Re: [PATCH v3 06/11] staging/android: turn fence_info into a __u64 pointer

[Maarten Lankhorst <maarten.lankhorst@linux.intel.com>]

Op 03-02-16 om 21:09 schreef Gustavo Padovan:
> Hi Maarten,
>
> 2016-02-03 Maarten Lankhorst <maarten.lankhorst@linux.intel.com>:
>
>> Op 03-02-16 om 14:25 schreef Gustavo Padovan:
>>> From: Gustavo Padovan <gustavo.padovan@collabora.co.uk>
>>>
>>> Turn sync_fence_info into __u64 type enable us to extend the struct in the
>>> future without breaking the ABI.
>>>
>>> v2: use type __u64 for fence_info
>>>
>>> v3: fix commit message to reflect the v2 change
>>>
>>> Signed-off-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk>
>>> ---
>>>  drivers/staging/android/sync.c      | 2 +-
>>>  drivers/staging/android/uapi/sync.h | 2 +-
>>>  2 files changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/staging/android/sync.c b/drivers/staging/android/sync.c
>>> index 2ab0c20..8425457 100644
>>> --- a/drivers/staging/android/sync.c
>>> +++ b/drivers/staging/android/sync.c
>>> @@ -525,7 +525,7 @@ static long sync_file_ioctl_fence_info(struct sync_file *sync_file,
>>>  	if (info->status >= 0)
>>>  		info->status = !info->status;
>>>  
>>> -	len = sizeof(struct sync_file_info);
>>> +	len = sizeof(struct sync_file_info) - sizeof(__u64);
>>>  
>>>  	for (i = 0; i < sync_file->num_fences; ++i) {
>>>  		struct fence *fence = sync_file->cbs[i].fence;
>>> diff --git a/drivers/staging/android/uapi/sync.h b/drivers/staging/android/uapi/sync.h
>>> index a0cf357..e649953 100644
>>> --- a/drivers/staging/android/uapi/sync.h
>>> +++ b/drivers/staging/android/uapi/sync.h
>>> @@ -54,7 +54,7 @@ struct sync_file_info {
>>>  	char	name[32];
>>>  	__s32	status;
>>>  
>>> -	__u8	sync_fence_info[0];
>>> +	__u64	sync_fence_info;
>>>  };
>>>  
>>>  #define SYNC_IOC_MAGIC		'>'
>> This still doesn't do what you expect it to.
>>
>> I think this is what you want is for userspace to do:
>>
>> struct sync_file_info info;
>>
>> info.flags = info.num_fences = 0;
>> ioctl(fd, SYNC_IOC_FENCE_INFO, &info);
>> if (info.num_fences) {
>> info.sync_fence_info = (uintptr)kcalloc(info.num_fences, sizeof(struct sync_fence_info));
>> ioctl(fd, SYNC_IOC_FENCE_INFO, &info);
>> }
>>
>> Maybe userspace could preallocate the max in advance and set num_fences higher,
>>
>> kernel would do something like:
>>
>> num_fences = min(info.num_fences, sync->num_fences);
>> struct sync_fence_info array[num_fences];
>>
>> info.num_fences = sync->num_fences;
>> if (num_fences &&
>>     copy_to_user((void * __user)(unsigned long)info.sync_fence_info, array, num_fences  * sizeof(array)))
>>  return -EFAULT;
> If we are going to call IOCTL twice I would actually have a new IOCTL only
> to fetch sync_fence_info.
>
> First we would call
>
> ioctl(fd, SYNC_IOC_FILE_INFO, &info);
>
> where info is:
>
> struct sync_file_info {    
>         char    name[32];  
>         __s32   status;    
>         __u32   flags;     
>         __u32   num_fences;
> };
>
> then we would allocate a buffer with
>
> size = info.num_fences * sizeof(struct sync_fence_info)
>
> and call the new ioctl
>
> ioctl(fd, SYNC_IOC_SYNC_FENCE_INFO, sync_fence_info);
>
> This looks like a cleaner solution and doesn't break ABI. What do you
> think?
I think it's good taste that userspace specifies the size of the buffer it passes, so former feels more clean to me,
since you need to pass num_fences anyway.
But Daniel knows more about designing ioctl's than I do, so for exact behavior it's best to ask him.

~Maarten