On 9/23/2021 1:33 PM, Loftus, Ciara wrote:
Doing basic operations like info_get or get_stats was broken
in af_xdp PMD. The info_get would crash because dev->device
was NULL in secondary process. Fix this by doing same initialization
as af_packet and tap devices.
The get_stats would crash because the XDP socket is not open in
primary process. As a workaround don't query kernel for dropped
packets when called from secondary process.
Note: this does not address the other bug which is that transmitting
in secondary process is broken because the send() in tx_kick
will fail because XDP socket fd is not valid in secondary process.
Hi Stephen,
Apologies for the delayed reply, I was on vacation.
In the Bugzilla report you suggest we:
"mark AF_XDP as broken in with primary/secondary
and return an error in probe in secondary process".
I agree with this suggestion. However with this patch we still permit
secondary, and just make sure it doesn't crash for get_stats. Did you
change
your mind?
Personally, I would prefer to have primary/secondary either working
100%
or else not allowed at all by throwing an error during probe. What do you
think? Do you have a reason/use case to permit secondary processes
despite
some features not being available eg. full stats, tx?
Thanks,
Ciara
There are two cases where secondary is useful even if send/receive can't
work from secondary process.
The pdump and proc-info applications can work with these patches.
I am using XDP over pdump as an easy way to get packets into the code
for
testing.
The flag in the documentation doesn't have a "limited" version.
If you want, will send another patch to disable secondary support.
Thanks for explaining. Since there are use cases for secondary, even if the
functionality is limited, I don't think it should be disabled.
Since we can't flag it as 'limited' in the feature matrix, could you please add
a note about the send/receive limitation in the AF_XDP PMD documentation
in a v2? There are already a number of limitations listed, which you can add
to.
Thanks,
Ciara
Supporting secondary, means adding a mechanism to pass the socket
around.
Looking at this in more detail, and my recommendation is:
For 21.11 release (and mark it as Fixes so it gets backported). Have the driver
return -ENOTSUP in secondary process.
For 22.03 add real secondary support using the rte_mp_msg to pass
necessary
state to secondary process. Includes socket (and other memory regions?).
The pdump and proc-info cases are only useful for developer testing, and
there are
other ways to do the same thing.
+1 that sounds reasonable to me.
This patch superseded by following which disables the secondary process support:
https://patches.dpdk.org/project/dpdk/patch/20210930134604.32585-1-ciara.lof...@intel.com/
