tcp: Make all TCP timers start and stop atomically under the inp_lock (work in progress)

[• hselasky]

Summary

Pass the RW-lock to be used when setting up callouts for the TCP stack.
This ensures that several lingering state checks can simply be removed.

MFC after: 1 week
Sponsored by: NVIDIA Networking

[rscheff]

Shouldn't these flags be cleared in the tcp_timer_stop() function?

[glebius]

I think this patch is in the right direction. Does it allow to assert that now tcp_freecb() is called only and only from tcp_discardcb()?

[• hselasky]

glebius@ : I will test a bit first. I wasn't sure what is guaranteed to be called in which order, but I think you are right, that the additional stop of the timers you commented about is not needed. I just checked where the two flags checked are set (DROPPED and TIMEDWAIT), and stopped all timers from there, but I guess it can't be dropped before going into TIMEDWAIT first ...

[glebius]

In D36622#831576, @hselasky wrote:

glebius@ : I will test a bit first. I wasn't sure what is guaranteed to be called in which order, but I think you are right, that the additional stop of the timers you commented about is not needed. I just checked where the two flags checked are set (DROPPED and TIMEDWAIT), and stopped all timers from there, but I guess it can't be dropped before going into TIMEDWAIT first ...

That's why I would prefer to not rush with your patch and go with my sequence of changes, that already is in the progress:

1. Remove compressed time wait.
2. Merge inpcb+tcpcb, so that the structure that contains the lock is the same structure we protect in the callouts.
3. Go with the patch that would be very close to this one.

[• hselasky]

glebius@: I have no rush with this patch.