lnet/klnds/o2iblnd/o2iblnd_modparams.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  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 static int ntx = 256;
  56 CFS_MODULE_PARM(ntx, "i", int, 0444,
  57                 "# of message descriptors");
  58
  59 static int credits = 64;
  60 CFS_MODULE_PARM(credits, "i", int, 0444,
  61                 "# concurrent sends");
  62
  63 static int peer_credits = 8;
  64 CFS_MODULE_PARM(peer_credits, "i", int, 0444,
  65                 "# concurrent sends to 1 peer");
  66
  67 static int peer_credits_hiw = 0;
  68 CFS_MODULE_PARM(peer_credits_hiw, "i", int, 0444,
  69                 "when eagerly to return credits");
  70
  71 static int peer_buffer_credits = 0;
  72 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
  73                 "# per-peer router buffer credits");
  74
  75 static int peer_timeout = 180;
  76 CFS_MODULE_PARM(peer_timeout, "i", int, 0444,
  77                 "Seconds without aliveness news to declare peer dead (<=0 to disable)");
  78
  79 static char *ipif_name = "ib0";
  80 CFS_MODULE_PARM(ipif_name, "s", charp, 0444,
  81                 "IPoIB interface name");
  82
  83 static int retry_count = 5;
  84 CFS_MODULE_PARM(retry_count, "i", int, 0644,
  85                 "Retransmissions when no ACK received");
  86
  87 static int rnr_retry_count = 6;
  88 CFS_MODULE_PARM(rnr_retry_count, "i", int, 0644,
  89                 "RNR retransmissions");
  90
  91 static int keepalive = 100;
  92 CFS_MODULE_PARM(keepalive, "i", int, 0644,
  93                 "Idle time in seconds before sending a keepalive");
  94
  95 static int ib_mtu = 0;
  96 CFS_MODULE_PARM(ib_mtu, "i", int, 0444,
  97                 "IB MTU 256/512/1024/2048/4096");
  98
  99 static int concurrent_sends = 0;
 100 CFS_MODULE_PARM(concurrent_sends, "i", int, 0444,
 101                 "send work-queue sizing");
 102
 103 static int map_on_demand = 0;
 104 CFS_MODULE_PARM(map_on_demand, "i", int, 0444,
 105                 "map on demand");
 106
 107 static int fmr_pool_size = 512;
 108 CFS_MODULE_PARM(fmr_pool_size, "i", int, 0444,
 109                 "size of the fmr pool (>= ntx / 4)");
 110
 111 static int fmr_flush_trigger = 384;
 112 CFS_MODULE_PARM(fmr_flush_trigger, "i", int, 0444,
 113                 "# dirty FMRs that triggers pool flush");
 114
 115 static int fmr_cache = 1;
 116 CFS_MODULE_PARM(fmr_cache, "i", int, 0444,
 117                 "non-zero to enable FMR caching");
 118
 119 static int pmr_pool_size = 512;
 120 CFS_MODULE_PARM(pmr_pool_size, "i", int, 0444,
 121                 "size of the MR cache pmr pool");
 122
 123 /*
 124  * 0: disable failover
 125  * 1: enable failover if necessary
 126  * 2: force to failover (for debug)
 127  */
 128 static int dev_failover = 0;
 129 CFS_MODULE_PARM(dev_failover, "i", int, 0444,
 130                "HCA failover for bonding (0 off, 1 on, other values reserved)");
 131
 132
 133 static int require_privileged_port = 0;
 134 CFS_MODULE_PARM(require_privileged_port, "i", int, 0644,
 135                 "require privileged port when accepting connection");
 136
 137 static int use_privileged_port = 1;
 138 CFS_MODULE_PARM(use_privileged_port, "i", int, 0644,
 139                 "use privileged port when initiating connection");
 140
 141 kib_tunables_t kiblnd_tunables = {
 142         .kib_dev_failover           = &dev_failover,
 143         .kib_service                = &service,
 144         .kib_cksum                  = &cksum,
 145         .kib_timeout                = &timeout,
 146         .kib_keepalive              = &keepalive,
 147         .kib_ntx                    = &ntx,
 148         .kib_credits                = &credits,
 149         .kib_peertxcredits          = &peer_credits,
 150         .kib_peercredits_hiw        = &peer_credits_hiw,
 151         .kib_peerrtrcredits         = &peer_buffer_credits,
 152         .kib_peertimeout            = &peer_timeout,
 153         .kib_default_ipif           = &ipif_name,
 154         .kib_retry_count            = &retry_count,
 155         .kib_rnr_retry_count        = &rnr_retry_count,
 156         .kib_concurrent_sends       = &concurrent_sends,
 157         .kib_ib_mtu                 = &ib_mtu,
 158         .kib_map_on_demand          = &map_on_demand,
 159         .kib_fmr_pool_size          = &fmr_pool_size,
 160         .kib_fmr_flush_trigger      = &fmr_flush_trigger,
 161         .kib_fmr_cache              = &fmr_cache,
 162         .kib_pmr_pool_size          = &pmr_pool_size,
 163         .kib_require_priv_port      = &require_privileged_port,
 164         .kib_use_priv_port          = &use_privileged_port
 165 };
 166
 167 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
 168
 169 static char ipif_basename_space[32];
 170
 171 #ifndef HAVE_SYSCTL_UNNUMBERED
 172
 173 enum {
 174         O2IBLND_SERVICE  = 1,
 175         O2IBLND_CKSUM,
 176         O2IBLND_TIMEOUT,
 177         O2IBLND_NTX,
 178         O2IBLND_CREDITS,
 179         O2IBLND_PEER_TXCREDITS,
 180         O2IBLND_PEER_CREDITS_HIW,
 181         O2IBLND_PEER_RTRCREDITS,
 182         O2IBLND_PEER_TIMEOUT,
 183         O2IBLND_IPIF_BASENAME,
 184         O2IBLND_RETRY_COUNT,
 185         O2IBLND_RNR_RETRY_COUNT,
 186         O2IBLND_KEEPALIVE,
 187         O2IBLND_CONCURRENT_SENDS,
 188         O2IBLND_IB_MTU,
 189         O2IBLND_MAP_ON_DEMAND,
 190         O2IBLND_FMR_POOL_SIZE,
 191         O2IBLND_FMR_FLUSH_TRIGGER,
 192         O2IBLND_FMR_CACHE,
 193         O2IBLND_PMR_POOL_SIZE,
 194         O2IBLND_DEV_FAILOVER
 195 };
 196 #else
 197
 198 #define O2IBLND_SERVICE          CTL_UNNUMBERED
 199 #define O2IBLND_CKSUM            CTL_UNNUMBERED
 200 #define O2IBLND_TIMEOUT          CTL_UNNUMBERED
 201 #define O2IBLND_NTX              CTL_UNNUMBERED
 202 #define O2IBLND_CREDITS          CTL_UNNUMBERED
 203 #define O2IBLND_PEER_TXCREDITS   CTL_UNNUMBERED
 204 #define O2IBLND_PEER_CREDITS_HIW CTL_UNNUMBERED
 205 #define O2IBLND_PEER_RTRCREDITS  CTL_UNNUMBERED
 206 #define O2IBLND_PEER_TIMEOUT     CTL_UNNUMBERED
 207 #define O2IBLND_IPIF_BASENAME    CTL_UNNUMBERED
 208 #define O2IBLND_RETRY_COUNT      CTL_UNNUMBERED
 209 #define O2IBLND_RNR_RETRY_COUNT  CTL_UNNUMBERED
 210 #define O2IBLND_KEEPALIVE        CTL_UNNUMBERED
 211 #define O2IBLND_CONCURRENT_SENDS CTL_UNNUMBERED
 212 #define O2IBLND_IB_MTU           CTL_UNNUMBERED
 213 #define O2IBLND_MAP_ON_DEMAND    CTL_UNNUMBERED
 214 #define O2IBLND_FMR_POOL_SIZE    CTL_UNNUMBERED
 215 #define O2IBLND_FMR_FLUSH_TRIGGER CTL_UNNUMBERED
 216 #define O2IBLND_FMR_CACHE        CTL_UNNUMBERED
 217 #define O2IBLND_PMR_POOL_SIZE    CTL_UNNUMBERED
 218 #define O2IBLND_DEV_FAILOVER     CTL_UNNUMBERED
 219
 220 #endif
 221
 222 static cfs_sysctl_table_t kiblnd_ctl_table[] = {
 223         {
 224                 .ctl_name = O2IBLND_SERVICE,
 225                 .procname = "service",
 226                 .data     = &service,
 227                 .maxlen   = sizeof(int),
 228                 .mode     = 0444,
 229                 .proc_handler = &proc_dointvec
 230         },
 231         {
 232                 .ctl_name = O2IBLND_CKSUM,
 233                 .procname = "cksum",
 234                 .data     = &cksum,
 235                 .maxlen   = sizeof(int),
 236                 .mode     = 0644,
 237                 .proc_handler = &proc_dointvec
 238         },
 239         {
 240                 .ctl_name = O2IBLND_TIMEOUT,
 241                 .procname = "timeout",
 242                 .data     = &timeout,
 243                 .maxlen   = sizeof(int),
 244                 .mode     = 0644,
 245                 .proc_handler = &proc_dointvec
 246         },
 247         {
 248                 .ctl_name = O2IBLND_NTX,
 249                 .procname = "ntx",
 250                 .data     = &ntx,
 251                 .maxlen   = sizeof(int),
 252                 .mode     = 0444,
 253                 .proc_handler = &proc_dointvec
 254         },
 255         {
 256                 .ctl_name = O2IBLND_CREDITS,
 257                 .procname = "credits",
 258                 .data     = &credits,
 259                 .maxlen   = sizeof(int),
 260                 .mode     = 0444,
 261                 .proc_handler = &proc_dointvec
 262         },
 263         {
 264                 .ctl_name = O2IBLND_PEER_TXCREDITS,
 265                 .procname = "peer_credits",
 266                 .data     = &peer_credits,
 267                 .maxlen   = sizeof(int),
 268                 .mode     = 0444,
 269                 .proc_handler = &proc_dointvec
 270         },
 271         {
 272                 .ctl_name = O2IBLND_PEER_CREDITS_HIW,
 273                 .procname = "peer_credits_hiw",
 274                 .data     = &peer_credits_hiw,
 275                 .maxlen   = sizeof(int),
 276                 .mode     = 0444,
 277                 .proc_handler = &proc_dointvec
 278         },
 279         {
 280                 .ctl_name = O2IBLND_PEER_RTRCREDITS,
 281                 .procname = "peer_buffer_credits",
 282                 .data     = &peer_buffer_credits,
 283                 .maxlen   = sizeof(int),
 284                 .mode     = 0444,
 285                 .proc_handler = &proc_dointvec
 286         },
 287         {
 288                 .ctl_name = O2IBLND_PEER_TIMEOUT,
 289                 .procname = "peer_timeout",
 290                 .data     = &peer_timeout,
 291                 .maxlen   = sizeof(int),
 292                 .mode     = 0444,
 293                 .proc_handler = &proc_dointvec
 294         },
 295         {
 296                 .ctl_name = O2IBLND_IPIF_BASENAME,
 297                 .procname = "ipif_name",
 298                 .data     = ipif_basename_space,
 299                 .maxlen   = sizeof(ipif_basename_space),
 300                 .mode     = 0444,
 301                 .proc_handler = &proc_dostring
 302         },
 303         {
 304                 .ctl_name = O2IBLND_RETRY_COUNT,
 305                 .procname = "retry_count",
 306                 .data     = &retry_count,
 307                 .maxlen   = sizeof(int),
 308                 .mode     = 0644,
 309                 .proc_handler = &proc_dointvec
 310         },
 311         {
 312                 .ctl_name = O2IBLND_RNR_RETRY_COUNT,
 313                 .procname = "rnr_retry_count",
 314                 .data     = &rnr_retry_count,
 315                 .maxlen   = sizeof(int),
 316                 .mode     = 0644,
 317                 .proc_handler = &proc_dointvec
 318         },
 319         {
 320                 .ctl_name = O2IBLND_KEEPALIVE,
 321                 .procname = "keepalive",
 322                 .data     = &keepalive,
 323                 .maxlen   = sizeof(int),
 324                 .mode     = 0644,
 325                 .proc_handler = &proc_dointvec
 326         },
 327         {
 328                 .ctl_name = O2IBLND_CONCURRENT_SENDS,
 329                 .procname = "concurrent_sends",
 330                 .data     = &concurrent_sends,
 331                 .maxlen   = sizeof(int),
 332                 .mode     = 0444,
 333                 .proc_handler = &proc_dointvec
 334         },
 335         {
 336                 .ctl_name = O2IBLND_IB_MTU,
 337                 .procname = "ib_mtu",
 338                 .data     = &ib_mtu,
 339                 .maxlen   = sizeof(int),
 340                 .mode     = 0444,
 341                 .proc_handler = &proc_dointvec
 342         },
 343         {
 344                 .ctl_name = O2IBLND_MAP_ON_DEMAND,
 345                 .procname = "map_on_demand",
 346                 .data     = &map_on_demand,
 347                 .maxlen   = sizeof(int),
 348                 .mode     = 0444,
 349                 .proc_handler = &proc_dointvec
 350         },
 351
 352         {
 353                 .ctl_name = O2IBLND_FMR_POOL_SIZE,
 354                 .procname = "fmr_pool_size",
 355                 .data     = &fmr_pool_size,
 356                 .maxlen   = sizeof(int),
 357                 .mode     = 0444,
 358                 .proc_handler = &proc_dointvec
 359         },
 360         {
 361                 .ctl_name = O2IBLND_FMR_FLUSH_TRIGGER,
 362                 .procname = "fmr_flush_trigger",
 363                 .data     = &fmr_flush_trigger,
 364                 .maxlen   = sizeof(int),
 365                 .mode     = 0444,
 366                 .proc_handler = &proc_dointvec
 367         },
 368         {
 369                 .ctl_name = O2IBLND_FMR_CACHE,
 370                 .procname = "fmr_cache",
 371                 .data     = &fmr_cache,
 372                 .maxlen   = sizeof(int),
 373                 .mode     = 0444,
 374                 .proc_handler = &proc_dointvec
 375         },
 376         {
 377                 .ctl_name = O2IBLND_PMR_POOL_SIZE,
 378                 .procname = "pmr_pool_size",
 379                 .data     = &pmr_pool_size,
 380                 .maxlen   = sizeof(int),
 381                 .mode     = 0444,
 382                 .proc_handler = &proc_dointvec
 383         },
 384         {
 385                 .ctl_name = O2IBLND_DEV_FAILOVER,
 386                 .procname = "dev_failover",
 387                 .data     = &dev_failover,
 388                 .maxlen   = sizeof(int),
 389                 .mode     = 0444,
 390                 .proc_handler = &proc_dointvec
 391         },
 392         {0}
 393 };
 394
 395 static cfs_sysctl_table_t kiblnd_top_ctl_table[] = {
 396         {
 397                 .ctl_name = CTL_O2IBLND,
 398                 .procname = "o2iblnd",
 399                 .data     = NULL,
 400                 .maxlen   = 0,
 401                 .mode     = 0555,
 402                 .child    = kiblnd_ctl_table
 403         },
 404         {0}
 405 };
 406
 407 void
 408 kiblnd_initstrtunable(char *space, char *str, int size)
 409 {
 410         strncpy(space, str, size);
 411         space[size-1] = 0;
 412 }
 413
 414 void
 415 kiblnd_sysctl_init (void)
 416 {
 417         kiblnd_initstrtunable(ipif_basename_space, ipif_name,
 418                               sizeof(ipif_basename_space));
 419
 420         kiblnd_tunables.kib_sysctl =
 421                 cfs_register_sysctl_table(kiblnd_top_ctl_table, 0);
 422
 423         if (kiblnd_tunables.kib_sysctl == NULL)
 424                 CWARN("Can't setup /proc tunables\n");
 425 }
 426
 427 void
 428 kiblnd_sysctl_fini (void)
 429 {
 430         if (kiblnd_tunables.kib_sysctl != NULL)
 431                 cfs_unregister_sysctl_table(kiblnd_tunables.kib_sysctl);
 432 }
 433
 434 #else
 435
 436 void
 437 kiblnd_sysctl_init (void)
 438 {
 439 }
 440
 441 void
 442 kiblnd_sysctl_fini (void)
 443 {
 444 }
 445
 446 #endif
 447
 448 int
 449 kiblnd_tunables_init (void)
 450 {
 451         if (kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu) < 0) {
 452                 CERROR("Invalid ib_mtu %d, expected 256/512/1024/2048/4096\n",
 453                        *kiblnd_tunables.kib_ib_mtu);
 454                 return -EINVAL;
 455         }
 456
 457         if (*kiblnd_tunables.kib_peertxcredits < IBLND_CREDITS_DEFAULT)
 458                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_DEFAULT;
 459
 460         if (*kiblnd_tunables.kib_peertxcredits > IBLND_CREDITS_MAX)
 461                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_MAX;
 462
 463         if (*kiblnd_tunables.kib_peertxcredits > *kiblnd_tunables.kib_credits)
 464                 *kiblnd_tunables.kib_peertxcredits = *kiblnd_tunables.kib_credits;
 465
 466         if (*kiblnd_tunables.kib_peercredits_hiw < *kiblnd_tunables.kib_peertxcredits / 2)
 467                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits / 2;
 468
 469         if (*kiblnd_tunables.kib_peercredits_hiw >= *kiblnd_tunables.kib_peertxcredits)
 470                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits - 1;
 471
 472         if (*kiblnd_tunables.kib_map_on_demand < 0 ||
 473             *kiblnd_tunables.kib_map_on_demand > IBLND_MAX_RDMA_FRAGS)
 474                 *kiblnd_tunables.kib_map_on_demand = 0; /* disable map-on-demand */
 475
 476         if (*kiblnd_tunables.kib_map_on_demand == 1)
 477                 *kiblnd_tunables.kib_map_on_demand = 2; /* don't make sense to create map if only one fragment */
 478
 479         if (*kiblnd_tunables.kib_concurrent_sends == 0) {
 480                 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
 481                     *kiblnd_tunables.kib_map_on_demand <= IBLND_MAX_RDMA_FRAGS / 8)
 482                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits) * 2;
 483                 else
 484                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits);
 485         }
 486
 487         if (*kiblnd_tunables.kib_concurrent_sends > *kiblnd_tunables.kib_peertxcredits * 2)
 488                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits * 2;
 489
 490         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits / 2)
 491                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits / 2;
 492
 493         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits) {
 494                 CWARN("Concurrent sends %d is lower than message queue size: %d, "
 495                       "performance may drop slightly.\n",
 496                       *kiblnd_tunables.kib_concurrent_sends, *kiblnd_tunables.kib_peertxcredits);
 497         }
 498
 499         kiblnd_sysctl_init();
 500         return 0;
 501 }
 502
 503 void
 504 kiblnd_tunables_fini (void)
 505 {
 506         kiblnd_sysctl_fini();
 507 }