Merge branch 'net-optmem_max-changes'

Eric Dumazet says:

====================
net: optmem_max changes

optmem_max default value is too small for tx zerocopy workloads.

First patch increases default from 20KB to 128 KB,
which is the value we have used for seven years.

Second patch makes optmem_max sysctl per netns.

Last patch tweaks two tests accordingly.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Documentation/admin-guide/sysctl/net.rst

@@ -345,7 +345,10 @@ optmem_max
 ----------
 
 Maximum ancillary buffer size allowed per socket. Ancillary data is a sequence
-of struct cmsghdr structures with appended data.
+of struct cmsghdr structures with appended data. TCP tx zerocopy also uses
+optmem_max as a limit for its internal structures.
+
+Default : 128 KB
 
 fb_tunnels_only_for_init_net
 ----------------------------

include/net/netns/core.h

@@ -13,6 +13,7 @@ struct netns_core {
 	struct ctl_table_header	*sysctl_hdr;
 
 	int	sysctl_somaxconn;
+	int	sysctl_optmem_max;
 	u8	sysctl_txrehash;
 
 #ifdef CONFIG_PROC_FS

include/net/sock.h

@@ -2920,7 +2920,6 @@ extern __u32 sysctl_wmem_max;
 extern __u32 sysctl_rmem_max;
 
 extern int sysctl_tstamp_allow_data;
-extern int sysctl_optmem_max;
 
 extern __u32 sysctl_wmem_default;
 extern __u32 sysctl_rmem_default;

net/core/bpf_sk_storage.c

@@ -275,9 +275,10 @@ BPF_CALL_2(bpf_sk_storage_delete, struct bpf_map *, map, struct sock *, sk)
 static int bpf_sk_storage_charge(struct bpf_local_storage_map *smap,
 				 void *owner, u32 size)
 {
-	int optmem_max = READ_ONCE(sysctl_optmem_max);
 	struct sock *sk = (struct sock *)owner;
+	int optmem_max;
 
+	optmem_max = READ_ONCE(sock_net(sk)->core.sysctl_optmem_max);
 	/* same check as in sock_kmalloc() */
 	if (size <= optmem_max &&
 	    atomic_read(&sk->sk_omem_alloc) + size < optmem_max) {

net/core/filter.c

@@ -1219,8 +1219,8 @@ void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
  */
 static bool __sk_filter_charge(struct sock *sk, struct sk_filter *fp)
 {
+	int optmem_max = READ_ONCE(sock_net(sk)->core.sysctl_optmem_max);
 	u32 filter_size = bpf_prog_size(fp->prog->len);
-	int optmem_max = READ_ONCE(sysctl_optmem_max);
 
 	/* same check as in sock_kmalloc() */
 	if (filter_size <= optmem_max &&
@@ -1550,12 +1550,13 @@ EXPORT_SYMBOL_GPL(sk_attach_filter);
 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
 {
 	struct bpf_prog *prog = __get_filter(fprog, sk);
-	int err;
+	int err, optmem_max;
 
 	if (IS_ERR(prog))
 		return PTR_ERR(prog);
 
-	if (bpf_prog_size(prog->len) > READ_ONCE(sysctl_optmem_max))
+	optmem_max = READ_ONCE(sock_net(sk)->core.sysctl_optmem_max);
+	if (bpf_prog_size(prog->len) > optmem_max)
 		err = -ENOMEM;
 	else
 		err = reuseport_attach_prog(sk, prog);
@@ -1594,7 +1595,7 @@ int sk_attach_bpf(u32 ufd, struct sock *sk)
 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
 {
 	struct bpf_prog *prog;
-	int err;
+	int err, optmem_max;
 
 	if (sock_flag(sk, SOCK_FILTER_LOCKED))
 		return -EPERM;
@@ -1622,7 +1623,8 @@ int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
 		}
 	} else {
 		/* BPF_PROG_TYPE_SOCKET_FILTER */
-		if (bpf_prog_size(prog->len) > READ_ONCE(sysctl_optmem_max)) {
+		optmem_max = READ_ONCE(sock_net(sk)->core.sysctl_optmem_max);
+		if (bpf_prog_size(prog->len) > optmem_max) {
 			err = -ENOMEM;
 			goto err_prog_put;
 		}

net/core/net_namespace.c

@@ -372,6 +372,10 @@ static __net_init int setup_net(struct net *net, struct user_namespace *user_ns)
 static int __net_init net_defaults_init_net(struct net *net)
 {
 	net->core.sysctl_somaxconn = SOMAXCONN;
+	/* Limits per socket sk_omem_alloc usage.
+	 * TCP zerocopy regular usage needs 128 KB.
+	 */
+	net->core.sysctl_optmem_max = 128 * 1024;
 	net->core.sysctl_txrehash = SOCK_TXREHASH_ENABLED;
 
 	return 0;

net/core/sock.c

@@ -283,10 +283,6 @@ EXPORT_SYMBOL(sysctl_rmem_max);
 __u32 sysctl_wmem_default __read_mostly = SK_WMEM_MAX;
 __u32 sysctl_rmem_default __read_mostly = SK_RMEM_MAX;
 
-/* Maximal space eaten by iovec or ancillary data plus some space */
-int sysctl_optmem_max __read_mostly = sizeof(unsigned long)*(2*UIO_MAXIOV+512);
-EXPORT_SYMBOL(sysctl_optmem_max);
-
 int sysctl_tstamp_allow_data __read_mostly = 1;
 
 DEFINE_STATIC_KEY_FALSE(memalloc_socks_key);
@@ -2651,7 +2647,7 @@ struct sk_buff *sock_omalloc(struct sock *sk, unsigned long size,
 
 	/* small safe race: SKB_TRUESIZE may differ from final skb->truesize */
 	if (atomic_read(&sk->sk_omem_alloc) + SKB_TRUESIZE(size) >
-	    READ_ONCE(sysctl_optmem_max))
+	    READ_ONCE(sock_net(sk)->core.sysctl_optmem_max))
 		return NULL;
 
 	skb = alloc_skb(size, priority);
@@ -2669,7 +2665,7 @@ struct sk_buff *sock_omalloc(struct sock *sk, unsigned long size,
  */
 void *sock_kmalloc(struct sock *sk, int size, gfp_t priority)
 {
-	int optmem_max = READ_ONCE(sysctl_optmem_max);
+	int optmem_max = READ_ONCE(sock_net(sk)->core.sysctl_optmem_max);
 
 	if ((unsigned int)size <= optmem_max &&
 	    atomic_read(&sk->sk_omem_alloc) + size < optmem_max) {

net/core/sysctl_net_core.c

@@ -508,13 +508,6 @@ static struct ctl_table net_core_table[] = {
 		.mode		= 0644,
 		.proc_handler	= proc_dointvec,
 	},
-	{
-		.procname	= "optmem_max",
-		.data		= &sysctl_optmem_max,
-		.maxlen		= sizeof(int),
-		.mode		= 0644,
-		.proc_handler	= proc_dointvec
-	},
 	{
 		.procname	= "tstamp_allow_data",
 		.data		= &sysctl_tstamp_allow_data,
@@ -673,6 +666,14 @@ static struct ctl_table netns_core_table[] = {
 		.extra1		= SYSCTL_ZERO,
 		.proc_handler	= proc_dointvec_minmax
 	},
+	{
+		.procname	= "optmem_max",
+		.data		= &init_net.core.sysctl_optmem_max,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.extra1		= SYSCTL_ZERO,
+		.proc_handler	= proc_dointvec_minmax
+	},
 	{
 		.procname	= "txrehash",
 		.data		= &init_net.core.sysctl_txrehash,

net/ipv4/ip_sockglue.c

@@ -775,7 +775,7 @@ static int ip_set_mcast_msfilter(struct sock *sk, sockptr_t optval, int optlen)
 
 	if (optlen < GROUP_FILTER_SIZE(0))
 		return -EINVAL;
-	if (optlen > READ_ONCE(sysctl_optmem_max))
+	if (optlen > READ_ONCE(sock_net(sk)->core.sysctl_optmem_max))
 		return -ENOBUFS;
 
 	gsf = memdup_sockptr(optval, optlen);
@@ -811,7 +811,7 @@ static int compat_ip_set_mcast_msfilter(struct sock *sk, sockptr_t optval,
 
 	if (optlen < size0)
 		return -EINVAL;
-	if (optlen > READ_ONCE(sysctl_optmem_max) - 4)
+	if (optlen > READ_ONCE(sock_net(sk)->core.sysctl_optmem_max) - 4)
 		return -ENOBUFS;
 
 	p = kmalloc(optlen + 4, GFP_KERNEL);
@@ -1254,7 +1254,7 @@ int do_ip_setsockopt(struct sock *sk, int level, int optname,
 
 		if (optlen < IP_MSFILTER_SIZE(0))
 			goto e_inval;
-		if (optlen > READ_ONCE(sysctl_optmem_max)) {
+		if (optlen > READ_ONCE(net->core.sysctl_optmem_max)) {
 			err = -ENOBUFS;
 			break;
 		}

net/ipv6/ipv6_sockglue.c

@@ -210,7 +210,7 @@ static int ipv6_set_mcast_msfilter(struct sock *sk, sockptr_t optval,
 
 	if (optlen < GROUP_FILTER_SIZE(0))
 		return -EINVAL;
-	if (optlen > READ_ONCE(sysctl_optmem_max))
+	if (optlen > READ_ONCE(sock_net(sk)->core.sysctl_optmem_max))
 		return -ENOBUFS;
 
 	gsf = memdup_sockptr(optval, optlen);
@@ -244,7 +244,7 @@ static int compat_ipv6_set_mcast_msfilter(struct sock *sk, sockptr_t optval,
 
 	if (optlen < size0)
 		return -EINVAL;
-	if (optlen > READ_ONCE(sysctl_optmem_max) - 4)
+	if (optlen > READ_ONCE(sock_net(sk)->core.sysctl_optmem_max) - 4)
 		return -ENOBUFS;
 
 	p = kmalloc(optlen + 4, GFP_KERNEL);

tools/testing/selftests/net/io_uring_zerocopy_tx.sh

@@ -76,23 +76,22 @@ case "${TXMODE}" in
 esac
 
 # Start of state changes: install cleanup handler
-save_sysctl_mem="$(sysctl -n ${path_sysctl_mem})"
 
 cleanup() {
 	ip netns del "${NS2}"
 	ip netns del "${NS1}"
-	sysctl -w -q "${path_sysctl_mem}=${save_sysctl_mem}"
 }
 
 trap cleanup EXIT
 
-# Configure system settings
-sysctl -w -q "${path_sysctl_mem}=1000000"
-
 # Create virtual ethernet pair between network namespaces
 ip netns add "${NS1}"
 ip netns add "${NS2}"
 
+# Configure system settings
+ip netns exec "${NS1}" sysctl -w -q "${path_sysctl_mem}=1000000"
+ip netns exec "${NS2}" sysctl -w -q "${path_sysctl_mem}=1000000"
+
 ip link add "${DEV}" mtu "${DEV_MTU}" netns "${NS1}" type veth \
   peer name "${DEV}" mtu "${DEV_MTU}" netns "${NS2}"
 

tools/testing/selftests/net/msg_zerocopy.sh

@@ -70,23 +70,22 @@ case "${TXMODE}" in
 esac
 
 # Start of state changes: install cleanup handler
-save_sysctl_mem="$(sysctl -n ${path_sysctl_mem})"
 
 cleanup() {
 	ip netns del "${NS2}"
 	ip netns del "${NS1}"
-	sysctl -w -q "${path_sysctl_mem}=${save_sysctl_mem}"
 }
 
 trap cleanup EXIT
 
-# Configure system settings
-sysctl -w -q "${path_sysctl_mem}=1000000"
-
 # Create virtual ethernet pair between network namespaces
 ip netns add "${NS1}"
 ip netns add "${NS2}"
 
+# Configure system settings
+ip netns exec "${NS1}" sysctl -w -q "${path_sysctl_mem}=1000000"
+ip netns exec "${NS2}" sysctl -w -q "${path_sysctl_mem}=1000000"
+
 ip link add "${DEV}" mtu "${DEV_MTU}" netns "${NS1}" type veth \
   peer name "${DEV}" mtu "${DEV_MTU}" netns "${NS2}"