Re: [PATCH 0 of 4] Add commands to automatically prep devices for pass-through

[George Dunlap <george.dunlap@eu.citrix.com>]

On 09/05/12 12:59, Ian Campbell wrote:
> Right, however it is strictly speaking a new feature which is not
> mentioned on the TODO list and has not previously been posted (AFAIK,
> please correct me if not) and we are currently supposed to be in feature
> freeze (and have been for several weeks, if not a month).
>
> IIRC this functionality was mooted when the pci permissive patch was
> being done as something which would be a 4.3 feature.
> We need to decide if we want to make an exception for this new feature
> or not. Although I'm sure this feature is very nice and handy, we've
> lived without it for years and people seem to be able to use the
> existing scheme.
My recollection was that I did "moot" the functionailty basically a day 
or two after the official feature freeze; my impression from those 
discussions was that we wouldn't add the feature to the list, but that 
it was reasonable to ask for an exception at such time as I actually had 
the patches.  (Quite possible that my understanding is wrong there.)  
Unfortunately due to other priorities, I didn't manage to actually start 
working on them until the end of last week.

Maybe part of the issue is how they're being presented.  My original 
plan was to add options to libxl_pci_{add,remove} do the rebinding, 
which would have looked less like a new feature and more like an 
improvement.  This version actually introduces new functions, so it 
looks much more like a "new feature", even though the functionality is 
the same, and arguably having a separate step is less of a risk of 
someone tripping over something.

Of course everyone thinks their pet feature is incredibly important. 
:-)  But we are planning on making a public push on some of the security 
features of Xen this summer, which will hopefully mean a lot of people 
investigate the idea of using pci pass-through functionality for network 
driver domains.  The problem with saying "people seem to be able to use 
the existing scheme" is that you only see those who have gone through it 
and succeeded; you don't see how many took at look at the instructions 
and said, "That sounds too complicated/dangerous for me."  It would be a 
shame if we tooted Xen's horn about security, got an extra several 
thousand people to look into it, and then had half of them go away 
because of something simple like this.  I think that's my main concern.

We could of course make the HOWTOs easier to follow even without 
including this functionality; including Anthony's (very useful) 
rebinding script would certainly be a lot better than having everyone 
manually doing the sysfs stuff.  But not nearly as good as having the 
commands in-tree.

If we decide not to take the new functions, can I propose that we at 
least take the one that renames "pci-device-list-assignable", so we 
won't have to rename it / deal with compatibility issues when these are 
implemented for 4.3?

  -George