Whamcloud - gitweb
LU-2275 mdt: Avoid setting positive dispositions too early
[fs/lustre-release.git] / lustre / mdt / mdt_recovery.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, 2013, Intel Corporation.
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/mdt/mdt_recovery.c
37  *
38  * Lustre Metadata Target (mdt) recovery-related methods
39  *
40  * Author: Huang Hua <huanghua@clusterfs.com>
41  * Author: Pershin Mike <tappro@clusterfs.com>
42  */
43
44 #define DEBUG_SUBSYSTEM S_MDS
45
46 #include "mdt_internal.h"
47
48 struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len)
49 {
50         struct lu_buf *buf;
51         struct mdt_thread_info *mti;
52
53         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
54         buf = &mti->mti_buf;
55         buf->lb_buf = area;
56         buf->lb_len = len;
57         return buf;
58 }
59
60 const struct lu_buf *mdt_buf_const(const struct lu_env *env,
61                                    const void *area, ssize_t len)
62 {
63         struct lu_buf *buf;
64         struct mdt_thread_info *mti;
65
66         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
67         buf = &mti->mti_buf;
68
69         buf->lb_buf = (void *)area;
70         buf->lb_len = len;
71         return buf;
72 }
73
74 void mdt_trans_stop(const struct lu_env *env,
75                     struct mdt_device *mdt, struct thandle *th)
76 {
77         dt_trans_stop(env, mdt->mdt_bottom, th);
78 }
79
80 static int mdt_clients_data_init(const struct lu_env *env,
81                                  struct mdt_device *mdt,
82                                  unsigned long last_size)
83 {
84         struct lr_server_data  *lsd = &mdt->mdt_lut.lut_lsd;
85         struct lsd_client_data *lcd;
86         struct obd_device      *obd = mdt2obd_dev(mdt);
87         loff_t off;
88         int cl_idx;
89         int rc = 0;
90         ENTRY;
91
92         OBD_ALLOC_PTR(lcd);
93         if (!lcd)
94                 RETURN(-ENOMEM);
95
96         /* When we do a clean MDS shutdown, we save the last_transno into
97          * the header.  If we find clients with higher last_transno values
98          * then those clients may need recovery done. */
99         LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
100         for (cl_idx = 0, off = lsd->lsd_client_start;
101              off < last_size; cl_idx++) {
102                 __u64 last_transno;
103                 struct obd_export *exp;
104                 struct mdt_thread_info *mti;
105
106                 off = lsd->lsd_client_start +
107                         cl_idx * lsd->lsd_client_size;
108
109                 rc = tgt_client_data_read(env, &mdt->mdt_lut, lcd, &off, cl_idx);
110                 if (rc) {
111                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
112                                LAST_RCVD, cl_idx, off, rc);
113                         rc = 0;
114                         break; /* read error shouldn't cause startup to fail */
115                 }
116
117                 if (lcd->lcd_uuid[0] == '\0') {
118                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
119                                cl_idx);
120                         continue;
121                 }
122
123                 last_transno = lcd_last_transno(lcd);
124
125                 /* These exports are cleaned up by mdt_obd_disconnect(), so
126                  * they need to be set up like real exports as
127                  * mdt_obd_connect() does.
128                  */
129                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
130                        " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
131                        last_transno, lsd->lsd_last_transno,
132                        lcd_last_xid(lcd));
133
134                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
135                 if (IS_ERR(exp)) {
136                         if (PTR_ERR(exp) == -EALREADY) {
137                                 /* export already exists, zero out this one */
138                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
139                                 continue;
140                         }
141                         GOTO(err_client, rc = PTR_ERR(exp));
142                 }
143
144                 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
145                 LASSERT(mti != NULL);
146                 mti->mti_exp = exp;
147                 /* copy on-disk lcd to the export */
148                 *exp->exp_target_data.ted_lcd = *lcd;
149                 rc = tgt_client_add(env, exp, cl_idx);
150                 /* can't fail existing */
151                 LASSERTF(rc == 0, "rc = %d\n", rc);
152                 /* VBR: set export last committed version */
153                 exp->exp_last_committed = last_transno;
154                 spin_lock(&exp->exp_lock);
155                 exp->exp_connecting = 0;
156                 exp->exp_in_recovery = 0;
157                 spin_unlock(&exp->exp_lock);
158                 obd->obd_max_recoverable_clients++;
159                 class_export_put(exp);
160
161                 CDEBUG(D_OTHER, "client at idx %d has last_transno ="LPU64"\n",
162                        cl_idx, last_transno);
163                 /* protect __u64 value update */
164                 spin_lock(&mdt->mdt_lut.lut_translock);
165                 mdt->mdt_lut.lut_last_transno = max(last_transno,
166                                                 mdt->mdt_lut.lut_last_transno);
167                 spin_unlock(&mdt->mdt_lut.lut_translock);
168         }
169
170 err_client:
171         OBD_FREE_PTR(lcd);
172         RETURN(rc);
173 }
174
175 static int mdt_server_data_init(const struct lu_env *env,
176                                 struct mdt_device *mdt,
177                                 struct lustre_sb_info *lsi)
178 {
179         struct lr_server_data  *lsd = &mdt->mdt_lut.lut_lsd;
180         struct lsd_client_data *lcd = NULL;
181         struct obd_device      *obd = mdt2obd_dev(mdt);
182         struct mdt_thread_info *mti;
183         struct dt_object       *obj;
184         struct lu_attr         *la;
185         unsigned long last_rcvd_size;
186         __u64 mount_count;
187         int rc;
188         ENTRY;
189
190         /* ensure padding in the struct is the correct size */
191         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
192                 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
193         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
194                 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
195
196         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
197         LASSERT(mti != NULL);
198         la = &mti->mti_attr.ma_attr;
199
200         obj = mdt->mdt_lut.lut_last_rcvd;
201         rc = dt_attr_get(env, obj, la, BYPASS_CAPA);
202         if (rc)
203                 RETURN(rc);
204
205         last_rcvd_size = (unsigned long)la->la_size;
206
207         if (last_rcvd_size == 0) {
208                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
209
210                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
211                        sizeof(lsd->lsd_uuid));
212                 lsd->lsd_last_transno = 0;
213                 lsd->lsd_mount_count = 0;
214                 lsd->lsd_server_size = LR_SERVER_SIZE;
215                 lsd->lsd_client_start = LR_CLIENT_START;
216                 lsd->lsd_client_size = LR_CLIENT_SIZE;
217                 lsd->lsd_feature_compat = OBD_COMPAT_MDT;
218                 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
219                 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
220                                             OBD_INCOMPAT_COMMON_LR |
221                                             OBD_INCOMPAT_MULTI_OI;
222         } else {
223                 LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name);
224                 rc = tgt_server_data_read(env, &mdt->mdt_lut);
225                 if (rc) {
226                         CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
227                         GOTO(out, rc);
228                 }
229                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
230                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
231                                            "the wrong disk %s. Were the /dev/ "
232                                            "assignments rearranged?\n",
233                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
234                         GOTO(out, rc = -EINVAL);
235                 }
236                 lsd->lsd_feature_compat |= OBD_COMPAT_MDT;
237                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_MDT |
238                                              OBD_INCOMPAT_COMMON_LR;
239         }
240         mount_count = lsd->lsd_mount_count;
241
242         if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) {
243                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
244                        obd->obd_name,
245                        lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP);
246                 GOTO(out, rc = -EINVAL);
247         }
248         if (lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP) {
249                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
250                        obd->obd_name,
251                        lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP);
252                 /* XXX: Do something like remount filesystem read-only */
253                 GOTO(out, rc = -EINVAL);
254         }
255         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
256         if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
257                 if (last_rcvd_size > lsd->lsd_client_start) {
258                         LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, "
259                                       "remove all clients for interop needs\n",
260                                       obd->obd_name);
261                         rc = tgt_truncate_last_rcvd(env, &mdt->mdt_lut,
262                                                     lsd->lsd_client_start);
263                         if (rc)
264                                 GOTO(out, rc);
265                         last_rcvd_size = lsd->lsd_client_start;
266                 }
267                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
268                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
269         }
270
271         lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
272
273         spin_lock(&mdt->mdt_lut.lut_translock);
274         mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno;
275         spin_unlock(&mdt->mdt_lut.lut_translock);
276
277         CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
278         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
279                obd->obd_name, mdt->mdt_lut.lut_last_transno);
280         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
281                obd->obd_name, mount_count + 1);
282         CDEBUG(D_INODE, "%s: server data size: %u\n",
283                obd->obd_name, lsd->lsd_server_size);
284         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
285                obd->obd_name, lsd->lsd_client_start);
286         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
287                obd->obd_name, lsd->lsd_client_size);
288         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
289                obd->obd_name, last_rcvd_size);
290         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
291                last_rcvd_size <= lsd->lsd_client_start ? 0 :
292                (last_rcvd_size - lsd->lsd_client_start) /
293                 lsd->lsd_client_size);
294         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
295
296         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
297             !lsd->lsd_client_size) {
298                 CERROR("Bad last_rcvd contents!\n");
299                 GOTO(out, rc = -EINVAL);
300         }
301
302         rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
303         if (rc)
304                 GOTO(err_client, rc);
305
306         spin_lock(&mdt->mdt_lut.lut_translock);
307         /* obd_last_committed is used for compatibility
308          * with other lustre recovery code */
309         obd->obd_last_committed = mdt->mdt_lut.lut_last_transno;
310         spin_unlock(&mdt->mdt_lut.lut_translock);
311
312         obd->u.obt.obt_mount_count = mount_count + 1;
313         obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
314         lsd->lsd_mount_count = obd->u.obt.obt_mount_count;
315
316         /* save it, so mount count and last_transno is current */
317         rc = tgt_server_data_update(env, &mdt->mdt_lut, 0);
318         if (rc)
319                 GOTO(err_client, rc);
320
321         RETURN(0);
322
323 err_client:
324         class_disconnect_exports(obd);
325 out:
326         return rc;
327 }
328
329 /*
330  * last_rcvd & last_committed update callbacks
331  */
332 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
333                                 struct thandle *th)
334 {
335         struct mdt_device *mdt = mti->mti_mdt;
336         struct ptlrpc_request *req = mdt_info_req(mti);
337         struct tg_export_data *ted;
338         struct lsd_client_data *lcd;
339         loff_t off;
340         int err;
341         __s32 rc = th->th_result;
342
343         ENTRY;
344         LASSERT(req);
345         LASSERT(req->rq_export);
346         LASSERT(mdt);
347         ted = &req->rq_export->exp_target_data;
348         LASSERT(ted);
349
350         mutex_lock(&ted->ted_lcd_lock);
351         lcd = ted->ted_lcd;
352         /* if the export has already been disconnected, we have no last_rcvd
353          * slot, update server data with latest transno then */
354         if (lcd == NULL) {
355                 mutex_unlock(&ted->ted_lcd_lock);
356                 CWARN("commit transaction for disconnected client %s: rc %d\n",
357                       req->rq_export->exp_client_uuid.uuid, rc);
358                 err = tgt_server_data_write(mti->mti_env, &mdt->mdt_lut, th);
359                 RETURN(err);
360         }
361
362         off = ted->ted_lr_off;
363         LASSERT(ergo(mti->mti_transno == 0, rc != 0));
364         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
365             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
366                 if (mti->mti_transno != 0) {
367                         if (lcd->lcd_last_close_transno > mti->mti_transno) {
368                                 CERROR("Trying to overwrite bigger transno:"
369                                        "on-disk: "LPU64", new: "LPU64" "
370                                        "replay: %d. see LU-617.\n",
371                                        lcd->lcd_last_close_transno,
372                                        mti->mti_transno, req_is_replay(req));
373                                 if (req_is_replay(req)) {
374                                         spin_lock(&req->rq_export->exp_lock);
375                                         req->rq_export->exp_vbr_failed = 1;
376                                         spin_unlock(&req->rq_export->exp_lock);
377                                 }
378                                 mutex_unlock(&ted->ted_lcd_lock);
379                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
380                         }
381                         lcd->lcd_last_close_transno = mti->mti_transno;
382                 }
383                 lcd->lcd_last_close_xid = req->rq_xid;
384                 lcd->lcd_last_close_result = rc;
385         } else {
386                 /* VBR: save versions in last_rcvd for reconstruct. */
387                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
388                 if (pre_versions) {
389                         lcd->lcd_pre_versions[0] = pre_versions[0];
390                         lcd->lcd_pre_versions[1] = pre_versions[1];
391                         lcd->lcd_pre_versions[2] = pre_versions[2];
392                         lcd->lcd_pre_versions[3] = pre_versions[3];
393                 }
394                 if (mti->mti_transno != 0) {
395                         if (lcd->lcd_last_transno > mti->mti_transno) {
396                                 CERROR("Trying to overwrite bigger transno:"
397                                        "on-disk: "LPU64", new: "LPU64" "
398                                        "replay: %d. see LU-617.\n",
399                                        lcd->lcd_last_transno,
400                                        mti->mti_transno, req_is_replay(req));
401                                 if (req_is_replay(req)) {
402                                         spin_lock(&req->rq_export->exp_lock);
403                                         req->rq_export->exp_vbr_failed = 1;
404                                         spin_unlock(&req->rq_export->exp_lock);
405                                 }
406                                 mutex_unlock(&ted->ted_lcd_lock);
407                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
408                         }
409                         lcd->lcd_last_transno = mti->mti_transno;
410                 }
411                 lcd->lcd_last_xid = req->rq_xid;
412                 lcd->lcd_last_result = rc;
413                 /*XXX: save intent_disposition in mdt_thread_info?
414                  * also there is bug - intent_dispostion is __u64,
415                  * see struct ldlm_reply->lock_policy_res1; */
416                 lcd->lcd_last_data = mti->mti_opdata;
417         }
418
419         if ((mti->mti_exp->exp_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0) {
420                 /* Although lightweight (LW) connections have no slot in
421                  * last_rcvd, we still want to maintain the in-memory
422                  * lsd_client_data structure in order to properly handle reply
423                  * reconstruction. */
424                 struct lu_target        *tg = &mdt->mdt_lut;
425                 bool                     update = false;
426
427                 mutex_unlock(&ted->ted_lcd_lock);
428                 err = 0;
429
430                 /* All operations performed by LW clients are synchronous and
431                  * we store the committed transno in the last_rcvd header */
432                 spin_lock(&tg->lut_translock);
433                 if (mti->mti_transno > tg->lut_lsd.lsd_last_transno) {
434                         tg->lut_lsd.lsd_last_transno = mti->mti_transno;
435                         update = true;
436                 }
437                 spin_unlock(&tg->lut_translock);
438
439                 if (update)
440                         err = tgt_server_data_write(mti->mti_env, tg, th);
441         } else if (off <= 0) {
442                 CERROR("%s: client idx %d has offset %lld\n",
443                        mdt2obd_dev(mdt)->obd_name, ted->ted_lr_idx, off);
444                 mutex_unlock(&ted->ted_lcd_lock);
445                 err = -EINVAL;
446         } else {
447                 err = tgt_client_data_write(mti->mti_env, &mdt->mdt_lut, lcd,
448                                             &off, th);
449                 mutex_unlock(&ted->ted_lcd_lock);
450         }
451         RETURN(err);
452 }
453
454 extern struct lu_context_key mdt_thread_key;
455
456 /* add credits for last_rcvd update */
457 static int mdt_txn_start_cb(const struct lu_env *env,
458                             struct thandle *th, void *cookie)
459 {
460         struct mdt_device *mdt = cookie;
461         struct mdt_thread_info *mti;
462         int rc;
463         ENTRY;
464
465         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
466
467         LASSERT(mdt->mdt_lut.lut_last_rcvd);
468         if (mti->mti_exp == NULL)
469                 RETURN(0);
470
471         rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
472                                      sizeof(struct lsd_client_data),
473                                      mti->mti_exp->exp_target_data.ted_lr_off,
474                                      th);
475         if (rc)
476                 return rc;
477
478         rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
479                                      sizeof(struct lr_server_data), 0, th);
480         if (rc)
481                 return rc;
482
483         if (mti->mti_mos != NULL)
484                 rc = dt_declare_version_set(env, mdt_obj2dt(mti->mti_mos), th);
485
486         return rc;
487 }
488
489 /* Update last_rcvd records with latests transaction data */
490 static int mdt_txn_stop_cb(const struct lu_env *env,
491                            struct thandle *txn, void *cookie)
492 {
493         struct mdt_device *mdt = cookie;
494         struct mdt_thread_info *mti;
495         struct ptlrpc_request *req;
496
497         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
498         req = mdt_info_req(mti);
499
500         if (mti->mti_mdt == NULL || req == NULL)
501                 return 0;
502
503         if (mti->mti_has_trans) {
504                 /* XXX: currently there are allowed cases, but the wrong cases
505                  * are also possible, so better check is needed here */
506                 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
507                        mti->mti_transno);
508                 return 0;
509         }
510
511         mti->mti_has_trans = 1;
512         spin_lock(&mdt->mdt_lut.lut_translock);
513         if (txn->th_result != 0) {
514                 if (mti->mti_transno != 0) {
515                         CERROR("Replay transno "LPU64" failed: rc %d\n",
516                                 mti->mti_transno, txn->th_result);
517                         return 0;
518                 }
519         } else if (mti->mti_transno == 0) {
520                 mti->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
521         } else {
522                 /* should be replay */
523                 if (mti->mti_transno > mdt->mdt_lut.lut_last_transno)
524                         mdt->mdt_lut.lut_last_transno = mti->mti_transno;
525         }
526         spin_unlock(&mdt->mdt_lut.lut_translock);
527         /* sometimes the reply message has not been successfully packed */
528         LASSERT(req != NULL && req->rq_repmsg != NULL);
529
530         /** VBR: set new versions */
531         if (txn->th_result == 0 && mti->mti_mos != NULL) {
532                 dt_version_set(env, mdt_obj2dt(mti->mti_mos),
533                                mti->mti_transno, txn);
534                 mti->mti_mos = NULL;
535         }
536
537         /* filling reply data */
538         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
539                mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
540
541         req->rq_transno = mti->mti_transno;
542         lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
543         /* if can't add callback, do sync write */
544         txn->th_sync |= !!tgt_last_commit_cb_add(txn, &mdt->mdt_lut,
545                                                  mti->mti_exp,
546                                                  mti->mti_transno);
547         return mdt_last_rcvd_update(mti, txn);
548 }
549
550 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
551                  struct obd_device *obd,
552                  struct lustre_sb_info *lsi)
553 {
554         int rc = 0;
555         ENTRY;
556
557         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
558                 RETURN(-ENOENT);
559
560         /* prepare transactions callbacks */
561         mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
562         mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
563         mdt->mdt_txn_cb.dtc_txn_commit = NULL;
564         mdt->mdt_txn_cb.dtc_cookie = mdt;
565         mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
566         CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
567
568         dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
569
570         rc = mdt_server_data_init(env, mdt, lsi);
571
572         RETURN(rc);
573 }
574
575 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
576 {
577         ENTRY;
578
579         /* Remove transaction callback */
580         dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
581         if (mdt->mdt_ck_obj)
582                 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
583         mdt->mdt_ck_obj = NULL;
584         EXIT;
585 }
586
587 /* reconstruction code */
588 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
589 {
590         struct ptlrpc_service_part *svcpt;
591         struct obd_export         *exp = req->rq_export;
592         cfs_list_t                *tmp;
593         struct ptlrpc_reply_state *oldrep;
594         int                        i;
595
596         /* CAVEAT EMPTOR: spinlock order */
597         spin_lock(&exp->exp_lock);
598         cfs_list_for_each (tmp, &exp->exp_outstanding_replies) {
599                 oldrep = cfs_list_entry(tmp, struct ptlrpc_reply_state,
600                                         rs_exp_list);
601
602                 if (oldrep->rs_xid != req->rq_xid)
603                         continue;
604
605                 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
606                         CERROR ("Resent req xid "LPU64" has mismatched opc: "
607                                 "new %d old %d\n", req->rq_xid,
608                                 lustre_msg_get_opc(req->rq_reqmsg),
609                                 oldrep->rs_opc);
610
611                 svcpt = oldrep->rs_svcpt;
612                 spin_lock(&svcpt->scp_rep_lock);
613
614                 cfs_list_del_init (&oldrep->rs_exp_list);
615
616                 CDEBUG(D_HA, "Stealing %d locks from rs %p x"LPD64".t"LPD64
617                        " o%d NID %s\n",
618                        oldrep->rs_nlocks, oldrep,
619                        oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
620                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
621
622                 for (i = 0; i < oldrep->rs_nlocks; i++)
623                         ptlrpc_save_lock(req, &oldrep->rs_locks[i],
624                                          oldrep->rs_modes[i], 0);
625                 oldrep->rs_nlocks = 0;
626
627                 DEBUG_REQ(D_HA, req, "stole locks for");
628                 spin_lock(&oldrep->rs_lock);
629                 ptlrpc_schedule_difficult_reply(oldrep);
630                 spin_unlock(&oldrep->rs_lock);
631
632                 spin_unlock(&svcpt->scp_rep_lock);
633                 break;
634         }
635         spin_unlock(&exp->exp_lock);
636 }
637
638 /**
639  * VBR: restore versions
640  */
641 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
642                          struct lsd_client_data *lcd)
643 {
644         __u64 pre_versions[4] = {0};
645         pre_versions[0] = lcd->lcd_pre_versions[0];
646         pre_versions[1] = lcd->lcd_pre_versions[1];
647         pre_versions[2] = lcd->lcd_pre_versions[2];
648         pre_versions[3] = lcd->lcd_pre_versions[3];
649         lustre_msg_set_versions(req->rq_repmsg, pre_versions);
650 }
651
652 void mdt_req_from_lcd(struct ptlrpc_request *req,
653                       struct lsd_client_data *lcd)
654 {
655         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
656                   lcd->lcd_last_transno, lcd->lcd_last_result);
657
658         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
659             lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
660                 req->rq_transno = lcd->lcd_last_close_transno;
661                 req->rq_status = lcd->lcd_last_close_result;
662         } else {
663                 req->rq_transno = lcd->lcd_last_transno;
664                 req->rq_status = lcd->lcd_last_result;
665                 mdt_vbr_reconstruct(req, lcd);
666         }
667         if (req->rq_status != 0)
668                 req->rq_transno = 0;
669         lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
670         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
671         DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
672                   req->rq_transno, req->rq_status);
673
674         mdt_steal_ack_locks(req);
675 }
676
677 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
678                              struct mdt_lock_handle *lhc)
679 {
680         struct ptlrpc_request *req = mdt_info_req(mti);
681         struct tg_export_data *ted = &req->rq_export->exp_target_data;
682
683         return mdt_req_from_lcd(req, ted->ted_lcd);
684 }
685
686 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
687                                    struct mdt_lock_handle *lhc)
688 {
689         struct ptlrpc_request  *req = mdt_info_req(mti);
690         struct obd_export *exp = req->rq_export;
691         struct tg_export_data *ted = &exp->exp_target_data;
692         struct mdt_device *mdt = mti->mti_mdt;
693         struct mdt_object *child;
694         struct mdt_body *body;
695         int rc;
696
697         mdt_req_from_lcd(req, ted->ted_lcd);
698         if (req->rq_status)
699                 return;
700
701         /* if no error, so child was created with requested fid */
702         child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
703         if (IS_ERR(child)) {
704                 rc = PTR_ERR(child);
705                 LCONSOLE_WARN("Child "DFID" lookup error %d."
706                               " Evicting client %s with export %s.\n",
707                               PFID(mdt_object_fid(child)), rc,
708                               obd_uuid2str(&exp->exp_client_uuid),
709                               obd_export_nid2str(exp));
710                 mdt_export_evict(exp);
711                 EXIT;
712                 return;
713         }
714
715         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
716         mti->mti_attr.ma_need = MA_INODE;
717         mti->mti_attr.ma_valid = 0;
718         rc = mdt_attr_get_complex(mti, child, &mti->mti_attr);
719         if (rc == -EREMOTE) {
720                 /* object was created on remote server */
721                 req->rq_status = rc;
722                 body->valid |= OBD_MD_MDS;
723         }
724         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
725                            mdt_object_fid(child));
726         mdt_object_put(mti->mti_env, child);
727 }
728
729 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
730                                     struct mdt_lock_handle *lhc)
731 {
732         struct ptlrpc_request  *req = mdt_info_req(mti);
733         struct obd_export *exp = req->rq_export;
734         struct mdt_export_data *med = &exp->exp_mdt_data;
735         struct mdt_device *mdt = mti->mti_mdt;
736         struct mdt_object *obj;
737         struct mdt_body *body;
738
739         mdt_req_from_lcd(req, med->med_ted.ted_lcd);
740         if (req->rq_status)
741                 return;
742
743         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
744         obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
745         if (IS_ERR(obj)) {
746                 int rc = PTR_ERR(obj);
747                 LCONSOLE_WARN(""DFID" lookup error %d."
748                               " Evicting client %s with export %s.\n",
749                               PFID(mdt_object_fid(obj)), rc,
750                               obd_uuid2str(&exp->exp_client_uuid),
751                               obd_export_nid2str(exp));
752                 mdt_export_evict(exp);
753                 EXIT;
754                 return;
755         }
756         mti->mti_attr.ma_need = MA_INODE;
757         mti->mti_attr.ma_valid = 0;
758         mdt_attr_get_complex(mti, obj, &mti->mti_attr);
759         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
760                            mdt_object_fid(obj));
761         if (mti->mti_ioepoch && (mti->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
762                 struct mdt_file_data *mfd;
763                 struct mdt_body *repbody;
764
765                 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
766                 repbody->ioepoch = obj->mot_ioepoch;
767                 spin_lock(&med->med_open_lock);
768                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
769                         if (mfd->mfd_xid == req->rq_xid)
770                                 break;
771                 }
772                 LASSERT(&mfd->mfd_list != &med->med_open_head);
773                 spin_unlock(&med->med_open_lock);
774                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
775         }
776
777         mdt_object_put(mti->mti_env, obj);
778 }
779
780 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
781                                   struct mdt_lock_handle *lhc);
782
783 static mdt_reconstructor reconstructors[REINT_MAX] = {
784         [REINT_SETATTR]  = mdt_reconstruct_setattr,
785         [REINT_CREATE]   = mdt_reconstruct_create,
786         [REINT_LINK]     = mdt_reconstruct_generic,
787         [REINT_UNLINK]   = mdt_reconstruct_generic,
788         [REINT_RENAME]   = mdt_reconstruct_generic,
789         [REINT_OPEN]     = mdt_reconstruct_open,
790         [REINT_SETXATTR] = mdt_reconstruct_generic
791 };
792
793 void mdt_reconstruct(struct mdt_thread_info *mti,
794                      struct mdt_lock_handle *lhc)
795 {
796         ENTRY;
797         reconstructors[mti->mti_rr.rr_opcode](mti, lhc);
798         EXIT;
799 }