Hi Alexandre,
On 12/04/2021 10:28, Alexandre Courbot wrote:
> Hi Hans,
>
> I have been able to reproduce the issue with the stateless part of the
> mtk-vcodec decoder driver (not merged as of today), which shows the
> exact same crash.
>
> With your patch applied I got the following output right before the
> crash, when decoding H.264 or VP9:
>
> H.264:
> [ 549.264784] v4l2-ctrls: v4l2_ctrl_request_complete 990a6b H264 Profile
> [ 976.630017] v4l2-ctrls: v4l2_ctrl_request_complete 980001 User Controls
> [ 125.152149] v4l2-ctrls: v4l2_ctrl_request_complete 980001 User Controls
> [ 330.049192] v4l2-ctrls: v4l2_ctrl_request_complete 990001 Codec Controls
>
> VP9:
> [ 215.945812] v4l2-ctrls: v4l2_ctrl_request_complete 990a6b H264 Profile
> [ 428.564501] v4l2-ctrls: v4l2_ctrl_request_complete 990a6b H264 Profile
> [ 830.970340] v4l2-ctrls: v4l2_ctrl_request_complete 990b00 VP9 Profile
>
> The other pr_info() in your patch was never seen.
>
> Interestingly these controls are not set by requests, and are not even
> supposed to be set on a decoder. I am not quite sure where this could
> come from...
I found the issue and Ezequiel successfully tested a patch I mailed him privately.
The patch needs a bit of cleanup, but I plan to post it today.
Regards,
Hans
>
> Cheers,
> Alex.
>
> On Thu, Apr 1, 2021 at 6:36 PM Hans Verkuil <hverkuil-cisco@xxxxxxxxx> wrote:
>>
>> On 01/04/2021 00:34, Ezequiel Garcia wrote:
>>> Hi Hans,
>>>
>>> On Wed, 31 Mar 2021 at 13:14, Ezequiel Garcia
>>> <ezequiel@xxxxxxxxxxxxxxxxxxxx> wrote:
>>>>
>>>> Hi Hans,
>>>>
>>>> We have found this crash in mainline:
>>>>
>>>> [ 258.174662] Unable to handle kernel NULL pointer dereference at
>>>> virtual address 0000000000000000
>>>> [ 258.182967] Mem abort info:
>>>> [ 258.187746] ESR = 0x96000004
>>>> [ 258.192100] EC = 0x25: DABT (current EL), IL = 32 bits
>>>> [ 258.199108] SET = 0, FnV = 0
>>>> [ 258.203732] EA = 0, S1PTW = 0
>>>> [ 258.208389] Data abort info:
>>>> [ 258.212743] ISV = 0, ISS = 0x00000004
>>>> [ 258.218322] CM = 0, WnR = 0
>>>> [ 258.222760] user pgtable: 4k pages, 48-bit VAs, pgdp=00000000444b0000
>>>> [ 258.230822] [0000000000000000] pgd=0000000000000000, p4d=0000000000000000
>>>> [ 258.239097] Internal error: Oops: 96000004 [#1] PREEMPT SMP
>>>> [ 258.246162] Modules linked in: hantro_vpu(C) videobuf2_vmalloc
>>>> v4l2_h264 videobuf2_dma_contig videobuf2_memops v4l2_mem2mem
>>>> videobuf2_v4l2 videobuf2_common videodev mc etnaviv fsl_imx8_ddr_perf
>>>> gpu_sched fuse
>>>> [ 258.263431] CPU: 0 PID: 34 Comm: kworker/0:1 Tainted: G C
>>>> 5.12.0-rc2+ #106
>>>> [ 258.270312] Hardware name: NXP i.MX8MQ EVK (DT)
>>>> [ 258.273542] Workqueue: events v4l2_m2m_device_run_work [v4l2_mem2mem]
>>>> [ 258.278716] pstate: 80000005 (Nzcv daif -PAN -UAO -TCO BTYPE=--)
>>>> [ 258.283426] pc : __memcpy+0x7c/0x180
>>>> [ 258.285709] lr : v4l2_ctrl_request_complete+0x7c/0x200 [videodev]
>>>> [ 258.290585] sp : ffff8000116fbc60
>>>> [ 258.292598] x29: ffff8000116fbc60 x28: ffff80001142b780
>>>> [ 258.296617] x27: ffff80001127bc80 x26: 00000000e9b00000
>>>> [ 258.300637] x25: 00000000e9a00000 x24: ffff0000067e2500
>>>> [ 258.304653] x23: ffff0000067e2538 x22: ffff00000387d238
>>>> [ 258.308671] x21: ffff0000067e2598 x20: ffff000004644e00
>>>> [ 258.312688] x19: ffff000006812a00 x18: 0000000000000000
>>>> [ 258.316708] x17: 0000000000000000 x16: 0000000000000000
>>>> [ 258.320724] x15: 0000000000000000 x14: 0000000000000000
>>>> [ 258.324744] x13: 0000000000000000 x12: 0000000000000000
>>>> [ 258.328761] x11: 0000000000000000 x10: 0000000000000000
>>>> [ 258.332776] x9 : 0000000000000000 x8 : 0000000000000000
>>>> [ 258.336793] x7 : 0000000000000000 x6 : ffff000006812a40
>>>> [ 258.340811] x5 : 0000000000000006 x4 : 0000000000000000
>>>> [ 258.344826] x3 : 0000000000000010 x2 : 0000000000000010
>>>> [ 258.348845] x1 : 0000000000000000 x0 : ffff000006812a40
>>>> [ 258.352864] Call trace:
>>>> [ 258.354010] __memcpy+0x7c/0x180
>>>> [ 258.355941] hantro_end_prepare_run+0x2c/0x60 [hantro_vpu]
>>>> [ 258.360150] hantro_g1_mpeg2_dec_run+0x3b0/0x7f0 [hantro_vpu]
>>>> [ 258.364615] device_run+0xa8/0xbc [hantro_vpu]
>>>> [ 258.367777] v4l2_m2m_try_run+0x84/0x134 [v4l2_mem2mem]
>>>> [ 258.371723] v4l2_m2m_device_run_work+0x14/0x20 [v4l2_mem2mem]
>>>> [ 258.376279] process_one_work+0x1c0/0x484
>>>> [ 258.378995] worker_thread+0x70/0x434
>>>> [ 258.381364] kthread+0x158/0x160
>>>> [ 258.383296] ret_from_fork+0x10/0x34
>>>> [ 258.385582] Code: a8c12027 a88120c7 a8c12027 a88120c7 (a8c12027)
>>>> [ 258.390377] ---[ end trace fea6ecb96dad642d ]---
>>>>
>>>> This needs several concurrent streams to run at the same time,
>>>> it doesn't happen with just one; it was reproduced in rkvdec
>>>> and hantro drivers. The above stacktrace is on MPEG-2,
>>>> but we have found it on H.264 as well.
>>>>
>>>> I've traced this to:
>>>>
>>>> v4l2_ctrl_lock(ctrl);
>>>> if (ref->req) {
>>>> ptr_to_ptr(ctrl, ref->req->p_req, ref->p_req);
>>>>
>>>> And:
>>>>
>>>> static void ptr_to_ptr(struct v4l2_ctrl *ctrl,
>>>> union v4l2_ctrl_ptr from, union v4l2_ctrl_ptr to)
>>>> {
>>>> if (ctrl == NULL)
>>>> return;
>>>> memcpy(to.p, from.p_const, ctrl->elems * ctrl->elem_size);
>>>> ^^^^^^^^^^^^
>>>> This is NULL
>>>>
>>>> I've been staring at this file and trying to understand how this can happen
>>>> but haven't been able to make any progress so far...
>>>>
>>>> Any ideas ?
>>
>> Can you test with this patch:
>>
>> ---------------------------------------
>> diff --git a/drivers/media/v4l2-core/v4l2-ctrls.c b/drivers/media/v4l2-core/v4l2-ctrls.c
>> index 757d215c2be4..19c56cb9f86a 100644
>> --- a/drivers/media/v4l2-core/v4l2-ctrls.c
>> +++ b/drivers/media/v4l2-core/v4l2-ctrls.c
>> @@ -2703,8 +2703,11 @@ static int handler_new_ref(struct v4l2_ctrl_handler *hdl,
>> if (hdl->error)
>> return hdl->error;
>>
>> - if (allocate_req)
>> + if (allocate_req) {
>> size_extra_req = ctrl->elems * ctrl->elem_size;
>> + if (!size_extra_req)
>> + pr_info("empty size_extra_req %x %s\n", ctrl->id, ctrl->name);
>> + }
>> new_ref = kzalloc(sizeof(*new_ref) + size_extra_req, GFP_KERNEL);
>> if (!new_ref)
>> return handler_set_err(hdl, -ENOMEM);
>> @@ -4610,6 +4613,8 @@ void v4l2_ctrl_request_complete(struct media_request *req,
>>
>> v4l2_ctrl_lock(ctrl);
>> if (ref->req) {
>> + if (!ref->req->p_req.p)
>> + pr_info("v4l2_ctrl_request_complete %x %s\n", ctrl->id, ctrl->name);
>> ptr_to_ptr(ctrl, ref->req->p_req, ref->p_req);
>> } else {
>> ptr_to_ptr(ctrl, ctrl->p_cur, ref->p_req);
>> ---------------------------------------
>>
>> I want to know if there are controls where size_extra_req == 0 (meaning
>> that p_req.p is also NULL), and which control encounters the NULL p_req.p.
>>
>> Regards,
>>
>> Hans
