Re: [PATCH nf-next 0/2] nf_tables: avoid retpoline overhead on set lookups

[Stefano Brivio <sbrivio@redhat.com>]

Hi Florian,

On Thu, 13 May 2021 22:29:54 +0200
Florian Westphal <fw@strlen.de> wrote:

> This adds a nft_set_do_lookup() helper, then extends it to use
> direct calls when RETPOLINE feature is enabled.
> 
> For non-retpoline builds, nft_set_do_lookup() inline helper
> does a indirect call.  INDIRECT_CALLABLE_SCOPE macro allows to
> keep the lookup functions static in this case.

Thanks for doing this! And sorry I looked into it more than one year
ago without ever finishing it ;)

I ran some quick tests, I was curious to see the impact of dropping
indirect calls on that path. With the 'performance' test cases of
nft_concat_range.sh, roughly estimating clock cycles as clock frequency
divided by packet rate, it looks like this offsets entirely the usage of
retpolines!

With a 'return true;' in the lookup function (I patched nft_set_pipapo),
on my usual single AMD Epyc 7351 thread, 2.9GHz, average of three runs,
I get:

                                               | packet |  est.  |
                                               |  rate  | cycles |
                                               | (Mpps) |        |
-----------------------------------------------|--------|--------|
Without retpolines, netdev drop                | 15.443 |    188 |
Without retpolines, dummy lookup function      |  9.995 |    292 |
-> Without retpolines, set lookup              |        |    104-|-.
- - - - - - - - - - - - - - - - - - - - - - - -|- - - - | - - - -|
With retpolines, netdev drop                   | 10.420 |    278 | |
With retpolines, dummy lookup function         |  7.038 |    412 |
-> With retpolines, set lookup                 |        |    134 | |
- - - - - - - - - - - - - - - - - - - - - - - -|- - - - | - - - -|
This series, retpolines, netdev drop           | 10.431 |    278 | |
This series, retpolines, dummy lookup function |  7.549 |    384 |
-> This series, retpolines, set lookup         |  ^ +7% |    106-|-'

estimated clock cycles for set lookup only are the difference between
cycles to hit the dummy lookup function and cycles to drop packets from
the netdev hook -- they're now approximately the same with and without
retpolines.

For context, I also ran the whole set of tests with actual matching.
This is indicative, just a single run:

 --------------.-----------------------------------.--------------------------.
AMD Epyc 7351  |          baselines, Mpps          |       this series        |
 1 thread      |___________________________________|__________________________|
 2.9GHz        |        |        |        |        |        |        |        |
 512KiB L1D$   | netdev |  hash  | rbtree |        |  hash  | rbtree |        |
 --------------|  hook  |   no   | single |        |   no   | single |        |
type   entries |  drop  | ranges | field  | pipapo | ranges | field  | pipapo |
 --------------|--------|--------|--------|--------|--------|-----------------|
net,port       |        |        |        |        | +15%   |  +4%   |  +4%   |
         1000  |   10.1 |    5.2 |    2.7 |    4.6 |    6.0 |    2.8 |    4.8 |
 --------------|--------|--------|--------|--------|--------|--------|--------|
port,net       |        |        |        |        | +11%   |  +5%   |  +4%   |
          100  |   10.4 |    5.4 |    4.1 |    5.0 |    6.0 |    4.3 |    5.2 |
 --------------|--------|--------|--------|--------|--------|--------|--------|
net6,port      |        |        |        |        | +15%   |  +9%   |  +6%   |
         1000  |   10.0 |    4.6 |    1.1 |    3.1 |    9.9 |    1.2 |    3.3 |
 --------------|--------|--------|--------|--------|--------|--------|--------|
port,proto     |        |        |        |        |  +7%   |  +3%   |  +3%   |
        10000  |   10.7 |    6.0 |    3.0 |    3.0 |    6.4 |    3.1 |    3.1 |
 --------------|--------|--------|--------|--------|--------|--------|--------|
net6,port,mac  |        |        |        |        |  +3%   |  +4%   |  +3%   |
           10  |    9.9 |    3.8 |    2.7 |    3.3 |    3.9 |    2.8 |    3.4 |
 --------------|--------|--------|--------|--------|--------|--------|--------|
net6,port,mac, |        |        |        |        |  +3%   |  +9%   |  +4%   |
proto    1000  |   10.0 |    3.6 |    1.1 |    2.4 |    3.7 |    1.2 |    2.5 |
 --------------|--------|--------|--------|--------|--------|--------|--------|
net,mac        |        |        |        |        |  +6%   |  +4%   |  +3%   |
         1000  |   10.5 |    4.8 |    2.7 |    4.0 |    5.1 |    2.8 |    4.1 |
 --------------'--------'--------'--------'--------'--------'--------'--------'

-- 
Stefano