lnet/klnds/o2iblnd/o2iblnd_modparams.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd_modparams.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include "o2iblnd.h"
  42
  43 static int service = 987;
  44 CFS_MODULE_PARM(service, "i", int, 0444,
  45                 "service number (within RDMA_PS_TCP)");
  46
  47 static int cksum = 0;
  48 CFS_MODULE_PARM(cksum, "i", int, 0644,
  49                 "set non-zero to enable message (not RDMA) checksums");
  50
  51 static int timeout = 50;
  52 CFS_MODULE_PARM(timeout, "i", int, 0644,
  53                 "timeout (seconds)");
  54
  55 /* Number of threads in each scheduler pool which is percpt,
  56  * we will estimate reasonable value based on CPUs if it's set to zero. */
  57 static int nscheds;
  58 CFS_MODULE_PARM(nscheds, "i", int, 0444,
  59                 "number of threads in each scheduler pool");
  60
  61 /* NB: this value is shared by all CPTs, it can grow at runtime */
  62 static int ntx = 512;
  63 CFS_MODULE_PARM(ntx, "i", int, 0444,
  64                 "# of message descriptors allocated for each pool");
  65
  66 /* NB: this value is shared by all CPTs */
  67 static int credits = 256;
  68 CFS_MODULE_PARM(credits, "i", int, 0444,
  69                 "# concurrent sends");
  70
  71 static int peer_credits = 8;
  72 CFS_MODULE_PARM(peer_credits, "i", int, 0444,
  73                 "# concurrent sends to 1 peer");
  74
  75 static int peer_credits_hiw = 0;
  76 CFS_MODULE_PARM(peer_credits_hiw, "i", int, 0444,
  77                 "when eagerly to return credits");
  78
  79 static int peer_buffer_credits = 0;
  80 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
  81                 "# per-peer router buffer credits");
  82
  83 static int peer_timeout = 180;
  84 CFS_MODULE_PARM(peer_timeout, "i", int, 0444,
  85                 "Seconds without aliveness news to declare peer dead (<=0 to disable)");
  86
  87 static char *ipif_name = "ib0";
  88 CFS_MODULE_PARM(ipif_name, "s", charp, 0444,
  89                 "IPoIB interface name");
  90
  91 static int retry_count = 5;
  92 CFS_MODULE_PARM(retry_count, "i", int, 0644,
  93                 "Retransmissions when no ACK received");
  94
  95 static int rnr_retry_count = 6;
  96 CFS_MODULE_PARM(rnr_retry_count, "i", int, 0644,
  97                 "RNR retransmissions");
  98
  99 static int keepalive = 100;
 100 CFS_MODULE_PARM(keepalive, "i", int, 0644,
 101                 "Idle time in seconds before sending a keepalive");
 102
 103 static int ib_mtu = 0;
 104 CFS_MODULE_PARM(ib_mtu, "i", int, 0444,
 105                 "IB MTU 256/512/1024/2048/4096");
 106
 107 static int concurrent_sends = 0;
 108 CFS_MODULE_PARM(concurrent_sends, "i", int, 0444,
 109                 "send work-queue sizing");
 110
 111 static int map_on_demand = 0;
 112 CFS_MODULE_PARM(map_on_demand, "i", int, 0444,
 113                 "map on demand");
 114
 115 /* NB: this value is shared by all CPTs, it can grow at runtime */
 116 static int fmr_pool_size = 512;
 117 CFS_MODULE_PARM(fmr_pool_size, "i", int, 0444,
 118                 "size of fmr pool on each CPT (>= ntx / 4)");
 119
 120 /* NB: this value is shared by all CPTs, it can grow at runtime */
 121 static int fmr_flush_trigger = 384;
 122 CFS_MODULE_PARM(fmr_flush_trigger, "i", int, 0444,
 123                 "# dirty FMRs that triggers pool flush");
 124
 125 static int fmr_cache = 1;
 126 CFS_MODULE_PARM(fmr_cache, "i", int, 0444,
 127                 "non-zero to enable FMR caching");
 128
 129 /* NB: this value is shared by all CPTs, it can grow at runtime */
 130 static int pmr_pool_size = 512;
 131 CFS_MODULE_PARM(pmr_pool_size, "i", int, 0444,
 132                 "size of MR cache pmr pool on each CPT");
 133
 134 /*
 135  * 0: disable failover
 136  * 1: enable failover if necessary
 137  * 2: force to failover (for debug)
 138  */
 139 static int dev_failover = 0;
 140 CFS_MODULE_PARM(dev_failover, "i", int, 0444,
 141                "HCA failover for bonding (0 off, 1 on, other values reserved)");
 142
 143
 144 static int require_privileged_port = 0;
 145 CFS_MODULE_PARM(require_privileged_port, "i", int, 0644,
 146                 "require privileged port when accepting connection");
 147
 148 static int use_privileged_port = 1;
 149 CFS_MODULE_PARM(use_privileged_port, "i", int, 0644,
 150                 "use privileged port when initiating connection");
 151
 152 kib_tunables_t kiblnd_tunables = {
 153         .kib_dev_failover           = &dev_failover,
 154         .kib_service                = &service,
 155         .kib_cksum                  = &cksum,
 156         .kib_timeout                = &timeout,
 157         .kib_keepalive              = &keepalive,
 158         .kib_ntx                    = &ntx,
 159         .kib_credits                = &credits,
 160         .kib_peertxcredits          = &peer_credits,
 161         .kib_peercredits_hiw        = &peer_credits_hiw,
 162         .kib_peerrtrcredits         = &peer_buffer_credits,
 163         .kib_peertimeout            = &peer_timeout,
 164         .kib_default_ipif           = &ipif_name,
 165         .kib_retry_count            = &retry_count,
 166         .kib_rnr_retry_count        = &rnr_retry_count,
 167         .kib_concurrent_sends       = &concurrent_sends,
 168         .kib_ib_mtu                 = &ib_mtu,
 169         .kib_map_on_demand          = &map_on_demand,
 170         .kib_fmr_pool_size          = &fmr_pool_size,
 171         .kib_fmr_flush_trigger      = &fmr_flush_trigger,
 172         .kib_fmr_cache              = &fmr_cache,
 173         .kib_pmr_pool_size          = &pmr_pool_size,
 174         .kib_require_priv_port      = &require_privileged_port,
 175         .kib_use_priv_port          = &use_privileged_port,
 176         .kib_nscheds                = &nscheds
 177 };
 178
 179 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
 180
 181 static char ipif_basename_space[32];
 182
 183 static struct ctl_table kiblnd_ctl_table[] = {
 184         {
 185                 INIT_CTL_NAME
 186                 .procname       = "service",
 187                 .data           = &service,
 188                 .maxlen         = sizeof(int),
 189                 .mode           = 0444,
 190                 .proc_handler   = &proc_dointvec
 191         },
 192         {
 193                 INIT_CTL_NAME
 194                 .procname       = "cksum",
 195                 .data           = &cksum,
 196                 .maxlen         = sizeof(int),
 197                 .mode           = 0644,
 198                 .proc_handler   = &proc_dointvec
 199         },
 200         {
 201                 INIT_CTL_NAME
 202                 .procname       = "timeout",
 203                 .data           = &timeout,
 204                 .maxlen         = sizeof(int),
 205                 .mode           = 0644,
 206                 .proc_handler   = &proc_dointvec
 207         },
 208         {
 209                 INIT_CTL_NAME
 210                 .procname       = "ntx",
 211                 .data           = &ntx,
 212                 .maxlen         = sizeof(int),
 213                 .mode           = 0444,
 214                 .proc_handler   = &proc_dointvec
 215         },
 216         {
 217                 INIT_CTL_NAME
 218                 .procname       = "credits",
 219                 .data           = &credits,
 220                 .maxlen         = sizeof(int),
 221                 .mode           = 0444,
 222                 .proc_handler   = &proc_dointvec
 223         },
 224         {
 225                 INIT_CTL_NAME
 226                 .procname       = "peer_credits",
 227                 .data           = &peer_credits,
 228                 .maxlen         = sizeof(int),
 229                 .mode           = 0444,
 230                 .proc_handler   = &proc_dointvec
 231         },
 232         {
 233                 INIT_CTL_NAME
 234                 .procname       = "peer_credits_hiw",
 235                 .data           = &peer_credits_hiw,
 236                 .maxlen         = sizeof(int),
 237                 .mode           = 0444,
 238                 .proc_handler   = &proc_dointvec
 239         },
 240         {
 241                 INIT_CTL_NAME
 242                 .procname       = "peer_buffer_credits",
 243                 .data           = &peer_buffer_credits,
 244                 .maxlen         = sizeof(int),
 245                 .mode           = 0444,
 246                 .proc_handler   = &proc_dointvec
 247         },
 248         {
 249                 INIT_CTL_NAME
 250                 .procname       = "peer_timeout",
 251                 .data           = &peer_timeout,
 252                 .maxlen         = sizeof(int),
 253                 .mode           = 0444,
 254                 .proc_handler   = &proc_dointvec
 255         },
 256         {
 257                 INIT_CTL_NAME
 258                 .procname       = "ipif_name",
 259                 .data           = ipif_basename_space,
 260                 .maxlen         = sizeof(ipif_basename_space),
 261                 .mode           = 0444,
 262                 .proc_handler   = &proc_dostring
 263         },
 264         {
 265                 INIT_CTL_NAME
 266                 .procname       = "retry_count",
 267                 .data           = &retry_count,
 268                 .maxlen         = sizeof(int),
 269                 .mode           = 0644,
 270                 .proc_handler   = &proc_dointvec
 271         },
 272         {
 273                 INIT_CTL_NAME
 274                 .procname       = "rnr_retry_count",
 275                 .data           = &rnr_retry_count,
 276                 .maxlen         = sizeof(int),
 277                 .mode           = 0644,
 278                 .proc_handler   = &proc_dointvec
 279         },
 280         {
 281                 INIT_CTL_NAME
 282                 .procname       = "keepalive",
 283                 .data           = &keepalive,
 284                 .maxlen         = sizeof(int),
 285                 .mode           = 0644,
 286                 .proc_handler   = &proc_dointvec
 287         },
 288         {
 289                 INIT_CTL_NAME
 290                 .procname       = "concurrent_sends",
 291                 .data           = &concurrent_sends,
 292                 .maxlen         = sizeof(int),
 293                 .mode           = 0444,
 294                 .proc_handler   = &proc_dointvec
 295         },
 296         {
 297                 INIT_CTL_NAME
 298                 .procname       = "ib_mtu",
 299                 .data           = &ib_mtu,
 300                 .maxlen         = sizeof(int),
 301                 .mode           = 0444,
 302                 .proc_handler   = &proc_dointvec
 303         },
 304         {
 305                 INIT_CTL_NAME
 306                 .procname       = "map_on_demand",
 307                 .data           = &map_on_demand,
 308                 .maxlen         = sizeof(int),
 309                 .mode           = 0444,
 310                 .proc_handler   = &proc_dointvec
 311         },
 312         {
 313                 INIT_CTL_NAME
 314                 .procname       = "fmr_pool_size",
 315                 .data           = &fmr_pool_size,
 316                 .maxlen         = sizeof(int),
 317                 .mode           = 0444,
 318                 .proc_handler   = &proc_dointvec
 319         },
 320         {
 321                 INIT_CTL_NAME
 322                 .procname       = "fmr_flush_trigger",
 323                 .data           = &fmr_flush_trigger,
 324                 .maxlen         = sizeof(int),
 325                 .mode           = 0444,
 326                 .proc_handler   = &proc_dointvec
 327         },
 328         {
 329                 INIT_CTL_NAME
 330                 .procname       = "fmr_cache",
 331                 .data           = &fmr_cache,
 332                 .maxlen         = sizeof(int),
 333                 .mode           = 0444,
 334                 .proc_handler   = &proc_dointvec
 335         },
 336         {
 337                 INIT_CTL_NAME
 338                 .procname       = "pmr_pool_size",
 339                 .data           = &pmr_pool_size,
 340                 .maxlen         = sizeof(int),
 341                 .mode           = 0444,
 342                 .proc_handler   = &proc_dointvec
 343         },
 344         {
 345                 INIT_CTL_NAME
 346                 .procname       = "dev_failover",
 347                 .data           = &dev_failover,
 348                 .maxlen         = sizeof(int),
 349                 .mode           = 0444,
 350                 .proc_handler   = &proc_dointvec
 351         },
 352         { 0 }
 353 };
 354
 355 static struct ctl_table kiblnd_top_ctl_table[] = {
 356         {
 357                 INIT_CTL_NAME
 358                 .procname       = "o2iblnd",
 359                 .data           = NULL,
 360                 .maxlen         = 0,
 361                 .mode           = 0555,
 362                 .child          = kiblnd_ctl_table
 363         },
 364         { 0 }
 365 };
 366
 367 void
 368 kiblnd_initstrtunable(char *space, char *str, int size)
 369 {
 370         strncpy(space, str, size);
 371         space[size-1] = 0;
 372 }
 373
 374 static void
 375 kiblnd_sysctl_init (void)
 376 {
 377         kiblnd_initstrtunable(ipif_basename_space, ipif_name,
 378                               sizeof(ipif_basename_space));
 379
 380         kiblnd_tunables.kib_sysctl =
 381                 register_sysctl_table(kiblnd_top_ctl_table);
 382
 383         if (kiblnd_tunables.kib_sysctl == NULL)
 384                 CWARN("Can't setup /proc tunables\n");
 385 }
 386
 387 static void
 388 kiblnd_sysctl_fini (void)
 389 {
 390         if (kiblnd_tunables.kib_sysctl != NULL)
 391                 unregister_sysctl_table(kiblnd_tunables.kib_sysctl);
 392 }
 393
 394 #else
 395
 396 static void
 397 kiblnd_sysctl_init (void)
 398 {
 399 }
 400
 401 static void
 402 kiblnd_sysctl_fini (void)
 403 {
 404 }
 405
 406 #endif
 407
 408 int
 409 kiblnd_tunables_init (void)
 410 {
 411         if (kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu) < 0) {
 412                 CERROR("Invalid ib_mtu %d, expected 256/512/1024/2048/4096\n",
 413                        *kiblnd_tunables.kib_ib_mtu);
 414                 return -EINVAL;
 415         }
 416
 417         if (*kiblnd_tunables.kib_peertxcredits < IBLND_CREDITS_DEFAULT)
 418                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_DEFAULT;
 419
 420         if (*kiblnd_tunables.kib_peertxcredits > IBLND_CREDITS_MAX)
 421                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_MAX;
 422
 423         if (*kiblnd_tunables.kib_peertxcredits > *kiblnd_tunables.kib_credits)
 424                 *kiblnd_tunables.kib_peertxcredits = *kiblnd_tunables.kib_credits;
 425
 426         if (*kiblnd_tunables.kib_peercredits_hiw < *kiblnd_tunables.kib_peertxcredits / 2)
 427                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits / 2;
 428
 429         if (*kiblnd_tunables.kib_peercredits_hiw >= *kiblnd_tunables.kib_peertxcredits)
 430                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits - 1;
 431
 432         if (*kiblnd_tunables.kib_map_on_demand < 0 ||
 433             *kiblnd_tunables.kib_map_on_demand > IBLND_MAX_RDMA_FRAGS)
 434                 *kiblnd_tunables.kib_map_on_demand = 0; /* disable map-on-demand */
 435
 436         if (*kiblnd_tunables.kib_map_on_demand == 1)
 437                 *kiblnd_tunables.kib_map_on_demand = 2; /* don't make sense to create map if only one fragment */
 438
 439         if (*kiblnd_tunables.kib_concurrent_sends == 0) {
 440                 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
 441                     *kiblnd_tunables.kib_map_on_demand <= IBLND_MAX_RDMA_FRAGS / 8)
 442                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits) * 2;
 443                 else
 444                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits);
 445         }
 446
 447         if (*kiblnd_tunables.kib_concurrent_sends > *kiblnd_tunables.kib_peertxcredits * 2)
 448                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits * 2;
 449
 450         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits / 2)
 451                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits / 2;
 452
 453         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits) {
 454                 CWARN("Concurrent sends %d is lower than message queue size: %d, "
 455                       "performance may drop slightly.\n",
 456                       *kiblnd_tunables.kib_concurrent_sends, *kiblnd_tunables.kib_peertxcredits);
 457         }
 458
 459         kiblnd_sysctl_init();
 460         return 0;
 461 }
 462
 463 void
 464 kiblnd_tunables_fini (void)
 465 {
 466         kiblnd_sysctl_fini();
 467 }