On Wed, 18 Dec 2019 18:08:04 +0100 David Hildenbrand [off-list ref] wrote:

On 01.12.19 00:21, Andrew Morton wrote:

quoted

On Sun, 27 Oct 2019 23:45:52 +0100 David Hildenbrand [off-list ref] wrote:

quoted

I think I just found an issue with try_offline_node(). 
try_offline_node() is pretty much broken already (touches garbage 
memmaps and will not considers mixed NIDs within sections), however, 
relies on the node span to look for memory sections to probe. So it 
seems to rely on the nodes getting shrunk when removing memory, not when 
offlining.

As we shrink the node span when offlining now and not when removing, 
this can go wrong once we offline the last memory block of the node and 
offline the last CPU. We could still have memory around that we could 
re-online, however, the node would already be offline. Unlikely, but 
possible.

Note that the same is also broken without this patch in case memory is 
never onlined. The "pfn_to_nid(pfn) != nid" can easily succeed on the 
garbage memmap, resulting in  no memory being detected as belonging to 
the node. Also, resize_pgdat_range() is called when onlining memory, not 
when adding it. :/ Oh this is so broken :)

The right fix is probably to walk over all memory blocks that could 
exist and test if they belong to the nid (if offline, check the 
block->nid, if online check all pageblocks). A fix we can then move in 
front of this patch.

Will look into this this week.

And this series shows almost no sign of having been reviewed.  I'll hold
it over for 5.6.

Hi Andrew, any chance we can get the (now at least reviewed - thx Oscar)
fix in patch #5 into 5.5? (I want to do the final stable backports for
the uninitialized memmap stuff)

Sure, I queued it for the next batch of 5.5 fixes.