[PATCH] mm/mempolicy: use one node set for interleave weights

[Gregory Price (Meta)]

From: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>

From: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>

Switching weighted interleave back to automatic mode while memory is
being onlined can cause a kernel divide error on x86, and a panic if
panic_on_oops is set.

reduce_interleave_weights() scans N_MEMORY three times, but hotplug
can cause N_MEMORY to change at runtime - and torn reads on that
nodemask can cause a total sum of zero bandwidth.  A node that
was hot-unplugged and then hotplugged back can then cause this
loop to generate a div/0.

Record the nodes visited while summing their bandwidth, then use
that same set when scaling and reducing the weights.  This gives
us a consistent state even if N_MEMORY changes.

Fixes: e341f9c3c841 ("mm/mempolicy: Weighted Interleave Auto-tuning")
Cc: <stable@xxxxxxxxxxxxxxx>
Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Signed-off-by: Gregory Price (Meta) <gourry@xxxxxxxxxx>
---
 mm/mempolicy.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

 [gregory: recieved off-list, separately validated and tested)

diff --git a/mm/mempolicy.c b/mm/mempolicy.c
index 6f92621e9135..26a1a21be65d 100644
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -207,15 +207,22 @@ static u8 get_il_weight(int node)
  */
 static void reduce_interleave_weights(unsigned int *bw, u8 *new_iw)
 {
+	nodemask_t nodes = NODE_MASK_NONE;
 	u64 sum_bw = 0;
 	unsigned int cast_sum_bw, scaling_factor = 1, iw_gcd = 0;
 	int nid;
 
-	for_each_node_state(nid, N_MEMORY)
+	/*
+	 * N_MEMORY can change while wi_state_lock is held. Remember the
+	 * nodes contributing to sum_bw so all passes use the same set.
+	 */
+	for_each_node_state(nid, N_MEMORY) {
+		node_set(nid, nodes);
 		sum_bw += bw[nid];
+	}
 
 	/* Scale bandwidths to whole numbers in the range [1, weightiness] */
-	for_each_node_state(nid, N_MEMORY) {
+	for_each_node_mask(nid, nodes) {
 		/*
 		 * Try not to perform 64-bit division.
 		 * If sum_bw < scaling_factor, then sum_bw < U32_MAX.
@@ -234,7 +241,7 @@ static void reduce_interleave_weights(unsigned int *bw, u8 *new_iw)
 	}
 
 	/* 1:2 is strictly better than 16:32. Reduce by the weights' GCD. */
-	for_each_node_state(nid, N_MEMORY)
+	for_each_node_mask(nid, nodes)
 		new_iw[nid] /= iw_gcd;
 }
 
-- 
2.56.0