locking/qspinlock: Add contended_release tracepoint

Unlike mutex and rw_semaphore, qspinlock has no owner field, so "perf
lock contention --lock-owner" cannot attribute a contended spinlock to
its holder. The waiter-side contention_begin event records that a
spinlock is contended, but not by whom. Firing contended_release in the
holder's context at unlock is the only way to capture the holder of a
contended spinlock.

Combine the contention check, trace call and release in an out-of-line
queued_spin_release_traced() so the compiler need not preserve the lock
pointer in a callee-saved register across the call.

The check in queued_spin_unlock() is paid on every unlock, even while
the tracepoint is disabled: a static-branch NOP on x86_64, and a few
more instructions to manage a stack frame elsewhere. Gate it behind
CONFIG_QUEUED_SPINLOCKS_TRACE_CONTENDED_RELEASE (default n) so nobody
pays for a tracepoint they do not use. Sleeping locks fire
contended_release regardless.

On x86 this generic path is used only with PARAVIRT_SPINLOCKS=n (e.g.
defconfig). PARAVIRT_SPINLOCKS=y kernels keep the paravirt static_call
unlock and are wired up separately.

All below are with the QUEUED_SPINLOCKS_TRACE_CONTENDED_RELEASE option
enabled.

_raw_spin_unlock(), x86_64 defconfig, GCC 11, tracepoint compiled in but
disabled. The unlock is the single 'movb'. The only instruction added to
the executed path is the 2-byte static-branch NOP. The CALL to the
traced helper and the JMP back are emitted out of line and are reached
only once the static branch is patched on:

          endbr64                            ; 4 bytes
          xchg   %ax,%ax                     ; 2 static-branch NOP
                                             ;   (added)
          movb   $0x0,(%rdi)                 ; 3 unlock (single store)
       A: decl   %gs:__preempt_count         ; 7
          je     B                           ; 2
          jmp    __x86_return_thunk          ; 5
          call   queued_spin_release_traced  ; 5 out of line, reached
                                             ;   only when the
                                             ;   tracepoint is on
          jmp    A                           ; 2 (added)
       B: call   __SCT__preempt_schedule     ; 5
          jmp    __x86_return_thunk          ; 5

Baseline is the same stream without the NOP and the out-of-line
CALL/JMP: 31 bytes vs 40 (+9 bytes).

Binary size impact on x86_64, defconfig: +680 bytes (+0.00%), since all
standard configs out-of-line unlock. Architectures with inlined unlock
(s390 (always), csky and loongarch (both when !PREEMPTION)) will see a
bigger increase in binary size.

On the same path (x86_64, PARAVIRT_SPINLOCKS=n) with the tracepoint
disabled, a _raw_spin_unlock()-heavy nginx workload [1] shows no
measurable difference between baseline and patched kernels in
throughput, latency, cycles, instructions, IPC, or L1 instruction-cache
misses (kernel and total): all deltas stay within run-to-run noise.

Unlike x86, on arm64 the frame setup code (STP, MOV and LDP) lands on
the executed path in addition to static-branch NOP. Binary size impact
on arm64, defconfig: +932 bytes (+0.00%).

The _raw_spin_unlock()-heavy nginx workload reflects the larger hot
path: L1 instruction-cache misses rise ~1.4% (kernel and total) and
instruction count ~0.4%, consistent with the per-unlock frame.
cpu_cycles, throughput and latency show no measurable change and are
within run-to-run noise.

Architectures with fully custom qspinlock implementations (e.g.
PowerPC) are not covered by this change.

[1]: https://lore.kernel.org/all/aiphFXe_TPNPxZ_n@shell.ilvokhin.com/

Signed-off-by: Dmitry Ilvokhin <d@ilvokhin.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Juergen Gross <jgross@suse.com>
Link: https://patch.msgid.link/0d998e22a0c595f670cfc6725bb683323aced5cb.1785778551.git.d@ilvokhin.com

0 files changed, 0 insertions, 0 deletions