On Thu, Mar 20, 2025 at 03:41:58PM +0800, Ming Lei wrote: > On Thu, Mar 20, 2025 at 12:08:19AM -0700, Christoph Hellwig wrote: > > On Tue, Mar 18, 2025 at 05:34:28PM +0800, Ming Lei wrote: > > > On Tue, Mar 18, 2025 at 12:57:17AM -0700, Christoph Hellwig wrote: > > > > On Tue, Mar 18, 2025 at 03:27:48PM +1100, Dave Chinner wrote: > > > > > Yes, NOWAIT may then add an incremental performance improvement on > > > > > top for optimal layout cases, but I'm still not yet convinced that > > > > > it is a generally applicable loop device optimisation that everyone > > > > > wants to always enable due to the potential for 100% NOWAIT > > > > > submission failure on any given loop device..... > > > > > > NOWAIT failure can be avoided actually: > > > > > > https://lore.kernel.org/linux-block/20250314021148.3081954-6-ming.lei@xxxxxxxxxx/ > > > > That's a very complex set of heuristics which doesn't match up > > with other uses of it. > > I'd suggest you to point them out in the patch review. Until you pointed them out here, I didn't know these patches existed. Please cc linux-fsdevel on any loop device changes you are proposing, Ming. It is as much a filesystem driver as it is a block device, and it changes need review from both sides of the fence. > > > > Yes, I think this is a really good first step: > > > > > > > > 1) switch loop to use a per-command work_item unconditionally, which also > > > > has the nice effect that it cleans up the horrible mess of the > > > > per-blkcg workers. (note that this is what the nvmet file backend has > > > > > > It could be worse to take per-command work, because IO handling crosses > > > all system wq worker contexts. > > > > So do other workloads with pretty good success. > > > > > > > > > always done with good result) > > > > > > per-command work does burn lots of CPU unnecessarily, it isn't good for > > > use case of container > > > > That does not match my observations in say nvmet. But if you have > > numbers please share them. > > Please see the result I posted: > > https://lore.kernel.org/linux-block/Z9FFTiuMC8WD6qMH@fedora/ You are arguing in circles about how we need to optimise for static file layouts. Please listen to the filesystem people when they tell you that static file layouts are a -secondary- optimisation target for loop devices. The primary optimisation target is the modification that makes all types of IO perform better in production, not just the one use case that overwrite-specific IO benchmarks exercise. If you want me to test your changes, I have a very loop device heavy workload here - it currently creates about 300 *sparse* loop devices totalling about 1.2TB of capacity, then does all sorts of IO to them through both the loop devices themselves and filesystems created on top of the loop devices. It typically generates 4-5GB/s of IO through the loop devices to the backing filesystem and it's physical storage. Speeding up or slowing down IO submission through the loop devices has direct impact on the speed of the workload. Using buffered IO through the loop device right now is about 25% faster than using aio+dio for the loop because there is some amount of re-read and re-write in the filesystem IO patterns. That said, AIO+DIO should be much faster than it is, hence my interest in making all the AIO+DIO IO submission independent of potential blocking operations. Hence if you have patch sets that improve loop device performance, then you need to make sure filesystem people like myself see those patch series so they can be tested and reviewed in a timely manner. That means you need to cc loop device patches to linux-fsdevel.... -Dave. -- Dave Chinner david@xxxxxxxxxxxxx
