| Message ID | cover.1696330098.git.petrm@nvidia.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | mlxsw: Control the order of blocks in ACL region | expand |
On Tue, Oct 03, 2023 at 01:25:25PM +0200, Petr Machata wrote: > Amit Cohen writes: > > For 12 key blocks in the A-TCAM, rules are split into two records, which > constitute two lookups. The two records are linked using a > "large entry key ID". > > Due to a Spectrum-4 hardware issue, KVD entries that correspond to key > blocks 0 to 5 of 12 key blocks will be placed in the same KVD pipe if they > only differ in their "large entry key ID", as it is ignored. This results > in a reduced scale, we can insert less than 20k filters and get an error: > > $ tc -b flower.batch > RTNETLINK answers: Input/output error > We have an error talking to the kernel > > To reduce the probability of this issue, we can place key blocks with > high entropy in blocks 0 to 5. The idea is to place blocks that are often > changed in blocks 0 to 5, for example, key blocks that match on IPv4 > addresses or the LSBs of IPv6 addresses. Such placement will reduce the > probability of these blocks to be same. > > Mark several blocks with 'high_entropy' flag and place them in blocks 0 > to 5. Note that the list of the blocks is just a suggestion, I will verify > it with architects. > > Currently, there is a one loop that chooses which blocks should be used > for a given list of elements and fills the blocks - when a block is > chosen, it fills it in the region. To be able to control the order of > the blocks, separate between searching blocks and filling them. Several > pre-changes are required. > > Patch set overview: > Patch #1 marks several blocks with 'high_entropy' flag. > Patches #2-#4 prepare the code for filling blocks at the end of the search. > Patch #5 changes the loop to just choose the blocks and fill the blocks at > the end. > > Amit Cohen (5): > mlxsw: Mark high entropy key blocks > mlxsw: core_acl_flex_keys: Add a bitmap to save which blocks are > chosen > mlxsw: core_acl_flex_keys: Save chosen elements per block > mlxsw: core_acl_flex_keys: Save chosen elements in all blocks per > search > mlxsw: core_acl_flex_keys: Fill blocks with high entropy first > > .../mellanox/mlxsw/core_acl_flex_keys.c | 64 +++++++++++++++++-- > .../mellanox/mlxsw/core_acl_flex_keys.h | 9 +++ > .../mellanox/mlxsw/spectrum_acl_flex_keys.c | 12 ++-- > 3 files changed, 72 insertions(+), 13 deletions(-) For series, Reviewed-by: Simon Horman <horms@kernel.org>
Hello: This series was applied to netdev/net-next.git (main) by David S. Miller <davem@davemloft.net>: On Tue, 3 Oct 2023 13:25:25 +0200 you wrote: > Amit Cohen writes: > > For 12 key blocks in the A-TCAM, rules are split into two records, which > constitute two lookups. The two records are linked using a > "large entry key ID". > > Due to a Spectrum-4 hardware issue, KVD entries that correspond to key > blocks 0 to 5 of 12 key blocks will be placed in the same KVD pipe if they > only differ in their "large entry key ID", as it is ignored. This results > in a reduced scale, we can insert less than 20k filters and get an error: > > [...] Here is the summary with links: - [net-next,1/5] mlxsw: Mark high entropy key blocks https://git.kernel.org/netdev/net-next/c/cad6431b8675 - [net-next,2/5] mlxsw: core_acl_flex_keys: Add a bitmap to save which blocks are chosen https://git.kernel.org/netdev/net-next/c/0a67b7a0ec36 - [net-next,3/5] mlxsw: core_acl_flex_keys: Save chosen elements per block https://git.kernel.org/netdev/net-next/c/545535fd30dc - [net-next,4/5] mlxsw: core_acl_flex_keys: Save chosen elements in all blocks per search https://git.kernel.org/netdev/net-next/c/900f4285bbc2 - [net-next,5/5] mlxsw: core_acl_flex_keys: Fill blocks with high entropy first https://git.kernel.org/netdev/net-next/c/c01e24936d16 You are awesome, thank you!
Amit Cohen writes: For 12 key blocks in the A-TCAM, rules are split into two records, which constitute two lookups. The two records are linked using a "large entry key ID". Due to a Spectrum-4 hardware issue, KVD entries that correspond to key blocks 0 to 5 of 12 key blocks will be placed in the same KVD pipe if they only differ in their "large entry key ID", as it is ignored. This results in a reduced scale, we can insert less than 20k filters and get an error: $ tc -b flower.batch RTNETLINK answers: Input/output error We have an error talking to the kernel To reduce the probability of this issue, we can place key blocks with high entropy in blocks 0 to 5. The idea is to place blocks that are often changed in blocks 0 to 5, for example, key blocks that match on IPv4 addresses or the LSBs of IPv6 addresses. Such placement will reduce the probability of these blocks to be same. Mark several blocks with 'high_entropy' flag and place them in blocks 0 to 5. Note that the list of the blocks is just a suggestion, I will verify it with architects. Currently, there is a one loop that chooses which blocks should be used for a given list of elements and fills the blocks - when a block is chosen, it fills it in the region. To be able to control the order of the blocks, separate between searching blocks and filling them. Several pre-changes are required. Patch set overview: Patch #1 marks several blocks with 'high_entropy' flag. Patches #2-#4 prepare the code for filling blocks at the end of the search. Patch #5 changes the loop to just choose the blocks and fill the blocks at the end. Amit Cohen (5): mlxsw: Mark high entropy key blocks mlxsw: core_acl_flex_keys: Add a bitmap to save which blocks are chosen mlxsw: core_acl_flex_keys: Save chosen elements per block mlxsw: core_acl_flex_keys: Save chosen elements in all blocks per search mlxsw: core_acl_flex_keys: Fill blocks with high entropy first .../mellanox/mlxsw/core_acl_flex_keys.c | 64 +++++++++++++++++-- .../mellanox/mlxsw/core_acl_flex_keys.h | 9 +++ .../mellanox/mlxsw/spectrum_acl_flex_keys.c | 12 ++-- 3 files changed, 72 insertions(+), 13 deletions(-)