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 peer_credits = 16;
33 CFS_MODULE_PARM(peer_credits, "i", int, 0444,
34 "# LNet peer credits");
36 /* NB - we'll not actually limit sends to this, we just size the mailbox buffer
37 * such that at most we'll have concurrent_sends * max_immediate messages
39 static int concurrent_sends = 0;
40 CFS_MODULE_PARM(concurrent_sends, "i", int, 0444,
41 "# concurrent HW sends to 1 peer");
43 /* default for 2k nodes @ 16 peer credits */
44 static int fma_cq_size = 32768;
45 CFS_MODULE_PARM(fma_cq_size, "i", int, 0444,
46 "size of the completion queue");
48 static int timeout = GNILND_BASE_TIMEOUT;
49 /* can't change @ runtime because LNet gets NI data at startup from
51 CFS_MODULE_PARM(timeout, "i", int, 0444,
52 "communications timeout (seconds)");
54 /* time to wait between datagram timeout and sending of next dgram */
55 static int min_reconnect_interval = GNILND_MIN_RECONNECT_TO;
56 CFS_MODULE_PARM(min_reconnect_interval, "i", int, 0644,
57 "minimum connection retry interval (seconds)");
59 /* if this goes longer than timeout, we'll timeout the TX before
61 static int max_reconnect_interval = GNILND_MAX_RECONNECT_TO;
62 CFS_MODULE_PARM(max_reconnect_interval, "i", int, 0644,
63 "maximum connection retry interval (seconds)");
65 static int max_immediate = (2<<10);
66 CFS_MODULE_PARM(max_immediate, "i", int, 0644,
67 "immediate/RDMA breakpoint");
69 static int checksum = GNILND_CHECKSUM_DEFAULT;
70 CFS_MODULE_PARM(checksum, "i", int, 0644,
71 "0: None, 1: headers, 2: short msg, 3: all traffic");
73 static int checksum_dump = 0;
74 CFS_MODULE_PARM(checksum_dump, "i", int, 0644,
75 "0: None, 1: dump log on failure, 2: payload data to D_INFO log");
77 static int bte_dlvr_mode = GNILND_RDMA_DLVR_OPTION;
78 CFS_MODULE_PARM(bte_dlvr_mode, "i", int, 0644,
79 "enable hashing for BTE (RDMA) transfers");
81 static int bte_relaxed_ordering = 1;
82 CFS_MODULE_PARM(bte_relaxed_ordering, "i", int, 0644,
83 "enable relaxed ordering (PASSPW) for BTE (RDMA) transfers");
85 static int ptag = GNI_PTAG_LND;
86 CFS_MODULE_PARM(ptag, "i", int, 0444,
87 "ptag for Gemini CDM");
89 static int max_retransmits = 1024;
90 CFS_MODULE_PARM(max_retransmits, "i", int, 0444,
91 "max retransmits for FMA");
93 static int nwildcard = 4;
94 CFS_MODULE_PARM(nwildcard, "i", int, 0444,
95 "# wildcard datagrams to post per net (interface)");
97 static int nice = -20;
98 CFS_MODULE_PARM(nice, "i", int, 0444,
99 "nice value for kgnilnd threads, default -20");
101 static int rdmaq_intervals = 4;
102 CFS_MODULE_PARM(rdmaq_intervals, "i", int, 0644,
103 "# intervals per second for rdmaq throttling, default 4, 0 to disable");
105 static int loops = 100;
106 CFS_MODULE_PARM(loops, "i", int, 0644,
107 "# of loops before scheduler is friendly, default 100");
109 static int hash_size = 503;
110 CFS_MODULE_PARM(hash_size, "i", int, 0444,
111 "prime number for peer/conn hash sizing, default 503");
113 static int peer_health = 0;
114 CFS_MODULE_PARM(peer_health, "i", int, 0444,
115 "Disable peer timeout for LNet peer health, default off, > 0 to enable");
117 static int peer_timeout = -1;
118 CFS_MODULE_PARM(peer_timeout, "i", int, 0444,
119 "Peer timeout used for peer_health, default based on gnilnd timeout, > -1 to manually set");
121 static int vmap_cksum = 0;
122 CFS_MODULE_PARM(vmap_cksum, "i", int, 0644,
123 "use vmap for all kiov checksumming, default off");
125 static int mbox_per_block = GNILND_FMABLK;
126 CFS_MODULE_PARM(mbox_per_block, "i", int, 0644,
127 "mailboxes per block");
129 static int nphys_mbox = 0;
130 CFS_MODULE_PARM(nphys_mbox, "i", int, 0444,
131 "# mbox to preallocate from physical memory, default 0");
133 static int mbox_credits = GNILND_MBOX_CREDITS;
134 CFS_MODULE_PARM(mbox_credits, "i", int, 0644,
135 "number of credits per mailbox");
137 static int sched_threads = GNILND_SCHED_THREADS;
138 CFS_MODULE_PARM(sched_threads, "i", int, 0444,
139 "number of threads for moving data");
141 static int net_hash_size = 11;
142 CFS_MODULE_PARM(net_hash_size, "i", int, 0444,
143 "prime number for net hash sizing, default 11");
145 static int hardware_timeout = GNILND_HARDWARE_TIMEOUT;
146 CFS_MODULE_PARM(hardware_timeout, "i", int, 0444,
147 "maximum time for traffic to get from one node to another");
149 static int mdd_timeout = GNILND_MDD_TIMEOUT;
150 CFS_MODULE_PARM(mdd_timeout, "i", int, 0644,
151 "maximum time (in minutes) for mdd to be held");
153 static int sched_timeout = GNILND_SCHED_TIMEOUT;
154 CFS_MODULE_PARM(sched_timeout, "i", int, 0644,
155 "scheduler aliveness in seconds max time");
157 static int sched_nice = GNILND_SCHED_NICE;
158 CFS_MODULE_PARM(sched_nice, "i", int, 0444,
159 "scheduler's nice setting, default compute 0 service -20");
161 static int reverse_rdma = GNILND_REVERSE_RDMA;
162 CFS_MODULE_PARM(reverse_rdma, "i", int, 0644,
163 "Normal 0: Reverse GET: 1 Reverse Put: 2 Reverse Both: 3");
165 static int dgram_timeout = GNILND_DGRAM_TIMEOUT;
166 CFS_MODULE_PARM(dgram_timeout, "i", int, 0644,
167 "dgram thread aliveness seconds max time");
169 kgn_tunables_t kgnilnd_tunables = {
170 .kgn_min_reconnect_interval = &min_reconnect_interval,
171 .kgn_max_reconnect_interval = &max_reconnect_interval,
172 .kgn_credits = &credits,
173 .kgn_peer_credits = &peer_credits,
174 .kgn_concurrent_sends = &concurrent_sends,
175 .kgn_fma_cq_size = &fma_cq_size,
176 .kgn_timeout = &timeout,
177 .kgn_max_immediate = &max_immediate,
178 .kgn_checksum = &checksum,
179 .kgn_checksum_dump = &checksum_dump,
180 .kgn_bte_dlvr_mode = &bte_dlvr_mode,
181 .kgn_bte_relaxed_ordering = &bte_relaxed_ordering,
183 .kgn_max_retransmits = &max_retransmits,
184 .kgn_nwildcard = &nwildcard,
186 .kgn_rdmaq_intervals = &rdmaq_intervals,
188 .kgn_peer_hash_size = &hash_size,
189 .kgn_peer_health = &peer_health,
190 .kgn_peer_timeout = &peer_timeout,
191 .kgn_vmap_cksum = &vmap_cksum,
192 .kgn_mbox_per_block = &mbox_per_block,
193 .kgn_nphys_mbox = &nphys_mbox,
194 .kgn_mbox_credits = &mbox_credits,
195 .kgn_sched_threads = &sched_threads,
196 .kgn_net_hash_size = &net_hash_size,
197 .kgn_hardware_timeout = &hardware_timeout,
198 .kgn_mdd_timeout = &mdd_timeout,
199 .kgn_sched_timeout = &sched_timeout,
200 .kgn_sched_nice = &sched_nice,
201 .kgn_reverse_rdma = &reverse_rdma,
202 .kgn_dgram_timeout = &dgram_timeout
205 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
206 static cfs_sysctl_table_t kgnilnd_ctl_table[] = {
209 .procname = "min_reconnect_interval",
210 .data = &min_reconnect_interval,
211 .maxlen = sizeof(int),
213 .proc_handler = &proc_dointvec
217 .procname = "max_reconnect_interval",
218 .data = &max_reconnect_interval,
219 .maxlen = sizeof(int),
221 .proc_handler = &proc_dointvec
225 .procname = "credits",
227 .maxlen = sizeof(int),
229 .proc_handler = &proc_dointvec
233 .procname = "peer_credits",
234 .data = &peer_credits,
235 .maxlen = sizeof(int),
237 .proc_handler = &proc_dointvec
241 .procname = "fma_cq_size",
242 .data = &fma_cq_size,
243 .maxlen = sizeof(int),
245 .proc_handler = &proc_dointvec
249 .procname = "timeout",
251 .maxlen = sizeof(int),
253 .proc_handler = &proc_dointvec
257 .procname = "max_immediate",
258 .data = &max_immediate,
259 .maxlen = sizeof(int),
261 .proc_handler = &proc_dointvec
265 .procname = "checksum",
267 .maxlen = sizeof(int),
269 .proc_handler = &proc_dointvec
273 .procname = "bte_dlvr_mode",
274 .data = &bte_dlvr_mode,
275 .maxlen = sizeof(int),
277 .proc_handler = &proc_dointvec
283 .maxlen = sizeof(int),
285 .proc_handler = &proc_dointvec
289 .procname = "nwildcard",
291 .maxlen = sizeof(int),
293 .proc_handler = &proc_dointvec
297 .procname = "bte_relaxed_ordering",
298 .data = &bte_relaxed_ordering,
299 .maxlen = sizeof(int),
301 .proc_handler = &proc_dointvec
305 .procname = "checksum_dump",
306 .data = &checksum_dump,
307 .maxlen = sizeof(int),
309 .proc_handler = &proc_dointvec
315 .maxlen = sizeof(int),
317 .proc_handler = &proc_dointvec
321 .procname = "rdmaq_intervals",
322 .data = &rdmaq_intervals,
323 .maxlen = sizeof(int),
325 .proc_handler = &proc_dointvec
331 .maxlen = sizeof(int),
333 .proc_handler = &proc_dointvec
337 .procname = "hash_size",
339 .maxlen = sizeof(int),
341 .proc_handler = &proc_dointvec
345 .procname = "peer_health",
346 .data = &peer_health,
347 .maxlen = sizeof(int),
349 .proc_handler = &proc_dointvec
353 .procname = "vmap_cksum",
355 .maxlen = sizeof(int),
357 .proc_handler = &proc_dointvec
361 .procname = "mbox_per_block",
362 .data = &mbox_per_block,
363 .maxlen = sizeof(int),
365 .proc_handler = &proc_dointvec
369 .procname = "mbox_credits"
370 .data = &mbox_credits,
371 .maxlen = sizeof(int),
373 .proc_handler = &proc_dointvec
377 .procname = "sched_threads"
378 .data = &sched_threads,
379 .maxlen = sizeof(int),
381 .proc_handler = &proc_dointvec
385 .procname = "net_hash_size",
386 .data = &net_hash_size,
387 .maxlen = sizeof(int),
389 .proc_handler = &proc_dointvec
393 .procname = "hardware_timeout",
394 .data = &hardware_timeout,
395 .maxlen = sizeof(int),
397 .proc_handler = &proc_dointvec
401 .procname = "mdd_timeout",
402 .data = &mdd_timeout,
403 .maxlen = sizeof(int),
405 .proc_handler = &proc_dointvec
409 .procname = "max_retransmits"
410 .data = &max_retransmits,
411 .maxlen = sizeof(int),
413 .proc_handler = &proc_dointvec
417 .procname = "concurrent_sends",
418 .data = &concurrent_sends,
419 .maxlen = sizeof(int),
421 .proc_handler = &proc_dointvec
425 .procname = "nphys_mbox",
427 .maxlen = sizeof(int),
429 .proc_handler = &proc_dointvec
433 .procname = "sched_timeout",
434 .data = &sched_timeout,
435 .maxlen = sizeof(int),
437 .proc_handler = &proc_dointvec
441 .procname = "sched_nice",
443 .maxlen = sizeof(int),
445 .proc_handler = &proc_dointvec
449 .procname = "reverse_rdma",
450 .data = &reverse_rdma,
451 .maxlen = sizeof(int),
453 .proc_handler = &proc_dointvec
456 .procname = "dgram_timeout"
457 .data = &dgram_timeout,
458 .maxlen = sizeof(int),
460 .proc_handler = &proc_dointvec
464 .procname = "peer_timeout"
465 .data = &peer_timeout,
466 .maxlen = sizeof(int),
468 .proc_handler = &proc_dointvec
473 static cfs_sysctl_table_t kgnilnd_top_ctl_table[] = {
476 .procname = "gnilnd",
480 .child = kgnilnd_ctl_table
487 kgnilnd_tunables_init()
491 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
492 kgnilnd_tunables.kgn_sysctl =
493 cfs_register_sysctl_table(kgnilnd_top_ctl_table, 0);
495 if (kgnilnd_tunables.kgn_sysctl == NULL)
496 CWARN("Can't setup /proc tunables\n");
498 switch (*kgnilnd_tunables.kgn_checksum) {
500 CERROR("Invalid checksum module parameter: %d\n",
501 *kgnilnd_tunables.kgn_checksum);
504 case GNILND_CHECKSUM_OFF:
505 /* no checksumming */
507 case GNILND_CHECKSUM_SMSG_HEADER:
508 LCONSOLE_INFO("SMSG header only checksumming enabled\n");
510 case GNILND_CHECKSUM_SMSG:
511 LCONSOLE_INFO("SMSG checksumming enabled\n");
513 case GNILND_CHECKSUM_SMSG_BTE:
514 LCONSOLE_INFO("SMSG + BTE checksumming enabled\n");
518 if (*kgnilnd_tunables.kgn_max_immediate > GNILND_MAX_IMMEDIATE) {
519 LCONSOLE_ERROR("kgnilnd module parameter 'max_immediate' too large %d > %d\n",
520 *kgnilnd_tunables.kgn_max_immediate, GNILND_MAX_IMMEDIATE);
525 if (*kgnilnd_tunables.kgn_mbox_per_block < 1) {
526 *kgnilnd_tunables.kgn_mbox_per_block = 1;
529 if (*kgnilnd_tunables.kgn_concurrent_sends == 0) {
530 *kgnilnd_tunables.kgn_concurrent_sends = *kgnilnd_tunables.kgn_peer_credits;
531 } else if (*kgnilnd_tunables.kgn_concurrent_sends > *kgnilnd_tunables.kgn_peer_credits) {
532 LCONSOLE_ERROR("kgnilnd parameter 'concurrent_sends' too large: %d > %d (peer_credits)\n",
533 *kgnilnd_tunables.kgn_concurrent_sends, *kgnilnd_tunables.kgn_peer_credits);
541 kgnilnd_tunables_fini()
543 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
544 if (kgnilnd_tunables.kgn_sysctl != NULL)
545 cfs_unregister_sysctl_table(kgnilnd_tunables.kgn_sysctl);