Here's one important thing to notice:
If the "fast start-up" (fast boot) is enabled in windows, the mlx vpi device will get the error 43, requiring reboot or disable/enable toggling. That is, with a patched kernel driver or a working linux ofed (if such a thing even exists anymore).
The microsoft driver should work, so it points in the direction of linux ofed :(
I did get that error 10, when I had winof driver on the VPI and microsoft one on the ethernet device.
Yes, after uninstalling them do the hw scan from the menu. The thing is, if you have driver version mismatch between the vpi & ethernet device, it will not work. So uninstall + remove driver and begin with the ethernet adapter.
@mimino
I think I know what's happening with your error code 10. There's two devices in windows, vpi and ethernet. If the VPI device is loaded with microsoft provided driver and the ethernet has winof driver, it'll give you the error code 10.
Uninstall the devices and check the delete driver...
This is very weird, as if Mellanox (nvidia) is trying to sabotage their older products. I downloaded the 4.9-3.1.5.0 LTS. By looking at the source code, it seems incompatible with the Windows driver. :oops:
Well, you've got the LTS version of OFED. The latest is 5.3-1, where yours is the LTS version 4.9-3 ?
Yes, it should work with 5.4 also!
I just realised, they've dropped support for connectx-3 on 5.1. Kind of weird :oops:
It's either patch the stock kernel driver or install the latest & greatest ofed. :(
That vhcr cmd, it failed executing, was MLX4_CMD_ALLOC_RES. So, before doing anything, I'd try that param.
I see a problem there, the windows driver doesn't play nice with that linux driver version.
You could try adding the log_num_mgm_entry_size=-1 parameter to mlx4_core.
I'd expect it to fail on port query, though.
Are you using some non-default parameters for the mlx4_core module?
Here's mine:
msi_x=1 enable_4k_uar=1 enable_qos=1 log_num_mac=7 log_num_mgm_entry_size=-1 log_mtts_per_seg=4
I did it last summer so, no. It was just a quick way to get it to work. It'd require extensive testing & cleaning up.
For me, it's been working fine.
<interface type='hostdev' managed='yes'>
<source>
<address type='pci' domain='0x0000' bus='0x83' slot='0x00' function='0x1'/>...
Are you trying to pass the main pci-device or a vf?
You should disable the host/vmware driver (nmlx4-core&net) anyways, it has no role in passthrough, if it's not a vf. It's probably preventing it.
The cause is right in front of you. The actual problem is the OpenZFS code quality. It uses the Solaris Porting Layer, which is basically a not-so-great wrapper around linux kernel threading api.
- It doesn't know NUMA at all
- It doesn't understand cache locality nor hierarchy
- It does...
Here's a dirty patch for the old (5.5.15) mainstream kernel driver.
It might apply successfully to later kernels as well.
From d4fbe94b4adb16a5cc094a50215dc39496c9fc66 Mon Sep 17 00:00:00 2001
From: Mikko Hissa <okrasit@gmail.com>
Date: Tue, 28 Jul 2020 05:35:13 +0000
Subject: [PATCH] MLX4...
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.