Re: [Linaro-mm-sig] [PATCHv2 3/4] staging: android: ion: Add an ioctl for ABI checking

[Laura Abbott <labbott@redhat.com>]

On 09/02/2016 02:33 PM, Arnd Bergmann wrote:
> On Friday, September 2, 2016 1:33:44 PM CEST Laura Abbott wrote:
>> On 09/02/2016 02:02 AM, Arnd Bergmann wrote:
>>> On Thursday, September 1, 2016 3:40:43 PM CEST Laura Abbott wrote:
>>>
>>>> --- a/drivers/staging/android/ion/ion-ioctl.c
>>>> +++ b/drivers/staging/android/ion/ion-ioctl.c
>>>> @@ -22,6 +22,29 @@
>>>>  #include "ion_priv.h"
>>>>  #include "compat_ion.h"
>>>>
>>>> +union ion_ioctl_arg {
>>>> +	struct ion_fd_data fd;
>>>> +	struct ion_allocation_data allocation;
>>>> +	struct ion_handle_data handle;
>>>> +	struct ion_custom_data custom;
>>>> +	struct ion_abi_version abi_version;
>>>> +};
>>>
>>> Are you introducing this, or just clarifying the defintion of the
>>> existing interface. For new interfaces, we should not have a union
>>> as an ioctl argument. Instead each ioctl command should have one
>>> specific structure (or better a scalar argument).
>>>
>>
>> This was just a structure inside ion_ioctl. I pulled it out for
>> the validate function. It's not an actual argument to any ioctl from
>> userspace. ion_ioctl copies using _IOC_SIZE.
>
> Ok, got it. This is fine from an interface point of view, just
> a bit unusual in the way it's written.
>
>>>> +static int validate_ioctl_arg(unsigned int cmd, union ion_ioctl_arg *arg)
>>>> +{
>>>> +	int ret = 0;
>>>> +
>>>> +	switch (cmd) {
>>>> +	case ION_IOC_ABI_VERSION:
>>>> +		ret = arg->abi_version.reserved != 0;
>>>> +		break;
>>>> +	default:
>>>> +		break;
>>>> +	}
>>>> +
>>>> +	return ret ? -EINVAL : 0;
>>>> +}
>>>
>>> I agree with Greg, ioctl interfaces should normally not be versioned,
>>> the usual way is to try a command and see if it fails or not.
>>>
>>
>> The concern was trying ioctls that wouldn't actually fail or would
>> have some other unexpected side effect.
>>
>> My conclusion from the other thread was that assuming we don't botch
>> up adding new ioctls in the future or make incompatible changes to
>> these in the future we shouldn't technically need it. I was still
>> trying to hedge my bets against the future but that might just be
>> making the problem worse?
>
> We've had a number of cases where versioned ABIs just didn't work out.
>
> The versions are either used to distinguish incompatible APIs, which
> we should avoid to start with, or they are used for backwards-compatible
> extensions that you should detect by checking whether an ioctl
> succeeds. Relying on the API version number breaks if you get a partial
> backport of features from a later version, and it's unclear what a
> user space tool should expect when the kernel reports a newer ABI
> than it knows.
>
> I think the wireless extensions and KVM are examples of versioned
> APIs that turned out to make things more complicated than they
> would have been otherwise.
>

Okay it sounds like the answer is to strive to never run into a case
where versioned ioctls are necessary. Shouldn't be too hard, right? ;)

>>>> +/**
>>>> + * struct ion_abi_version
>>>> + *
>>>> + *  @version - current ABI version
>>>> + */
>>>> +
>>>> +#define ION_ABI_VERSION                KERNEL_VERSION(0, 1, 0)
>>>> +
>>>> +struct ion_abi_version {
>>>> +	__u32 abi_version;
>>>> +	__u32 reserved;
>>>> +};
>>>> +
>>>
>>> This interface doesn't really need a "reserved" field, you could
>>> as well use a __u32 by itself. If you ever need a second field,
>>> just add a new command number.
>>>
>>
>> The botching-ioctls.txt document suggested everything should be aligned
>> to 64-bits. Was I interpreting that too literally?
>
> I didn't even know that file existed ;-)
>
> I'm pretty sure the paragraph refers to the problem of x86 of having
> a structure like
>
> 	struct ioctl_arg {
> 		__u64 first;
> 		__u32 second;
> 	};
>
> which is 12 bytes long on x86, but 16 bytes long including implied
> padding on all 64-bit architectures and most (maybe all) 32-bit ones
> other than x86.
>

Right, that's the problem it's trying to avoid.

> If there is no 64-bit member in the struct, there is no need for padding
> at the end.
>

That's what I thought as well. I think I'll submit a patch to the docs
clarifying a few things.


> 	Arnd
>

Thanks,
Laura