On Sat, Jan 25, 2020 at 10:45:25AM -0800, Linus Torvalds wrote: > On Sat, Jan 25, 2020 at 5:06 AM Alexey Gladkov <gladkov.alexey@xxxxxxxxx> wrote: > > > > This allows to flush dcache entries of a task on multiple procfs mounts > > per pid namespace. > > From a quick read-through, this is the only one I really react negatively to. > > The locking looks odd. It only seems to protect the new proc_mounts > list, but then it's a whole big rwsem, and it's taken over all of > proc_flush_task_mnt(), and the locking is exported to all over as a > result of that - including the dummy functions for "there is no proc" > case. > > And proc_flush_task_mnt() itself should need no locking over any of > it, so it's all just for the silly looping over the list. Thank you, I will rework this part. > So > > (a) this looks fishy and feels wrong - I get a very strong feeling > that the locking is wrong to begin with, and could/should have been > done differently > > (b) all the locking should have been internal to /proc, and those > wrappers shouldn't exist in a common header file (and certainly not > for the non-proc case). > > Yes, (a) is just a feeling, and I don't have any great suggestions. > Maybe make it an RCU list and use a spinlock for updating it? I’m thinking, is it possible to get rid of proc_flush_task at all ? Maybe we can try to flush dcache during readdir for example. > But (b) is pretty much a non-starter in this form. Those wrappers > shouldn't be in a globally exported core header file. No way. > > Linus > -- Rgrds, legion
