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_timeout = 0;
  68 CFS_MODULE_PARM(peer_timeout, "i", int, 0444,
  69                 "Seconds without aliveness news to declare peer dead (<=0 to disable)");
  70
  71 static char *ipif_name = "ib0";
  72 CFS_MODULE_PARM(ipif_name, "s", charp, 0444,
  73                 "IPoIB interface name");
  74
  75 static int retry_count = 5;
  76 CFS_MODULE_PARM(retry_count, "i", int, 0644,
  77                 "Retransmissions when no ACK received");
  78
  79 static int rnr_retry_count = 6;
  80 CFS_MODULE_PARM(rnr_retry_count, "i", int, 0644,
  81                 "RNR retransmissions");
  82
  83 static int keepalive = 100;
  84 CFS_MODULE_PARM(keepalive, "i", int, 0644,
  85                 "Idle time in seconds before sending a keepalive");
  86
  87 static int ib_mtu = 0;
  88 CFS_MODULE_PARM(ib_mtu, "i", int, 0444,
  89                 "IB MTU 256/512/1024/2048/4096");
  90
  91 #if IBLND_MAP_ON_DEMAND
  92 static int concurrent_sends = IBLND_RX_MSGS;
  93 #else
  94 static int concurrent_sends = IBLND_MSG_QUEUE_SIZE;
  95 #endif
  96 CFS_MODULE_PARM(concurrent_sends, "i", int, 0444,
  97                 "send work-queue sizing");
  98
  99 #if IBLND_MAP_ON_DEMAND
 100 static int fmr_pool_size = 512;
 101 CFS_MODULE_PARM(fmr_pool_size, "i", int, 0444,
 102                 "size of the fmr pool (>= ntx)");
 103
 104 static int fmr_flush_trigger = 384;
 105 CFS_MODULE_PARM(fmr_flush_trigger, "i", int, 0444,
 106                 "# dirty FMRs that triggers pool flush");
 107
 108 static int fmr_cache = 1;
 109 CFS_MODULE_PARM(fmr_cache, "i", int, 0444,
 110                 "non-zero to enable FMR caching");
 111 #endif
 112
 113 kib_tunables_t kiblnd_tunables = {
 114         .kib_service                = &service,
 115         .kib_cksum                  = &cksum,
 116         .kib_timeout                = &timeout,
 117         .kib_keepalive              = &keepalive,
 118         .kib_ntx                    = &ntx,
 119         .kib_credits                = &credits,
 120         .kib_peercredits            = &peer_credits,
 121         .kib_peertimeout            = &peer_timeout,
 122         .kib_default_ipif           = &ipif_name,
 123         .kib_retry_count            = &retry_count,
 124         .kib_rnr_retry_count        = &rnr_retry_count,
 125         .kib_concurrent_sends       = &concurrent_sends,
 126         .kib_ib_mtu                 = &ib_mtu,
 127 #if IBLND_MAP_ON_DEMAND
 128         .kib_fmr_pool_size          = &fmr_pool_size,
 129         .kib_fmr_flush_trigger      = &fmr_flush_trigger,
 130         .kib_fmr_cache              = &fmr_cache,
 131 #endif
 132 };
 133
 134 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
 135
 136 static char ipif_basename_space[32];
 137
 138 #ifndef HAVE_SYSCTL_UNNUMBERED
 139
 140 enum {
 141         O2IBLND_SERVICE  = 1,
 142         O2IBLND_CKSUM,
 143         O2IBLND_TIMEOUT,
 144         O2IBLND_NTX,
 145         O2IBLND_CREDITS,
 146         O2IBLND_PEER_CREDITS,
 147         O2IBLND_PEER_TIMEOUT,
 148         O2IBLND_IPIF_BASENAME,
 149         O2IBLND_RETRY_COUNT,
 150         O2IBLND_RNR_RETRY_COUNT,
 151         O2IBLND_KEEPALIVE,
 152         O2IBLND_CONCURRENT_SENDS,
 153         O2IBLND_IB_MTU,
 154         O2IBLND_FMR_POOL_SIZE,
 155         O2IBLND_FMR_FLUSH_TRIGGER,
 156         O2IBLND_FMR_CACHE
 157 };
 158 #else
 159
 160 #define O2IBLND_SERVICE          CTL_UNNUMBERED
 161 #define O2IBLND_CKSUM            CTL_UNNUMBERED
 162 #define O2IBLND_TIMEOUT          CTL_UNNUMBERED
 163 #define O2IBLND_NTX              CTL_UNNUMBERED
 164 #define O2IBLND_CREDITS          CTL_UNNUMBERED
 165 #define O2IBLND_PEER_CREDITS     CTL_UNNUMBERED
 166 #define O2IBLND_PEER_TIMEOUT     CTL_UNNUMBERED
 167 #define O2IBLND_IPIF_BASENAME    CTL_UNNUMBERED
 168 #define O2IBLND_RETRY_COUNT      CTL_UNNUMBERED
 169 #define O2IBLND_RNR_RETRY_COUNT  CTL_UNNUMBERED
 170 #define O2IBLND_KEEPALIVE        CTL_UNNUMBERED
 171 #define O2IBLND_CONCURRENT_SENDS CTL_UNNUMBERED
 172 #define O2IBLND_IB_MTU           CTL_UNNUMBERED
 173 #define O2IBLND_FMR_POOL_SIZE    CTL_UNNUMBERED
 174 #define O2IBLND_FMR_FLUSH_TRIGGER CTL_UNNUMBERED
 175 #define O2IBLND_FMR_CACHE        CTL_UNNUMBERED
 176
 177 #endif
 178
 179 static cfs_sysctl_table_t kiblnd_ctl_table[] = {
 180         {
 181                 .ctl_name = O2IBLND_SERVICE,
 182                 .procname = "service",
 183                 .data     = &service,
 184                 .maxlen   = sizeof(int),
 185                 .mode     = 0444,
 186                 .proc_handler = &proc_dointvec
 187         },
 188         {
 189                 .ctl_name = O2IBLND_CKSUM,
 190                 .procname = "cksum",
 191                 .data     = &cksum,
 192                 .maxlen   = sizeof(int),
 193                 .mode     = 0644,
 194                 .proc_handler = &proc_dointvec
 195         },
 196         {
 197                 .ctl_name = O2IBLND_TIMEOUT,
 198                 .procname = "timeout",
 199                 .data     = &timeout,
 200                 .maxlen   = sizeof(int),
 201                 .mode     = 0644,
 202                 .proc_handler = &proc_dointvec
 203         },
 204         {
 205                 .ctl_name = O2IBLND_NTX,
 206                 .procname = "ntx",
 207                 .data     = &ntx,
 208                 .maxlen   = sizeof(int),
 209                 .mode     = 0444,
 210                 .proc_handler = &proc_dointvec
 211         },
 212         {
 213                 .ctl_name = O2IBLND_CREDITS,
 214                 .procname = "credits",
 215                 .data     = &credits,
 216                 .maxlen   = sizeof(int),
 217                 .mode     = 0444,
 218                 .proc_handler = &proc_dointvec
 219         },
 220         {
 221                 .ctl_name = O2IBLND_PEER_CREDITS,
 222                 .procname = "peer_credits",
 223                 .data     = &peer_credits,
 224                 .maxlen   = sizeof(int),
 225                 .mode     = 0444,
 226                 .proc_handler = &proc_dointvec
 227         },
 228         {
 229                 .ctl_name = O2IBLND_PEER_TIMEOUT,
 230                 .procname = "peer_timeout",
 231                 .data     = &peer_timeout,
 232                 .maxlen   = sizeof(int),
 233                 .mode     = 0444,
 234                 .proc_handler = &proc_dointvec
 235         },
 236         {
 237                 .ctl_name = O2IBLND_IPIF_BASENAME,
 238                 .procname = "ipif_name",
 239                 .data     = ipif_basename_space,
 240                 .maxlen   = sizeof(ipif_basename_space),
 241                 .mode     = 0444,
 242                 .proc_handler = &proc_dostring
 243         },
 244         {
 245                 .ctl_name = O2IBLND_RETRY_COUNT,
 246                 .procname = "retry_count",
 247                 .data     = &retry_count,
 248                 .maxlen   = sizeof(int),
 249                 .mode     = 0644,
 250                 .proc_handler = &proc_dointvec
 251         },
 252         {
 253                 .ctl_name = O2IBLND_RNR_RETRY_COUNT,
 254                 .procname = "rnr_retry_count",
 255                 .data     = &rnr_retry_count,
 256                 .maxlen   = sizeof(int),
 257                 .mode     = 0644,
 258                 .proc_handler = &proc_dointvec
 259         },
 260         {
 261                 .ctl_name = O2IBLND_KEEPALIVE,
 262                 .procname = "keepalive",
 263                 .data     = &keepalive,
 264                 .maxlen   = sizeof(int),
 265                 .mode     = 0644,
 266                 .proc_handler = &proc_dointvec
 267         },
 268         {
 269                 .ctl_name = O2IBLND_CONCURRENT_SENDS,
 270                 .procname = "concurrent_sends",
 271                 .data     = &concurrent_sends,
 272                 .maxlen   = sizeof(int),
 273                 .mode     = 0444,
 274                 .proc_handler = &proc_dointvec
 275         },
 276         {
 277                 .ctl_name = O2IBLND_IB_MTU,
 278                 .procname = "ib_mtu",
 279                 .data     = &ib_mtu,
 280                 .maxlen   = sizeof(int),
 281                 .mode     = 0444,
 282                 .proc_handler = &proc_dointvec
 283         },
 284 #if IBLND_MAP_ON_DEMAND
 285         {
 286                 .ctl_name = O2IBLND_FMR_POOL_SIZE,
 287                 .procname = "fmr_pool_size",
 288                 .data     = &fmr_pool_size,
 289                 .maxlen   = sizeof(int),
 290                 .mode     = 0444,
 291                 .proc_handler = &proc_dointvec
 292         },
 293         {
 294                 .ctl_name = O2IBLND_FMR_FLUSH_TRIGGER,
 295                 .procname = "fmr_flush_trigger",
 296                 .data     = &fmr_flush_trigger,
 297                 .maxlen   = sizeof(int),
 298                 .mode     = 0444,
 299                 .proc_handler = &proc_dointvec
 300         },
 301         {
 302                 .ctl_name = O2IBLND_FMR_CACHE,
 303                 .procname = "fmr_cache",
 304                 .data     = &fmr_cache,
 305                 .maxlen   = sizeof(int),
 306                 .mode     = 0444,
 307                 .proc_handler = &proc_dointvec
 308         },
 309 #endif
 310         {0}
 311 };
 312
 313 static cfs_sysctl_table_t kiblnd_top_ctl_table[] = {
 314         {
 315                 .ctl_name = CTL_O2IBLND,
 316                 .procname = "o2iblnd",
 317                 .data     = NULL,
 318                 .maxlen   = 0,
 319                 .mode     = 0555,
 320                 .child    = kiblnd_ctl_table
 321         },
 322         {0}
 323 };
 324
 325 void
 326 kiblnd_initstrtunable(char *space, char *str, int size)
 327 {
 328         strncpy(space, str, size);
 329         space[size-1] = 0;
 330 }
 331
 332 void
 333 kiblnd_sysctl_init (void)
 334 {
 335         kiblnd_initstrtunable(ipif_basename_space, ipif_name,
 336                               sizeof(ipif_basename_space));
 337
 338         kiblnd_tunables.kib_sysctl =
 339                 cfs_register_sysctl_table(kiblnd_top_ctl_table, 0);
 340
 341         if (kiblnd_tunables.kib_sysctl == NULL)
 342                 CWARN("Can't setup /proc tunables\n");
 343 }
 344
 345 void
 346 kiblnd_sysctl_fini (void)
 347 {
 348         if (kiblnd_tunables.kib_sysctl != NULL)
 349                 cfs_unregister_sysctl_table(kiblnd_tunables.kib_sysctl);
 350 }
 351
 352 #else
 353
 354 void
 355 kiblnd_sysctl_init (void)
 356 {
 357 }
 358
 359 void
 360 kiblnd_sysctl_fini (void)
 361 {
 362 }
 363
 364 #endif
 365
 366 int
 367 kiblnd_tunables_init (void)
 368 {
 369         if (*kiblnd_tunables.kib_credits > *kiblnd_tunables.kib_ntx) {
 370                 CERROR("Can't set credits(%d) > ntx(%d)\n",
 371                        *kiblnd_tunables.kib_credits,
 372                        *kiblnd_tunables.kib_ntx);
 373                 return -EINVAL;
 374         }
 375
 376         if (*kiblnd_tunables.kib_ib_mtu != 0 &&
 377             *kiblnd_tunables.kib_ib_mtu != 256 &&
 378             *kiblnd_tunables.kib_ib_mtu != 512 &&
 379             *kiblnd_tunables.kib_ib_mtu != 1024 &&
 380             *kiblnd_tunables.kib_ib_mtu != 2048 &&
 381             *kiblnd_tunables.kib_ib_mtu != 4096) {
 382                 CERROR("Invalid ib_mtu %d, expected 256/512/1024/2048/4096\n",
 383                        *kiblnd_tunables.kib_ib_mtu);
 384                 return -EINVAL;
 385         }
 386
 387         if (*kiblnd_tunables.kib_concurrent_sends > IBLND_RX_MSGS)
 388                 *kiblnd_tunables.kib_concurrent_sends = IBLND_RX_MSGS;
 389         if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE / 2)
 390                 *kiblnd_tunables.kib_concurrent_sends = IBLND_MSG_QUEUE_SIZE / 2;
 391
 392         if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE) {
 393                 CWARN("Concurrent sends %d is lower than message queue size: %d, "
 394                       "performance may drop slightly.\n",
 395                       *kiblnd_tunables.kib_concurrent_sends, IBLND_MSG_QUEUE_SIZE);
 396         }
 397
 398         kiblnd_sysctl_init();
 399         return 0;
 400 }
 401
 402 void
 403 kiblnd_tunables_fini (void)
 404 {
 405         kiblnd_sysctl_fini();
 406 }