Merge tag 'rcu.release.v6.13' of git://git.kernel.org/pub/scm/linux/kernel/git/rcu/linux

Pull RCU updates from Frederic Weisbecker:
 "SRCU:

   - Introduction of the new SRCU-lite flavour with a new pair of
     srcu_read_[un]lock_lite() APIs. In practice the read side using
     this flavour becomes lighter by removing a full memory barrier on
     LOCK and a full memory barrier on UNLOCK. This comes at the expense
     of a higher latency write side with two (in the best case of a
     snaphot of unused read-sides) or more RCU grace periods on the
     update side which now assumes by itself the whole full ordering
     guarantee against the LOCK/UNLOCK counters on both indexes, along
     with the accesses performed inside.

     Uretprobes is a known potential user.

     Note this doesn't replace the default normal flavour of SRCU which
     still behaves the same as usual.

   - Add testing of SRCU-lite through rcutorture and rcuscale

   - Various cleanups on the way.

  Fixes:

   - Allow short-circuiting RCU-TASKS-RUDE grace periods on
     architectures that have sane noinstr boundaries forbidding tracing
     on low-level idle and kernel entry code. RCU-TASKS is enough on
     such configurations because it involves an RCU grace period that
     waits for all idle tasks to either schedule out voluntarily or
     enter into RCU unwatched noinstr code.

   - Allow and test start_poll_synchronize_rcu() with IRQs disabled.

   - Mention rcuog kthreads in relevant documentation and Kconfig help

   - Various fixes and consolidations

  rcutorture:

   - Add --no-affinity on tools to leave the affinity setting of guests
     up to the user.

   - Add guest_os_delay parameter to rcuscale for better warm-up
     control.

   - Fix and improve some rcuscale error handling.

   - Various cleanups and fixes

  stall:

   - Remove dead code

   - Stop dumping tasks if a stalled grace period eventually ended
     midway as that only produces confusing output.

   - Optimize detection of stalling CPUs and avoid useless node locking
     otherwise.

  NOCB:

   - Fix rcu_barrier() hang due to a race against callbacks
     deoffloading. This is not yet used, except by rcutorture, and waits
     for its promised cpusets interface.

   - Remove leftover function declaration"

* tag 'rcu.release.v6.13' of git://git.kernel.org/pub/scm/linux/kernel/git/rcu/linux: (42 commits)
  rcuscale: Remove redundant WARN_ON_ONCE() splat
  rcuscale: Do a proper cleanup if kfree_scale_init() fails
  srcu: Unconditionally record srcu_read_lock_lite() in ->srcu_reader_flavor
  srcu: Check for srcu_read_lock_lite() across all CPUs
  srcu: Remove smp_mb() from srcu_read_unlock_lite()
  rcutorture: Avoid printing cpu=-1 for no-fault RCU boost failure
  rcuscale: Add guest_os_delay module parameter
  refscale: Correct affinity check
  torture: Add --no-affinity parameter to kvm.sh
  rcu/nocb: Fix missed RCU barrier on deoffloading
  rcu/kvfree: Fix data-race in __mod_timer / kvfree_call_rcu
  rcu/srcutiny: don't return before reenabling preemption
  rcu-tasks: Remove open-coded one-byte cmpxchg() emulation
  doc: Remove kernel-parameters.txt entry for rcutorture.read_exit
  rcutorture: Test start-poll primitives with interrupts disabled
  rcu: Permit start_poll_synchronize_rcu*() with interrupts disabled
  rcu: Allow short-circuiting of synchronize_rcu_tasks_rude()
  doc: Add rcuog kthreads to kernel-per-CPU-kthreads.rst
  rcu: Add rcuog kthreads to RCU_NOCB_CPU help text
  rcu: Use the BITS_PER_LONG macro
  ...

1 files changed, 18 insertions, 39 deletions

diff --git a/kernel/rcu/tree_stall.h b/kernel/rcu/tree_stall.h
index 4432db6d0b99..925fcdad5dea 100644
--- a/kernel/rcu/tree_stall.h
+++ b/kernel/rcu/tree_stall.h
@@ -76,36 +76,6 @@ int rcu_jiffies_till_stall_check(void)
 }
 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
 
-/**
- * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
- *
- * Returns @true if the current grace period is sufficiently old that
- * it is reasonable to assume that it might be stalled.  This can be
- * useful when deciding whether to allocate memory to enable RCU-mediated
- * freeing on the one hand or just invoking synchronize_rcu() on the other.
- * The latter is preferable when the grace period is stalled.
- *
- * Note that sampling of the .gp_start and .gp_seq fields must be done
- * carefully to avoid false positives at the beginnings and ends of
- * grace periods.
- */
-bool rcu_gp_might_be_stalled(void)
-{
-	unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
-	unsigned long j = jiffies;
-
-	if (d < RCU_STALL_MIGHT_MIN)
-		d = RCU_STALL_MIGHT_MIN;
-	smp_mb(); // jiffies before .gp_seq to avoid false positives.
-	if (!rcu_gp_in_progress())
-		return false;
-	// Long delays at this point avoids false positive, but a delay
-	// of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
-	smp_mb(); // .gp_seq before second .gp_start
-	// And ditto here.
-	return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
-}
-
 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
 void rcu_sysrq_start(void)
 {
@@ -365,7 +335,7 @@ static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
  * that don't support NMI-based stack dumps.  The NMI-triggered stack
  * traces are more accurate because they are printed by the target CPU.
  */
-static void rcu_dump_cpu_stacks(void)
+static void rcu_dump_cpu_stacks(unsigned long gp_seq)
 {
 	int cpu;
 	unsigned long flags;
@@ -373,15 +343,23 @@ static void rcu_dump_cpu_stacks(void)
 
 	rcu_for_each_leaf_node(rnp) {
 		printk_deferred_enter();
-		raw_spin_lock_irqsave_rcu_node(rnp, flags);
-		for_each_leaf_node_possible_cpu(rnp, cpu)
+		for_each_leaf_node_possible_cpu(rnp, cpu) {
+			if (gp_seq != data_race(rcu_state.gp_seq)) {
+				printk_deferred_exit();
+				pr_err("INFO: Stall ended during stack backtracing.\n");
+				return;
+			}
+			if (!(data_race(rnp->qsmask) & leaf_node_cpu_bit(rnp, cpu)))
+				continue;
+			raw_spin_lock_irqsave_rcu_node(rnp, flags);
 			if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 				if (cpu_is_offline(cpu))
 					pr_err("Offline CPU %d blocking current GP.\n", cpu);
 				else
 					dump_cpu_task(cpu);
 			}
-		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
+			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
+		}
 		printk_deferred_exit();
 	}
 }
@@ -638,7 +616,7 @@ static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 	       (long)rcu_seq_current(&rcu_state.gp_seq), totqlen,
 	       data_race(rcu_state.n_online_cpus)); // Diagnostic read
 	if (ndetected) {
-		rcu_dump_cpu_stacks();
+		rcu_dump_cpu_stacks(gp_seq);
 
 		/* Complain about tasks blocking the grace period. */
 		rcu_for_each_leaf_node(rnp)
@@ -670,7 +648,7 @@ static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 	rcu_force_quiescent_state();  /* Kick them all. */
 }
 
-static void print_cpu_stall(unsigned long gps)
+static void print_cpu_stall(unsigned long gp_seq, unsigned long gps)
 {
 	int cpu;
 	unsigned long flags;
@@ -707,7 +685,7 @@ static void print_cpu_stall(unsigned long gps)
 	rcu_check_gp_kthread_expired_fqs_timer();
 	rcu_check_gp_kthread_starvation();
 
-	rcu_dump_cpu_stacks();
+	rcu_dump_cpu_stacks(gp_seq);
 
 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
 	/* Rewrite if needed in case of slow consoles. */
@@ -789,7 +767,8 @@ static void check_cpu_stall(struct rcu_data *rdp)
 	gs2 = READ_ONCE(rcu_state.gp_seq);
 	if (gs1 != gs2 ||
 	    ULONG_CMP_LT(j, js) ||
-	    ULONG_CMP_GE(gps, js))
+	    ULONG_CMP_GE(gps, js) ||
+	    !rcu_seq_state(gs2))
 		return; /* No stall or GP completed since entering function. */
 	rnp = rdp->mynode;
 	jn = jiffies + ULONG_MAX / 2;
@@ -810,7 +789,7 @@ static void check_cpu_stall(struct rcu_data *rdp)
 			pr_err("INFO: %s detected stall, but suppressed full report due to a stuck CSD-lock.\n", rcu_state.name);
 		} else if (self_detected) {
 			/* We haven't checked in, so go dump stack. */
-			print_cpu_stall(gps);
+			print_cpu_stall(gs2, gps);
 		} else {
 			/* They had a few time units to dump stack, so complain. */
 			print_other_cpu_stall(gs2, gps);