Re: + fix-zero-copy-i-o-on-__get_user_pages-allocated-pages.patch added to mm-hotfixes-unstable branch

[Jason Gunthorpe <jgg@ziepe.ca>]

On Thu, May 08, 2025 at 09:11:36PM +0300, Pantelis Antoniou wrote:

> 3. While we go about fixing it, this has caused a pretty significant
> userspace regression, where the address space that those BOs reside
> cannot be used for I/O when a network filesystem is involved. I think
> it's a matter of time when regular filesystem start using the same
> method of pining and doing I/O instead of using the filecache on fast
> memory mediums.

Regular file systems already uses GUP on O_DIRECT paths and already
didn't work basically forever. It seems like a kernel bug in the net
filesystems to have done something different in their O_DIRECT for so
long :\

Anyhow, the fixes must come from DRM using the mm properly, not by
hacking up the mm to ignore the well defined API rules we have.

I don't know enough about DRM to say exactly what that means, but that
is where you should be focusing your attention to fix it.

Somehow I suspect the number of places actually using O_DIRECT from a
network filesystem to a DRM buffer is going to be pretty small since
it never worked on a normal filesystem. Meaning this isn't some
general common code that is feeding generic files into GPUs, but
something custom built to only use a network that happened to stumble
onto this kernel bug and abuse it. Do you know differently?

Jason