempool_read_unlock();
1387 void rte_mempool_walk(void (*func)(struct rte_mempool *, void *),
1390 struct rte_tailq_entry *te = NULL;
1391 struct rte_mempool_list *mempool_list;
1394 mempool_list = RTE_TAILQ_CAST(rte_mempool_tailq.head, rte_mempool_list);
1396 rte_mcfg_mempool_read_lock();
1398 TAILQ_FOREACH_SAFE(te, mempool_list, next, tmp_te) {
1399 (*func)((struct rte_mempool *) te->data, arg);
1402 rte_mcfg_mempool_read_unlock();