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 kib_tunables_t kiblnd_tunables = {
 133         .kib_dev_failover           = &dev_failover,
 134         .kib_service                = &service,
 135         .kib_cksum                  = &cksum,
 136         .kib_timeout                = &timeout,
 137         .kib_keepalive              = &keepalive,
 138         .kib_ntx                    = &ntx,
 139         .kib_credits                = &credits,
 140         .kib_peertxcredits          = &peer_credits,
 141         .kib_peercredits_hiw        = &peer_credits_hiw,
 142         .kib_peerrtrcredits         = &peer_buffer_credits,
 143         .kib_peertimeout            = &peer_timeout,
 144         .kib_default_ipif           = &ipif_name,
 145         .kib_retry_count            = &retry_count,
 146         .kib_rnr_retry_count        = &rnr_retry_count,
 147         .kib_concurrent_sends       = &concurrent_sends,
 148         .kib_ib_mtu                 = &ib_mtu,
 149         .kib_map_on_demand          = &map_on_demand,
 150         .kib_fmr_pool_size          = &fmr_pool_size,
 151         .kib_fmr_flush_trigger      = &fmr_flush_trigger,
 152         .kib_fmr_cache              = &fmr_cache,
 153         .kib_pmr_pool_size          = &pmr_pool_size,
 154 };
 155
 156 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
 157
 158 static char ipif_basename_space[32];
 159
 160 #ifndef HAVE_SYSCTL_UNNUMBERED
 161
 162 enum {
 163         O2IBLND_SERVICE  = 1,
 164         O2IBLND_CKSUM,
 165         O2IBLND_TIMEOUT,
 166         O2IBLND_NTX,
 167         O2IBLND_CREDITS,
 168         O2IBLND_PEER_TXCREDITS,
 169         O2IBLND_PEER_CREDITS_HIW,
 170         O2IBLND_PEER_RTRCREDITS,
 171         O2IBLND_PEER_TIMEOUT,
 172         O2IBLND_IPIF_BASENAME,
 173         O2IBLND_RETRY_COUNT,
 174         O2IBLND_RNR_RETRY_COUNT,
 175         O2IBLND_KEEPALIVE,
 176         O2IBLND_CONCURRENT_SENDS,
 177         O2IBLND_IB_MTU,
 178         O2IBLND_MAP_ON_DEMAND,
 179         O2IBLND_FMR_POOL_SIZE,
 180         O2IBLND_FMR_FLUSH_TRIGGER,
 181         O2IBLND_FMR_CACHE,
 182         O2IBLND_PMR_POOL_SIZE,
 183         O2IBLND_DEV_FAILOVER
 184 };
 185 #else
 186
 187 #define O2IBLND_SERVICE          CTL_UNNUMBERED
 188 #define O2IBLND_CKSUM            CTL_UNNUMBERED
 189 #define O2IBLND_TIMEOUT          CTL_UNNUMBERED
 190 #define O2IBLND_NTX              CTL_UNNUMBERED
 191 #define O2IBLND_CREDITS          CTL_UNNUMBERED
 192 #define O2IBLND_PEER_TXCREDITS   CTL_UNNUMBERED
 193 #define O2IBLND_PEER_CREDITS_HIW CTL_UNNUMBERED
 194 #define O2IBLND_PEER_RTRCREDITS  CTL_UNNUMBERED
 195 #define O2IBLND_PEER_TIMEOUT     CTL_UNNUMBERED
 196 #define O2IBLND_IPIF_BASENAME    CTL_UNNUMBERED
 197 #define O2IBLND_RETRY_COUNT      CTL_UNNUMBERED
 198 #define O2IBLND_RNR_RETRY_COUNT  CTL_UNNUMBERED
 199 #define O2IBLND_KEEPALIVE        CTL_UNNUMBERED
 200 #define O2IBLND_CONCURRENT_SENDS CTL_UNNUMBERED
 201 #define O2IBLND_IB_MTU           CTL_UNNUMBERED
 202 #define O2IBLND_MAP_ON_DEMAND    CTL_UNNUMBERED
 203 #define O2IBLND_FMR_POOL_SIZE    CTL_UNNUMBERED
 204 #define O2IBLND_FMR_FLUSH_TRIGGER CTL_UNNUMBERED
 205 #define O2IBLND_FMR_CACHE        CTL_UNNUMBERED
 206 #define O2IBLND_PMR_POOL_SIZE    CTL_UNNUMBERED
 207 #define O2IBLND_DEV_FAILOVER     CTL_UNNUMBERED
 208
 209 #endif
 210
 211 static cfs_sysctl_table_t kiblnd_ctl_table[] = {
 212         {
 213                 .ctl_name = O2IBLND_SERVICE,
 214                 .procname = "service",
 215                 .data     = &service,
 216                 .maxlen   = sizeof(int),
 217                 .mode     = 0444,
 218                 .proc_handler = &proc_dointvec
 219         },
 220         {
 221                 .ctl_name = O2IBLND_CKSUM,
 222                 .procname = "cksum",
 223                 .data     = &cksum,
 224                 .maxlen   = sizeof(int),
 225                 .mode     = 0644,
 226                 .proc_handler = &proc_dointvec
 227         },
 228         {
 229                 .ctl_name = O2IBLND_TIMEOUT,
 230                 .procname = "timeout",
 231                 .data     = &timeout,
 232                 .maxlen   = sizeof(int),
 233                 .mode     = 0644,
 234                 .proc_handler = &proc_dointvec
 235         },
 236         {
 237                 .ctl_name = O2IBLND_NTX,
 238                 .procname = "ntx",
 239                 .data     = &ntx,
 240                 .maxlen   = sizeof(int),
 241                 .mode     = 0444,
 242                 .proc_handler = &proc_dointvec
 243         },
 244         {
 245                 .ctl_name = O2IBLND_CREDITS,
 246                 .procname = "credits",
 247                 .data     = &credits,
 248                 .maxlen   = sizeof(int),
 249                 .mode     = 0444,
 250                 .proc_handler = &proc_dointvec
 251         },
 252         {
 253                 .ctl_name = O2IBLND_PEER_TXCREDITS,
 254                 .procname = "peer_credits",
 255                 .data     = &peer_credits,
 256                 .maxlen   = sizeof(int),
 257                 .mode     = 0444,
 258                 .proc_handler = &proc_dointvec
 259         },
 260         {
 261                 .ctl_name = O2IBLND_PEER_CREDITS_HIW,
 262                 .procname = "peer_credits_hiw",
 263                 .data     = &peer_credits_hiw,
 264                 .maxlen   = sizeof(int),
 265                 .mode     = 0444,
 266                 .proc_handler = &proc_dointvec
 267         },
 268         {
 269                 .ctl_name = O2IBLND_PEER_RTRCREDITS,
 270                 .procname = "peer_buffer_credits",
 271                 .data     = &peer_buffer_credits,
 272                 .maxlen   = sizeof(int),
 273                 .mode     = 0444,
 274                 .proc_handler = &proc_dointvec
 275         },
 276         {
 277                 .ctl_name = O2IBLND_PEER_TIMEOUT,
 278                 .procname = "peer_timeout",
 279                 .data     = &peer_timeout,
 280                 .maxlen   = sizeof(int),
 281                 .mode     = 0444,
 282                 .proc_handler = &proc_dointvec
 283         },
 284         {
 285                 .ctl_name = O2IBLND_IPIF_BASENAME,
 286                 .procname = "ipif_name",
 287                 .data     = ipif_basename_space,
 288                 .maxlen   = sizeof(ipif_basename_space),
 289                 .mode     = 0444,
 290                 .proc_handler = &proc_dostring
 291         },
 292         {
 293                 .ctl_name = O2IBLND_RETRY_COUNT,
 294                 .procname = "retry_count",
 295                 .data     = &retry_count,
 296                 .maxlen   = sizeof(int),
 297                 .mode     = 0644,
 298                 .proc_handler = &proc_dointvec
 299         },
 300         {
 301                 .ctl_name = O2IBLND_RNR_RETRY_COUNT,
 302                 .procname = "rnr_retry_count",
 303                 .data     = &rnr_retry_count,
 304                 .maxlen   = sizeof(int),
 305                 .mode     = 0644,
 306                 .proc_handler = &proc_dointvec
 307         },
 308         {
 309                 .ctl_name = O2IBLND_KEEPALIVE,
 310                 .procname = "keepalive",
 311                 .data     = &keepalive,
 312                 .maxlen   = sizeof(int),
 313                 .mode     = 0644,
 314                 .proc_handler = &proc_dointvec
 315         },
 316         {
 317                 .ctl_name = O2IBLND_CONCURRENT_SENDS,
 318                 .procname = "concurrent_sends",
 319                 .data     = &concurrent_sends,
 320                 .maxlen   = sizeof(int),
 321                 .mode     = 0444,
 322                 .proc_handler = &proc_dointvec
 323         },
 324         {
 325                 .ctl_name = O2IBLND_IB_MTU,
 326                 .procname = "ib_mtu",
 327                 .data     = &ib_mtu,
 328                 .maxlen   = sizeof(int),
 329                 .mode     = 0444,
 330                 .proc_handler = &proc_dointvec
 331         },
 332         {
 333                 .ctl_name = O2IBLND_MAP_ON_DEMAND,
 334                 .procname = "map_on_demand",
 335                 .data     = &map_on_demand,
 336                 .maxlen   = sizeof(int),
 337                 .mode     = 0444,
 338                 .proc_handler = &proc_dointvec
 339         },
 340
 341         {
 342                 .ctl_name = O2IBLND_FMR_POOL_SIZE,
 343                 .procname = "fmr_pool_size",
 344                 .data     = &fmr_pool_size,
 345                 .maxlen   = sizeof(int),
 346                 .mode     = 0444,
 347                 .proc_handler = &proc_dointvec
 348         },
 349         {
 350                 .ctl_name = O2IBLND_FMR_FLUSH_TRIGGER,
 351                 .procname = "fmr_flush_trigger",
 352                 .data     = &fmr_flush_trigger,
 353                 .maxlen   = sizeof(int),
 354                 .mode     = 0444,
 355                 .proc_handler = &proc_dointvec
 356         },
 357         {
 358                 .ctl_name = O2IBLND_FMR_CACHE,
 359                 .procname = "fmr_cache",
 360                 .data     = &fmr_cache,
 361                 .maxlen   = sizeof(int),
 362                 .mode     = 0444,
 363                 .proc_handler = &proc_dointvec
 364         },
 365         {
 366                 .ctl_name = O2IBLND_PMR_POOL_SIZE,
 367                 .procname = "pmr_pool_size",
 368                 .data     = &pmr_pool_size,
 369                 .maxlen   = sizeof(int),
 370                 .mode     = 0444,
 371                 .proc_handler = &proc_dointvec
 372         },
 373         {
 374                 .ctl_name = O2IBLND_DEV_FAILOVER,
 375                 .procname = "dev_failover",
 376                 .data     = &dev_failover,
 377                 .maxlen   = sizeof(int),
 378                 .mode     = 0444,
 379                 .proc_handler = &proc_dointvec
 380         },
 381         {0}
 382 };
 383
 384 static cfs_sysctl_table_t kiblnd_top_ctl_table[] = {
 385         {
 386                 .ctl_name = CTL_O2IBLND,
 387                 .procname = "o2iblnd",
 388                 .data     = NULL,
 389                 .maxlen   = 0,
 390                 .mode     = 0555,
 391                 .child    = kiblnd_ctl_table
 392         },
 393         {0}
 394 };
 395
 396 void
 397 kiblnd_initstrtunable(char *space, char *str, int size)
 398 {
 399         strncpy(space, str, size);
 400         space[size-1] = 0;
 401 }
 402
 403 void
 404 kiblnd_sysctl_init (void)
 405 {
 406         kiblnd_initstrtunable(ipif_basename_space, ipif_name,
 407                               sizeof(ipif_basename_space));
 408
 409         kiblnd_tunables.kib_sysctl =
 410                 cfs_register_sysctl_table(kiblnd_top_ctl_table, 0);
 411
 412         if (kiblnd_tunables.kib_sysctl == NULL)
 413                 CWARN("Can't setup /proc tunables\n");
 414 }
 415
 416 void
 417 kiblnd_sysctl_fini (void)
 418 {
 419         if (kiblnd_tunables.kib_sysctl != NULL)
 420                 cfs_unregister_sysctl_table(kiblnd_tunables.kib_sysctl);
 421 }
 422
 423 #else
 424
 425 void
 426 kiblnd_sysctl_init (void)
 427 {
 428 }
 429
 430 void
 431 kiblnd_sysctl_fini (void)
 432 {
 433 }
 434
 435 #endif
 436
 437 int
 438 kiblnd_tunables_init (void)
 439 {
 440         if (kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu) < 0) {
 441                 CERROR("Invalid ib_mtu %d, expected 256/512/1024/2048/4096\n",
 442                        *kiblnd_tunables.kib_ib_mtu);
 443                 return -EINVAL;
 444         }
 445
 446         if (*kiblnd_tunables.kib_peertxcredits < IBLND_CREDITS_DEFAULT)
 447                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_DEFAULT;
 448
 449         if (*kiblnd_tunables.kib_peertxcredits > IBLND_CREDITS_MAX)
 450                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_MAX;
 451
 452         if (*kiblnd_tunables.kib_peertxcredits > *kiblnd_tunables.kib_credits)
 453                 *kiblnd_tunables.kib_peertxcredits = *kiblnd_tunables.kib_credits;
 454
 455         if (*kiblnd_tunables.kib_peercredits_hiw < *kiblnd_tunables.kib_peertxcredits / 2)
 456                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits / 2;
 457
 458         if (*kiblnd_tunables.kib_peercredits_hiw >= *kiblnd_tunables.kib_peertxcredits)
 459                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits - 1;
 460
 461         if (*kiblnd_tunables.kib_map_on_demand < 0 ||
 462             *kiblnd_tunables.kib_map_on_demand > IBLND_MAX_RDMA_FRAGS)
 463                 *kiblnd_tunables.kib_map_on_demand = 0; /* disable map-on-demand */
 464
 465         if (*kiblnd_tunables.kib_map_on_demand == 1)
 466                 *kiblnd_tunables.kib_map_on_demand = 2; /* don't make sense to create map if only one fragment */
 467
 468         if (*kiblnd_tunables.kib_concurrent_sends == 0) {
 469                 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
 470                     *kiblnd_tunables.kib_map_on_demand <= IBLND_MAX_RDMA_FRAGS / 8)
 471                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits) * 2;
 472                 else
 473                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits);
 474         }
 475
 476         if (*kiblnd_tunables.kib_concurrent_sends > *kiblnd_tunables.kib_peertxcredits * 2)
 477                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits * 2;
 478
 479         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits / 2)
 480                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits / 2;
 481
 482         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits) {
 483                 CWARN("Concurrent sends %d is lower than message queue size: %d, "
 484                       "performance may drop slightly.\n",
 485                       *kiblnd_tunables.kib_concurrent_sends, *kiblnd_tunables.kib_peertxcredits);
 486         }
 487
 488         kiblnd_sysctl_init();
 489         return 0;
 490 }
 491
 492 void
 493 kiblnd_tunables_fini (void)
 494 {
 495         kiblnd_sysctl_fini();
 496 }