cvs commit: src/sys/dev/usb ugen.c

[Poul-Henning Kamp]

In message <20040908.145846.71090050.imp at bsdimp.com>, "M. Warner Losh" writes:
>It turns out that Brian's change typically is the same as the old
>code.  However, there are a few cases where it isn't:
>
>(1) Where the application doesn't close the device when it gets an
>    error.  Stupid application, I know, but in this case the device
>    never detaches.  This is one of the things that caused our
>    appliction to die with the new code, but not the old code.  phk's
>    work with deadfs and dev_t ref counting may help here.

The intent is that devfs will act as isolator for drivers which
disintegrate, in other words, even if the application does not close
and go away, devfs will make it go away from the drivers point of
view.   The refcounting necessary is not in place currently, I ran
into a funky detail and got derailed.

Problem is that we can't just declare that destroy_dev() will evict
all threads because we do not know what they might be sleeping
on/for while they are inside the driver.   A two phase scheme is
necesary where the driver first blocks future entry, then evicts
any threads already in there and then calls destroy_dev().

I've done some work on this based on the phk_tty branch in p4, but
it was pretty exploratory.   The challenge with ttys is that the
struct tty might referenced as controlling tty even after the last
close.

Once I get this sorted out, I'll aim at getting phk_tty MFP4'ed.

-- 
Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
phk at FreeBSD.ORG         | TCP/IP since RFC 956
FreeBSD committer       | BSD since 4.3-tahoe    
Never attribute to malice what can adequately be explained by incompetence.