Whamcloud - gitweb
LU-6880 update: after reply move dtrq to finish list
[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 sub_thandle      *master_st = NULL;
1127         struct top_thandle      *top_th;
1128         struct thandle          *sub_th = NULL;
1129         int                     rc = 0;
1130         ENTRY;
1131
1132         top_th = container_of(th, struct top_thandle, tt_super);
1133
1134         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
1135                 RETURN(top_th->tt_master_sub_thandle);
1136
1137         if (top_th->tt_multiple_thandle != NULL) {
1138                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
1139                 if (st != NULL)
1140                         RETURN(st->st_sub_th);
1141         }
1142
1143         sub_th = dt_trans_create(env, sub_dt);
1144         if (IS_ERR(sub_th))
1145                 RETURN(sub_th);
1146
1147         /* Create top_multiple_thandle if necessary */
1148         if (top_th->tt_multiple_thandle == NULL) {
1149                 struct top_multiple_thandle *tmt;
1150
1151                 rc = top_trans_create_tmt(env, top_th);
1152                 if (rc < 0)
1153                         GOTO(stop_trans, rc);
1154
1155                 tmt = top_th->tt_multiple_thandle;
1156
1157                 /* Add master sub th to the top trans list */
1158                 tmt->tmt_master_sub_dt =
1159                         top_th->tt_master_sub_thandle->th_dev;
1160                 master_st = create_sub_thandle_with_thandle(top_th,
1161                                         top_th->tt_master_sub_thandle);
1162                 if (IS_ERR(master_st)) {
1163                         rc = PTR_ERR(master_st);
1164                         master_st = NULL;
1165                         GOTO(stop_trans, rc);
1166                 }
1167         }
1168
1169         /* create and init sub th to the top trans list */
1170         st = create_sub_thandle_with_thandle(top_th, sub_th);
1171         if (IS_ERR(st)) {
1172                 rc = PTR_ERR(st);
1173                 st = NULL;
1174                 GOTO(stop_trans, rc);
1175         }
1176         st->st_sub_th->th_wait_submit = 1;
1177 stop_trans:
1178         if (rc < 0) {
1179                 if (master_st != NULL) {
1180                         list_del(&master_st->st_sub_list);
1181                         OBD_FREE_PTR(master_st);
1182                 }
1183                 sub_th->th_result = rc;
1184                 dt_trans_stop(env, sub_dt, sub_th);
1185                 sub_th = ERR_PTR(rc);
1186         }
1187
1188         RETURN(sub_th);
1189 }
1190 EXPORT_SYMBOL(thandle_get_sub_by_dt);
1191
1192 /**
1193  * Top multiple thandle destroy
1194  *
1195  * Destroy multiple thandle and all its sub thandle.
1196  *
1197  * \param[in] tmt       top_multiple_thandle to be destroyed.
1198  */
1199 void top_multiple_thandle_destroy(struct top_multiple_thandle *tmt)
1200 {
1201         struct sub_thandle *st;
1202         struct sub_thandle *tmp;
1203
1204         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
1205         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
1206                                  st_sub_list) {
1207                 struct sub_thandle_cookie *stc;
1208                 struct sub_thandle_cookie *tmp;
1209
1210                 list_del(&st->st_sub_list);
1211                 list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
1212                                          stc_list) {
1213                         list_del(&stc->stc_list);
1214                         OBD_FREE_PTR(stc);
1215                 }
1216                 OBD_FREE_PTR(st);
1217         }
1218         OBD_FREE_PTR(tmt);
1219 }
1220 EXPORT_SYMBOL(top_multiple_thandle_destroy);
1221
1222 /**
1223  * Cancel the update log on MDTs
1224  *
1225  * Cancel the update log on MDTs then destroy the thandle.
1226  *
1227  * \param[in] env       execution environment
1228  * \param[in] tmt       the top multiple thandle whose updates records
1229  *                      will be cancelled.
1230  *
1231  * \retval              0 if cancellation succeeds.
1232  * \retval              negative errno if cancellation fails.
1233  */
1234 static int distribute_txn_cancel_records(const struct lu_env *env,
1235                                          struct top_multiple_thandle *tmt)
1236 {
1237         struct sub_thandle *st;
1238         ENTRY;
1239
1240         top_multiple_thandle_dump(tmt, D_INFO);
1241         /* Cancel update logs on other MDTs */
1242         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1243                 struct llog_ctxt        *ctxt;
1244                 struct obd_device       *obd;
1245                 struct llog_cookie      *cookie;
1246                 struct sub_thandle_cookie *stc;
1247                 int rc;
1248
1249                 obd = st->st_dt->dd_lu_dev.ld_obd;
1250                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
1251                 if (ctxt == NULL)
1252                         continue;
1253                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
1254                         cookie = &stc->stc_cookie;
1255                         if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
1256                                 continue;
1257
1258                         rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
1259                                                      cookie);
1260                         CDEBUG(D_HA, "%s: batchid %llu cancel update log "
1261                                DOSTID ".%u : rc = %d\n", obd->obd_name,
1262                                tmt->tmt_batchid,
1263                                POSTID(&cookie->lgc_lgl.lgl_oi),
1264                                cookie->lgc_index, rc);
1265                 }
1266
1267                 llog_ctxt_put(ctxt);
1268         }
1269
1270         RETURN(0);
1271 }
1272
1273 /**
1274  * Check if there are committed transaction
1275  *
1276  * Check if there are committed transaction in the distribute transaction
1277  * list, then cancel the update records for those committed transaction.
1278  * Because the distribute transaction in the list are sorted by batchid,
1279  * and cancellation will be done by batchid order, so we only check the first
1280  * the transaction(with lowest batchid) in the list.
1281  *
1282  * \param[in] lod       lod device where cancel thread is
1283  *
1284  * \retval              true if it is ready
1285  * \retval              false if it is not ready
1286  */
1287 static bool tdtd_ready_for_cancel_log(struct target_distribute_txn_data *tdtd)
1288 {
1289         struct top_multiple_thandle     *tmt = NULL;
1290         struct obd_device               *obd = tdtd->tdtd_lut->lut_obd;
1291         bool    ready = false;
1292
1293         spin_lock(&tdtd->tdtd_batchid_lock);
1294         if (!list_empty(&tdtd->tdtd_list)) {
1295                 tmt = list_entry(tdtd->tdtd_list.next,
1296                                  struct top_multiple_thandle, tmt_commit_list);
1297                 if (tmt->tmt_committed &&
1298                     (!obd->obd_recovering || (obd->obd_recovering &&
1299                     tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)))
1300                         ready = true;
1301         }
1302         spin_unlock(&tdtd->tdtd_batchid_lock);
1303
1304         return ready;
1305 }
1306
1307 struct distribute_txn_bid_data {
1308         struct dt_txn_commit_cb  dtbd_cb;
1309         struct target_distribute_txn_data      *dtbd_tdtd;
1310         __u64                    dtbd_batchid;
1311 };
1312
1313 /**
1314  * callback of updating commit batchid
1315  *
1316  * Updating commit batchid then wake up the commit thread to cancel the
1317  * records.
1318  *
1319  * \param[in]env        execution environment
1320  * \param[in]th         thandle to updating commit batchid
1321  * \param[in]cb         commit callback
1322  * \param[in]err        result of thandle
1323  */
1324 static void distribute_txn_batchid_cb(struct lu_env *env,
1325                                       struct thandle *th,
1326                                       struct dt_txn_commit_cb *cb,
1327                                       int err)
1328 {
1329         struct distribute_txn_bid_data          *dtbd = NULL;
1330         struct target_distribute_txn_data       *tdtd;
1331
1332         dtbd = container_of0(cb, struct distribute_txn_bid_data, dtbd_cb);
1333         tdtd = dtbd->dtbd_tdtd;
1334
1335         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1336               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1337         spin_lock(&tdtd->tdtd_batchid_lock);
1338         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1339             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1340                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1341         spin_unlock(&tdtd->tdtd_batchid_lock);
1342         atomic_dec(&tdtd->tdtd_refcount);
1343         wake_up(&tdtd->tdtd_commit_thread_waitq);
1344
1345         OBD_FREE_PTR(dtbd);
1346 }
1347
1348 /**
1349  * Update the commit batchid in disk
1350  *
1351  * Update commit batchid in the disk, after this is committed, it can start
1352  * to cancel the update records.
1353  *
1354  * \param[in] env       execution environment
1355  * \param[in] tdtd      distribute transaction structure
1356  * \param[in] batchid   commit batchid to be updated
1357  *
1358  * \retval              0 if update succeeds.
1359  * \retval              negative errno if update fails.
1360  */
1361 static int
1362 distribute_txn_commit_batchid_update(const struct lu_env *env,
1363                               struct target_distribute_txn_data *tdtd,
1364                               __u64 batchid)
1365 {
1366         struct distribute_txn_bid_data  *dtbd = NULL;
1367         struct thandle          *th;
1368         struct lu_buf            buf;
1369         __u64                    tmp;
1370         __u64                    off;
1371         int                      rc;
1372         ENTRY;
1373
1374         OBD_ALLOC_PTR(dtbd);
1375         if (dtbd == NULL)
1376                 RETURN(-ENOMEM);
1377         dtbd->dtbd_batchid = batchid;
1378         dtbd->dtbd_tdtd = tdtd;
1379         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1380         atomic_inc(&tdtd->tdtd_refcount);
1381
1382         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1383         if (IS_ERR(th)) {
1384                 OBD_FREE_PTR(dtbd);
1385                 RETURN(PTR_ERR(th));
1386         }
1387
1388         tmp = cpu_to_le64(batchid);
1389         buf.lb_buf = &tmp;
1390         buf.lb_len = sizeof(tmp);
1391         off = 0;
1392
1393         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1394                                      th);
1395         if (rc < 0)
1396                 GOTO(stop, rc);
1397
1398         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1399         if (rc < 0)
1400                 GOTO(stop, rc);
1401
1402         rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
1403         if (rc < 0)
1404                 GOTO(stop, rc);
1405
1406         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1407                              &off, th);
1408
1409         CDEBUG(D_INFO, "%s: update batchid "LPU64": rc = %d\n",
1410                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1411
1412 stop:
1413         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1414         if (rc < 0)
1415                 OBD_FREE_PTR(dtbd);
1416         RETURN(rc);
1417 }
1418
1419 /**
1420  * Init commit batchid for distribute transaction.
1421  *
1422  * Initialize the batchid object and get commit batchid from the object.
1423  *
1424  * \param[in] env       execution environment
1425  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1426  *
1427  * \retval              0 if initialization succeeds.
1428  * \retval              negative errno if initialization fails.
1429  **/
1430 static int
1431 distribute_txn_commit_batchid_init(const struct lu_env *env,
1432                                    struct target_distribute_txn_data *tdtd)
1433 {
1434         struct tgt_thread_info  *tti = tgt_th_info(env);
1435         struct lu_target        *lut = tdtd->tdtd_lut;
1436         struct lu_attr          *attr = &tti->tti_attr;
1437         struct lu_fid           *fid = &tti->tti_fid1;
1438         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1439         struct dt_object        *dt_obj = NULL;
1440         struct lu_buf           buf;
1441         __u64                   tmp;
1442         __u64                   off;
1443         int                     rc;
1444         ENTRY;
1445
1446         memset(attr, 0, sizeof(*attr));
1447         attr->la_valid = LA_MODE;
1448         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1449         dof->dof_type = dt_mode_to_dft(S_IFREG);
1450
1451         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1452
1453         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1454                                    attr);
1455         if (IS_ERR(dt_obj)) {
1456                 rc = PTR_ERR(dt_obj);
1457                 dt_obj = NULL;
1458                 GOTO(out_put, rc);
1459         }
1460
1461         tdtd->tdtd_batchid_obj = dt_obj;
1462
1463         buf.lb_buf = &tmp;
1464         buf.lb_len = sizeof(tmp);
1465         off = 0;
1466         rc = dt_read(env, dt_obj, &buf, &off);
1467         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1468                 CERROR("%s can't read last committed batchid: rc = %d\n",
1469                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1470                 if (rc > 0)
1471                         rc = -EINVAL;
1472                 GOTO(out_put, rc);
1473         } else if (rc == buf.lb_len) {
1474                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1475                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1476                        tdtd->tdtd_lut->lut_obd->obd_name,
1477                        tdtd->tdtd_committed_batchid);
1478                 rc = 0;
1479         }
1480
1481 out_put:
1482         if (rc < 0 && dt_obj != NULL) {
1483                 lu_object_put(env, &dt_obj->do_lu);
1484                 tdtd->tdtd_batchid_obj = NULL;
1485         }
1486         return rc;
1487 }
1488
1489 /**
1490  * manage the distribute transaction thread
1491  *
1492  * Distribute transaction are linked to the list, and once the distribute
1493  * transaction is committed, it will update the last committed batchid first,
1494  * after it is committed, it will cancel the records.
1495  *
1496  * \param[in] _arg      argument for commit thread
1497  *
1498  * \retval              0 if thread is running successfully
1499  * \retval              negative errno if the thread can not be run.
1500  */
1501 static int distribute_txn_commit_thread(void *_arg)
1502 {
1503         struct target_distribute_txn_data *tdtd = _arg;
1504         struct lu_target        *lut = tdtd->tdtd_lut;
1505         struct ptlrpc_thread    *thread = &lut->lut_tdtd_commit_thread;
1506         struct l_wait_info       lwi = { 0 };
1507         struct lu_env            env;
1508         struct list_head         list;
1509         int                      rc;
1510         struct top_multiple_thandle *tmt;
1511         struct top_multiple_thandle *tmp;
1512         __u64                    batchid = 0, committed;
1513
1514         ENTRY;
1515
1516         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
1517         if (rc != 0)
1518                 RETURN(rc);
1519
1520         spin_lock(&tdtd->tdtd_batchid_lock);
1521         thread->t_flags = SVC_RUNNING;
1522         spin_unlock(&tdtd->tdtd_batchid_lock);
1523         wake_up(&thread->t_ctl_waitq);
1524         INIT_LIST_HEAD(&list);
1525
1526         CDEBUG(D_HA, "%s: start commit thread committed batchid "LPU64"\n",
1527                tdtd->tdtd_lut->lut_obd->obd_name,
1528                tdtd->tdtd_committed_batchid);
1529
1530         while (distribute_txn_commit_thread_running(lut)) {
1531                 spin_lock(&tdtd->tdtd_batchid_lock);
1532                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1533                                          tmt_commit_list) {
1534                         if (tmt->tmt_committed == 0)
1535                                 break;
1536
1537                         /* Note: right now, replay is based on master MDT
1538                          * transno, but cancellation is based on batchid.
1539                          * so we do not try to cancel the update log until
1540                          * the recoverying is done, unless the update records
1541                          * batchid < committed_batchid. */
1542                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1543                                 list_move_tail(&tmt->tmt_commit_list, &list);
1544                         } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
1545                                 LASSERTF(tmt->tmt_batchid >= batchid,
1546                                          "tmt %p tmt_batchid: "LPU64", batchid "
1547                                           LPU64"\n", tmt, tmt->tmt_batchid,
1548                                          batchid);
1549                                 /* There are three types of distribution
1550                                  * transaction result
1551                                  *
1552                                  * 1. If tmt_result < 0, it means the
1553                                  * distribution transaction fails, which should
1554                                  * be rare, because once declare phase succeeds,
1555                                  * the operation should succeeds anyway. Note in
1556                                  * this case, we will still update batchid so
1557                                  * cancellation would be stopped.
1558                                  *
1559                                  * 2. If tmt_result == 0, it means the
1560                                  * distribution transaction succeeds, and we
1561                                  * will update batchid.
1562                                  *
1563                                  * 3. If tmt_result > 0, it means distribute
1564                                  * transaction is not yet committed on every
1565                                  * node, but we need release this tmt before
1566                                  * that, which usuually happens during umount.
1567                                  */
1568                                 if (tmt->tmt_result <= 0)
1569                                         batchid = tmt->tmt_batchid;
1570                                 list_move_tail(&tmt->tmt_commit_list, &list);
1571                         }
1572                 }
1573                 spin_unlock(&tdtd->tdtd_batchid_lock);
1574
1575                 CDEBUG(D_HA, "%s: batchid: "LPU64" committed batchid "
1576                        LPU64"\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1577                        tdtd->tdtd_committed_batchid);
1578                 /* update globally committed on a storage */
1579                 if (batchid > tdtd->tdtd_committed_batchid) {
1580                         rc = distribute_txn_commit_batchid_update(&env, tdtd,
1581                                                              batchid);
1582                         if (rc == 0)
1583                                 batchid = 0;
1584                 }
1585                 /* cancel the records for committed batchid's */
1586                 /* XXX: should we postpone cancel's till the end of recovery? */
1587                 committed = tdtd->tdtd_committed_batchid;
1588                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1589                         if (tmt->tmt_batchid > committed)
1590                                 break;
1591                         list_del_init(&tmt->tmt_commit_list);
1592                         if (tmt->tmt_result <= 0)
1593                                 distribute_txn_cancel_records(&env, tmt);
1594                         top_multiple_thandle_put(tmt);
1595                 }
1596
1597                 l_wait_event(tdtd->tdtd_commit_thread_waitq,
1598                              !distribute_txn_commit_thread_running(lut) ||
1599                              committed < tdtd->tdtd_committed_batchid ||
1600                              tdtd_ready_for_cancel_log(tdtd), &lwi);
1601         };
1602
1603         l_wait_event(tdtd->tdtd_commit_thread_waitq,
1604                      atomic_read(&tdtd->tdtd_refcount) == 0, &lwi);
1605
1606         spin_lock(&tdtd->tdtd_batchid_lock);
1607         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1608                                  tmt_commit_list)
1609                 list_move_tail(&tmt->tmt_commit_list, &list);
1610         spin_unlock(&tdtd->tdtd_batchid_lock);
1611
1612         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1613                tdtd->tdtd_lut->lut_obd->obd_name);
1614         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1615                 list_del_init(&tmt->tmt_commit_list);
1616                 top_multiple_thandle_dump(tmt, D_HA);
1617                 top_multiple_thandle_put(tmt);
1618         }
1619
1620         thread->t_flags = SVC_STOPPED;
1621         lu_env_fini(&env);
1622         wake_up(&thread->t_ctl_waitq);
1623
1624         RETURN(0);
1625 }
1626
1627 /**
1628  * Start llog cancel thread
1629  *
1630  * Start llog cancel(master/slave) thread on LOD
1631  *
1632  * \param[in]lclt       cancel log thread to be started.
1633  *
1634  * \retval              0 if the thread is started successfully.
1635  * \retval              negative errno if the thread is not being
1636  *                      started.
1637  */
1638 int distribute_txn_init(const struct lu_env *env,
1639                         struct lu_target *lut,
1640                         struct target_distribute_txn_data *tdtd,
1641                         __u32 index)
1642 {
1643         struct task_struct      *task;
1644         struct l_wait_info       lwi = { 0 };
1645         int                     rc;
1646         ENTRY;
1647
1648         INIT_LIST_HEAD(&tdtd->tdtd_list);
1649         INIT_LIST_HEAD(&tdtd->tdtd_replay_finish_list);
1650         INIT_LIST_HEAD(&tdtd->tdtd_replay_list);
1651         spin_lock_init(&tdtd->tdtd_batchid_lock);
1652         spin_lock_init(&tdtd->tdtd_replay_list_lock);
1653         tdtd->tdtd_replay_handler = distribute_txn_replay_handle;
1654         tdtd->tdtd_replay_ready = 0;
1655
1656         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1657
1658         init_waitqueue_head(&lut->lut_tdtd_commit_thread.t_ctl_waitq);
1659         init_waitqueue_head(&tdtd->tdtd_commit_thread_waitq);
1660         atomic_set(&tdtd->tdtd_refcount, 0);
1661
1662         tdtd->tdtd_lut = lut;
1663         rc = distribute_txn_commit_batchid_init(env, tdtd);
1664         if (rc != 0)
1665                 RETURN(rc);
1666
1667         task = kthread_run(distribute_txn_commit_thread, tdtd, "tdtd-%u",
1668                            index);
1669         if (IS_ERR(task))
1670                 RETURN(PTR_ERR(task));
1671
1672         l_wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1673                      distribute_txn_commit_thread_running(lut) ||
1674                      distribute_txn_commit_thread_stopped(lut), &lwi);
1675         RETURN(0);
1676 }
1677 EXPORT_SYMBOL(distribute_txn_init);
1678
1679 /**
1680  * Stop llog cancel thread
1681  *
1682  * Stop llog cancel(master/slave) thread on LOD and also destory
1683  * all of transaction in the list.
1684  *
1685  * \param[in]lclt       cancel log thread to be stopped.
1686  */
1687 void distribute_txn_fini(const struct lu_env *env,
1688                          struct target_distribute_txn_data *tdtd)
1689 {
1690         struct lu_target *lut = tdtd->tdtd_lut;
1691
1692         /* Stop cancel thread */
1693         if (lut == NULL || !distribute_txn_commit_thread_running(lut))
1694                 return;
1695
1696         spin_lock(&tdtd->tdtd_batchid_lock);
1697         lut->lut_tdtd_commit_thread.t_flags = SVC_STOPPING;
1698         spin_unlock(&tdtd->tdtd_batchid_lock);
1699         wake_up(&tdtd->tdtd_commit_thread_waitq);
1700         wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1701                    lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED);
1702
1703         dtrq_list_destroy(tdtd);
1704         if (tdtd->tdtd_batchid_obj != NULL) {
1705                 lu_object_put(env, &tdtd->tdtd_batchid_obj->do_lu);
1706                 tdtd->tdtd_batchid_obj = NULL;
1707         }
1708 }
1709 EXPORT_SYMBOL(distribute_txn_fini);