[aarch64-port-dev ] RFR(xs): Aarch64: ReservedStackAccess may leave stack guard in inconsistent state

[David Holmes]

On 20/12/2018 5:14 am, dean.long at oracle.com wrote:
> Is there a bug ID for this?  It's not clear to me which repo(s) this is 
> targeted for.

There are also now two threads on this - one with [aarch64-port-dev ] 
in the subject (I do wish mailing lists would not do that as it totally 
messes up threads posted to multiple lists!).

> 
> The aarch64 part looks good.
> 
> The shared part could be simplified.  There are only 4 guard states, so 
> checking
> for != to 3 of them is the same as checking for == to the one that is 
> allowed.
> Also,
> 
> 2576   assert(_stack_guard_state != stack_guard_unused, "must be using 
> guard pages.");
> 2577   assert(_stack_guard_state != stack_guard_reserved_disabled, 
> "already disabled");
> 2578   assert(_stack_guard_state != 
> stack_guard_yellow_reserved_disabled, "loosing information");
> 2579
> 2580   // Simply return if called for a thread that does not use guard 
> pages.
> 2581   if (_stack_guard_state == stack_guard_unused) return;
> 2582
> 2583   // Already disabled together with yellow pages.
> 2584   if (_stack_guard_state == stack_guard_yellow_reserved_disabled) 
> return;
> 
> the if-returns don't seems very useful after the asserts.

They are there to ensure that in product mode if we hit a case not 
exposed by testing then we don't crash. Should never happen but ...

Cheers,
David

> dl
> 
> On 12/19/18 6:08 AM, Andrey Petushkov wrote:
>> Hi All,
>>
>> Please review small fix for for the following problem: during 
>> execution of
>> ReservedStackAccess annotated method it might happen that reserved pages
>> are reguarded while yellow pages left unguarded. At the same time the 
>> state
>> of stack guard is recorded as fully protected (stack_guard_enabled).
>> Subsequent execution of ReservedStackAccess annotated method then may 
>> lead
>> to VM crash instead of mere StackOverflowError.
>> Reproduced with ReservedStackTestCompiler jtreg test on aarch32-jdk11 
>> port
>> codebase. The bug seem to be apparent and exists in aarch64-port as well,
>> hence the review request.
>> The actual fix is in aarch64-specific code but the shared code is changed
>> as well to help to catch such bugs earlier in the future
>>
>> the link http://cr.openjdk.java.net/~apetushkov/ReservedStackAccess/
>>
>> Thanks,
>> Andrey
>