1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2016 6WIND S.A.
3 * Copyright 2016 Mellanox Technologies, Ltd
7 #pragma GCC diagnostic ignored "-Wpedantic"
9 #include <infiniband/verbs.h>
11 #pragma GCC diagnostic error "-Wpedantic"
14 #include <rte_mempool.h>
15 #include <rte_malloc.h>
16 #include <rte_rwlock.h>
20 #include "mlx5_rxtx.h"
21 #include "mlx5_glue.h"
23 struct mr_find_contig_memsegs_data {
27 const struct rte_memseg_list *msl;
30 struct mr_update_mp_data {
31 struct rte_eth_dev *dev;
32 struct mlx5_mr_ctrl *mr_ctrl;
37 * Expand B-tree table to a given size. Can't be called with holding
38 * memory_hotplug_lock or priv->mr.rwlock due to rte_realloc().
41 * Pointer to B-tree structure.
43 * Number of entries for expansion.
46 * 0 on success, -1 on failure.
49 mr_btree_expand(struct mlx5_mr_btree *bt, int n)
57 * Downside of directly using rte_realloc() is that SOCKET_ID_ANY is
58 * used inside if there's no room to expand. Because this is a quite
59 * rare case and a part of very slow path, it is very acceptable.
60 * Initially cache_bh[] will be given practically enough space and once
61 * it is expanded, expansion wouldn't be needed again ever.
63 mem = rte_realloc(bt->table, n * sizeof(struct mlx5_mr_cache), 0);
65 /* Not an error, B-tree search will be skipped. */
66 DRV_LOG(WARNING, "failed to expand MR B-tree (%p) table",
70 DRV_LOG(DEBUG, "expanded MR B-tree table (size=%u)", n);
78 * Look up LKey from given B-tree lookup table, store the last index and return
82 * Pointer to B-tree structure.
84 * Pointer to index. Even on search failure, returns index where it stops
85 * searching so that index can be used when inserting a new entry.
90 * Searched LKey on success, UINT32_MAX on no match.
93 mr_btree_lookup(struct mlx5_mr_btree *bt, uint16_t *idx, uintptr_t addr)
95 struct mlx5_mr_cache *lkp_tbl;
100 lkp_tbl = *bt->table;
102 /* First entry must be NULL for comparison. */
103 assert(bt->len > 0 || (lkp_tbl[0].start == 0 &&
104 lkp_tbl[0].lkey == UINT32_MAX));
107 register uint16_t delta = n >> 1;
109 if (addr < lkp_tbl[base + delta].start) {
116 assert(addr >= lkp_tbl[base].start);
118 if (addr < lkp_tbl[base].end)
119 return lkp_tbl[base].lkey;
125 * Insert an entry to B-tree lookup table.
128 * Pointer to B-tree structure.
130 * Pointer to new entry to insert.
133 * 0 on success, -1 on failure.
136 mr_btree_insert(struct mlx5_mr_btree *bt, struct mlx5_mr_cache *entry)
138 struct mlx5_mr_cache *lkp_tbl;
143 assert(bt->len <= bt->size);
145 lkp_tbl = *bt->table;
146 /* Find out the slot for insertion. */
147 if (mr_btree_lookup(bt, &idx, entry->start) != UINT32_MAX) {
149 "abort insertion to B-tree(%p): already exist at"
150 " idx=%u [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
151 (void *)bt, idx, entry->start, entry->end, entry->lkey);
152 /* Already exist, return. */
155 /* If table is full, return error. */
156 if (unlikely(bt->len == bt->size)) {
162 shift = (bt->len - idx) * sizeof(struct mlx5_mr_cache);
164 memmove(&lkp_tbl[idx + 1], &lkp_tbl[idx], shift);
165 lkp_tbl[idx] = *entry;
168 "inserted B-tree(%p)[%u],"
169 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
170 (void *)bt, idx, entry->start, entry->end, entry->lkey);
175 * Initialize B-tree and allocate memory for lookup table.
178 * Pointer to B-tree structure.
180 * Number of entries to allocate.
182 * NUMA socket on which memory must be allocated.
185 * 0 on success, a negative errno value otherwise and rte_errno is set.
188 mlx5_mr_btree_init(struct mlx5_mr_btree *bt, int n, int socket)
194 assert(!bt->table && !bt->size);
195 memset(bt, 0, sizeof(*bt));
196 bt->table = rte_calloc_socket("B-tree table",
197 n, sizeof(struct mlx5_mr_cache),
199 if (bt->table == NULL) {
201 DEBUG("failed to allocate memory for btree cache on socket %d",
206 /* First entry must be NULL for binary search. */
207 (*bt->table)[bt->len++] = (struct mlx5_mr_cache) {
210 DEBUG("initialized B-tree %p with table %p",
211 (void *)bt, (void *)bt->table);
216 * Free B-tree resources.
219 * Pointer to B-tree structure.
222 mlx5_mr_btree_free(struct mlx5_mr_btree *bt)
226 DEBUG("freeing B-tree %p with table %p",
227 (void *)bt, (void *)bt->table);
229 memset(bt, 0, sizeof(*bt));
233 * Dump all the entries in a B-tree
236 * Pointer to B-tree structure.
239 mlx5_mr_btree_dump(struct mlx5_mr_btree *bt __rte_unused)
243 struct mlx5_mr_cache *lkp_tbl;
247 lkp_tbl = *bt->table;
248 for (idx = 0; idx < bt->len; ++idx) {
249 struct mlx5_mr_cache *entry = &lkp_tbl[idx];
251 DEBUG("B-tree(%p)[%u],"
252 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
253 (void *)bt, idx, entry->start, entry->end, entry->lkey);
259 * Find virtually contiguous memory chunk in a given MR.
262 * Pointer to MR structure.
264 * Pointer to returning MR cache entry. If not found, this will not be
267 * Start index of the memseg bitmap.
270 * Next index to go on lookup.
273 mr_find_next_chunk(struct mlx5_mr *mr, struct mlx5_mr_cache *entry,
280 /* MR for external memory doesn't have memseg list. */
281 if (mr->msl == NULL) {
282 struct ibv_mr *ibv_mr = mr->ibv_mr;
284 assert(mr->ms_bmp_n == 1);
285 assert(mr->ms_n == 1);
286 assert(base_idx == 0);
288 * Can't search it from memseg list but get it directly from
289 * verbs MR as there's only one chunk.
291 entry->start = (uintptr_t)ibv_mr->addr;
292 entry->end = (uintptr_t)ibv_mr->addr + mr->ibv_mr->length;
293 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey);
294 /* Returning 1 ends iteration. */
297 for (idx = base_idx; idx < mr->ms_bmp_n; ++idx) {
298 if (rte_bitmap_get(mr->ms_bmp, idx)) {
299 const struct rte_memseg_list *msl;
300 const struct rte_memseg *ms;
303 ms = rte_fbarray_get(&msl->memseg_arr,
304 mr->ms_base_idx + idx);
305 assert(msl->page_sz == ms->hugepage_sz);
308 end = ms->addr_64 + ms->hugepage_sz;
310 /* Passed the end of a fragment. */
315 /* Found one chunk. */
316 entry->start = start;
318 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey);
324 * Insert a MR to the global B-tree cache. It may fail due to low-on-memory.
325 * Then, this entry will have to be searched by mr_lookup_dev_list() in
326 * mlx5_mr_create() on miss.
329 * Pointer to Ethernet device.
331 * Pointer to MR to insert.
334 * 0 on success, -1 on failure.
337 mr_insert_dev_cache(struct rte_eth_dev *dev, struct mlx5_mr *mr)
339 struct priv *priv = dev->data->dev_private;
342 DRV_LOG(DEBUG, "port %u inserting MR(%p) to global cache",
343 dev->data->port_id, (void *)mr);
344 for (n = 0; n < mr->ms_bmp_n; ) {
345 struct mlx5_mr_cache entry = { 0, };
347 /* Find a contiguous chunk and advance the index. */
348 n = mr_find_next_chunk(mr, &entry, n);
351 if (mr_btree_insert(&priv->mr.cache, &entry) < 0) {
353 * Overflowed, but the global table cannot be expanded
354 * because of deadlock.
363 * Look up address in the original global MR list.
366 * Pointer to Ethernet device.
368 * Pointer to returning MR cache entry. If no match, this will not be updated.
373 * Found MR on match, NULL otherwise.
375 static struct mlx5_mr *
376 mr_lookup_dev_list(struct rte_eth_dev *dev, struct mlx5_mr_cache *entry,
379 struct priv *priv = dev->data->dev_private;
382 /* Iterate all the existing MRs. */
383 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
388 for (n = 0; n < mr->ms_bmp_n; ) {
389 struct mlx5_mr_cache ret = { 0, };
391 n = mr_find_next_chunk(mr, &ret, n);
392 if (addr >= ret.start && addr < ret.end) {
403 * Look up address on device.
406 * Pointer to Ethernet device.
408 * Pointer to returning MR cache entry. If no match, this will not be updated.
413 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
416 mr_lookup_dev(struct rte_eth_dev *dev, struct mlx5_mr_cache *entry,
419 struct priv *priv = dev->data->dev_private;
421 uint32_t lkey = UINT32_MAX;
425 * If the global cache has overflowed since it failed to expand the
426 * B-tree table, it can't have all the existing MRs. Then, the address
427 * has to be searched by traversing the original MR list instead, which
428 * is very slow path. Otherwise, the global cache is all inclusive.
430 if (!unlikely(priv->mr.cache.overflow)) {
431 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
432 if (lkey != UINT32_MAX)
433 *entry = (*priv->mr.cache.table)[idx];
435 /* Falling back to the slowest path. */
436 mr = mr_lookup_dev_list(dev, entry, addr);
440 assert(lkey == UINT32_MAX || (addr >= entry->start &&
446 * Free MR resources. MR lock must not be held to avoid a deadlock. rte_free()
447 * can raise memory free event and the callback function will spin on the lock.
450 * Pointer to MR to free.
453 mr_free(struct mlx5_mr *mr)
457 DRV_LOG(DEBUG, "freeing MR(%p):", (void *)mr);
458 if (mr->ibv_mr != NULL)
459 claim_zero(mlx5_glue->dereg_mr(mr->ibv_mr));
460 if (mr->ms_bmp != NULL)
461 rte_bitmap_free(mr->ms_bmp);
466 * Releass resources of detached MR having no online entry.
469 * Pointer to Ethernet device.
472 mlx5_mr_garbage_collect(struct rte_eth_dev *dev)
474 struct priv *priv = dev->data->dev_private;
475 struct mlx5_mr *mr_next;
476 struct mlx5_mr_list free_list = LIST_HEAD_INITIALIZER(free_list);
478 /* Must be called from the primary process. */
479 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
481 * MR can't be freed with holding the lock because rte_free() could call
482 * memory free callback function. This will be a deadlock situation.
484 rte_rwlock_write_lock(&priv->mr.rwlock);
485 /* Detach the whole free list and release it after unlocking. */
486 free_list = priv->mr.mr_free_list;
487 LIST_INIT(&priv->mr.mr_free_list);
488 rte_rwlock_write_unlock(&priv->mr.rwlock);
489 /* Release resources. */
490 mr_next = LIST_FIRST(&free_list);
491 while (mr_next != NULL) {
492 struct mlx5_mr *mr = mr_next;
494 mr_next = LIST_NEXT(mr, mr);
499 /* Called during rte_memseg_contig_walk() by mlx5_mr_create(). */
501 mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl,
502 const struct rte_memseg *ms, size_t len, void *arg)
504 struct mr_find_contig_memsegs_data *data = arg;
506 if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len)
508 /* Found, save it and stop walking. */
509 data->start = ms->addr_64;
510 data->end = ms->addr_64 + len;
516 * Create a new global Memroy Region (MR) for a missing virtual address.
517 * Register entire virtually contiguous memory chunk around the address.
520 * Pointer to Ethernet device.
522 * Pointer to returning MR cache entry, found in the global cache or newly
523 * created. If failed to create one, this will not be updated.
525 * Target virtual address to register.
528 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
531 mlx5_mr_create(struct rte_eth_dev *dev, struct mlx5_mr_cache *entry,
534 struct priv *priv = dev->data->dev_private;
535 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
536 const struct rte_memseg_list *msl;
537 const struct rte_memseg *ms;
538 struct mlx5_mr *mr = NULL;
543 int ms_idx_shift = -1;
545 struct mr_find_contig_memsegs_data data = {
548 struct mr_find_contig_memsegs_data data_re;
550 DRV_LOG(DEBUG, "port %u creating a MR using address (%p)",
551 dev->data->port_id, (void *)addr);
552 if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
554 "port %u using address (%p) of unregistered mempool"
555 " in secondary process, please create mempool"
556 " before rte_eth_dev_start()",
557 dev->data->port_id, (void *)addr);
562 * Release detached MRs if any. This can't be called with holding either
563 * memory_hotplug_lock or priv->mr.rwlock. MRs on the free list have
564 * been detached by the memory free event but it couldn't be released
565 * inside the callback due to deadlock. As a result, releasing resources
566 * is quite opportunistic.
568 mlx5_mr_garbage_collect(dev);
570 * Find out a contiguous virtual address chunk in use, to which the
571 * given address belongs, in order to register maximum range. In the
572 * best case where mempools are not dynamically recreated and
573 * '--socket-mem' is speicified as an EAL option, it is very likely to
574 * have only one MR(LKey) per a socket and per a hugepage-size even
575 * though the system memory is highly fragmented.
577 if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) {
579 "port %u unable to find virtually contiguous"
580 " chunk for address (%p)."
581 " rte_memseg_contig_walk() failed.",
582 dev->data->port_id, (void *)addr);
587 /* Addresses must be page-aligned. */
588 assert(rte_is_aligned((void *)data.start, data.msl->page_sz));
589 assert(rte_is_aligned((void *)data.end, data.msl->page_sz));
591 ms = rte_mem_virt2memseg((void *)data.start, msl);
592 len = data.end - data.start;
593 assert(msl->page_sz == ms->hugepage_sz);
594 /* Number of memsegs in the range. */
595 ms_n = len / msl->page_sz;
596 DEBUG("port %u extending %p to [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
597 " page_sz=0x%" PRIx64 ", ms_n=%u",
598 dev->data->port_id, (void *)addr,
599 data.start, data.end, msl->page_sz, ms_n);
600 /* Size of memory for bitmap. */
601 bmp_size = rte_bitmap_get_memory_footprint(ms_n);
602 mr = rte_zmalloc_socket(NULL,
603 RTE_ALIGN_CEIL(sizeof(*mr),
604 RTE_CACHE_LINE_SIZE) +
606 RTE_CACHE_LINE_SIZE, msl->socket_id);
608 DEBUG("port %u unable to allocate memory for a new MR of"
610 dev->data->port_id, (void *)addr);
616 * Save the index of the first memseg and initialize memseg bitmap. To
617 * see if a memseg of ms_idx in the memseg-list is still valid, check:
618 * rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx)
620 mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
621 bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE);
622 mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size);
623 if (mr->ms_bmp == NULL) {
624 DEBUG("port %u unable to initialize bitamp for a new MR of"
626 dev->data->port_id, (void *)addr);
631 * Should recheck whether the extended contiguous chunk is still valid.
632 * Because memory_hotplug_lock can't be held if there's any memory
633 * related calls in a critical path, resource allocation above can't be
634 * locked. If the memory has been changed at this point, try again with
635 * just single page. If not, go on with the big chunk atomically from
638 rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
640 if (len > msl->page_sz &&
641 !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) {
642 DEBUG("port %u unable to find virtually contiguous"
643 " chunk for address (%p)."
644 " rte_memseg_contig_walk() failed.",
645 dev->data->port_id, (void *)addr);
649 if (data.start != data_re.start || data.end != data_re.end) {
651 * The extended contiguous chunk has been changed. Try again
652 * with single memseg instead.
654 data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz);
655 data.end = data.start + msl->page_sz;
656 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
658 goto alloc_resources;
660 assert(data.msl == data_re.msl);
661 rte_rwlock_write_lock(&priv->mr.rwlock);
663 * Check the address is really missing. If other thread already created
664 * one or it is not found due to overflow, abort and return.
666 if (mr_lookup_dev(dev, entry, addr) != UINT32_MAX) {
668 * Insert to the global cache table. It may fail due to
669 * low-on-memory. Then, this entry will have to be searched
672 mr_btree_insert(&priv->mr.cache, entry);
673 DEBUG("port %u found MR for %p on final lookup, abort",
674 dev->data->port_id, (void *)addr);
675 rte_rwlock_write_unlock(&priv->mr.rwlock);
676 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
678 * Must be unlocked before calling rte_free() because
679 * mlx5_mr_mem_event_free_cb() can be called inside.
685 * Trim start and end addresses for verbs MR. Set bits for registering
686 * memsegs but exclude already registered ones. Bitmap can be
689 for (n = 0; n < ms_n; ++n) {
691 struct mlx5_mr_cache ret = { 0, };
693 start = data_re.start + n * msl->page_sz;
694 /* Exclude memsegs already registered by other MRs. */
695 if (mr_lookup_dev(dev, &ret, start) == UINT32_MAX) {
697 * Start from the first unregistered memseg in the
700 if (ms_idx_shift == -1) {
701 mr->ms_base_idx += n;
705 data.end = start + msl->page_sz;
706 rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift);
710 len = data.end - data.start;
711 mr->ms_bmp_n = len / msl->page_sz;
712 assert(ms_idx_shift + mr->ms_bmp_n <= ms_n);
714 * Finally create a verbs MR for the memory chunk. ibv_reg_mr() can be
715 * called with holding the memory lock because it doesn't use
716 * mlx5_alloc_buf_extern() which eventually calls rte_malloc_socket()
717 * through mlx5_alloc_verbs_buf().
719 mr->ibv_mr = mlx5_glue->reg_mr(priv->pd, (void *)data.start, len,
720 IBV_ACCESS_LOCAL_WRITE);
721 if (mr->ibv_mr == NULL) {
722 DEBUG("port %u fail to create a verbs MR for address (%p)",
723 dev->data->port_id, (void *)addr);
727 assert((uintptr_t)mr->ibv_mr->addr == data.start);
728 assert(mr->ibv_mr->length == len);
729 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
730 DEBUG("port %u MR CREATED (%p) for %p:\n"
731 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
732 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
733 dev->data->port_id, (void *)mr, (void *)addr,
734 data.start, data.end, rte_cpu_to_be_32(mr->ibv_mr->lkey),
735 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
736 /* Insert to the global cache table. */
737 mr_insert_dev_cache(dev, mr);
738 /* Fill in output data. */
739 mr_lookup_dev(dev, entry, addr);
740 /* Lookup can't fail. */
741 assert(entry->lkey != UINT32_MAX);
742 rte_rwlock_write_unlock(&priv->mr.rwlock);
743 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
746 rte_rwlock_write_unlock(&priv->mr.rwlock);
748 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
751 * In case of error, as this can be called in a datapath, a warning
752 * message per an error is preferable instead. Must be unlocked before
753 * calling rte_free() because mlx5_mr_mem_event_free_cb() can be called
761 * Rebuild the global B-tree cache of device from the original MR list.
764 * Pointer to Ethernet device.
767 mr_rebuild_dev_cache(struct rte_eth_dev *dev)
769 struct priv *priv = dev->data->dev_private;
772 DRV_LOG(DEBUG, "port %u rebuild dev cache[]", dev->data->port_id);
773 /* Flush cache to rebuild. */
774 priv->mr.cache.len = 1;
775 priv->mr.cache.overflow = 0;
776 /* Iterate all the existing MRs. */
777 LIST_FOREACH(mr, &priv->mr.mr_list, mr)
778 if (mr_insert_dev_cache(dev, mr) < 0)
783 * Callback for memory free event. Iterate freed memsegs and check whether it
784 * belongs to an existing MR. If found, clear the bit from bitmap of MR. As a
785 * result, the MR would be fragmented. If it becomes empty, the MR will be freed
786 * later by mlx5_mr_garbage_collect(). Even if this callback is called from a
787 * secondary process, the garbage collector will be called in primary process
788 * as the secondary process can't call mlx5_mr_create().
790 * The global cache must be rebuilt if there's any change and this event has to
791 * be propagated to dataplane threads to flush the local caches.
794 * Pointer to Ethernet device.
796 * Address of freed memory.
798 * Size of freed memory.
801 mlx5_mr_mem_event_free_cb(struct rte_eth_dev *dev, const void *addr, size_t len)
803 struct priv *priv = dev->data->dev_private;
804 const struct rte_memseg_list *msl;
810 DEBUG("port %u free callback: addr=%p, len=%zu",
811 dev->data->port_id, addr, len);
812 msl = rte_mem_virt2memseg_list(addr);
813 /* addr and len must be page-aligned. */
814 assert((uintptr_t)addr == RTE_ALIGN((uintptr_t)addr, msl->page_sz));
815 assert(len == RTE_ALIGN(len, msl->page_sz));
816 ms_n = len / msl->page_sz;
817 rte_rwlock_write_lock(&priv->mr.rwlock);
818 /* Clear bits of freed memsegs from MR. */
819 for (i = 0; i < ms_n; ++i) {
820 const struct rte_memseg *ms;
821 struct mlx5_mr_cache entry;
826 /* Find MR having this memseg. */
827 start = (uintptr_t)addr + i * msl->page_sz;
828 mr = mr_lookup_dev_list(dev, &entry, start);
831 assert(mr->msl); /* Can't be external memory. */
832 ms = rte_mem_virt2memseg((void *)start, msl);
834 assert(msl->page_sz == ms->hugepage_sz);
835 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
836 pos = ms_idx - mr->ms_base_idx;
837 assert(rte_bitmap_get(mr->ms_bmp, pos));
838 assert(pos < mr->ms_bmp_n);
839 DEBUG("port %u MR(%p): clear bitmap[%u] for addr %p",
840 dev->data->port_id, (void *)mr, pos, (void *)start);
841 rte_bitmap_clear(mr->ms_bmp, pos);
842 if (--mr->ms_n == 0) {
844 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
845 DEBUG("port %u remove MR(%p) from list",
846 dev->data->port_id, (void *)mr);
849 * MR is fragmented or will be freed. the global cache must be
855 mr_rebuild_dev_cache(dev);
857 * Flush local caches by propagating invalidation across cores.
858 * rte_smp_wmb() is enough to synchronize this event. If one of
859 * freed memsegs is seen by other core, that means the memseg
860 * has been allocated by allocator, which will come after this
861 * free call. Therefore, this store instruction (incrementing
862 * generation below) will be guaranteed to be seen by other core
863 * before the core sees the newly allocated memory.
866 DEBUG("broadcasting local cache flush, gen=%d",
870 rte_rwlock_write_unlock(&priv->mr.rwlock);
874 * Callback for memory event. This can be called from both primary and secondary
885 mlx5_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
886 size_t len, void *arg __rte_unused)
889 struct mlx5_dev_list *dev_list = &mlx5_shared_data->mem_event_cb_list;
891 switch (event_type) {
892 case RTE_MEM_EVENT_FREE:
893 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
894 /* Iterate all the existing mlx5 devices. */
895 LIST_FOREACH(priv, dev_list, mem_event_cb)
896 mlx5_mr_mem_event_free_cb(ETH_DEV(priv), addr, len);
897 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
899 case RTE_MEM_EVENT_ALLOC:
906 * Look up address in the global MR cache table. If not found, create a new MR.
907 * Insert the found/created entry to local bottom-half cache table.
910 * Pointer to Ethernet device.
912 * Pointer to per-queue MR control structure.
914 * Pointer to returning MR cache entry, found in the global cache or newly
915 * created. If failed to create one, this is not written.
920 * Searched LKey on success, UINT32_MAX on no match.
923 mlx5_mr_lookup_dev(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
924 struct mlx5_mr_cache *entry, uintptr_t addr)
926 struct priv *priv = dev->data->dev_private;
927 struct mlx5_mr_btree *bt = &mr_ctrl->cache_bh;
931 /* If local cache table is full, try to double it. */
932 if (unlikely(bt->len == bt->size))
933 mr_btree_expand(bt, bt->size << 1);
934 /* Look up in the global cache. */
935 rte_rwlock_read_lock(&priv->mr.rwlock);
936 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
937 if (lkey != UINT32_MAX) {
939 *entry = (*priv->mr.cache.table)[idx];
940 rte_rwlock_read_unlock(&priv->mr.rwlock);
942 * Update local cache. Even if it fails, return the found entry
943 * to update top-half cache. Next time, this entry will be found
944 * in the global cache.
946 mr_btree_insert(bt, entry);
949 rte_rwlock_read_unlock(&priv->mr.rwlock);
950 /* First time to see the address? Create a new MR. */
951 lkey = mlx5_mr_create(dev, entry, addr);
953 * Update the local cache if successfully created a new global MR. Even
954 * if failed to create one, there's no action to take in this datapath
955 * code. As returning LKey is invalid, this will eventually make HW
958 if (lkey != UINT32_MAX)
959 mr_btree_insert(bt, entry);
964 * Bottom-half of LKey search on datapath. Firstly search in cache_bh[] and if
965 * misses, search in the global MR cache table and update the new entry to
966 * per-queue local caches.
969 * Pointer to Ethernet device.
971 * Pointer to per-queue MR control structure.
976 * Searched LKey on success, UINT32_MAX on no match.
979 mlx5_mr_addr2mr_bh(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
984 /* Victim in top-half cache to replace with new entry. */
985 struct mlx5_mr_cache *repl = &mr_ctrl->cache[mr_ctrl->head];
987 /* Binary-search MR translation table. */
988 lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr);
989 /* Update top-half cache. */
990 if (likely(lkey != UINT32_MAX)) {
991 *repl = (*mr_ctrl->cache_bh.table)[bh_idx];
994 * If missed in local lookup table, search in the global cache
995 * and local cache_bh[] will be updated inside if possible.
996 * Top-half cache entry will also be updated.
998 lkey = mlx5_mr_lookup_dev(dev, mr_ctrl, repl, addr);
999 if (unlikely(lkey == UINT32_MAX))
1002 /* Update the most recently used entry. */
1003 mr_ctrl->mru = mr_ctrl->head;
1004 /* Point to the next victim, the oldest. */
1005 mr_ctrl->head = (mr_ctrl->head + 1) % MLX5_MR_CACHE_N;
1010 * Bottom-half of LKey search on Rx.
1013 * Pointer to Rx queue structure.
1018 * Searched LKey on success, UINT32_MAX on no match.
1021 mlx5_rx_addr2mr_bh(struct mlx5_rxq_data *rxq, uintptr_t addr)
1023 struct mlx5_rxq_ctrl *rxq_ctrl =
1024 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
1025 struct mlx5_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
1026 struct priv *priv = rxq_ctrl->priv;
1029 "Rx queue %u: miss on top-half, mru=%u, head=%u, addr=%p",
1030 rxq_ctrl->idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr);
1031 return mlx5_mr_addr2mr_bh(ETH_DEV(priv), 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 mlx5_tx_addr2mr_bh(struct mlx5_txq_data *txq, uintptr_t addr)
1048 struct mlx5_txq_ctrl *txq_ctrl =
1049 container_of(txq, struct mlx5_txq_ctrl, txq);
1050 struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1051 struct priv *priv = txq_ctrl->priv;
1054 "Tx queue %u: miss on top-half, mru=%u, head=%u, addr=%p",
1055 txq_ctrl->idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr);
1056 return mlx5_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
1060 * Flush all of the local cache entries.
1063 * Pointer to per-queue MR control structure.
1066 mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl)
1068 /* Reset the most-recently-used index. */
1070 /* Reset the linear search array. */
1072 memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache));
1073 /* Reset the B-tree table. */
1074 mr_ctrl->cache_bh.len = 1;
1075 mr_ctrl->cache_bh.overflow = 0;
1076 /* Update the generation number. */
1077 mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr;
1078 DRV_LOG(DEBUG, "mr_ctrl(%p): flushed, cur_gen=%d",
1079 (void *)mr_ctrl, mr_ctrl->cur_gen);
1083 * Called during rte_mempool_mem_iter() by mlx5_mr_update_ext_mp().
1085 * Externally allocated chunk is registered and a MR is created for the chunk.
1086 * The MR object is added to the global list. If memseg list of a MR object
1087 * (mr->msl) is null, the MR object can be regarded as externally allocated
1090 * Once external memory is registered, it should be static. If the memory is
1091 * freed and the virtual address range has different physical memory mapped
1092 * again, it may cause crash on device due to the wrong translation entry. PMD
1093 * can't track the free event of the external memory for now.
1096 mlx5_mr_update_ext_mp_cb(struct rte_mempool *mp, void *opaque,
1097 struct rte_mempool_memhdr *memhdr,
1098 unsigned mem_idx __rte_unused)
1100 struct mr_update_mp_data *data = opaque;
1101 struct rte_eth_dev *dev = data->dev;
1102 struct priv *priv = dev->data->dev_private;
1103 struct mlx5_mr_ctrl *mr_ctrl = data->mr_ctrl;
1104 struct mlx5_mr *mr = NULL;
1105 uintptr_t addr = (uintptr_t)memhdr->addr;
1106 size_t len = memhdr->len;
1107 struct mlx5_mr_cache entry;
1110 /* If already registered, it should return. */
1111 rte_rwlock_read_lock(&priv->mr.rwlock);
1112 lkey = mr_lookup_dev(dev, &entry, addr);
1113 rte_rwlock_read_unlock(&priv->mr.rwlock);
1114 if (lkey != UINT32_MAX)
1116 mr = rte_zmalloc_socket(NULL,
1117 RTE_ALIGN_CEIL(sizeof(*mr),
1118 RTE_CACHE_LINE_SIZE),
1119 RTE_CACHE_LINE_SIZE, mp->socket_id);
1122 "port %u unable to allocate memory for a new MR of"
1124 dev->data->port_id, mp->name);
1128 DRV_LOG(DEBUG, "port %u register MR for chunk #%d of mempool (%s)",
1129 dev->data->port_id, mem_idx, mp->name);
1130 mr->ibv_mr = mlx5_glue->reg_mr(priv->pd, (void *)addr, len,
1131 IBV_ACCESS_LOCAL_WRITE);
1132 if (mr->ibv_mr == NULL) {
1134 "port %u fail to create a verbs MR for address (%p)",
1135 dev->data->port_id, (void *)addr);
1140 mr->msl = NULL; /* Mark it is external memory. */
1144 rte_rwlock_write_lock(&priv->mr.rwlock);
1145 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
1147 "port %u MR CREATED (%p) for external memory %p:\n"
1148 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
1149 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
1150 dev->data->port_id, (void *)mr, (void *)addr,
1151 addr, addr + len, rte_cpu_to_be_32(mr->ibv_mr->lkey),
1152 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
1153 /* Insert to the global cache table. */
1154 mr_insert_dev_cache(dev, mr);
1155 rte_rwlock_write_unlock(&priv->mr.rwlock);
1156 /* Insert to the local cache table */
1157 mlx5_mr_addr2mr_bh(dev, mr_ctrl, addr);
1161 * Register MR for entire memory chunks in a Mempool having externally allocated
1162 * memory and fill in local cache.
1165 * Pointer to Ethernet device.
1167 * Pointer to per-queue MR control structure.
1169 * Pointer to registering Mempool.
1172 * 0 on success, -1 on failure.
1175 mlx5_mr_update_ext_mp(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
1176 struct rte_mempool *mp)
1178 struct mr_update_mp_data data = {
1184 rte_mempool_mem_iter(mp, mlx5_mr_update_ext_mp_cb, &data);
1189 * Register MR entire memory chunks in a Mempool having externally allocated
1190 * memory and search LKey of the address to return.
1193 * Pointer to Ethernet device.
1197 * Pointer to registering Mempool where addr belongs.
1200 * LKey for address on success, UINT32_MAX on failure.
1203 mlx5_tx_update_ext_mp(struct mlx5_txq_data *txq, uintptr_t addr,
1204 struct rte_mempool *mp)
1206 struct mlx5_txq_ctrl *txq_ctrl =
1207 container_of(txq, struct mlx5_txq_ctrl, txq);
1208 struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1209 struct priv *priv = txq_ctrl->priv;
1211 mlx5_mr_update_ext_mp(ETH_DEV(priv), mr_ctrl, mp);
1212 return mlx5_tx_addr2mr_bh(txq, addr);
1215 /* Called during rte_mempool_mem_iter() by mlx5_mr_update_mp(). */
1217 mlx5_mr_update_mp_cb(struct rte_mempool *mp __rte_unused, void *opaque,
1218 struct rte_mempool_memhdr *memhdr,
1219 unsigned mem_idx __rte_unused)
1221 struct mr_update_mp_data *data = opaque;
1224 /* Stop iteration if failed in the previous walk. */
1227 /* Register address of the chunk and update local caches. */
1228 lkey = mlx5_mr_addr2mr_bh(data->dev, data->mr_ctrl,
1229 (uintptr_t)memhdr->addr);
1230 if (lkey == UINT32_MAX)
1235 * Register entire memory chunks in a Mempool.
1238 * Pointer to Ethernet device.
1240 * Pointer to per-queue MR control structure.
1242 * Pointer to registering Mempool.
1245 * 0 on success, -1 on failure.
1248 mlx5_mr_update_mp(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
1249 struct rte_mempool *mp)
1251 struct mr_update_mp_data data = {
1257 rte_mempool_mem_iter(mp, mlx5_mr_update_mp_cb, &data);
1258 if (data.ret < 0 && rte_errno == ENXIO) {
1259 /* Mempool may have externally allocated memory. */
1260 return mlx5_mr_update_ext_mp(dev, mr_ctrl, mp);
1266 * Dump all the created MRs and the global cache entries.
1269 * Pointer to Ethernet device.
1272 mlx5_mr_dump_dev(struct rte_eth_dev *dev __rte_unused)
1275 struct priv *priv = dev->data->dev_private;
1280 rte_rwlock_read_lock(&priv->mr.rwlock);
1281 /* Iterate all the existing MRs. */
1282 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
1285 DEBUG("port %u MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u",
1286 dev->data->port_id, mr_n++,
1287 rte_cpu_to_be_32(mr->ibv_mr->lkey),
1288 mr->ms_n, mr->ms_bmp_n);
1291 for (n = 0; n < mr->ms_bmp_n; ) {
1292 struct mlx5_mr_cache ret = { 0, };
1294 n = mr_find_next_chunk(mr, &ret, n);
1297 DEBUG(" chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")",
1298 chunk_n++, ret.start, ret.end);
1301 DEBUG("port %u dumping global cache", dev->data->port_id);
1302 mlx5_mr_btree_dump(&priv->mr.cache);
1303 rte_rwlock_read_unlock(&priv->mr.rwlock);
1308 * Release all the created MRs and resources. Remove device from memory callback
1312 * Pointer to Ethernet device.
1315 mlx5_mr_release(struct rte_eth_dev *dev)
1317 struct priv *priv = dev->data->dev_private;
1318 struct mlx5_mr *mr_next = LIST_FIRST(&priv->mr.mr_list);
1320 /* Remove from memory callback device list. */
1321 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1322 LIST_REMOVE(priv, mem_event_cb);
1323 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1324 if (rte_log_get_level(mlx5_logtype) == RTE_LOG_DEBUG)
1325 mlx5_mr_dump_dev(dev);
1326 rte_rwlock_write_lock(&priv->mr.rwlock);
1327 /* Detach from MR list and move to free list. */
1328 while (mr_next != NULL) {
1329 struct mlx5_mr *mr = mr_next;
1331 mr_next = LIST_NEXT(mr, mr);
1332 LIST_REMOVE(mr, mr);
1333 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
1335 LIST_INIT(&priv->mr.mr_list);
1336 /* Free global cache. */
1337 mlx5_mr_btree_free(&priv->mr.cache);
1338 rte_rwlock_write_unlock(&priv->mr.rwlock);
1339 /* Free all remaining MRs. */
1340 mlx5_mr_garbage_collect(dev);