On Sun, 11 Feb 2007, Heiko Carstens wrote:

quoted text
> On Sun, Feb 11, 2007 at 12:15:24PM -0800, Davide Libenzi wrote:
> > 
> > Add epoll compat_ code to kernel/compat.c. IA64 and ARM-OABI are currently 
> > using their own version of epoll compat_ code and they could probably wire 
> > to the new common code. Patch over 2.6.20.
> > + * epoll (fs/eventpoll.c) compat bits follow ...
> > + */
> > +struct compat_epoll_event {
> > +	u32 events;
> > +	u32 data[2];
> > +};
> > +
> >[...]
> > +
> > + * We need the compat layer over the epoll_event structure, only if the offset
> > + * of the __u64 data member is not 4 (size of the events member that precedes the
> > + * data one).
> > + */
> > +#define EPOLL_NEED_EVENT_COMPAT() (offsetof(struct epoll_event, data) != 4)
> 
> With
> 
> struct epoll_event {
>         __u32 events;
>         __u64 data;
> };
> 
> this won't work on s390. offsetof(struct epoll_event, data) is 8 on both
> 31 bit and 64 bit. So it will do the conversion and corrupt all the data.
> Actually we would only need the compat conversion for the sigset_t stuff.

Yup, that's broken not only on s390, but on every arch with alignof(u64) == 8
in 32 bits mode.
The assumption was that for cases like the above, you simply wouldn't wire 
the compat_ version. That is true for epoll_wait and epoll_ctl, where the 
only need for compat was the "struct epoll_event". But that's not true for 
epoll_pwait, since this one needs to be wired because of the sigset_t.
On top of sigset_t, epoll_pwait may need "struct epoll_event" translation.
Now, that *really* sux because two versions of compat_epoll_pwait are 
needed, once that does sigset_t translation only, and one that does 
sigset_t + "struct epoll_event".



quoted text
> But then again I thought most 32 bit architectures would add a 4 byte
> pad between events and data, no?

i386 does not, for example ;)



quoted text
> Maybe we need some arch dependent struct compat_epoll_event and have
> something like
> #define EPOLL_NEED_EVENT_COMPAT() \
> (offsetof(struct epoll_event, data) != offsetof(struct compat_epoll_event, data))
> 
> ?

No, it won't work. Unless there is (or we define) a per-arch macro that 
tells us how the 32 bits mode align an u64, I'm afraid we can't do any 
smart tricks and we need to have the double compat_epoll_pwait.



- Davide


-
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/