contrib/libpcap/pcap-rdmasniff.c

   1 /*
   2  * Copyright (c) 2017 Pure Storage, Inc.
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  *
   9  * 1. Redistributions of source code must retain the above copyright
  10  * notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  * notice, this list of conditions and the following disclaimer in the
  13  * documentation and/or other materials provided with the distribution.
  14  * 3. The name of the author may not be used to endorse or promote
  15  * products derived from this software without specific prior written
  16  * permission.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  19  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  21  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  22  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  23  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  24  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  28  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  29  */
  30
  31 #ifdef HAVE_CONFIG_H
  32 #include "config.h"
  33 #endif
  34
  35 #include "pcap-int.h"
  36 #include "pcap-rdmasniff.h"
  37
  38 #include <infiniband/verbs.h>
  39 #include <stdlib.h>
  40 #include <string.h>
  41 #include <sys/time.h>
  42
  43 #if !defined(IBV_FLOW_ATTR_SNIFFER)
  44 #define IBV_FLOW_ATTR_SNIFFER   3
  45 #endif
  46
  47 static const int RDMASNIFF_NUM_RECEIVES = 128;
  48 static const int RDMASNIFF_RECEIVE_SIZE = 10000;
  49
  50 struct pcap_rdmasniff {
  51         struct ibv_device *             rdma_device;
  52         struct ibv_context *            context;
  53         struct ibv_comp_channel *       channel;
  54         struct ibv_pd *                 pd;
  55         struct ibv_cq *                 cq;
  56         struct ibv_qp *                 qp;
  57         struct ibv_flow *               flow;
  58         struct ibv_mr *                 mr;
  59         u_char *                        oneshot_buffer;
  60         unsigned                        port_num;
  61         int                             cq_event;
  62         u_int                           packets_recv;
  63 };
  64
  65 static int
  66 rdmasniff_stats(pcap_t *handle, struct pcap_stat *stat)
  67 {
  68         struct pcap_rdmasniff *priv = handle->priv;
  69
  70         stat->ps_recv = priv->packets_recv;
  71         stat->ps_drop = 0;
  72         stat->ps_ifdrop = 0;
  73
  74         return 0;
  75 }
  76
  77 static void
  78 rdmasniff_cleanup(pcap_t *handle)
  79 {
  80         struct pcap_rdmasniff *priv = handle->priv;
  81
  82         ibv_dereg_mr(priv->mr);
  83         ibv_destroy_flow(priv->flow);
  84         ibv_destroy_qp(priv->qp);
  85         ibv_destroy_cq(priv->cq);
  86         ibv_dealloc_pd(priv->pd);
  87         ibv_destroy_comp_channel(priv->channel);
  88         ibv_close_device(priv->context);
  89         free(priv->oneshot_buffer);
  90
  91         pcap_cleanup_live_common(handle);
  92 }
  93
  94 static void
  95 rdmasniff_post_recv(pcap_t *handle, uint64_t wr_id)
  96 {
  97         struct pcap_rdmasniff *priv = handle->priv;
  98         struct ibv_sge sg_entry;
  99         struct ibv_recv_wr wr, *bad_wr;
 100
 101         sg_entry.length = RDMASNIFF_RECEIVE_SIZE;
 102         sg_entry.addr = (uintptr_t) handle->buffer + RDMASNIFF_RECEIVE_SIZE * wr_id;
 103         sg_entry.lkey = priv->mr->lkey;
 104
 105         wr.wr_id = wr_id;
 106         wr.num_sge = 1;
 107         wr.sg_list = &sg_entry;
 108         wr.next = NULL;
 109
 110         ibv_post_recv(priv->qp, &wr, &bad_wr);
 111 }
 112
 113 static int
 114 rdmasniff_read(pcap_t *handle, int max_packets, pcap_handler callback, u_char *user)
 115 {
 116         struct pcap_rdmasniff *priv = handle->priv;
 117         struct ibv_cq *ev_cq;
 118         void *ev_ctx;
 119         struct ibv_wc wc;
 120         struct pcap_pkthdr pkth;
 121         u_char *pktd;
 122         int count = 0;
 123
 124         if (!priv->cq_event) {
 125                 while (ibv_get_cq_event(priv->channel, &ev_cq, &ev_ctx) < 0) {
 126                         if (errno != EINTR) {
 127                                 return PCAP_ERROR;
 128                         }
 129                         if (handle->break_loop) {
 130                                 handle->break_loop = 0;
 131                                 return PCAP_ERROR_BREAK;
 132                         }
 133                 }
 134                 ibv_ack_cq_events(priv->cq, 1);
 135                 ibv_req_notify_cq(priv->cq, 0);
 136                 priv->cq_event = 1;
 137         }
 138
 139         while (count < max_packets || PACKET_COUNT_IS_UNLIMITED(max_packets)) {
 140                 if (ibv_poll_cq(priv->cq, 1, &wc) != 1) {
 141                         priv->cq_event = 0;
 142                         break;
 143                 }
 144
 145                 if (wc.status != IBV_WC_SUCCESS) {
 146                         fprintf(stderr, "failed WC wr_id %lld status %d/%s\n",
 147                                 (unsigned long long) wc.wr_id,
 148                                 wc.status, ibv_wc_status_str(wc.status));
 149                         continue;
 150                 }
 151
 152                 pkth.len = wc.byte_len;
 153                 pkth.caplen = min(pkth.len, (u_int)handle->snapshot);
 154                 gettimeofday(&pkth.ts, NULL);
 155
 156                 pktd = (u_char *) handle->buffer + wc.wr_id * RDMASNIFF_RECEIVE_SIZE;
 157
 158                 if (handle->fcode.bf_insns == NULL ||
 159                     bpf_filter(handle->fcode.bf_insns, pktd, pkth.len, pkth.caplen)) {
 160                         callback(user, &pkth, pktd);
 161                         ++priv->packets_recv;
 162                         ++count;
 163                 }
 164
 165                 rdmasniff_post_recv(handle, wc.wr_id);
 166
 167                 if (handle->break_loop) {
 168                         handle->break_loop = 0;
 169                         return PCAP_ERROR_BREAK;
 170                 }
 171         }
 172
 173         return count;
 174 }
 175
 176 static void
 177 rdmasniff_oneshot(u_char *user, const struct pcap_pkthdr *h, const u_char *bytes)
 178 {
 179         struct oneshot_userdata *sp = (struct oneshot_userdata *) user;
 180         pcap_t *handle = sp->pd;
 181         struct pcap_rdmasniff *priv = handle->priv;
 182
 183         *sp->hdr = *h;
 184         memcpy(priv->oneshot_buffer, bytes, h->caplen);
 185         *sp->pkt = priv->oneshot_buffer;
 186 }
 187
 188 static int
 189 rdmasniff_activate(pcap_t *handle)
 190 {
 191         struct pcap_rdmasniff *priv = handle->priv;
 192         struct ibv_qp_init_attr qp_init_attr;
 193         struct ibv_qp_attr qp_attr;
 194         struct ibv_flow_attr flow_attr;
 195         struct ibv_port_attr port_attr;
 196         int i;
 197
 198         priv->context = ibv_open_device(priv->rdma_device);
 199         if (!priv->context) {
 200                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 201                               "Failed to open device %s", handle->opt.device);
 202                 goto error;
 203         }
 204
 205         priv->pd = ibv_alloc_pd(priv->context);
 206         if (!priv->pd) {
 207                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 208                               "Failed to alloc PD for device %s", handle->opt.device);
 209                 goto error;
 210         }
 211
 212         priv->channel = ibv_create_comp_channel(priv->context);
 213         if (!priv->channel) {
 214                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 215                               "Failed to create comp channel for device %s", handle->opt.device);
 216                 goto error;
 217         }
 218
 219         priv->cq = ibv_create_cq(priv->context, RDMASNIFF_NUM_RECEIVES,
 220                                  NULL, priv->channel, 0);
 221         if (!priv->cq) {
 222                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 223                               "Failed to create CQ for device %s", handle->opt.device);
 224                 goto error;
 225         }
 226
 227         ibv_req_notify_cq(priv->cq, 0);
 228
 229         memset(&qp_init_attr, 0, sizeof qp_init_attr);
 230         qp_init_attr.send_cq = qp_init_attr.recv_cq = priv->cq;
 231         qp_init_attr.cap.max_recv_wr = RDMASNIFF_NUM_RECEIVES;
 232         qp_init_attr.cap.max_recv_sge = 1;
 233         qp_init_attr.qp_type = IBV_QPT_RAW_PACKET;
 234         priv->qp = ibv_create_qp(priv->pd, &qp_init_attr);
 235         if (!priv->qp) {
 236                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 237                               "Failed to create QP for device %s", handle->opt.device);
 238                 goto error;
 239         }
 240
 241         memset(&qp_attr, 0, sizeof qp_attr);
 242         qp_attr.qp_state = IBV_QPS_INIT;
 243         qp_attr.port_num = priv->port_num;
 244         if (ibv_modify_qp(priv->qp, &qp_attr, IBV_QP_STATE | IBV_QP_PORT)) {
 245                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 246                               "Failed to modify QP to INIT for device %s", handle->opt.device);
 247                 goto error;
 248         }
 249
 250         memset(&qp_attr, 0, sizeof qp_attr);
 251         qp_attr.qp_state = IBV_QPS_RTR;
 252         if (ibv_modify_qp(priv->qp, &qp_attr, IBV_QP_STATE)) {
 253                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 254                               "Failed to modify QP to RTR for device %s", handle->opt.device);
 255                 goto error;
 256         }
 257
 258         memset(&flow_attr, 0, sizeof flow_attr);
 259         flow_attr.type = IBV_FLOW_ATTR_SNIFFER;
 260         flow_attr.size = sizeof flow_attr;
 261         flow_attr.port = priv->port_num;
 262         priv->flow = ibv_create_flow(priv->qp, &flow_attr);
 263         if (!priv->flow) {
 264                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 265                               "Failed to create flow for device %s", handle->opt.device);
 266                 goto error;
 267         }
 268
 269         handle->bufsize = RDMASNIFF_NUM_RECEIVES * RDMASNIFF_RECEIVE_SIZE;
 270         handle->buffer = malloc(handle->bufsize);
 271         if (!handle->buffer) {
 272                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 273                               "Failed to allocate receive buffer for device %s", handle->opt.device);
 274                 goto error;
 275         }
 276
 277         priv->oneshot_buffer = malloc(RDMASNIFF_RECEIVE_SIZE);
 278         if (!priv->oneshot_buffer) {
 279                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 280                               "Failed to allocate oneshot buffer for device %s", handle->opt.device);
 281                 goto error;
 282         }
 283
 284         priv->mr = ibv_reg_mr(priv->pd, handle->buffer, handle->bufsize, IBV_ACCESS_LOCAL_WRITE);
 285         if (!priv->mr) {
 286                 pcap_snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
 287                               "Failed to register MR for device %s", handle->opt.device);
 288                 goto error;
 289         }
 290
 291
 292         for (i = 0; i < RDMASNIFF_NUM_RECEIVES; ++i) {
 293                 rdmasniff_post_recv(handle, i);
 294         }
 295
 296         if (!ibv_query_port(priv->context, priv->port_num, &port_attr) &&
 297             port_attr.link_layer == IBV_LINK_LAYER_INFINIBAND) {
 298                 handle->linktype = DLT_INFINIBAND;
 299         } else {
 300                 handle->linktype = DLT_EN10MB;
 301         }
 302
 303         if (handle->snapshot <= 0 || handle->snapshot > RDMASNIFF_RECEIVE_SIZE)
 304                 handle->snapshot = RDMASNIFF_RECEIVE_SIZE;
 305
 306         handle->offset = 0;
 307         handle->read_op = rdmasniff_read;
 308         handle->stats_op = rdmasniff_stats;
 309         handle->cleanup_op = rdmasniff_cleanup;
 310         handle->setfilter_op = install_bpf_program;
 311         handle->setdirection_op = NULL;
 312         handle->set_datalink_op = NULL;
 313         handle->getnonblock_op = pcap_getnonblock_fd;
 314         handle->setnonblock_op = pcap_setnonblock_fd;
 315         handle->oneshot_callback = rdmasniff_oneshot;
 316         handle->selectable_fd = priv->channel->fd;
 317
 318         return 0;
 319
 320 error:
 321         if (priv->mr) {
 322                 ibv_dereg_mr(priv->mr);
 323         }
 324
 325         if (priv->flow) {
 326                 ibv_destroy_flow(priv->flow);
 327         }
 328
 329         if (priv->qp) {
 330                 ibv_destroy_qp(priv->qp);
 331         }
 332
 333         if (priv->cq) {
 334                 ibv_destroy_cq(priv->cq);
 335         }
 336
 337         if (priv->channel) {
 338                 ibv_destroy_comp_channel(priv->channel);
 339         }
 340
 341         if (priv->pd) {
 342                 ibv_dealloc_pd(priv->pd);
 343         }
 344
 345         if (priv->context) {
 346                 ibv_close_device(priv->context);
 347         }
 348
 349         if (priv->oneshot_buffer) {
 350                 free(priv->oneshot_buffer);
 351         }
 352
 353         return PCAP_ERROR;
 354 }
 355
 356 pcap_t *
 357 rdmasniff_create(const char *device, char *ebuf, int *is_ours)
 358 {
 359         struct pcap_rdmasniff *priv;
 360         struct ibv_device **dev_list;
 361         int numdev;
 362         size_t namelen;
 363         const char *port;
 364         unsigned port_num;
 365         int i;
 366         pcap_t *p = NULL;
 367
 368         *is_ours = 0;
 369
 370         dev_list = ibv_get_device_list(&numdev);
 371         if (!dev_list || !numdev) {
 372                 return NULL;
 373         }
 374
 375         namelen = strlen(device);
 376
 377         port = strchr(device, ':');
 378         if (port) {
 379                 port_num = strtoul(port + 1, NULL, 10);
 380                 if (port_num > 0) {
 381                         namelen = port - device;
 382                 } else {
 383                         port_num = 1;
 384                 }
 385         } else {
 386                 port_num = 1;
 387         }
 388
 389         for (i = 0; i < numdev; ++i) {
 390                 if (strlen(dev_list[i]->name) == namelen &&
 391                     !strncmp(device, dev_list[i]->name, namelen)) {
 392                         *is_ours = 1;
 393
 394                         p = pcap_create_common(ebuf, sizeof (struct pcap_rdmasniff));
 395                         if (p) {
 396                                 p->activate_op = rdmasniff_activate;
 397                                 priv = p->priv;
 398                                 priv->rdma_device = dev_list[i];
 399                                 priv->port_num = port_num;
 400                         }
 401                         break;
 402                 }
 403         }
 404
 405         ibv_free_device_list(dev_list);
 406         return p;
 407 }
 408
 409 int
 410 rdmasniff_findalldevs(pcap_if_list_t *devlistp, char *err_str)
 411 {
 412         struct ibv_device **dev_list;
 413         int numdev;
 414         int i;
 415         int ret = 0;
 416
 417         dev_list = ibv_get_device_list(&numdev);
 418         if (!dev_list || !numdev) {
 419                 return 0;
 420         }
 421
 422         for (i = 0; i < numdev; ++i) {
 423                 /*
 424                  * XXX - do the notions of "up", "running", or
 425                  * "connected" apply here?
 426                  */
 427                 if (!add_dev(devlistp, dev_list[i]->name, 0, "RDMA sniffer", err_str)) {
 428                         ret = -1;
 429                         goto out;
 430                 }
 431         }
 432
 433 out:
 434         ibv_free_device_list(dev_list);
 435         return ret;
 436 }