1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2019 Mellanox Technologies, Ltd
8 #include <sys/eventfd.h>
10 #include <rte_malloc.h>
11 #include <rte_memory.h>
12 #include <rte_errno.h>
13 #include <rte_lcore.h>
14 #include <rte_atomic.h>
15 #include <rte_common.h>
17 #include <rte_alarm.h>
19 #include <mlx5_common.h>
20 #include <mlx5_common_os.h>
21 #include <mlx5_common_devx.h>
22 #include <mlx5_glue.h>
24 #include "mlx5_vdpa_utils.h"
25 #include "mlx5_vdpa.h"
28 #define MLX5_VDPA_ERROR_TIME_SEC 3u
31 mlx5_vdpa_event_qp_global_release(struct mlx5_vdpa_priv *priv)
33 mlx5_devx_uar_release(&priv->uar);
34 #ifdef HAVE_IBV_DEVX_EVENT
36 mlx5_os_devx_destroy_event_channel(priv->eventc);
42 /* Prepare all the global resources for all the event objects.*/
44 mlx5_vdpa_event_qp_global_prepare(struct mlx5_vdpa_priv *priv)
46 priv->eventc = mlx5_os_devx_create_event_channel(priv->cdev->ctx,
47 MLX5DV_DEVX_CREATE_EVENT_CHANNEL_FLAGS_OMIT_EV_DATA);
50 DRV_LOG(ERR, "Failed to create event channel %d.",
54 if (mlx5_devx_uar_prepare(priv->cdev, &priv->uar) != 0) {
55 DRV_LOG(ERR, "Failed to allocate UAR.");
60 mlx5_vdpa_event_qp_global_release(priv);
65 mlx5_vdpa_cq_destroy(struct mlx5_vdpa_cq *cq)
67 mlx5_devx_cq_destroy(&cq->cq_obj);
68 memset(cq, 0, sizeof(*cq));
71 static inline void __rte_unused
72 mlx5_vdpa_cq_arm(struct mlx5_vdpa_priv *priv, struct mlx5_vdpa_cq *cq)
74 uint32_t arm_sn = cq->arm_sn << MLX5_CQ_SQN_OFFSET;
75 uint32_t cq_ci = cq->cq_ci & MLX5_CI_MASK;
76 uint32_t doorbell_hi = arm_sn | MLX5_CQ_DBR_CMD_ALL | cq_ci;
77 uint64_t doorbell = ((uint64_t)doorbell_hi << 32) | cq->cq_obj.cq->id;
78 uint64_t db_be = rte_cpu_to_be_64(doorbell);
80 mlx5_doorbell_ring(&priv->uar.cq_db, db_be, doorbell_hi,
81 &cq->cq_obj.db_rec[MLX5_CQ_ARM_DB], 0);
87 mlx5_vdpa_cq_create(struct mlx5_vdpa_priv *priv, uint16_t log_desc_n,
88 int callfd, struct mlx5_vdpa_cq *cq)
90 struct mlx5_devx_cq_attr attr = {
92 .uar_page_id = mlx5_os_get_devx_uar_page_id(priv->uar.obj),
94 uint16_t event_nums[1] = {0};
97 ret = mlx5_devx_cq_create(priv->cdev->ctx, &cq->cq_obj, log_desc_n,
98 &attr, SOCKET_ID_ANY);
102 cq->log_desc_n = log_desc_n;
103 rte_spinlock_init(&cq->sl);
104 /* Subscribe CQ event to the event channel controlled by the driver. */
105 ret = mlx5_os_devx_subscribe_devx_event(priv->eventc,
107 sizeof(event_nums), event_nums,
108 (uint64_t)(uintptr_t)cq);
110 DRV_LOG(ERR, "Failed to subscribe CQE event.");
115 /* Init CQ to ones to be in HW owner in the start. */
116 cq->cq_obj.cqes[0].op_own = MLX5_CQE_OWNER_MASK;
117 cq->cq_obj.cqes[0].wqe_counter = rte_cpu_to_be_16(UINT16_MAX);
119 mlx5_vdpa_cq_arm(priv, cq);
122 mlx5_vdpa_cq_destroy(cq);
126 static inline uint32_t
127 mlx5_vdpa_cq_poll(struct mlx5_vdpa_cq *cq)
129 struct mlx5_vdpa_event_qp *eqp =
130 container_of(cq, struct mlx5_vdpa_event_qp, cq);
131 const unsigned int cq_size = 1 << cq->log_desc_n;
134 uint16_t wqe_counter;
140 uint16_t next_wqe_counter = eqp->qp_pi;
141 uint16_t cur_wqe_counter;
144 last_word.word = rte_read32(&cq->cq_obj.cqes[0].wqe_counter);
145 cur_wqe_counter = rte_be_to_cpu_16(last_word.wqe_counter);
146 comp = cur_wqe_counter + (uint16_t)1 - next_wqe_counter;
149 MLX5_ASSERT(MLX5_CQE_OPCODE(last_word.op_own) !=
151 if (unlikely(!(MLX5_CQE_OPCODE(last_word.op_own) ==
153 MLX5_CQE_OPCODE(last_word.op_own) ==
157 /* Ring CQ doorbell record. */
158 cq->cq_obj.db_rec[0] = rte_cpu_to_be_32(cq->cq_ci);
161 /* Ring SW QP doorbell record. */
162 eqp->sw_qp.db_rec[0] = rte_cpu_to_be_32(eqp->qp_pi + cq_size);
168 mlx5_vdpa_arm_all_cqs(struct mlx5_vdpa_priv *priv)
170 struct mlx5_vdpa_cq *cq;
173 for (i = 0; i < priv->nr_virtqs; i++) {
174 cq = &priv->virtqs[i].eqp.cq;
175 if (cq->cq_obj.cq && !cq->armed)
176 mlx5_vdpa_cq_arm(priv, cq);
181 mlx5_vdpa_timer_sleep(struct mlx5_vdpa_priv *priv, uint32_t max)
183 if (priv->event_mode == MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER) {
186 priv->timer_delay_us += priv->event_us;
191 priv->timer_delay_us /= max;
195 if (priv->timer_delay_us)
196 usleep(priv->timer_delay_us);
198 /* Give-up CPU to improve polling threads scheduling. */
202 /* Notify virtio device for specific virtq new traffic. */
204 mlx5_vdpa_queue_complete(struct mlx5_vdpa_cq *cq)
209 comp = mlx5_vdpa_cq_poll(cq);
211 if (cq->callfd != -1)
212 eventfd_write(cq->callfd, (eventfd_t)1);
219 /* Notify virtio device for any virtq new traffic. */
221 mlx5_vdpa_queues_complete(struct mlx5_vdpa_priv *priv)
226 for (i = 0; i < priv->nr_virtqs; i++) {
227 struct mlx5_vdpa_cq *cq = &priv->virtqs[i].eqp.cq;
228 uint32_t comp = mlx5_vdpa_queue_complete(cq);
237 mlx5_vdpa_drain_cq(struct mlx5_vdpa_priv *priv)
241 for (i = 0; i < priv->caps.max_num_virtio_queues * 2; i++) {
242 struct mlx5_vdpa_cq *cq = &priv->virtqs[i].eqp.cq;
244 mlx5_vdpa_queue_complete(cq);
246 cq->cq_obj.cqes[0].wqe_counter =
247 rte_cpu_to_be_16(UINT16_MAX);
248 priv->virtqs[i].eqp.qp_pi = 0;
250 mlx5_vdpa_cq_arm(priv, cq);
255 /* Wait on all CQs channel for completion event. */
256 static struct mlx5_vdpa_cq *
257 mlx5_vdpa_event_wait(struct mlx5_vdpa_priv *priv __rte_unused)
259 #ifdef HAVE_IBV_DEVX_EVENT
261 struct mlx5dv_devx_async_event_hdr event_resp;
262 uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) + 128];
264 int ret = mlx5_glue->devx_get_event(priv->eventc, &out.event_resp,
268 return (struct mlx5_vdpa_cq *)(uintptr_t)out.event_resp.cookie;
269 DRV_LOG(INFO, "Got error in devx_get_event, ret = %d, errno = %d.",
276 mlx5_vdpa_event_handle(void *arg)
278 struct mlx5_vdpa_priv *priv = arg;
279 struct mlx5_vdpa_cq *cq;
282 switch (priv->event_mode) {
283 case MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER:
284 case MLX5_VDPA_EVENT_MODE_FIXED_TIMER:
285 priv->timer_delay_us = priv->event_us;
287 pthread_mutex_lock(&priv->vq_config_lock);
288 max = mlx5_vdpa_queues_complete(priv);
289 if (max == 0 && priv->no_traffic_counter++ >=
290 priv->no_traffic_max) {
291 DRV_LOG(DEBUG, "Device %s traffic was stopped.",
292 priv->vdev->device->name);
293 mlx5_vdpa_arm_all_cqs(priv);
296 (&priv->vq_config_lock);
297 cq = mlx5_vdpa_event_wait(priv);
299 (&priv->vq_config_lock);
301 mlx5_vdpa_queue_complete(cq) > 0)
304 priv->timer_delay_us = priv->event_us;
305 priv->no_traffic_counter = 0;
306 } else if (max != 0) {
307 priv->no_traffic_counter = 0;
309 pthread_mutex_unlock(&priv->vq_config_lock);
310 mlx5_vdpa_timer_sleep(priv, max);
313 case MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT:
315 cq = mlx5_vdpa_event_wait(priv);
317 pthread_mutex_lock(&priv->vq_config_lock);
318 if (mlx5_vdpa_queue_complete(cq) > 0)
319 mlx5_vdpa_cq_arm(priv, cq);
320 pthread_mutex_unlock(&priv->vq_config_lock);
330 mlx5_vdpa_err_interrupt_handler(void *cb_arg __rte_unused)
332 #ifdef HAVE_IBV_DEVX_EVENT
333 struct mlx5_vdpa_priv *priv = cb_arg;
335 struct mlx5dv_devx_async_event_hdr event_resp;
336 uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) + 128];
338 uint32_t vq_index, i, version;
339 struct mlx5_vdpa_virtq *virtq;
342 pthread_mutex_lock(&priv->vq_config_lock);
343 while (mlx5_glue->devx_get_event(priv->err_chnl, &out.event_resp,
345 (ssize_t)sizeof(out.event_resp.cookie)) {
346 vq_index = out.event_resp.cookie & UINT32_MAX;
347 version = out.event_resp.cookie >> 32;
348 if (vq_index >= priv->nr_virtqs) {
349 DRV_LOG(ERR, "Invalid device %s error event virtq %d.",
350 priv->vdev->device->name, vq_index);
353 virtq = &priv->virtqs[vq_index];
354 if (!virtq->enable || virtq->version != version)
356 if (rte_rdtsc() / rte_get_tsc_hz() < MLX5_VDPA_ERROR_TIME_SEC)
358 virtq->stopped = true;
359 /* Query error info. */
360 if (mlx5_vdpa_virtq_query(priv, vq_index))
363 if (mlx5_vdpa_virtq_enable(priv, vq_index, 0)) {
364 DRV_LOG(ERR, "Failed to disable virtq %d.", vq_index);
367 /* Retry if error happens less than N times in 3 seconds. */
368 sec = (rte_rdtsc() - virtq->err_time[0]) / rte_get_tsc_hz();
369 if (sec > MLX5_VDPA_ERROR_TIME_SEC) {
371 if (mlx5_vdpa_virtq_enable(priv, vq_index, 1))
372 DRV_LOG(ERR, "Failed to enable virtq %d.",
375 DRV_LOG(WARNING, "Recover virtq %d: %u.",
376 vq_index, ++virtq->n_retry);
378 /* Retry timeout, give up. */
379 DRV_LOG(ERR, "Device %s virtq %d failed to recover.",
380 priv->vdev->device->name, vq_index);
383 /* Shift in current time to error time log end. */
384 for (i = 1; i < RTE_DIM(virtq->err_time); i++)
385 virtq->err_time[i - 1] = virtq->err_time[i];
386 virtq->err_time[RTE_DIM(virtq->err_time) - 1] = rte_rdtsc();
388 pthread_mutex_unlock(&priv->vq_config_lock);
393 mlx5_vdpa_err_event_setup(struct mlx5_vdpa_priv *priv)
398 /* Setup device event channel. */
399 priv->err_chnl = mlx5_glue->devx_create_event_channel(priv->cdev->ctx,
401 if (!priv->err_chnl) {
403 DRV_LOG(ERR, "Failed to create device event channel %d.",
407 flags = fcntl(priv->err_chnl->fd, F_GETFL);
408 ret = fcntl(priv->err_chnl->fd, F_SETFL, flags | O_NONBLOCK);
411 DRV_LOG(ERR, "Failed to change device event channel FD.");
414 priv->err_intr_handle =
415 rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
416 if (priv->err_intr_handle == NULL) {
417 DRV_LOG(ERR, "Fail to allocate intr_handle");
420 if (rte_intr_fd_set(priv->err_intr_handle, priv->err_chnl->fd))
423 if (rte_intr_type_set(priv->err_intr_handle, RTE_INTR_HANDLE_EXT))
426 ret = rte_intr_callback_register(priv->err_intr_handle,
427 mlx5_vdpa_err_interrupt_handler,
430 rte_intr_fd_set(priv->err_intr_handle, 0);
431 DRV_LOG(ERR, "Failed to register error interrupt for device %d.",
436 DRV_LOG(DEBUG, "Registered error interrupt for device%d.",
441 mlx5_vdpa_err_event_unset(priv);
446 mlx5_vdpa_err_event_unset(struct mlx5_vdpa_priv *priv)
448 int retries = MLX5_VDPA_INTR_RETRIES;
451 if (!rte_intr_fd_get(priv->err_intr_handle))
453 while (retries-- && ret == -EAGAIN) {
454 ret = rte_intr_callback_unregister(priv->err_intr_handle,
455 mlx5_vdpa_err_interrupt_handler,
457 if (ret == -EAGAIN) {
458 DRV_LOG(DEBUG, "Try again to unregister fd %d "
459 "of error interrupt, retries = %d.",
460 rte_intr_fd_get(priv->err_intr_handle),
465 if (priv->err_chnl) {
466 #ifdef HAVE_IBV_DEVX_EVENT
468 struct mlx5dv_devx_async_event_hdr event_resp;
469 uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) +
473 /* Clean all pending events. */
474 while (mlx5_glue->devx_get_event(priv->err_chnl,
475 &out.event_resp, sizeof(out.buf)) >=
476 (ssize_t)sizeof(out.event_resp.cookie))
479 mlx5_glue->devx_destroy_event_channel(priv->err_chnl);
480 priv->err_chnl = NULL;
482 rte_intr_instance_free(priv->err_intr_handle);
486 mlx5_vdpa_cqe_event_setup(struct mlx5_vdpa_priv *priv)
492 const struct sched_param sp = {
493 .sched_priority = sched_get_priority_max(SCHED_RR),
497 /* All virtqs are in poll mode. */
499 pthread_attr_init(&attr);
500 ret = pthread_attr_setschedpolicy(&attr, SCHED_RR);
502 DRV_LOG(ERR, "Failed to set thread sched policy = RR.");
505 ret = pthread_attr_setschedparam(&attr, &sp);
507 DRV_LOG(ERR, "Failed to set thread priority.");
510 ret = pthread_create(&priv->timer_tid, &attr, mlx5_vdpa_event_handle,
513 DRV_LOG(ERR, "Failed to create timer thread.");
517 if (priv->event_core != -1)
518 CPU_SET(priv->event_core, &cpuset);
520 cpuset = rte_lcore_cpuset(rte_get_main_lcore());
521 ret = pthread_setaffinity_np(priv->timer_tid, sizeof(cpuset), &cpuset);
523 DRV_LOG(ERR, "Failed to set thread affinity.");
526 snprintf(name, sizeof(name), "vDPA-mlx5-%d", priv->vid);
527 ret = rte_thread_setname(priv->timer_tid, name);
529 DRV_LOG(DEBUG, "Cannot set timer thread name.");
534 mlx5_vdpa_cqe_event_unset(struct mlx5_vdpa_priv *priv)
538 if (priv->timer_tid) {
539 pthread_cancel(priv->timer_tid);
540 pthread_join(priv->timer_tid, &status);
546 mlx5_vdpa_event_qp_destroy(struct mlx5_vdpa_event_qp *eqp)
548 mlx5_devx_qp_destroy(&eqp->sw_qp);
550 claim_zero(mlx5_devx_cmd_destroy(eqp->fw_qp));
551 mlx5_vdpa_cq_destroy(&eqp->cq);
552 memset(eqp, 0, sizeof(*eqp));
556 mlx5_vdpa_qps2rts(struct mlx5_vdpa_event_qp *eqp)
558 if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_RST2INIT_QP,
559 eqp->sw_qp.qp->id)) {
560 DRV_LOG(ERR, "Failed to modify FW QP to INIT state(%u).",
564 if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp.qp,
565 MLX5_CMD_OP_RST2INIT_QP, eqp->fw_qp->id)) {
566 DRV_LOG(ERR, "Failed to modify SW QP to INIT state(%u).",
570 if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_INIT2RTR_QP,
571 eqp->sw_qp.qp->id)) {
572 DRV_LOG(ERR, "Failed to modify FW QP to RTR state(%u).",
576 if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp.qp,
577 MLX5_CMD_OP_INIT2RTR_QP, eqp->fw_qp->id)) {
578 DRV_LOG(ERR, "Failed to modify SW QP to RTR state(%u).",
582 if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_RTR2RTS_QP,
583 eqp->sw_qp.qp->id)) {
584 DRV_LOG(ERR, "Failed to modify FW QP to RTS state(%u).",
588 if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp.qp, MLX5_CMD_OP_RTR2RTS_QP,
590 DRV_LOG(ERR, "Failed to modify SW QP to RTS state(%u).",
598 mlx5_vdpa_qps2rst2rts(struct mlx5_vdpa_event_qp *eqp)
600 if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_QP_2RST,
601 eqp->sw_qp.qp->id)) {
602 DRV_LOG(ERR, "Failed to modify FW QP to RST state(%u).",
606 if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp.qp,
607 MLX5_CMD_OP_QP_2RST, eqp->fw_qp->id)) {
608 DRV_LOG(ERR, "Failed to modify SW QP to RST state(%u).",
612 return mlx5_vdpa_qps2rts(eqp);
616 mlx5_vdpa_event_qp_prepare(struct mlx5_vdpa_priv *priv, uint16_t desc_n,
617 int callfd, struct mlx5_vdpa_event_qp *eqp)
619 struct mlx5_devx_qp_attr attr = {0};
620 uint16_t log_desc_n = rte_log2_u32(desc_n);
623 if (eqp->cq.cq_obj.cq != NULL && log_desc_n == eqp->cq.log_desc_n) {
624 /* Reuse existing resources. */
625 eqp->cq.callfd = callfd;
626 /* FW will set event qp to error state in q destroy. */
627 if (!mlx5_vdpa_qps2rst2rts(eqp)) {
628 rte_write32(rte_cpu_to_be_32(RTE_BIT32(log_desc_n)),
629 &eqp->sw_qp.db_rec[0]);
634 mlx5_vdpa_event_qp_destroy(eqp);
635 if (mlx5_vdpa_cq_create(priv, log_desc_n, callfd, &eqp->cq))
637 attr.pd = priv->cdev->pdn;
639 mlx5_ts_format_conv(priv->cdev->config.hca_attr.qp_ts_format);
640 eqp->fw_qp = mlx5_devx_cmd_create_qp(priv->cdev->ctx, &attr);
642 DRV_LOG(ERR, "Failed to create FW QP(%u).", rte_errno);
645 attr.uar_index = mlx5_os_get_devx_uar_page_id(priv->uar.obj);
646 attr.cqn = eqp->cq.cq_obj.cq->id;
647 attr.num_of_receive_wqes = RTE_BIT32(log_desc_n);
648 attr.log_rq_stride = rte_log2_u32(MLX5_WSEG_SIZE);
649 attr.num_of_send_wqbbs = 0; /* No need SQ. */
651 mlx5_ts_format_conv(priv->cdev->config.hca_attr.qp_ts_format);
652 ret = mlx5_devx_qp_create(priv->cdev->ctx, &(eqp->sw_qp),
653 attr.num_of_receive_wqes *
654 MLX5_WSEG_SIZE, &attr, SOCKET_ID_ANY);
656 DRV_LOG(ERR, "Failed to create SW QP(%u).", rte_errno);
659 if (mlx5_vdpa_qps2rts(eqp))
663 if (eqp->sw_qp.db_rec)
664 rte_write32(rte_cpu_to_be_32(RTE_BIT32(log_desc_n)),
665 &eqp->sw_qp.db_rec[0]);
668 mlx5_vdpa_event_qp_destroy(eqp);