You are attempting a fairly hard task, the time to recover is probably
going to be a function of how much data was dirty at the time of crash,
you need a fairly deterministic test setup to get any sort of reliable
results. For instance, in XFS, 300 changes to the same inode one after
the other will be a lot cheaper to recover from than 300 changes to
different inodes, and with different inodes it would depend on where
those inodes were allocated, they could all end up in 10 inode clusters
(i.e. 10 disk reads and writes) or 300 different inode clusters. So lots
of test runs and averaged results (+ standard deviation etc).
That aside, there is really no simple way to simulate a failed unmount
without modifying the kernel code. In XFS you would have to modify the
linvfs_put_super() function in fs/xfs/linux/xfs_super.c so that it
just returned immediately rather than executing any code.
Steve
> Hi all...
>
> I am new to this list. Excuse me if the question
> is somehow off-topic. I know you're all busy, so
> I'll try not to disturb you much.
>
> I am trying to design several tests to run with
> different file systems for a linux gazette article.
> Since all of them are journal file systems,
> I would like to measure the time to recover for
> each file system once a crash occurs.
> My question is the following:
> is there anyway (from user space) to crash a
> partition, I mean unmount it uncleanly as it
> happens when you power off the computer without
> unmounting. I would like to unmount uncleanly just
> the partition I choose, in order not to crash the
> root partition too.
> If not, is it possible to do that writing a
> user app?...not hacking the modules code...
> because I need the buffers not to be flushed to
> disk.
>
> Suggestions and ideas for tests are welcome.
>
> Thanx in advance.
>
>
> --
> Ta luegoz
>
> _ __ _______ __ ___ _____ ___
> | / /( |\ \/ /| _)|_ _|| |
> |_\_\/_/|_| |__| |_| |_| |___|
>
> krypto mailto:krypto@xxxxxxxxxxxx
>
|