On 06/22, Christian Brauner wrote: > > On Mon, Jun 22, 2020 at 12:24:01PM +0200, Dominique Martinet wrote: > > Christian Brauner wrote on Mon, Jun 22, 2020: > > > On Mon, Jun 22, 2020 at 08:25:28AM +0200, Oleg Nesterov wrote: > > >> current->sighand is stable and can't go away. Unless "current" is > > >> exiting and > > >> has already passed exit_notify(). So I don't think net/9p needs this > > >> helper. > > > > > > From what I can gather from the thread (cf. [1]) that is linked in the > > > commit message the main motivation for all of this is sparse not being > > > happy and not some bug. (Maybe I'm not seeing something though.) > > > > > > The patch itself linked here doesn't seem to buy anything. I agree with > > > Oleg. Afaict, lock_task_sighand() would only be needed here if the task > > > wouldn't be current. So maybe it should just be dropped from the series. > > > > Sure. I honestly have no idea on what guarantees we have from the task > > being current here as opposed to any other task -- I guess that another > > thread calling exit for exemple would have to wait? > > When a thread in a non-trivial thread-group (sorry for the math > reference :)) execs it'll unshare its struct sighand.
Well, not really... The execing threads will kill other other threads, then it will check if ->sighand should be unshared. The latter is very unlikely, I don't think CLONE_SIGHAND without CLONE_THREAD is actually used today. But this doesn't really matter. I mean, even if you race with another thread doing exec/exit/whatever, current->sighand is stable. Unless, again, current has already exited (called exit_notify()). > The new struct > sighand will be assigned using rcu_assign_pointer() so afaik (Paul or > Oleg can yell at me if I'm talking nonsense) any prior callers will see > the prior sighand value. Yes, but see above. If tsk is not current, then (in general) it is not safe to use tsk->sighand directly. It can can be changed by exec (as you explained), or you can hit tsk->sighand == NULL if you race with exit. Oleg.