Whamcloud - gitweb
387223c061c79a1ee5dccd075c050f7964f376c6
[fs/lustre-release.git] / lustre / target / tgt_lastrcvd.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) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  *
31  * Lustre Unified Target
32  * These are common function to work with last_received file
33  *
34  * Author: Mikhail Pershin <mike.pershin@intel.com>
35  */
36 #include <obd.h>
37 #include <obd_class.h>
38 #include <lustre_fid.h>
39
40 #include "tgt_internal.h"
41
42 /** version recovery epoch */
43 #define LR_EPOCH_BITS   32
44
45 /* Allocate a bitmap for a chunk of reply data slots */
46 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
47 {
48         unsigned long *bm;
49
50         OBD_ALLOC_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
51                         sizeof(long));
52         if (bm == NULL)
53                 return -ENOMEM;
54
55         spin_lock(&lut->lut_client_bitmap_lock);
56
57         if (lut->lut_reply_bitmap[chunk] != NULL) {
58                 /* someone else already allocated the bitmap for this chunk */
59                 spin_unlock(&lut->lut_client_bitmap_lock);
60                 OBD_FREE_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
61                          sizeof(long));
62                 return 0;
63         }
64
65         lut->lut_reply_bitmap[chunk] = bm;
66
67         spin_unlock(&lut->lut_client_bitmap_lock);
68
69         return 0;
70 }
71
72 /* Look for an available reply data slot in the bitmap
73  * of the target @lut
74  * Allocate bitmap chunk when first used
75  * XXX algo could be improved if this routine limits performance
76  */
77 static int tgt_find_free_reply_slot(struct lu_target *lut)
78 {
79         unsigned long *bmp;
80         int chunk = 0;
81         int rc;
82         int b;
83
84         for (chunk = 0; chunk < LUT_REPLY_SLOTS_MAX_CHUNKS; chunk++) {
85                 /* allocate the bitmap chunk if necessary */
86                 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
87                         rc = tgt_bitmap_chunk_alloc(lut, chunk);
88                         if (rc != 0)
89                                 return rc;
90                 }
91                 bmp = lut->lut_reply_bitmap[chunk];
92
93                 /* look for an available slot in this chunk */
94                 do {
95                         b = find_first_zero_bit(bmp, LUT_REPLY_SLOTS_PER_CHUNK);
96                         if (b >= LUT_REPLY_SLOTS_PER_CHUNK)
97                                 break;
98
99                         /* found one */
100                         if (test_and_set_bit(b, bmp) == 0)
101                                 return chunk * LUT_REPLY_SLOTS_PER_CHUNK + b;
102                 } while (true);
103         }
104
105         return -ENOSPC;
106 }
107
108 /* Mark the reply data slot @idx 'used' in the corresponding bitmap chunk
109  * of the target @lut
110  * Allocate the bitmap chunk if necessary
111  */
112 static int tgt_set_reply_slot(struct lu_target *lut, int idx)
113 {
114         int chunk;
115         int b;
116         int rc;
117
118         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
119         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
120
121         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
122         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
123
124         /* allocate the bitmap chunk if necessary */
125         if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
126                 rc = tgt_bitmap_chunk_alloc(lut, chunk);
127                 if (rc != 0)
128                         return rc;
129         }
130
131         /* mark the slot 'used' in this chunk */
132         if (test_and_set_bit(b, lut->lut_reply_bitmap[chunk]) != 0) {
133                 CERROR("%s: slot %d already set in bitmap\n",
134                        tgt_name(lut), idx);
135                 return -EALREADY;
136         }
137
138         return 0;
139 }
140
141
142 /* Mark the reply data slot @idx 'unused' in the corresponding bitmap chunk
143  * of the target @lut
144  */
145 static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
146 {
147         int chunk;
148         int b;
149
150         if (lut->lut_obd->obd_stopping)
151                 /*
152                  * in case of failover keep the bit set in order to
153                  * avoid overwriting slots in reply_data which might
154                  * be required by resent rpcs
155                  */
156                 return 0;
157         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
158         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
159
160         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
161         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
162
163         if (lut->lut_reply_bitmap[chunk] == NULL) {
164                 CERROR("%s: slot %d not allocated\n",
165                        tgt_name(lut), idx);
166                 return -ENOENT;
167         }
168
169         if (test_and_clear_bit(b, lut->lut_reply_bitmap[chunk]) == 0) {
170                 CERROR("%s: slot %d already clear in bitmap\n",
171                        tgt_name(lut), idx);
172                 return -EALREADY;
173         }
174
175         return 0;
176 }
177
178
179 /* Read header of reply_data file of target @tgt into structure @lrh */
180 static int tgt_reply_header_read(const struct lu_env *env,
181                                  struct lu_target *tgt,
182                                  struct lsd_reply_header *lrh)
183 {
184         int                      rc;
185         struct lsd_reply_header  buf;
186         struct tgt_thread_info  *tti = tgt_th_info(env);
187
188         tti->tti_off = 0;
189         tti->tti_buf.lb_buf = &buf;
190         tti->tti_buf.lb_len = sizeof(buf);
191
192         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
193                             &tti->tti_off);
194         if (rc != 0)
195                 return rc;
196
197         lrh->lrh_magic = le32_to_cpu(buf.lrh_magic);
198         lrh->lrh_header_size = le32_to_cpu(buf.lrh_header_size);
199         lrh->lrh_reply_size = le32_to_cpu(buf.lrh_reply_size);
200
201         CDEBUG(D_HA, "%s: read %s header. magic=0x%08x "
202                "header_size=%d reply_size=%d\n",
203                 tgt->lut_obd->obd_name, REPLY_DATA,
204                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
205
206         return 0;
207 }
208
209 /* Write header into replay_data file of target @tgt from structure @lrh */
210 static int tgt_reply_header_write(const struct lu_env *env,
211                                   struct lu_target *tgt,
212                                   struct lsd_reply_header *lrh)
213 {
214         int                      rc;
215         struct lsd_reply_header  buf;
216         struct tgt_thread_info  *tti = tgt_th_info(env);
217         struct thandle          *th;
218         struct dt_object        *dto;
219
220         CDEBUG(D_HA, "%s: write %s header. magic=0x%08x "
221                "header_size=%d reply_size=%d\n",
222                 tgt->lut_obd->obd_name, REPLY_DATA,
223                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
224
225         if (tgt->lut_bottom->dd_rdonly)
226                 RETURN(0);
227
228         buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
229         buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
230         buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
231
232         th = dt_trans_create(env, tgt->lut_bottom);
233         if (IS_ERR(th))
234                 return PTR_ERR(th);
235         th->th_sync = 1;
236
237         tti->tti_off = 0;
238         tti->tti_buf.lb_buf = &buf;
239         tti->tti_buf.lb_len = sizeof(buf);
240
241         rc = dt_declare_record_write(env, tgt->lut_reply_data,
242                                      &tti->tti_buf, tti->tti_off, th);
243         if (rc)
244                 GOTO(out, rc);
245
246         rc = dt_trans_start(env, tgt->lut_bottom, th);
247         if (rc)
248                 GOTO(out, rc);
249
250         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
251         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
252 out:
253         dt_trans_stop(env, tgt->lut_bottom, th);
254         return rc;
255 }
256
257 /* Write the reply data @lrd into reply_data file of target @tgt
258  * at offset @off
259  */
260 static int tgt_reply_data_write(const struct lu_env *env, struct lu_target *tgt,
261                                 struct lsd_reply_data *lrd, loff_t off,
262                                 struct thandle *th)
263 {
264         struct tgt_thread_info  *tti = tgt_th_info(env);
265         struct dt_object        *dto;
266         struct lsd_reply_data   *buf = &tti->tti_lrd;
267
268         lrd->lrd_result = ptlrpc_status_hton(lrd->lrd_result);
269
270         buf->lrd_transno         = cpu_to_le64(lrd->lrd_transno);
271         buf->lrd_xid             = cpu_to_le64(lrd->lrd_xid);
272         buf->lrd_data            = cpu_to_le64(lrd->lrd_data);
273         buf->lrd_result          = cpu_to_le32(lrd->lrd_result);
274         buf->lrd_client_gen      = cpu_to_le32(lrd->lrd_client_gen);
275
276         lrd->lrd_result = ptlrpc_status_ntoh(lrd->lrd_result);
277
278         tti->tti_off = off;
279         tti->tti_buf.lb_buf = buf;
280         tti->tti_buf.lb_len = sizeof(*buf);
281
282         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
283         return dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
284 }
285
286 /* Read the reply data from reply_data file of target @tgt at offset @off
287  * into structure @lrd
288  */
289 static int tgt_reply_data_read(const struct lu_env *env, struct lu_target *tgt,
290                                struct lsd_reply_data *lrd, loff_t off)
291 {
292         int                      rc;
293         struct tgt_thread_info  *tti = tgt_th_info(env);
294         struct lsd_reply_data   *buf = &tti->tti_lrd;
295
296         tti->tti_off = off;
297         tti->tti_buf.lb_buf = buf;
298         tti->tti_buf.lb_len = sizeof(*buf);
299
300         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
301                             &tti->tti_off);
302         if (rc != 0)
303                 return rc;
304
305         lrd->lrd_transno         = le64_to_cpu(buf->lrd_transno);
306         lrd->lrd_xid             = le64_to_cpu(buf->lrd_xid);
307         lrd->lrd_data            = le64_to_cpu(buf->lrd_data);
308         lrd->lrd_result          = le32_to_cpu(buf->lrd_result);
309         lrd->lrd_client_gen      = le32_to_cpu(buf->lrd_client_gen);
310
311         return 0;
312 }
313
314
315 /* Free the in-memory reply data structure @trd and release
316  * the corresponding slot in the reply_data file of target @lut
317  * Called with ted_lcd_lock held
318  */
319 static void tgt_free_reply_data(struct lu_target *lut,
320                                 struct tg_export_data *ted,
321                                 struct tg_reply_data *trd)
322 {
323         CDEBUG(D_TRACE, "%s: free reply data %p: xid %llu, transno %llu, "
324                "client gen %u, slot idx %d\n",
325                lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
326                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
327                trd->trd_index);
328
329         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
330
331         list_del(&trd->trd_list);
332         ted->ted_reply_cnt--;
333         if (lut != NULL && trd->trd_index != TRD_INDEX_MEMORY)
334                 tgt_clear_reply_slot(lut, trd->trd_index);
335         OBD_FREE_PTR(trd);
336 }
337
338 /* Release the reply data @trd from target @lut
339  * The reply data with the highest transno for this export
340  * is retained to ensure correctness of target recovery
341  * Called with ted_lcd_lock held
342  */
343 static void tgt_release_reply_data(struct lu_target *lut,
344                                    struct tg_export_data *ted,
345                                    struct tg_reply_data *trd)
346 {
347         CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
348                "client gen %u, slot idx %d\n",
349                lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
350                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
351                trd->trd_index);
352
353         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
354
355         /* Do not free the reply data corresponding to the
356          * highest transno of this export.
357          * This ensures on-disk reply data is kept and
358          * last committed transno can be restored from disk in case
359          * of target recovery
360          */
361         if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
362                 /* free previous retained reply */
363                 if (ted->ted_reply_last != NULL)
364                         tgt_free_reply_data(lut, ted, ted->ted_reply_last);
365                 /* retain the reply */
366                 list_del_init(&trd->trd_list);
367                 ted->ted_reply_last = trd;
368         } else {
369                 tgt_free_reply_data(lut, ted, trd);
370         }
371 }
372
373 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
374 {
375         tti->tti_buf.lb_buf = &tti->tti_lsd;
376         tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
377         return &tti->tti_buf;
378 }
379
380 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
381 {
382         tti->tti_buf.lb_buf = &tti->tti_lcd;
383         tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
384         return &tti->tti_buf;
385 }
386
387 /**
388  * Allocate in-memory data for client slot related to export.
389  */
390 int tgt_client_alloc(struct obd_export *exp)
391 {
392         ENTRY;
393         LASSERT(exp != exp->exp_obd->obd_self_export);
394
395         spin_lock_init(&exp->exp_target_data.ted_nodemap_lock);
396         INIT_LIST_HEAD(&exp->exp_target_data.ted_nodemap_member);
397         spin_lock_init(&exp->exp_target_data.ted_fmd_lock);
398         INIT_LIST_HEAD(&exp->exp_target_data.ted_fmd_list);
399
400         OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
401         if (exp->exp_target_data.ted_lcd == NULL)
402                 RETURN(-ENOMEM);
403         /* Mark that slot is not yet valid, 0 doesn't work here */
404         exp->exp_target_data.ted_lr_idx = -1;
405         INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
406         mutex_init(&exp->exp_target_data.ted_lcd_lock);
407         RETURN(0);
408 }
409 EXPORT_SYMBOL(tgt_client_alloc);
410
411 /**
412  * Free in-memory data for client slot related to export.
413  */
414 void tgt_client_free(struct obd_export *exp)
415 {
416         struct tg_export_data   *ted = &exp->exp_target_data;
417         struct lu_target        *lut = class_exp2tgt(exp);
418         struct tg_reply_data    *trd, *tmp;
419
420         LASSERT(exp != exp->exp_obd->obd_self_export);
421
422         tgt_fmd_cleanup(exp);
423
424         /* free reply data */
425         mutex_lock(&ted->ted_lcd_lock);
426         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
427                 tgt_release_reply_data(lut, ted, trd);
428         }
429         if (ted->ted_reply_last != NULL) {
430                 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
431                 ted->ted_reply_last = NULL;
432         }
433         mutex_unlock(&ted->ted_lcd_lock);
434
435         if (!hlist_unhashed(&exp->exp_gen_hash))
436                 cfs_hash_del(exp->exp_obd->obd_gen_hash,
437                              &ted->ted_lcd->lcd_generation,
438                              &exp->exp_gen_hash);
439
440         OBD_FREE_PTR(ted->ted_lcd);
441         ted->ted_lcd = NULL;
442
443         /* Target may have been freed (see LU-7430)
444          * Slot may be not yet assigned */
445         if (exp->exp_obd->u.obt.obt_magic != OBT_MAGIC ||
446             ted->ted_lr_idx < 0)
447                 return;
448
449         /* Clear bit when lcd is freed */
450         LASSERT(lut && lut->lut_client_bitmap);
451         if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
452                 CERROR("%s: client %u bit already clear in bitmap\n",
453                        exp->exp_obd->obd_name, ted->ted_lr_idx);
454                 LBUG();
455         }
456
457         if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
458                 atomic_dec(&lut->lut_num_clients);
459 }
460 EXPORT_SYMBOL(tgt_client_free);
461
462 static inline void tgt_check_lcd(const char *obd_name, int index,
463                                  struct lsd_client_data *lcd)
464 {
465         size_t uuid_size = sizeof(lcd->lcd_uuid);
466
467         if (strnlen((char*)lcd->lcd_uuid, uuid_size) == uuid_size) {
468                 lcd->lcd_uuid[uuid_size - 1] = '\0';
469
470                 LCONSOLE_ERROR("the client UUID (%s) on %s for exports stored in last_rcvd(index = %d) is bad!\n",
471                                lcd->lcd_uuid, obd_name, index);
472         }
473 }
474
475 static int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
476                                 struct lsd_client_data *lcd,
477                                 loff_t *off, int index)
478 {
479         struct tgt_thread_info  *tti = tgt_th_info(env);
480         int                      rc;
481
482         tti_buf_lcd(tti);
483         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
484         if (rc == 0) {
485                 tgt_check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
486                 lcd_le_to_cpu(&tti->tti_lcd, lcd);
487                 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
488                 lcd->lcd_last_close_result =
489                         ptlrpc_status_ntoh(lcd->lcd_last_close_result);
490         }
491
492         CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = %llu"
493                ", last_xid = %llu, last_result = %u, last_data = %u, "
494                "last_close_transno = %llu, last_close_xid = %llu, "
495                "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
496                *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
497                lcd->lcd_last_result, lcd->lcd_last_data,
498                lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
499                lcd->lcd_last_close_result, rc);
500         return rc;
501 }
502
503 static int tgt_client_data_write(const struct lu_env *env,
504                                  struct lu_target *tgt,
505                                  struct lsd_client_data *lcd,
506                                  loff_t *off, struct thandle *th)
507 {
508         struct tgt_thread_info *tti = tgt_th_info(env);
509         struct dt_object        *dto;
510
511         lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
512         lcd->lcd_last_close_result =
513                 ptlrpc_status_hton(lcd->lcd_last_close_result);
514         lcd_cpu_to_le(lcd, &tti->tti_lcd);
515         tti_buf_lcd(tti);
516
517         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
518         return dt_record_write(env, dto, &tti->tti_buf, off, th);
519 }
520
521 struct tgt_new_client_callback {
522         struct dt_txn_commit_cb  lncc_cb;
523         struct obd_export       *lncc_exp;
524 };
525
526 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
527                               struct dt_txn_commit_cb *cb, int err)
528 {
529         struct tgt_new_client_callback *ccb;
530
531         ccb = container_of(cb, struct tgt_new_client_callback, lncc_cb);
532
533         LASSERT(ccb->lncc_exp->exp_obd);
534
535         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
536                ccb->lncc_exp->exp_obd->obd_name,
537                ccb->lncc_exp->exp_client_uuid.uuid);
538
539         spin_lock(&ccb->lncc_exp->exp_lock);
540
541         ccb->lncc_exp->exp_need_sync = 0;
542
543         spin_unlock(&ccb->lncc_exp->exp_lock);
544         class_export_cb_put(ccb->lncc_exp);
545
546         OBD_FREE_PTR(ccb);
547 }
548
549 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
550 {
551         struct tgt_new_client_callback  *ccb;
552         struct dt_txn_commit_cb         *dcb;
553         int                              rc;
554
555         OBD_ALLOC_PTR(ccb);
556         if (ccb == NULL)
557                 return -ENOMEM;
558
559         ccb->lncc_exp = class_export_cb_get(exp);
560
561         dcb = &ccb->lncc_cb;
562         dcb->dcb_func = tgt_cb_new_client;
563         INIT_LIST_HEAD(&dcb->dcb_linkage);
564         strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
565
566         rc = dt_trans_cb_add(th, dcb);
567         if (rc) {
568                 class_export_cb_put(exp);
569                 OBD_FREE_PTR(ccb);
570         }
571         return rc;
572 }
573
574 /**
575  * Update client data in last_rcvd
576  */
577 static int tgt_client_data_update(const struct lu_env *env,
578                                   struct obd_export *exp)
579 {
580         struct tg_export_data   *ted = &exp->exp_target_data;
581         struct lu_target        *tgt = class_exp2tgt(exp);
582         struct tgt_thread_info  *tti = tgt_th_info(env);
583         struct thandle          *th;
584         int                      rc = 0;
585
586         ENTRY;
587
588         if (unlikely(tgt == NULL)) {
589                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
590                           class_exp2obd(exp)->obd_name);
591                 RETURN(-EINVAL);
592         }
593
594         if (tgt->lut_bottom->dd_rdonly)
595                 RETURN(0);
596
597         th = dt_trans_create(env, tgt->lut_bottom);
598         if (IS_ERR(th))
599                 RETURN(PTR_ERR(th));
600
601         tti_buf_lcd(tti);
602         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
603                                      &tti->tti_buf,
604                                      ted->ted_lr_off, th);
605         if (rc)
606                 GOTO(out, rc);
607
608         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
609         if (rc)
610                 GOTO(out, rc);
611
612         mutex_lock(&ted->ted_lcd_lock);
613
614         /*
615          * Until this operations will be committed the sync is needed
616          * for this export. This should be done _after_ starting the
617          * transaction so that many connecting clients will not bring
618          * server down with lots of sync writes.
619          */
620         rc = tgt_new_client_cb_add(th, exp);
621         if (rc) {
622                 /* can't add callback, do sync now */
623                 th->th_sync = 1;
624         } else {
625                 spin_lock(&exp->exp_lock);
626                 exp->exp_need_sync = 1;
627                 spin_unlock(&exp->exp_lock);
628         }
629
630         tti->tti_off = ted->ted_lr_off;
631         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
632
633         mutex_unlock(&ted->ted_lcd_lock);
634
635         EXIT;
636 out:
637         dt_trans_stop(env, tgt->lut_bottom, th);
638         CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
639                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
640                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
641
642         return rc;
643 }
644
645 static int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
646 {
647         struct tgt_thread_info  *tti = tgt_th_info(env);
648         int                      rc;
649
650         tti->tti_off = 0;
651         tti_buf_lsd(tti);
652         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
653                             &tti->tti_off);
654         if (rc == 0)
655                 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
656
657         CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
658                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
659                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
660         return rc;
661 }
662
663 static int tgt_server_data_write(const struct lu_env *env,
664                                  struct lu_target *tgt, struct thandle *th)
665 {
666         struct tgt_thread_info  *tti = tgt_th_info(env);
667         struct dt_object        *dto;
668         int                      rc;
669
670         ENTRY;
671
672         tti->tti_off = 0;
673         tti_buf_lsd(tti);
674         lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
675
676         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
677         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
678
679         CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
680                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
681                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
682
683         RETURN(rc);
684 }
685
686 /**
687  * Update server data in last_rcvd
688  */
689 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
690                            int sync)
691 {
692         struct tgt_thread_info  *tti = tgt_th_info(env);
693         struct thandle          *th;
694         int                      rc = 0;
695
696         ENTRY;
697
698         CDEBUG(D_SUPER,
699                "%s: mount_count is %llu, last_transno is %llu\n",
700                tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
701                tgt->lut_last_transno);
702
703         /* Always save latest transno to keep it fresh */
704         spin_lock(&tgt->lut_translock);
705         tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
706         spin_unlock(&tgt->lut_translock);
707
708         if (tgt->lut_bottom->dd_rdonly)
709                 RETURN(0);
710
711         th = dt_trans_create(env, tgt->lut_bottom);
712         if (IS_ERR(th))
713                 RETURN(PTR_ERR(th));
714
715         th->th_sync = sync;
716
717         tti_buf_lsd(tti);
718         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
719                                      &tti->tti_buf, tti->tti_off, th);
720         if (rc)
721                 GOTO(out, rc);
722
723         rc = dt_trans_start(env, tgt->lut_bottom, th);
724         if (rc)
725                 GOTO(out, rc);
726
727         rc = tgt_server_data_write(env, tgt, th);
728 out:
729         dt_trans_stop(env, tgt->lut_bottom, th);
730
731         CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
732                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
733                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
734         RETURN(rc);
735 }
736 EXPORT_SYMBOL(tgt_server_data_update);
737
738 static int tgt_truncate_last_rcvd(const struct lu_env *env,
739                                   struct lu_target *tgt, loff_t size)
740 {
741         struct dt_object *dt = tgt->lut_last_rcvd;
742         struct thandle   *th;
743         struct lu_attr    attr;
744         int               rc;
745
746         ENTRY;
747
748         if (tgt->lut_bottom->dd_rdonly)
749                 RETURN(0);
750
751         attr.la_size = size;
752         attr.la_valid = LA_SIZE;
753
754         th = dt_trans_create(env, tgt->lut_bottom);
755         if (IS_ERR(th))
756                 RETURN(PTR_ERR(th));
757         rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
758         if (rc)
759                 GOTO(cleanup, rc);
760         rc = dt_declare_attr_set(env, dt, &attr, th);
761         if (rc)
762                 GOTO(cleanup, rc);
763         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
764         if (rc)
765                 GOTO(cleanup, rc);
766
767         rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
768         if (rc == 0)
769                 rc = dt_attr_set(env, dt, &attr, th);
770
771 cleanup:
772         dt_trans_stop(env, tgt->lut_bottom, th);
773
774         RETURN(rc);
775 }
776
777 static void tgt_client_epoch_update(const struct lu_env *env,
778                                     struct obd_export *exp)
779 {
780         struct lsd_client_data  *lcd = exp->exp_target_data.ted_lcd;
781         struct lu_target        *tgt = class_exp2tgt(exp);
782
783         LASSERT(tgt && tgt->lut_bottom);
784         /** VBR: set client last_epoch to current epoch */
785         if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
786                 return;
787         lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
788         tgt_client_data_update(env, exp);
789 }
790
791 /**
792  * Update boot epoch when recovery ends
793  */
794 void tgt_boot_epoch_update(struct lu_target *tgt)
795 {
796         struct lu_env            env;
797         struct ptlrpc_request   *req;
798         __u32                    start_epoch;
799         LIST_HEAD(client_list);
800         int                      rc;
801
802         if (tgt->lut_obd->obd_stopping)
803                 return;
804
805         rc = lu_env_init(&env, LCT_DT_THREAD);
806         if (rc) {
807                 CERROR("%s: can't initialize environment: rc = %d\n",
808                         tgt->lut_obd->obd_name, rc);
809                 return;
810         }
811
812         spin_lock(&tgt->lut_translock);
813         start_epoch = (tgt->lut_last_transno >> LR_EPOCH_BITS) + 1;
814         tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
815         tgt->lut_lsd.lsd_start_epoch = start_epoch;
816         spin_unlock(&tgt->lut_translock);
817
818         /**
819          * The recovery is not yet finished and final queue can still be updated
820          * with resend requests. Move final list to separate one for processing
821          */
822         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
823         list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
824         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
825
826         /**
827          * go through list of exports participated in recovery and
828          * set new epoch for them
829          */
830         list_for_each_entry(req, &client_list, rq_list) {
831                 LASSERT(!req->rq_export->exp_delayed);
832                 if (!req->rq_export->exp_vbr_failed)
833                         tgt_client_epoch_update(&env, req->rq_export);
834         }
835         /** return list back at once */
836         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
837         list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
838         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
839
840         /** Clear MULTI RPCS incompatibility flag if
841          * - target is MDT and
842          * - there is no client to recover or the recovery was aborted
843          */
844         if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
845             (atomic_read(&tgt->lut_obd->obd_max_recoverable_clients) == 0 ||
846             tgt->lut_obd->obd_abort_recovery))
847                 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
848
849         /** update server epoch */
850         tgt_server_data_update(&env, tgt, 1);
851         lu_env_fini(&env);
852 }
853
854 /**
855  * commit callback, need to update last_committed value
856  */
857 struct tgt_last_committed_callback {
858         struct dt_txn_commit_cb  llcc_cb;
859         struct lu_target        *llcc_tgt;
860         struct obd_export       *llcc_exp;
861         __u64                    llcc_transno;
862 };
863
864 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
865                                   struct dt_txn_commit_cb *cb, int err)
866 {
867         struct tgt_last_committed_callback *ccb;
868
869         ccb = container_of(cb, struct tgt_last_committed_callback, llcc_cb);
870
871         LASSERT(ccb->llcc_exp);
872         LASSERT(ccb->llcc_tgt != NULL);
873         LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
874
875         if (th->th_reserved_quota.lqi_space > 0) {
876                 CDEBUG(D_QUOTA, "free quota %llu %llu\n",
877                        th->th_reserved_quota.lqi_id.qid_gid,
878                        th->th_reserved_quota.lqi_space);
879
880                 /* env can be NULL for freeing reserved quota */
881                 th->th_reserved_quota.lqi_space *= -1;
882                 dt_reserve_or_free_quota(NULL, th->th_dev,
883                                          &th->th_reserved_quota);
884         }
885
886         /* error hit, don't update last committed to provide chance to
887          * replay data after fail */
888         if (err != 0)
889                 goto out;
890
891         /* Fast path w/o spinlock, if exp_last_committed was updated
892          * with higher transno, no need to take spinlock and check,
893          * also no need to update obd_last_committed. */
894         if (ccb->llcc_transno <= ccb->llcc_exp->exp_last_committed)
895                 goto out;
896         spin_lock(&ccb->llcc_tgt->lut_translock);
897         if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
898                 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
899
900         if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
901                 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
902                 spin_unlock(&ccb->llcc_tgt->lut_translock);
903
904                 ptlrpc_commit_replies(ccb->llcc_exp);
905                 tgt_cancel_slc_locks(ccb->llcc_tgt, ccb->llcc_transno);
906         } else {
907                 spin_unlock(&ccb->llcc_tgt->lut_translock);
908         }
909
910         CDEBUG(D_HA, "%s: transno %lld is committed\n",
911                ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
912
913 out:
914         class_export_cb_put(ccb->llcc_exp);
915         OBD_FREE_PTR(ccb);
916 }
917
918 /**
919  * Add commit callback function, it returns a non-zero value to inform
920  * caller to use sync transaction if necessary.
921  */
922 static int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
923                                   struct obd_export *exp, __u64 transno)
924 {
925         struct tgt_last_committed_callback      *ccb;
926         struct dt_txn_commit_cb                 *dcb;
927         int                                      rc;
928
929         OBD_ALLOC_PTR(ccb);
930         if (ccb == NULL)
931                 return -ENOMEM;
932
933         ccb->llcc_tgt = tgt;
934         ccb->llcc_exp = class_export_cb_get(exp);
935         ccb->llcc_transno = transno;
936
937         dcb = &ccb->llcc_cb;
938         dcb->dcb_func = tgt_cb_last_committed;
939         INIT_LIST_HEAD(&dcb->dcb_linkage);
940         strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
941
942         rc = dt_trans_cb_add(th, dcb);
943         if (rc) {
944                 class_export_cb_put(exp);
945                 OBD_FREE_PTR(ccb);
946         }
947
948         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
949                 /* report failure to force synchronous operation */
950                 return -EPERM;
951
952         /* if exp_need_sync is set, return non-zero value to force
953          * a sync transaction. */
954         return rc ? rc : exp->exp_need_sync;
955 }
956
957 static int tgt_is_local_client(const struct lu_env *env,
958                                       struct obd_export *exp)
959 {
960         struct lu_target        *tgt = class_exp2tgt(exp);
961         struct tgt_session_info *tsi = tgt_ses_info(env);
962         struct ptlrpc_request   *req = tgt_ses_req(tsi);
963         struct lnet_nid nid;
964
965         if (exp_connect_flags(exp) & OBD_CONNECT_MDS ||
966             exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS)
967                 return 0;
968         if (tgt->lut_local_recovery)
969                 return 0;
970         if (!req)
971                 return 0;
972         lnet_nid4_to_nid(req->rq_peer.nid, &nid);
973         if (!LNetIsPeerLocal(&nid))
974                 return 0;
975
976         return 1;
977 }
978
979 /**
980  * Add new client to the last_rcvd upon new connection.
981  *
982  * We use a bitmap to locate a free space in the last_rcvd file and initialize
983  * tg_export_data.
984  */
985 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
986 {
987         struct tg_export_data   *ted = &exp->exp_target_data;
988         struct lu_target        *tgt = class_exp2tgt(exp);
989         int                      rc = 0, idx;
990
991         ENTRY;
992
993         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
994         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
995                 RETURN(0);
996
997         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
998                 RETURN(0);
999
1000         if (tgt_is_local_client(env, exp)) {
1001                 LCONSOLE_WARN("%s: local client %s w/o recovery\n",
1002                               exp->exp_obd->obd_name, ted->ted_lcd->lcd_uuid);
1003                 exp->exp_no_recovery = 1;
1004                 RETURN(0);
1005         }
1006
1007         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
1008          * there's no need for extra complication here
1009          */
1010         idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
1011 repeat:
1012         if (idx >= LR_MAX_CLIENTS ||
1013             OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
1014                 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
1015                        tgt->lut_obd->obd_name,  idx);
1016                 RETURN(-EOVERFLOW);
1017         }
1018         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
1019                 idx = find_next_zero_bit(tgt->lut_client_bitmap,
1020                                              LR_MAX_CLIENTS, idx);
1021                 goto repeat;
1022         }
1023
1024         ted->ted_lr_idx = idx;
1025         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1026                           idx * tgt->lut_lsd.lsd_client_size;
1027
1028         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1029
1030         if (tgt_is_multimodrpcs_client(exp)) {
1031                 /* Set MULTI RPCS incompatibility flag to prevent previous
1032                  * Lustre versions to mount a target with reply_data file */
1033                 atomic_inc(&tgt->lut_num_clients);
1034                 if (!(tgt->lut_lsd.lsd_feature_incompat &
1035                       OBD_INCOMPAT_MULTI_RPCS)) {
1036                         tgt->lut_lsd.lsd_feature_incompat |=
1037                                                         OBD_INCOMPAT_MULTI_RPCS;
1038                         rc = tgt_server_data_update(env, tgt, 1);
1039                         if (rc < 0) {
1040                                 CERROR("%s: unable to set MULTI RPCS "
1041                                        "incompatibility flag\n",
1042                                        exp->exp_obd->obd_name);
1043                                 RETURN(rc);
1044                         }
1045                 }
1046
1047                 /* assign client slot generation */
1048                 ted->ted_lcd->lcd_generation =
1049                                 atomic_inc_return(&tgt->lut_client_generation);
1050         } else {
1051                 ted->ted_lcd->lcd_generation = 0;
1052         }
1053
1054         CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
1055                "generation %d\n",
1056                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1057                ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
1058
1059         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
1060                 RETURN(-ENOSPC);
1061
1062         rc = tgt_client_data_update(env, exp);
1063         if (rc)
1064                 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
1065                        tgt->lut_obd->obd_name, idx, rc);
1066
1067         RETURN(rc);
1068 }
1069 EXPORT_SYMBOL(tgt_client_new);
1070
1071 /* Add an existing client to the MDS in-memory state based on
1072  * a client that was previously found in the last_rcvd file and
1073  * already has an assigned slot (idx >= 0).
1074  *
1075  * It should not be possible to fail adding an existing client - otherwise
1076  * mdt_init_server_data() callsite needs to be fixed.
1077  */
1078 int tgt_client_add(const struct lu_env *env,  struct obd_export *exp, int idx)
1079 {
1080         struct tg_export_data   *ted = &exp->exp_target_data;
1081         struct lu_target        *tgt = class_exp2tgt(exp);
1082
1083         ENTRY;
1084
1085         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
1086         LASSERTF(idx >= 0, "%d\n", idx);
1087
1088         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
1089             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1090                 RETURN(0);
1091
1092         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
1093                 CERROR("%s: client %d: bit already set in bitmap!!\n",
1094                        tgt->lut_obd->obd_name,  idx);
1095                 LBUG();
1096         }
1097         atomic_inc(&tgt->lut_num_clients);
1098
1099         CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
1100                "generation %d\n",
1101                tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
1102                ted->ted_lcd->lcd_generation);
1103
1104         ted->ted_lr_idx = idx;
1105         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1106                           idx * tgt->lut_lsd.lsd_client_size;
1107
1108         mutex_init(&ted->ted_lcd_lock);
1109
1110         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1111
1112         RETURN(0);
1113 }
1114
1115 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1116 {
1117         struct tg_export_data   *ted = &exp->exp_target_data;
1118         struct lu_target        *tgt = class_exp2tgt(exp);
1119         int                      rc;
1120
1121         ENTRY;
1122
1123         LASSERT(ted->ted_lcd);
1124
1125         if (unlikely(tgt == NULL)) {
1126                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1127                        class_exp2obd(exp)->obd_name);
1128                 RETURN(-EINVAL);
1129         }
1130
1131         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1132         if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1133                     (char *)tgt->lut_obd->obd_uuid.uuid) ||
1134             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT ||
1135             exp->exp_no_recovery)
1136                 RETURN(0);
1137
1138         /* Slot may be not yet assigned, use case is race between Client
1139          * reconnect and forced eviction */
1140         if (ted->ted_lr_idx < 0) {
1141                 CWARN("%s: client with UUID '%s' not in bitmap\n",
1142                       tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid);
1143                 RETURN(0);
1144         }
1145
1146         CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1147                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1148                ted->ted_lcd->lcd_uuid);
1149
1150         /* Clear the bit _after_ zeroing out the client so we don't
1151            race with filter_client_add and zero out new clients.*/
1152         if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1153                 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1154                        tgt->lut_obd->obd_name, ted->ted_lr_idx);
1155                 LBUG();
1156         }
1157
1158         /* Do not erase record for recoverable client. */
1159         if (exp->exp_flags & OBD_OPT_FAILOVER)
1160                 RETURN(0);
1161
1162         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_DEL))
1163                 RETURN(0);
1164
1165         /* Make sure the server's last_transno is up to date.
1166          * This should be done before zeroing client slot so last_transno will
1167          * be in server data or in client data in case of failure */
1168         rc = tgt_server_data_update(env, tgt, 0);
1169         if (rc != 0) {
1170                 CERROR("%s: failed to update server data, skip client %s "
1171                        "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1172                        ted->ted_lcd->lcd_uuid, rc);
1173                 RETURN(rc);
1174         }
1175
1176         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1177         rc = tgt_client_data_update(env, exp);
1178
1179         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1180                "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1181                tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1182                ted->ted_lr_idx, ted->ted_lr_off, rc);
1183         RETURN(rc);
1184 }
1185 EXPORT_SYMBOL(tgt_client_del);
1186
1187 static void tgt_clean_by_tag(struct obd_export *exp, __u64 xid, __u16 tag)
1188 {
1189         struct tg_export_data   *ted = &exp->exp_target_data;
1190         struct lu_target        *lut = class_exp2tgt(exp);
1191         struct tg_reply_data    *trd, *tmp;
1192
1193         if (tag == 0)
1194                 return;
1195
1196         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1197                 if (trd->trd_tag != tag)
1198                         continue;
1199
1200                 LASSERT(ergo(tgt_is_increasing_xid_client(exp),
1201                              trd->trd_reply.lrd_xid <= xid));
1202
1203                 ted->ted_release_tag++;
1204                 tgt_release_reply_data(lut, ted, trd);
1205         }
1206 }
1207
1208 static int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1209                        struct tg_export_data *ted, struct tg_reply_data *trd,
1210                        struct ptlrpc_request *req,
1211                        struct thandle *th, bool update_lrd_file)
1212 {
1213         struct lsd_reply_data   *lrd;
1214         int     i;
1215         int     rc;
1216
1217         lrd = &trd->trd_reply;
1218         /* update export last transno */
1219         mutex_lock(&ted->ted_lcd_lock);
1220         if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1221                 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1222         mutex_unlock(&ted->ted_lcd_lock);
1223
1224         if (tgt != NULL) {
1225                 /* find a empty slot */
1226                 i = tgt_find_free_reply_slot(tgt);
1227                 if (unlikely(i < 0)) {
1228                         CERROR("%s: couldn't find a slot for reply data: "
1229                                "rc = %d\n", tgt_name(tgt), i);
1230                         RETURN(i);
1231                 }
1232                 trd->trd_index = i;
1233
1234                 if (update_lrd_file) {
1235                         loff_t  off;
1236
1237                         /* write reply data to disk */
1238                         off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1239                         rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1240                         if (unlikely(rc != 0)) {
1241                                 CERROR("%s: can't update %s file: rc = %d\n",
1242                                        tgt_name(tgt), REPLY_DATA, rc);
1243                                 GOTO(free_slot, rc);
1244                         }
1245                 }
1246         } else {
1247                 trd->trd_index = TRD_INDEX_MEMORY;
1248         }
1249
1250         /* add reply data to target export's reply list */
1251         mutex_lock(&ted->ted_lcd_lock);
1252         if (req != NULL) {
1253                 int exclude = tgt_is_increasing_xid_client(req->rq_export) ?
1254                               MSG_REPLAY : MSG_REPLAY|MSG_RESENT;
1255
1256                 if (req->rq_obsolete) {
1257                         CDEBUG(D_INFO,
1258                                "drop reply data update for obsolete req xid=%llu,"
1259                                "transno=%llu, tag=%hu\n", req->rq_xid,
1260                                lrd->lrd_transno, trd->trd_tag);
1261                         mutex_unlock(&ted->ted_lcd_lock);
1262                         GOTO(free_slot, rc = -EBADR);
1263                 }
1264
1265                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & exclude))
1266                         tgt_clean_by_tag(req->rq_export, req->rq_xid,
1267                                          trd->trd_tag);
1268         }
1269         list_add(&trd->trd_list, &ted->ted_reply_list);
1270         ted->ted_reply_cnt++;
1271         if (ted->ted_reply_cnt > ted->ted_reply_max)
1272                 ted->ted_reply_max = ted->ted_reply_cnt;
1273         mutex_unlock(&ted->ted_lcd_lock);
1274
1275         CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1276                "tag %hu, client gen %u, slot idx %d\n",
1277                trd, lrd->lrd_xid, lrd->lrd_transno,
1278                trd->trd_tag, lrd->lrd_client_gen, trd->trd_index);
1279
1280         RETURN(0);
1281
1282 free_slot:
1283         if (tgt != NULL)
1284                 tgt_clear_reply_slot(tgt, trd->trd_index);
1285         return rc;
1286 }
1287
1288 int tgt_mk_reply_data(const struct lu_env *env,
1289                       struct lu_target *tgt,
1290                       struct tg_export_data *ted,
1291                       struct ptlrpc_request *req,
1292                       __u64 opdata,
1293                       struct thandle *th,
1294                       bool write_update,
1295                       __u64 transno)
1296 {
1297         struct tg_reply_data    *trd;
1298         struct lsd_reply_data   *lrd;
1299         __u64                   *pre_versions = NULL;
1300         int                     rc;
1301         struct tgt_session_info *tsi = NULL;
1302
1303         OBD_ALLOC_PTR(trd);
1304         if (unlikely(trd == NULL))
1305                 RETURN(-ENOMEM);
1306
1307         if (env != NULL)
1308                 tsi = tgt_ses_info(env);
1309
1310         /* fill reply data information */
1311         lrd = &trd->trd_reply;
1312         lrd->lrd_transno = transno;
1313         if (req != NULL) {
1314                 lrd->lrd_xid = req->rq_xid;
1315                 trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1316                 lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1317                 if (write_update) {
1318                         pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1319                         lrd->lrd_result = th->th_result;
1320                 }
1321         } else {
1322                 LASSERT(env != NULL);
1323                 LASSERT(tsi->tsi_xid != 0);
1324
1325                 lrd->lrd_xid = tsi->tsi_xid;
1326                 lrd->lrd_result = tsi->tsi_result;
1327                 lrd->lrd_client_gen = tsi->tsi_client_gen;
1328         }
1329
1330         lrd->lrd_data = opdata;
1331         if (pre_versions) {
1332                 trd->trd_pre_versions[0] = pre_versions[0];
1333                 trd->trd_pre_versions[1] = pre_versions[1];
1334                 trd->trd_pre_versions[2] = pre_versions[2];
1335                 trd->trd_pre_versions[3] = pre_versions[3];
1336         }
1337
1338         if (tsi && tsi->tsi_open_obj)
1339                 trd->trd_object = *lu_object_fid(&tsi->tsi_open_obj->do_lu);
1340
1341         rc = tgt_add_reply_data(env, tgt, ted, trd, req,
1342                                 th, write_update);
1343         if (rc < 0) {
1344                 OBD_FREE_PTR(trd);
1345                 if (rc == -EBADR)
1346                         rc = 0;
1347         }
1348         return rc;
1349
1350 }
1351 EXPORT_SYMBOL(tgt_mk_reply_data);
1352
1353 /*
1354  * last_rcvd & last_committed update callbacks
1355  */
1356 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1357                                 struct dt_object *obj, __u64 opdata,
1358                                 struct thandle *th, struct ptlrpc_request *req)
1359 {
1360         struct tgt_thread_info  *tti = tgt_th_info(env);
1361         struct tgt_session_info *tsi = tgt_ses_info(env);
1362         struct obd_export *exp = tsi->tsi_exp;
1363         struct tg_export_data *ted;
1364         __u64 *transno_p;
1365         bool nolcd = false;
1366         int rc = 0;
1367
1368         ENTRY;
1369
1370
1371         LASSERT(exp != NULL);
1372         ted = &exp->exp_target_data;
1373
1374         /* Some clients don't support recovery, and they don't have last_rcvd
1375          * client data:
1376          * 1. lightweight clients.
1377          * 2. local clients on MDS which doesn't enable "localrecov".
1378          * 3. OFD connect may cause transaction before export has last_rcvd
1379          *    slot.
1380          */
1381         if (ted->ted_lr_idx < 0)
1382                 nolcd = true;
1383
1384         if (req != NULL)
1385                 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1386         else
1387                 /* From update replay, tti_transno should be set already */
1388                 LASSERT(tti->tti_transno != 0);
1389
1390         spin_lock(&tgt->lut_translock);
1391         if (th->th_result != 0) {
1392                 if (tti->tti_transno != 0) {
1393                         CERROR("%s: replay transno %llu failed: rc = %d\n",
1394                                tgt_name(tgt), tti->tti_transno, th->th_result);
1395                 }
1396         } else if (tti->tti_transno == 0) {
1397                 tti->tti_transno = ++tgt->lut_last_transno;
1398         } else {
1399                 /* should be replay */
1400                 if (tti->tti_transno > tgt->lut_last_transno)
1401                         tgt->lut_last_transno = tti->tti_transno;
1402         }
1403         spin_unlock(&tgt->lut_translock);
1404
1405         /** VBR: set new versions */
1406         if (th->th_result == 0 && obj != NULL) {
1407                 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1408                 dt_version_set(env, dto, tti->tti_transno, th);
1409         }
1410
1411         /* filling reply data */
1412         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
1413                tti->tti_transno, tgt->lut_obd->obd_last_committed);
1414
1415         if (req != NULL) {
1416                 req->rq_transno = tti->tti_transno;
1417                 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1418         }
1419
1420         /* if can't add callback, do sync write */
1421         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1422
1423         if (nolcd) {
1424                 /* store transno in the last_rcvd header */
1425                 spin_lock(&tgt->lut_translock);
1426                 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1427                         tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1428                         spin_unlock(&tgt->lut_translock);
1429                         /* Although current connection doesn't have slot
1430                          * in the last_rcvd, we still want to maintain
1431                          * the in-memory lsd_client_data structure in order to
1432                          * properly handle reply reconstruction. */
1433                         rc = tgt_server_data_write(env, tgt, th);
1434                 } else {
1435                         spin_unlock(&tgt->lut_translock);
1436                 }
1437         } else if (ted->ted_lr_off == 0) {
1438                 CERROR("%s: client idx %d has offset %lld\n",
1439                        tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1440                 RETURN(-EINVAL);
1441         }
1442
1443         /* Target that supports multiple reply data */
1444         if (tgt_is_multimodrpcs_client(exp)) {
1445                 return tgt_mk_reply_data(env, tgt, ted, req, opdata, th,
1446                                          !!(req != NULL), tti->tti_transno);
1447         }
1448
1449         /* Enough for update replay, let's return */
1450         if (req == NULL)
1451                 RETURN(rc);
1452
1453         mutex_lock(&ted->ted_lcd_lock);
1454         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1455         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1456                 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1457                 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1458                 ted->ted_lcd->lcd_last_close_result = th->th_result;
1459         } else {
1460                 /* VBR: save versions in last_rcvd for reconstruct. */
1461                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1462
1463                 if (pre_versions) {
1464                         ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1465                         ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1466                         ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1467                         ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1468                 }
1469                 transno_p = &ted->ted_lcd->lcd_last_transno;
1470                 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1471                 ted->ted_lcd->lcd_last_result = th->th_result;
1472                 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1473                  * see struct ldlm_reply->lock_policy_res1; */
1474                 ted->ted_lcd->lcd_last_data = opdata;
1475         }
1476
1477         /* Update transno in slot only if non-zero number, i.e. no errors */
1478         if (likely(tti->tti_transno != 0)) {
1479                 /* Don't overwrite bigger transaction number with lower one.
1480                  * That is not sign of problem in all cases, but in any case
1481                  * this value should be monotonically increased only. */
1482                 if (*transno_p > tti->tti_transno) {
1483                         if (!tgt->lut_no_reconstruct) {
1484                                 CERROR("%s: trying to overwrite bigger transno:"
1485                                        "on-disk: %llu, new: %llu replay: "
1486                                        "%d. See LU-617.\n", tgt_name(tgt),
1487                                        *transno_p, tti->tti_transno,
1488                                        req_is_replay(req));
1489                                 if (req_is_replay(req)) {
1490                                         spin_lock(&req->rq_export->exp_lock);
1491                                         req->rq_export->exp_vbr_failed = 1;
1492                                         spin_unlock(&req->rq_export->exp_lock);
1493                                 }
1494                                 mutex_unlock(&ted->ted_lcd_lock);
1495                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1496                         }
1497                 } else {
1498                         *transno_p = tti->tti_transno;
1499                 }
1500         }
1501
1502         if (!nolcd) {
1503                 tti->tti_off = ted->ted_lr_off;
1504                 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_RCVD_EIO))
1505                         rc = -EIO;
1506                 else
1507                         rc = tgt_client_data_write(env, tgt, ted->ted_lcd,
1508                                                    &tti->tti_off, th);
1509                 if (rc < 0) {
1510                         mutex_unlock(&ted->ted_lcd_lock);
1511                         RETURN(rc);
1512                 }
1513         }
1514         mutex_unlock(&ted->ted_lcd_lock);
1515         RETURN(rc);
1516 }
1517
1518 /*
1519  * last_rcvd update for echo client simulation.
1520  * It updates last_rcvd client slot and version of object in
1521  * simple way but with all locks to simulate all drawbacks
1522  */
1523 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1524                                      struct lu_target *tgt,
1525                                      struct dt_object *obj,
1526                                      struct thandle *th,
1527                                      struct obd_export *exp)
1528 {
1529         struct tgt_thread_info  *tti = tgt_th_info(env);
1530         struct tg_export_data   *ted = &exp->exp_target_data;
1531         int                      rc = 0;
1532
1533         ENTRY;
1534
1535         tti->tti_transno = 0;
1536
1537         spin_lock(&tgt->lut_translock);
1538         if (th->th_result == 0)
1539                 tti->tti_transno = ++tgt->lut_last_transno;
1540         spin_unlock(&tgt->lut_translock);
1541
1542         /** VBR: set new versions */
1543         if (th->th_result == 0 && obj != NULL)
1544                 dt_version_set(env, obj, tti->tti_transno, th);
1545
1546         /* if can't add callback, do sync write */
1547         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1548                                                 tti->tti_transno);
1549
1550         LASSERT(ted->ted_lr_off > 0);
1551
1552         mutex_lock(&ted->ted_lcd_lock);
1553         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1554         ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1555         ted->ted_lcd->lcd_last_result = th->th_result;
1556
1557         tti->tti_off = ted->ted_lr_off;
1558         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1559         mutex_unlock(&ted->ted_lcd_lock);
1560         RETURN(rc);
1561 }
1562
1563 static int tgt_clients_data_init(const struct lu_env *env,
1564                                  struct lu_target *tgt,
1565                                  unsigned long last_size)
1566 {
1567         struct obd_device       *obd = tgt->lut_obd;
1568         struct lr_server_data   *lsd = &tgt->lut_lsd;
1569         struct lsd_client_data  *lcd = NULL;
1570         struct tg_export_data   *ted;
1571         int                      cl_idx;
1572         int                      rc = 0;
1573         loff_t                   off = lsd->lsd_client_start;
1574         __u32                    generation = 0;
1575         struct cfs_hash         *hash = NULL;
1576
1577         ENTRY;
1578
1579         if (tgt->lut_bottom->dd_rdonly)
1580                 RETURN(0);
1581
1582         BUILD_BUG_ON(offsetof(struct lsd_client_data, lcd_padding) +
1583                      sizeof(lcd->lcd_padding) != LR_CLIENT_SIZE);
1584
1585         OBD_ALLOC_PTR(lcd);
1586         if (lcd == NULL)
1587                 RETURN(-ENOMEM);
1588
1589         hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1590         if (hash == NULL)
1591                 GOTO(err_out, rc = -ENODEV);
1592
1593         for (cl_idx = 0; off < last_size; cl_idx++) {
1594                 struct obd_export       *exp;
1595                 __u64                    last_transno;
1596
1597                 /* Don't assume off is incremented properly by
1598                  * read_record(), in case sizeof(*lcd)
1599                  * isn't the same as fsd->lsd_client_size.  */
1600                 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1601                 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1602                 if (rc) {
1603                         CERROR("%s: error reading last_rcvd %s idx %d off "
1604                                "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1605                                cl_idx, off, rc);
1606                         rc = 0;
1607                         break; /* read error shouldn't cause startup to fail */
1608                 }
1609
1610                 if (lcd->lcd_uuid[0] == '\0') {
1611                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1612                                cl_idx);
1613                         continue;
1614                 }
1615
1616                 last_transno = lcd_last_transno(lcd);
1617
1618                 /* These exports are cleaned up by disconnect, so they
1619                  * need to be set up like real exports as connect does.
1620                  */
1621                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: %llu"
1622                        " srv lr: %llu lx: %llu gen %u\n", lcd->lcd_uuid,
1623                        cl_idx, last_transno, lsd->lsd_last_transno,
1624                        lcd_last_xid(lcd), lcd->lcd_generation);
1625
1626                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1627                 if (IS_ERR(exp)) {
1628                         if (PTR_ERR(exp) == -EALREADY) {
1629                                 /* export already exists, zero out this one */
1630                                 CERROR("%s: Duplicate export %s!\n",
1631                                        tgt_name(tgt), lcd->lcd_uuid);
1632                                 continue;
1633                         }
1634                         GOTO(err_out, rc = PTR_ERR(exp));
1635                 }
1636
1637                 ted = &exp->exp_target_data;
1638                 *ted->ted_lcd = *lcd;
1639
1640                 rc = tgt_client_add(env, exp, cl_idx);
1641                 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1642                 /* VBR: set export last committed version */
1643                 exp->exp_last_committed = last_transno;
1644                 spin_lock(&exp->exp_lock);
1645                 exp->exp_connecting = 0;
1646                 exp->exp_in_recovery = 0;
1647                 spin_unlock(&exp->exp_lock);
1648                 atomic_inc(&obd->obd_max_recoverable_clients);
1649
1650                 if (tgt->lut_lsd.lsd_feature_incompat &
1651                     OBD_INCOMPAT_MULTI_RPCS &&
1652                     lcd->lcd_generation != 0) {
1653                         /* compute the highest valid client generation */
1654                         generation = max(generation, lcd->lcd_generation);
1655                         /* fill client_generation <-> export hash table */
1656                         rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1657                                                  &exp->exp_gen_hash);
1658                         if (rc != 0) {
1659                                 CERROR("%s: duplicate export for client "
1660                                        "generation %u\n",
1661                                        tgt_name(tgt), lcd->lcd_generation);
1662                                 class_export_put(exp);
1663                                 GOTO(err_out, rc);
1664                         }
1665                 }
1666
1667                 class_export_put(exp);
1668
1669                 rc = rev_import_init(exp);
1670                 if (rc != 0) {
1671                         class_unlink_export(exp);
1672                         GOTO(err_out, rc);
1673                 }
1674
1675                 /* Need to check last_rcvd even for duplicated exports. */
1676                 CDEBUG(D_OTHER, "client at idx %d has last_transno = %llu\n",
1677                        cl_idx, last_transno);
1678
1679                 spin_lock(&tgt->lut_translock);
1680                 tgt->lut_last_transno = max(last_transno,
1681                                             tgt->lut_last_transno);
1682                 spin_unlock(&tgt->lut_translock);
1683         }
1684
1685         /* record highest valid client generation */
1686         atomic_set(&tgt->lut_client_generation, generation);
1687
1688 err_out:
1689         if (hash != NULL)
1690                 cfs_hash_putref(hash);
1691         OBD_FREE_PTR(lcd);
1692         RETURN(rc);
1693 }
1694
1695 struct server_compat_data {
1696         __u32 rocompat;
1697         __u32 incompat;
1698         __u32 rocinit;
1699         __u32 incinit;
1700 };
1701
1702 static struct server_compat_data tgt_scd[] = {
1703         [LDD_F_SV_TYPE_MDT] = {
1704                 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1705                 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1706                             OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1707                             OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1708                             OBD_INCOMPAT_MULTI_RPCS,
1709                 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1710                 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1711                            OBD_INCOMPAT_MULTI_OI,
1712         },
1713         [LDD_F_SV_TYPE_OST] = {
1714                 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1715                 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1716                             OBD_INCOMPAT_FID,
1717                 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1718                 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1719         }
1720 };
1721
1722 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1723 {
1724         struct tgt_thread_info          *tti = tgt_th_info(env);
1725         struct lr_server_data           *lsd = &tgt->lut_lsd;
1726         unsigned long                    last_rcvd_size;
1727         __u32                            index;
1728         int                              rc, type;
1729
1730         rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1731         if (rc)
1732                 RETURN(rc);
1733
1734         last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1735
1736         /* ensure padding in the struct is the correct size */
1737         BUILD_BUG_ON(offsetof(struct lr_server_data, lsd_padding) +
1738                      sizeof(lsd->lsd_padding) != LR_SERVER_SIZE);
1739
1740         rc = server_name2index(tgt_name(tgt), &index, NULL);
1741         if (rc < 0) {
1742                 CERROR("%s: Can not get index from name: rc = %d\n",
1743                        tgt_name(tgt), rc);
1744                 RETURN(rc);
1745         }
1746         /* server_name2index() returns type */
1747         type = rc;
1748         if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1749                 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1750                 RETURN(-EINVAL);
1751         }
1752
1753         /* last_rcvd on OST doesn't provide reconstruct support because there
1754          * may be up to 8 in-flight write requests per single slot in
1755          * last_rcvd client data
1756          */
1757         tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1758
1759         if (last_rcvd_size == 0) {
1760                 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1761
1762                 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1763                        sizeof(lsd->lsd_uuid));
1764                 lsd->lsd_last_transno = 0;
1765                 lsd->lsd_mount_count = 0;
1766                 lsd->lsd_server_size = LR_SERVER_SIZE;
1767                 lsd->lsd_client_start = LR_CLIENT_START;
1768                 lsd->lsd_client_size = LR_CLIENT_SIZE;
1769                 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1770                 lsd->lsd_osd_index = index;
1771                 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1772                 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1773         } else {
1774                 rc = tgt_server_data_read(env, tgt);
1775                 if (rc) {
1776                         CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1777                                tgt_name(tgt), rc);
1778                         RETURN(rc);
1779                 }
1780                 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1781                         if (tgt->lut_bottom->dd_rdonly) {
1782                                 /* Such difference may be caused by mounting
1783                                  * up snapshot with new fsname under rd_only
1784                                  * mode. But even if it was NOT, it will not
1785                                  * damage the system because of "rd_only". */
1786                                 memcpy(lsd->lsd_uuid,
1787                                        tgt->lut_obd->obd_uuid.uuid,
1788                                        sizeof(lsd->lsd_uuid));
1789                         } else {
1790                                 LCONSOLE_ERROR_MSG(0x157, "Trying to start "
1791                                                    "OBD %s using the wrong "
1792                                                    "disk %s. Were the /dev/ "
1793                                                    "assignments rearranged?\n",
1794                                                    tgt->lut_obd->obd_uuid.uuid,
1795                                                    lsd->lsd_uuid);
1796                                 RETURN(-EINVAL);
1797                         }
1798                 }
1799
1800                 if (lsd->lsd_osd_index != index) {
1801                         LCONSOLE_ERROR_MSG(0x157,
1802                                            "%s: index %d in last rcvd is different with the index %d in config log, It might be disk corruption!\n",
1803                                            tgt_name(tgt),
1804                                            lsd->lsd_osd_index, index);
1805                         RETURN(-EINVAL);
1806                 }
1807         }
1808
1809         if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1810                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1811                        tgt_name(tgt),
1812                        lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1813                 RETURN(-EINVAL);
1814         }
1815
1816         if (type == LDD_F_SV_TYPE_MDT)
1817                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1818
1819         if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1820                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1821                        tgt_name(tgt),
1822                        lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1823                 RETURN(-EINVAL);
1824         }
1825         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1826         if (type == LDD_F_SV_TYPE_MDT &&
1827             !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1828                 if (last_rcvd_size > lsd->lsd_client_start) {
1829                         LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1830                                       "remove all clients for interop needs\n",
1831                                       tgt_name(tgt));
1832                         rc = tgt_truncate_last_rcvd(env, tgt,
1833                                                     lsd->lsd_client_start);
1834                         if (rc)
1835                                 RETURN(rc);
1836                         last_rcvd_size = lsd->lsd_client_start;
1837                 }
1838                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1839                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1840         }
1841
1842         spin_lock(&tgt->lut_translock);
1843         tgt->lut_last_transno = lsd->lsd_last_transno;
1844         spin_unlock(&tgt->lut_translock);
1845
1846         lsd->lsd_mount_count++;
1847
1848         CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1849         CDEBUG(D_INODE, "%s: server last_transno: %llu\n",
1850                tgt_name(tgt), tgt->lut_last_transno);
1851         CDEBUG(D_INODE, "%s: server mount_count: %llu\n",
1852                tgt_name(tgt), lsd->lsd_mount_count);
1853         CDEBUG(D_INODE, "%s: server data size: %u\n",
1854                tgt_name(tgt), lsd->lsd_server_size);
1855         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1856                tgt_name(tgt), lsd->lsd_client_start);
1857         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1858                tgt_name(tgt), lsd->lsd_client_size);
1859         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1860                tgt_name(tgt), last_rcvd_size);
1861         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1862                tgt_name(tgt), lsd->lsd_subdir_count);
1863         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1864                last_rcvd_size <= lsd->lsd_client_start ? 0 :
1865                (last_rcvd_size - lsd->lsd_client_start) /
1866                 lsd->lsd_client_size);
1867         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1868
1869         if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1870             lsd->lsd_client_size == 0) {
1871                 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1872                 RETURN(-EINVAL);
1873         }
1874
1875         if (!tgt->lut_obd->obd_replayable)
1876                 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1877
1878         rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1879         if (rc < 0)
1880                 GOTO(err_client, rc);
1881
1882         spin_lock(&tgt->lut_translock);
1883         /* obd_last_committed is used for compatibility
1884          * with other lustre recovery code */
1885         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1886         spin_unlock(&tgt->lut_translock);
1887
1888         tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1889         tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1890
1891         /* save it, so mount count and last_transno is current */
1892         rc = tgt_server_data_update(env, tgt, 0);
1893         if (rc < 0)
1894                 GOTO(err_client, rc);
1895
1896         RETURN(0);
1897
1898 err_client:
1899         class_disconnect_exports(tgt->lut_obd);
1900         return rc;
1901 }
1902
1903 /* add credits for last_rcvd update */
1904 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1905                      void *cookie)
1906 {
1907         struct lu_target        *tgt = cookie;
1908         struct tgt_session_info *tsi;
1909         struct tgt_thread_info  *tti = tgt_th_info(env);
1910         struct dt_object        *dto;
1911         int                      rc;
1912
1913         /* For readonly case, the caller should have got failure
1914          * when start the transaction. If the logic comes here,
1915          * there must be something wrong. */
1916         if (unlikely(tgt->lut_bottom->dd_rdonly)) {
1917                 dump_stack();
1918                 LBUG();
1919         }
1920
1921         /* if there is no session, then this transaction is not result of
1922          * request processing but some local operation */
1923         if (env->le_ses == NULL)
1924                 return 0;
1925
1926         LASSERT(tgt->lut_last_rcvd);
1927         tsi = tgt_ses_info(env);
1928         /* OFD may start transaction without export assigned */
1929         if (tsi->tsi_exp == NULL)
1930                 return 0;
1931
1932         if (tgt_is_multimodrpcs_client(tsi->tsi_exp)) {
1933                 /*
1934                  * Use maximum possible file offset for declaration to ensure
1935                  * ZFS will reserve enough credits for a write anywhere in this
1936                  * file, since we don't know where in the file the write will be
1937                  * because a replay slot has not been assigned.  This should be
1938                  * replaced by dmu_tx_hold_append() when available.
1939                  */
1940                 tti->tti_buf.lb_buf = NULL;
1941                 tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
1942                 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
1943                 rc = dt_declare_record_write(env, dto, &tti->tti_buf, -1, th);
1944                 if (rc)
1945                         return rc;
1946         } else {
1947                 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1948                 tti_buf_lcd(tti);
1949                 tti->tti_off = tsi->tsi_exp->exp_target_data.ted_lr_off;
1950                 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1951                                              tti->tti_off, th);
1952                 if (rc)
1953                         return rc;
1954         }
1955
1956         if (tsi->tsi_vbr_obj != NULL &&
1957             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1958                 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1959                 rc = dt_declare_version_set(env, dto, th);
1960         }
1961
1962         return rc;
1963 }
1964
1965 /* Update last_rcvd records with latests transaction data */
1966 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1967                     void *cookie)
1968 {
1969         struct lu_target        *tgt = cookie;
1970         struct tgt_session_info *tsi;
1971         struct tgt_thread_info  *tti = tgt_th_info(env);
1972         struct dt_object        *obj = NULL;
1973         int                      rc;
1974         bool                     echo_client;
1975
1976         if (env->le_ses == NULL)
1977                 return 0;
1978
1979         tsi = tgt_ses_info(env);
1980         /* OFD may start transaction without export assigned */
1981         if (tsi->tsi_exp == NULL)
1982                 return 0;
1983
1984         echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
1985
1986         if (tti->tti_has_trans && !echo_client) {
1987                 if (tti->tti_mult_trans == 0) {
1988                         CDEBUG(D_HA, "More than one transaction %llu\n",
1989                                tti->tti_transno);
1990                         RETURN(0);
1991                 }
1992                 /* we need another transno to be assigned */
1993                 tti->tti_transno = 0;
1994         } else if (th->th_result == 0) {
1995                 tti->tti_has_trans = 1;
1996         }
1997
1998         if (tsi->tsi_vbr_obj != NULL &&
1999             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
2000                 obj = tsi->tsi_vbr_obj;
2001         }
2002
2003         if (unlikely(echo_client)) /* echo client special case */
2004                 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
2005                                                tsi->tsi_exp);
2006         else
2007                 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
2008                                           tgt_ses_req(tsi));
2009         return rc;
2010 }
2011
2012 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
2013 {
2014         struct tgt_thread_info  *tti = tgt_th_info(env);
2015         struct lsd_reply_data   *lrd = &tti->tti_lrd;
2016         unsigned long            reply_data_size;
2017         int                      rc;
2018         struct lsd_reply_header *lrh = NULL;
2019         struct tg_reply_data    *trd = NULL;
2020         int                      idx;
2021         loff_t                   off;
2022         struct cfs_hash         *hash = NULL;
2023         struct obd_export       *exp;
2024         struct tg_export_data   *ted;
2025         int                      reply_data_recovered = 0;
2026
2027         rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
2028         if (rc)
2029                 GOTO(out, rc);
2030         reply_data_size = (unsigned long)tti->tti_attr.la_size;
2031
2032         OBD_ALLOC_PTR(lrh);
2033         if (lrh == NULL)
2034                 GOTO(out, rc = -ENOMEM);
2035
2036         if (reply_data_size == 0) {
2037                 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
2038                        tgt_name(tgt));
2039                 lrh->lrh_magic = LRH_MAGIC;
2040                 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
2041                 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
2042                 rc = tgt_reply_header_write(env, tgt, lrh);
2043                 if (rc) {
2044                         CERROR("%s: error writing %s: rc = %d\n",
2045                                tgt_name(tgt), REPLY_DATA, rc);
2046                         GOTO(out, rc);
2047                 }
2048         } else {
2049                 rc = tgt_reply_header_read(env, tgt, lrh);
2050                 if (rc) {
2051                         CERROR("%s: error reading %s: rc = %d\n",
2052                                tgt_name(tgt), REPLY_DATA, rc);
2053                         GOTO(out, rc);
2054                 }
2055                 if (lrh->lrh_magic != LRH_MAGIC ||
2056                     lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
2057                     lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
2058                         CERROR("%s: invalid header in %s\n",
2059                                tgt_name(tgt), REPLY_DATA);
2060                         GOTO(out, rc = -EINVAL);
2061                 }
2062
2063                 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
2064                 if (hash == NULL)
2065                         GOTO(out, rc = -ENODEV);
2066
2067                 OBD_ALLOC_PTR(trd);
2068                 if (trd == NULL)
2069                         GOTO(out, rc = -ENOMEM);
2070
2071                 /* Load reply_data from disk */
2072                 for (idx = 0, off = sizeof(struct lsd_reply_header);
2073                      off < reply_data_size;
2074                      idx++, off += sizeof(struct lsd_reply_data)) {
2075                         rc = tgt_reply_data_read(env, tgt, lrd, off);
2076                         if (rc) {
2077                                 CERROR("%s: error reading %s: rc = %d\n",
2078                                        tgt_name(tgt), REPLY_DATA, rc);
2079                                 GOTO(out, rc);
2080                         }
2081
2082                         exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
2083                         if (exp == NULL) {
2084                                 /* old reply data from a disconnected client */
2085                                 continue;
2086                         }
2087                         ted = &exp->exp_target_data;
2088                         mutex_lock(&ted->ted_lcd_lock);
2089
2090                         /* create in-memory reply_data and link it to
2091                          * target export's reply list */
2092                         rc = tgt_set_reply_slot(tgt, idx);
2093                         if (rc != 0) {
2094                                 mutex_unlock(&ted->ted_lcd_lock);
2095                                 GOTO(out, rc);
2096                         }
2097                         trd->trd_reply = *lrd;
2098                         trd->trd_pre_versions[0] = 0;
2099                         trd->trd_pre_versions[1] = 0;
2100                         trd->trd_pre_versions[2] = 0;
2101                         trd->trd_pre_versions[3] = 0;
2102                         trd->trd_index = idx;
2103                         trd->trd_tag = 0;
2104                         fid_zero(&trd->trd_object);
2105                         list_add(&trd->trd_list, &ted->ted_reply_list);
2106                         ted->ted_reply_cnt++;
2107                         if (ted->ted_reply_cnt > ted->ted_reply_max)
2108                                 ted->ted_reply_max = ted->ted_reply_cnt;
2109
2110                         CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
2111                                "transno %llu, client gen %u, slot idx %d\n",
2112                                tgt_name(tgt), trd, lrd->lrd_xid,
2113                                lrd->lrd_transno, lrd->lrd_client_gen,
2114                                trd->trd_index);
2115
2116                         /* update export last committed transation */
2117                         exp->exp_last_committed = max(exp->exp_last_committed,
2118                                                       lrd->lrd_transno);
2119                         /* Update lcd_last_transno as well for check in
2120                          * tgt_release_reply_data() or the latest client
2121                          * transno can be lost.
2122                          */
2123                         ted->ted_lcd->lcd_last_transno =
2124                                 max(ted->ted_lcd->lcd_last_transno,
2125                                     exp->exp_last_committed);
2126
2127                         mutex_unlock(&ted->ted_lcd_lock);
2128                         class_export_put(exp);
2129
2130                         /* update target last committed transaction */
2131                         spin_lock(&tgt->lut_translock);
2132                         tgt->lut_last_transno = max(tgt->lut_last_transno,
2133                                                     lrd->lrd_transno);
2134                         spin_unlock(&tgt->lut_translock);
2135
2136                         reply_data_recovered++;
2137
2138                         OBD_ALLOC_PTR(trd);
2139                         if (trd == NULL)
2140                                 GOTO(out, rc = -ENOMEM);
2141                 }
2142                 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
2143                        tgt_name(tgt), reply_data_recovered);
2144         }
2145
2146         spin_lock(&tgt->lut_translock);
2147         /* obd_last_committed is used for compatibility
2148          * with other lustre recovery code */
2149         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
2150         spin_unlock(&tgt->lut_translock);
2151
2152         rc = 0;
2153
2154 out:
2155         if (hash != NULL)
2156                 cfs_hash_putref(hash);
2157         if (trd != NULL)
2158                 OBD_FREE_PTR(trd);
2159         if (lrh != NULL)
2160                 OBD_FREE_PTR(lrh);
2161         return rc;
2162 }
2163
2164 static int tgt_check_lookup_req(struct ptlrpc_request *req, int lookup,
2165                                 struct tg_reply_data *trd)
2166 {
2167         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2168         struct lu_target *lut = class_exp2tgt(req->rq_export);
2169         __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
2170         int rc = 0;
2171         struct tg_reply_data *reply;
2172         bool check_increasing;
2173
2174         if (tag == 0)
2175                 return 0;
2176
2177         check_increasing = tgt_is_increasing_xid_client(req->rq_export) &&
2178                            !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
2179         if (!lookup && !check_increasing)
2180                 return 0;
2181
2182         list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
2183                 if (lookup && reply->trd_reply.lrd_xid == req->rq_xid) {
2184                         rc = 1;
2185                         if (trd != NULL)
2186                                 *trd = *reply;
2187                         break;
2188                 } else if (check_increasing && reply->trd_tag == tag &&
2189                            reply->trd_reply.lrd_xid > req->rq_xid) {
2190                         rc = -EPROTO;
2191                         CERROR("%s: busy tag=%u req_xid=%llu, trd=%p: xid=%llu transno=%llu client_gen=%u slot_idx=%d: rc = %d\n",
2192                                tgt_name(lut), tag, req->rq_xid, trd,
2193                                reply->trd_reply.lrd_xid,
2194                                reply->trd_reply.lrd_transno,
2195                                reply->trd_reply.lrd_client_gen,
2196                                reply->trd_index, rc);
2197                         break;
2198                 }
2199         }
2200
2201         return rc;
2202 }
2203
2204 /* Look for a reply data matching specified request @req
2205  * A copy is returned in @trd if the pointer is not NULL
2206  */
2207 int tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
2208 {
2209         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2210         int found = 0;
2211         bool not_replay = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
2212
2213         mutex_lock(&ted->ted_lcd_lock);
2214         if (not_replay && req->rq_xid <= req->rq_export->exp_last_xid) {
2215                 /* A check for the last_xid is needed here in case there is
2216                  * no reply data is left in the list. It may happen if another
2217                  * RPC on another slot increased the last_xid between our
2218                  * process_req_last_xid & tgt_lookup_reply calls */
2219                 found = -EPROTO;
2220         } else {
2221                 found = tgt_check_lookup_req(req, 1, trd);
2222         }
2223         mutex_unlock(&ted->ted_lcd_lock);
2224
2225         CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d last_xid %llu\n",
2226                tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid, found,
2227                req->rq_export->exp_last_xid);
2228
2229         return found;
2230 }
2231 EXPORT_SYMBOL(tgt_lookup_reply);
2232
2233 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
2234 {
2235         struct tg_export_data   *ted = &exp->exp_target_data;
2236         struct lu_target        *lut = class_exp2tgt(exp);
2237         struct tg_reply_data    *trd, *tmp;
2238
2239
2240         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2241                 if (trd->trd_reply.lrd_xid > rcvd_xid)
2242                         continue;
2243                 ted->ted_release_xid++;
2244                 tgt_release_reply_data(lut, ted, trd);
2245         }
2246
2247         return 0;
2248 }
2249
2250 int tgt_handle_tag(struct ptlrpc_request *req)
2251 {
2252         return tgt_check_lookup_req(req, 0, NULL);
2253 }
2254