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 = (2<<10);
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 max_retransmits = 1024;
98 CFS_MODULE_PARM(max_retransmits, "i", int, 0444,
99 "max retransmits for FMA");
101 static int nwildcard = 4;
102 CFS_MODULE_PARM(nwildcard, "i", int, 0444,
103 "# wildcard datagrams to post per net (interface)");
105 static int nice = -20;
106 CFS_MODULE_PARM(nice, "i", int, 0444,
107 "nice value for kgnilnd threads, default -20");
109 static int rdmaq_intervals = 4;
110 CFS_MODULE_PARM(rdmaq_intervals, "i", int, 0644,
111 "# intervals per second for rdmaq throttling, default 4, 0 to disable");
113 static int loops = 100;
114 CFS_MODULE_PARM(loops, "i", int, 0644,
115 "# of loops before scheduler is friendly, default 100");
117 static int hash_size = 503;
118 CFS_MODULE_PARM(hash_size, "i", int, 0444,
119 "prime number for peer/conn hash sizing, default 503");
121 static int peer_health = 0;
122 CFS_MODULE_PARM(peer_health, "i", int, 0444,
123 "Disable peer timeout for LNet peer health, default off, > 0 to enable");
125 static int peer_timeout = -1;
126 CFS_MODULE_PARM(peer_timeout, "i", int, 0444,
127 "Peer timeout used for peer_health, default based on gnilnd timeout, > -1 to manually set");
129 static int vmap_cksum = 0;
130 CFS_MODULE_PARM(vmap_cksum, "i", int, 0644,
131 "use vmap for all kiov checksumming, default off");
133 static int mbox_per_block = GNILND_FMABLK;
134 CFS_MODULE_PARM(mbox_per_block, "i", int, 0644,
135 "mailboxes per block");
137 static int nphys_mbox = 0;
138 CFS_MODULE_PARM(nphys_mbox, "i", int, 0444,
139 "# mbox to preallocate from physical memory, default 0");
141 static int mbox_credits = GNILND_MBOX_CREDITS;
142 CFS_MODULE_PARM(mbox_credits, "i", int, 0644,
143 "number of credits per mailbox");
145 static int sched_threads = GNILND_SCHED_THREADS;
146 CFS_MODULE_PARM(sched_threads, "i", int, 0444,
147 "number of threads for moving data");
149 static int net_hash_size = 11;
150 CFS_MODULE_PARM(net_hash_size, "i", int, 0444,
151 "prime number for net hash sizing, default 11");
153 static int hardware_timeout = GNILND_HARDWARE_TIMEOUT;
154 CFS_MODULE_PARM(hardware_timeout, "i", int, 0444,
155 "maximum time for traffic to get from one node to another");
157 static int mdd_timeout = GNILND_MDD_TIMEOUT;
158 CFS_MODULE_PARM(mdd_timeout, "i", int, 0644,
159 "maximum time (in minutes) for mdd to be held");
161 static int sched_timeout = GNILND_SCHED_TIMEOUT;
162 CFS_MODULE_PARM(sched_timeout, "i", int, 0644,
163 "scheduler aliveness in seconds max time");
165 static int sched_nice = GNILND_SCHED_NICE;
166 CFS_MODULE_PARM(sched_nice, "i", int, 0444,
167 "scheduler's nice setting, default compute 0 service -20");
169 static int reverse_rdma = GNILND_REVERSE_RDMA;
170 CFS_MODULE_PARM(reverse_rdma, "i", int, 0644,
171 "Normal 0: Reverse GET: 1 Reverse Put: 2 Reverse Both: 3");
173 static int dgram_timeout = GNILND_DGRAM_TIMEOUT;
174 CFS_MODULE_PARM(dgram_timeout, "i", int, 0644,
175 "dgram thread aliveness seconds max time");
177 static int efault_lbug = 0;
178 CFS_MODULE_PARM(efault_lbug, "i", int, 0644,
179 "If a compute receives an EFAULT in"
180 " a message should it LBUG. 0 off 1 on");
182 kgn_tunables_t kgnilnd_tunables = {
183 .kgn_min_reconnect_interval = &min_reconnect_interval,
184 .kgn_max_reconnect_interval = &max_reconnect_interval,
185 .kgn_credits = &credits,
186 .kgn_peer_credits = &peer_credits,
187 .kgn_concurrent_sends = &concurrent_sends,
188 .kgn_fma_cq_size = &fma_cq_size,
189 .kgn_timeout = &timeout,
190 .kgn_max_immediate = &max_immediate,
191 .kgn_checksum = &checksum,
192 .kgn_checksum_dump = &checksum_dump,
193 .kgn_bte_dlvr_mode = &bte_dlvr_mode,
194 .kgn_bte_relaxed_ordering = &bte_relaxed_ordering,
196 .kgn_max_retransmits = &max_retransmits,
197 .kgn_nwildcard = &nwildcard,
199 .kgn_rdmaq_intervals = &rdmaq_intervals,
201 .kgn_peer_hash_size = &hash_size,
202 .kgn_peer_health = &peer_health,
203 .kgn_peer_timeout = &peer_timeout,
204 .kgn_vmap_cksum = &vmap_cksum,
205 .kgn_mbox_per_block = &mbox_per_block,
206 .kgn_nphys_mbox = &nphys_mbox,
207 .kgn_mbox_credits = &mbox_credits,
208 .kgn_sched_threads = &sched_threads,
209 .kgn_net_hash_size = &net_hash_size,
210 .kgn_hardware_timeout = &hardware_timeout,
211 .kgn_mdd_timeout = &mdd_timeout,
212 .kgn_sched_timeout = &sched_timeout,
213 .kgn_sched_nice = &sched_nice,
214 .kgn_reverse_rdma = &reverse_rdma,
215 .kgn_dgram_timeout = &dgram_timeout,
216 .kgn_eager_credits = &eager_credits,
217 .kgn_efault_lbug = &efault_lbug
220 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
221 static struct ctl_table kgnilnd_ctl_table[] = {
224 .procname = "min_reconnect_interval",
225 .data = &min_reconnect_interval,
226 .maxlen = sizeof(int),
228 .proc_handler = &proc_dointvec
232 .procname = "max_reconnect_interval",
233 .data = &max_reconnect_interval,
234 .maxlen = sizeof(int),
236 .proc_handler = &proc_dointvec
240 .procname = "credits",
242 .maxlen = sizeof(int),
244 .proc_handler = &proc_dointvec
248 .procname = "peer_credits",
249 .data = &peer_credits,
250 .maxlen = sizeof(int),
252 .proc_handler = &proc_dointvec
256 .procname = "fma_cq_size",
257 .data = &fma_cq_size,
258 .maxlen = sizeof(int),
260 .proc_handler = &proc_dointvec
264 .procname = "timeout",
266 .maxlen = sizeof(int),
268 .proc_handler = &proc_dointvec
272 .procname = "max_immediate",
273 .data = &max_immediate,
274 .maxlen = sizeof(int),
276 .proc_handler = &proc_dointvec
280 .procname = "checksum",
282 .maxlen = sizeof(int),
284 .proc_handler = &proc_dointvec
288 .procname = "bte_dlvr_mode",
289 .data = &bte_dlvr_mode,
290 .maxlen = sizeof(int),
292 .proc_handler = &proc_dointvec
298 .maxlen = sizeof(int),
300 .proc_handler = &proc_dointvec
304 .procname = "nwildcard",
306 .maxlen = sizeof(int),
308 .proc_handler = &proc_dointvec
312 .procname = "bte_relaxed_ordering",
313 .data = &bte_relaxed_ordering,
314 .maxlen = sizeof(int),
316 .proc_handler = &proc_dointvec
320 .procname = "checksum_dump",
321 .data = &checksum_dump,
322 .maxlen = sizeof(int),
324 .proc_handler = &proc_dointvec
330 .maxlen = sizeof(int),
332 .proc_handler = &proc_dointvec
336 .procname = "rdmaq_intervals",
337 .data = &rdmaq_intervals,
338 .maxlen = sizeof(int),
340 .proc_handler = &proc_dointvec
346 .maxlen = sizeof(int),
348 .proc_handler = &proc_dointvec
352 .procname = "hash_size",
354 .maxlen = sizeof(int),
356 .proc_handler = &proc_dointvec
360 .procname = "peer_health",
361 .data = &peer_health,
362 .maxlen = sizeof(int),
364 .proc_handler = &proc_dointvec
368 .procname = "vmap_cksum",
370 .maxlen = sizeof(int),
372 .proc_handler = &proc_dointvec
376 .procname = "mbox_per_block",
377 .data = &mbox_per_block,
378 .maxlen = sizeof(int),
380 .proc_handler = &proc_dointvec
384 .procname = "mbox_credits"
385 .data = &mbox_credits,
386 .maxlen = sizeof(int),
388 .proc_handler = &proc_dointvec
392 .procname = "sched_threads"
393 .data = &sched_threads,
394 .maxlen = sizeof(int),
396 .proc_handler = &proc_dointvec
400 .procname = "net_hash_size",
401 .data = &net_hash_size,
402 .maxlen = sizeof(int),
404 .proc_handler = &proc_dointvec
408 .procname = "hardware_timeout",
409 .data = &hardware_timeout,
410 .maxlen = sizeof(int),
412 .proc_handler = &proc_dointvec
416 .procname = "mdd_timeout",
417 .data = &mdd_timeout,
418 .maxlen = sizeof(int),
420 .proc_handler = &proc_dointvec
424 .procname = "max_retransmits"
425 .data = &max_retransmits,
426 .maxlen = sizeof(int),
428 .proc_handler = &proc_dointvec
432 .procname = "concurrent_sends",
433 .data = &concurrent_sends,
434 .maxlen = sizeof(int),
436 .proc_handler = &proc_dointvec
440 .procname = "nphys_mbox",
442 .maxlen = sizeof(int),
444 .proc_handler = &proc_dointvec
448 .procname = "sched_timeout",
449 .data = &sched_timeout,
450 .maxlen = sizeof(int),
452 .proc_handler = &proc_dointvec
456 .procname = "sched_nice",
458 .maxlen = sizeof(int),
460 .proc_handler = &proc_dointvec
464 .procname = "reverse_rdma",
465 .data = &reverse_rdma,
466 .maxlen = sizeof(int),
468 .proc_handler = &proc_dointvec
471 .procname = "dgram_timeout"
472 .data = &dgram_timeout,
473 .maxlen = sizeof(int),
475 .proc_handler = &proc_dointvec
479 .procname = "peer_timeout"
480 .data = &peer_timeout,
481 .maxlen = sizeof(int),
483 .proc_handler = &proc_dointvec
487 .procname = "eager_credits",
488 .data = &eager_credits,
489 .maxlen = sizeof(int),
491 .proc_handler = &proc_dointvec
495 .procname = "efault_lbug"
496 .data = &efault_lbug,
497 .maxlen = sizeof(int),
499 .proc_handler = &proc_dointvec
504 static struct ctl_table kgnilnd_top_ctl_table[] = {
507 .procname = "gnilnd",
511 .child = kgnilnd_ctl_table
518 kgnilnd_tunables_init()
522 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
523 kgnilnd_tunables.kgn_sysctl =
524 cfs_register_sysctl_table(kgnilnd_top_ctl_table, 0);
526 if (kgnilnd_tunables.kgn_sysctl == NULL)
527 CWARN("Can't setup /proc tunables\n");
529 switch (*kgnilnd_tunables.kgn_checksum) {
531 CERROR("Invalid checksum module parameter: %d\n",
532 *kgnilnd_tunables.kgn_checksum);
535 case GNILND_CHECKSUM_OFF:
536 /* no checksumming */
538 case GNILND_CHECKSUM_SMSG_HEADER:
539 LCONSOLE_INFO("SMSG header only checksumming enabled\n");
541 case GNILND_CHECKSUM_SMSG:
542 LCONSOLE_INFO("SMSG checksumming enabled\n");
544 case GNILND_CHECKSUM_SMSG_BTE:
545 LCONSOLE_INFO("SMSG + BTE checksumming enabled\n");
549 if (*kgnilnd_tunables.kgn_max_immediate > GNILND_MAX_IMMEDIATE) {
550 LCONSOLE_ERROR("kgnilnd module parameter 'max_immediate' too large %d > %d\n",
551 *kgnilnd_tunables.kgn_max_immediate, GNILND_MAX_IMMEDIATE);
556 if (*kgnilnd_tunables.kgn_mbox_per_block < 1) {
557 *kgnilnd_tunables.kgn_mbox_per_block = 1;
560 if (*kgnilnd_tunables.kgn_concurrent_sends == 0) {
561 *kgnilnd_tunables.kgn_concurrent_sends = *kgnilnd_tunables.kgn_peer_credits;
562 } else if (*kgnilnd_tunables.kgn_concurrent_sends > *kgnilnd_tunables.kgn_peer_credits) {
563 LCONSOLE_ERROR("kgnilnd parameter 'concurrent_sends' too large: %d > %d (peer_credits)\n",
564 *kgnilnd_tunables.kgn_concurrent_sends, *kgnilnd_tunables.kgn_peer_credits);
572 kgnilnd_tunables_fini()
574 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
575 if (kgnilnd_tunables.kgn_sysctl != NULL)
576 cfs_unregister_sysctl_table(kgnilnd_tunables.kgn_sysctl);