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 "-Wpedantic"
45 #include <infiniband/verbs.h>
47 #pragma GCC diagnostic error "-Wpedantic"
50 /* DPDK headers don't like -pedantic. */
52 #pragma GCC diagnostic ignored "-Wpedantic"
55 #include <rte_malloc.h>
56 #include <rte_ethdev.h>
57 #include <rte_common.h>
59 #pragma GCC diagnostic error "-Wpedantic"
62 #include "mlx5_utils.h"
63 #include "mlx5_defs.h"
65 #include "mlx5_rxtx.h"
66 #include "mlx5_autoconf.h"
67 #include "mlx5_defs.h"
70 * Allocate TX queue elements.
73 * Pointer to TX queue structure.
75 * Number of elements to allocate.
78 txq_alloc_elts(struct txq_ctrl *txq_ctrl, unsigned int elts_n)
82 for (i = 0; (i != elts_n); ++i)
83 (*txq_ctrl->txq.elts)[i] = NULL;
84 for (i = 0; (i != (1u << txq_ctrl->txq.wqe_n)); ++i) {
85 volatile struct mlx5_wqe64 *wqe = &(*txq_ctrl->txq.wqes)[i];
87 memset((void *)(uintptr_t)wqe, 0x0, sizeof(*wqe));
89 DEBUG("%p: allocated and configured %u WRs", (void *)txq_ctrl, elts_n);
90 txq_ctrl->txq.elts_head = 0;
91 txq_ctrl->txq.elts_tail = 0;
92 txq_ctrl->txq.elts_comp = 0;
96 * Free TX queue elements.
99 * Pointer to TX queue structure.
102 txq_free_elts(struct txq_ctrl *txq_ctrl)
104 unsigned int elts_n = 1 << txq_ctrl->txq.elts_n;
105 unsigned int elts_head = txq_ctrl->txq.elts_head;
106 unsigned int elts_tail = txq_ctrl->txq.elts_tail;
107 struct rte_mbuf *(*elts)[elts_n] = txq_ctrl->txq.elts;
109 DEBUG("%p: freeing WRs", (void *)txq_ctrl);
110 txq_ctrl->txq.elts_head = 0;
111 txq_ctrl->txq.elts_tail = 0;
112 txq_ctrl->txq.elts_comp = 0;
114 while (elts_tail != elts_head) {
115 struct rte_mbuf *elt = (*elts)[elts_tail];
118 rte_pktmbuf_free(elt);
121 memset(&(*elts)[elts_tail],
123 sizeof((*elts)[elts_tail]));
125 if (++elts_tail == elts_n)
131 * Clean up a TX queue.
133 * Destroy objects, free allocated memory and reset the structure for reuse.
136 * Pointer to TX queue structure.
139 txq_cleanup(struct txq_ctrl *txq_ctrl)
141 struct ibv_exp_release_intf_params params;
144 DEBUG("cleaning up %p", (void *)txq_ctrl);
145 txq_free_elts(txq_ctrl);
146 if (txq_ctrl->if_qp != NULL) {
147 assert(txq_ctrl->priv != NULL);
148 assert(txq_ctrl->priv->ctx != NULL);
149 assert(txq_ctrl->qp != NULL);
150 params = (struct ibv_exp_release_intf_params){
153 claim_zero(ibv_exp_release_intf(txq_ctrl->priv->ctx,
157 if (txq_ctrl->if_cq != NULL) {
158 assert(txq_ctrl->priv != NULL);
159 assert(txq_ctrl->priv->ctx != NULL);
160 assert(txq_ctrl->cq != NULL);
161 params = (struct ibv_exp_release_intf_params){
164 claim_zero(ibv_exp_release_intf(txq_ctrl->priv->ctx,
168 if (txq_ctrl->qp != NULL)
169 claim_zero(ibv_destroy_qp(txq_ctrl->qp));
170 if (txq_ctrl->cq != NULL)
171 claim_zero(ibv_destroy_cq(txq_ctrl->cq));
172 if (txq_ctrl->rd != NULL) {
173 struct ibv_exp_destroy_res_domain_attr attr = {
177 assert(txq_ctrl->priv != NULL);
178 assert(txq_ctrl->priv->ctx != NULL);
179 claim_zero(ibv_exp_destroy_res_domain(txq_ctrl->priv->ctx,
183 for (i = 0; (i != RTE_DIM(txq_ctrl->txq.mp2mr)); ++i) {
184 if (txq_ctrl->txq.mp2mr[i].mp == NULL)
186 assert(txq_ctrl->txq.mp2mr[i].mr != NULL);
187 claim_zero(ibv_dereg_mr(txq_ctrl->txq.mp2mr[i].mr));
189 memset(txq_ctrl, 0, sizeof(*txq_ctrl));
193 * Initialize TX queue.
196 * Pointer to TX queue control template.
198 * Pointer to TX queue control.
201 * 0 on success, errno value on failure.
204 txq_setup(struct txq_ctrl *tmpl, struct txq_ctrl *txq_ctrl)
206 struct mlx5_qp *qp = to_mqp(tmpl->qp);
207 struct ibv_cq *ibcq = tmpl->cq;
208 struct mlx5_cq *cq = to_mxxx(cq, cq);
210 if (cq->cqe_sz != RTE_CACHE_LINE_SIZE) {
211 ERROR("Wrong MLX5_CQE_SIZE environment variable value: "
212 "it should be set to %u", RTE_CACHE_LINE_SIZE);
215 tmpl->txq.cqe_n = log2above(ibcq->cqe);
216 tmpl->txq.qp_num_8s = qp->ctrl_seg.qp_num << 8;
218 (volatile struct mlx5_wqe64 (*)[])
219 (uintptr_t)qp->gen_data.sqstart;
220 tmpl->txq.wqe_n = log2above(qp->sq.wqe_cnt);
221 tmpl->txq.qp_db = &qp->gen_data.db[MLX5_SND_DBR];
222 tmpl->txq.bf_reg = qp->gen_data.bf->reg;
223 tmpl->txq.bf_offset = qp->gen_data.bf->offset;
224 tmpl->txq.bf_buf_size = log2above(qp->gen_data.bf->buf_size);
225 tmpl->txq.cq_db = cq->dbrec;
227 (volatile struct mlx5_cqe (*)[])
228 (uintptr_t)cq->active_buf->buf;
230 (struct rte_mbuf *(*)[1 << tmpl->txq.elts_n])
231 ((uintptr_t)txq_ctrl + sizeof(*txq_ctrl));
236 * Configure a TX queue.
239 * Pointer to Ethernet device structure.
241 * Pointer to TX queue structure.
243 * Number of descriptors to configure in queue.
245 * NUMA socket on which memory must be allocated.
247 * Thresholds parameters.
250 * 0 on success, errno value on failure.
253 txq_ctrl_setup(struct rte_eth_dev *dev, struct txq_ctrl *txq_ctrl,
254 uint16_t desc, unsigned int socket,
255 const struct rte_eth_txconf *conf)
257 struct priv *priv = mlx5_get_priv(dev);
258 struct txq_ctrl tmpl = {
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;
268 struct ibv_exp_cq_attr cq_attr;
270 enum ibv_exp_query_intf_status status;
273 if (mlx5_getenv_int("MLX5_ENABLE_CQE_COMPRESSION")) {
275 ERROR("MLX5_ENABLE_CQE_COMPRESSION must never be set");
278 (void)conf; /* Thresholds configuration (ignored). */
279 assert(desc > MLX5_TX_COMP_THRESH);
280 tmpl.txq.elts_n = log2above(desc);
281 /* MRs will be registered in mp2mr[] later. */
282 attr.rd = (struct ibv_exp_res_domain_init_attr){
283 .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
284 IBV_EXP_RES_DOMAIN_MSG_MODEL),
285 .thread_model = IBV_EXP_THREAD_SINGLE,
286 .msg_model = IBV_EXP_MSG_HIGH_BW,
288 tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
289 if (tmpl.rd == NULL) {
291 ERROR("%p: RD creation failure: %s",
292 (void *)dev, strerror(ret));
295 attr.cq = (struct ibv_exp_cq_init_attr){
296 .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
297 .res_domain = tmpl.rd,
299 tmpl.cq = ibv_exp_create_cq(priv->ctx,
300 (((desc / MLX5_TX_COMP_THRESH) - 1) ?
301 ((desc / MLX5_TX_COMP_THRESH) - 1) : 1),
302 NULL, NULL, 0, &attr.cq);
303 if (tmpl.cq == NULL) {
305 ERROR("%p: CQ creation failure: %s",
306 (void *)dev, strerror(ret));
309 DEBUG("priv->device_attr.max_qp_wr is %d",
310 priv->device_attr.max_qp_wr);
311 DEBUG("priv->device_attr.max_sge is %d",
312 priv->device_attr.max_sge);
313 attr.init = (struct ibv_exp_qp_init_attr){
314 /* CQ to be associated with the send queue. */
316 /* CQ to be associated with the receive queue. */
319 /* Max number of outstanding WRs. */
320 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
321 priv->device_attr.max_qp_wr :
324 * Max number of scatter/gather elements in a WR,
325 * must be 1 to prevent libmlx5 from trying to affect
326 * too much memory. TX gather is not impacted by the
327 * priv->device_attr.max_sge limit and will still work
332 .qp_type = IBV_QPT_RAW_PACKET,
333 /* Do *NOT* enable this, completions events are managed per
337 .res_domain = tmpl.rd,
338 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
339 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
341 if (priv->txq_inline && (priv->txqs_n >= priv->txqs_inline)) {
342 tmpl.txq.max_inline =
343 ((priv->txq_inline + (RTE_CACHE_LINE_SIZE - 1)) /
344 RTE_CACHE_LINE_SIZE);
345 attr.init.cap.max_inline_data =
346 tmpl.txq.max_inline * RTE_CACHE_LINE_SIZE;
348 tmpl.qp = ibv_exp_create_qp(priv->ctx, &attr.init);
349 if (tmpl.qp == NULL) {
350 ret = (errno ? errno : EINVAL);
351 ERROR("%p: QP creation failure: %s",
352 (void *)dev, strerror(ret));
355 DEBUG("TX queue capabilities: max_send_wr=%u, max_send_sge=%u,"
356 " max_inline_data=%u",
357 attr.init.cap.max_send_wr,
358 attr.init.cap.max_send_sge,
359 attr.init.cap.max_inline_data);
360 attr.mod = (struct ibv_exp_qp_attr){
361 /* Move the QP to this state. */
362 .qp_state = IBV_QPS_INIT,
363 /* Primary port number. */
364 .port_num = priv->port
366 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
367 (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
369 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
370 (void *)dev, strerror(ret));
373 ret = txq_setup(&tmpl, txq_ctrl);
375 ERROR("%p: cannot initialize TX queue structure: %s",
376 (void *)dev, strerror(ret));
379 txq_alloc_elts(&tmpl, desc);
380 attr.mod = (struct ibv_exp_qp_attr){
381 .qp_state = IBV_QPS_RTR
383 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
385 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
386 (void *)dev, strerror(ret));
389 attr.mod.qp_state = IBV_QPS_RTS;
390 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
392 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
393 (void *)dev, strerror(ret));
396 attr.params = (struct ibv_exp_query_intf_params){
397 .intf_scope = IBV_EXP_INTF_GLOBAL,
398 .intf = IBV_EXP_INTF_CQ,
401 tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
402 if (tmpl.if_cq == NULL) {
404 ERROR("%p: CQ interface family query failed with status %d",
405 (void *)dev, status);
408 attr.params = (struct ibv_exp_query_intf_params){
409 .intf_scope = IBV_EXP_INTF_GLOBAL,
410 .intf = IBV_EXP_INTF_QP_BURST,
413 /* Enable multi-packet send if supported. */
415 ((priv->mps && !priv->sriov) ?
416 IBV_EXP_QP_BURST_CREATE_ENABLE_MULTI_PACKET_SEND_WR :
419 tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
420 if (tmpl.if_qp == NULL) {
422 ERROR("%p: QP interface family query failed with status %d",
423 (void *)dev, status);
426 /* Clean up txq in case we're reinitializing it. */
427 DEBUG("%p: cleaning-up old txq just in case", (void *)txq_ctrl);
428 txq_cleanup(txq_ctrl);
430 DEBUG("%p: txq updated with %p", (void *)txq_ctrl, (void *)&tmpl);
431 /* Pre-register known mempools. */
432 rte_mempool_walk(txq_mp2mr_iter, txq_ctrl);
442 * DPDK callback to configure a TX queue.
445 * Pointer to Ethernet device structure.
449 * Number of descriptors to configure in queue.
451 * NUMA socket on which memory must be allocated.
453 * Thresholds parameters.
456 * 0 on success, negative errno value on failure.
459 mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
460 unsigned int socket, const struct rte_eth_txconf *conf)
462 struct priv *priv = dev->data->dev_private;
463 struct txq *txq = (*priv->txqs)[idx];
464 struct txq_ctrl *txq_ctrl = container_of(txq, struct txq_ctrl, txq);
467 if (mlx5_is_secondary())
468 return -E_RTE_SECONDARY;
471 if (desc <= MLX5_TX_COMP_THRESH) {
472 WARN("%p: number of descriptors requested for TX queue %u"
473 " must be higher than MLX5_TX_COMP_THRESH, using"
475 (void *)dev, idx, MLX5_TX_COMP_THRESH + 1, desc);
476 desc = MLX5_TX_COMP_THRESH + 1;
478 if (!rte_is_power_of_2(desc)) {
479 desc = 1 << log2above(desc);
480 WARN("%p: increased number of descriptors in TX queue %u"
481 " to the next power of two (%d)",
482 (void *)dev, idx, desc);
484 DEBUG("%p: configuring queue %u for %u descriptors",
485 (void *)dev, idx, desc);
486 if (idx >= priv->txqs_n) {
487 ERROR("%p: queue index out of range (%u >= %u)",
488 (void *)dev, idx, priv->txqs_n);
493 DEBUG("%p: reusing already allocated queue index %u (%p)",
494 (void *)dev, idx, (void *)txq);
499 (*priv->txqs)[idx] = NULL;
500 txq_cleanup(txq_ctrl);
503 rte_calloc_socket("TXQ", 1,
505 desc * sizeof(struct rte_mbuf *),
507 if (txq_ctrl == NULL) {
508 ERROR("%p: unable to allocate queue index %u",
514 ret = txq_ctrl_setup(dev, txq_ctrl, desc, socket, conf);
518 txq_ctrl->txq.stats.idx = idx;
519 DEBUG("%p: adding TX queue %p to list",
520 (void *)dev, (void *)txq_ctrl);
521 (*priv->txqs)[idx] = &txq_ctrl->txq;
522 /* Update send callback. */
523 priv_select_tx_function(priv);
530 * DPDK callback to release a TX queue.
533 * Generic TX queue pointer.
536 mlx5_tx_queue_release(void *dpdk_txq)
538 struct txq *txq = (struct txq *)dpdk_txq;
539 struct txq_ctrl *txq_ctrl;
543 if (mlx5_is_secondary())
548 txq_ctrl = container_of(txq, struct txq_ctrl, txq);
549 priv = txq_ctrl->priv;
551 for (i = 0; (i != priv->txqs_n); ++i)
552 if ((*priv->txqs)[i] == txq) {
553 DEBUG("%p: removing TX queue %p from list",
554 (void *)priv->dev, (void *)txq_ctrl);
555 (*priv->txqs)[i] = NULL;
558 txq_cleanup(txq_ctrl);
564 * DPDK callback for TX in secondary processes.
566 * This function configures all queues from primary process information
567 * if necessary before reverting to the normal TX burst callback.
570 * Generic pointer to TX queue structure.
572 * Packets to transmit.
574 * Number of packets in array.
577 * Number of packets successfully transmitted (<= pkts_n).
580 mlx5_tx_burst_secondary_setup(void *dpdk_txq, struct rte_mbuf **pkts,
583 struct txq *txq = dpdk_txq;
584 struct txq_ctrl *txq_ctrl = container_of(txq, struct txq_ctrl, txq);
585 struct priv *priv = mlx5_secondary_data_setup(txq_ctrl->priv);
586 struct priv *primary_priv;
592 mlx5_secondary_data[priv->dev->data->port_id].primary_priv;
593 /* Look for queue index in both private structures. */
594 for (index = 0; index != priv->txqs_n; ++index)
595 if (((*primary_priv->txqs)[index] == txq) ||
596 ((*priv->txqs)[index] == txq))
598 if (index == priv->txqs_n)
600 txq = (*priv->txqs)[index];
601 return priv->dev->tx_pkt_burst(txq, pkts, pkts_n);