Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Andrew Morton <akpm@osdl.org>]

Andi Kleen <ak@suse.de> wrote:
>
> > Still.  It seems that what's happened is that we took a pagefault while
>  > reiserfs had a transaction open.  The fault is against a mmapped ext3 file
>  > and we ended up in the recently-reworked ext3_get_block() which tests
>  > journal_current_handle() to work out whether we're in a write or a read. 
>  > oops.  The presence of reiserfs journal_info makes it decide it's a write,
>  > not a read so it starts treating a reiserfs journal_info as an ext3 one.
>  > 
>  > The code used to work OK because it was only for direct-IO, which doesn't
>  > get recurred into like this.  But it got used for regular I/O in -mm.
> 
>  Oops. Can this happen in more situations?

I don't _think_ so, but it's pretty scary.  The code's been this way for a
while.

Typical scenario:

	reiserfs_journal_start		- sets current->journal_info
	  copy_from_user
	    pagefault
	      ext3_readpages()

As long as ext3_readpage[s]() doesn't try to start a transaction we're OK. 
And it shouldn't, if create==0.

Fortunately filemap_nopage() doesn't do atime updates.  If it did, things
would get messy.

We do have deadlock possibilities in there - filemap_nopage() does
lock_page() inside journal_start(), whereas generic_file_write() does
journal_start() inside lock_page().  Chris Mason and I have stared
unhappily at that a few times.  Hard to fix.

But I don't _think_ we have any more journal_start() recursions like this -
ext3 tends to get pretty noisy if it detects that in unexpected places.