Re: [PATCH] AFS: Implement shared-writable mmap [try #2]

[Nick Piggin <nickpiggin@yahoo.com.au>]

David Howells wrote:
> Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> 
> 
>>You can drop the lock, do the invalidation,
> 
> 
> Hmmm...  There's a danger of incurring a race by doing that.  Consider two
> processes both trying to write to a dirty page for which writeback will be
> rejected:
> 
>  (1) The first process gets EKEYREJECTED from the server, drops its lock and
>      is then preempted.
> 
>  (2) The second process gets EKEYREJECTED from the server, drops its lock,
>      truncates the page, reloads the page and modifies it.
> 
>  (3) The first process resumes and truncates the page, thereby splatting the
>      second process's write.
> 
> Or:
> 
>  (1) The first process gets EKEYREJECTED from the server, clears the writeback
>      information from the page, drops its lock and is then preempted.
> 
>  (2) The second process attaches its own writeback information to the page and
>      modifies it.
> 
>  (3) The first process resumes and truncates the page, thereby splatting the
>      second process's write.

If there are race issues involving concurrent invalidations, then you'd
fix that up by taking a filesystem specific lock to prevent them.

Generic write path should be holding i_mutex, but I don't think you can
take that from page_mkwrite... Just add one of your own.


> Really, what I want to do is pass the page lock to truncate to deal with.
> Better still, I want truncate to be selective, based on whether or not a page
> is still associated with the rejected writeback.  I wonder if I should call
> truncate_complete_page() or invalidate_complete_page() directly.

No, you shouldn't. We could theoretically introduce a new API for this,
but I think it would be preferable if you can fix the race in the fs.

-- 
SUSE Labs, Novell Inc.