Re: The Serial Layer

["Theodore Ts'o" <tytso@mit.edu>]

On Thu, Sep 09, 2004 at 05:28:46PM +0100, Alan Cox wrote:
> On Iau, 2004-09-09 at 13:57, Theodore Ts'o wrote:
> > Calling ldisc.receive_buf directly() should be OK as long as you're
> > not in an interrupt handler.  (For example the comtrol driver polls in
> > a timer bottom-half, so it's OK to call ldisc.receive_buf).
> > Unfortunately, some drivers don't quite follow the rules.
> 
> If the driver calls ldisc->receive_buf it means it bypasses the
> TTY_DONT_FLIP locking used by the n_tty driver. Am I missing something
> here.

Um, yes, you're right.  The direct calls to ldisc->receive_buf predate
TTY_DONT_FLIP; TTY_DONT_FLIP was added by Bill Hawes (if I remember
correctly) in an attempt to fix various locking problems, but
unfortunately the direct callers weren't fixed.

> Most of that I think comes out in the wash with the ldisc locking. 
> Once the driver uses tty_ldisc_ref() it'll get NULL back in the case
> where the hangup raced the driver. I'm also no longer dropping back
> to N_TTY in the hangup path. I couldn't see any reason this was
> neccessary since the release will clean it up anyway ?

We needed to close the line discpline in order to prevent a line
discipline (such as ppp) from trying to write to the tty.  Given there
was an invariant that a tty always had a line discpline always, we
reassigned it back to N_TTY.  The right answer is probably to be to
add checks to the line discpline code before they attempt to send data
to the tty.

						- Ted