Re: [PATCH 1 of 2] block_page_mkwrite() Implementation V2

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

David Howells wrote:
> Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> 
> 
>>Dave is using prepare_write here to ensure blocks are allocated in the
>>given range. The filesystem's ->nopage function must ensure it is uptodate
>>before allowing it to be mapped.
> 
> 
> Which is fine... assuming it's called.  For blockdev-based filesystems, this
> is probably true.  But I'm not sure you can guarantee it.
> 
> I've seen Ext3, for example, unlocking a page that isn't yet uptodate.
> nopage() won't get called on it again, but prepare_write() might.  I don't
> know why this happens, but it's something I've fallen over in doing
> CacheFiles.  When reading, readpage() is just called on it again and again
> until it is up to date.  When writing, prepare_write() is called correctly.

There are bugs in the core VM and block filesystem code where !uptodate pages
are left in pagetables. Some of these are fixed in -mm.

But they aren't a good reason to invent completely different ways to do things.


>>Consider that the code currently works OK today _without_ page_mkwrite.
>>page_mkwrite is being added to do block allocation / reservation.
> 
> 
> Which doesn't prove anything.  All it means is that PG_uptodate being unset is
> handled elsewhere.

It means that Dave's page_mkwrite function will do the block allocation
and everything else continues as it is. Your suggested change to pass in
offset == to is just completely wrong for this.

PG_uptodate being unset should be done via pagecache invalidation or truncation
APIs, which (sometimes... modulo bugs) tear down pagetables first.

-- 
SUSE Labs, Novell Inc.