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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/o2iblnd/o2iblnd_modparams.c
  33  *
  34  * Author: Eric Barton <eric@bartonsoftware.com>
  35  */
  36
  37 #include "o2iblnd.h"
  38
  39 #define CURRENT_LND_VERSION 1
  40
  41 static int service = 987;
  42 module_param(service, int, 0444);
  43 MODULE_PARM_DESC(service, "service number (within RDMA_PS_TCP)");
  44
  45 static int cksum = 0;
  46 module_param(cksum, int, 0644);
  47 MODULE_PARM_DESC(cksum, "set non-zero to enable message (not RDMA) checksums");
  48
  49 static int timeout = 50;
  50 module_param(timeout, int, 0644);
  51 MODULE_PARM_DESC(timeout, "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 module_param(nscheds, int, 0444);
  57 MODULE_PARM_DESC(nscheds, "number of threads in each scheduler pool");
  58
  59 static unsigned int conns_per_peer = 1;
  60 module_param(conns_per_peer, uint, 0444);
  61 MODULE_PARM_DESC(conns_per_peer, "number of connections per peer");
  62
  63 /* NB: this value is shared by all CPTs, it can grow at runtime */
  64 static int ntx = 512;
  65 module_param(ntx, int, 0444);
  66 MODULE_PARM_DESC(ntx, "# of message descriptors allocated for each pool");
  67
  68 /* NB: this value is shared by all CPTs */
  69 static int credits = 256;
  70 module_param(credits, int, 0444);
  71 MODULE_PARM_DESC(credits, "# concurrent sends");
  72
  73 static int peer_credits = 8;
  74 module_param(peer_credits, int, 0444);
  75 MODULE_PARM_DESC(peer_credits, "# concurrent sends to 1 peer");
  76
  77 static int peer_credits_hiw = 0;
  78 module_param(peer_credits_hiw, int, 0444);
  79 MODULE_PARM_DESC(peer_credits_hiw, "when eagerly to return credits");
  80
  81 static int peer_buffer_credits = 0;
  82 module_param(peer_buffer_credits, int, 0444);
  83 MODULE_PARM_DESC(peer_buffer_credits, "# per-peer router buffer credits");
  84
  85 static int peer_timeout = 180;
  86 module_param(peer_timeout, int, 0444);
  87 MODULE_PARM_DESC(peer_timeout, "Seconds without aliveness news to declare peer dead (<=0 to disable)");
  88
  89 static char *ipif_name = "ib0";
  90 module_param(ipif_name, charp, 0444);
  91 MODULE_PARM_DESC(ipif_name, "IPoIB interface name");
  92
  93 static int retry_count = 5;
  94 module_param(retry_count, int, 0644);
  95 MODULE_PARM_DESC(retry_count, "Retransmissions when no ACK received");
  96
  97 static int rnr_retry_count = 6;
  98 module_param(rnr_retry_count, int, 0644);
  99 MODULE_PARM_DESC(rnr_retry_count, "RNR retransmissions");
 100
 101 static int keepalive = 100;
 102 module_param(keepalive, int, 0644);
 103 MODULE_PARM_DESC(keepalive, "Idle time in seconds before sending a keepalive");
 104
 105 static int ib_mtu;
 106 module_param(ib_mtu, int, 0444);
 107 MODULE_PARM_DESC(ib_mtu, "IB MTU 256/512/1024/2048/4096");
 108
 109 static int concurrent_sends;
 110 module_param(concurrent_sends, int, 0444);
 111 MODULE_PARM_DESC(concurrent_sends, "send work-queue sizing");
 112
 113 #ifdef HAVE_IB_GET_DMA_MR
 114 #define IBLND_DEFAULT_MAP_ON_DEMAND 0
 115 #define IBLND_MIN_MAP_ON_DEMAND 0
 116 #else
 117 #define IBLND_DEFAULT_MAP_ON_DEMAND IBLND_MAX_RDMA_FRAGS
 118 #define IBLND_MIN_MAP_ON_DEMAND 1
 119 #endif
 120 static int map_on_demand = IBLND_DEFAULT_MAP_ON_DEMAND;
 121 module_param(map_on_demand, int, 0444);
 122 MODULE_PARM_DESC(map_on_demand, "map on demand");
 123
 124 /* NB: this value is shared by all CPTs, it can grow at runtime */
 125 static int fmr_pool_size = 512;
 126 module_param(fmr_pool_size, int, 0444);
 127 MODULE_PARM_DESC(fmr_pool_size, "size of fmr pool on each CPT (>= ntx / 4)");
 128
 129 /* NB: this value is shared by all CPTs, it can grow at runtime */
 130 static int fmr_flush_trigger = 384;
 131 module_param(fmr_flush_trigger, int, 0444);
 132 MODULE_PARM_DESC(fmr_flush_trigger, "# dirty FMRs that triggers pool flush");
 133
 134 static int fmr_cache = 1;
 135 module_param(fmr_cache, int, 0444);
 136 MODULE_PARM_DESC(fmr_cache, "non-zero to enable FMR caching");
 137
 138 /*
 139  * 0: disable failover
 140  * 1: enable failover if necessary
 141  * 2: force to failover (for debug)
 142  */
 143 static int dev_failover = 0;
 144 module_param(dev_failover, int, 0444);
 145 MODULE_PARM_DESC(dev_failover, "HCA failover for bonding (0 off, 1 on, other values reserved)");
 146
 147 static int require_privileged_port;
 148 module_param(require_privileged_port, int, 0644);
 149 MODULE_PARM_DESC(require_privileged_port, "require privileged port when accepting connection");
 150
 151 static int use_privileged_port = 1;
 152 module_param(use_privileged_port, int, 0644);
 153 MODULE_PARM_DESC(use_privileged_port, "use privileged port when initiating connection");
 154
 155 static unsigned int wrq_sge = 1;
 156 module_param(wrq_sge, uint, 0444);
 157 MODULE_PARM_DESC(wrq_sge, "# scatter/gather element per work request");
 158
 159 kib_tunables_t kiblnd_tunables = {
 160         .kib_dev_failover           = &dev_failover,
 161         .kib_service                = &service,
 162         .kib_cksum                  = &cksum,
 163         .kib_timeout                = &timeout,
 164         .kib_keepalive              = &keepalive,
 165         .kib_ntx                    = &ntx,
 166         .kib_default_ipif           = &ipif_name,
 167         .kib_retry_count            = &retry_count,
 168         .kib_rnr_retry_count        = &rnr_retry_count,
 169         .kib_ib_mtu                 = &ib_mtu,
 170         .kib_require_priv_port      = &require_privileged_port,
 171         .kib_use_priv_port          = &use_privileged_port,
 172         .kib_nscheds                = &nscheds,
 173         .kib_wrq_sge                = &wrq_sge,
 174 };
 175
 176 static struct lnet_ioctl_config_o2iblnd_tunables default_tunables;
 177
 178 /* # messages/RDMAs in-flight */
 179 int
 180 kiblnd_msg_queue_size(int version, struct lnet_ni *ni)
 181 {
 182         if (version == IBLND_MSG_VERSION_1)
 183                 return IBLND_MSG_QUEUE_SIZE_V1;
 184         else if (ni)
 185                 return ni->ni_net->net_tunables.lct_peer_tx_credits;
 186         else
 187                 return peer_credits;
 188 }
 189
 190 int
 191 kiblnd_tunables_setup(struct lnet_ni *ni)
 192 {
 193         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
 194         struct lnet_ioctl_config_lnd_cmn_tunables *net_tunables;
 195
 196         /*
 197          * if there was no tunables specified, setup the tunables to be
 198          * defaulted
 199          */
 200         if (!ni->ni_lnd_tunables_set)
 201                 memcpy(&ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib,
 202                        &default_tunables, sizeof(*tunables));
 203
 204         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
 205
 206         /* Current API version */
 207         tunables->lnd_version = CURRENT_LND_VERSION;
 208
 209         if (kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu) < 0) {
 210                 CERROR("Invalid ib_mtu %d, expected 256/512/1024/2048/4096\n",
 211                        *kiblnd_tunables.kib_ib_mtu);
 212                 return -EINVAL;
 213         }
 214
 215         net_tunables = &ni->ni_net->net_tunables;
 216
 217         if (net_tunables->lct_peer_timeout == -1)
 218                 net_tunables->lct_peer_timeout = peer_timeout;
 219
 220         if (net_tunables->lct_max_tx_credits == -1)
 221                 net_tunables->lct_max_tx_credits = credits;
 222
 223         if (net_tunables->lct_peer_tx_credits == -1)
 224                 net_tunables->lct_peer_tx_credits = peer_credits;
 225
 226         if (net_tunables->lct_peer_rtr_credits == -1)
 227                 net_tunables->lct_peer_rtr_credits = peer_buffer_credits;
 228
 229         if (net_tunables->lct_peer_tx_credits < IBLND_CREDITS_DEFAULT)
 230                 net_tunables->lct_peer_tx_credits = IBLND_CREDITS_DEFAULT;
 231
 232         if (net_tunables->lct_peer_tx_credits > IBLND_CREDITS_MAX)
 233                 net_tunables->lct_peer_tx_credits = IBLND_CREDITS_MAX;
 234
 235         if (net_tunables->lct_peer_tx_credits >
 236             net_tunables->lct_max_tx_credits)
 237                 net_tunables->lct_peer_tx_credits =
 238                         net_tunables->lct_max_tx_credits;
 239
 240         if (!tunables->lnd_peercredits_hiw)
 241                 tunables->lnd_peercredits_hiw = peer_credits_hiw;
 242
 243         if (tunables->lnd_peercredits_hiw < net_tunables->lct_peer_tx_credits / 2)
 244                 tunables->lnd_peercredits_hiw = net_tunables->lct_peer_tx_credits / 2;
 245
 246         if (tunables->lnd_peercredits_hiw >= net_tunables->lct_peer_tx_credits)
 247                 tunables->lnd_peercredits_hiw = net_tunables->lct_peer_tx_credits - 1;
 248
 249         if (tunables->lnd_map_on_demand < IBLND_MIN_MAP_ON_DEMAND ||
 250             tunables->lnd_map_on_demand > IBLND_MAX_RDMA_FRAGS) {
 251                 /* Use the default */
 252                 CWARN("Invalid map_on_demand (%d), expects %d - %d. Using default of %d\n",
 253                       tunables->lnd_map_on_demand, IBLND_MIN_MAP_ON_DEMAND,
 254                       IBLND_MAX_RDMA_FRAGS, IBLND_DEFAULT_MAP_ON_DEMAND);
 255                 tunables->lnd_map_on_demand = IBLND_DEFAULT_MAP_ON_DEMAND;
 256         }
 257
 258         if (tunables->lnd_map_on_demand == 1) {
 259                 /* don't make sense to create map if only one fragment */
 260                 tunables->lnd_map_on_demand = 2;
 261         }
 262
 263         if (tunables->lnd_concurrent_sends == 0) {
 264                 if (tunables->lnd_map_on_demand > 0 &&
 265                     tunables->lnd_map_on_demand <= IBLND_MAX_RDMA_FRAGS / 8) {
 266                         tunables->lnd_concurrent_sends =
 267                                         net_tunables->lct_peer_tx_credits * 2;
 268                 } else {
 269                         tunables->lnd_concurrent_sends =
 270                                 net_tunables->lct_peer_tx_credits;
 271                 }
 272         }
 273
 274         if (tunables->lnd_concurrent_sends > net_tunables->lct_peer_tx_credits * 2)
 275                 tunables->lnd_concurrent_sends = net_tunables->lct_peer_tx_credits * 2;
 276
 277         if (tunables->lnd_concurrent_sends < net_tunables->lct_peer_tx_credits / 2)
 278                 tunables->lnd_concurrent_sends = net_tunables->lct_peer_tx_credits / 2;
 279
 280         if (tunables->lnd_concurrent_sends < net_tunables->lct_peer_tx_credits) {
 281                 CWARN("Concurrent sends %d is lower than message "
 282                       "queue size: %d, performance may drop slightly.\n",
 283                       tunables->lnd_concurrent_sends,
 284                       net_tunables->lct_peer_tx_credits);
 285         }
 286
 287         if (!tunables->lnd_fmr_pool_size)
 288                 tunables->lnd_fmr_pool_size = fmr_pool_size;
 289         if (!tunables->lnd_fmr_flush_trigger)
 290                 tunables->lnd_fmr_flush_trigger = fmr_flush_trigger;
 291         if (!tunables->lnd_fmr_cache)
 292                 tunables->lnd_fmr_cache = fmr_cache;
 293         if (!tunables->lnd_conns_per_peer) {
 294                 tunables->lnd_conns_per_peer = (conns_per_peer) ?
 295                         conns_per_peer : 1;
 296         }
 297
 298         return 0;
 299 }
 300
 301 int
 302 kiblnd_tunables_init(void)
 303 {
 304         default_tunables.lnd_version = CURRENT_LND_VERSION;
 305         default_tunables.lnd_peercredits_hiw = peer_credits_hiw,
 306         default_tunables.lnd_map_on_demand = map_on_demand;
 307         default_tunables.lnd_concurrent_sends = concurrent_sends;
 308         default_tunables.lnd_fmr_pool_size = fmr_pool_size;
 309         default_tunables.lnd_fmr_flush_trigger = fmr_flush_trigger;
 310         default_tunables.lnd_fmr_cache = fmr_cache;
 311         default_tunables.lnd_conns_per_peer = conns_per_peer;
 312         return 0;
 313 }