@@ -2,6 +2,8 @@
* Copyright (c) 2023 Marvell.
*/
+#include <rte_service_component.h>
+
#include "rte_eventdev.h"
#include "eventdev_pmd.h"
#include "rte_event_dma_adapter.h"
@@ -69,6 +71,10 @@ struct dma_device_info {
/* Number of vchans configured for a DMA device. */
uint16_t num_dma_dev_vchan;
+
+ /* Next queue pair to be processed */
+ uint16_t next_vchan_id;
+
} __rte_cache_aligned;
struct event_dma_adapter {
@@ -90,6 +96,9 @@ struct event_dma_adapter {
/* Lock to serialize config updates with service function */
rte_spinlock_t lock;
+ /* Next dma device to be processed */
+ uint16_t next_dmadev_id;
+
/* DMA device structure array */
struct dma_device_info *dma_devs;
@@ -107,6 +116,26 @@ struct event_dma_adapter {
/* No. of vchan queue configured */
uint16_t nb_vchanq;
+
+ /* Per adapter EAL service ID */
+ uint32_t service_id;
+
+ /* Service initialization state */
+ uint8_t service_initialized;
+
+ /* Max DMA ops processed in any service function invocation */
+ uint32_t max_nb;
+
+ /* Store event port's implicit release capability */
+ uint8_t implicit_release_disabled;
+
+ /* Flag to indicate backpressure at dma_dev
+ * Stop further dequeuing events from eventdev
+ */
+ bool stop_enq_to_dma_dev;
+
+ /* Loop counter to flush dma ops */
+ uint16_t transmit_loop_count;
} __rte_cache_aligned;
static struct event_dma_adapter **event_dma_adapter;
@@ -148,6 +177,18 @@ edma_array_init(void)
return 0;
}
+static inline bool
+edma_circular_buffer_batch_ready(struct dma_ops_circular_buffer *bufp)
+{
+ return bufp->count >= DMA_BATCH_SIZE;
+}
+
+static inline bool
+edma_circular_buffer_space_for_batch(struct dma_ops_circular_buffer *bufp)
+{
+ return (bufp->size - bufp->count) >= DMA_BATCH_SIZE;
+}
+
static inline int
edma_circular_buffer_init(const char *name, struct dma_ops_circular_buffer *buf, uint16_t sz)
{
@@ -166,6 +207,67 @@ edma_circular_buffer_free(struct dma_ops_circular_buffer *buf)
rte_free(buf->op_buffer);
}
+static inline int
+edma_circular_buffer_add(struct dma_ops_circular_buffer *bufp, struct rte_event_dma_adapter_op *op)
+{
+ uint16_t *tail = &bufp->tail;
+
+ bufp->op_buffer[*tail] = op;
+
+ /* circular buffer, go round */
+ *tail = (*tail + 1) % bufp->size;
+ bufp->count++;
+
+ return 0;
+}
+
+static inline int
+edma_circular_buffer_flush_to_dma_dev(struct event_dma_adapter *adapter,
+ struct dma_ops_circular_buffer *bufp, uint8_t dma_dev_id,
+ uint16_t vchan, uint16_t *nb_ops_flushed)
+{
+ struct rte_event_dma_adapter_op *op;
+ struct dma_vchan_info *tq;
+ uint16_t *head = &bufp->head;
+ uint16_t *tail = &bufp->tail;
+ uint16_t n;
+ uint16_t i;
+ int ret;
+
+ if (*tail > *head)
+ n = *tail - *head;
+ else if (*tail < *head)
+ n = bufp->size - *head;
+ else {
+ *nb_ops_flushed = 0;
+ return 0; /* buffer empty */
+ }
+
+ tq = &adapter->dma_devs[dma_dev_id].tqmap[vchan];
+
+ for (i = 0; i < n; i++) {
+ op = bufp->op_buffer[*head];
+ ret = rte_dma_copy_sg(dma_dev_id, vchan, op->src_seg, op->dst_seg,
+ op->nb_src, op->nb_dst, op->flags);
+ if (ret < 0)
+ break;
+
+ /* Enqueue in transaction queue. */
+ edma_circular_buffer_add(&tq->dma_buf, op);
+
+ *head = (*head + 1) % bufp->size;
+ }
+
+ *nb_ops_flushed = i;
+ bufp->count -= *nb_ops_flushed;
+ if (!bufp->count) {
+ *head = 0;
+ *tail = 0;
+ }
+
+ return *nb_ops_flushed == n ? 0 : -1;
+}
+
static int
edma_default_config_cb(uint8_t id, uint8_t evdev_id, struct rte_event_dma_adapter_conf *conf,
void *arg)
@@ -360,6 +462,406 @@ rte_event_dma_adapter_free(uint8_t id)
return 0;
}
+static inline unsigned int
+edma_enq_to_dma_dev(struct event_dma_adapter *adapter, struct rte_event *ev, unsigned int cnt)
+{
+ struct dma_vchan_info *vchan_qinfo = NULL;
+ struct rte_event_dma_adapter_op *dma_op;
+ uint16_t vchan, nb_enqueued = 0;
+ int16_t dma_dev_id;
+ unsigned int i, n;
+ int ret;
+
+ ret = 0;
+ n = 0;
+
+ for (i = 0; i < cnt; i++) {
+ dma_op = ev[i].event_ptr;
+ if (dma_op == NULL)
+ continue;
+
+ /* Expected to have reponse info appended to dma_op. */
+
+ dma_dev_id = dma_op->dma_dev_id;
+ vchan = dma_op->vchan;
+ vchan_qinfo = &adapter->dma_devs[dma_dev_id].vchanq[vchan];
+ if (!vchan_qinfo->vq_enabled) {
+ if (dma_op != NULL && dma_op->op_mp != NULL)
+ rte_mempool_put(dma_op->op_mp, dma_op);
+ continue;
+ }
+ edma_circular_buffer_add(&vchan_qinfo->dma_buf, dma_op);
+
+ if (edma_circular_buffer_batch_ready(&vchan_qinfo->dma_buf)) {
+ ret = edma_circular_buffer_flush_to_dma_dev(adapter, &vchan_qinfo->dma_buf,
+ dma_dev_id, vchan,
+ &nb_enqueued);
+ n += nb_enqueued;
+
+ /**
+ * If some dma ops failed to flush to dma_dev and
+ * space for another batch is not available, stop
+ * dequeue from eventdev momentarily
+ */
+ if (unlikely(ret < 0 &&
+ !edma_circular_buffer_space_for_batch(&vchan_qinfo->dma_buf)))
+ adapter->stop_enq_to_dma_dev = true;
+ }
+ }
+
+ return n;
+}
+
+static unsigned int
+edma_adapter_dev_flush(struct event_dma_adapter *adapter, int16_t dma_dev_id,
+ uint16_t *nb_ops_flushed)
+{
+ struct dma_vchan_info *vchan_info;
+ struct dma_device_info *dev_info;
+ uint16_t nb = 0, nb_enqueued = 0;
+ uint16_t vchan, nb_vchans;
+
+ dev_info = &adapter->dma_devs[dma_dev_id];
+ nb_vchans = dev_info->num_vchanq;
+
+ for (vchan = 0; vchan < nb_vchans; vchan++) {
+
+ vchan_info = &dev_info->vchanq[vchan];
+ if (unlikely(vchan_info == NULL || !vchan_info->vq_enabled))
+ continue;
+
+ edma_circular_buffer_flush_to_dma_dev(adapter, &vchan_info->dma_buf, dma_dev_id,
+ vchan, &nb_enqueued);
+ *nb_ops_flushed += vchan_info->dma_buf.count;
+ nb += nb_enqueued;
+ }
+
+ return nb;
+}
+
+static unsigned int
+edma_adapter_enq_flush(struct event_dma_adapter *adapter)
+{
+ int16_t dma_dev_id;
+ uint16_t nb_enqueued = 0;
+ uint16_t nb_ops_flushed = 0;
+ uint16_t num_dma_dev = rte_dma_count_avail();
+
+ for (dma_dev_id = 0; dma_dev_id < num_dma_dev; dma_dev_id++)
+ nb_enqueued += edma_adapter_dev_flush(adapter, dma_dev_id, &nb_ops_flushed);
+ /**
+ * Enable dequeue from eventdev if all ops from circular
+ * buffer flushed to dma_dev
+ */
+ if (!nb_ops_flushed)
+ adapter->stop_enq_to_dma_dev = false;
+
+ return nb_enqueued;
+}
+
+/* Flush an instance's enqueue buffers every DMA_ENQ_FLUSH_THRESHOLD
+ * iterations of edma_adapter_enq_run()
+ */
+#define DMA_ENQ_FLUSH_THRESHOLD 1024
+
+static int
+edma_adapter_enq_run(struct event_dma_adapter *adapter, unsigned int max_enq)
+{
+ uint8_t event_port_id = adapter->event_port_id;
+ uint8_t event_dev_id = adapter->eventdev_id;
+ struct rte_event ev[DMA_BATCH_SIZE];
+ unsigned int nb_enq, nb_enqueued;
+ uint16_t n;
+
+ if (adapter->mode == RTE_EVENT_DMA_ADAPTER_OP_NEW)
+ return 0;
+
+ nb_enqueued = 0;
+ for (nb_enq = 0; nb_enq < max_enq; nb_enq += n) {
+
+ if (unlikely(adapter->stop_enq_to_dma_dev)) {
+ nb_enqueued += edma_adapter_enq_flush(adapter);
+
+ if (unlikely(adapter->stop_enq_to_dma_dev))
+ break;
+ }
+
+ n = rte_event_dequeue_burst(event_dev_id, event_port_id, ev, DMA_BATCH_SIZE, 0);
+
+ if (!n)
+ break;
+
+ nb_enqueued += edma_enq_to_dma_dev(adapter, ev, n);
+ }
+
+ if ((++adapter->transmit_loop_count & (DMA_ENQ_FLUSH_THRESHOLD - 1)) == 0)
+ nb_enqueued += edma_adapter_enq_flush(adapter);
+
+ return nb_enqueued;
+}
+
+#define DMA_ADAPTER_MAX_EV_ENQ_RETRIES 100
+
+static inline uint16_t
+edma_ops_enqueue_burst(struct event_dma_adapter *adapter, struct rte_event_dma_adapter_op **ops,
+ uint16_t num)
+{
+ uint8_t event_port_id = adapter->event_port_id;
+ uint8_t event_dev_id = adapter->eventdev_id;
+ struct rte_event events[DMA_BATCH_SIZE];
+ struct rte_event *response_info;
+ uint16_t nb_enqueued, nb_ev;
+ uint8_t retry;
+ uint8_t i;
+
+ nb_ev = 0;
+ retry = 0;
+ nb_enqueued = 0;
+ num = RTE_MIN(num, DMA_BATCH_SIZE);
+ for (i = 0; i < num; i++) {
+ struct rte_event *ev = &events[nb_ev++];
+
+ /* Expected to have response info appended to dma_op. */
+ response_info = (struct rte_event *)((uint8_t *)ops[i] +
+ sizeof(struct rte_event_dma_adapter_op));
+ if (unlikely(response_info == NULL)) {
+ if (ops[i] != NULL && ops[i]->op_mp != NULL)
+ rte_mempool_put(ops[i]->op_mp, ops[i]);
+ continue;
+ }
+
+ rte_memcpy(ev, response_info, sizeof(struct rte_event));
+ ev->event_ptr = ops[i];
+ ev->event_type = RTE_EVENT_TYPE_DMADEV;
+ if (adapter->implicit_release_disabled)
+ ev->op = RTE_EVENT_OP_FORWARD;
+ else
+ ev->op = RTE_EVENT_OP_NEW;
+ }
+
+ do {
+ nb_enqueued += rte_event_enqueue_burst(event_dev_id, event_port_id,
+ &events[nb_enqueued], nb_ev - nb_enqueued);
+
+ } while (retry++ < DMA_ADAPTER_MAX_EV_ENQ_RETRIES && nb_enqueued < nb_ev);
+
+ return nb_enqueued;
+}
+
+static int
+edma_circular_buffer_flush_to_evdev(struct event_dma_adapter *adapter,
+ struct dma_ops_circular_buffer *bufp,
+ uint16_t *enqueue_count)
+{
+ struct rte_event_dma_adapter_op **ops = bufp->op_buffer;
+ uint16_t n = 0, nb_ops_flushed;
+ uint16_t *head = &bufp->head;
+ uint16_t *tail = &bufp->tail;
+
+ if (*tail > *head)
+ n = *tail - *head;
+ else if (*tail < *head)
+ n = bufp->size - *head;
+ else {
+ if (enqueue_count)
+ *enqueue_count = 0;
+ return 0; /* buffer empty */
+ }
+
+ if (enqueue_count && n > *enqueue_count)
+ n = *enqueue_count;
+
+ nb_ops_flushed = edma_ops_enqueue_burst(adapter, &ops[*head], n);
+ if (enqueue_count)
+ *enqueue_count = nb_ops_flushed;
+
+ bufp->count -= nb_ops_flushed;
+ if (!bufp->count) {
+ *head = 0;
+ *tail = 0;
+ return 0; /* buffer empty */
+ }
+
+ *head = (*head + nb_ops_flushed) % bufp->size;
+ return 1;
+}
+
+static void
+edma_ops_buffer_flush(struct event_dma_adapter *adapter)
+{
+ if (likely(adapter->ebuf.count == 0))
+ return;
+
+ while (edma_circular_buffer_flush_to_evdev(adapter, &adapter->ebuf, NULL))
+ ;
+}
+
+static inline unsigned int
+edma_adapter_deq_run(struct event_dma_adapter *adapter, unsigned int max_deq)
+{
+ struct dma_vchan_info *vchan_info;
+ struct dma_ops_circular_buffer *tq_buf;
+ struct rte_event_dma_adapter_op *ops;
+ uint16_t n, nb_deq, nb_enqueued, i;
+ struct dma_device_info *dev_info;
+ uint16_t vchan, num_vchan;
+ uint16_t num_dma_dev;
+ int16_t dma_dev_id;
+ uint16_t index;
+ bool done;
+ bool err;
+
+ nb_deq = 0;
+ edma_ops_buffer_flush(adapter);
+
+ num_dma_dev = rte_dma_count_avail();
+ do {
+ done = true;
+
+ for (dma_dev_id = adapter->next_dmadev_id; dma_dev_id < num_dma_dev; dma_dev_id++) {
+ uint16_t queues = 0;
+ dev_info = &adapter->dma_devs[dma_dev_id];
+ num_vchan = dev_info->num_vchanq;
+
+ for (vchan = dev_info->next_vchan_id; queues < num_vchan;
+ vchan = (vchan + 1) % num_vchan, queues++) {
+
+ vchan_info = &dev_info->vchanq[vchan];
+ if (unlikely(vchan_info == NULL || !vchan_info->vq_enabled))
+ continue;
+
+ n = rte_dma_completed(dma_dev_id, vchan, DMA_BATCH_SIZE,
+ &index, &err);
+ if (!n)
+ continue;
+
+ done = false;
+
+ tq_buf = &dev_info->tqmap[vchan].dma_buf;
+
+ nb_enqueued = n;
+ if (unlikely(!adapter->ebuf.count))
+ edma_circular_buffer_flush_to_evdev(adapter, tq_buf,
+ &nb_enqueued);
+
+ if (likely(nb_enqueued == n))
+ goto check;
+
+ /* Failed to enqueue events case */
+ for (i = nb_enqueued; i < n; i++) {
+ ops = tq_buf->op_buffer[tq_buf->head];
+ edma_circular_buffer_add(&adapter->ebuf, ops);
+ tq_buf->head = (tq_buf->head + 1) % tq_buf->size;
+ }
+
+check:
+ nb_deq += n;
+ if (nb_deq >= max_deq) {
+ if ((vchan + 1) == num_vchan)
+ adapter->next_dmadev_id =
+ (dma_dev_id + 1) % num_dma_dev;
+
+ dev_info->next_vchan_id = (vchan + 1) % num_vchan;
+
+ return nb_deq;
+ }
+ }
+ }
+ adapter->next_dmadev_id = 0;
+
+ } while (done == false);
+
+ return nb_deq;
+}
+
+static int
+edma_adapter_run(struct event_dma_adapter *adapter, unsigned int max_ops)
+{
+ unsigned int ops_left = max_ops;
+
+ while (ops_left > 0) {
+ unsigned int e_cnt, d_cnt;
+
+ e_cnt = edma_adapter_deq_run(adapter, ops_left);
+ ops_left -= RTE_MIN(ops_left, e_cnt);
+
+ d_cnt = edma_adapter_enq_run(adapter, ops_left);
+ ops_left -= RTE_MIN(ops_left, d_cnt);
+
+ if (e_cnt == 0 && d_cnt == 0)
+ break;
+ }
+
+ if (ops_left == max_ops) {
+ rte_event_maintain(adapter->eventdev_id, adapter->event_port_id, 0);
+ return -EAGAIN;
+ } else
+ return 0;
+}
+
+static int
+edma_service_func(void *args)
+{
+ struct event_dma_adapter *adapter = args;
+ int ret;
+
+ if (rte_spinlock_trylock(&adapter->lock) == 0)
+ return 0;
+ ret = edma_adapter_run(adapter, adapter->max_nb);
+ rte_spinlock_unlock(&adapter->lock);
+
+ return ret;
+}
+
+static int
+edma_init_service(struct event_dma_adapter *adapter, uint8_t id)
+{
+ struct rte_event_dma_adapter_conf adapter_conf;
+ struct rte_service_spec service;
+ uint32_t impl_rel;
+ int ret;
+
+ if (adapter->service_initialized)
+ return 0;
+
+ memset(&service, 0, sizeof(service));
+ snprintf(service.name, DMA_ADAPTER_NAME_LEN, "rte_event_dma_adapter_%d", id);
+ service.socket_id = adapter->socket_id;
+ service.callback = edma_service_func;
+ service.callback_userdata = adapter;
+
+ /* Service function handles locking for queue add/del updates */
+ service.capabilities = RTE_SERVICE_CAP_MT_SAFE;
+ ret = rte_service_component_register(&service, &adapter->service_id);
+ if (ret) {
+ RTE_EDEV_LOG_ERR("failed to register service %s err = %" PRId32, service.name, ret);
+ return ret;
+ }
+
+ ret = adapter->conf_cb(id, adapter->eventdev_id, &adapter_conf, adapter->conf_arg);
+ if (ret) {
+ RTE_EDEV_LOG_ERR("configuration callback failed err = %" PRId32, ret);
+ return ret;
+ }
+
+ adapter->max_nb = adapter_conf.max_nb;
+ adapter->event_port_id = adapter_conf.event_port_id;
+
+ if (rte_event_port_attr_get(adapter->eventdev_id, adapter->event_port_id,
+ RTE_EVENT_PORT_ATTR_IMPLICIT_RELEASE_DISABLE, &impl_rel)) {
+ RTE_EDEV_LOG_ERR("Failed to get port info for eventdev %" PRId32,
+ adapter->eventdev_id);
+ edma_circular_buffer_free(&adapter->ebuf);
+ rte_free(adapter);
+ return -EINVAL;
+ }
+
+ adapter->implicit_release_disabled = (uint8_t)impl_rel;
+ adapter->service_initialized = 1;
+
+ return ret;
+}
+
static void
edma_update_vchanq_info(struct event_dma_adapter *adapter, struct dma_device_info *dev_info,
uint16_t vchan, uint8_t add)
@@ -391,6 +893,60 @@ edma_update_vchanq_info(struct event_dma_adapter *adapter, struct dma_device_inf
}
}
+static int
+edma_add_vchan(struct event_dma_adapter *adapter, int16_t dma_dev_id, uint16_t vchan)
+{
+ struct dma_device_info *dev_info = &adapter->dma_devs[dma_dev_id];
+ struct dma_vchan_info *vchanq;
+ struct dma_vchan_info *tqmap;
+ uint16_t nb_vchans;
+ uint32_t i;
+
+ if (dev_info->vchanq == NULL) {
+ nb_vchans = dev_info->num_dma_dev_vchan;
+
+ dev_info->vchanq = rte_zmalloc_socket(adapter->mem_name,
+ nb_vchans * sizeof(struct dma_vchan_info),
+ 0, adapter->socket_id);
+ if (dev_info->vchanq == NULL)
+ return -ENOMEM;
+
+ dev_info->tqmap = rte_zmalloc_socket(adapter->mem_name,
+ nb_vchans * sizeof(struct dma_vchan_info),
+ 0, adapter->socket_id);
+ if (dev_info->tqmap == NULL)
+ return -ENOMEM;
+
+ for (i = 0; i < nb_vchans; i++) {
+ vchanq = &dev_info->vchanq[i];
+
+ if (edma_circular_buffer_init("dma_dev_circular_buffer", &vchanq->dma_buf,
+ DMA_ADAPTER_OPS_BUFFER_SIZE)) {
+ RTE_EDEV_LOG_ERR("Failed to get memory for dma_dev buffer");
+ rte_free(vchanq);
+ return -ENOMEM;
+ }
+
+ tqmap = &dev_info->tqmap[i];
+ if (edma_circular_buffer_init("dma_dev_circular_trans_buf", &tqmap->dma_buf,
+ DMA_ADAPTER_OPS_BUFFER_SIZE)) {
+ RTE_EDEV_LOG_ERR(
+ "Failed to get memory for dma_dev transction buffer");
+ rte_free(tqmap);
+ return -ENOMEM;
+ }
+ }
+ }
+
+ if (vchan == RTE_DMA_ALL_VCHAN) {
+ for (i = 0; i < dev_info->num_dma_dev_vchan; i++)
+ edma_update_vchanq_info(adapter, dev_info, i, 1);
+ } else
+ edma_update_vchanq_info(adapter, dev_info, vchan, 1);
+
+ return 0;
+}
+
int
rte_event_dma_adapter_vchan_add(uint8_t id, int16_t dma_dev_id, uint16_t vchan,
const struct rte_event *event)
@@ -470,6 +1026,38 @@ rte_event_dma_adapter_vchan_add(uint8_t id, int16_t dma_dev_id, uint16_t vchan,
edma_update_vchanq_info(adapter, &adapter->dma_devs[dma_dev_id], vchan, 1);
}
+ /* In case HW cap is RTE_EVENT_DMA_ADAPTER_CAP_INTERNAL_PORT_OP_NEW, or SW adapter, initiate
+ * services so the application can choose which ever way it wants to use the adapter.
+ *
+ * Case 1: RTE_EVENT_DMA_ADAPTER_CAP_INTERNAL_PORT_OP_NEW. Application may wants to use one
+ * of below two modes
+ *
+ * a. OP_FORWARD mode -> HW Dequeue + SW enqueue
+ * b. OP_NEW mode -> HW Dequeue
+ *
+ * Case 2: No HW caps, use SW adapter
+ *
+ * a. OP_FORWARD mode -> SW enqueue & dequeue
+ * b. OP_NEW mode -> SW Dequeue
+ */
+ if ((cap & RTE_EVENT_DMA_ADAPTER_CAP_INTERNAL_PORT_OP_NEW &&
+ !(cap & RTE_EVENT_DMA_ADAPTER_CAP_INTERNAL_PORT_OP_FWD) &&
+ adapter->mode == RTE_EVENT_DMA_ADAPTER_OP_FORWARD) ||
+ (!(cap & RTE_EVENT_DMA_ADAPTER_CAP_INTERNAL_PORT_OP_NEW) &&
+ !(cap & RTE_EVENT_DMA_ADAPTER_CAP_INTERNAL_PORT_OP_FWD) &&
+ !(cap & RTE_EVENT_DMA_ADAPTER_CAP_INTERNAL_PORT_VCHAN_EV_BIND))) {
+ rte_spinlock_lock(&adapter->lock);
+ ret = edma_init_service(adapter, id);
+ if (ret == 0)
+ ret = edma_add_vchan(adapter, dma_dev_id, vchan);
+ rte_spinlock_unlock(&adapter->lock);
+
+ if (ret)
+ return ret;
+
+ rte_service_component_runstate_set(adapter->service_id, 1);
+ }
+
return 0;
}
@@ -533,6 +1121,7 @@ rte_event_dma_adapter_vchan_del(uint8_t id, int16_t dma_dev_id, uint16_t vchan)
}
rte_spinlock_unlock(&adapter->lock);
+ rte_service_component_runstate_set(adapter->service_id, adapter->nb_vchanq);
}
return ret;