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>
17 #include <rte_bus_pci.h>
21 #include "mlx5_rxtx.h"
22 #include "mlx5_glue.h"
24 struct mr_find_contig_memsegs_data {
28 const struct rte_memseg_list *msl;
31 struct mr_update_mp_data {
32 struct rte_eth_dev *dev;
33 struct mlx5_mr_ctrl *mr_ctrl;
38 * Expand B-tree table to a given size. Can't be called with holding
39 * memory_hotplug_lock or sh->mr.rwlock due to rte_realloc().
42 * Pointer to B-tree structure.
44 * Number of entries for expansion.
47 * 0 on success, -1 on failure.
50 mr_btree_expand(struct mlx5_mr_btree *bt, int n)
58 * Downside of directly using rte_realloc() is that SOCKET_ID_ANY is
59 * used inside if there's no room to expand. Because this is a quite
60 * rare case and a part of very slow path, it is very acceptable.
61 * Initially cache_bh[] will be given practically enough space and once
62 * it is expanded, expansion wouldn't be needed again ever.
64 mem = rte_realloc(bt->table, n * sizeof(struct mlx5_mr_cache), 0);
66 /* Not an error, B-tree search will be skipped. */
67 DRV_LOG(WARNING, "failed to expand MR B-tree (%p) table",
71 DRV_LOG(DEBUG, "expanded MR B-tree table (size=%u)", n);
79 * Look up LKey from given B-tree lookup table, store the last index and return
83 * Pointer to B-tree structure.
85 * Pointer to index. Even on search failure, returns index where it stops
86 * searching so that index can be used when inserting a new entry.
91 * Searched LKey on success, UINT32_MAX on no match.
94 mr_btree_lookup(struct mlx5_mr_btree *bt, uint16_t *idx, uintptr_t addr)
96 struct mlx5_mr_cache *lkp_tbl;
101 lkp_tbl = *bt->table;
103 /* First entry must be NULL for comparison. */
104 assert(bt->len > 0 || (lkp_tbl[0].start == 0 &&
105 lkp_tbl[0].lkey == UINT32_MAX));
108 register uint16_t delta = n >> 1;
110 if (addr < lkp_tbl[base + delta].start) {
117 assert(addr >= lkp_tbl[base].start);
119 if (addr < lkp_tbl[base].end)
120 return lkp_tbl[base].lkey;
126 * Insert an entry to B-tree lookup table.
129 * Pointer to B-tree structure.
131 * Pointer to new entry to insert.
134 * 0 on success, -1 on failure.
137 mr_btree_insert(struct mlx5_mr_btree *bt, struct mlx5_mr_cache *entry)
139 struct mlx5_mr_cache *lkp_tbl;
144 assert(bt->len <= bt->size);
146 lkp_tbl = *bt->table;
147 /* Find out the slot for insertion. */
148 if (mr_btree_lookup(bt, &idx, entry->start) != UINT32_MAX) {
150 "abort insertion to B-tree(%p): already exist at"
151 " idx=%u [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
152 (void *)bt, idx, entry->start, entry->end, entry->lkey);
153 /* Already exist, return. */
156 /* If table is full, return error. */
157 if (unlikely(bt->len == bt->size)) {
163 shift = (bt->len - idx) * sizeof(struct mlx5_mr_cache);
165 memmove(&lkp_tbl[idx + 1], &lkp_tbl[idx], shift);
166 lkp_tbl[idx] = *entry;
169 "inserted B-tree(%p)[%u],"
170 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
171 (void *)bt, idx, entry->start, entry->end, entry->lkey);
176 * Initialize B-tree and allocate memory for lookup table.
179 * Pointer to B-tree structure.
181 * Number of entries to allocate.
183 * NUMA socket on which memory must be allocated.
186 * 0 on success, a negative errno value otherwise and rte_errno is set.
189 mlx5_mr_btree_init(struct mlx5_mr_btree *bt, int n, int socket)
195 assert(!bt->table && !bt->size);
196 memset(bt, 0, sizeof(*bt));
197 bt->table = rte_calloc_socket("B-tree table",
198 n, sizeof(struct mlx5_mr_cache),
200 if (bt->table == NULL) {
202 DEBUG("failed to allocate memory for btree cache on socket %d",
207 /* First entry must be NULL for binary search. */
208 (*bt->table)[bt->len++] = (struct mlx5_mr_cache) {
211 DEBUG("initialized B-tree %p with table %p",
212 (void *)bt, (void *)bt->table);
217 * Free B-tree resources.
220 * Pointer to B-tree structure.
223 mlx5_mr_btree_free(struct mlx5_mr_btree *bt)
227 DEBUG("freeing B-tree %p with table %p",
228 (void *)bt, (void *)bt->table);
230 memset(bt, 0, sizeof(*bt));
234 * Dump all the entries in a B-tree
237 * Pointer to B-tree structure.
240 mlx5_mr_btree_dump(struct mlx5_mr_btree *bt __rte_unused)
244 struct mlx5_mr_cache *lkp_tbl;
248 lkp_tbl = *bt->table;
249 for (idx = 0; idx < bt->len; ++idx) {
250 struct mlx5_mr_cache *entry = &lkp_tbl[idx];
252 DEBUG("B-tree(%p)[%u],"
253 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
254 (void *)bt, idx, entry->start, entry->end, entry->lkey);
260 * Find virtually contiguous memory chunk in a given MR.
263 * Pointer to MR structure.
265 * Pointer to returning MR cache entry. If not found, this will not be
268 * Start index of the memseg bitmap.
271 * Next index to go on lookup.
274 mr_find_next_chunk(struct mlx5_mr *mr, struct mlx5_mr_cache *entry,
281 /* MR for external memory doesn't have memseg list. */
282 if (mr->msl == NULL) {
283 struct ibv_mr *ibv_mr = mr->ibv_mr;
285 assert(mr->ms_bmp_n == 1);
286 assert(mr->ms_n == 1);
287 assert(base_idx == 0);
289 * Can't search it from memseg list but get it directly from
290 * verbs MR as there's only one chunk.
292 entry->start = (uintptr_t)ibv_mr->addr;
293 entry->end = (uintptr_t)ibv_mr->addr + mr->ibv_mr->length;
294 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey);
295 /* Returning 1 ends iteration. */
298 for (idx = base_idx; idx < mr->ms_bmp_n; ++idx) {
299 if (rte_bitmap_get(mr->ms_bmp, idx)) {
300 const struct rte_memseg_list *msl;
301 const struct rte_memseg *ms;
304 ms = rte_fbarray_get(&msl->memseg_arr,
305 mr->ms_base_idx + idx);
306 assert(msl->page_sz == ms->hugepage_sz);
309 end = ms->addr_64 + ms->hugepage_sz;
311 /* Passed the end of a fragment. */
316 /* Found one chunk. */
317 entry->start = start;
319 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey);
325 * Insert a MR to the global B-tree cache. It may fail due to low-on-memory.
326 * Then, this entry will have to be searched by mr_lookup_dev_list() in
327 * mlx5_mr_create() on miss.
330 * Pointer to Ethernet device shared context.
332 * Pointer to MR to insert.
335 * 0 on success, -1 on failure.
338 mr_insert_dev_cache(struct mlx5_ibv_shared *sh, struct mlx5_mr *mr)
342 DRV_LOG(DEBUG, "device %s inserting MR(%p) to global cache",
343 sh->ibdev_name, (void *)mr);
344 for (n = 0; n < mr->ms_bmp_n; ) {
345 struct mlx5_mr_cache entry;
347 memset(&entry, 0, sizeof(entry));
348 /* Find a contiguous chunk and advance the index. */
349 n = mr_find_next_chunk(mr, &entry, n);
352 if (mr_btree_insert(&sh->mr.cache, &entry) < 0) {
354 * Overflowed, but the global table cannot be expanded
355 * because of deadlock.
364 * Look up address in the original global MR list.
367 * Pointer to Ethernet device shared context.
369 * Pointer to returning MR cache entry. If no match, this will not be updated.
374 * Found MR on match, NULL otherwise.
376 static struct mlx5_mr *
377 mr_lookup_dev_list(struct mlx5_ibv_shared *sh, struct mlx5_mr_cache *entry,
382 /* Iterate all the existing MRs. */
383 LIST_FOREACH(mr, &sh->mr.mr_list, mr) {
388 for (n = 0; n < mr->ms_bmp_n; ) {
389 struct mlx5_mr_cache ret;
391 memset(&ret, 0, sizeof(ret));
392 n = mr_find_next_chunk(mr, &ret, n);
393 if (addr >= ret.start && addr < ret.end) {
404 * Look up address on device.
407 * Pointer to Ethernet device shared context.
409 * Pointer to returning MR cache entry. If no match, this will not be updated.
414 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
417 mr_lookup_dev(struct mlx5_ibv_shared *sh, struct mlx5_mr_cache *entry,
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(sh->mr.cache.overflow)) {
431 lkey = mr_btree_lookup(&sh->mr.cache, &idx, addr);
432 if (lkey != UINT32_MAX)
433 *entry = (*sh->mr.cache.table)[idx];
435 /* Falling back to the slowest path. */
436 mr = mr_lookup_dev_list(sh, 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 * Release resources of detached MR having no online entry.
469 * Pointer to Ethernet device shared context.
472 mlx5_mr_garbage_collect(struct mlx5_ibv_shared *sh)
474 struct mlx5_mr *mr_next;
475 struct mlx5_mr_list free_list = LIST_HEAD_INITIALIZER(free_list);
477 /* Must be called from the primary process. */
478 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
480 * MR can't be freed with holding the lock because rte_free() could call
481 * memory free callback function. This will be a deadlock situation.
483 rte_rwlock_write_lock(&sh->mr.rwlock);
484 /* Detach the whole free list and release it after unlocking. */
485 free_list = sh->mr.mr_free_list;
486 LIST_INIT(&sh->mr.mr_free_list);
487 rte_rwlock_write_unlock(&sh->mr.rwlock);
488 /* Release resources. */
489 mr_next = LIST_FIRST(&free_list);
490 while (mr_next != NULL) {
491 struct mlx5_mr *mr = mr_next;
493 mr_next = LIST_NEXT(mr, mr);
498 /* Called during rte_memseg_contig_walk() by mlx5_mr_create(). */
500 mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl,
501 const struct rte_memseg *ms, size_t len, void *arg)
503 struct mr_find_contig_memsegs_data *data = arg;
505 if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len)
507 /* Found, save it and stop walking. */
508 data->start = ms->addr_64;
509 data->end = ms->addr_64 + len;
515 * Create a new global Memory Region (MR) for a missing virtual address.
516 * This API should be called on a secondary process, then a request is sent to
517 * the primary process in order to create a MR for the address. As the global MR
518 * list is on the shared memory, following LKey lookup should succeed unless the
522 * Pointer to Ethernet device.
524 * Pointer to returning MR cache entry, found in the global cache or newly
525 * created. If failed to create one, this will not be updated.
527 * Target virtual address to register.
530 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
533 mlx5_mr_create_secondary(struct rte_eth_dev *dev, struct mlx5_mr_cache *entry,
536 struct mlx5_priv *priv = dev->data->dev_private;
539 DEBUG("port %u requesting MR creation for address (%p)",
540 dev->data->port_id, (void *)addr);
541 ret = mlx5_mp_req_mr_create(dev, addr);
543 DEBUG("port %u fail to request MR creation for address (%p)",
544 dev->data->port_id, (void *)addr);
547 rte_rwlock_read_lock(&priv->sh->mr.rwlock);
548 /* Fill in output data. */
549 mr_lookup_dev(priv->sh, entry, addr);
550 /* Lookup can't fail. */
551 assert(entry->lkey != UINT32_MAX);
552 rte_rwlock_read_unlock(&priv->sh->mr.rwlock);
553 DEBUG("port %u MR CREATED by primary process for %p:\n"
554 " [0x%" PRIxPTR ", 0x%" PRIxPTR "), lkey=0x%x",
555 dev->data->port_id, (void *)addr,
556 entry->start, entry->end, entry->lkey);
561 * Create a new global Memory Region (MR) for a missing virtual address.
562 * Register entire virtually contiguous memory chunk around the address.
563 * This must be called from the primary process.
566 * Pointer to Ethernet device.
568 * Pointer to returning MR cache entry, found in the global cache or newly
569 * created. If failed to create one, this will not be updated.
571 * Target virtual address to register.
574 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
577 mlx5_mr_create_primary(struct rte_eth_dev *dev, struct mlx5_mr_cache *entry,
580 struct mlx5_priv *priv = dev->data->dev_private;
581 struct mlx5_ibv_shared *sh = priv->sh;
582 struct mlx5_dev_config *config = &priv->config;
583 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
584 const struct rte_memseg_list *msl;
585 const struct rte_memseg *ms;
586 struct mlx5_mr *mr = NULL;
591 int ms_idx_shift = -1;
593 struct mr_find_contig_memsegs_data data = {
596 struct mr_find_contig_memsegs_data data_re;
598 DRV_LOG(DEBUG, "port %u creating a MR using address (%p)",
599 dev->data->port_id, (void *)addr);
601 * Release detached MRs if any. This can't be called with holding either
602 * memory_hotplug_lock or sh->mr.rwlock. MRs on the free list have
603 * been detached by the memory free event but it couldn't be released
604 * inside the callback due to deadlock. As a result, releasing resources
605 * is quite opportunistic.
607 mlx5_mr_garbage_collect(sh);
609 * If enabled, find out a contiguous virtual address chunk in use, to
610 * which the given address belongs, in order to register maximum range.
611 * In the best case where mempools are not dynamically recreated and
612 * '--socket-mem' is specified as an EAL option, it is very likely to
613 * have only one MR(LKey) per a socket and per a hugepage-size even
614 * though the system memory is highly fragmented. As the whole memory
615 * chunk will be pinned by kernel, it can't be reused unless entire
616 * chunk is freed from EAL.
618 * If disabled, just register one memseg (page). Then, memory
619 * consumption will be minimized but it may drop performance if there
620 * are many MRs to lookup on the datapath.
622 if (!config->mr_ext_memseg_en) {
623 data.msl = rte_mem_virt2memseg_list((void *)addr);
624 data.start = RTE_ALIGN_FLOOR(addr, data.msl->page_sz);
625 data.end = data.start + data.msl->page_sz;
626 } else if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) {
628 "port %u unable to find virtually contiguous"
629 " chunk for address (%p)."
630 " rte_memseg_contig_walk() failed.",
631 dev->data->port_id, (void *)addr);
636 /* Addresses must be page-aligned. */
637 assert(rte_is_aligned((void *)data.start, data.msl->page_sz));
638 assert(rte_is_aligned((void *)data.end, data.msl->page_sz));
640 ms = rte_mem_virt2memseg((void *)data.start, msl);
641 len = data.end - data.start;
642 assert(msl->page_sz == ms->hugepage_sz);
643 /* Number of memsegs in the range. */
644 ms_n = len / msl->page_sz;
645 DEBUG("port %u extending %p to [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
646 " page_sz=0x%" PRIx64 ", ms_n=%u",
647 dev->data->port_id, (void *)addr,
648 data.start, data.end, msl->page_sz, ms_n);
649 /* Size of memory for bitmap. */
650 bmp_size = rte_bitmap_get_memory_footprint(ms_n);
651 mr = rte_zmalloc_socket(NULL,
652 RTE_ALIGN_CEIL(sizeof(*mr),
653 RTE_CACHE_LINE_SIZE) +
655 RTE_CACHE_LINE_SIZE, msl->socket_id);
657 DEBUG("port %u unable to allocate memory for a new MR of"
659 dev->data->port_id, (void *)addr);
665 * Save the index of the first memseg and initialize memseg bitmap. To
666 * see if a memseg of ms_idx in the memseg-list is still valid, check:
667 * rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx)
669 mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
670 bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE);
671 mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size);
672 if (mr->ms_bmp == NULL) {
673 DEBUG("port %u unable to initialize bitmap for a new MR of"
675 dev->data->port_id, (void *)addr);
680 * Should recheck whether the extended contiguous chunk is still valid.
681 * Because memory_hotplug_lock can't be held if there's any memory
682 * related calls in a critical path, resource allocation above can't be
683 * locked. If the memory has been changed at this point, try again with
684 * just single page. If not, go on with the big chunk atomically from
687 rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
689 if (len > msl->page_sz &&
690 !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) {
691 DEBUG("port %u unable to find virtually contiguous"
692 " chunk for address (%p)."
693 " rte_memseg_contig_walk() failed.",
694 dev->data->port_id, (void *)addr);
698 if (data.start != data_re.start || data.end != data_re.end) {
700 * The extended contiguous chunk has been changed. Try again
701 * with single memseg instead.
703 data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz);
704 data.end = data.start + msl->page_sz;
705 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
707 goto alloc_resources;
709 assert(data.msl == data_re.msl);
710 rte_rwlock_write_lock(&sh->mr.rwlock);
712 * Check the address is really missing. If other thread already created
713 * one or it is not found due to overflow, abort and return.
715 if (mr_lookup_dev(sh, entry, addr) != UINT32_MAX) {
717 * Insert to the global cache table. It may fail due to
718 * low-on-memory. Then, this entry will have to be searched
721 mr_btree_insert(&sh->mr.cache, entry);
722 DEBUG("port %u found MR for %p on final lookup, abort",
723 dev->data->port_id, (void *)addr);
724 rte_rwlock_write_unlock(&sh->mr.rwlock);
725 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
727 * Must be unlocked before calling rte_free() because
728 * mlx5_mr_mem_event_free_cb() can be called inside.
734 * Trim start and end addresses for verbs MR. Set bits for registering
735 * memsegs but exclude already registered ones. Bitmap can be
738 for (n = 0; n < ms_n; ++n) {
740 struct mlx5_mr_cache ret;
742 memset(&ret, 0, sizeof(ret));
743 start = data_re.start + n * msl->page_sz;
744 /* Exclude memsegs already registered by other MRs. */
745 if (mr_lookup_dev(sh, &ret, start) == UINT32_MAX) {
747 * Start from the first unregistered memseg in the
750 if (ms_idx_shift == -1) {
751 mr->ms_base_idx += n;
755 data.end = start + msl->page_sz;
756 rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift);
760 len = data.end - data.start;
761 mr->ms_bmp_n = len / msl->page_sz;
762 assert(ms_idx_shift + mr->ms_bmp_n <= ms_n);
764 * Finally create a verbs MR for the memory chunk. ibv_reg_mr() can be
765 * called with holding the memory lock because it doesn't use
766 * mlx5_alloc_buf_extern() which eventually calls rte_malloc_socket()
767 * through mlx5_alloc_verbs_buf().
769 mr->ibv_mr = mlx5_glue->reg_mr(sh->pd, (void *)data.start, len,
770 IBV_ACCESS_LOCAL_WRITE);
771 if (mr->ibv_mr == NULL) {
772 DEBUG("port %u fail to create a verbs MR for address (%p)",
773 dev->data->port_id, (void *)addr);
777 assert((uintptr_t)mr->ibv_mr->addr == data.start);
778 assert(mr->ibv_mr->length == len);
779 LIST_INSERT_HEAD(&sh->mr.mr_list, mr, mr);
780 DEBUG("port %u MR CREATED (%p) for %p:\n"
781 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
782 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
783 dev->data->port_id, (void *)mr, (void *)addr,
784 data.start, data.end, rte_cpu_to_be_32(mr->ibv_mr->lkey),
785 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
786 /* Insert to the global cache table. */
787 mr_insert_dev_cache(sh, mr);
788 /* Fill in output data. */
789 mr_lookup_dev(sh, entry, addr);
790 /* Lookup can't fail. */
791 assert(entry->lkey != UINT32_MAX);
792 rte_rwlock_write_unlock(&sh->mr.rwlock);
793 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
796 rte_rwlock_write_unlock(&sh->mr.rwlock);
798 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
801 * In case of error, as this can be called in a datapath, a warning
802 * message per an error is preferable instead. Must be unlocked before
803 * calling rte_free() because mlx5_mr_mem_event_free_cb() can be called
811 * Create a new global Memory Region (MR) for a missing virtual address.
812 * This can be called from primary and secondary process.
815 * Pointer to Ethernet device.
817 * Pointer to returning MR cache entry, found in the global cache or newly
818 * created. If failed to create one, this will not be updated.
820 * Target virtual address to register.
823 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
826 mlx5_mr_create(struct rte_eth_dev *dev, struct mlx5_mr_cache *entry,
831 switch (rte_eal_process_type()) {
832 case RTE_PROC_PRIMARY:
833 ret = mlx5_mr_create_primary(dev, entry, addr);
835 case RTE_PROC_SECONDARY:
836 ret = mlx5_mr_create_secondary(dev, entry, addr);
845 * Rebuild the global B-tree cache of device from the original MR list.
848 * Pointer to Ethernet device shared context.
851 mr_rebuild_dev_cache(struct mlx5_ibv_shared *sh)
855 DRV_LOG(DEBUG, "device %s rebuild dev cache[]", sh->ibdev_name);
856 /* Flush cache to rebuild. */
857 sh->mr.cache.len = 1;
858 sh->mr.cache.overflow = 0;
859 /* Iterate all the existing MRs. */
860 LIST_FOREACH(mr, &sh->mr.mr_list, mr)
861 if (mr_insert_dev_cache(sh, mr) < 0)
866 * Callback for memory free event. Iterate freed memsegs and check whether it
867 * belongs to an existing MR. If found, clear the bit from bitmap of MR. As a
868 * result, the MR would be fragmented. If it becomes empty, the MR will be freed
869 * later by mlx5_mr_garbage_collect(). Even if this callback is called from a
870 * secondary process, the garbage collector will be called in primary process
871 * as the secondary process can't call mlx5_mr_create().
873 * The global cache must be rebuilt if there's any change and this event has to
874 * be propagated to dataplane threads to flush the local caches.
877 * Pointer to the Ethernet device shared context.
879 * Address of freed memory.
881 * Size of freed memory.
884 mlx5_mr_mem_event_free_cb(struct mlx5_ibv_shared *sh,
885 const void *addr, size_t len)
887 const struct rte_memseg_list *msl;
893 DEBUG("device %s free callback: addr=%p, len=%zu",
894 sh->ibdev_name, addr, len);
895 msl = rte_mem_virt2memseg_list(addr);
896 /* addr and len must be page-aligned. */
897 assert((uintptr_t)addr == RTE_ALIGN((uintptr_t)addr, msl->page_sz));
898 assert(len == RTE_ALIGN(len, msl->page_sz));
899 ms_n = len / msl->page_sz;
900 rte_rwlock_write_lock(&sh->mr.rwlock);
901 /* Clear bits of freed memsegs from MR. */
902 for (i = 0; i < ms_n; ++i) {
903 const struct rte_memseg *ms;
904 struct mlx5_mr_cache entry;
909 /* Find MR having this memseg. */
910 start = (uintptr_t)addr + i * msl->page_sz;
911 mr = mr_lookup_dev_list(sh, &entry, start);
914 assert(mr->msl); /* Can't be external memory. */
915 ms = rte_mem_virt2memseg((void *)start, msl);
917 assert(msl->page_sz == ms->hugepage_sz);
918 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
919 pos = ms_idx - mr->ms_base_idx;
920 assert(rte_bitmap_get(mr->ms_bmp, pos));
921 assert(pos < mr->ms_bmp_n);
922 DEBUG("device %s MR(%p): clear bitmap[%u] for addr %p",
923 sh->ibdev_name, (void *)mr, pos, (void *)start);
924 rte_bitmap_clear(mr->ms_bmp, pos);
925 if (--mr->ms_n == 0) {
927 LIST_INSERT_HEAD(&sh->mr.mr_free_list, mr, mr);
928 DEBUG("device %s remove MR(%p) from list",
929 sh->ibdev_name, (void *)mr);
932 * MR is fragmented or will be freed. the global cache must be
938 mr_rebuild_dev_cache(sh);
940 * Flush local caches by propagating invalidation across cores.
941 * rte_smp_wmb() is enough to synchronize this event. If one of
942 * freed memsegs is seen by other core, that means the memseg
943 * has been allocated by allocator, which will come after this
944 * free call. Therefore, this store instruction (incrementing
945 * generation below) will be guaranteed to be seen by other core
946 * before the core sees the newly allocated memory.
949 DEBUG("broadcasting local cache flush, gen=%d",
953 rte_rwlock_write_unlock(&sh->mr.rwlock);
957 * Callback for memory event. This can be called from both primary and secondary
968 mlx5_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
969 size_t len, void *arg __rte_unused)
971 struct mlx5_ibv_shared *sh;
972 struct mlx5_dev_list *dev_list = &mlx5_shared_data->mem_event_cb_list;
974 /* Must be called from the primary process. */
975 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
976 switch (event_type) {
977 case RTE_MEM_EVENT_FREE:
978 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
979 /* Iterate all the existing mlx5 devices. */
980 LIST_FOREACH(sh, dev_list, mem_event_cb)
981 mlx5_mr_mem_event_free_cb(sh, addr, len);
982 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
984 case RTE_MEM_EVENT_ALLOC:
991 * Look up address in the global MR cache table. If not found, create a new MR.
992 * Insert the found/created entry to local bottom-half cache table.
995 * Pointer to Ethernet device.
997 * Pointer to per-queue MR control structure.
999 * Pointer to returning MR cache entry, found in the global cache or newly
1000 * created. If failed to create one, this is not written.
1005 * Searched LKey on success, UINT32_MAX on no match.
1008 mlx5_mr_lookup_dev(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
1009 struct mlx5_mr_cache *entry, uintptr_t addr)
1011 struct mlx5_priv *priv = dev->data->dev_private;
1012 struct mlx5_ibv_shared *sh = priv->sh;
1013 struct mlx5_mr_btree *bt = &mr_ctrl->cache_bh;
1017 /* If local cache table is full, try to double it. */
1018 if (unlikely(bt->len == bt->size))
1019 mr_btree_expand(bt, bt->size << 1);
1020 /* Look up in the global cache. */
1021 rte_rwlock_read_lock(&sh->mr.rwlock);
1022 lkey = mr_btree_lookup(&sh->mr.cache, &idx, addr);
1023 if (lkey != UINT32_MAX) {
1025 *entry = (*sh->mr.cache.table)[idx];
1026 rte_rwlock_read_unlock(&sh->mr.rwlock);
1028 * Update local cache. Even if it fails, return the found entry
1029 * to update top-half cache. Next time, this entry will be found
1030 * in the global cache.
1032 mr_btree_insert(bt, entry);
1035 rte_rwlock_read_unlock(&sh->mr.rwlock);
1036 /* First time to see the address? Create a new MR. */
1037 lkey = mlx5_mr_create(dev, entry, addr);
1039 * Update the local cache if successfully created a new global MR. Even
1040 * if failed to create one, there's no action to take in this datapath
1041 * code. As returning LKey is invalid, this will eventually make HW
1044 if (lkey != UINT32_MAX)
1045 mr_btree_insert(bt, entry);
1050 * Bottom-half of LKey search on datapath. Firstly search in cache_bh[] and if
1051 * misses, search in the global MR cache table and update the new entry to
1052 * per-queue local caches.
1055 * Pointer to Ethernet device.
1057 * Pointer to per-queue MR control structure.
1062 * Searched LKey on success, UINT32_MAX on no match.
1065 mlx5_mr_addr2mr_bh(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
1069 uint16_t bh_idx = 0;
1070 /* Victim in top-half cache to replace with new entry. */
1071 struct mlx5_mr_cache *repl = &mr_ctrl->cache[mr_ctrl->head];
1073 /* Binary-search MR translation table. */
1074 lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr);
1075 /* Update top-half cache. */
1076 if (likely(lkey != UINT32_MAX)) {
1077 *repl = (*mr_ctrl->cache_bh.table)[bh_idx];
1080 * If missed in local lookup table, search in the global cache
1081 * and local cache_bh[] will be updated inside if possible.
1082 * Top-half cache entry will also be updated.
1084 lkey = mlx5_mr_lookup_dev(dev, mr_ctrl, repl, addr);
1085 if (unlikely(lkey == UINT32_MAX))
1088 /* Update the most recently used entry. */
1089 mr_ctrl->mru = mr_ctrl->head;
1090 /* Point to the next victim, the oldest. */
1091 mr_ctrl->head = (mr_ctrl->head + 1) % MLX5_MR_CACHE_N;
1096 * Bottom-half of LKey search on Rx.
1099 * Pointer to Rx queue structure.
1104 * Searched LKey on success, UINT32_MAX on no match.
1107 mlx5_rx_addr2mr_bh(struct mlx5_rxq_data *rxq, uintptr_t addr)
1109 struct mlx5_rxq_ctrl *rxq_ctrl =
1110 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
1111 struct mlx5_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
1112 struct mlx5_priv *priv = rxq_ctrl->priv;
1114 return mlx5_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
1118 * Bottom-half of LKey search on Tx.
1121 * Pointer to Tx queue structure.
1126 * Searched LKey on success, UINT32_MAX on no match.
1129 mlx5_tx_addr2mr_bh(struct mlx5_txq_data *txq, uintptr_t addr)
1131 struct mlx5_txq_ctrl *txq_ctrl =
1132 container_of(txq, struct mlx5_txq_ctrl, txq);
1133 struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1134 struct mlx5_priv *priv = txq_ctrl->priv;
1136 return mlx5_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
1140 * Bottom-half of LKey search on Tx. If it can't be searched in the memseg
1141 * list, register the mempool of the mbuf as externally allocated memory.
1144 * Pointer to Tx queue structure.
1149 * Searched LKey on success, UINT32_MAX on no match.
1152 mlx5_tx_mb2mr_bh(struct mlx5_txq_data *txq, struct rte_mbuf *mb)
1154 uintptr_t addr = (uintptr_t)mb->buf_addr;
1157 lkey = mlx5_tx_addr2mr_bh(txq, addr);
1158 if (lkey == UINT32_MAX && rte_errno == ENXIO) {
1159 /* Mempool may have externally allocated memory. */
1160 return mlx5_tx_update_ext_mp(txq, addr, mlx5_mb2mp(mb));
1166 * Flush all of the local cache entries.
1169 * Pointer to per-queue MR control structure.
1172 mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl)
1174 /* Reset the most-recently-used index. */
1176 /* Reset the linear search array. */
1178 memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache));
1179 /* Reset the B-tree table. */
1180 mr_ctrl->cache_bh.len = 1;
1181 mr_ctrl->cache_bh.overflow = 0;
1182 /* Update the generation number. */
1183 mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr;
1184 DRV_LOG(DEBUG, "mr_ctrl(%p): flushed, cur_gen=%d",
1185 (void *)mr_ctrl, mr_ctrl->cur_gen);
1189 * Creates a memory region for external memory, that is memory which is not
1190 * part of the DPDK memory segments.
1193 * Pointer to the ethernet device.
1195 * Starting virtual address of memory.
1197 * Length of memory segment being mapped.
1199 * Socket to allocate heap memory for the control structures.
1202 * Pointer to MR structure on success, NULL otherwise.
1204 static struct mlx5_mr *
1205 mlx5_create_mr_ext(struct rte_eth_dev *dev, uintptr_t addr, size_t len,
1208 struct mlx5_priv *priv = dev->data->dev_private;
1209 struct mlx5_mr *mr = NULL;
1211 mr = rte_zmalloc_socket(NULL,
1212 RTE_ALIGN_CEIL(sizeof(*mr),
1213 RTE_CACHE_LINE_SIZE),
1214 RTE_CACHE_LINE_SIZE, socket_id);
1217 mr->ibv_mr = mlx5_glue->reg_mr(priv->sh->pd, (void *)addr, len,
1218 IBV_ACCESS_LOCAL_WRITE);
1219 if (mr->ibv_mr == NULL) {
1221 "port %u fail to create a verbs MR for address (%p)",
1222 dev->data->port_id, (void *)addr);
1226 mr->msl = NULL; /* Mark it is external memory. */
1231 "port %u MR CREATED (%p) for external memory %p:\n"
1232 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
1233 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
1234 dev->data->port_id, (void *)mr, (void *)addr,
1235 addr, addr + len, rte_cpu_to_be_32(mr->ibv_mr->lkey),
1236 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
1241 * Called during rte_mempool_mem_iter() by mlx5_mr_update_ext_mp().
1243 * Externally allocated chunk is registered and a MR is created for the chunk.
1244 * The MR object is added to the global list. If memseg list of a MR object
1245 * (mr->msl) is null, the MR object can be regarded as externally allocated
1248 * Once external memory is registered, it should be static. If the memory is
1249 * freed and the virtual address range has different physical memory mapped
1250 * again, it may cause crash on device due to the wrong translation entry. PMD
1251 * can't track the free event of the external memory for now.
1254 mlx5_mr_update_ext_mp_cb(struct rte_mempool *mp, void *opaque,
1255 struct rte_mempool_memhdr *memhdr,
1256 unsigned mem_idx __rte_unused)
1258 struct mr_update_mp_data *data = opaque;
1259 struct rte_eth_dev *dev = data->dev;
1260 struct mlx5_priv *priv = dev->data->dev_private;
1261 struct mlx5_ibv_shared *sh = priv->sh;
1262 struct mlx5_mr_ctrl *mr_ctrl = data->mr_ctrl;
1263 struct mlx5_mr *mr = NULL;
1264 uintptr_t addr = (uintptr_t)memhdr->addr;
1265 size_t len = memhdr->len;
1266 struct mlx5_mr_cache entry;
1269 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
1270 /* If already registered, it should return. */
1271 rte_rwlock_read_lock(&sh->mr.rwlock);
1272 lkey = mr_lookup_dev(sh, &entry, addr);
1273 rte_rwlock_read_unlock(&sh->mr.rwlock);
1274 if (lkey != UINT32_MAX)
1276 DRV_LOG(DEBUG, "port %u register MR for chunk #%d of mempool (%s)",
1277 dev->data->port_id, mem_idx, mp->name);
1278 mr = mlx5_create_mr_ext(dev, addr, len, mp->socket_id);
1281 "port %u unable to allocate a new MR of"
1283 dev->data->port_id, mp->name);
1287 rte_rwlock_write_lock(&sh->mr.rwlock);
1288 LIST_INSERT_HEAD(&sh->mr.mr_list, mr, mr);
1289 /* Insert to the global cache table. */
1290 mr_insert_dev_cache(sh, mr);
1291 rte_rwlock_write_unlock(&sh->mr.rwlock);
1292 /* Insert to the local cache table */
1293 mlx5_mr_addr2mr_bh(dev, mr_ctrl, addr);
1297 * Finds the first ethdev that match the pci device.
1298 * The existence of multiple ethdev per pci device is only with representors.
1299 * On such case, it is enough to get only one of the ports as they all share
1300 * the same ibv context.
1303 * Pointer to the PCI device.
1306 * Pointer to the ethdev if found, NULL otherwise.
1308 static struct rte_eth_dev *
1309 pci_dev_to_eth_dev(struct rte_pci_device *pdev)
1313 RTE_ETH_FOREACH_DEV_OF(port_id, &pdev->device)
1314 return &rte_eth_devices[port_id];
1319 * DPDK callback to DMA map external memory to a PCI device.
1322 * Pointer to the PCI device.
1324 * Starting virtual address of memory to be mapped.
1326 * Starting IOVA address of memory to be mapped.
1328 * Length of memory segment being mapped.
1331 * 0 on success, negative value on error.
1334 mlx5_dma_map(struct rte_pci_device *pdev, void *addr,
1335 uint64_t iova __rte_unused, size_t len)
1337 struct rte_eth_dev *dev;
1339 struct mlx5_priv *priv;
1340 struct mlx5_ibv_shared *sh;
1342 dev = pci_dev_to_eth_dev(pdev);
1344 DRV_LOG(WARNING, "unable to find matching ethdev "
1345 "to PCI device %p", (void *)pdev);
1349 priv = dev->data->dev_private;
1350 mr = mlx5_create_mr_ext(dev, (uintptr_t)addr, len, SOCKET_ID_ANY);
1353 "port %u unable to dma map", dev->data->port_id);
1358 rte_rwlock_write_lock(&sh->mr.rwlock);
1359 LIST_INSERT_HEAD(&sh->mr.mr_list, mr, mr);
1360 /* Insert to the global cache table. */
1361 mr_insert_dev_cache(sh, mr);
1362 rte_rwlock_write_unlock(&sh->mr.rwlock);
1367 * DPDK callback to DMA unmap external memory to a PCI device.
1370 * Pointer to the PCI device.
1372 * Starting virtual address of memory to be unmapped.
1374 * Starting IOVA address of memory to be unmapped.
1376 * Length of memory segment being unmapped.
1379 * 0 on success, negative value on error.
1382 mlx5_dma_unmap(struct rte_pci_device *pdev, void *addr,
1383 uint64_t iova __rte_unused, size_t len __rte_unused)
1385 struct rte_eth_dev *dev;
1386 struct mlx5_priv *priv;
1387 struct mlx5_ibv_shared *sh;
1389 struct mlx5_mr_cache entry;
1391 dev = pci_dev_to_eth_dev(pdev);
1393 DRV_LOG(WARNING, "unable to find matching ethdev "
1394 "to PCI device %p", (void *)pdev);
1398 priv = dev->data->dev_private;
1400 rte_rwlock_read_lock(&sh->mr.rwlock);
1401 mr = mr_lookup_dev_list(sh, &entry, (uintptr_t)addr);
1403 rte_rwlock_read_unlock(&sh->mr.rwlock);
1404 DRV_LOG(WARNING, "address 0x%" PRIxPTR " wasn't registered "
1405 "to PCI device %p", (uintptr_t)addr,
1410 LIST_REMOVE(mr, mr);
1411 LIST_INSERT_HEAD(&sh->mr.mr_free_list, mr, mr);
1412 DEBUG("port %u remove MR(%p) from list", dev->data->port_id,
1414 mr_rebuild_dev_cache(sh);
1416 * Flush local caches by propagating invalidation across cores.
1417 * rte_smp_wmb() is enough to synchronize this event. If one of
1418 * freed memsegs is seen by other core, that means the memseg
1419 * has been allocated by allocator, which will come after this
1420 * free call. Therefore, this store instruction (incrementing
1421 * generation below) will be guaranteed to be seen by other core
1422 * before the core sees the newly allocated memory.
1425 DEBUG("broadcasting local cache flush, gen=%d", sh->mr.dev_gen);
1427 rte_rwlock_read_unlock(&sh->mr.rwlock);
1432 * Register MR for entire memory chunks in a Mempool having externally allocated
1433 * memory and fill in local cache.
1436 * Pointer to Ethernet device.
1438 * Pointer to per-queue MR control structure.
1440 * Pointer to registering Mempool.
1443 * 0 on success, -1 on failure.
1446 mlx5_mr_update_ext_mp(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
1447 struct rte_mempool *mp)
1449 struct mr_update_mp_data data = {
1455 rte_mempool_mem_iter(mp, mlx5_mr_update_ext_mp_cb, &data);
1460 * Register MR entire memory chunks in a Mempool having externally allocated
1461 * memory and search LKey of the address to return.
1464 * Pointer to Ethernet device.
1468 * Pointer to registering Mempool where addr belongs.
1471 * LKey for address on success, UINT32_MAX on failure.
1474 mlx5_tx_update_ext_mp(struct mlx5_txq_data *txq, uintptr_t addr,
1475 struct rte_mempool *mp)
1477 struct mlx5_txq_ctrl *txq_ctrl =
1478 container_of(txq, struct mlx5_txq_ctrl, txq);
1479 struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1480 struct mlx5_priv *priv = txq_ctrl->priv;
1482 if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
1484 "port %u using address (%p) from unregistered mempool"
1485 " having externally allocated memory"
1486 " in secondary process, please create mempool"
1487 " prior to rte_eth_dev_start()",
1488 PORT_ID(priv), (void *)addr);
1491 mlx5_mr_update_ext_mp(ETH_DEV(priv), mr_ctrl, mp);
1492 return mlx5_tx_addr2mr_bh(txq, addr);
1495 /* Called during rte_mempool_mem_iter() by mlx5_mr_update_mp(). */
1497 mlx5_mr_update_mp_cb(struct rte_mempool *mp __rte_unused, void *opaque,
1498 struct rte_mempool_memhdr *memhdr,
1499 unsigned mem_idx __rte_unused)
1501 struct mr_update_mp_data *data = opaque;
1504 /* Stop iteration if failed in the previous walk. */
1507 /* Register address of the chunk and update local caches. */
1508 lkey = mlx5_mr_addr2mr_bh(data->dev, data->mr_ctrl,
1509 (uintptr_t)memhdr->addr);
1510 if (lkey == UINT32_MAX)
1515 * Register entire memory chunks in a Mempool.
1518 * Pointer to Ethernet device.
1520 * Pointer to per-queue MR control structure.
1522 * Pointer to registering Mempool.
1525 * 0 on success, -1 on failure.
1528 mlx5_mr_update_mp(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
1529 struct rte_mempool *mp)
1531 struct mr_update_mp_data data = {
1537 rte_mempool_mem_iter(mp, mlx5_mr_update_mp_cb, &data);
1538 if (data.ret < 0 && rte_errno == ENXIO) {
1539 /* Mempool may have externally allocated memory. */
1540 return mlx5_mr_update_ext_mp(dev, mr_ctrl, mp);
1546 * Dump all the created MRs and the global cache entries.
1549 * Pointer to Ethernet device shared context.
1552 mlx5_mr_dump_dev(struct mlx5_ibv_shared *sh __rte_unused)
1559 rte_rwlock_read_lock(&sh->mr.rwlock);
1560 /* Iterate all the existing MRs. */
1561 LIST_FOREACH(mr, &sh->mr.mr_list, mr) {
1564 DEBUG("device %s MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u",
1565 sh->ibdev_name, mr_n++,
1566 rte_cpu_to_be_32(mr->ibv_mr->lkey),
1567 mr->ms_n, mr->ms_bmp_n);
1570 for (n = 0; n < mr->ms_bmp_n; ) {
1571 struct mlx5_mr_cache ret = { 0, };
1573 n = mr_find_next_chunk(mr, &ret, n);
1576 DEBUG(" chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")",
1577 chunk_n++, ret.start, ret.end);
1580 DEBUG("device %s dumping global cache", sh->ibdev_name);
1581 mlx5_mr_btree_dump(&sh->mr.cache);
1582 rte_rwlock_read_unlock(&sh->mr.rwlock);
1587 * Release all the created MRs and resources for shared device context.
1591 * Pointer to Ethernet device shared context.
1594 mlx5_mr_release(struct mlx5_ibv_shared *sh)
1596 struct mlx5_mr *mr_next;
1598 if (rte_log_get_level(mlx5_logtype) == RTE_LOG_DEBUG)
1599 mlx5_mr_dump_dev(sh);
1600 rte_rwlock_write_lock(&sh->mr.rwlock);
1601 /* Detach from MR list and move to free list. */
1602 mr_next = LIST_FIRST(&sh->mr.mr_list);
1603 while (mr_next != NULL) {
1604 struct mlx5_mr *mr = mr_next;
1606 mr_next = LIST_NEXT(mr, mr);
1607 LIST_REMOVE(mr, mr);
1608 LIST_INSERT_HEAD(&sh->mr.mr_free_list, mr, mr);
1610 LIST_INIT(&sh->mr.mr_list);
1611 /* Free global cache. */
1612 mlx5_mr_btree_free(&sh->mr.cache);
1613 rte_rwlock_write_unlock(&sh->mr.rwlock);
1614 /* Free all remaining MRs. */
1615 mlx5_mr_garbage_collect(sh);