[PATCH] md/raid10: protect head_position accesses consistency

From: Chen Cheng

Date: Wed Jun 17 2026 - 22:52:46 EST


From: Chen Cheng <chencheng@xxxxxxxxx>

the completion paths update conf->mirrors[x].head_position
without a lock, while read_balance() reads the same field
locklessly.

protect conf->mirror[x].head_position consistency for
multi-CPUs by READ_ONCE()/WRITE_ONCE().

the report #1:
=====================
BUG: KCSAN: data-race in find_bio_disk.isra.0 / read_balance

write to 0xffff8f0309679fd0 of 8 bytes by interrupt on cpu 9:
find_bio_disk.isra.0+0x108/0x180
raid10_end_write_request+0xd2/0x530
bio_endio+0x3c9/0x3e0
[...]

read to 0xffff8f0309679fd0 of 8 bytes by task 338766 on cpu 7:
read_balance+0x2ec/0x700
raid10_read_request+0x17b/0x550
raid10_make_request+0x27e/0x1080
md_handle_request+0x2c5/0x700
[...]

value changed: 0x000000000032f440 -> 0x000000000014f8e0

the report #2:
=======================
BUG: KCSAN: data-race in raid10_end_read_request / read_balance

write to 0xffff8f03291091b0 of 8 bytes by interrupt on cpu 9:
raid10_end_read_request+0xf0/0x400
bio_endio+0x3c9/0x3e0
[...]

read to 0xffff8f03291091b0 of 8 bytes by task 336943 on cpu 8:
read_balance+0x2ec/0x700
raid10_read_request+0x17b/0x550
raid10_make_request+0x27e/0x1080
md_handle_request+0x2c5/0x700
[...]

value changed: 0x000000000041c188 -> 0x000000000041c1b8

Signed-off-by: Chen Cheng <chencheng@xxxxxxxxx>
---
drivers/md/raid10.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
index caf1da15aa87..55aa052a957b 100644
--- a/drivers/md/raid10.c
+++ b/drivers/md/raid10.c
@@ -343,12 +343,12 @@ static void raid_end_bio_io(struct r10bio *r10_bio)
*/
static inline void update_head_pos(int slot, struct r10bio *r10_bio)
{
struct r10conf *conf = r10_bio->mddev->private;

- conf->mirrors[r10_bio->devs[slot].devnum].head_position =
- r10_bio->devs[slot].addr + (r10_bio->sectors);
+ WRITE_ONCE(conf->mirrors[r10_bio->devs[slot].devnum].head_position,
+ r10_bio->devs[slot].addr + (r10_bio->sectors));
}

/*
* Find the disk number which triggered given bio
*/
@@ -841,11 +841,11 @@ static struct md_rdev *read_balance(struct r10conf *conf,
/* for far > 1 always use the lowest address */
else if (geo->far_copies > 1)
new_distance = r10_bio->devs[slot].addr;
else
new_distance = abs(r10_bio->devs[slot].addr -
- conf->mirrors[disk].head_position);
+ READ_ONCE(conf->mirrors[disk].head_position));

if (new_distance < best_dist) {
best_dist = new_distance;
best_dist_slot = slot;
best_dist_rdev = rdev;
--
2.54.0