DPDK patches and discussions
 help / color / mirror / Atom feed
* [dpdk-dev] [PATCH v1 1/1] compress/octeontx: async burst mode feature support
@ 2022-12-15 18:55 Mahipal Challa
  2023-01-04 11:41 ` [EXT] " Akhil Goyal
  0 siblings, 1 reply; 2+ messages in thread
From: Mahipal Challa @ 2022-12-15 18:55 UTC (permalink / raw)
  To: dev; +Cc: jerinj, pathreya, ashishg, mchalla

Async burst mode feature support is added, improves the single thread
compression/decompression throughput.

Signed-off-by: Mahipal Challa <mchalla@marvell.com>
---
 drivers/compress/octeontx/include/zip_regs.h |   1 +
 drivers/compress/octeontx/otx_zip.h          |  37 +--
 drivers/compress/octeontx/otx_zip_pmd.c      | 258 ++++++++++---------
 3 files changed, 160 insertions(+), 136 deletions(-)

diff --git a/drivers/compress/octeontx/include/zip_regs.h b/drivers/compress/octeontx/include/zip_regs.h
index a7fcccc055..a07483b045 100644
--- a/drivers/compress/octeontx/include/zip_regs.h
+++ b/drivers/compress/octeontx/include/zip_regs.h
@@ -687,6 +687,7 @@ union zip_zptr_s {
 #define ZIP_COMP_E_TIMEOUT       (0xc)
 #define ZIP_COMP_E_INSTR_ERR     (0xd)
 #define ZIP_COMP_E_HCTX_ERR      (0xe)
+#define ZIP_COMP_E_PTR_ERR       (0xf)
 #define ZIP_COMP_E_STOP          (3)
 
 /**
diff --git a/drivers/compress/octeontx/otx_zip.h b/drivers/compress/octeontx/otx_zip.h
index cdef8cc6cb..ee14deb03d 100644
--- a/drivers/compress/octeontx/otx_zip.h
+++ b/drivers/compress/octeontx/otx_zip.h
@@ -55,6 +55,10 @@ extern int octtx_zip_logtype_driver;
 				ZIP_MAX_NCBP_SIZE)/* ~8072ull */
 
 #define ZIP_BUF_SIZE	256
+#define ZIP_BURST_SIZE	64
+
+#define ZIP_MAXSEG_SIZE      59460
+#define ZIP_EXTRABUF_SIZE    4096
 
 #define ZIP_SGPTR_ALIGN	16
 #define ZIP_CMDQ_ALIGN	128
@@ -67,7 +71,7 @@ extern int octtx_zip_logtype_driver;
 	((_align) * (((x) + (_align) - 1) / (_align)))
 
 /**< ZIP PMD device name */
-#define COMPRESSDEV_NAME_ZIP_PMD	compress_octeonx
+#define COMPRESSDEV_NAME_ZIP_PMD	compress_octeontx
 
 #define ZIP_PMD_LOG(level, fmt, args...) \
 	rte_log(RTE_LOG_ ## level, \
@@ -95,18 +99,18 @@ struct zip_stream;
 struct zipvf_qp;
 
 /* Algorithm handler function prototype */
-typedef int (*comp_func_t)(struct rte_comp_op *op,
-			   struct zipvf_qp *qp, struct zip_stream *zstrm);
+typedef int (*comp_func_t)(struct rte_comp_op *op, struct zipvf_qp *qp,
+			   struct zip_stream *zstrm, int num);
 
 /**
  * ZIP private stream structure
  */
 struct zip_stream {
-	union zip_inst_s *inst;
+	union zip_inst_s *inst[ZIP_BURST_SIZE];
 	/* zip instruction pointer */
 	comp_func_t func;
 	/* function to process comp operation */
-	void *bufs[MAX_BUFS_PER_STREAM];
+	void *bufs[MAX_BUFS_PER_STREAM * ZIP_BURST_SIZE];
 } __rte_cache_aligned;
 
 
@@ -162,11 +166,10 @@ struct zip_vf {
 
 
 static inline void
-zipvf_prepare_in_buf(struct zip_stream *zstrm, struct rte_comp_op *op)
+zipvf_prepare_in_buf(union zip_inst_s *inst, struct rte_comp_op *op)
 {
 	uint32_t offset, inlen;
 	struct rte_mbuf *m_src;
-	union zip_inst_s *inst = zstrm->inst;
 
 	inlen = op->src.length;
 	offset = op->src.offset;
@@ -180,11 +183,10 @@ zipvf_prepare_in_buf(struct zip_stream *zstrm, struct rte_comp_op *op)
 }
 
 static inline void
-zipvf_prepare_out_buf(struct zip_stream *zstrm, struct rte_comp_op *op)
+zipvf_prepare_out_buf(union zip_inst_s *inst, struct rte_comp_op *op)
 {
 	uint32_t offset;
 	struct rte_mbuf *m_dst;
-	union zip_inst_s *inst = zstrm->inst;
 
 	offset = op->dst.offset;
 	m_dst = op->m_dst;
@@ -195,14 +197,15 @@ zipvf_prepare_out_buf(struct zip_stream *zstrm, struct rte_comp_op *op)
 			rte_pktmbuf_iova_offset(m_dst, offset);
 	inst->s.totaloutputlength = rte_pktmbuf_pkt_len(m_dst) -
 			op->dst.offset;
+	if (inst->s.totaloutputlength == ZIP_MAXSEG_SIZE)
+		inst->s.totaloutputlength += ZIP_EXTRABUF_SIZE; /* DSTOP */
+
 	inst->s.out_ptr_ctl.s.length = inst->s.totaloutputlength;
 }
 
 static inline void
-zipvf_prepare_cmd_stateless(struct rte_comp_op *op, struct zip_stream *zstrm)
+zipvf_prepare_cmd_stateless(struct rte_comp_op *op, union zip_inst_s *inst)
 {
-	union zip_inst_s *inst = zstrm->inst;
-
 	/* set flush flag to always 1*/
 	inst->s.ef = 1;
 
@@ -215,8 +218,8 @@ zipvf_prepare_cmd_stateless(struct rte_comp_op *op, struct zip_stream *zstrm)
 	inst->s.adlercrc32 = op->input_chksum;
 
 	/* Prepare gather buffers */
-	zipvf_prepare_in_buf(zstrm, op);
-	zipvf_prepare_out_buf(zstrm, op);
+	zipvf_prepare_in_buf(inst, op);
+	zipvf_prepare_out_buf(inst, op);
 }
 
 #ifdef ZIP_DBG
@@ -224,6 +227,7 @@ static inline void
 zip_dump_instruction(void *inst)
 {
 	union zip_inst_s *cmd83 = (union zip_inst_s *)inst;
+
 	printf("####### START ########\n");
 	printf("doneint:%d totaloutputlength:%d\n", cmd83->s.doneint,
 		cmd83->s.totaloutputlength);
@@ -265,9 +269,8 @@ void
 zipvf_push_command(struct zipvf_qp *qp, union zip_inst_s *zcmd);
 
 int
-zip_process_op(struct rte_comp_op *op,
-				struct zipvf_qp *qp,
-				struct zip_stream *zstrm);
+zip_process_op(struct rte_comp_op *op, struct zipvf_qp *qp,
+	       struct zip_stream *zstrm, int num);
 
 uint64_t
 zip_reg_read64(uint8_t *hw_addr, uint64_t offset);
diff --git a/drivers/compress/octeontx/otx_zip_pmd.c b/drivers/compress/octeontx/otx_zip_pmd.c
index dff188e223..2d856b19bb 100644
--- a/drivers/compress/octeontx/otx_zip_pmd.c
+++ b/drivers/compress/octeontx/otx_zip_pmd.c
@@ -32,10 +32,8 @@ static const struct rte_compressdev_capabilities
  * Reset session to default state for next set of stateless operation
  */
 static inline void
-reset_stream(struct zip_stream *z_stream)
+reset_stream(union zip_inst_s *inst)
 {
-	union zip_inst_s *inst = (union zip_inst_s *)(z_stream->inst);
-
 	inst->s.bf = 1;
 	inst->s.ef = 0;
 }
@@ -43,12 +41,11 @@ reset_stream(struct zip_stream *z_stream)
 int
 zip_process_op(struct rte_comp_op *op,
 		struct zipvf_qp *qp,
-		struct zip_stream *zstrm)
+		struct zip_stream *zstrm, int num)
 {
-	union zip_inst_s *inst = zstrm->inst;
+	union zip_inst_s *inst = zstrm->inst[num];
 	volatile union zip_zres_s *zresult = NULL;
 
-
 	if ((op->m_src->nb_segs > 1) || (op->m_dst->nb_segs > 1) ||
 			(op->src.offset > rte_pktmbuf_pkt_len(op->m_src)) ||
 			(op->dst.offset > rte_pktmbuf_pkt_len(op->m_dst))) {
@@ -57,9 +54,9 @@ zip_process_op(struct rte_comp_op *op,
 		return 0;
 	}
 
-	zipvf_prepare_cmd_stateless(op, zstrm);
+	zipvf_prepare_cmd_stateless(op, inst);
 
-	zresult = (union zip_zres_s *)zstrm->bufs[RES_BUF];
+	zresult = (union zip_zres_s *)zstrm->bufs[RES_BUF + num];
 	zresult->s.compcode = 0;
 
 #ifdef ZIP_DBG
@@ -69,43 +66,6 @@ zip_process_op(struct rte_comp_op *op,
 	/* Submit zip command */
 	zipvf_push_command(qp, (void *)inst);
 
-	/* Check and Process results in sync mode */
-	do {
-	} while (!zresult->s.compcode);
-
-	if (zresult->s.compcode == ZIP_COMP_E_SUCCESS) {
-		op->status = RTE_COMP_OP_STATUS_SUCCESS;
-	} else {
-		/* FATAL error cannot do anything */
-		ZIP_PMD_ERR("operation failed with error code:%d\n",
-			zresult->s.compcode);
-		if (zresult->s.compcode == ZIP_COMP_E_DSTOP)
-			op->status = RTE_COMP_OP_STATUS_OUT_OF_SPACE_TERMINATED;
-		else
-			op->status = RTE_COMP_OP_STATUS_ERROR;
-	}
-
-#ifdef ZIP_DBG
-	ZIP_PMD_INFO("written %d\n", zresult->s.totalbyteswritten);
-#endif
-
-	/* Update op stats */
-	switch (op->status) {
-	case RTE_COMP_OP_STATUS_SUCCESS:
-		op->consumed = zresult->s.totalbytesread;
-	/* Fall-through */
-	case RTE_COMP_OP_STATUS_OUT_OF_SPACE_TERMINATED:
-		op->produced = zresult->s.totalbyteswritten;
-		break;
-	default:
-		ZIP_PMD_ERR("stats not updated for status:%d\n",
-				op->status);
-		break;
-	}
-	/* zstream is reset irrespective of result */
-	reset_stream(zstrm);
-
-	zresult->s.compcode = ZIP_COMP_E_NOTDONE;
 	return 0;
 }
 
@@ -115,86 +75,88 @@ zip_set_stream_parameters(struct rte_compressdev *dev,
 			const struct rte_comp_xform *xform,
 			struct zip_stream *z_stream)
 {
-	int ret;
-	union zip_inst_s *inst;
 	struct zip_vf *vf = (struct zip_vf *)dev->data->dev_private;
+	union zip_inst_s *inst;
 	void *res;
+	int ret, i;
 
 	/* Allocate resources required by a stream */
 	ret = rte_mempool_get_bulk(vf->zip_mp,
-			z_stream->bufs, MAX_BUFS_PER_STREAM);
+			z_stream->bufs, (MAX_BUFS_PER_STREAM * ZIP_BURST_SIZE));
 	if (ret < 0)
 		return -1;
 
-	/* get one command buffer from pool and set up */
-	inst = (union zip_inst_s *)z_stream->bufs[CMD_BUF];
-	res = z_stream->bufs[RES_BUF];
-
-	memset(inst->u, 0, sizeof(inst->u));
-
-	/* set bf for only first ops of stream */
-	inst->s.bf = 1;
-
-	if (xform->type == RTE_COMP_COMPRESS) {
-		inst->s.op = ZIP_OP_E_COMP;
+	for (i = 0; i < ZIP_BURST_SIZE; i++) {
+		/* get one command buffer from pool and set up */
+		inst = (union zip_inst_s *)z_stream->bufs[(CMD_BUF * ZIP_BURST_SIZE) + i];
+		res = z_stream->bufs[(RES_BUF * ZIP_BURST_SIZE) + i];
+		memset(inst->u, 0, sizeof(inst->u));
+
+		/* set bf for only first ops of stream */
+		inst->s.bf = 1;
+
+		if (xform->type == RTE_COMP_COMPRESS) {
+			inst->s.op = ZIP_OP_E_COMP;
+
+			switch (xform->compress.deflate.huffman) {
+			case RTE_COMP_HUFFMAN_DEFAULT:
+				inst->s.cc = ZIP_CC_DEFAULT;
+				break;
+			case RTE_COMP_HUFFMAN_FIXED:
+				inst->s.cc = ZIP_CC_FIXED_HUFF;
+				break;
+			case RTE_COMP_HUFFMAN_DYNAMIC:
+				inst->s.cc = ZIP_CC_DYN_HUFF;
+				break;
+			default:
+				ret = -1;
+				goto err;
+			}
+
+			switch (xform->compress.level) {
+			case RTE_COMP_LEVEL_MIN:
+				inst->s.ss = ZIP_COMP_E_LEVEL_MIN;
+				break;
+			case RTE_COMP_LEVEL_MAX:
+				inst->s.ss = ZIP_COMP_E_LEVEL_MAX;
+				break;
+			case RTE_COMP_LEVEL_NONE:
+				ZIP_PMD_ERR("Compression level not supported");
+				ret = -1;
+				goto err;
+			default:
+				/* for any value between min and max , choose
+				 * PMD default.
+				 */
+				inst->s.ss = ZIP_COMP_E_LEVEL_MED; /** PMD default **/
+				break;
+			}
+		} else if (xform->type == RTE_COMP_DECOMPRESS) {
+			inst->s.op = ZIP_OP_E_DECOMP;
+			/* from HRM,
+			 * For DEFLATE decompression, [CC] must be 0x0.
+			 * For decompression, [SS] must be 0x0
+			 */
+			inst->s.cc = 0;
+			/* Speed bit should not be set for decompression */
+			inst->s.ss = 0;
+			/* decompression context is supported only for STATEFUL
+			 * operations. Currently we support STATELESS ONLY so
+			 * skip setting of ctx pointer
+			 */
 
-		switch (xform->compress.deflate.huffman) {
-		case RTE_COMP_HUFFMAN_DEFAULT:
-			inst->s.cc = ZIP_CC_DEFAULT;
-			break;
-		case RTE_COMP_HUFFMAN_FIXED:
-			inst->s.cc = ZIP_CC_FIXED_HUFF;
-			break;
-		case RTE_COMP_HUFFMAN_DYNAMIC:
-			inst->s.cc = ZIP_CC_DYN_HUFF;
-			break;
-		default:
+		} else {
+			ZIP_PMD_ERR("\nxform type not supported");
 			ret = -1;
 			goto err;
 		}
 
-		switch (xform->compress.level) {
-		case RTE_COMP_LEVEL_MIN:
-			inst->s.ss = ZIP_COMP_E_LEVEL_MIN;
-			break;
-		case RTE_COMP_LEVEL_MAX:
-			inst->s.ss = ZIP_COMP_E_LEVEL_MAX;
-			break;
-		case RTE_COMP_LEVEL_NONE:
-			ZIP_PMD_ERR("Compression level not supported");
-			ret = -1;
-			goto err;
-		default:
-			/* for any value between min and max , choose
-			 * PMD default.
-			 */
-			inst->s.ss = ZIP_COMP_E_LEVEL_MED; /** PMD default **/
-			break;
-		}
-	} else if (xform->type == RTE_COMP_DECOMPRESS) {
-		inst->s.op = ZIP_OP_E_DECOMP;
-		/* from HRM,
-		 * For DEFLATE decompression, [CC] must be 0x0.
-		 * For decompression, [SS] must be 0x0
-		 */
-		inst->s.cc = 0;
-		/* Speed bit should not be set for decompression */
-		inst->s.ss = 0;
-		/* decompression context is supported only for STATEFUL
-		 * operations. Currently we support STATELESS ONLY so
-		 * skip setting of ctx pointer
-		 */
+		inst->s.res_ptr_addr.s.addr = rte_mempool_virt2iova(res);
+		inst->s.res_ptr_ctl.s.length = 0;
 
-	} else {
-		ZIP_PMD_ERR("\nxform type not supported");
-		ret = -1;
-		goto err;
+		z_stream->inst[i] = inst;
 	}
 
-	inst->s.res_ptr_addr.s.addr = rte_mempool_virt2iova(res);
-	inst->s.res_ptr_ctl.s.length = 0;
-
-	z_stream->inst = inst;
 	z_stream->func = zip_process_op;
 
 	return 0;
@@ -202,7 +164,7 @@ zip_set_stream_parameters(struct rte_compressdev *dev,
 err:
 	rte_mempool_put_bulk(vf->zip_mp,
 			     (void *)&(z_stream->bufs[0]),
-			     MAX_BUFS_PER_STREAM);
+			     (MAX_BUFS_PER_STREAM * ZIP_BURST_SIZE));
 
 	return ret;
 }
@@ -235,7 +197,7 @@ zip_pmd_config(struct rte_compressdev *dev,
 	/** TBD Should we use the per core object cache for stream resources */
 	zip_buf_mp = rte_mempool_create(
 			res_pool,
-			nb_streams * MAX_BUFS_PER_STREAM,
+			(nb_streams * MAX_BUFS_PER_STREAM * ZIP_BURST_SIZE),
 			ZIP_BUF_SIZE,
 			0,
 			0,
@@ -453,6 +415,7 @@ zip_pmd_stream_create(struct rte_compressdev *dev,
 		return ret;
 	}
 	*stream = strm;
+
 	return 0;
 }
 
@@ -470,7 +433,7 @@ zip_pmd_stream_free(struct rte_compressdev *dev, void *stream)
 	/* Free resources back to pool */
 	rte_mempool_put_bulk(vf->zip_mp,
 				(void *)&(z_stream->bufs[0]),
-				MAX_BUFS_PER_STREAM);
+				(MAX_BUFS_PER_STREAM * ZIP_BURST_SIZE));
 
 	/* Zero out the whole structure */
 	memset(stream, 0, sizeof(struct zip_stream));
@@ -481,12 +444,12 @@ zip_pmd_stream_free(struct rte_compressdev *dev, void *stream)
 
 
 static uint16_t
-zip_pmd_enqueue_burst_sync(void *queue_pair,
+zip_pmd_enqueue_burst(void *queue_pair,
 		struct rte_comp_op **ops, uint16_t nb_ops)
 {
 	struct zipvf_qp *qp = queue_pair;
-	struct rte_comp_op *op;
 	struct zip_stream *zstrm;
+	struct rte_comp_op *op;
 	int i, ret = 0;
 	uint16_t enqd = 0;
 
@@ -501,7 +464,7 @@ zip_pmd_enqueue_burst_sync(void *queue_pair,
 			if (unlikely(zstrm == NULL))
 				op->status = RTE_COMP_OP_STATUS_INVALID_ARGS;
 			else
-				ret = zstrm->func(op, qp, zstrm);
+				ret = zstrm->func(op, qp, zstrm, i);
 		}
 
 		/* Whatever is out of op, put it into completion queue with
@@ -518,21 +481,78 @@ zip_pmd_enqueue_burst_sync(void *queue_pair,
 			enqd++;
 		}
 	}
+
+#ifdef ZIP_DBG
+	ZIP_PMD_INFO("ops_enqd[nb_ops:%d]:%d\n", nb_ops, enqd);
+#endif
 	return enqd;
 }
 
 static uint16_t
-zip_pmd_dequeue_burst_sync(void *queue_pair,
+zip_pmd_dequeue_burst(void *queue_pair,
 		struct rte_comp_op **ops, uint16_t nb_ops)
 {
+	volatile union zip_zres_s *zresult = NULL;
 	struct zipvf_qp *qp = queue_pair;
 
 	unsigned int nb_dequeued = 0;
+	struct zip_stream *zstrm;
+	struct rte_comp_op *op;
+	unsigned int i;
 
 	nb_dequeued = rte_ring_dequeue_burst(qp->processed_pkts,
-			(void **)ops, nb_ops, NULL);
+				(void **)ops, nb_ops, NULL);
 	qp->qp_stats.dequeued_count += nb_dequeued;
 
+	/* Dequeue all the submitted ops */
+	for (i = 0; i < nb_dequeued; i++) {
+		op = ops[i];
+		/* process stateless ops */
+		zstrm = (struct zip_stream *)op->private_xform;
+		zresult = (union zip_zres_s *)zstrm->bufs[RES_BUF + i];
+
+		/* Check and Process results in sync mode */
+		do {
+		} while (!zresult->s.compcode);
+
+		if (zresult->s.compcode == ZIP_COMP_E_SUCCESS) {
+			op->status = RTE_COMP_OP_STATUS_SUCCESS;
+		} else {
+			/* FATAL error cannot do anything */
+			ZIP_PMD_ERR("operation failed with error code:%d\n",
+				zresult->s.compcode);
+			if (zresult->s.compcode == ZIP_COMP_E_DSTOP)
+				op->status = RTE_COMP_OP_STATUS_OUT_OF_SPACE_TERMINATED;
+			else
+				op->status = RTE_COMP_OP_STATUS_ERROR;
+		}
+
+	#ifdef ZIP_DBG
+		ZIP_PMD_INFO("written %d\n", zresult->s.totalbyteswritten);
+	#endif
+
+		/* Update op stats */
+		switch (op->status) {
+		case RTE_COMP_OP_STATUS_SUCCESS:
+			op->consumed = zresult->s.totalbytesread;
+		/* Fall-through */
+		case RTE_COMP_OP_STATUS_OUT_OF_SPACE_TERMINATED:
+			op->produced = zresult->s.totalbyteswritten;
+			break;
+		default:
+			ZIP_PMD_ERR("stats not updated for status:%d\n",
+				    op->status);
+			break;
+		}
+
+		/* zstream is reset irrespective of result */
+		reset_stream(zstrm->inst[i]);
+		zresult->s.compcode = ZIP_COMP_E_NOTDONE;
+	}
+
+#ifdef ZIP_DBG
+	ZIP_PMD_INFO("ops_deqd[nb_ops:%d]: %d\n", nb_ops, nb_dequeued);
+#endif
 	return nb_dequeued;
 }
 
@@ -597,8 +617,8 @@ zip_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 
 	compressdev->dev_ops = &octtx_zip_pmd_ops;
 	/* register rx/tx burst functions for data path */
-	compressdev->dequeue_burst = zip_pmd_dequeue_burst_sync;
-	compressdev->enqueue_burst = zip_pmd_enqueue_burst_sync;
+	compressdev->dequeue_burst = zip_pmd_dequeue_burst;
+	compressdev->enqueue_burst = zip_pmd_enqueue_burst;
 	compressdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
 	return ret;
 }
-- 
2.25.1


^ permalink raw reply	[flat|nested] 2+ messages in thread

* RE: [EXT] [dpdk-dev] [PATCH v1 1/1] compress/octeontx: async burst mode feature support
  2022-12-15 18:55 [dpdk-dev] [PATCH v1 1/1] compress/octeontx: async burst mode feature support Mahipal Challa
@ 2023-01-04 11:41 ` Akhil Goyal
  0 siblings, 0 replies; 2+ messages in thread
From: Akhil Goyal @ 2023-01-04 11:41 UTC (permalink / raw)
  To: Mahipal Challa, dev
  Cc: Jerin Jacob Kollanukkaran, Narayana Prasad Raju Athreya,
	Ashish Gupta, Mahipal Challa

> Async burst mode feature support is added, improves the single thread
> compression/decompression throughput.
> 
> Signed-off-by: Mahipal Challa <mchalla@marvell.com>
Applied to dpdk-next-crypto

Thanks.

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2023-01-04 11:41 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-12-15 18:55 [dpdk-dev] [PATCH v1 1/1] compress/octeontx: async burst mode feature support Mahipal Challa
2023-01-04 11:41 ` [EXT] " Akhil Goyal

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).