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  2008 Sun Microsystems, Inc. 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_timeout = 0;
  72 CFS_MODULE_PARM(peer_timeout, "i", int, 0444,
  73                 "Seconds without aliveness news to declare peer dead (<=0 to disable)");
  74
  75 static char *ipif_name = "ib0";
  76 CFS_MODULE_PARM(ipif_name, "s", charp, 0444,
  77                 "IPoIB interface name");
  78
  79 static int retry_count = 5;
  80 CFS_MODULE_PARM(retry_count, "i", int, 0644,
  81                 "Retransmissions when no ACK received");
  82
  83 static int rnr_retry_count = 6;
  84 CFS_MODULE_PARM(rnr_retry_count, "i", int, 0644,
  85                 "RNR retransmissions");
  86
  87 static int keepalive = 100;
  88 CFS_MODULE_PARM(keepalive, "i", int, 0644,
  89                 "Idle time in seconds before sending a keepalive");
  90
  91 static int ib_mtu = 0;
  92 CFS_MODULE_PARM(ib_mtu, "i", int, 0444,
  93                 "IB MTU 256/512/1024/2048/4096");
  94
  95 static int concurrent_sends = 0;
  96 CFS_MODULE_PARM(concurrent_sends, "i", int, 0444,
  97                 "send work-queue sizing");
  98
  99 static int map_on_demand = 0;
 100 CFS_MODULE_PARM(map_on_demand, "i", int, 0444,
 101                 "map on demand");
 102
 103 static int fmr_pool_size = 512;
 104 CFS_MODULE_PARM(fmr_pool_size, "i", int, 0444,
 105                 "size of the fmr pool (>= ntx)");
 106
 107 static int fmr_flush_trigger = 384;
 108 CFS_MODULE_PARM(fmr_flush_trigger, "i", int, 0444,
 109                 "# dirty FMRs that triggers pool flush");
 110
 111 static int fmr_cache = 1;
 112 CFS_MODULE_PARM(fmr_cache, "i", int, 0444,
 113                 "non-zero to enable FMR caching");
 114
 115 static int pmr_pool_size = 512;
 116 CFS_MODULE_PARM(pmr_pool_size, "i", int, 0444,
 117                 "size of the MR cache pmr pool");
 118
 119 kib_tunables_t kiblnd_tunables = {
 120         .kib_service                = &service,
 121         .kib_cksum                  = &cksum,
 122         .kib_timeout                = &timeout,
 123         .kib_keepalive              = &keepalive,
 124         .kib_ntx                    = &ntx,
 125         .kib_credits                = &credits,
 126         .kib_peercredits            = &peer_credits,
 127         .kib_peercredits_hiw        = &peer_credits_hiw,
 128         .kib_peertimeout            = &peer_timeout,
 129         .kib_default_ipif           = &ipif_name,
 130         .kib_retry_count            = &retry_count,
 131         .kib_rnr_retry_count        = &rnr_retry_count,
 132         .kib_concurrent_sends       = &concurrent_sends,
 133         .kib_ib_mtu                 = &ib_mtu,
 134         .kib_map_on_demand          = &map_on_demand,
 135         .kib_fmr_pool_size          = &fmr_pool_size,
 136         .kib_fmr_flush_trigger      = &fmr_flush_trigger,
 137         .kib_fmr_cache              = &fmr_cache,
 138         .kib_pmr_pool_size          = &pmr_pool_size,
 139 };
 140
 141 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
 142
 143 static char ipif_basename_space[32];
 144
 145 #ifndef HAVE_SYSCTL_UNNUMBERED
 146
 147 enum {
 148         O2IBLND_SERVICE  = 1,
 149         O2IBLND_CKSUM,
 150         O2IBLND_TIMEOUT,
 151         O2IBLND_NTX,
 152         O2IBLND_CREDITS,
 153         O2IBLND_PEER_CREDITS,
 154         O2IBLND_PEER_CREDITS_HIW,
 155         O2IBLND_PEER_TIMEOUT,
 156         O2IBLND_IPIF_BASENAME,
 157         O2IBLND_RETRY_COUNT,
 158         O2IBLND_RNR_RETRY_COUNT,
 159         O2IBLND_KEEPALIVE,
 160         O2IBLND_CONCURRENT_SENDS,
 161         O2IBLND_IB_MTU,
 162         O2IBLND_MAP_ON_DEMAND,
 163         O2IBLND_FMR_POOL_SIZE,
 164         O2IBLND_FMR_FLUSH_TRIGGER,
 165         O2IBLND_FMR_CACHE,
 166         O2IBLND_PMR_POOL_SIZE
 167 };
 168 #else
 169
 170 #define O2IBLND_SERVICE          CTL_UNNUMBERED
 171 #define O2IBLND_CKSUM            CTL_UNNUMBERED
 172 #define O2IBLND_TIMEOUT          CTL_UNNUMBERED
 173 #define O2IBLND_NTX              CTL_UNNUMBERED
 174 #define O2IBLND_CREDITS          CTL_UNNUMBERED
 175 #define O2IBLND_PEER_CREDITS     CTL_UNNUMBERED
 176 #define O2IBLND_PEER_CREDITS_HIW CTL_UNNUMBERED
 177 #define O2IBLND_PEER_TIMEOUT     CTL_UNNUMBERED
 178 #define O2IBLND_IPIF_BASENAME    CTL_UNNUMBERED
 179 #define O2IBLND_RETRY_COUNT      CTL_UNNUMBERED
 180 #define O2IBLND_RNR_RETRY_COUNT  CTL_UNNUMBERED
 181 #define O2IBLND_KEEPALIVE        CTL_UNNUMBERED
 182 #define O2IBLND_CONCURRENT_SENDS CTL_UNNUMBERED
 183 #define O2IBLND_IB_MTU           CTL_UNNUMBERED
 184 #define O2IBLND_MAP_ON_DEMAND    CTL_UNNUMBERED
 185 #define O2IBLND_FMR_POOL_SIZE    CTL_UNNUMBERED
 186 #define O2IBLND_FMR_FLUSH_TRIGGER CTL_UNNUMBERED
 187 #define O2IBLND_FMR_CACHE        CTL_UNNUMBERED
 188 #define O2IBLND_PMR_POOL_SIZE    CTL_UNNUMBERED
 189
 190 #endif
 191
 192 static cfs_sysctl_table_t kiblnd_ctl_table[] = {
 193         {
 194                 .ctl_name = O2IBLND_SERVICE,
 195                 .procname = "service",
 196                 .data     = &service,
 197                 .maxlen   = sizeof(int),
 198                 .mode     = 0444,
 199                 .proc_handler = &proc_dointvec
 200         },
 201         {
 202                 .ctl_name = O2IBLND_CKSUM,
 203                 .procname = "cksum",
 204                 .data     = &cksum,
 205                 .maxlen   = sizeof(int),
 206                 .mode     = 0644,
 207                 .proc_handler = &proc_dointvec
 208         },
 209         {
 210                 .ctl_name = O2IBLND_TIMEOUT,
 211                 .procname = "timeout",
 212                 .data     = &timeout,
 213                 .maxlen   = sizeof(int),
 214                 .mode     = 0644,
 215                 .proc_handler = &proc_dointvec
 216         },
 217         {
 218                 .ctl_name = O2IBLND_NTX,
 219                 .procname = "ntx",
 220                 .data     = &ntx,
 221                 .maxlen   = sizeof(int),
 222                 .mode     = 0444,
 223                 .proc_handler = &proc_dointvec
 224         },
 225         {
 226                 .ctl_name = O2IBLND_CREDITS,
 227                 .procname = "credits",
 228                 .data     = &credits,
 229                 .maxlen   = sizeof(int),
 230                 .mode     = 0444,
 231                 .proc_handler = &proc_dointvec
 232         },
 233         {
 234                 .ctl_name = O2IBLND_PEER_CREDITS,
 235                 .procname = "peer_credits",
 236                 .data     = &peer_credits,
 237                 .maxlen   = sizeof(int),
 238                 .mode     = 0444,
 239                 .proc_handler = &proc_dointvec
 240         },
 241         {
 242                 .ctl_name = O2IBLND_PEER_CREDITS_HIW,
 243                 .procname = "peer_credits_hiw",
 244                 .data     = &peer_credits_hiw,
 245                 .maxlen   = sizeof(int),
 246                 .mode     = 0444,
 247                 .proc_handler = &proc_dointvec
 248         },
 249         {
 250                 .ctl_name = O2IBLND_PEER_TIMEOUT,
 251                 .procname = "peer_timeout",
 252                 .data     = &peer_timeout,
 253                 .maxlen   = sizeof(int),
 254                 .mode     = 0444,
 255                 .proc_handler = &proc_dointvec
 256         },
 257         {
 258                 .ctl_name = O2IBLND_IPIF_BASENAME,
 259                 .procname = "ipif_name",
 260                 .data     = ipif_basename_space,
 261                 .maxlen   = sizeof(ipif_basename_space),
 262                 .mode     = 0444,
 263                 .proc_handler = &proc_dostring
 264         },
 265         {
 266                 .ctl_name = O2IBLND_RETRY_COUNT,
 267                 .procname = "retry_count",
 268                 .data     = &retry_count,
 269                 .maxlen   = sizeof(int),
 270                 .mode     = 0644,
 271                 .proc_handler = &proc_dointvec
 272         },
 273         {
 274                 .ctl_name = O2IBLND_RNR_RETRY_COUNT,
 275                 .procname = "rnr_retry_count",
 276                 .data     = &rnr_retry_count,
 277                 .maxlen   = sizeof(int),
 278                 .mode     = 0644,
 279                 .proc_handler = &proc_dointvec
 280         },
 281         {
 282                 .ctl_name = O2IBLND_KEEPALIVE,
 283                 .procname = "keepalive",
 284                 .data     = &keepalive,
 285                 .maxlen   = sizeof(int),
 286                 .mode     = 0644,
 287                 .proc_handler = &proc_dointvec
 288         },
 289         {
 290                 .ctl_name = O2IBLND_CONCURRENT_SENDS,
 291                 .procname = "concurrent_sends",
 292                 .data     = &concurrent_sends,
 293                 .maxlen   = sizeof(int),
 294                 .mode     = 0444,
 295                 .proc_handler = &proc_dointvec
 296         },
 297         {
 298                 .ctl_name = O2IBLND_IB_MTU,
 299                 .procname = "ib_mtu",
 300                 .data     = &ib_mtu,
 301                 .maxlen   = sizeof(int),
 302                 .mode     = 0444,
 303                 .proc_handler = &proc_dointvec
 304         },
 305         {
 306                 .ctl_name = O2IBLND_MAP_ON_DEMAND,
 307                 .procname = "map_on_demand",
 308                 .data     = &map_on_demand,
 309                 .maxlen   = sizeof(int),
 310                 .mode     = 0444,
 311                 .proc_handler = &proc_dointvec
 312         },
 313
 314         {
 315                 .ctl_name = O2IBLND_FMR_POOL_SIZE,
 316                 .procname = "fmr_pool_size",
 317                 .data     = &fmr_pool_size,
 318                 .maxlen   = sizeof(int),
 319                 .mode     = 0444,
 320                 .proc_handler = &proc_dointvec
 321         },
 322         {
 323                 .ctl_name = O2IBLND_FMR_FLUSH_TRIGGER,
 324                 .procname = "fmr_flush_trigger",
 325                 .data     = &fmr_flush_trigger,
 326                 .maxlen   = sizeof(int),
 327                 .mode     = 0444,
 328                 .proc_handler = &proc_dointvec
 329         },
 330         {
 331                 .ctl_name = O2IBLND_FMR_CACHE,
 332                 .procname = "fmr_cache",
 333                 .data     = &fmr_cache,
 334                 .maxlen   = sizeof(int),
 335                 .mode     = 0444,
 336                 .proc_handler = &proc_dointvec
 337         },
 338         {
 339                 .ctl_name = O2IBLND_PMR_POOL_SIZE,
 340                 .procname = "pmr_pool_size",
 341                 .data     = &pmr_pool_size,
 342                 .maxlen   = sizeof(int),
 343                 .mode     = 0444,
 344                 .proc_handler = &proc_dointvec
 345         },
 346         {0}
 347 };
 348
 349 static cfs_sysctl_table_t kiblnd_top_ctl_table[] = {
 350         {
 351                 .ctl_name = CTL_O2IBLND,
 352                 .procname = "o2iblnd",
 353                 .data     = NULL,
 354                 .maxlen   = 0,
 355                 .mode     = 0555,
 356                 .child    = kiblnd_ctl_table
 357         },
 358         {0}
 359 };
 360
 361 void
 362 kiblnd_initstrtunable(char *space, char *str, int size)
 363 {
 364         strncpy(space, str, size);
 365         space[size-1] = 0;
 366 }
 367
 368 void
 369 kiblnd_sysctl_init (void)
 370 {
 371         kiblnd_initstrtunable(ipif_basename_space, ipif_name,
 372                               sizeof(ipif_basename_space));
 373
 374         kiblnd_tunables.kib_sysctl =
 375                 cfs_register_sysctl_table(kiblnd_top_ctl_table, 0);
 376
 377         if (kiblnd_tunables.kib_sysctl == NULL)
 378                 CWARN("Can't setup /proc tunables\n");
 379 }
 380
 381 void
 382 kiblnd_sysctl_fini (void)
 383 {
 384         if (kiblnd_tunables.kib_sysctl != NULL)
 385                 cfs_unregister_sysctl_table(kiblnd_tunables.kib_sysctl);
 386 }
 387
 388 #else
 389
 390 void
 391 kiblnd_sysctl_init (void)
 392 {
 393 }
 394
 395 void
 396 kiblnd_sysctl_fini (void)
 397 {
 398 }
 399
 400 #endif
 401
 402 int
 403 kiblnd_tunables_init (void)
 404 {
 405         if (*kiblnd_tunables.kib_credits > *kiblnd_tunables.kib_ntx) {
 406                 CERROR("Can't set credits(%d) > ntx(%d)\n",
 407                        *kiblnd_tunables.kib_credits,
 408                        *kiblnd_tunables.kib_ntx);
 409                 return -EINVAL;
 410         }
 411
 412         if (kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu) < 0) {
 413                 CERROR("Invalid ib_mtu %d, expected 256/512/1024/2048/4096\n",
 414                        *kiblnd_tunables.kib_ib_mtu);
 415                 return -EINVAL;
 416         }
 417
 418         if (*kiblnd_tunables.kib_peercredits < IBLND_CREDITS_DEFAULT)
 419                 *kiblnd_tunables.kib_peercredits = IBLND_CREDITS_DEFAULT;
 420
 421         if (*kiblnd_tunables.kib_peercredits > IBLND_CREDITS_MAX)
 422                 *kiblnd_tunables.kib_peercredits = IBLND_CREDITS_MAX;
 423
 424         if (*kiblnd_tunables.kib_peercredits_hiw < *kiblnd_tunables.kib_peercredits / 2)
 425                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peercredits / 2;
 426
 427         if (*kiblnd_tunables.kib_peercredits_hiw >= *kiblnd_tunables.kib_peercredits)
 428                 *kiblnd_tunables.kib_peercredits_hiw = *kiblnd_tunables.kib_peercredits - 1;
 429
 430         if (*kiblnd_tunables.kib_map_on_demand < 0 ||
 431             *kiblnd_tunables.kib_map_on_demand >= IBLND_MAX_RDMA_FRAGS)
 432                 *kiblnd_tunables.kib_map_on_demand = 0; /* disable map-on-demand */
 433
 434         if (*kiblnd_tunables.kib_concurrent_sends == 0) {
 435                 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
 436                     *kiblnd_tunables.kib_map_on_demand <= IBLND_MAX_RDMA_FRAGS / 8)
 437                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peercredits) * 2;
 438                 else
 439                         *kiblnd_tunables.kib_concurrent_sends = (*kiblnd_tunables.kib_peercredits);
 440         }
 441
 442         if (*kiblnd_tunables.kib_concurrent_sends > *kiblnd_tunables.kib_peercredits * 2)
 443                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peercredits * 2;
 444
 445         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peercredits / 2)
 446                 *kiblnd_tunables.kib_concurrent_sends = *kiblnd_tunables.kib_peercredits / 2;
 447
 448         if (*kiblnd_tunables.kib_concurrent_sends < *kiblnd_tunables.kib_peercredits) {
 449                 CWARN("Concurrent sends %d is lower than message queue size: %d, "
 450                       "performance may drop slightly.\n",
 451                       *kiblnd_tunables.kib_concurrent_sends, *kiblnd_tunables.kib_peercredits);
 452         }
 453
 454         kiblnd_sysctl_init();
 455         return 0;
 456 }
 457
 458 void
 459 kiblnd_tunables_fini (void)
 460 {
 461         kiblnd_sysctl_fini();
 462 }