1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2017 6WIND S.A.
3 * Copyright 2017 Mellanox Technologies, Ltd
8 * Memory management functions for mlx4 driver.
18 /* Verbs headers do not support -pedantic. */
20 #pragma GCC diagnostic ignored "-Wpedantic"
22 #include <infiniband/verbs.h>
24 #pragma GCC diagnostic error "-Wpedantic"
27 #include <rte_branch_prediction.h>
28 #include <rte_common.h>
29 #include <rte_errno.h>
30 #include <rte_malloc.h>
31 #include <rte_memory.h>
32 #include <rte_mempool.h>
33 #include <rte_rwlock.h>
35 #include "mlx4_glue.h"
37 #include "mlx4_rxtx.h"
38 #include "mlx4_utils.h"
40 struct mr_find_contig_memsegs_data {
44 const struct rte_memseg_list *msl;
47 struct mr_update_mp_data {
48 struct rte_eth_dev *dev;
49 struct mlx4_mr_ctrl *mr_ctrl;
54 * Expand B-tree table to a given size. Can't be called with holding
55 * memory_hotplug_lock or priv->mr.rwlock due to rte_realloc().
58 * Pointer to B-tree structure.
60 * Number of entries for expansion.
63 * 0 on success, -1 on failure.
66 mr_btree_expand(struct mlx4_mr_btree *bt, int n)
74 * Downside of directly using rte_realloc() is that SOCKET_ID_ANY is
75 * used inside if there's no room to expand. Because this is a quite
76 * rare case and a part of very slow path, it is very acceptable.
77 * Initially cache_bh[] will be given practically enough space and once
78 * it is expanded, expansion wouldn't be needed again ever.
80 mem = rte_realloc(bt->table, n * sizeof(struct mlx4_mr_cache), 0);
82 /* Not an error, B-tree search will be skipped. */
83 WARN("failed to expand MR B-tree (%p) table", (void *)bt);
86 DEBUG("expanded MR B-tree table (size=%u)", n);
94 * Look up LKey from given B-tree lookup table, store the last index and return
98 * Pointer to B-tree structure.
100 * Pointer to index. Even on search failure, returns index where it stops
101 * searching so that index can be used when inserting a new entry.
106 * Searched LKey on success, UINT32_MAX on no match.
109 mr_btree_lookup(struct mlx4_mr_btree *bt, uint16_t *idx, uintptr_t addr)
111 struct mlx4_mr_cache *lkp_tbl;
116 lkp_tbl = *bt->table;
118 /* First entry must be NULL for comparison. */
119 assert(bt->len > 0 || (lkp_tbl[0].start == 0 &&
120 lkp_tbl[0].lkey == UINT32_MAX));
123 register uint16_t delta = n >> 1;
125 if (addr < lkp_tbl[base + delta].start) {
132 assert(addr >= lkp_tbl[base].start);
134 if (addr < lkp_tbl[base].end)
135 return lkp_tbl[base].lkey;
141 * Insert an entry to B-tree lookup table.
144 * Pointer to B-tree structure.
146 * Pointer to new entry to insert.
149 * 0 on success, -1 on failure.
152 mr_btree_insert(struct mlx4_mr_btree *bt, struct mlx4_mr_cache *entry)
154 struct mlx4_mr_cache *lkp_tbl;
159 assert(bt->len <= bt->size);
161 lkp_tbl = *bt->table;
162 /* Find out the slot for insertion. */
163 if (mr_btree_lookup(bt, &idx, entry->start) != UINT32_MAX) {
164 DEBUG("abort insertion to B-tree(%p): already exist at"
165 " idx=%u [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
166 (void *)bt, idx, entry->start, entry->end, entry->lkey);
167 /* Already exist, return. */
170 /* If table is full, return error. */
171 if (unlikely(bt->len == bt->size)) {
177 shift = (bt->len - idx) * sizeof(struct mlx4_mr_cache);
179 memmove(&lkp_tbl[idx + 1], &lkp_tbl[idx], shift);
180 lkp_tbl[idx] = *entry;
182 DEBUG("inserted B-tree(%p)[%u],"
183 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
184 (void *)bt, idx, entry->start, entry->end, entry->lkey);
189 * Initialize B-tree and allocate memory for lookup table.
192 * Pointer to B-tree structure.
194 * Number of entries to allocate.
196 * NUMA socket on which memory must be allocated.
199 * 0 on success, a negative errno value otherwise and rte_errno is set.
202 mlx4_mr_btree_init(struct mlx4_mr_btree *bt, int n, int socket)
208 memset(bt, 0, sizeof(*bt));
209 bt->table = rte_calloc_socket("B-tree table",
210 n, sizeof(struct mlx4_mr_cache),
212 if (bt->table == NULL) {
214 ERROR("failed to allocate memory for btree cache on socket %d",
219 /* First entry must be NULL for binary search. */
220 (*bt->table)[bt->len++] = (struct mlx4_mr_cache) {
223 DEBUG("initialized B-tree %p with table %p",
224 (void *)bt, (void *)bt->table);
229 * Free B-tree resources.
232 * Pointer to B-tree structure.
235 mlx4_mr_btree_free(struct mlx4_mr_btree *bt)
239 DEBUG("freeing B-tree %p with table %p", (void *)bt, (void *)bt->table);
241 memset(bt, 0, sizeof(*bt));
246 * Dump all the entries in a B-tree
249 * Pointer to B-tree structure.
252 mlx4_mr_btree_dump(struct mlx4_mr_btree *bt)
255 struct mlx4_mr_cache *lkp_tbl;
259 lkp_tbl = *bt->table;
260 for (idx = 0; idx < bt->len; ++idx) {
261 struct mlx4_mr_cache *entry = &lkp_tbl[idx];
263 DEBUG("B-tree(%p)[%u],"
264 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
265 (void *)bt, idx, entry->start, entry->end, entry->lkey);
271 * Find virtually contiguous memory chunk in a given MR.
274 * Pointer to MR structure.
276 * Pointer to returning MR cache entry. If not found, this will not be
279 * Start index of the memseg bitmap.
282 * Next index to go on lookup.
285 mr_find_next_chunk(struct mlx4_mr *mr, struct mlx4_mr_cache *entry,
292 /* MR for external memory doesn't have memseg list. */
293 if (mr->msl == NULL) {
294 struct ibv_mr *ibv_mr = mr->ibv_mr;
296 assert(mr->ms_bmp_n == 1);
297 assert(mr->ms_n == 1);
298 assert(base_idx == 0);
300 * Can't search it from memseg list but get it directly from
301 * verbs MR as there's only one chunk.
303 entry->start = (uintptr_t)ibv_mr->addr;
304 entry->end = (uintptr_t)ibv_mr->addr + mr->ibv_mr->length;
305 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey);
306 /* Returning 1 ends iteration. */
309 for (idx = base_idx; idx < mr->ms_bmp_n; ++idx) {
310 if (rte_bitmap_get(mr->ms_bmp, idx)) {
311 const struct rte_memseg_list *msl;
312 const struct rte_memseg *ms;
315 ms = rte_fbarray_get(&msl->memseg_arr,
316 mr->ms_base_idx + idx);
317 assert(msl->page_sz == ms->hugepage_sz);
320 end = ms->addr_64 + ms->hugepage_sz;
322 /* Passed the end of a fragment. */
327 /* Found one chunk. */
328 entry->start = start;
330 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey);
336 * Insert a MR to the global B-tree cache. It may fail due to low-on-memory.
337 * Then, this entry will have to be searched by mr_lookup_dev_list() in
338 * mlx4_mr_create() on miss.
341 * Pointer to Ethernet device.
343 * Pointer to MR to insert.
346 * 0 on success, -1 on failure.
349 mr_insert_dev_cache(struct rte_eth_dev *dev, struct mlx4_mr *mr)
351 struct priv *priv = dev->data->dev_private;
354 DEBUG("port %u inserting MR(%p) to global cache",
355 dev->data->port_id, (void *)mr);
356 for (n = 0; n < mr->ms_bmp_n; ) {
357 struct mlx4_mr_cache entry;
359 memset(&entry, 0, sizeof(entry));
360 /* Find a contiguous chunk and advance the index. */
361 n = mr_find_next_chunk(mr, &entry, n);
364 if (mr_btree_insert(&priv->mr.cache, &entry) < 0) {
366 * Overflowed, but the global table cannot be expanded
367 * because of deadlock.
376 * Look up address in the original global MR list.
379 * Pointer to Ethernet device.
381 * Pointer to returning MR cache entry. If no match, this will not be updated.
386 * Found MR on match, NULL otherwise.
388 static struct mlx4_mr *
389 mr_lookup_dev_list(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
392 struct priv *priv = dev->data->dev_private;
395 /* Iterate all the existing MRs. */
396 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
401 for (n = 0; n < mr->ms_bmp_n; ) {
402 struct mlx4_mr_cache ret;
404 memset(&ret, 0, sizeof(ret));
405 n = mr_find_next_chunk(mr, &ret, n);
406 if (addr >= ret.start && addr < ret.end) {
417 * Look up address on device.
420 * Pointer to Ethernet device.
422 * Pointer to returning MR cache entry. If no match, this will not be updated.
427 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
430 mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
433 struct priv *priv = dev->data->dev_private;
435 uint32_t lkey = UINT32_MAX;
439 * If the global cache has overflowed since it failed to expand the
440 * B-tree table, it can't have all the existing MRs. Then, the address
441 * has to be searched by traversing the original MR list instead, which
442 * is very slow path. Otherwise, the global cache is all inclusive.
444 if (!unlikely(priv->mr.cache.overflow)) {
445 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
446 if (lkey != UINT32_MAX)
447 *entry = (*priv->mr.cache.table)[idx];
449 /* Falling back to the slowest path. */
450 mr = mr_lookup_dev_list(dev, entry, addr);
454 assert(lkey == UINT32_MAX || (addr >= entry->start &&
460 * Free MR resources. MR lock must not be held to avoid a deadlock. rte_free()
461 * can raise memory free event and the callback function will spin on the lock.
464 * Pointer to MR to free.
467 mr_free(struct mlx4_mr *mr)
471 DEBUG("freeing MR(%p):", (void *)mr);
472 if (mr->ibv_mr != NULL)
473 claim_zero(mlx4_glue->dereg_mr(mr->ibv_mr));
474 if (mr->ms_bmp != NULL)
475 rte_bitmap_free(mr->ms_bmp);
480 * Releass resources of detached MR having no online entry.
483 * Pointer to Ethernet device.
486 mlx4_mr_garbage_collect(struct rte_eth_dev *dev)
488 struct priv *priv = dev->data->dev_private;
489 struct mlx4_mr *mr_next;
490 struct mlx4_mr_list free_list = LIST_HEAD_INITIALIZER(free_list);
493 * MR can't be freed with holding the lock because rte_free() could call
494 * memory free callback function. This will be a deadlock situation.
496 rte_rwlock_write_lock(&priv->mr.rwlock);
497 /* Detach the whole free list and release it after unlocking. */
498 free_list = priv->mr.mr_free_list;
499 LIST_INIT(&priv->mr.mr_free_list);
500 rte_rwlock_write_unlock(&priv->mr.rwlock);
501 /* Release resources. */
502 mr_next = LIST_FIRST(&free_list);
503 while (mr_next != NULL) {
504 struct mlx4_mr *mr = mr_next;
506 mr_next = LIST_NEXT(mr, mr);
511 /* Called during rte_memseg_contig_walk() by mlx4_mr_create(). */
513 mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl,
514 const struct rte_memseg *ms, size_t len, void *arg)
516 struct mr_find_contig_memsegs_data *data = arg;
518 if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len)
520 /* Found, save it and stop walking. */
521 data->start = ms->addr_64;
522 data->end = ms->addr_64 + len;
528 * Create a new global Memroy Region (MR) for a missing virtual address.
529 * Register entire virtually contiguous memory chunk around the address.
532 * Pointer to Ethernet device.
534 * Pointer to returning MR cache entry, found in the global cache or newly
535 * created. If failed to create one, this will not be updated.
537 * Target virtual address to register.
540 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
543 mlx4_mr_create(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
546 struct priv *priv = dev->data->dev_private;
547 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
548 const struct rte_memseg_list *msl;
549 const struct rte_memseg *ms;
550 struct mlx4_mr *mr = NULL;
555 int ms_idx_shift = -1;
557 struct mr_find_contig_memsegs_data data = {
560 struct mr_find_contig_memsegs_data data_re;
562 DEBUG("port %u creating a MR using address (%p)",
563 dev->data->port_id, (void *)addr);
565 * Release detached MRs if any. This can't be called with holding either
566 * memory_hotplug_lock or priv->mr.rwlock. MRs on the free list have
567 * been detached by the memory free event but it couldn't be released
568 * inside the callback due to deadlock. As a result, releasing resources
569 * is quite opportunistic.
571 mlx4_mr_garbage_collect(dev);
573 * Find out a contiguous virtual address chunk in use, to which the
574 * given address belongs, in order to register maximum range. In the
575 * best case where mempools are not dynamically recreated and
576 * '--socket-mem' is specified as an EAL option, it is very likely to
577 * have only one MR(LKey) per a socket and per a hugepage-size even
578 * though the system memory is highly fragmented.
580 if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) {
581 WARN("port %u unable to find virtually contiguous"
582 " chunk for address (%p)."
583 " rte_memseg_contig_walk() failed.",
584 dev->data->port_id, (void *)addr);
589 /* Addresses must be page-aligned. */
590 assert(rte_is_aligned((void *)data.start, data.msl->page_sz));
591 assert(rte_is_aligned((void *)data.end, data.msl->page_sz));
593 ms = rte_mem_virt2memseg((void *)data.start, msl);
594 len = data.end - data.start;
595 assert(msl->page_sz == ms->hugepage_sz);
596 /* Number of memsegs in the range. */
597 ms_n = len / msl->page_sz;
598 DEBUG("port %u extending %p to [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
599 " page_sz=0x%" PRIx64 ", ms_n=%u",
600 dev->data->port_id, (void *)addr,
601 data.start, data.end, msl->page_sz, ms_n);
602 /* Size of memory for bitmap. */
603 bmp_size = rte_bitmap_get_memory_footprint(ms_n);
604 mr = rte_zmalloc_socket(NULL,
605 RTE_ALIGN_CEIL(sizeof(*mr),
606 RTE_CACHE_LINE_SIZE) +
608 RTE_CACHE_LINE_SIZE, msl->socket_id);
610 WARN("port %u unable to allocate memory for a new MR of"
612 dev->data->port_id, (void *)addr);
618 * Save the index of the first memseg and initialize memseg bitmap. To
619 * see if a memseg of ms_idx in the memseg-list is still valid, check:
620 * rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx)
622 mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
623 bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE);
624 mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size);
625 if (mr->ms_bmp == NULL) {
626 WARN("port %u unable to initialize bitamp for a new MR of"
628 dev->data->port_id, (void *)addr);
633 * Should recheck whether the extended contiguous chunk is still valid.
634 * Because memory_hotplug_lock can't be held if there's any memory
635 * related calls in a critical path, resource allocation above can't be
636 * locked. If the memory has been changed at this point, try again with
637 * just single page. If not, go on with the big chunk atomically from
640 rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
642 if (len > msl->page_sz &&
643 !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) {
644 WARN("port %u unable to find virtually contiguous"
645 " chunk for address (%p)."
646 " rte_memseg_contig_walk() failed.",
647 dev->data->port_id, (void *)addr);
651 if (data.start != data_re.start || data.end != data_re.end) {
653 * The extended contiguous chunk has been changed. Try again
654 * with single memseg instead.
656 data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz);
657 data.end = data.start + msl->page_sz;
658 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
660 goto alloc_resources;
662 assert(data.msl == data_re.msl);
663 rte_rwlock_write_lock(&priv->mr.rwlock);
665 * Check the address is really missing. If other thread already created
666 * one or it is not found due to overflow, abort and return.
668 if (mr_lookup_dev(dev, entry, addr) != UINT32_MAX) {
670 * Insert to the global cache table. It may fail due to
671 * low-on-memory. Then, this entry will have to be searched
674 mr_btree_insert(&priv->mr.cache, entry);
675 DEBUG("port %u found MR for %p on final lookup, abort",
676 dev->data->port_id, (void *)addr);
677 rte_rwlock_write_unlock(&priv->mr.rwlock);
678 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
680 * Must be unlocked before calling rte_free() because
681 * mlx4_mr_mem_event_free_cb() can be called inside.
687 * Trim start and end addresses for verbs MR. Set bits for registering
688 * memsegs but exclude already registered ones. Bitmap can be
691 for (n = 0; n < ms_n; ++n) {
693 struct mlx4_mr_cache ret;
695 memset(&ret, 0, sizeof(ret));
696 start = data_re.start + n * msl->page_sz;
697 /* Exclude memsegs already registered by other MRs. */
698 if (mr_lookup_dev(dev, &ret, start) == UINT32_MAX) {
700 * Start from the first unregistered memseg in the
703 if (ms_idx_shift == -1) {
704 mr->ms_base_idx += n;
708 data.end = start + msl->page_sz;
709 rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift);
713 len = data.end - data.start;
714 mr->ms_bmp_n = len / msl->page_sz;
715 assert(ms_idx_shift + mr->ms_bmp_n <= ms_n);
717 * Finally create a verbs MR for the memory chunk. ibv_reg_mr() can be
718 * called with holding the memory lock because it doesn't use
719 * mlx4_alloc_buf_extern() which eventually calls rte_malloc_socket()
720 * through mlx4_alloc_verbs_buf().
722 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)data.start, len,
723 IBV_ACCESS_LOCAL_WRITE);
724 if (mr->ibv_mr == NULL) {
725 WARN("port %u fail to create a verbs MR for address (%p)",
726 dev->data->port_id, (void *)addr);
730 assert((uintptr_t)mr->ibv_mr->addr == data.start);
731 assert(mr->ibv_mr->length == len);
732 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
733 DEBUG("port %u MR CREATED (%p) for %p:\n"
734 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
735 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
736 dev->data->port_id, (void *)mr, (void *)addr,
737 data.start, data.end, rte_cpu_to_be_32(mr->ibv_mr->lkey),
738 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
739 /* Insert to the global cache table. */
740 mr_insert_dev_cache(dev, mr);
741 /* Fill in output data. */
742 mr_lookup_dev(dev, entry, addr);
743 /* Lookup can't fail. */
744 assert(entry->lkey != UINT32_MAX);
745 rte_rwlock_write_unlock(&priv->mr.rwlock);
746 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
749 rte_rwlock_write_unlock(&priv->mr.rwlock);
751 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
754 * In case of error, as this can be called in a datapath, a warning
755 * message per an error is preferable instead. Must be unlocked before
756 * calling rte_free() because mlx4_mr_mem_event_free_cb() can be called
764 * Rebuild the global B-tree cache of device from the original MR list.
767 * Pointer to Ethernet device.
770 mr_rebuild_dev_cache(struct rte_eth_dev *dev)
772 struct priv *priv = dev->data->dev_private;
775 DEBUG("port %u rebuild dev cache[]", dev->data->port_id);
776 /* Flush cache to rebuild. */
777 priv->mr.cache.len = 1;
778 priv->mr.cache.overflow = 0;
779 /* Iterate all the existing MRs. */
780 LIST_FOREACH(mr, &priv->mr.mr_list, mr)
781 if (mr_insert_dev_cache(dev, mr) < 0)
786 * Callback for memory free event. Iterate freed memsegs and check whether it
787 * belongs to an existing MR. If found, clear the bit from bitmap of MR. As a
788 * result, the MR would be fragmented. If it becomes empty, the MR will be freed
789 * later by mlx4_mr_garbage_collect().
791 * The global cache must be rebuilt if there's any change and this event has to
792 * be propagated to dataplane threads to flush the local caches.
795 * Pointer to Ethernet device.
797 * Address of freed memory.
799 * Size of freed memory.
802 mlx4_mr_mem_event_free_cb(struct rte_eth_dev *dev, const void *addr, size_t len)
804 struct priv *priv = dev->data->dev_private;
805 const struct rte_memseg_list *msl;
811 DEBUG("port %u free callback: addr=%p, len=%zu",
812 dev->data->port_id, addr, len);
813 msl = rte_mem_virt2memseg_list(addr);
814 /* addr and len must be page-aligned. */
815 assert((uintptr_t)addr == RTE_ALIGN((uintptr_t)addr, msl->page_sz));
816 assert(len == RTE_ALIGN(len, msl->page_sz));
817 ms_n = len / msl->page_sz;
818 rte_rwlock_write_lock(&priv->mr.rwlock);
819 /* Clear bits of freed memsegs from MR. */
820 for (i = 0; i < ms_n; ++i) {
821 const struct rte_memseg *ms;
822 struct mlx4_mr_cache entry;
827 /* Find MR having this memseg. */
828 start = (uintptr_t)addr + i * msl->page_sz;
829 mr = mr_lookup_dev_list(dev, &entry, start);
832 assert(mr->msl); /* Can't be external memory. */
833 ms = rte_mem_virt2memseg((void *)start, msl);
835 assert(msl->page_sz == ms->hugepage_sz);
836 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
837 pos = ms_idx - mr->ms_base_idx;
838 assert(rte_bitmap_get(mr->ms_bmp, pos));
839 assert(pos < mr->ms_bmp_n);
840 DEBUG("port %u MR(%p): clear bitmap[%u] for addr %p",
841 dev->data->port_id, (void *)mr, pos, (void *)start);
842 rte_bitmap_clear(mr->ms_bmp, pos);
843 if (--mr->ms_n == 0) {
845 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
846 DEBUG("port %u remove MR(%p) from list",
847 dev->data->port_id, (void *)mr);
850 * MR is fragmented or will be freed. the global cache must be
856 mr_rebuild_dev_cache(dev);
858 * Flush local caches by propagating invalidation across cores.
859 * rte_smp_wmb() is enough to synchronize this event. If one of
860 * freed memsegs is seen by other core, that means the memseg
861 * has been allocated by allocator, which will come after this
862 * free call. Therefore, this store instruction (incrementing
863 * generation below) will be guaranteed to be seen by other core
864 * before the core sees the newly allocated memory.
867 DEBUG("broadcasting local cache flush, gen=%d",
871 rte_rwlock_write_unlock(&priv->mr.rwlock);
874 mlx4_mr_dump_dev(dev);
879 * Callback for memory event.
889 mlx4_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
890 size_t len, void *arg __rte_unused)
894 switch (event_type) {
895 case RTE_MEM_EVENT_FREE:
896 rte_rwlock_read_lock(&mlx4_mem_event_rwlock);
897 /* Iterate all the existing mlx4 devices. */
898 LIST_FOREACH(priv, &mlx4_mem_event_cb_list, mem_event_cb)
899 mlx4_mr_mem_event_free_cb(priv->dev, addr, len);
900 rte_rwlock_read_unlock(&mlx4_mem_event_rwlock);
902 case RTE_MEM_EVENT_ALLOC:
909 * Look up address in the global MR cache table. If not found, create a new MR.
910 * Insert the found/created entry to local bottom-half cache table.
913 * Pointer to Ethernet device.
915 * Pointer to per-queue MR control structure.
917 * Pointer to returning MR cache entry, found in the global cache or newly
918 * created. If failed to create one, this is not written.
923 * Searched LKey on success, UINT32_MAX on no match.
926 mlx4_mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
927 struct mlx4_mr_cache *entry, uintptr_t addr)
929 struct priv *priv = dev->data->dev_private;
930 struct mlx4_mr_btree *bt = &mr_ctrl->cache_bh;
934 /* If local cache table is full, try to double it. */
935 if (unlikely(bt->len == bt->size))
936 mr_btree_expand(bt, bt->size << 1);
937 /* Look up in the global cache. */
938 rte_rwlock_read_lock(&priv->mr.rwlock);
939 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
940 if (lkey != UINT32_MAX) {
942 *entry = (*priv->mr.cache.table)[idx];
943 rte_rwlock_read_unlock(&priv->mr.rwlock);
945 * Update local cache. Even if it fails, return the found entry
946 * to update top-half cache. Next time, this entry will be found
947 * in the global cache.
949 mr_btree_insert(bt, entry);
952 rte_rwlock_read_unlock(&priv->mr.rwlock);
953 /* First time to see the address? Create a new MR. */
954 lkey = mlx4_mr_create(dev, entry, addr);
956 * Update the local cache if successfully created a new global MR. Even
957 * if failed to create one, there's no action to take in this datapath
958 * code. As returning LKey is invalid, this will eventually make HW
961 if (lkey != UINT32_MAX)
962 mr_btree_insert(bt, entry);
967 * Bottom-half of LKey search on datapath. Firstly search in cache_bh[] and if
968 * misses, search in the global MR cache table and update the new entry to
969 * per-queue local caches.
972 * Pointer to Ethernet device.
974 * Pointer to per-queue MR control structure.
979 * Searched LKey on success, UINT32_MAX on no match.
982 mlx4_mr_addr2mr_bh(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
987 /* Victim in top-half cache to replace with new entry. */
988 struct mlx4_mr_cache *repl = &mr_ctrl->cache[mr_ctrl->head];
990 /* Binary-search MR translation table. */
991 lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr);
992 /* Update top-half cache. */
993 if (likely(lkey != UINT32_MAX)) {
994 *repl = (*mr_ctrl->cache_bh.table)[bh_idx];
997 * If missed in local lookup table, search in the global cache
998 * and local cache_bh[] will be updated inside if possible.
999 * Top-half cache entry will also be updated.
1001 lkey = mlx4_mr_lookup_dev(dev, mr_ctrl, repl, addr);
1002 if (unlikely(lkey == UINT32_MAX))
1005 /* Update the most recently used entry. */
1006 mr_ctrl->mru = mr_ctrl->head;
1007 /* Point to the next victim, the oldest. */
1008 mr_ctrl->head = (mr_ctrl->head + 1) % MLX4_MR_CACHE_N;
1013 * Bottom-half of LKey search on Rx.
1016 * Pointer to Rx queue structure.
1021 * Searched LKey on success, UINT32_MAX on no match.
1024 mlx4_rx_addr2mr_bh(struct rxq *rxq, uintptr_t addr)
1026 struct mlx4_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
1027 struct priv *priv = rxq->priv;
1029 DEBUG("Rx queue %u: miss on top-half, mru=%u, head=%u, addr=%p",
1030 rxq->stats.idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr);
1031 return mlx4_mr_addr2mr_bh(priv->dev, mr_ctrl, addr);
1035 * Bottom-half of LKey search on Tx.
1038 * Pointer to Tx queue structure.
1043 * Searched LKey on success, UINT32_MAX on no match.
1046 mlx4_tx_addr2mr_bh(struct txq *txq, uintptr_t addr)
1048 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1049 struct priv *priv = txq->priv;
1051 DEBUG("Tx queue %u: miss on top-half, mru=%u, head=%u, addr=%p",
1052 txq->stats.idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr);
1053 return mlx4_mr_addr2mr_bh(priv->dev, mr_ctrl, addr);
1057 * Bottom-half of LKey search on Tx. If it can't be searched in the memseg
1058 * list, register the mempool of the mbuf as externally allocated memory.
1061 * Pointer to Tx queue structure.
1066 * Searched LKey on success, UINT32_MAX on no match.
1069 mlx4_tx_mb2mr_bh(struct txq *txq, struct rte_mbuf *mb)
1071 uintptr_t addr = (uintptr_t)mb->buf_addr;
1074 lkey = mlx4_tx_addr2mr_bh(txq, addr);
1075 if (lkey == UINT32_MAX && rte_errno == ENXIO) {
1076 /* Mempool may have externally allocated memory. */
1077 return mlx4_tx_update_ext_mp(txq, addr, mlx4_mb2mp(mb));
1083 * Flush all of the local cache entries.
1086 * Pointer to per-queue MR control structure.
1089 mlx4_mr_flush_local_cache(struct mlx4_mr_ctrl *mr_ctrl)
1091 /* Reset the most-recently-used index. */
1093 /* Reset the linear search array. */
1095 memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache));
1096 /* Reset the B-tree table. */
1097 mr_ctrl->cache_bh.len = 1;
1098 mr_ctrl->cache_bh.overflow = 0;
1099 /* Update the generation number. */
1100 mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr;
1101 DEBUG("mr_ctrl(%p): flushed, cur_gen=%d",
1102 (void *)mr_ctrl, mr_ctrl->cur_gen);
1106 * Called during rte_mempool_mem_iter() by mlx4_mr_update_ext_mp().
1108 * Externally allocated chunk is registered and a MR is created for the chunk.
1109 * The MR object is added to the global list. If memseg list of a MR object
1110 * (mr->msl) is null, the MR object can be regarded as externally allocated
1113 * Once external memory is registered, it should be static. If the memory is
1114 * freed and the virtual address range has different physical memory mapped
1115 * again, it may cause crash on device due to the wrong translation entry. PMD
1116 * can't track the free event of the external memory for now.
1119 mlx4_mr_update_ext_mp_cb(struct rte_mempool *mp, void *opaque,
1120 struct rte_mempool_memhdr *memhdr,
1121 unsigned mem_idx __rte_unused)
1123 struct mr_update_mp_data *data = opaque;
1124 struct rte_eth_dev *dev = data->dev;
1125 struct priv *priv = dev->data->dev_private;
1126 struct mlx4_mr_ctrl *mr_ctrl = data->mr_ctrl;
1127 struct mlx4_mr *mr = NULL;
1128 uintptr_t addr = (uintptr_t)memhdr->addr;
1129 size_t len = memhdr->len;
1130 struct mlx4_mr_cache entry;
1133 /* If already registered, it should return. */
1134 rte_rwlock_read_lock(&priv->mr.rwlock);
1135 lkey = mr_lookup_dev(dev, &entry, addr);
1136 rte_rwlock_read_unlock(&priv->mr.rwlock);
1137 if (lkey != UINT32_MAX)
1139 mr = rte_zmalloc_socket(NULL,
1140 RTE_ALIGN_CEIL(sizeof(*mr),
1141 RTE_CACHE_LINE_SIZE),
1142 RTE_CACHE_LINE_SIZE, mp->socket_id);
1144 WARN("port %u unable to allocate memory for a new MR of"
1146 dev->data->port_id, mp->name);
1150 DEBUG("port %u register MR for chunk #%d of mempool (%s)",
1151 dev->data->port_id, mem_idx, mp->name);
1152 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)addr, len,
1153 IBV_ACCESS_LOCAL_WRITE);
1154 if (mr->ibv_mr == NULL) {
1155 WARN("port %u fail to create a verbs MR for address (%p)",
1156 dev->data->port_id, (void *)addr);
1161 mr->msl = NULL; /* Mark it is external memory. */
1165 rte_rwlock_write_lock(&priv->mr.rwlock);
1166 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
1167 DEBUG("port %u MR CREATED (%p) for external memory %p:\n"
1168 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
1169 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
1170 dev->data->port_id, (void *)mr, (void *)addr,
1171 addr, addr + len, rte_cpu_to_be_32(mr->ibv_mr->lkey),
1172 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
1173 /* Insert to the global cache table. */
1174 mr_insert_dev_cache(dev, mr);
1175 rte_rwlock_write_unlock(&priv->mr.rwlock);
1176 /* Insert to the local cache table */
1177 mlx4_mr_addr2mr_bh(dev, mr_ctrl, addr);
1181 * Register MR for entire memory chunks in a Mempool having externally allocated
1182 * memory and fill in local cache.
1185 * Pointer to Ethernet device.
1187 * Pointer to per-queue MR control structure.
1189 * Pointer to registering Mempool.
1192 * 0 on success, -1 on failure.
1195 mlx4_mr_update_ext_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1196 struct rte_mempool *mp)
1198 struct mr_update_mp_data data = {
1204 rte_mempool_mem_iter(mp, mlx4_mr_update_ext_mp_cb, &data);
1209 * Register MR entire memory chunks in a Mempool having externally allocated
1210 * memory and search LKey of the address to return.
1213 * Pointer to Ethernet device.
1217 * Pointer to registering Mempool where addr belongs.
1220 * LKey for address on success, UINT32_MAX on failure.
1223 mlx4_tx_update_ext_mp(struct txq *txq, uintptr_t addr, struct rte_mempool *mp)
1225 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1226 struct priv *priv = txq->priv;
1228 mlx4_mr_update_ext_mp(priv->dev, mr_ctrl, mp);
1229 return mlx4_tx_addr2mr_bh(txq, addr);
1232 /* Called during rte_mempool_mem_iter() by mlx4_mr_update_mp(). */
1234 mlx4_mr_update_mp_cb(struct rte_mempool *mp __rte_unused, void *opaque,
1235 struct rte_mempool_memhdr *memhdr,
1236 unsigned mem_idx __rte_unused)
1238 struct mr_update_mp_data *data = opaque;
1241 /* Stop iteration if failed in the previous walk. */
1244 /* Register address of the chunk and update local caches. */
1245 lkey = mlx4_mr_addr2mr_bh(data->dev, data->mr_ctrl,
1246 (uintptr_t)memhdr->addr);
1247 if (lkey == UINT32_MAX)
1252 * Register entire memory chunks in a Mempool.
1255 * Pointer to Ethernet device.
1257 * Pointer to per-queue MR control structure.
1259 * Pointer to registering Mempool.
1262 * 0 on success, -1 on failure.
1265 mlx4_mr_update_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1266 struct rte_mempool *mp)
1268 struct mr_update_mp_data data = {
1274 rte_mempool_mem_iter(mp, mlx4_mr_update_mp_cb, &data);
1275 if (data.ret < 0 && rte_errno == ENXIO) {
1276 /* Mempool may have externally allocated memory. */
1277 return mlx4_mr_update_ext_mp(dev, mr_ctrl, mp);
1284 * Dump all the created MRs and the global cache entries.
1287 * Pointer to Ethernet device.
1290 mlx4_mr_dump_dev(struct rte_eth_dev *dev)
1292 struct priv *priv = dev->data->dev_private;
1297 rte_rwlock_read_lock(&priv->mr.rwlock);
1298 /* Iterate all the existing MRs. */
1299 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
1302 DEBUG("port %u MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u",
1303 dev->data->port_id, mr_n++,
1304 rte_cpu_to_be_32(mr->ibv_mr->lkey),
1305 mr->ms_n, mr->ms_bmp_n);
1308 for (n = 0; n < mr->ms_bmp_n; ) {
1309 struct mlx4_mr_cache ret;
1311 memset(&ret, 0, sizeof(ret));
1312 n = mr_find_next_chunk(mr, &ret, n);
1315 DEBUG(" chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")",
1316 chunk_n++, ret.start, ret.end);
1319 DEBUG("port %u dumping global cache", dev->data->port_id);
1320 mlx4_mr_btree_dump(&priv->mr.cache);
1321 rte_rwlock_read_unlock(&priv->mr.rwlock);
1326 * Release all the created MRs and resources. Remove device from memory callback
1330 * Pointer to Ethernet device.
1333 mlx4_mr_release(struct rte_eth_dev *dev)
1335 struct priv *priv = dev->data->dev_private;
1336 struct mlx4_mr *mr_next = LIST_FIRST(&priv->mr.mr_list);
1338 /* Remove from memory callback device list. */
1339 rte_rwlock_write_lock(&mlx4_mem_event_rwlock);
1340 LIST_REMOVE(priv, mem_event_cb);
1341 rte_rwlock_write_unlock(&mlx4_mem_event_rwlock);
1343 mlx4_mr_dump_dev(dev);
1345 rte_rwlock_write_lock(&priv->mr.rwlock);
1346 /* Detach from MR list and move to free list. */
1347 while (mr_next != NULL) {
1348 struct mlx4_mr *mr = mr_next;
1350 mr_next = LIST_NEXT(mr, mr);
1351 LIST_REMOVE(mr, mr);
1352 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
1354 LIST_INIT(&priv->mr.mr_list);
1355 /* Free global cache. */
1356 mlx4_mr_btree_free(&priv->mr.cache);
1357 rte_rwlock_write_unlock(&priv->mr.rwlock);
1358 /* Free all remaining MRs. */
1359 mlx4_mr_garbage_collect(dev);