On 2024/06/06 16:59, Daniel P. Berrangé wrote:
On Thu, Jun 06, 2024 at 04:19:11PM +0900, Akihiko Odaki wrote:
On 2024/06/06 16:14, Daniel P. Berrangé wrote:
On Thu, Jun 06, 2024 at 05:14:20AM +0900, Akihiko Odaki wrote:
On 2024/06/05 19:23, Daniel P. Berrangé wrote:
On Tue, Jun 04, 2024 at 03:37:42PM +0800, Jason Wang wrote:
From: Akihiko Odaki <akihiko.od...@daynix.com>
Propagating ebpf-rss-fds errors has several problems.
First, it makes device realization fail and disables the fallback to the
conventional eBPF loading.
AFAICT, this is not a bug - this is desired behaviour.
If the user/mgmt app has told QEMU to use FDs it has passed
in, then any failure to do this *MUST* be treated as a fatal
error. Falling back to other codepaths is ignoring a direct
user request.
The FD options are more like an assistance rather than a request. When QEMU
does not have a permission to load eBPF programs, a user can get the eBPF
programs with the request-ebpf command of QMP, load it, and pass the FDs to
QEMU.
That still doesn't alter the fact that if the user has chosen to pass FDs
and QEMU fails to use them, it *MUST* report that error back to the user.
The user should be more interested in whether the eBPF functionality is
successfully enabled or not, and that is irrelevant from whether the eBPF
program is loaded by QEMU or someone else.
No, this is wrong. A mgmt application or user will have made a decision
about *how* it wants QEMU to configure a particular feature. QEMU must
always honour the mgmt application's request, and not try to do something
different.
If the mgmt app did not want the FDs to be used, it would not have
passed them to QEMU in the first place. Ignoring the FDs is not likely
to work, because QEMU is unlikely to have permission to open the FDs
itself.
Ignoring the errors when creating the FDs, makes it much much harder
to detect and diagnose deployment problems, because the root cause
error is being discarded, and replaced by a later error which misleads
the app managing QEMU.
Always honouring the user requested config, or giving an error back
when it fails, is standard QEMU practice.
I see.
I'll append a follow-up patch to the series "[PATCH 0/3] virtio-net:
Convert feature properties to OnOffAuto" to remove the fallback path. We
can keep this for now to remove the flawed error handling code.
Regards,
Akihiko Odaki
