Re: [PATCH] maintainers: claim responsability for xen

Hi  Thomas and all,

Apologize for being an unqualified maintainer.

-----Original Message-----
From: Thomas Monjalon [mailto:thomas@monjalon.net]
Sent: Friday, May 5, 2017 6:04 AM
To: Joao Martins; Konrad Rzeszutek Wilk; Tan, Jianfeng
Cc: Konrad Rzeszutek Wilk; dev@dpdk.org; Xen-devel
Subject: Re: [dpdk-dev] [PATCH] maintainers: claim responsability for xen

Ping

The Xen dom0 support in DPDK seems dead.

Reminder:
Last time we talked about, it was because of a severe bug which is not
fixed yet:
        http://dpdk.org/ml/archives/dev/2016-July/044207.html

For this bug, we removed the userspace memset(0) and suppose it has been done by kernel, however, xen0 uses __get_free_pages() kernel API to map hugepages and reseve memseg, I think it makes sense to zero the hugepage for xen0 in rte_dom0_mm kernel module (instead of some special code for xen0 in userspace) to keep aligned behavior.

        http://dpdk.org/ml/archives/dev/2016-July/044376.html

It does not make any sense to upstream a netfront PMD before we have a netback PMD, as the legacy netback driver would be the bottleneck. Anyone has plan on this? And a question mark keeps in my mind that is it a must to implement netback in dom0?

From another perspective, instead of using netfront/netback, we can also use virtio/vhost as the device model; however, xl tool in xen only supports vhost-kernel backend instead of vhost-user backend. So anyone has plan to enhance xl tool so that we can accelerate dom0 just using vswitch like OVS-DPDK?

A third solution is to use xenvirtio as the frontend, and vhost_xen as the backend. This solution is to use virtio ring on grant table mechanism of xen. Honestly, I don't even know if it still work now. And to make it more usable, better to upstream vhost_xen inside popular vswitch like OVS-DPDK.

The request (9 months ago) was to give more time for feedbacks:
        http://dpdk.org/ml/archives/dev/2016-July/044847.html

Apologize again that I volunteer to maintain these files, but spend very few time on this.

Thanks,
Jianfeng