RE: [PATCH] intel-iommu: Set status bit after operation completed

[Duan, Zhenzhong]

Hi Peter

>-----Original Message-----
>From: Peter Xu <peterx@redhat.com>
>Sent: Thursday, March 9, 2023 10:56 PM
>To: Duan, Zhenzhong <zhenzhong.duan@intel.com>
>Cc: qemu-devel@nongnu.org; mst@redhat.com; jasowang@redhat.com;
>pbonzini@redhat.com; richard.henderson@linaro.org; eduardo@habkost.net;
>marcel.apfelbaum@gmail.com
>Subject: Re: [PATCH] intel-iommu: Set status bit after operation completed
>
>Hi, Zhenzhong,
>
>On Thu, Mar 09, 2023 at 05:23:19PM +0800, Zhenzhong Duan wrote:
>> According to SDM 11.4.4.2 Global Status Register:
>> "This field is cleared by hardware when software sets the SRTP field
>> in the Global Command register. This field is set by hardware when
>> hardware completes the ‘Set Root Table Pointer’ operation using the
>> value provided in the Root Table Address register"
>>
>> Follow above spec to clear then set RTPS after finish all works, this
>> way helps avoiding potential race with guest kernel. Though linux
>> kernel is single threaded in writing GCMD_REG and checking GSTS_REG.
>>
>> Same reasion for GSTS_REG.TES
>
>Is this a real bug?  Or, when it'll make a difference to the guest?
Not a real bug, just for robustness. Sorry, I should have made it clear in 
comments.

I think it may break with special designed guest OS,
E.x: Imagine a guest write GCMD_REG and start a new thread to do further work.
New thread find status bit in GTS_REG set and go ahead, but the address space 
switch
may not finish yet if guest memory is big, which may trigger a potential race.

Original code lives for a long history so it looks all the practiced os aren't 
designed as above.

Thanks
Zhenzhong