scsi: bfa: use 64-bit times in bfa_aen_entry_s ABI

bfa_aen_entry_s is passed through a netlink socket that can be read by
either 32-bit or 64-bit processes, but the data format is different
between the two on current implementations.

Originally, this was using a 'struct timeval', which also suffers from
getting redefined with a new libc implementation.

With this patch, the layout gets fixed to having two 64-bit members for
the time, making it the same on 32-bit kernels and 64-bit kernels
running either compat or native user space including x32.

Provided that the new header file gets used to recompile any 32-bit
application binaries, this will fix running those on a 64-bit kernel
(with or without this patch) e.g. in a container environment, and it
will make binaries work that will be built against a future 32-bit glibc
that uses a 64-bit time_t, and avoid the y2038 overflow there.

However, this also breaks compatibility with any existing 32-bit binary
running on a native 32-bit kernel, those must be recompiled against the
new header, which in turn makes them incompatible with older kernels
unless the same change gets applied there.

Obviously this patch should only be applied when the benefits outweigh
the possible breakage. I'm posting it under the assumption that there
are no open-source tools using the netlink interface, and that users of
the binaries provided by qlogic for SLES10/11 and RHEL5/6 are not
actually being used on new future systems with 32-bit x86 kernels.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Anil Gurumurthy <Anil.Gurumurthy@cavium.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

drivers/scsi/bfa/bfa_defs_svc.h

@@ -1455,8 +1455,8 @@ struct bfa_aen_entry_s {
 	enum bfa_aen_category   aen_category;
 	u32                     aen_type;
 	union bfa_aen_data_u    aen_data;
-	unsigned long		aen_tv_sec;
-	unsigned long		aen_tv_usec;
+	u64			aen_tv_sec;
+	u64			aen_tv_usec;
 	u32                     seq_num;
 	u32                     bfad_num;
 };