#include <rte_alarm.h>
#include <mlx5_common.h>
+#include <mlx5_common_os.h>
+#include <mlx5_glue.h>
#include "mlx5_vdpa_utils.h"
#include "mlx5_vdpa.h"
-#define MLX5_VDPA_DEFAULT_TIMER_DELAY_US 500u
-#define MLX5_VDPA_NO_TRAFFIC_TIME_S 2LLU
+#define MLX5_VDPA_ERROR_TIME_SEC 3u
void
mlx5_vdpa_event_qp_global_release(struct mlx5_vdpa_priv *priv)
sizeof(out.buf)) >=
(ssize_t)sizeof(out.event_resp.cookie))
;
- mlx5_glue->devx_destroy_event_channel(priv->eventc);
+ mlx5_os_devx_destroy_event_channel(priv->eventc);
priv->eventc = NULL;
}
#endif
static int
mlx5_vdpa_event_qp_global_prepare(struct mlx5_vdpa_priv *priv)
{
- uint32_t lcore;
+ int flags, ret;
if (priv->eventc)
return 0;
- lcore = (uint32_t)rte_lcore_to_cpu_id(-1);
- if (mlx5_glue->devx_query_eqn(priv->ctx, lcore, &priv->eqn)) {
+ if (mlx5_glue->devx_query_eqn(priv->ctx, 0, &priv->eqn)) {
rte_errno = errno;
DRV_LOG(ERR, "Failed to query EQ number %d.", rte_errno);
return -1;
}
- priv->eventc = mlx5_glue->devx_create_event_channel(priv->ctx,
+ priv->eventc = mlx5_os_devx_create_event_channel(priv->ctx,
MLX5DV_DEVX_CREATE_EVENT_CHANNEL_FLAGS_OMIT_EV_DATA);
if (!priv->eventc) {
rte_errno = errno;
rte_errno);
goto error;
}
- priv->uar = mlx5_glue->devx_alloc_uar(priv->ctx, 0);
+ flags = fcntl(priv->eventc->fd, F_GETFL);
+ ret = fcntl(priv->eventc->fd, F_SETFL, flags | O_NONBLOCK);
+ if (ret) {
+ DRV_LOG(ERR, "Failed to change event channel FD.");
+ goto error;
+ }
+ /*
+ * This PMD always claims the write memory barrier on UAR
+ * registers writings, it is safe to allocate UAR with any
+ * memory mapping type.
+ */
+ priv->uar = mlx5_devx_alloc_uar(priv->ctx, -1);
if (!priv->uar) {
rte_errno = errno;
DRV_LOG(ERR, "Failed to allocate UAR.");
mlx5_vdpa_cq_create(struct mlx5_vdpa_priv *priv, uint16_t log_desc_n,
int callfd, struct mlx5_vdpa_cq *cq)
{
- struct mlx5_devx_cq_attr attr;
+ struct mlx5_devx_cq_attr attr = {0};
size_t pgsize = sysconf(_SC_PAGESIZE);
uint32_t umem_size;
- int ret;
uint16_t event_nums[1] = {0};
+ uint16_t cq_size = 1 << log_desc_n;
+ int ret;
cq->log_desc_n = log_desc_n;
- umem_size = sizeof(struct mlx5_cqe) * (1 << log_desc_n) +
- sizeof(*cq->db_rec) * 2;
+ umem_size = sizeof(struct mlx5_cqe) * cq_size + sizeof(*cq->db_rec) * 2;
cq->umem_buf = rte_zmalloc(__func__, umem_size, 4096);
if (!cq->umem_buf) {
DRV_LOG(ERR, "Failed to allocate memory for CQ.");
}
attr.q_umem_valid = 1;
attr.db_umem_valid = 1;
- attr.use_first_only = 0;
+ attr.use_first_only = 1;
attr.overrun_ignore = 0;
attr.uar_page_id = priv->uar->page_id;
attr.q_umem_id = cq->umem_obj->umem_id;
attr.q_umem_offset = 0;
attr.db_umem_id = cq->umem_obj->umem_id;
- attr.db_umem_offset = sizeof(struct mlx5_cqe) * (1 << log_desc_n);
+ attr.db_umem_offset = sizeof(struct mlx5_cqe) * cq_size;
attr.eqn = priv->eqn;
attr.log_cq_size = log_desc_n;
attr.log_page_size = rte_log2_u32(pgsize);
cq->cq_ci = 0;
rte_spinlock_init(&cq->sl);
/* Subscribe CQ event to the event channel controlled by the driver. */
- ret = mlx5_glue->devx_subscribe_devx_event(priv->eventc, cq->cq->obj,
+ ret = mlx5_os_devx_subscribe_devx_event(priv->eventc, cq->cq->obj,
sizeof(event_nums),
event_nums,
(uint64_t)(uintptr_t)cq);
rte_errno = errno;
goto error;
}
- if (callfd != -1) {
- ret = mlx5_glue->devx_subscribe_devx_event_fd(priv->eventc,
- callfd,
- cq->cq->obj, 0);
- if (ret) {
- DRV_LOG(ERR, "Failed to subscribe CQE event fd.");
- rte_errno = errno;
- goto error;
- }
- }
cq->callfd = callfd;
/* Init CQ to ones to be in HW owner in the start. */
- memset((void *)(uintptr_t)cq->umem_buf, 0xFF, attr.db_umem_offset);
+ cq->cqes[0].op_own = MLX5_CQE_OWNER_MASK;
+ cq->cqes[0].wqe_counter = rte_cpu_to_be_16(UINT16_MAX);
/* First arming. */
mlx5_vdpa_cq_arm(priv, cq);
return 0;
struct mlx5_vdpa_event_qp *eqp =
container_of(cq, struct mlx5_vdpa_event_qp, cq);
const unsigned int cq_size = 1 << cq->log_desc_n;
- const unsigned int cq_mask = cq_size - 1;
- uint32_t total = 0;
- int ret;
-
- do {
- volatile struct mlx5_cqe *cqe = cq->cqes + ((cq->cq_ci + total)
- & cq_mask);
-
- ret = check_cqe(cqe, cq_size, cq->cq_ci + total);
- switch (ret) {
- case MLX5_CQE_STATUS_ERR:
+ union {
+ struct {
+ uint16_t wqe_counter;
+ uint8_t rsvd5;
+ uint8_t op_own;
+ };
+ uint32_t word;
+ } last_word;
+ uint16_t next_wqe_counter = cq->cq_ci;
+ uint16_t cur_wqe_counter;
+ uint16_t comp;
+
+ last_word.word = rte_read32(&cq->cqes[0].wqe_counter);
+ cur_wqe_counter = rte_be_to_cpu_16(last_word.wqe_counter);
+ comp = cur_wqe_counter + (uint16_t)1 - next_wqe_counter;
+ if (comp) {
+ cq->cq_ci += comp;
+ MLX5_ASSERT(MLX5_CQE_OPCODE(last_word.op_own) !=
+ MLX5_CQE_INVALID);
+ if (unlikely(!(MLX5_CQE_OPCODE(last_word.op_own) ==
+ MLX5_CQE_RESP_ERR ||
+ MLX5_CQE_OPCODE(last_word.op_own) ==
+ MLX5_CQE_REQ_ERR)))
cq->errors++;
- /*fall-through*/
- case MLX5_CQE_STATUS_SW_OWN:
- total++;
- break;
- case MLX5_CQE_STATUS_HW_OWN:
- default:
- break;
- }
- } while (ret != MLX5_CQE_STATUS_HW_OWN);
- rte_io_wmb();
- cq->cq_ci += total;
- /* Ring CQ doorbell record. */
- cq->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci);
- rte_io_wmb();
- /* Ring SW QP doorbell record. */
- eqp->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci + cq_size);
- return total;
+ rte_io_wmb();
+ /* Ring CQ doorbell record. */
+ cq->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci);
+ rte_io_wmb();
+ /* Ring SW QP doorbell record. */
+ eqp->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci + cq_size);
+ }
+ return comp;
}
static void
}
}
+static void
+mlx5_vdpa_timer_sleep(struct mlx5_vdpa_priv *priv, uint32_t max)
+{
+ if (priv->event_mode == MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER) {
+ switch (max) {
+ case 0:
+ priv->timer_delay_us += priv->event_us;
+ break;
+ case 1:
+ break;
+ default:
+ priv->timer_delay_us /= max;
+ break;
+ }
+ }
+ if (priv->timer_delay_us)
+ usleep(priv->timer_delay_us);
+}
+
static void *
mlx5_vdpa_poll_handle(void *arg)
{
struct mlx5_vdpa_priv *priv = arg;
int i;
struct mlx5_vdpa_cq *cq;
- uint32_t total;
+ uint32_t max;
uint64_t current_tic;
pthread_mutex_lock(&priv->timer_lock);
while (!priv->timer_on)
pthread_cond_wait(&priv->timer_cond, &priv->timer_lock);
pthread_mutex_unlock(&priv->timer_lock);
+ priv->timer_delay_us = priv->event_mode ==
+ MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER ?
+ MLX5_VDPA_DEFAULT_TIMER_DELAY_US :
+ priv->event_us;
while (1) {
- total = 0;
+ max = 0;
+ pthread_mutex_lock(&priv->vq_config_lock);
for (i = 0; i < priv->nr_virtqs; i++) {
cq = &priv->virtqs[i].eqp.cq;
if (cq->cq && !cq->armed) {
if (cq->callfd != -1)
eventfd_write(cq->callfd,
(eventfd_t)1);
- total += comp;
+ if (comp > max)
+ max = comp;
}
}
}
current_tic = rte_rdtsc();
- if (!total) {
+ if (!max) {
/* No traffic ? stop timer and load interrupts. */
if (current_tic - priv->last_traffic_tic >=
- rte_get_timer_hz() * MLX5_VDPA_NO_TRAFFIC_TIME_S) {
+ rte_get_timer_hz() * priv->no_traffic_time_s) {
DRV_LOG(DEBUG, "Device %s traffic was stopped.",
priv->vdev->device->name);
mlx5_vdpa_arm_all_cqs(priv);
+ pthread_mutex_unlock(&priv->vq_config_lock);
pthread_mutex_lock(&priv->timer_lock);
priv->timer_on = 0;
while (!priv->timer_on)
pthread_cond_wait(&priv->timer_cond,
&priv->timer_lock);
pthread_mutex_unlock(&priv->timer_lock);
+ priv->timer_delay_us = priv->event_mode ==
+ MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER ?
+ MLX5_VDPA_DEFAULT_TIMER_DELAY_US :
+ priv->event_us;
continue;
}
} else {
priv->last_traffic_tic = current_tic;
}
- usleep(priv->timer_delay_us);
+ pthread_mutex_unlock(&priv->vq_config_lock);
+ mlx5_vdpa_timer_sleep(priv, max);
}
return NULL;
}
uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) + 128];
} out;
+ pthread_mutex_lock(&priv->vq_config_lock);
while (mlx5_glue->devx_get_event(priv->eventc, &out.event_resp,
sizeof(out.buf)) >=
(ssize_t)sizeof(out.event_resp.cookie)) {
struct mlx5_vdpa_virtq *virtq = container_of(eqp,
struct mlx5_vdpa_virtq, eqp);
+ if (!virtq->enable)
+ continue;
mlx5_vdpa_cq_poll(cq);
+ /* Notify guest for descs consuming. */
+ if (cq->callfd != -1)
+ eventfd_write(cq->callfd, (eventfd_t)1);
+ if (priv->event_mode == MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT) {
+ mlx5_vdpa_cq_arm(priv, cq);
+ pthread_mutex_unlock(&priv->vq_config_lock);
+ return;
+ }
/* Don't arm again - timer will take control. */
DRV_LOG(DEBUG, "Device %s virtq %d cq %d event was captured."
" Timer is %s, cq ci is %u.\n",
pthread_cond_signal(&priv->timer_cond);
}
pthread_mutex_unlock(&priv->timer_lock);
+ pthread_mutex_unlock(&priv->vq_config_lock);
+}
+
+static void
+mlx5_vdpa_err_interrupt_handler(void *cb_arg __rte_unused)
+{
+#ifdef HAVE_IBV_DEVX_EVENT
+ struct mlx5_vdpa_priv *priv = cb_arg;
+ union {
+ struct mlx5dv_devx_async_event_hdr event_resp;
+ uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) + 128];
+ } out;
+ uint32_t vq_index, i, version;
+ struct mlx5_vdpa_virtq *virtq;
+ uint64_t sec;
+
+ pthread_mutex_lock(&priv->vq_config_lock);
+ while (mlx5_glue->devx_get_event(priv->err_chnl, &out.event_resp,
+ sizeof(out.buf)) >=
+ (ssize_t)sizeof(out.event_resp.cookie)) {
+ vq_index = out.event_resp.cookie & UINT32_MAX;
+ version = out.event_resp.cookie >> 32;
+ if (vq_index >= priv->nr_virtqs) {
+ DRV_LOG(ERR, "Invalid device %s error event virtq %d.",
+ priv->vdev->device->name, vq_index);
+ continue;
+ }
+ virtq = &priv->virtqs[vq_index];
+ if (!virtq->enable || virtq->version != version)
+ continue;
+ if (rte_rdtsc() / rte_get_tsc_hz() < MLX5_VDPA_ERROR_TIME_SEC)
+ continue;
+ virtq->stopped = true;
+ /* Query error info. */
+ if (mlx5_vdpa_virtq_query(priv, vq_index))
+ goto log;
+ /* Disable vq. */
+ if (mlx5_vdpa_virtq_enable(priv, vq_index, 0)) {
+ DRV_LOG(ERR, "Failed to disable virtq %d.", vq_index);
+ goto log;
+ }
+ /* Retry if error happens less than N times in 3 seconds. */
+ sec = (rte_rdtsc() - virtq->err_time[0]) / rte_get_tsc_hz();
+ if (sec > MLX5_VDPA_ERROR_TIME_SEC) {
+ /* Retry. */
+ if (mlx5_vdpa_virtq_enable(priv, vq_index, 1))
+ DRV_LOG(ERR, "Failed to enable virtq %d.",
+ vq_index);
+ else
+ DRV_LOG(WARNING, "Recover virtq %d: %u.",
+ vq_index, ++virtq->n_retry);
+ } else {
+ /* Retry timeout, give up. */
+ DRV_LOG(ERR, "Device %s virtq %d failed to recover.",
+ priv->vdev->device->name, vq_index);
+ }
+log:
+ /* Shift in current time to error time log end. */
+ for (i = 1; i < RTE_DIM(virtq->err_time); i++)
+ virtq->err_time[i - 1] = virtq->err_time[i];
+ virtq->err_time[RTE_DIM(virtq->err_time) - 1] = rte_rdtsc();
+ }
+ pthread_mutex_unlock(&priv->vq_config_lock);
+#endif
}
int
-mlx5_vdpa_cqe_event_setup(struct mlx5_vdpa_priv *priv)
+mlx5_vdpa_err_event_setup(struct mlx5_vdpa_priv *priv)
{
+ int ret;
int flags;
+
+ /* Setup device event channel. */
+ priv->err_chnl = mlx5_glue->devx_create_event_channel(priv->ctx, 0);
+ if (!priv->err_chnl) {
+ rte_errno = errno;
+ DRV_LOG(ERR, "Failed to create device event channel %d.",
+ rte_errno);
+ goto error;
+ }
+ flags = fcntl(priv->err_chnl->fd, F_GETFL);
+ ret = fcntl(priv->err_chnl->fd, F_SETFL, flags | O_NONBLOCK);
+ if (ret) {
+ DRV_LOG(ERR, "Failed to change device event channel FD.");
+ goto error;
+ }
+ priv->err_intr_handle.fd = priv->err_chnl->fd;
+ priv->err_intr_handle.type = RTE_INTR_HANDLE_EXT;
+ if (rte_intr_callback_register(&priv->err_intr_handle,
+ mlx5_vdpa_err_interrupt_handler,
+ priv)) {
+ priv->err_intr_handle.fd = 0;
+ DRV_LOG(ERR, "Failed to register error interrupt for device %d.",
+ priv->vid);
+ goto error;
+ } else {
+ DRV_LOG(DEBUG, "Registered error interrupt for device%d.",
+ priv->vid);
+ }
+ return 0;
+error:
+ mlx5_vdpa_err_event_unset(priv);
+ return -1;
+}
+
+void
+mlx5_vdpa_err_event_unset(struct mlx5_vdpa_priv *priv)
+{
+ int retries = MLX5_VDPA_INTR_RETRIES;
+ int ret = -EAGAIN;
+
+ if (!priv->err_intr_handle.fd)
+ return;
+ while (retries-- && ret == -EAGAIN) {
+ ret = rte_intr_callback_unregister(&priv->err_intr_handle,
+ mlx5_vdpa_err_interrupt_handler,
+ priv);
+ if (ret == -EAGAIN) {
+ DRV_LOG(DEBUG, "Try again to unregister fd %d "
+ "of error interrupt, retries = %d.",
+ priv->err_intr_handle.fd, retries);
+ rte_pause();
+ }
+ }
+ memset(&priv->err_intr_handle, 0, sizeof(priv->err_intr_handle));
+ if (priv->err_chnl) {
+#ifdef HAVE_IBV_DEVX_EVENT
+ union {
+ struct mlx5dv_devx_async_event_hdr event_resp;
+ uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) +
+ 128];
+ } out;
+
+ /* Clean all pending events. */
+ while (mlx5_glue->devx_get_event(priv->err_chnl,
+ &out.event_resp, sizeof(out.buf)) >=
+ (ssize_t)sizeof(out.event_resp.cookie))
+ ;
+#endif
+ mlx5_glue->devx_destroy_event_channel(priv->err_chnl);
+ priv->err_chnl = NULL;
+ }
+}
+
+int
+mlx5_vdpa_cqe_event_setup(struct mlx5_vdpa_priv *priv)
+{
int ret;
+ rte_cpuset_t cpuset;
+ pthread_attr_t attr;
+ char name[16];
if (!priv->eventc)
/* All virtqs are in poll mode. */
return 0;
- pthread_mutex_init(&priv->timer_lock, NULL);
- pthread_cond_init(&priv->timer_cond, NULL);
- priv->timer_on = 0;
- priv->timer_delay_us = MLX5_VDPA_DEFAULT_TIMER_DELAY_US;
- ret = pthread_create(&priv->timer_tid, NULL, mlx5_vdpa_poll_handle,
- (void *)priv);
- if (ret) {
- DRV_LOG(ERR, "Failed to create timer thread.");
- return -1;
- }
- flags = fcntl(priv->eventc->fd, F_GETFL);
- ret = fcntl(priv->eventc->fd, F_SETFL, flags | O_NONBLOCK);
- if (ret) {
- DRV_LOG(ERR, "Failed to change event channel FD.");
- goto error;
+ if (priv->event_mode != MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT) {
+ pthread_mutex_init(&priv->timer_lock, NULL);
+ pthread_cond_init(&priv->timer_cond, NULL);
+ priv->timer_on = 0;
+ pthread_attr_init(&attr);
+ CPU_ZERO(&cpuset);
+ if (priv->event_core != -1)
+ CPU_SET(priv->event_core, &cpuset);
+ else
+ cpuset = rte_lcore_cpuset(rte_get_main_lcore());
+ ret = pthread_attr_setaffinity_np(&attr, sizeof(cpuset),
+ &cpuset);
+ if (ret) {
+ DRV_LOG(ERR, "Failed to set thread affinity.");
+ return -1;
+ }
+ ret = pthread_create(&priv->timer_tid, &attr,
+ mlx5_vdpa_poll_handle, (void *)priv);
+ if (ret) {
+ DRV_LOG(ERR, "Failed to create timer thread.");
+ return -1;
+ }
+ snprintf(name, sizeof(name), "vDPA-mlx5-%d", priv->vid);
+ ret = pthread_setname_np(priv->timer_tid, name);
+ if (ret) {
+ DRV_LOG(ERR, "Failed to set timer thread name.");
+ return -1;
+ }
}
priv->intr_handle.fd = priv->eventc->fd;
priv->intr_handle.type = RTE_INTR_HANDLE_EXT;