[PATCH 0/2] mm/mempolicy: stop copying state in the interleave paths

[PATCH 0/2] mm/mempolicy: stop copying state in the interleave paths

[Gregory Price]

The interleave node selectors and bulk allocators take copies of
nodemasks and node weights (for weighted interleave) in the fault path.
Both of these copies can be entirely eliminated.

For node weights, use SRCU to pin the weights in place.  This eliminates
a copy and a kmalloc from the bulk allocator path.

For nodemasks, we can operate directly on pol->nodes as long as we bounds
check the walk.  A concurrent rebind can shrink the mask, or tear the read
of it so the mask appears empty.

 - The interleave node selectors fall back to numa_node_id() when that
   happens, which is what they already did when a copy came back empty.

 - The bulk allocator simply returns what it managed to allocate.

The node count and weight totals are read separately from the nodemask
walk that consumes them - creating a time-of-check / time-of-use race.
Just clamp the walk to a single pass (number of nodes), and clamp each
bulk allocation chunk to the space left in the request.

The cost is distribution accuracy during a rebind.  The copies never
corrected for that either - they only kept the code from dividing by
zero and overrunning the allocation request.

Gregory Price (2):
  mm/mempolicy: use SRCU for the weighted interleave state
  mm/mempolicy: stop copying the nodemask in the interleave paths

 mm/mempolicy.c | 156 ++++++++++++++++++++++++++-----------------------
 1 file changed, 83 insertions(+), 73 deletions(-)

-- 
2.55.0

[PATCH 1/2] mm/mempolicy: use SRCU for the weighted interleave state

[Gregory Price]

alloc_pages_bulk_weighted_interleave() copies iw_table into a scratch
array on every call so it can walk the weights outside of RCU.  The copy
exists only because the loop may sleep in the page allocator and so cannot
hold rcu_read_lock().

Use SRCU to pin the global iw_table object and use it in-place instead.

Retire through both flavors - call_srcu() for the sleeping readers, then
kfree_rcu() for the reference-less ones - so writers no longer block on
synchronize_rcu() either.

Tested in a VM with KASAN, PROVE_LOCKING and DEBUG_OBJECTS_RCU_HEAD,
with a udelay() injected into the read section to widen the race against
concurrent sysfs weight writers, and placement checked against the
configured weights.

Every retired state reached its callback. Swapping the deferred free for
a bare kfree() in the same test reports a use-after-free immediately.

Suggested-by: Andrew Morton <akpm@linux-foundation.org>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Assisted-by: Claude:claude-opus-5
Signed-off-by: Gregory Price (Meta) <gourry@gourry.net>
---
 mm/mempolicy.c | 70 ++++++++++++++++++++++++--------------------------
 1 file changed, 34 insertions(+), 36 deletions(-)

diff --git a/mm/mempolicy.c b/mm/mempolicy.c
index 060a0eb26917..2643915dc966 100644
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -112,6 +112,7 @@
 #include <linux/printk.h>
 #include <linux/leafops.h>
 #include <linux/gcd.h>
+#include <linux/srcu.h>
 
 #include <asm/tlbflush.h>
 #include <asm/tlb.h>
@@ -157,6 +158,7 @@ static const int weightiness = 32;
  */
 struct weighted_interleave_state {
 	bool mode_auto;
+	struct rcu_head rcu;
 	u8 iw_table[];
 };
 static struct weighted_interleave_state __rcu *wi_state;
@@ -168,6 +170,24 @@ static unsigned int *node_bw_table;
  */
 static DEFINE_MUTEX(wi_state_lock);
 
+/* Readers that sleep while walking iw_table hold this instead */
+DEFINE_STATIC_SRCU_FAST(wi_srcu);
+
+static void wi_state_free_rcu(struct rcu_head *head)
+{
+	struct weighted_interleave_state *state =
+		container_of(head, struct weighted_interleave_state, rcu);
+
+	kfree_rcu(state, rcu);
+}
+
+/* Retire through both flavors: sleeping readers use SRCU, the rest RCU */
+static void wi_state_retire(struct weighted_interleave_state *state)
+{
+	if (state)
+		call_srcu(&wi_srcu, &state->rcu, wi_state_free_rcu);
+}
+
 static u8 get_il_weight(int node)
 {
 	struct weighted_interleave_state *state;
@@ -266,10 +286,7 @@ int mempolicy_set_node_perf(unsigned int node, struct access_coordinate *coords)
 	rcu_assign_pointer(wi_state, new_wi_state);
 
 	mutex_unlock(&wi_state_lock);
-	if (old_wi_state) {
-		synchronize_rcu();
-		kfree(old_wi_state);
-	}
+	wi_state_retire(old_wi_state);
 out:
 	kfree(old_bw);
 	return 0;
@@ -2644,7 +2661,8 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	unsigned long nr_allocated = 0;
 	unsigned long rounds;
 	unsigned long node_pages, delta;
-	u8 *weights, weight;
+	struct srcu_ctr __percpu *scp;
+	u8 *table, weight;
 	unsigned int weight_total = 0;
 	unsigned long rem_pages = nr_pages;
 	nodemask_t nodes;
@@ -2688,25 +2706,14 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	me->il_weight = 0;
 	prev_node = node;
 
-	/* create a local copy of node weights to operate on outside rcu */
-	weights = kmalloc(nr_node_ids, gfp & GFP_RECLAIM_MASK);
-	if (!weights)
-		return total_allocated;
-
-	rcu_read_lock();
-	state = rcu_dereference(wi_state);
-	if (state) {
-		memcpy(weights, state->iw_table, nr_node_ids * sizeof(u8));
-		rcu_read_unlock();
-	} else {
-		rcu_read_unlock();
-		for (i = 0; i < nr_node_ids; i++)
-			weights[i] = 1;
-	}
+	/* The page allocator may sleep, pin the weight table with SRCU */
+	scp = srcu_read_lock_fast(&wi_srcu);
+	state = srcu_dereference(wi_state, &wi_srcu);
+	table = state ? state->iw_table : NULL;
 
 	/* calculate total, detect system default usage */
 	for_each_node_mask(node, nodes)
-		weight_total += weights[node];
+		weight_total += table ? table[node] : 1;
 
 	/*
 	 * Calculate rounds/partial rounds to minimize __alloc_pages_bulk calls.
@@ -2718,10 +2725,10 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	rounds = rem_pages / weight_total;
 	delta = rem_pages % weight_total;
 	resume_node = next_node_in(prev_node, nodes);
-	resume_weight = weights[resume_node];
+	resume_weight = table ? table[resume_node] : 1;
 	for (i = 0; i < nnodes; i++) {
 		node = next_node_in(prev_node, nodes);
-		weight = weights[node];
+		weight = table ? table[node] : 1;
 		node_pages = weight * rounds;
 		/* If a delta exists, add this node's portion of the delta */
 		if (delta > weight) {
@@ -2747,7 +2754,7 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	}
 	me->il_prev = resume_node;
 	me->il_weight = resume_weight;
-	kfree(weights);
+	srcu_read_unlock_fast(&wi_srcu, scp);
 	return total_allocated;
 }
 
@@ -3673,10 +3680,7 @@ static ssize_t node_store(struct kobject *kobj, struct kobj_attribute *attr,
 
 	rcu_assign_pointer(wi_state, new_wi_state);
 	mutex_unlock(&wi_state_lock);
-	if (old_wi_state) {
-		synchronize_rcu();
-		kfree(old_wi_state);
-	}
+	wi_state_retire(old_wi_state);
 	return count;
 }
 
@@ -3742,10 +3746,7 @@ static ssize_t weighted_interleave_auto_store(struct kobject *kobj,
 update_wi_state:
 	rcu_assign_pointer(wi_state, new_wi_state);
 	mutex_unlock(&wi_state_lock);
-	if (old_wi_state) {
-		synchronize_rcu();
-		kfree(old_wi_state);
-	}
+	wi_state_retire(old_wi_state);
 	return count;
 }
 
@@ -3789,10 +3790,7 @@ static void wi_state_free(void)
 	rcu_assign_pointer(wi_state, NULL);
 	mutex_unlock(&wi_state_lock);
 
-	if (old_wi_state) {
-		synchronize_rcu();
-		kfree(old_wi_state);
-	}
+	wi_state_retire(old_wi_state);
 }
 
 static struct kobj_attribute wi_auto_attr = {
-- 
2.55.0

[PATCH 2/2] mm/mempolicy: stop copying the nodemask in the interleave paths

[Gregory Price]

The interleave node selectors copy pol->nodes onto the stack so the mask
cannot change while they walk it.  nodemask_t is 128 bytes at
MAX_NUMNODES=1024, and two of the three run per folio fault.

The copy only buys consistency between the node count and the walk.
Drop the consistency and just bounds check the walk instead.

If an empty nodelist or weight is perceived, fall back to numa_node_id(),
which is what the functions already did when the copy came back empty.

weighted_interleave_nid() counts the nodes as we sum the weights. We use
that node count to limit the maximum skew a single node can host.

interleave_nid() walks with next_node_in() rather than next_node(), so a
mask that shrank mid-walk wraps to a node still in the policy.

alloc_pages_bulk_weighted_interleave() derives per-node counts from a
weight total summed over the mask, so a changing mask can make them exceed
the request.  Clamp each chunk to the space left in page_array.

A cpuset cookie will not work here: two of these take VMA policies, which
mpol_rebind_mm() rebinds under mmap_write_lock(), not mems_allowed_seq.

Cost is distribution accuracy during a rebind - but the copy never
corrected this anyway, it was just a safety mechanism to prevent div/0
and overrunning the alloc request buffer.

Remove read_once_policy_nodemask(), now unused.

-fstack-usage at MAX_NUMNODES=1024:

  weighted_interleave_nid            184 -> 56
  interleave_nid                     168 -> 32
  alloc_pages_bulk_mempolicy_noprof  360 -> 136

Assisted-by: Claude:claude-opus-5
Signed-off-by: Gregory Price (Meta) <gourry@gourry.net>
---
 mm/mempolicy.c | 86 ++++++++++++++++++++++++++++----------------------
 1 file changed, 49 insertions(+), 37 deletions(-)

diff --git a/mm/mempolicy.c b/mm/mempolicy.c
index 2643915dc966..296129126109 100644
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -2197,34 +2197,15 @@ unsigned int mempolicy_slab_node(void)
 	}
 }
 
-static unsigned int read_once_policy_nodemask(struct mempolicy *pol,
-					      nodemask_t *mask)
-{
-	/*
-	 * barrier stabilizes the nodemask locally so that it can be iterated
-	 * over safely without concern for changes. Allocators validate node
-	 * selection does not violate mems_allowed, so this is safe.
-	 */
-	barrier();
-	memcpy(mask, &pol->nodes, sizeof(nodemask_t));
-	barrier();
-	return nodes_weight(*mask);
-}
-
 static unsigned int weighted_interleave_nid(struct mempolicy *pol, pgoff_t ilx)
 {
 	struct weighted_interleave_state *state;
-	nodemask_t nodemask;
-	unsigned int target, nr_nodes;
+	unsigned int target, nnodes = 0;
 	u8 *table = NULL;
 	unsigned int weight_total = 0;
 	u8 weight;
 	int nid = 0;
 
-	nr_nodes = read_once_policy_nodemask(pol, &nodemask);
-	if (!nr_nodes)
-		return numa_node_id();
-
 	rcu_read_lock();
 
 	state = rcu_dereference(wi_state);
@@ -2232,22 +2213,40 @@ static unsigned int weighted_interleave_nid(struct mempolicy *pol, pgoff_t ilx)
 	if (state)
 		table = state->iw_table;
 
-	/* calculate the total weight */
-	for_each_node_mask(nid, nodemask)
+	/* calculate the total weight and the node count */
+	for_each_node_mask(nid, pol->nodes) {
 		weight_total += table ? table[nid] : 1;
+		nnodes++;
+	}
+
+	/* the mask is empty */
+	if (!weight_total) {
+		rcu_read_unlock();
+		return numa_node_id();
+	}
 
 	/* Calculate the node offset based on totals */
 	target = ilx % weight_total;
-	nid = first_node(nodemask);
-	while (target) {
+	nid = first_node(pol->nodes);
+
+	/*
+	 * The target was calculated in a separate loop, and a concurrent
+	 * rebind can change the total number of nodes.  Clamp this loop to
+	 * a single pass (nnodes) to keep the walk bounded by node count.
+	 */
+	while (target && nnodes-- && nid < MAX_NUMNODES) {
 		/* detect system default usage */
 		weight = table ? table[nid] : 1;
 		if (target < weight)
 			break;
 		target -= weight;
-		nid = next_node_in(nid, nodemask);
+		nid = next_node_in(nid, pol->nodes);
 	}
 	rcu_read_unlock();
+
+	/* the mask emptied under the walk */
+	if (nid >= MAX_NUMNODES)
+		return numa_node_id();
 	return nid;
 }
 
@@ -2258,18 +2257,21 @@ static unsigned int weighted_interleave_nid(struct mempolicy *pol, pgoff_t ilx)
  */
 static unsigned int interleave_nid(struct mempolicy *pol, pgoff_t ilx)
 {
-	nodemask_t nodemask;
 	unsigned int target, nnodes;
 	int i;
 	int nid;
 
-	nnodes = read_once_policy_nodemask(pol, &nodemask);
+	nnodes = nodes_weight(pol->nodes);
 	if (!nnodes)
 		return numa_node_id();
 	target = ilx % nnodes;
-	nid = first_node(nodemask);
-	for (i = 0; i < target; i++)
-		nid = next_node(nid, nodemask);
+	nid = first_node(pol->nodes);
+	for (i = 0; i < target && nid < MAX_NUMNODES; i++)
+		nid = next_node_in(nid, pol->nodes);
+
+	/* the mask emptied under the walk */
+	if (nid >= MAX_NUMNODES)
+		return numa_node_id();
 	return nid;
 }
 
@@ -2665,7 +2667,6 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	u8 *table, weight;
 	unsigned int weight_total = 0;
 	unsigned long rem_pages = nr_pages;
-	nodemask_t nodes;
 	int nnodes, node;
 	int resume_node = MAX_NUMNODES - 1;
 	u8 resume_weight = 0;
@@ -2675,10 +2676,10 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	if (!nr_pages)
 		return 0;
 
-	/* read the nodes onto the stack, retry if done during rebind */
+	/* count the nodes, retry if a rebind happened during the read */
 	do {
 		cpuset_mems_cookie = read_mems_allowed_begin();
-		nnodes = read_once_policy_nodemask(pol, &nodes);
+		nnodes = nodes_weight(pol->nodes);
 	} while (read_mems_allowed_retry(cpuset_mems_cookie));
 
 	/* if the nodemask has become invalid, we cannot do anything */
@@ -2688,7 +2689,7 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	/* Continue allocating from most recent node and adjust the nr_pages */
 	node = me->il_prev;
 	weight = me->il_weight;
-	if (weight && node_isset(node, nodes)) {
+	if (weight && node_isset(node, pol->nodes)) {
 		node_pages = min(rem_pages, weight);
 		nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
 						  page_array);
@@ -2712,9 +2713,13 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	table = state ? state->iw_table : NULL;
 
 	/* calculate total, detect system default usage */
-	for_each_node_mask(node, nodes)
+	for_each_node_mask(node, pol->nodes)
 		weight_total += table ? table[node] : 1;
 
+	/* the mask emptied since it was counted */
+	if (!weight_total)
+		goto out;
+
 	/*
 	 * Calculate rounds/partial rounds to minimize __alloc_pages_bulk calls.
 	 * Track which node weighted interleave should resume from.
@@ -2724,10 +2729,14 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	 */
 	rounds = rem_pages / weight_total;
 	delta = rem_pages % weight_total;
-	resume_node = next_node_in(prev_node, nodes);
+	resume_node = next_node_in(prev_node, pol->nodes);
+	if (resume_node >= MAX_NUMNODES)
+		goto out;
 	resume_weight = table ? table[resume_node] : 1;
 	for (i = 0; i < nnodes; i++) {
-		node = next_node_in(prev_node, nodes);
+		node = next_node_in(prev_node, pol->nodes);
+		if (node >= MAX_NUMNODES)
+			break;
 		weight = table ? table[node] : 1;
 		node_pages = weight * rounds;
 		/* If a delta exists, add this node's portion of the delta */
@@ -2744,6 +2753,8 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 		/* node_pages can be 0 if an allocation fails and rounds == 0 */
 		if (!node_pages)
 			break;
+		/* a rebind can invalidate the counts: never overrun page_array */
+		node_pages = min(node_pages, nr_pages - total_allocated);
 		nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
 						  page_array);
 		page_array += nr_allocated;
@@ -2754,6 +2765,7 @@ static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
 	}
 	me->il_prev = resume_node;
 	me->il_weight = resume_weight;
+out:
 	srcu_read_unlock_fast(&wi_srcu, scp);
 	return total_allocated;
 }
-- 
2.55.0

Re: [PATCH 0/2] mm/mempolicy: stop copying state in the interleave paths

[Andrew Morton]

On Fri, 28 Aug 2026 21:59:41 -0400 Gregory Price <gourry@gourry.net> wrote:

> The interleave node selectors and bulk allocators take copies of
> nodemasks and node weights (for weighted interleave) in the fault path.
> Both of these copies can be entirely eliminated.
> 
> For node weights, use SRCU to pin the weights in place.  This eliminates
> a copy and a kmalloc from the bulk allocator path.
> 
> For nodemasks, we can operate directly on pol->nodes as long as we bounds
> check the walk.  A concurrent rebind can shrink the mask, or tear the read
> of it so the mask appears empty.
> 
>  - The interleave node selectors fall back to numa_node_id() when that
>    happens, which is what they already did when a copy came back empty.
> 
>  - The bulk allocator simply returns what it managed to allocate.
> 
> The node count and weight totals are read separately from the nodemask
> walk that consumes them - creating a time-of-check / time-of-use race.
> Just clamp the walk to a single pass (number of nodes), and clamp each
> bulk allocation chunk to the space left in the request.
> 
> The cost is distribution accuracy during a rebind.  The copies never
> corrected for that either - they only kept the code from dividing by
> zero and overrunning the allocation request.

Not very well, it seems.  Sashiko thinks there's a div-by-zero in
alloc_pages_bulk_interleave().

	https://sashiko.dev/#/patchset/20260829015943.1258774-1-gourry@gourry.net

> Gregory Price (2):
>   mm/mempolicy: use SRCU for the weighted interleave state
>   mm/mempolicy: stop copying the nodemask in the interleave paths

Looks nice, thanks - I'll queue it for testing.