Re: setting up CDB filters in udev (was Re: [PATCH v2 0/3] block: add queue-private command filter, editable via sysfs)

[Paolo Bonzini <pbonzini@redhat.com>]

Ping?

Paolo

Il 25/10/2012 20:35, Paolo Bonzini ha scritto:
>> On Thu, Oct 25, 2012 at 09:37:39AM +0200, Paolo Bonzini wrote:
>>> Il 24/10/2012 18:47, Tejun Heo ha scritto:
>>>> So, I'm still not convinced we need to go forward with full
>>>> configurability. All use cases you described can be covered with
>>>> per-class static filters + simple override switch to disable all,
>>>> which would result in a lot simpler implementation w/ much
>>>> smaller userland interface.
>>>
>>> I'm not sure the userland interface would be smaller, and it would
>>> be more complex to get right:
>>>
>>> 1) how do you override the default?  ioctl+SCM_RIGHTS or sysfs?
>>
>> Disabling filters if opened by root and tranfering via SCM_RIGHTS
>> would be the simplest interface-wise (there's no new interface at
>> all).  Would that be too dangerous security-wise?
> 
> That would be a change with respect to what we have now.  After
> transferring a root-opened (better: CAP_SYS_RAWIO-opened) file
> descriptor to an unprivileged process your SG_IO commands get
> filtered.  So a ioctl is needed if you want to rely on SCM_RIGHTS.
> 
>>> 2) do you need to override the default to "no access", "full
>>> access" and "default access", or is a binary knob (default
>>> access/full access) sufficient?
>>
>> Default / full should be enough, no?
> 
> If a ioctl has to be added, I'd rather have at least none/full/default.
> 
>>> 3) what capabilities control the setting?
>>
>> CAP_SYS_RAWIO seems to be a pretty good fit.
> 
> Yes, for a ioctl it is (for sysfs CAP_SYS_ADMIN is better IMHO).
> 
>> I guess I just feel quite reluctant to expose another rather obscure
>> userland configurable in-kernel filter and at the same time I'm not
>> sure whether this is flexible enough.  What if a device is shared by
>> multiple virtual machines which are trusted at different levels?
> 
> No, you just don't do that.  If a device is passed through to virtual
> machines, it is between similar virtual machines (for some definition
> of similar).  The only case where you have this sharing is in practice
> if either the device is read-only (my patch does give you a basic
> two-level filtering, with two separate bitmaps for RO and RW) or if you
> allow persistent reservations (which is as close to full trust as you
> can get).
> 
>> I'm not trying to block it at all cost but let's make sure we looked
>> into most possibilities before (re)adding this userland visible
>> interface.
> 
> Sure, understood. :)
> 
>> Jens, James, what do you guys think?
> 
> Paolo
>