echo: write error: Invalid argument while setting io_poll for NVMe SSD

1
votes

I am using the following command as a root user to enable Linux kernel polling for a NVMe SSD device.

# echo 1 > /sys/block/nvme2n1/queue/io_poll

I get the following error: bash: echo: write error: Invalid argument

OS details : Ubuntu 16.04, Linux kernel 5.0.0+.

Any pointers?

References for the command: https://itpeernetwork.intel.com/tuning-performance-intel-optane-ssds-linux-operating-systems/#gs.1yu3o1

https://lwn.net/Articles/663543/

2
linuxlinux-kernelubuntu-16.04sysfsnvme
This looks like a usage question not a development one. You should probably read the documentation, or if approaching this from a developer perspective find and read the code that generates that error and try to figure out if the meaning is that you are asking in the wrong way, or that you are asking for something which is not supported. - Chris Stratton
The command works for other machines which are using kernel 4.15. The usage is correct, as also mentioned in the references. - Nikhil Sambhus
That doesn't mean it's going to work with this particular hardware or its drivers. - Chris Stratton

2 Answers

1
votes

Nikhil posted "Error while enabling io_poll for NVMe SSD" to the Linux block mailing list and got a reply from Keith Busch:

Be sure to turn on the polling queues in the nvme driver. There are none by default. The kernel parameter to that [sic] enable them is:

nvme.poll_queues=X

Where 'X' is the number of polling queues. I'd recommend at least 1 per CPU socket, but more is better.

From the looks of Nikhil's reply, it seems io_poll could be set to 1 after the above change had been made.

For context, browsing the kernel source you can the 5.0 kernel introduced the NVMe poll_queues option (this is also alluded to in the storage section of Kernel Newbies 5.0 kernel changelog). Further, a 5.0 commit defaulted poll_queues in to 0 (at the time of writing in late 2019 we're currently up to 5.5-rc2 and this is still the case).

Further, it looks like Frank Ober (who posted an answer to this question as @FrankO) asked "why the change was made" in the "Polled io for Linux kernel 5.x" Linux block mailing list thread and also received a reply from Keith:

The original polling implementation shared resources that generate interrupts. This prevents it from running as fast as it can, so dedicated polling queues are used now.

0
votes

This blog you are referencing was born in the time of 4.x kernels, since then io_uring IO submission concepts have come about in the 5.x kernel timeframe.

Please read this document about io_uring: https://kernel.dk/io_uring.pdf

You should test a Optane media drive with: ioengine=io_uring

With or without hipri in case you want to use bio_poll or not.

Here is something that should work as far as passing to fio:

[global]
direct=1
filename=/dev/nvme1n1
log_avg_msec=500
time_based
percentile_list=1:5:10:20:30:40:50:60:70:80:90:95:99:99.5:99.9:99.95:99.99:99.999:99.9999


[rand-read-4k-qd1]
runtime=120
bs=4K
iodepth=1
numjobs=1
cpus_allowed=0
ioengine=io_uring
hipri
rw=randread

uname -a 5.4.1-1.el8.elrepo.x86_64

CentOS 8

Have fun, Frank Ober