tcp: bring back NUMA dispersion in inet_ehash_locks_alloc()

We have platforms with 6 NUMA nodes and 480 cpus.

inet_ehash_locks_alloc() currently allocates a single 64KB page
to hold all ehash spinlocks. This adds more pressure on a single node.

Change inet_ehash_locks_alloc() to use vmalloc() to spread
the spinlocks on all online nodes, driven by NUMA policies.

At boot time, NUMA policy is interleave=all, meaning that
tcp_hashinfo.ehash_locks gets hash dispersion on all nodes.

Tested:

lack5:~# grep inet_ehash_locks_alloc /proc/vmallocinfo
0x00000000d9aec4d1-0x00000000a828b652   69632 inet_ehash_locks_alloc+0x90/0x100 pages=16 vmalloc N0=2 N1=3 N2=3 N3=3 N4=3 N5=2

lack5:~# echo 8192 >/proc/sys/net/ipv4/tcp_child_ehash_entries
lack5:~# numactl --interleave=all unshare -n bash -c "grep inet_ehash_locks_alloc /proc/vmallocinfo"
0x000000004e99d30c-0x00000000763f3279   36864 inet_ehash_locks_alloc+0x90/0x100 pages=8 vmalloc N0=1 N1=2 N2=2 N3=1 N4=1 N5=1
0x00000000d9aec4d1-0x00000000a828b652   69632 inet_ehash_locks_alloc+0x90/0x100 pages=16 vmalloc N0=2 N1=3 N2=3 N3=3 N4=3 N5=2

lack5:~# numactl --interleave=0,5 unshare -n bash -c "grep inet_ehash_locks_alloc /proc/vmallocinfo"
0x00000000fd73a33e-0x0000000004b9a177   36864 inet_ehash_locks_alloc+0x90/0x100 pages=8 vmalloc N0=4 N5=4
0x00000000d9aec4d1-0x00000000a828b652   69632 inet_ehash_locks_alloc+0x90/0x100 pages=16 vmalloc N0=2 N1=3 N2=3 N3=3 N4=3 N5=2

lack5:~# echo 1024 >/proc/sys/net/ipv4/tcp_child_ehash_entries
lack5:~# numactl --interleave=all unshare -n bash -c "grep inet_ehash_locks_alloc /proc/vmallocinfo"
0x00000000db07d7a2-0x00000000ad697d29    8192 inet_ehash_locks_alloc+0x90/0x100 pages=1 vmalloc N2=1
0x00000000d9aec4d1-0x00000000a828b652   69632 inet_ehash_locks_alloc+0x90/0x100 pages=16 vmalloc N0=2 N1=3 N2=3 N3=3 N4=3 N5=2

Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250305130550.1865988-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

1 files changed, 26 insertions, 11 deletions

diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
index d1b5f45ee718..3edbe2dad8ba 100644
--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -1261,22 +1261,37 @@ int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
 {
 	unsigned int locksz = sizeof(spinlock_t);
 	unsigned int i, nblocks = 1;
+	spinlock_t *ptr = NULL;
 
-	if (locksz != 0) {
-		/* allocate 2 cache lines or at least one spinlock per cpu */
-		nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
-		nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
+	if (locksz == 0)
+		goto set_mask;
 
-		/* no more locks than number of hash buckets */
-		nblocks = min(nblocks, hashinfo->ehash_mask + 1);
+	/* Allocate 2 cache lines or at least one spinlock per cpu. */
+	nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U) * num_possible_cpus();
 
-		hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
-		if (!hashinfo->ehash_locks)
-			return -ENOMEM;
+	/* At least one page per NUMA node. */
+	nblocks = max(nblocks, num_online_nodes() * PAGE_SIZE / locksz);
+
+	nblocks = roundup_pow_of_two(nblocks);
+
+	/* No more locks than number of hash buckets. */
+	nblocks = min(nblocks, hashinfo->ehash_mask + 1);
 
-		for (i = 0; i < nblocks; i++)
-			spin_lock_init(&hashinfo->ehash_locks[i]);
+	if (num_online_nodes() > 1) {
+		/* Use vmalloc() to allow NUMA policy to spread pages
+		 * on all available nodes if desired.
+		 */
+		ptr = vmalloc_array(nblocks, locksz);
+	}
+	if (!ptr) {
+		ptr = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
+		if (!ptr)
+			return -ENOMEM;
 	}
+	for (i = 0; i < nblocks; i++)
+		spin_lock_init(&ptr[i]);
+	hashinfo->ehash_locks = ptr;
+set_mask:
 	hashinfo->ehash_locks_mask = nblocks - 1;
 	return 0;
 }