Re: [PATCH] PCI: hv: Only reuse existing IRTE allocation for Multi-MSI

From: Lorenzo Pieralisi
Date: Wed Aug 17 2022 - 03:51:43 EST


On Tue, Aug 16, 2022 at 09:13:26PM +0000, Dexuan Cui wrote:
> > From: Bjorn Helgaas <helgaas@xxxxxxxxxx>
> > Sent: Tuesday, August 16, 2022 8:51 AM
> > To: Dexuan Cui <decui@xxxxxxxxxxxxx>
> >
> > This has only observations with no explanations, and I don't see how
> > it will be useful to future readers of the git history.
> Please see the below.
>
> > I assume you bisected the problem to b4b77778ecc5?
> Yes.
>
> > Can you just revert that? A revert requires no more explanation than
> > "this broke something."
>
> It's better to not revert b4b77778ecc5, which is required by Jeff's
> Multi-MSI device, which doesn't seem to be affected by the interrupt
> issue I described.

You must debug it, there are no two ways about it.

We can't apply fixes on a hunch, more so given that I am not convinced
at all this patch is fixing anything, it is just papering over an
underlying bug that is still to be pinpointed.

> > I guess this is a fine distinction, but I really don't like random
> > code changes that "seem to avoid a problem but we don't know how."
> > A revert at least has the advantage that we can cover our eyes and
> > pretend the commit never happened. This patch feels like future
> > readers will have to try to understand the code even though we
> > clearly don't understand why it makes a difference.
>
> I just replied to Lorenzo's email with more details. FYI, this is the link
> to my reply:
> https://lwn.net/ml/linux-kernel/SA1PR21MB1335D08F987BBAE08EADF010BF6B9@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
>
> I just felt the commit message might be too long if I had put all the
> details there. :-) Can we add a Links: tag?

Commit logs must describe the issue you are fixing, thouroughly and
concisely. To start with "Jeffrey's 4 recent patches" is a very bad
start for a commit log, it means nothing, try to read your log as
someone who needs to understand the commit years down the line please.

Now, back to this patch: we are at -rc1, unless Bjorn is willing to
do so I am not inclined to apply this patch till next merge window
(and actually I am not inclined to merge it at all).

This gives you folks time to debug it (and it must be debugged), the
fact that it works for one multi-MSI device does not mean that the
bug isn't still there - I am worried that the issue is with
b4b77778ecc5 and the interaction with core MSI/IOMMU.

Thanks,
Lorenzo