> -----Original Message-----
> From: David Marchand <[email protected]>
> Sent: Tuesday, August 17, 2021 4:17 PM
> To: Joyce Kong <[email protected]>
> Cc: Burakov, Anatoly <[email protected]>; Olivier Matz
> <[email protected]>; Andrew Rybchenko
> <[email protected]>; Wang, Yipeng1
> <[email protected]>; Gobriel, Sameh <[email protected]>;
> Bruce Richardson <[email protected]>; Vladimir Medvedkin
> <[email protected]>; Ananyev, Konstantin
> <[email protected]>; Honnappa Nagarahalli
> <[email protected]>; Ruifeng Wang
> <[email protected]>; dev <[email protected]>; nd <[email protected]>; dpdk
> stable <[email protected]>; Aaron Conole <[email protected]>
> Subject: Re: [dpdk-dev] [PATCH v2] test/func_reentrancy: free memzones
> after creating test case
>
> On Sat, Jul 31, 2021 at 2:04 PM Joyce Kong <[email protected]> wrote:
> >
> > Function reentrancy test limits maximum number of iterations
> > simultaneously, however it doesn't free the 'fr_test_once'
> > memzones after the fact, so introduce freeing 'fr_test_once'
> > in ring/mempool/hash/fbk/lpm_clean.
> >
> > Meanwhile, add the missing free for test case on main thread.
> >
> > Fixes: 104a92bd026f ("app: add reentrancy tests")
> > Fixes: 995eec619024 ("test: clean up memory for function reentrancy
> > test")
> > Cc: [email protected]
> >
> > Signed-off-by: Joyce Kong <[email protected]>
> > Reviewed-by: Ruifeng Wang <[email protected]>
> > Reviewed-by: Feifei Wang <[email protected]>
>
> This patch actually breaks the test (we are lucky, the failure happens often
> ;)).
>
> 28/94 DPDK:fast-tests / func_reentrancy_autotest FAIL
> 0.22s (exit status 255 or signal 127 SIGinvalid)
>
> --- command ---
> 16:13:45 DPDK_TEST='func_reentrancy_autotest'
> /home-local/jenkins-local/jenkins-agent/workspace/Generic-Unit-Test-
> DPDK/dpdk/build/app/test/dpdk-test
> -l 0-15 --file-prefix=func_reentrancy_autotest
> --- stdout ---
> RTE>>func_reentrancy_autotest
> Func-ReEnt CASE 0: eal init once PASS
> ring create/lookup: common object allocated 2 times (should be 1) Func-
> ReEnt CASE 1: ring create/lookup FAIL Test Failed
> RTE>>
> --- stderr ---
>
>
> I guess, this is what happens:
>
> main lcore worker lcore 1 ...
> worker lcore X
> enters ring_create_lookup()
>
> enters ring_create_lookup()
> rte_eal_wait_lcore(worker lcore 1);
> leaves ring_create_lookup()
> ring_clean(worker lcore 1);
>
> leaves ring_create_lookup()
>
> There is no synchronisation point for the main lcore to know the worker
> lcores are finished invoking the func callback.
> With this patch, the "common" object is freed by the main lcore
> *potentially* before some workers start trying to create it.
> And we end up with multiple workers successfully creating this object, hence
> the obj_count being incremented.
>
>
> --
> David Marchand
I think add rte_eal_mp_wait_lcore() like below can ensure the lcores to free
objects
after all func callback finished.
Shall do the change in next version.
RTE_LCORE_FOREACH_WORKER(lcore_id) {
if (cores == 1)
break;
cores--;
rte_eal_remote_launch(pt_case->func, pt_case->arg, lcore_id);
}
rte_atomic32_set(&synchro, 1);
if (pt_case->func(pt_case->arg) < 0)
ret = -1;
+ rte_eal_mp_wait_lcore();
cores = cores_save;
RTE_LCORE_FOREACH_WORKER(lcore_id) {
if (cores == 1)
break;
cores--;
- if (rte_eal_wait_lcore(lcore_id) < 0)
- ret = -1;
if (pt_case->clean != NULL)
pt_case->clean(lcore_id);
}
