[dpdk-dev] [PATCH v2] compress/isal: add chained mbuf support

[De Lara Guarch, Pablo]

Hi Lee,

> -----Original Message-----
> From: Daly, Lee
> Sent: Wednesday, July 11, 2018 1:40 PM
> To: De Lara Guarch, Pablo <pablo.de.lara.guarch at intel.com>
> Cc: dev at dpdk.org; Daly, Lee <lee.daly at intel.com>
> Subject: [PATCH v2] compress/isal: add chained mbuf support
> 
> This patch adds chained mbuf support for input or output buffers during
> compression/decompression operations.
> 
> V2 - Fixed using chained mbufs with offsets.
> 
> Signed-off-by: Lee Daly <lee.daly at intel.com>
> ---
>  doc/guides/compressdevs/features/isal.ini     |   1 +
>  doc/guides/compressdevs/isal.rst              |   2 -
>  drivers/compress/isal/isal_compress_pmd.c     | 397 ++++++++++++++++++++---
> ---
>  drivers/compress/isal/isal_compress_pmd_ops.c |   5 +-
>  4 files changed, 321 insertions(+), 84 deletions(-)
> 
> diff --git a/doc/guides/compressdevs/features/isal.ini
> b/doc/guides/compressdevs/features/isal.ini
> index 7183d10..06a95fb 100644
> --- a/doc/guides/compressdevs/features/isal.ini
> +++ b/doc/guides/compressdevs/features/isal.ini
> @@ -8,6 +8,7 @@ CPU SSE        = Y
>  CPU AVX        = Y
>  CPU AVX2       = Y
>  CPU AVX512     = Y
> +Chained mbufs  = Y

There are new feature flags replacing this one.

>  Deflate        = Y
>  Fixed          = Y
>  Dynamic        = Y
> diff --git a/doc/guides/compressdevs/isal.rst
> b/doc/guides/compressdevs/isal.rst
> index 276ff06..3bc3022 100644
> --- a/doc/guides/compressdevs/isal.rst
> +++ b/doc/guides/compressdevs/isal.rst
> @@ -78,8 +78,6 @@ As a result the level mappings from the API to the PMD are
> shown below.
>  Limitations
>  -----------
> 
> -* Chained mbufs will not be supported until a future release, meaning max data
> size being passed to PMD through a single mbuf is 64K - 1. If data is larger than
> this it will need to be split up and sent as multiple operations.
> -
>  * Compressdev level 0, no compression, is not supported.
> 
>  * Checksums will not be supported until future release.
> diff --git a/drivers/compress/isal/isal_compress_pmd.c
> b/drivers/compress/isal/isal_compress_pmd.c
> index 747ded1..70231c4 100644
> --- a/drivers/compress/isal/isal_compress_pmd.c
> +++ b/drivers/compress/isal/isal_compress_pmd.c
> @@ -188,6 +188,211 @@ isal_comp_set_priv_xform_parameters(struct
> isal_priv_xform *priv_xform,
>  	return 0;
>  }
> 
> +/* Compression using chained mbufs for input/output data */ static int
> +chained_mbuf_compression(struct rte_comp_op *op, struct isal_comp_qp
> +*qp) {
> +	int ret, i;

"i" should be unsigned int.

> +	uint8_t segs;
> +	uint32_t consumed_data;
> +	uint32_t remaining_data = op->src.length;
> +	struct rte_mbuf *src = op->m_src;
> +	struct rte_mbuf *dst = op->m_dst;
> +
> +	/* check for offset passing multiple segments
> +	 * and point compression stream to input/output buffer
> +	 */
> +	if (op->src.offset > src->data_len) {

>=

> +		segs = op->src.offset / src->data_len;

This is assuming that each segment has the same data length.
Better to do a while loop, decreasing the offset (remaining_offset?),
and moving to the next segment, until it is less than the current segment.

> +		for (i = 0; i < segs; i++)
> +			src = src->next;
> +
> +		qp->stream->avail_in = src->data_len -
> +				(op->src.offset % src->data_len);

Doing what I suggested above simplifies this, as you can substract "remaining_offset" to data_len.

> +		qp->stream->next_in = rte_pktmbuf_mtod_offset(src, uint8_t *,
> +				(op->src.offset % src->data_len));
> +	} else {
> +		qp->stream->avail_in = src->data_len - op->src.offset;
> +		qp->stream->next_in = rte_pktmbuf_mtod_offset(src, uint8_t *,
> +				op->src.offset);
> +	}
> +
> +	if (op->dst.offset > dst->data_len) {
> +		segs = op->dst.offset / dst->data_len;
> +		for (int i = 0; i < segs; i++)
> +			dst = dst->next;

Same as above.

> +
> +		qp->stream->avail_out = dst->data_len -
> +				(op->dst.offset % dst->data_len);
> +		qp->stream->next_out = rte_pktmbuf_mtod_offset(dst, uint8_t
> *,
> +				(op->dst.offset % dst->data_len));

...

> +		if (remaining_data == op->src.length) {
> +			consumed_data = src->data_len - qp->stream->avail_in
> -
> +					(op->src.offset % src->data_len);
> +		} else
> +			consumed_data = src->data_len - qp->stream->avail_in;
> +
> +		remaining_data -= consumed_data;

What about "remaining_data = op->src.length - qp->stream->total_in"?
This way, you don't need the condition (nor "consumed_data").

> +
> +		if (ret != COMP_OK) {
> +			ISAL_PMD_LOG(ERR, "Compression operation
> failed\n");
> +			op->status = RTE_COMP_OP_STATUS_ERROR;
> +			return ret;
> +		}

...

> +			} else {
> +				ISAL_PMD_LOG(ERR,
> +				"Not enough input buffer segments\n"
> +					"Output Produced: %d\n", op-
> >produced);

This will always be 0, as it is only updated at the end of the function.

> +				op->status =
> RTE_COMP_OP_STATUS_INVALID_ARGS;
> +				return -1;
> +			}
> +		}
> +

...

> +				ISAL_PMD_LOG(ERR,
> +				"Not enough output buffer segments\n"
> +					"Output Produced: %d\n", op-
> >produced);

Same as above.

> +				op->status =
> +
> 	RTE_COMP_OP_STATUS_OUT_OF_SPACE_TERMINATED;
> +				return -1;
> +			}
> +		}
> +	}
> +	op->consumed = qp->stream->total_in;
> +	op->produced = qp->stream->total_out;
> +
> +	return 0;
> +}
> +
> +/* Decompression using chained mbufs for input/output data */ static
> +int chained_mbuf_decompression(struct rte_comp_op *op, struct
> +isal_comp_qp *qp) {

Most comments from the compression function applies to the decompression one.

>  /* Stateless Compression Function */
>  static int
>  process_isal_deflate(struct rte_comp_op *op, struct isal_comp_qp *qp, @@ -

...

> +	/* Chained mbufs */
> +	if (op->m_src->nb_segs > 1 || op->m_dst->nb_segs > 1) {
> +		ret = chained_mbuf_compression(op, qp);
> +		if (ret < 0) //check the returns... ensure op status changed.
> +			return ret;
> +	} else {
> +	/* Linear buffer */
> +		qp->stream->end_of_stream = 1; /* All input consumed in one
> go */
> +		if ((op->src.length + op->src.offset) > op->m_src->data_len) {

This is unnecessary, as the check that you are doing above is effectively the same.

> +			ISAL_PMD_LOG(ERR, "Input mbuf not big enough for
> offset.\n");
> +			op->status = RTE_COMP_OP_STATUS_INVALID_ARGS;

...

> +		/*  Point compression stream to output buffer */
> +		qp->stream->avail_out = op->m_dst->data_len;

Data_len - dst.offset?

> +		qp->stream->next_out  = rte_pktmbuf_mtod_offset(op->m_dst,
> uint8_t *,
> +			op->dst.offset);
> 

...

> @@ -291,57 +511,72 @@ process_isal_inflate(struct rte_comp_op *op, struct
> isal_comp_qp *qp)

...

> +	/* Chained mbufs */
> +	if (op->m_src->nb_segs > 1 || op->m_dst->nb_segs > 1) {
> +		ret = chained_mbuf_decompression(op, qp);
> +		if (ret !=  0)
> +			return ret;
> +	} else {
> +		/* Linear buffer */
> +		if ((op->src.length + op->src.offset) > op->m_src->data_len) {

This is unnecessary, as the check that you are doing above is effectively the same.

> +			ISAL_PMD_LOG(ERR, "Input mbuf not big enough for
> offset.\n");
> +			op->status = RTE_COMP_OP_STATUS_INVALID_ARGS;
>  		return -1;
> -	}

...

> +		/* Point decompression state to output buffer */
> +		qp->state->avail_out = op->m_dst->data_len;

Data_len - dst.offset?

...

> -	if (ret != ISAL_DECOMP_OK) {
> -		op->status = RTE_COMP_OP_STATUS_ERROR;
> -		return ret;
> +		op->consumed = op->src.length - qp->state->avail_in;

What about "total_in"?

> +		op->produced = qp->state->total_out;
>  	}
> 
> -	op->consumed = op->src.length - qp->state->avail_in;
> -	op->produced = qp->state->total_out;
> -
> -return ret;
> +	return ret;
>  }