Whamcloud - gitweb
LU-15653 client: able to cleanup devices manually
[fs/lustre-release.git] / lustre / lmv / lmv_obd.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  */
31
32 #define DEBUG_SUBSYSTEM S_LMV
33
34 #include <linux/file.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/user_namespace.h>
38 #include <linux/uidgid.h>
39 #include <linux/slab.h>
40 #include <linux/pagemap.h>
41 #include <linux/mm.h>
42 #include <linux/math64.h>
43 #include <linux/seq_file.h>
44 #include <linux/namei.h>
45
46 #include <obd_support.h>
47 #include <lustre_lib.h>
48 #include <lustre_net.h>
49 #include <obd_class.h>
50 #include <lustre_lmv.h>
51 #include <lprocfs_status.h>
52 #include <cl_object.h>
53 #include <lustre_fid.h>
54 #include <uapi/linux/lustre/lustre_ioctl.h>
55 #include <lustre_kernelcomm.h>
56 #include "lmv_internal.h"
57
58 static int lmv_check_connect(struct obd_device *obd);
59
60 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
61                          int activate)
62 {
63         if (tgt->ltd_active == activate)
64                 return;
65
66         tgt->ltd_active = activate;
67         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
68                 (activate ? 1 : -1);
69
70         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
71 }
72
73 /**
74  * Error codes:
75  *
76  *  -EINVAL  : UUID can't be found in the LMV's target list
77  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
78  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
79  */
80 static int lmv_set_mdc_active(struct lmv_obd *lmv,
81                               const struct obd_uuid *uuid,
82                               int activate)
83 {
84         struct lu_tgt_desc *tgt = NULL;
85         struct obd_device *obd;
86         int rc = 0;
87
88         ENTRY;
89
90         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
91                         lmv, uuid->uuid, activate);
92
93         spin_lock(&lmv->lmv_lock);
94         lmv_foreach_connected_tgt(lmv, tgt) {
95                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
96                        tgt->ltd_index, tgt->ltd_uuid.uuid,
97                        tgt->ltd_exp->exp_handle.h_cookie);
98
99                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
100                         break;
101         }
102
103         if (!tgt)
104                 GOTO(out_lmv_lock, rc = -EINVAL);
105
106         obd = class_exp2obd(tgt->ltd_exp);
107         if (obd == NULL)
108                 GOTO(out_lmv_lock, rc = -ENOTCONN);
109
110         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
111                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
112                obd->obd_type->typ_name, tgt->ltd_index);
113         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
114
115         if (tgt->ltd_active == activate) {
116                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
117                        activate ? "" : "in");
118                 GOTO(out_lmv_lock, rc);
119         }
120
121         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
122                activate ? "" : "in");
123         lmv_activate_target(lmv, tgt, activate);
124         EXIT;
125
126  out_lmv_lock:
127         spin_unlock(&lmv->lmv_lock);
128         return rc;
129 }
130
131 static struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
132 {
133         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
134         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
135
136         return tgt ? obd_get_uuid(tgt->ltd_exp) : NULL;
137 }
138
139 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
140                       enum obd_notify_event ev)
141 {
142         struct obd_connect_data *conn_data;
143         struct lmv_obd          *lmv = &obd->u.lmv;
144         struct obd_uuid         *uuid;
145         int                      rc = 0;
146         ENTRY;
147
148         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
149                 CERROR("unexpected notification of %s %s!\n",
150                        watched->obd_type->typ_name,
151                        watched->obd_name);
152                 RETURN(-EINVAL);
153         }
154
155         uuid = &watched->u.cli.cl_target_uuid;
156         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
157                 /*
158                  * Set MDC as active before notifying the observer, so the
159                  * observer can use the MDC normally.
160                  */
161                 rc = lmv_set_mdc_active(lmv, uuid,
162                                         ev == OBD_NOTIFY_ACTIVE);
163                 if (rc) {
164                         CERROR("%sactivation of %s failed: %d\n",
165                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
166                                uuid->uuid, rc);
167                         RETURN(rc);
168                 }
169         } else if (ev == OBD_NOTIFY_OCD) {
170                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
171                 /*
172                  * XXX: Make sure that ocd_connect_flags from all targets are
173                  * the same. Otherwise one of MDTs runs wrong version or
174                  * something like this.  --umka
175                  */
176                 obd->obd_self_export->exp_connect_data = *conn_data;
177         }
178
179         /*
180          * Pass the notification up the chain.
181          */
182         if (obd->obd_observer)
183                 rc = obd_notify(obd->obd_observer, watched, ev);
184
185         RETURN(rc);
186 }
187
188 static int lmv_connect(const struct lu_env *env,
189                        struct obd_export **pexp, struct obd_device *obd,
190                        struct obd_uuid *cluuid, struct obd_connect_data *data,
191                        void *localdata)
192 {
193         struct lmv_obd *lmv = &obd->u.lmv;
194         struct lustre_handle conn = { 0 };
195         struct obd_export *exp;
196         int rc;
197         ENTRY;
198
199         rc = class_connect(&conn, obd, cluuid);
200         if (rc) {
201                 CERROR("class_connection() returned %d\n", rc);
202                 RETURN(rc);
203         }
204
205         exp = class_conn2export(&conn);
206
207         lmv->connected = 0;
208         lmv->conn_data = *data;
209         lmv->lmv_cache = localdata;
210
211         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
212                                                     &obd->obd_kset.kobj);
213         if (!lmv->lmv_tgts_kobj) {
214                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
215                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
216         }
217
218         rc = lmv_check_connect(obd);
219         if (rc != 0)
220                 GOTO(out_sysfs, rc);
221
222         *pexp = exp;
223
224         RETURN(rc);
225
226 out_sysfs:
227         if (lmv->lmv_tgts_kobj)
228                 kobject_put(lmv->lmv_tgts_kobj);
229
230         class_disconnect(exp);
231
232         return rc;
233 }
234
235 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
236                             __u32 def_easize)
237 {
238         struct obd_device *obd = exp->exp_obd;
239         struct lmv_obd *lmv = &obd->u.lmv;
240         struct lmv_tgt_desc *tgt;
241         int change = 0;
242         int rc = 0;
243
244         ENTRY;
245
246         if (lmv->max_easize < easize) {
247                 lmv->max_easize = easize;
248                 change = 1;
249         }
250         if (lmv->max_def_easize < def_easize) {
251                 lmv->max_def_easize = def_easize;
252                 change = 1;
253         }
254
255         if (change == 0)
256                 RETURN(0);
257
258         if (lmv->connected == 0)
259                 RETURN(0);
260
261         lmv_foreach_connected_tgt(lmv, tgt) {
262                 if (!tgt->ltd_active)
263                         continue;
264
265                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
266                 if (rc) {
267                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
268                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
269                         break;
270                 }
271         }
272         RETURN(rc);
273 }
274
275 #define MAX_STRING_SIZE 128
276
277 static int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
278 {
279         struct lmv_obd *lmv = &obd->u.lmv;
280         struct obd_device *mdc_obd;
281         struct obd_export *mdc_exp;
282         struct lu_fld_target target;
283         int  rc;
284         ENTRY;
285
286         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
287                                         &obd->obd_uuid);
288         if (!mdc_obd) {
289                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
290                 RETURN(-EINVAL);
291         }
292
293         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
294                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
295                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
296
297         if (!mdc_obd->obd_set_up) {
298                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
299                 RETURN(-EINVAL);
300         }
301
302         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
303                          &lmv->conn_data, lmv->lmv_cache);
304         if (rc) {
305                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
306                 RETURN(rc);
307         }
308
309         /*
310          * Init fid sequence client for this mdc and add new fld target.
311          */
312         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
313         if (rc)
314                 RETURN(rc);
315
316         target.ft_srv = NULL;
317         target.ft_exp = mdc_exp;
318         target.ft_idx = tgt->ltd_index;
319
320         fld_client_add_target(&lmv->lmv_fld, &target);
321
322         rc = obd_register_observer(mdc_obd, obd);
323         if (rc) {
324                 obd_disconnect(mdc_exp);
325                 CERROR("target %s register_observer error %d\n",
326                        tgt->ltd_uuid.uuid, rc);
327                 RETURN(rc);
328         }
329
330         if (obd->obd_observer) {
331                 /*
332                  * Tell the observer about the new target.
333                  */
334                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
335                                 OBD_NOTIFY_ACTIVE);
336                 if (rc) {
337                         obd_disconnect(mdc_exp);
338                         RETURN(rc);
339                 }
340         }
341
342         tgt->ltd_active = 1;
343         tgt->ltd_exp = mdc_exp;
344         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
345
346         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
347
348         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
349         if (rc) {
350                 obd_disconnect(mdc_exp);
351                 RETURN(rc);
352         }
353
354         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
355                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
356                atomic_read(&obd->obd_refcount));
357
358         lmv_statfs_check_update(obd, tgt);
359
360         if (lmv->lmv_tgts_kobj)
361                 /* Even if we failed to create the link, that's fine */
362                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
363                                        &mdc_obd->obd_kset.kobj,
364                                        mdc_obd->obd_name);
365         RETURN(0);
366 }
367
368 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
369 {
370         LASSERT(tgt);
371         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
372         OBD_FREE_PTR(tgt);
373 }
374
375 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
376                            __u32 index, int gen)
377 {
378         struct obd_device *mdc_obd;
379         struct lmv_obd *lmv = &obd->u.lmv;
380         struct lmv_tgt_desc *tgt;
381         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
382         int rc = 0;
383
384         ENTRY;
385
386         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
387         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
388                                         &obd->obd_uuid);
389         if (!mdc_obd) {
390                 CERROR("%s: Target %s not attached: rc = %d\n",
391                        obd->obd_name, uuidp->uuid, -EINVAL);
392                 RETURN(-EINVAL);
393         }
394
395         OBD_ALLOC_PTR(tgt);
396         if (!tgt)
397                 RETURN(-ENOMEM);
398
399         mutex_init(&tgt->ltd_fid_mutex);
400         tgt->ltd_index = index;
401         tgt->ltd_uuid = *uuidp;
402         tgt->ltd_active = 0;
403
404         mutex_lock(&ltd->ltd_mutex);
405         rc = ltd_add_tgt(ltd, tgt);
406         mutex_unlock(&ltd->ltd_mutex);
407
408         if (rc)
409                 GOTO(out_tgt, rc);
410
411         if (!lmv->connected)
412                 /* lmv_check_connect() will connect this target. */
413                 RETURN(0);
414
415         rc = lmv_connect_mdc(obd, tgt);
416         if (!rc) {
417                 int easize = sizeof(struct lmv_stripe_md) +
418                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
419
420                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
421         }
422
423         RETURN(rc);
424
425 out_tgt:
426         OBD_FREE_PTR(tgt);
427         return rc;
428 }
429
430 static int lmv_check_connect(struct obd_device *obd)
431 {
432         struct lmv_obd *lmv = &obd->u.lmv;
433         struct lmv_tgt_desc *tgt;
434         int easize;
435         int rc;
436
437         ENTRY;
438
439         if (lmv->connected)
440                 RETURN(0);
441
442         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
443         if (lmv->connected)
444                 GOTO(unlock, rc = 0);
445
446         if (!lmv->lmv_mdt_count) {
447                 CERROR("%s: no targets configured: rc = -EINVAL\n",
448                        obd->obd_name);
449                 GOTO(unlock, rc = -EINVAL);
450         }
451
452         if (!lmv_mdt0_inited(lmv)) {
453                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
454                        obd->obd_name);
455                 GOTO(unlock, rc = -EINVAL);
456         }
457
458         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
459                obd->obd_uuid.uuid, obd->obd_name);
460
461         lmv_foreach_tgt(lmv, tgt) {
462                 rc = lmv_connect_mdc(obd, tgt);
463                 if (rc)
464                         GOTO(out_disc, rc);
465         }
466
467         lmv->connected = 1;
468         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
469         lmv_init_ea_size(obd->obd_self_export, easize, 0);
470         EXIT;
471 unlock:
472         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
473
474         return rc;
475
476 out_disc:
477         lmv_foreach_tgt(lmv, tgt) {
478                 tgt->ltd_active = 0;
479                 if (!tgt->ltd_exp)
480                         continue;
481
482                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
483                 obd_disconnect(tgt->ltd_exp);
484         }
485
486         goto unlock;
487 }
488
489 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
490 {
491         struct lmv_obd *lmv = &obd->u.lmv;
492         struct obd_device *mdc_obd;
493         int rc;
494         ENTRY;
495
496         LASSERT(tgt != NULL);
497         LASSERT(obd != NULL);
498
499         mdc_obd = class_exp2obd(tgt->ltd_exp);
500
501         if (mdc_obd) {
502                 mdc_obd->obd_force = obd->obd_force;
503                 mdc_obd->obd_fail = obd->obd_fail;
504                 mdc_obd->obd_no_recov = obd->obd_no_recov;
505
506                 if (lmv->lmv_tgts_kobj)
507                         sysfs_remove_link(lmv->lmv_tgts_kobj,
508                                           mdc_obd->obd_name);
509         }
510
511         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
512         if (rc)
513                 CERROR("Can't finalize fids factory\n");
514
515         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
516                tgt->ltd_exp->exp_obd->obd_name,
517                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
518
519         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
520         rc = obd_disconnect(tgt->ltd_exp);
521         if (rc) {
522                 if (tgt->ltd_active) {
523                         CERROR("Target %s disconnect error %d\n",
524                                tgt->ltd_uuid.uuid, rc);
525                 }
526         }
527
528         lmv_activate_target(lmv, tgt, 0);
529         tgt->ltd_exp = NULL;
530         RETURN(0);
531 }
532
533 static int lmv_disconnect(struct obd_export *exp)
534 {
535         struct obd_device *obd = class_exp2obd(exp);
536         struct lmv_obd *lmv = &obd->u.lmv;
537         struct lmv_tgt_desc *tgt;
538         int rc;
539
540         ENTRY;
541
542         lmv_foreach_connected_tgt(lmv, tgt)
543                 lmv_disconnect_mdc(obd, tgt);
544
545         if (lmv->lmv_tgts_kobj)
546                 kobject_put(lmv->lmv_tgts_kobj);
547
548         lmv->connected = 0;
549         rc = class_disconnect(exp);
550
551         RETURN(rc);
552 }
553
554 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
555                         void __user *uarg)
556 {
557         struct obd_device *obd = class_exp2obd(exp);
558         struct lmv_obd *lmv = &obd->u.lmv;
559         struct getinfo_fid2path *gf;
560         struct lmv_tgt_desc *tgt;
561         struct getinfo_fid2path *remote_gf = NULL;
562         struct lu_fid root_fid;
563         int remote_gf_size = 0;
564         int rc;
565
566         gf = karg;
567         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
568         if (IS_ERR(tgt))
569                 RETURN(PTR_ERR(tgt));
570
571         root_fid = *gf->gf_u.gf_root_fid;
572         LASSERT(fid_is_sane(&root_fid));
573
574 repeat_fid2path:
575         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
576         if (rc != 0 && rc != -EREMOTE)
577                 GOTO(out_fid2path, rc);
578
579         /* If remote_gf != NULL, it means just building the
580          * path on the remote MDT, copy this path segement to gf */
581         if (remote_gf != NULL) {
582                 struct getinfo_fid2path *ori_gf;
583                 char *ptr;
584                 int len;
585
586                 ori_gf = (struct getinfo_fid2path *)karg;
587                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
588                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
589                         GOTO(out_fid2path, rc = -EOVERFLOW);
590
591                 ptr = ori_gf->gf_u.gf_path;
592
593                 len = strlen(gf->gf_u.gf_path);
594                 /* move the current path to the right to release space
595                  * for closer-to-root part */
596                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
597                 memcpy(ptr, gf->gf_u.gf_path, len);
598                 ptr[len] = '/';
599         }
600
601         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
602                tgt->ltd_exp->exp_obd->obd_name,
603                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
604                gf->gf_linkno);
605
606         if (rc == 0)
607                 GOTO(out_fid2path, rc);
608
609         /* sigh, has to go to another MDT to do path building further */
610         if (remote_gf == NULL) {
611                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
612                 OBD_ALLOC(remote_gf, remote_gf_size);
613                 if (remote_gf == NULL)
614                         GOTO(out_fid2path, rc = -ENOMEM);
615                 remote_gf->gf_pathlen = PATH_MAX;
616         }
617
618         if (!fid_is_sane(&gf->gf_fid)) {
619                 CERROR("%s: invalid FID "DFID": rc = %d\n",
620                        tgt->ltd_exp->exp_obd->obd_name,
621                        PFID(&gf->gf_fid), -EINVAL);
622                 GOTO(out_fid2path, rc = -EINVAL);
623         }
624
625         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
626         if (IS_ERR(tgt))
627                 GOTO(out_fid2path, rc = -EINVAL);
628
629         remote_gf->gf_fid = gf->gf_fid;
630         remote_gf->gf_recno = -1;
631         remote_gf->gf_linkno = -1;
632         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
633         *remote_gf->gf_u.gf_root_fid = root_fid;
634         gf = remote_gf;
635         goto repeat_fid2path;
636
637 out_fid2path:
638         if (remote_gf != NULL)
639                 OBD_FREE(remote_gf, remote_gf_size);
640         RETURN(rc);
641 }
642
643 static int lmv_hsm_req_count(struct lmv_obd *lmv,
644                              const struct hsm_user_request *hur,
645                              const struct lmv_tgt_desc *tgt_mds)
646 {
647         struct lmv_tgt_desc *curr_tgt;
648         __u32 i;
649         int nr = 0;
650
651         /* count how many requests must be sent to the given target */
652         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
653                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
654                 if (IS_ERR(curr_tgt))
655                         RETURN(PTR_ERR(curr_tgt));
656                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
657                         nr++;
658         }
659         return nr;
660 }
661
662 static int lmv_hsm_req_build(struct lmv_obd *lmv,
663                               struct hsm_user_request *hur_in,
664                               const struct lmv_tgt_desc *tgt_mds,
665                               struct hsm_user_request *hur_out)
666 {
667         __u32 i, nr_out;
668         struct lmv_tgt_desc *curr_tgt;
669
670         /* build the hsm_user_request for the given target */
671         hur_out->hur_request = hur_in->hur_request;
672         nr_out = 0;
673         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
674                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
675                 if (IS_ERR(curr_tgt))
676                         RETURN(PTR_ERR(curr_tgt));
677                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
678                         hur_out->hur_user_item[nr_out] =
679                                                 hur_in->hur_user_item[i];
680                         nr_out++;
681                 }
682         }
683         hur_out->hur_request.hr_itemcount = nr_out;
684         memcpy(hur_data(hur_out), hur_data(hur_in),
685                hur_in->hur_request.hr_data_len);
686
687         RETURN(0);
688 }
689
690 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
691                                  int len, struct lustre_kernelcomm *lk,
692                                  void __user *uarg)
693 {
694         struct lmv_obd *lmv = &obd->u.lmv;
695         struct lu_tgt_desc *tgt;
696         int rc;
697
698         ENTRY;
699
700         /* unregister request (call from llapi_hsm_copytool_fini) */
701         lmv_foreach_connected_tgt(lmv, tgt)
702                 /* best effort: try to clean as much as possible
703                  * (continue on error) */
704                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
705
706         /* Whatever the result, remove copytool from kuc groups.
707          * Unreached coordinators will get EPIPE on next requests
708          * and will unregister automatically.
709          */
710         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
711
712         RETURN(rc);
713 }
714
715 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
716                                int len, struct lustre_kernelcomm *lk,
717                                void __user *uarg)
718 {
719         struct lmv_obd *lmv = &obd->u.lmv;
720         struct file *filp;
721         bool any_set = false;
722         struct kkuc_ct_data *kcd;
723         size_t kcd_size;
724         struct lu_tgt_desc *tgt;
725         __u32 i;
726         int err;
727         int rc = 0;
728
729         ENTRY;
730
731         filp = fget(lk->lk_wfd);
732         if (!filp)
733                 RETURN(-EBADF);
734
735         if (lk->lk_flags & LK_FLG_DATANR)
736                 kcd_size = offsetof(struct kkuc_ct_data,
737                                     kcd_archives[lk->lk_data_count]);
738         else
739                 kcd_size = sizeof(*kcd);
740
741         OBD_ALLOC(kcd, kcd_size);
742         if (kcd == NULL)
743                 GOTO(err_fput, rc = -ENOMEM);
744
745         kcd->kcd_nr_archives = lk->lk_data_count;
746         if (lk->lk_flags & LK_FLG_DATANR) {
747                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
748                 if (lk->lk_data_count > 0)
749                         memcpy(kcd->kcd_archives, lk->lk_data,
750                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
751         } else {
752                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
753         }
754
755         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
756                                    lk->lk_group, kcd, kcd_size);
757         OBD_FREE(kcd, kcd_size);
758         if (rc)
759                 GOTO(err_fput, rc);
760
761         /* All or nothing: try to register to all MDS.
762          * In case of failure, unregister from previous MDS,
763          * except if it because of inactive target. */
764         lmv_foreach_connected_tgt(lmv, tgt) {
765                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
766                 if (err) {
767                         if (tgt->ltd_active) {
768                                 /* permanent error */
769                                 CERROR("%s: iocontrol MDC %s on MDT"
770                                        " idx %d cmd %x: err = %d\n",
771                                        lmv2obd_dev(lmv)->obd_name,
772                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
773                                        err);
774                                 rc = err;
775                                 lk->lk_flags |= LK_FLG_STOP;
776                                 i = tgt->ltd_index;
777                                 /* unregister from previous MDS */
778                                 lmv_foreach_connected_tgt(lmv, tgt) {
779                                         if (tgt->ltd_index >= i)
780                                                 break;
781
782                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
783                                                       lk, uarg);
784                                 }
785                                 GOTO(err_kkuc_rem, rc);
786                         }
787                         /* else: transient error.
788                          * kuc will register to the missing MDT
789                          * when it is back */
790                 } else {
791                         any_set = true;
792                 }
793         }
794
795         if (!any_set)
796                 /* no registration done: return error */
797                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
798
799         RETURN(0);
800
801 err_kkuc_rem:
802         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
803
804 err_fput:
805         fput(filp);
806         return rc;
807 }
808
809 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
810                          int len, void *karg, void __user *uarg)
811 {
812         struct obd_device *obd = class_exp2obd(exp);
813         struct lmv_obd *lmv = &obd->u.lmv;
814         struct lu_tgt_desc *tgt = NULL;
815         int set = 0;
816         __u32 count = lmv->lmv_mdt_count;
817         int rc = 0;
818
819         ENTRY;
820
821         if (count == 0)
822                 RETURN(-ENOTTY);
823
824         switch (cmd) {
825         case IOC_OBD_STATFS: {
826                 struct obd_ioctl_data *data = karg;
827                 struct obd_device *mdc_obd;
828                 struct obd_statfs stat_buf = {0};
829                 __u32 index;
830
831                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
832
833                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
834                         RETURN(-ENODEV);
835
836                 tgt = lmv_tgt(lmv, index);
837                 if (!tgt)
838                         RETURN(-EAGAIN);
839
840                 if (!tgt->ltd_active)
841                         RETURN(-ENODATA);
842
843                 mdc_obd = class_exp2obd(tgt->ltd_exp);
844                 if (!mdc_obd)
845                         RETURN(-EINVAL);
846
847                 /* copy UUID */
848                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
849                                  min((int) data->ioc_plen2,
850                                      (int) sizeof(struct obd_uuid))))
851                         RETURN(-EFAULT);
852
853                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
854                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
855                                 0);
856                 if (rc)
857                         RETURN(rc);
858                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
859                                  min((int) data->ioc_plen1,
860                                      (int) sizeof(stat_buf))))
861                         RETURN(-EFAULT);
862                 break;
863         }
864         case OBD_IOC_QUOTACTL: {
865                 struct if_quotactl *qctl = karg;
866                 struct obd_quotactl *oqctl;
867
868                 if (qctl->qc_valid == QC_MDTIDX) {
869                         tgt = lmv_tgt(lmv, qctl->qc_idx);
870                 } else if (qctl->qc_valid == QC_UUID) {
871                         lmv_foreach_tgt(lmv, tgt) {
872                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
873                                                      &qctl->obd_uuid))
874                                         continue;
875
876                                 if (!tgt->ltd_exp)
877                                         RETURN(-EINVAL);
878
879                                 break;
880                         }
881                 } else {
882                         RETURN(-EINVAL);
883                 }
884
885                 if (!tgt || !tgt->ltd_exp)
886                         RETURN(-EINVAL);
887
888                 OBD_ALLOC_PTR(oqctl);
889                 if (!oqctl)
890                         RETURN(-ENOMEM);
891
892                 QCTL_COPY(oqctl, qctl);
893                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
894                 if (rc == 0) {
895                         QCTL_COPY(qctl, oqctl);
896                         qctl->qc_valid = QC_MDTIDX;
897                         qctl->obd_uuid = tgt->ltd_uuid;
898                 }
899                 OBD_FREE_PTR(oqctl);
900                 break;
901         }
902         case LL_IOC_GET_CONNECT_FLAGS: {
903                 tgt = lmv_tgt(lmv, 0);
904                 rc = -ENODATA;
905                 if (tgt && tgt->ltd_exp)
906                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
907                 break;
908         }
909         case LL_IOC_FID2MDTIDX: {
910                 struct lu_fid *fid = karg;
911                 int             mdt_index;
912
913                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
914                 if (rc != 0)
915                         RETURN(rc);
916
917                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
918                  * point to user space memory for FID2MDTIDX. */
919                 *(__u32 *)uarg = mdt_index;
920                 break;
921         }
922         case OBD_IOC_FID2PATH: {
923                 rc = lmv_fid2path(exp, len, karg, uarg);
924                 break;
925         }
926         case LL_IOC_HSM_STATE_GET:
927         case LL_IOC_HSM_STATE_SET:
928         case LL_IOC_HSM_ACTION: {
929                 struct md_op_data *op_data = karg;
930
931                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
932                 if (IS_ERR(tgt))
933                         RETURN(PTR_ERR(tgt));
934
935                 if (tgt->ltd_exp == NULL)
936                         RETURN(-EINVAL);
937
938                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
939                 break;
940         }
941         case LL_IOC_HSM_PROGRESS: {
942                 const struct hsm_progress_kernel *hpk = karg;
943
944                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
945                 if (IS_ERR(tgt))
946                         RETURN(PTR_ERR(tgt));
947                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
948                 break;
949         }
950         case LL_IOC_HSM_REQUEST: {
951                 struct hsm_user_request *hur = karg;
952                 unsigned int reqcount = hur->hur_request.hr_itemcount;
953
954                 if (reqcount == 0)
955                         RETURN(0);
956
957                 /* if the request is about a single fid
958                  * or if there is a single MDS, no need to split
959                  * the request. */
960                 if (reqcount == 1 || count == 1) {
961                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
962                         if (IS_ERR(tgt))
963                                 RETURN(PTR_ERR(tgt));
964                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
965                 } else {
966                         /* split fid list to their respective MDS */
967                         lmv_foreach_connected_tgt(lmv, tgt) {
968                                 int nr, rc1;
969                                 size_t reqlen;
970                                 struct hsm_user_request *req;
971
972                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
973                                 if (nr < 0)
974                                         RETURN(nr);
975                                 if (nr == 0) /* nothing for this MDS */
976                                         continue;
977
978                                 /* build a request with fids for this MDS */
979                                 reqlen = offsetof(typeof(*hur),
980                                                   hur_user_item[nr])
981                                                 + hur->hur_request.hr_data_len;
982                                 OBD_ALLOC_LARGE(req, reqlen);
983                                 if (req == NULL)
984                                         RETURN(-ENOMEM);
985                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
986                                 if (rc1 < 0)
987                                         GOTO(hsm_req_err, rc1);
988                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
989                                                     req, uarg);
990 hsm_req_err:
991                                 if (rc1 != 0 && rc == 0)
992                                         rc = rc1;
993                                 OBD_FREE_LARGE(req, reqlen);
994                         }
995                 }
996                 break;
997         }
998         case LL_IOC_LOV_SWAP_LAYOUTS: {
999                 struct md_op_data *op_data = karg;
1000                 struct lmv_tgt_desc *tgt1, *tgt2;
1001
1002                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1003                 if (IS_ERR(tgt1))
1004                         RETURN(PTR_ERR(tgt1));
1005
1006                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1007                 if (IS_ERR(tgt2))
1008                         RETURN(PTR_ERR(tgt2));
1009
1010                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1011                         RETURN(-EINVAL);
1012
1013                 /* only files on same MDT can have their layouts swapped */
1014                 if (tgt1->ltd_index != tgt2->ltd_index)
1015                         RETURN(-EPERM);
1016
1017                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1018                 break;
1019         }
1020         case LL_IOC_HSM_CT_START: {
1021                 struct lustre_kernelcomm *lk = karg;
1022                 if (lk->lk_flags & LK_FLG_STOP)
1023                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1024                 else
1025                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1026                 break;
1027         }
1028         default:
1029                 lmv_foreach_connected_tgt(lmv, tgt) {
1030                         struct obd_device *mdc_obd;
1031                         int err;
1032
1033                         /* ll_umount_begin() sets force flag but for lmv, not
1034                          * mdc. Let's pass it through */
1035                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1036                         mdc_obd->obd_force = obd->obd_force;
1037                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1038                         if (err) {
1039                                 if (tgt->ltd_active) {
1040                                         CERROR("error: iocontrol MDC %s on MDT"
1041                                                " idx %d cmd %x: err = %d\n",
1042                                                tgt->ltd_uuid.uuid,
1043                                                tgt->ltd_index, cmd, err);
1044                                         if (!rc)
1045                                                 rc = err;
1046                                 }
1047                         } else
1048                                 set = 1;
1049                 }
1050                 if (!set && !rc)
1051                         rc = -EIO;
1052         }
1053         RETURN(rc);
1054 }
1055
1056 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1057                   struct lu_fid *fid, struct md_op_data *op_data)
1058 {
1059         struct obd_device *obd = class_exp2obd(exp);
1060         struct lmv_obd *lmv = &obd->u.lmv;
1061         struct lmv_tgt_desc *tgt;
1062         int rc;
1063
1064         ENTRY;
1065
1066         LASSERT(op_data);
1067         LASSERT(fid);
1068
1069         tgt = lmv_tgt(lmv, op_data->op_mds);
1070         if (!tgt)
1071                 RETURN(-ENODEV);
1072
1073         if (!tgt->ltd_active || !tgt->ltd_exp)
1074                 RETURN(-ENODEV);
1075
1076         /*
1077          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1078          * on server that seq in new allocated fid is not yet known.
1079          */
1080         mutex_lock(&tgt->ltd_fid_mutex);
1081         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1082         mutex_unlock(&tgt->ltd_fid_mutex);
1083         if (rc > 0) {
1084                 LASSERT(fid_is_sane(fid));
1085                 rc = 0;
1086         }
1087
1088         RETURN(rc);
1089 }
1090
1091 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1092 {
1093         struct lmv_obd *lmv = &obd->u.lmv;
1094         struct lmv_desc *desc;
1095         struct lnet_processid lnet_id;
1096         int i = 0;
1097         int rc;
1098
1099         ENTRY;
1100
1101         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1102                 CERROR("LMV setup requires a descriptor\n");
1103                 RETURN(-EINVAL);
1104         }
1105
1106         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1107         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1108                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1109                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1110                 RETURN(-EINVAL);
1111         }
1112
1113         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1114                      desc->ld_uuid.uuid);
1115         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1116         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1117         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1118                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1119         lmv->max_def_easize = 0;
1120         lmv->max_easize = 0;
1121
1122         spin_lock_init(&lmv->lmv_lock);
1123
1124         /*
1125          * initialize rr_index to lower 32bit of netid, so that client
1126          * can distribute subdirs evenly from the beginning.
1127          */
1128         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1129                 if (!nid_is_lo0(&lnet_id.nid)) {
1130                         lmv->lmv_qos_rr_index = ntohl(lnet_id.nid.nid_addr[0]);
1131                         break;
1132                 }
1133         }
1134
1135         rc = lmv_tunables_init(obd);
1136         if (rc)
1137                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1138                       obd->obd_name, rc);
1139
1140         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1141                              LUSTRE_CLI_FLD_HASH_DHT);
1142         if (rc)
1143                 CERROR("Can't init FLD, err %d\n", rc);
1144
1145         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1146         if (rc)
1147                 CWARN("%s: error initialize target table: rc = %d\n",
1148                       obd->obd_name, rc);
1149
1150         RETURN(rc);
1151 }
1152
1153 static int lmv_cleanup(struct obd_device *obd)
1154 {
1155         struct lmv_obd *lmv = &obd->u.lmv;
1156         struct lu_tgt_desc *tgt;
1157         struct lu_tgt_desc *tmp;
1158
1159         ENTRY;
1160
1161         fld_client_fini(&lmv->lmv_fld);
1162         fld_client_debugfs_fini(&lmv->lmv_fld);
1163
1164         lprocfs_obd_cleanup(obd);
1165         lprocfs_free_md_stats(obd);
1166
1167         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1168                 lmv_del_target(lmv, tgt);
1169         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1170
1171         RETURN(0);
1172 }
1173
1174 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1175 {
1176         struct lustre_cfg       *lcfg = buf;
1177         struct obd_uuid         obd_uuid;
1178         int                     gen;
1179         __u32                   index;
1180         int                     rc;
1181         ENTRY;
1182
1183         switch (lcfg->lcfg_command) {
1184         case LCFG_ADD_MDC:
1185                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1186                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1187                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1188                         GOTO(out, rc = -EINVAL);
1189
1190                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1191
1192                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1193                         GOTO(out, rc = -EINVAL);
1194                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1195                         GOTO(out, rc = -EINVAL);
1196                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1197                 GOTO(out, rc);
1198         default:
1199                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1200                 GOTO(out, rc = -EINVAL);
1201         }
1202 out:
1203         RETURN(rc);
1204 }
1205
1206 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1207 {
1208         int i;
1209
1210         if (flags & OBD_STATFS_FOR_MDT0)
1211                 return 0;
1212
1213         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1214                 return lmv->lmv_statfs_start;
1215
1216         /* choose initial MDT for this client */
1217         for (i = 0;; i++) {
1218                 struct lnet_processid lnet_id;
1219                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1220                         break;
1221
1222                 if (!nid_is_lo0(&lnet_id.nid)) {
1223                         /* We dont need a full 64-bit modulus, just enough
1224                          * to distribute the requests across MDTs evenly.
1225                          */
1226                         lmv->lmv_statfs_start = nidhash(&lnet_id.nid) %
1227                                                 lmv->lmv_mdt_count;
1228                         break;
1229                 }
1230         }
1231
1232         return lmv->lmv_statfs_start;
1233 }
1234
1235 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1236                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1237 {
1238         struct obd_device *obd = class_exp2obd(exp);
1239         struct lmv_obd *lmv = &obd->u.lmv;
1240         struct obd_statfs *temp;
1241         struct lu_tgt_desc *tgt;
1242         __u32 i;
1243         __u32 idx;
1244         int rc = 0;
1245         int err = 0;
1246
1247         ENTRY;
1248
1249         OBD_ALLOC(temp, sizeof(*temp));
1250         if (temp == NULL)
1251                 RETURN(-ENOMEM);
1252
1253         /* distribute statfs among MDTs */
1254         idx = lmv_select_statfs_mdt(lmv, flags);
1255
1256         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1257                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1258                 tgt = lmv_tgt(lmv, idx);
1259                 if (!tgt || !tgt->ltd_exp)
1260                         continue;
1261
1262                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1263                                 flags | OBD_STATFS_NESTED);
1264                 if (rc) {
1265                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1266                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1267                         err = rc;
1268                         /* Try another MDT */
1269                         if (flags & OBD_STATFS_SUM)
1270                                 continue;
1271                         GOTO(out_free_temp, rc);
1272                 }
1273
1274                 if (temp->os_state & OS_STATFS_SUM ||
1275                     flags == OBD_STATFS_FOR_MDT0) {
1276                         /* reset to the last aggregated values
1277                          * and don't sum with non-aggrated data */
1278                         /* If the statfs is from mount, it needs to retrieve
1279                          * necessary information from MDT0. i.e. mount does
1280                          * not need the merged osfs from all of MDT. Also
1281                          * clients can be mounted as long as MDT0 is in
1282                          * service */
1283                         *osfs = *temp;
1284                         GOTO(out_free_temp, rc);
1285                 }
1286
1287                 if (i == 0) {
1288                         *osfs = *temp;
1289                 } else {
1290                         osfs->os_bavail += temp->os_bavail;
1291                         osfs->os_blocks += temp->os_blocks;
1292                         osfs->os_ffree += temp->os_ffree;
1293                         osfs->os_files += temp->os_files;
1294                         osfs->os_granted += temp->os_granted;
1295                 }
1296         }
1297         /* There is no stats from some MDTs, data incomplete */
1298         if (err)
1299                 rc = err;
1300 out_free_temp:
1301         OBD_FREE(temp, sizeof(*temp));
1302         RETURN(rc);
1303 }
1304
1305 static int lmv_statfs_update(void *cookie, int rc)
1306 {
1307         struct obd_info *oinfo = cookie;
1308         struct obd_device *obd = oinfo->oi_obd;
1309         struct lmv_obd *lmv = &obd->u.lmv;
1310         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1311         struct obd_statfs *osfs = oinfo->oi_osfs;
1312
1313         /*
1314          * NB: don't deactivate TGT upon error, because we may not trigger async
1315          * statfs any longer, then there is no chance to activate TGT.
1316          */
1317         if (!rc) {
1318                 spin_lock(&lmv->lmv_lock);
1319                 tgt->ltd_statfs = *osfs;
1320                 tgt->ltd_statfs_age = ktime_get_seconds();
1321                 spin_unlock(&lmv->lmv_lock);
1322                 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1323         }
1324
1325         return rc;
1326 }
1327
1328 /* update tgt statfs async if it's ld_qos_maxage old */
1329 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1330 {
1331         struct obd_info oinfo = {
1332                 .oi_obd = obd,
1333                 .oi_tgt = tgt,
1334                 .oi_cb_up = lmv_statfs_update,
1335         };
1336         int rc;
1337
1338         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1339             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1340                 return 0;
1341
1342         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1343
1344         return rc;
1345 }
1346
1347 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1348                         struct lu_fid *fid)
1349 {
1350         struct obd_device *obd = exp->exp_obd;
1351         struct lmv_obd *lmv = &obd->u.lmv;
1352         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1353         int rc;
1354
1355         ENTRY;
1356
1357         if (!tgt)
1358                 RETURN(-ENODEV);
1359
1360         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1361         RETURN(rc);
1362 }
1363
1364 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1365                         u64 obd_md_valid, const char *name, size_t buf_size,
1366                         struct ptlrpc_request **req)
1367 {
1368         struct obd_device *obd = exp->exp_obd;
1369         struct lmv_obd *lmv = &obd->u.lmv;
1370         struct lmv_tgt_desc *tgt;
1371         int rc;
1372
1373         ENTRY;
1374
1375         tgt = lmv_fid2tgt(lmv, fid);
1376         if (IS_ERR(tgt))
1377                 RETURN(PTR_ERR(tgt));
1378
1379         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1380
1381         RETURN(rc);
1382 }
1383
1384 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1385                         u64 obd_md_valid, const char *name,
1386                         const void *value, size_t value_size,
1387                         unsigned int xattr_flags, u32 suppgid,
1388                         struct ptlrpc_request **req)
1389 {
1390         struct obd_device *obd = exp->exp_obd;
1391         struct lmv_obd *lmv = &obd->u.lmv;
1392         struct lmv_tgt_desc *tgt;
1393         int rc;
1394
1395         ENTRY;
1396
1397         tgt = lmv_fid2tgt(lmv, fid);
1398         if (IS_ERR(tgt))
1399                 RETURN(PTR_ERR(tgt));
1400
1401         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1402                          value, value_size, xattr_flags, suppgid, req);
1403
1404         RETURN(rc);
1405 }
1406
1407 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1408                        struct ptlrpc_request **request)
1409 {
1410         struct obd_device *obd = exp->exp_obd;
1411         struct lmv_obd *lmv = &obd->u.lmv;
1412         struct lmv_tgt_desc *tgt;
1413         int rc;
1414
1415         ENTRY;
1416
1417         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1418         if (IS_ERR(tgt))
1419                 RETURN(PTR_ERR(tgt));
1420
1421         if (op_data->op_flags & MF_GET_MDT_IDX) {
1422                 op_data->op_mds = tgt->ltd_index;
1423                 RETURN(0);
1424         }
1425
1426         rc = md_getattr(tgt->ltd_exp, op_data, request);
1427
1428         RETURN(rc);
1429 }
1430
1431 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1432 {
1433         struct obd_device *obd = exp->exp_obd;
1434         struct lmv_obd *lmv = &obd->u.lmv;
1435         struct lu_tgt_desc *tgt;
1436
1437         ENTRY;
1438
1439         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1440
1441         /*
1442          * With DNE every object can have two locks in different namespaces:
1443          * lookup lock in space of MDT storing direntry and update/open lock in
1444          * space of MDT storing inode.
1445          */
1446         lmv_foreach_connected_tgt(lmv, tgt)
1447                 md_null_inode(tgt->ltd_exp, fid);
1448
1449         RETURN(0);
1450 }
1451
1452 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1453                      struct md_open_data *mod, struct ptlrpc_request **request)
1454 {
1455         struct obd_device *obd = exp->exp_obd;
1456         struct lmv_obd *lmv = &obd->u.lmv;
1457         struct lmv_tgt_desc *tgt;
1458         int rc;
1459
1460         ENTRY;
1461
1462         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1463         if (IS_ERR(tgt))
1464                 RETURN(PTR_ERR(tgt));
1465
1466         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1467         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1468         RETURN(rc);
1469 }
1470
1471 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 mdt,
1472                                               unsigned short dir_depth)
1473 {
1474         struct lu_tgt_desc *tgt, *cur = NULL;
1475         __u64 total_avail = 0;
1476         __u64 total_weight = 0;
1477         __u64 cur_weight = 0;
1478         int total_usable = 0;
1479         __u64 rand;
1480         int rc;
1481
1482         ENTRY;
1483
1484         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1485                 RETURN(ERR_PTR(-EAGAIN));
1486
1487         down_write(&lmv->lmv_qos.lq_rw_sem);
1488
1489         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1490                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1491
1492         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1493         if (rc)
1494                 GOTO(unlock, tgt = ERR_PTR(rc));
1495
1496         lmv_foreach_tgt(lmv, tgt) {
1497                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1498                         tgt->ltd_qos.ltq_usable = 0;
1499                         continue;
1500                 }
1501
1502                 tgt->ltd_qos.ltq_usable = 1;
1503                 lu_tgt_qos_weight_calc(tgt);
1504                 if (tgt->ltd_index == mdt)
1505                         cur = tgt;
1506                 total_avail += tgt->ltd_qos.ltq_avail;
1507                 total_weight += tgt->ltd_qos.ltq_weight;
1508                 total_usable++;
1509         }
1510
1511         /* if current MDT has above-average space, within range of the QOS
1512          * threshold, stay on the same MDT to avoid creating needless remote
1513          * MDT directories. It's more likely for low level directories
1514          * "16 / (dir_depth + 10)" is the factor to make it more unlikely for
1515          * top level directories, while more likely for low levels.
1516          */
1517         rand = total_avail * 16 / (total_usable * (dir_depth + 10));
1518         if (cur && cur->ltd_qos.ltq_avail >= rand) {
1519                 tgt = cur;
1520                 GOTO(unlock, tgt);
1521         }
1522
1523         rand = lu_prandom_u64_max(total_weight);
1524
1525         lmv_foreach_connected_tgt(lmv, tgt) {
1526                 if (!tgt->ltd_qos.ltq_usable)
1527                         continue;
1528
1529                 cur_weight += tgt->ltd_qos.ltq_weight;
1530                 if (cur_weight < rand)
1531                         continue;
1532
1533                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1534                 GOTO(unlock, tgt);
1535         }
1536
1537         /* no proper target found */
1538         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1539 unlock:
1540         up_write(&lmv->lmv_qos.lq_rw_sem);
1541
1542         return tgt;
1543 }
1544
1545 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv)
1546 {
1547         struct lu_tgt_desc *tgt;
1548         int i;
1549         int index;
1550
1551         ENTRY;
1552
1553         spin_lock(&lmv->lmv_lock);
1554         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1555                 index = (i + lmv->lmv_qos_rr_index) %
1556                         lmv->lmv_mdt_descs.ltd_tgts_size;
1557                 tgt = lmv_tgt(lmv, index);
1558                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1559                         continue;
1560
1561                 lmv->lmv_qos_rr_index = (tgt->ltd_index + 1) %
1562                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1563                 spin_unlock(&lmv->lmv_lock);
1564
1565                 RETURN(tgt);
1566         }
1567         spin_unlock(&lmv->lmv_lock);
1568
1569         RETURN(ERR_PTR(-ENODEV));
1570 }
1571
1572 /* locate MDT which is less full (avoid the most full MDT) */
1573 static struct lu_tgt_desc *lmv_locate_tgt_lf(struct lmv_obd *lmv)
1574 {
1575         struct lu_tgt_desc *min = NULL;
1576         struct lu_tgt_desc *tgt;
1577         __u64 avail = 0;
1578         __u64 rand;
1579
1580         ENTRY;
1581
1582         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1583                 RETURN(ERR_PTR(-EAGAIN));
1584
1585         down_write(&lmv->lmv_qos.lq_rw_sem);
1586
1587         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1588                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1589
1590         lmv_foreach_tgt(lmv, tgt) {
1591                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1592                         tgt->ltd_qos.ltq_usable = 0;
1593                         continue;
1594                 }
1595
1596                 tgt->ltd_qos.ltq_usable = 1;
1597                 lu_tgt_qos_weight_calc(tgt);
1598                 avail += tgt->ltd_qos.ltq_avail;
1599                 if (!min || min->ltd_qos.ltq_avail > tgt->ltd_qos.ltq_avail)
1600                         min = tgt;
1601         }
1602
1603         /* avoid the most full MDT */
1604         if (min)
1605                 avail -= min->ltd_qos.ltq_avail;
1606
1607         rand = lu_prandom_u64_max(avail);
1608         avail = 0;
1609         lmv_foreach_connected_tgt(lmv, tgt) {
1610                 if (!tgt->ltd_qos.ltq_usable)
1611                         continue;
1612
1613                 if (tgt == min)
1614                         continue;
1615
1616                 avail += tgt->ltd_qos.ltq_avail;
1617                 if (avail < rand)
1618                         continue;
1619
1620                 GOTO(unlock, tgt);
1621         }
1622
1623         /* no proper target found */
1624         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1625 unlock:
1626         up_write(&lmv->lmv_qos.lq_rw_sem);
1627
1628         RETURN(tgt);
1629 }
1630
1631 /* locate MDT by file name, for striped directory, the file name hash decides
1632  * which stripe its dirent is stored.
1633  */
1634 static struct lmv_tgt_desc *
1635 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1636                        const char *name, int namelen, struct lu_fid *fid,
1637                        __u32 *mds, bool new_layout)
1638 {
1639         struct lmv_tgt_desc *tgt;
1640         const struct lmv_oinfo *oinfo;
1641
1642         if (!lmv_dir_striped(lsm) || !namelen) {
1643                 tgt = lmv_fid2tgt(lmv, fid);
1644                 if (IS_ERR(tgt))
1645                         return tgt;
1646
1647                 *mds = tgt->ltd_index;
1648                 return tgt;
1649         }
1650
1651         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1652                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1653                         return ERR_PTR(-EBADF);
1654                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1655         } else {
1656                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1657                 if (IS_ERR(oinfo))
1658                         return ERR_CAST(oinfo);
1659         }
1660
1661         /* check stripe FID is sane */
1662         if (!fid_is_sane(&oinfo->lmo_fid))
1663                 return ERR_PTR(-ENODEV);
1664
1665         *fid = oinfo->lmo_fid;
1666         *mds = oinfo->lmo_mds;
1667         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1668
1669         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1670
1671         return tgt ? tgt : ERR_PTR(-ENODEV);
1672 }
1673
1674 /**
1675  * Locate MDT of op_data->op_fid1
1676  *
1677  * For striped directory, it will locate the stripe by name hash, if hash_type
1678  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1679  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1680  * indicates whether old or new layout is used to locate.
1681  *
1682  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1683  *
1684  * \param[in] lmv               LMV device
1685  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1686  *                              op_mds and op_fid1 will be updated if op_mea1
1687  *                              indicates fid1 represents a striped directory.
1688  *
1689  * retval               pointer to the lmv_tgt_desc if succeed.
1690  *                      ERR_PTR(errno) if failed.
1691  */
1692 struct lmv_tgt_desc *
1693 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1694 {
1695         struct lmv_stripe_md *lsm = op_data->op_mea1;
1696         struct lmv_oinfo *oinfo;
1697         struct lmv_tgt_desc *tgt;
1698
1699         if (lmv_dir_foreign(lsm))
1700                 return ERR_PTR(-ENODATA);
1701
1702         /* During creating VOLATILE file, it should honor the mdt
1703          * index if the file under striped dir is being restored, see
1704          * ct_restore(). */
1705         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1706             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1707                 tgt = lmv_tgt(lmv, op_data->op_mds);
1708                 if (!tgt)
1709                         return ERR_PTR(-ENODEV);
1710
1711                 if (lmv_dir_striped(lsm)) {
1712                         int i;
1713
1714                         /* refill the right parent fid */
1715                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1716                                 oinfo = &lsm->lsm_md_oinfo[i];
1717                                 if (oinfo->lmo_mds == op_data->op_mds) {
1718                                         op_data->op_fid1 = oinfo->lmo_fid;
1719                                         break;
1720                                 }
1721                         }
1722
1723                         if (i == lsm->lsm_md_stripe_count)
1724                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1725                 }
1726         } else if (lmv_dir_bad_hash(lsm)) {
1727                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1728                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1729
1730                 op_data->op_fid1 = oinfo->lmo_fid;
1731                 op_data->op_mds = oinfo->lmo_mds;
1732                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1733                 if (!tgt)
1734                         return ERR_PTR(-ENODEV);
1735         } else {
1736                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1737                                 op_data->op_name, op_data->op_namelen,
1738                                 &op_data->op_fid1, &op_data->op_mds,
1739                                 op_data->op_new_layout);
1740         }
1741
1742         return tgt;
1743 }
1744
1745 /* Locate MDT of op_data->op_fid2 for link/rename */
1746 static struct lmv_tgt_desc *
1747 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1748 {
1749         struct lmv_tgt_desc *tgt;
1750         int rc;
1751
1752         LASSERT(op_data->op_name);
1753         if (lmv_dir_layout_changing(op_data->op_mea2)) {
1754                 struct lu_fid fid1 = op_data->op_fid1;
1755                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1756                 struct ptlrpc_request *request = NULL;
1757
1758                 /*
1759                  * avoid creating new file under old layout of migrating
1760                  * directory, check it here.
1761                  */
1762                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1763                                 op_data->op_name, op_data->op_namelen,
1764                                 &op_data->op_fid2, &op_data->op_mds, false);
1765                 if (IS_ERR(tgt))
1766                         RETURN(tgt);
1767
1768                 op_data->op_fid1 = op_data->op_fid2;
1769                 op_data->op_mea1 = op_data->op_mea2;
1770                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1771                 op_data->op_fid1 = fid1;
1772                 op_data->op_mea1 = lsm1;
1773                 if (!rc) {
1774                         ptlrpc_req_finished(request);
1775                         RETURN(ERR_PTR(-EEXIST));
1776                 }
1777
1778                 if (rc != -ENOENT)
1779                         RETURN(ERR_PTR(rc));
1780         }
1781
1782         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1783                                 op_data->op_namelen, &op_data->op_fid2,
1784                                 &op_data->op_mds, true);
1785 }
1786
1787 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1788 {
1789         struct lu_tgt_desc *tgt;
1790         struct ptlrpc_request *request;
1791         int rc;
1792
1793         LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1794         LASSERT(!op_data->op_new_layout);
1795
1796         tgt = lmv_locate_tgt(lmv, op_data);
1797         if (IS_ERR(tgt))
1798                 return PTR_ERR(tgt);
1799
1800         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1801         if (!rc) {
1802                 ptlrpc_req_finished(request);
1803                 return -EEXIST;
1804         }
1805
1806         return rc;
1807 }
1808
1809 /* mkdir by QoS upon 'lfs mkdir -i -1'.
1810  *
1811  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1812  * directories under striped directory.
1813  */
1814 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1815 {
1816         const struct lmv_user_md *lum = op_data->op_data;
1817
1818         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1819                 return false;
1820
1821         if (lmv_dir_striped(op_data->op_mea1))
1822                 return false;
1823
1824         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1825                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1826                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1827 }
1828
1829 /* mkdir by QoS if either ROOT or parent default LMV is space balanced. */
1830 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1831 {
1832         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1833
1834         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1835                 return false;
1836
1837         if (lmv_dir_striped(op_data->op_mea1))
1838                 return false;
1839
1840         return (op_data->op_flags & MF_QOS_MKDIR) ||
1841                (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT);
1842 }
1843
1844 /* if parent default LMV is space balanced, and
1845  * 1. max_inherit_rr is set
1846  * 2. or parent is ROOT
1847  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1848  * LMV requests roundrobin mkdir, do the same.
1849  * NB, this needs to check server is balanced, which is done by caller.
1850  */
1851 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1852 {
1853         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1854
1855         if (!lmv_op_default_qos_mkdir(op_data))
1856                 return false;
1857
1858         return (op_data->op_flags & MF_RR_MKDIR) ||
1859                (lsm && lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE) ||
1860                fid_is_root(&op_data->op_fid1);
1861 }
1862
1863 /* 'lfs mkdir -i <specific_MDT>' */
1864 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1865 {
1866         const struct lmv_user_md *lum = op_data->op_data;
1867
1868         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1869                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1870                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1871                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1872                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1873 }
1874
1875 /* parent default LMV master_mdt_index is not -1. */
1876 static inline bool
1877 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1878 {
1879         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1880                op_data->op_default_mea1 &&
1881                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1882                         LMV_OFFSET_DEFAULT;
1883 }
1884
1885 /* locate MDT by space usage */
1886 static struct lu_tgt_desc *lmv_locate_tgt_by_space(struct lmv_obd *lmv,
1887                                                    struct md_op_data *op_data,
1888                                                    struct lmv_tgt_desc *tgt)
1889 {
1890         struct lmv_tgt_desc *tmp = tgt;
1891
1892         tgt = lmv_locate_tgt_qos(lmv, op_data->op_mds, op_data->op_dir_depth);
1893         if (tgt == ERR_PTR(-EAGAIN)) {
1894                 if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1895                     !lmv_op_default_rr_mkdir(op_data) &&
1896                     !lmv_op_user_qos_mkdir(op_data))
1897                         /* if not necessary, don't create remote directory. */
1898                         tgt = tmp;
1899                 else
1900                         tgt = lmv_locate_tgt_rr(lmv);
1901         }
1902
1903         /*
1904          * only update statfs after QoS mkdir, this means the cached statfs may
1905          * be stale, and current mkdir may not follow QoS accurately, but it's
1906          * not serious, and avoids periodic statfs when client doesn't mkdir by
1907          * QoS.
1908          */
1909         if (!IS_ERR(tgt)) {
1910                 op_data->op_mds = tgt->ltd_index;
1911                 lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1912         }
1913
1914         return tgt;
1915 }
1916
1917 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1918                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1919                 gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
1920                 struct ptlrpc_request **request)
1921 {
1922         struct obd_device *obd = exp->exp_obd;
1923         struct lmv_obd *lmv = &obd->u.lmv;
1924         struct lmv_tgt_desc *tgt;
1925         struct mdt_body *repbody;
1926         int rc;
1927
1928         ENTRY;
1929
1930         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1931                 RETURN(-EIO);
1932
1933         if (lmv_dir_bad_hash(op_data->op_mea1))
1934                 RETURN(-EBADF);
1935
1936         if (lmv_dir_layout_changing(op_data->op_mea1)) {
1937                 /*
1938                  * if parent is migrating, create() needs to lookup existing
1939                  * name in both old and new layout, check old layout on client.
1940                  */
1941                 rc = lmv_old_layout_lookup(lmv, op_data);
1942                 if (rc != -ENOENT)
1943                         RETURN(rc);
1944
1945                 op_data->op_new_layout = true;
1946         }
1947
1948         tgt = lmv_locate_tgt(lmv, op_data);
1949         if (IS_ERR(tgt))
1950                 RETURN(PTR_ERR(tgt));
1951
1952         /* the order to apply policy in mkdir:
1953          * 1. is "lfs mkdir -i N"? mkdir on MDT N.
1954          * 2. is "lfs mkdir -i -1"? mkdir by space usage.
1955          * 3. is starting MDT specified in default LMV? mkdir on MDT N.
1956          * 4. is default LMV space balanced? mkdir by space usage.
1957          */
1958         if (lmv_op_user_specific_mkdir(op_data)) {
1959                 struct lmv_user_md *lum = op_data->op_data;
1960
1961                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1962                 tgt = lmv_tgt(lmv, op_data->op_mds);
1963                 if (!tgt)
1964                         RETURN(-ENODEV);
1965         } else if (lmv_op_user_qos_mkdir(op_data)) {
1966                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
1967                 if (IS_ERR(tgt))
1968                         RETURN(PTR_ERR(tgt));
1969         } else if (lmv_op_default_specific_mkdir(op_data)) {
1970                 op_data->op_mds =
1971                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1972                 tgt = lmv_tgt(lmv, op_data->op_mds);
1973                 if (!tgt)
1974                         RETURN(-ENODEV);
1975         } else if (lmv_op_default_qos_mkdir(op_data)) {
1976                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
1977                 if (IS_ERR(tgt))
1978                         RETURN(PTR_ERR(tgt));
1979         }
1980
1981 retry:
1982         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1983         if (rc)
1984                 RETURN(rc);
1985
1986         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1987                 (int)op_data->op_namelen, op_data->op_name,
1988                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1989                 op_data->op_mds);
1990
1991         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1992         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1993                        cap_effective, rdev, request);
1994         if (rc == 0) {
1995                 if (*request == NULL)
1996                         RETURN(rc);
1997                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1998         }
1999
2000         /* dir restripe needs to send to MDT where dir is located */
2001         if (rc != -EREMOTE ||
2002             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
2003                 RETURN(rc);
2004
2005         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2006         if (repbody == NULL)
2007                 RETURN(-EPROTO);
2008
2009         /* Not cross-ref case, just get out of here. */
2010         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
2011                 RETURN(rc);
2012
2013         op_data->op_fid2 = repbody->mbo_fid1;
2014         ptlrpc_req_finished(*request);
2015         *request = NULL;
2016
2017         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2018         if (IS_ERR(tgt))
2019                 RETURN(PTR_ERR(tgt));
2020
2021         op_data->op_mds = tgt->ltd_index;
2022         goto retry;
2023 }
2024
2025 static int
2026 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
2027             const union ldlm_policy_data *policy, struct md_op_data *op_data,
2028             struct lustre_handle *lockh, __u64 extra_lock_flags)
2029 {
2030         struct obd_device *obd = exp->exp_obd;
2031         struct lmv_obd *lmv = &obd->u.lmv;
2032         struct lmv_tgt_desc *tgt;
2033         int rc;
2034
2035         ENTRY;
2036
2037         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
2038
2039         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2040         if (IS_ERR(tgt))
2041                 RETURN(PTR_ERR(tgt));
2042
2043         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
2044                PFID(&op_data->op_fid1), tgt->ltd_index);
2045
2046         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
2047                         extra_lock_flags);
2048
2049         RETURN(rc);
2050 }
2051
2052 int
2053 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
2054                  struct ptlrpc_request **preq)
2055 {
2056         struct obd_device *obd = exp->exp_obd;
2057         struct lmv_obd *lmv = &obd->u.lmv;
2058         struct lmv_tgt_desc *tgt;
2059         struct mdt_body *body;
2060         int rc;
2061
2062         ENTRY;
2063
2064 retry:
2065         if (op_data->op_namelen == 2 &&
2066             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
2067                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2068         else
2069                 tgt = lmv_locate_tgt(lmv, op_data);
2070         if (IS_ERR(tgt))
2071                 RETURN(PTR_ERR(tgt));
2072
2073         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
2074                 (int)op_data->op_namelen, op_data->op_name,
2075                 PFID(&op_data->op_fid1), tgt->ltd_index);
2076
2077         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2078         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2079                 ptlrpc_req_finished(*preq);
2080                 *preq = NULL;
2081                 goto retry;
2082         }
2083
2084         if (rc)
2085                 RETURN(rc);
2086
2087         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2088         LASSERT(body != NULL);
2089
2090         if (body->mbo_valid & OBD_MD_MDS) {
2091                 op_data->op_fid1 = body->mbo_fid1;
2092                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2093                 op_data->op_namelen = 0;
2094                 op_data->op_name = NULL;
2095
2096                 ptlrpc_req_finished(*preq);
2097                 *preq = NULL;
2098
2099                 goto retry;
2100         }
2101
2102         RETURN(rc);
2103 }
2104
2105 #define md_op_data_fid(op_data, fl)                     \
2106         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2107          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2108          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2109          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2110          NULL)
2111
2112 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2113                             struct md_op_data *op_data, __u32 op_tgt,
2114                             enum ldlm_mode mode, int bits, int flag)
2115 {
2116         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2117         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2118         union ldlm_policy_data policy = { { 0 } };
2119         int rc = 0;
2120         ENTRY;
2121
2122         if (!fid_is_sane(fid))
2123                 RETURN(0);
2124
2125         if (tgt == NULL) {
2126                 tgt = lmv_fid2tgt(lmv, fid);
2127                 if (IS_ERR(tgt))
2128                         RETURN(PTR_ERR(tgt));
2129         }
2130
2131         if (tgt->ltd_index != op_tgt) {
2132                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2133                 policy.l_inodebits.bits = bits;
2134                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2135                                       mode, LCF_ASYNC, NULL);
2136         } else {
2137                 CDEBUG(D_INODE,
2138                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2139                        op_tgt, PFID(fid));
2140                 op_data->op_flags |= flag;
2141                 rc = 0;
2142         }
2143
2144         RETURN(rc);
2145 }
2146
2147 /*
2148  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2149  * op_data->op_fid2
2150  */
2151 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2152                     struct ptlrpc_request **request)
2153 {
2154         struct obd_device       *obd = exp->exp_obd;
2155         struct lmv_obd          *lmv = &obd->u.lmv;
2156         struct lmv_tgt_desc     *tgt;
2157         int                      rc;
2158         ENTRY;
2159
2160         LASSERT(op_data->op_namelen != 0);
2161
2162         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2163                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2164                op_data->op_name, PFID(&op_data->op_fid1));
2165
2166         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2167         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2168         op_data->op_cap = current_cap();
2169
2170         tgt = lmv_locate_tgt2(lmv, op_data);
2171         if (IS_ERR(tgt))
2172                 RETURN(PTR_ERR(tgt));
2173
2174         /*
2175          * Cancel UPDATE lock on child (fid1).
2176          */
2177         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2178         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2179                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2180         if (rc != 0)
2181                 RETURN(rc);
2182
2183         rc = md_link(tgt->ltd_exp, op_data, request);
2184
2185         RETURN(rc);
2186 }
2187
2188 /* migrate the top directory */
2189 static inline bool lmv_op_topdir_migrate(const struct md_op_data *op_data)
2190 {
2191         if (!S_ISDIR(op_data->op_mode))
2192                 return false;
2193
2194         if (lmv_dir_layout_changing(op_data->op_mea1))
2195                 return false;
2196
2197         return true;
2198 }
2199
2200 /* migrate top dir to specific MDTs */
2201 static inline bool lmv_topdir_specific_migrate(const struct md_op_data *op_data)
2202 {
2203         const struct lmv_user_md *lum = op_data->op_data;
2204
2205         if (!lmv_op_topdir_migrate(op_data))
2206                 return false;
2207
2208         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2209 }
2210
2211 /* migrate top dir in QoS mode if user issued "lfs migrate -m -1..." */
2212 static inline bool lmv_topdir_qos_migrate(const struct md_op_data *op_data)
2213 {
2214         const struct lmv_user_md *lum = op_data->op_data;
2215
2216         if (!lmv_op_topdir_migrate(op_data))
2217                 return false;
2218
2219         return le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
2220 }
2221
2222 static inline bool lmv_subdir_specific_migrate(const struct md_op_data *op_data)
2223 {
2224         const struct lmv_user_md *lum = op_data->op_data;
2225
2226         if (!S_ISDIR(op_data->op_mode))
2227                 return false;
2228
2229         if (!lmv_dir_layout_changing(op_data->op_mea1))
2230                 return false;
2231
2232         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2233 }
2234
2235 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2236                         const char *name, size_t namelen,
2237                         struct ptlrpc_request **request)
2238 {
2239         struct obd_device *obd = exp->exp_obd;
2240         struct lmv_obd *lmv = &obd->u.lmv;
2241         struct lmv_stripe_md *lsm = op_data->op_mea1;
2242         struct lmv_tgt_desc *parent_tgt;
2243         struct lmv_tgt_desc *sp_tgt;
2244         struct lmv_tgt_desc *tp_tgt = NULL;
2245         struct lmv_tgt_desc *child_tgt;
2246         struct lmv_tgt_desc *tgt;
2247         struct lu_fid target_fid = { 0 };
2248         int rc;
2249
2250         ENTRY;
2251
2252         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2253
2254         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2255                PFID(&op_data->op_fid1), (int)namelen, name);
2256
2257         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2258         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2259         op_data->op_cap = current_cap();
2260
2261         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2262         if (IS_ERR(parent_tgt))
2263                 RETURN(PTR_ERR(parent_tgt));
2264
2265         if (lmv_dir_striped(lsm)) {
2266                 const struct lmv_oinfo *oinfo;
2267
2268                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2269                 if (IS_ERR(oinfo))
2270                         RETURN(PTR_ERR(oinfo));
2271
2272                 /* save source stripe FID in fid4 temporarily for ELC */
2273                 op_data->op_fid4 = oinfo->lmo_fid;
2274                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2275                 if (!sp_tgt)
2276                         RETURN(-ENODEV);
2277
2278                 /*
2279                  * if parent is being migrated too, fill op_fid2 with target
2280                  * stripe fid, otherwise the target stripe is not created yet.
2281                  */
2282                 if (lmv_dir_layout_changing(lsm)) {
2283                         oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2284                                                         true);
2285                         if (IS_ERR(oinfo))
2286                                 RETURN(PTR_ERR(oinfo));
2287
2288                         op_data->op_fid2 = oinfo->lmo_fid;
2289                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2290                         if (!tp_tgt)
2291                                 RETURN(-ENODEV);
2292
2293                         /* parent unchanged and update namespace only */
2294                         if (lu_fid_eq(&op_data->op_fid4, &op_data->op_fid2) &&
2295                             op_data->op_bias & MDS_MIGRATE_NSONLY)
2296                                 RETURN(-EALREADY);
2297                 }
2298         } else {
2299                 sp_tgt = parent_tgt;
2300         }
2301
2302         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2303         if (IS_ERR(child_tgt))
2304                 RETURN(PTR_ERR(child_tgt));
2305
2306         if (lmv_topdir_specific_migrate(op_data)) {
2307                 struct lmv_user_md *lum = op_data->op_data;
2308
2309                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2310         } else if (lmv_topdir_qos_migrate(op_data)) {
2311                 tgt = lmv_locate_tgt_lf(lmv);
2312                 if (tgt == ERR_PTR(-EAGAIN))
2313                         tgt = lmv_locate_tgt_rr(lmv);
2314                 if (IS_ERR(tgt))
2315                         RETURN(PTR_ERR(tgt));
2316
2317                 op_data->op_mds = tgt->ltd_index;
2318         } else if (lmv_subdir_specific_migrate(op_data)) {
2319                 struct lmv_user_md *lum = op_data->op_data;
2320                 __u32 i;
2321
2322                 LASSERT(tp_tgt);
2323                 if (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) {
2324                         /* adjust MDTs in lum, since subdir is located on where
2325                          * its parent stripe is, not the first specified MDT.
2326                          */
2327                         for (i = 0; i < le32_to_cpu(lum->lum_stripe_count);
2328                              i++) {
2329                                 if (le32_to_cpu(lum->lum_objects[i].lum_mds) ==
2330                                     tp_tgt->ltd_index)
2331                                         break;
2332                         }
2333
2334                         if (i == le32_to_cpu(lum->lum_stripe_count))
2335                                 RETURN(-ENODEV);
2336
2337                         lum->lum_objects[i].lum_mds =
2338                                 lum->lum_objects[0].lum_mds;
2339                         lum->lum_objects[0].lum_mds =
2340                                 cpu_to_le32(tp_tgt->ltd_index);
2341                 }
2342                 /* NB, the above adjusts subdir migration for command like
2343                  * "lfs migrate -m 0,1,2 ...", but for migration like
2344                  * "lfs migrate -m 0 -c 2 ...", the top dir is migrated to MDT0
2345                  * and MDT1, however its subdir may be migrated to MDT1 and MDT2
2346                  */
2347
2348                 lum->lum_stripe_offset = cpu_to_le32(tp_tgt->ltd_index);
2349                 op_data->op_mds = tp_tgt->ltd_index;
2350         } else if (tp_tgt) {
2351                 op_data->op_mds = tp_tgt->ltd_index;
2352         } else {
2353                 op_data->op_mds = sp_tgt->ltd_index;
2354         }
2355
2356         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2357         if (rc)
2358                 RETURN(rc);
2359
2360         /*
2361          * for directory, send migrate request to the MDT where the object will
2362          * be migrated to, because we can't create a striped directory remotely.
2363          *
2364          * otherwise, send to the MDT where source is located because regular
2365          * file may open lease.
2366          *
2367          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2368          * backward compatibility.
2369          */
2370         if (S_ISDIR(op_data->op_mode) &&
2371             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2372                 tgt = lmv_fid2tgt(lmv, &target_fid);
2373                 if (IS_ERR(tgt))
2374                         RETURN(PTR_ERR(tgt));
2375         } else {
2376                 tgt = child_tgt;
2377         }
2378
2379         /* cancel UPDATE lock of parent master object */
2380         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2381                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2382         if (rc)
2383                 RETURN(rc);
2384
2385         /* cancel UPDATE lock of source parent */
2386         if (sp_tgt != parent_tgt) {
2387                 /*
2388                  * migrate RPC packs master object FID, because we can only pack
2389                  * two FIDs in reint RPC, but MDS needs to know both source
2390                  * parent and target parent, and it will obtain them from master
2391                  * FID and LMV, the other FID in RPC is kept for target.
2392                  *
2393                  * since this FID is not passed to MDC, cancel it anyway.
2394                  */
2395                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2396                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2397                 if (rc)
2398                         RETURN(rc);
2399
2400                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2401         }
2402         op_data->op_fid4 = target_fid;
2403
2404         /* cancel UPDATE locks of target parent */
2405         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2406                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2407         if (rc)
2408                 RETURN(rc);
2409
2410         /* cancel LOOKUP lock of source if source is remote object */
2411         if (child_tgt != sp_tgt) {
2412                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2413                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2414                                       MF_MDC_CANCEL_FID3);
2415                 if (rc)
2416                         RETURN(rc);
2417         }
2418
2419         /* cancel ELC locks of source */
2420         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2421                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2422         if (rc)
2423                 RETURN(rc);
2424
2425         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2426
2427         RETURN(rc);
2428 }
2429
2430 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2431                       const char *old, size_t oldlen,
2432                       const char *new, size_t newlen,
2433                       struct ptlrpc_request **request)
2434 {
2435         struct obd_device *obd = exp->exp_obd;
2436         struct lmv_obd *lmv = &obd->u.lmv;
2437         struct lmv_tgt_desc *sp_tgt;
2438         struct lmv_tgt_desc *tp_tgt = NULL;
2439         struct lmv_tgt_desc *src_tgt = NULL;
2440         struct lmv_tgt_desc *tgt;
2441         struct mdt_body *body;
2442         int rc;
2443
2444         ENTRY;
2445
2446         LASSERT(oldlen != 0);
2447
2448         if (op_data->op_cli_flags & CLI_MIGRATE) {
2449                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2450                 RETURN(rc);
2451         }
2452
2453         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2454         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2455         op_data->op_cap = current_cap();
2456
2457         op_data->op_name = new;
2458         op_data->op_namelen = newlen;
2459
2460         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2461         if (IS_ERR(tp_tgt))
2462                 RETURN(PTR_ERR(tp_tgt));
2463
2464         /* Since the target child might be destroyed, and it might become
2465          * orphan, and we can only check orphan on the local MDT right now, so
2466          * we send rename request to the MDT where target child is located. If
2467          * target child does not exist, then it will send the request to the
2468          * target parent */
2469         if (fid_is_sane(&op_data->op_fid4)) {
2470                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2471                 if (IS_ERR(tgt))
2472                         RETURN(PTR_ERR(tgt));
2473         } else {
2474                 tgt = tp_tgt;
2475         }
2476
2477         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2478
2479         /* cancel UPDATE locks of target parent */
2480         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2481                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2482         if (rc != 0)
2483                 RETURN(rc);
2484
2485         if (fid_is_sane(&op_data->op_fid4)) {
2486                 /* cancel LOOKUP lock of target on target parent */
2487                 if (tgt != tp_tgt) {
2488                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2489                                               tgt->ltd_index, LCK_EX,
2490                                               MDS_INODELOCK_LOOKUP,
2491                                               MF_MDC_CANCEL_FID4);
2492                         if (rc != 0)
2493                                 RETURN(rc);
2494                 }
2495         }
2496
2497         if (fid_is_sane(&op_data->op_fid3)) {
2498                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2499                 if (IS_ERR(src_tgt))
2500                         RETURN(PTR_ERR(src_tgt));
2501
2502                 /* cancel ELC locks of source */
2503                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2504                                       LCK_EX, MDS_INODELOCK_ELC,
2505                                       MF_MDC_CANCEL_FID3);
2506                 if (rc != 0)
2507                         RETURN(rc);
2508         }
2509
2510         op_data->op_name = old;
2511         op_data->op_namelen = oldlen;
2512 retry:
2513         sp_tgt = lmv_locate_tgt(lmv, op_data);
2514         if (IS_ERR(sp_tgt))
2515                 RETURN(PTR_ERR(sp_tgt));
2516
2517         /* cancel UPDATE locks of source parent */
2518         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2519                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2520         if (rc != 0)
2521                 RETURN(rc);
2522
2523         if (fid_is_sane(&op_data->op_fid3)) {
2524                 /* cancel LOOKUP lock of source on source parent */
2525                 if (src_tgt != sp_tgt) {
2526                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2527                                               tgt->ltd_index, LCK_EX,
2528                                               MDS_INODELOCK_LOOKUP,
2529                                               MF_MDC_CANCEL_FID3);
2530                         if (rc != 0)
2531                                 RETURN(rc);
2532                 }
2533         }
2534
2535 rename:
2536         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2537                 PFID(&op_data->op_fid1), (int)oldlen, old,
2538                 PFID(&op_data->op_fid2), (int)newlen, new);
2539
2540         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2541                         request);
2542         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2543                 ptlrpc_req_finished(*request);
2544                 *request = NULL;
2545                 goto retry;
2546         }
2547
2548         if (rc && rc != -EXDEV)
2549                 RETURN(rc);
2550
2551         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2552         if (body == NULL)
2553                 RETURN(-EPROTO);
2554
2555         /* Not cross-ref case, just get out of here. */
2556         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2557                 RETURN(rc);
2558
2559         op_data->op_fid4 = body->mbo_fid1;
2560
2561         ptlrpc_req_finished(*request);
2562         *request = NULL;
2563
2564         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2565         if (IS_ERR(tgt))
2566                 RETURN(PTR_ERR(tgt));
2567
2568         if (fid_is_sane(&op_data->op_fid4)) {
2569                 /* cancel LOOKUP lock of target on target parent */
2570                 if (tgt != tp_tgt) {
2571                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2572                                               tgt->ltd_index, LCK_EX,
2573                                               MDS_INODELOCK_LOOKUP,
2574                                               MF_MDC_CANCEL_FID4);
2575                         if (rc != 0)
2576                                 RETURN(rc);
2577                 }
2578         }
2579
2580         goto rename;
2581 }
2582
2583 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2584                        void *ea, size_t ealen, struct ptlrpc_request **request)
2585 {
2586         struct obd_device *obd = exp->exp_obd;
2587         struct lmv_obd *lmv = &obd->u.lmv;
2588         struct lmv_tgt_desc *tgt;
2589         int rc = 0;
2590
2591         ENTRY;
2592
2593         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2594                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2595                op_data->op_xvalid);
2596
2597         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2598         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2599         if (IS_ERR(tgt))
2600                 RETURN(PTR_ERR(tgt));
2601
2602         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2603
2604         RETURN(rc);
2605 }
2606
2607 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2608                      struct ptlrpc_request **request)
2609 {
2610         struct obd_device *obd = exp->exp_obd;
2611         struct lmv_obd *lmv = &obd->u.lmv;
2612         struct lmv_tgt_desc *tgt;
2613         int rc;
2614
2615         ENTRY;
2616
2617         tgt = lmv_fid2tgt(lmv, fid);
2618         if (IS_ERR(tgt))
2619                 RETURN(PTR_ERR(tgt));
2620
2621         rc = md_fsync(tgt->ltd_exp, fid, request);
2622         RETURN(rc);
2623 }
2624
2625 struct stripe_dirent {
2626         struct page             *sd_page;
2627         struct lu_dirpage       *sd_dp;
2628         struct lu_dirent        *sd_ent;
2629         bool                     sd_eof;
2630 };
2631
2632 struct lmv_dir_ctxt {
2633         struct lmv_obd          *ldc_lmv;
2634         struct md_op_data       *ldc_op_data;
2635         struct md_readdir_info  *ldc_mrinfo;
2636         __u64                    ldc_hash;
2637         int                      ldc_count;
2638         struct stripe_dirent     ldc_stripes[0];
2639 };
2640
2641 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2642 {
2643         if (stripe->sd_page) {
2644                 kunmap(stripe->sd_page);
2645                 put_page(stripe->sd_page);
2646                 stripe->sd_page = NULL;
2647                 stripe->sd_ent = NULL;
2648         }
2649 }
2650
2651 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2652 {
2653         int i;
2654
2655         for (i = 0; i < ctxt->ldc_count; i++)
2656                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2657 }
2658
2659 /* if @ent is dummy, or . .., get next */
2660 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2661                                            struct lu_dirent *ent,
2662                                            int stripe_index)
2663 {
2664         for (; ent; ent = lu_dirent_next(ent)) {
2665                 /* Skip dummy entry */
2666                 if (le16_to_cpu(ent->lde_namelen) == 0)
2667                         continue;
2668
2669                 /* skip . and .. for other stripes */
2670                 if (stripe_index &&
2671                     (strncmp(ent->lde_name, ".",
2672                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2673                      strncmp(ent->lde_name, "..",
2674                              le16_to_cpu(ent->lde_namelen)) == 0))
2675                         continue;
2676
2677                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2678                         break;
2679         }
2680
2681         return ent;
2682 }
2683
2684 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2685                                             struct stripe_dirent *stripe,
2686                                             int stripe_index)
2687 {
2688         struct md_op_data *op_data = ctxt->ldc_op_data;
2689         struct lmv_oinfo *oinfo;
2690         struct lu_fid fid = op_data->op_fid1;
2691         struct inode *inode = op_data->op_data;
2692         struct lmv_tgt_desc *tgt;
2693         struct lu_dirent *ent = stripe->sd_ent;
2694         __u64 hash = ctxt->ldc_hash;
2695         int rc = 0;
2696
2697         ENTRY;
2698
2699         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2700         LASSERT(!ent);
2701
2702         do {
2703                 if (stripe->sd_page) {
2704                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2705
2706                         /* @hash should be the last dirent hash */
2707                         LASSERTF(hash <= end,
2708                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2709                                  ctxt, stripe, hash, end);
2710                         /* unload last page */
2711                         stripe_dirent_unload(stripe);
2712                         /* eof */
2713                         if (end == MDS_DIR_END_OFF) {
2714                                 stripe->sd_eof = true;
2715                                 break;
2716                         }
2717                         hash = end;
2718                 }
2719
2720                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2721                 if (!oinfo->lmo_root) {
2722                         rc = -ENOENT;
2723                         break;
2724                 }
2725
2726                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2727                 if (!tgt) {
2728                         rc = -ENODEV;
2729                         break;
2730                 }
2731
2732                 /* op_data is shared by stripes, reset after use */
2733                 op_data->op_fid1 = oinfo->lmo_fid;
2734                 op_data->op_fid2 = oinfo->lmo_fid;
2735                 op_data->op_data = oinfo->lmo_root;
2736
2737                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_mrinfo, hash,
2738                                   &stripe->sd_page);
2739
2740                 op_data->op_fid1 = fid;
2741                 op_data->op_fid2 = fid;
2742                 op_data->op_data = inode;
2743
2744                 if (rc)
2745                         break;
2746
2747                 stripe->sd_dp = page_address(stripe->sd_page);
2748                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2749                                         stripe_index);
2750                 /* in case a page filled with ., .. and dummy, read next */
2751         } while (!ent);
2752
2753         stripe->sd_ent = ent;
2754         if (rc) {
2755                 LASSERT(!ent);
2756                 /* treat error as eof, so dir can be partially accessed */
2757                 stripe->sd_eof = true;
2758                 ctxt->ldc_mrinfo->mr_partial_readdir_rc = rc;
2759                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2760                               "directory is partially accessed!\n",
2761                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2762                               rc);
2763         }
2764
2765         RETURN(ent);
2766 }
2767
2768 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2769 {
2770         struct obd_device *obd = exp->exp_obd;
2771         struct lmv_obd *lmv = &obd->u.lmv;
2772         struct lmv_tgt_desc *tgt;
2773         int rc;
2774
2775         ENTRY;
2776
2777         rc = lmv_check_connect(obd);
2778         if (rc != 0)
2779                 RETURN(rc);
2780
2781         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2782         if (IS_ERR(tgt))
2783                 RETURN(PTR_ERR(tgt));
2784
2785         data->op_flags |= MF_MDC_CANCEL_FID1;
2786         rc = md_file_resync(tgt->ltd_exp, data);
2787         RETURN(rc);
2788 }
2789
2790 /**
2791  * Get dirent with the closest hash for striped directory
2792  *
2793  * This function will search the dir entry, whose hash value is the
2794  * closest(>=) to hash from all of sub-stripes, and it is only being called
2795  * for striped directory.
2796  *
2797  * \param[in] ctxt              dir read context
2798  *
2799  * \retval                      dirent get the entry successfully
2800  *                              NULL does not get the entry, normally it means
2801  *                              it reaches the end of the directory, while read
2802  *                              stripe dirent error is ignored to allow partial
2803  *                              access.
2804  */
2805 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2806 {
2807         struct stripe_dirent *stripe;
2808         struct lu_dirent *ent = NULL;
2809         int i;
2810         int min = -1;
2811
2812         /* TODO: optimize with k-way merge sort */
2813         for (i = 0; i < ctxt->ldc_count; i++) {
2814                 stripe = &ctxt->ldc_stripes[i];
2815                 if (stripe->sd_eof)
2816                         continue;
2817
2818                 if (!stripe->sd_ent) {
2819                         stripe_dirent_load(ctxt, stripe, i);
2820                         if (!stripe->sd_ent) {
2821                                 LASSERT(stripe->sd_eof);
2822                                 continue;
2823                         }
2824                 }
2825
2826                 if (min == -1 ||
2827                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2828                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2829                         min = i;
2830                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2831                             ctxt->ldc_hash)
2832                                 break;
2833                 }
2834         }
2835
2836         if (min != -1) {
2837                 stripe = &ctxt->ldc_stripes[min];
2838                 ent = stripe->sd_ent;
2839                 /* pop found dirent */
2840                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2841                                                    min);
2842         }
2843
2844         return ent;
2845 }
2846
2847 /**
2848  * Build dir entry page for striped directory
2849  *
2850  * This function gets one entry by @offset from a striped directory. It will
2851  * read entries from all of stripes, and choose one closest to the required
2852  * offset(&offset). A few notes
2853  * 1. skip . and .. for non-zero stripes, because there can only have one .
2854  * and .. in a directory.
2855  * 2. op_data will be shared by all of stripes, instead of allocating new
2856  * one, so need to restore before reusing.
2857  *
2858  * \param[in] exp       obd export refer to LMV
2859  * \param[in] op_data   hold those MD parameters of read_entry
2860  * \param[in] mrinfo    ldlm callback being used in enqueue in mdc_read_entry,
2861  *                      and partial readdir result will be stored in it.
2862  * \param[in] offset    starting hash offset
2863  * \param[out] ppage    the page holding the entry. Note: because the entry
2864  *                      will be accessed in upper layer, so we need hold the
2865  *                      page until the usages of entry is finished, see
2866  *                      ll_dir_entry_next.
2867  *
2868  * retval               =0 if get entry successfully
2869  *                      <0 cannot get entry
2870  */
2871 static int lmv_striped_read_page(struct obd_export *exp,
2872                                  struct md_op_data *op_data,
2873                                  struct md_readdir_info *mrinfo, __u64 offset,
2874                                  struct page **ppage)
2875 {
2876         struct page *page = NULL;
2877         struct lu_dirpage *dp;
2878         void *start;
2879         struct lu_dirent *ent;
2880         struct lu_dirent *last_ent;
2881         int stripe_count;
2882         struct lmv_dir_ctxt *ctxt;
2883         struct lu_dirent *next = NULL;
2884         __u16 ent_size;
2885         size_t left_bytes;
2886         int rc = 0;
2887         ENTRY;
2888
2889         /* Allocate a page and read entries from all of stripes and fill
2890          * the page by hash order */
2891         page = alloc_page(GFP_KERNEL);
2892         if (!page)
2893                 RETURN(-ENOMEM);
2894
2895         /* Initialize the entry page */
2896         dp = kmap(page);
2897         memset(dp, 0, sizeof(*dp));
2898         dp->ldp_hash_start = cpu_to_le64(offset);
2899
2900         start = dp + 1;
2901         left_bytes = PAGE_SIZE - sizeof(*dp);
2902         ent = start;
2903         last_ent = ent;
2904
2905         /* initalize dir read context */
2906         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2907         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2908         if (!ctxt)
2909                 GOTO(free_page, rc = -ENOMEM);
2910         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2911         ctxt->ldc_op_data = op_data;
2912         ctxt->ldc_mrinfo = mrinfo;
2913         ctxt->ldc_hash = offset;
2914         ctxt->ldc_count = stripe_count;
2915
2916         while (1) {
2917                 next = lmv_dirent_next(ctxt);
2918
2919                 /* end of directory */
2920                 if (!next) {
2921                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2922                         break;
2923                 }
2924                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2925
2926                 ent_size = le16_to_cpu(next->lde_reclen);
2927
2928                 /* the last entry lde_reclen is 0, but it might not be the last
2929                  * one of this temporay dir page */
2930                 if (!ent_size)
2931                         ent_size = lu_dirent_calc_size(
2932                                         le16_to_cpu(next->lde_namelen),
2933                                         le32_to_cpu(next->lde_attrs));
2934                 /* page full */
2935                 if (ent_size > left_bytes)
2936                         break;
2937
2938                 memcpy(ent, next, ent_size);
2939
2940                 /* Replace . with master FID and Replace .. with the parent FID
2941                  * of master object */
2942                 if (strncmp(ent->lde_name, ".",
2943                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2944                     le16_to_cpu(ent->lde_namelen) == 1)
2945                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2946                 else if (strncmp(ent->lde_name, "..",
2947                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2948                            le16_to_cpu(ent->lde_namelen) == 2)
2949                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2950
2951                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2952                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2953                        le64_to_cpu(ent->lde_hash));
2954
2955                 left_bytes -= ent_size;
2956                 ent->lde_reclen = cpu_to_le16(ent_size);
2957                 last_ent = ent;
2958                 ent = (void *)ent + ent_size;
2959         };
2960
2961         last_ent->lde_reclen = 0;
2962
2963         if (ent == start)
2964                 dp->ldp_flags |= LDF_EMPTY;
2965         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2966                 dp->ldp_flags |= LDF_COLLIDE;
2967         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2968         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2969
2970         put_lmv_dir_ctxt(ctxt);
2971         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2972
2973         *ppage = page;
2974
2975         RETURN(0);
2976
2977 free_page:
2978         kunmap(page);
2979         __free_page(page);
2980
2981         return rc;
2982 }
2983
2984 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2985                          struct md_readdir_info *mrinfo, __u64 offset,
2986                          struct page **ppage)
2987 {
2988         struct obd_device *obd = exp->exp_obd;
2989         struct lmv_obd *lmv = &obd->u.lmv;
2990         struct lmv_tgt_desc *tgt;
2991         int rc;
2992
2993         ENTRY;
2994
2995         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2996                 RETURN(-ENODATA);
2997
2998         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2999                 rc = lmv_striped_read_page(exp, op_data, mrinfo, offset, ppage);
3000                 RETURN(rc);
3001         }
3002
3003         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
3004         if (IS_ERR(tgt))
3005                 RETURN(PTR_ERR(tgt));
3006
3007         rc = md_read_page(tgt->ltd_exp, op_data, mrinfo, offset, ppage);
3008
3009         RETURN(rc);
3010 }
3011
3012 /**
3013  * Unlink a file/directory
3014  *
3015  * Unlink a file or directory under the parent dir. The unlink request
3016  * usually will be sent to the MDT where the child is located, but if
3017  * the client does not have the child FID then request will be sent to the
3018  * MDT where the parent is located.
3019  *
3020  * If the parent is a striped directory then it also needs to locate which
3021  * stripe the name of the child is located, and replace the parent FID
3022  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
3023  * it will walk through all of sub-stripes until the child is being
3024  * unlinked finally.
3025  *
3026  * \param[in] exp       export refer to LMV
3027  * \param[in] op_data   different parameters transferred beween client
3028  *                      MD stacks, name, namelen, FIDs etc.
3029  *                      op_fid1 is the parent FID, op_fid2 is the child
3030  *                      FID.
3031  * \param[out] request  point to the request of unlink.
3032  *
3033  * retval               0 if succeed
3034  *                      negative errno if failed.
3035  */
3036 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
3037                       struct ptlrpc_request **request)
3038 {
3039         struct obd_device *obd = exp->exp_obd;
3040         struct lmv_obd *lmv = &obd->u.lmv;
3041         struct lmv_tgt_desc *tgt;
3042         struct lmv_tgt_desc *parent_tgt;
3043         struct mdt_body *body;
3044         int rc;
3045
3046         ENTRY;
3047
3048         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
3049         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
3050         op_data->op_cap = current_cap();
3051
3052 retry:
3053         parent_tgt = lmv_locate_tgt(lmv, op_data);
3054         if (IS_ERR(parent_tgt))
3055                 RETURN(PTR_ERR(parent_tgt));
3056
3057         if (likely(!fid_is_zero(&op_data->op_fid2))) {
3058                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3059                 if (IS_ERR(tgt))
3060                         RETURN(PTR_ERR(tgt));
3061         } else {
3062                 tgt = parent_tgt;
3063         }
3064
3065         /*
3066          * If child's fid is given, cancel unused locks for it if it is from
3067          * another export than parent.
3068          *
3069          * LOOKUP lock for child (fid3) should also be cancelled on parent
3070          * tgt_tgt in mdc_unlink().
3071          */
3072         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
3073
3074         if (parent_tgt != tgt)
3075                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
3076                                       LCK_EX, MDS_INODELOCK_LOOKUP,
3077                                       MF_MDC_CANCEL_FID3);
3078
3079         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
3080                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
3081         if (rc)
3082                 RETURN(rc);
3083
3084         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
3085                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
3086                tgt->ltd_index);
3087
3088         rc = md_unlink(tgt->ltd_exp, op_data, request);
3089         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
3090                 ptlrpc_req_finished(*request);
3091                 *request = NULL;
3092                 goto retry;
3093         }
3094
3095         if (rc != -EREMOTE)
3096                 RETURN(rc);
3097
3098         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
3099         if (body == NULL)
3100                 RETURN(-EPROTO);
3101
3102         /* Not cross-ref case, just get out of here. */
3103         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
3104                 RETURN(rc);
3105
3106         /* This is a remote object, try remote MDT. */
3107         op_data->op_fid2 = body->mbo_fid1;
3108         ptlrpc_req_finished(*request);
3109         *request = NULL;
3110
3111         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3112         if (IS_ERR(tgt))
3113                 RETURN(PTR_ERR(tgt));
3114
3115         goto retry;
3116 }
3117
3118 static int lmv_precleanup(struct obd_device *obd)
3119 {
3120         ENTRY;
3121         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
3122         RETURN(0);
3123 }
3124
3125 /**
3126  * Get by key a value associated with a LMV device.
3127  *
3128  * Dispatch request to lower-layer devices as needed.
3129  *
3130  * \param[in] env               execution environment for this thread
3131  * \param[in] exp               export for the LMV device
3132  * \param[in] keylen            length of key identifier
3133  * \param[in] key               identifier of key to get value for
3134  * \param[in] vallen            size of \a val
3135  * \param[out] val              pointer to storage location for value
3136  * \param[in] lsm               optional striping metadata of object
3137  *
3138  * \retval 0            on success
3139  * \retval negative     negated errno on failure
3140  */
3141 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
3142                         __u32 keylen, void *key, __u32 *vallen, void *val)
3143 {
3144         struct obd_device *obd;
3145         struct lmv_obd *lmv;
3146         struct lu_tgt_desc *tgt;
3147         int rc = 0;
3148
3149         ENTRY;
3150
3151         obd = class_exp2obd(exp);
3152         if (obd == NULL) {
3153                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3154                        exp->exp_handle.h_cookie);
3155                 RETURN(-EINVAL);
3156         }
3157
3158         lmv = &obd->u.lmv;
3159         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
3160                 LASSERT(*vallen == sizeof(__u32));
3161                 lmv_foreach_connected_tgt(lmv, tgt) {
3162                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
3163                                           vallen, val))
3164                                 RETURN(0);
3165                 }
3166                 RETURN(-EINVAL);
3167         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3168                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3169                    KEY_IS(KEY_CONN_DATA)) {
3170                 /*
3171                  * Forwarding this request to first MDS, it should know LOV
3172                  * desc.
3173                  */
3174                 tgt = lmv_tgt(lmv, 0);
3175                 if (!tgt)
3176                         RETURN(-ENODEV);
3177
3178                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3179                 if (!rc && KEY_IS(KEY_CONN_DATA))
3180                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3181                 RETURN(rc);
3182         } else if (KEY_IS(KEY_TGT_COUNT)) {
3183                 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
3184                 RETURN(0);
3185         }
3186
3187         CDEBUG(D_IOCTL, "Invalid key\n");
3188         RETURN(-EINVAL);
3189 }
3190
3191 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3192                      int *__rcs, struct ptlrpc_request_set *_set)
3193 {
3194         struct obd_device *obd = class_exp2obd(exp);
3195         struct ptlrpc_request_set *set = _set;
3196         struct lmv_obd *lmv = &obd->u.lmv;
3197         int tgt_count = lmv->lmv_mdt_count;
3198         struct lu_tgt_desc *tgt;
3199         struct fid_array *fat, **fas = NULL;
3200         int i, rc, **rcs = NULL;
3201
3202         if (!set) {
3203                 set = ptlrpc_prep_set();
3204                 if (!set)
3205                         RETURN(-ENOMEM);
3206         }
3207
3208         /* split FIDs by targets */
3209         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3210         if (fas == NULL)
3211                 GOTO(out, rc = -ENOMEM);
3212         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3213         if (rcs == NULL)
3214                 GOTO(out_fas, rc = -ENOMEM);
3215
3216         for (i = 0; i < fa->fa_nr; i++) {
3217                 unsigned int idx;
3218
3219                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3220                 if (rc) {
3221                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3222                                PFID(&fa->fa_fids[i]), rc);
3223                         continue;
3224                 }
3225                 LASSERT(idx < tgt_count);
3226                 if (!fas[idx])
3227                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3228                                   fa_fids[fa->fa_nr]));
3229                 if (!fas[idx])
3230                         GOTO(out, rc = -ENOMEM);
3231                 if (!rcs[idx])
3232                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3233                 if (!rcs[idx])
3234                         GOTO(out, rc = -ENOMEM);
3235
3236                 fat = fas[idx];
3237                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3238         }
3239
3240         lmv_foreach_connected_tgt(lmv, tgt) {
3241                 fat = fas[tgt->ltd_index];
3242                 if (!fat || fat->fa_nr == 0)
3243                         continue;
3244                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3245         }
3246
3247         rc = ptlrpc_set_wait(NULL, set);
3248         if (rc == 0) {
3249                 int j = 0;
3250                 for (i = 0; i < tgt_count; i++) {
3251                         fat = fas[i];
3252                         if (!fat || fat->fa_nr == 0)
3253                                 continue;
3254                         /* copy FIDs back */
3255                         memcpy(fa->fa_fids + j, fat->fa_fids,
3256                                fat->fa_nr * sizeof(struct lu_fid));
3257                         /* copy rcs back */
3258                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3259                         j += fat->fa_nr;
3260                 }
3261         }
3262         if (set != _set)
3263                 ptlrpc_set_destroy(set);
3264
3265 out:
3266         for (i = 0; i < tgt_count; i++) {
3267                 if (fas && fas[i])
3268                         OBD_FREE(fas[i], offsetof(struct fid_array,
3269                                                 fa_fids[fa->fa_nr]));
3270                 if (rcs && rcs[i])
3271                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3272         }
3273         if (rcs)
3274                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3275 out_fas:
3276         if (fas)
3277                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3278
3279         RETURN(rc);
3280 }
3281
3282 /**
3283  * Asynchronously set by key a value associated with a LMV device.
3284  *
3285  * Dispatch request to lower-layer devices as needed.
3286  *
3287  * \param[in] env       execution environment for this thread
3288  * \param[in] exp       export for the LMV device
3289  * \param[in] keylen    length of key identifier
3290  * \param[in] key       identifier of key to store value for
3291  * \param[in] vallen    size of value to store
3292  * \param[in] val       pointer to data to be stored
3293  * \param[in] set       optional list of related ptlrpc requests
3294  *
3295  * \retval 0            on success
3296  * \retval negative     negated errno on failure
3297  */
3298 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3299                               __u32 keylen, void *key, __u32 vallen, void *val,
3300                               struct ptlrpc_request_set *set)
3301 {
3302         struct lmv_tgt_desc *tgt;
3303         struct obd_device *obd;
3304         struct lmv_obd *lmv;