Re: Induced crashes

[tb@becket.net (Thomas Bushnell, BSG)]

neal@cs.uml.edu (Neal H. Walfield) writes:

> > ext2fs should be quite robust: Even pulling the plug at any time should not
> > corrupt the filesystem beyond what e2fsck can repair.
> 
> Let us assume that ext2fs writes a block of metadata to disk.  In the
> kernel, in the middle of the DMA operation, the kernel panics.  There
> is no guarantee that e2fsck will be able to handle this case.

But it *can* make sure that e2fsck can repair the result.

> In other words, although a write operation may appear atomic to
> ext2fs, in the kernel, that operation is composed of many smaller
> atomic operations (with respect to the hardware) during any of which,
> the kernel may crash.

Disk hardware guarantees that a sector write can always be completed
even if the power goes out partway through.  That means that writing a
single sector *is* always atomic.

Thomas