Re: 答复: [????] Re: [PATCH] KVM: irqchip: allocate routing entries in chunks

[Yanfei Xu]

On 2026/9/10 11:26, Li,Rongqing wrote:
> > Please look at Yanfei's series and help come to an agreement on how best to fix
> > this.  I am trying to get to Yanfei's series, and normally would take a close look
> > at both, but I am extremely short on cycles at the moment.
> >
> > https://lore.kernel.org/all/20260525035242.107264-1-yanfei.xu@bytedance.
> > com
> Hi Sean,
>
> Thanks for pointing me to Yanfei's series. I agree that both approaches
> aim to reduce the allocation overhead in KVM_SET_GSI_ROUTING().
>
> One main concern with using kvalloc() for the entire routing entry array is
> that, under severe memory fragmentation, it may fall back to vmalloc().
> The vmalloc() path can have relatively high and less predictable overhead,
> especially when routing tables are updated frequently when guest OS is
> running irqbalance.
Hi Rongqing,

You are right, fallbacking to vmalloc is not friendly to frequently
updating senarios. I assume Sean also noticed that and suggested to
introduce the patch2 at v2. It could avoid to allocate memory when
the rounded size of total entries is same with previous cached
allocation. I think irqblance could benefit from patch2.

Regards,
Yanfei

> Thanks
> [Li,Rongqing]