1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
10 #include <sys/queue.h>
12 #include <rte_memory.h>
14 #include <rte_launch.h>
15 #include <rte_per_lcore.h>
16 #include <rte_lcore.h>
17 #include <rte_debug.h>
18 #include <rte_common.h>
19 #include <rte_spinlock.h>
21 #include "eal_internal_cfg.h"
22 #include "eal_memalloc.h"
23 #include "malloc_elem.h"
24 #include "malloc_heap.h"
27 * If debugging is enabled, freed memory is set to poison value
28 * to catch buggy programs. Otherwise, freed memory is set to zero
29 * to avoid having to zero in zmalloc
31 #ifdef RTE_MALLOC_DEBUG
32 #define MALLOC_POISON 0x6b
34 #define MALLOC_POISON 0
38 malloc_elem_find_max_iova_contig(struct malloc_elem *elem, size_t align)
40 void *cur_page, *contig_seg_start, *page_end, *cur_seg_end;
41 void *data_start, *data_end;
42 rte_iova_t expected_iova;
43 struct rte_memseg *ms;
44 size_t page_sz, cur, max;
46 page_sz = (size_t)elem->msl->page_sz;
47 data_start = RTE_PTR_ADD(elem, MALLOC_ELEM_HEADER_LEN);
48 data_end = RTE_PTR_ADD(elem, elem->size - MALLOC_ELEM_TRAILER_LEN);
49 /* segment must start after header and with specified alignment */
50 contig_seg_start = RTE_PTR_ALIGN_CEIL(data_start, align);
52 /* return if aligned address is already out of malloc element */
53 if (contig_seg_start > data_end)
56 /* if we're in IOVA as VA mode, or if we're in legacy mode with
57 * hugepages, all elements are IOVA-contiguous. however, we can only
58 * make these assumptions about internal memory - externally allocated
59 * segments have to be checked.
61 if (!elem->msl->external &&
62 (rte_eal_iova_mode() == RTE_IOVA_VA ||
63 (internal_config.legacy_mem &&
64 rte_eal_has_hugepages())))
65 return RTE_PTR_DIFF(data_end, contig_seg_start);
67 cur_page = RTE_PTR_ALIGN_FLOOR(contig_seg_start, page_sz);
68 ms = rte_mem_virt2memseg(cur_page, elem->msl);
70 /* do first iteration outside the loop */
71 page_end = RTE_PTR_ADD(cur_page, page_sz);
72 cur_seg_end = RTE_MIN(page_end, data_end);
73 cur = RTE_PTR_DIFF(cur_seg_end, contig_seg_start) -
74 MALLOC_ELEM_TRAILER_LEN;
76 expected_iova = ms->iova + page_sz;
77 /* memsegs are contiguous in memory */
80 cur_page = RTE_PTR_ADD(cur_page, page_sz);
82 while (cur_page < data_end) {
83 page_end = RTE_PTR_ADD(cur_page, page_sz);
84 cur_seg_end = RTE_MIN(page_end, data_end);
86 /* reset start of contiguous segment if unexpected iova */
87 if (ms->iova != expected_iova) {
88 /* next contiguous segment must start at specified
91 contig_seg_start = RTE_PTR_ALIGN(cur_page, align);
92 /* new segment start may be on a different page, so find
93 * the page and skip to next iteration to make sure
94 * we're not blowing past data end.
96 ms = rte_mem_virt2memseg(contig_seg_start, elem->msl);
98 /* don't trigger another recalculation */
99 expected_iova = ms->iova;
102 /* cur_seg_end ends on a page boundary or on data end. if we're
103 * looking at data end, then malloc trailer is already included
104 * in the calculations. if we're looking at page end, then we
105 * know there's more data past this page and thus there's space
106 * for malloc element trailer, so don't count it here.
108 cur = RTE_PTR_DIFF(cur_seg_end, contig_seg_start);
109 /* update max if cur value is bigger */
113 /* move to next page */
115 expected_iova = ms->iova + page_sz;
116 /* memsegs are contiguous in memory */
124 * Initialize a general malloc_elem header structure
127 malloc_elem_init(struct malloc_elem *elem, struct malloc_heap *heap,
128 struct rte_memseg_list *msl, size_t size,
129 struct malloc_elem *orig_elem, size_t orig_size)
135 memset(&elem->free_list, 0, sizeof(elem->free_list));
136 elem->state = ELEM_FREE;
139 elem->orig_elem = orig_elem;
140 elem->orig_size = orig_size;
146 malloc_elem_insert(struct malloc_elem *elem)
148 struct malloc_elem *prev_elem, *next_elem;
149 struct malloc_heap *heap = elem->heap;
151 /* first and last elements must be both NULL or both non-NULL */
152 if ((heap->first == NULL) != (heap->last == NULL)) {
153 RTE_LOG(ERR, EAL, "Heap is probably corrupt\n");
157 if (heap->first == NULL && heap->last == NULL) {
163 } else if (elem < heap->first) {
164 /* if lower than start */
166 next_elem = heap->first;
168 } else if (elem > heap->last) {
169 /* if higher than end */
170 prev_elem = heap->last;
174 /* the new memory is somewhere inbetween start and end */
175 uint64_t dist_from_start, dist_from_end;
177 dist_from_end = RTE_PTR_DIFF(heap->last, elem);
178 dist_from_start = RTE_PTR_DIFF(elem, heap->first);
180 /* check which is closer, and find closest list entries */
181 if (dist_from_start < dist_from_end) {
182 prev_elem = heap->first;
183 while (prev_elem->next < elem)
184 prev_elem = prev_elem->next;
185 next_elem = prev_elem->next;
187 next_elem = heap->last;
188 while (next_elem->prev > elem)
189 next_elem = next_elem->prev;
190 prev_elem = next_elem->prev;
194 /* insert new element */
195 elem->prev = prev_elem;
196 elem->next = next_elem;
198 prev_elem->next = elem;
200 next_elem->prev = elem;
204 * Attempt to find enough physically contiguous memory in this block to store
205 * our data. Assume that element has at least enough space to fit in the data,
206 * so we just check the page addresses.
209 elem_check_phys_contig(const struct rte_memseg_list *msl,
210 void *start, size_t size)
212 return eal_memalloc_is_contig(msl, start, size);
216 * calculate the starting point of where data of the requested size
217 * and alignment would fit in the current element. If the data doesn't
221 elem_start_pt(struct malloc_elem *elem, size_t size, unsigned align,
222 size_t bound, bool contig)
224 size_t elem_size = elem->size;
227 * we're allocating from the end, so adjust the size of element by
230 while (elem_size >= size) {
231 const size_t bmask = ~(bound - 1);
232 uintptr_t end_pt = (uintptr_t)elem +
233 elem_size - MALLOC_ELEM_TRAILER_LEN;
234 uintptr_t new_data_start = RTE_ALIGN_FLOOR((end_pt - size),
236 uintptr_t new_elem_start;
239 if ((new_data_start & bmask) != ((end_pt - 1) & bmask)) {
240 end_pt = RTE_ALIGN_FLOOR(end_pt, bound);
241 new_data_start = RTE_ALIGN_FLOOR((end_pt - size),
243 end_pt = new_data_start + size;
245 if (((end_pt - 1) & bmask) != (new_data_start & bmask))
249 new_elem_start = new_data_start - MALLOC_ELEM_HEADER_LEN;
251 /* if the new start point is before the exist start,
254 if (new_elem_start < (uintptr_t)elem)
258 size_t new_data_size = end_pt - new_data_start;
261 * if physical contiguousness was requested and we
262 * couldn't fit all data into one physically contiguous
263 * block, try again with lower addresses.
265 if (!elem_check_phys_contig(elem->msl,
266 (void *)new_data_start,
272 return (void *)new_elem_start;
278 * use elem_start_pt to determine if we get meet the size and
279 * alignment request from the current element
282 malloc_elem_can_hold(struct malloc_elem *elem, size_t size, unsigned align,
283 size_t bound, bool contig)
285 return elem_start_pt(elem, size, align, bound, contig) != NULL;
289 * split an existing element into two smaller elements at the given
290 * split_pt parameter.
293 split_elem(struct malloc_elem *elem, struct malloc_elem *split_pt)
295 struct malloc_elem *next_elem = elem->next;
296 const size_t old_elem_size = (uintptr_t)split_pt - (uintptr_t)elem;
297 const size_t new_elem_size = elem->size - old_elem_size;
299 malloc_elem_init(split_pt, elem->heap, elem->msl, new_elem_size,
300 elem->orig_elem, elem->orig_size);
301 split_pt->prev = elem;
302 split_pt->next = next_elem;
304 next_elem->prev = split_pt;
306 elem->heap->last = split_pt;
307 elem->next = split_pt;
308 elem->size = old_elem_size;
311 /* Update inner padding inner element size. */
312 elem = RTE_PTR_ADD(elem, elem->pad);
313 elem->size = old_elem_size - elem->pad;
318 * our malloc heap is a doubly linked list, so doubly remove our element.
320 static void __rte_unused
321 remove_elem(struct malloc_elem *elem)
323 struct malloc_elem *next, *prev;
330 elem->heap->last = prev;
334 elem->heap->first = next;
341 next_elem_is_adjacent(struct malloc_elem *elem)
343 return elem->next == RTE_PTR_ADD(elem, elem->size) &&
344 elem->next->msl == elem->msl &&
345 (!internal_config.match_allocations ||
346 elem->orig_elem == elem->next->orig_elem);
350 prev_elem_is_adjacent(struct malloc_elem *elem)
352 return elem == RTE_PTR_ADD(elem->prev, elem->prev->size) &&
353 elem->prev->msl == elem->msl &&
354 (!internal_config.match_allocations ||
355 elem->orig_elem == elem->prev->orig_elem);
359 * Given an element size, compute its freelist index.
360 * We free an element into the freelist containing similarly-sized elements.
361 * We try to allocate elements starting with the freelist containing
362 * similarly-sized elements, and if necessary, we search freelists
363 * containing larger elements.
365 * Example element size ranges for a heap with five free lists:
366 * heap->free_head[0] - (0 , 2^8]
367 * heap->free_head[1] - (2^8 , 2^10]
368 * heap->free_head[2] - (2^10 ,2^12]
369 * heap->free_head[3] - (2^12, 2^14]
370 * heap->free_head[4] - (2^14, MAX_SIZE]
373 malloc_elem_free_list_index(size_t size)
375 #define MALLOC_MINSIZE_LOG2 8
376 #define MALLOC_LOG2_INCREMENT 2
381 if (size <= (1UL << MALLOC_MINSIZE_LOG2))
384 /* Find next power of 2 >= size. */
385 log2 = sizeof(size) * 8 - __builtin_clzl(size-1);
387 /* Compute freelist index, based on log2(size). */
388 index = (log2 - MALLOC_MINSIZE_LOG2 + MALLOC_LOG2_INCREMENT - 1) /
389 MALLOC_LOG2_INCREMENT;
391 return index <= RTE_HEAP_NUM_FREELISTS-1?
392 index: RTE_HEAP_NUM_FREELISTS-1;
396 * Add the specified element to its heap's free list.
399 malloc_elem_free_list_insert(struct malloc_elem *elem)
403 idx = malloc_elem_free_list_index(elem->size - MALLOC_ELEM_HEADER_LEN);
404 elem->state = ELEM_FREE;
405 LIST_INSERT_HEAD(&elem->heap->free_head[idx], elem, free_list);
409 * Remove the specified element from its heap's free list.
412 malloc_elem_free_list_remove(struct malloc_elem *elem)
414 LIST_REMOVE(elem, free_list);
418 * reserve a block of data in an existing malloc_elem. If the malloc_elem
419 * is much larger than the data block requested, we split the element in two.
420 * This function is only called from malloc_heap_alloc so parameter checking
421 * is not done here, as it's done there previously.
424 malloc_elem_alloc(struct malloc_elem *elem, size_t size, unsigned align,
425 size_t bound, bool contig)
427 struct malloc_elem *new_elem = elem_start_pt(elem, size, align, bound,
429 const size_t old_elem_size = (uintptr_t)new_elem - (uintptr_t)elem;
430 const size_t trailer_size = elem->size - old_elem_size - size -
431 MALLOC_ELEM_OVERHEAD;
433 malloc_elem_free_list_remove(elem);
435 if (trailer_size > MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
436 /* split it, too much free space after elem */
437 struct malloc_elem *new_free_elem =
438 RTE_PTR_ADD(new_elem, size + MALLOC_ELEM_OVERHEAD);
440 split_elem(elem, new_free_elem);
441 malloc_elem_free_list_insert(new_free_elem);
443 if (elem == elem->heap->last)
444 elem->heap->last = new_free_elem;
447 if (old_elem_size < MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
448 /* don't split it, pad the element instead */
449 elem->state = ELEM_BUSY;
450 elem->pad = old_elem_size;
452 /* put a dummy header in padding, to point to real element header */
453 if (elem->pad > 0) { /* pad will be at least 64-bytes, as everything
454 * is cache-line aligned */
455 new_elem->pad = elem->pad;
456 new_elem->state = ELEM_PAD;
457 new_elem->size = elem->size - elem->pad;
458 set_header(new_elem);
464 /* we are going to split the element in two. The original element
465 * remains free, and the new element is the one allocated.
466 * Re-insert original element, in case its new size makes it
467 * belong on a different list.
469 split_elem(elem, new_elem);
470 new_elem->state = ELEM_BUSY;
471 malloc_elem_free_list_insert(elem);
477 * join two struct malloc_elem together. elem1 and elem2 must
478 * be contiguous in memory.
481 join_elem(struct malloc_elem *elem1, struct malloc_elem *elem2)
483 struct malloc_elem *next = elem2->next;
484 elem1->size += elem2->size;
488 elem1->heap->last = elem1;
493 malloc_elem_join_adjacent_free(struct malloc_elem *elem)
496 * check if next element exists, is adjacent and is free, if so join
497 * with it, need to remove from free list.
499 if (elem->next != NULL && elem->next->state == ELEM_FREE &&
500 next_elem_is_adjacent(elem)) {
504 /* we will want to erase the trailer and header */
505 erase = RTE_PTR_SUB(elem->next, MALLOC_ELEM_TRAILER_LEN);
506 erase_len = MALLOC_ELEM_OVERHEAD + elem->next->pad;
508 /* remove from free list, join to this one */
509 malloc_elem_free_list_remove(elem->next);
510 join_elem(elem, elem->next);
512 /* erase header, trailer and pad */
513 memset(erase, MALLOC_POISON, erase_len);
517 * check if prev element exists, is adjacent and is free, if so join
518 * with it, need to remove from free list.
520 if (elem->prev != NULL && elem->prev->state == ELEM_FREE &&
521 prev_elem_is_adjacent(elem)) {
522 struct malloc_elem *new_elem;
526 /* we will want to erase trailer and header */
527 erase = RTE_PTR_SUB(elem, MALLOC_ELEM_TRAILER_LEN);
528 erase_len = MALLOC_ELEM_OVERHEAD + elem->pad;
530 /* remove from free list, join to this one */
531 malloc_elem_free_list_remove(elem->prev);
533 new_elem = elem->prev;
534 join_elem(new_elem, elem);
536 /* erase header, trailer and pad */
537 memset(erase, MALLOC_POISON, erase_len);
546 * free a malloc_elem block by adding it to the free list. If the
547 * blocks either immediately before or immediately after newly freed block
548 * are also free, the blocks are merged together.
551 malloc_elem_free(struct malloc_elem *elem)
556 ptr = RTE_PTR_ADD(elem, MALLOC_ELEM_HEADER_LEN);
557 data_len = elem->size - MALLOC_ELEM_OVERHEAD;
559 elem = malloc_elem_join_adjacent_free(elem);
561 malloc_elem_free_list_insert(elem);
565 /* decrease heap's count of allocated elements */
566 elem->heap->alloc_count--;
569 memset(ptr, MALLOC_POISON, data_len);
574 /* assume all checks were already done */
576 malloc_elem_hide_region(struct malloc_elem *elem, void *start, size_t len)
578 struct malloc_elem *hide_start, *hide_end, *prev, *next;
579 size_t len_before, len_after;
582 hide_end = RTE_PTR_ADD(start, len);
587 /* we cannot do anything with non-adjacent elements */
588 if (next && next_elem_is_adjacent(elem)) {
589 len_after = RTE_PTR_DIFF(next, hide_end);
590 if (len_after >= MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
592 split_elem(elem, hide_end);
594 malloc_elem_free_list_insert(hide_end);
595 } else if (len_after > 0) {
596 RTE_LOG(ERR, EAL, "Unaligned element, heap is probably corrupt\n");
601 /* we cannot do anything with non-adjacent elements */
602 if (prev && prev_elem_is_adjacent(elem)) {
603 len_before = RTE_PTR_DIFF(hide_start, elem);
604 if (len_before >= MALLOC_ELEM_OVERHEAD + MIN_DATA_SIZE) {
606 split_elem(elem, hide_start);
611 malloc_elem_free_list_insert(prev);
612 } else if (len_before > 0) {
613 RTE_LOG(ERR, EAL, "Unaligned element, heap is probably corrupt\n");
622 * attempt to resize a malloc_elem by expanding into any free space
623 * immediately after it in memory.
626 malloc_elem_resize(struct malloc_elem *elem, size_t size)
628 const size_t new_size = size + elem->pad + MALLOC_ELEM_OVERHEAD;
630 /* if we request a smaller size, then always return ok */
631 if (elem->size >= new_size)
634 /* check if there is a next element, it's free and adjacent */
635 if (!elem->next || elem->next->state != ELEM_FREE ||
636 !next_elem_is_adjacent(elem))
638 if (elem->size + elem->next->size < new_size)
641 /* we now know the element fits, so remove from free list,
644 malloc_elem_free_list_remove(elem->next);
645 join_elem(elem, elem->next);
647 if (elem->size - new_size >= MIN_DATA_SIZE + MALLOC_ELEM_OVERHEAD) {
648 /* now we have a big block together. Lets cut it down a bit, by splitting */
649 struct malloc_elem *split_pt = RTE_PTR_ADD(elem, new_size);
650 split_pt = RTE_PTR_ALIGN_CEIL(split_pt, RTE_CACHE_LINE_SIZE);
651 split_elem(elem, split_pt);
652 malloc_elem_free_list_insert(split_pt);
657 static inline const char *
658 elem_state_to_str(enum elem_state state)
672 malloc_elem_dump(const struct malloc_elem *elem, FILE *f)
674 fprintf(f, "Malloc element at %p (%s)\n", elem,
675 elem_state_to_str(elem->state));
676 fprintf(f, " len: 0x%zx pad: 0x%" PRIx32 "\n", elem->size, elem->pad);
677 fprintf(f, " prev: %p next: %p\n", elem->prev, elem->next);