1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2016 6WIND S.A.
3 * Copyright 2020 Mellanox Technologies, Ltd
7 #include <rte_eal_memconfig.h>
8 #include <rte_eal_paging.h>
10 #include <rte_mempool.h>
11 #include <rte_malloc.h>
12 #include <rte_rwlock.h>
14 #include "mlx5_glue.h"
15 #include "mlx5_common.h"
16 #include "mlx5_common_mp.h"
17 #include "mlx5_common_mr.h"
18 #include "mlx5_common_os.h"
19 #include "mlx5_common_log.h"
20 #include "mlx5_malloc.h"
22 struct mr_find_contig_memsegs_data {
26 const struct rte_memseg_list *msl;
29 /* Virtual memory range. */
35 /** Memory region for a mempool. */
36 struct mlx5_mempool_mr {
37 struct mlx5_pmd_mr pmd_mr;
38 uint32_t refcnt; /**< Number of mempools sharing this MR. */
41 /* Mempool registration. */
42 struct mlx5_mempool_reg {
43 LIST_ENTRY(mlx5_mempool_reg) next;
44 /** Registered mempool, used to designate registrations. */
45 struct rte_mempool *mp;
46 /** Memory regions for the address ranges of the mempool. */
47 struct mlx5_mempool_mr *mrs;
48 /** Number of memory regions. */
53 mlx5_mprq_buf_free_cb(void *addr __rte_unused, void *opaque)
55 struct mlx5_mprq_buf *buf = opaque;
57 if (__atomic_load_n(&buf->refcnt, __ATOMIC_RELAXED) == 1) {
58 rte_mempool_put(buf->mp, buf);
59 } else if (unlikely(__atomic_sub_fetch(&buf->refcnt, 1,
60 __ATOMIC_RELAXED) == 0)) {
61 __atomic_store_n(&buf->refcnt, 1, __ATOMIC_RELAXED);
62 rte_mempool_put(buf->mp, buf);
67 * Expand B-tree table to a given size. Can't be called with holding
68 * memory_hotplug_lock or share_cache.rwlock due to rte_realloc().
71 * Pointer to B-tree structure.
73 * Number of entries for expansion.
76 * 0 on success, -1 on failure.
79 mr_btree_expand(struct mlx5_mr_btree *bt, int n)
87 * Downside of directly using rte_realloc() is that SOCKET_ID_ANY is
88 * used inside if there's no room to expand. Because this is a quite
89 * rare case and a part of very slow path, it is very acceptable.
90 * Initially cache_bh[] will be given practically enough space and once
91 * it is expanded, expansion wouldn't be needed again ever.
93 mem = mlx5_realloc(bt->table, MLX5_MEM_RTE | MLX5_MEM_ZERO,
94 n * sizeof(struct mr_cache_entry), 0, SOCKET_ID_ANY);
96 /* Not an error, B-tree search will be skipped. */
97 DRV_LOG(WARNING, "failed to expand MR B-tree (%p) table",
101 DRV_LOG(DEBUG, "expanded MR B-tree table (size=%u)", n);
109 * Look up LKey from given B-tree lookup table, store the last index and return
113 * Pointer to B-tree structure.
115 * Pointer to index. Even on search failure, returns index where it stops
116 * searching so that index can be used when inserting a new entry.
121 * Searched LKey on success, UINT32_MAX on no match.
124 mr_btree_lookup(struct mlx5_mr_btree *bt, uint16_t *idx, uintptr_t addr)
126 struct mr_cache_entry *lkp_tbl;
130 MLX5_ASSERT(bt != NULL);
131 lkp_tbl = *bt->table;
133 /* First entry must be NULL for comparison. */
134 MLX5_ASSERT(bt->len > 0 || (lkp_tbl[0].start == 0 &&
135 lkp_tbl[0].lkey == UINT32_MAX));
138 register uint16_t delta = n >> 1;
140 if (addr < lkp_tbl[base + delta].start) {
147 MLX5_ASSERT(addr >= lkp_tbl[base].start);
149 if (addr < lkp_tbl[base].end)
150 return lkp_tbl[base].lkey;
156 * Insert an entry to B-tree lookup table.
159 * Pointer to B-tree structure.
161 * Pointer to new entry to insert.
164 * 0 on success, -1 on failure.
167 mr_btree_insert(struct mlx5_mr_btree *bt, struct mr_cache_entry *entry)
169 struct mr_cache_entry *lkp_tbl;
173 MLX5_ASSERT(bt != NULL);
174 MLX5_ASSERT(bt->len <= bt->size);
175 MLX5_ASSERT(bt->len > 0);
176 lkp_tbl = *bt->table;
177 /* Find out the slot for insertion. */
178 if (mr_btree_lookup(bt, &idx, entry->start) != UINT32_MAX) {
180 "abort insertion to B-tree(%p): already exist at"
181 " idx=%u [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
182 (void *)bt, idx, entry->start, entry->end, entry->lkey);
183 /* Already exist, return. */
186 /* If table is full, return error. */
187 if (unlikely(bt->len == bt->size)) {
193 shift = (bt->len - idx) * sizeof(struct mr_cache_entry);
195 memmove(&lkp_tbl[idx + 1], &lkp_tbl[idx], shift);
196 lkp_tbl[idx] = *entry;
199 "inserted B-tree(%p)[%u],"
200 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
201 (void *)bt, idx, entry->start, entry->end, entry->lkey);
206 * Initialize B-tree and allocate memory for lookup table.
209 * Pointer to B-tree structure.
211 * Number of entries to allocate.
213 * NUMA socket on which memory must be allocated.
216 * 0 on success, a negative errno value otherwise and rte_errno is set.
219 mlx5_mr_btree_init(struct mlx5_mr_btree *bt, int n, int socket)
225 MLX5_ASSERT(!bt->table && !bt->size);
226 memset(bt, 0, sizeof(*bt));
227 bt->table = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
228 sizeof(struct mr_cache_entry) * n,
230 if (bt->table == NULL) {
233 "failed to allocate memory for btree cache on socket "
238 /* First entry must be NULL for binary search. */
239 (*bt->table)[bt->len++] = (struct mr_cache_entry) {
242 DRV_LOG(DEBUG, "initialized B-tree %p with table %p",
243 (void *)bt, (void *)bt->table);
248 * Free B-tree resources.
251 * Pointer to B-tree structure.
254 mlx5_mr_btree_free(struct mlx5_mr_btree *bt)
258 DRV_LOG(DEBUG, "freeing B-tree %p with table %p",
259 (void *)bt, (void *)bt->table);
260 mlx5_free(bt->table);
261 memset(bt, 0, sizeof(*bt));
265 * Dump all the entries in a B-tree
268 * Pointer to B-tree structure.
271 mlx5_mr_btree_dump(struct mlx5_mr_btree *bt __rte_unused)
273 #ifdef RTE_LIBRTE_MLX5_DEBUG
275 struct mr_cache_entry *lkp_tbl;
279 lkp_tbl = *bt->table;
280 for (idx = 0; idx < bt->len; ++idx) {
281 struct mr_cache_entry *entry = &lkp_tbl[idx];
283 DRV_LOG(DEBUG, "B-tree(%p)[%u],"
284 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
285 (void *)bt, idx, entry->start, entry->end, entry->lkey);
291 * Initialize per-queue MR control descriptor.
294 * Pointer to MR control structure.
296 * Pointer to the mlx5 device structure.
298 * NUMA socket on which memory must be allocated.
301 * 0 on success, a negative errno value otherwise and rte_errno is set.
304 mlx5_mr_ctrl_init(struct mlx5_mr_ctrl *mr_ctrl, struct mlx5_common_device *cdev,
307 if (mr_ctrl == NULL) {
311 mr_ctrl->cdev = cdev;
312 /* Save pointer of global generation number to check memory event. */
313 mr_ctrl->dev_gen_ptr = &cdev->mr_scache.dev_gen;
314 /* Initialize B-tree and allocate memory for bottom-half cache table. */
315 return mlx5_mr_btree_init(&mr_ctrl->cache_bh, MLX5_MR_BTREE_CACHE_N,
320 * Find virtually contiguous memory chunk in a given MR.
323 * Pointer to MR structure.
325 * Pointer to returning MR cache entry. If not found, this will not be
328 * Start index of the memseg bitmap.
331 * Next index to go on lookup.
334 mr_find_next_chunk(struct mlx5_mr *mr, struct mr_cache_entry *entry,
341 /* MR for external memory doesn't have memseg list. */
342 if (mr->msl == NULL) {
343 MLX5_ASSERT(mr->ms_bmp_n == 1);
344 MLX5_ASSERT(mr->ms_n == 1);
345 MLX5_ASSERT(base_idx == 0);
347 * Can't search it from memseg list but get it directly from
348 * pmd_mr as there's only one chunk.
350 entry->start = (uintptr_t)mr->pmd_mr.addr;
351 entry->end = (uintptr_t)mr->pmd_mr.addr + mr->pmd_mr.len;
352 entry->lkey = rte_cpu_to_be_32(mr->pmd_mr.lkey);
353 /* Returning 1 ends iteration. */
356 for (idx = base_idx; idx < mr->ms_bmp_n; ++idx) {
357 if (rte_bitmap_get(mr->ms_bmp, idx)) {
358 const struct rte_memseg_list *msl;
359 const struct rte_memseg *ms;
362 ms = rte_fbarray_get(&msl->memseg_arr,
363 mr->ms_base_idx + idx);
364 MLX5_ASSERT(msl->page_sz == ms->hugepage_sz);
367 end = ms->addr_64 + ms->hugepage_sz;
369 /* Passed the end of a fragment. */
374 /* Found one chunk. */
375 entry->start = start;
377 entry->lkey = rte_cpu_to_be_32(mr->pmd_mr.lkey);
383 * Insert a MR to the global B-tree cache. It may fail due to low-on-memory.
384 * Then, this entry will have to be searched by mr_lookup_list() in
385 * mlx5_mr_create() on miss.
388 * Pointer to a global shared MR cache.
390 * Pointer to MR to insert.
393 * 0 on success, -1 on failure.
396 mlx5_mr_insert_cache(struct mlx5_mr_share_cache *share_cache,
401 DRV_LOG(DEBUG, "Inserting MR(%p) to global cache(%p)",
402 (void *)mr, (void *)share_cache);
403 for (n = 0; n < mr->ms_bmp_n; ) {
404 struct mr_cache_entry entry;
406 memset(&entry, 0, sizeof(entry));
407 /* Find a contiguous chunk and advance the index. */
408 n = mr_find_next_chunk(mr, &entry, n);
411 if (mr_btree_insert(&share_cache->cache, &entry) < 0) {
413 * Overflowed, but the global table cannot be expanded
414 * because of deadlock.
423 * Look up address in the original global MR list.
426 * Pointer to a global shared MR cache.
428 * Pointer to returning MR cache entry. If no match, this will not be updated.
433 * Found MR on match, NULL otherwise.
436 mlx5_mr_lookup_list(struct mlx5_mr_share_cache *share_cache,
437 struct mr_cache_entry *entry, uintptr_t addr)
441 /* Iterate all the existing MRs. */
442 LIST_FOREACH(mr, &share_cache->mr_list, mr) {
447 for (n = 0; n < mr->ms_bmp_n; ) {
448 struct mr_cache_entry ret;
450 memset(&ret, 0, sizeof(ret));
451 n = mr_find_next_chunk(mr, &ret, n);
452 if (addr >= ret.start && addr < ret.end) {
463 * Look up address on global MR cache.
466 * Pointer to a global shared MR cache.
468 * Pointer to returning MR cache entry. If no match, this will not be updated.
473 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
476 mlx5_mr_lookup_cache(struct mlx5_mr_share_cache *share_cache,
477 struct mr_cache_entry *entry, uintptr_t addr)
480 uint32_t lkey = UINT32_MAX;
484 * If the global cache has overflowed since it failed to expand the
485 * B-tree table, it can't have all the existing MRs. Then, the address
486 * has to be searched by traversing the original MR list instead, which
487 * is very slow path. Otherwise, the global cache is all inclusive.
489 if (!unlikely(share_cache->cache.overflow)) {
490 lkey = mr_btree_lookup(&share_cache->cache, &idx, addr);
491 if (lkey != UINT32_MAX)
492 *entry = (*share_cache->cache.table)[idx];
494 /* Falling back to the slowest path. */
495 mr = mlx5_mr_lookup_list(share_cache, entry, addr);
499 MLX5_ASSERT(lkey == UINT32_MAX || (addr >= entry->start &&
505 * Free MR resources. MR lock must not be held to avoid a deadlock. rte_free()
506 * can raise memory free event and the callback function will spin on the lock.
509 * Pointer to MR to free.
512 mlx5_mr_free(struct mlx5_mr *mr, mlx5_dereg_mr_t dereg_mr_cb)
516 DRV_LOG(DEBUG, "freeing MR(%p):", (void *)mr);
517 dereg_mr_cb(&mr->pmd_mr);
518 if (mr->ms_bmp != NULL)
519 rte_bitmap_free(mr->ms_bmp);
524 mlx5_mr_rebuild_cache(struct mlx5_mr_share_cache *share_cache)
528 DRV_LOG(DEBUG, "Rebuild dev cache[] %p", (void *)share_cache);
529 /* Flush cache to rebuild. */
530 share_cache->cache.len = 1;
531 share_cache->cache.overflow = 0;
532 /* Iterate all the existing MRs. */
533 LIST_FOREACH(mr, &share_cache->mr_list, mr)
534 if (mlx5_mr_insert_cache(share_cache, mr) < 0)
539 * Release resources of detached MR having no online entry.
542 * Pointer to a global shared MR cache.
545 mlx5_mr_garbage_collect(struct mlx5_mr_share_cache *share_cache)
547 struct mlx5_mr *mr_next;
548 struct mlx5_mr_list free_list = LIST_HEAD_INITIALIZER(free_list);
550 /* Must be called from the primary process. */
551 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
553 * MR can't be freed with holding the lock because rte_free() could call
554 * memory free callback function. This will be a deadlock situation.
556 rte_rwlock_write_lock(&share_cache->rwlock);
557 /* Detach the whole free list and release it after unlocking. */
558 free_list = share_cache->mr_free_list;
559 LIST_INIT(&share_cache->mr_free_list);
560 rte_rwlock_write_unlock(&share_cache->rwlock);
561 /* Release resources. */
562 mr_next = LIST_FIRST(&free_list);
563 while (mr_next != NULL) {
564 struct mlx5_mr *mr = mr_next;
566 mr_next = LIST_NEXT(mr, mr);
567 mlx5_mr_free(mr, share_cache->dereg_mr_cb);
571 /* Called during rte_memseg_contig_walk() by mlx5_mr_create(). */
573 mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl,
574 const struct rte_memseg *ms, size_t len, void *arg)
576 struct mr_find_contig_memsegs_data *data = arg;
578 if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len)
580 /* Found, save it and stop walking. */
581 data->start = ms->addr_64;
582 data->end = ms->addr_64 + len;
588 * Create a new global Memory Region (MR) for a missing virtual address.
589 * This API should be called on a secondary process, then a request is sent to
590 * the primary process in order to create a MR for the address. As the global MR
591 * list is on the shared memory, following LKey lookup should succeed unless the
595 * Pointer to the mlx5 common device.
597 * Pointer to a global shared MR cache.
599 * Pointer to returning MR cache entry, found in the global cache or newly
600 * created. If failed to create one, this will not be updated.
602 * Target virtual address to register.
605 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
608 mlx5_mr_create_secondary(struct mlx5_common_device *cdev,
609 struct mlx5_mr_share_cache *share_cache,
610 struct mr_cache_entry *entry, uintptr_t addr)
614 DRV_LOG(DEBUG, "Requesting MR creation for address (%p)", (void *)addr);
615 ret = mlx5_mp_req_mr_create(cdev, addr);
617 DRV_LOG(DEBUG, "Fail to request MR creation for address (%p)",
621 rte_rwlock_read_lock(&share_cache->rwlock);
622 /* Fill in output data. */
623 mlx5_mr_lookup_cache(share_cache, entry, addr);
624 /* Lookup can't fail. */
625 MLX5_ASSERT(entry->lkey != UINT32_MAX);
626 rte_rwlock_read_unlock(&share_cache->rwlock);
627 DRV_LOG(DEBUG, "MR CREATED by primary process for %p:\n"
628 " [0x%" PRIxPTR ", 0x%" PRIxPTR "), lkey=0x%x",
629 (void *)addr, entry->start, entry->end, entry->lkey);
634 * Create a new global Memory Region (MR) for a missing virtual address.
635 * Register entire virtually contiguous memory chunk around the address.
638 * Pointer to pd of a device (net, regex, vdpa,...).
640 * Pointer to a global shared MR cache.
642 * Pointer to returning MR cache entry, found in the global cache or newly
643 * created. If failed to create one, this will not be updated.
645 * Target virtual address to register.
646 * @param mr_ext_memseg_en
647 * Configurable flag about external memory segment enable or not.
650 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
653 mlx5_mr_create_primary(void *pd,
654 struct mlx5_mr_share_cache *share_cache,
655 struct mr_cache_entry *entry, uintptr_t addr,
656 unsigned int mr_ext_memseg_en)
658 struct mr_find_contig_memsegs_data data = {.addr = addr, };
659 struct mr_find_contig_memsegs_data data_re;
660 const struct rte_memseg_list *msl;
661 const struct rte_memseg *ms;
662 struct mlx5_mr *mr = NULL;
663 int ms_idx_shift = -1;
670 DRV_LOG(DEBUG, "Creating a MR using address (%p)", (void *)addr);
672 * Release detached MRs if any. This can't be called with holding either
673 * memory_hotplug_lock or share_cache->rwlock. MRs on the free list have
674 * been detached by the memory free event but it couldn't be released
675 * inside the callback due to deadlock. As a result, releasing resources
676 * is quite opportunistic.
678 mlx5_mr_garbage_collect(share_cache);
680 * If enabled, find out a contiguous virtual address chunk in use, to
681 * which the given address belongs, in order to register maximum range.
682 * In the best case where mempools are not dynamically recreated and
683 * '--socket-mem' is specified as an EAL option, it is very likely to
684 * have only one MR(LKey) per a socket and per a hugepage-size even
685 * though the system memory is highly fragmented. As the whole memory
686 * chunk will be pinned by kernel, it can't be reused unless entire
687 * chunk is freed from EAL.
689 * If disabled, just register one memseg (page). Then, memory
690 * consumption will be minimized but it may drop performance if there
691 * are many MRs to lookup on the datapath.
693 if (!mr_ext_memseg_en) {
694 data.msl = rte_mem_virt2memseg_list((void *)addr);
695 data.start = RTE_ALIGN_FLOOR(addr, data.msl->page_sz);
696 data.end = data.start + data.msl->page_sz;
697 } else if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) {
699 "Unable to find virtually contiguous"
700 " chunk for address (%p)."
701 " rte_memseg_contig_walk() failed.", (void *)addr);
706 /* Addresses must be page-aligned. */
707 MLX5_ASSERT(data.msl);
708 MLX5_ASSERT(rte_is_aligned((void *)data.start, data.msl->page_sz));
709 MLX5_ASSERT(rte_is_aligned((void *)data.end, data.msl->page_sz));
711 ms = rte_mem_virt2memseg((void *)data.start, msl);
712 len = data.end - data.start;
714 MLX5_ASSERT(msl->page_sz == ms->hugepage_sz);
715 /* Number of memsegs in the range. */
716 ms_n = len / msl->page_sz;
717 DRV_LOG(DEBUG, "Extending %p to [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
718 " page_sz=0x%" PRIx64 ", ms_n=%u",
719 (void *)addr, data.start, data.end, msl->page_sz, ms_n);
720 /* Size of memory for bitmap. */
721 bmp_size = rte_bitmap_get_memory_footprint(ms_n);
722 mr = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
723 RTE_ALIGN_CEIL(sizeof(*mr), RTE_CACHE_LINE_SIZE) +
724 bmp_size, RTE_CACHE_LINE_SIZE, msl->socket_id);
726 DRV_LOG(DEBUG, "Unable to allocate memory for a new MR of"
727 " address (%p).", (void *)addr);
733 * Save the index of the first memseg and initialize memseg bitmap. To
734 * see if a memseg of ms_idx in the memseg-list is still valid, check:
735 * rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx)
737 mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
738 bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE);
739 mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size);
740 if (mr->ms_bmp == NULL) {
741 DRV_LOG(DEBUG, "Unable to initialize bitmap for a new MR of"
742 " address (%p).", (void *)addr);
747 * Should recheck whether the extended contiguous chunk is still valid.
748 * Because memory_hotplug_lock can't be held if there's any memory
749 * related calls in a critical path, resource allocation above can't be
750 * locked. If the memory has been changed at this point, try again with
751 * just single page. If not, go on with the big chunk atomically from
754 rte_mcfg_mem_read_lock();
756 if (len > msl->page_sz &&
757 !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) {
759 "Unable to find virtually contiguous chunk for address "
760 "(%p). rte_memseg_contig_walk() failed.", (void *)addr);
764 if (data.start != data_re.start || data.end != data_re.end) {
766 * The extended contiguous chunk has been changed. Try again
767 * with single memseg instead.
769 data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz);
770 data.end = data.start + msl->page_sz;
771 rte_mcfg_mem_read_unlock();
772 mlx5_mr_free(mr, share_cache->dereg_mr_cb);
773 goto alloc_resources;
775 MLX5_ASSERT(data.msl == data_re.msl);
776 rte_rwlock_write_lock(&share_cache->rwlock);
778 * Check the address is really missing. If other thread already created
779 * one or it is not found due to overflow, abort and return.
781 if (mlx5_mr_lookup_cache(share_cache, entry, addr) != UINT32_MAX) {
783 * Insert to the global cache table. It may fail due to
784 * low-on-memory. Then, this entry will have to be searched
787 mr_btree_insert(&share_cache->cache, entry);
788 DRV_LOG(DEBUG, "Found MR for %p on final lookup, abort",
790 rte_rwlock_write_unlock(&share_cache->rwlock);
791 rte_mcfg_mem_read_unlock();
793 * Must be unlocked before calling rte_free() because
794 * mlx5_mr_mem_event_free_cb() can be called inside.
796 mlx5_mr_free(mr, share_cache->dereg_mr_cb);
800 * Trim start and end addresses for verbs MR. Set bits for registering
801 * memsegs but exclude already registered ones. Bitmap can be
804 for (n = 0; n < ms_n; ++n) {
806 struct mr_cache_entry ret;
808 memset(&ret, 0, sizeof(ret));
809 start = data_re.start + n * msl->page_sz;
810 /* Exclude memsegs already registered by other MRs. */
811 if (mlx5_mr_lookup_cache(share_cache, &ret, start) ==
814 * Start from the first unregistered memseg in the
817 if (ms_idx_shift == -1) {
818 mr->ms_base_idx += n;
822 data.end = start + msl->page_sz;
823 rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift);
827 len = data.end - data.start;
828 mr->ms_bmp_n = len / msl->page_sz;
829 MLX5_ASSERT(ms_idx_shift + mr->ms_bmp_n <= ms_n);
831 * Finally create an MR for the memory chunk. Verbs: ibv_reg_mr() can
832 * be called with holding the memory lock because it doesn't use
833 * mlx5_alloc_buf_extern() which eventually calls rte_malloc_socket()
834 * through mlx5_alloc_verbs_buf().
836 share_cache->reg_mr_cb(pd, (void *)data.start, len, &mr->pmd_mr);
837 if (mr->pmd_mr.obj == NULL) {
838 DRV_LOG(DEBUG, "Fail to create an MR for address (%p)",
843 MLX5_ASSERT((uintptr_t)mr->pmd_mr.addr == data.start);
844 MLX5_ASSERT(mr->pmd_mr.len);
845 LIST_INSERT_HEAD(&share_cache->mr_list, mr, mr);
846 DRV_LOG(DEBUG, "MR CREATED (%p) for %p:\n"
847 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
848 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
849 (void *)mr, (void *)addr, data.start, data.end,
850 rte_cpu_to_be_32(mr->pmd_mr.lkey),
851 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
852 /* Insert to the global cache table. */
853 mlx5_mr_insert_cache(share_cache, mr);
854 /* Fill in output data. */
855 mlx5_mr_lookup_cache(share_cache, entry, addr);
856 /* Lookup can't fail. */
857 MLX5_ASSERT(entry->lkey != UINT32_MAX);
858 rte_rwlock_write_unlock(&share_cache->rwlock);
859 rte_mcfg_mem_read_unlock();
862 rte_rwlock_write_unlock(&share_cache->rwlock);
864 rte_mcfg_mem_read_unlock();
867 * In case of error, as this can be called in a datapath, a warning
868 * message per an error is preferable instead. Must be unlocked before
869 * calling rte_free() because mlx5_mr_mem_event_free_cb() can be called
872 mlx5_mr_free(mr, share_cache->dereg_mr_cb);
877 * Create a new global Memory Region (MR) for a missing virtual address.
878 * This can be called from primary and secondary process.
881 * Pointer to the mlx5 common device.
883 * Pointer to a global shared MR cache.
885 * Pointer to returning MR cache entry, found in the global cache or newly
886 * created. If failed to create one, this will not be updated.
888 * Target virtual address to register.
891 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
894 mlx5_mr_create(struct mlx5_common_device *cdev,
895 struct mlx5_mr_share_cache *share_cache,
896 struct mr_cache_entry *entry, uintptr_t addr)
900 switch (rte_eal_process_type()) {
901 case RTE_PROC_PRIMARY:
902 ret = mlx5_mr_create_primary(cdev->pd, share_cache, entry, addr,
903 cdev->config.mr_ext_memseg_en);
905 case RTE_PROC_SECONDARY:
906 ret = mlx5_mr_create_secondary(cdev, share_cache, entry, addr);
915 * Look up address in the global MR cache table. If not found, create a new MR.
916 * Insert the found/created entry to local bottom-half cache table.
919 * Pointer to per-queue MR control structure.
921 * Pointer to returning MR cache entry, found in the global cache or newly
922 * created. If failed to create one, this is not written.
927 * Searched LKey on success, UINT32_MAX on no match.
930 mr_lookup_caches(struct mlx5_mr_ctrl *mr_ctrl,
931 struct mr_cache_entry *entry, uintptr_t addr)
933 struct mlx5_mr_share_cache *share_cache = &mr_ctrl->cdev->mr_scache;
934 struct mlx5_mr_btree *bt = &mr_ctrl->cache_bh;
938 /* If local cache table is full, try to double it. */
939 if (unlikely(bt->len == bt->size))
940 mr_btree_expand(bt, bt->size << 1);
941 /* Look up in the global cache. */
942 rte_rwlock_read_lock(&share_cache->rwlock);
943 lkey = mr_btree_lookup(&share_cache->cache, &idx, addr);
944 if (lkey != UINT32_MAX) {
946 *entry = (*share_cache->cache.table)[idx];
947 rte_rwlock_read_unlock(&share_cache->rwlock);
949 * Update local cache. Even if it fails, return the found entry
950 * to update top-half cache. Next time, this entry will be found
951 * in the global cache.
953 mr_btree_insert(bt, entry);
956 rte_rwlock_read_unlock(&share_cache->rwlock);
957 /* First time to see the address? Create a new MR. */
958 lkey = mlx5_mr_create(mr_ctrl->cdev, share_cache, entry, addr);
960 * Update the local cache if successfully created a new global MR. Even
961 * if failed to create one, there's no action to take in this datapath
962 * code. As returning LKey is invalid, this will eventually make HW
965 if (lkey != UINT32_MAX)
966 mr_btree_insert(bt, entry);
971 * Bottom-half of LKey search on datapath. First search in cache_bh[] and if
972 * misses, search in the global MR cache table and update the new entry to
973 * per-queue local caches.
976 * Pointer to per-queue MR control structure.
981 * Searched LKey on success, UINT32_MAX on no match.
984 mlx5_mr_addr2mr_bh(struct mlx5_mr_ctrl *mr_ctrl, uintptr_t addr)
988 /* Victim in top-half cache to replace with new entry. */
989 struct mr_cache_entry *repl = &mr_ctrl->cache[mr_ctrl->head];
991 /* Binary-search MR translation table. */
992 lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr);
993 /* Update top-half cache. */
994 if (likely(lkey != UINT32_MAX)) {
995 *repl = (*mr_ctrl->cache_bh.table)[bh_idx];
998 * If missed in local lookup table, search in the global cache
999 * and local cache_bh[] will be updated inside if possible.
1000 * Top-half cache entry will also be updated.
1002 lkey = mr_lookup_caches(mr_ctrl, repl, addr);
1003 if (unlikely(lkey == UINT32_MAX))
1006 /* Update the most recently used entry. */
1007 mr_ctrl->mru = mr_ctrl->head;
1008 /* Point to the next victim, the oldest. */
1009 mr_ctrl->head = (mr_ctrl->head + 1) % MLX5_MR_CACHE_N;
1014 * Release all the created MRs and resources on global MR cache of a device
1017 * @param share_cache
1018 * Pointer to a global shared MR cache.
1021 mlx5_mr_release_cache(struct mlx5_mr_share_cache *share_cache)
1023 struct mlx5_mr *mr_next;
1025 rte_rwlock_write_lock(&share_cache->rwlock);
1026 /* Detach from MR list and move to free list. */
1027 mr_next = LIST_FIRST(&share_cache->mr_list);
1028 while (mr_next != NULL) {
1029 struct mlx5_mr *mr = mr_next;
1031 mr_next = LIST_NEXT(mr, mr);
1032 LIST_REMOVE(mr, mr);
1033 LIST_INSERT_HEAD(&share_cache->mr_free_list, mr, mr);
1035 LIST_INIT(&share_cache->mr_list);
1036 /* Free global cache. */
1037 mlx5_mr_btree_free(&share_cache->cache);
1038 rte_rwlock_write_unlock(&share_cache->rwlock);
1039 /* Free all remaining MRs. */
1040 mlx5_mr_garbage_collect(share_cache);
1044 * Initialize global MR cache of a device.
1046 * @param share_cache
1047 * Pointer to a global shared MR cache.
1049 * NUMA socket on which memory must be allocated.
1052 * 0 on success, a negative errno value otherwise and rte_errno is set.
1055 mlx5_mr_create_cache(struct mlx5_mr_share_cache *share_cache, int socket)
1057 /* Set the reg_mr and dereg_mr callback functions */
1058 mlx5_os_set_reg_mr_cb(&share_cache->reg_mr_cb,
1059 &share_cache->dereg_mr_cb);
1060 rte_rwlock_init(&share_cache->rwlock);
1061 rte_rwlock_init(&share_cache->mprwlock);
1062 share_cache->mp_cb_registered = 0;
1063 /* Initialize B-tree and allocate memory for global MR cache table. */
1064 return mlx5_mr_btree_init(&share_cache->cache,
1065 MLX5_MR_BTREE_CACHE_N * 2, socket);
1069 * Flush all of the local cache entries.
1072 * Pointer to per-queue MR local cache.
1075 mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl)
1077 /* Reset the most-recently-used index. */
1079 /* Reset the linear search array. */
1081 memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache));
1082 /* Reset the B-tree table. */
1083 mr_ctrl->cache_bh.len = 1;
1084 mr_ctrl->cache_bh.overflow = 0;
1085 /* Update the generation number. */
1086 mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr;
1087 DRV_LOG(DEBUG, "mr_ctrl(%p): flushed, cur_gen=%d",
1088 (void *)mr_ctrl, mr_ctrl->cur_gen);
1092 * Creates a memory region for external memory, that is memory which is not
1093 * part of the DPDK memory segments.
1096 * Pointer to pd of a device (net, regex, vdpa,...).
1098 * Starting virtual address of memory.
1100 * Length of memory segment being mapped.
1102 * Socket to allocate heap memory for the control structures.
1105 * Pointer to MR structure on success, NULL otherwise.
1108 mlx5_create_mr_ext(void *pd, uintptr_t addr, size_t len, int socket_id,
1109 mlx5_reg_mr_t reg_mr_cb)
1111 struct mlx5_mr *mr = NULL;
1113 mr = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
1114 RTE_ALIGN_CEIL(sizeof(*mr), RTE_CACHE_LINE_SIZE),
1115 RTE_CACHE_LINE_SIZE, socket_id);
1118 reg_mr_cb(pd, (void *)addr, len, &mr->pmd_mr);
1119 if (mr->pmd_mr.obj == NULL) {
1121 "Fail to create MR for address (%p)",
1126 mr->msl = NULL; /* Mark it is external memory. */
1131 "MR CREATED (%p) for external memory %p:\n"
1132 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
1133 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
1134 (void *)mr, (void *)addr,
1135 addr, addr + len, rte_cpu_to_be_32(mr->pmd_mr.lkey),
1136 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
1141 * Callback for memory free event. Iterate freed memsegs and check whether it
1142 * belongs to an existing MR. If found, clear the bit from bitmap of MR. As a
1143 * result, the MR would be fragmented. If it becomes empty, the MR will be freed
1144 * later by mlx5_mr_garbage_collect(). Even if this callback is called from a
1145 * secondary process, the garbage collector will be called in primary process
1146 * as the secondary process can't call mlx5_mr_create().
1148 * The global cache must be rebuilt if there's any change and this event has to
1149 * be propagated to dataplane threads to flush the local caches.
1151 * @param share_cache
1152 * Pointer to a global shared MR cache.
1154 * Name of ibv device.
1156 * Address of freed memory.
1158 * Size of freed memory.
1161 mlx5_free_mr_by_addr(struct mlx5_mr_share_cache *share_cache,
1162 const char *ibdev_name, const void *addr, size_t len)
1164 const struct rte_memseg_list *msl;
1170 DRV_LOG(DEBUG, "device %s free callback: addr=%p, len=%zu",
1171 ibdev_name, addr, len);
1172 msl = rte_mem_virt2memseg_list(addr);
1173 /* addr and len must be page-aligned. */
1174 MLX5_ASSERT((uintptr_t)addr ==
1175 RTE_ALIGN((uintptr_t)addr, msl->page_sz));
1176 MLX5_ASSERT(len == RTE_ALIGN(len, msl->page_sz));
1177 ms_n = len / msl->page_sz;
1178 rte_rwlock_write_lock(&share_cache->rwlock);
1179 /* Clear bits of freed memsegs from MR. */
1180 for (i = 0; i < ms_n; ++i) {
1181 const struct rte_memseg *ms;
1182 struct mr_cache_entry entry;
1187 /* Find MR having this memseg. */
1188 start = (uintptr_t)addr + i * msl->page_sz;
1189 mr = mlx5_mr_lookup_list(share_cache, &entry, start);
1192 MLX5_ASSERT(mr->msl); /* Can't be external memory. */
1193 ms = rte_mem_virt2memseg((void *)start, msl);
1194 MLX5_ASSERT(ms != NULL);
1195 MLX5_ASSERT(msl->page_sz == ms->hugepage_sz);
1196 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
1197 pos = ms_idx - mr->ms_base_idx;
1198 MLX5_ASSERT(rte_bitmap_get(mr->ms_bmp, pos));
1199 MLX5_ASSERT(pos < mr->ms_bmp_n);
1200 DRV_LOG(DEBUG, "device %s MR(%p): clear bitmap[%u] for addr %p",
1201 ibdev_name, (void *)mr, pos, (void *)start);
1202 rte_bitmap_clear(mr->ms_bmp, pos);
1203 if (--mr->ms_n == 0) {
1204 LIST_REMOVE(mr, mr);
1205 LIST_INSERT_HEAD(&share_cache->mr_free_list, mr, mr);
1206 DRV_LOG(DEBUG, "device %s remove MR(%p) from list",
1207 ibdev_name, (void *)mr);
1210 * MR is fragmented or will be freed. the global cache must be
1216 mlx5_mr_rebuild_cache(share_cache);
1218 * No explicit wmb is needed after updating dev_gen due to
1219 * store-release ordering in unlock that provides the
1220 * implicit barrier at the software visible level.
1222 ++share_cache->dev_gen;
1223 DRV_LOG(DEBUG, "broadcasting local cache flush, gen=%d",
1224 share_cache->dev_gen);
1226 rte_rwlock_write_unlock(&share_cache->rwlock);
1230 * Dump all the created MRs and the global cache entries.
1232 * @param share_cache
1233 * Pointer to a global shared MR cache.
1236 mlx5_mr_dump_cache(struct mlx5_mr_share_cache *share_cache __rte_unused)
1238 #ifdef RTE_LIBRTE_MLX5_DEBUG
1243 rte_rwlock_read_lock(&share_cache->rwlock);
1244 /* Iterate all the existing MRs. */
1245 LIST_FOREACH(mr, &share_cache->mr_list, mr) {
1248 DRV_LOG(DEBUG, "MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u",
1249 mr_n++, rte_cpu_to_be_32(mr->pmd_mr.lkey),
1250 mr->ms_n, mr->ms_bmp_n);
1253 for (n = 0; n < mr->ms_bmp_n; ) {
1254 struct mr_cache_entry ret = { 0, };
1256 n = mr_find_next_chunk(mr, &ret, n);
1260 " chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")",
1261 chunk_n++, ret.start, ret.end);
1264 DRV_LOG(DEBUG, "Dumping global cache %p", (void *)share_cache);
1265 mlx5_mr_btree_dump(&share_cache->cache);
1266 rte_rwlock_read_unlock(&share_cache->rwlock);
1271 mlx5_range_compare_start(const void *lhs, const void *rhs)
1273 const struct mlx5_range *r1 = lhs, *r2 = rhs;
1275 if (r1->start > r2->start)
1277 else if (r1->start < r2->start)
1283 mlx5_range_from_mempool_chunk(struct rte_mempool *mp, void *opaque,
1284 struct rte_mempool_memhdr *memhdr,
1287 struct mlx5_range *ranges = opaque, *range = &ranges[idx];
1288 uint64_t page_size = rte_mem_page_size();
1291 range->start = RTE_ALIGN_FLOOR((uintptr_t)memhdr->addr, page_size);
1292 range->end = RTE_ALIGN_CEIL(range->start + memhdr->len, page_size);
1296 * Get VA-contiguous ranges of the mempool memory.
1297 * Each range start and end is aligned to the system page size.
1302 * Receives the ranges, caller must release it with free().
1303 * @param[out] ount_n
1304 * Receives the number of @p out elements.
1307 * 0 on success, (-1) on failure.
1310 mlx5_get_mempool_ranges(struct rte_mempool *mp, struct mlx5_range **out,
1311 unsigned int *out_n)
1313 struct mlx5_range *chunks;
1314 unsigned int chunks_n = mp->nb_mem_chunks, contig_n, i;
1316 /* Collect page-aligned memory ranges of the mempool. */
1317 chunks = calloc(sizeof(chunks[0]), chunks_n);
1320 rte_mempool_mem_iter(mp, mlx5_range_from_mempool_chunk, chunks);
1321 /* Merge adjacent chunks and place them at the beginning. */
1322 qsort(chunks, chunks_n, sizeof(chunks[0]), mlx5_range_compare_start);
1324 for (i = 1; i < chunks_n; i++)
1325 if (chunks[i - 1].end != chunks[i].start) {
1326 chunks[contig_n - 1].end = chunks[i - 1].end;
1327 chunks[contig_n] = chunks[i];
1330 /* Extend the last contiguous chunk to the end of the mempool. */
1331 chunks[contig_n - 1].end = chunks[i - 1].end;
1338 * Analyze mempool memory to select memory ranges to register.
1341 * Mempool to analyze.
1343 * Receives memory ranges to register, aligned to the system page size.
1344 * The caller must release them with free().
1346 * Receives the number of @p out items.
1347 * @param[out] share_hugepage
1348 * Receives True if the entire pool resides within a single hugepage.
1351 * 0 on success, (-1) on failure.
1354 mlx5_mempool_reg_analyze(struct rte_mempool *mp, struct mlx5_range **out,
1355 unsigned int *out_n, bool *share_hugepage)
1357 struct mlx5_range *ranges = NULL;
1358 unsigned int i, ranges_n = 0;
1359 struct rte_memseg_list *msl;
1361 if (mlx5_get_mempool_ranges(mp, &ranges, &ranges_n) < 0) {
1362 DRV_LOG(ERR, "Cannot get address ranges for mempool %s",
1366 /* Check if the hugepage of the pool can be shared. */
1367 *share_hugepage = false;
1368 msl = rte_mem_virt2memseg_list((void *)ranges[0].start);
1370 uint64_t hugepage_sz = 0;
1372 /* Check that all ranges are on pages of the same size. */
1373 for (i = 0; i < ranges_n; i++) {
1374 if (hugepage_sz != 0 && hugepage_sz != msl->page_sz)
1376 hugepage_sz = msl->page_sz;
1378 if (i == ranges_n) {
1380 * If the entire pool is within one hugepage,
1381 * combine all ranges into one of the hugepage size.
1383 uintptr_t reg_start = ranges[0].start;
1384 uintptr_t reg_end = ranges[ranges_n - 1].end;
1385 uintptr_t hugepage_start =
1386 RTE_ALIGN_FLOOR(reg_start, hugepage_sz);
1387 uintptr_t hugepage_end = hugepage_start + hugepage_sz;
1388 if (reg_end < hugepage_end) {
1389 ranges[0].start = hugepage_start;
1390 ranges[0].end = hugepage_end;
1392 *share_hugepage = true;
1401 /** Create a registration object for the mempool. */
1402 static struct mlx5_mempool_reg *
1403 mlx5_mempool_reg_create(struct rte_mempool *mp, unsigned int mrs_n)
1405 struct mlx5_mempool_reg *mpr = NULL;
1407 mpr = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
1408 sizeof(*mpr) + mrs_n * sizeof(mpr->mrs[0]),
1409 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1411 DRV_LOG(ERR, "Cannot allocate mempool %s registration object",
1416 mpr->mrs = (struct mlx5_mempool_mr *)(mpr + 1);
1422 * Destroy a mempool registration object.
1425 * Whether @p mpr owns its MRs excludively, i.e. they are not shared.
1428 mlx5_mempool_reg_destroy(struct mlx5_mr_share_cache *share_cache,
1429 struct mlx5_mempool_reg *mpr, bool standalone)
1434 for (i = 0; i < mpr->mrs_n; i++)
1435 share_cache->dereg_mr_cb(&mpr->mrs[i].pmd_mr);
1440 /** Find registration object of a mempool. */
1441 static struct mlx5_mempool_reg *
1442 mlx5_mempool_reg_lookup(struct mlx5_mr_share_cache *share_cache,
1443 struct rte_mempool *mp)
1445 struct mlx5_mempool_reg *mpr;
1447 LIST_FOREACH(mpr, &share_cache->mempool_reg_list, next)
1453 /** Increment reference counters of MRs used in the registration. */
1455 mlx5_mempool_reg_attach(struct mlx5_mempool_reg *mpr)
1459 for (i = 0; i < mpr->mrs_n; i++)
1460 __atomic_add_fetch(&mpr->mrs[i].refcnt, 1, __ATOMIC_RELAXED);
1464 * Decrement reference counters of MRs used in the registration.
1466 * @return True if no more references to @p mpr MRs exist, False otherwise.
1469 mlx5_mempool_reg_detach(struct mlx5_mempool_reg *mpr)
1474 for (i = 0; i < mpr->mrs_n; i++)
1475 ret |= __atomic_sub_fetch(&mpr->mrs[i].refcnt, 1,
1476 __ATOMIC_RELAXED) == 0;
1481 mlx5_mr_mempool_register_primary(struct mlx5_mr_share_cache *share_cache,
1482 void *pd, struct rte_mempool *mp)
1484 struct mlx5_range *ranges = NULL;
1485 struct mlx5_mempool_reg *mpr, *new_mpr;
1486 unsigned int i, ranges_n;
1487 bool share_hugepage;
1490 /* Early check to avoid unnecessary creation of MRs. */
1491 rte_rwlock_read_lock(&share_cache->rwlock);
1492 mpr = mlx5_mempool_reg_lookup(share_cache, mp);
1493 rte_rwlock_read_unlock(&share_cache->rwlock);
1495 DRV_LOG(DEBUG, "Mempool %s is already registered for PD %p",
1500 if (mlx5_mempool_reg_analyze(mp, &ranges, &ranges_n,
1501 &share_hugepage) < 0) {
1502 DRV_LOG(ERR, "Cannot get mempool %s memory ranges", mp->name);
1506 new_mpr = mlx5_mempool_reg_create(mp, ranges_n);
1507 if (new_mpr == NULL) {
1509 "Cannot create a registration object for mempool %s in PD %p",
1515 * If the entire mempool fits in a single hugepage, the MR for this
1516 * hugepage can be shared across mempools that also fit in it.
1518 if (share_hugepage) {
1519 rte_rwlock_write_lock(&share_cache->rwlock);
1520 LIST_FOREACH(mpr, &share_cache->mempool_reg_list, next) {
1521 if (mpr->mrs[0].pmd_mr.addr == (void *)ranges[0].start)
1525 new_mpr->mrs = mpr->mrs;
1526 mlx5_mempool_reg_attach(new_mpr);
1527 LIST_INSERT_HEAD(&share_cache->mempool_reg_list,
1530 rte_rwlock_write_unlock(&share_cache->rwlock);
1532 DRV_LOG(DEBUG, "Shared MR %#x in PD %p for mempool %s with mempool %s",
1533 mpr->mrs[0].pmd_mr.lkey, pd, mp->name,
1539 for (i = 0; i < ranges_n; i++) {
1540 struct mlx5_mempool_mr *mr = &new_mpr->mrs[i];
1541 const struct mlx5_range *range = &ranges[i];
1542 size_t len = range->end - range->start;
1544 if (share_cache->reg_mr_cb(pd, (void *)range->start, len,
1547 "Failed to create an MR in PD %p for address range "
1548 "[0x%" PRIxPTR ", 0x%" PRIxPTR "] (%zu bytes) for mempool %s",
1549 pd, range->start, range->end, len, mp->name);
1553 "Created a new MR %#x in PD %p for address range "
1554 "[0x%" PRIxPTR ", 0x%" PRIxPTR "] (%zu bytes) for mempool %s",
1555 mr->pmd_mr.lkey, pd, range->start, range->end, len,
1558 if (i != ranges_n) {
1559 mlx5_mempool_reg_destroy(share_cache, new_mpr, true);
1563 /* Concurrent registration is not supposed to happen. */
1564 rte_rwlock_write_lock(&share_cache->rwlock);
1565 mpr = mlx5_mempool_reg_lookup(share_cache, mp);
1567 mlx5_mempool_reg_attach(new_mpr);
1568 LIST_INSERT_HEAD(&share_cache->mempool_reg_list, new_mpr, next);
1571 rte_rwlock_write_unlock(&share_cache->rwlock);
1573 DRV_LOG(DEBUG, "Mempool %s is already registered for PD %p",
1575 mlx5_mempool_reg_destroy(share_cache, new_mpr, true);
1585 mlx5_mr_mempool_register_secondary(struct mlx5_common_device *cdev,
1586 struct rte_mempool *mp)
1588 return mlx5_mp_req_mempool_reg(cdev, mp, true);
1592 * Register the memory of a mempool in the protection domain.
1595 * Pointer to the mlx5 common device.
1597 * Mempool to register.
1600 * 0 on success, (-1) on failure and rte_errno is set.
1603 mlx5_mr_mempool_register(struct mlx5_common_device *cdev,
1604 struct rte_mempool *mp)
1606 if (mp->flags & RTE_MEMPOOL_F_NON_IO)
1608 switch (rte_eal_process_type()) {
1609 case RTE_PROC_PRIMARY:
1610 return mlx5_mr_mempool_register_primary(&cdev->mr_scache,
1612 case RTE_PROC_SECONDARY:
1613 return mlx5_mr_mempool_register_secondary(cdev, mp);
1620 mlx5_mr_mempool_unregister_primary(struct mlx5_mr_share_cache *share_cache,
1621 struct rte_mempool *mp)
1623 struct mlx5_mempool_reg *mpr;
1624 bool standalone = false;
1626 rte_rwlock_write_lock(&share_cache->rwlock);
1627 LIST_FOREACH(mpr, &share_cache->mempool_reg_list, next)
1628 if (mpr->mp == mp) {
1629 LIST_REMOVE(mpr, next);
1630 standalone = mlx5_mempool_reg_detach(mpr);
1633 * The unlock operation below provides a memory
1634 * barrier due to its store-release semantics.
1636 ++share_cache->dev_gen;
1639 rte_rwlock_write_unlock(&share_cache->rwlock);
1644 mlx5_mempool_reg_destroy(share_cache, mpr, standalone);
1649 mlx5_mr_mempool_unregister_secondary(struct mlx5_common_device *cdev,
1650 struct rte_mempool *mp)
1652 return mlx5_mp_req_mempool_reg(cdev, mp, false);
1656 * Unregister the memory of a mempool from the protection domain.
1659 * Pointer to the mlx5 common device.
1661 * Mempool to unregister.
1664 * 0 on success, (-1) on failure and rte_errno is set.
1667 mlx5_mr_mempool_unregister(struct mlx5_common_device *cdev,
1668 struct rte_mempool *mp)
1670 if (mp->flags & RTE_MEMPOOL_F_NON_IO)
1672 switch (rte_eal_process_type()) {
1673 case RTE_PROC_PRIMARY:
1674 return mlx5_mr_mempool_unregister_primary(&cdev->mr_scache, mp);
1675 case RTE_PROC_SECONDARY:
1676 return mlx5_mr_mempool_unregister_secondary(cdev, mp);
1683 * Lookup a MR key by and address in a registered mempool.
1686 * Mempool registration object.
1688 * Address within the mempool.
1690 * Bottom-half cache entry to fill.
1693 * MR key or UINT32_MAX on failure, which can only happen
1694 * if the address is not from within the mempool.
1697 mlx5_mempool_reg_addr2mr(struct mlx5_mempool_reg *mpr, uintptr_t addr,
1698 struct mr_cache_entry *entry)
1700 uint32_t lkey = UINT32_MAX;
1703 for (i = 0; i < mpr->mrs_n; i++) {
1704 const struct mlx5_pmd_mr *mr = &mpr->mrs[i].pmd_mr;
1705 uintptr_t mr_addr = (uintptr_t)mr->addr;
1707 if (mr_addr <= addr) {
1708 lkey = rte_cpu_to_be_32(mr->lkey);
1709 entry->start = mr_addr;
1710 entry->end = mr_addr + mr->len;
1719 * Update bottom-half cache from the list of mempool registrations.
1721 * @param share_cache
1722 * Pointer to a global shared MR cache.
1724 * Per-queue MR control handle.
1726 * Pointer to an entry in the bottom-half cache to update
1727 * with the MR lkey looked up.
1729 * Mempool containing the address.
1731 * Address to lookup.
1733 * MR lkey on success, UINT32_MAX on failure.
1736 mlx5_lookup_mempool_regs(struct mlx5_mr_share_cache *share_cache,
1737 struct mlx5_mr_ctrl *mr_ctrl,
1738 struct mr_cache_entry *entry,
1739 struct rte_mempool *mp, uintptr_t addr)
1741 struct mlx5_mr_btree *bt = &mr_ctrl->cache_bh;
1742 struct mlx5_mempool_reg *mpr;
1743 uint32_t lkey = UINT32_MAX;
1745 /* If local cache table is full, try to double it. */
1746 if (unlikely(bt->len == bt->size))
1747 mr_btree_expand(bt, bt->size << 1);
1748 /* Look up in mempool registrations. */
1749 rte_rwlock_read_lock(&share_cache->rwlock);
1750 mpr = mlx5_mempool_reg_lookup(share_cache, mp);
1752 lkey = mlx5_mempool_reg_addr2mr(mpr, addr, entry);
1753 rte_rwlock_read_unlock(&share_cache->rwlock);
1755 * Update local cache. Even if it fails, return the found entry
1756 * to update top-half cache. Next time, this entry will be found
1757 * in the global cache.
1759 if (lkey != UINT32_MAX)
1760 mr_btree_insert(bt, entry);
1765 * Bottom-half lookup for the address from the mempool.
1767 * @param share_cache
1768 * Pointer to a global shared MR cache.
1770 * Per-queue MR control handle.
1772 * Mempool containing the address.
1774 * Address to lookup.
1776 * MR lkey on success, UINT32_MAX on failure.
1779 mlx5_mr_mempool2mr_bh(struct mlx5_mr_share_cache *share_cache,
1780 struct mlx5_mr_ctrl *mr_ctrl,
1781 struct rte_mempool *mp, uintptr_t addr)
1783 struct mr_cache_entry *repl = &mr_ctrl->cache[mr_ctrl->head];
1785 uint16_t bh_idx = 0;
1787 /* Binary-search MR translation table. */
1788 lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr);
1789 /* Update top-half cache. */
1790 if (likely(lkey != UINT32_MAX)) {
1791 *repl = (*mr_ctrl->cache_bh.table)[bh_idx];
1793 lkey = mlx5_lookup_mempool_regs(share_cache, mr_ctrl, repl,
1795 /* Can only fail if the address is not from the mempool. */
1796 if (unlikely(lkey == UINT32_MAX))
1799 /* Update the most recently used entry. */
1800 mr_ctrl->mru = mr_ctrl->head;
1801 /* Point to the next victim, the oldest. */
1802 mr_ctrl->head = (mr_ctrl->head + 1) % MLX5_MR_CACHE_N;
1807 mlx5_mr_mb2mr_bh(struct mlx5_mr_ctrl *mr_ctrl, struct rte_mbuf *mb)
1810 uintptr_t addr = (uintptr_t)mb->buf_addr;
1811 struct mlx5_common_device *cdev = mr_ctrl->cdev;
1813 if (cdev->config.mr_mempool_reg_en) {
1814 struct rte_mempool *mp = NULL;
1815 struct mlx5_mprq_buf *buf;
1817 if (!RTE_MBUF_HAS_EXTBUF(mb)) {
1818 mp = mlx5_mb2mp(mb);
1819 } else if (mb->shinfo->free_cb == mlx5_mprq_buf_free_cb) {
1820 /* Recover MPRQ mempool. */
1821 buf = mb->shinfo->fcb_opaque;
1825 lkey = mlx5_mr_mempool2mr_bh(&cdev->mr_scache,
1828 * Lookup can only fail on invalid input, e.g. "addr"
1829 * is not from "mp" or "mp" has MEMPOOL_F_NON_IO set.
1831 if (lkey != UINT32_MAX)
1834 /* Fallback for generic mechanism in corner cases. */
1836 return mlx5_mr_addr2mr_bh(mr_ctrl, addr);