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 /*
  31  * This file is part of Lustre, http://www.lustre.org/
  32  * Lustre is a trademark of Sun Microsystems, Inc.
  33  *
  34  * lnet/klnds/o2iblnd/o2iblnd_modparams.c
  35  *
  36  * Author: Eric Barton <eric@bartonsoftware.com>
  37  */
  38
  39 #include "o2iblnd.h"
  40
  41 static int service = 987;
  42 CFS_MODULE_PARM(service, "i", int, 0444,
  43                 "service number (within RDMA_PS_TCP)");
  44
  45 static int cksum = 0;
  46 CFS_MODULE_PARM(cksum, "i", int, 0644,
  47                 "set non-zero to enable message (not RDMA) checksums");
  48
  49 static int timeout = 50;
  50 CFS_MODULE_PARM(timeout, "i", int, 0644,
  51                 "timeout (seconds)");
  52
  53 /* Number of threads in each scheduler pool which is percpt,
  54  * we will estimate reasonable value based on CPUs if it's set to zero. */
  55 static int nscheds;
  56 CFS_MODULE_PARM(nscheds, "i", int, 0444,
  57                 "number of threads in each scheduler pool");
  58
  59 /* NB: this value is shared by all CPTs, it can grow at runtime */
  60 static int ntx = 512;
  61 CFS_MODULE_PARM(ntx, "i", int, 0444,
  62                 "# of message descriptors allocated for each pool");
  63
  64 /* NB: this value is shared by all CPTs */
  65 static int credits = 256;
  66 CFS_MODULE_PARM(credits, "i", int, 0444,
  67                 "# concurrent sends");
  68
  69 static int peer_credits = 8;
  70 CFS_MODULE_PARM(peer_credits, "i", int, 0444,
  71                 "# concurrent sends to 1 peer");
  72
  73 static int peer_credits_hiw = 0;
  74 CFS_MODULE_PARM(peer_credits_hiw, "i", int, 0444,
  75                 "when eagerly to return credits");
  76
  77 static int peer_buffer_credits = 0;
  78 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
  79                 "# per-peer router buffer credits");
  80
  81 static int peer_timeout = 180;
  82 CFS_MODULE_PARM(peer_timeout, "i", int, 0444,
  83                 "Seconds without aliveness news to declare peer dead (<=0 to disable)");
  84
  85 static char *ipif_name = "ib0";
  86 CFS_MODULE_PARM(ipif_name, "s", charp, 0444,
  87                 "IPoIB interface name");
  88
  89 static int retry_count = 5;
  90 CFS_MODULE_PARM(retry_count, "i", int, 0644,
  91                 "Retransmissions when no ACK received");
  92
  93 static int rnr_retry_count = 6;
  94 CFS_MODULE_PARM(rnr_retry_count, "i", int, 0644,
  95                 "RNR retransmissions");
  96
  97 static int keepalive = 100;
  98 CFS_MODULE_PARM(keepalive, "i", int, 0644,
  99                 "Idle time in seconds before sending a keepalive");
 100
 101 static int ib_mtu = 0;
 102 CFS_MODULE_PARM(ib_mtu, "i", int, 0444,
 103                 "IB MTU 256/512/1024/2048/4096");
 104
 105 static int concurrent_sends = 0;
 106 CFS_MODULE_PARM(concurrent_sends, "i", int, 0444,
 107                 "send work-queue sizing");
 108
 109 static int map_on_demand = 0;
 110 CFS_MODULE_PARM(map_on_demand, "i", int, 0444,
 111                 "map on demand");
 112
 113 /* NB: this value is shared by all CPTs, it can grow at runtime */
 114 static int fmr_pool_size = 512;
 115 CFS_MODULE_PARM(fmr_pool_size, "i", int, 0444,
 116                 "size of fmr pool on each CPT (>= ntx / 4)");
 117
 118 /* NB: this value is shared by all CPTs, it can grow at runtime */
 119 static int fmr_flush_trigger = 384;
 120 CFS_MODULE_PARM(fmr_flush_trigger, "i", int, 0444,
 121                 "# dirty FMRs that triggers pool flush");
 122
 123 static int fmr_cache = 1;
 124 CFS_MODULE_PARM(fmr_cache, "i", int, 0444,
 125                 "non-zero to enable FMR caching");
 126
 127 /* NB: this value is shared by all CPTs, it can grow at runtime */
 128 static int pmr_pool_size = 512;
 129 CFS_MODULE_PARM(pmr_pool_size, "i", int, 0444,
 130                 "size of MR cache pmr pool on each CPT");
 131
 132 /*
 133  * 0: disable failover
 134  * 1: enable failover if necessary
 135  * 2: force to failover (for debug)
 136  */
 137 static int dev_failover = 0;
 138 CFS_MODULE_PARM(dev_failover, "i", int, 0444,
 139                "HCA failover for bonding (0 off, 1 on, other values reserved)");
 140
 141
 142 static int require_privileged_port = 0;
 143 CFS_MODULE_PARM(require_privileged_port, "i", int, 0644,
 144                 "require privileged port when accepting connection");
 145
 146 static int use_privileged_port = 1;
 147 CFS_MODULE_PARM(use_privileged_port, "i", int, 0644,
 148                 "use privileged port when initiating connection");
 149
 150 kib_tunables_t kiblnd_tunables = {
 151         .kib_dev_failover           = &dev_failover,
 152         .kib_service                = &service,
 153         .kib_cksum                  = &cksum,
 154         .kib_timeout                = &timeout,
 155         .kib_keepalive              = &keepalive,
 156         .kib_ntx                    = &ntx,
 157         .kib_credits                = &credits,
 158         .kib_peertxcredits          = &peer_credits,
 159         .kib_peercredits_hiw        = &peer_credits_hiw,
 160         .kib_peerrtrcredits         = &peer_buffer_credits,
 161         .kib_peertimeout            = &peer_timeout,
 162         .kib_default_ipif           = &ipif_name,
 163         .kib_retry_count            = &retry_count,
 164         .kib_rnr_retry_count        = &rnr_retry_count,
 165         .kib_concurrent_sends       = &concurrent_sends,
 166         .kib_ib_mtu                 = &ib_mtu,
 167         .kib_map_on_demand          = &map_on_demand,
 168         .kib_fmr_pool_size          = &fmr_pool_size,
 169         .kib_fmr_flush_trigger      = &fmr_flush_trigger,
 170         .kib_fmr_cache              = &fmr_cache,
 171         .kib_pmr_pool_size          = &pmr_pool_size,
 172         .kib_require_priv_port      = &require_privileged_port,
 173         .kib_use_priv_port          = &use_privileged_port,
 174         .kib_nscheds                = &nscheds
 175 };
 176
 177 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
 178
 179 static char ipif_basename_space[32];
 180
 181 #ifndef HAVE_SYSCTL_UNNUMBERED
 182
 183 enum {
 184         O2IBLND_SERVICE  = 1,
 185         O2IBLND_CKSUM,
 186         O2IBLND_TIMEOUT,
 187         O2IBLND_NTX,
 188         O2IBLND_CREDITS,
 189         O2IBLND_PEER_TXCREDITS,
 190         O2IBLND_PEER_CREDITS_HIW,
 191         O2IBLND_PEER_RTRCREDITS,
 192         O2IBLND_PEER_TIMEOUT,
 193         O2IBLND_IPIF_BASENAME,
 194         O2IBLND_RETRY_COUNT,
 195         O2IBLND_RNR_RETRY_COUNT,
 196         O2IBLND_KEEPALIVE,
 197         O2IBLND_CONCURRENT_SENDS,
 198         O2IBLND_IB_MTU,
 199         O2IBLND_MAP_ON_DEMAND,
 200         O2IBLND_FMR_POOL_SIZE,
 201         O2IBLND_FMR_FLUSH_TRIGGER,
 202         O2IBLND_FMR_CACHE,
 203         O2IBLND_PMR_POOL_SIZE,
 204         O2IBLND_DEV_FAILOVER
 205 };
 206 #else
 207
 208 #define O2IBLND_SERVICE          CTL_UNNUMBERED
 209 #define O2IBLND_CKSUM            CTL_UNNUMBERED
 210 #define O2IBLND_TIMEOUT          CTL_UNNUMBERED
 211 #define O2IBLND_NTX              CTL_UNNUMBERED
 212 #define O2IBLND_CREDITS          CTL_UNNUMBERED
 213 #define O2IBLND_PEER_TXCREDITS   CTL_UNNUMBERED
 214 #define O2IBLND_PEER_CREDITS_HIW CTL_UNNUMBERED
 215 #define O2IBLND_PEER_RTRCREDITS  CTL_UNNUMBERED
 216 #define O2IBLND_PEER_TIMEOUT     CTL_UNNUMBERED
 217 #define O2IBLND_IPIF_BASENAME    CTL_UNNUMBERED
 218 #define O2IBLND_RETRY_COUNT      CTL_UNNUMBERED
 219 #define O2IBLND_RNR_RETRY_COUNT  CTL_UNNUMBERED
 220 #define O2IBLND_KEEPALIVE        CTL_UNNUMBERED
 221 #define O2IBLND_CONCURRENT_SENDS CTL_UNNUMBERED
 222 #define O2IBLND_IB_MTU           CTL_UNNUMBERED
 223 #define O2IBLND_MAP_ON_DEMAND    CTL_UNNUMBERED
 224 #define O2IBLND_FMR_POOL_SIZE    CTL_UNNUMBERED
 225 #define O2IBLND_FMR_FLUSH_TRIGGER CTL_UNNUMBERED
 226 #define O2IBLND_FMR_CACHE        CTL_UNNUMBERED
 227 #define O2IBLND_PMR_POOL_SIZE    CTL_UNNUMBERED
 228 #define O2IBLND_DEV_FAILOVER     CTL_UNNUMBERED
 229
 230 #endif
 231
 232 static cfs_sysctl_table_t kiblnd_ctl_table[] = {
 233         {
 234                 .ctl_name = O2IBLND_SERVICE,
 235                 .procname = "service",
 236                 .data     = &service,
 237                 .maxlen   = sizeof(int),
 238                 .mode     = 0444,
 239                 .proc_handler = &proc_dointvec
 240         },
 241         {
 242                 .ctl_name = O2IBLND_CKSUM,
 243                 .procname = "cksum",
 244                 .data     = &cksum,
 245                 .maxlen   = sizeof(int),
 246                 .mode     = 0644,
 247                 .proc_handler = &proc_dointvec
 248         },
 249         {
 250                 .ctl_name = O2IBLND_TIMEOUT,
 251                 .procname = "timeout",
 252                 .data     = &timeout,
 253                 .maxlen   = sizeof(int),
 254                 .mode     = 0644,
 255                 .proc_handler = &proc_dointvec
 256         },
 257         {
 258                 .ctl_name = O2IBLND_NTX,
 259                 .procname = "ntx",
 260                 .data     = &ntx,
 261                 .maxlen   = sizeof(int),
 262                 .mode     = 0444,
 263                 .proc_handler = &proc_dointvec
 264         },
 265         {
 266                 .ctl_name = O2IBLND_CREDITS,
 267                 .procname = "credits",
 268                 .data     = &credits,
 269                 .maxlen   = sizeof(int),
 270                 .mode     = 0444,
 271                 .proc_handler = &proc_dointvec
 272         },
 273         {
 274                 .ctl_name = O2IBLND_PEER_TXCREDITS,
 275                 .procname = "peer_credits",
 276                 .data     = &peer_credits,
 277                 .maxlen   = sizeof(int),
 278                 .mode     = 0444,
 279                 .proc_handler = &proc_dointvec
 280         },
 281         {
 282                 .ctl_name = O2IBLND_PEER_CREDITS_HIW,
 283                 .procname = "peer_credits_hiw",
 284                 .data     = &peer_credits_hiw,
 285                 .maxlen   = sizeof(int),
 286                 .mode     = 0444,
 287                 .proc_handler = &proc_dointvec
 288         },
 289         {
 290                 .ctl_name = O2IBLND_PEER_RTRCREDITS,
 291                 .procname = "peer_buffer_credits",
 292                 .data     = &peer_buffer_credits,
 293                 .maxlen   = sizeof(int),
 294                 .mode     = 0444,
 295                 .proc_handler = &proc_dointvec
 296         },
 297         {
 298                 .ctl_name = O2IBLND_PEER_TIMEOUT,
 299                 .procname = "peer_timeout",
 300                 .data     = &peer_timeout,
 301                 .maxlen   = sizeof(int),
 302                 .mode     = 0444,
 303                 .proc_handler = &proc_dointvec
 304         },
 305         {
 306                 .ctl_name = O2IBLND_IPIF_BASENAME,
 307                 .procname = "ipif_name",
 308                 .data     = ipif_basename_space,
 309                 .maxlen   = sizeof(ipif_basename_space),
 310                 .mode     = 0444,
 311                 .proc_handler = &proc_dostring
 312         },
 313         {
 314                 .ctl_name = O2IBLND_RETRY_COUNT,
 315                 .procname = "retry_count",
 316                 .data     = &retry_count,
 317                 .maxlen   = sizeof(int),
 318                 .mode     = 0644,
 319                 .proc_handler = &proc_dointvec
 320         },
 321         {
 322                 .ctl_name = O2IBLND_RNR_RETRY_COUNT,
 323                 .procname = "rnr_retry_count",
 324                 .data     = &rnr_retry_count,
 325                 .maxlen   = sizeof(int),
 326                 .mode     = 0644,
 327                 .proc_handler = &proc_dointvec
 328         },
 329         {
 330                 .ctl_name = O2IBLND_KEEPALIVE,
 331                 .procname = "keepalive",
 332                 .data     = &keepalive,
 333                 .maxlen   = sizeof(int),
 334                 .mode     = 0644,
 335                 .proc_handler = &proc_dointvec
 336         },
 337         {
 338                 .ctl_name = O2IBLND_CONCURRENT_SENDS,
 339                 .procname = "concurrent_sends",
 340                 .data     = &concurrent_sends,
 341                 .maxlen   = sizeof(int),
 342                 .mode     = 0444,
 343                 .proc_handler = &proc_dointvec
 344         },
 345         {
 346                 .ctl_name = O2IBLND_IB_MTU,
 347                 .procname = "ib_mtu",
 348                 .data     = &ib_mtu,
 349                 .maxlen   = sizeof(int),
 350                 .mode     = 0444,
 351                 .proc_handler = &proc_dointvec
 352         },
 353         {
 354                 .ctl_name = O2IBLND_MAP_ON_DEMAND,
 355                 .procname = "map_on_demand",
 356                 .data     = &map_on_demand,
 357                 .maxlen   = sizeof(int),
 358                 .mode     = 0444,
 359                 .proc_handler = &proc_dointvec
 360         },
 361
 362         {
 363                 .ctl_name = O2IBLND_FMR_POOL_SIZE,
 364                 .procname = "fmr_pool_size",
 365                 .data     = &fmr_pool_size,
 366                 .maxlen   = sizeof(int),
 367                 .mode     = 0444,
 368                 .proc_handler = &proc_dointvec
 369         },
 370         {
 371                 .ctl_name = O2IBLND_FMR_FLUSH_TRIGGER,
 372                 .procname = "fmr_flush_trigger",
 373                 .data     = &fmr_flush_trigger,
 374                 .maxlen   = sizeof(int),
 375                 .mode     = 0444,
 376                 .proc_handler = &proc_dointvec
 377         },
 378         {
 379                 .ctl_name = O2IBLND_FMR_CACHE,
 380                 .procname = "fmr_cache",
 381                 .data     = &fmr_cache,
 382                 .maxlen   = sizeof(int),
 383                 .mode     = 0444,
 384                 .proc_handler = &proc_dointvec
 385         },
 386         {
 387                 .ctl_name = O2IBLND_PMR_POOL_SIZE,
 388                 .procname = "pmr_pool_size",
 389                 .data     = &pmr_pool_size,
 390                 .maxlen   = sizeof(int),
 391                 .mode     = 0444,
 392                 .proc_handler = &proc_dointvec
 393         },
 394         {
 395                 .ctl_name = O2IBLND_DEV_FAILOVER,
 396                 .procname = "dev_failover",
 397                 .data     = &dev_failover,
 398                 .maxlen   = sizeof(int),
 399                 .mode     = 0444,
 400                 .proc_handler = &proc_dointvec
 401         },
 402         {0}
 403 };
 404
 405 static cfs_sysctl_table_t kiblnd_top_ctl_table[] = {
 406         {
 407                 .ctl_name = CTL_O2IBLND,
 408                 .procname = "o2iblnd",
 409                 .data     = NULL,
 410                 .maxlen   = 0,
 411                 .mode     = 0555,
 412                 .child    = kiblnd_ctl_table
 413         },
 414         {0}
 415 };
 416
 417 void
 418 kiblnd_initstrtunable(char *space, char *str, int size)
 419 {
 420         strncpy(space, str, size);
 421         space[size-1] = 0;
 422 }
 423
 424 void
 425 kiblnd_sysctl_init (void)
 426 {
 427         kiblnd_initstrtunable(ipif_basename_space, ipif_name,
 428                               sizeof(ipif_basename_space));
 429
 430         kiblnd_tunables.kib_sysctl =
 431                 cfs_register_sysctl_table(kiblnd_top_ctl_table, 0);
 432
 433         if (kiblnd_tunables.kib_sysctl == NULL)
 434                 CWARN("Can't setup /proc tunables\n");
 435 }
 436
 437 void
 438 kiblnd_sysctl_fini (void)
 439 {
 440         if (kiblnd_tunables.kib_sysctl != NULL)
 441                 cfs_unregister_sysctl_table(kiblnd_tunables.kib_sysctl);
 442 }
 443
 444 #else
 445
 446 void
 447 kiblnd_sysctl_init (void)
 448 {
 449 }
 450
 451 void
 452 kiblnd_sysctl_fini (void)
 453 {
 454 }
 455
 456 #endif
 457
 458 int
 459 kiblnd_tunables_init (void)
 460 {
 461         if (kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu) < 0) {
 462                 CERROR("Invalid ib_mtu %d, expected 256/512/1024/2048/4096\n",
 463                        *kiblnd_tunables.kib_ib_mtu);
 464                 return -EINVAL;
 465         }
 466
 467         if (*kiblnd_tunables.kib_peertxcredits < IBLND_CREDITS_DEFAULT)
 468                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_DEFAULT;
 469
 470         if (*kiblnd_tunables.kib_peertxcredits > IBLND_CREDITS_MAX)
 471                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_MAX;
 472
 473         if (*kiblnd_tunables.kib_peertxcredits > *kiblnd_tunables.kib_credits)
 474                 *kiblnd_tunables.kib_peertxcredits = *kiblnd_tunables.kib_credits;
 475
 476         if (*kiblnd_tunables.kib_peercredits_hiw < *kiblnd_tunables.kib_peertxcredits / 2)
 477                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits / 2;
 478
 479         if (*kiblnd_tunables.kib_peercredits_hiw >= *kiblnd_tunables.kib_peertxcredits)
 480                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits - 1;
 481
 482         if (*kiblnd_tunables.kib_map_on_demand < 0 ||
 483             *kiblnd_tunables.kib_map_on_demand > IBLND_MAX_RDMA_FRAGS)
 484                 *kiblnd_tunables.kib_map_on_demand = 0; /* disable map-on-demand */
 485
 486         if (*kiblnd_tunables.kib_map_on_demand == 1)
 487                 *kiblnd_tunables.kib_map_on_demand = 2; /* don't make sense to create map if only one fragment */
 488
 489         if (*kiblnd_tunables.kib_concurrent_sends == 0) {
 490                 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
 491                     *kiblnd_tunables.kib_map_on_demand <= IBLND_MAX_RDMA_FRAGS / 8)
 492                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits) * 2;
 493                 else
 494                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits);
 495         }
 496
 497         if (*kiblnd_tunables.kib_concurrent_sends > *kiblnd_tunables.kib_peertxcredits * 2)
 498                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits * 2;
 499
 500         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits / 2)
 501                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits / 2;
 502
 503         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits) {
 504                 CWARN("Concurrent sends %d is lower than message queue size: %d, "
 505                       "performance may drop slightly.\n",
 506                       *kiblnd_tunables.kib_concurrent_sends, *kiblnd_tunables.kib_peertxcredits);
 507         }
 508
 509         kiblnd_sysctl_init();
 510         return 0;
 511 }
 512
 513 void
 514 kiblnd_tunables_fini (void)
 515 {
 516         kiblnd_sysctl_fini();
 517 }