[RFC] y2038: HCI_TIME_STAMP with time64 - Y2038

10 Jan 2020


      I noticed earlier this week that the HCI_CMSG_TSTAMP/HCI_TIME_STAMP
interface has no time64 equivalent, as we apparently missed that when
converting the normal socket timestamps to support both time32 and time64
variants of the sockopt and cmsg data.
The interface was originally added back in 2002 by Maksim Krasnyanskiy
when bluetooth support first became non-experimental.
When using HCI_TIME_STAMP on a 32-bit system with a time64
libc, users will interpret the { s32 tv_sec; s32 tv_usec } layout of
the kernel as { s64 tv_sec; ... }, which puts complete garbage
into the timestamp regardless of whether this code runs before or
after y2038. From looking at codesearch.debian.org, I found two
users of this: libpcap and hcidump. There are probably others that
are not part of Debian.
Fixing this the same was as normal socket timestamps is not possible
because include/net/bluetooth/hci.h is not an exported UAPI header.
This means any changes to it for defining HCI_TIME_STAMP conditionally
would be ignored by applications that use a different copy of the
header.
I can see three possible ways forward:
1. move include/net/bluetooth/hci.h to include/uapi/, add a conditional
   definition of HCI_TIME_STAMP and make the kernel code support
   both formats. Then change applications to rely on that version of
   header file to get the correct definition but not change application code.
2. Leave the kernel completely unchanged and modify only the users
    to not expect the output to be a 'struct timeval' but interpret as
    as { uint32_t tv_sec; int32_t tv_usec; } structure on 32-bit architectures,
    which will work until the unsigned time overflows 86 years from now
    in 2106 (same as the libpcap on-disk format).
3. Add support for the normal SO_TIMESTAMPNS_NEW sockopt in
   HCI, providing timestamps in the unambiguous { long long tv_sec;
   long long tv_nsec; } format to user space, and change applications
   to use that if supported by the kernel.
Arnd