net/ipv4/tcp_probe.c

   1 /*
   2  * tcpprobe - Observe the TCP flow with kprobes.
   3  *
   4  * The idea for this came from Werner Almesberger's umlsim
   5  * Copyright (C) 2004, Stephen Hemminger <shemminger@osdl.org>
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License as published by
   9  * the Free Software Foundation; either version 2 of the License.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 #include <linux/kernel.h>
  22 #include <linux/kprobes.h>
  23 #include <linux/socket.h>
  24 #include <linux/tcp.h>
  25 #include <linux/proc_fs.h>
  26 #include <linux/module.h>
  27 #include <linux/ktime.h>
  28 #include <linux/time.h>
  29
  30 #include <net/tcp.h>
  31
  32 MODULE_AUTHOR("Stephen Hemminger <shemminger@linux-foundation.org>");
  33 MODULE_DESCRIPTION("TCP cwnd snooper");
  34 MODULE_LICENSE("GPL");
  35 MODULE_VERSION("1.1");
  36
  37 static int port __read_mostly = 0;
  38 MODULE_PARM_DESC(port, "Port to match (0=all)");
  39 module_param(port, int, 0);
  40
  41 static int bufsize __read_mostly = 4096;
  42 MODULE_PARM_DESC(bufsize, "Log buffer size in packets (4096)");
  43 module_param(bufsize, int, 0);
  44
  45 static int full __read_mostly;
  46 MODULE_PARM_DESC(full, "Full log (1=every ack packet received,  0=only cwnd changes)");
  47 module_param(full, int, 0);
  48
  49 static const char procname[] = "tcpprobe";
  50
  51 struct tcp_log {
  52         ktime_t tstamp;
  53         __be32  saddr, daddr;
  54         __be16  sport, dport;
  55         u16     length;
  56         u32     snd_nxt;
  57         u32     snd_una;
  58         u32     snd_wnd;
  59         u32     snd_cwnd;
  60         u32     ssthresh;
  61         u32     srtt;
  62 };
  63
  64 static struct {
  65         spinlock_t      lock;
  66         wait_queue_head_t wait;
  67         ktime_t         start;
  68         u32             lastcwnd;
  69
  70         unsigned long   head, tail;
  71         struct tcp_log  *log;
  72 } tcp_probe;
  73
  74
  75 static inline int tcp_probe_used(void)
  76 {
  77         return (tcp_probe.head - tcp_probe.tail) % bufsize;
  78 }
  79
  80 static inline int tcp_probe_avail(void)
  81 {
  82         return bufsize - tcp_probe_used();
  83 }
  84
  85 /*
  86  * Hook inserted to be called before each receive packet.
  87  * Note: arguments must match tcp_rcv_established()!
  88  */
  89 static int jtcp_rcv_established(struct sock *sk, struct sk_buff *skb,
  90                                struct tcphdr *th, unsigned len)
  91 {
  92         const struct tcp_sock *tp = tcp_sk(sk);
  93         const struct inet_sock *inet = inet_sk(sk);
  94
  95         /* Only update if port matches */
  96         if ((port == 0 || ntohs(inet->dport) == port || ntohs(inet->sport) == port)
  97             && (full || tp->snd_cwnd != tcp_probe.lastcwnd)) {
  98
  99                 spin_lock(&tcp_probe.lock);
 100                 /* If log fills, just silently drop */
 101                 if (tcp_probe_avail() > 1) {
 102                         struct tcp_log *p = tcp_probe.log + tcp_probe.head;
 103
 104                         p->tstamp = ktime_get();
 105                         p->saddr = inet->saddr;
 106                         p->sport = inet->sport;
 107                         p->daddr = inet->daddr;
 108                         p->dport = inet->dport;
 109                         p->length = skb->len;
 110                         p->snd_nxt = tp->snd_nxt;
 111                         p->snd_una = tp->snd_una;
 112                         p->snd_cwnd = tp->snd_cwnd;
 113                         p->snd_wnd = tp->snd_wnd;
 114                         p->srtt = tp->srtt >> 3;
 115
 116                         tcp_probe.head = (tcp_probe.head + 1) % bufsize;
 117                 }
 118                 tcp_probe.lastcwnd = tp->snd_cwnd;
 119                 spin_unlock(&tcp_probe.lock);
 120
 121                 wake_up(&tcp_probe.wait);
 122         }
 123
 124         jprobe_return();
 125         return 0;
 126 }
 127
 128 static struct jprobe tcp_jprobe = {
 129         .kp = {
 130                 .symbol_name    = "tcp_rcv_established",
 131         },
 132         .entry  = JPROBE_ENTRY(jtcp_rcv_established),
 133 };
 134
 135 static int tcpprobe_open(struct inode * inode, struct file * file)
 136 {
 137         /* Reset (empty) log */
 138         spin_lock_bh(&tcp_probe.lock);
 139         tcp_probe.head = tcp_probe.tail = 0;
 140         tcp_probe.start = ktime_get();
 141         spin_unlock_bh(&tcp_probe.lock);
 142
 143         return 0;
 144 }
 145
 146 static int tcpprobe_sprint(char *tbuf, int n)
 147 {
 148         const struct tcp_log *p
 149                 = tcp_probe.log + tcp_probe.tail % bufsize;
 150         struct timespec tv
 151                 = ktime_to_timespec(ktime_sub(p->tstamp, tcp_probe.start));
 152
 153         return snprintf(tbuf, n,
 154                         "%lu.%09lu %d.%d.%d.%d:%u %d.%d.%d.%d:%u"
 155                         " %d %#x %#x %u %u %u %u\n",
 156                         (unsigned long) tv.tv_sec,
 157                         (unsigned long) tv.tv_nsec,
 158                         NIPQUAD(p->saddr), ntohs(p->sport),
 159                         NIPQUAD(p->daddr), ntohs(p->dport),
 160                         p->length, p->snd_nxt, p->snd_una,
 161                         p->snd_cwnd, p->ssthresh, p->snd_wnd, p->srtt);
 162 }
 163
 164 static ssize_t tcpprobe_read(struct file *file, char __user *buf,
 165                              size_t len, loff_t *ppos)
 166 {
 167         int error = 0, cnt = 0;
 168
 169         if (!buf || len < 0)
 170                 return -EINVAL;
 171
 172         while (cnt < len) {
 173                 char tbuf[128];
 174                 int width;
 175
 176                 /* Wait for data in buffer */
 177                 error = wait_event_interruptible(tcp_probe.wait,
 178                                                  tcp_probe_used() > 0);
 179                 if (error)
 180                         break;
 181
 182                 spin_lock_bh(&tcp_probe.lock);
 183                 if (tcp_probe.head == tcp_probe.tail) {
 184                         /* multiple readers race? */
 185                         spin_unlock_bh(&tcp_probe.lock);
 186                         continue;
 187                 }
 188
 189                 width = tcpprobe_sprint(tbuf, sizeof(tbuf));
 190
 191                 if (width < len)
 192                         tcp_probe.tail = (tcp_probe.tail + 1) % bufsize;
 193
 194                 spin_unlock_bh(&tcp_probe.lock);
 195
 196                 /* if record greater than space available
 197                    return partial buffer (so far) */
 198                 if (width >= len)
 199                         break;
 200
 201                 error = copy_to_user(buf + cnt, tbuf, width);
 202                 if (error)
 203                         break;
 204                 cnt += width;
 205         }
 206
 207         return cnt == 0 ? error : cnt;
 208 }
 209
 210 static const struct file_operations tcpprobe_fops = {
 211         .owner   = THIS_MODULE,
 212         .open    = tcpprobe_open,
 213         .read    = tcpprobe_read,
 214 };
 215
 216 static __init int tcpprobe_init(void)
 217 {
 218         int ret = -ENOMEM;
 219
 220         init_waitqueue_head(&tcp_probe.wait);
 221         spin_lock_init(&tcp_probe.lock);
 222
 223         if (bufsize < 0)
 224                 return -EINVAL;
 225
 226         tcp_probe.log = kcalloc(sizeof(struct tcp_log), bufsize, GFP_KERNEL);
 227         if (!tcp_probe.log)
 228                 goto err0;
 229
 230         if (!proc_net_fops_create(procname, S_IRUSR, &tcpprobe_fops))
 231                 goto err0;
 232
 233         ret = register_jprobe(&tcp_jprobe);
 234         if (ret)
 235                 goto err1;
 236
 237         pr_info("TCP probe registered (port=%d)\n", port);
 238         return 0;
 239  err1:
 240         proc_net_remove(procname);
 241  err0:
 242         kfree(tcp_probe.log);
 243         return ret;
 244 }
 245 module_init(tcpprobe_init);
 246
 247 static __exit void tcpprobe_exit(void)
 248 {
 249         proc_net_remove(procname);
 250         unregister_jprobe(&tcp_jprobe);
 251         kfree(tcp_probe.log);
 252 }
 253 module_exit(tcpprobe_exit);