Petr Vandrovec wrote:
> Adar Dembo wrote:
>
>
>> I'm playing around with guestd in QEMU and I've noticed that iopl/ioperm
>>
>> aren't used by anything before doing PIO operations.
>>
>> I figure this works in VMware b/c you guys are intercepting the backdoor
>> io port regardless of CPL/IOPL. While this is useful for OSes like
>> Windows that have no way to change iopl, it would be nice on Posix
>> platforms if you did actually use iopl appropriately.
>>
>> It's a whole lot easier to just take a vmexit for PIO than it is to
>> intercept #gp and try to decode whether it was caused by a ring 3 PIO
>> instruction.
>>
>
> Hello Anthony,
> there are two reasons why we allow backdoor port to be accessed from
> CPL3:
>
> (1) We want to be able to access it even from non-suid applications -
> for example copy/paste daemon runs under normal user account, and so it
> cannot do iopl(3).
>
> (2) When not running in hardware assisted mode binary translation (or
> simulation) has to be used for CPL3 level code, which causes huge
> performance impact.
>
> If it is problem for qemu then perhaps creating kernel module to provide
> access to backdoor is simplest way to address problem, and it will be
> compatible with all VMware products as well.
Well, long term, I'd like to move to a virtio based socket. It's hard
to say how that would intersect with something like VMCI because I don't
know anything about it :-) I think in the interim, a simple root daemon
that allows backdoor operations via a domain socket or something would
suffice. Then non root users can still issue backdoor operations but
the daemon can still use iopl().
Regards,
Anthony Liguori
> Only restriction is that
> some of our backdoors are accessible to CPL0 code only already, so
> driver just cannot blindly issue backdoor call with registers it
> received from userspace, but that should not be complicated thing to
> address.
>
> For future products we want to use VMCI, which comes with regular kernel
> driver, and userspace library which can be accessed without IOPL
> elevation or any tricks in the emulation (well, except that it still
> does not use regular I/O instructions). Unfortunately latest released
> product (WS6) does not provide any interesting service over VMCI, and
> VMCI even did not make into this opensource release (you can take a look
> at WS6, but that one is *not* GPLed).
>
> You can add 'monitor_control.restrict_backdoor = "TRUE"' to VM's
> configuration file when using VMware, and then you should observe
> exactly same behavior you see with QEMU - guestd and everybody else
> crashing.
> Best regards,
> Petr Vandrovec
>
> -------------------------------------------------------------------------
> This SF.net email is sponsored by: Microsoft
> Defy all challenges. Microsoft(R) Visual Studio 2005.
> http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
> _______________________________________________
> open-vm-tools-devel mailing list
> ope...@li...
> https://lists.sourceforge.net/lists/listinfo/open-vm-tools-devel
>
>
|
