Whamcloud - gitweb
LU-5371 lmv: Do not ignore ENOENT in lmv_unlink
[fs/lustre-release.git] / lustre / lmv / lmv_obd.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2013, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 #define DEBUG_SUBSYSTEM S_LMV
38 #include <linux/slab.h>
39 #include <linux/module.h>
40 #include <linux/init.h>
41 #include <linux/slab.h>
42 #include <linux/pagemap.h>
43 #include <linux/mm.h>
44 #include <linux/math64.h>
45 #include <linux/seq_file.h>
46 #include <linux/namei.h>
47
48 #include <lustre/lustre_idl.h>
49 #include <obd_support.h>
50 #include <lustre_lib.h>
51 #include <lustre_net.h>
52 #include <obd_class.h>
53 #include <lustre_lmv.h>
54 #include <lprocfs_status.h>
55 #include <cl_object.h>
56 #include <lclient.h>
57 #include <lustre_lite.h>
58 #include <lustre_fid.h>
59 #include <lustre_ioctl.h>
60 #include "lmv_internal.h"
61
62 /* This hash is only for testing purpose */
63 static inline unsigned int
64 lmv_hash_all_chars(unsigned int count, const char *name, int namelen)
65 {
66         unsigned int c = 0;
67         const unsigned char *p = (const unsigned char *)name;
68
69         while (--namelen >= 0)
70                 c += p[namelen];
71
72         c = c % count;
73
74         return c;
75 }
76
77 static inline unsigned int
78 lmv_hash_fnv1a(unsigned int count, const char *name, int namelen)
79 {
80         __u64   hash;
81
82         hash = lustre_hash_fnv_1a_64(name, namelen);
83
84         hash = hash % count;
85
86         return hash;
87 }
88
89 int lmv_name_to_stripe_index(__u32 lmv_hash_type, unsigned int stripe_count,
90                              const char *name, int namelen)
91 {
92         int     idx;
93         __u32   hash_type = lmv_hash_type & LMV_HASH_TYPE_MASK;
94
95         LASSERT(namelen > 0);
96         if (stripe_count <= 1)
97                 return 0;
98
99         /* for migrating object, always start from 0 stripe */
100         if (lmv_hash_type & LMV_HASH_FLAG_MIGRATION)
101                 return 0;
102
103         switch (hash_type) {
104         case LMV_HASH_TYPE_ALL_CHARS:
105                 idx = lmv_hash_all_chars(stripe_count, name, namelen);
106                 break;
107         case LMV_HASH_TYPE_FNV_1A_64:
108                 idx = lmv_hash_fnv1a(stripe_count, name, namelen);
109                 break;
110         default:
111                 idx = -EBADFD;
112                 break;
113         }
114
115         CDEBUG(D_INFO, "name %.*s hash_type %d idx %d\n", namelen, name,
116                hash_type, idx);
117
118         return idx;
119 }
120
121 static void lmv_activate_target(struct lmv_obd *lmv,
122                                 struct lmv_tgt_desc *tgt,
123                                 int activate)
124 {
125         if (tgt->ltd_active == activate)
126                 return;
127
128         tgt->ltd_active = activate;
129         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
130 }
131
132 /**
133  * Error codes:
134  *
135  *  -EINVAL  : UUID can't be found in the LMV's target list
136  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
137  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
138  */
139 static int lmv_set_mdc_active(struct lmv_obd *lmv,
140                               const struct obd_uuid *uuid,
141                               int activate)
142 {
143         struct lmv_tgt_desc     *tgt = NULL;
144         struct obd_device       *obd;
145         __u32                    i;
146         int                      rc = 0;
147         ENTRY;
148
149         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
150                         lmv, uuid->uuid, activate);
151
152         spin_lock(&lmv->lmv_lock);
153         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
154                 tgt = lmv->tgts[i];
155                 if (tgt == NULL || tgt->ltd_exp == NULL)
156                         continue;
157
158                 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n", i,
159                        tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
160
161                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
162                         break;
163         }
164
165         if (i == lmv->desc.ld_tgt_count)
166                 GOTO(out_lmv_lock, rc = -EINVAL);
167
168         obd = class_exp2obd(tgt->ltd_exp);
169         if (obd == NULL)
170                 GOTO(out_lmv_lock, rc = -ENOTCONN);
171
172         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
173                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
174                obd->obd_type->typ_name, i);
175         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
176
177         if (tgt->ltd_active == activate) {
178                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
179                        activate ? "" : "in");
180                 GOTO(out_lmv_lock, rc);
181         }
182
183         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
184                activate ? "" : "in");
185         lmv_activate_target(lmv, tgt, activate);
186         EXIT;
187
188  out_lmv_lock:
189         spin_unlock(&lmv->lmv_lock);
190         return rc;
191 }
192
193 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
194 {
195         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
196         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
197
198         return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
199 }
200
201 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
202                       enum obd_notify_event ev, void *data)
203 {
204         struct obd_connect_data *conn_data;
205         struct lmv_obd          *lmv = &obd->u.lmv;
206         struct obd_uuid         *uuid;
207         int                      rc = 0;
208         ENTRY;
209
210         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
211                 CERROR("unexpected notification of %s %s!\n",
212                        watched->obd_type->typ_name,
213                        watched->obd_name);
214                 RETURN(-EINVAL);
215         }
216
217         uuid = &watched->u.cli.cl_target_uuid;
218         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
219                 /*
220                  * Set MDC as active before notifying the observer, so the
221                  * observer can use the MDC normally.
222                  */
223                 rc = lmv_set_mdc_active(lmv, uuid,
224                                         ev == OBD_NOTIFY_ACTIVE);
225                 if (rc) {
226                         CERROR("%sactivation of %s failed: %d\n",
227                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
228                                uuid->uuid, rc);
229                         RETURN(rc);
230                 }
231         } else if (ev == OBD_NOTIFY_OCD) {
232                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
233                 /*
234                  * XXX: Make sure that ocd_connect_flags from all targets are
235                  * the same. Otherwise one of MDTs runs wrong version or
236                  * something like this.  --umka
237                  */
238                 obd->obd_self_export->exp_connect_data = *conn_data;
239         }
240 #if 0
241         else if (ev == OBD_NOTIFY_DISCON) {
242                 /*
243                  * For disconnect event, flush fld cache for failout MDS case.
244                  */
245                 fld_client_flush(&lmv->lmv_fld);
246         }
247 #endif
248         /*
249          * Pass the notification up the chain.
250          */
251         if (obd->obd_observer)
252                 rc = obd_notify(obd->obd_observer, watched, ev, data);
253
254         RETURN(rc);
255 }
256
257 /**
258  * This is fake connect function. Its purpose is to initialize lmv and say
259  * caller that everything is okay. Real connection will be performed later.
260  */
261 static int lmv_connect(const struct lu_env *env,
262                        struct obd_export **exp, struct obd_device *obd,
263                        struct obd_uuid *cluuid, struct obd_connect_data *data,
264                        void *localdata)
265 {
266         struct lmv_obd        *lmv = &obd->u.lmv;
267         struct lustre_handle  conn = { 0 };
268         int                    rc = 0;
269         ENTRY;
270
271         /*
272          * We don't want to actually do the underlying connections more than
273          * once, so keep track.
274          */
275         lmv->refcount++;
276         if (lmv->refcount > 1) {
277                 *exp = NULL;
278                 RETURN(0);
279         }
280
281         rc = class_connect(&conn, obd, cluuid);
282         if (rc) {
283                 CERROR("class_connection() returned %d\n", rc);
284                 RETURN(rc);
285         }
286
287         *exp = class_conn2export(&conn);
288         class_export_get(*exp);
289
290         lmv->exp = *exp;
291         lmv->connected = 0;
292         lmv->cluuid = *cluuid;
293
294         if (data)
295                 lmv->conn_data = *data;
296
297         if (lmv->targets_proc_entry == NULL) {
298                 lmv->targets_proc_entry = lprocfs_seq_register("target_obds",
299                                                         obd->obd_proc_entry,
300                                                         NULL, NULL);
301                 if (IS_ERR(lmv->targets_proc_entry)) {
302                         CERROR("%s: cannot register "
303                                "/proc/fs/lustre/%s/%s/target_obds\n",
304                                obd->obd_name, obd->obd_type->typ_name,
305                                obd->obd_name);
306                         lmv->targets_proc_entry = NULL;
307                 }
308         }
309
310         /*
311          * All real clients should perform actual connection right away, because
312          * it is possible, that LMV will not have opportunity to connect targets
313          * and MDC stuff will be called directly, for instance while reading
314          * ../mdc/../kbytesfree procfs file, etc.
315          */
316         if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_REAL))
317                 rc = lmv_check_connect(obd);
318
319         if (rc && lmv->targets_proc_entry != NULL)
320                 lprocfs_remove(&lmv->targets_proc_entry);
321         RETURN(rc);
322 }
323
324 static void lmv_set_timeouts(struct obd_device *obd)
325 {
326         struct lmv_obd          *lmv;
327         __u32                    i;
328
329         lmv = &obd->u.lmv;
330         if (lmv->server_timeout == 0)
331                 return;
332
333         if (lmv->connected == 0)
334                 return;
335
336         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
337                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
338
339                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
340                         continue;
341
342                 obd_set_info_async(NULL, tgt->ltd_exp, sizeof(KEY_INTERMDS),
343                                    KEY_INTERMDS, 0, NULL, NULL);
344         }
345 }
346
347 static int lmv_init_ea_size(struct obd_export *exp, int easize,
348                             int def_easize, int cookiesize, int def_cookiesize)
349 {
350         struct obd_device       *obd = exp->exp_obd;
351         struct lmv_obd          *lmv = &obd->u.lmv;
352         __u32                    i;
353         int                      rc = 0;
354         int                      change = 0;
355         ENTRY;
356
357         if (lmv->max_easize < easize) {
358                 lmv->max_easize = easize;
359                 change = 1;
360         }
361         if (lmv->max_def_easize < def_easize) {
362                 lmv->max_def_easize = def_easize;
363                 change = 1;
364         }
365         if (lmv->max_cookiesize < cookiesize) {
366                 lmv->max_cookiesize = cookiesize;
367                 change = 1;
368         }
369         if (lmv->max_def_cookiesize < def_cookiesize) {
370                 lmv->max_def_cookiesize = def_cookiesize;
371                 change = 1;
372         }
373         if (change == 0)
374                 RETURN(0);
375
376         if (lmv->connected == 0)
377                 RETURN(0);
378
379         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
380                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
381
382                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active) {
383                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
384                         continue;
385                 }
386
387                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize,
388                                      cookiesize, def_cookiesize);
389                 if (rc) {
390                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
391                                " rc = %d\n", obd->obd_name, i, rc);
392                         break;
393                 }
394         }
395         RETURN(rc);
396 }
397
398 #define MAX_STRING_SIZE 128
399
400 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
401 {
402         struct lmv_obd          *lmv = &obd->u.lmv;
403         struct obd_uuid         *cluuid = &lmv->cluuid;
404         struct obd_uuid          lmv_mdc_uuid = { "LMV_MDC_UUID" };
405         struct obd_device       *mdc_obd;
406         struct obd_export       *mdc_exp;
407         struct lu_fld_target     target;
408         int                      rc;
409         ENTRY;
410
411         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
412                                         &obd->obd_uuid);
413         if (!mdc_obd) {
414                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
415                 RETURN(-EINVAL);
416         }
417
418         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
419                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
420                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
421                 cluuid->uuid);
422
423         if (!mdc_obd->obd_set_up) {
424                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
425                 RETURN(-EINVAL);
426         }
427
428         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
429                          &lmv->conn_data, NULL);
430         if (rc) {
431                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
432                 RETURN(rc);
433         }
434
435         /*
436          * Init fid sequence client for this mdc and add new fld target.
437          */
438         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
439         if (rc)
440                 RETURN(rc);
441
442         target.ft_srv = NULL;
443         target.ft_exp = mdc_exp;
444         target.ft_idx = tgt->ltd_idx;
445
446         fld_client_add_target(&lmv->lmv_fld, &target);
447
448         rc = obd_register_observer(mdc_obd, obd);
449         if (rc) {
450                 obd_disconnect(mdc_exp);
451                 CERROR("target %s register_observer error %d\n",
452                        tgt->ltd_uuid.uuid, rc);
453                 RETURN(rc);
454         }
455
456         if (obd->obd_observer) {
457                 /*
458                  * Tell the observer about the new target.
459                  */
460                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
461                                 OBD_NOTIFY_ACTIVE,
462                                 (void *)(tgt - lmv->tgts[0]));
463                 if (rc) {
464                         obd_disconnect(mdc_exp);
465                         RETURN(rc);
466                 }
467         }
468
469         tgt->ltd_active = 1;
470         tgt->ltd_exp = mdc_exp;
471         lmv->desc.ld_active_tgt_count++;
472
473         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize,
474                         lmv->max_cookiesize, lmv->max_def_cookiesize);
475
476         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
477                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
478                 atomic_read(&obd->obd_refcount));
479
480         if (lmv->targets_proc_entry != NULL) {
481                 struct proc_dir_entry *mdc_symlink;
482
483                 LASSERT(mdc_obd->obd_type != NULL);
484                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
485                 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
486                                                   lmv->targets_proc_entry,
487                                                   "../../../%s/%s",
488                                                   mdc_obd->obd_type->typ_name,
489                                                   mdc_obd->obd_name);
490                 if (mdc_symlink == NULL) {
491                         CERROR("cannot register LMV target "
492                                "/proc/fs/lustre/%s/%s/target_obds/%s\n",
493                                obd->obd_type->typ_name, obd->obd_name,
494                                mdc_obd->obd_name);
495                 }
496         }
497         RETURN(0);
498 }
499
500 static void lmv_del_target(struct lmv_obd *lmv, int index)
501 {
502         if (lmv->tgts[index] == NULL)
503                 return;
504
505         OBD_FREE_PTR(lmv->tgts[index]);
506         lmv->tgts[index] = NULL;
507         return;
508 }
509
510 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
511                            __u32 index, int gen)
512 {
513         struct lmv_obd      *lmv = &obd->u.lmv;
514         struct lmv_tgt_desc *tgt;
515         int                  orig_tgt_count = 0;
516         int                  rc = 0;
517         ENTRY;
518
519         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
520
521         lmv_init_lock(lmv);
522
523         if (lmv->desc.ld_tgt_count == 0) {
524                 struct obd_device *mdc_obd;
525
526                 mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
527                                                 &obd->obd_uuid);
528                 if (!mdc_obd) {
529                         lmv_init_unlock(lmv);
530                         CERROR("%s: Target %s not attached: rc = %d\n",
531                                obd->obd_name, uuidp->uuid, -EINVAL);
532                         RETURN(-EINVAL);
533                 }
534         }
535
536         if ((index < lmv->tgts_size) && (lmv->tgts[index] != NULL)) {
537                 tgt = lmv->tgts[index];
538                 CERROR("%s: UUID %s already assigned at LOV target index %d:"
539                        " rc = %d\n", obd->obd_name,
540                        obd_uuid2str(&tgt->ltd_uuid), index, -EEXIST);
541                 lmv_init_unlock(lmv);
542                 RETURN(-EEXIST);
543         }
544
545         if (index >= lmv->tgts_size) {
546                 /* We need to reallocate the lmv target array. */
547                 struct lmv_tgt_desc **newtgts, **old = NULL;
548                 __u32 newsize = 1;
549                 __u32 oldsize = 0;
550
551                 while (newsize < index + 1)
552                         newsize = newsize << 1;
553                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
554                 if (newtgts == NULL) {
555                         lmv_init_unlock(lmv);
556                         RETURN(-ENOMEM);
557                 }
558
559                 if (lmv->tgts_size) {
560                         memcpy(newtgts, lmv->tgts,
561                                sizeof(*newtgts) * lmv->tgts_size);
562                         old = lmv->tgts;
563                         oldsize = lmv->tgts_size;
564                 }
565
566                 lmv->tgts = newtgts;
567                 lmv->tgts_size = newsize;
568                 smp_rmb();
569                 if (old)
570                         OBD_FREE(old, sizeof(*old) * oldsize);
571
572                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n", lmv->tgts,
573                        lmv->tgts_size);
574         }
575
576         OBD_ALLOC_PTR(tgt);
577         if (!tgt) {
578                 lmv_init_unlock(lmv);
579                 RETURN(-ENOMEM);
580         }
581
582         mutex_init(&tgt->ltd_fid_mutex);
583         tgt->ltd_idx = index;
584         tgt->ltd_uuid = *uuidp;
585         tgt->ltd_active = 0;
586         lmv->tgts[index] = tgt;
587         if (index >= lmv->desc.ld_tgt_count) {
588                 orig_tgt_count = lmv->desc.ld_tgt_count;
589                 lmv->desc.ld_tgt_count = index + 1;
590         }
591
592         if (lmv->connected) {
593                 rc = lmv_connect_mdc(obd, tgt);
594                 if (rc != 0) {
595                         spin_lock(&lmv->lmv_lock);
596                         if (lmv->desc.ld_tgt_count == index + 1)
597                                 lmv->desc.ld_tgt_count = orig_tgt_count;
598                         memset(tgt, 0, sizeof(*tgt));
599                         spin_unlock(&lmv->lmv_lock);
600                 } else {
601                         int easize = sizeof(struct lmv_stripe_md) +
602                                 lmv->desc.ld_tgt_count * sizeof(struct lu_fid);
603                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0, 0);
604                 }
605         }
606
607         lmv_init_unlock(lmv);
608         RETURN(rc);
609 }
610
611 int lmv_check_connect(struct obd_device *obd)
612 {
613         struct lmv_obd          *lmv = &obd->u.lmv;
614         struct lmv_tgt_desc     *tgt;
615         __u32                    i;
616         int                      rc;
617         int                      easize;
618         ENTRY;
619
620         if (lmv->connected)
621                 RETURN(0);
622
623         lmv_init_lock(lmv);
624         if (lmv->connected) {
625                 lmv_init_unlock(lmv);
626                 RETURN(0);
627         }
628
629         if (lmv->desc.ld_tgt_count == 0) {
630                 lmv_init_unlock(lmv);
631                 CERROR("%s: no targets configured.\n", obd->obd_name);
632                 RETURN(-EINVAL);
633         }
634
635         LASSERT(lmv->tgts != NULL);
636
637         if (lmv->tgts[0] == NULL) {
638                 lmv_init_unlock(lmv);
639                 CERROR("%s: no target configured for index 0.\n",
640                        obd->obd_name);
641                 RETURN(-EINVAL);
642         }
643
644         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
645                lmv->cluuid.uuid, obd->obd_name);
646
647         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
648                 tgt = lmv->tgts[i];
649                 if (tgt == NULL)
650                         continue;
651                 rc = lmv_connect_mdc(obd, tgt);
652                 if (rc)
653                         GOTO(out_disc, rc);
654         }
655
656         lmv_set_timeouts(obd);
657         class_export_put(lmv->exp);
658         lmv->connected = 1;
659         easize = lmv_mds_md_size(lmv->desc.ld_tgt_count, LMV_MAGIC);
660         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0, 0);
661         lmv_init_unlock(lmv);
662         RETURN(0);
663
664  out_disc:
665         while (i-- > 0) {
666                 int rc2;
667                 tgt = lmv->tgts[i];
668                 if (tgt == NULL)
669                         continue;
670                 tgt->ltd_active = 0;
671                 if (tgt->ltd_exp) {
672                         --lmv->desc.ld_active_tgt_count;
673                         rc2 = obd_disconnect(tgt->ltd_exp);
674                         if (rc2) {
675                                 CERROR("LMV target %s disconnect on "
676                                        "MDC idx %d: error %d\n",
677                                        tgt->ltd_uuid.uuid, i, rc2);
678                         }
679                 }
680         }
681         class_disconnect(lmv->exp);
682         lmv_init_unlock(lmv);
683         RETURN(rc);
684 }
685
686 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
687 {
688         struct lmv_obd         *lmv = &obd->u.lmv;
689         struct obd_device      *mdc_obd;
690         int                     rc;
691         ENTRY;
692
693         LASSERT(tgt != NULL);
694         LASSERT(obd != NULL);
695
696         mdc_obd = class_exp2obd(tgt->ltd_exp);
697
698         if (mdc_obd) {
699                 mdc_obd->obd_force = obd->obd_force;
700                 mdc_obd->obd_fail = obd->obd_fail;
701                 mdc_obd->obd_no_recov = obd->obd_no_recov;
702         }
703
704         if (lmv->targets_proc_entry != NULL)
705                 lprocfs_remove_proc_entry(mdc_obd->obd_name,
706                                           lmv->targets_proc_entry);
707
708         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
709         if (rc)
710                 CERROR("Can't finanize fids factory\n");
711
712         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
713                tgt->ltd_exp->exp_obd->obd_name,
714                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
715
716         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
717         rc = obd_disconnect(tgt->ltd_exp);
718         if (rc) {
719                 if (tgt->ltd_active) {
720                         CERROR("Target %s disconnect error %d\n",
721                                tgt->ltd_uuid.uuid, rc);
722                 }
723         }
724
725         lmv_activate_target(lmv, tgt, 0);
726         tgt->ltd_exp = NULL;
727         RETURN(0);
728 }
729
730 static int lmv_disconnect(struct obd_export *exp)
731 {
732         struct obd_device       *obd = class_exp2obd(exp);
733         struct lmv_obd          *lmv = &obd->u.lmv;
734         int                      rc;
735         __u32                    i;
736         ENTRY;
737
738         if (!lmv->tgts)
739                 goto out_local;
740
741         /*
742          * Only disconnect the underlying layers on the final disconnect.
743          */
744         lmv->refcount--;
745         if (lmv->refcount != 0)
746                 goto out_local;
747
748         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
749                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
750                         continue;
751
752                 lmv_disconnect_mdc(obd, lmv->tgts[i]);
753         }
754
755         if (lmv->targets_proc_entry != NULL)
756                 lprocfs_remove(&lmv->targets_proc_entry);
757         else
758                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
759                        obd->obd_type->typ_name, obd->obd_name);
760
761 out_local:
762         /*
763          * This is the case when no real connection is established by
764          * lmv_check_connect().
765          */
766         if (!lmv->connected)
767                 class_export_put(exp);
768         rc = class_disconnect(exp);
769         if (lmv->refcount == 0)
770                 lmv->connected = 0;
771         RETURN(rc);
772 }
773
774 static int lmv_fid2path(struct obd_export *exp, int len, void *karg, void *uarg)
775 {
776         struct obd_device       *obddev = class_exp2obd(exp);
777         struct lmv_obd          *lmv = &obddev->u.lmv;
778         struct getinfo_fid2path *gf;
779         struct lmv_tgt_desc     *tgt;
780         struct getinfo_fid2path *remote_gf = NULL;
781         int                     remote_gf_size = 0;
782         int                     rc;
783
784         gf = (struct getinfo_fid2path *)karg;
785         tgt = lmv_find_target(lmv, &gf->gf_fid);
786         if (IS_ERR(tgt))
787                 RETURN(PTR_ERR(tgt));
788
789 repeat_fid2path:
790         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
791         if (rc != 0 && rc != -EREMOTE)
792                 GOTO(out_fid2path, rc);
793
794         /* If remote_gf != NULL, it means just building the
795          * path on the remote MDT, copy this path segement to gf */
796         if (remote_gf != NULL) {
797                 struct getinfo_fid2path *ori_gf;
798                 char *ptr;
799
800                 ori_gf = (struct getinfo_fid2path *)karg;
801                 if (strlen(ori_gf->gf_path) +
802                     strlen(gf->gf_path) > ori_gf->gf_pathlen)
803                         GOTO(out_fid2path, rc = -EOVERFLOW);
804
805                 ptr = ori_gf->gf_path;
806
807                 memmove(ptr + strlen(gf->gf_path) + 1, ptr,
808                         strlen(ori_gf->gf_path));
809
810                 strncpy(ptr, gf->gf_path, strlen(gf->gf_path));
811                 ptr += strlen(gf->gf_path);
812                 *ptr = '/';
813         }
814
815         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: "LPU64" ln: %u\n",
816                tgt->ltd_exp->exp_obd->obd_name,
817                gf->gf_path, PFID(&gf->gf_fid), gf->gf_recno,
818                gf->gf_linkno);
819
820         if (rc == 0)
821                 GOTO(out_fid2path, rc);
822
823         /* sigh, has to go to another MDT to do path building further */
824         if (remote_gf == NULL) {
825                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
826                 OBD_ALLOC(remote_gf, remote_gf_size);
827                 if (remote_gf == NULL)
828                         GOTO(out_fid2path, rc = -ENOMEM);
829                 remote_gf->gf_pathlen = PATH_MAX;
830         }
831
832         if (!fid_is_sane(&gf->gf_fid)) {
833                 CERROR("%s: invalid FID "DFID": rc = %d\n",
834                        tgt->ltd_exp->exp_obd->obd_name,
835                        PFID(&gf->gf_fid), -EINVAL);
836                 GOTO(out_fid2path, rc = -EINVAL);
837         }
838
839         tgt = lmv_find_target(lmv, &gf->gf_fid);
840         if (IS_ERR(tgt))
841                 GOTO(out_fid2path, rc = -EINVAL);
842
843         remote_gf->gf_fid = gf->gf_fid;
844         remote_gf->gf_recno = -1;
845         remote_gf->gf_linkno = -1;
846         memset(remote_gf->gf_path, 0, remote_gf->gf_pathlen);
847         gf = remote_gf;
848         goto repeat_fid2path;
849
850 out_fid2path:
851         if (remote_gf != NULL)
852                 OBD_FREE(remote_gf, remote_gf_size);
853         RETURN(rc);
854 }
855
856 static int lmv_hsm_req_count(struct lmv_obd *lmv,
857                              const struct hsm_user_request *hur,
858                              const struct lmv_tgt_desc *tgt_mds)
859 {
860         __u32                    i;
861         int                      nr = 0;
862         struct lmv_tgt_desc     *curr_tgt;
863
864         /* count how many requests must be sent to the given target */
865         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
866                 curr_tgt = lmv_find_target(lmv, &hur->hur_user_item[i].hui_fid);
867                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
868                         nr++;
869         }
870         return nr;
871 }
872
873 static void lmv_hsm_req_build(struct lmv_obd *lmv,
874                               struct hsm_user_request *hur_in,
875                               const struct lmv_tgt_desc *tgt_mds,
876                               struct hsm_user_request *hur_out)
877 {
878         __u32                    i, nr_out;
879         struct lmv_tgt_desc     *curr_tgt;
880
881         /* build the hsm_user_request for the given target */
882         hur_out->hur_request = hur_in->hur_request;
883         nr_out = 0;
884         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
885                 curr_tgt = lmv_find_target(lmv,
886                                            &hur_in->hur_user_item[i].hui_fid);
887                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
888                         hur_out->hur_user_item[nr_out] =
889                                                 hur_in->hur_user_item[i];
890                         nr_out++;
891                 }
892         }
893         hur_out->hur_request.hr_itemcount = nr_out;
894         memcpy(hur_data(hur_out), hur_data(hur_in),
895                hur_in->hur_request.hr_data_len);
896 }
897
898 static int lmv_hsm_ct_unregister(struct lmv_obd *lmv, unsigned int cmd, int len,
899                                  struct lustre_kernelcomm *lk, void *uarg)
900 {
901         __u32                    i;
902         int                      rc;
903         struct kkuc_ct_data     *kcd = NULL;
904         ENTRY;
905
906         /* unregister request (call from llapi_hsm_copytool_fini) */
907         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
908                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
909
910                 if (tgt == NULL || tgt->ltd_exp == NULL)
911                         continue;
912                 /* best effort: try to clean as much as possible
913                  * (continue on error) */
914                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
915         }
916
917         /* Whatever the result, remove copytool from kuc groups.
918          * Unreached coordinators will get EPIPE on next requests
919          * and will unregister automatically.
920          */
921         rc = libcfs_kkuc_group_rem(lk->lk_uid, lk->lk_group, (void **)&kcd);
922         if (kcd != NULL)
923                 OBD_FREE_PTR(kcd);
924
925         RETURN(rc);
926 }
927
928 static int lmv_hsm_ct_register(struct lmv_obd *lmv, unsigned int cmd, int len,
929                                struct lustre_kernelcomm *lk, void *uarg)
930 {
931         struct file             *filp;
932         __u32                    i, j;
933         int                      err, rc;
934         bool                     any_set = false;
935         struct kkuc_ct_data     *kcd;
936         ENTRY;
937
938         /* All or nothing: try to register to all MDS.
939          * In case of failure, unregister from previous MDS,
940          * except if it because of inactive target. */
941         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
942                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
943
944                 if (tgt == NULL || tgt->ltd_exp == NULL)
945                         continue;
946                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
947                 if (err) {
948                         if (tgt->ltd_active) {
949                                 /* permanent error */
950                                 CERROR("%s: iocontrol MDC %s on MDT"
951                                        " idx %d cmd %x: err = %d\n",
952                                        class_exp2obd(lmv->exp)->obd_name,
953                                        tgt->ltd_uuid.uuid, i, cmd, err);
954                                 rc = err;
955                                 lk->lk_flags |= LK_FLG_STOP;
956                                 /* unregister from previous MDS */
957                                 for (j = 0; j < i; j++) {
958                                         tgt = lmv->tgts[j];
959                                         if (tgt == NULL || tgt->ltd_exp == NULL)
960                                                 continue;
961                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
962                                                       lk, uarg);
963                                 }
964                                 RETURN(rc);
965                         }
966                         /* else: transient error.
967                          * kuc will register to the missing MDT
968                          * when it is back */
969                 } else {
970                         any_set = true;
971                 }
972         }
973
974         if (!any_set)
975                 /* no registration done: return error */
976                 RETURN(-ENOTCONN);
977
978         /* at least one registration done, with no failure */
979         filp = fget(lk->lk_wfd);
980         if (filp == NULL)
981                 RETURN(-EBADF);
982
983         OBD_ALLOC_PTR(kcd);
984         if (kcd == NULL) {
985                 fput(filp);
986                 RETURN(-ENOMEM);
987         }
988         kcd->kcd_magic = KKUC_CT_DATA_MAGIC;
989         kcd->kcd_uuid = lmv->cluuid;
990         kcd->kcd_archive = lk->lk_data;
991
992         rc = libcfs_kkuc_group_add(filp, lk->lk_uid, lk->lk_group, kcd);
993         if (rc != 0) {
994                 if (filp != NULL)
995                         fput(filp);
996                 OBD_FREE_PTR(kcd);
997         }
998
999         RETURN(rc);
1000 }
1001
1002
1003
1004
1005 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
1006                          int len, void *karg, void *uarg)
1007 {
1008         struct obd_device       *obddev = class_exp2obd(exp);
1009         struct lmv_obd          *lmv = &obddev->u.lmv;
1010         struct lmv_tgt_desc     *tgt = NULL;
1011         __u32                    i = 0;
1012         int                      rc = 0;
1013         int                      set = 0;
1014         __u32                    count = lmv->desc.ld_tgt_count;
1015         ENTRY;
1016
1017         if (count == 0)
1018                 RETURN(-ENOTTY);
1019
1020         switch (cmd) {
1021         case IOC_OBD_STATFS: {
1022                 struct obd_ioctl_data *data = karg;
1023                 struct obd_device *mdc_obd;
1024                 struct obd_statfs stat_buf = {0};
1025                 __u32 index;
1026
1027                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
1028                 if ((index >= count))
1029                         RETURN(-ENODEV);
1030
1031                 tgt = lmv->tgts[index];
1032                 if (tgt == NULL || !tgt->ltd_active)
1033                         RETURN(-ENODATA);
1034
1035                 mdc_obd = class_exp2obd(tgt->ltd_exp);
1036                 if (!mdc_obd)
1037                         RETURN(-EINVAL);
1038
1039                 /* copy UUID */
1040                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
1041                                  min((int) data->ioc_plen2,
1042                                      (int) sizeof(struct obd_uuid))))
1043                         RETURN(-EFAULT);
1044
1045                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
1046                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
1047                                 0);
1048                 if (rc)
1049                         RETURN(rc);
1050                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1051                                  min((int) data->ioc_plen1,
1052                                      (int) sizeof(stat_buf))))
1053                         RETURN(-EFAULT);
1054                 break;
1055         }
1056         case OBD_IOC_QUOTACTL: {
1057                 struct if_quotactl *qctl = karg;
1058                 struct obd_quotactl *oqctl;
1059
1060                 if (qctl->qc_valid == QC_MDTIDX) {
1061                         if (count <= qctl->qc_idx)
1062                                 RETURN(-EINVAL);
1063
1064                         tgt = lmv->tgts[qctl->qc_idx];
1065                         if (tgt == NULL || tgt->ltd_exp == NULL)
1066                                 RETURN(-EINVAL);
1067                 } else if (qctl->qc_valid == QC_UUID) {
1068                         for (i = 0; i < count; i++) {
1069                                 tgt = lmv->tgts[i];
1070                                 if (tgt == NULL)
1071                                         continue;
1072                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
1073                                                      &qctl->obd_uuid))
1074                                         continue;
1075
1076                                 if (tgt->ltd_exp == NULL)
1077                                         RETURN(-EINVAL);
1078
1079                                 break;
1080                         }
1081                 } else {
1082                         RETURN(-EINVAL);
1083                 }
1084
1085                 if (i >= count)
1086                         RETURN(-EAGAIN);
1087
1088                 LASSERT(tgt != NULL && tgt->ltd_exp != NULL);
1089                 OBD_ALLOC_PTR(oqctl);
1090                 if (!oqctl)
1091                         RETURN(-ENOMEM);
1092
1093                 QCTL_COPY(oqctl, qctl);
1094                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
1095                 if (rc == 0) {
1096                         QCTL_COPY(qctl, oqctl);
1097                         qctl->qc_valid = QC_MDTIDX;
1098                         qctl->obd_uuid = tgt->ltd_uuid;
1099                 }
1100                 OBD_FREE_PTR(oqctl);
1101                 break;
1102         }
1103         case OBD_IOC_CHANGELOG_SEND:
1104         case OBD_IOC_CHANGELOG_CLEAR: {
1105                 struct ioc_changelog *icc = karg;
1106
1107                 if (icc->icc_mdtindex >= count)
1108                         RETURN(-ENODEV);
1109
1110                 tgt = lmv->tgts[icc->icc_mdtindex];
1111                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
1112                         RETURN(-ENODEV);
1113                 rc = obd_iocontrol(cmd, tgt->ltd_exp, sizeof(*icc), icc, NULL);
1114                 break;
1115         }
1116         case LL_IOC_GET_CONNECT_FLAGS: {
1117                 tgt = lmv->tgts[0];
1118                 if (tgt == NULL || tgt->ltd_exp == NULL)
1119                         RETURN(-ENODATA);
1120                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1121                 break;
1122         }
1123         case LL_IOC_FID2MDTIDX: {
1124                 struct lu_fid *fid = karg;
1125                 int             mdt_index;
1126
1127                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
1128                 if (rc != 0)
1129                         RETURN(rc);
1130
1131                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
1132                  * point to user space memory for FID2MDTIDX. */
1133                 *(__u32 *)uarg = mdt_index;
1134                 break;
1135         }
1136         case OBD_IOC_FID2PATH: {
1137                 rc = lmv_fid2path(exp, len, karg, uarg);
1138                 break;
1139         }
1140         case LL_IOC_HSM_STATE_GET:
1141         case LL_IOC_HSM_STATE_SET:
1142         case LL_IOC_HSM_ACTION: {
1143                 struct md_op_data       *op_data = karg;
1144
1145                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1146                 if (IS_ERR(tgt))
1147                         RETURN(PTR_ERR(tgt));
1148
1149                 if (tgt->ltd_exp == NULL)
1150                         RETURN(-EINVAL);
1151
1152                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1153                 break;
1154         }
1155         case LL_IOC_HSM_PROGRESS: {
1156                 const struct hsm_progress_kernel *hpk = karg;
1157
1158                 tgt = lmv_find_target(lmv, &hpk->hpk_fid);
1159                 if (IS_ERR(tgt))
1160                         RETURN(PTR_ERR(tgt));
1161                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1162                 break;
1163         }
1164         case LL_IOC_HSM_REQUEST: {
1165                 struct hsm_user_request *hur = karg;
1166                 unsigned int reqcount = hur->hur_request.hr_itemcount;
1167
1168                 if (reqcount == 0)
1169                         RETURN(0);
1170
1171                 /* if the request is about a single fid
1172                  * or if there is a single MDS, no need to split
1173                  * the request. */
1174                 if (reqcount == 1 || count == 1) {
1175                         tgt = lmv_find_target(lmv,
1176                                               &hur->hur_user_item[0].hui_fid);
1177                         if (IS_ERR(tgt))
1178                                 RETURN(PTR_ERR(tgt));
1179                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1180                 } else {
1181                         /* split fid list to their respective MDS */
1182                         for (i = 0; i < count; i++) {
1183                                 unsigned int            nr, reqlen;
1184                                 int                     rc1;
1185                                 struct hsm_user_request *req;
1186
1187                                 tgt = lmv->tgts[i];
1188                                 if (tgt == NULL || tgt->ltd_exp == NULL)
1189                                         continue;
1190
1191                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
1192                                 if (nr == 0) /* nothing for this MDS */
1193                                         continue;
1194
1195                                 /* build a request with fids for this MDS */
1196                                 reqlen = offsetof(typeof(*hur),
1197                                                   hur_user_item[nr])
1198                                                 + hur->hur_request.hr_data_len;
1199                                 OBD_ALLOC_LARGE(req, reqlen);
1200                                 if (req == NULL)
1201                                         RETURN(-ENOMEM);
1202
1203                                 lmv_hsm_req_build(lmv, hur, tgt, req);
1204
1205                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1206                                                     req, uarg);
1207                                 if (rc1 != 0 && rc == 0)
1208                                         rc = rc1;
1209                                 OBD_FREE_LARGE(req, reqlen);
1210                         }
1211                 }
1212                 break;
1213         }
1214         case LL_IOC_LOV_SWAP_LAYOUTS: {
1215                 struct md_op_data       *op_data = karg;
1216                 struct lmv_tgt_desc     *tgt1, *tgt2;
1217
1218                 tgt1 = lmv_find_target(lmv, &op_data->op_fid1);
1219                 if (IS_ERR(tgt1))
1220                         RETURN(PTR_ERR(tgt1));
1221
1222                 tgt2 = lmv_find_target(lmv, &op_data->op_fid2);
1223                 if (IS_ERR(tgt2))
1224                         RETURN(PTR_ERR(tgt2));
1225
1226                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1227                         RETURN(-EINVAL);
1228
1229                 /* only files on same MDT can have their layouts swapped */
1230                 if (tgt1->ltd_idx != tgt2->ltd_idx)
1231                         RETURN(-EPERM);
1232
1233                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1234                 break;
1235         }
1236         case LL_IOC_HSM_CT_START: {
1237                 struct lustre_kernelcomm *lk = karg;
1238                 if (lk->lk_flags & LK_FLG_STOP)
1239                         rc = lmv_hsm_ct_unregister(lmv, cmd, len, lk, uarg);
1240                 else
1241                         rc = lmv_hsm_ct_register(lmv, cmd, len, lk, uarg);
1242                 break;
1243         }
1244         default:
1245                 for (i = 0; i < count; i++) {
1246                         struct obd_device *mdc_obd;
1247                         int err;
1248
1249                         tgt = lmv->tgts[i];
1250                         if (tgt == NULL || tgt->ltd_exp == NULL)
1251                                 continue;
1252                         /* ll_umount_begin() sets force flag but for lmv, not
1253                          * mdc. Let's pass it through */
1254                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1255                         mdc_obd->obd_force = obddev->obd_force;
1256                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1257                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
1258                                 RETURN(err);
1259                         } else if (err) {
1260                                 if (tgt->ltd_active) {
1261                                         CERROR("error: iocontrol MDC %s on MDT"
1262                                                " idx %d cmd %x: err = %d\n",
1263                                                tgt->ltd_uuid.uuid, i, cmd, err);
1264                                         if (!rc)
1265                                                 rc = err;
1266                                 }
1267                         } else
1268                                 set = 1;
1269                 }
1270                 if (!set && !rc)
1271                         rc = -EIO;
1272         }
1273         RETURN(rc);
1274 }
1275
1276 #if 0
1277 static int lmv_all_chars_policy(int count, const char *name,
1278                                 int len)
1279 {
1280         unsigned int c = 0;
1281
1282         while (len > 0)
1283                 c += name[--len];
1284         c = c % count;
1285         return c;
1286 }
1287
1288 static int lmv_nid_policy(struct lmv_obd *lmv)
1289 {
1290         struct obd_import *imp;
1291         __u32              id;
1292
1293         /*
1294          * XXX: To get nid we assume that underlying obd device is mdc.
1295          */
1296         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
1297         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
1298         return id % lmv->desc.ld_tgt_count;
1299 }
1300
1301 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
1302                           placement_policy_t placement)
1303 {
1304         switch (placement) {
1305         case PLACEMENT_CHAR_POLICY:
1306                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
1307                                             op_data->op_name,
1308                                             op_data->op_namelen);
1309         case PLACEMENT_NID_POLICY:
1310                 return lmv_nid_policy(lmv);
1311
1312         default:
1313                 break;
1314         }
1315
1316         CERROR("Unsupported placement policy %x\n", placement);
1317         return -EINVAL;
1318 }
1319 #endif
1320
1321 /**
1322  * This is _inode_ placement policy function (not name).
1323  */
1324 static int lmv_placement_policy(struct obd_device *obd,
1325                                 struct md_op_data *op_data,
1326                                 mdsno_t *mds)
1327 {
1328         struct lmv_obd          *lmv = &obd->u.lmv;
1329         ENTRY;
1330
1331         LASSERT(mds != NULL);
1332
1333         if (lmv->desc.ld_tgt_count == 1) {
1334                 *mds = 0;
1335                 RETURN(0);
1336         }
1337
1338         /**
1339          * If stripe_offset is provided during setdirstripe
1340          * (setdirstripe -i xx), xx MDS will be choosen.
1341          */
1342         if (op_data->op_cli_flags & CLI_SET_MEA && op_data->op_data != NULL) {
1343                 struct lmv_user_md *lum;
1344
1345                 lum = op_data->op_data;
1346
1347                 if (le32_to_cpu(lum->lum_stripe_offset) != (__u32)-1) {
1348                         *mds = le32_to_cpu(lum->lum_stripe_offset);
1349                 } else {
1350                         /* -1 means default, which will be in the same MDT with
1351                          * the stripe */
1352                         *mds = op_data->op_mds;
1353                         lum->lum_stripe_offset = cpu_to_le32(op_data->op_mds);
1354                 }
1355         } else {
1356                 /* Allocate new fid on target according to operation type and
1357                  * parent home mds. */
1358                 *mds = op_data->op_mds;
1359         }
1360
1361         RETURN(0);
1362 }
1363
1364 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
1365                     mdsno_t mds)
1366 {
1367         struct lmv_tgt_desc     *tgt;
1368         int                      rc;
1369         ENTRY;
1370
1371         tgt = lmv_get_target(lmv, mds, NULL);
1372         if (IS_ERR(tgt))
1373                 RETURN(PTR_ERR(tgt));
1374
1375         /*
1376          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1377          * on server that seq in new allocated fid is not yet known.
1378          */
1379         mutex_lock(&tgt->ltd_fid_mutex);
1380
1381         if (tgt->ltd_active == 0 || tgt->ltd_exp == NULL)
1382                 GOTO(out, rc = -ENODEV);
1383
1384         /*
1385          * Asking underlying tgt layer to allocate new fid.
1386          */
1387         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1388         if (rc > 0) {
1389                 LASSERT(fid_is_sane(fid));
1390                 rc = 0;
1391         }
1392
1393         EXIT;
1394 out:
1395         mutex_unlock(&tgt->ltd_fid_mutex);
1396         return rc;
1397 }
1398
1399 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1400                   struct lu_fid *fid, struct md_op_data *op_data)
1401 {
1402         struct obd_device     *obd = class_exp2obd(exp);
1403         struct lmv_obd        *lmv = &obd->u.lmv;
1404         mdsno_t                mds = 0;
1405         int                    rc;
1406         ENTRY;
1407
1408         LASSERT(op_data != NULL);
1409         LASSERT(fid != NULL);
1410
1411         rc = lmv_placement_policy(obd, op_data, &mds);
1412         if (rc) {
1413                 CERROR("Can't get target for allocating fid, "
1414                        "rc %d\n", rc);
1415                 RETURN(rc);
1416         }
1417
1418         rc = __lmv_fid_alloc(lmv, fid, mds);
1419         if (rc) {
1420                 CERROR("Can't alloc new fid, rc %d\n", rc);
1421                 RETURN(rc);
1422         }
1423
1424         RETURN(rc);
1425 }
1426
1427 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1428 {
1429         struct lmv_obd  *lmv = &obd->u.lmv;
1430         struct lmv_desc *desc;
1431         int             rc;
1432         ENTRY;
1433
1434         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1435                 CERROR("LMV setup requires a descriptor\n");
1436                 RETURN(-EINVAL);
1437         }
1438
1439         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1440         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1441                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1442                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1443                 RETURN(-EINVAL);
1444         }
1445
1446         OBD_ALLOC(lmv->tgts, sizeof(*lmv->tgts) * 32);
1447         if (lmv->tgts == NULL)
1448                 RETURN(-ENOMEM);
1449         lmv->tgts_size = 32;
1450
1451         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1452         lmv->desc.ld_tgt_count = 0;
1453         lmv->desc.ld_active_tgt_count = 0;
1454         lmv->max_cookiesize = 0;
1455         lmv->max_def_easize = 0;
1456         lmv->max_easize = 0;
1457         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1458
1459         spin_lock_init(&lmv->lmv_lock);
1460         mutex_init(&lmv->init_mutex);
1461
1462 #ifdef LPROCFS
1463         obd->obd_vars = lprocfs_lmv_obd_vars;
1464         lprocfs_seq_obd_setup(obd);
1465         lprocfs_alloc_md_stats(obd, 0);
1466         rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
1467                                 0444, &lmv_proc_target_fops, obd);
1468         if (rc)
1469                 CWARN("%s: error adding LMV target_obd file: rc = %d\n",
1470                       obd->obd_name, rc);
1471 #endif
1472         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1473                              LUSTRE_CLI_FLD_HASH_DHT);
1474         if (rc) {
1475                 CERROR("Can't init FLD, err %d\n", rc);
1476                 GOTO(out, rc);
1477         }
1478
1479         RETURN(0);
1480
1481 out:
1482         return rc;
1483 }
1484
1485 static int lmv_cleanup(struct obd_device *obd)
1486 {
1487         struct lmv_obd   *lmv = &obd->u.lmv;
1488         ENTRY;
1489
1490         fld_client_fini(&lmv->lmv_fld);
1491         if (lmv->tgts != NULL) {
1492                 int i;
1493                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1494                         if (lmv->tgts[i] == NULL)
1495                                 continue;
1496                         lmv_del_target(lmv, i);
1497                 }
1498                 OBD_FREE(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1499                 lmv->tgts_size = 0;
1500         }
1501         RETURN(0);
1502 }
1503
1504 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1505 {
1506         struct lustre_cfg       *lcfg = buf;
1507         struct obd_uuid         obd_uuid;
1508         int                     gen;
1509         __u32                   index;
1510         int                     rc;
1511         ENTRY;
1512
1513         switch (lcfg->lcfg_command) {
1514         case LCFG_ADD_MDC:
1515                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1516                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1517                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1518                         GOTO(out, rc = -EINVAL);
1519
1520                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1521
1522                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1523                         GOTO(out, rc = -EINVAL);
1524                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1525                         GOTO(out, rc = -EINVAL);
1526                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1527                 GOTO(out, rc);
1528         default:
1529                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1530                 GOTO(out, rc = -EINVAL);
1531         }
1532 out:
1533         RETURN(rc);
1534 }
1535
1536 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1537                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1538 {
1539         struct obd_device       *obd = class_exp2obd(exp);
1540         struct lmv_obd          *lmv = &obd->u.lmv;
1541         struct obd_statfs       *temp;
1542         int                      rc = 0;
1543         __u32                    i;
1544         ENTRY;
1545
1546         rc = lmv_check_connect(obd);
1547         if (rc)
1548                 RETURN(rc);
1549
1550         OBD_ALLOC(temp, sizeof(*temp));
1551         if (temp == NULL)
1552                 RETURN(-ENOMEM);
1553
1554         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1555                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1556                         continue;
1557
1558                 rc = obd_statfs(env, lmv->tgts[i]->ltd_exp, temp,
1559                                 max_age, flags);
1560                 if (rc) {
1561                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1562                                lmv->tgts[i]->ltd_exp->exp_obd->obd_name,
1563                                rc);
1564                         GOTO(out_free_temp, rc);
1565                 }
1566
1567                 if (i == 0) {
1568                         *osfs = *temp;
1569                         /* If the statfs is from mount, it will needs
1570                          * retrieve necessary information from MDT0.
1571                          * i.e. mount does not need the merged osfs
1572                          * from all of MDT.
1573                          * And also clients can be mounted as long as
1574                          * MDT0 is in service*/
1575                         if (flags & OBD_STATFS_FOR_MDT0)
1576                                 GOTO(out_free_temp, rc);
1577                 } else {
1578                         osfs->os_bavail += temp->os_bavail;
1579                         osfs->os_blocks += temp->os_blocks;
1580                         osfs->os_ffree += temp->os_ffree;
1581                         osfs->os_files += temp->os_files;
1582                 }
1583         }
1584
1585         EXIT;
1586 out_free_temp:
1587         OBD_FREE(temp, sizeof(*temp));
1588         return rc;
1589 }
1590
1591 static int lmv_getstatus(struct obd_export *exp,
1592                          struct lu_fid *fid,
1593                          struct obd_capa **pc)
1594 {
1595         struct obd_device    *obd = exp->exp_obd;
1596         struct lmv_obd       *lmv = &obd->u.lmv;
1597         int                   rc;
1598         ENTRY;
1599
1600         rc = lmv_check_connect(obd);
1601         if (rc)
1602                 RETURN(rc);
1603
1604         rc = md_getstatus(lmv->tgts[0]->ltd_exp, fid, pc);
1605         RETURN(rc);
1606 }
1607
1608 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1609                         struct obd_capa *oc, obd_valid valid, const char *name,
1610                         const char *input, int input_size, int output_size,
1611                         int flags, struct ptlrpc_request **request)
1612 {
1613         struct obd_device      *obd = exp->exp_obd;
1614         struct lmv_obd         *lmv = &obd->u.lmv;
1615         struct lmv_tgt_desc    *tgt;
1616         int                     rc;
1617         ENTRY;
1618
1619         rc = lmv_check_connect(obd);
1620         if (rc)
1621                 RETURN(rc);
1622
1623         tgt = lmv_find_target(lmv, fid);
1624         if (IS_ERR(tgt))
1625                 RETURN(PTR_ERR(tgt));
1626
1627         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1628                          input_size, output_size, flags, request);
1629
1630         RETURN(rc);
1631 }
1632
1633 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1634                         struct obd_capa *oc, obd_valid valid, const char *name,
1635                         const char *input, int input_size, int output_size,
1636                         int flags, __u32 suppgid,
1637                         struct ptlrpc_request **request)
1638 {
1639         struct obd_device      *obd = exp->exp_obd;
1640         struct lmv_obd         *lmv = &obd->u.lmv;
1641         struct lmv_tgt_desc    *tgt;
1642         int                     rc;
1643         ENTRY;
1644
1645         rc = lmv_check_connect(obd);
1646         if (rc)
1647                 RETURN(rc);
1648
1649         tgt = lmv_find_target(lmv, fid);
1650         if (IS_ERR(tgt))
1651                 RETURN(PTR_ERR(tgt));
1652
1653         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1654                          input_size, output_size, flags, suppgid,
1655                          request);
1656
1657         RETURN(rc);
1658 }
1659
1660 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1661                        struct ptlrpc_request **request)
1662 {
1663         struct obd_device       *obd = exp->exp_obd;
1664         struct lmv_obd          *lmv = &obd->u.lmv;
1665         struct lmv_tgt_desc     *tgt;
1666         int                      rc;
1667         ENTRY;
1668
1669         rc = lmv_check_connect(obd);
1670         if (rc)
1671                 RETURN(rc);
1672
1673         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1674         if (IS_ERR(tgt))
1675                 RETURN(PTR_ERR(tgt));
1676
1677         if (op_data->op_flags & MF_GET_MDT_IDX) {
1678                 op_data->op_mds = tgt->ltd_idx;
1679                 RETURN(0);
1680         }
1681
1682         rc = md_getattr(tgt->ltd_exp, op_data, request);
1683
1684         RETURN(rc);
1685 }
1686
1687 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1688 {
1689         struct obd_device   *obd = exp->exp_obd;
1690         struct lmv_obd      *lmv = &obd->u.lmv;
1691         __u32                i;
1692         int                  rc;
1693         ENTRY;
1694
1695         rc = lmv_check_connect(obd);
1696         if (rc)
1697                 RETURN(rc);
1698
1699         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1700
1701         /*
1702          * With DNE every object can have two locks in different namespaces:
1703          * lookup lock in space of MDT storing direntry and update/open lock in
1704          * space of MDT storing inode.
1705          */
1706         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1707                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1708                         continue;
1709                 md_null_inode(lmv->tgts[i]->ltd_exp, fid);
1710         }
1711
1712         RETURN(0);
1713 }
1714
1715 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1716                            ldlm_iterator_t it, void *data)
1717 {
1718         struct obd_device       *obd = exp->exp_obd;
1719         struct lmv_obd          *lmv = &obd->u.lmv;
1720         int                     i;
1721         int                     tgt;
1722         int                     rc;
1723         ENTRY;
1724
1725         rc = lmv_check_connect(obd);
1726         if (rc)
1727                 RETURN(rc);
1728
1729         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1730
1731         /*
1732          * With DNE every object can have two locks in different namespaces:
1733          * lookup lock in space of MDT storing direntry and update/open lock in
1734          * space of MDT storing inode.  Try the MDT that the FID maps to first,
1735          * since this can be easily found, and only try others if that fails.
1736          */
1737         for (i = 0, tgt = lmv_find_target_index(lmv, fid);
1738              i < lmv->desc.ld_tgt_count;
1739              i++, tgt = (tgt + 1) % lmv->desc.ld_tgt_count) {
1740                 if (tgt < 0) {
1741                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
1742                                obd->obd_name, PFID(fid), tgt);
1743                         tgt = 0;
1744                 }
1745
1746                 if (lmv->tgts[tgt] == NULL ||
1747                     lmv->tgts[tgt]->ltd_exp == NULL)
1748                         continue;
1749
1750                 rc = md_find_cbdata(lmv->tgts[tgt]->ltd_exp, fid, it, data);
1751                 if (rc)
1752                         RETURN(rc);
1753         }
1754
1755         RETURN(rc);
1756 }
1757
1758
1759 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1760                      struct md_open_data *mod, struct ptlrpc_request **request)
1761 {
1762         struct obd_device     *obd = exp->exp_obd;
1763         struct lmv_obd        *lmv = &obd->u.lmv;
1764         struct lmv_tgt_desc   *tgt;
1765         int                    rc;
1766         ENTRY;
1767
1768         rc = lmv_check_connect(obd);
1769         if (rc)
1770                 RETURN(rc);
1771
1772         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1773         if (IS_ERR(tgt))
1774                 RETURN(PTR_ERR(tgt));
1775
1776         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1777         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1778         RETURN(rc);
1779 }
1780
1781 /**
1782  * Choosing the MDT by name or FID in @op_data.
1783  * For non-striped directory, it will locate MDT by fid.
1784  * For striped-directory, it will locate MDT by name. And also
1785  * it will reset op_fid1 with the FID of the choosen stripe.
1786  **/
1787 struct lmv_tgt_desc *
1788 lmv_locate_target_for_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1789                            const char *name, int namelen, struct lu_fid *fid,
1790                            mdsno_t *mds)
1791 {
1792         struct lmv_tgt_desc     *tgt;
1793         const struct lmv_oinfo  *oinfo;
1794
1795         oinfo = lsm_name_to_stripe_info(lsm, name, namelen);
1796         if (IS_ERR(oinfo))
1797                 RETURN(ERR_CAST(oinfo));
1798         *fid = oinfo->lmo_fid;
1799         *mds = oinfo->lmo_mds;
1800         tgt = lmv_get_target(lmv, *mds, NULL);
1801
1802         CDEBUG(D_INFO, "locate on mds %u "DFID"\n", *mds, PFID(fid));
1803         return tgt;
1804 }
1805
1806 /**
1807  * Locate mds by fid or name
1808  *
1809  * For striped directory (lsm != NULL), it will locate the stripe
1810  * by name hash (see lsm_name_to_stripe_info()). Note: if the hash_type
1811  * is unknown, it will return -EBADFD, and lmv_intent_lookup might need
1812  * walk through all of stripes to locate the entry.
1813  *
1814  * For normal direcotry, it will locate MDS by FID directly.
1815  * \param[in] lmv       LMV device
1816  * \param[in] op_data   client MD stack parameters, name, namelen
1817  *                      mds_num etc.
1818  * \param[in] fid       object FID used to locate MDS.
1819  *
1820  * retval               pointer to the lmv_tgt_desc if succeed.
1821  *                      ERR_PTR(errno) if failed.
1822  */
1823 struct lmv_tgt_desc*
1824 lmv_locate_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
1825                struct lu_fid *fid)
1826 {
1827         struct lmv_stripe_md    *lsm = op_data->op_mea1;
1828         struct lmv_tgt_desc     *tgt;
1829
1830         /* During creating VOLATILE file, it should honor the mdt
1831          * index if the file under striped dir is being restored, see
1832          * ct_restore(). */
1833         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1834             (int)op_data->op_mds != -1 && lsm != NULL) {
1835                 int i;
1836                 tgt = lmv_get_target(lmv, op_data->op_mds, NULL);
1837                 if (IS_ERR(tgt))
1838                         return tgt;
1839
1840                 /* refill the right parent fid */
1841                 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1842                         struct lmv_oinfo *oinfo;
1843
1844                         oinfo = &lsm->lsm_md_oinfo[i];
1845                         if (oinfo->lmo_mds == op_data->op_mds) {
1846                                 *fid = oinfo->lmo_fid;
1847                                 break;
1848                         }
1849                 }
1850
1851                 /* Hmm, can not find the stripe by mdt_index(op_mds) */
1852                 if (i == lsm->lsm_md_stripe_count)
1853                         tgt = ERR_PTR(-EINVAL);
1854
1855                 return tgt;
1856         }
1857
1858         if (lsm == NULL || op_data->op_namelen == 0) {
1859                 tgt = lmv_find_target(lmv, fid);
1860                 if (IS_ERR(tgt))
1861                         return tgt;
1862
1863                 op_data->op_mds = tgt->ltd_idx;
1864                 return tgt;
1865         }
1866
1867         return lmv_locate_target_for_name(lmv, lsm, op_data->op_name,
1868                                           op_data->op_namelen, fid,
1869                                           &op_data->op_mds);
1870 }
1871
1872 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1873                const void *data, int datalen, int mode, __u32 uid,
1874                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1875                struct ptlrpc_request **request)
1876 {
1877         struct obd_device       *obd = exp->exp_obd;
1878         struct lmv_obd          *lmv = &obd->u.lmv;
1879         struct lmv_tgt_desc     *tgt;
1880         int                      rc;
1881         ENTRY;
1882
1883         rc = lmv_check_connect(obd);
1884         if (rc)
1885                 RETURN(rc);
1886
1887         if (!lmv->desc.ld_active_tgt_count)
1888                 RETURN(-EIO);
1889
1890         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1891         if (IS_ERR(tgt))
1892                 RETURN(PTR_ERR(tgt));
1893
1894         CDEBUG(D_INODE, "CREATE name '%.*s' on "DFID" -> mds #%x\n",
1895                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1896                op_data->op_mds);
1897
1898         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1899         if (rc)
1900                 RETURN(rc);
1901
1902         /* Send the create request to the MDT where the object
1903          * will be located */
1904         tgt = lmv_find_target(lmv, &op_data->op_fid2);
1905         if (IS_ERR(tgt))
1906                 RETURN(PTR_ERR(tgt));
1907
1908         op_data->op_mds = tgt->ltd_idx;
1909
1910         CDEBUG(D_INODE, "CREATE obj "DFID" -> mds #%x\n",
1911                PFID(&op_data->op_fid2), op_data->op_mds);
1912
1913         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1914         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1915                        cap_effective, rdev, request);
1916         if (rc == 0) {
1917                 if (*request == NULL)
1918                         RETURN(rc);
1919                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1920         }
1921         RETURN(rc);
1922 }
1923
1924 static int lmv_done_writing(struct obd_export *exp,
1925                             struct md_op_data *op_data,
1926                             struct md_open_data *mod)
1927 {
1928         struct obd_device     *obd = exp->exp_obd;
1929         struct lmv_obd        *lmv = &obd->u.lmv;
1930         struct lmv_tgt_desc   *tgt;
1931         int                    rc;
1932         ENTRY;
1933
1934         rc = lmv_check_connect(obd);
1935         if (rc)
1936                 RETURN(rc);
1937
1938         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1939         if (IS_ERR(tgt))
1940                 RETURN(PTR_ERR(tgt));
1941
1942         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1943         RETURN(rc);
1944 }
1945
1946 static int
1947 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1948             const union ldlm_policy_data *policy,
1949             struct lookup_intent *it, struct md_op_data *op_data,
1950             struct lustre_handle *lockh, __u64 extra_lock_flags)
1951 {
1952         struct obd_device        *obd = exp->exp_obd;
1953         struct lmv_obd           *lmv = &obd->u.lmv;
1954         struct lmv_tgt_desc      *tgt;
1955         int                       rc;
1956         ENTRY;
1957
1958         rc = lmv_check_connect(obd);
1959         if (rc)
1960                 RETURN(rc);
1961
1962         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1963                LL_IT2STR(it), PFID(&op_data->op_fid1));
1964
1965         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1966         if (IS_ERR(tgt))
1967                 RETURN(PTR_ERR(tgt));
1968
1969         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1970                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1971
1972         rc = md_enqueue(tgt->ltd_exp, einfo, policy, it, op_data, lockh,
1973                         extra_lock_flags);
1974
1975         RETURN(rc);
1976 }
1977
1978 static int
1979 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1980                  struct ptlrpc_request **preq)
1981 {
1982         struct ptlrpc_request   *req = NULL;
1983         struct obd_device       *obd = exp->exp_obd;
1984         struct lmv_obd          *lmv = &obd->u.lmv;
1985         struct lmv_tgt_desc     *tgt;
1986         struct mdt_body         *body;
1987         int                      rc;
1988         ENTRY;
1989
1990         rc = lmv_check_connect(obd);
1991         if (rc)
1992                 RETURN(rc);
1993
1994         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1995         if (IS_ERR(tgt))
1996                 RETURN(PTR_ERR(tgt));
1997
1998         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1999                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
2000                tgt->ltd_idx);
2001
2002         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2003         if (rc != 0)
2004                 RETURN(rc);
2005
2006         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2007         LASSERT(body != NULL);
2008
2009         if (body->mbo_valid & OBD_MD_MDS) {
2010                 struct lu_fid rid = body->mbo_fid1;
2011                 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
2012                        PFID(&rid));
2013
2014                 tgt = lmv_find_target(lmv, &rid);
2015                 if (IS_ERR(tgt)) {
2016                         ptlrpc_req_finished(*preq);
2017                         preq = NULL;
2018                         RETURN(PTR_ERR(tgt));
2019                 }
2020
2021                 op_data->op_fid1 = rid;
2022                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2023                 op_data->op_namelen = 0;
2024                 op_data->op_name = NULL;
2025                 rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
2026                 ptlrpc_req_finished(*preq);
2027                 *preq = req;
2028         }
2029
2030         RETURN(rc);
2031 }
2032
2033 #define md_op_data_fid(op_data, fl)                     \
2034         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2035          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2036          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2037          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2038          NULL)
2039
2040 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2041                             struct md_op_data *op_data,
2042                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
2043 {
2044         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
2045         struct obd_device      *obd = exp->exp_obd;
2046         struct lmv_obd         *lmv = &obd->u.lmv;
2047         ldlm_policy_data_t      policy = {{ 0 }};
2048         int                     rc = 0;
2049         ENTRY;
2050
2051         if (!fid_is_sane(fid))
2052                 RETURN(0);
2053
2054         if (tgt == NULL) {
2055                 tgt = lmv_find_target(lmv, fid);
2056                 if (IS_ERR(tgt))
2057                         RETURN(PTR_ERR(tgt));
2058         }
2059
2060         if (tgt->ltd_idx != op_tgt) {
2061                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2062                 policy.l_inodebits.bits = bits;
2063                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2064                                       mode, LCF_ASYNC, NULL);
2065         } else {
2066                 CDEBUG(D_INODE,
2067                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2068                        op_tgt, PFID(fid));
2069                 op_data->op_flags |= flag;
2070                 rc = 0;
2071         }
2072
2073         RETURN(rc);
2074 }
2075
2076 /*
2077  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2078  * op_data->op_fid2
2079  */
2080 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2081                     struct ptlrpc_request **request)
2082 {
2083         struct obd_device       *obd = exp->exp_obd;
2084         struct lmv_obd          *lmv = &obd->u.lmv;
2085         struct lmv_tgt_desc     *tgt;
2086         int                      rc;
2087         ENTRY;
2088
2089         rc = lmv_check_connect(obd);
2090         if (rc)
2091                 RETURN(rc);
2092
2093         LASSERT(op_data->op_namelen != 0);
2094
2095         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2096                PFID(&op_data->op_fid2), op_data->op_namelen,
2097                op_data->op_name, PFID(&op_data->op_fid1));
2098
2099         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2100         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2101         op_data->op_cap = cfs_curproc_cap_pack();
2102         if (op_data->op_mea2 != NULL) {
2103                 struct lmv_stripe_md    *lsm = op_data->op_mea2;
2104                 const struct lmv_oinfo  *oinfo;
2105
2106                 oinfo = lsm_name_to_stripe_info(lsm, op_data->op_name,
2107                                                 op_data->op_namelen);
2108                 if (IS_ERR(oinfo))
2109                         RETURN(PTR_ERR(oinfo));
2110
2111                 op_data->op_fid2 = oinfo->lmo_fid;
2112         }
2113
2114         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid2);
2115         if (IS_ERR(tgt))
2116                 RETURN(PTR_ERR(tgt));
2117
2118         /*
2119          * Cancel UPDATE lock on child (fid1).
2120          */
2121         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2122         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
2123                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2124         if (rc != 0)
2125                 RETURN(rc);
2126
2127         rc = md_link(tgt->ltd_exp, op_data, request);
2128
2129         RETURN(rc);
2130 }
2131
2132 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2133                       const char *old, int oldlen, const char *new, int newlen,
2134                       struct ptlrpc_request **request)
2135 {
2136         struct obd_device       *obd = exp->exp_obd;
2137         struct lmv_obd          *lmv = &obd->u.lmv;
2138         struct lmv_tgt_desc     *src_tgt;
2139         int                     rc;
2140         ENTRY;
2141
2142         LASSERT(oldlen != 0);
2143
2144         CDEBUG(D_INODE, "RENAME %.*s in "DFID":%d to %.*s in "DFID":%d\n",
2145                oldlen, old, PFID(&op_data->op_fid1),
2146                op_data->op_mea1 ? op_data->op_mea1->lsm_md_stripe_count : 0,
2147                newlen, new, PFID(&op_data->op_fid2),
2148                op_data->op_mea2 ? op_data->op_mea2->lsm_md_stripe_count : 0);
2149
2150         rc = lmv_check_connect(obd);
2151         if (rc)
2152                 RETURN(rc);
2153
2154         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2155         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2156         op_data->op_cap = cfs_curproc_cap_pack();
2157         if (op_data->op_cli_flags & CLI_MIGRATE) {
2158                 LASSERTF(fid_is_sane(&op_data->op_fid3), "invalid FID "DFID"\n",
2159                          PFID(&op_data->op_fid3));
2160                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
2161                 if (rc)
2162                         RETURN(rc);
2163                 src_tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid3);
2164         } else {
2165                 if (op_data->op_mea1 != NULL) {
2166                         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2167
2168                         src_tgt = lmv_locate_target_for_name(lmv, lsm, old,
2169                                                              oldlen,
2170                                                              &op_data->op_fid1,
2171                                                              &op_data->op_mds);
2172                         if (IS_ERR(src_tgt))
2173                                 RETURN(PTR_ERR(src_tgt));
2174                 } else {
2175                         src_tgt = lmv_find_target(lmv, &op_data->op_fid1);
2176                         if (IS_ERR(src_tgt))
2177                                 RETURN(PTR_ERR(src_tgt));
2178
2179                         op_data->op_mds = src_tgt->ltd_idx;
2180                 }
2181
2182                 if (op_data->op_mea2) {
2183                         struct lmv_stripe_md    *lsm = op_data->op_mea2;
2184                         const struct lmv_oinfo  *oinfo;
2185
2186                         oinfo = lsm_name_to_stripe_info(lsm, new, newlen);
2187                         if (IS_ERR(oinfo))
2188                                 RETURN(PTR_ERR(oinfo));
2189
2190                         op_data->op_fid2 = oinfo->lmo_fid;
2191                 }
2192         }
2193         if (IS_ERR(src_tgt))
2194                 RETURN(PTR_ERR(src_tgt));
2195
2196         /*
2197          * LOOKUP lock on src child (fid3) should also be cancelled for
2198          * src_tgt in mdc_rename.
2199          */
2200         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2201
2202         /*
2203          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2204          * own target.
2205          */
2206         rc = lmv_early_cancel(exp, NULL, op_data, src_tgt->ltd_idx,
2207                               LCK_EX, MDS_INODELOCK_UPDATE,
2208                               MF_MDC_CANCEL_FID2);
2209
2210         if (rc != 0)
2211                 RETURN(rc);
2212         /*
2213          * Cancel LOOKUP locks on source child (fid3) for parent tgt_tgt.
2214          */
2215         if (fid_is_sane(&op_data->op_fid3)) {
2216                 struct lmv_tgt_desc *tgt;
2217
2218                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2219                 if (IS_ERR(tgt))
2220                         RETURN(PTR_ERR(tgt));
2221
2222                 /* Cancel LOOKUP lock on its parent */
2223                 rc = lmv_early_cancel(exp, tgt, op_data, src_tgt->ltd_idx,
2224                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2225                                       MF_MDC_CANCEL_FID3);
2226                 if (rc != 0)
2227                         RETURN(rc);
2228
2229                 rc = lmv_early_cancel(exp, NULL, op_data, src_tgt->ltd_idx,
2230                                       LCK_EX, MDS_INODELOCK_FULL,
2231                                       MF_MDC_CANCEL_FID3);
2232                 if (rc != 0)
2233                         RETURN(rc);
2234         }
2235
2236         /*
2237          * Cancel all the locks on tgt child (fid4).
2238          */
2239         if (fid_is_sane(&op_data->op_fid4))
2240                 rc = lmv_early_cancel(exp, NULL, op_data, src_tgt->ltd_idx,
2241                                       LCK_EX, MDS_INODELOCK_FULL,
2242                                       MF_MDC_CANCEL_FID4);
2243
2244         CDEBUG(D_INODE, DFID":m%d to "DFID"\n", PFID(&op_data->op_fid1),
2245                op_data->op_mds, PFID(&op_data->op_fid2));
2246
2247         rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2248                        request);
2249
2250         RETURN(rc);
2251 }
2252
2253 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2254                        void *ea, int ealen, void *ea2, int ea2len,
2255                        struct ptlrpc_request **request,
2256                        struct md_open_data **mod)
2257 {
2258         struct obd_device       *obd = exp->exp_obd;
2259         struct lmv_obd          *lmv = &obd->u.lmv;
2260         struct lmv_tgt_desc     *tgt;
2261         int                      rc = 0;
2262         ENTRY;
2263
2264         rc = lmv_check_connect(obd);
2265         if (rc)
2266                 RETURN(rc);
2267
2268         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x\n",
2269                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid);
2270
2271         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2272         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2273         if (IS_ERR(tgt))
2274                 RETURN(PTR_ERR(tgt));
2275
2276         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2277                         ea2len, request, mod);
2278
2279         RETURN(rc);
2280 }
2281
2282 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2283                      struct obd_capa *oc, struct ptlrpc_request **request)
2284 {
2285         struct obd_device       *obd = exp->exp_obd;
2286         struct lmv_obd          *lmv = &obd->u.lmv;
2287         struct lmv_tgt_desc     *tgt;
2288         int                      rc;
2289         ENTRY;
2290
2291         rc = lmv_check_connect(obd);
2292         if (rc != 0)
2293                 RETURN(rc);
2294
2295         tgt = lmv_find_target(lmv, fid);
2296         if (IS_ERR(tgt))
2297                 RETURN(PTR_ERR(tgt));
2298
2299         rc = md_fsync(tgt->ltd_exp, fid, oc, request);
2300         RETURN(rc);
2301 }
2302
2303 /**
2304  * Get current minimum entry from striped directory
2305  *
2306  * This function will search the dir entry, whose hash value is the
2307  * closest(>=) to @hash_offset, from all of sub-stripes, and it is
2308  * only being called for striped directory.
2309  *
2310  * \param[in] exp               export of LMV
2311  * \param[in] op_data           parameters transferred beween client MD stack
2312  *                              stripe_information will be included in this
2313  *                              parameter
2314  * \param[in] cb_op             ldlm callback being used in enqueue in
2315  *                              mdc_read_page
2316  * \param[in] hash_offset       the hash value, which is used to locate
2317  *                              minum(closet) dir entry
2318  * \param[in|out] stripe_offset the caller use this to indicate the stripe
2319  *                              index of last entry, so to avoid hash conflict
2320  *                              between stripes. It will also be used to
2321  *                              return the stripe index of current dir entry.
2322  * \param[in|out] entp          the minum entry and it also is being used
2323  *                              to input the last dir entry to resolve the
2324  *                              hash conflict
2325  *
2326  * \param[out] ppage            the page which holds the minum entry
2327  *
2328  * \retval                      = 0 get the entry successfully
2329  *                              negative errno (< 0) does not get the entry
2330  */
2331 static int lmv_get_min_striped_entry(struct obd_export *exp,
2332                                      struct md_op_data *op_data,
2333                                      struct md_callback *cb_op,
2334                                      __u64 hash_offset, int *stripe_offset,
2335                                      struct lu_dirent **entp,
2336                                      struct page **ppage)
2337 {
2338         struct obd_device       *obd = exp->exp_obd;
2339         struct lmv_obd          *lmv = &obd->u.lmv;
2340         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2341         struct lmv_tgt_desc     *tgt;
2342         int                     stripe_count;
2343         struct lu_dirent        *min_ent = NULL;
2344         struct page             *min_page = NULL;
2345         int                     min_idx = 0;
2346         int                     i;
2347         int                     rc = 0;
2348         ENTRY;
2349
2350         stripe_count = lsm->lsm_md_stripe_count;
2351         for (i = 0; i < stripe_count; i++) {
2352                 struct lu_dirent        *ent = NULL;
2353                 struct page             *page = NULL;
2354                 struct lu_dirpage       *dp;
2355                 __u64                   stripe_hash = hash_offset;
2356
2357                 tgt = lmv_get_target(lmv, lsm->lsm_md_oinfo[i].lmo_mds, NULL);
2358                 if (IS_ERR(tgt))
2359                         GOTO(out, rc = PTR_ERR(tgt));
2360
2361                 /* op_data will be shared by each stripe, so we need
2362                  * reset these value for each stripe */
2363                 op_data->op_stripe_offset = i;
2364                 op_data->op_fid1 = lsm->lsm_md_oinfo[i].lmo_fid;
2365                 op_data->op_fid2 = lsm->lsm_md_oinfo[i].lmo_fid;
2366                 op_data->op_data = lsm->lsm_md_oinfo[i].lmo_root;
2367 next:
2368                 rc = md_read_page(tgt->ltd_exp, op_data, cb_op, stripe_hash,
2369                                   &page);
2370                 if (rc != 0)
2371                         GOTO(out, rc);
2372
2373                 dp = page_address(page);
2374                 for (ent = lu_dirent_start(dp); ent != NULL;
2375                      ent = lu_dirent_next(ent)) {
2376                         /* Skip dummy entry */
2377                         if (le16_to_cpu(ent->lde_namelen) == 0)
2378                                 continue;
2379
2380                         if (le64_to_cpu(ent->lde_hash) < hash_offset)
2381                                 continue;
2382
2383                         if (le64_to_cpu(ent->lde_hash) == hash_offset &&
2384                             (*entp == ent || i < *stripe_offset))
2385                                 continue;
2386
2387                         /* skip . and .. for other stripes */
2388                         if (i != 0 &&
2389                             (strncmp(ent->lde_name, ".",
2390                                      le16_to_cpu(ent->lde_namelen)) == 0 ||
2391                              strncmp(ent->lde_name, "..",
2392                                      le16_to_cpu(ent->lde_namelen)) == 0))
2393                                 continue;
2394                         break;
2395                 }
2396
2397                 if (ent == NULL) {
2398                         stripe_hash = le64_to_cpu(dp->ldp_hash_end);
2399
2400                         kunmap(page);
2401                         page_cache_release(page);
2402                         page = NULL;
2403
2404                         /* reach the end of current stripe, go to next stripe */
2405                         if (stripe_hash == MDS_DIR_END_OFF)
2406                                 continue;
2407                         else
2408                                 goto next;
2409                 }
2410
2411                 if (min_ent != NULL) {
2412                         if (le64_to_cpu(min_ent->lde_hash) >
2413                             le64_to_cpu(ent->lde_hash)) {
2414                                 min_ent = ent;
2415                                 kunmap(min_page);
2416                                 page_cache_release(min_page);
2417                                 min_idx = i;
2418                                 min_page = page;
2419                         } else {
2420                                 kunmap(page);
2421                                 page_cache_release(page);
2422                                 page = NULL;
2423                         }
2424                 } else {
2425                         min_ent = ent;
2426                         min_page = page;
2427                         min_idx = i;
2428                 }
2429         }
2430
2431 out:
2432         if (*ppage != NULL) {
2433                 kunmap(*ppage);
2434                 page_cache_release(*ppage);
2435         }
2436         *stripe_offset = min_idx;
2437         *entp = min_ent;
2438         *ppage = min_page;
2439         RETURN(rc);
2440 }
2441
2442 /**
2443  * Build dir entry page from a striped directory
2444  *
2445  * This function gets one entry by @offset from a striped directory. It will
2446  * read entries from all of stripes, and choose one closest to the required
2447  * offset(&offset). A few notes
2448  * 1. skip . and .. for non-zero stripes, because there can only have one .
2449  * and .. in a directory.
2450  * 2. op_data will be shared by all of stripes, instead of allocating new
2451  * one, so need to restore before reusing.
2452  * 3. release the entry page if that is not being chosen.
2453  *
2454  * \param[in] exp       obd export refer to LMV
2455  * \param[in] op_data   hold those MD parameters of read_entry
2456  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2457  * \param[out] ldp      the entry being read
2458  * \param[out] ppage    the page holding the entry. Note: because the entry
2459  *                      will be accessed in upper layer, so we need hold the
2460  *                      page until the usages of entry is finished, see
2461  *                      ll_dir_entry_next.
2462  *
2463  * retval               =0 if get entry successfully
2464  *                      <0 cannot get entry
2465  */
2466 static int lmv_read_striped_page(struct obd_export *exp,
2467                                  struct md_op_data *op_data,
2468                                  struct md_callback *cb_op,
2469                                  __u64 offset, struct page **ppage)
2470 {
2471         struct obd_device       *obd = exp->exp_obd;
2472         struct lu_fid           master_fid = op_data->op_fid1;
2473         struct inode            *master_inode = op_data->op_data;
2474         __u64                   hash_offset = offset;
2475         struct lu_dirpage       *dp;
2476         struct page             *min_ent_page = NULL;
2477         struct page             *ent_page = NULL;
2478         struct lu_dirent        *ent;
2479         void                    *area;
2480         int                     ent_idx = 0;
2481         struct lu_dirent        *min_ent = NULL;
2482         struct lu_dirent        *last_ent;
2483         int                     left_bytes;
2484         int                     rc;
2485         ENTRY;
2486
2487         rc = lmv_check_connect(obd);
2488         if (rc)
2489                 RETURN(rc);
2490
2491         /* Allocate a page and read entries from all of stripes and fill
2492          * the page by hash order */
2493         ent_page = alloc_page(GFP_KERNEL);
2494         if (ent_page == NULL)
2495                 RETURN(-ENOMEM);
2496
2497         /* Initialize the entry page */
2498         dp = kmap(ent_page);
2499         memset(dp, 0, sizeof(*dp));
2500         dp->ldp_hash_start = cpu_to_le64(offset);
2501         dp->ldp_flags |= LDF_COLLIDE;
2502
2503         area = dp + 1;
2504         left_bytes = PAGE_CACHE_SIZE - sizeof(*dp);
2505         ent = area;
2506         last_ent = ent;
2507         do {
2508                 __u16   ent_size;
2509
2510                 /* Find the minum entry from all sub-stripes */
2511                 rc = lmv_get_min_striped_entry(exp, op_data, cb_op, hash_offset,
2512                                                &ent_idx, &min_ent,
2513                                                &min_ent_page);
2514                 if (rc != 0)
2515                         GOTO(out, rc);
2516
2517                 /* If it can not get minum entry, it means it already reaches
2518                  * the end of this directory */
2519                 if (min_ent == NULL) {
2520                         last_ent->lde_reclen = 0;
2521                         hash_offset = MDS_DIR_END_OFF;
2522                         GOTO(out, rc);
2523                 }
2524
2525                 ent_size = le16_to_cpu(min_ent->lde_reclen);
2526
2527                 /* the last entry lde_reclen is 0, but it might not
2528                  * the end of this entry of this temporay entry */
2529                 if (ent_size == 0)
2530                         ent_size = lu_dirent_calc_size(
2531                                         le16_to_cpu(min_ent->lde_namelen),
2532                                         le32_to_cpu(min_ent->lde_attrs));
2533                 if (ent_size > left_bytes) {
2534                         last_ent->lde_reclen = cpu_to_le16(0);
2535                         hash_offset = le64_to_cpu(min_ent->lde_hash);
2536                         GOTO(out, rc);
2537                 }
2538
2539                 memcpy(ent, min_ent, ent_size);
2540
2541                 /* Replace . with master FID and Replace .. with the parent FID
2542                  * of master object */
2543                 if (strncmp(ent->lde_name, ".",
2544                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2545                     le16_to_cpu(ent->lde_namelen) == 1)
2546                         fid_cpu_to_le(&ent->lde_fid, &master_fid);
2547                 else if (strncmp(ent->lde_name, "..",
2548                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2549                            le16_to_cpu(ent->lde_namelen) == 2)
2550                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2551
2552                 left_bytes -= ent_size;
2553                 ent->lde_reclen = cpu_to_le16(ent_size);
2554                 last_ent = ent;
2555                 ent = (void *)ent + ent_size;
2556                 hash_offset = le64_to_cpu(min_ent->lde_hash);
2557                 if (hash_offset == MDS_DIR_END_OFF) {
2558                         last_ent->lde_reclen = 0;
2559                         break;
2560                 }
2561         } while (1);
2562 out:
2563         if (min_ent_page != NULL) {
2564                 kunmap(min_ent_page);
2565                 page_cache_release(min_ent_page);
2566         }
2567
2568         if (unlikely(rc != 0)) {
2569                 __free_page(ent_page);
2570                 ent_page = NULL;
2571         } else {
2572                 if (ent == area)
2573                         dp->ldp_flags |= LDF_EMPTY;
2574                 dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2575                 dp->ldp_hash_end = cpu_to_le64(hash_offset);
2576         }
2577
2578         /* We do not want to allocate md_op_data during each
2579          * dir entry reading, so op_data will be shared by every stripe,
2580          * then we need to restore it back to original value before
2581          * return to the upper layer */
2582         op_data->op_fid1 = master_fid;
2583         op_data->op_fid2 = master_fid;
2584         op_data->op_data = master_inode;
2585
2586         *ppage = ent_page;
2587
2588         RETURN(rc);
2589 }
2590
2591 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2592                   struct md_callback *cb_op, __u64 offset,
2593                   struct page **ppage)
2594 {
2595         struct obd_device       *obd = exp->exp_obd;
2596         struct lmv_obd          *lmv = &obd->u.lmv;
2597         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2598         struct lmv_tgt_desc     *tgt;
2599         int                     rc;
2600         ENTRY;
2601
2602         rc = lmv_check_connect(obd);
2603         if (rc != 0)
2604                 RETURN(rc);
2605
2606         if (unlikely(lsm != NULL)) {
2607                 rc = lmv_read_striped_page(exp, op_data, cb_op, offset, ppage);
2608                 RETURN(rc);
2609         }
2610
2611         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2612         if (IS_ERR(tgt))
2613                 RETURN(PTR_ERR(tgt));
2614
2615         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2616
2617         RETURN(rc);
2618 }
2619
2620 /**
2621  * Unlink a file/directory
2622  *
2623  * Unlink a file or directory under the parent dir. The unlink request
2624  * usually will be sent to the MDT where the child is located, but if
2625  * the client does not have the child FID then request will be sent to the
2626  * MDT where the parent is located.
2627  *
2628  * If the parent is a striped directory then it also needs to locate which
2629  * stripe the name of the child is located, and replace the parent FID
2630  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2631  * it will walk through all of sub-stripes until the child is being
2632  * unlinked finally.
2633  *
2634  * \param[in] exp       export refer to LMV
2635  * \param[in] op_data   different parameters transferred beween client
2636  *                      MD stacks, name, namelen, FIDs etc.
2637  *                      op_fid1 is the parent FID, op_fid2 is the child
2638  *                      FID.
2639  * \param[out] request  point to the request of unlink.
2640  *
2641  * retval               0 if succeed
2642  *                      negative errno if failed.
2643  */
2644 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2645                       struct ptlrpc_request **request)
2646 {
2647         struct obd_device       *obd = exp->exp_obd;
2648         struct lmv_obd          *lmv = &obd->u.lmv;
2649         struct lmv_tgt_desc     *tgt = NULL;
2650         struct lmv_tgt_desc     *parent_tgt = NULL;
2651         struct mdt_body         *body;
2652         int                     rc;
2653         int                     stripe_index = 0;
2654         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2655         ENTRY;
2656
2657         rc = lmv_check_connect(obd);
2658         if (rc)
2659                 RETURN(rc);
2660 retry_unlink:
2661         /* For striped dir, we need to locate the parent as well */
2662         if (lsm != NULL) {
2663                 struct lmv_tgt_desc *tmp;
2664
2665                 LASSERT(op_data->op_name != NULL &&
2666                         op_data->op_namelen != 0);
2667
2668                 tmp = lmv_locate_target_for_name(lmv, lsm,
2669                                                  op_data->op_name,
2670                                                  op_data->op_namelen,
2671                                                  &op_data->op_fid1,
2672                                                  &op_data->op_mds);
2673
2674                 /* return -EBADFD means unknown hash type, might
2675                  * need try all sub-stripe here */
2676                 if (IS_ERR(tmp) && PTR_ERR(tmp) != -EBADFD)
2677                         RETURN(PTR_ERR(tmp));
2678
2679                 /* Note: both migrating dir and unknown hash dir need to
2680                  * try all of sub-stripes, so we need start search the
2681                  * name from stripe 0, but migrating dir is already handled
2682                  * inside lmv_locate_target_for_name(), so we only check
2683                  * unknown hash type directory here */
2684                 if (!lmv_is_known_hash_type(lsm)) {
2685                         struct lmv_oinfo *oinfo;
2686
2687                         oinfo = &lsm->lsm_md_oinfo[stripe_index];
2688
2689                         op_data->op_fid1 = oinfo->lmo_fid;
2690                         op_data->op_mds = oinfo->lmo_mds;
2691                 }
2692         }
2693
2694 try_next_stripe:
2695         /* Send unlink requests to the MDT where the child is located */
2696         if (likely(!fid_is_zero(&op_data->op_fid2)))
2697                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2698         else if (lsm != NULL)
2699                 tgt = lmv_get_target(lmv, op_data->op_mds, NULL);
2700         else
2701                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
2702
2703         if (IS_ERR(tgt))
2704                 RETURN(PTR_ERR(tgt));
2705
2706         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2707         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2708         op_data->op_cap = cfs_curproc_cap_pack();
2709
2710         /*
2711          * If child's fid is given, cancel unused locks for it if it is from
2712          * another export than parent.
2713          *
2714          * LOOKUP lock for child (fid3) should also be cancelled on parent
2715          * tgt_tgt in mdc_unlink().
2716          */
2717         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2718
2719         /*
2720          * Cancel FULL locks on child (fid3).
2721          */
2722         parent_tgt = lmv_find_target(lmv, &op_data->op_fid1);
2723         if (IS_ERR(parent_tgt))
2724                 RETURN(PTR_ERR(parent_tgt));
2725
2726         if (parent_tgt != tgt) {
2727                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_idx,
2728                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2729                                       MF_MDC_CANCEL_FID3);
2730         }
2731
2732         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
2733                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2734         if (rc != 0)
2735                 RETURN(rc);
2736
2737         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%d\n",
2738                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2), tgt->ltd_idx);
2739
2740         rc = md_unlink(tgt->ltd_exp, op_data, request);
2741         if (rc != 0 && rc != -EREMOTE && rc != -ENOENT)
2742                 RETURN(rc);
2743
2744         /* Try next stripe if it is needed. */
2745         if (rc == -ENOENT && lsm != NULL && lmv_need_try_all_stripes(lsm)) {
2746                 struct lmv_oinfo *oinfo;
2747
2748                 stripe_index++;
2749                 if (stripe_index >= lsm->lsm_md_stripe_count)
2750                         RETURN(rc);
2751
2752                 oinfo = &lsm->lsm_md_oinfo[stripe_index];
2753
2754                 op_data->op_fid1 = oinfo->lmo_fid;
2755                 op_data->op_mds = oinfo->lmo_mds;
2756
2757                 ptlrpc_req_finished(*request);
2758                 *request = NULL;
2759
2760                 goto try_next_stripe;
2761         }
2762
2763         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2764         if (body == NULL)
2765                 RETURN(-EPROTO);
2766
2767         /* Not cross-ref case, just get out of here. */
2768         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2769                 RETURN(rc);
2770
2771         CDEBUG(D_INODE, "%s: try unlink to another MDT for "DFID"\n",
2772                exp->exp_obd->obd_name, PFID(&body->mbo_fid1));
2773
2774         /* This is a remote object, try remote MDT, Note: it may
2775          * try more than 1 time here, Considering following case
2776          * /mnt/lustre is root on MDT0, remote1 is on MDT1
2777          * 1. Initially A does not know where remote1 is, it send
2778          *    unlink RPC to MDT0, MDT0 return -EREMOTE, it will
2779          *    resend unlink RPC to MDT1 (retry 1st time).
2780          *
2781          * 2. During the unlink RPC in flight,
2782          *    client B mv /mnt/lustre/remote1 /mnt/lustre/remote2
2783          *    and create new remote1, but on MDT0
2784          *
2785          * 3. MDT1 get unlink RPC(from A), then do remote lock on
2786          *    /mnt/lustre, then lookup get fid of remote1, and find
2787          *    it is remote dir again, and replay -EREMOTE again.
2788          *
2789          * 4. Then A will resend unlink RPC to MDT0. (retry 2nd times).
2790          *
2791          * In theory, it might try unlimited time here, but it should
2792          * be very rare case.  */
2793         op_data->op_fid2 = body->mbo_fid1;
2794         ptlrpc_req_finished(*request);
2795         *request = NULL;
2796
2797         goto retry_unlink;
2798 }
2799
2800 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2801 {
2802         struct lmv_obd *lmv = &obd->u.lmv;
2803         int rc = 0;
2804
2805         switch (stage) {
2806         case OBD_CLEANUP_EARLY:
2807                 /* XXX: here should be calling obd_precleanup() down to
2808                  * stack. */
2809                 break;
2810         case OBD_CLEANUP_EXPORTS:
2811                 fld_client_proc_fini(&lmv->lmv_fld);
2812                 lprocfs_obd_cleanup(obd);
2813                 lprocfs_free_md_stats(obd);
2814                 break;
2815         default:
2816                 break;
2817         }
2818         RETURN(rc);
2819 }
2820
2821 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2822                         __u32 keylen, void *key, __u32 *vallen, void *val,
2823                         struct lov_stripe_md *lsm)
2824 {
2825         struct obd_device       *obd;
2826         struct lmv_obd          *lmv;
2827         int                      rc = 0;
2828         ENTRY;
2829
2830         obd = class_exp2obd(exp);
2831         if (obd == NULL) {
2832                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2833                        exp->exp_handle.h_cookie);
2834                 RETURN(-EINVAL);
2835         }
2836
2837         lmv = &obd->u.lmv;
2838         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2839                 int i;
2840
2841                 rc = lmv_check_connect(obd);
2842                 if (rc)
2843                         RETURN(rc);
2844
2845                 LASSERT(*vallen == sizeof(__u32));
2846                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2847                         struct lmv_tgt_desc *tgt = lmv->tgts[i];
2848                         /*
2849                          * All tgts should be connected when this gets called.
2850                          */
2851                         if (tgt == NULL || tgt->ltd_exp == NULL)
2852                                 continue;
2853
2854                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2855                                           vallen, val, NULL))
2856                                 RETURN(0);
2857                 }
2858                 RETURN(-EINVAL);
2859         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2860                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2861                    KEY_IS(KEY_MAX_COOKIESIZE) ||
2862                    KEY_IS(KEY_DEFAULT_COOKIESIZE) ||
2863                    KEY_IS(KEY_CONN_DATA)) {
2864                 rc = lmv_check_connect(obd);
2865                 if (rc)
2866                         RETURN(rc);
2867
2868                 /*
2869                  * Forwarding this request to first MDS, it should know LOV
2870                  * desc.
2871                  */
2872                 rc = obd_get_info(env, lmv->tgts[0]->ltd_exp, keylen, key,
2873                                   vallen, val, NULL);
2874                 if (!rc && KEY_IS(KEY_CONN_DATA))
2875                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2876                 RETURN(rc);
2877         } else if (KEY_IS(KEY_TGT_COUNT)) {
2878                 *((int *)val) = lmv->desc.ld_tgt_count;
2879                 RETURN(0);
2880         }
2881
2882         CDEBUG(D_IOCTL, "Invalid key\n");
2883         RETURN(-EINVAL);
2884 }
2885
2886 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
2887                        obd_count keylen, void *key, obd_count vallen,
2888                        void *val, struct ptlrpc_request_set *set)
2889 {
2890         struct lmv_tgt_desc    *tgt = NULL;
2891         struct obd_device      *obd;
2892         struct lmv_obd         *lmv;
2893         int rc = 0;
2894         ENTRY;
2895
2896         obd = class_exp2obd(exp);
2897         if (obd == NULL) {
2898                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2899                        exp->exp_handle.h_cookie);
2900                 RETURN(-EINVAL);
2901         }
2902         lmv = &obd->u.lmv;
2903
2904         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX)) {
2905                 int i, err = 0;
2906
2907                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2908                         tgt = lmv->tgts[i];
2909
2910                         if (tgt == NULL || tgt->ltd_exp == NULL)
2911                                 continue;
2912
2913                         err = obd_set_info_async(env, tgt->ltd_exp,
2914                                                  keylen, key, vallen, val, set);
2915                         if (err && rc == 0)
2916                                 rc = err;
2917                 }
2918
2919                 RETURN(rc);
2920         }
2921
2922         RETURN(-EINVAL);
2923 }
2924
2925 static int lmv_pack_md_v1(const struct lmv_stripe_md *lsm,
2926                           struct lmv_mds_md_v1 *lmm1)
2927 {
2928         int     cplen;
2929         int     i;
2930
2931         lmm1->lmv_magic = cpu_to_le32(lsm->lsm_md_magic);
2932         lmm1->lmv_stripe_count = cpu_to_le32(lsm->lsm_md_stripe_count);
2933         lmm1->lmv_master_mdt_index = cpu_to_le32(lsm->lsm_md_master_mdt_index);
2934         lmm1->lmv_hash_type = cpu_to_le32(lsm->lsm_md_hash_type);
2935         cplen = strlcpy(lmm1->lmv_pool_name, lsm->lsm_md_pool_name,
2936                         sizeof(lmm1->lmv_pool_name));
2937         if (cplen >= sizeof(lmm1->lmv_pool_name))
2938                 return -E2BIG;
2939
2940         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
2941                 fid_cpu_to_le(&lmm1->lmv_stripe_fids[i],
2942                               &lsm->lsm_md_oinfo[i].lmo_fid);
2943         return 0;
2944 }
2945
2946 int lmv_pack_md(union lmv_mds_md **lmmp, const struct lmv_stripe_md *lsm,
2947                 int stripe_count)
2948 {
2949         int     lmm_size = 0;
2950         bool    allocated = false;
2951         int     rc = 0;
2952         ENTRY;
2953
2954         LASSERT(lmmp != NULL);
2955         /* Free lmm */
2956         if (*lmmp != NULL && lsm == NULL) {
2957                 int stripe_count;
2958
2959                 stripe_count = lmv_mds_md_stripe_count_get(*lmmp);
2960                 lmm_size = lmv_mds_md_size(stripe_count,
2961                                            le32_to_cpu((*lmmp)->lmv_magic));
2962                 if (lmm_size == 0)
2963                         RETURN(-EINVAL);
2964                 OBD_FREE(*lmmp, lmm_size);
2965                 *lmmp = NULL;
2966                 RETURN(0);
2967         }
2968
2969         /* Alloc lmm */
2970         if (*lmmp == NULL && lsm == NULL) {
2971                 lmm_size = lmv_mds_md_size(stripe_count, LMV_MAGIC);
2972                 LASSERT(lmm_size > 0);
2973                 OBD_ALLOC(*lmmp, lmm_size);
2974                 if (*lmmp == NULL)
2975                         RETURN(-ENOMEM);
2976                 lmv_mds_md_stripe_count_set(*lmmp, stripe_count);
2977                 (*lmmp)->lmv_magic = cpu_to_le32(LMV_MAGIC);
2978                 RETURN(lmm_size);
2979         }
2980
2981         /* pack lmm */
2982         LASSERT(lsm != NULL);
2983         lmm_size = lmv_mds_md_size(lsm->lsm_md_stripe_count, lsm->lsm_md_magic);
2984         if (*lmmp == NULL) {
2985                 OBD_ALLOC(*lmmp, lmm_size);
2986                 if (*lmmp == NULL)
2987                         RETURN(-ENOMEM);
2988                 allocated = true;
2989         }
2990
2991         switch (lsm->lsm_md_magic) {
2992         case LMV_MAGIC_V1:
2993                 rc = lmv_pack_md_v1(lsm, &(*lmmp)->lmv_md_v1);
2994                 break;
2995         default:
2996                 rc = -EINVAL;
2997                 break;
2998         }
2999
3000         if (rc != 0 && allocated) {
3001                 OBD_FREE(*lmmp, lmm_size);
3002                 *lmmp = NULL;
3003         }
3004
3005         RETURN(lmm_size);
3006 }
3007 EXPORT_SYMBOL(lmv_pack_md);
3008
3009 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3010                             const struct lmv_mds_md_v1 *lmm1)
3011 {
3012         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3013         int             stripe_count;
3014         int             cplen;
3015         int             i;
3016         int             rc = 0;
3017         ENTRY;
3018
3019         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3020         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3021         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3022         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3023                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3024         else
3025                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3026         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3027         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3028                         sizeof(lsm->lsm_md_pool_name));
3029
3030         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3031                 RETURN(-E2BIG);
3032
3033         CDEBUG(D_INFO, "unpack lsm count %d, master %d hash_type %d"
3034                "layout_version %d\n", lsm->lsm_md_stripe_count,
3035                lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
3036                lsm->lsm_md_layout_version);
3037
3038         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3039         for (i = 0; i < le32_to_cpu(stripe_count); i++) {
3040                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3041                               &lmm1->lmv_stripe_fids[i]);
3042                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3043                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3044                 if (rc != 0)
3045                         RETURN(rc);
3046                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3047                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3048         }
3049
3050         RETURN(rc);
3051 }
3052
3053 int lmv_unpack_md(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3054                   const union lmv_mds_md *lmm, int stripe_count)
3055 {
3056         struct lmv_stripe_md     *lsm;
3057         int                      lsm_size;
3058         int                      rc;
3059         bool                     allocated = false;
3060         ENTRY;
3061
3062         LASSERT(lsmp != NULL);
3063
3064         lsm = *lsmp;
3065         /* Free memmd */
3066         if (lsm != NULL && lmm == NULL) {
3067                 int i;
3068                 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3069                         /* For migrating inode, the master stripe and master
3070                          * object will be the same, so do not need iput, see
3071                          * ll_update_lsm_md */
3072                         if (!(lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION &&
3073                               i == 0) && lsm->lsm_md_oinfo[i].lmo_root != NULL)
3074                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
3075                 }
3076                 lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3077                 OBD_FREE(lsm, lsm_size);
3078                 *lsmp = NULL;
3079                 RETURN(0);
3080         }
3081
3082         /* Alloc memmd */
3083         if (lsm == NULL && lmm == NULL) {
3084                 lsm_size = lmv_stripe_md_size(stripe_count);
3085                 OBD_ALLOC(lsm, lsm_size);
3086                 if (lsm == NULL)
3087                         RETURN(-ENOMEM);
3088                 lsm->lsm_md_stripe_count = stripe_count;
3089                 *lsmp = lsm;
3090                 RETURN(0);
3091         }
3092
3093         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3094                 RETURN(-EPERM);
3095
3096         /* Unpack memmd */
3097         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3098             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3099                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3100                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3101                        -EIO);
3102                 RETURN(-EIO);
3103         }
3104
3105         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3106                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3107         else
3108                 /**
3109                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3110                  * stripecount should be 0 then.
3111                  */
3112                 lsm_size = lmv_stripe_md_size(0);
3113
3114         lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3115         if (lsm == NULL) {
3116                 OBD_ALLOC(lsm, lsm_size);
3117                 if (lsm == NULL)
3118                         RETURN(-ENOMEM);
3119                 allocated = true;
3120                 *lsmp = lsm;
3121         }
3122
3123         switch (le32_to_cpu(lmm->lmv_magic)) {
3124         case LMV_MAGIC_V1:
3125                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3126                 break;
3127         default:
3128                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3129                        le32_to_cpu(lmm->lmv_magic));
3130                 rc = -EINVAL;
3131                 break;
3132         }
3133
3134         if (rc != 0 && allocated) {
3135                 OBD_FREE(lsm, lsm_size);
3136                 *lsmp = NULL;
3137                 lsm_size = rc;
3138         }
3139         RETURN(lsm_size);
3140 }
3141
3142 int lmv_alloc_memmd(struct lmv_stripe_md **lsmp, int stripes)
3143 {
3144         return lmv_unpack_md(NULL, lsmp, NULL, stripes);
3145 }
3146 EXPORT_SYMBOL(lmv_alloc_memmd);
3147
3148 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3149 {
3150         lmv_unpack_md(NULL, &lsm, NULL, 0);
3151 }
3152 EXPORT_SYMBOL(lmv_free_memmd);
3153
3154 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3155                  struct lov_mds_md *lmm, int disk_len)
3156 {
3157         return lmv_unpack_md(exp, (struct lmv_stripe_md **)lsmp,
3158                              (union lmv_mds_md *)lmm, disk_len);
3159 }
3160
3161 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
3162                struct lov_stripe_md *lsm)
3163 {
3164         struct obd_device               *obd = exp->exp_obd;
3165         struct lmv_obd                  *lmv_obd = &obd->u.lmv;
3166         const struct lmv_stripe_md      *lmv = (struct lmv_stripe_md *)lsm;
3167         int                             stripe_count;
3168
3169         if (lmmp == NULL) {
3170                 if (lsm != NULL)
3171                         stripe_count = lmv->lsm_md_stripe_count;
3172                 else
3173                         stripe_count = lmv_obd->desc.ld_tgt_count;
3174
3175                 return lmv_mds_md_size(stripe_count, LMV_MAGIC_V1);
3176         }
3177
3178         return lmv_pack_md((union lmv_mds_md **)lmmp, lmv, 0);
3179 }
3180
3181 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3182                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
3183                              ldlm_cancel_flags_t flags, void *opaque)
3184 {
3185         struct obd_device       *obd = exp->exp_obd;
3186         struct lmv_obd          *lmv = &obd->u.lmv;
3187         int                      rc = 0;
3188         int                      err;
3189         __u32                    i;
3190         ENTRY;
3191
3192         LASSERT(fid != NULL);
3193
3194         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3195                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
3196
3197                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3198                         continue;
3199
3200                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3201                                        opaque);
3202                 if (!rc)
3203                         rc = err;
3204         }
3205         RETURN(rc);
3206 }
3207
3208 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
3209                       __u64 *bits)
3210 {
3211         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3212         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3213         int                      rc;
3214         ENTRY;
3215
3216         if (tgt == NULL || tgt->ltd_exp == NULL)