Re: [patch 4/9] x86/io: Speedup schedule out of I/O bitmap user

From: Andy Lutomirski
Date: Fri Nov 08 2019 - 17:41:41 EST


On 11/7/19 6:08 AM, Thomas Gleixner wrote:
> On Thu, 7 Nov 2019, Thomas Gleixner wrote:
>> On Thu, 7 Nov 2019, Peter Zijlstra wrote:
>>> On Wed, Nov 06, 2019 at 08:35:03PM +0100, Thomas Gleixner wrote:
>>>> There is no requirement to update the TSS I/O bitmap when a thread using it is
>>>> scheduled out and the incoming thread does not use it.
>>>>
>>>> For the permission check based on the TSS I/O bitmap the CPU calculates the memory
>>>> location of the I/O bitmap by the address of the TSS and the io_bitmap_base member
>>>> of the tss_struct. The easiest way to invalidate the I/O bitmap is to switch the
>>>> offset to an address outside of the TSS limit.
>>>>
>>>> If an I/O instruction is issued from user space the TSS limit causes #GP to be
>>>> raised in the same was as valid I/O bitmap with all bits set to 1 would do.
>>>>
>>>> This removes the extra work when an I/O bitmap using task is scheduled out
>>>> and puts the burden on the rare I/O bitmap users when they are scheduled
>>>> in.
>>>
>>> This also nicely aligns with that the context switch time is accounted
>>> to the next task. So by doing the expensive part on switch-in gets it
>>> all accounted to the task that caused it.
>>
>> Just that I can't add the storage to tss_struct due to the VMX insanity of
>> setting TSS limit hard to 0x67 on vmexit instead of restoring the host
>> value.
>
> Well, I can. The build bugon in vmx.c is just bogus.

SDM vol 3 27.5.2 says the BUILD_BUG_ON is right. Or am I
misunderstanding you?

I'm reasonably confident that the TSS limit is indeed 0x67 after VM
exit, and I wrote the existing code that tries to optimize this to avoid
LTR when not needed.