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         spin_lock_init(&queue->rskq_lock);
  37
  38         spin_lock_init(&queue->fastopenq.lock);
  39         queue->fastopenq.rskq_rst_head = NULL;
  40         queue->fastopenq.rskq_rst_tail = NULL;
  41         queue->fastopenq.qlen = 0;
  42
  43         queue->rskq_accept_head = NULL;
  44 }
  45
  46 /*
  47  * This function is called to set a Fast Open socket's "fastopen_rsk" field
  48  * to NULL when a TFO socket no longer needs to access the request_sock.
  49  * This happens only after 3WHS has been either completed or aborted (e.g.,
  50  * RST is received).
  51  *
  52  * Before TFO, a child socket is created only after 3WHS is completed,
  53  * hence it never needs to access the request_sock. things get a lot more
  54  * complex with TFO. A child socket, accepted or not, has to access its
  55  * request_sock for 3WHS processing, e.g., to retransmit SYN-ACK pkts,
  56  * until 3WHS is either completed or aborted. Afterwards the req will stay
  57  * until either the child socket is accepted, or in the rare case when the
  58  * listener is closed before the child is accepted.
  59  *
  60  * In short, a request socket is only freed after BOTH 3WHS has completed
  61  * (or aborted) and the child socket has been accepted (or listener closed).
  62  * When a child socket is accepted, its corresponding req->sk is set to
  63  * NULL since it's no longer needed. More importantly, "req->sk == NULL"
  64  * will be used by the code below to determine if a child socket has been
  65  * accepted or not, and the check is protected by the fastopenq->lock
  66  * described below.
  67  *
  68  * Note that fastopen_rsk is only accessed from the child socket's context
  69  * with its socket lock held. But a request_sock (req) can be accessed by
  70  * both its child socket through fastopen_rsk, and a listener socket through
  71  * icsk_accept_queue.rskq_accept_head. To protect the access a simple spin
  72  * lock per listener "icsk->icsk_accept_queue.fastopenq->lock" is created.
  73  * only in the rare case when both the listener and the child locks are held,
  74  * e.g., in inet_csk_listen_stop() do we not need to acquire the lock.
  75  * The lock also protects other fields such as fastopenq->qlen, which is
  76  * decremented by this function when fastopen_rsk is no longer needed.
  77  *
  78  * Note that another solution was to simply use the existing socket lock
  79  * from the listener. But first socket lock is difficult to use. It is not
  80  * a simple spin lock - one must consider sock_owned_by_user() and arrange
  81  * to use sk_add_backlog() stuff. But what really makes it infeasible is the
  82  * locking hierarchy violation. E.g., inet_csk_listen_stop() may try to
  83  * acquire a child's lock while holding listener's socket lock. A corner
  84  * case might also exist in tcp_v4_hnd_req() that will trigger this locking
  85  * order.
  86  *
  87  * This function also sets "treq->tfo_listener" to false.
  88  * treq->tfo_listener is used by the listener so it is protected by the
  89  * fastopenq->lock in this function.
  90  */
  91 void reqsk_fastopen_remove(struct sock *sk, struct request_sock *req,
  92                            bool reset)
  93 {
  94         struct sock *lsk = req->rsk_listener;
  95         struct fastopen_queue *fastopenq;
  96
  97         fastopenq = &inet_csk(lsk)->icsk_accept_queue.fastopenq;
  98
  99         RCU_INIT_POINTER(tcp_sk(sk)->fastopen_rsk, NULL);
 100         spin_lock_bh(&fastopenq->lock);
 101         fastopenq->qlen--;
 102         tcp_rsk(req)->tfo_listener = false;
 103         if (req->sk)    /* the child socket hasn't been accepted yet */
 104                 goto out;
 105
 106         if (!reset || lsk->sk_state != TCP_LISTEN) {
 107                 /* If the listener has been closed don't bother with the
 108                  * special RST handling below.
 109                  */
 110                 spin_unlock_bh(&fastopenq->lock);
 111                 reqsk_put(req);
 112                 return;
 113         }
 114         /* Wait for 60secs before removing a req that has triggered RST.
 115          * This is a simple defense against TFO spoofing attack - by
 116          * counting the req against fastopen.max_qlen, and disabling
 117          * TFO when the qlen exceeds max_qlen.
 118          *
 119          * For more details see CoNext'11 "TCP Fast Open" paper.
 120          */
 121         req->rsk_timer.expires = jiffies + 60*HZ;
 122         if (fastopenq->rskq_rst_head == NULL)
 123                 fastopenq->rskq_rst_head = req;
 124         else
 125                 fastopenq->rskq_rst_tail->dl_next = req;
 126
 127         req->dl_next = NULL;
 128         fastopenq->rskq_rst_tail = req;
 129         fastopenq->qlen++;
 130 out:
 131         spin_unlock_bh(&fastopenq->lock);
 132 }