Message-Id: <20061116011351.1401a00f.akpm@osdl.org>
Date:	Thu, 16 Nov 2006 01:13:51 -0800
From:	Andrew Morton <akpm@...l.org>
To:	cmm@...ibm.com
Cc:	Hugh Dickins <hugh@...itas.com>, Mel Gorman <mel@...net.ie>,
	"Martin J. Bligh" <mbligh@...igh.org>,
	linux-kernel@...r.kernel.org,
	"linux-ext4@...r.kernel.org" <linux-ext4@...r.kernel.org>
Subject: Re: Boot failure with ext2 and initrds

On Thu, 16 Nov 2006 00:49:20 -0800
Mingming Cao <cmm@...ibm.com> wrote:

> On Wed, 2006-11-15 at 23:22 -0800, Andrew Morton wrote:
> > On Wed, 15 Nov 2006 22:55:43 -0800
> > Mingming Cao <cmm@...ibm.com> wrote:
> > 
> > > Hmm, maxblocks, in bitmap_search_next_usable_block(),  is the end block 
> > > number of the range  to search, not the lengh of the range. maxblocks 
> > > get passed to ext2_find_next_zero_bit(), where it expecting to take the 
> > > _size_ of the range to search instead...
> > > 
> > > Something like this: (this is not a patch)
> > >   @@ -524,7 +524,7 @@ bitmap_search_next_usable_block(ext2_grp
> > >    	ext2_grpblk_t next;
> > > 
> > >    -  	next = ext2_find_next_zero_bit(bh->b_data, maxblocks, start);
> > >    +  	next = ext2_find_next_zero_bit(bh->b_data, maxblocks-start + 1, start);
> > > 	if (next >= maxblocks)
> > >    		return -1;
> > >    	return next;
> > >    }
> > 
> > yes, the `size' arg to find_next_zero_bit() represents the number of bits
> > to scan at `offset'.
> > 
> > So I think your change is correctish.  But we don't want the "+ 1", do we?
> > 
> I think we still need the "+1", maxblocks here is the ending block of
> the reservation window, so the number of bits to scan =end-start+1.
> 
> > If we're right then this bug could cause the code to scan off the end of the
> > bitmap.  But it won't explain Hugh's bug, because of the if (next >= maxblocks).
> > 
> 
> Yeah.. at first I thought it might be related, then, thinked it over,
> the bug only makes the bits to scan larger, so if find_next_zero_bit()
> returns something off the end of bitmap, that is fine, it just
> indicating that there is no free bit left in the rest of bitmap, which
> is expected behavior. So bitmap_search_next_usable_block() fail is the
> expected. It will move on to next block group and try to create a new
> reservation window there.

I wonder why it's never oopsed.  Perhaps there's always a zero in there for
some reason.

> That does not explain the repeated reservation window add and remove
> behavior Huge has reported. 

I spent quite some time comparing with ext3.  I'm a bit stumped and I'm
suspecting that the simplistic porting the code is now OK, but something's
just wrong.

I assume that the while (1) loop in ext3_try_to_allocate_with_rsv() has
gone infinite.  I don't see why, but more staring is needed.

What lock protects the fields in struct ext[234]_reserve_window from being
concurrently modified by two CPUs?  None, it seems.  Ditto
ext[234]_reserve_window_node.  i_mutex will cover it for write(), but not
for pageout over a file hole.  If we end up with a zero- or negative-sized
window then odd things might happen.

> > btw, how come try_to_extend_reservation() uses spin_trylock?
> 
> Since locks are all allocated from reservation window, when ext3
> multiple blocks allocation was added, we added try_to_extend_reservation
> () to ext3, which trying to extend the reservation window size to at
> least match the number of blocks to allocate. So we have better chance
> to allocating multiple blocks from the window at a time.
> 
> Since all the multiple block allocation is based on best effort basis,
> the same applied to try_to_extend_reservation(). It seems no need to
> wait for the reservation tree lock if it's not avaible at that moment.
> 

I suspect that was not a good idea - it has added rather different
behaviour in the once-in-a-million case.


-
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists