Thanks for CCing me on this thread. Indeed this seems to be the exact same issue that I (and Vincent Chen) reported in the other thread [1]. Ming Lei <ming.lei@xxxxxxxxxx> writes: > diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c > index 2cafcf11ee8b..3cc14fc76546 100644 I can confirm that this patch fixes these issues. Prior to this patch, I managed to reproduce both the "Unexpected reply" and "Double reply" bugs reliably via a TCP/IP connection between two AMD64 VMs on different hosts, two aarch64 VMs, and an aarch64 Raspberry Pi 5 SBC connecting to an NBD server on another host. With this patch, at least in the AMD64 VM setup, it survived an hour-long stress test using the `stress-ng` command from [1]. Also, sending a SIGINT to a process actively using the disk often manifested this bug. I could not reproduce this faulty behavior with this patch. I'm still working on backporting this to and compiling the v6.6 Raspberry Pi vendor kernel and will report my observations there. I'll also try Ming's v2 from [2]. Thanks! -Leon [1]: https://lore.kernel.org/linux-block/87jzea4w2n.fsf@xxxxxxxxxxxxxxxxxxxxxxxxxx/ [2]: https://lore.kernel.org/linux-block/20241018140831.3064135-1-ming.lei@xxxxxxxxxx/