Re: [PATCH] nvme: Fix nvme_setup_command metadata trace event for cdw10

[Keith Busch]

On Wed, Jul 06, 2022 at 06:34:34PM +0200, Christoph Hellwig wrote:
> On Wed, Jul 06, 2022 at 10:26:09AM -0600, Keith Busch wrote:
> > On Wed, Jul 06, 2022 at 06:18:25PM +0200, Christoph Hellwig wrote:
> > > On Wed, Jul 06, 2022 at 10:13:22AM -0600, Keith Busch wrote:
> > > > Did you test what the trace looks like afte this? We're losing valuable trace
> > > > data here. The field is supposed to get CDW's 10 - 15, so that's 24 bytes. I
> > > > don't know why it cares that the address of the field being read is only 4
> > > > bytes; we want everything that comes after it too.
> > > 
> > > Because accesses should not spawn boundaries of members in structs unless
> > > copying the entire struct.  If we want to trace the various fields we
> > > need to individually assign them.
> > > 
> > > Anyway, I'm dropping this patch from nvme-5.19 for now to let the
> > > discussion conclude.
> > 
> > How about this instead?
> 
> Maybe a better option would be to use struct_group().

Good call, I'd never used that macro before. The result produces anonymous
unions like I just proposed, so yes, I like that option.