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