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, 0644,
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 = 2048;
70 CFS_MODULE_PARM(max_immediate, "i", int, 0444,
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 (disabled)");
197 static int thread_safe = GNILND_TS_ENABLE;
198 CFS_MODULE_PARM(thread_safe, "i", int, 0444,
199 "Use kgni thread safe API if available");
201 static int reg_fail_timeout = GNILND_REGFAILTO_DISABLE;
202 CFS_MODULE_PARM(reg_fail_timeout, "i", int, 0644,
203 "fmablk registration timeout LBUG");
205 kgn_tunables_t kgnilnd_tunables = {
206 .kgn_min_reconnect_interval = &min_reconnect_interval,
207 .kgn_max_reconnect_interval = &max_reconnect_interval,
208 .kgn_credits = &credits,
209 .kgn_peer_credits = &peer_credits,
210 .kgn_concurrent_sends = &concurrent_sends,
211 .kgn_fma_cq_size = &fma_cq_size,
212 .kgn_timeout = &timeout,
213 .kgn_max_immediate = &max_immediate,
214 .kgn_checksum = &checksum,
215 .kgn_checksum_dump = &checksum_dump,
216 .kgn_bte_dlvr_mode = &bte_dlvr_mode,
217 .kgn_bte_relaxed_ordering = &bte_relaxed_ordering,
220 .kgn_max_retransmits = &max_retransmits,
221 .kgn_nwildcard = &nwildcard,
223 .kgn_rdmaq_intervals = &rdmaq_intervals,
225 .kgn_peer_hash_size = &hash_size,
226 .kgn_peer_health = &peer_health,
227 .kgn_peer_timeout = &peer_timeout,
228 .kgn_vmap_cksum = &vmap_cksum,
229 .kgn_mbox_per_block = &mbox_per_block,
230 .kgn_nphys_mbox = &nphys_mbox,
231 .kgn_mbox_credits = &mbox_credits,
232 .kgn_sched_threads = &sched_threads,
233 .kgn_net_hash_size = &net_hash_size,
234 .kgn_hardware_timeout = &hardware_timeout,
235 .kgn_mdd_timeout = &mdd_timeout,
236 .kgn_sched_timeout = &sched_timeout,
237 .kgn_sched_nice = &sched_nice,
238 .kgn_reverse_rdma = &reverse_rdma,
239 .kgn_dgram_timeout = &dgram_timeout,
240 .kgn_eager_credits = &eager_credits,
241 .kgn_fast_reconn = &fast_reconn,
242 .kgn_efault_lbug = &efault_lbug,
243 .kgn_thread_affinity = &thread_affinity,
244 .kgn_thread_safe = &thread_safe,
245 .kgn_reg_fail_timeout = ®_fail_timeout,
246 .kgn_max_purgatory = &max_conn_purg
249 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
250 static struct ctl_table kgnilnd_ctl_table[] = {
253 .procname = "min_reconnect_interval",
254 .data = &min_reconnect_interval,
255 .maxlen = sizeof(int),
257 .proc_handler = &proc_dointvec
261 .procname = "max_reconnect_interval",
262 .data = &max_reconnect_interval,
263 .maxlen = sizeof(int),
265 .proc_handler = &proc_dointvec
269 .procname = "credits",
271 .maxlen = sizeof(int),
273 .proc_handler = &proc_dointvec
277 .procname = "peer_credits",
278 .data = &peer_credits,
279 .maxlen = sizeof(int),
281 .proc_handler = &proc_dointvec
285 .procname = "fma_cq_size",
286 .data = &fma_cq_size,
287 .maxlen = sizeof(int),
289 .proc_handler = &proc_dointvec
293 .procname = "timeout",
295 .maxlen = sizeof(int),
297 .proc_handler = &proc_dointvec
301 .procname = "max_immediate",
302 .data = &max_immediate,
303 .maxlen = sizeof(int),
305 .proc_handler = &proc_dointvec
309 .procname = "checksum",
311 .maxlen = sizeof(int),
313 .proc_handler = &proc_dointvec
317 .procname = "bte_dlvr_mode",
318 .data = &bte_dlvr_mode,
319 .maxlen = sizeof(int),
321 .proc_handler = &proc_dointvec
327 .maxlen = sizeof(int),
329 .proc_handler = &proc_dointvec
335 .maxlen = sizeof(int),
337 .proc_handler = &proc_dointvec
341 .procname = "nwildcard",
343 .maxlen = sizeof(int),
345 .proc_handler = &proc_dointvec
349 .procname = "bte_relaxed_ordering",
350 .data = &bte_relaxed_ordering,
351 .maxlen = sizeof(int),
353 .proc_handler = &proc_dointvec
357 .procname = "checksum_dump",
358 .data = &checksum_dump,
359 .maxlen = sizeof(int),
361 .proc_handler = &proc_dointvec
367 .maxlen = sizeof(int),
369 .proc_handler = &proc_dointvec
373 .procname = "rdmaq_intervals",
374 .data = &rdmaq_intervals,
375 .maxlen = sizeof(int),
377 .proc_handler = &proc_dointvec
383 .maxlen = sizeof(int),
385 .proc_handler = &proc_dointvec
389 .procname = "hash_size",
391 .maxlen = sizeof(int),
393 .proc_handler = &proc_dointvec
397 .procname = "peer_health",
398 .data = &peer_health,
399 .maxlen = sizeof(int),
401 .proc_handler = &proc_dointvec
405 .procname = "vmap_cksum",
407 .maxlen = sizeof(int),
409 .proc_handler = &proc_dointvec
413 .procname = "mbox_per_block",
414 .data = &mbox_per_block,
415 .maxlen = sizeof(int),
417 .proc_handler = &proc_dointvec
421 .procname = "mbox_credits"
422 .data = &mbox_credits,
423 .maxlen = sizeof(int),
425 .proc_handler = &proc_dointvec
429 .procname = "sched_threads"
430 .data = &sched_threads,
431 .maxlen = sizeof(int),
433 .proc_handler = &proc_dointvec
437 .procname = "net_hash_size",
438 .data = &net_hash_size,
439 .maxlen = sizeof(int),
441 .proc_handler = &proc_dointvec
445 .procname = "hardware_timeout",
446 .data = &hardware_timeout,
447 .maxlen = sizeof(int),
449 .proc_handler = &proc_dointvec
453 .procname = "mdd_timeout",
454 .data = &mdd_timeout,
455 .maxlen = sizeof(int),
457 .proc_handler = &proc_dointvec
461 .procname = "max_retransmits"
462 .data = &max_retransmits,
463 .maxlen = sizeof(int),
465 .proc_handler = &proc_dointvec
469 .procname = "concurrent_sends",
470 .data = &concurrent_sends,
471 .maxlen = sizeof(int),
473 .proc_handler = &proc_dointvec
477 .procname = "nphys_mbox",
479 .maxlen = sizeof(int),
481 .proc_handler = &proc_dointvec
485 .procname = "sched_timeout",
486 .data = &sched_timeout,
487 .maxlen = sizeof(int),
489 .proc_handler = &proc_dointvec
493 .procname = "sched_nice",
495 .maxlen = sizeof(int),
497 .proc_handler = &proc_dointvec
501 .procname = "reverse_rdma",
502 .data = &reverse_rdma,
503 .maxlen = sizeof(int),
505 .proc_handler = &proc_dointvec
508 .procname = "dgram_timeout"
509 .data = &dgram_timeout,
510 .maxlen = sizeof(int),
512 .proc_handler = &proc_dointvec
516 .procname = "peer_timeout"
517 .data = &peer_timeout,
518 .maxlen = sizeof(int),
520 .proc_handler = &proc_dointvec
524 .procname = "eager_credits",
525 .data = &eager_credits,
526 .maxlen = sizeof(int),
528 .proc_handler = &proc_dointvec
532 .procname = "efault_lbug"
533 .data = &efault_lbug,
534 .maxlen = sizeof(int),
536 .proc_handler = &proc_dointvec
540 .procname = "thread_affinity"
541 .data = &thread_affinity,
542 .maxlen = sizeof(int),
544 .proc_handler = &proc_dointvec
548 .procname = "thread_safe"
549 .data = &thread_safe,
550 .maxlen = sizeof(int),
552 .proc_handler = &proc_dointvec
556 .procname = "reg_fail_timeout"
557 .data = ®_fail_timeout,
558 .maxlen = sizeof(int),
560 .proc_handler = &proc_dointvec
564 .procname = "max_conn_purg"
565 .data = &max_conn_purg,
566 .maxlen = sizeof(int),
568 .proc_handler = &proc_dointvec
573 static struct ctl_table kgnilnd_top_ctl_table[] = {
576 .procname = "gnilnd",
580 .child = kgnilnd_ctl_table
587 kgnilnd_tunables_init()
591 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
592 kgnilnd_tunables.kgn_sysctl =
593 register_sysctl_table(kgnilnd_top_ctl_table, 0);
595 if (kgnilnd_tunables.kgn_sysctl == NULL)
596 CWARN("Can't setup /proc tunables\n");
598 switch (*kgnilnd_tunables.kgn_checksum) {
600 CERROR("Invalid checksum module parameter: %d\n",
601 *kgnilnd_tunables.kgn_checksum);
604 case GNILND_CHECKSUM_OFF:
605 /* no checksumming */
607 case GNILND_CHECKSUM_SMSG_HEADER:
608 LCONSOLE_INFO("SMSG header only checksumming enabled\n");
610 case GNILND_CHECKSUM_SMSG:
611 LCONSOLE_INFO("SMSG checksumming enabled\n");
613 case GNILND_CHECKSUM_SMSG_BTE:
614 LCONSOLE_INFO("SMSG + BTE checksumming enabled\n");
618 if (*kgnilnd_tunables.kgn_max_immediate > GNILND_MAX_IMMEDIATE) {
619 LCONSOLE_ERROR("kgnilnd module parameter 'max_immediate' too large %d > %d\n",
620 *kgnilnd_tunables.kgn_max_immediate, GNILND_MAX_IMMEDIATE);
625 if (*kgnilnd_tunables.kgn_mbox_per_block < 1) {
626 *kgnilnd_tunables.kgn_mbox_per_block = 1;
629 if (*kgnilnd_tunables.kgn_concurrent_sends == 0) {
630 *kgnilnd_tunables.kgn_concurrent_sends = *kgnilnd_tunables.kgn_peer_credits;
631 } else if (*kgnilnd_tunables.kgn_concurrent_sends > *kgnilnd_tunables.kgn_peer_credits) {
632 LCONSOLE_ERROR("kgnilnd parameter 'concurrent_sends' too large: %d > %d (peer_credits)\n",
633 *kgnilnd_tunables.kgn_concurrent_sends, *kgnilnd_tunables.kgn_peer_credits);
641 kgnilnd_tunables_fini()
643 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
644 if (kgnilnd_tunables.kgn_sysctl != NULL)
645 unregister_sysctl_table(kgnilnd_tunables.kgn_sysctl);