Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
Merge branch 'tcp-remove-net-core-request_sock-c-and-no-longer-inline-__reqsk_free'
Eric Dumazet says:
====================
tcp: remove net/core/request_sock.c and no longer inline __reqsk_free()
After DCCP removal, net/core/request_sock.c makes no more sense.
Move reqsk_queue_alloc() and reqsk_fastopen_remove() to TCP files.
Then put __reqsk_free() out of line to save ~2 Kbytes of text.
====================
Link: https://patch.msgid.link/20260204055147.1682705-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
+107
-138
+1
-10
include/net/request_sock.h
+1
-10
include/net/request_sock.h
···
123
123
return sk;
124
124
}
125
125
126
-
static inline void __reqsk_free(struct request_sock *req)
127
-
{
128
-
req->rsk_ops->destructor(req);
129
-
if (req->rsk_listener)
130
-
sock_put(req->rsk_listener);
131
-
kfree(req->saved_syn);
132
-
kmem_cache_free(req->rsk_ops->slab, req);
133
-
}
126
+
void __reqsk_free(struct request_sock *req);
134
127
135
128
static inline void reqsk_free(struct request_sock *req)
136
129
{
···
188
195
* if TFO is enabled.
189
196
*/
190
197
};
191
-
192
-
void reqsk_queue_alloc(struct request_sock_queue *queue);
193
198
194
199
void reqsk_fastopen_remove(struct sock *sk, struct request_sock *req,
195
200
bool reset);
+1
-1
net/core/Makefile
+1
-1
net/core/Makefile
-127
net/core/request_sock.c
-127
net/core/request_sock.c
···
1
-
// SPDX-License-Identifier: GPL-2.0-or-later
2
-
/*
3
-
* NET Generic infrastructure for Network protocols.
4
-
*
5
-
* Authors: Arnaldo Carvalho de Melo <acme@conectiva.com.br>
6
-
*
7
-
* From code originally in include/net/tcp.h
8
-
*/
9
-
10
-
#include <linux/module.h>
11
-
#include <linux/random.h>
12
-
#include <linux/slab.h>
13
-
#include <linux/string.h>
14
-
#include <linux/tcp.h>
15
-
#include <linux/vmalloc.h>
16
-
17
-
#include <net/request_sock.h>
18
-
19
-
/*
20
-
* Maximum number of SYN_RECV sockets in queue per LISTEN socket.
21
-
* One SYN_RECV socket costs about 80bytes on a 32bit machine.
22
-
* It would be better to replace it with a global counter for all sockets
23
-
* but then some measure against one socket starving all other sockets
24
-
* would be needed.
25
-
*
26
-
* The minimum value of it is 128. Experiments with real servers show that
27
-
* it is absolutely not enough even at 100conn/sec. 256 cures most
28
-
* of problems.
29
-
* This value is adjusted to 128 for low memory machines,
30
-
* and it will increase in proportion to the memory of machine.
31
-
* Note : Dont forget somaxconn that may limit backlog too.
32
-
*/
33
-
34
-
void reqsk_queue_alloc(struct request_sock_queue *queue)
35
-
{
36
-
queue->fastopenq.rskq_rst_head = NULL;
37
-
queue->fastopenq.rskq_rst_tail = NULL;
38
-
queue->fastopenq.qlen = 0;
39
-
40
-
queue->rskq_accept_head = NULL;
41
-
}
42
-
43
-
/*
44
-
* This function is called to set a Fast Open socket's "fastopen_rsk" field
45
-
* to NULL when a TFO socket no longer needs to access the request_sock.
46
-
* This happens only after 3WHS has been either completed or aborted (e.g.,
47
-
* RST is received).
48
-
*
49
-
* Before TFO, a child socket is created only after 3WHS is completed,
50
-
* hence it never needs to access the request_sock. things get a lot more
51
-
* complex with TFO. A child socket, accepted or not, has to access its
52
-
* request_sock for 3WHS processing, e.g., to retransmit SYN-ACK pkts,
53
-
* until 3WHS is either completed or aborted. Afterwards the req will stay
54
-
* until either the child socket is accepted, or in the rare case when the
55
-
* listener is closed before the child is accepted.
56
-
*
57
-
* In short, a request socket is only freed after BOTH 3WHS has completed
58
-
* (or aborted) and the child socket has been accepted (or listener closed).
59
-
* When a child socket is accepted, its corresponding req->sk is set to
60
-
* NULL since it's no longer needed. More importantly, "req->sk == NULL"
61
-
* will be used by the code below to determine if a child socket has been
62
-
* accepted or not, and the check is protected by the fastopenq->lock
63
-
* described below.
64
-
*
65
-
* Note that fastopen_rsk is only accessed from the child socket's context
66
-
* with its socket lock held. But a request_sock (req) can be accessed by
67
-
* both its child socket through fastopen_rsk, and a listener socket through
68
-
* icsk_accept_queue.rskq_accept_head. To protect the access a simple spin
69
-
* lock per listener "icsk->icsk_accept_queue.fastopenq->lock" is created.
70
-
* only in the rare case when both the listener and the child locks are held,
71
-
* e.g., in inet_csk_listen_stop() do we not need to acquire the lock.
72
-
* The lock also protects other fields such as fastopenq->qlen, which is
73
-
* decremented by this function when fastopen_rsk is no longer needed.
74
-
*
75
-
* Note that another solution was to simply use the existing socket lock
76
-
* from the listener. But first socket lock is difficult to use. It is not
77
-
* a simple spin lock - one must consider sock_owned_by_user() and arrange
78
-
* to use sk_add_backlog() stuff. But what really makes it infeasible is the
79
-
* locking hierarchy violation. E.g., inet_csk_listen_stop() may try to
80
-
* acquire a child's lock while holding listener's socket lock.
81
-
*
82
-
* This function also sets "treq->tfo_listener" to false.
83
-
* treq->tfo_listener is used by the listener so it is protected by the
84
-
* fastopenq->lock in this function.
85
-
*/
86
-
void reqsk_fastopen_remove(struct sock *sk, struct request_sock *req,
87
-
bool reset)
88
-
{
89
-
struct sock *lsk = req->rsk_listener;
90
-
struct fastopen_queue *fastopenq;
91
-
92
-
fastopenq = &inet_csk(lsk)->icsk_accept_queue.fastopenq;
93
-
94
-
RCU_INIT_POINTER(tcp_sk(sk)->fastopen_rsk, NULL);
95
-
spin_lock_bh(&fastopenq->lock);
96
-
fastopenq->qlen--;
97
-
tcp_rsk(req)->tfo_listener = false;
98
-
if (req->sk) /* the child socket hasn't been accepted yet */
99
-
goto out;
100
-
101
-
if (!reset || lsk->sk_state != TCP_LISTEN) {
102
-
/* If the listener has been closed don't bother with the
103
-
* special RST handling below.
104
-
*/
105
-
spin_unlock_bh(&fastopenq->lock);
106
-
reqsk_put(req);
107
-
return;
108
-
}
109
-
/* Wait for 60secs before removing a req that has triggered RST.
110
-
* This is a simple defense against TFO spoofing attack - by
111
-
* counting the req against fastopen.max_qlen, and disabling
112
-
* TFO when the qlen exceeds max_qlen.
113
-
*
114
-
* For more details see CoNext'11 "TCP Fast Open" paper.
115
-
*/
116
-
req->rsk_timer.expires = jiffies + 60*HZ;
117
-
if (fastopenq->rskq_rst_head == NULL)
118
-
fastopenq->rskq_rst_head = req;
119
-
else
120
-
fastopenq->rskq_rst_tail->dl_next = req;
121
-
122
-
req->dl_next = NULL;
123
-
fastopenq->rskq_rst_tail = req;
124
-
fastopenq->qlen++;
125
-
out:
126
-
spin_unlock_bh(&fastopenq->lock);
127
-
}
+19
net/ipv4/inet_connection_sock.c
+19
net/ipv4/inet_connection_sock.c
···
919
919
}
920
920
EXPORT_SYMBOL(inet_reqsk_alloc);
921
921
922
+
void __reqsk_free(struct request_sock *req)
923
+
{
924
+
req->rsk_ops->destructor(req);
925
+
if (req->rsk_listener)
926
+
sock_put(req->rsk_listener);
927
+
kfree(req->saved_syn);
928
+
kmem_cache_free(req->rsk_ops->slab, req);
929
+
}
930
+
EXPORT_SYMBOL_GPL(__reqsk_free);
931
+
922
932
static struct request_sock *inet_reqsk_clone(struct request_sock *req,
923
933
struct sock *sk)
924
934
{
···
1322
1312
return -EINVAL;
1323
1313
1324
1314
return 0;
1315
+
}
1316
+
1317
+
static void reqsk_queue_alloc(struct request_sock_queue *queue)
1318
+
{
1319
+
queue->fastopenq.rskq_rst_head = NULL;
1320
+
queue->fastopenq.rskq_rst_tail = NULL;
1321
+
queue->fastopenq.qlen = 0;
1322
+
1323
+
queue->rskq_accept_head = NULL;
1325
1324
}
1326
1325
1327
1326
int inet_csk_listen_start(struct sock *sk)
+86
net/ipv4/tcp_fastopen.c
+86
net/ipv4/tcp_fastopen.c
···
5
5
#include <net/tcp.h>
6
6
#include <net/busy_poll.h>
7
7
8
+
/*
9
+
* This function is called to set a Fast Open socket's "fastopen_rsk" field
10
+
* to NULL when a TFO socket no longer needs to access the request_sock.
11
+
* This happens only after 3WHS has been either completed or aborted (e.g.,
12
+
* RST is received).
13
+
*
14
+
* Before TFO, a child socket is created only after 3WHS is completed,
15
+
* hence it never needs to access the request_sock. things get a lot more
16
+
* complex with TFO. A child socket, accepted or not, has to access its
17
+
* request_sock for 3WHS processing, e.g., to retransmit SYN-ACK pkts,
18
+
* until 3WHS is either completed or aborted. Afterwards the req will stay
19
+
* until either the child socket is accepted, or in the rare case when the
20
+
* listener is closed before the child is accepted.
21
+
*
22
+
* In short, a request socket is only freed after BOTH 3WHS has completed
23
+
* (or aborted) and the child socket has been accepted (or listener closed).
24
+
* When a child socket is accepted, its corresponding req->sk is set to
25
+
* NULL since it's no longer needed. More importantly, "req->sk == NULL"
26
+
* will be used by the code below to determine if a child socket has been
27
+
* accepted or not, and the check is protected by the fastopenq->lock
28
+
* described below.
29
+
*
30
+
* Note that fastopen_rsk is only accessed from the child socket's context
31
+
* with its socket lock held. But a request_sock (req) can be accessed by
32
+
* both its child socket through fastopen_rsk, and a listener socket through
33
+
* icsk_accept_queue.rskq_accept_head. To protect the access a simple spin
34
+
* lock per listener "icsk->icsk_accept_queue.fastopenq->lock" is created.
35
+
* only in the rare case when both the listener and the child locks are held,
36
+
* e.g., in inet_csk_listen_stop() do we not need to acquire the lock.
37
+
* The lock also protects other fields such as fastopenq->qlen, which is
38
+
* decremented by this function when fastopen_rsk is no longer needed.
39
+
*
40
+
* Note that another solution was to simply use the existing socket lock
41
+
* from the listener. But first socket lock is difficult to use. It is not
42
+
* a simple spin lock - one must consider sock_owned_by_user() and arrange
43
+
* to use sk_add_backlog() stuff. But what really makes it infeasible is the
44
+
* locking hierarchy violation. E.g., inet_csk_listen_stop() may try to
45
+
* acquire a child's lock while holding listener's socket lock.
46
+
*
47
+
* This function also sets "treq->tfo_listener" to false.
48
+
* treq->tfo_listener is used by the listener so it is protected by the
49
+
* fastopenq->lock in this function.
50
+
*/
51
+
void reqsk_fastopen_remove(struct sock *sk, struct request_sock *req,
52
+
bool reset)
53
+
{
54
+
struct sock *lsk = req->rsk_listener;
55
+
struct fastopen_queue *fastopenq;
56
+
57
+
fastopenq = &inet_csk(lsk)->icsk_accept_queue.fastopenq;
58
+
59
+
RCU_INIT_POINTER(tcp_sk(sk)->fastopen_rsk, NULL);
60
+
spin_lock_bh(&fastopenq->lock);
61
+
fastopenq->qlen--;
62
+
tcp_rsk(req)->tfo_listener = false;
63
+
if (req->sk) /* the child socket hasn't been accepted yet */
64
+
goto out;
65
+
66
+
if (!reset || lsk->sk_state != TCP_LISTEN) {
67
+
/* If the listener has been closed don't bother with the
68
+
* special RST handling below.
69
+
*/
70
+
spin_unlock_bh(&fastopenq->lock);
71
+
reqsk_put(req);
72
+
return;
73
+
}
74
+
/* Wait for 60secs before removing a req that has triggered RST.
75
+
* This is a simple defense against TFO spoofing attack - by
76
+
* counting the req against fastopen.max_qlen, and disabling
77
+
* TFO when the qlen exceeds max_qlen.
78
+
*
79
+
* For more details see CoNext'11 "TCP Fast Open" paper.
80
+
*/
81
+
req->rsk_timer.expires = jiffies + 60*HZ;
82
+
if (fastopenq->rskq_rst_head == NULL)
83
+
fastopenq->rskq_rst_head = req;
84
+
else
85
+
fastopenq->rskq_rst_tail->dl_next = req;
86
+
87
+
req->dl_next = NULL;
88
+
fastopenq->rskq_rst_tail = req;
89
+
fastopenq->qlen++;
90
+
out:
91
+
spin_unlock_bh(&fastopenq->lock);
92
+
}
93
+
8
94
void tcp_fastopen_init_key_once(struct net *net)
9
95
{
10
96
u8 key[TCP_FASTOPEN_KEY_LENGTH];