Re: Question about TTY_DO_WRITE_WAKEUP

[Thomas Pfaff <tpfaff@pcs.com>]

On Thu, 20 Nov 2008, Alan Cox wrote:

> > But n_tty_write_wakeup is only called from tty_wakeup when this bit is already 
> > set, therefore it makes no sense to set this bit in n_tty_write_wakeup again.
> 
> The base code should probably really use test_and_clear_bit() when
> calling that method.

Why should you test it, clear it and set it again in n_tty_write_wakeup ?

> > 
> > The flow looks to me as
> > 
> > If the tty driver sets TTY_DO_WRITE_WAKEUP a SIGIO is generated on every 
> > tty_wakeup.
> > 
> > If it is not set then n_tty_write_wakeup is never called and a SIGIO is not 
> > generated.
> 
> Which isn't perfect (excess SIGIO cases) but doesn't seem incorrect. If
> you've not blocked the tty output buffer then write() has not returned a
> short write and no SIGIO is due.
> 

Of course this is not incorrect, but this does not solve my problem with the 
TTY_DO_WRITE_WAKEUP flag.

IMHO a SIGIO on write possible should always be generated if the user wants it, 
currently it is generated when the user wants it and the tty driver enables the 
TTY_DO_WRITE_WAKEUP flag. Unfortunately most drivers don't set it.

Regarding excess SIGIO cases:

Once a write fails with EAGAIN a flag can be set and only in that case a SIGIO is 
generated, afterwards the bit is cleared. Maybe that is what TTY_DO_WRITE_WAKEUP 
was intended for. 

Thomas