Messages in this thread | | | From | Linus Torvalds <> | Date | Sat, 2 Jun 2012 18:37:47 -0700 | Subject | Re: [PATCH] tty: add lockdep annotations |
| |
On Sat, Jun 2, 2012 at 3:25 PM, Linus Torvalds <torvalds@linux-foundation.org> wrote: > > Hmm. Ok. Looking at it, the ".shutdown" and ".remove" functions are > all very limited, so I suspect we could just make the rule be that the > install/lookup functions are serialized against each other by the > pty_mutex (true today), and then we just add a small spinlock for the > actual driver array insert/lookup.
Actually, I think we could probably make it really trivial by forcing the free'ing of the tty itself to be RCU-delayed.
Then shutdown/remove would remove the entry with no locking what-so-ever (which is really nice if you're in an interrupt - because now *other* users don't need to use those annoying irq-safe versions), and simply just clear the ttys[] array index.
The lookup side would need to just do a RCU read lock, read the ->ttys[index] thing using ACCESS_ONCE, and then just do the atomic_inc_not_zero() dance I already did to validate that the thing is still alive.
Voila - very cheap locking, and the part that could possibly happen from interrupts (shutdown/remove) needs no locking at all.
Making the tty freeing be rcu-delayed sounds pretty dang simple too.
What do you think?
Anyway, I'm closing the merge window now (doing the tagging, booting and checking that allmodconfig/allyesconfig/allnoconfigs all compile fine) so it's 3.6 material, but it doesn't sound bad.
Linus
| |