[PATCH RT 16/22] block: blk-mq: move blk_queue_usage_counter_release() into process context

[Julia Cartwright]

From: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>

4.9.115-rt94-rc1 stable review patch.
If you have any objection to the inclusion of this patch, let me know.

--- 8< --- 8< --- 8< ---
| BUG: sleeping function called from invalid context at kernel/locking/rtmutex.c:914
| in_atomic(): 1, irqs_disabled(): 0, pid: 255, name: kworker/u257:6
| 5 locks held by kworker/u257:6/255:
|  #0:  ("events_unbound"){.+.+.+}, at: [<ffffffff8108edf1>] process_one_work+0x171/0x5e0
|  #1:  ((&entry->work)){+.+.+.}, at: [<ffffffff8108edf1>] process_one_work+0x171/0x5e0
|  #2:  (&shost->scan_mutex){+.+.+.}, at: [<ffffffffa000faa3>] __scsi_add_device+0xa3/0x130 [scsi_mod]
|  #3:  (&set->tag_list_lock){+.+...}, at: [<ffffffff812f09fa>] blk_mq_init_queue+0x96a/0xa50
|  #4:  (rcu_read_lock_sched){......}, at: [<ffffffff8132887d>] percpu_ref_kill_and_confirm+0x1d/0x120
| Preemption disabled at:[<ffffffff812eff76>] blk_mq_freeze_queue_start+0x56/0x70
|
| CPU: 2 PID: 255 Comm: kworker/u257:6 Not tainted 3.18.7-rt0+ #1
| Workqueue: events_unbound async_run_entry_fn
|  0000000000000003 ffff8800bc29f998 ffffffff815b3a12 0000000000000000
|  0000000000000000 ffff8800bc29f9b8 ffffffff8109aa16 ffff8800bc29fa28
|  ffff8800bc5d1bc8 ffff8800bc29f9e8 ffffffff815b8dd4 ffff880000000000
| Call Trace:
|  [<ffffffff815b3a12>] dump_stack+0x4f/0x7c
|  [<ffffffff8109aa16>] __might_sleep+0x116/0x190
|  [<ffffffff815b8dd4>] rt_spin_lock+0x24/0x60
|  [<ffffffff810b6089>] __wake_up+0x29/0x60
|  [<ffffffff812ee06e>] blk_mq_usage_counter_release+0x1e/0x20
|  [<ffffffff81328966>] percpu_ref_kill_and_confirm+0x106/0x120
|  [<ffffffff812eff76>] blk_mq_freeze_queue_start+0x56/0x70
|  [<ffffffff812f0000>] blk_mq_update_tag_set_depth+0x40/0xd0
|  [<ffffffff812f0a1c>] blk_mq_init_queue+0x98c/0xa50
|  [<ffffffffa000dcf0>] scsi_mq_alloc_queue+0x20/0x60 [scsi_mod]
|  [<ffffffffa000ea35>] scsi_alloc_sdev+0x2f5/0x370 [scsi_mod]
|  [<ffffffffa000f494>] scsi_probe_and_add_lun+0x9e4/0xdd0 [scsi_mod]
|  [<ffffffffa000fb26>] __scsi_add_device+0x126/0x130 [scsi_mod]
|  [<ffffffffa013033f>] ata_scsi_scan_host+0xaf/0x200 [libata]
|  [<ffffffffa012b5b6>] async_port_probe+0x46/0x60 [libata]
|  [<ffffffff810978fb>] async_run_entry_fn+0x3b/0xf0
|  [<ffffffff8108ee81>] process_one_work+0x201/0x5e0

percpu_ref_kill_and_confirm() invokes blk_mq_usage_counter_release() in
a rcu-sched region. swait based wake queue can't be used due to
wake_up_all() usage and disabled interrupts in !RT configs (as reported
by Corey Minyard).
The wq_has_sleeper() check has been suggested by Peter Zijlstra.

Cc: stable-rt@xxxxxxxxxxxxxxx
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
(cherry picked from commit 2d701058d614554cce412a787f00568b9fdffade)
Signed-off-by: Julia Cartwright <julia@xxxxxx>
---
 block/blk-core.c       | 14 +++++++++++++-
 include/linux/blkdev.h |  2 ++
 2 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 87d3e0a503e5..346d5bba3948 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -675,12 +675,21 @@ void blk_queue_exit(struct request_queue *q)
 	percpu_ref_put(&q->q_usage_counter);
 }
 
+static void blk_queue_usage_counter_release_swork(struct swork_event *sev)
+{
+	struct request_queue *q =
+		container_of(sev, struct request_queue, mq_pcpu_wake);
+
+	wake_up_all(&q->mq_freeze_wq);
+}
+
 static void blk_queue_usage_counter_release(struct percpu_ref *ref)
 {
 	struct request_queue *q =
 		container_of(ref, struct request_queue, q_usage_counter);
 
-	wake_up_all(&q->mq_freeze_wq);
+	if (wq_has_sleeper(&q->mq_freeze_wq))
+		swork_queue(&q->mq_pcpu_wake);
 }
 
 static void blk_rq_timed_out_timer(unsigned long data)
@@ -751,6 +760,7 @@ struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
 	__set_bit(QUEUE_FLAG_BYPASS, &q->queue_flags);
 
 	init_waitqueue_head(&q->mq_freeze_wq);
+	INIT_SWORK(&q->mq_pcpu_wake, blk_queue_usage_counter_release_swork);
 
 	/*
 	 * Init percpu_ref in atomic mode so that it's faster to shutdown.
@@ -3556,6 +3566,8 @@ int __init blk_dev_init(void)
 	if (!kblockd_workqueue)
 		panic("Failed to create kblockd\n");
 
+	BUG_ON(swork_get());
+
 	request_cachep = kmem_cache_create("blkdev_requests",
 			sizeof(struct request), 0, SLAB_PANIC, NULL);
 
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index fdb449fe3ff7..ab039211ab9f 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -24,6 +24,7 @@
 #include <linux/rcupdate.h>
 #include <linux/percpu-refcount.h>
 #include <linux/scatterlist.h>
+#include <linux/swork.h>
 
 struct module;
 struct scsi_ioctl_command;
@@ -469,6 +470,7 @@ struct request_queue {
 #endif
 	struct rcu_head		rcu_head;
 	wait_queue_head_t	mq_freeze_wq;
+	struct swork_event	mq_pcpu_wake;
 	struct percpu_ref	q_usage_counter;
 	struct list_head	all_q_node;
 
-- 
2.18.0