Thursday, April 6, 2017

Applications crash on SLES12 due to lock elision

This issue has been discussed in other places, but mostly related to specific applications and I think it needs its own post here for those who would stumble on this following a Google search.

glibc 2.18, released in 2013, came with a new feature named TSX Lock Elision.

Briefly, this feature changes the behaviour of in the way it handles mutexes on some specific processors that support hardware lock elision. Intel Xeon CPUs, in particular, support TSX since around 2013 or so. Lock elision offers significant performance gains for some software such as databases.

You can see if your Linux server's CPU supports lock elision by checking /proc/cpuinfo. If it mentions "hle" (hardware lock elision), it does.

RHEL 7 does not support this as of now. It comes with glibc 2.17, so lock elision is not enabled on these systems. As for SLES12, it comes with glibc 2.19, which means that SLES12 systems will use lock elision if the CPU supports it.

However, if an application unlocks a mutex twice, this can cause problems if lock elision is enabled. This is explained in detail in an LWN article. Let me quote an important paragraph in this article:

pthread_mutex_unlock() detects whether the current lock is executed transactionally by checking if the lock is free. If it is free it commits the transaction, otherwise the lock is unlocked normally. This implies that if a broken program unlocks a free lock, it may attempt to commit outside a transaction, an error which causes a fault in RTM. In POSIX, unlocking a free lock is undefined (so any behavior, including starting World War 3 is acceptable). It is possible to detect this situation by adding an additional check in the unlock path. The current glibc implementation does not do this, but if this programming mistake is common, the implementation may add this check in the future.

The "programming mistake" here is double-unlocking mutexes. I've made a sample C program that does exactly this, and although it works fine with glibc 2.17, it will crash on glibc 2.19 with a segmentation fault in __lll_unlock_elision(), if, and only if, the server's cpuinfo reports "hle".

I've stumbled upon a few applications, which I will not name here, that crash on SLES12. Upon analyzing their cores, I found that they have this same exact problem with __lll_unlock_elision(). So, one can assume that they might double-unlock some mutexes.

The bottom line is that if you have an app that does this, your best bet is to contact the vendor, and ask them to remove double mutex unlocks in their code, if they have any.

If that is not possible, there are two workarounds:

1. The first is to patch /etc/ to override libpthread 2.19 with a version that is compiled with lock elision disabled. This is documented in Novell's KB here.

2. The second (and preferred) solution is to adjust LD_LIBRARY_PATH to override it on a per-application basis. You could therefore change its startup script to add this:


Hope this helps.

No comments: