Hi Xiao,
I was looking at the issue of parallel flush requests being slow as a
customer reported it against a kernel of ours that doesn't have your
patch.
I reviewed you patch more carefully and noticed a problem. It calls
submit_bio() multiple times from within a make_request_fn function.
Calling submit_bio once is perfectly safe. Calling it twice requires
care and usually involves one of the call submitting part of the
original, and the other resubmitting the original with that part
removed.
Calling it more often is rarely safe. The problem is that all these
bios get queued on current->biolist and don't even start being
processed until the top level make_request_fn completes. This can
result in deadlocks.
Also, with your patch there can be lots of flush requests all being
sent at much the same time, which all do essentially the same
thing. This is fairly wasteful.
I came up with a different approach. I restored the old because of
flush requests being serialized, but now if a caller will not bother
sending a flush if some other caller submitted a flush request while
it waited. This should result in many fewer flush requests, and must
less delay.
Would you be able to run you same test on this version to confirm
that it fixes the performance issue, and if so review the code to see
if you are happy with it.
Thanks a lot,
NeilBrown
---
NeilBrown (2):
Revert "MD: fix lock contention for flush bios"
md: batch flush requests.
drivers/md/md.c | 172 +++++++++++++++++++++++--------------------------------
drivers/md/md.h | 25 +++-----
2 files changed, 81 insertions(+), 116 deletions(-)
--
Signature
