[PATCH 0/19] Add atomic SRCU

[PATCH 0/19] Add atomic SRCU

[Paul E. McKenney]

Hello!

This series adds an atomic SRCU and some additional adjustments:

1.	Make a Tiny SRCU grace period imply an RCU grace period.

2.	Suppress srcu_advance_state() mutex_lock in atomic.

3.	Suppress to-big transition for atomic SRCU.

4.	Add an atomic Tree SRCU.

5.	Add an atomic Tiny SRCU.

6.	Add support for testing synchronize_srcu_atomic().

7.	Disable preemption across synchronize_srcu_atomic().

8.	Use IRQ_WORK_INIT_HARD for srcu's irq_work, courtesy of Sebastian
	Andrzej Siewior.

9.	Fix WARN_ON() for rcu_segcblist_n_cbs() in cleanup_srcu_struct(),
	courtesy of Sunho Park.

10.	Warn if Tiny SRCU readers are preempted.

11.	Explicitly note DEFINE_SRCU() needs for srcu_barrier().

12.	Add atomic-SRCU support to torture.sh, courtesy of Kunwu Chan.

13.	Add reader-free fastpath to synchronize_srcu_atomic(), courtesy
	of Kunwu Chan.

14.	Skip callback scheduling for atomic SRCU grace periods, courtesy
	of Kunwu Chan.

15.	Remove srcu_barrier() sleep for atomic SRCU, courtesy of Kunwu
	Chan.

16.	Restrict atomic-SRCU non_block annotation to task context,
	courtesy of Kunwu Chan.

17.	Make init_srcu_struct_atomic() prevent transition to big,
	courtesy of Kunwu Chan.

18.	Don't transition atomic SRCU to big in srcu_gp_end(), courtesy
	of Kunwu Chan.

19.	Skip torture to-big transition for atomic SRCU, courtesy of
	Kunwu Chan.

						Thanx, Paul

------------------------------------------------------------------------

 b/include/linux/srcu.h                              |   81 +++++
 b/include/linux/srcutiny.h                          |    6 
 b/include/linux/srcutree.h                          |    9 
 b/kernel/rcu/rcutorture.c                           |   46 ++-
 b/kernel/rcu/srcutiny.c                             |    2 
 b/kernel/rcu/srcutree.c                             |   23 +
 b/tools/testing/selftests/rcutorture/bin/torture.sh |   24 +
 include/linux/srcu.h                                |   10 
 include/linux/srcutree.h                            |    6 
 kernel/rcu/srcutiny.c                               |   87 ++++++
 kernel/rcu/srcutree.c                               |  285 +++++++++++++++++---
 11 files changed, 511 insertions(+), 68 deletions(-)

[PATCH 01/19] srcutiny: Make a Tiny SRCU grace period imply an RCU grace period

[Paul E. McKenney]

In non-preemptible kernels, a Tiny SRCU grace period implies an RCU
grace period because any context switch suffices.  But in preemptible
kernels, it is possible for a Tiny SRCU grace period to elapse without
a corresponding RCU grace period.

Which was OK until RCU Tasks Trace was re-implemented in terms
of SRCU-fast, which in Tiny SRCU is implemented as SRCU, which is
already fast.  And RCU Tasks Trace grace periods are required to imply
RCU grace periods.  This commit therefore adds a synchronize_rcu(),
but only in preemptible kernels.

Because preemptible Tiny SRCU is not in mainline, this added call to
synchronize_rcu() will not slow anything down:  The comparison would
instead be with TREE SRCU.  But if this added call ever becomes a problem,
the Tiny SRCU srcu_struct structure could track whether or not this is for
SRCU-fast, and to add the synchronize_rcu() only in the SRCU-fast case.
However, at the moment, this is seen as unnecessary complexity.

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutiny.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/kernel/rcu/srcutiny.c b/kernel/rcu/srcutiny.c
index 5de9a6905838..32b37d63d58a 100644
--- a/kernel/rcu/srcutiny.c
+++ b/kernel/rcu/srcutiny.c
@@ -159,6 +159,8 @@ void srcu_drive_gp(struct work_struct *wp)
 	WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1);
 	WRITE_ONCE(ssp->srcu_gp_waiting, true);  /* srcu_read_unlock() wakes! */
 	preempt_enable();
+	if (IS_ENABLED(CONFIG_PREEMPTION))
+		synchronize_rcu(); // Needed for RCU Tasks Trace to imply RCU grace period
 	do {
 		// Deadlock issues prevent __srcu_read_unlock() from
 		// doing an unconditional wakeup, so polling is required.
-- 
2.40.1

[PATCH 02/19] srcutree: Suppress srcu_advance_state() mutex_lock in atomic

[Paul E. McKenney]

The new SRCU_READ_FLAVOR_ATOMIC prohibits sleeping in both readers and
grace-period waits in order to permit waiting on an SRCU grace period
within an OOM notifier.  This means that srcu_advance_state() cannot
acquire ->srcu_gp_mutex in this case, because mutex_lock() can sleep.

This commit therefore adds an is_atomic parameter to srcu_advance_state().
When this parameter is false, current behavior is maintained, in other
words, ->srcu_gp_mutex is acquired and released as before.  But when this
is_atomic parameter is false, the caller is responsible for excluding
other callers.

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutree.c | 23 +++++++++++++++--------
 1 file changed, 15 insertions(+), 8 deletions(-)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index d32c374ee72a..dc063eb49b0d 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -1941,13 +1941,16 @@ EXPORT_SYMBOL_GPL(srcu_batches_completed);
 /*
  * Core SRCU state machine.  Push state bits of ->srcu_gp_seq
  * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
- * completed in that state.
+ * completed in that state.  Set is_atomic to indicate that
+ * the caller is excluding other calls so that ->srcu_gp_mutex
+ * is not needed, and to indicate that blocking is forbidden.
  */
-static void srcu_advance_state(struct srcu_struct *ssp)
+static void srcu_advance_state(struct srcu_struct *ssp, bool is_atomic)
 {
 	int idx;
 
-	mutex_lock(&ssp->srcu_sup->srcu_gp_mutex);
+	if (!is_atomic)
+		mutex_lock(&ssp->srcu_sup->srcu_gp_mutex);
 
 	/*
 	 * Because readers might be delayed for an extended period after
@@ -1965,7 +1968,8 @@ static void srcu_advance_state(struct srcu_struct *ssp)
 		if (ULONG_CMP_GE(ssp->srcu_sup->srcu_gp_seq, ssp->srcu_sup->srcu_gp_seq_needed)) {
 			WARN_ON_ONCE(rcu_seq_state(ssp->srcu_sup->srcu_gp_seq));
 			raw_spin_unlock_irq_rcu_node(ssp->srcu_sup);
-			mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
+			if (!is_atomic)
+				mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
 			return;
 		}
 		idx = rcu_seq_state(READ_ONCE(ssp->srcu_sup->srcu_gp_seq));
@@ -1973,7 +1977,8 @@ static void srcu_advance_state(struct srcu_struct *ssp)
 			srcu_gp_start(ssp);
 		raw_spin_unlock_irq_rcu_node(ssp->srcu_sup);
 		if (idx != SRCU_STATE_IDLE) {
-			mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
+			if (!is_atomic)
+				mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
 			return; /* Someone else started the grace period. */
 		}
 	}
@@ -1981,7 +1986,8 @@ static void srcu_advance_state(struct srcu_struct *ssp)
 	if (rcu_seq_state(READ_ONCE(ssp->srcu_sup->srcu_gp_seq)) == SRCU_STATE_SCAN1) {
 		idx = !(ssp->srcu_ctrp - &ssp->sda->srcu_ctrs[0]);
 		if (!try_check_zero(ssp, idx, 1)) {
-			mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
+			if (!is_atomic)
+				mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
 			return; /* readers present, retry later. */
 		}
 		srcu_flip(ssp);
@@ -1999,7 +2005,8 @@ static void srcu_advance_state(struct srcu_struct *ssp)
 		 */
 		idx = !(ssp->srcu_ctrp - &ssp->sda->srcu_ctrs[0]);
 		if (!try_check_zero(ssp, idx, 2)) {
-			mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
+			if (!is_atomic)
+				mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
 			return; /* readers present, retry later. */
 		}
 		ssp->srcu_sup->srcu_n_exp_nodelay = 0;
@@ -2107,7 +2114,7 @@ static void process_srcu(struct work_struct *work)
 	sup = container_of(work, struct srcu_usage, work.work);
 	ssp = sup->srcu_ssp;
 
-	srcu_advance_state(ssp);
+	srcu_advance_state(ssp, false);
 	raw_spin_lock_irq_rcu_node(ssp->srcu_sup);
 	curdelay = srcu_get_delay(ssp);
 	raw_spin_unlock_irq_rcu_node(ssp->srcu_sup);
-- 
2.40.1

[PATCH 03/19] srcutree: Suppress to-big transition for atomic SRCU

[Paul E. McKenney]

Initially (and perhaps forever), call_srcu() will not be available for
atomic srcu_struct structures.  There is therefore no reason to transition
such a structure to big, because the main purpose of such a transition
is to reduce lock contention for concurrent SRCU callback queueing.

This commit therefore adds an is_atomic parameter to both the
check_init_srcu_struct() and init_srcu_struct_fields() functions, which
suppresses the initialization-time transition to big that is enabled by
default on large systems.

It will still be possible to force a transition using rcutorture as a
destructive test.  This might (or might not) be adjusted later.

[ paulmck: Apply feedback from Kunwu Chan. ]

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutree.c | 31 ++++++++++++++++++-------------
 1 file changed, 18 insertions(+), 13 deletions(-)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index dc063eb49b0d..c611a7168c70 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -238,8 +238,10 @@ static bool init_srcu_struct_nodes(struct srcu_struct *ssp, gfp_t gfp_flags)
  * Initialize non-compile-time initialized fields, including the
  * associated srcu_node and srcu_data structures.  The is_static parameter
  * tells us that ->sda has already been wired up to srcu_data.
+ * The is_atomic parameter tells us that there is no reason to
+ * ever transition to big.
  */
-static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
+static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static, bool is_atomic)
 {
 	if (!is_static)
 		ssp->srcu_sup = kzalloc_obj(*ssp->srcu_sup);
@@ -267,7 +269,8 @@ static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
 	init_srcu_struct_data(ssp);
 	ssp->srcu_sup->srcu_gp_seq_needed_exp = SRCU_GP_SEQ_INITIAL_VAL;
 	ssp->srcu_sup->srcu_last_gp_end = ktime_get_mono_fast_ns();
-	if (READ_ONCE(ssp->srcu_sup->srcu_size_state) == SRCU_SIZE_SMALL && SRCU_SIZING_IS_INIT()) {
+	if (!is_atomic &&
+	    READ_ONCE(ssp->srcu_sup->srcu_size_state) == SRCU_SIZE_SMALL && SRCU_SIZING_IS_INIT()) {
 		if (!preemptible())
 			WRITE_ONCE(ssp->srcu_sup->srcu_size_state, SRCU_SIZE_ALLOC);
 		else if (init_srcu_struct_nodes(ssp, GFP_KERNEL))
@@ -301,7 +304,7 @@ __init_srcu_struct_common(struct srcu_struct *ssp, const char *name, struct lock
 	/* Don't re-initialize a lock while it is held. */
 	debug_check_no_locks_freed((void *)ssp, sizeof(*ssp));
 	lockdep_init_map(&ssp->dep_map, name, key, 0);
-	return init_srcu_struct_fields(ssp, false);
+	return init_srcu_struct_fields(ssp, false, false);
 }
 
 int init_srcu_struct_lockdep(struct srcu_struct *ssp, const char *name,
@@ -343,7 +346,7 @@ EXPORT_SYMBOL_GPL(__init_srcu_struct_fast_updown);
 int init_srcu_struct_generic(struct srcu_struct *ssp)
 {
 	ssp->srcu_reader_flavor = 0;
-	return init_srcu_struct_fields(ssp, false);
+	return init_srcu_struct_fields(ssp, false, false);
 }
 EXPORT_SYMBOL_GPL(init_srcu_struct_generic);
 
@@ -360,7 +363,7 @@ EXPORT_SYMBOL_GPL(init_srcu_struct_generic);
 int init_srcu_struct_fast(struct srcu_struct *ssp)
 {
 	ssp->srcu_reader_flavor = SRCU_READ_FLAVOR_FAST;
-	return init_srcu_struct_fields(ssp, false);
+	return init_srcu_struct_fields(ssp, false, false);
 }
 EXPORT_SYMBOL_GPL(init_srcu_struct_fast);
 
@@ -378,7 +381,7 @@ EXPORT_SYMBOL_GPL(init_srcu_struct_fast);
 int init_srcu_struct_fast_updown(struct srcu_struct *ssp)
 {
 	ssp->srcu_reader_flavor = SRCU_READ_FLAVOR_FAST_UPDOWN;
-	return init_srcu_struct_fields(ssp, false);
+	return init_srcu_struct_fields(ssp, false, false);
 }
 EXPORT_SYMBOL_GPL(init_srcu_struct_fast_updown);
 
@@ -470,9 +473,11 @@ static void raw_spin_lock_irqsave_ssp_contention(struct srcu_struct *ssp, unsign
  * done with compile-time initialization, so this check is added
  * to each update-side SRCU primitive.  Use ssp->lock, which -is-
  * compile-time initialized, to resolve races involving multiple
- * CPUs trying to garner first-use privileges.
+ * CPUs trying to garner first-use privileges.  The is_atomic
+ * parameter tells us that there will never be a reason to
+ * transition to big.
  */
-static void check_init_srcu_struct(struct srcu_struct *ssp)
+static void check_init_srcu_struct(struct srcu_struct *ssp, bool is_atomic)
 {
 	unsigned long flags;
 
@@ -484,7 +489,7 @@ static void check_init_srcu_struct(struct srcu_struct *ssp)
 		raw_spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
 		return;
 	}
-	init_srcu_struct_fields(ssp, true);
+	init_srcu_struct_fields(ssp, true, is_atomic);
 	raw_spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
 }
 
@@ -1279,7 +1284,7 @@ static bool srcu_should_expedite(struct srcu_struct *ssp)
 	unsigned long t;
 	unsigned long tlast;
 
-	check_init_srcu_struct(ssp);
+	check_init_srcu_struct(ssp, false);
 	/* If _lite() readers, don't do unsolicited expediting. */
 	if (this_cpu_read(ssp->sda->srcu_reader_flavor) & SRCU_READ_FLAVOR_SLOWGP)
 		return false;
@@ -1338,7 +1343,7 @@ static unsigned long srcu_gp_start_if_needed(struct srcu_struct *ssp,
 	struct srcu_node *sdp_mynode;
 	int ss_state;
 
-	check_init_srcu_struct(ssp);
+	check_init_srcu_struct(ssp, false);
 	/*
 	 * While starting a new grace period, make sure we are in an
 	 * SRCU read-side critical section so that the grace-period
@@ -1617,7 +1622,7 @@ static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm)
 	if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
 		return;
 	might_sleep();
-	check_init_srcu_struct(ssp);
+	check_init_srcu_struct(ssp, false);
 	init_completion(&rcu.completion);
 	init_rcu_head_on_stack(&rcu.head);
 	__call_srcu(ssp, &rcu.head, wakeme_after_rcu, do_norm);
@@ -1827,7 +1832,7 @@ void srcu_barrier(struct srcu_struct *ssp)
 	int idx;
 	unsigned long s;
 
-	check_init_srcu_struct(ssp);
+	check_init_srcu_struct(ssp, false);
 
 	/*
 	 * Register any deferred callbacks before snapshotting the sequence.  The
-- 
2.40.1

[PATCH 04/19] srcutree: Add an atomic Tree SRCU

[Paul E. McKenney]

Some dedicated srcu_struct users have read-side critical sections
which are short, never sleep, and never block on anything which may
itself depend on memory allocation — because they were, until
recently, spinlock or rwlock critical sections. For such a domain the
update side can safely wait for readers by spinning, from contexts
where sleeping is undesirable or the grace-period machinery's
latency (workqueue scheduling, jiffy-paced retries) dominates the
actual reader drain time.

But that is only safe if every reader keeps the promise. Make the
promise explicit and machine-checkable:

 - srcu_read_lock_atomic() / srcu_read_unlock_atomic() enter the
   usual (smp_mb-based) read-side critical section with preemption
   disabled and (except in hardirq, where it is redundant)
   non_block_start() armed, recording SRCU_READ_FLAVOR_ATOMIC in the
   per-CPU reader flavor. Disabling preemption enforces the
   no-sleeping promise on every configuration and bounds the section,
   so it is always running on some CPU; non_block_start() extends the
   enforcement to even potentially-sleeping calls on paths which
   happen not to block. The existing reader-flavor consistency checks
   complain about any mixing with other flavors.

 - synchronize_srcu_atomic() waits for all pre-existing readers by
   repeating the try_synchronize_srcu() both-epoch counter proof with
   cpu_relax() until it succeeds: no sleeping, no index flip, no
   grace-period sequence update, and therefore no interaction with
   concurrent call_srcu(), synchronize_srcu() or srcu_barrier(). It
   always provides the full grace-period guarantee: if the domain
   turns out to have had readers of any other flavor — a caller bug,
   since such a reader may be asleep and spinning on it would be
   unbounded — it complains and falls back to a real (sleeping) grace
   period internally, that being the only correct wait for a
   possibly-sleeping reader. The flavor mask is rechecked on every
   iteration so a first non-atomic reader appearing mid-spin takes
   the same path.

The immediate motivation is the proposed conversion of KVM's
gfn_to_pfn_cache to SRCU¹, whose mmu_notifier invalidation path drains
readers before the primary MMU zaps a page. With the readers declared
atomic, that drain becomes spin-only: no sleeping at all in the
notifier, bounded by the longest reader section, satisfying even the
strictest reading of the OOM-reaper non-blocking requirement without
needing to touch the non_block_start() annotation².

This commit also adds atomic-SRCU-specific initializers:
init_srcu_struct_atomic(), DEFINE_SRCU_ATOMIC(), and
DEFINE_STATIC_SRCU_ATOMIC().

This commit implements only Tree SRCU.  Tiny SRCU will follow.

¹ https://lore.kernel.org/all/20260811132237.102400-1-dwmw2@infradead.org/
² https://lore.kernel.org/all/20260812134934.GC662699@ziepe.ca/

[ paulmck: Apply Kunwu Chan feedback. ]

Co-developed-by: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
Assisted-by: Claude:claude-mythos-5
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 include/linux/srcu.h     |  81 ++++++++++++++++++--
 include/linux/srcutree.h |   9 ++-
 kernel/rcu/srcutree.c    | 154 +++++++++++++++++++++++++++++++++++++--
 3 files changed, 227 insertions(+), 17 deletions(-)

diff --git a/include/linux/srcu.h b/include/linux/srcu.h
index 7d9bc06df98d..3f232f2e0524 100644
--- a/include/linux/srcu.h
+++ b/include/linux/srcu.h
@@ -36,6 +36,8 @@ static inline int __init_srcu_struct(struct srcu_struct *ssp, const char *name,
 int __init_srcu_struct_fast(struct srcu_struct *ssp, const char *name, struct lock_class_key *key);
 int __init_srcu_struct_fast_updown(struct srcu_struct *ssp, const char *name,
 				   struct lock_class_key *key);
+int __init_srcu_struct_atomic(struct srcu_struct *ssp, const char *name,
+			      struct lock_class_key *key);
 #endif // #ifndef CONFIG_TINY_SRCU
 
 #define init_srcu_struct_fast(ssp) \
@@ -52,6 +54,13 @@ int __init_srcu_struct_fast_updown(struct srcu_struct *ssp, const char *name,
 	__init_srcu_struct_fast_updown((ssp), #ssp, &__srcu_key); \
 })
 
+#define init_srcu_struct_atomic(ssp) \
+({ \
+	static struct lock_class_key __srcu_key; \
+	\
+	__init_srcu_struct_atomic((ssp), #ssp, &__srcu_key); \
+})
+
 #define __SRCU_DEP_MAP_INIT(srcu_name)	.dep_map = { .name = #srcu_name },
 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 
@@ -64,6 +73,7 @@ static inline int __init_srcu_struct(struct srcu_struct *ssp, const char *name,
 #ifndef CONFIG_TINY_SRCU
 int init_srcu_struct_fast(struct srcu_struct *ssp);
 int init_srcu_struct_fast_updown(struct srcu_struct *ssp);
+int init_srcu_struct_atomic(struct srcu_struct *ssp);
 #endif // #ifndef CONFIG_TINY_SRCU
 
 #define __SRCU_DEP_MAP_INIT(srcu_name)
@@ -77,14 +87,18 @@ int init_srcu_struct_fast_updown(struct srcu_struct *ssp);
 })
 
 /* Values for SRCU Tree srcu_data ->srcu_reader_flavor, but also used by rcutorture. */
-#define SRCU_READ_FLAVOR_NORMAL		0x1		// srcu_read_lock().
-#define SRCU_READ_FLAVOR_NMI		0x2		// srcu_read_lock_nmisafe().
-//					0x4		// SRCU-lite is no longer with us.
-#define SRCU_READ_FLAVOR_FAST		0x4		// srcu_read_lock_fast(), also NMI-safe.
-#define SRCU_READ_FLAVOR_FAST_UPDOWN	0x8		// srcu_read_lock_fast_updown().
+#define SRCU_READ_FLAVOR_NORMAL		0x01		// srcu_read_lock().
+#define SRCU_READ_FLAVOR_NMI		0x02		// srcu_read_lock_nmisafe().
+//					0x04		// SRCU-lite is no longer with us.
+#define SRCU_READ_FLAVOR_FAST		0x04		// srcu_read_lock_fast(), also NMI-safe.
+#define SRCU_READ_FLAVOR_FAST_UPDOWN	0x08		// srcu_read_lock_fast_updown().
+#define SRCU_READ_FLAVOR_ATOMIC		0x10		// srcu_read_lock_atomic().
 #define SRCU_READ_FLAVOR_ALL		(SRCU_READ_FLAVOR_NORMAL | SRCU_READ_FLAVOR_NMI | \
-					 SRCU_READ_FLAVOR_FAST | SRCU_READ_FLAVOR_FAST_UPDOWN)
+					 SRCU_READ_FLAVOR_FAST | SRCU_READ_FLAVOR_FAST_UPDOWN | \
+					 SRCU_READ_FLAVOR_ATOMIC)
 						// All of the above.
+#define SRCU_READ_FLAVOR_PREDEF		(SRCU_READ_FLAVOR_FAST | SRCU_READ_FLAVOR_ATOMIC)
+						// Flavors special DEFINE_SRCU() flavors.
 #define SRCU_READ_FLAVOR_SLOWGP		(SRCU_READ_FLAVOR_FAST | SRCU_READ_FLAVOR_FAST_UPDOWN)
 						// Flavors requiring synchronize_rcu()
 						// instead of smp_mb().
@@ -102,6 +116,7 @@ void call_srcu(struct srcu_struct *ssp, struct rcu_head *head,
 		void (*func)(struct rcu_head *head));
 void cleanup_srcu_struct(struct srcu_struct *ssp);
 void synchronize_srcu(struct srcu_struct *ssp);
+void synchronize_srcu_atomic(struct srcu_struct *ssp);
 
 #define SRCU_GET_STATE_COMPLETED 0x1
 
@@ -306,6 +321,43 @@ static inline int srcu_read_lock(struct srcu_struct *ssp)
 	return retval;
 }
 
+/**
+ * srcu_read_lock_atomic - register a new reader promising an atomic section
+ * @ssp: srcu_struct in which to register the new reader.
+ *
+ * As srcu_read_lock(), but the caller promises that the read-side
+ * critical section never sleeps and never blocks on anything which
+ * may itself depend on memory allocation to make progress. Preemption
+ * is disabled for the duration, which both enforces that promise (any
+ * sleepable call in the section will splat on every configuration)
+ * and bounds the section so that the update side may spin rather
+ * than sleep when waiting for readers: see synchronize_srcu_atomic().
+ *
+ * The lock and matching srcu_read_unlock_atomic() must be invoked on
+ * the same CPU, from the same context; passing the return value to
+ * another task is not permitted for this flavor.
+ */
+static inline int srcu_read_lock_atomic(struct srcu_struct *ssp)
+	__acquires_shared(ssp)
+{
+	int retval;
+
+	preempt_disable();
+	/*
+	 * Arm might_sleep() to catch even a *potentially* sleeping call
+	 * in the section, not just an actual schedule: the atomic-domain
+	 * promise must hold on every path, contended or not. In hardirq
+	 * the annotation would land on the interrupted task; it is also
+	 * redundant there, so skip it.
+	 */
+	if (!in_hardirq())
+		non_block_start();
+	srcu_check_read_flavor(ssp, SRCU_READ_FLAVOR_ATOMIC);
+	retval = __srcu_read_lock(ssp);
+	srcu_lock_acquire(&ssp->dep_map);
+	return retval;
+}
+
 /**
  * srcu_read_lock_fast - register a new reader for an SRCU-protected structure.
  * @ssp: srcu_struct in which to register the new reader.
@@ -498,6 +550,23 @@ static inline void srcu_read_unlock(struct srcu_struct *ssp, int idx)
 	__srcu_read_unlock(ssp, idx);
 }
 
+/**
+ * srcu_read_unlock_atomic - unregister an atomic-section reader
+ * @ssp: srcu_struct from which to unregister the old reader.
+ * @idx: return value from corresponding srcu_read_lock_atomic().
+ */
+static inline void srcu_read_unlock_atomic(struct srcu_struct *ssp, int idx)
+	__releases_shared(ssp)
+{
+	WARN_ON_ONCE(idx & ~0x1);
+	srcu_check_read_flavor(ssp, SRCU_READ_FLAVOR_ATOMIC);
+	srcu_lock_release(&ssp->dep_map);
+	__srcu_read_unlock(ssp, idx);
+	if (!in_hardirq())
+		non_block_end();
+	preempt_enable();
+}
+
 /**
  * srcu_read_unlock_fast - unregister a old reader from an SRCU-protected structure.
  * @ssp: srcu_struct in which to unregister the old reader.
diff --git a/include/linux/srcutree.h b/include/linux/srcutree.h
index 1ce759fb7094..ad9d9658b0a2 100644
--- a/include/linux/srcutree.h
+++ b/include/linux/srcutree.h
@@ -80,6 +80,7 @@ struct srcu_usage {
 	struct mutex srcu_cb_mutex;		/* Serialize CB preparation. */
 	raw_spinlock_t __private lock;		/* Protect counters and size state. */
 	struct mutex srcu_gp_mutex;		/* Serialize GP work. */
+	atomic_t srcu_atomic_gp_flag;		/* Serialize atomic GP work. */
 	unsigned long srcu_gp_seq;		/* Grace-period seq #. */
 	unsigned long srcu_gp_seq_needed;	/* Latest gp_seq needed. */
 	unsigned long srcu_gp_seq_needed_exp;	/* Furthest future exp GP. */
@@ -229,14 +230,16 @@ struct srcu_struct {
 	is_static struct srcu_struct name =							\
 		__SRCU_STRUCT_INIT(name, name##_srcu_usage, name##_srcu_data, fast)
 #endif
-#define DEFINE_SRCU(name)		__DEFINE_SRCU(name, 0, /* not static */)
+#define DEFINE_SRCU(name)		__DEFINE_SRCU(name, 0, /* !static */)
 #define DEFINE_STATIC_SRCU(name)	__DEFINE_SRCU(name, 0, static)
-#define DEFINE_SRCU_FAST(name)		__DEFINE_SRCU(name, SRCU_READ_FLAVOR_FAST, /* not static */)
+#define DEFINE_SRCU_FAST(name)		__DEFINE_SRCU(name, SRCU_READ_FLAVOR_FAST, /* !static */)
 #define DEFINE_STATIC_SRCU_FAST(name)	__DEFINE_SRCU(name, SRCU_READ_FLAVOR_FAST, static)
 #define DEFINE_SRCU_FAST_UPDOWN(name)	__DEFINE_SRCU(name, SRCU_READ_FLAVOR_FAST_UPDOWN, \
-						      /* not static */)
+						      /* !static */)
 #define DEFINE_STATIC_SRCU_FAST_UPDOWN(name) \
 					__DEFINE_SRCU(name, SRCU_READ_FLAVOR_FAST_UPDOWN, static)
+#define DEFINE_SRCU_ATOMIC(name)	__DEFINE_SRCU(name, SRCU_READ_FLAVOR_ATOMIC, /* !static */)
+#define DEFINE_STATIC_SRCU_ATOMIC(name)	__DEFINE_SRCU(name, SRCU_READ_FLAVOR_ATOMIC, static)
 
 int __srcu_read_lock(struct srcu_struct *ssp) __acquires_shared(ssp);
 void synchronize_srcu_expedited(struct srcu_struct *ssp);
diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index c611a7168c70..570d068d1840 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -253,6 +253,7 @@ static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static, bool
 	ssp->srcu_sup->node = NULL;
 	mutex_init(&ssp->srcu_sup->srcu_cb_mutex);
 	mutex_init(&ssp->srcu_sup->srcu_gp_mutex);
+	atomic_set(&ssp->srcu_sup->srcu_atomic_gp_flag, 0);
 	ssp->srcu_sup->srcu_gp_seq = SRCU_GP_SEQ_INITIAL_VAL;
 	ssp->srcu_sup->srcu_barrier_seq = 0;
 	mutex_init(&ssp->srcu_sup->srcu_barrier_mutex);
@@ -330,6 +331,13 @@ int __init_srcu_struct_fast_updown(struct srcu_struct *ssp, const char *name,
 }
 EXPORT_SYMBOL_GPL(__init_srcu_struct_fast_updown);
 
+int __init_srcu_struct_atomic(struct srcu_struct *ssp, const char *name, struct lock_class_key *key)
+{
+	ssp->srcu_reader_flavor = SRCU_READ_FLAVOR_ATOMIC;
+	return __init_srcu_struct_common(ssp, name, key);
+}
+EXPORT_SYMBOL_GPL(__init_srcu_struct_atomic);
+
 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 
 /**
@@ -385,6 +393,26 @@ int init_srcu_struct_fast_updown(struct srcu_struct *ssp)
 }
 EXPORT_SYMBOL_GPL(init_srcu_struct_fast_updown);
 
+/**
+ * init_srcu_struct_atomic - initialize an atomic sleep-RCU structure
+ * @ssp: structure to initialize.
+ *
+ * Use this in place of DEFINE_SRCU_ATOMIC() and DEFINE_STATIC_SRCU_ATOMIC()
+ * for non-static srcu_struct structures that are to be passed to
+ * srcu_read_lock_atomic() and friends.  It is necessary to invoke this on a
+ * given srcu_struct before passing that srcu_struct to any other function.
+ * Each srcu_struct represents a separate domain of SRCU protection.
+ *
+ * And yes, we really are defining a sleepable RCU implementation that
+ * cannot sleep.  Strange universe we live in, isn't it?
+ */
+int init_srcu_struct_atomic(struct srcu_struct *ssp)
+{
+	ssp->srcu_reader_flavor = SRCU_READ_FLAVOR_ATOMIC;
+	return init_srcu_struct_fields(ssp, false, false);
+}
+EXPORT_SYMBOL_GPL(init_srcu_struct_atomic);
+
 #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 
 /*
@@ -802,7 +830,10 @@ void __srcu_check_read_flavor(struct srcu_struct *ssp, int read_flavor)
 	WARN_ON_ONCE(ssp->srcu_reader_flavor && read_flavor != ssp->srcu_reader_flavor);
 	WARN_ON_ONCE(old_read_flavor && ssp->srcu_reader_flavor &&
 		     old_read_flavor != ssp->srcu_reader_flavor);
-	WARN_ON_ONCE(read_flavor == SRCU_READ_FLAVOR_FAST && !ssp->srcu_reader_flavor);
+	WARN_ON_ONCE(!!(read_flavor & SRCU_READ_FLAVOR_PREDEF) &&
+		     read_flavor != ssp->srcu_reader_flavor);
+	WARN_ON_ONCE(!!(ssp->srcu_reader_flavor & SRCU_READ_FLAVOR_PREDEF) &&
+		     read_flavor != ssp->srcu_reader_flavor);
 	if (!old_read_flavor) {
 		old_read_flavor = cmpxchg(&sdp->srcu_reader_flavor, 0, read_flavor);
 		if (!old_read_flavor)
@@ -942,7 +973,7 @@ static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp
  * are initiating callback invocation.  This allows the ->srcu_have_cbs[]
  * array to have a finite number of elements.
  */
-static void srcu_gp_end(struct srcu_struct *ssp)
+static void srcu_gp_end(struct srcu_struct *ssp, bool is_atomic)
 {
 	unsigned long cbdelay = 1;
 	bool cbs;
@@ -958,7 +989,8 @@ static void srcu_gp_end(struct srcu_struct *ssp)
 	struct srcu_usage *sup = ssp->srcu_sup;
 
 	/* Prevent more than one additional grace period. */
-	mutex_lock(&sup->srcu_cb_mutex);
+	if (!is_atomic)
+		mutex_lock(&sup->srcu_cb_mutex);
 
 	/* End the current grace period. */
 	raw_spin_lock_irq_rcu_node(sup);
@@ -973,7 +1005,8 @@ static void srcu_gp_end(struct srcu_struct *ssp)
 	if (ULONG_CMP_LT(sup->srcu_gp_seq_needed_exp, gpseq))
 		WRITE_ONCE(sup->srcu_gp_seq_needed_exp, gpseq);
 	raw_spin_unlock_irq_rcu_node(sup);
-	mutex_unlock(&sup->srcu_gp_mutex);
+	if (!is_atomic)
+		mutex_unlock(&sup->srcu_gp_mutex);
 	/* A new grace period can start at this point.  But only one. */
 
 	/* Initiate callback invocation as needed. */
@@ -1018,13 +1051,15 @@ static void srcu_gp_end(struct srcu_struct *ssp)
 		}
 
 	/* Callback initiation done, allow grace periods after next. */
-	mutex_unlock(&sup->srcu_cb_mutex);
+	if (!is_atomic)
+		mutex_unlock(&sup->srcu_cb_mutex);
 
 	/* Start a new grace period if needed. */
 	raw_spin_lock_irq_rcu_node(sup);
 	gpseq = rcu_seq_current(&sup->srcu_gp_seq);
 	if (!rcu_seq_state(gpseq) &&
 	    ULONG_CMP_LT(gpseq, sup->srcu_gp_seq_needed)) {
+		WARN_ON_ONCE(ssp->srcu_reader_flavor & SRCU_READ_FLAVOR_ATOMIC);
 		srcu_gp_start(ssp);
 		raw_spin_unlock_irq_rcu_node(sup);
 		srcu_reschedule(ssp, 0);
@@ -1148,6 +1183,7 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 	/* If grace period not already in progress, start it. */
 	if (!WARN_ON_ONCE(rcu_seq_done(&sup->srcu_gp_seq, s)) &&
 	    rcu_seq_state(sup->srcu_gp_seq) == SRCU_STATE_IDLE) {
+		WARN_ON_ONCE(ssp->srcu_reader_flavor & SRCU_READ_FLAVOR_ATOMIC);
 		srcu_gp_start(ssp);
 
 		// And how can that list_add() in the "else" clause
@@ -1482,6 +1518,8 @@ static void srcu_do_enqueue(struct srcu_struct *ssp, struct rcu_head *rhp,
 static void __call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
 			rcu_callback_t func, bool do_norm)
 {
+	if (WARN_ON_ONCE(ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC))
+		return; // Leak the callback rather than corrupt SRCU state.
 	if (should_rcu_defer()) {
 		struct srcu_defer *sndp = this_cpu_ptr(&srcu_defer);
 		struct srcu_data *sdp;
@@ -1621,6 +1659,11 @@ static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm)
 
 	if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
 		return;
+	if (WARN_ON_ONCE(ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC)) {
+		// This works, and exposes a possible bug.
+		synchronize_srcu_atomic(ssp);
+		return;
+	}
 	might_sleep();
 	check_init_srcu_struct(ssp, false);
 	init_completion(&rcu.completion);
@@ -1741,10 +1784,21 @@ EXPORT_SYMBOL_GPL(get_state_synchronize_srcu);
  * period has elapsed in the meantime.  Unlike get_state_synchronize_srcu(),
  * this function also ensures that any needed SRCU grace period will be
  * started.  This convenience does come at a cost in terms of CPU overhead.
+ *
+ * This function cannot be used with atomic SRCU, which only has
+ * atomic grace periods.  Give a warning if someone tries, and return
+ * the same cookie that would have been returned, but refrain from
+ * messing up state by starting a grace period.  If someone somewhere
+ * somehow invokes synchronize_srcu_atomic(), passing this cookie to
+ * poll_state_synchronize_srcu() will return true.  If no one ever invokes
+ * synchronize_srcu_atomic(), too bad.
  */
 unsigned long start_poll_synchronize_srcu(struct srcu_struct *ssp)
 {
-	return srcu_gp_start_if_needed(ssp, NULL, true);
+	if (WARN_ON_ONCE(ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC))
+		return get_state_synchronize_srcu(ssp);
+	else
+		return srcu_gp_start_if_needed(ssp, NULL, true);
 }
 EXPORT_SYMBOL_GPL(start_poll_synchronize_srcu);
 
@@ -1833,6 +1887,12 @@ void srcu_barrier(struct srcu_struct *ssp)
 	unsigned long s;
 
 	check_init_srcu_struct(ssp, false);
+	if (WARN_ON_ONCE(ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC)) {
+		// There shouldn't be any callbacks for atomic SRCU,
+		// but just in case.
+		schedule_timeout_uninterruptible(HZ/10);
+		return;
+	}
 
 	/*
 	 * Register any deferred callbacks before snapshotting the sequence.  The
@@ -1904,6 +1964,9 @@ static void srcu_expedite_current_cb(struct rcu_head *rhp)
  * no current grace period, one might be created.  If the current grace
  * period is currently sleeping, that sleep will complete before expediting
  * will take effect.
+ *
+ * This function must not be invoked on srcu_struct structures that are
+ * used with srcu_read_lock_atomic() and synchronize_srcu_atomic().
  */
 void srcu_expedite_current(struct srcu_struct *ssp)
 {
@@ -1911,6 +1974,9 @@ void srcu_expedite_current(struct srcu_struct *ssp)
 	bool needcb = false;
 	struct srcu_data *sdp;
 
+	// Atomic SRCU has no callbacks, so there is nothing to expedite.
+	if (WARN_ON_ONCE(ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC))
+		return;
 	migrate_disable();
 	sdp = this_cpu_ptr(ssp->sda);
 	raw_spin_lock_irqsave_sdp_contention(sdp, &flags);
@@ -1978,8 +2044,10 @@ static void srcu_advance_state(struct srcu_struct *ssp, bool is_atomic)
 			return;
 		}
 		idx = rcu_seq_state(READ_ONCE(ssp->srcu_sup->srcu_gp_seq));
-		if (idx == SRCU_STATE_IDLE)
+		if (idx == SRCU_STATE_IDLE) {
+			WARN_ON_ONCE(ssp->srcu_reader_flavor & SRCU_READ_FLAVOR_ATOMIC);
 			srcu_gp_start(ssp);
+		}
 		raw_spin_unlock_irq_rcu_node(ssp->srcu_sup);
 		if (idx != SRCU_STATE_IDLE) {
 			if (!is_atomic)
@@ -2015,9 +2083,78 @@ static void srcu_advance_state(struct srcu_struct *ssp, bool is_atomic)
 			return; /* readers present, retry later. */
 		}
 		ssp->srcu_sup->srcu_n_exp_nodelay = 0;
-		srcu_gp_end(ssp);  /* Releases ->srcu_gp_mutex. */
+		srcu_gp_end(ssp, is_atomic);  /* Releases ->srcu_gp_mutex. */
+	}
+}
+
+/**
+ * synchronize_srcu_atomic - spin for prior SRCU read-side critical-section completion
+ * @ssp: srcu_struct with which to synchronize.
+ *
+ * Similar to synchronize_srcu(), but spins rather than blocking.
+ * Use only with srcu_read_lock_atomic() and srcu_read_unlock_atomic(),
+ * which are forbidden from voluntarily context switching.  If
+ * synchronize_srcu_atomic() is invoked from a more restrictive context
+ * (for example, interrupts disabled) for a given srcu_struct structure,
+ * then for that structure, all calls to both srcu_read_lock_atomic()
+ * and srcu_read_unlock_atomic() must be invoked from that same context,
+ * or one that is even more strict.
+ *
+ * If synchronize_srcu_atomic() is invoked on a given srcu_struct
+ * structure, then none of call_srcu(), synchronize_srcu(),
+ * synchronize_srcu_expedited(), start_poll_synchronize_srcu(),
+ * srcu_barrier(), or srcu_expedite_current() may be invoked on that
+ * same structure.
+ *
+ * Because synchronize_srcu_atomic() is even more expedited than is
+ * synchronize_srcu_expedited(), there is no expedited counterpart to
+ * this function.
+ */
+void synchronize_srcu_atomic(struct srcu_struct *ssp)
+{
+	unsigned long srcu_state;
+	struct srcu_usage *sup = ssp->srcu_sup;
+
+	// Initialize.	Either init_srcu_struct() was invoked or
+	// DEFINE_SRCU() or similar was used.  Therefore, no allocation
+	// will be done here.
+	check_init_srcu_struct(ssp, true);
+	srcu_check_read_flavor(ssp, SRCU_READ_FLAVOR_ATOMIC);
+
+	// Perhaps others will do our work for us.
+	srcu_state = get_state_synchronize_srcu(ssp);
+	while (atomic_read(&sup->srcu_atomic_gp_flag) ||
+	       atomic_xchg(&sup->srcu_atomic_gp_flag, 1)) {
+		if (poll_state_synchronize_srcu(ssp, srcu_state))
+			return;
+		cpu_relax();
+	}
+
+	// One last check for others doing our work for us under the lock.
+	raw_spin_lock_irq_rcu_node(sup);
+	if (poll_state_synchronize_srcu(ssp, srcu_state)) {
+		raw_spin_unlock_irq_rcu_node(sup);
+		atomic_set(&sup->srcu_atomic_gp_flag, 0);
+		return;
+	}
+
+	// OK, we really have to do it ourselves.  Start the grace period.
+	non_block_start();  // We must not voluntarily block!
+	smp_store_release(&sup->srcu_gp_seq_needed, srcu_state); // See srcu_funnel_gp_start().
+	ASSERT_EXCLUSIVE_WRITER(ssp->srcu_sup->srcu_gp_seq);
+	srcu_gp_start(ssp);
+	raw_spin_unlock_irq_rcu_node(sup);
+
+	// Wait for it to complete, helping it along.
+	while (!poll_state_synchronize_srcu(ssp, srcu_state)) {
+		cpu_relax();
+		srcu_advance_state(ssp, true);
 	}
+	ASSERT_EXCLUSIVE_WRITER(sup->srcu_atomic_gp_flag);
+	atomic_set_release(&sup->srcu_atomic_gp_flag, 0);
+	non_block_end();
 }
+EXPORT_SYMBOL_GPL(synchronize_srcu_atomic);
 
 /*
  * Invoke a limited number of SRCU callbacks that have passed through
@@ -2098,6 +2235,7 @@ static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay)
 		}
 	} else if (!rcu_seq_state(ssp->srcu_sup->srcu_gp_seq)) {
 		/* Outstanding request and no GP.  Start one. */
+		WARN_ON_ONCE(ssp->srcu_reader_flavor & SRCU_READ_FLAVOR_ATOMIC);
 		srcu_gp_start(ssp);
 	}
 	raw_spin_unlock_irq_rcu_node(ssp->srcu_sup);
-- 
2.40.1

[PATCH 05/19] srcutiny: Add an atomic Tiny SRCU

[Paul E. McKenney]

This commit adds the Tiny SRCU counterpart to Tree SRCU's
synchronize_srcu_atomic().  One might hope that this could be as trivial
as Tiny RCU's synchronize_rcu(), and there was a time when it would
have been.  But lazy preemption really can preempt an SRCU read-side
critical section, which means that synchronize_srcu_atomic() really must
be prepared to spin waiting for it.

This spinning currently consists of cond_resched_tasks_rcu_qs()
and cpu_relax().  It would be better to have some way of telling the
scheduler that there is nothing useful for us to do.  We cannot use
the traditional wait_event() approach because synchronize_srcu_atomic()
is not permitted to block.

[ paulmck: Apply Kunwu Chan feedback. ]

Co-developed-by: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
Assisted-by: Claude:claude-mythos-5
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 include/linux/srcutiny.h |  6 +++
 kernel/rcu/srcutiny.c    | 79 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 85 insertions(+)

diff --git a/include/linux/srcutiny.h b/include/linux/srcutiny.h
index 85b5de438450..47a368f945e3 100644
--- a/include/linux/srcutiny.h
+++ b/include/linux/srcutiny.h
@@ -19,6 +19,7 @@ struct srcu_struct {
 	short srcu_lock_nesting[2];	/* srcu_read_lock() nesting depth. */
 	u8 srcu_gp_running;		/* GP workqueue running? */
 	u8 srcu_gp_waiting;		/* GP waiting for readers? */
+	u8 srcu_atomic_gp_flag;		/* Serialize atomic GP work.*/
 	unsigned long srcu_idx;		/* Current reader array element in bit 0x2. */
 	unsigned long srcu_idx_max;	/* Furthest future srcu_idx request. */
 	struct swait_queue_head srcu_wq;
@@ -64,15 +65,20 @@ void srcu_defer_drain(struct irq_work *irq_work);
 #define DEFINE_SRCU_FAST_UPDOWN(name) DEFINE_SRCU(name)
 #define DEFINE_STATIC_SRCU_FAST_UPDOWN(name) \
 	static struct srcu_struct name = __SRCU_STRUCT_INIT(name, name, name, name)
+#define DEFINE_SRCU_ATOMIC(name) DEFINE_SRCU(name)
+#define DEFINE_STATIC_SRCU_ATOMIC(name) \
+	static struct srcu_struct name = __SRCU_STRUCT_INIT(name, name, name, name)
 
 // Dummy structure for srcu_notifier_head.
 struct srcu_usage { };
 #define __SRCU_USAGE_INIT(name) { }
 #define __init_srcu_struct_fast __init_srcu_struct
 #define __init_srcu_struct_fast_updown __init_srcu_struct
+#define __init_srcu_struct_atomic __init_srcu_struct
 #ifndef CONFIG_DEBUG_LOCK_ALLOC
 #define init_srcu_struct_fast init_srcu_struct
 #define init_srcu_struct_fast_updown init_srcu_struct
+#define init_srcu_struct_atomic init_srcu_struct
 #endif // #ifndef CONFIG_DEBUG_LOCK_ALLOC
 
 void synchronize_srcu(struct srcu_struct *ssp);
diff --git a/kernel/rcu/srcutiny.c b/kernel/rcu/srcutiny.c
index 32b37d63d58a..c6a2b74ae9d6 100644
--- a/kernel/rcu/srcutiny.c
+++ b/kernel/rcu/srcutiny.c
@@ -41,6 +41,7 @@ static int init_srcu_struct_fields(struct srcu_struct *ssp)
 	ssp->srcu_cb_tail = &ssp->srcu_cb_head;
 	ssp->srcu_gp_running = false;
 	ssp->srcu_gp_waiting = false;
+	ssp->srcu_atomic_gp_flag = 0;
 	ssp->srcu_idx = 0;
 	ssp->srcu_idx_max = 0;
 	INIT_WORK(&ssp->srcu_work, srcu_drive_gp);
@@ -339,6 +340,79 @@ void synchronize_srcu(struct srcu_struct *ssp)
 }
 EXPORT_SYMBOL_GPL(synchronize_srcu);
 
+/*
+ * synchronize_srcu_atomic - spinning grace period for atomic-reader domains
+ * @ssp: srcu_struct with which to synchronize.
+ *
+ * On !SMP this cannot spin: a reader observed mid-section is preempted
+ * or interrupted-out, and can only finish if we yield the CPU. But it
+ * is also never needed: an atomic-flavor reader (preemption disabled)
+ * cannot be observed mid-section from process context on the sole CPU.
+ * So a reader observed here has broken the atomic-domain promise, and
+ * the only correct wait for it is a real grace period.
+ *
+ * (Actual kernel-doc header is in Tree SRCU.)
+ */
+void synchronize_srcu_atomic(struct srcu_struct *ssp)
+{
+	int idx;
+	bool ret;
+	unsigned long srcu_state = get_state_synchronize_srcu(ssp);
+
+	srcu_lock_sync(&ssp->dep_map);
+
+	if (IS_ENABLED(CONFIG_PREEMPTION))
+		synchronize_rcu(); // Needed for RCU Tasks Trace to imply RCU grace period.
+				   // And in Tiny RCU, it is near zero cost and doesn't block.
+
+	// Usually, there will be no readers.
+	preempt_disable();  // Guard against lazy preemption and some other grace period.
+	ret = !READ_ONCE(ssp->srcu_lock_nesting[0]) && !READ_ONCE(ssp->srcu_lock_nesting[1]);
+	if (ret) {
+		WRITE_ONCE(ssp->srcu_idx_max, ssp->srcu_idx + 2);
+		WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 2);
+		preempt_enable();
+		return;
+	}
+
+	// Wait to drive a grace period or for someone else to do it
+	// for us while we are lazily preempted.
+	while (ssp->srcu_atomic_gp_flag) {
+		if (poll_state_synchronize_srcu(ssp, srcu_state)) {
+			preempt_enable();
+			return;
+		}
+		preempt_enable();
+		cpu_relax();
+		cond_resched_tasks_rcu_qs();
+		preempt_disable();
+	}
+	ssp->srcu_atomic_gp_flag = 1;
+	preempt_enable();
+
+	// We get here if a reader has been lazily preempted.
+	// First, wait for old readers, which are quite unlikely.
+	WRITE_ONCE(ssp->srcu_idx_max, get_state_synchronize_srcu(ssp));
+	idx = !(((READ_ONCE(ssp->srcu_idx) + 1) & 0x2) >> 1);
+	while (READ_ONCE(ssp->srcu_lock_nesting[idx])) {
+		cond_resched_tasks_rcu_qs();
+		cpu_relax();
+	}
+
+	// Next, flip the index and wait for the other group of readers.
+	WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1);
+	idx = !idx;
+	while (READ_ONCE(ssp->srcu_lock_nesting[idx])) {
+		cond_resched_tasks_rcu_qs();
+		cpu_relax();
+	}
+
+	// Finally, flip the index again for poll_state_synchronize_srcu().
+	WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1);
+	WARN_ON_ONCE(!poll_state_synchronize_srcu(ssp, srcu_state));
+}
+EXPORT_SYMBOL_GPL(synchronize_srcu_atomic);
+
 /* Register any deferred callbacks, then wait for all in-flight ones. */
 void srcu_barrier(struct srcu_struct *ssp)
 {
@@ -367,6 +441,11 @@ EXPORT_SYMBOL_GPL(get_state_synchronize_srcu);
  * The difference between this and get_state_synchronize_srcu() is that
  * this function ensures that the poll_state_synchronize_srcu() will
  * eventually return the value true.
+ *
+ * This function cannot be used with atomic SRCU, which only has
+ * atomic grace periods.  Doing so will silently corrupt internal
+ * SRCU state.  Tree SRCU has appropriate checking with splats,
+ * so please test with CONFIG_SMP=y as well as CONFIG_SMP=n.
  */
 unsigned long start_poll_synchronize_srcu(struct srcu_struct *ssp)
 {
-- 
2.40.1

[PATCH 06/19] rcutorture: Add support for testing synchronize_srcu_atomic()

[Paul E. McKenney]

This commit adds support for the value 0x10 for reader_flavor, which
specifies SRCU_READ_FLAVOR_ATOMIC, that is, srcu_read_lock_atomic(),
srcu_read_unlock_atomic(), and synchronize_srcu_atomic().

[ paulmck: Apply Kunwu Chan feedback. ]

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/rcutorture.c | 46 +++++++++++++++++++++++++++++++++++------
 1 file changed, 40 insertions(+), 6 deletions(-)

diff --git a/kernel/rcu/rcutorture.c b/kernel/rcu/rcutorture.c
index 4d1be2a49f01..182d47975efd 100644
--- a/kernel/rcu/rcutorture.c
+++ b/kernel/rcu/rcutorture.c
@@ -710,10 +710,21 @@ static struct rcu_torture_ops rcu_busted_ops = {
 DEFINE_STATIC_SRCU(srcu_ctl);
 DEFINE_STATIC_SRCU_FAST(srcu_ctlf);
 DEFINE_STATIC_SRCU_FAST_UPDOWN(srcu_ctlfud);
+DEFINE_STATIC_SRCU_ATOMIC(srcu_ctla);
 static struct srcu_struct srcu_ctld;
 static struct srcu_struct *srcu_ctlp = &srcu_ctl;
 static struct rcu_torture_ops srcud_ops;
 
+// Restrict APIs permitted for atomic SRCU.
+static void srcu_torture_init_forbidden_apis(void)
+{
+	cur_ops->call = NULL;
+	cur_ops->cb_barrier = NULL;
+	cur_ops->deferred_free = NULL;
+	cur_ops->exp_current = NULL;
+	cur_ops->start_gp_poll = NULL;
+}
+
 static void srcu_torture_init(void)
 {
 	rcu_sync_torture_init();
@@ -729,6 +740,11 @@ static void srcu_torture_init(void)
 		srcu_ctlp = &srcu_ctlfud;
 		VERBOSE_TOROUT_STRING("srcu_torture_init fast-up/down SRCU");
 	}
+	if (reader_flavor & SRCU_READ_FLAVOR_ATOMIC) {
+		srcu_ctlp = &srcu_ctla;
+		VERBOSE_TOROUT_STRING("srcu_torture_init atomic SRCU");
+		srcu_torture_init_forbidden_apis();
+	}
 }
 
 static void srcu_get_gp_data(int *flags, unsigned long *gp_seq)
@@ -766,6 +782,11 @@ static int srcu_torture_read_lock(void)
 		WARN_ON_ONCE(idx & ~0x1);
 		ret += idx << 3;
 	}
+	if (reader_flavor & SRCU_READ_FLAVOR_ATOMIC) {
+		idx = srcu_read_lock_atomic(srcu_ctlp);
+		WARN_ON_ONCE(idx & ~0x1);
+		ret += idx << 4;
+	}
 	return ret;
 }
 
@@ -785,7 +806,8 @@ srcu_read_delay(struct torture_random_state *rrsp, struct rt_read_seg *rtrsp)
 
 	delay = torture_random(rrsp) %
 		(nrealreaders * 2 * longdelay * uspertick);
-	if (!delay && !in_atomic() && !rcu_preempt_depth() && !irqs_disabled()) {
+	if (!delay && !in_atomic() && !rcu_preempt_depth() && !irqs_disabled() &&
+	    !(reader_flavor & SRCU_READ_FLAVOR_ATOMIC)) {
 		schedule_timeout_interruptible(longdelay);
 		rtrsp->rt_delay_jiffies = longdelay;
 	} else {
@@ -796,6 +818,8 @@ srcu_read_delay(struct torture_random_state *rrsp, struct rt_read_seg *rtrsp)
 static void srcu_torture_read_unlock(int idx)
 {
 	WARN_ON_ONCE((reader_flavor && (idx & ~reader_flavor)) || (!reader_flavor && (idx & ~0x1)));
+	if (reader_flavor & SRCU_READ_FLAVOR_ATOMIC)
+		srcu_read_unlock_atomic(srcu_ctlp, (idx & 0x10) >> 4);
 	if (reader_flavor & SRCU_READ_FLAVOR_FAST_UPDOWN)
 		srcu_read_unlock_fast_updown(srcu_ctlp,
 					     __srcu_ctr_to_ptr(srcu_ctlp, (idx & 0x8) >> 3));
@@ -875,7 +899,10 @@ static void srcu_torture_deferred_free(struct rcu_torture *rp)
 
 static void srcu_torture_synchronize(void)
 {
-	synchronize_srcu(srcu_ctlp);
+	if (reader_flavor & SRCU_READ_FLAVOR_ATOMIC)
+		synchronize_srcu_atomic(srcu_ctlp);
+	else
+		synchronize_srcu(srcu_ctlp);
 }
 
 static unsigned long srcu_torture_get_gp_state(void)
@@ -911,7 +938,10 @@ static void srcu_torture_stats(void)
 
 static void srcu_torture_synchronize_expedited(void)
 {
-	synchronize_srcu_expedited(srcu_ctlp);
+	if (reader_flavor & SRCU_READ_FLAVOR_ATOMIC)
+		synchronize_srcu_atomic(srcu_ctlp);
+	else
+		synchronize_srcu_expedited(srcu_ctlp);
 }
 
 static void srcu_torture_expedite_current(void)
@@ -969,6 +999,10 @@ static void srcud_torture_init(void)
 	} else if (reader_flavor & SRCU_READ_FLAVOR_FAST_UPDOWN) {
 		WARN_ON(init_srcu_struct_fast_updown(&srcu_ctld));
 		VERBOSE_TOROUT_STRING("srcud_torture_init fast-up/down SRCU");
+	} else if (reader_flavor & SRCU_READ_FLAVOR_ATOMIC) {
+		WARN_ON(init_srcu_struct_atomic(&srcu_ctld));
+		VERBOSE_TOROUT_STRING("srcud_torture_init atomic SRCU");
+		srcu_torture_init_forbidden_apis();
 	} else {
 		WARN_ON(init_srcu_struct(&srcu_ctld));
 	}
@@ -1749,7 +1783,7 @@ rcu_torture_writer(void *arg)
 		pr_alert("%s" TORTURE_FLAG " Waited %lu jiffies for boot to complete.\n",
 			 torture_type, jiffies - j);
 
-	if (IS_ENABLED(CONFIG_RCU_LAZY))
+	if (IS_ENABLED(CONFIG_RCU_LAZY) && cur_ops->call)
 		INIT_WORK_ONSTACK(&lazy_work, rcu_torture_writer_work);
 
 	do {
@@ -1944,7 +1978,7 @@ rcu_torture_writer(void *arg)
 				       !rcu_gp_is_normal();
 		}
 		rcu_torture_writer_state = RTWS_STUTTER;
-		if (IS_ENABLED(CONFIG_RCU_LAZY))
+		if (IS_ENABLED(CONFIG_RCU_LAZY) && cur_ops->call)
 			queue_work(system_percpu_wq, &lazy_work);
 		stutter_waited = stutter_wait("rcu_torture_writer");
 		if (stutter_waited &&
@@ -1977,7 +2011,7 @@ rcu_torture_writer(void *arg)
 			 " Dynamic grace-period expediting was disabled.\n",
 			 torture_type);
 
-	if (IS_ENABLED(CONFIG_RCU_LAZY)) {
+	if (IS_ENABLED(CONFIG_RCU_LAZY) && cur_ops->call) {
 		cancel_work_sync(&lazy_work);
 		destroy_work_on_stack(&lazy_work);
 	}
-- 
2.40.1

[PATCH 07/19] srcutree: Disable preemption across synchronize_srcu_atomic()

[Paul E. McKenney]

Because synchronize_srcu_atomic() cannot sleep, if a pair of them run
concurrently pinned to the same CPU, it is possible that one will spin
uselessly waiting for the other while at the same time preventing that
other from running.  This commit therefore disables preemption to prevent
this failure mode.

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutree.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 570d068d1840..61c2375ba2ec 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -2123,11 +2123,14 @@ void synchronize_srcu_atomic(struct srcu_struct *ssp)
 
 	// Perhaps others will do our work for us.
 	srcu_state = get_state_synchronize_srcu(ssp);
+	preempt_disable();
 	while (atomic_read(&sup->srcu_atomic_gp_flag) ||
 	       atomic_xchg(&sup->srcu_atomic_gp_flag, 1)) {
+		preempt_enable();
 		if (poll_state_synchronize_srcu(ssp, srcu_state))
 			return;
 		cpu_relax();
+		preempt_disable();
 	}
 
 	// One last check for others doing our work for us under the lock.
@@ -2135,6 +2138,7 @@ void synchronize_srcu_atomic(struct srcu_struct *ssp)
 	if (poll_state_synchronize_srcu(ssp, srcu_state)) {
 		raw_spin_unlock_irq_rcu_node(sup);
 		atomic_set(&sup->srcu_atomic_gp_flag, 0);
+		preempt_enable();
 		return;
 	}
 
@@ -2152,6 +2156,7 @@ void synchronize_srcu_atomic(struct srcu_struct *ssp)
 	}
 	ASSERT_EXCLUSIVE_WRITER(sup->srcu_atomic_gp_flag);
 	atomic_set_release(&sup->srcu_atomic_gp_flag, 0);
+	preempt_enable();
 	non_block_end();
 }
 EXPORT_SYMBOL_GPL(synchronize_srcu_atomic);
-- 
2.40.1

[PATCH 08/19] srcu: Use IRQ_WORK_INIT_HARD for srcu's irq_work

[Paul E. McKenney]

From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

The irq_work in srcu is used to schedule a delayed work. The work item
is not scheduled directly because it is not always possible wake a
thread directly.

On PREEMPT_RT the default irq_work is initialized with IRQ_WORK_LAZY and
is delayed to the irq_work thread. A system booted with the command line
"trace_event=…" will freeze during boot because the irq_work thread is
not yet deployed (and tracing uses synchronize_srcu() in
tp_rcu_cond_sync()).

The irq_work performs just a wakeup a thread, there is nothing wrong
with doing this from hardirq context on PREEMPT_RT.

Use IRQ_WORK_INIT_HARD for srcu's irq_work.

Fixes: 7c405fb3279b3 ("rcu: Use an intermediate irq_work to start process_srcu()")
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutree.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 61c2375ba2ec..6625420cfb43 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -259,7 +259,12 @@ static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static, bool
 	mutex_init(&ssp->srcu_sup->srcu_barrier_mutex);
 	atomic_set(&ssp->srcu_sup->srcu_barrier_cpu_cnt, 0);
 	INIT_DELAYED_WORK(&ssp->srcu_sup->work, process_srcu);
-	init_irq_work(&ssp->srcu_sup->irq_work, srcu_irq_work);
+	/*
+	 * trace events started on the command line require SRCU before
+	 * the irq_work kthread starts. Since all it does is a simple
+	 * wakeup, having it as a hard irq, even on PREEMPT_RT is fine.
+	 */
+	ssp->srcu_sup->irq_work = IRQ_WORK_INIT_HARD(srcu_irq_work);
 	ssp->srcu_sup->sda_is_static = is_static;
 	if (!is_static) {
 		ssp->sda = alloc_percpu(struct srcu_data);
-- 
2.40.1

[PATCH 09/19] srcu: Fix WARN_ON() for rcu_segcblist_n_cbs() in cleanup_srcu_struct()

[Paul E. McKenney]

From: Sunho Park <shpark061104@gmail.com>

The WARN_ON() added by commit 78a38cbf6f20 ("srcu: Queue sdp->work
when the delay timer is successfully deleted") uses rcu_segcblist_n_cbs()
to detect callbacks that srcu_barrier() failed to wait for. However, the
->len counter is decremented only at the end of srcu_invoke_callbacks(),
after the invoking loop has finished. Since srcu_barrier() can return
right after the barrier callback is invoked, cleanup_srcu_struct() can see
a non-zero n_cbs even though the cblist is already physically empty,
falsely triggering the WARN_ON() together with a still-pending delay_work
timer.

This can be triggered as follows, as seen in the syzbot report against
kvm_destroy_vm() -> cleanup_srcu_struct(&kvm->srcu):

1. call_srcu(&kvm->srcu, &bus->rcu, __free_bus) starts SRCU grace
   period GP1.

2. GP1 ends: a delay timer is armed and sdp->work is queued, but
   sdp->work has not run yet.

3. Another call_srcu(&kvm->srcu, &bus->rcu, __free_bus) call invokes
   srcu_segcblist_advance(), which moves the GP1 callback to
   RCU_DONE_TAIL, and starts SRCU grace period GP2.

4. srcu_barrier() is called. It queues its barrier callback after the
   GP2 callback and waits for srcu_invoke_callbacks() to invoke it.

5. GP2 ends: another delay timer is armed, and the sdp->work queued in
   step 2 begins to run. Its srcu_invoke_callbacks() call invokes
   srcu_segcblist_advance() again, moving the GP2 and barrier callbacks
   to RCU_DONE_TAIL as well, and then invokes all of them. However,
   rcu_segcblist_add_len(), which updates srcu_cblist's ->len, has not
   run yet at this point.

6. srcu_barrier() returns once its callback has been invoked, and
   cleanup_srcu_struct() starts running. It finds the delay timer
   armed in step 5 still pending and srcu_cblist's ->len still
   non-zero (because step 5 has not reached rcu_segcblist_add_len()
   yet), and WARN_ON() fires even though every callback has actually
   been invoked.

Use rcu_segcblist_empty(), which checks the actual head of the cblist,
instead of rcu_segcblist_n_cbs(), which checks the racy ->len counter.
Callbacks that have genuinely not been invoked yet still leave the list
non-empty, so the WARN_ON() still catches callers that skip srcu_barrier()
or queue callbacks after it.

Link: https://lore.kernel.org/rcu/e6350377085ddd85d6ef00d8e9a67bd50c762d3c@linux.dev/T/#t
Reported-by: syzbot+d4faf7db59e11f6fd1ab@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=d4faf7db59e11f6fd1ab
Fixes: 78a38cbf6f20 ("srcu: Queue sdp->work when the delay timer is successfully deleted")
Suggested-by: Zqiang <qiang.zhang@linux.dev>
Signed-off-by: Sunho Park <shpark061104@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutree.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 6625420cfb43..64081c8eacc8 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -787,7 +787,7 @@ void cleanup_srcu_struct(struct srcu_struct *ssp)
 		// Call srcu_barrier() before this cleanup_srcu_struct()
 		// to avoid triggering this WARN_ON().
 		if (WARN_ON(timer_delete_sync(&sdp->delay_work) &&
-			    rcu_segcblist_n_cbs(&sdp->srcu_cblist)) &&
+			    !rcu_segcblist_empty(&sdp->srcu_cblist)) &&
 		    rcu_cpu_beenfullyonline(sdp->cpu))
 			queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
 		flush_work(&sdp->work);
-- 
2.40.1

Re: [PATCH 09/19] srcu: Fix WARN_ON() for rcu_segcblist_n_cbs() in cleanup_srcu_struct()

[Sean Christopherson]

On Fri, Sep 18, 2026, Paul E. McKenney wrote:
> From: Sunho Park <shpark061104@gmail.com>
> 
> The WARN_ON() added by commit 78a38cbf6f20 ("srcu: Queue sdp->work
> when the delay timer is successfully deleted") uses rcu_segcblist_n_cbs()
> to detect callbacks that srcu_barrier() failed to wait for. However, the
> ->len counter is decremented only at the end of srcu_invoke_callbacks(),
> after the invoking loop has finished. Since srcu_barrier() can return
> right after the barrier callback is invoked, cleanup_srcu_struct() can see
> a non-zero n_cbs even though the cblist is already physically empty,
> falsely triggering the WARN_ON() together with a still-pending delay_work
> timer.
> 
> This can be triggered as follows, as seen in the syzbot report against
> kvm_destroy_vm() -> cleanup_srcu_struct(&kvm->srcu):
> 
> 1. call_srcu(&kvm->srcu, &bus->rcu, __free_bus) starts SRCU grace
>    period GP1.
> 
> 2. GP1 ends: a delay timer is armed and sdp->work is queued, but
>    sdp->work has not run yet.
> 
> 3. Another call_srcu(&kvm->srcu, &bus->rcu, __free_bus) call invokes
>    srcu_segcblist_advance(), which moves the GP1 callback to
>    RCU_DONE_TAIL, and starts SRCU grace period GP2.
> 
> 4. srcu_barrier() is called. It queues its barrier callback after the
>    GP2 callback and waits for srcu_invoke_callbacks() to invoke it.
> 
> 5. GP2 ends: another delay timer is armed, and the sdp->work queued in
>    step 2 begins to run. Its srcu_invoke_callbacks() call invokes
>    srcu_segcblist_advance() again, moving the GP2 and barrier callbacks
>    to RCU_DONE_TAIL as well, and then invokes all of them. However,
>    rcu_segcblist_add_len(), which updates srcu_cblist's ->len, has not
>    run yet at this point.
> 
> 6. srcu_barrier() returns once its callback has been invoked, and
>    cleanup_srcu_struct() starts running. It finds the delay timer
>    armed in step 5 still pending and srcu_cblist's ->len still
>    non-zero (because step 5 has not reached rcu_segcblist_add_len()
>    yet), and WARN_ON() fires even though every callback has actually
>    been invoked.
> 
> Use rcu_segcblist_empty(), which checks the actual head of the cblist,
> instead of rcu_segcblist_n_cbs(), which checks the racy ->len counter.
> Callbacks that have genuinely not been invoked yet still leave the list
> non-empty, so the WARN_ON() still catches callers that skip srcu_barrier()
> or queue callbacks after it.
> 
> Link: https://lore.kernel.org/rcu/e6350377085ddd85d6ef00d8e9a67bd50c762d3c@linux.dev/T/#t
> Reported-by: syzbot+d4faf7db59e11f6fd1ab@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=d4faf7db59e11f6fd1ab
> Fixes: 78a38cbf6f20 ("srcu: Queue sdp->work when the delay timer is successfully deleted")
> Suggested-by: Zqiang <qiang.zhang@linux.dev>
> Signed-off-by: Sunho Park <shpark061104@gmail.com>
> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
> ---

Is this going to be grabbed for 7.3?  It shows up relatively frequently in KVM.
Not so much that it's truly problematic/alarming, but enough that it definitely
stands out.

Tested-by: Sean Christopherson <seanjc@google.com>

Re: [PATCH 09/19] srcu: Fix WARN_ON() for rcu_segcblist_n_cbs() in cleanup_srcu_struct()

[Paul E. McKenney]

On Wed, Sep 30, 2026 at 05:27:19PM -0700, Sean Christopherson wrote:
> On Fri, Sep 18, 2026, Paul E. McKenney wrote:
> > From: Sunho Park <shpark061104@gmail.com>
> > 
> > The WARN_ON() added by commit 78a38cbf6f20 ("srcu: Queue sdp->work
> > when the delay timer is successfully deleted") uses rcu_segcblist_n_cbs()
> > to detect callbacks that srcu_barrier() failed to wait for. However, the
> > ->len counter is decremented only at the end of srcu_invoke_callbacks(),
> > after the invoking loop has finished. Since srcu_barrier() can return
> > right after the barrier callback is invoked, cleanup_srcu_struct() can see
> > a non-zero n_cbs even though the cblist is already physically empty,
> > falsely triggering the WARN_ON() together with a still-pending delay_work
> > timer.
> > 
> > This can be triggered as follows, as seen in the syzbot report against
> > kvm_destroy_vm() -> cleanup_srcu_struct(&kvm->srcu):
> > 
> > 1. call_srcu(&kvm->srcu, &bus->rcu, __free_bus) starts SRCU grace
> >    period GP1.
> > 
> > 2. GP1 ends: a delay timer is armed and sdp->work is queued, but
> >    sdp->work has not run yet.
> > 
> > 3. Another call_srcu(&kvm->srcu, &bus->rcu, __free_bus) call invokes
> >    srcu_segcblist_advance(), which moves the GP1 callback to
> >    RCU_DONE_TAIL, and starts SRCU grace period GP2.
> > 
> > 4. srcu_barrier() is called. It queues its barrier callback after the
> >    GP2 callback and waits for srcu_invoke_callbacks() to invoke it.
> > 
> > 5. GP2 ends: another delay timer is armed, and the sdp->work queued in
> >    step 2 begins to run. Its srcu_invoke_callbacks() call invokes
> >    srcu_segcblist_advance() again, moving the GP2 and barrier callbacks
> >    to RCU_DONE_TAIL as well, and then invokes all of them. However,
> >    rcu_segcblist_add_len(), which updates srcu_cblist's ->len, has not
> >    run yet at this point.
> > 
> > 6. srcu_barrier() returns once its callback has been invoked, and
> >    cleanup_srcu_struct() starts running. It finds the delay timer
> >    armed in step 5 still pending and srcu_cblist's ->len still
> >    non-zero (because step 5 has not reached rcu_segcblist_add_len()
> >    yet), and WARN_ON() fires even though every callback has actually
> >    been invoked.
> > 
> > Use rcu_segcblist_empty(), which checks the actual head of the cblist,
> > instead of rcu_segcblist_n_cbs(), which checks the racy ->len counter.
> > Callbacks that have genuinely not been invoked yet still leave the list
> > non-empty, so the WARN_ON() still catches callers that skip srcu_barrier()
> > or queue callbacks after it.
> > 
> > Link: https://lore.kernel.org/rcu/e6350377085ddd85d6ef00d8e9a67bd50c762d3c@linux.dev/T/#t
> > Reported-by: syzbot+d4faf7db59e11f6fd1ab@syzkaller.appspotmail.com
> > Closes: https://syzkaller.appspot.com/bug?extid=d4faf7db59e11f6fd1ab
> > Fixes: 78a38cbf6f20 ("srcu: Queue sdp->work when the delay timer is successfully deleted")
> > Suggested-by: Zqiang <qiang.zhang@linux.dev>
> > Signed-off-by: Sunho Park <shpark061104@gmail.com>
> > Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
> > ---
> 
> Is this going to be grabbed for 7.3?  It shows up relatively frequently in KVM.
> Not so much that it's truly problematic/alarming, but enough that it definitely
> stands out.
> 
> Tested-by: Sean Christopherson <seanjc@google.com>

I wasn't thinking in terms of doing so, figuring that the almost-here
merge window would be good enough.  But it does seem to apply cleanly.
Any other votes for pushing it into 7.3?  If I do so, it must be soon.

Either way, I will of course apply your Tested-by.

							Thanx, Paul

Re: [PATCH 09/19] srcu: Fix WARN_ON() for rcu_segcblist_n_cbs() in cleanup_srcu_struct()

[Paul E. McKenney]

On Wed, Sep 30, 2026 at 05:57:21PM -0700, Paul E. McKenney wrote:
> On Wed, Sep 30, 2026 at 05:27:19PM -0700, Sean Christopherson wrote:
> > On Fri, Sep 18, 2026, Paul E. McKenney wrote:
> > > From: Sunho Park <shpark061104@gmail.com>
> > > 
> > > The WARN_ON() added by commit 78a38cbf6f20 ("srcu: Queue sdp->work
> > > when the delay timer is successfully deleted") uses rcu_segcblist_n_cbs()
> > > to detect callbacks that srcu_barrier() failed to wait for. However, the
> > > ->len counter is decremented only at the end of srcu_invoke_callbacks(),
> > > after the invoking loop has finished. Since srcu_barrier() can return
> > > right after the barrier callback is invoked, cleanup_srcu_struct() can see
> > > a non-zero n_cbs even though the cblist is already physically empty,
> > > falsely triggering the WARN_ON() together with a still-pending delay_work
> > > timer.
> > > 
> > > This can be triggered as follows, as seen in the syzbot report against
> > > kvm_destroy_vm() -> cleanup_srcu_struct(&kvm->srcu):
> > > 
> > > 1. call_srcu(&kvm->srcu, &bus->rcu, __free_bus) starts SRCU grace
> > >    period GP1.
> > > 
> > > 2. GP1 ends: a delay timer is armed and sdp->work is queued, but
> > >    sdp->work has not run yet.
> > > 
> > > 3. Another call_srcu(&kvm->srcu, &bus->rcu, __free_bus) call invokes
> > >    srcu_segcblist_advance(), which moves the GP1 callback to
> > >    RCU_DONE_TAIL, and starts SRCU grace period GP2.
> > > 
> > > 4. srcu_barrier() is called. It queues its barrier callback after the
> > >    GP2 callback and waits for srcu_invoke_callbacks() to invoke it.
> > > 
> > > 5. GP2 ends: another delay timer is armed, and the sdp->work queued in
> > >    step 2 begins to run. Its srcu_invoke_callbacks() call invokes
> > >    srcu_segcblist_advance() again, moving the GP2 and barrier callbacks
> > >    to RCU_DONE_TAIL as well, and then invokes all of them. However,
> > >    rcu_segcblist_add_len(), which updates srcu_cblist's ->len, has not
> > >    run yet at this point.
> > > 
> > > 6. srcu_barrier() returns once its callback has been invoked, and
> > >    cleanup_srcu_struct() starts running. It finds the delay timer
> > >    armed in step 5 still pending and srcu_cblist's ->len still
> > >    non-zero (because step 5 has not reached rcu_segcblist_add_len()
> > >    yet), and WARN_ON() fires even though every callback has actually
> > >    been invoked.
> > > 
> > > Use rcu_segcblist_empty(), which checks the actual head of the cblist,
> > > instead of rcu_segcblist_n_cbs(), which checks the racy ->len counter.
> > > Callbacks that have genuinely not been invoked yet still leave the list
> > > non-empty, so the WARN_ON() still catches callers that skip srcu_barrier()
> > > or queue callbacks after it.
> > > 
> > > Link: https://lore.kernel.org/rcu/e6350377085ddd85d6ef00d8e9a67bd50c762d3c@linux.dev/T/#t
> > > Reported-by: syzbot+d4faf7db59e11f6fd1ab@syzkaller.appspotmail.com
> > > Closes: https://syzkaller.appspot.com/bug?extid=d4faf7db59e11f6fd1ab
> > > Fixes: 78a38cbf6f20 ("srcu: Queue sdp->work when the delay timer is successfully deleted")
> > > Suggested-by: Zqiang <qiang.zhang@linux.dev>
> > > Signed-off-by: Sunho Park <shpark061104@gmail.com>
> > > Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
> > > ---
> > 
> > Is this going to be grabbed for 7.3?  It shows up relatively frequently in KVM.
> > Not so much that it's truly problematic/alarming, but enough that it definitely
> > stands out.
> > 
> > Tested-by: Sean Christopherson <seanjc@google.com>
> 
> I wasn't thinking in terms of doing so, figuring that the almost-here
> merge window would be good enough.  But it does seem to apply cleanly.
> Any other votes for pushing it into 7.3?  If I do so, it must be soon.

I take it back.  Unless there are objections to my pushing this into v7.3
(instead of waiting until the v7.4 merge window), I will push it tomorrow
(Friday October 2), Pacific Time.

> Either way, I will of course apply your Tested-by.

And done!

							Thanx, Paul

[PATCH 10/19] srcutree: Warn if Tiny SRCU readers are preempted

[Paul E. McKenney]

The fastpath of synchronize_srcu_atomic() should always be taken
because readers always disable preemption.  Therefore, the fact that
synchronize_srcu_atomic() is running at all should mean that all readers
have ended.

But there are always bugs.  And responding to a usage bug with a
too-short SRCU grace period, and thus possibly corrupting memory
is at best a sadistic response so such a bug.  For this reason,
synchronize_srcu_atomic() explicitly waits for readers.  Except that
it does so silently, possibly failing to flag this bug.  This commit
therefore adds a splat if synchronize_srcu_atomic() fails to take the
early exit.

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutiny.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/kernel/rcu/srcutiny.c b/kernel/rcu/srcutiny.c
index c6a2b74ae9d6..99f8bfd98b04 100644
--- a/kernel/rcu/srcutiny.c
+++ b/kernel/rcu/srcutiny.c
@@ -375,6 +375,14 @@ void synchronize_srcu_atomic(struct srcu_struct *ssp)
 		return;
 	}
 
+	// Because readers disable preemption, we should never get here.
+	// However, a splat and some spinning is usually preferable to
+	// memory corruption due to a too-short grace period.  There is
+	// the possibility that this will hang if the preempted reader is
+	// not looked upon favorably by the scheduler, but this is still
+	// preferable to memory corruption.
+	WARN_ON_ONCE(1);
+
 	// Wait to drive a grace period or for someone else to do it
 	// for us while we are lazily preempted.
 	while (ssp->srcu_atomic_gp_flag) {
-- 
2.40.1

[PATCH 11/19] srcutree: Explicitly note DEFINE_SRCU() needs for srcu_barrier()

[Paul E. McKenney]

In the core kernel, an srcu_struct structure created by DEFINE_SRCU()
or friends lives as long as the kernel does, so there are no particular
requirements surrounding the end of that structure's life.  In contrast,
when DEFINE_SRCU() and friends are used within a module, the corresponding
srcu_struct structures' lifetimes end when that module exits.  This in
turn means that if such a structure was passed to call_srcu(), then
srcu_barrier() must be invoked after the last call_srcu() invocation
but before the module exits.

This commit therefore adds a comment stating this.

Reported-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 include/linux/srcutree.h | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/include/linux/srcutree.h b/include/linux/srcutree.h
index ad9d9658b0a2..93b76e543894 100644
--- a/include/linux/srcutree.h
+++ b/include/linux/srcutree.h
@@ -214,6 +214,12 @@ struct srcu_struct {
  * instead of smp_mb(), and given that the first (for example)
  * srcu_read_lock_fast() might race with the first synchronize_srcu(),
  * this different must be specified at initialization time.
+ *
+ * If you use any of the DEFINE_SRCU() functions within a module, the
+ * module-entry code will invoke init_srcu_struct() and the module-exit
+ * code will invoke cleanup_srcu_struct().  This means that if your module
+ * passes the resulting srcu_struct structure to call_srcu(), you will
+ * need to also pass this structure to srcu_barrier() prior to module exit.
  */
 #ifdef MODULE
 # define __DEFINE_SRCU(name, fast, is_static)							\
-- 
2.40.1

[PATCH 12/19] rcutorture: Add atomic-SRCU support to torture.sh

[Paul E. McKenney]

From: Kunwu Chan <kunwu.chan@gmail.com>

Add the --do-atomic-srcu argument to torture.sh, which runs the
SRCU-N, SRCU-P, and SRCU-T scenarios, thus covering both Tree SRCU
(SRCU-N and SRCU-P) and Tiny SRCU (SRCU-T), with
rcutorture.reader_flavor=0x10 appended to the boot parameters so
that it takes precedence over each scenario's own reader-flavor
setting.  This exercises srcu_read_lock_atomic(),
srcu_read_unlock_atomic(), and synchronize_srcu_atomic().

As with other torture.sh tests, the --do-kcsan argument runs a
KCSAN+PROVE_LOCKING variant of this test.

[ paulmck: Make --do-atomic-srcu be default-on. ]

Signed-off-by: Kunwu Chan <kunwu.chan@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 .../selftests/rcutorture/bin/torture.sh       | 24 +++++++++++++++++++
 1 file changed, 24 insertions(+)

diff --git a/tools/testing/selftests/rcutorture/bin/torture.sh b/tools/testing/selftests/rcutorture/bin/torture.sh
index f0083891ee81..e7b578ec152f 100755
--- a/tools/testing/selftests/rcutorture/bin/torture.sh
+++ b/tools/testing/selftests/rcutorture/bin/torture.sh
@@ -68,6 +68,7 @@ do_clocksourcewd="${ifnotaarch64}"
 do_rt=yes
 do_rcutasksflavors="${ifnotaarch64}" # FIXME: Back to "yes" when SMP=n auto-avoided
 do_srcu_lockdep=yes
+do_atomic_srcu=yes
 do_rcu_rust=no
 
 # doyesno - Helper function for yes/no arguments
@@ -103,6 +104,7 @@ usage () {
 	echo "       --do-rcu-rust / --do-no-rcu-rust / --no-rcu-rust"
 	echo "       --do-scftorture / --do-no-scftorture / --no-scftorture"
 	echo "       --do-srcu-lockdep / --do-no-srcu-lockdep / --no-srcu-lockdep"
+	echo "       --do-atomic-srcu / --do-no-atomic-srcu / --no-atomic-srcu"
 	echo "       --duration [ <minutes> | <hours>h | <days>d ]"
 	echo "       --guest-cpu-limit N"
 	echo "       --kcsan-kmake-arg kernel-make-arguments"
@@ -148,6 +150,7 @@ do
 		do_kcsan=yes
 		do_clocksourcewd="${ifnotaarch64}"
 		do_srcu_lockdep=yes
+		do_atomic_srcu=yes
 		;;
 	--do-allmodconfig|--do-no-allmodconfig|--no-allmodconfig)
 		do_allmodconfig=`doyesno "$1" --do-allmodconfig`
@@ -183,6 +186,7 @@ do
 		do_kcsan=no
 		do_clocksourcewd=no
 		do_srcu_lockdep=no
+		do_atomic_srcu=no
 		;;
 	--do-normal|--do-norm|--do-no-normal|--do-no-norm|--no-normal|--no-norm)
 		do_normal=`doyesno "$1" --do-normal`
@@ -212,6 +216,9 @@ do
 	--do-srcu-lockdep|--do-no-srcu-lockdep|--no-srcu-lockdep)
 		do_srcu_lockdep=`doyesno "$1" --do-srcu-lockdep`
 		;;
+	--do-atomic-srcu|--do-no-atomic-srcu|--no-atomic-srcu)
+		do_atomic_srcu=`doyesno "$1" --do-atomic-srcu`
+		;;
 	--duration)
 		checkarg --duration "(minutes)" $# "$2" '^[0-9][0-9]*\(m\|h\|d\|\)$' '^error'
 		mult=1
@@ -497,6 +504,23 @@ then
 	torture_set "rcutorture" tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration "$duration_rcutorture" --configs "$configs_rcutorture" --trust-make
 fi
 
+# Test atomic SRCU across Tree SRCU (SRCU-N and SRCU-P) and Tiny SRCU
+# (SRCU-T).  The reader flavor selects srcu_read_lock_atomic() and
+# synchronize_srcu_atomic().  Tiny SRCU requires SMP=n, which aarch64
+# does not support.
+if test "$do_atomic_srcu" = "yes"
+then
+	torture_bootargs="rcutorture.reader_flavor=0x10"
+	configs_atomic_srcu="SRCU-N SRCU-P"
+	if test "$ifnotaarch64" = yes
+	then
+		configs_atomic_srcu="$configs_atomic_srcu SRCU-T"
+	fi
+	torture_set "atomic-srcu" tools/testing/selftests/rcutorture/bin/kvm.sh \
+		--allcpus --duration "$duration_rcutorture" \
+		--configs "$configs_atomic_srcu" --trust-make
+fi
+
 if test "$do_locktorture" = "yes"
 then
 	torture_bootargs="torture.disable_onoff_at_boot"
-- 
2.40.1

[PATCH 13/19] srcutree: Add reader-free fastpath to synchronize_srcu_atomic()

[Paul E. McKenney]

From: Kunwu Chan <kunwu.chan@gmail.com>

synchronize_srcu_atomic() is restricted to srcu_read_lock_atomic() and
srcu_read_unlock_atomic(), whose read-side critical sections disable
preemption.  In the common case where there are no readers at all, the
grace period therefore need not do the index flip.  Add a fastpath
that sums both ranks of the per-CPU ->srcu_ctrs[] counters and, if the
lock counts match the unlock counts on both ranks, ends the grace
period immediately, skipping the srcu_advance_state() scans, mirroring
the similar Tiny SRCU fastpath.

Correctness requires the counter-sum proof to follow the grace-period
anchor written by srcu_gp_start(); placing it before the anchor could
let this grace period miss a pre-existing reader and return without
waiting for it.  The smp_mb() between the unlock and lock sums pairs
with the smp_mb() in __srcu_read_lock().  The grace period is ended
manually under ->lock and ->srcu_atomic_gp_flag.

In theory, this is slower than David Woodhouse's earlier patch, but
David's measurements showed that the performance was close enough
that it makes sense to keep the get_state_synchronize_srcu() and
poll_state_synchronize_srcu semantics.

Link: https://lore.kernel.org/all/20260907075829.2073224-4-kunwu.chan@linux.dev/
Link: https://lore.kernel.org/all/0f4dea21bac43685d3286df329401177e4452b36.camel@infradead.org/
Link: https://lore.kernel.org/all/0d4af6318ac67486858be1df8d436147b444a2d2.camel@infradead.org/
Signed-off-by: Kunwu Chan <kunwu.chan@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Cc: David Woodhouse <dwmw2@infradead.org>
---
 kernel/rcu/srcutree.c | 48 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 48 insertions(+)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 64081c8eacc8..93b8d1088abe 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -2119,6 +2119,8 @@ void synchronize_srcu_atomic(struct srcu_struct *ssp)
 {
 	unsigned long srcu_state;
 	struct srcu_usage *sup = ssp->srcu_sup;
+	unsigned long rdm0, rdm1;
+	unsigned long unlocks0, unlocks1;
 
 	// Initialize.	Either init_srcu_struct() was invoked or
 	// DEFINE_SRCU() or similar was used.  Therefore, no allocation
@@ -2154,6 +2156,52 @@ void synchronize_srcu_atomic(struct srcu_struct *ssp)
 	srcu_gp_start(ssp);
 	raw_spin_unlock_irq_rcu_node(sup);
 
+	//
+	// Fastpath:  If there are no readers at all, neither grace-period
+	// scan need wait, so both can be satisfied at once without doing
+	// the index flip.  The counter-sum proof is the same as that of
+	// srcu_readers_active_idx_check(), but spanning both indices.
+	// Atomic SRCU guarantees that all readers are of
+	// SRCU_READ_FLAVOR_ATOMIC, so the SLOWGP check never triggers and
+	// the ->srcu_reader_flavor masks returned by
+	// srcu_readers_unlock_idx() are unused.
+	//
+	// This proof must follow the grace-period anchor written by the
+	// srcu_gp_start() above, never precede it.  With the anchor first,
+	// a reader whose lock increment is missed by the sums below cannot
+	// have incremented its lock counter before the anchor, and therefore
+	// cannot be a pre-existing reader of this grace period.  Placing the
+	// proof before the anchor would let this grace period miss a
+	// pre-existing reader and return without waiting for it.
+	//
+	// The smp_mb() pairs with the smp_mb() in __srcu_read_lock()
+	// (store-buffering pattern), which guarantees that a lock is always
+	// counted if the corresponding unlock is counted, the same
+	// memory-ordering guarantee as is provided by
+	// srcu_readers_active_idx_check().
+	//
+	unlocks0 = srcu_readers_unlock_idx(ssp, 0, &rdm0);
+	unlocks1 = srcu_readers_unlock_idx(ssp, 1, &rdm1);
+	smp_mb(); /* A */
+	if (srcu_readers_lock_idx(ssp, 0, false, unlocks0) &&
+	    srcu_readers_lock_idx(ssp, 1, false, unlocks1)) {
+		// No readers, so end this grace period manually, skipping
+		// the index flip.  Advancing the sequence number via
+		// rcu_seq_start() in srcu_gp_start() above and rcu_seq_end()
+		// below keeps get_state_synchronize_srcu() and
+		// poll_state_synchronize_srcu() working, all under ->lock
+		// and ->srcu_atomic_gp_flag, which excludes concurrent
+		// sequence-number updates.
+		raw_spin_lock_irq_rcu_node(sup);
+		rcu_seq_end(&sup->srcu_gp_seq);
+		raw_spin_unlock_irq_rcu_node(sup);
+		WARN_ON_ONCE(!poll_state_synchronize_srcu(ssp, srcu_state));
+		atomic_set_release(&sup->srcu_atomic_gp_flag, 0);
+		preempt_enable();
+		non_block_end();
+		return;
+	}
+
 	// Wait for it to complete, helping it along.
 	while (!poll_state_synchronize_srcu(ssp, srcu_state)) {
 		cpu_relax();
-- 
2.40.1

[PATCH 14/19] srcutree: Skip callback scheduling for atomic SRCU grace periods

[Paul E. McKenney]

From: Kunwu Chan <kunwu.chan@gmail.com>

call_srcu() is forbidden on atomic srcu_struct, so srcu_gp_end() never
has callbacks to invoke for them.  Yet it schedules callback invocation,
which for atomic SRCU's SRCU_SIZE_SMALL state arms the boot CPU's
->delay_work timer every grace period, only for srcu_invoke_callbacks()
to find nothing to do.

Skip this for atomic SRCU.

Signed-off-by: Kunwu Chan <kunwu.chan@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutree.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 93b8d1088abe..0954486880c7 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -1016,10 +1016,10 @@ static void srcu_gp_end(struct srcu_struct *ssp, bool is_atomic)
 
 	/* Initiate callback invocation as needed. */
 	ss_state = smp_load_acquire(&sup->srcu_size_state);
-	if (ss_state < SRCU_SIZE_WAIT_BARRIER) {
+	if (!is_atomic && ss_state < SRCU_SIZE_WAIT_BARRIER) {
 		srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, get_boot_cpu_id()),
 					cbdelay);
-	} else {
+	} else if (!is_atomic) {
 		idx = rcu_seq_ctr(gpseq) % ARRAY_SIZE(snp->srcu_have_cbs);
 		srcu_for_each_node_breadth_first(ssp, snp) {
 			raw_spin_lock_irq_rcu_node(snp);
-- 
2.40.1

[PATCH 15/19] srcutree: Remove srcu_barrier() sleep for atomic SRCU

[Paul E. McKenney]

From: Kunwu Chan <kunwu.chan@gmail.com>

The atomic-SRCU path in srcu_barrier() sleeps for 100 milliseconds just
in case there are callbacks to wait for.  But call_srcu() refuses
atomic SRCU with a WARN_ON_ONCE() before reaching the deferred-enqueue
path, so there can be no callbacks, deferred or otherwise.

Drop the sleep.

Signed-off-by: Kunwu Chan <kunwu.chan@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutree.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 0954486880c7..82e61421142e 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -1892,12 +1892,9 @@ void srcu_barrier(struct srcu_struct *ssp)
 	unsigned long s;
 
 	check_init_srcu_struct(ssp, false);
-	if (WARN_ON_ONCE(ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC)) {
-		// There shouldn't be any callbacks for atomic SRCU,
-		// but just in case.
-		schedule_timeout_uninterruptible(HZ/10);
+	// Atomic SRCU has no callbacks, so there is nothing to wait on.
+	if (WARN_ON_ONCE(ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC))
 		return;
-	}
 
 	/*
 	 * Register any deferred callbacks before snapshotting the sequence.  The
-- 
2.40.1

[PATCH 16/19] srcu: Restrict atomic-SRCU non_block annotation to task context

[Paul E. McKenney]

From: Kunwu Chan <kunwu.chan@gmail.com>

srcu_read_lock_atomic() and srcu_read_unlock_atomic() arm and disarm
might_sleep() checks via non_block_start()/non_block_end(), skipping
hardirq so the update does not land on the interrupted task's
->non_block_count.

Inline softirqs run on the interrupted task's stack as well, so a
timer callback running atomic-SRCU readers races with the interrupted
task's own ->non_block_count updates, as KCSAN reports:

	BUG: KCSAN: data-race in srcu_torture_read_lock / srcu_torture_read_unlock

	write to 0xffffa00f818ea418 of 4 bytes by interrupt on cpu 0:
		srcu_torture_read_lock+0x422/0x470
		rcutorture_one_extend+0xdc/0x600
		rcu_torture_one_read+0xd1/0x330
		rcu_torture_timer+0x75/0x140
		call_timer_fn+0xe6/0x2f0
		...
		run_timer_softirq+0xb7/0x130
		handle_softirqs+0xfc/0x3f0
		__irq_exit_rcu+0x8e/0x100

Use in_task() so the annotation is applied only in task context; it
is redundant elsewhere because might_sleep() already warns about
sleeping from atomic context.

Signed-off-by: Kunwu Chan <kunwu.chan@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 include/linux/srcu.h | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/include/linux/srcu.h b/include/linux/srcu.h
index 3f232f2e0524..1a8a465a5650 100644
--- a/include/linux/srcu.h
+++ b/include/linux/srcu.h
@@ -346,11 +346,13 @@ static inline int srcu_read_lock_atomic(struct srcu_struct *ssp)
 	/*
 	 * Arm might_sleep() to catch even a *potentially* sleeping call
 	 * in the section, not just an actual schedule: the atomic-domain
-	 * promise must hold on every path, contended or not. In hardirq
-	 * the annotation would land on the interrupted task; it is also
+	 * promise must hold on every path, contended or not. In hardirq,
+	 * softirq, or NMI the annotation would land on the interrupted
+	 * task, and can also result in data races against that task's
+	 * own non_block_start()/non_block_end() invocations; it is also
 	 * redundant there, so skip it.
 	 */
-	if (!in_hardirq())
+	if (in_task())
 		non_block_start();
 	srcu_check_read_flavor(ssp, SRCU_READ_FLAVOR_ATOMIC);
 	retval = __srcu_read_lock(ssp);
@@ -562,7 +564,7 @@ static inline void srcu_read_unlock_atomic(struct srcu_struct *ssp, int idx)
 	srcu_check_read_flavor(ssp, SRCU_READ_FLAVOR_ATOMIC);
 	srcu_lock_release(&ssp->dep_map);
 	__srcu_read_unlock(ssp, idx);
-	if (!in_hardirq())
+	if (in_task())
 		non_block_end();
 	preempt_enable();
 }
-- 
2.40.1

[PATCH 17/19] srcutree: Make init_srcu_struct_atomic() prevent transition to big

[Paul E. McKenney]

From: Kunwu Chan <kunwu.chan@gmail.com>

The is_atomic parameter of init_srcu_struct_fields() exists so that
atomic SRCU never transitions to big, but neither
init_srcu_struct_atomic() nor its lockdep counterpart
__init_srcu_struct_atomic() sets it.

On systems where srcutree.convert_to_big selects SRCU_SIZING_INIT,
this needlessly allocates a full srcu_node combining tree for any
dynamically initialized atomic srcu_struct, despite atomic SRCU
having neither callbacks nor srcu_barrier() operations.

Pass true from both atomic entry points, adding an is_atomic parameter
to __init_srcu_struct_common().

Signed-off-by: Kunwu Chan <kunwu.chan@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutree.c | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 82e61421142e..4e9a0b8ea344 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -305,26 +305,27 @@ static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static, bool
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 
 static int
-__init_srcu_struct_common(struct srcu_struct *ssp, const char *name, struct lock_class_key *key)
+__init_srcu_struct_common(struct srcu_struct *ssp, const char *name,
+			  struct lock_class_key *key, bool is_atomic)
 {
 	/* Don't re-initialize a lock while it is held. */
 	debug_check_no_locks_freed((void *)ssp, sizeof(*ssp));
 	lockdep_init_map(&ssp->dep_map, name, key, 0);
-	return init_srcu_struct_fields(ssp, false, false);
+	return init_srcu_struct_fields(ssp, false, is_atomic);
 }
 
 int init_srcu_struct_lockdep(struct srcu_struct *ssp, const char *name,
 			     struct lock_class_key *key)
 {
 	ssp->srcu_reader_flavor = 0;
-	return __init_srcu_struct_common(ssp, name, key);
+	return __init_srcu_struct_common(ssp, name, key, false);
 }
 EXPORT_SYMBOL_GPL(init_srcu_struct_lockdep);
 
 int __init_srcu_struct_fast(struct srcu_struct *ssp, const char *name, struct lock_class_key *key)
 {
 	ssp->srcu_reader_flavor = SRCU_READ_FLAVOR_FAST;
-	return __init_srcu_struct_common(ssp, name, key);
+	return __init_srcu_struct_common(ssp, name, key, false);
 }
 EXPORT_SYMBOL_GPL(__init_srcu_struct_fast);
 
@@ -332,14 +333,14 @@ int __init_srcu_struct_fast_updown(struct srcu_struct *ssp, const char *name,
 				   struct lock_class_key *key)
 {
 	ssp->srcu_reader_flavor = SRCU_READ_FLAVOR_FAST_UPDOWN;
-	return __init_srcu_struct_common(ssp, name, key);
+	return __init_srcu_struct_common(ssp, name, key, false);
 }
 EXPORT_SYMBOL_GPL(__init_srcu_struct_fast_updown);
 
 int __init_srcu_struct_atomic(struct srcu_struct *ssp, const char *name, struct lock_class_key *key)
 {
 	ssp->srcu_reader_flavor = SRCU_READ_FLAVOR_ATOMIC;
-	return __init_srcu_struct_common(ssp, name, key);
+	return __init_srcu_struct_common(ssp, name, key, true);
 }
 EXPORT_SYMBOL_GPL(__init_srcu_struct_atomic);
 
@@ -414,7 +415,7 @@ EXPORT_SYMBOL_GPL(init_srcu_struct_fast_updown);
 int init_srcu_struct_atomic(struct srcu_struct *ssp)
 {
 	ssp->srcu_reader_flavor = SRCU_READ_FLAVOR_ATOMIC;
-	return init_srcu_struct_fields(ssp, false, false);
+	return init_srcu_struct_fields(ssp, false, true);
 }
 EXPORT_SYMBOL_GPL(init_srcu_struct_atomic);
 
-- 
2.40.1

[PATCH 18/19] srcutree: Don't transition atomic SRCU to big in srcu_gp_end()

[Paul E. McKenney]

From: Kunwu Chan <kunwu.chan@gmail.com>

Atomic SRCU must remain in the small size state. Warn if this
invariant is violated and avoid transitioning to big in that case.

[ paulmck: Folded "if" onto one line for 100-character limit. ]

Signed-off-by: Kunwu Chan <kunwu.chan@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Reviewed-by: Bradley Morgan <brads@mainlining.org>
---
 kernel/rcu/srcutree.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 4e9a0b8ea344..6a05bd8a6f39 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -1073,8 +1073,10 @@ static void srcu_gp_end(struct srcu_struct *ssp, bool is_atomic)
 		raw_spin_unlock_irq_rcu_node(sup);
 	}
 
-	/* Transition to big if needed. */
-	if (ss_state != SRCU_SIZE_SMALL && ss_state != SRCU_SIZE_BIG) {
+	/* Transition to big if needed, but never for atomic SRCU. */
+	if (ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC && ss_state != SRCU_SIZE_SMALL) {
+		WARN_ON_ONCE(1);
+	} else if (ss_state != SRCU_SIZE_SMALL && ss_state != SRCU_SIZE_BIG) {
 		if (ss_state == SRCU_SIZE_ALLOC)
 			init_srcu_struct_nodes(ssp, GFP_KERNEL);
 		else
-- 
2.40.1

[PATCH 19/19] srcutree: Skip torture to-big transition for atomic SRCU

[Paul E. McKenney]

From: Kunwu Chan <kunwu.chan@gmail.com>

Atomic SRCU does not use the srcu_node combining tree. Exclude it
from the SRCU_SIZING_IS_TORTURE() transition in
srcu_torture_stats_print(), and warn if such a transition is requested.

Signed-off-by: Kunwu Chan <kunwu.chan@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/srcutree.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 6a05bd8a6f39..6e37d53b412e 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -2422,8 +2422,10 @@ void srcu_torture_stats_print(struct srcu_struct *ssp, char *tt, char *tf)
 		}
 		pr_cont(" T(%ld,%ld)\n", s0, s1);
 	}
-	if (SRCU_SIZING_IS_TORTURE())
-		srcu_transition_to_big(ssp);
+	if (SRCU_SIZING_IS_TORTURE()) {
+		if (!WARN_ON_ONCE(ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC))
+			srcu_transition_to_big(ssp);
+	}
 }
 EXPORT_SYMBOL_GPL(srcu_torture_stats_print);
 
-- 
2.40.1