[1/2] eventdev/dma: reorganize event DMA ops

Message ID 20240406101311.11044-1-pbhagavatula@marvell.com (mailing list archive)
State Superseded
Delegated to: Jerin Jacob
Headers
Series [1/2] eventdev/dma: reorganize event DMA ops |

Checks

Context Check Description
ci/checkpatch success coding style OK

Commit Message

Pavan Nikhilesh Bhagavatula April 6, 2024, 10:13 a.m. UTC
  From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Re-organize event DMA ops structure to allow holding
source and destination pointers without the need for
additional memory, the mempool allocating memory for
rte_event_dma_adapter_ops can size the structure to
accommodate all the needed source and destination
pointers.

Add multiple words for holding user metadata, adapter
implementation specific metadata and event metadata.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
 app/test-eventdev/test_perf_common.c        | 26 ++++--------
 app/test/test_event_dma_adapter.c           | 20 +++------
 doc/guides/prog_guide/event_dma_adapter.rst |  2 +-
 drivers/dma/cnxk/cnxk_dmadev_fp.c           | 39 +++++++----------
 lib/eventdev/rte_event_dma_adapter.c        | 27 ++++--------
 lib/eventdev/rte_event_dma_adapter.h        | 46 +++++++++++++++------
 6 files changed, 72 insertions(+), 88 deletions(-)
  

Comments

Amit Prakash Shukla May 16, 2024, 7:36 a.m. UTC | #1
> -----Original Message-----
> From: pbhagavatula@marvell.com <pbhagavatula@marvell.com>
> Sent: Saturday, April 6, 2024 3:43 PM
> To: Jerin Jacob <jerinj@marvell.com>; Amit Prakash Shukla
> <amitprakashs@marvell.com>; Vamsi Krishna Attunuru
> <vattunuru@marvell.com>
> Cc: dev@dpdk.org; Pavan Nikhilesh Bhagavatula
> <pbhagavatula@marvell.com>
> Subject: [PATCH 1/2] eventdev/dma: reorganize event DMA ops
> 
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> 
> Re-organize event DMA ops structure to allow holding source and destination
> pointers without the need for additional memory, the mempool allocating
> memory for rte_event_dma_adapter_ops can size the structure to
> accommodate all the needed source and destination pointers.
> 
> Add multiple words for holding user metadata, adapter implementation
> specific metadata and event metadata.
> 
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>

Acked-by: Amit Prakash Shukla <amitprakashs@marvell.com>
  

Patch

diff --git a/app/test-eventdev/test_perf_common.c b/app/test-eventdev/test_perf_common.c
index 93e6132de8de..00fcac716304 100644
--- a/app/test-eventdev/test_perf_common.c
+++ b/app/test-eventdev/test_perf_common.c
@@ -1503,7 +1503,6 @@  perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt,
 		prod = 0;
 		for (; port < perf_nb_event_ports(opt); port++) {
 			struct prod_data *p = &t->prod[port];
-			struct rte_event *response_info;
 			uint32_t flow_id;
 
 			p->dev_id = opt->dev_id;
@@ -1523,13 +1522,10 @@  perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt,
 			for (flow_id = 0; flow_id < t->nb_flows; flow_id++) {
 				rte_mempool_get(t->da_op_pool, (void **)&op);
 
-				op->src_seg = rte_malloc(NULL, sizeof(struct rte_dma_sge), 0);
-				op->dst_seg = rte_malloc(NULL, sizeof(struct rte_dma_sge), 0);
-
-				op->src_seg->addr = rte_pktmbuf_iova(rte_pktmbuf_alloc(pool));
-				op->dst_seg->addr = rte_pktmbuf_iova(rte_pktmbuf_alloc(pool));
-				op->src_seg->length = 1024;
-				op->dst_seg->length = 1024;
+				op->src_dst_seg[0].addr = rte_pktmbuf_iova(rte_pktmbuf_alloc(pool));
+				op->src_dst_seg[1].addr = rte_pktmbuf_iova(rte_pktmbuf_alloc(pool));
+				op->src_dst_seg[0].length = 1024;
+				op->src_dst_seg[1].length = 1024;
 				op->nb_src = 1;
 				op->nb_dst = 1;
 				op->flags = RTE_DMA_OP_FLAG_SUBMIT;
@@ -1537,12 +1533,6 @@  perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt,
 				op->dma_dev_id = dma_dev_id;
 				op->vchan = vchan_id;
 
-				response_info = (struct rte_event *)((uint8_t *)op +
-						 sizeof(struct rte_event_dma_adapter_op));
-				response_info->queue_id = p->queue_id;
-				response_info->sched_type = RTE_SCHED_TYPE_ATOMIC;
-				response_info->flow_id = flow_id;
-
 				p->da.dma_op[flow_id] = op;
 			}
 
@@ -2036,7 +2026,7 @@  perf_dmadev_setup(struct evt_test *test, struct evt_options *opt)
 		return -ENODEV;
 	}
 
-	elt_size = sizeof(struct rte_event_dma_adapter_op) + sizeof(struct rte_event);
+	elt_size = sizeof(struct rte_event_dma_adapter_op) + (sizeof(struct rte_dma_sge) * 2);
 	t->da_op_pool = rte_mempool_create("dma_op_pool", opt->pool_sz, elt_size, 256,
 					   0, NULL, NULL, NULL, NULL, rte_socket_id(), 0);
 	if (t->da_op_pool == NULL) {
@@ -2085,10 +2075,8 @@  perf_dmadev_destroy(struct evt_test *test, struct evt_options *opt)
 		for (flow_id = 0; flow_id < t->nb_flows; flow_id++) {
 			op = p->da.dma_op[flow_id];
 
-			rte_pktmbuf_free((struct rte_mbuf *)(uintptr_t)op->src_seg->addr);
-			rte_pktmbuf_free((struct rte_mbuf *)(uintptr_t)op->dst_seg->addr);
-			rte_free(op->src_seg);
-			rte_free(op->dst_seg);
+			rte_pktmbuf_free((struct rte_mbuf *)op->src_dst_seg[0].addr);
+			rte_pktmbuf_free((struct rte_mbuf *)op->src_dst_seg[1].addr);
 			rte_mempool_put(op->op_mp, op);
 		}
 
diff --git a/app/test/test_event_dma_adapter.c b/app/test/test_event_dma_adapter.c
index 35b417b69f7b..d9dff4ff7d3f 100644
--- a/app/test/test_event_dma_adapter.c
+++ b/app/test/test_event_dma_adapter.c
@@ -235,7 +235,6 @@  test_op_forward_mode(void)
 	struct rte_mbuf *dst_mbuf[TEST_MAX_OP];
 	struct rte_event_dma_adapter_op *op;
 	struct rte_event ev[TEST_MAX_OP];
-	struct rte_event response_info;
 	int ret, i;
 
 	ret = rte_pktmbuf_alloc_bulk(params.src_mbuf_pool, src_mbuf, TEST_MAX_OP);
@@ -253,14 +252,11 @@  test_op_forward_mode(void)
 		rte_mempool_get(params.op_mpool, (void **)&op);
 		TEST_ASSERT_NOT_NULL(op, "Failed to allocate dma operation struct\n");
 
-		op->src_seg = rte_malloc(NULL, sizeof(struct rte_dma_sge), 0);
-		op->dst_seg = rte_malloc(NULL, sizeof(struct rte_dma_sge), 0);
-
 		/* Update Op */
-		op->src_seg->addr = rte_pktmbuf_iova(src_mbuf[i]);
-		op->dst_seg->addr = rte_pktmbuf_iova(dst_mbuf[i]);
-		op->src_seg->length = PACKET_LENGTH;
-		op->dst_seg->length = PACKET_LENGTH;
+		op->src_dst_seg[0].addr = rte_pktmbuf_iova(src_mbuf[i]);
+		op->src_dst_seg[1].addr = rte_pktmbuf_iova(dst_mbuf[i]);
+		op->src_dst_seg[0].length = PACKET_LENGTH;
+		op->src_dst_seg[1].length = PACKET_LENGTH;
 		op->nb_src = 1;
 		op->nb_dst = 1;
 		op->flags = RTE_DMA_OP_FLAG_SUBMIT;
@@ -268,10 +264,6 @@  test_op_forward_mode(void)
 		op->dma_dev_id = TEST_DMA_DEV_ID;
 		op->vchan = TEST_DMA_VCHAN_ID;
 
-		response_info.event = dma_response_info.event;
-		rte_memcpy((uint8_t *)op + sizeof(struct rte_event_dma_adapter_op), &response_info,
-			   sizeof(struct rte_event));
-
 		/* Fill in event info and update event_ptr with rte_event_dma_adapter_op */
 		memset(&ev[i], 0, sizeof(struct rte_event));
 		ev[i].event = 0;
@@ -294,8 +286,6 @@  test_op_forward_mode(void)
 
 		TEST_ASSERT_EQUAL(ret, 0, "Data mismatch for dma adapter\n");
 
-		rte_free(op->src_seg);
-		rte_free(op->dst_seg);
 		rte_mempool_put(op->op_mp, op);
 	}
 
@@ -400,7 +390,7 @@  configure_dmadev(void)
 						       rte_socket_id());
 	RTE_TEST_ASSERT_NOT_NULL(params.dst_mbuf_pool, "Can't create DMA_DST_MBUFPOOL\n");
 
-	elt_size = sizeof(struct rte_event_dma_adapter_op) + sizeof(struct rte_event);
+	elt_size = sizeof(struct rte_event_dma_adapter_op) + (sizeof(struct rte_dma_sge) * 2);
 	params.op_mpool = rte_mempool_create("EVENT_DMA_OP_POOL", DMA_OP_POOL_SIZE, elt_size, 0,
 					     0, NULL, NULL, NULL, NULL, rte_socket_id(), 0);
 	RTE_TEST_ASSERT_NOT_NULL(params.op_mpool, "Can't create DMA_OP_POOL\n");
diff --git a/doc/guides/prog_guide/event_dma_adapter.rst b/doc/guides/prog_guide/event_dma_adapter.rst
index 3443b6a8032e..1fb9b0a07b87 100644
--- a/doc/guides/prog_guide/event_dma_adapter.rst
+++ b/doc/guides/prog_guide/event_dma_adapter.rst
@@ -144,7 +144,7 @@  on which it enqueues events towards the DMA adapter using ``rte_event_enqueue_bu
    uint32_t cap;
    int ret;
 
-   /* Fill in event info and update event_ptr with rte_dma_op */
+   /* Fill in event info and update event_ptr with rte_event_dma_adapter_op */
    memset(&ev, 0, sizeof(ev));
    .
    .
diff --git a/drivers/dma/cnxk/cnxk_dmadev_fp.c b/drivers/dma/cnxk/cnxk_dmadev_fp.c
index f6562b603e45..9f7f9b2eed0e 100644
--- a/drivers/dma/cnxk/cnxk_dmadev_fp.c
+++ b/drivers/dma/cnxk/cnxk_dmadev_fp.c
@@ -457,7 +457,6 @@  cn10k_dma_adapter_enqueue(void *ws, struct rte_event ev[], uint16_t nb_events)
 	struct cnxk_dpi_compl_s *comp_ptr;
 	struct cnxk_dpi_conf *dpi_conf;
 	struct cnxk_dpi_vf_s *dpivf;
-	struct rte_event *rsp_info;
 	struct cn10k_sso_hws *work;
 	uint16_t nb_src, nb_dst;
 	rte_mcslock_t mcs_lock_me;
@@ -469,9 +468,7 @@  cn10k_dma_adapter_enqueue(void *ws, struct rte_event ev[], uint16_t nb_events)
 
 	for (count = 0; count < nb_events; count++) {
 		op = ev[count].event_ptr;
-		rsp_info = (struct rte_event *)((uint8_t *)op +
-			     sizeof(struct rte_event_dma_adapter_op));
-		dpivf =	rte_dma_fp_objs[op->dma_dev_id].dev_private;
+		dpivf = rte_dma_fp_objs[op->dma_dev_id].dev_private;
 		dpi_conf = &dpivf->conf[op->vchan];
 
 		if (unlikely(rte_mempool_get(dpi_conf->adapter_info.req_mp, (void **)&comp_ptr)))
@@ -488,15 +485,14 @@  cn10k_dma_adapter_enqueue(void *ws, struct rte_event ev[], uint16_t nb_events)
 		hdr[0] = dpi_conf->cmd.u | ((uint64_t)DPI_HDR_PT_WQP << 54);
 		hdr[0] |= (nb_dst << 6) | nb_src;
 		hdr[1] = ((uint64_t)comp_ptr);
-		hdr[2] = cnxk_dma_adapter_format_event(rsp_info->event);
+		hdr[2] = cnxk_dma_adapter_format_event(ev[count].event);
 
-		src = &op->src_seg[0];
-		dst = &op->dst_seg[0];
+		src = &op->src_dst_seg[0];
+		dst = &op->src_dst_seg[op->nb_src];
 
 		if (CNXK_TAG_IS_HEAD(work->gw_rdata) ||
 		    ((CNXK_TT_FROM_TAG(work->gw_rdata) == SSO_TT_ORDERED) &&
-		    (rsp_info->sched_type & DPI_HDR_TT_MASK) ==
-			    RTE_SCHED_TYPE_ORDERED))
+		     (ev[count].sched_type & DPI_HDR_TT_MASK) == RTE_SCHED_TYPE_ORDERED))
 			roc_sso_hws_head_wait(work->base);
 
 		rte_mcslock_lock(&dpivf->mcs_lock, &mcs_lock_me);
@@ -566,12 +562,12 @@  cn9k_dma_adapter_dual_enqueue(void *ws, struct rte_event ev[], uint16_t nb_event
 		 * For all other cases, src pointers are first pointers.
 		 */
 		if (((dpi_conf->cmd.u >> 48) & DPI_HDR_XTYPE_MASK) == DPI_XTYPE_INBOUND) {
-			fptr = &op->dst_seg[0];
-			lptr = &op->src_seg[0];
+			fptr = &op->src_dst_seg[nb_src];
+			lptr = &op->src_dst_seg[0];
 			RTE_SWAP(nb_src, nb_dst);
 		} else {
-			fptr = &op->src_seg[0];
-			lptr = &op->dst_seg[0];
+			fptr = &op->src_dst_seg[0];
+			lptr = &op->src_dst_seg[nb_src];
 		}
 
 		hdr[0] = ((uint64_t)nb_dst << 54) | (uint64_t)nb_src << 48;
@@ -612,7 +608,6 @@  cn9k_dma_adapter_enqueue(void *ws, struct rte_event ev[], uint16_t nb_events)
 	struct cnxk_dpi_compl_s *comp_ptr;
 	struct cnxk_dpi_conf *dpi_conf;
 	struct cnxk_dpi_vf_s *dpivf;
-	struct rte_event *rsp_info;
 	struct cn9k_sso_hws *work;
 	uint16_t nb_src, nb_dst;
 	rte_mcslock_t mcs_lock_me;
@@ -624,9 +619,7 @@  cn9k_dma_adapter_enqueue(void *ws, struct rte_event ev[], uint16_t nb_events)
 
 	for (count = 0; count < nb_events; count++) {
 		op = ev[count].event_ptr;
-		rsp_info = (struct rte_event *)((uint8_t *)op +
-			    sizeof(struct rte_event_dma_adapter_op));
-		dpivf =	rte_dma_fp_objs[op->dma_dev_id].dev_private;
+		dpivf = rte_dma_fp_objs[op->dma_dev_id].dev_private;
 		dpi_conf = &dpivf->conf[op->vchan];
 
 		if (unlikely(rte_mempool_get(dpi_conf->adapter_info.req_mp, (void **)&comp_ptr)))
@@ -647,18 +640,18 @@  cn9k_dma_adapter_enqueue(void *ws, struct rte_event ev[], uint16_t nb_events)
 		 * For all other cases, src pointers are first pointers.
 		 */
 		if (((dpi_conf->cmd.u >> 48) & DPI_HDR_XTYPE_MASK) == DPI_XTYPE_INBOUND) {
-			fptr = &op->dst_seg[0];
-			lptr = &op->src_seg[0];
+			fptr = &op->src_dst_seg[nb_src];
+			lptr = &op->src_dst_seg[0];
 			RTE_SWAP(nb_src, nb_dst);
 		} else {
-			fptr = &op->src_seg[0];
-			lptr = &op->dst_seg[0];
+			fptr = &op->src_dst_seg[0];
+			lptr = &op->src_dst_seg[nb_src];
 		}
 
 		hdr[0] = ((uint64_t)nb_dst << 54) | (uint64_t)nb_src << 48;
-		hdr[0] |= cnxk_dma_adapter_format_event(rsp_info->event);
+		hdr[0] |= cnxk_dma_adapter_format_event(ev[count].event);
 
-		if ((rsp_info->sched_type & DPI_HDR_TT_MASK) == RTE_SCHED_TYPE_ORDERED)
+		if ((ev[count].sched_type & DPI_HDR_TT_MASK) == RTE_SCHED_TYPE_ORDERED)
 			roc_sso_hws_head_wait(work->base);
 
 		rte_mcslock_lock(&dpivf->mcs_lock, &mcs_lock_me);
diff --git a/lib/eventdev/rte_event_dma_adapter.c b/lib/eventdev/rte_event_dma_adapter.c
index 24dff556dbfe..e52ef46a1b36 100644
--- a/lib/eventdev/rte_event_dma_adapter.c
+++ b/lib/eventdev/rte_event_dma_adapter.c
@@ -236,9 +236,9 @@  edma_circular_buffer_flush_to_dma_dev(struct event_dma_adapter *adapter,
 				      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;
+	struct dma_vchan_info *tq;
 	uint16_t n;
 	uint16_t i;
 	int ret;
@@ -257,11 +257,13 @@  edma_circular_buffer_flush_to_dma_dev(struct event_dma_adapter *adapter,
 	for (i = 0; i < n; i++)	{
 		op = bufp->op_buffer[*head];
 		if (op->nb_src == 1 && op->nb_dst == 1)
-			ret = rte_dma_copy(dma_dev_id, vchan, op->src_seg->addr, op->dst_seg->addr,
-					   op->src_seg->length, op->flags);
+			ret = rte_dma_copy(dma_dev_id, vchan, op->src_dst_seg[0].addr,
+					   op->src_dst_seg[1].addr, op->src_dst_seg[0].length,
+					   op->flags);
 		else
-			ret = rte_dma_copy_sg(dma_dev_id, vchan, op->src_seg, op->dst_seg,
-					      op->nb_src, op->nb_dst, op->flags);
+			ret = rte_dma_copy_sg(dma_dev_id, vchan, &op->src_dst_seg[0],
+					      &op->src_dst_seg[op->nb_src], op->nb_src, op->nb_dst,
+					      op->flags);
 		if (ret < 0)
 			break;
 
@@ -511,8 +513,7 @@  edma_enq_to_dma_dev(struct event_dma_adapter *adapter, struct rte_event *ev, uns
 		if (dma_op == NULL)
 			continue;
 
-		/* Expected to have response info appended to dma_op. */
-
+		dma_op->impl_opaque[0] = ev[i].event;
 		dma_dev_id = dma_op->dma_dev_id;
 		vchan = dma_op->vchan;
 		vchan_qinfo = &adapter->dma_devs[dma_dev_id].vchanq[vchan];
@@ -647,7 +648,6 @@  edma_ops_enqueue_burst(struct event_dma_adapter *adapter, struct rte_event_dma_a
 	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;
@@ -659,16 +659,7 @@  edma_ops_enqueue_burst(struct event_dma_adapter *adapter, struct rte_event_dma_a
 	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 = ops[i]->impl_opaque[0];
 		ev->event_ptr = ops[i];
 		ev->event_type = RTE_EVENT_TYPE_DMADEV;
 		if (adapter->implicit_release_disabled)
diff --git a/lib/eventdev/rte_event_dma_adapter.h b/lib/eventdev/rte_event_dma_adapter.h
index e924ab673df7..048ddba3f354 100644
--- a/lib/eventdev/rte_event_dma_adapter.h
+++ b/lib/eventdev/rte_event_dma_adapter.h
@@ -157,24 +157,46 @@  extern "C" {
  * instance.
  */
 struct rte_event_dma_adapter_op {
-	struct rte_dma_sge *src_seg;
-	/**< Source segments. */
-	struct rte_dma_sge *dst_seg;
-	/**< Destination segments. */
-	uint16_t nb_src;
-	/**< Number of source segments. */
-	uint16_t nb_dst;
-	/**< Number of destination segments. */
 	uint64_t flags;
 	/**< Flags related to the operation.
 	 * @see RTE_DMA_OP_FLAG_*
 	 */
-	int16_t dma_dev_id;
-	/**< DMA device ID to be used */
-	uint16_t vchan;
-	/**< DMA vchan ID to be used */
 	struct rte_mempool *op_mp;
 	/**< Mempool from which op is allocated. */
+	enum rte_dma_status_code status;
+	/**< Status code for this operation. */
+	uint32_t rsvd;
+	/**< Reserved for future use. */
+	uint64_t impl_opaque[2];
+	/**< Implementation-specific opaque data.
+	 * An dma device implementation use this field to hold
+	 * implementation specific values to share between dequeue and enqueue
+	 * operations.
+	 * The application should not modify this field.
+	 */
+	uint64_t user_meta;
+	/**<  Memory to store user specific metadata.
+	 * The dma device implementation should not modify this area.
+	 */
+	uint64_t event_meta;
+	/**< Event metadata that defines event attributes when used in OP_NEW mode.
+	 * @see rte_event_dma_adapter_mode::RTE_EVENT_DMA_ADAPTER_OP_NEW
+	 * @see struct rte_event::event
+	 */
+	int16_t dma_dev_id;
+	/**< DMA device ID to be used with OP_FORWARD mode.
+	 * @see rte_event_dma_adapter_mode::RTE_EVENT_DMA_ADAPTER_OP_FORWARD
+	 */
+	uint16_t vchan;
+	/**< DMA vchan ID to be used with OP_FORWARD mode
+	 * @see rte_event_dma_adapter_mode::RTE_EVENT_DMA_ADAPTER_OP_FORWARD
+	 */
+	uint16_t nb_src;
+	/**< Number of source segments. */
+	uint16_t nb_dst;
+	/**< Number of destination segments. */
+	struct rte_dma_sge src_dst_seg[0];
+	/**< Source and destination segments. */
 };
 
 /**