[RESEND3,0/2] fuse: don't stuck clients on retrieve_notify with size > max_write

Message

Kirill Smelkov March 14, 2019, 10:45 a.m. UTC

Miklos,

On Thu, Feb 28, 2019 at 02:47:57PM +0300, Kirill Smelkov wrote:
> On Thu, Feb 28, 2019 at 09:10:15AM +0100, Miklos Szeredi wrote:
> > On Wed, Feb 27, 2019 at 9:39 PM Kirill Smelkov <[e-mail]> wrote:
> > 
> > > I more or less agree with this statement. However can we please make the
> > > breakage to be explicitly visible with an error instead of exhibiting it
> > > via harder to debug stucks/deadlocks? For example sys_read < max_write
> > > -> error instead of getting stuck. And if notify_retrieve requests
> > > buffer larger than max_write -> error or cut to max_write, but don't
> > > return OK when we know we will never send what was requested to
> > > filesystem even if it uses max_write sized reads. What is the point of
> > > breaking in hard to diagnose way when we can make the breakage showing
> > > itself explicitly? Would a patch for such behaviour accepted?
> > 
> > Sure, if it's only adds a couple of lines.   Adding more than say ten
> > lines for such a non-bug fix is definitely excessive.
>
> Ok, thanks. Please consider applying the following patch. (It's a bit
> pity to hear the problem is not considered to be a bug, but anyway).
>
> I will also send the second patch as another mail, since I could not
> made `git am --scissors` to apply several patched extracted from one
> mail successfully.

[...]

On Thu, Mar 07, 2019 at 12:34:21PM +0300, Kirill Smelkov wrote:
> Ping. Miklos, is there anything wrong with this patch and its
> second counterpart?

As we were talking here are those patches. The first one cuts notify_retrieve
request to max_write and is one line only. The second one returns error to
filesystem server if it is buggy and does sys_read with buffer size <
max_write. It is 2 lines of code and 7 lines of comments.

I still think that the patches fix real bugs. It is a bug if server behaviour
is a bit non-confirming or simply on an edge of being correct or questionable,
and instead of properly getting plain error from kernel, the whole system gets
stuck. It is a bug because bug amplification factor here is at least one order
of magnitude instead of staying ~1x.

I'm sending the patches for the third time already, but did not get any
feedback. Could you please have a look?

Thanks beforehand,
Kirill


Kirill Smelkov (2):
  fuse: retrieve: cap requested size to negotiated max_write
  fuse: require /dev/fuse reads to have enough buffer capacity as negotiated

 fs/fuse/dev.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)