[systemd-devel] [PATCH] Fix broken syscall(__NR_fanotify_mark... on 32bit mips.

[David Daney]

On 04/20/2011 11:34 AM, Lennart Poettering wrote:
> On Wed, 20.04.11 11:19, David Daney (ddaney at caviumnetworks.com) wrote:
>
>>
>> On 04/20/2011 11:09 AM, Lennart Poettering wrote:
>>> On Wed, 20.04.11 10:36, David Daney (ddaney at caviumnetworks.com) wrote:
>>>
>>>> You would have to do something like this (untested):
>>>>
>>>> int foo_fanotify_mark(int fanotify_fd, unsigned int flags, u64 mask,
>>>> int dfd, const char  __user * pathname)
>>>> {
>>>> 	u32 mask_low = (u32)mask;
>>>> 	u32 mask_high = (u32)(mask>>   32);
>>>>
>>>> 	return syscall(4337, fanotify_fd, flags, mask_low, mask_high, dfd,
>>>> pathname);
>>>>
>>>> }
>>>>
>>>> The order of mask_low, mask_high in the syscall argument list
>>>> depends on the endianness.  Figuring out the correct order is left
>>>> as an exercise for the reader.
>>>
>>> That's basically the same patch as this one, right?
>>>
>>> http://lists.freedesktop.org/archives/systemd-devel/attachments/20110420/be2d393b/attachment.obj
>>>
>>
>> It looks like it might do the same thing.  Someone should try it on
>> an o32 MIPS userland running on a mips64 kernel.
>>
>> I think the behavior of the union thing is undefined, but should
>> work on GCC.
>
> Hmm, but unions is the official C99 way to do these things, isn't it?
> Instead of doing casts here and there which create aliasing probs?
>
> Any comment whether this will break non-MIPS 32bit archs, like x86?

It would break the MIPS n32 ABI userspace.

On MIPS n32 we are still __LP64__, but 64-bit values are passed in a 
single register.

I expect that the experimental x86_64 x32 ABI would suffer the same way.

So the change would have to be gated by ABI rather than __LP64__

David Daney