On Tue, Nov 29, 2022 at 11:47:47AM -0500, Benjamin Coddington wrote:
> On 29 Nov 2022, at 9:02, Christoph Hellwig wrote:
>
> > Hmm. Having to set a flag to not accidentally corrupt per-task
> > state seems a bit fragile. Wouldn't it make sense to find a way to opt
> > into the feature only for sockets created from the syscall layer?
>
> It's totally fragile, and that's why it's currently broken in production.
> The fragile ship sailed when networking decided to depend on users setting
> the socket's GFP_ flags correctly to avoid corruption.
>
> Meantime, this problem needs fixing in a way that makes everyone happy.
> This fix doesn't make it less fragile, but it may (hopefully) address the
> previous criticisms enough that something gets done to fix it.
Also, let's remember that while we're discussing how the kernel sould
work in an ideal world, the reality is that production NFS systems
crash randomly upon memory reclaim since commit a1231fda7e94 ("SUNRPC:
Set memalloc_nofs_save() on all rpciod/xprtiod jobs"). Fixing that is
just a matter of re-introducing GFP_NOFS on SUNRPC sockets (which has
been proposed several times already). Then we'll have plenty of time
to argue about how networking should use the per-task page_frag and
how to remove GFP_NOFS in the long term.
