> On Mon, Apr 10, 2000 at 10:52:10PM -0500, Jim Mostek wrote:
> > I have a few fixes (not fully tested) that eliminate a bunch
> > of file system corruption when multiple processes are hitting
> > the same file. ext2 dies pretty quickly running doio with multiple
> > threads. XFS is getting better, but there is still corruption.
>
> Could you expand a bit on the ext2 problems. Does it crash ?
>
> -Andi
doio is an internal SGI (originally Cray) test for verifying read/write
correctness. In this mode it has a 50 Mbyte file which four processes
are walking through doing non-overlapping, but sequential write operations
(via the write system call and via mmapped write).
What appears to happen usually is two threads are doing writes which are
next to each other in the file, the page which they share gets corrupted,
either the start of one write gets zeroed out, or the end of the other
write gets old data over the top of it.
The bug may perhaps be related to the handling of BH_New.
I think we should maybe look into getting doio out into the community.
Steve
|