On 1/20/20 4:51 PM, Dmitry V. Levin wrote:
> On Wed, Jan 15, 2020 at 09:53:27AM -0700, Jens Axboe wrote:
>> On 1/15/20 9:50 AM, Eugene Syromiatnikov wrote:
>>> On Wed, Jan 15, 2020 at 09:41:58AM -0700, Jens Axboe wrote:
>>>> On 1/15/20 9:35 AM, Eugene Syromiatnikov wrote:
>>>>> fds field of struct io_uring_files_update is problematic with regards
>>>>> to compat user space, as pointer size is different in 32-bit, 32-on-64-bit,
>>>>> and 64-bit user space. In order to avoid custom handling of compat in
>>>>> the syscall implementation, make fds __u64 and use u64_to_user_ptr in
>>>>> order to retrieve it. Also, align the field naturally and check that
>>>>> no garbage is passed there.
>>>>
>>>> Good point, it's an s32 pointer so won't align nicely. But how about
>>>> just having it be:
>>>>
>>>> struct io_uring_files_update {
>>>> __u32 offset;
>>>> __u32 resv;
>>>> __s32 *fds;
>>>> };
>>>>
>>>> which should align nicely on both 32 and 64-bit?
>>>
>>> The issue is that 32-bit user space would pass a 12-byte structure with
>>> a 4-byte pointer in it to the 64-bit kernel, that, in turn, would treat it
>>> as a 8-byte value (which might sometimes work on little-endian architectures,
>>> if there are happen to be zeroes after the pointer, but will be always broken
>>> on big-endian ones). __u64 is used in order to avoid special compat wrapper;
>>> see, for example, __u64 usage in btrfs or BPF for similar purposes.
>>
>> Ah yes, I'm an idiot, apparently not enough coffee yet. We'd need it in
>> a union for this to work. I'll just go with yours, it'll work just fine.
>> I will fold it in, I need to make some updates and rebase anyway.
>
> I see the patch has missed v5.5-rc7.
> Jens, please make sure a fix is merged before v5.5 is out.
Ah shoot, I actually thought I added it for 5.6 only, but you are right,
it's in 5.5-rc as well. I'll ship a patch this week for 5.5.
--
Jens Axboe
