Re: [PATCH stable 5.10] mm/memory_hotplug: extend offline_and_remove_memory() to handle more than one memory block

[David Hildenbrand]

On 19.06.23 09:22, mawupeng wrote:
> On 2023/6/19 15:16, Greg KH wrote:
> > On Mon, Jun 19, 2023 at 02:51:21PM +0800, Wupeng Ma wrote:
> > > From: David Hildenbrand <david@xxxxxxxxxx>
> > >
> > > commit 8dc4bb58a146655eb057247d7c9d19e73928715b upstream.
> > >
> > > virtio-mem soon wants to use offline_and_remove_memory() memory that
> > > exceeds a single Linux memory block (memory_block_size_bytes()). Let's
> > > remove that restriction.
> > >
> > > Let's remember the old state and try to restore that if anything goes
> > > wrong. While re-onlining can, in general, fail, it's highly unlikely to
> > > happen (usually only when a notifier fails to allocate memory, and these
> > > are rather rare).
> > >
> > > This will be used by virtio-mem to offline+remove memory ranges that are
> > > bigger than a single memory block - for example, with a device block
> > > size of 1 GiB (e.g., gigantic pages in the hypervisor) and a Linux memory
> > > block size of 128MB.
> > >
> > > While we could compress the state into 2 bit, using 8 bit is much
> > > easier.
> > >
> > > This handling is similar, but different to acpi_scan_try_to_offline():
> > >
> > > a) We don't try to offline twice. I am not sure if this CONFIG_MEMCG
> > > optimization is still relevant - it should only apply to ZONE_NORMAL
> > > (where we have no guarantees). If relevant, we can always add it.
> > >
> > > b) acpi_scan_try_to_offline() simply onlines all memory in case
> > > something goes wrong. It doesn't restore previous online type. Let's do
> > > that, so we won't overwrite what e.g., user space configured.
> > >
> > > Reviewed-by: Wei Yang <richard.weiyang@xxxxxxxxxxxxxxxxx>
> > > Cc: "Michael S. Tsirkin" <mst@xxxxxxxxxx>
> > > Cc: Jason Wang <jasowang@xxxxxxxxxx>
> > > Cc: Pankaj Gupta <pankaj.gupta.linux@xxxxxxxxx>
> > > Cc: Michal Hocko <mhocko@xxxxxxxxxx>
> > > Cc: Oscar Salvador <osalvador@xxxxxxx>
> > > Cc: Wei Yang <richard.weiyang@xxxxxxxxxxxxxxxxx>
> > > Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> > > Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> > > Link: https://lore.kernel.org/r/20201112133815.13332-28-david@xxxxxxxxxx
> > > Signed-off-by: Michael S. Tsirkin <mst@xxxxxxxxxx>
> > > Acked-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> > > Signed-off-by: Ma Wupeng <mawupeng1@xxxxxxxxxx>
> > > ---
> > >   mm/memory_hotplug.c | 105 +++++++++++++++++++++++++++++++++++++-------
> > >   1 file changed, 89 insertions(+), 16 deletions(-)
> >
> > Why is this needed in 5.10.y?  Looks like a new feature to me, what
> > problem does it solve there?
> >
> > thanks,
> >
> > greg k-h
>
> It do introduce a new feature. But at the same time, it fix a memleak introduced
> in Commit 08b3acd7a68f ("mm/memory_hotplug: Introduce offline_and_remove_memory()"
>
> Our test find a memleak in init_memory_block, it is clear that mem is never
> been released due to wrong refcount. Commit 08b3acd7a68f ("mm/memory_hotplug:
> Introduce offline_and_remove_memory()") failed to dec refcount after
> find_memory_block which fail to dec refcount to zero in remove memory
> causing the leak.
>
> Commit 8dc4bb58a146 ("mm/memory_hotplug: extend offline_and_remove_memory()
> to handle more than one memory block") introduce walk_memory_blocks to
> replace find_memory_block which dec refcount by calling put_device after
> find_memory_block_by_id. In the way, the memleak is fixed.
>
> Here is the simplified calltrace:
>
>    kmem_cache_alloc_trace+0x664/0xed0
>    init_memory_block+0x8c/0x170
>    create_memory_block_devices+0xa4/0x150
>    add_memory_resource+0x188/0x530
>    __add_memory+0x78/0x104
>    add_memory+0x6c/0xb0

Makes sense to me. Of course, we could think about a simplified stable 
fix that only drops the ref.

--
Cheers,

David / dhildenb