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 #ifndef HAVE_SYSCTL_UNNUMBERED
 184
 185 enum {
 186         O2IBLND_SERVICE  = 1,
 187         O2IBLND_CKSUM,
 188         O2IBLND_TIMEOUT,
 189         O2IBLND_NTX,
 190         O2IBLND_CREDITS,
 191         O2IBLND_PEER_TXCREDITS,
 192         O2IBLND_PEER_CREDITS_HIW,
 193         O2IBLND_PEER_RTRCREDITS,
 194         O2IBLND_PEER_TIMEOUT,
 195         O2IBLND_IPIF_BASENAME,
 196         O2IBLND_RETRY_COUNT,
 197         O2IBLND_RNR_RETRY_COUNT,
 198         O2IBLND_KEEPALIVE,
 199         O2IBLND_CONCURRENT_SENDS,
 200         O2IBLND_IB_MTU,
 201         O2IBLND_MAP_ON_DEMAND,
 202         O2IBLND_FMR_POOL_SIZE,
 203         O2IBLND_FMR_FLUSH_TRIGGER,
 204         O2IBLND_FMR_CACHE,
 205         O2IBLND_PMR_POOL_SIZE,
 206         O2IBLND_DEV_FAILOVER
 207 };
 208 #else
 209
 210 #define O2IBLND_SERVICE          CTL_UNNUMBERED
 211 #define O2IBLND_CKSUM            CTL_UNNUMBERED
 212 #define O2IBLND_TIMEOUT          CTL_UNNUMBERED
 213 #define O2IBLND_NTX              CTL_UNNUMBERED
 214 #define O2IBLND_CREDITS          CTL_UNNUMBERED
 215 #define O2IBLND_PEER_TXCREDITS   CTL_UNNUMBERED
 216 #define O2IBLND_PEER_CREDITS_HIW CTL_UNNUMBERED
 217 #define O2IBLND_PEER_RTRCREDITS  CTL_UNNUMBERED
 218 #define O2IBLND_PEER_TIMEOUT     CTL_UNNUMBERED
 219 #define O2IBLND_IPIF_BASENAME    CTL_UNNUMBERED
 220 #define O2IBLND_RETRY_COUNT      CTL_UNNUMBERED
 221 #define O2IBLND_RNR_RETRY_COUNT  CTL_UNNUMBERED
 222 #define O2IBLND_KEEPALIVE        CTL_UNNUMBERED
 223 #define O2IBLND_CONCURRENT_SENDS CTL_UNNUMBERED
 224 #define O2IBLND_IB_MTU           CTL_UNNUMBERED
 225 #define O2IBLND_MAP_ON_DEMAND    CTL_UNNUMBERED
 226 #define O2IBLND_FMR_POOL_SIZE    CTL_UNNUMBERED
 227 #define O2IBLND_FMR_FLUSH_TRIGGER CTL_UNNUMBERED
 228 #define O2IBLND_FMR_CACHE        CTL_UNNUMBERED
 229 #define O2IBLND_PMR_POOL_SIZE    CTL_UNNUMBERED
 230 #define O2IBLND_DEV_FAILOVER     CTL_UNNUMBERED
 231
 232 #endif
 233
 234 static cfs_sysctl_table_t kiblnd_ctl_table[] = {
 235         {
 236                 .ctl_name = O2IBLND_SERVICE,
 237                 .procname = "service",
 238                 .data     = &service,
 239                 .maxlen   = sizeof(int),
 240                 .mode     = 0444,
 241                 .proc_handler = &proc_dointvec
 242         },
 243         {
 244                 .ctl_name = O2IBLND_CKSUM,
 245                 .procname = "cksum",
 246                 .data     = &cksum,
 247                 .maxlen   = sizeof(int),
 248                 .mode     = 0644,
 249                 .proc_handler = &proc_dointvec
 250         },
 251         {
 252                 .ctl_name = O2IBLND_TIMEOUT,
 253                 .procname = "timeout",
 254                 .data     = &timeout,
 255                 .maxlen   = sizeof(int),
 256                 .mode     = 0644,
 257                 .proc_handler = &proc_dointvec
 258         },
 259         {
 260                 .ctl_name = O2IBLND_NTX,
 261                 .procname = "ntx",
 262                 .data     = &ntx,
 263                 .maxlen   = sizeof(int),
 264                 .mode     = 0444,
 265                 .proc_handler = &proc_dointvec
 266         },
 267         {
 268                 .ctl_name = O2IBLND_CREDITS,
 269                 .procname = "credits",
 270                 .data     = &credits,
 271                 .maxlen   = sizeof(int),
 272                 .mode     = 0444,
 273                 .proc_handler = &proc_dointvec
 274         },
 275         {
 276                 .ctl_name = O2IBLND_PEER_TXCREDITS,
 277                 .procname = "peer_credits",
 278                 .data     = &peer_credits,
 279                 .maxlen   = sizeof(int),
 280                 .mode     = 0444,
 281                 .proc_handler = &proc_dointvec
 282         },
 283         {
 284                 .ctl_name = O2IBLND_PEER_CREDITS_HIW,
 285                 .procname = "peer_credits_hiw",
 286                 .data     = &peer_credits_hiw,
 287                 .maxlen   = sizeof(int),
 288                 .mode     = 0444,
 289                 .proc_handler = &proc_dointvec
 290         },
 291         {
 292                 .ctl_name = O2IBLND_PEER_RTRCREDITS,
 293                 .procname = "peer_buffer_credits",
 294                 .data     = &peer_buffer_credits,
 295                 .maxlen   = sizeof(int),
 296                 .mode     = 0444,
 297                 .proc_handler = &proc_dointvec
 298         },
 299         {
 300                 .ctl_name = O2IBLND_PEER_TIMEOUT,
 301                 .procname = "peer_timeout",
 302                 .data     = &peer_timeout,
 303                 .maxlen   = sizeof(int),
 304                 .mode     = 0444,
 305                 .proc_handler = &proc_dointvec
 306         },
 307         {
 308                 .ctl_name = O2IBLND_IPIF_BASENAME,
 309                 .procname = "ipif_name",
 310                 .data     = ipif_basename_space,
 311                 .maxlen   = sizeof(ipif_basename_space),
 312                 .mode     = 0444,
 313                 .proc_handler = &proc_dostring
 314         },
 315         {
 316                 .ctl_name = O2IBLND_RETRY_COUNT,
 317                 .procname = "retry_count",
 318                 .data     = &retry_count,
 319                 .maxlen   = sizeof(int),
 320                 .mode     = 0644,
 321                 .proc_handler = &proc_dointvec
 322         },
 323         {
 324                 .ctl_name = O2IBLND_RNR_RETRY_COUNT,
 325                 .procname = "rnr_retry_count",
 326                 .data     = &rnr_retry_count,
 327                 .maxlen   = sizeof(int),
 328                 .mode     = 0644,
 329                 .proc_handler = &proc_dointvec
 330         },
 331         {
 332                 .ctl_name = O2IBLND_KEEPALIVE,
 333                 .procname = "keepalive",
 334                 .data     = &keepalive,
 335                 .maxlen   = sizeof(int),
 336                 .mode     = 0644,
 337                 .proc_handler = &proc_dointvec
 338         },
 339         {
 340                 .ctl_name = O2IBLND_CONCURRENT_SENDS,
 341                 .procname = "concurrent_sends",
 342                 .data     = &concurrent_sends,
 343                 .maxlen   = sizeof(int),
 344                 .mode     = 0444,
 345                 .proc_handler = &proc_dointvec
 346         },
 347         {
 348                 .ctl_name = O2IBLND_IB_MTU,
 349                 .procname = "ib_mtu",
 350                 .data     = &ib_mtu,
 351                 .maxlen   = sizeof(int),
 352                 .mode     = 0444,
 353                 .proc_handler = &proc_dointvec
 354         },
 355         {
 356                 .ctl_name = O2IBLND_MAP_ON_DEMAND,
 357                 .procname = "map_on_demand",
 358                 .data     = &map_on_demand,
 359                 .maxlen   = sizeof(int),
 360                 .mode     = 0444,
 361                 .proc_handler = &proc_dointvec
 362         },
 363
 364         {
 365                 .ctl_name = O2IBLND_FMR_POOL_SIZE,
 366                 .procname = "fmr_pool_size",
 367                 .data     = &fmr_pool_size,
 368                 .maxlen   = sizeof(int),
 369                 .mode     = 0444,
 370                 .proc_handler = &proc_dointvec
 371         },
 372         {
 373                 .ctl_name = O2IBLND_FMR_FLUSH_TRIGGER,
 374                 .procname = "fmr_flush_trigger",
 375                 .data     = &fmr_flush_trigger,
 376                 .maxlen   = sizeof(int),
 377                 .mode     = 0444,
 378                 .proc_handler = &proc_dointvec
 379         },
 380         {
 381                 .ctl_name = O2IBLND_FMR_CACHE,
 382                 .procname = "fmr_cache",
 383                 .data     = &fmr_cache,
 384                 .maxlen   = sizeof(int),
 385                 .mode     = 0444,
 386                 .proc_handler = &proc_dointvec
 387         },
 388         {
 389                 .ctl_name = O2IBLND_PMR_POOL_SIZE,
 390                 .procname = "pmr_pool_size",
 391                 .data     = &pmr_pool_size,
 392                 .maxlen   = sizeof(int),
 393                 .mode     = 0444,
 394                 .proc_handler = &proc_dointvec
 395         },
 396         {
 397                 .ctl_name = O2IBLND_DEV_FAILOVER,
 398                 .procname = "dev_failover",
 399                 .data     = &dev_failover,
 400                 .maxlen   = sizeof(int),
 401                 .mode     = 0444,
 402                 .proc_handler = &proc_dointvec
 403         },
 404         {0}
 405 };
 406
 407 static cfs_sysctl_table_t kiblnd_top_ctl_table[] = {
 408         {
 409                 .ctl_name = CTL_O2IBLND,
 410                 .procname = "o2iblnd",
 411                 .data     = NULL,
 412                 .maxlen   = 0,
 413                 .mode     = 0555,
 414                 .child    = kiblnd_ctl_table
 415         },
 416         {0}
 417 };
 418
 419 void
 420 kiblnd_initstrtunable(char *space, char *str, int size)
 421 {
 422         strncpy(space, str, size);
 423         space[size-1] = 0;
 424 }
 425
 426 void
 427 kiblnd_sysctl_init (void)
 428 {
 429         kiblnd_initstrtunable(ipif_basename_space, ipif_name,
 430                               sizeof(ipif_basename_space));
 431
 432         kiblnd_tunables.kib_sysctl =
 433                 cfs_register_sysctl_table(kiblnd_top_ctl_table, 0);
 434
 435         if (kiblnd_tunables.kib_sysctl == NULL)
 436                 CWARN("Can't setup /proc tunables\n");
 437 }
 438
 439 void
 440 kiblnd_sysctl_fini (void)
 441 {
 442         if (kiblnd_tunables.kib_sysctl != NULL)
 443                 cfs_unregister_sysctl_table(kiblnd_tunables.kib_sysctl);
 444 }
 445
 446 #else
 447
 448 void
 449 kiblnd_sysctl_init (void)
 450 {
 451 }
 452
 453 void
 454 kiblnd_sysctl_fini (void)
 455 {
 456 }
 457
 458 #endif
 459
 460 int
 461 kiblnd_tunables_init (void)
 462 {
 463         if (kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu) < 0) {
 464                 CERROR("Invalid ib_mtu %d, expected 256/512/1024/2048/4096\n",
 465                        *kiblnd_tunables.kib_ib_mtu);
 466                 return -EINVAL;
 467         }
 468
 469         if (*kiblnd_tunables.kib_peertxcredits < IBLND_CREDITS_DEFAULT)
 470                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_DEFAULT;
 471
 472         if (*kiblnd_tunables.kib_peertxcredits > IBLND_CREDITS_MAX)
 473                 *kiblnd_tunables.kib_peertxcredits = IBLND_CREDITS_MAX;
 474
 475         if (*kiblnd_tunables.kib_peertxcredits > *kiblnd_tunables.kib_credits)
 476                 *kiblnd_tunables.kib_peertxcredits = *kiblnd_tunables.kib_credits;
 477
 478         if (*kiblnd_tunables.kib_peercredits_hiw < *kiblnd_tunables.kib_peertxcredits / 2)
 479                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits / 2;
 480
 481         if (*kiblnd_tunables.kib_peercredits_hiw >= *kiblnd_tunables.kib_peertxcredits)
 482                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peertxcredits - 1;
 483
 484         if (*kiblnd_tunables.kib_map_on_demand < 0 ||
 485             *kiblnd_tunables.kib_map_on_demand > IBLND_MAX_RDMA_FRAGS)
 486                 *kiblnd_tunables.kib_map_on_demand = 0; /* disable map-on-demand */
 487
 488         if (*kiblnd_tunables.kib_map_on_demand == 1)
 489                 *kiblnd_tunables.kib_map_on_demand = 2; /* don't make sense to create map if only one fragment */
 490
 491         if (*kiblnd_tunables.kib_concurrent_sends == 0) {
 492                 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
 493                     *kiblnd_tunables.kib_map_on_demand <= IBLND_MAX_RDMA_FRAGS / 8)
 494                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits) * 2;
 495                 else
 496                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peertxcredits);
 497         }
 498
 499         if (*kiblnd_tunables.kib_concurrent_sends > *kiblnd_tunables.kib_peertxcredits * 2)
 500                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits * 2;
 501
 502         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits / 2)
 503                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peertxcredits / 2;
 504
 505         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peertxcredits) {
 506                 CWARN("Concurrent sends %d is lower than message queue size: %d, "
 507                       "performance may drop slightly.\n",
 508                       *kiblnd_tunables.kib_concurrent_sends, *kiblnd_tunables.kib_peertxcredits);
 509         }
 510
 511         kiblnd_sysctl_init();
 512         return 0;
 513 }
 514
 515 void
 516 kiblnd_tunables_fini (void)
 517 {
 518         kiblnd_sysctl_fini();
 519 }