Re: [PATCH v2 0/7] srcu: NMI-safe atomic SRCU readers, and their irqsave forms

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 2313 bytes --]

On Wed, 2026-11-11 at 09:59 -0500, Mathieu Desnoyers wrote:
> 
> On architectures without NMI-safe this_cpu_inc(), every atomic reader
> now pays an atomic increment at lock and at unlock.  The update side is
> unchanged.  On architectures with NMI-safe this_cpu_inc(), such as x86,
> the readers use the same counter increments as before.  

The latter claim is only true in __srcu_read_lock_fast()/_fast_updown()
isn't it? Not in the atomic-SRCU. Unless we do something like this on
top?

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk> in case it actually
makes sense and you use it as anything other than a hint.

diff --git a/include/linux/srcu.h b/include/linux/srcu.h
index b6e6171b4cd4..a6db7bf55cde 100644
--- a/include/linux/srcu.h
+++ b/include/linux/srcu.h
@@ -342,7 +342,18 @@ static inline int __srcu_read_lock_atomic(struct srcu_struct *ssp)
 	if (in_task())
 		non_block_start();
 	srcu_check_read_flavor(ssp, SRCU_READ_FLAVOR_ATOMIC);
-	retval = __srcu_read_lock_nmisafe(ssp);
+	/*
+	 * Where this_cpu_inc() is NMI-safe, the this_cpu_inc() + smp_mb()
+	 * of __srcu_read_lock() is NMI-safe too, and is exactly what this
+	 * flavor's readers always paid. Only the architectures which
+	 * genuinely need the atomic increments take them; the pairing
+	 * with synchronize_srcu_atomic()'s counter scan is unchanged
+	 * either way (full barrier after the increment in both forms).
+	 */
+	if (!IS_ENABLED(CONFIG_NEED_SRCU_NMI_SAFE))
+		retval = __srcu_read_lock(ssp);
+	else
+		retval = __srcu_read_lock_nmisafe(ssp);
 	/*
 	 * An NMI may interrupt lockdep itself.  Annotate it as a trylock
 	 * there, as srcu_read_lock_nmisafe() does.
@@ -600,7 +611,10 @@ static inline void __srcu_read_unlock_atomic(struct srcu_struct *ssp, int idx)
 	WARN_ON_ONCE(idx & ~0x1);
 	srcu_check_read_flavor(ssp, SRCU_READ_FLAVOR_ATOMIC);
 	srcu_lock_release(&ssp->dep_map);
-	__srcu_read_unlock_nmisafe(ssp, idx);
+	if (!IS_ENABLED(CONFIG_NEED_SRCU_NMI_SAFE))
+		__srcu_read_unlock(ssp, idx);
+	else
+		__srcu_read_unlock_nmisafe(ssp, idx);
 	if (in_task())
 		non_block_end();
 }



> David agreed to
> the change, as long as synchronize_srcu_atomic() stays within the 4-16us
> it was designed for [3].

I set a test running (*without* the above). 

[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 6179 bytes --]