On Wed, Sep 4, 2019 at 2:49 PM Aleksa Sarai [off-list ref] wrote:

Hinting to userspace to do a retry (with -EAGAIN as you mention in your
other mail) wouldn't be a bad thing at all, though you'd almost
certainly get quite a few spurious -EAGAINs -- &{mount,rename}_lock are
global for the entire machine, after all.

I'd hope that we have some future (possibly very long-term)
alternative that is not quite system-global, but yes, right now they
are.

Which is one reason I'd rather see EAGAIN in user space - yes, it
probably makes it even easier to trigger, but it also means that user
space might be able to do something about it when it does trigger.

For example, maybe user space can first just use an untrusted path
as-is, and if it gets EAGAIN or EXDEV, it may be that user space can
simplify the path (ie turn "xyz/.../abc" into just "abc".

And even if user space doesn't do anything like that, I suspect a
performance problem is going to be a whole lot easier to debug and
report when somebody ends up seeing excessive retries happening. As a
developer you'll see it in profiles or in system call traces, rather
than it resulting in very odd possible slowdowns for the kernel.

And yeah, it would probably be best to then at least delay doing
option 3 indefinitely, just to make sure user space knows about and
actually has a test-case for that EAGAIN happening.

              Linus