On 3 November 2021 13:05:11 GMT, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>On 11/3/21 13:56, David Woodhouse wrote:
>>> No need to resubmit, thanks! I'll review the code later and
>>> decide whether to include this in 5.16 or go for the "good"
>>> solution in 5.16 and submit this one for 5.15 only.
>> I would call this the good solution for steal time. We really do
>> always have a userspace HVA for that when it matters, and we should
>> use it.
>>
>> For Xen event channel delivery we have to do it from hardware
>> interrupts under arbitrary current->mm and we need a kernel mapping,
>> and we need the MMU notifiers and all that stuff. But for every
>> mapping we do that way, we need extra checks in the MMU notifiers.
>>
>> For steal time there's just no need.
>
>Yes, but doing things by hand that it is slightly harder to get right,
>between the asm and the manual user_access_{begin,end}.
Right. When I embarked on this I had a fantasy that I could just use the futex asm helpers which do most of it for us. But it didn't turn out that way. On the other hand, it's only needed for the *atomic* accesses (xchg, bit set) and most of the time we can just use normal uaccess stuff (and remember to mark the gfn dirty!)
>The good solution would be to handle the remapping of _all_ gfn-to-pfn
>caches from the MMU notifiers, so that you can still do map/unmap, keep
>the code simple, and get for free the KVM-specific details such as
>marking the gfn as dirty.
>
>When I was working on it before, I got stuck with wanting to do it not
>just good but perfect, including the eVMCS page in it. But that makes
>no sense because really all that needs to be fixed is the _current_
>users of the gfn-to-pfn cache.
Yeah.
Well, let's take a look at how I've done it for the Xen event channel delivery, in particular the rwlock that has to be held *while* doing the access via the mapped kernel address. Then we can ponder whether we want to offer something along those lines as a generic facility.
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.
