On 08/12/15 10:19, Daniel Vetter wrote:
On Mon, Dec 07, 2015 at 07:32:42PM +0200, Tomi Valkeinen wrote:
On 25/08/15 16:45, Daniel Vetter wrote:
When the usual fbcon legacy options are enabled we have ->register_framebuffer ->fb notifier chain calls into fbcon ->fbcon sets up console on new fbi ->fbi->set_par ->drm_fb_helper_set_par exercises full kms api
And because of locking inversion hilarity all of register_framebuffer is done with the console lock held. Which means that the first time on driver load we exercise _all_ the kms code (all probe paths and modeset paths for everything connected) is under the console lock. That means if anything goes belly-up in that big pile of code nothing ever reaches logfiles (and the machine is dead).
Usual tactic to debug that is to temporarily remove those console_lock calls to be able to capture backtraces. I'm fed up writing this patch and recompiling kernels. Hence this patch here to add an unsafe, kernel-taining option to do this at runtime.
I think this was never merged. This was part 4 of 4, were there dependencies or...? Should I apply this for 4.5?
Patches 1-3 have all already landed in drm, and patch 4 is free standing. Would be great if you can pull it in.
Ok, I'll apply for 4.5.
But then... I think my issues with console lock have been later at runtime, not at register. Maybe we need a module option to disable the console lock altogether. I wonder how much havoc that might create, though.
Hm, where in fbdev do you hold the console_lock outside of registering/unregistering an fbdev (because of fbcon)? There's of course general trouble with console_lock deadlocks and fun like that, but ime that all got a lot more manageable since I added lockdep annotations to console_lock.
I don't know... I just have a vague recollection that I was having trouble with the lock with... crashes during blanking, perhaps. I really can't remember, so possibly things are better now, or I just remember wrong.
Tomi