[dpdk-dev] [PATCH 2/5] hash: enable gfni thash implementation

[Medvedkin, Vladimir]

Hi Konstantin,

On 08/10/2021 13:31, Ananyev, Konstantin wrote:
> 
>> This patch enables new GFNI Toeplitz hash in
>> predictable RSS library.
>>
>> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin at intel.com>
>> ---
>>   lib/hash/rte_thash.c | 43 +++++++++++++++++++++++++++++++++++++++----
>>   lib/hash/rte_thash.h | 19 +++++++++++++++++++
>>   lib/hash/version.map |  1 +
>>   3 files changed, 59 insertions(+), 4 deletions(-)
>>
>> diff --git a/lib/hash/rte_thash.c b/lib/hash/rte_thash.c
>> index 07447f7..86a0e96 100644
>> --- a/lib/hash/rte_thash.c
>> +++ b/lib/hash/rte_thash.c
>> @@ -86,6 +86,8 @@ struct rte_thash_ctx {
>>   	uint32_t	reta_sz_log;	/** < size of the RSS ReTa in bits */
>>   	uint32_t	subtuples_nb;	/** < number of subtuples */
>>   	uint32_t	flags;
>> +	uint64_t	*matrices;
> 
> Comment, what is that, etc.
> 

I'll rephrase the comment below.

>> +	/**< rte_thash_complete_matrix generated matrices */
>>   	uint8_t		hash_key[0];
>>   };
>>
>> @@ -253,12 +255,25 @@ rte_thash_init_ctx(const char *name, uint32_t key_len, uint32_t reta_sz,
>>   			ctx->hash_key[i] = rte_rand();
>>   	}
>>
>> +	if (rte_thash_gfni_supported) {
> 
> I think it should be:
> if (rte_thash_gfni_supported && rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
> 
> 

Agree

>> +		ctx->matrices = rte_zmalloc(NULL, key_len * sizeof(uint64_t),
>> +			RTE_CACHE_LINE_SIZE);
> 
> You can do it probably before allocation ctx, at the same place where te is allocated.
> Might be a bit nicer.
> 

I'd prefer to keep allocation and initialization of matrices in one 
place, below there is rte_thash_complete_matrix() which uses previously 
generated ctx->hash_key.

>> +		if (ctx->matrices == NULL)
> 
> 		RTE_LOG(ERR, ...);
> 		rte_ernno = ENOMEM;
> 

Agree

>> +			goto free_ctx;
>> +
>> +		rte_thash_complete_matrix(ctx->matrices, ctx->hash_key,
>> +			key_len);
>> +	}
>> +
>>   	te->data = (void *)ctx;
>>   	TAILQ_INSERT_TAIL(thash_list, te, next);
>>
>>   	rte_mcfg_tailq_write_unlock();
>>
>>   	return ctx;
>> +
>> +free_ctx:
>> +	rte_free(ctx);
>>   free_te:
>>   	rte_free(te);
>>   exit:
>> @@ -372,6 +387,10 @@ generate_subkey(struct rte_thash_ctx *ctx, struct thash_lfsr *lfsr,
>>   			set_bit(ctx->hash_key, get_rev_bit_lfsr(lfsr), i);
>>   	}
>>
>> +	if (rte_thash_gfni_supported)
> 
> Here and in data-path functions, I think it would be better:
> if (ctx->matrices != NULL)

Agree

>> +		rte_thash_complete_matrix(ctx->matrices, ctx->hash_key,
>> +			ctx->key_len);
>> +
>>   	return 0;
>>   }
>>
>> @@ -628,6 +647,16 @@ rte_thash_get_key(struct rte_thash_ctx *ctx)
>>   	return ctx->hash_key;
>>   }
>>
>> +const uint64_t *
>> +rte_thash_get_gfni_matrices(struct rte_thash_ctx *ctx)
>> +{
>> +	if (rte_thash_gfni_supported)
>> +		return ctx->matrices;
> 
> Why not just always:
> return ctx->matices;
> ?
> 

Agree

>> +
>> +	rte_errno = ENOTSUP;
>> +	return NULL;
>> +}
>> +
>>   static inline uint8_t
>>   read_unaligned_byte(uint8_t *ptr, unsigned int len, unsigned int offset)
>>   {
>> @@ -739,11 +768,17 @@ rte_thash_adjust_tuple(struct rte_thash_ctx *ctx,
>>   	attempts = RTE_MIN(attempts, 1U << (h->tuple_len - ctx->reta_sz_log));
>>
>>   	for (i = 0; i < attempts; i++) {
>> -		for (j = 0; j < (tuple_len / 4); j++)
>> -			tmp_tuple[j] =
>> -				rte_be_to_cpu_32(*(uint32_t *)&tuple[j * 4]);
>> +		if (rte_thash_gfni_supported)
> if (ctx->matrices)
> 
>> +			hash = rte_thash_gfni(ctx->matrices, tuple, tuple_len);
>> +		else {
>> +			for (j = 0; j < (tuple_len / 4); j++)
>> +				tmp_tuple[j] =
>> +					rte_be_to_cpu_32(
>> +						*(uint32_t *)&tuple[j * 4]);
>> +
>> +			hash = rte_softrss(tmp_tuple, tuple_len / 4, hash_key);
>> +		}
>>
>> -		hash = rte_softrss(tmp_tuple, tuple_len / 4, hash_key);
>>   		adj_bits = rte_thash_get_complement(h, hash, desired_value);
>>
>>   		/*
>> diff --git a/lib/hash/rte_thash.h b/lib/hash/rte_thash.h
>> index e3f1fc6..6e6861c 100644
>> --- a/lib/hash/rte_thash.h
>> +++ b/lib/hash/rte_thash.h
>> @@ -410,6 +410,25 @@ const uint8_t *
>>   rte_thash_get_key(struct rte_thash_ctx *ctx);
>>
>>   /**
>> + * Get a pointer to the toeplitz hash matrices contained in the context.
>> + * These matrices could be used with fast toeplitz hash implementation if
>> + * CPU supports GFNI.
>> + * Matrices changes after each addition of a helper.
>> + *
>> + * @warning
>> + * @b EXPERIMENTAL: this API may change without prior notice.
>> + *
>> + * @param ctx
>> + *  Thash context
>> + * @return
>> + *  A pointer to the toeplitz hash key matrices on success
>> + *  NULL if GFNI is not supported.
>> + */
>> +__rte_experimental
>> +const uint64_t *
>> +rte_thash_get_gfni_matrices(struct rte_thash_ctx *ctx);
>> +
>> +/**
>>    * Function prototype for the rte_thash_adjust_tuple
>>    * to check if adjusted tuple could be used.
>>    * Generally it is some kind of lookup function to check
>> diff --git a/lib/hash/version.map b/lib/hash/version.map
>> index cecf922..3eda695 100644
>> --- a/lib/hash/version.map
>> +++ b/lib/hash/version.map
>> @@ -43,6 +43,7 @@ EXPERIMENTAL {
>>   	rte_thash_find_existing;
>>   	rte_thash_free_ctx;
>>   	rte_thash_get_complement;
>> +	rte_thash_get_gfni_matrices;
>>   	rte_thash_get_helper;
>>   	rte_thash_get_key;
>>   	rte_thash_gfni_supported;
>> --
>> 2.7.4
> 

-- 
Regards,
Vladimir