Re: Buffered I/O broken on s390x with page faults disabled (gfs2)

[Linus Torvalds <torvalds@linux-foundation.org>]

On Wed, Mar 9, 2022 at 10:42 AM Andreas Gruenbacher <agruenba@redhat.com> wrote:
>
> From what I took from the previous discussion, probing at a sub-page
> granularity won't be necessary for bytewise copying: when the address
> we're trying to access is poisoned, fault_in_*() will fail; when we get
> a short result, that will take us to the poisoned address in the next
> iteration.

Sadly, that isn't actually the case.

It's not the case for GUP (that page aligns things), and it's not the
case for fault_in_writeable() itself (that also page aligns things).

But more importantly, it's not actually the case for the *users*
either. Not all of the users are byte-stream oriented, and I think it
was btrfs that had a case of "copy a struct at the beginning of the
stream". And if that copy failed, it wouldn't advance by as many bytes
as it got - it would require that struct to be all fetched, and start
from the beginning.

So we do need to probe at least a minimum set of bytes. Probably a
fairly small minimum, but still...


> With a large enough buffer, a simple malloc() will return unmapped
> pages, and reading into such a buffer will result in fault-in.  So page
> faults during read() are actually pretty normal, and it's not the user's
> fault.

Agreed. But that wasn't the case here:

> In my test case, the buffer was pre-initialized with memset() to avoid
> those kinds of page faults, which meant that the page faults in
> gfs2_file_read_iter() only started to happen when we were out of memory.
> But that's not the common case.

Exactly. I do not think this is a case that we should - or need to -
optimize for.

And doing too much pre-faulting is actually counter-productive.

> * Get rid of max_size: it really makes no sense to second-guess what the
>   caller needs.

It's not about "what caller needs". It's literally about latency
issues. If you can force a busy loop in kernel space by having one
unmapped page and then do a 2GB read(), that's a *PROBLEM*.

Now, we can try this thing, because I think we end up having other
size limitations in the IO subsystem that means that the filesystem
won't actually do that, but the moment I hear somebody talk about
latencies, that max_size goes back.

                Linus