On Fri, May 24, 2019 at 10:06 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote: > > On Fri, May 24, 2019 at 09:11:46AM -0700, Matthew Wilcox wrote: > > On Thu, May 23, 2019 at 03:59:33PM -0400, Johannes Weiner wrote: > > > My point is that we cannot have random drivers' internal data > > > structures charge to and pin cgroups indefinitely just because they > > > happen to do the modprobing or otherwise interact with the driver. > > > > > > It makes no sense in terms of performance or cgroup semantics. > > > > But according to Roman, you already have that problem with the page > > cache. > > https://lore.kernel.org/linux-mm/20190522222254.GA5700@castle/T/ > > > > So this argument doesn't make sense to me. > > You haven't addressed the rest of the argument though: why every user > of the xarray, and data structures based on it, should incur the > performance cost of charging memory to a cgroup, even when we have no > interest in tracking those allocations on behalf of a cgroup. > > Which brings me to repeating the semantics argument: it doesn't make > sense to charge e.g. driver memory, which is arguably a shared system > resource, to whoever cgroup happens to do the modprobe / ioctl etc. > > Anyway, this seems like a fairly serious regression, and it would make > sense to find a self-contained, backportable fix instead of something > that has subtle implications for every user of the xarray / ida code. Adding to Johannes point, one concrete example of xarray we don't want to charge is swapper_spaces. Swap is a system level resource. It does not make any sense to charge the swap overhead to a job and also it will have negative consequences like pinning zombies. Shakeel
