Skip to contentSkip to conversation

lib/group_cpus.c: avoid to acquire cpu hotplug lock in group_cpus_evenly

PATCH Superseded Ming Lei 2 participants 5 messages Aug 18, 2023 – Aug 31, 2023

Subsystems (4)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Ming Lei

author
#0 lore raw

PATCH v3 1/1 lib/group_cpus.c: avoid to acquire cpu hotplug lock in group_cpus_evenly

group_cpus_evenly() could be part of storage driver's error handler, such as nvme driver, when may happen during CPU hotplug, in which storage queue has to drain its pending IOs because all CPUs associated with the queue are offline and the queue is becoming inactive. And handling IO needs error handler to provide forward progress.

Then dead lock is caused:

1) inside CPU hotplug handler, CPU hotplug lock is held, and blk-mq's handler is waiting for inflight IO

2) error handler is waiting for CPU hotplug lock

3) inflight IO can't be completed in blk-mq's CPU hotplug handler because error handling can't provide forward progress.

Solve the deadlock by not holding CPU hotplug lock in group_cpus_evenly(), in which two stage spreads are taken: 1) the 1st stage is over all present CPUs; 2) the end stage is over all other CPUs.

Turns out the two stage spread just needs consistent 'cpu_present_mask', and remove the CPU hotplug lock by storing it into one local cache. This way doesn't change correctness, because all CPUs are still covered.

---
V3:
- reuse npresmsk, and avoid to allocate new variable, suggested by Chengming Zhou

V2:
- fix "Cc: block list"
- add tested-by tag

 lib/group_cpus.c | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)
lib/group_cpus.c +9 -6
index aa3f6815bb12..fffe8a893597 100644
@@ -366,13 +366,18 @@ struct cpumask *group_cpus_evenly(unsigned int numgrps)
 	if (!masks)
 		goto fail_node_to_cpumask;
 
-	/* Stabilize the cpumasks */
-	cpus_read_lock();
 	build_node_to_cpumask(node_to_cpumask);
 
+	/*
+	 * Make a local cache of 'cpu_present_mask', so the two stages
+	 * spread can observe consistent 'cpu_present_mask' without holding
+	 * cpu hotplug lock.
+	 */
+	cpumask_copy(npresmsk, cpu_present_mask);
+
 	/* grouping present CPUs first */
 	ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
-				  cpu_present_mask, nmsk, masks);
+				  npresmsk, nmsk, masks);
 	if (ret < 0)
 		goto fail_build_affinity;
 	nr_present = ret;
@@ -387,15 +392,13 @@ struct cpumask *group_cpus_evenly(unsigned int numgrps)
 		curgrp = 0;
 	else
 		curgrp = nr_present;
-	cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
+	cpumask_andnot(npresmsk, cpu_possible_mask, npresmsk);
 	ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
 				  npresmsk, nmsk, masks);
 	if (ret >= 0)
 		nr_others = ret;
 
  fail_build_affinity:
-	cpus_read_unlock();
-
 	if (ret >= 0)
 		WARN_ON(nr_present + nr_others < numgrps);
 
Signature · 1 line
2.40.1
author
#1 lore raw
Quoted 33 lines · Ming Lei

group_cpus_evenly() could be part of storage driver's error handler, such as nvme driver, when may happen during CPU hotplug, in which storage queue has to drain its pending IOs because all CPUs associated with the queue are offline and the queue is becoming inactive. And handling IO needs error handler to provide forward progress.

Then dead lock is caused:

1) inside CPU hotplug handler, CPU hotplug lock is held, and blk-mq's handler is waiting for inflight IO

2) error handler is waiting for CPU hotplug lock

3) inflight IO can't be completed in blk-mq's CPU hotplug handler because error handling can't provide forward progress.

Solve the deadlock by not holding CPU hotplug lock in group_cpus_evenly(), in which two stage spreads are taken: 1) the 1st stage is over all present CPUs; 2) the end stage is over all other CPUs.

Turns out the two stage spread just needs consistent 'cpu_present_mask', and remove the CPU hotplug lock by storing it into one local cache. This way doesn't change correctness, because all CPUs are still covered.

Quoted 4 lines

---
V3:
- reuse npresmsk, and avoid to allocate new variable, suggested by Chengming Zhou

Hello Thomas and Jens,

Ping...

Thanks,
Ming

#2 lore raw
In reply to Ming Lei

Ming!

Quoted 11 lines · Ming Lei

group_cpus_evenly() could be part of storage driver's error handler, such as nvme driver, when may happen during CPU hotplug, in which storage queue has to drain its pending IOs because all CPUs associated with the queue are offline and the queue is becoming inactive. And handling IO needs error handler to provide forward progress.

Then dead lock is caused:

1) inside CPU hotplug handler, CPU hotplug lock is held, and blk-mq's handler is waiting for inflight IO

Quoted 8 lines

2) error handler is waiting for CPU hotplug lock

3) inflight IO can't be completed in blk-mq's CPU hotplug handler because error handling can't provide forward progress.

Solve the deadlock by not holding CPU hotplug lock in group_cpus_evenly(), in which two stage spreads are taken: 1) the 1st stage is over all present CPUs; 2) the end stage is over all other CPUs.

That solves the deadlock, but makes the code racy against a concurrent hot-add operation which modifies cpu_present_mask. IOW, it introduces a data race.

The changelog does not explain why this does not matter nor does the comment.

Thanks,

tglx

author
#3 lore raw

Hello Thomas,

Quoted 20 lines · Thomas Gleixner

Ming!

Quoted 16 lines · Ming Lei

group_cpus_evenly() could be part of storage driver's error handler, such as nvme driver, when may happen during CPU hotplug, in which storage queue has to drain its pending IOs because all CPUs associated with the queue are offline and the queue is becoming inactive. And handling IO needs error handler to provide forward progress.

Then dead lock is caused:

1) inside CPU hotplug handler, CPU hotplug lock is held, and blk-mq's handler is waiting for inflight IO

2) error handler is waiting for CPU hotplug lock

3) inflight IO can't be completed in blk-mq's CPU hotplug handler because error handling can't provide forward progress.

Quoted 7 lines
Quoted 3 lines

Solve the deadlock by not holding CPU hotplug lock in group_cpus_evenly(), in which two stage spreads are taken: 1) the 1st stage is over all present CPUs; 2) the end stage is over all other CPUs.

That solves the deadlock, but makes the code racy against a concurrent hot-add operation which modifies cpu_present_mask. IOW, it introduces a data race.

Good catch, we can annotate it with data_race(), and I think here the readonly access to cpu_present_mask is fine, and we have many such usage, such as for_each_present_cpu().

Quoted 3 lines

The changelog does not explain why this does not matter nor does the comment.

OK, I will document the change, such as:

There isn't difference compared with holding cpu hotplug lock, because the cpu hp state is always changed even though cpu_present_mask isn't updated with the lock. And all CPUs are always covered during the spread, the difference is just that it is done in 1st stage or 2nd stage.

If you don't object, I will post next version with above change.

Thanks,
Ming

#4 lore raw

Ming!

Quoted 4 lines · Ming Lei

On Wed, Aug 30, 2023 at 07:44:03PM +0200, Thomas Gleixner wrote:
There isn't difference compared with holding cpu hotplug lock, because the cpu hp state is always changed even though cpu_present_mask isn't updated with the lock.

Updates to cpu_present_mask happen with the cpus lock write locked:

acpi_processor_hotadd_init()
cpu_maps_update_begin()
cpus_write_lock()
acpi_map_cpu()

    generic_processor_info()
      set_cpu_present(cpu, true);

The only CPU mask which is lockless race free accessible is cpus_possible_mask because that one is set up at boot once and never modified afterwards.

Quoted 2 lines

And all CPUs are always covered during the spread, the difference is just that it is done in 1st stage or 2nd stage.

I'm not objecting against the approach per se. It just needs proper documentation why this is correct.

Thanks,

tglx