Re: [PATCH v3] drm/ssd130x: Fix an oops when attempting to update a disabled plane

[Javier Martinez Canillas <javierm@redhat.com>]

Maxime Ripard <mripard@redhat.com> writes:

Hello Maxime,

Thanks a lot for your feedback.

> Hi,
>
> On Mon, Jul 17, 2023 at 06:30:22PM +0200, Javier Martinez Canillas wrote:
>> Geert reports that the following NULL pointer dereference happens for him
>> after commit 49d7d581ceaf ("drm/ssd130x: Don't allocate buffers on each
>> plane update"):
>> 
>>     [drm] Initialized ssd130x 1.0.0 20220131 for 0-003c on minor 0
>>     ssd130x-i2c 0-003c: [drm] surface width(128), height(32), bpp(1)
>>     and format(R1   little-endian (0x20203152))
>>     Unable to handle kernel NULL pointer dereference at virtual address 00000000
>>     Oops [#1]
>>     CPU: 0 PID: 1 Comm: swapper Not tainted
>>     6.5.0-rc1-orangecrab-02219-g0a529a1e4bf4 #565
>>     epc : ssd130x_update_rect.isra.0+0x13c/0x340
>>      ra : ssd130x_update_rect.isra.0+0x2bc/0x340
>>     ...
>>     status: 00000120 badaddr: 00000000 cause: 0000000f
>>     [<c0303d90>] ssd130x_update_rect.isra.0+0x13c/0x340
>>     [<c0304200>] ssd130x_primary_plane_helper_atomic_update+0x26c/0x284
>>     [<c02f8d54>] drm_atomic_helper_commit_planes+0xfc/0x27c
>>     [<c02f9314>] drm_atomic_helper_commit_tail+0x5c/0xb4
>>     [<c02f94fc>] commit_tail+0x190/0x1b8
>>     [<c02f99fc>] drm_atomic_helper_commit+0x194/0x1c0
>>     [<c02c5d00>] drm_atomic_commit+0xa4/0xe4
>>     [<c02cce40>] drm_client_modeset_commit_atomic+0x244/0x278
>>     [<c02ccef0>] drm_client_modeset_commit_locked+0x7c/0x1bc
>>     [<c02cd064>] drm_client_modeset_commit+0x34/0x64
>>     [<c0301a78>] __drm_fb_helper_restore_fbdev_mode_unlocked+0xc4/0xe8
>>     [<c0303424>] drm_fb_helper_set_par+0x38/0x58
>>     [<c027c410>] fbcon_init+0x294/0x534
>>     ...
>> 
>> The problem is that fbcon calls fbcon_init() which triggers a DRM modeset
>> and this leads to drm_atomic_helper_commit_planes() attempting to commit
>> the atomic state for all planes, even the ones whose CRTC is not enabled.
>> 
>> Since the primary plane buffer is allocated in the encoder .atomic_enable
>> callback, this happens after that initial modeset commit and leads to the
>> mentioned NULL pointer dereference.
>
> I think that's where the problem lies: you must not allocate a buffer in
> atomic_enable.
>
> After drm_atomic_helper_swap_state(), the new commit is being applied
> and you aren't allowed to fail, and an allocation can fail.
>
> Everything needs to be prepared by the time _swap_state is called, and
> it's one of the point of atomic_check.
>
> So you need to allocate your buffer there, and use it in whatever
> atomic_commit related hook you need it in.
>
> The typical way of doing this would be to create a custom state
> structure that embeds the global one, create your own reset,
> atomic_duplicate_state and atomic_destroy_state implementations, and
> store the buffer pointer there.
>

I see. That makes totally sense to me. I'll do that in v4 then!

> Maxime

-- 
Best regards,

Javier Martinez Canillas
Core Platforms
Red Hat