Whamcloud - gitweb
LU-7400 lod: register stop callbacks at create
[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, 2015, 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         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
576         return 0;
577 }
578
579 /**
580  * Create the top transaction.
581  *
582  * Create the top transaction on the master device. It will create a top
583  * thandle and a sub thandle on the master device.
584  *
585  * \param[in] env               execution environment
586  * \param[in] master_dev        master_dev the top thandle will be created
587  *
588  * \retval                      pointer to the created thandle.
589  * \retval                      ERR_PTR(errno) if creation failed.
590  */
591 struct thandle *
592 top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
593 {
594         struct top_thandle      *top_th;
595         struct thandle          *child_th;
596
597         OBD_ALLOC_GFP(top_th, sizeof(*top_th), __GFP_IO);
598         if (top_th == NULL)
599                 return ERR_PTR(-ENOMEM);
600
601         top_th->tt_super.th_top = &top_th->tt_super;
602
603         if (master_dev != NULL) {
604                 child_th = dt_trans_create(env, master_dev);
605                 if (IS_ERR(child_th)) {
606                         OBD_FREE_PTR(top_th);
607                         return child_th;
608                 }
609
610                 child_th->th_top = &top_th->tt_super;
611                 child_th->th_wait_submit = 1;
612                 top_th->tt_master_sub_thandle = child_th;
613
614                 top_th->tt_super.th_tags |= child_th->th_tags;
615         }
616         return &top_th->tt_super;
617 }
618 EXPORT_SYMBOL(top_trans_create);
619
620 /**
621  * Declare write update transaction
622  *
623  * Check if there are updates being recorded in this transaction,
624  * it will write the record into the disk.
625  *
626  * \param[in] env       execution environment
627  * \param[in] tmt       top multiple transaction handle
628  *
629  * \retval              0 if writing succeeds
630  * \retval              negative errno if writing fails
631  */
632 static int declare_updates_write(const struct lu_env *env,
633                                  struct top_multiple_thandle *tmt)
634 {
635         struct llog_update_record *record;
636         struct sub_thandle *st;
637         int rc = 0;
638
639         record = tmt->tmt_update_records->tur_update_records;
640         /* Declare update write for all other target */
641         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
642                 if (st->st_sub_th == NULL)
643                         continue;
644
645                 rc = sub_declare_updates_write(env, record, st->st_sub_th,
646                                                tmt->tmt_record_size);
647                 if (rc < 0)
648                         break;
649         }
650
651         return rc;
652 }
653
654 /**
655  * Assign batchid to the distribute transaction.
656  *
657  * Assign batchid to the distribute transaction
658  *
659  * \param[in] tmt       distribute transaction
660  */
661 static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
662 {
663         struct target_distribute_txn_data *tdtd;
664         struct dt_device *dt = new->tmt_master_sub_dt;
665         struct sub_thandle *st;
666
667         LASSERT(dt != NULL);
668         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
669         spin_lock(&tdtd->tdtd_batchid_lock);
670         new->tmt_batchid = tdtd->tdtd_batchid++;
671         list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
672         spin_unlock(&tdtd->tdtd_batchid_lock);
673         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
674                 if (st->st_sub_th != NULL)
675                         sub_thandle_register_commit_cb(st, new);
676         }
677         top_multiple_thandle_get(new);
678         top_multiple_thandle_dump(new, D_INFO);
679 }
680
681 /**
682  * Insert distribute transaction to the distribute txn list.
683  *
684  * Insert distribute transaction to the distribute txn list.
685  *
686  * \param[in] new       the distribute txn to be inserted.
687  */
688 void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
689 {
690         struct dt_device *dt = new->tmt_master_sub_dt;
691         struct top_multiple_thandle *tmt;
692         struct target_distribute_txn_data *tdtd;
693         struct sub_thandle *st;
694         bool    at_head = false;
695
696         LASSERT(dt != NULL);
697         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
698
699         spin_lock(&tdtd->tdtd_batchid_lock);
700         list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
701                 if (new->tmt_batchid > tmt->tmt_batchid) {
702                         list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
703                         break;
704                 }
705         }
706         if (list_empty(&new->tmt_commit_list)) {
707                 at_head = true;
708                 list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
709         }
710         spin_unlock(&tdtd->tdtd_batchid_lock);
711
712         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
713                 if (st->st_sub_th != NULL)
714                         sub_thandle_register_commit_cb(st, new);
715         }
716
717         top_multiple_thandle_get(new);
718         top_multiple_thandle_dump(new, D_INFO);
719         if (new->tmt_committed && at_head)
720                 wake_up(&tdtd->tdtd_commit_thread_waitq);
721 }
722
723 /**
724  * Prepare cross-MDT operation.
725  *
726  * Create the update record buffer to record updates for cross-MDT operation,
727  * add master sub transaction to tt_sub_trans_list, and declare the update
728  * writes.
729  *
730  * During updates packing, all of parameters will be packed in
731  * tur_update_params, and updates will be packed in tur_update_records.
732  * Then in transaction stop, parameters and updates will be merged
733  * into one updates buffer.
734  *
735  * And also master thandle will be added to the sub_th list, so it will be
736  * easy to track the commit status.
737  *
738  * \param[in] env       execution environment
739  * \param[in] th        top transaction handle
740  *
741  * \retval              0 if preparation succeeds.
742  * \retval              negative errno if preparation fails.
743  */
744 static int prepare_multiple_node_trans(const struct lu_env *env,
745                                        struct top_multiple_thandle *tmt)
746 {
747         struct thandle_update_records   *tur;
748         int                             rc;
749         ENTRY;
750
751         if (tmt->tmt_update_records == NULL) {
752                 tur = &update_env_info(env)->uti_tur;
753                 rc = check_and_prepare_update_record(env, tur);
754                 if (rc < 0)
755                         RETURN(rc);
756
757                 tmt->tmt_update_records = tur;
758                 distribute_txn_assign_batchid(tmt);
759         }
760
761         rc = declare_updates_write(env, tmt);
762
763         RETURN(rc);
764 }
765
766 /**
767  * start the top transaction.
768  *
769  * Start all of its sub transactions, then start master sub transaction.
770  *
771  * \param[in] env               execution environment
772  * \param[in] master_dev        master_dev the top thandle will be start
773  * \param[in] th                top thandle
774  *
775  * \retval                      0 if transaction start succeeds.
776  * \retval                      negative errno if start fails.
777  */
778 int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
779                     struct thandle *th)
780 {
781         struct top_thandle      *top_th = container_of(th, struct top_thandle,
782                                                        tt_super);
783         struct sub_thandle              *st;
784         struct top_multiple_thandle     *tmt = top_th->tt_multiple_thandle;
785         int                             rc = 0;
786         ENTRY;
787
788         if (tmt == NULL) {
789                 if (th->th_sync)
790                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
791                 if (th->th_local)
792                         top_th->tt_master_sub_thandle->th_local = th->th_local;
793                 top_th->tt_master_sub_thandle->th_tags = th->th_tags;
794                 rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
795                                     top_th->tt_master_sub_thandle);
796                 RETURN(rc);
797         }
798
799         tmt = top_th->tt_multiple_thandle;
800         rc = prepare_multiple_node_trans(env, tmt);
801         if (rc < 0)
802                 RETURN(rc);
803
804         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
805                 if (st->st_sub_th == NULL)
806                         continue;
807                 if (th->th_sync)
808                         st->st_sub_th->th_sync = th->th_sync;
809                 if (th->th_local)
810                         st->st_sub_th->th_local = th->th_local;
811                 st->st_sub_th->th_tags = th->th_tags;
812                 rc = dt_trans_start(env, st->st_sub_th->th_dev,
813                                     st->st_sub_th);
814                 if (rc != 0)
815                         GOTO(out, rc);
816         }
817 out:
818         th->th_result = rc;
819         RETURN(rc);
820 }
821 EXPORT_SYMBOL(top_trans_start);
822
823 /**
824  * Check whether we need write updates record
825  *
826  * Check if the updates for the top_thandle needs to be writen
827  * to all targets. Only if the transaction succeeds and the updates
828  * number > 2, it will write the updates,
829  *
830  * \params [in] top_th  top thandle.
831  *
832  * \retval              true if it needs to write updates
833  * \retval              false if it does not need to write updates
834  **/
835 static bool top_check_write_updates(struct top_thandle *top_th)
836 {
837         struct top_multiple_thandle     *tmt;
838         struct thandle_update_records   *tur;
839
840         /* Do not write updates to records if the transaction fails */
841         if (top_th->tt_super.th_result != 0)
842                 return false;
843
844         tmt = top_th->tt_multiple_thandle;
845         if (tmt == NULL)
846                 return false;
847
848         tur = tmt->tmt_update_records;
849         if (tur == NULL)
850                 return false;
851
852         /* Hmm, false update records, since the cross-MDT operation
853          * should includes both local and remote updates, so the
854          * updates count should >= 2 */
855         if (tur->tur_update_records == NULL ||
856             tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
857                 return false;
858
859         return true;
860 }
861
862 /**
863  * Check if top transaction is stopped
864  *
865  * Check if top transaction is stopped, only if all sub transaction
866  * is stopped, then the top transaction is stopped.
867  *
868  * \param [in] top_th   top thandle
869  *
870  * \retval              true if the top transaction is stopped.
871  * \retval              false if the top transaction is not stopped.
872  */
873 static bool top_trans_is_stopped(struct top_thandle *top_th)
874 {
875         struct top_multiple_thandle     *tmt;
876         struct sub_thandle              *st;
877         bool                    all_stopped = true;
878
879         tmt = top_th->tt_multiple_thandle;
880         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
881                 if (!st->st_stopped && st->st_sub_th != NULL) {
882                         all_stopped = false;
883                         break;
884                 }
885
886                 if (st->st_result != 0 &&
887                     top_th->tt_super.th_result == 0)
888                         top_th->tt_super.th_result = st->st_result;
889         }
890
891         return all_stopped;
892 }
893
894 /**
895  * Wait result of top transaction
896  *
897  * Wait until all sub transaction get its result.
898  *
899  * \param [in] top_th   top thandle.
900  *
901  * \retval              the result of top thandle.
902  */
903 static int top_trans_wait_result(struct top_thandle *top_th)
904 {
905         struct l_wait_info      lwi = {0};
906
907         l_wait_event(top_th->tt_multiple_thandle->tmt_stop_waitq,
908                      top_trans_is_stopped(top_th), &lwi);
909
910         RETURN(top_th->tt_super.th_result);
911 }
912
913 /**
914  * Stop the top transaction.
915  *
916  * Stop the transaction on the master device first, then stop transactions
917  * on other sub devices.
918  *
919  * \param[in] env               execution environment
920  * \param[in] master_dev        master_dev the top thandle will be created
921  * \param[in] th                top thandle
922  *
923  * \retval                      0 if stop transaction succeeds.
924  * \retval                      negative errno if stop transaction fails.
925  */
926 int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
927                    struct thandle *th)
928 {
929         struct top_thandle      *top_th = container_of(th, struct top_thandle,
930                                                        tt_super);
931         struct sub_thandle              *st;
932         struct sub_thandle              *master_st;
933         struct top_multiple_thandle     *tmt;
934         struct thandle_update_records   *tur;
935         bool                            write_updates = false;
936         int                     rc = 0;
937         ENTRY;
938
939         if (likely(top_th->tt_multiple_thandle == NULL)) {
940                 LASSERT(master_dev != NULL);
941
942                 if (th->th_sync)
943                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
944                 if (th->th_local)
945                         top_th->tt_master_sub_thandle->th_local = th->th_local;
946                 top_th->tt_master_sub_thandle->th_tags = th->th_tags;
947                 rc = dt_trans_stop(env, master_dev,
948                                    top_th->tt_master_sub_thandle);
949                 OBD_FREE_PTR(top_th);
950                 RETURN(rc);
951         }
952
953         tmt = top_th->tt_multiple_thandle;
954         tur = tmt->tmt_update_records;
955
956         /* Note: we need stop the master thandle first, then the stop
957          * callback will fill the master transno in the update logs,
958          * then these update logs will be sent to other MDTs */
959         /* get the master sub thandle */
960         master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
961         write_updates = top_check_write_updates(top_th);
962
963         /* Step 1: write the updates log on Master MDT */
964         if (master_st != NULL && master_st->st_sub_th != NULL &&
965             write_updates) {
966                 struct llog_update_record *lur;
967
968                 /* Merge the parameters and updates into one buffer */
969                 rc = prepare_writing_updates(env, tmt);
970                 if (rc < 0) {
971                         CERROR("%s: cannot prepare updates: rc = %d\n",
972                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
973                         th->th_result = rc;
974                         GOTO(stop_master_trans, rc);
975                 }
976
977                 lur = tur->tur_update_records;
978                 /* Write updates to the master MDT */
979                 rc = sub_updates_write(env, lur, master_st);
980
981                 /* Cleanup the common parameters in the update records,
982                  * master transno callback might add more parameters.
983                  * and we need merge the update records again in the
984                  * following */
985                 if (tur->tur_update_params != NULL)
986                         lur->lur_update_rec.ur_param_count = 0;
987
988                 if (rc < 0) {
989                         CERROR("%s: write updates failed: rc = %d\n",
990                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
991                         th->th_result = rc;
992                         GOTO(stop_master_trans, rc);
993                 }
994         }
995
996 stop_master_trans:
997         /* Step 2: Stop the transaction on the master MDT, and fill the
998          * master transno in the update logs to other MDT. */
999         if (master_st != NULL && master_st->st_sub_th != NULL) {
1000                 if (th->th_local)
1001                         master_st->st_sub_th->th_local = th->th_local;
1002                 if (th->th_sync)
1003                         master_st->st_sub_th->th_sync = th->th_sync;
1004                 master_st->st_sub_th->th_tags = th->th_tags;
1005                 master_st->st_sub_th->th_result = th->th_result;
1006                 rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
1007                 if (rc < 0) {
1008                         th->th_result = rc;
1009                         GOTO(stop_other_trans, rc);
1010                 } else if (tur != NULL && tur->tur_update_records != NULL) {
1011                         struct llog_update_record *lur;
1012
1013                         lur = tur->tur_update_records;
1014                         if (lur->lur_update_rec.ur_master_transno == 0)
1015                                 /* Update master transno after master stop
1016                                  * callback */
1017                                 lur->lur_update_rec.ur_master_transno =
1018                                                 tgt_th_info(env)->tti_transno;
1019                 }
1020         }
1021
1022         /* Step 3: write updates to other MDTs */
1023         if (write_updates) {
1024                 struct llog_update_record *lur;
1025
1026                 /* Stop callback of master will add more updates and also update
1027                  * master transno, so merge the parameters and updates into one
1028                  * buffer again */
1029                 rc = prepare_writing_updates(env, tmt);
1030                 if (rc < 0) {
1031                         CERROR("%s: prepare updates failed: rc = %d\n",
1032                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
1033                         th->th_result = rc;
1034                         GOTO(stop_other_trans, rc);
1035                 }
1036                 lur = tur->tur_update_records;
1037                 list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
1038                                     st_sub_list) {
1039                         if (st->st_sub_th == NULL || st == master_st ||
1040                             st->st_sub_th->th_result < 0)
1041                                 continue;
1042
1043                         rc = sub_updates_write(env, lur, st);
1044                         if (rc < 0) {
1045                                 th->th_result = rc;
1046                                 break;
1047                         }
1048                 }
1049         }
1050
1051 stop_other_trans:
1052         /* Step 4: Stop the transaction on other MDTs */
1053         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1054                 if (st == master_st || st->st_sub_th == NULL)
1055                         continue;
1056
1057                 if (th->th_sync)
1058                         st->st_sub_th->th_sync = th->th_sync;
1059                 if (th->th_local)
1060                         st->st_sub_th->th_local = th->th_local;
1061                 st->st_sub_th->th_tags = th->th_tags;
1062                 st->st_sub_th->th_result = th->th_result;
1063                 rc = dt_trans_stop(env, st->st_sub_th->th_dev,
1064                                    st->st_sub_th);
1065                 if (unlikely(rc < 0 && th->th_result == 0))
1066                         th->th_result = rc;
1067         }
1068
1069         rc = top_trans_wait_result(top_th);
1070
1071         tmt->tmt_result = rc;
1072
1073         /* Balance for the refcount in top_trans_create, Note: if it is NOT
1074          * multiple node transaction, the top transaction will be destroyed. */
1075         top_multiple_thandle_put(tmt);
1076         OBD_FREE_PTR(top_th);
1077         RETURN(rc);
1078 }
1079 EXPORT_SYMBOL(top_trans_stop);
1080
1081 /**
1082  * Create top_multiple_thandle for top_thandle
1083  *
1084  * Create top_mutilple_thandle to manage the mutiple node transaction
1085  * for top_thandle, and it also needs to add master sub thandle to the
1086  * sub trans list now.
1087  *
1088  * \param[in] env       execution environment
1089  * \param[in] top_th    the top thandle
1090  *
1091  * \retval      0 if creation succeeds
1092  * \retval      negative errno if creation fails
1093  */
1094 int top_trans_create_tmt(const struct lu_env *env,
1095                          struct top_thandle *top_th)
1096 {
1097         struct top_multiple_thandle *tmt;
1098
1099         OBD_ALLOC_PTR(tmt);
1100         if (tmt == NULL)
1101                 return -ENOMEM;
1102
1103         tmt->tmt_magic = TOP_THANDLE_MAGIC;
1104         INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
1105         INIT_LIST_HEAD(&tmt->tmt_commit_list);
1106         atomic_set(&tmt->tmt_refcount, 1);
1107         spin_lock_init(&tmt->tmt_sub_lock);
1108         init_waitqueue_head(&tmt->tmt_stop_waitq);
1109
1110         top_th->tt_multiple_thandle = tmt;
1111
1112         return 0;
1113 }
1114
1115 static struct sub_thandle *
1116 create_sub_thandle_with_thandle(struct top_thandle *top_th,
1117                                 struct thandle *sub_th)
1118 {
1119         struct sub_thandle *st;
1120
1121         /* create and init sub th to the top trans list */
1122         st = create_sub_thandle(top_th->tt_multiple_thandle,
1123                                 sub_th->th_dev);
1124         if (IS_ERR(st))
1125                 return st;
1126
1127         st->st_sub_th = sub_th;
1128
1129         sub_th->th_top = &top_th->tt_super;
1130         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
1131         return st;
1132 }
1133
1134 /**
1135  * Get sub thandle.
1136  *
1137  * Get sub thandle from the top thandle according to the sub dt_device.
1138  *
1139  * \param[in] env       execution environment
1140  * \param[in] th        thandle on the top layer.
1141  * \param[in] sub_dt    sub dt_device used to get sub transaction
1142  *
1143  * \retval              thandle of sub transaction if succeed
1144  * \retval              PTR_ERR(errno) if failed
1145  */
1146 struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
1147                                       struct thandle *th,
1148                                       struct dt_device *sub_dt)
1149 {
1150         struct sub_thandle      *st = NULL;
1151         struct sub_thandle      *master_st = NULL;
1152         struct top_thandle      *top_th;
1153         struct thandle          *sub_th = NULL;
1154         int                     rc = 0;
1155         ENTRY;
1156
1157         top_th = container_of(th, struct top_thandle, tt_super);
1158
1159         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
1160                 RETURN(top_th->tt_master_sub_thandle);
1161
1162         if (top_th->tt_multiple_thandle != NULL) {
1163                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
1164                 if (st != NULL)
1165                         RETURN(st->st_sub_th);
1166         }
1167
1168         sub_th = dt_trans_create(env, sub_dt);
1169         if (IS_ERR(sub_th))
1170                 RETURN(sub_th);
1171
1172         /* Create top_multiple_thandle if necessary */
1173         if (top_th->tt_multiple_thandle == NULL) {
1174                 struct top_multiple_thandle *tmt;
1175
1176                 rc = top_trans_create_tmt(env, top_th);
1177                 if (rc < 0)
1178                         GOTO(stop_trans, rc);
1179
1180                 tmt = top_th->tt_multiple_thandle;
1181
1182                 /* Add master sub th to the top trans list */
1183                 tmt->tmt_master_sub_dt =
1184                         top_th->tt_master_sub_thandle->th_dev;
1185                 master_st = create_sub_thandle_with_thandle(top_th,
1186                                         top_th->tt_master_sub_thandle);
1187                 if (IS_ERR(master_st)) {
1188                         rc = PTR_ERR(master_st);
1189                         master_st = NULL;
1190                         GOTO(stop_trans, rc);
1191                 }
1192         }
1193
1194         /* create and init sub th to the top trans list */
1195         st = create_sub_thandle_with_thandle(top_th, sub_th);
1196         if (IS_ERR(st)) {
1197                 rc = PTR_ERR(st);
1198                 st = NULL;
1199                 GOTO(stop_trans, rc);
1200         }
1201         st->st_sub_th->th_wait_submit = 1;
1202 stop_trans:
1203         if (rc < 0) {
1204                 if (master_st != NULL) {
1205                         list_del(&master_st->st_sub_list);
1206                         OBD_FREE_PTR(master_st);
1207                 }
1208                 sub_th->th_result = rc;
1209                 dt_trans_stop(env, sub_dt, sub_th);
1210                 sub_th = ERR_PTR(rc);
1211         }
1212
1213         RETURN(sub_th);
1214 }
1215 EXPORT_SYMBOL(thandle_get_sub_by_dt);
1216
1217 /**
1218  * Top multiple thandle destroy
1219  *
1220  * Destroy multiple thandle and all its sub thandle.
1221  *
1222  * \param[in] tmt       top_multiple_thandle to be destroyed.
1223  */
1224 void top_multiple_thandle_destroy(struct top_multiple_thandle *tmt)
1225 {
1226         struct sub_thandle *st;
1227         struct sub_thandle *tmp;
1228
1229         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
1230         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
1231                                  st_sub_list) {
1232                 struct sub_thandle_cookie *stc;
1233                 struct sub_thandle_cookie *tmp;
1234
1235                 list_del(&st->st_sub_list);
1236                 list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
1237                                          stc_list) {
1238                         list_del(&stc->stc_list);
1239                         OBD_FREE_PTR(stc);
1240                 }
1241                 OBD_FREE_PTR(st);
1242         }
1243         OBD_FREE_PTR(tmt);
1244 }
1245 EXPORT_SYMBOL(top_multiple_thandle_destroy);
1246
1247 /**
1248  * Cancel the update log on MDTs
1249  *
1250  * Cancel the update log on MDTs then destroy the thandle.
1251  *
1252  * \param[in] env       execution environment
1253  * \param[in] tmt       the top multiple thandle whose updates records
1254  *                      will be cancelled.
1255  *
1256  * \retval              0 if cancellation succeeds.
1257  * \retval              negative errno if cancellation fails.
1258  */
1259 static int distribute_txn_cancel_records(const struct lu_env *env,
1260                                          struct top_multiple_thandle *tmt)
1261 {
1262         struct sub_thandle *st;
1263         ENTRY;
1264
1265         top_multiple_thandle_dump(tmt, D_INFO);
1266         /* Cancel update logs on other MDTs */
1267         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1268                 struct llog_ctxt        *ctxt;
1269                 struct obd_device       *obd;
1270                 struct llog_cookie      *cookie;
1271                 struct sub_thandle_cookie *stc;
1272                 int rc;
1273
1274                 obd = st->st_dt->dd_lu_dev.ld_obd;
1275                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
1276                 if (ctxt == NULL)
1277                         continue;
1278                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
1279                         cookie = &stc->stc_cookie;
1280                         if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
1281                                 continue;
1282
1283                         rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
1284                                                      cookie);
1285                         CDEBUG(D_HA, "%s: batchid %llu cancel update log "
1286                                DOSTID ".%u : rc = %d\n", obd->obd_name,
1287                                tmt->tmt_batchid,
1288                                POSTID(&cookie->lgc_lgl.lgl_oi),
1289                                cookie->lgc_index, rc);
1290                 }
1291
1292                 llog_ctxt_put(ctxt);
1293         }
1294
1295         RETURN(0);
1296 }
1297
1298 /**
1299  * Check if there are committed transaction
1300  *
1301  * Check if there are committed transaction in the distribute transaction
1302  * list, then cancel the update records for those committed transaction.
1303  * Because the distribute transaction in the list are sorted by batchid,
1304  * and cancellation will be done by batchid order, so we only check the first
1305  * the transaction(with lowest batchid) in the list.
1306  *
1307  * \param[in] lod       lod device where cancel thread is
1308  *
1309  * \retval              true if it is ready
1310  * \retval              false if it is not ready
1311  */
1312 static bool tdtd_ready_for_cancel_log(struct target_distribute_txn_data *tdtd)
1313 {
1314         struct top_multiple_thandle     *tmt = NULL;
1315         struct obd_device               *obd = tdtd->tdtd_lut->lut_obd;
1316         bool    ready = false;
1317
1318         spin_lock(&tdtd->tdtd_batchid_lock);
1319         if (!list_empty(&tdtd->tdtd_list)) {
1320                 tmt = list_entry(tdtd->tdtd_list.next,
1321                                  struct top_multiple_thandle, tmt_commit_list);
1322                 if (tmt->tmt_committed &&
1323                     (!obd->obd_recovering || (obd->obd_recovering &&
1324                     tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)))
1325                         ready = true;
1326         }
1327         spin_unlock(&tdtd->tdtd_batchid_lock);
1328
1329         return ready;
1330 }
1331
1332 struct distribute_txn_bid_data {
1333         struct dt_txn_commit_cb  dtbd_cb;
1334         struct target_distribute_txn_data      *dtbd_tdtd;
1335         __u64                    dtbd_batchid;
1336 };
1337
1338 /**
1339  * callback of updating commit batchid
1340  *
1341  * Updating commit batchid then wake up the commit thread to cancel the
1342  * records.
1343  *
1344  * \param[in]env        execution environment
1345  * \param[in]th         thandle to updating commit batchid
1346  * \param[in]cb         commit callback
1347  * \param[in]err        result of thandle
1348  */
1349 static void distribute_txn_batchid_cb(struct lu_env *env,
1350                                       struct thandle *th,
1351                                       struct dt_txn_commit_cb *cb,
1352                                       int err)
1353 {
1354         struct distribute_txn_bid_data          *dtbd = NULL;
1355         struct target_distribute_txn_data       *tdtd;
1356
1357         dtbd = container_of0(cb, struct distribute_txn_bid_data, dtbd_cb);
1358         tdtd = dtbd->dtbd_tdtd;
1359
1360         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1361               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1362         spin_lock(&tdtd->tdtd_batchid_lock);
1363         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1364             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1365                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1366         spin_unlock(&tdtd->tdtd_batchid_lock);
1367         atomic_dec(&tdtd->tdtd_refcount);
1368         wake_up(&tdtd->tdtd_commit_thread_waitq);
1369
1370         OBD_FREE_PTR(dtbd);
1371 }
1372
1373 /**
1374  * Update the commit batchid in disk
1375  *
1376  * Update commit batchid in the disk, after this is committed, it can start
1377  * to cancel the update records.
1378  *
1379  * \param[in] env       execution environment
1380  * \param[in] tdtd      distribute transaction structure
1381  * \param[in] batchid   commit batchid to be updated
1382  *
1383  * \retval              0 if update succeeds.
1384  * \retval              negative errno if update fails.
1385  */
1386 static int
1387 distribute_txn_commit_batchid_update(const struct lu_env *env,
1388                               struct target_distribute_txn_data *tdtd,
1389                               __u64 batchid)
1390 {
1391         struct distribute_txn_bid_data  *dtbd = NULL;
1392         struct thandle          *th;
1393         struct lu_buf            buf;
1394         __u64                    tmp;
1395         __u64                    off;
1396         int                      rc;
1397         ENTRY;
1398
1399         OBD_ALLOC_PTR(dtbd);
1400         if (dtbd == NULL)
1401                 RETURN(-ENOMEM);
1402         dtbd->dtbd_batchid = batchid;
1403         dtbd->dtbd_tdtd = tdtd;
1404         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1405         atomic_inc(&tdtd->tdtd_refcount);
1406
1407         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1408         if (IS_ERR(th)) {
1409                 OBD_FREE_PTR(dtbd);
1410                 RETURN(PTR_ERR(th));
1411         }
1412
1413         tmp = cpu_to_le64(batchid);
1414         buf.lb_buf = &tmp;
1415         buf.lb_len = sizeof(tmp);
1416         off = 0;
1417
1418         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1419                                      th);
1420         if (rc < 0)
1421                 GOTO(stop, rc);
1422
1423         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1424         if (rc < 0)
1425                 GOTO(stop, rc);
1426
1427         rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
1428         if (rc < 0)
1429                 GOTO(stop, rc);
1430
1431         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1432                              &off, th);
1433
1434         CDEBUG(D_INFO, "%s: update batchid "LPU64": rc = %d\n",
1435                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1436
1437 stop:
1438         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1439         if (rc < 0)
1440                 OBD_FREE_PTR(dtbd);
1441         RETURN(rc);
1442 }
1443
1444 /**
1445  * Init commit batchid for distribute transaction.
1446  *
1447  * Initialize the batchid object and get commit batchid from the object.
1448  *
1449  * \param[in] env       execution environment
1450  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1451  *
1452  * \retval              0 if initialization succeeds.
1453  * \retval              negative errno if initialization fails.
1454  **/
1455 static int
1456 distribute_txn_commit_batchid_init(const struct lu_env *env,
1457                                    struct target_distribute_txn_data *tdtd)
1458 {
1459         struct tgt_thread_info  *tti = tgt_th_info(env);
1460         struct lu_target        *lut = tdtd->tdtd_lut;
1461         struct lu_attr          *attr = &tti->tti_attr;
1462         struct lu_fid           *fid = &tti->tti_fid1;
1463         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1464         struct dt_object        *dt_obj = NULL;
1465         struct lu_buf           buf;
1466         __u64                   tmp;
1467         __u64                   off;
1468         int                     rc;
1469         ENTRY;
1470
1471         memset(attr, 0, sizeof(*attr));
1472         attr->la_valid = LA_MODE;
1473         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1474         dof->dof_type = dt_mode_to_dft(S_IFREG);
1475
1476         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1477
1478         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1479                                    attr);
1480         if (IS_ERR(dt_obj)) {
1481                 rc = PTR_ERR(dt_obj);
1482                 dt_obj = NULL;
1483                 GOTO(out_put, rc);
1484         }
1485
1486         tdtd->tdtd_batchid_obj = dt_obj;
1487
1488         buf.lb_buf = &tmp;
1489         buf.lb_len = sizeof(tmp);
1490         off = 0;
1491         rc = dt_read(env, dt_obj, &buf, &off);
1492         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1493                 CERROR("%s can't read last committed batchid: rc = %d\n",
1494                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1495                 if (rc > 0)
1496                         rc = -EINVAL;
1497                 GOTO(out_put, rc);
1498         } else if (rc == buf.lb_len) {
1499                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1500                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1501                        tdtd->tdtd_lut->lut_obd->obd_name,
1502                        tdtd->tdtd_committed_batchid);
1503                 rc = 0;
1504         }
1505
1506 out_put:
1507         if (rc < 0 && dt_obj != NULL) {
1508                 lu_object_put(env, &dt_obj->do_lu);
1509                 tdtd->tdtd_batchid_obj = NULL;
1510         }
1511         return rc;
1512 }
1513
1514 /**
1515  * manage the distribute transaction thread
1516  *
1517  * Distribute transaction are linked to the list, and once the distribute
1518  * transaction is committed, it will update the last committed batchid first,
1519  * after it is committed, it will cancel the records.
1520  *
1521  * \param[in] _arg      argument for commit thread
1522  *
1523  * \retval              0 if thread is running successfully
1524  * \retval              negative errno if the thread can not be run.
1525  */
1526 static int distribute_txn_commit_thread(void *_arg)
1527 {
1528         struct target_distribute_txn_data *tdtd = _arg;
1529         struct lu_target        *lut = tdtd->tdtd_lut;
1530         struct ptlrpc_thread    *thread = &lut->lut_tdtd_commit_thread;
1531         struct l_wait_info       lwi = { 0 };
1532         struct lu_env            env;
1533         struct list_head         list;
1534         int                      rc;
1535         struct top_multiple_thandle *tmt;
1536         struct top_multiple_thandle *tmp;
1537         __u64                    batchid = 0, committed;
1538
1539         ENTRY;
1540
1541         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
1542         if (rc != 0)
1543                 RETURN(rc);
1544
1545         spin_lock(&tdtd->tdtd_batchid_lock);
1546         thread->t_flags = SVC_RUNNING;
1547         spin_unlock(&tdtd->tdtd_batchid_lock);
1548         wake_up(&thread->t_ctl_waitq);
1549         INIT_LIST_HEAD(&list);
1550
1551         CDEBUG(D_HA, "%s: start commit thread committed batchid "LPU64"\n",
1552                tdtd->tdtd_lut->lut_obd->obd_name,
1553                tdtd->tdtd_committed_batchid);
1554
1555         while (distribute_txn_commit_thread_running(lut)) {
1556                 spin_lock(&tdtd->tdtd_batchid_lock);
1557                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1558                                          tmt_commit_list) {
1559                         if (tmt->tmt_committed == 0)
1560                                 break;
1561
1562                         /* Note: right now, replay is based on master MDT
1563                          * transno, but cancellation is based on batchid.
1564                          * so we do not try to cancel the update log until
1565                          * the recoverying is done, unless the update records
1566                          * batchid < committed_batchid. */
1567                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1568                                 list_move_tail(&tmt->tmt_commit_list, &list);
1569                         } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
1570                                 LASSERTF(tmt->tmt_batchid >= batchid,
1571                                          "tmt %p tmt_batchid: "LPU64", batchid "
1572                                           LPU64"\n", tmt, tmt->tmt_batchid,
1573                                          batchid);
1574                                 /* There are three types of distribution
1575                                  * transaction result
1576                                  *
1577                                  * 1. If tmt_result < 0, it means the
1578                                  * distribution transaction fails, which should
1579                                  * be rare, because once declare phase succeeds,
1580                                  * the operation should succeeds anyway. Note in
1581                                  * this case, we will still update batchid so
1582                                  * cancellation would be stopped.
1583                                  *
1584                                  * 2. If tmt_result == 0, it means the
1585                                  * distribution transaction succeeds, and we
1586                                  * will update batchid.
1587                                  *
1588                                  * 3. If tmt_result > 0, it means distribute
1589                                  * transaction is not yet committed on every
1590                                  * node, but we need release this tmt before
1591                                  * that, which usuually happens during umount.
1592                                  */
1593                                 if (tmt->tmt_result <= 0)
1594                                         batchid = tmt->tmt_batchid;
1595                                 list_move_tail(&tmt->tmt_commit_list, &list);
1596                         }
1597                 }
1598                 spin_unlock(&tdtd->tdtd_batchid_lock);
1599
1600                 CDEBUG(D_HA, "%s: batchid: "LPU64" committed batchid "
1601                        LPU64"\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1602                        tdtd->tdtd_committed_batchid);
1603                 /* update globally committed on a storage */
1604                 if (batchid > tdtd->tdtd_committed_batchid) {
1605                         rc = distribute_txn_commit_batchid_update(&env, tdtd,
1606                                                              batchid);
1607                         if (rc == 0)
1608                                 batchid = 0;
1609                 }
1610                 /* cancel the records for committed batchid's */
1611                 /* XXX: should we postpone cancel's till the end of recovery? */
1612                 committed = tdtd->tdtd_committed_batchid;
1613                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1614                         if (tmt->tmt_batchid > committed)
1615                                 break;
1616                         list_del_init(&tmt->tmt_commit_list);
1617                         if (tmt->tmt_result <= 0)
1618                                 distribute_txn_cancel_records(&env, tmt);
1619                         top_multiple_thandle_put(tmt);
1620                 }
1621
1622                 l_wait_event(tdtd->tdtd_commit_thread_waitq,
1623                              !distribute_txn_commit_thread_running(lut) ||
1624                              committed < tdtd->tdtd_committed_batchid ||
1625                              tdtd_ready_for_cancel_log(tdtd), &lwi);
1626         };
1627
1628         l_wait_event(tdtd->tdtd_commit_thread_waitq,
1629                      atomic_read(&tdtd->tdtd_refcount) == 0, &lwi);
1630
1631         spin_lock(&tdtd->tdtd_batchid_lock);
1632         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1633                                  tmt_commit_list)
1634                 list_move_tail(&tmt->tmt_commit_list, &list);
1635         spin_unlock(&tdtd->tdtd_batchid_lock);
1636
1637         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1638                tdtd->tdtd_lut->lut_obd->obd_name);
1639         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1640                 list_del_init(&tmt->tmt_commit_list);
1641                 top_multiple_thandle_dump(tmt, D_HA);
1642                 top_multiple_thandle_put(tmt);
1643         }
1644
1645         thread->t_flags = SVC_STOPPED;
1646         lu_env_fini(&env);
1647         wake_up(&thread->t_ctl_waitq);
1648
1649         RETURN(0);
1650 }
1651
1652 /**
1653  * Start llog cancel thread
1654  *
1655  * Start llog cancel(master/slave) thread on LOD
1656  *
1657  * \param[in]lclt       cancel log thread to be started.
1658  *
1659  * \retval              0 if the thread is started successfully.
1660  * \retval              negative errno if the thread is not being
1661  *                      started.
1662  */
1663 int distribute_txn_init(const struct lu_env *env,
1664                         struct lu_target *lut,
1665                         struct target_distribute_txn_data *tdtd,
1666                         __u32 index)
1667 {
1668         struct task_struct      *task;
1669         struct l_wait_info       lwi = { 0 };
1670         int                     rc;
1671         ENTRY;
1672
1673         INIT_LIST_HEAD(&tdtd->tdtd_list);
1674         INIT_LIST_HEAD(&tdtd->tdtd_replay_finish_list);
1675         INIT_LIST_HEAD(&tdtd->tdtd_replay_list);
1676         spin_lock_init(&tdtd->tdtd_batchid_lock);
1677         spin_lock_init(&tdtd->tdtd_replay_list_lock);
1678         tdtd->tdtd_replay_handler = distribute_txn_replay_handle;
1679         tdtd->tdtd_replay_ready = 0;
1680
1681         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1682
1683         init_waitqueue_head(&lut->lut_tdtd_commit_thread.t_ctl_waitq);
1684         init_waitqueue_head(&tdtd->tdtd_commit_thread_waitq);
1685         atomic_set(&tdtd->tdtd_refcount, 0);
1686
1687         tdtd->tdtd_lut = lut;
1688         rc = distribute_txn_commit_batchid_init(env, tdtd);
1689         if (rc != 0)
1690                 RETURN(rc);
1691
1692         task = kthread_run(distribute_txn_commit_thread, tdtd, "tdtd-%u",
1693                            index);
1694         if (IS_ERR(task))
1695                 RETURN(PTR_ERR(task));
1696
1697         l_wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1698                      distribute_txn_commit_thread_running(lut) ||
1699                      distribute_txn_commit_thread_stopped(lut), &lwi);
1700         RETURN(0);
1701 }
1702 EXPORT_SYMBOL(distribute_txn_init);
1703
1704 /**
1705  * Stop llog cancel thread
1706  *
1707  * Stop llog cancel(master/slave) thread on LOD and also destory
1708  * all of transaction in the list.
1709  *
1710  * \param[in]lclt       cancel log thread to be stopped.
1711  */
1712 void distribute_txn_fini(const struct lu_env *env,
1713                          struct target_distribute_txn_data *tdtd)
1714 {
1715         struct lu_target *lut = tdtd->tdtd_lut;
1716
1717         /* Stop cancel thread */
1718         if (lut == NULL || !distribute_txn_commit_thread_running(lut))
1719                 return;
1720
1721         spin_lock(&tdtd->tdtd_batchid_lock);
1722         lut->lut_tdtd_commit_thread.t_flags = SVC_STOPPING;
1723         spin_unlock(&tdtd->tdtd_batchid_lock);
1724         wake_up(&tdtd->tdtd_commit_thread_waitq);
1725         wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1726                    lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED);
1727
1728         dtrq_list_destroy(tdtd);
1729         if (tdtd->tdtd_batchid_obj != NULL) {
1730                 lu_object_put(env, &tdtd->tdtd_batchid_obj->do_lu);
1731                 tdtd->tdtd_batchid_obj = NULL;
1732         }
1733 }
1734 EXPORT_SYMBOL(distribute_txn_fini);