Re: [PATCH] mm/alloc: fallback to first node if the wanted node offline

[Michal Hocko]

On Wed 12-12-18 16:31:35, Pingfan Liu wrote:
> On Mon, Dec 10, 2018 at 8:37 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> >
> [...]
> >
> > In other words. Does the following work? I am sorry to wildguess this
> > way but I am not able to recreate your setups to play with this myself.
> >
> > diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c
> > index 1308f5408bf7..d51643e10d00 100644
> > --- a/arch/x86/mm/numa.c
> > +++ b/arch/x86/mm/numa.c
> > @@ -216,8 +216,6 @@ static void __init alloc_node_data(int nid)
> >
> >         node_data[nid] = nd;
> >         memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
> > -
> > -       node_set_online(nid);
> >  }
> >
> >  /**
> > @@ -527,6 +525,19 @@ static void __init numa_clear_kernel_node_hotplug(void)
> >         }
> >  }
> >
> > +static void __init init_memory_less_node(int nid)
> > +{
> > +       unsigned long zones_size[MAX_NR_ZONES] = {0};
> > +       unsigned long zholes_size[MAX_NR_ZONES] = {0};
> > +
> > +       free_area_init_node(nid, zones_size, 0, zholes_size);
> > +
> > +       /*
> > +        * All zonelists will be built later in start_kernel() after per cpu
> > +        * areas are initialized.
> > +        */
> > +}
> > +
> >  static int __init numa_register_memblks(struct numa_meminfo *mi)
> >  {
> >         unsigned long uninitialized_var(pfn_align);
> > @@ -570,7 +581,7 @@ static int __init numa_register_memblks(struct numa_meminfo *mi)
> >                 return -EINVAL;
> >
> >         /* Finally register nodes. */
> > -       for_each_node_mask(nid, node_possible_map) {
> > +       for_each_node(nid) {
> >                 u64 start = PFN_PHYS(max_pfn);
> >                 u64 end = 0;
> >
> > @@ -592,6 +603,10 @@ static int __init numa_register_memblks(struct numa_meminfo *mi)
> >                         continue;
> >
> >                 alloc_node_data(nid);
> > +               if (!end)
> 
> Here comes the bug, since !end can not reach here.

You are right. I am dumb. I've just completely missed that. Sigh.
Anyway, I think the code is more complicated than necessary and we can
simply drop the check. I do not think we really have to worry about
the start overflowing end. So the end patch should look as follows.
Btw. I believe it is better to pull alloc_node_data out of init_memory_less_node
because a) there is no need to duplicate the call and moreover we want
to pull node_set_online as well. The code also seems cleaner this way.

Thanks for your testing and your patience with me here.

diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c
index 1308f5408bf7..a5548fe668fb 100644
--- a/arch/x86/mm/numa.c
+++ b/arch/x86/mm/numa.c
@@ -216,8 +216,6 @@ static void __init alloc_node_data(int nid)
 
 	node_data[nid] = nd;
 	memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
-
-	node_set_online(nid);
 }
 
 /**
@@ -527,6 +525,19 @@ static void __init numa_clear_kernel_node_hotplug(void)
 	}
 }
 
+static void __init init_memory_less_node(int nid)
+{
+	unsigned long zones_size[MAX_NR_ZONES] = {0};
+	unsigned long zholes_size[MAX_NR_ZONES] = {0};
+
+	free_area_init_node(nid, zones_size, 0, zholes_size);
+
+	/*
+	 * All zonelists will be built later in start_kernel() after per cpu
+	 * areas are initialized.
+	 */
+}
+
 static int __init numa_register_memblks(struct numa_meminfo *mi)
 {
 	unsigned long uninitialized_var(pfn_align);
@@ -570,7 +581,7 @@ static int __init numa_register_memblks(struct numa_meminfo *mi)
 		return -EINVAL;
 
 	/* Finally register nodes. */
-	for_each_node_mask(nid, node_possible_map) {
+	for_each_node(nid) {
 		u64 start = PFN_PHYS(max_pfn);
 		u64 end = 0;
 
@@ -581,9 +592,6 @@ static int __init numa_register_memblks(struct numa_meminfo *mi)
 			end = max(mi->blk[i].end, end);
 		}
 
-		if (start >= end)
-			continue;
-
 		/*
 		 * Don't confuse VM with a node that doesn't have the
 		 * minimum amount of memory:
@@ -592,6 +600,10 @@ static int __init numa_register_memblks(struct numa_meminfo *mi)
 			continue;
 
 		alloc_node_data(nid);
+		if (!end)
+			init_memory_less_node(nid);
+		else
+			node_set_online(nid);
 	}
 
 	/* Dump memblock with node info and return. */
@@ -721,21 +733,6 @@ void __init x86_numa_init(void)
 	numa_init(dummy_numa_init);
 }
 
-static void __init init_memory_less_node(int nid)
-{
-	unsigned long zones_size[MAX_NR_ZONES] = {0};
-	unsigned long zholes_size[MAX_NR_ZONES] = {0};
-
-	/* Allocate and initialize node data. Memory-less node is now online.*/
-	alloc_node_data(nid);
-	free_area_init_node(nid, zones_size, 0, zholes_size);
-
-	/*
-	 * All zonelists will be built later in start_kernel() after per cpu
-	 * areas are initialized.
-	 */
-}
-
 /*
  * Setup early cpu_to_node.
  *
@@ -763,9 +760,6 @@ void __init init_cpu_to_node(void)
 		if (node == NUMA_NO_NODE)
 			continue;
 
-		if (!node_online(node))
-			init_memory_less_node(node);
-
 		numa_set_node(cpu, node);
 	}
 }
-- 
Michal Hocko
SUSE Labs