sys/netinet/siftr.c

   1 /*-
   2  * Copyright (c) 2007-2009
   3  *      Swinburne University of Technology, Melbourne, Australia.
   4  * Copyright (c) 2009-2010, The FreeBSD Foundation
   5  * All rights reserved.
   6  *
   7  * Portions of this software were developed at the Centre for Advanced
   8  * Internet Architectures, Swinburne University of Technology, Melbourne,
   9  * Australia by Lawrence Stewart under sponsorship from the FreeBSD Foundation.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  */
  32
  33 /******************************************************
  34  * Statistical Information For TCP Research (SIFTR)
  35  *
  36  * A FreeBSD kernel module that adds very basic intrumentation to the
  37  * TCP stack, allowing internal stats to be recorded to a log file
  38  * for experimental, debugging and performance analysis purposes.
  39  *
  40  * SIFTR was first released in 2007 by James Healy and Lawrence Stewart whilst
  41  * working on the NewTCP research project at Swinburne University of
  42  * Technology's Centre for Advanced Internet Architectures, Melbourne,
  43  * Australia, which was made possible in part by a grant from the Cisco
  44  * University Research Program Fund at Community Foundation Silicon Valley.
  45  * More details are available at:
  46  *   http://caia.swin.edu.au/urp/newtcp/
  47  *
  48  * Work on SIFTR v1.2.x was sponsored by the FreeBSD Foundation as part of
  49  * the "Enhancing the FreeBSD TCP Implementation" project 2008-2009.
  50  * More details are available at:
  51  *   http://www.freebsdfoundation.org/
  52  *   http://caia.swin.edu.au/freebsd/etcp09/
  53  *
  54  * Lawrence Stewart is the current maintainer, and all contact regarding
  55  * SIFTR should be directed to him via email: lastewart@swin.edu.au
  56  *
  57  * Initial release date: June 2007
  58  * Most recent update: September 2010
  59  ******************************************************/
  60
  61 #include <sys/cdefs.h>
  62 __FBSDID("$FreeBSD$");
  63
  64 #include <sys/param.h>
  65 #include <sys/alq.h>
  66 #include <sys/errno.h>
  67 #include <sys/hash.h>
  68 #include <sys/kernel.h>
  69 #include <sys/kthread.h>
  70 #include <sys/lock.h>
  71 #include <sys/mbuf.h>
  72 #include <sys/module.h>
  73 #include <sys/mutex.h>
  74 #include <sys/pcpu.h>
  75 #include <sys/proc.h>
  76 #include <sys/sbuf.h>
  77 #include <sys/smp.h>
  78 #include <sys/socket.h>
  79 #include <sys/socketvar.h>
  80 #include <sys/sysctl.h>
  81 #include <sys/unistd.h>
  82
  83 #include <net/if.h>
  84 #include <net/pfil.h>
  85
  86 #include <netinet/in.h>
  87 #include <netinet/in_pcb.h>
  88 #include <netinet/in_systm.h>
  89 #include <netinet/in_var.h>
  90 #include <netinet/ip.h>
  91 #include <netinet/tcp_var.h>
  92
  93 #ifdef SIFTR_IPV6
  94 #include <netinet/ip6.h>
  95 #include <netinet6/in6_pcb.h>
  96 #endif /* SIFTR_IPV6 */
  97
  98 #include <machine/in_cksum.h>
  99
 100 /*
 101  * Three digit version number refers to X.Y.Z where:
 102  * X is the major version number
 103  * Y is bumped to mark backwards incompatible changes
 104  * Z is bumped to mark backwards compatible changes
 105  */
 106 #define V_MAJOR         1
 107 #define V_BACKBREAK     2
 108 #define V_BACKCOMPAT    4
 109 #define MODVERSION      __CONCAT(V_MAJOR, __CONCAT(V_BACKBREAK, V_BACKCOMPAT))
 110 #define MODVERSION_STR  __XSTRING(V_MAJOR) "." __XSTRING(V_BACKBREAK) "." \
 111     __XSTRING(V_BACKCOMPAT)
 112
 113 #define HOOK 0
 114 #define UNHOOK 1
 115 #define SIFTR_EXPECTED_MAX_TCP_FLOWS 65536
 116 #define SYS_NAME "FreeBSD"
 117 #define PACKET_TAG_SIFTR 100
 118 #define PACKET_COOKIE_SIFTR 21749576
 119 #define SIFTR_LOG_FILE_MODE 0644
 120 #define SIFTR_DISABLE 0
 121 #define SIFTR_ENABLE 1
 122
 123 /*
 124  * Hard upper limit on the length of log messages. Bump this up if you add new
 125  * data fields such that the line length could exceed the below value.
 126  */
 127 #define MAX_LOG_MSG_LEN 200
 128 /* XXX: Make this a sysctl tunable. */
 129 #define SIFTR_ALQ_BUFLEN (1000*MAX_LOG_MSG_LEN)
 130
 131 /*
 132  * 1 byte for IP version
 133  * IPv4: src/dst IP (4+4) + src/dst port (2+2) = 12 bytes
 134  * IPv6: src/dst IP (16+16) + src/dst port (2+2) = 36 bytes
 135  */
 136 #ifdef SIFTR_IPV6
 137 #define FLOW_KEY_LEN 37
 138 #else
 139 #define FLOW_KEY_LEN 13
 140 #endif
 141
 142 #ifdef SIFTR_IPV6
 143 #define SIFTR_IPMODE 6
 144 #else
 145 #define SIFTR_IPMODE 4
 146 #endif
 147
 148 /* useful macros */
 149 #define CAST_PTR_INT(X) (*((int*)(X)))
 150
 151 #define UPPER_SHORT(X)  (((X) & 0xFFFF0000) >> 16)
 152 #define LOWER_SHORT(X)  ((X) & 0x0000FFFF)
 153
 154 #define FIRST_OCTET(X)  (((X) & 0xFF000000) >> 24)
 155 #define SECOND_OCTET(X) (((X) & 0x00FF0000) >> 16)
 156 #define THIRD_OCTET(X)  (((X) & 0x0000FF00) >> 8)
 157 #define FOURTH_OCTET(X) ((X) & 0x000000FF)
 158
 159 static MALLOC_DEFINE(M_SIFTR, "siftr", "dynamic memory used by SIFTR");
 160 static MALLOC_DEFINE(M_SIFTR_PKTNODE, "siftr_pktnode",
 161     "SIFTR pkt_node struct");
 162 static MALLOC_DEFINE(M_SIFTR_HASHNODE, "siftr_hashnode",
 163     "SIFTR flow_hash_node struct");
 164
 165 /* Used as links in the pkt manager queue. */
 166 struct pkt_node {
 167         /* Timestamp of pkt as noted in the pfil hook. */
 168         struct timeval          tval;
 169         /* Direction pkt is travelling; either PFIL_IN or PFIL_OUT. */
 170         uint8_t                 direction;
 171         /* IP version pkt_node relates to; either INP_IPV4 or INP_IPV6. */
 172         uint8_t                 ipver;
 173         /* Hash of the pkt which triggered the log message. */
 174         uint32_t                hash;
 175         /* Local/foreign IP address. */
 176 #ifdef SIFTR_IPV6
 177         uint32_t                ip_laddr[4];
 178         uint32_t                ip_faddr[4];
 179 #else
 180         uint8_t                 ip_laddr[4];
 181         uint8_t                 ip_faddr[4];
 182 #endif
 183         /* Local TCP port. */
 184         uint16_t                tcp_localport;
 185         /* Foreign TCP port. */
 186         uint16_t                tcp_foreignport;
 187         /* Congestion Window (bytes). */
 188         u_long                  snd_cwnd;
 189         /* Sending Window (bytes). */
 190         u_long                  snd_wnd;
 191         /* Receive Window (bytes). */
 192         u_long                  rcv_wnd;
 193         /* Bandwidth Controlled Window (bytes). */
 194         u_long                  snd_bwnd;
 195         /* Slow Start Threshold (bytes). */
 196         u_long                  snd_ssthresh;
 197         /* Current state of the TCP FSM. */
 198         int                     conn_state;
 199         /* Max Segment Size (bytes). */
 200         u_int                   max_seg_size;
 201         /*
 202          * Smoothed RTT stored as found in the TCP control block
 203          * in units of (TCP_RTT_SCALE*hz).
 204          */
 205         int                     smoothed_rtt;
 206         /* Is SACK enabled? */
 207         u_char                  sack_enabled;
 208         /* Window scaling for snd window. */
 209         u_char                  snd_scale;
 210         /* Window scaling for recv window. */
 211         u_char                  rcv_scale;
 212         /* TCP control block flags. */
 213         u_int                   flags;
 214         /* Retransmit timeout length. */
 215         int                     rxt_length;
 216         /* Size of the TCP send buffer in bytes. */
 217         u_int                   snd_buf_hiwater;
 218         /* Current num bytes in the send socket buffer. */
 219         u_int                   snd_buf_cc;
 220         /* Size of the TCP receive buffer in bytes. */
 221         u_int                   rcv_buf_hiwater;
 222         /* Current num bytes in the receive socket buffer. */
 223         u_int                   rcv_buf_cc;
 224         /* Number of bytes inflight that we are waiting on ACKs for. */
 225         u_int                   sent_inflight_bytes;
 226         /* Number of segments currently in the reassembly queue. */
 227         int                     t_segqlen;
 228         /* Link to next pkt_node in the list. */
 229         STAILQ_ENTRY(pkt_node)  nodes;
 230 };
 231
 232 struct flow_hash_node
 233 {
 234         uint16_t counter;
 235         uint8_t key[FLOW_KEY_LEN];
 236         LIST_ENTRY(flow_hash_node) nodes;
 237 };
 238
 239 struct siftr_stats
 240 {
 241         /* # TCP pkts seen by the SIFTR PFIL hooks, including any skipped. */
 242         uint64_t n_in;
 243         uint64_t n_out;
 244         /* # pkts skipped due to failed malloc calls. */
 245         uint32_t nskip_in_malloc;
 246         uint32_t nskip_out_malloc;
 247         /* # pkts skipped due to failed mtx acquisition. */
 248         uint32_t nskip_in_mtx;
 249         uint32_t nskip_out_mtx;
 250         /* # pkts skipped due to failed inpcb lookups. */
 251         uint32_t nskip_in_inpcb;
 252         uint32_t nskip_out_inpcb;
 253         /* # pkts skipped due to failed tcpcb lookups. */
 254         uint32_t nskip_in_tcpcb;
 255         uint32_t nskip_out_tcpcb;
 256         /* # pkts skipped due to stack reinjection. */
 257         uint32_t nskip_in_dejavu;
 258         uint32_t nskip_out_dejavu;
 259 };
 260
 261 static DPCPU_DEFINE(struct siftr_stats, ss);
 262
 263 static volatile unsigned int siftr_exit_pkt_manager_thread = 0;
 264 static unsigned int siftr_enabled = 0;
 265 static unsigned int siftr_pkts_per_log = 1;
 266 static unsigned int siftr_generate_hashes = 0;
 267 /* static unsigned int siftr_binary_log = 0; */
 268 static char siftr_logfile[PATH_MAX] = "/var/log/siftr.log";
 269 static u_long siftr_hashmask;
 270 STAILQ_HEAD(pkthead, pkt_node) pkt_queue = STAILQ_HEAD_INITIALIZER(pkt_queue);
 271 LIST_HEAD(listhead, flow_hash_node) *counter_hash;
 272 static int wait_for_pkt;
 273 static struct alq *siftr_alq = NULL;
 274 static struct mtx siftr_pkt_queue_mtx;
 275 static struct mtx siftr_pkt_mgr_mtx;
 276 static struct thread *siftr_pkt_manager_thr = NULL;
 277 /*
 278  * pfil.h defines PFIL_IN as 1 and PFIL_OUT as 2,
 279  * which we use as an index into this array.
 280  */
 281 static char direction[3] = {'\0', 'i','o'};
 282
 283 /* Required function prototypes. */
 284 static int siftr_sysctl_enabled_handler(SYSCTL_HANDLER_ARGS);
 285 static int siftr_sysctl_logfile_name_handler(SYSCTL_HANDLER_ARGS);
 286
 287
 288 /* Declare the net.inet.siftr sysctl tree and populate it. */
 289
 290 SYSCTL_DECL(_net_inet_siftr);
 291
 292 SYSCTL_NODE(_net_inet, OID_AUTO, siftr, CTLFLAG_RW, NULL,
 293     "siftr related settings");
 294
 295 SYSCTL_PROC(_net_inet_siftr, OID_AUTO, enabled, CTLTYPE_UINT|CTLFLAG_RW,
 296     &siftr_enabled, 0, &siftr_sysctl_enabled_handler, "IU",
 297     "switch siftr module operations on/off");
 298
 299 SYSCTL_PROC(_net_inet_siftr, OID_AUTO, logfile, CTLTYPE_STRING|CTLFLAG_RW,
 300     &siftr_logfile, sizeof(siftr_logfile), &siftr_sysctl_logfile_name_handler,
 301     "A", "file to save siftr log messages to");
 302
 303 SYSCTL_UINT(_net_inet_siftr, OID_AUTO, ppl, CTLFLAG_RW,
 304     &siftr_pkts_per_log, 1,
 305     "number of packets between generating a log message");
 306
 307 SYSCTL_UINT(_net_inet_siftr, OID_AUTO, genhashes, CTLFLAG_RW,
 308     &siftr_generate_hashes, 0,
 309     "enable packet hash generation");
 310
 311 /* XXX: TODO
 312 SYSCTL_UINT(_net_inet_siftr, OID_AUTO, binary, CTLFLAG_RW,
 313     &siftr_binary_log, 0,
 314     "write log files in binary instead of ascii");
 315 */
 316
 317
 318 /* Begin functions. */
 319
 320 static void
 321 siftr_process_pkt(struct pkt_node * pkt_node)
 322 {
 323         struct flow_hash_node *hash_node;
 324         struct listhead *counter_list;
 325         struct siftr_stats *ss;
 326         struct ale *log_buf;
 327         uint8_t key[FLOW_KEY_LEN];
 328         uint8_t found_match, key_offset;
 329
 330         hash_node = NULL;
 331         ss = DPCPU_PTR(ss);
 332         found_match = 0;
 333         key_offset = 1;
 334
 335         /*
 336          * Create the key that will be used to create a hash index
 337          * into our hash table. Our key consists of:
 338          * ipversion, localip, localport, foreignip, foreignport
 339          */
 340         key[0] = pkt_node->ipver;
 341         memcpy(key + key_offset, &pkt_node->ip_laddr,
 342             sizeof(pkt_node->ip_laddr));
 343         key_offset += sizeof(pkt_node->ip_laddr);
 344         memcpy(key + key_offset, &pkt_node->tcp_localport,
 345             sizeof(pkt_node->tcp_localport));
 346         key_offset += sizeof(pkt_node->tcp_localport);
 347         memcpy(key + key_offset, &pkt_node->ip_faddr,
 348             sizeof(pkt_node->ip_faddr));
 349         key_offset += sizeof(pkt_node->ip_faddr);
 350         memcpy(key + key_offset, &pkt_node->tcp_foreignport,
 351             sizeof(pkt_node->tcp_foreignport));
 352
 353         counter_list = counter_hash +
 354             (hash32_buf(key, sizeof(key), 0) & siftr_hashmask);
 355
 356         /*
 357          * If the list is not empty i.e. the hash index has
 358          * been used by another flow previously.
 359          */
 360         if (LIST_FIRST(counter_list) != NULL) {
 361                 /*
 362                  * Loop through the hash nodes in the list.
 363                  * There should normally only be 1 hash node in the list,
 364                  * except if there have been collisions at the hash index
 365                  * computed by hash32_buf().
 366                  */
 367                 LIST_FOREACH(hash_node, counter_list, nodes) {
 368                         /*
 369                          * Check if the key for the pkt we are currently
 370                          * processing is the same as the key stored in the
 371                          * hash node we are currently processing.
 372                          * If they are the same, then we've found the
 373                          * hash node that stores the counter for the flow
 374                          * the pkt belongs to.
 375                          */
 376                         if (memcmp(hash_node->key, key, sizeof(key)) == 0) {
 377                                 found_match = 1;
 378                                 break;
 379                         }
 380                 }
 381         }
 382
 383         /* If this flow hash hasn't been seen before or we have a collision. */
 384         if (hash_node == NULL || !found_match) {
 385                 /* Create a new hash node to store the flow's counter. */
 386                 hash_node = malloc(sizeof(struct flow_hash_node),
 387                     M_SIFTR_HASHNODE, M_WAITOK);
 388
 389                 if (hash_node != NULL) {
 390                         /* Initialise our new hash node list entry. */
 391                         hash_node->counter = 0;
 392                         memcpy(hash_node->key, key, sizeof(key));
 393                         LIST_INSERT_HEAD(counter_list, hash_node, nodes);
 394                 } else {
 395                         /* Malloc failed. */
 396                         if (pkt_node->direction == PFIL_IN)
 397                                 ss->nskip_in_malloc++;
 398                         else
 399                                 ss->nskip_out_malloc++;
 400
 401                         return;
 402                 }
 403         } else if (siftr_pkts_per_log > 1) {
 404                 /*
 405                  * Taking the remainder of the counter divided
 406                  * by the current value of siftr_pkts_per_log
 407                  * and storing that in counter provides a neat
 408                  * way to modulate the frequency of log
 409                  * messages being written to the log file.
 410                  */
 411                 hash_node->counter = (hash_node->counter + 1) %
 412                     siftr_pkts_per_log;
 413
 414                 /*
 415                  * If we have not seen enough packets since the last time
 416                  * we wrote a log message for this connection, return.
 417                  */
 418                 if (hash_node->counter > 0)
 419                         return;
 420         }
 421
 422         log_buf = alq_getn(siftr_alq, MAX_LOG_MSG_LEN, ALQ_WAITOK);
 423
 424         if (log_buf == NULL)
 425                 return; /* Should only happen if the ALQ is shutting down. */
 426
 427 #ifdef SIFTR_IPV6
 428         pkt_node->ip_laddr[3] = ntohl(pkt_node->ip_laddr[3]);
 429         pkt_node->ip_faddr[3] = ntohl(pkt_node->ip_faddr[3]);
 430
 431         if (pkt_node->ipver == INP_IPV6) { /* IPv6 packet */
 432                 pkt_node->ip_laddr[0] = ntohl(pkt_node->ip_laddr[0]);
 433                 pkt_node->ip_laddr[1] = ntohl(pkt_node->ip_laddr[1]);
 434                 pkt_node->ip_laddr[2] = ntohl(pkt_node->ip_laddr[2]);
 435                 pkt_node->ip_faddr[0] = ntohl(pkt_node->ip_faddr[0]);
 436                 pkt_node->ip_faddr[1] = ntohl(pkt_node->ip_faddr[1]);
 437                 pkt_node->ip_faddr[2] = ntohl(pkt_node->ip_faddr[2]);
 438
 439                 /* Construct an IPv6 log message. */
 440                 log_buf->ae_bytesused = snprintf(log_buf->ae_data,
 441                     MAX_LOG_MSG_LEN,
 442                     "%c,0x%08x,%zd.%06ld,%x:%x:%x:%x:%x:%x:%x:%x,%u,%x:%x:%x:"
 443                     "%x:%x:%x:%x:%x,%u,%ld,%ld,%ld,%ld,%ld,%u,%u,%u,%u,%u,%u,"
 444                     "%u,%d,%u,%u,%u,%u,%u,%u\n",
 445                     direction[pkt_node->direction],
 446                     pkt_node->hash,
 447                     pkt_node->tval.tv_sec,
 448                     pkt_node->tval.tv_usec,
 449                     UPPER_SHORT(pkt_node->ip_laddr[0]),
 450                     LOWER_SHORT(pkt_node->ip_laddr[0]),
 451                     UPPER_SHORT(pkt_node->ip_laddr[1]),
 452                     LOWER_SHORT(pkt_node->ip_laddr[1]),
 453                     UPPER_SHORT(pkt_node->ip_laddr[2]),
 454                     LOWER_SHORT(pkt_node->ip_laddr[2]),
 455                     UPPER_SHORT(pkt_node->ip_laddr[3]),
 456                     LOWER_SHORT(pkt_node->ip_laddr[3]),
 457                     ntohs(pkt_node->tcp_localport),
 458                     UPPER_SHORT(pkt_node->ip_faddr[0]),
 459                     LOWER_SHORT(pkt_node->ip_faddr[0]),
 460                     UPPER_SHORT(pkt_node->ip_faddr[1]),
 461                     LOWER_SHORT(pkt_node->ip_faddr[1]),
 462                     UPPER_SHORT(pkt_node->ip_faddr[2]),
 463                     LOWER_SHORT(pkt_node->ip_faddr[2]),
 464                     UPPER_SHORT(pkt_node->ip_faddr[3]),
 465                     LOWER_SHORT(pkt_node->ip_faddr[3]),
 466                     ntohs(pkt_node->tcp_foreignport),
 467                     pkt_node->snd_ssthresh,
 468                     pkt_node->snd_cwnd,
 469                     pkt_node->snd_bwnd,
 470                     pkt_node->snd_wnd,
 471                     pkt_node->rcv_wnd,
 472                     pkt_node->snd_scale,
 473                     pkt_node->rcv_scale,
 474                     pkt_node->conn_state,
 475                     pkt_node->max_seg_size,
 476                     pkt_node->smoothed_rtt,
 477                     pkt_node->sack_enabled,
 478                     pkt_node->flags,
 479                     pkt_node->rxt_length,
 480                     pkt_node->snd_buf_hiwater,
 481                     pkt_node->snd_buf_cc,
 482                     pkt_node->rcv_buf_hiwater,
 483                     pkt_node->rcv_buf_cc,
 484                     pkt_node->sent_inflight_bytes,
 485                     pkt_node->t_segqlen);
 486         } else { /* IPv4 packet */
 487                 pkt_node->ip_laddr[0] = FIRST_OCTET(pkt_node->ip_laddr[3]);
 488                 pkt_node->ip_laddr[1] = SECOND_OCTET(pkt_node->ip_laddr[3]);
 489                 pkt_node->ip_laddr[2] = THIRD_OCTET(pkt_node->ip_laddr[3]);
 490                 pkt_node->ip_laddr[3] = FOURTH_OCTET(pkt_node->ip_laddr[3]);
 491                 pkt_node->ip_faddr[0] = FIRST_OCTET(pkt_node->ip_faddr[3]);
 492                 pkt_node->ip_faddr[1] = SECOND_OCTET(pkt_node->ip_faddr[3]);
 493                 pkt_node->ip_faddr[2] = THIRD_OCTET(pkt_node->ip_faddr[3]);
 494                 pkt_node->ip_faddr[3] = FOURTH_OCTET(pkt_node->ip_faddr[3]);
 495 #endif /* SIFTR_IPV6 */
 496
 497                 /* Construct an IPv4 log message. */
 498                 log_buf->ae_bytesused = snprintf(log_buf->ae_data,
 499                     MAX_LOG_MSG_LEN,
 500                     "%c,0x%08x,%jd.%06ld,%u.%u.%u.%u,%u,%u.%u.%u.%u,%u,%ld,%ld,"
 501                     "%ld,%ld,%ld,%u,%u,%u,%u,%u,%u,%u,%d,%u,%u,%u,%u,%u,%u\n",
 502                     direction[pkt_node->direction],
 503                     pkt_node->hash,
 504                     (intmax_t)pkt_node->tval.tv_sec,
 505                     pkt_node->tval.tv_usec,
 506                     pkt_node->ip_laddr[0],
 507                     pkt_node->ip_laddr[1],
 508                     pkt_node->ip_laddr[2],
 509                     pkt_node->ip_laddr[3],
 510                     ntohs(pkt_node->tcp_localport),
 511                     pkt_node->ip_faddr[0],
 512                     pkt_node->ip_faddr[1],
 513                     pkt_node->ip_faddr[2],
 514                     pkt_node->ip_faddr[3],
 515                     ntohs(pkt_node->tcp_foreignport),
 516                     pkt_node->snd_ssthresh,
 517                     pkt_node->snd_cwnd,
 518                     pkt_node->snd_bwnd,
 519                     pkt_node->snd_wnd,
 520                     pkt_node->rcv_wnd,
 521                     pkt_node->snd_scale,
 522                     pkt_node->rcv_scale,
 523                     pkt_node->conn_state,
 524                     pkt_node->max_seg_size,
 525                     pkt_node->smoothed_rtt,
 526                     pkt_node->sack_enabled,
 527                     pkt_node->flags,
 528                     pkt_node->rxt_length,
 529                     pkt_node->snd_buf_hiwater,
 530                     pkt_node->snd_buf_cc,
 531                     pkt_node->rcv_buf_hiwater,
 532                     pkt_node->rcv_buf_cc,
 533                     pkt_node->sent_inflight_bytes,
 534                     pkt_node->t_segqlen);
 535 #ifdef SIFTR_IPV6
 536         }
 537 #endif
 538
 539         alq_post_flags(siftr_alq, log_buf, 0);
 540 }
 541
 542
 543 static void
 544 siftr_pkt_manager_thread(void *arg)
 545 {
 546         STAILQ_HEAD(pkthead, pkt_node) tmp_pkt_queue =
 547             STAILQ_HEAD_INITIALIZER(tmp_pkt_queue);
 548         struct pkt_node *pkt_node, *pkt_node_temp;
 549         uint8_t draining;
 550
 551         draining = 2;
 552
 553         mtx_lock(&siftr_pkt_mgr_mtx);
 554
 555         /* draining == 0 when queue has been flushed and it's safe to exit. */
 556         while (draining) {
 557                 /*
 558                  * Sleep until we are signalled to wake because thread has
 559                  * been told to exit or until 1 tick has passed.
 560                  */
 561                 mtx_sleep(&wait_for_pkt, &siftr_pkt_mgr_mtx, PWAIT, "pktwait",
 562                     1);
 563
 564                 /* Gain exclusive access to the pkt_node queue. */
 565                 mtx_lock(&siftr_pkt_queue_mtx);
 566
 567                 /*
 568                  * Move pkt_queue to tmp_pkt_queue, which leaves
 569                  * pkt_queue empty and ready to receive more pkt_nodes.
 570                  */
 571                 STAILQ_CONCAT(&tmp_pkt_queue, &pkt_queue);
 572
 573                 /*
 574                  * We've finished making changes to the list. Unlock it
 575                  * so the pfil hooks can continue queuing pkt_nodes.
 576                  */
 577                 mtx_unlock(&siftr_pkt_queue_mtx);
 578
 579                 /*
 580                  * We can't hold a mutex whilst calling siftr_process_pkt
 581                  * because ALQ might sleep waiting for buffer space.
 582                  */
 583                 mtx_unlock(&siftr_pkt_mgr_mtx);
 584
 585                 /* Flush all pkt_nodes to the log file. */
 586                 STAILQ_FOREACH_SAFE(pkt_node, &tmp_pkt_queue, nodes,
 587                     pkt_node_temp) {
 588                         siftr_process_pkt(pkt_node);
 589                         STAILQ_REMOVE_HEAD(&tmp_pkt_queue, nodes);
 590                         free(pkt_node, M_SIFTR_PKTNODE);
 591                 }
 592
 593                 KASSERT(STAILQ_EMPTY(&tmp_pkt_queue),
 594                     ("SIFTR tmp_pkt_queue not empty after flush"));
 595
 596                 mtx_lock(&siftr_pkt_mgr_mtx);
 597
 598                 /*
 599                  * If siftr_exit_pkt_manager_thread gets set during the window
 600                  * where we are draining the tmp_pkt_queue above, there might
 601                  * still be pkts in pkt_queue that need to be drained.
 602                  * Allow one further iteration to occur after
 603                  * siftr_exit_pkt_manager_thread has been set to ensure
 604                  * pkt_queue is completely empty before we kill the thread.
 605                  *
 606                  * siftr_exit_pkt_manager_thread is set only after the pfil
 607                  * hooks have been removed, so only 1 extra iteration
 608                  * is needed to drain the queue.
 609                  */
 610                 if (siftr_exit_pkt_manager_thread)
 611                         draining--;
 612         }
 613
 614         mtx_unlock(&siftr_pkt_mgr_mtx);
 615
 616         /* Calls wakeup on this thread's struct thread ptr. */
 617         kthread_exit();
 618 }
 619
 620
 621 static uint32_t
 622 hash_pkt(struct mbuf *m, uint32_t offset)
 623 {
 624         uint32_t hash;
 625
 626         hash = 0;
 627
 628         while (m != NULL && offset > m->m_len) {
 629                 /*
 630                  * The IP packet payload does not start in this mbuf, so
 631                  * need to figure out which mbuf it starts in and what offset
 632                  * into the mbuf's data region the payload starts at.
 633                  */
 634                 offset -= m->m_len;
 635                 m = m->m_next;
 636         }
 637
 638         while (m != NULL) {
 639                 /* Ensure there is data in the mbuf */
 640                 if ((m->m_len - offset) > 0)
 641                         hash = hash32_buf(m->m_data + offset,
 642                             m->m_len - offset, hash);
 643
 644                 m = m->m_next;
 645                 offset = 0;
 646         }
 647
 648         return (hash);
 649 }
 650
 651
 652 /*
 653  * Check if a given mbuf has the SIFTR mbuf tag. If it does, log the fact that
 654  * it's a reinjected packet and return. If it doesn't, tag the mbuf and return.
 655  * Return value >0 means the caller should skip processing this mbuf.
 656  */
 657 static inline int
 658 siftr_chkreinject(struct mbuf *m, int dir, struct siftr_stats *ss)
 659 {
 660         if (m_tag_locate(m, PACKET_COOKIE_SIFTR, PACKET_TAG_SIFTR, NULL)
 661             != NULL) {
 662                 if (dir == PFIL_IN)
 663                         ss->nskip_in_dejavu++;
 664                 else
 665                         ss->nskip_out_dejavu++;
 666
 667                 return (1);
 668         } else {
 669                 struct m_tag *tag = m_tag_alloc(PACKET_COOKIE_SIFTR,
 670                     PACKET_TAG_SIFTR, 0, M_NOWAIT);
 671                 if (tag == NULL) {
 672                         if (dir == PFIL_IN)
 673                                 ss->nskip_in_malloc++;
 674                         else
 675                                 ss->nskip_out_malloc++;
 676
 677                         return (1);
 678                 }
 679
 680                 m_tag_prepend(m, tag);
 681         }
 682
 683         return (0);
 684 }
 685
 686
 687 /*
 688  * Look up an inpcb for a packet. Return the inpcb pointer if found, or NULL
 689  * otherwise.
 690  */
 691 static inline struct inpcb *
 692 siftr_findinpcb(int ipver, struct ip *ip, struct mbuf *m, uint16_t sport,
 693     uint16_t dport, int dir, struct siftr_stats *ss)
 694 {
 695         struct inpcb *inp;
 696
 697         /* We need the tcbinfo lock. */
 698         INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
 699         INP_INFO_RLOCK(&V_tcbinfo);
 700
 701         if (dir == PFIL_IN)
 702                 inp = (ipver == INP_IPV4 ?
 703                     in_pcblookup_hash(&V_tcbinfo, ip->ip_src, sport, ip->ip_dst,
 704                     dport, 0, m->m_pkthdr.rcvif)
 705                     :
 706 #ifdef SIFTR_IPV6
 707                     in6_pcblookup_hash(&V_tcbinfo,
 708                     &((struct ip6_hdr *)ip)->ip6_src, sport,
 709                     &((struct ip6_hdr *)ip)->ip6_dst, dport, 0,
 710                     m->m_pkthdr.rcvif)
 711 #else
 712                     NULL
 713 #endif
 714                     );
 715
 716         else
 717                 inp = (ipver == INP_IPV4 ?
 718                     in_pcblookup_hash(&V_tcbinfo, ip->ip_dst, dport, ip->ip_src,
 719                     sport, 0, m->m_pkthdr.rcvif)
 720                     :
 721 #ifdef SIFTR_IPV6
 722                     in6_pcblookup_hash(&V_tcbinfo,
 723                     &((struct ip6_hdr *)ip)->ip6_dst, dport,
 724                     &((struct ip6_hdr *)ip)->ip6_src, sport, 0,
 725                     m->m_pkthdr.rcvif)
 726 #else
 727                     NULL
 728 #endif
 729                     );
 730
 731         /* If we can't find the inpcb, bail. */
 732         if (inp == NULL) {
 733                 if (dir == PFIL_IN)
 734                         ss->nskip_in_inpcb++;
 735                 else
 736                         ss->nskip_out_inpcb++;
 737         } else {
 738                 /* Acquire the inpcb lock. */
 739                 INP_UNLOCK_ASSERT(inp);
 740                 INP_RLOCK(inp);
 741         }
 742         INP_INFO_RUNLOCK(&V_tcbinfo);
 743
 744         return (inp);
 745 }
 746
 747
 748 static inline void
 749 siftr_siftdata(struct pkt_node *pn, struct inpcb *inp, struct tcpcb *tp,
 750     int ipver, int dir, int inp_locally_locked)
 751 {
 752 #ifdef SIFTR_IPV6
 753         if (ipver == INP_IPV4) {
 754                 pn->ip_laddr[3] = inp->inp_laddr.s_addr;
 755                 pn->ip_faddr[3] = inp->inp_faddr.s_addr;
 756 #else
 757                 *((uint32_t *)pn->ip_laddr) = inp->inp_laddr.s_addr;
 758                 *((uint32_t *)pn->ip_faddr) = inp->inp_faddr.s_addr;
 759 #endif
 760 #ifdef SIFTR_IPV6
 761         } else {
 762                 pn->ip_laddr[0] = inp->in6p_laddr.s6_addr32[0];
 763                 pn->ip_laddr[1] = inp->in6p_laddr.s6_addr32[1];
 764                 pn->ip_laddr[2] = inp->in6p_laddr.s6_addr32[2];
 765                 pn->ip_laddr[3] = inp->in6p_laddr.s6_addr32[3];
 766                 pn->ip_faddr[0] = inp->in6p_faddr.s6_addr32[0];
 767                 pn->ip_faddr[1] = inp->in6p_faddr.s6_addr32[1];
 768                 pn->ip_faddr[2] = inp->in6p_faddr.s6_addr32[2];
 769                 pn->ip_faddr[3] = inp->in6p_faddr.s6_addr32[3];
 770         }
 771 #endif
 772         pn->tcp_localport = inp->inp_lport;
 773         pn->tcp_foreignport = inp->inp_fport;
 774         pn->snd_cwnd = tp->snd_cwnd;
 775         pn->snd_wnd = tp->snd_wnd;
 776         pn->rcv_wnd = tp->rcv_wnd;
 777         pn->snd_bwnd = tp->snd_bwnd;
 778         pn->snd_ssthresh = tp->snd_ssthresh;
 779         pn->snd_scale = tp->snd_scale;
 780         pn->rcv_scale = tp->rcv_scale;
 781         pn->conn_state = tp->t_state;
 782         pn->max_seg_size = tp->t_maxseg;
 783         pn->smoothed_rtt = tp->t_srtt;
 784         pn->sack_enabled = (tp->t_flags & TF_SACK_PERMIT) != 0;
 785         pn->flags = tp->t_flags;
 786         pn->rxt_length = tp->t_rxtcur;
 787         pn->snd_buf_hiwater = inp->inp_socket->so_snd.sb_hiwat;
 788         pn->snd_buf_cc = inp->inp_socket->so_snd.sb_cc;
 789         pn->rcv_buf_hiwater = inp->inp_socket->so_rcv.sb_hiwat;
 790         pn->rcv_buf_cc = inp->inp_socket->so_rcv.sb_cc;
 791         pn->sent_inflight_bytes = tp->snd_max - tp->snd_una;
 792         pn->t_segqlen = tp->t_segqlen;
 793
 794         /* We've finished accessing the tcb so release the lock. */
 795         if (inp_locally_locked)
 796                 INP_RUNLOCK(inp);
 797
 798         pn->ipver = ipver;
 799         pn->direction = dir;
 800
 801         /*
 802          * Significantly more accurate than using getmicrotime(), but slower!
 803          * Gives true microsecond resolution at the expense of a hit to
 804          * maximum pps throughput processing when SIFTR is loaded and enabled.
 805          */
 806         microtime(&pn->tval);
 807 }
 808
 809
 810 /*
 811  * pfil hook that is called for each IPv4 packet making its way through the
 812  * stack in either direction.
 813  * The pfil subsystem holds a non-sleepable mutex somewhere when
 814  * calling our hook function, so we can't sleep at all.
 815  * It's very important to use the M_NOWAIT flag with all function calls
 816  * that support it so that they won't sleep, otherwise you get a panic.
 817  */
 818 static int
 819 siftr_chkpkt(void *arg, struct mbuf **m, struct ifnet *ifp, int dir,
 820     struct inpcb *inp)
 821 {
 822         struct pkt_node *pn;
 823         struct ip *ip;
 824         struct tcphdr *th;
 825         struct tcpcb *tp;
 826         struct siftr_stats *ss;
 827         unsigned int ip_hl;
 828         int inp_locally_locked;
 829
 830         inp_locally_locked = 0;
 831         ss = DPCPU_PTR(ss);
 832
 833         /*
 834          * m_pullup is not required here because ip_{input|output}
 835          * already do the heavy lifting for us.
 836          */
 837
 838         ip = mtod(*m, struct ip *);
 839
 840         /* Only continue processing if the packet is TCP. */
 841         if (ip->ip_p != IPPROTO_TCP)
 842                 goto ret;
 843
 844         /*
 845          * If a kernel subsystem reinjects packets into the stack, our pfil
 846          * hook will be called multiple times for the same packet.
 847          * Make sure we only process unique packets.
 848          */
 849         if (siftr_chkreinject(*m, dir, ss))
 850                 goto ret;
 851
 852         if (dir == PFIL_IN)
 853                 ss->n_in++;
 854         else
 855                 ss->n_out++;
 856
 857         /*
 858          * Create a tcphdr struct starting at the correct offset
 859          * in the IP packet. ip->ip_hl gives the ip header length
 860          * in 4-byte words, so multiply it to get the size in bytes.
 861          */
 862         ip_hl = (ip->ip_hl << 2);
 863         th = (struct tcphdr *)((caddr_t)ip + ip_hl);
 864
 865         /*
 866          * If the pfil hooks don't provide a pointer to the
 867          * inpcb, we need to find it ourselves and lock it.
 868          */
 869         if (!inp) {
 870                 /* Find the corresponding inpcb for this pkt. */
 871                 inp = siftr_findinpcb(INP_IPV4, ip, *m, th->th_sport,
 872                     th->th_dport, dir, ss);
 873
 874                 if (inp == NULL)
 875                         goto ret;
 876                 else
 877                         inp_locally_locked = 1;
 878         }
 879
 880         INP_LOCK_ASSERT(inp);
 881
 882         /* Find the TCP control block that corresponds with this packet */
 883         tp = intotcpcb(inp);
 884
 885         /*
 886          * If we can't find the TCP control block (happens occasionaly for a
 887          * packet sent during the shutdown phase of a TCP connection),
 888          * or we're in the timewait state, bail
 889          */
 890         if (tp == NULL || inp->inp_flags & INP_TIMEWAIT) {
 891                 if (dir == PFIL_IN)
 892                         ss->nskip_in_tcpcb++;
 893                 else
 894                         ss->nskip_out_tcpcb++;
 895
 896                 goto inp_unlock;
 897         }
 898
 899         pn = malloc(sizeof(struct pkt_node), M_SIFTR_PKTNODE, M_NOWAIT|M_ZERO);
 900
 901         if (pn == NULL) {
 902                 if (dir == PFIL_IN)
 903                         ss->nskip_in_malloc++;
 904                 else
 905                         ss->nskip_out_malloc++;
 906
 907                 goto inp_unlock;
 908         }
 909
 910         siftr_siftdata(pn, inp, tp, INP_IPV4, dir, inp_locally_locked);
 911
 912         if (siftr_generate_hashes) {
 913                 if ((*m)->m_pkthdr.csum_flags & CSUM_TCP) {
 914                         /*
 915                          * For outbound packets, the TCP checksum isn't
 916                          * calculated yet. This is a problem for our packet
 917                          * hashing as the receiver will calc a different hash
 918                          * to ours if we don't include the correct TCP checksum
 919                          * in the bytes being hashed. To work around this
 920                          * problem, we manually calc the TCP checksum here in
 921                          * software. We unset the CSUM_TCP flag so the lower
 922                          * layers don't recalc it.
 923                          */
 924                         (*m)->m_pkthdr.csum_flags &= ~CSUM_TCP;
 925
 926                         /*
 927                          * Calculate the TCP checksum in software and assign
 928                          * to correct TCP header field, which will follow the
 929                          * packet mbuf down the stack. The trick here is that
 930                          * tcp_output() sets th->th_sum to the checksum of the
 931                          * pseudo header for us already. Because of the nature
 932                          * of the checksumming algorithm, we can sum over the
 933                          * entire IP payload (i.e. TCP header and data), which
 934                          * will include the already calculated pseduo header
 935                          * checksum, thus giving us the complete TCP checksum.
 936                          *
 937                          * To put it in simple terms, if checksum(1,2,3,4)=10,
 938                          * then checksum(1,2,3,4,5) == checksum(10,5).
 939                          * This property is what allows us to "cheat" and
 940                          * checksum only the IP payload which has the TCP
 941                          * th_sum field populated with the pseudo header's
 942                          * checksum, and not need to futz around checksumming
 943                          * pseudo header bytes and TCP header/data in one hit.
 944                          * Refer to RFC 1071 for more info.
 945                          *
 946                          * NB: in_cksum_skip(struct mbuf *m, int len, int skip)
 947                          * in_cksum_skip 2nd argument is NOT the number of
 948                          * bytes to read from the mbuf at "skip" bytes offset
 949                          * from the start of the mbuf (very counter intuitive!).
 950                          * The number of bytes to read is calculated internally
 951                          * by the function as len-skip i.e. to sum over the IP
 952                          * payload (TCP header + data) bytes, it is INCORRECT
 953                          * to call the function like this:
 954                          * in_cksum_skip(at, ip->ip_len - offset, offset)
 955                          * Rather, it should be called like this:
 956                          * in_cksum_skip(at, ip->ip_len, offset)
 957                          * which means read "ip->ip_len - offset" bytes from
 958                          * the mbuf cluster "at" at offset "offset" bytes from
 959                          * the beginning of the "at" mbuf's data pointer.
 960                          */
 961                         th->th_sum = in_cksum_skip(*m, ip->ip_len, ip_hl);
 962                 }
 963
 964                 /*
 965                  * XXX: Having to calculate the checksum in software and then
 966                  * hash over all bytes is really inefficient. Would be nice to
 967                  * find a way to create the hash and checksum in the same pass
 968                  * over the bytes.
 969                  */
 970                 pn->hash = hash_pkt(*m, ip_hl);
 971         }
 972
 973         mtx_lock(&siftr_pkt_queue_mtx);
 974         STAILQ_INSERT_TAIL(&pkt_queue, pn, nodes);
 975         mtx_unlock(&siftr_pkt_queue_mtx);
 976         goto ret;
 977
 978 inp_unlock:
 979         if (inp_locally_locked)
 980                 INP_RUNLOCK(inp);
 981
 982 ret:
 983         /* Returning 0 ensures pfil will not discard the pkt */
 984         return (0);
 985 }
 986
 987
 988 #ifdef SIFTR_IPV6
 989 static int
 990 siftr_chkpkt6(void *arg, struct mbuf **m, struct ifnet *ifp, int dir,
 991     struct inpcb *inp)
 992 {
 993         struct pkt_node *pn;
 994         struct ip6_hdr *ip6;
 995         struct tcphdr *th;
 996         struct tcpcb *tp;
 997         struct siftr_stats *ss;
 998         unsigned int ip6_hl;
 999         int inp_locally_locked;
1000
1001         inp_locally_locked = 0;
1002         ss = DPCPU_PTR(ss);
1003
1004         /*
1005          * m_pullup is not required here because ip6_{input|output}
1006          * already do the heavy lifting for us.
1007          */
1008
1009         ip6 = mtod(*m, struct ip6_hdr *);
1010
1011         /*
1012          * Only continue processing if the packet is TCP
1013          * XXX: We should follow the next header fields
1014          * as shown on Pg 6 RFC 2460, but right now we'll
1015          * only check pkts that have no extension headers.
1016          */
1017         if (ip6->ip6_nxt != IPPROTO_TCP)
1018                 goto ret6;
1019
1020         /*
1021          * If a kernel subsystem reinjects packets into the stack, our pfil
1022          * hook will be called multiple times for the same packet.
1023          * Make sure we only process unique packets.
1024          */
1025         if (siftr_chkreinject(*m, dir, ss))
1026                 goto ret6;
1027
1028         if (dir == PFIL_IN)
1029                 ss->n_in++;
1030         else
1031                 ss->n_out++;
1032
1033         ip6_hl = sizeof(struct ip6_hdr);
1034
1035         /*
1036          * Create a tcphdr struct starting at the correct offset
1037          * in the ipv6 packet. ip->ip_hl gives the ip header length
1038          * in 4-byte words, so multiply it to get the size in bytes.
1039          */
1040         th = (struct tcphdr *)((caddr_t)ip6 + ip6_hl);
1041
1042         /*
1043          * For inbound packets, the pfil hooks don't provide a pointer to the
1044          * inpcb, so we need to find it ourselves and lock it.
1045          */
1046         if (!inp) {
1047                 /* Find the corresponding inpcb for this pkt. */
1048                 inp = siftr_findinpcb(INP_IPV6, (struct ip *)ip6, *m,
1049                     th->th_sport, th->th_dport, dir, ss);
1050
1051                 if (inp == NULL)
1052                         goto ret6;
1053                 else
1054                         inp_locally_locked = 1;
1055         }
1056
1057         /* Find the TCP control block that corresponds with this packet. */
1058         tp = intotcpcb(inp);
1059
1060         /*
1061          * If we can't find the TCP control block (happens occasionaly for a
1062          * packet sent during the shutdown phase of a TCP connection),
1063          * or we're in the timewait state, bail.
1064          */
1065         if (tp == NULL || inp->inp_flags & INP_TIMEWAIT) {
1066                 if (dir == PFIL_IN)
1067                         ss->nskip_in_tcpcb++;
1068                 else
1069                         ss->nskip_out_tcpcb++;
1070
1071                 goto inp_unlock6;
1072         }
1073
1074         pn = malloc(sizeof(struct pkt_node), M_SIFTR_PKTNODE, M_NOWAIT|M_ZERO);
1075
1076         if (pn == NULL) {
1077                 if (dir == PFIL_IN)
1078                         ss->nskip_in_malloc++;
1079                 else
1080                         ss->nskip_out_malloc++;
1081
1082                 goto inp_unlock6;
1083         }
1084
1085         siftr_siftdata(pn, inp, tp, INP_IPV6, dir, inp_locally_locked);
1086
1087         /* XXX: Figure out how to generate hashes for IPv6 packets. */
1088
1089         mtx_lock(&siftr_pkt_queue_mtx);
1090         STAILQ_INSERT_TAIL(&pkt_queue, pn, nodes);
1091         mtx_unlock(&siftr_pkt_queue_mtx);
1092         goto ret6;
1093
1094 inp_unlock6:
1095         if (inp_locally_locked)
1096                 INP_RUNLOCK(inp);
1097
1098 ret6:
1099         /* Returning 0 ensures pfil will not discard the pkt. */
1100         return (0);
1101 }
1102 #endif /* #ifdef SIFTR_IPV6 */
1103
1104
1105 static int
1106 siftr_pfil(int action)
1107 {
1108         struct pfil_head *pfh_inet;
1109 #ifdef SIFTR_IPV6
1110         struct pfil_head *pfh_inet6;
1111 #endif
1112         VNET_ITERATOR_DECL(vnet_iter);
1113
1114         VNET_LIST_RLOCK();
1115         VNET_FOREACH(vnet_iter) {
1116                 CURVNET_SET(vnet_iter);
1117                 pfh_inet = pfil_head_get(PFIL_TYPE_AF, AF_INET);
1118 #ifdef SIFTR_IPV6
1119                 pfh_inet6 = pfil_head_get(PFIL_TYPE_AF, AF_INET6);
1120 #endif
1121
1122                 if (action == HOOK) {
1123                         pfil_add_hook(siftr_chkpkt, NULL,
1124                             PFIL_IN | PFIL_OUT | PFIL_WAITOK, pfh_inet);
1125 #ifdef SIFTR_IPV6
1126                         pfil_add_hook(siftr_chkpkt6, NULL,
1127                             PFIL_IN | PFIL_OUT | PFIL_WAITOK, pfh_inet6);
1128 #endif
1129                 } else if (action == UNHOOK) {
1130                         pfil_remove_hook(siftr_chkpkt, NULL,
1131                             PFIL_IN | PFIL_OUT | PFIL_WAITOK, pfh_inet);
1132 #ifdef SIFTR_IPV6
1133                         pfil_remove_hook(siftr_chkpkt6, NULL,
1134                             PFIL_IN | PFIL_OUT | PFIL_WAITOK, pfh_inet6);
1135 #endif
1136                 }
1137                 CURVNET_RESTORE();
1138         }
1139         VNET_LIST_RUNLOCK();
1140
1141         return (0);
1142 }
1143
1144
1145 static int
1146 siftr_sysctl_logfile_name_handler(SYSCTL_HANDLER_ARGS)
1147 {
1148         struct alq *new_alq;
1149         int error;
1150
1151         if (req->newptr == NULL)
1152                 goto skip;
1153
1154         /* If old filename and new filename are different. */
1155         if (strncmp(siftr_logfile, (char *)req->newptr, PATH_MAX)) {
1156
1157                 error = alq_open(&new_alq, req->newptr, curthread->td_ucred,
1158                     SIFTR_LOG_FILE_MODE, SIFTR_ALQ_BUFLEN, 0);
1159
1160                 /* Bail if unable to create new alq. */
1161                 if (error)
1162                         return (1);
1163
1164                 /*
1165                  * If disabled, siftr_alq == NULL so we simply close
1166                  * the alq as we've proved it can be opened.
1167                  * If enabled, close the existing alq and switch the old
1168                  * for the new.
1169                  */
1170                 if (siftr_alq == NULL)
1171                         alq_close(new_alq);
1172                 else {
1173                         alq_close(siftr_alq);
1174                         siftr_alq = new_alq;
1175                 }
1176         }
1177
1178 skip:
1179         return (sysctl_handle_string(oidp, arg1, arg2, req));
1180 }
1181
1182
1183 static int
1184 siftr_manage_ops(uint8_t action)
1185 {
1186         struct siftr_stats totalss;
1187         struct timeval tval;
1188         struct flow_hash_node *counter, *tmp_counter;
1189         struct sbuf *s;
1190         int i, key_index, ret, error;
1191         uint32_t bytes_to_write, total_skipped_pkts;
1192         uint16_t lport, fport;
1193         uint8_t *key, ipver;
1194
1195 #ifdef SIFTR_IPV6
1196         uint32_t laddr[4];
1197         uint32_t faddr[4];
1198 #else
1199         uint8_t laddr[4];
1200         uint8_t faddr[4];
1201 #endif
1202
1203         error = 0;
1204         total_skipped_pkts = 0;
1205
1206         /* Init an autosizing sbuf that initially holds 200 chars. */
1207         if ((s = sbuf_new(NULL, NULL, 200, SBUF_AUTOEXTEND)) == NULL)
1208                 return (-1);
1209
1210         if (action == SIFTR_ENABLE) {
1211                 /*
1212                  * Create our alq
1213                  * XXX: We should abort if alq_open fails!
1214                  */
1215                 alq_open(&siftr_alq, siftr_logfile, curthread->td_ucred,
1216                     SIFTR_LOG_FILE_MODE, SIFTR_ALQ_BUFLEN, 0);
1217
1218                 STAILQ_INIT(&pkt_queue);
1219
1220                 DPCPU_ZERO(ss);
1221
1222                 siftr_exit_pkt_manager_thread = 0;
1223
1224                 ret = kthread_add(&siftr_pkt_manager_thread, NULL, NULL,
1225                     &siftr_pkt_manager_thr, RFNOWAIT, 0,
1226                     "siftr_pkt_manager_thr");
1227
1228                 siftr_pfil(HOOK);
1229
1230                 microtime(&tval);
1231
1232                 sbuf_printf(s,
1233                     "enable_time_secs=%jd\tenable_time_usecs=%06ld\t"
1234                     "siftrver=%s\thz=%u\ttcp_rtt_scale=%u\tsysname=%s\t"
1235                     "sysver=%u\tipmode=%u\n",
1236                     (intmax_t)tval.tv_sec, tval.tv_usec, MODVERSION_STR, hz,
1237                     TCP_RTT_SCALE, SYS_NAME, __FreeBSD_version, SIFTR_IPMODE);
1238
1239                 sbuf_finish(s);
1240                 alq_writen(siftr_alq, sbuf_data(s), sbuf_len(s), ALQ_WAITOK);
1241
1242         } else if (action == SIFTR_DISABLE && siftr_pkt_manager_thr != NULL) {
1243                 /*
1244                  * Remove the pfil hook functions. All threads currently in
1245                  * the hook functions are allowed to exit before siftr_pfil()
1246                  * returns.
1247                  */
1248                 siftr_pfil(UNHOOK);
1249
1250                 /* This will block until the pkt manager thread unlocks it. */
1251                 mtx_lock(&siftr_pkt_mgr_mtx);
1252
1253                 /* Tell the pkt manager thread that it should exit now. */
1254                 siftr_exit_pkt_manager_thread = 1;
1255
1256                 /*
1257                  * Wake the pkt_manager thread so it realises that
1258                  * siftr_exit_pkt_manager_thread == 1 and exits gracefully.
1259                  * The wakeup won't be delivered until we unlock
1260                  * siftr_pkt_mgr_mtx so this isn't racy.
1261                  */
1262                 wakeup(&wait_for_pkt);
1263
1264                 /* Wait for the pkt_manager thread to exit. */
1265                 mtx_sleep(siftr_pkt_manager_thr, &siftr_pkt_mgr_mtx, PWAIT,
1266                     "thrwait", 0);
1267
1268                 siftr_pkt_manager_thr = NULL;
1269                 mtx_unlock(&siftr_pkt_mgr_mtx);
1270
1271                 totalss.n_in = DPCPU_VARSUM(ss, n_in);
1272                 totalss.n_out = DPCPU_VARSUM(ss, n_out);
1273                 totalss.nskip_in_malloc = DPCPU_VARSUM(ss, nskip_in_malloc);
1274                 totalss.nskip_out_malloc = DPCPU_VARSUM(ss, nskip_out_malloc);
1275                 totalss.nskip_in_mtx = DPCPU_VARSUM(ss, nskip_in_mtx);
1276                 totalss.nskip_out_mtx = DPCPU_VARSUM(ss, nskip_out_mtx);
1277                 totalss.nskip_in_tcpcb = DPCPU_VARSUM(ss, nskip_in_tcpcb);
1278                 totalss.nskip_out_tcpcb = DPCPU_VARSUM(ss, nskip_out_tcpcb);
1279                 totalss.nskip_in_inpcb = DPCPU_VARSUM(ss, nskip_in_inpcb);
1280                 totalss.nskip_out_inpcb = DPCPU_VARSUM(ss, nskip_out_inpcb);
1281
1282                 total_skipped_pkts = totalss.nskip_in_malloc +
1283                     totalss.nskip_out_malloc + totalss.nskip_in_mtx +
1284                     totalss.nskip_out_mtx + totalss.nskip_in_tcpcb +
1285                     totalss.nskip_out_tcpcb + totalss.nskip_in_inpcb +
1286                     totalss.nskip_out_inpcb;
1287
1288                 microtime(&tval);
1289
1290                 sbuf_printf(s,
1291                     "disable_time_secs=%jd\tdisable_time_usecs=%06ld\t"
1292                     "num_inbound_tcp_pkts=%ju\tnum_outbound_tcp_pkts=%ju\t"
1293                     "total_tcp_pkts=%ju\tnum_inbound_skipped_pkts_malloc=%u\t"
1294                     "num_outbound_skipped_pkts_malloc=%u\t"
1295                     "num_inbound_skipped_pkts_mtx=%u\t"
1296                     "num_outbound_skipped_pkts_mtx=%u\t"
1297                     "num_inbound_skipped_pkts_tcpcb=%u\t"
1298                     "num_outbound_skipped_pkts_tcpcb=%u\t"
1299                     "num_inbound_skipped_pkts_inpcb=%u\t"
1300                     "num_outbound_skipped_pkts_inpcb=%u\t"
1301                     "total_skipped_tcp_pkts=%u\tflow_list=",
1302                     (intmax_t)tval.tv_sec,
1303                     tval.tv_usec,
1304                     (uintmax_t)totalss.n_in,
1305                     (uintmax_t)totalss.n_out,
1306                     (uintmax_t)(totalss.n_in + totalss.n_out),
1307                     totalss.nskip_in_malloc,
1308                     totalss.nskip_out_malloc,
1309                     totalss.nskip_in_mtx,
1310                     totalss.nskip_out_mtx,
1311                     totalss.nskip_in_tcpcb,
1312                     totalss.nskip_out_tcpcb,
1313                     totalss.nskip_in_inpcb,
1314                     totalss.nskip_out_inpcb,
1315                     total_skipped_pkts);
1316
1317                 /*
1318                  * Iterate over the flow hash, printing a summary of each
1319                  * flow seen and freeing any malloc'd memory.
1320                  * The hash consists of an array of LISTs (man 3 queue).
1321                  */
1322                 for (i = 0; i <= siftr_hashmask; i++) {
1323                         LIST_FOREACH_SAFE(counter, counter_hash + i, nodes,
1324                             tmp_counter) {
1325                                 key = counter->key;
1326                                 key_index = 1;
1327
1328                                 ipver = key[0];
1329
1330                                 memcpy(laddr, key + key_index, sizeof(laddr));
1331                                 key_index += sizeof(laddr);
1332                                 memcpy(&lport, key + key_index, sizeof(lport));
1333                                 key_index += sizeof(lport);
1334                                 memcpy(faddr, key + key_index, sizeof(faddr));
1335                                 key_index += sizeof(faddr);
1336                                 memcpy(&fport, key + key_index, sizeof(fport));
1337
1338 #ifdef SIFTR_IPV6
1339                                 laddr[3] = ntohl(laddr[3]);
1340                                 faddr[3] = ntohl(faddr[3]);
1341
1342                                 if (ipver == INP_IPV6) {
1343                                         laddr[0] = ntohl(laddr[0]);
1344                                         laddr[1] = ntohl(laddr[1]);
1345                                         laddr[2] = ntohl(laddr[2]);
1346                                         faddr[0] = ntohl(faddr[0]);
1347                                         faddr[1] = ntohl(faddr[1]);
1348                                         faddr[2] = ntohl(faddr[2]);
1349
1350                                         sbuf_printf(s,
1351                                             "%x:%x:%x:%x:%x:%x:%x:%x;%u-"
1352                                             "%x:%x:%x:%x:%x:%x:%x:%x;%u,",
1353                                             UPPER_SHORT(laddr[0]),
1354                                             LOWER_SHORT(laddr[0]),
1355                                             UPPER_SHORT(laddr[1]),
1356                                             LOWER_SHORT(laddr[1]),
1357                                             UPPER_SHORT(laddr[2]),
1358                                             LOWER_SHORT(laddr[2]),
1359                                             UPPER_SHORT(laddr[3]),
1360                                             LOWER_SHORT(laddr[3]),
1361                                             ntohs(lport),
1362                                             UPPER_SHORT(faddr[0]),
1363                                             LOWER_SHORT(faddr[0]),
1364                                             UPPER_SHORT(faddr[1]),
1365                                             LOWER_SHORT(faddr[1]),
1366                                             UPPER_SHORT(faddr[2]),
1367                                             LOWER_SHORT(faddr[2]),
1368                                             UPPER_SHORT(faddr[3]),
1369                                             LOWER_SHORT(faddr[3]),
1370                                             ntohs(fport));
1371                                 } else {
1372                                         laddr[0] = FIRST_OCTET(laddr[3]);
1373                                         laddr[1] = SECOND_OCTET(laddr[3]);
1374                                         laddr[2] = THIRD_OCTET(laddr[3]);
1375                                         laddr[3] = FOURTH_OCTET(laddr[3]);
1376                                         faddr[0] = FIRST_OCTET(faddr[3]);
1377                                         faddr[1] = SECOND_OCTET(faddr[3]);
1378                                         faddr[2] = THIRD_OCTET(faddr[3]);
1379                                         faddr[3] = FOURTH_OCTET(faddr[3]);
1380 #endif
1381                                         sbuf_printf(s,
1382                                             "%u.%u.%u.%u;%u-%u.%u.%u.%u;%u,",
1383                                             laddr[0],
1384                                             laddr[1],
1385                                             laddr[2],
1386                                             laddr[3],
1387                                             ntohs(lport),
1388                                             faddr[0],
1389                                             faddr[1],
1390                                             faddr[2],
1391                                             faddr[3],
1392                                             ntohs(fport));
1393 #ifdef SIFTR_IPV6
1394                                 }
1395 #endif
1396
1397                                 free(counter, M_SIFTR_HASHNODE);
1398                         }
1399
1400                         LIST_INIT(counter_hash + i);
1401                 }
1402
1403                 sbuf_printf(s, "\n");
1404                 sbuf_finish(s);
1405
1406                 i = 0;
1407                 do {
1408                         bytes_to_write = min(SIFTR_ALQ_BUFLEN, sbuf_len(s)-i);
1409                         alq_writen(siftr_alq, sbuf_data(s)+i, bytes_to_write, ALQ_WAITOK);
1410                         i += bytes_to_write;
1411                 } while (i < sbuf_len(s));
1412
1413                 alq_close(siftr_alq);
1414                 siftr_alq = NULL;
1415         }
1416
1417         sbuf_delete(s);
1418
1419         /*
1420          * XXX: Should be using ret to check if any functions fail
1421          * and set error appropriately
1422          */
1423
1424         return (error);
1425 }
1426
1427
1428 static int
1429 siftr_sysctl_enabled_handler(SYSCTL_HANDLER_ARGS)
1430 {
1431         if (req->newptr == NULL)
1432                 goto skip;
1433
1434         /* If the value passed in isn't 0 or 1, return an error. */
1435         if (CAST_PTR_INT(req->newptr) != 0 && CAST_PTR_INT(req->newptr) != 1)
1436                 return (1);
1437
1438         /* If we are changing state (0 to 1 or 1 to 0). */
1439         if (CAST_PTR_INT(req->newptr) != siftr_enabled )
1440                 if (siftr_manage_ops(CAST_PTR_INT(req->newptr))) {
1441                         siftr_manage_ops(SIFTR_DISABLE);
1442                         return (1);
1443                 }
1444
1445 skip:
1446         return (sysctl_handle_int(oidp, arg1, arg2, req));
1447 }
1448
1449
1450 static void
1451 siftr_shutdown_handler(void *arg)
1452 {
1453         siftr_manage_ops(SIFTR_DISABLE);
1454 }
1455
1456
1457 /*
1458  * Module is being unloaded or machine is shutting down. Take care of cleanup.
1459  */
1460 static int
1461 deinit_siftr(void)
1462 {
1463         /* Cleanup. */
1464         siftr_manage_ops(SIFTR_DISABLE);
1465         hashdestroy(counter_hash, M_SIFTR, siftr_hashmask);
1466         mtx_destroy(&siftr_pkt_queue_mtx);
1467         mtx_destroy(&siftr_pkt_mgr_mtx);
1468
1469         return (0);
1470 }
1471
1472
1473 /*
1474  * Module has just been loaded into the kernel.
1475  */
1476 static int
1477 init_siftr(void)
1478 {
1479         EVENTHANDLER_REGISTER(shutdown_pre_sync, siftr_shutdown_handler, NULL,
1480             SHUTDOWN_PRI_FIRST);
1481
1482         /* Initialise our flow counter hash table. */
1483         counter_hash = hashinit(SIFTR_EXPECTED_MAX_TCP_FLOWS, M_SIFTR,
1484             &siftr_hashmask);
1485
1486         mtx_init(&siftr_pkt_queue_mtx, "siftr_pkt_queue_mtx", NULL, MTX_DEF);
1487         mtx_init(&siftr_pkt_mgr_mtx, "siftr_pkt_mgr_mtx", NULL, MTX_DEF);
1488
1489         /* Print message to the user's current terminal. */
1490         uprintf("\nStatistical Information For TCP Research (SIFTR) %s\n"
1491             "          http://caia.swin.edu.au/urp/newtcp\n\n",
1492             MODVERSION_STR);
1493
1494         return (0);
1495 }
1496
1497
1498 /*
1499  * This is the function that is called to load and unload the module.
1500  * When the module is loaded, this function is called once with
1501  * "what" == MOD_LOAD
1502  * When the module is unloaded, this function is called twice with
1503  * "what" = MOD_QUIESCE first, followed by "what" = MOD_UNLOAD second
1504  * When the system is shut down e.g. CTRL-ALT-DEL or using the shutdown command,
1505  * this function is called once with "what" = MOD_SHUTDOWN
1506  * When the system is shut down, the handler isn't called until the very end
1507  * of the shutdown sequence i.e. after the disks have been synced.
1508  */
1509 static int
1510 siftr_load_handler(module_t mod, int what, void *arg)
1511 {
1512         int ret;
1513
1514         switch (what) {
1515         case MOD_LOAD:
1516                 ret = init_siftr();
1517                 break;
1518
1519         case MOD_QUIESCE:
1520         case MOD_SHUTDOWN:
1521                 ret = deinit_siftr();
1522                 break;
1523
1524         case MOD_UNLOAD:
1525                 ret = 0;
1526                 break;
1527
1528         default:
1529                 ret = EINVAL;
1530                 break;
1531         }
1532
1533         return (ret);
1534 }
1535
1536
1537 static moduledata_t siftr_mod = {
1538         .name = "siftr",
1539         .evhand = siftr_load_handler,
1540 };
1541
1542 /*
1543  * Param 1: name of the kernel module
1544  * Param 2: moduledata_t struct containing info about the kernel module
1545  *          and the execution entry point for the module
1546  * Param 3: From sysinit_sub_id enumeration in /usr/include/sys/kernel.h
1547  *          Defines the module initialisation order
1548  * Param 4: From sysinit_elem_order enumeration in /usr/include/sys/kernel.h
1549  *          Defines the initialisation order of this kld relative to others
1550  *          within the same subsystem as defined by param 3
1551  */
1552 DECLARE_MODULE(siftr, siftr_mod, SI_SUB_SMP, SI_ORDER_ANY);
1553 MODULE_DEPEND(siftr, alq, 1, 1, 1);
1554 MODULE_VERSION(siftr, MODVERSION);