Bug #21046: Backport: TLS fix for ARM64

• Status: Closed
• Assignee: -
• Target version: -
• Backport: 3.1: REQUIRED, 3.2: REQUIRED, 3.3: REQUIRED, 3.4: REQUIRED

[nobu (Nobuyoshi Nakada)]

f7059af50a31a4d27a04ace0beadb60616f3f971

Use no-inline version rb_current_ec on Arm64

The TLS across .so issue seems related to Arm64, but not Darwin.

[Eregon (Benoit Daloze)]

Does this mean all current releases of CRuby might hit this issue on linux-aarch64?
Do you know if it happens every time rb_current_ec() is called or only in some cases?

[kjtsanaktsidis (KJ Tsanaktsidis)]

Heya - sorry I've been away from Ruby for a little while, but this caught my eye and I wanted to bring your attention to a similar bug I opened about rb_current_ec with M:N threading last year: https://bugs.ruby-lang.org/issues/20243

I don't know what reproduction you found for this particular problem that led to your fix, but I wonder: did it involve M:N threading? If so, I wonder if it's the same issue that I reported.

In that case we actually probably need to force un-inlining of rb_current_ec if M:N threading is enabled in general, I would have thought.

[nagachika (Tomoyuki Chikanaga)]

Backporting commit git:f7059af50a31a4d27a04ace0beadb60616f3f971 into ruby_3_2 depends on the changeset for the M:N Threads scheduler for Ractor (git:be1bbd5b7d40ad863ab35097765d3754726bbd54).
I would appreciate it if you could create a backport PR for ruby_3_2.