Whamcloud - gitweb
LU-6475 mdt: race between open and migrate
[fs/lustre-release.git] / lustre / target / update_trans.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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2014, Intel Corporation.
24  */
25 /*
26  * lustre/target/update_trans.c
27  *
28  * This file implements the update distribute transaction API.
29  *
30  * To manage the cross-MDT operation (distribute operation) transaction,
31  * the transaction will also be separated two layers on MD stack, top
32  * transaction and sub transaction.
33  *
34  * During the distribute operation, top transaction is created in the LOD
35  * layer, and represent the operation. Sub transaction is created by
36  * each OSD or OSP. Top transaction start/stop will trigger all of its sub
37  * transaction start/stop. Top transaction (the whole operation) is committed
38  * only all of its sub transaction are committed.
39  *
40  * there are three kinds of transactions
41  * 1. local transaction: All updates are in a single local OSD.
42  * 2. Remote transaction: All Updates are only in the remote OSD,
43  *    i.e. locally all updates are in OSP.
44  * 3. Mixed transaction: Updates are both in local OSD and remote
45  *    OSD.
46  *
47  * Author: Di Wang <di.wang@intel.com>
48  */
49
50 #define DEBUG_SUBSYSTEM S_CLASS
51
52 #include <linux/kthread.h>
53 #include <lu_target.h>
54 #include <lustre_log.h>
55 #include <lustre_update.h>
56 #include <obd.h>
57 #include <obd_class.h>
58 #include <tgt_internal.h>
59
60 #include <tgt_internal.h>
61 /**
62  * Dump top mulitple thandle
63  *
64  * Dump top multiple thandle and all of its sub thandle to the debug log.
65  *
66  * \param[in]mask       debug mask
67  * \param[in]top_th     top_thandle to be dumped
68  */
69 static void top_multiple_thandle_dump(struct top_multiple_thandle *tmt,
70                                       __u32 mask)
71 {
72         struct sub_thandle      *st;
73
74         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
75         CDEBUG(mask, "%s tmt %p refcount %d committed %d result %d"
76                "batchid "LPU64"\n",
77                tmt->tmt_master_sub_dt ?
78                tmt->tmt_master_sub_dt->dd_lu_dev.ld_obd->obd_name :
79                "NULL",
80                tmt, atomic_read(&tmt->tmt_refcount), tmt->tmt_committed,
81                tmt->tmt_result, tmt->tmt_batchid);
82
83         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
84                 struct sub_thandle_cookie *stc;
85
86                 CDEBUG(mask, "st %p obd %s committed %d sub_th %p\n",
87                        st, st->st_dt->dd_lu_dev.ld_obd->obd_name,
88                        st->st_committed, st->st_sub_th);
89
90                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
91                         CDEBUG(mask, " cookie "DOSTID": %u\n",
92                                POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
93                                stc->stc_cookie.lgc_index);
94                 }
95         }
96 }
97
98 /**
99  * Declare write update to sub device
100  *
101  * Declare Write updates llog records to the sub device during distribute
102  * transaction.
103  *
104  * \param[in] env       execution environment
105  * \param[in] record    update records being written
106  * \param[in] sub_th    sub transaction handle
107  * \param[in] record_size total update record size
108  *
109  * \retval              0 if writing succeeds
110  * \retval              negative errno if writing fails
111  */
112 static int sub_declare_updates_write(const struct lu_env *env,
113                                      struct llog_update_record *record,
114                                      struct thandle *sub_th, size_t record_size)
115 {
116         struct llog_ctxt        *ctxt;
117         struct dt_device        *dt = sub_th->th_dev;
118         int                     left = record_size;
119         int rc;
120
121         /* If ctxt is NULL, it means not need to write update,
122          * for example if the the OSP is used to connect to OST */
123         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
124                                 LLOG_UPDATELOG_ORIG_CTXT);
125
126         /* Not ready to record updates yet. */
127         if (ctxt == NULL || ctxt->loc_handle == NULL) {
128                 llog_ctxt_put(ctxt);
129                 return 0;
130         }
131
132         rc = llog_declare_add(env, ctxt->loc_handle,
133                               &record->lur_hdr, sub_th);
134         if (rc < 0)
135                 GOTO(out_put, rc);
136
137         while (left > ctxt->loc_chunk_size) {
138                 rc = llog_declare_add(env, ctxt->loc_handle,
139                                       &record->lur_hdr, sub_th);
140                 if (rc < 0)
141                         GOTO(out_put, rc);
142
143                 left -= ctxt->loc_chunk_size;
144         }
145
146 out_put:
147         llog_ctxt_put(ctxt);
148
149         return rc;
150 }
151
152 /**
153  * write update to sub device
154  *
155  * Write llog update record to the sub device during distribute
156  * transaction. If it succeeds, llog cookie of the record will be
157  * returned by @cookie.
158  *
159  * \param[in] env       execution environment
160  * \param[in] record    update records being written
161  * \param[in] sub_th    sub transaction handle
162  * \param[out] cookie   llog cookie of the update record.
163  *
164  * \retval              1 if writing succeeds
165  * \retval              negative errno if writing fails
166  */
167 static int sub_updates_write(const struct lu_env *env,
168                              struct llog_update_record *record,
169                              struct sub_thandle *sub_th)
170 {
171         struct dt_device        *dt = sub_th->st_dt;
172         struct llog_ctxt        *ctxt;
173         int                     rc;
174         struct llog_update_record *lur = NULL;
175         struct update_params    *params = NULL;
176         __u32                   update_count = 0;
177         __u32                   param_count = 0;
178         __u32                   last_update_count = 0;
179         __u32                   last_param_count = 0;
180         void                    *src;
181         void                    *start;
182         void                    *next;
183         struct sub_thandle_cookie *stc;
184         ENTRY;
185
186         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
187                                 LLOG_UPDATELOG_ORIG_CTXT);
188         /* If ctxt == NULL, then it means updates on OST (only happens
189          * during migration), and we do not track those updates for now */
190         /* If ctxt->loc_handle == NULL, then it does not need to record
191          * update, usually happens in error handler path */
192         if (ctxt == NULL || ctxt->loc_handle == NULL) {
193                 llog_ctxt_put(ctxt);
194                 RETURN(0);
195         }
196
197         /* Since the cross-MDT updates will includes both local
198          * and remote updates, the update ops count must > 1 */
199         LASSERT(record->lur_update_rec.ur_update_count > 1);
200         LASSERTF(record->lur_hdr.lrh_len == llog_update_record_size(record),
201                  "lrh_len %u record_size %zu\n", record->lur_hdr.lrh_len,
202                  llog_update_record_size(record));
203
204         if (likely(record->lur_hdr.lrh_len <= ctxt->loc_chunk_size)) {
205                 OBD_ALLOC_PTR(stc);
206                 if (stc == NULL)
207                         GOTO(llog_put, rc = -ENOMEM);
208                 INIT_LIST_HEAD(&stc->stc_list);
209
210                 rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
211                               &stc->stc_cookie, sub_th->st_sub_th);
212
213                 CDEBUG(D_INFO, "%s: Add update log "DOSTID":%u: rc = %d\n",
214                        dt->dd_lu_dev.ld_obd->obd_name,
215                        POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
216                        stc->stc_cookie.lgc_index, rc);
217
218                 if (rc > 0) {
219                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
220                         rc = 0;
221                 } else {
222                         OBD_FREE_PTR(stc);
223                 }
224
225                 GOTO(llog_put, rc);
226         }
227
228         /* Split the records into chunk_size update record */
229         OBD_ALLOC_LARGE(lur, ctxt->loc_chunk_size);
230         if (lur == NULL)
231                 GOTO(llog_put, rc = -ENOMEM);
232
233         memcpy(lur, &record->lur_hdr, sizeof(record->lur_hdr));
234         lur->lur_update_rec.ur_update_count = 0;
235         lur->lur_update_rec.ur_param_count = 0;
236         src = &record->lur_update_rec.ur_ops;
237         start = next = src;
238         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
239         params = update_records_get_params(&record->lur_update_rec);
240         do {
241                 size_t rec_len;
242
243                 if (update_count < record->lur_update_rec.ur_update_count) {
244                         next = update_op_next_op((struct update_op *)src);
245                 } else {
246                         if (param_count == 0)
247                                 next = update_records_get_params(
248                                                 &record->lur_update_rec);
249                         else
250                                 next = (char *)src +
251                                         object_update_param_size(
252                                         (struct object_update_param *)src);
253                 }
254
255                 rec_len = cfs_size_round((unsigned long)(next - src));
256                 /* If its size > llog chunk_size, then write current chunk to
257                  * the update llog. */
258                 if (lur->lur_hdr.lrh_len + rec_len + LLOG_MIN_REC_SIZE >
259                     ctxt->loc_chunk_size ||
260                     param_count == record->lur_update_rec.ur_param_count) {
261                         lur->lur_update_rec.ur_update_count =
262                                 update_count > last_update_count ?
263                                 update_count - last_update_count : 0;
264                         lur->lur_update_rec.ur_param_count = param_count -
265                                                              last_param_count;
266
267                         memcpy(&lur->lur_update_rec.ur_ops, start,
268                                (unsigned long)(src - start));
269                         if (last_update_count != 0)
270                                 lur->lur_update_rec.ur_flags |=
271                                                 UPDATE_RECORD_CONTINUE;
272
273                         update_records_dump(&lur->lur_update_rec, D_INFO, true);
274                         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
275                         LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
276
277                         OBD_ALLOC_PTR(stc);
278                         if (stc == NULL)
279                                 GOTO(llog_put, rc = -ENOMEM);
280                         INIT_LIST_HEAD(&stc->stc_list);
281
282                         rc = llog_add(env, ctxt->loc_handle,
283                                       &lur->lur_hdr,
284                                       &stc->stc_cookie, sub_th->st_sub_th);
285
286                         CDEBUG(D_INFO, "%s: Add update log "DOSTID":%u"
287                                " rc = %d\n", dt->dd_lu_dev.ld_obd->obd_name,
288                                POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
289                                stc->stc_cookie.lgc_index, rc);
290
291                         if (rc > 0) {
292                                 list_add(&stc->stc_list,
293                                          &sub_th->st_cookie_list);
294                                 rc = 0;
295                         } else {
296                                 OBD_FREE_PTR(stc);
297                                 GOTO(llog_put, rc);
298                         }
299
300                         last_update_count = update_count;
301                         last_param_count = param_count;
302                         start = src;
303                         lur->lur_update_rec.ur_update_count = 0;
304                         lur->lur_update_rec.ur_param_count = 0;
305                         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
306                 }
307
308                 src = next;
309                 lur->lur_hdr.lrh_len += cfs_size_round(rec_len);
310                 if (update_count < record->lur_update_rec.ur_update_count)
311                         update_count++;
312                 else if (param_count < record->lur_update_rec.ur_param_count)
313                         param_count++;
314                 else
315                         break;
316         } while (1);
317
318 llog_put:
319         if (lur != NULL)
320                 OBD_FREE_LARGE(lur, ctxt->loc_chunk_size);
321         llog_ctxt_put(ctxt);
322
323         RETURN(rc);
324 }
325
326 /**
327  * Prepare the update records.
328  *
329  * Merge params and ops into the update records, then initializing
330  * the update buffer.
331  *
332  * During transaction execution phase, parameters and update ops
333  * are collected in two different buffers (see lod_updates_pack()),
334  * during transaction stop, it needs to be merged in one buffer,
335  * so it will be written in the update log.
336  *
337  * \param[in] env       execution environment
338  * \param[in] tmt       top_multiple_thandle for distribute txn
339  *
340  * \retval              0 if merging succeeds.
341  * \retval              negaitive errno if merging fails.
342  */
343 static int prepare_writing_updates(const struct lu_env *env,
344                                    struct top_multiple_thandle *tmt)
345 {
346         struct thandle_update_records   *tur = tmt->tmt_update_records;
347         struct llog_update_record       *lur;
348         struct update_params *params;
349         size_t params_size;
350         size_t update_size;
351
352         if (tur == NULL || tur->tur_update_records == NULL ||
353             tur->tur_update_params == NULL)
354                 return 0;
355
356         lur = tur->tur_update_records;
357         /* Extends the update records buffer if needed */
358         params_size = update_params_size(tur->tur_update_params,
359                                          tur->tur_update_param_count);
360         LASSERT(lur->lur_update_rec.ur_param_count == 0);
361         update_size = llog_update_record_size(lur);
362         if (cfs_size_round(update_size + params_size) >
363             tur->tur_update_records_buf_size) {
364                 int rc;
365
366                 rc = tur_update_records_extend(tur,
367                         cfs_size_round(update_size + params_size));
368                 if (rc < 0)
369                         return rc;
370
371                 lur = tur->tur_update_records;
372         }
373
374         params = update_records_get_params(&lur->lur_update_rec);
375         memcpy(params, tur->tur_update_params, params_size);
376
377         lur->lur_update_rec.ur_param_count = tur->tur_update_param_count;
378         lur->lur_update_rec.ur_batchid = tmt->tmt_batchid;
379         /* Init update record header */
380         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
381         lur->lur_hdr.lrh_type = UPDATE_REC;
382
383         /* Dump updates for debugging purpose */
384         update_records_dump(&lur->lur_update_rec, D_INFO, true);
385
386         return 0;
387 }
388
389 static inline int
390 distribute_txn_commit_thread_running(struct lu_target *lut)
391 {
392         return lut->lut_tdtd_commit_thread.t_flags & SVC_RUNNING;
393 }
394
395 static inline int
396 distribute_txn_commit_thread_stopped(struct lu_target *lut)
397 {
398         return lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED;
399 }
400
401 /**
402  * Top thandle commit callback
403  *
404  * This callback will be called when all of sub transactions are committed.
405  *
406  * \param[in] th        top thandle to be committed.
407  */
408 static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
409 {
410         struct lu_target *lut;
411         ENTRY;
412
413         LASSERT(atomic_read(&tmt->tmt_refcount) > 0);
414
415         top_multiple_thandle_dump(tmt, D_HA);
416         tmt->tmt_committed = 1;
417         lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
418         if (distribute_txn_commit_thread_running(lut))
419                 wake_up(&lut->lut_tdtd->tdtd_commit_thread_waitq);
420         RETURN_EXIT;
421 }
422
423 struct sub_thandle *lookup_sub_thandle(struct top_multiple_thandle *tmt,
424                                        struct dt_device *dt_dev)
425 {
426         struct sub_thandle *st;
427
428         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
429                 if (st->st_dt == dt_dev)
430                         return st;
431         }
432         return NULL;
433 }
434 EXPORT_SYMBOL(lookup_sub_thandle);
435
436 struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
437                                        struct dt_device *dt_dev)
438 {
439         struct sub_thandle *st;
440
441         OBD_ALLOC_PTR(st);
442         if (st == NULL)
443                 RETURN(ERR_PTR(-ENOMEM));
444
445         INIT_LIST_HEAD(&st->st_sub_list);
446         INIT_LIST_HEAD(&st->st_cookie_list);
447         st->st_dt = dt_dev;
448
449         list_add(&st->st_sub_list, &tmt->tmt_sub_thandle_list);
450         return st;
451 }
452
453 /**
454  * sub thandle commit callback
455  *
456  * Mark the sub thandle to be committed and if all sub thandle are committed
457  * notify the top thandle.
458  *
459  * \param[in] env       execution environment
460  * \param[in] sub_th    sub thandle being committed
461  * \param[in] cb        commit callback
462  * \param[in] err       trans result
463  */
464 static void sub_trans_commit_cb(struct lu_env *env,
465                                 struct thandle *sub_th,
466                                 struct dt_txn_commit_cb *cb, int err)
467 {
468         struct sub_thandle      *st;
469         struct top_multiple_thandle *tmt = cb->dcb_data;
470         bool                    all_committed = true;
471         ENTRY;
472
473         /* Check if all sub thandles are committed */
474         spin_lock(&tmt->tmt_sub_lock);
475         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
476                 if (st->st_sub_th == sub_th) {
477                         st->st_committed = 1;
478                         st->st_result = err;
479                 }
480                 if (!st->st_committed)
481                         all_committed = false;
482         }
483         spin_unlock(&tmt->tmt_sub_lock);
484
485         if (tmt->tmt_result == 0)
486                 tmt->tmt_result = err;
487
488         if (all_committed)
489                 top_trans_committed_cb(tmt);
490
491         top_multiple_thandle_dump(tmt, D_INFO);
492         top_multiple_thandle_put(tmt);
493         RETURN_EXIT;
494 }
495
496 static void sub_thandle_register_commit_cb(struct sub_thandle *st,
497                                     struct top_multiple_thandle *tmt)
498 {
499         LASSERT(st->st_sub_th != NULL);
500         top_multiple_thandle_get(tmt);
501         st->st_commit_dcb.dcb_func = sub_trans_commit_cb;
502         st->st_commit_dcb.dcb_data = tmt;
503         INIT_LIST_HEAD(&st->st_commit_dcb.dcb_linkage);
504         dt_trans_cb_add(st->st_sub_th, &st->st_commit_dcb);
505 }
506
507 /**
508  * Sub thandle stop call back
509  *
510  * After sub thandle is stopped, it will call this callback to notify
511  * the top thandle.
512  *
513  * \param[in] th        sub thandle to be stopped
514  * \param[in] rc        result of sub trans
515  */
516 static void sub_trans_stop_cb(struct lu_env *env,
517                               struct thandle *sub_th,
518                               struct dt_txn_commit_cb *cb, int err)
519 {
520         struct sub_thandle              *st;
521         struct top_multiple_thandle     *tmt = cb->dcb_data;
522         ENTRY;
523
524         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
525                 if (st->st_stopped)
526                         continue;
527
528                 if (st->st_dt == sub_th->th_dev) {
529                         st->st_stopped = 1;
530                         st->st_result = err;
531                         break;
532                 }
533         }
534
535         wake_up(&tmt->tmt_stop_waitq);
536         RETURN_EXIT;
537 }
538
539 static void sub_thandle_register_stop_cb(struct sub_thandle *st,
540                                          struct top_multiple_thandle *tmt)
541 {
542         st->st_stop_dcb.dcb_func = sub_trans_stop_cb;
543         st->st_stop_dcb.dcb_data = tmt;
544         st->st_stop_dcb.dcb_flags = DCB_TRANS_STOP;
545         INIT_LIST_HEAD(&st->st_stop_dcb.dcb_linkage);
546         dt_trans_cb_add(st->st_sub_th, &st->st_stop_dcb);
547 }
548
549 /**
550  * Create sub thandle
551  *
552  * Create transaction handle for sub_thandle
553  *
554  * \param[in] env       execution environment
555  * \param[in] th        top thandle
556  * \param[in] st        sub_thandle
557  *
558  * \retval              0 if creation succeeds.
559  * \retval              negative errno if creation fails.
560  */
561 int sub_thandle_trans_create(const struct lu_env *env,
562                              struct top_thandle *top_th,
563                              struct sub_thandle *st)
564 {
565         struct thandle *sub_th;
566
567         sub_th = dt_trans_create(env, st->st_dt);
568         if (IS_ERR(sub_th))
569                 return PTR_ERR(sub_th);
570
571         sub_th->th_top = &top_th->tt_super;
572         st->st_sub_th = sub_th;
573
574         sub_th->th_wait_submit = 1;
575         return 0;
576 }
577
578 /**
579  * Create the top transaction.
580  *
581  * Create the top transaction on the master device. It will create a top
582  * thandle and a sub thandle on the master device.
583  *
584  * \param[in] env               execution environment
585  * \param[in] master_dev        master_dev the top thandle will be created
586  *
587  * \retval                      pointer to the created thandle.
588  * \retval                      ERR_PTR(errno) if creation failed.
589  */
590 struct thandle *
591 top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
592 {
593         struct top_thandle      *top_th;
594         struct thandle          *child_th;
595
596         OBD_ALLOC_GFP(top_th, sizeof(*top_th), __GFP_IO);
597         if (top_th == NULL)
598                 return ERR_PTR(-ENOMEM);
599
600         top_th->tt_super.th_top = &top_th->tt_super;
601
602         if (master_dev != NULL) {
603                 child_th = dt_trans_create(env, master_dev);
604                 if (IS_ERR(child_th)) {
605                         OBD_FREE_PTR(top_th);
606                         return child_th;
607                 }
608
609                 child_th->th_top = &top_th->tt_super;
610                 child_th->th_wait_submit = 1;
611                 top_th->tt_master_sub_thandle = child_th;
612
613                 top_th->tt_super.th_tags |= child_th->th_tags;
614         }
615         return &top_th->tt_super;
616 }
617 EXPORT_SYMBOL(top_trans_create);
618
619 /**
620  * Declare write update transaction
621  *
622  * Check if there are updates being recorded in this transaction,
623  * it will write the record into the disk.
624  *
625  * \param[in] env       execution environment
626  * \param[in] tmt       top multiple transaction handle
627  *
628  * \retval              0 if writing succeeds
629  * \retval              negative errno if writing fails
630  */
631 static int declare_updates_write(const struct lu_env *env,
632                                  struct top_multiple_thandle *tmt)
633 {
634         struct llog_update_record *record;
635         struct sub_thandle *st;
636         int rc = 0;
637
638         record = tmt->tmt_update_records->tur_update_records;
639         /* Declare update write for all other target */
640         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
641                 if (st->st_sub_th == NULL)
642                         continue;
643
644                 rc = sub_declare_updates_write(env, record, st->st_sub_th,
645                                                tmt->tmt_record_size);
646                 if (rc < 0)
647                         break;
648         }
649
650         return rc;
651 }
652
653 /**
654  * Assign batchid to the distribute transaction.
655  *
656  * Assign batchid to the distribute transaction
657  *
658  * \param[in] tmt       distribute transaction
659  */
660 static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
661 {
662         struct target_distribute_txn_data *tdtd;
663         struct dt_device *dt = new->tmt_master_sub_dt;
664
665         LASSERT(dt != NULL);
666         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
667         spin_lock(&tdtd->tdtd_batchid_lock);
668         new->tmt_batchid = tdtd->tdtd_batchid++;
669         list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
670         spin_unlock(&tdtd->tdtd_batchid_lock);
671         top_multiple_thandle_get(new);
672         top_multiple_thandle_dump(new, D_INFO);
673 }
674
675 /**
676  * Insert distribute transaction to the distribute txn list.
677  *
678  * Insert distribute transaction to the distribute txn list.
679  *
680  * \param[in] new       the distribute txn to be inserted.
681  */
682 void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
683 {
684         struct dt_device *dt = new->tmt_master_sub_dt;
685         struct top_multiple_thandle *tmt;
686         struct target_distribute_txn_data *tdtd;
687         bool    at_head = false;
688
689         LASSERT(dt != NULL);
690         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
691
692         spin_lock(&tdtd->tdtd_batchid_lock);
693         list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
694                 if (new->tmt_batchid > tmt->tmt_batchid) {
695                         list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
696                         break;
697                 }
698         }
699         if (list_empty(&new->tmt_commit_list)) {
700                 at_head = true;
701                 list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
702         }
703         spin_unlock(&tdtd->tdtd_batchid_lock);
704         top_multiple_thandle_get(new);
705         top_multiple_thandle_dump(new, D_INFO);
706         if (new->tmt_committed && at_head)
707                 wake_up(&tdtd->tdtd_commit_thread_waitq);
708 }
709
710 /**
711  * Prepare cross-MDT operation.
712  *
713  * Create the update record buffer to record updates for cross-MDT operation,
714  * add master sub transaction to tt_sub_trans_list, and declare the update
715  * writes.
716  *
717  * During updates packing, all of parameters will be packed in
718  * tur_update_params, and updates will be packed in tur_update_records.
719  * Then in transaction stop, parameters and updates will be merged
720  * into one updates buffer.
721  *
722  * And also master thandle will be added to the sub_th list, so it will be
723  * easy to track the commit status.
724  *
725  * \param[in] env       execution environment
726  * \param[in] th        top transaction handle
727  *
728  * \retval              0 if preparation succeeds.
729  * \retval              negative errno if preparation fails.
730  */
731 static int prepare_multiple_node_trans(const struct lu_env *env,
732                                        struct top_multiple_thandle *tmt)
733 {
734         struct thandle_update_records   *tur;
735         int                             rc;
736         ENTRY;
737
738         if (tmt->tmt_update_records == NULL) {
739                 tur = &update_env_info(env)->uti_tur;
740                 rc = check_and_prepare_update_record(env, tur);
741                 if (rc < 0)
742                         RETURN(rc);
743
744                 tmt->tmt_update_records = tur;
745                 distribute_txn_assign_batchid(tmt);
746         }
747
748         rc = declare_updates_write(env, tmt);
749
750         RETURN(rc);
751 }
752
753 /**
754  * start the top transaction.
755  *
756  * Start all of its sub transactions, then start master sub transaction.
757  *
758  * \param[in] env               execution environment
759  * \param[in] master_dev        master_dev the top thandle will be start
760  * \param[in] th                top thandle
761  *
762  * \retval                      0 if transaction start succeeds.
763  * \retval                      negative errno if start fails.
764  */
765 int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
766                     struct thandle *th)
767 {
768         struct top_thandle      *top_th = container_of(th, struct top_thandle,
769                                                        tt_super);
770         struct sub_thandle              *st;
771         struct top_multiple_thandle     *tmt = top_th->tt_multiple_thandle;
772         int                             rc = 0;
773         ENTRY;
774
775         if (tmt == NULL) {
776                 rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
777                                     top_th->tt_master_sub_thandle);
778                 RETURN(rc);
779         }
780
781         tmt = top_th->tt_multiple_thandle;
782         rc = prepare_multiple_node_trans(env, tmt);
783         if (rc < 0)
784                 RETURN(rc);
785
786         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
787                 if (st->st_sub_th == NULL)
788                         continue;
789                 if (th->th_sync)
790                         st->st_sub_th->th_sync = th->th_sync;
791                 st->st_sub_th->th_local = th->th_local;
792                 st->st_sub_th->th_tags = th->th_tags;
793                 rc = dt_trans_start(env, st->st_sub_th->th_dev,
794                                     st->st_sub_th);
795                 if (rc != 0)
796                         GOTO(out, rc);
797
798                 sub_thandle_register_stop_cb(st, tmt);
799                 sub_thandle_register_commit_cb(st, tmt);
800         }
801 out:
802         th->th_result = rc;
803         RETURN(rc);
804 }
805 EXPORT_SYMBOL(top_trans_start);
806
807 /**
808  * Check whether we need write updates record
809  *
810  * Check if the updates for the top_thandle needs to be writen
811  * to all targets. Only if the transaction succeeds and the updates
812  * number > 2, it will write the updates,
813  *
814  * \params [in] top_th  top thandle.
815  *
816  * \retval              true if it needs to write updates
817  * \retval              false if it does not need to write updates
818  **/
819 static bool top_check_write_updates(struct top_thandle *top_th)
820 {
821         struct top_multiple_thandle     *tmt;
822         struct thandle_update_records   *tur;
823
824         /* Do not write updates to records if the transaction fails */
825         if (top_th->tt_super.th_result != 0)
826                 return false;
827
828         tmt = top_th->tt_multiple_thandle;
829         if (tmt == NULL)
830                 return false;
831
832         tur = tmt->tmt_update_records;
833         if (tur == NULL)
834                 return false;
835
836         /* Hmm, false update records, since the cross-MDT operation
837          * should includes both local and remote updates, so the
838          * updates count should >= 2 */
839         if (tur->tur_update_records == NULL ||
840             tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
841                 return false;
842
843         return true;
844 }
845
846 /**
847  * Check if top transaction is stopped
848  *
849  * Check if top transaction is stopped, only if all sub transaction
850  * is stopped, then the top transaction is stopped.
851  *
852  * \param [in] top_th   top thandle
853  *
854  * \retval              true if the top transaction is stopped.
855  * \retval              false if the top transaction is not stopped.
856  */
857 static bool top_trans_is_stopped(struct top_thandle *top_th)
858 {
859         struct top_multiple_thandle     *tmt;
860         struct sub_thandle              *st;
861         bool                    all_stopped = true;
862
863         tmt = top_th->tt_multiple_thandle;
864         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
865                 if (!st->st_stopped && st->st_sub_th != NULL) {
866                         all_stopped = false;
867                         break;
868                 }
869
870                 if (st->st_result != 0 &&
871                     top_th->tt_super.th_result == 0)
872                         top_th->tt_super.th_result = st->st_result;
873         }
874
875         return all_stopped;
876 }
877
878 /**
879  * Wait result of top transaction
880  *
881  * Wait until all sub transaction get its result.
882  *
883  * \param [in] top_th   top thandle.
884  *
885  * \retval              the result of top thandle.
886  */
887 static int top_trans_wait_result(struct top_thandle *top_th)
888 {
889         struct l_wait_info      lwi = {0};
890
891         l_wait_event(top_th->tt_multiple_thandle->tmt_stop_waitq,
892                      top_trans_is_stopped(top_th), &lwi);
893
894         RETURN(top_th->tt_super.th_result);
895 }
896
897 /**
898  * Stop the top transaction.
899  *
900  * Stop the transaction on the master device first, then stop transactions
901  * on other sub devices.
902  *
903  * \param[in] env               execution environment
904  * \param[in] master_dev        master_dev the top thandle will be created
905  * \param[in] th                top thandle
906  *
907  * \retval                      0 if stop transaction succeeds.
908  * \retval                      negative errno if stop transaction fails.
909  */
910 int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
911                    struct thandle *th)
912 {
913         struct top_thandle      *top_th = container_of(th, struct top_thandle,
914                                                        tt_super);
915         struct sub_thandle              *st;
916         struct sub_thandle              *master_st;
917         struct top_multiple_thandle     *tmt;
918         struct thandle_update_records   *tur;
919         bool                            write_updates = false;
920         int                     rc = 0;
921         ENTRY;
922
923         if (likely(top_th->tt_multiple_thandle == NULL)) {
924                 LASSERT(master_dev != NULL);
925                 rc = dt_trans_stop(env, master_dev,
926                                    top_th->tt_master_sub_thandle);
927                 OBD_FREE_PTR(top_th);
928                 RETURN(rc);
929         }
930
931         tmt = top_th->tt_multiple_thandle;
932         tur = tmt->tmt_update_records;
933
934         /* Note: we need stop the master thandle first, then the stop
935          * callback will fill the master transno in the update logs,
936          * then these update logs will be sent to other MDTs */
937         /* get the master sub thandle */
938         master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
939         write_updates = top_check_write_updates(top_th);
940
941         /* Step 1: write the updates log on Master MDT */
942         if (master_st != NULL && master_st->st_sub_th != NULL &&
943             write_updates) {
944                 struct llog_update_record *lur;
945
946                 /* Merge the parameters and updates into one buffer */
947                 rc = prepare_writing_updates(env, tmt);
948                 if (rc < 0) {
949                         CERROR("%s: cannot prepare updates: rc = %d\n",
950                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
951                         th->th_result = rc;
952                         GOTO(stop_master_trans, rc);
953                 }
954
955                 lur = tur->tur_update_records;
956                 /* Write updates to the master MDT */
957                 rc = sub_updates_write(env, lur, master_st);
958
959                 /* Cleanup the common parameters in the update records,
960                  * master transno callback might add more parameters.
961                  * and we need merge the update records again in the
962                  * following */
963                 if (tur->tur_update_params != NULL)
964                         lur->lur_update_rec.ur_param_count = 0;
965
966                 if (rc < 0) {
967                         CERROR("%s: write updates failed: rc = %d\n",
968                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
969                         th->th_result = rc;
970                         GOTO(stop_master_trans, rc);
971                 }
972         }
973
974 stop_master_trans:
975         /* Step 2: Stop the transaction on the master MDT, and fill the
976          * master transno in the update logs to other MDT. */
977         if (master_st != NULL && master_st->st_sub_th != NULL) {
978                 master_st->st_sub_th->th_local = th->th_local;
979                 if (th->th_sync)
980                         master_st->st_sub_th->th_sync = th->th_sync;
981                 master_st->st_sub_th->th_tags = th->th_tags;
982                 master_st->st_sub_th->th_result = th->th_result;
983                 rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
984                 if (rc < 0) {
985                         th->th_result = rc;
986                         GOTO(stop_other_trans, rc);
987                 } else if (tur != NULL && tur->tur_update_records != NULL) {
988                         struct llog_update_record *lur;
989
990                         lur = tur->tur_update_records;
991                         if (lur->lur_update_rec.ur_master_transno == 0)
992                                 /* Update master transno after master stop
993                                  * callback */
994                                 lur->lur_update_rec.ur_master_transno =
995                                                 tgt_th_info(env)->tti_transno;
996                 }
997         }
998
999         /* Step 3: write updates to other MDTs */
1000         if (write_updates) {
1001                 struct llog_update_record *lur;
1002
1003                 /* Stop callback of master will add more updates and also update
1004                  * master transno, so merge the parameters and updates into one
1005                  * buffer again */
1006                 rc = prepare_writing_updates(env, tmt);
1007                 if (rc < 0) {
1008                         CERROR("%s: prepare updates failed: rc = %d\n",
1009                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
1010                         th->th_result = rc;
1011                         GOTO(stop_other_trans, rc);
1012                 }
1013                 lur = tur->tur_update_records;
1014                 list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
1015                                     st_sub_list) {
1016                         if (st->st_sub_th == NULL || st == master_st ||
1017                             st->st_sub_th->th_result < 0)
1018                                 continue;
1019
1020                         rc = sub_updates_write(env, lur, st);
1021                         if (rc < 0) {
1022                                 th->th_result = rc;
1023                                 break;
1024                         }
1025                 }
1026         }
1027
1028 stop_other_trans:
1029         /* Step 4: Stop the transaction on other MDTs */
1030         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1031                 if (st == master_st || st->st_sub_th == NULL)
1032                         continue;
1033
1034                 if (th->th_sync)
1035                         st->st_sub_th->th_sync = th->th_sync;
1036                 st->st_sub_th->th_local = th->th_local;
1037                 st->st_sub_th->th_tags = th->th_tags;
1038                 st->st_sub_th->th_result = th->th_result;
1039                 rc = dt_trans_stop(env, st->st_sub_th->th_dev,
1040                                    st->st_sub_th);
1041                 if (unlikely(rc < 0 && th->th_result == 0))
1042                         th->th_result = rc;
1043         }
1044
1045         rc = top_trans_wait_result(top_th);
1046
1047         tmt->tmt_result = rc;
1048
1049         /* Balance for the refcount in top_trans_create, Note: if it is NOT
1050          * multiple node transaction, the top transaction will be destroyed. */
1051         top_multiple_thandle_put(tmt);
1052         OBD_FREE_PTR(top_th);
1053         RETURN(rc);
1054 }
1055 EXPORT_SYMBOL(top_trans_stop);
1056
1057 /**
1058  * Create top_multiple_thandle for top_thandle
1059  *
1060  * Create top_mutilple_thandle to manage the mutiple node transaction
1061  * for top_thandle, and it also needs to add master sub thandle to the
1062  * sub trans list now.
1063  *
1064  * \param[in] env       execution environment
1065  * \param[in] top_th    the top thandle
1066  *
1067  * \retval      0 if creation succeeds
1068  * \retval      negative errno if creation fails
1069  */
1070 int top_trans_create_tmt(const struct lu_env *env,
1071                          struct top_thandle *top_th)
1072 {
1073         struct top_multiple_thandle *tmt;
1074
1075         OBD_ALLOC_PTR(tmt);
1076         if (tmt == NULL)
1077                 return -ENOMEM;
1078
1079         tmt->tmt_magic = TOP_THANDLE_MAGIC;
1080         INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
1081         INIT_LIST_HEAD(&tmt->tmt_commit_list);
1082         atomic_set(&tmt->tmt_refcount, 1);
1083         spin_lock_init(&tmt->tmt_sub_lock);
1084         init_waitqueue_head(&tmt->tmt_stop_waitq);
1085
1086         top_th->tt_multiple_thandle = tmt;
1087
1088         return 0;
1089 }
1090
1091 static struct sub_thandle *
1092 create_sub_thandle_with_thandle(struct top_thandle *top_th,
1093                                 struct thandle *sub_th)
1094 {
1095         struct sub_thandle *st;
1096
1097         /* create and init sub th to the top trans list */
1098         st = create_sub_thandle(top_th->tt_multiple_thandle,
1099                                 sub_th->th_dev);
1100         if (IS_ERR(st))
1101                 return st;
1102
1103         st->st_sub_th = sub_th;
1104
1105         sub_th->th_top = &top_th->tt_super;
1106         return st;
1107 }
1108
1109 /**
1110  * Get sub thandle.
1111  *
1112  * Get sub thandle from the top thandle according to the sub dt_device.
1113  *
1114  * \param[in] env       execution environment
1115  * \param[in] th        thandle on the top layer.
1116  * \param[in] sub_dt    sub dt_device used to get sub transaction
1117  *
1118  * \retval              thandle of sub transaction if succeed
1119  * \retval              PTR_ERR(errno) if failed
1120  */
1121 struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
1122                                       struct thandle *th,
1123                                       struct dt_device *sub_dt)
1124 {
1125         struct sub_thandle      *st = NULL;
1126         struct top_thandle      *top_th;
1127         struct thandle          *sub_th = NULL;
1128         int                     rc = 0;
1129         ENTRY;
1130
1131         top_th = container_of(th, struct top_thandle, tt_super);
1132
1133         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
1134                 RETURN(top_th->tt_master_sub_thandle);
1135
1136         if (top_th->tt_multiple_thandle != NULL) {
1137                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
1138                 if (st != NULL)
1139                         RETURN(st->st_sub_th);
1140         }
1141
1142         sub_th = dt_trans_create(env, sub_dt);
1143         if (IS_ERR(sub_th))
1144                 RETURN(sub_th);
1145
1146         /* Create top_multiple_thandle if necessary */
1147         if (top_th->tt_multiple_thandle == NULL) {
1148                 struct top_multiple_thandle *tmt;
1149
1150                 rc = top_trans_create_tmt(env, top_th);
1151                 if (rc < 0)
1152                         GOTO(stop_trans, rc);
1153
1154                 tmt = top_th->tt_multiple_thandle;
1155
1156                 /* Add master sub th to the top trans list */
1157                 tmt->tmt_master_sub_dt =
1158                         top_th->tt_master_sub_thandle->th_dev;
1159                 st = create_sub_thandle_with_thandle(top_th,
1160                                 top_th->tt_master_sub_thandle);
1161                 if (IS_ERR(st))
1162                         GOTO(stop_trans, rc = PTR_ERR(st));
1163         }
1164
1165         /* create and init sub th to the top trans list */
1166         st = create_sub_thandle_with_thandle(top_th, sub_th);
1167         st->st_sub_th->th_wait_submit = 1;
1168 stop_trans:
1169         if (rc < 0) {
1170                 if (st != NULL)
1171                         OBD_FREE_PTR(st);
1172                 sub_th->th_result = rc;
1173                 dt_trans_stop(env, sub_dt, sub_th);
1174                 sub_th = ERR_PTR(rc);
1175         }
1176
1177         RETURN(sub_th);
1178 }
1179 EXPORT_SYMBOL(thandle_get_sub_by_dt);
1180
1181 /**
1182  * Top multiple thandle destroy
1183  *
1184  * Destroy multiple thandle and all its sub thandle.
1185  *
1186  * \param[in] tmt       top_multiple_thandle to be destroyed.
1187  */
1188 void top_multiple_thandle_destroy(struct top_multiple_thandle *tmt)
1189 {
1190         struct sub_thandle *st;
1191         struct sub_thandle *tmp;
1192
1193         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
1194         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
1195                                  st_sub_list) {
1196                 struct sub_thandle_cookie *stc;
1197                 struct sub_thandle_cookie *tmp;
1198
1199                 list_del(&st->st_sub_list);
1200                 list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
1201                                          stc_list) {
1202                         list_del(&stc->stc_list);
1203                         OBD_FREE_PTR(stc);
1204                 }
1205                 OBD_FREE_PTR(st);
1206         }
1207         OBD_FREE_PTR(tmt);
1208 }
1209 EXPORT_SYMBOL(top_multiple_thandle_destroy);
1210
1211 /**
1212  * Cancel the update log on MDTs
1213  *
1214  * Cancel the update log on MDTs then destroy the thandle.
1215  *
1216  * \param[in] env       execution environment
1217  * \param[in] tmt       the top multiple thandle whose updates records
1218  *                      will be cancelled.
1219  *
1220  * \retval              0 if cancellation succeeds.
1221  * \retval              negative errno if cancellation fails.
1222  */
1223 static int distribute_txn_cancel_records(const struct lu_env *env,
1224                                          struct top_multiple_thandle *tmt)
1225 {
1226         struct sub_thandle *st;
1227         ENTRY;
1228
1229         top_multiple_thandle_dump(tmt, D_INFO);
1230         /* Cancel update logs on other MDTs */
1231         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1232                 struct llog_ctxt        *ctxt;
1233                 struct obd_device       *obd;
1234                 struct llog_cookie      *cookie;
1235                 struct sub_thandle_cookie *stc;
1236                 int rc;
1237
1238                 obd = st->st_dt->dd_lu_dev.ld_obd;
1239                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
1240                 if (ctxt == NULL)
1241                         continue;
1242                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
1243                         cookie = &stc->stc_cookie;
1244                         if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
1245                                 continue;
1246
1247                         rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
1248                                                      cookie);
1249                         CDEBUG(D_HA, "%s: batchid %llu cancel update log "
1250                                DOSTID ".%u : rc = %d\n", obd->obd_name,
1251                                tmt->tmt_batchid,
1252                                POSTID(&cookie->lgc_lgl.lgl_oi),
1253                                cookie->lgc_index, rc);
1254                 }
1255
1256                 llog_ctxt_put(ctxt);
1257         }
1258
1259         RETURN(0);
1260 }
1261
1262 /**
1263  * Check if there are committed transaction
1264  *
1265  * Check if there are committed transaction in the distribute transaction
1266  * list, then cancel the update records for those committed transaction.
1267  * Because the distribute transaction in the list are sorted by batchid,
1268  * and cancellation will be done by batchid order, so we only check the first
1269  * the transaction(with lowest batchid) in the list.
1270  *
1271  * \param[in] lod       lod device where cancel thread is
1272  *
1273  * \retval              true if it is ready
1274  * \retval              false if it is not ready
1275  */
1276 static bool tdtd_ready_for_cancel_log(struct target_distribute_txn_data *tdtd)
1277 {
1278         struct top_multiple_thandle     *tmt = NULL;
1279         struct obd_device               *obd = tdtd->tdtd_lut->lut_obd;
1280         bool    ready = false;
1281
1282         spin_lock(&tdtd->tdtd_batchid_lock);
1283         if (!list_empty(&tdtd->tdtd_list)) {
1284                 tmt = list_entry(tdtd->tdtd_list.next,
1285                                  struct top_multiple_thandle, tmt_commit_list);
1286                 if (tmt->tmt_committed &&
1287                     (!obd->obd_recovering || (obd->obd_recovering &&
1288                     tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)))
1289                         ready = true;
1290         }
1291         spin_unlock(&tdtd->tdtd_batchid_lock);
1292
1293         return ready;
1294 }
1295
1296 struct distribute_txn_bid_data {
1297         struct dt_txn_commit_cb  dtbd_cb;
1298         struct target_distribute_txn_data      *dtbd_tdtd;
1299         __u64                    dtbd_batchid;
1300 };
1301
1302 /**
1303  * callback of updating commit batchid
1304  *
1305  * Updating commit batchid then wake up the commit thread to cancel the
1306  * records.
1307  *
1308  * \param[in]env        execution environment
1309  * \param[in]th         thandle to updating commit batchid
1310  * \param[in]cb         commit callback
1311  * \param[in]err        result of thandle
1312  */
1313 static void distribute_txn_batchid_cb(struct lu_env *env,
1314                                       struct thandle *th,
1315                                       struct dt_txn_commit_cb *cb,
1316                                       int err)
1317 {
1318         struct distribute_txn_bid_data          *dtbd = NULL;
1319         struct target_distribute_txn_data       *tdtd;
1320
1321         dtbd = container_of0(cb, struct distribute_txn_bid_data, dtbd_cb);
1322         tdtd = dtbd->dtbd_tdtd;
1323
1324         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1325               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1326         spin_lock(&tdtd->tdtd_batchid_lock);
1327         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1328             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1329                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1330         spin_unlock(&tdtd->tdtd_batchid_lock);
1331         atomic_dec(&tdtd->tdtd_refcount);
1332         wake_up(&tdtd->tdtd_commit_thread_waitq);
1333
1334         OBD_FREE_PTR(dtbd);
1335 }
1336
1337 /**
1338  * Update the commit batchid in disk
1339  *
1340  * Update commit batchid in the disk, after this is committed, it can start
1341  * to cancel the update records.
1342  *
1343  * \param[in] env       execution environment
1344  * \param[in] tdtd      distribute transaction structure
1345  * \param[in] batchid   commit batchid to be updated
1346  *
1347  * \retval              0 if update succeeds.
1348  * \retval              negative errno if update fails.
1349  */
1350 static int
1351 distribute_txn_commit_batchid_update(const struct lu_env *env,
1352                               struct target_distribute_txn_data *tdtd,
1353                               __u64 batchid)
1354 {
1355         struct distribute_txn_bid_data  *dtbd = NULL;
1356         struct thandle          *th;
1357         struct lu_buf            buf;
1358         __u64                    tmp;
1359         __u64                    off;
1360         int                      rc;
1361         ENTRY;
1362
1363         OBD_ALLOC_PTR(dtbd);
1364         if (dtbd == NULL)
1365                 RETURN(-ENOMEM);
1366         dtbd->dtbd_batchid = batchid;
1367         dtbd->dtbd_tdtd = tdtd;
1368         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1369         atomic_inc(&tdtd->tdtd_refcount);
1370
1371         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1372         if (IS_ERR(th)) {
1373                 OBD_FREE_PTR(dtbd);
1374                 RETURN(PTR_ERR(th));
1375         }
1376
1377         tmp = cpu_to_le64(batchid);
1378         buf.lb_buf = &tmp;
1379         buf.lb_len = sizeof(tmp);
1380         off = 0;
1381
1382         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1383                                      th);
1384         if (rc < 0)
1385                 GOTO(stop, rc);
1386
1387         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1388         if (rc < 0)
1389                 GOTO(stop, rc);
1390
1391         rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
1392         if (rc < 0)
1393                 GOTO(stop, rc);
1394
1395         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1396                              &off, th);
1397
1398         CDEBUG(D_INFO, "%s: update batchid "LPU64": rc = %d\n",
1399                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1400
1401 stop:
1402         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1403         if (rc < 0)
1404                 OBD_FREE_PTR(dtbd);
1405         RETURN(rc);
1406 }
1407
1408 /**
1409  * Init commit batchid for distribute transaction.
1410  *
1411  * Initialize the batchid object and get commit batchid from the object.
1412  *
1413  * \param[in] env       execution environment
1414  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1415  *
1416  * \retval              0 if initialization succeeds.
1417  * \retval              negative errno if initialization fails.
1418  **/
1419 static int
1420 distribute_txn_commit_batchid_init(const struct lu_env *env,
1421                                    struct target_distribute_txn_data *tdtd)
1422 {
1423         struct tgt_thread_info  *tti = tgt_th_info(env);
1424         struct lu_target        *lut = tdtd->tdtd_lut;
1425         struct lu_attr          *attr = &tti->tti_attr;
1426         struct lu_fid           *fid = &tti->tti_fid1;
1427         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1428         struct dt_object        *dt_obj = NULL;
1429         struct lu_buf           buf;
1430         __u64                   tmp;
1431         __u64                   off;
1432         int                     rc;
1433         ENTRY;
1434
1435         memset(attr, 0, sizeof(*attr));
1436         attr->la_valid = LA_MODE;
1437         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1438         dof->dof_type = dt_mode_to_dft(S_IFREG);
1439
1440         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1441
1442         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1443                                    attr);
1444         if (IS_ERR(dt_obj)) {
1445                 rc = PTR_ERR(dt_obj);
1446                 dt_obj = NULL;
1447                 GOTO(out_put, rc);
1448         }
1449
1450         tdtd->tdtd_batchid_obj = dt_obj;
1451
1452         buf.lb_buf = &tmp;
1453         buf.lb_len = sizeof(tmp);
1454         off = 0;
1455         rc = dt_read(env, dt_obj, &buf, &off);
1456         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1457                 CERROR("%s can't read last committed batchid: rc = %d\n",
1458                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1459                 if (rc > 0)
1460                         rc = -EINVAL;
1461                 GOTO(out_put, rc);
1462         } else if (rc == buf.lb_len) {
1463                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1464                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1465                        tdtd->tdtd_lut->lut_obd->obd_name,
1466                        tdtd->tdtd_committed_batchid);
1467                 rc = 0;
1468         }
1469
1470 out_put:
1471         if (rc < 0 && dt_obj != NULL) {
1472                 lu_object_put(env, &dt_obj->do_lu);
1473                 tdtd->tdtd_batchid_obj = NULL;
1474         }
1475         return rc;
1476 }
1477
1478 /**
1479  * manage the distribute transaction thread
1480  *
1481  * Distribute transaction are linked to the list, and once the distribute
1482  * transaction is committed, it will update the last committed batchid first,
1483  * after it is committed, it will cancel the records.
1484  *
1485  * \param[in] _arg      argument for commit thread
1486  *
1487  * \retval              0 if thread is running successfully
1488  * \retval              negative errno if the thread can not be run.
1489  */
1490 static int distribute_txn_commit_thread(void *_arg)
1491 {
1492         struct target_distribute_txn_data *tdtd = _arg;
1493         struct lu_target        *lut = tdtd->tdtd_lut;
1494         struct ptlrpc_thread    *thread = &lut->lut_tdtd_commit_thread;
1495         struct l_wait_info       lwi = { 0 };
1496         struct lu_env            env;
1497         struct list_head         list;
1498         int                      rc;
1499         struct top_multiple_thandle *tmt;
1500         struct top_multiple_thandle *tmp;
1501         __u64                    batchid = 0, committed;
1502
1503         ENTRY;
1504
1505         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
1506         if (rc != 0)
1507                 RETURN(rc);
1508
1509         spin_lock(&tdtd->tdtd_batchid_lock);
1510         thread->t_flags = SVC_RUNNING;
1511         spin_unlock(&tdtd->tdtd_batchid_lock);
1512         wake_up(&thread->t_ctl_waitq);
1513         INIT_LIST_HEAD(&list);
1514
1515         CDEBUG(D_HA, "%s: start commit thread committed batchid "LPU64"\n",
1516                tdtd->tdtd_lut->lut_obd->obd_name,
1517                tdtd->tdtd_committed_batchid);
1518
1519         while (distribute_txn_commit_thread_running(lut)) {
1520                 spin_lock(&tdtd->tdtd_batchid_lock);
1521                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1522                                          tmt_commit_list) {
1523                         if (tmt->tmt_committed == 0)
1524                                 break;
1525
1526                         /* Note: right now, replay is based on master MDT
1527                          * transno, but cancellation is based on batchid.
1528                          * so we do not try to cancel the update log until
1529                          * the recoverying is done, unless the update records
1530                          * batchid < committed_batchid. */
1531                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1532                                 list_move_tail(&tmt->tmt_commit_list, &list);
1533                         } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
1534                                 LASSERTF(tmt->tmt_batchid >= batchid,
1535                                          "tmt %p tmt_batchid: "LPU64", batchid "
1536                                           LPU64"\n", tmt, tmt->tmt_batchid,
1537                                          batchid);
1538                                 /* There are three types of distribution
1539                                  * transaction result
1540                                  *
1541                                  * 1. If tmt_result < 0, it means the
1542                                  * distribution transaction fails, which should
1543                                  * be rare, because once declare phase succeeds,
1544                                  * the operation should succeeds anyway. Note in
1545                                  * this case, we will still update batchid so
1546                                  * cancellation would be stopped.
1547                                  *
1548                                  * 2. If tmt_result == 0, it means the
1549                                  * distribution transaction succeeds, and we
1550                                  * will update batchid.
1551                                  *
1552                                  * 3. If tmt_result > 0, it means distribute
1553                                  * transaction is not yet committed on every
1554                                  * node, but we need release this tmt before
1555                                  * that, which usuually happens during umount.
1556                                  */
1557                                 if (tmt->tmt_result <= 0)
1558                                         batchid = tmt->tmt_batchid;
1559                                 list_move_tail(&tmt->tmt_commit_list, &list);
1560                         }
1561                 }
1562                 spin_unlock(&tdtd->tdtd_batchid_lock);
1563
1564                 CDEBUG(D_HA, "%s: batchid: "LPU64" committed batchid "
1565                        LPU64"\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1566                        tdtd->tdtd_committed_batchid);
1567                 /* update globally committed on a storage */
1568                 if (batchid > tdtd->tdtd_committed_batchid) {
1569                         distribute_txn_commit_batchid_update(&env, tdtd,
1570                                                              batchid);
1571                         spin_lock(&tdtd->tdtd_batchid_lock);
1572                         if (batchid > tdtd->tdtd_batchid) {
1573                                 /* This might happen during recovery,
1574                                  * batchid is initialized as last transno,
1575                                  * and the batchid in the update records
1576                                  * on other MDTs might be bigger than
1577                                  * the batchid, so we need update it to
1578                                  * avoid duplicate batchid. */
1579                                 CDEBUG(D_HA, "%s update batchid from "LPU64
1580                                        " to "LPU64"\n",
1581                                        tdtd->tdtd_lut->lut_obd->obd_name,
1582                                        tdtd->tdtd_batchid, batchid);
1583                                 tdtd->tdtd_batchid = batchid;
1584                         }
1585                         spin_unlock(&tdtd->tdtd_batchid_lock);
1586                 }
1587                 /* cancel the records for committed batchid's */
1588                 /* XXX: should we postpone cancel's till the end of recovery? */
1589                 committed = tdtd->tdtd_committed_batchid;
1590                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1591                         if (tmt->tmt_batchid > committed)
1592                                 break;
1593                         list_del_init(&tmt->tmt_commit_list);
1594                         if (tmt->tmt_result <= 0)
1595                                 distribute_txn_cancel_records(&env, tmt);
1596                         top_multiple_thandle_put(tmt);
1597                 }
1598
1599                 l_wait_event(tdtd->tdtd_commit_thread_waitq,
1600                              !distribute_txn_commit_thread_running(lut) ||
1601                              committed < tdtd->tdtd_committed_batchid ||
1602                              tdtd_ready_for_cancel_log(tdtd), &lwi);
1603         };
1604
1605         l_wait_event(tdtd->tdtd_commit_thread_waitq,
1606                      atomic_read(&tdtd->tdtd_refcount) == 0, &lwi);
1607
1608         spin_lock(&tdtd->tdtd_batchid_lock);
1609         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1610                                  tmt_commit_list)
1611                 list_move_tail(&tmt->tmt_commit_list, &list);
1612         spin_unlock(&tdtd->tdtd_batchid_lock);
1613
1614         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1615                tdtd->tdtd_lut->lut_obd->obd_name);
1616         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1617                 list_del_init(&tmt->tmt_commit_list);
1618                 top_multiple_thandle_dump(tmt, D_HA);
1619                 top_multiple_thandle_put(tmt);
1620         }
1621
1622         thread->t_flags = SVC_STOPPED;
1623         lu_env_fini(&env);
1624         wake_up(&thread->t_ctl_waitq);
1625
1626         RETURN(0);
1627 }
1628
1629 /**
1630  * Start llog cancel thread
1631  *
1632  * Start llog cancel(master/slave) thread on LOD
1633  *
1634  * \param[in]lclt       cancel log thread to be started.
1635  *
1636  * \retval              0 if the thread is started successfully.
1637  * \retval              negative errno if the thread is not being
1638  *                      started.
1639  */
1640 int distribute_txn_init(const struct lu_env *env,
1641                         struct lu_target *lut,
1642                         struct target_distribute_txn_data *tdtd,
1643                         __u32 index)
1644 {
1645         struct task_struct      *task;
1646         struct l_wait_info       lwi = { 0 };
1647         int                     rc;
1648         ENTRY;
1649
1650         spin_lock_init(&tdtd->tdtd_batchid_lock);
1651         INIT_LIST_HEAD(&tdtd->tdtd_list);
1652
1653         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1654
1655         init_waitqueue_head(&lut->lut_tdtd_commit_thread.t_ctl_waitq);
1656         init_waitqueue_head(&tdtd->tdtd_commit_thread_waitq);
1657         atomic_set(&tdtd->tdtd_refcount, 0);
1658
1659         tdtd->tdtd_lut = lut;
1660         rc = distribute_txn_commit_batchid_init(env, tdtd);
1661         if (rc != 0)
1662                 RETURN(rc);
1663
1664         task = kthread_run(distribute_txn_commit_thread, tdtd, "tdtd-%u",
1665                            index);
1666         if (IS_ERR(task))
1667                 RETURN(PTR_ERR(task));
1668
1669         l_wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1670                      distribute_txn_commit_thread_running(lut) ||
1671                      distribute_txn_commit_thread_stopped(lut), &lwi);
1672         RETURN(0);
1673 }
1674 EXPORT_SYMBOL(distribute_txn_init);
1675
1676 /**
1677  * Stop llog cancel thread
1678  *
1679  * Stop llog cancel(master/slave) thread on LOD and also destory
1680  * all of transaction in the list.
1681  *
1682  * \param[in]lclt       cancel log thread to be stopped.
1683  */
1684 void distribute_txn_fini(const struct lu_env *env,
1685                          struct target_distribute_txn_data *tdtd)
1686 {
1687         struct lu_target *lut = tdtd->tdtd_lut;
1688
1689         /* Stop cancel thread */
1690         if (lut == NULL || !distribute_txn_commit_thread_running(lut))
1691                 return;
1692
1693         spin_lock(&tdtd->tdtd_batchid_lock);
1694         lut->lut_tdtd_commit_thread.t_flags = SVC_STOPPING;
1695         spin_unlock(&tdtd->tdtd_batchid_lock);
1696         wake_up(&tdtd->tdtd_commit_thread_waitq);
1697         wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1698                    lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED);
1699
1700         dtrq_list_destroy(tdtd);
1701         if (tdtd->tdtd_batchid_obj != NULL) {
1702                 lu_object_put(env, &tdtd->tdtd_batchid_obj->do_lu);
1703                 tdtd->tdtd_batchid_obj = NULL;
1704         }
1705 }
1706 EXPORT_SYMBOL(distribute_txn_fini);