Quelle  send.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
 */

#include "queueing.h"
#include "timers.h"
#include "device.h"
#include "peer.h"
#include "socket.h"
#include "messages.h"
#include "cookie.h"

#include <linux/uio.h>
#include <linux/inetdevice.h>
#include <linux/socket.h>
#include <net/ip_tunnels.h>
#include <net/udp.h>
#include <net/sock.h>

static void wg_packet_send_handshake_initiation(struct wg_peer *peer)
{
 struct message_handshake_initiation packet;

 if (!wg_birthdate_has_expired(atomic64_read(&peer->last_sent_handshake),
          REKEY_TIMEOUT))
  return; /* This function is rate limited. */

 atomic64_set(&peer->last_sent_handshake, ktime_get_coarse_boottime_ns());
 net_dbg_ratelimited("%s: Sending handshake initiation to peer %llu (%pISpfsc)\n",
       peer->device->dev->name, peer->internal_id,
       &peer->endpoint.addr);

 if (wg_noise_handshake_create_initiation(&packet, &peer->handshake)) {
  wg_cookie_add_mac_to_packet(&packet, sizeof(packet), peer);
  wg_timers_any_authenticated_packet_traversal(peer);
  wg_timers_any_authenticated_packet_sent(peer);
  atomic64_set(&peer->last_sent_handshake,
        ktime_get_coarse_boottime_ns());
  wg_socket_send_buffer_to_peer(peer, &packet, sizeof(packet),
           HANDSHAKE_DSCP);
  wg_timers_handshake_initiated(peer);
 }
}

void wg_packet_handshake_send_worker(struct work_struct *work)
{
 struct wg_peer *peer = container_of(work, struct wg_peer,
         transmit_handshake_work);

 wg_packet_send_handshake_initiation(peer);
 wg_peer_put(peer);
}

void wg_packet_send_queued_handshake_initiation(struct wg_peer *peer,
      bool is_retry)
{
 if (!is_retry)
  peer->timer_handshake_attempts = 0;

 rcu_read_lock_bh();
 /* We check last_sent_handshake here in addition to the actual function
 * we're queueing up, so that we don't queue things if not strictly
 * necessary:
 */
 if (!wg_birthdate_has_expired(atomic64_read(&peer->last_sent_handshake),
          REKEY_TIMEOUT) ||
   unlikely(READ_ONCE(peer->is_dead)))
  goto out;

 wg_peer_get(peer);
 /* Queues up calling packet_send_queued_handshakes(peer), where we do a
 * peer_put(peer) after:
 */
 if (!queue_work(peer->device->handshake_send_wq,
   &peer->transmit_handshake_work))
  /* If the work was already queued, we want to drop the
 * extra reference:
 */
  wg_peer_put(peer);
out:
 rcu_read_unlock_bh();
}

void wg_packet_send_handshake_response(struct wg_peer *peer)
{
 struct message_handshake_response packet;

 atomic64_set(&peer->last_sent_handshake, ktime_get_coarse_boottime_ns());
 net_dbg_ratelimited("%s: Sending handshake response to peer %llu (%pISpfsc)\n",
       peer->device->dev->name, peer->internal_id,
       &peer->endpoint.addr);

 if (wg_noise_handshake_create_response(&packet, &peer->handshake)) {
  wg_cookie_add_mac_to_packet(&packet, sizeof(packet), peer);
  if (wg_noise_handshake_begin_session(&peer->handshake,
           &peer->keypairs)) {
   wg_timers_session_derived(peer);
   wg_timers_any_authenticated_packet_traversal(peer);
   wg_timers_any_authenticated_packet_sent(peer);
   atomic64_set(&peer->last_sent_handshake,
         ktime_get_coarse_boottime_ns());
   wg_socket_send_buffer_to_peer(peer, &packet,
            sizeof(packet),
            HANDSHAKE_DSCP);
  }
 }
}

void wg_packet_send_handshake_cookie(struct wg_device *wg,
         struct sk_buff *initiating_skb,
         __le32 sender_index)
{
 struct message_handshake_cookie packet;

 net_dbg_skb_ratelimited("%s: Sending cookie response for denied handshake message for %pISpfsc\n",
    wg->dev->name, initiating_skb);
 wg_cookie_message_create(&packet, initiating_skb, sender_index,
     &wg->cookie_checker);
 wg_socket_send_buffer_as_reply_to_skb(wg, initiating_skb, &packet,
           sizeof(packet));
}

static void keep_key_fresh(struct wg_peer *peer)
{
 struct noise_keypair *keypair;
 bool send;

 rcu_read_lock_bh();
 keypair = rcu_dereference_bh(peer->keypairs.current_keypair);
 send = keypair && READ_ONCE(keypair->sending.is_valid) &&
        (atomic64_read(&keypair->sending_counter) > REKEY_AFTER_MESSAGES ||
  (keypair->i_am_the_initiator &&
   wg_birthdate_has_expired(keypair->sending.birthdate, REKEY_AFTER_TIME)));
 rcu_read_unlock_bh();

 if (unlikely(send))
  wg_packet_send_queued_handshake_initiation(peer, false);
}

static unsigned int calculate_skb_padding(struct sk_buff *skb)
{
 unsigned int padded_size, last_unit = skb->len;

 if (unlikely(!PACKET_CB(skb)->mtu))
  return ALIGN(last_unit, MESSAGE_PADDING_MULTIPLE) - last_unit;

 /* We do this modulo business with the MTU, just in case the networking
 * layer gives us a packet that's bigger than the MTU. In that case, we
 * wouldn't want the final subtraction to overflow in the case of the
 * padded_size being clamped. Fortunately, that's very rarely the case,
 * so we optimize for that not happening.
 */
 if (unlikely(last_unit > PACKET_CB(skb)->mtu))
  last_unit %= PACKET_CB(skb)->mtu;

 padded_size = min(PACKET_CB(skb)->mtu,
     ALIGN(last_unit, MESSAGE_PADDING_MULTIPLE));
 return padded_size - last_unit;
}

static bool encrypt_packet(struct sk_buff *skb, struct noise_keypair *keypair)
{
 unsigned int padding_len, plaintext_len, trailer_len;
 struct scatterlist sg[MAX_SKB_FRAGS + 8];
 struct message_data *header;
 struct sk_buff *trailer;
 int num_frags;

 /* Force hash calculation before encryption so that flow analysis is
 * consistent over the inner packet.
 */
 skb_get_hash(skb);

 /* Calculate lengths. */
 padding_len = calculate_skb_padding(skb);
 trailer_len = padding_len + noise_encrypted_len(0);
 plaintext_len = skb->len + padding_len;

 /* Expand data section to have room for padding and auth tag. */
 num_frags = skb_cow_data(skb, trailer_len, &trailer);
 if (unlikely(num_frags < 0 || num_frags > ARRAY_SIZE(sg)))
  return false;

 /* Set the padding to zeros, and make sure it and the auth tag are part
 * of the skb.
 */
 memset(skb_tail_pointer(trailer), 0, padding_len);

 /* Expand head section to have room for our header and the network
 * stack's headers.
 */
 if (unlikely(skb_cow_head(skb, DATA_PACKET_HEAD_ROOM) < 0))
  return false;

 /* Finalize checksum calculation for the inner packet, if required. */
 if (unlikely(skb->ip_summed == CHECKSUM_PARTIAL &&
       skb_checksum_help(skb)))
  return false;

 /* Only after checksumming can we safely add on the padding at the end
 * and the header.
 */
 skb_set_inner_network_header(skb, 0);
 header = (struct message_data *)skb_push(skb, sizeof(*header));
 header->header.type = cpu_to_le32(MESSAGE_DATA);
 header->key_idx = keypair->remote_index;
 header->counter = cpu_to_le64(PACKET_CB(skb)->nonce);
 pskb_put(skb, trailer, trailer_len);

 /* Now we can encrypt the scattergather segments */
 sg_init_table(sg, num_frags);
 if (skb_to_sgvec(skb, sg, sizeof(struct message_data),
    noise_encrypted_len(plaintext_len)) <= 0)
  return false;
 return chacha20poly1305_encrypt_sg_inplace(sg, plaintext_len, NULL, 0,
         PACKET_CB(skb)->nonce,
         keypair->sending.key);
}

void wg_packet_send_keepalive(struct wg_peer *peer)
{
 struct sk_buff *skb;

 if (skb_queue_empty_lockless(&peer->staged_packet_queue)) {
  skb = alloc_skb(DATA_PACKET_HEAD_ROOM + MESSAGE_MINIMUM_LENGTH,
    GFP_ATOMIC);
  if (unlikely(!skb))
   return;
  skb_reserve(skb, DATA_PACKET_HEAD_ROOM);
  skb->dev = peer->device->dev;
  PACKET_CB(skb)->mtu = skb->dev->mtu;
  skb_queue_tail(&peer->staged_packet_queue, skb);
  net_dbg_ratelimited("%s: Sending keepalive packet to peer %llu (%pISpfsc)\n",
        peer->device->dev->name, peer->internal_id,
        &peer->endpoint.addr);
 }

 wg_packet_send_staged_packets(peer);
}

static void wg_packet_create_data_done(struct wg_peer *peer, struct sk_buff *first)
{
 struct sk_buff *skb, *next;
 bool is_keepalive, data_sent = false;

 wg_timers_any_authenticated_packet_traversal(peer);
 wg_timers_any_authenticated_packet_sent(peer);
 skb_list_walk_safe(first, skb, next) {
  is_keepalive = skb->len == message_data_len(0);
  if (likely(!wg_socket_send_skb_to_peer(peer, skb,
    PACKET_CB(skb)->ds) && !is_keepalive))
   data_sent = true;
 }

 if (likely(data_sent))
  wg_timers_data_sent(peer);

 keep_key_fresh(peer);
}

void wg_packet_tx_worker(struct work_struct *work)
{
 struct wg_peer *peer = container_of(work, struct wg_peer, transmit_packet_work);
 struct noise_keypair *keypair;
 enum packet_state state;
 struct sk_buff *first;

 while ((first = wg_prev_queue_peek(&peer->tx_queue)) != NULL &&
        (state = atomic_read_acquire(&PACKET_CB(first)->state)) !=
         PACKET_STATE_UNCRYPTED) {
  wg_prev_queue_drop_peeked(&peer->tx_queue);
  keypair = PACKET_CB(first)->keypair;

  if (likely(state == PACKET_STATE_CRYPTED))
   wg_packet_create_data_done(peer, first);
  else
   kfree_skb_list(first);

  wg_noise_keypair_put(keypair, false);
  wg_peer_put(peer);
  if (need_resched())
   cond_resched();
 }
}

void wg_packet_encrypt_worker(struct work_struct *work)
{
 struct crypt_queue *queue = container_of(work, struct multicore_worker,
       work)->ptr;
 struct sk_buff *first, *skb, *next;

 while ((first = ptr_ring_consume_bh(&queue->ring)) != NULL) {
  enum packet_state state = PACKET_STATE_CRYPTED;

  skb_list_walk_safe(first, skb, next) {
   if (likely(encrypt_packet(skb,
     PACKET_CB(first)->keypair))) {
    wg_reset_packet(skb, true);
   } else {
    state = PACKET_STATE_DEAD;
    break;
   }
  }
  wg_queue_enqueue_per_peer_tx(first, state);
  if (need_resched())
   cond_resched();
 }
}

static void wg_packet_create_data(struct wg_peer *peer, struct sk_buff *first)
{
 struct wg_device *wg = peer->device;
 int ret = -EINVAL;

 rcu_read_lock_bh();
 if (unlikely(READ_ONCE(peer->is_dead)))
  goto err;

 ret = wg_queue_enqueue_per_device_and_peer(&wg->encrypt_queue, &peer->tx_queue, first,
         wg->packet_crypt_wq);
 if (unlikely(ret == -EPIPE))
  wg_queue_enqueue_per_peer_tx(first, PACKET_STATE_DEAD);
err:
 rcu_read_unlock_bh();
 if (likely(!ret || ret == -EPIPE))
  return;
 wg_noise_keypair_put(PACKET_CB(first)->keypair, false);
 wg_peer_put(peer);
 kfree_skb_list(first);
}

void wg_packet_purge_staged_packets(struct wg_peer *peer)
{
 spin_lock_bh(&peer->staged_packet_queue.lock);
 DEV_STATS_ADD(peer->device->dev, tx_dropped,
        peer->staged_packet_queue.qlen);
 __skb_queue_purge(&peer->staged_packet_queue);
 spin_unlock_bh(&peer->staged_packet_queue.lock);
}

void wg_packet_send_staged_packets(struct wg_peer *peer)
{
 struct noise_keypair *keypair;
 struct sk_buff_head packets;
 struct sk_buff *skb;

 /* Steal the current queue into our local one. */
 __skb_queue_head_init(&packets);
 spin_lock_bh(&peer->staged_packet_queue.lock);
 skb_queue_splice_init(&peer->staged_packet_queue, &packets);
 spin_unlock_bh(&peer->staged_packet_queue.lock);
 if (unlikely(skb_queue_empty(&packets)))
  return;

 /* First we make sure we have a valid reference to a valid key. */
 rcu_read_lock_bh();
 keypair = wg_noise_keypair_get(
  rcu_dereference_bh(peer->keypairs.current_keypair));
 rcu_read_unlock_bh();
 if (unlikely(!keypair))
  goto out_nokey;
 if (unlikely(!READ_ONCE(keypair->sending.is_valid)))
  goto out_nokey;
 if (unlikely(wg_birthdate_has_expired(keypair->sending.birthdate,
           REJECT_AFTER_TIME)))
  goto out_invalid;

 /* After we know we have a somewhat valid key, we now try to assign
 * nonces to all of the packets in the queue. If we can't assign nonces
 * for all of them, we just consider it a failure and wait for the next
 * handshake.
 */
 skb_queue_walk(&packets, skb) {
  /* 0 for no outer TOS: no leak. TODO: at some later point, we
 * might consider using flowi->tos as outer instead.
 */
  PACKET_CB(skb)->ds = ip_tunnel_ecn_encap(0, ip_hdr(skb), skb);
  PACKET_CB(skb)->nonce =
    atomic64_inc_return(&keypair->sending_counter) - 1;
  if (unlikely(PACKET_CB(skb)->nonce >= REJECT_AFTER_MESSAGES))
   goto out_invalid;
 }

 packets.prev->next = NULL;
 wg_peer_get(keypair->entry.peer);
 PACKET_CB(packets.next)->keypair = keypair;
 wg_packet_create_data(peer, packets.next);
 return;

out_invalid:
 WRITE_ONCE(keypair->sending.is_valid, false);
out_nokey:
 wg_noise_keypair_put(keypair, false);

 /* We orphan the packets if we're waiting on a handshake, so that they
 * don't block a socket's pool.
 */
 skb_queue_walk(&packets, skb)
  skb_orphan(skb);
 /* Then we put them back on the top of the queue. We're not too
 * concerned about accidentally getting things a little out of order if
 * packets are being added really fast, because this queue is for before
 * packets can even be sent and it's small anyway.
 */
 spin_lock_bh(&peer->staged_packet_queue.lock);
 skb_queue_splice(&packets, &peer->staged_packet_queue);
 spin_unlock_bh(&peer->staged_packet_queue.lock);

 /* If we're exiting because there's something wrong with the key, it
 * means we should initiate a new handshake.
 */
 wg_packet_send_queued_handshake_initiation(peer, false);
}