On Fri, Jan 18, 2019 at 02:46:53PM -0500, Josef Bacik wrote: > On Fri, Jan 18, 2019 at 07:44:03PM +0100, Andrea Righi wrote: > > On Fri, Jan 18, 2019 at 11:35:31AM -0500, Josef Bacik wrote: > > > On Fri, Jan 18, 2019 at 11:31:24AM +0100, Andrea Righi wrote: > > > > This is a redesign of my old cgroup-io-throttle controller: > > > > https://lwn.net/Articles/330531/ > > > > > > > > I'm resuming this old patch to point out a problem that I think is still > > > > not solved completely. > > > > > > > > = Problem = > > > > > > > > The io.max controller works really well at limiting synchronous I/O > > > > (READs), but a lot of I/O requests are initiated outside the context of > > > > the process that is ultimately responsible for its creation (e.g., > > > > WRITEs). > > > > > > > > Throttling at the block layer in some cases is too late and we may end > > > > up slowing down processes that are not responsible for the I/O that > > > > is being processed at that level. > > > > > > How so? The writeback threads are per-cgroup and have the cgroup stuff set > > > properly. So if you dirty a bunch of pages, they are associated with your > > > cgroup, and then writeback happens and it's done in the writeback thread > > > associated with your cgroup and then that is throttled. Then you are throttled > > > at balance_dirty_pages() because the writeout is taking longer. > > > > Right, writeback is per-cgroup and slowing down writeback affects only > > that specific cgroup, but, there are cases where other processes from > > other cgroups may require to wait on that writeback to complete before > > doing I/O (for example an fsync() to a file shared among different > > cgroups). In this case we may end up blocking cgroups that shouldn't be > > blocked, that looks like a priority-inversion problem. This is the > > problem that I'm trying to address. > > Well this case is a misconfiguration, you shouldn't be sharing files between > cgroups. But even if you are, fsync() is synchronous, we should be getting the > context from the process itself and thus should have its own rules applied. > There's nothing we can do for outstanding IO, but that shouldn't be that much. > That would need to be dealt with on a per-contoller basis. OK, fair point. We shouldn't be sharing files between cgroups. I'm still not sure if we can have similar issues with metadata I/O (that may introduce latencies like the sync() scenario), I have to investigate more and do more tests. > > > > > > > > > I introduced the blk_cgroup_congested() stuff for paths that it's not easy to > > > clearly tie IO to the thing generating the IO, such as readahead and such. If > > > you are running into this case that may be something worth using. Course it > > > only works for io.latency now but there's no reason you can't add support to it > > > for io.max or whatever. > > > > IIUC blk_cgroup_congested() is used in readahead I/O (and swap with > > memcg), something like this: if the cgroup is already congested don't > > generate extra I/O due to readahead. Am I right? > > Yeah, but that's just how it's currently used, it can be used any which way we > feel like. I think it'd be very interesting to have the possibility to either throttle I/O before writeback or during writeback. Right now we can only throttle writeback. Maybe we can try to introduce some kind of dirty page throttling controller using blk_cgroup_congested()... Opinions? > > > > > > > > > > > > > > = Proposed solution = > > > > > > > > The main idea of this controller is to split I/O measurement and I/O > > > > throttling: I/O is measured at the block layer for READS, at page cache > > > > (dirty pages) for WRITEs, and processes are limited while they're > > > > generating I/O at the VFS level, based on the measured I/O. > > > > > > > > > > This is what blk_cgroup_congested() is meant to accomplish, I would suggest > > > looking into that route and simply changing the existing io controller you are > > > using to take advantage of that so it will actually throttle things. Then just > > > sprinkle it around the areas where we indirectly generate IO. Thanks, > > > > Absolutely, I can probably use blk_cgroup_congested() as a method to > > determine when a cgroup should be throttled (instead of doing my own > > I/O measuring), but to prevent the "slow writeback slowing down other > > cgroups" issue I still need to apply throttling when pages are dirtied > > in page cache. > > Again this is just a fuckup from a configuration stand point. The argument > could be made that sync() is probably broken here, but I think the right > solution here is to just pass the cgroup context along with the writeback > information and use that if it's set instead. Thanks, Alright, let's skip the root cgroup for now. I think the point here is if we want to provide sync() isolation among cgroups or not. According to the manpage: sync() causes all pending modifications to filesystem metadata and cached file data to be written to the underlying filesystems. And: According to the standard specification (e.g., POSIX.1-2001), sync() schedules the writes, but may return before the actual writing is done. However Linux waits for I/O completions, and thus sync() or syncfs() provide the same guarantees as fsync called on every file in the sys‐ tem or filesystem respectively. Excluding the root cgroup, do you think a sync() issued inside a specific cgroup should wait for I/O completions only for the writes that have been generated by that cgroup? Thanks, -Andrea