Whamcloud - gitweb
LU-1419 lnet: Add support for Cray's Gemini interconnect
[fs/lustre-release.git] / lnet / klnds / gnilnd / gnilnd_modparams.c
1 /*
2  * Copyright (C) 2004 Cluster File Systems, Inc.
3  *
4  * Copyright (C) 2009-2012 Cray, Inc.
5  *
6  *   Derived from work by: Eric Barton <eric@bartonsoftware.com>
7  *   Author: Nic Henke <nic@cray.com>
8  *
9  *   This file is part of Lustre, http://www.lustre.org.
10  *
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.
14  *
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.
19  *
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.
23  *
24  */
25
26 #include "gnilnd.h"
27
28 static int credits = 256;
29 CFS_MODULE_PARM(credits, "i", int, 0444,
30                 "# concurrent sends");
31
32 static int peer_credits = 16;
33 CFS_MODULE_PARM(peer_credits, "i", int, 0444,
34                 "# LNet peer credits");
35
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
38  * in the mailbox */
39 static int concurrent_sends = 0;
40 CFS_MODULE_PARM(concurrent_sends, "i", int, 0444,
41                 "# concurrent HW sends to 1 peer");
42
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");
47
48 static int timeout = GNILND_BASE_TIMEOUT;
49 /* can't change @ runtime because LNet gets NI data at startup from
50  * this value */
51 CFS_MODULE_PARM(timeout, "i", int, 0444,
52                 "communications timeout (seconds)");
53
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)");
58
59 /* if this goes longer than timeout, we'll timeout the TX before
60  * the dgram */
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)");
64
65 static int max_immediate = (2<<10);
66 CFS_MODULE_PARM(max_immediate, "i", int, 0644,
67                 "immediate/RDMA breakpoint");
68
69 #ifdef CONFIG_CRAY_GEMINI
70 static int checksum = GNILND_CHECKSUM_SMSG_BTE;
71 #else
72 static int checksum = 0;
73 #endif
74 CFS_MODULE_PARM(checksum, "i", int, 0644,
75                 "0: None, 1: headers, 2: short msg, 3: all traffic");
76
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");
80
81 static int bte_hash = 1;
82 CFS_MODULE_PARM(bte_hash, "i", int, 0644,
83                 "enable hashing for BTE (RDMA) transfers");
84
85 static int bte_adapt = 1;
86 CFS_MODULE_PARM(bte_adapt, "i", int, 0644,
87                 "enable adaptive request and response for BTE (RDMA) transfers");
88
89 static int bte_relaxed_ordering = 1;
90 CFS_MODULE_PARM(bte_relaxed_ordering, "i", int, 0644,
91                 "enable relaxed ordering (PASSPW) for BTE (RDMA) transfers");
92
93 static int ptag = GNI_PTAG_LND;
94 CFS_MODULE_PARM(ptag, "i", int, 0444,
95                 "ptag for Gemini CDM");
96
97 static int max_retransmits = 1024;
98 CFS_MODULE_PARM(max_retransmits, "i", int, 0644,
99                 "max retransmits for FMA");
100
101 static int nwildcard = 4;
102 CFS_MODULE_PARM(nwildcard, "i", int, 0444,
103                 "# wildcard datagrams to post per net (interface)");
104
105 static int nice = -20;
106 CFS_MODULE_PARM(nice, "i", int, 0444,
107                 "nice value for kgnilnd threads, default -20");
108
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");
112
113 static int loops = 100;
114 CFS_MODULE_PARM(loops, "i", int, 0644,
115                 "# of loops before scheduler is friendly, default 100");
116
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");
120
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");
124
125 static int vmap_cksum = 0;
126 CFS_MODULE_PARM(vmap_cksum, "i", int, 0644,
127                 "use vmap for all kiov checksumming, default off");
128
129 static int mbox_per_block = GNILND_FMABLK;
130 CFS_MODULE_PARM(mbox_per_block, "i", int, 0644,
131                 "mailboxes per block");
132
133 static int nphys_mbox = 0;
134 CFS_MODULE_PARM(nphys_mbox, "i", int, 0444,
135                 "# mbox to preallocate from physical memory, default 0");
136
137 static int mbox_credits = GNILND_MBOX_CREDITS;
138 CFS_MODULE_PARM(mbox_credits, "i", int, 0644,
139                 "number of credits per mailbox");
140
141 static int sched_threads = GNILND_SCHED_THREADS;
142 CFS_MODULE_PARM(sched_threads, "i", int, 0444,
143                 "number of threads for moving data");
144
145 static int net_hash_size = 11;
146 CFS_MODULE_PARM(net_hash_size, "i", int, 0444,
147                 "prime number for net hash sizing, default 11");
148
149 static int hardware_timeout = GNILND_HARDWARE_TIMEOUT;
150 CFS_MODULE_PARM(hardware_timeout, "i", int, 0444,
151                 "maximum time for traffic to get from one node to another");
152
153 static int mdd_timeout = GNILND_MDD_TIMEOUT;
154 CFS_MODULE_PARM(mdd_timeout, "i", int, 0644,
155                 "maximum time (in minutes) for mdd to be held");
156
157 kgn_tunables_t kgnilnd_tunables = {
158         .kgn_min_reconnect_interval = &min_reconnect_interval,
159         .kgn_max_reconnect_interval = &max_reconnect_interval,
160         .kgn_credits                = &credits,
161         .kgn_peer_credits           = &peer_credits,
162         .kgn_concurrent_sends       = &concurrent_sends,
163         .kgn_fma_cq_size            = &fma_cq_size,
164         .kgn_timeout                = &timeout,
165         .kgn_max_immediate          = &max_immediate,
166         .kgn_checksum               = &checksum,
167         .kgn_checksum_dump          = &checksum_dump,
168         .kgn_bte_hash               = &bte_hash,
169         .kgn_bte_adapt              = &bte_adapt,
170         .kgn_bte_relaxed_ordering   = &bte_relaxed_ordering,
171         .kgn_ptag                   = &ptag,
172         .kgn_max_retransmits        = &max_retransmits,
173         .kgn_nwildcard              = &nwildcard,
174         .kgn_nice                   = &nice,
175         .kgn_rdmaq_intervals        = &rdmaq_intervals,
176         .kgn_loops                  = &loops,
177         .kgn_peer_hash_size         = &hash_size,
178         .kgn_peer_health            = &peer_health,
179         .kgn_vmap_cksum             = &vmap_cksum,
180         .kgn_mbox_per_block         = &mbox_per_block,
181         .kgn_nphys_mbox             = &nphys_mbox,
182         .kgn_mbox_credits           = &mbox_credits,
183         .kgn_sched_threads          = &sched_threads,
184         .kgn_net_hash_size          = &net_hash_size,
185         .kgn_hardware_timeout       = &hardware_timeout,
186         .kgn_mdd_timeout            = &mdd_timeout
187 };
188
189 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
190 static cfs_sysctl_table_t kgnilnd_ctl_table[] = {
191         {
192                 INIT_CTL_NAME(2)
193                 .procname = "min_reconnect_interval",
194                 .data     = &min_reconnect_interval,
195                 .maxlen   = sizeof(int),
196                 .mode     = 0644,
197                 .proc_handler = &proc_dointvec
198         },
199         {
200                 INIT_CTL_NAME(3)
201                 .procname = "max_reconnect_interval",
202                 .data     = &max_reconnect_interval,
203                 .maxlen   = sizeof(int),
204                 .mode     = 0644,
205                 .proc_handler = &proc_dointvec
206         },
207         {
208                 INIT_CTL_NAME(5)
209                 .procname = "credits",
210                 .data     = &credits,
211                 .maxlen   = sizeof(int),
212                 .mode     = 0444,
213                 .proc_handler = &proc_dointvec
214         },
215         {
216                 INIT_CTL_NAME(6)
217                 .procname = "peer_credits",
218                 .data     = &peer_credits,
219                 .maxlen   = sizeof(int),
220                 .mode     = 0444,
221                 .proc_handler = &proc_dointvec
222         },
223         {
224                 INIT_CTL_NAME(7)
225                 .procname = "fma_cq_size",
226                 .data     = &fma_cq_size,
227                 .maxlen   = sizeof(int),
228                 .mode     = 0444,
229                 .proc_handler = &proc_dointvec
230         },
231         {
232                 INIT_CTL_NAME(8)
233                 .procname = "timeout",
234                 .data     = &timeout,
235                 .maxlen   = sizeof(int),
236                 .mode     = 0444,
237                 .proc_handler = &proc_dointvec
238         },
239         {
240                 INIT_CTL_NAME(9)
241                 .procname = "max_immediate",
242                 .data     = &max_immediate,
243                 .maxlen   = sizeof(int),
244                 .mode     = 0444,
245                 .proc_handler = &proc_dointvec
246         },
247         {
248                 INIT_CTL_NAME(10)
249                 .procname = "checksum",
250                 .data     = &checksum,
251                 .maxlen   = sizeof(int),
252                 .mode     = 0644,
253                 .proc_handler = &proc_dointvec
254         },
255         {
256                 INIT_CTL_NAME(11)
257                 .procname = "bte_hash",
258                 .data     = &bte_hash,
259                 .maxlen   = sizeof(int),
260                 .mode     = 0644,
261                 .proc_handler = &proc_dointvec
262         },
263         {
264                 INIT_CTL_NAME(12)
265                 .procname = "bte_adapt",
266                 .data     = &bte_adapt,
267                 .maxlen   = sizeof(int),
268                 .mode     = 0644,
269                 .proc_handler = &proc_dointvec
270         },
271         {
272                 INIT_CTL_NAME(13)
273                 .procname = "ptag",
274                 .data     = &ptag,
275                 .maxlen   = sizeof(int),
276                 .mode     = 0444,
277                 .proc_handler = &proc_dointvec
278         },
279         {
280                 INIT_CTL_NAME(14)
281                 .procname = "nwildcard",
282                 .data     = &nwildcard,
283                 .maxlen   = sizeof(int),
284                 .mode     = 0444,
285                 .proc_handler = &proc_dointvec
286         },
287         {
288                 INIT_CTL_NAME(15)
289                 .procname = "bte_relaxed_ordering",
290                 .data     = &bte_relaxed_ordering,
291                 .maxlen   = sizeof(int),
292                 .mode     = 0644,
293                 .proc_handler = &proc_dointvec
294         },
295         {
296                 INIT_CTL_NAME(16)
297                 .procname = "checksum_dump",
298                 .data     = &checksum_dump,
299                 .maxlen   = sizeof(int),
300                 .mode     = 0644,
301                 .proc_handler = &proc_dointvec
302         },
303         {
304                 INIT_CTL_NAME(17)
305                 .procname = "nice",
306                 .data     = &nice,
307                 .maxlen   = sizeof(int),
308                 .mode     = 0444,
309                 .proc_handler = &proc_dointvec
310         },
311         {
312                 INIT_CTL_NAME(18)
313                 .procname = "rdmaq_intervals",
314                 .data     = &rdmaq_intervals,
315                 .maxlen   = sizeof(int),
316                 .mode     = 0644,
317                 .proc_handler = &proc_dointvec
318         },
319         {
320                 INIT_CTL_NAME(19)
321                 .procname = "loops",
322                 .data     = &loops,
323                 .maxlen   = sizeof(int),
324                 .mode     = 0644,
325                 .proc_handler = &proc_dointvec
326         },
327         {
328                 INIT_CTL_NAME(20)
329                 .procname = "hash_size",
330                 .data     = &hash_size,
331                 .maxlen   = sizeof(int),
332                 .mode     = 0444,
333                 .proc_handler = &proc_dointvec
334         },
335         {
336                 INIT_CTL_NAME(21)
337                 .procname = "peer_health",
338                 .data     = &peer_health,
339                 .maxlen   = sizeof(int),
340                 .mode     = 0444,
341                 .proc_handler = &proc_dointvec
342         },
343         {
344                 INIT_CTL_NAME(22)
345                 .procname = "vmap_cksum",
346                 .data     = &vmap_cksum,
347                 .maxlen   = sizeof(int),
348                 .mode     = 0644,
349                 .proc_handler = &proc_dointvec
350         },
351         {
352                 INIT_CTL_NAME(23)
353                 .procname = "mbox_per_block",
354                 .data     = &mbox_per_block,
355                 .maxlen   = sizeof(int),
356                 .mode     = 0644,
357                 .proc_handler = &proc_dointvec
358         },
359         {
360                 INIT_CTL_NAME(24)
361                 .procname = "mbox_credits"
362                 .data     = &mbox_credits,
363                 .maxlen   = sizeof(int),
364                 .mode     = 0644,
365                 .proc_handler = &proc_dointvec
366         },
367         {
368                 INIT_CTL_NAME(25)
369                 .procname = "sched_threads"
370                 .data     = &sched_threads,
371                 .maxlen   = sizeof(int),
372                 .mode     = 0444,
373                 .proc_handler = &proc_dointvec
374         },
375         {
376                 INIT_CTL_NAME(26)
377                 .procname = "net_hash_size",
378                 .data     = &net_hash_size,
379                 .maxlen   = sizeof(int),
380                 .mode     = 0444,
381                 .proc_handler = &proc_dointvec
382         },
383         {
384                 INIT_CTL_NAME(27)
385                 .procname = "hardware_timeout",
386                 .data     = &hardware_timeout,
387                 .maxlen   = sizeof(int),
388                 .mode     = 0444,
389                 .proc_handler = &proc_dointvec
390         },
391         {
392                 INIT_CTL_NAME(28)
393                 .procname = "mdd_timeout",
394                 .data     = &mdd_timeout,
395                 .maxlen   = sizeof(int),
396                 .mode     = 0644,
397                 .proc_handler = &proc_dointvec
398         },
399         {
400                 INIT_CTL_NAME(29)
401                 .procname = "max_retransmits"
402                 .data     = &max_retransmits,
403                 .maxlen   = sizeof(int),
404                 .mode     = 0444,
405                 .proc_handler = &proc_dointvec
406         },
407         {
408                 INIT_CTL_NAME(30)
409                 .procname = "concurrent_sends",
410                 .data     = &concurrent_sends,
411                 .maxlen   = sizeof(int),
412                 .mode     = 0444,
413                 .proc_handler = &proc_dointvec
414         },
415         {
416                 INIT_CTL_NAME(31)
417                 .procname = "nphys_mbox",
418                 .data     = &nphys_mbox,
419                 .maxlen   = sizeof(int),
420                 .mode     = 0444,
421                 .proc_handler = &proc_dointvec
422         },
423         {0}
424 };
425
426 static cfs_sysctl_table_t kgnilnd_top_ctl_table[] = {
427         {
428                 INIT_CTL_NAME(202)
429                 .procname = "gnilnd",
430                 .data     = NULL,
431                 .maxlen   = 0,
432                 .mode     = 0555,
433                 .child    = kgnilnd_ctl_table
434         },
435         {       INIT_CTL_NAME(0)   }
436 };
437 #endif
438
439 int
440 kgnilnd_tunables_init()
441 {
442         int rc = 0;
443
444 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
445         kgnilnd_tunables.kgn_sysctl =
446                 cfs_register_sysctl_table(kgnilnd_top_ctl_table, 0);
447
448         if (kgnilnd_tunables.kgn_sysctl == NULL)
449                 CWARN("Can't setup /proc tunables\n");
450 #endif
451         switch (*kgnilnd_tunables.kgn_checksum) {
452         default:
453                 CERROR("Invalid checksum module parameter: %d\n",
454                        *kgnilnd_tunables.kgn_checksum);
455                 rc = -EINVAL;
456                 GOTO(out, rc);
457         case GNILND_CHECKSUM_OFF:
458                 /* no checksumming */
459                 break;
460         case GNILND_CHECKSUM_SMSG_HEADER:
461                 LCONSOLE_INFO("SMSG header only checksumming enabled\n");
462                 break;
463         case GNILND_CHECKSUM_SMSG:
464                 LCONSOLE_INFO("SMSG checksumming enabled\n");
465                 break;
466         case GNILND_CHECKSUM_SMSG_BTE:
467                 LCONSOLE_INFO("SMSG + BTE checksumming enabled\n");
468                 break;
469         }
470
471         if (*kgnilnd_tunables.kgn_max_immediate > GNILND_MAX_IMMEDIATE) {
472                 LCONSOLE_ERROR("kgnilnd module parameter 'max_immediate' too large %d > %d\n",
473                 *kgnilnd_tunables.kgn_max_immediate, GNILND_MAX_IMMEDIATE);
474                 rc = -EINVAL;
475                 GOTO(out, rc);
476         }
477
478         if (*kgnilnd_tunables.kgn_mbox_per_block < 1) {
479                 *kgnilnd_tunables.kgn_mbox_per_block = 1;
480         }
481
482         if (*kgnilnd_tunables.kgn_concurrent_sends == 0) {
483                 *kgnilnd_tunables.kgn_concurrent_sends = *kgnilnd_tunables.kgn_peer_credits;
484         } else if (*kgnilnd_tunables.kgn_concurrent_sends > *kgnilnd_tunables.kgn_peer_credits) {
485                 LCONSOLE_ERROR("kgnilnd parameter 'concurrent_sends' too large: %d > %d (peer_credits)\n",
486                                *kgnilnd_tunables.kgn_concurrent_sends, *kgnilnd_tunables.kgn_peer_credits);
487                 rc = -EINVAL;
488         }
489 out:
490         return rc;
491 }
492
493 void
494 kgnilnd_tunables_fini()
495 {
496 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
497         if (kgnilnd_tunables.kgn_sysctl != NULL)
498                 cfs_unregister_sysctl_table(kgnilnd_tunables.kgn_sysctl);
499 #endif
500 }