Re: [Xen-devel] [PATCH v1 04/10] vfio/type1: Prepare is_invalid_reserved_pfn() for PG_reserved changes

[Dan Williams <dan.j.williams@xxxxxxxxx>]

On Fri, Nov 8, 2019 at 2:22 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 08.11.19 08:14, David Hildenbrand wrote:
> > On 08.11.19 06:09, Dan Williams wrote:
> >> On Thu, Nov 7, 2019 at 2:07 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >>>
> >>> On 07.11.19 19:22, David Hildenbrand wrote:
> >>>>
> >>>>
> >>>>> Am 07.11.2019 um 16:40 schrieb Dan Williams <dan.j.williams@xxxxxxxxx>:
> >>>>>
> >>>>> On Thu, Oct 24, 2019 at 5:12 AM David Hildenbrand <david@xxxxxxxxxx> 
> >>>>> wrote:
> >>>>>>
> >>>>>> Right now, ZONE_DEVICE memory is always set PG_reserved. We want to
> >>>>>> change that.
> >>>>>>
> >>>>>> KVM has this weird use case that you can map anything from /dev/mem
> >>>>>> into the guest. pfn_valid() is not a reliable check whether the memmap
> >>>>>> was initialized and can be touched. pfn_to_online_page() makes sure
> >>>>>> that we have an initialized memmap (and don't have ZONE_DEVICE memory).
> >>>>>>
> >>>>>> Rewrite is_invalid_reserved_pfn() similar to kvm_is_reserved_pfn() to 
> >>>>>> make
> >>>>>> sure the function produces the same result once we stop setting 
> >>>>>> ZONE_DEVICE
> >>>>>> pages PG_reserved.
> >>>>>>
> >>>>>> Cc: Alex Williamson <alex.williamson@xxxxxxxxxx>
> >>>>>> Cc: Cornelia Huck <cohuck@xxxxxxxxxx>
> >>>>>> Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> >>>>>> ---
> >>>>>> drivers/vfio/vfio_iommu_type1.c | 10 ++++++++--
> >>>>>> 1 file changed, 8 insertions(+), 2 deletions(-)
> >>>>>>
> >>>>>> diff --git a/drivers/vfio/vfio_iommu_type1.c 
> >>>>>> b/drivers/vfio/vfio_iommu_type1.c
> >>>>>> index 2ada8e6cdb88..f8ce8c408ba8 100644
> >>>>>> --- a/drivers/vfio/vfio_iommu_type1.c
> >>>>>> +++ b/drivers/vfio/vfio_iommu_type1.c
> >>>>>> @@ -299,9 +299,15 @@ static int vfio_lock_acct(struct vfio_dma *dma, 
> >>>>>> long npage, bool async)
> >>>>>>     */
> >>>>>> static bool is_invalid_reserved_pfn(unsigned long pfn)
> >>>>>> {
> >>>>>> -       if (pfn_valid(pfn))
> >>>>>> -               return PageReserved(pfn_to_page(pfn));
> >>>>>> +       struct page *page = pfn_to_online_page(pfn);
> >>>>>
> >>>>> Ugh, I just realized this is not a safe conversion until
> >>>>> pfn_to_online_page() is moved over to subsection granularity. As it
> >>>>> stands it will return true for any ZONE_DEVICE pages that share a
> >>>>> section with boot memory.
> >>>>
> >>>> That should not happen right now and I commented back when you 
> >>>> introduced subsection support that I don’t want to have ZONE_DEVICE 
> >>>> mixed with online pages in a section. Having memory block devices that 
> >>>> partially span ZONE_DEVICE would be ... really weird. With something 
> >>>> like pfn_active() - as discussed - we could at least make this check 
> >>>> work - but I am not sure if we really want to go down that path. In the 
> >>>> worst case, some MB of RAM are lost ... I guess this needs more thought.
> >>>>
> >>>
> >>> I just realized the "boot memory" part. Is that a real thing? IOW, can
> >>> we have ZONE_DEVICE falling into a memory block (with holes)? I somewhat
> >>> have doubts that this would work ...
> >>
> >> One of the real world failure cases that started the subsection effect
> >> is that Persistent Memory collides with System RAM on a 64MB boundary
> >> on shipping platforms. System RAM ends on a 64MB boundary and due to a
> >> lack of memory controller resources PMEM is mapped contiguously at the
> >> end of that boundary. Some more details in the subsection cover letter
> >> / changelogs [1] [2]. It's not sufficient to just lose some memory,
> >> that's the broken implementation that lead to the subsection work
> >> because the lost memory may change from one boot to the next and
> >> software can't reliably inject a padding that conforms to the x86
> >> 128MB section constraint.
> >
> > Thanks, I thought it was mostly for weird alignment where other parts of
> > the section are basically "holes" and not memory.
> >
> > Yes, it is a real bug that ZONE_DEVICE pages fall into sections that are
> > marked SECTION_IS_ONLINE.
> >
> >>
> >> Suffice to say I think we need your pfn_active() to get subsection
> >> granularity pfn_to_online_page() before PageReserved() can be removed.
> >
> > I agree that we have to fix this. I don't like ZONE_DEVICE pages falling
> > into memory device blocks (e.g., cannot get offlined), but I guess that
> > train is gone :) As long as it's not for memory hotplug, I can most
> > probably live with this.
> >
> > Also, I'd like to get Michals opinion on this and the pfn_active()
> > approach, but I can understand he's busy.
> >
> > This patch set can wait, I won't be working next week besides
> > reading/writing mails either way.
> >
> > Is anybody looking into the pfn_active() thingy?
> >
>
> I wonder if we should do something like this right now to fix this
> (exclude the false positive ZONE_DEVICE pages we could have within an
> online section, which was not possible before subsection hotplug):
>
> diff --git a/include/linux/memory_hotplug.h b/include/linux/memory_hotplug.h
> index 384ffb3d69ab..490a9e9358b3 100644
> --- a/include/linux/memory_hotplug.h
> +++ b/include/linux/memory_hotplug.h
> @@ -30,6 +30,8 @@ struct vmem_altmap;
>          if (___nr < NR_MEM_SECTIONS && online_section_nr(___nr) && \
>              pfn_valid_within(___pfn))                              \
>                  ___page = pfn_to_page(___pfn);                     \
> +       if (unlikely(___page && is_zone_device_page(___page)))     \
> +               ___page = NULL;                                    \
>          ___page;                                                   \
>   })
>
>
> Yeah, it's another is_zone_device_page(), but it should not be racy
> here, as we want to exclude, not include ZONE_DEVICE.
>
> I don't have time to look into this right now, unfortunately.

I don't want to band-aid without an actual bug report. I'll take a
look at a subsection-map for the online state.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel