On Sat, Aug 03, 2024 at 11:58:13AM -0700, Andrew Morton wrote:
> On Fri, 2 Aug 2024 10:49:22 +0100 Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> wrote:
>
> > > --- a/mm/mm_init.c
> > > +++ b/mm/mm_init.c
> > > @@ -1838,11 +1838,10 @@ void __init free_area_init(unsigned long *max_zone_pfn)
> > >
> > > if (!node_online(nid)) {
> > > /* Allocator not initialized yet */
> > > - pgdat = arch_alloc_nodedata(nid);
> > > + pgdat = memblock_alloc(sizeof(*pgdat), SMP_CACHE_BYTES);
> > > if (!pgdat)
> > > panic("Cannot allocate %zuB for node %d.\n",
> > > sizeof(*pgdat), nid);
> > > - arch_refresh_nodedata(nid, pgdat);
> >
> > This allocates pgdat but never sets node_data[nid] to it
> > and promptly leaks it on the line below.
> >
> > Just to sanity check this I spun up a qemu machine with no memory
> > initially present on some nodes and it went boom as you'd expect.
> >
> > I tested with addition of
> > NODE_DATA(nid) = pgdat;
> > and it all seems to work as expected.
>
> Thanks, I added that. It blew up on x86_64 allnoconfig because
> node_data[] (and hence NODE_DATA()) isn't an lvalue when CONFIG_NUMA=n.
>
> I'll put some #ifdef CONFIG_NUMAs in there for now but
>
> a) NODE_DATA() is upper-case. Implies "constant". Shouldn't be assigned to.
>
> b) NODE_DATA() should be non-lvalue when CONFIG_NUMA=y also. But no,
> we insist on implementing things in cpp instead of in C.
This looks like a candidate for a separate tree-wide cleanup.
> c) In fact assigning to anything which ends in "()" is nuts. Please
> clean up my tempfix.
>
> c) Mike, generally I'm wondering if there's a bunch of code here
> which isn't needed on CONFIG_NUMA=n. Please check all of this for
> unneeded bloatiness.
I believe the patch addresses your concerns, just with this the commit log
needs update. Instead of
Replace the call to arch_alloc_nodedata() in free_area_init() with
memblock_alloc(), remove arch_refresh_nodedata() and cleanup
include/linux/memory_hotplug.h from the associated ifdefery.
it should be
Replace the call to arch_alloc_nodedata() in free_area_init() with a
new helper alloc_offline_node_data(), remove arch_refresh_nodedata()
and cleanup include/linux/memory_hotplug.h from the associated
ifdefery.
I can send an updated patch if you prefer.
diff --git a/include/linux/numa.h b/include/linux/numa.h
index 3b12d8ca0afd..5a749fd67f39 100644
--- a/include/linux/numa.h
+++ b/include/linux/numa.h
@@ -34,6 +34,7 @@ extern struct pglist_data *node_data[];
#define NODE_DATA(nid) (node_data[nid])
void __init alloc_node_data(int nid);
+void __init alloc_offline_node_data(int nit);
/* Generic implementation available */
int numa_nearest_node(int node, unsigned int state);
@@ -62,6 +63,8 @@ static inline int phys_to_target_node(u64 start)
{
return 0;
}
+
+static inline void alloc_offline_node_data(int nit) {}
#endif
#define numa_map_to_online_node(node) numa_nearest_node(node, N_ONLINE)
diff --git a/mm/mm_init.c b/mm/mm_init.c
index bcc2f2dd8021..2785be04e7bb 100644
--- a/mm/mm_init.c
+++ b/mm/mm_init.c
@@ -1836,13 +1836,8 @@ void __init free_area_init(unsigned long *max_zone_pfn)
for_each_node(nid) {
pg_data_t *pgdat;
- if (!node_online(nid)) {
- /* Allocator not initialized yet */
- pgdat = memblock_alloc(sizeof(*pgdat), SMP_CACHE_BYTES);
- if (!pgdat)
- panic("Cannot allocate %zuB for node %d.\n",
- sizeof(*pgdat), nid);
- }
+ if (!node_online(nid))
+ alloc_offline_node_data(nid);
pgdat = NODE_DATA(nid);
free_area_init_node(nid);
diff --git a/mm/numa.c b/mm/numa.c
index da27eb151dc5..07e486a977c7 100644
--- a/mm/numa.c
+++ b/mm/numa.c
@@ -34,6 +34,18 @@ void __init alloc_node_data(int nid)
memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
}
+void __init alloc_offline_node_data(int nit)
+{
+ pg_data_t *pgdat;
+
+ pgdat = memblock_alloc(sizeof(*pgdat), SMP_CACHE_BYTES);
+ if (!pgdat)
+ panic("Cannot allocate %zuB for node %d.\n",
+ sizeof(*pgdat), nid);
+
+ node_data[nid] = pgdat;
+}
+
/* Stub functions: */
#ifndef memory_add_physaddr_to_nid
--
Sincerely yours,
Mike.
