linux-kernel - Re: [PATCH 1/7] KVM: userspace interface

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:	Thu, 19 Oct 2006 14:31:36 -0500
From:	Anthony Liguori <aliguori@...ibm.com>
To:	Avi Kivity <avi@...ranet.com>
CC:	linux-kernel <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 1/7] KVM: userspace interface

Avi Kivity wrote:
> Anthony Liguori wrote:
>>>
>>> We started using VT only for 64 bit, then added 32 bit, then 16-bit 
>>> protected, then vm86 and real mode.  We'd transfer the x86 state on 
>>> each mode change, but it was (a) fragile (b) considered unclean.
>>>
>>> You're right that "big real" mode is not supported, but so far that 
>>> hasn't been a problem.  Do you know of an OS that needs big real mode?
>>
>> AFAIK the SLES boot splash patches to grub use it.  It's definitely a 
>> requirement.  Currently, there is an effort in Xen to use QEMU for 
>> partial emulation.  Hopefully, it will be there for the next release.
>>
>
> Ouch.  Boot splash patches.
>
> Well, we had real mode in qemu once, we can put it there again.
>
>> Allowing QEMU to do emulation also will help with IO performance.  
>> Instead of having to take many trips to userspace for MMIO 
>> especially, you can allow QEMU to execute a certain number of basic 
>> blocks and then return.  Minimizing trips between userspace and the 
>> kernel is going to be critical performance wise.
>>
>
> My plan was to allow userspace to register certain mmio addresses for 
> cacheing, so that if the guest code had a code sequence like
>
>  writel(dst_x_reg, x);
>  writel(dst_y_reg, y)
>  writel(width_reg, w);
>  writel(height_reg, h);
>  writel(blt_cmd_reg, fill);
>
> then kvm would cache the first four in a mmap()able memory area and 
> only exit to userspace on the fifth.  Userspace would then read the 
> cached registers from memory and emulate the command.

Letting QEMU do a certain amount of emulation after every transition 
would the problem in a more elegant and generic way.

> This saves userspace transitions but not guest/host switches.  I'm 
> counting on Intel and AMD to reduce the cost of these, but it will 
> probably never be cheap.
>
> Paravirtualized drivers are also an option; we may try to keep 
> compatibility with Xen's.

Please no!  With proper device emulation, paravirtualized drivers 
shouldn't be necessary.

>
>>
>> I've been tossing around the idea of doing partial IO emulation in 
>> the kernel.  If you could sync the device states between userspace 
>> and kernel, it should be possible.  Given that the you're already in 
>> the kernel at VMEXIT time, if you could feed something right to the 
>> block driver or network driver, you ought to be able to get pretty 
>> darn good performance.
>>
>
> Do you mean putting the device model into the kernel?

Perhaps part of it.  Still an idea at this point.

Regards,

Anthony Liguori
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/