[ Sending again with a trimmed CC list to just the lists. Jeff - cc lists that large get blocked by mailing lists... ] On Tue, Apr 07, 2015 at 02:55:13PM -0400, Jeff Moyer wrote: > The way the on-stack plugging currently works, each nesting level > flushes its own list of I/Os. This can be less than optimal (read > awful) for certain workloads. For example, consider an application > that issues asynchronous O_DIRECT I/Os. It can send down a bunch of > I/Os together in a single io_submit call, only to have each of them > dispatched individually down in the bowels of the dirct I/O code. > The reason is that there are blk_plug-s instantiated both at the upper > call site in do_io_submit and down in do_direct_IO. The latter will > submit as little as 1 I/O at a time (if you have a small enough I/O > size) instead of performing the batching that the plugging > infrastructure is supposed to provide. I'm wondering what impact this will have on filesystem metadata IO that needs to be issued immediately. e.g. we are doing writeback, so there is a high level plug in place and we need to page in btree blocks to do extent allocation. We do readahead at this point, but it looks like this change will prevent the readahead from being issued by the unplug in xfs_buf_iosubmit(). So while I can see how this can make your single microbenchmark better (because it's only doing concurrent direct IO to the block device and hence there are no dependencies between individual IOs), I have significant reservations that it's actually a win for filesystem-based workloads where we need direct control of flushing to minimise IO latency due to IO dependencies... Patches like this one: https://lkml.org/lkml/2015/3/20/442 show similar real-world workload improvements to your patchset by being smarter about using high level plugging to enable cross-file merging of IO, but it still relies on the lower layers of plugging to resolve latency bubbles caused by IO dependencies in the filesystems. > NOTE TO SUBSYSTEM MAINTAINERS: Before this patch, blk_finish_plug > would always flush the plug list. After this patch, this is only the > case for the outer-most plug. If you require the plug list to be > flushed, you should be calling blk_flush_plug(current). Btrfs and dm > maintainers should take a close look at this patch and ensure they get > the right behavior in the end. IOWs, you are saying we need to change all our current unplugs to blk_flush_plug(current) to *try* to maintain the same behaviour as we currently have? I say *try*, because no instead of just flushing the readahead IO on the plug, we'll also flush all the queued data writeback IO onthe high level plug. We don't actually want to do that; we only want to submit the readahead and not the bulk IO that will delay the latency sensitive dependent IOs.... If that is the case, shouldn't you actually be trying to fix the specific plugging problem you've identified (i.e. do_direct_IO() is flushing far too frequently) rather than making a sweeping generalisation that the IO stack plugging infrastructure needs fundamental change? Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
