Re: [PATCH v2 06/11] staging/android: turn fence_info into a __u64 pointer

From: Maarten Lankhorst
Date: Wed Feb 03 2016 - 03:32:30 EST


Op 02-02-16 om 21:28 schreef Gustavo Padovan:
> Hi Maarten,
>
> 2016-02-02 Maarten Lankhorst <maarten.lankhorst@xxxxxxxxxxxxxxx>:
>
>> Op 02-02-16 om 14:23 schreef Gustavo Padovan:
>>> From: Gustavo Padovan <gustavo.padovan@xxxxxxxxxxxxxxx>
>>>
>>> Making fence_info a pointer enables us to extend the struct in the future
>>> without breaking the ABI.
>>>
>>> v2: use type __u64 for fence_info
>>> Signed-off-by: Gustavo Padovan <gustavo.padovan@xxxxxxxxxxxxxxx>
>>> ---
>>> 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 f7530f0..03b1214 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 ed281fc..8e2ed32 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 fence_info[0];
>>> + __u64 fence_info;
>>>
>> With the information from the commit I would expect fence_info would be a true pointer.
> I missed that. It is fixed for v3 now.
>
>> so if (sync_file_info->num_fences == 0) not copying fence_info, but set all members including this one.
>> else for_all_fences() copy_to_user(fence_info struct)
>>
>> Otherwise you still can't extend the struct without breaking the abi.
> all sync_files have at least 1 fence so we are left to the else case,
> coupy all fences to the user. So we always have variable length at the
> end of the struct. But I still don't see how this will prevent us from
> breaking the API.
Ok this was with fence_info not being a pointer. :)