On Wed 21-02-18 11:57:47, Christopher Lameter wrote: > On Wed, 21 Feb 2018, Shakeel Butt wrote: > > > On Wed, Feb 21, 2018 at 8:09 AM, Christopher Lameter <cl@xxxxxxxxx> wrote: > > > Another way to solve this is to switch the user context right? > > > > > > Isnt it possible to avoid these patches if do the allocation in another > > > task context instead? > > > > > > > Sorry, can you please explain what you mean by 'switch the user > > context'. Is there any example in kernel which does something similar? > > See include/linux/task_work.h. One use case is in mntput_no_expire() in > linux/fs/namespace.c > > > > Are there really any other use cases beyond fsnotify? > > > > > > > Another use case I have in mind and plan to upstream is to bind a > > filesystem mount with a memcg. So, all the file pages (or anon pages > > for shmem) and kmem (like inodes and dentry) will be charged to that > > memcg. > > The mount logic already uses task_work.h. That may be the approach to > expand there. I don't see how task work can be used here. Firstly I don't know of a case where task work would be used for something else than the current task - and that is substantial because otherwise you have to deal with lots of problems like races with task exit, when work gets executed (normally it gets executed once task exits to userspace) etc. Or do you mean that you'd queue task work for current task and then somehow magically switch memcg there? In that case this magic switching isn't clear to me... Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
