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;
240 size_t user_size = size;
242 size = RTE_CACHE_LINE_ROUNDUP(size);
243 align = RTE_CACHE_LINE_ROUNDUP(align);
245 /* roundup might cause an overflow */
248 elem = find_suitable_element(heap, size, flags, align, bound, contig);
250 elem = malloc_elem_alloc(elem, size, align, bound, contig);
252 /* increase heap's count of allocated elements */
255 asan_set_redzone(elem, user_size);
258 return elem == NULL ? NULL : (void *)(&elem[1]);
262 heap_alloc_biggest(struct malloc_heap *heap, const char *type __rte_unused,
263 unsigned int flags, size_t align, bool contig)
265 struct malloc_elem *elem;
268 align = RTE_CACHE_LINE_ROUNDUP(align);
270 elem = find_biggest_element(heap, &size, flags, align, contig);
272 elem = malloc_elem_alloc(elem, size, align, 0, contig);
274 /* increase heap's count of allocated elements */
277 asan_set_redzone(elem, size);
280 return elem == NULL ? NULL : (void *)(&elem[1]);
283 /* this function is exposed in malloc_mp.h */
285 rollback_expand_heap(struct rte_memseg **ms, int n_segs,
286 struct malloc_elem *elem, void *map_addr, size_t map_len)
289 malloc_elem_free_list_remove(elem);
290 malloc_elem_hide_region(elem, map_addr, map_len);
293 eal_memalloc_free_seg_bulk(ms, n_segs);
296 /* this function is exposed in malloc_mp.h */
298 alloc_pages_on_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
299 int socket, unsigned int flags, size_t align, size_t bound,
300 bool contig, struct rte_memseg **ms, int n_segs)
302 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
303 struct rte_memseg_list *msl;
304 struct malloc_elem *elem = NULL;
307 void *ret, *map_addr;
309 alloc_sz = (size_t)pg_sz * n_segs;
311 /* first, check if we're allowed to allocate this memory */
312 if (eal_memalloc_mem_alloc_validate(socket,
313 heap->total_size + alloc_sz) < 0) {
314 RTE_LOG(DEBUG, EAL, "User has disallowed allocation\n");
318 allocd_pages = eal_memalloc_alloc_seg_bulk(ms, n_segs, pg_sz,
321 /* make sure we've allocated our pages... */
322 if (allocd_pages < 0)
325 map_addr = ms[0]->addr;
326 msl = rte_mem_virt2memseg_list(map_addr);
328 /* check if we wanted contiguous memory but didn't get it */
329 if (contig && !eal_memalloc_is_contig(msl, map_addr, alloc_sz)) {
330 RTE_LOG(DEBUG, EAL, "%s(): couldn't allocate physically contiguous space\n",
336 * Once we have all the memseg lists configured, if there is a dma mask
337 * set, check iova addresses are not out of range. Otherwise the device
338 * setting the dma mask could have problems with the mapped memory.
340 * There are two situations when this can happen:
341 * 1) memory initialization
342 * 2) dynamic memory allocation
344 * For 1), an error when checking dma mask implies app can not be
345 * executed. For 2) implies the new memory can not be added.
347 if (mcfg->dma_maskbits &&
348 rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
350 * Currently this can only happen if IOMMU is enabled
351 * and the address width supported by the IOMMU hw is
352 * not enough for using the memory mapped IOVAs.
354 * If IOVA is VA, advice to try with '--iova-mode pa'
355 * which could solve some situations when IOVA VA is not
359 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask\n",
363 * If IOVA is VA and it is possible to run with IOVA PA,
364 * because user is root, give and advice for solving the
367 if ((rte_eal_iova_mode() == RTE_IOVA_VA) &&
368 rte_eal_using_phys_addrs())
370 "%s(): Please try initializing EAL with --iova-mode=pa parameter\n",
375 /* add newly minted memsegs to malloc heap */
376 elem = malloc_heap_add_memory(heap, msl, map_addr, alloc_sz);
378 /* try once more, as now we have allocated new memory */
379 ret = find_suitable_element(heap, elt_size, flags, align, bound,
388 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
393 try_expand_heap_primary(struct malloc_heap *heap, uint64_t pg_sz,
394 size_t elt_size, int socket, unsigned int flags, size_t align,
395 size_t bound, bool contig)
397 struct malloc_elem *elem;
398 struct rte_memseg **ms;
402 bool callback_triggered = false;
404 alloc_sz = RTE_ALIGN_CEIL(align + elt_size +
405 MALLOC_ELEM_TRAILER_LEN, pg_sz);
406 n_segs = alloc_sz / pg_sz;
408 /* we can't know in advance how many pages we'll need, so we malloc */
409 ms = malloc(sizeof(*ms) * n_segs);
412 memset(ms, 0, sizeof(*ms) * n_segs);
414 elem = alloc_pages_on_heap(heap, pg_sz, elt_size, socket, flags, align,
415 bound, contig, ms, n_segs);
420 map_addr = ms[0]->addr;
422 /* notify user about changes in memory map */
423 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC, map_addr, alloc_sz);
425 /* notify other processes that this has happened */
426 if (request_sync()) {
427 /* we couldn't ensure all processes have mapped memory,
428 * so free it back and notify everyone that it's been
431 * technically, we could've avoided adding memory addresses to
432 * the map, but that would've led to inconsistent behavior
433 * between primary and secondary processes, as those get
434 * callbacks during sync. therefore, force primary process to
435 * do alloc-and-rollback syncs as well.
437 callback_triggered = true;
440 heap->total_size += alloc_sz;
442 RTE_LOG(DEBUG, EAL, "Heap on socket %d was expanded by %zdMB\n",
443 socket, alloc_sz >> 20ULL);
450 if (callback_triggered)
451 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
454 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
464 try_expand_heap_secondary(struct malloc_heap *heap, uint64_t pg_sz,
465 size_t elt_size, int socket, unsigned int flags, size_t align,
466 size_t bound, bool contig)
468 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
469 struct malloc_mp_req req;
472 memset(&req, 0, sizeof(req));
474 req.t = REQ_TYPE_ALLOC;
475 req.alloc_req.align = align;
476 req.alloc_req.bound = bound;
477 req.alloc_req.contig = contig;
478 req.alloc_req.flags = flags;
479 req.alloc_req.elt_size = elt_size;
480 req.alloc_req.page_sz = pg_sz;
481 req.alloc_req.socket = socket;
482 req.alloc_req.malloc_heap_idx = heap - mcfg->malloc_heaps;
484 req_result = request_to_primary(&req);
489 if (req.result != REQ_RESULT_SUCCESS)
496 try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
497 int socket, unsigned int flags, size_t align, size_t bound,
502 rte_mcfg_mem_write_lock();
504 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
505 ret = try_expand_heap_primary(heap, pg_sz, elt_size, socket,
506 flags, align, bound, contig);
508 ret = try_expand_heap_secondary(heap, pg_sz, elt_size, socket,
509 flags, align, bound, contig);
512 rte_mcfg_mem_write_unlock();
517 compare_pagesz(const void *a, const void *b)
519 const struct rte_memseg_list * const*mpa = a;
520 const struct rte_memseg_list * const*mpb = b;
521 const struct rte_memseg_list *msla = *mpa;
522 const struct rte_memseg_list *mslb = *mpb;
523 uint64_t pg_sz_a = msla->page_sz;
524 uint64_t pg_sz_b = mslb->page_sz;
526 if (pg_sz_a < pg_sz_b)
528 if (pg_sz_a > pg_sz_b)
534 alloc_more_mem_on_socket(struct malloc_heap *heap, size_t size, int socket,
535 unsigned int flags, size_t align, size_t bound, bool contig)
537 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
538 struct rte_memseg_list *requested_msls[RTE_MAX_MEMSEG_LISTS];
539 struct rte_memseg_list *other_msls[RTE_MAX_MEMSEG_LISTS];
540 uint64_t requested_pg_sz[RTE_MAX_MEMSEG_LISTS];
541 uint64_t other_pg_sz[RTE_MAX_MEMSEG_LISTS];
543 int i, n_other_msls, n_other_pg_sz, n_requested_msls, n_requested_pg_sz;
544 bool size_hint = (flags & RTE_MEMZONE_SIZE_HINT_ONLY) > 0;
545 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
548 memset(requested_msls, 0, sizeof(requested_msls));
549 memset(other_msls, 0, sizeof(other_msls));
550 memset(requested_pg_sz, 0, sizeof(requested_pg_sz));
551 memset(other_pg_sz, 0, sizeof(other_pg_sz));
554 * go through memseg list and take note of all the page sizes available,
555 * and if any of them were specifically requested by the user.
557 n_requested_msls = 0;
559 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
560 struct rte_memseg_list *msl = &mcfg->memsegs[i];
562 if (msl->socket_id != socket)
565 if (msl->base_va == NULL)
568 /* if pages of specific size were requested */
569 if (size_flags != 0 && check_hugepage_sz(size_flags,
571 requested_msls[n_requested_msls++] = msl;
572 else if (size_flags == 0 || size_hint)
573 other_msls[n_other_msls++] = msl;
576 /* sort the lists, smallest first */
577 qsort(requested_msls, n_requested_msls, sizeof(requested_msls[0]),
579 qsort(other_msls, n_other_msls, sizeof(other_msls[0]),
582 /* now, extract page sizes we are supposed to try */
584 n_requested_pg_sz = 0;
585 for (i = 0; i < n_requested_msls; i++) {
586 uint64_t pg_sz = requested_msls[i]->page_sz;
588 if (prev_pg_sz != pg_sz) {
589 requested_pg_sz[n_requested_pg_sz++] = pg_sz;
595 for (i = 0; i < n_other_msls; i++) {
596 uint64_t pg_sz = other_msls[i]->page_sz;
598 if (prev_pg_sz != pg_sz) {
599 other_pg_sz[n_other_pg_sz++] = pg_sz;
604 /* finally, try allocating memory of specified page sizes, starting from
607 for (i = 0; i < n_requested_pg_sz; i++) {
608 uint64_t pg_sz = requested_pg_sz[i];
611 * do not pass the size hint here, as user expects other page
612 * sizes first, before resorting to best effort allocation.
614 if (!try_expand_heap(heap, pg_sz, size, socket, size_flags,
615 align, bound, contig))
618 if (n_other_pg_sz == 0)
621 /* now, check if we can reserve anything with size hint */
622 ret = find_suitable_element(heap, size, flags, align, bound, contig);
627 * we still couldn't reserve memory, so try expanding heap with other
628 * page sizes, if there are any
630 for (i = 0; i < n_other_pg_sz; i++) {
631 uint64_t pg_sz = other_pg_sz[i];
633 if (!try_expand_heap(heap, pg_sz, size, socket, flags,
634 align, bound, contig))
640 /* this will try lower page sizes first */
642 malloc_heap_alloc_on_heap_id(const char *type, size_t size,
643 unsigned int heap_id, unsigned int flags, size_t align,
644 size_t bound, bool contig)
646 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
647 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
648 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
651 const struct internal_config *internal_conf =
652 eal_get_internal_configuration();
654 rte_spinlock_lock(&(heap->lock));
656 align = align == 0 ? 1 : align;
658 /* for legacy mode, try once and with all flags */
659 if (internal_conf->legacy_mem) {
660 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
665 * we do not pass the size hint here, because even if allocation fails,
666 * we may still be able to allocate memory from appropriate page sizes,
667 * we just need to request more memory first.
670 socket_id = rte_socket_id_by_idx(heap_id);
672 * if socket ID is negative, we cannot find a socket ID for this heap -
673 * which means it's an external heap. those can have unexpected page
674 * sizes, so if the user asked to allocate from there - assume user
675 * knows what they're doing, and allow allocating from there with any
679 size_flags |= RTE_MEMZONE_SIZE_HINT_ONLY;
681 ret = heap_alloc(heap, type, size, size_flags, align, bound, contig);
685 /* if socket ID is invalid, this is an external heap */
689 if (!alloc_more_mem_on_socket(heap, size, socket_id, flags, align,
691 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
693 /* this should have succeeded */
695 RTE_LOG(ERR, EAL, "Error allocating from heap\n");
698 rte_spinlock_unlock(&(heap->lock));
703 malloc_get_numa_socket(void)
705 const struct internal_config *conf = eal_get_internal_configuration();
706 unsigned int socket_id = rte_socket_id();
709 if (socket_id != (unsigned int)SOCKET_ID_ANY)
712 /* for control threads, return first socket where memory is available */
713 for (idx = 0; idx < rte_socket_count(); idx++) {
714 socket_id = rte_socket_id_by_idx(idx);
715 if (conf->socket_mem[socket_id] != 0)
719 return rte_socket_id_by_idx(0);
723 malloc_heap_alloc(const char *type, size_t size, int socket_arg,
724 unsigned int flags, size_t align, size_t bound, bool contig)
726 int socket, heap_id, i;
729 /* return NULL if size is 0 or alignment is not power-of-2 */
730 if (size == 0 || (align && !rte_is_power_of_2(align)))
733 if (!rte_eal_has_hugepages() && socket_arg < RTE_MAX_NUMA_NODES)
734 socket_arg = SOCKET_ID_ANY;
736 if (socket_arg == SOCKET_ID_ANY)
737 socket = malloc_get_numa_socket();
741 /* turn socket ID into heap ID */
742 heap_id = malloc_socket_to_heap_id(socket);
743 /* if heap id is negative, socket ID was invalid */
747 ret = malloc_heap_alloc_on_heap_id(type, size, heap_id, flags, align,
749 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
752 /* try other heaps. we are only iterating through native DPDK sockets,
753 * so external heaps won't be included.
755 for (i = 0; i < (int) rte_socket_count(); i++) {
758 ret = malloc_heap_alloc_on_heap_id(type, size, i, flags, align,
767 heap_alloc_biggest_on_heap_id(const char *type, unsigned int heap_id,
768 unsigned int flags, size_t align, bool contig)
770 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
771 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
774 rte_spinlock_lock(&(heap->lock));
776 align = align == 0 ? 1 : align;
778 ret = heap_alloc_biggest(heap, type, flags, align, contig);
780 rte_spinlock_unlock(&(heap->lock));
786 malloc_heap_alloc_biggest(const char *type, int socket_arg, unsigned int flags,
787 size_t align, bool contig)
789 int socket, i, cur_socket, heap_id;
792 /* return NULL if align is not power-of-2 */
793 if ((align && !rte_is_power_of_2(align)))
796 if (!rte_eal_has_hugepages())
797 socket_arg = SOCKET_ID_ANY;
799 if (socket_arg == SOCKET_ID_ANY)
800 socket = malloc_get_numa_socket();
804 /* turn socket ID into heap ID */
805 heap_id = malloc_socket_to_heap_id(socket);
806 /* if heap id is negative, socket ID was invalid */
810 ret = heap_alloc_biggest_on_heap_id(type, heap_id, flags, align,
812 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
815 /* try other heaps */
816 for (i = 0; i < (int) rte_socket_count(); i++) {
817 cur_socket = rte_socket_id_by_idx(i);
818 if (cur_socket == socket)
820 ret = heap_alloc_biggest_on_heap_id(type, i, flags, align,
828 /* this function is exposed in malloc_mp.h */
830 malloc_heap_free_pages(void *aligned_start, size_t aligned_len)
832 int n_segs, seg_idx, max_seg_idx;
833 struct rte_memseg_list *msl;
836 msl = rte_mem_virt2memseg_list(aligned_start);
840 page_sz = (size_t)msl->page_sz;
841 n_segs = aligned_len / page_sz;
842 seg_idx = RTE_PTR_DIFF(aligned_start, msl->base_va) / page_sz;
843 max_seg_idx = seg_idx + n_segs;
845 for (; seg_idx < max_seg_idx; seg_idx++) {
846 struct rte_memseg *ms;
848 ms = rte_fbarray_get(&msl->memseg_arr, seg_idx);
849 eal_memalloc_free_seg(ms);
855 malloc_heap_free(struct malloc_elem *elem)
857 struct malloc_heap *heap;
858 void *start, *aligned_start, *end, *aligned_end;
859 size_t len, aligned_len, page_sz;
860 struct rte_memseg_list *msl;
861 unsigned int i, n_segs, before_space, after_space;
863 const struct internal_config *internal_conf =
864 eal_get_internal_configuration();
866 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
869 asan_clear_redzone(elem);
871 /* elem may be merged with previous element, so keep heap address */
874 page_sz = (size_t)msl->page_sz;
876 rte_spinlock_lock(&(heap->lock));
878 void *asan_ptr = RTE_PTR_ADD(elem, MALLOC_ELEM_HEADER_LEN + elem->pad);
879 size_t asan_data_len = elem->size - MALLOC_ELEM_OVERHEAD - elem->pad;
881 /* mark element as free */
882 elem->state = ELEM_FREE;
884 elem = malloc_elem_free(elem);
886 /* anything after this is a bonus */
889 /* ...of which we can't avail if we are in legacy mode, or if this is an
890 * externally allocated segment.
892 if (internal_conf->legacy_mem || (msl->external > 0))
895 /* check if we can free any memory back to the system */
896 if (elem->size < page_sz)
899 /* if user requested to match allocations, the sizes must match - if not,
900 * we will defer freeing these hugepages until the entire original allocation
903 if (internal_conf->match_allocations && elem->size != elem->orig_size)
906 /* probably, but let's make sure, as we may not be using up full page */
909 aligned_start = RTE_PTR_ALIGN_CEIL(start, page_sz);
910 end = RTE_PTR_ADD(elem, len);
911 aligned_end = RTE_PTR_ALIGN_FLOOR(end, page_sz);
913 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
915 /* can't free anything */
916 if (aligned_len < page_sz)
919 /* we can free something. however, some of these pages may be marked as
920 * unfreeable, so also check that as well
922 n_segs = aligned_len / page_sz;
923 for (i = 0; i < n_segs; i++) {
924 const struct rte_memseg *tmp =
925 rte_mem_virt2memseg(aligned_start, msl);
927 if (tmp->flags & RTE_MEMSEG_FLAG_DO_NOT_FREE) {
928 /* this is an unfreeable segment, so move start */
929 aligned_start = RTE_PTR_ADD(tmp->addr, tmp->len);
933 /* recalculate length and number of segments */
934 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
935 n_segs = aligned_len / page_sz;
937 /* check if we can still free some pages */
941 /* We're not done yet. We also have to check if by freeing space we will
942 * be leaving free elements that are too small to store new elements.
943 * Check if we have enough space in the beginning and at the end, or if
944 * start/end are exactly page aligned.
946 before_space = RTE_PTR_DIFF(aligned_start, elem);
947 after_space = RTE_PTR_DIFF(end, aligned_end);
948 if (before_space != 0 &&
949 before_space < MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
950 /* There is not enough space before start, but we may be able to
951 * move the start forward by one page.
957 aligned_start = RTE_PTR_ADD(aligned_start, page_sz);
958 aligned_len -= page_sz;
961 if (after_space != 0 && after_space <
962 MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
963 /* There is not enough space after end, but we may be able to
964 * move the end backwards by one page.
970 aligned_end = RTE_PTR_SUB(aligned_end, page_sz);
971 aligned_len -= page_sz;
975 /* now we can finally free us some pages */
977 rte_mcfg_mem_write_lock();
980 * we allow secondary processes to clear the heap of this allocated
981 * memory because it is safe to do so, as even if notifications about
982 * unmapped pages don't make it to other processes, heap is shared
983 * across all processes, and will become empty of this memory anyway,
984 * and nothing can allocate it back unless primary process will be able
985 * to deliver allocation message to every single running process.
988 malloc_elem_free_list_remove(elem);
990 malloc_elem_hide_region(elem, (void *) aligned_start, aligned_len);
992 heap->total_size -= aligned_len;
994 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
995 /* notify user about changes in memory map */
996 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
997 aligned_start, aligned_len);
999 /* don't care if any of this fails */
1000 malloc_heap_free_pages(aligned_start, aligned_len);
1004 struct malloc_mp_req req;
1006 memset(&req, 0, sizeof(req));
1008 req.t = REQ_TYPE_FREE;
1009 req.free_req.addr = aligned_start;
1010 req.free_req.len = aligned_len;
1013 * we request primary to deallocate pages, but we don't do it
1014 * in this thread. instead, we notify primary that we would like
1015 * to deallocate pages, and this process will receive another
1016 * request (in parallel) that will do it for us on another
1019 * we also don't really care if this succeeds - the data is
1020 * already removed from the heap, so it is, for all intents and
1021 * purposes, hidden from the rest of DPDK even if some other
1022 * process (including this one) may have these pages mapped.
1024 * notifications about deallocated memory happen during sync.
1026 request_to_primary(&req);
1029 RTE_LOG(DEBUG, EAL, "Heap on socket %d was shrunk by %zdMB\n",
1030 msl->socket_id, aligned_len >> 20ULL);
1032 rte_mcfg_mem_write_unlock();
1034 asan_set_freezone(asan_ptr, asan_data_len);
1036 rte_spinlock_unlock(&(heap->lock));
1041 malloc_heap_resize(struct malloc_elem *elem, size_t size)
1045 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
1048 rte_spinlock_lock(&(elem->heap->lock));
1050 ret = malloc_elem_resize(elem, size);
1052 rte_spinlock_unlock(&(elem->heap->lock));
1058 * Function to retrieve data for a given heap
1061 malloc_heap_get_stats(struct malloc_heap *heap,
1062 struct rte_malloc_socket_stats *socket_stats)
1065 struct malloc_elem *elem;
1067 rte_spinlock_lock(&heap->lock);
1069 /* Initialise variables for heap */
1070 socket_stats->free_count = 0;
1071 socket_stats->heap_freesz_bytes = 0;
1072 socket_stats->greatest_free_size = 0;
1074 /* Iterate through free list */
1075 for (idx = 0; idx < RTE_HEAP_NUM_FREELISTS; idx++) {
1076 for (elem = LIST_FIRST(&heap->free_head[idx]);
1077 !!elem; elem = LIST_NEXT(elem, free_list))
1079 socket_stats->free_count++;
1080 socket_stats->heap_freesz_bytes += elem->size;
1081 if (elem->size > socket_stats->greatest_free_size)
1082 socket_stats->greatest_free_size = elem->size;
1085 /* Get stats on overall heap and allocated memory on this heap */
1086 socket_stats->heap_totalsz_bytes = heap->total_size;
1087 socket_stats->heap_allocsz_bytes = (socket_stats->heap_totalsz_bytes -
1088 socket_stats->heap_freesz_bytes);
1089 socket_stats->alloc_count = heap->alloc_count;
1091 rte_spinlock_unlock(&heap->lock);
1096 * Function to retrieve data for a given heap
1099 malloc_heap_dump(struct malloc_heap *heap, FILE *f)
1101 struct malloc_elem *elem;
1103 rte_spinlock_lock(&heap->lock);
1105 fprintf(f, "Heap size: 0x%zx\n", heap->total_size);
1106 fprintf(f, "Heap alloc count: %u\n", heap->alloc_count);
1110 malloc_elem_dump(elem, f);
1114 rte_spinlock_unlock(&heap->lock);
1118 destroy_elem(struct malloc_elem *elem, size_t len)
1120 struct malloc_heap *heap = elem->heap;
1122 /* notify all subscribers that a memory area is going to be removed */
1123 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE, elem, len);
1125 /* this element can be removed */
1126 malloc_elem_free_list_remove(elem);
1127 malloc_elem_hide_region(elem, elem, len);
1129 heap->total_size -= len;
1131 memset(elem, 0, sizeof(*elem));
1136 struct rte_memseg_list *
1137 malloc_heap_create_external_seg(void *va_addr, rte_iova_t iova_addrs[],
1138 unsigned int n_pages, size_t page_sz, const char *seg_name,
1139 unsigned int socket_id)
1141 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1142 char fbarray_name[RTE_FBARRAY_NAME_LEN];
1143 struct rte_memseg_list *msl = NULL;
1144 struct rte_fbarray *arr;
1145 size_t seg_len = n_pages * page_sz;
1148 /* first, find a free memseg list */
1149 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1150 struct rte_memseg_list *tmp = &mcfg->memsegs[i];
1151 if (tmp->base_va == NULL) {
1157 RTE_LOG(ERR, EAL, "Couldn't find empty memseg list\n");
1162 snprintf(fbarray_name, sizeof(fbarray_name), "%s_%p",
1165 /* create the backing fbarray */
1166 if (rte_fbarray_init(&msl->memseg_arr, fbarray_name, n_pages,
1167 sizeof(struct rte_memseg)) < 0) {
1168 RTE_LOG(ERR, EAL, "Couldn't create fbarray backing the memseg list\n");
1171 arr = &msl->memseg_arr;
1173 /* fbarray created, fill it up */
1174 for (i = 0; i < n_pages; i++) {
1175 struct rte_memseg *ms;
1177 rte_fbarray_set_used(arr, i);
1178 ms = rte_fbarray_get(arr, i);
1179 ms->addr = RTE_PTR_ADD(va_addr, i * page_sz);
1180 ms->iova = iova_addrs == NULL ? RTE_BAD_IOVA : iova_addrs[i];
1181 ms->hugepage_sz = page_sz;
1183 ms->nchannel = rte_memory_get_nchannel();
1184 ms->nrank = rte_memory_get_nrank();
1185 ms->socket_id = socket_id;
1188 /* set up the memseg list */
1189 msl->base_va = va_addr;
1190 msl->page_sz = page_sz;
1191 msl->socket_id = socket_id;
1199 struct extseg_walk_arg {
1202 struct rte_memseg_list *msl;
1206 extseg_walk(const struct rte_memseg_list *msl, void *arg)
1208 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1209 struct extseg_walk_arg *wa = arg;
1211 if (msl->base_va == wa->va_addr && msl->len == wa->len) {
1212 unsigned int found_idx;
1215 found_idx = msl - mcfg->memsegs;
1216 wa->msl = &mcfg->memsegs[found_idx];
1222 struct rte_memseg_list *
1223 malloc_heap_find_external_seg(void *va_addr, size_t len)
1225 struct extseg_walk_arg wa;
1228 wa.va_addr = va_addr;
1231 res = rte_memseg_list_walk_thread_unsafe(extseg_walk, &wa);
1234 /* 0 means nothing was found, -1 shouldn't happen */
1243 malloc_heap_destroy_external_seg(struct rte_memseg_list *msl)
1245 /* destroy the fbarray backing this memory */
1246 if (rte_fbarray_destroy(&msl->memseg_arr) < 0)
1249 /* reset the memseg list */
1250 memset(msl, 0, sizeof(*msl));
1256 malloc_heap_add_external_memory(struct malloc_heap *heap,
1257 struct rte_memseg_list *msl)
1259 /* erase contents of new memory */
1260 memset(msl->base_va, 0, msl->len);
1262 /* now, add newly minted memory to the malloc heap */
1263 malloc_heap_add_memory(heap, msl, msl->base_va, msl->len);
1265 heap->total_size += msl->len;
1268 RTE_LOG(DEBUG, EAL, "Added segment for heap %s starting at %p\n",
1269 heap->name, msl->base_va);
1271 /* notify all subscribers that a new memory area has been added */
1272 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC,
1273 msl->base_va, msl->len);
1279 malloc_heap_remove_external_memory(struct malloc_heap *heap, void *va_addr,
1282 struct malloc_elem *elem = heap->first;
1284 /* find element with specified va address */
1285 while (elem != NULL && elem != va_addr) {
1287 /* stop if we've blown past our VA */
1288 if (elem > (struct malloc_elem *)va_addr) {
1293 /* check if element was found */
1294 if (elem == NULL || elem->msl->len != len) {
1298 /* if element's size is not equal to segment len, segment is busy */
1299 if (elem->state == ELEM_BUSY || elem->size != len) {
1303 return destroy_elem(elem, len);
1307 malloc_heap_create(struct malloc_heap *heap, const char *heap_name)
1309 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1310 uint32_t next_socket_id = mcfg->next_socket_id;
1312 /* prevent overflow. did you really create 2 billion heaps??? */
1313 if (next_socket_id > INT32_MAX) {
1314 RTE_LOG(ERR, EAL, "Cannot assign new socket ID's\n");
1319 /* initialize empty heap */
1320 heap->alloc_count = 0;
1323 LIST_INIT(heap->free_head);
1324 rte_spinlock_init(&heap->lock);
1325 heap->total_size = 0;
1326 heap->socket_id = next_socket_id;
1328 /* we hold a global mem hotplug writelock, so it's safe to increment */
1329 mcfg->next_socket_id++;
1332 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1337 malloc_heap_destroy(struct malloc_heap *heap)
1339 if (heap->alloc_count != 0) {
1340 RTE_LOG(ERR, EAL, "Heap is still in use\n");
1344 if (heap->first != NULL || heap->last != NULL) {
1345 RTE_LOG(ERR, EAL, "Heap still contains memory segments\n");
1349 if (heap->total_size != 0)
1350 RTE_LOG(ERR, EAL, "Total size not zero, heap is likely corrupt\n");
1352 /* after this, the lock will be dropped */
1353 memset(heap, 0, sizeof(*heap));
1359 rte_eal_malloc_heap_init(void)
1361 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1363 const struct internal_config *internal_conf =
1364 eal_get_internal_configuration();
1366 if (internal_conf->match_allocations)
1367 RTE_LOG(DEBUG, EAL, "Hugepages will be freed exactly as allocated.\n");
1369 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1370 /* assign min socket ID to external heaps */
1371 mcfg->next_socket_id = EXTERNAL_HEAP_MIN_SOCKET_ID;
1373 /* assign names to default DPDK heaps */
1374 for (i = 0; i < rte_socket_count(); i++) {
1375 struct malloc_heap *heap = &mcfg->malloc_heaps[i];
1376 char heap_name[RTE_HEAP_NAME_MAX_LEN];
1377 int socket_id = rte_socket_id_by_idx(i);
1379 snprintf(heap_name, sizeof(heap_name),
1380 "socket_%i", socket_id);
1381 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1382 heap->socket_id = socket_id;
1387 if (register_mp_requests()) {
1388 RTE_LOG(ERR, EAL, "Couldn't register malloc multiprocess actions\n");
1389 rte_mcfg_mem_read_unlock();
1393 /* unlock mem hotplug here. it's safe for primary as no requests can
1394 * even come before primary itself is fully initialized, and secondaries
1395 * do not need to initialize the heap.
1397 rte_mcfg_mem_read_unlock();
1399 /* secondary process does not need to initialize anything */
1400 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1403 /* add all IOVA-contiguous areas to the heap */
1404 return rte_memseg_contig_walk(malloc_add_seg, NULL);