Re: [PATCH v3 2/4] KVM: x86: Use vector-hashing to deliver lowest-priority interrupts

From: rkrcmar@xxxxxxxxxx
Date: Fri Jan 22 2016 - 08:49:26 EST

2016-01-22 12:00+0800, Yang Zhang:
> On 2016/1/22 1:21, rkrcmar@xxxxxxxxxx wrote:
>>(I think there isn't a practical difference between *r=-1 and *r=0.)
> Currently, if *r == -1, the remote_irr may get set. But it seems wrong. I

Yeah ...

> need to have a double check to see whether it is a bug in current code.

Looking forward to the patch!


>>'ret = true' is the better one. We know that the interrupt is not
>>deliverable [1], so there's no point in trying to deliver with the slow
>>path. We behave similarly when the interrupt targets a single disabled
>>1: Well ... it's possible that slowpath would deliver it thanks to
>> different handling of disabled APICs, but it's undefined behavior,
> why it is undefined behavior? Besides, why we will keep two different
> handling logic for the fast path and slow path? It looks weird.

It does look very weird ... the slow path would require refactoring,
though, so we save effort without a considerable drawback.
(I would love if it behaved identically, but I don't want to force it on
someone and likely won't do it myself ...)

I consider it undefined because SMD says that an OS musn't configure
this behavior and doesn't say what should happen if the OS does => we
could do anything. (Killing the guest would be great for debugging OS
issues, but ours behavior is fairly conservative.)