RFR: 8305403: Shenandoah evacuation workers may deadlock

[William Kemper]

On Mon, 3 Apr 2023 21:45:39 GMT, William Kemper <wkemper at openjdk.org> wrote:

> Shenandoah evacuation workers may deadlock if a safepoint begins during concurrent evacuation _and_ a GC worker thread experiences an out of memory error during the evacuation.
> 
> This situation occurs because some of the workers have set the cancelled state to `NOT_CANCELLED` and then yielded to the suspendible thread set because a safepoint is starting. Other workers who then experience an OOM during evacuation attempt to transition the heap from `CANCELLABLE` to `CANCELLED` in a CAS loop that will never succeed (because the cancelled state is `NOT_CANCELLED`). These workers are unable to join the suspendible thread set, so the threads which have yielded are unable to resume and reset the heap to `CANCELLABLE`. The VM thread cannot enter the safepoint and eventually all of the mutator threads block when they are unable to allocate.
> 
> The changes here remove the `NOT_CANCELLED` state and remove the CAS loop used to transition from `CANCELLABLE` to `CANCELLED`. Additionally, worker threads that need to suspend no longer put the heap in the `NOT_CANCELLED` state before suspending.
> 
> In order to test this, the behavior of the diagnostic flag was modified to provoke this scenario:
> 
> diff --git a/src/hotspot/share/gc/shenandoah/shenandoahHeap.inline.hpp b/src/hotspot/share/gc/shenandoah/shenandoahHeap.inline.hpp
> index 4158f4bee22..e261dd3a81b 100644
> --- a/src/hotspot/share/gc/shenandoah/shenandoahHeap.inline.hpp
> +++ b/src/hotspot/share/gc/shenandoah/shenandoahHeap.inline.hpp
> @@ -292,8 +292,7 @@ inline oop ShenandoahHeap::evacuate_object(oop p, Thread* thread) {
>    HeapWord* copy = nullptr;
>  
>  #ifdef ASSERT
> -  if (ShenandoahOOMDuringEvacALot &&
> -      (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call
> +  if (ShenandoahOOMDuringEvacALot && thread->is_Worker_thread() && SuspendibleThreadSet::should_yield() && uint(os::random()) % 1000 < 1) {
>          copy = nullptr;
>    } else {
>  #endif
> 
> 
> This applies the failure only to worker threads and only when they are expected to suspend and only with a 1/1000 chance of failing for each evacuation. Without these fixes, the modification to `ShenandoahOOMDuringEvacALot` causes a deadlock in under two minutes of execution on [the extremem benchmark](https://github.com/corretto/heapothesys/tree/master/Extremem) with these additional diagnostic flags:
> 
> -XX:+SafepointALot -XX:+ShenandoahOOMDuringEvacALot -XX:ConcGCThreads=12 -XX:GuaranteedSafepointInterval=50
> 
> After these changes, the benchmark no longer deadlocks.
> 
> Additional testing includes jtreg `test/hotspot/jtreg:hotspot_gc_shenandoah`, the Dacapo benchmark suite, HyperAlloc and specjbb 2015.

>  Is there a reason why this piece of code that aided testing isn't in the PR

I felt it was too specific (only targeted worker threads that should suspend). I can add it to the PR. I would also like to change the `AllocFailALot` and `OOMDuringEvacALot` flags to express a percentage, but that would be a separate PR. (And might need a CSR?)

-------------

PR Comment: https://git.openjdk.org/jdk/pull/13309#issuecomment-1509075428