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 rte_bitmap_free(mr->ms_bmp);
479 * Release resources of detached MR having no online entry.
482 * Pointer to Ethernet device.
485 mlx4_mr_garbage_collect(struct rte_eth_dev *dev)
487 struct mlx4_priv *priv = dev->data->dev_private;
488 struct mlx4_mr *mr_next;
489 struct mlx4_mr_list free_list = LIST_HEAD_INITIALIZER(free_list);
491 /* Must be called from the primary process. */
492 MLX4_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
494 * MR can't be freed with holding the lock because rte_free() could call
495 * memory free callback function. This will be a deadlock situation.
497 rte_rwlock_write_lock(&priv->mr.rwlock);
498 /* Detach the whole free list and release it after unlocking. */
499 free_list = priv->mr.mr_free_list;
500 LIST_INIT(&priv->mr.mr_free_list);
501 rte_rwlock_write_unlock(&priv->mr.rwlock);
502 /* Release resources. */
503 mr_next = LIST_FIRST(&free_list);
504 while (mr_next != NULL) {
505 struct mlx4_mr *mr = mr_next;
507 mr_next = LIST_NEXT(mr, mr);
512 /* Called during rte_memseg_contig_walk() by mlx4_mr_create(). */
514 mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl,
515 const struct rte_memseg *ms, size_t len, void *arg)
517 struct mr_find_contig_memsegs_data *data = arg;
519 if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len)
521 /* Found, save it and stop walking. */
522 data->start = ms->addr_64;
523 data->end = ms->addr_64 + len;
529 * Create a new global Memory Region (MR) for a missing virtual address.
530 * This API should be called on a secondary process, then a request is sent to
531 * the primary process in order to create a MR for the address. As the global MR
532 * list is on the shared memory, following LKey lookup should succeed unless the
536 * Pointer to Ethernet device.
538 * Pointer to returning MR cache entry, found in the global cache or newly
539 * created. If failed to create one, this will not be updated.
541 * Target virtual address to register.
544 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
547 mlx4_mr_create_secondary(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
550 struct mlx4_priv *priv = dev->data->dev_private;
553 DEBUG("port %u requesting MR creation for address (%p)",
554 dev->data->port_id, (void *)addr);
555 ret = mlx4_mp_req_mr_create(dev, addr);
557 DEBUG("port %u fail to request MR creation for address (%p)",
558 dev->data->port_id, (void *)addr);
561 rte_rwlock_read_lock(&priv->mr.rwlock);
562 /* Fill in output data. */
563 mr_lookup_dev(dev, entry, addr);
564 /* Lookup can't fail. */
565 MLX4_ASSERT(entry->lkey != UINT32_MAX);
566 rte_rwlock_read_unlock(&priv->mr.rwlock);
567 DEBUG("port %u MR CREATED by primary process for %p:\n"
568 " [0x%" PRIxPTR ", 0x%" PRIxPTR "), lkey=0x%x",
569 dev->data->port_id, (void *)addr,
570 entry->start, entry->end, entry->lkey);
575 * Create a new global Memory Region (MR) for a missing virtual address.
576 * Register entire virtually contiguous memory chunk around the address.
577 * This must be called from the primary process.
580 * Pointer to Ethernet device.
582 * Pointer to returning MR cache entry, found in the global cache or newly
583 * created. If failed to create one, this will not be updated.
585 * Target virtual address to register.
588 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
591 mlx4_mr_create_primary(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
594 struct mlx4_priv *priv = dev->data->dev_private;
595 const struct rte_memseg_list *msl;
596 const struct rte_memseg *ms;
597 struct mlx4_mr *mr = NULL;
602 int ms_idx_shift = -1;
604 struct mr_find_contig_memsegs_data data = {
607 struct mr_find_contig_memsegs_data data_re;
609 DEBUG("port %u creating a MR using address (%p)",
610 dev->data->port_id, (void *)addr);
612 * Release detached MRs if any. This can't be called with holding either
613 * memory_hotplug_lock or priv->mr.rwlock. MRs on the free list have
614 * been detached by the memory free event but it couldn't be released
615 * inside the callback due to deadlock. As a result, releasing resources
616 * is quite opportunistic.
618 mlx4_mr_garbage_collect(dev);
620 * If enabled, find out a contiguous virtual address chunk in use, to
621 * which the given address belongs, in order to register maximum range.
622 * In the best case where mempools are not dynamically recreated and
623 * '--socket-mem' is specified as an EAL option, it is very likely to
624 * have only one MR(LKey) per a socket and per a hugepage-size even
625 * though the system memory is highly fragmented. As the whole memory
626 * chunk will be pinned by kernel, it can't be reused unless entire
627 * chunk is freed from EAL.
629 * If disabled, just register one memseg (page). Then, memory
630 * consumption will be minimized but it may drop performance if there
631 * are many MRs to lookup on the datapath.
633 if (!priv->mr_ext_memseg_en) {
634 data.msl = rte_mem_virt2memseg_list((void *)addr);
635 data.start = RTE_ALIGN_FLOOR(addr, data.msl->page_sz);
636 data.end = data.start + data.msl->page_sz;
637 } else if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) {
638 WARN("port %u unable to find virtually contiguous"
639 " chunk for address (%p)."
640 " rte_memseg_contig_walk() failed.",
641 dev->data->port_id, (void *)addr);
646 /* Addresses must be page-aligned. */
647 MLX4_ASSERT(rte_is_aligned((void *)data.start, data.msl->page_sz));
648 MLX4_ASSERT(rte_is_aligned((void *)data.end, data.msl->page_sz));
650 ms = rte_mem_virt2memseg((void *)data.start, msl);
651 len = data.end - data.start;
652 MLX4_ASSERT(msl->page_sz == ms->hugepage_sz);
653 /* Number of memsegs in the range. */
654 ms_n = len / msl->page_sz;
655 DEBUG("port %u extending %p to [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
656 " page_sz=0x%" PRIx64 ", ms_n=%u",
657 dev->data->port_id, (void *)addr,
658 data.start, data.end, msl->page_sz, ms_n);
659 /* Size of memory for bitmap. */
660 bmp_size = rte_bitmap_get_memory_footprint(ms_n);
661 mr = rte_zmalloc_socket(NULL,
662 RTE_ALIGN_CEIL(sizeof(*mr),
663 RTE_CACHE_LINE_SIZE) +
665 RTE_CACHE_LINE_SIZE, msl->socket_id);
667 WARN("port %u unable to allocate memory for a new MR of"
669 dev->data->port_id, (void *)addr);
675 * Save the index of the first memseg and initialize memseg bitmap. To
676 * see if a memseg of ms_idx in the memseg-list is still valid, check:
677 * rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx)
679 mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
680 bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE);
681 mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size);
682 if (mr->ms_bmp == NULL) {
683 WARN("port %u unable to initialize bitmap for a new MR of"
685 dev->data->port_id, (void *)addr);
690 * Should recheck whether the extended contiguous chunk is still valid.
691 * Because memory_hotplug_lock can't be held if there's any memory
692 * related calls in a critical path, resource allocation above can't be
693 * locked. If the memory has been changed at this point, try again with
694 * just single page. If not, go on with the big chunk atomically from
697 rte_mcfg_mem_read_lock();
699 if (len > msl->page_sz &&
700 !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) {
701 WARN("port %u unable to find virtually contiguous"
702 " chunk for address (%p)."
703 " rte_memseg_contig_walk() failed.",
704 dev->data->port_id, (void *)addr);
708 if (data.start != data_re.start || data.end != data_re.end) {
710 * The extended contiguous chunk has been changed. Try again
711 * with single memseg instead.
713 data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz);
714 data.end = data.start + msl->page_sz;
715 rte_mcfg_mem_read_unlock();
717 goto alloc_resources;
719 MLX4_ASSERT(data.msl == data_re.msl);
720 rte_rwlock_write_lock(&priv->mr.rwlock);
722 * Check the address is really missing. If other thread already created
723 * one or it is not found due to overflow, abort and return.
725 if (mr_lookup_dev(dev, entry, addr) != UINT32_MAX) {
727 * Insert to the global cache table. It may fail due to
728 * low-on-memory. Then, this entry will have to be searched
731 mr_btree_insert(&priv->mr.cache, entry);
732 DEBUG("port %u found MR for %p on final lookup, abort",
733 dev->data->port_id, (void *)addr);
734 rte_rwlock_write_unlock(&priv->mr.rwlock);
735 rte_mcfg_mem_read_unlock();
737 * Must be unlocked before calling rte_free() because
738 * mlx4_mr_mem_event_free_cb() can be called inside.
744 * Trim start and end addresses for verbs MR. Set bits for registering
745 * memsegs but exclude already registered ones. Bitmap can be
748 for (n = 0; n < ms_n; ++n) {
750 struct mlx4_mr_cache ret;
752 memset(&ret, 0, sizeof(ret));
753 start = data_re.start + n * msl->page_sz;
754 /* Exclude memsegs already registered by other MRs. */
755 if (mr_lookup_dev(dev, &ret, start) == UINT32_MAX) {
757 * Start from the first unregistered memseg in the
760 if (ms_idx_shift == -1) {
761 mr->ms_base_idx += n;
765 data.end = start + msl->page_sz;
766 rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift);
770 len = data.end - data.start;
771 mr->ms_bmp_n = len / msl->page_sz;
772 MLX4_ASSERT(ms_idx_shift + mr->ms_bmp_n <= ms_n);
774 * Finally create a verbs MR for the memory chunk. ibv_reg_mr() can be
775 * called with holding the memory lock because it doesn't use
776 * mlx4_alloc_buf_extern() which eventually calls rte_malloc_socket()
777 * through mlx4_alloc_verbs_buf().
779 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)data.start, len,
780 IBV_ACCESS_LOCAL_WRITE);
781 if (mr->ibv_mr == NULL) {
782 WARN("port %u fail to create a verbs MR for address (%p)",
783 dev->data->port_id, (void *)addr);
787 MLX4_ASSERT((uintptr_t)mr->ibv_mr->addr == data.start);
788 MLX4_ASSERT(mr->ibv_mr->length == len);
789 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
790 DEBUG("port %u MR CREATED (%p) for %p:\n"
791 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
792 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
793 dev->data->port_id, (void *)mr, (void *)addr,
794 data.start, data.end, rte_cpu_to_be_32(mr->ibv_mr->lkey),
795 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
796 /* Insert to the global cache table. */
797 mr_insert_dev_cache(dev, mr);
798 /* Fill in output data. */
799 mr_lookup_dev(dev, entry, addr);
800 /* Lookup can't fail. */
801 MLX4_ASSERT(entry->lkey != UINT32_MAX);
802 rte_rwlock_write_unlock(&priv->mr.rwlock);
803 rte_mcfg_mem_read_unlock();
806 rte_rwlock_write_unlock(&priv->mr.rwlock);
808 rte_mcfg_mem_read_unlock();
811 * In case of error, as this can be called in a datapath, a warning
812 * message per an error is preferable instead. Must be unlocked before
813 * calling rte_free() because mlx4_mr_mem_event_free_cb() can be called
821 * Create a new global Memory Region (MR) for a missing virtual address.
822 * This can be called from primary and secondary process.
825 * Pointer to Ethernet device.
827 * Pointer to returning MR cache entry, found in the global cache or newly
828 * created. If failed to create one, this will not be updated.
830 * Target virtual address to register.
833 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
836 mlx4_mr_create(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
841 switch (rte_eal_process_type()) {
842 case RTE_PROC_PRIMARY:
843 ret = mlx4_mr_create_primary(dev, entry, addr);
845 case RTE_PROC_SECONDARY:
846 ret = mlx4_mr_create_secondary(dev, entry, addr);
855 * Rebuild the global B-tree cache of device from the original MR list.
858 * Pointer to Ethernet device.
861 mr_rebuild_dev_cache(struct rte_eth_dev *dev)
863 struct mlx4_priv *priv = dev->data->dev_private;
866 DEBUG("port %u rebuild dev cache[]", dev->data->port_id);
867 /* Flush cache to rebuild. */
868 priv->mr.cache.len = 1;
869 priv->mr.cache.overflow = 0;
870 /* Iterate all the existing MRs. */
871 LIST_FOREACH(mr, &priv->mr.mr_list, mr)
872 if (mr_insert_dev_cache(dev, mr) < 0)
877 * Callback for memory free event. Iterate freed memsegs and check whether it
878 * belongs to an existing MR. If found, clear the bit from bitmap of MR. As a
879 * result, the MR would be fragmented. If it becomes empty, the MR will be freed
880 * later by mlx4_mr_garbage_collect().
882 * The global cache must be rebuilt if there's any change and this event has to
883 * be propagated to dataplane threads to flush the local caches.
886 * Pointer to Ethernet device.
888 * Address of freed memory.
890 * Size of freed memory.
893 mlx4_mr_mem_event_free_cb(struct rte_eth_dev *dev, const void *addr, size_t len)
895 struct mlx4_priv *priv = dev->data->dev_private;
896 const struct rte_memseg_list *msl;
902 DEBUG("port %u free callback: addr=%p, len=%zu",
903 dev->data->port_id, addr, len);
904 msl = rte_mem_virt2memseg_list(addr);
905 /* addr and len must be page-aligned. */
906 MLX4_ASSERT((uintptr_t)addr ==
907 RTE_ALIGN((uintptr_t)addr, msl->page_sz));
908 MLX4_ASSERT(len == RTE_ALIGN(len, msl->page_sz));
909 ms_n = len / msl->page_sz;
910 rte_rwlock_write_lock(&priv->mr.rwlock);
911 /* Clear bits of freed memsegs from MR. */
912 for (i = 0; i < ms_n; ++i) {
913 const struct rte_memseg *ms;
914 struct mlx4_mr_cache entry;
919 /* Find MR having this memseg. */
920 start = (uintptr_t)addr + i * msl->page_sz;
921 mr = mr_lookup_dev_list(dev, &entry, start);
924 MLX4_ASSERT(mr->msl); /* Can't be external memory. */
925 ms = rte_mem_virt2memseg((void *)start, msl);
926 MLX4_ASSERT(ms != NULL);
927 MLX4_ASSERT(msl->page_sz == ms->hugepage_sz);
928 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
929 pos = ms_idx - mr->ms_base_idx;
930 MLX4_ASSERT(rte_bitmap_get(mr->ms_bmp, pos));
931 MLX4_ASSERT(pos < mr->ms_bmp_n);
932 DEBUG("port %u MR(%p): clear bitmap[%u] for addr %p",
933 dev->data->port_id, (void *)mr, pos, (void *)start);
934 rte_bitmap_clear(mr->ms_bmp, pos);
935 if (--mr->ms_n == 0) {
937 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
938 DEBUG("port %u remove MR(%p) from list",
939 dev->data->port_id, (void *)mr);
942 * MR is fragmented or will be freed. the global cache must be
948 mr_rebuild_dev_cache(dev);
950 * No explicit wmb is needed after updating dev_gen due to
951 * store-release ordering in unlock that provides the
952 * implicit barrier at the software visible level.
955 DEBUG("broadcasting local cache flush, gen=%d",
958 rte_rwlock_write_unlock(&priv->mr.rwlock);
959 #ifdef RTE_LIBRTE_MLX4_DEBUG
961 mlx4_mr_dump_dev(dev);
966 * Callback for memory event.
976 mlx4_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
977 size_t len, void *arg __rte_unused)
979 struct mlx4_priv *priv;
980 struct mlx4_dev_list *dev_list = &mlx4_shared_data->mem_event_cb_list;
982 /* Must be called from the primary process. */
983 MLX4_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
984 switch (event_type) {
985 case RTE_MEM_EVENT_FREE:
986 rte_rwlock_read_lock(&mlx4_shared_data->mem_event_rwlock);
987 /* Iterate all the existing mlx4 devices. */
988 LIST_FOREACH(priv, dev_list, mem_event_cb)
989 mlx4_mr_mem_event_free_cb(ETH_DEV(priv), addr, len);
990 rte_rwlock_read_unlock(&mlx4_shared_data->mem_event_rwlock);
992 case RTE_MEM_EVENT_ALLOC:
999 * Look up address in the global MR cache table. If not found, create a new MR.
1000 * Insert the found/created entry to local bottom-half cache table.
1003 * Pointer to Ethernet device.
1005 * Pointer to per-queue MR control structure.
1007 * Pointer to returning MR cache entry, found in the global cache or newly
1008 * created. If failed to create one, this is not written.
1013 * Searched LKey on success, UINT32_MAX on no match.
1016 mlx4_mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1017 struct mlx4_mr_cache *entry, uintptr_t addr)
1019 struct mlx4_priv *priv = dev->data->dev_private;
1020 struct mlx4_mr_btree *bt = &mr_ctrl->cache_bh;
1024 /* If local cache table is full, try to double it. */
1025 if (unlikely(bt->len == bt->size))
1026 mr_btree_expand(bt, bt->size << 1);
1027 /* Look up in the global cache. */
1028 rte_rwlock_read_lock(&priv->mr.rwlock);
1029 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
1030 if (lkey != UINT32_MAX) {
1032 *entry = (*priv->mr.cache.table)[idx];
1033 rte_rwlock_read_unlock(&priv->mr.rwlock);
1035 * Update local cache. Even if it fails, return the found entry
1036 * to update top-half cache. Next time, this entry will be found
1037 * in the global cache.
1039 mr_btree_insert(bt, entry);
1042 rte_rwlock_read_unlock(&priv->mr.rwlock);
1043 /* First time to see the address? Create a new MR. */
1044 lkey = mlx4_mr_create(dev, entry, addr);
1046 * Update the local cache if successfully created a new global MR. Even
1047 * if failed to create one, there's no action to take in this datapath
1048 * code. As returning LKey is invalid, this will eventually make HW
1051 if (lkey != UINT32_MAX)
1052 mr_btree_insert(bt, entry);
1057 * Bottom-half of LKey search on datapath. Firstly search in cache_bh[] and if
1058 * misses, search in the global MR cache table and update the new entry to
1059 * per-queue local caches.
1062 * Pointer to Ethernet device.
1064 * Pointer to per-queue MR control structure.
1069 * Searched LKey on success, UINT32_MAX on no match.
1072 mlx4_mr_addr2mr_bh(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1076 uint16_t bh_idx = 0;
1077 /* Victim in top-half cache to replace with new entry. */
1078 struct mlx4_mr_cache *repl = &mr_ctrl->cache[mr_ctrl->head];
1080 /* Binary-search MR translation table. */
1081 lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr);
1082 /* Update top-half cache. */
1083 if (likely(lkey != UINT32_MAX)) {
1084 *repl = (*mr_ctrl->cache_bh.table)[bh_idx];
1087 * If missed in local lookup table, search in the global cache
1088 * and local cache_bh[] will be updated inside if possible.
1089 * Top-half cache entry will also be updated.
1091 lkey = mlx4_mr_lookup_dev(dev, mr_ctrl, repl, addr);
1092 if (unlikely(lkey == UINT32_MAX))
1095 /* Update the most recently used entry. */
1096 mr_ctrl->mru = mr_ctrl->head;
1097 /* Point to the next victim, the oldest. */
1098 mr_ctrl->head = (mr_ctrl->head + 1) % MLX4_MR_CACHE_N;
1103 * Bottom-half of LKey search on Rx.
1106 * Pointer to Rx queue structure.
1111 * Searched LKey on success, UINT32_MAX on no match.
1114 mlx4_rx_addr2mr_bh(struct rxq *rxq, uintptr_t addr)
1116 struct mlx4_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
1117 struct mlx4_priv *priv = rxq->priv;
1119 return mlx4_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
1123 * Bottom-half of LKey search on Tx.
1126 * Pointer to Tx queue structure.
1131 * Searched LKey on success, UINT32_MAX on no match.
1134 mlx4_tx_addr2mr_bh(struct txq *txq, uintptr_t addr)
1136 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1137 struct mlx4_priv *priv = txq->priv;
1139 return mlx4_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
1143 * Bottom-half of LKey search on Tx. If it can't be searched in the memseg
1144 * list, register the mempool of the mbuf as externally allocated memory.
1147 * Pointer to Tx queue structure.
1152 * Searched LKey on success, UINT32_MAX on no match.
1155 mlx4_tx_mb2mr_bh(struct txq *txq, struct rte_mbuf *mb)
1157 uintptr_t addr = (uintptr_t)mb->buf_addr;
1160 lkey = mlx4_tx_addr2mr_bh(txq, addr);
1161 if (lkey == UINT32_MAX && rte_errno == ENXIO) {
1162 /* Mempool may have externally allocated memory. */
1163 return mlx4_tx_update_ext_mp(txq, addr, mlx4_mb2mp(mb));
1169 * Flush all of the local cache entries.
1172 * Pointer to per-queue MR control structure.
1175 mlx4_mr_flush_local_cache(struct mlx4_mr_ctrl *mr_ctrl)
1177 /* Reset the most-recently-used index. */
1179 /* Reset the linear search array. */
1181 memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache));
1182 /* Reset the B-tree table. */
1183 mr_ctrl->cache_bh.len = 1;
1184 mr_ctrl->cache_bh.overflow = 0;
1185 /* Update the generation number. */
1186 mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr;
1187 DEBUG("mr_ctrl(%p): flushed, cur_gen=%d",
1188 (void *)mr_ctrl, mr_ctrl->cur_gen);
1192 * Called during rte_mempool_mem_iter() by mlx4_mr_update_ext_mp().
1194 * Externally allocated chunk is registered and a MR is created for the chunk.
1195 * The MR object is added to the global list. If memseg list of a MR object
1196 * (mr->msl) is null, the MR object can be regarded as externally allocated
1199 * Once external memory is registered, it should be static. If the memory is
1200 * freed and the virtual address range has different physical memory mapped
1201 * again, it may cause crash on device due to the wrong translation entry. PMD
1202 * can't track the free event of the external memory for now.
1205 mlx4_mr_update_ext_mp_cb(struct rte_mempool *mp, void *opaque,
1206 struct rte_mempool_memhdr *memhdr,
1207 unsigned mem_idx __rte_unused)
1209 struct mr_update_mp_data *data = opaque;
1210 struct rte_eth_dev *dev = data->dev;
1211 struct mlx4_priv *priv = dev->data->dev_private;
1212 struct mlx4_mr_ctrl *mr_ctrl = data->mr_ctrl;
1213 struct mlx4_mr *mr = NULL;
1214 uintptr_t addr = (uintptr_t)memhdr->addr;
1215 size_t len = memhdr->len;
1216 struct mlx4_mr_cache entry;
1219 MLX4_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1220 /* If already registered, it should return. */
1221 rte_rwlock_read_lock(&priv->mr.rwlock);
1222 lkey = mr_lookup_dev(dev, &entry, addr);
1223 rte_rwlock_read_unlock(&priv->mr.rwlock);
1224 if (lkey != UINT32_MAX)
1226 mr = rte_zmalloc_socket(NULL,
1227 RTE_ALIGN_CEIL(sizeof(*mr),
1228 RTE_CACHE_LINE_SIZE),
1229 RTE_CACHE_LINE_SIZE, mp->socket_id);
1231 WARN("port %u unable to allocate memory for a new MR of"
1233 dev->data->port_id, mp->name);
1237 DEBUG("port %u register MR for chunk #%d of mempool (%s)",
1238 dev->data->port_id, mem_idx, mp->name);
1239 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)addr, len,
1240 IBV_ACCESS_LOCAL_WRITE);
1241 if (mr->ibv_mr == NULL) {
1242 WARN("port %u fail to create a verbs MR for address (%p)",
1243 dev->data->port_id, (void *)addr);
1248 mr->msl = NULL; /* Mark it is external memory. */
1252 rte_rwlock_write_lock(&priv->mr.rwlock);
1253 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
1254 DEBUG("port %u MR CREATED (%p) for external memory %p:\n"
1255 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
1256 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
1257 dev->data->port_id, (void *)mr, (void *)addr,
1258 addr, addr + len, rte_cpu_to_be_32(mr->ibv_mr->lkey),
1259 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
1260 /* Insert to the global cache table. */
1261 mr_insert_dev_cache(dev, mr);
1262 rte_rwlock_write_unlock(&priv->mr.rwlock);
1263 /* Insert to the local cache table */
1264 mlx4_mr_addr2mr_bh(dev, mr_ctrl, addr);
1268 * Register MR for entire memory chunks in a Mempool having externally allocated
1269 * memory and fill in local cache.
1272 * Pointer to Ethernet device.
1274 * Pointer to per-queue MR control structure.
1276 * Pointer to registering Mempool.
1279 * 0 on success, -1 on failure.
1282 mlx4_mr_update_ext_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1283 struct rte_mempool *mp)
1285 struct mr_update_mp_data data = {
1291 rte_mempool_mem_iter(mp, mlx4_mr_update_ext_mp_cb, &data);
1296 * Register MR entire memory chunks in a Mempool having externally allocated
1297 * memory and search LKey of the address to return.
1300 * Pointer to Ethernet device.
1304 * Pointer to registering Mempool where addr belongs.
1307 * LKey for address on success, UINT32_MAX on failure.
1310 mlx4_tx_update_ext_mp(struct txq *txq, uintptr_t addr, struct rte_mempool *mp)
1312 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1313 struct mlx4_priv *priv = txq->priv;
1315 if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
1316 WARN("port %u using address (%p) from unregistered mempool"
1317 " having externally allocated memory"
1318 " in secondary process, please create mempool"
1319 " prior to rte_eth_dev_start()",
1320 PORT_ID(priv), (void *)addr);
1323 mlx4_mr_update_ext_mp(ETH_DEV(priv), mr_ctrl, mp);
1324 return mlx4_tx_addr2mr_bh(txq, addr);
1327 /* Called during rte_mempool_mem_iter() by mlx4_mr_update_mp(). */
1329 mlx4_mr_update_mp_cb(struct rte_mempool *mp __rte_unused, void *opaque,
1330 struct rte_mempool_memhdr *memhdr,
1331 unsigned mem_idx __rte_unused)
1333 struct mr_update_mp_data *data = opaque;
1336 /* Stop iteration if failed in the previous walk. */
1339 /* Register address of the chunk and update local caches. */
1340 lkey = mlx4_mr_addr2mr_bh(data->dev, data->mr_ctrl,
1341 (uintptr_t)memhdr->addr);
1342 if (lkey == UINT32_MAX)
1347 * Register entire memory chunks in a Mempool.
1350 * Pointer to Ethernet device.
1352 * Pointer to per-queue MR control structure.
1354 * Pointer to registering Mempool.
1357 * 0 on success, -1 on failure.
1360 mlx4_mr_update_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1361 struct rte_mempool *mp)
1363 struct mr_update_mp_data data = {
1369 rte_mempool_mem_iter(mp, mlx4_mr_update_mp_cb, &data);
1370 if (data.ret < 0 && rte_errno == ENXIO) {
1371 /* Mempool may have externally allocated memory. */
1372 return mlx4_mr_update_ext_mp(dev, mr_ctrl, mp);
1377 #ifdef RTE_LIBRTE_MLX4_DEBUG
1379 * Dump all the created MRs and the global cache entries.
1382 * Pointer to Ethernet device.
1385 mlx4_mr_dump_dev(struct rte_eth_dev *dev)
1387 struct mlx4_priv *priv = dev->data->dev_private;
1392 rte_rwlock_read_lock(&priv->mr.rwlock);
1393 /* Iterate all the existing MRs. */
1394 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
1397 DEBUG("port %u MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u",
1398 dev->data->port_id, mr_n++,
1399 rte_cpu_to_be_32(mr->ibv_mr->lkey),
1400 mr->ms_n, mr->ms_bmp_n);
1403 for (n = 0; n < mr->ms_bmp_n; ) {
1404 struct mlx4_mr_cache ret;
1406 memset(&ret, 0, sizeof(ret));
1407 n = mr_find_next_chunk(mr, &ret, n);
1410 DEBUG(" chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")",
1411 chunk_n++, ret.start, ret.end);
1414 DEBUG("port %u dumping global cache", dev->data->port_id);
1415 mlx4_mr_btree_dump(&priv->mr.cache);
1416 rte_rwlock_read_unlock(&priv->mr.rwlock);
1421 * Release all the created MRs and resources. Remove device from memory callback
1425 * Pointer to Ethernet device.
1428 mlx4_mr_release(struct rte_eth_dev *dev)
1430 struct mlx4_priv *priv = dev->data->dev_private;
1431 struct mlx4_mr *mr_next;
1433 /* Remove from memory callback device list. */
1434 rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock);
1435 LIST_REMOVE(priv, mem_event_cb);
1436 rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock);
1437 #ifdef RTE_LIBRTE_MLX4_DEBUG
1438 mlx4_mr_dump_dev(dev);
1440 rte_rwlock_write_lock(&priv->mr.rwlock);
1441 /* Detach from MR list and move to free list. */
1442 mr_next = LIST_FIRST(&priv->mr.mr_list);
1443 while (mr_next != NULL) {
1444 struct mlx4_mr *mr = mr_next;
1446 mr_next = LIST_NEXT(mr, mr);
1447 LIST_REMOVE(mr, mr);
1448 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
1450 LIST_INIT(&priv->mr.mr_list);
1451 /* Free global cache. */
1452 mlx4_mr_btree_free(&priv->mr.cache);
1453 rte_rwlock_write_unlock(&priv->mr.rwlock);
1454 /* Free all remaining MRs. */
1455 mlx4_mr_garbage_collect(dev);