Whamcloud - gitweb
LU-2150 ost: ost_brw_read() to ptlrpc_free_bulk_nopin()
[fs/lustre-release.git] / lustre / mgs / mgs_nids.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, Whamcloud, Inc.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/mgs/mgs_nids.c
37  *
38  * NID table management for lustre.
39  *
40  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
41  */
42
43 #define DEBUG_SUBSYSTEM S_MGS
44 #define D_MGS D_CONFIG
45
46 #ifdef __KERNEL__
47 #include <linux/pagemap.h>
48 #endif
49
50 #include <obd.h>
51 #include <obd_class.h>
52 #include <lustre_disk.h>
53
54 #include "mgs_internal.h"
55
56 static unsigned int ir_timeout;
57
58 static int nidtbl_is_sane(struct mgs_nidtbl *tbl)
59 {
60         struct mgs_nidtbl_target *tgt;
61         int version = 0;
62
63         LASSERT(cfs_mutex_is_locked(&tbl->mn_lock));
64         cfs_list_for_each_entry(tgt, &tbl->mn_targets, mnt_list) {
65                 if (!tgt->mnt_version)
66                         continue;
67
68                 if (version >= tgt->mnt_version)
69                         return 0;
70
71                 version = tgt->mnt_version;
72         }
73         return 1;
74 }
75
76 /**
77  * Fetch nidtbl entries whose version are not less than @version
78  * nidtbl entries will be packed in @pages by @unit_size units - entries
79  * shouldn't cross unit boundaries.
80  */
81 static int mgs_nidtbl_read(struct obd_export *exp, struct mgs_nidtbl *tbl,
82                            struct mgs_config_res *res, cfs_page_t **pages,
83                            int nrpages, int units_total, int unit_size)
84 {
85         struct mgs_nidtbl_target *tgt;
86         struct mgs_nidtbl_entry  *entry;
87         struct mgs_nidtbl_entry  *last_in_unit = NULL;
88         struct mgs_target_info   *mti;
89         __u64 version = res->mcr_offset;
90         bool nobuf = false;
91         void *buf = NULL;
92         int bytes_in_unit = 0;
93         int units_in_page = 0;
94         int index = 0;
95         int rc = 0;
96         ENTRY;
97
98         /* make sure unit_size is power 2 */
99         LASSERT((unit_size & (unit_size - 1)) == 0);
100         LASSERT(nrpages << CFS_PAGE_SHIFT >= units_total * unit_size);
101
102         cfs_mutex_lock(&tbl->mn_lock);
103         LASSERT(nidtbl_is_sane(tbl));
104
105         /* no more entries ? */
106         if (version > tbl->mn_version) {
107                 version = tbl->mn_version;
108                 goto out;
109         }
110
111         /* iterate over all targets to compose a bitmap by the type of llog.
112          * If the llog is for MDTs, llog entries for OSTs will be returned;
113          * otherwise, it's for clients, then llog entries for both OSTs and
114          * MDTs will be returned.
115          */
116         cfs_list_for_each_entry(tgt, &tbl->mn_targets, mnt_list) {
117                 int entry_len = sizeof(*entry);
118
119                 if (tgt->mnt_version < version)
120                         continue;
121
122                 /* write target recover information */
123                 mti  = &tgt->mnt_mti;
124                 LASSERT(mti->mti_nid_count < MTI_NIDS_MAX);
125                 entry_len += mti->mti_nid_count * sizeof(lnet_nid_t);
126
127                 if (entry_len > unit_size) {
128                         CWARN("nidtbl: too large entry: entry length %d,"
129                               "unit size: %d\n", entry_len, unit_size);
130                         GOTO(out, rc = -EOVERFLOW);
131                 }
132
133                 if (bytes_in_unit < entry_len) {
134                         if (units_total == 0) {
135                                 nobuf = true;
136                                 break;
137                         }
138
139                         /* check if we need to consume remaining bytes. */
140                         if (last_in_unit != NULL && bytes_in_unit) {
141 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
142                                 /* May need to swab back to update the length.*/
143                                 if (exp->exp_need_mne_swab)
144                                         lustre_swab_mgs_nidtbl_entry(last_in_unit);
145 #endif
146                                 last_in_unit->mne_length += bytes_in_unit;
147 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
148                                 if (exp->exp_need_mne_swab)
149                                         lustre_swab_mgs_nidtbl_entry(last_in_unit);
150 #endif
151                                 rc  += bytes_in_unit;
152                                 buf += bytes_in_unit;
153                                 last_in_unit = NULL;
154                         }
155                         LASSERT((rc & (unit_size - 1)) == 0);
156
157                         if (units_in_page == 0) {
158                                 /* allocate a new page */
159                                 pages[index] = cfs_alloc_page(CFS_ALLOC_STD);
160                                 if (pages[index] == NULL) {
161                                         rc = -ENOMEM;
162                                         break;
163                                 }
164
165                                 /* destroy previous map */
166                                 if (index > 0)
167                                         cfs_kunmap(pages[index - 1]);
168
169                                 /* reassign buffer */
170                                 buf = cfs_kmap(pages[index]);
171                                 ++index;
172
173                                 units_in_page = CFS_PAGE_SIZE / unit_size;
174                                 LASSERT(units_in_page > 0);
175                         }
176
177                         /* allocate an unit */
178                         LASSERT(((long)buf & (unit_size - 1)) == 0);
179                         bytes_in_unit = unit_size;
180                         --units_in_page;
181                         --units_total;
182                 }
183
184                 /* fill in entry. */
185                 entry = (struct mgs_nidtbl_entry *)buf;
186                 entry->mne_version   = tgt->mnt_version;
187                 entry->mne_instance  = mti->mti_instance;
188                 entry->mne_index     = mti->mti_stripe_index;
189                 entry->mne_length    = entry_len;
190                 entry->mne_type      = tgt->mnt_type;
191                 entry->mne_nid_type  = 0;
192                 entry->mne_nid_size  = sizeof(lnet_nid_t);
193                 entry->mne_nid_count = mti->mti_nid_count;
194                 memcpy(entry->u.nids, mti->mti_nids,
195                        mti->mti_nid_count * sizeof(lnet_nid_t));
196
197 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
198                 /* For LU-1644, swab entry for 2.2 clients. */
199                 if (exp->exp_need_mne_swab)
200                         lustre_swab_mgs_nidtbl_entry(entry);
201 #endif
202
203                 version = tgt->mnt_version;
204                 rc     += entry_len;
205                 buf    += entry_len;
206
207                 bytes_in_unit -= entry_len;
208                 last_in_unit   = entry;
209
210                 CDEBUG(D_MGS, "fsname %s, entry size %d, pages %d/%d/%d/%d.\n",
211                        tbl->mn_fsdb->fsdb_name, entry_len,
212                        bytes_in_unit, index, nrpages, units_total);
213         }
214         if (index > 0)
215                 cfs_kunmap(pages[index - 1]);
216 out:
217         LASSERT(version <= tbl->mn_version);
218         res->mcr_size = tbl->mn_version;
219         res->mcr_offset = nobuf ? version : tbl->mn_version;
220         cfs_mutex_unlock(&tbl->mn_lock);
221         LASSERT(ergo(version == 1, rc == 0)); /* get the log first time */
222
223         CDEBUG(D_MGS, "Read IR logs %s return with %d, version %llu\n",
224                tbl->mn_fsdb->fsdb_name, rc, version);
225         RETURN(rc);
226 }
227
228 static int nidtbl_update_version(const struct lu_env *env,
229                                  struct mgs_device *mgs,
230                                  struct mgs_nidtbl *tbl)
231 {
232         struct dt_object *fsdb;
233         struct thandle   *th;
234         u64               version;
235         struct lu_buf     buf = {
236                                 .lb_buf = &version,
237                                 .lb_len = sizeof(version)
238                           };
239         loff_t            off = 0;
240         int               rc;
241         ENTRY;
242
243         LASSERT(cfs_mutex_is_locked(&tbl->mn_lock));
244
245         fsdb = local_file_find_or_create(env, mgs->mgs_los, mgs->mgs_nidtbl_dir,
246                                          tbl->mn_fsdb->fsdb_name,
247                                          S_IFREG | S_IRUGO | S_IWUSR);
248         if (IS_ERR(fsdb))
249                 RETURN(PTR_ERR(fsdb));
250
251         th = dt_trans_create(env, mgs->mgs_bottom);
252         if (IS_ERR(th))
253                 GOTO(out_put, rc = PTR_ERR(th));
254
255         th->th_sync = 1; /* update table synchronously */
256         rc = dt_declare_record_write(env, fsdb, buf.lb_len, off, th);
257         if (rc)
258                 GOTO(out, rc);
259
260         rc = dt_trans_start_local(env, mgs->mgs_bottom, th);
261         if (rc)
262                 GOTO(out, rc);
263
264         version = cpu_to_le64(tbl->mn_version);
265         rc = dt_record_write(env, fsdb, &buf, &off, th);
266
267 out:
268         dt_trans_stop(env, mgs->mgs_bottom, th);
269 out_put:
270         lu_object_put(env, &fsdb->do_lu);
271         RETURN(rc);
272 }
273
274 #define MGS_NIDTBL_VERSION_INIT 2
275
276 static int nidtbl_read_version(const struct lu_env *env,
277                                struct mgs_device *mgs, struct mgs_nidtbl *tbl,
278                                u64 *version)
279 {
280         struct dt_object *fsdb;
281         struct lu_fid     fid;
282         u64               tmpver;
283         struct lu_buf     buf = {
284                                 .lb_buf = &tmpver,
285                                 .lb_len = sizeof(tmpver)
286                           };
287         loff_t               off = 0;
288         int                  rc;
289         ENTRY;
290
291         LASSERT(cfs_mutex_is_locked(&tbl->mn_lock));
292
293         LASSERT(mgs->mgs_nidtbl_dir);
294         rc = dt_lookup_dir(env, mgs->mgs_nidtbl_dir, tbl->mn_fsdb->fsdb_name,
295                            &fid);
296         if (rc == -ENOENT) {
297                 *version = MGS_NIDTBL_VERSION_INIT;
298                 RETURN(0);
299         } else if (rc < 0) {
300                 RETURN(rc);
301         }
302
303         fsdb = dt_locate_at(env, mgs->mgs_bottom, &fid,
304                             &mgs->mgs_dt_dev.dd_lu_dev);
305         if (IS_ERR(fsdb))
306                 RETURN(PTR_ERR(fsdb));
307
308         rc = dt_read(env, fsdb, &buf, &off);
309         if (rc == buf.lb_len) {
310                 *version = le64_to_cpu(tmpver);
311                 rc = 0;
312         } else if (rc == 0) {
313                 *version = MGS_NIDTBL_VERSION_INIT;
314         } else {
315                 CERROR("%s: read version file %s error %d\n",
316                        mgs->mgs_obd->obd_name, tbl->mn_fsdb->fsdb_name, rc);
317         }
318         lu_object_put(env, &fsdb->do_lu);
319         RETURN(rc);
320 }
321
322 static int mgs_nidtbl_write(const struct lu_env *env, struct fs_db *fsdb,
323                             struct mgs_target_info *mti)
324 {
325         struct mgs_nidtbl        *tbl;
326         struct mgs_nidtbl_target *tgt;
327         bool found = false;
328         int type   = mti->mti_flags & LDD_F_SV_TYPE_MASK;
329         int rc     = 0;
330         ENTRY;
331
332         type &= ~LDD_F_SV_TYPE_MGS;
333         LASSERT(type != 0);
334
335         tbl = &fsdb->fsdb_nidtbl;
336         cfs_mutex_lock(&tbl->mn_lock);
337         cfs_list_for_each_entry(tgt, &tbl->mn_targets, mnt_list) {
338                 struct mgs_target_info *info = &tgt->mnt_mti;
339                 if (type == tgt->mnt_type &&
340                     mti->mti_stripe_index == info->mti_stripe_index) {
341                         found = true;
342                         break;
343                 }
344         }
345         if (!found) {
346                 OBD_ALLOC_PTR(tgt);
347                 if (tgt == NULL)
348                         GOTO(out, rc = -ENOMEM);
349
350                 CFS_INIT_LIST_HEAD(&tgt->mnt_list);
351                 tgt->mnt_fs      = tbl;
352                 tgt->mnt_version = 0;       /* 0 means invalid */
353                 tgt->mnt_type    = type;
354
355                 ++tbl->mn_nr_targets;
356         }
357
358         tgt->mnt_version = ++tbl->mn_version;
359         tgt->mnt_mti     = *mti;
360
361         cfs_list_move_tail(&tgt->mnt_list, &tbl->mn_targets);
362
363         rc = nidtbl_update_version(env, fsdb->fsdb_mgs, tbl);
364         EXIT;
365
366 out:
367         cfs_mutex_unlock(&tbl->mn_lock);
368         if (rc)
369                 CERROR("Write NID table version for file system %s error %d\n",
370                        fsdb->fsdb_name, rc);
371         return rc;
372 }
373
374 static void mgs_nidtbl_fini_fs(struct fs_db *fsdb)
375 {
376         struct mgs_nidtbl *tbl = &fsdb->fsdb_nidtbl;
377         CFS_LIST_HEAD(head);
378
379         cfs_mutex_lock(&tbl->mn_lock);
380         tbl->mn_nr_targets = 0;
381         cfs_list_splice_init(&tbl->mn_targets, &head);
382         cfs_mutex_unlock(&tbl->mn_lock);
383
384         while (!cfs_list_empty(&head)) {
385                 struct mgs_nidtbl_target *tgt;
386                 tgt = list_entry(head.next, struct mgs_nidtbl_target, mnt_list);
387                 cfs_list_del(&tgt->mnt_list);
388                 OBD_FREE_PTR(tgt);
389         }
390 }
391
392 static int mgs_nidtbl_init_fs(const struct lu_env *env, struct fs_db *fsdb)
393 {
394         struct mgs_nidtbl *tbl = &fsdb->fsdb_nidtbl;
395         int rc;
396
397         CFS_INIT_LIST_HEAD(&tbl->mn_targets);
398         cfs_mutex_init(&tbl->mn_lock);
399         tbl->mn_nr_targets = 0;
400         tbl->mn_fsdb = fsdb;
401         cfs_mutex_lock(&tbl->mn_lock);
402         rc = nidtbl_read_version(env, fsdb->fsdb_mgs, tbl, &tbl->mn_version);
403         cfs_mutex_unlock(&tbl->mn_lock);
404         if (rc < 0)
405                 CERROR("%s: IR: failed to read current version, rc = %d\n",
406                        fsdb->fsdb_mgs->mgs_obd->obd_name, rc);
407         else
408                 CDEBUG(D_MGS, "IR: current version is %llu\n",
409                        tbl->mn_version);
410
411         return rc;
412 }
413
414 /* --------- Imperative Recovery relies on nidtbl stuff ------- */
415 void mgs_ir_notify_complete(struct fs_db *fsdb)
416 {
417         struct timeval tv;
418         cfs_duration_t delta;
419
420         cfs_atomic_set(&fsdb->fsdb_notify_phase, 0);
421
422         /* do statistic */
423         fsdb->fsdb_notify_count++;
424         delta = cfs_time_sub(cfs_time_current(), fsdb->fsdb_notify_start);
425         fsdb->fsdb_notify_total += delta;
426         if (delta > fsdb->fsdb_notify_max)
427                 fsdb->fsdb_notify_max = delta;
428
429         cfs_duration_usec(delta, &tv);
430         CDEBUG(D_MGS, "Revoke recover lock of %s completed after %ld.%06lds\n",
431                fsdb->fsdb_name, tv.tv_sec, tv.tv_usec);
432 }
433
434 static int mgs_ir_notify(void *arg)
435 {
436         struct fs_db      *fsdb   = arg;
437         struct ldlm_res_id resid;
438
439         char name[sizeof(fsdb->fsdb_name) + 20];
440
441         LASSERTF(sizeof(name) < 32, "name is too large to be in stack.\n");
442         sprintf(name, "mgs_%s_notify", fsdb->fsdb_name);
443         cfs_daemonize(name);
444
445         cfs_complete(&fsdb->fsdb_notify_comp);
446
447         set_user_nice(current, -2);
448
449         mgc_fsname2resid(fsdb->fsdb_name, &resid, CONFIG_T_RECOVER);
450         while (1) {
451                 struct l_wait_info   lwi = { 0 };
452
453                 l_wait_event(fsdb->fsdb_notify_waitq,
454                              fsdb->fsdb_notify_stop ||
455                              cfs_atomic_read(&fsdb->fsdb_notify_phase),
456                              &lwi);
457                 if (fsdb->fsdb_notify_stop)
458                         break;
459
460                 CDEBUG(D_MGS, "%s woken up, phase is %d\n",
461                        name, cfs_atomic_read(&fsdb->fsdb_notify_phase));
462
463                 fsdb->fsdb_notify_start = cfs_time_current();
464                 mgs_revoke_lock(fsdb->fsdb_mgs, fsdb, CONFIG_T_RECOVER);
465         }
466
467         cfs_complete(&fsdb->fsdb_notify_comp);
468         return 0;
469 }
470
471 int mgs_ir_init_fs(const struct lu_env *env, struct mgs_device *mgs,
472                    struct fs_db *fsdb)
473 {
474         int rc;
475
476         if (!ir_timeout)
477                 ir_timeout = OBD_IR_MGS_TIMEOUT;
478
479         fsdb->fsdb_ir_state = IR_FULL;
480         if (cfs_time_before(cfs_time_current_sec(),
481                             mgs->mgs_start_time + ir_timeout))
482                 fsdb->fsdb_ir_state = IR_STARTUP;
483         fsdb->fsdb_nonir_clients = 0;
484         CFS_INIT_LIST_HEAD(&fsdb->fsdb_clients);
485
486         /* start notify thread */
487         fsdb->fsdb_mgs = mgs;
488         cfs_atomic_set(&fsdb->fsdb_notify_phase, 0);
489         cfs_waitq_init(&fsdb->fsdb_notify_waitq);
490         cfs_init_completion(&fsdb->fsdb_notify_comp);
491         rc = cfs_create_thread(mgs_ir_notify, fsdb, CFS_DAEMON_FLAGS);
492         if (rc > 0)
493                 cfs_wait_for_completion(&fsdb->fsdb_notify_comp);
494         else
495                 CERROR("Start notify thread error %d\n", rc);
496
497         mgs_nidtbl_init_fs(env, fsdb);
498         return 0;
499 }
500
501 void mgs_ir_fini_fs(struct mgs_device *mgs, struct fs_db *fsdb)
502 {
503         if (cfs_test_bit(FSDB_MGS_SELF, &fsdb->fsdb_flags))
504                 return;
505
506         mgs_fsc_cleanup_by_fsdb(fsdb);
507
508         mgs_nidtbl_fini_fs(fsdb);
509
510         LASSERT(cfs_list_empty(&fsdb->fsdb_clients));
511
512         fsdb->fsdb_notify_stop = 1;
513         cfs_waitq_signal(&fsdb->fsdb_notify_waitq);
514         cfs_wait_for_completion(&fsdb->fsdb_notify_comp);
515 }
516
517 /* caller must have held fsdb_mutex */
518 static inline void ir_state_graduate(struct fs_db *fsdb)
519 {
520         if (fsdb->fsdb_ir_state == IR_STARTUP) {
521                 if (cfs_time_before(fsdb->fsdb_mgs->mgs_start_time + ir_timeout,
522                                     cfs_time_current_sec())) {
523                         fsdb->fsdb_ir_state = IR_FULL;
524                         if (fsdb->fsdb_nonir_clients)
525                                 fsdb->fsdb_ir_state = IR_PARTIAL;
526                 }
527         }
528 }
529
530 int mgs_ir_update(const struct lu_env *env, struct mgs_device *mgs,
531                   struct mgs_target_info *mti)
532 {
533         struct fs_db *fsdb;
534         bool notify = true;
535         int rc;
536
537         if (mti->mti_instance == 0)
538                 return -EINVAL;
539
540         rc = mgs_find_or_make_fsdb(env, mgs, mti->mti_fsname, &fsdb);
541         if (rc)
542                 return rc;
543
544         rc = mgs_nidtbl_write(env, fsdb, mti);
545         if (rc)
546                 return rc;
547
548         /* check ir state */
549         cfs_mutex_lock(&fsdb->fsdb_mutex);
550         ir_state_graduate(fsdb);
551         switch (fsdb->fsdb_ir_state) {
552         case IR_FULL:
553                 mti->mti_flags |= LDD_F_IR_CAPABLE;
554                 break;
555         case IR_DISABLED:
556                 notify = false;
557         case IR_STARTUP:
558         case IR_PARTIAL:
559                 break;
560         default:
561                 LBUG();
562         }
563         cfs_mutex_unlock(&fsdb->fsdb_mutex);
564
565         LASSERT(ergo(mti->mti_flags & LDD_F_IR_CAPABLE, notify));
566         if (notify) {
567                 CDEBUG(D_MGS, "Try to revoke recover lock of %s\n",
568                        fsdb->fsdb_name);
569                 cfs_atomic_inc(&fsdb->fsdb_notify_phase);
570                 cfs_waitq_signal(&fsdb->fsdb_notify_waitq);
571         }
572         return 0;
573 }
574
575 /* NID table can be cached by two entities: Clients and MDTs */
576 enum {
577         IR_CLIENT  = 1,
578         IR_MDT     = 2
579 };
580
581 static int delogname(char *logname, char *fsname, int *typ)
582 {
583         char *ptr;
584         int   type;
585         int   len;
586
587         ptr = strrchr(logname, '-');
588         if (ptr == NULL)
589                 return -EINVAL;
590
591         /* decouple file system name. The llog name may be:
592          * - "prefix-fsname", prefix is "cliir" or "mdtir"
593          */
594         if (strncmp(ptr, "-mdtir", 6) == 0)
595                 type = IR_MDT;
596         else if (strncmp(ptr, "-cliir", 6) == 0)
597                 type = IR_CLIENT;
598         else
599                 return -EINVAL;
600
601         len = ptr - logname;
602         if (len == 0)
603                 return -EINVAL;
604
605         memcpy(fsname, logname, len);
606         fsname[len] = 0;
607         if (typ)
608                 *typ = type;
609         return 0;
610 }
611
612 int mgs_get_ir_logs(struct ptlrpc_request *req)
613 {
614         struct lu_env     *env = req->rq_svc_thread->t_env;
615         struct mgs_device *mgs = exp2mgs_dev(req->rq_export);
616         struct fs_db      *fsdb;
617         struct mgs_config_body  *body;
618         struct mgs_config_res   *res;
619         struct ptlrpc_bulk_desc *desc;
620         struct l_wait_info lwi;
621         char               fsname[16];
622         long               bufsize;
623         int                unit_size;
624         int                type;
625         int                rc = 0;
626         int                i;
627         int                bytes;
628         int                page_count;
629         int                nrpages;
630         cfs_page_t       **pages = NULL;
631         ENTRY;
632
633         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
634         if (body == NULL)
635                 RETURN(-EINVAL);
636
637         if (body->mcb_type != CONFIG_T_RECOVER)
638                 RETURN(-EINVAL);
639
640         rc = delogname(body->mcb_name, fsname, &type);
641         if (rc)
642                 RETURN(rc);
643
644         rc = mgs_find_or_make_fsdb(env, mgs, fsname, &fsdb);
645         if (rc)
646                 RETURN(rc);
647
648         bufsize = body->mcb_units << body->mcb_bits;
649         nrpages = (bufsize + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
650         if (nrpages > PTLRPC_MAX_BRW_PAGES)
651                 RETURN(-EINVAL);
652
653         CDEBUG(D_MGS, "Reading IR log %s bufsize %ld.\n",
654                body->mcb_name, bufsize);
655
656         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
657         if (pages == NULL)
658                 RETURN(-ENOMEM);
659
660         rc = req_capsule_server_pack(&req->rq_pill);
661         if (rc)
662                 GOTO(out, rc);
663
664         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
665         if (res == NULL)
666                 GOTO(out, rc = -EINVAL);
667
668         res->mcr_offset = body->mcb_offset;
669         unit_size = min_t(int, 1 << body->mcb_bits, CFS_PAGE_SIZE);
670         bytes = mgs_nidtbl_read(req->rq_export, &fsdb->fsdb_nidtbl, res,
671                                 pages, nrpages, bufsize / unit_size, unit_size);
672         if (bytes < 0)
673                 GOTO(out, rc = bytes);
674
675         /* start bulk transfer */
676         page_count = (bytes + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
677         LASSERT(page_count <= nrpages);
678         desc = ptlrpc_prep_bulk_exp(req, page_count,
679                                     BULK_PUT_SOURCE, MGS_BULK_PORTAL);
680         if (desc == NULL)
681                 GOTO(out, rc = -ENOMEM);
682
683         for (i = 0; i < page_count && bytes > 0; i++) {
684                 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0,
685                                           min_t(int, bytes, CFS_PAGE_SIZE));
686                 bytes -= CFS_PAGE_SIZE;
687         }
688
689         rc = target_bulk_io(req->rq_export, desc, &lwi);
690         ptlrpc_free_bulk_pin(desc);
691
692 out:
693         for (i = 0; i < nrpages; i++) {
694                 if (pages[i] == NULL)
695                         break;
696                 cfs_free_page(pages[i]);
697         }
698         OBD_FREE(pages, sizeof(*pages) * nrpages);
699         return rc;
700 }
701
702 static int lprocfs_ir_set_state(struct fs_db *fsdb, const char *buf)
703 {
704         const char *strings[] = IR_STRINGS;
705         int         state = -1;
706         int         i;
707
708         for (i = 0; i < ARRAY_SIZE(strings); i++) {
709                 if (strcmp(strings[i], buf) == 0) {
710                         state = i;
711                         break;
712                 }
713         }
714         if (state < 0)
715                 return -EINVAL;
716
717         CDEBUG(D_MGS, "change fsr state of %s from %s to %s\n",
718                fsdb->fsdb_name, strings[fsdb->fsdb_ir_state], strings[state]);
719         cfs_mutex_lock(&fsdb->fsdb_mutex);
720         if (state == IR_FULL && fsdb->fsdb_nonir_clients)
721                 state = IR_PARTIAL;
722         fsdb->fsdb_ir_state = state;
723         cfs_mutex_unlock(&fsdb->fsdb_mutex);
724
725         return 0;
726 }
727
728 static int lprocfs_ir_set_timeout(struct fs_db *fsdb, const char *buf)
729 {
730         return -EINVAL;
731 }
732
733 static int lprocfs_ir_clear_stats(struct fs_db *fsdb, const char *buf)
734 {
735         if (*buf)
736                 return -EINVAL;
737
738         fsdb->fsdb_notify_total = 0;
739         fsdb->fsdb_notify_max   = 0;
740         fsdb->fsdb_notify_count = 0;
741         return 0;
742 }
743
744 static struct lproc_ir_cmd {
745         char *name;
746         int   namelen;
747         int (*handler)(struct fs_db *, const char *);
748 } ir_cmds[] = {
749         { "state=",   6, lprocfs_ir_set_state },
750         { "timeout=", 8, lprocfs_ir_set_timeout },
751         { "0",        1, lprocfs_ir_clear_stats }
752 };
753
754 int lprocfs_wr_ir_state(struct file *file, const char *buffer,
755                          unsigned long count, void *data)
756 {
757         struct fs_db *fsdb = data;
758         char *kbuf;
759         char *ptr;
760         int rc = 0;
761
762         if (count > CFS_PAGE_SIZE)
763                 return -EINVAL;
764
765         OBD_ALLOC(kbuf, count + 1);
766         if (kbuf == NULL)
767                 return -ENOMEM;
768
769         if (copy_from_user(kbuf, buffer, count)) {
770                 OBD_FREE(kbuf, count);
771                 return -EFAULT;
772         }
773
774         kbuf[count] = 0; /* buffer is supposed to end with 0 */
775         if (kbuf[count - 1] == '\n')
776                 kbuf[count - 1] = 0;
777         ptr = kbuf;
778
779         /* fsname=<file system name> must be the 1st entry */
780         while (ptr != NULL) {
781                 char *tmpptr;
782                 int i;
783
784                 tmpptr = strchr(ptr, ';');
785                 if (tmpptr)
786                         *tmpptr++ = 0;
787
788                 rc = -EINVAL;
789                 for (i = 0; i < ARRAY_SIZE(ir_cmds); i++) {
790                         struct lproc_ir_cmd *cmd;
791                         int cmdlen;
792
793                         cmd    = &ir_cmds[i];
794                         cmdlen = cmd->namelen;
795                         if (strncmp(cmd->name, ptr, cmdlen) == 0) {
796                                 ptr += cmdlen;
797                                 rc = cmd->handler(fsdb, ptr);
798                                 break;
799                         }
800                 }
801                 if (rc)
802                         break;
803
804                 ptr = tmpptr;
805         }
806         if (rc)
807                 CERROR("Unable to process command: %s(%d)\n", ptr, rc);
808         OBD_FREE(kbuf, count + 1);
809         return rc ?: count;
810 }
811
812 int lprocfs_rd_ir_state(struct seq_file *seq, void *data)
813 {
814         struct fs_db      *fsdb = data;
815         struct mgs_nidtbl *tbl  = &fsdb->fsdb_nidtbl;
816         const char        *ir_strings[] = IR_STRINGS;
817         struct timeval     tv_max;
818         struct timeval     tv;
819
820         /* mgs_live_seq_show() already holds fsdb_mutex. */
821         ir_state_graduate(fsdb);
822
823         seq_printf(seq, "\nimperative_recovery_state:\n");
824         seq_printf(seq,
825                    "    state: %s\n"
826                    "    nonir_clients: %d\n"
827                    "    nidtbl_version: %lld\n",
828                    ir_strings[fsdb->fsdb_ir_state], fsdb->fsdb_nonir_clients,
829                    tbl->mn_version);
830
831         cfs_duration_usec(fsdb->fsdb_notify_total, &tv);
832         cfs_duration_usec(fsdb->fsdb_notify_max, &tv_max);
833
834         seq_printf(seq, "    notify_duration_total: %lu.%06lu\n"
835                         "    notify_duation_max: %lu.%06lu\n"
836                         "    notify_count: %u\n",
837                    tv.tv_sec, tv.tv_usec,
838                    tv_max.tv_sec, tv_max.tv_usec,
839                    fsdb->fsdb_notify_count);
840
841         return 0;
842 }
843
844 int lprocfs_rd_ir_timeout(char *page, char **start, off_t off, int count,
845                           int *eof, void *data)
846 {
847         *eof = 1;
848         return snprintf(page, count, "%d\n", ir_timeout);
849 }
850
851 int lprocfs_wr_ir_timeout(struct file *file, const char *buffer,
852                           unsigned long count, void *data)
853 {
854         return lprocfs_wr_uint(file, buffer, count, &ir_timeout);
855 }
856
857 /* --------------- Handle non IR support clients --------------- */
858 /* attach a lustre file system to an export */
859 int mgs_fsc_attach(const struct lu_env *env, struct obd_export *exp,
860                    char *fsname)
861 {
862         struct mgs_export_data *data = &exp->u.eu_mgs_data;
863         struct mgs_device *mgs = exp2mgs_dev(exp);
864         struct fs_db      *fsdb;
865         struct mgs_fsc    *fsc     = NULL;
866         struct mgs_fsc    *new_fsc = NULL;
867         bool               found   = false;
868         int                rc;
869         ENTRY;
870
871         rc = mgs_find_or_make_fsdb(env, mgs, fsname, &fsdb);
872         if (rc)
873                 RETURN(rc);
874
875         /* allocate a new fsc in case we need it in spinlock. */
876         OBD_ALLOC_PTR(new_fsc);
877         if (new_fsc == NULL)
878                 RETURN(-ENOMEM);
879
880         CFS_INIT_LIST_HEAD(&new_fsc->mfc_export_list);
881         CFS_INIT_LIST_HEAD(&new_fsc->mfc_fsdb_list);
882         new_fsc->mfc_fsdb       = fsdb;
883         new_fsc->mfc_export     = class_export_get(exp);
884         new_fsc->mfc_ir_capable =
885                         !!(exp->exp_connect_flags & OBD_CONNECT_IMP_RECOV);
886
887         rc = -EEXIST;
888         cfs_mutex_lock(&fsdb->fsdb_mutex);
889
890         /* tend to find it in export list because this list is shorter. */
891         cfs_spin_lock(&data->med_lock);
892         cfs_list_for_each_entry(fsc, &data->med_clients, mfc_export_list) {
893                 if (strcmp(fsname, fsc->mfc_fsdb->fsdb_name) == 0) {
894                         found = true;
895                         break;
896                 }
897         }
898         if (!found) {
899                 fsc = new_fsc;
900                 new_fsc = NULL;
901
902                 /* add it into export list. */
903                 cfs_list_add(&fsc->mfc_export_list, &data->med_clients);
904
905                 /* add into fsdb list. */
906                 cfs_list_add(&fsc->mfc_fsdb_list, &fsdb->fsdb_clients);
907                 if (!fsc->mfc_ir_capable) {
908                         ++fsdb->fsdb_nonir_clients;
909                         if (fsdb->fsdb_ir_state == IR_FULL)
910                                 fsdb->fsdb_ir_state = IR_PARTIAL;
911                 }
912                 rc = 0;
913         }
914         cfs_spin_unlock(&data->med_lock);
915         cfs_mutex_unlock(&fsdb->fsdb_mutex);
916
917         if (new_fsc) {
918                 class_export_put(new_fsc->mfc_export);
919                 OBD_FREE_PTR(new_fsc);
920         }
921         RETURN(rc);
922 }
923
924 void mgs_fsc_cleanup(struct obd_export *exp)
925 {
926         struct mgs_export_data *data = &exp->u.eu_mgs_data;
927         struct mgs_fsc *fsc, *tmp;
928         CFS_LIST_HEAD(head);
929
930         cfs_spin_lock(&data->med_lock);
931         cfs_list_splice_init(&data->med_clients, &head);
932         cfs_spin_unlock(&data->med_lock);
933
934         cfs_list_for_each_entry_safe(fsc, tmp, &head, mfc_export_list) {
935                 struct fs_db *fsdb = fsc->mfc_fsdb;
936
937                 LASSERT(fsc->mfc_export == exp);
938
939                 cfs_mutex_lock(&fsdb->fsdb_mutex);
940                 cfs_list_del_init(&fsc->mfc_fsdb_list);
941                 if (fsc->mfc_ir_capable == 0) {
942                         --fsdb->fsdb_nonir_clients;
943                         LASSERT(fsdb->fsdb_ir_state != IR_FULL);
944                         if (fsdb->fsdb_nonir_clients == 0 &&
945                             fsdb->fsdb_ir_state == IR_PARTIAL)
946                                 fsdb->fsdb_ir_state = IR_FULL;
947                 }
948                 cfs_mutex_unlock(&fsdb->fsdb_mutex);
949                 cfs_list_del_init(&fsc->mfc_export_list);
950                 class_export_put(fsc->mfc_export);
951                 OBD_FREE_PTR(fsc);
952         }
953 }
954
955 /* must be called with fsdb->fsdb_mutex held */
956 void mgs_fsc_cleanup_by_fsdb(struct fs_db *fsdb)
957 {
958         struct mgs_fsc *fsc, *tmp;
959
960         cfs_list_for_each_entry_safe(fsc, tmp, &fsdb->fsdb_clients,
961                                      mfc_fsdb_list) {
962                 struct mgs_export_data *data = &fsc->mfc_export->u.eu_mgs_data;
963
964                 LASSERT(fsdb == fsc->mfc_fsdb);
965                 cfs_list_del_init(&fsc->mfc_fsdb_list);
966
967                 cfs_spin_lock(&data->med_lock);
968                 cfs_list_del_init(&fsc->mfc_export_list);
969                 cfs_spin_unlock(&data->med_lock);
970                 class_export_put(fsc->mfc_export);
971                 OBD_FREE_PTR(fsc);
972         }
973
974         fsdb->fsdb_nonir_clients = 0;
975         if (fsdb->fsdb_ir_state == IR_PARTIAL)
976                 fsdb->fsdb_ir_state = IR_FULL;
977 }