Re: [dpdk-dev] [PATCH v2] Add toeplitz hash algorithm used by RSS

[Bruce Richardson <bruce.richardson@intel.com>]

On Fri, May 08, 2015 at 10:58:12AM -0400, Vladimir Medvedkin wrote:
> Software implementation of the Toeplitz hash function used by RSS.
> Can be used either for packet distribution on single queue NIC
> or for simulating of RSS computation on specific NIC (for example
> after GRE header decapsulating).
> 
> v3 changes
> - Rework API to be more generic
> - Add sctp_tag into tuple
> 
> v2 changes
> - Add ipv6 support
> - Various style fixes
> 
> Signed-off-by: Vladimir Medvedkin <medvedkinv@gmail.com>

Hi Vladimir,

first off, thanks for this patch, it looks like something we want into DPDK.
However, as Thomas has already pointed out, I think we could really do with
some unit tests for this code. As it stands right now, we don't even know if this
header compiles, as the header file is not used by any C file in DPDK.

Some other comments are inline below.

Regards,
/Bruce
> ---
>  lib/librte_hash/Makefile    |   1 +
>  lib/librte_hash/rte_thash.h | 207 ++++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 208 insertions(+)
>  create mode 100644 lib/librte_hash/rte_thash.h
> 
> diff --git a/lib/librte_hash/Makefile b/lib/librte_hash/Makefile
> index 3696cb1..981230b 100644
> --- a/lib/librte_hash/Makefile
> +++ b/lib/librte_hash/Makefile
> @@ -49,6 +49,7 @@ SRCS-$(CONFIG_RTE_LIBRTE_HASH) += rte_fbk_hash.c
>  SYMLINK-$(CONFIG_RTE_LIBRTE_HASH)-include := rte_hash.h
>  SYMLINK-$(CONFIG_RTE_LIBRTE_HASH)-include += rte_hash_crc.h
>  SYMLINK-$(CONFIG_RTE_LIBRTE_HASH)-include += rte_jhash.h
> +SYMLINK-$(CONFIG_RTE_LIBRTE_HASH)-include += rte_thash.h
>  SYMLINK-$(CONFIG_RTE_LIBRTE_HASH)-include += rte_fbk_hash.h
>  
>  # this lib needs eal
> diff --git a/lib/librte_hash/rte_thash.h b/lib/librte_hash/rte_thash.h
> new file mode 100644
> index 0000000..5d5111b
> --- /dev/null
> +++ b/lib/librte_hash/rte_thash.h
> @@ -0,0 +1,207 @@
> +/*-
> + *   BSD LICENSE
> + *
> + *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.

Date and copyright owner look to need update here.

> + *   All rights reserved.
> + *
> + *   Redistribution and use in source and binary forms, with or without
> + *   modification, are permitted provided that the following conditions
> + *   are met:
> + *
> + *     * Redistributions of source code must retain the above copyright
> + *       notice, this list of conditions and the following disclaimer.
> + *     * Redistributions in binary form must reproduce the above copyright
> + *       notice, this list of conditions and the following disclaimer in
> + *       the documentation and/or other materials provided with the
> + *       distribution.
> + *     * Neither the name of Intel Corporation nor the names of its
> + *       contributors may be used to endorse or promote products derived
> + *       from this software without specific prior written permission.
> + *
> + *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
> + *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
> + *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
> + *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
> + *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
> + *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
> + *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
> + *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
> + *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
> + *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
> + *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
> + */
> +
> +#ifndef _RTE_THASH_H
> +#define _RTE_THASH_H
> +
> +/**
> + * @file
> + *
> + * toeplitz hash functions.
> + */
> +
> +#ifdef __cplusplus
> +extern "C" {
> +#endif
> +
> +/**
> + * Software implementation of the Toeplitz hash function used by RSS.
> + * Can be used either for packet distribution on single queue NIC
> + * or for simulating of RSS computation on specific NIC (for example
> + * after GRE header decapsulating)
> + */
> +
> +#include <stdint.h>
> +#include <rte_byteorder.h>
> +#include <rte_vect.h>
> +
> +#ifdef __SSE3__
> +static const __m128i bswap_mask = {0x0405060700010203, 0x0C0D0E0F08090A0B};
> +#endif

I think a more descriptive name for this might help. Yes, it's a mask for
byteswapping, but I think the name could do with being more descriptive. I see
below it's used for IPv6 IP extraction, so maybe that could be put in the name
somehow. A comment should also be added on what it is used for.
Finally, since this is a public symbol in the including C file, it should have
an rte_ prefix - or perhaps just an underscore prefix "_".

> +
> +#define RTE_THASH_V4_L3	 2	/*calculate hash of ipv4 header only*/
> +#define RTE_THASH_V4_L4	 3	/*calculate hash of ipv4 + transport headers*/
> +#define RTE_THASH_V6_L3	 8	/*calculate hash of ipv6 header only */
> +#define RTE_THASH_V6_L4	 9	/*calculate hash of ipv6 + transport headers */
> +

Are these #defines used anywhere? If not, they should be removed.

> +struct rte_ports {
> +        uint16_t dport;
> +        uint16_t sport;
> +};
> +
> +union rte_thash_l4 {
> +        struct          rte_ports ports;
> +        uint32_t        sctp_tag;
> +};
> +

Any reason for the rte_ports struct being separated out from the rte_thash_l4
structure?

> +/**
> + * IPv4 tuple
> + * addreses and ports/sctp_tag have to be CPU byte order
> + */
> +struct rte_ipv4_tuple {
> +	uint32_t	src_addr;
> +	uint32_t	dst_addr;
> +	union rte_thash_l4 l4;
> +};
> +
> +/**
> + * IPv6 tuple
> + * Addresses have to be filled by rte_thash_load_v6_addr()
> + * ports/sctp_tag have to be CPU byte order
> + */
> +struct rte_ipv6_tuple {
> +	uint8_t		src_addr[16];
> +	uint8_t		dst_addr[16];
> +	union rte_thash_l4 l4;
> +};
> +
> +union rte_thash_tuple {
> +	struct rte_ipv4_tuple	v4;
> +	struct rte_ipv6_tuple	v6;
> +} __attribute__((aligned(16)));
> +

I can see the reason for these structure, but they are not actually used
anywhere in the code below. Should some of the functions, e.g. load_v6_addr fn
not use these as type parameters?

> +/**
> + * Prepare special converted key to use with rte_softrss_be()
> + * @param orig
> + *   pointer to original RSS key
> + * @param targ
> + *   pointer to target RSS key
> + * @param len
> + *   RSS key length
> + */
> +static inline void
> +rte_convert_rss_key(const uint32_t *orig, uint32_t *targ, int len)
> +{
> +	int i;
> +
> +	for (i = 0; i < (len >> 2); i++) {
> +		targ[i] = rte_be_to_cpu_32(orig[i]);
> +	}
> +}
> +
> +/**
> + * Prepare and load IPv6 address
> + * @param orig
> + *   Pointer to ipv6 address inside ipv6_hdr
> + * @param targ
> + *   Pointer to ipv6 address inside rte_ipv6_tuple
> + */
> +static inline void
> +rte_thash_load_v6_addr(const uint8_t *orig, uint8_t *targ)

Why not take a struct ipv6_hdr (from ip.h) and struct rte_ipv6_tuple parameters
directly, rather than uint8_t parameters to somewhere in the middle of the
structures? Specifying that way allows the compile to check the user is doing
things correctly.

> +{
> +#ifdef __SSE3__

I actually think the minimum supported SSE instruction set level for DPDK is
SSE3, so perhaps these #idefs could be removed.

> +	__m128i ipv6 = _mm_loadu_si128((const __m128i *)orig);
> +	*(__m128i *)targ = _mm_shuffle_epi8(ipv6, bswap_mask);
> +#else
> +	int i;
> +
> +	for (i = 0; i < 4; i++) {
> +		*((uint32_t *)targ + i) =
> +			rte_be_to_cpu_32(*((const uint32_t *)orig + i));
> +	}
> +#endif
> +}
> +
> +/**
> + * Generic implementation. Can be used with original rss_key
> + * @param input_tuple
> + *   Pointer to input tuple
> + * @param input_len
> + *   Length of input_tuple in 4-bytes chunks
> + * @param rss_key
> + *   Pointer to RSS hash key.
> + * @return
> + *   Calculated hash value.
> + */
> +static inline uint32_t
> +rte_softrss(uint32_t *input_tuple, uint32_t input_len,
> +		const uint8_t *rss_key)
> +{
> +	uint32_t i, j, ret = 0;
> +
> +	for (j = 0; j < input_len; j++) {
> +		for (i = 0; i < 32; i++) {
> +			if (input_tuple[j] & (1 << (31 - i))) {
> +				ret ^= rte_cpu_to_be_32(((const uint32_t *)rss_key)[j]) << i |
> +					(uint32_t)((uint64_t)(rte_cpu_to_be_32(((const uint32_t *)rss_key)[j + 1])) >> (32 - i));

Line is rather long and not terribly readable. Can it be split up into a number
of easier to read statements (without affecting performance)?

> +			}
> +		}
> +	}
> +	return ret;
> +}
> +
> +/**
> + * Optimized implementation.
> + * If you want the calculated hash value matches NIC RSS value
> + * you have to use special converted key.

Put in reference to function used for the conversion.
Can you also document when you might want to use this version over the 
previous one?

> + * @param input_tuple
> + *   Pointer to input tuple
> + * @param input_len
> + *   Length of input_tuple in 4-bytes chunks
> + * @param *rss_key
> + *   Pointer to RSS hash key.
> + * @return
> + *   Calculated hash value.
> + */
> +static inline uint32_t
> +rte_softrss_be(uint32_t *input_tuple, uint32_t input_len,
> +		const uint8_t *rss_key)
> +{
> +	uint32_t i, j, ret = 0;
> +
> +	for (j = 0; j < input_len; j++) {
> +		for (i = 0; i < 32; i++) {
> +			if (input_tuple[j] & (1 << (31 - i))) {
> +				ret ^= ((const uint32_t *)rss_key)[j] << i |
> +					(uint32_t)((uint64_t)(((const uint32_t *)rss_key)[j + 1]) >> (32 - i));
> +			}
> +		}
> +	}
> +	return ret;
> +}
> +
> +#ifdef __cplusplus
> +}
> +#endif
> +
> +#endif /* _RTE_THASH_H */
> -- 
> 1.8.3.2
>