Re: device_attach(9) and driver initialization

On Monday, April 09, 2012 4:05:29 pm Konstantin Belousov wrote:
> On Mon, Apr 09, 2012 at 03:36:08PM -0400, John Baldwin wrote:
> > On Monday, April 09, 2012 3:10:00 pm Konstantin Belousov wrote:
> > > On Mon, Apr 09, 2012 at 11:01:03AM -0400, John Baldwin wrote:
> > > > On Saturday, April 07, 2012 11:08:41 pm Warner Losh wrote:
> > > > > 
> > > > > On Apr 7, 2012, at 8:46 AM, Ian Lepore wrote:
> > > > > 
> > > > > > On Sat, 2012-04-07 at 15:50 +0300, Konstantin Belousov wrote:
> > > > > >> Hello,
> > > > > >> there seems to be a problem with device attach sequence offered by 
> > > > newbus.
> > > > > >> Basically, when device attach method is executing, device is not fully
> > > > > >> initialized yet. Also the device state in the newbus part of the world
> > > > > >> is DS_ALIVE. There is definitely no shattering news in the statements,
> > > > > >> but drivers that e.g. create devfs node to communicate with consumers
> > > > > >> are prone to a race.
> > > > > >> 
> > > > > >> If /dev node is created inside device attach method, then usermode
> > > > > >> can start calling cdevsw methods before device fully initialized itself.
> > > > > >> Even more, if device tries to use newbus helpers in cdevsw methods,
> > > > > >> like device_busy(9), then panic occurs "called for unatteched device".
> > > > > >> I get reports from users about this issues, to it is not something
> > > > > >> that only could happen.
> > > > > >> 
> > > > > >> I propose to add DEVICE_AFTER_ATTACH() driver method, to be called
> > > > > >> from newbus right after device attach finished and newbus considers
> > > > > >> the device fully initialized. Driver then could create devfs node
> > > > > >> in the after_attach method instead of attach. Please see the patch below.
> > > > > >> 
> > > > > >> diff --git a/sys/kern/device_if.m b/sys/kern/device_if.m
> > > > > >> index eb720eb..9db74e2 100644
> > > > > >> --- a/sys/kern/device_if.m
> > > > > >> +++ b/sys/kern/device_if.m
> > > > > >> _at__at_ -43,6 +43,10 _at__at_ INTERFACE device;
> > > > > >> # Default implementations of some methods.
> > > > > >> #
> > > > > >> CODE {
> > > > > >> +	static void null_after_attach(device_t dev)
> > > > > >> +	{
> > > > > >> +	}
> > > > > >> +
> > > > > >> 	static int null_shutdown(device_t dev)
> > > > > >> 	{
> > > > > >> 	    return 0;
> > > > > >> _at__at_ -199,6 +203,21 _at__at_ METHOD int attach {
> > > > > >> };
> > > > > >> 
> > > > > >> /**
> > > > > >> + * _at_brief Notify the driver that device is in attached state
> > > > > >> + *
> > > > > >> + * Called after driver is successfully attached to the device and
> > > > > >> + * corresponding device_t is fully operational. Driver now may expose
> > > > > >> + * the device to the consumers, e.g. create devfs nodes.
> > > > > >> + *
> > > > > >> + * _at_param dev		the device to probe
> > > > > >> + *
> > > > > >> + * _at_see DEVICE_ATTACH()
> > > > > >> + */
> > > > > >> +METHOD void after_attach {
> > > > > >> +	device_t dev;
> > > > > >> +} DEFAULT null_after_attach;
> > > > > >> +
> > > > > >> +/**
> > > > > >>  * _at_brief Detach a driver from a device.
> > > > > >>  *
> > > > > >>  * This can be called if the user is replacing the
> > > > > >> diff --git a/sys/kern/subr_bus.c b/sys/kern/subr_bus.c
> > > > > >> index d485b9f..6d849cb 100644
> > > > > >> --- a/sys/kern/subr_bus.c
> > > > > >> +++ b/sys/kern/subr_bus.c
> > > > > >> _at__at_ -2743,6 +2743,7 _at__at_ device_attach(device_t dev)
> > > > > >> 	dev->state = DS_ATTACHED;
> > > > > >> 	dev->flags &= ~DF_DONENOMATCH;
> > > > > >> 	devadded(dev);
> > > > > >> +	DEVICE_AFTER_ATTACH(dev);
> > > > > >> 	return (0);
> > > > > >> }
> > > > > >> 
> > > > > > 
> > > > > > Does device_get_softc() work before attach is completed?  (I don't have
> > > > > > time to go look in the code right now).  If so, then a mutex initialized
> > > > > > and acquired early in the driver's attach routine, and also acquired in
> > > > > > the driver's cdev implementation routines before using any newbus
> > > > > > functions other than device_get_softc(), would solve the problem without
> > > > > > a driver api change that would make it harder to backport/MFC driver
> > > > > > changes.
> > > > > 
> > > > > Also, more generally, don't create the dev nodes before you are ready to 
> > > > deal with requests.  Why do we need to uglify everything here?  If you can't 
> > > > do that, you can check a bit in the softc and return EBUSY or ENXIO on open if 
> > > > that bit says that your driver isn't ready to accept requests.
> > > > 
> > > > I agree, this dosen't actually fix anything as the decision for what to put
> > > > in your foo_attach() method rather than foo_after_attach() is non-obvious and 
> > > > very arbitrary.
> > > > 
> > > > The actual bug appears to only be with using 'device_busy()'. I think
> > > > this should be fixed by making device_busy() better, not by adding
> > > > this type of obfuscation to attach. It should be trivial to make
> > > > device_busy() safe to use on a device that is currently being attached
> > > > which will not require any changes to drivers.
> > > 
> > > Could you, please, elaborate your proposal ? How do you think device_busy()
> > > can be enchanced ?
> > > 
> > > Obvious idea to sleep inside device_busy() until dev->state becomes !=
> > > DS_ATTACHED is no go, IMO. The issue is that this causes immediate deadlocks
> > > if device_attach() method needs to call destroy_dev() to rollback.
> > 
> > I think you could have a DS_ATTACHING state and allow device_busy() to work
> > for DS_ATTACHING.  The idea being that it is a bug for a driver to invoke
> > device_busy() if it is going to fail attach.  You may then need to do a fixup
> > in device_attach() to promote the state from DS_ATTACHED to DS_BUSY when it
> > returns if there is a non-zero busy count.
> This is quite good idea, but it still adds burden to device author,
> although I agree that this is manageable. A scenario I have in mind now
> is the following:
> assume that driver needs to create two devfs nodes, lets name them
> dri/card0 and dri/forcewake0. Driver would perform two make_dev_p(9)
> calls, and while creation of dri/card0 succeed, consequent creation
> of dri/forcewake0 could fail for numerous reasons.
> 
> Now, the driver needs to ensure that cdesvw->d_open() on dri/card0
> would return ENXIO until dri/forcewake0 is created. This can be implemented
> with flag, indeed. But still somewhat muddy, and probably leads to
> user-visible errors (I mostly worry about graphical login managers).

You could also sleep on the flag in d_open() (you can imagine a two-step
process where you set a "adding cdev's flag", then all d_open() calls block
on it).  Then when finished adding cdev's, you set a flag if an error
occurred and wake up all the waiters.  If no error occured the waiters can
have the first open() work fine.  But even with other proposals you still
have to deal with this problem if you want to fail out entirely if you
have problems creating cdevs.

> But for single-node drivers it is indeed a nice solution.

I think we are somewhat stuck with this for other reasons as well.  Note
that device_busy() propagates up the tree to parent devices, so imagine
kldloading a driver that creates a tree (e.g. a bus with a few consumers)
where the leaf devices are attached by a call to bus_generic_attach() from
the device_attach() method for the parent device.  Even if you add a
DEVICE_AFTER_ATTACH() hook, while it may allow device_busy() to be invoked
on the leaf device, when it tries to propagate device_busy() up to the
parent device it would still be in the middle of attach and blow up anyway.

-- 
John Baldwin