On 11/10/20 2:06 AM, Song Liu wrote:

On Sun, Nov 8, 2020 at 6:02 PM heming.zhao@suse.com
[off-list ref] wrote:

quoted

Please note, I gave two solutions for this bug in cover-letter.
This patch uses solution 2. For detail, please check cover-letter.

Thank you.

[...]

quoted

quoted

How to fix:

There are two sides to fix (or break the dead loop):
1. on sending msg side, modify lock_comm, change it to return
     success/failed.
     This will make mdadm cmd return error when lock_comm is timeout.
2. on receiving msg side, process_metadata_update need to add error
     handling.
     currently, other msg types won't trigger error or error doesn't need
     to return sender. So only process_metadata_update need to modify.

Ether of 1 & 2 can fix the hunging issue, but I prefer fix on both side.

Similar comments on how to make the commit log easy to understand.
Besides that, please split the change into two commits, for fix #1 and #2
respectively.

My comment meaning is that solution 2 also has two sub-solutions: sending side or receiving side.
(but in fact, there are 3 sub-solutions: sending, receiving & both sides)

sending side, related with patch 2 functions: sendmsg & lock_comm
 (code flow: sendmsg => lock_comm)

receiving side, related with patch 2 functions: process_recvd_msg & process_metadata_update
 (code flow: process_recvd_msg => process_metadata_update)

To break any side waiting can break deadlock. In the patch 2, my fix is both sides.