diff options
author | Zack Weinberg <zackw@panix.com> | 2018-03-23 09:16:59 -0400 |
---|---|---|
committer | Zack Weinberg <zackw@panix.com> | 2018-03-26 08:30:46 -0400 |
commit | 9ea49e16c79bd2acd0d0648ca0163f26dd1c3dae (patch) | |
tree | bc5d3c3735dd1ee96b43f1dee3001dfb165d1fd8 /sysdeps/mach/hurd/fdatasync.c | |
parent | 3d8eb8099425ae4f474e97082e04784c2984ec48 (diff) | |
download | glibc-zack/wip-pthread-no-dupe-defns.tar.gz |
[Bug 15368] Move pthread_kill to libc and use it to implement raise.zack/wip-pthread-no-dupe-defns
The fix for bug #15368 was unnecessarily Linux-specific. To recap,
POSIX specifies raise to be async-signal-safe, but also specifies it
to be equivalent to pthread_kill(pthread_self(), sig), which is not
an async-signal-safe sequence of operations; a signal handler could
run in between pthread_self and pthread_kill, and do something (such
as calling fork, which is also async-signal-safe) that would invalidate
the thread descriptor. This is even true in the hypothetical case of
a port that doesn't implement multithreading: kill(getpid(), sig) will
fire the signal twice if a signal handler runs in between, calls fork,
and then returns on both sides of the fork. I don't see anything in
the standards to forbid that.
The Linux-specific fix was to override the definitions of raise in
both libpthread and libc to the same unitary function that blocks
signals, retrieves TID and PID directly from the kernel, calls tgkill,
and only then unblocks signals. This patch generalizes that to any
port: pthread_kill is moved from libpthread to libc, with a forwarding
stub left behind. The definition of raise in libpthread is also
replaced with a forwarding stub. The Linux-specific definition of
raise is deleted; those ports will now use sysdeps/pthread/raise.c,
which blocks signals first, then calls pthread_self and pthread_kill,
and then unblocks signals. Similarly, sysdeps/posix/raise.c (which
would be used on a port that didn't implement multithreading) blocks
signals, calls getpid and kill, and then unblocks signals. Thus,
ports need only implement the primitives correctly and do not need to
worry about making raise async-signal-safe.
The only wrinkle was that up till now, we did not bother initializing
the ->tid field of the initial thread's descriptor unless libpthread
was loaded; now that raise calls pthread_kill even in a single-
threaded environment, that won't fly. This is abstractly easy to fix;
the tricky part was figuring out _where_ to put the calls (two of
them, as it happens) to __pthread_initialize_pids, and I'd appreciate
careful eyes on those changes.
You might be wondering why it's safe to rely on the TID in the thread
descriptor, rather than calling gettid directly. Since all signals
are blocked from before calling pthread_self until after pthread_kill
uses the TID to call tgkill, the question is whether some _other_
thread could do something that would invalidate the calling thread's
descriptor, and I believe there is no such thing.
While I was at it I fixed another bug: raise was returning an error
code on failure (like pthread_kill does) instead of setting errno as
specified. This is user-visible but I don't think it's worth recording
as a fixed bug, nobody bothers checking whether raise failed anyway.
* nptl/pt-raise.c
* sysdeps/unix/sysv/linux/pt-raise.c
* sysdeps/unix/sysv/linux/raise.c:
Remove file.
* sysdeps/unix/sysv/linux/pthread_kill.c: Use __is_internal_signal
to check for forbidden signals. Use INTERNAL_SYSCALL_CALL to call
getpid. Provide __libc_pthread_kill, with __pthread_kill as
strong alias and pthread_kill as weak alias.
* sysdeps/posix/raise.c: Block signals around the calls to
__getpid and __kill. Provide __libc_raise, with raise as strong
alias, libc_hidden_def for raise, and gsignal as weak alias.
* sysdeps/pthread/raise.c: New file. Implement by blocking
signals, calling pthread_self and pthread_kill, and then
unblocking signals again. Provide same symbols as above.
* sysdeps/generic/internal-signals.h: Define all of the same
functions that sysdeps/unix/sysv/linux/internal-signals.h does,
with sensible default definitions.
* sysdeps/unix/sysv/linux/internal-signals.h: Clarify comments.
* nptl/pthread_kill.c: Define __libc_pthread_kill, with
__pthread_kill as strong alias and pthread_kill as weak alias.
* nptl/pthread_self.c: Define __pthread_self, with
pthread_self as weak alias.
* signal/raise.c: Define __libc_raise, with raise as strong alias,
libc_hidden_def for raise, and gsignal as weak alias.
* nptl/Makefile: Move pthread_kill from libpthread-routines to
routines. Remove pt-raise from libpthread-routines.
* nptl/Versions (libc/GLIBC_2.28): Add pthread_kill.
(libc/GLIBC_PRIVATE): Add __libc_pthread_kill and __libc_raise.
* sysdeps/generic/pt-compat-stubs.S: Add stubs for raise and
pthread_kill.
* nptl/nptl-init.c (__pthread_initialize_minimal_internal):
Don't call __pthread_initialize_pids here.
* csu/libc-tls.c (__libc_setup_tls):
Call __pthread_initialize_pids after all other setup.
* elf/rtld.c (init_tls): Likewise.
* include/pthreadP.h: New forwarder.
* include/pthread.h: Add multiple inclusion guard. Declare
__pthread_self.
* include/signal.h: Declare __pthread_kill.
* sysdeps/**/libc.abilist (GLIBC_2.28): Add pthread_kill.
Diffstat (limited to 'sysdeps/mach/hurd/fdatasync.c')
0 files changed, 0 insertions, 0 deletions