2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Derived from work by: Eric Barton <eric@bartonsoftware.com>
7 * Author: Nic Henke <nic@cray.com>
9 * This file is part of Lustre, http://www.lustre.org.
11 * Lustre is free software; you can redistribute it and/or
12 * modify it under the terms of version 2 of the GNU General Public
13 * License as published by the Free Software Foundation.
15 * Lustre is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with Lustre; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 static int credits = 256;
29 CFS_MODULE_PARM(credits, "i", int, 0444,
30 "# concurrent sends");
32 static int eager_credits = 256 * 1024;
33 CFS_MODULE_PARM(eager_credits, "i", int, 0444,
36 static int peer_credits = 16;
37 CFS_MODULE_PARM(peer_credits, "i", int, 0444,
38 "# LNet peer credits");
40 /* NB - we'll not actually limit sends to this, we just size the mailbox buffer
41 * such that at most we'll have concurrent_sends * max_immediate messages
43 static int concurrent_sends = 0;
44 CFS_MODULE_PARM(concurrent_sends, "i", int, 0444,
45 "# concurrent HW sends to 1 peer");
47 /* default for 2k nodes @ 16 peer credits */
48 static int fma_cq_size = 32768;
49 CFS_MODULE_PARM(fma_cq_size, "i", int, 0444,
50 "size of the completion queue");
52 static int timeout = GNILND_BASE_TIMEOUT;
53 /* can't change @ runtime because LNet gets NI data at startup from
55 CFS_MODULE_PARM(timeout, "i", int, 0444,
56 "communications timeout (seconds)");
58 /* time to wait between datagram timeout and sending of next dgram */
59 static int min_reconnect_interval = GNILND_MIN_RECONNECT_TO;
60 CFS_MODULE_PARM(min_reconnect_interval, "i", int, 0644,
61 "minimum connection retry interval (seconds)");
63 /* if this goes longer than timeout, we'll timeout the TX before
65 static int max_reconnect_interval = GNILND_MAX_RECONNECT_TO;
66 CFS_MODULE_PARM(max_reconnect_interval, "i", int, 0644,
67 "maximum connection retry interval (seconds)");
69 static int max_immediate = 8192;
70 CFS_MODULE_PARM(max_immediate, "i", int, 0644,
71 "immediate/RDMA breakpoint");
73 static int checksum = GNILND_CHECKSUM_DEFAULT;
74 CFS_MODULE_PARM(checksum, "i", int, 0644,
75 "0: None, 1: headers, 2: short msg, 3: all traffic");
77 static int checksum_dump = 0;
78 CFS_MODULE_PARM(checksum_dump, "i", int, 0644,
79 "0: None, 1: dump log on failure, 2: payload data to D_INFO log");
81 static int bte_dlvr_mode = GNILND_RDMA_DLVR_OPTION;
82 CFS_MODULE_PARM(bte_dlvr_mode, "i", int, 0644,
83 "enable hashing for BTE (RDMA) transfers");
85 static int bte_relaxed_ordering = 1;
86 CFS_MODULE_PARM(bte_relaxed_ordering, "i", int, 0644,
87 "enable relaxed ordering (PASSPW) for BTE (RDMA) transfers");
90 static int ptag = GNI_PTAG_LND_KNC;
92 static int ptag = GNI_PTAG_LND;
94 CFS_MODULE_PARM(ptag, "i", int, 0444,
95 "ptag for Gemini CDM");
97 static int pkey = GNI_JOB_CREATE_COOKIE(GNI_PKEY_LND, 0);
98 CFS_MODULE_PARM(pkey, "i", int, 0444, "pkey for CDM");
100 static int max_retransmits = 1024;
101 CFS_MODULE_PARM(max_retransmits, "i", int, 0444,
102 "max retransmits for FMA");
104 static int nwildcard = 4;
105 CFS_MODULE_PARM(nwildcard, "i", int, 0444,
106 "# wildcard datagrams to post per net (interface)");
108 static int nice = -20;
109 CFS_MODULE_PARM(nice, "i", int, 0444,
110 "nice value for kgnilnd threads, default -20");
112 static int rdmaq_intervals = 4;
113 CFS_MODULE_PARM(rdmaq_intervals, "i", int, 0644,
114 "# intervals per second for rdmaq throttling, default 4, 0 to disable");
116 static int loops = 100;
117 CFS_MODULE_PARM(loops, "i", int, 0644,
118 "# of loops before scheduler is friendly, default 100");
120 static int hash_size = 503;
121 CFS_MODULE_PARM(hash_size, "i", int, 0444,
122 "prime number for peer/conn hash sizing, default 503");
124 static int peer_health = 0;
125 CFS_MODULE_PARM(peer_health, "i", int, 0444,
126 "Disable peer timeout for LNet peer health, default off, > 0 to enable");
128 static int peer_timeout = -1;
129 CFS_MODULE_PARM(peer_timeout, "i", int, 0444,
130 "Peer timeout used for peer_health, default based on gnilnd timeout, > -1 to manually set");
132 static int vmap_cksum = 0;
133 CFS_MODULE_PARM(vmap_cksum, "i", int, 0644,
134 "use vmap for all kiov checksumming, default off");
136 static int mbox_per_block = GNILND_FMABLK;
137 CFS_MODULE_PARM(mbox_per_block, "i", int, 0644,
138 "mailboxes per block");
140 static int nphys_mbox = 0;
141 CFS_MODULE_PARM(nphys_mbox, "i", int, 0444,
142 "# mbox to preallocate from physical memory, default 0");
144 static int mbox_credits = GNILND_MBOX_CREDITS;
145 CFS_MODULE_PARM(mbox_credits, "i", int, 0644,
146 "number of credits per mailbox");
148 static int sched_threads = GNILND_SCHED_THREADS;
149 CFS_MODULE_PARM(sched_threads, "i", int, 0444,
150 "number of threads for moving data");
152 static int net_hash_size = 11;
153 CFS_MODULE_PARM(net_hash_size, "i", int, 0444,
154 "prime number for net hash sizing, default 11");
156 static int hardware_timeout = GNILND_HARDWARE_TIMEOUT;
157 CFS_MODULE_PARM(hardware_timeout, "i", int, 0444,
158 "maximum time for traffic to get from one node to another");
160 static int mdd_timeout = GNILND_MDD_TIMEOUT;
161 CFS_MODULE_PARM(mdd_timeout, "i", int, 0644,
162 "maximum time (in minutes) for mdd to be held");
164 static int sched_timeout = GNILND_SCHED_TIMEOUT;
165 CFS_MODULE_PARM(sched_timeout, "i", int, 0644,
166 "scheduler aliveness in seconds max time");
168 static int sched_nice = GNILND_SCHED_NICE;
169 CFS_MODULE_PARM(sched_nice, "i", int, 0444,
170 "scheduler's nice setting, default compute 0 service -20");
172 static int reverse_rdma = GNILND_REVERSE_RDMA;
173 CFS_MODULE_PARM(reverse_rdma, "i", int, 0644,
174 "Normal 0: Reverse GET: 1 Reverse Put: 2 Reverse Both: 3");
176 static int dgram_timeout = GNILND_DGRAM_TIMEOUT;
177 CFS_MODULE_PARM(dgram_timeout, "i", int, 0644,
178 "dgram thread aliveness seconds max time");
180 static int efault_lbug = 0;
181 CFS_MODULE_PARM(efault_lbug, "i", int, 0644,
182 "If a compute receives an EFAULT in"
183 " a message should it LBUG. 0 off 1 on");
185 static int fast_reconn = GNILND_FAST_RECONNECT;
186 CFS_MODULE_PARM(fast_reconn, "i", int, 0644,
187 "fast reconnect on connection timeout");
189 static int max_conn_purg = GNILND_PURGATORY_MAX;
190 CFS_MODULE_PARM(max_conn_purg, "i", int, 0644,
191 "Max number of connections per peer in purgatory");
193 static int thread_affinity = 0;
194 CFS_MODULE_PARM(thread_affinity, "i", int, 0444,
195 "scheduler thread affinity default 0 (diabled)");
197 kgn_tunables_t kgnilnd_tunables = {
198 .kgn_min_reconnect_interval = &min_reconnect_interval,
199 .kgn_max_reconnect_interval = &max_reconnect_interval,
200 .kgn_credits = &credits,
201 .kgn_peer_credits = &peer_credits,
202 .kgn_concurrent_sends = &concurrent_sends,
203 .kgn_fma_cq_size = &fma_cq_size,
204 .kgn_timeout = &timeout,
205 .kgn_max_immediate = &max_immediate,
206 .kgn_checksum = &checksum,
207 .kgn_checksum_dump = &checksum_dump,
208 .kgn_bte_dlvr_mode = &bte_dlvr_mode,
209 .kgn_bte_relaxed_ordering = &bte_relaxed_ordering,
212 .kgn_max_retransmits = &max_retransmits,
213 .kgn_nwildcard = &nwildcard,
215 .kgn_rdmaq_intervals = &rdmaq_intervals,
217 .kgn_peer_hash_size = &hash_size,
218 .kgn_peer_health = &peer_health,
219 .kgn_peer_timeout = &peer_timeout,
220 .kgn_vmap_cksum = &vmap_cksum,
221 .kgn_mbox_per_block = &mbox_per_block,
222 .kgn_nphys_mbox = &nphys_mbox,
223 .kgn_mbox_credits = &mbox_credits,
224 .kgn_sched_threads = &sched_threads,
225 .kgn_net_hash_size = &net_hash_size,
226 .kgn_hardware_timeout = &hardware_timeout,
227 .kgn_mdd_timeout = &mdd_timeout,
228 .kgn_sched_timeout = &sched_timeout,
229 .kgn_sched_nice = &sched_nice,
230 .kgn_reverse_rdma = &reverse_rdma,
231 .kgn_dgram_timeout = &dgram_timeout,
232 .kgn_eager_credits = &eager_credits,
233 .kgn_fast_reconn = &fast_reconn,
234 .kgn_efault_lbug = &efault_lbug,
235 .kgn_thread_affinity = &thread_affinity,
236 .kgn_max_purgatory = &max_conn_purg
239 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
240 static struct ctl_table kgnilnd_ctl_table[] = {
243 .procname = "min_reconnect_interval",
244 .data = &min_reconnect_interval,
245 .maxlen = sizeof(int),
247 .proc_handler = &proc_dointvec
251 .procname = "max_reconnect_interval",
252 .data = &max_reconnect_interval,
253 .maxlen = sizeof(int),
255 .proc_handler = &proc_dointvec
259 .procname = "credits",
261 .maxlen = sizeof(int),
263 .proc_handler = &proc_dointvec
267 .procname = "peer_credits",
268 .data = &peer_credits,
269 .maxlen = sizeof(int),
271 .proc_handler = &proc_dointvec
275 .procname = "fma_cq_size",
276 .data = &fma_cq_size,
277 .maxlen = sizeof(int),
279 .proc_handler = &proc_dointvec
283 .procname = "timeout",
285 .maxlen = sizeof(int),
287 .proc_handler = &proc_dointvec
291 .procname = "max_immediate",
292 .data = &max_immediate,
293 .maxlen = sizeof(int),
295 .proc_handler = &proc_dointvec
299 .procname = "checksum",
301 .maxlen = sizeof(int),
303 .proc_handler = &proc_dointvec
307 .procname = "bte_dlvr_mode",
308 .data = &bte_dlvr_mode,
309 .maxlen = sizeof(int),
311 .proc_handler = &proc_dointvec
317 .maxlen = sizeof(int),
319 .proc_handler = &proc_dointvec
325 .maxlen = sizeof(int),
327 .proc_handler = &proc_dointvec
331 .procname = "nwildcard",
333 .maxlen = sizeof(int),
335 .proc_handler = &proc_dointvec
339 .procname = "bte_relaxed_ordering",
340 .data = &bte_relaxed_ordering,
341 .maxlen = sizeof(int),
343 .proc_handler = &proc_dointvec
347 .procname = "checksum_dump",
348 .data = &checksum_dump,
349 .maxlen = sizeof(int),
351 .proc_handler = &proc_dointvec
357 .maxlen = sizeof(int),
359 .proc_handler = &proc_dointvec
363 .procname = "rdmaq_intervals",
364 .data = &rdmaq_intervals,
365 .maxlen = sizeof(int),
367 .proc_handler = &proc_dointvec
373 .maxlen = sizeof(int),
375 .proc_handler = &proc_dointvec
379 .procname = "hash_size",
381 .maxlen = sizeof(int),
383 .proc_handler = &proc_dointvec
387 .procname = "peer_health",
388 .data = &peer_health,
389 .maxlen = sizeof(int),
391 .proc_handler = &proc_dointvec
395 .procname = "vmap_cksum",
397 .maxlen = sizeof(int),
399 .proc_handler = &proc_dointvec
403 .procname = "mbox_per_block",
404 .data = &mbox_per_block,
405 .maxlen = sizeof(int),
407 .proc_handler = &proc_dointvec
411 .procname = "mbox_credits"
412 .data = &mbox_credits,
413 .maxlen = sizeof(int),
415 .proc_handler = &proc_dointvec
419 .procname = "sched_threads"
420 .data = &sched_threads,
421 .maxlen = sizeof(int),
423 .proc_handler = &proc_dointvec
427 .procname = "net_hash_size",
428 .data = &net_hash_size,
429 .maxlen = sizeof(int),
431 .proc_handler = &proc_dointvec
435 .procname = "hardware_timeout",
436 .data = &hardware_timeout,
437 .maxlen = sizeof(int),
439 .proc_handler = &proc_dointvec
443 .procname = "mdd_timeout",
444 .data = &mdd_timeout,
445 .maxlen = sizeof(int),
447 .proc_handler = &proc_dointvec
451 .procname = "max_retransmits"
452 .data = &max_retransmits,
453 .maxlen = sizeof(int),
455 .proc_handler = &proc_dointvec
459 .procname = "concurrent_sends",
460 .data = &concurrent_sends,
461 .maxlen = sizeof(int),
463 .proc_handler = &proc_dointvec
467 .procname = "nphys_mbox",
469 .maxlen = sizeof(int),
471 .proc_handler = &proc_dointvec
475 .procname = "sched_timeout",
476 .data = &sched_timeout,
477 .maxlen = sizeof(int),
479 .proc_handler = &proc_dointvec
483 .procname = "sched_nice",
485 .maxlen = sizeof(int),
487 .proc_handler = &proc_dointvec
491 .procname = "reverse_rdma",
492 .data = &reverse_rdma,
493 .maxlen = sizeof(int),
495 .proc_handler = &proc_dointvec
498 .procname = "dgram_timeout"
499 .data = &dgram_timeout,
500 .maxlen = sizeof(int),
502 .proc_handler = &proc_dointvec
506 .procname = "peer_timeout"
507 .data = &peer_timeout,
508 .maxlen = sizeof(int),
510 .proc_handler = &proc_dointvec
514 .procname = "eager_credits",
515 .data = &eager_credits,
516 .maxlen = sizeof(int),
518 .proc_handler = &proc_dointvec
522 .procname = "efault_lbug"
523 .data = &efault_lbug,
524 .maxlen = sizeof(int),
526 .proc_handler = &proc_dointvec
530 .procname = "thread_affinity"
531 .data = &thread_affinity,
532 .maxlen = sizeof(int),
534 .proc_handler = &proc_dointvec
538 .procname = "max_conn_purg"
539 .data = &max_conn_purg,
540 .maxlen = sizeof(int),
542 .proc_handler = &proc_dointvec
547 static struct ctl_table kgnilnd_top_ctl_table[] = {
550 .procname = "gnilnd",
554 .child = kgnilnd_ctl_table
561 kgnilnd_tunables_init()
565 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
566 kgnilnd_tunables.kgn_sysctl =
567 cfs_register_sysctl_table(kgnilnd_top_ctl_table, 0);
569 if (kgnilnd_tunables.kgn_sysctl == NULL)
570 CWARN("Can't setup /proc tunables\n");
572 switch (*kgnilnd_tunables.kgn_checksum) {
574 CERROR("Invalid checksum module parameter: %d\n",
575 *kgnilnd_tunables.kgn_checksum);
578 case GNILND_CHECKSUM_OFF:
579 /* no checksumming */
581 case GNILND_CHECKSUM_SMSG_HEADER:
582 LCONSOLE_INFO("SMSG header only checksumming enabled\n");
584 case GNILND_CHECKSUM_SMSG:
585 LCONSOLE_INFO("SMSG checksumming enabled\n");
587 case GNILND_CHECKSUM_SMSG_BTE:
588 LCONSOLE_INFO("SMSG + BTE checksumming enabled\n");
592 if (*kgnilnd_tunables.kgn_max_immediate > GNILND_MAX_IMMEDIATE) {
593 LCONSOLE_ERROR("kgnilnd module parameter 'max_immediate' too large %d > %d\n",
594 *kgnilnd_tunables.kgn_max_immediate, GNILND_MAX_IMMEDIATE);
599 if (*kgnilnd_tunables.kgn_mbox_per_block < 1) {
600 *kgnilnd_tunables.kgn_mbox_per_block = 1;
603 if (*kgnilnd_tunables.kgn_concurrent_sends == 0) {
604 *kgnilnd_tunables.kgn_concurrent_sends = *kgnilnd_tunables.kgn_peer_credits;
605 } else if (*kgnilnd_tunables.kgn_concurrent_sends > *kgnilnd_tunables.kgn_peer_credits) {
606 LCONSOLE_ERROR("kgnilnd parameter 'concurrent_sends' too large: %d > %d (peer_credits)\n",
607 *kgnilnd_tunables.kgn_concurrent_sends, *kgnilnd_tunables.kgn_peer_credits);
615 kgnilnd_tunables_fini()
617 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
618 if (kgnilnd_tunables.kgn_sysctl != NULL)
619 cfs_unregister_sysctl_table(kgnilnd_tunables.kgn_sysctl);