Whamcloud - gitweb
96b6b5d7f765e52882b49af3de17e8253397ab73
[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  */
31
32 #define DEBUG_SUBSYSTEM S_LMV
33
34 #include <linux/file.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/user_namespace.h>
38 #include <linux/uidgid.h>
39 #include <linux/slab.h>
40 #include <linux/pagemap.h>
41 #include <linux/mm.h>
42 #include <linux/math64.h>
43 #include <linux/seq_file.h>
44 #include <linux/namei.h>
45
46 #include <obd_support.h>
47 #include <lustre_lib.h>
48 #include <lustre_net.h>
49 #include <obd_class.h>
50 #include <lustre_lmv.h>
51 #include <lprocfs_status.h>
52 #include <cl_object.h>
53 #include <lustre_fid.h>
54 #include <uapi/linux/lustre/lustre_ioctl.h>
55 #include <lustre_kernelcomm.h>
56 #include "lmv_internal.h"
57
58 static int lmv_check_connect(struct obd_device *obd);
59
60 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
61                          int activate)
62 {
63         if (tgt->ltd_active == activate)
64                 return;
65
66         tgt->ltd_active = activate;
67         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
68                 (activate ? 1 : -1);
69
70         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
71 }
72
73 /**
74  * Error codes:
75  *
76  *  -EINVAL  : UUID can't be found in the LMV's target list
77  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
78  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
79  */
80 static int lmv_set_mdc_active(struct lmv_obd *lmv,
81                               const struct obd_uuid *uuid,
82                               int activate)
83 {
84         struct lu_tgt_desc *tgt = NULL;
85         struct obd_device *obd;
86         int rc = 0;
87
88         ENTRY;
89
90         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
91                         lmv, uuid->uuid, activate);
92
93         spin_lock(&lmv->lmv_lock);
94         lmv_foreach_connected_tgt(lmv, tgt) {
95                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
96                        tgt->ltd_index, tgt->ltd_uuid.uuid,
97                        tgt->ltd_exp->exp_handle.h_cookie);
98
99                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
100                         break;
101         }
102
103         if (!tgt)
104                 GOTO(out_lmv_lock, rc = -EINVAL);
105
106         obd = class_exp2obd(tgt->ltd_exp);
107         if (obd == NULL)
108                 GOTO(out_lmv_lock, rc = -ENOTCONN);
109
110         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
111                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
112                obd->obd_type->typ_name, tgt->ltd_index);
113         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
114
115         if (tgt->ltd_active == activate) {
116                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
117                        activate ? "" : "in");
118                 GOTO(out_lmv_lock, rc);
119         }
120
121         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
122                activate ? "" : "in");
123         lmv_activate_target(lmv, tgt, activate);
124         EXIT;
125
126  out_lmv_lock:
127         spin_unlock(&lmv->lmv_lock);
128         return rc;
129 }
130
131 static struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
132 {
133         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
134         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
135
136         return tgt ? obd_get_uuid(tgt->ltd_exp) : NULL;
137 }
138
139 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
140                       enum obd_notify_event ev)
141 {
142         struct obd_connect_data *conn_data;
143         struct lmv_obd          *lmv = &obd->u.lmv;
144         struct obd_uuid         *uuid;
145         int                      rc = 0;
146         ENTRY;
147
148         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
149                 CERROR("unexpected notification of %s %s!\n",
150                        watched->obd_type->typ_name,
151                        watched->obd_name);
152                 RETURN(-EINVAL);
153         }
154
155         uuid = &watched->u.cli.cl_target_uuid;
156         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
157                 /*
158                  * Set MDC as active before notifying the observer, so the
159                  * observer can use the MDC normally.
160                  */
161                 rc = lmv_set_mdc_active(lmv, uuid,
162                                         ev == OBD_NOTIFY_ACTIVE);
163                 if (rc) {
164                         CERROR("%sactivation of %s failed: %d\n",
165                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
166                                uuid->uuid, rc);
167                         RETURN(rc);
168                 }
169         } else if (ev == OBD_NOTIFY_OCD) {
170                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
171                 /*
172                  * XXX: Make sure that ocd_connect_flags from all targets are
173                  * the same. Otherwise one of MDTs runs wrong version or
174                  * something like this.  --umka
175                  */
176                 obd->obd_self_export->exp_connect_data = *conn_data;
177         }
178
179         /*
180          * Pass the notification up the chain.
181          */
182         if (obd->obd_observer)
183                 rc = obd_notify(obd->obd_observer, watched, ev);
184
185         RETURN(rc);
186 }
187
188 static int lmv_connect(const struct lu_env *env,
189                        struct obd_export **pexp, struct obd_device *obd,
190                        struct obd_uuid *cluuid, struct obd_connect_data *data,
191                        void *localdata)
192 {
193         struct lmv_obd *lmv = &obd->u.lmv;
194         struct lustre_handle conn = { 0 };
195         struct obd_export *exp;
196         int rc;
197         ENTRY;
198
199         rc = class_connect(&conn, obd, cluuid);
200         if (rc) {
201                 CERROR("class_connection() returned %d\n", rc);
202                 RETURN(rc);
203         }
204
205         exp = class_conn2export(&conn);
206
207         lmv->connected = 0;
208         lmv->conn_data = *data;
209         lmv->lmv_cache = localdata;
210
211         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
212                                                     &obd->obd_kset.kobj);
213         if (!lmv->lmv_tgts_kobj) {
214                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
215                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
216         }
217
218         rc = lmv_check_connect(obd);
219         if (rc != 0)
220                 GOTO(out_sysfs, rc);
221
222         *pexp = exp;
223
224         RETURN(rc);
225
226 out_sysfs:
227         if (lmv->lmv_tgts_kobj)
228                 kobject_put(lmv->lmv_tgts_kobj);
229
230         class_disconnect(exp);
231
232         return rc;
233 }
234
235 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
236                             __u32 def_easize)
237 {
238         struct obd_device *obd = exp->exp_obd;
239         struct lmv_obd *lmv = &obd->u.lmv;
240         struct lmv_tgt_desc *tgt;
241         int change = 0;
242         int rc = 0;
243
244         ENTRY;
245
246         if (lmv->max_easize < easize) {
247                 lmv->max_easize = easize;
248                 change = 1;
249         }
250         if (lmv->max_def_easize < def_easize) {
251                 lmv->max_def_easize = def_easize;
252                 change = 1;
253         }
254
255         if (change == 0)
256                 RETURN(0);
257
258         if (lmv->connected == 0)
259                 RETURN(0);
260
261         lmv_foreach_connected_tgt(lmv, tgt) {
262                 if (!tgt->ltd_active)
263                         continue;
264
265                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
266                 if (rc) {
267                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
268                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
269                         break;
270                 }
271         }
272         RETURN(rc);
273 }
274
275 #define MAX_STRING_SIZE 128
276
277 static int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
278 {
279         struct lmv_obd *lmv = &obd->u.lmv;
280         struct obd_device *mdc_obd;
281         struct obd_export *mdc_exp;
282         struct lu_fld_target target;
283         int  rc;
284         ENTRY;
285
286         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
287                                         &obd->obd_uuid);
288         if (!mdc_obd) {
289                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
290                 RETURN(-EINVAL);
291         }
292
293         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
294                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
295                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
296
297         if (!mdc_obd->obd_set_up) {
298                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
299                 RETURN(-EINVAL);
300         }
301
302         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
303                          &lmv->conn_data, lmv->lmv_cache);
304         if (rc) {
305                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
306                 RETURN(rc);
307         }
308
309         /*
310          * Init fid sequence client for this mdc and add new fld target.
311          */
312         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
313         if (rc)
314                 RETURN(rc);
315
316         target.ft_srv = NULL;
317         target.ft_exp = mdc_exp;
318         target.ft_idx = tgt->ltd_index;
319
320         fld_client_add_target(&lmv->lmv_fld, &target);
321
322         rc = obd_register_observer(mdc_obd, obd);
323         if (rc) {
324                 obd_disconnect(mdc_exp);
325                 CERROR("target %s register_observer error %d\n",
326                        tgt->ltd_uuid.uuid, rc);
327                 RETURN(rc);
328         }
329
330         if (obd->obd_observer) {
331                 /*
332                  * Tell the observer about the new target.
333                  */
334                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
335                                 OBD_NOTIFY_ACTIVE);
336                 if (rc) {
337                         obd_disconnect(mdc_exp);
338                         RETURN(rc);
339                 }
340         }
341
342         tgt->ltd_active = 1;
343         tgt->ltd_exp = mdc_exp;
344         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
345
346         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
347
348         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
349         if (rc) {
350                 obd_disconnect(mdc_exp);
351                 RETURN(rc);
352         }
353
354         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
355                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
356                atomic_read(&obd->obd_refcount));
357
358         lmv_statfs_check_update(obd, tgt);
359
360         if (lmv->lmv_tgts_kobj)
361                 /* Even if we failed to create the link, that's fine */
362                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
363                                        &mdc_obd->obd_kset.kobj,
364                                        mdc_obd->obd_name);
365         RETURN(0);
366 }
367
368 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
369 {
370         LASSERT(tgt);
371         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
372         OBD_FREE_PTR(tgt);
373 }
374
375 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
376                            __u32 index, int gen)
377 {
378         struct obd_device *mdc_obd;
379         struct lmv_obd *lmv = &obd->u.lmv;
380         struct lmv_tgt_desc *tgt;
381         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
382         int rc = 0;
383
384         ENTRY;
385
386         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
387         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
388                                         &obd->obd_uuid);
389         if (!mdc_obd) {
390                 CERROR("%s: Target %s not attached: rc = %d\n",
391                        obd->obd_name, uuidp->uuid, -EINVAL);
392                 RETURN(-EINVAL);
393         }
394
395         OBD_ALLOC_PTR(tgt);
396         if (!tgt)
397                 RETURN(-ENOMEM);
398
399         mutex_init(&tgt->ltd_fid_mutex);
400         tgt->ltd_index = index;
401         tgt->ltd_uuid = *uuidp;
402         tgt->ltd_active = 0;
403
404         mutex_lock(&ltd->ltd_mutex);
405         rc = ltd_add_tgt(ltd, tgt);
406         mutex_unlock(&ltd->ltd_mutex);
407
408         if (rc)
409                 GOTO(out_tgt, rc);
410
411         if (!lmv->connected)
412                 /* lmv_check_connect() will connect this target. */
413                 RETURN(0);
414
415         rc = lmv_connect_mdc(obd, tgt);
416         if (!rc) {
417                 int easize = sizeof(struct lmv_stripe_md) +
418                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
419
420                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
421         }
422
423         RETURN(rc);
424
425 out_tgt:
426         OBD_FREE_PTR(tgt);
427         return rc;
428 }
429
430 static int lmv_check_connect(struct obd_device *obd)
431 {
432         struct lmv_obd *lmv = &obd->u.lmv;
433         struct lmv_tgt_desc *tgt;
434         int easize;
435         int rc;
436
437         ENTRY;
438
439         if (lmv->connected)
440                 RETURN(0);
441
442         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
443         if (lmv->connected)
444                 GOTO(unlock, rc = 0);
445
446         if (!lmv->lmv_mdt_count) {
447                 CERROR("%s: no targets configured: rc = -EINVAL\n",
448                        obd->obd_name);
449                 GOTO(unlock, rc = -EINVAL);
450         }
451
452         if (!lmv_mdt0_inited(lmv)) {
453                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
454                        obd->obd_name);
455                 GOTO(unlock, rc = -EINVAL);
456         }
457
458         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
459                obd->obd_uuid.uuid, obd->obd_name);
460
461         lmv_foreach_tgt(lmv, tgt) {
462                 rc = lmv_connect_mdc(obd, tgt);
463                 if (rc)
464                         GOTO(out_disc, rc);
465         }
466
467         lmv->connected = 1;
468         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
469         lmv_init_ea_size(obd->obd_self_export, easize, 0);
470         EXIT;
471 unlock:
472         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
473
474         return rc;
475
476 out_disc:
477         lmv_foreach_tgt(lmv, tgt) {
478                 tgt->ltd_active = 0;
479                 if (!tgt->ltd_exp)
480                         continue;
481
482                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
483                 obd_disconnect(tgt->ltd_exp);
484         }
485
486         goto unlock;
487 }
488
489 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
490 {
491         struct lmv_obd *lmv = &obd->u.lmv;
492         struct obd_device *mdc_obd;
493         int rc;
494         ENTRY;
495
496         LASSERT(tgt != NULL);
497         LASSERT(obd != NULL);
498
499         mdc_obd = class_exp2obd(tgt->ltd_exp);
500
501         if (mdc_obd) {
502                 mdc_obd->obd_force = obd->obd_force;
503                 mdc_obd->obd_fail = obd->obd_fail;
504                 mdc_obd->obd_no_recov = obd->obd_no_recov;
505
506                 if (lmv->lmv_tgts_kobj)
507                         sysfs_remove_link(lmv->lmv_tgts_kobj,
508                                           mdc_obd->obd_name);
509         }
510
511         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
512         if (rc)
513                 CERROR("Can't finalize fids factory\n");
514
515         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
516                tgt->ltd_exp->exp_obd->obd_name,
517                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
518
519         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
520         rc = obd_disconnect(tgt->ltd_exp);
521         if (rc) {
522                 if (tgt->ltd_active) {
523                         CERROR("Target %s disconnect error %d\n",
524                                tgt->ltd_uuid.uuid, rc);
525                 }
526         }
527
528         lmv_activate_target(lmv, tgt, 0);
529         tgt->ltd_exp = NULL;
530         RETURN(0);
531 }
532
533 static int lmv_disconnect(struct obd_export *exp)
534 {
535         struct obd_device *obd = class_exp2obd(exp);
536         struct lmv_obd *lmv = &obd->u.lmv;
537         struct lmv_tgt_desc *tgt;
538         int rc;
539
540         ENTRY;
541
542         lmv_foreach_connected_tgt(lmv, tgt)
543                 lmv_disconnect_mdc(obd, tgt);
544
545         if (lmv->lmv_tgts_kobj)
546                 kobject_put(lmv->lmv_tgts_kobj);
547
548         if (!lmv->connected)
549                 class_export_put(exp);
550         rc = class_disconnect(exp);
551         lmv->connected = 0;
552
553         RETURN(rc);
554 }
555
556 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
557                         void __user *uarg)
558 {
559         struct obd_device *obd = class_exp2obd(exp);
560         struct lmv_obd *lmv = &obd->u.lmv;
561         struct getinfo_fid2path *gf;
562         struct lmv_tgt_desc *tgt;
563         struct getinfo_fid2path *remote_gf = NULL;
564         struct lu_fid root_fid;
565         int remote_gf_size = 0;
566         int rc;
567
568         gf = karg;
569         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
570         if (IS_ERR(tgt))
571                 RETURN(PTR_ERR(tgt));
572
573         root_fid = *gf->gf_u.gf_root_fid;
574         LASSERT(fid_is_sane(&root_fid));
575
576 repeat_fid2path:
577         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
578         if (rc != 0 && rc != -EREMOTE)
579                 GOTO(out_fid2path, rc);
580
581         /* If remote_gf != NULL, it means just building the
582          * path on the remote MDT, copy this path segement to gf */
583         if (remote_gf != NULL) {
584                 struct getinfo_fid2path *ori_gf;
585                 char *ptr;
586                 int len;
587
588                 ori_gf = (struct getinfo_fid2path *)karg;
589                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
590                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
591                         GOTO(out_fid2path, rc = -EOVERFLOW);
592
593                 ptr = ori_gf->gf_u.gf_path;
594
595                 len = strlen(gf->gf_u.gf_path);
596                 /* move the current path to the right to release space
597                  * for closer-to-root part */
598                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
599                 memcpy(ptr, gf->gf_u.gf_path, len);
600                 ptr[len] = '/';
601         }
602
603         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
604                tgt->ltd_exp->exp_obd->obd_name,
605                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
606                gf->gf_linkno);
607
608         if (rc == 0)
609                 GOTO(out_fid2path, rc);
610
611         /* sigh, has to go to another MDT to do path building further */
612         if (remote_gf == NULL) {
613                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
614                 OBD_ALLOC(remote_gf, remote_gf_size);
615                 if (remote_gf == NULL)
616                         GOTO(out_fid2path, rc = -ENOMEM);
617                 remote_gf->gf_pathlen = PATH_MAX;
618         }
619
620         if (!fid_is_sane(&gf->gf_fid)) {
621                 CERROR("%s: invalid FID "DFID": rc = %d\n",
622                        tgt->ltd_exp->exp_obd->obd_name,
623                        PFID(&gf->gf_fid), -EINVAL);
624                 GOTO(out_fid2path, rc = -EINVAL);
625         }
626
627         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
628         if (IS_ERR(tgt))
629                 GOTO(out_fid2path, rc = -EINVAL);
630
631         remote_gf->gf_fid = gf->gf_fid;
632         remote_gf->gf_recno = -1;
633         remote_gf->gf_linkno = -1;
634         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
635         *remote_gf->gf_u.gf_root_fid = root_fid;
636         gf = remote_gf;
637         goto repeat_fid2path;
638
639 out_fid2path:
640         if (remote_gf != NULL)
641                 OBD_FREE(remote_gf, remote_gf_size);
642         RETURN(rc);
643 }
644
645 static int lmv_hsm_req_count(struct lmv_obd *lmv,
646                              const struct hsm_user_request *hur,
647                              const struct lmv_tgt_desc *tgt_mds)
648 {
649         struct lmv_tgt_desc *curr_tgt;
650         __u32 i;
651         int nr = 0;
652
653         /* count how many requests must be sent to the given target */
654         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
655                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
656                 if (IS_ERR(curr_tgt))
657                         RETURN(PTR_ERR(curr_tgt));
658                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
659                         nr++;
660         }
661         return nr;
662 }
663
664 static int lmv_hsm_req_build(struct lmv_obd *lmv,
665                               struct hsm_user_request *hur_in,
666                               const struct lmv_tgt_desc *tgt_mds,
667                               struct hsm_user_request *hur_out)
668 {
669         __u32 i, nr_out;
670         struct lmv_tgt_desc *curr_tgt;
671
672         /* build the hsm_user_request for the given target */
673         hur_out->hur_request = hur_in->hur_request;
674         nr_out = 0;
675         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
676                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
677                 if (IS_ERR(curr_tgt))
678                         RETURN(PTR_ERR(curr_tgt));
679                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
680                         hur_out->hur_user_item[nr_out] =
681                                                 hur_in->hur_user_item[i];
682                         nr_out++;
683                 }
684         }
685         hur_out->hur_request.hr_itemcount = nr_out;
686         memcpy(hur_data(hur_out), hur_data(hur_in),
687                hur_in->hur_request.hr_data_len);
688
689         RETURN(0);
690 }
691
692 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
693                                  int len, struct lustre_kernelcomm *lk,
694                                  void __user *uarg)
695 {
696         struct lmv_obd *lmv = &obd->u.lmv;
697         struct lu_tgt_desc *tgt;
698         int rc;
699
700         ENTRY;
701
702         /* unregister request (call from llapi_hsm_copytool_fini) */
703         lmv_foreach_connected_tgt(lmv, tgt)
704                 /* best effort: try to clean as much as possible
705                  * (continue on error) */
706                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
707
708         /* Whatever the result, remove copytool from kuc groups.
709          * Unreached coordinators will get EPIPE on next requests
710          * and will unregister automatically.
711          */
712         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
713
714         RETURN(rc);
715 }
716
717 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
718                                int len, struct lustre_kernelcomm *lk,
719                                void __user *uarg)
720 {
721         struct lmv_obd *lmv = &obd->u.lmv;
722         struct file *filp;
723         bool any_set = false;
724         struct kkuc_ct_data *kcd;
725         size_t kcd_size;
726         struct lu_tgt_desc *tgt;
727         __u32 i;
728         int err;
729         int rc = 0;
730
731         ENTRY;
732
733         filp = fget(lk->lk_wfd);
734         if (!filp)
735                 RETURN(-EBADF);
736
737         if (lk->lk_flags & LK_FLG_DATANR)
738                 kcd_size = offsetof(struct kkuc_ct_data,
739                                     kcd_archives[lk->lk_data_count]);
740         else
741                 kcd_size = sizeof(*kcd);
742
743         OBD_ALLOC(kcd, kcd_size);
744         if (kcd == NULL)
745                 GOTO(err_fput, rc = -ENOMEM);
746
747         kcd->kcd_nr_archives = lk->lk_data_count;
748         if (lk->lk_flags & LK_FLG_DATANR) {
749                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
750                 if (lk->lk_data_count > 0)
751                         memcpy(kcd->kcd_archives, lk->lk_data,
752                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
753         } else {
754                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
755         }
756
757         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
758                                    lk->lk_group, kcd, kcd_size);
759         OBD_FREE(kcd, kcd_size);
760         if (rc)
761                 GOTO(err_fput, rc);
762
763         /* All or nothing: try to register to all MDS.
764          * In case of failure, unregister from previous MDS,
765          * except if it because of inactive target. */
766         lmv_foreach_connected_tgt(lmv, tgt) {
767                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
768                 if (err) {
769                         if (tgt->ltd_active) {
770                                 /* permanent error */
771                                 CERROR("%s: iocontrol MDC %s on MDT"
772                                        " idx %d cmd %x: err = %d\n",
773                                        lmv2obd_dev(lmv)->obd_name,
774                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
775                                        err);
776                                 rc = err;
777                                 lk->lk_flags |= LK_FLG_STOP;
778                                 i = tgt->ltd_index;
779                                 /* unregister from previous MDS */
780                                 lmv_foreach_connected_tgt(lmv, tgt) {
781                                         if (tgt->ltd_index >= i)
782                                                 break;
783
784                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
785                                                       lk, uarg);
786                                 }
787                                 GOTO(err_kkuc_rem, rc);
788                         }
789                         /* else: transient error.
790                          * kuc will register to the missing MDT
791                          * when it is back */
792                 } else {
793                         any_set = true;
794                 }
795         }
796
797         if (!any_set)
798                 /* no registration done: return error */
799                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
800
801         RETURN(0);
802
803 err_kkuc_rem:
804         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
805
806 err_fput:
807         fput(filp);
808         return rc;
809 }
810
811 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
812                          int len, void *karg, void __user *uarg)
813 {
814         struct obd_device *obd = class_exp2obd(exp);
815         struct lmv_obd *lmv = &obd->u.lmv;
816         struct lu_tgt_desc *tgt = NULL;
817         int set = 0;
818         __u32 count = lmv->lmv_mdt_count;
819         int rc = 0;
820
821         ENTRY;
822
823         if (count == 0)
824                 RETURN(-ENOTTY);
825
826         switch (cmd) {
827         case IOC_OBD_STATFS: {
828                 struct obd_ioctl_data *data = karg;
829                 struct obd_device *mdc_obd;
830                 struct obd_statfs stat_buf = {0};
831                 __u32 index;
832
833                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
834
835                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
836                         RETURN(-ENODEV);
837
838                 tgt = lmv_tgt(lmv, index);
839                 if (!tgt)
840                         RETURN(-EAGAIN);
841
842                 if (!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(obd, cmd, len, lk, uarg);
1026                 else
1027                         rc = lmv_hsm_ct_register(obd, 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 = obd->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 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1059                   struct lu_fid *fid, struct md_op_data *op_data)
1060 {
1061         struct obd_device *obd = class_exp2obd(exp);
1062         struct lmv_obd *lmv = &obd->u.lmv;
1063         struct lmv_tgt_desc *tgt;
1064         int rc;
1065
1066         ENTRY;
1067
1068         LASSERT(op_data);
1069         LASSERT(fid);
1070
1071         tgt = lmv_tgt(lmv, op_data->op_mds);
1072         if (!tgt)
1073                 RETURN(-ENODEV);
1074
1075         if (!tgt->ltd_active || !tgt->ltd_exp)
1076                 RETURN(-ENODEV);
1077
1078         /*
1079          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1080          * on server that seq in new allocated fid is not yet known.
1081          */
1082         mutex_lock(&tgt->ltd_fid_mutex);
1083         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1084         mutex_unlock(&tgt->ltd_fid_mutex);
1085         if (rc > 0) {
1086                 LASSERT(fid_is_sane(fid));
1087                 rc = 0;
1088         }
1089
1090         RETURN(rc);
1091 }
1092
1093 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1094 {
1095         struct lmv_obd *lmv = &obd->u.lmv;
1096         struct lmv_desc *desc;
1097         struct lnet_process_id lnet_id;
1098         int i = 0;
1099         int rc;
1100
1101         ENTRY;
1102
1103         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1104                 CERROR("LMV setup requires a descriptor\n");
1105                 RETURN(-EINVAL);
1106         }
1107
1108         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1109         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1110                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1111                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1112                 RETURN(-EINVAL);
1113         }
1114
1115         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1116                      desc->ld_uuid.uuid);
1117         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1118         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1119         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1120                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1121         lmv->max_def_easize = 0;
1122         lmv->max_easize = 0;
1123
1124         spin_lock_init(&lmv->lmv_lock);
1125
1126         /*
1127          * initialize rr_index to lower 32bit of netid, so that client
1128          * can distribute subdirs evenly from the beginning.
1129          */
1130         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1131                 if (lnet_id.nid != LNET_NID_LO_0) {
1132                         lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1133                         break;
1134                 }
1135         }
1136
1137         rc = lmv_tunables_init(obd);
1138         if (rc)
1139                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1140                       obd->obd_name, rc);
1141
1142         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1143                              LUSTRE_CLI_FLD_HASH_DHT);
1144         if (rc)
1145                 CERROR("Can't init FLD, err %d\n", rc);
1146
1147         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1148         if (rc)
1149                 CWARN("%s: error initialize target table: rc = %d\n",
1150                       obd->obd_name, rc);
1151
1152         RETURN(rc);
1153 }
1154
1155 static int lmv_cleanup(struct obd_device *obd)
1156 {
1157         struct lmv_obd *lmv = &obd->u.lmv;
1158         struct lu_tgt_desc *tgt;
1159         struct lu_tgt_desc *tmp;
1160
1161         ENTRY;
1162
1163         fld_client_fini(&lmv->lmv_fld);
1164         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1165                 lmv_del_target(lmv, tgt);
1166         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1167
1168         RETURN(0);
1169 }
1170
1171 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1172 {
1173         struct lustre_cfg       *lcfg = buf;
1174         struct obd_uuid         obd_uuid;
1175         int                     gen;
1176         __u32                   index;
1177         int                     rc;
1178         ENTRY;
1179
1180         switch (lcfg->lcfg_command) {
1181         case LCFG_ADD_MDC:
1182                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1183                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1184                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1185                         GOTO(out, rc = -EINVAL);
1186
1187                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1188
1189                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1190                         GOTO(out, rc = -EINVAL);
1191                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1192                         GOTO(out, rc = -EINVAL);
1193                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1194                 GOTO(out, rc);
1195         default:
1196                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1197                 GOTO(out, rc = -EINVAL);
1198         }
1199 out:
1200         RETURN(rc);
1201 }
1202
1203 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1204 {
1205         int i;
1206
1207         if (flags & OBD_STATFS_FOR_MDT0)
1208                 return 0;
1209
1210         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1211                 return lmv->lmv_statfs_start;
1212
1213         /* choose initial MDT for this client */
1214         for (i = 0;; i++) {
1215                 struct lnet_process_id lnet_id;
1216                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1217                         break;
1218
1219                 if (lnet_id.nid != LNET_NID_LO_0) {
1220                         /* We dont need a full 64-bit modulus, just enough
1221                          * to distribute the requests across MDTs evenly.
1222                          */
1223                         lmv->lmv_statfs_start = (u32)lnet_id.nid %
1224                                                 lmv->lmv_mdt_count;
1225                         break;
1226                 }
1227         }
1228
1229         return lmv->lmv_statfs_start;
1230 }
1231
1232 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1233                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1234 {
1235         struct obd_device *obd = class_exp2obd(exp);
1236         struct lmv_obd *lmv = &obd->u.lmv;
1237         struct obd_statfs *temp;
1238         struct lu_tgt_desc *tgt;
1239         __u32 i;
1240         __u32 idx;
1241         int rc = 0;
1242
1243         ENTRY;
1244
1245         OBD_ALLOC(temp, sizeof(*temp));
1246         if (temp == NULL)
1247                 RETURN(-ENOMEM);
1248
1249         /* distribute statfs among MDTs */
1250         idx = lmv_select_statfs_mdt(lmv, flags);
1251
1252         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1253                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1254                 tgt = lmv_tgt(lmv, idx);
1255                 if (!tgt || !tgt->ltd_exp)
1256                         continue;
1257
1258                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1259                                 flags | OBD_STATFS_NESTED);
1260                 if (rc) {
1261                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1262                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1263                         GOTO(out_free_temp, rc);
1264                 }
1265
1266                 if (temp->os_state & OS_STATFS_SUM ||
1267                     flags == OBD_STATFS_FOR_MDT0) {
1268                         /* reset to the last aggregated values
1269                          * and don't sum with non-aggrated data */
1270                         /* If the statfs is from mount, it needs to retrieve
1271                          * necessary information from MDT0. i.e. mount does
1272                          * not need the merged osfs from all of MDT. Also
1273                          * clients can be mounted as long as MDT0 is in
1274                          * service */
1275                         *osfs = *temp;
1276                         break;
1277                 }
1278
1279                 if (i == 0) {
1280                         *osfs = *temp;
1281                 } else {
1282                         osfs->os_bavail += temp->os_bavail;
1283                         osfs->os_blocks += temp->os_blocks;
1284                         osfs->os_ffree += temp->os_ffree;
1285                         osfs->os_files += temp->os_files;
1286                         osfs->os_granted += temp->os_granted;
1287                 }
1288         }
1289
1290         EXIT;
1291 out_free_temp:
1292         OBD_FREE(temp, sizeof(*temp));
1293         return rc;
1294 }
1295
1296 static int lmv_statfs_update(void *cookie, int rc)
1297 {
1298         struct obd_info *oinfo = cookie;
1299         struct obd_device *obd = oinfo->oi_obd;
1300         struct lmv_obd *lmv = &obd->u.lmv;
1301         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1302         struct obd_statfs *osfs = oinfo->oi_osfs;
1303
1304         /*
1305          * NB: don't deactivate TGT upon error, because we may not trigger async
1306          * statfs any longer, then there is no chance to activate TGT.
1307          */
1308         if (!rc) {
1309                 spin_lock(&lmv->lmv_lock);
1310                 tgt->ltd_statfs = *osfs;
1311                 tgt->ltd_statfs_age = ktime_get_seconds();
1312                 spin_unlock(&lmv->lmv_lock);
1313                 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1314         }
1315
1316         return rc;
1317 }
1318
1319 /* update tgt statfs async if it's ld_qos_maxage old */
1320 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1321 {
1322         struct obd_info oinfo = {
1323                 .oi_obd = obd,
1324                 .oi_tgt = tgt,
1325                 .oi_cb_up = lmv_statfs_update,
1326         };
1327         int rc;
1328
1329         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1330             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1331                 return 0;
1332
1333         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1334
1335         return rc;
1336 }
1337
1338 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1339                         struct lu_fid *fid)
1340 {
1341         struct obd_device *obd = exp->exp_obd;
1342         struct lmv_obd *lmv = &obd->u.lmv;
1343         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1344         int rc;
1345
1346         ENTRY;
1347
1348         if (!tgt)
1349                 RETURN(-ENODEV);
1350
1351         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1352         RETURN(rc);
1353 }
1354
1355 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1356                         u64 obd_md_valid, const char *name, size_t buf_size,
1357                         struct ptlrpc_request **req)
1358 {
1359         struct obd_device *obd = exp->exp_obd;
1360         struct lmv_obd *lmv = &obd->u.lmv;
1361         struct lmv_tgt_desc *tgt;
1362         int rc;
1363
1364         ENTRY;
1365
1366         tgt = lmv_fid2tgt(lmv, fid);
1367         if (IS_ERR(tgt))
1368                 RETURN(PTR_ERR(tgt));
1369
1370         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1371
1372         RETURN(rc);
1373 }
1374
1375 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1376                         u64 obd_md_valid, const char *name,
1377                         const void *value, size_t value_size,
1378                         unsigned int xattr_flags, u32 suppgid,
1379                         struct ptlrpc_request **req)
1380 {
1381         struct obd_device *obd = exp->exp_obd;
1382         struct lmv_obd *lmv = &obd->u.lmv;
1383         struct lmv_tgt_desc *tgt;
1384         int rc;
1385
1386         ENTRY;
1387
1388         tgt = lmv_fid2tgt(lmv, fid);
1389         if (IS_ERR(tgt))
1390                 RETURN(PTR_ERR(tgt));
1391
1392         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1393                          value, value_size, xattr_flags, suppgid, req);
1394
1395         RETURN(rc);
1396 }
1397
1398 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1399                        struct ptlrpc_request **request)
1400 {
1401         struct obd_device *obd = exp->exp_obd;
1402         struct lmv_obd *lmv = &obd->u.lmv;
1403         struct lmv_tgt_desc *tgt;
1404         int rc;
1405
1406         ENTRY;
1407
1408         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1409         if (IS_ERR(tgt))
1410                 RETURN(PTR_ERR(tgt));
1411
1412         if (op_data->op_flags & MF_GET_MDT_IDX) {
1413                 op_data->op_mds = tgt->ltd_index;
1414                 RETURN(0);
1415         }
1416
1417         rc = md_getattr(tgt->ltd_exp, op_data, request);
1418
1419         RETURN(rc);
1420 }
1421
1422 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1423 {
1424         struct obd_device *obd = exp->exp_obd;
1425         struct lmv_obd *lmv = &obd->u.lmv;
1426         struct lu_tgt_desc *tgt;
1427
1428         ENTRY;
1429
1430         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1431
1432         /*
1433          * With DNE every object can have two locks in different namespaces:
1434          * lookup lock in space of MDT storing direntry and update/open lock in
1435          * space of MDT storing inode.
1436          */
1437         lmv_foreach_connected_tgt(lmv, tgt)
1438                 md_null_inode(tgt->ltd_exp, fid);
1439
1440         RETURN(0);
1441 }
1442
1443 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1444                      struct md_open_data *mod, struct ptlrpc_request **request)
1445 {
1446         struct obd_device *obd = exp->exp_obd;
1447         struct lmv_obd *lmv = &obd->u.lmv;
1448         struct lmv_tgt_desc *tgt;
1449         int rc;
1450
1451         ENTRY;
1452
1453         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1454         if (IS_ERR(tgt))
1455                 RETURN(PTR_ERR(tgt));
1456
1457         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1458         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1459         RETURN(rc);
1460 }
1461
1462 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt,
1463                                               unsigned short dir_depth)
1464 {
1465         struct lu_tgt_desc *tgt, *cur = NULL;
1466         __u64 total_avail = 0;
1467         __u64 total_weight = 0;
1468         __u64 cur_weight = 0;
1469         int total_usable = 0;
1470         __u64 rand;
1471         int rc;
1472
1473         ENTRY;
1474
1475         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1476                 RETURN(ERR_PTR(-EAGAIN));
1477
1478         down_write(&lmv->lmv_qos.lq_rw_sem);
1479
1480         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1481                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1482
1483         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1484         if (rc)
1485                 GOTO(unlock, tgt = ERR_PTR(rc));
1486
1487         lmv_foreach_tgt(lmv, tgt) {
1488                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1489                         tgt->ltd_qos.ltq_usable = 0;
1490                         continue;
1491                 }
1492
1493                 tgt->ltd_qos.ltq_usable = 1;
1494                 lu_tgt_qos_weight_calc(tgt);
1495                 if (tgt->ltd_index == *mdt)
1496                         cur = tgt;
1497                 total_avail += tgt->ltd_qos.ltq_avail;
1498                 total_weight += tgt->ltd_qos.ltq_weight;
1499                 total_usable++;
1500         }
1501
1502         /* if current MDT has above-average space, within range of the QOS
1503          * threshold, stay on the same MDT to avoid creating needless remote
1504          * MDT directories. It's more likely for low level directories.
1505          */
1506         rand = total_avail * (256 - lmv->lmv_qos.lq_threshold_rr) /
1507                (total_usable * 256 * (1 + dir_depth / 4));
1508         if (cur && cur->ltd_qos.ltq_avail >= rand) {
1509                 tgt = cur;
1510                 GOTO(unlock, rc = 0);
1511         }
1512
1513         rand = lu_prandom_u64_max(total_weight);
1514
1515         lmv_foreach_connected_tgt(lmv, tgt) {
1516                 if (!tgt->ltd_qos.ltq_usable)
1517                         continue;
1518
1519                 cur_weight += tgt->ltd_qos.ltq_weight;
1520                 if (cur_weight < rand)
1521                         continue;
1522
1523                 *mdt = tgt->ltd_index;
1524                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1525                 GOTO(unlock, rc = 0);
1526         }
1527
1528         /* no proper target found */
1529         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1530 unlock:
1531         up_write(&lmv->lmv_qos.lq_rw_sem);
1532
1533         return tgt;
1534 }
1535
1536 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
1537 {
1538         struct lu_tgt_desc *tgt;
1539         int i;
1540         int index;
1541
1542         ENTRY;
1543
1544         spin_lock(&lmv->lmv_lock);
1545         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1546                 index = (i + lmv->lmv_qos_rr_index) %
1547                         lmv->lmv_mdt_descs.ltd_tgts_size;
1548                 tgt = lmv_tgt(lmv, index);
1549                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1550                         continue;
1551
1552                 *mdt = tgt->ltd_index;
1553                 lmv->lmv_qos_rr_index = (*mdt + 1) %
1554                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1555                 spin_unlock(&lmv->lmv_lock);
1556
1557                 RETURN(tgt);
1558         }
1559         spin_unlock(&lmv->lmv_lock);
1560
1561         RETURN(ERR_PTR(-ENODEV));
1562 }
1563
1564 /* locate MDT by file name, for striped directory, the file name hash decides
1565  * which stripe its dirent is stored.
1566  */
1567 static struct lmv_tgt_desc *
1568 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1569                        const char *name, int namelen, struct lu_fid *fid,
1570                        __u32 *mds, bool new_layout)
1571 {
1572         struct lmv_tgt_desc *tgt;
1573         const struct lmv_oinfo *oinfo;
1574
1575         if (!lmv_dir_striped(lsm) || !namelen) {
1576                 tgt = lmv_fid2tgt(lmv, fid);
1577                 if (IS_ERR(tgt))
1578                         return tgt;
1579
1580                 *mds = tgt->ltd_index;
1581                 return tgt;
1582         }
1583
1584         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1585                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1586                         return ERR_PTR(-EBADF);
1587                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1588         } else {
1589                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1590                 if (IS_ERR(oinfo))
1591                         return ERR_CAST(oinfo);
1592         }
1593
1594         /* check stripe FID is sane */
1595         if (!fid_is_sane(&oinfo->lmo_fid))
1596                 return ERR_PTR(-ENODEV);
1597
1598         *fid = oinfo->lmo_fid;
1599         *mds = oinfo->lmo_mds;
1600         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1601
1602         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1603
1604         return tgt ? tgt : ERR_PTR(-ENODEV);
1605 }
1606
1607 /**
1608  * Locate MDT of op_data->op_fid1
1609  *
1610  * For striped directory, it will locate the stripe by name hash, if hash_type
1611  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1612  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1613  * indicates whether old or new layout is used to locate.
1614  *
1615  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1616  *
1617  * \param[in] lmv               LMV device
1618  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1619  *                              op_mds and op_fid1 will be updated if op_mea1
1620  *                              indicates fid1 represents a striped directory.
1621  *
1622  * retval               pointer to the lmv_tgt_desc if succeed.
1623  *                      ERR_PTR(errno) if failed.
1624  */
1625 struct lmv_tgt_desc *
1626 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1627 {
1628         struct lmv_stripe_md *lsm = op_data->op_mea1;
1629         struct lmv_oinfo *oinfo;
1630         struct lmv_tgt_desc *tgt;
1631
1632         if (lmv_dir_foreign(lsm))
1633                 return ERR_PTR(-ENODATA);
1634
1635         /* During creating VOLATILE file, it should honor the mdt
1636          * index if the file under striped dir is being restored, see
1637          * ct_restore(). */
1638         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1639             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1640                 tgt = lmv_tgt(lmv, op_data->op_mds);
1641                 if (!tgt)
1642                         return ERR_PTR(-ENODEV);
1643
1644                 if (lmv_dir_striped(lsm)) {
1645                         int i;
1646
1647                         /* refill the right parent fid */
1648                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1649                                 oinfo = &lsm->lsm_md_oinfo[i];
1650                                 if (oinfo->lmo_mds == op_data->op_mds) {
1651                                         op_data->op_fid1 = oinfo->lmo_fid;
1652                                         break;
1653                                 }
1654                         }
1655
1656                         if (i == lsm->lsm_md_stripe_count)
1657                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1658                 }
1659         } else if (lmv_dir_bad_hash(lsm)) {
1660                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1661                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1662
1663                 op_data->op_fid1 = oinfo->lmo_fid;
1664                 op_data->op_mds = oinfo->lmo_mds;
1665                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1666                 if (!tgt)
1667                         return ERR_PTR(-ENODEV);
1668         } else {
1669                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1670                                 op_data->op_name, op_data->op_namelen,
1671                                 &op_data->op_fid1, &op_data->op_mds,
1672                                 op_data->op_new_layout);
1673         }
1674
1675         return tgt;
1676 }
1677
1678 /* Locate MDT of op_data->op_fid2 for link/rename */
1679 static struct lmv_tgt_desc *
1680 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1681 {
1682         struct lmv_tgt_desc *tgt;
1683         int rc;
1684
1685         LASSERT(op_data->op_name);
1686         if (lmv_dir_layout_changing(op_data->op_mea2)) {
1687                 struct lu_fid fid1 = op_data->op_fid1;
1688                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1689                 struct ptlrpc_request *request = NULL;
1690
1691                 /*
1692                  * avoid creating new file under old layout of migrating
1693                  * directory, check it here.
1694                  */
1695                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1696                                 op_data->op_name, op_data->op_namelen,
1697                                 &op_data->op_fid2, &op_data->op_mds, false);
1698                 if (IS_ERR(tgt))
1699                         RETURN(tgt);
1700
1701                 op_data->op_fid1 = op_data->op_fid2;
1702                 op_data->op_mea1 = op_data->op_mea2;
1703                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1704                 op_data->op_fid1 = fid1;
1705                 op_data->op_mea1 = lsm1;
1706                 if (!rc) {
1707                         ptlrpc_req_finished(request);
1708                         RETURN(ERR_PTR(-EEXIST));
1709                 }
1710
1711                 if (rc != -ENOENT)
1712                         RETURN(ERR_PTR(rc));
1713         }
1714
1715         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1716                                 op_data->op_namelen, &op_data->op_fid2,
1717                                 &op_data->op_mds, true);
1718 }
1719
1720 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1721 {
1722         struct lu_tgt_desc *tgt;
1723         struct ptlrpc_request *request;
1724         int rc;
1725
1726         LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1727         LASSERT(!op_data->op_new_layout);
1728
1729         tgt = lmv_locate_tgt(lmv, op_data);
1730         if (IS_ERR(tgt))
1731                 return PTR_ERR(tgt);
1732
1733         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1734         if (!rc) {
1735                 ptlrpc_req_finished(request);
1736                 return -EEXIST;
1737         }
1738
1739         return rc;
1740 }
1741
1742 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1743 {
1744         const struct lmv_user_md *lum = op_data->op_data;
1745
1746         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1747                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1748                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1749 }
1750
1751 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1752 {
1753         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1754
1755         return (op_data->op_flags & MF_QOS_MKDIR) ||
1756                (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT);
1757 }
1758
1759 /* mkdir by QoS in three cases:
1760  * 1. ROOT default LMV is space balanced.
1761  * 2. 'lfs mkdir -i -1'
1762  * 3. parent default LMV master_mdt_index is -1
1763  *
1764  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1765  * directories under striped directory.
1766  */
1767 static inline bool lmv_op_qos_mkdir(const struct md_op_data *op_data)
1768 {
1769         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1770                 return false;
1771
1772         if (lmv_dir_striped(op_data->op_mea1))
1773                 return false;
1774
1775         if (lmv_op_user_qos_mkdir(op_data))
1776                 return true;
1777
1778         if (lmv_op_default_qos_mkdir(op_data))
1779                 return true;
1780
1781         return false;
1782 }
1783
1784 /* if parent default LMV is space balanced, and
1785  * 1. max_inherit_rr is set
1786  * 2. or parent is ROOT
1787  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1788  * LMV requests roundrobin mkdir, do the same.
1789  * NB, this needs to check server is balanced, which is done by caller.
1790  */
1791 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1792 {
1793         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1794
1795         if (!lmv_op_default_qos_mkdir(op_data))
1796                 return false;
1797
1798         return (op_data->op_flags & MF_RR_MKDIR) ||
1799                (lsm && lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE) ||
1800                fid_is_root(&op_data->op_fid1);
1801 }
1802
1803 /* 'lfs mkdir -i <specific_MDT>' */
1804 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1805 {
1806         const struct lmv_user_md *lum = op_data->op_data;
1807
1808         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1809                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1810                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1811                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1812                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1813 }
1814
1815 /* parent default LMV master_mdt_index is not -1. */
1816 static inline bool
1817 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1818 {
1819         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1820                op_data->op_default_mea1 &&
1821                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1822                         LMV_OFFSET_DEFAULT;
1823 }
1824
1825 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1826                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1827                 gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
1828                 struct ptlrpc_request **request)
1829 {
1830         struct obd_device *obd = exp->exp_obd;
1831         struct lmv_obd *lmv = &obd->u.lmv;
1832         struct lmv_tgt_desc *tgt;
1833         struct mdt_body *repbody;
1834         int rc;
1835
1836         ENTRY;
1837
1838         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1839                 RETURN(-EIO);
1840
1841         if (lmv_dir_bad_hash(op_data->op_mea1))
1842                 RETURN(-EBADF);
1843
1844         if (lmv_dir_layout_changing(op_data->op_mea1)) {
1845                 /*
1846                  * if parent is migrating, create() needs to lookup existing
1847                  * name in both old and new layout, check old layout on client.
1848                  */
1849                 rc = lmv_old_layout_lookup(lmv, op_data);
1850                 if (rc != -ENOENT)
1851                         RETURN(rc);
1852
1853                 op_data->op_new_layout = true;
1854         }
1855
1856         tgt = lmv_locate_tgt(lmv, op_data);
1857         if (IS_ERR(tgt))
1858                 RETURN(PTR_ERR(tgt));
1859
1860         if (lmv_op_user_specific_mkdir(op_data)) {
1861                 struct lmv_user_md *lum = op_data->op_data;
1862
1863                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1864                 tgt = lmv_tgt(lmv, op_data->op_mds);
1865                 if (!tgt)
1866                         RETURN(-ENODEV);
1867         } else if (lmv_op_default_specific_mkdir(op_data)) {
1868                 op_data->op_mds =
1869                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1870                 tgt = lmv_tgt(lmv, op_data->op_mds);
1871                 if (!tgt)
1872                         RETURN(-ENODEV);
1873         } else if (lmv_op_qos_mkdir(op_data)) {
1874                 struct lmv_tgt_desc *tmp = tgt;
1875
1876                 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds,
1877                                          op_data->op_dir_depth);
1878                 if (tgt == ERR_PTR(-EAGAIN)) {
1879                         if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1880                             !lmv_op_default_rr_mkdir(op_data) &&
1881                             !lmv_op_user_qos_mkdir(op_data))
1882                                 /* if it's not necessary, don't create remote
1883                                  * directory.
1884                                  */
1885                                 tgt = tmp;
1886                         else
1887                                 tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1888                 }
1889                 if (IS_ERR(tgt))
1890                         RETURN(PTR_ERR(tgt));
1891
1892                 /*
1893                  * only update statfs after QoS mkdir, this means the cached
1894                  * statfs may be stale, and current mkdir may not follow QoS
1895                  * accurately, but it's not serious, and avoids periodic statfs
1896                  * when client doesn't mkdir by QoS.
1897                  */
1898                 lmv_statfs_check_update(obd, tgt);
1899         }
1900
1901 retry:
1902         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1903         if (rc)
1904                 RETURN(rc);
1905
1906         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1907                 (int)op_data->op_namelen, op_data->op_name,
1908                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1909                 op_data->op_mds);
1910
1911         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1912         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1913                        cap_effective, rdev, request);
1914         if (rc == 0) {
1915                 if (*request == NULL)
1916                         RETURN(rc);
1917                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1918         }
1919
1920         /* dir restripe needs to send to MDT where dir is located */
1921         if (rc != -EREMOTE ||
1922             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
1923                 RETURN(rc);
1924
1925         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
1926         if (repbody == NULL)
1927                 RETURN(-EPROTO);
1928
1929         /* Not cross-ref case, just get out of here. */
1930         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
1931                 RETURN(rc);
1932
1933         op_data->op_fid2 = repbody->mbo_fid1;
1934         ptlrpc_req_finished(*request);
1935         *request = NULL;
1936
1937         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
1938         if (IS_ERR(tgt))
1939                 RETURN(PTR_ERR(tgt));
1940
1941         op_data->op_mds = tgt->ltd_index;
1942         goto retry;
1943 }
1944
1945 static int
1946 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1947             const union ldlm_policy_data *policy, struct md_op_data *op_data,
1948             struct lustre_handle *lockh, __u64 extra_lock_flags)
1949 {
1950         struct obd_device *obd = exp->exp_obd;
1951         struct lmv_obd *lmv = &obd->u.lmv;
1952         struct lmv_tgt_desc *tgt;
1953         int rc;
1954
1955         ENTRY;
1956
1957         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1958
1959         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1960         if (IS_ERR(tgt))
1961                 RETURN(PTR_ERR(tgt));
1962
1963         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1964                PFID(&op_data->op_fid1), tgt->ltd_index);
1965
1966         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1967                         extra_lock_flags);
1968
1969         RETURN(rc);
1970 }
1971
1972 int
1973 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1974                  struct ptlrpc_request **preq)
1975 {
1976         struct obd_device *obd = exp->exp_obd;
1977         struct lmv_obd *lmv = &obd->u.lmv;
1978         struct lmv_tgt_desc *tgt;
1979         struct mdt_body *body;
1980         int rc;
1981
1982         ENTRY;
1983
1984 retry:
1985         if (op_data->op_namelen == 2 &&
1986             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
1987                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1988         else
1989                 tgt = lmv_locate_tgt(lmv, op_data);
1990         if (IS_ERR(tgt))
1991                 RETURN(PTR_ERR(tgt));
1992
1993         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1994                 (int)op_data->op_namelen, op_data->op_name,
1995                 PFID(&op_data->op_fid1), tgt->ltd_index);
1996
1997         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1998         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1999                 ptlrpc_req_finished(*preq);
2000                 *preq = NULL;
2001                 goto retry;
2002         }
2003
2004         if (rc)
2005                 RETURN(rc);
2006
2007         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2008         LASSERT(body != NULL);
2009
2010         if (body->mbo_valid & OBD_MD_MDS) {
2011                 op_data->op_fid1 = body->mbo_fid1;
2012                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2013                 op_data->op_namelen = 0;
2014                 op_data->op_name = NULL;
2015
2016                 ptlrpc_req_finished(*preq);
2017                 *preq = NULL;
2018
2019                 goto retry;
2020         }
2021
2022         RETURN(rc);
2023 }
2024
2025 #define md_op_data_fid(op_data, fl)                     \
2026         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2027          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2028          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2029          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2030          NULL)
2031
2032 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2033                             struct md_op_data *op_data, __u32 op_tgt,
2034                             enum ldlm_mode mode, int bits, int flag)
2035 {
2036         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2037         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2038         union ldlm_policy_data policy = { { 0 } };
2039         int rc = 0;
2040         ENTRY;
2041
2042         if (!fid_is_sane(fid))
2043                 RETURN(0);
2044
2045         if (tgt == NULL) {
2046                 tgt = lmv_fid2tgt(lmv, fid);
2047                 if (IS_ERR(tgt))
2048                         RETURN(PTR_ERR(tgt));
2049         }
2050
2051         if (tgt->ltd_index != op_tgt) {
2052                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2053                 policy.l_inodebits.bits = bits;
2054                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2055                                       mode, LCF_ASYNC, NULL);
2056         } else {
2057                 CDEBUG(D_INODE,
2058                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2059                        op_tgt, PFID(fid));
2060                 op_data->op_flags |= flag;
2061                 rc = 0;
2062         }
2063
2064         RETURN(rc);
2065 }
2066
2067 /*
2068  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2069  * op_data->op_fid2
2070  */
2071 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2072                     struct ptlrpc_request **request)
2073 {
2074         struct obd_device       *obd = exp->exp_obd;
2075         struct lmv_obd          *lmv = &obd->u.lmv;
2076         struct lmv_tgt_desc     *tgt;
2077         int                      rc;
2078         ENTRY;
2079
2080         LASSERT(op_data->op_namelen != 0);
2081
2082         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2083                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2084                op_data->op_name, PFID(&op_data->op_fid1));
2085
2086         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2087         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2088         op_data->op_cap = current_cap();
2089
2090         tgt = lmv_locate_tgt2(lmv, op_data);
2091         if (IS_ERR(tgt))
2092                 RETURN(PTR_ERR(tgt));
2093
2094         /*
2095          * Cancel UPDATE lock on child (fid1).
2096          */
2097         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2098         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2099                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2100         if (rc != 0)
2101                 RETURN(rc);
2102
2103         rc = md_link(tgt->ltd_exp, op_data, request);
2104
2105         RETURN(rc);
2106 }
2107
2108 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2109                         const char *name, size_t namelen,
2110                         struct ptlrpc_request **request)
2111 {
2112         struct obd_device *obd = exp->exp_obd;
2113         struct lmv_obd *lmv = &obd->u.lmv;
2114         struct lmv_stripe_md *lsm = op_data->op_mea1;
2115         struct lmv_tgt_desc *parent_tgt;
2116         struct lmv_tgt_desc *sp_tgt;
2117         struct lmv_tgt_desc *tp_tgt = NULL;
2118         struct lmv_tgt_desc *child_tgt;
2119         struct lmv_tgt_desc *tgt;
2120         struct lu_fid target_fid = { 0 };
2121         int rc;
2122
2123         ENTRY;
2124
2125         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2126
2127         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2128                PFID(&op_data->op_fid1), (int)namelen, name);
2129
2130         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2131         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2132         op_data->op_cap = current_cap();
2133
2134         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2135         if (IS_ERR(parent_tgt))
2136                 RETURN(PTR_ERR(parent_tgt));
2137
2138         if (lmv_dir_striped(lsm)) {
2139                 const struct lmv_oinfo *oinfo;
2140
2141                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2142                 if (IS_ERR(oinfo))
2143                         RETURN(PTR_ERR(oinfo));
2144
2145                 /* save source stripe FID in fid4 temporarily for ELC */
2146                 op_data->op_fid4 = oinfo->lmo_fid;
2147                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2148                 if (!sp_tgt)
2149                         RETURN(-ENODEV);
2150
2151                 /*
2152                  * if parent is being migrated too, fill op_fid2 with target
2153                  * stripe fid, otherwise the target stripe is not created yet.
2154                  */
2155                 if (lmv_dir_layout_changing(lsm)) {
2156                         oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2157                                                         true);
2158                         if (IS_ERR(oinfo))
2159                                 RETURN(PTR_ERR(oinfo));
2160
2161                         op_data->op_fid2 = oinfo->lmo_fid;
2162                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2163                         if (!tp_tgt)
2164                                 RETURN(-ENODEV);
2165                 }
2166         } else {
2167                 sp_tgt = parent_tgt;
2168         }
2169
2170         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2171         if (IS_ERR(child_tgt))
2172                 RETURN(PTR_ERR(child_tgt));
2173
2174         /* for directory, migrate to MDT specified by lum_stripe_offset;
2175          * otherwise migrate to the target stripe of parent, but parent
2176          * directory may have finished migration (normally current file too),
2177          * allocate FID on MDT lum_stripe_offset, and server will check
2178          * whether file was migrated already.
2179          */
2180         if (S_ISDIR(op_data->op_mode) || !tp_tgt) {
2181                 struct lmv_user_md *lum = op_data->op_data;
2182
2183                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2184         } else  {
2185                 op_data->op_mds = tp_tgt->ltd_index;
2186         }
2187         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2188         if (rc)
2189                 RETURN(rc);
2190
2191         /*
2192          * for directory, send migrate request to the MDT where the object will
2193          * be migrated to, because we can't create a striped directory remotely.
2194          *
2195          * otherwise, send to the MDT where source is located because regular
2196          * file may open lease.
2197          *
2198          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2199          * backward compatibility.
2200          */
2201         if (S_ISDIR(op_data->op_mode) &&
2202             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2203                 tgt = lmv_fid2tgt(lmv, &target_fid);
2204                 if (IS_ERR(tgt))
2205                         RETURN(PTR_ERR(tgt));
2206         } else {
2207                 tgt = child_tgt;
2208         }
2209
2210         /* cancel UPDATE lock of parent master object */
2211         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2212                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2213         if (rc)
2214                 RETURN(rc);
2215
2216         /* cancel UPDATE lock of source parent */
2217         if (sp_tgt != parent_tgt) {
2218                 /*
2219                  * migrate RPC packs master object FID, because we can only pack
2220                  * two FIDs in reint RPC, but MDS needs to know both source
2221                  * parent and target parent, and it will obtain them from master
2222                  * FID and LMV, the other FID in RPC is kept for target.
2223                  *
2224                  * since this FID is not passed to MDC, cancel it anyway.
2225                  */
2226                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2227                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2228                 if (rc)
2229                         RETURN(rc);
2230
2231                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2232         }
2233         op_data->op_fid4 = target_fid;
2234
2235         /* cancel UPDATE locks of target parent */
2236         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2237                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2238         if (rc)
2239                 RETURN(rc);
2240
2241         /* cancel LOOKUP lock of source if source is remote object */
2242         if (child_tgt != sp_tgt) {
2243                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2244                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2245                                       MF_MDC_CANCEL_FID3);
2246                 if (rc)
2247                         RETURN(rc);
2248         }
2249
2250         /* cancel ELC locks of source */
2251         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2252                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2253         if (rc)
2254                 RETURN(rc);
2255
2256         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2257
2258         RETURN(rc);
2259 }
2260
2261 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2262                       const char *old, size_t oldlen,
2263                       const char *new, size_t newlen,
2264                       struct ptlrpc_request **request)
2265 {
2266         struct obd_device *obd = exp->exp_obd;
2267         struct lmv_obd *lmv = &obd->u.lmv;
2268         struct lmv_tgt_desc *sp_tgt;
2269         struct lmv_tgt_desc *tp_tgt = NULL;
2270         struct lmv_tgt_desc *src_tgt = NULL;
2271         struct lmv_tgt_desc *tgt;
2272         struct mdt_body *body;
2273         int rc;
2274
2275         ENTRY;
2276
2277         LASSERT(oldlen != 0);
2278
2279         if (op_data->op_cli_flags & CLI_MIGRATE) {
2280                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2281                 RETURN(rc);
2282         }
2283
2284         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2285         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2286         op_data->op_cap = current_cap();
2287
2288         op_data->op_name = new;
2289         op_data->op_namelen = newlen;
2290
2291         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2292         if (IS_ERR(tp_tgt))
2293                 RETURN(PTR_ERR(tp_tgt));
2294
2295         /* Since the target child might be destroyed, and it might become
2296          * orphan, and we can only check orphan on the local MDT right now, so
2297          * we send rename request to the MDT where target child is located. If
2298          * target child does not exist, then it will send the request to the
2299          * target parent */
2300         if (fid_is_sane(&op_data->op_fid4)) {
2301                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2302                 if (IS_ERR(tgt))
2303                         RETURN(PTR_ERR(tgt));
2304         } else {
2305                 tgt = tp_tgt;
2306         }
2307
2308         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2309
2310         /* cancel UPDATE locks of target parent */
2311         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2312                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2313         if (rc != 0)
2314                 RETURN(rc);
2315
2316         if (fid_is_sane(&op_data->op_fid4)) {
2317                 /* cancel LOOKUP lock of target on target parent */
2318                 if (tgt != tp_tgt) {
2319                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2320                                               tgt->ltd_index, LCK_EX,
2321                                               MDS_INODELOCK_LOOKUP,
2322                                               MF_MDC_CANCEL_FID4);
2323                         if (rc != 0)
2324                                 RETURN(rc);
2325                 }
2326         }
2327
2328         if (fid_is_sane(&op_data->op_fid3)) {
2329                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2330                 if (IS_ERR(src_tgt))
2331                         RETURN(PTR_ERR(src_tgt));
2332
2333                 /* cancel ELC locks of source */
2334                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2335                                       LCK_EX, MDS_INODELOCK_ELC,
2336                                       MF_MDC_CANCEL_FID3);
2337                 if (rc != 0)
2338                         RETURN(rc);
2339         }
2340
2341         op_data->op_name = old;
2342         op_data->op_namelen = oldlen;
2343 retry:
2344         sp_tgt = lmv_locate_tgt(lmv, op_data);
2345         if (IS_ERR(sp_tgt))
2346                 RETURN(PTR_ERR(sp_tgt));
2347
2348         /* cancel UPDATE locks of source parent */
2349         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2350                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2351         if (rc != 0)
2352                 RETURN(rc);
2353
2354         if (fid_is_sane(&op_data->op_fid3)) {
2355                 /* cancel LOOKUP lock of source on source parent */
2356                 if (src_tgt != sp_tgt) {
2357                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2358                                               tgt->ltd_index, LCK_EX,
2359                                               MDS_INODELOCK_LOOKUP,
2360                                               MF_MDC_CANCEL_FID3);
2361                         if (rc != 0)
2362                                 RETURN(rc);
2363                 }
2364         }
2365
2366 rename:
2367         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2368                 PFID(&op_data->op_fid1), (int)oldlen, old,
2369                 PFID(&op_data->op_fid2), (int)newlen, new);
2370
2371         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2372                         request);
2373         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2374                 ptlrpc_req_finished(*request);
2375                 *request = NULL;
2376                 goto retry;
2377         }
2378
2379         if (rc && rc != -EXDEV)
2380                 RETURN(rc);
2381
2382         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2383         if (body == NULL)
2384                 RETURN(-EPROTO);
2385
2386         /* Not cross-ref case, just get out of here. */
2387         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2388                 RETURN(rc);
2389
2390         op_data->op_fid4 = body->mbo_fid1;
2391
2392         ptlrpc_req_finished(*request);
2393         *request = NULL;
2394
2395         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2396         if (IS_ERR(tgt))
2397                 RETURN(PTR_ERR(tgt));
2398
2399         if (fid_is_sane(&op_data->op_fid4)) {
2400                 /* cancel LOOKUP lock of target on target parent */
2401                 if (tgt != tp_tgt) {
2402                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2403                                               tgt->ltd_index, LCK_EX,
2404                                               MDS_INODELOCK_LOOKUP,
2405                                               MF_MDC_CANCEL_FID4);
2406                         if (rc != 0)
2407                                 RETURN(rc);
2408                 }
2409         }
2410
2411         goto rename;
2412 }
2413
2414 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2415                        void *ea, size_t ealen, struct ptlrpc_request **request)
2416 {
2417         struct obd_device *obd = exp->exp_obd;
2418         struct lmv_obd *lmv = &obd->u.lmv;
2419         struct lmv_tgt_desc *tgt;
2420         int rc = 0;
2421
2422         ENTRY;
2423
2424         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2425                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2426                op_data->op_xvalid);
2427
2428         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2429         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2430         if (IS_ERR(tgt))
2431                 RETURN(PTR_ERR(tgt));
2432
2433         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2434
2435         RETURN(rc);
2436 }
2437
2438 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2439                      struct ptlrpc_request **request)
2440 {
2441         struct obd_device *obd = exp->exp_obd;
2442         struct lmv_obd *lmv = &obd->u.lmv;
2443         struct lmv_tgt_desc *tgt;
2444         int rc;
2445
2446         ENTRY;
2447
2448         tgt = lmv_fid2tgt(lmv, fid);
2449         if (IS_ERR(tgt))
2450                 RETURN(PTR_ERR(tgt));
2451
2452         rc = md_fsync(tgt->ltd_exp, fid, request);
2453         RETURN(rc);
2454 }
2455
2456 struct stripe_dirent {
2457         struct page             *sd_page;
2458         struct lu_dirpage       *sd_dp;
2459         struct lu_dirent        *sd_ent;
2460         bool                     sd_eof;
2461 };
2462
2463 struct lmv_dir_ctxt {
2464         struct lmv_obd          *ldc_lmv;
2465         struct md_op_data       *ldc_op_data;
2466         struct md_callback      *ldc_cb_op;
2467         __u64                    ldc_hash;
2468         int                      ldc_count;
2469         struct stripe_dirent     ldc_stripes[0];
2470 };
2471
2472 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2473 {
2474         if (stripe->sd_page) {
2475                 kunmap(stripe->sd_page);
2476                 put_page(stripe->sd_page);
2477                 stripe->sd_page = NULL;
2478                 stripe->sd_ent = NULL;
2479         }
2480 }
2481
2482 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2483 {
2484         int i;
2485
2486         for (i = 0; i < ctxt->ldc_count; i++)
2487                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2488 }
2489
2490 /* if @ent is dummy, or . .., get next */
2491 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2492                                            struct lu_dirent *ent,
2493                                            int stripe_index)
2494 {
2495         for (; ent; ent = lu_dirent_next(ent)) {
2496                 /* Skip dummy entry */
2497                 if (le16_to_cpu(ent->lde_namelen) == 0)
2498                         continue;
2499
2500                 /* skip . and .. for other stripes */
2501                 if (stripe_index &&
2502                     (strncmp(ent->lde_name, ".",
2503                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2504                      strncmp(ent->lde_name, "..",
2505                              le16_to_cpu(ent->lde_namelen)) == 0))
2506                         continue;
2507
2508                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2509                         break;
2510         }
2511
2512         return ent;
2513 }
2514
2515 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2516                                             struct stripe_dirent *stripe,
2517                                             int stripe_index)
2518 {
2519         struct md_op_data *op_data = ctxt->ldc_op_data;
2520         struct lmv_oinfo *oinfo;
2521         struct lu_fid fid = op_data->op_fid1;
2522         struct inode *inode = op_data->op_data;
2523         struct lmv_tgt_desc *tgt;
2524         struct lu_dirent *ent = stripe->sd_ent;
2525         __u64 hash = ctxt->ldc_hash;
2526         int rc = 0;
2527
2528         ENTRY;
2529
2530         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2531         LASSERT(!ent);
2532
2533         do {
2534                 if (stripe->sd_page) {
2535                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2536
2537                         /* @hash should be the last dirent hash */
2538                         LASSERTF(hash <= end,
2539                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2540                                  ctxt, stripe, hash, end);
2541                         /* unload last page */
2542                         stripe_dirent_unload(stripe);
2543                         /* eof */
2544                         if (end == MDS_DIR_END_OFF) {
2545                                 stripe->sd_eof = true;
2546                                 break;
2547                         }
2548                         hash = end;
2549                 }
2550
2551                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2552                 if (!oinfo->lmo_root) {
2553                         rc = -ENOENT;
2554                         break;
2555                 }
2556
2557                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2558                 if (!tgt) {
2559                         rc = -ENODEV;
2560                         break;
2561                 }
2562
2563                 /* op_data is shared by stripes, reset after use */
2564                 op_data->op_fid1 = oinfo->lmo_fid;
2565                 op_data->op_fid2 = oinfo->lmo_fid;
2566                 op_data->op_data = oinfo->lmo_root;
2567
2568                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2569                                   &stripe->sd_page);
2570
2571                 op_data->op_fid1 = fid;
2572                 op_data->op_fid2 = fid;
2573                 op_data->op_data = inode;
2574
2575                 if (rc)
2576                         break;
2577
2578                 stripe->sd_dp = page_address(stripe->sd_page);
2579                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2580                                         stripe_index);
2581                 /* in case a page filled with ., .. and dummy, read next */
2582         } while (!ent);
2583
2584         stripe->sd_ent = ent;
2585         if (rc) {
2586                 LASSERT(!ent);
2587                 /* treat error as eof, so dir can be partially accessed */
2588                 stripe->sd_eof = true;
2589                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2590                               "directory is partially accessed!\n",
2591                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2592                               rc);
2593         }
2594
2595         RETURN(ent);
2596 }
2597
2598 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2599 {
2600         struct obd_device *obd = exp->exp_obd;
2601         struct lmv_obd *lmv = &obd->u.lmv;
2602         struct lmv_tgt_desc *tgt;
2603         int rc;
2604
2605         ENTRY;
2606
2607         rc = lmv_check_connect(obd);
2608         if (rc != 0)
2609                 RETURN(rc);
2610
2611         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2612         if (IS_ERR(tgt))
2613                 RETURN(PTR_ERR(tgt));
2614
2615         data->op_flags |= MF_MDC_CANCEL_FID1;
2616         rc = md_file_resync(tgt->ltd_exp, data);
2617         RETURN(rc);
2618 }
2619
2620 /**
2621  * Get dirent with the closest hash for striped directory
2622  *
2623  * This function will search the dir entry, whose hash value is the
2624  * closest(>=) to hash from all of sub-stripes, and it is only being called
2625  * for striped directory.
2626  *
2627  * \param[in] ctxt              dir read context
2628  *
2629  * \retval                      dirent get the entry successfully
2630  *                              NULL does not get the entry, normally it means
2631  *                              it reaches the end of the directory, while read
2632  *                              stripe dirent error is ignored to allow partial
2633  *                              access.
2634  */
2635 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2636 {
2637         struct stripe_dirent *stripe;
2638         struct lu_dirent *ent = NULL;
2639         int i;
2640         int min = -1;
2641
2642         /* TODO: optimize with k-way merge sort */
2643         for (i = 0; i < ctxt->ldc_count; i++) {
2644                 stripe = &ctxt->ldc_stripes[i];
2645                 if (stripe->sd_eof)
2646                         continue;
2647
2648                 if (!stripe->sd_ent) {
2649                         stripe_dirent_load(ctxt, stripe, i);
2650                         if (!stripe->sd_ent) {
2651                                 LASSERT(stripe->sd_eof);
2652                                 continue;
2653                         }
2654                 }
2655
2656                 if (min == -1 ||
2657                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2658                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2659                         min = i;
2660                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2661                             ctxt->ldc_hash)
2662                                 break;
2663                 }
2664         }
2665
2666         if (min != -1) {
2667                 stripe = &ctxt->ldc_stripes[min];
2668                 ent = stripe->sd_ent;
2669                 /* pop found dirent */
2670                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2671                                                    min);
2672         }
2673
2674         return ent;
2675 }
2676
2677 /**
2678  * Build dir entry page for striped directory
2679  *
2680  * This function gets one entry by @offset from a striped directory. It will
2681  * read entries from all of stripes, and choose one closest to the required
2682  * offset(&offset). A few notes
2683  * 1. skip . and .. for non-zero stripes, because there can only have one .
2684  * and .. in a directory.
2685  * 2. op_data will be shared by all of stripes, instead of allocating new
2686  * one, so need to restore before reusing.
2687  *
2688  * \param[in] exp       obd export refer to LMV
2689  * \param[in] op_data   hold those MD parameters of read_entry
2690  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2691  * \param[in] offset    starting hash offset
2692  * \param[out] ppage    the page holding the entry. Note: because the entry
2693  *                      will be accessed in upper layer, so we need hold the
2694  *                      page until the usages of entry is finished, see
2695  *                      ll_dir_entry_next.
2696  *
2697  * retval               =0 if get entry successfully
2698  *                      <0 cannot get entry
2699  */
2700 static int lmv_striped_read_page(struct obd_export *exp,
2701                                  struct md_op_data *op_data,
2702                                  struct md_callback *cb_op,
2703                                  __u64 offset, struct page **ppage)
2704 {
2705         struct page *page = NULL;
2706         struct lu_dirpage *dp;
2707         void *start;
2708         struct lu_dirent *ent;
2709         struct lu_dirent *last_ent;
2710         int stripe_count;
2711         struct lmv_dir_ctxt *ctxt;
2712         struct lu_dirent *next = NULL;
2713         __u16 ent_size;
2714         size_t left_bytes;
2715         int rc = 0;
2716         ENTRY;
2717
2718         /* Allocate a page and read entries from all of stripes and fill
2719          * the page by hash order */
2720         page = alloc_page(GFP_KERNEL);
2721         if (!page)
2722                 RETURN(-ENOMEM);
2723
2724         /* Initialize the entry page */
2725         dp = kmap(page);
2726         memset(dp, 0, sizeof(*dp));
2727         dp->ldp_hash_start = cpu_to_le64(offset);
2728
2729         start = dp + 1;
2730         left_bytes = PAGE_SIZE - sizeof(*dp);
2731         ent = start;
2732         last_ent = ent;
2733
2734         /* initalize dir read context */
2735         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2736         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2737         if (!ctxt)
2738                 GOTO(free_page, rc = -ENOMEM);
2739         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2740         ctxt->ldc_op_data = op_data;
2741         ctxt->ldc_cb_op = cb_op;
2742         ctxt->ldc_hash = offset;
2743         ctxt->ldc_count = stripe_count;
2744
2745         while (1) {
2746                 next = lmv_dirent_next(ctxt);
2747
2748                 /* end of directory */
2749                 if (!next) {
2750                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2751                         break;
2752                 }
2753                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2754
2755                 ent_size = le16_to_cpu(next->lde_reclen);
2756
2757                 /* the last entry lde_reclen is 0, but it might not be the last
2758                  * one of this temporay dir page */
2759                 if (!ent_size)
2760                         ent_size = lu_dirent_calc_size(
2761                                         le16_to_cpu(next->lde_namelen),
2762                                         le32_to_cpu(next->lde_attrs));
2763                 /* page full */
2764                 if (ent_size > left_bytes)
2765                         break;
2766
2767                 memcpy(ent, next, ent_size);
2768
2769                 /* Replace . with master FID and Replace .. with the parent FID
2770                  * of master object */
2771                 if (strncmp(ent->lde_name, ".",
2772                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2773                     le16_to_cpu(ent->lde_namelen) == 1)
2774                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2775                 else if (strncmp(ent->lde_name, "..",
2776                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2777                            le16_to_cpu(ent->lde_namelen) == 2)
2778                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2779
2780                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2781                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2782                        le64_to_cpu(ent->lde_hash));
2783
2784                 left_bytes -= ent_size;
2785                 ent->lde_reclen = cpu_to_le16(ent_size);
2786                 last_ent = ent;
2787                 ent = (void *)ent + ent_size;
2788         };
2789
2790         last_ent->lde_reclen = 0;
2791
2792         if (ent == start)
2793                 dp->ldp_flags |= LDF_EMPTY;
2794         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2795                 dp->ldp_flags |= LDF_COLLIDE;
2796         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2797         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2798
2799         put_lmv_dir_ctxt(ctxt);
2800         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2801
2802         *ppage = page;
2803
2804         RETURN(0);
2805
2806 free_page:
2807         kunmap(page);
2808         __free_page(page);
2809
2810         return rc;
2811 }
2812
2813 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2814                          struct md_callback *cb_op, __u64 offset,
2815                          struct page **ppage)
2816 {
2817         struct obd_device *obd = exp->exp_obd;
2818         struct lmv_obd *lmv = &obd->u.lmv;
2819         struct lmv_tgt_desc *tgt;
2820         int rc;
2821
2822         ENTRY;
2823
2824         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2825                 RETURN(-ENODATA);
2826
2827         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2828                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2829                 RETURN(rc);
2830         }
2831
2832         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2833         if (IS_ERR(tgt))
2834                 RETURN(PTR_ERR(tgt));
2835
2836         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2837
2838         RETURN(rc);
2839 }
2840
2841 /**
2842  * Unlink a file/directory
2843  *
2844  * Unlink a file or directory under the parent dir. The unlink request
2845  * usually will be sent to the MDT where the child is located, but if
2846  * the client does not have the child FID then request will be sent to the
2847  * MDT where the parent is located.
2848  *
2849  * If the parent is a striped directory then it also needs to locate which
2850  * stripe the name of the child is located, and replace the parent FID
2851  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2852  * it will walk through all of sub-stripes until the child is being
2853  * unlinked finally.
2854  *
2855  * \param[in] exp       export refer to LMV
2856  * \param[in] op_data   different parameters transferred beween client
2857  *                      MD stacks, name, namelen, FIDs etc.
2858  *                      op_fid1 is the parent FID, op_fid2 is the child
2859  *                      FID.
2860  * \param[out] request  point to the request of unlink.
2861  *
2862  * retval               0 if succeed
2863  *                      negative errno if failed.
2864  */
2865 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2866                       struct ptlrpc_request **request)
2867 {
2868         struct obd_device *obd = exp->exp_obd;
2869         struct lmv_obd *lmv = &obd->u.lmv;
2870         struct lmv_tgt_desc *tgt;
2871         struct lmv_tgt_desc *parent_tgt;
2872         struct mdt_body *body;
2873         int rc;
2874
2875         ENTRY;
2876
2877         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2878         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2879         op_data->op_cap = current_cap();
2880
2881 retry:
2882         parent_tgt = lmv_locate_tgt(lmv, op_data);
2883         if (IS_ERR(parent_tgt))
2884                 RETURN(PTR_ERR(parent_tgt));
2885
2886         if (likely(!fid_is_zero(&op_data->op_fid2))) {
2887                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2888                 if (IS_ERR(tgt))
2889                         RETURN(PTR_ERR(tgt));
2890         } else {
2891                 tgt = parent_tgt;
2892         }
2893
2894         /*
2895          * If child's fid is given, cancel unused locks for it if it is from
2896          * another export than parent.
2897          *
2898          * LOOKUP lock for child (fid3) should also be cancelled on parent
2899          * tgt_tgt in mdc_unlink().
2900          */
2901         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2902
2903         if (parent_tgt != tgt)
2904                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2905                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2906                                       MF_MDC_CANCEL_FID3);
2907
2908         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2909                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2910         if (rc)
2911                 RETURN(rc);
2912
2913         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2914                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2915                tgt->ltd_index);
2916
2917         rc = md_unlink(tgt->ltd_exp, op_data, request);
2918         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2919                 ptlrpc_req_finished(*request);
2920                 *request = NULL;
2921                 goto retry;
2922         }
2923
2924         if (rc != -EREMOTE)
2925                 RETURN(rc);
2926
2927         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2928         if (body == NULL)
2929                 RETURN(-EPROTO);
2930
2931         /* Not cross-ref case, just get out of here. */
2932         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2933                 RETURN(rc);
2934
2935         /* This is a remote object, try remote MDT. */
2936         op_data->op_fid2 = body->mbo_fid1;
2937         ptlrpc_req_finished(*request);
2938         *request = NULL;
2939
2940         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2941         if (IS_ERR(tgt))
2942                 RETURN(PTR_ERR(tgt));
2943
2944         goto retry;
2945 }
2946
2947 static int lmv_precleanup(struct obd_device *obd)
2948 {
2949         ENTRY;
2950         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2951         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2952         lprocfs_obd_cleanup(obd);
2953         lprocfs_free_md_stats(obd);
2954         RETURN(0);
2955 }
2956
2957 /**
2958  * Get by key a value associated with a LMV device.
2959  *
2960  * Dispatch request to lower-layer devices as needed.
2961  *
2962  * \param[in] env               execution environment for this thread
2963  * \param[in] exp               export for the LMV device
2964  * \param[in] keylen            length of key identifier
2965  * \param[in] key               identifier of key to get value for
2966  * \param[in] vallen            size of \a val
2967  * \param[out] val              pointer to storage location for value
2968  * \param[in] lsm               optional striping metadata of object
2969  *
2970  * \retval 0            on success
2971  * \retval negative     negated errno on failure
2972  */
2973 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2974                         __u32 keylen, void *key, __u32 *vallen, void *val)
2975 {
2976         struct obd_device *obd;
2977         struct lmv_obd *lmv;
2978         struct lu_tgt_desc *tgt;
2979         int rc = 0;
2980
2981         ENTRY;
2982
2983         obd = class_exp2obd(exp);
2984         if (obd == NULL) {
2985                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2986                        exp->exp_handle.h_cookie);
2987                 RETURN(-EINVAL);
2988         }
2989
2990         lmv = &obd->u.lmv;
2991         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2992                 LASSERT(*vallen == sizeof(__u32));
2993                 lmv_foreach_connected_tgt(lmv, tgt) {
2994                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2995                                           vallen, val))
2996                                 RETURN(0);
2997                 }
2998                 RETURN(-EINVAL);
2999         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3000                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3001                    KEY_IS(KEY_CONN_DATA)) {
3002                 /*
3003                  * Forwarding this request to first MDS, it should know LOV
3004                  * desc.
3005                  */
3006                 tgt = lmv_tgt(lmv, 0);
3007                 if (!tgt)
3008                         RETURN(-ENODEV);
3009
3010                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3011                 if (!rc && KEY_IS(KEY_CONN_DATA))
3012                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3013                 RETURN(rc);
3014         } else if (KEY_IS(KEY_TGT_COUNT)) {
3015                 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
3016                 RETURN(0);
3017         }
3018
3019         CDEBUG(D_IOCTL, "Invalid key\n");
3020         RETURN(-EINVAL);
3021 }
3022
3023 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3024                      int *__rcs, struct ptlrpc_request_set *_set)
3025 {
3026         struct obd_device *obd = class_exp2obd(exp);
3027         struct ptlrpc_request_set *set = _set;
3028         struct lmv_obd *lmv = &obd->u.lmv;
3029         int tgt_count = lmv->lmv_mdt_count;
3030         struct lu_tgt_desc *tgt;
3031         struct fid_array *fat, **fas = NULL;
3032         int i, rc, **rcs = NULL;
3033
3034         if (!set) {
3035                 set = ptlrpc_prep_set();
3036                 if (!set)
3037                         RETURN(-ENOMEM);
3038         }
3039
3040         /* split FIDs by targets */
3041         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3042         if (fas == NULL)
3043                 GOTO(out, rc = -ENOMEM);
3044         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3045         if (rcs == NULL)
3046                 GOTO(out_fas, rc = -ENOMEM);
3047
3048         for (i = 0; i < fa->fa_nr; i++) {
3049                 unsigned int idx;
3050
3051                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3052                 if (rc) {
3053                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3054                                PFID(&fa->fa_fids[i]), rc);
3055                         continue;
3056                 }
3057                 LASSERT(idx < tgt_count);
3058                 if (!fas[idx])
3059                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3060                                   fa_fids[fa->fa_nr]));
3061                 if (!fas[idx])
3062                         GOTO(out, rc = -ENOMEM);
3063                 if (!rcs[idx])
3064                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3065                 if (!rcs[idx])
3066                         GOTO(out, rc = -ENOMEM);
3067
3068                 fat = fas[idx];
3069                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3070         }
3071
3072         lmv_foreach_connected_tgt(lmv, tgt) {
3073                 fat = fas[tgt->ltd_index];
3074                 if (!fat || fat->fa_nr == 0)
3075                         continue;
3076                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3077         }
3078
3079         rc = ptlrpc_set_wait(NULL, set);
3080         if (rc == 0) {
3081                 int j = 0;
3082                 for (i = 0; i < tgt_count; i++) {
3083                         fat = fas[i];
3084                         if (!fat || fat->fa_nr == 0)
3085                                 continue;
3086                         /* copy FIDs back */
3087                         memcpy(fa->fa_fids + j, fat->fa_fids,
3088                                fat->fa_nr * sizeof(struct lu_fid));
3089                         /* copy rcs back */
3090                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3091                         j += fat->fa_nr;
3092                 }
3093         }
3094         if (set != _set)
3095                 ptlrpc_set_destroy(set);
3096
3097 out:
3098         for (i = 0; i < tgt_count; i++) {
3099                 if (fas && fas[i])
3100                         OBD_FREE(fas[i], offsetof(struct fid_array,
3101                                                 fa_fids[fa->fa_nr]));
3102                 if (rcs && rcs[i])
3103                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3104         }
3105         if (rcs)
3106                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3107 out_fas:
3108         if (fas)
3109                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3110
3111         RETURN(rc);
3112 }
3113
3114 /**
3115  * Asynchronously set by key a value associated with a LMV device.
3116  *
3117  * Dispatch request to lower-layer devices as needed.
3118  *
3119  * \param[in] env       execution environment for this thread
3120  * \param[in] exp       export for the LMV device
3121  * \param[in] keylen    length of key identifier
3122  * \param[in] key       identifier of key to store value for
3123  * \param[in] vallen    size of value to store
3124  * \param[in] val       pointer to data to be stored
3125  * \param[in] set       optional list of related ptlrpc requests
3126  *
3127  * \retval 0            on success
3128  * \retval negative     negated errno on failure
3129  */
3130 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3131                               __u32 keylen, void *key, __u32 vallen, void *val,
3132                               struct ptlrpc_request_set *set)
3133 {
3134         struct lmv_tgt_desc *tgt;
3135         struct obd_device *obd;
3136         struct lmv_obd *lmv;
3137         int rc = 0;
3138         ENTRY;
3139
3140         obd = class_exp2obd(exp);
3141         if (obd == NULL) {
3142                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3143                        exp->exp_handle.h_cookie);
3144                 RETURN(-EINVAL);
3145         }
3146         lmv = &obd->u.lmv;
3147
3148         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3149             KEY_IS(KEY_DEFAULT_EASIZE)) {
3150                 int err = 0;
3151
3152                 lmv_foreach_connected_tgt(lmv, tgt) {
3153                         err = obd_set_info_async(env, tgt->ltd_exp,
3154                                                  keylen, key, vallen, val, set);
3155                         if (err && rc == 0)
3156                                 rc = err;
3157                 }
3158
3159                 RETURN(rc);
3160         }
3161
3162         RETURN(-EINVAL);
3163 }
3164
3165 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3166                             const struct lmv_mds_md_v1 *lmm1)
3167 {
3168         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3169         int             stripe_count;
3170         int             cplen;
3171         int             i;
3172         int             rc = 0;
3173         ENTRY;
3174
3175         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3176         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3177         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3178         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3179                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3180         else
3181                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3182         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3183         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3184         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3185         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3186                         sizeof(lsm->lsm_md_pool_name));
3187
3188         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3189                 RETURN(-E2BIG);
3190
3191         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3192                "layout_version %d\n", lsm->lsm_md_stripe_count,
3193                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3194                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3195                lsm->lsm_md_layout_version);
3196
3197         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3198         for (i = 0; i < stripe_count; i++) {
3199                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3200                               &lmm1->lmv_stripe_fids[i]);
3201                 /*
3202                  * set default value -1, so lmv_locate_tgt() knows this stripe
3203                  * target is not initialized.
3204                  */
3205                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3206                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3207                         continue;
3208
3209                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3210                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3211                 if (rc == -ENOENT)
3212                         continue;
3213
3214                 if (rc)
3215                         RETURN(rc);
3216
3217                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3218                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3219         }
3220
3221         RETURN(rc);
3222 }
3223
3224 static inline int lmv_unpack_user_md(struct obd_export *exp,
3225                                      struct lmv_stripe_md *lsm,
3226                                      const struct lmv_user_md *lmu)
3227 {
3228         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3229         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3230         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3231         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3232         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3233         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3234         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3235
3236         return 0;
3237 }
3238
3239 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3240                         const union lmv_mds_md *lmm, size_t lmm_size)
3241 {
3242         struct lmv_stripe_md     *lsm;
3243         int                      lsm_size;
3244         int                      rc;
3245         bool                     allocated = false;
3246         ENTRY;
3247
3248         LASSERT(lsmp != NULL);
3249
3250         lsm = *lsmp;
3251         /* Free memmd */
3252         if (lsm != NULL && lmm == NULL) {
3253                 int i;
3254                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3255
3256                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3257                         size_t lfm_size;
3258
3259                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3260                                                               lfm_value[0]);
3261                         OBD_FREE_LARGE(lfm, lfm_size);
3262                         RETURN(0);
3263                 }
3264
3265                 if (lmv_dir_striped(lsm)) {
3266                         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3267                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
3268                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3269                 } else {
3270                         lsm_size = lmv_stripe_md_size(0);
3271                 }
3272                 OBD_FREE(lsm, lsm_size);
3273                 *lsmp = NULL;
3274                 RETURN(0);
3275         }
3276
3277         /* foreign lmv case */
3278         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3279                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3280
3281                 if (lfm == NULL) {
3282                         OBD_ALLOC_LARGE(lfm, lmm_size);
3283                         if (lfm == NULL)
3284                                 RETURN(-ENOMEM);
3285                         *lsmp = (struct lmv_stripe_md *)lfm;
3286                 }
3287                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3288                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3289                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3290                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3291                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3292                        lfm->lfm_length);
3293                 RETURN(lmm_size);
3294         }
3295
3296         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3297                 RETURN(-EPERM);
3298
3299         /* Unpack memmd */
3300