[PATCH 3/5] locking/qspinlock: Add contended_release tracepoint

[Dmitry Ilvokhin <d@ilvokhin.com>]

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>
---
 include/asm-generic/qspinlock.h | 21 +++++++++++++++++++++
 kernel/Kconfig.locks            | 20 ++++++++++++++++++++
 kernel/locking/qspinlock.c      | 22 ++++++++++++++++++++++
 3 files changed, 63 insertions(+)

diff --git a/include/asm-generic/qspinlock.h b/include/asm-generic/qspinlock.h
index ae45289e8ec7..2ca94e41823b 100644
--- a/include/asm-generic/qspinlock.h
+++ b/include/asm-generic/qspinlock.h
@@ -41,6 +41,7 @@
 
 #include <asm-generic/qspinlock_types.h>
 #include <linux/atomic.h>
+#include <linux/tracepoint-defs.h>
 
 #ifndef queued_spin_is_locked
 /**
@@ -130,12 +131,32 @@ static __always_inline void queued_spin_release(struct qspinlock *lock)
 #endif
 
 #ifndef queued_spin_unlock
+
+DECLARE_TRACEPOINT(contended_release);
+
+extern void queued_spin_release_traced(struct qspinlock *lock);
+
 /**
  * queued_spin_unlock - unlock a queued spinlock
  * @lock : Pointer to queued spinlock structure
+ *
+ * Generic tracing wrapper around the arch-overridable
+ * queued_spin_release().
  */
 static __always_inline void queued_spin_unlock(struct qspinlock *lock)
 {
+	/*
+	 * Trace and release are combined in queued_spin_release_traced() so
+	 * the compiler does not need to preserve the lock pointer across the
+	 * function call, avoiding callee-saved register save/restore on the
+	 * hot path. queued_spin_release() is therefore called both here and in
+	 * queued_spin_release_traced(). Keep the two in sync.
+	 */
+	if (IS_ENABLED(CONFIG_QUEUED_SPINLOCKS_TRACE_CONTENDED_RELEASE) &&
+	    tracepoint_enabled(contended_release)) {
+		queued_spin_release_traced(lock);
+		return;
+	}
 	queued_spin_release(lock);
 }
 #endif
diff --git a/kernel/Kconfig.locks b/kernel/Kconfig.locks
index 4198f0273ecd..1c6423aafcd4 100644
--- a/kernel/Kconfig.locks
+++ b/kernel/Kconfig.locks
@@ -243,6 +243,26 @@ config QUEUED_SPINLOCKS
 	def_bool y if ARCH_USE_QUEUED_SPINLOCKS
 	depends on SMP
 
+config QUEUED_SPINLOCKS_TRACE_CONTENDED_RELEASE
+	bool "Trace contended_release on queued spinlocks"
+	depends on QUEUED_SPINLOCKS && TRACEPOINTS
+	help
+	  Fire the lock:contended_release tracepoint when a contended queued
+	  spinlock is released, so it is possible to attribute a contended
+	  spinlock to its holder.
+
+	  Architectures that can patch the unlock site do this at no cost and
+	  do not need this option.
+
+	  Everywhere else the check is compiled into queued_spin_unlock() and
+	  a small cost is paid on every unlock even when the tracepoint is
+	  disabled: a static-branch NOP and possibly a few more instructions
+	  to manage a stack frame.
+
+	  Sleeping locks fire lock:contended_release regardless of this option.
+
+	  If unsure, say N.
+
 config BPF_ARCH_SPINLOCK
 	bool
 
diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c
index af8d122bb649..33fe6d437c8f 100644
--- a/kernel/locking/qspinlock.c
+++ b/kernel/locking/qspinlock.c
@@ -104,6 +104,28 @@ static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
 #define queued_spin_lock_slowpath	native_queued_spin_lock_slowpath
 #endif
 
+#if !defined(queued_spin_unlock) && \
+	IS_ENABLED(CONFIG_QUEUED_SPINLOCKS_TRACE_CONTENDED_RELEASE)
+/*
+ * Out-of-line trace-and-release path for queued_spin_unlock(), used when
+ * the contended_release tracepoint is enabled.
+ *
+ * queued_spin_release() is duplicated here on purpose: doing the release
+ * in this function (rather than tracing here and releasing in the caller)
+ * lets queued_spin_unlock() return right after the call, so the
+ * tracepoint-disabled hot path never has to keep lock live across a call
+ * in a callee-saved register. Keep this release in sync with the one in
+ * queued_spin_unlock().
+ */
+void __lockfunc queued_spin_release_traced(struct qspinlock *lock)
+{
+	if (queued_spin_is_contended(lock))
+		trace_call__contended_release(lock);
+	queued_spin_release(lock);
+}
+EXPORT_SYMBOL(queued_spin_release_traced);
+#endif
+
 #endif /* _GEN_PV_LOCK_SLOWPATH */
 
 /**
-- 
2.53.0-Meta