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.
17 /* Verbs headers do not support -pedantic. */
19 #pragma GCC diagnostic ignored "-Wpedantic"
21 #include <infiniband/verbs.h>
23 #pragma GCC diagnostic error "-Wpedantic"
26 #include <rte_branch_prediction.h>
27 #include <rte_common.h>
28 #include <rte_eal_memconfig.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;
115 MLX4_ASSERT(bt != NULL);
116 lkp_tbl = *bt->table;
118 /* First entry must be NULL for comparison. */
119 MLX4_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 MLX4_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;
158 MLX4_ASSERT(bt != NULL);
159 MLX4_ASSERT(bt->len <= bt->size);
160 MLX4_ASSERT(bt->len > 0);
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));
244 #ifdef RTE_LIBRTE_MLX4_DEBUG
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 MLX4_ASSERT(mr->ms_bmp_n == 1);
297 MLX4_ASSERT(mr->ms_n == 1);
298 MLX4_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 MLX4_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 mlx4_priv *priv = dev->data->dev_private;
354 DEBUG("port %u inserting MR(%p) to global cache",
355 dev->data->port_id, (void *)mr);
356 for (n = 0; n < mr->ms_bmp_n; ) {
357 struct mlx4_mr_cache entry;
359 memset(&entry, 0, sizeof(entry));
360 /* Find a contiguous chunk and advance the index. */
361 n = mr_find_next_chunk(mr, &entry, n);
364 if (mr_btree_insert(&priv->mr.cache, &entry) < 0) {
366 * Overflowed, but the global table cannot be expanded
367 * because of deadlock.
376 * Look up address in the original global MR list.
379 * Pointer to Ethernet device.
381 * Pointer to returning MR cache entry. If no match, this will not be updated.
386 * Found MR on match, NULL otherwise.
388 static struct mlx4_mr *
389 mr_lookup_dev_list(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
392 struct mlx4_priv *priv = dev->data->dev_private;
395 /* Iterate all the existing MRs. */
396 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
401 for (n = 0; n < mr->ms_bmp_n; ) {
402 struct mlx4_mr_cache ret;
404 memset(&ret, 0, sizeof(ret));
405 n = mr_find_next_chunk(mr, &ret, n);
406 if (addr >= ret.start && addr < ret.end) {
417 * Look up address on device.
420 * Pointer to Ethernet device.
422 * Pointer to returning MR cache entry. If no match, this will not be updated.
427 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
430 mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
433 struct mlx4_priv *priv = dev->data->dev_private;
435 uint32_t lkey = UINT32_MAX;
439 * If the global cache has overflowed since it failed to expand the
440 * B-tree table, it can't have all the existing MRs. Then, the address
441 * has to be searched by traversing the original MR list instead, which
442 * is very slow path. Otherwise, the global cache is all inclusive.
444 if (!unlikely(priv->mr.cache.overflow)) {
445 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
446 if (lkey != UINT32_MAX)
447 *entry = (*priv->mr.cache.table)[idx];
449 /* Falling back to the slowest path. */
450 mr = mr_lookup_dev_list(dev, entry, addr);
454 MLX4_ASSERT(lkey == UINT32_MAX || (addr >= entry->start &&
460 * Free MR resources. MR lock must not be held to avoid a deadlock. rte_free()
461 * can raise memory free event and the callback function will spin on the lock.
464 * Pointer to MR to free.
467 mr_free(struct mlx4_mr *mr)
471 DEBUG("freeing MR(%p):", (void *)mr);
472 if (mr->ibv_mr != NULL)
473 claim_zero(mlx4_glue->dereg_mr(mr->ibv_mr));
474 if (mr->ms_bmp != NULL)
475 rte_bitmap_free(mr->ms_bmp);
480 * Release resources of detached MR having no online entry.
483 * Pointer to Ethernet device.
486 mlx4_mr_garbage_collect(struct rte_eth_dev *dev)
488 struct mlx4_priv *priv = dev->data->dev_private;
489 struct mlx4_mr *mr_next;
490 struct mlx4_mr_list free_list = LIST_HEAD_INITIALIZER(free_list);
492 /* Must be called from the primary process. */
493 MLX4_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
495 * MR can't be freed with holding the lock because rte_free() could call
496 * memory free callback function. This will be a deadlock situation.
498 rte_rwlock_write_lock(&priv->mr.rwlock);
499 /* Detach the whole free list and release it after unlocking. */
500 free_list = priv->mr.mr_free_list;
501 LIST_INIT(&priv->mr.mr_free_list);
502 rte_rwlock_write_unlock(&priv->mr.rwlock);
503 /* Release resources. */
504 mr_next = LIST_FIRST(&free_list);
505 while (mr_next != NULL) {
506 struct mlx4_mr *mr = mr_next;
508 mr_next = LIST_NEXT(mr, mr);
513 /* Called during rte_memseg_contig_walk() by mlx4_mr_create(). */
515 mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl,
516 const struct rte_memseg *ms, size_t len, void *arg)
518 struct mr_find_contig_memsegs_data *data = arg;
520 if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len)
522 /* Found, save it and stop walking. */
523 data->start = ms->addr_64;
524 data->end = ms->addr_64 + len;
530 * Create a new global Memory Region (MR) for a missing virtual address.
531 * This API should be called on a secondary process, then a request is sent to
532 * the primary process in order to create a MR for the address. As the global MR
533 * list is on the shared memory, following LKey lookup should succeed unless the
537 * Pointer to Ethernet device.
539 * Pointer to returning MR cache entry, found in the global cache or newly
540 * created. If failed to create one, this will not be updated.
542 * Target virtual address to register.
545 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
548 mlx4_mr_create_secondary(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
551 struct mlx4_priv *priv = dev->data->dev_private;
554 DEBUG("port %u requesting MR creation for address (%p)",
555 dev->data->port_id, (void *)addr);
556 ret = mlx4_mp_req_mr_create(dev, addr);
558 DEBUG("port %u fail to request MR creation for address (%p)",
559 dev->data->port_id, (void *)addr);
562 rte_rwlock_read_lock(&priv->mr.rwlock);
563 /* Fill in output data. */
564 mr_lookup_dev(dev, entry, addr);
565 /* Lookup can't fail. */
566 MLX4_ASSERT(entry->lkey != UINT32_MAX);
567 rte_rwlock_read_unlock(&priv->mr.rwlock);
568 DEBUG("port %u MR CREATED by primary process for %p:\n"
569 " [0x%" PRIxPTR ", 0x%" PRIxPTR "), lkey=0x%x",
570 dev->data->port_id, (void *)addr,
571 entry->start, entry->end, entry->lkey);
576 * Create a new global Memory Region (MR) for a missing virtual address.
577 * Register entire virtually contiguous memory chunk around the address.
578 * This must be called from the primary process.
581 * Pointer to Ethernet device.
583 * Pointer to returning MR cache entry, found in the global cache or newly
584 * created. If failed to create one, this will not be updated.
586 * Target virtual address to register.
589 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
592 mlx4_mr_create_primary(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
595 struct mlx4_priv *priv = dev->data->dev_private;
596 const struct rte_memseg_list *msl;
597 const struct rte_memseg *ms;
598 struct mlx4_mr *mr = NULL;
603 int ms_idx_shift = -1;
605 struct mr_find_contig_memsegs_data data = {
608 struct mr_find_contig_memsegs_data data_re;
610 DEBUG("port %u creating a MR using address (%p)",
611 dev->data->port_id, (void *)addr);
613 * Release detached MRs if any. This can't be called with holding either
614 * memory_hotplug_lock or priv->mr.rwlock. MRs on the free list have
615 * been detached by the memory free event but it couldn't be released
616 * inside the callback due to deadlock. As a result, releasing resources
617 * is quite opportunistic.
619 mlx4_mr_garbage_collect(dev);
621 * If enabled, find out a contiguous virtual address chunk in use, to
622 * which the given address belongs, in order to register maximum range.
623 * In the best case where mempools are not dynamically recreated and
624 * '--socket-mem' is specified as an EAL option, it is very likely to
625 * have only one MR(LKey) per a socket and per a hugepage-size even
626 * though the system memory is highly fragmented. As the whole memory
627 * chunk will be pinned by kernel, it can't be reused unless entire
628 * chunk is freed from EAL.
630 * If disabled, just register one memseg (page). Then, memory
631 * consumption will be minimized but it may drop performance if there
632 * are many MRs to lookup on the datapath.
634 if (!priv->mr_ext_memseg_en) {
635 data.msl = rte_mem_virt2memseg_list((void *)addr);
636 data.start = RTE_ALIGN_FLOOR(addr, data.msl->page_sz);
637 data.end = data.start + data.msl->page_sz;
638 } else if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) {
639 WARN("port %u unable to find virtually contiguous"
640 " chunk for address (%p)."
641 " rte_memseg_contig_walk() failed.",
642 dev->data->port_id, (void *)addr);
647 /* Addresses must be page-aligned. */
648 MLX4_ASSERT(rte_is_aligned((void *)data.start, data.msl->page_sz));
649 MLX4_ASSERT(rte_is_aligned((void *)data.end, data.msl->page_sz));
651 ms = rte_mem_virt2memseg((void *)data.start, msl);
652 len = data.end - data.start;
653 MLX4_ASSERT(msl->page_sz == ms->hugepage_sz);
654 /* Number of memsegs in the range. */
655 ms_n = len / msl->page_sz;
656 DEBUG("port %u extending %p to [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
657 " page_sz=0x%" PRIx64 ", ms_n=%u",
658 dev->data->port_id, (void *)addr,
659 data.start, data.end, msl->page_sz, ms_n);
660 /* Size of memory for bitmap. */
661 bmp_size = rte_bitmap_get_memory_footprint(ms_n);
662 mr = rte_zmalloc_socket(NULL,
663 RTE_ALIGN_CEIL(sizeof(*mr),
664 RTE_CACHE_LINE_SIZE) +
666 RTE_CACHE_LINE_SIZE, msl->socket_id);
668 WARN("port %u unable to allocate memory for a new MR of"
670 dev->data->port_id, (void *)addr);
676 * Save the index of the first memseg and initialize memseg bitmap. To
677 * see if a memseg of ms_idx in the memseg-list is still valid, check:
678 * rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx)
680 mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
681 bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE);
682 mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size);
683 if (mr->ms_bmp == NULL) {
684 WARN("port %u unable to initialize bitmap for a new MR of"
686 dev->data->port_id, (void *)addr);
691 * Should recheck whether the extended contiguous chunk is still valid.
692 * Because memory_hotplug_lock can't be held if there's any memory
693 * related calls in a critical path, resource allocation above can't be
694 * locked. If the memory has been changed at this point, try again with
695 * just single page. If not, go on with the big chunk atomically from
698 rte_mcfg_mem_read_lock();
700 if (len > msl->page_sz &&
701 !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) {
702 WARN("port %u unable to find virtually contiguous"
703 " chunk for address (%p)."
704 " rte_memseg_contig_walk() failed.",
705 dev->data->port_id, (void *)addr);
709 if (data.start != data_re.start || data.end != data_re.end) {
711 * The extended contiguous chunk has been changed. Try again
712 * with single memseg instead.
714 data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz);
715 data.end = data.start + msl->page_sz;
716 rte_mcfg_mem_read_unlock();
718 goto alloc_resources;
720 MLX4_ASSERT(data.msl == data_re.msl);
721 rte_rwlock_write_lock(&priv->mr.rwlock);
723 * Check the address is really missing. If other thread already created
724 * one or it is not found due to overflow, abort and return.
726 if (mr_lookup_dev(dev, entry, addr) != UINT32_MAX) {
728 * Insert to the global cache table. It may fail due to
729 * low-on-memory. Then, this entry will have to be searched
732 mr_btree_insert(&priv->mr.cache, entry);
733 DEBUG("port %u found MR for %p on final lookup, abort",
734 dev->data->port_id, (void *)addr);
735 rte_rwlock_write_unlock(&priv->mr.rwlock);
736 rte_mcfg_mem_read_unlock();
738 * Must be unlocked before calling rte_free() because
739 * mlx4_mr_mem_event_free_cb() can be called inside.
745 * Trim start and end addresses for verbs MR. Set bits for registering
746 * memsegs but exclude already registered ones. Bitmap can be
749 for (n = 0; n < ms_n; ++n) {
751 struct mlx4_mr_cache ret;
753 memset(&ret, 0, sizeof(ret));
754 start = data_re.start + n * msl->page_sz;
755 /* Exclude memsegs already registered by other MRs. */
756 if (mr_lookup_dev(dev, &ret, start) == UINT32_MAX) {
758 * Start from the first unregistered memseg in the
761 if (ms_idx_shift == -1) {
762 mr->ms_base_idx += n;
766 data.end = start + msl->page_sz;
767 rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift);
771 len = data.end - data.start;
772 mr->ms_bmp_n = len / msl->page_sz;
773 MLX4_ASSERT(ms_idx_shift + mr->ms_bmp_n <= ms_n);
775 * Finally create a verbs MR for the memory chunk. ibv_reg_mr() can be
776 * called with holding the memory lock because it doesn't use
777 * mlx4_alloc_buf_extern() which eventually calls rte_malloc_socket()
778 * through mlx4_alloc_verbs_buf().
780 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)data.start, len,
781 IBV_ACCESS_LOCAL_WRITE);
782 if (mr->ibv_mr == NULL) {
783 WARN("port %u fail to create a verbs MR for address (%p)",
784 dev->data->port_id, (void *)addr);
788 MLX4_ASSERT((uintptr_t)mr->ibv_mr->addr == data.start);
789 MLX4_ASSERT(mr->ibv_mr->length == len);
790 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
791 DEBUG("port %u MR CREATED (%p) for %p:\n"
792 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
793 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
794 dev->data->port_id, (void *)mr, (void *)addr,
795 data.start, data.end, rte_cpu_to_be_32(mr->ibv_mr->lkey),
796 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
797 /* Insert to the global cache table. */
798 mr_insert_dev_cache(dev, mr);
799 /* Fill in output data. */
800 mr_lookup_dev(dev, entry, addr);
801 /* Lookup can't fail. */
802 MLX4_ASSERT(entry->lkey != UINT32_MAX);
803 rte_rwlock_write_unlock(&priv->mr.rwlock);
804 rte_mcfg_mem_read_unlock();
807 rte_rwlock_write_unlock(&priv->mr.rwlock);
809 rte_mcfg_mem_read_unlock();
812 * In case of error, as this can be called in a datapath, a warning
813 * message per an error is preferable instead. Must be unlocked before
814 * calling rte_free() because mlx4_mr_mem_event_free_cb() can be called
822 * Create a new global Memory Region (MR) for a missing virtual address.
823 * This can be called from primary and secondary process.
826 * Pointer to Ethernet device.
828 * Pointer to returning MR cache entry, found in the global cache or newly
829 * created. If failed to create one, this will not be updated.
831 * Target virtual address to register.
834 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
837 mlx4_mr_create(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
842 switch (rte_eal_process_type()) {
843 case RTE_PROC_PRIMARY:
844 ret = mlx4_mr_create_primary(dev, entry, addr);
846 case RTE_PROC_SECONDARY:
847 ret = mlx4_mr_create_secondary(dev, entry, addr);
856 * Rebuild the global B-tree cache of device from the original MR list.
859 * Pointer to Ethernet device.
862 mr_rebuild_dev_cache(struct rte_eth_dev *dev)
864 struct mlx4_priv *priv = dev->data->dev_private;
867 DEBUG("port %u rebuild dev cache[]", dev->data->port_id);
868 /* Flush cache to rebuild. */
869 priv->mr.cache.len = 1;
870 priv->mr.cache.overflow = 0;
871 /* Iterate all the existing MRs. */
872 LIST_FOREACH(mr, &priv->mr.mr_list, mr)
873 if (mr_insert_dev_cache(dev, mr) < 0)
878 * Callback for memory free event. Iterate freed memsegs and check whether it
879 * belongs to an existing MR. If found, clear the bit from bitmap of MR. As a
880 * result, the MR would be fragmented. If it becomes empty, the MR will be freed
881 * later by mlx4_mr_garbage_collect().
883 * The global cache must be rebuilt if there's any change and this event has to
884 * be propagated to dataplane threads to flush the local caches.
887 * Pointer to Ethernet device.
889 * Address of freed memory.
891 * Size of freed memory.
894 mlx4_mr_mem_event_free_cb(struct rte_eth_dev *dev, const void *addr, size_t len)
896 struct mlx4_priv *priv = dev->data->dev_private;
897 const struct rte_memseg_list *msl;
903 DEBUG("port %u free callback: addr=%p, len=%zu",
904 dev->data->port_id, addr, len);
905 msl = rte_mem_virt2memseg_list(addr);
906 /* addr and len must be page-aligned. */
907 MLX4_ASSERT((uintptr_t)addr ==
908 RTE_ALIGN((uintptr_t)addr, msl->page_sz));
909 MLX4_ASSERT(len == RTE_ALIGN(len, msl->page_sz));
910 ms_n = len / msl->page_sz;
911 rte_rwlock_write_lock(&priv->mr.rwlock);
912 /* Clear bits of freed memsegs from MR. */
913 for (i = 0; i < ms_n; ++i) {
914 const struct rte_memseg *ms;
915 struct mlx4_mr_cache entry;
920 /* Find MR having this memseg. */
921 start = (uintptr_t)addr + i * msl->page_sz;
922 mr = mr_lookup_dev_list(dev, &entry, start);
925 MLX4_ASSERT(mr->msl); /* Can't be external memory. */
926 ms = rte_mem_virt2memseg((void *)start, msl);
927 MLX4_ASSERT(ms != NULL);
928 MLX4_ASSERT(msl->page_sz == ms->hugepage_sz);
929 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
930 pos = ms_idx - mr->ms_base_idx;
931 MLX4_ASSERT(rte_bitmap_get(mr->ms_bmp, pos));
932 MLX4_ASSERT(pos < mr->ms_bmp_n);
933 DEBUG("port %u MR(%p): clear bitmap[%u] for addr %p",
934 dev->data->port_id, (void *)mr, pos, (void *)start);
935 rte_bitmap_clear(mr->ms_bmp, pos);
936 if (--mr->ms_n == 0) {
938 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
939 DEBUG("port %u remove MR(%p) from list",
940 dev->data->port_id, (void *)mr);
943 * MR is fragmented or will be freed. the global cache must be
949 mr_rebuild_dev_cache(dev);
951 * Flush local caches by propagating invalidation across cores.
952 * rte_smp_wmb() is enough to synchronize this event. If one of
953 * freed memsegs is seen by other core, that means the memseg
954 * has been allocated by allocator, which will come after this
955 * free call. Therefore, this store instruction (incrementing
956 * generation below) will be guaranteed to be seen by other core
957 * before the core sees the newly allocated memory.
960 DEBUG("broadcasting local cache flush, gen=%d",
964 rte_rwlock_write_unlock(&priv->mr.rwlock);
965 #ifdef RTE_LIBRTE_MLX4_DEBUG
967 mlx4_mr_dump_dev(dev);
972 * Callback for memory event.
982 mlx4_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
983 size_t len, void *arg __rte_unused)
985 struct mlx4_priv *priv;
986 struct mlx4_dev_list *dev_list = &mlx4_shared_data->mem_event_cb_list;
988 /* Must be called from the primary process. */
989 MLX4_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
990 switch (event_type) {
991 case RTE_MEM_EVENT_FREE:
992 rte_rwlock_read_lock(&mlx4_shared_data->mem_event_rwlock);
993 /* Iterate all the existing mlx4 devices. */
994 LIST_FOREACH(priv, dev_list, mem_event_cb)
995 mlx4_mr_mem_event_free_cb(ETH_DEV(priv), addr, len);
996 rte_rwlock_read_unlock(&mlx4_shared_data->mem_event_rwlock);
998 case RTE_MEM_EVENT_ALLOC:
1005 * Look up address in the global MR cache table. If not found, create a new MR.
1006 * Insert the found/created entry to local bottom-half cache table.
1009 * Pointer to Ethernet device.
1011 * Pointer to per-queue MR control structure.
1013 * Pointer to returning MR cache entry, found in the global cache or newly
1014 * created. If failed to create one, this is not written.
1019 * Searched LKey on success, UINT32_MAX on no match.
1022 mlx4_mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1023 struct mlx4_mr_cache *entry, uintptr_t addr)
1025 struct mlx4_priv *priv = dev->data->dev_private;
1026 struct mlx4_mr_btree *bt = &mr_ctrl->cache_bh;
1030 /* If local cache table is full, try to double it. */
1031 if (unlikely(bt->len == bt->size))
1032 mr_btree_expand(bt, bt->size << 1);
1033 /* Look up in the global cache. */
1034 rte_rwlock_read_lock(&priv->mr.rwlock);
1035 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
1036 if (lkey != UINT32_MAX) {
1038 *entry = (*priv->mr.cache.table)[idx];
1039 rte_rwlock_read_unlock(&priv->mr.rwlock);
1041 * Update local cache. Even if it fails, return the found entry
1042 * to update top-half cache. Next time, this entry will be found
1043 * in the global cache.
1045 mr_btree_insert(bt, entry);
1048 rte_rwlock_read_unlock(&priv->mr.rwlock);
1049 /* First time to see the address? Create a new MR. */
1050 lkey = mlx4_mr_create(dev, entry, addr);
1052 * Update the local cache if successfully created a new global MR. Even
1053 * if failed to create one, there's no action to take in this datapath
1054 * code. As returning LKey is invalid, this will eventually make HW
1057 if (lkey != UINT32_MAX)
1058 mr_btree_insert(bt, entry);
1063 * Bottom-half of LKey search on datapath. Firstly search in cache_bh[] and if
1064 * misses, search in the global MR cache table and update the new entry to
1065 * per-queue local caches.
1068 * Pointer to Ethernet device.
1070 * Pointer to per-queue MR control structure.
1075 * Searched LKey on success, UINT32_MAX on no match.
1078 mlx4_mr_addr2mr_bh(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1082 uint16_t bh_idx = 0;
1083 /* Victim in top-half cache to replace with new entry. */
1084 struct mlx4_mr_cache *repl = &mr_ctrl->cache[mr_ctrl->head];
1086 /* Binary-search MR translation table. */
1087 lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr);
1088 /* Update top-half cache. */
1089 if (likely(lkey != UINT32_MAX)) {
1090 *repl = (*mr_ctrl->cache_bh.table)[bh_idx];
1093 * If missed in local lookup table, search in the global cache
1094 * and local cache_bh[] will be updated inside if possible.
1095 * Top-half cache entry will also be updated.
1097 lkey = mlx4_mr_lookup_dev(dev, mr_ctrl, repl, addr);
1098 if (unlikely(lkey == UINT32_MAX))
1101 /* Update the most recently used entry. */
1102 mr_ctrl->mru = mr_ctrl->head;
1103 /* Point to the next victim, the oldest. */
1104 mr_ctrl->head = (mr_ctrl->head + 1) % MLX4_MR_CACHE_N;
1109 * Bottom-half of LKey search on Rx.
1112 * Pointer to Rx queue structure.
1117 * Searched LKey on success, UINT32_MAX on no match.
1120 mlx4_rx_addr2mr_bh(struct rxq *rxq, uintptr_t addr)
1122 struct mlx4_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
1123 struct mlx4_priv *priv = rxq->priv;
1125 return mlx4_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
1129 * Bottom-half of LKey search on Tx.
1132 * Pointer to Tx queue structure.
1137 * Searched LKey on success, UINT32_MAX on no match.
1140 mlx4_tx_addr2mr_bh(struct txq *txq, uintptr_t addr)
1142 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1143 struct mlx4_priv *priv = txq->priv;
1145 return mlx4_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
1149 * Bottom-half of LKey search on Tx. If it can't be searched in the memseg
1150 * list, register the mempool of the mbuf as externally allocated memory.
1153 * Pointer to Tx queue structure.
1158 * Searched LKey on success, UINT32_MAX on no match.
1161 mlx4_tx_mb2mr_bh(struct txq *txq, struct rte_mbuf *mb)
1163 uintptr_t addr = (uintptr_t)mb->buf_addr;
1166 lkey = mlx4_tx_addr2mr_bh(txq, addr);
1167 if (lkey == UINT32_MAX && rte_errno == ENXIO) {
1168 /* Mempool may have externally allocated memory. */
1169 return mlx4_tx_update_ext_mp(txq, addr, mlx4_mb2mp(mb));
1175 * Flush all of the local cache entries.
1178 * Pointer to per-queue MR control structure.
1181 mlx4_mr_flush_local_cache(struct mlx4_mr_ctrl *mr_ctrl)
1183 /* Reset the most-recently-used index. */
1185 /* Reset the linear search array. */
1187 memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache));
1188 /* Reset the B-tree table. */
1189 mr_ctrl->cache_bh.len = 1;
1190 mr_ctrl->cache_bh.overflow = 0;
1191 /* Update the generation number. */
1192 mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr;
1193 DEBUG("mr_ctrl(%p): flushed, cur_gen=%d",
1194 (void *)mr_ctrl, mr_ctrl->cur_gen);
1198 * Called during rte_mempool_mem_iter() by mlx4_mr_update_ext_mp().
1200 * Externally allocated chunk is registered and a MR is created for the chunk.
1201 * The MR object is added to the global list. If memseg list of a MR object
1202 * (mr->msl) is null, the MR object can be regarded as externally allocated
1205 * Once external memory is registered, it should be static. If the memory is
1206 * freed and the virtual address range has different physical memory mapped
1207 * again, it may cause crash on device due to the wrong translation entry. PMD
1208 * can't track the free event of the external memory for now.
1211 mlx4_mr_update_ext_mp_cb(struct rte_mempool *mp, void *opaque,
1212 struct rte_mempool_memhdr *memhdr,
1213 unsigned mem_idx __rte_unused)
1215 struct mr_update_mp_data *data = opaque;
1216 struct rte_eth_dev *dev = data->dev;
1217 struct mlx4_priv *priv = dev->data->dev_private;
1218 struct mlx4_mr_ctrl *mr_ctrl = data->mr_ctrl;
1219 struct mlx4_mr *mr = NULL;
1220 uintptr_t addr = (uintptr_t)memhdr->addr;
1221 size_t len = memhdr->len;
1222 struct mlx4_mr_cache entry;
1225 MLX4_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1226 /* If already registered, it should return. */
1227 rte_rwlock_read_lock(&priv->mr.rwlock);
1228 lkey = mr_lookup_dev(dev, &entry, addr);
1229 rte_rwlock_read_unlock(&priv->mr.rwlock);
1230 if (lkey != UINT32_MAX)
1232 mr = rte_zmalloc_socket(NULL,
1233 RTE_ALIGN_CEIL(sizeof(*mr),
1234 RTE_CACHE_LINE_SIZE),
1235 RTE_CACHE_LINE_SIZE, mp->socket_id);
1237 WARN("port %u unable to allocate memory for a new MR of"
1239 dev->data->port_id, mp->name);
1243 DEBUG("port %u register MR for chunk #%d of mempool (%s)",
1244 dev->data->port_id, mem_idx, mp->name);
1245 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)addr, len,
1246 IBV_ACCESS_LOCAL_WRITE);
1247 if (mr->ibv_mr == NULL) {
1248 WARN("port %u fail to create a verbs MR for address (%p)",
1249 dev->data->port_id, (void *)addr);
1254 mr->msl = NULL; /* Mark it is external memory. */
1258 rte_rwlock_write_lock(&priv->mr.rwlock);
1259 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
1260 DEBUG("port %u MR CREATED (%p) for external memory %p:\n"
1261 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
1262 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
1263 dev->data->port_id, (void *)mr, (void *)addr,
1264 addr, addr + len, rte_cpu_to_be_32(mr->ibv_mr->lkey),
1265 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
1266 /* Insert to the global cache table. */
1267 mr_insert_dev_cache(dev, mr);
1268 rte_rwlock_write_unlock(&priv->mr.rwlock);
1269 /* Insert to the local cache table */
1270 mlx4_mr_addr2mr_bh(dev, mr_ctrl, addr);
1274 * Register MR for entire memory chunks in a Mempool having externally allocated
1275 * memory and fill in local cache.
1278 * Pointer to Ethernet device.
1280 * Pointer to per-queue MR control structure.
1282 * Pointer to registering Mempool.
1285 * 0 on success, -1 on failure.
1288 mlx4_mr_update_ext_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1289 struct rte_mempool *mp)
1291 struct mr_update_mp_data data = {
1297 rte_mempool_mem_iter(mp, mlx4_mr_update_ext_mp_cb, &data);
1302 * Register MR entire memory chunks in a Mempool having externally allocated
1303 * memory and search LKey of the address to return.
1306 * Pointer to Ethernet device.
1310 * Pointer to registering Mempool where addr belongs.
1313 * LKey for address on success, UINT32_MAX on failure.
1316 mlx4_tx_update_ext_mp(struct txq *txq, uintptr_t addr, struct rte_mempool *mp)
1318 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1319 struct mlx4_priv *priv = txq->priv;
1321 if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
1322 WARN("port %u using address (%p) from unregistered mempool"
1323 " having externally allocated memory"
1324 " in secondary process, please create mempool"
1325 " prior to rte_eth_dev_start()",
1326 PORT_ID(priv), (void *)addr);
1329 mlx4_mr_update_ext_mp(ETH_DEV(priv), mr_ctrl, mp);
1330 return mlx4_tx_addr2mr_bh(txq, addr);
1333 /* Called during rte_mempool_mem_iter() by mlx4_mr_update_mp(). */
1335 mlx4_mr_update_mp_cb(struct rte_mempool *mp __rte_unused, void *opaque,
1336 struct rte_mempool_memhdr *memhdr,
1337 unsigned mem_idx __rte_unused)
1339 struct mr_update_mp_data *data = opaque;
1342 /* Stop iteration if failed in the previous walk. */
1345 /* Register address of the chunk and update local caches. */
1346 lkey = mlx4_mr_addr2mr_bh(data->dev, data->mr_ctrl,
1347 (uintptr_t)memhdr->addr);
1348 if (lkey == UINT32_MAX)
1353 * Register entire memory chunks in a Mempool.
1356 * Pointer to Ethernet device.
1358 * Pointer to per-queue MR control structure.
1360 * Pointer to registering Mempool.
1363 * 0 on success, -1 on failure.
1366 mlx4_mr_update_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1367 struct rte_mempool *mp)
1369 struct mr_update_mp_data data = {
1375 rte_mempool_mem_iter(mp, mlx4_mr_update_mp_cb, &data);
1376 if (data.ret < 0 && rte_errno == ENXIO) {
1377 /* Mempool may have externally allocated memory. */
1378 return mlx4_mr_update_ext_mp(dev, mr_ctrl, mp);
1383 #ifdef RTE_LIBRTE_MLX4_DEBUG
1385 * Dump all the created MRs and the global cache entries.
1388 * Pointer to Ethernet device.
1391 mlx4_mr_dump_dev(struct rte_eth_dev *dev)
1393 struct mlx4_priv *priv = dev->data->dev_private;
1398 rte_rwlock_read_lock(&priv->mr.rwlock);
1399 /* Iterate all the existing MRs. */
1400 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
1403 DEBUG("port %u MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u",
1404 dev->data->port_id, mr_n++,
1405 rte_cpu_to_be_32(mr->ibv_mr->lkey),
1406 mr->ms_n, mr->ms_bmp_n);
1409 for (n = 0; n < mr->ms_bmp_n; ) {
1410 struct mlx4_mr_cache ret;
1412 memset(&ret, 0, sizeof(ret));
1413 n = mr_find_next_chunk(mr, &ret, n);
1416 DEBUG(" chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")",
1417 chunk_n++, ret.start, ret.end);
1420 DEBUG("port %u dumping global cache", dev->data->port_id);
1421 mlx4_mr_btree_dump(&priv->mr.cache);
1422 rte_rwlock_read_unlock(&priv->mr.rwlock);
1427 * Release all the created MRs and resources. Remove device from memory callback
1431 * Pointer to Ethernet device.
1434 mlx4_mr_release(struct rte_eth_dev *dev)
1436 struct mlx4_priv *priv = dev->data->dev_private;
1437 struct mlx4_mr *mr_next;
1439 /* Remove from memory callback device list. */
1440 rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock);
1441 LIST_REMOVE(priv, mem_event_cb);
1442 rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock);
1443 #ifdef RTE_LIBRTE_MLX4_DEBUG
1444 mlx4_mr_dump_dev(dev);
1446 rte_rwlock_write_lock(&priv->mr.rwlock);
1447 /* Detach from MR list and move to free list. */
1448 mr_next = LIST_FIRST(&priv->mr.mr_list);
1449 while (mr_next != NULL) {
1450 struct mlx4_mr *mr = mr_next;
1452 mr_next = LIST_NEXT(mr, mr);
1453 LIST_REMOVE(mr, mr);
1454 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
1456 LIST_INIT(&priv->mr.mr_list);
1457 /* Free global cache. */
1458 mlx4_mr_btree_free(&priv->mr.cache);
1459 rte_rwlock_write_unlock(&priv->mr.rwlock);
1460 /* Free all remaining MRs. */
1461 mlx4_mr_garbage_collect(dev);