net/core/request_sock.c

   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>
  18 #include <linux/tcp.h>
  19 #include <linux/vmalloc.h>
  20
  21 #include <net/request_sock.h>
  22
  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  *
  30  * The minimum value of it is 128. Experiments with real servers show that
  31  * it is absolutely not enough even at 100conn/sec. 256 cures most
  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.
  35  * Note : Dont forget somaxconn that may limit backlog too.
  36  */
  37 int sysctl_max_syn_backlog = 256;
  38 EXPORT_SYMBOL(sysctl_max_syn_backlog);
  39
  40 int reqsk_queue_alloc(struct request_sock_queue *queue,
  41                       unsigned int nr_table_entries)
  42 {
  43         size_t lopt_size = sizeof(struct listen_sock);
  44         struct listen_sock *lopt = NULL;
  45
  46         nr_table_entries = min_t(u32, nr_table_entries, sysctl_max_syn_backlog);
  47         nr_table_entries = max_t(u32, nr_table_entries, 8);
  48         nr_table_entries = roundup_pow_of_two(nr_table_entries + 1);
  49         lopt_size += nr_table_entries * sizeof(struct request_sock *);
  50
  51         if (lopt_size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER))
  52                 lopt = kzalloc(lopt_size, GFP_KERNEL |
  53                                           __GFP_NOWARN |
  54                                           __GFP_NORETRY);
  55         if (!lopt)
  56                 lopt = vzalloc(lopt_size);
  57         if (!lopt)
  58                 return -ENOMEM;
  59
  60         get_random_bytes(&lopt->hash_rnd, sizeof(lopt->hash_rnd));
  61         spin_lock_init(&queue->syn_wait_lock);
  62         queue->rskq_accept_head = NULL;
  63         lopt->nr_table_entries = nr_table_entries;
  64         lopt->max_qlen_log = ilog2(nr_table_entries);
  65
  66         spin_lock_bh(&queue->syn_wait_lock);
  67         queue->listen_opt = lopt;
  68         spin_unlock_bh(&queue->syn_wait_lock);
  69
  70         return 0;
  71 }
  72
  73 void __reqsk_queue_destroy(struct request_sock_queue *queue)
  74 {
  75         /* This is an error recovery path only, no locking needed */
  76         kvfree(queue->listen_opt);
  77 }
  78
  79 static inline struct listen_sock *reqsk_queue_yank_listen_sk(
  80                 struct request_sock_queue *queue)
  81 {
  82         struct listen_sock *lopt;
  83
  84         spin_lock_bh(&queue->syn_wait_lock);
  85         lopt = queue->listen_opt;
  86         queue->listen_opt = NULL;
  87         spin_unlock_bh(&queue->syn_wait_lock);
  88
  89         return lopt;
  90 }
  91
  92 void reqsk_queue_destroy(struct request_sock_queue *queue)
  93 {
  94         /* make all the listen_opt local to us */
  95         struct listen_sock *lopt = reqsk_queue_yank_listen_sk(queue);
  96
  97         if (listen_sock_qlen(lopt) != 0) {
  98                 unsigned int i;
  99
 100                 for (i = 0; i < lopt->nr_table_entries; i++) {
 101                         struct request_sock *req;
 102
 103                         spin_lock_bh(&queue->syn_wait_lock);
 104                         while ((req = lopt->syn_table[i]) != NULL) {
 105                                 lopt->syn_table[i] = req->dl_next;
 106                                 /* Because of following del_timer_sync(),
 107                                  * we must release the spinlock here
 108                                  * or risk a dead lock.
 109                                  */
 110                                 spin_unlock_bh(&queue->syn_wait_lock);
 111                                 atomic_inc(&lopt->qlen_dec);
 112                                 if (del_timer_sync(&req->rsk_timer))
 113                                         reqsk_put(req);
 114                                 reqsk_put(req);
 115                                 spin_lock_bh(&queue->syn_wait_lock);
 116                         }
 117                         spin_unlock_bh(&queue->syn_wait_lock);
 118                 }
 119         }
 120
 121         if (WARN_ON(listen_sock_qlen(lopt) != 0))
 122                 pr_err("qlen %u\n", listen_sock_qlen(lopt));
 123         kvfree(lopt);
 124 }
 125
 126 /*
 127  * This function is called to set a Fast Open socket's "fastopen_rsk" field
 128  * to NULL when a TFO socket no longer needs to access the request_sock.
 129  * This happens only after 3WHS has been either completed or aborted (e.g.,
 130  * RST is received).
 131  *
 132  * Before TFO, a child socket is created only after 3WHS is completed,
 133  * hence it never needs to access the request_sock. things get a lot more
 134  * complex with TFO. A child socket, accepted or not, has to access its
 135  * request_sock for 3WHS processing, e.g., to retransmit SYN-ACK pkts,
 136  * until 3WHS is either completed or aborted. Afterwards the req will stay
 137  * until either the child socket is accepted, or in the rare case when the
 138  * listener is closed before the child is accepted.
 139  *
 140  * In short, a request socket is only freed after BOTH 3WHS has completed
 141  * (or aborted) and the child socket has been accepted (or listener closed).
 142  * When a child socket is accepted, its corresponding req->sk is set to
 143  * NULL since it's no longer needed. More importantly, "req->sk == NULL"
 144  * will be used by the code below to determine if a child socket has been
 145  * accepted or not, and the check is protected by the fastopenq->lock
 146  * described below.
 147  *
 148  * Note that fastopen_rsk is only accessed from the child socket's context
 149  * with its socket lock held. But a request_sock (req) can be accessed by
 150  * both its child socket through fastopen_rsk, and a listener socket through
 151  * icsk_accept_queue.rskq_accept_head. To protect the access a simple spin
 152  * lock per listener "icsk->icsk_accept_queue.fastopenq->lock" is created.
 153  * only in the rare case when both the listener and the child locks are held,
 154  * e.g., in inet_csk_listen_stop() do we not need to acquire the lock.
 155  * The lock also protects other fields such as fastopenq->qlen, which is
 156  * decremented by this function when fastopen_rsk is no longer needed.
 157  *
 158  * Note that another solution was to simply use the existing socket lock
 159  * from the listener. But first socket lock is difficult to use. It is not
 160  * a simple spin lock - one must consider sock_owned_by_user() and arrange
 161  * to use sk_add_backlog() stuff. But what really makes it infeasible is the
 162  * locking hierarchy violation. E.g., inet_csk_listen_stop() may try to
 163  * acquire a child's lock while holding listener's socket lock. A corner
 164  * case might also exist in tcp_v4_hnd_req() that will trigger this locking
 165  * order.
 166  *
 167  * This function also sets "treq->tfo_listener" to false.
 168  * treq->tfo_listener is used by the listener so it is protected by the
 169  * fastopenq->lock in this function.
 170  */
 171 void reqsk_fastopen_remove(struct sock *sk, struct request_sock *req,
 172                            bool reset)
 173 {
 174         struct sock *lsk = req->rsk_listener;
 175         struct fastopen_queue *fastopenq;
 176
 177         fastopenq = inet_csk(lsk)->icsk_accept_queue.fastopenq;
 178
 179         tcp_sk(sk)->fastopen_rsk = NULL;
 180         spin_lock_bh(&fastopenq->lock);
 181         fastopenq->qlen--;
 182         tcp_rsk(req)->tfo_listener = false;
 183         if (req->sk)    /* the child socket hasn't been accepted yet */
 184                 goto out;
 185
 186         if (!reset || lsk->sk_state != TCP_LISTEN) {
 187                 /* If the listener has been closed don't bother with the
 188                  * special RST handling below.
 189                  */
 190                 spin_unlock_bh(&fastopenq->lock);
 191                 reqsk_put(req);
 192                 return;
 193         }
 194         /* Wait for 60secs before removing a req that has triggered RST.
 195          * This is a simple defense against TFO spoofing attack - by
 196          * counting the req against fastopen.max_qlen, and disabling
 197          * TFO when the qlen exceeds max_qlen.
 198          *
 199          * For more details see CoNext'11 "TCP Fast Open" paper.
 200          */
 201         req->rsk_timer.expires = jiffies + 60*HZ;
 202         if (fastopenq->rskq_rst_head == NULL)
 203                 fastopenq->rskq_rst_head = req;
 204         else
 205                 fastopenq->rskq_rst_tail->dl_next = req;
 206
 207         req->dl_next = NULL;
 208         fastopenq->rskq_rst_tail = req;
 209         fastopenq->qlen++;
 210 out:
 211         spin_unlock_bh(&fastopenq->lock);
 212 }