Public bug reported:
firebuild 0.8.6.1-3 fails to build in stonking-proposed on all architectures
that run the test suite (amd64, amd64v3, arm64, armhf, ppc64el, s390x; riscv64
only built because it uses nocheck):
https://launchpad.net/ubuntu/+source/firebuild/0.8.6.1-3
The integration test 25 "env fixup" hangs and the build is killed after
150 minutes of inactivity. 0.8.6.1-2 passed that test in stonking in
May, and Debian sid passes it with 0.8.6.1-3+b1.
Root cause: the Ubuntu fakeroot delta 2.1.4-1ubuntu1 (d/p/intercept-raw-
fchmodat2-syscalls.patch, LP: #1915924) adds a syscall() wrapper to
libfakeroot. When a program that uses jemalloc (such as firebuild) runs
under fakeroot, it deadlocks at startup.
Reproducer (stonking-proposed, inside the firebuild build tree):
cd obj-*/test
fakeroot ./run-firebuild -- /usr/bin/true # hangs forever
It does not matter which command firebuild runs: GNU id, true, echo and sh -c
true all hang.
Backtrace of the hung firebuild process:
#0-3 pthread_mutex_lock (jemalloc init mutex, already held)
#4-6 libjemalloc.so.2
#7 malloc
#8 _dl_exception_create_format (objname=libfakeroot-sysv.so, "undefined
symbol: %s%s%s")
#9 _dl_lookup_symbol_x (undef_name="acl_get_fd")
#11 _dl_sym / dlsym
#18 load_library_symbols () libfakeroot-sysv.so
#19 tmp_syscall () libfakeroot-sysv.so
#20 syscall () libfakeroot-sysv.so <- new Ubuntu wrapper
#21-23 libjemalloc.so.2 (init, holding its init mutex)
#24 libstdc++ static initializer -> malloc
#27 _dl_init
So jemalloc calls syscall() during initialisation while it holds its
init mutex. The new libfakeroot syscall() wrapper then lazily resolves
all of its real symbols. dlsym("acl_get_fd") fails, and glibc calls
malloc() to format the error message. That re-enters jemalloc and blocks
on the same non-recursive mutex.
This affects any program linked with jemalloc (and probably other
allocators that call syscall() during init) when it runs under fakeroot,
not just firebuild.
Possible fixes in fakeroot:
- In the syscall() wrapper, resolve only the real syscall with
dlsym(RTLD_NEXT, "syscall") instead of calling load_library_symbols(). That
lookup succeeds, so glibc does not allocate an error string.
- Or add a guard against re-entry and forward straight to the real syscall
while symbols are being loaded.
- Or drop the delta once uutils stops issuing raw fchmodat2 (LP: #2164536).
firebuild does not need a change. Its migration is blocked until
fakeroot is fixed.
** Affects: fakeroot (Ubuntu)
Importance: Undecided
Assignee: Nadzeya Hutsko (nadzeya)
Status: New
** Affects: firebuild (Ubuntu)
Importance: Undecided
Status: New
** Tags: ftbfs update-excuse
** Tags added: ftbfs update-excuse
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2169262
Title:
syscall() wrapper from intercept-raw-fchmodat2-syscalls.patch
deadlocks jemalloc-linked programs under fakeroot (firebuild FTBFS)
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/fakeroot/+bug/2169262/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
