Re: Questions answered by Neil Brown

On Wednesday February 26, ptb@it.uc3m.es wrote:

What is also puzzling me is that despite the horrible potential for
what might happen from doing the original users end_io early, I
can't see any consequences in actual tests!

I am writing stuff to a file on a raid1 mirror mount, then dismantling
the mount, the device, and rmmoding the driver. Then I put it all back
again. On remount the data in the file is all perfect. Yet it was built
with plenty of async writes! Surely the buffers in some of those writes
should have been pointing nowhere and been full of rubbish?

I suspect that mostly you are writing from a cache, and the data will
probably stay around.
To be able to demonstrate a problem you probably need very high memory
pressure so things don't stay in cache long, lots of metadata updates,
and probably some for of journalling filesystem like I mentioned
previously.  Have very long latencies for the delayed write would also
make the problem more likely.

Even if you cannot demonstrate a problem, I'm sure one would be
noticed sooner or later if you released this sort of code into the
wild and people used it.

NeilBrown