Re: 8aeb879baf12 - significant system call latency regression, bisected

[Peter Zijlstra]

On Fri, Jun 19, 2026 at 03:23:52AM -0700, H. Peter Anvin wrote:
> On June 19, 2026 1:14:27 AM PDT, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >On Wed, Jun 17, 2026 at 02:37:18PM +0200, Peter Zijlstra wrote:
> >>  - makes -fno-jump-tables unconditional
> >>  - removes array_index_nospec() from the syscall dispatch
> >
> >FWIW, this also allows making all SYSCALLs __noendbr, very much
> >including the 'legacy' sys_call_table :-)
> >
> >Compile tested with IA32_EMULATION=n and reliably yields:
> >
> >vmlinux.o: warning: objtool: sys_call_table+0x0: data relocation to !ENDBR: __x64_sys_read+0x0
> >...
> >vmlinux.o: warning: objtool: sys_call_table+0xeb0: data relocation to !ENDBR: __x64_sys_listns+0x0
> >
> >(which is just one little objtool patch away from being fixed)
> >
> >and boots fine (in kvm).

> *Very* nice indeed. 
> 
> I was definitely hoping this would be the next step.

A few patches here:

  git://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git x86/syscall

That now includes IA32_EMULATION and a few terse changelogs. The
alignment thing went missing since that didn't seem to be very good.