Whamcloud - gitweb
LU-8837 lmv: don't use lqr_alloc spinlock in lmv
[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 static 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 ? obd_get_uuid(tgt->ltd_exp) : NULL;
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 static 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 finalize fids factory\n");
515
516         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
517                tgt->ltd_exp->exp_obd->obd_name,
518                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
519
520         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
521         rc = obd_disconnect(tgt->ltd_exp);
522         if (rc) {
523                 if (tgt->ltd_active) {
524                         CERROR("Target %s disconnect error %d\n",
525                                tgt->ltd_uuid.uuid, rc);
526                 }
527         }
528
529         lmv_activate_target(lmv, tgt, 0);
530         tgt->ltd_exp = NULL;
531         RETURN(0);
532 }
533
534 static int lmv_disconnect(struct obd_export *exp)
535 {
536         struct obd_device *obd = class_exp2obd(exp);
537         struct lmv_obd *lmv = &obd->u.lmv;
538         struct lmv_tgt_desc *tgt;
539         int rc;
540
541         ENTRY;
542
543         lmv_foreach_connected_tgt(lmv, tgt)
544                 lmv_disconnect_mdc(obd, tgt);
545
546         if (lmv->lmv_tgts_kobj)
547                 kobject_put(lmv->lmv_tgts_kobj);
548
549         if (!lmv->connected)
550                 class_export_put(exp);
551         rc = class_disconnect(exp);
552         lmv->connected = 0;
553
554         RETURN(rc);
555 }
556
557 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
558                         void __user *uarg)
559 {
560         struct obd_device *obd = class_exp2obd(exp);
561         struct lmv_obd *lmv = &obd->u.lmv;
562         struct getinfo_fid2path *gf;
563         struct lmv_tgt_desc *tgt;
564         struct getinfo_fid2path *remote_gf = NULL;
565         struct lu_fid root_fid;
566         int remote_gf_size = 0;
567         int rc;
568
569         gf = karg;
570         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
571         if (IS_ERR(tgt))
572                 RETURN(PTR_ERR(tgt));
573
574         root_fid = *gf->gf_u.gf_root_fid;
575         LASSERT(fid_is_sane(&root_fid));
576
577 repeat_fid2path:
578         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
579         if (rc != 0 && rc != -EREMOTE)
580                 GOTO(out_fid2path, rc);
581
582         /* If remote_gf != NULL, it means just building the
583          * path on the remote MDT, copy this path segement to gf */
584         if (remote_gf != NULL) {
585                 struct getinfo_fid2path *ori_gf;
586                 char *ptr;
587                 int len;
588
589                 ori_gf = (struct getinfo_fid2path *)karg;
590                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
591                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
592                         GOTO(out_fid2path, rc = -EOVERFLOW);
593
594                 ptr = ori_gf->gf_u.gf_path;
595
596                 len = strlen(gf->gf_u.gf_path);
597                 /* move the current path to the right to release space
598                  * for closer-to-root part */
599                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
600                 memcpy(ptr, gf->gf_u.gf_path, len);
601                 ptr[len] = '/';
602         }
603
604         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
605                tgt->ltd_exp->exp_obd->obd_name,
606                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
607                gf->gf_linkno);
608
609         if (rc == 0)
610                 GOTO(out_fid2path, rc);
611
612         /* sigh, has to go to another MDT to do path building further */
613         if (remote_gf == NULL) {
614                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
615                 OBD_ALLOC(remote_gf, remote_gf_size);
616                 if (remote_gf == NULL)
617                         GOTO(out_fid2path, rc = -ENOMEM);
618                 remote_gf->gf_pathlen = PATH_MAX;
619         }
620
621         if (!fid_is_sane(&gf->gf_fid)) {
622                 CERROR("%s: invalid FID "DFID": rc = %d\n",
623                        tgt->ltd_exp->exp_obd->obd_name,
624                        PFID(&gf->gf_fid), -EINVAL);
625                 GOTO(out_fid2path, rc = -EINVAL);
626         }
627
628         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
629         if (IS_ERR(tgt))
630                 GOTO(out_fid2path, rc = -EINVAL);
631
632         remote_gf->gf_fid = gf->gf_fid;
633         remote_gf->gf_recno = -1;
634         remote_gf->gf_linkno = -1;
635         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
636         *remote_gf->gf_u.gf_root_fid = root_fid;
637         gf = remote_gf;
638         goto repeat_fid2path;
639
640 out_fid2path:
641         if (remote_gf != NULL)
642                 OBD_FREE(remote_gf, remote_gf_size);
643         RETURN(rc);
644 }
645
646 static int lmv_hsm_req_count(struct lmv_obd *lmv,
647                              const struct hsm_user_request *hur,
648                              const struct lmv_tgt_desc *tgt_mds)
649 {
650         struct lmv_tgt_desc *curr_tgt;
651         __u32 i;
652         int nr = 0;
653
654         /* count how many requests must be sent to the given target */
655         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
656                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
657                 if (IS_ERR(curr_tgt))
658                         RETURN(PTR_ERR(curr_tgt));
659                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
660                         nr++;
661         }
662         return nr;
663 }
664
665 static int lmv_hsm_req_build(struct lmv_obd *lmv,
666                               struct hsm_user_request *hur_in,
667                               const struct lmv_tgt_desc *tgt_mds,
668                               struct hsm_user_request *hur_out)
669 {
670         __u32 i, nr_out;
671         struct lmv_tgt_desc *curr_tgt;
672
673         /* build the hsm_user_request for the given target */
674         hur_out->hur_request = hur_in->hur_request;
675         nr_out = 0;
676         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
677                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
678                 if (IS_ERR(curr_tgt))
679                         RETURN(PTR_ERR(curr_tgt));
680                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
681                         hur_out->hur_user_item[nr_out] =
682                                                 hur_in->hur_user_item[i];
683                         nr_out++;
684                 }
685         }
686         hur_out->hur_request.hr_itemcount = nr_out;
687         memcpy(hur_data(hur_out), hur_data(hur_in),
688                hur_in->hur_request.hr_data_len);
689
690         RETURN(0);
691 }
692
693 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
694                                  int len, struct lustre_kernelcomm *lk,
695                                  void __user *uarg)
696 {
697         struct lmv_obd *lmv = &obd->u.lmv;
698         struct lu_tgt_desc *tgt;
699         int rc;
700
701         ENTRY;
702
703         /* unregister request (call from llapi_hsm_copytool_fini) */
704         lmv_foreach_connected_tgt(lmv, tgt)
705                 /* best effort: try to clean as much as possible
706                  * (continue on error) */
707                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
708
709         /* Whatever the result, remove copytool from kuc groups.
710          * Unreached coordinators will get EPIPE on next requests
711          * and will unregister automatically.
712          */
713         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
714
715         RETURN(rc);
716 }
717
718 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
719                                int len, struct lustre_kernelcomm *lk,
720                                void __user *uarg)
721 {
722         struct lmv_obd *lmv = &obd->u.lmv;
723         struct file *filp;
724         bool any_set = false;
725         struct kkuc_ct_data *kcd;
726         size_t kcd_size;
727         struct lu_tgt_desc *tgt;
728         __u32 i;
729         int err;
730         int rc = 0;
731
732         ENTRY;
733
734         filp = fget(lk->lk_wfd);
735         if (!filp)
736                 RETURN(-EBADF);
737
738         if (lk->lk_flags & LK_FLG_DATANR)
739                 kcd_size = offsetof(struct kkuc_ct_data,
740                                     kcd_archives[lk->lk_data_count]);
741         else
742                 kcd_size = sizeof(*kcd);
743
744         OBD_ALLOC(kcd, kcd_size);
745         if (kcd == NULL)
746                 GOTO(err_fput, rc = -ENOMEM);
747
748         kcd->kcd_nr_archives = lk->lk_data_count;
749         if (lk->lk_flags & LK_FLG_DATANR) {
750                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
751                 if (lk->lk_data_count > 0)
752                         memcpy(kcd->kcd_archives, lk->lk_data,
753                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
754         } else {
755                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
756         }
757
758         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
759                                    lk->lk_group, kcd, kcd_size);
760         OBD_FREE(kcd, kcd_size);
761         if (rc)
762                 GOTO(err_fput, rc);
763
764         /* All or nothing: try to register to all MDS.
765          * In case of failure, unregister from previous MDS,
766          * except if it because of inactive target. */
767         lmv_foreach_connected_tgt(lmv, tgt) {
768                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
769                 if (err) {
770                         if (tgt->ltd_active) {
771                                 /* permanent error */
772                                 CERROR("%s: iocontrol MDC %s on MDT"
773                                        " idx %d cmd %x: err = %d\n",
774                                        lmv2obd_dev(lmv)->obd_name,
775                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
776                                        err);
777                                 rc = err;
778                                 lk->lk_flags |= LK_FLG_STOP;
779                                 i = tgt->ltd_index;
780                                 /* unregister from previous MDS */
781                                 lmv_foreach_connected_tgt(lmv, tgt) {
782                                         if (tgt->ltd_index >= i)
783                                                 break;
784
785                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
786                                                       lk, uarg);
787                                 }
788                                 GOTO(err_kkuc_rem, rc);
789                         }
790                         /* else: transient error.
791                          * kuc will register to the missing MDT
792                          * when it is back */
793                 } else {
794                         any_set = true;
795                 }
796         }
797
798         if (!any_set)
799                 /* no registration done: return error */
800                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
801
802         RETURN(0);
803
804 err_kkuc_rem:
805         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
806
807 err_fput:
808         fput(filp);
809         return rc;
810 }
811
812 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
813                          int len, void *karg, void __user *uarg)
814 {
815         struct obd_device *obd = class_exp2obd(exp);
816         struct lmv_obd *lmv = &obd->u.lmv;
817         struct lu_tgt_desc *tgt = NULL;
818         int set = 0;
819         __u32 count = lmv->lmv_mdt_count;
820         int rc = 0;
821
822         ENTRY;
823
824         if (count == 0)
825                 RETURN(-ENOTTY);
826
827         switch (cmd) {
828         case IOC_OBD_STATFS: {
829                 struct obd_ioctl_data *data = karg;
830                 struct obd_device *mdc_obd;
831                 struct obd_statfs stat_buf = {0};
832                 __u32 index;
833
834                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
835
836                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
837                         RETURN(-ENODEV);
838
839                 tgt = lmv_tgt(lmv, index);
840                 if (!tgt)
841                         RETURN(-EAGAIN);
842
843                 if (!tgt->ltd_active)
844                         RETURN(-ENODATA);
845
846                 mdc_obd = class_exp2obd(tgt->ltd_exp);
847                 if (!mdc_obd)
848                         RETURN(-EINVAL);
849
850                 /* copy UUID */
851                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
852                                  min((int) data->ioc_plen2,
853                                      (int) sizeof(struct obd_uuid))))
854                         RETURN(-EFAULT);
855
856                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
857                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
858                                 0);
859                 if (rc)
860                         RETURN(rc);
861                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
862                                  min((int) data->ioc_plen1,
863                                      (int) sizeof(stat_buf))))
864                         RETURN(-EFAULT);
865                 break;
866         }
867         case OBD_IOC_QUOTACTL: {
868                 struct if_quotactl *qctl = karg;
869                 struct obd_quotactl *oqctl;
870
871                 if (qctl->qc_valid == QC_MDTIDX) {
872                         tgt = lmv_tgt(lmv, qctl->qc_idx);
873                 } else if (qctl->qc_valid == QC_UUID) {
874                         lmv_foreach_tgt(lmv, tgt) {
875                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
876                                                      &qctl->obd_uuid))
877                                         continue;
878
879                                 if (!tgt->ltd_exp)
880                                         RETURN(-EINVAL);
881
882                                 break;
883                         }
884                 } else {
885                         RETURN(-EINVAL);
886                 }
887
888                 if (!tgt || !tgt->ltd_exp)
889                         RETURN(-EINVAL);
890
891                 OBD_ALLOC_PTR(oqctl);
892                 if (!oqctl)
893                         RETURN(-ENOMEM);
894
895                 QCTL_COPY(oqctl, qctl);
896                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
897                 if (rc == 0) {
898                         QCTL_COPY(qctl, oqctl);
899                         qctl->qc_valid = QC_MDTIDX;
900                         qctl->obd_uuid = tgt->ltd_uuid;
901                 }
902                 OBD_FREE_PTR(oqctl);
903                 break;
904         }
905         case LL_IOC_GET_CONNECT_FLAGS: {
906                 tgt = lmv_tgt(lmv, 0);
907                 rc = -ENODATA;
908                 if (tgt && tgt->ltd_exp)
909                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
910                 break;
911         }
912         case LL_IOC_FID2MDTIDX: {
913                 struct lu_fid *fid = karg;
914                 int             mdt_index;
915
916                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
917                 if (rc != 0)
918                         RETURN(rc);
919
920                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
921                  * point to user space memory for FID2MDTIDX. */
922                 *(__u32 *)uarg = mdt_index;
923                 break;
924         }
925         case OBD_IOC_FID2PATH: {
926                 rc = lmv_fid2path(exp, len, karg, uarg);
927                 break;
928         }
929         case LL_IOC_HSM_STATE_GET:
930         case LL_IOC_HSM_STATE_SET:
931         case LL_IOC_HSM_ACTION: {
932                 struct md_op_data *op_data = karg;
933
934                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
935                 if (IS_ERR(tgt))
936                         RETURN(PTR_ERR(tgt));
937
938                 if (tgt->ltd_exp == NULL)
939                         RETURN(-EINVAL);
940
941                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
942                 break;
943         }
944         case LL_IOC_HSM_PROGRESS: {
945                 const struct hsm_progress_kernel *hpk = karg;
946
947                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
948                 if (IS_ERR(tgt))
949                         RETURN(PTR_ERR(tgt));
950                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
951                 break;
952         }
953         case LL_IOC_HSM_REQUEST: {
954                 struct hsm_user_request *hur = karg;
955                 unsigned int reqcount = hur->hur_request.hr_itemcount;
956
957                 if (reqcount == 0)
958                         RETURN(0);
959
960                 /* if the request is about a single fid
961                  * or if there is a single MDS, no need to split
962                  * the request. */
963                 if (reqcount == 1 || count == 1) {
964                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
965                         if (IS_ERR(tgt))
966                                 RETURN(PTR_ERR(tgt));
967                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
968                 } else {
969                         /* split fid list to their respective MDS */
970                         lmv_foreach_connected_tgt(lmv, tgt) {
971                                 int nr, rc1;
972                                 size_t reqlen;
973                                 struct hsm_user_request *req;
974
975                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
976                                 if (nr < 0)
977                                         RETURN(nr);
978                                 if (nr == 0) /* nothing for this MDS */
979                                         continue;
980
981                                 /* build a request with fids for this MDS */
982                                 reqlen = offsetof(typeof(*hur),
983                                                   hur_user_item[nr])
984                                                 + hur->hur_request.hr_data_len;
985                                 OBD_ALLOC_LARGE(req, reqlen);
986                                 if (req == NULL)
987                                         RETURN(-ENOMEM);
988                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
989                                 if (rc1 < 0)
990                                         GOTO(hsm_req_err, rc1);
991                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
992                                                     req, uarg);
993 hsm_req_err:
994                                 if (rc1 != 0 && rc == 0)
995                                         rc = rc1;
996                                 OBD_FREE_LARGE(req, reqlen);
997                         }
998                 }
999                 break;
1000         }
1001         case LL_IOC_LOV_SWAP_LAYOUTS: {
1002                 struct md_op_data *op_data = karg;
1003                 struct lmv_tgt_desc *tgt1, *tgt2;
1004
1005                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1006                 if (IS_ERR(tgt1))
1007                         RETURN(PTR_ERR(tgt1));
1008
1009                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1010                 if (IS_ERR(tgt2))
1011                         RETURN(PTR_ERR(tgt2));
1012
1013                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1014                         RETURN(-EINVAL);
1015
1016                 /* only files on same MDT can have their layouts swapped */
1017                 if (tgt1->ltd_index != tgt2->ltd_index)
1018                         RETURN(-EPERM);
1019
1020                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1021                 break;
1022         }
1023         case LL_IOC_HSM_CT_START: {
1024                 struct lustre_kernelcomm *lk = karg;
1025                 if (lk->lk_flags & LK_FLG_STOP)
1026                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1027                 else
1028                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1029                 break;
1030         }
1031         default:
1032                 lmv_foreach_connected_tgt(lmv, tgt) {
1033                         struct obd_device *mdc_obd;
1034                         int err;
1035
1036                         /* ll_umount_begin() sets force flag but for lmv, not
1037                          * mdc. Let's pass it through */
1038                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1039                         mdc_obd->obd_force = obd->obd_force;
1040                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1041                         if (err) {
1042                                 if (tgt->ltd_active) {
1043                                         CERROR("error: iocontrol MDC %s on MDT"
1044                                                " idx %d cmd %x: err = %d\n",
1045                                                tgt->ltd_uuid.uuid,
1046                                                tgt->ltd_index, cmd, err);
1047                                         if (!rc)
1048                                                 rc = err;
1049                                 }
1050                         } else
1051                                 set = 1;
1052                 }
1053                 if (!set && !rc)
1054                         rc = -EIO;
1055         }
1056         RETURN(rc);
1057 }
1058
1059 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1060                   struct lu_fid *fid, struct md_op_data *op_data)
1061 {
1062         struct obd_device *obd = class_exp2obd(exp);
1063         struct lmv_obd *lmv = &obd->u.lmv;
1064         struct lmv_tgt_desc *tgt;
1065         int rc;
1066
1067         ENTRY;
1068
1069         LASSERT(op_data);
1070         LASSERT(fid);
1071
1072         tgt = lmv_tgt(lmv, op_data->op_mds);
1073         if (!tgt)
1074                 RETURN(-ENODEV);
1075
1076         if (!tgt->ltd_active || !tgt->ltd_exp)
1077                 RETURN(-ENODEV);
1078
1079         /*
1080          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1081          * on server that seq in new allocated fid is not yet known.
1082          */
1083         mutex_lock(&tgt->ltd_fid_mutex);
1084         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1085         mutex_unlock(&tgt->ltd_fid_mutex);
1086         if (rc > 0) {
1087                 LASSERT(fid_is_sane(fid));
1088                 rc = 0;
1089         }
1090
1091         RETURN(rc);
1092 }
1093
1094 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1095 {
1096         struct lmv_obd *lmv = &obd->u.lmv;
1097         struct lmv_desc *desc;
1098         struct lnet_process_id lnet_id;
1099         int i = 0;
1100         int rc;
1101
1102         ENTRY;
1103
1104         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1105                 CERROR("LMV setup requires a descriptor\n");
1106                 RETURN(-EINVAL);
1107         }
1108
1109         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1110         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1111                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1112                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1113                 RETURN(-EINVAL);
1114         }
1115
1116         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1117                      desc->ld_uuid.uuid);
1118         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1119         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1120         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1121                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1122         lmv->max_def_easize = 0;
1123         lmv->max_easize = 0;
1124
1125         spin_lock_init(&lmv->lmv_lock);
1126
1127         /*
1128          * initialize rr_index to lower 32bit of netid, so that client
1129          * can distribute subdirs evenly from the beginning.
1130          */
1131         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1132                 if (lnet_id.nid != LNET_NID_LO_0) {
1133                         lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1134                         break;
1135                 }
1136         }
1137
1138         rc = lmv_tunables_init(obd);
1139         if (rc)
1140                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1141                       obd->obd_name, rc);
1142
1143         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1144                              LUSTRE_CLI_FLD_HASH_DHT);
1145         if (rc)
1146                 CERROR("Can't init FLD, err %d\n", rc);
1147
1148         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1149         if (rc)
1150                 CWARN("%s: error initialize target table: rc = %d\n",
1151                       obd->obd_name, rc);
1152
1153         RETURN(rc);
1154 }
1155
1156 static int lmv_cleanup(struct obd_device *obd)
1157 {
1158         struct lmv_obd *lmv = &obd->u.lmv;
1159         struct lu_tgt_desc *tgt;
1160         struct lu_tgt_desc *tmp;
1161
1162         ENTRY;
1163
1164         fld_client_fini(&lmv->lmv_fld);
1165         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1166                 lmv_del_target(lmv, tgt);
1167         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1168
1169         RETURN(0);
1170 }
1171
1172 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1173 {
1174         struct lustre_cfg       *lcfg = buf;
1175         struct obd_uuid         obd_uuid;
1176         int                     gen;
1177         __u32                   index;
1178         int                     rc;
1179         ENTRY;
1180
1181         switch (lcfg->lcfg_command) {
1182         case LCFG_ADD_MDC:
1183                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1184                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1185                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1186                         GOTO(out, rc = -EINVAL);
1187
1188                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1189
1190                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1191                         GOTO(out, rc = -EINVAL);
1192                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1193                         GOTO(out, rc = -EINVAL);
1194                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1195                 GOTO(out, rc);
1196         default:
1197                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1198                 GOTO(out, rc = -EINVAL);
1199         }
1200 out:
1201         RETURN(rc);
1202 }
1203
1204 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1205 {
1206         int i;
1207
1208         if (flags & OBD_STATFS_FOR_MDT0)
1209                 return 0;
1210
1211         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1212                 return lmv->lmv_statfs_start;
1213
1214         /* choose initial MDT for this client */
1215         for (i = 0;; i++) {
1216                 struct lnet_process_id lnet_id;
1217                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1218                         break;
1219
1220                 if (lnet_id.nid != LNET_NID_LO_0) {
1221                         /* We dont need a full 64-bit modulus, just enough
1222                          * to distribute the requests across MDTs evenly.
1223                          */
1224                         lmv->lmv_statfs_start = (u32)lnet_id.nid %
1225                                                 lmv->lmv_mdt_count;
1226                         break;
1227                 }
1228         }
1229
1230         return lmv->lmv_statfs_start;
1231 }
1232
1233 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1234                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1235 {
1236         struct obd_device *obd = class_exp2obd(exp);
1237         struct lmv_obd *lmv = &obd->u.lmv;
1238         struct obd_statfs *temp;
1239         struct lu_tgt_desc *tgt;
1240         __u32 i;
1241         __u32 idx;
1242         int rc = 0;
1243
1244         ENTRY;
1245
1246         OBD_ALLOC(temp, sizeof(*temp));
1247         if (temp == NULL)
1248                 RETURN(-ENOMEM);
1249
1250         /* distribute statfs among MDTs */
1251         idx = lmv_select_statfs_mdt(lmv, flags);
1252
1253         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1254                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1255                 tgt = lmv_tgt(lmv, idx);
1256                 if (!tgt || !tgt->ltd_exp)
1257                         continue;
1258
1259                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1260                                 flags | OBD_STATFS_NESTED);
1261                 if (rc) {
1262                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1263                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1264                         GOTO(out_free_temp, rc);
1265                 }
1266
1267                 if (temp->os_state & OS_STATFS_SUM ||
1268                     flags == OBD_STATFS_FOR_MDT0) {
1269                         /* reset to the last aggregated values
1270                          * and don't sum with non-aggrated data */
1271                         /* If the statfs is from mount, it needs to retrieve
1272                          * necessary information from MDT0. i.e. mount does
1273                          * not need the merged osfs from all of MDT. Also
1274                          * clients can be mounted as long as MDT0 is in
1275                          * service */
1276                         *osfs = *temp;
1277                         break;
1278                 }
1279
1280                 if (i == 0) {
1281                         *osfs = *temp;
1282                 } else {
1283                         osfs->os_bavail += temp->os_bavail;
1284                         osfs->os_blocks += temp->os_blocks;
1285                         osfs->os_ffree += temp->os_ffree;
1286                         osfs->os_files += temp->os_files;
1287                         osfs->os_granted += temp->os_granted;
1288                 }
1289         }
1290
1291         EXIT;
1292 out_free_temp:
1293         OBD_FREE(temp, sizeof(*temp));
1294         return rc;
1295 }
1296
1297 static int lmv_statfs_update(void *cookie, int rc)
1298 {
1299         struct obd_info *oinfo = cookie;
1300         struct obd_device *obd = oinfo->oi_obd;
1301         struct lmv_obd *lmv = &obd->u.lmv;
1302         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1303         struct obd_statfs *osfs = oinfo->oi_osfs;
1304
1305         /*
1306          * NB: don't deactivate TGT upon error, because we may not trigger async
1307          * statfs any longer, then there is no chance to activate TGT.
1308          */
1309         if (!rc) {
1310                 spin_lock(&lmv->lmv_lock);
1311                 tgt->ltd_statfs = *osfs;
1312                 tgt->ltd_statfs_age = ktime_get_seconds();
1313                 spin_unlock(&lmv->lmv_lock);
1314                 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1315         }
1316
1317         return rc;
1318 }
1319
1320 /* update tgt statfs async if it's ld_qos_maxage old */
1321 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1322 {
1323         struct obd_info oinfo = {
1324                 .oi_obd = obd,
1325                 .oi_tgt = tgt,
1326                 .oi_cb_up = lmv_statfs_update,
1327         };
1328         int rc;
1329
1330         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1331             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1332                 return 0;
1333
1334         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1335
1336         return rc;
1337 }
1338
1339 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1340                         struct lu_fid *fid)
1341 {
1342         struct obd_device *obd = exp->exp_obd;
1343         struct lmv_obd *lmv = &obd->u.lmv;
1344         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1345         int rc;
1346
1347         ENTRY;
1348
1349         if (!tgt)
1350                 RETURN(-ENODEV);
1351
1352         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1353         RETURN(rc);
1354 }
1355
1356 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1357                         u64 obd_md_valid, const char *name, size_t buf_size,
1358                         struct ptlrpc_request **req)
1359 {
1360         struct obd_device *obd = exp->exp_obd;
1361         struct lmv_obd *lmv = &obd->u.lmv;
1362         struct lmv_tgt_desc *tgt;
1363         int rc;
1364
1365         ENTRY;
1366
1367         tgt = lmv_fid2tgt(lmv, fid);
1368         if (IS_ERR(tgt))
1369                 RETURN(PTR_ERR(tgt));
1370
1371         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1372
1373         RETURN(rc);
1374 }
1375
1376 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1377                         u64 obd_md_valid, const char *name,
1378                         const void *value, size_t value_size,
1379                         unsigned int xattr_flags, u32 suppgid,
1380                         struct ptlrpc_request **req)
1381 {
1382         struct obd_device *obd = exp->exp_obd;
1383         struct lmv_obd *lmv = &obd->u.lmv;
1384         struct lmv_tgt_desc *tgt;
1385         int rc;
1386
1387         ENTRY;
1388
1389         tgt = lmv_fid2tgt(lmv, fid);
1390         if (IS_ERR(tgt))
1391                 RETURN(PTR_ERR(tgt));
1392
1393         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1394                          value, value_size, xattr_flags, suppgid, req);
1395
1396         RETURN(rc);
1397 }
1398
1399 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1400                        struct ptlrpc_request **request)
1401 {
1402         struct obd_device *obd = exp->exp_obd;
1403         struct lmv_obd *lmv = &obd->u.lmv;
1404         struct lmv_tgt_desc *tgt;
1405         int rc;
1406
1407         ENTRY;
1408
1409         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1410         if (IS_ERR(tgt))
1411                 RETURN(PTR_ERR(tgt));
1412
1413         if (op_data->op_flags & MF_GET_MDT_IDX) {
1414                 op_data->op_mds = tgt->ltd_index;
1415                 RETURN(0);
1416         }
1417
1418         rc = md_getattr(tgt->ltd_exp, op_data, request);
1419
1420         RETURN(rc);
1421 }
1422
1423 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1424 {
1425         struct obd_device *obd = exp->exp_obd;
1426         struct lmv_obd *lmv = &obd->u.lmv;
1427         struct lu_tgt_desc *tgt;
1428
1429         ENTRY;
1430
1431         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1432
1433         /*
1434          * With DNE every object can have two locks in different namespaces:
1435          * lookup lock in space of MDT storing direntry and update/open lock in
1436          * space of MDT storing inode.
1437          */
1438         lmv_foreach_connected_tgt(lmv, tgt)
1439                 md_null_inode(tgt->ltd_exp, fid);
1440
1441         RETURN(0);
1442 }
1443
1444 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1445                      struct md_open_data *mod, struct ptlrpc_request **request)
1446 {
1447         struct obd_device *obd = exp->exp_obd;
1448         struct lmv_obd *lmv = &obd->u.lmv;
1449         struct lmv_tgt_desc *tgt;
1450         int rc;
1451
1452         ENTRY;
1453
1454         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1455         if (IS_ERR(tgt))
1456                 RETURN(PTR_ERR(tgt));
1457
1458         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1459         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1460         RETURN(rc);
1461 }
1462
1463 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
1464 {
1465         struct lu_tgt_desc *tgt;
1466         __u64 total_weight = 0;
1467         __u64 cur_weight = 0;
1468         __u64 rand;
1469         int rc;
1470
1471         ENTRY;
1472
1473         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1474                 RETURN(ERR_PTR(-EAGAIN));
1475
1476         down_write(&lmv->lmv_qos.lq_rw_sem);
1477
1478         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1479                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1480
1481         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1482         if (rc)
1483                 GOTO(unlock, tgt = ERR_PTR(rc));
1484
1485         lmv_foreach_tgt(lmv, tgt) {
1486                 tgt->ltd_qos.ltq_usable = 0;
1487                 if (!tgt->ltd_exp || !tgt->ltd_active)
1488                         continue;
1489
1490                 tgt->ltd_qos.ltq_usable = 1;
1491                 lu_tgt_qos_weight_calc(tgt);
1492                 total_weight += tgt->ltd_qos.ltq_weight;
1493         }
1494
1495         rand = lu_prandom_u64_max(total_weight);
1496
1497         lmv_foreach_connected_tgt(lmv, tgt) {
1498                 if (!tgt->ltd_qos.ltq_usable)
1499                         continue;
1500
1501                 cur_weight += tgt->ltd_qos.ltq_weight;
1502                 if (cur_weight < rand)
1503                         continue;
1504
1505                 *mdt = tgt->ltd_index;
1506                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1507                 GOTO(unlock, rc = 0);
1508         }
1509
1510         /* no proper target found */
1511         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1512 unlock:
1513         up_write(&lmv->lmv_qos.lq_rw_sem);
1514
1515         return tgt;
1516 }
1517
1518 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
1519 {
1520         struct lu_tgt_desc *tgt;
1521         int i;
1522         int index;
1523
1524         ENTRY;
1525
1526         spin_lock(&lmv->lmv_lock);
1527         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1528                 index = (i + lmv->lmv_qos_rr_index) %
1529                         lmv->lmv_mdt_descs.ltd_tgts_size;
1530                 tgt = lmv_tgt(lmv, index);
1531                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1532                         continue;
1533
1534                 *mdt = tgt->ltd_index;
1535                 lmv->lmv_qos_rr_index = (*mdt + 1) %
1536                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1537                 spin_unlock(&lmv->lmv_lock);
1538
1539                 RETURN(tgt);
1540         }
1541         spin_unlock(&lmv->lmv_lock);
1542
1543         RETURN(ERR_PTR(-ENODEV));
1544 }
1545
1546 /* locate MDT by file name, for striped directory, the file name hash decides
1547  * which stripe its dirent is stored.
1548  */
1549 static struct lmv_tgt_desc *
1550 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1551                        const char *name, int namelen, struct lu_fid *fid,
1552                        __u32 *mds, bool new_layout)
1553 {
1554         struct lmv_tgt_desc *tgt;
1555         const struct lmv_oinfo *oinfo;
1556
1557         if (!lmv_dir_striped(lsm) || !namelen) {
1558                 tgt = lmv_fid2tgt(lmv, fid);
1559                 if (IS_ERR(tgt))
1560                         return tgt;
1561
1562                 *mds = tgt->ltd_index;
1563                 return tgt;
1564         }
1565
1566         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1567                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1568                         return ERR_PTR(-EBADF);
1569                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1570         } else {
1571                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1572                 if (IS_ERR(oinfo))
1573                         return ERR_CAST(oinfo);
1574         }
1575
1576         /* check stripe FID is sane */
1577         if (!fid_is_sane(&oinfo->lmo_fid))
1578                 return ERR_PTR(-ENODEV);
1579
1580         *fid = oinfo->lmo_fid;
1581         *mds = oinfo->lmo_mds;
1582         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1583
1584         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1585
1586         return tgt ? tgt : ERR_PTR(-ENODEV);
1587 }
1588
1589 /**
1590  * Locate MDT of op_data->op_fid1
1591  *
1592  * For striped directory, it will locate the stripe by name hash, if hash_type
1593  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1594  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1595  * indicates whether old or new layout is used to locate.
1596  *
1597  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1598  *
1599  * \param[in] lmv               LMV device
1600  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1601  *                              op_mds and op_fid1 will be updated if op_mea1
1602  *                              indicates fid1 represents a striped directory.
1603  *
1604  * retval               pointer to the lmv_tgt_desc if succeed.
1605  *                      ERR_PTR(errno) if failed.
1606  */
1607 struct lmv_tgt_desc *
1608 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1609 {
1610         struct lmv_stripe_md *lsm = op_data->op_mea1;
1611         struct lmv_oinfo *oinfo;
1612         struct lmv_tgt_desc *tgt;
1613
1614         if (lmv_dir_foreign(lsm))
1615                 return ERR_PTR(-ENODATA);
1616
1617         /* During creating VOLATILE file, it should honor the mdt
1618          * index if the file under striped dir is being restored, see
1619          * ct_restore(). */
1620         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1621             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1622                 tgt = lmv_tgt(lmv, op_data->op_mds);
1623                 if (!tgt)
1624                         return ERR_PTR(-ENODEV);
1625
1626                 if (lmv_dir_striped(lsm)) {
1627                         int i;
1628
1629                         /* refill the right parent fid */
1630                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1631                                 oinfo = &lsm->lsm_md_oinfo[i];
1632                                 if (oinfo->lmo_mds == op_data->op_mds) {
1633                                         op_data->op_fid1 = oinfo->lmo_fid;
1634                                         break;
1635                                 }
1636                         }
1637
1638                         if (i == lsm->lsm_md_stripe_count)
1639                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1640                 }
1641         } else if (lmv_dir_bad_hash(lsm)) {
1642                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1643                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1644
1645                 op_data->op_fid1 = oinfo->lmo_fid;
1646                 op_data->op_mds = oinfo->lmo_mds;
1647                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1648                 if (!tgt)
1649                         return ERR_PTR(-ENODEV);
1650         } else {
1651                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1652                                 op_data->op_name, op_data->op_namelen,
1653                                 &op_data->op_fid1, &op_data->op_mds,
1654                                 op_data->op_new_layout);
1655         }
1656
1657         return tgt;
1658 }
1659
1660 /* Locate MDT of op_data->op_fid2 for link/rename */
1661 static struct lmv_tgt_desc *
1662 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1663 {
1664         struct lmv_tgt_desc *tgt;
1665         int rc;
1666
1667         LASSERT(op_data->op_name);
1668         if (lmv_dir_layout_changing(op_data->op_mea2)) {
1669                 struct lu_fid fid1 = op_data->op_fid1;
1670                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1671                 struct ptlrpc_request *request = NULL;
1672
1673                 /*
1674                  * avoid creating new file under old layout of migrating
1675                  * directory, check it here.
1676                  */
1677                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1678                                 op_data->op_name, op_data->op_namelen,
1679                                 &op_data->op_fid2, &op_data->op_mds, false);
1680                 if (IS_ERR(tgt))
1681                         RETURN(tgt);
1682
1683                 op_data->op_fid1 = op_data->op_fid2;
1684                 op_data->op_mea1 = op_data->op_mea2;
1685                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1686                 op_data->op_fid1 = fid1;
1687                 op_data->op_mea1 = lsm1;
1688                 if (!rc) {
1689                         ptlrpc_req_finished(request);
1690                         RETURN(ERR_PTR(-EEXIST));
1691                 }
1692
1693                 if (rc != -ENOENT)
1694                         RETURN(ERR_PTR(rc));
1695         }
1696
1697         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1698                                 op_data->op_namelen, &op_data->op_fid2,
1699                                 &op_data->op_mds, true);
1700 }
1701
1702 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1703 {
1704         struct lu_tgt_desc *tgt;
1705         struct ptlrpc_request *request;
1706         int rc;
1707
1708         LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1709         LASSERT(!op_data->op_new_layout);
1710
1711         tgt = lmv_locate_tgt(lmv, op_data);
1712         if (IS_ERR(tgt))
1713                 return PTR_ERR(tgt);
1714
1715         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1716         if (!rc) {
1717                 ptlrpc_req_finished(request);
1718                 return -EEXIST;
1719         }
1720
1721         return rc;
1722 }
1723
1724 /* mkdir by QoS in two cases:
1725  * 1. 'lfs mkdir -i -1'
1726  * 2. parent default LMV master_mdt_index is -1
1727  *
1728  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1729  * directories under striped directory.
1730  */
1731 static inline bool lmv_op_qos_mkdir(const struct md_op_data *op_data)
1732 {
1733         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1734         const struct lmv_user_md *lum = op_data->op_data;
1735
1736         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1737                 return false;
1738
1739         if (lmv_dir_striped(op_data->op_mea1))
1740                 return false;
1741
1742         if (op_data->op_cli_flags & CLI_SET_MEA && lum &&
1743             (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1744              le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1745             le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT)
1746                 return true;
1747
1748         if (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT)
1749                 return true;
1750
1751         return false;
1752 }
1753
1754 /* 'lfs mkdir -i <specific_MDT>' */
1755 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1756 {
1757         const struct lmv_user_md *lum = op_data->op_data;
1758
1759         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1760                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1761                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1762                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1763                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1764 }
1765
1766 /* parent default LMV master_mdt_index is not -1. */
1767 static inline bool
1768 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1769 {
1770         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1771                op_data->op_default_mea1 &&
1772                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1773                         LMV_OFFSET_DEFAULT;
1774 }
1775 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1776                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1777                 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1778                 struct ptlrpc_request **request)
1779 {
1780         struct obd_device *obd = exp->exp_obd;
1781         struct lmv_obd *lmv = &obd->u.lmv;
1782         struct lmv_tgt_desc *tgt;
1783         struct mdt_body *repbody;
1784         int rc;
1785
1786         ENTRY;
1787
1788         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1789                 RETURN(-EIO);
1790
1791         if (lmv_dir_bad_hash(op_data->op_mea1))
1792                 RETURN(-EBADF);
1793
1794         if (lmv_dir_layout_changing(op_data->op_mea1)) {
1795                 /*
1796                  * if parent is migrating, create() needs to lookup existing
1797                  * name in both old and new layout, check old layout on client.
1798                  */
1799                 rc = lmv_old_layout_lookup(lmv, op_data);
1800                 if (rc != -ENOENT)
1801                         RETURN(rc);
1802
1803                 op_data->op_new_layout = true;
1804         }
1805
1806         tgt = lmv_locate_tgt(lmv, op_data);
1807         if (IS_ERR(tgt))
1808                 RETURN(PTR_ERR(tgt));
1809
1810         if (lmv_op_user_specific_mkdir(op_data)) {
1811                 struct lmv_user_md *lum = op_data->op_data;
1812
1813                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1814                 tgt = lmv_tgt(lmv, op_data->op_mds);
1815                 if (!tgt)
1816                         RETURN(-ENODEV);
1817         } else if (lmv_op_default_specific_mkdir(op_data)) {
1818                 op_data->op_mds =
1819                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1820                 tgt = lmv_tgt(lmv, op_data->op_mds);
1821                 if (!tgt)
1822                         RETURN(-ENODEV);
1823         } else if (lmv_op_qos_mkdir(op_data)) {
1824                 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds);
1825                 if (tgt == ERR_PTR(-EAGAIN))
1826                         tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1827                 if (IS_ERR(tgt))
1828                         RETURN(PTR_ERR(tgt));
1829                 /*
1830                  * only update statfs after QoS mkdir, this means the cached
1831                  * statfs may be stale, and current mkdir may not follow QoS
1832                  * accurately, but it's not serious, and avoids periodic statfs
1833                  * when client doesn't mkdir by QoS.
1834                  */
1835                 lmv_statfs_check_update(obd, tgt);
1836         }
1837
1838 retry:
1839         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1840         if (rc)
1841                 RETURN(rc);
1842
1843         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1844                 (int)op_data->op_namelen, op_data->op_name,
1845                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1846                 op_data->op_mds);
1847
1848         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1849         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1850                        cap_effective, rdev, request);
1851         if (rc == 0) {
1852                 if (*request == NULL)
1853                         RETURN(rc);
1854                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1855         }
1856
1857         /* dir restripe needs to send to MDT where dir is located */
1858         if (rc != -EREMOTE ||
1859             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
1860                 RETURN(rc);
1861
1862         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
1863         if (repbody == NULL)
1864                 RETURN(-EPROTO);
1865
1866         /* Not cross-ref case, just get out of here. */
1867         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
1868                 RETURN(rc);
1869
1870         op_data->op_fid2 = repbody->mbo_fid1;
1871         ptlrpc_req_finished(*request);
1872         *request = NULL;
1873
1874         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
1875         if (IS_ERR(tgt))
1876                 RETURN(PTR_ERR(tgt));
1877
1878         op_data->op_mds = tgt->ltd_index;
1879         goto retry;
1880 }
1881
1882 static int
1883 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1884             const union ldlm_policy_data *policy, struct md_op_data *op_data,
1885             struct lustre_handle *lockh, __u64 extra_lock_flags)
1886 {
1887         struct obd_device *obd = exp->exp_obd;
1888         struct lmv_obd *lmv = &obd->u.lmv;
1889         struct lmv_tgt_desc *tgt;
1890         int rc;
1891
1892         ENTRY;
1893
1894         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1895
1896         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1897         if (IS_ERR(tgt))
1898                 RETURN(PTR_ERR(tgt));
1899
1900         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1901                PFID(&op_data->op_fid1), tgt->ltd_index);
1902
1903         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1904                         extra_lock_flags);
1905
1906         RETURN(rc);
1907 }
1908
1909 int
1910 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1911                  struct ptlrpc_request **preq)
1912 {
1913         struct obd_device *obd = exp->exp_obd;
1914         struct lmv_obd *lmv = &obd->u.lmv;
1915         struct lmv_tgt_desc *tgt;
1916         struct mdt_body *body;
1917         int rc;
1918
1919         ENTRY;
1920
1921 retry:
1922         tgt = lmv_locate_tgt(lmv, op_data);
1923         if (IS_ERR(tgt))
1924                 RETURN(PTR_ERR(tgt));
1925
1926         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1927                 (int)op_data->op_namelen, op_data->op_name,
1928                 PFID(&op_data->op_fid1), tgt->ltd_index);
1929
1930         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1931         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1932                 ptlrpc_req_finished(*preq);
1933                 *preq = NULL;
1934                 goto retry;
1935         }
1936
1937         if (rc)
1938                 RETURN(rc);
1939
1940         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1941         LASSERT(body != NULL);
1942
1943         if (body->mbo_valid & OBD_MD_MDS) {
1944                 op_data->op_fid1 = body->mbo_fid1;
1945                 op_data->op_valid |= OBD_MD_FLCROSSREF;
1946                 op_data->op_namelen = 0;
1947                 op_data->op_name = NULL;
1948
1949                 ptlrpc_req_finished(*preq);
1950                 *preq = NULL;
1951
1952                 goto retry;
1953         }
1954
1955         RETURN(rc);
1956 }
1957
1958 #define md_op_data_fid(op_data, fl)                     \
1959         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1960          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1961          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1962          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1963          NULL)
1964
1965 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1966                             struct md_op_data *op_data, __u32 op_tgt,
1967                             enum ldlm_mode mode, int bits, int flag)
1968 {
1969         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1970         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1971         union ldlm_policy_data policy = { { 0 } };
1972         int rc = 0;
1973         ENTRY;
1974
1975         if (!fid_is_sane(fid))
1976                 RETURN(0);
1977
1978         if (tgt == NULL) {
1979                 tgt = lmv_fid2tgt(lmv, fid);
1980                 if (IS_ERR(tgt))
1981                         RETURN(PTR_ERR(tgt));
1982         }
1983
1984         if (tgt->ltd_index != op_tgt) {
1985                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1986                 policy.l_inodebits.bits = bits;
1987                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1988                                       mode, LCF_ASYNC, NULL);
1989         } else {
1990                 CDEBUG(D_INODE,
1991                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
1992                        op_tgt, PFID(fid));
1993                 op_data->op_flags |= flag;
1994                 rc = 0;
1995         }
1996
1997         RETURN(rc);
1998 }
1999
2000 /*
2001  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2002  * op_data->op_fid2
2003  */
2004 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2005                     struct ptlrpc_request **request)
2006 {
2007         struct obd_device       *obd = exp->exp_obd;
2008         struct lmv_obd          *lmv = &obd->u.lmv;
2009         struct lmv_tgt_desc     *tgt;
2010         int                      rc;
2011         ENTRY;
2012
2013         LASSERT(op_data->op_namelen != 0);
2014
2015         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2016                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2017                op_data->op_name, PFID(&op_data->op_fid1));
2018
2019         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2020         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2021         op_data->op_cap = cfs_curproc_cap_pack();
2022
2023         tgt = lmv_locate_tgt2(lmv, op_data);
2024         if (IS_ERR(tgt))
2025                 RETURN(PTR_ERR(tgt));
2026
2027         /*
2028          * Cancel UPDATE lock on child (fid1).
2029          */
2030         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2031         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2032                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2033         if (rc != 0)
2034                 RETURN(rc);
2035
2036         rc = md_link(tgt->ltd_exp, op_data, request);
2037
2038         RETURN(rc);
2039 }
2040
2041 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2042                         const char *name, size_t namelen,
2043                         struct ptlrpc_request **request)
2044 {
2045         struct obd_device *obd = exp->exp_obd;
2046         struct lmv_obd *lmv = &obd->u.lmv;
2047         struct lmv_stripe_md *lsm = op_data->op_mea1;
2048         struct lmv_tgt_desc *parent_tgt;
2049         struct lmv_tgt_desc *sp_tgt;
2050         struct lmv_tgt_desc *tp_tgt = NULL;
2051         struct lmv_tgt_desc *child_tgt;
2052         struct lmv_tgt_desc *tgt;
2053         struct lu_fid target_fid = { 0 };
2054         int rc;
2055
2056         ENTRY;
2057
2058         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2059
2060         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2061                PFID(&op_data->op_fid1), (int)namelen, name);
2062
2063         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2064         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2065         op_data->op_cap = cfs_curproc_cap_pack();
2066
2067         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2068         if (IS_ERR(parent_tgt))
2069                 RETURN(PTR_ERR(parent_tgt));
2070
2071         if (lmv_dir_striped(lsm)) {
2072                 const struct lmv_oinfo *oinfo;
2073
2074                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2075                 if (IS_ERR(oinfo))
2076                         RETURN(PTR_ERR(oinfo));
2077
2078                 /* save source stripe FID in fid4 temporarily for ELC */
2079                 op_data->op_fid4 = oinfo->lmo_fid;
2080                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2081                 if (!sp_tgt)
2082                         RETURN(-ENODEV);
2083
2084                 /*
2085                  * if parent is being migrated too, fill op_fid2 with target
2086                  * stripe fid, otherwise the target stripe is not created yet.
2087                  */
2088                 if (lmv_dir_layout_changing(lsm)) {
2089                         oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2090                                                         true);
2091                         if (IS_ERR(oinfo))
2092                                 RETURN(PTR_ERR(oinfo));
2093
2094                         op_data->op_fid2 = oinfo->lmo_fid;
2095                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2096                         if (!tp_tgt)
2097                                 RETURN(-ENODEV);
2098                 }
2099         } else {
2100                 sp_tgt = parent_tgt;
2101         }
2102
2103         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2104         if (IS_ERR(child_tgt))
2105                 RETURN(PTR_ERR(child_tgt));
2106
2107         /* for directory, migrate to MDT specified by lum_stripe_offset;
2108          * otherwise migrate to the target stripe of parent, but parent
2109          * directory may have finished migration (normally current file too),
2110          * allocate FID on MDT lum_stripe_offset, and server will check
2111          * whether file was migrated already.
2112          */
2113         if (S_ISDIR(op_data->op_mode) || !tp_tgt) {
2114                 struct lmv_user_md *lum = op_data->op_data;
2115
2116                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2117         } else  {
2118                 op_data->op_mds = tp_tgt->ltd_index;
2119         }
2120         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2121         if (rc)
2122                 RETURN(rc);
2123
2124         /*
2125          * for directory, send migrate request to the MDT where the object will
2126          * be migrated to, because we can't create a striped directory remotely.
2127          *
2128          * otherwise, send to the MDT where source is located because regular
2129          * file may open lease.
2130          *
2131          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2132          * backward compatibility.
2133          */
2134         if (S_ISDIR(op_data->op_mode) &&
2135             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2136                 tgt = lmv_fid2tgt(lmv, &target_fid);
2137                 if (IS_ERR(tgt))
2138                         RETURN(PTR_ERR(tgt));
2139         } else {
2140                 tgt = child_tgt;
2141         }
2142
2143         /* cancel UPDATE lock of parent master object */
2144         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2145                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2146         if (rc)
2147                 RETURN(rc);
2148
2149         /* cancel UPDATE lock of source parent */
2150         if (sp_tgt != parent_tgt) {
2151                 /*
2152                  * migrate RPC packs master object FID, because we can only pack
2153                  * two FIDs in reint RPC, but MDS needs to know both source
2154                  * parent and target parent, and it will obtain them from master
2155                  * FID and LMV, the other FID in RPC is kept for target.
2156                  *
2157                  * since this FID is not passed to MDC, cancel it anyway.
2158                  */
2159                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2160                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2161                 if (rc)
2162                         RETURN(rc);
2163
2164                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2165         }
2166         op_data->op_fid4 = target_fid;
2167
2168         /* cancel UPDATE locks of target parent */
2169         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2170                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2171         if (rc)
2172                 RETURN(rc);
2173
2174         /* cancel LOOKUP lock of source if source is remote object */
2175         if (child_tgt != sp_tgt) {
2176                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2177                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2178                                       MF_MDC_CANCEL_FID3);
2179                 if (rc)
2180                         RETURN(rc);
2181         }
2182
2183         /* cancel ELC locks of source */
2184         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2185                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2186         if (rc)
2187                 RETURN(rc);
2188
2189         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2190
2191         RETURN(rc);
2192 }
2193
2194 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2195                       const char *old, size_t oldlen,
2196                       const char *new, size_t newlen,
2197                       struct ptlrpc_request **request)
2198 {
2199         struct obd_device *obd = exp->exp_obd;
2200         struct lmv_obd *lmv = &obd->u.lmv;
2201         struct lmv_tgt_desc *sp_tgt;
2202         struct lmv_tgt_desc *tp_tgt = NULL;
2203         struct lmv_tgt_desc *src_tgt = NULL;
2204         struct lmv_tgt_desc *tgt;
2205         struct mdt_body *body;
2206         int rc;
2207
2208         ENTRY;
2209
2210         LASSERT(oldlen != 0);
2211
2212         if (op_data->op_cli_flags & CLI_MIGRATE) {
2213                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2214                 RETURN(rc);
2215         }
2216
2217         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2218         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2219         op_data->op_cap = cfs_curproc_cap_pack();
2220
2221         op_data->op_name = new;
2222         op_data->op_namelen = newlen;
2223
2224         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2225         if (IS_ERR(tp_tgt))
2226                 RETURN(PTR_ERR(tp_tgt));
2227
2228         /* Since the target child might be destroyed, and it might become
2229          * orphan, and we can only check orphan on the local MDT right now, so
2230          * we send rename request to the MDT where target child is located. If
2231          * target child does not exist, then it will send the request to the
2232          * target parent */
2233         if (fid_is_sane(&op_data->op_fid4)) {
2234                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2235                 if (IS_ERR(tgt))
2236                         RETURN(PTR_ERR(tgt));
2237         } else {
2238                 tgt = tp_tgt;
2239         }
2240
2241         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2242
2243         /* cancel UPDATE locks of target parent */
2244         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2245                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2246         if (rc != 0)
2247                 RETURN(rc);
2248
2249         if (fid_is_sane(&op_data->op_fid4)) {
2250                 /* cancel LOOKUP lock of target on target parent */
2251                 if (tgt != tp_tgt) {
2252                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2253                                               tgt->ltd_index, LCK_EX,
2254                                               MDS_INODELOCK_LOOKUP,
2255                                               MF_MDC_CANCEL_FID4);
2256                         if (rc != 0)
2257                                 RETURN(rc);
2258                 }
2259         }
2260
2261         if (fid_is_sane(&op_data->op_fid3)) {
2262                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2263                 if (IS_ERR(src_tgt))
2264                         RETURN(PTR_ERR(src_tgt));
2265
2266                 /* cancel ELC locks of source */
2267                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2268                                       LCK_EX, MDS_INODELOCK_ELC,
2269                                       MF_MDC_CANCEL_FID3);
2270                 if (rc != 0)
2271                         RETURN(rc);
2272         }
2273
2274         op_data->op_name = old;
2275         op_data->op_namelen = oldlen;
2276 retry:
2277         sp_tgt = lmv_locate_tgt(lmv, op_data);
2278         if (IS_ERR(sp_tgt))
2279                 RETURN(PTR_ERR(sp_tgt));
2280
2281         /* cancel UPDATE locks of source parent */
2282         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2283                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2284         if (rc != 0)
2285                 RETURN(rc);
2286
2287         if (fid_is_sane(&op_data->op_fid3)) {
2288                 /* cancel LOOKUP lock of source on source parent */
2289                 if (src_tgt != sp_tgt) {
2290                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2291                                               tgt->ltd_index, LCK_EX,
2292                                               MDS_INODELOCK_LOOKUP,
2293                                               MF_MDC_CANCEL_FID3);
2294                         if (rc != 0)
2295                                 RETURN(rc);
2296                 }
2297         }
2298
2299 rename:
2300         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2301                 PFID(&op_data->op_fid1), (int)oldlen, old,
2302                 PFID(&op_data->op_fid2), (int)newlen, new);
2303
2304         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2305                         request);
2306         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2307                 ptlrpc_req_finished(*request);
2308                 *request = NULL;
2309                 goto retry;
2310         }
2311
2312         if (rc && rc != -EXDEV)
2313                 RETURN(rc);
2314
2315         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2316         if (body == NULL)
2317                 RETURN(-EPROTO);
2318
2319         /* Not cross-ref case, just get out of here. */
2320         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2321                 RETURN(rc);
2322
2323         op_data->op_fid4 = body->mbo_fid1;
2324
2325         ptlrpc_req_finished(*request);
2326         *request = NULL;
2327
2328         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2329         if (IS_ERR(tgt))
2330                 RETURN(PTR_ERR(tgt));
2331
2332         if (fid_is_sane(&op_data->op_fid4)) {
2333                 /* cancel LOOKUP lock of target on target parent */
2334                 if (tgt != tp_tgt) {
2335                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2336                                               tgt->ltd_index, LCK_EX,
2337                                               MDS_INODELOCK_LOOKUP,
2338                                               MF_MDC_CANCEL_FID4);
2339                         if (rc != 0)
2340                                 RETURN(rc);
2341                 }
2342         }
2343
2344         goto rename;
2345 }
2346
2347 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2348                        void *ea, size_t ealen, struct ptlrpc_request **request)
2349 {
2350         struct obd_device *obd = exp->exp_obd;
2351         struct lmv_obd *lmv = &obd->u.lmv;
2352         struct lmv_tgt_desc *tgt;
2353         int rc = 0;
2354
2355         ENTRY;
2356
2357         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2358                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2359                op_data->op_xvalid);
2360
2361         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2362         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2363         if (IS_ERR(tgt))
2364                 RETURN(PTR_ERR(tgt));
2365
2366         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2367
2368         RETURN(rc);
2369 }
2370
2371 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2372                      struct ptlrpc_request **request)
2373 {
2374         struct obd_device *obd = exp->exp_obd;
2375         struct lmv_obd *lmv = &obd->u.lmv;
2376         struct lmv_tgt_desc *tgt;
2377         int rc;
2378
2379         ENTRY;
2380
2381         tgt = lmv_fid2tgt(lmv, fid);
2382         if (IS_ERR(tgt))
2383                 RETURN(PTR_ERR(tgt));
2384
2385         rc = md_fsync(tgt->ltd_exp, fid, request);
2386         RETURN(rc);
2387 }
2388
2389 struct stripe_dirent {
2390         struct page             *sd_page;
2391         struct lu_dirpage       *sd_dp;
2392         struct lu_dirent        *sd_ent;
2393         bool                     sd_eof;
2394 };
2395
2396 struct lmv_dir_ctxt {
2397         struct lmv_obd          *ldc_lmv;
2398         struct md_op_data       *ldc_op_data;
2399         struct md_callback      *ldc_cb_op;
2400         __u64                    ldc_hash;
2401         int                      ldc_count;
2402         struct stripe_dirent     ldc_stripes[0];
2403 };
2404
2405 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2406 {
2407         if (stripe->sd_page) {
2408                 kunmap(stripe->sd_page);
2409                 put_page(stripe->sd_page);
2410                 stripe->sd_page = NULL;
2411                 stripe->sd_ent = NULL;
2412         }
2413 }
2414
2415 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2416 {
2417         int i;
2418
2419         for (i = 0; i < ctxt->ldc_count; i++)
2420                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2421 }
2422
2423 /* if @ent is dummy, or . .., get next */
2424 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2425                                            struct lu_dirent *ent,
2426                                            int stripe_index)
2427 {
2428         for (; ent; ent = lu_dirent_next(ent)) {
2429                 /* Skip dummy entry */
2430                 if (le16_to_cpu(ent->lde_namelen) == 0)
2431                         continue;
2432
2433                 /* skip . and .. for other stripes */
2434                 if (stripe_index &&
2435                     (strncmp(ent->lde_name, ".",
2436                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2437                      strncmp(ent->lde_name, "..",
2438                              le16_to_cpu(ent->lde_namelen)) == 0))
2439                         continue;
2440
2441                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2442                         break;
2443         }
2444
2445         return ent;
2446 }
2447
2448 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2449                                             struct stripe_dirent *stripe,
2450                                             int stripe_index)
2451 {
2452         struct md_op_data *op_data = ctxt->ldc_op_data;
2453         struct lmv_oinfo *oinfo;
2454         struct lu_fid fid = op_data->op_fid1;
2455         struct inode *inode = op_data->op_data;
2456         struct lmv_tgt_desc *tgt;
2457         struct lu_dirent *ent = stripe->sd_ent;
2458         __u64 hash = ctxt->ldc_hash;
2459         int rc = 0;
2460
2461         ENTRY;
2462
2463         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2464         LASSERT(!ent);
2465
2466         do {
2467                 if (stripe->sd_page) {
2468                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2469
2470                         /* @hash should be the last dirent hash */
2471                         LASSERTF(hash <= end,
2472                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2473                                  ctxt, stripe, hash, end);
2474                         /* unload last page */
2475                         stripe_dirent_unload(stripe);
2476                         /* eof */
2477                         if (end == MDS_DIR_END_OFF) {
2478                                 stripe->sd_eof = true;
2479                                 break;
2480                         }
2481                         hash = end;
2482                 }
2483
2484                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2485                 if (!oinfo->lmo_root) {
2486                         rc = -ENOENT;
2487                         break;
2488                 }
2489
2490                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2491                 if (!tgt) {
2492                         rc = -ENODEV;
2493                         break;
2494                 }
2495
2496                 /* op_data is shared by stripes, reset after use */
2497                 op_data->op_fid1 = oinfo->lmo_fid;
2498                 op_data->op_fid2 = oinfo->lmo_fid;
2499                 op_data->op_data = oinfo->lmo_root;
2500
2501                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2502                                   &stripe->sd_page);
2503
2504                 op_data->op_fid1 = fid;
2505                 op_data->op_fid2 = fid;
2506                 op_data->op_data = inode;
2507
2508                 if (rc)
2509                         break;
2510
2511                 stripe->sd_dp = page_address(stripe->sd_page);
2512                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2513                                         stripe_index);
2514                 /* in case a page filled with ., .. and dummy, read next */
2515         } while (!ent);
2516
2517         stripe->sd_ent = ent;
2518         if (rc) {
2519                 LASSERT(!ent);
2520                 /* treat error as eof, so dir can be partially accessed */
2521                 stripe->sd_eof = true;
2522                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2523                               "directory is partially accessed!\n",
2524                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2525                               rc);
2526         }
2527
2528         RETURN(ent);
2529 }
2530
2531 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2532 {
2533         struct obd_device *obd = exp->exp_obd;
2534         struct lmv_obd *lmv = &obd->u.lmv;
2535         struct lmv_tgt_desc *tgt;
2536         int rc;
2537
2538         ENTRY;
2539
2540         rc = lmv_check_connect(obd);
2541         if (rc != 0)
2542                 RETURN(rc);
2543
2544         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2545         if (IS_ERR(tgt))
2546                 RETURN(PTR_ERR(tgt));
2547
2548         data->op_flags |= MF_MDC_CANCEL_FID1;
2549         rc = md_file_resync(tgt->ltd_exp, data);
2550         RETURN(rc);
2551 }
2552
2553 /**
2554  * Get dirent with the closest hash for striped directory
2555  *
2556  * This function will search the dir entry, whose hash value is the
2557  * closest(>=) to hash from all of sub-stripes, and it is only being called
2558  * for striped directory.
2559  *
2560  * \param[in] ctxt              dir read context
2561  *
2562  * \retval                      dirent get the entry successfully
2563  *                              NULL does not get the entry, normally it means
2564  *                              it reaches the end of the directory, while read
2565  *                              stripe dirent error is ignored to allow partial
2566  *                              access.
2567  */
2568 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2569 {
2570         struct stripe_dirent *stripe;
2571         struct lu_dirent *ent = NULL;
2572         int i;
2573         int min = -1;
2574
2575         /* TODO: optimize with k-way merge sort */
2576         for (i = 0; i < ctxt->ldc_count; i++) {
2577                 stripe = &ctxt->ldc_stripes[i];
2578                 if (stripe->sd_eof)
2579                         continue;
2580
2581                 if (!stripe->sd_ent) {
2582                         stripe_dirent_load(ctxt, stripe, i);
2583                         if (!stripe->sd_ent) {
2584                                 LASSERT(stripe->sd_eof);
2585                                 continue;
2586                         }
2587                 }
2588
2589                 if (min == -1 ||
2590                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2591                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2592                         min = i;
2593                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2594                             ctxt->ldc_hash)
2595                                 break;
2596                 }
2597         }
2598
2599         if (min != -1) {
2600                 stripe = &ctxt->ldc_stripes[min];
2601                 ent = stripe->sd_ent;
2602                 /* pop found dirent */
2603                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2604                                                    min);
2605         }
2606
2607         return ent;
2608 }
2609
2610 /**
2611  * Build dir entry page for striped directory
2612  *
2613  * This function gets one entry by @offset from a striped directory. It will
2614  * read entries from all of stripes, and choose one closest to the required
2615  * offset(&offset). A few notes
2616  * 1. skip . and .. for non-zero stripes, because there can only have one .
2617  * and .. in a directory.
2618  * 2. op_data will be shared by all of stripes, instead of allocating new
2619  * one, so need to restore before reusing.
2620  *
2621  * \param[in] exp       obd export refer to LMV
2622  * \param[in] op_data   hold those MD parameters of read_entry
2623  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2624  * \param[in] offset    starting hash offset
2625  * \param[out] ppage    the page holding the entry. Note: because the entry
2626  *                      will be accessed in upper layer, so we need hold the
2627  *                      page until the usages of entry is finished, see
2628  *                      ll_dir_entry_next.
2629  *
2630  * retval               =0 if get entry successfully
2631  *                      <0 cannot get entry
2632  */
2633 static int lmv_striped_read_page(struct obd_export *exp,
2634                                  struct md_op_data *op_data,
2635                                  struct md_callback *cb_op,
2636                                  __u64 offset, struct page **ppage)
2637 {
2638         struct page *page = NULL;
2639         struct lu_dirpage *dp;
2640         void *start;
2641         struct lu_dirent *ent;
2642         struct lu_dirent *last_ent;
2643         int stripe_count;
2644         struct lmv_dir_ctxt *ctxt;
2645         struct lu_dirent *next = NULL;
2646         __u16 ent_size;
2647         size_t left_bytes;
2648         int rc = 0;
2649         ENTRY;
2650
2651         /* Allocate a page and read entries from all of stripes and fill
2652          * the page by hash order */
2653         page = alloc_page(GFP_KERNEL);
2654         if (!page)
2655                 RETURN(-ENOMEM);
2656
2657         /* Initialize the entry page */
2658         dp = kmap(page);
2659         memset(dp, 0, sizeof(*dp));
2660         dp->ldp_hash_start = cpu_to_le64(offset);
2661
2662         start = dp + 1;
2663         left_bytes = PAGE_SIZE - sizeof(*dp);
2664         ent = start;
2665         last_ent = ent;
2666
2667         /* initalize dir read context */
2668         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2669         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2670         if (!ctxt)
2671                 GOTO(free_page, rc = -ENOMEM);
2672         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2673         ctxt->ldc_op_data = op_data;
2674         ctxt->ldc_cb_op = cb_op;
2675         ctxt->ldc_hash = offset;
2676         ctxt->ldc_count = stripe_count;
2677
2678         while (1) {
2679                 next = lmv_dirent_next(ctxt);
2680
2681                 /* end of directory */
2682                 if (!next) {
2683                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2684                         break;
2685                 }
2686                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2687
2688                 ent_size = le16_to_cpu(next->lde_reclen);
2689
2690                 /* the last entry lde_reclen is 0, but it might not be the last
2691                  * one of this temporay dir page */
2692                 if (!ent_size)
2693                         ent_size = lu_dirent_calc_size(
2694                                         le16_to_cpu(next->lde_namelen),
2695                                         le32_to_cpu(next->lde_attrs));
2696                 /* page full */
2697                 if (ent_size > left_bytes)
2698                         break;
2699
2700                 memcpy(ent, next, ent_size);
2701
2702                 /* Replace . with master FID and Replace .. with the parent FID
2703                  * of master object */
2704                 if (strncmp(ent->lde_name, ".",
2705                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2706                     le16_to_cpu(ent->lde_namelen) == 1)
2707                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2708                 else if (strncmp(ent->lde_name, "..",
2709                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2710                            le16_to_cpu(ent->lde_namelen) == 2)
2711                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2712
2713                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2714                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2715                        le64_to_cpu(ent->lde_hash));
2716
2717                 left_bytes -= ent_size;
2718                 ent->lde_reclen = cpu_to_le16(ent_size);
2719                 last_ent = ent;
2720                 ent = (void *)ent + ent_size;
2721         };
2722
2723         last_ent->lde_reclen = 0;
2724
2725         if (ent == start)
2726                 dp->ldp_flags |= LDF_EMPTY;
2727         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2728                 dp->ldp_flags |= LDF_COLLIDE;
2729         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2730         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2731
2732         put_lmv_dir_ctxt(ctxt);
2733         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2734
2735         *ppage = page;
2736
2737         RETURN(0);
2738
2739 free_page:
2740         kunmap(page);
2741         __free_page(page);
2742
2743         return rc;
2744 }
2745
2746 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2747                          struct md_callback *cb_op, __u64 offset,
2748                          struct page **ppage)
2749 {
2750         struct obd_device *obd = exp->exp_obd;
2751         struct lmv_obd *lmv = &obd->u.lmv;
2752         struct lmv_tgt_desc *tgt;
2753         int rc;
2754
2755         ENTRY;
2756
2757         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2758                 RETURN(-ENODATA);
2759
2760         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2761                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2762                 RETURN(rc);
2763         }
2764
2765         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2766         if (IS_ERR(tgt))
2767                 RETURN(PTR_ERR(tgt));
2768
2769         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2770
2771         RETURN(rc);
2772 }
2773
2774 /**
2775  * Unlink a file/directory
2776  *
2777  * Unlink a file or directory under the parent dir. The unlink request
2778  * usually will be sent to the MDT where the child is located, but if
2779  * the client does not have the child FID then request will be sent to the
2780  * MDT where the parent is located.
2781  *
2782  * If the parent is a striped directory then it also needs to locate which
2783  * stripe the name of the child is located, and replace the parent FID
2784  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2785  * it will walk through all of sub-stripes until the child is being
2786  * unlinked finally.
2787  *
2788  * \param[in] exp       export refer to LMV
2789  * \param[in] op_data   different parameters transferred beween client
2790  *                      MD stacks, name, namelen, FIDs etc.
2791  *                      op_fid1 is the parent FID, op_fid2 is the child
2792  *                      FID.
2793  * \param[out] request  point to the request of unlink.
2794  *
2795  * retval               0 if succeed
2796  *                      negative errno if failed.
2797  */
2798 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2799                       struct ptlrpc_request **request)
2800 {
2801         struct obd_device *obd = exp->exp_obd;
2802         struct lmv_obd *lmv = &obd->u.lmv;
2803         struct lmv_tgt_desc *tgt;
2804         struct lmv_tgt_desc *parent_tgt;
2805         struct mdt_body *body;
2806         int rc;
2807
2808         ENTRY;
2809
2810         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2811         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2812         op_data->op_cap = cfs_curproc_cap_pack();
2813
2814 retry:
2815         parent_tgt = lmv_locate_tgt(lmv, op_data);
2816         if (IS_ERR(parent_tgt))
2817                 RETURN(PTR_ERR(parent_tgt));
2818
2819         if (likely(!fid_is_zero(&op_data->op_fid2))) {
2820                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2821                 if (IS_ERR(tgt))
2822                         RETURN(PTR_ERR(tgt));
2823         } else {
2824                 tgt = parent_tgt;
2825         }
2826
2827         /*
2828          * If child's fid is given, cancel unused locks for it if it is from
2829          * another export than parent.
2830          *
2831          * LOOKUP lock for child (fid3) should also be cancelled on parent
2832          * tgt_tgt in mdc_unlink().
2833          */
2834         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2835
2836         if (parent_tgt != tgt)
2837                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2838                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2839                                       MF_MDC_CANCEL_FID3);
2840
2841         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2842                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2843         if (rc)
2844                 RETURN(rc);
2845
2846         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2847                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2848                tgt->ltd_index);
2849
2850         rc = md_unlink(tgt->ltd_exp, op_data, request);
2851         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2852                 ptlrpc_req_finished(*request);
2853                 *request = NULL;
2854                 goto retry;
2855         }
2856
2857         if (rc != -EREMOTE)
2858                 RETURN(rc);
2859
2860         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2861         if (body == NULL)
2862                 RETURN(-EPROTO);
2863
2864         /* Not cross-ref case, just get out of here. */
2865         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2866                 RETURN(rc);
2867
2868         /* This is a remote object, try remote MDT. */
2869         op_data->op_fid2 = body->mbo_fid1;
2870         ptlrpc_req_finished(*request);
2871         *request = NULL;
2872
2873         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2874         if (IS_ERR(tgt))
2875                 RETURN(PTR_ERR(tgt));
2876
2877         goto retry;
2878 }
2879
2880 static int lmv_precleanup(struct obd_device *obd)
2881 {
2882         ENTRY;
2883         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2884         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2885         lprocfs_obd_cleanup(obd);
2886         lprocfs_free_md_stats(obd);
2887         RETURN(0);
2888 }
2889
2890 /**
2891  * Get by key a value associated with a LMV device.
2892  *
2893  * Dispatch request to lower-layer devices as needed.
2894  *
2895  * \param[in] env               execution environment for this thread
2896  * \param[in] exp               export for the LMV device
2897  * \param[in] keylen            length of key identifier
2898  * \param[in] key               identifier of key to get value for
2899  * \param[in] vallen            size of \a val
2900  * \param[out] val              pointer to storage location for value
2901  * \param[in] lsm               optional striping metadata of object
2902  *
2903  * \retval 0            on success
2904  * \retval negative     negated errno on failure
2905  */
2906 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2907                         __u32 keylen, void *key, __u32 *vallen, void *val)
2908 {
2909         struct obd_device *obd;
2910         struct lmv_obd *lmv;
2911         struct lu_tgt_desc *tgt;
2912         int rc = 0;
2913
2914         ENTRY;
2915
2916         obd = class_exp2obd(exp);
2917         if (obd == NULL) {
2918                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2919                        exp->exp_handle.h_cookie);
2920                 RETURN(-EINVAL);
2921         }
2922
2923         lmv = &obd->u.lmv;
2924         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2925                 LASSERT(*vallen == sizeof(__u32));
2926                 lmv_foreach_connected_tgt(lmv, tgt) {
2927                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2928                                           vallen, val))
2929                                 RETURN(0);
2930                 }
2931                 RETURN(-EINVAL);
2932         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2933                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2934                    KEY_IS(KEY_CONN_DATA)) {
2935                 /*
2936                  * Forwarding this request to first MDS, it should know LOV
2937                  * desc.
2938                  */
2939                 tgt = lmv_tgt(lmv, 0);
2940                 if (!tgt)
2941                         RETURN(-ENODEV);
2942
2943                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
2944                 if (!rc && KEY_IS(KEY_CONN_DATA))
2945                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2946                 RETURN(rc);
2947         } else if (KEY_IS(KEY_TGT_COUNT)) {
2948                 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
2949                 RETURN(0);
2950         }
2951
2952         CDEBUG(D_IOCTL, "Invalid key\n");
2953         RETURN(-EINVAL);
2954 }
2955
2956 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
2957                      int *__rcs, struct ptlrpc_request_set *_set)
2958 {
2959         struct obd_device *obd = class_exp2obd(exp);
2960         struct ptlrpc_request_set *set = _set;
2961         struct lmv_obd *lmv = &obd->u.lmv;
2962         int tgt_count = lmv->lmv_mdt_count;
2963         struct lu_tgt_desc *tgt;
2964         struct fid_array *fat, **fas = NULL;
2965         int i, rc, **rcs = NULL;
2966
2967         if (!set) {
2968                 set = ptlrpc_prep_set();
2969                 if (!set)
2970                         RETURN(-ENOMEM);
2971         }
2972
2973         /* split FIDs by targets */
2974         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
2975         if (fas == NULL)
2976                 GOTO(out, rc = -ENOMEM);
2977         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
2978         if (rcs == NULL)
2979                 GOTO(out_fas, rc = -ENOMEM);
2980
2981         for (i = 0; i < fa->fa_nr; i++) {
2982                 unsigned int idx;
2983
2984                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
2985                 if (rc) {
2986                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
2987                                PFID(&fa->fa_fids[i]), rc);
2988                         continue;
2989                 }
2990                 LASSERT(idx < tgt_count);
2991                 if (!fas[idx])
2992                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
2993                                   fa_fids[fa->fa_nr]));
2994                 if (!fas[idx])
2995                         GOTO(out, rc = -ENOMEM);
2996                 if (!rcs[idx])
2997                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
2998                 if (!rcs[idx])
2999                         GOTO(out, rc = -ENOMEM);
3000
3001                 fat = fas[idx];
3002                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3003         }
3004
3005         lmv_foreach_connected_tgt(lmv, tgt) {
3006                 fat = fas[tgt->ltd_index];
3007                 if (!fat || fat->fa_nr == 0)
3008                         continue;
3009                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3010         }
3011
3012         rc = ptlrpc_set_wait(NULL, set);
3013         if (rc == 0) {
3014                 int j = 0;
3015                 for (i = 0; i < tgt_count; i++) {
3016                         fat = fas[i];
3017                         if (!fat || fat->fa_nr == 0)
3018                                 continue;
3019                         /* copy FIDs back */
3020                         memcpy(fa->fa_fids + j, fat->fa_fids,
3021                                fat->fa_nr * sizeof(struct lu_fid));
3022                         /* copy rcs back */
3023                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3024                         j += fat->fa_nr;
3025                 }
3026         }
3027         if (set != _set)
3028                 ptlrpc_set_destroy(set);
3029
3030 out:
3031         for (i = 0; i < tgt_count; i++) {
3032                 if (fas && fas[i])
3033                         OBD_FREE(fas[i], offsetof(struct fid_array,
3034                                                 fa_fids[fa->fa_nr]));
3035                 if (rcs && rcs[i])
3036                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3037         }
3038         if (rcs)
3039                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3040 out_fas:
3041         if (fas)
3042                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3043
3044         RETURN(rc);
3045 }
3046
3047 /**
3048  * Asynchronously set by key a value associated with a LMV device.
3049  *
3050  * Dispatch request to lower-layer devices as needed.
3051  *
3052  * \param[in] env       execution environment for this thread
3053  * \param[in] exp       export for the LMV device
3054  * \param[in] keylen    length of key identifier
3055  * \param[in] key       identifier of key to store value for
3056  * \param[in] vallen    size of value to store
3057  * \param[in] val       pointer to data to be stored
3058  * \param[in] set       optional list of related ptlrpc requests
3059  *
3060  * \retval 0            on success
3061  * \retval negative     negated errno on failure
3062  */
3063 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3064                               __u32 keylen, void *key, __u32 vallen, void *val,
3065                               struct ptlrpc_request_set *set)
3066 {
3067         struct lmv_tgt_desc *tgt;
3068         struct obd_device *obd;
3069         struct lmv_obd *lmv;
3070         int rc = 0;
3071         ENTRY;
3072
3073         obd = class_exp2obd(exp);
3074         if (obd == NULL) {
3075                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3076                        exp->exp_handle.h_cookie);
3077                 RETURN(-EINVAL);
3078         }
3079         lmv = &obd->u.lmv;
3080
3081         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3082             KEY_IS(KEY_DEFAULT_EASIZE)) {
3083                 int err = 0;
3084
3085                 lmv_foreach_connected_tgt(lmv, tgt) {
3086                         err = obd_set_info_async(env, tgt->ltd_exp,
3087                                                  keylen, key, vallen, val, set);
3088                         if (err && rc == 0)
3089                                 rc = err;
3090                 }
3091
3092                 RETURN(rc);
3093         }
3094
3095         RETURN(-EINVAL);
3096 }
3097
3098 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3099                             const struct lmv_mds_md_v1 *lmm1)
3100 {
3101         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3102         int             stripe_count;
3103         int             cplen;
3104         int             i;
3105         int             rc = 0;
3106         ENTRY;
3107
3108         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3109         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3110         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3111         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3112                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3113         else
3114                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3115         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3116         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3117         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3118         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3119                         sizeof(lsm->lsm_md_pool_name));
3120
3121         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3122                 RETURN(-E2BIG);
3123
3124         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3125                "layout_version %d\n", lsm->lsm_md_stripe_count,
3126                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3127                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3128                lsm->lsm_md_layout_version);
3129
3130         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3131         for (i = 0; i < stripe_count; i++) {
3132                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3133                               &lmm1->lmv_stripe_fids[i]);
3134                 /*
3135                  * set default value -1, so lmv_locate_tgt() knows this stripe
3136                  * target is not initialized.
3137                  */
3138                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3139                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3140                         continue;
3141
3142                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3143                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3144                 if (rc == -ENOENT)
3145                         continue;
3146
3147                 if (rc)
3148                         RETURN(rc);
3149
3150                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3151                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3152         }
3153
3154         RETURN(rc);
3155 }
3156
3157 static inline int lmv_unpack_user_md(struct obd_export *exp,
3158                                      struct lmv_stripe_md *lsm,
3159                                      const struct lmv_user_md *lmu)
3160 {
3161         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3162         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3163         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3164         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3165         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3166
3167         return 0;
3168 }
3169
3170 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3171                         const union lmv_mds_md *lmm, size_t lmm_size)
3172 {
3173         struct lmv_stripe_md     *lsm;
3174         int                      lsm_size;
3175         int                      rc;
3176         bool                     allocated = false;
3177         ENTRY;
3178
3179         LASSERT(lsmp != NULL);
3180
3181         lsm = *lsmp;
3182         /* Free memmd */
3183         if (lsm != NULL && lmm == NULL) {
3184                 int i;
3185                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3186
3187                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3188                         size_t lfm_size;
3189
3190                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3191                                                               lfm_value[0]);
3192                         OBD_FREE_LARGE(lfm, lfm_size);
3193                         RETURN(0);
3194                 }
3195
3196                 if (lmv_dir_striped(lsm)) {
3197                         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3198                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
3199                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3200                 } else {
3201                         lsm_size = lmv_stripe_md_size(0);
3202                 }
3203                 OBD_FREE(lsm, lsm_size);
3204                 *lsmp = NULL;
3205                 RETURN(0);
3206         }
3207
3208         /* foreign lmv case */
3209         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3210                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3211
3212                 if (lfm == NULL) {
3213                         OBD_ALLOC_LARGE(lfm, lmm_size);
3214                         if (lfm == NULL)
3215                                 RETURN(-ENOMEM);
3216                         *lsmp = (struct lmv_stripe_md *)lfm;
3217                 }
3218                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3219                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3220                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3221                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3222                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3223                        lfm->lfm_length);
3224                 RETURN(lmm_size);
3225         }
3226
3227         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3228                 RETURN(-EPERM);
3229
3230         /* Unpack memmd */
3231         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3232             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3233                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3234                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3235                        -EIO);
3236                 RETURN(-EIO);
3237         }
3238
3239         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3240                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3241         else
3242                 /**
3243                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3244                  * stripecount should be 0 then.
3245                  */
3246                 lsm_size = lmv_stripe_md_size(0);
3247
3248         if (lsm == NULL) {
3249                 OBD_ALLOC(lsm, lsm_size);
3250                 if (lsm == NULL)
3251                         RETURN(-ENOMEM);
3252                 allocated = true;
3253                 *lsmp = lsm;
3254         }
3255
3256         switch (le32_to_cpu(lmm->lmv_magic)) {
3257         case LMV_MAGIC_V1:
3258                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3259                 break;
3260         case LMV_USER_MAGIC:
3261                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3262                 break;
3263         default:
3264                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3265                        le32_to_cpu(lmm->lmv_magic));
3266                 rc = -EINVAL;
3267                 break;
3268         }
3269
3270         if (rc != 0 && allocated) {
3271                 OBD_FREE(lsm, lsm_size);
3272                 *lsmp = NULL;
3273                 lsm_size = rc;
3274         }
3275         RETURN(lsm_size);
3276 }
3277
3278 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3279 {
3280         lmv_unpackmd(NULL, &lsm, NULL, 0);
3281 }
3282 EXPORT_SYMBOL(lmv_free_memmd);
3283
3284 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3285                              union ldlm_policy_data *policy,
3286                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3287                              void *opaque)
3288 {
3289         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3290         struct lu_tgt_desc *tgt;
3291         int err;
3292         int rc = 0;
3293
3294         ENTRY;
3295
3296         LASSERT(fid != NULL);
3297
3298         lmv_foreach_connected_tgt(lmv, tgt) {
3299                 if (!tgt->ltd_active)