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, elem, 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",
325 * Once we have all the memseg lists configured, if there is a dma mask
326 * set, check iova addresses are not out of range. Otherwise the device
327 * setting the dma mask could have problems with the mapped memory.
329 * There are two situations when this can happen:
330 * 1) memory initialization
331 * 2) dynamic memory allocation
333 * For 1), an error when checking dma mask implies app can not be
334 * executed. For 2) implies the new memory can not be added.
336 if (mcfg->dma_maskbits &&
337 rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
339 * Currently this can only happen if IOMMU is enabled
340 * and the address width supported by the IOMMU hw is
341 * not enough for using the memory mapped IOVAs.
343 * If IOVA is VA, advice to try with '--iova-mode pa'
344 * which could solve some situations when IOVA VA is not
348 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask\n",
352 * If IOVA is VA and it is possible to run with IOVA PA,
353 * because user is root, give and advice for solving the
356 if ((rte_eal_iova_mode() == RTE_IOVA_VA) &&
357 rte_eal_using_phys_addrs())
359 "%s(): Please try initializing EAL with --iova-mode=pa parameter\n",
364 /* add newly minted memsegs to malloc heap */
365 elem = malloc_heap_add_memory(heap, msl, map_addr, alloc_sz);
367 /* try once more, as now we have allocated new memory */
368 ret = find_suitable_element(heap, elt_size, flags, align, bound,
377 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
382 try_expand_heap_primary(struct malloc_heap *heap, uint64_t pg_sz,
383 size_t elt_size, int socket, unsigned int flags, size_t align,
384 size_t bound, bool contig)
386 struct malloc_elem *elem;
387 struct rte_memseg **ms;
391 bool callback_triggered = false;
393 alloc_sz = RTE_ALIGN_CEIL(align + elt_size +
394 MALLOC_ELEM_TRAILER_LEN, pg_sz);
395 n_segs = alloc_sz / pg_sz;
397 /* we can't know in advance how many pages we'll need, so we malloc */
398 ms = malloc(sizeof(*ms) * n_segs);
401 memset(ms, 0, sizeof(*ms) * n_segs);
403 elem = alloc_pages_on_heap(heap, pg_sz, elt_size, socket, flags, align,
404 bound, contig, ms, n_segs);
409 map_addr = ms[0]->addr;
411 /* notify user about changes in memory map */
412 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC, map_addr, alloc_sz);
414 /* notify other processes that this has happened */
415 if (request_sync()) {
416 /* we couldn't ensure all processes have mapped memory,
417 * so free it back and notify everyone that it's been
420 * technically, we could've avoided adding memory addresses to
421 * the map, but that would've led to inconsistent behavior
422 * between primary and secondary processes, as those get
423 * callbacks during sync. therefore, force primary process to
424 * do alloc-and-rollback syncs as well.
426 callback_triggered = true;
429 heap->total_size += alloc_sz;
431 RTE_LOG(DEBUG, EAL, "Heap on socket %d was expanded by %zdMB\n",
432 socket, alloc_sz >> 20ULL);
439 if (callback_triggered)
440 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
443 rollback_expand_heap(ms, n_segs, elem, map_addr, alloc_sz);
453 try_expand_heap_secondary(struct malloc_heap *heap, uint64_t pg_sz,
454 size_t elt_size, int socket, unsigned int flags, size_t align,
455 size_t bound, bool contig)
457 struct malloc_mp_req req;
460 memset(&req, 0, sizeof(req));
462 req.t = REQ_TYPE_ALLOC;
463 req.alloc_req.align = align;
464 req.alloc_req.bound = bound;
465 req.alloc_req.contig = contig;
466 req.alloc_req.flags = flags;
467 req.alloc_req.elt_size = elt_size;
468 req.alloc_req.page_sz = pg_sz;
469 req.alloc_req.socket = socket;
470 req.alloc_req.heap = heap; /* it's in shared memory */
472 req_result = request_to_primary(&req);
477 if (req.result != REQ_RESULT_SUCCESS)
484 try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
485 int socket, unsigned int flags, size_t align, size_t bound,
488 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
491 rte_rwlock_write_lock(&mcfg->memory_hotplug_lock);
493 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
494 ret = try_expand_heap_primary(heap, pg_sz, elt_size, socket,
495 flags, align, bound, contig);
497 ret = try_expand_heap_secondary(heap, pg_sz, elt_size, socket,
498 flags, align, bound, contig);
501 rte_rwlock_write_unlock(&mcfg->memory_hotplug_lock);
506 compare_pagesz(const void *a, const void *b)
508 const struct rte_memseg_list * const*mpa = a;
509 const struct rte_memseg_list * const*mpb = b;
510 const struct rte_memseg_list *msla = *mpa;
511 const struct rte_memseg_list *mslb = *mpb;
512 uint64_t pg_sz_a = msla->page_sz;
513 uint64_t pg_sz_b = mslb->page_sz;
515 if (pg_sz_a < pg_sz_b)
517 if (pg_sz_a > pg_sz_b)
523 alloc_more_mem_on_socket(struct malloc_heap *heap, size_t size, int socket,
524 unsigned int flags, size_t align, size_t bound, bool contig)
526 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
527 struct rte_memseg_list *requested_msls[RTE_MAX_MEMSEG_LISTS];
528 struct rte_memseg_list *other_msls[RTE_MAX_MEMSEG_LISTS];
529 uint64_t requested_pg_sz[RTE_MAX_MEMSEG_LISTS];
530 uint64_t other_pg_sz[RTE_MAX_MEMSEG_LISTS];
532 int i, n_other_msls, n_other_pg_sz, n_requested_msls, n_requested_pg_sz;
533 bool size_hint = (flags & RTE_MEMZONE_SIZE_HINT_ONLY) > 0;
534 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
537 memset(requested_msls, 0, sizeof(requested_msls));
538 memset(other_msls, 0, sizeof(other_msls));
539 memset(requested_pg_sz, 0, sizeof(requested_pg_sz));
540 memset(other_pg_sz, 0, sizeof(other_pg_sz));
543 * go through memseg list and take note of all the page sizes available,
544 * and if any of them were specifically requested by the user.
546 n_requested_msls = 0;
548 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
549 struct rte_memseg_list *msl = &mcfg->memsegs[i];
551 if (msl->socket_id != socket)
554 if (msl->base_va == NULL)
557 /* if pages of specific size were requested */
558 if (size_flags != 0 && check_hugepage_sz(size_flags,
560 requested_msls[n_requested_msls++] = msl;
561 else if (size_flags == 0 || size_hint)
562 other_msls[n_other_msls++] = msl;
565 /* sort the lists, smallest first */
566 qsort(requested_msls, n_requested_msls, sizeof(requested_msls[0]),
568 qsort(other_msls, n_other_msls, sizeof(other_msls[0]),
571 /* now, extract page sizes we are supposed to try */
573 n_requested_pg_sz = 0;
574 for (i = 0; i < n_requested_msls; i++) {
575 uint64_t pg_sz = requested_msls[i]->page_sz;
577 if (prev_pg_sz != pg_sz) {
578 requested_pg_sz[n_requested_pg_sz++] = pg_sz;
584 for (i = 0; i < n_other_msls; i++) {
585 uint64_t pg_sz = other_msls[i]->page_sz;
587 if (prev_pg_sz != pg_sz) {
588 other_pg_sz[n_other_pg_sz++] = pg_sz;
593 /* finally, try allocating memory of specified page sizes, starting from
596 for (i = 0; i < n_requested_pg_sz; i++) {
597 uint64_t pg_sz = requested_pg_sz[i];
600 * do not pass the size hint here, as user expects other page
601 * sizes first, before resorting to best effort allocation.
603 if (!try_expand_heap(heap, pg_sz, size, socket, size_flags,
604 align, bound, contig))
607 if (n_other_pg_sz == 0)
610 /* now, check if we can reserve anything with size hint */
611 ret = find_suitable_element(heap, size, flags, align, bound, contig);
616 * we still couldn't reserve memory, so try expanding heap with other
617 * page sizes, if there are any
619 for (i = 0; i < n_other_pg_sz; i++) {
620 uint64_t pg_sz = other_pg_sz[i];
622 if (!try_expand_heap(heap, pg_sz, size, socket, flags,
623 align, bound, contig))
629 /* this will try lower page sizes first */
631 malloc_heap_alloc_on_heap_id(const char *type, size_t size,
632 unsigned int heap_id, unsigned int flags, size_t align,
633 size_t bound, bool contig)
635 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
636 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
637 unsigned int size_flags = flags & ~RTE_MEMZONE_SIZE_HINT_ONLY;
641 rte_spinlock_lock(&(heap->lock));
643 align = align == 0 ? 1 : align;
645 /* for legacy mode, try once and with all flags */
646 if (internal_config.legacy_mem) {
647 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
652 * we do not pass the size hint here, because even if allocation fails,
653 * we may still be able to allocate memory from appropriate page sizes,
654 * we just need to request more memory first.
657 socket_id = rte_socket_id_by_idx(heap_id);
659 * if socket ID is negative, we cannot find a socket ID for this heap -
660 * which means it's an external heap. those can have unexpected page
661 * sizes, so if the user asked to allocate from there - assume user
662 * knows what they're doing, and allow allocating from there with any
666 size_flags |= RTE_MEMZONE_SIZE_HINT_ONLY;
668 ret = heap_alloc(heap, type, size, size_flags, align, bound, contig);
672 /* if socket ID is invalid, this is an external heap */
676 if (!alloc_more_mem_on_socket(heap, size, socket_id, flags, align,
678 ret = heap_alloc(heap, type, size, flags, align, bound, contig);
680 /* this should have succeeded */
682 RTE_LOG(ERR, EAL, "Error allocating from heap\n");
685 rte_spinlock_unlock(&(heap->lock));
690 malloc_heap_alloc(const char *type, size_t size, int socket_arg,
691 unsigned int flags, size_t align, size_t bound, bool contig)
693 int socket, heap_id, i;
696 /* return NULL if size is 0 or alignment is not power-of-2 */
697 if (size == 0 || (align && !rte_is_power_of_2(align)))
700 if (!rte_eal_has_hugepages() && socket_arg < RTE_MAX_NUMA_NODES)
701 socket_arg = SOCKET_ID_ANY;
703 if (socket_arg == SOCKET_ID_ANY)
704 socket = malloc_get_numa_socket();
708 /* turn socket ID into heap ID */
709 heap_id = malloc_socket_to_heap_id(socket);
710 /* if heap id is negative, socket ID was invalid */
714 ret = malloc_heap_alloc_on_heap_id(type, size, heap_id, flags, align,
716 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
719 /* try other heaps. we are only iterating through native DPDK sockets,
720 * so external heaps won't be included.
722 for (i = 0; i < (int) rte_socket_count(); i++) {
725 ret = malloc_heap_alloc_on_heap_id(type, size, i, flags, align,
734 heap_alloc_biggest_on_heap_id(const char *type, unsigned int heap_id,
735 unsigned int flags, size_t align, bool contig)
737 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
738 struct malloc_heap *heap = &mcfg->malloc_heaps[heap_id];
741 rte_spinlock_lock(&(heap->lock));
743 align = align == 0 ? 1 : align;
745 ret = heap_alloc_biggest(heap, type, flags, align, contig);
747 rte_spinlock_unlock(&(heap->lock));
753 malloc_heap_alloc_biggest(const char *type, int socket_arg, unsigned int flags,
754 size_t align, bool contig)
756 int socket, i, cur_socket, heap_id;
759 /* return NULL if align is not power-of-2 */
760 if ((align && !rte_is_power_of_2(align)))
763 if (!rte_eal_has_hugepages())
764 socket_arg = SOCKET_ID_ANY;
766 if (socket_arg == SOCKET_ID_ANY)
767 socket = malloc_get_numa_socket();
771 /* turn socket ID into heap ID */
772 heap_id = malloc_socket_to_heap_id(socket);
773 /* if heap id is negative, socket ID was invalid */
777 ret = heap_alloc_biggest_on_heap_id(type, heap_id, flags, align,
779 if (ret != NULL || socket_arg != SOCKET_ID_ANY)
782 /* try other heaps */
783 for (i = 0; i < (int) rte_socket_count(); i++) {
784 cur_socket = rte_socket_id_by_idx(i);
785 if (cur_socket == socket)
787 ret = heap_alloc_biggest_on_heap_id(type, i, flags, align,
795 /* this function is exposed in malloc_mp.h */
797 malloc_heap_free_pages(void *aligned_start, size_t aligned_len)
799 int n_segs, seg_idx, max_seg_idx;
800 struct rte_memseg_list *msl;
803 msl = rte_mem_virt2memseg_list(aligned_start);
807 page_sz = (size_t)msl->page_sz;
808 n_segs = aligned_len / page_sz;
809 seg_idx = RTE_PTR_DIFF(aligned_start, msl->base_va) / page_sz;
810 max_seg_idx = seg_idx + n_segs;
812 for (; seg_idx < max_seg_idx; seg_idx++) {
813 struct rte_memseg *ms;
815 ms = rte_fbarray_get(&msl->memseg_arr, seg_idx);
816 eal_memalloc_free_seg(ms);
822 malloc_heap_free(struct malloc_elem *elem)
824 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
825 struct malloc_heap *heap;
826 void *start, *aligned_start, *end, *aligned_end;
827 size_t len, aligned_len, page_sz;
828 struct rte_memseg_list *msl;
829 unsigned int i, n_segs, before_space, after_space;
832 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
835 /* elem may be merged with previous element, so keep heap address */
838 page_sz = (size_t)msl->page_sz;
840 rte_spinlock_lock(&(heap->lock));
842 /* mark element as free */
843 elem->state = ELEM_FREE;
845 elem = malloc_elem_free(elem);
847 /* anything after this is a bonus */
850 /* ...of which we can't avail if we are in legacy mode, or if this is an
851 * externally allocated segment.
853 if (internal_config.legacy_mem || (msl->external > 0))
856 /* check if we can free any memory back to the system */
857 if (elem->size < page_sz)
860 /* if user requested to match allocations, the sizes must match - if not,
861 * we will defer freeing these hugepages until the entire original allocation
864 if (internal_config.match_allocations && elem->size != elem->orig_size)
867 /* probably, but let's make sure, as we may not be using up full page */
870 aligned_start = RTE_PTR_ALIGN_CEIL(start, page_sz);
871 end = RTE_PTR_ADD(elem, len);
872 aligned_end = RTE_PTR_ALIGN_FLOOR(end, page_sz);
874 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
876 /* can't free anything */
877 if (aligned_len < page_sz)
880 /* we can free something. however, some of these pages may be marked as
881 * unfreeable, so also check that as well
883 n_segs = aligned_len / page_sz;
884 for (i = 0; i < n_segs; i++) {
885 const struct rte_memseg *tmp =
886 rte_mem_virt2memseg(aligned_start, msl);
888 if (tmp->flags & RTE_MEMSEG_FLAG_DO_NOT_FREE) {
889 /* this is an unfreeable segment, so move start */
890 aligned_start = RTE_PTR_ADD(tmp->addr, tmp->len);
894 /* recalculate length and number of segments */
895 aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
896 n_segs = aligned_len / page_sz;
898 /* check if we can still free some pages */
902 /* We're not done yet. We also have to check if by freeing space we will
903 * be leaving free elements that are too small to store new elements.
904 * Check if we have enough space in the beginning and at the end, or if
905 * start/end are exactly page aligned.
907 before_space = RTE_PTR_DIFF(aligned_start, elem);
908 after_space = RTE_PTR_DIFF(end, aligned_end);
909 if (before_space != 0 &&
910 before_space < MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
911 /* There is not enough space before start, but we may be able to
912 * move the start forward by one page.
918 aligned_start = RTE_PTR_ADD(aligned_start, page_sz);
919 aligned_len -= page_sz;
922 if (after_space != 0 && after_space <
923 MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
924 /* There is not enough space after end, but we may be able to
925 * move the end backwards by one page.
931 aligned_end = RTE_PTR_SUB(aligned_end, page_sz);
932 aligned_len -= page_sz;
936 /* now we can finally free us some pages */
938 rte_rwlock_write_lock(&mcfg->memory_hotplug_lock);
941 * we allow secondary processes to clear the heap of this allocated
942 * memory because it is safe to do so, as even if notifications about
943 * unmapped pages don't make it to other processes, heap is shared
944 * across all processes, and will become empty of this memory anyway,
945 * and nothing can allocate it back unless primary process will be able
946 * to deliver allocation message to every single running process.
949 malloc_elem_free_list_remove(elem);
951 malloc_elem_hide_region(elem, (void *) aligned_start, aligned_len);
953 heap->total_size -= aligned_len;
955 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
956 /* notify user about changes in memory map */
957 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
958 aligned_start, aligned_len);
960 /* don't care if any of this fails */
961 malloc_heap_free_pages(aligned_start, aligned_len);
965 struct malloc_mp_req req;
967 memset(&req, 0, sizeof(req));
969 req.t = REQ_TYPE_FREE;
970 req.free_req.addr = aligned_start;
971 req.free_req.len = aligned_len;
974 * we request primary to deallocate pages, but we don't do it
975 * in this thread. instead, we notify primary that we would like
976 * to deallocate pages, and this process will receive another
977 * request (in parallel) that will do it for us on another
980 * we also don't really care if this succeeds - the data is
981 * already removed from the heap, so it is, for all intents and
982 * purposes, hidden from the rest of DPDK even if some other
983 * process (including this one) may have these pages mapped.
985 * notifications about deallocated memory happen during sync.
987 request_to_primary(&req);
990 RTE_LOG(DEBUG, EAL, "Heap on socket %d was shrunk by %zdMB\n",
991 msl->socket_id, aligned_len >> 20ULL);
993 rte_rwlock_write_unlock(&mcfg->memory_hotplug_lock);
995 rte_spinlock_unlock(&(heap->lock));
1000 malloc_heap_resize(struct malloc_elem *elem, size_t size)
1004 if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
1007 rte_spinlock_lock(&(elem->heap->lock));
1009 ret = malloc_elem_resize(elem, size);
1011 rte_spinlock_unlock(&(elem->heap->lock));
1017 * Function to retrieve data for a given heap
1020 malloc_heap_get_stats(struct malloc_heap *heap,
1021 struct rte_malloc_socket_stats *socket_stats)
1024 struct malloc_elem *elem;
1026 rte_spinlock_lock(&heap->lock);
1028 /* Initialise variables for heap */
1029 socket_stats->free_count = 0;
1030 socket_stats->heap_freesz_bytes = 0;
1031 socket_stats->greatest_free_size = 0;
1033 /* Iterate through free list */
1034 for (idx = 0; idx < RTE_HEAP_NUM_FREELISTS; idx++) {
1035 for (elem = LIST_FIRST(&heap->free_head[idx]);
1036 !!elem; elem = LIST_NEXT(elem, free_list))
1038 socket_stats->free_count++;
1039 socket_stats->heap_freesz_bytes += elem->size;
1040 if (elem->size > socket_stats->greatest_free_size)
1041 socket_stats->greatest_free_size = elem->size;
1044 /* Get stats on overall heap and allocated memory on this heap */
1045 socket_stats->heap_totalsz_bytes = heap->total_size;
1046 socket_stats->heap_allocsz_bytes = (socket_stats->heap_totalsz_bytes -
1047 socket_stats->heap_freesz_bytes);
1048 socket_stats->alloc_count = heap->alloc_count;
1050 rte_spinlock_unlock(&heap->lock);
1055 * Function to retrieve data for a given heap
1058 malloc_heap_dump(struct malloc_heap *heap, FILE *f)
1060 struct malloc_elem *elem;
1062 rte_spinlock_lock(&heap->lock);
1064 fprintf(f, "Heap size: 0x%zx\n", heap->total_size);
1065 fprintf(f, "Heap alloc count: %u\n", heap->alloc_count);
1069 malloc_elem_dump(elem, f);
1073 rte_spinlock_unlock(&heap->lock);
1077 destroy_seg(struct malloc_elem *elem, size_t len)
1079 struct malloc_heap *heap = elem->heap;
1080 struct rte_memseg_list *msl;
1084 /* notify all subscribers that a memory area is going to be removed */
1085 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE, elem, len);
1087 /* this element can be removed */
1088 malloc_elem_free_list_remove(elem);
1089 malloc_elem_hide_region(elem, elem, len);
1091 heap->total_size -= len;
1093 memset(elem, 0, sizeof(*elem));
1095 /* destroy the fbarray backing this memory */
1096 if (rte_fbarray_destroy(&msl->memseg_arr) < 0)
1099 /* reset the memseg list */
1100 memset(msl, 0, sizeof(*msl));
1105 struct rte_memseg_list *
1106 malloc_heap_create_external_seg(void *va_addr, rte_iova_t iova_addrs[],
1107 unsigned int n_pages, size_t page_sz, const char *seg_name,
1108 unsigned int socket_id)
1110 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1111 char fbarray_name[RTE_FBARRAY_NAME_LEN];
1112 struct rte_memseg_list *msl = NULL;
1113 struct rte_fbarray *arr;
1114 size_t seg_len = n_pages * page_sz;
1117 /* first, find a free memseg list */
1118 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1119 struct rte_memseg_list *tmp = &mcfg->memsegs[i];
1120 if (tmp->base_va == NULL) {
1126 RTE_LOG(ERR, EAL, "Couldn't find empty memseg list\n");
1131 snprintf(fbarray_name, sizeof(fbarray_name) - 1, "%s_%p",
1134 /* create the backing fbarray */
1135 if (rte_fbarray_init(&msl->memseg_arr, fbarray_name, n_pages,
1136 sizeof(struct rte_memseg)) < 0) {
1137 RTE_LOG(ERR, EAL, "Couldn't create fbarray backing the memseg list\n");
1140 arr = &msl->memseg_arr;
1142 /* fbarray created, fill it up */
1143 for (i = 0; i < n_pages; i++) {
1144 struct rte_memseg *ms;
1146 rte_fbarray_set_used(arr, i);
1147 ms = rte_fbarray_get(arr, i);
1148 ms->addr = RTE_PTR_ADD(va_addr, i * page_sz);
1149 ms->iova = iova_addrs == NULL ? RTE_BAD_IOVA : iova_addrs[i];
1150 ms->hugepage_sz = page_sz;
1152 ms->nchannel = rte_memory_get_nchannel();
1153 ms->nrank = rte_memory_get_nrank();
1154 ms->socket_id = socket_id;
1157 /* set up the memseg list */
1158 msl->base_va = va_addr;
1159 msl->page_sz = page_sz;
1160 msl->socket_id = socket_id;
1169 malloc_heap_add_external_memory(struct malloc_heap *heap,
1170 struct rte_memseg_list *msl)
1172 /* erase contents of new memory */
1173 memset(msl->base_va, 0, msl->len);
1175 /* now, add newly minted memory to the malloc heap */
1176 malloc_heap_add_memory(heap, msl, msl->base_va, msl->len);
1178 heap->total_size += msl->len;
1181 RTE_LOG(DEBUG, EAL, "Added segment for heap %s starting at %p\n",
1182 heap->name, msl->base_va);
1184 /* notify all subscribers that a new memory area has been added */
1185 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC,
1186 msl->base_va, msl->len);
1192 malloc_heap_remove_external_memory(struct malloc_heap *heap, void *va_addr,
1195 struct malloc_elem *elem = heap->first;
1197 /* find element with specified va address */
1198 while (elem != NULL && elem != va_addr) {
1200 /* stop if we've blown past our VA */
1201 if (elem > (struct malloc_elem *)va_addr) {
1206 /* check if element was found */
1207 if (elem == NULL || elem->msl->len != len) {
1211 /* if element's size is not equal to segment len, segment is busy */
1212 if (elem->state == ELEM_BUSY || elem->size != len) {
1216 return destroy_seg(elem, len);
1220 malloc_heap_create(struct malloc_heap *heap, const char *heap_name)
1222 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1223 uint32_t next_socket_id = mcfg->next_socket_id;
1225 /* prevent overflow. did you really create 2 billion heaps??? */
1226 if (next_socket_id > INT32_MAX) {
1227 RTE_LOG(ERR, EAL, "Cannot assign new socket ID's\n");
1232 /* initialize empty heap */
1233 heap->alloc_count = 0;
1236 LIST_INIT(heap->free_head);
1237 rte_spinlock_init(&heap->lock);
1238 heap->total_size = 0;
1239 heap->socket_id = next_socket_id;
1241 /* we hold a global mem hotplug writelock, so it's safe to increment */
1242 mcfg->next_socket_id++;
1245 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1250 malloc_heap_destroy(struct malloc_heap *heap)
1252 if (heap->alloc_count != 0) {
1253 RTE_LOG(ERR, EAL, "Heap is still in use\n");
1257 if (heap->first != NULL || heap->last != NULL) {
1258 RTE_LOG(ERR, EAL, "Heap still contains memory segments\n");
1262 if (heap->total_size != 0)
1263 RTE_LOG(ERR, EAL, "Total size not zero, heap is likely corrupt\n");
1265 /* after this, the lock will be dropped */
1266 memset(heap, 0, sizeof(*heap));
1272 rte_eal_malloc_heap_init(void)
1274 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1277 if (internal_config.match_allocations) {
1278 RTE_LOG(DEBUG, EAL, "Hugepages will be freed exactly as allocated.\n");
1281 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1282 /* assign min socket ID to external heaps */
1283 mcfg->next_socket_id = EXTERNAL_HEAP_MIN_SOCKET_ID;
1285 /* assign names to default DPDK heaps */
1286 for (i = 0; i < rte_socket_count(); i++) {
1287 struct malloc_heap *heap = &mcfg->malloc_heaps[i];
1288 char heap_name[RTE_HEAP_NAME_MAX_LEN];
1289 int socket_id = rte_socket_id_by_idx(i);
1291 snprintf(heap_name, sizeof(heap_name) - 1,
1292 "socket_%i", socket_id);
1293 strlcpy(heap->name, heap_name, RTE_HEAP_NAME_MAX_LEN);
1294 heap->socket_id = socket_id;
1299 if (register_mp_requests()) {
1300 RTE_LOG(ERR, EAL, "Couldn't register malloc multiprocess actions\n");
1301 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
1305 /* unlock mem hotplug here. it's safe for primary as no requests can
1306 * even come before primary itself is fully initialized, and secondaries
1307 * do not need to initialize the heap.
1309 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
1311 /* secondary process does not need to initialize anything */
1312 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1315 /* add all IOVA-contiguous areas to the heap */
1316 return rte_memseg_contig_walk(malloc_add_seg, NULL);