Re: efinet SNP issue affects iscsi boot

[Micah Parrish]

On 09/20/2018 07:47 AM, address@hidden wrote:

On Thu, Sep 20, 2018 at 06:38:07PM +0800, Michael Chang wrote:

On Thu, Sep 13, 2018 at 06:06:15PM -0600, Micah Parrish wrote:

Hi, new subscriber here.? We have a problem with Grub 2 and its use of SNP
instead of MNP.? Our UEFI driver for a network card parses the relevant DHCP
options for iSCSI boot, generates an iBFT table, then gets closed by Grub
when it opens the SNP interface. The driver removes the iBFT table as part
of normal unload cleanup.? I think this should happen with the Tianocore
UEFI reference driver as well.? The problem is often masked or does not
occur when there are multiple network ports enabled.? It occurs with several
different vendors NICs.

Possible solutions I see:

1. Grub parses the DHCP options and creates its own iBFT.

2. Grub copies the already generated iBFT before destroying the interface.

3. Grub opens the network interface MNP instead of SNP.

Although I am a neophyte at grub and UEFI development, I would like to start
a discussion on possible solutions.? Has anyone else seen this?

For possible solution 3, I managed to work out patch to use MNP but is not
polished, although it survived my testing. If you don't mind and willing to
give it go I will post it here as RFC patch for review.

That would be perfect. However, there are a few things worth mentioning here.

The issue is never ending story. So, please look for relevant discussions
in grub-devel archives and take them into account if it is possible/make sense.
If you have any difficulties with finding them drop me a line.

Please do not drop SNP driver. I think that we should make MNP driver a new
default and SNP should stay as a backup. Just in case.

Additionally, a few days ago I have started looking for people interested
in the project. There are some. Hence, if you are going to take a stab at
it I will ask them to do some reviews of your work. I will drop you their
emails if they are happy to do so.

Thank you for the context.  I read through the archives, searching for SNP.  I'm willing to test whatever we come up with. 

I also read through the efinet.c from OpenSolaris.  It is quite different from the current SNP efinet.c, since it uses a ring buffer and filters.  I assume a naive switch over to MNP might run into the same lost packet problem seen before.  Was it seen across many platforms, or specific to a few?

-Micah