Re: [PATCH] bcachefs: fix incorrect usage of REQ_OP_FLUSH

[Dave Chinner <david@fromorbit.com>]

On Mon, Jan 22, 2024 at 07:41:47PM +0100, Christoph Hellwig wrote:
> On Mon, Jan 22, 2024 at 12:42:24PM -0500, Kent Overstreet wrote:
> > updating my tests for the MD_FAULTY removal, then will do. Is there
> > anything you want me to look for?
> 
> Nothing fancy.  Just that you do data integrity operations and do
> not see an error.
> 
> > considering most tests won't break or won't clearly break if flush/fua
> > is being dropped (even generic/388 was passing reliably, of course,
> > since virtual block devices aren't going to reorder writes...) maybe we
> > could do some print statement sanity checking...
> 
> Well, xfstests is not very good about power fail testing, because it
> can't inject a power fail..  Which is a problem in it's own, but
> would need hardware power failing or at least some pretty good VM
> emulation of the same outside the scope of the actual xfstests runner.

We do actually have stuff in fstests that checks write vs flush
ordering as issued by the filesystem. We use dm-logwrites for
tracking the writes and flushes and to be able to replay arbitrary
groups of writes between flushes.  generic/482 is one of those
tests.

These are the ordering problems that power failures expose,
so we do actually have tests that cover the same conditions that
pulling the power from a device would exercise.

I even wrote a debugging script (tools/dm-logwrite-replay) to
iterate write+fua groups in the test corpse to isolate the write
groups where the consistency failure occurs when doing work to
optimise flushes being issued by the XFS journal checkpoint writes.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com