Re: [LSF/MM TOPIC] mmap_sem in ->fault and ->page_mkwrite

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

On Thu, Jan 31, 2013 at 11:23:35PM +0100, Jan Kara wrote:
>   Hi,
> 
>   I'm not sure if this is such a great topic but it's a question which
> I came across a few times already and LSF/MM is a good place for
> brainstorming somewhat crazy ideas ;).
> 
> So currently ->fault() and ->page_mkwrite() are called under mmap_sem held
> for reading. Now this creates sometimes unpleasant locking dependencies for
> filesystems (modern filesystems have to do an equivalent of ->write_begin
> in ->page_mkwrite and that is a non-trivial operation). Just to mention my
> last itch, I had to split reader side of filesystem freezing lock into two
> locks - one which ranks above mmap_sem and one which ranks below it. Then
> writer side has to wait for both locks. It works but ...
> 
> So I was wondering: Would it be somehow possible we could drop mmap_sem in
> these two callbacks (especially ->page_mkwrite())? I understand process'
> mapping can change under us once we drop the semaphore so we'd have to
> somehow recheck we have still the right page after re-taking mmap_sem. Like
> if we protected VMAs with SRCU so that they don't disappear under us once
> we drop mmap_sem and after retaking mmap_sem we would recheck whether VMA
> still applies to our fault.
> 
> And I know there's VM_FAULT_RETRY but that really seems like a special hack
> for x86 architecture page fault code. Making it work for all architectures
> and callers such as get_user_pages() didn't really seem plausible to me.

Please, *please*, don't.  VMA locking is complete horror without SRCU
mess thrown in.  It's a bloody bad idea, at least without a very massive
cleanup prior to that thing.

Start with drawing the call graph for vma-related code - at least the
parts from relevant locks grabbed to accesses of fields protected by
said locks.