Whamcloud - gitweb
LU-9855 lustre: rename obd_device variables to obd
[fs/lustre-release.git] / lustre / lmv / lmv_obd.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  */
32
33 #define DEBUG_SUBSYSTEM S_LMV
34
35 #include <linux/file.h>
36 #include <linux/module.h>
37 #include <linux/init.h>
38 #include <linux/user_namespace.h>
39 #include <linux/uidgid.h>
40 #include <linux/slab.h>
41 #include <linux/pagemap.h>
42 #include <linux/mm.h>
43 #include <linux/math64.h>
44 #include <linux/seq_file.h>
45 #include <linux/namei.h>
46
47 #include <obd_support.h>
48 #include <lustre_lib.h>
49 #include <lustre_net.h>
50 #include <obd_class.h>
51 #include <lustre_lmv.h>
52 #include <lprocfs_status.h>
53 #include <cl_object.h>
54 #include <lustre_fid.h>
55 #include <uapi/linux/lustre/lustre_ioctl.h>
56 #include <lustre_kernelcomm.h>
57 #include "lmv_internal.h"
58
59 static int lmv_check_connect(struct obd_device *obd);
60
61 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
62                          int activate)
63 {
64         if (tgt->ltd_active == activate)
65                 return;
66
67         tgt->ltd_active = activate;
68         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
69                 (activate ? 1 : -1);
70
71         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
72 }
73
74 /**
75  * Error codes:
76  *
77  *  -EINVAL  : UUID can't be found in the LMV's target list
78  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
79  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
80  */
81 static int lmv_set_mdc_active(struct lmv_obd *lmv,
82                               const struct obd_uuid *uuid,
83                               int activate)
84 {
85         struct lu_tgt_desc *tgt = NULL;
86         struct obd_device *obd;
87         int rc = 0;
88
89         ENTRY;
90
91         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
92                         lmv, uuid->uuid, activate);
93
94         spin_lock(&lmv->lmv_lock);
95         lmv_foreach_connected_tgt(lmv, tgt) {
96                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
97                        tgt->ltd_index, tgt->ltd_uuid.uuid,
98                        tgt->ltd_exp->exp_handle.h_cookie);
99
100                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
101                         break;
102         }
103
104         if (!tgt)
105                 GOTO(out_lmv_lock, rc = -EINVAL);
106
107         obd = class_exp2obd(tgt->ltd_exp);
108         if (obd == NULL)
109                 GOTO(out_lmv_lock, rc = -ENOTCONN);
110
111         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
112                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
113                obd->obd_type->typ_name, tgt->ltd_index);
114         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
115
116         if (tgt->ltd_active == activate) {
117                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
118                        activate ? "" : "in");
119                 GOTO(out_lmv_lock, rc);
120         }
121
122         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
123                activate ? "" : "in");
124         lmv_activate_target(lmv, tgt, activate);
125         EXIT;
126
127  out_lmv_lock:
128         spin_unlock(&lmv->lmv_lock);
129         return rc;
130 }
131
132 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
133 {
134         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
135         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
136
137         return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
138 }
139
140 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
141                       enum obd_notify_event ev)
142 {
143         struct obd_connect_data *conn_data;
144         struct lmv_obd          *lmv = &obd->u.lmv;
145         struct obd_uuid         *uuid;
146         int                      rc = 0;
147         ENTRY;
148
149         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
150                 CERROR("unexpected notification of %s %s!\n",
151                        watched->obd_type->typ_name,
152                        watched->obd_name);
153                 RETURN(-EINVAL);
154         }
155
156         uuid = &watched->u.cli.cl_target_uuid;
157         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
158                 /*
159                  * Set MDC as active before notifying the observer, so the
160                  * observer can use the MDC normally.
161                  */
162                 rc = lmv_set_mdc_active(lmv, uuid,
163                                         ev == OBD_NOTIFY_ACTIVE);
164                 if (rc) {
165                         CERROR("%sactivation of %s failed: %d\n",
166                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
167                                uuid->uuid, rc);
168                         RETURN(rc);
169                 }
170         } else if (ev == OBD_NOTIFY_OCD) {
171                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
172                 /*
173                  * XXX: Make sure that ocd_connect_flags from all targets are
174                  * the same. Otherwise one of MDTs runs wrong version or
175                  * something like this.  --umka
176                  */
177                 obd->obd_self_export->exp_connect_data = *conn_data;
178         }
179
180         /*
181          * Pass the notification up the chain.
182          */
183         if (obd->obd_observer)
184                 rc = obd_notify(obd->obd_observer, watched, ev);
185
186         RETURN(rc);
187 }
188
189 static int lmv_connect(const struct lu_env *env,
190                        struct obd_export **pexp, struct obd_device *obd,
191                        struct obd_uuid *cluuid, struct obd_connect_data *data,
192                        void *localdata)
193 {
194         struct lmv_obd *lmv = &obd->u.lmv;
195         struct lustre_handle conn = { 0 };
196         struct obd_export *exp;
197         int rc;
198         ENTRY;
199
200         rc = class_connect(&conn, obd, cluuid);
201         if (rc) {
202                 CERROR("class_connection() returned %d\n", rc);
203                 RETURN(rc);
204         }
205
206         exp = class_conn2export(&conn);
207
208         lmv->connected = 0;
209         lmv->conn_data = *data;
210         lmv->lmv_cache = localdata;
211
212         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
213                                                     &obd->obd_kset.kobj);
214         if (!lmv->lmv_tgts_kobj) {
215                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
216                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
217         }
218
219         rc = lmv_check_connect(obd);
220         if (rc != 0)
221                 GOTO(out_sysfs, rc);
222
223         *pexp = exp;
224
225         RETURN(rc);
226
227 out_sysfs:
228         if (lmv->lmv_tgts_kobj)
229                 kobject_put(lmv->lmv_tgts_kobj);
230
231         class_disconnect(exp);
232
233         return rc;
234 }
235
236 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
237                             __u32 def_easize)
238 {
239         struct obd_device *obd = exp->exp_obd;
240         struct lmv_obd *lmv = &obd->u.lmv;
241         struct lmv_tgt_desc *tgt;
242         int change = 0;
243         int rc = 0;
244
245         ENTRY;
246
247         if (lmv->max_easize < easize) {
248                 lmv->max_easize = easize;
249                 change = 1;
250         }
251         if (lmv->max_def_easize < def_easize) {
252                 lmv->max_def_easize = def_easize;
253                 change = 1;
254         }
255
256         if (change == 0)
257                 RETURN(0);
258
259         if (lmv->connected == 0)
260                 RETURN(0);
261
262         lmv_foreach_connected_tgt(lmv, tgt) {
263                 if (!tgt->ltd_active)
264                         continue;
265
266                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
267                 if (rc) {
268                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
269                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
270                         break;
271                 }
272         }
273         RETURN(rc);
274 }
275
276 #define MAX_STRING_SIZE 128
277
278 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
279 {
280         struct lmv_obd          *lmv = &obd->u.lmv;
281         struct obd_device       *mdc_obd;
282         struct obd_export       *mdc_exp;
283         struct lu_fld_target     target;
284         int                      rc;
285         ENTRY;
286
287         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
288                                         &obd->obd_uuid);
289         if (!mdc_obd) {
290                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
291                 RETURN(-EINVAL);
292         }
293
294         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
295                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
296                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
297
298         if (!mdc_obd->obd_set_up) {
299                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
300                 RETURN(-EINVAL);
301         }
302
303         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
304                          &lmv->conn_data, lmv->lmv_cache);
305         if (rc) {
306                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
307                 RETURN(rc);
308         }
309
310         /*
311          * Init fid sequence client for this mdc and add new fld target.
312          */
313         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
314         if (rc)
315                 RETURN(rc);
316
317         target.ft_srv = NULL;
318         target.ft_exp = mdc_exp;
319         target.ft_idx = tgt->ltd_index;
320
321         fld_client_add_target(&lmv->lmv_fld, &target);
322
323         rc = obd_register_observer(mdc_obd, obd);
324         if (rc) {
325                 obd_disconnect(mdc_exp);
326                 CERROR("target %s register_observer error %d\n",
327                        tgt->ltd_uuid.uuid, rc);
328                 RETURN(rc);
329         }
330
331         if (obd->obd_observer) {
332                 /*
333                  * Tell the observer about the new target.
334                  */
335                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
336                                 OBD_NOTIFY_ACTIVE);
337                 if (rc) {
338                         obd_disconnect(mdc_exp);
339                         RETURN(rc);
340                 }
341         }
342
343         tgt->ltd_active = 1;
344         tgt->ltd_exp = mdc_exp;
345         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
346
347         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
348
349         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
350         if (rc) {
351                 obd_disconnect(mdc_exp);
352                 RETURN(rc);
353         }
354
355         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
356                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
357                 atomic_read(&obd->obd_refcount));
358
359         lmv_statfs_check_update(obd, tgt);
360
361         if (lmv->lmv_tgts_kobj)
362                 /* Even if we failed to create the link, that's fine */
363                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
364                                        &mdc_obd->obd_kset.kobj,
365                                        mdc_obd->obd_name);
366         RETURN(0);
367 }
368
369 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
370 {
371         LASSERT(tgt);
372         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
373         OBD_FREE_PTR(tgt);
374 }
375
376 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
377                            __u32 index, int gen)
378 {
379         struct obd_device *mdc_obd;
380         struct lmv_obd *lmv = &obd->u.lmv;
381         struct lmv_tgt_desc *tgt;
382         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
383         int rc = 0;
384
385         ENTRY;
386
387         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
388         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
389                                         &obd->obd_uuid);
390         if (!mdc_obd) {
391                 CERROR("%s: Target %s not attached: rc = %d\n",
392                        obd->obd_name, uuidp->uuid, -EINVAL);
393                 RETURN(-EINVAL);
394         }
395
396         OBD_ALLOC_PTR(tgt);
397         if (!tgt)
398                 RETURN(-ENOMEM);
399
400         mutex_init(&tgt->ltd_fid_mutex);
401         tgt->ltd_index = index;
402         tgt->ltd_uuid = *uuidp;
403         tgt->ltd_active = 0;
404
405         mutex_lock(&ltd->ltd_mutex);
406         rc = ltd_add_tgt(ltd, tgt);
407         mutex_unlock(&ltd->ltd_mutex);
408
409         if (rc)
410                 GOTO(out_tgt, rc);
411
412         if (!lmv->connected)
413                 /* lmv_check_connect() will connect this target. */
414                 RETURN(0);
415
416         rc = lmv_connect_mdc(obd, tgt);
417         if (!rc) {
418                 int easize = sizeof(struct lmv_stripe_md) +
419                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
420
421                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
422         }
423
424         RETURN(rc);
425
426 out_tgt:
427         OBD_FREE_PTR(tgt);
428         return rc;
429 }
430
431 static int lmv_check_connect(struct obd_device *obd)
432 {
433         struct lmv_obd *lmv = &obd->u.lmv;
434         struct lmv_tgt_desc *tgt;
435         int easize;
436         int rc;
437
438         ENTRY;
439
440         if (lmv->connected)
441                 RETURN(0);
442
443         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
444         if (lmv->connected)
445                 GOTO(unlock, rc = 0);
446
447         if (!lmv->lmv_mdt_count) {
448                 CERROR("%s: no targets configured: rc = -EINVAL\n",
449                        obd->obd_name);
450                 GOTO(unlock, rc = -EINVAL);
451         }
452
453         if (!lmv_mdt0_inited(lmv)) {
454                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
455                        obd->obd_name);
456                 GOTO(unlock, rc = -EINVAL);
457         }
458
459         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
460                obd->obd_uuid.uuid, obd->obd_name);
461
462         lmv_foreach_tgt(lmv, tgt) {
463                 rc = lmv_connect_mdc(obd, tgt);
464                 if (rc)
465                         GOTO(out_disc, rc);
466         }
467
468         lmv->connected = 1;
469         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
470         lmv_init_ea_size(obd->obd_self_export, easize, 0);
471         EXIT;
472 unlock:
473         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
474
475         return rc;
476
477 out_disc:
478         lmv_foreach_tgt(lmv, tgt) {
479                 tgt->ltd_active = 0;
480                 if (!tgt->ltd_exp)
481                         continue;
482
483                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
484                 obd_disconnect(tgt->ltd_exp);
485         }
486
487         goto unlock;
488 }
489
490 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
491 {
492         struct lmv_obd         *lmv = &obd->u.lmv;
493         struct obd_device      *mdc_obd;
494         int                     rc;
495         ENTRY;
496
497         LASSERT(tgt != NULL);
498         LASSERT(obd != NULL);
499
500         mdc_obd = class_exp2obd(tgt->ltd_exp);
501
502         if (mdc_obd) {
503                 mdc_obd->obd_force = obd->obd_force;
504                 mdc_obd->obd_fail = obd->obd_fail;
505                 mdc_obd->obd_no_recov = obd->obd_no_recov;
506
507                 if (lmv->lmv_tgts_kobj)
508                         sysfs_remove_link(lmv->lmv_tgts_kobj,
509                                           mdc_obd->obd_name);
510         }
511
512         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
513         if (rc)
514                 CERROR("Can't finanize fids factory\n");
515
516         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
517                tgt->ltd_exp->exp_obd->obd_name,
518                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
519
520         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
521         rc = obd_disconnect(tgt->ltd_exp);
522         if (rc) {
523                 if (tgt->ltd_active) {
524                         CERROR("Target %s disconnect error %d\n",
525                                tgt->ltd_uuid.uuid, rc);
526                 }
527         }
528
529         lmv_activate_target(lmv, tgt, 0);
530         tgt->ltd_exp = NULL;
531         RETURN(0);
532 }
533
534 static int lmv_disconnect(struct obd_export *exp)
535 {
536         struct obd_device *obd = class_exp2obd(exp);
537         struct lmv_obd *lmv = &obd->u.lmv;
538         struct lmv_tgt_desc *tgt;
539         int rc;
540
541         ENTRY;
542
543         lmv_foreach_connected_tgt(lmv, tgt)
544                 lmv_disconnect_mdc(obd, tgt);
545
546         if (lmv->lmv_tgts_kobj)
547                 kobject_put(lmv->lmv_tgts_kobj);
548
549         if (!lmv->connected)
550                 class_export_put(exp);
551         rc = class_disconnect(exp);
552         lmv->connected = 0;
553
554         RETURN(rc);
555 }
556
557 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
558                         void __user *uarg)
559 {
560         struct obd_device *obd = class_exp2obd(exp);
561         struct lmv_obd *lmv = &obd->u.lmv;
562         struct getinfo_fid2path *gf;
563         struct lmv_tgt_desc *tgt;
564         struct getinfo_fid2path *remote_gf = NULL;
565         struct lu_fid root_fid;
566         int remote_gf_size = 0;
567         int rc;
568
569         gf = karg;
570         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
571         if (IS_ERR(tgt))
572                 RETURN(PTR_ERR(tgt));
573
574         root_fid = *gf->gf_u.gf_root_fid;
575         LASSERT(fid_is_sane(&root_fid));
576
577 repeat_fid2path:
578         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
579         if (rc != 0 && rc != -EREMOTE)
580                 GOTO(out_fid2path, rc);
581
582         /* If remote_gf != NULL, it means just building the
583          * path on the remote MDT, copy this path segement to gf */
584         if (remote_gf != NULL) {
585                 struct getinfo_fid2path *ori_gf;
586                 char *ptr;
587                 int len;
588
589                 ori_gf = (struct getinfo_fid2path *)karg;
590                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
591                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
592                         GOTO(out_fid2path, rc = -EOVERFLOW);
593
594                 ptr = ori_gf->gf_u.gf_path;
595
596                 len = strlen(gf->gf_u.gf_path);
597                 /* move the current path to the right to release space
598                  * for closer-to-root part */
599                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
600                 memcpy(ptr, gf->gf_u.gf_path, len);
601                 ptr[len] = '/';
602         }
603
604         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
605                tgt->ltd_exp->exp_obd->obd_name,
606                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
607                gf->gf_linkno);
608
609         if (rc == 0)
610                 GOTO(out_fid2path, rc);
611
612         /* sigh, has to go to another MDT to do path building further */
613         if (remote_gf == NULL) {
614                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
615                 OBD_ALLOC(remote_gf, remote_gf_size);
616                 if (remote_gf == NULL)
617                         GOTO(out_fid2path, rc = -ENOMEM);
618                 remote_gf->gf_pathlen = PATH_MAX;
619         }
620
621         if (!fid_is_sane(&gf->gf_fid)) {
622                 CERROR("%s: invalid FID "DFID": rc = %d\n",
623                        tgt->ltd_exp->exp_obd->obd_name,
624                        PFID(&gf->gf_fid), -EINVAL);
625                 GOTO(out_fid2path, rc = -EINVAL);
626         }
627
628         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
629         if (IS_ERR(tgt))
630                 GOTO(out_fid2path, rc = -EINVAL);
631
632         remote_gf->gf_fid = gf->gf_fid;
633         remote_gf->gf_recno = -1;
634         remote_gf->gf_linkno = -1;
635         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
636         *remote_gf->gf_u.gf_root_fid = root_fid;
637         gf = remote_gf;
638         goto repeat_fid2path;
639
640 out_fid2path:
641         if (remote_gf != NULL)
642                 OBD_FREE(remote_gf, remote_gf_size);
643         RETURN(rc);
644 }
645
646 static int lmv_hsm_req_count(struct lmv_obd *lmv,
647                              const struct hsm_user_request *hur,
648                              const struct lmv_tgt_desc *tgt_mds)
649 {
650         struct lmv_tgt_desc *curr_tgt;
651         __u32 i;
652         int nr = 0;
653
654         /* count how many requests must be sent to the given target */
655         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
656                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
657                 if (IS_ERR(curr_tgt))
658                         RETURN(PTR_ERR(curr_tgt));
659                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
660                         nr++;
661         }
662         return nr;
663 }
664
665 static int lmv_hsm_req_build(struct lmv_obd *lmv,
666                               struct hsm_user_request *hur_in,
667                               const struct lmv_tgt_desc *tgt_mds,
668                               struct hsm_user_request *hur_out)
669 {
670         __u32 i, nr_out;
671         struct lmv_tgt_desc *curr_tgt;
672
673         /* build the hsm_user_request for the given target */
674         hur_out->hur_request = hur_in->hur_request;
675         nr_out = 0;
676         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
677                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
678                 if (IS_ERR(curr_tgt))
679                         RETURN(PTR_ERR(curr_tgt));
680                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
681                         hur_out->hur_user_item[nr_out] =
682                                                 hur_in->hur_user_item[i];
683                         nr_out++;
684                 }
685         }
686         hur_out->hur_request.hr_itemcount = nr_out;
687         memcpy(hur_data(hur_out), hur_data(hur_in),
688                hur_in->hur_request.hr_data_len);
689
690         RETURN(0);
691 }
692
693 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
694                                  int len, struct lustre_kernelcomm *lk,
695                                  void __user *uarg)
696 {
697         struct lmv_obd *lmv = &obd->u.lmv;
698         struct lu_tgt_desc *tgt;
699         int rc;
700
701         ENTRY;
702
703         /* unregister request (call from llapi_hsm_copytool_fini) */
704         lmv_foreach_connected_tgt(lmv, tgt)
705                 /* best effort: try to clean as much as possible
706                  * (continue on error) */
707                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
708
709         /* Whatever the result, remove copytool from kuc groups.
710          * Unreached coordinators will get EPIPE on next requests
711          * and will unregister automatically.
712          */
713         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
714
715         RETURN(rc);
716 }
717
718 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
719                                int len, struct lustre_kernelcomm *lk,
720                                void __user *uarg)
721 {
722         struct lmv_obd *lmv = &obd->u.lmv;
723         struct file *filp;
724         bool any_set = false;
725         struct kkuc_ct_data *kcd;
726         size_t kcd_size;
727         struct lu_tgt_desc *tgt;
728         __u32 i;
729         int err;
730         int rc = 0;
731
732         ENTRY;
733
734         filp = fget(lk->lk_wfd);
735         if (!filp)
736                 RETURN(-EBADF);
737
738         if (lk->lk_flags & LK_FLG_DATANR)
739                 kcd_size = offsetof(struct kkuc_ct_data,
740                                     kcd_archives[lk->lk_data_count]);
741         else
742                 kcd_size = sizeof(*kcd);
743
744         OBD_ALLOC(kcd, kcd_size);
745         if (kcd == NULL)
746                 GOTO(err_fput, rc = -ENOMEM);
747
748         kcd->kcd_nr_archives = lk->lk_data_count;
749         if (lk->lk_flags & LK_FLG_DATANR) {
750                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
751                 if (lk->lk_data_count > 0)
752                         memcpy(kcd->kcd_archives, lk->lk_data,
753                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
754         } else {
755                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
756         }
757
758         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
759                                    lk->lk_group, kcd, kcd_size);
760         OBD_FREE(kcd, kcd_size);
761         if (rc)
762                 GOTO(err_fput, rc);
763
764         /* All or nothing: try to register to all MDS.
765          * In case of failure, unregister from previous MDS,
766          * except if it because of inactive target. */
767         lmv_foreach_connected_tgt(lmv, tgt) {
768                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
769                 if (err) {
770                         if (tgt->ltd_active) {
771                                 /* permanent error */
772                                 CERROR("%s: iocontrol MDC %s on MDT"
773                                        " idx %d cmd %x: err = %d\n",
774                                        lmv2obd_dev(lmv)->obd_name,
775                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
776                                        err);
777                                 rc = err;
778                                 lk->lk_flags |= LK_FLG_STOP;
779                                 i = tgt->ltd_index;
780                                 /* unregister from previous MDS */
781                                 lmv_foreach_connected_tgt(lmv, tgt) {
782                                         if (tgt->ltd_index >= i)
783                                                 break;
784
785                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
786                                                       lk, uarg);
787                                 }
788                                 GOTO(err_kkuc_rem, rc);
789                         }
790                         /* else: transient error.
791                          * kuc will register to the missing MDT
792                          * when it is back */
793                 } else {
794                         any_set = true;
795                 }
796         }
797
798         if (!any_set)
799                 /* no registration done: return error */
800                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
801
802         RETURN(0);
803
804 err_kkuc_rem:
805         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
806
807 err_fput:
808         fput(filp);
809         return rc;
810 }
811
812 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
813                          int len, void *karg, void __user *uarg)
814 {
815         struct obd_device *obd = class_exp2obd(exp);
816         struct lmv_obd *lmv = &obd->u.lmv;
817         struct lu_tgt_desc *tgt = NULL;
818         int set = 0;
819         __u32 count = lmv->lmv_mdt_count;
820         int rc = 0;
821
822         ENTRY;
823
824         if (count == 0)
825                 RETURN(-ENOTTY);
826
827         switch (cmd) {
828         case IOC_OBD_STATFS: {
829                 struct obd_ioctl_data *data = karg;
830                 struct obd_device *mdc_obd;
831                 struct obd_statfs stat_buf = {0};
832                 __u32 index;
833
834                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
835
836                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
837                         RETURN(-ENODEV);
838
839                 tgt = lmv_tgt(lmv, index);
840                 if (!tgt || !tgt->ltd_active)
841                         RETURN(-ENODATA);
842
843                 mdc_obd = class_exp2obd(tgt->ltd_exp);
844                 if (!mdc_obd)
845                         RETURN(-EINVAL);
846
847                 /* copy UUID */
848                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
849                                  min((int) data->ioc_plen2,
850                                      (int) sizeof(struct obd_uuid))))
851                         RETURN(-EFAULT);
852
853                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
854                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
855                                 0);
856                 if (rc)
857                         RETURN(rc);
858                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
859                                  min((int) data->ioc_plen1,
860                                      (int) sizeof(stat_buf))))
861                         RETURN(-EFAULT);
862                 break;
863         }
864         case OBD_IOC_QUOTACTL: {
865                 struct if_quotactl *qctl = karg;
866                 struct obd_quotactl *oqctl;
867
868                 if (qctl->qc_valid == QC_MDTIDX) {
869                         tgt = lmv_tgt(lmv, qctl->qc_idx);
870                 } else if (qctl->qc_valid == QC_UUID) {
871                         lmv_foreach_tgt(lmv, tgt) {
872                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
873                                                      &qctl->obd_uuid))
874                                         continue;
875
876                                 if (!tgt->ltd_exp)
877                                         RETURN(-EINVAL);
878
879                                 break;
880                         }
881                 } else {
882                         RETURN(-EINVAL);
883                 }
884
885                 if (!tgt || !tgt->ltd_exp)
886                         RETURN(-EINVAL);
887
888                 OBD_ALLOC_PTR(oqctl);
889                 if (!oqctl)
890                         RETURN(-ENOMEM);
891
892                 QCTL_COPY(oqctl, qctl);
893                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
894                 if (rc == 0) {
895                         QCTL_COPY(qctl, oqctl);
896                         qctl->qc_valid = QC_MDTIDX;
897                         qctl->obd_uuid = tgt->ltd_uuid;
898                 }
899                 OBD_FREE_PTR(oqctl);
900                 break;
901         }
902         case LL_IOC_GET_CONNECT_FLAGS: {
903                 tgt = lmv_tgt(lmv, 0);
904                 rc = -ENODATA;
905                 if (tgt && tgt->ltd_exp)
906                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
907                 break;
908         }
909         case LL_IOC_FID2MDTIDX: {
910                 struct lu_fid *fid = karg;
911                 int             mdt_index;
912
913                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
914                 if (rc != 0)
915                         RETURN(rc);
916
917                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
918                  * point to user space memory for FID2MDTIDX. */
919                 *(__u32 *)uarg = mdt_index;
920                 break;
921         }
922         case OBD_IOC_FID2PATH: {
923                 rc = lmv_fid2path(exp, len, karg, uarg);
924                 break;
925         }
926         case LL_IOC_HSM_STATE_GET:
927         case LL_IOC_HSM_STATE_SET:
928         case LL_IOC_HSM_ACTION: {
929                 struct md_op_data *op_data = karg;
930
931                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
932                 if (IS_ERR(tgt))
933                         RETURN(PTR_ERR(tgt));
934
935                 if (tgt->ltd_exp == NULL)
936                         RETURN(-EINVAL);
937
938                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
939                 break;
940         }
941         case LL_IOC_HSM_PROGRESS: {
942                 const struct hsm_progress_kernel *hpk = karg;
943
944                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
945                 if (IS_ERR(tgt))
946                         RETURN(PTR_ERR(tgt));
947                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
948                 break;
949         }
950         case LL_IOC_HSM_REQUEST: {
951                 struct hsm_user_request *hur = karg;
952                 unsigned int reqcount = hur->hur_request.hr_itemcount;
953
954                 if (reqcount == 0)
955                         RETURN(0);
956
957                 /* if the request is about a single fid
958                  * or if there is a single MDS, no need to split
959                  * the request. */
960                 if (reqcount == 1 || count == 1) {
961                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
962                         if (IS_ERR(tgt))
963                                 RETURN(PTR_ERR(tgt));
964                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
965                 } else {
966                         /* split fid list to their respective MDS */
967                         lmv_foreach_connected_tgt(lmv, tgt) {
968                                 int nr, rc1;
969                                 size_t reqlen;
970                                 struct hsm_user_request *req;
971
972                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
973                                 if (nr < 0)
974                                         RETURN(nr);
975                                 if (nr == 0) /* nothing for this MDS */
976                                         continue;
977
978                                 /* build a request with fids for this MDS */
979                                 reqlen = offsetof(typeof(*hur),
980                                                   hur_user_item[nr])
981                                                 + hur->hur_request.hr_data_len;
982                                 OBD_ALLOC_LARGE(req, reqlen);
983                                 if (req == NULL)
984                                         RETURN(-ENOMEM);
985                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
986                                 if (rc1 < 0)
987                                         GOTO(hsm_req_err, rc1);
988                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
989                                                     req, uarg);
990 hsm_req_err:
991                                 if (rc1 != 0 && rc == 0)
992                                         rc = rc1;
993                                 OBD_FREE_LARGE(req, reqlen);
994                         }
995                 }
996                 break;
997         }
998         case LL_IOC_LOV_SWAP_LAYOUTS: {
999                 struct md_op_data *op_data = karg;
1000                 struct lmv_tgt_desc *tgt1, *tgt2;
1001
1002                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1003                 if (IS_ERR(tgt1))
1004                         RETURN(PTR_ERR(tgt1));
1005
1006                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1007                 if (IS_ERR(tgt2))
1008                         RETURN(PTR_ERR(tgt2));
1009
1010                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1011                         RETURN(-EINVAL);
1012
1013                 /* only files on same MDT can have their layouts swapped */
1014                 if (tgt1->ltd_index != tgt2->ltd_index)
1015                         RETURN(-EPERM);
1016
1017                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1018                 break;
1019         }
1020         case LL_IOC_HSM_CT_START: {
1021                 struct lustre_kernelcomm *lk = karg;
1022                 if (lk->lk_flags & LK_FLG_STOP)
1023                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1024                 else
1025                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1026                 break;
1027         }
1028         default:
1029                 lmv_foreach_connected_tgt(lmv, tgt) {
1030                         struct obd_device *mdc_obd;
1031                         int err;
1032
1033                         /* ll_umount_begin() sets force flag but for lmv, not
1034                          * mdc. Let's pass it through */
1035                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1036                         mdc_obd->obd_force = obd->obd_force;
1037                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1038                         if (err) {
1039                                 if (tgt->ltd_active) {
1040                                         CERROR("error: iocontrol MDC %s on MDT"
1041                                                " idx %d cmd %x: err = %d\n",
1042                                                tgt->ltd_uuid.uuid,
1043                                                tgt->ltd_index, cmd, err);
1044                                         if (!rc)
1045                                                 rc = err;
1046                                 }
1047                         } else
1048                                 set = 1;
1049                 }
1050                 if (!set && !rc)
1051                         rc = -EIO;
1052         }
1053         RETURN(rc);
1054 }
1055
1056 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1057                   struct lu_fid *fid, struct md_op_data *op_data)
1058 {
1059         struct obd_device *obd = class_exp2obd(exp);
1060         struct lmv_obd *lmv = &obd->u.lmv;
1061         struct lmv_tgt_desc *tgt;
1062         int rc;
1063
1064         ENTRY;
1065
1066         LASSERT(op_data);
1067         LASSERT(fid);
1068
1069         tgt = lmv_tgt(lmv, op_data->op_mds);
1070         if (!tgt)
1071                 RETURN(-ENODEV);
1072
1073         if (!tgt->ltd_active || !tgt->ltd_exp)
1074                 RETURN(-ENODEV);
1075
1076         /*
1077          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1078          * on server that seq in new allocated fid is not yet known.
1079          */
1080         mutex_lock(&tgt->ltd_fid_mutex);
1081         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1082         mutex_unlock(&tgt->ltd_fid_mutex);
1083         if (rc > 0) {
1084                 LASSERT(fid_is_sane(fid));
1085                 rc = 0;
1086         }
1087
1088         RETURN(rc);
1089 }
1090
1091 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1092 {
1093         struct lmv_obd *lmv = &obd->u.lmv;
1094         struct lmv_desc *desc;
1095         struct lnet_process_id lnet_id;
1096         int i = 0;
1097         int rc;
1098
1099         ENTRY;
1100
1101         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1102                 CERROR("LMV setup requires a descriptor\n");
1103                 RETURN(-EINVAL);
1104         }
1105
1106         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1107         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1108                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1109                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1110                 RETURN(-EINVAL);
1111         }
1112
1113         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1114                      desc->ld_uuid.uuid);
1115         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1116         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1117         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1118                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1119         lmv->max_def_easize = 0;
1120         lmv->max_easize = 0;
1121
1122         spin_lock_init(&lmv->lmv_lock);
1123
1124         /*
1125          * initialize rr_index to lower 32bit of netid, so that client
1126          * can distribute subdirs evenly from the beginning.
1127          */
1128         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1129                 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1130                         lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1131                         break;
1132                 }
1133         }
1134
1135         rc = lmv_tunables_init(obd);
1136         if (rc)
1137                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1138                       obd->obd_name, rc);
1139
1140         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1141                              LUSTRE_CLI_FLD_HASH_DHT);
1142         if (rc)
1143                 CERROR("Can't init FLD, err %d\n", rc);
1144
1145         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1146         if (rc)
1147                 CWARN("%s: error initialize target table: rc = %d\n",
1148                       obd->obd_name, rc);
1149
1150         RETURN(rc);
1151 }
1152
1153 static int lmv_cleanup(struct obd_device *obd)
1154 {
1155         struct lmv_obd *lmv = &obd->u.lmv;
1156         struct lu_tgt_desc *tgt;
1157         struct lu_tgt_desc *tmp;
1158
1159         ENTRY;
1160
1161         fld_client_fini(&lmv->lmv_fld);
1162         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1163                 lmv_del_target(lmv, tgt);
1164         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1165
1166         RETURN(0);
1167 }
1168
1169 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1170 {
1171         struct lustre_cfg       *lcfg = buf;
1172         struct obd_uuid         obd_uuid;
1173         int                     gen;
1174         __u32                   index;
1175         int                     rc;
1176         ENTRY;
1177
1178         switch (lcfg->lcfg_command) {
1179         case LCFG_ADD_MDC:
1180                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1181                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1182                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1183                         GOTO(out, rc = -EINVAL);
1184
1185                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1186
1187                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1188                         GOTO(out, rc = -EINVAL);
1189                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1190                         GOTO(out, rc = -EINVAL);
1191                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1192                 GOTO(out, rc);
1193         default:
1194                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1195                 GOTO(out, rc = -EINVAL);
1196         }
1197 out:
1198         RETURN(rc);
1199 }
1200
1201 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1202 {
1203         int i;
1204
1205         if (flags & OBD_STATFS_FOR_MDT0)
1206                 return 0;
1207
1208         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1209                 return lmv->lmv_statfs_start;
1210
1211         /* choose initial MDT for this client */
1212         for (i = 0;; i++) {
1213                 struct lnet_process_id lnet_id;
1214                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1215                         break;
1216
1217                 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1218                         /* We dont need a full 64-bit modulus, just enough
1219                          * to distribute the requests across MDTs evenly.
1220                          */
1221                         lmv->lmv_statfs_start = (u32)lnet_id.nid %
1222                                                 lmv->lmv_mdt_count;
1223                         break;
1224                 }
1225         }
1226
1227         return lmv->lmv_statfs_start;
1228 }
1229
1230 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1231                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1232 {
1233         struct obd_device *obd = class_exp2obd(exp);
1234         struct lmv_obd *lmv = &obd->u.lmv;
1235         struct obd_statfs *temp;
1236         struct lu_tgt_desc *tgt;
1237         __u32 i;
1238         __u32 idx;
1239         int rc = 0;
1240
1241         ENTRY;
1242
1243         OBD_ALLOC(temp, sizeof(*temp));
1244         if (temp == NULL)
1245                 RETURN(-ENOMEM);
1246
1247         /* distribute statfs among MDTs */
1248         idx = lmv_select_statfs_mdt(lmv, flags);
1249
1250         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1251                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1252                 tgt = lmv_tgt(lmv, idx);
1253                 if (!tgt || !tgt->ltd_exp)
1254                         continue;
1255
1256                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age, flags);
1257                 if (rc) {
1258                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1259                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1260                         GOTO(out_free_temp, rc);
1261                 }
1262
1263                 if (temp->os_state & OS_STATE_SUM ||
1264                     flags == OBD_STATFS_FOR_MDT0) {
1265                         /* reset to the last aggregated values
1266                          * and don't sum with non-aggrated data */
1267                         /* If the statfs is from mount, it needs to retrieve
1268                          * necessary information from MDT0. i.e. mount does
1269                          * not need the merged osfs from all of MDT. Also
1270                          * clients can be mounted as long as MDT0 is in
1271                          * service */
1272                         *osfs = *temp;
1273                         break;
1274                 }
1275
1276                 if (i == 0) {
1277                         *osfs = *temp;
1278                 } else {
1279                         osfs->os_bavail += temp->os_bavail;
1280                         osfs->os_blocks += temp->os_blocks;
1281                         osfs->os_ffree += temp->os_ffree;
1282                         osfs->os_files += temp->os_files;
1283                         osfs->os_granted += temp->os_granted;
1284                 }
1285         }
1286
1287         EXIT;
1288 out_free_temp:
1289         OBD_FREE(temp, sizeof(*temp));
1290         return rc;
1291 }
1292
1293 static int lmv_statfs_update(void *cookie, int rc)
1294 {
1295         struct obd_info *oinfo = cookie;
1296         struct obd_device *obd = oinfo->oi_obd;
1297         struct lmv_obd *lmv = &obd->u.lmv;
1298         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1299         struct obd_statfs *osfs = oinfo->oi_osfs;
1300
1301         /*
1302          * NB: don't deactivate TGT upon error, because we may not trigger async
1303          * statfs any longer, then there is no chance to activate TGT.
1304          */
1305         if (!rc) {
1306                 spin_lock(&lmv->lmv_lock);
1307                 tgt->ltd_statfs = *osfs;
1308                 tgt->ltd_statfs_age = ktime_get_seconds();
1309                 spin_unlock(&lmv->lmv_lock);
1310                 lmv->lmv_qos.lq_dirty = 1;
1311         }
1312
1313         return rc;
1314 }
1315
1316 /* update tgt statfs async if it's ld_qos_maxage old */
1317 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1318 {
1319         struct obd_info oinfo = {
1320                 .oi_obd = obd,
1321                 .oi_tgt = tgt,
1322                 .oi_cb_up = lmv_statfs_update,
1323         };
1324         int rc;
1325
1326         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1327             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1328                 return 0;
1329
1330         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1331
1332         return rc;
1333 }
1334
1335 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1336                         struct lu_fid *fid)
1337 {
1338         struct obd_device *obd = exp->exp_obd;
1339         struct lmv_obd *lmv = &obd->u.lmv;
1340         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1341         int rc;
1342
1343         ENTRY;
1344
1345         if (!tgt)
1346                 RETURN(-ENODEV);
1347
1348         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1349         RETURN(rc);
1350 }
1351
1352 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1353                         u64 obd_md_valid, const char *name, size_t buf_size,
1354                         struct ptlrpc_request **req)
1355 {
1356         struct obd_device *obd = exp->exp_obd;
1357         struct lmv_obd *lmv = &obd->u.lmv;
1358         struct lmv_tgt_desc *tgt;
1359         int rc;
1360
1361         ENTRY;
1362
1363         tgt = lmv_fid2tgt(lmv, fid);
1364         if (IS_ERR(tgt))
1365                 RETURN(PTR_ERR(tgt));
1366
1367         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1368
1369         RETURN(rc);
1370 }
1371
1372 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1373                         u64 obd_md_valid, const char *name,
1374                         const void *value, size_t value_size,
1375                         unsigned int xattr_flags, u32 suppgid,
1376                         struct ptlrpc_request **req)
1377 {
1378         struct obd_device *obd = exp->exp_obd;
1379         struct lmv_obd *lmv = &obd->u.lmv;
1380         struct lmv_tgt_desc *tgt;
1381         int rc;
1382
1383         ENTRY;
1384
1385         tgt = lmv_fid2tgt(lmv, fid);
1386         if (IS_ERR(tgt))
1387                 RETURN(PTR_ERR(tgt));
1388
1389         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1390                          value, value_size, xattr_flags, suppgid, req);
1391
1392         RETURN(rc);
1393 }
1394
1395 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1396                        struct ptlrpc_request **request)
1397 {
1398         struct obd_device *obd = exp->exp_obd;
1399         struct lmv_obd *lmv = &obd->u.lmv;
1400         struct lmv_tgt_desc *tgt;
1401         int rc;
1402
1403         ENTRY;
1404
1405         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1406         if (IS_ERR(tgt))
1407                 RETURN(PTR_ERR(tgt));
1408
1409         if (op_data->op_flags & MF_GET_MDT_IDX) {
1410                 op_data->op_mds = tgt->ltd_index;
1411                 RETURN(0);
1412         }
1413
1414         rc = md_getattr(tgt->ltd_exp, op_data, request);
1415
1416         RETURN(rc);
1417 }
1418
1419 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1420 {
1421         struct obd_device *obd = exp->exp_obd;
1422         struct lmv_obd *lmv = &obd->u.lmv;
1423         struct lu_tgt_desc *tgt;
1424
1425         ENTRY;
1426
1427         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1428
1429         /*
1430          * With DNE every object can have two locks in different namespaces:
1431          * lookup lock in space of MDT storing direntry and update/open lock in
1432          * space of MDT storing inode.
1433          */
1434         lmv_foreach_connected_tgt(lmv, tgt)
1435                 md_null_inode(tgt->ltd_exp, fid);
1436
1437         RETURN(0);
1438 }
1439
1440 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1441                      struct md_open_data *mod, struct ptlrpc_request **request)
1442 {
1443         struct obd_device *obd = exp->exp_obd;
1444         struct lmv_obd *lmv = &obd->u.lmv;
1445         struct lmv_tgt_desc *tgt;
1446         int rc;
1447
1448         ENTRY;
1449
1450         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1451         if (IS_ERR(tgt))
1452                 RETURN(PTR_ERR(tgt));
1453
1454         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1455         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1456         RETURN(rc);
1457 }
1458
1459 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
1460 {
1461         struct lu_tgt_desc *tgt;
1462         __u64 total_weight = 0;
1463         __u64 cur_weight = 0;
1464         __u64 rand;
1465         int rc;
1466
1467         ENTRY;
1468
1469         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1470                 RETURN(ERR_PTR(-EAGAIN));
1471
1472         down_write(&lmv->lmv_qos.lq_rw_sem);
1473
1474         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1475                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1476
1477         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1478         if (rc)
1479                 GOTO(unlock, tgt = ERR_PTR(rc));
1480
1481         lmv_foreach_tgt(lmv, tgt) {
1482                 tgt->ltd_qos.ltq_usable = 0;
1483                 if (!tgt->ltd_exp || !tgt->ltd_active)
1484                         continue;
1485
1486                 tgt->ltd_qos.ltq_usable = 1;
1487                 lu_tgt_qos_weight_calc(tgt);
1488                 total_weight += tgt->ltd_qos.ltq_weight;
1489         }
1490
1491         rand = lu_prandom_u64_max(total_weight);
1492
1493         lmv_foreach_connected_tgt(lmv, tgt) {
1494                 if (!tgt->ltd_qos.ltq_usable)
1495                         continue;
1496
1497                 cur_weight += tgt->ltd_qos.ltq_weight;
1498                 if (cur_weight < rand)
1499                         continue;
1500
1501                 *mdt = tgt->ltd_index;
1502                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1503                 GOTO(unlock, rc = 0);
1504         }
1505
1506         /* no proper target found */
1507         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1508 unlock:
1509         up_write(&lmv->lmv_qos.lq_rw_sem);
1510
1511         return tgt;
1512 }
1513
1514 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
1515 {
1516         struct lu_tgt_desc *tgt;
1517         int i;
1518         int index;
1519
1520         ENTRY;
1521
1522         spin_lock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1523         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1524                 index = (i + lmv->lmv_qos_rr_index) %
1525                         lmv->lmv_mdt_descs.ltd_tgts_size;
1526                 tgt = lmv_tgt(lmv, index);
1527                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1528                         continue;
1529
1530                 *mdt = tgt->ltd_index;
1531                 lmv->lmv_qos_rr_index = (*mdt + 1) %
1532                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1533                 spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1534
1535                 RETURN(tgt);
1536         }
1537         spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1538
1539         RETURN(ERR_PTR(-ENODEV));
1540 }
1541
1542 static struct lmv_tgt_desc *
1543 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1544                        const char *name, int namelen, struct lu_fid *fid,
1545                        __u32 *mds, bool post_migrate)
1546 {
1547         struct lmv_tgt_desc *tgt;
1548         const struct lmv_oinfo *oinfo;
1549
1550         if (!lmv_dir_striped(lsm) || !namelen) {
1551                 tgt = lmv_fid2tgt(lmv, fid);
1552                 if (IS_ERR(tgt))
1553                         return tgt;
1554
1555                 *mds = tgt->ltd_index;
1556                 return tgt;
1557         }
1558
1559         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1560                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1561                         return ERR_PTR(-EBADF);
1562                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1563         } else {
1564                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
1565                                                 post_migrate);
1566                 if (IS_ERR(oinfo))
1567                         return ERR_CAST(oinfo);
1568         }
1569
1570         *fid = oinfo->lmo_fid;
1571         *mds = oinfo->lmo_mds;
1572         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1573
1574         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1575
1576         return tgt ? tgt : ERR_PTR(-ENODEV);
1577 }
1578
1579 /**
1580  * Locate MDT of op_data->op_fid1
1581  *
1582  * For striped directory, it will locate the stripe by name hash, if hash_type
1583  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1584  * which is set outside, and if dir is migrating, 'op_data->op_post_migrate'
1585  * indicates whether old or new layout is used to locate.
1586  *
1587  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1588  *
1589  * \param[in] lmv       LMV device
1590  * \param[in] op_data   client MD stack parameters, name, namelen
1591  *                      mds_num etc.
1592  *
1593  * retval               pointer to the lmv_tgt_desc if succeed.
1594  *                      ERR_PTR(errno) if failed.
1595  */
1596 struct lmv_tgt_desc *
1597 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1598 {
1599         struct lmv_stripe_md *lsm = op_data->op_mea1;
1600         struct lmv_oinfo *oinfo;
1601         struct lmv_tgt_desc *tgt;
1602
1603         if (lmv_dir_foreign(lsm))
1604                 return ERR_PTR(-ENODATA);
1605
1606         /* During creating VOLATILE file, it should honor the mdt
1607          * index if the file under striped dir is being restored, see
1608          * ct_restore(). */
1609         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1610             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1611                 tgt = lmv_tgt(lmv, op_data->op_mds);
1612                 if (!tgt)
1613                         return ERR_PTR(-ENODEV);
1614
1615                 if (lmv_dir_striped(lsm)) {
1616                         int i;
1617
1618                         /* refill the right parent fid */
1619                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1620                                 oinfo = &lsm->lsm_md_oinfo[i];
1621                                 if (oinfo->lmo_mds == op_data->op_mds) {
1622                                         op_data->op_fid1 = oinfo->lmo_fid;
1623                                         break;
1624                                 }
1625                         }
1626
1627                         if (i == lsm->lsm_md_stripe_count)
1628                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1629                 }
1630         } else if (lmv_dir_bad_hash(lsm)) {
1631                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1632                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1633
1634                 op_data->op_fid1 = oinfo->lmo_fid;
1635                 op_data->op_mds = oinfo->lmo_mds;
1636                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1637                 if (!tgt)
1638                         return ERR_PTR(-ENODEV);
1639         } else {
1640                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1641                                 op_data->op_name, op_data->op_namelen,
1642                                 &op_data->op_fid1, &op_data->op_mds,
1643                                 op_data->op_post_migrate);
1644         }
1645
1646         return tgt;
1647 }
1648
1649 /* Locate MDT of op_data->op_fid2 for link/rename */
1650 static struct lmv_tgt_desc *
1651 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1652 {
1653         struct lmv_tgt_desc *tgt;
1654         int rc;
1655
1656         LASSERT(op_data->op_name);
1657         if (lmv_dir_migrating(op_data->op_mea2)) {
1658                 struct lu_fid fid1 = op_data->op_fid1;
1659                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1660                 struct ptlrpc_request *request = NULL;
1661
1662                 /*
1663                  * avoid creating new file under old layout of migrating
1664                  * directory, check it here.
1665                  */
1666                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1667                                 op_data->op_name, op_data->op_namelen,
1668                                 &op_data->op_fid2, &op_data->op_mds, false);
1669                 if (IS_ERR(tgt))
1670                         RETURN(tgt);
1671
1672                 op_data->op_fid1 = op_data->op_fid2;
1673                 op_data->op_mea1 = op_data->op_mea2;
1674                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1675                 op_data->op_fid1 = fid1;
1676                 op_data->op_mea1 = lsm1;
1677                 if (!rc) {
1678                         ptlrpc_req_finished(request);
1679                         RETURN(ERR_PTR(-EEXIST));
1680                 }
1681
1682                 if (rc != -ENOENT)
1683                         RETURN(ERR_PTR(rc));
1684         }
1685
1686         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1687                                 op_data->op_namelen, &op_data->op_fid2,
1688                                 &op_data->op_mds, true);
1689 }
1690
1691 int lmv_migrate_existence_check(struct lmv_obd *lmv, struct md_op_data *op_data)
1692 {
1693         struct lu_tgt_desc *tgt;
1694         struct ptlrpc_request *request;
1695         int rc;
1696
1697         LASSERT(lmv_dir_migrating(op_data->op_mea1));
1698
1699         tgt = lmv_locate_tgt(lmv, op_data);
1700         if (IS_ERR(tgt))
1701                 return PTR_ERR(tgt);
1702
1703         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1704         if (!rc) {
1705                 ptlrpc_req_finished(request);
1706                 return -EEXIST;
1707         }
1708
1709         return rc;
1710 }
1711
1712 /* mkdir by QoS in two cases:
1713  * 1. 'lfs mkdir -i -1'
1714  * 2. parent default LMV master_mdt_index is -1
1715  *
1716  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1717  * directories under striped directory.
1718  */
1719 static inline bool lmv_op_qos_mkdir(const struct md_op_data *op_data)
1720 {
1721         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1722         const struct lmv_user_md *lum = op_data->op_data;
1723
1724         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1725                 return false;
1726
1727         if (lmv_dir_striped(op_data->op_mea1))
1728                 return false;
1729
1730         if (op_data->op_cli_flags & CLI_SET_MEA && lum &&
1731             (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1732              le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1733             le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT)
1734                 return true;
1735
1736         if (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT)
1737                 return true;
1738
1739         return false;
1740 }
1741
1742 /* 'lfs mkdir -i <specific_MDT>' */
1743 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1744 {
1745         const struct lmv_user_md *lum = op_data->op_data;
1746
1747         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1748                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1749                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1750                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1751                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1752 }
1753
1754 /* parent default LMV master_mdt_index is not -1. */
1755 static inline bool
1756 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1757 {
1758         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1759                op_data->op_default_mea1 &&
1760                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1761                         LMV_OFFSET_DEFAULT;
1762 }
1763 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1764                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1765                 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1766                 struct ptlrpc_request **request)
1767 {
1768         struct obd_device *obd = exp->exp_obd;
1769         struct lmv_obd *lmv = &obd->u.lmv;
1770         struct lmv_tgt_desc *tgt;
1771         int rc;
1772
1773         ENTRY;
1774
1775         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1776                 RETURN(-EIO);
1777
1778         if (lmv_dir_bad_hash(op_data->op_mea1))
1779                 RETURN(-EBADF);
1780
1781         if (lmv_dir_migrating(op_data->op_mea1)) {
1782                 /*
1783                  * if parent is migrating, create() needs to lookup existing
1784                  * name in both old and new layout, check old layout on client.
1785                  */
1786                 rc = lmv_migrate_existence_check(lmv, op_data);
1787                 if (rc != -ENOENT)
1788                         RETURN(rc);
1789
1790                 op_data->op_post_migrate = true;
1791         }
1792
1793         tgt = lmv_locate_tgt(lmv, op_data);
1794         if (IS_ERR(tgt))
1795                 RETURN(PTR_ERR(tgt));
1796
1797         if (lmv_op_qos_mkdir(op_data)) {
1798                 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds);
1799                 if (tgt == ERR_PTR(-EAGAIN))
1800                         tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1801                 /*
1802                  * only update statfs after QoS mkdir, this means the cached
1803                  * statfs may be stale, and current mkdir may not follow QoS
1804                  * accurately, but it's not serious, and avoids periodic statfs
1805                  * when client doesn't mkdir by QoS.
1806                  */
1807                 if (!IS_ERR(tgt))
1808                         lmv_statfs_check_update(obd, tgt);
1809         } else if (lmv_op_user_specific_mkdir(op_data)) {
1810                 struct lmv_user_md *lum = op_data->op_data;
1811
1812                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1813                 tgt = lmv_tgt(lmv, op_data->op_mds);
1814                 if (!tgt)
1815                         RETURN(-ENODEV);
1816         } else if (lmv_op_default_specific_mkdir(op_data)) {
1817                 op_data->op_mds =
1818                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1819                 tgt = lmv_tgt(lmv, op_data->op_mds);
1820                 if (!tgt)
1821                         RETURN(-ENODEV);
1822         }
1823
1824         if (IS_ERR(tgt))
1825                 RETURN(PTR_ERR(tgt));
1826
1827         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1828         if (rc)
1829                 RETURN(rc);
1830
1831         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1832                 (int)op_data->op_namelen, op_data->op_name,
1833                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1834                 op_data->op_mds);
1835
1836         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1837         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1838                        cap_effective, rdev, request);
1839         if (rc == 0) {
1840                 if (*request == NULL)
1841                         RETURN(rc);
1842                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1843         }
1844         RETURN(rc);
1845 }
1846
1847 static int
1848 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1849             const union ldlm_policy_data *policy, struct md_op_data *op_data,
1850             struct lustre_handle *lockh, __u64 extra_lock_flags)
1851 {
1852         struct obd_device *obd = exp->exp_obd;
1853         struct lmv_obd *lmv = &obd->u.lmv;
1854         struct lmv_tgt_desc *tgt;
1855         int rc;
1856
1857         ENTRY;
1858
1859         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1860
1861         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1862         if (IS_ERR(tgt))
1863                 RETURN(PTR_ERR(tgt));
1864
1865         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1866                PFID(&op_data->op_fid1), tgt->ltd_index);
1867
1868         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1869                         extra_lock_flags);
1870
1871         RETURN(rc);
1872 }
1873
1874 int
1875 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1876                  struct ptlrpc_request **preq)
1877 {
1878         struct obd_device *obd = exp->exp_obd;
1879         struct lmv_obd *lmv = &obd->u.lmv;
1880         struct lmv_tgt_desc *tgt;
1881         struct mdt_body *body;
1882         int rc;
1883
1884         ENTRY;
1885
1886 retry:
1887         tgt = lmv_locate_tgt(lmv, op_data);
1888         if (IS_ERR(tgt))
1889                 RETURN(PTR_ERR(tgt));
1890
1891         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1892                 (int)op_data->op_namelen, op_data->op_name,
1893                 PFID(&op_data->op_fid1), tgt->ltd_index);
1894
1895         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1896         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1897                 ptlrpc_req_finished(*preq);
1898                 *preq = NULL;
1899                 goto retry;
1900         }
1901
1902         if (rc)
1903                 RETURN(rc);
1904
1905         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1906         LASSERT(body != NULL);
1907
1908         if (body->mbo_valid & OBD_MD_MDS) {
1909                 op_data->op_fid1 = body->mbo_fid1;
1910                 op_data->op_valid |= OBD_MD_FLCROSSREF;
1911                 op_data->op_namelen = 0;
1912                 op_data->op_name = NULL;
1913
1914                 ptlrpc_req_finished(*preq);
1915                 *preq = NULL;
1916
1917                 goto retry;
1918         }
1919
1920         RETURN(rc);
1921 }
1922
1923 #define md_op_data_fid(op_data, fl)                     \
1924         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1925          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1926          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1927          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1928          NULL)
1929
1930 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1931                             struct md_op_data *op_data, __u32 op_tgt,
1932                             enum ldlm_mode mode, int bits, int flag)
1933 {
1934         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1935         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1936         union ldlm_policy_data policy = { { 0 } };
1937         int rc = 0;
1938         ENTRY;
1939
1940         if (!fid_is_sane(fid))
1941                 RETURN(0);
1942
1943         if (tgt == NULL) {
1944                 tgt = lmv_fid2tgt(lmv, fid);
1945                 if (IS_ERR(tgt))
1946                         RETURN(PTR_ERR(tgt));
1947         }
1948
1949         if (tgt->ltd_index != op_tgt) {
1950                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1951                 policy.l_inodebits.bits = bits;
1952                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1953                                       mode, LCF_ASYNC, NULL);
1954         } else {
1955                 CDEBUG(D_INODE,
1956                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
1957                        op_tgt, PFID(fid));
1958                 op_data->op_flags |= flag;
1959                 rc = 0;
1960         }
1961
1962         RETURN(rc);
1963 }
1964
1965 /*
1966  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1967  * op_data->op_fid2
1968  */
1969 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1970                     struct ptlrpc_request **request)
1971 {
1972         struct obd_device       *obd = exp->exp_obd;
1973         struct lmv_obd          *lmv = &obd->u.lmv;
1974         struct lmv_tgt_desc     *tgt;
1975         int                      rc;
1976         ENTRY;
1977
1978         LASSERT(op_data->op_namelen != 0);
1979
1980         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1981                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
1982                op_data->op_name, PFID(&op_data->op_fid1));
1983
1984         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
1985         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
1986         op_data->op_cap = cfs_curproc_cap_pack();
1987
1988         tgt = lmv_locate_tgt2(lmv, op_data);
1989         if (IS_ERR(tgt))
1990                 RETURN(PTR_ERR(tgt));
1991
1992         /*
1993          * Cancel UPDATE lock on child (fid1).
1994          */
1995         op_data->op_flags |= MF_MDC_CANCEL_FID2;
1996         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
1997                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1998         if (rc != 0)
1999                 RETURN(rc);
2000
2001         rc = md_link(tgt->ltd_exp, op_data, request);
2002
2003         RETURN(rc);
2004 }
2005
2006 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2007                         const char *name, size_t namelen,
2008                         struct ptlrpc_request **request)
2009 {
2010         struct obd_device *obd = exp->exp_obd;
2011         struct lmv_obd *lmv = &obd->u.lmv;
2012         struct lmv_stripe_md *lsm = op_data->op_mea1;
2013         struct lmv_tgt_desc *parent_tgt;
2014         struct lmv_tgt_desc *sp_tgt;
2015         struct lmv_tgt_desc *tp_tgt = NULL;
2016         struct lmv_tgt_desc *child_tgt;
2017         struct lmv_tgt_desc *tgt;
2018         struct lu_fid target_fid;
2019         int rc;
2020
2021         ENTRY;
2022
2023         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2024
2025         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2026                PFID(&op_data->op_fid1), (int)namelen, name);
2027
2028         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2029         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2030         op_data->op_cap = cfs_curproc_cap_pack();
2031
2032         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2033         if (IS_ERR(parent_tgt))
2034                 RETURN(PTR_ERR(parent_tgt));
2035
2036         if (lmv_dir_striped(lsm)) {
2037                 __u32 hash_type = lsm->lsm_md_hash_type;
2038                 __u32 stripe_count = lsm->lsm_md_stripe_count;
2039
2040                 /*
2041                  * old stripes are appended after new stripes for migrating
2042                  * directory.
2043                  */
2044                 if (lmv_dir_migrating(lsm)) {
2045                         hash_type = lsm->lsm_md_migrate_hash;
2046                         stripe_count -= lsm->lsm_md_migrate_offset;
2047                 }
2048
2049                 rc = lmv_name_to_stripe_index(hash_type, stripe_count, name,
2050                                               namelen);
2051                 if (rc < 0)
2052                         RETURN(rc);
2053
2054                 if (lmv_dir_migrating(lsm))
2055                         rc += lsm->lsm_md_migrate_offset;
2056
2057                 /* save it in fid4 temporarily for early cancel */
2058                 op_data->op_fid4 = lsm->lsm_md_oinfo[rc].lmo_fid;
2059                 sp_tgt = lmv_tgt(lmv, lsm->lsm_md_oinfo[rc].lmo_mds);
2060                 if (!sp_tgt)
2061                         RETURN(-ENODEV);
2062
2063                 /*
2064                  * if parent is being migrated too, fill op_fid2 with target
2065                  * stripe fid, otherwise the target stripe is not created yet.
2066                  */
2067                 if (lmv_dir_migrating(lsm)) {
2068                         hash_type = lsm->lsm_md_hash_type &
2069                                     ~LMV_HASH_FLAG_MIGRATION;
2070                         stripe_count = lsm->lsm_md_migrate_offset;
2071
2072                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
2073                                                       name, namelen);
2074                         if (rc < 0)
2075                                 RETURN(rc);
2076
2077                         op_data->op_fid2 = lsm->lsm_md_oinfo[rc].lmo_fid;
2078                         tp_tgt = lmv_tgt(lmv, lsm->lsm_md_oinfo[rc].lmo_mds);
2079                         if (!tp_tgt)
2080                                 RETURN(-ENODEV);
2081                 }
2082         } else {
2083                 sp_tgt = parent_tgt;
2084         }
2085
2086         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2087         if (IS_ERR(child_tgt))
2088                 RETURN(PTR_ERR(child_tgt));
2089
2090         /* for directory, migrate to MDT specified by lum_stripe_offset;
2091          * otherwise migrate to the target stripe of parent, but parent
2092          * directory may have finished migration (normally current file too),
2093          * allocate FID on MDT lum_stripe_offset, and server will check
2094          * whether file was migrated already.
2095          */
2096         if (S_ISDIR(op_data->op_mode) || !tp_tgt) {
2097                 struct lmv_user_md *lum = op_data->op_data;
2098
2099                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2100         } else  {
2101                 op_data->op_mds = tp_tgt->ltd_index;
2102         }
2103         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2104         if (rc)
2105                 RETURN(rc);
2106
2107         /*
2108          * for directory, send migrate request to the MDT where the object will
2109          * be migrated to, because we can't create a striped directory remotely.
2110          *
2111          * otherwise, send to the MDT where source is located because regular
2112          * file may open lease.
2113          *
2114          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2115          * backward compatibility.
2116          */
2117         if (S_ISDIR(op_data->op_mode) &&
2118             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2119                 tgt = lmv_fid2tgt(lmv, &target_fid);
2120                 if (IS_ERR(tgt))
2121                         RETURN(PTR_ERR(tgt));
2122         } else {
2123                 tgt = child_tgt;
2124         }
2125
2126         /* cancel UPDATE lock of parent master object */
2127         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2128                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2129         if (rc)
2130                 RETURN(rc);
2131
2132         /* cancel UPDATE lock of source parent */
2133         if (sp_tgt != parent_tgt) {
2134                 /*
2135                  * migrate RPC packs master object FID, because we can only pack
2136                  * two FIDs in reint RPC, but MDS needs to know both source
2137                  * parent and target parent, and it will obtain them from master
2138                  * FID and LMV, the other FID in RPC is kept for target.
2139                  *
2140                  * since this FID is not passed to MDC, cancel it anyway.
2141                  */
2142                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2143                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2144                 if (rc)
2145                         RETURN(rc);
2146
2147                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2148         }
2149         op_data->op_fid4 = target_fid;
2150
2151         /* cancel UPDATE locks of target parent */
2152         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2153                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2154         if (rc)
2155                 RETURN(rc);
2156
2157         /* cancel LOOKUP lock of source if source is remote object */
2158         if (child_tgt != sp_tgt) {
2159                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2160                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2161                                       MF_MDC_CANCEL_FID3);
2162                 if (rc)
2163                         RETURN(rc);
2164         }
2165
2166         /* cancel ELC locks of source */
2167         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2168                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2169         if (rc)
2170                 RETURN(rc);
2171
2172         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2173
2174         RETURN(rc);
2175 }
2176
2177 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2178                       const char *old, size_t oldlen,
2179                       const char *new, size_t newlen,
2180                       struct ptlrpc_request **request)
2181 {
2182         struct obd_device *obd = exp->exp_obd;
2183         struct lmv_obd *lmv = &obd->u.lmv;
2184         struct lmv_tgt_desc *sp_tgt;
2185         struct lmv_tgt_desc *tp_tgt = NULL;
2186         struct lmv_tgt_desc *src_tgt = NULL;
2187         struct lmv_tgt_desc *tgt;
2188         struct mdt_body *body;
2189         int rc;
2190
2191         ENTRY;
2192
2193         LASSERT(oldlen != 0);
2194
2195         if (op_data->op_cli_flags & CLI_MIGRATE) {
2196                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2197                 RETURN(rc);
2198         }
2199
2200         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2201         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2202         op_data->op_cap = cfs_curproc_cap_pack();
2203
2204         op_data->op_name = new;
2205         op_data->op_namelen = newlen;
2206
2207         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2208         if (IS_ERR(tp_tgt))
2209                 RETURN(PTR_ERR(tp_tgt));
2210
2211         /* Since the target child might be destroyed, and it might become
2212          * orphan, and we can only check orphan on the local MDT right now, so
2213          * we send rename request to the MDT where target child is located. If
2214          * target child does not exist, then it will send the request to the
2215          * target parent */
2216         if (fid_is_sane(&op_data->op_fid4)) {
2217                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2218                 if (IS_ERR(tgt))
2219                         RETURN(PTR_ERR(tgt));
2220         } else {
2221                 tgt = tp_tgt;
2222         }
2223
2224         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2225
2226         /* cancel UPDATE locks of target parent */
2227         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2228                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2229         if (rc != 0)
2230                 RETURN(rc);
2231
2232         if (fid_is_sane(&op_data->op_fid4)) {
2233                 /* cancel LOOKUP lock of target on target parent */
2234                 if (tgt != tp_tgt) {
2235                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2236                                               tgt->ltd_index, LCK_EX,
2237                                               MDS_INODELOCK_LOOKUP,
2238                                               MF_MDC_CANCEL_FID4);
2239                         if (rc != 0)
2240                                 RETURN(rc);
2241                 }
2242         }
2243
2244         if (fid_is_sane(&op_data->op_fid3)) {
2245                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2246                 if (IS_ERR(src_tgt))
2247                         RETURN(PTR_ERR(src_tgt));
2248
2249                 /* cancel ELC locks of source */
2250                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2251                                       LCK_EX, MDS_INODELOCK_ELC,
2252                                       MF_MDC_CANCEL_FID3);
2253                 if (rc != 0)
2254                         RETURN(rc);
2255         }
2256
2257         op_data->op_name = old;
2258         op_data->op_namelen = oldlen;
2259 retry:
2260         sp_tgt = lmv_locate_tgt(lmv, op_data);
2261         if (IS_ERR(sp_tgt))
2262                 RETURN(PTR_ERR(sp_tgt));
2263
2264         /* cancel UPDATE locks of source parent */
2265         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2266                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2267         if (rc != 0)
2268                 RETURN(rc);
2269
2270         if (fid_is_sane(&op_data->op_fid3)) {
2271                 /* cancel LOOKUP lock of source on source parent */
2272                 if (src_tgt != sp_tgt) {
2273                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2274                                               tgt->ltd_index, LCK_EX,
2275                                               MDS_INODELOCK_LOOKUP,
2276                                               MF_MDC_CANCEL_FID3);
2277                         if (rc != 0)
2278                                 RETURN(rc);
2279                 }
2280         }
2281
2282 rename:
2283         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2284                 PFID(&op_data->op_fid1), (int)oldlen, old,
2285                 PFID(&op_data->op_fid2), (int)newlen, new);
2286
2287         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2288                         request);
2289         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2290                 ptlrpc_req_finished(*request);
2291                 *request = NULL;
2292                 goto retry;
2293         }
2294
2295         if (rc && rc != -EXDEV)
2296                 RETURN(rc);
2297
2298         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2299         if (body == NULL)
2300                 RETURN(-EPROTO);
2301
2302         /* Not cross-ref case, just get out of here. */
2303         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2304                 RETURN(rc);
2305
2306         op_data->op_fid4 = body->mbo_fid1;
2307
2308         ptlrpc_req_finished(*request);
2309         *request = NULL;
2310
2311         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2312         if (IS_ERR(tgt))
2313                 RETURN(PTR_ERR(tgt));
2314
2315         if (fid_is_sane(&op_data->op_fid4)) {
2316                 /* cancel LOOKUP lock of target on target parent */
2317                 if (tgt != tp_tgt) {
2318                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2319                                               tgt->ltd_index, LCK_EX,
2320                                               MDS_INODELOCK_LOOKUP,
2321                                               MF_MDC_CANCEL_FID4);
2322                         if (rc != 0)
2323                                 RETURN(rc);
2324                 }
2325         }
2326
2327         goto rename;
2328 }
2329
2330 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2331                        void *ea, size_t ealen, struct ptlrpc_request **request)
2332 {
2333         struct obd_device *obd = exp->exp_obd;
2334         struct lmv_obd *lmv = &obd->u.lmv;
2335         struct lmv_tgt_desc *tgt;
2336         int rc = 0;
2337
2338         ENTRY;
2339
2340         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2341                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2342                op_data->op_xvalid);
2343
2344         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2345         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2346         if (IS_ERR(tgt))
2347                 RETURN(PTR_ERR(tgt));
2348
2349         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2350
2351         RETURN(rc);
2352 }
2353
2354 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2355                      struct ptlrpc_request **request)
2356 {
2357         struct obd_device *obd = exp->exp_obd;
2358         struct lmv_obd *lmv = &obd->u.lmv;
2359         struct lmv_tgt_desc *tgt;
2360         int rc;
2361
2362         ENTRY;
2363
2364         tgt = lmv_fid2tgt(lmv, fid);
2365         if (IS_ERR(tgt))
2366                 RETURN(PTR_ERR(tgt));
2367
2368         rc = md_fsync(tgt->ltd_exp, fid, request);
2369         RETURN(rc);
2370 }
2371
2372 struct stripe_dirent {
2373         struct page             *sd_page;
2374         struct lu_dirpage       *sd_dp;
2375         struct lu_dirent        *sd_ent;
2376         bool                     sd_eof;
2377 };
2378
2379 struct lmv_dir_ctxt {
2380         struct lmv_obd          *ldc_lmv;
2381         struct md_op_data       *ldc_op_data;
2382         struct md_callback      *ldc_cb_op;
2383         __u64                    ldc_hash;
2384         int                      ldc_count;
2385         struct stripe_dirent     ldc_stripes[0];
2386 };
2387
2388 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2389 {
2390         if (stripe->sd_page) {
2391                 kunmap(stripe->sd_page);
2392                 put_page(stripe->sd_page);
2393                 stripe->sd_page = NULL;
2394                 stripe->sd_ent = NULL;
2395         }
2396 }
2397
2398 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2399 {
2400         int i;
2401
2402         for (i = 0; i < ctxt->ldc_count; i++)
2403                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2404 }
2405
2406 /* if @ent is dummy, or . .., get next */
2407 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2408                                            struct lu_dirent *ent,
2409                                            int stripe_index)
2410 {
2411         for (; ent; ent = lu_dirent_next(ent)) {
2412                 /* Skip dummy entry */
2413                 if (le16_to_cpu(ent->lde_namelen) == 0)
2414                         continue;
2415
2416                 /* skip . and .. for other stripes */
2417                 if (stripe_index &&
2418                     (strncmp(ent->lde_name, ".",
2419                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2420                      strncmp(ent->lde_name, "..",
2421                              le16_to_cpu(ent->lde_namelen)) == 0))
2422                         continue;
2423
2424                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2425                         break;
2426         }
2427
2428         return ent;
2429 }
2430
2431 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2432                                             struct stripe_dirent *stripe,
2433                                             int stripe_index)
2434 {
2435         struct md_op_data *op_data = ctxt->ldc_op_data;
2436         struct lmv_oinfo *oinfo;
2437         struct lu_fid fid = op_data->op_fid1;
2438         struct inode *inode = op_data->op_data;
2439         struct lmv_tgt_desc *tgt;
2440         struct lu_dirent *ent = stripe->sd_ent;
2441         __u64 hash = ctxt->ldc_hash;
2442         int rc = 0;
2443
2444         ENTRY;
2445
2446         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2447         LASSERT(!ent);
2448
2449         do {
2450                 if (stripe->sd_page) {
2451                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2452
2453                         /* @hash should be the last dirent hash */
2454                         LASSERTF(hash <= end,
2455                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2456                                  ctxt, stripe, hash, end);
2457                         /* unload last page */
2458                         stripe_dirent_unload(stripe);
2459                         /* eof */
2460                         if (end == MDS_DIR_END_OFF) {
2461                                 stripe->sd_eof = true;
2462                                 break;
2463                         }
2464                         hash = end;
2465                 }
2466
2467                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2468                 if (!oinfo->lmo_root) {
2469                         rc = -ENOENT;
2470                         break;
2471                 }
2472
2473                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2474                 if (!tgt) {
2475                         rc = -ENODEV;
2476                         break;
2477                 }
2478
2479                 /* op_data is shared by stripes, reset after use */
2480                 op_data->op_fid1 = oinfo->lmo_fid;
2481                 op_data->op_fid2 = oinfo->lmo_fid;
2482                 op_data->op_data = oinfo->lmo_root;
2483
2484                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2485                                   &stripe->sd_page);
2486
2487                 op_data->op_fid1 = fid;
2488                 op_data->op_fid2 = fid;
2489                 op_data->op_data = inode;
2490
2491                 if (rc)
2492                         break;
2493
2494                 stripe->sd_dp = page_address(stripe->sd_page);
2495                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2496                                         stripe_index);
2497                 /* in case a page filled with ., .. and dummy, read next */
2498         } while (!ent);
2499
2500         stripe->sd_ent = ent;
2501         if (rc) {
2502                 LASSERT(!ent);
2503                 /* treat error as eof, so dir can be partially accessed */
2504                 stripe->sd_eof = true;
2505                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2506                               "directory is partially accessed!\n",
2507                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2508                               rc);
2509         }
2510
2511         RETURN(ent);
2512 }
2513
2514 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2515 {
2516         struct obd_device *obd = exp->exp_obd;
2517         struct lmv_obd *lmv = &obd->u.lmv;
2518         struct lmv_tgt_desc *tgt;
2519         int rc;
2520
2521         ENTRY;
2522
2523         rc = lmv_check_connect(obd);
2524         if (rc != 0)
2525                 RETURN(rc);
2526
2527         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2528         if (IS_ERR(tgt))
2529                 RETURN(PTR_ERR(tgt));
2530
2531         data->op_flags |= MF_MDC_CANCEL_FID1;
2532         rc = md_file_resync(tgt->ltd_exp, data);
2533         RETURN(rc);
2534 }
2535
2536 /**
2537  * Get dirent with the closest hash for striped directory
2538  *
2539  * This function will search the dir entry, whose hash value is the
2540  * closest(>=) to hash from all of sub-stripes, and it is only being called
2541  * for striped directory.
2542  *
2543  * \param[in] ctxt              dir read context
2544  *
2545  * \retval                      dirent get the entry successfully
2546  *                              NULL does not get the entry, normally it means
2547  *                              it reaches the end of the directory, while read
2548  *                              stripe dirent error is ignored to allow partial
2549  *                              access.
2550  */
2551 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2552 {
2553         struct stripe_dirent *stripe;
2554         struct lu_dirent *ent = NULL;
2555         int i;
2556         int min = -1;
2557
2558         /* TODO: optimize with k-way merge sort */
2559         for (i = 0; i < ctxt->ldc_count; i++) {
2560                 stripe = &ctxt->ldc_stripes[i];
2561                 if (stripe->sd_eof)
2562                         continue;
2563
2564                 if (!stripe->sd_ent) {
2565                         stripe_dirent_load(ctxt, stripe, i);
2566                         if (!stripe->sd_ent) {
2567                                 LASSERT(stripe->sd_eof);
2568                                 continue;
2569                         }
2570                 }
2571
2572                 if (min == -1 ||
2573                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2574                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2575                         min = i;
2576                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2577                             ctxt->ldc_hash)
2578                                 break;
2579                 }
2580         }
2581
2582         if (min != -1) {
2583                 stripe = &ctxt->ldc_stripes[min];
2584                 ent = stripe->sd_ent;
2585                 /* pop found dirent */
2586                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2587                                                    min);
2588         }
2589
2590         return ent;
2591 }
2592
2593 /**
2594  * Build dir entry page for striped directory
2595  *
2596  * This function gets one entry by @offset from a striped directory. It will
2597  * read entries from all of stripes, and choose one closest to the required
2598  * offset(&offset). A few notes
2599  * 1. skip . and .. for non-zero stripes, because there can only have one .
2600  * and .. in a directory.
2601  * 2. op_data will be shared by all of stripes, instead of allocating new
2602  * one, so need to restore before reusing.
2603  *
2604  * \param[in] exp       obd export refer to LMV
2605  * \param[in] op_data   hold those MD parameters of read_entry
2606  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2607  * \param[in] offset    starting hash offset
2608  * \param[out] ppage    the page holding the entry. Note: because the entry
2609  *                      will be accessed in upper layer, so we need hold the
2610  *                      page until the usages of entry is finished, see
2611  *                      ll_dir_entry_next.
2612  *
2613  * retval               =0 if get entry successfully
2614  *                      <0 cannot get entry
2615  */
2616 static int lmv_striped_read_page(struct obd_export *exp,
2617                                  struct md_op_data *op_data,
2618                                  struct md_callback *cb_op,
2619                                  __u64 offset, struct page **ppage)
2620 {
2621         struct page *page = NULL;
2622         struct lu_dirpage *dp;
2623         void *start;
2624         struct lu_dirent *ent;
2625         struct lu_dirent *last_ent;
2626         int stripe_count;
2627         struct lmv_dir_ctxt *ctxt;
2628         struct lu_dirent *next = NULL;
2629         __u16 ent_size;
2630         size_t left_bytes;
2631         int rc = 0;
2632         ENTRY;
2633
2634         /* Allocate a page and read entries from all of stripes and fill
2635          * the page by hash order */
2636         page = alloc_page(GFP_KERNEL);
2637         if (!page)
2638                 RETURN(-ENOMEM);
2639
2640         /* Initialize the entry page */
2641         dp = kmap(page);
2642         memset(dp, 0, sizeof(*dp));
2643         dp->ldp_hash_start = cpu_to_le64(offset);
2644
2645         start = dp + 1;
2646         left_bytes = PAGE_SIZE - sizeof(*dp);
2647         ent = start;
2648         last_ent = ent;
2649
2650         /* initalize dir read context */
2651         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2652         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2653         if (!ctxt)
2654                 GOTO(free_page, rc = -ENOMEM);
2655         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2656         ctxt->ldc_op_data = op_data;
2657         ctxt->ldc_cb_op = cb_op;
2658         ctxt->ldc_hash = offset;
2659         ctxt->ldc_count = stripe_count;
2660
2661         while (1) {
2662                 next = lmv_dirent_next(ctxt);
2663
2664                 /* end of directory */
2665                 if (!next) {
2666                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2667                         break;
2668                 }
2669                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2670
2671                 ent_size = le16_to_cpu(next->lde_reclen);
2672
2673                 /* the last entry lde_reclen is 0, but it might not be the last
2674                  * one of this temporay dir page */
2675                 if (!ent_size)
2676                         ent_size = lu_dirent_calc_size(
2677                                         le16_to_cpu(next->lde_namelen),
2678                                         le32_to_cpu(next->lde_attrs));
2679                 /* page full */
2680                 if (ent_size > left_bytes)
2681                         break;
2682
2683                 memcpy(ent, next, ent_size);
2684
2685                 /* Replace . with master FID and Replace .. with the parent FID
2686                  * of master object */
2687                 if (strncmp(ent->lde_name, ".",
2688                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2689                     le16_to_cpu(ent->lde_namelen) == 1)
2690                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2691                 else if (strncmp(ent->lde_name, "..",
2692                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2693                            le16_to_cpu(ent->lde_namelen) == 2)
2694                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2695
2696                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2697                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2698                        le64_to_cpu(ent->lde_hash));
2699
2700                 left_bytes -= ent_size;
2701                 ent->lde_reclen = cpu_to_le16(ent_size);
2702                 last_ent = ent;
2703                 ent = (void *)ent + ent_size;
2704         };
2705
2706         last_ent->lde_reclen = 0;
2707
2708         if (ent == start)
2709                 dp->ldp_flags |= LDF_EMPTY;
2710         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2711                 dp->ldp_flags |= LDF_COLLIDE;
2712         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2713         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2714
2715         put_lmv_dir_ctxt(ctxt);
2716         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2717
2718         *ppage = page;
2719
2720         RETURN(0);
2721
2722 free_page:
2723         kunmap(page);
2724         __free_page(page);
2725
2726         return rc;
2727 }
2728
2729 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2730                   struct md_callback *cb_op, __u64 offset,
2731                   struct page **ppage)
2732 {
2733         struct obd_device *obd = exp->exp_obd;
2734         struct lmv_obd *lmv = &obd->u.lmv;
2735         struct lmv_tgt_desc *tgt;
2736         int rc;
2737
2738         ENTRY;
2739
2740         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2741                 RETURN(-ENODATA);
2742
2743         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2744                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2745                 RETURN(rc);
2746         }
2747
2748         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2749         if (IS_ERR(tgt))
2750                 RETURN(PTR_ERR(tgt));
2751
2752         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2753
2754         RETURN(rc);
2755 }
2756
2757 /**
2758  * Unlink a file/directory
2759  *
2760  * Unlink a file or directory under the parent dir. The unlink request
2761  * usually will be sent to the MDT where the child is located, but if
2762  * the client does not have the child FID then request will be sent to the
2763  * MDT where the parent is located.
2764  *
2765  * If the parent is a striped directory then it also needs to locate which
2766  * stripe the name of the child is located, and replace the parent FID
2767  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2768  * it will walk through all of sub-stripes until the child is being
2769  * unlinked finally.
2770  *
2771  * \param[in] exp       export refer to LMV
2772  * \param[in] op_data   different parameters transferred beween client
2773  *                      MD stacks, name, namelen, FIDs etc.
2774  *                      op_fid1 is the parent FID, op_fid2 is the child
2775  *                      FID.
2776  * \param[out] request  point to the request of unlink.
2777  *
2778  * retval               0 if succeed
2779  *                      negative errno if failed.
2780  */
2781 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2782                       struct ptlrpc_request **request)
2783 {
2784         struct obd_device *obd = exp->exp_obd;
2785         struct lmv_obd *lmv = &obd->u.lmv;
2786         struct lmv_tgt_desc *tgt;
2787         struct lmv_tgt_desc *parent_tgt;
2788         struct mdt_body *body;
2789         int rc;
2790
2791         ENTRY;
2792
2793         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2794         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2795         op_data->op_cap = cfs_curproc_cap_pack();
2796
2797 retry:
2798         parent_tgt = lmv_locate_tgt(lmv, op_data);
2799         if (IS_ERR(parent_tgt))
2800                 RETURN(PTR_ERR(parent_tgt));
2801
2802         if (likely(!fid_is_zero(&op_data->op_fid2))) {
2803                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2804                 if (IS_ERR(tgt))
2805                         RETURN(PTR_ERR(tgt));
2806         } else {
2807                 tgt = parent_tgt;
2808         }
2809
2810         /*
2811          * If child's fid is given, cancel unused locks for it if it is from
2812          * another export than parent.
2813          *
2814          * LOOKUP lock for child (fid3) should also be cancelled on parent
2815          * tgt_tgt in mdc_unlink().
2816          */
2817         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2818
2819         if (parent_tgt != tgt)
2820                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2821                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2822                                       MF_MDC_CANCEL_FID3);
2823
2824         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2825                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2826         if (rc)
2827                 RETURN(rc);
2828
2829         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2830                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2831                tgt->ltd_index);
2832
2833         rc = md_unlink(tgt->ltd_exp, op_data, request);
2834         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2835                 ptlrpc_req_finished(*request);
2836                 *request = NULL;
2837                 goto retry;
2838         }
2839
2840         if (rc != -EREMOTE)
2841                 RETURN(rc);
2842
2843         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2844         if (body == NULL)
2845                 RETURN(-EPROTO);
2846
2847         /* Not cross-ref case, just get out of here. */
2848         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2849                 RETURN(rc);
2850
2851         /* This is a remote object, try remote MDT. */
2852         op_data->op_fid2 = body->mbo_fid1;
2853         ptlrpc_req_finished(*request);
2854         *request = NULL;
2855
2856         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2857         if (IS_ERR(tgt))
2858                 RETURN(PTR_ERR(tgt));
2859
2860         goto retry;
2861 }
2862
2863 static int lmv_precleanup(struct obd_device *obd)
2864 {
2865         ENTRY;
2866         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2867         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2868         lprocfs_obd_cleanup(obd);
2869         lprocfs_free_md_stats(obd);
2870         RETURN(0);
2871 }
2872
2873 /**
2874  * Get by key a value associated with a LMV device.
2875  *
2876  * Dispatch request to lower-layer devices as needed.
2877  *
2878  * \param[in] env               execution environment for this thread
2879  * \param[in] exp               export for the LMV device
2880  * \param[in] keylen            length of key identifier
2881  * \param[in] key               identifier of key to get value for
2882  * \param[in] vallen            size of \a val
2883  * \param[out] val              pointer to storage location for value
2884  * \param[in] lsm               optional striping metadata of object
2885  *
2886  * \retval 0            on success
2887  * \retval negative     negated errno on failure
2888  */
2889 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2890                         __u32 keylen, void *key, __u32 *vallen, void *val)
2891 {
2892         struct obd_device *obd;
2893         struct lmv_obd *lmv;
2894         struct lu_tgt_desc *tgt;
2895         int rc = 0;
2896
2897         ENTRY;
2898
2899         obd = class_exp2obd(exp);
2900         if (obd == NULL) {
2901                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2902                        exp->exp_handle.h_cookie);
2903                 RETURN(-EINVAL);
2904         }
2905
2906         lmv = &obd->u.lmv;
2907         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2908                 LASSERT(*vallen == sizeof(__u32));
2909                 lmv_foreach_connected_tgt(lmv, tgt) {
2910                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2911                                           vallen, val))
2912                                 RETURN(0);
2913                 }
2914                 RETURN(-EINVAL);
2915         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2916                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2917                    KEY_IS(KEY_CONN_DATA)) {
2918                 /*
2919                  * Forwarding this request to first MDS, it should know LOV
2920                  * desc.
2921                  */
2922                 tgt = lmv_tgt(lmv, 0);
2923                 if (!tgt)
2924                         RETURN(-ENODEV);
2925
2926                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
2927                 if (!rc && KEY_IS(KEY_CONN_DATA))
2928                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2929                 RETURN(rc);
2930         } else if (KEY_IS(KEY_TGT_COUNT)) {
2931                 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
2932                 RETURN(0);
2933         }
2934
2935         CDEBUG(D_IOCTL, "Invalid key\n");
2936         RETURN(-EINVAL);
2937 }
2938
2939 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
2940                      int *__rcs, struct ptlrpc_request_set *_set)
2941 {
2942         struct obd_device *obd = class_exp2obd(exp);
2943         struct ptlrpc_request_set *set = _set;
2944         struct lmv_obd *lmv = &obd->u.lmv;
2945         int tgt_count = lmv->lmv_mdt_count;
2946         struct lu_tgt_desc *tgt;
2947         struct fid_array *fat, **fas = NULL;
2948         int i, rc, **rcs = NULL;
2949
2950         if (!set) {
2951                 set = ptlrpc_prep_set();
2952                 if (!set)
2953                         RETURN(-ENOMEM);
2954         }
2955
2956         /* split FIDs by targets */
2957         OBD_ALLOC(fas, sizeof(fas) * tgt_count);
2958         if (fas == NULL)
2959                 GOTO(out, rc = -ENOMEM);
2960         OBD_ALLOC(rcs, sizeof(int *) * tgt_count);
2961         if (rcs == NULL)
2962                 GOTO(out_fas, rc = -ENOMEM);
2963
2964         for (i = 0; i < fa->fa_nr; i++) {
2965                 unsigned int idx;
2966
2967                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
2968                 if (rc) {
2969                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
2970                                PFID(&fa->fa_fids[i]), rc);
2971                         continue;
2972                 }
2973                 LASSERT(idx < tgt_count);
2974                 if (!fas[idx])
2975                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
2976                                   fa_fids[fa->fa_nr]));
2977                 if (!fas[idx])
2978                         GOTO(out, rc = -ENOMEM);
2979                 if (!rcs[idx])
2980                         OBD_ALLOC(rcs[idx], sizeof(int) * fa->fa_nr);
2981                 if (!rcs[idx])
2982                         GOTO(out, rc = -ENOMEM);
2983
2984                 fat = fas[idx];
2985                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
2986         }
2987
2988         lmv_foreach_connected_tgt(lmv, tgt) {
2989                 fat = fas[tgt->ltd_index];
2990                 if (!fat || fat->fa_nr == 0)
2991                         continue;
2992                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
2993         }
2994
2995         rc = ptlrpc_set_wait(NULL, set);
2996         if (rc == 0) {
2997                 int j = 0;
2998                 for (i = 0; i < tgt_count; i++) {
2999                         fat = fas[i];
3000                         if (!fat || fat->fa_nr == 0)
3001                                 continue;
3002                         /* copy FIDs back */
3003                         memcpy(fa->fa_fids + j, fat->fa_fids,
3004                                fat->fa_nr * sizeof(struct lu_fid));
3005                         /* copy rcs back */
3006                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3007                         j += fat->fa_nr;
3008                 }
3009         }
3010         if (set != _set)
3011                 ptlrpc_set_destroy(set);
3012
3013 out:
3014         for (i = 0; i < tgt_count; i++) {
3015                 if (fas && fas[i])
3016                         OBD_FREE(fas[i], offsetof(struct fid_array,
3017                                                 fa_fids[fa->fa_nr]));
3018                 if (rcs && rcs[i])
3019                         OBD_FREE(rcs[i], sizeof(int) * fa->fa_nr);
3020         }
3021         if (rcs)
3022                 OBD_FREE(rcs, sizeof(int *) * tgt_count);
3023 out_fas:
3024         if (fas)
3025                 OBD_FREE(fas, sizeof(fas) * tgt_count);
3026
3027         RETURN(rc);
3028 }
3029
3030 /**
3031  * Asynchronously set by key a value associated with a LMV device.
3032  *
3033  * Dispatch request to lower-layer devices as needed.
3034  *
3035  * \param[in] env       execution environment for this thread
3036  * \param[in] exp       export for the LMV device
3037  * \param[in] keylen    length of key identifier
3038  * \param[in] key       identifier of key to store value for
3039  * \param[in] vallen    size of value to store
3040  * \param[in] val       pointer to data to be stored
3041  * \param[in] set       optional list of related ptlrpc requests
3042  *
3043  * \retval 0            on success
3044  * \retval negative     negated errno on failure
3045  */
3046 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3047                         __u32 keylen, void *key, __u32 vallen, void *val,
3048                         struct ptlrpc_request_set *set)
3049 {
3050         struct lmv_tgt_desc     *tgt = NULL;
3051         struct obd_device       *obd;
3052         struct lmv_obd          *lmv;
3053         int rc = 0;
3054         ENTRY;
3055
3056         obd = class_exp2obd(exp);
3057         if (obd == NULL) {
3058                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3059                        exp->exp_handle.h_cookie);
3060                 RETURN(-EINVAL);
3061         }
3062         lmv = &obd->u.lmv;
3063
3064         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3065             KEY_IS(KEY_DEFAULT_EASIZE)) {
3066                 int err = 0;
3067
3068                 lmv_foreach_connected_tgt(lmv, tgt) {
3069                         err = obd_set_info_async(env, tgt->ltd_exp,
3070                                                  keylen, key, vallen, val, set);
3071                         if (err && rc == 0)
3072                                 rc = err;
3073                 }
3074
3075                 RETURN(rc);
3076         }
3077
3078         RETURN(-EINVAL);
3079 }
3080
3081 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3082                             const struct lmv_mds_md_v1 *lmm1)
3083 {
3084         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3085         int             stripe_count;
3086         int             cplen;
3087         int             i;
3088         int             rc = 0;
3089         ENTRY;
3090
3091         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3092         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3093         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3094         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3095                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3096         else
3097                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3098         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3099         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3100         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3101         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3102                         sizeof(lsm->lsm_md_pool_name));
3103
3104         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3105                 RETURN(-E2BIG);
3106
3107         CDEBUG(D_INFO, "unpack lsm count %d, master %d hash_type %#x "
3108                "layout_version %d\n", lsm->lsm_md_stripe_count,
3109                lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
3110                lsm->lsm_md_layout_version);
3111
3112         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3113         for (i = 0; i < stripe_count; i++) {
3114                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3115                               &lmm1->lmv_stripe_fids[i]);
3116                 /*
3117                  * set default value -1, so lmv_locate_tgt() knows this stripe
3118                  * target is not initialized.
3119                  */
3120                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3121                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3122                         continue;
3123
3124                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3125                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3126                 if (rc == -ENOENT)
3127                         continue;
3128
3129                 if (rc)
3130                         RETURN(rc);
3131
3132                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3133                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3134         }
3135
3136         RETURN(rc);
3137 }
3138
3139 static inline int lmv_unpack_user_md(struct obd_export *exp,
3140                                      struct lmv_stripe_md *lsm,
3141                                      const struct lmv_user_md *lmu)
3142 {
3143         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3144         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3145         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3146         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3147         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3148
3149         return 0;
3150 }
3151
3152 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3153                         const union lmv_mds_md *lmm, size_t lmm_size)
3154 {
3155         struct lmv_stripe_md     *lsm;
3156         int                      lsm_size;
3157         int                      rc;
3158         bool                     allocated = false;
3159         ENTRY;
3160
3161         LASSERT(lsmp != NULL);
3162
3163         lsm = *lsmp;
3164         /* Free memmd */
3165         if (lsm != NULL && lmm == NULL) {
3166                 int i;
3167                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3168
3169                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3170                         size_t lfm_size;
3171
3172                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3173                                                               lfm_value[0]);
3174                         OBD_FREE_LARGE(lfm, lfm_size);
3175                         RETURN(0);
3176                 }
3177
3178                 if (lmv_dir_striped(lsm)) {
3179                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3180                                 if (lsm->lsm_md_oinfo[i].lmo_root)
3181                                         iput(lsm->lsm_md_oinfo[i].lmo_root);
3182                         }
3183                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3184                 } else {
3185                         lsm_size = lmv_stripe_md_size(0);
3186                 }
3187                 OBD_FREE(lsm, lsm_size);
3188                 *lsmp = NULL;
3189                 RETURN(0);
3190         }
3191
3192         /* foreign lmv case */
3193         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3194                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3195
3196                 if (lfm == NULL) {
3197                         OBD_ALLOC_LARGE(lfm, lmm_size);
3198                         if (lfm == NULL)
3199                                 RETURN(-ENOMEM);
3200                         *lsmp = (struct lmv_stripe_md *)lfm;
3201                 }
3202                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3203                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3204                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3205                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3206                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3207                        lfm->lfm_length);
3208                 RETURN(lmm_size);
3209         }
3210
3211         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3212                 RETURN(-EPERM);
3213
3214         /* Unpack memmd */
3215         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3216             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3217                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3218                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3219                        -EIO);
3220                 RETURN(-EIO);
3221         }
3222
3223         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3224                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3225         else
3226                 /**
3227                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3228                  * stripecount should be 0 then.
3229                  */
3230                 lsm_size = lmv_stripe_md_size(0);
3231
3232         if (lsm == NULL) {
3233                 OBD_ALLOC(lsm, lsm_size);
3234                 if (lsm == NULL)
3235                         RETURN(-ENOMEM);
3236                 allocated = true;
3237                 *lsmp = lsm;
3238         }
3239
3240         switch (le32_to_cpu(lmm->lmv_magic)) {
3241         case LMV_MAGIC_V1:
3242                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3243                 break;
3244         case LMV_USER_MAGIC:
3245                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3246                 break;
3247         default:
3248                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3249                        le32_to_cpu(lmm->lmv_magic));
3250                 rc = -EINVAL;
3251                 break;
3252         }
3253
3254         if (rc != 0 && allocated) {
3255                 OBD_FREE(lsm, lsm_size);
3256                 *lsmp = NULL;
3257                 lsm_size = rc;
3258         }
3259         RETURN(lsm_size);
3260 }
3261
3262 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3263 {
3264         lmv_unpackmd(NULL, &lsm, NULL, 0);
3265 }
3266 EXPORT_SYMBOL(lmv_free_memmd);
3267
3268 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3269                              union ldlm_policy_data *policy,
3270                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3271                              void *opaque)
3272 {
3273         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3274         struct lu_tgt_desc *tgt;
3275         int err;
3276         int rc = 0;
3277
3278         ENTRY;
3279
3280         LASSERT(fid != NULL);
3281
3282         lmv_foreach_connected_tgt(lmv, tgt) {
3283                 if (!tgt->ltd_active)
3284                         continue;
3285
3286                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3287                                        opaque);
3288                 if (!rc)
3289                         rc = err;
3290         }
3291         RETURN(rc);
3292 }
3293