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 module_param(service, int, 0444);
  45 MODULE_PARM_DESC(service, "service number (within RDMA_PS_TCP)");
  46
  47 static int cksum = 0;
  48 module_param(cksum, int, 0644);
  49 MODULE_PARM_DESC(cksum, "set non-zero to enable message (not RDMA) checksums");
  50
  51 static int timeout = 50;
  52 module_param(timeout, int, 0644);
  53 MODULE_PARM_DESC(timeout, "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 module_param(nscheds, int, 0444);
  59 MODULE_PARM_DESC(nscheds, "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 module_param(ntx, int, 0444);
  64 MODULE_PARM_DESC(ntx, "# of message descriptors allocated for each pool");
  65
  66 /* NB: this value is shared by all CPTs */
  67 static int credits = 256;
  68 module_param(credits, int, 0444);
  69 MODULE_PARM_DESC(credits, "# concurrent sends");
  70
  71 static int peer_credits = 8;
  72 module_param(peer_credits, int, 0444);
  73 MODULE_PARM_DESC(peer_credits, "# concurrent sends to 1 peer");
  74
  75 static int peer_credits_hiw = 0;
  76 module_param(peer_credits_hiw, int, 0444);
  77 MODULE_PARM_DESC(peer_credits_hiw, "when eagerly to return credits");
  78
  79 static int peer_buffer_credits = 0;
  80 module_param(peer_buffer_credits, int, 0444);
  81 MODULE_PARM_DESC(peer_buffer_credits, "# per-peer router buffer credits");
  82
  83 static int peer_timeout = 180;
  84 module_param(peer_timeout, int, 0444);
  85 MODULE_PARM_DESC(peer_timeout, "Seconds without aliveness news to declare peer dead (<=0 to disable)");
  86
  87 static char *ipif_name = "ib0";
  88 module_param(ipif_name, charp, 0444);
  89 MODULE_PARM_DESC(ipif_name, "IPoIB interface name");
  90
  91 static int retry_count = 5;
  92 module_param(retry_count, int, 0644);
  93 MODULE_PARM_DESC(retry_count, "Retransmissions when no ACK received");
  94
  95 static int rnr_retry_count = 6;
  96 module_param(rnr_retry_count, int, 0644);
  97 MODULE_PARM_DESC(rnr_retry_count, "RNR retransmissions");
  98
  99 static int keepalive = 100;
 100 module_param(keepalive, int, 0644);
 101 MODULE_PARM_DESC(keepalive, "Idle time in seconds before sending a keepalive");
 102
 103 static int ib_mtu;
 104 module_param(ib_mtu, int, 0444);
 105 MODULE_PARM_DESC(ib_mtu, "IB MTU 256/512/1024/2048/4096");
 106
 107 static int concurrent_sends;
 108 module_param(concurrent_sends, int, 0444);
 109 MODULE_PARM_DESC(concurrent_sends, "send work-queue sizing");
 110
 111 static int map_on_demand;
 112 module_param(map_on_demand, int, 0444);
 113 MODULE_PARM_DESC(map_on_demand, "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 module_param(fmr_pool_size, int, 0444);
 118 MODULE_PARM_DESC(fmr_pool_size, "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 module_param(fmr_flush_trigger, int, 0444);
 123 MODULE_PARM_DESC(fmr_flush_trigger, "# dirty FMRs that triggers pool flush");
 124
 125 static int fmr_cache = 1;
 126 module_param(fmr_cache, int, 0444);
 127 MODULE_PARM_DESC(fmr_cache, "non-zero to enable FMR caching");
 128
 129 /*
 130  * 0: disable failover
 131  * 1: enable failover if necessary
 132  * 2: force to failover (for debug)
 133  */
 134 static int dev_failover = 0;
 135 module_param(dev_failover, int, 0444);
 136 MODULE_PARM_DESC(dev_failover, "HCA failover for bonding (0 off, 1 on, other values reserved)");
 137
 138 static int require_privileged_port;
 139 module_param(require_privileged_port, int, 0644);
 140 MODULE_PARM_DESC(require_privileged_port, "require privileged port when accepting connection");
 141
 142 static int use_privileged_port = 1;
 143 module_param(use_privileged_port, int, 0644);
 144 MODULE_PARM_DESC(use_privileged_port, "use privileged port when initiating connection");
 145
 146 kib_tunables_t kiblnd_tunables = {
 147         .kib_dev_failover           = &dev_failover,
 148         .kib_service                = &service,
 149         .kib_cksum                  = &cksum,
 150         .kib_timeout                = &timeout,
 151         .kib_keepalive              = &keepalive,
 152         .kib_ntx                    = &ntx,
 153         .kib_default_ipif           = &ipif_name,
 154         .kib_retry_count            = &retry_count,
 155         .kib_rnr_retry_count        = &rnr_retry_count,
 156         .kib_ib_mtu                 = &ib_mtu,
 157         .kib_require_priv_port      = &require_privileged_port,
 158         .kib_use_priv_port          = &use_privileged_port,
 159         .kib_nscheds                = &nscheds
 160 };
 161
 162 static struct lnet_ioctl_config_o2iblnd_tunables default_tunables;
 163
 164 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
 165
 166 static char ipif_basename_space[32];
 167
 168 static struct ctl_table kiblnd_ctl_table[] = {
 169         {
 170                 INIT_CTL_NAME
 171                 .procname       = "service",
 172                 .data           = &service,
 173                 .maxlen         = sizeof(int),
 174                 .mode           = 0444,
 175                 .proc_handler   = &proc_dointvec
 176         },
 177         {
 178                 INIT_CTL_NAME
 179                 .procname       = "cksum",
 180                 .data           = &cksum,
 181                 .maxlen         = sizeof(int),
 182                 .mode           = 0644,
 183                 .proc_handler   = &proc_dointvec
 184         },
 185         {
 186                 INIT_CTL_NAME
 187                 .procname       = "timeout",
 188                 .data           = &timeout,
 189                 .maxlen         = sizeof(int),
 190                 .mode           = 0644,
 191                 .proc_handler   = &proc_dointvec
 192         },
 193         {
 194                 INIT_CTL_NAME
 195                 .procname       = "ntx",
 196                 .data           = &ntx,
 197                 .maxlen         = sizeof(int),
 198                 .mode           = 0444,
 199                 .proc_handler   = &proc_dointvec
 200         },
 201         {
 202                 INIT_CTL_NAME
 203                 .procname       = "credits",
 204                 .data           = &credits,
 205                 .maxlen         = sizeof(int),
 206                 .mode           = 0444,
 207                 .proc_handler   = &proc_dointvec
 208         },
 209         {
 210                 INIT_CTL_NAME
 211                 .procname       = "peer_credits",
 212                 .data           = &peer_credits,
 213                 .maxlen         = sizeof(int),
 214                 .mode           = 0444,
 215                 .proc_handler   = &proc_dointvec
 216         },
 217         {
 218                 INIT_CTL_NAME
 219                 .procname       = "peer_credits_hiw",
 220                 .data           = &peer_credits_hiw,
 221                 .maxlen         = sizeof(int),
 222                 .mode           = 0444,
 223                 .proc_handler   = &proc_dointvec
 224         },
 225         {
 226                 INIT_CTL_NAME
 227                 .procname       = "peer_buffer_credits",
 228                 .data           = &peer_buffer_credits,
 229                 .maxlen         = sizeof(int),
 230                 .mode           = 0444,
 231                 .proc_handler   = &proc_dointvec
 232         },
 233         {
 234                 INIT_CTL_NAME
 235                 .procname       = "peer_timeout",
 236                 .data           = &peer_timeout,
 237                 .maxlen         = sizeof(int),
 238                 .mode           = 0444,
 239                 .proc_handler   = &proc_dointvec
 240         },
 241         {
 242                 INIT_CTL_NAME
 243                 .procname       = "ipif_name",
 244                 .data           = ipif_basename_space,
 245                 .maxlen         = sizeof(ipif_basename_space),
 246                 .mode           = 0444,
 247                 .proc_handler   = &proc_dostring
 248         },
 249         {
 250                 INIT_CTL_NAME
 251                 .procname       = "retry_count",
 252                 .data           = &retry_count,
 253                 .maxlen         = sizeof(int),
 254                 .mode           = 0644,
 255                 .proc_handler   = &proc_dointvec
 256         },
 257         {
 258                 INIT_CTL_NAME
 259                 .procname       = "rnr_retry_count",
 260                 .data           = &rnr_retry_count,
 261                 .maxlen         = sizeof(int),
 262                 .mode           = 0644,
 263                 .proc_handler   = &proc_dointvec
 264         },
 265         {
 266                 INIT_CTL_NAME
 267                 .procname       = "keepalive",
 268                 .data           = &keepalive,
 269                 .maxlen         = sizeof(int),
 270                 .mode           = 0644,
 271                 .proc_handler   = &proc_dointvec
 272         },
 273         {
 274                 INIT_CTL_NAME
 275                 .procname       = "concurrent_sends",
 276                 .data           = &concurrent_sends,
 277                 .maxlen         = sizeof(int),
 278                 .mode           = 0444,
 279                 .proc_handler   = &proc_dointvec
 280         },
 281         {
 282                 INIT_CTL_NAME
 283                 .procname       = "ib_mtu",
 284                 .data           = &ib_mtu,
 285                 .maxlen         = sizeof(int),
 286                 .mode           = 0444,
 287                 .proc_handler   = &proc_dointvec
 288         },
 289         {
 290                 INIT_CTL_NAME
 291                 .procname       = "map_on_demand",
 292                 .data           = &map_on_demand,
 293                 .maxlen         = sizeof(int),
 294                 .mode           = 0444,
 295                 .proc_handler   = &proc_dointvec
 296         },
 297         {
 298                 INIT_CTL_NAME
 299                 .procname       = "fmr_pool_size",
 300                 .data           = &fmr_pool_size,
 301                 .maxlen         = sizeof(int),
 302                 .mode           = 0444,
 303                 .proc_handler   = &proc_dointvec
 304         },
 305         {
 306                 INIT_CTL_NAME
 307                 .procname       = "fmr_flush_trigger",
 308                 .data           = &fmr_flush_trigger,
 309                 .maxlen         = sizeof(int),
 310                 .mode           = 0444,
 311                 .proc_handler   = &proc_dointvec
 312         },
 313         {
 314                 INIT_CTL_NAME
 315                 .procname       = "fmr_cache",
 316                 .data           = &fmr_cache,
 317                 .maxlen         = sizeof(int),
 318                 .mode           = 0444,
 319                 .proc_handler   = &proc_dointvec
 320         },
 321         {
 322                 INIT_CTL_NAME
 323                 .procname       = "dev_failover",
 324                 .data           = &dev_failover,
 325                 .maxlen         = sizeof(int),
 326                 .mode           = 0444,
 327                 .proc_handler   = &proc_dointvec
 328         },
 329         { 0 }
 330 };
 331
 332 static struct ctl_table kiblnd_top_ctl_table[] = {
 333         {
 334                 INIT_CTL_NAME
 335                 .procname       = "o2iblnd",
 336                 .data           = NULL,
 337                 .maxlen         = 0,
 338                 .mode           = 0555,
 339                 .child          = kiblnd_ctl_table
 340         },
 341         { 0 }
 342 };
 343
 344 void
 345 kiblnd_initstrtunable(char *space, char *str, int size)
 346 {
 347         strncpy(space, str, size);
 348         space[size-1] = 0;
 349 }
 350
 351 static void
 352 kiblnd_sysctl_init (void)
 353 {
 354         kiblnd_initstrtunable(ipif_basename_space, ipif_name,
 355                               sizeof(ipif_basename_space));
 356
 357         kiblnd_tunables.kib_sysctl =
 358                 register_sysctl_table(kiblnd_top_ctl_table);
 359
 360         if (kiblnd_tunables.kib_sysctl == NULL)
 361                 CWARN("Can't setup /proc tunables\n");
 362 }
 363
 364 static void
 365 kiblnd_sysctl_fini (void)
 366 {
 367         if (kiblnd_tunables.kib_sysctl != NULL)
 368                 unregister_sysctl_table(kiblnd_tunables.kib_sysctl);
 369 }
 370
 371 #else
 372
 373 static void
 374 kiblnd_sysctl_init (void)
 375 {
 376 }
 377
 378 static void
 379 kiblnd_sysctl_fini (void)
 380 {
 381 }
 382 #endif
 383
 384 /* # messages/RDMAs in-flight */
 385 int
 386 kiblnd_msg_queue_size(int version, lnet_ni_t *ni)
 387 {
 388         if (version == IBLND_MSG_VERSION_1)
 389                 return IBLND_MSG_QUEUE_SIZE_V1;
 390         else if (ni)
 391                 return ni->ni_peertxcredits;
 392         else
 393                 return peer_credits;
 394 }
 395
 396 int
 397 kiblnd_tunables_setup(lnet_ni_t *ni)
 398 {
 399         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
 400
 401         /*
 402          * if there was no tunables specified, setup the tunables to be
 403          * defaulted
 404          */
 405         if (!ni->ni_lnd_tunables) {
 406                 LIBCFS_ALLOC(ni->ni_lnd_tunables,
 407                              sizeof(*ni->ni_lnd_tunables));
 408                 if (!ni->ni_lnd_tunables)
 409                         return -ENOMEM;
 410
 411                 memcpy(&ni->ni_lnd_tunables->lt_tun_u.lt_o2ib,
 412                        &default_tunables, sizeof(*tunables));
 413         }
 414         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
 415
 416         /* Current API version */
 417         tunables->lnd_version = 0;
 418
 419         if (kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu) < 0) {
 420                 CERROR("Invalid ib_mtu %d, expected 256/512/1024/2048/4096\n",
 421                        *kiblnd_tunables.kib_ib_mtu);
 422                 return -EINVAL;
 423         }
 424
 425         if (!ni->ni_peertimeout)
 426                 ni->ni_peertimeout = peer_timeout;
 427
 428         if (!ni->ni_maxtxcredits)
 429                 ni->ni_maxtxcredits = credits;
 430
 431         if (!ni->ni_peertxcredits)
 432                 ni->ni_peertxcredits = peer_credits;
 433
 434         if (!ni->ni_peerrtrcredits)
 435                 ni->ni_peerrtrcredits = peer_buffer_credits;
 436
 437         if (ni->ni_peertxcredits < IBLND_CREDITS_DEFAULT)
 438                 ni->ni_peertxcredits = IBLND_CREDITS_DEFAULT;
 439
 440         if (ni->ni_peertxcredits > IBLND_CREDITS_MAX)
 441                 ni->ni_peertxcredits = IBLND_CREDITS_MAX;
 442
 443         if (ni->ni_peertxcredits > credits)
 444                 ni->ni_peertxcredits = credits;
 445
 446         if (!tunables->lnd_peercredits_hiw)
 447                 tunables->lnd_peercredits_hiw = peer_credits_hiw;
 448
 449         if (tunables->lnd_peercredits_hiw < ni->ni_peertxcredits / 2)
 450                 tunables->lnd_peercredits_hiw = ni->ni_peertxcredits / 2;
 451
 452         if (tunables->lnd_peercredits_hiw >= ni->ni_peertxcredits)
 453                 tunables->lnd_peercredits_hiw = ni->ni_peertxcredits - 1;
 454
 455         if (tunables->lnd_map_on_demand < 0 ||
 456             tunables->lnd_map_on_demand > IBLND_MAX_RDMA_FRAGS) {
 457                 /* disable map-on-demand */
 458                 tunables->lnd_map_on_demand = 0;
 459         }
 460
 461         if (tunables->lnd_map_on_demand == 1) {
 462                 /* don't make sense to create map if only one fragment */
 463                 tunables->lnd_map_on_demand = 2;
 464         }
 465
 466         if (tunables->lnd_concurrent_sends == 0) {
 467                 if (tunables->lnd_map_on_demand > 0 &&
 468                     tunables->lnd_map_on_demand <= IBLND_MAX_RDMA_FRAGS / 8) {
 469                         tunables->lnd_concurrent_sends =
 470                                                 ni->ni_peertxcredits * 2;
 471                 } else {
 472                         tunables->lnd_concurrent_sends = ni->ni_peertxcredits;
 473                 }
 474         }
 475
 476         if (tunables->lnd_concurrent_sends > ni->ni_peertxcredits * 2)
 477                 tunables->lnd_concurrent_sends = ni->ni_peertxcredits * 2;
 478
 479         if (tunables->lnd_concurrent_sends < ni->ni_peertxcredits / 2)
 480                 tunables->lnd_concurrent_sends = ni->ni_peertxcredits / 2;
 481
 482         if (tunables->lnd_concurrent_sends < ni->ni_peertxcredits) {
 483                 CWARN("Concurrent sends %d is lower than message "
 484                       "queue size: %d, performance may drop slightly.\n",
 485                       tunables->lnd_concurrent_sends, ni->ni_peertxcredits);
 486         }
 487
 488         if (!tunables->lnd_fmr_pool_size)
 489                 tunables->lnd_fmr_pool_size = fmr_pool_size;
 490         if (!tunables->lnd_fmr_flush_trigger)
 491                 tunables->lnd_fmr_flush_trigger = fmr_flush_trigger;
 492         if (!tunables->lnd_fmr_cache)
 493                 tunables->lnd_fmr_cache = fmr_cache;
 494
 495         return 0;
 496 }
 497
 498 int
 499 kiblnd_tunables_init(void)
 500 {
 501         default_tunables.lnd_version = 0;
 502         default_tunables.lnd_peercredits_hiw = peer_credits_hiw,
 503         default_tunables.lnd_map_on_demand = map_on_demand;
 504         default_tunables.lnd_concurrent_sends = concurrent_sends;
 505         default_tunables.lnd_fmr_pool_size = fmr_pool_size;
 506         default_tunables.lnd_fmr_flush_trigger = fmr_flush_trigger;
 507         default_tunables.lnd_fmr_cache = fmr_cache;
 508
 509         kiblnd_sysctl_init();
 510         return 0;
 511 }
 512
 513 void
 514 kiblnd_tunables_fini(void)
 515 {
 516         kiblnd_sysctl_fini();
 517 }