Skip to contentSkip to conversation

Locking question for DRM

DISCUSSION Alan Cox 2 participants 5 messages Apr 23, 2010

lore
Share & tools
lore

Alan Cox

author
#0 lore raw

Various bits of the DRM deal with minor->master:

In the case of the open helper its protected by the struct mutex.

In the release path it's protected on some paths, but not this one ...

     /* if the master has gone away we can't do anything with the lock */
        if (file_priv->minor->master)
                drm_master_release(dev, filp);

and I can't see what makes this safe if the drm_release for the master and a client occur at the same time ?

The setmaster/dropmaster ioctl seems similar - the various conditional checks are not protected from parallel changes occuring during their execution.

Is this a bug or is something clever afoot ?

Alan

#1 lore raw
Quoted 5 lines · Alan Cox

Various bits of the DRM deal with minor->master:

In the case of the open helper its protected by the struct mutex.

Quoted 8 lines

In the release path it's protected on some paths, but not this one ...

    /* if the master has gone away we can't do anything with the lock */ if (file_priv->minor->master)
               drm_master_release(dev, filp);

and I can't see what makes this safe if the drm_release for the master and a client occur at the same time ?

lock_kernel in drm_release. We probably need to clean that up.

Quoted 6 lines

The setmaster/dropmaster ioctl seems similar - the various conditional checks are not protected from parallel changes occuring during their execution.

Is this a bug or is something clever afoot ?

These ioctls are also under the BKL.

So yes its nasty, and we should probably grow a minor lock to protect that.

Dave.

author
#2 lore raw
Quoted 4 lines
Quoted 2 lines

and I can't see what makes this safe if the drm_release for the master and a client occur at the same time ?

lock_kernel in drm_release. We probably need to clean that up.

I don't think that works. drm_open_helper doesn't appear to be under the BKL merely the struct mutex.

Quoted 7 lines
Quoted 5 lines

The setmaster/dropmaster ioctl seems similar - the various conditional checks are not protected from parallel changes occuring during their execution.

Is this a bug or is something clever afoot ?

These ioctls are also under the BKL.

But setmaster can sleep so the BKL is dropped on contention of the struct_mutex, ditto dropmaster

Alan

#3 lore raw
Quoted 7 lines · Alan Cox
Quoted 4 lines
Quoted 2 lines

and I can't see what makes this safe if the drm_release for the master and a client occur at the same time ?

lock_kernel in drm_release. We probably need to clean that up.

I don't think that works. drm_open_helper doesn't appear to be under the BKL merely the struct mutex.

It blocks the case you specified of two releases happening together.

Quoted 11 lines
Quoted 8 lines
Quoted 6 lines

The setmaster/dropmaster ioctl seems similar - the various conditional checks are not protected from parallel changes occuring during their execution.

Is this a bug or is something clever afoot ?

These ioctls are also under the BKL.

But setmaster can sleep so the BKL is dropped on contention of the struct_mutex, ditto dropmaster

they should only sleep in the mutex lock nuless the driver callback is allocating memory. but yeah its a bit of a mess.

Dave.

author
#4 lore raw
Quoted 4 lines
Quoted 2 lines

I don't think that works. drm_open_helper doesn't appear to be under the BKL merely the struct mutex.

It blocks the case you specified of two releases happening together.

But not parallel open/release

Quoted 5 lines
Quoted 2 lines

But setmaster can sleep so the BKL is dropped on contention of the struct_mutex, ditto dropmaster

they should only sleep in the mutex lock nuless the driver callback is allocating memory. but yeah its a bit of a mess.

With the mutex alone the damage is done. Consider two setmasters and some other action which is making the mutex contend

		CPU1				CPU2

file->priv->minor->master == NULL ?
file_priv->minor->master != file_priv->master

mutex_lock (drop BKL)

				file->priv->minor->master === NULL
				master != file_priv->master
				mutex_lock   [drop BKL]

takes mutex
minor->master = drm_master_get
is_master = 1
master_set
drop mutex
return 0

				takes mutex
				minor->master = drm_master_get
				is_master = 1
				master_set
				drops mutex
				return 0

Alan