diff mbox series

[v2] scsi: use GFP_NOIO to avoid circular locking dependency

Message ID 20250129104525.0ae8421e@fangorn (mailing list archive)
State New
Headers show
Series [v2] scsi: use GFP_NOIO to avoid circular locking dependency | expand

Commit Message

Rik van Riel Jan. 29, 2025, 3:45 p.m. UTC 
On Tue, 28 Jan 2025 21:35:18 -0800
Christoph Hellwig <hch@infradead.org> wrote:

> GFP_NOFS is never the right thing for block layer allocations.
> The right thing here is GFP_NOIO which is a superset of GFP_NOFS.
> Otherwise you could reproduce the same deadlock when using swap
> instead of a file system to reproduce basically the same deadlock.

Duh, you are right of course!

The fixed up patch with GFP_NOIO is below.

---8<---

From 74272b4537415fd7d94c216e422510c27aa88fa0 Mon Sep 17 00:00:00 2001
From: Rik van Riel <riel@surriel.com>
Date: Tue, 28 Jan 2025 16:35:39 -0500
Subject: [PATCH] scsi: use GFP_NOIO to avoid circular locking dependency
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Filesystems can write to disk from page reclaim with __GFP_FS
set. Marc found a case where scsi_realloc_sdev_budget_map
ends up in page reclaim with GFP_KERNEL, where it could try
to take filesystem locks again, leading to a deadlock.

WARNING: possible circular locking dependency detected
6.13.0 #1 Not tainted
------------------------------------------------------
kswapd0/70 is trying to acquire lock:
ffff8881025d5d78 (&q->q_usage_counter(io)){++++}-{0:0}, at: blk_mq_submit_bio+0x461/0x6e0

but task is already holding lock:
ffffffff81ef5f40 (fs_reclaim){+.+.}-{0:0}, at: balance_pgdat+0x9f/0x760

The full lockdep splat can be found in Marc's report:

https://lkml.org/lkml/2025/1/24/1101

Avoid the potential deadlock by doing the allocation with GFP_NOIO,
which prevents both filesystem and block layer recursion.

Reported-by: Marc Aurèle La France <tsi@tuyoix.net>
Signed-off-by: Rik van Riel <riel@surriel.com>
---
 drivers/scsi/scsi_scan.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Christoph Hellwig Jan. 31, 2025, 7:10 a.m. UTC | #1 
On Wed, Jan 29, 2025 at 10:45:25AM -0500, Rik van Riel wrote:
> On Tue, 28 Jan 2025 21:35:18 -0800
> Christoph Hellwig <hch@infradead.org> wrote:
> 
> > GFP_NOFS is never the right thing for block layer allocations.
> > The right thing here is GFP_NOIO which is a superset of GFP_NOFS.
> > Otherwise you could reproduce the same deadlock when using swap
> > instead of a file system to reproduce basically the same deadlock.
> 
> Duh, you are right of course!
> 
> The fixed up patch with GFP_NOIO is below.

Looks good:

Reviewed-by: Christoph Hellwig <hch@lst.de>
diff mbox series

Patch

diff --git a/drivers/scsi/scsi_scan.c b/drivers/scsi/scsi_scan.c
index f2093982b3db..b0964b6dd646 100644
--- a/drivers/scsi/scsi_scan.c
+++ b/drivers/scsi/scsi_scan.c
@@ -245,7 +245,7 @@  static int scsi_realloc_sdev_budget_map(struct scsi_device *sdev,
 	}
 	ret = sbitmap_init_node(&sdev->budget_map,
 				scsi_device_max_queue_depth(sdev),
-				new_shift, GFP_KERNEL,
+				new_shift, GFP_NOIO,
 				sdev->request_queue->node, false, true);
 	if (!ret)
 		sbitmap_resize(&sdev->budget_map, depth);