[PATCH v2 0/8] mm/memory_hotplug: Factor out memory block device handling

[Dan Williams]

On Tue, May 7, 2019 at 12:38 PM David Hildenbrand <david at redhat.com> wrote:
>
> On 07.05.19 21:21, David Hildenbrand wrote:
> > On 07.05.19 21:04, Dan Williams wrote:
> >> On Tue, May 7, 2019 at 11:38 AM David Hildenbrand <david at redhat.com> wrote:
> >>>
> >>> We only want memory block devices for memory to be onlined/offlined
> >>> (add/remove from the buddy). This is required so user space can
> >>> online/offline memory and kdump gets notified about newly onlined memory.
> >>>
> >>> Only such memory has the requirement of having to span whole memory blocks.
> >>> Let's factor out creation/removal of memory block devices. This helps
> >>> to further cleanup arch_add_memory/arch_remove_memory() and to make
> >>> implementation of new features easier. E.g. supplying a driver for
> >>> memory block devices becomes way easier (so user space is able to
> >>> distinguish different types of added memory to properly online it).
> >>>
> >>> Patch 1 makes sure the memory block size granularity is always respected.
> >>> Patch 2 implements arch_remove_memory() on s390x. Patch 3 prepares
> >>> arch_remove_memory() to be also called without CONFIG_MEMORY_HOTREMOVE.
> >>> Patch 4,5 and 6 factor out creation/removal of memory block devices.
> >>> Patch 7 gets rid of some unlikely errors that could have happened, not
> >>> removing links between memory block devices and nodes, previously brought
> >>> up by Oscar.
> >>>
> >>> Did a quick sanity test with DIMM plug/unplug, making sure all devices
> >>> and sysfs links properly get added/removed. Compile tested on s390x and
> >>> x86-64.
> >>>
> >>> Based on git://git.cmpxchg.org/linux-mmots.git
> >>>
> >>> Next refactoring on my list will be making sure that remove_memory()
> >>> will never deal with zones / access "struct pages". Any kind of zone
> >>> handling will have to be done when offlining system memory / before
> >>> removing device memory. I am thinking about remove_pfn_range_from_zone()",
> >>> du undo everything "move_pfn_range_to_zone()" did.
> >>>
> >>> v1 -> v2:
> >>> - s390x/mm: Implement arch_remove_memory()
> >>> -- remove mapping after "__remove_pages"
> >>>
> >>>
> >>> David Hildenbrand (8):
> >>>   mm/memory_hotplug: Simplify and fix check_hotplug_memory_range()
> >>>   s390x/mm: Implement arch_remove_memory()
> >>>   mm/memory_hotplug: arch_remove_memory() and __remove_pages() with
> >>>     CONFIG_MEMORY_HOTPLUG
> >>>   mm/memory_hotplug: Create memory block devices after arch_add_memory()
> >>>   mm/memory_hotplug: Drop MHP_MEMBLOCK_API
> >>
> >> So at a minimum we need a bit of patch staging guidance because this
> >> obviously collides with the subsection bits that are built on top of
> >> the existence of MHP_MEMBLOCK_API. What trigger do you envision as a
> >> replacement that arch_add_memory() use to determine that subsection
> >> operations should be disallowed?
> >>
> >
> > Looks like we now have time to sort it out :)
> >
> >
> > Looking at your series
> >
> > [PATCH v8 08/12] mm/sparsemem: Prepare for sub-section ranges
> >
> > is the "single" effectively place using MHP_MEMBLOCK_API, namely
> > "subsection_check()". Used when adding/removing memory.
> >
> >
> > +static int subsection_check(unsigned long pfn, unsigned long nr_pages,
> > +             unsigned long flags, const char *reason)
> > +{
> > +     /*
> > +      * Only allow partial section hotplug for !memblock ranges,
> > +      * since register_new_memory() requires section alignment, and
> > +      * CONFIG_SPARSEMEM_VMEMMAP=n requires sections to be fully
> > +      * populated.
> > +      */
> > +     if ((!IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP)
> > +                             || (flags & MHP_MEMBLOCK_API))
> > +                     && ((pfn & ~PAGE_SECTION_MASK)
> > +                             || (nr_pages & ~PAGE_SECTION_MASK))) {
> > +             WARN(1, "Sub-section hot-%s incompatible with %s\n", reason,
> > +                             (flags & MHP_MEMBLOCK_API)
> > +                             ? "memblock api" : "!CONFIG_SPARSEMEM_VMEMMAP");
> > +             return -EINVAL;
> > +     }
> > +     return 0;
> >  }
> >
> >
> > (flags & MHP_MEMBLOCK_API)) && ((pfn & ~PAGE_SECTION_MASK) || (nr_pages
> > & ~PAGE_SECTION_MASK)))
> >
> > sounds like something the caller (add_memory()) always has to take care
> > of. No need to check. The one imposing this restriction is the only caller.
> >
> > In my opinion, that check/function can go completely.
> >
> > Am I missing something / missing another user?
> >
>
> In other word, this series moves the restriction out of
> arch_add_memory() and therefore you don't need subsection_check() at all
> anymore. At least if I am not missing something :)

Ah, ok. Only direct arch_add_memory() users need to be worried about
subsection hotplug and the add_memory_resource() + __remove_memory()
paths are already protected by check_hotplug_memory_range(). Ok, I can
get on board with the removal.

Let me go ahead and review this series so Andrew can get it pulled in
and I can rebase.