A quick follow-up: I now see that 'self.queue' is in fact a misnomer for an
array of queues and not a single queue index specification. It would be better
to name it 'queues'. But the issue stands: this array differs from GLOBAL table.
As far as I can tell from reading documentation for action 'rss', it does not
specifically assume a per-flow 'reta' to be derived from the queue array. If the
meaning of queue indices in the action is solely to give a subset of queues
where a packet can land, but not in a strict, 'reta'-like manner, then the
assumption of the test that the GLOBAL 'reta' should be consulted is still odd.
It in fact checks where the the packet would have landed were it not for the
flow rule to intercept it. Once again, I apologise in case I've got this wrong.
Thank you.
On Fri, 18 Jul 2025, Ivan Malov wrote:
Hi Thomas,
On Fri, 18 Jul 2025, Thomas Wilks wrote:
Hi all,
This is v3 of the RSS test suites which has been rebased onto the latest
patches.
I’m also requesting comments and some help with an issue where the packet
RSS queue and
the predicted RSS queue differ when running on ConnectX-6 NICs while being
the same on E810-C NICs.
Looking at patch [2/2], 'test_key_hash_default_hash_algorithm_queues' might
do
the following:
1) insert a flow rule with action 'rss' and a specific/single random queue
"A",
which is expressed as 'rss types ipv4-udp end queues {self.queue} end';
2) set the GLOBAL rss table ('reta') for the port, with the queue indices
going in some random order (say, [ACBBAAADCDAA..C], of size 512);
3) send/receive the packet;
4) based on the packet mbuf hash value, offset into the GLOBAL table ('reta')
and retrieve the "expected" queue index;
5) compare the actual receive queue index with the predicted one.
Is this understanding correct?
If my understanding is correct, then I'm afraid this design might be flawed.
Why does it assume the GLOBAL table effect on action 'rss" with some specific
queue index listed in it? In fact, the 'queue' array of the action, at least,
when non-empty, defines exact subset of queues where the packet can land,
which
might effectively form a distinct, per-flow 'reta'. In this case, where only
one
queue index "A" is specified, the effective 'reta' for this particular flow
match is going to be [AAAAAAAAA..A] (say, of size 512 by default), isn't it?
If so, then why does the code try to peek at the GLOBAL reta instead, when it
has in fact wittingly reduced action 'rss' behaviour to such of action
'queue'?
Once again, I apologise in case I've got something wrong.
I’m currently in the process of refactoring the test suites and addressing
the review comments from v2. Below is a summary of the changes made so far
that are included in this version:
As a separate question, doesn't one also want to verify the "expected" hash
value (SW hash based on the IP/UDP 2/4-tuple) versus the "actual" value?
Or is this too inconvenient to automate in testpmd environment?
Thank you.
Changes in v3:
- Merged the RSS test suites into a single file for better maintainability.
- Moved the supporting functions into the same test suite file and removed
the previous separate function file.
- Combined the reta_key_reta_queues test cases into a single test case.
- Added new test cases to support flow rule creation on ConnectX-6 NICs.
- Removed requirements for specific hashing algorithms which has been
replaced the flow_validate function.
Regards,
Thomas Wilks
Alex Chapman (1):
dts: add RSS functions to testpmd
Thomas Wilks (1):
dts: add PMD RSS testsuite
dts/framework/remote_session/testpmd_shell.py | 132 +++-
dts/tests/TestSuite_pmd_rss.py | 610 ++++++++++++++++++
2 files changed, 729 insertions(+), 13 deletions(-)
create mode 100644 dts/tests/TestSuite_pmd_rss.py
--
2.43.0
