On Thu, 18 Jan 2024 22:54:26 GMT, Chris Plummer <cjplum...@openjdk.org> wrote:
> In PointerFinder.java we have some code to determine if a pointer is in a
> TLAB, but it only executes for the SerialGC. It should work for all GCs, so I
> moved the code out of the SerialGC block.
>
> I also cleaned up the printing in PointerLocation. java a bit so when not
> using verbose mode not as much info about the tlab address is printed. This
> is consistent with other addresses, such as java stack addresses, which is
> what I modeled this change on.
>
> It's hard to test this change since it is hard to consistently get an address
> to be in the tlab. I wrote a little test program that just sits in a loop
> doing allocations. I attached to it with clhsdb and ran the threadcontext
> command, which does a fincpc on each register. About half the time the main
> thread was suspended in a frame where some registers where pointing into the
> tlab, and I confirmed this was the case for both SerialGC and G1. Here's an
> example of one register with verbose off and verbose on:
>
> rsi: 0x000000008a5d4448: In TLAB for thread "main"
> sun.jvm.hotspot.runtime.JavaThread@0x00007ffa24029000
>
> rsi: 0x000000008a5d4448: In TLAB for thread ("main" #1 prio=5
> tid=0x00007ffa24029000 nid=25392 runnable [0x0000000000000000]
> java.lang.Thread.State: RUNNABLE
> JavaThread state: _thread_in_java
> )
> [0x000000008a5d4448,0x000000008ab724b8,0x000000008b0c0250,{0x000000008b0c0490})
>
> For testing I ran all tier1, tier2, and tier5 svc tests (still in progress)
Looks good to me.
(Not critical, but maybe PointerLocation doesn't need the boolean inTlab. It
could use that its tlab is not null, like the isInHeap/stack/metadata checks
that something is not null.)
-------------
Marked as reviewed by kevinw (Committer).
PR Review: https://git.openjdk.org/jdk/pull/17494#pullrequestreview-1846152245
