[dpdk-dev] [RFC v3 1/1] lib: add compressdev API

[Trahe, Fiona]

Hi Ahmed,

> -----Original Message-----
> From: Ahmed Mansour [mailto:ahmed.mansour at nxp.com]
> Sent: Monday, December 18, 2017 9:44 PM
> To: dev at dpdk.org; Shally.Verma at cavium.com
> Cc: Mahipal.Challa at cavium.com; NarayanaPrasad.Athreya at cavium.com; De Lara Guarch, Pablo
> <pablo.de.lara.guarch at intel.com>; Trahe, Fiona <fiona.trahe at intel.com>; Roy Pledge
> <roy.pledge at nxp.com>; Youri Querry <youri.querry_1 at nxp.com>; Hemant Agrawal
> <hemant.agrawal at nxp.com>
> Subject: Re: [RFC v3 1/1] lib: add compressdev API
> 
> On 12/15/2017 11:19 PM, Trahe, Fiona wrote:
> .. <snip>
> 
> > +
> > +/** Compression Algorithms */
> > +enum rte_comp_algorithm {
> > +	RTE_COMP_NULL = 0,
> > +	/**< No compression.
> > +	 * Pass-through, data is copied unchanged from source buffer to
> > +	 * destination buffer.
> > +	 */
> > +	RTE_COMP_DEFLATE,
> > +	/**< DEFLATE compression algorithm
> > +	 *
> https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Ftools.ietf.org%2Fhtml%2Frfc195
> 1&data=02%7C01%7Cahmed.mansour%40nxp.com%7Cf3edbd70b38b49eb1f0308d54634e444%7C686ea
> 1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C636492115333128845&sdata=B3G0aIncVAK17dXlnSivXi0
> e56h2D7pEQZ9gK%2Fh3qZQ%3D&reserved=0
> > +	 */
> > +	RTE_COMP_LZS,
> > +	/**< LZS compression algorithm
> > +	 *
> https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Ftools.ietf.org%2Fhtml%2Frfc239
> 5&data=02%7C01%7Cahmed.mansour%40nxp.com%7Cf3edbd70b38b49eb1f0308d54634e444%7C686ea
> 1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C636492115333128845&sdata=aNRFIfkelXlCUUgpp%2BzC
> YaTu28tp6fF0m6k7F13w1Ps%3D&reserved=0
> > +	 */
> > +	RTE_COMP_ALGO_LIST_END
> > +};
> > +
> > +/**< Compression Level.
> > + * The number is interpreted by each PMD differently. However, lower numbers
> > + * give fastest compression, at the expense of compression ratio while
> > + * higher numbers may give better compression ratios but are likely slower.
> > + */
> > +#define	RTE_COMP_LEVEL_PMD_DEFAULT	(-1)
> > +/** Use PMD Default */
> > +#define	RTE_COMP_LEVEL_NONE		(0)
> > +/** Output uncompressed blocks if supported by the specified algorithm */
> > +#define RTE_COMP_LEVEL_MIN		(1)
> > +/** Use minimum compression level supported by the PMD */
> > +#define RTE_COMP_LEVEL_MAX		(9)
> > +/** Use maximum compression level supported by the PMD */
> > +
> > +/** Compression checksum types */
> > +enum rte_comp_checksum_type {
> > +	RTE_COMP_NONE,
> > +	/**< No checksum generated */
> > +	RTE_COMP_CRC32,
> > +	/**< Generates a CRC32 checksum, as used by gzip */
> > +	RTE_COMP_ADLER32,
> > +	/**< Generates an Adler-32 checksum, as used by zlib */
> > +	RTE_COMP_CRC32_ADLER32,
> > +	/**< Generates both Adler-32 and CRC32 checksums, concatenated.
> > +	 * CRC32 is in the lower 32bits, Adler-32 in the upper 32 bits.
> > +	 */
> 
> What would be a real life use case for returning both CRC32 and ADLER32?
> Packaging the data once as Gzip and once as zlib?
[Fiona] We've had requests for this from customers.

> 
> > +};
> > +
> > +/*
> > + * enum rte_comp_hash_algo {
> > + *   RTE_COMP_HASH_NONE,
> > + *   RTE_COMP_HASH_SHA1,
> > + *   RTE_COMP_HASH_SHA256,
> > + * };
> > + * Need further input from cavium on this
> > + * xform will need a flag with above enum value
> > + * op will need to provide a virt/phys ptr to a data buffer of appropriate size.
> > + * And via capability PMD can say whether supported or not.
> > + */
> > +
> > +/** Compression Huffman Type - used by DEFLATE algorithm */
> > +enum rte_comp_huffman {
> > +	RTE_COMP_DEFAULT,
> > +	/**< PMD may choose which Huffman codes to use */
> > +	RTE_COMP_FIXED,
> > +	/**< Use Fixed Huffman codes */
> > +	RTE_COMP_DYNAMIC,
> > +	/**< Use Dynamic Huffman codes */
> > +};
> > +
> > +
> > +enum rte_comp_flush_flag {
> > +	RTE_COMP_FLUSH_NONE,
> > +	/**< Data is not flushed. Output may remain in the compressor and be
> > +	 * processed during a following op. It may not be possible to decompress
> > +	 * output until a later op with some other flush flag has been sent.
> > +	 */
> > +	RTE_COMP_FLUSH_SYNC,
> > +	/**< All data should be flushed to output buffer. Output data can be
> > +	 * decompressed. However state and history is not cleared, so future
> > +	 * ops may use history from this op */
> > +	RTE_COMP_FLUSH_FULL,
> > +	/**< All data should be flushed to output buffer. Output data can be
> > +	 * decompressed. State and history data is cleared, so future
> > +	 * ops will be independent of ops processed before this.
> > +	 */
> > +	RTE_COMP_FLUSH_FINAL
> > +	/**< Same as RTE_COMP_FLUSH_FULL but also bfinal bit is set in last block
> > +	 */
> > +	 /* TODO:
> > +	  * describe flag meanings for decompression.
> > +	  * describe behavous in OUT_OF_SPACE case.
> > +	  * At least the last flag is specific to deflate algo. Should this be
> > +	  * called rte_comp_deflate_flush_flag? And should there be
> > +	  * comp_op_deflate_params in the op? */
> 
> What about Z_BLOCK and Z_TREES? Those are needed for sfwr zlib.net
> replacements.
[Fiona] We haven't seen a need for those. I would suggest proposing a patch later once the initial API is stabilised. 

> > +};
> > +
> > +/** Compression transform types */
> > +enum rte_comp_xform_type {
> > +	RTE_COMP_COMPRESS,
> > +	/**< Compression service - compress */
> > +	RTE_COMP_DECOMPRESS,
> > +	/**< Compression service - decompress */
> > +};
> > +
> 
> ... <snip>
> 
> > +struct rte_comp_session;
> > +/**
> > + * Compression Operation.
> > + *
> > + * This structure contains data relating to performing a compression
> > + * operation on the referenced mbuf data buffers.
> > + *
> > + * All compression operations are Out-of-place (OOP) operations,
> > + * as the size of the output data is different to the size of the input data.
> > + *
> > + * Comp operations are enqueued and dequeued in comp PMDs using the
> > + * rte_compressdev_enqueue_burst() / rte_compressdev_dequeue_burst() APIs
> > + */
> > +struct rte_comp_op {
> > +
> > +	enum rte_comp_op_type op_type;
> > +	void * stream_private;
> > +	/* location where PMD maintains stream state
> > +	 * only required if op_type is STATEFUL, else should be NULL
> > +	 */
> > +	struct rte_comp_session *session;
> > +	/**< Handle for the initialised session context */
> > +	struct rte_mempool *mempool;
> > +	/**< mempool from which operation is allocated */
> > +	phys_addr_t phys_addr;
> > +	/**< physical address of this operation */
> 
> iova_addr?
[Fiona] Yes, this should be rte_iova_t.  will fix this in the v1.
> 
> > +	struct rte_mbuf *m_src;
> > +	/**< source mbuf
> > +	 * The total size of the input buffer(s) can be retrieved using
> > +	 * rte_pktmbuf_data_len(m_src)
> > +	 */
> > +	struct rte_mbuf *m_dst;
> > +	/**< destination mbuf
> > +	 * The total size of the output buffer(s) can be retrieved using
> > +	 * rte_pktmbuf_data_len(m_dst)
> > +	 */
> > +
> > +	struct {
> > +		uint32_t offset;
> > +		/**< Starting point for compression or decompression,
> > +		 * specified as number of bytes from start of packet in
> > +		 * source buffer.
> > +		 * Starting point for checksum generation in compress direction.
> > +		 */
> 
> Why should offset have two different meanings for compression and
> decompression. It seems that the use case of offset on input is
> applicable to both use modes 
[Fiona] It does have the same meaning for compression/decompression?
Just for checksum it's different. Checksum is usually required on the uncompressed data.
In decompression the input buffer contains compressed data.
  

> <snip>
>  ...