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 = { 0, };
359 /* Find a contiguous chunk and advance the index. */
360 n = mr_find_next_chunk(mr, &entry, n);
363 if (mr_btree_insert(&priv->mr.cache, &entry) < 0) {
365 * Overflowed, but the global table cannot be expanded
366 * because of deadlock.
375 * Look up address in the original global MR list.
378 * Pointer to Ethernet device.
380 * Pointer to returning MR cache entry. If no match, this will not be updated.
385 * Found MR on match, NULL otherwise.
387 static struct mlx4_mr *
388 mr_lookup_dev_list(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
391 struct priv *priv = dev->data->dev_private;
394 /* Iterate all the existing MRs. */
395 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
400 for (n = 0; n < mr->ms_bmp_n; ) {
401 struct mlx4_mr_cache ret = { 0, };
403 n = mr_find_next_chunk(mr, &ret, n);
404 if (addr >= ret.start && addr < ret.end) {
415 * Look up address on device.
418 * Pointer to Ethernet device.
420 * Pointer to returning MR cache entry. If no match, this will not be updated.
425 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
428 mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
431 struct priv *priv = dev->data->dev_private;
433 uint32_t lkey = UINT32_MAX;
437 * If the global cache has overflowed since it failed to expand the
438 * B-tree table, it can't have all the existing MRs. Then, the address
439 * has to be searched by traversing the original MR list instead, which
440 * is very slow path. Otherwise, the global cache is all inclusive.
442 if (!unlikely(priv->mr.cache.overflow)) {
443 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
444 if (lkey != UINT32_MAX)
445 *entry = (*priv->mr.cache.table)[idx];
447 /* Falling back to the slowest path. */
448 mr = mr_lookup_dev_list(dev, entry, addr);
452 assert(lkey == UINT32_MAX || (addr >= entry->start &&
458 * Free MR resources. MR lock must not be held to avoid a deadlock. rte_free()
459 * can raise memory free event and the callback function will spin on the lock.
462 * Pointer to MR to free.
465 mr_free(struct mlx4_mr *mr)
469 DEBUG("freeing MR(%p):", (void *)mr);
470 if (mr->ibv_mr != NULL)
471 claim_zero(mlx4_glue->dereg_mr(mr->ibv_mr));
472 if (mr->ms_bmp != NULL)
473 rte_bitmap_free(mr->ms_bmp);
478 * Releass resources of detached MR having no online entry.
481 * Pointer to Ethernet device.
484 mlx4_mr_garbage_collect(struct rte_eth_dev *dev)
486 struct priv *priv = dev->data->dev_private;
487 struct mlx4_mr *mr_next;
488 struct mlx4_mr_list free_list = LIST_HEAD_INITIALIZER(free_list);
491 * MR can't be freed with holding the lock because rte_free() could call
492 * memory free callback function. This will be a deadlock situation.
494 rte_rwlock_write_lock(&priv->mr.rwlock);
495 /* Detach the whole free list and release it after unlocking. */
496 free_list = priv->mr.mr_free_list;
497 LIST_INIT(&priv->mr.mr_free_list);
498 rte_rwlock_write_unlock(&priv->mr.rwlock);
499 /* Release resources. */
500 mr_next = LIST_FIRST(&free_list);
501 while (mr_next != NULL) {
502 struct mlx4_mr *mr = mr_next;
504 mr_next = LIST_NEXT(mr, mr);
509 /* Called during rte_memseg_contig_walk() by mlx4_mr_create(). */
511 mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl,
512 const struct rte_memseg *ms, size_t len, void *arg)
514 struct mr_find_contig_memsegs_data *data = arg;
516 if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len)
518 /* Found, save it and stop walking. */
519 data->start = ms->addr_64;
520 data->end = ms->addr_64 + len;
526 * Create a new global Memroy Region (MR) for a missing virtual address.
527 * Register entire virtually contiguous memory chunk around the address.
530 * Pointer to Ethernet device.
532 * Pointer to returning MR cache entry, found in the global cache or newly
533 * created. If failed to create one, this will not be updated.
535 * Target virtual address to register.
538 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
541 mlx4_mr_create(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
544 struct priv *priv = dev->data->dev_private;
545 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
546 const struct rte_memseg_list *msl;
547 const struct rte_memseg *ms;
548 struct mlx4_mr *mr = NULL;
553 int ms_idx_shift = -1;
555 struct mr_find_contig_memsegs_data data = {
558 struct mr_find_contig_memsegs_data data_re;
560 DEBUG("port %u creating a MR using address (%p)",
561 dev->data->port_id, (void *)addr);
563 * Release detached MRs if any. This can't be called with holding either
564 * memory_hotplug_lock or priv->mr.rwlock. MRs on the free list have
565 * been detached by the memory free event but it couldn't be released
566 * inside the callback due to deadlock. As a result, releasing resources
567 * is quite opportunistic.
569 mlx4_mr_garbage_collect(dev);
571 * Find out a contiguous virtual address chunk in use, to which the
572 * given address belongs, in order to register maximum range. In the
573 * best case where mempools are not dynamically recreated and
574 * '--socket-mem' is specified as an EAL option, it is very likely to
575 * have only one MR(LKey) per a socket and per a hugepage-size even
576 * though the system memory is highly fragmented.
578 if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) {
579 WARN("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 WARN("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 WARN("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 WARN("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 * mlx4_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 mlx4_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 * mlx4_alloc_buf_extern() which eventually calls rte_malloc_socket()
717 * through mlx4_alloc_verbs_buf().
719 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)data.start, len,
720 IBV_ACCESS_LOCAL_WRITE);
721 if (mr->ibv_mr == NULL) {
722 WARN("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 mlx4_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 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 mlx4_mr_garbage_collect().
788 * The global cache must be rebuilt if there's any change and this event has to
789 * be propagated to dataplane threads to flush the local caches.
792 * Pointer to Ethernet device.
794 * Address of freed memory.
796 * Size of freed memory.
799 mlx4_mr_mem_event_free_cb(struct rte_eth_dev *dev, const void *addr, size_t len)
801 struct priv *priv = dev->data->dev_private;
802 const struct rte_memseg_list *msl;
808 DEBUG("port %u free callback: addr=%p, len=%zu",
809 dev->data->port_id, addr, len);
810 msl = rte_mem_virt2memseg_list(addr);
811 /* addr and len must be page-aligned. */
812 assert((uintptr_t)addr == RTE_ALIGN((uintptr_t)addr, msl->page_sz));
813 assert(len == RTE_ALIGN(len, msl->page_sz));
814 ms_n = len / msl->page_sz;
815 rte_rwlock_write_lock(&priv->mr.rwlock);
816 /* Clear bits of freed memsegs from MR. */
817 for (i = 0; i < ms_n; ++i) {
818 const struct rte_memseg *ms;
819 struct mlx4_mr_cache entry;
824 /* Find MR having this memseg. */
825 start = (uintptr_t)addr + i * msl->page_sz;
826 mr = mr_lookup_dev_list(dev, &entry, start);
829 assert(mr->msl); /* Can't be external memory. */
830 ms = rte_mem_virt2memseg((void *)start, msl);
832 assert(msl->page_sz == ms->hugepage_sz);
833 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
834 pos = ms_idx - mr->ms_base_idx;
835 assert(rte_bitmap_get(mr->ms_bmp, pos));
836 assert(pos < mr->ms_bmp_n);
837 DEBUG("port %u MR(%p): clear bitmap[%u] for addr %p",
838 dev->data->port_id, (void *)mr, pos, (void *)start);
839 rte_bitmap_clear(mr->ms_bmp, pos);
840 if (--mr->ms_n == 0) {
842 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
843 DEBUG("port %u remove MR(%p) from list",
844 dev->data->port_id, (void *)mr);
847 * MR is fragmented or will be freed. the global cache must be
853 mr_rebuild_dev_cache(dev);
855 * Flush local caches by propagating invalidation across cores.
856 * rte_smp_wmb() is enough to synchronize this event. If one of
857 * freed memsegs is seen by other core, that means the memseg
858 * has been allocated by allocator, which will come after this
859 * free call. Therefore, this store instruction (incrementing
860 * generation below) will be guaranteed to be seen by other core
861 * before the core sees the newly allocated memory.
864 DEBUG("broadcasting local cache flush, gen=%d",
868 rte_rwlock_write_unlock(&priv->mr.rwlock);
871 mlx4_mr_dump_dev(dev);
876 * Callback for memory event.
886 mlx4_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
887 size_t len, void *arg __rte_unused)
891 switch (event_type) {
892 case RTE_MEM_EVENT_FREE:
893 rte_rwlock_read_lock(&mlx4_mem_event_rwlock);
894 /* Iterate all the existing mlx4 devices. */
895 LIST_FOREACH(priv, &mlx4_mem_event_cb_list, mem_event_cb)
896 mlx4_mr_mem_event_free_cb(priv->dev, addr, len);
897 rte_rwlock_read_unlock(&mlx4_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 mlx4_mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
924 struct mlx4_mr_cache *entry, uintptr_t addr)
926 struct priv *priv = dev->data->dev_private;
927 struct mlx4_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 = mlx4_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 mlx4_mr_addr2mr_bh(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
984 /* Victim in top-half cache to replace with new entry. */
985 struct mlx4_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 = mlx4_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) % MLX4_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 mlx4_rx_addr2mr_bh(struct rxq *rxq, uintptr_t addr)
1023 struct mlx4_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
1024 struct priv *priv = rxq->priv;
1026 DEBUG("Rx queue %u: miss on top-half, mru=%u, head=%u, addr=%p",
1027 rxq->stats.idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr);
1028 return mlx4_mr_addr2mr_bh(priv->dev, mr_ctrl, addr);
1032 * Bottom-half of LKey search on Tx.
1035 * Pointer to Tx queue structure.
1040 * Searched LKey on success, UINT32_MAX on no match.
1043 mlx4_tx_addr2mr_bh(struct txq *txq, uintptr_t addr)
1045 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1046 struct priv *priv = txq->priv;
1048 DEBUG("Tx queue %u: miss on top-half, mru=%u, head=%u, addr=%p",
1049 txq->stats.idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr);
1050 return mlx4_mr_addr2mr_bh(priv->dev, mr_ctrl, addr);
1054 * Flush all of the local cache entries.
1057 * Pointer to per-queue MR control structure.
1060 mlx4_mr_flush_local_cache(struct mlx4_mr_ctrl *mr_ctrl)
1062 /* Reset the most-recently-used index. */
1064 /* Reset the linear search array. */
1066 memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache));
1067 /* Reset the B-tree table. */
1068 mr_ctrl->cache_bh.len = 1;
1069 mr_ctrl->cache_bh.overflow = 0;
1070 /* Update the generation number. */
1071 mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr;
1072 DEBUG("mr_ctrl(%p): flushed, cur_gen=%d",
1073 (void *)mr_ctrl, mr_ctrl->cur_gen);
1077 * Called during rte_mempool_mem_iter() by mlx4_mr_update_ext_mp().
1079 * Externally allocated chunk is registered and a MR is created for the chunk.
1080 * The MR object is added to the global list. If memseg list of a MR object
1081 * (mr->msl) is null, the MR object can be regarded as externally allocated
1084 * Once external memory is registered, it should be static. If the memory is
1085 * freed and the virtual address range has different physical memory mapped
1086 * again, it may cause crash on device due to the wrong translation entry. PMD
1087 * can't track the free event of the external memory for now.
1090 mlx4_mr_update_ext_mp_cb(struct rte_mempool *mp, void *opaque,
1091 struct rte_mempool_memhdr *memhdr,
1092 unsigned mem_idx __rte_unused)
1094 struct mr_update_mp_data *data = opaque;
1095 struct rte_eth_dev *dev = data->dev;
1096 struct priv *priv = dev->data->dev_private;
1097 struct mlx4_mr_ctrl *mr_ctrl = data->mr_ctrl;
1098 struct mlx4_mr *mr = NULL;
1099 uintptr_t addr = (uintptr_t)memhdr->addr;
1100 size_t len = memhdr->len;
1101 struct mlx4_mr_cache entry;
1104 /* If already registered, it should return. */
1105 rte_rwlock_read_lock(&priv->mr.rwlock);
1106 lkey = mr_lookup_dev(dev, &entry, addr);
1107 rte_rwlock_read_unlock(&priv->mr.rwlock);
1108 if (lkey != UINT32_MAX)
1110 mr = rte_zmalloc_socket(NULL,
1111 RTE_ALIGN_CEIL(sizeof(*mr),
1112 RTE_CACHE_LINE_SIZE),
1113 RTE_CACHE_LINE_SIZE, mp->socket_id);
1115 WARN("port %u unable to allocate memory for a new MR of"
1117 dev->data->port_id, mp->name);
1121 DEBUG("port %u register MR for chunk #%d of mempool (%s)",
1122 dev->data->port_id, mem_idx, mp->name);
1123 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)addr, len,
1124 IBV_ACCESS_LOCAL_WRITE);
1125 if (mr->ibv_mr == NULL) {
1126 WARN("port %u fail to create a verbs MR for address (%p)",
1127 dev->data->port_id, (void *)addr);
1132 mr->msl = NULL; /* Mark it is external memory. */
1136 rte_rwlock_write_lock(&priv->mr.rwlock);
1137 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
1138 DEBUG("port %u MR CREATED (%p) for external memory %p:\n"
1139 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
1140 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
1141 dev->data->port_id, (void *)mr, (void *)addr,
1142 addr, addr + len, rte_cpu_to_be_32(mr->ibv_mr->lkey),
1143 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
1144 /* Insert to the global cache table. */
1145 mr_insert_dev_cache(dev, mr);
1146 rte_rwlock_write_unlock(&priv->mr.rwlock);
1147 /* Insert to the local cache table */
1148 mlx4_mr_addr2mr_bh(dev, mr_ctrl, addr);
1152 * Register MR for entire memory chunks in a Mempool having externally allocated
1153 * memory and fill in local cache.
1156 * Pointer to Ethernet device.
1158 * Pointer to per-queue MR control structure.
1160 * Pointer to registering Mempool.
1163 * 0 on success, -1 on failure.
1166 mlx4_mr_update_ext_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1167 struct rte_mempool *mp)
1169 struct mr_update_mp_data data = {
1175 rte_mempool_mem_iter(mp, mlx4_mr_update_ext_mp_cb, &data);
1180 * Register MR entire memory chunks in a Mempool having externally allocated
1181 * memory and search LKey of the address to return.
1184 * Pointer to Ethernet device.
1188 * Pointer to registering Mempool where addr belongs.
1191 * LKey for address on success, UINT32_MAX on failure.
1194 mlx4_tx_update_ext_mp(struct txq *txq, uintptr_t addr, struct rte_mempool *mp)
1196 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1197 struct priv *priv = txq->priv;
1199 mlx4_mr_update_ext_mp(priv->dev, mr_ctrl, mp);
1200 return mlx4_tx_addr2mr_bh(txq, addr);
1203 /* Called during rte_mempool_mem_iter() by mlx4_mr_update_mp(). */
1205 mlx4_mr_update_mp_cb(struct rte_mempool *mp __rte_unused, void *opaque,
1206 struct rte_mempool_memhdr *memhdr,
1207 unsigned mem_idx __rte_unused)
1209 struct mr_update_mp_data *data = opaque;
1212 /* Stop iteration if failed in the previous walk. */
1215 /* Register address of the chunk and update local caches. */
1216 lkey = mlx4_mr_addr2mr_bh(data->dev, data->mr_ctrl,
1217 (uintptr_t)memhdr->addr);
1218 if (lkey == UINT32_MAX)
1223 * Register entire memory chunks in a Mempool.
1226 * Pointer to Ethernet device.
1228 * Pointer to per-queue MR control structure.
1230 * Pointer to registering Mempool.
1233 * 0 on success, -1 on failure.
1236 mlx4_mr_update_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1237 struct rte_mempool *mp)
1239 struct mr_update_mp_data data = {
1245 rte_mempool_mem_iter(mp, mlx4_mr_update_mp_cb, &data);
1246 if (data.ret < 0 && rte_errno == ENXIO) {
1247 /* Mempool may have externally allocated memory. */
1248 return mlx4_mr_update_ext_mp(dev, mr_ctrl, mp);
1255 * Dump all the created MRs and the global cache entries.
1258 * Pointer to Ethernet device.
1261 mlx4_mr_dump_dev(struct rte_eth_dev *dev)
1263 struct priv *priv = dev->data->dev_private;
1268 rte_rwlock_read_lock(&priv->mr.rwlock);
1269 /* Iterate all the existing MRs. */
1270 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
1273 DEBUG("port %u MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u",
1274 dev->data->port_id, mr_n++,
1275 rte_cpu_to_be_32(mr->ibv_mr->lkey),
1276 mr->ms_n, mr->ms_bmp_n);
1279 for (n = 0; n < mr->ms_bmp_n; ) {
1280 struct mlx4_mr_cache ret = { 0, };
1282 n = mr_find_next_chunk(mr, &ret, n);
1285 DEBUG(" chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")",
1286 chunk_n++, ret.start, ret.end);
1289 DEBUG("port %u dumping global cache", dev->data->port_id);
1290 mlx4_mr_btree_dump(&priv->mr.cache);
1291 rte_rwlock_read_unlock(&priv->mr.rwlock);
1296 * Release all the created MRs and resources. Remove device from memory callback
1300 * Pointer to Ethernet device.
1303 mlx4_mr_release(struct rte_eth_dev *dev)
1305 struct priv *priv = dev->data->dev_private;
1306 struct mlx4_mr *mr_next = LIST_FIRST(&priv->mr.mr_list);
1308 /* Remove from memory callback device list. */
1309 rte_rwlock_write_lock(&mlx4_mem_event_rwlock);
1310 LIST_REMOVE(priv, mem_event_cb);
1311 rte_rwlock_write_unlock(&mlx4_mem_event_rwlock);
1313 mlx4_mr_dump_dev(dev);
1315 rte_rwlock_write_lock(&priv->mr.rwlock);
1316 /* Detach from MR list and move to free list. */
1317 while (mr_next != NULL) {
1318 struct mlx4_mr *mr = mr_next;
1320 mr_next = LIST_NEXT(mr, mr);
1321 LIST_REMOVE(mr, mr);
1322 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
1324 LIST_INIT(&priv->mr.mr_list);
1325 /* Free global cache. */
1326 mlx4_mr_btree_free(&priv->mr.cache);
1327 rte_rwlock_write_unlock(&priv->mr.rwlock);
1328 /* Free all remaining MRs. */
1329 mlx4_mr_garbage_collect(dev);