librdmacm-ssa.patch

   1 diff -Nur librdmacm-1.0.19/examples/rping.c librdmacm-1.0.19.ssa1/examples/rping.c
   2 --- librdmacm-1.0.19/examples/rping.c   2014-07-17 00:30:15.000000000 +0200
   3 +++ librdmacm-1.0.19.ssa1/examples/rping.c      2015-01-22 18:11:28.000000000 +0100
   4 @@ -277,15 +277,20 @@
   5         struct ibv_wc wc;
   6         struct ibv_recv_wr *bad_wr;
   7         int ret;
   8 +       int flushed = 0;
   9
  10         while ((ret = ibv_poll_cq(cb->cq, 1, &wc)) == 1) {
  11                 ret = 0;
  12
  13                 if (wc.status) {
  14 -                       if (wc.status != IBV_WC_WR_FLUSH_ERR)
  15 -                               fprintf(stderr,
  16 -                                       "cq completion failed status %d\n",
  17 -                                       wc.status);
  18 +                       if (wc.status == IBV_WC_WR_FLUSH_ERR) {
  19 +                               flushed = 1;
  20 +                               continue;
  21 +
  22 +                       }
  23 +                       fprintf(stderr,
  24 +                               "cq completion failed status %d\n",
  25 +                               wc.status);
  26                         ret = -1;
  27                         goto error;
  28                 }
  29 @@ -334,7 +339,7 @@
  30                 fprintf(stderr, "poll error %d\n", ret);
  31                 goto error;
  32         }
  33 -       return 0;
  34 +       return flushed;
  35
  36  error:
  37         cb->state = ERROR;
  38 @@ -1055,18 +1060,19 @@
  39         ret = rping_connect_client(cb);
  40         if (ret) {
  41                 fprintf(stderr, "connect error %d\n", ret);
  42 -               goto err2;
  43 +               goto err3;
  44         }
  45
  46         ret = rping_test_client(cb);
  47         if (ret) {
  48                 fprintf(stderr, "rping client failed: %d\n", ret);
  49 -               goto err3;
  50 +               goto err4;
  51         }
  52
  53         ret = 0;
  54 -err3:
  55 +err4:
  56         rdma_disconnect(cb->cm_id);
  57 +err3:
  58         pthread_join(cb->cqthread, NULL);
  59  err2:
  60         rping_free_buffers(cb);
  61 diff -Nur librdmacm-1.0.19/src/cma.c librdmacm-1.0.19.ssa1/src/cma.c
  62 --- librdmacm-1.0.19/src/cma.c  2014-07-17 00:30:15.000000000 +0200
  63 +++ librdmacm-1.0.19.ssa1/src/cma.c     2015-01-22 19:46:41.000000000 +0100
  64 @@ -49,6 +49,7 @@
  65  #include <stddef.h>
  66  #include <netdb.h>
  67  #include <syslog.h>
  68 +#include <limits.h>
  69
  70  #include "cma.h"
  71  #include "indexer.h"
  72 @@ -1035,7 +1036,9 @@
  73  static int ucma_modify_qp_rtr(struct rdma_cm_id *id, uint8_t resp_res)
  74  {
  75         struct ibv_qp_attr qp_attr;
  76 +       struct ibv_port_attr port_attr;
  77         int qp_attr_mask, ret;
  78 +       uint8_t link_layer;
  79
  80         if (!id->qp)
  81                 return ERR(EINVAL);
  82 @@ -1055,6 +1058,17 @@
  83         if (ret)
  84                 return ret;
  85
  86 +       /* Workaround for rdma_ucm kernel bug */
  87 +       ret = ibv_query_port(id->verbs, id->port_num, &port_attr);
  88 +       if (ret)
  89 +               link_layer = IBV_LINK_LAYER_UNSPECIFIED;
  90 +       else
  91 +               link_layer = port_attr.link_layer;
  92 +
  93 +       if (id->verbs->device->transport_type == IBV_TRANSPORT_IB &&
  94 +           link_layer == IBV_LINK_LAYER_INFINIBAND)
  95 +               qp_attr_mask &= UINT_MAX ^ 0xe00000;    /* mask off bits 21-24 which are used for RoCE */
  96 +
  97         if (resp_res != RDMA_MAX_RESP_RES)
  98                 qp_attr.max_dest_rd_atomic = resp_res;
  99         return rdma_seterrno(ibv_modify_qp(id->qp, &qp_attr, qp_attr_mask));
 100 diff -Nur librdmacm-1.0.19/src/preload.c librdmacm-1.0.19.ssa1/src/preload.c
 101 --- librdmacm-1.0.19/src/preload.c      2014-07-17 00:30:15.000000000 +0200
 102 +++ librdmacm-1.0.19.ssa1/src/preload.c 2015-01-22 18:11:29.000000000 +0100
 103 @@ -50,6 +50,9 @@
 104  #include <netinet/tcp.h>
 105  #include <unistd.h>
 106  #include <semaphore.h>
 107 +#include <ctype.h>
 108 +#include <stdlib.h>
 109 +#include <stdio.h>
 110
 111  #include <rdma/rdma_cma.h>
 112  #include <rdma/rdma_verbs.h>
 113 @@ -122,6 +125,136 @@
 114         atomic_t refcnt;
 115  };
 116
 117 +struct config_entry {
 118 +       char *name;
 119 +       int domain;
 120 +       int type;
 121 +       int protocol;
 122 +};
 123 +
 124 +static struct config_entry *config;
 125 +static int config_cnt;
 126 +extern char *program_invocation_short_name;
 127 +
 128 +
 129 +static void free_config(void)
 130 +{
 131 +       while (config_cnt)
 132 +               free(config[--config_cnt].name);
 133 +
 134 +       free(config);
 135 +}
 136 +
 137 +/*
 138 + * Config file format:
 139 + * # Starting '#' indicates comment
 140 + * # wild card values are supported using '*'
 141 + * # domain - *, INET, INET6, IB
 142 + * # type - *, STREAM, DGRAM
 143 + * # protocol - *, TCP, UDP
 144 + * program_name domain type protocol
 145 + */
 146 +static void scan_config(void)
 147 +{
 148 +       struct config_entry *new_config;
 149 +       FILE *fp;
 150 +       char line[120], prog[64], dom[16], type[16], proto[16];
 151 +
 152 +       fp = fopen(RS_CONF_DIR "/preload_config", "r");
 153 +       if (!fp)
 154 +               return;
 155 +
 156 +       while (fgets(line, sizeof(line), fp)) {
 157 +               if (line[0] == '#')
 158 +                       continue;
 159 +
 160 +               if (sscanf(line, "%64s%16s%16s%16s", prog, dom, type, proto) != 4)
 161 +                       continue;
 162 +
 163 +               new_config = realloc(config, (config_cnt + 1) *
 164 +                                            sizeof(struct config_entry));
 165 +               if (!new_config)
 166 +                       break;
 167 +
 168 +               config = new_config;
 169 +               memset(&config[config_cnt], 0, sizeof(struct config_entry));
 170 +
 171 +               if (!strcasecmp(dom, "INET") ||
 172 +                   !strcasecmp(dom, "AF_INET") ||
 173 +                   !strcasecmp(dom, "PF_INET")) {
 174 +                       config[config_cnt].domain = AF_INET;
 175 +               } else if (!strcasecmp(dom, "INET6") ||
 176 +                          !strcasecmp(dom, "AF_INET6") ||
 177 +                          !strcasecmp(dom, "PF_INET6")) {
 178 +                       config[config_cnt].domain = AF_INET6;
 179 +               } else if (!strcasecmp(dom, "IB") ||
 180 +                          !strcasecmp(dom, "AF_IB") ||
 181 +                          !strcasecmp(dom, "PF_IB")) {
 182 +                       config[config_cnt].domain = AF_IB;
 183 +               } else if (strcmp(dom, "*")) {
 184 +                       continue;
 185 +               }
 186 +
 187 +               if (!strcasecmp(type, "STREAM") ||
 188 +                   !strcasecmp(type, "SOCK_STREAM")) {
 189 +                       config[config_cnt].type = SOCK_STREAM;
 190 +               } else if (!strcasecmp(type, "DGRAM") ||
 191 +                          !strcasecmp(type, "SOCK_DGRAM")) {
 192 +                       config[config_cnt].type = SOCK_DGRAM;
 193 +               } else if (strcmp(type, "*")) {
 194 +                       continue;
 195 +               }
 196 +
 197 +               if (!strcasecmp(proto, "TCP") ||
 198 +                   !strcasecmp(proto, "IPPROTO_TCP")) {
 199 +                       config[config_cnt].protocol = IPPROTO_TCP;
 200 +               } else if (!strcasecmp(proto, "UDP") ||
 201 +                          !strcasecmp(proto, "IPPROTO_UDP")) {
 202 +                       config[config_cnt].protocol = IPPROTO_UDP;
 203 +               } else if (strcmp(proto, "*")) {
 204 +                       continue;
 205 +               }
 206 +
 207 +               if (strcmp(prog, "*")) {
 208 +                   if (!(config[config_cnt].name = strdup(prog)))
 209 +                           continue;
 210 +               }
 211 +
 212 +               config_cnt++;
 213 +       }
 214 +
 215 +       fclose(fp);
 216 +       if (config_cnt)
 217 +               atexit(free_config);
 218 +}
 219 +
 220 +static int intercept_socket(int domain, int type, int protocol)
 221 +{
 222 +       int i;
 223 +
 224 +       if (!config_cnt)
 225 +               return 1;
 226 +
 227 +       if (!protocol) {
 228 +               if (type == SOCK_STREAM)
 229 +                       protocol = IPPROTO_TCP;
 230 +               else if (type == SOCK_DGRAM)
 231 +                       protocol = IPPROTO_UDP;
 232 +       }
 233 +
 234 +       for (i = 0; i < config_cnt; i++) {
 235 +               if ((!config[i].name ||
 236 +                    !strncasecmp(config[i].name, program_invocation_short_name,
 237 +                                 strlen(config[i].name))) &&
 238 +                   (!config[i].domain || config[i].domain == domain) &&
 239 +                   (!config[i].type || config[i].type == type) &&
 240 +                   (!config[i].protocol || config[i].protocol == protocol))
 241 +                       return 1;
 242 +       }
 243 +
 244 +       return 0;
 245 +}
 246 +
 247  static int fd_open(void)
 248  {
 249         struct fd_info *fdi;
 250 @@ -308,6 +441,7 @@
 251         rs.fcntl = dlsym(RTLD_DEFAULT, "rfcntl");
 252
 253         getenv_options();
 254 +       scan_config();
 255         init = 1;
 256  out:
 257         pthread_mutex_unlock(&mut);
 258 @@ -404,10 +538,11 @@
 259         static __thread int recursive;
 260         int index, ret;
 261
 262 -       if (recursive)
 263 +       init_preload();
 264 +
 265 +       if (recursive || !intercept_socket(domain, type, protocol))
 266                 goto real;
 267
 268 -       init_preload();
 269         index = fd_open();
 270         if (index < 0)
 271                 return index;
 272 diff -Nur librdmacm-1.0.19/src/rsocket.c librdmacm-1.0.19.ssa1/src/rsocket.c
 273 --- librdmacm-1.0.19/src/rsocket.c      2014-07-17 00:30:15.000000000 +0200
 274 +++ librdmacm-1.0.19.ssa1/src/rsocket.c 2015-01-22 18:11:29.000000000 +0100
 275 @@ -1,5 +1,5 @@
 276  /*
 277 - * Copyright (c) 2008-2013 Intel Corporation.  All rights reserved.
 278 + * Copyright (c) 2008-2014 Intel Corporation.  All rights reserved.
 279   *
 280   * This software is available to you under a choice of one of two
 281   * licenses.  You may choose to be licensed under the terms of the GNU
 282 @@ -381,6 +381,7 @@
 283         dlist_entry       iomap_list;
 284         dlist_entry       iomap_queue;
 285         int               iomap_pending;
 286 +       int               unack_cqe;
 287  };
 288
 289  #define DS_UDP_TAG 0x55555555
 290 @@ -967,9 +968,6 @@
 291                 return;
 292         }
 293
 294 -       if (rs->index >= 0)
 295 -               rs_remove(rs);
 296 -
 297         if (rs->rmsg)
 298                 free(rs->rmsg);
 299
 300 @@ -993,11 +991,16 @@
 301
 302         if (rs->cm_id) {
 303                 rs_free_iomappings(rs);
 304 -               if (rs->cm_id->qp)
 305 +               if (rs->cm_id->qp) {
 306 +                       ibv_ack_cq_events(rs->cm_id->recv_cq, rs->unack_cqe);
 307                         rdma_destroy_qp(rs->cm_id);
 308 +               }
 309                 rdma_destroy_id(rs->cm_id);
 310         }
 311
 312 +       if (rs->index >= 0)
 313 +               rs_remove(rs);
 314 +
 315         fastlock_destroy(&rs->map_lock);
 316         fastlock_destroy(&rs->cq_wait_lock);
 317         fastlock_destroy(&rs->cq_lock);
 318 @@ -1174,9 +1177,14 @@
 319         rs = idm_lookup(&idm, socket);
 320         if (!rs)
 321                 return ERR(EBADF);
 322 -       ret = rdma_listen(rs->cm_id, backlog);
 323 -       if (!ret)
 324 -               rs->state = rs_listening;
 325 +
 326 +       if (rs->state != rs_listening) {
 327 +               ret = rdma_listen(rs->cm_id, backlog);
 328 +               if (!ret)
 329 +                       rs->state = rs_listening;
 330 +       } else {
 331 +               ret = 0;
 332 +       }
 333         return ret;
 334  }
 335
 336 @@ -1965,9 +1973,12 @@
 337
 338         ret = ibv_get_cq_event(rs->cm_id->recv_cq_channel, &cq, &context);
 339         if (!ret) {
 340 -               ibv_ack_cq_events(rs->cm_id->recv_cq, 1);
 341 +               if (++rs->unack_cqe >= rs->sq_size + rs->rq_size) {
 342 +                       ibv_ack_cq_events(rs->cm_id->recv_cq, rs->unack_cqe);
 343 +                       rs->unack_cqe = 0;
 344 +               }
 345                 rs->cq_armed = 0;
 346 -       } else if (errno != EAGAIN) {
 347 +       } else if (!(errno == EAGAIN || errno == EINTR)) {
 348                 rs->state = rs_error;
 349         }
 350
 351 @@ -2383,7 +2394,7 @@
 352         struct rsocket *rs;
 353         size_t left = len;
 354         uint32_t end_size, rsize;
 355 -       int ret;
 356 +       int ret = 0;
 357
 358         rs = idm_at(&idm, socket);
 359         if (rs->type == SOCK_DGRAM) {
 360 @@ -2410,7 +2421,6 @@
 361                                 break;
 362                 }
 363
 364 -               ret = 0;
 365                 if (flags & MSG_PEEK) {
 366                         left = len - rs_peek(rs, buf, left);
 367                         break;
 368 @@ -2445,7 +2455,7 @@
 369         } while (left && (flags & MSG_WAITALL) && (rs->state & rs_readable));
 370
 371         fastlock_release(&rs->rlock);
 372 -       return ret ? ret : len - left;
 373 +       return (ret && left == len) ? ret : len - left;
 374  }
 375
 376  ssize_t rrecvfrom(int socket, void *buf, size_t len, int flags,