Re: [REGRESSION] x86: Use RCU in all users of,__module_address() slows kTLS+wrk workload on KASAN+PROVE_LOCKING+kmemleak

From: Peter Zijlstra

Date: Fri Nov 20 2026 - 06:59:03 EST


On Thu, Nov 19, 2026 at 09:54:37PM +0200, Yael Chemla wrote:
> Hi,
>
> A kTLS stress test started failing in v6.15-rc1 on kernels built with
> KASAN and PROVE_LOCKING. It was bisected to commit 14daa3bca217 ("x86:
> Use RCU in all users of __module_address().")
>
> The test runs wrk (1000 connections, 15s, --timeout 5s) against nginx
> serving a 1 MB file over kTLS, with 1% packet loss. It fails because too
> few requests complete: wrk reports only timeouts, no connect/read/write
> errors. With --timeout 10s it passes.
>
> Reverting 14daa3bca217 on top of 01d5b167dc23 (modules-6.15-rc1 merge)
> or v7.4-rc2 (7475ea8b5861), with the same config, makes the test pass;
> without the revert it fails on both.
>
> The failing config has KASAN, PROVE_LOCKING, kmemleak and UBSAN enabled.
> Building without KASAN, or without PROVE_LOCKING, also passes.
>
> I have not looked into where the time goes.
>
> #regzbot introduced: 14daa3bca217

Anything *SAN is not a performance target, not a regression.