According to the discussion in this thread: https://protect2.fireeye.com/v1/url?k=24ecce33-7b77f77c-24ed457c-0cc47a30 d446-e317a6beb8cfa273&q=1&e=f8bb12df-3698-4bce-a7b7-d72e22b91431&u=https% 3A%2F%2Finbox.dpdk.org%2Fdev%2FCAOE1vsOehF4ZMOWffpEv%3DQF6YOc5wXtg23PV83B 9CLiTMn8wQA%40mail.gmail.com%2F%23r I was able to reproduce the distributor test failure in the exactly same way as described, but on x86_64 machine with 32 cores. So it does not seem to be the problem related to the ARM architecture. IMO issue occurs when there are many worker threads returning at the same time packets. I was not able to observe the issue on ARM devices, but I used only machines with 4 cores. So that is max 3 worker cores, so maximum of 32*3 = 96 packets processed at the same time which is less than 127 , so the issue cannot occur. Can anyone verify this patch on a machine similar to one used in CI lab, on which the issue occurred? Lukasz Wojciechowski (1): test/distributor: prevent return buffer overload app/test/test_distributor.c | 20 +++++++++++--------- 1 file changed, 11 insertions(+), 9 deletions(-) -- 2.17.1
The distributor library implementation uses a cyclic queue to store packets returned from workers. These packets can be later collected with rte_distributor_returned_pkts() call. However the queue has limited capacity. It is able to contain only 127 packets (RTE_DISTRIB_RETURNS_MASK). Big burst tests sent 1024 packets in 32 packets bursts without waiting until they are processed by the distributor. In case when tests were run with big number of worker threads, it happened that more than 127 packets were returned from workers and put into cyclic queue. This caused packets to be dropped by the queue, making them impossible to be collected later with rte_distributor_returned_pkts() calls. However the test waited for all packets to be returned infinitely. This patch fixes the big burst test by not allowing more than queue capacity packets to be processed at the same time, making impossible to drop any packets. It also cleans up duplicated code in the same test. Bugzilla ID: 612 Fixes: c0de0eb82e40 ("distributor: switch over to new API") Cc: david.hunt@intel.com Cc: stable@dpdk.org Signed-off-by: Lukasz Wojciechowski <l.wojciechow@partner.samsung.com> --- app/test/test_distributor.c | 20 +++++++++++--------- 1 file changed, 11 insertions(+), 9 deletions(-) diff --git a/app/test/test_distributor.c b/app/test/test_distributor.c index f4c6229f1..961f326cd 100644 --- a/app/test/test_distributor.c +++ b/app/test/test_distributor.c @@ -217,6 +217,8 @@ sanity_test(struct worker_params *wp, struct rte_mempool *p) clear_packet_count(); struct rte_mbuf *many_bufs[BIG_BATCH], *return_bufs[BIG_BATCH]; unsigned num_returned = 0; + unsigned int num_being_processed = 0; + unsigned int return_buffer_capacity = 127;/* RTE_DISTRIB_RETURNS_MASK */ /* flush out any remaining packets */ rte_distributor_flush(db); @@ -233,16 +235,16 @@ sanity_test(struct worker_params *wp, struct rte_mempool *p) for (i = 0; i < BIG_BATCH/BURST; i++) { rte_distributor_process(db, &many_bufs[i*BURST], BURST); - count = rte_distributor_returned_pkts(db, - &return_bufs[num_returned], - BIG_BATCH - num_returned); - num_returned += count; + num_being_processed += BURST; + do { + count = rte_distributor_returned_pkts(db, + &return_bufs[num_returned], + BIG_BATCH - num_returned); + num_being_processed -= count; + num_returned += count; + rte_distributor_flush(db); + } while (num_being_processed + BURST > return_buffer_capacity); } - rte_distributor_flush(db); - count = rte_distributor_returned_pkts(db, - &return_bufs[num_returned], - BIG_BATCH - num_returned); - num_returned += count; retries = 0; do { rte_distributor_flush(db); -- 2.17.1
On Tue, Jan 19, 2021 at 4:59 AM Lukasz Wojciechowski
<l.wojciechow@partner.samsung.com> wrote:
>
> According to the discussion in this thread:
> https://protect2.fireeye.com/v1/url?k=24ecce33-7b77f77c-24ed457c-0cc47a30
> d446-e317a6beb8cfa273&q=1&e=f8bb12df-3698-4bce-a7b7-d72e22b91431&u=https%
> 3A%2F%2Finbox.dpdk.org%2Fdev%2FCAOE1vsOehF4ZMOWffpEv%3DQF6YOc5wXtg23PV83B
> 9CLiTMn8wQA%40mail.gmail.com%2F%23r
>
> I was able to reproduce the distributor test failure in the exactly same
> way as described, but on x86_64 machine with 32 cores.
> So it does not seem to be the problem related to the ARM architecture.
> IMO issue occurs when there are many worker threads returning at the same
> time packets.
>
> I was not able to observe the issue on ARM devices, but I used only
> machines with 4 cores. So that is max 3 worker cores,
> so maximum of 32*3 = 96 packets processed at the same time
> which is less than 127 , so the issue cannot occur.
>
> Can anyone verify this patch on a machine similar to one used in CI lab,
> on which the issue occurred?
Thanks for looking at it, Lukasz.
Unfortunately, I can't reproduce it on my x86 system (26 workers in
the test) and I don't have a ARM machine.
--
David Marchand
Thank you David,
If you have the possibility you can try on some emulated virtual
machine, where cores are much slower, so the workers don't return
packages immediately.
It reproduces in 100% cases in such environment.
Best regards
Lukasz
W dniu 19.01.2021 o 09:44, David Marchand pisze:
> On Tue, Jan 19, 2021 at 4:59 AM Lukasz Wojciechowski
> <l.wojciechow@partner.samsung.com> wrote:
>> According to the discussion in this thread:
>> https://protect2.fireeye.com/v1/url?k=24ecce33-7b77f77c-24ed457c-0cc47a30
>> d446-e317a6beb8cfa273&q=1&e=f8bb12df-3698-4bce-a7b7-d72e22b91431&u=https%
>> 3A%2F%2Finbox.dpdk.org%2Fdev%2FCAOE1vsOehF4ZMOWffpEv%3DQF6YOc5wXtg23PV83B
>> 9CLiTMn8wQA%40mail.gmail.com%2F%23r
>>
>> I was able to reproduce the distributor test failure in the exactly same
>> way as described, but on x86_64 machine with 32 cores.
>> So it does not seem to be the problem related to the ARM architecture.
>> IMO issue occurs when there are many worker threads returning at the same
>> time packets.
>>
>> I was not able to observe the issue on ARM devices, but I used only
>> machines with 4 cores. So that is max 3 worker cores,
>> so maximum of 32*3 = 96 packets processed at the same time
>> which is less than 127 , so the issue cannot occur.
>>
>> Can anyone verify this patch on a machine similar to one used in CI lab,
>> on which the issue occurred?
> Thanks for looking at it, Lukasz.
> Unfortunately, I can't reproduce it on my x86 system (26 workers in
> the test) and I don't have a ARM machine.
>
--
Lukasz Wojciechowski
Principal Software Engineer
Samsung R&D Institute Poland
Samsung Electronics
Office +48 22 377 88 25
l.wojciechow@partner.samsung.com
On Tue, Jan 19, 2021 at 2:07 PM Lukasz Wojciechowski
<l.wojciechow@partner.samsung.com> wrote:
>
> Thank you David,
> If you have the possibility you can try on some emulated virtual
> machine, where cores are much slower, so the workers don't return
> packages immediately.
> It reproduces in 100% cases in such environment.
I reproduced the issue with starting a testpmd on the same cores in this system.
I usually reproduce it after 1-2 minutes of continuously running the
distributor_autotest unit test.
I've applied your fix in my tree and I will let this loop run for a while.
--
David Marchand
On Tue, Jan 19, 2021 at 4:59 AM Lukasz Wojciechowski <l.wojciechow@partner.samsung.com> wrote: > > The distributor library implementation uses a cyclic queue to store > packets returned from workers. These packets can be later collected > with rte_distributor_returned_pkts() call. > However the queue has limited capacity. It is able to contain only > 127 packets (RTE_DISTRIB_RETURNS_MASK). > > Big burst tests sent 1024 packets in 32 packets bursts without waiting > until they are processed by the distributor. In case when tests were > run with big number of worker threads, it happened that more than > 127 packets were returned from workers and put into cyclic queue. > This caused packets to be dropped by the queue, making them impossible > to be collected later with rte_distributor_returned_pkts() calls. > However the test waited for all packets to be returned infinitely. > > This patch fixes the big burst test by not allowing more than > queue capacity packets to be processed at the same time, making > impossible to drop any packets. > It also cleans up duplicated code in the same test. > > Bugzilla ID: 612 > Fixes: c0de0eb82e40 ("distributor: switch over to new API") > Cc: david.hunt@intel.com > Cc: stable@dpdk.org > > Signed-off-by: Lukasz Wojciechowski <l.wojciechow@partner.samsung.com> Pasting from my reply to the cover letter: > I reproduced the issue with starting a testpmd on the same cores in this system. > I usually reproduce it after 1-2 minutes of continuously running the > distributor_autotest unit test. > > I've applied your fix in my tree and I will let this loop run for a while. This has been running fine for more than 30 minutes on my x86 28 cores system. Tested-by: David Marchand <david.marchand@redhat.com> -- David Marchand
Hi Lukasz,
On 19/1/2021 3:59 AM, Lukasz Wojciechowski wrote:
> The distributor library implementation uses a cyclic queue to store
> packets returned from workers. These packets can be later collected
> with rte_distributor_returned_pkts() call.
> However the queue has limited capacity. It is able to contain only
> 127 packets (RTE_DISTRIB_RETURNS_MASK).
>
> Big burst tests sent 1024 packets in 32 packets bursts without waiting
> until they are processed by the distributor. In case when tests were
> run with big number of worker threads, it happened that more than
> 127 packets were returned from workers and put into cyclic queue.
> This caused packets to be dropped by the queue, making them impossible
> to be collected later with rte_distributor_returned_pkts() calls.
> However the test waited for all packets to be returned infinitely.
>
> This patch fixes the big burst test by not allowing more than
> queue capacity packets to be processed at the same time, making
> impossible to drop any packets.
> It also cleans up duplicated code in the same test.
>
> Bugzilla ID: 612
> Fixes: c0de0eb82e40 ("distributor: switch over to new API")
> Cc: david.hunt@intel.com
> Cc: stable@dpdk.org
>
> Signed-off-by: Lukasz Wojciechowski <l.wojciechow@partner.samsung.com>
> ---
This patch cleans up the code nicely, and it makes sense to return
packets in the do..while. LGTM.
Reviewed-by: David Hunt <david.hunt@intel.com>
On Thu, Jan 28, 2021 at 3:10 PM David Marchand
<david.marchand@redhat.com> wrote:
>
> On Tue, Jan 19, 2021 at 4:59 AM Lukasz Wojciechowski
> <l.wojciechow@partner.samsung.com> wrote:
> >
> > The distributor library implementation uses a cyclic queue to store
> > packets returned from workers. These packets can be later collected
> > with rte_distributor_returned_pkts() call.
> > However the queue has limited capacity. It is able to contain only
> > 127 packets (RTE_DISTRIB_RETURNS_MASK).
> >
> > Big burst tests sent 1024 packets in 32 packets bursts without waiting
> > until they are processed by the distributor. In case when tests were
> > run with big number of worker threads, it happened that more than
> > 127 packets were returned from workers and put into cyclic queue.
> > This caused packets to be dropped by the queue, making them impossible
> > to be collected later with rte_distributor_returned_pkts() calls.
> > However the test waited for all packets to be returned infinitely.
> >
> > This patch fixes the big burst test by not allowing more than
> > queue capacity packets to be processed at the same time, making
> > impossible to drop any packets.
> > It also cleans up duplicated code in the same test.
> >
> > Bugzilla ID: 612
> > Fixes: c0de0eb82e40 ("distributor: switch over to new API")
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Lukasz Wojciechowski <l.wojciechow@partner.samsung.com>
> Tested-by: David Marchand <david.marchand@redhat.com>
Reviewed-by: David Hunt <david.hunt@intel.com>
Applied, thanks Lukasz.
This should fix the issue seen at UNH on the ARM server.
--
David Marchand
Thank you guys!
W dniu 29.01.2021 o 09:03, David Marchand pisze:
> On Thu, Jan 28, 2021 at 3:10 PM David Marchand
> <david.marchand@redhat.com> wrote:
>> On Tue, Jan 19, 2021 at 4:59 AM Lukasz Wojciechowski
>> <l.wojciechow@partner.samsung.com> wrote:
>>> The distributor library implementation uses a cyclic queue to store
>>> packets returned from workers. These packets can be later collected
>>> with rte_distributor_returned_pkts() call.
>>> However the queue has limited capacity. It is able to contain only
>>> 127 packets (RTE_DISTRIB_RETURNS_MASK).
>>>
>>> Big burst tests sent 1024 packets in 32 packets bursts without waiting
>>> until they are processed by the distributor. In case when tests were
>>> run with big number of worker threads, it happened that more than
>>> 127 packets were returned from workers and put into cyclic queue.
>>> This caused packets to be dropped by the queue, making them impossible
>>> to be collected later with rte_distributor_returned_pkts() calls.
>>> However the test waited for all packets to be returned infinitely.
>>>
>>> This patch fixes the big burst test by not allowing more than
>>> queue capacity packets to be processed at the same time, making
>>> impossible to drop any packets.
>>> It also cleans up duplicated code in the same test.
>>>
>>> Bugzilla ID: 612
>>> Fixes: c0de0eb82e40 ("distributor: switch over to new API")
>>> Cc: stable@dpdk.org
>>>
>>> Signed-off-by: Lukasz Wojciechowski <l.wojciechow@partner.samsung.com>
>> Tested-by: David Marchand <david.marchand@redhat.com>
> Reviewed-by: David Hunt <david.hunt@intel.com>
>
> Applied, thanks Lukasz.
>
> This should fix the issue seen at UNH on the ARM server.
>
--
Lukasz Wojciechowski
Principal Software Engineer
Samsung R&D Institute Poland
Samsung Electronics
Office +48 22 377 88 25
l.wojciechow@partner.samsung.com