Panicked during epoch::pin()

[siyuan0322]

I came across a weird panic recently, which seems is the guard_count: usize has overflowed during calling epoch::pin(). But I'm sure I haven't call epoch::pin() that many times.

I'm wondering if this ever occurred you guys, and could kindly provides some insights on this issue.

thread 'reactor 0' panicked at 'called `Option::unwrap()` on a `None` value', /home/graphscope/.cargo/registry/src/index.crates.io-6f17d22bba15001f/crossbeam-epoch-0.9.15/src/internal.rs:386:57

The relevant code context are

crossbeam/crossbeam-epoch/src/internal.rs

Lines 380 to 387 in 18afbb6

	#[inline]
	pub(crate) fn pin(&self) -> Guard {
	let guard = Guard { local: self };
	let guard_count = self.guard_count.get();
	self.guard_count.set(guard_count.checked_add(1).unwrap());
	if guard_count == 0 {

[taiki-e]

In what environment did you encounter this problem? Do you have an example to reproduce this problem?

I'm wondering if this ever occurred you guys,

I have never seen this problem, but assuming you are using a 64-bit system, it would take years to cause an overflow due to addition, so I guess it is more likely an overflow due to subtraction caused by an incorrect invocation of unpin. If so, it's either a bug on our code, a bug in the standard library or platform (especially around thread-local), or a bug in your code, but if you're not using unsafe code, it's one of the former two.

[taiki-e]

If you have encountered rust-lang/rust#47949 (rustc bug) in some way, I think it is possible to trigger this problem because the destructor of the guard in repin_after will not be called. I can't immediately think of an example that would trigger it, though.

[siyuan0322]

Thanks you very much, I'm trying to get a backtrace of the panic currently. I can't reproduce it with a small example, which makes it hard to debug.

[siyuan0322]

Hi, I want to thanks for your help in previous communications. After a clear inspect of the source code, I have another two question about the pin() and unpin() here,

• The guard_count is a usize that at least should be 0, but why there it checks for if guard_count == 0 after the self.guard_count.set(guard_count.checked_add(1).unwrap());

    pub(crate) fn pin(&self) -> Guard {
        let guard = Guard { local: self };

        let guard_count = self.guard_count.get();
        self.guard_count.set(guard_count.checked_add(1).unwrap());

        if guard_count == 0 {

• Why do not check the guard_count against 0 but 1, in the unpin(), I see it checked it against 0 in the finalize() afterward.

    pub(crate) fn unpin(&self) {
        let guard_count = self.guard_count.get();
        self.guard_count.set(guard_count - 1);

        if guard_count == 1 {
            self.epoch.store(Epoch::starting(), Ordering::Release);

            if self.handle_count.get() == 0 {
                self.finalize();
            }
        }
    }

[siyuan0322]

Now I managed to reproduce this but several times, sometimes it panicked in the overflow of checked_add(1) during pin(), sometimes it panicked in the overflow of subtract of guard_count - 1 during unpin(), which is in the drop() of the Guard.

😢 totally confused.

[longbinlai]

Hi, I want to thanks for your help in previous communications. After a clear inspect of the source code, I have another two question about the pin() and unpin() here,

• The guard_count is a usize that at least should be 0, but why there it checks for if guard_count == 0 after the self.guard_count.set(guard_count.checked_add(1).unwrap());

    pub(crate) fn pin(&self) -> Guard {
        let guard = Guard { local: self };

        let guard_count = self.guard_count.get();
        self.guard_count.set(guard_count.checked_add(1).unwrap());

        if guard_count == 0 {

• Why do not check the guard_count against 0 but 1, in the unpin(), I see it checked it against 0 in the finalize() afterward.

    pub(crate) fn unpin(&self) {
        let guard_count = self.guard_count.get();
        self.guard_count.set(guard_count - 1);

        if guard_count == 1 {
            self.epoch.store(Epoch::starting(), Ordering::Release);

            if self.handle_count.get() == 0 {
                self.finalize();
            }
        }
    }

@taiki-e Could you please help us with the above question? Thank you so much.

[taiki-e]

• The guard_count is a usize that at least should be 0, but why there it checks for if guard_count == 0 after the self.guard_count.set(guard_count.checked_add(1).unwrap());

• Why do not check the guard_count against 0 but 1, in the unpin(),

Getting a value, then adding or subtracting it, and then checking the old value is the same idiom as fetch_add or fetch_sub.

In the former case, guard_count == 0 means that there was no pinned one; in the latter, guard_count == 1 means that it was the last one (so check if it is needed to call the finalize).

I see it checked it against 0 in the finalize() afterward.

It checks handle_count, not guard_count. (If there is a live handle, finalize cannot be called.)

[taiki-e]

Now I managed to reproduce this

Is it possible to provide this reproduction?

If this is difficult, at least please provide us information about your build environment, build configuration, other dependencies, and any unsafe code you may have.

TBH I feel it is impossible to help you with just the information currently provided.

[longbinlai]

Now I managed to reproduce this

Is it possible to provide this reproduction?

If this is difficult, at least please provide us information about your build environment, build configuration, other dependencies, and any unsafe code you may have.

TBH I feel it is impossible to help you with just the information currently provided.

Indeed, we are diligently analyzing our code but have yet to uncover significant findings. Concurrently, we're delving into the intricacies of Crossbeam's code to gain a deeper understanding. During this process, we may occasionally seek your assistance with questions like those previously mentioned. We genuinely appreciate your support and guidance in these matters.