Hi Dongli,

-----Original Message-----
From: Dongli Zhang <redacted>
Sent: Tuesday, June 25, 2019 9:30 AM
To: Wenbin Zeng <redacted>
Cc: axboe@kernel.dk; keith.busch@intel.com; hare@suse.com; ming.lei@redhat.com;
osandov@fb.com; sagi@grimberg.me; bvanassche@acm.org;
linux-block@vger.kernel.org; linux-kernel@vger.kernel.org; wenbinzeng(曾文斌)
[off-list ref]
Subject: Re: [PATCH] blk-mq: update hctx->cpumask at cpu-hotplug(Internet mail)

Hi Wenbin,

On 6/24/19 11:24 PM, Wenbin Zeng wrote:

quoted

Currently hctx->cpumask is not updated when hot-plugging new cpus,
as there are many chances kblockd_mod_delayed_work_on() getting
called with WORK_CPU_UNBOUND, workqueue blk_mq_run_work_fn may run
on the newly-plugged cpus, consequently __blk_mq_run_hw_queue()
reporting excessive "run queue from wrong CPU" messages because
cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) returns false.

This patch added a cpu-hotplug handler into blk-mq, updating
hctx->cpumask at cpu-hotplug.

Signed-off-by: Wenbin Zeng <redacted>
---
 block/blk-mq.c         | 29 +++++++++++++++++++++++++++++
 include/linux/blk-mq.h |  1 +
 2 files changed, 30 insertions(+)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index ce0f5f4..2e465fc 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -39,6 +39,8 @@
 #include "blk-mq-sched.h"
 #include "blk-rq-qos.h"

+static enum cpuhp_state cpuhp_blk_mq_online;
+
 static void blk_mq_poll_stats_start(struct request_queue *q);
 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);

@@ -2215,6 +2217,21 @@ int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct

blk_mq_tags *tags,

quoted

 	return -ENOMEM;
 }

+static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
+{
+	struct blk_mq_hw_ctx *hctx;
+
+	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_online);
+
+	if (!cpumask_test_cpu(cpu, hctx->cpumask)) {
+		mutex_lock(&hctx->queue->sysfs_lock);
+		cpumask_set_cpu(cpu, hctx->cpumask);
+		mutex_unlock(&hctx->queue->sysfs_lock);
+	}
+
+	return 0;
+}
+

As this callback is registered for each hctx, when a cpu is online, it is called
for each hctx.

Just taking a 4-queue nvme as example (regardless about other block like loop).
Suppose cpu=2 (out of 0, 1, 2 and 3) is offline. When we online cpu=2,

blk_mq_hctx_notify_online() called: cpu=2 and blk_mq_hw_ctx->queue_num=3
blk_mq_hctx_notify_online() called: cpu=2 and blk_mq_hw_ctx->queue_num=2
blk_mq_hctx_notify_online() called: cpu=2 and blk_mq_hw_ctx->queue_num=1
blk_mq_hctx_notify_online() called: cpu=2 and blk_mq_hw_ctx->queue_num=0

There is no need to set cpu 2 for blk_mq_hw_ctx->queue_num=[3, 1, 0]. I am
afraid this patch would erroneously set cpumask for blk_mq_hw_ctx->queue_num=[3,
1, 0].

I used to submit the below patch explaining above for removing a cpu and it is
unfortunately not merged yet.

https://patchwork.kernel.org/patch/10889307/


Another thing is during initialization, the hctx->cpumask should already been
set and even the cpu is offline. Would you please explain the case hctx->cpumask
is not set correctly, e.g., how to reproduce with a kvm guest running
scsi/virtio/nvme/loop?

My scenario is:
A kvm guest started with single cpu, during initialization only one cpu was visible by kernel.
After boot, I hot-add some cpus via qemu monitor (I believe virsh setvcpus --live can do the same thing), for example:
(qemu) cpu-add 1
(qemu) cpu-add 2
(qemu) cpu-add 3

In such scenario, hctx->cpumask doesn't get updated when these cpus are added.

Dongli Zhang