sigchain: recover from concurrent double-pops of the same signal

From: ian4hu <[email protected]>

sigchain_push()/sigchain_pop() update signals[sig].n and the handler stack non-atomically. When two copies of the same signal are delivered in quick succession to different threads, both instances of the handler run the usual "cleanup; sigchain_pop(sig); raise(sig)" pattern concurrently, and both read the same s->n before either one decrements it. Each then installs s->old[s->n - 1], so a two-entry chain unwinds both entries while the dispatcher keeps pointing at a handler that was already popped: n ends at 0 while the effective disposition is still old[1].

The surviving dispatcher then runs that popped handler, whose own sigchain_pop() hits the "n < 1" guard and does nothing, and the raise() at the end redelivers to the same stuck dispatcher forever. The process neither dies nor continues; every signal-handling thread spins at 100% CPU and only SIGKILL recovers it.

This is reachable in practice: "git clone --depth 1" over ssh holds two SIGTERM entries in the chain at the index-pack parking point (clone's remove_junk_on_signal, then tempfile's remove_tempfiles_on_signal from the first lockfile), and two SIGTERMs a fraction of a millisecond apart land on two threads and corrupt the chain.

Rather than serializing the chain mutation, teach the "n < 1" guard in sigchain_pop() to repair the damage. Track how many pushes were ever performed for the signal; when a pop finds the chain empty but the chain bottom is known, reinstall s->old[0], the disposition that was in place before the chain was built. The stuck handler's raise() then lands on that disposition -- typically SIG_DFL -- and the process dies by it, exactly as a single delivery would have.

The repair is inert on every legitimate path to the guard: if the signal was never pushed, nothing is installed; if the chain was fully unwound by regular pops, s->old[0] is already installed and reinstalling it changes nothing. The push/pop fast path is untouched.

A fault-injected A/B run of the minimal reproducer, compiled from identical sources with and without this change (the gap between the "read n" and "decrement n" steps of pop widened so both racing pops reliably read the same value; 300 interleaved pairs per arm) froze the unpatched arm 121 times out of 300. The patched arm never froze: all 300 runs reinstated the pre-chain disposition and died by it, and 20/20 single-delivery control runs never reached the repair code.

---
sigchain: recover from concurrent double-pops of the same signal

During git clone --depth 1 over ssh (or any transport that parks the client with an in-flight packfile), two SIGTERMs delivered a fraction of a millisecond apart can corrupt the sigchain unwind: both fetch threads end up looping inside remove_junk_on_signal() at ~200% CPU, the process survives both signals, and further SIGTERMs have no effect (only SIGKILL recovers it).

Mechanism (git 2.56.0):

1. clone pushes two SIGTERM handlers: remove_junk_on_signal (builtin/clone.c:1134), then on the first tempfile activation (e.g. shallow.lock) remove_tempfiles_on_signal (tempfile.c:117):
old[0]=SIG_DFL, old[1]=remove_junk_on_signal, n=2.
2. Two deliveries that land on two threads run the top handler concurrently; both sigchain_pop() calls read the same s->n and both install s->old[s->n - 1]: the chain unwinds two levels for two deliveries, ending at n=0 while the dispatcher still points at old[1].
3. The surviving dispatcher then chains to a handler that was already popped, whose own sigchain_pop() hits the n < 1 guard (a silent no-op), and whose raise() redelivers to the same stuck dispatcher forever.

This patch takes the "repair after the fact" route instead of serializing the chain mutation: sigchain_pop() remembers how many pushes were ever performed for the signal, and when it finds the chain empty while the chain bottom is known, it reinstalls s->old[0] -- the disposition that was in place before the chain was built. The stuck handler's raise() then lands on that disposition (typically SIG_DFL) and the process dies by it, exactly as a single delivery would have.

The repair is inert on every legitimate path to the n < 1 guard: if the signal was never pushed nothing is installed, and if the chain was fully unwound by regular pops, s->old[0] is already installed so reinstalling it changes nothing. The push/pop fast path is untouched -- no locking, no signal-mask manipulation, no Windows-specific code.

Validation: a minimal reproducer (two threads parked in read(), the same push/pop chain, first handler entry doing real close()+unlink() work), compiled from identical sources with and without this change. With a fault injector widening the gap between the "read n" and "decrement n" steps of pop so both racing pops reliably read the same value (300 interleaved pairs per arm), the unpatched arm froze 121/300 runs while the patched arm never froze: all 300 runs reinstated the pre-chain disposition and died by it. 20/20 single-delivery control runs never reached the repair code.

An alternative fix that serializes sigchain_push()/sigchain_pop() with a per-signal spinlock exists as well; both eliminate the freeze. The lock prevents the corruption, this patch heals it; they are independent and either one suffices.

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-2263%2Fian4hu%2Fsigchain-recover-from-concurrent-pops-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-2263/ian4hu/sigchain-recover-from-concurrent-pops-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/2263

 sigchain.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)

index 66123bdbab..f6da68ce4d 100644
@@ -7,6 +7,7 @@ struct sigchain_signal {
 	sigchain_fun *old;
 	int n;
 	int alloc;
+	int pushed;
 };
 static struct sigchain_signal signals[SIGCHAIN_MAX_SIGNALS];
 

@@ -26,6 +27,7 @@ int sigchain_push(int sig, sigchain_fun f)
 	if (s->old[s->n] == SIG_ERR)
 		return -1;
 	s->n++;
+	s->pushed++;
 	return 0;
 }
 

@@ -33,8 +35,21 @@ int sigchain_pop(int sig)
 {
 	struct sigchain_signal *s = signals + sig;
 	check_signum(sig);
-	if (s->n < 1)
+	if (s->n < 1) {
+		/*
+		 * We are unwinding a chain that is already empty.  If we
+		 * got here from a signal handler, two concurrent pops of
+		 * the same signal both decremented "n", leaving the
+		 * dispatcher stuck on a handler that no longer has a
+		 * chain entry: its raise() would re-enter it forever.
+		 * Reinstall the disposition that was in place before the
+		 * chain was built so the raise() that follows lands on
+		 * that instead.
+		 */
+		if (s->pushed)
+			signal(sig, s->old[0]);
 		return 0;
+	}
 
 	if (signal(sig, s->old[s->n - 1]) == SIG_ERR)
 		return -1;

base-commit: f57c9a73884400380bddd172acc88b569834f909

gitgitgadget

Hi,

Gentle ping on this one. It is queued in 'seen' as iy/sigchain-recover-concurrent-pops, and the latest "What's cooking" marks it as "Needs review".

A short recap: when two copies of the same signal reach two threads nearly at the same time, the concurrent sigchain_pop() calls in the handlers can unwind the chain by two levels instead of one, leaving the process spinning in a handler that keeps re-raising the signal. This is reachable with "git clone --depth 1" over ssh receiving two SIGTERMs in quick succession, and only SIGKILL gets rid of the process.

The patch leaves the push/pop fast path alone and only teaches the "n < 1" guard in sigchain_pop() to fall back to the pre-chain disposition, so the stuck handler's raise() terminates the process just as a single delivery would.

Peff, I have Cc'ed you since you wrote the sigchain code. I would appreciate your thoughts on whether this "repair after the fact" approach is acceptable, or whether you would rather see the chain mutation serialized instead. I also have a version that uses a per-signal spinlock around push/pop, and can send it if that is preferred.

Thanks,
ian4hu

It sounds to me like the bug is that we're unwinding the same chain of signal handlers in two different threads at the same time. Doesn't that mean we're potentially calling the same handler twice, once from each thread, which doesn't sound like it'll lead to anything good. I wonder how hard it would be to fix the code to ensure that signals are blocked when we spawn a new thread so that they are only handled in a single thread. It would probably be good idea to start using sigaction() as well so that we don't interrupt a signal handler if we receive another instance of the same signal while the handler is running.

Thanks

Phillip

Hi Phillip,

Thanks for the review. I agree with both halves of your mail, and I think the difference between the two approaches is worth spelling out.

What I sent only repairs the aftermath. It leaves the corrupted state reachable and merely turns the wedge into the normal death, so it is small and inert on the legitimate paths, but it is a band-aid and, frankly, a bit dirty: it heals a state that should never have been entered, and by the time it heals it the two threads have already run the handler's close()/unlink() cleanup concurrently.

I did first try the serializing route. My earlier attempt
(https://github.com/gitgitgadget/git/pull/2262, never sent to the list) took a per-signal spinlock around the push/pop update, with the signal blocked while the lock is held. That keeps signals[sig].n and the installed disposition consistent under any interleaving, so the corrupt state cannot be constructed at all, and it also protects sigchain_push/pop called from worker threads. I dropped it because of the edges:

- two threads can still run the handler body concurrently; the lock only makes the unwind atomic, it does not serialize the cleanup itself;
- the POSIX corner is awkward: sigprocmask() is unspecified in a multithreaded process, pthread_sigmask() is not async-signal-safe, and taking the lock in a handler needs one of them;
- a thread that dies inside the critical section leaves the lock held and the next delivery spins forever, trading a hang for a hang;
- the Windows branch needs its own InterlockedExchange path that cannot be exercised in CI here.

So the question I would like your guidance on: do you consider a locked/serialized chain a complete fix for what you described, or does this have to be prevented at the thread level - block the signals when a thread is spawned, and use sigaction() for the chain? My own reading is that serializing is a complete fix for the chain state but not for the handler bodies, so if the latter is what you want, your suggestion is the way to go and the lock does not buy much beyond it.

If you prefer the thread-level fix, I am happy to build it: a git_pthread_create() wrapper that blocks SIGINT/SIGHUP/SIGTERM/ SIGQUIT around pthread_create() - leaving SIGPIPE alone, since it is a synchronous signal and blocking it in workers turns deaths into EPIPE - plus a sigaction()-based chain that masks all signals while a handler runs. I would then drop this patch, or keep it as a separate last-resort series if you think the belt and braces are worth it.

Thanks,
ian4hu

I think that is probably the safest solution (though it does not help for threads created external library code - I'm not sure if we have any instances of that). I'd be interested to hear what others think, so I'd wait a couple of days to see if anyone else comments before spending much time on it.

Does it? I thought that happened when its disposition is set to SIG_IGN. Blocking a signal delays its delivery until it is unblocked, it does not ignore the signal.

When I wrote my last email I was thinking about blocking the signal that is being handled as sigaction() defers the signal that is being handled unless SA_NODEFER is passed. However, sigchain_push_common() sets the same handler for all the common signals so it is probably safer to block at least those as well as I'm not sure all our handlers are reentrant.

I'd be inclined to drop it - if we're handling signals in a single thread that should avoid the issue.

Thanks

Phillip

A couple more thoughts ...

We have a number of places where we spin up a bunch of threads to parallelize some work, for example in preload_index() we create some threads in a loop to read the index. In cases like that, where the loop body is just setting up some per-thread data and creating a thread, it would be more efficient to block signals before the loop and unblock them after, rather than blocking and unblocking each time we create a thread.

To clarify, in a multi threaded process the delivery of the signal is only delayed if it is blocked in all threads. If there's one or more threads where the signal isn't blocked it will be delivered to one of those threads.

Thanks

Phillip