[jdk11u-dev] RFR: 8191278: MappedByteBuffer bulk access memory failures are not handled gracefully

[Andrew Haley]

On 8/5/21 8:15 PM, Sergey Nazarkin wrote:
> The patch looks scary but in general it just adds a guard at the beginning of unsafe copy block. So a bunch of code just shifted right without single line modification.
>
> We got a few reports from the users about JVM crash when they modify file being mapped. It was  definitely application issue but Unsafe contains protection for primitive types so it might be worth to have the same for copy operation.

Sure, I understand. However, the only thing this patch changes is that
you get a Java stack trace instead of a VM dump. Your program exits,
either way. So it really is just about being graceful.

The default response to enhancements in old releases is no, and there's
a good reason for that.

We need to learn from experience. Take, for example, JDK-8267689. That
was a fix for a rare and hard-to-provoke crash when accessing
misaligned Unsafe memory. The backport patch looked reasonable enough,
and was working fine on the head release, but it was wrong. In fact,
the head release had a subtle but latent bug, which wasn't so latent
in 8u. This bug was worse than the original bug that JDK-8267689 was
supposed to fix.

I think we shouldn't risk regressions like that unless we actually
need to.

-- 
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671


-- 
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671