input_event size mismatch between 64bit kernel and 32bit userland ...

input_event size mismatch between 64bit kernel and 32bit userland ...

[Jun Sun]

Pavel Kiryukhin discovered this problem while he is developing
64bit MIPS kernel.  With his permission I am forwarding this question
to lkml and we would be interested to know how x86_64 solves this
problem.

Thanks.

Jun

----------------------------------------------------------------------
Hi all,
I stuck in simple situation:
USB mouse (or keyboard). n64 kernel (2.4.20), n32 userland.

Userspace application tries to read "input_event" (16 bytes) from  
"/dev/input/event0" [ read(fd,&key_ev, sizeof(key_ev)) ],
input core driver treats "input_event" as 24 bytes structure. It is due
to different size of  "timeval" (and finally  "long") in n64 kernel and
n32 userland.

Application gets some garbage as mouse events . No solutions like "ioctl
wrappers" applicable in this case.

I don't want to change any arch independent files, but can not find any
acceptable solution. It looks like headers "/usr/include/linux/input.h"
in root file system and "/include/linux/input.h" in kernel should be the
same,
 (All works fine as soon as I declare a new "input_event" structure in
user application that corresponds in size to kernel
structure - but this is not acceptable).

Can anybody advice me what to do with the difference in "input_event"
structure sizes in o32/n32 userland and n64 kernel? Just a general
approach that can be used when  driver's read/write operation treat some
values as 64 bit while user application tries to read/write 32-bit
values (based on the same headers).

Please, don't kick me if solution is simple and obvious.

Re: input_event size mismatch between 64bit kernel and 32bit userland ...

[David S. Miller]

On Thu, 3 Jun 2004 18:07:28 -0700
Jun Sun <jsun@mvista.com> wrote:

> Pavel Kiryukhin discovered this problem while he is developing
> 64bit MIPS kernel.  With his permission I am forwarding this question
> to lkml and we would be interested to know how x86_64 solves this
> problem.

x86_64 doesn't solve this problem, it's broken there too

Some CONFIG_COMPAT code needs to be added to the input layer
I guess