Hi David
We currently allocate a fixed size 512 (TCP_SYNQ_HSIZE) slots hash table for
each LISTEN socket, regardless of various parameters (listen backlog for
example)
On x86_64, this means order-1 allocations (might fail), even for 'small'
sockets, expecting few connections. On the contrary, a huge server wanting a
backlog of 50000 is slowed down a bit because of this fixed limit.
This patch makes the sizing of listen hash table a dynamic parameter,
depending of :
- net.core.somaxconn tunable (/proc/sys/net/core/somaxconn , default is 128)
- net.ipv4.tcp_max_syn_backlog tunable (default : 256, 1024 or 128)
- backlog value given by user application (2nd parameter of listen())
- and available LOWMEM ram
reqsk_queue_alloc() goal is to use a power of two size for the whole
listen_sock structure, to avoid wasting memory for large backlogs, meaning
the hash table nr_table_entries is not anymore a power of two. (Hence one AND
(nr_table_entries - 1) must be replaced by MODULO nr_table_entries)
We still limit memory allocation with the two existing tunables (somaxconn &
tcp_max_syn_backlog).
In case memory allocation has problems, reqsk_queue_alloc() reduces the size
of the hash table to allow a successfull listen() call, without giving
feedback to user application, as this 'backlog' was advisory.
Thank you
include/net/request_sock.h | 8 ++++----
include/net/tcp.h | 1 -
net/core/request_sock.c | 39
+++++++++++++++++++++++++++++----------
net/dccp/ipv4.c | 2 +-
net/dccp/proto.c | 6 +++---
net/ipv4/af_inet.c | 2 +-
net/ipv4/inet_connection_sock.c | 8 +++++---
net/ipv4/tcp_ipv4.c | 6 +++---
net/ipv6/tcp_ipv6.c | 2 +-
9 files changed, 47 insertions(+), 27 deletions(-)
Signed-off-by: Eric Dumazet <[EMAIL PROTECTED]>
--- linux-2.6.19-rc2/net/core/request_sock.c 2006-10-13 18:25:04.000000000
+0200
+++ linux-2.6.19-rc2-ed/net/core/request_sock.c 2006-10-17 14:47:48.000000000
+0200
@@ -29,29 +29,48 @@
* it is absolutely not enough even at 100conn/sec. 256 cures most
* of problems. This value is adjusted to 128 for very small machines
* (<=32Mb of memory) and to 1024 on normal or better ones (>=256Mb).
- * Further increasing requires to change hash table size.
*/
int sysctl_max_syn_backlog = 256;
int reqsk_queue_alloc(struct request_sock_queue *queue,
- const int nr_table_entries)
+ u32 nr_entries)
{
- const int lopt_size = sizeof(struct listen_sock) +
- nr_table_entries * sizeof(struct request_sock *);
- struct listen_sock *lopt = kzalloc(lopt_size, GFP_KERNEL);
+ struct listen_sock *lopt;
+ size_t size = sizeof(struct listen_sock);
- if (lopt == NULL)
- return -ENOMEM;
+ nr_entries = min_t(u32, nr_entries, sysctl_max_syn_backlog);
+ nr_entries = max_t(u32, nr_entries, 8);
+ size += nr_entries*sizeof(struct request_sock *);
+ size = roundup_pow_of_two(size);
+ while (1) {
+ lopt = kzalloc(size, GFP_KERNEL);
+ if (lopt != NULL)
+ break;
+ size >>= 1;
+ if (size < sizeof(struct listen_sock) +
+ 8 * sizeof(struct request_sock *))
+ return -ENOMEM;
+ }
+ lopt->nr_table_entries = (size - sizeof(struct listen_sock)) /
+ sizeof(struct request_sock *);
- for (lopt->max_qlen_log = 6;
- (1 << lopt->max_qlen_log) < sysctl_max_syn_backlog;
+ /*
+ * max_qlen_log computation is based on the backlog (nr_entries),
+ * not on actual hash size (lopt->nr_table_entries).
+ */
+ for (lopt->max_qlen_log = 3;
+ (1 << lopt->max_qlen_log) < nr_entries;
lopt->max_qlen_log++);
get_random_bytes(&lopt->hash_rnd, sizeof(lopt->hash_rnd));
rwlock_init(&queue->syn_wait_lock);
queue->rskq_accept_head = NULL;
- lopt->nr_table_entries = nr_table_entries;
+ /*
+ * This write_lock_bh()/write_unlock_bh() pair forces this CPU to commit
+ * its memory changes and let readers (which acquire syn_wait_lock in
+ * reader mode) operate without seeing random content.
+ */
write_lock_bh(&queue->syn_wait_lock);
queue->listen_opt = lopt;
write_unlock_bh(&queue->syn_wait_lock);
--- linux-2.6.19-rc2/net/ipv4/af_inet.c 2006-10-13 18:25:04.000000000 +0200
+++ linux-2.6.19-rc2-ed/net/ipv4/af_inet.c 2006-10-17 10:32:22.000000000
+0200
@@ -204,7 +204,7 @@
* we can only allow the backlog to be adjusted.
*/
if (old_state != TCP_LISTEN) {
- err = inet_csk_listen_start(sk, TCP_SYNQ_HSIZE);
+ err = inet_csk_listen_start(sk, backlog);
if (err)
goto out;
}
--- linux-2.6.19-rc2/net/ipv4/tcp_ipv4.c 2006-10-13 18:25:04.000000000
+0200
+++ linux-2.6.19-rc2-ed/net/ipv4/tcp_ipv4.c 2006-10-17 12:19:38.000000000
+0200
@@ -715,7 +715,7 @@
return dopt;
}
-struct request_sock_ops tcp_request_sock_ops = {
+struct request_sock_ops tcp_request_sock_ops __read_mostly = {
.family = PF_INET,
.obj_size = sizeof(struct tcp_request_sock),
.rtx_syn_ack = tcp_v4_send_synack,
@@ -1385,7 +1385,7 @@
if (st->state == TCP_SEQ_STATE_OPENREQ) {
struct request_sock *req = cur;
- icsk = inet_csk(st->syn_wait_sk);
+ icsk = inet_csk(st->syn_wait_sk);
req = req->dl_next;
while (1) {
while (req) {
@@ -1395,7 +1395,7 @@
}
req = req->dl_next;
}
- if (++st->sbucket >= TCP_SYNQ_HSIZE)
+ if (++st->sbucket >=
icsk->icsk_accept_queue.listen_opt->nr_table_entries)
break;
get_req:
req =
icsk->icsk_accept_queue.listen_opt->syn_table[st->sbucket];
--- linux-2.6.19-rc2/net/dccp/proto.c 2006-10-13 18:25:04.000000000 +0200
+++ linux-2.6.19-rc2-ed/net/dccp/proto.c 2006-10-17 10:32:22.000000000
+0200
@@ -262,12 +262,12 @@
EXPORT_SYMBOL_GPL(dccp_destroy_sock);
-static inline int dccp_listen_start(struct sock *sk)
+static inline int dccp_listen_start(struct sock *sk, int backlog)
{
struct dccp_sock *dp = dccp_sk(sk);
dp->dccps_role = DCCP_ROLE_LISTEN;
- return inet_csk_listen_start(sk, TCP_SYNQ_HSIZE);
+ return inet_csk_listen_start(sk, backlog);
}
int dccp_disconnect(struct sock *sk, int flags)
@@ -788,7 +788,7 @@
* FIXME: here it probably should be sk->sk_prot->listen_start
* see tcp_listen_start
*/
- err = dccp_listen_start(sk);
+ err = dccp_listen_start(sk, backlog);
if (err)
goto out;
}
--- linux-2.6.19-rc2/net/dccp/ipv4.c 2006-10-13 18:25:04.000000000 +0200
+++ linux-2.6.19-rc2-ed/net/dccp/ipv4.c 2006-10-17 10:44:21.000000000 +0200
@@ -1020,7 +1020,7 @@
kfree(inet_rsk(req)->opt);
}
-static struct request_sock_ops dccp_request_sock_ops = {
+static struct request_sock_ops dccp_request_sock_ops _read_mostly = {
.family = PF_INET,
.obj_size = sizeof(struct dccp_request_sock),
.rtx_syn_ack = dccp_v4_send_response,
--- linux-2.6.19-rc2/net/ipv6/tcp_ipv6.c 2006-10-13 18:25:04.000000000
+0200
+++ linux-2.6.19-rc2-ed/net/ipv6/tcp_ipv6.c 2006-10-17 10:44:21.000000000
+0200
@@ -526,7 +526,7 @@
kfree_skb(inet6_rsk(req)->pktopts);
}
-static struct request_sock_ops tcp6_request_sock_ops = {
+static struct request_sock_ops tcp6_request_sock_ops _read_mostly = {
.family = AF_INET6,
.obj_size = sizeof(struct tcp6_request_sock),
.rtx_syn_ack = tcp_v6_send_synack,
--- linux-2.6.19-rc2/net/ipv4/inet_connection_sock.c 2006-10-13
18:25:04.000000000 +0200
+++ linux-2.6.19-rc2-ed/net/ipv4/inet_connection_sock.c 2006-10-17
10:32:22.000000000 +0200
@@ -343,9 +343,9 @@
EXPORT_SYMBOL_GPL(inet_csk_route_req);
static inline u32 inet_synq_hash(const __be32 raddr, const __be16 rport,
- const u32 rnd, const u16 synq_hsize)
+ const u32 rnd, const u32 synq_hsize)
{
- return jhash_2words((__force u32)raddr, (__force u32)rport, rnd) &
(synq_hsize - 1);
+ return jhash_2words((__force u32)raddr, (__force u32)rport, rnd) %
synq_hsize;
}
#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
@@ -478,7 +478,9 @@
reqp = &req->dl_next;
}
- i = (i + 1) & (lopt->nr_table_entries - 1);
+ i++;
+ if (i == lopt->nr_table_entries)
+ i = 0;
} while (--budget > 0);
--- linux-2.6.19-rc2/include/net/tcp.h 2006-10-13 18:25:04.000000000 +0200
+++ linux-2.6.19-rc2-ed/include/net/tcp.h 2006-10-17 10:51:51.000000000
+0200
@@ -138,7 +138,6 @@
#define MAX_TCP_SYNCNT 127
#define TCP_SYNQ_INTERVAL (HZ/5) /* Period of SYNACK timer */
-#define TCP_SYNQ_HSIZE 512 /* Size of SYNACK hash table */
#define TCP_PAWS_24DAYS (60 * 60 * 24 * 24)
#define TCP_PAWS_MSL 60 /* Per-host timestamps are invalidated
--- linux-2.6.19-rc2/include/net/request_sock.h 2006-10-13 18:25:04.000000000
+0200
+++ linux-2.6.19-rc2-ed/include/net/request_sock.h 2006-10-17
12:33:18.000000000 +0200
@@ -28,8 +28,8 @@
struct request_sock_ops {
int family;
- kmem_cache_t *slab;
int obj_size;
+ kmem_cache_t *slab;
int (*rtx_syn_ack)(struct sock *sk,
struct request_sock *req,
struct dst_entry *dst);
@@ -51,12 +51,12 @@
u32 rcv_wnd; /* rcv_wnd offered
first time */
u32 ts_recent;
unsigned long expires;
- struct request_sock_ops *rsk_ops;
+ const struct request_sock_ops *rsk_ops;
struct sock *sk;
u32 secid;
};
-static inline struct request_sock *reqsk_alloc(struct request_sock_ops *ops)
+static inline struct request_sock *reqsk_alloc(const struct request_sock_ops
*ops)
{
struct request_sock *req = kmem_cache_alloc(ops->slab, SLAB_ATOMIC);
@@ -120,7 +120,7 @@
};
extern int reqsk_queue_alloc(struct request_sock_queue *queue,
- const int nr_table_entries);
+ unsigned int nr_table_entries);
static inline struct listen_sock *reqsk_queue_yank_listen_sk(struct
request_sock_queue *queue)
{
