On Thu, Aug 01, 2019 at 02:29:21PM +0200, David Marchand wrote:
> On Thu, Aug 1, 2019 at 2:28 PM David Marchand <[email protected]>
> wrote:
> >
> > The memory tests currently check that, for normal mode (not legacy mode),
> > there is no memory left behind when exiting.
> >
> > The problem is that if a ethdev port is allocated when scanning pci
> > devices (even if the driver probe fails like when you have a virtio
> > management interface attached to the kernel), on exit, dpdk won't free
> > the associated memory since ethdev never frees the ethdev memzone.
> >
> > Workaround this by disabling pci scan.
>
> Not entirely happy with this patch.
> I am open to suggestions :-)
>
Hello David,
Why not cleanup on .fini the ethdev subsystem?
There is RTE_UNREGISTER_CLASS() that was added just for that. I
think I remember someone (maybe Thomas) not liking it. Why are we
keeping the memzones allocated on exit? Debug ? Primary/secondary snafu?
I would take this opportunity to troll once again about the default
blacklist-mode of the PCI bus, which is still an issue for downstream
consumers of DPDK and seems to be kept only to avoid changing the CI
infrastructure of upstream contributors. Unfortunately the root-cause
might be elsewhere.
> >
> > Fixes: 651cc78f83b5 ("test: fix hugepage file handling in EAL flags
> > autotest")
> > Fixes: 690fd3577e90 ("test/eal: add cases for in-memory and
> > single-file-segments")
> > Cc: [email protected]
>
> And we might want to drop stable.
>
>
> --
> David Marchand
--
Gaëtan Rivet
6WIND
