Sagi,

Just wanted to give you an update on where we're at with this.

All tests run with your earlier patch removing the inline dispatch from
nvme_tcp_queue_request() are successful. At this point, I am leaning to
remove that optimization from mainline.

Thanks Keith,

Did you run it with the extra information debug patch I sent you? What
I'm concerned about is that given that you have the only environment
where this reproduces, and this is removed it will be very difficult
to add it back in.

Also, what about the read issue? that one is still unresolved from
my PoV.

I added additional tracing to see what is going on, but we eventually
hit a memory issue after some hours of runtime. I've never seen an issue
like this before, It triggers in nvme_tcp_advance_req() when tracing the
rq->tag and req->data_sent:

   WARNING: CPU: 1 PID: 3428 at arch/x86/include/asm/kfence.h:44 kfence_protect_page+0x33/0xa0

I think the above is a distraction, but I can provide the full stack
trace and patch adding the tracepoing if you think it's helpful.

That is... odd..

_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme