Re: [PATCH blktests] create a test for direct io offsets

[Keith Busch <kbusch@kernel.org>]

On Fri, Oct 17, 2025 at 04:13:32AM -0700, Christoph Hellwig wrote:
> > +static void init_kernel_version()
> > +{
> > +	struct utsname buffer;
> > +	char *major_version_str;
> > +	char *minor_version_str;
> > +	if (uname(&buffer) != 0)
> > +		err(errno, "uname");
> > +
> > +	major_version_str = strtok(buffer.release, ".");
> > +	minor_version_str = strtok(NULL, ".");
> > +
> > +	kernel_major = strtol(major_version_str, NULL, 0);
> > +	kernel_minor = strtol(minor_version_str, NULL, 0);
> > +}
> 
> Testing for specific kernel versions is probably going to fall
> flat when this stuff gets backported..  I just realize that maybe
> we just need a statx / fsxattr flag to report that this is supported
> to make everyones life easier?  statx probably won't make Christian
> happy, but now that the fsxattr stuff is in common code that seems
> like an easy enough option to rush into 6.18 still.

In addition to a flag to say direct-io can use block data split among
multiple vectors, there's two currently unreported attributes you need
to know in order to successfuly use direct-io like this: max sectors and
virtual boundary mask. The max sectors attribute is needed to know
exactly how many vectors you can insert before the total length needs to
add up to a block size.

Adding these to statx or fsxattr will help remove a ton of crap from
this test since I could remove all the stuff extracting it out of sysfs
from any given file.