On Mon 30-03-20 12:21:27, Mike Rapoport wrote:
> On Mon, Mar 30, 2020 at 09:42:46AM +0200, Michal Hocko wrote:
> > On Sat 28-03-20 11:31:17, Hoan Tran wrote:
> > > In NUMA layout which nodes have memory ranges that span across other nodes,
> > > the mm driver can detect the memory node id incorrectly.
> > >
> > > For example, with layout below
> > > Node 0 address: 0000 xxxx 0000 xxxx
> > > Node 1 address: xxxx 1111 xxxx 1111
> > >
> > > Note:
> > > - Memory from low to high
> > > - 0/1: Node id
> > > - x: Invalid memory of a node
> > >
> > > When mm probes the memory map, without CONFIG_NODES_SPAN_OTHER_NODES
> > > config, mm only checks the memory validity but not the node id.
> > > Because of that, Node 1 also detects the memory from node 0 as below
> > > when it scans from the start address to the end address of node 1.
> > >
> > > Node 0 address: 0000 xxxx xxxx xxxx
> > > Node 1 address: xxxx 1111 1111 1111
> > >
> > > This layout could occur on any architecture. Most of them enables
> > > this config by default with CONFIG_NUMA. This patch, by default, enables
> > > CONFIG_NODES_SPAN_OTHER_NODES or uses early_pfn_in_nid() for NUMA.
> >
> > I am not opposed to this at all. It reduces the config space and that is
> > a good thing on its own. The history has shown that meory layout might
> > be really wild wrt NUMA. The config is only used for early_pfn_in_nid
> > which is clearly an overkill.
> >
> > Your description doesn't really explain why this is safe though. The
> > history of this config is somehow messy, though. Mike has tried
> > to remove it a94b3ab7eab4 ("[PATCH] mm: remove arch independent
> > NODES_SPAN_OTHER_NODES") just to be reintroduced by 7516795739bd
> > ("[PATCH] Reintroduce NODES_SPAN_OTHER_NODES for powerpc") without any
> > reasoning what so ever. This doesn't make it really easy see whether
> > reasons for reintroduction are still there. Maybe there are some subtle
> > dependencies. I do not see any TBH but that might be burried deep in an
> > arch specific code.
>
> Well, back then early_pfn_in_nid() was arch-dependant, today everyone
> except ia64 rely on HAVE_MEMBLOCK_NODE_MAP.
What would it take to make ia64 use HAVE_MEMBLOCK_NODE_MAP? I would
really love to see that thing go away. It is causing problems when
people try to use memblock api.
> So, if the memblock node map
> is correct, that using CONFIG_NUMA instead of CONFIG_NODES_SPAN_OTHER_NODES
> would only mean that early_pfn_in_nid() will cost several cycles more on
> architectures that didn't select CONFIG_NODES_SPAN_OTHER_NODES (i.e. arm64
> and sh).
Do we have any idea on how much of an overhead that is? Because this is
per each pfn so it can accumulate a lot!
> Agian, ia64 is an exception here.
Thanks for the clarification!
--
Michal Hocko
SUSE Labs
