atom feed26 messages in org.freebsd.freebsd-currentRe: device_attach(9) and driver initi...
FromSent OnAttachments
Konstantin BelousovApr 7, 2012 5:50 am 
Ian LeporeApr 7, 2012 7:46 am 
Konstantin BelousovApr 7, 2012 7:57 am 
Nathan WhitehornApr 7, 2012 8:04 am 
Konstantin BelousovApr 7, 2012 8:14 am 
Nathan WhitehornApr 7, 2012 8:48 am 
Julian ElischerApr 7, 2012 2:31 pm 
Julian ElischerApr 7, 2012 2:34 pm 
Warner LoshApr 7, 2012 8:08 pm 
Warner LoshApr 7, 2012 8:10 pm 
Warner LoshApr 7, 2012 8:12 pm 
Warner LoshApr 7, 2012 8:14 pm 
Konstantin BelousovApr 7, 2012 8:58 pm 
Ian LeporeApr 9, 2012 7:07 am 
Ian LeporeApr 9, 2012 7:40 am 
Konstantin BelousovApr 9, 2012 7:58 am 
John BaldwinApr 9, 2012 8:00 am 
Ian LeporeApr 9, 2012 8:41 am 
Warner LoshApr 9, 2012 8:57 am 
Konstantin BelousovApr 9, 2012 12:09 pm 
John BaldwinApr 9, 2012 12:35 pm 
Konstantin BelousovApr 9, 2012 1:05 pm 
John BaldwinApr 10, 2012 6:55 am 
Konstantin BelousovApr 10, 2012 7:38 am 
John BaldwinApr 10, 2012 11:47 am 
Konstantin BelousovApr 10, 2012 11:31 pm 
Subject:Re: device_attach(9) and driver initialization
From:Konstantin Belousov (kost@gmail.com)
Date:Apr 9, 2012 12:09:37 pm
List:org.freebsd.freebsd-current

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 @@ -43,6 +43,10 @@ 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; @@ -199,6 +203,21 @@ METHOD int attach { };

/** + * @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. + * + * @param dev the device to probe + * + * @see DEVICE_ATTACH() + */ +METHOD void after_attach { + device_t dev; +} DEFAULT null_after_attach; + +/** * @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 @@ -2743,6 +2743,7 @@ 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.

Pointing driver authors to destroy_dev_sched() for this purpose is overkill, since average driver author, me included, could not use destroy_dev_sched() properly, at least without lot of works.