On Fri, Aug 20, 2021 at 9:57 AM Joyce Kong [off-list ref] wrote:

quoted

-----Original Message-----
From: David Marchand <redacted>
Sent: Tuesday, August 17, 2021 4:17 PM
To: Joyce Kong <redacted>
Cc: Burakov, Anatoly <redacted>; Olivier Matz
[off-list ref]; Andrew Rybchenko
[off-list ref]; Wang, Yipeng1
[off-list ref]; Gobriel, Sameh [off-list ref];
Bruce Richardson [off-list ref]; Vladimir Medvedkin
[off-list ref]; Ananyev, Konstantin
[off-list ref]; Honnappa Nagarahalli
[off-list ref]; Ruifeng Wang
[off-list ref]; dev [off-list ref]; nd [off-list ref]; dpdk
stable [off-list ref]; Aaron Conole [off-list ref]
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 [off-list ref] wrote:

quoted

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: stable@dpdk.org

Signed-off-by: Joyce Kong <redacted>
Reviewed-by: Ruifeng Wang <redacted>
Reviewed-by: Feifei Wang <redacted>

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);
        }

Using mp_wait_lcore, the test can't tell if a lcore returned an error
after executing the passed callback.

An alternative is to split the current loop to first have the per
lcore rte_eal_wait_lcore() calls + ret code check, and then a second
loop calls the clean() callback.


-- 
David Marchand