[RFC] cdrom weirdness

[RFC] cdrom weirdness

[Al Viro]

1) CDROM_LOCKDOOR sets a global variable (keeplocked) that affects all
cdroms.  Intentional?

2) cdrom_dvd_rw_close_write() call can be delayed indefinitely by keeping
an ioctl-only (opened with O_NDELAY) descriptor.

3) open cdrom for data, have the door locked, keep fd opened.
   open it again for write, have the open fail and cleanup in cdrom_open()
   will happily unlock the door for you.  I'd change that to "lock if we
   had no lockers, unlock on failure exit if we did lock", but there's
   an interesting comment:
        /* Something failed.  Try to unlock the drive, because some drivers
        (notably ide-cd) lock the drive after every command.
	...
   What the hell is that about?  It's not "some drivers", AFAICT - it's
   been done explicitly in open_for_data().  Or is there something
   really driver-specific in it?

4) while we are at it, if you clear lockdoor via sysctl while something has
   cdrom opened - no unlock on close for you.

5) autoeject happens on the last close *IF* the last file happens to be
   opened for data.  IOW, if some crap has opened it ioctl-only and kept
   that opened after everyone else has closed - no autoeject for you.

6)	/*
	 * flush cache on last write release
	 */
	if (CDROM_CAN(CDC_RAM) && !cdi->use_count && cdi->for_data)
		cdrom_close_write(cdi);
   is interesting, seeing that nothing has ever touched ->for_data, for
   values of "ever" including "since the code in question had been merged
   into the tree"...

Re: [RFC] cdrom weirdness

[Jens Axboe]

On Thu, Dec 04 2008, Al Viro wrote:
> 1) CDROM_LOCKDOOR sets a global variable (keeplocked) that affects all
> cdroms.  Intentional?

It's always been so, predates me. Pretty ugly, I don't think anyone
would mind if that was changed to be per-device when the ioctl is ussed
:-)

> 2) cdrom_dvd_rw_close_write() call can be delayed indefinitely by keeping
> an ioctl-only (opened with O_NDELAY) descriptor.

Yep, not sure what you want me to say here...

> 3) open cdrom for data, have the door locked, keep fd opened.
>    open it again for write, have the open fail and cleanup in cdrom_open()
>    will happily unlock the door for you.  I'd change that to "lock if we
>    had no lockers, unlock on failure exit if we did lock", but there's
>    an interesting comment:
>         /* Something failed.  Try to unlock the drive, because some drivers
>         (notably ide-cd) lock the drive after every command.
> 	...
>    What the hell is that about?  It's not "some drivers", AFAICT - it's
>    been done explicitly in open_for_data().  Or is there something
>    really driver-specific in it?
> 
> 4) while we are at it, if you clear lockdoor via sysctl while something has
>    cdrom opened - no unlock on close for you.
> 
> 5) autoeject happens on the last close *IF* the last file happens to be
>    opened for data.  IOW, if some crap has opened it ioctl-only and kept
>    that opened after everyone else has closed - no autoeject for you.

Most of the above are long known issues with not counting
write/non-write/ioctls opens, since it was tricky/impossible to do
because of fcntl().

> 6)	/*
> 	 * flush cache on last write release
> 	 */
> 	if (CDROM_CAN(CDC_RAM) && !cdi->use_count && cdi->for_data)
> 		cdrom_close_write(cdi);
>    is interesting, seeing that nothing has ever touched ->for_data, for
>    values of "ever" including "since the code in question had been merged
>    into the tree"...

Hmm weird, you are right. The member was added in 2.6.2, but never used
except here. I guess this just needs to use opened_for_data.

-- 
Jens Axboe