[RFC PATCH v1] stop_machine: Add stop_housekeeping_cpuslocked()

[Costa Shulyupin]

CPU hotplug interferes with CPU isolation and introduces latency to
real-time tasks.

The test:

  rtla timerlat hist -c 1 -a 500 &
  echo 0 > /sys/devices/system/cpu/cpu2/online

The RTLA tool reveals the following blocking thread stack trace:

               -> multi_cpu_stop
               -> cpu_stopper_thread
               -> smpboot_thread_fn

This happens because multi_cpu_stop() disables interrupts for EACH online
CPU since takedown_cpu() indirectly invokes take_cpu_down() through
stop_machine_cpuslocked(). I'm omitting the detailed description of the
call chain.

Proposal: Limit the stop operation to housekeeping CPUs.

take_cpu_down() invokes with cpuhp_invoke_callback_range_nofail:
 - tick_cpu_dying()
 - hrtimers_cpu_dying()
 - smpcfd_dying_cpu()
 - x86_pmu_dying_cpu()
 - rcutree_dying_cpu()
 - sched_cpu_dying()
 - cache_ap_offline()

Which synchronizations do these functions require instead of stop_machine?

Passed standard regression tests:
 - https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/6042
 - https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-10/-/merge_requests/89

Passes primitive CPU hotplug test without warnings.

What tests and test suites do you recommend?

Signed-off-by: Costa Shulyupin <costa.shul@xxxxxxxxxx>
---
 include/linux/stop_machine.h | 12 ++++++++++++
 kernel/cpu.c                 |  2 +-
 kernel/stop_machine.c        | 23 +++++++++++++++++++++++
 3 files changed, 36 insertions(+), 1 deletion(-)

diff --git a/include/linux/stop_machine.h b/include/linux/stop_machine.h
index 3132262a404dc..4c9e709f174c8 100644
--- a/include/linux/stop_machine.h
+++ b/include/linux/stop_machine.h
@@ -140,6 +140,18 @@ int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data, const struct cpumask *
  */
 int stop_core_cpuslocked(unsigned int cpu, cpu_stop_fn_t fn, void *data);
 
+/**
+ * stop_housekeeping_cpuslocked: freeze housekeeping CPUs and run specified function.
+ * Unlike stop_machine_cpuslocked, it doesn't stop isolated CPUs.
+ * @fn: the function to run
+ * @data: the data ptr for the @fn()
+ * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
+ *
+ * Must be called from with in a cpus_read_lock() protected
+ * region. Avoids nested calls to cpus_read_lock().
+ */
+int stop_housekeeping_cpuslocked(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus);
+
 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
 				   const struct cpumask *cpus);
 #else	/* CONFIG_SMP || CONFIG_HOTPLUG_CPU */
diff --git a/kernel/cpu.c b/kernel/cpu.c
index 6d7c51e7c366c..052f48bebf816 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -1310,7 +1310,7 @@ static int takedown_cpu(unsigned int cpu)
 	/*
 	 * So now all preempt/rcu users must observe !cpu_active().
 	 */
-	err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
+	err = stop_housekeeping_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 	if (err) {
 		/* CPU refused to die */
 		irq_unlock_sparse();
diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
index da821ce258ea7..1f36e44d42476 100644
--- a/kernel/stop_machine.c
+++ b/kernel/stop_machine.c
@@ -22,6 +22,7 @@
 #include <linux/atomic.h>
 #include <linux/nmi.h>
 #include <linux/sched/wake_q.h>
+#include <linux/sched/isolation.h>
 
 /*
  * Structure to determine completion condition and record errors.  May
@@ -619,6 +620,28 @@ int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
 	return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
 }
 
+int stop_housekeeping_cpuslocked(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
+{
+	struct multi_stop_data msdata = { .fn = fn, .data = data, .active_cpus = cpus };
+	cpumask_var_t stop_mask;
+	int ret;
+
+	lockdep_assert_cpus_held();
+
+	if (!alloc_cpumask_var(&stop_mask, GFP_KERNEL))
+		return -ENOMEM;
+	cpumask_and(stop_mask, cpu_online_mask, housekeeping_cpumask(HK_TYPE_MANAGED_IRQ));
+	cpumask_or(stop_mask, stop_mask, cpus);
+	msdata.num_threads = cpumask_weight(stop_mask);
+
+	/* Set the initial state and stop online housekeeping cpus. */
+	set_state(&msdata, MULTI_STOP_PREPARE);
+	ret = stop_cpus(stop_mask, multi_cpu_stop, &msdata);
+	free_cpumask_var(stop_mask);
+
+	return ret;
+}
+
 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
 {
 	int ret;
-- 
2.47.0