Now we support sharing one page if PAGE_SIZE is not equal stripe size. To support this,
it needs to support calculating xor value with different offsets for each r5dev. One
offset array is used to record those offsets.
In RMW mode, parity page is used as a source page. It sets ASYNC_TX_XOR_DROP_DST before
calculating xor value in ops_run_prexor5. So it needs to add src_list and src_offs at
the same time. Now it only needs src_list. So the xor value which is calculated is wrong.
It can cause data corruption problem.
I can reproduce this problem 100% on a POWER8 machine. The steps are:
mdadm -CR /dev/md0 -l5 -n3 /dev/sdb1 /dev/sdc1 /dev/sdd1 --size=3G
mkfs.xfs /dev/md0
mount /dev/md0 /mnt/test
mount: /mnt/test: mount(2) system call failed: Structure needs cleaning.
Fixes: 29bcff787 ("md/raid5: add new xor function to support different page offset")
Signed-off-by: Xiao Ni <xni@xxxxxxxxxx>
---
crypto/async_tx/async_xor.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/crypto/async_tx/async_xor.c b/crypto/async_tx/async_xor.c
index a057ecb..6cd7f70 100644
--- a/crypto/async_tx/async_xor.c
+++ b/crypto/async_tx/async_xor.c
@@ -233,6 +233,7 @@ async_xor_offs(struct page *dest, unsigned int offset,
if (submit->flags & ASYNC_TX_XOR_DROP_DST) {
src_cnt--;
src_list++;
+ src_offs++;
}
/* wait for any prerequisite operations */
--
2.7.5
