| Summary: | Marking all functions which don't have callbacks with the leaf attribute breaks pthread applications. | ||
|---|---|---|---|
| Product: | glibc | Reporter: | Octoploid <cryptooctoploid> |
| Component: | nptl | Assignee: | Ulrich Drepper <drepper.fsp> |
| Status: | RESOLVED FIXED | ||
| Severity: | normal | CC: | hubicka, jakub, jim |
| Priority: | P2 | Flags: | fweimer:
security-
|
| Version: | 2.14 | ||
| Target Milestone: | --- | ||
| Host: | Target: | ||
| Build: | Last reconfirmed: | ||
I think the leaf attribute in gcc is currently primarily used by ipa-reference to find out what variables are supposed not to be read resp. written during the call to that function (with leaf attribute all static vars in the current translation unit are in that set if the leaf function is defined in some other compilation unit) and for checking whether labels can be reached by non-local goto. While the synchronization primitives don't call any callbacks, from this POV it is undesirable to treat them that way, they need to be considered as full barriers. I think the GCC documentation of leaf attribute should be improved and the pthread.h/sem.h functions that are supposed to work as memory synchronization points should use __THROWNL or similar macro which would be the old __THROW, without leaf attribute in it. There are other functions that shouldn't be marked leaf BTW, although they are __THROW - e.g. sprintf/snprintf/vsprintf/vsnprintf and their checking equivalents - it is possible to register a printf hook and the hook certainly can be defined in the current function. And the wide counterparts too. I'd say sscanf/vsscanf should be treated similarly just in case there will be scanf hooks implemented in the future. at least per POSIX, http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_11, here's the full list: The following functions synchronize memory with respect to other threads: fork pthread_barrier_wait pthread_cond_broadcast pthread_cond_signal pthread_cond_timedwait pthread_cond_wait pthread_create pthread_join pthread_mutex_lock pthread_mutex_timedlock pthread_mutex_trylock pthread_mutex_unlock pthread_spin_lock pthread_spin_trylock pthread_spin_unlock pthread_rwlock_rdlock pthread_rwlock_timedrdlock pthread_rwlock_timedwrlock pthread_rwlock_tryrdlock pthread_rwlock_trywrlock pthread_rwlock_unlock pthread_rwlock_wrlock sem_post sem_timedwait sem_trywait sem_wait semctl semop wait waitpid Fixed. Thanks. |
Commit aa78043a4aafe5 marks all functions, which don't have callbacks, with the leaf attribute. This may break pthreaded apps (for example "git grep"), because pthread_mutex_{lock,unlock,trylock,timedlock} and pthread_cond_{wait,timedwait} , etc. are also marked and therefore gcc happily optimizes them. See: https://bugzilla.redhat.com/show_bug.cgi?id=747377 or: http://thread.gmane.org/gmane.comp.version-control.git/184184 There are two possibilties to solve this issue; either don't mark the pthread related functions in glibc or teach gcc to unmark those functions.