proposed-migration for faketime 0.9.10-2.1ubuntu1

Unsurprisingly, there are some symbol differences.

# diff -u <(objdump -T /usr/lib/arm-linux-gnueabihf/faketime/libfaketime.so.1 | grep '\.text' | cut -b34- | sort -u) <(objdump -T ./src/libfaketime.so.1 | grep '\.text' | cut -b34- | sort -u)@@ -1,50 +1,51 @@
 Base __clock_gettime
+Base __clock_gettime64
+Base __clock_nanosleep_time64
+Base __fstat64_time64
 Base __ftime
+Base __futimens64
 Base __fxstat
 Base __fxstat64
 Base __fxstatat
 Base __fxstatat64
 Base __gettimeofday
+Base __lstat64_time64
 Base __lxstat
 Base __lxstat64
+Base __nanosleep64
+Base __ppoll64
+Base __pselect64
+Base __select64
+Base __sem_timedwait64
+Base __stat64_time64
 Base __time
+Base __time64
+Base __timerfd_gettime64
+Base __timerfd_settime64
+Base __timespec_get64
+Base __utimensat64
 Base __xstat
 Base __xstat64
 Base alarm
-Base clock_gettime
-Base clock_nanosleep
 Base epoll_pwait
 Base epoll_wait
-Base fstat
 Base ftime
-Base futimens
 Base getentropy
 Base getpid
 Base getrandom
 Base gettimeofday
 Base lock_for_stat
-Base lstat
-Base nanosleep
 Base needs_forced_monotonic_fix
 Base poll
-Base ppoll
-Base pselect
 Base pthread_cond_destroy
 Base pthread_cond_init
 Base pthread_cond_timedwait
 Base read_config_file
-Base select
-Base sem_timedwait
 Base sleep
-Base stat
-Base time
-Base timerfd_gettime
-Base timerfd_settime
-Base timespec_get
+Base statx
 Base unlock_for_stat
 Base usleep
 Base utime
-Base utimensat
 Base utimes
 GLIBC_2.2) timer_gettime
 GLIBC_2.2) timer_settime
#

One or more of these functions may be broken under TIME_BITS=64. But we don't get far enough to figure that out?

The failure is because bash, for some reason, still links against gettimeofday rather than __gettimeofday64, and calls that to seed its internal random number generator. I still can't figure out why it's using the old version, though.

ubuntu@noble-armhf:~$ readelf -W -s a.out | grep timeof # a simple test program that just calls gettimeofday()
     9: 00000000 0 FUNC GLOBAL DEFAULT UND __gettimeofday64@GLIBC_2.34 (2)
    98: 00000000 0 FUNC GLOBAL DEFAULT UND __gettimeofday64@GLIBC_2.34
ubuntu@noble-armhf:~$ readelf -W -s /bin/bash | grep timeof
   105: 00000000 0 FUNC GLOBAL DEFAULT UND gettimeofday@GLIBC_2.4 (2)

faketime redefines gettimeofday, except that it uses the timeval definition from glibc which is of course using 64bit time_t. Mayhem ensues.

Oh, hang on. The bash build has apparently been uploaded just a day after the t64 gcc, which means gcc was presumably still building when the bash build started:

gcc-13 armhf 13.2.0-13ubuntu1 (from the bash build logs)

A bash rebuild should "fix" this somewhat. Well, at least a little bit.

What now remains to be done is to heavily patch faketime to, when on armhf:

1/ use the proper symbols from glibc (e.g. __clock_gettime64 instead of __clock_gettime)
2/ expose those symbols instead of the legacy 32-bit ones.

I added an sssd task due to the workaround we had to add to it in https://bugs.launchpad.net/ubuntu/+source/sssd/+bug/2058576 (we don't install faketime). Should faketime be fixed, then we can revert that change in sssd.

This bug was fixed in the package bash - 5.2.21-2ubuntu4

---------------
bash (5.2.21-2ubuntu4) noble; urgency=medium

  * No-change rebuild for CVE-2024-3094

 -- Steve Langasek <email address hidden> Sun, 31 Mar 2024 08:41:03 +0000