Re: [git pull] drm for 4.19-rc1
From: John Stultz
Date: Thu Aug 16 2018 - 16:38:52 EST
On Thu, Aug 16, 2018 at 12:16 AM, Daniel Vetter <daniel@xxxxxxxx> wrote:
> On Thu, Aug 16, 2018 at 8:04 AM, John Stultz <john.stultz@xxxxxxxxxx> wrote:
>> On Tue, Aug 14, 2018 at 7:53 PM, Dave Airlie <airlied@xxxxxxxxx> wrote:
>>> This is the main drm pull request for 4.19.
>>>
>>> Rob has some new hardware support for new qualcomm hw that I'll send along
>>> separately. This has the display part of it, the remaining pull is for the
>>> acceleration engine.
>>>
>>> This also contains a wound-wait/wait-die mutex rework, Peter has acked it
>>> for merging via my tree.
>>>
>>> Otherwise mostly the usual level of activity.
>>
>> Hey Folks,
>> Since this branch landed, I've been seeing the following panic on
>> bootup w/ the HiKey board (which uses the hisilicon/kirin drm driver):
>>
>> [ 8.088388] Unable to handle kernel read from unreadable memory at
>> virtual address 0000000000000030
>> [ 8.088393] Mem abort info:
>> [ 8.088397] ESR = 0x96000005
>> [ 8.088402] Exception class = DABT (current EL), IL = 32 bits
>> [ 8.088406] SET = 0, FnV = 0
>> [ 8.088410] EA = 0, S1PTW = 0
>> [ 8.088413] Data abort info:
>> [ 8.088417] ISV = 0, ISS = 0x00000005
>> [ 8.088421] CM = 0, WnR = 0
>> [ 8.088427] user pgtable: 4k pages, 39-bit VAs, pgdp = (____ptrval____)
>> [ 8.088432] [0000000000000030] pgd=0000000000000000, pud=0000000000000000
>> [ 8.088443] Internal error: Oops: 96000005 [#1] PREEMPT SMP
>> [ 8.088453] CPU: 5 PID: 1414 Comm: kworker/5:2 Tainted: G W
>> 4.18.0-07439-gbf1fba4 #633
>> [ 8.088457] Hardware name: HiKey Development Board (DT)
>> [ 8.088474] Workqueue: events adv7511_hpd_work
>> [ 8.088482] pstate: 40400005 (nZcv daif +PAN -UAO)
>> [ 8.088493] pc : drm_sysfs_hotplug_event+0x40/0x78
>> [ 8.088499] lr : drm_sysfs_hotplug_event+0x40/0x78
>> [ 8.088502] sp : ffffff800ba73d20
>> [ 8.088506] x29: ffffff800ba73d20 x28: 0000000000000000
>> [ 8.088514] x27: ffffff8009293cd8 x26: ffffffc074e55938
>> [ 8.088522] x25: 0000000000000000 x24: ffffffc07ff85000
>> [ 8.088530] x23: ffffffc0742c4a78 x22: ffffffc07ff86c00
>> [ 8.088537] x21: ffffffc0750d0e00 x20: 0000000000000000
>> [ 8.088545] x19: ffffff8009009a48 x18: 0000000000000000
>> [ 8.088552] x17: 0000000000000000 x16: ffffffc074fbde80
>> [ 8.088560] x15: 0000000000000000 x14: ffffffc005f96c00
>> [ 8.088568] x13: 00000040770c9000 x12: 0000000034d5d91d
>> [ 8.088575] x11: 0000000000000000 x10: 0000000000000990
>> [ 8.088582] x9 : ffffff800ba739b0 x8 : ffffff800913e000
>> [ 8.088589] x7 : 0000000000000000 x6 : ffffff8009009a48
>> [ 8.088596] x5 : ffffff80090588d0 x4 : 0000000000000000
>> [ 8.088602] x3 : ffffff8009009a48 x2 : 0000000000000000
>> [ 8.088608] x1 : 18701cfc97cf1200 x0 : 0000000000000000
>> [ 8.120775] Process kworker/5:2 (pid: 1414, stack limit = 0x(____ptrval____))
>> [ 8.120778] Call trace:
>> [ 8.120787] drm_sysfs_hotplug_event+0x40/0x78
>> [ 8.120794] drm_kms_helper_hotplug_event+0x14/0x40
>> [ 8.120800] adv7511_hpd_work+0x64/0xe0
>> [ 8.120807] process_one_work+0x12c/0x320
>> [ 8.120814] worker_thread+0x48/0x458
>> [ 8.126654] kthread+0xf8/0x128
>> [ 8.126661] ret_from_fork+0x10/0x18
>> [ 8.126672] Code: aa0003f4 52800020 a902ffa2 94006637 (f9401a80)
>> [ 8.135638] ---[ end trace cf7120942e6f40fa ]---
>>
>> And earlier in boot we see:
>>
>> [ 4.620909] kirin-drm f4100000.ade: bound f4107800.dsi (ops dsi_ops)
>> [ 4.627304] [drm] Supports vblank timestamp caching Rev 2 (21.10.2013).
>> [ 4.633935] [drm] No driver support for vblank timestamp query.
>> [ 4.732910] kirin-drm f4100000.ade: [drm:drm_fb_helper_fbdev_setup]
>> *ERROR* Failed to set fbdev configuration
>> [ 4.742948] [drm:kirin_drm_bind] *ERROR* failed to initialize fbdev.
>> [ 4.749585] kirin-drm f4100000.ade: master bind failed: -22
>> [ 4.755218] dw-dsi: probe of f4107800.dsi failed with error -22
>>
>> I've also seen similar trouble w/ the HiKey960 which uses a similar
>> but still out of tree driver that also utilizes the cma fbhelper code,
>> which makes me suspect it has to do with the drm/cma-helper changes
>> below:
>>
>>> Noralf TrÃnnes (15):
>>> drm/file: Don't set master on in-kernel clients
>>> drm: Make ioctls available for in-kernel clients
>>> drm: Begin an API for in-kernel clients
>>> drm/fb-helper: Add generic fbdev emulation .fb_probe function
>>> drm/pl111: Set .gem_prime_vmap and .gem_prime_mmap
>>> drm/cma-helper: Use the generic fbdev emulation
>>> drm/debugfs: Add internal client debugfs file
>>> drm/fb-helper: Finish the generic fbdev emulation
>>> drm/tinydrm: Use drm_fbdev_generic_setup()
>>> drm/cma-helper: Remove drm_fb_cma_fbdev_init_with_funcs()
>>
>> Though I've not yet had time to bisect this down tonight.
>>
>> I'll spend some more time on this tomorrow, but wanted to give folks a
>> heads up in the meantime.
>
> Hm, not immediately seeing what's going boom here. Bisect would indeed
> be good, but maybe we need to chase the callchain to figure out where
> exactly that -EINVAL is coming from in the reworked code (and why
> hikey is the first to hit that, there's lots of cma based drivers
> after all).
So I can confirm reverting:
85b5bafb86e6 ("drm/cma-helper: Remove drm_fb_cma_fbdev_init_with_funcs()")
and
894a677f4b3e ("drm/cma-helper: Use the generic fbdev emulation")
makes things work again.
Though without the reverts, I was digging to understand what was
failing, and it seems when we call into
drm_client_framebuffer_create(), as Noralf suggested, we're eventually
failing at drm_internal_framebuffer_create() when checking
config->max_height. The kirin driver is somewhat limited w/ max
width/height values at 2048 each.
Though I'm on a 1920x1080 monitor (max resolution supported by the
board), it seems we're trying to allocate a 1920x2160 buffer? I know
Android tries to allocate 2x Y for page-flipping, but I'm not sure yet
why the old code didn't hit this problem.
Unfortunately bumping the max width/height values to 4096 cause the
system to hard hang as userspace starts up (and setting it to
2048/2160 does the same).
So yea, I'm going to continue to dig at the hard-hang issue, and
hopefully we can just bump the max width/height, but if there are
other suggestions, please let me know.
thanks
-john