Skip to contentSkip to conversation

blk-cgroup: Remove unnecessary rcu_read_lock/unlock()

PATCH Applied [email protected] 5 participants 7 messages May 16, 2022 – May 18, 2022

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 blk-cgroup: Remove unnecessary rcu_read_lock/unlock()

From: Fanjun Kong <[email protected]>

spin_lock_irq/spin_unlock_irq contains preempt_disable/enable(). Which can serve as RCU read-side critical region, so remove rcu_read_lock/unlock().

---
 block/blk-cgroup.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
block/blk-cgroup.c +1 -3
index a91f8ae18b49..7bdc16a36560 100644
@@ -1283,14 +1283,13 @@ int blkcg_init_queue(struct request_queue *q)
 	preloaded = !radix_tree_preload(GFP_KERNEL);
 
 	/* Make sure the root blkg exists. */
-	rcu_read_lock();
+	/* spin_lock_irq can serve as RCU read-side critical section. */
 	spin_lock_irq(&q->queue_lock);
 	blkg = blkg_create(&blkcg_root, q, new_blkg);
 	if (IS_ERR(blkg))
 		goto err_unlock;
 	q->root_blkg = blkg;
 	spin_unlock_irq(&q->queue_lock);
-	rcu_read_unlock();
 
 	if (preloaded)
 		radix_tree_preload_end();
@@ -1316,7 +1315,6 @@ int blkcg_init_queue(struct request_queue *q)
 	return ret;
 err_unlock:
 	spin_unlock_irq(&q->queue_lock);
-	rcu_read_unlock();
 	if (preloaded)
 		radix_tree_preload_end();
 	return PTR_ERR(blkg);
Signature · 1 line
2.36.0
maintainer Applied
#3 lore raw
In reply to [email protected]
Quoted 8 lines · On Tue, 17 May 2022 01:39:30 +0800, [email protected] wrote:

From: Fanjun Kong <[email protected]>

spin_lock_irq/spin_unlock_irq contains preempt_disable/enable(). Which can serve as RCU read-side critical region, so remove rcu_read_lock/unlock().

[...]

Applied, thanks!

[1/1] blk-cgroup: Remove unnecessary rcu_read_lock/unlock()

      commit: 77c570a1ea85ba4ab135c61a028420a6e9fe77f3

Best regards,

Signature · 1 line
Jens Axboe
#4 lore raw
In reply to [email protected]
Quoted 7 lines · On 16.05.2022 19:39, [email protected] wrote:

From: Fanjun Kong <[email protected]>

spin_lock_irq/spin_unlock_irq contains preempt_disable/enable(). Which can serve as RCU read-side critical region, so remove rcu_read_lock/unlock().

This patch landed in today's linux next-20220518 as commit 77c570a1ea85 ("blk-cgroup: Remove unnecessary rcu_read_lock/unlock()").

Unfortunately it triggers the following warning on ARM64 based Raspberry Pi 4B board:

------------[ cut here ]------------
WARNING: CPU: 0 PID: 1 at block/blk-cgroup.c:301 blkg_create+0x398/0x4e0
Modules linked in:
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.18.0-rc3+ #5080
Hardware name: Raspberry Pi 4 Model B (DT)
pstate: 600000c5 (nZCv daIF -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
pc : blkg_create+0x398/0x4e0
...
Call trace:
  blkg_create+0x398/0x4e0
  blkcg_init_queue+0x74/0x204
  __alloc_disk_node+0xf8/0x1f0
  __blk_alloc_disk+0x38/0x140
  brd_alloc.part.0+0xf8/0x220
  brd_init+0xe8/0x164
  do_one_initcall+0x74/0x400
  kernel_init_freeable+0x2f4/0x37c
  kernel_init+0x28/0x130
  ret_from_fork+0x10/0x20
irq event stamp: 218372
hardirqs last  enabled at (218371): [<ffff80000914b99c>] 
_raw_spin_unlock_irqrestore+0x98/0x9c
hardirqs last disabled at (218372): [<ffff80000914bcbc>] 
_raw_spin_lock_irq+0xac/0xb0
softirqs last  enabled at (216732): [<ffff800008010470>] _stext+0x470/0x5e8
softirqs last disabled at (216723): [<ffff8000080a0ec4>] 
__irq_exit_rcu+0x180/0x1ac
---[ end trace 0000000000000000 ]---

If this is a false positive, then the check in the code needs to be adjusted.

Quoted 25 lines
---
  block/blk-cgroup.c | 4 +---
  1 file changed, 1 insertion(+), 3 deletions(-)
block/blk-cgroup.c +1 -2
index a91f8ae18b49..7bdc16a36560 100644
@@ -1283,14 +1283,13 @@ int blkcg_init_queue(struct request_queue *q)
  	preloaded = !radix_tree_preload(GFP_KERNEL);
 
  	/* Make sure the root blkg exists. */
-	rcu_read_lock();
+	/* spin_lock_irq can serve as RCU read-side critical section. */
  	spin_lock_irq(&q->queue_lock);
  	blkg = blkg_create(&blkcg_root, q, new_blkg);
  	if (IS_ERR(blkg))
  		goto err_unlock;
  	q->root_blkg = blkg;
  	spin_unlock_irq(&q->queue_lock);
-	rcu_read_unlock();
 
  	if (preloaded)
  		radix_tree_preload_end();

block/blk-cgroup.c int blkcg_init_queue(struct request_queue *q) · line 1320

Quoted 7 lines
block/blk-cgroup.c +0 -1
@@ -1316,7 +1315,6 @@ int blkcg_init_queue(struct request_queue *q)
  	return ret;
  err_unlock:
  	spin_unlock_irq(&q->queue_lock);
-	rcu_read_unlock();
  	if (preloaded)
  		radix_tree_preload_end();
  	return PTR_ERR(blkg);

Best regards

Signature · 2 lines
Marek Szyprowski, PhD
Samsung R&D Institute Poland
maintainer
#5 lore raw
Quoted 9 lines · Marek Szyprowski
Quoted 7 lines · On 16.05.2022 19:39, [email protected] wrote:

From: Fanjun Kong <[email protected]>

spin_lock_irq/spin_unlock_irq contains preempt_disable/enable(). Which can serve as RCU read-side critical region, so remove rcu_read_lock/unlock().

Quoted 7 lines

This patch landed in today's linux next-20220518 as commit 77c570a1ea85 ("blk-cgroup: Remove unnecessary rcu_read_lock/unlock()").

Unfortunately it triggers the following warning on ARM64 based Raspberry Pi 4B board:>

------------[ cut here ]------------
WARNING: CPU: 0 PID: 1 at block/blk-cgroup.c:301 blkg_create+0x398/0x4e0

Should this use rcu_read_lock_any_held() rather than rcu_read_lock_held()?

Signature · 1 line
Jens Axboe
maintainer
#6 lore raw
Quoted 13 lines · Jens Axboe
Quoted 12 lines · Marek Szyprowski
Quoted 7 lines · On 16.05.2022 19:39, [email protected] wrote:

From: Fanjun Kong <[email protected]>

spin_lock_irq/spin_unlock_irq contains preempt_disable/enable(). Which can serve as RCU read-side critical region, so remove rcu_read_lock/unlock().

This patch landed in today's linux next-20220518 as commit 77c570a1ea85 ("blk-cgroup: Remove unnecessary rcu_read_lock/unlock()").

Quoted 7 lines
Quoted 4 lines

Unfortunately it triggers the following warning on ARM64 based Raspberry Pi 4B board:>

------------[ cut here ]------------
WARNING: CPU: 0 PID: 1 at block/blk-cgroup.c:301 blkg_create+0x398/0x4e0

Should this use rcu_read_lock_any_held() rather than rcu_read_lock_held()?

I think the better alternative is just to delete the WARN_ON(), we have a:

lockdep_assert_held(&q->queue_lock);

right after it. Since the queue_lock is IRQ disabling, having two checks serves no purpose. I'll kill the line.

Signature · 1 line
Jens Axboe