RFR(S): 8152947: VM crash with assert(!removed || is_in_use()) failed: unused osr nmethod should be invalidated
tobias.hartmann at oracle.com
Tue Mar 29 15:28:48 UTC 2016
On 29.03.2016 15:05, Dmitry Dmitriev wrote:
> Hi Tobias,
> I think that "#ifdef ASSERT" must be instead of "#ifndef ASSERT" on lines 1338 and 1516. And it seems that the whole fix should be retested in this case.
Oh, thanks for catching this!
Here is the new webrev:
I verified that this (still) fixes the problem and re-submitted the tests.
> On 29.03.2016 15:01, Tobias Hartmann wrote:
>> please review the following patch:
>> JDK-8023191 added an assert to check if osr nmethods are invalidated (i.e. removed from the osr list). The assert is too strong because nmethod::invalidate_osr_method() may be executed by multiple threads that compete for making an osr nmethod not-entrant:
>> Thread 1: Invalidates the osr nmethod and blocks just before executing the assert in line 1397.
>> Thread 2: nmethod is already invalidated. Acquires patching lock and sets the nmethod to not-entrant.
>> Thread 1: Continues execution and hits the assert because the nmethod was on the list and is (now) not-entrant.
>> Thread 1 should ignore this and return. See also comment in nmethod::make_not_entrant_or_zombie():
>> // another thread already performed this transition so nothing
>> // to do, but return false to indicate this.
>> I changed the verification code to check for osr invalidation *after* the state of the nmethod was set.
>> Tested with failing test, JPRT and RBT (in progress).
More information about the hotspot-compiler-dev