Re: [PATCH] Fix ext2 error reporting on fsync

[Andrew Morton <akpm@osdl.org>]

Jan Kara <jack@suse.cz> wrote:
>
> > Jan Kara <jack@suse.cz> wrote:
> > >
> > > > (If you're proposing that we walk the buffer_head list until we somehow
> > >  > find the list_head or hlist_head which is embedded within the desired
> > >  > address_space then yes2, but how do we know where to terminate the search?)
> > >    Yes, that's it. If the list is not circular, then it's easy to find
> > >  it's head. So what I suggest is to change private_list from a circular
> > >  list of list_head structures to the non-circular one. And because
> > >  non-circular list is already implemented in hlist macros I'd simply use
> > >  those. The only minor hack is that mapping would not contain hlist_head
> > >  but hlist_node with pprev set to NULL so that we easily recognize it when
> > >  traversing backwards. Is it clearer now?
> > 
> > Ah, yes.  It could get pretty inefficient in some corner cases, I guess.  A
> > 4GB file will have up to 1000 buffers on that list so we're looking at
> > potentially O(1000000) operations.
> > 
> > Plan B is to stick with a doubly-linked list, but mark the address_space's
> > list_head by setting bit 0 of its list_head.next to 1.
> > 
> > The challenge is to implement this in a manner which Linus doesn't notice ;)
>   In the mean time I though of an improvement of our cunning plan ;) We
> could reserve a bit in bh_state. Something like BH_Neighbor_EIO. If
> buffer is going to be thrown out from private_list (and has BH_Write_EIO
> or BH_Neighbor_EIO set) we find another victim (previous buffer on the
> list) and mark it by BH_Neighbor_EIO. If the buffer was the first one,
> we can set error on the mapping. By this we get rid of going through the
> long link list.
>   I guess I like this solution enough so that I start coding ;).

hm.

Another approach would be to allocate a new buffer_head which "belongs" to
the address_space.  So instead of doing:

		bh<->bh<->address_space<->bh<->bh

we do:

                     address_space
                           ^
                           |
                           v
		bh<->bh<->bh'<->bh<->bh

and set a magic bit in bh' which says "this is not a real bh; your
address_space is at *b_private".