On Wed, Jun 03, 2026 at 02:08:11PM +0200, Peter Zijlstra wrote:
> On Thu, May 14, 2026 at 12:34:55PM +0000, Dmitry Ilvokhin wrote:
> 
> > Baseline, in best case scenario of least number of executed
> > instructions.
> > 
> >     3e0:  endbr64                          ; 4 bytes (always executed)
> >     3e4:  movb $0x0,(%rdi)                 ; 3 bytes (unlock,
> >                                            ; always executed)
> >     3e7:  decl %gs:__preempt_count         ; 7 bytes (always executed)
> >     3ee:  je   3f5                         ; 2 bytes (always executed)
> >     3f0:  jmp  __x86_return_thunk          ; 5 bytes (executed if above
> >                                            ; je is not taken)
> >                                            ; rest is not executed
> >     3f5:  call __SCT__preempt_schedule     ; 5 bytes
> >     3fa:  jmp  __x86_return_thunk          ; 5 bytes
> > 
> > Tracepoint (again same case of least number of executed instructions).
> > 
> >     bc0:  endbr64                          ; 4 bytes (always executed)
> >     bc4:  xchg %ax,%ax                     ; 2 bytes (always executed, this 
> > is an
> >                                            ; only addition on the execution 
> > path).
> >     bc6:  movb $0x0,(%rdi)                 ; 3 bytes (unlock, always 
> > executed)
> >     bc9:  decl %gs:__preempt_count         ; 7 bytes (always executed)
> >     bd0:  je   bde                         ; 2 bytes (always executed)
> >     bd2:  jmp  __x86_return_thunk          ; 5 bytes (executed if above
> >                                            ; je is not taken)
> >                                            ; rest is not executed
> >     bd7:  call queued_spin_release_traced  ; 5 bytes
> >     bdc:  jmp  bc9                         ; 2 bytes
> >     bde:  call __SCT__preempt_schedule     ; 5 bytes
> >     be3:  jmp  __x86_return_thunk          ; 5 bytes
> > 
> 
> So I've been playing with this a bit, and it is all really sad.
> 
> Now, since pretty much everybody+dog will have PARAVIRT_SPINLOCK=y, the
> 'best' solution would be changing that paravirt call with a
> static_call(), that actually shrinks the code by 1 byte.
> 
> And then this tracepoint nonsense can simply use a different unlock
> function, just like paravirt.
> 
> 0000 00000000000001d0 <_raw_spin_unlock>:
> 0000  1d0:    f3 0f 1e fa             endbr64
> 0004  1d4:    ff 15 00 00 00 00       call   *0x0(%rip)        # 1da 
> <_raw_spin_unlock+0xa>   1d6: R_X86_64_PC32      pv_ops_lock+0x4
> 000a  1da:    65 ff 0d 00 00 00 00    decl   %gs:0x0(%rip)        # 1e1 
> <_raw_spin_unlock+0x11>       1dd: R_X86_64_PC32      __preempt_count-0x4
> 0011  1e1:    74 06                   je     1e9 <_raw_spin_unlock+0x19>
> 0013  1e3:    2e e9 00 00 00 00       cs jmp 1e9 <_raw_spin_unlock+0x19>      
> 1e5: R_X86_64_PLT32     __x86_return_thunk-0x4
> 0019  1e9:    e8 00 00 00 00          call   1ee <_raw_spin_unlock+0x1e>      
> 1ea: R_X86_64_PLT32     __SCT__preempt_schedule-0x4
> 001e  1ee:    2e e9 00 00 00 00       cs jmp 1f4 <_raw_spin_unlock+0x24>      
> 1f0: R_X86_64_PLT32     __x86_return_thunk-0x4
> 
> 
> 0000 00000000000001d0 <_raw_spin_unlock>:
> 0000  1d0:    f3 0f 1e fa             endbr64
> 0004  1d4:    e8 00 00 00 00          call   1d9 <_raw_spin_unlock+0x9>       
> 1d5: R_X86_64_PLT32     __SCT__queued_spin_unlock-0x4
> 0009  1d9:    65 ff 0d 00 00 00 00    decl   %gs:0x0(%rip)        # 1e0 
> <_raw_spin_unlock+0x10>       1dc: R_X86_64_PC32      __preempt_count-0x4
> 0010  1e0:    74 06                   je     1e8 <_raw_spin_unlock+0x18>
> 0012  1e2:    2e e9 00 00 00 00       cs jmp 1e8 <_raw_spin_unlock+0x18>      
> 1e4: R_X86_64_PLT32     __x86_return_thunk-0x4
> 0018  1e8:    e8 00 00 00 00          call   1ed <_raw_spin_unlock+0x1d>      
> 1e9: R_X86_64_PLT32     __SCT__preempt_schedule-0x4
> 001d  1ed:    2e e9 00 00 00 00       cs jmp 1f3 <_raw_spin_unlock+0x23>      
> 1ef: R_X86_64_PLT32     __x86_return_thunk-0x4
> 
> 
> Something a little like so, which is completely untested, except to
> build kernel/locking/spinlock.o (with clang-23).

Thanks a lot for taking a look, Peter.

I like the static_call idea. It's truly zero cost on x86 (and, as you
note, even a byte smaller). The one caveat is that it relies on
HAVE_STATIC_CALL_INLINE to stay free. 

So my plan would be: static_call where HAVE_STATIC_CALL_INLINE is
available (x86), and a static branch fallback elsewhere, gated behind a
default-off config so it imposes nothing on arches/kernels that don't
opt in. I'm mostly interested in x86, but would like arm64 to work too,
which would use the fallback.

Concretely:

1. Split the sleepable-lock patches out and send them separately.
   They're independent of the static call work and look far less
   controversial.

2. Convert the paravirt spinlock unlock to a static_call, as the
   foundation for the unlock tracepoint. I'm happy to take a stab at it.
   Let me know if you'd rather do it yourself.

3. Build the unlock tracepoint on top: static_call where it's cheap,
   config-gated static_branch fallback where it isn't.

Does this plan sound reasonable to you?

> 
> Also, I think someone should go do some performance runs with
> ARCH_INLINE_SPIN_* set for x86 just like for s390.

That's a good point, I'll run benchmarks and report back with the
results.

Reply via email to