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)
147 unsigned int elts_n = txq->elts_n;
148 unsigned int elts_head = txq->elts_head;
149 unsigned int elts_tail = txq->elts_tail;
150 struct txq_elt (*elts)[elts_n] = txq->elts;
151 linear_t (*elts_linear)[elts_n] = txq->elts_linear;
152 struct ibv_mr *mr_linear = txq->mr_linear;
154 DEBUG("%p: freeing WRs", (void *)txq);
159 txq->elts_comp_cd = 0;
160 txq->elts_comp_cd_init = 0;
162 txq->elts_linear = NULL;
163 txq->mr_linear = NULL;
164 if (mr_linear != NULL)
165 claim_zero(ibv_dereg_mr(mr_linear));
167 rte_free(elts_linear);
170 while (elts_tail != elts_head) {
171 struct txq_elt *elt = &(*elts)[elts_tail];
173 assert(elt->buf != NULL);
174 rte_pktmbuf_free(elt->buf);
177 memset(elt, 0x77, sizeof(*elt));
179 if (++elts_tail == elts_n)
186 * Clean up a TX queue.
188 * Destroy objects, free allocated memory and reset the structure for reuse.
191 * Pointer to TX queue structure.
194 txq_cleanup(struct txq *txq)
196 struct ibv_exp_release_intf_params params;
199 DEBUG("cleaning up %p", (void *)txq);
201 txq->poll_cnt = NULL;
202 #if MLX5_PMD_MAX_INLINE > 0
203 txq->send_pending_inline = NULL;
205 txq->send_flush = NULL;
206 if (txq->if_qp != NULL) {
207 assert(txq->priv != NULL);
208 assert(txq->priv->ctx != NULL);
209 assert(txq->qp != NULL);
210 params = (struct ibv_exp_release_intf_params){
213 claim_zero(ibv_exp_release_intf(txq->priv->ctx,
217 if (txq->if_cq != NULL) {
218 assert(txq->priv != NULL);
219 assert(txq->priv->ctx != NULL);
220 assert(txq->cq != NULL);
221 params = (struct ibv_exp_release_intf_params){
224 claim_zero(ibv_exp_release_intf(txq->priv->ctx,
229 claim_zero(ibv_destroy_qp(txq->qp));
231 claim_zero(ibv_destroy_cq(txq->cq));
232 if (txq->rd != NULL) {
233 struct ibv_exp_destroy_res_domain_attr attr = {
237 assert(txq->priv != NULL);
238 assert(txq->priv->ctx != NULL);
239 claim_zero(ibv_exp_destroy_res_domain(txq->priv->ctx,
243 for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
244 if (txq->mp2mr[i].mp == NULL)
246 assert(txq->mp2mr[i].mr != NULL);
247 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
249 memset(txq, 0, sizeof(*txq));
253 * Configure a TX queue.
256 * Pointer to Ethernet device structure.
258 * Pointer to TX queue structure.
260 * Number of descriptors to configure in queue.
262 * NUMA socket on which memory must be allocated.
264 * Thresholds parameters.
267 * 0 on success, errno value on failure.
270 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
271 unsigned int socket, const struct rte_eth_txconf *conf)
273 struct priv *priv = mlx5_get_priv(dev);
279 struct ibv_exp_query_intf_params params;
280 struct ibv_exp_qp_init_attr init;
281 struct ibv_exp_res_domain_init_attr rd;
282 struct ibv_exp_cq_init_attr cq;
283 struct ibv_exp_qp_attr mod;
285 enum ibv_exp_query_intf_status status;
288 (void)conf; /* Thresholds configuration (ignored). */
289 if ((desc == 0) || (desc % MLX5_PMD_SGE_WR_N)) {
290 ERROR("%p: invalid number of TX descriptors (must be a"
291 " multiple of %d)", (void *)dev, MLX5_PMD_SGE_WR_N);
294 desc /= MLX5_PMD_SGE_WR_N;
295 /* MRs will be registered in mp2mr[] later. */
296 attr.rd = (struct ibv_exp_res_domain_init_attr){
297 .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
298 IBV_EXP_RES_DOMAIN_MSG_MODEL),
299 .thread_model = IBV_EXP_THREAD_SINGLE,
300 .msg_model = IBV_EXP_MSG_HIGH_BW,
302 tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
303 if (tmpl.rd == NULL) {
305 ERROR("%p: RD creation failure: %s",
306 (void *)dev, strerror(ret));
309 attr.cq = (struct ibv_exp_cq_init_attr){
310 .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
311 .res_domain = tmpl.rd,
313 tmpl.cq = ibv_exp_create_cq(priv->ctx, desc, NULL, NULL, 0, &attr.cq);
314 if (tmpl.cq == NULL) {
316 ERROR("%p: CQ creation failure: %s",
317 (void *)dev, strerror(ret));
320 DEBUG("priv->device_attr.max_qp_wr is %d",
321 priv->device_attr.max_qp_wr);
322 DEBUG("priv->device_attr.max_sge is %d",
323 priv->device_attr.max_sge);
324 attr.init = (struct ibv_exp_qp_init_attr){
325 /* CQ to be associated with the send queue. */
327 /* CQ to be associated with the receive queue. */
330 /* Max number of outstanding WRs. */
331 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
332 priv->device_attr.max_qp_wr :
334 /* Max number of scatter/gather elements in a WR. */
335 .max_send_sge = ((priv->device_attr.max_sge <
337 priv->device_attr.max_sge :
339 #if MLX5_PMD_MAX_INLINE > 0
340 .max_inline_data = MLX5_PMD_MAX_INLINE,
343 .qp_type = IBV_QPT_RAW_PACKET,
344 /* Do *NOT* enable this, completions events are managed per
348 .res_domain = tmpl.rd,
349 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
350 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
352 tmpl.qp = ibv_exp_create_qp(priv->ctx, &attr.init);
353 if (tmpl.qp == NULL) {
354 ret = (errno ? errno : EINVAL);
355 ERROR("%p: QP creation failure: %s",
356 (void *)dev, strerror(ret));
359 #if MLX5_PMD_MAX_INLINE > 0
360 /* ibv_create_qp() updates this value. */
361 tmpl.max_inline = attr.init.cap.max_inline_data;
363 attr.mod = (struct ibv_exp_qp_attr){
364 /* Move the QP to this state. */
365 .qp_state = IBV_QPS_INIT,
366 /* Primary port number. */
367 .port_num = priv->port
369 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
370 (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
372 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
373 (void *)dev, strerror(ret));
376 ret = txq_alloc_elts(&tmpl, desc);
378 ERROR("%p: TXQ allocation failed: %s",
379 (void *)dev, strerror(ret));
382 attr.mod = (struct ibv_exp_qp_attr){
383 .qp_state = IBV_QPS_RTR
385 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
387 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
388 (void *)dev, strerror(ret));
391 attr.mod.qp_state = IBV_QPS_RTS;
392 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
394 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
395 (void *)dev, strerror(ret));
398 attr.params = (struct ibv_exp_query_intf_params){
399 .intf_scope = IBV_EXP_INTF_GLOBAL,
400 .intf = IBV_EXP_INTF_CQ,
403 tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
404 if (tmpl.if_cq == NULL) {
406 ERROR("%p: CQ interface family query failed with status %d",
407 (void *)dev, status);
410 attr.params = (struct ibv_exp_query_intf_params){
411 .intf_scope = IBV_EXP_INTF_GLOBAL,
412 .intf = IBV_EXP_INTF_QP_BURST,
414 #ifdef HAVE_VERBS_VLAN_INSERTION
417 #ifdef HAVE_EXP_QP_BURST_CREATE_ENABLE_MULTI_PACKET_SEND_WR
418 /* Enable multi-packet send if supported. */
421 IBV_EXP_QP_BURST_CREATE_ENABLE_MULTI_PACKET_SEND_WR :
425 tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
426 if (tmpl.if_qp == NULL) {
428 ERROR("%p: QP interface family query failed with status %d",
429 (void *)dev, status);
432 /* Clean up txq in case we're reinitializing it. */
433 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
436 txq->poll_cnt = txq->if_cq->poll_cnt;
437 #if MLX5_PMD_MAX_INLINE > 0
438 txq->send_pending_inline = txq->if_qp->send_pending_inline;
439 #ifdef HAVE_VERBS_VLAN_INSERTION
440 txq->send_pending_inline_vlan = txq->if_qp->send_pending_inline_vlan;
443 #if MLX5_PMD_SGE_WR_N > 1
444 txq->send_pending_sg_list = txq->if_qp->send_pending_sg_list;
445 #ifdef HAVE_VERBS_VLAN_INSERTION
446 txq->send_pending_sg_list_vlan = txq->if_qp->send_pending_sg_list_vlan;
449 txq->send_pending = txq->if_qp->send_pending;
450 #ifdef HAVE_VERBS_VLAN_INSERTION
451 txq->send_pending_vlan = txq->if_qp->send_pending_vlan;
453 txq->send_flush = txq->if_qp->send_flush;
454 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
455 /* Pre-register known mempools. */
456 rte_mempool_walk(txq_mp2mr_iter, txq);
466 * DPDK callback to configure a TX queue.
469 * Pointer to Ethernet device structure.
473 * Number of descriptors to configure in queue.
475 * NUMA socket on which memory must be allocated.
477 * Thresholds parameters.
480 * 0 on success, negative errno value on failure.
483 mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
484 unsigned int socket, const struct rte_eth_txconf *conf)
486 struct priv *priv = dev->data->dev_private;
487 struct txq *txq = (*priv->txqs)[idx];
490 if (mlx5_is_secondary())
491 return -E_RTE_SECONDARY;
494 DEBUG("%p: configuring queue %u for %u descriptors",
495 (void *)dev, idx, desc);
496 if (idx >= priv->txqs_n) {
497 ERROR("%p: queue index out of range (%u >= %u)",
498 (void *)dev, idx, priv->txqs_n);
503 DEBUG("%p: reusing already allocated queue index %u (%p)",
504 (void *)dev, idx, (void *)txq);
509 (*priv->txqs)[idx] = NULL;
512 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
514 ERROR("%p: unable to allocate queue index %u",
520 ret = txq_setup(dev, txq, desc, socket, conf);
524 txq->stats.idx = idx;
525 DEBUG("%p: adding TX queue %p to list",
526 (void *)dev, (void *)txq);
527 (*priv->txqs)[idx] = txq;
528 /* Update send callback. */
529 dev->tx_pkt_burst = mlx5_tx_burst;
536 * DPDK callback to release a TX queue.
539 * Generic TX queue pointer.
542 mlx5_tx_queue_release(void *dpdk_txq)
544 struct txq *txq = (struct txq *)dpdk_txq;
548 if (mlx5_is_secondary())
555 for (i = 0; (i != priv->txqs_n); ++i)
556 if ((*priv->txqs)[i] == txq) {
557 DEBUG("%p: removing TX queue %p from list",
558 (void *)priv->dev, (void *)txq);
559 (*priv->txqs)[i] = NULL;
568 * DPDK callback for TX in secondary processes.
570 * This function configures all queues from primary process information
571 * if necessary before reverting to the normal TX burst callback.
574 * Generic pointer to TX queue structure.
576 * Packets to transmit.
578 * Number of packets in array.
581 * Number of packets successfully transmitted (<= pkts_n).
584 mlx5_tx_burst_secondary_setup(void *dpdk_txq, struct rte_mbuf **pkts,
587 struct txq *txq = dpdk_txq;
588 struct priv *priv = mlx5_secondary_data_setup(txq->priv);
589 struct priv *primary_priv;
595 mlx5_secondary_data[priv->dev->data->port_id].primary_priv;
596 /* Look for queue index in both private structures. */
597 for (index = 0; index != priv->txqs_n; ++index)
598 if (((*primary_priv->txqs)[index] == txq) ||
599 ((*priv->txqs)[index] == txq))
601 if (index == priv->txqs_n)
603 txq = (*priv->txqs)[index];
604 return priv->dev->tx_pkt_burst(txq, pkts, pkts_n);