[16]RFR(S):8249092:InternalError: a fault occurred in a recent unsafe memory access operation in compiled Java code

[David Holmes]

Hi,

On 3/09/2020 12:07 pm, Wang Zhuo(Zhuoren) wrote:
> Hi Patric,
> The original problem(https://bugs.openjdk.java.net/browse/JDK-8246051) 
> is architecture specific. When running TestUnsafeUnalignedSwap.java on 
> aarch64 platforms, JVM crashes without the fix because aarch64 does not 
> support unaligned compare_and_swap. On X86 platforms the crash cannot be 
> reproduced because X86 support unaligned compare_and_swap.

Patric is asking about the original situation where an unaligned CAS was 
performed, which motivated the change made by JDK-8246051.

The GuardUnsafeAccess mechanism (or more specifically the signal 
handling tricks underneath it) was not intended for general use, but was 
specifically created to deal with the case of page faults related to 
mapped ByteBuffers where we wanted application code using JDK exported 
APIs to get an InternalError when they did the wrong thing with their 
mapped files, instead of crashing. No part of the JDK should be calling 
Unsafe.compareAndSwap* with unaligned data - if it does that is a bug. 
If third-party code is using Unsafe directly then that is their problem 
and we do not try to make things easier for them.

The use of GuardUnsafeAccess with the CAS primitives on some platforms 
can result in an infinite loop as the mechanism cannot be applied to 
arbitrary code sequences. Somewhat ironically Aarch64 is one platform 
that can suffer from this.

Cheers,
David
-----

> 
> Regards,
> Zhuoren
> 
>     ------------------------------------------------------------------
>     From:Patric Hedlin <patric.hedlin at oracle.com>
>     Sent At:2020 Sep. 2 (Wed.) 20:22
>     To:Sandler <zhuoren.wz at alibaba-inc.com>; aarch64-port-dev
>     <aarch64-port-dev at openjdk.java.net>; hotspot-runtime-dev
>     <hotspot-runtime-dev at openjdk.java.net>
>     Cc:david.holmes <david.holmes at oracle.com>; rahul.v.raghavan
>     <rahul.v.raghavan at oracle.com>
>     Subject:Re: [16]RFR(S):8249092:InternalError: a fault occurred in a
>     recent unsafe memory access operation in compiled Java code
> 
>     Hi Zhuoren,
> 
>     I don't actually know what behaviour to expect from the Unsafe
>     atomics in this test-case but perhaps you could re-cap the original
>     problem (addressed in JDK-8246051) since it seems to raise some
>     questions. Did you have an example (real code) where this behaviour
>     is essential?
> 
>     Best regards,
>     Patric Hedlin
> 
>     (Including hotspot-runtime-dev at openjdk.java.net)
> 
> 
>     On 2020-09-01 07:35, Wang Zhuo(Zhuoren) wrote:
>     Hi, this is a fix for a test case.
>     In -Xcomp mode, compiler/unsafe/TestUnsafeUnalignedSwap.java will
>     fail because the catch misses the error due to async exception.
>     This patch uses a loop to make sure the error can be caught. Also
>     -Xcomp is added in test.
>     BUG: https://bugs.openjdk.java.net/browse/JDK-8249092
>     Patch: http://cr.openjdk.java.net/~wzhuo/8249092/webrev.00/
> 
> 
>     Regards,
>     Zhuoren
> 
>