Re: [dpdk-dev] [PATCH v2 2/3] hash: add predictable RSS implementation

["Medvedkin, Vladimir" <vladimir.medvedkin@intel.com>]

Hi Konstantin,

Thanks for the review,

On 07/04/2021 15:53, Ananyev, Konstantin wrote:
> Hi Vladimir,
> 
> Few comments below, mostly minor.
> One generic one - doc seems missing.
> With that in place:
> Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> 
>>
>> This patch implements predictable RSS functionality.
>>
>> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>

<snip>

>> +#defineRETA_SZ_MIN2U
>> +#defineRETA_SZ_MAX16U
> 
> Should these RETA_SZ defines be in public header?
> So user can know what are allowed values?
> 

I don't think this is necessary, because the user chooses it not 
arbitrary, but depending on the NIC.

>> +#define RETA_SZ_IN_RANGE(reta_sz)((reta_sz >= RETA_SZ_MIN) && \

<snip>

>> +uint32_t i;
> 
> Empty line is  missing.
> 

Thanks

>> +if ((name == NULL) || (key_len == 0) || !RETA_SZ_IN_RANGE(reta_sz)) {
>> +rte_errno = EINVAL;
>> +return NULL;
>> +}

<snip>

>> +static inline void
>> +set_bit(uint8_t *ptr, uint32_t bit, uint32_t pos)
>> +{
>> +uint32_t byte_idx = pos >> 3;
> 
> Just as a nit to be consistent with the line below:
> pos / CHAR_BIT;
> 

Fixed

>> +uint32_t bit_idx = (CHAR_BIT - 1) - (pos & (CHAR_BIT - 1));
>> +uint8_t tmp;

<snip>

>> +ent = rte_zmalloc(NULL, sizeof(struct rte_thash_subtuple_helper) +
>> +sizeof(uint32_t) * (1 << ctx->reta_sz_log), 0);
> 
> Helper can be used by data-path code (via rte_thash_get_compliment()) right?
> Then might be better to align it at cache-line.
> 

Agree, I'll fix it

>> +if (ent == NULL)
>> +return -ENOMEM;

<snip>

>>   uint32_t
>> -rte_thash_get_compliment(struct rte_thash_subtuple_helper *h __rte_unused,
>> -uint32_t hash __rte_unused, uint32_t desired_hash __rte_unused)
>> +rte_thash_get_compliment(struct rte_thash_subtuple_helper *h,
>> +uint32_t hash, uint32_t desired_hash)
>>   {
>> -return 0;
>> +return h->compl_table[(hash ^ desired_hash) & h->lsb_msk];
>>   }
> 
> Would it make sense to add another-one for multi values:
> rte_thash_get_compliment(uint32_t hash, const uint32_t desired_hashes[], uint32_t adj_hash[], uint32_t num);
> So user can get adjustment values for multiple queues at once?
> 

At the moment I can't find scenarios why do we need to have a bulk 
version for this function

>>
>>   const uint8_t *
>> -rte_thash_get_key(struct rte_thash_ctx *ctx __rte_unused)
>> +rte_thash_get_key(struct rte_thash_ctx *ctx)
>>   {
>> -return NULL;
>> +return ctx->hash_key;
>> +}
>> +
>> +static inline void
>> +xor_bit(uint8_t *ptr, uint32_t bit, uint32_t pos)
>> +{
>> +uint32_t byte_idx = pos >> 3;
>> +uint32_t bit_idx = (CHAR_BIT - 1) - (pos & (CHAR_BIT - 1));
>> +uint8_t tmp;
>> +
>> +tmp = ptr[byte_idx];
>> +tmp ^= bit << bit_idx;
>> +ptr[byte_idx] = tmp;
>> +}
>> +
>> +int
>> +rte_thash_adjust_tuple(struct rte_thash_subtuple_helper *h,
>> +uint8_t *orig_tuple, uint32_t adj_bits,
>> +rte_thash_check_tuple_t fn, void *userdata)
>> +{
>> +unsigned i;
>> +
>> +if ((h == NULL) || (orig_tuple == NULL))
>> +return -EINVAL;
>> +
>> +adj_bits &= h->lsb_msk;
>> +/* Hint: LSB of adj_bits corresponds to offset + len bit of tuple */
>> +for (i = 0; i < sizeof(uint32_t) * CHAR_BIT; i++) {
>> +uint8_t bit = (adj_bits >> i) & 0x1;
>> +if (bit)
>> +xor_bit(orig_tuple, bit,
>> +h->tuple_offset + h->tuple_len - 1 - i);
>> +}
>> +
>> +if (fn != NULL)
>> +return (fn(userdata, orig_tuple)) ? 0 : -EEXIST;
>> +
>> +return 0;
>>   }
> 
> Not sure is there much point to have a callback that is called only once.
> Might be better to rework the function in a way that user to provide 2 callbacks -
> one to generate new value, second to check.
> Something like that:
> 
> int
> rte_thash_gen_tuple(struct rte_thash_subtuple_helper *h,
> uint8_t *tuple, uint32_t desired_hash,
> int (*cb_gen_tuple)(uint8_t *, void *),
> int (*cb_check_tuple)(const uint8_t *, void *),
> void *userdata)
> {
> do {
> rc = cb_gen_tuple(tuple, userdata);
> if (rc != 0)
> return rc;
> hash = rte_softrss(tuple, ...);
> adj = rte_thash_get_compliment(h, hash, desired_hash);
> update_tuple(tuple, adj, ...);
> rc = cb_check_tuple(tuple, userdata);
> } while(rc != 0);
> 
>               return rc;
> }

Agree, there is no point to call the callback for a single function 
call. I'll rewrite rte_thash_adjust_tuple() and send a new version an 
v3. As for gen_tuple, I think we don't need to have a separate callback,
new rte_thash_adjust_tuple implementation randomly changes corresponding 
bits (based on configured offset and length in the helper) in the tuple.

> 
>> diff --git a/lib/librte_hash/rte_thash.h b/lib/librte_hash/rte_thash.h
>> index 38a641b..fd67931 100644
>> --- a/lib/librte_hash/rte_thash.h
>> +++ b/lib/librte_hash/rte_thash.h
>> @@ -360,6 +360,48 @@ __rte_experimental
>>   const uint8_t *
>>   rte_thash_get_key(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
>> + * if adjusted tuple is already in use.
>> + *
>> + * @param userdata
>> + *  Pointer to the userdata. It could be a pointer to the
>> + *  table with used tuples to search.
>> + * @param tuple
>> + *  Pointer to the tuple to check
>> + *
>> + * @return
>> + *  1 on success
>> + *  0 otherwise
>> + */
>> +typedef int (*rte_thash_check_tuple_t)(void *userdata, uint8_t *tuple);
>> +
>> +/**
>> + * Adjust tuple with complimentary bits.
>> + *
>> + * @param h
>> + *  Pointer to the helper struct
>> + * @param orig_tuple
>> + *  Pointer to the tuple to be adjusted
>> + * @param adj_bits
>> + *  Valure returned by rte_thash_get_compliment
>> + * @param fn
>> + *  Callback function to check adjusted tuple. Could be NULL
>> + * @param userdata
>> + *  Pointer to the userdata to be passed to fn(). Could be NULL
>> + *
>> + * @return
>> + *  0 on success
>> + *  negative otherwise
>> + */
>> +__rte_experimental
>> +int
>> +rte_thash_adjust_tuple(struct rte_thash_subtuple_helper *h,
>> +uint8_t *orig_tuple, uint32_t adj_bits,
>> +rte_thash_check_tuple_t fn, void *userdata);
>> +
>>   #ifdef __cplusplus
>>   }
>>   #endif
>> diff --git a/lib/librte_hash/version.map b/lib/librte_hash/version.map
>> index 93cb230..a992a1e 100644
>> --- a/lib/librte_hash/version.map
>> +++ b/lib/librte_hash/version.map
>> @@ -32,6 +32,7 @@ DPDK_21 {
>>   EXPERIMENTAL {
>>   global:
>>
>> +rte_thash_adjust_tuple;
>>   rte_hash_free_key_with_position;
>>   rte_hash_lookup_with_hash_bulk;
>>   rte_hash_lookup_with_hash_bulk_data;
>> --
>> 2.7.4
> 

-- 
Regards,
Vladimir