1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2019 Mellanox Technologies, Ltd
7 #include <sys/eventfd.h>
9 #include <rte_malloc.h>
10 #include <rte_errno.h>
11 #include <rte_lcore.h>
12 #include <rte_atomic.h>
13 #include <rte_common.h>
15 #include <rte_alarm.h>
17 #include <mlx5_common.h>
18 #include <mlx5_glue.h>
20 #include "mlx5_vdpa_utils.h"
21 #include "mlx5_vdpa.h"
24 #define MLX5_VDPA_ERROR_TIME_SEC 3u
27 mlx5_vdpa_event_qp_global_release(struct mlx5_vdpa_priv *priv)
30 mlx5_glue->devx_free_uar(priv->uar);
33 #ifdef HAVE_IBV_DEVX_EVENT
36 struct mlx5dv_devx_async_event_hdr event_resp;
37 uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr)
41 /* Clean all pending events. */
42 while (mlx5_glue->devx_get_event(priv->eventc, &out.event_resp,
44 (ssize_t)sizeof(out.event_resp.cookie))
46 mlx5_glue->devx_destroy_event_channel(priv->eventc);
53 /* Prepare all the global resources for all the event objects.*/
55 mlx5_vdpa_event_qp_global_prepare(struct mlx5_vdpa_priv *priv)
61 if (mlx5_glue->devx_query_eqn(priv->ctx, 0, &priv->eqn)) {
63 DRV_LOG(ERR, "Failed to query EQ number %d.", rte_errno);
66 priv->eventc = mlx5_glue->devx_create_event_channel(priv->ctx,
67 MLX5DV_DEVX_CREATE_EVENT_CHANNEL_FLAGS_OMIT_EV_DATA);
70 DRV_LOG(ERR, "Failed to create event channel %d.",
74 flags = fcntl(priv->eventc->fd, F_GETFL);
75 ret = fcntl(priv->eventc->fd, F_SETFL, flags | O_NONBLOCK);
77 DRV_LOG(ERR, "Failed to change event channel FD.");
81 * This PMD always claims the write memory barrier on UAR
82 * registers writings, it is safe to allocate UAR with any
83 * memory mapping type.
85 priv->uar = mlx5_devx_alloc_uar(priv->ctx, -1);
88 DRV_LOG(ERR, "Failed to allocate UAR.");
93 mlx5_vdpa_event_qp_global_release(priv);
98 mlx5_vdpa_cq_destroy(struct mlx5_vdpa_cq *cq)
101 claim_zero(mlx5_devx_cmd_destroy(cq->cq));
103 claim_zero(mlx5_glue->devx_umem_dereg(cq->umem_obj));
105 rte_free((void *)(uintptr_t)cq->umem_buf);
106 memset(cq, 0, sizeof(*cq));
109 static inline void __rte_unused
110 mlx5_vdpa_cq_arm(struct mlx5_vdpa_priv *priv, struct mlx5_vdpa_cq *cq)
112 uint32_t arm_sn = cq->arm_sn << MLX5_CQ_SQN_OFFSET;
113 uint32_t cq_ci = cq->cq_ci & MLX5_CI_MASK;
114 uint32_t doorbell_hi = arm_sn | MLX5_CQ_DBR_CMD_ALL | cq_ci;
115 uint64_t doorbell = ((uint64_t)doorbell_hi << 32) | cq->cq->id;
116 uint64_t db_be = rte_cpu_to_be_64(doorbell);
117 uint32_t *addr = RTE_PTR_ADD(priv->uar->base_addr, MLX5_CQ_DOORBELL);
120 cq->db_rec[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
123 *(uint64_t *)addr = db_be;
125 *(uint32_t *)addr = db_be;
127 *((uint32_t *)addr + 1) = db_be >> 32;
134 mlx5_vdpa_cq_create(struct mlx5_vdpa_priv *priv, uint16_t log_desc_n,
135 int callfd, struct mlx5_vdpa_cq *cq)
137 struct mlx5_devx_cq_attr attr = {0};
138 size_t pgsize = sysconf(_SC_PAGESIZE);
140 uint16_t event_nums[1] = {0};
141 uint16_t cq_size = 1 << log_desc_n;
144 cq->log_desc_n = log_desc_n;
145 umem_size = sizeof(struct mlx5_cqe) * cq_size + sizeof(*cq->db_rec) * 2;
146 cq->umem_buf = rte_zmalloc(__func__, umem_size, 4096);
148 DRV_LOG(ERR, "Failed to allocate memory for CQ.");
152 cq->umem_obj = mlx5_glue->devx_umem_reg(priv->ctx,
153 (void *)(uintptr_t)cq->umem_buf,
155 IBV_ACCESS_LOCAL_WRITE);
157 DRV_LOG(ERR, "Failed to register umem for CQ.");
160 attr.q_umem_valid = 1;
161 attr.db_umem_valid = 1;
162 attr.use_first_only = 1;
163 attr.overrun_ignore = 0;
164 attr.uar_page_id = priv->uar->page_id;
165 attr.q_umem_id = cq->umem_obj->umem_id;
166 attr.q_umem_offset = 0;
167 attr.db_umem_id = cq->umem_obj->umem_id;
168 attr.db_umem_offset = sizeof(struct mlx5_cqe) * cq_size;
169 attr.eqn = priv->eqn;
170 attr.log_cq_size = log_desc_n;
171 attr.log_page_size = rte_log2_u32(pgsize);
172 cq->cq = mlx5_devx_cmd_create_cq(priv->ctx, &attr);
175 cq->db_rec = RTE_PTR_ADD(cq->umem_buf, (uintptr_t)attr.db_umem_offset);
177 rte_spinlock_init(&cq->sl);
178 /* Subscribe CQ event to the event channel controlled by the driver. */
179 ret = mlx5_glue->devx_subscribe_devx_event(priv->eventc, cq->cq->obj,
182 (uint64_t)(uintptr_t)cq);
184 DRV_LOG(ERR, "Failed to subscribe CQE event.");
189 /* Init CQ to ones to be in HW owner in the start. */
190 cq->cqes[0].op_own = MLX5_CQE_OWNER_MASK;
191 cq->cqes[0].wqe_counter = rte_cpu_to_be_16(UINT16_MAX);
193 mlx5_vdpa_cq_arm(priv, cq);
196 mlx5_vdpa_cq_destroy(cq);
200 static inline uint32_t
201 mlx5_vdpa_cq_poll(struct mlx5_vdpa_cq *cq)
203 struct mlx5_vdpa_event_qp *eqp =
204 container_of(cq, struct mlx5_vdpa_event_qp, cq);
205 const unsigned int cq_size = 1 << cq->log_desc_n;
208 uint16_t wqe_counter;
214 uint16_t next_wqe_counter = cq->cq_ci;
215 uint16_t cur_wqe_counter;
218 last_word.word = rte_read32(&cq->cqes[0].wqe_counter);
219 cur_wqe_counter = rte_be_to_cpu_16(last_word.wqe_counter);
220 comp = cur_wqe_counter + (uint16_t)1 - next_wqe_counter;
223 MLX5_ASSERT(MLX5_CQE_OPCODE(last_word.op_own) !=
225 if (unlikely(!(MLX5_CQE_OPCODE(last_word.op_own) ==
227 MLX5_CQE_OPCODE(last_word.op_own) ==
231 /* Ring CQ doorbell record. */
232 cq->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci);
234 /* Ring SW QP doorbell record. */
235 eqp->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci + cq_size);
241 mlx5_vdpa_arm_all_cqs(struct mlx5_vdpa_priv *priv)
243 struct mlx5_vdpa_cq *cq;
246 for (i = 0; i < priv->nr_virtqs; i++) {
247 cq = &priv->virtqs[i].eqp.cq;
248 if (cq->cq && !cq->armed)
249 mlx5_vdpa_cq_arm(priv, cq);
254 mlx5_vdpa_timer_sleep(struct mlx5_vdpa_priv *priv, uint32_t max)
256 if (priv->event_mode == MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER) {
259 priv->timer_delay_us += priv->event_us;
264 priv->timer_delay_us /= max;
268 usleep(priv->timer_delay_us);
272 mlx5_vdpa_poll_handle(void *arg)
274 struct mlx5_vdpa_priv *priv = arg;
276 struct mlx5_vdpa_cq *cq;
278 uint64_t current_tic;
280 pthread_mutex_lock(&priv->timer_lock);
281 while (!priv->timer_on)
282 pthread_cond_wait(&priv->timer_cond, &priv->timer_lock);
283 pthread_mutex_unlock(&priv->timer_lock);
284 priv->timer_delay_us = priv->event_mode ==
285 MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER ?
286 MLX5_VDPA_DEFAULT_TIMER_DELAY_US :
290 pthread_mutex_lock(&priv->vq_config_lock);
291 for (i = 0; i < priv->nr_virtqs; i++) {
292 cq = &priv->virtqs[i].eqp.cq;
293 if (cq->cq && !cq->armed) {
294 uint32_t comp = mlx5_vdpa_cq_poll(cq);
297 /* Notify guest for descs consuming. */
298 if (cq->callfd != -1)
299 eventfd_write(cq->callfd,
306 current_tic = rte_rdtsc();
308 /* No traffic ? stop timer and load interrupts. */
309 if (current_tic - priv->last_traffic_tic >=
310 rte_get_timer_hz() * priv->no_traffic_time_s) {
311 DRV_LOG(DEBUG, "Device %s traffic was stopped.",
312 priv->vdev->device->name);
313 mlx5_vdpa_arm_all_cqs(priv);
314 pthread_mutex_unlock(&priv->vq_config_lock);
315 pthread_mutex_lock(&priv->timer_lock);
317 while (!priv->timer_on)
318 pthread_cond_wait(&priv->timer_cond,
320 pthread_mutex_unlock(&priv->timer_lock);
321 priv->timer_delay_us = priv->event_mode ==
322 MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER ?
323 MLX5_VDPA_DEFAULT_TIMER_DELAY_US :
328 priv->last_traffic_tic = current_tic;
330 pthread_mutex_unlock(&priv->vq_config_lock);
331 mlx5_vdpa_timer_sleep(priv, max);
337 mlx5_vdpa_interrupt_handler(void *cb_arg)
339 struct mlx5_vdpa_priv *priv = cb_arg;
340 #ifdef HAVE_IBV_DEVX_EVENT
342 struct mlx5dv_devx_async_event_hdr event_resp;
343 uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) + 128];
346 pthread_mutex_lock(&priv->vq_config_lock);
347 while (mlx5_glue->devx_get_event(priv->eventc, &out.event_resp,
349 (ssize_t)sizeof(out.event_resp.cookie)) {
350 struct mlx5_vdpa_cq *cq = (struct mlx5_vdpa_cq *)
351 (uintptr_t)out.event_resp.cookie;
352 struct mlx5_vdpa_event_qp *eqp = container_of(cq,
353 struct mlx5_vdpa_event_qp, cq);
354 struct mlx5_vdpa_virtq *virtq = container_of(eqp,
355 struct mlx5_vdpa_virtq, eqp);
359 mlx5_vdpa_cq_poll(cq);
360 /* Notify guest for descs consuming. */
361 if (cq->callfd != -1)
362 eventfd_write(cq->callfd, (eventfd_t)1);
363 if (priv->event_mode == MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT) {
364 mlx5_vdpa_cq_arm(priv, cq);
365 pthread_mutex_unlock(&priv->vq_config_lock);
368 /* Don't arm again - timer will take control. */
369 DRV_LOG(DEBUG, "Device %s virtq %d cq %d event was captured."
370 " Timer is %s, cq ci is %u.\n",
371 priv->vdev->device->name,
372 (int)virtq->index, cq->cq->id,
373 priv->timer_on ? "on" : "off", cq->cq_ci);
378 /* Traffic detected: make sure timer is on. */
379 priv->last_traffic_tic = rte_rdtsc();
380 pthread_mutex_lock(&priv->timer_lock);
381 if (!priv->timer_on) {
383 pthread_cond_signal(&priv->timer_cond);
385 pthread_mutex_unlock(&priv->timer_lock);
386 pthread_mutex_unlock(&priv->vq_config_lock);
390 mlx5_vdpa_err_interrupt_handler(void *cb_arg __rte_unused)
392 #ifdef HAVE_IBV_DEVX_EVENT
393 struct mlx5_vdpa_priv *priv = cb_arg;
395 struct mlx5dv_devx_async_event_hdr event_resp;
396 uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) + 128];
398 uint32_t vq_index, i, version;
399 struct mlx5_vdpa_virtq *virtq;
402 pthread_mutex_lock(&priv->vq_config_lock);
403 while (mlx5_glue->devx_get_event(priv->err_chnl, &out.event_resp,
405 (ssize_t)sizeof(out.event_resp.cookie)) {
406 vq_index = out.event_resp.cookie & UINT32_MAX;
407 version = out.event_resp.cookie >> 32;
408 if (vq_index >= priv->nr_virtqs) {
409 DRV_LOG(ERR, "Invalid device %s error event virtq %d.",
410 priv->vdev->device->name, vq_index);
413 virtq = &priv->virtqs[vq_index];
414 if (!virtq->enable || virtq->version != version)
416 if (rte_rdtsc() / rte_get_tsc_hz() < MLX5_VDPA_ERROR_TIME_SEC)
418 virtq->stopped = true;
419 /* Query error info. */
420 if (mlx5_vdpa_virtq_query(priv, vq_index))
423 if (mlx5_vdpa_virtq_enable(priv, vq_index, 0)) {
424 DRV_LOG(ERR, "Failed to disable virtq %d.", vq_index);
427 /* Retry if error happens less than N times in 3 seconds. */
428 sec = (rte_rdtsc() - virtq->err_time[0]) / rte_get_tsc_hz();
429 if (sec > MLX5_VDPA_ERROR_TIME_SEC) {
431 if (mlx5_vdpa_virtq_enable(priv, vq_index, 1))
432 DRV_LOG(ERR, "Failed to enable virtq %d.",
435 DRV_LOG(WARNING, "Recover virtq %d: %u.",
436 vq_index, ++virtq->n_retry);
438 /* Retry timeout, give up. */
439 DRV_LOG(ERR, "Device %s virtq %d failed to recover.",
440 priv->vdev->device->name, vq_index);
443 /* Shift in current time to error time log end. */
444 for (i = 1; i < RTE_DIM(virtq->err_time); i++)
445 virtq->err_time[i - 1] = virtq->err_time[i];
446 virtq->err_time[RTE_DIM(virtq->err_time) - 1] = rte_rdtsc();
448 pthread_mutex_unlock(&priv->vq_config_lock);
453 mlx5_vdpa_err_event_setup(struct mlx5_vdpa_priv *priv)
458 /* Setup device event channel. */
459 priv->err_chnl = mlx5_glue->devx_create_event_channel(priv->ctx, 0);
460 if (!priv->err_chnl) {
462 DRV_LOG(ERR, "Failed to create device event channel %d.",
466 flags = fcntl(priv->err_chnl->fd, F_GETFL);
467 ret = fcntl(priv->err_chnl->fd, F_SETFL, flags | O_NONBLOCK);
469 DRV_LOG(ERR, "Failed to change device event channel FD.");
472 priv->err_intr_handle.fd = priv->err_chnl->fd;
473 priv->err_intr_handle.type = RTE_INTR_HANDLE_EXT;
474 if (rte_intr_callback_register(&priv->err_intr_handle,
475 mlx5_vdpa_err_interrupt_handler,
477 priv->err_intr_handle.fd = 0;
478 DRV_LOG(ERR, "Failed to register error interrupt for device %d.",
482 DRV_LOG(DEBUG, "Registered error interrupt for device%d.",
487 mlx5_vdpa_err_event_unset(priv);
492 mlx5_vdpa_err_event_unset(struct mlx5_vdpa_priv *priv)
494 int retries = MLX5_VDPA_INTR_RETRIES;
497 if (!priv->err_intr_handle.fd)
499 while (retries-- && ret == -EAGAIN) {
500 ret = rte_intr_callback_unregister(&priv->err_intr_handle,
501 mlx5_vdpa_err_interrupt_handler,
503 if (ret == -EAGAIN) {
504 DRV_LOG(DEBUG, "Try again to unregister fd %d "
505 "of error interrupt, retries = %d.",
506 priv->err_intr_handle.fd, retries);
510 memset(&priv->err_intr_handle, 0, sizeof(priv->err_intr_handle));
511 if (priv->err_chnl) {
512 #ifdef HAVE_IBV_DEVX_EVENT
514 struct mlx5dv_devx_async_event_hdr event_resp;
515 uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) +
519 /* Clean all pending events. */
520 while (mlx5_glue->devx_get_event(priv->err_chnl,
521 &out.event_resp, sizeof(out.buf)) >=
522 (ssize_t)sizeof(out.event_resp.cookie))
525 mlx5_glue->devx_destroy_event_channel(priv->err_chnl);
526 priv->err_chnl = NULL;
531 mlx5_vdpa_cqe_event_setup(struct mlx5_vdpa_priv *priv)
536 /* All virtqs are in poll mode. */
538 if (priv->event_mode != MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT) {
539 pthread_mutex_init(&priv->timer_lock, NULL);
540 pthread_cond_init(&priv->timer_cond, NULL);
542 ret = pthread_create(&priv->timer_tid, NULL,
543 mlx5_vdpa_poll_handle, (void *)priv);
545 DRV_LOG(ERR, "Failed to create timer thread.");
549 priv->intr_handle.fd = priv->eventc->fd;
550 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
551 if (rte_intr_callback_register(&priv->intr_handle,
552 mlx5_vdpa_interrupt_handler, priv)) {
553 priv->intr_handle.fd = 0;
554 DRV_LOG(ERR, "Failed to register CQE interrupt %d.", rte_errno);
559 mlx5_vdpa_cqe_event_unset(priv);
564 mlx5_vdpa_cqe_event_unset(struct mlx5_vdpa_priv *priv)
566 int retries = MLX5_VDPA_INTR_RETRIES;
570 if (priv->intr_handle.fd) {
571 while (retries-- && ret == -EAGAIN) {
572 ret = rte_intr_callback_unregister(&priv->intr_handle,
573 mlx5_vdpa_interrupt_handler,
575 if (ret == -EAGAIN) {
576 DRV_LOG(DEBUG, "Try again to unregister fd %d "
577 "of CQ interrupt, retries = %d.",
578 priv->intr_handle.fd, retries);
582 memset(&priv->intr_handle, 0, sizeof(priv->intr_handle));
584 if (priv->timer_tid) {
585 pthread_cancel(priv->timer_tid);
586 pthread_join(priv->timer_tid, &status);
592 mlx5_vdpa_event_qp_destroy(struct mlx5_vdpa_event_qp *eqp)
595 claim_zero(mlx5_devx_cmd_destroy(eqp->sw_qp));
597 claim_zero(mlx5_glue->devx_umem_dereg(eqp->umem_obj));
599 rte_free(eqp->umem_buf);
601 claim_zero(mlx5_devx_cmd_destroy(eqp->fw_qp));
602 mlx5_vdpa_cq_destroy(&eqp->cq);
603 memset(eqp, 0, sizeof(*eqp));
607 mlx5_vdpa_qps2rts(struct mlx5_vdpa_event_qp *eqp)
609 if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_RST2INIT_QP,
611 DRV_LOG(ERR, "Failed to modify FW QP to INIT state(%u).",
615 if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp, MLX5_CMD_OP_RST2INIT_QP,
617 DRV_LOG(ERR, "Failed to modify SW QP to INIT state(%u).",
621 if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_INIT2RTR_QP,
623 DRV_LOG(ERR, "Failed to modify FW QP to RTR state(%u).",
627 if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp, MLX5_CMD_OP_INIT2RTR_QP,
629 DRV_LOG(ERR, "Failed to modify SW QP to RTR state(%u).",
633 if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_RTR2RTS_QP,
635 DRV_LOG(ERR, "Failed to modify FW QP to RTS state(%u).",
639 if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp, MLX5_CMD_OP_RTR2RTS_QP,
641 DRV_LOG(ERR, "Failed to modify SW QP to RTS state(%u).",
649 mlx5_vdpa_event_qp_create(struct mlx5_vdpa_priv *priv, uint16_t desc_n,
650 int callfd, struct mlx5_vdpa_event_qp *eqp)
652 struct mlx5_devx_qp_attr attr = {0};
653 uint16_t log_desc_n = rte_log2_u32(desc_n);
654 uint32_t umem_size = (1 << log_desc_n) * MLX5_WSEG_SIZE +
655 sizeof(*eqp->db_rec) * 2;
657 if (mlx5_vdpa_event_qp_global_prepare(priv))
659 if (mlx5_vdpa_cq_create(priv, log_desc_n, callfd, &eqp->cq))
662 eqp->fw_qp = mlx5_devx_cmd_create_qp(priv->ctx, &attr);
664 DRV_LOG(ERR, "Failed to create FW QP(%u).", rte_errno);
667 eqp->umem_buf = rte_zmalloc(__func__, umem_size, 4096);
668 if (!eqp->umem_buf) {
669 DRV_LOG(ERR, "Failed to allocate memory for SW QP.");
673 eqp->umem_obj = mlx5_glue->devx_umem_reg(priv->ctx,
674 (void *)(uintptr_t)eqp->umem_buf,
676 IBV_ACCESS_LOCAL_WRITE);
677 if (!eqp->umem_obj) {
678 DRV_LOG(ERR, "Failed to register umem for SW QP.");
681 attr.uar_index = priv->uar->page_id;
682 attr.cqn = eqp->cq.cq->id;
683 attr.log_page_size = rte_log2_u32(sysconf(_SC_PAGESIZE));
684 attr.rq_size = 1 << log_desc_n;
685 attr.log_rq_stride = rte_log2_u32(MLX5_WSEG_SIZE);
686 attr.sq_size = 0; /* No need SQ. */
687 attr.dbr_umem_valid = 1;
688 attr.wq_umem_id = eqp->umem_obj->umem_id;
689 attr.wq_umem_offset = 0;
690 attr.dbr_umem_id = eqp->umem_obj->umem_id;
691 attr.dbr_address = (1 << log_desc_n) * MLX5_WSEG_SIZE;
692 eqp->sw_qp = mlx5_devx_cmd_create_qp(priv->ctx, &attr);
694 DRV_LOG(ERR, "Failed to create SW QP(%u).", rte_errno);
697 eqp->db_rec = RTE_PTR_ADD(eqp->umem_buf, (uintptr_t)attr.dbr_address);
698 if (mlx5_vdpa_qps2rts(eqp))
701 rte_write32(rte_cpu_to_be_32(1 << log_desc_n), &eqp->db_rec[0]);
704 mlx5_vdpa_event_qp_destroy(eqp);