Re: [RFC PATCH v3 2/4] drm: Create DRM_IOCTL_GET_RESET

["André Almeida" <andrealmeid@igalia.com>]

Em 21/06/2023 05:09, Pekka Paalanen escreveu:
> On Tue, 20 Jun 2023 21:57:17 -0300
> André Almeida <andrealmeid@igalia.com> wrote:
> 
>> Create a new DRM ioctl operation to get the numbers of resets for a
>> given context. The numbers reflect just the resets that happened after
>> the context was created, and not since the machine was booted.
>>
>> Create a debugfs interface to make easier to test the API without real
>> resets.
>>
>> Signed-off-by: André Almeida <andrealmeid@igalia.com>
>> ---
>>   drivers/gpu/drm/drm_debugfs.c |  2 ++
>>   drivers/gpu/drm/drm_ioctl.c   | 58 +++++++++++++++++++++++++++++++++++
>>   include/drm/drm_device.h      |  3 ++
>>   include/drm/drm_drv.h         |  3 ++
>>   include/uapi/drm/drm.h        | 21 +++++++++++++
>>   include/uapi/drm/drm_mode.h   | 15 +++++++++
>>   6 files changed, 102 insertions(+)
> 
> ...
> 
>> diff --git a/include/uapi/drm/drm.h b/include/uapi/drm/drm.h
>> index a87bbbbca2d4..a84559aa0d77 100644
>> --- a/include/uapi/drm/drm.h
>> +++ b/include/uapi/drm/drm.h
>> @@ -1169,6 +1169,27 @@ extern "C" {
>>    */
>>   #define DRM_IOCTL_MODE_GETFB2		DRM_IOWR(0xCE, struct drm_mode_fb_cmd2)
>>   
>> +/**
>> + * DRM_IOCTL_GET_RESET - Get information about device resets
>> + *
>> + * This operation requests from the device information about resets. It should
>> + * consider only resets that happens after the context is created, therefore,
>> + * the counter should be zero during context creation.
>> + *
>> + * dev_reset_count tells how many resets have happened on this device, and
>> + * ctx_reset_count tells how many of such resets were caused by this context.
>> + *
>> + * Flags can be used to tell if a reset is in progress, and userspace should
>> + * wait until it's not in progress anymore to be able to create a new context;
>> + * and to tell if the VRAM is considered lost. There's no safe way to clean this
>> + * flag so if a context see this flag set, it should be like that until the end
>> + * of the context.
> 
> Is "this flag" the VRAM_LOST? Or any flag?
> 
> Does this mean that not all resets are fatal to the context? Is there
> any kind of reset that should not be fatal to a context? All the
> rendering APIs seem to assume that any reset is fatal and the context
> must be destroyed.

I got this flag from the `AMDGPU_CTX_OP_QUERY_STATE2` operation, and 
it's used to notify that the reset was fatal for a giving context, 
although the idea of non-fatal resets seems to be a bit controversial 
for now, so I think it will be better if I leave this flag for latter 
improvements of the API.

> 
>> + */
>> +#define DRM_IOCTL_GET_RESET		DRM_IOWR(0xCF, struct drm_get_reset)
>> +
>> +#define DRM_RESET_IN_PROGRESS	0x1
>> +#define DRM_RESET_VRAM_LOST	0x2
> 
> Ok, so the dmabuf lost is being communicated here, but how would a
> userspace process know on which device a dmabuf resides on?
> 
> Let's assume process A uses device 1 to draw, exports a dmabuf, sends
> it to process B which imports it to device 2. Device 1 resets and loses
> VRAM contents. How would process B notice that the dmabuf is lost when
> it never touches device 1 itself?
> 
>> +
>>   /*
>>    * Device specific ioctls should only be in their respective headers
>>    * The device specific ioctl range is from 0x40 to 0x9f.
>> diff --git a/include/uapi/drm/drm_mode.h b/include/uapi/drm/drm_mode.h
>> index 43691058d28f..c3257bd1af9c 100644
>> --- a/include/uapi/drm/drm_mode.h
>> +++ b/include/uapi/drm/drm_mode.h
>> @@ -1308,6 +1308,21 @@ struct drm_mode_rect {
>>   	__s32 y2;
>>   };
>>   
>> +/**
>> + * struct drm_get_reset - Get information about a DRM device resets
>> + * @ctx_id: the context id to be queried about resets
>> + * @flags: flags
>> + * @dev_reset_count: global counter of resets for a given DRM device
>> + * @ctx_reset_count: of all the resets counted by this device, how many were
>> + * caused by this context.
>> + */
>> +struct drm_get_reset {
>> +	__u32 ctx_id;
>> +	__u32 flags;
>> +	__u64 dev_reset_count;
>> +	__u64 ctx_reset_count;
>> +};
>> +
>>   #if defined(__cplusplus)
>>   }
>>   #endif
> 
> Thanks,
> pq