Re: [PATCH] nvme-tcp: fix io_work priority inversion when send_mutex is held

[Seiji Nishikawa]

Hello Keith,

A short follow-up on this patch.

The problem is a regression in v6.16 and later, and this is the minimal
fix for it. It applies cleanly on the current nvme-7.4 branch. On the
customer system, the unpatched kernel hung every time. With this patch,
there has been no hang since October 15. blktests nvme over tcp passes
with it, with the same results as without it.

Hannes suggested removing the direct send path instead. I agree with
that as a follow-up, and I will send it as a separate RFC patch for the
next merge window, as I wrote in my previous mail. I can only test it on
VMs, so the performance side needs testing by someone with a real
nvme-tcp setup. It is a bigger change and not something for stable.

Could this fix go into nvme-7.4, so that it can reach stable and the
distro kernels? Please let me know if anything is missing.

Thanks,
Seiji