From: Peter Zijlstra <peterz@infradead.org>

Extend the futex2 interface to be numa aware.

When FUTEX2_NUMA is specified for a futex, the user value is extended
to two words (of the same size). The first is the user value we all
know, the second one will be the node to place this futex on.

  struct futex_numa_32 {
	u32 val;
	u32 node;
  };

When node is set to ~0, WAIT will set it to the current node_id such
that WAKE knows where to find it. If userspace corrupts the node value
between WAIT and WAKE, the futex will not be found and no wakeup will
happen.

When FUTEX2_NUMA is not set, the node is simply an extension of the
hash, such that traditional futexes are still interleaved over the
nodes.

This is done to avoid having to have a separate !numa hash-table.

[bigeasy: ensure to have at least hashsize of 4 in futex_init(), add
pr_info() for size and allocation information.]

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 include/linux/futex.h      |   3 ++
 include/uapi/linux/futex.h |   8 +++
 kernel/futex/core.c        | 100 ++++++++++++++++++++++++++++++-------
 kernel/futex/futex.h       |  33 ++++++++++--
 4 files changed, 124 insertions(+), 20 deletions(-)

diff --git a/include/linux/futex.h b/include/linux/futex.h
index 0cb4c6d512dfb..ee48dcfbfe59d 100644
--- a/include/linux/futex.h
+++ b/include/linux/futex.h
@@ -34,6 +34,7 @@ union futex_key {
 		u64 i_seq;
 		unsigned long pgoff;
 		unsigned int offset;
+		/* unsigned int node; */
 	} shared;
 	struct {
 		union {
@@ -42,11 +43,13 @@ union futex_key {
 		};
 		unsigned long address;
 		unsigned int offset;
+		/* unsigned int node; */
 	} private;
 	struct {
 		u64 ptr;
 		unsigned long word;
 		unsigned int offset;
+		unsigned int node;	/* NOT hashed! */
 	} both;
 };
 
diff --git a/include/uapi/linux/futex.h b/include/uapi/linux/futex.h
index d2ee625ea1890..0435025beaae8 100644
--- a/include/uapi/linux/futex.h
+++ b/include/uapi/linux/futex.h
@@ -74,6 +74,14 @@
 /* do not use */
 #define FUTEX_32		FUTEX2_SIZE_U32 /* historical accident :-( */
 
+
+/*
+ * When FUTEX2_NUMA doubles the futex word, the second word is a node value.
+ * The special value -1 indicates no-node. This is the same value as
+ * NUMA_NO_NODE, except that value is not ABI, this is.
+ */
+#define FUTEX_NO_NODE		(-1)
+
 /*
  * Max numbers of elements in a futex_waitv array
  */
diff --git a/kernel/futex/core.c b/kernel/futex/core.c
index 220ddc7620e49..5ab2713ac906f 100644
--- a/kernel/futex/core.c
+++ b/kernel/futex/core.c
@@ -36,6 +36,8 @@
 #include <linux/pagemap.h>
 #include <linux/debugfs.h>
 #include <linux/plist.h>
+#include <linux/gfp.h>
+#include <linux/vmalloc.h>
 #include <linux/memblock.h>
 #include <linux/fault-inject.h>
 #include <linux/slab.h>
@@ -51,11 +53,14 @@
  * reside in the same cacheline.
  */
 static struct {
-	struct futex_hash_bucket *queues;
 	unsigned long            hashmask;
+	unsigned int		 hashshift;
+	struct futex_hash_bucket *queues[MAX_NUMNODES];
 } __futex_data __read_mostly __aligned(2*sizeof(long));
-#define futex_queues   (__futex_data.queues)
-#define futex_hashmask (__futex_data.hashmask)
+
+#define futex_hashmask	(__futex_data.hashmask)
+#define futex_hashshift	(__futex_data.hashshift)
+#define futex_queues	(__futex_data.queues)
 
 struct futex_private_hash {
 	rcuref_t	users;
@@ -332,15 +337,35 @@ __futex_hash(union futex_key *key, struct futex_private_hash *fph)
 {
 	struct futex_hash_bucket *hb;
 	u32 hash;
+	int node;
 
 	hb = __futex_hash_private(key, fph);
 	if (hb)
 		return hb;
 
 	hash = jhash2((u32 *)key,
-		      offsetof(typeof(*key), both.offset) / 4,
+		      offsetof(typeof(*key), both.offset) / sizeof(u32),
 		      key->both.offset);
-	return &futex_queues[hash & futex_hashmask];
+	node = key->both.node;
+
+	if (node == FUTEX_NO_NODE) {
+		/*
+		 * In case of !FLAGS_NUMA, use some unused hash bits to pick a
+		 * node -- this ensures regular futexes are interleaved across
+		 * the nodes and avoids having to allocate multiple
+		 * hash-tables.
+		 *
+		 * NOTE: this isn't perfectly uniform, but it is fast and
+		 * handles sparse node masks.
+		 */
+		node = (hash >> futex_hashshift) % nr_node_ids;
+		if (!node_possible(node)) {
+			node = find_next_bit_wrap(node_possible_map.bits,
+						  nr_node_ids, node);
+		}
+	}
+
+	return &futex_queues[node][hash & futex_hashmask];
 }
 
 /**
@@ -447,25 +472,49 @@ int get_futex_key(u32 __user *uaddr, unsigned int flags, union futex_key *key,
 	struct page *page;
 	struct folio *folio;
 	struct address_space *mapping;
-	int err, ro = 0;
+	int node, err, size, ro = 0;
 	bool fshared;
 
 	fshared = flags & FLAGS_SHARED;
+	size = futex_size(flags);
+	if (flags & FLAGS_NUMA)
+		size *= 2;
 
 	/*
 	 * The futex address must be "naturally" aligned.
 	 */
 	key->both.offset = address % PAGE_SIZE;
-	if (unlikely((address % sizeof(u32)) != 0))
+	if (unlikely((address % size) != 0))
 		return -EINVAL;
 	address -= key->both.offset;
 
-	if (unlikely(!access_ok(uaddr, sizeof(u32))))
+	if (unlikely(!access_ok(uaddr, size)))
 		return -EFAULT;
 
 	if (unlikely(should_fail_futex(fshared)))
 		return -EFAULT;
 
+	if (flags & FLAGS_NUMA) {
+		u32 __user *naddr = uaddr + size / 2;
+
+		if (futex_get_value(&node, naddr))
+			return -EFAULT;
+
+		if (node == FUTEX_NO_NODE) {
+			node = numa_node_id();
+			if (futex_put_value(node, naddr))
+				return -EFAULT;
+
+		} else if (node >= MAX_NUMNODES || !node_possible(node)) {
+			return -EINVAL;
+		}
+
+		key->both.node = node;
+
+	} else {
+		key->both.node = FUTEX_NO_NODE;
+	}
+
 	/*
 	 * PROCESS_PRIVATE futexes are fast.
 	 * As the mm cannot disappear under us and the 'key' only needs
@@ -1603,24 +1652,41 @@ int futex_hash_prctl(unsigned long arg2, unsigned long arg3)
 static int __init futex_init(void)
 {
 	unsigned long hashsize, i;
-	unsigned int futex_shift;
+	unsigned int order, n;
+	unsigned long size;
 
 #ifdef CONFIG_BASE_SMALL
 	hashsize = 16;
 #else
-	hashsize = roundup_pow_of_two(256 * num_possible_cpus());
+	hashsize = 256 * num_possible_cpus();
+	hashsize /= num_possible_nodes();
+	hashsize = max(4, hashsize);
+	hashsize = roundup_pow_of_two(hashsize);
 #endif
+	futex_hashshift = ilog2(hashsize);
+	size = sizeof(struct futex_hash_bucket) * hashsize;
+	order = get_order(size);
 
-	futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
-					       hashsize, 0, 0,
-					       &futex_shift, NULL,
-					       hashsize, hashsize);
-	hashsize = 1UL << futex_shift;
+	for_each_node(n) {
+		struct futex_hash_bucket *table;
 
-	for (i = 0; i < hashsize; i++)
-		futex_hash_bucket_init(&futex_queues[i], NULL);
+		if (order > MAX_PAGE_ORDER)
+			table = vmalloc_huge_node(size, GFP_KERNEL, n);
+		else
+			table = alloc_pages_exact_nid(n, size, GFP_KERNEL);
+
+		BUG_ON(!table);
+
+		for (i = 0; i < hashsize; i++)
+			futex_hash_bucket_init(&table[i], NULL);
+
+		futex_queues[n] = table;
+	}
 
 	futex_hashmask = hashsize - 1;
+	pr_info("futex hash table entries: %lu (%lu bytes on %d NUMA nodes, total %lu KiB, %s).\n",
+		hashsize, size, num_possible_nodes(), size * num_possible_nodes() / 1024,
+		order > MAX_PAGE_ORDER ? "vmalloc" : "linear");
 	return 0;
 }
 core_initcall(futex_init);
diff --git a/kernel/futex/futex.h b/kernel/futex/futex.h
index 899aed5acde12..acc7953678898 100644
--- a/kernel/futex/futex.h
+++ b/kernel/futex/futex.h
@@ -54,7 +54,7 @@ static inline unsigned int futex_to_flags(unsigned int op)
 	return flags;
 }
 
-#define FUTEX2_VALID_MASK (FUTEX2_SIZE_MASK | FUTEX2_PRIVATE)
+#define FUTEX2_VALID_MASK (FUTEX2_SIZE_MASK | FUTEX2_NUMA | FUTEX2_PRIVATE)
 
 /* FUTEX2_ to FLAGS_ */
 static inline unsigned int futex2_to_flags(unsigned int flags2)
@@ -87,6 +87,19 @@ static inline bool futex_flags_valid(unsigned int flags)
 	if ((flags & FLAGS_SIZE_MASK) != FLAGS_SIZE_32)
 		return false;
 
+	/*
+	 * Must be able to represent both FUTEX_NO_NODE and every valid nodeid
+	 * in a futex word.
+	 */
+	if (flags & FLAGS_NUMA) {
+		int bits = 8 * futex_size(flags);
+		u64 max = ~0ULL;
+
+		max >>= 64 - bits;
+		if (nr_node_ids >= max)
+			return false;
+	}
+
 	return true;
 }
 
@@ -282,7 +295,7 @@ static inline int futex_cmpxchg_value_locked(u32 *curval, u32 __user *uaddr, u32
  * This looks a bit overkill, but generally just results in a couple
  * of instructions.
  */
-static __always_inline int futex_read_inatomic(u32 *dest, u32 __user *from)
+static __always_inline int futex_get_value(u32 *dest, u32 __user *from)
 {
 	u32 val;
 
@@ -299,12 +312,26 @@ static __always_inline int futex_read_inatomic(u32 *dest, u32 __user *from)
 	return -EFAULT;
 }
 
+static __always_inline int futex_put_value(u32 val, u32 __user *to)
+{
+	if (can_do_masked_user_access())
+		to = masked_user_access_begin(to);
+	else if (!user_read_access_begin(to, sizeof(*to)))
+		return -EFAULT;
+	unsafe_put_user(val, to, Efault);
+	user_read_access_end();
+	return 0;
+Efault:
+	user_read_access_end();
+	return -EFAULT;
+}
+
 static inline int futex_get_value_locked(u32 *dest, u32 __user *from)
 {
 	int ret;
 
 	pagefault_disable();
-	ret = futex_read_inatomic(dest, from);
+	ret = futex_get_value(dest, from);
 	pagefault_enable();
 
 	return ret;
-- 
2.49.0

On 2025-04-07 17:57:38 [+0200], To linux-kernel@vger.kernel.org wrote:
> --- a/kernel/futex/core.c
> +++ b/kernel/futex/core.c
> @@ -332,15 +337,35 @@ __futex_hash(union futex_key *key, struct futex_private_hash *fph)
…
> +	if (node == FUTEX_NO_NODE) {
> +		/*
> +		 * In case of !FLAGS_NUMA, use some unused hash bits to pick a
> +		 * node -- this ensures regular futexes are interleaved across
> +		 * the nodes and avoids having to allocate multiple
> +		 * hash-tables.
> +		 *
> +		 * NOTE: this isn't perfectly uniform, but it is fast and
> +		 * handles sparse node masks.
> +		 */
> +		node = (hash >> futex_hashshift) % nr_node_ids;

forgot to mention earlier: This % nr_node_ids turns into div and it is
visible in perf top while looking at __futex_hash(). We could round it
down to a power-of-two (which should be the case in my 1, 2 and 4 based
NUMA world) and then we could use AND instead.
ARM does not support NUMA or div so it is not a concern.

Maybe a fast path for 1/2/4 would make sense since it is the most common
one. In case you consider it I could run test to see how significant it
is. It might be that it pops up in "perf bench futex hash" but not be
significant in general use case. I had some hacks and those did not
improve the numbers as much as I hoped for.

> +		if (!node_possible(node)) {
> +			node = find_next_bit_wrap(node_possible_map.bits,
> +						  nr_node_ids, node);
> +		}
> +	}
> +
> +	return &futex_queues[node][hash & futex_hashmask];
>  }
>  
>  /**

Sebastian

On 2025-04-07 18:52:25 [+0200], To linux-kernel@vger.kernel.org wrote:
> On 2025-04-07 17:57:38 [+0200], To linux-kernel@vger.kernel.org wrote:
> > --- a/kernel/futex/core.c
> > +++ b/kernel/futex/core.c
> > @@ -332,15 +337,35 @@ __futex_hash(union futex_key *key, struct futex_private_hash *fph)
> …
> > +	if (node == FUTEX_NO_NODE) {
> > +		/*
> > +		 * In case of !FLAGS_NUMA, use some unused hash bits to pick a
> > +		 * node -- this ensures regular futexes are interleaved across
> > +		 * the nodes and avoids having to allocate multiple
> > +		 * hash-tables.
> > +		 *
> > +		 * NOTE: this isn't perfectly uniform, but it is fast and
> > +		 * handles sparse node masks.
> > +		 */
> > +		node = (hash >> futex_hashshift) % nr_node_ids;
> 
> forgot to mention earlier: This % nr_node_ids turns into div and it is
> visible in perf top while looking at __futex_hash(). We could round it
> down to a power-of-two (which should be the case in my 1, 2 and 4 based
> NUMA world) and then we could use AND instead.
> ARM does not support NUMA or div so it is not a concern.
> 
> Maybe a fast path for 1/2/4 would make sense since it is the most common
> one. In case you consider it I could run test to see how significant it
> is. It might be that it pops up in "perf bench futex hash" but not be
> significant in general use case. I had some hacks and those did not
> improve the numbers as much as I hoped for.

Since I'm cleaning up: I'm getting approx 1% improvement with this so I
am not considering it.

Sebastian