Whamcloud - gitweb
LU-13437 mdt: don't fetch LOOKUP lock for remote object
[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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  */
32
33 #define DEBUG_SUBSYSTEM S_LMV
34
35 #include <linux/file.h>
36 #include <linux/module.h>
37 #include <linux/init.h>
38 #include <linux/user_namespace.h>
39 #include <linux/uidgid.h>
40 #include <linux/slab.h>
41 #include <linux/pagemap.h>
42 #include <linux/mm.h>
43 #include <linux/math64.h>
44 #include <linux/seq_file.h>
45 #include <linux/namei.h>
46
47 #include <obd_support.h>
48 #include <lustre_lib.h>
49 #include <lustre_net.h>
50 #include <obd_class.h>
51 #include <lustre_lmv.h>
52 #include <lprocfs_status.h>
53 #include <cl_object.h>
54 #include <lustre_fid.h>
55 #include <uapi/linux/lustre/lustre_ioctl.h>
56 #include <lustre_kernelcomm.h>
57 #include "lmv_internal.h"
58
59 static int lmv_check_connect(struct obd_device *obd);
60
61 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
62                          int activate)
63 {
64         if (tgt->ltd_active == activate)
65                 return;
66
67         tgt->ltd_active = activate;
68         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
69                 (activate ? 1 : -1);
70
71         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
72 }
73
74 /**
75  * Error codes:
76  *
77  *  -EINVAL  : UUID can't be found in the LMV's target list
78  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
79  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
80  */
81 static int lmv_set_mdc_active(struct lmv_obd *lmv,
82                               const struct obd_uuid *uuid,
83                               int activate)
84 {
85         struct lu_tgt_desc *tgt = NULL;
86         struct obd_device *obd;
87         int rc = 0;
88
89         ENTRY;
90
91         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
92                         lmv, uuid->uuid, activate);
93
94         spin_lock(&lmv->lmv_lock);
95         lmv_foreach_connected_tgt(lmv, tgt) {
96                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
97                        tgt->ltd_index, tgt->ltd_uuid.uuid,
98                        tgt->ltd_exp->exp_handle.h_cookie);
99
100                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
101                         break;
102         }
103
104         if (!tgt)
105                 GOTO(out_lmv_lock, rc = -EINVAL);
106
107         obd = class_exp2obd(tgt->ltd_exp);
108         if (obd == NULL)
109                 GOTO(out_lmv_lock, rc = -ENOTCONN);
110
111         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
112                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
113                obd->obd_type->typ_name, tgt->ltd_index);
114         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
115
116         if (tgt->ltd_active == activate) {
117                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
118                        activate ? "" : "in");
119                 GOTO(out_lmv_lock, rc);
120         }
121
122         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
123                activate ? "" : "in");
124         lmv_activate_target(lmv, tgt, activate);
125         EXIT;
126
127  out_lmv_lock:
128         spin_unlock(&lmv->lmv_lock);
129         return rc;
130 }
131
132 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
133 {
134         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
135         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
136
137         return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
138 }
139
140 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
141                       enum obd_notify_event ev)
142 {
143         struct obd_connect_data *conn_data;
144         struct lmv_obd          *lmv = &obd->u.lmv;
145         struct obd_uuid         *uuid;
146         int                      rc = 0;
147         ENTRY;
148
149         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
150                 CERROR("unexpected notification of %s %s!\n",
151                        watched->obd_type->typ_name,
152                        watched->obd_name);
153                 RETURN(-EINVAL);
154         }
155
156         uuid = &watched->u.cli.cl_target_uuid;
157         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
158                 /*
159                  * Set MDC as active before notifying the observer, so the
160                  * observer can use the MDC normally.
161                  */
162                 rc = lmv_set_mdc_active(lmv, uuid,
163                                         ev == OBD_NOTIFY_ACTIVE);
164                 if (rc) {
165                         CERROR("%sactivation of %s failed: %d\n",
166                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
167                                uuid->uuid, rc);
168                         RETURN(rc);
169                 }
170         } else if (ev == OBD_NOTIFY_OCD) {
171                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
172                 /*
173                  * XXX: Make sure that ocd_connect_flags from all targets are
174                  * the same. Otherwise one of MDTs runs wrong version or
175                  * something like this.  --umka
176                  */
177                 obd->obd_self_export->exp_connect_data = *conn_data;
178         }
179
180         /*
181          * Pass the notification up the chain.
182          */
183         if (obd->obd_observer)
184                 rc = obd_notify(obd->obd_observer, watched, ev);
185
186         RETURN(rc);
187 }
188
189 static int lmv_connect(const struct lu_env *env,
190                        struct obd_export **pexp, struct obd_device *obd,
191                        struct obd_uuid *cluuid, struct obd_connect_data *data,
192                        void *localdata)
193 {
194         struct lmv_obd *lmv = &obd->u.lmv;
195         struct lustre_handle conn = { 0 };
196         struct obd_export *exp;
197         int rc;
198         ENTRY;
199
200         rc = class_connect(&conn, obd, cluuid);
201         if (rc) {
202                 CERROR("class_connection() returned %d\n", rc);
203                 RETURN(rc);
204         }
205
206         exp = class_conn2export(&conn);
207
208         lmv->connected = 0;
209         lmv->conn_data = *data;
210         lmv->lmv_cache = localdata;
211
212         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
213                                                     &obd->obd_kset.kobj);
214         if (!lmv->lmv_tgts_kobj) {
215                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
216                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
217         }
218
219         rc = lmv_check_connect(obd);
220         if (rc != 0)
221                 GOTO(out_sysfs, rc);
222
223         *pexp = exp;
224
225         RETURN(rc);
226
227 out_sysfs:
228         if (lmv->lmv_tgts_kobj)
229                 kobject_put(lmv->lmv_tgts_kobj);
230
231         class_disconnect(exp);
232
233         return rc;
234 }
235
236 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
237                             __u32 def_easize)
238 {
239         struct obd_device *obd = exp->exp_obd;
240         struct lmv_obd *lmv = &obd->u.lmv;
241         struct lmv_tgt_desc *tgt;
242         int change = 0;
243         int rc = 0;
244
245         ENTRY;
246
247         if (lmv->max_easize < easize) {
248                 lmv->max_easize = easize;
249                 change = 1;
250         }
251         if (lmv->max_def_easize < def_easize) {
252                 lmv->max_def_easize = def_easize;
253                 change = 1;
254         }
255
256         if (change == 0)
257                 RETURN(0);
258
259         if (lmv->connected == 0)
260                 RETURN(0);
261
262         lmv_foreach_connected_tgt(lmv, tgt) {
263                 if (!tgt->ltd_active)
264                         continue;
265
266                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
267                 if (rc) {
268                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
269                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
270                         break;
271                 }
272         }
273         RETURN(rc);
274 }
275
276 #define MAX_STRING_SIZE 128
277
278 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
279 {
280         struct lmv_obd          *lmv = &obd->u.lmv;
281         struct obd_device       *mdc_obd;
282         struct obd_export       *mdc_exp;
283         struct lu_fld_target     target;
284         int                      rc;
285         ENTRY;
286
287         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
288                                         &obd->obd_uuid);
289         if (!mdc_obd) {
290                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
291                 RETURN(-EINVAL);
292         }
293
294         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
295                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
296                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
297
298         if (!mdc_obd->obd_set_up) {
299                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
300                 RETURN(-EINVAL);
301         }
302
303         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
304                          &lmv->conn_data, lmv->lmv_cache);
305         if (rc) {
306                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
307                 RETURN(rc);
308         }
309
310         /*
311          * Init fid sequence client for this mdc and add new fld target.
312          */
313         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
314         if (rc)
315                 RETURN(rc);
316
317         target.ft_srv = NULL;
318         target.ft_exp = mdc_exp;
319         target.ft_idx = tgt->ltd_index;
320
321         fld_client_add_target(&lmv->lmv_fld, &target);
322
323         rc = obd_register_observer(mdc_obd, obd);
324         if (rc) {
325                 obd_disconnect(mdc_exp);
326                 CERROR("target %s register_observer error %d\n",
327                        tgt->ltd_uuid.uuid, rc);
328                 RETURN(rc);
329         }
330
331         if (obd->obd_observer) {
332                 /*
333                  * Tell the observer about the new target.
334                  */
335                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
336                                 OBD_NOTIFY_ACTIVE);
337                 if (rc) {
338                         obd_disconnect(mdc_exp);
339                         RETURN(rc);
340                 }
341         }
342
343         tgt->ltd_active = 1;
344         tgt->ltd_exp = mdc_exp;
345         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
346
347         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
348
349         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
350         if (rc) {
351                 obd_disconnect(mdc_exp);
352                 RETURN(rc);
353         }
354
355         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
356                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
357                 atomic_read(&obd->obd_refcount));
358
359         lmv_statfs_check_update(obd, tgt);
360
361         if (lmv->lmv_tgts_kobj)
362                 /* Even if we failed to create the link, that's fine */
363                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
364                                        &mdc_obd->obd_kset.kobj,
365                                        mdc_obd->obd_name);
366         RETURN(0);
367 }
368
369 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
370 {
371         LASSERT(tgt);
372         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
373         OBD_FREE_PTR(tgt);
374 }
375
376 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
377                            __u32 index, int gen)
378 {
379         struct obd_device *mdc_obd;
380         struct lmv_obd *lmv = &obd->u.lmv;
381         struct lmv_tgt_desc *tgt;
382         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
383         int rc = 0;
384
385         ENTRY;
386
387         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
388         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
389                                         &obd->obd_uuid);
390         if (!mdc_obd) {
391                 CERROR("%s: Target %s not attached: rc = %d\n",
392                        obd->obd_name, uuidp->uuid, -EINVAL);
393                 RETURN(-EINVAL);
394         }
395
396         OBD_ALLOC_PTR(tgt);
397         if (!tgt)
398                 RETURN(-ENOMEM);
399
400         mutex_init(&tgt->ltd_fid_mutex);
401         tgt->ltd_index = index;
402         tgt->ltd_uuid = *uuidp;
403         tgt->ltd_active = 0;
404
405         mutex_lock(&ltd->ltd_mutex);
406         rc = ltd_add_tgt(ltd, tgt);
407         mutex_unlock(&ltd->ltd_mutex);
408
409         if (rc)
410                 GOTO(out_tgt, rc);
411
412         if (!lmv->connected)
413                 /* lmv_check_connect() will connect this target. */
414                 RETURN(0);
415
416         rc = lmv_connect_mdc(obd, tgt);
417         if (!rc) {
418                 int easize = sizeof(struct lmv_stripe_md) +
419                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
420
421                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
422         }
423
424         RETURN(rc);
425
426 out_tgt:
427         OBD_FREE_PTR(tgt);
428         return rc;
429 }
430
431 static int lmv_check_connect(struct obd_device *obd)
432 {
433         struct lmv_obd *lmv = &obd->u.lmv;
434         struct lmv_tgt_desc *tgt;
435         int easize;
436         int rc;
437
438         ENTRY;
439
440         if (lmv->connected)
441                 RETURN(0);
442
443         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
444         if (lmv->connected)
445                 GOTO(unlock, rc = 0);
446
447         if (!lmv->lmv_mdt_count) {
448                 CERROR("%s: no targets configured: rc = -EINVAL\n",
449                        obd->obd_name);
450                 GOTO(unlock, rc = -EINVAL);
451         }
452
453         if (!lmv_mdt0_inited(lmv)) {
454                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
455                        obd->obd_name);
456                 GOTO(unlock, rc = -EINVAL);
457         }
458
459         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
460                obd->obd_uuid.uuid, obd->obd_name);
461
462         lmv_foreach_tgt(lmv, tgt) {
463                 rc = lmv_connect_mdc(obd, tgt);
464                 if (rc)
465                         GOTO(out_disc, rc);
466         }
467
468         lmv->connected = 1;
469         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
470         lmv_init_ea_size(obd->obd_self_export, easize, 0);
471         EXIT;
472 unlock:
473         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
474
475         return rc;
476
477 out_disc:
478         lmv_foreach_tgt(lmv, tgt) {
479                 tgt->ltd_active = 0;
480                 if (!tgt->ltd_exp)
481                         continue;
482
483                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
484                 obd_disconnect(tgt->ltd_exp);
485         }
486
487         goto unlock;
488 }
489
490 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
491 {
492         struct lmv_obd         *lmv = &obd->u.lmv;
493         struct obd_device      *mdc_obd;
494         int                     rc;
495         ENTRY;
496
497         LASSERT(tgt != NULL);
498         LASSERT(obd != NULL);
499
500         mdc_obd = class_exp2obd(tgt->ltd_exp);
501
502         if (mdc_obd) {
503                 mdc_obd->obd_force = obd->obd_force;
504                 mdc_obd->obd_fail = obd->obd_fail;
505                 mdc_obd->obd_no_recov = obd->obd_no_recov;
506
507                 if (lmv->lmv_tgts_kobj)
508                         sysfs_remove_link(lmv->lmv_tgts_kobj,
509                                           mdc_obd->obd_name);
510         }
511
512         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
513         if (rc)
514                 CERROR("Can't finanize fids factory\n");
515
516         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
517                tgt->ltd_exp->exp_obd->obd_name,
518                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
519
520         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
521         rc = obd_disconnect(tgt->ltd_exp);
522         if (rc) {
523                 if (tgt->ltd_active) {
524                         CERROR("Target %s disconnect error %d\n",
525                                tgt->ltd_uuid.uuid, rc);
526                 }
527         }
528
529         lmv_activate_target(lmv, tgt, 0);
530         tgt->ltd_exp = NULL;
531         RETURN(0);
532 }
533
534 static int lmv_disconnect(struct obd_export *exp)
535 {
536         struct obd_device *obd = class_exp2obd(exp);
537         struct lmv_obd *lmv = &obd->u.lmv;
538         struct lmv_tgt_desc *tgt;
539         int rc;
540
541         ENTRY;
542
543         lmv_foreach_connected_tgt(lmv, tgt)
544                 lmv_disconnect_mdc(obd, tgt);
545
546         if (lmv->lmv_tgts_kobj)
547                 kobject_put(lmv->lmv_tgts_kobj);
548
549         if (!lmv->connected)
550                 class_export_put(exp);
551         rc = class_disconnect(exp);
552         lmv->connected = 0;
553
554         RETURN(rc);
555 }
556
557 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
558                         void __user *uarg)
559 {
560         struct obd_device *obd = class_exp2obd(exp);
561         struct lmv_obd *lmv = &obd->u.lmv;
562         struct getinfo_fid2path *gf;
563         struct lmv_tgt_desc *tgt;
564         struct getinfo_fid2path *remote_gf = NULL;
565         struct lu_fid root_fid;
566         int remote_gf_size = 0;
567         int rc;
568
569         gf = karg;
570         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
571         if (IS_ERR(tgt))
572                 RETURN(PTR_ERR(tgt));
573
574         root_fid = *gf->gf_u.gf_root_fid;
575         LASSERT(fid_is_sane(&root_fid));
576
577 repeat_fid2path:
578         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
579         if (rc != 0 && rc != -EREMOTE)
580                 GOTO(out_fid2path, rc);
581
582         /* If remote_gf != NULL, it means just building the
583          * path on the remote MDT, copy this path segement to gf */
584         if (remote_gf != NULL) {
585                 struct getinfo_fid2path *ori_gf;
586                 char *ptr;
587                 int len;
588
589                 ori_gf = (struct getinfo_fid2path *)karg;
590                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
591                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
592                         GOTO(out_fid2path, rc = -EOVERFLOW);
593
594                 ptr = ori_gf->gf_u.gf_path;
595
596                 len = strlen(gf->gf_u.gf_path);
597                 /* move the current path to the right to release space
598                  * for closer-to-root part */
599                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
600                 memcpy(ptr, gf->gf_u.gf_path, len);
601                 ptr[len] = '/';
602         }
603
604         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
605                tgt->ltd_exp->exp_obd->obd_name,
606                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
607                gf->gf_linkno);
608
609         if (rc == 0)
610                 GOTO(out_fid2path, rc);
611
612         /* sigh, has to go to another MDT to do path building further */
613         if (remote_gf == NULL) {
614                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
615                 OBD_ALLOC(remote_gf, remote_gf_size);
616                 if (remote_gf == NULL)
617                         GOTO(out_fid2path, rc = -ENOMEM);
618                 remote_gf->gf_pathlen = PATH_MAX;
619         }
620
621         if (!fid_is_sane(&gf->gf_fid)) {
622                 CERROR("%s: invalid FID "DFID": rc = %d\n",
623                        tgt->ltd_exp->exp_obd->obd_name,
624                        PFID(&gf->gf_fid), -EINVAL);
625                 GOTO(out_fid2path, rc = -EINVAL);
626         }
627
628         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
629         if (IS_ERR(tgt))
630                 GOTO(out_fid2path, rc = -EINVAL);
631
632         remote_gf->gf_fid = gf->gf_fid;
633         remote_gf->gf_recno = -1;
634         remote_gf->gf_linkno = -1;
635         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
636         *remote_gf->gf_u.gf_root_fid = root_fid;
637         gf = remote_gf;
638         goto repeat_fid2path;
639
640 out_fid2path:
641         if (remote_gf != NULL)
642                 OBD_FREE(remote_gf, remote_gf_size);
643         RETURN(rc);
644 }
645
646 static int lmv_hsm_req_count(struct lmv_obd *lmv,
647                              const struct hsm_user_request *hur,
648                              const struct lmv_tgt_desc *tgt_mds)
649 {
650         struct lmv_tgt_desc *curr_tgt;
651         __u32 i;
652         int nr = 0;
653
654         /* count how many requests must be sent to the given target */
655         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
656                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
657                 if (IS_ERR(curr_tgt))
658                         RETURN(PTR_ERR(curr_tgt));
659                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
660                         nr++;
661         }
662         return nr;
663 }
664
665 static int lmv_hsm_req_build(struct lmv_obd *lmv,
666                               struct hsm_user_request *hur_in,
667                               const struct lmv_tgt_desc *tgt_mds,
668                               struct hsm_user_request *hur_out)
669 {
670         __u32 i, nr_out;
671         struct lmv_tgt_desc *curr_tgt;
672
673         /* build the hsm_user_request for the given target */
674         hur_out->hur_request = hur_in->hur_request;
675         nr_out = 0;
676         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
677                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
678                 if (IS_ERR(curr_tgt))
679                         RETURN(PTR_ERR(curr_tgt));
680                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
681                         hur_out->hur_user_item[nr_out] =
682                                                 hur_in->hur_user_item[i];
683                         nr_out++;
684                 }
685         }
686         hur_out->hur_request.hr_itemcount = nr_out;
687         memcpy(hur_data(hur_out), hur_data(hur_in),
688                hur_in->hur_request.hr_data_len);
689
690         RETURN(0);
691 }
692
693 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
694                                  int len, struct lustre_kernelcomm *lk,
695                                  void __user *uarg)
696 {
697         struct lmv_obd *lmv = &obd->u.lmv;
698         struct lu_tgt_desc *tgt;
699         int rc;
700
701         ENTRY;
702
703         /* unregister request (call from llapi_hsm_copytool_fini) */
704         lmv_foreach_connected_tgt(lmv, tgt)
705                 /* best effort: try to clean as much as possible
706                  * (continue on error) */
707                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
708
709         /* Whatever the result, remove copytool from kuc groups.
710          * Unreached coordinators will get EPIPE on next requests
711          * and will unregister automatically.
712          */
713         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
714
715         RETURN(rc);
716 }
717
718 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
719                                int len, struct lustre_kernelcomm *lk,
720                                void __user *uarg)
721 {
722         struct lmv_obd *lmv = &obd->u.lmv;
723         struct file *filp;
724         bool any_set = false;
725         struct kkuc_ct_data *kcd;
726         size_t kcd_size;
727         struct lu_tgt_desc *tgt;
728         __u32 i;
729         int err;
730         int rc = 0;
731
732         ENTRY;
733
734         filp = fget(lk->lk_wfd);
735         if (!filp)
736                 RETURN(-EBADF);
737
738         if (lk->lk_flags & LK_FLG_DATANR)
739                 kcd_size = offsetof(struct kkuc_ct_data,
740                                     kcd_archives[lk->lk_data_count]);
741         else
742                 kcd_size = sizeof(*kcd);
743
744         OBD_ALLOC(kcd, kcd_size);
745         if (kcd == NULL)
746                 GOTO(err_fput, rc = -ENOMEM);
747
748         kcd->kcd_nr_archives = lk->lk_data_count;
749         if (lk->lk_flags & LK_FLG_DATANR) {
750                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
751                 if (lk->lk_data_count > 0)
752                         memcpy(kcd->kcd_archives, lk->lk_data,
753                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
754         } else {
755                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
756         }
757
758         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
759                                    lk->lk_group, kcd, kcd_size);
760         OBD_FREE(kcd, kcd_size);
761         if (rc)
762                 GOTO(err_fput, rc);
763
764         /* All or nothing: try to register to all MDS.
765          * In case of failure, unregister from previous MDS,
766          * except if it because of inactive target. */
767         lmv_foreach_connected_tgt(lmv, tgt) {
768                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
769                 if (err) {
770                         if (tgt->ltd_active) {
771                                 /* permanent error */
772                                 CERROR("%s: iocontrol MDC %s on MDT"
773                                        " idx %d cmd %x: err = %d\n",
774                                        lmv2obd_dev(lmv)->obd_name,
775                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
776                                        err);
777                                 rc = err;
778                                 lk->lk_flags |= LK_FLG_STOP;
779                                 i = tgt->ltd_index;
780                                 /* unregister from previous MDS */
781                                 lmv_foreach_connected_tgt(lmv, tgt) {
782                                         if (tgt->ltd_index >= i)
783                                                 break;
784
785                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
786                                                       lk, uarg);
787                                 }
788                                 GOTO(err_kkuc_rem, rc);
789                         }
790                         /* else: transient error.
791                          * kuc will register to the missing MDT
792                          * when it is back */
793                 } else {
794                         any_set = true;
795                 }
796         }
797
798         if (!any_set)
799                 /* no registration done: return error */
800                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
801
802         RETURN(0);
803
804 err_kkuc_rem:
805         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
806
807 err_fput:
808         fput(filp);
809         return rc;
810 }
811
812 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
813                          int len, void *karg, void __user *uarg)
814 {
815         struct obd_device *obd = class_exp2obd(exp);
816         struct lmv_obd *lmv = &obd->u.lmv;
817         struct lu_tgt_desc *tgt = NULL;
818         int set = 0;
819         __u32 count = lmv->lmv_mdt_count;
820         int rc = 0;
821
822         ENTRY;
823
824         if (count == 0)
825                 RETURN(-ENOTTY);
826
827         switch (cmd) {
828         case IOC_OBD_STATFS: {
829                 struct obd_ioctl_data *data = karg;
830                 struct obd_device *mdc_obd;
831                 struct obd_statfs stat_buf = {0};
832                 __u32 index;
833
834                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
835
836                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
837                         RETURN(-ENODEV);
838
839                 tgt = lmv_tgt(lmv, index);
840                 if (!tgt)
841                         RETURN(-EAGAIN);
842
843                 if (!tgt->ltd_active)
844                         RETURN(-ENODATA);
845
846                 mdc_obd = class_exp2obd(tgt->ltd_exp);
847                 if (!mdc_obd)
848                         RETURN(-EINVAL);
849
850                 /* copy UUID */
851                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
852                                  min((int) data->ioc_plen2,
853                                      (int) sizeof(struct obd_uuid))))
854                         RETURN(-EFAULT);
855
856                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
857                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
858                                 0);
859                 if (rc)
860                         RETURN(rc);
861                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
862                                  min((int) data->ioc_plen1,
863                                      (int) sizeof(stat_buf))))
864                         RETURN(-EFAULT);
865                 break;
866         }
867         case OBD_IOC_QUOTACTL: {
868                 struct if_quotactl *qctl = karg;
869                 struct obd_quotactl *oqctl;
870
871                 if (qctl->qc_valid == QC_MDTIDX) {
872                         tgt = lmv_tgt(lmv, qctl->qc_idx);
873                 } else if (qctl->qc_valid == QC_UUID) {
874                         lmv_foreach_tgt(lmv, tgt) {
875                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
876                                                      &qctl->obd_uuid))
877                                         continue;
878
879                                 if (!tgt->ltd_exp)
880                                         RETURN(-EINVAL);
881
882                                 break;
883                         }
884                 } else {
885                         RETURN(-EINVAL);
886                 }
887
888                 if (!tgt || !tgt->ltd_exp)
889                         RETURN(-EINVAL);
890
891                 OBD_ALLOC_PTR(oqctl);
892                 if (!oqctl)
893                         RETURN(-ENOMEM);
894
895                 QCTL_COPY(oqctl, qctl);
896                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
897                 if (rc == 0) {
898                         QCTL_COPY(qctl, oqctl);
899                         qctl->qc_valid = QC_MDTIDX;
900                         qctl->obd_uuid = tgt->ltd_uuid;
901                 }
902                 OBD_FREE_PTR(oqctl);
903                 break;
904         }
905         case LL_IOC_GET_CONNECT_FLAGS: {
906                 tgt = lmv_tgt(lmv, 0);
907                 rc = -ENODATA;
908                 if (tgt && tgt->ltd_exp)
909                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
910                 break;
911         }
912         case LL_IOC_FID2MDTIDX: {
913                 struct lu_fid *fid = karg;
914                 int             mdt_index;
915
916                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
917                 if (rc != 0)
918                         RETURN(rc);
919
920                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
921                  * point to user space memory for FID2MDTIDX. */
922                 *(__u32 *)uarg = mdt_index;
923                 break;
924         }
925         case OBD_IOC_FID2PATH: {
926                 rc = lmv_fid2path(exp, len, karg, uarg);
927                 break;
928         }
929         case LL_IOC_HSM_STATE_GET:
930         case LL_IOC_HSM_STATE_SET:
931         case LL_IOC_HSM_ACTION: {
932                 struct md_op_data *op_data = karg;
933
934                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
935                 if (IS_ERR(tgt))
936                         RETURN(PTR_ERR(tgt));
937
938                 if (tgt->ltd_exp == NULL)
939                         RETURN(-EINVAL);
940
941                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
942                 break;
943         }
944         case LL_IOC_HSM_PROGRESS: {
945                 const struct hsm_progress_kernel *hpk = karg;
946
947                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
948                 if (IS_ERR(tgt))
949                         RETURN(PTR_ERR(tgt));
950                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
951                 break;
952         }
953         case LL_IOC_HSM_REQUEST: {
954                 struct hsm_user_request *hur = karg;
955                 unsigned int reqcount = hur->hur_request.hr_itemcount;
956
957                 if (reqcount == 0)
958                         RETURN(0);
959
960                 /* if the request is about a single fid
961                  * or if there is a single MDS, no need to split
962                  * the request. */
963                 if (reqcount == 1 || count == 1) {
964                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
965                         if (IS_ERR(tgt))
966                                 RETURN(PTR_ERR(tgt));
967                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
968                 } else {
969                         /* split fid list to their respective MDS */
970                         lmv_foreach_connected_tgt(lmv, tgt) {
971                                 int nr, rc1;
972                                 size_t reqlen;
973                                 struct hsm_user_request *req;
974
975                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
976                                 if (nr < 0)
977                                         RETURN(nr);
978                                 if (nr == 0) /* nothing for this MDS */
979                                         continue;
980
981                                 /* build a request with fids for this MDS */
982                                 reqlen = offsetof(typeof(*hur),
983                                                   hur_user_item[nr])
984                                                 + hur->hur_request.hr_data_len;
985                                 OBD_ALLOC_LARGE(req, reqlen);
986                                 if (req == NULL)
987                                         RETURN(-ENOMEM);
988                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
989                                 if (rc1 < 0)
990                                         GOTO(hsm_req_err, rc1);
991                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
992                                                     req, uarg);
993 hsm_req_err:
994                                 if (rc1 != 0 && rc == 0)
995                                         rc = rc1;
996                                 OBD_FREE_LARGE(req, reqlen);
997                         }
998                 }
999                 break;
1000         }
1001         case LL_IOC_LOV_SWAP_LAYOUTS: {
1002                 struct md_op_data *op_data = karg;
1003                 struct lmv_tgt_desc *tgt1, *tgt2;
1004
1005                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1006                 if (IS_ERR(tgt1))
1007                         RETURN(PTR_ERR(tgt1));
1008
1009                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1010                 if (IS_ERR(tgt2))
1011                         RETURN(PTR_ERR(tgt2));
1012
1013                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1014                         RETURN(-EINVAL);
1015
1016                 /* only files on same MDT can have their layouts swapped */
1017                 if (tgt1->ltd_index != tgt2->ltd_index)
1018                         RETURN(-EPERM);
1019
1020                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1021                 break;
1022         }
1023         case LL_IOC_HSM_CT_START: {
1024                 struct lustre_kernelcomm *lk = karg;
1025                 if (lk->lk_flags & LK_FLG_STOP)
1026                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1027                 else
1028                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1029                 break;
1030         }
1031         default:
1032                 lmv_foreach_connected_tgt(lmv, tgt) {
1033                         struct obd_device *mdc_obd;
1034                         int err;
1035
1036                         /* ll_umount_begin() sets force flag but for lmv, not
1037                          * mdc. Let's pass it through */
1038                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1039                         mdc_obd->obd_force = obd->obd_force;
1040                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1041                         if (err) {
1042                                 if (tgt->ltd_active) {
1043                                         CERROR("error: iocontrol MDC %s on MDT"
1044                                                " idx %d cmd %x: err = %d\n",
1045                                                tgt->ltd_uuid.uuid,
1046                                                tgt->ltd_index, cmd, err);
1047                                         if (!rc)
1048                                                 rc = err;
1049                                 }
1050                         } else
1051                                 set = 1;
1052                 }
1053                 if (!set && !rc)
1054                         rc = -EIO;
1055         }
1056         RETURN(rc);
1057 }
1058
1059 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1060                   struct lu_fid *fid, struct md_op_data *op_data)
1061 {
1062         struct obd_device *obd = class_exp2obd(exp);
1063         struct lmv_obd *lmv = &obd->u.lmv;
1064         struct lmv_tgt_desc *tgt;
1065         int rc;
1066
1067         ENTRY;
1068
1069         LASSERT(op_data);
1070         LASSERT(fid);
1071
1072         tgt = lmv_tgt(lmv, op_data->op_mds);
1073         if (!tgt)
1074                 RETURN(-ENODEV);
1075
1076         if (!tgt->ltd_active || !tgt->ltd_exp)
1077                 RETURN(-ENODEV);
1078
1079         /*
1080          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1081          * on server that seq in new allocated fid is not yet known.
1082          */
1083         mutex_lock(&tgt->ltd_fid_mutex);
1084         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1085         mutex_unlock(&tgt->ltd_fid_mutex);
1086         if (rc > 0) {
1087                 LASSERT(fid_is_sane(fid));
1088                 rc = 0;
1089         }
1090
1091         RETURN(rc);
1092 }
1093
1094 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1095 {
1096         struct lmv_obd *lmv = &obd->u.lmv;
1097         struct lmv_desc *desc;
1098         struct lnet_process_id lnet_id;
1099         int i = 0;
1100         int rc;
1101
1102         ENTRY;
1103
1104         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1105                 CERROR("LMV setup requires a descriptor\n");
1106                 RETURN(-EINVAL);
1107         }
1108
1109         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1110         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1111                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1112                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1113                 RETURN(-EINVAL);
1114         }
1115
1116         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1117                      desc->ld_uuid.uuid);
1118         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1119         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1120         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1121                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1122         lmv->max_def_easize = 0;
1123         lmv->max_easize = 0;
1124
1125         spin_lock_init(&lmv->lmv_lock);
1126
1127         /*
1128          * initialize rr_index to lower 32bit of netid, so that client
1129          * can distribute subdirs evenly from the beginning.
1130          */
1131         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1132                 if (lnet_id.nid != LNET_NID_LO_0) {
1133                         lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1134                         break;
1135                 }
1136         }
1137
1138         rc = lmv_tunables_init(obd);
1139         if (rc)
1140                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1141                       obd->obd_name, rc);
1142
1143         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1144                              LUSTRE_CLI_FLD_HASH_DHT);
1145         if (rc)
1146                 CERROR("Can't init FLD, err %d\n", rc);
1147
1148         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1149         if (rc)
1150                 CWARN("%s: error initialize target table: rc = %d\n",
1151                       obd->obd_name, rc);
1152
1153         RETURN(rc);
1154 }
1155
1156 static int lmv_cleanup(struct obd_device *obd)
1157 {
1158         struct lmv_obd *lmv = &obd->u.lmv;
1159         struct lu_tgt_desc *tgt;
1160         struct lu_tgt_desc *tmp;
1161
1162         ENTRY;
1163
1164         fld_client_fini(&lmv->lmv_fld);
1165         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1166                 lmv_del_target(lmv, tgt);
1167         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1168
1169         RETURN(0);
1170 }
1171
1172 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1173 {
1174         struct lustre_cfg       *lcfg = buf;
1175         struct obd_uuid         obd_uuid;
1176         int                     gen;
1177         __u32                   index;
1178         int                     rc;
1179         ENTRY;
1180
1181         switch (lcfg->lcfg_command) {
1182         case LCFG_ADD_MDC:
1183                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1184                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1185                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1186                         GOTO(out, rc = -EINVAL);
1187
1188                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1189
1190                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1191                         GOTO(out, rc = -EINVAL);
1192                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1193                         GOTO(out, rc = -EINVAL);
1194                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1195                 GOTO(out, rc);
1196         default:
1197                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1198                 GOTO(out, rc = -EINVAL);
1199         }
1200 out:
1201         RETURN(rc);
1202 }
1203
1204 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1205 {
1206         int i;
1207
1208         if (flags & OBD_STATFS_FOR_MDT0)
1209                 return 0;
1210
1211         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1212                 return lmv->lmv_statfs_start;
1213
1214         /* choose initial MDT for this client */
1215         for (i = 0;; i++) {
1216                 struct lnet_process_id lnet_id;
1217                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1218                         break;
1219
1220                 if (lnet_id.nid != LNET_NID_LO_0) {
1221                         /* We dont need a full 64-bit modulus, just enough
1222                          * to distribute the requests across MDTs evenly.
1223                          */
1224                         lmv->lmv_statfs_start = (u32)lnet_id.nid %
1225                                                 lmv->lmv_mdt_count;
1226                         break;
1227                 }
1228         }
1229
1230         return lmv->lmv_statfs_start;
1231 }
1232
1233 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1234                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1235 {
1236         struct obd_device *obd = class_exp2obd(exp);
1237         struct lmv_obd *lmv = &obd->u.lmv;
1238         struct obd_statfs *temp;
1239         struct lu_tgt_desc *tgt;
1240         __u32 i;
1241         __u32 idx;
1242         int rc = 0;
1243
1244         ENTRY;
1245
1246         OBD_ALLOC(temp, sizeof(*temp));
1247         if (temp == NULL)
1248                 RETURN(-ENOMEM);
1249
1250         /* distribute statfs among MDTs */
1251         idx = lmv_select_statfs_mdt(lmv, flags);
1252
1253         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1254                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1255                 tgt = lmv_tgt(lmv, idx);
1256                 if (!tgt || !tgt->ltd_exp)
1257                         continue;
1258
1259                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age, flags);
1260                 if (rc) {
1261                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1262                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1263                         GOTO(out_free_temp, rc);
1264                 }
1265
1266                 if (temp->os_state & OS_STATFS_SUM ||
1267                     flags == OBD_STATFS_FOR_MDT0) {
1268                         /* reset to the last aggregated values
1269                          * and don't sum with non-aggrated data */
1270                         /* If the statfs is from mount, it needs to retrieve
1271                          * necessary information from MDT0. i.e. mount does
1272                          * not need the merged osfs from all of MDT. Also
1273                          * clients can be mounted as long as MDT0 is in
1274                          * service */
1275                         *osfs = *temp;
1276                         break;
1277                 }
1278
1279                 if (i == 0) {
1280                         *osfs = *temp;
1281                 } else {
1282                         osfs->os_bavail += temp->os_bavail;
1283                         osfs->os_blocks += temp->os_blocks;
1284                         osfs->os_ffree += temp->os_ffree;
1285                         osfs->os_files += temp->os_files;
1286                         osfs->os_granted += temp->os_granted;
1287                 }
1288         }
1289
1290         EXIT;
1291 out_free_temp:
1292         OBD_FREE(temp, sizeof(*temp));
1293         return rc;
1294 }
1295
1296 static int lmv_statfs_update(void *cookie, int rc)
1297 {
1298         struct obd_info *oinfo = cookie;
1299         struct obd_device *obd = oinfo->oi_obd;
1300         struct lmv_obd *lmv = &obd->u.lmv;
1301         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1302         struct obd_statfs *osfs = oinfo->oi_osfs;
1303
1304         /*
1305          * NB: don't deactivate TGT upon error, because we may not trigger async
1306          * statfs any longer, then there is no chance to activate TGT.
1307          */
1308         if (!rc) {
1309                 spin_lock(&lmv->lmv_lock);
1310                 tgt->ltd_statfs = *osfs;
1311                 tgt->ltd_statfs_age = ktime_get_seconds();
1312                 spin_unlock(&lmv->lmv_lock);
1313                 lmv->lmv_qos.lq_dirty = 1;
1314         }
1315
1316         return rc;
1317 }
1318
1319 /* update tgt statfs async if it's ld_qos_maxage old */
1320 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1321 {
1322         struct obd_info oinfo = {
1323                 .oi_obd = obd,
1324                 .oi_tgt = tgt,
1325                 .oi_cb_up = lmv_statfs_update,
1326         };
1327         int rc;
1328
1329         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1330             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1331                 return 0;
1332
1333         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1334
1335         return rc;
1336 }
1337
1338 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1339                         struct lu_fid *fid)
1340 {
1341         struct obd_device *obd = exp->exp_obd;
1342         struct lmv_obd *lmv = &obd->u.lmv;
1343         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1344         int rc;
1345
1346         ENTRY;
1347
1348         if (!tgt)
1349                 RETURN(-ENODEV);
1350
1351         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1352         RETURN(rc);
1353 }
1354
1355 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1356                         u64 obd_md_valid, const char *name, size_t buf_size,
1357                         struct ptlrpc_request **req)
1358 {
1359         struct obd_device *obd = exp->exp_obd;
1360         struct lmv_obd *lmv = &obd->u.lmv;
1361         struct lmv_tgt_desc *tgt;
1362         int rc;
1363
1364         ENTRY;
1365
1366         tgt = lmv_fid2tgt(lmv, fid);
1367         if (IS_ERR(tgt))
1368                 RETURN(PTR_ERR(tgt));
1369
1370         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1371
1372         RETURN(rc);
1373 }
1374
1375 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1376                         u64 obd_md_valid, const char *name,
1377                         const void *value, size_t value_size,
1378                         unsigned int xattr_flags, u32 suppgid,
1379                         struct ptlrpc_request **req)
1380 {
1381         struct obd_device *obd = exp->exp_obd;
1382         struct lmv_obd *lmv = &obd->u.lmv;
1383         struct lmv_tgt_desc *tgt;
1384         int rc;
1385
1386         ENTRY;
1387
1388         tgt = lmv_fid2tgt(lmv, fid);
1389         if (IS_ERR(tgt))
1390                 RETURN(PTR_ERR(tgt));
1391
1392         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1393                          value, value_size, xattr_flags, suppgid, req);
1394
1395         RETURN(rc);
1396 }
1397
1398 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1399                        struct ptlrpc_request **request)
1400 {
1401         struct obd_device *obd = exp->exp_obd;
1402         struct lmv_obd *lmv = &obd->u.lmv;
1403         struct lmv_tgt_desc *tgt;
1404         int rc;
1405
1406         ENTRY;
1407
1408         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1409         if (IS_ERR(tgt))
1410                 RETURN(PTR_ERR(tgt));
1411
1412         if (op_data->op_flags & MF_GET_MDT_IDX) {
1413                 op_data->op_mds = tgt->ltd_index;
1414                 RETURN(0);
1415         }
1416
1417         rc = md_getattr(tgt->ltd_exp, op_data, request);
1418
1419         RETURN(rc);
1420 }
1421
1422 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1423 {
1424         struct obd_device *obd = exp->exp_obd;
1425         struct lmv_obd *lmv = &obd->u.lmv;
1426         struct lu_tgt_desc *tgt;
1427
1428         ENTRY;
1429
1430         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1431
1432         /*
1433          * With DNE every object can have two locks in different namespaces:
1434          * lookup lock in space of MDT storing direntry and update/open lock in
1435          * space of MDT storing inode.
1436          */
1437         lmv_foreach_connected_tgt(lmv, tgt)
1438                 md_null_inode(tgt->ltd_exp, fid);
1439
1440         RETURN(0);
1441 }
1442
1443 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1444                      struct md_open_data *mod, struct ptlrpc_request **request)
1445 {
1446         struct obd_device *obd = exp->exp_obd;
1447         struct lmv_obd *lmv = &obd->u.lmv;
1448         struct lmv_tgt_desc *tgt;
1449         int rc;
1450
1451         ENTRY;
1452
1453         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1454         if (IS_ERR(tgt))
1455                 RETURN(PTR_ERR(tgt));
1456
1457         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1458         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1459         RETURN(rc);
1460 }
1461
1462 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
1463 {
1464         struct lu_tgt_desc *tgt;
1465         __u64 total_weight = 0;
1466         __u64 cur_weight = 0;
1467         __u64 rand;
1468         int rc;
1469
1470         ENTRY;
1471
1472         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1473                 RETURN(ERR_PTR(-EAGAIN));
1474
1475         down_write(&lmv->lmv_qos.lq_rw_sem);
1476
1477         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1478                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1479
1480         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1481         if (rc)
1482                 GOTO(unlock, tgt = ERR_PTR(rc));
1483
1484         lmv_foreach_tgt(lmv, tgt) {
1485                 tgt->ltd_qos.ltq_usable = 0;
1486                 if (!tgt->ltd_exp || !tgt->ltd_active)
1487                         continue;
1488
1489                 tgt->ltd_qos.ltq_usable = 1;
1490                 lu_tgt_qos_weight_calc(tgt);
1491                 total_weight += tgt->ltd_qos.ltq_weight;
1492         }
1493
1494         rand = lu_prandom_u64_max(total_weight);
1495
1496         lmv_foreach_connected_tgt(lmv, tgt) {
1497                 if (!tgt->ltd_qos.ltq_usable)
1498                         continue;
1499
1500                 cur_weight += tgt->ltd_qos.ltq_weight;
1501                 if (cur_weight < rand)
1502                         continue;
1503
1504                 *mdt = tgt->ltd_index;
1505                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1506                 GOTO(unlock, rc = 0);
1507         }
1508
1509         /* no proper target found */
1510         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1511 unlock:
1512         up_write(&lmv->lmv_qos.lq_rw_sem);
1513
1514         return tgt;
1515 }
1516
1517 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
1518 {
1519         struct lu_tgt_desc *tgt;
1520         int i;
1521         int index;
1522
1523         ENTRY;
1524
1525         spin_lock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1526         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1527                 index = (i + lmv->lmv_qos_rr_index) %
1528                         lmv->lmv_mdt_descs.ltd_tgts_size;
1529                 tgt = lmv_tgt(lmv, index);
1530                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1531                         continue;
1532
1533                 *mdt = tgt->ltd_index;
1534                 lmv->lmv_qos_rr_index = (*mdt + 1) %
1535                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1536                 spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1537
1538                 RETURN(tgt);
1539         }
1540         spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1541
1542         RETURN(ERR_PTR(-ENODEV));
1543 }
1544
1545 static struct lmv_tgt_desc *
1546 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1547                        const char *name, int namelen, struct lu_fid *fid,
1548                        __u32 *mds, bool new_layout)
1549 {
1550         struct lmv_tgt_desc *tgt;
1551         const struct lmv_oinfo *oinfo;
1552
1553         if (!lmv_dir_striped(lsm) || !namelen) {
1554                 tgt = lmv_fid2tgt(lmv, fid);
1555                 if (IS_ERR(tgt))
1556                         return tgt;
1557
1558                 *mds = tgt->ltd_index;
1559                 return tgt;
1560         }
1561
1562         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1563                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1564                         return ERR_PTR(-EBADF);
1565                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1566         } else {
1567                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1568                 if (IS_ERR(oinfo))
1569                         return ERR_CAST(oinfo);
1570         }
1571
1572         /* check stripe FID is sane */
1573         if (!fid_is_sane(&oinfo->lmo_fid))
1574                 return ERR_PTR(-ENODEV);
1575
1576         *fid = oinfo->lmo_fid;
1577         *mds = oinfo->lmo_mds;
1578         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1579
1580         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1581
1582         return tgt ? tgt : ERR_PTR(-ENODEV);
1583 }
1584
1585 /**
1586  * Locate MDT of op_data->op_fid1
1587  *
1588  * For striped directory, it will locate the stripe by name hash, if hash_type
1589  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1590  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1591  * indicates whether old or new layout is used to locate.
1592  *
1593  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1594  *
1595  * \param[in] lmv       LMV device
1596  * \param[in] op_data   client MD stack parameters, name, namelen
1597  *                      mds_num etc.
1598  *
1599  * retval               pointer to the lmv_tgt_desc if succeed.
1600  *                      ERR_PTR(errno) if failed.
1601  */
1602 struct lmv_tgt_desc *
1603 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1604 {
1605         struct lmv_stripe_md *lsm = op_data->op_mea1;
1606         struct lmv_oinfo *oinfo;
1607         struct lmv_tgt_desc *tgt;
1608
1609         if (lmv_dir_foreign(lsm))
1610                 return ERR_PTR(-ENODATA);
1611
1612         /* During creating VOLATILE file, it should honor the mdt
1613          * index if the file under striped dir is being restored, see
1614          * ct_restore(). */
1615         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1616             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1617                 tgt = lmv_tgt(lmv, op_data->op_mds);
1618                 if (!tgt)
1619                         return ERR_PTR(-ENODEV);
1620
1621                 if (lmv_dir_striped(lsm)) {
1622                         int i;
1623
1624                         /* refill the right parent fid */
1625                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1626                                 oinfo = &lsm->lsm_md_oinfo[i];
1627                                 if (oinfo->lmo_mds == op_data->op_mds) {
1628                                         op_data->op_fid1 = oinfo->lmo_fid;
1629                                         break;
1630                                 }
1631                         }
1632
1633                         if (i == lsm->lsm_md_stripe_count)
1634                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1635                 }
1636         } else if (lmv_dir_bad_hash(lsm)) {
1637                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1638                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1639
1640                 op_data->op_fid1 = oinfo->lmo_fid;
1641                 op_data->op_mds = oinfo->lmo_mds;
1642                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1643                 if (!tgt)
1644                         return ERR_PTR(-ENODEV);
1645         } else {
1646                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1647                                 op_data->op_name, op_data->op_namelen,
1648                                 &op_data->op_fid1, &op_data->op_mds,
1649                                 op_data->op_new_layout);
1650         }
1651
1652         return tgt;
1653 }
1654
1655 /* Locate MDT of op_data->op_fid2 for link/rename */
1656 static struct lmv_tgt_desc *
1657 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1658 {
1659         struct lmv_tgt_desc *tgt;
1660         int rc;
1661
1662         LASSERT(op_data->op_name);
1663         if (lmv_dir_layout_changing(op_data->op_mea2)) {
1664                 struct lu_fid fid1 = op_data->op_fid1;
1665                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1666                 struct ptlrpc_request *request = NULL;
1667
1668                 /*
1669                  * avoid creating new file under old layout of migrating
1670                  * directory, check it here.
1671                  */
1672                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1673                                 op_data->op_name, op_data->op_namelen,
1674                                 &op_data->op_fid2, &op_data->op_mds, false);
1675                 if (IS_ERR(tgt))
1676                         RETURN(tgt);
1677
1678                 op_data->op_fid1 = op_data->op_fid2;
1679                 op_data->op_mea1 = op_data->op_mea2;
1680                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1681                 op_data->op_fid1 = fid1;
1682                 op_data->op_mea1 = lsm1;
1683                 if (!rc) {
1684                         ptlrpc_req_finished(request);
1685                         RETURN(ERR_PTR(-EEXIST));
1686                 }
1687
1688                 if (rc != -ENOENT)
1689                         RETURN(ERR_PTR(rc));
1690         }
1691
1692         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1693                                 op_data->op_namelen, &op_data->op_fid2,
1694                                 &op_data->op_mds, true);
1695 }
1696
1697 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1698 {
1699         struct lu_tgt_desc *tgt;
1700         struct ptlrpc_request *request;
1701         int rc;
1702
1703         LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1704         LASSERT(!op_data->op_new_layout);
1705
1706         tgt = lmv_locate_tgt(lmv, op_data);
1707         if (IS_ERR(tgt))
1708                 return PTR_ERR(tgt);
1709
1710         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1711         if (!rc) {
1712                 ptlrpc_req_finished(request);
1713                 return -EEXIST;
1714         }
1715
1716         return rc;
1717 }
1718
1719 /* mkdir by QoS in two cases:
1720  * 1. 'lfs mkdir -i -1'
1721  * 2. parent default LMV master_mdt_index is -1
1722  *
1723  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1724  * directories under striped directory.
1725  */
1726 static inline bool lmv_op_qos_mkdir(const struct md_op_data *op_data)
1727 {
1728         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1729         const struct lmv_user_md *lum = op_data->op_data;
1730
1731         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1732                 return false;
1733
1734         if (lmv_dir_striped(op_data->op_mea1))
1735                 return false;
1736
1737         if (op_data->op_cli_flags & CLI_SET_MEA && lum &&
1738             (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1739              le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1740             le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT)
1741                 return true;
1742
1743         if (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT)
1744                 return true;
1745
1746         return false;
1747 }
1748
1749 /* 'lfs mkdir -i <specific_MDT>' */
1750 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1751 {
1752         const struct lmv_user_md *lum = op_data->op_data;
1753
1754         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1755                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1756                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1757                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1758                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1759 }
1760
1761 /* parent default LMV master_mdt_index is not -1. */
1762 static inline bool
1763 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1764 {
1765         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1766                op_data->op_default_mea1 &&
1767                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1768                         LMV_OFFSET_DEFAULT;
1769 }
1770 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1771                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1772                 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1773                 struct ptlrpc_request **request)
1774 {
1775         struct obd_device *obd = exp->exp_obd;
1776         struct lmv_obd *lmv = &obd->u.lmv;
1777         struct lmv_tgt_desc *tgt;
1778         struct mdt_body *repbody;
1779         int rc;
1780
1781         ENTRY;
1782
1783         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1784                 RETURN(-EIO);
1785
1786         if (lmv_dir_bad_hash(op_data->op_mea1))
1787                 RETURN(-EBADF);
1788
1789         if (lmv_dir_layout_changing(op_data->op_mea1)) {
1790                 /*
1791                  * if parent is migrating, create() needs to lookup existing
1792                  * name in both old and new layout, check old layout on client.
1793                  */
1794                 rc = lmv_old_layout_lookup(lmv, op_data);
1795                 if (rc != -ENOENT)
1796                         RETURN(rc);
1797
1798                 op_data->op_new_layout = true;
1799         }
1800
1801         tgt = lmv_locate_tgt(lmv, op_data);
1802         if (IS_ERR(tgt))
1803                 RETURN(PTR_ERR(tgt));
1804
1805         if (lmv_op_user_specific_mkdir(op_data)) {
1806                 struct lmv_user_md *lum = op_data->op_data;
1807
1808                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1809                 tgt = lmv_tgt(lmv, op_data->op_mds);
1810                 if (!tgt)
1811                         RETURN(-ENODEV);
1812         } else if (lmv_op_default_specific_mkdir(op_data)) {
1813                 op_data->op_mds =
1814                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1815                 tgt = lmv_tgt(lmv, op_data->op_mds);
1816                 if (!tgt)
1817                         RETURN(-ENODEV);
1818         } else if (lmv_op_qos_mkdir(op_data)) {
1819                 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds);
1820                 if (tgt == ERR_PTR(-EAGAIN))
1821                         tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1822                 if (IS_ERR(tgt))
1823                         RETURN(PTR_ERR(tgt));
1824                 /*
1825                  * only update statfs after QoS mkdir, this means the cached
1826                  * statfs may be stale, and current mkdir may not follow QoS
1827                  * accurately, but it's not serious, and avoids periodic statfs
1828                  * when client doesn't mkdir by QoS.
1829                  */
1830                 lmv_statfs_check_update(obd, tgt);
1831         }
1832
1833 retry:
1834         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1835         if (rc)
1836                 RETURN(rc);
1837
1838         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1839                 (int)op_data->op_namelen, op_data->op_name,
1840                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1841                 op_data->op_mds);
1842
1843         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1844         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1845                        cap_effective, rdev, request);
1846         if (rc == 0) {
1847                 if (*request == NULL)
1848                         RETURN(rc);
1849                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1850         }
1851
1852         /* dir restripe needs to send to MDT where dir is located */
1853         if (rc != -EREMOTE ||
1854             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
1855                 RETURN(rc);
1856
1857         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
1858         if (repbody == NULL)
1859                 RETURN(-EPROTO);
1860
1861         /* Not cross-ref case, just get out of here. */
1862         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
1863                 RETURN(rc);
1864
1865         op_data->op_fid2 = repbody->mbo_fid1;
1866         ptlrpc_req_finished(*request);
1867         *request = NULL;
1868
1869         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
1870         if (IS_ERR(tgt))
1871                 RETURN(PTR_ERR(tgt));
1872
1873         op_data->op_mds = tgt->ltd_index;
1874         goto retry;
1875 }
1876
1877 static int
1878 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1879             const union ldlm_policy_data *policy, struct md_op_data *op_data,
1880             struct lustre_handle *lockh, __u64 extra_lock_flags)
1881 {
1882         struct obd_device *obd = exp->exp_obd;
1883         struct lmv_obd *lmv = &obd->u.lmv;
1884         struct lmv_tgt_desc *tgt;
1885         int rc;
1886
1887         ENTRY;
1888
1889         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1890
1891         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1892         if (IS_ERR(tgt))
1893                 RETURN(PTR_ERR(tgt));
1894
1895         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1896                PFID(&op_data->op_fid1), tgt->ltd_index);
1897
1898         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1899                         extra_lock_flags);
1900
1901         RETURN(rc);
1902 }
1903
1904 int
1905 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1906                  struct ptlrpc_request **preq)
1907 {
1908         struct obd_device *obd = exp->exp_obd;
1909         struct lmv_obd *lmv = &obd->u.lmv;
1910         struct lmv_tgt_desc *tgt;
1911         struct mdt_body *body;
1912         int rc;
1913
1914         ENTRY;
1915
1916 retry:
1917         tgt = lmv_locate_tgt(lmv, op_data);
1918         if (IS_ERR(tgt))
1919                 RETURN(PTR_ERR(tgt));
1920
1921         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1922                 (int)op_data->op_namelen, op_data->op_name,
1923                 PFID(&op_data->op_fid1), tgt->ltd_index);
1924
1925         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1926         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1927                 ptlrpc_req_finished(*preq);
1928                 *preq = NULL;
1929                 goto retry;
1930         }
1931
1932         if (rc)
1933                 RETURN(rc);
1934
1935         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1936         LASSERT(body != NULL);
1937
1938         if (body->mbo_valid & OBD_MD_MDS) {
1939                 op_data->op_fid1 = body->mbo_fid1;
1940                 op_data->op_valid |= OBD_MD_FLCROSSREF;
1941                 op_data->op_namelen = 0;
1942                 op_data->op_name = NULL;
1943
1944                 ptlrpc_req_finished(*preq);
1945                 *preq = NULL;
1946
1947                 goto retry;
1948         }
1949
1950         RETURN(rc);
1951 }
1952
1953 #define md_op_data_fid(op_data, fl)                     \
1954         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1955          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1956          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1957          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1958          NULL)
1959
1960 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1961                             struct md_op_data *op_data, __u32 op_tgt,
1962                             enum ldlm_mode mode, int bits, int flag)
1963 {
1964         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1965         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1966         union ldlm_policy_data policy = { { 0 } };
1967         int rc = 0;
1968         ENTRY;
1969
1970         if (!fid_is_sane(fid))
1971                 RETURN(0);
1972
1973         if (tgt == NULL) {
1974                 tgt = lmv_fid2tgt(lmv, fid);
1975                 if (IS_ERR(tgt))
1976                         RETURN(PTR_ERR(tgt));
1977         }
1978
1979         if (tgt->ltd_index != op_tgt) {
1980                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1981                 policy.l_inodebits.bits = bits;
1982                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1983                                       mode, LCF_ASYNC, NULL);
1984         } else {
1985                 CDEBUG(D_INODE,
1986                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
1987                        op_tgt, PFID(fid));
1988                 op_data->op_flags |= flag;
1989                 rc = 0;
1990         }
1991
1992         RETURN(rc);
1993 }
1994
1995 /*
1996  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1997  * op_data->op_fid2
1998  */
1999 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2000                     struct ptlrpc_request **request)
2001 {
2002         struct obd_device       *obd = exp->exp_obd;
2003         struct lmv_obd          *lmv = &obd->u.lmv;
2004         struct lmv_tgt_desc     *tgt;
2005         int                      rc;
2006         ENTRY;
2007
2008         LASSERT(op_data->op_namelen != 0);
2009
2010         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2011                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2012                op_data->op_name, PFID(&op_data->op_fid1));
2013
2014         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2015         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2016         op_data->op_cap = cfs_curproc_cap_pack();
2017
2018         tgt = lmv_locate_tgt2(lmv, op_data);
2019         if (IS_ERR(tgt))
2020                 RETURN(PTR_ERR(tgt));
2021
2022         /*
2023          * Cancel UPDATE lock on child (fid1).
2024          */
2025         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2026         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2027                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2028         if (rc != 0)
2029                 RETURN(rc);
2030
2031         rc = md_link(tgt->ltd_exp, op_data, request);
2032
2033         RETURN(rc);
2034 }
2035
2036 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2037                         const char *name, size_t namelen,
2038                         struct ptlrpc_request **request)
2039 {
2040         struct obd_device *obd = exp->exp_obd;
2041         struct lmv_obd *lmv = &obd->u.lmv;
2042         struct lmv_stripe_md *lsm = op_data->op_mea1;
2043         struct lmv_tgt_desc *parent_tgt;
2044         struct lmv_tgt_desc *sp_tgt;
2045         struct lmv_tgt_desc *tp_tgt = NULL;
2046         struct lmv_tgt_desc *child_tgt;
2047         struct lmv_tgt_desc *tgt;
2048         struct lu_fid target_fid = { 0 };
2049         int rc;
2050
2051         ENTRY;
2052
2053         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2054
2055         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2056                PFID(&op_data->op_fid1), (int)namelen, name);
2057
2058         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2059         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2060         op_data->op_cap = cfs_curproc_cap_pack();
2061
2062         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2063         if (IS_ERR(parent_tgt))
2064                 RETURN(PTR_ERR(parent_tgt));
2065
2066         if (lmv_dir_striped(lsm)) {
2067                 const struct lmv_oinfo *oinfo;
2068
2069                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2070                 if (IS_ERR(oinfo))
2071                         RETURN(PTR_ERR(oinfo));
2072
2073                 /* save source stripe FID in fid4 temporarily for ELC */
2074                 op_data->op_fid4 = oinfo->lmo_fid;
2075                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2076                 if (!sp_tgt)
2077                         RETURN(-ENODEV);
2078
2079                 /*
2080                  * if parent is being migrated too, fill op_fid2 with target
2081                  * stripe fid, otherwise the target stripe is not created yet.
2082                  */
2083                 if (lmv_dir_layout_changing(lsm)) {
2084                         oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2085                                                         true);
2086                         if (IS_ERR(oinfo))
2087                                 RETURN(PTR_ERR(oinfo));
2088
2089                         op_data->op_fid2 = oinfo->lmo_fid;
2090                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2091                         if (!tp_tgt)
2092                                 RETURN(-ENODEV);
2093                 }
2094         } else {
2095                 sp_tgt = parent_tgt;
2096         }
2097
2098         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2099         if (IS_ERR(child_tgt))
2100                 RETURN(PTR_ERR(child_tgt));
2101
2102         /* for directory, migrate to MDT specified by lum_stripe_offset;
2103          * otherwise migrate to the target stripe of parent, but parent
2104          * directory may have finished migration (normally current file too),
2105          * allocate FID on MDT lum_stripe_offset, and server will check
2106          * whether file was migrated already.
2107          */
2108         if (S_ISDIR(op_data->op_mode) || !tp_tgt) {
2109                 struct lmv_user_md *lum = op_data->op_data;
2110
2111                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2112         } else  {
2113                 op_data->op_mds = tp_tgt->ltd_index;
2114         }
2115         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2116         if (rc)
2117                 RETURN(rc);
2118
2119         /*
2120          * for directory, send migrate request to the MDT where the object will
2121          * be migrated to, because we can't create a striped directory remotely.
2122          *
2123          * otherwise, send to the MDT where source is located because regular
2124          * file may open lease.
2125          *
2126          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2127          * backward compatibility.
2128          */
2129         if (S_ISDIR(op_data->op_mode) &&
2130             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2131                 tgt = lmv_fid2tgt(lmv, &target_fid);
2132                 if (IS_ERR(tgt))
2133                         RETURN(PTR_ERR(tgt));
2134         } else {
2135                 tgt = child_tgt;
2136         }
2137
2138         /* cancel UPDATE lock of parent master object */
2139         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2140                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2141         if (rc)
2142                 RETURN(rc);
2143
2144         /* cancel UPDATE lock of source parent */
2145         if (sp_tgt != parent_tgt) {
2146                 /*
2147                  * migrate RPC packs master object FID, because we can only pack
2148                  * two FIDs in reint RPC, but MDS needs to know both source
2149                  * parent and target parent, and it will obtain them from master
2150                  * FID and LMV, the other FID in RPC is kept for target.
2151                  *
2152                  * since this FID is not passed to MDC, cancel it anyway.
2153                  */
2154                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2155                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2156                 if (rc)
2157                         RETURN(rc);
2158
2159                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2160         }
2161         op_data->op_fid4 = target_fid;
2162
2163         /* cancel UPDATE locks of target parent */
2164         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2165                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2166         if (rc)
2167                 RETURN(rc);
2168
2169         /* cancel LOOKUP lock of source if source is remote object */
2170         if (child_tgt != sp_tgt) {
2171                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2172                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2173                                       MF_MDC_CANCEL_FID3);
2174                 if (rc)
2175                         RETURN(rc);
2176         }
2177
2178         /* cancel ELC locks of source */
2179         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2180                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2181         if (rc)
2182                 RETURN(rc);
2183
2184         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2185
2186         RETURN(rc);
2187 }
2188
2189 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2190                       const char *old, size_t oldlen,
2191                       const char *new, size_t newlen,
2192                       struct ptlrpc_request **request)
2193 {
2194         struct obd_device *obd = exp->exp_obd;
2195         struct lmv_obd *lmv = &obd->u.lmv;
2196         struct lmv_tgt_desc *sp_tgt;
2197         struct lmv_tgt_desc *tp_tgt = NULL;
2198         struct lmv_tgt_desc *src_tgt = NULL;
2199         struct lmv_tgt_desc *tgt;
2200         struct mdt_body *body;
2201         int rc;
2202
2203         ENTRY;
2204
2205         LASSERT(oldlen != 0);
2206
2207         if (op_data->op_cli_flags & CLI_MIGRATE) {
2208                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2209                 RETURN(rc);
2210         }
2211
2212         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2213         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2214         op_data->op_cap = cfs_curproc_cap_pack();
2215
2216         op_data->op_name = new;
2217         op_data->op_namelen = newlen;
2218
2219         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2220         if (IS_ERR(tp_tgt))
2221                 RETURN(PTR_ERR(tp_tgt));
2222
2223         /* Since the target child might be destroyed, and it might become
2224          * orphan, and we can only check orphan on the local MDT right now, so
2225          * we send rename request to the MDT where target child is located. If
2226          * target child does not exist, then it will send the request to the
2227          * target parent */
2228         if (fid_is_sane(&op_data->op_fid4)) {
2229                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2230                 if (IS_ERR(tgt))
2231                         RETURN(PTR_ERR(tgt));
2232         } else {
2233                 tgt = tp_tgt;
2234         }
2235
2236         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2237
2238         /* cancel UPDATE locks of target parent */
2239         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2240                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2241         if (rc != 0)
2242                 RETURN(rc);
2243
2244         if (fid_is_sane(&op_data->op_fid4)) {
2245                 /* cancel LOOKUP lock of target on target parent */
2246                 if (tgt != tp_tgt) {
2247                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2248                                               tgt->ltd_index, LCK_EX,
2249                                               MDS_INODELOCK_LOOKUP,
2250                                               MF_MDC_CANCEL_FID4);
2251                         if (rc != 0)
2252                                 RETURN(rc);
2253                 }
2254         }
2255
2256         if (fid_is_sane(&op_data->op_fid3)) {
2257                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2258                 if (IS_ERR(src_tgt))
2259                         RETURN(PTR_ERR(src_tgt));
2260
2261                 /* cancel ELC locks of source */
2262                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2263                                       LCK_EX, MDS_INODELOCK_ELC,
2264                                       MF_MDC_CANCEL_FID3);
2265                 if (rc != 0)
2266                         RETURN(rc);
2267         }
2268
2269         op_data->op_name = old;
2270         op_data->op_namelen = oldlen;
2271 retry:
2272         sp_tgt = lmv_locate_tgt(lmv, op_data);
2273         if (IS_ERR(sp_tgt))
2274                 RETURN(PTR_ERR(sp_tgt));
2275
2276         /* cancel UPDATE locks of source parent */
2277         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2278                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2279         if (rc != 0)
2280                 RETURN(rc);
2281
2282         if (fid_is_sane(&op_data->op_fid3)) {
2283                 /* cancel LOOKUP lock of source on source parent */
2284                 if (src_tgt != sp_tgt) {
2285                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2286                                               tgt->ltd_index, LCK_EX,
2287                                               MDS_INODELOCK_LOOKUP,
2288                                               MF_MDC_CANCEL_FID3);
2289                         if (rc != 0)
2290                                 RETURN(rc);
2291                 }
2292         }
2293
2294 rename:
2295         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2296                 PFID(&op_data->op_fid1), (int)oldlen, old,
2297                 PFID(&op_data->op_fid2), (int)newlen, new);
2298
2299         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2300                         request);
2301         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2302                 ptlrpc_req_finished(*request);
2303                 *request = NULL;
2304                 goto retry;
2305         }
2306
2307         if (rc && rc != -EXDEV)
2308                 RETURN(rc);
2309
2310         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2311         if (body == NULL)
2312                 RETURN(-EPROTO);
2313
2314         /* Not cross-ref case, just get out of here. */
2315         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2316                 RETURN(rc);
2317
2318         op_data->op_fid4 = body->mbo_fid1;
2319
2320         ptlrpc_req_finished(*request);
2321         *request = NULL;
2322
2323         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2324         if (IS_ERR(tgt))
2325                 RETURN(PTR_ERR(tgt));
2326
2327         if (fid_is_sane(&op_data->op_fid4)) {
2328                 /* cancel LOOKUP lock of target on target parent */
2329                 if (tgt != tp_tgt) {
2330                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2331                                               tgt->ltd_index, LCK_EX,
2332                                               MDS_INODELOCK_LOOKUP,
2333                                               MF_MDC_CANCEL_FID4);
2334                         if (rc != 0)
2335                                 RETURN(rc);
2336                 }
2337         }
2338
2339         goto rename;
2340 }
2341
2342 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2343                        void *ea, size_t ealen, struct ptlrpc_request **request)
2344 {
2345         struct obd_device *obd = exp->exp_obd;
2346         struct lmv_obd *lmv = &obd->u.lmv;
2347         struct lmv_tgt_desc *tgt;
2348         int rc = 0;
2349
2350         ENTRY;
2351
2352         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2353                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2354                op_data->op_xvalid);
2355
2356         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2357         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2358         if (IS_ERR(tgt))
2359                 RETURN(PTR_ERR(tgt));
2360
2361         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2362
2363         RETURN(rc);
2364 }
2365
2366 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2367                      struct ptlrpc_request **request)
2368 {
2369         struct obd_device *obd = exp->exp_obd;
2370         struct lmv_obd *lmv = &obd->u.lmv;
2371         struct lmv_tgt_desc *tgt;
2372         int rc;
2373
2374         ENTRY;
2375
2376         tgt = lmv_fid2tgt(lmv, fid);
2377         if (IS_ERR(tgt))
2378                 RETURN(PTR_ERR(tgt));
2379
2380         rc = md_fsync(tgt->ltd_exp, fid, request);
2381         RETURN(rc);
2382 }
2383
2384 struct stripe_dirent {
2385         struct page             *sd_page;
2386         struct lu_dirpage       *sd_dp;
2387         struct lu_dirent        *sd_ent;
2388         bool                     sd_eof;
2389 };
2390
2391 struct lmv_dir_ctxt {
2392         struct lmv_obd          *ldc_lmv;
2393         struct md_op_data       *ldc_op_data;
2394         struct md_callback      *ldc_cb_op;
2395         __u64                    ldc_hash;
2396         int                      ldc_count;
2397         struct stripe_dirent     ldc_stripes[0];
2398 };
2399
2400 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2401 {
2402         if (stripe->sd_page) {
2403                 kunmap(stripe->sd_page);
2404                 put_page(stripe->sd_page);
2405                 stripe->sd_page = NULL;
2406                 stripe->sd_ent = NULL;
2407         }
2408 }
2409
2410 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2411 {
2412         int i;
2413
2414         for (i = 0; i < ctxt->ldc_count; i++)
2415                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2416 }
2417
2418 /* if @ent is dummy, or . .., get next */
2419 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2420                                            struct lu_dirent *ent,
2421                                            int stripe_index)
2422 {
2423         for (; ent; ent = lu_dirent_next(ent)) {
2424                 /* Skip dummy entry */
2425                 if (le16_to_cpu(ent->lde_namelen) == 0)
2426                         continue;
2427
2428                 /* skip . and .. for other stripes */
2429                 if (stripe_index &&
2430                     (strncmp(ent->lde_name, ".",
2431                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2432                      strncmp(ent->lde_name, "..",
2433                              le16_to_cpu(ent->lde_namelen)) == 0))
2434                         continue;
2435
2436                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2437                         break;
2438         }
2439
2440         return ent;
2441 }
2442
2443 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2444                                             struct stripe_dirent *stripe,
2445                                             int stripe_index)
2446 {
2447         struct md_op_data *op_data = ctxt->ldc_op_data;
2448         struct lmv_oinfo *oinfo;
2449         struct lu_fid fid = op_data->op_fid1;
2450         struct inode *inode = op_data->op_data;
2451         struct lmv_tgt_desc *tgt;
2452         struct lu_dirent *ent = stripe->sd_ent;
2453         __u64 hash = ctxt->ldc_hash;
2454         int rc = 0;
2455
2456         ENTRY;
2457
2458         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2459         LASSERT(!ent);
2460
2461         do {
2462                 if (stripe->sd_page) {
2463                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2464
2465                         /* @hash should be the last dirent hash */
2466                         LASSERTF(hash <= end,
2467                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2468                                  ctxt, stripe, hash, end);
2469                         /* unload last page */
2470                         stripe_dirent_unload(stripe);
2471                         /* eof */
2472                         if (end == MDS_DIR_END_OFF) {
2473                                 stripe->sd_eof = true;
2474                                 break;
2475                         }
2476                         hash = end;
2477                 }
2478
2479                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2480                 if (!oinfo->lmo_root) {
2481                         rc = -ENOENT;
2482                         break;
2483                 }
2484
2485                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2486                 if (!tgt) {
2487                         rc = -ENODEV;
2488                         break;
2489                 }
2490
2491                 /* op_data is shared by stripes, reset after use */
2492                 op_data->op_fid1 = oinfo->lmo_fid;
2493                 op_data->op_fid2 = oinfo->lmo_fid;
2494                 op_data->op_data = oinfo->lmo_root;
2495
2496                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2497                                   &stripe->sd_page);
2498
2499                 op_data->op_fid1 = fid;
2500                 op_data->op_fid2 = fid;
2501                 op_data->op_data = inode;
2502
2503                 if (rc)
2504                         break;
2505
2506                 stripe->sd_dp = page_address(stripe->sd_page);
2507                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2508                                         stripe_index);
2509                 /* in case a page filled with ., .. and dummy, read next */
2510         } while (!ent);
2511
2512         stripe->sd_ent = ent;
2513         if (rc) {
2514                 LASSERT(!ent);
2515                 /* treat error as eof, so dir can be partially accessed */
2516                 stripe->sd_eof = true;
2517                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2518                               "directory is partially accessed!\n",
2519                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2520                               rc);
2521         }
2522
2523         RETURN(ent);
2524 }
2525
2526 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2527 {
2528         struct obd_device *obd = exp->exp_obd;
2529         struct lmv_obd *lmv = &obd->u.lmv;
2530         struct lmv_tgt_desc *tgt;
2531         int rc;
2532
2533         ENTRY;
2534
2535         rc = lmv_check_connect(obd);
2536         if (rc != 0)
2537                 RETURN(rc);
2538
2539         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2540         if (IS_ERR(tgt))
2541                 RETURN(PTR_ERR(tgt));
2542
2543         data->op_flags |= MF_MDC_CANCEL_FID1;
2544         rc = md_file_resync(tgt->ltd_exp, data);
2545         RETURN(rc);
2546 }
2547
2548 /**
2549  * Get dirent with the closest hash for striped directory
2550  *
2551  * This function will search the dir entry, whose hash value is the
2552  * closest(>=) to hash from all of sub-stripes, and it is only being called
2553  * for striped directory.
2554  *
2555  * \param[in] ctxt              dir read context
2556  *
2557  * \retval                      dirent get the entry successfully
2558  *                              NULL does not get the entry, normally it means
2559  *                              it reaches the end of the directory, while read
2560  *                              stripe dirent error is ignored to allow partial
2561  *                              access.
2562  */
2563 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2564 {
2565         struct stripe_dirent *stripe;
2566         struct lu_dirent *ent = NULL;
2567         int i;
2568         int min = -1;
2569
2570         /* TODO: optimize with k-way merge sort */
2571         for (i = 0; i < ctxt->ldc_count; i++) {
2572                 stripe = &ctxt->ldc_stripes[i];
2573                 if (stripe->sd_eof)
2574                         continue;
2575
2576                 if (!stripe->sd_ent) {
2577                         stripe_dirent_load(ctxt, stripe, i);
2578                         if (!stripe->sd_ent) {
2579                                 LASSERT(stripe->sd_eof);
2580                                 continue;
2581                         }
2582                 }
2583
2584                 if (min == -1 ||
2585                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2586                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2587                         min = i;
2588                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2589                             ctxt->ldc_hash)
2590                                 break;
2591                 }
2592         }
2593
2594         if (min != -1) {
2595                 stripe = &ctxt->ldc_stripes[min];
2596                 ent = stripe->sd_ent;
2597                 /* pop found dirent */
2598                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2599                                                    min);
2600         }
2601
2602         return ent;
2603 }
2604
2605 /**
2606  * Build dir entry page for striped directory
2607  *
2608  * This function gets one entry by @offset from a striped directory. It will
2609  * read entries from all of stripes, and choose one closest to the required
2610  * offset(&offset). A few notes
2611  * 1. skip . and .. for non-zero stripes, because there can only have one .
2612  * and .. in a directory.
2613  * 2. op_data will be shared by all of stripes, instead of allocating new
2614  * one, so need to restore before reusing.
2615  *
2616  * \param[in] exp       obd export refer to LMV
2617  * \param[in] op_data   hold those MD parameters of read_entry
2618  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2619  * \param[in] offset    starting hash offset
2620  * \param[out] ppage    the page holding the entry. Note: because the entry
2621  *                      will be accessed in upper layer, so we need hold the
2622  *                      page until the usages of entry is finished, see
2623  *                      ll_dir_entry_next.
2624  *
2625  * retval               =0 if get entry successfully
2626  *                      <0 cannot get entry
2627  */
2628 static int lmv_striped_read_page(struct obd_export *exp,
2629                                  struct md_op_data *op_data,
2630                                  struct md_callback *cb_op,
2631                                  __u64 offset, struct page **ppage)
2632 {
2633         struct page *page = NULL;
2634         struct lu_dirpage *dp;
2635         void *start;
2636         struct lu_dirent *ent;
2637         struct lu_dirent *last_ent;
2638         int stripe_count;
2639         struct lmv_dir_ctxt *ctxt;
2640         struct lu_dirent *next = NULL;
2641         __u16 ent_size;
2642         size_t left_bytes;
2643         int rc = 0;
2644         ENTRY;
2645
2646         /* Allocate a page and read entries from all of stripes and fill
2647          * the page by hash order */
2648         page = alloc_page(GFP_KERNEL);
2649         if (!page)
2650                 RETURN(-ENOMEM);
2651
2652         /* Initialize the entry page */
2653         dp = kmap(page);
2654         memset(dp, 0, sizeof(*dp));
2655         dp->ldp_hash_start = cpu_to_le64(offset);
2656
2657         start = dp + 1;
2658         left_bytes = PAGE_SIZE - sizeof(*dp);
2659         ent = start;
2660         last_ent = ent;
2661
2662         /* initalize dir read context */
2663         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2664         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2665         if (!ctxt)
2666                 GOTO(free_page, rc = -ENOMEM);
2667         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2668         ctxt->ldc_op_data = op_data;
2669         ctxt->ldc_cb_op = cb_op;
2670         ctxt->ldc_hash = offset;
2671         ctxt->ldc_count = stripe_count;
2672
2673         while (1) {
2674                 next = lmv_dirent_next(ctxt);
2675
2676                 /* end of directory */
2677                 if (!next) {
2678                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2679                         break;
2680                 }
2681                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2682
2683                 ent_size = le16_to_cpu(next->lde_reclen);
2684
2685                 /* the last entry lde_reclen is 0, but it might not be the last
2686                  * one of this temporay dir page */
2687                 if (!ent_size)
2688                         ent_size = lu_dirent_calc_size(
2689                                         le16_to_cpu(next->lde_namelen),
2690                                         le32_to_cpu(next->lde_attrs));
2691                 /* page full */
2692                 if (ent_size > left_bytes)
2693                         break;
2694
2695                 memcpy(ent, next, ent_size);
2696
2697                 /* Replace . with master FID and Replace .. with the parent FID
2698                  * of master object */
2699                 if (strncmp(ent->lde_name, ".",
2700                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2701                     le16_to_cpu(ent->lde_namelen) == 1)
2702                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2703                 else if (strncmp(ent->lde_name, "..",
2704                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2705                            le16_to_cpu(ent->lde_namelen) == 2)
2706                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2707
2708                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2709                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2710                        le64_to_cpu(ent->lde_hash));
2711
2712                 left_bytes -= ent_size;
2713                 ent->lde_reclen = cpu_to_le16(ent_size);
2714                 last_ent = ent;
2715                 ent = (void *)ent + ent_size;
2716         };
2717
2718         last_ent->lde_reclen = 0;
2719
2720         if (ent == start)
2721                 dp->ldp_flags |= LDF_EMPTY;
2722         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2723                 dp->ldp_flags |= LDF_COLLIDE;
2724         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2725         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2726
2727         put_lmv_dir_ctxt(ctxt);
2728         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2729
2730         *ppage = page;
2731
2732         RETURN(0);
2733
2734 free_page:
2735         kunmap(page);
2736         __free_page(page);
2737
2738         return rc;
2739 }
2740
2741 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2742                   struct md_callback *cb_op, __u64 offset,
2743                   struct page **ppage)
2744 {
2745         struct obd_device *obd = exp->exp_obd;
2746         struct lmv_obd *lmv = &obd->u.lmv;
2747         struct lmv_tgt_desc *tgt;
2748         int rc;
2749
2750         ENTRY;
2751
2752         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2753                 RETURN(-ENODATA);
2754
2755         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2756                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2757                 RETURN(rc);
2758         }
2759
2760         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2761         if (IS_ERR(tgt))
2762                 RETURN(PTR_ERR(tgt));
2763
2764         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2765
2766         RETURN(rc);
2767 }
2768
2769 /**
2770  * Unlink a file/directory
2771  *
2772  * Unlink a file or directory under the parent dir. The unlink request
2773  * usually will be sent to the MDT where the child is located, but if
2774  * the client does not have the child FID then request will be sent to the
2775  * MDT where the parent is located.
2776  *
2777  * If the parent is a striped directory then it also needs to locate which
2778  * stripe the name of the child is located, and replace the parent FID
2779  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2780  * it will walk through all of sub-stripes until the child is being
2781  * unlinked finally.
2782  *
2783  * \param[in] exp       export refer to LMV
2784  * \param[in] op_data   different parameters transferred beween client
2785  *                      MD stacks, name, namelen, FIDs etc.
2786  *                      op_fid1 is the parent FID, op_fid2 is the child
2787  *                      FID.
2788  * \param[out] request  point to the request of unlink.
2789  *
2790  * retval               0 if succeed
2791  *                      negative errno if failed.
2792  */
2793 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2794                       struct ptlrpc_request **request)
2795 {
2796         struct obd_device *obd = exp->exp_obd;
2797         struct lmv_obd *lmv = &obd->u.lmv;
2798         struct lmv_tgt_desc *tgt;
2799         struct lmv_tgt_desc *parent_tgt;
2800         struct mdt_body *body;
2801         int rc;
2802
2803         ENTRY;
2804
2805         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2806         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2807         op_data->op_cap = cfs_curproc_cap_pack();
2808
2809 retry:
2810         parent_tgt = lmv_locate_tgt(lmv, op_data);
2811         if (IS_ERR(parent_tgt))
2812                 RETURN(PTR_ERR(parent_tgt));
2813
2814         if (likely(!fid_is_zero(&op_data->op_fid2))) {
2815                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2816                 if (IS_ERR(tgt))
2817                         RETURN(PTR_ERR(tgt));
2818         } else {
2819                 tgt = parent_tgt;
2820         }
2821
2822         /*
2823          * If child's fid is given, cancel unused locks for it if it is from
2824          * another export than parent.
2825          *
2826          * LOOKUP lock for child (fid3) should also be cancelled on parent
2827          * tgt_tgt in mdc_unlink().
2828          */
2829         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2830
2831         if (parent_tgt != tgt)
2832                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2833                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2834                                       MF_MDC_CANCEL_FID3);
2835
2836         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2837                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2838         if (rc)
2839                 RETURN(rc);
2840
2841         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2842                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2843                tgt->ltd_index);
2844
2845         rc = md_unlink(tgt->ltd_exp, op_data, request);
2846         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2847                 ptlrpc_req_finished(*request);
2848                 *request = NULL;
2849                 goto retry;
2850         }
2851
2852         if (rc != -EREMOTE)
2853                 RETURN(rc);
2854
2855         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2856         if (body == NULL)
2857                 RETURN(-EPROTO);
2858
2859         /* Not cross-ref case, just get out of here. */
2860         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2861                 RETURN(rc);
2862
2863         /* This is a remote object, try remote MDT. */
2864         op_data->op_fid2 = body->mbo_fid1;
2865         ptlrpc_req_finished(*request);
2866         *request = NULL;
2867
2868         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2869         if (IS_ERR(tgt))
2870                 RETURN(PTR_ERR(tgt));
2871
2872         goto retry;
2873 }
2874
2875 static int lmv_precleanup(struct obd_device *obd)
2876 {
2877         ENTRY;
2878         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2879         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2880         lprocfs_obd_cleanup(obd);
2881         lprocfs_free_md_stats(obd);
2882         RETURN(0);
2883 }
2884
2885 /**
2886  * Get by key a value associated with a LMV device.
2887  *
2888  * Dispatch request to lower-layer devices as needed.
2889  *
2890  * \param[in] env               execution environment for this thread
2891  * \param[in] exp               export for the LMV device
2892  * \param[in] keylen            length of key identifier
2893  * \param[in] key               identifier of key to get value for
2894  * \param[in] vallen            size of \a val
2895  * \param[out] val              pointer to storage location for value
2896  * \param[in] lsm               optional striping metadata of object
2897  *
2898  * \retval 0            on success
2899  * \retval negative     negated errno on failure
2900  */
2901 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2902                         __u32 keylen, void *key, __u32 *vallen, void *val)
2903 {
2904         struct obd_device *obd;
2905         struct lmv_obd *lmv;
2906         struct lu_tgt_desc *tgt;
2907         int rc = 0;
2908
2909         ENTRY;
2910
2911         obd = class_exp2obd(exp);
2912         if (obd == NULL) {
2913                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2914                        exp->exp_handle.h_cookie);
2915                 RETURN(-EINVAL);
2916         }
2917
2918         lmv = &obd->u.lmv;
2919         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2920                 LASSERT(*vallen == sizeof(__u32));
2921                 lmv_foreach_connected_tgt(lmv, tgt) {
2922                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2923                                           vallen, val))
2924                                 RETURN(0);
2925                 }
2926                 RETURN(-EINVAL);
2927         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2928                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2929                    KEY_IS(KEY_CONN_DATA)) {
2930                 /*
2931                  * Forwarding this request to first MDS, it should know LOV
2932                  * desc.
2933                  */
2934                 tgt = lmv_tgt(lmv, 0);
2935                 if (!tgt)
2936                         RETURN(-ENODEV);
2937
2938                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
2939                 if (!rc && KEY_IS(KEY_CONN_DATA))
2940                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2941                 RETURN(rc);
2942         } else if (KEY_IS(KEY_TGT_COUNT)) {
2943                 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
2944                 RETURN(0);
2945         }
2946
2947         CDEBUG(D_IOCTL, "Invalid key\n");
2948         RETURN(-EINVAL);
2949 }
2950
2951 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
2952                      int *__rcs, struct ptlrpc_request_set *_set)
2953 {
2954         struct obd_device *obd = class_exp2obd(exp);
2955         struct ptlrpc_request_set *set = _set;
2956         struct lmv_obd *lmv = &obd->u.lmv;
2957         int tgt_count = lmv->lmv_mdt_count;
2958         struct lu_tgt_desc *tgt;
2959         struct fid_array *fat, **fas = NULL;
2960         int i, rc, **rcs = NULL;
2961
2962         if (!set) {
2963                 set = ptlrpc_prep_set();
2964                 if (!set)
2965                         RETURN(-ENOMEM);
2966         }
2967
2968         /* split FIDs by targets */
2969         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
2970         if (fas == NULL)
2971                 GOTO(out, rc = -ENOMEM);
2972         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
2973         if (rcs == NULL)
2974                 GOTO(out_fas, rc = -ENOMEM);
2975
2976         for (i = 0; i < fa->fa_nr; i++) {
2977                 unsigned int idx;
2978
2979                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
2980                 if (rc) {
2981                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
2982                                PFID(&fa->fa_fids[i]), rc);
2983                         continue;
2984                 }
2985                 LASSERT(idx < tgt_count);
2986                 if (!fas[idx])
2987                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
2988                                   fa_fids[fa->fa_nr]));
2989                 if (!fas[idx])
2990                         GOTO(out, rc = -ENOMEM);
2991                 if (!rcs[idx])
2992                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
2993                 if (!rcs[idx])
2994                         GOTO(out, rc = -ENOMEM);
2995
2996                 fat = fas[idx];
2997                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
2998         }
2999
3000         lmv_foreach_connected_tgt(lmv, tgt) {
3001                 fat = fas[tgt->ltd_index];
3002                 if (!fat || fat->fa_nr == 0)
3003                         continue;
3004                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3005         }
3006
3007         rc = ptlrpc_set_wait(NULL, set);
3008         if (rc == 0) {
3009                 int j = 0;
3010                 for (i = 0; i < tgt_count; i++) {
3011                         fat = fas[i];
3012                         if (!fat || fat->fa_nr == 0)
3013                                 continue;
3014                         /* copy FIDs back */
3015                         memcpy(fa->fa_fids + j, fat->fa_fids,
3016                                fat->fa_nr * sizeof(struct lu_fid));
3017                         /* copy rcs back */
3018                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3019                         j += fat->fa_nr;
3020                 }
3021         }
3022         if (set != _set)
3023                 ptlrpc_set_destroy(set);
3024
3025 out:
3026         for (i = 0; i < tgt_count; i++) {
3027                 if (fas && fas[i])
3028                         OBD_FREE(fas[i], offsetof(struct fid_array,
3029                                                 fa_fids[fa->fa_nr]));
3030                 if (rcs && rcs[i])
3031                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3032         }
3033         if (rcs)
3034                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3035 out_fas:
3036         if (fas)
3037                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3038
3039         RETURN(rc);
3040 }
3041
3042 /**
3043  * Asynchronously set by key a value associated with a LMV device.
3044  *
3045  * Dispatch request to lower-layer devices as needed.
3046  *
3047  * \param[in] env       execution environment for this thread
3048  * \param[in] exp       export for the LMV device
3049  * \param[in] keylen    length of key identifier
3050  * \param[in] key       identifier of key to store value for
3051  * \param[in] vallen    size of value to store
3052  * \param[in] val       pointer to data to be stored
3053  * \param[in] set       optional list of related ptlrpc requests
3054  *
3055  * \retval 0            on success
3056  * \retval negative     negated errno on failure
3057  */
3058 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3059                         __u32 keylen, void *key, __u32 vallen, void *val,
3060                         struct ptlrpc_request_set *set)
3061 {
3062         struct lmv_tgt_desc     *tgt = NULL;
3063         struct obd_device       *obd;
3064         struct lmv_obd          *lmv;
3065         int rc = 0;
3066         ENTRY;
3067
3068         obd = class_exp2obd(exp);
3069         if (obd == NULL) {
3070                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3071                        exp->exp_handle.h_cookie);
3072                 RETURN(-EINVAL);
3073         }
3074         lmv = &obd->u.lmv;
3075
3076         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3077             KEY_IS(KEY_DEFAULT_EASIZE)) {
3078                 int err = 0;
3079
3080                 lmv_foreach_connected_tgt(lmv, tgt) {
3081                         err = obd_set_info_async(env, tgt->ltd_exp,
3082                                                  keylen, key, vallen, val, set);
3083                         if (err && rc == 0)
3084                                 rc = err;
3085                 }
3086
3087                 RETURN(rc);
3088         }
3089
3090         RETURN(-EINVAL);
3091 }
3092
3093 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3094                             const struct lmv_mds_md_v1 *lmm1)
3095 {
3096         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3097         int             stripe_count;
3098         int             cplen;
3099         int             i;
3100         int             rc = 0;
3101         ENTRY;
3102
3103         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3104         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3105         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3106         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3107                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3108         else
3109                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3110         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3111         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3112         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3113         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3114                         sizeof(lsm->lsm_md_pool_name));
3115
3116         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3117                 RETURN(-E2BIG);
3118
3119         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3120                "layout_version %d\n", lsm->lsm_md_stripe_count,
3121                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3122                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3123                lsm->lsm_md_layout_version);
3124
3125         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3126         for (i = 0; i < stripe_count; i++) {
3127                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3128                               &lmm1->lmv_stripe_fids[i]);
3129                 /*
3130                  * set default value -1, so lmv_locate_tgt() knows this stripe
3131                  * target is not initialized.
3132                  */
3133                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3134                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3135                         continue;
3136
3137                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3138                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3139                 if (rc == -ENOENT)
3140                         continue;
3141
3142                 if (rc)
3143                         RETURN(rc);
3144
3145                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3146                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3147         }
3148
3149         RETURN(rc);
3150 }
3151
3152 static inline int lmv_unpack_user_md(struct obd_export *exp,
3153                                      struct lmv_stripe_md *lsm,
3154                                      const struct lmv_user_md *lmu)
3155 {
3156         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3157         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3158         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3159         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3160         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3161
3162         return 0;
3163 }
3164
3165 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3166                         const union lmv_mds_md *lmm, size_t lmm_size)
3167 {
3168         struct lmv_stripe_md     *lsm;
3169         int                      lsm_size;
3170         int                      rc;
3171         bool                     allocated = false;
3172         ENTRY;
3173
3174         LASSERT(lsmp != NULL);
3175
3176         lsm = *lsmp;
3177         /* Free memmd */
3178         if (lsm != NULL && lmm == NULL) {
3179                 int i;
3180                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3181
3182                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3183                         size_t lfm_size;
3184
3185                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3186                                                               lfm_value[0]);
3187                         OBD_FREE_LARGE(lfm, lfm_size);
3188                         RETURN(0);
3189                 }
3190
3191                 if (lmv_dir_striped(lsm)) {
3192                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3193                                 if (lsm->lsm_md_oinfo[i].lmo_root)
3194                                         iput(lsm->lsm_md_oinfo[i].lmo_root);
3195                         }
3196                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3197                 } else {
3198                         lsm_size = lmv_stripe_md_size(0);
3199                 }
3200                 OBD_FREE(lsm, lsm_size);
3201                 *lsmp = NULL;
3202                 RETURN(0);
3203         }
3204
3205         /* foreign lmv case */
3206         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3207                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3208
3209                 if (lfm == NULL) {
3210                         OBD_ALLOC_LARGE(lfm, lmm_size);
3211                         if (lfm == NULL)
3212                                 RETURN(-ENOMEM);
3213                         *lsmp = (struct lmv_stripe_md *)lfm;
3214                 }
3215                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3216                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3217                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3218                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3219                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3220                        lfm->lfm_length);
3221                 RETURN(lmm_size);
3222         }
3223
3224         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3225                 RETURN(-EPERM);
3226
3227         /* Unpack memmd */
3228         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3229             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3230                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3231                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3232                        -EIO);
3233                 RETURN(-EIO);
3234         }
3235
3236         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3237                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3238         else
3239                 /**
3240                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3241                  * stripecount should be 0 then.
3242                  */
3243                 lsm_size = lmv_stripe_md_size(0);
3244
3245         if (lsm == NULL) {
3246                 OBD_ALLOC(lsm, lsm_size);
3247                 if (lsm == NULL)
3248                         RETURN(-ENOMEM);
3249                 allocated = true;
3250                 *lsmp = lsm;
3251         }
3252
3253         switch (le32_to_cpu(lmm->lmv_magic)) {
3254         case LMV_MAGIC_V1:
3255                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3256                 break;
3257         case LMV_USER_MAGIC:
3258                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3259                 break;
3260         default:
3261                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3262                        le32_to_cpu(lmm->lmv_magic));
3263                 rc = -EINVAL;
3264                 break;
3265         }
3266
3267         if (rc != 0 && allocated) {
3268                 OBD_FREE(lsm, lsm_size);
3269                 *lsmp = NULL;
3270                 lsm_size = rc;
3271         }
3272         RETURN(lsm_size);
3273 }
3274
3275 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3276 {
3277         lmv_unpackmd(NULL, &lsm, NULL, 0);
3278 }
3279 EXPORT_SYMBOL(lmv_free_memmd);
3280
3281 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3282                              union ldlm_policy_data *policy,
3283                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3284                              void *opaque)
3285 {
3286         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3287         struct lu_tgt_desc *tgt;
3288         int err;
3289         int rc = 0;
3290
3291         ENTRY;
3292
3293         LASSERT(fid != NULL);
3294
3295         lmv_foreach_connected_tgt(lmv, tgt) {
3296                 if (!tgt->ltd_active)
3297                         continue;
3298