On 27.02.20 10:20, David Hildenbrand wrote:
>> Yes. Changes to mm/gup.c really should normally go through linux-mm and
>> Andrew's tree, if at all possible. This would have been caught, and figured out
>> on linux-mm, had that been done--instead of leaving the linux-next maintainer
>> trying to guess at how to resolve the conflict.
>>
>> +Cc David Hildenbrand, who I see looked at the kvms390 proposed patch a bit.
>> Maybe he has some opinions, especially about my questions below.
>
> I'll leave figuring out the details to Christian/Claudio (-EBUSY) :)
>
>>
>> The fix-up below may (or may not) need some changes:
>>
>>
>> diff --cc mm/gup.c
>> index 354bcfbd844b,f589299b0d4a..000000000000
>> --- a/mm/gup.c
>> +++ b/mm/gup.c
>> @@@ -269,18 -470,11 +468,19 @@@ retry
>> goto retry;
>> }
>>
>> + /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
>> + if (unlikely(!try_grab_page(page, flags))) {
>> + page = ERR_PTR(-ENOMEM);
>> + goto out;
>> + }
>> + if (flags & FOLL_GET) {
>>
>>
>> If I'm reading the diff correctly, I believe that line should *maybe* be changed to:
>>
>> if (flags & (FOLL_GET | FOLL_PIN)) {
>>
>> ...because each of those flags has a similar effect: pinned pages for DMA or RDMA
>> use. So either flag will require a call to arch_make_page_accessible()...except that
>> I'm not sure that's what you want. Would the absence of a call to
>> arch_make_page_accessible() cause things like pin_user_pages() to not work correctly?
>> Seems like it would, to me.
>
> Yes, it's required. From the commit message "enable paging, file backing
> etc, it is also necessary to protect the host against a malicious user
> space. For example a bad QEMU could simply start direct I/O on such
> protected memory.". So we really want to convert the page from
> unencrypted/inaccessible to encrypted/accessible at this point (iow,
> make it definitely accessible, and make sure it stays accessible).
>
>>
>> (I'm pretty unhappy that we have to ask this at the linux-next level.)
>
> Yeah, I *think* this fell through the cracks (on linux-mm, but also in
> Andrew's inbox) because the series has a big fat "KVM: s390:" as prefix.
> Christian decided to pull it in to give it some churn yesterday (I think
> he originally wanted to have this patch and the other KVM protvirt
> patches in 5.7 [2] ... but not sure what will happen due to this conflict).
Yes, I would like to have this patch in 5.7. Depending on the schedule of the
FOLL_PIN patches that means:
1. Claudios callback patch _before_ the FOLL_PIN patches + Claudio will provide a fixup.
2. Claudios callback patch on top of the FOLL_PIN patches (Claudio will provide a
version that combines the first patch + fixup)
>
> At least now this patch has attention ... although it would have been
> better if linux-next admins wouldn't have to mess with this :)
>
> [1] https://lkml.kernel.org/r/20200224114107.4646-2-borntraeger@xxxxxxxxxx
> [2] https://lkml.kernel.org/r/20200224114107.4646-1-borntraeger@xxxxxxxxxx
>
