Re: [GIT PULL] Orangefs (text only resend)

[Al Viro <viro@ZenIV.linux.org.uk>]

On Fri, Sep 11, 2015 at 05:12:08PM -0400, Mike Marshall wrote:
> I'm about to leave for the day...
> 
> I haven't found any problems with the GFP_KERNEL allocations that
> Al warned me about... that doesn't mean there aren't any...

*IF* you have nothing that would require locks in any of the pathways
related to memory pressure (and can guarantee that no such thing will
appear), GFP_KERNEL should be OK.  Still, doing that under the system-wide
mutex taken whenever you need to send a request looks like a Bad Idea(tm) -
too easy to introduce such deadlocks on subsequent changes.

> I'm using copy_page_to_iter in my new branch as Al suggested. I've
> changed the code quite a bit from any samples I've posted, there were
> regressions with it. I finally stole some code from
> cifs/file.c/cifs_readdata_to_iov
> and everything works... but I'm not happy with it yet... Linus once posted
> a message to the effect that "you don't fix bugs by thrashing around until
> stuff seems to work, you fix them by doing the right thing on purpose..."
> and I'm working towards that end...

Could you tell where does the current code live?  What's in -next appears
to be unchanged...

BTW, as for passing all your tests...  Do those include fuzzing it by
misbehaving server?  And getdents() from a directory that has a bunch of
long names *and* a short one in the very end looks like it would misbehave
even on correctly working server...