1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
10 #include <sys/queue.h>
12 #include <rte_memory.h>
13 #include <rte_errno.h>
15 #include <rte_eal_memconfig.h>
16 #include <rte_launch.h>
17 #include <rte_per_lcore.h>
18 #include <rte_lcore.h>
19 #include <rte_common.h>
20 #include <rte_string_fns.h>
21 #include <rte_spinlock.h>
22 #include <rte_memcpy.h>
23 #include <rte_memzone.h>
24 #include <rte_atomic.h>
25 #include <rte_fbarray.h>
27 #include "eal_internal_cfg.h"
28 #include "eal_memalloc.h"
29 #include "eal_memcfg.h"
30 #include "eal_private.h"
31 #include "malloc_elem.h"
32 #include "malloc_heap.h"
33 #include "malloc_mp.h"
35 /* start external socket ID's at a very high number */
36 #define CONST_MAX(a, b) (a > b ? a : b) /* RTE_MAX is not a constant */
37 #define EXTERNAL_HEAP_MIN_SOCKET_ID (CONST_MAX((1 << 8), RTE_MAX_NUMA_NODES))
40 check_hugepage_sz(unsigned flags, uint64_t hugepage_sz)
42 unsigned check_flag = 0;
44 if (!(flags & ~RTE_MEMZONE_SIZE_HINT_ONLY))
47 switch (hugepage_sz) {
49 check_flag = RTE_MEMZONE_256KB;
52 check_flag = RTE_MEMZONE_2MB;
55 check_flag = RTE_MEMZONE_16MB;
58 check_flag = RTE_MEMZONE_256MB;
61 check_flag = RTE_MEMZONE_512MB;
64 check_flag = RTE_MEMZONE_1GB;
67 check_flag = RTE_MEMZONE_4GB;
70 check_flag = RTE_MEMZONE_16GB;
73 return check_flag & flags;
77 malloc_socket_to_heap_id(unsigned int socket_id)
79 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
82 for (i = 0; i < RTE_MAX_HEAPS; i++) {
83 struct malloc_heap *heap = &mcfg->malloc_heaps[i];
85 if (heap->socket_id == socket_id)
92 * Expand the heap with a memory area.
94 static struct malloc_elem *
95 malloc_heap_add_memory(struct malloc_heap *heap, struct rte_memseg_list *msl,
96 void *start, size_t len)
98 struct malloc_elem *elem = start;
100 malloc_elem_init(elem, heap, msl, len, elem, len);
102 malloc_elem_insert(elem);
104 elem = malloc_elem_join_adjacent_free(elem);
106 malloc_elem_free_list_insert(elem);
112 malloc_add_seg(const struct rte_memseg_list *msl,
113 const struct rte_memseg *ms, size_t len, void *arg __rte_unused)
115 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
116 struct rte_memseg_list *found_msl;
117 struct malloc_heap *heap;
118 int msl_idx, heap_idx;
123 heap_idx = malloc_socket_to_heap_id(msl->socket_id);
125 RTE_LOG(ERR, EAL, "Memseg list has invalid socket id\n");
128 heap = &mcfg->malloc_heaps[heap_idx];
130 /* msl is const, so find it */
131 msl_idx = msl - mcfg->memsegs;
133 if (msl_idx < 0 || msl_idx >= RTE_MAX_MEMSEG_LISTS)
136 found_msl = &mcfg->memsegs[msl_idx];
138 malloc_heap_add_memory(heap, found_msl, ms->addr, len);
140 heap->total_size += len;
142 RTE_LOG(DEBUG, EAL, "Added %zuM to heap on socket %i\n", len >> 20,
148 * Iterates through the freelist for a heap to find a free element
149 * which can store data of the required size and with the requested alignment.
150 * If size is 0, find the biggest available elem.
151 * Returns null on failure, or pointer to element on success.
153 static struct malloc_elem *
154 find_suitable_element(struct malloc_heap *heap, size_t size,
155 unsigned int flags, size_t align, size_t bound, bool contig)
158 struct malloc_elem *elem, *alt_elem = NULL;
160 for (idx = malloc_elem_free_list_index(size);
161 idx < RTE_HEAP_NUM_FREELISTS; idx++) {
162 for (elem = LIST_FIRST(&heap->free_head[idx]);
163 !!elem; elem = LIST_NEXT(elem, free_list)) {
164 if (malloc_elem_can_hold(elem, size, align, bound,
166 if (check_hugepage_sz(flags,
169 if (alt_elem == NULL)
175 if ((alt_elem != NULL) && (flags & RTE_MEMZONE_SIZE_HINT_ONLY))
182 * Iterates through the freelist for a heap to find a free element with the
183 * biggest size and requested alignment. Will also set size to whatever element
184 * size that was found.
185 * Returns null on failure, or pointer to element on success.
187 static struct malloc_elem *
188 find_biggest_element(struct malloc_heap *heap, size_t *size,
189 unsigned int flags, size_t align, bool contig)
191 struct malloc_elem *elem, *max_elem = NULL;
192 size_t idx, max_size = 0;
194 for (idx = 0; idx < RTE_HEAP_NUM_FREELISTS; idx++) {
195 for (elem = LIST_FIRST(&heap->free_head[idx]);
196 !!elem; elem = LIST_NEXT(elem, free_list)) {
198 if ((flags & RTE_MEMZONE_SIZE_HINT_ONLY) == 0 &&
199 !check_hugepage_sz(flags,
204 malloc_elem_find_max_iova_contig(elem,
207 void *data_start = RTE_PTR_ADD(elem,
208 MALLOC_ELEM_HEADER_LEN);
209 void *data_end = RTE_PTR_ADD(elem, elem->size -
210 MALLOC_ELEM_TRAILER_LEN);
211 void *aligned = RTE_PTR_ALIGN_CEIL(data_start,
213 /* check if aligned data start is beyond end */
214 if (aligned >= data_end)
216 cur_size = RTE_PTR_DIFF(data_end, aligned);
218 if (cur_size > max_size) {
230 * Main function to allocate a block of memory from the heap.
231 * It locks the free list, scans it, and adds a new memseg if the
232 * scan fails. Once the new memseg is added, it re-scans and should return
233 * the new element after releasing the lock.
236 heap_alloc(struct malloc_heap *heap, const char *type __rte_unused, size_t size,
237 unsigned int flags, size_t align, size_t bound, bool contig)
239 struct malloc_elem *elem;
241 size = RTE_CACHE_LINE_ROUNDUP(size);
242 align = RTE_CACHE_LINE_ROUNDUP(align);
244 /* roundup might cause an overflow */
247 elem = find_suitable_element(heap, size, flags, align, bound, contig);
249 elem = malloc_elem_alloc(elem, size, align, bound, contig);
251 /* increase heap's count of allocated elements */
255 return elem == NULL ? NULL : (void *)(&elem[1]);
259 heap_alloc_biggest(struct malloc_heap *heap, const char *type __rte_unused,
260 unsigned int flags, size_t align, bool contig)
262 struct malloc_elem *elem;
265 align = RTE_CACHE_LINE_ROUNDUP(align);
267 elem = find_biggest_element(heap, &size, flags, align, contig);
269 elem = malloc_elem_alloc(elem, size, align, 0, contig);
271 /* increase heap's count of allocated elements */
275 return elem == NULL ? NULL : (void *)(&elem[1]);
278 /* this function is exposed in malloc_mp.h */
280 rollback_expand_heap(struct rte_memseg **ms, int n_segs,
281 struct malloc_elem *elem, void *map_addr, size_t map_len)
284 malloc_elem_free_list_remove(elem);
285 malloc_elem_hide_region(elem, map_addr, map_len);
288 eal_memalloc_free_seg_bulk(ms, n_segs);
291 /* this function is exposed in malloc_mp.h */
293 alloc_pages_on_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
294 int socket, unsigned int flags, size_t align, size_t bound,
295 bool contig, struct rte_memseg **ms, int n_segs)
297 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
298 struct rte_memseg_list *msl;
299 struct malloc_elem *elem = NULL;
302 void *ret, *map_addr;
304 alloc_sz = (size_t)pg_sz * n_segs;
306 /* first, check if we're allowed to allocate this memory */
307 if (eal_memalloc_mem_alloc_validate(socket,
308 heap->total_size + alloc_sz) < 0) {
309 RTE_LOG(DEBUG, EAL, "User has disallowed allocation\n");
313 allocd_pages = eal_memalloc_alloc_seg_bulk(ms, n_segs, pg_sz,
316 /* make sure we've allocated our pages... */
317 if (allocd_pages < 0)
320 map_addr = ms[0]->addr;
321 msl = rte_mem_virt2memseg_list(map_addr);
323 /* check if we wanted contiguous memory but didn't get it */
324 if (contig && !eal_memalloc_is_contig(msl, map_addr, alloc_sz)) {
325 RTE_LOG(DEBUG, EAL, "%s(): couldn't allocate physically contiguous space\n",
331 * Once we have all the memseg lists configured, if there is a dma mask
332 * set, check iova addresses are not out of range. Otherwise the device
333 * setting the dma mask could have problems with the mapped memory.
335 * There are two situations when this can happen:
336 * 1) memory initialization
337 * 2) dynamic memory allocation
339 * For 1), an error when checking dma mask implies app can not be
340 * executed. For 2) implies the new memory can not be added.
342 if (mcfg->dma_maskbits &&
343 rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
345 * Currently this can only happen if IOMMU is enabled
346 * and the address width supported by the IOMMU hw is
347 * not enough for using the memory mapped IOVAs.
349 * If IOVA is VA, advice to try with '--iova-mode pa'
350 * which could solve some situations when IOVA VA is not
354 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask\n",
358 * If IOVA is VA and it is possible to run with IOVA PA,
359 * because user is root, give and advice for solving the
362 if ((rte_eal_iova_mode() == RTE_IOVA_VA) &&
363 rte_eal_using_phys_addrs())
365 "%s(): Please try initializing EAL with --iova-mode=pa parameter\n",
370 /* add newly minted memsegs to malloc heap */
371 elem = malloc_heap_add_memory(heap, msl, map_addr, alloc_sz);
373 /* try once more, as now we have allocated new memory */
374 ret = find_suitable_element(heap, elt_size, flags, align, bound,
383 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
388 try_expand_heap_primary(struct malloc_heap *heap, uint64_t pg_sz,
389 size_t elt_size, int socket, unsigned int flags, size_t align,
390 size_t bound, bool contig)
392 struct malloc_elem *elem;
393 struct rte_memseg **ms;
397 bool callback_triggered = false;
399 alloc_sz = RTE_ALIGN_CEIL(align + elt_size +
400 MALLOC_ELEM_TRAILER_LEN, pg_sz);
401 n_segs = alloc_sz / pg_sz;
403 /* we can't know in advance how many pages we'll need, so we malloc */
404 ms = malloc(sizeof(*ms) * n_segs);
407 memset(ms, 0, sizeof(*ms) * n_segs);
409 elem = alloc_pages_on_heap(heap, pg_sz, elt_size, socket, flags, align,
410 bound, contig, ms, n_segs);
415 map_addr = ms[0]->addr;
417 /* notify user about changes in memory map */
418 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC, map_addr, alloc_sz);
420 /* notify other processes that this has happened */
421 if (request_sync()) {
422 /* we couldn't ensure all processes have mapped memory,
423 * so free it back and notify everyone that it's been
426 * technically, we could've avoided adding memory addresses to
427 * the map, but that would've led to inconsistent behavior
428 * between primary and secondary processes, as those get
429 * callbacks during sync. therefore, force primary process to
430 * do alloc-and-rollback syncs as well.
432 callback_triggered = true;
435 heap->total_size += alloc_sz;
437 RTE_LOG(DEBUG, EAL, "Heap on socket %d was expanded by %zdMB\n",
438 socket, alloc_sz >> 20ULL);
445 if (callback_triggered)
446 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
449 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
459 try_expand_heap_secondary(struct malloc_heap *heap, uint64_t pg_sz,
460 size_t elt_size, int socket, unsigned int flags, size_t align,
461 size_t bound, bool contig)
463 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
464 struct malloc_mp_req req;
467 memset(&req, 0, sizeof(req));
469 req.t = REQ_TYPE_ALLOC;
470 req.alloc_req.align = align;
471 req.alloc_req.bound = bound;
472 req.alloc_req.contig = contig;
473 req.alloc_req.flags = flags;
474 req.alloc_req.elt_size = elt_size;
475 req.alloc_req.page_sz = pg_sz;
476 req.alloc_req.socket = socket;
477 req.alloc_req.malloc_heap_idx = heap - mcfg->malloc_heaps;
479 req_result = request_to_primary(&req);
484 if (req.result != REQ_RESULT_SUCCESS)
491 try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
492 int socket, unsigned int flags, size_t align, size_t bound,
497 rte_mcfg_mem_write_lock();
499 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
500 ret = try_expand_heap_primary(heap, pg_sz, elt_size, socket,
501 flags, align, bound, contig);
503 ret = try_expand_heap_secondary(heap, pg_sz, elt_size, socket,
504 flags, align, bound, contig);
507 rte_mcfg_mem_write_unlock();
512 compare_pagesz(const void *a, const void *b)
514 const struct rte_memseg_list * const*mpa = a;
515 const struct rte_memseg_list * const*mpb = b;
516 const struct rte_memseg_list *msla = *mpa;
517 const struct rte_memseg_list *mslb = *mpb;
518 uint64_t pg_sz_a = msla->page_sz;
519 uint64_t pg_sz_b = mslb->page_sz;
521 if (pg_sz_a < pg_sz_b)
523 if (pg_sz_a > pg_sz_b)
529 alloc_more_mem_on_socket(struct malloc_heap *heap, size_t size, int socket,
530 unsigned int flags, size_t align, size_t bound, bool contig)
532 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
533 struct rte_memseg_list *requested_msls[RTE_MAX_MEMSEG_LISTS];
534 struct rte_memseg_list *other_msls[RTE_MAX_MEMSEG_LISTS];
535 uint64_t requested_pg_sz[RTE_MAX_MEMSEG_LISTS];
536 uint64_t other_pg_sz[RTE_MAX_MEMSEG_LISTS];
538 int i, n_other_msls, n_other_pg_sz, n_requested_msls, n_requested_pg_sz;
539 bool size_hint = (flags & RTE_MEMZONE_SIZE_HINT_ONLY) > 0;
540 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
543 memset(requested_msls, 0, sizeof(requested_msls));
544 memset(other_msls, 0, sizeof(other_msls));
545 memset(requested_pg_sz, 0, sizeof(requested_pg_sz));
546 memset(other_pg_sz, 0, sizeof(other_pg_sz));
549 * go through memseg list and take note of all the page sizes available,
550 * and if any of them were specifically requested by the user.
552 n_requested_msls = 0;
554 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
555 struct rte_memseg_list *msl = &mcfg->memsegs[i];
557 if (msl->socket_id != socket)
560 if (msl->base_va == NULL)
563 /* if pages of specific size were requested */
564 if (size_flags != 0 && check_hugepage_sz(size_flags,
566 requested_msls[n_requested_msls++] = msl;
567 else if (size_flags == 0 || size_hint)
568 other_msls[n_other_msls++] = msl;
571 /* sort the lists, smallest first */
572 qsort(requested_msls, n_requested_msls, sizeof(requested_msls[0]),
574 qsort(other_msls, n_other_msls, sizeof(other_msls[0]),
577 /* now, extract page sizes we are supposed to try */
579 n_requested_pg_sz = 0;
580 for (i = 0; i < n_requested_msls; i++) {
581 uint64_t pg_sz = requested_msls[i]->page_sz;
583 if (prev_pg_sz != pg_sz) {
584 requested_pg_sz[n_requested_pg_sz++] = pg_sz;
590 for (i = 0; i < n_other_msls; i++) {
591 uint64_t pg_sz = other_msls[i]->page_sz;
593 if (prev_pg_sz != pg_sz) {
594 other_pg_sz[n_other_pg_sz++] = pg_sz;
599 /* finally, try allocating memory of specified page sizes, starting from
602 for (i = 0; i < n_requested_pg_sz; i++) {
603 uint64_t pg_sz = requested_pg_sz[i];
606 * do not pass the size hint here, as user expects other page
607 * sizes first, before resorting to best effort allocation.
609 if (!try_expand_heap(heap, pg_sz, size, socket, size_flags,
610 align, bound, contig))
613 if (n_other_pg_sz == 0)
616 /* now, check if we can reserve anything with size hint */
617 ret = find_suitable_element(heap, size, flags, align, bound, contig);
622 * we still couldn't reserve memory, so try expanding heap with other
623 * page sizes, if there are any
625 for (i = 0; i < n_other_pg_sz; i++) {
626 uint64_t pg_sz = other_pg_sz[i];
628 if (!try_expand_heap(heap, pg_sz, size, socket, flags,
629 align, bound, contig))
635 /* this will try lower page sizes first */
637 malloc_heap_alloc_on_heap_id(const char *type, size_t size,
638 unsigned int heap_id, unsigned int flags, size_t align,
639 size_t bound, bool contig)
641 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
642 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
643 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
646 const struct internal_config *internal_conf =
647 eal_get_internal_configuration();
649 rte_spinlock_lock(&(heap->lock));
651 align = align == 0 ? 1 : align;
653 /* for legacy mode, try once and with all flags */
654 if (internal_conf->legacy_mem) {
655 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
660 * we do not pass the size hint here, because even if allocation fails,
661 * we may still be able to allocate memory from appropriate page sizes,
662 * we just need to request more memory first.
665 socket_id = rte_socket_id_by_idx(heap_id);
667 * if socket ID is negative, we cannot find a socket ID for this heap -
668 * which means it's an external heap. those can have unexpected page
669 * sizes, so if the user asked to allocate from there - assume user
670 * knows what they're doing, and allow allocating from there with any
674 size_flags |= RTE_MEMZONE_SIZE_HINT_ONLY;
676 ret = heap_alloc(heap, type, size, size_flags, align, bound, contig);
680 /* if socket ID is invalid, this is an external heap */
684 if (!alloc_more_mem_on_socket(heap, size, socket_id, flags, align,
686 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
688 /* this should have succeeded */
690 RTE_LOG(ERR, EAL, "Error allocating from heap\n");
693 rte_spinlock_unlock(&(heap->lock));
698 malloc_get_numa_socket(void)
700 const struct internal_config *internal_conf =
701 eal_get_internal_configuration();
702 unsigned socket_id = rte_socket_id();
704 if (socket_id != (unsigned)SOCKET_ID_ANY)
707 /* return first id where memory is available, otherwise 0 */
708 for (socket_id = 0; socket_id < RTE_MAX_NUMA_NODES; ++socket_id)
709 if (internal_conf->socket_mem[socket_id] != 0)
716 malloc_heap_alloc(const char *type, size_t size, int socket_arg,
717 unsigned int flags, size_t align, size_t bound, bool contig)
719 int socket, heap_id, i;
722 /* return NULL if size is 0 or alignment is not power-of-2 */
723 if (size == 0 || (align && !rte_is_power_of_2(align)))
726 if (!rte_eal_has_hugepages() && socket_arg < RTE_MAX_NUMA_NODES)
727 socket_arg = SOCKET_ID_ANY;
729 if (socket_arg == SOCKET_ID_ANY)
730 socket = malloc_get_numa_socket();
734 /* turn socket ID into heap ID */
735 heap_id = malloc_socket_to_heap_id(socket);
736 /* if heap id is negative, socket ID was invalid */
740 ret = malloc_heap_alloc_on_heap_id(type, size, heap_id, flags, align,
742 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
745 /* try other heaps. we are only iterating through native DPDK sockets,
746 * so external heaps won't be included.
748 for (i = 0; i < (int) rte_socket_count(); i++) {
751 ret = malloc_heap_alloc_on_heap_id(type, size, i, flags, align,
760 heap_alloc_biggest_on_heap_id(const char *type, unsigned int heap_id,
761 unsigned int flags, size_t align, bool contig)
763 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
764 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
767 rte_spinlock_lock(&(heap->lock));
769 align = align == 0 ? 1 : align;
771 ret = heap_alloc_biggest(heap, type, flags, align, contig);
773 rte_spinlock_unlock(&(heap->lock));
779 malloc_heap_alloc_biggest(const char *type, int socket_arg, unsigned int flags,
780 size_t align, bool contig)
782 int socket, i, cur_socket, heap_id;
785 /* return NULL if align is not power-of-2 */
786 if ((align && !rte_is_power_of_2(align)))
789 if (!rte_eal_has_hugepages())
790 socket_arg = SOCKET_ID_ANY;
792 if (socket_arg == SOCKET_ID_ANY)
793 socket = malloc_get_numa_socket();
797 /* turn socket ID into heap ID */
798 heap_id = malloc_socket_to_heap_id(socket);
799 /* if heap id is negative, socket ID was invalid */
803 ret = heap_alloc_biggest_on_heap_id(type, heap_id, flags, align,
805 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
808 /* try other heaps */
809 for (i = 0; i < (int) rte_socket_count(); i++) {
810 cur_socket = rte_socket_id_by_idx(i);
811 if (cur_socket == socket)
813 ret = heap_alloc_biggest_on_heap_id(type, i, flags, align,
821 /* this function is exposed in malloc_mp.h */
823 malloc_heap_free_pages(void *aligned_start, size_t aligned_len)
825 int n_segs, seg_idx, max_seg_idx;
826 struct rte_memseg_list *msl;
829 msl = rte_mem_virt2memseg_list(aligned_start);
833 page_sz = (size_t)msl->page_sz;
834 n_segs = aligned_len / page_sz;
835 seg_idx = RTE_PTR_DIFF(aligned_start, msl->base_va) / page_sz;
836 max_seg_idx = seg_idx + n_segs;
838 for (; seg_idx < max_seg_idx; seg_idx++) {
839 struct rte_memseg *ms;
841 ms = rte_fbarray_get(&msl->memseg_arr, seg_idx);
842 eal_memalloc_free_seg(ms);
848 malloc_heap_free(struct malloc_elem *elem)
850 struct malloc_heap *heap;
851 void *start, *aligned_start, *end, *aligned_end;
852 size_t len, aligned_len, page_sz;
853 struct rte_memseg_list *msl;
854 unsigned int i, n_segs, before_space, after_space;
856 const struct internal_config *internal_conf =
857 eal_get_internal_configuration();
859 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
862 /* elem may be merged with previous element, so keep heap address */
865 page_sz = (size_t)msl->page_sz;
867 rte_spinlock_lock(&(heap->lock));
869 /* mark element as free */
870 elem->state = ELEM_FREE;
872 elem = malloc_elem_free(elem);
874 /* anything after this is a bonus */
877 /* ...of which we can't avail if we are in legacy mode, or if this is an
878 * externally allocated segment.
880 if (internal_conf->legacy_mem || (msl->external > 0))
883 /* check if we can free any memory back to the system */
884 if (elem->size < page_sz)
887 /* if user requested to match allocations, the sizes must match - if not,
888 * we will defer freeing these hugepages until the entire original allocation
891 if (internal_conf->match_allocations && elem->size != elem->orig_size)
894 /* probably, but let's make sure, as we may not be using up full page */
897 aligned_start = RTE_PTR_ALIGN_CEIL(start, page_sz);
898 end = RTE_PTR_ADD(elem, len);
899 aligned_end = RTE_PTR_ALIGN_FLOOR(end, page_sz);
901 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
903 /* can't free anything */
904 if (aligned_len < page_sz)
907 /* we can free something. however, some of these pages may be marked as
908 * unfreeable, so also check that as well
910 n_segs = aligned_len / page_sz;
911 for (i = 0; i < n_segs; i++) {
912 const struct rte_memseg *tmp =
913 rte_mem_virt2memseg(aligned_start, msl);
915 if (tmp->flags & RTE_MEMSEG_FLAG_DO_NOT_FREE) {
916 /* this is an unfreeable segment, so move start */
917 aligned_start = RTE_PTR_ADD(tmp->addr, tmp->len);
921 /* recalculate length and number of segments */
922 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
923 n_segs = aligned_len / page_sz;
925 /* check if we can still free some pages */
929 /* We're not done yet. We also have to check if by freeing space we will
930 * be leaving free elements that are too small to store new elements.
931 * Check if we have enough space in the beginning and at the end, or if
932 * start/end are exactly page aligned.
934 before_space = RTE_PTR_DIFF(aligned_start, elem);
935 after_space = RTE_PTR_DIFF(end, aligned_end);
936 if (before_space != 0 &&
937 before_space < MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
938 /* There is not enough space before start, but we may be able to
939 * move the start forward by one page.
945 aligned_start = RTE_PTR_ADD(aligned_start, page_sz);
946 aligned_len -= page_sz;
949 if (after_space != 0 && after_space <
950 MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
951 /* There is not enough space after end, but we may be able to
952 * move the end backwards by one page.
958 aligned_end = RTE_PTR_SUB(aligned_end, page_sz);
959 aligned_len -= page_sz;
963 /* now we can finally free us some pages */
965 rte_mcfg_mem_write_lock();
968 * we allow secondary processes to clear the heap of this allocated
969 * memory because it is safe to do so, as even if notifications about
970 * unmapped pages don't make it to other processes, heap is shared
971 * across all processes, and will become empty of this memory anyway,
972 * and nothing can allocate it back unless primary process will be able
973 * to deliver allocation message to every single running process.
976 malloc_elem_free_list_remove(elem);
978 malloc_elem_hide_region(elem, (void *) aligned_start, aligned_len);
980 heap->total_size -= aligned_len;
982 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
983 /* notify user about changes in memory map */
984 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
985 aligned_start, aligned_len);
987 /* don't care if any of this fails */
988 malloc_heap_free_pages(aligned_start, aligned_len);
992 struct malloc_mp_req req;
994 memset(&req, 0, sizeof(req));
996 req.t = REQ_TYPE_FREE;
997 req.free_req.addr = aligned_start;
998 req.free_req.len = aligned_len;
1001 * we request primary to deallocate pages, but we don't do it
1002 * in this thread. instead, we notify primary that we would like
1003 * to deallocate pages, and this process will receive another
1004 * request (in parallel) that will do it for us on another
1007 * we also don't really care if this succeeds - the data is
1008 * already removed from the heap, so it is, for all intents and
1009 * purposes, hidden from the rest of DPDK even if some other
1010 * process (including this one) may have these pages mapped.
1012 * notifications about deallocated memory happen during sync.
1014 request_to_primary(&req);
1017 RTE_LOG(DEBUG, EAL, "Heap on socket %d was shrunk by %zdMB\n",
1018 msl->socket_id, aligned_len >> 20ULL);
1020 rte_mcfg_mem_write_unlock();
1022 rte_spinlock_unlock(&(heap->lock));
1027 malloc_heap_resize(struct malloc_elem *elem, size_t size)
1031 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
1034 rte_spinlock_lock(&(elem->heap->lock));
1036 ret = malloc_elem_resize(elem, size);
1038 rte_spinlock_unlock(&(elem->heap->lock));
1044 * Function to retrieve data for a given heap
1047 malloc_heap_get_stats(struct malloc_heap *heap,
1048 struct rte_malloc_socket_stats *socket_stats)
1051 struct malloc_elem *elem;
1053 rte_spinlock_lock(&heap->lock);
1055 /* Initialise variables for heap */
1056 socket_stats->free_count = 0;
1057 socket_stats->heap_freesz_bytes = 0;
1058 socket_stats->greatest_free_size = 0;
1060 /* Iterate through free list */
1061 for (idx = 0; idx < RTE_HEAP_NUM_FREELISTS; idx++) {
1062 for (elem = LIST_FIRST(&heap->free_head[idx]);
1063 !!elem; elem = LIST_NEXT(elem, free_list))
1065 socket_stats->free_count++;
1066 socket_stats->heap_freesz_bytes += elem->size;
1067 if (elem->size > socket_stats->greatest_free_size)
1068 socket_stats->greatest_free_size = elem->size;
1071 /* Get stats on overall heap and allocated memory on this heap */
1072 socket_stats->heap_totalsz_bytes = heap->total_size;
1073 socket_stats->heap_allocsz_bytes = (socket_stats->heap_totalsz_bytes -
1074 socket_stats->heap_freesz_bytes);
1075 socket_stats->alloc_count = heap->alloc_count;
1077 rte_spinlock_unlock(&heap->lock);
1082 * Function to retrieve data for a given heap
1085 malloc_heap_dump(struct malloc_heap *heap, FILE *f)
1087 struct malloc_elem *elem;
1089 rte_spinlock_lock(&heap->lock);
1091 fprintf(f, "Heap size: 0x%zx\n", heap->total_size);
1092 fprintf(f, "Heap alloc count: %u\n", heap->alloc_count);
1096 malloc_elem_dump(elem, f);
1100 rte_spinlock_unlock(&heap->lock);
1104 destroy_elem(struct malloc_elem *elem, size_t len)
1106 struct malloc_heap *heap = elem->heap;
1108 /* notify all subscribers that a memory area is going to be removed */
1109 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE, elem, len);
1111 /* this element can be removed */
1112 malloc_elem_free_list_remove(elem);
1113 malloc_elem_hide_region(elem, elem, len);
1115 heap->total_size -= len;
1117 memset(elem, 0, sizeof(*elem));
1122 struct rte_memseg_list *
1123 malloc_heap_create_external_seg(void *va_addr, rte_iova_t iova_addrs[],
1124 unsigned int n_pages, size_t page_sz, const char *seg_name,
1125 unsigned int socket_id)
1127 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1128 char fbarray_name[RTE_FBARRAY_NAME_LEN];
1129 struct rte_memseg_list *msl = NULL;
1130 struct rte_fbarray *arr;
1131 size_t seg_len = n_pages * page_sz;
1134 /* first, find a free memseg list */
1135 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1136 struct rte_memseg_list *tmp = &mcfg->memsegs[i];
1137 if (tmp->base_va == NULL) {
1143 RTE_LOG(ERR, EAL, "Couldn't find empty memseg list\n");
1148 snprintf(fbarray_name, sizeof(fbarray_name), "%s_%p",
1151 /* create the backing fbarray */
1152 if (rte_fbarray_init(&msl->memseg_arr, fbarray_name, n_pages,
1153 sizeof(struct rte_memseg)) < 0) {
1154 RTE_LOG(ERR, EAL, "Couldn't create fbarray backing the memseg list\n");
1157 arr = &msl->memseg_arr;
1159 /* fbarray created, fill it up */
1160 for (i = 0; i < n_pages; i++) {
1161 struct rte_memseg *ms;
1163 rte_fbarray_set_used(arr, i);
1164 ms = rte_fbarray_get(arr, i);
1165 ms->addr = RTE_PTR_ADD(va_addr, i * page_sz);
1166 ms->iova = iova_addrs == NULL ? RTE_BAD_IOVA : iova_addrs[i];
1167 ms->hugepage_sz = page_sz;
1169 ms->nchannel = rte_memory_get_nchannel();
1170 ms->nrank = rte_memory_get_nrank();
1171 ms->socket_id = socket_id;
1174 /* set up the memseg list */
1175 msl->base_va = va_addr;
1176 msl->page_sz = page_sz;
1177 msl->socket_id = socket_id;
1185 struct extseg_walk_arg {
1188 struct rte_memseg_list *msl;
1192 extseg_walk(const struct rte_memseg_list *msl, void *arg)
1194 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1195 struct extseg_walk_arg *wa = arg;
1197 if (msl->base_va == wa->va_addr && msl->len == wa->len) {
1198 unsigned int found_idx;
1201 found_idx = msl - mcfg->memsegs;
1202 wa->msl = &mcfg->memsegs[found_idx];
1208 struct rte_memseg_list *
1209 malloc_heap_find_external_seg(void *va_addr, size_t len)
1211 struct extseg_walk_arg wa;
1214 wa.va_addr = va_addr;
1217 res = rte_memseg_list_walk_thread_unsafe(extseg_walk, &wa);
1220 /* 0 means nothing was found, -1 shouldn't happen */
1229 malloc_heap_destroy_external_seg(struct rte_memseg_list *msl)
1231 /* destroy the fbarray backing this memory */
1232 if (rte_fbarray_destroy(&msl->memseg_arr) < 0)
1235 /* reset the memseg list */
1236 memset(msl, 0, sizeof(*msl));
1242 malloc_heap_add_external_memory(struct malloc_heap *heap,
1243 struct rte_memseg_list *msl)
1245 /* erase contents of new memory */
1246 memset(msl->base_va, 0, msl->len);
1248 /* now, add newly minted memory to the malloc heap */
1249 malloc_heap_add_memory(heap, msl, msl->base_va, msl->len);
1251 heap->total_size += msl->len;
1254 RTE_LOG(DEBUG, EAL, "Added segment for heap %s starting at %p\n",
1255 heap->name, msl->base_va);
1257 /* notify all subscribers that a new memory area has been added */
1258 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC,
1259 msl->base_va, msl->len);
1265 malloc_heap_remove_external_memory(struct malloc_heap *heap, void *va_addr,
1268 struct malloc_elem *elem = heap->first;
1270 /* find element with specified va address */
1271 while (elem != NULL && elem != va_addr) {
1273 /* stop if we've blown past our VA */
1274 if (elem > (struct malloc_elem *)va_addr) {
1279 /* check if element was found */
1280 if (elem == NULL || elem->msl->len != len) {
1284 /* if element's size is not equal to segment len, segment is busy */
1285 if (elem->state == ELEM_BUSY || elem->size != len) {
1289 return destroy_elem(elem, len);
1293 malloc_heap_create(struct malloc_heap *heap, const char *heap_name)
1295 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1296 uint32_t next_socket_id = mcfg->next_socket_id;
1298 /* prevent overflow. did you really create 2 billion heaps??? */
1299 if (next_socket_id > INT32_MAX) {
1300 RTE_LOG(ERR, EAL, "Cannot assign new socket ID's\n");
1305 /* initialize empty heap */
1306 heap->alloc_count = 0;
1309 LIST_INIT(heap->free_head);
1310 rte_spinlock_init(&heap->lock);
1311 heap->total_size = 0;
1312 heap->socket_id = next_socket_id;
1314 /* we hold a global mem hotplug writelock, so it's safe to increment */
1315 mcfg->next_socket_id++;
1318 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1323 malloc_heap_destroy(struct malloc_heap *heap)
1325 if (heap->alloc_count != 0) {
1326 RTE_LOG(ERR, EAL, "Heap is still in use\n");
1330 if (heap->first != NULL || heap->last != NULL) {
1331 RTE_LOG(ERR, EAL, "Heap still contains memory segments\n");
1335 if (heap->total_size != 0)
1336 RTE_LOG(ERR, EAL, "Total size not zero, heap is likely corrupt\n");
1338 /* after this, the lock will be dropped */
1339 memset(heap, 0, sizeof(*heap));
1345 rte_eal_malloc_heap_init(void)
1347 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1349 const struct internal_config *internal_conf =
1350 eal_get_internal_configuration();
1352 if (internal_conf->match_allocations)
1353 RTE_LOG(DEBUG, EAL, "Hugepages will be freed exactly as allocated.\n");
1355 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1356 /* assign min socket ID to external heaps */
1357 mcfg->next_socket_id = EXTERNAL_HEAP_MIN_SOCKET_ID;
1359 /* assign names to default DPDK heaps */
1360 for (i = 0; i < rte_socket_count(); i++) {
1361 struct malloc_heap *heap = &mcfg->malloc_heaps[i];
1362 char heap_name[RTE_HEAP_NAME_MAX_LEN];
1363 int socket_id = rte_socket_id_by_idx(i);
1365 snprintf(heap_name, sizeof(heap_name),
1366 "socket_%i", socket_id);
1367 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1368 heap->socket_id = socket_id;
1373 if (register_mp_requests()) {
1374 RTE_LOG(ERR, EAL, "Couldn't register malloc multiprocess actions\n");
1375 rte_mcfg_mem_read_unlock();
1379 /* unlock mem hotplug here. it's safe for primary as no requests can
1380 * even come before primary itself is fully initialized, and secondaries
1381 * do not need to initialize the heap.
1383 rte_mcfg_mem_read_unlock();
1385 /* secondary process does not need to initialize anything */
1386 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1389 /* add all IOVA-contiguous areas to the heap */
1390 return rte_memseg_contig_walk(malloc_add_seg, NULL);