Re: [patch] fs: improved handling of page and buffer IO errors

[steve@chygwyn.com]

Hi,

On Tue, Oct 21, 2008 at 04:32:21PM +0200, Miklos Szeredi wrote:
> On Tue, 21 Oct 2008, steve@chygwyn.com
> > Once thats done, the dlm/glock is dropped (again notification is via
> > the dlm) and if Node A has outstanding requests queued up, it
> > re-requests the glock. This is a slightly simplified explanation
> > but, I hope it gives the general drift.
> 
> Yes, thanks.
> 
> > So to return to the original subject, in order to allow all
> > this locking to occur with no lock ordering problems, we have
> > to define a suitable ordering of page locks vs. glocks, and the
> > ordering that we use is that glocks must come before page locks. The
> > full ordering of locks in GFS2 is in Documentation/filesystems/gfs2-glocks.txt
> > 
> > As a result of that, the VFS needs reads (and page_mkwrite) to
> > retry when !PageUptodate() in case the returned page has been
> > invalidated at any time when the page lock has been dropped.
> 
> Since this commit PG_uptodate isn't cleared on invalidate:
> 
>   commit 84209e02de48d72289650cc5a7ae8dd18223620f
>   Author: Miklos Szeredi <mszeredi@suse.cz>
>   Date:   Fri Aug 1 20:28:47 2008 +0200
>   
>       mm: dont clear PG_uptodate on truncate/invalidate
> 
> Testing for !page->mapping, however, is a reliable way to detect both
> truncation and invalidation.
> 
> So the page can have the following states:
> 
>   !PG_uptodate                    -> page has not been read
>   PG_uptodate && page->mapping    -> page has been read and is valid
>   PG_uptodate && !page->mapping   -> page has been read but no longer valid
> 
> So PG_uptodate does not reflect the validity of the data, only whether
> the data was ever made up-to-date.
> 
> Does this make sense?  Should it be documented somewhere?
>
Well I'm not sure why we'd need to distinguish between "page has not been read"
and "page has been read but no longer valid". I guess I don't understand why
those two cases are not the same from the vfs and filesystem points of
view.

I'm sure it should be documented :-) it certainly seems confusing and if we
want to keep this scheme, can we change PG_uptodate to PG_wasread or
PG_usedonce or something like that which more clearly reflects its
purpose in that case,

Steve.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>