Re: [patch 9/9] mm: fix pagecache write deadlocks

[Nick Piggin <npiggin@suse.de>]

On Sun, Feb 04, 2007 at 05:40:35PM +0000, Anton Altaparmakov wrote:
> On Sun, 4 Feb 2007, Andrew Morton wrote:
> > truncate's OK: we're holding i_mutex.
> 
> How about excluding readpage() (in addition to truncate if Nick is right  
> and some cases of truncate do not hold i_mutex) with an extra page flag as
> I proposed for truncate exclusion?  Then it would not matter that
> prepare_write might have allocated blocks and might expose stale data.    
> It would go to sleep and wait on the bit to be cleared instead of trying  
> to bring the page uptodate.  It can then lock the page and either find it 
> uptodate (because commit_write did it) or not and then bring it uptodate.
> 
> Then we could safely fault in the page, copy from it into a temporary 
> page, then lock the destination page again and copy into it.
> 
> This is getting more involved as a patch again...  )-:  But at least it   
> does not affect the common case except for having to check the new page 
> flag in every readpage() and truncate() call.  But at least the checks 
> could be with an "if (unlikely(newpageflag()))" so should not be too bad.
> 
> Have I missed anything this time?

Yes. If you have a flag to exclude readpage(), then you must also
exclude filemap_nopage, in which case it is still deadlocky.