[dpdk-dev] [RFC PATCH] test/distributor: fix burst flush on worker quit

David Marchand david.marchand at redhat.com
Wed Apr 28 15:11:10 CEST 2021


On Mon, Apr 26, 2021 at 6:33 PM Stanislaw Kardach <kda at semihalf.com> wrote:
>
> While working on RISC-V port I have encountered a situation where worker
> threads get stuck in the rte_distributor_return_pkt() function in the
> burst test.
> After investigation some of the threads enter this function with
> flag RTE_DISTRIB_GET_BUF set in the d->retptr64[0]. At the same time
> main thread has already passed rte_distributor_process() so nobody will
> clear this flag and hence workers can't return.
>
> What I've noticed is that adding a flush just after the last _process(),
> similarly to how quit_workers() function is written in the
> test_distributor.c fixes the issue.
> Additionally the issue disappears when I remove the rdtsc delay code
> inside the rte_distributor_request_pkt().
> However I can't get this to reproduce on x86 (even with SIMD forced
> off) and with artificial delays, which is why I wonder whether I'm not
> actually hiding some other issue.
>
> Looking at the implementation of the distributor, it is based on
> __atomic_* builtins and the only platform related bit in the fast-path
> is the rte_rdtsc() and rte_pause(). There may be some issues in the
> toolchain (I've tried so far with the Ubuntu one - 10.2.0-8ubuntu1).
> I should add that all unit tests for distributor are passing so either
> there's some coverage corner case or the implementation works on RISC-V.
> As for RDTSC I'm using a sleep-stable time counter with 1MHz frequency
> and switching to high resolution cycle counter also removes the issue
> but that's the same as removing the rdtsc delay as mentioned above.
>
> I'd love to hear from You if this fix makes any sense.
>
> While modifying this test, I've also pulled in a fix from
> test_distributor.c which ensures that each thread gets his own wakeup
> packet as it's possible that when sending a burst of packets, they won't
> be spread over all the workers.

Do we have two fixes in a single patch?
This is ok if both point at the same originating commit, but else, we
need two patches with proper Fixes: line.

I see we have reviews, can you submit non-rfc patches?

Thanks.
-- 
David Marchand



More information about the dev mailing list