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 (!check_hugepage_sz(flags, elem->msl->page_sz))
199 malloc_elem_find_max_iova_contig(elem,
202 void *data_start = RTE_PTR_ADD(elem,
203 MALLOC_ELEM_HEADER_LEN);
204 void *data_end = RTE_PTR_ADD(elem, elem->size -
205 MALLOC_ELEM_TRAILER_LEN);
206 void *aligned = RTE_PTR_ALIGN_CEIL(data_start,
208 /* check if aligned data start is beyond end */
209 if (aligned >= data_end)
211 cur_size = RTE_PTR_DIFF(data_end, aligned);
213 if (cur_size > max_size) {
225 * Main function to allocate a block of memory from the heap.
226 * It locks the free list, scans it, and adds a new memseg if the
227 * scan fails. Once the new memseg is added, it re-scans and should return
228 * the new element after releasing the lock.
231 heap_alloc(struct malloc_heap *heap, const char *type __rte_unused, size_t size,
232 unsigned int flags, size_t align, size_t bound, bool contig)
234 struct malloc_elem *elem;
236 size = RTE_CACHE_LINE_ROUNDUP(size);
237 align = RTE_CACHE_LINE_ROUNDUP(align);
239 elem = find_suitable_element(heap, size, flags, align, bound, contig);
241 elem = malloc_elem_alloc(elem, size, align, bound, contig);
243 /* increase heap's count of allocated elements */
247 return elem == NULL ? NULL : (void *)(&elem[1]);
251 heap_alloc_biggest(struct malloc_heap *heap, const char *type __rte_unused,
252 unsigned int flags, size_t align, bool contig)
254 struct malloc_elem *elem;
257 align = RTE_CACHE_LINE_ROUNDUP(align);
259 elem = find_biggest_element(heap, &size, flags, align, contig);
261 elem = malloc_elem_alloc(elem, size, align, 0, contig);
263 /* increase heap's count of allocated elements */
267 return elem == NULL ? NULL : (void *)(&elem[1]);
270 /* this function is exposed in malloc_mp.h */
272 rollback_expand_heap(struct rte_memseg **ms, int n_segs,
273 struct malloc_elem *elem, void *map_addr, size_t map_len)
276 malloc_elem_free_list_remove(elem);
277 malloc_elem_hide_region(elem, map_addr, map_len);
280 eal_memalloc_free_seg_bulk(ms, n_segs);
283 /* this function is exposed in malloc_mp.h */
285 alloc_pages_on_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
286 int socket, unsigned int flags, size_t align, size_t bound,
287 bool contig, struct rte_memseg **ms, int n_segs)
289 struct rte_memseg_list *msl;
290 struct malloc_elem *elem = NULL;
293 void *ret, *map_addr;
295 alloc_sz = (size_t)pg_sz * n_segs;
297 /* first, check if we're allowed to allocate this memory */
298 if (eal_memalloc_mem_alloc_validate(socket,
299 heap->total_size + alloc_sz) < 0) {
300 RTE_LOG(DEBUG, EAL, "User has disallowed allocation\n");
304 allocd_pages = eal_memalloc_alloc_seg_bulk(ms, n_segs, pg_sz,
307 /* make sure we've allocated our pages... */
308 if (allocd_pages < 0)
311 map_addr = ms[0]->addr;
312 msl = rte_mem_virt2memseg_list(map_addr);
314 /* check if we wanted contiguous memory but didn't get it */
315 if (contig && !eal_memalloc_is_contig(msl, map_addr, alloc_sz)) {
316 RTE_LOG(DEBUG, EAL, "%s(): couldn't allocate physically contiguous space\n",
321 /* add newly minted memsegs to malloc heap */
322 elem = malloc_heap_add_memory(heap, msl, map_addr, alloc_sz);
324 /* try once more, as now we have allocated new memory */
325 ret = find_suitable_element(heap, elt_size, flags, align, bound,
334 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
339 try_expand_heap_primary(struct malloc_heap *heap, uint64_t pg_sz,
340 size_t elt_size, int socket, unsigned int flags, size_t align,
341 size_t bound, bool contig)
343 struct malloc_elem *elem;
344 struct rte_memseg **ms;
348 bool callback_triggered = false;
350 alloc_sz = RTE_ALIGN_CEIL(align + elt_size +
351 MALLOC_ELEM_TRAILER_LEN, pg_sz);
352 n_segs = alloc_sz / pg_sz;
354 /* we can't know in advance how many pages we'll need, so we malloc */
355 ms = malloc(sizeof(*ms) * n_segs);
358 memset(ms, 0, sizeof(*ms) * n_segs);
360 elem = alloc_pages_on_heap(heap, pg_sz, elt_size, socket, flags, align,
361 bound, contig, ms, n_segs);
366 map_addr = ms[0]->addr;
368 /* notify user about changes in memory map */
369 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC, map_addr, alloc_sz);
371 /* notify other processes that this has happened */
372 if (request_sync()) {
373 /* we couldn't ensure all processes have mapped memory,
374 * so free it back and notify everyone that it's been
377 * technically, we could've avoided adding memory addresses to
378 * the map, but that would've led to inconsistent behavior
379 * between primary and secondary processes, as those get
380 * callbacks during sync. therefore, force primary process to
381 * do alloc-and-rollback syncs as well.
383 callback_triggered = true;
386 heap->total_size += alloc_sz;
388 RTE_LOG(DEBUG, EAL, "Heap on socket %d was expanded by %zdMB\n",
389 socket, alloc_sz >> 20ULL);
396 if (callback_triggered)
397 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
400 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
410 try_expand_heap_secondary(struct malloc_heap *heap, uint64_t pg_sz,
411 size_t elt_size, int socket, unsigned int flags, size_t align,
412 size_t bound, bool contig)
414 struct malloc_mp_req req;
417 memset(&req, 0, sizeof(req));
419 req.t = REQ_TYPE_ALLOC;
420 req.alloc_req.align = align;
421 req.alloc_req.bound = bound;
422 req.alloc_req.contig = contig;
423 req.alloc_req.flags = flags;
424 req.alloc_req.elt_size = elt_size;
425 req.alloc_req.page_sz = pg_sz;
426 req.alloc_req.socket = socket;
427 req.alloc_req.heap = heap; /* it's in shared memory */
429 req_result = request_to_primary(&req);
434 if (req.result != REQ_RESULT_SUCCESS)
441 try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
442 int socket, unsigned int flags, size_t align, size_t bound,
445 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
448 rte_rwlock_write_lock(&mcfg->memory_hotplug_lock);
450 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
451 ret = try_expand_heap_primary(heap, pg_sz, elt_size, socket,
452 flags, align, bound, contig);
454 ret = try_expand_heap_secondary(heap, pg_sz, elt_size, socket,
455 flags, align, bound, contig);
458 rte_rwlock_write_unlock(&mcfg->memory_hotplug_lock);
463 compare_pagesz(const void *a, const void *b)
465 const struct rte_memseg_list * const*mpa = a;
466 const struct rte_memseg_list * const*mpb = b;
467 const struct rte_memseg_list *msla = *mpa;
468 const struct rte_memseg_list *mslb = *mpb;
469 uint64_t pg_sz_a = msla->page_sz;
470 uint64_t pg_sz_b = mslb->page_sz;
472 if (pg_sz_a < pg_sz_b)
474 if (pg_sz_a > pg_sz_b)
480 alloc_more_mem_on_socket(struct malloc_heap *heap, size_t size, int socket,
481 unsigned int flags, size_t align, size_t bound, bool contig)
483 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
484 struct rte_memseg_list *requested_msls[RTE_MAX_MEMSEG_LISTS];
485 struct rte_memseg_list *other_msls[RTE_MAX_MEMSEG_LISTS];
486 uint64_t requested_pg_sz[RTE_MAX_MEMSEG_LISTS];
487 uint64_t other_pg_sz[RTE_MAX_MEMSEG_LISTS];
489 int i, n_other_msls, n_other_pg_sz, n_requested_msls, n_requested_pg_sz;
490 bool size_hint = (flags & RTE_MEMZONE_SIZE_HINT_ONLY) > 0;
491 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
494 memset(requested_msls, 0, sizeof(requested_msls));
495 memset(other_msls, 0, sizeof(other_msls));
496 memset(requested_pg_sz, 0, sizeof(requested_pg_sz));
497 memset(other_pg_sz, 0, sizeof(other_pg_sz));
500 * go through memseg list and take note of all the page sizes available,
501 * and if any of them were specifically requested by the user.
503 n_requested_msls = 0;
505 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
506 struct rte_memseg_list *msl = &mcfg->memsegs[i];
508 if (msl->socket_id != socket)
511 if (msl->base_va == NULL)
514 /* if pages of specific size were requested */
515 if (size_flags != 0 && check_hugepage_sz(size_flags,
517 requested_msls[n_requested_msls++] = msl;
518 else if (size_flags == 0 || size_hint)
519 other_msls[n_other_msls++] = msl;
522 /* sort the lists, smallest first */
523 qsort(requested_msls, n_requested_msls, sizeof(requested_msls[0]),
525 qsort(other_msls, n_other_msls, sizeof(other_msls[0]),
528 /* now, extract page sizes we are supposed to try */
530 n_requested_pg_sz = 0;
531 for (i = 0; i < n_requested_msls; i++) {
532 uint64_t pg_sz = requested_msls[i]->page_sz;
534 if (prev_pg_sz != pg_sz) {
535 requested_pg_sz[n_requested_pg_sz++] = pg_sz;
541 for (i = 0; i < n_other_msls; i++) {
542 uint64_t pg_sz = other_msls[i]->page_sz;
544 if (prev_pg_sz != pg_sz) {
545 other_pg_sz[n_other_pg_sz++] = pg_sz;
550 /* finally, try allocating memory of specified page sizes, starting from
553 for (i = 0; i < n_requested_pg_sz; i++) {
554 uint64_t pg_sz = requested_pg_sz[i];
557 * do not pass the size hint here, as user expects other page
558 * sizes first, before resorting to best effort allocation.
560 if (!try_expand_heap(heap, pg_sz, size, socket, size_flags,
561 align, bound, contig))
564 if (n_other_pg_sz == 0)
567 /* now, check if we can reserve anything with size hint */
568 ret = find_suitable_element(heap, size, flags, align, bound, contig);
573 * we still couldn't reserve memory, so try expanding heap with other
574 * page sizes, if there are any
576 for (i = 0; i < n_other_pg_sz; i++) {
577 uint64_t pg_sz = other_pg_sz[i];
579 if (!try_expand_heap(heap, pg_sz, size, socket, flags,
580 align, bound, contig))
586 /* this will try lower page sizes first */
588 malloc_heap_alloc_on_heap_id(const char *type, size_t size,
589 unsigned int heap_id, unsigned int flags, size_t align,
590 size_t bound, bool contig)
592 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
593 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
594 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
598 rte_spinlock_lock(&(heap->lock));
600 align = align == 0 ? 1 : align;
602 /* for legacy mode, try once and with all flags */
603 if (internal_config.legacy_mem) {
604 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
609 * we do not pass the size hint here, because even if allocation fails,
610 * we may still be able to allocate memory from appropriate page sizes,
611 * we just need to request more memory first.
614 socket_id = rte_socket_id_by_idx(heap_id);
616 * if socket ID is negative, we cannot find a socket ID for this heap -
617 * which means it's an external heap. those can have unexpected page
618 * sizes, so if the user asked to allocate from there - assume user
619 * knows what they're doing, and allow allocating from there with any
623 size_flags |= RTE_MEMZONE_SIZE_HINT_ONLY;
625 ret = heap_alloc(heap, type, size, size_flags, align, bound, contig);
629 /* if socket ID is invalid, this is an external heap */
633 if (!alloc_more_mem_on_socket(heap, size, socket_id, flags, align,
635 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
637 /* this should have succeeded */
639 RTE_LOG(ERR, EAL, "Error allocating from heap\n");
642 rte_spinlock_unlock(&(heap->lock));
647 malloc_heap_alloc(const char *type, size_t size, int socket_arg,
648 unsigned int flags, size_t align, size_t bound, bool contig)
650 int socket, heap_id, i;
653 /* return NULL if size is 0 or alignment is not power-of-2 */
654 if (size == 0 || (align && !rte_is_power_of_2(align)))
657 if (!rte_eal_has_hugepages() && socket_arg < RTE_MAX_NUMA_NODES)
658 socket_arg = SOCKET_ID_ANY;
660 if (socket_arg == SOCKET_ID_ANY)
661 socket = malloc_get_numa_socket();
665 /* turn socket ID into heap ID */
666 heap_id = malloc_socket_to_heap_id(socket);
667 /* if heap id is negative, socket ID was invalid */
671 ret = malloc_heap_alloc_on_heap_id(type, size, heap_id, flags, align,
673 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
676 /* try other heaps. we are only iterating through native DPDK sockets,
677 * so external heaps won't be included.
679 for (i = 0; i < (int) rte_socket_count(); i++) {
682 ret = malloc_heap_alloc_on_heap_id(type, size, i, flags, align,
691 heap_alloc_biggest_on_heap_id(const char *type, unsigned int heap_id,
692 unsigned int flags, size_t align, bool contig)
694 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
695 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
698 rte_spinlock_lock(&(heap->lock));
700 align = align == 0 ? 1 : align;
702 ret = heap_alloc_biggest(heap, type, flags, align, contig);
704 rte_spinlock_unlock(&(heap->lock));
710 malloc_heap_alloc_biggest(const char *type, int socket_arg, unsigned int flags,
711 size_t align, bool contig)
713 int socket, i, cur_socket, heap_id;
716 /* return NULL if align is not power-of-2 */
717 if ((align && !rte_is_power_of_2(align)))
720 if (!rte_eal_has_hugepages())
721 socket_arg = SOCKET_ID_ANY;
723 if (socket_arg == SOCKET_ID_ANY)
724 socket = malloc_get_numa_socket();
728 /* turn socket ID into heap ID */
729 heap_id = malloc_socket_to_heap_id(socket);
730 /* if heap id is negative, socket ID was invalid */
734 ret = heap_alloc_biggest_on_heap_id(type, heap_id, flags, align,
736 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
739 /* try other heaps */
740 for (i = 0; i < (int) rte_socket_count(); i++) {
741 cur_socket = rte_socket_id_by_idx(i);
742 if (cur_socket == socket)
744 ret = heap_alloc_biggest_on_heap_id(type, i, flags, align,
752 /* this function is exposed in malloc_mp.h */
754 malloc_heap_free_pages(void *aligned_start, size_t aligned_len)
756 int n_segs, seg_idx, max_seg_idx;
757 struct rte_memseg_list *msl;
760 msl = rte_mem_virt2memseg_list(aligned_start);
764 page_sz = (size_t)msl->page_sz;
765 n_segs = aligned_len / page_sz;
766 seg_idx = RTE_PTR_DIFF(aligned_start, msl->base_va) / page_sz;
767 max_seg_idx = seg_idx + n_segs;
769 for (; seg_idx < max_seg_idx; seg_idx++) {
770 struct rte_memseg *ms;
772 ms = rte_fbarray_get(&msl->memseg_arr, seg_idx);
773 eal_memalloc_free_seg(ms);
779 malloc_heap_free(struct malloc_elem *elem)
781 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
782 struct malloc_heap *heap;
783 void *start, *aligned_start, *end, *aligned_end;
784 size_t len, aligned_len, page_sz;
785 struct rte_memseg_list *msl;
786 unsigned int i, n_segs, before_space, after_space;
789 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
792 /* elem may be merged with previous element, so keep heap address */
795 page_sz = (size_t)msl->page_sz;
797 rte_spinlock_lock(&(heap->lock));
799 /* mark element as free */
800 elem->state = ELEM_FREE;
802 elem = malloc_elem_free(elem);
804 /* anything after this is a bonus */
807 /* ...of which we can't avail if we are in legacy mode, or if this is an
808 * externally allocated segment.
810 if (internal_config.legacy_mem || (msl->external > 0))
813 /* check if we can free any memory back to the system */
814 if (elem->size < page_sz)
817 /* probably, but let's make sure, as we may not be using up full page */
820 aligned_start = RTE_PTR_ALIGN_CEIL(start, page_sz);
821 end = RTE_PTR_ADD(elem, len);
822 aligned_end = RTE_PTR_ALIGN_FLOOR(end, page_sz);
824 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
826 /* can't free anything */
827 if (aligned_len < page_sz)
830 /* we can free something. however, some of these pages may be marked as
831 * unfreeable, so also check that as well
833 n_segs = aligned_len / page_sz;
834 for (i = 0; i < n_segs; i++) {
835 const struct rte_memseg *tmp =
836 rte_mem_virt2memseg(aligned_start, msl);
838 if (tmp->flags & RTE_MEMSEG_FLAG_DO_NOT_FREE) {
839 /* this is an unfreeable segment, so move start */
840 aligned_start = RTE_PTR_ADD(tmp->addr, tmp->len);
844 /* recalculate length and number of segments */
845 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
846 n_segs = aligned_len / page_sz;
848 /* check if we can still free some pages */
852 /* We're not done yet. We also have to check if by freeing space we will
853 * be leaving free elements that are too small to store new elements.
854 * Check if we have enough space in the beginning and at the end, or if
855 * start/end are exactly page aligned.
857 before_space = RTE_PTR_DIFF(aligned_start, elem);
858 after_space = RTE_PTR_DIFF(end, aligned_end);
859 if (before_space != 0 &&
860 before_space < MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
861 /* There is not enough space before start, but we may be able to
862 * move the start forward by one page.
868 aligned_start = RTE_PTR_ADD(aligned_start, page_sz);
869 aligned_len -= page_sz;
872 if (after_space != 0 && after_space <
873 MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
874 /* There is not enough space after end, but we may be able to
875 * move the end backwards by one page.
881 aligned_end = RTE_PTR_SUB(aligned_end, page_sz);
882 aligned_len -= page_sz;
886 /* now we can finally free us some pages */
888 rte_rwlock_write_lock(&mcfg->memory_hotplug_lock);
891 * we allow secondary processes to clear the heap of this allocated
892 * memory because it is safe to do so, as even if notifications about
893 * unmapped pages don't make it to other processes, heap is shared
894 * across all processes, and will become empty of this memory anyway,
895 * and nothing can allocate it back unless primary process will be able
896 * to deliver allocation message to every single running process.
899 malloc_elem_free_list_remove(elem);
901 malloc_elem_hide_region(elem, (void *) aligned_start, aligned_len);
903 heap->total_size -= aligned_len;
905 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
906 /* notify user about changes in memory map */
907 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
908 aligned_start, aligned_len);
910 /* don't care if any of this fails */
911 malloc_heap_free_pages(aligned_start, aligned_len);
915 struct malloc_mp_req req;
917 memset(&req, 0, sizeof(req));
919 req.t = REQ_TYPE_FREE;
920 req.free_req.addr = aligned_start;
921 req.free_req.len = aligned_len;
924 * we request primary to deallocate pages, but we don't do it
925 * in this thread. instead, we notify primary that we would like
926 * to deallocate pages, and this process will receive another
927 * request (in parallel) that will do it for us on another
930 * we also don't really care if this succeeds - the data is
931 * already removed from the heap, so it is, for all intents and
932 * purposes, hidden from the rest of DPDK even if some other
933 * process (including this one) may have these pages mapped.
935 * notifications about deallocated memory happen during sync.
937 request_to_primary(&req);
940 RTE_LOG(DEBUG, EAL, "Heap on socket %d was shrunk by %zdMB\n",
941 msl->socket_id, aligned_len >> 20ULL);
943 rte_rwlock_write_unlock(&mcfg->memory_hotplug_lock);
945 rte_spinlock_unlock(&(heap->lock));
950 malloc_heap_resize(struct malloc_elem *elem, size_t size)
954 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
957 rte_spinlock_lock(&(elem->heap->lock));
959 ret = malloc_elem_resize(elem, size);
961 rte_spinlock_unlock(&(elem->heap->lock));
967 * Function to retrieve data for a given heap
970 malloc_heap_get_stats(struct malloc_heap *heap,
971 struct rte_malloc_socket_stats *socket_stats)
974 struct malloc_elem *elem;
976 rte_spinlock_lock(&heap->lock);
978 /* Initialise variables for heap */
979 socket_stats->free_count = 0;
980 socket_stats->heap_freesz_bytes = 0;
981 socket_stats->greatest_free_size = 0;
983 /* Iterate through free list */
984 for (idx = 0; idx < RTE_HEAP_NUM_FREELISTS; idx++) {
985 for (elem = LIST_FIRST(&heap->free_head[idx]);
986 !!elem; elem = LIST_NEXT(elem, free_list))
988 socket_stats->free_count++;
989 socket_stats->heap_freesz_bytes += elem->size;
990 if (elem->size > socket_stats->greatest_free_size)
991 socket_stats->greatest_free_size = elem->size;
994 /* Get stats on overall heap and allocated memory on this heap */
995 socket_stats->heap_totalsz_bytes = heap->total_size;
996 socket_stats->heap_allocsz_bytes = (socket_stats->heap_totalsz_bytes -
997 socket_stats->heap_freesz_bytes);
998 socket_stats->alloc_count = heap->alloc_count;
1000 rte_spinlock_unlock(&heap->lock);
1005 * Function to retrieve data for a given heap
1008 malloc_heap_dump(struct malloc_heap *heap, FILE *f)
1010 struct malloc_elem *elem;
1012 rte_spinlock_lock(&heap->lock);
1014 fprintf(f, "Heap size: 0x%zx\n", heap->total_size);
1015 fprintf(f, "Heap alloc count: %u\n", heap->alloc_count);
1019 malloc_elem_dump(elem, f);
1023 rte_spinlock_unlock(&heap->lock);
1027 destroy_seg(struct malloc_elem *elem, size_t len)
1029 struct malloc_heap *heap = elem->heap;
1030 struct rte_memseg_list *msl;
1034 /* notify all subscribers that a memory area is going to be removed */
1035 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE, elem, len);
1037 /* this element can be removed */
1038 malloc_elem_free_list_remove(elem);
1039 malloc_elem_hide_region(elem, elem, len);
1041 heap->total_size -= len;
1043 memset(elem, 0, sizeof(*elem));
1045 /* destroy the fbarray backing this memory */
1046 if (rte_fbarray_destroy(&msl->memseg_arr) < 0)
1049 /* reset the memseg list */
1050 memset(msl, 0, sizeof(*msl));
1056 malloc_heap_add_external_memory(struct malloc_heap *heap, void *va_addr,
1057 rte_iova_t iova_addrs[], unsigned int n_pages, size_t page_sz)
1059 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1060 char fbarray_name[RTE_FBARRAY_NAME_LEN];
1061 struct rte_memseg_list *msl = NULL;
1062 struct rte_fbarray *arr;
1063 size_t seg_len = n_pages * page_sz;
1066 /* first, find a free memseg list */
1067 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1068 struct rte_memseg_list *tmp = &mcfg->memsegs[i];
1069 if (tmp->base_va == NULL) {
1075 RTE_LOG(ERR, EAL, "Couldn't find empty memseg list\n");
1080 snprintf(fbarray_name, sizeof(fbarray_name) - 1, "%s_%p",
1081 heap->name, va_addr);
1083 /* create the backing fbarray */
1084 if (rte_fbarray_init(&msl->memseg_arr, fbarray_name, n_pages,
1085 sizeof(struct rte_memseg)) < 0) {
1086 RTE_LOG(ERR, EAL, "Couldn't create fbarray backing the memseg list\n");
1089 arr = &msl->memseg_arr;
1091 /* fbarray created, fill it up */
1092 for (i = 0; i < n_pages; i++) {
1093 struct rte_memseg *ms;
1095 rte_fbarray_set_used(arr, i);
1096 ms = rte_fbarray_get(arr, i);
1097 ms->addr = RTE_PTR_ADD(va_addr, i * page_sz);
1098 ms->iova = iova_addrs == NULL ? RTE_BAD_IOVA : iova_addrs[i];
1099 ms->hugepage_sz = page_sz;
1101 ms->nchannel = rte_memory_get_nchannel();
1102 ms->nrank = rte_memory_get_nrank();
1103 ms->socket_id = heap->socket_id;
1106 /* set up the memseg list */
1107 msl->base_va = va_addr;
1108 msl->page_sz = page_sz;
1109 msl->socket_id = heap->socket_id;
1114 /* erase contents of new memory */
1115 memset(va_addr, 0, seg_len);
1117 /* now, add newly minted memory to the malloc heap */
1118 malloc_heap_add_memory(heap, msl, va_addr, seg_len);
1120 heap->total_size += seg_len;
1123 RTE_LOG(DEBUG, EAL, "Added segment for heap %s starting at %p\n",
1124 heap->name, va_addr);
1126 /* notify all subscribers that a new memory area has been added */
1127 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC,
1134 malloc_heap_remove_external_memory(struct malloc_heap *heap, void *va_addr,
1137 struct malloc_elem *elem = heap->first;
1139 /* find element with specified va address */
1140 while (elem != NULL && elem != va_addr) {
1142 /* stop if we've blown past our VA */
1143 if (elem > (struct malloc_elem *)va_addr) {
1148 /* check if element was found */
1149 if (elem == NULL || elem->msl->len != len) {
1153 /* if element's size is not equal to segment len, segment is busy */
1154 if (elem->state == ELEM_BUSY || elem->size != len) {
1158 return destroy_seg(elem, len);
1162 malloc_heap_create(struct malloc_heap *heap, const char *heap_name)
1164 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1165 uint32_t next_socket_id = mcfg->next_socket_id;
1167 /* prevent overflow. did you really create 2 billion heaps??? */
1168 if (next_socket_id > INT32_MAX) {
1169 RTE_LOG(ERR, EAL, "Cannot assign new socket ID's\n");
1174 /* initialize empty heap */
1175 heap->alloc_count = 0;
1178 LIST_INIT(heap->free_head);
1179 rte_spinlock_init(&heap->lock);
1180 heap->total_size = 0;
1181 heap->socket_id = next_socket_id;
1183 /* we hold a global mem hotplug writelock, so it's safe to increment */
1184 mcfg->next_socket_id++;
1187 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1192 malloc_heap_destroy(struct malloc_heap *heap)
1194 if (heap->alloc_count != 0) {
1195 RTE_LOG(ERR, EAL, "Heap is still in use\n");
1199 if (heap->first != NULL || heap->last != NULL) {
1200 RTE_LOG(ERR, EAL, "Heap still contains memory segments\n");
1204 if (heap->total_size != 0)
1205 RTE_LOG(ERR, EAL, "Total size not zero, heap is likely corrupt\n");
1207 /* after this, the lock will be dropped */
1208 memset(heap, 0, sizeof(*heap));
1214 rte_eal_malloc_heap_init(void)
1216 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1219 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1220 /* assign min socket ID to external heaps */
1221 mcfg->next_socket_id = EXTERNAL_HEAP_MIN_SOCKET_ID;
1223 /* assign names to default DPDK heaps */
1224 for (i = 0; i < rte_socket_count(); i++) {
1225 struct malloc_heap *heap = &mcfg->malloc_heaps[i];
1226 char heap_name[RTE_HEAP_NAME_MAX_LEN];
1227 int socket_id = rte_socket_id_by_idx(i);
1229 snprintf(heap_name, sizeof(heap_name) - 1,
1230 "socket_%i", socket_id);
1231 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1232 heap->socket_id = socket_id;
1237 if (register_mp_requests()) {
1238 RTE_LOG(ERR, EAL, "Couldn't register malloc multiprocess actions\n");
1239 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
1243 /* unlock mem hotplug here. it's safe for primary as no requests can
1244 * even come before primary itself is fully initialized, and secondaries
1245 * do not need to initialize the heap.
1247 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
1249 /* secondary process does not need to initialize anything */
1250 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1253 /* add all IOVA-contiguous areas to the heap */
1254 return rte_memseg_contig_walk(malloc_add_seg, NULL);