DPDK patches and discussions
 help / color / mirror / Atom feed
From: Stanislaw Kardach <kda@semihalf.com>
To: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
Cc: dev@dpdk.org, konstantin.ananyev@intel.com,
	andrey.chilikin@intel.com, ray.kinsella@intel.com,
	yipeng1.wang@intel.com, sameh.gobriel@intel.com,
	bruce.richardson@intel.com, david.marchand@redhat.com
Subject: Re: [dpdk-dev] [PATCH v2] hash: fix tuple adjustment
Date: Wed, 5 May 2021 21:13:22 +0200	[thread overview]
Message-ID: <20210505191322.idkyj5q66ne7cz6q@toster> (raw)
In-Reply-To: <1620138304-203463-1-git-send-email-vladimir.medvedkin@intel.com>

On Tue, May 04, 2021 at 03:25:04PM +0100, Vladimir Medvedkin wrote:
> rte_thash_adjust_tuple() uses random to generate a new subtuple if
> fn() callback reports about collision. In some cases random changes
> the subtuple in a way that after complementary bits are applied the
> original tuple is obtained. This patch replaces random with subtuple
> increment.
> 
> Fixes: 28ebff11c2dc ("hash: add predictable RSS")
> Cc: vladimir.medvedkin@intel.com
> 
> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
> ---
>  lib/hash/rte_thash.c | 121 ++++++++++++++++++++++++++++++++++++++++++---------
>  1 file changed, 100 insertions(+), 21 deletions(-)
> 
> diff --git a/lib/hash/rte_thash.c b/lib/hash/rte_thash.c
> index 135a26d..58129df 100644
> --- a/lib/hash/rte_thash.c
> +++ b/lib/hash/rte_thash.c
> @@ -610,16 +610,91 @@ rte_thash_get_key(struct rte_thash_ctx *ctx)
>  	return ctx->hash_key;
>  }
>  
> +static inline uint8_t
> +read_unaligned_byte(uint8_t *ptr, unsigned int len, unsigned int offset)
> +{
> +	uint8_t ret = 0;
> +
> +	ret = ptr[offset / CHAR_BIT];
> +	if (offset % CHAR_BIT) {
> +		ret <<= (offset % CHAR_BIT);
> +		ret |= ptr[(offset / CHAR_BIT) + 1] >>
> +			(CHAR_BIT - (offset % CHAR_BIT));
> +	}
> +
> +	return ret >> (CHAR_BIT - len);
> +}
> +
> +static inline uint32_t
> +read_unaligned_bits(uint8_t *ptr, int len, int offset)
> +{
> +	uint32_t ret = 0;
> +
> +	len = RTE_MAX(len, 0);
> +	len = RTE_MIN(len, (int)(sizeof(uint32_t) * CHAR_BIT));
> +
> +	while (len > 0) {
> +		ret <<= CHAR_BIT;
> +
> +		ret |= read_unaligned_byte(ptr, RTE_MIN(len, CHAR_BIT),
> +			offset);
> +		offset += CHAR_BIT;
> +		len -= CHAR_BIT;
> +	}
> +
> +	return ret;
> +}
> +
> +/* returns mask for len bits with given offset inside byte */
> +static inline uint8_t
> +get_bits_mask(unsigned int len, unsigned int offset)
> +{
> +	unsigned int last_bit;
> +
> +	offset %= CHAR_BIT;
> +	/* last bit within byte */
> +	last_bit = RTE_MIN((unsigned int)CHAR_BIT, offset + len);
> +
> +	return ((1 << (CHAR_BIT - offset)) - 1) ^
> +		((1 << (CHAR_BIT - last_bit)) - 1);
> +}
> +
> +static inline void
> +write_unaligned_byte(uint8_t *ptr, unsigned int len,
> +	unsigned int offset, uint8_t val)
> +{
> +	uint8_t tmp;
> +
> +	tmp = ptr[offset / CHAR_BIT];
> +	tmp &= ~get_bits_mask(len, offset);
> +	tmp |= ((val << (CHAR_BIT - len)) >> (offset % CHAR_BIT));
> +	ptr[offset / CHAR_BIT] = tmp;
> +	if (((offset + len) / CHAR_BIT) != (offset / CHAR_BIT)) {
> +		int rest_len = (offset + len) % CHAR_BIT;
> +		tmp = ptr[(offset + len) / CHAR_BIT];
> +		tmp &= ~get_bits_mask(rest_len, 0);
> +		tmp |= val << (CHAR_BIT - rest_len);
> +		ptr[(offset + len) / CHAR_BIT] = tmp;
> +	}
> +}
> +
>  static inline void
> -xor_bit(uint8_t *ptr, uint32_t bit, uint32_t pos)
> +write_unaligned_bits(uint8_t *ptr, int len, int offset, uint32_t val)
>  {
> -	uint32_t byte_idx = pos >> 3;
> -	uint32_t bit_idx = (CHAR_BIT - 1) - (pos & (CHAR_BIT - 1));
>  	uint8_t tmp;
> +	unsigned int part_len;
> +
> +	len = RTE_MAX(len, 0);
> +	len = RTE_MIN(len, (int)(sizeof(uint32_t) * CHAR_BIT));
>  
> -	tmp = ptr[byte_idx];
> -	tmp ^= bit << bit_idx;
> -	ptr[byte_idx] = tmp;
> +	while (len > 0) {
> +		part_len = RTE_MIN(CHAR_BIT, len);
> +		tmp = (uint8_t)val & ((1 << part_len) - 1);
> +		write_unaligned_byte(ptr, part_len,
> +			offset + len - part_len, tmp);
> +		len -= CHAR_BIT;
> +		val >>= CHAR_BIT;
> +	}
>  }
>  
>  int
> @@ -632,8 +707,10 @@ rte_thash_adjust_tuple(struct rte_thash_ctx *ctx,
>  	uint32_t tmp_tuple[tuple_len / sizeof(uint32_t)];
>  	unsigned int i, j, ret = 0;
>  	uint32_t hash, adj_bits;
> -	uint8_t bit;
>  	const uint8_t *hash_key;
> +	uint32_t tmp;
> +	int offset;
> +	int tmp_len;
>  
>  	if ((ctx == NULL) || (h == NULL) || (tuple == NULL) ||
>  			(tuple_len % sizeof(uint32_t) != 0) || (attempts <= 0))
> @@ -641,6 +718,8 @@ rte_thash_adjust_tuple(struct rte_thash_ctx *ctx,
>  
>  	hash_key = rte_thash_get_key(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] =
> @@ -651,14 +730,12 @@ rte_thash_adjust_tuple(struct rte_thash_ctx *ctx,
>  
>  		/*
>  		 * Hint: LSB of adj_bits corresponds to
> -		 * offset + len bit of tuple
> +		 * offset + len bit of the subtuple
>  		 */
> -		for (j = 0; j < sizeof(uint32_t) * CHAR_BIT; j++) {
> -			bit = (adj_bits >> j) & 0x1;
> -			if (bit)
> -				xor_bit(tuple, bit, h->tuple_offset +
> -					h->tuple_len - 1 - j);
> -		}
> +		offset =  h->tuple_offset + h->tuple_len - ctx->reta_sz_log;
> +		tmp = read_unaligned_bits(tuple, ctx->reta_sz_log, offset);
> +		tmp ^= adj_bits;
> +		write_unaligned_bits(tuple, ctx->reta_sz_log, offset, tmp);
>  
>  		if (fn != NULL) {
>  			ret = (fn(userdata, tuple)) ? 0 : -EEXIST;
> @@ -666,13 +743,15 @@ rte_thash_adjust_tuple(struct rte_thash_ctx *ctx,
>  				return 0;
>  			else if (i < (attempts - 1)) {
Small nit. This comment should be updated as the bits aren't random
anymore, just incremented.
>  				/* Update tuple with random bits */
> -				for (j = 0; j < h->tuple_len; j++) {
> -					bit = rte_rand() & 0x1;
> -					if (bit)
> -						xor_bit(tuple, bit,
> -							h->tuple_offset +
> -							h->tuple_len - 1 - j);
> -				}
> +				tmp_len = RTE_MIN(sizeof(uint32_t) * CHAR_BIT,
> +					h->tuple_len - ctx->reta_sz_log);
> +				offset -= tmp_len;
> +				tmp = read_unaligned_bits(tuple, tmp_len,
> +					offset);
> +				tmp++;
> +				tmp &= (1 << tmp_len) - 1;
> +				write_unaligned_bits(tuple, tmp_len, offset,
> +					tmp);
>  			}
>  		} else
>  			return 0;
> -- 
> 2.7.4
>
Makes the issue not visible on both x86 and RISC-V.

Tested-by: Stanislaw Kardach <kda@semihalf.com>
Reviewed-by: Stanislaw Kardach <kda@semihalf.com>

-- 
Best Regards,
Stanislaw Kardach

  parent reply	other threads:[~2021-05-05 19:13 UTC|newest]

Thread overview: 7+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-05-04 14:07 [dpdk-dev] [PATCH] " Vladimir Medvedkin
2021-05-04 14:25 ` [dpdk-dev] [PATCH v2] " Vladimir Medvedkin
2021-05-05 10:16   ` Thomas Monjalon
2021-05-05 17:57   ` Wang, Yipeng1
2021-05-05 19:13   ` Stanislaw Kardach [this message]
2021-05-06 11:40   ` [dpdk-dev] [PATCH v3] " Vladimir Medvedkin
2021-05-10 13:25     ` Thomas Monjalon

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210505191322.idkyj5q66ne7cz6q@toster \
    --to=kda@semihalf.com \
    --cc=andrey.chilikin@intel.com \
    --cc=bruce.richardson@intel.com \
    --cc=david.marchand@redhat.com \
    --cc=dev@dpdk.org \
    --cc=konstantin.ananyev@intel.com \
    --cc=ray.kinsella@intel.com \
    --cc=sameh.gobriel@intel.com \
    --cc=vladimir.medvedkin@intel.com \
    --cc=yipeng1.wang@intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).