Re: [PATCH v1 2/3] mm/memory_hotplug: initialize memmap of !ZONE_DEVICE with PageOffline() instead of PageReserved()

[David Hildenbrand <david@xxxxxxxxxx>]

On 10.06.24 06:23, Oscar Salvador wrote:
> On Fri, Jun 07, 2024 at 11:09:37AM +0200, David Hildenbrand wrote:
> > We currently initialize the memmap such that PG_reserved is set and the
> > refcount of the page is 1. In virtio-mem code, we have to manually clear
> > that PG_reserved flag to make memory offlining with partially hotplugged
> > memory blocks possible: has_unmovable_pages() would otherwise bail out on
> > such pages.
> >
> > We want to avoid PG_reserved where possible and move to typed pages
> > instead. Further, we want to further enlighten memory offlining code about
> > PG_offline: offline pages in an online memory section. One example is
> > handling managed page count adjustments in a cleaner way during memory
> > offlining.
> >
> > So let's initialize the pages with PG_offline instead of PG_reserved.
> > generic_online_page()->__free_pages_core() will now clear that flag before
> > handing that memory to the buddy.
> >
> > Note that the page refcount is still 1 and would forbid offlining of such
> > memory except when special care is take during GOING_OFFLINE as
> > currently only implemented by virtio-mem.
> >
> > With this change, we can now get non-PageReserved() pages in the XEN
> > balloon list. From what I can tell, that can already happen via
> > decrease_reservation(), so that should be fine.
> >
> > HV-balloon should not really observe a change: partial online memory
> > blocks still cannot get surprise-offlined, because the refcount of these
> > PageOffline() pages is 1.
> >
> > Update virtio-mem, HV-balloon and XEN-balloon code to be aware that
> > hotplugged pages are now PageOffline() instead of PageReserved() before
> > they are handed over to the buddy.
> >
> > We'll leave the ZONE_DEVICE case alone for now.
> >
> > Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> > diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> > index 27e3be75edcf7..0254059efcbe1 100644
> > --- a/mm/memory_hotplug.c
> > +++ b/mm/memory_hotplug.c
> > @@ -734,7 +734,7 @@ static inline void section_taint_zone_device(unsigned long 
> > pfn)
> >   /*
> >    * Associate the pfn range with the given zone, initializing the memmaps
> >    * and resizing the pgdat/zone data to span the added pages. After this
> > - * call, all affected pages are PG_reserved.
> > + * call, all affected pages are PageOffline().
> >    *
> >    * All aligned pageblocks are initialized to the specified migratetype
> >    * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
> > @@ -1100,8 +1100,12 @@ int mhp_init_memmap_on_memory(unsigned long pfn, 
> > unsigned long nr_pages,
> >   
> >   	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
> >   
> > -	for (i = 0; i < nr_pages; i++)
> > -               SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
> > +       for (i = 0; i < nr_pages; i++) {
> > +               struct page *page = pfn_to_page(pfn + i);
> > +
> > +               __ClearPageOffline(page);
> > +               SetPageVmemmapSelfHosted(page);
> So, refresh my memory here please.
> AFAIR, those VmemmapSelfHosted pages were marked Reserved before, but now,
> memmap_init_range() will not mark them reserved anymore.
Correct.

> I do not think that is ok? I am worried about walkers getting this wrong.
>
> We usually skip PageReserved pages in walkers because are pages we cannot deal
> with for those purposes, but with this change, we will leak
> PageVmemmapSelfHosted, and I am not sure whether are ready for that.
There are fortunately not that many left.

I'd even say marking them (vmemmap) reserved is more wrong than right: 
note that ordinary vmemmap pages after memory hotplug are not reserved! 
Only bootmem should be reserved.
Let's take at the relevant core-mm ones (arch stuff is mostly just for 
MMIO remapping)
fs/proc/task_mmu.c:     if (PageReserved(page))
fs/proc/task_mmu.c:     if (PageReserved(page))

-> If we find vmemmap pages mapped into user space we already messed up
   seriously

kernel/power/snapshot.c:        if (PageReserved(page) ||
kernel/power/snapshot.c:        if (PageReserved(page)

-> There should be no change (saveable_page() would still allow saving
   them, highmem does not apply)

mm/hugetlb_vmemmap.c:           if (!PageReserved(head))
mm/hugetlb_vmemmap.c:   if (PageReserved(page))

-> Wants to identify bootmem, but we exclude these
   PageVmemmapSelfHosted() on the splitting part already properly


mm/page_alloc.c:                VM_WARN_ON_ONCE(PageReserved(p));
mm/page_alloc.c:                if (PageReserved(page))

-> pfn_range_valid_contig() would scan them, just like for ordinary
   vmemmap pages during hotplug. We'll simply fail isolating/migrating
   them similarly (like any unmovable allocations) later

mm/page_ext.c:          BUG_ON(PageReserved(page));

-> free_page_ext handling, does not apply

mm/page_isolation.c:            if (PageReserved(page))

-> has_unmovable_pages() should still detect them as unmovable (e.g.,
   neither movable nor LRU).

mm/page_owner.c:                        if (PageReserved(page))
mm/page_owner.c:                        if (PageReserved(page))

-> Simply page_ext_get() will return NULL instead and we'll similarly
   skip them

mm/sparse.c:            if (!PageReserved(virt_to_page(ms->usage))) {

-> Detecting boot memory for ms->usage allocation, does not apply to
   vmemmap.

virt/kvm/kvm_main.c:    if (!PageReserved(page))
virt/kvm/kvm_main.c:    return !PageReserved(page);

-> For MMIO remapping purposes, does not apply to vmemmap


> Moreover, boot memmap pages are marked as PageReserved, which would be
> now inconsistent with those added during hotplug operations.
Just like vmemmap pages allocated dynamically during memory hotplug. 
Now, really only bootmem-ones are PageReserved.
> All in all, I feel uneasy about this change.
I really don't want to mark these pages here PageReserved for the sake 
of it.
Any PageReserved user that I am missing, or why we should handle these 
vmemmap pages differently than the ones allocated during ordinary memory 
hotplug?
In the future, we might want to consider using a dedicated page type for 
them, so we can stop using a bit that doesn't allow to reliably identify 
them. (we should mark all vmemmap with that type then)
Thanks!

--
Cheers,

David / dhildenb