On 26-Jul-19 10:50 AM, Ananyev, Konstantin wrote:
-----Original Message-----
From: dev [mailto:dev-boun...@dpdk.org] On Behalf Of Bruce Richardson
Sent: Friday, July 26, 2019 10:40 AM
To: Burakov, Anatoly <anatoly.bura...@intel.com>
Cc: dev@dpdk.org; Thomas Monjalon <tho...@monjalon.net>
Subject: Re: [dpdk-dev] Should we disallow running secondaries after primary
has died?
On Fri, Jul 26, 2019 at 10:05:02AM +0100, Burakov, Anatoly wrote:
Hi all,
While investigating this bug:
https://bugs.dpdk.org/show_bug.cgi?id=284
I came across a realization that, when primary process dies, very little
actually works. There are some documented issues that are already present
when secondary processes keep running, like memory map becoming static, and
hotplug not working any more.
What is less known (and documented) is that VFIO also completely stops
working when initializing processes, because some time since 18.xx releases,
we've fixed a long standing VFIO-related bug that had to do with creating
new containers every time a secondary is run - secondary processes will now
reuse the primary process's container instead.
Meaning, for VFIO devices, secondary process *initialization* will fail
after primary process has died, because there is no longer a process from
which we can get the VFIO container from. Things will still sort-of work
with igb_uio or in vfio-noiommu mode, but again - no memory map updates, no
hotplug, potentially other things that i don't even know about.
Therefore, while ideally we would like people to have primary process always
running, the least we can do to avoid documenting a complex matrix of "what
is supported in which case" is to disallow secondary process initialization
after primary process has died.
("disallow" as in "explicitly document it as unsupported", although we can
also outright prevent it if we want - rte_eal_primary_proc_alive() will tell
us that)
Documenting this limitation seems a good thing to do. I'm not sure that
it's worthwhile trying to make the scenario (of running a secondary after a
primary has terminated) supported.
/Bruce
NP to disallow it.
In fact, I think it would be easier for everyone just to drop current DPDK MP
model,
and keep just standalone DPDK instances.
That's the dream, but i don't think it'll ever come to fruition, at
least not without a huge push from the community.
--
Thanks,
Anatoly
