4 * Copyright 2015 6WIND S.A.
5 * Copyright 2015 Mellanox.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of 6WIND S.A. nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
41 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
43 #pragma GCC diagnostic ignored "-pedantic"
45 #include <infiniband/verbs.h>
47 #pragma GCC diagnostic error "-pedantic"
50 /* DPDK headers don't like -pedantic. */
52 #pragma GCC diagnostic ignored "-pedantic"
55 #include <rte_malloc.h>
56 #include <rte_ethdev.h>
57 #include <rte_common.h>
59 #pragma GCC diagnostic error "-pedantic"
62 #include "mlx5_utils.h"
64 #include "mlx5_rxtx.h"
65 #include "mlx5_autoconf.h"
66 #include "mlx5_defs.h"
69 * Allocate TX queue elements.
72 * Pointer to TX queue structure.
74 * Number of elements to allocate.
77 * 0 on success, errno value on failure.
80 txq_alloc_elts(struct txq *txq, unsigned int elts_n)
83 struct txq_elt (*elts)[elts_n] =
84 rte_calloc_socket("TXQ", 1, sizeof(*elts), 0, txq->socket);
85 linear_t (*elts_linear)[elts_n] =
86 rte_calloc_socket("TXQ", 1, sizeof(*elts_linear), 0,
88 struct ibv_mr *mr_linear = NULL;
91 if ((elts == NULL) || (elts_linear == NULL)) {
92 ERROR("%p: can't allocate packets array", (void *)txq);
97 ibv_reg_mr(txq->priv->pd, elts_linear, sizeof(*elts_linear),
98 (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
99 if (mr_linear == NULL) {
100 ERROR("%p: unable to configure MR, ibv_reg_mr() failed",
105 for (i = 0; (i != elts_n); ++i) {
106 struct txq_elt *elt = &(*elts)[i];
110 DEBUG("%p: allocated and configured %u WRs", (void *)txq, elts_n);
111 txq->elts_n = elts_n;
116 /* Request send completion every MLX5_PMD_TX_PER_COMP_REQ packets or
117 * at least 4 times per ring. */
118 txq->elts_comp_cd_init =
119 ((MLX5_PMD_TX_PER_COMP_REQ < (elts_n / 4)) ?
120 MLX5_PMD_TX_PER_COMP_REQ : (elts_n / 4));
121 txq->elts_comp_cd = txq->elts_comp_cd_init;
122 txq->elts_linear = elts_linear;
123 txq->mr_linear = mr_linear;
127 if (mr_linear != NULL)
128 claim_zero(ibv_dereg_mr(mr_linear));
130 rte_free(elts_linear);
133 DEBUG("%p: failed, freed everything", (void *)txq);
139 * Free TX queue elements.
142 * Pointer to TX queue structure.
145 txq_free_elts(struct txq *txq)
148 unsigned int elts_n = txq->elts_n;
149 struct txq_elt (*elts)[elts_n] = txq->elts;
150 linear_t (*elts_linear)[elts_n] = txq->elts_linear;
151 struct ibv_mr *mr_linear = txq->mr_linear;
153 DEBUG("%p: freeing WRs", (void *)txq);
156 txq->elts_linear = NULL;
157 txq->mr_linear = NULL;
158 if (mr_linear != NULL)
159 claim_zero(ibv_dereg_mr(mr_linear));
161 rte_free(elts_linear);
164 for (i = 0; (i != RTE_DIM(*elts)); ++i) {
165 struct txq_elt *elt = &(*elts)[i];
167 if (elt->buf == NULL)
169 rte_pktmbuf_free(elt->buf);
175 * Clean up a TX queue.
177 * Destroy objects, free allocated memory and reset the structure for reuse.
180 * Pointer to TX queue structure.
183 txq_cleanup(struct txq *txq)
185 struct ibv_exp_release_intf_params params;
188 DEBUG("cleaning up %p", (void *)txq);
190 if (txq->if_qp != NULL) {
191 assert(txq->priv != NULL);
192 assert(txq->priv->ctx != NULL);
193 assert(txq->qp != NULL);
194 params = (struct ibv_exp_release_intf_params){
197 claim_zero(ibv_exp_release_intf(txq->priv->ctx,
201 if (txq->if_cq != NULL) {
202 assert(txq->priv != NULL);
203 assert(txq->priv->ctx != NULL);
204 assert(txq->cq != NULL);
205 params = (struct ibv_exp_release_intf_params){
208 claim_zero(ibv_exp_release_intf(txq->priv->ctx,
213 claim_zero(ibv_destroy_qp(txq->qp));
215 claim_zero(ibv_destroy_cq(txq->cq));
216 if (txq->rd != NULL) {
217 struct ibv_exp_destroy_res_domain_attr attr = {
221 assert(txq->priv != NULL);
222 assert(txq->priv->ctx != NULL);
223 claim_zero(ibv_exp_destroy_res_domain(txq->priv->ctx,
227 for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
228 if (txq->mp2mr[i].mp == NULL)
230 assert(txq->mp2mr[i].mr != NULL);
231 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
233 memset(txq, 0, sizeof(*txq));
237 * Configure a TX queue.
240 * Pointer to Ethernet device structure.
242 * Pointer to TX queue structure.
244 * Number of descriptors to configure in queue.
246 * NUMA socket on which memory must be allocated.
248 * Thresholds parameters.
251 * 0 on success, errno value on failure.
254 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
255 unsigned int socket, const struct rte_eth_txconf *conf)
257 struct priv *priv = dev->data->dev_private;
263 struct ibv_exp_query_intf_params params;
264 struct ibv_exp_qp_init_attr init;
265 struct ibv_exp_res_domain_init_attr rd;
266 struct ibv_exp_cq_init_attr cq;
267 struct ibv_exp_qp_attr mod;
269 enum ibv_exp_query_intf_status status;
272 (void)conf; /* Thresholds configuration (ignored). */
273 if ((desc == 0) || (desc % MLX5_PMD_SGE_WR_N)) {
274 ERROR("%p: invalid number of TX descriptors (must be a"
275 " multiple of %d)", (void *)dev, MLX5_PMD_SGE_WR_N);
278 desc /= MLX5_PMD_SGE_WR_N;
279 /* MRs will be registered in mp2mr[] later. */
280 attr.rd = (struct ibv_exp_res_domain_init_attr){
281 .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
282 IBV_EXP_RES_DOMAIN_MSG_MODEL),
283 .thread_model = IBV_EXP_THREAD_SINGLE,
284 .msg_model = IBV_EXP_MSG_HIGH_BW,
286 tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
287 if (tmpl.rd == NULL) {
289 ERROR("%p: RD creation failure: %s",
290 (void *)dev, strerror(ret));
293 attr.cq = (struct ibv_exp_cq_init_attr){
294 .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
295 .res_domain = tmpl.rd,
297 tmpl.cq = ibv_exp_create_cq(priv->ctx, desc, NULL, NULL, 0, &attr.cq);
298 if (tmpl.cq == NULL) {
300 ERROR("%p: CQ creation failure: %s",
301 (void *)dev, strerror(ret));
304 DEBUG("priv->device_attr.max_qp_wr is %d",
305 priv->device_attr.max_qp_wr);
306 DEBUG("priv->device_attr.max_sge is %d",
307 priv->device_attr.max_sge);
308 attr.init = (struct ibv_exp_qp_init_attr){
309 /* CQ to be associated with the send queue. */
311 /* CQ to be associated with the receive queue. */
314 /* Max number of outstanding WRs. */
315 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
316 priv->device_attr.max_qp_wr :
318 /* Max number of scatter/gather elements in a WR. */
319 .max_send_sge = ((priv->device_attr.max_sge <
321 priv->device_attr.max_sge :
323 #if MLX5_PMD_MAX_INLINE > 0
324 .max_inline_data = MLX5_PMD_MAX_INLINE,
327 .qp_type = IBV_QPT_RAW_PACKET,
328 /* Do *NOT* enable this, completions events are managed per
332 .res_domain = tmpl.rd,
333 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
334 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
336 tmpl.qp = ibv_exp_create_qp(priv->ctx, &attr.init);
337 if (tmpl.qp == NULL) {
338 ret = (errno ? errno : EINVAL);
339 ERROR("%p: QP creation failure: %s",
340 (void *)dev, strerror(ret));
343 #if MLX5_PMD_MAX_INLINE > 0
344 /* ibv_create_qp() updates this value. */
345 tmpl.max_inline = attr.init.cap.max_inline_data;
347 attr.mod = (struct ibv_exp_qp_attr){
348 /* Move the QP to this state. */
349 .qp_state = IBV_QPS_INIT,
350 /* Primary port number. */
351 .port_num = priv->port
353 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
354 (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
356 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
357 (void *)dev, strerror(ret));
360 ret = txq_alloc_elts(&tmpl, desc);
362 ERROR("%p: TXQ allocation failed: %s",
363 (void *)dev, strerror(ret));
366 attr.mod = (struct ibv_exp_qp_attr){
367 .qp_state = IBV_QPS_RTR
369 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
371 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
372 (void *)dev, strerror(ret));
375 attr.mod.qp_state = IBV_QPS_RTS;
376 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
378 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
379 (void *)dev, strerror(ret));
382 attr.params = (struct ibv_exp_query_intf_params){
383 .intf_scope = IBV_EXP_INTF_GLOBAL,
384 .intf = IBV_EXP_INTF_CQ,
387 tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
388 if (tmpl.if_cq == NULL) {
390 ERROR("%p: CQ interface family query failed with status %d",
391 (void *)dev, status);
394 attr.params = (struct ibv_exp_query_intf_params){
395 .intf_scope = IBV_EXP_INTF_GLOBAL,
396 .intf = IBV_EXP_INTF_QP_BURST,
398 #ifdef HAVE_EXP_QP_BURST_CREATE_ENABLE_MULTI_PACKET_SEND_WR
399 /* Multi packet send WR can only be used outside of VF. */
402 IBV_EXP_QP_BURST_CREATE_ENABLE_MULTI_PACKET_SEND_WR :
406 tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
407 if (tmpl.if_qp == NULL) {
409 ERROR("%p: QP interface family query failed with status %d",
410 (void *)dev, status);
413 /* Clean up txq in case we're reinitializing it. */
414 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
417 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
418 /* Pre-register known mempools. */
419 rte_mempool_walk(txq_mp2mr_iter, txq);
429 * DPDK callback to configure a TX queue.
432 * Pointer to Ethernet device structure.
436 * Number of descriptors to configure in queue.
438 * NUMA socket on which memory must be allocated.
440 * Thresholds parameters.
443 * 0 on success, negative errno value on failure.
446 mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
447 unsigned int socket, const struct rte_eth_txconf *conf)
449 struct priv *priv = dev->data->dev_private;
450 struct txq *txq = (*priv->txqs)[idx];
454 DEBUG("%p: configuring queue %u for %u descriptors",
455 (void *)dev, idx, desc);
456 if (idx >= priv->txqs_n) {
457 ERROR("%p: queue index out of range (%u >= %u)",
458 (void *)dev, idx, priv->txqs_n);
463 DEBUG("%p: reusing already allocated queue index %u (%p)",
464 (void *)dev, idx, (void *)txq);
469 (*priv->txqs)[idx] = NULL;
472 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
474 ERROR("%p: unable to allocate queue index %u",
480 ret = txq_setup(dev, txq, desc, socket, conf);
484 txq->stats.idx = idx;
485 DEBUG("%p: adding TX queue %p to list",
486 (void *)dev, (void *)txq);
487 (*priv->txqs)[idx] = txq;
488 /* Update send callback. */
489 dev->tx_pkt_burst = mlx5_tx_burst;
496 * DPDK callback to release a TX queue.
499 * Generic TX queue pointer.
502 mlx5_tx_queue_release(void *dpdk_txq)
504 struct txq *txq = (struct txq *)dpdk_txq;
512 for (i = 0; (i != priv->txqs_n); ++i)
513 if ((*priv->txqs)[i] == txq) {
514 DEBUG("%p: removing TX queue %p from list",
515 (void *)priv->dev, (void *)txq);
516 (*priv->txqs)[i] = NULL;