actually my fix seems cleaner because it puts the "clearing" in one single
> > So for now I did an hack to cure the other two (writepage can still
> > corrupt the fs as said). I think the right fix (ala 2.5) is to change
> > the API so we can use the last blocks too, but the below will cure 2.4
> > and for writepage the right fix IMHO is to do the reservation of the
> > space.
> This is better in a way because it reclaims the eztra few blocks. But
> the -ac approach also works for writepage().
yes, it can solve the metadata corruption (assuming the locking is
right, I can as well call ->truncate within writepage but it's not
obvious at all that it won't race because we don't hold the i_sem within
writepage), but the data corruption still holds. I mean, there's no
failure path to notify userspace that a certain page fault is writing
into a page over an hole, that we don't have space to later allocate on
disk. so to me it sounds like MAP_SHARED should preallocate the space of
the holes so you will know that the writes into the MAP_SHARED segments
won't be lost (current state of things will lead to silent corruption
and pinned dirty pages in ram, aka broken allocate on flush like
> Why was that code not brought across?
Who developed that code? Can the author of the code forward port it to
2.4.18pre and post a patch to the list so we can review? thanks,
Avoiding the matadata corruption would be a good start at least for 2.4,
then we should just focus on the writepage locking that could race with the
other "create=1" get_blocks. If it doesn't race I will certainly agree
on that approch for 2.4.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to firstname.lastname@example.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/