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_atomic.h>
24 #include <rte_fbarray.h>
26 #include "eal_internal_cfg.h"
27 #include "eal_memalloc.h"
28 #include "malloc_elem.h"
29 #include "malloc_heap.h"
30 #include "malloc_mp.h"
32 /* start external socket ID's at a very high number */
33 #define CONST_MAX(a, b) (a > b ? a : b) /* RTE_MAX is not a constant */
34 #define EXTERNAL_HEAP_MIN_SOCKET_ID (CONST_MAX((1 << 8), RTE_MAX_NUMA_NODES))
37 check_hugepage_sz(unsigned flags, uint64_t hugepage_sz)
39 unsigned check_flag = 0;
41 if (!(flags & ~RTE_MEMZONE_SIZE_HINT_ONLY))
44 switch (hugepage_sz) {
46 check_flag = RTE_MEMZONE_256KB;
49 check_flag = RTE_MEMZONE_2MB;
52 check_flag = RTE_MEMZONE_16MB;
55 check_flag = RTE_MEMZONE_256MB;
58 check_flag = RTE_MEMZONE_512MB;
61 check_flag = RTE_MEMZONE_1GB;
64 check_flag = RTE_MEMZONE_4GB;
67 check_flag = RTE_MEMZONE_16GB;
70 return check_flag & flags;
74 malloc_socket_to_heap_id(unsigned int socket_id)
76 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
79 for (i = 0; i < RTE_MAX_HEAPS; i++) {
80 struct malloc_heap *heap = &mcfg->malloc_heaps[i];
82 if (heap->socket_id == socket_id)
89 * Expand the heap with a memory area.
91 static struct malloc_elem *
92 malloc_heap_add_memory(struct malloc_heap *heap, struct rte_memseg_list *msl,
93 void *start, size_t len)
95 struct malloc_elem *elem = start;
97 malloc_elem_init(elem, heap, msl, len);
99 malloc_elem_insert(elem);
101 elem = malloc_elem_join_adjacent_free(elem);
103 malloc_elem_free_list_insert(elem);
109 malloc_add_seg(const struct rte_memseg_list *msl,
110 const struct rte_memseg *ms, size_t len, void *arg __rte_unused)
112 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
113 struct rte_memseg_list *found_msl;
114 struct malloc_heap *heap;
115 int msl_idx, heap_idx;
120 heap_idx = malloc_socket_to_heap_id(msl->socket_id);
122 RTE_LOG(ERR, EAL, "Memseg list has invalid socket id\n");
125 heap = &mcfg->malloc_heaps[heap_idx];
127 /* msl is const, so find it */
128 msl_idx = msl - mcfg->memsegs;
130 if (msl_idx < 0 || msl_idx >= RTE_MAX_MEMSEG_LISTS)
133 found_msl = &mcfg->memsegs[msl_idx];
135 malloc_heap_add_memory(heap, found_msl, ms->addr, len);
137 heap->total_size += len;
139 RTE_LOG(DEBUG, EAL, "Added %zuM to heap on socket %i\n", len >> 20,
145 * Iterates through the freelist for a heap to find a free element
146 * which can store data of the required size and with the requested alignment.
147 * If size is 0, find the biggest available elem.
148 * Returns null on failure, or pointer to element on success.
150 static struct malloc_elem *
151 find_suitable_element(struct malloc_heap *heap, size_t size,
152 unsigned int flags, size_t align, size_t bound, bool contig)
155 struct malloc_elem *elem, *alt_elem = NULL;
157 for (idx = malloc_elem_free_list_index(size);
158 idx < RTE_HEAP_NUM_FREELISTS; idx++) {
159 for (elem = LIST_FIRST(&heap->free_head[idx]);
160 !!elem; elem = LIST_NEXT(elem, free_list)) {
161 if (malloc_elem_can_hold(elem, size, align, bound,
163 if (check_hugepage_sz(flags,
166 if (alt_elem == NULL)
172 if ((alt_elem != NULL) && (flags & RTE_MEMZONE_SIZE_HINT_ONLY))
179 * Iterates through the freelist for a heap to find a free element with the
180 * biggest size and requested alignment. Will also set size to whatever element
181 * size that was found.
182 * Returns null on failure, or pointer to element on success.
184 static struct malloc_elem *
185 find_biggest_element(struct malloc_heap *heap, size_t *size,
186 unsigned int flags, size_t align, bool contig)
188 struct malloc_elem *elem, *max_elem = NULL;
189 size_t idx, max_size = 0;
191 for (idx = 0; idx < RTE_HEAP_NUM_FREELISTS; idx++) {
192 for (elem = LIST_FIRST(&heap->free_head[idx]);
193 !!elem; elem = LIST_NEXT(elem, free_list)) {
195 if ((flags & RTE_MEMZONE_SIZE_HINT_ONLY) == 0 &&
196 !check_hugepage_sz(flags,
201 malloc_elem_find_max_iova_contig(elem,
204 void *data_start = RTE_PTR_ADD(elem,
205 MALLOC_ELEM_HEADER_LEN);
206 void *data_end = RTE_PTR_ADD(elem, elem->size -
207 MALLOC_ELEM_TRAILER_LEN);
208 void *aligned = RTE_PTR_ALIGN_CEIL(data_start,
210 /* check if aligned data start is beyond end */
211 if (aligned >= data_end)
213 cur_size = RTE_PTR_DIFF(data_end, aligned);
215 if (cur_size > max_size) {
227 * Main function to allocate a block of memory from the heap.
228 * It locks the free list, scans it, and adds a new memseg if the
229 * scan fails. Once the new memseg is added, it re-scans and should return
230 * the new element after releasing the lock.
233 heap_alloc(struct malloc_heap *heap, const char *type __rte_unused, size_t size,
234 unsigned int flags, size_t align, size_t bound, bool contig)
236 struct malloc_elem *elem;
238 size = RTE_CACHE_LINE_ROUNDUP(size);
239 align = RTE_CACHE_LINE_ROUNDUP(align);
241 elem = find_suitable_element(heap, size, flags, align, bound, contig);
243 elem = malloc_elem_alloc(elem, size, align, bound, contig);
245 /* increase heap's count of allocated elements */
249 return elem == NULL ? NULL : (void *)(&elem[1]);
253 heap_alloc_biggest(struct malloc_heap *heap, const char *type __rte_unused,
254 unsigned int flags, size_t align, bool contig)
256 struct malloc_elem *elem;
259 align = RTE_CACHE_LINE_ROUNDUP(align);
261 elem = find_biggest_element(heap, &size, flags, align, contig);
263 elem = malloc_elem_alloc(elem, size, align, 0, contig);
265 /* increase heap's count of allocated elements */
269 return elem == NULL ? NULL : (void *)(&elem[1]);
272 /* this function is exposed in malloc_mp.h */
274 rollback_expand_heap(struct rte_memseg **ms, int n_segs,
275 struct malloc_elem *elem, void *map_addr, size_t map_len)
278 malloc_elem_free_list_remove(elem);
279 malloc_elem_hide_region(elem, map_addr, map_len);
282 eal_memalloc_free_seg_bulk(ms, n_segs);
285 /* this function is exposed in malloc_mp.h */
287 alloc_pages_on_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
288 int socket, unsigned int flags, size_t align, size_t bound,
289 bool contig, struct rte_memseg **ms, int n_segs)
291 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
292 struct rte_memseg_list *msl;
293 struct malloc_elem *elem = NULL;
296 void *ret, *map_addr;
298 alloc_sz = (size_t)pg_sz * n_segs;
300 /* first, check if we're allowed to allocate this memory */
301 if (eal_memalloc_mem_alloc_validate(socket,
302 heap->total_size + alloc_sz) < 0) {
303 RTE_LOG(DEBUG, EAL, "User has disallowed allocation\n");
307 allocd_pages = eal_memalloc_alloc_seg_bulk(ms, n_segs, pg_sz,
310 /* make sure we've allocated our pages... */
311 if (allocd_pages < 0)
314 map_addr = ms[0]->addr;
315 msl = rte_mem_virt2memseg_list(map_addr);
317 /* check if we wanted contiguous memory but didn't get it */
318 if (contig && !eal_memalloc_is_contig(msl, map_addr, alloc_sz)) {
319 RTE_LOG(DEBUG, EAL, "%s(): couldn't allocate physically contiguous space\n",
324 if (mcfg->dma_maskbits) {
325 if (rte_eal_check_dma_mask(mcfg->dma_maskbits)) {
327 "%s(): couldn't allocate memory due to DMA mask\n",
333 /* add newly minted memsegs to malloc heap */
334 elem = malloc_heap_add_memory(heap, msl, map_addr, alloc_sz);
336 /* try once more, as now we have allocated new memory */
337 ret = find_suitable_element(heap, elt_size, flags, align, bound,
346 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
351 try_expand_heap_primary(struct malloc_heap *heap, uint64_t pg_sz,
352 size_t elt_size, int socket, unsigned int flags, size_t align,
353 size_t bound, bool contig)
355 struct malloc_elem *elem;
356 struct rte_memseg **ms;
360 bool callback_triggered = false;
362 alloc_sz = RTE_ALIGN_CEIL(align + elt_size +
363 MALLOC_ELEM_TRAILER_LEN, pg_sz);
364 n_segs = alloc_sz / pg_sz;
366 /* we can't know in advance how many pages we'll need, so we malloc */
367 ms = malloc(sizeof(*ms) * n_segs);
370 memset(ms, 0, sizeof(*ms) * n_segs);
372 elem = alloc_pages_on_heap(heap, pg_sz, elt_size, socket, flags, align,
373 bound, contig, ms, n_segs);
378 map_addr = ms[0]->addr;
380 /* notify user about changes in memory map */
381 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC, map_addr, alloc_sz);
383 /* notify other processes that this has happened */
384 if (request_sync()) {
385 /* we couldn't ensure all processes have mapped memory,
386 * so free it back and notify everyone that it's been
389 * technically, we could've avoided adding memory addresses to
390 * the map, but that would've led to inconsistent behavior
391 * between primary and secondary processes, as those get
392 * callbacks during sync. therefore, force primary process to
393 * do alloc-and-rollback syncs as well.
395 callback_triggered = true;
398 heap->total_size += alloc_sz;
400 RTE_LOG(DEBUG, EAL, "Heap on socket %d was expanded by %zdMB\n",
401 socket, alloc_sz >> 20ULL);
408 if (callback_triggered)
409 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
412 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
422 try_expand_heap_secondary(struct malloc_heap *heap, uint64_t pg_sz,
423 size_t elt_size, int socket, unsigned int flags, size_t align,
424 size_t bound, bool contig)
426 struct malloc_mp_req req;
429 memset(&req, 0, sizeof(req));
431 req.t = REQ_TYPE_ALLOC;
432 req.alloc_req.align = align;
433 req.alloc_req.bound = bound;
434 req.alloc_req.contig = contig;
435 req.alloc_req.flags = flags;
436 req.alloc_req.elt_size = elt_size;
437 req.alloc_req.page_sz = pg_sz;
438 req.alloc_req.socket = socket;
439 req.alloc_req.heap = heap; /* it's in shared memory */
441 req_result = request_to_primary(&req);
446 if (req.result != REQ_RESULT_SUCCESS)
453 try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
454 int socket, unsigned int flags, size_t align, size_t bound,
457 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
460 rte_rwlock_write_lock(&mcfg->memory_hotplug_lock);
462 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
463 ret = try_expand_heap_primary(heap, pg_sz, elt_size, socket,
464 flags, align, bound, contig);
466 ret = try_expand_heap_secondary(heap, pg_sz, elt_size, socket,
467 flags, align, bound, contig);
470 rte_rwlock_write_unlock(&mcfg->memory_hotplug_lock);
475 compare_pagesz(const void *a, const void *b)
477 const struct rte_memseg_list * const*mpa = a;
478 const struct rte_memseg_list * const*mpb = b;
479 const struct rte_memseg_list *msla = *mpa;
480 const struct rte_memseg_list *mslb = *mpb;
481 uint64_t pg_sz_a = msla->page_sz;
482 uint64_t pg_sz_b = mslb->page_sz;
484 if (pg_sz_a < pg_sz_b)
486 if (pg_sz_a > pg_sz_b)
492 alloc_more_mem_on_socket(struct malloc_heap *heap, size_t size, int socket,
493 unsigned int flags, size_t align, size_t bound, bool contig)
495 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
496 struct rte_memseg_list *requested_msls[RTE_MAX_MEMSEG_LISTS];
497 struct rte_memseg_list *other_msls[RTE_MAX_MEMSEG_LISTS];
498 uint64_t requested_pg_sz[RTE_MAX_MEMSEG_LISTS];
499 uint64_t other_pg_sz[RTE_MAX_MEMSEG_LISTS];
501 int i, n_other_msls, n_other_pg_sz, n_requested_msls, n_requested_pg_sz;
502 bool size_hint = (flags & RTE_MEMZONE_SIZE_HINT_ONLY) > 0;
503 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
506 memset(requested_msls, 0, sizeof(requested_msls));
507 memset(other_msls, 0, sizeof(other_msls));
508 memset(requested_pg_sz, 0, sizeof(requested_pg_sz));
509 memset(other_pg_sz, 0, sizeof(other_pg_sz));
512 * go through memseg list and take note of all the page sizes available,
513 * and if any of them were specifically requested by the user.
515 n_requested_msls = 0;
517 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
518 struct rte_memseg_list *msl = &mcfg->memsegs[i];
520 if (msl->socket_id != socket)
523 if (msl->base_va == NULL)
526 /* if pages of specific size were requested */
527 if (size_flags != 0 && check_hugepage_sz(size_flags,
529 requested_msls[n_requested_msls++] = msl;
530 else if (size_flags == 0 || size_hint)
531 other_msls[n_other_msls++] = msl;
534 /* sort the lists, smallest first */
535 qsort(requested_msls, n_requested_msls, sizeof(requested_msls[0]),
537 qsort(other_msls, n_other_msls, sizeof(other_msls[0]),
540 /* now, extract page sizes we are supposed to try */
542 n_requested_pg_sz = 0;
543 for (i = 0; i < n_requested_msls; i++) {
544 uint64_t pg_sz = requested_msls[i]->page_sz;
546 if (prev_pg_sz != pg_sz) {
547 requested_pg_sz[n_requested_pg_sz++] = pg_sz;
553 for (i = 0; i < n_other_msls; i++) {
554 uint64_t pg_sz = other_msls[i]->page_sz;
556 if (prev_pg_sz != pg_sz) {
557 other_pg_sz[n_other_pg_sz++] = pg_sz;
562 /* finally, try allocating memory of specified page sizes, starting from
565 for (i = 0; i < n_requested_pg_sz; i++) {
566 uint64_t pg_sz = requested_pg_sz[i];
569 * do not pass the size hint here, as user expects other page
570 * sizes first, before resorting to best effort allocation.
572 if (!try_expand_heap(heap, pg_sz, size, socket, size_flags,
573 align, bound, contig))
576 if (n_other_pg_sz == 0)
579 /* now, check if we can reserve anything with size hint */
580 ret = find_suitable_element(heap, size, flags, align, bound, contig);
585 * we still couldn't reserve memory, so try expanding heap with other
586 * page sizes, if there are any
588 for (i = 0; i < n_other_pg_sz; i++) {
589 uint64_t pg_sz = other_pg_sz[i];
591 if (!try_expand_heap(heap, pg_sz, size, socket, flags,
592 align, bound, contig))
598 /* this will try lower page sizes first */
600 malloc_heap_alloc_on_heap_id(const char *type, size_t size,
601 unsigned int heap_id, unsigned int flags, size_t align,
602 size_t bound, bool contig)
604 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
605 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
606 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
610 rte_spinlock_lock(&(heap->lock));
612 align = align == 0 ? 1 : align;
614 /* for legacy mode, try once and with all flags */
615 if (internal_config.legacy_mem) {
616 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
621 * we do not pass the size hint here, because even if allocation fails,
622 * we may still be able to allocate memory from appropriate page sizes,
623 * we just need to request more memory first.
626 socket_id = rte_socket_id_by_idx(heap_id);
628 * if socket ID is negative, we cannot find a socket ID for this heap -
629 * which means it's an external heap. those can have unexpected page
630 * sizes, so if the user asked to allocate from there - assume user
631 * knows what they're doing, and allow allocating from there with any
635 size_flags |= RTE_MEMZONE_SIZE_HINT_ONLY;
637 ret = heap_alloc(heap, type, size, size_flags, align, bound, contig);
641 /* if socket ID is invalid, this is an external heap */
645 if (!alloc_more_mem_on_socket(heap, size, socket_id, flags, align,
647 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
649 /* this should have succeeded */
651 RTE_LOG(ERR, EAL, "Error allocating from heap\n");
654 rte_spinlock_unlock(&(heap->lock));
659 malloc_heap_alloc(const char *type, size_t size, int socket_arg,
660 unsigned int flags, size_t align, size_t bound, bool contig)
662 int socket, heap_id, i;
665 /* return NULL if size is 0 or alignment is not power-of-2 */
666 if (size == 0 || (align && !rte_is_power_of_2(align)))
669 if (!rte_eal_has_hugepages() && socket_arg < RTE_MAX_NUMA_NODES)
670 socket_arg = SOCKET_ID_ANY;
672 if (socket_arg == SOCKET_ID_ANY)
673 socket = malloc_get_numa_socket();
677 /* turn socket ID into heap ID */
678 heap_id = malloc_socket_to_heap_id(socket);
679 /* if heap id is negative, socket ID was invalid */
683 ret = malloc_heap_alloc_on_heap_id(type, size, heap_id, flags, align,
685 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
688 /* try other heaps. we are only iterating through native DPDK sockets,
689 * so external heaps won't be included.
691 for (i = 0; i < (int) rte_socket_count(); i++) {
694 ret = malloc_heap_alloc_on_heap_id(type, size, i, flags, align,
703 heap_alloc_biggest_on_heap_id(const char *type, unsigned int heap_id,
704 unsigned int flags, size_t align, bool contig)
706 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
707 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
710 rte_spinlock_lock(&(heap->lock));
712 align = align == 0 ? 1 : align;
714 ret = heap_alloc_biggest(heap, type, flags, align, contig);
716 rte_spinlock_unlock(&(heap->lock));
722 malloc_heap_alloc_biggest(const char *type, int socket_arg, unsigned int flags,
723 size_t align, bool contig)
725 int socket, i, cur_socket, heap_id;
728 /* return NULL if align is not power-of-2 */
729 if ((align && !rte_is_power_of_2(align)))
732 if (!rte_eal_has_hugepages())
733 socket_arg = SOCKET_ID_ANY;
735 if (socket_arg == SOCKET_ID_ANY)
736 socket = malloc_get_numa_socket();
740 /* turn socket ID into heap ID */
741 heap_id = malloc_socket_to_heap_id(socket);
742 /* if heap id is negative, socket ID was invalid */
746 ret = heap_alloc_biggest_on_heap_id(type, heap_id, flags, align,
748 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
751 /* try other heaps */
752 for (i = 0; i < (int) rte_socket_count(); i++) {
753 cur_socket = rte_socket_id_by_idx(i);
754 if (cur_socket == socket)
756 ret = heap_alloc_biggest_on_heap_id(type, i, flags, align,
764 /* this function is exposed in malloc_mp.h */
766 malloc_heap_free_pages(void *aligned_start, size_t aligned_len)
768 int n_segs, seg_idx, max_seg_idx;
769 struct rte_memseg_list *msl;
772 msl = rte_mem_virt2memseg_list(aligned_start);
776 page_sz = (size_t)msl->page_sz;
777 n_segs = aligned_len / page_sz;
778 seg_idx = RTE_PTR_DIFF(aligned_start, msl->base_va) / page_sz;
779 max_seg_idx = seg_idx + n_segs;
781 for (; seg_idx < max_seg_idx; seg_idx++) {
782 struct rte_memseg *ms;
784 ms = rte_fbarray_get(&msl->memseg_arr, seg_idx);
785 eal_memalloc_free_seg(ms);
791 malloc_heap_free(struct malloc_elem *elem)
793 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
794 struct malloc_heap *heap;
795 void *start, *aligned_start, *end, *aligned_end;
796 size_t len, aligned_len, page_sz;
797 struct rte_memseg_list *msl;
798 unsigned int i, n_segs, before_space, after_space;
801 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
804 /* elem may be merged with previous element, so keep heap address */
807 page_sz = (size_t)msl->page_sz;
809 rte_spinlock_lock(&(heap->lock));
811 /* mark element as free */
812 elem->state = ELEM_FREE;
814 elem = malloc_elem_free(elem);
816 /* anything after this is a bonus */
819 /* ...of which we can't avail if we are in legacy mode, or if this is an
820 * externally allocated segment.
822 if (internal_config.legacy_mem || (msl->external > 0))
825 /* check if we can free any memory back to the system */
826 if (elem->size < page_sz)
829 /* probably, but let's make sure, as we may not be using up full page */
832 aligned_start = RTE_PTR_ALIGN_CEIL(start, page_sz);
833 end = RTE_PTR_ADD(elem, len);
834 aligned_end = RTE_PTR_ALIGN_FLOOR(end, page_sz);
836 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
838 /* can't free anything */
839 if (aligned_len < page_sz)
842 /* we can free something. however, some of these pages may be marked as
843 * unfreeable, so also check that as well
845 n_segs = aligned_len / page_sz;
846 for (i = 0; i < n_segs; i++) {
847 const struct rte_memseg *tmp =
848 rte_mem_virt2memseg(aligned_start, msl);
850 if (tmp->flags & RTE_MEMSEG_FLAG_DO_NOT_FREE) {
851 /* this is an unfreeable segment, so move start */
852 aligned_start = RTE_PTR_ADD(tmp->addr, tmp->len);
856 /* recalculate length and number of segments */
857 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
858 n_segs = aligned_len / page_sz;
860 /* check if we can still free some pages */
864 /* We're not done yet. We also have to check if by freeing space we will
865 * be leaving free elements that are too small to store new elements.
866 * Check if we have enough space in the beginning and at the end, or if
867 * start/end are exactly page aligned.
869 before_space = RTE_PTR_DIFF(aligned_start, elem);
870 after_space = RTE_PTR_DIFF(end, aligned_end);
871 if (before_space != 0 &&
872 before_space < MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
873 /* There is not enough space before start, but we may be able to
874 * move the start forward by one page.
880 aligned_start = RTE_PTR_ADD(aligned_start, page_sz);
881 aligned_len -= page_sz;
884 if (after_space != 0 && after_space <
885 MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
886 /* There is not enough space after end, but we may be able to
887 * move the end backwards by one page.
893 aligned_end = RTE_PTR_SUB(aligned_end, page_sz);
894 aligned_len -= page_sz;
898 /* now we can finally free us some pages */
900 rte_rwlock_write_lock(&mcfg->memory_hotplug_lock);
903 * we allow secondary processes to clear the heap of this allocated
904 * memory because it is safe to do so, as even if notifications about
905 * unmapped pages don't make it to other processes, heap is shared
906 * across all processes, and will become empty of this memory anyway,
907 * and nothing can allocate it back unless primary process will be able
908 * to deliver allocation message to every single running process.
911 malloc_elem_free_list_remove(elem);
913 malloc_elem_hide_region(elem, (void *) aligned_start, aligned_len);
915 heap->total_size -= aligned_len;
917 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
918 /* notify user about changes in memory map */
919 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
920 aligned_start, aligned_len);
922 /* don't care if any of this fails */
923 malloc_heap_free_pages(aligned_start, aligned_len);
927 struct malloc_mp_req req;
929 memset(&req, 0, sizeof(req));
931 req.t = REQ_TYPE_FREE;
932 req.free_req.addr = aligned_start;
933 req.free_req.len = aligned_len;
936 * we request primary to deallocate pages, but we don't do it
937 * in this thread. instead, we notify primary that we would like
938 * to deallocate pages, and this process will receive another
939 * request (in parallel) that will do it for us on another
942 * we also don't really care if this succeeds - the data is
943 * already removed from the heap, so it is, for all intents and
944 * purposes, hidden from the rest of DPDK even if some other
945 * process (including this one) may have these pages mapped.
947 * notifications about deallocated memory happen during sync.
949 request_to_primary(&req);
952 RTE_LOG(DEBUG, EAL, "Heap on socket %d was shrunk by %zdMB\n",
953 msl->socket_id, aligned_len >> 20ULL);
955 rte_rwlock_write_unlock(&mcfg->memory_hotplug_lock);
957 rte_spinlock_unlock(&(heap->lock));
962 malloc_heap_resize(struct malloc_elem *elem, size_t size)
966 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
969 rte_spinlock_lock(&(elem->heap->lock));
971 ret = malloc_elem_resize(elem, size);
973 rte_spinlock_unlock(&(elem->heap->lock));
979 * Function to retrieve data for a given heap
982 malloc_heap_get_stats(struct malloc_heap *heap,
983 struct rte_malloc_socket_stats *socket_stats)
986 struct malloc_elem *elem;
988 rte_spinlock_lock(&heap->lock);
990 /* Initialise variables for heap */
991 socket_stats->free_count = 0;
992 socket_stats->heap_freesz_bytes = 0;
993 socket_stats->greatest_free_size = 0;
995 /* Iterate through free list */
996 for (idx = 0; idx < RTE_HEAP_NUM_FREELISTS; idx++) {
997 for (elem = LIST_FIRST(&heap->free_head[idx]);
998 !!elem; elem = LIST_NEXT(elem, free_list))
1000 socket_stats->free_count++;
1001 socket_stats->heap_freesz_bytes += elem->size;
1002 if (elem->size > socket_stats->greatest_free_size)
1003 socket_stats->greatest_free_size = elem->size;
1006 /* Get stats on overall heap and allocated memory on this heap */
1007 socket_stats->heap_totalsz_bytes = heap->total_size;
1008 socket_stats->heap_allocsz_bytes = (socket_stats->heap_totalsz_bytes -
1009 socket_stats->heap_freesz_bytes);
1010 socket_stats->alloc_count = heap->alloc_count;
1012 rte_spinlock_unlock(&heap->lock);
1017 * Function to retrieve data for a given heap
1020 malloc_heap_dump(struct malloc_heap *heap, FILE *f)
1022 struct malloc_elem *elem;
1024 rte_spinlock_lock(&heap->lock);
1026 fprintf(f, "Heap size: 0x%zx\n", heap->total_size);
1027 fprintf(f, "Heap alloc count: %u\n", heap->alloc_count);
1031 malloc_elem_dump(elem, f);
1035 rte_spinlock_unlock(&heap->lock);
1039 destroy_seg(struct malloc_elem *elem, size_t len)
1041 struct malloc_heap *heap = elem->heap;
1042 struct rte_memseg_list *msl;
1046 /* notify all subscribers that a memory area is going to be removed */
1047 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE, elem, len);
1049 /* this element can be removed */
1050 malloc_elem_free_list_remove(elem);
1051 malloc_elem_hide_region(elem, elem, len);
1053 heap->total_size -= len;
1055 memset(elem, 0, sizeof(*elem));
1057 /* destroy the fbarray backing this memory */
1058 if (rte_fbarray_destroy(&msl->memseg_arr) < 0)
1061 /* reset the memseg list */
1062 memset(msl, 0, sizeof(*msl));
1068 malloc_heap_add_external_memory(struct malloc_heap *heap, void *va_addr,
1069 rte_iova_t iova_addrs[], unsigned int n_pages, size_t page_sz)
1071 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1072 char fbarray_name[RTE_FBARRAY_NAME_LEN];
1073 struct rte_memseg_list *msl = NULL;
1074 struct rte_fbarray *arr;
1075 size_t seg_len = n_pages * page_sz;
1078 /* first, find a free memseg list */
1079 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1080 struct rte_memseg_list *tmp = &mcfg->memsegs[i];
1081 if (tmp->base_va == NULL) {
1087 RTE_LOG(ERR, EAL, "Couldn't find empty memseg list\n");
1092 snprintf(fbarray_name, sizeof(fbarray_name) - 1, "%s_%p",
1093 heap->name, va_addr);
1095 /* create the backing fbarray */
1096 if (rte_fbarray_init(&msl->memseg_arr, fbarray_name, n_pages,
1097 sizeof(struct rte_memseg)) < 0) {
1098 RTE_LOG(ERR, EAL, "Couldn't create fbarray backing the memseg list\n");
1101 arr = &msl->memseg_arr;
1103 /* fbarray created, fill it up */
1104 for (i = 0; i < n_pages; i++) {
1105 struct rte_memseg *ms;
1107 rte_fbarray_set_used(arr, i);
1108 ms = rte_fbarray_get(arr, i);
1109 ms->addr = RTE_PTR_ADD(va_addr, i * page_sz);
1110 ms->iova = iova_addrs == NULL ? RTE_BAD_IOVA : iova_addrs[i];
1111 ms->hugepage_sz = page_sz;
1113 ms->nchannel = rte_memory_get_nchannel();
1114 ms->nrank = rte_memory_get_nrank();
1115 ms->socket_id = heap->socket_id;
1118 /* set up the memseg list */
1119 msl->base_va = va_addr;
1120 msl->page_sz = page_sz;
1121 msl->socket_id = heap->socket_id;
1126 /* erase contents of new memory */
1127 memset(va_addr, 0, seg_len);
1129 /* now, add newly minted memory to the malloc heap */
1130 malloc_heap_add_memory(heap, msl, va_addr, seg_len);
1132 heap->total_size += seg_len;
1135 RTE_LOG(DEBUG, EAL, "Added segment for heap %s starting at %p\n",
1136 heap->name, va_addr);
1138 /* notify all subscribers that a new memory area has been added */
1139 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC,
1146 malloc_heap_remove_external_memory(struct malloc_heap *heap, void *va_addr,
1149 struct malloc_elem *elem = heap->first;
1151 /* find element with specified va address */
1152 while (elem != NULL && elem != va_addr) {
1154 /* stop if we've blown past our VA */
1155 if (elem > (struct malloc_elem *)va_addr) {
1160 /* check if element was found */
1161 if (elem == NULL || elem->msl->len != len) {
1165 /* if element's size is not equal to segment len, segment is busy */
1166 if (elem->state == ELEM_BUSY || elem->size != len) {
1170 return destroy_seg(elem, len);
1174 malloc_heap_create(struct malloc_heap *heap, const char *heap_name)
1176 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1177 uint32_t next_socket_id = mcfg->next_socket_id;
1179 /* prevent overflow. did you really create 2 billion heaps??? */
1180 if (next_socket_id > INT32_MAX) {
1181 RTE_LOG(ERR, EAL, "Cannot assign new socket ID's\n");
1186 /* initialize empty heap */
1187 heap->alloc_count = 0;
1190 LIST_INIT(heap->free_head);
1191 rte_spinlock_init(&heap->lock);
1192 heap->total_size = 0;
1193 heap->socket_id = next_socket_id;
1195 /* we hold a global mem hotplug writelock, so it's safe to increment */
1196 mcfg->next_socket_id++;
1199 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1204 malloc_heap_destroy(struct malloc_heap *heap)
1206 if (heap->alloc_count != 0) {
1207 RTE_LOG(ERR, EAL, "Heap is still in use\n");
1211 if (heap->first != NULL || heap->last != NULL) {
1212 RTE_LOG(ERR, EAL, "Heap still contains memory segments\n");
1216 if (heap->total_size != 0)
1217 RTE_LOG(ERR, EAL, "Total size not zero, heap is likely corrupt\n");
1219 /* after this, the lock will be dropped */
1220 memset(heap, 0, sizeof(*heap));
1226 rte_eal_malloc_heap_init(void)
1228 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1231 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1232 /* assign min socket ID to external heaps */
1233 mcfg->next_socket_id = EXTERNAL_HEAP_MIN_SOCKET_ID;
1235 /* assign names to default DPDK heaps */
1236 for (i = 0; i < rte_socket_count(); i++) {
1237 struct malloc_heap *heap = &mcfg->malloc_heaps[i];
1238 char heap_name[RTE_HEAP_NAME_MAX_LEN];
1239 int socket_id = rte_socket_id_by_idx(i);
1241 snprintf(heap_name, sizeof(heap_name) - 1,
1242 "socket_%i", socket_id);
1243 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1244 heap->socket_id = socket_id;
1249 if (register_mp_requests()) {
1250 RTE_LOG(ERR, EAL, "Couldn't register malloc multiprocess actions\n");
1251 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
1255 /* unlock mem hotplug here. it's safe for primary as no requests can
1256 * even come before primary itself is fully initialized, and secondaries
1257 * do not need to initialize the heap.
1259 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
1261 /* secondary process does not need to initialize anything */
1262 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1265 /* add all IOVA-contiguous areas to the heap */
1266 return rte_memseg_contig_walk(malloc_add_seg, NULL);