On Thu, Sep 15, 2011 at 05:25:57PM -0700, Darrick J. Wong wrote:
> On Thu, Sep 15, 2011 at 10:47:48AM -0500, Shawn Bohrer wrote:
> > Thanks Christoph,
> > On Thu, Sep 15, 2011 at 10:55:57AM -0400, Christoph Hellwig wrote:
> > > On Thu, Sep 15, 2011 at 09:47:55AM -0500, Shawn Bohrer wrote:
> > > > I've got a workload that is latency sensitive that writes data to a
> > > > memory mapped file on XFS. With the 3.0 kernel I'm seeing stalls of
> > > > up to 100ms that occur during writeback that we did not see with older
> > > > kernels. I've traced the stalls and it looks like they are blocking
> > > > on wait_on_page_writeback() introduced in
> > > > d76ee18a8551e33ad7dbd55cac38bc7b094f3abb "fs: block_page_mkwrite
> > > > should wait for writeback to finish"
> > > >
> > > > Reading the commit description doesn't really explain to me why this
> > > > change was needed.
> > >
> > > It it there to avoid pages beeing modified while they are under
> > > writeback, which defeats various checksumming like DIF/DIX, the iscsi
> > > CRCs, or even just the RAID parity calculations. All of these either
> > > failed before, or had to work around it by copying all data was
> > > written.
> > I'm assuming you mean software RAID here? We do have a hardware RAID
> > controller. Also for anything that was working around this issue
> > before by copying the data, are those workarounds still in place?
> I suspect iscsi and md-raid5 are still making shadow copies of data blocks
> before writing them out. However, there was no previous workaround for
> errors -- this ("*_page_mkwrite should wait...") patch series _is_ the fix for
> DIF/DIX. I recall that we rejected the shadow buffer approach for DIF/DIX
> because allocating new pages is expensive if we do it for each disk write in
> anticipation of future page writes...
So for the most part it sounds like this change is needed for DIF/DIX.
Could we only enable the wait_on_page_writeback() if
CONFIG_BLK_DEV_INTEGRITY is set? Does it make sense to tie these
> > > If you don't use any of these you can remove the call and things
> > > will work like they did before.
> > I may do this for now.
> > In the longer term is there any chance this could be made better? I'm
> > not an expert here so my suggestions may be naive. Could a mechanism
> > be made to check if the page needs to be checksummed and only block in
> ...however, one could replace that wait_on_page_writeback with some sort of
> call that would duplicate the page, update each program's page table to point
> to the new page, and then somehow reap the page that's under IO when the IO
> completes. That might also be complicated to implement, I don't know. If
> there aren't any free pages, then this scheme (and the one I mentioned in the
> previous paragraph) will block a thread while the system tries to reclaim some
> I think we also talked about a block device flag to signal that the device
> requires stable page writes, which would let us turn off the waits on devices
> that don't care. That at least could defer this discussion until you
> one of these devices that wants stable page writes.
I would be in favor of something like this.
> I'm curious, is this program writing to the mmap region while another program
> is trying to fsync/fdatasync/sync dirty pages to disk? Is that how you
> the jittery latency? We'd figured that not many programs would notice the
> latency unless there was something that was causing a lot of dirty page writes
> concurrent to something else dirtying a lot of pages. Clearly we failed in
> your case. Sorry. :/
The other thread in this case is the [flush-8:0] daemon writing back
the pages. So in our case you could see the spikes every time it wakes
up to write back dirty pages. While we can control this to some
extent with vm.dirty_writeback_centisecs and vm.dirty_expire_centisecs
it essentially impossible to ensure the writeback doesn't coincide
with us writing to the page again.
> That said, imagine if we revert to the pre-3.0 mechanism (or add that flag):
> we start transferring page A to the disk for writing and your program comes in
> and changes A to A' before that transfer completes, then the disk will see a
> data blob that is partly A and partly A', and the proportions of A/A' are
> ill-defined. I agree that ~100ms latency is not good, however. :(
In our use case I don't _think_ we care too much about the part A part
A' problem. For the most part if we cared about not getting a mix we
would fsync/msync the changes.
> What are your program's mmap write latency requirements?
In these use cases I'm in the "as fast as possible" business. We
don't have hard latency requirements, but we generally don't want to
see things get worse.
This email, along with any attachments, is confidential. If you
believe you received this message in error, please contact the
sender immediately and delete all copies of the message.