Messages in this thread | | | Date | Fri, 24 May 2013 10:31:34 +0200 | From | Paolo Bonzini <> | Subject | Re: PING^7 (was Re: [PATCH v2 00/14] Corrections and customization of the SG_IO command whitelist (CVE-2012-4542)) |
| |
Il 24/05/2013 10:02, Tejun Heo ha scritto: > On Fri, May 24, 2013 at 4:13 PM, Paolo Bonzini <pbonzini@redhat.com> wrote: >>> The same filtering table being applied to different classes of >>> hardware is a software bug, but my point is that the practive >>> essentially entrusts non-insignificant part of security enforcement to >>> the hardware itself. The variety of hardware in question is very wide >>> and significant portion has historically been known to be flaky. >> >> Unproven theory, and contradicted by actual practice. Bugs are more >> common in the handling of borderline conditions, not in the handling of >> unimplemented commands. >> >> If you want to be secure aginst buggy firmware, the commands you have to >> block are READ and WRITE. Check out the list of existing USB quirks. > > Well, I'd actually much prefer disabling CDB whitelisting for all !MMC > devices if at all possible. You're basically arguing that because what > we have is already broken, it should be okay to break it further. > Also, RW commands having more quirks doesn't necessarily indicate that > they tend to be more broken. They just get hammered on a lot in > various ways so problems on those commands tend to be more noticeable.
I agree intuition may not count, and it's perfectly possible that firmware writers forgot a "break;" or put the wrong location in a jump table, so that unimplemented commands give interesting results.
However, the _fact_ is that this might happen anyway with the buttload of commands that are already enabled by the whitelist and that most disks will never implement.
>> You need to allow more commands. >> The count-me-out knob allows all commands. >> You cannot always allow all commands. >> Ergo, you cannot always use the count-me-out knob. > > The thing is that both approaches aren't perfect here so you can make > similar type of argument from the other side. If the system wants to > give out raw hardware access to VMs, requiring it to delegate the > device fully could be reasonable.
No, it is not unfortunately. Allowing to do discards is one thing, allowing to disrupt the settings of a SAN is another. You can only delegate the device fully in these cases:
(a) of course, if the guest is trusted;
(b) if QEMU is running as a confined user, then you can get by with a userspace whitelist. (Otherwise you're just a ptrace away from arbitrary access, as you pointed out).
Unfortunately, there are _real_ problems that this patches fix, such as log spews due to blocked commands, and these happen even if neither of the above conditions is true.
Is there anything else I can do? Sure, I can check for the presence of the whitelist and hack the VPD pages to hide features that I know the whitelist will block. Is it the right thing to do? In my opinion no. It makes no sense to have raw device access _disable_ features compared to emulation.
> Not ideal but widening direct > command access by default is pretty nasty too.
I actually agree, and that's why I'm trying to balance the widening and restricting.
Paolo
| |