On Sat, May 17, 2008 at 08:48:33PM -0400, Chris Mason wrote:The workload we used was to run our existing Centera application on a rack of boxes. The application is a bit special in that it does a digital signature on each file and never returns success for the client until an fsync is done on the server (kind of like synchronous NFS).
Well, the barriers happen like so (even if we actually only do one
barrier in submit_bh, it turns into two)
write log blocks
flush #1
write commit block
flush #2
write metadata blocks
I'd agree with Ted, there's a fairly small chance of things get reordered around flush #1. flush #2 is likely to have lots of reordering though. It should be easy to create situations where the metadata for a transaction is written before the log blocks ever see the disk.
True, but even with a very heavy fsync() workload, a commit doesn't
cause the metadata blocks to be written until we have to do a journal
truncate operation. A heavy fsync() workload would increase how
quickly we would use up the journal and need to do a journal truncate,
though.
EMC did a ton of automated testing around this when Jens and I did
the initial barrier implementations, and they were able to trigger
corruptions in fsync heavy workloads with randomized power offs.
I'll dig up the workload they used.
I could imagine a mode which forces a barrier operation for commits
triggered by fsync()'s, but not commits that occur due to a natural
closing of transactions. I'm not sure it's worth it, though, since
many of the benchmarks that we care about (like Postmark) do use
fsync() fairly heavily.
The really annoying thing is that what is really needed is a way to
make write barriers cheaper; we don't need to do a synchronous flush,
but unfortunately for most drives there isn't any other way of keeping
disk writes from getting reordered.