Whamcloud - gitweb
LU-12624 obdclass: lu_tgt_descs cleanup
[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 #ifdef HAVE_UIDGID_HEADER
40 # include <linux/uidgid.h>
41 #endif
42 #include <linux/slab.h>
43 #include <linux/pagemap.h>
44 #include <linux/mm.h>
45 #include <linux/math64.h>
46 #include <linux/seq_file.h>
47 #include <linux/namei.h>
48
49 #include <obd_support.h>
50 #include <lustre_lib.h>
51 #include <lustre_net.h>
52 #include <obd_class.h>
53 #include <lustre_lmv.h>
54 #include <lprocfs_status.h>
55 #include <cl_object.h>
56 #include <lustre_fid.h>
57 #include <uapi/linux/lustre/lustre_ioctl.h>
58 #include <lustre_kernelcomm.h>
59 #include "lmv_internal.h"
60
61 static int lmv_check_connect(struct obd_device *obd);
62
63 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
64                          int activate)
65 {
66         if (tgt->ltd_active == activate)
67                 return;
68
69         tgt->ltd_active = activate;
70         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
71                 (activate ? 1 : -1);
72
73         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
74 }
75
76 /**
77  * Error codes:
78  *
79  *  -EINVAL  : UUID can't be found in the LMV's target list
80  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
81  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
82  */
83 static int lmv_set_mdc_active(struct lmv_obd *lmv,
84                               const struct obd_uuid *uuid,
85                               int activate)
86 {
87         struct lu_tgt_desc *tgt = NULL;
88         struct obd_device *obd;
89         int rc = 0;
90
91         ENTRY;
92
93         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
94                         lmv, uuid->uuid, activate);
95
96         spin_lock(&lmv->lmv_lock);
97         lmv_foreach_connected_tgt(lmv, tgt) {
98                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
99                        tgt->ltd_index, tgt->ltd_uuid.uuid,
100                        tgt->ltd_exp->exp_handle.h_cookie);
101
102                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
103                         break;
104         }
105
106         if (!tgt)
107                 GOTO(out_lmv_lock, rc = -EINVAL);
108
109         obd = class_exp2obd(tgt->ltd_exp);
110         if (obd == NULL)
111                 GOTO(out_lmv_lock, rc = -ENOTCONN);
112
113         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
114                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
115                obd->obd_type->typ_name, tgt->ltd_index);
116         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
117
118         if (tgt->ltd_active == activate) {
119                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
120                        activate ? "" : "in");
121                 GOTO(out_lmv_lock, rc);
122         }
123
124         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
125                activate ? "" : "in");
126         lmv_activate_target(lmv, tgt, activate);
127         EXIT;
128
129  out_lmv_lock:
130         spin_unlock(&lmv->lmv_lock);
131         return rc;
132 }
133
134 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
135 {
136         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
137         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
138
139         return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
140 }
141
142 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
143                       enum obd_notify_event ev)
144 {
145         struct obd_connect_data *conn_data;
146         struct lmv_obd          *lmv = &obd->u.lmv;
147         struct obd_uuid         *uuid;
148         int                      rc = 0;
149         ENTRY;
150
151         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
152                 CERROR("unexpected notification of %s %s!\n",
153                        watched->obd_type->typ_name,
154                        watched->obd_name);
155                 RETURN(-EINVAL);
156         }
157
158         uuid = &watched->u.cli.cl_target_uuid;
159         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
160                 /*
161                  * Set MDC as active before notifying the observer, so the
162                  * observer can use the MDC normally.
163                  */
164                 rc = lmv_set_mdc_active(lmv, uuid,
165                                         ev == OBD_NOTIFY_ACTIVE);
166                 if (rc) {
167                         CERROR("%sactivation of %s failed: %d\n",
168                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
169                                uuid->uuid, rc);
170                         RETURN(rc);
171                 }
172         } else if (ev == OBD_NOTIFY_OCD) {
173                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
174                 /*
175                  * XXX: Make sure that ocd_connect_flags from all targets are
176                  * the same. Otherwise one of MDTs runs wrong version or
177                  * something like this.  --umka
178                  */
179                 obd->obd_self_export->exp_connect_data = *conn_data;
180         }
181
182         /*
183          * Pass the notification up the chain.
184          */
185         if (obd->obd_observer)
186                 rc = obd_notify(obd->obd_observer, watched, ev);
187
188         RETURN(rc);
189 }
190
191 static int lmv_connect(const struct lu_env *env,
192                        struct obd_export **pexp, struct obd_device *obd,
193                        struct obd_uuid *cluuid, struct obd_connect_data *data,
194                        void *localdata)
195 {
196         struct lmv_obd *lmv = &obd->u.lmv;
197         struct lustre_handle conn = { 0 };
198         struct obd_export *exp;
199         int rc;
200         ENTRY;
201
202         rc = class_connect(&conn, obd, cluuid);
203         if (rc) {
204                 CERROR("class_connection() returned %d\n", rc);
205                 RETURN(rc);
206         }
207
208         exp = class_conn2export(&conn);
209
210         lmv->connected = 0;
211         lmv->conn_data = *data;
212         lmv->lmv_cache = localdata;
213
214         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
215                                                     &obd->obd_kset.kobj);
216         if (!lmv->lmv_tgts_kobj) {
217                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
218                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
219         }
220
221         rc = lmv_check_connect(obd);
222         if (rc != 0)
223                 GOTO(out_sysfs, rc);
224
225         *pexp = exp;
226
227         RETURN(rc);
228
229 out_sysfs:
230         if (lmv->lmv_tgts_kobj)
231                 kobject_put(lmv->lmv_tgts_kobj);
232
233         class_disconnect(exp);
234
235         return rc;
236 }
237
238 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
239                             __u32 def_easize)
240 {
241         struct obd_device *obd = exp->exp_obd;
242         struct lmv_obd *lmv = &obd->u.lmv;
243         struct lmv_tgt_desc *tgt;
244         int change = 0;
245         int rc = 0;
246
247         ENTRY;
248
249         if (lmv->max_easize < easize) {
250                 lmv->max_easize = easize;
251                 change = 1;
252         }
253         if (lmv->max_def_easize < def_easize) {
254                 lmv->max_def_easize = def_easize;
255                 change = 1;
256         }
257
258         if (change == 0)
259                 RETURN(0);
260
261         if (lmv->connected == 0)
262                 RETURN(0);
263
264         lmv_foreach_connected_tgt(lmv, tgt) {
265                 if (!tgt->ltd_active)
266                         continue;
267
268                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
269                 if (rc) {
270                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
271                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
272                         break;
273                 }
274         }
275         RETURN(rc);
276 }
277
278 #define MAX_STRING_SIZE 128
279
280 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
281 {
282         struct lmv_obd          *lmv = &obd->u.lmv;
283         struct obd_device       *mdc_obd;
284         struct obd_export       *mdc_exp;
285         struct lu_fld_target     target;
286         int                      rc;
287         ENTRY;
288
289         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
290                                         &obd->obd_uuid);
291         if (!mdc_obd) {
292                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
293                 RETURN(-EINVAL);
294         }
295
296         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
297                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
298                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
299
300         if (!mdc_obd->obd_set_up) {
301                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
302                 RETURN(-EINVAL);
303         }
304
305         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
306                          &lmv->conn_data, lmv->lmv_cache);
307         if (rc) {
308                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
309                 RETURN(rc);
310         }
311
312         /*
313          * Init fid sequence client for this mdc and add new fld target.
314          */
315         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
316         if (rc)
317                 RETURN(rc);
318
319         target.ft_srv = NULL;
320         target.ft_exp = mdc_exp;
321         target.ft_idx = tgt->ltd_index;
322
323         fld_client_add_target(&lmv->lmv_fld, &target);
324
325         rc = obd_register_observer(mdc_obd, obd);
326         if (rc) {
327                 obd_disconnect(mdc_exp);
328                 CERROR("target %s register_observer error %d\n",
329                        tgt->ltd_uuid.uuid, rc);
330                 RETURN(rc);
331         }
332
333         if (obd->obd_observer) {
334                 /*
335                  * Tell the observer about the new target.
336                  */
337                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
338                                 OBD_NOTIFY_ACTIVE);
339                 if (rc) {
340                         obd_disconnect(mdc_exp);
341                         RETURN(rc);
342                 }
343         }
344
345         tgt->ltd_active = 1;
346         tgt->ltd_exp = mdc_exp;
347         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
348
349         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
350
351         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
352         if (rc) {
353                 obd_disconnect(mdc_exp);
354                 RETURN(rc);
355         }
356
357         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
358                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
359                 atomic_read(&obd->obd_refcount));
360
361         lmv_statfs_check_update(obd, tgt);
362
363         if (lmv->lmv_tgts_kobj)
364                 /* Even if we failed to create the link, that's fine */
365                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
366                                        &mdc_obd->obd_kset.kobj,
367                                        mdc_obd->obd_name);
368         RETURN(0);
369 }
370
371 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
372 {
373         LASSERT(tgt);
374         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
375         OBD_FREE_PTR(tgt);
376 }
377
378 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
379                            __u32 index, int gen)
380 {
381         struct obd_device *mdc_obd;
382         struct lmv_obd *lmv = &obd->u.lmv;
383         struct lmv_tgt_desc *tgt;
384         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
385         int rc = 0;
386
387         ENTRY;
388
389         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
390         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
391                                         &obd->obd_uuid);
392         if (!mdc_obd) {
393                 CERROR("%s: Target %s not attached: rc = %d\n",
394                        obd->obd_name, uuidp->uuid, -EINVAL);
395                 RETURN(-EINVAL);
396         }
397
398         OBD_ALLOC_PTR(tgt);
399         if (!tgt)
400                 RETURN(-ENOMEM);
401
402         mutex_init(&tgt->ltd_fid_mutex);
403         tgt->ltd_index = index;
404         tgt->ltd_uuid = *uuidp;
405         tgt->ltd_active = 0;
406
407         mutex_lock(&ltd->ltd_mutex);
408         rc = ltd_add_tgt(ltd, tgt);
409         mutex_unlock(&ltd->ltd_mutex);
410
411         if (rc)
412                 GOTO(out_tgt, rc);
413
414         if (!lmv->connected)
415                 /* lmv_check_connect() will connect this target. */
416                 RETURN(0);
417
418         rc = lmv_connect_mdc(obd, tgt);
419         if (!rc) {
420                 int easize = sizeof(struct lmv_stripe_md) +
421                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
422
423                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
424         }
425
426         RETURN(rc);
427
428 out_tgt:
429         OBD_FREE_PTR(tgt);
430         return rc;
431 }
432
433 static int lmv_check_connect(struct obd_device *obd)
434 {
435         struct lmv_obd *lmv = &obd->u.lmv;
436         struct lmv_tgt_desc *tgt;
437         int easize;
438         int rc;
439
440         ENTRY;
441
442         if (lmv->connected)
443                 RETURN(0);
444
445         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
446         if (lmv->connected)
447                 GOTO(unlock, rc = 0);
448
449         if (!lmv->lmv_mdt_count) {
450                 CERROR("%s: no targets configured: rc = -EINVAL\n",
451                        obd->obd_name);
452                 GOTO(unlock, rc = -EINVAL);
453         }
454
455         if (!lmv_mdt0_inited(lmv)) {
456                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
457                        obd->obd_name);
458                 GOTO(unlock, rc = -EINVAL);
459         }
460
461         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
462                obd->obd_uuid.uuid, obd->obd_name);
463
464         lmv_foreach_tgt(lmv, tgt) {
465                 rc = lmv_connect_mdc(obd, tgt);
466                 if (rc)
467                         GOTO(out_disc, rc);
468         }
469
470         lmv->connected = 1;
471         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
472         lmv_init_ea_size(obd->obd_self_export, easize, 0);
473         EXIT;
474 unlock:
475         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
476
477         return rc;
478
479 out_disc:
480         lmv_foreach_tgt(lmv, tgt) {
481                 tgt->ltd_active = 0;
482                 if (!tgt->ltd_exp)
483                         continue;
484
485                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
486                 obd_disconnect(tgt->ltd_exp);
487         }
488
489         goto unlock;
490 }
491
492 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
493 {
494         struct lmv_obd         *lmv = &obd->u.lmv;
495         struct obd_device      *mdc_obd;
496         int                     rc;
497         ENTRY;
498
499         LASSERT(tgt != NULL);
500         LASSERT(obd != NULL);
501
502         mdc_obd = class_exp2obd(tgt->ltd_exp);
503
504         if (mdc_obd) {
505                 mdc_obd->obd_force = obd->obd_force;
506                 mdc_obd->obd_fail = obd->obd_fail;
507                 mdc_obd->obd_no_recov = obd->obd_no_recov;
508
509                 if (lmv->lmv_tgts_kobj)
510                         sysfs_remove_link(lmv->lmv_tgts_kobj,
511                                           mdc_obd->obd_name);
512         }
513
514         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
515         if (rc)
516                 CERROR("Can't finanize fids factory\n");
517
518         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
519                tgt->ltd_exp->exp_obd->obd_name,
520                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
521
522         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
523         rc = obd_disconnect(tgt->ltd_exp);
524         if (rc) {
525                 if (tgt->ltd_active) {
526                         CERROR("Target %s disconnect error %d\n",
527                                tgt->ltd_uuid.uuid, rc);
528                 }
529         }
530
531         lmv_activate_target(lmv, tgt, 0);
532         tgt->ltd_exp = NULL;
533         RETURN(0);
534 }
535
536 static int lmv_disconnect(struct obd_export *exp)
537 {
538         struct obd_device *obd = class_exp2obd(exp);
539         struct lmv_obd *lmv = &obd->u.lmv;
540         struct lmv_tgt_desc *tgt;
541         int rc;
542
543         ENTRY;
544
545         lmv_foreach_connected_tgt(lmv, tgt)
546                 lmv_disconnect_mdc(obd, tgt);
547
548         if (lmv->lmv_tgts_kobj)
549                 kobject_put(lmv->lmv_tgts_kobj);
550
551         if (!lmv->connected)
552                 class_export_put(exp);
553         rc = class_disconnect(exp);
554         lmv->connected = 0;
555
556         RETURN(rc);
557 }
558
559 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
560                         void __user *uarg)
561 {
562         struct obd_device *obddev = class_exp2obd(exp);
563         struct lmv_obd *lmv = &obddev->u.lmv;
564         struct getinfo_fid2path *gf;
565         struct lmv_tgt_desc *tgt;
566         struct getinfo_fid2path *remote_gf = NULL;
567         struct lu_fid root_fid;
568         int remote_gf_size = 0;
569         int rc;
570
571         gf = karg;
572         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
573         if (IS_ERR(tgt))
574                 RETURN(PTR_ERR(tgt));
575
576         root_fid = *gf->gf_u.gf_root_fid;
577         LASSERT(fid_is_sane(&root_fid));
578
579 repeat_fid2path:
580         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
581         if (rc != 0 && rc != -EREMOTE)
582                 GOTO(out_fid2path, rc);
583
584         /* If remote_gf != NULL, it means just building the
585          * path on the remote MDT, copy this path segement to gf */
586         if (remote_gf != NULL) {
587                 struct getinfo_fid2path *ori_gf;
588                 char *ptr;
589                 int len;
590
591                 ori_gf = (struct getinfo_fid2path *)karg;
592                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
593                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
594                         GOTO(out_fid2path, rc = -EOVERFLOW);
595
596                 ptr = ori_gf->gf_u.gf_path;
597
598                 len = strlen(gf->gf_u.gf_path);
599                 /* move the current path to the right to release space
600                  * for closer-to-root part */
601                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
602                 memcpy(ptr, gf->gf_u.gf_path, len);
603                 ptr[len] = '/';
604         }
605
606         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
607                tgt->ltd_exp->exp_obd->obd_name,
608                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
609                gf->gf_linkno);
610
611         if (rc == 0)
612                 GOTO(out_fid2path, rc);
613
614         /* sigh, has to go to another MDT to do path building further */
615         if (remote_gf == NULL) {
616                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
617                 OBD_ALLOC(remote_gf, remote_gf_size);
618                 if (remote_gf == NULL)
619                         GOTO(out_fid2path, rc = -ENOMEM);
620                 remote_gf->gf_pathlen = PATH_MAX;
621         }
622
623         if (!fid_is_sane(&gf->gf_fid)) {
624                 CERROR("%s: invalid FID "DFID": rc = %d\n",
625                        tgt->ltd_exp->exp_obd->obd_name,
626                        PFID(&gf->gf_fid), -EINVAL);
627                 GOTO(out_fid2path, rc = -EINVAL);
628         }
629
630         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
631         if (IS_ERR(tgt))
632                 GOTO(out_fid2path, rc = -EINVAL);
633
634         remote_gf->gf_fid = gf->gf_fid;
635         remote_gf->gf_recno = -1;
636         remote_gf->gf_linkno = -1;
637         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
638         *remote_gf->gf_u.gf_root_fid = root_fid;
639         gf = remote_gf;
640         goto repeat_fid2path;
641
642 out_fid2path:
643         if (remote_gf != NULL)
644                 OBD_FREE(remote_gf, remote_gf_size);
645         RETURN(rc);
646 }
647
648 static int lmv_hsm_req_count(struct lmv_obd *lmv,
649                              const struct hsm_user_request *hur,
650                              const struct lmv_tgt_desc *tgt_mds)
651 {
652         struct lmv_tgt_desc *curr_tgt;
653         __u32 i;
654         int nr = 0;
655
656         /* count how many requests must be sent to the given target */
657         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
658                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
659                 if (IS_ERR(curr_tgt))
660                         RETURN(PTR_ERR(curr_tgt));
661                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
662                         nr++;
663         }
664         return nr;
665 }
666
667 static int lmv_hsm_req_build(struct lmv_obd *lmv,
668                               struct hsm_user_request *hur_in,
669                               const struct lmv_tgt_desc *tgt_mds,
670                               struct hsm_user_request *hur_out)
671 {
672         __u32 i, nr_out;
673         struct lmv_tgt_desc *curr_tgt;
674
675         /* build the hsm_user_request for the given target */
676         hur_out->hur_request = hur_in->hur_request;
677         nr_out = 0;
678         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
679                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
680                 if (IS_ERR(curr_tgt))
681                         RETURN(PTR_ERR(curr_tgt));
682                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
683                         hur_out->hur_user_item[nr_out] =
684                                                 hur_in->hur_user_item[i];
685                         nr_out++;
686                 }
687         }
688         hur_out->hur_request.hr_itemcount = nr_out;
689         memcpy(hur_data(hur_out), hur_data(hur_in),
690                hur_in->hur_request.hr_data_len);
691
692         RETURN(0);
693 }
694
695 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
696                                  int len, struct lustre_kernelcomm *lk,
697                                  void __user *uarg)
698 {
699         struct lmv_obd *lmv = &obd->u.lmv;
700         struct lu_tgt_desc *tgt;
701         int rc;
702
703         ENTRY;
704
705         /* unregister request (call from llapi_hsm_copytool_fini) */
706         lmv_foreach_connected_tgt(lmv, tgt)
707                 /* best effort: try to clean as much as possible
708                  * (continue on error) */
709                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
710
711         /* Whatever the result, remove copytool from kuc groups.
712          * Unreached coordinators will get EPIPE on next requests
713          * and will unregister automatically.
714          */
715         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
716
717         RETURN(rc);
718 }
719
720 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
721                                int len, struct lustre_kernelcomm *lk,
722                                void __user *uarg)
723 {
724         struct lmv_obd *lmv = &obd->u.lmv;
725         struct file *filp;
726         bool any_set = false;
727         struct kkuc_ct_data *kcd;
728         size_t kcd_size;
729         struct lu_tgt_desc *tgt;
730         __u32 i;
731         int err;
732         int rc = 0;
733
734         ENTRY;
735
736         filp = fget(lk->lk_wfd);
737         if (!filp)
738                 RETURN(-EBADF);
739
740         if (lk->lk_flags & LK_FLG_DATANR)
741                 kcd_size = offsetof(struct kkuc_ct_data,
742                                     kcd_archives[lk->lk_data_count]);
743         else
744                 kcd_size = sizeof(*kcd);
745
746         OBD_ALLOC(kcd, kcd_size);
747         if (kcd == NULL)
748                 GOTO(err_fput, rc = -ENOMEM);
749
750         kcd->kcd_nr_archives = lk->lk_data_count;
751         if (lk->lk_flags & LK_FLG_DATANR) {
752                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
753                 if (lk->lk_data_count > 0)
754                         memcpy(kcd->kcd_archives, lk->lk_data,
755                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
756         } else {
757                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
758         }
759
760         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
761                                    lk->lk_group, kcd, kcd_size);
762         OBD_FREE(kcd, kcd_size);
763         if (rc)
764                 GOTO(err_fput, rc);
765
766         /* All or nothing: try to register to all MDS.
767          * In case of failure, unregister from previous MDS,
768          * except if it because of inactive target. */
769         lmv_foreach_connected_tgt(lmv, tgt) {
770                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
771                 if (err) {
772                         if (tgt->ltd_active) {
773                                 /* permanent error */
774                                 CERROR("%s: iocontrol MDC %s on MDT"
775                                        " idx %d cmd %x: err = %d\n",
776                                        lmv2obd_dev(lmv)->obd_name,
777                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
778                                        err);
779                                 rc = err;
780                                 lk->lk_flags |= LK_FLG_STOP;
781                                 i = tgt->ltd_index;
782                                 /* unregister from previous MDS */
783                                 lmv_foreach_connected_tgt(lmv, tgt) {
784                                         if (tgt->ltd_index >= i)
785                                                 break;
786
787                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
788                                                       lk, uarg);
789                                 }
790                                 GOTO(err_kkuc_rem, rc);
791                         }
792                         /* else: transient error.
793                          * kuc will register to the missing MDT
794                          * when it is back */
795                 } else {
796                         any_set = true;
797                 }
798         }
799
800         if (!any_set)
801                 /* no registration done: return error */
802                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
803
804         RETURN(0);
805
806 err_kkuc_rem:
807         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
808
809 err_fput:
810         fput(filp);
811         return rc;
812 }
813
814 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
815                          int len, void *karg, void __user *uarg)
816 {
817         struct obd_device *obddev = class_exp2obd(exp);
818         struct lmv_obd *lmv = &obddev->u.lmv;
819         struct lu_tgt_desc *tgt = NULL;
820         int set = 0;
821         __u32 count = lmv->lmv_mdt_count;
822         int rc = 0;
823
824         ENTRY;
825
826         if (count == 0)
827                 RETURN(-ENOTTY);
828
829         switch (cmd) {
830         case IOC_OBD_STATFS: {
831                 struct obd_ioctl_data *data = karg;
832                 struct obd_device *mdc_obd;
833                 struct obd_statfs stat_buf = {0};
834                 __u32 index;
835
836                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
837
838                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
839                         RETURN(-ENODEV);
840
841                 tgt = lmv_tgt(lmv, index);
842                 if (!tgt || !tgt->ltd_active)
843                         RETURN(-ENODATA);
844
845                 mdc_obd = class_exp2obd(tgt->ltd_exp);
846                 if (!mdc_obd)
847                         RETURN(-EINVAL);
848
849                 /* copy UUID */
850                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
851                                  min((int) data->ioc_plen2,
852                                      (int) sizeof(struct obd_uuid))))
853                         RETURN(-EFAULT);
854
855                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
856                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
857                                 0);
858                 if (rc)
859                         RETURN(rc);
860                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
861                                  min((int) data->ioc_plen1,
862                                      (int) sizeof(stat_buf))))
863                         RETURN(-EFAULT);
864                 break;
865         }
866         case OBD_IOC_QUOTACTL: {
867                 struct if_quotactl *qctl = karg;
868                 struct obd_quotactl *oqctl;
869
870                 if (qctl->qc_valid == QC_MDTIDX) {
871                         tgt = lmv_tgt(lmv, qctl->qc_idx);
872                 } else if (qctl->qc_valid == QC_UUID) {
873                         lmv_foreach_tgt(lmv, tgt) {
874                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
875                                                      &qctl->obd_uuid))
876                                         continue;
877
878                                 if (!tgt->ltd_exp)
879                                         RETURN(-EINVAL);
880
881                                 break;
882                         }
883                 } else {
884                         RETURN(-EINVAL);
885                 }
886
887                 if (!tgt || !tgt->ltd_exp)
888                         RETURN(-EINVAL);
889
890                 OBD_ALLOC_PTR(oqctl);
891                 if (!oqctl)
892                         RETURN(-ENOMEM);
893
894                 QCTL_COPY(oqctl, qctl);
895                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
896                 if (rc == 0) {
897                         QCTL_COPY(qctl, oqctl);
898                         qctl->qc_valid = QC_MDTIDX;
899                         qctl->obd_uuid = tgt->ltd_uuid;
900                 }
901                 OBD_FREE_PTR(oqctl);
902                 break;
903         }
904         case LL_IOC_GET_CONNECT_FLAGS: {
905                 tgt = lmv_tgt(lmv, 0);
906                 rc = -ENODATA;
907                 if (tgt && tgt->ltd_exp)
908                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
909                 break;
910         }
911         case LL_IOC_FID2MDTIDX: {
912                 struct lu_fid *fid = karg;
913                 int             mdt_index;
914
915                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
916                 if (rc != 0)
917                         RETURN(rc);
918
919                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
920                  * point to user space memory for FID2MDTIDX. */
921                 *(__u32 *)uarg = mdt_index;
922                 break;
923         }
924         case OBD_IOC_FID2PATH: {
925                 rc = lmv_fid2path(exp, len, karg, uarg);
926                 break;
927         }
928         case LL_IOC_HSM_STATE_GET:
929         case LL_IOC_HSM_STATE_SET:
930         case LL_IOC_HSM_ACTION: {
931                 struct md_op_data *op_data = karg;
932
933                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
934                 if (IS_ERR(tgt))
935                         RETURN(PTR_ERR(tgt));
936
937                 if (tgt->ltd_exp == NULL)
938                         RETURN(-EINVAL);
939
940                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
941                 break;
942         }
943         case LL_IOC_HSM_PROGRESS: {
944                 const struct hsm_progress_kernel *hpk = karg;
945
946                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
947                 if (IS_ERR(tgt))
948                         RETURN(PTR_ERR(tgt));
949                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
950                 break;
951         }
952         case LL_IOC_HSM_REQUEST: {
953                 struct hsm_user_request *hur = karg;
954                 unsigned int reqcount = hur->hur_request.hr_itemcount;
955
956                 if (reqcount == 0)
957                         RETURN(0);
958
959                 /* if the request is about a single fid
960                  * or if there is a single MDS, no need to split
961                  * the request. */
962                 if (reqcount == 1 || count == 1) {
963                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
964                         if (IS_ERR(tgt))
965                                 RETURN(PTR_ERR(tgt));
966                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
967                 } else {
968                         /* split fid list to their respective MDS */
969                         lmv_foreach_connected_tgt(lmv, tgt) {
970                                 int nr, rc1;
971                                 size_t reqlen;
972                                 struct hsm_user_request *req;
973
974                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
975                                 if (nr < 0)
976                                         RETURN(nr);
977                                 if (nr == 0) /* nothing for this MDS */
978                                         continue;
979
980                                 /* build a request with fids for this MDS */
981                                 reqlen = offsetof(typeof(*hur),
982                                                   hur_user_item[nr])
983                                                 + hur->hur_request.hr_data_len;
984                                 OBD_ALLOC_LARGE(req, reqlen);
985                                 if (req == NULL)
986                                         RETURN(-ENOMEM);
987                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
988                                 if (rc1 < 0)
989                                         GOTO(hsm_req_err, rc1);
990                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
991                                                     req, uarg);
992 hsm_req_err:
993                                 if (rc1 != 0 && rc == 0)
994                                         rc = rc1;
995                                 OBD_FREE_LARGE(req, reqlen);
996                         }
997                 }
998                 break;
999         }
1000         case LL_IOC_LOV_SWAP_LAYOUTS: {
1001                 struct md_op_data *op_data = karg;
1002                 struct lmv_tgt_desc *tgt1, *tgt2;
1003
1004                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1005                 if (IS_ERR(tgt1))
1006                         RETURN(PTR_ERR(tgt1));
1007
1008                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1009                 if (IS_ERR(tgt2))
1010                         RETURN(PTR_ERR(tgt2));
1011
1012                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1013                         RETURN(-EINVAL);
1014
1015                 /* only files on same MDT can have their layouts swapped */
1016                 if (tgt1->ltd_index != tgt2->ltd_index)
1017                         RETURN(-EPERM);
1018
1019                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1020                 break;
1021         }
1022         case LL_IOC_HSM_CT_START: {
1023                 struct lustre_kernelcomm *lk = karg;
1024                 if (lk->lk_flags & LK_FLG_STOP)
1025                         rc = lmv_hsm_ct_unregister(obddev, cmd, len, lk, uarg);
1026                 else
1027                         rc = lmv_hsm_ct_register(obddev, cmd, len, lk, uarg);
1028                 break;
1029         }
1030         default:
1031                 lmv_foreach_connected_tgt(lmv, tgt) {
1032                         struct obd_device *mdc_obd;
1033                         int err;
1034
1035                         /* ll_umount_begin() sets force flag but for lmv, not
1036                          * mdc. Let's pass it through */
1037                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1038                         mdc_obd->obd_force = obddev->obd_force;
1039                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1040                         if (err) {
1041                                 if (tgt->ltd_active) {
1042                                         CERROR("error: iocontrol MDC %s on MDT"
1043                                                " idx %d cmd %x: err = %d\n",
1044                                                tgt->ltd_uuid.uuid,
1045                                                tgt->ltd_index, cmd, err);
1046                                         if (!rc)
1047                                                 rc = err;
1048                                 }
1049                         } else
1050                                 set = 1;
1051                 }
1052                 if (!set && !rc)
1053                         rc = -EIO;
1054         }
1055         RETURN(rc);
1056 }
1057
1058 /**
1059  * This is _inode_ placement policy function (not name).
1060  */
1061 static u32 lmv_placement_policy(struct obd_device *obd,
1062                                 struct md_op_data *op_data)
1063 {
1064         struct lmv_obd *lmv = &obd->u.lmv;
1065         struct lmv_user_md *lum;
1066         u32 mdt;
1067
1068         ENTRY;
1069
1070         if (lmv->lmv_mdt_count == 1)
1071                 RETURN(0);
1072
1073         lum = op_data->op_data;
1074         /*
1075          * Choose MDT by
1076          * 1. See if the stripe offset is specified by lum.
1077          * 2. If parent has default LMV, and its hash type is "space", choose
1078          *    MDT with QoS. (see lmv_locate_tgt_qos()).
1079          * 3. Then check if default LMV stripe offset is not -1.
1080          * 4. Finally choose MDS by name hash if the parent
1081          *    is striped directory. (see lmv_locate_tgt()).
1082          *
1083          * presently explicit MDT location is not supported
1084          * for foreign dirs (as it can't be embedded into free
1085          * format LMV, like with lum_stripe_offset), so we only
1086          * rely on default stripe offset or then name hashing.
1087          */
1088         if (op_data->op_cli_flags & CLI_SET_MEA && lum != NULL &&
1089             le32_to_cpu(lum->lum_magic != LMV_MAGIC_FOREIGN) &&
1090             le32_to_cpu(lum->lum_stripe_offset) != (__u32)-1) {
1091                 mdt = le32_to_cpu(lum->lum_stripe_offset);
1092         } else if (op_data->op_code == LUSTRE_OPC_MKDIR &&
1093                    !lmv_dir_striped(op_data->op_mea1) &&
1094                    lmv_dir_qos_mkdir(op_data->op_default_mea1)) {
1095                 mdt = op_data->op_mds;
1096         } else if (op_data->op_code == LUSTRE_OPC_MKDIR &&
1097                    op_data->op_default_mea1 &&
1098                    op_data->op_default_mea1->lsm_md_master_mdt_index !=
1099                         (__u32)-1) {
1100                 mdt = op_data->op_default_mea1->lsm_md_master_mdt_index;
1101                 op_data->op_mds = mdt;
1102         } else {
1103                 mdt = op_data->op_mds;
1104         }
1105
1106         RETURN(mdt);
1107 }
1108
1109 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid, u32 mds)
1110 {
1111         struct lmv_tgt_desc *tgt;
1112         int rc;
1113
1114         ENTRY;
1115
1116         tgt = lmv_tgt(lmv, mds);
1117         if (!tgt)
1118                 RETURN(-ENODEV);
1119
1120         /*
1121          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1122          * on server that seq in new allocated fid is not yet known.
1123          */
1124         mutex_lock(&tgt->ltd_fid_mutex);
1125
1126         if (tgt->ltd_active == 0 || tgt->ltd_exp == NULL)
1127                 GOTO(out, rc = -ENODEV);
1128
1129         /*
1130          * Asking underlying tgt layer to allocate new fid.
1131          */
1132         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1133         if (rc > 0) {
1134                 LASSERT(fid_is_sane(fid));
1135                 rc = 0;
1136         }
1137
1138         EXIT;
1139 out:
1140         mutex_unlock(&tgt->ltd_fid_mutex);
1141         return rc;
1142 }
1143
1144 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1145                   struct lu_fid *fid, struct md_op_data *op_data)
1146 {
1147         struct obd_device *obd = class_exp2obd(exp);
1148         struct lmv_obd *lmv = &obd->u.lmv;
1149         u32 mds;
1150         int rc;
1151
1152         ENTRY;
1153
1154         LASSERT(op_data != NULL);
1155         LASSERT(fid != NULL);
1156
1157         mds = lmv_placement_policy(obd, op_data);
1158
1159         rc = __lmv_fid_alloc(lmv, fid, mds);
1160         if (rc)
1161                 CERROR("Can't alloc new fid, rc %d\n", rc);
1162
1163         RETURN(rc);
1164 }
1165
1166 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1167 {
1168         struct lmv_obd *lmv = &obd->u.lmv;
1169         struct lmv_desc *desc;
1170         struct lnet_process_id lnet_id;
1171         int i = 0;
1172         int rc;
1173
1174         ENTRY;
1175
1176         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1177                 CERROR("LMV setup requires a descriptor\n");
1178                 RETURN(-EINVAL);
1179         }
1180
1181         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1182         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1183                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1184                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1185                 RETURN(-EINVAL);
1186         }
1187
1188         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1189                      desc->ld_uuid.uuid);
1190         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1191         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1192         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1193                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1194         lmv->max_def_easize = 0;
1195         lmv->max_easize = 0;
1196
1197         spin_lock_init(&lmv->lmv_lock);
1198
1199         /*
1200          * initialize rr_index to lower 32bit of netid, so that client
1201          * can distribute subdirs evenly from the beginning.
1202          */
1203         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1204                 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1205                         lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1206                         break;
1207                 }
1208         }
1209
1210         rc = lmv_tunables_init(obd);
1211         if (rc)
1212                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1213                       obd->obd_name, rc);
1214
1215         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1216                              LUSTRE_CLI_FLD_HASH_DHT);
1217         if (rc)
1218                 CERROR("Can't init FLD, err %d\n", rc);
1219
1220         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1221         if (rc)
1222                 CWARN("%s: error initialize target table: rc = %d\n",
1223                       obd->obd_name, rc);
1224
1225         RETURN(rc);
1226 }
1227
1228 static int lmv_cleanup(struct obd_device *obd)
1229 {
1230         struct lmv_obd *lmv = &obd->u.lmv;
1231         struct lu_tgt_desc *tgt;
1232         struct lu_tgt_desc *tmp;
1233
1234         ENTRY;
1235
1236         fld_client_fini(&lmv->lmv_fld);
1237         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1238                 lmv_del_target(lmv, tgt);
1239         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1240
1241         RETURN(0);
1242 }
1243
1244 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1245 {
1246         struct lustre_cfg       *lcfg = buf;
1247         struct obd_uuid         obd_uuid;
1248         int                     gen;
1249         __u32                   index;
1250         int                     rc;
1251         ENTRY;
1252
1253         switch (lcfg->lcfg_command) {
1254         case LCFG_ADD_MDC:
1255                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1256                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1257                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1258                         GOTO(out, rc = -EINVAL);
1259
1260                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1261
1262                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1263                         GOTO(out, rc = -EINVAL);
1264                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1265                         GOTO(out, rc = -EINVAL);
1266                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1267                 GOTO(out, rc);
1268         default:
1269                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1270                 GOTO(out, rc = -EINVAL);
1271         }
1272 out:
1273         RETURN(rc);
1274 }
1275
1276 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1277 {
1278         int i;
1279
1280         if (flags & OBD_STATFS_FOR_MDT0)
1281                 return 0;
1282
1283         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1284                 return lmv->lmv_statfs_start;
1285
1286         /* choose initial MDT for this client */
1287         for (i = 0;; i++) {
1288                 struct lnet_process_id lnet_id;
1289                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1290                         break;
1291
1292                 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1293                         /* We dont need a full 64-bit modulus, just enough
1294                          * to distribute the requests across MDTs evenly.
1295                          */
1296                         lmv->lmv_statfs_start = (u32)lnet_id.nid %
1297                                                 lmv->lmv_mdt_count;
1298                         break;
1299                 }
1300         }
1301
1302         return lmv->lmv_statfs_start;
1303 }
1304
1305 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1306                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1307 {
1308         struct obd_device *obd = class_exp2obd(exp);
1309         struct lmv_obd *lmv = &obd->u.lmv;
1310         struct obd_statfs *temp;
1311         struct lu_tgt_desc *tgt;
1312         __u32 i;
1313         __u32 idx;
1314         int rc = 0;
1315
1316         ENTRY;
1317
1318         OBD_ALLOC(temp, sizeof(*temp));
1319         if (temp == NULL)
1320                 RETURN(-ENOMEM);
1321
1322         /* distribute statfs among MDTs */
1323         idx = lmv_select_statfs_mdt(lmv, flags);
1324
1325         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1326                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1327                 tgt = lmv_tgt(lmv, idx);
1328                 if (!tgt || !tgt->ltd_exp)
1329                         continue;
1330
1331                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age, flags);
1332                 if (rc) {
1333                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1334                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1335                         GOTO(out_free_temp, rc);
1336                 }
1337
1338                 if (temp->os_state & OS_STATE_SUM ||
1339                     flags == OBD_STATFS_FOR_MDT0) {
1340                         /* reset to the last aggregated values
1341                          * and don't sum with non-aggrated data */
1342                         /* If the statfs is from mount, it needs to retrieve
1343                          * necessary information from MDT0. i.e. mount does
1344                          * not need the merged osfs from all of MDT. Also
1345                          * clients can be mounted as long as MDT0 is in
1346                          * service */
1347                         *osfs = *temp;
1348                         break;
1349                 }
1350
1351                 if (i == 0) {
1352                         *osfs = *temp;
1353                 } else {
1354                         osfs->os_bavail += temp->os_bavail;
1355                         osfs->os_blocks += temp->os_blocks;
1356                         osfs->os_ffree += temp->os_ffree;
1357                         osfs->os_files += temp->os_files;
1358                         osfs->os_granted += temp->os_granted;
1359                 }
1360         }
1361
1362         EXIT;
1363 out_free_temp:
1364         OBD_FREE(temp, sizeof(*temp));
1365         return rc;
1366 }
1367
1368 static int lmv_statfs_update(void *cookie, int rc)
1369 {
1370         struct obd_info *oinfo = cookie;
1371         struct obd_device *obd = oinfo->oi_obd;
1372         struct lmv_obd *lmv = &obd->u.lmv;
1373         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1374         struct obd_statfs *osfs = oinfo->oi_osfs;
1375
1376         /*
1377          * NB: don't deactivate TGT upon error, because we may not trigger async
1378          * statfs any longer, then there is no chance to activate TGT.
1379          */
1380         if (!rc) {
1381                 spin_lock(&lmv->lmv_lock);
1382                 tgt->ltd_statfs = *osfs;
1383                 tgt->ltd_statfs_age = ktime_get_seconds();
1384                 spin_unlock(&lmv->lmv_lock);
1385                 lmv->lmv_qos.lq_dirty = 1;
1386         }
1387
1388         return rc;
1389 }
1390
1391 /* update tgt statfs async if it's ld_qos_maxage old */
1392 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1393 {
1394         struct obd_info oinfo = {
1395                 .oi_obd = obd,
1396                 .oi_tgt = tgt,
1397                 .oi_cb_up = lmv_statfs_update,
1398         };
1399         int rc;
1400
1401         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1402             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1403                 return 0;
1404
1405         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1406
1407         return rc;
1408 }
1409
1410 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1411                         struct lu_fid *fid)
1412 {
1413         struct obd_device *obd = exp->exp_obd;
1414         struct lmv_obd *lmv = &obd->u.lmv;
1415         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1416         int rc;
1417
1418         ENTRY;
1419
1420         if (!tgt)
1421                 RETURN(-ENODEV);
1422
1423         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1424         RETURN(rc);
1425 }
1426
1427 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1428                         u64 obd_md_valid, const char *name, size_t buf_size,
1429                         struct ptlrpc_request **req)
1430 {
1431         struct obd_device *obd = exp->exp_obd;
1432         struct lmv_obd *lmv = &obd->u.lmv;
1433         struct lmv_tgt_desc *tgt;
1434         int rc;
1435
1436         ENTRY;
1437
1438         tgt = lmv_fid2tgt(lmv, fid);
1439         if (IS_ERR(tgt))
1440                 RETURN(PTR_ERR(tgt));
1441
1442         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1443
1444         RETURN(rc);
1445 }
1446
1447 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1448                         u64 obd_md_valid, const char *name,
1449                         const void *value, size_t value_size,
1450                         unsigned int xattr_flags, u32 suppgid,
1451                         struct ptlrpc_request **req)
1452 {
1453         struct obd_device *obd = exp->exp_obd;
1454         struct lmv_obd *lmv = &obd->u.lmv;
1455         struct lmv_tgt_desc *tgt;
1456         int rc;
1457
1458         ENTRY;
1459
1460         tgt = lmv_fid2tgt(lmv, fid);
1461         if (IS_ERR(tgt))
1462                 RETURN(PTR_ERR(tgt));
1463
1464         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1465                          value, value_size, xattr_flags, suppgid, req);
1466
1467         RETURN(rc);
1468 }
1469
1470 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1471                        struct ptlrpc_request **request)
1472 {
1473         struct obd_device *obd = exp->exp_obd;
1474         struct lmv_obd *lmv = &obd->u.lmv;
1475         struct lmv_tgt_desc *tgt;
1476         int rc;
1477
1478         ENTRY;
1479
1480         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1481         if (IS_ERR(tgt))
1482                 RETURN(PTR_ERR(tgt));
1483
1484         if (op_data->op_flags & MF_GET_MDT_IDX) {
1485                 op_data->op_mds = tgt->ltd_index;
1486                 RETURN(0);
1487         }
1488
1489         rc = md_getattr(tgt->ltd_exp, op_data, request);
1490
1491         RETURN(rc);
1492 }
1493
1494 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1495 {
1496         struct obd_device *obd = exp->exp_obd;
1497         struct lmv_obd *lmv = &obd->u.lmv;
1498         struct lu_tgt_desc *tgt;
1499
1500         ENTRY;
1501
1502         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1503
1504         /*
1505          * With DNE every object can have two locks in different namespaces:
1506          * lookup lock in space of MDT storing direntry and update/open lock in
1507          * space of MDT storing inode.
1508          */
1509         lmv_foreach_connected_tgt(lmv, tgt)
1510                 md_null_inode(tgt->ltd_exp, fid);
1511
1512         RETURN(0);
1513 }
1514
1515 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1516                      struct md_open_data *mod, struct ptlrpc_request **request)
1517 {
1518         struct obd_device *obd = exp->exp_obd;
1519         struct lmv_obd *lmv = &obd->u.lmv;
1520         struct lmv_tgt_desc *tgt;
1521         int rc;
1522
1523         ENTRY;
1524
1525         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1526         if (IS_ERR(tgt))
1527                 RETURN(PTR_ERR(tgt));
1528
1529         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1530         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1531         RETURN(rc);
1532 }
1533
1534 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
1535 {
1536         struct lu_tgt_desc *tgt;
1537         __u64 total_weight = 0;
1538         __u64 cur_weight = 0;
1539         __u64 rand;
1540         int rc;
1541
1542         ENTRY;
1543
1544         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1545                 RETURN(ERR_PTR(-EAGAIN));
1546
1547         down_write(&lmv->lmv_qos.lq_rw_sem);
1548
1549         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1550                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1551
1552         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1553         if (rc)
1554                 GOTO(unlock, tgt = ERR_PTR(rc));
1555
1556         lmv_foreach_tgt(lmv, tgt) {
1557                 tgt->ltd_qos.ltq_usable = 0;
1558                 if (!tgt->ltd_exp || !tgt->ltd_active)
1559                         continue;
1560
1561                 tgt->ltd_qos.ltq_usable = 1;
1562                 lu_tgt_qos_weight_calc(tgt);
1563                 total_weight += tgt->ltd_qos.ltq_weight;
1564         }
1565
1566         rand = lu_prandom_u64_max(total_weight);
1567
1568         lmv_foreach_connected_tgt(lmv, tgt) {
1569                 if (!tgt->ltd_qos.ltq_usable)
1570                         continue;
1571
1572                 cur_weight += tgt->ltd_qos.ltq_weight;
1573                 if (cur_weight < rand)
1574                         continue;
1575
1576                 *mdt = tgt->ltd_index;
1577                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1578                 GOTO(unlock, rc = 0);
1579         }
1580
1581         /* no proper target found */
1582         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1583 unlock:
1584         up_write(&lmv->lmv_qos.lq_rw_sem);
1585
1586         return tgt;
1587 }
1588
1589 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
1590 {
1591         struct lu_tgt_desc *tgt;
1592         int i;
1593         int index;
1594
1595         ENTRY;
1596
1597         spin_lock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1598         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1599                 index = (i + lmv->lmv_qos_rr_index) %
1600                         lmv->lmv_mdt_descs.ltd_tgts_size;
1601                 tgt = lmv_tgt(lmv, index);
1602                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1603                         continue;
1604
1605                 *mdt = tgt->ltd_index;
1606                 lmv->lmv_qos_rr_index = (*mdt + 1) %
1607                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1608                 spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1609
1610                 RETURN(tgt);
1611         }
1612         spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1613
1614         RETURN(ERR_PTR(-ENODEV));
1615 }
1616
1617 static struct lmv_tgt_desc *
1618 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1619                        const char *name, int namelen, struct lu_fid *fid,
1620                        __u32 *mds, bool post_migrate)
1621 {
1622         struct lmv_tgt_desc *tgt;
1623         const struct lmv_oinfo *oinfo;
1624
1625         if (!lmv_dir_striped(lsm) || !namelen) {
1626                 tgt = lmv_fid2tgt(lmv, fid);
1627                 if (IS_ERR(tgt))
1628                         return tgt;
1629
1630                 *mds = tgt->ltd_index;
1631                 return tgt;
1632         }
1633
1634         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1635                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1636                         return ERR_PTR(-EBADF);
1637                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1638         } else {
1639                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
1640                                                 post_migrate);
1641                 if (IS_ERR(oinfo))
1642                         return ERR_CAST(oinfo);
1643         }
1644
1645         *fid = oinfo->lmo_fid;
1646         *mds = oinfo->lmo_mds;
1647         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1648
1649         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1650
1651         return tgt ? tgt : ERR_PTR(-ENODEV);
1652 }
1653
1654 /**
1655  * Locate MDT of op_data->op_fid1
1656  *
1657  * For striped directory, it will locate the stripe by name hash, if hash_type
1658  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1659  * which is set outside, and if dir is migrating, 'op_data->op_post_migrate'
1660  * indicates whether old or new layout is used to locate.
1661  *
1662  * For plain direcotry, normally it will locate MDT by FID, but if this
1663  * directory has default LMV, and its hash type is "space", locate MDT with QoS.
1664  *
1665  * \param[in] lmv       LMV device
1666  * \param[in] op_data   client MD stack parameters, name, namelen
1667  *                      mds_num etc.
1668  *
1669  * retval               pointer to the lmv_tgt_desc if succeed.
1670  *                      ERR_PTR(errno) if failed.
1671  */
1672 struct lmv_tgt_desc *
1673 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1674 {
1675         struct lmv_stripe_md *lsm = op_data->op_mea1;
1676         struct lmv_oinfo *oinfo;
1677         struct lmv_tgt_desc *tgt;
1678
1679         if (lmv_dir_foreign(lsm))
1680                 return ERR_PTR(-ENODATA);
1681
1682         /* During creating VOLATILE file, it should honor the mdt
1683          * index if the file under striped dir is being restored, see
1684          * ct_restore(). */
1685         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1686             (int)op_data->op_mds != -1) {
1687                 tgt = lmv_tgt(lmv, op_data->op_mds);
1688                 if (!tgt)
1689                         return ERR_PTR(-ENODEV);
1690
1691                 if (lmv_dir_striped(lsm)) {
1692                         int i;
1693
1694                         /* refill the right parent fid */
1695                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1696                                 oinfo = &lsm->lsm_md_oinfo[i];
1697                                 if (oinfo->lmo_mds == op_data->op_mds) {
1698                                         op_data->op_fid1 = oinfo->lmo_fid;
1699                                         break;
1700                                 }
1701                         }
1702
1703                         if (i == lsm->lsm_md_stripe_count)
1704                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1705                 }
1706         } else if (lmv_dir_bad_hash(lsm)) {
1707                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1708                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1709
1710                 op_data->op_fid1 = oinfo->lmo_fid;
1711                 op_data->op_mds = oinfo->lmo_mds;
1712                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1713                 if (!tgt)
1714                         tgt = ERR_PTR(-ENODEV);
1715         } else if (op_data->op_code == LUSTRE_OPC_MKDIR &&
1716                    lmv_dir_qos_mkdir(op_data->op_default_mea1) &&
1717                    !lmv_dir_striped(lsm)) {
1718                 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds);
1719                 if (tgt == ERR_PTR(-EAGAIN))
1720                         tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1721                 /*
1722                  * only update statfs when mkdir under dir with "space" hash,
1723                  * this means the cached statfs may be stale, and current mkdir
1724                  * may not follow QoS accurately, but it's not serious, and it
1725                  * avoids periodic statfs when client doesn't mkdir under
1726                  * "space" hashed directories.
1727                  *
1728                  * TODO: after MDT support QoS object allocation, also update
1729                  * statfs for 'lfs mkdir -i -1 ...", currently it's done in user
1730                  * space.
1731                  */
1732                 if (!IS_ERR(tgt)) {
1733                         struct obd_device *obd;
1734
1735                         obd = container_of(lmv, struct obd_device, u.lmv);
1736                         lmv_statfs_check_update(obd, tgt);
1737                 }
1738         } else {
1739                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1740                                 op_data->op_name, op_data->op_namelen,
1741                                 &op_data->op_fid1, &op_data->op_mds,
1742                                 op_data->op_post_migrate);
1743         }
1744
1745         return tgt;
1746 }
1747
1748 /* Locate MDT of op_data->op_fid2 for link/rename */
1749 static struct lmv_tgt_desc *
1750 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1751 {
1752         struct lmv_tgt_desc *tgt;
1753         int rc;
1754
1755         LASSERT(op_data->op_name);
1756         if (lmv_dir_migrating(op_data->op_mea2)) {
1757                 struct lu_fid fid1 = op_data->op_fid1;
1758                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1759                 struct ptlrpc_request *request = NULL;
1760
1761                 /*
1762                  * avoid creating new file under old layout of migrating
1763                  * directory, check it here.
1764                  */
1765                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1766                                 op_data->op_name, op_data->op_namelen,
1767                                 &op_data->op_fid2, &op_data->op_mds, false);
1768                 if (IS_ERR(tgt))
1769                         RETURN(tgt);
1770
1771                 op_data->op_fid1 = op_data->op_fid2;
1772                 op_data->op_mea1 = op_data->op_mea2;
1773                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1774                 op_data->op_fid1 = fid1;
1775                 op_data->op_mea1 = lsm1;
1776                 if (!rc) {
1777                         ptlrpc_req_finished(request);
1778                         RETURN(ERR_PTR(-EEXIST));
1779                 }
1780
1781                 if (rc != -ENOENT)
1782                         RETURN(ERR_PTR(rc));
1783         }
1784
1785         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1786                                 op_data->op_namelen, &op_data->op_fid2,
1787                                 &op_data->op_mds, true);
1788 }
1789
1790 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1791                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1792                 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1793                 struct ptlrpc_request **request)
1794 {
1795         struct obd_device *obd = exp->exp_obd;
1796         struct lmv_obd *lmv = &obd->u.lmv;
1797         struct lmv_tgt_desc *tgt;
1798         int rc;
1799
1800         ENTRY;
1801
1802         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1803                 RETURN(-EIO);
1804
1805         if (lmv_dir_bad_hash(op_data->op_mea1))
1806                 RETURN(-EBADF);
1807
1808         if (lmv_dir_migrating(op_data->op_mea1)) {
1809                 /*
1810                  * if parent is migrating, create() needs to lookup existing
1811                  * name, to avoid creating new file under old layout of
1812                  * migrating directory, check old layout here.
1813                  */
1814                 tgt = lmv_locate_tgt(lmv, op_data);
1815                 if (IS_ERR(tgt))
1816                         RETURN(PTR_ERR(tgt));
1817
1818                 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1819                 if (!rc) {
1820                         ptlrpc_req_finished(*request);
1821                         *request = NULL;
1822                         RETURN(-EEXIST);
1823                 }
1824
1825                 if (rc != -ENOENT)
1826                         RETURN(rc);
1827
1828                 op_data->op_post_migrate = true;
1829         }
1830
1831         tgt = lmv_locate_tgt(lmv, op_data);
1832         if (IS_ERR(tgt))
1833                 RETURN(PTR_ERR(tgt));
1834
1835         CDEBUG(D_INODE, "CREATE name '%.*s' on "DFID" -> mds #%x\n",
1836                 (int)op_data->op_namelen, op_data->op_name,
1837                 PFID(&op_data->op_fid1), op_data->op_mds);
1838
1839         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1840         if (rc)
1841                 RETURN(rc);
1842
1843         if (exp_connect_flags(exp) & OBD_CONNECT_DIR_STRIPE) {
1844                 /* Send the create request to the MDT where the object
1845                  * will be located */
1846                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
1847                 if (IS_ERR(tgt))
1848                         RETURN(PTR_ERR(tgt));
1849
1850                 op_data->op_mds = tgt->ltd_index;
1851         }
1852
1853         CDEBUG(D_INODE, "CREATE obj "DFID" -> mds #%x\n",
1854                PFID(&op_data->op_fid2), op_data->op_mds);
1855
1856         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1857         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1858                        cap_effective, rdev, request);
1859         if (rc == 0) {
1860                 if (*request == NULL)
1861                         RETURN(rc);
1862                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1863         }
1864         RETURN(rc);
1865 }
1866
1867 static int
1868 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1869             const union ldlm_policy_data *policy, struct md_op_data *op_data,
1870             struct lustre_handle *lockh, __u64 extra_lock_flags)
1871 {
1872         struct obd_device *obd = exp->exp_obd;
1873         struct lmv_obd *lmv = &obd->u.lmv;
1874         struct lmv_tgt_desc *tgt;
1875         int rc;
1876
1877         ENTRY;
1878
1879         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1880
1881         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1882         if (IS_ERR(tgt))
1883                 RETURN(PTR_ERR(tgt));
1884
1885         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1886                PFID(&op_data->op_fid1), tgt->ltd_index);
1887
1888         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1889                         extra_lock_flags);
1890
1891         RETURN(rc);
1892 }
1893
1894 int
1895 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1896                  struct ptlrpc_request **preq)
1897 {
1898         struct obd_device *obd = exp->exp_obd;
1899         struct lmv_obd *lmv = &obd->u.lmv;
1900         struct lmv_tgt_desc *tgt;
1901         struct mdt_body *body;
1902         int rc;
1903
1904         ENTRY;
1905
1906 retry:
1907         tgt = lmv_locate_tgt(lmv, op_data);
1908         if (IS_ERR(tgt))
1909                 RETURN(PTR_ERR(tgt));
1910
1911         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1912                 (int)op_data->op_namelen, op_data->op_name,
1913                 PFID(&op_data->op_fid1), tgt->ltd_index);
1914
1915         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1916         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1917                 ptlrpc_req_finished(*preq);
1918                 *preq = NULL;
1919                 goto retry;
1920         }
1921
1922         if (rc)
1923                 RETURN(rc);
1924
1925         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1926         LASSERT(body != NULL);
1927
1928         if (body->mbo_valid & OBD_MD_MDS) {
1929                 op_data->op_fid1 = body->mbo_fid1;
1930                 op_data->op_valid |= OBD_MD_FLCROSSREF;
1931                 op_data->op_namelen = 0;
1932                 op_data->op_name = NULL;
1933
1934                 ptlrpc_req_finished(*preq);
1935                 *preq = NULL;
1936
1937                 goto retry;
1938         }
1939
1940         RETURN(rc);
1941 }
1942
1943 #define md_op_data_fid(op_data, fl)                     \
1944         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1945          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1946          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1947          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1948          NULL)
1949
1950 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1951                             struct md_op_data *op_data, __u32 op_tgt,
1952                             enum ldlm_mode mode, int bits, int flag)
1953 {
1954         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1955         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1956         union ldlm_policy_data policy = { { 0 } };
1957         int rc = 0;
1958         ENTRY;
1959
1960         if (!fid_is_sane(fid))
1961                 RETURN(0);
1962
1963         if (tgt == NULL) {
1964                 tgt = lmv_fid2tgt(lmv, fid);
1965                 if (IS_ERR(tgt))
1966                         RETURN(PTR_ERR(tgt));
1967         }
1968
1969         if (tgt->ltd_index != op_tgt) {
1970                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1971                 policy.l_inodebits.bits = bits;
1972                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1973                                       mode, LCF_ASYNC, NULL);
1974         } else {
1975                 CDEBUG(D_INODE,
1976                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
1977                        op_tgt, PFID(fid));
1978                 op_data->op_flags |= flag;
1979                 rc = 0;
1980         }
1981
1982         RETURN(rc);
1983 }
1984
1985 /*
1986  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1987  * op_data->op_fid2
1988  */
1989 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1990                     struct ptlrpc_request **request)
1991 {
1992         struct obd_device       *obd = exp->exp_obd;
1993         struct lmv_obd          *lmv = &obd->u.lmv;
1994         struct lmv_tgt_desc     *tgt;
1995         int                      rc;
1996         ENTRY;
1997
1998         LASSERT(op_data->op_namelen != 0);
1999
2000         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2001                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2002                op_data->op_name, PFID(&op_data->op_fid1));
2003
2004         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2005         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2006         op_data->op_cap = cfs_curproc_cap_pack();
2007
2008         tgt = lmv_locate_tgt2(lmv, op_data);
2009         if (IS_ERR(tgt))
2010                 RETURN(PTR_ERR(tgt));
2011
2012         /*
2013          * Cancel UPDATE lock on child (fid1).
2014          */
2015         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2016         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2017                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2018         if (rc != 0)
2019                 RETURN(rc);
2020
2021         rc = md_link(tgt->ltd_exp, op_data, request);
2022
2023         RETURN(rc);
2024 }
2025
2026 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2027                         const char *name, size_t namelen,
2028                         struct ptlrpc_request **request)
2029 {
2030         struct obd_device *obd = exp->exp_obd;
2031         struct lmv_obd *lmv = &obd->u.lmv;
2032         struct lmv_stripe_md *lsm = op_data->op_mea1;
2033         struct lmv_tgt_desc *parent_tgt;
2034         struct lmv_tgt_desc *sp_tgt;
2035         struct lmv_tgt_desc *tp_tgt = NULL;
2036         struct lmv_tgt_desc *child_tgt;
2037         struct lmv_tgt_desc *tgt;
2038         struct lu_fid target_fid;
2039         int rc;
2040
2041         ENTRY;
2042
2043         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2044
2045         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2046                PFID(&op_data->op_fid1), (int)namelen, name);
2047
2048         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2049         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2050         op_data->op_cap = cfs_curproc_cap_pack();
2051
2052         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2053         if (IS_ERR(parent_tgt))
2054                 RETURN(PTR_ERR(parent_tgt));
2055
2056         if (lmv_dir_striped(lsm)) {
2057                 __u32 hash_type = lsm->lsm_md_hash_type;
2058                 __u32 stripe_count = lsm->lsm_md_stripe_count;
2059
2060                 /*
2061                  * old stripes are appended after new stripes for migrating
2062                  * directory.
2063                  */
2064                 if (lmv_dir_migrating(lsm)) {
2065                         hash_type = lsm->lsm_md_migrate_hash;
2066                         stripe_count -= lsm->lsm_md_migrate_offset;
2067                 }
2068
2069                 rc = lmv_name_to_stripe_index(hash_type, stripe_count, name,
2070                                               namelen);
2071                 if (rc < 0)
2072                         RETURN(rc);
2073
2074                 if (lmv_dir_migrating(lsm))
2075                         rc += lsm->lsm_md_migrate_offset;
2076
2077                 /* save it in fid4 temporarily for early cancel */
2078                 op_data->op_fid4 = lsm->lsm_md_oinfo[rc].lmo_fid;
2079                 sp_tgt = lmv_tgt(lmv, lsm->lsm_md_oinfo[rc].lmo_mds);
2080                 if (!sp_tgt)
2081                         RETURN(-ENODEV);
2082
2083                 /*
2084                  * if parent is being migrated too, fill op_fid2 with target
2085                  * stripe fid, otherwise the target stripe is not created yet.
2086                  */
2087                 if (lmv_dir_migrating(lsm)) {
2088                         hash_type = lsm->lsm_md_hash_type &
2089                                     ~LMV_HASH_FLAG_MIGRATION;
2090                         stripe_count = lsm->lsm_md_migrate_offset;
2091
2092                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
2093                                                       name, namelen);
2094                         if (rc < 0)
2095                                 RETURN(rc);
2096
2097                         op_data->op_fid2 = lsm->lsm_md_oinfo[rc].lmo_fid;
2098                         tp_tgt = lmv_tgt(lmv, lsm->lsm_md_oinfo[rc].lmo_mds);
2099                         if (!tp_tgt)
2100                                 RETURN(-ENODEV);
2101                 }
2102         } else {
2103                 sp_tgt = parent_tgt;
2104         }
2105
2106         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2107         if (IS_ERR(child_tgt))
2108                 RETURN(PTR_ERR(child_tgt));
2109
2110         if (!S_ISDIR(op_data->op_mode) && tp_tgt)
2111                 rc = __lmv_fid_alloc(lmv, &target_fid, tp_tgt->ltd_index);
2112         else
2113                 rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2114         if (rc)
2115                 RETURN(rc);
2116
2117         /*
2118          * for directory, send migrate request to the MDT where the object will
2119          * be migrated to, because we can't create a striped directory remotely.
2120          *
2121          * otherwise, send to the MDT where source is located because regular
2122          * file may open lease.
2123          *
2124          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2125          * backward compatibility.
2126          */
2127         if (S_ISDIR(op_data->op_mode) &&
2128             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2129                 tgt = lmv_fid2tgt(lmv, &target_fid);
2130                 if (IS_ERR(tgt))
2131                         RETURN(PTR_ERR(tgt));
2132         } else {
2133                 tgt = child_tgt;
2134         }
2135
2136         /* cancel UPDATE lock of parent master object */
2137         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2138                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2139         if (rc)
2140                 RETURN(rc);
2141
2142         /* cancel UPDATE lock of source parent */
2143         if (sp_tgt != parent_tgt) {
2144                 /*
2145                  * migrate RPC packs master object FID, because we can only pack
2146                  * two FIDs in reint RPC, but MDS needs to know both source
2147                  * parent and target parent, and it will obtain them from master
2148                  * FID and LMV, the other FID in RPC is kept for target.
2149                  *
2150                  * since this FID is not passed to MDC, cancel it anyway.
2151                  */
2152                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2153                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2154                 if (rc)
2155                         RETURN(rc);
2156
2157                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2158         }
2159         op_data->op_fid4 = target_fid;
2160
2161         /* cancel UPDATE locks of target parent */
2162         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2163                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2164         if (rc)
2165                 RETURN(rc);
2166
2167         /* cancel LOOKUP lock of source if source is remote object */
2168         if (child_tgt != sp_tgt) {
2169                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2170                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2171                                       MF_MDC_CANCEL_FID3);
2172                 if (rc)
2173                         RETURN(rc);
2174         }
2175
2176         /* cancel ELC locks of source */
2177         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2178                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2179         if (rc)
2180                 RETURN(rc);
2181
2182         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2183
2184         RETURN(rc);
2185 }
2186
2187 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2188                       const char *old, size_t oldlen,
2189                       const char *new, size_t newlen,
2190                       struct ptlrpc_request **request)
2191 {
2192         struct obd_device *obd = exp->exp_obd;
2193         struct lmv_obd *lmv = &obd->u.lmv;
2194         struct lmv_tgt_desc *sp_tgt;
2195         struct lmv_tgt_desc *tp_tgt = NULL;
2196         struct lmv_tgt_desc *src_tgt = NULL;
2197         struct lmv_tgt_desc *tgt;
2198         struct mdt_body *body;
2199         int rc;
2200
2201         ENTRY;
2202
2203         LASSERT(oldlen != 0);
2204
2205         if (op_data->op_cli_flags & CLI_MIGRATE) {
2206                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2207                 RETURN(rc);
2208         }
2209
2210         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2211         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2212         op_data->op_cap = cfs_curproc_cap_pack();
2213
2214         op_data->op_name = new;
2215         op_data->op_namelen = newlen;
2216
2217         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2218         if (IS_ERR(tp_tgt))
2219                 RETURN(PTR_ERR(tp_tgt));
2220
2221         /* Since the target child might be destroyed, and it might become
2222          * orphan, and we can only check orphan on the local MDT right now, so
2223          * we send rename request to the MDT where target child is located. If
2224          * target child does not exist, then it will send the request to the
2225          * target parent */
2226         if (fid_is_sane(&op_data->op_fid4)) {
2227                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2228                 if (IS_ERR(tgt))
2229                         RETURN(PTR_ERR(tgt));
2230         } else {
2231                 tgt = tp_tgt;
2232         }
2233
2234         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2235
2236         /* cancel UPDATE locks of target parent */
2237         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2238                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2239         if (rc != 0)
2240                 RETURN(rc);
2241
2242         if (fid_is_sane(&op_data->op_fid4)) {
2243                 /* cancel LOOKUP lock of target on target parent */
2244                 if (tgt != tp_tgt) {
2245                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2246                                               tgt->ltd_index, LCK_EX,
2247                                               MDS_INODELOCK_LOOKUP,
2248                                               MF_MDC_CANCEL_FID4);
2249                         if (rc != 0)
2250                                 RETURN(rc);
2251                 }
2252         }
2253
2254         if (fid_is_sane(&op_data->op_fid3)) {
2255                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2256                 if (IS_ERR(src_tgt))
2257                         RETURN(PTR_ERR(src_tgt));
2258
2259                 /* cancel ELC locks of source */
2260                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2261                                       LCK_EX, MDS_INODELOCK_ELC,
2262                                       MF_MDC_CANCEL_FID3);
2263                 if (rc != 0)
2264                         RETURN(rc);
2265         }
2266
2267         op_data->op_name = old;
2268         op_data->op_namelen = oldlen;
2269 retry:
2270         sp_tgt = lmv_locate_tgt(lmv, op_data);
2271         if (IS_ERR(sp_tgt))
2272                 RETURN(PTR_ERR(sp_tgt));
2273
2274         /* cancel UPDATE locks of source parent */
2275         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2276                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2277         if (rc != 0)
2278                 RETURN(rc);
2279
2280         if (fid_is_sane(&op_data->op_fid3)) {
2281                 /* cancel LOOKUP lock of source on source parent */
2282                 if (src_tgt != sp_tgt) {
2283                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2284                                               tgt->ltd_index, LCK_EX,
2285                                               MDS_INODELOCK_LOOKUP,
2286                                               MF_MDC_CANCEL_FID3);
2287                         if (rc != 0)
2288                                 RETURN(rc);
2289                 }
2290         }
2291
2292 rename:
2293         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2294                 PFID(&op_data->op_fid1), (int)oldlen, old,
2295                 PFID(&op_data->op_fid2), (int)newlen, new);
2296
2297         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2298                         request);
2299         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2300                 ptlrpc_req_finished(*request);
2301                 *request = NULL;
2302                 goto retry;
2303         }
2304
2305         if (rc && rc != -EXDEV)
2306                 RETURN(rc);
2307
2308         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2309         if (body == NULL)
2310                 RETURN(-EPROTO);
2311
2312         /* Not cross-ref case, just get out of here. */
2313         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2314                 RETURN(rc);
2315
2316         op_data->op_fid4 = body->mbo_fid1;
2317
2318         ptlrpc_req_finished(*request);
2319         *request = NULL;
2320
2321         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2322         if (IS_ERR(tgt))
2323                 RETURN(PTR_ERR(tgt));
2324
2325         if (fid_is_sane(&op_data->op_fid4)) {
2326                 /* cancel LOOKUP lock of target on target parent */
2327                 if (tgt != tp_tgt) {
2328                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2329                                               tgt->ltd_index, LCK_EX,
2330                                               MDS_INODELOCK_LOOKUP,
2331                                               MF_MDC_CANCEL_FID4);
2332                         if (rc != 0)
2333                                 RETURN(rc);
2334                 }
2335         }
2336
2337         goto rename;
2338 }
2339
2340 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2341                        void *ea, size_t ealen, struct ptlrpc_request **request)
2342 {
2343         struct obd_device *obd = exp->exp_obd;
2344         struct lmv_obd *lmv = &obd->u.lmv;
2345         struct lmv_tgt_desc *tgt;
2346         int rc = 0;
2347
2348         ENTRY;
2349
2350         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2351                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2352                op_data->op_xvalid);
2353
2354         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2355         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2356         if (IS_ERR(tgt))
2357                 RETURN(PTR_ERR(tgt));
2358
2359         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2360
2361         RETURN(rc);
2362 }
2363
2364 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2365                      struct ptlrpc_request **request)
2366 {
2367         struct obd_device *obd = exp->exp_obd;
2368         struct lmv_obd *lmv = &obd->u.lmv;
2369         struct lmv_tgt_desc *tgt;
2370         int rc;
2371
2372         ENTRY;
2373
2374         tgt = lmv_fid2tgt(lmv, fid);
2375         if (IS_ERR(tgt))
2376                 RETURN(PTR_ERR(tgt));
2377
2378         rc = md_fsync(tgt->ltd_exp, fid, request);
2379         RETURN(rc);
2380 }
2381
2382 struct stripe_dirent {
2383         struct page             *sd_page;
2384         struct lu_dirpage       *sd_dp;
2385         struct lu_dirent        *sd_ent;
2386         bool                     sd_eof;
2387 };
2388
2389 struct lmv_dir_ctxt {
2390         struct lmv_obd          *ldc_lmv;
2391         struct md_op_data       *ldc_op_data;
2392         struct md_callback      *ldc_cb_op;
2393         __u64                    ldc_hash;
2394         int                      ldc_count;
2395         struct stripe_dirent     ldc_stripes[0];
2396 };
2397
2398 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2399 {
2400         if (stripe->sd_page) {
2401                 kunmap(stripe->sd_page);
2402                 put_page(stripe->sd_page);
2403                 stripe->sd_page = NULL;
2404                 stripe->sd_ent = NULL;
2405         }
2406 }
2407
2408 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2409 {
2410         int i;
2411
2412         for (i = 0; i < ctxt->ldc_count; i++)
2413                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2414 }
2415
2416 /* if @ent is dummy, or . .., get next */
2417 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2418                                            struct lu_dirent *ent,
2419                                            int stripe_index)
2420 {
2421         for (; ent; ent = lu_dirent_next(ent)) {
2422                 /* Skip dummy entry */
2423                 if (le16_to_cpu(ent->lde_namelen) == 0)
2424                         continue;
2425
2426                 /* skip . and .. for other stripes */
2427                 if (stripe_index &&
2428                     (strncmp(ent->lde_name, ".",
2429                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2430                      strncmp(ent->lde_name, "..",
2431                              le16_to_cpu(ent->lde_namelen)) == 0))
2432                         continue;
2433
2434                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2435                         break;
2436         }
2437
2438         return ent;
2439 }
2440
2441 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2442                                             struct stripe_dirent *stripe,
2443                                             int stripe_index)
2444 {
2445         struct md_op_data *op_data = ctxt->ldc_op_data;
2446         struct lmv_oinfo *oinfo;
2447         struct lu_fid fid = op_data->op_fid1;
2448         struct inode *inode = op_data->op_data;
2449         struct lmv_tgt_desc *tgt;
2450         struct lu_dirent *ent = stripe->sd_ent;
2451         __u64 hash = ctxt->ldc_hash;
2452         int rc = 0;
2453
2454         ENTRY;
2455
2456         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2457         LASSERT(!ent);
2458
2459         do {
2460                 if (stripe->sd_page) {
2461                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2462
2463                         /* @hash should be the last dirent hash */
2464                         LASSERTF(hash <= end,
2465                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2466                                  ctxt, stripe, hash, end);
2467                         /* unload last page */
2468                         stripe_dirent_unload(stripe);
2469                         /* eof */
2470                         if (end == MDS_DIR_END_OFF) {
2471                                 stripe->sd_eof = true;
2472                                 break;
2473                         }
2474                         hash = end;
2475                 }
2476
2477                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2478                 if (!oinfo->lmo_root) {
2479                         rc = -ENOENT;
2480                         break;
2481                 }
2482
2483                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2484                 if (!tgt) {
2485                         rc = -ENODEV;
2486                         break;
2487                 }
2488
2489                 /* op_data is shared by stripes, reset after use */
2490                 op_data->op_fid1 = oinfo->lmo_fid;
2491                 op_data->op_fid2 = oinfo->lmo_fid;
2492                 op_data->op_data = oinfo->lmo_root;
2493
2494                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2495                                   &stripe->sd_page);
2496
2497                 op_data->op_fid1 = fid;
2498                 op_data->op_fid2 = fid;
2499                 op_data->op_data = inode;
2500
2501                 if (rc)
2502                         break;
2503
2504                 stripe->sd_dp = page_address(stripe->sd_page);
2505                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2506                                         stripe_index);
2507                 /* in case a page filled with ., .. and dummy, read next */
2508         } while (!ent);
2509
2510         stripe->sd_ent = ent;
2511         if (rc) {
2512                 LASSERT(!ent);
2513                 /* treat error as eof, so dir can be partially accessed */
2514                 stripe->sd_eof = true;
2515                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2516                               "directory is partially accessed!\n",
2517                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2518                               rc);
2519         }
2520
2521         RETURN(ent);
2522 }
2523
2524 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2525 {
2526         struct obd_device *obd = exp->exp_obd;
2527         struct lmv_obd *lmv = &obd->u.lmv;
2528         struct lmv_tgt_desc *tgt;
2529         int rc;
2530
2531         ENTRY;
2532
2533         rc = lmv_check_connect(obd);
2534         if (rc != 0)
2535                 RETURN(rc);
2536
2537         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2538         if (IS_ERR(tgt))
2539                 RETURN(PTR_ERR(tgt));
2540
2541         data->op_flags |= MF_MDC_CANCEL_FID1;
2542         rc = md_file_resync(tgt->ltd_exp, data);
2543         RETURN(rc);
2544 }
2545
2546 /**
2547  * Get dirent with the closest hash for striped directory
2548  *
2549  * This function will search the dir entry, whose hash value is the
2550  * closest(>=) to hash from all of sub-stripes, and it is only being called
2551  * for striped directory.
2552  *
2553  * \param[in] ctxt              dir read context
2554  *
2555  * \retval                      dirent get the entry successfully
2556  *                              NULL does not get the entry, normally it means
2557  *                              it reaches the end of the directory, while read
2558  *                              stripe dirent error is ignored to allow partial
2559  *                              access.
2560  */
2561 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2562 {
2563         struct stripe_dirent *stripe;
2564         struct lu_dirent *ent = NULL;
2565         int i;
2566         int min = -1;
2567
2568         /* TODO: optimize with k-way merge sort */
2569         for (i = 0; i < ctxt->ldc_count; i++) {
2570                 stripe = &ctxt->ldc_stripes[i];
2571                 if (stripe->sd_eof)
2572                         continue;
2573
2574                 if (!stripe->sd_ent) {
2575                         stripe_dirent_load(ctxt, stripe, i);
2576                         if (!stripe->sd_ent) {
2577                                 LASSERT(stripe->sd_eof);
2578                                 continue;
2579                         }
2580                 }
2581
2582                 if (min == -1 ||
2583                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2584                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2585                         min = i;
2586                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2587                             ctxt->ldc_hash)
2588                                 break;
2589                 }
2590         }
2591
2592         if (min != -1) {
2593                 stripe = &ctxt->ldc_stripes[min];
2594                 ent = stripe->sd_ent;
2595                 /* pop found dirent */
2596                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2597                                                    min);
2598         }
2599
2600         return ent;
2601 }
2602
2603 /**
2604  * Build dir entry page for striped directory
2605  *
2606  * This function gets one entry by @offset from a striped directory. It will
2607  * read entries from all of stripes, and choose one closest to the required
2608  * offset(&offset). A few notes
2609  * 1. skip . and .. for non-zero stripes, because there can only have one .
2610  * and .. in a directory.
2611  * 2. op_data will be shared by all of stripes, instead of allocating new
2612  * one, so need to restore before reusing.
2613  *
2614  * \param[in] exp       obd export refer to LMV
2615  * \param[in] op_data   hold those MD parameters of read_entry
2616  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2617  * \param[in] offset    starting hash offset
2618  * \param[out] ppage    the page holding the entry. Note: because the entry
2619  *                      will be accessed in upper layer, so we need hold the
2620  *                      page until the usages of entry is finished, see
2621  *                      ll_dir_entry_next.
2622  *
2623  * retval               =0 if get entry successfully
2624  *                      <0 cannot get entry
2625  */
2626 static int lmv_striped_read_page(struct obd_export *exp,
2627                                  struct md_op_data *op_data,
2628                                  struct md_callback *cb_op,
2629                                  __u64 offset, struct page **ppage)
2630 {
2631         struct page *page = NULL;
2632         struct lu_dirpage *dp;
2633         void *start;
2634         struct lu_dirent *ent;
2635         struct lu_dirent *last_ent;
2636         int stripe_count;
2637         struct lmv_dir_ctxt *ctxt;
2638         struct lu_dirent *next = NULL;
2639         __u16 ent_size;
2640         size_t left_bytes;
2641         int rc = 0;
2642         ENTRY;
2643
2644         /* Allocate a page and read entries from all of stripes and fill
2645          * the page by hash order */
2646         page = alloc_page(GFP_KERNEL);
2647         if (!page)
2648                 RETURN(-ENOMEM);
2649
2650         /* Initialize the entry page */
2651         dp = kmap(page);
2652         memset(dp, 0, sizeof(*dp));
2653         dp->ldp_hash_start = cpu_to_le64(offset);
2654
2655         start = dp + 1;
2656         left_bytes = PAGE_SIZE - sizeof(*dp);
2657         ent = start;
2658         last_ent = ent;
2659
2660         /* initalize dir read context */
2661         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2662         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2663         if (!ctxt)
2664                 GOTO(free_page, rc = -ENOMEM);
2665         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2666         ctxt->ldc_op_data = op_data;
2667         ctxt->ldc_cb_op = cb_op;
2668         ctxt->ldc_hash = offset;
2669         ctxt->ldc_count = stripe_count;
2670
2671         while (1) {
2672                 next = lmv_dirent_next(ctxt);
2673
2674                 /* end of directory */
2675                 if (!next) {
2676                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2677                         break;
2678                 }
2679                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2680
2681                 ent_size = le16_to_cpu(next->lde_reclen);
2682
2683                 /* the last entry lde_reclen is 0, but it might not be the last
2684                  * one of this temporay dir page */
2685                 if (!ent_size)
2686                         ent_size = lu_dirent_calc_size(
2687                                         le16_to_cpu(next->lde_namelen),
2688                                         le32_to_cpu(next->lde_attrs));
2689                 /* page full */
2690                 if (ent_size > left_bytes)
2691                         break;
2692
2693                 memcpy(ent, next, ent_size);
2694
2695                 /* Replace . with master FID and Replace .. with the parent FID
2696                  * of master object */
2697                 if (strncmp(ent->lde_name, ".",
2698                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2699                     le16_to_cpu(ent->lde_namelen) == 1)
2700                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2701                 else if (strncmp(ent->lde_name, "..",
2702                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2703                            le16_to_cpu(ent->lde_namelen) == 2)
2704                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2705
2706                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2707                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2708                        le64_to_cpu(ent->lde_hash));
2709
2710                 left_bytes -= ent_size;
2711                 ent->lde_reclen = cpu_to_le16(ent_size);
2712                 last_ent = ent;
2713                 ent = (void *)ent + ent_size;
2714         };
2715
2716         last_ent->lde_reclen = 0;
2717
2718         if (ent == start)
2719                 dp->ldp_flags |= LDF_EMPTY;
2720         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2721                 dp->ldp_flags |= LDF_COLLIDE;
2722         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2723         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2724
2725         put_lmv_dir_ctxt(ctxt);
2726         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2727
2728         *ppage = page;
2729
2730         RETURN(0);
2731
2732 free_page:
2733         kunmap(page);
2734         __free_page(page);
2735
2736         return rc;
2737 }
2738
2739 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2740                   struct md_callback *cb_op, __u64 offset,
2741                   struct page **ppage)
2742 {
2743         struct obd_device *obd = exp->exp_obd;
2744         struct lmv_obd *lmv = &obd->u.lmv;
2745         struct lmv_tgt_desc *tgt;
2746         int rc;
2747
2748         ENTRY;
2749
2750         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2751                 RETURN(-ENODATA);
2752
2753         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2754                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2755                 RETURN(rc);
2756         }
2757
2758         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2759         if (IS_ERR(tgt))
2760                 RETURN(PTR_ERR(tgt));
2761
2762         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2763
2764         RETURN(rc);
2765 }
2766
2767 /**
2768  * Unlink a file/directory
2769  *
2770  * Unlink a file or directory under the parent dir. The unlink request
2771  * usually will be sent to the MDT where the child is located, but if
2772  * the client does not have the child FID then request will be sent to the
2773  * MDT where the parent is located.
2774  *
2775  * If the parent is a striped directory then it also needs to locate which
2776  * stripe the name of the child is located, and replace the parent FID
2777  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2778  * it will walk through all of sub-stripes until the child is being
2779  * unlinked finally.
2780  *
2781  * \param[in] exp       export refer to LMV
2782  * \param[in] op_data   different parameters transferred beween client
2783  *                      MD stacks, name, namelen, FIDs etc.
2784  *                      op_fid1 is the parent FID, op_fid2 is the child
2785  *                      FID.
2786  * \param[out] request  point to the request of unlink.
2787  *
2788  * retval               0 if succeed
2789  *                      negative errno if failed.
2790  */
2791 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2792                       struct ptlrpc_request **request)
2793 {
2794         struct obd_device *obd = exp->exp_obd;
2795         struct lmv_obd *lmv = &obd->u.lmv;
2796         struct lmv_tgt_desc *tgt;
2797         struct lmv_tgt_desc *parent_tgt;
2798         struct mdt_body *body;
2799         int rc;
2800
2801         ENTRY;
2802
2803         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2804         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2805         op_data->op_cap = cfs_curproc_cap_pack();
2806
2807 retry:
2808         parent_tgt = lmv_locate_tgt(lmv, op_data);
2809         if (IS_ERR(parent_tgt))
2810                 RETURN(PTR_ERR(parent_tgt));
2811
2812         if (likely(!fid_is_zero(&op_data->op_fid2))) {
2813                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2814                 if (IS_ERR(tgt))
2815                         RETURN(PTR_ERR(tgt));
2816         } else {
2817                 tgt = parent_tgt;
2818         }
2819
2820         /*
2821          * If child's fid is given, cancel unused locks for it if it is from
2822          * another export than parent.
2823          *
2824          * LOOKUP lock for child (fid3) should also be cancelled on parent
2825          * tgt_tgt in mdc_unlink().
2826          */
2827         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2828
2829         if (parent_tgt != tgt)
2830                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2831                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2832                                       MF_MDC_CANCEL_FID3);
2833
2834         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2835                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2836         if (rc)
2837                 RETURN(rc);
2838
2839         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2840                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2841                tgt->ltd_index);
2842
2843         rc = md_unlink(tgt->ltd_exp, op_data, request);
2844         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2845                 ptlrpc_req_finished(*request);
2846                 *request = NULL;
2847                 goto retry;
2848         }
2849
2850         if (rc != -EREMOTE)
2851                 RETURN(rc);
2852
2853         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2854         if (body == NULL)
2855                 RETURN(-EPROTO);
2856
2857         /* Not cross-ref case, just get out of here. */
2858         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2859                 RETURN(rc);
2860
2861         /* This is a remote object, try remote MDT. */
2862         op_data->op_fid2 = body->mbo_fid1;
2863         ptlrpc_req_finished(*request);
2864         *request = NULL;
2865
2866         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2867         if (IS_ERR(tgt))
2868                 RETURN(PTR_ERR(tgt));
2869
2870         goto retry;
2871 }
2872
2873 static int lmv_precleanup(struct obd_device *obd)
2874 {
2875         ENTRY;
2876         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2877         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2878         lprocfs_obd_cleanup(obd);
2879         lprocfs_free_md_stats(obd);
2880         RETURN(0);
2881 }
2882
2883 /**
2884  * Get by key a value associated with a LMV device.
2885  *
2886  * Dispatch request to lower-layer devices as needed.
2887  *
2888  * \param[in] env               execution environment for this thread
2889  * \param[in] exp               export for the LMV device
2890  * \param[in] keylen            length of key identifier
2891  * \param[in] key               identifier of key to get value for
2892  * \param[in] vallen            size of \a val
2893  * \param[out] val              pointer to storage location for value
2894  * \param[in] lsm               optional striping metadata of object
2895  *
2896  * \retval 0            on success
2897  * \retval negative     negated errno on failure
2898  */
2899 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2900                         __u32 keylen, void *key, __u32 *vallen, void *val)
2901 {
2902         struct obd_device *obd;
2903         struct lmv_obd *lmv;
2904         struct lu_tgt_desc *tgt;
2905         int rc = 0;
2906
2907         ENTRY;
2908
2909         obd = class_exp2obd(exp);
2910         if (obd == NULL) {
2911                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2912                        exp->exp_handle.h_cookie);
2913                 RETURN(-EINVAL);
2914         }
2915
2916         lmv = &obd->u.lmv;
2917         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2918                 LASSERT(*vallen == sizeof(__u32));
2919                 lmv_foreach_connected_tgt(lmv, tgt) {
2920                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2921                                           vallen, val))
2922                                 RETURN(0);
2923                 }
2924                 RETURN(-EINVAL);
2925         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2926                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2927                    KEY_IS(KEY_CONN_DATA)) {
2928                 /*
2929                  * Forwarding this request to first MDS, it should know LOV
2930                  * desc.
2931                  */
2932                 tgt = lmv_tgt(lmv, 0);
2933                 if (!tgt)
2934                         RETURN(-ENODEV);
2935
2936                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
2937                 if (!rc && KEY_IS(KEY_CONN_DATA))
2938                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2939                 RETURN(rc);
2940         } else if (KEY_IS(KEY_TGT_COUNT)) {
2941                 *((int *)val) = lmv->lmv_mdt_descs.ltd_tgts_size;
2942                 RETURN(0);
2943         }
2944
2945         CDEBUG(D_IOCTL, "Invalid key\n");
2946         RETURN(-EINVAL);
2947 }
2948
2949 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
2950                      int *__rcs, struct ptlrpc_request_set *_set)
2951 {
2952         struct obd_device *obddev = class_exp2obd(exp);
2953         struct ptlrpc_request_set *set = _set;
2954         struct lmv_obd *lmv = &obddev->u.lmv;
2955         int tgt_count = lmv->lmv_mdt_count;
2956         struct lu_tgt_desc *tgt;
2957         struct fid_array *fat, **fas = NULL;
2958         int i, rc, **rcs = NULL;
2959
2960         if (!set) {
2961                 set = ptlrpc_prep_set();
2962                 if (!set)
2963                         RETURN(-ENOMEM);
2964         }
2965
2966         /* split FIDs by targets */
2967         OBD_ALLOC(fas, sizeof(fas) * tgt_count);
2968         if (fas == NULL)
2969                 GOTO(out, rc = -ENOMEM);
2970         OBD_ALLOC(rcs, sizeof(int *) * tgt_count);
2971         if (rcs == NULL)
2972                 GOTO(out_fas, rc = -ENOMEM);
2973
2974         for (i = 0; i < fa->fa_nr; i++) {
2975                 unsigned int idx;
2976
2977                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
2978                 if (rc) {
2979                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
2980                                PFID(&fa->fa_fids[i]), rc);
2981                         continue;
2982                 }
2983                 LASSERT(idx < tgt_count);
2984                 if (!fas[idx])
2985                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
2986                                   fa_fids[fa->fa_nr]));
2987                 if (!fas[idx])
2988                         GOTO(out, rc = -ENOMEM);
2989                 if (!rcs[idx])
2990                         OBD_ALLOC(rcs[idx], sizeof(int) * fa->fa_nr);
2991                 if (!rcs[idx])
2992                         GOTO(out, rc = -ENOMEM);
2993
2994                 fat = fas[idx];
2995                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
2996         }
2997
2998         lmv_foreach_connected_tgt(lmv, tgt) {
2999                 fat = fas[tgt->ltd_index];
3000                 if (!fat || fat->fa_nr == 0)
3001                         continue;
3002                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3003         }
3004
3005         rc = ptlrpc_set_wait(NULL, set);
3006         if (rc == 0) {
3007                 int j = 0;
3008                 for (i = 0; i < tgt_count; i++) {
3009                         fat = fas[i];
3010                         if (!fat || fat->fa_nr == 0)
3011                                 continue;
3012                         /* copy FIDs back */
3013                         memcpy(fa->fa_fids + j, fat->fa_fids,
3014                                fat->fa_nr * sizeof(struct lu_fid));
3015                         /* copy rcs back */
3016                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3017                         j += fat->fa_nr;
3018                 }
3019         }
3020         if (set != _set)
3021                 ptlrpc_set_destroy(set);
3022
3023 out:
3024         for (i = 0; i < tgt_count; i++) {
3025                 if (fas && fas[i])
3026                         OBD_FREE(fas[i], offsetof(struct fid_array,
3027                                                 fa_fids[fa->fa_nr]));
3028                 if (rcs && rcs[i])
3029                         OBD_FREE(rcs[i], sizeof(int) * fa->fa_nr);
3030         }
3031         if (rcs)
3032                 OBD_FREE(rcs, sizeof(int *) * tgt_count);
3033 out_fas:
3034         if (fas)
3035                 OBD_FREE(fas, sizeof(fas) * tgt_count);
3036
3037         RETURN(rc);
3038 }
3039
3040 /**
3041  * Asynchronously set by key a value associated with a LMV device.
3042  *
3043  * Dispatch request to lower-layer devices as needed.
3044  *
3045  * \param[in] env       execution environment for this thread
3046  * \param[in] exp       export for the LMV device
3047  * \param[in] keylen    length of key identifier
3048  * \param[in] key       identifier of key to store value for
3049  * \param[in] vallen    size of value to store
3050  * \param[in] val       pointer to data to be stored
3051  * \param[in] set       optional list of related ptlrpc requests
3052  *
3053  * \retval 0            on success
3054  * \retval negative     negated errno on failure
3055  */
3056 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3057                         __u32 keylen, void *key, __u32 vallen, void *val,
3058                         struct ptlrpc_request_set *set)
3059 {
3060         struct lmv_tgt_desc     *tgt = NULL;
3061         struct obd_device       *obd;
3062         struct lmv_obd          *lmv;
3063         int rc = 0;
3064         ENTRY;
3065
3066         obd = class_exp2obd(exp);
3067         if (obd == NULL) {
3068                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3069                        exp->exp_handle.h_cookie);
3070                 RETURN(-EINVAL);
3071         }
3072         lmv = &obd->u.lmv;
3073
3074         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3075             KEY_IS(KEY_DEFAULT_EASIZE)) {
3076                 int err = 0;
3077
3078                 lmv_foreach_connected_tgt(lmv, tgt) {
3079                         err = obd_set_info_async(env, tgt->ltd_exp,
3080                                                  keylen, key, vallen, val, set);
3081                         if (err && rc == 0)
3082                                 rc = err;
3083                 }
3084
3085                 RETURN(rc);
3086         }
3087
3088         RETURN(-EINVAL);
3089 }
3090
3091 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3092                             const struct lmv_mds_md_v1 *lmm1)
3093 {
3094         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3095         int             stripe_count;
3096         int             cplen;
3097         int             i;
3098         int             rc = 0;
3099         ENTRY;
3100
3101         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3102         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3103         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3104         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3105                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3106         else
3107                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3108         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3109         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3110         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3111         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3112                         sizeof(lsm->lsm_md_pool_name));
3113
3114         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3115                 RETURN(-E2BIG);
3116
3117         CDEBUG(D_INFO, "unpack lsm count %d, master %d hash_type %#x "
3118                "layout_version %d\n", lsm->lsm_md_stripe_count,
3119                lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
3120                lsm->lsm_md_layout_version);
3121
3122         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3123         for (i = 0; i < stripe_count; i++) {
3124                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3125                               &lmm1->lmv_stripe_fids[i]);
3126                 /*
3127                  * set default value -1, so lmv_locate_tgt() knows this stripe
3128                  * target is not initialized.
3129                  */
3130                 lsm->lsm_md_oinfo[i].lmo_mds = (u32)-1;
3131                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3132                         continue;
3133
3134                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3135                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3136                 if (rc == -ENOENT)
3137                         continue;
3138
3139                 if (rc)
3140                         RETURN(rc);
3141
3142                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3143                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3144         }
3145
3146         RETURN(rc);
3147 }
3148
3149 static inline int lmv_unpack_user_md(struct obd_export *exp,
3150                                      struct lmv_stripe_md *lsm,
3151                                      const struct lmv_user_md *lmu)
3152 {
3153         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3154         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3155         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3156         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3157
3158         return 0;
3159 }
3160
3161 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3162                         const union lmv_mds_md *lmm, size_t lmm_size)
3163 {
3164         struct lmv_stripe_md     *lsm;
3165         int                      lsm_size;
3166         int                      rc;
3167         bool                     allocated = false;
3168         ENTRY;
3169
3170         LASSERT(lsmp != NULL);
3171
3172         lsm = *lsmp;
3173         /* Free memmd */
3174         if (lsm != NULL && lmm == NULL) {
3175                 int i;
3176                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3177
3178                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3179                         size_t lfm_size;
3180
3181                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3182                                                               lfm_value[0]);
3183                         OBD_FREE_LARGE(lfm, lfm_size);
3184                         RETURN(0);
3185                 }
3186
3187                 if (lmv_dir_striped(lsm)) {
3188                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3189                                 if (lsm->lsm_md_oinfo[i].lmo_root)
3190                                         iput(lsm->lsm_md_oinfo[i].lmo_root);
3191                         }
3192                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3193                 } else {
3194                         lsm_size = lmv_stripe_md_size(0);
3195                 }
3196                 OBD_FREE(lsm, lsm_size);
3197                 *lsmp = NULL;
3198                 RETURN(0);
3199         }
3200
3201         /* foreign lmv case */
3202         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3203                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3204
3205                 if (lfm == NULL) {
3206                         OBD_ALLOC_LARGE(lfm, lmm_size);
3207                         if (lfm == NULL)
3208                                 RETURN(-ENOMEM);
3209                         *lsmp = (struct lmv_stripe_md *)lfm;
3210                 }
3211                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3212                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3213                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3214                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3215                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3216                        lfm->lfm_length);
3217                 RETURN(lmm_size);
3218         }
3219
3220         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3221                 RETURN(-EPERM);
3222
3223         /* Unpack memmd */
3224         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3225             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3226                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3227                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3228                        -EIO);
3229                 RETURN(-EIO);
3230         }
3231
3232         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3233                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3234         else
3235                 /**
3236                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3237                  * stripecount should be 0 then.
3238                  */
3239                 lsm_size = lmv_stripe_md_size(0);
3240
3241         if (lsm == NULL) {
3242                 OBD_ALLOC(lsm, lsm_size);
3243                 if (lsm == NULL)
3244                         RETURN(-ENOMEM);
3245                 allocated = true;
3246                 *lsmp = lsm;
3247         }
3248
3249         switch (le32_to_cpu(lmm->lmv_magic)) {
3250         case LMV_MAGIC_V1:
3251                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3252                 break;
3253         case LMV_USER_MAGIC:
3254                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3255                 break;
3256         default:
3257                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3258                        le32_to_cpu(lmm->lmv_magic));
3259                 rc = -EINVAL;
3260                 break;
3261         }
3262
3263         if (rc != 0 && allocated) {
3264                 OBD_FREE(lsm, lsm_size);
3265                 *lsmp = NULL;
3266                 lsm_size = rc;
3267         }
3268         RETURN(lsm_size);
3269 }
3270
3271 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3272 {
3273         lmv_unpackmd(NULL, &lsm, NULL, 0);
3274 }
3275 EXPORT_SYMBOL(lmv_free_memmd);
3276
3277 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3278                              union ldlm_policy_data *policy,
3279                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3280                              void *opaque)
3281 {
3282         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3283         struct lu_tgt_desc *tgt;
3284         int err;
3285         int rc = 0;
3286
3287         ENTRY;
3288