Commit | Line | Data |
---|---|---|
0e87506f ACM |
1 | /* |
2 | * NET Generic infrastructure for Network protocols. | |
3 | * | |
4 | * Authors: Arnaldo Carvalho de Melo <acme@conectiva.com.br> | |
5 | * | |
6 | * From code originally in include/net/tcp.h | |
7 | * | |
8 | * This program is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU General Public License | |
10 | * as published by the Free Software Foundation; either version | |
11 | * 2 of the License, or (at your option) any later version. | |
12 | */ | |
13 | ||
14 | #include <linux/module.h> | |
15 | #include <linux/random.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/string.h> | |
8336886f | 18 | #include <linux/tcp.h> |
72a3effa | 19 | #include <linux/vmalloc.h> |
0e87506f ACM |
20 | |
21 | #include <net/request_sock.h> | |
22 | ||
e52c1f17 DM |
23 | /* |
24 | * Maximum number of SYN_RECV sockets in queue per LISTEN socket. | |
25 | * One SYN_RECV socket costs about 80bytes on a 32bit machine. | |
26 | * It would be better to replace it with a global counter for all sockets | |
27 | * but then some measure against one socket starving all other sockets | |
28 | * would be needed. | |
29 | * | |
99b53bdd | 30 | * The minimum value of it is 128. Experiments with real servers show that |
e52c1f17 | 31 | * it is absolutely not enough even at 100conn/sec. 256 cures most |
99b53bdd PP |
32 | * of problems. |
33 | * This value is adjusted to 128 for low memory machines, | |
34 | * and it will increase in proportion to the memory of machine. | |
72a3effa | 35 | * Note : Dont forget somaxconn that may limit backlog too. |
e52c1f17 DM |
36 | */ |
37 | int sysctl_max_syn_backlog = 256; | |
493f377d | 38 | EXPORT_SYMBOL(sysctl_max_syn_backlog); |
e52c1f17 | 39 | |
10cbc8f1 ED |
40 | void reqsk_queue_alloc(struct request_sock_queue *queue, |
41 | unsigned int nr_table_entries) | |
0e87506f | 42 | { |
72a3effa ED |
43 | nr_table_entries = min_t(u32, nr_table_entries, sysctl_max_syn_backlog); |
44 | nr_table_entries = max_t(u32, nr_table_entries, 8); | |
45 | nr_table_entries = roundup_pow_of_two(nr_table_entries + 1); | |
f6d8cb2e | 46 | |
fff1f300 | 47 | spin_lock_init(&queue->rskq_lock); |
0536fcc0 ED |
48 | |
49 | spin_lock_init(&queue->fastopenq.lock); | |
50 | queue->fastopenq.rskq_rst_head = NULL; | |
51 | queue->fastopenq.rskq_rst_tail = NULL; | |
52 | queue->fastopenq.qlen = 0; | |
53 | queue->fastopenq.max_qlen = 0; | |
54 | ||
3eb4801d | 55 | queue->rskq_accept_head = NULL; |
10cbc8f1 | 56 | queue->max_qlen_log = ilog2(nr_table_entries); |
83e3609e ACM |
57 | } |
58 | ||
8336886f JC |
59 | /* |
60 | * This function is called to set a Fast Open socket's "fastopen_rsk" field | |
61 | * to NULL when a TFO socket no longer needs to access the request_sock. | |
62 | * This happens only after 3WHS has been either completed or aborted (e.g., | |
63 | * RST is received). | |
64 | * | |
65 | * Before TFO, a child socket is created only after 3WHS is completed, | |
66 | * hence it never needs to access the request_sock. things get a lot more | |
67 | * complex with TFO. A child socket, accepted or not, has to access its | |
68 | * request_sock for 3WHS processing, e.g., to retransmit SYN-ACK pkts, | |
69 | * until 3WHS is either completed or aborted. Afterwards the req will stay | |
70 | * until either the child socket is accepted, or in the rare case when the | |
71 | * listener is closed before the child is accepted. | |
72 | * | |
73 | * In short, a request socket is only freed after BOTH 3WHS has completed | |
74 | * (or aborted) and the child socket has been accepted (or listener closed). | |
75 | * When a child socket is accepted, its corresponding req->sk is set to | |
76 | * NULL since it's no longer needed. More importantly, "req->sk == NULL" | |
77 | * will be used by the code below to determine if a child socket has been | |
78 | * accepted or not, and the check is protected by the fastopenq->lock | |
79 | * described below. | |
80 | * | |
81 | * Note that fastopen_rsk is only accessed from the child socket's context | |
82 | * with its socket lock held. But a request_sock (req) can be accessed by | |
83 | * both its child socket through fastopen_rsk, and a listener socket through | |
84 | * icsk_accept_queue.rskq_accept_head. To protect the access a simple spin | |
85 | * lock per listener "icsk->icsk_accept_queue.fastopenq->lock" is created. | |
86 | * only in the rare case when both the listener and the child locks are held, | |
87 | * e.g., in inet_csk_listen_stop() do we not need to acquire the lock. | |
88 | * The lock also protects other fields such as fastopenq->qlen, which is | |
89 | * decremented by this function when fastopen_rsk is no longer needed. | |
90 | * | |
91 | * Note that another solution was to simply use the existing socket lock | |
92 | * from the listener. But first socket lock is difficult to use. It is not | |
93 | * a simple spin lock - one must consider sock_owned_by_user() and arrange | |
94 | * to use sk_add_backlog() stuff. But what really makes it infeasible is the | |
95 | * locking hierarchy violation. E.g., inet_csk_listen_stop() may try to | |
96 | * acquire a child's lock while holding listener's socket lock. A corner | |
97 | * case might also exist in tcp_v4_hnd_req() that will trigger this locking | |
98 | * order. | |
99 | * | |
9439ce00 ED |
100 | * This function also sets "treq->tfo_listener" to false. |
101 | * treq->tfo_listener is used by the listener so it is protected by the | |
8336886f JC |
102 | * fastopenq->lock in this function. |
103 | */ | |
104 | void reqsk_fastopen_remove(struct sock *sk, struct request_sock *req, | |
105 | bool reset) | |
106 | { | |
9439ce00 ED |
107 | struct sock *lsk = req->rsk_listener; |
108 | struct fastopen_queue *fastopenq; | |
109 | ||
0536fcc0 | 110 | fastopenq = &inet_csk(lsk)->icsk_accept_queue.fastopenq; |
8336886f | 111 | |
8336886f JC |
112 | tcp_sk(sk)->fastopen_rsk = NULL; |
113 | spin_lock_bh(&fastopenq->lock); | |
114 | fastopenq->qlen--; | |
9439ce00 | 115 | tcp_rsk(req)->tfo_listener = false; |
8336886f JC |
116 | if (req->sk) /* the child socket hasn't been accepted yet */ |
117 | goto out; | |
118 | ||
119 | if (!reset || lsk->sk_state != TCP_LISTEN) { | |
120 | /* If the listener has been closed don't bother with the | |
121 | * special RST handling below. | |
122 | */ | |
123 | spin_unlock_bh(&fastopenq->lock); | |
13854e5a | 124 | reqsk_put(req); |
8336886f JC |
125 | return; |
126 | } | |
127 | /* Wait for 60secs before removing a req that has triggered RST. | |
128 | * This is a simple defense against TFO spoofing attack - by | |
129 | * counting the req against fastopen.max_qlen, and disabling | |
130 | * TFO when the qlen exceeds max_qlen. | |
131 | * | |
132 | * For more details see CoNext'11 "TCP Fast Open" paper. | |
133 | */ | |
fa76ce73 | 134 | req->rsk_timer.expires = jiffies + 60*HZ; |
8336886f JC |
135 | if (fastopenq->rskq_rst_head == NULL) |
136 | fastopenq->rskq_rst_head = req; | |
137 | else | |
138 | fastopenq->rskq_rst_tail->dl_next = req; | |
139 | ||
140 | req->dl_next = NULL; | |
141 | fastopenq->rskq_rst_tail = req; | |
142 | fastopenq->qlen++; | |
143 | out: | |
144 | spin_unlock_bh(&fastopenq->lock); | |
8336886f | 145 | } |