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, bool dirty)
98 struct malloc_elem *elem = start;
100 malloc_elem_init(elem, heap, msl, len, elem, len, dirty);
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,
139 ms->flags & RTE_MEMSEG_FLAG_DIRTY);
141 heap->total_size += len;
143 RTE_LOG(DEBUG, EAL, "Added %zuM to heap on socket %i\n", len >> 20,
149 * Iterates through the freelist for a heap to find a free element
150 * which can store data of the required size and with the requested alignment.
151 * If size is 0, find the biggest available elem.
152 * Returns null on failure, or pointer to element on success.
154 static struct malloc_elem *
155 find_suitable_element(struct malloc_heap *heap, size_t size,
156 unsigned int flags, size_t align, size_t bound, bool contig)
159 struct malloc_elem *elem, *alt_elem = NULL;
161 for (idx = malloc_elem_free_list_index(size);
162 idx < RTE_HEAP_NUM_FREELISTS; idx++) {
163 for (elem = LIST_FIRST(&heap->free_head[idx]);
164 !!elem; elem = LIST_NEXT(elem, free_list)) {
165 if (malloc_elem_can_hold(elem, size, align, bound,
167 if (check_hugepage_sz(flags,
170 if (alt_elem == NULL)
176 if ((alt_elem != NULL) && (flags & RTE_MEMZONE_SIZE_HINT_ONLY))
183 * Iterates through the freelist for a heap to find a free element with the
184 * biggest size and requested alignment. Will also set size to whatever element
185 * size that was found.
186 * Returns null on failure, or pointer to element on success.
188 static struct malloc_elem *
189 find_biggest_element(struct malloc_heap *heap, size_t *size,
190 unsigned int flags, size_t align, bool contig)
192 struct malloc_elem *elem, *max_elem = NULL;
193 size_t idx, max_size = 0;
195 for (idx = 0; idx < RTE_HEAP_NUM_FREELISTS; idx++) {
196 for (elem = LIST_FIRST(&heap->free_head[idx]);
197 !!elem; elem = LIST_NEXT(elem, free_list)) {
199 if ((flags & RTE_MEMZONE_SIZE_HINT_ONLY) == 0 &&
200 !check_hugepage_sz(flags,
205 malloc_elem_find_max_iova_contig(elem,
208 void *data_start = RTE_PTR_ADD(elem,
209 MALLOC_ELEM_HEADER_LEN);
210 void *data_end = RTE_PTR_ADD(elem, elem->size -
211 MALLOC_ELEM_TRAILER_LEN);
212 void *aligned = RTE_PTR_ALIGN_CEIL(data_start,
214 /* check if aligned data start is beyond end */
215 if (aligned >= data_end)
217 cur_size = RTE_PTR_DIFF(data_end, aligned);
219 if (cur_size > max_size) {
231 * Main function to allocate a block of memory from the heap.
232 * It locks the free list, scans it, and adds a new memseg if the
233 * scan fails. Once the new memseg is added, it re-scans and should return
234 * the new element after releasing the lock.
237 heap_alloc(struct malloc_heap *heap, const char *type __rte_unused, size_t size,
238 unsigned int flags, size_t align, size_t bound, bool contig)
240 struct malloc_elem *elem;
241 size_t user_size = size;
243 size = RTE_CACHE_LINE_ROUNDUP(size);
244 align = RTE_CACHE_LINE_ROUNDUP(align);
246 /* roundup might cause an overflow */
249 elem = find_suitable_element(heap, size, flags, align, bound, contig);
251 elem = malloc_elem_alloc(elem, size, align, bound, contig);
253 /* increase heap's count of allocated elements */
256 asan_set_redzone(elem, user_size);
259 return elem == NULL ? NULL : (void *)(&elem[1]);
263 heap_alloc_biggest(struct malloc_heap *heap, const char *type __rte_unused,
264 unsigned int flags, size_t align, bool contig)
266 struct malloc_elem *elem;
269 align = RTE_CACHE_LINE_ROUNDUP(align);
271 elem = find_biggest_element(heap, &size, flags, align, contig);
273 elem = malloc_elem_alloc(elem, size, align, 0, contig);
275 /* increase heap's count of allocated elements */
278 asan_set_redzone(elem, size);
281 return elem == NULL ? NULL : (void *)(&elem[1]);
284 /* this function is exposed in malloc_mp.h */
286 rollback_expand_heap(struct rte_memseg **ms, int n_segs,
287 struct malloc_elem *elem, void *map_addr, size_t map_len)
290 malloc_elem_free_list_remove(elem);
291 malloc_elem_hide_region(elem, map_addr, map_len);
294 eal_memalloc_free_seg_bulk(ms, n_segs);
297 /* this function is exposed in malloc_mp.h */
299 alloc_pages_on_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
300 int socket, unsigned int flags, size_t align, size_t bound,
301 bool contig, struct rte_memseg **ms, int n_segs)
303 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
304 struct rte_memseg_list *msl;
305 struct malloc_elem *elem = NULL;
309 void *ret, *map_addr;
311 alloc_sz = (size_t)pg_sz * n_segs;
313 /* first, check if we're allowed to allocate this memory */
314 if (eal_memalloc_mem_alloc_validate(socket,
315 heap->total_size + alloc_sz) < 0) {
316 RTE_LOG(DEBUG, EAL, "User has disallowed allocation\n");
320 allocd_pages = eal_memalloc_alloc_seg_bulk(ms, n_segs, pg_sz,
323 /* make sure we've allocated our pages... */
324 if (allocd_pages < 0)
327 map_addr = ms[0]->addr;
328 msl = rte_mem_virt2memseg_list(map_addr);
330 /* check if we wanted contiguous memory but didn't get it */
331 if (contig && !eal_memalloc_is_contig(msl, map_addr, alloc_sz)) {
332 RTE_LOG(DEBUG, EAL, "%s(): couldn't allocate physically contiguous space\n",
338 * Once we have all the memseg lists configured, if there is a dma mask
339 * set, check iova addresses are not out of range. Otherwise the device
340 * setting the dma mask could have problems with the mapped memory.
342 * There are two situations when this can happen:
343 * 1) memory initialization
344 * 2) dynamic memory allocation
346 * For 1), an error when checking dma mask implies app can not be
347 * executed. For 2) implies the new memory can not be added.
349 if (mcfg->dma_maskbits &&
350 rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
352 * Currently this can only happen if IOMMU is enabled
353 * and the address width supported by the IOMMU hw is
354 * not enough for using the memory mapped IOVAs.
356 * If IOVA is VA, advice to try with '--iova-mode pa'
357 * which could solve some situations when IOVA VA is not
361 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask\n",
365 * If IOVA is VA and it is possible to run with IOVA PA,
366 * because user is root, give and advice for solving the
369 if ((rte_eal_iova_mode() == RTE_IOVA_VA) &&
370 rte_eal_using_phys_addrs())
372 "%s(): Please try initializing EAL with --iova-mode=pa parameter\n",
377 /* Element is dirty if it contains at least one dirty page. */
378 for (i = 0; i < allocd_pages; i++)
379 dirty |= ms[i]->flags & RTE_MEMSEG_FLAG_DIRTY;
381 /* add newly minted memsegs to malloc heap */
382 elem = malloc_heap_add_memory(heap, msl, map_addr, alloc_sz, dirty);
384 /* try once more, as now we have allocated new memory */
385 ret = find_suitable_element(heap, elt_size, flags, align, bound,
394 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
399 try_expand_heap_primary(struct malloc_heap *heap, uint64_t pg_sz,
400 size_t elt_size, int socket, unsigned int flags, size_t align,
401 size_t bound, bool contig)
403 struct malloc_elem *elem;
404 struct rte_memseg **ms;
408 bool callback_triggered = false;
410 alloc_sz = RTE_ALIGN_CEIL(align + elt_size +
411 MALLOC_ELEM_TRAILER_LEN, pg_sz);
412 n_segs = alloc_sz / pg_sz;
414 /* we can't know in advance how many pages we'll need, so we malloc */
415 ms = malloc(sizeof(*ms) * n_segs);
418 memset(ms, 0, sizeof(*ms) * n_segs);
420 elem = alloc_pages_on_heap(heap, pg_sz, elt_size, socket, flags, align,
421 bound, contig, ms, n_segs);
426 map_addr = ms[0]->addr;
428 /* notify user about changes in memory map */
429 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC, map_addr, alloc_sz);
431 /* notify other processes that this has happened */
432 if (request_sync()) {
433 /* we couldn't ensure all processes have mapped memory,
434 * so free it back and notify everyone that it's been
437 * technically, we could've avoided adding memory addresses to
438 * the map, but that would've led to inconsistent behavior
439 * between primary and secondary processes, as those get
440 * callbacks during sync. therefore, force primary process to
441 * do alloc-and-rollback syncs as well.
443 callback_triggered = true;
446 heap->total_size += alloc_sz;
448 RTE_LOG(DEBUG, EAL, "Heap on socket %d was expanded by %zdMB\n",
449 socket, alloc_sz >> 20ULL);
456 if (callback_triggered)
457 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
460 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
470 try_expand_heap_secondary(struct malloc_heap *heap, uint64_t pg_sz,
471 size_t elt_size, int socket, unsigned int flags, size_t align,
472 size_t bound, bool contig)
474 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
475 struct malloc_mp_req req;
478 memset(&req, 0, sizeof(req));
480 req.t = REQ_TYPE_ALLOC;
481 req.alloc_req.align = align;
482 req.alloc_req.bound = bound;
483 req.alloc_req.contig = contig;
484 req.alloc_req.flags = flags;
485 req.alloc_req.elt_size = elt_size;
486 req.alloc_req.page_sz = pg_sz;
487 req.alloc_req.socket = socket;
488 req.alloc_req.malloc_heap_idx = heap - mcfg->malloc_heaps;
490 req_result = request_to_primary(&req);
495 if (req.result != REQ_RESULT_SUCCESS)
502 try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
503 int socket, unsigned int flags, size_t align, size_t bound,
508 rte_mcfg_mem_write_lock();
510 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
511 ret = try_expand_heap_primary(heap, pg_sz, elt_size, socket,
512 flags, align, bound, contig);
514 ret = try_expand_heap_secondary(heap, pg_sz, elt_size, socket,
515 flags, align, bound, contig);
518 rte_mcfg_mem_write_unlock();
523 compare_pagesz(const void *a, const void *b)
525 const struct rte_memseg_list * const*mpa = a;
526 const struct rte_memseg_list * const*mpb = b;
527 const struct rte_memseg_list *msla = *mpa;
528 const struct rte_memseg_list *mslb = *mpb;
529 uint64_t pg_sz_a = msla->page_sz;
530 uint64_t pg_sz_b = mslb->page_sz;
532 if (pg_sz_a < pg_sz_b)
534 if (pg_sz_a > pg_sz_b)
540 alloc_more_mem_on_socket(struct malloc_heap *heap, size_t size, int socket,
541 unsigned int flags, size_t align, size_t bound, bool contig)
543 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
544 struct rte_memseg_list *requested_msls[RTE_MAX_MEMSEG_LISTS];
545 struct rte_memseg_list *other_msls[RTE_MAX_MEMSEG_LISTS];
546 uint64_t requested_pg_sz[RTE_MAX_MEMSEG_LISTS];
547 uint64_t other_pg_sz[RTE_MAX_MEMSEG_LISTS];
549 int i, n_other_msls, n_other_pg_sz, n_requested_msls, n_requested_pg_sz;
550 bool size_hint = (flags & RTE_MEMZONE_SIZE_HINT_ONLY) > 0;
551 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
554 memset(requested_msls, 0, sizeof(requested_msls));
555 memset(other_msls, 0, sizeof(other_msls));
556 memset(requested_pg_sz, 0, sizeof(requested_pg_sz));
557 memset(other_pg_sz, 0, sizeof(other_pg_sz));
560 * go through memseg list and take note of all the page sizes available,
561 * and if any of them were specifically requested by the user.
563 n_requested_msls = 0;
565 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
566 struct rte_memseg_list *msl = &mcfg->memsegs[i];
568 if (msl->socket_id != socket)
571 if (msl->base_va == NULL)
574 /* if pages of specific size were requested */
575 if (size_flags != 0 && check_hugepage_sz(size_flags,
577 requested_msls[n_requested_msls++] = msl;
578 else if (size_flags == 0 || size_hint)
579 other_msls[n_other_msls++] = msl;
582 /* sort the lists, smallest first */
583 qsort(requested_msls, n_requested_msls, sizeof(requested_msls[0]),
585 qsort(other_msls, n_other_msls, sizeof(other_msls[0]),
588 /* now, extract page sizes we are supposed to try */
590 n_requested_pg_sz = 0;
591 for (i = 0; i < n_requested_msls; i++) {
592 uint64_t pg_sz = requested_msls[i]->page_sz;
594 if (prev_pg_sz != pg_sz) {
595 requested_pg_sz[n_requested_pg_sz++] = pg_sz;
601 for (i = 0; i < n_other_msls; i++) {
602 uint64_t pg_sz = other_msls[i]->page_sz;
604 if (prev_pg_sz != pg_sz) {
605 other_pg_sz[n_other_pg_sz++] = pg_sz;
610 /* finally, try allocating memory of specified page sizes, starting from
613 for (i = 0; i < n_requested_pg_sz; i++) {
614 uint64_t pg_sz = requested_pg_sz[i];
617 * do not pass the size hint here, as user expects other page
618 * sizes first, before resorting to best effort allocation.
620 if (!try_expand_heap(heap, pg_sz, size, socket, size_flags,
621 align, bound, contig))
624 if (n_other_pg_sz == 0)
627 /* now, check if we can reserve anything with size hint */
628 ret = find_suitable_element(heap, size, flags, align, bound, contig);
633 * we still couldn't reserve memory, so try expanding heap with other
634 * page sizes, if there are any
636 for (i = 0; i < n_other_pg_sz; i++) {
637 uint64_t pg_sz = other_pg_sz[i];
639 if (!try_expand_heap(heap, pg_sz, size, socket, flags,
640 align, bound, contig))
646 /* this will try lower page sizes first */
648 malloc_heap_alloc_on_heap_id(const char *type, size_t size,
649 unsigned int heap_id, unsigned int flags, size_t align,
650 size_t bound, bool contig)
652 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
653 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
654 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
657 const struct internal_config *internal_conf =
658 eal_get_internal_configuration();
660 rte_spinlock_lock(&(heap->lock));
662 align = align == 0 ? 1 : align;
664 /* for legacy mode, try once and with all flags */
665 if (internal_conf->legacy_mem) {
666 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
671 * we do not pass the size hint here, because even if allocation fails,
672 * we may still be able to allocate memory from appropriate page sizes,
673 * we just need to request more memory first.
676 socket_id = rte_socket_id_by_idx(heap_id);
678 * if socket ID is negative, we cannot find a socket ID for this heap -
679 * which means it's an external heap. those can have unexpected page
680 * sizes, so if the user asked to allocate from there - assume user
681 * knows what they're doing, and allow allocating from there with any
685 size_flags |= RTE_MEMZONE_SIZE_HINT_ONLY;
687 ret = heap_alloc(heap, type, size, size_flags, align, bound, contig);
691 /* if socket ID is invalid, this is an external heap */
695 if (!alloc_more_mem_on_socket(heap, size, socket_id, flags, align,
697 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
699 /* this should have succeeded */
701 RTE_LOG(ERR, EAL, "Error allocating from heap\n");
704 rte_spinlock_unlock(&(heap->lock));
709 malloc_get_numa_socket(void)
711 const struct internal_config *conf = eal_get_internal_configuration();
712 unsigned int socket_id = rte_socket_id();
715 if (socket_id != (unsigned int)SOCKET_ID_ANY)
718 /* for control threads, return first socket where memory is available */
719 for (idx = 0; idx < rte_socket_count(); idx++) {
720 socket_id = rte_socket_id_by_idx(idx);
721 if (conf->socket_mem[socket_id] != 0)
725 return rte_socket_id_by_idx(0);
729 malloc_heap_alloc(const char *type, size_t size, int socket_arg,
730 unsigned int flags, size_t align, size_t bound, bool contig)
732 int socket, heap_id, i;
735 /* return NULL if size is 0 or alignment is not power-of-2 */
736 if (size == 0 || (align && !rte_is_power_of_2(align)))
739 if (!rte_eal_has_hugepages() && socket_arg < RTE_MAX_NUMA_NODES)
740 socket_arg = SOCKET_ID_ANY;
742 if (socket_arg == SOCKET_ID_ANY)
743 socket = malloc_get_numa_socket();
747 /* turn socket ID into heap ID */
748 heap_id = malloc_socket_to_heap_id(socket);
749 /* if heap id is negative, socket ID was invalid */
753 ret = malloc_heap_alloc_on_heap_id(type, size, heap_id, flags, align,
755 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
758 /* try other heaps. we are only iterating through native DPDK sockets,
759 * so external heaps won't be included.
761 for (i = 0; i < (int) rte_socket_count(); i++) {
764 ret = malloc_heap_alloc_on_heap_id(type, size, i, flags, align,
773 heap_alloc_biggest_on_heap_id(const char *type, unsigned int heap_id,
774 unsigned int flags, size_t align, bool contig)
776 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
777 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
780 rte_spinlock_lock(&(heap->lock));
782 align = align == 0 ? 1 : align;
784 ret = heap_alloc_biggest(heap, type, flags, align, contig);
786 rte_spinlock_unlock(&(heap->lock));
792 malloc_heap_alloc_biggest(const char *type, int socket_arg, unsigned int flags,
793 size_t align, bool contig)
795 int socket, i, cur_socket, heap_id;
798 /* return NULL if align is not power-of-2 */
799 if ((align && !rte_is_power_of_2(align)))
802 if (!rte_eal_has_hugepages())
803 socket_arg = SOCKET_ID_ANY;
805 if (socket_arg == SOCKET_ID_ANY)
806 socket = malloc_get_numa_socket();
810 /* turn socket ID into heap ID */
811 heap_id = malloc_socket_to_heap_id(socket);
812 /* if heap id is negative, socket ID was invalid */
816 ret = heap_alloc_biggest_on_heap_id(type, heap_id, flags, align,
818 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
821 /* try other heaps */
822 for (i = 0; i < (int) rte_socket_count(); i++) {
823 cur_socket = rte_socket_id_by_idx(i);
824 if (cur_socket == socket)
826 ret = heap_alloc_biggest_on_heap_id(type, i, flags, align,
834 /* this function is exposed in malloc_mp.h */
836 malloc_heap_free_pages(void *aligned_start, size_t aligned_len)
838 int n_segs, seg_idx, max_seg_idx;
839 struct rte_memseg_list *msl;
842 msl = rte_mem_virt2memseg_list(aligned_start);
846 page_sz = (size_t)msl->page_sz;
847 n_segs = aligned_len / page_sz;
848 seg_idx = RTE_PTR_DIFF(aligned_start, msl->base_va) / page_sz;
849 max_seg_idx = seg_idx + n_segs;
851 for (; seg_idx < max_seg_idx; seg_idx++) {
852 struct rte_memseg *ms;
854 ms = rte_fbarray_get(&msl->memseg_arr, seg_idx);
855 eal_memalloc_free_seg(ms);
861 malloc_heap_free(struct malloc_elem *elem)
863 struct malloc_heap *heap;
864 void *start, *aligned_start, *end, *aligned_end;
865 size_t len, aligned_len, page_sz;
866 struct rte_memseg_list *msl;
867 unsigned int i, n_segs, before_space, after_space;
869 const struct internal_config *internal_conf =
870 eal_get_internal_configuration();
872 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
875 asan_clear_redzone(elem);
877 /* elem may be merged with previous element, so keep heap address */
880 page_sz = (size_t)msl->page_sz;
882 rte_spinlock_lock(&(heap->lock));
884 void *asan_ptr = RTE_PTR_ADD(elem, MALLOC_ELEM_HEADER_LEN + elem->pad);
885 size_t asan_data_len = elem->size - MALLOC_ELEM_OVERHEAD - elem->pad;
887 /* mark element as free */
888 elem->state = ELEM_FREE;
890 elem = malloc_elem_free(elem);
892 /* anything after this is a bonus */
895 /* ...of which we can't avail if we are in legacy mode, or if this is an
896 * externally allocated segment.
898 if (internal_conf->legacy_mem || (msl->external > 0))
901 /* check if we can free any memory back to the system */
902 if (elem->size < page_sz)
905 /* if user requested to match allocations, the sizes must match - if not,
906 * we will defer freeing these hugepages until the entire original allocation
909 if (internal_conf->match_allocations && elem->size != elem->orig_size)
912 /* probably, but let's make sure, as we may not be using up full page */
915 aligned_start = RTE_PTR_ALIGN_CEIL(start, page_sz);
916 end = RTE_PTR_ADD(elem, len);
917 aligned_end = RTE_PTR_ALIGN_FLOOR(end, page_sz);
919 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
921 /* can't free anything */
922 if (aligned_len < page_sz)
925 /* we can free something. however, some of these pages may be marked as
926 * unfreeable, so also check that as well
928 n_segs = aligned_len / page_sz;
929 for (i = 0; i < n_segs; i++) {
930 const struct rte_memseg *tmp =
931 rte_mem_virt2memseg(aligned_start, msl);
933 if (tmp->flags & RTE_MEMSEG_FLAG_DO_NOT_FREE) {
934 /* this is an unfreeable segment, so move start */
935 aligned_start = RTE_PTR_ADD(tmp->addr, tmp->len);
939 /* recalculate length and number of segments */
940 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
941 n_segs = aligned_len / page_sz;
943 /* check if we can still free some pages */
947 /* We're not done yet. We also have to check if by freeing space we will
948 * be leaving free elements that are too small to store new elements.
949 * Check if we have enough space in the beginning and at the end, or if
950 * start/end are exactly page aligned.
952 before_space = RTE_PTR_DIFF(aligned_start, elem);
953 after_space = RTE_PTR_DIFF(end, aligned_end);
954 if (before_space != 0 &&
955 before_space < MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
956 /* There is not enough space before start, but we may be able to
957 * move the start forward by one page.
963 aligned_start = RTE_PTR_ADD(aligned_start, page_sz);
964 aligned_len -= page_sz;
967 if (after_space != 0 && after_space <
968 MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
969 /* There is not enough space after end, but we may be able to
970 * move the end backwards by one page.
976 aligned_end = RTE_PTR_SUB(aligned_end, page_sz);
977 aligned_len -= page_sz;
981 /* now we can finally free us some pages */
983 rte_mcfg_mem_write_lock();
986 * we allow secondary processes to clear the heap of this allocated
987 * memory because it is safe to do so, as even if notifications about
988 * unmapped pages don't make it to other processes, heap is shared
989 * across all processes, and will become empty of this memory anyway,
990 * and nothing can allocate it back unless primary process will be able
991 * to deliver allocation message to every single running process.
994 malloc_elem_free_list_remove(elem);
996 malloc_elem_hide_region(elem, (void *) aligned_start, aligned_len);
998 heap->total_size -= aligned_len;
1000 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1001 /* notify user about changes in memory map */
1002 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
1003 aligned_start, aligned_len);
1005 /* don't care if any of this fails */
1006 malloc_heap_free_pages(aligned_start, aligned_len);
1010 struct malloc_mp_req req;
1012 memset(&req, 0, sizeof(req));
1014 req.t = REQ_TYPE_FREE;
1015 req.free_req.addr = aligned_start;
1016 req.free_req.len = aligned_len;
1019 * we request primary to deallocate pages, but we don't do it
1020 * in this thread. instead, we notify primary that we would like
1021 * to deallocate pages, and this process will receive another
1022 * request (in parallel) that will do it for us on another
1025 * we also don't really care if this succeeds - the data is
1026 * already removed from the heap, so it is, for all intents and
1027 * purposes, hidden from the rest of DPDK even if some other
1028 * process (including this one) may have these pages mapped.
1030 * notifications about deallocated memory happen during sync.
1032 request_to_primary(&req);
1035 RTE_LOG(DEBUG, EAL, "Heap on socket %d was shrunk by %zdMB\n",
1036 msl->socket_id, aligned_len >> 20ULL);
1038 rte_mcfg_mem_write_unlock();
1040 asan_set_freezone(asan_ptr, asan_data_len);
1042 rte_spinlock_unlock(&(heap->lock));
1047 malloc_heap_resize(struct malloc_elem *elem, size_t size)
1051 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
1054 rte_spinlock_lock(&(elem->heap->lock));
1056 ret = malloc_elem_resize(elem, size);
1058 rte_spinlock_unlock(&(elem->heap->lock));
1064 * Function to retrieve data for a given heap
1067 malloc_heap_get_stats(struct malloc_heap *heap,
1068 struct rte_malloc_socket_stats *socket_stats)
1071 struct malloc_elem *elem;
1073 rte_spinlock_lock(&heap->lock);
1075 /* Initialise variables for heap */
1076 socket_stats->free_count = 0;
1077 socket_stats->heap_freesz_bytes = 0;
1078 socket_stats->greatest_free_size = 0;
1080 /* Iterate through free list */
1081 for (idx = 0; idx < RTE_HEAP_NUM_FREELISTS; idx++) {
1082 for (elem = LIST_FIRST(&heap->free_head[idx]);
1083 !!elem; elem = LIST_NEXT(elem, free_list))
1085 socket_stats->free_count++;
1086 socket_stats->heap_freesz_bytes += elem->size;
1087 if (elem->size > socket_stats->greatest_free_size)
1088 socket_stats->greatest_free_size = elem->size;
1091 /* Get stats on overall heap and allocated memory on this heap */
1092 socket_stats->heap_totalsz_bytes = heap->total_size;
1093 socket_stats->heap_allocsz_bytes = (socket_stats->heap_totalsz_bytes -
1094 socket_stats->heap_freesz_bytes);
1095 socket_stats->alloc_count = heap->alloc_count;
1097 rte_spinlock_unlock(&heap->lock);
1102 * Function to retrieve data for a given heap
1105 malloc_heap_dump(struct malloc_heap *heap, FILE *f)
1107 struct malloc_elem *elem;
1109 rte_spinlock_lock(&heap->lock);
1111 fprintf(f, "Heap size: 0x%zx\n", heap->total_size);
1112 fprintf(f, "Heap alloc count: %u\n", heap->alloc_count);
1116 malloc_elem_dump(elem, f);
1120 rte_spinlock_unlock(&heap->lock);
1124 destroy_elem(struct malloc_elem *elem, size_t len)
1126 struct malloc_heap *heap = elem->heap;
1128 /* notify all subscribers that a memory area is going to be removed */
1129 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE, elem, len);
1131 /* this element can be removed */
1132 malloc_elem_free_list_remove(elem);
1133 malloc_elem_hide_region(elem, elem, len);
1135 heap->total_size -= len;
1137 memset(elem, 0, sizeof(*elem));
1142 struct rte_memseg_list *
1143 malloc_heap_create_external_seg(void *va_addr, rte_iova_t iova_addrs[],
1144 unsigned int n_pages, size_t page_sz, const char *seg_name,
1145 unsigned int socket_id)
1147 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1148 char fbarray_name[RTE_FBARRAY_NAME_LEN];
1149 struct rte_memseg_list *msl = NULL;
1150 struct rte_fbarray *arr;
1151 size_t seg_len = n_pages * page_sz;
1154 /* first, find a free memseg list */
1155 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1156 struct rte_memseg_list *tmp = &mcfg->memsegs[i];
1157 if (tmp->base_va == NULL) {
1163 RTE_LOG(ERR, EAL, "Couldn't find empty memseg list\n");
1168 snprintf(fbarray_name, sizeof(fbarray_name), "%s_%p",
1171 /* create the backing fbarray */
1172 if (rte_fbarray_init(&msl->memseg_arr, fbarray_name, n_pages,
1173 sizeof(struct rte_memseg)) < 0) {
1174 RTE_LOG(ERR, EAL, "Couldn't create fbarray backing the memseg list\n");
1177 arr = &msl->memseg_arr;
1179 /* fbarray created, fill it up */
1180 for (i = 0; i < n_pages; i++) {
1181 struct rte_memseg *ms;
1183 rte_fbarray_set_used(arr, i);
1184 ms = rte_fbarray_get(arr, i);
1185 ms->addr = RTE_PTR_ADD(va_addr, i * page_sz);
1186 ms->iova = iova_addrs == NULL ? RTE_BAD_IOVA : iova_addrs[i];
1187 ms->hugepage_sz = page_sz;
1189 ms->nchannel = rte_memory_get_nchannel();
1190 ms->nrank = rte_memory_get_nrank();
1191 ms->socket_id = socket_id;
1194 /* set up the memseg list */
1195 msl->base_va = va_addr;
1196 msl->page_sz = page_sz;
1197 msl->socket_id = socket_id;
1205 struct extseg_walk_arg {
1208 struct rte_memseg_list *msl;
1212 extseg_walk(const struct rte_memseg_list *msl, void *arg)
1214 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1215 struct extseg_walk_arg *wa = arg;
1217 if (msl->base_va == wa->va_addr && msl->len == wa->len) {
1218 unsigned int found_idx;
1221 found_idx = msl - mcfg->memsegs;
1222 wa->msl = &mcfg->memsegs[found_idx];
1228 struct rte_memseg_list *
1229 malloc_heap_find_external_seg(void *va_addr, size_t len)
1231 struct extseg_walk_arg wa;
1234 wa.va_addr = va_addr;
1237 res = rte_memseg_list_walk_thread_unsafe(extseg_walk, &wa);
1240 /* 0 means nothing was found, -1 shouldn't happen */
1249 malloc_heap_destroy_external_seg(struct rte_memseg_list *msl)
1251 /* destroy the fbarray backing this memory */
1252 if (rte_fbarray_destroy(&msl->memseg_arr) < 0)
1255 /* reset the memseg list */
1256 memset(msl, 0, sizeof(*msl));
1262 malloc_heap_add_external_memory(struct malloc_heap *heap,
1263 struct rte_memseg_list *msl)
1265 /* erase contents of new memory */
1266 memset(msl->base_va, 0, msl->len);
1268 /* now, add newly minted memory to the malloc heap */
1269 malloc_heap_add_memory(heap, msl, msl->base_va, msl->len, false);
1271 heap->total_size += msl->len;
1274 RTE_LOG(DEBUG, EAL, "Added segment for heap %s starting at %p\n",
1275 heap->name, msl->base_va);
1277 /* notify all subscribers that a new memory area has been added */
1278 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC,
1279 msl->base_va, msl->len);
1285 malloc_heap_remove_external_memory(struct malloc_heap *heap, void *va_addr,
1288 struct malloc_elem *elem = heap->first;
1290 /* find element with specified va address */
1291 while (elem != NULL && elem != va_addr) {
1293 /* stop if we've blown past our VA */
1294 if (elem > (struct malloc_elem *)va_addr) {
1299 /* check if element was found */
1300 if (elem == NULL || elem->msl->len != len) {
1304 /* if element's size is not equal to segment len, segment is busy */
1305 if (elem->state == ELEM_BUSY || elem->size != len) {
1309 return destroy_elem(elem, len);
1313 malloc_heap_create(struct malloc_heap *heap, const char *heap_name)
1315 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1316 uint32_t next_socket_id = mcfg->next_socket_id;
1318 /* prevent overflow. did you really create 2 billion heaps??? */
1319 if (next_socket_id > INT32_MAX) {
1320 RTE_LOG(ERR, EAL, "Cannot assign new socket ID's\n");
1325 /* initialize empty heap */
1326 heap->alloc_count = 0;
1329 LIST_INIT(heap->free_head);
1330 rte_spinlock_init(&heap->lock);
1331 heap->total_size = 0;
1332 heap->socket_id = next_socket_id;
1334 /* we hold a global mem hotplug writelock, so it's safe to increment */
1335 mcfg->next_socket_id++;
1338 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1343 malloc_heap_destroy(struct malloc_heap *heap)
1345 if (heap->alloc_count != 0) {
1346 RTE_LOG(ERR, EAL, "Heap is still in use\n");
1350 if (heap->first != NULL || heap->last != NULL) {
1351 RTE_LOG(ERR, EAL, "Heap still contains memory segments\n");
1355 if (heap->total_size != 0)
1356 RTE_LOG(ERR, EAL, "Total size not zero, heap is likely corrupt\n");
1358 /* after this, the lock will be dropped */
1359 memset(heap, 0, sizeof(*heap));
1365 rte_eal_malloc_heap_init(void)
1367 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1369 const struct internal_config *internal_conf =
1370 eal_get_internal_configuration();
1372 if (internal_conf->match_allocations)
1373 RTE_LOG(DEBUG, EAL, "Hugepages will be freed exactly as allocated.\n");
1375 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1376 /* assign min socket ID to external heaps */
1377 mcfg->next_socket_id = EXTERNAL_HEAP_MIN_SOCKET_ID;
1379 /* assign names to default DPDK heaps */
1380 for (i = 0; i < rte_socket_count(); i++) {
1381 struct malloc_heap *heap = &mcfg->malloc_heaps[i];
1382 char heap_name[RTE_HEAP_NAME_MAX_LEN];
1383 int socket_id = rte_socket_id_by_idx(i);
1385 snprintf(heap_name, sizeof(heap_name),
1386 "socket_%i", socket_id);
1387 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1388 heap->socket_id = socket_id;
1393 if (register_mp_requests()) {
1394 RTE_LOG(ERR, EAL, "Couldn't register malloc multiprocess actions\n");
1395 rte_mcfg_mem_read_unlock();
1399 /* unlock mem hotplug here. it's safe for primary as no requests can
1400 * even come before primary itself is fully initialized, and secondaries
1401 * do not need to initialize the heap.
1403 rte_mcfg_mem_read_unlock();
1405 /* secondary process does not need to initialize anything */
1406 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1409 /* add all IOVA-contiguous areas to the heap */
1410 return rte_memseg_contig_walk(malloc_add_seg, NULL);