1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2016 6WIND S.A.
3 * Copyright 2016 Mellanox.
7 #pragma GCC diagnostic ignored "-Wpedantic"
9 #include <infiniband/verbs.h>
11 #pragma GCC diagnostic error "-Wpedantic"
14 #include <rte_mempool.h>
15 #include <rte_malloc.h>
18 #include "mlx5_rxtx.h"
19 #include "mlx5_glue.h"
21 struct mlx5_check_mempool_data {
27 /* Called by mlx5_check_mempool() when iterating the memory chunks. */
29 mlx5_check_mempool_cb(struct rte_mempool *mp __rte_unused,
30 void *opaque, struct rte_mempool_memhdr *memhdr,
31 unsigned int mem_idx __rte_unused)
33 struct mlx5_check_mempool_data *data = opaque;
35 /* It already failed, skip the next chunks. */
38 /* It is the first chunk. */
39 if (data->start == NULL && data->end == NULL) {
40 data->start = memhdr->addr;
41 data->end = data->start + memhdr->len;
44 if (data->end == memhdr->addr) {
45 data->end += memhdr->len;
48 if (data->start == (char *)memhdr->addr + memhdr->len) {
49 data->start -= memhdr->len;
52 /* Error, mempool is not virtually contiguous. */
57 * Check if a mempool can be used: it must be virtually contiguous.
60 * Pointer to memory pool.
62 * Pointer to the start address of the mempool virtual memory area
64 * Pointer to the end address of the mempool virtual memory area
67 * 0 on success (mempool is virtually contiguous), -1 on error.
70 mlx5_check_mempool(struct rte_mempool *mp, uintptr_t *start,
73 struct mlx5_check_mempool_data data;
75 memset(&data, 0, sizeof(data));
76 rte_mempool_mem_iter(mp, mlx5_check_mempool_cb, &data);
77 *start = (uintptr_t)data.start;
78 *end = (uintptr_t)data.end;
83 * Register a Memory Region (MR) <-> Memory Pool (MP) association in
84 * txq->mp2mr[]. If mp2mr[] is full, remove an entry first.
86 * This function should only be called by txq_mp2mr().
89 * Pointer to private structure.
91 * Pointer to TX queue structure.
93 * Memory Pool for which a Memory Region lkey must be returned.
95 * Index of the next available entry.
98 * mr on success, NULL on failure.
101 priv_txq_mp2mr_reg(struct priv *priv, struct mlx5_txq_data *txq,
102 struct rte_mempool *mp, unsigned int idx)
104 struct mlx5_txq_ctrl *txq_ctrl =
105 container_of(txq, struct mlx5_txq_ctrl, txq);
108 /* Add a new entry, register MR first. */
109 DEBUG("%p: discovered new memory pool \"%s\" (%p)",
110 (void *)txq_ctrl, mp->name, (void *)mp);
111 mr = priv_mr_get(priv, mp);
113 if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
114 DEBUG("Using unregistered mempool 0x%p(%s) in secondary process,"
115 " please create mempool before rte_eth_dev_start()",
116 (void *)mp, mp->name);
119 mr = priv_mr_new(priv, mp);
121 if (unlikely(mr == NULL)) {
122 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
126 if (unlikely(idx == RTE_DIM(txq->mp2mr))) {
127 /* Table is full, remove oldest entry. */
128 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
131 priv_mr_release(priv, txq->mp2mr[0]);
132 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
133 (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
135 /* Store the new entry. */
136 txq_ctrl->txq.mp2mr[idx] = mr;
137 DEBUG("%p: new MR lkey for MP \"%s\" (%p): 0x%08" PRIu32,
138 (void *)txq_ctrl, mp->name, (void *)mp,
139 txq_ctrl->txq.mp2mr[idx]->lkey);
144 * Register a Memory Region (MR) <-> Memory Pool (MP) association in
145 * txq->mp2mr[]. If mp2mr[] is full, remove an entry first.
147 * This function should only be called by txq_mp2mr().
150 * Pointer to TX queue structure.
152 * Memory Pool for which a Memory Region lkey must be returned.
154 * Index of the next available entry.
157 * mr on success, NULL on failure.
160 mlx5_txq_mp2mr_reg(struct mlx5_txq_data *txq, struct rte_mempool *mp,
163 struct mlx5_txq_ctrl *txq_ctrl =
164 container_of(txq, struct mlx5_txq_ctrl, txq);
167 priv_lock(txq_ctrl->priv);
168 mr = priv_txq_mp2mr_reg(txq_ctrl->priv, txq, mp, idx);
169 priv_unlock(txq_ctrl->priv);
173 struct mlx5_mp2mr_mbuf_check_data {
178 * Callback function for rte_mempool_obj_iter() to check whether a given
179 * mempool object looks like a mbuf.
182 * The mempool pointer
184 * Context data (struct txq_mp2mr_mbuf_check_data). Contains the
189 * Object index, unused.
192 txq_mp2mr_mbuf_check(struct rte_mempool *mp, void *arg, void *obj,
193 uint32_t index __rte_unused)
195 struct mlx5_mp2mr_mbuf_check_data *data = arg;
196 struct rte_mbuf *buf = obj;
199 * Check whether mbuf structure fits element size and whether mempool
202 if (sizeof(*buf) > mp->elt_size || buf->pool != mp)
207 * Iterator function for rte_mempool_walk() to register existing mempools and
208 * fill the MP to MR cache of a TX queue.
211 * Memory Pool to register.
213 * Pointer to TX queue structure.
216 mlx5_mp2mr_iter(struct rte_mempool *mp, void *arg)
218 struct priv *priv = (struct priv *)arg;
219 struct mlx5_mp2mr_mbuf_check_data data = {
224 /* Register mempool only if the first element looks like a mbuf. */
225 if (rte_mempool_obj_iter(mp, txq_mp2mr_mbuf_check, &data) == 0 ||
228 mr = priv_mr_get(priv, mp);
230 priv_mr_release(priv, mr);
233 priv_mr_new(priv, mp);
237 * Register a new memory region from the mempool and store it in the memory
241 * Pointer to private structure.
243 * Pointer to the memory pool to register.
246 * The memory region on success.
249 priv_mr_new(struct priv *priv, struct rte_mempool *mp)
251 const struct rte_memseg *ms = rte_eal_get_physmem_layout();
257 mr = rte_zmalloc_socket(__func__, sizeof(*mr), 0, mp->socket_id);
259 DEBUG("unable to configure MR, ibv_reg_mr() failed.");
262 if (mlx5_check_mempool(mp, &start, &end) != 0) {
263 ERROR("mempool %p: not virtually contiguous",
267 DEBUG("mempool %p area start=%p end=%p size=%zu",
268 (void *)mp, (void *)start, (void *)end,
269 (size_t)(end - start));
270 /* Save original addresses for exact MR lookup. */
273 /* Round start and end to page boundary if found in memory segments. */
274 for (i = 0; (i < RTE_MAX_MEMSEG) && (ms[i].addr != NULL); ++i) {
275 uintptr_t addr = (uintptr_t)ms[i].addr;
276 size_t len = ms[i].len;
277 unsigned int align = ms[i].hugepage_sz;
279 if ((start > addr) && (start < addr + len))
280 start = RTE_ALIGN_FLOOR(start, align);
281 if ((end > addr) && (end < addr + len))
282 end = RTE_ALIGN_CEIL(end, align);
284 DEBUG("mempool %p using start=%p end=%p size=%zu for MR",
285 (void *)mp, (void *)start, (void *)end,
286 (size_t)(end - start));
287 mr->mr = mlx5_glue->reg_mr(priv->pd, (void *)start, end - start,
288 IBV_ACCESS_LOCAL_WRITE);
290 mr->lkey = rte_cpu_to_be_32(mr->mr->lkey);
291 rte_atomic32_inc(&mr->refcnt);
292 DEBUG("%p: new Memory Region %p refcnt: %d", (void *)priv,
293 (void *)mr, rte_atomic32_read(&mr->refcnt));
294 LIST_INSERT_HEAD(&priv->mr, mr, next);
299 * Search the memory region object in the memory region list.
302 * Pointer to private structure.
304 * Pointer to the memory pool to register.
307 * The memory region on success.
310 priv_mr_get(struct priv *priv, struct rte_mempool *mp)
315 if (LIST_EMPTY(&priv->mr))
317 LIST_FOREACH(mr, &priv->mr, next) {
319 rte_atomic32_inc(&mr->refcnt);
320 DEBUG("Memory Region %p refcnt: %d",
321 (void *)mr, rte_atomic32_read(&mr->refcnt));
329 * Release the memory region object.
332 * Pointer to memory region to release.
335 * 0 on success, errno on failure.
338 priv_mr_release(struct priv *priv __rte_unused, struct mlx5_mr *mr)
341 DEBUG("Memory Region %p refcnt: %d",
342 (void *)mr, rte_atomic32_read(&mr->refcnt));
343 if (rte_atomic32_dec_and_test(&mr->refcnt)) {
344 claim_zero(mlx5_glue->dereg_mr(mr->mr));
345 LIST_REMOVE(mr, next);
353 * Verify the flow list is empty
356 * Pointer to private structure.
359 * The number of object not released.
362 priv_mr_verify(struct priv *priv)
367 LIST_FOREACH(mr, &priv->mr, next) {
368 DEBUG("%p: mr %p still referenced", (void *)priv,