To let you know, me and Tom R. did review these changes and agreed that it is the least intrusive
changes for Hotspot shared code.
Thanks,
Vladimir
On 9/25/18 8:11 AM, Daniel D. Daugherty wrote:
Adding serviceability-dev@... since this is JVM/TI...
Dan
On 9/25/18 10:48 AM, Doug Simon wrote:
A major design point of Graal is to treat allocations as non-side effecting to give more freedom
to the optimizer by reducing the number of distinct FrameStates that need to be managed. When
failing an allocation, Graal will deoptimize to the last side effecting instruction before the
allocation. This mean the VM code for heap allocation will potentially be executed twice, once
from Graal compiled code and then again in the interpreter. While this is perfectly fine according
to the JVM specification, it can cause confusing behavior for JVMTI based tools. They will receive
2 ResourceExhausted events for a single allocation. Furthermore, the first ResourceExhausted event
(on the Graal allocation slow path) might denote a bytecode instruction that performs no
allocation, making it hard to debug the memory failure.
The proposed solution is to add an extra set of JVMCI VM runtime calls for allocation. These entry
points will attempt the allocation and upon failure,
skip side-effects such as posting JVMTI events or handling -XX:OnOutOfMemoryError. The compiled
code using these entry points is expected deoptmize on null.
The path from these new entry points to where allocation can fail goes through quite a bit of VM
code. One could modify all these paths by:
* Returning null instead of throwing an exception on failure.
* Adding a `bool null_on_fail` argument to all relevant methods.
* Adding extra null checking where necessary after each call to these methods when `null_on_fail
== true`.
This represents a significant number of changes.
Instead, the proposed solution introduces a new _in_retryable_allocation thread-local. This way,
only the entry points and allocation routines that raise an exception need to be modified. Failure
is communicated back to the new entry points by throwing a special pre-allocated OOME object
(i.e., Universe::out_of_memory_error_retry()) which must not propagate back to Java code. Use of
this object is not strictly necessary; it is introduced to highlight/document the special
allocation mode.
The proposed solution is at http://cr.openjdk.java.net/~dnsimon/8208686.
THE JBS bug is: https://bugs.openjdk.java.net/browse/JDK-8208686
-Doug
