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_ioctl_old.h>
  56 #include <lustre_kernelcomm.h>
  57 #include "lmv_internal.h"
  58
  59 static int lmv_check_connect(struct obd_device *obd);
  60 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data);
  61
  62 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
  63                          int activate)
  64 {
  65         if (tgt->ltd_active == activate)
  66                 return;
  67
  68         tgt->ltd_active = activate;
  69         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
  70                 (activate ? 1 : -1);
  71
  72         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
  73 }
  74
  75 /**
  76  * Error codes:
  77  *
  78  *  -EINVAL  : UUID can't be found in the LMV's target list
  79  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  80  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  81  */
  82 static int lmv_set_mdc_active(struct lmv_obd *lmv,
  83                               const struct obd_uuid *uuid,
  84                               int activate)
  85 {
  86         struct lu_tgt_desc *tgt = NULL;
  87         struct obd_device *obd;
  88         int rc = 0;
  89
  90         ENTRY;
  91
  92         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  93                         lmv, uuid->uuid, activate);
  94
  95         spin_lock(&lmv->lmv_lock);
  96         lmv_foreach_connected_tgt(lmv, tgt) {
  97                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
  98                        tgt->ltd_index, tgt->ltd_uuid.uuid,
  99                        tgt->ltd_exp->exp_handle.h_cookie);
 100
 101                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 102                         break;
 103         }
 104
 105         if (!tgt)
 106                 GOTO(out_lmv_lock, rc = -EINVAL);
 107
 108         obd = class_exp2obd(tgt->ltd_exp);
 109         if (obd == NULL)
 110                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 111
 112         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 113                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 114                obd->obd_type->typ_name, tgt->ltd_index);
 115         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 116
 117         if (tgt->ltd_active == activate) {
 118                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 119                        activate ? "" : "in");
 120                 GOTO(out_lmv_lock, rc);
 121         }
 122
 123         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 124                activate ? "" : "in");
 125         lmv_activate_target(lmv, tgt, activate);
 126         EXIT;
 127
 128  out_lmv_lock:
 129         spin_unlock(&lmv->lmv_lock);
 130         return rc;
 131 }
 132
 133 struct lu_tgt_desc *lmv_tgt_retry(struct lmv_obd *lmv, __u32 index)
 134 {
 135         struct obd_device *obd = lmv2obd_dev(lmv);
 136         struct lu_tgt_desc *tgt;
 137         static time64_t next_print;
 138         time64_t retry_limit = 0;
 139         time64_t now;
 140         unsigned int level;
 141         int rc;
 142
 143         might_sleep();
 144 retry:
 145         tgt = lmv_tgt(lmv, index);
 146         if (likely(tgt && tgt->ltd_exp))
 147                 return tgt;
 148
 149         now = ktime_get_seconds();
 150         if (retry_limit == 0) {
 151                 level = now > next_print ? D_WARNING : D_INFO;
 152                 retry_limit = now + RECONNECT_DELAY_MAX;
 153         } else if (now > retry_limit) {
 154                 level = D_ERROR;
 155         } else {
 156                 level = D_INFO;
 157         }
 158         CDEBUG_LIMIT(level, index < lmv->lmv_mdt_count ?
 159                      "%s: MDT index %u/%u not configured\n" :
 160                      "%s: MDT index %u more than MDT count %u\n",
 161                      obd->obd_name, index, lmv->lmv_mdt_count);
 162         if (now > next_print) {
 163                 LCONSOLE_INFO("%s: wait %ds while client connects to new MDT\n",
 164                               obd->obd_name, (int)(retry_limit - now));
 165                 next_print = retry_limit + 600;
 166         }
 167         if (now < retry_limit) {
 168                 rc = schedule_timeout_interruptible(cfs_time_seconds(1));
 169                 if (rc == 0)
 170                         goto retry;
 171         }
 172
 173         return NULL;
 174 }
 175
 176 static struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 177 {
 178         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
 179         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
 180
 181         return tgt ? obd_get_uuid(tgt->ltd_exp) : NULL;
 182 }
 183
 184 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 185                       enum obd_notify_event ev)
 186 {
 187         struct obd_connect_data *conn_data;
 188         struct lmv_obd          *lmv = &obd->u.lmv;
 189         struct obd_uuid         *uuid;
 190         int                      rc = 0;
 191         ENTRY;
 192
 193         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 194                 CERROR("unexpected notification of %s %s!\n",
 195                        watched->obd_type->typ_name,
 196                        watched->obd_name);
 197                 RETURN(-EINVAL);
 198         }
 199
 200         uuid = &watched->u.cli.cl_target_uuid;
 201         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 202                 /*
 203                  * Set MDC as active before notifying the observer, so the
 204                  * observer can use the MDC normally.
 205                  */
 206                 rc = lmv_set_mdc_active(lmv, uuid,
 207                                         ev == OBD_NOTIFY_ACTIVE);
 208                 if (rc) {
 209                         CERROR("%sactivation of %s failed: %d\n",
 210                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 211                                uuid->uuid, rc);
 212                         RETURN(rc);
 213                 }
 214         } else if (ev == OBD_NOTIFY_OCD) {
 215                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 216                 /*
 217                  * XXX: Make sure that ocd_connect_flags from all targets are
 218                  * the same. Otherwise one of MDTs runs wrong version or
 219                  * something like this.  --umka
 220                  */
 221                 obd->obd_self_export->exp_connect_data = *conn_data;
 222         }
 223
 224         /*
 225          * Pass the notification up the chain.
 226          */
 227         if (obd->obd_observer)
 228                 rc = obd_notify(obd->obd_observer, watched, ev);
 229
 230         RETURN(rc);
 231 }
 232
 233 static int lmv_connect(const struct lu_env *env,
 234                        struct obd_export **pexp, struct obd_device *obd,
 235                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 236                        void *localdata)
 237 {
 238         struct lmv_obd *lmv = &obd->u.lmv;
 239         struct lustre_handle conn = { 0 };
 240         struct obd_export *exp;
 241         int rc;
 242         ENTRY;
 243
 244         rc = class_connect(&conn, obd, cluuid);
 245         if (rc) {
 246                 CERROR("class_connection() returned %d\n", rc);
 247                 RETURN(rc);
 248         }
 249
 250         exp = class_conn2export(&conn);
 251
 252         lmv->connected = 0;
 253         lmv->conn_data = *data;
 254         lmv->lmv_cache = localdata;
 255
 256         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
 257                                                     &obd->obd_kset.kobj);
 258         if (!lmv->lmv_tgts_kobj) {
 259                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
 260                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 261         }
 262
 263         rc = lmv_check_connect(obd);
 264         if (rc != 0)
 265                 GOTO(out_sysfs, rc);
 266
 267         *pexp = exp;
 268
 269         RETURN(rc);
 270
 271 out_sysfs:
 272         if (lmv->lmv_tgts_kobj)
 273                 kobject_put(lmv->lmv_tgts_kobj);
 274
 275         class_disconnect(exp);
 276
 277         return rc;
 278 }
 279
 280 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
 281                             __u32 def_easize)
 282 {
 283         struct obd_device *obd = exp->exp_obd;
 284         struct lmv_obd *lmv = &obd->u.lmv;
 285         struct lmv_tgt_desc *tgt;
 286         int change = 0;
 287         int rc = 0;
 288
 289         ENTRY;
 290
 291         if (lmv->max_easize < easize) {
 292                 lmv->max_easize = easize;
 293                 change = 1;
 294         }
 295         if (lmv->max_def_easize < def_easize) {
 296                 lmv->max_def_easize = def_easize;
 297                 change = 1;
 298         }
 299
 300         if (change == 0)
 301                 RETURN(0);
 302
 303         if (lmv->connected == 0)
 304                 RETURN(0);
 305
 306         lmv_foreach_connected_tgt(lmv, tgt) {
 307                 if (!tgt->ltd_active)
 308                         continue;
 309
 310                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 311                 if (rc) {
 312                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 313                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
 314                         break;
 315                 }
 316         }
 317         RETURN(rc);
 318 }
 319
 320 #define MAX_STRING_SIZE 128
 321
 322 static int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 323 {
 324         struct lmv_obd *lmv = &obd->u.lmv;
 325         struct obd_device *mdc_obd;
 326         struct obd_export *mdc_exp;
 327         struct lu_fld_target target;
 328         int  rc;
 329         ENTRY;
 330
 331         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 332                                         &obd->obd_uuid);
 333         if (!mdc_obd) {
 334                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 335                 RETURN(-EINVAL);
 336         }
 337
 338         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
 339                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 340                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
 341
 342         if (!mdc_obd->obd_set_up) {
 343                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 344                 RETURN(-EINVAL);
 345         }
 346
 347         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
 348                          &lmv->conn_data, lmv->lmv_cache);
 349         if (rc) {
 350                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 351                 RETURN(rc);
 352         }
 353
 354         /*
 355          * Init fid sequence client for this mdc and add new fld target.
 356          */
 357         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 358         if (rc)
 359                 RETURN(rc);
 360
 361         target.ft_srv = NULL;
 362         target.ft_exp = mdc_exp;
 363         target.ft_idx = tgt->ltd_index;
 364
 365         fld_client_add_target(&lmv->lmv_fld, &target);
 366
 367         rc = obd_register_observer(mdc_obd, obd);
 368         if (rc) {
 369                 obd_disconnect(mdc_exp);
 370                 CERROR("target %s register_observer error %d\n",
 371                        tgt->ltd_uuid.uuid, rc);
 372                 RETURN(rc);
 373         }
 374
 375         if (obd->obd_observer) {
 376                 /*
 377                  * Tell the observer about the new target.
 378                  */
 379                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 380                                 OBD_NOTIFY_ACTIVE);
 381                 if (rc) {
 382                         obd_disconnect(mdc_exp);
 383                         RETURN(rc);
 384                 }
 385         }
 386
 387         tgt->ltd_active = 1;
 388         tgt->ltd_exp = mdc_exp;
 389         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
 390
 391         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 392
 393         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
 394         if (rc) {
 395                 obd_disconnect(mdc_exp);
 396                 RETURN(rc);
 397         }
 398
 399         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 400                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 401                kref_read(&obd->obd_refcount));
 402
 403         lmv_statfs_check_update(obd, tgt);
 404
 405         if (lmv->lmv_tgts_kobj)
 406                 /* Even if we failed to create the link, that's fine */
 407                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
 408                                        &mdc_obd->obd_kset.kobj,
 409                                        mdc_obd->obd_name);
 410         RETURN(0);
 411 }
 412
 413 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
 414 {
 415         LASSERT(tgt);
 416         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
 417         OBD_FREE_PTR(tgt);
 418 }
 419
 420 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 421                            __u32 index, int gen)
 422 {
 423         struct obd_device *mdc_obd;
 424         struct lmv_obd *lmv = &obd->u.lmv;
 425         struct lmv_tgt_desc *tgt;
 426         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
 427         int rc = 0;
 428
 429         ENTRY;
 430
 431         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 432         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 433                                         &obd->obd_uuid);
 434         if (!mdc_obd) {
 435                 CERROR("%s: Target %s not attached: rc = %d\n",
 436                        obd->obd_name, uuidp->uuid, -EINVAL);
 437                 RETURN(-EINVAL);
 438         }
 439
 440         OBD_ALLOC_PTR(tgt);
 441         if (!tgt)
 442                 RETURN(-ENOMEM);
 443
 444         mutex_init(&tgt->ltd_fid_mutex);
 445         tgt->ltd_index = index;
 446         tgt->ltd_uuid = *uuidp;
 447         tgt->ltd_active = 0;
 448
 449         mutex_lock(&ltd->ltd_mutex);
 450         rc = ltd_add_tgt(ltd, tgt);
 451         mutex_unlock(&ltd->ltd_mutex);
 452
 453         if (rc)
 454                 GOTO(out_tgt, rc);
 455
 456         if (!lmv->connected)
 457                 /* lmv_check_connect() will connect this target. */
 458                 RETURN(0);
 459
 460         rc = lmv_connect_mdc(obd, tgt);
 461         if (!rc) {
 462                 int easize = sizeof(struct lmv_stripe_md) +
 463                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
 464
 465                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 466         }
 467
 468         RETURN(rc);
 469
 470 out_tgt:
 471         OBD_FREE_PTR(tgt);
 472         return rc;
 473 }
 474
 475 static int lmv_check_connect(struct obd_device *obd)
 476 {
 477         struct lmv_obd *lmv = &obd->u.lmv;
 478         struct lmv_tgt_desc *tgt;
 479         int easize;
 480         int rc;
 481
 482         ENTRY;
 483
 484         if (lmv->connected)
 485                 RETURN(0);
 486
 487         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
 488         if (lmv->connected)
 489                 GOTO(unlock, rc = 0);
 490
 491         if (!lmv->lmv_mdt_count) {
 492                 CERROR("%s: no targets configured: rc = -EINVAL\n",
 493                        obd->obd_name);
 494                 GOTO(unlock, rc = -EINVAL);
 495         }
 496
 497         if (!lmv_mdt0_inited(lmv)) {
 498                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
 499                        obd->obd_name);
 500                 GOTO(unlock, rc = -EINVAL);
 501         }
 502
 503         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 504                obd->obd_uuid.uuid, obd->obd_name);
 505
 506         lmv_foreach_tgt(lmv, tgt) {
 507                 rc = lmv_connect_mdc(obd, tgt);
 508                 if (rc)
 509                         GOTO(out_disc, rc);
 510         }
 511
 512         lmv->connected = 1;
 513         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
 514         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 515         EXIT;
 516 unlock:
 517         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
 518
 519         return rc;
 520
 521 out_disc:
 522         lmv_foreach_tgt(lmv, tgt) {
 523                 tgt->ltd_active = 0;
 524                 if (!tgt->ltd_exp)
 525                         continue;
 526
 527                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
 528                 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 529                 obd_disconnect(tgt->ltd_exp);
 530         }
 531
 532         goto unlock;
 533 }
 534
 535 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 536 {
 537         struct lmv_obd *lmv = &obd->u.lmv;
 538         struct obd_device *mdc_obd;
 539         int rc;
 540         ENTRY;
 541
 542         LASSERT(tgt != NULL);
 543         LASSERT(obd != NULL);
 544
 545         mdc_obd = class_exp2obd(tgt->ltd_exp);
 546
 547         if (mdc_obd) {
 548                 mdc_obd->obd_force = obd->obd_force;
 549                 mdc_obd->obd_fail = obd->obd_fail;
 550                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 551
 552                 if (lmv->lmv_tgts_kobj)
 553                         sysfs_remove_link(lmv->lmv_tgts_kobj,
 554                                           mdc_obd->obd_name);
 555         }
 556
 557         rc = lu_qos_del_tgt(&lmv->lmv_qos, tgt);
 558         if (rc)
 559                 CERROR("%s: Can't del target from QoS table: rc = %d\n",
 560                        tgt->ltd_exp->exp_obd->obd_name, rc);
 561
 562         rc = fld_client_del_target(&lmv->lmv_fld, tgt->ltd_index);
 563         if (rc)
 564                 CERROR("%s: Can't del fld targets: rc = %d\n",
 565                        tgt->ltd_exp->exp_obd->obd_name, rc);
 566
 567         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 568         if (rc)
 569                 CERROR("%s: Can't finalize fids factory: rc = %d\n",
 570                        tgt->ltd_exp->exp_obd->obd_name, rc);
 571
 572         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 573                tgt->ltd_exp->exp_obd->obd_name,
 574                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 575
 576         lmv_activate_target(lmv, tgt, 0);
 577         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 578         rc = obd_disconnect(tgt->ltd_exp);
 579         if (rc) {
 580                 CERROR("%s: Target %s disconnect error: rc = %d\n",
 581                        tgt->ltd_exp->exp_obd->obd_name,
 582                        tgt->ltd_uuid.uuid, rc);
 583         }
 584         tgt->ltd_exp = NULL;
 585         RETURN(0);
 586 }
 587
 588 static int lmv_disconnect(struct obd_export *exp)
 589 {
 590         struct obd_device *obd = class_exp2obd(exp);
 591         struct lmv_obd *lmv = &obd->u.lmv;
 592         struct lmv_tgt_desc *tgt;
 593         int rc;
 594
 595         ENTRY;
 596
 597         lmv_foreach_connected_tgt(lmv, tgt)
 598                 lmv_disconnect_mdc(obd, tgt);
 599
 600         if (lmv->lmv_tgts_kobj)
 601                 kobject_put(lmv->lmv_tgts_kobj);
 602
 603         lmv->connected = 0;
 604         rc = class_disconnect(exp);
 605
 606         RETURN(rc);
 607 }
 608
 609 static void lmv_statfs_update(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
 610                               struct obd_statfs *osfs)
 611 {
 612         spin_lock(&lmv->lmv_lock);
 613         tgt->ltd_statfs = *osfs;
 614         tgt->ltd_statfs_age = ktime_get_seconds();
 615         spin_unlock(&lmv->lmv_lock);
 616         set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
 617 }
 618
 619 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 620                         void __user *uarg)
 621 {
 622         struct obd_device *obd = class_exp2obd(exp);
 623         struct lmv_obd *lmv = &obd->u.lmv;
 624         struct getinfo_fid2path *gf;
 625         struct lmv_tgt_desc *tgt;
 626         struct getinfo_fid2path *remote_gf = NULL;
 627         struct lu_fid root_fid;
 628         int remote_gf_size = 0;
 629         int currentisenc = 0;
 630         int globalisenc = 0;
 631         int rc;
 632
 633         gf = karg;
 634         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 635         if (IS_ERR(tgt))
 636                 RETURN(PTR_ERR(tgt));
 637
 638         root_fid = *gf->gf_u.gf_root_fid;
 639         LASSERT(fid_is_sane(&root_fid));
 640
 641 repeat_fid2path:
 642         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 643         if (rc != 0 && rc != -EREMOTE)
 644                 GOTO(out_fid2path, rc);
 645
 646         if (gf->gf_u.gf_path[0] == '/') {
 647                 /* by convention, server side (mdt_path_current()) puts
 648                  * a leading '/' to tell client that we are dealing with
 649                  * an encrypted file
 650                  */
 651                 currentisenc = 1;
 652                 globalisenc = 1;
 653         } else {
 654                 currentisenc = 0;
 655         }
 656
 657         /* If remote_gf != NULL, it means just building the
 658          * path on the remote MDT, copy this path segment to gf.
 659          */
 660         if (remote_gf != NULL) {
 661                 struct getinfo_fid2path *ori_gf;
 662                 int oldisenc = 0;
 663                 char *ptr;
 664                 int len;
 665
 666                 ori_gf = (struct getinfo_fid2path *)karg;
 667                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 668                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 669                         GOTO(out_fid2path, rc = -EOVERFLOW);
 670
 671                 ptr = ori_gf->gf_u.gf_path;
 672                 oldisenc = ptr[0] == '/';
 673
 674                 len = strlen(gf->gf_u.gf_path);
 675                 if (len) {
 676                         /* move the current path to the right to release space
 677                          * for closer-to-root part
 678                          */
 679                         memmove(ptr + len - currentisenc + 1 + globalisenc,
 680                                 ptr + oldisenc,
 681                                 strlen(ori_gf->gf_u.gf_path) - oldisenc + 1);
 682                         if (globalisenc)
 683                                 *(ptr++) = '/';
 684                         memcpy(ptr, gf->gf_u.gf_path + currentisenc,
 685                                len - currentisenc);
 686                         ptr[len - currentisenc] = '/';
 687                 }
 688         }
 689
 690         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 691                tgt->ltd_exp->exp_obd->obd_name,
 692                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 693                gf->gf_linkno);
 694
 695         if (rc == 0)
 696                 GOTO(out_fid2path, rc);
 697
 698         /* sigh, has to go to another MDT to do path building further */
 699         if (remote_gf == NULL) {
 700                 remote_gf_size = sizeof(*remote_gf) + len - sizeof(*gf);
 701                 OBD_ALLOC(remote_gf, remote_gf_size);
 702                 if (remote_gf == NULL)
 703                         GOTO(out_fid2path, rc = -ENOMEM);
 704                 remote_gf->gf_pathlen = len - sizeof(*gf);
 705         }
 706
 707         if (!fid_is_sane(&gf->gf_fid)) {
 708                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 709                        tgt->ltd_exp->exp_obd->obd_name,
 710                        PFID(&gf->gf_fid), -EINVAL);
 711                 GOTO(out_fid2path, rc = -EINVAL);
 712         }
 713
 714         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 715         if (IS_ERR(tgt))
 716                 GOTO(out_fid2path, rc = -EINVAL);
 717
 718         remote_gf->gf_fid = gf->gf_fid;
 719         remote_gf->gf_recno = -1;
 720         remote_gf->gf_linkno = -1;
 721         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 722         *remote_gf->gf_u.gf_root_fid = root_fid;
 723         gf = remote_gf;
 724         goto repeat_fid2path;
 725
 726 out_fid2path:
 727         if (remote_gf != NULL)
 728                 OBD_FREE(remote_gf, remote_gf_size);
 729         RETURN(rc);
 730 }
 731
 732 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 733                              const struct hsm_user_request *hur,
 734                              const struct lmv_tgt_desc *tgt_mds)
 735 {
 736         struct lmv_tgt_desc *curr_tgt;
 737         __u32 i;
 738         int nr = 0;
 739
 740         /* count how many requests must be sent to the given target */
 741         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 742                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
 743                 if (IS_ERR(curr_tgt))
 744                         RETURN(PTR_ERR(curr_tgt));
 745                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 746                         nr++;
 747         }
 748         return nr;
 749 }
 750
 751 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 752                               struct hsm_user_request *hur_in,
 753                               const struct lmv_tgt_desc *tgt_mds,
 754                               struct hsm_user_request *hur_out)
 755 {
 756         __u32 i, nr_out;
 757         struct lmv_tgt_desc *curr_tgt;
 758
 759         /* build the hsm_user_request for the given target */
 760         hur_out->hur_request = hur_in->hur_request;
 761         nr_out = 0;
 762         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 763                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
 764                 if (IS_ERR(curr_tgt))
 765                         RETURN(PTR_ERR(curr_tgt));
 766                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 767                         hur_out->hur_user_item[nr_out] =
 768                                                 hur_in->hur_user_item[i];
 769                         nr_out++;
 770                 }
 771         }
 772         hur_out->hur_request.hr_itemcount = nr_out;
 773         memcpy(hur_data(hur_out), hur_data(hur_in),
 774                hur_in->hur_request.hr_data_len);
 775
 776         RETURN(0);
 777 }
 778
 779 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 780                                  int len, struct lustre_kernelcomm *lk,
 781                                  void __user *uarg)
 782 {
 783         struct lmv_obd *lmv = &obd->u.lmv;
 784         struct lu_tgt_desc *tgt;
 785         int rc;
 786
 787         ENTRY;
 788
 789         /* unregister request (call from llapi_hsm_copytool_fini) */
 790         lmv_foreach_connected_tgt(lmv, tgt)
 791                 /* best effort: try to clean as much as possible
 792                  * (continue on error) */
 793                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 794
 795         /* Whatever the result, remove copytool from kuc groups.
 796          * Unreached coordinators will get EPIPE on next requests
 797          * and will unregister automatically.
 798          */
 799         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 800
 801         RETURN(rc);
 802 }
 803
 804 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 805                                int len, struct lustre_kernelcomm *lk,
 806                                void __user *uarg)
 807 {
 808         struct lmv_obd *lmv = &obd->u.lmv;
 809         struct file *filp;
 810         bool any_set = false;
 811         struct kkuc_ct_data *kcd;
 812         size_t kcd_size;
 813         struct lu_tgt_desc *tgt;
 814         __u32 i;
 815         int err;
 816         int rc = 0;
 817
 818         ENTRY;
 819
 820         filp = fget(lk->lk_wfd);
 821         if (!filp)
 822                 RETURN(-EBADF);
 823
 824         if (lk->lk_flags & LK_FLG_DATANR)
 825                 kcd_size = offsetof(struct kkuc_ct_data,
 826                                     kcd_archives[lk->lk_data_count]);
 827         else
 828                 kcd_size = sizeof(*kcd);
 829
 830         OBD_ALLOC(kcd, kcd_size);
 831         if (kcd == NULL)
 832                 GOTO(err_fput, rc = -ENOMEM);
 833
 834         kcd->kcd_nr_archives = lk->lk_data_count;
 835         if (lk->lk_flags & LK_FLG_DATANR) {
 836                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
 837                 if (lk->lk_data_count > 0)
 838                         memcpy(kcd->kcd_archives, lk->lk_data,
 839                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
 840         } else {
 841                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
 842         }
 843
 844         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 845                                    lk->lk_group, kcd, kcd_size);
 846         OBD_FREE(kcd, kcd_size);
 847         if (rc)
 848                 GOTO(err_fput, rc);
 849
 850         /* All or nothing: try to register to all MDS.
 851          * In case of failure, unregister from previous MDS,
 852          * except if it because of inactive target. */
 853         lmv_foreach_connected_tgt(lmv, tgt) {
 854                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 855                 if (err) {
 856                         if (tgt->ltd_active) {
 857                                 /* permanent error */
 858                                 CERROR("%s: iocontrol MDC %s on MDT"
 859                                        " idx %d cmd %x: err = %d\n",
 860                                        lmv2obd_dev(lmv)->obd_name,
 861                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
 862                                        err);
 863                                 rc = err;
 864                                 lk->lk_flags |= LK_FLG_STOP;
 865                                 i = tgt->ltd_index;
 866                                 /* unregister from previous MDS */
 867                                 lmv_foreach_connected_tgt(lmv, tgt) {
 868                                         if (tgt->ltd_index >= i)
 869                                                 break;
 870
 871                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 872                                                       lk, uarg);
 873                                 }
 874                                 GOTO(err_kkuc_rem, rc);
 875                         }
 876                         /* else: transient error.
 877                          * kuc will register to the missing MDT
 878                          * when it is back */
 879                 } else {
 880                         any_set = true;
 881                 }
 882         }
 883
 884         if (!any_set)
 885                 /* no registration done: return error */
 886                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 887
 888         RETURN(0);
 889
 890 err_kkuc_rem:
 891         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 892
 893 err_fput:
 894         fput(filp);
 895         return rc;
 896 }
 897
 898 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 899                          int len, void *karg, void __user *uarg)
 900 {
 901         struct obd_device *obd = class_exp2obd(exp);
 902         struct lmv_obd *lmv = &obd->u.lmv;
 903         struct lu_tgt_desc *tgt = NULL;
 904         int set = 0;
 905         __u32 count = lmv->lmv_mdt_count;
 906         int rc = 0;
 907
 908         ENTRY;
 909         CDEBUG(D_IOCTL, "%s: cmd=%x len=%u karg=%pK uarg=%pK\n",
 910                exp->exp_obd->obd_name, cmd, len, karg, uarg);
 911         if (count == 0)
 912                 RETURN(-ENOTTY);
 913
 914         /* exit early for unknown ioctl types */
 915         if (unlikely(_IOC_TYPE(cmd) != 'f' && !IOC_OSC_SET_ACTIVE_ALLOW(cmd)))
 916                 RETURN(OBD_IOC_ERROR(obd->obd_name, cmd, "unknown", -ENOTTY));
 917
 918         /* handle commands that don't use @karg first */
 919         switch (cmd) {
 920         case LL_IOC_GET_CONNECT_FLAGS:
 921                 tgt = lmv_tgt(lmv, 0);
 922                 rc = -ENODATA;
 923                 if (tgt && tgt->ltd_exp)
 924                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, NULL, uarg);
 925                 RETURN(rc);
 926         }
 927
 928         if (unlikely(karg == NULL))
 929                 RETURN(OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL", -EINVAL));
 930
 931         switch (cmd) {
 932         case IOC_OBD_STATFS: {
 933                 struct obd_ioctl_data *data = karg;
 934                 struct obd_device *mdc_obd;
 935                 struct obd_statfs stat_buf = {0};
 936                 __u32 index;
 937
 938                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 939
 940                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
 941                         RETURN(-ENODEV);
 942
 943                 tgt = lmv_tgt(lmv, index);
 944                 if (!tgt)
 945                         RETURN(-EAGAIN);
 946
 947                 if (!tgt->ltd_active)
 948                         RETURN(-ENODATA);
 949
 950                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 951                 if (!mdc_obd)
 952                         RETURN(-EINVAL);
 953
 954                 /* copy UUID */
 955                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 956                                  min((int) data->ioc_plen2,
 957                                      (int) sizeof(struct obd_uuid))))
 958                         RETURN(-EFAULT);
 959
 960                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 961                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 962                                 0);
 963                 if (rc)
 964                         RETURN(rc);
 965                 lmv_statfs_update(lmv, tgt, &stat_buf);
 966                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 967                                  min_t(int, data->ioc_plen1, sizeof(stat_buf))))
 968                         RETURN(-EFAULT);
 969                 break;
 970         }
 971         case OBD_IOC_QUOTACTL: {
 972                 struct if_quotactl *qctl = karg;
 973                 struct obd_quotactl *oqctl;
 974                 struct obd_import *imp;
 975
 976                 if (qctl->qc_valid == QC_MDTIDX) {
 977                         tgt = lmv_tgt(lmv, qctl->qc_idx);
 978                 } else if (qctl->qc_valid == QC_UUID) {
 979                         lmv_foreach_tgt(lmv, tgt) {
 980                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 981                                                      &qctl->obd_uuid))
 982                                         continue;
 983
 984                                 if (!tgt->ltd_exp)
 985                                         RETURN(-EINVAL);
 986
 987                                 break;
 988                         }
 989                 } else {
 990                         RETURN(-EINVAL);
 991                 }
 992
 993                 if (!tgt)
 994                         RETURN(-ENODEV);
 995
 996                 if (!tgt->ltd_exp)
 997                         RETURN(-EINVAL);
 998
 999                 imp = class_exp2cliimp(tgt->ltd_exp);
1000                 if (!tgt->ltd_active && imp->imp_state != LUSTRE_IMP_IDLE) {
1001                         qctl->qc_valid = QC_MDTIDX;
1002                         qctl->obd_uuid = tgt->ltd_uuid;
1003                         RETURN(-ENODATA);
1004                 }
1005
1006                 OBD_ALLOC_PTR(oqctl);
1007                 if (!oqctl)
1008                         RETURN(-ENOMEM);
1009
1010                 QCTL_COPY(oqctl, qctl);
1011                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
1012                 if (rc == 0) {
1013                         QCTL_COPY_NO_PNAME(qctl, oqctl);
1014                         qctl->qc_valid = QC_MDTIDX;
1015                         qctl->obd_uuid = tgt->ltd_uuid;
1016                 }
1017                 OBD_FREE_PTR(oqctl);
1018                 break;
1019         }
1020         case LL_IOC_FID2MDTIDX: {
1021                 struct lu_fid *fid = karg;
1022                 int             mdt_index;
1023
1024                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
1025                 if (rc != 0)
1026                         RETURN(rc);
1027
1028                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
1029                  * point to user space memory for FID2MDTIDX. */
1030                 *(__u32 *)uarg = mdt_index;
1031                 break;
1032         }
1033         case OBD_IOC_FID2PATH: {
1034                 rc = lmv_fid2path(exp, len, karg, uarg);
1035                 break;
1036         }
1037         case LL_IOC_HSM_STATE_GET:
1038         case LL_IOC_HSM_STATE_SET:
1039         case LL_IOC_HSM_ACTION: {
1040                 struct md_op_data *op_data = karg;
1041
1042                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1043                 if (IS_ERR(tgt))
1044                         RETURN(PTR_ERR(tgt));
1045
1046                 if (tgt->ltd_exp == NULL)
1047                         RETURN(-EINVAL);
1048
1049                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1050                 break;
1051         }
1052         case LL_IOC_HSM_PROGRESS: {
1053                 const struct hsm_progress_kernel *hpk = karg;
1054
1055                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
1056                 if (IS_ERR(tgt))
1057                         RETURN(PTR_ERR(tgt));
1058                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1059                 break;
1060         }
1061         case LL_IOC_HSM_REQUEST: {
1062                 struct hsm_user_request *hur = karg;
1063                 unsigned int reqcount = hur->hur_request.hr_itemcount;
1064
1065                 if (reqcount == 0)
1066                         RETURN(0);
1067
1068                 /* if the request is about a single fid
1069                  * or if there is a single MDS, no need to split
1070                  * the request. */
1071                 if (reqcount == 1 || count == 1) {
1072                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
1073                         if (IS_ERR(tgt))
1074                                 RETURN(PTR_ERR(tgt));
1075                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1076                 } else {
1077                         /* split fid list to their respective MDS */
1078                         lmv_foreach_connected_tgt(lmv, tgt) {
1079                                 int nr, rc1;
1080                                 size_t reqlen;
1081                                 struct hsm_user_request *req;
1082
1083                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
1084                                 if (nr < 0)
1085                                         RETURN(nr);
1086                                 if (nr == 0) /* nothing for this MDS */
1087                                         continue;
1088
1089                                 /* build a request with fids for this MDS */
1090                                 reqlen = offsetof(typeof(*hur),
1091                                                   hur_user_item[nr])
1092                                                 + hur->hur_request.hr_data_len;
1093                                 OBD_ALLOC_LARGE(req, reqlen);
1094                                 if (req == NULL)
1095                                         RETURN(-ENOMEM);
1096                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
1097                                 if (rc1 < 0)
1098                                         GOTO(hsm_req_err, rc1);
1099                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1100                                                     req, uarg);
1101 hsm_req_err:
1102                                 if (rc1 != 0 && rc == 0)
1103                                         rc = rc1;
1104                                 OBD_FREE_LARGE(req, reqlen);
1105                         }
1106                 }
1107                 break;
1108         }
1109         case LL_IOC_LOV_SWAP_LAYOUTS: {
1110                 struct md_op_data *op_data = karg;
1111                 struct lmv_tgt_desc *tgt1, *tgt2;
1112
1113                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1114                 if (IS_ERR(tgt1))
1115                         RETURN(PTR_ERR(tgt1));
1116
1117                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1118                 if (IS_ERR(tgt2))
1119                         RETURN(PTR_ERR(tgt2));
1120
1121                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1122                         RETURN(-EINVAL);
1123
1124                 /* only files on same MDT can have their layouts swapped */
1125                 if (tgt1->ltd_index != tgt2->ltd_index)
1126                         RETURN(-EPERM);
1127
1128                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1129                 break;
1130         }
1131         case LL_IOC_HSM_CT_START: {
1132                 struct lustre_kernelcomm *lk = karg;
1133
1134                 if (lk->lk_flags & LK_FLG_STOP)
1135                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1136                 else
1137                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1138                 break;
1139         }
1140         default:
1141                 lmv_foreach_connected_tgt(lmv, tgt) {
1142                         struct obd_device *mdc_obd;
1143                         int err;
1144
1145                         /* ll_umount_begin() sets force flag but for lmv, not
1146                          * mdc. Let's pass it through */
1147                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1148                         mdc_obd->obd_force = obd->obd_force;
1149                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1150                         if (err) {
1151                                 if (tgt->ltd_active) {
1152                                         OBD_IOC_ERROR(obd->obd_name, cmd,
1153                                                       tgt->ltd_uuid.uuid, err);
1154                                         if (!rc)
1155                                                 rc = err;
1156                                         if (unlikely(err == -ENOTTY))
1157                                                 break;
1158                                 }
1159                         } else {
1160                                 set = 1;
1161                         }
1162                 }
1163                 if (!set && !rc)
1164                         rc = -EIO;
1165                 break;
1166         }
1167         RETURN(rc);
1168 }
1169
1170 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1171                   struct lu_fid *fid, struct md_op_data *op_data)
1172 {
1173         struct obd_device *obd = class_exp2obd(exp);
1174         struct lmv_obd *lmv = &obd->u.lmv;
1175         struct lmv_tgt_desc *tgt;
1176         int rc;
1177
1178         ENTRY;
1179
1180         LASSERT(op_data);
1181         LASSERT(fid);
1182
1183         tgt = lmv_tgt(lmv, op_data->op_mds);
1184         if (!tgt)
1185                 RETURN(-ENODEV);
1186
1187         if (!tgt->ltd_active || !tgt->ltd_exp)
1188                 RETURN(-ENODEV);
1189
1190         /*
1191          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1192          * on server that seq in new allocated fid is not yet known.
1193          */
1194         mutex_lock(&tgt->ltd_fid_mutex);
1195         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1196         mutex_unlock(&tgt->ltd_fid_mutex);
1197         if (rc > 0) {
1198                 LASSERT(fid_is_sane(fid));
1199                 rc = 0;
1200         }
1201
1202         RETURN(rc);
1203 }
1204
1205 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1206 {
1207         struct lmv_obd *lmv = &obd->u.lmv;
1208         struct lmv_desc *desc;
1209         struct lnet_processid lnet_id;
1210         int i = 0;
1211         int rc;
1212
1213         ENTRY;
1214
1215         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1216                 CERROR("LMV setup requires a descriptor\n");
1217                 RETURN(-EINVAL);
1218         }
1219
1220         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1221         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1222                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1223                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1224                 RETURN(-EINVAL);
1225         }
1226
1227         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1228                      desc->ld_uuid.uuid);
1229         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1230         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1231         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1232                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1233         lmv->max_def_easize = 0;
1234         lmv->max_easize = 0;
1235
1236         spin_lock_init(&lmv->lmv_lock);
1237
1238         /*
1239          * initialize rr_index to lower 32bit of netid, so that client
1240          * can distribute subdirs evenly from the beginning.
1241          */
1242         while (LNetGetId(i++, &lnet_id, true) != -ENOENT) {
1243                 if (!nid_is_lo0(&lnet_id.nid)) {
1244                         lmv->lmv_qos_rr_index = nidhash(&lnet_id.nid);
1245                         break;
1246                 }
1247         }
1248
1249         rc = lmv_tunables_init(obd);
1250         if (rc)
1251                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1252                       obd->obd_name, rc);
1253
1254         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1255                              LUSTRE_CLI_FLD_HASH_DHT);
1256         if (rc)
1257                 CERROR("Can't init FLD, err %d\n", rc);
1258
1259         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1260         if (rc)
1261                 CWARN("%s: error initialize target table: rc = %d\n",
1262                       obd->obd_name, rc);
1263
1264         RETURN(rc);
1265 }
1266
1267 static int lmv_cleanup(struct obd_device *obd)
1268 {
1269         struct lmv_obd *lmv = &obd->u.lmv;
1270         struct lu_tgt_desc *tgt;
1271         struct lu_tgt_desc *tmp;
1272
1273         ENTRY;
1274
1275         fld_client_fini(&lmv->lmv_fld);
1276         fld_client_debugfs_fini(&lmv->lmv_fld);
1277
1278         lprocfs_obd_cleanup(obd);
1279         lprocfs_free_md_stats(obd);
1280
1281         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1282                 lmv_del_target(lmv, tgt);
1283         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1284
1285         RETURN(0);
1286 }
1287
1288 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1289 {
1290         struct lustre_cfg       *lcfg = buf;
1291         struct obd_uuid         obd_uuid;
1292         int                     gen;
1293         __u32                   index;
1294         int                     rc;
1295         ENTRY;
1296
1297         switch (lcfg->lcfg_command) {
1298         case LCFG_ADD_MDC:
1299                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1300                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1301                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1302                         GOTO(out, rc = -EINVAL);
1303
1304                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1305
1306                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1307                         GOTO(out, rc = -EINVAL);
1308                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1309                         GOTO(out, rc = -EINVAL);
1310                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1311                 GOTO(out, rc);
1312         default:
1313                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1314                 GOTO(out, rc = -EINVAL);
1315         }
1316 out:
1317         RETURN(rc);
1318 }
1319
1320 static int lmv_select_statfs_mdt(struct obd_export *exp, struct lmv_obd *lmv,
1321                                  u32 flags)
1322 {
1323         bool large_nid = exp_connect_flags2(exp) & OBD_CONNECT2_LARGE_NID;
1324         int i;
1325
1326         if (flags & OBD_STATFS_FOR_MDT0)
1327                 return 0;
1328
1329         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1330                 return lmv->lmv_statfs_start;
1331
1332         /* choose initial MDT for this client */
1333         for (i = 0;; i++) {
1334                 struct lnet_processid lnet_id;
1335
1336                 if (LNetGetId(i, &lnet_id, large_nid) == -ENOENT)
1337                         break;
1338
1339                 if (!nid_is_lo0(&lnet_id.nid)) {
1340                         /* We dont need a full 64-bit modulus, just enough
1341                          * to distribute the requests across MDTs evenly.
1342                          */
1343                         lmv->lmv_statfs_start = nidhash(&lnet_id.nid) %
1344                                                 lmv->lmv_mdt_count;
1345                         break;
1346                 }
1347         }
1348
1349         return lmv->lmv_statfs_start;
1350 }
1351
1352 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1353                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1354 {
1355         struct obd_device *obd = class_exp2obd(exp);
1356         struct lmv_obd *lmv = &obd->u.lmv;
1357         struct obd_statfs *temp;
1358         struct lu_tgt_desc *tgt;
1359         __u32 i;
1360         __u32 idx;
1361         int rc = 0;
1362         int err = 0;
1363
1364         ENTRY;
1365
1366         OBD_ALLOC(temp, sizeof(*temp));
1367         if (temp == NULL)
1368                 RETURN(-ENOMEM);
1369
1370         /* distribute statfs among MDTs */
1371         idx = lmv_select_statfs_mdt(exp, lmv, flags);
1372
1373         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1374                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1375                 tgt = lmv_tgt(lmv, idx);
1376                 if (!tgt || !tgt->ltd_exp)
1377                         continue;
1378
1379                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1380                                 flags | OBD_STATFS_NESTED);
1381                 if (rc) {
1382                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1383                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1384                         err = rc;
1385                         /* Try another MDT */
1386                         if (flags & OBD_STATFS_SUM)
1387                                 continue;
1388                         GOTO(out_free_temp, rc);
1389                 }
1390
1391                 if (temp->os_state & OS_STATFS_SUM ||
1392                     flags == OBD_STATFS_FOR_MDT0) {
1393                         /* reset to the last aggregated values
1394                          * and don't sum with non-aggrated data */
1395                         /* If the statfs is from mount, it needs to retrieve
1396                          * necessary information from MDT0. i.e. mount does
1397                          * not need the merged osfs from all of MDT. Also
1398                          * clients can be mounted as long as MDT0 is in
1399                          * service */
1400                         *osfs = *temp;
1401                         GOTO(out_free_temp, rc);
1402                 }
1403
1404                 if (i == 0) {
1405                         *osfs = *temp;
1406                 } else {
1407                         osfs->os_bavail += temp->os_bavail;
1408                         osfs->os_blocks += temp->os_blocks;
1409                         osfs->os_ffree += temp->os_ffree;
1410                         osfs->os_files += temp->os_files;
1411                         osfs->os_granted += temp->os_granted;
1412                 }
1413         }
1414         /* There is no stats from some MDTs, data incomplete */
1415         if (err)
1416                 rc = err;
1417 out_free_temp:
1418         OBD_FREE(temp, sizeof(*temp));
1419         RETURN(rc);
1420 }
1421
1422 static int lmv_statfs_cb(void *cookie, int rc)
1423 {
1424         struct obd_info *oinfo = cookie;
1425         struct obd_device *obd = oinfo->oi_obd;
1426         struct lmv_obd *lmv = &obd->u.lmv;
1427         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1428         struct obd_statfs *osfs = oinfo->oi_osfs;
1429
1430         /*
1431          * NB: don't deactivate TGT upon error, because we may not trigger async
1432          * statfs any longer, then there is no chance to activate TGT.
1433          */
1434         if (!rc)
1435                 lmv_statfs_update(lmv, tgt, osfs);
1436
1437         return rc;
1438 }
1439
1440 /* update tgt statfs async if it's ld_qos_maxage old */
1441 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1442 {
1443         struct obd_info oinfo = {
1444                 .oi_obd = obd,
1445                 .oi_tgt = tgt,
1446                 .oi_cb_up = lmv_statfs_cb,
1447         };
1448         int rc;
1449
1450         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1451             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1452                 return 0;
1453
1454         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1455
1456         return rc;
1457 }
1458
1459 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1460                         struct lu_fid *fid)
1461 {
1462         struct obd_device *obd = exp->exp_obd;
1463         struct lmv_obd *lmv = &obd->u.lmv;
1464         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1465         int rc;
1466
1467         ENTRY;
1468
1469         if (!tgt)
1470                 RETURN(-ENODEV);
1471
1472         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1473         RETURN(rc);
1474 }
1475
1476 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1477                         u64 obd_md_valid, const char *name, size_t buf_size,
1478                         struct ptlrpc_request **req)
1479 {
1480         struct obd_device *obd = exp->exp_obd;
1481         struct lmv_obd *lmv = &obd->u.lmv;
1482         struct lmv_tgt_desc *tgt;
1483         int rc;
1484
1485         ENTRY;
1486
1487         tgt = lmv_fid2tgt(lmv, fid);
1488         if (IS_ERR(tgt))
1489                 RETURN(PTR_ERR(tgt));
1490
1491         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1492
1493         RETURN(rc);
1494 }
1495
1496 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1497                         u64 obd_md_valid, const char *name,
1498                         const void *value, size_t value_size,
1499                         unsigned int xattr_flags, u32 suppgid,
1500                         struct ptlrpc_request **req)
1501 {
1502         struct obd_device *obd = exp->exp_obd;
1503         struct lmv_obd *lmv = &obd->u.lmv;
1504         struct lmv_tgt_desc *tgt;
1505         int rc;
1506
1507         ENTRY;
1508
1509         tgt = lmv_fid2tgt(lmv, fid);
1510         if (IS_ERR(tgt))
1511                 RETURN(PTR_ERR(tgt));
1512
1513         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1514                          value, value_size, xattr_flags, suppgid, req);
1515
1516         RETURN(rc);
1517 }
1518
1519 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1520                        struct ptlrpc_request **request)
1521 {
1522         struct obd_device *obd = exp->exp_obd;
1523         struct lmv_obd *lmv = &obd->u.lmv;
1524         struct lmv_tgt_desc *tgt;
1525         int rc;
1526
1527         ENTRY;
1528
1529         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1530         if (IS_ERR(tgt))
1531                 RETURN(PTR_ERR(tgt));
1532
1533         if (op_data->op_flags & MF_GET_MDT_IDX) {
1534                 op_data->op_mds = tgt->ltd_index;
1535                 RETURN(0);
1536         }
1537
1538         rc = md_getattr(tgt->ltd_exp, op_data, request);
1539
1540         RETURN(rc);
1541 }
1542
1543 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1544 {
1545         struct obd_device *obd = exp->exp_obd;
1546         struct lmv_obd *lmv = &obd->u.lmv;
1547         struct lu_tgt_desc *tgt;
1548
1549         ENTRY;
1550
1551         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1552
1553         /*
1554          * With DNE every object can have two locks in different namespaces:
1555          * lookup lock in space of MDT storing direntry and update/open lock in
1556          * space of MDT storing inode.
1557          */
1558         lmv_foreach_connected_tgt(lmv, tgt)
1559                 md_null_inode(tgt->ltd_exp, fid);
1560
1561         RETURN(0);
1562 }
1563
1564 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1565                      struct md_open_data *mod, struct ptlrpc_request **request)
1566 {
1567         struct obd_device *obd = exp->exp_obd;
1568         struct lmv_obd *lmv = &obd->u.lmv;
1569         struct lmv_tgt_desc *tgt;
1570         int rc;
1571
1572         ENTRY;
1573
1574         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1575         if (IS_ERR(tgt))
1576                 RETURN(PTR_ERR(tgt));
1577
1578         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1579         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1580         RETURN(rc);
1581 }
1582
1583 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv,
1584                                               struct md_op_data *op_data)
1585 {
1586         struct lu_tgt_desc *tgt, *cur = NULL;
1587         __u64 total_avail = 0;
1588         __u64 total_weight = 0;
1589         __u64 cur_weight = 0;
1590         int total_usable = 0;
1591         __u64 rand;
1592         int rc;
1593
1594         ENTRY;
1595
1596         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1597                 RETURN(ERR_PTR(-EAGAIN));
1598
1599         down_write(&lmv->lmv_qos.lq_rw_sem);
1600
1601         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1602                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1603
1604         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1605         if (rc)
1606                 GOTO(unlock, tgt = ERR_PTR(rc));
1607
1608         lmv_foreach_tgt(lmv, tgt) {
1609                 if (!tgt->ltd_exp || !tgt->ltd_active ||
1610                     (tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE)) {
1611                         tgt->ltd_qos.ltq_usable = 0;
1612                         continue;
1613                 }
1614                 /* update one hour overdue statfs */
1615                 if (ktime_get_seconds() - tgt->ltd_statfs_age >
1616                     60 * lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1617                         lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1618                 tgt->ltd_qos.ltq_usable = 1;
1619                 lu_tgt_qos_weight_calc(tgt, true);
1620                 if (tgt->ltd_index == op_data->op_mds)
1621                         cur = tgt;
1622                 total_avail += tgt->ltd_qos.ltq_avail;
1623                 total_weight += tgt->ltd_qos.ltq_weight;
1624                 total_usable++;
1625         }
1626
1627         /* If current MDT has above-average space and dir is not already using
1628          * round-robin to spread across more MDTs, stay on the parent MDT
1629          * to avoid creating needless remote MDT directories.  Remote dirs
1630          * close to the root balance space more effectively than bottom dirs,
1631          * so prefer to create remote dirs at top level of directory tree.
1632          * "16 / (dir_depth + 10)" is the factor to make it less likely
1633          * for top-level directories to stay local unless they have more than
1634          * average free space, while deep dirs prefer local until more full.
1635          *    depth=0 -> 160%, depth=3 -> 123%, depth=6 -> 100%,
1636          *    depth=9 -> 84%, depth=12 -> 73%, depth=15 -> 64%
1637          */
1638         if (!lmv_op_default_rr_mkdir(op_data)) {
1639                 rand = total_avail * 16 /
1640                         (total_usable * (op_data->op_dir_depth + 10));
1641                 if (cur && cur->ltd_qos.ltq_avail >= rand) {
1642                         tgt = cur;
1643                         GOTO(unlock, tgt);
1644                 }
1645         }
1646
1647         rand = lu_prandom_u64_max(total_weight);
1648
1649         lmv_foreach_connected_tgt(lmv, tgt) {
1650                 if (!tgt->ltd_qos.ltq_usable)
1651                         continue;
1652
1653                 cur_weight += tgt->ltd_qos.ltq_weight;
1654                 if (cur_weight < rand)
1655                         continue;
1656
1657                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1658                 GOTO(unlock, tgt);
1659         }
1660
1661         /* no proper target found */
1662         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1663 unlock:
1664         up_write(&lmv->lmv_qos.lq_rw_sem);
1665
1666         return tgt;
1667 }
1668
1669 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv)
1670 {
1671         struct lu_tgt_desc *tgt;
1672         int i;
1673         int index;
1674
1675         ENTRY;
1676
1677         spin_lock(&lmv->lmv_lock);
1678         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1679                 index = (i + lmv->lmv_qos_rr_index) %
1680                         lmv->lmv_mdt_descs.ltd_tgts_size;
1681                 tgt = lmv_tgt(lmv, index);
1682                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active ||
1683                     (tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE))
1684                         continue;
1685
1686                 lmv->lmv_qos_rr_index = (tgt->ltd_index + 1) %
1687                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1688                 spin_unlock(&lmv->lmv_lock);
1689
1690                 RETURN(tgt);
1691         }
1692         spin_unlock(&lmv->lmv_lock);
1693
1694         RETURN(ERR_PTR(-ENODEV));
1695 }
1696
1697 /* locate MDT which is less full (avoid the most full MDT) */
1698 static struct lu_tgt_desc *lmv_locate_tgt_lf(struct lmv_obd *lmv)
1699 {
1700         struct lu_tgt_desc *min = NULL;
1701         struct lu_tgt_desc *tgt;
1702         __u64 avail = 0;
1703         __u64 rand;
1704
1705         ENTRY;
1706
1707         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1708                 RETURN(ERR_PTR(-EAGAIN));
1709
1710         down_write(&lmv->lmv_qos.lq_rw_sem);
1711
1712         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1713                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1714
1715         lmv_foreach_tgt(lmv, tgt) {
1716                 if (!tgt->ltd_exp || !tgt->ltd_active ||
1717                     (tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE)) {
1718                         tgt->ltd_qos.ltq_usable = 0;
1719                         continue;
1720                 }
1721
1722                 tgt->ltd_qos.ltq_usable = 1;
1723                 lu_tgt_qos_weight_calc(tgt, true);
1724                 avail += tgt->ltd_qos.ltq_avail;
1725                 if (!min || min->ltd_qos.ltq_avail > tgt->ltd_qos.ltq_avail)
1726                         min = tgt;
1727         }
1728
1729         /* avoid the most full MDT */
1730         if (min)
1731                 avail -= min->ltd_qos.ltq_avail;
1732
1733         rand = lu_prandom_u64_max(avail);
1734         avail = 0;
1735         lmv_foreach_connected_tgt(lmv, tgt) {
1736                 if (!tgt->ltd_qos.ltq_usable)
1737                         continue;
1738
1739                 if (tgt == min)
1740                         continue;
1741
1742                 avail += tgt->ltd_qos.ltq_avail;
1743                 if (avail < rand)
1744                         continue;
1745
1746                 GOTO(unlock, tgt);
1747         }
1748
1749         /* no proper target found */
1750         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1751 unlock:
1752         up_write(&lmv->lmv_qos.lq_rw_sem);
1753
1754         RETURN(tgt);
1755 }
1756
1757 /* locate MDT by file name, for striped directory, the file name hash decides
1758  * which stripe its dirent is stored.
1759  */
1760 static struct lmv_tgt_desc *
1761 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_object *lso,
1762                        const char *name, int namelen, struct lu_fid *fid,
1763                        __u32 *mds, bool new_layout)
1764 {
1765         struct lmv_tgt_desc *tgt;
1766         const struct lmv_oinfo *oinfo;
1767
1768         if (!lmv_dir_striped(lso) || !namelen) {
1769                 tgt = lmv_fid2tgt(lmv, fid);
1770                 if (IS_ERR(tgt))
1771                         return tgt;
1772
1773                 *mds = tgt->ltd_index;
1774                 return tgt;
1775         }
1776
1777         if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1778                 if (cfs_fail_val >= lso->lso_lsm.lsm_md_stripe_count)
1779                         return ERR_PTR(-EBADF);
1780                 oinfo = &lso->lso_lsm.lsm_md_oinfo[cfs_fail_val];
1781         } else {
1782                 oinfo = lsm_name_to_stripe_info(lso, name, namelen, new_layout);
1783                 if (IS_ERR(oinfo))
1784                         return ERR_CAST(oinfo);
1785         }
1786
1787         /* check stripe FID is sane */
1788         if (!fid_is_sane(&oinfo->lmo_fid))
1789                 return ERR_PTR(-ENODEV);
1790
1791         *fid = oinfo->lmo_fid;
1792         *mds = oinfo->lmo_mds;
1793         tgt = lmv_tgt_retry(lmv, oinfo->lmo_mds);
1794
1795         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1796
1797         return tgt ? tgt : ERR_PTR(-ENODEV);
1798 }
1799
1800 /**
1801  * Locate MDT of op_data->op_fid1
1802  *
1803  * For striped directory, it will locate the stripe by name hash, if hash_type
1804  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1805  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1806  * indicates whether old or new layout is used to locate.
1807  *
1808  * For plain directory, it just locate the MDT of op_data->op_fid1.
1809  *
1810  * \param[in] lmv               LMV device
1811  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1812  *                              op_mds and op_fid1 will be updated if op_lso1
1813  *                              indicates fid1 represents a striped directory.
1814  *
1815  * retval               pointer to the lmv_tgt_desc if succeed.
1816  *                      ERR_PTR(errno) if failed.
1817  */
1818 struct lmv_tgt_desc *
1819 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1820 {
1821         struct lmv_stripe_md *lsm;
1822         struct lmv_oinfo *oinfo;
1823         struct lmv_tgt_desc *tgt;
1824
1825         if (lmv_dir_foreign(op_data->op_lso1))
1826                 return ERR_PTR(-ENODATA);
1827
1828         /* During creating VOLATILE file, it should honor the mdt
1829          * index if the file under striped dir is being restored, see
1830          * ct_restore(). */
1831         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1832             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1833                 tgt = lmv_tgt(lmv, op_data->op_mds);
1834                 if (!tgt)
1835                         return ERR_PTR(-ENODEV);
1836
1837                 if (lmv_dir_striped(op_data->op_lso1)) {
1838                         int i;
1839
1840                         /* refill the right parent fid */
1841                         lsm = &op_data->op_lso1->lso_lsm;
1842                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1843                                 oinfo = &lsm->lsm_md_oinfo[i];
1844                                 if (oinfo->lmo_mds == op_data->op_mds) {
1845                                         op_data->op_fid1 = oinfo->lmo_fid;
1846                                         break;
1847                                 }
1848                         }
1849
1850                         if (i == lsm->lsm_md_stripe_count)
1851                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1852                 }
1853         } else if (lmv_dir_bad_hash(op_data->op_lso1)) {
1854                 lsm = &op_data->op_lso1->lso_lsm;
1855
1856                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1857                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1858
1859                 op_data->op_fid1 = oinfo->lmo_fid;
1860                 op_data->op_mds = oinfo->lmo_mds;
1861                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1862                 if (!tgt)
1863                         return ERR_PTR(-ENODEV);
1864         } else {
1865                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_lso1,
1866                                 op_data->op_name, op_data->op_namelen,
1867                                 &op_data->op_fid1, &op_data->op_mds,
1868                                 op_data->op_new_layout);
1869         }
1870
1871         return tgt;
1872 }
1873
1874 /* Locate MDT of op_data->op_fid2 for link/rename */
1875 static struct lmv_tgt_desc *
1876 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1877 {
1878         struct lmv_tgt_desc *tgt;
1879         int rc;
1880
1881         LASSERT(op_data->op_name);
1882         if (lmv_dir_layout_changing(op_data->op_lso2)) {
1883                 struct lu_fid fid1 = op_data->op_fid1;
1884                 struct lmv_stripe_object *lso1 = op_data->op_lso1;
1885                 struct ptlrpc_request *request = NULL;
1886
1887                 /*
1888                  * avoid creating new file under old layout of migrating
1889                  * directory, check it here.
1890                  */
1891                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_lso2,
1892                                 op_data->op_name, op_data->op_namelen,
1893                                 &op_data->op_fid2, &op_data->op_mds, false);
1894                 if (IS_ERR(tgt))
1895                         RETURN(tgt);
1896
1897                 op_data->op_fid1 = op_data->op_fid2;
1898                 op_data->op_lso1 = op_data->op_lso2;
1899                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1900                 op_data->op_fid1 = fid1;
1901                 op_data->op_lso1 = lso1;
1902                 if (!rc) {
1903                         ptlrpc_req_finished(request);
1904                         RETURN(ERR_PTR(-EEXIST));
1905                 }
1906
1907                 if (rc != -ENOENT)
1908                         RETURN(ERR_PTR(rc));
1909         }
1910
1911         return lmv_locate_tgt_by_name(lmv, op_data->op_lso2,
1912                                       op_data->op_name, op_data->op_namelen,
1913                                       &op_data->op_fid2, &op_data->op_mds,
1914                                       true);
1915 }
1916
1917 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1918 {
1919         struct lu_tgt_desc *tgt;
1920         struct ptlrpc_request *request;
1921         int rc;
1922
1923         LASSERT(lmv_dir_layout_changing(op_data->op_lso1));
1924         LASSERT(!op_data->op_new_layout);
1925
1926         tgt = lmv_locate_tgt(lmv, op_data);
1927         if (IS_ERR(tgt))
1928                 return PTR_ERR(tgt);
1929
1930         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1931         if (!rc) {
1932                 ptlrpc_req_finished(request);
1933                 return -EEXIST;
1934         }
1935
1936         return rc;
1937 }
1938
1939 /* mkdir by QoS upon 'lfs mkdir -i -1'.
1940  *
1941  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1942  * directories under striped directory.
1943  */
1944 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1945 {
1946         const struct lmv_user_md *lum = op_data->op_data;
1947
1948         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1949                 return false;
1950
1951         if (lmv_dir_striped(op_data->op_lso1))
1952                 return false;
1953
1954         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1955                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1956                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1957 }
1958
1959 /* mkdir by QoS if either ROOT or parent default LMV is space balanced. */
1960 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1961 {
1962         const struct lmv_stripe_object *lso = op_data->op_default_lso1;
1963
1964         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1965                 return false;
1966
1967         if (lmv_dir_striped(op_data->op_lso1))
1968                 return false;
1969
1970         return (op_data->op_flags & MF_QOS_MKDIR) ||
1971                (lso && lso->lso_lsm.lsm_md_master_mdt_index ==
1972                 LMV_OFFSET_DEFAULT);
1973 }
1974
1975 /* if parent default LMV is space balanced, and
1976  * 1. max_inherit_rr is set
1977  * 2. or parent is ROOT
1978  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1979  * LMV requests roundrobin mkdir, do the same.
1980  * NB, this needs to check server is balanced, which is done by caller.
1981  */
1982 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1983 {
1984         const struct lmv_stripe_object *lso = op_data->op_default_lso1;
1985
1986         return (op_data->op_flags & MF_RR_MKDIR) ||
1987                (lso && lso->lso_lsm.lsm_md_max_inherit_rr !=
1988                 LMV_INHERIT_RR_NONE) || fid_is_root(&op_data->op_fid1);
1989 }
1990
1991 /* 'lfs mkdir -i <specific_MDT>' */
1992 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1993 {
1994         const struct lmv_user_md *lum = op_data->op_data;
1995
1996         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1997                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1998                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1999                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
2000                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2001 }
2002
2003 /* parent default LMV master_mdt_index is not -1. */
2004 static inline bool
2005 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
2006 {
2007         return op_data->op_code == LUSTRE_OPC_MKDIR &&
2008                op_data->op_default_lso1 &&
2009                op_data->op_default_lso1->lso_lsm.lsm_md_master_mdt_index !=
2010                         LMV_OFFSET_DEFAULT;
2011 }
2012
2013 /* locate MDT by space usage */
2014 static struct lu_tgt_desc *lmv_locate_tgt_by_space(struct lmv_obd *lmv,
2015                                                    struct md_op_data *op_data,
2016                                                    struct lmv_tgt_desc *tgt)
2017 {
2018         struct lmv_tgt_desc *tmp = tgt;
2019
2020         tgt = lmv_locate_tgt_qos(lmv, op_data);
2021         if (tgt == ERR_PTR(-EAGAIN)) {
2022                 if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
2023                     !lmv_op_default_rr_mkdir(op_data) &&
2024                     !lmv_op_user_qos_mkdir(op_data) &&
2025                     !(tmp->ltd_statfs.os_state & OS_STATFS_NOCREATE))
2026                         /* if not necessary, don't create remote directory. */
2027                         tgt = tmp;
2028                 else
2029                         tgt = lmv_locate_tgt_rr(lmv);
2030                 if (!IS_ERR(tgt))
2031                         lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
2032         }
2033
2034         if (!IS_ERR(tgt))
2035                 op_data->op_mds = tgt->ltd_index;
2036
2037         /* If space balance was called because the original target was marked
2038          * NOCREATE, periodically check whether the state has changed.
2039          */
2040         if (tmp != tgt && tmp->ltd_statfs.os_state & OS_STATFS_NOCREATE)
2041                 lmv_statfs_check_update(lmv2obd_dev(lmv), tmp);
2042
2043         return tgt;
2044 }
2045
2046 static int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
2047                       const void *data, size_t datalen, umode_t mode, uid_t uid,
2048                       gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
2049                       struct ptlrpc_request **request)
2050 {
2051         struct obd_device *obd = exp->exp_obd;
2052         struct lmv_obd *lmv = &obd->u.lmv;
2053         struct lmv_tgt_desc *tgt;
2054         struct mdt_body *repbody;
2055         int rc;
2056
2057         ENTRY;
2058
2059         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
2060                 RETURN(-EIO);
2061
2062         if (lmv_dir_bad_hash(op_data->op_lso1))
2063                 RETURN(-EBADF);
2064
2065         if (lmv_dir_layout_changing(op_data->op_lso1)) {
2066                 /*
2067                  * if parent is migrating, create() needs to lookup existing
2068                  * name in both old and new layout, check old layout on client.
2069                  */
2070                 rc = lmv_old_layout_lookup(lmv, op_data);
2071                 if (rc != -ENOENT)
2072                         RETURN(rc);
2073
2074                 op_data->op_new_layout = true;
2075         }
2076
2077         tgt = lmv_locate_tgt(lmv, op_data);
2078         if (IS_ERR(tgt))
2079                 RETURN(PTR_ERR(tgt));
2080
2081         /* the order to apply policy in mkdir:
2082          * 1. is "lfs mkdir -i N"? mkdir on MDT N.
2083          * 2. is "lfs mkdir -i -1"? mkdir by space usage.
2084          * 3. is starting MDT specified in default LMV? mkdir on MDT N.
2085          * 4. is default LMV space balanced? mkdir by space usage.
2086          *
2087          * If the existing parent or specific MDT selected is deactivated
2088          * with OS_STATFS_NOCREATE then select a different MDT by QOS.
2089          */
2090         if (lmv_op_user_specific_mkdir(op_data)) {
2091                 struct lmv_user_md *lum = op_data->op_data;
2092
2093                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2094                 tgt = lmv_tgt(lmv, op_data->op_mds);
2095                 if (!tgt)
2096                         RETURN(-ENODEV);
2097                 if (unlikely(tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE))
2098                         GOTO(new_tgt, -EAGAIN);
2099         } else if (lmv_op_user_qos_mkdir(op_data)) {
2100                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
2101                 if (IS_ERR(tgt))
2102                         RETURN(PTR_ERR(tgt));
2103         } else if (lmv_op_default_specific_mkdir(op_data)) {
2104                 struct lmv_stripe_md *lsm = &op_data->op_default_lso1->lso_lsm;
2105
2106                 op_data->op_mds = lsm->lsm_md_master_mdt_index;
2107                 tgt = lmv_tgt(lmv, op_data->op_mds);
2108                 if (!tgt)
2109                         RETURN(-ENODEV);
2110                 if (unlikely(tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE))
2111                         GOTO(new_tgt, -EAGAIN);
2112         } else if (lmv_op_default_qos_mkdir(op_data) ||
2113                    tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE) {
2114 new_tgt:
2115                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
2116                 if (IS_ERR(tgt))
2117                         RETURN(PTR_ERR(tgt));
2118         }
2119
2120 retry:
2121         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
2122         if (rc)
2123                 RETURN(rc);
2124
2125         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
2126                 (int)op_data->op_namelen, op_data->op_name,
2127                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
2128                 op_data->op_mds);
2129
2130         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2131         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
2132                        cap_effective, rdev, request);
2133         if (rc == 0) {
2134                 if (*request == NULL)
2135                         RETURN(rc);
2136                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
2137         }
2138
2139         /* dir restripe needs to send to MDT where dir is located */
2140         if (rc != -EREMOTE ||
2141             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
2142                 RETURN(rc);
2143
2144         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2145         if (repbody == NULL)
2146                 RETURN(-EPROTO);
2147
2148         /* Not cross-ref case, just get out of here. */
2149         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
2150                 RETURN(rc);
2151
2152         op_data->op_fid2 = repbody->mbo_fid1;
2153         ptlrpc_req_finished(*request);
2154         *request = NULL;
2155
2156         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2157         if (IS_ERR(tgt))
2158                 RETURN(PTR_ERR(tgt));
2159
2160         op_data->op_mds = tgt->ltd_index;
2161         goto retry;
2162 }
2163
2164 static int
2165 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
2166             const union ldlm_policy_data *policy, struct md_op_data *op_data,
2167             struct lustre_handle *lockh, __u64 extra_lock_flags)
2168 {
2169         struct obd_device *obd = exp->exp_obd;
2170         struct lmv_obd *lmv = &obd->u.lmv;
2171         struct lmv_tgt_desc *tgt;
2172         int rc;
2173
2174         ENTRY;
2175
2176         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
2177
2178         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2179         if (IS_ERR(tgt))
2180                 RETURN(PTR_ERR(tgt));
2181
2182         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
2183                PFID(&op_data->op_fid1), tgt->ltd_index);
2184
2185         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
2186                         extra_lock_flags);
2187
2188         RETURN(rc);
2189 }
2190
2191 int
2192 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
2193                  struct ptlrpc_request **preq)
2194 {
2195         struct obd_device *obd = exp->exp_obd;
2196         struct lmv_obd *lmv = &obd->u.lmv;
2197         struct lmv_tgt_desc *tgt;
2198         struct mdt_body *body;
2199         int rc;
2200
2201         ENTRY;
2202
2203 retry:
2204         if (op_data->op_namelen == 2 &&
2205             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
2206                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2207         else
2208                 tgt = lmv_locate_tgt(lmv, op_data);
2209         if (IS_ERR(tgt))
2210                 RETURN(PTR_ERR(tgt));
2211
2212         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
2213                 (int)op_data->op_namelen, op_data->op_name,
2214                 PFID(&op_data->op_fid1), tgt->ltd_index);
2215
2216         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2217         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2218                 ptlrpc_req_finished(*preq);
2219                 *preq = NULL;
2220                 goto retry;
2221         }
2222
2223         if (rc)
2224                 RETURN(rc);
2225
2226         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2227         LASSERT(body != NULL);
2228
2229         if (body->mbo_valid & OBD_MD_MDS) {
2230                 op_data->op_fid1 = body->mbo_fid1;
2231                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2232                 op_data->op_namelen = 0;
2233                 op_data->op_name = NULL;
2234
2235                 ptlrpc_req_finished(*preq);
2236                 *preq = NULL;
2237
2238                 goto retry;
2239         }
2240
2241         RETURN(rc);
2242 }
2243
2244 #define md_op_data_fid(op_data, fl)                     \
2245         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2246          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2247          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2248          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2249          NULL)
2250
2251 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2252                             struct md_op_data *op_data, __u32 op_tgt,
2253                             enum ldlm_mode mode, int bits, int flag)
2254 {
2255         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2256         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2257         union ldlm_policy_data policy = { { 0 } };
2258         int rc = 0;
2259         ENTRY;
2260
2261         if (!fid_is_sane(fid))
2262                 RETURN(0);
2263
2264         if (tgt == NULL) {
2265                 tgt = lmv_fid2tgt(lmv, fid);
2266                 if (IS_ERR(tgt))
2267                         RETURN(PTR_ERR(tgt));
2268         }
2269
2270         if (tgt->ltd_index != op_tgt) {
2271                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2272                 policy.l_inodebits.bits = bits;
2273                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2274                                       mode, LCF_ASYNC, NULL);
2275         } else {
2276                 CDEBUG(D_INODE,
2277                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2278                        op_tgt, PFID(fid));
2279                 op_data->op_flags |= flag;
2280                 rc = 0;
2281         }
2282
2283         RETURN(rc);
2284 }
2285
2286 /*
2287  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2288  * op_data->op_fid2
2289  */
2290 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2291                     struct ptlrpc_request **request)
2292 {
2293         struct obd_device       *obd = exp->exp_obd;
2294         struct lmv_obd          *lmv = &obd->u.lmv;
2295         struct lmv_tgt_desc     *tgt;
2296         int                      rc;
2297         ENTRY;
2298
2299         LASSERT(op_data->op_namelen != 0);
2300
2301         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2302                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2303                op_data->op_name, PFID(&op_data->op_fid1));
2304
2305         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2306         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2307         op_data->op_cap = current_cap();
2308
2309         tgt = lmv_locate_tgt2(lmv, op_data);
2310         if (IS_ERR(tgt))
2311                 RETURN(PTR_ERR(tgt));
2312
2313         /*
2314          * Cancel UPDATE lock on child (fid1).
2315          */
2316         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2317         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2318                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2319         if (rc != 0)
2320                 RETURN(rc);
2321
2322         rc = md_link(tgt->ltd_exp, op_data, request);
2323
2324         RETURN(rc);
2325 }
2326
2327 /* migrate the top directory */
2328 static inline bool lmv_op_topdir_migrate(const struct md_op_data *op_data)
2329 {
2330         if (!S_ISDIR(op_data->op_mode))
2331                 return false;
2332
2333         if (lmv_dir_layout_changing(op_data->op_lso1))
2334                 return false;
2335
2336         return true;
2337 }
2338
2339 /* migrate top dir to specific MDTs */
2340 static inline bool lmv_topdir_specific_migrate(const struct md_op_data *op_data)
2341 {
2342         const struct lmv_user_md *lum = op_data->op_data;
2343
2344         if (!lmv_op_topdir_migrate(op_data))
2345                 return false;
2346
2347         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2348 }
2349
2350 /* migrate top dir in QoS mode if user issued "lfs migrate -m -1..." */
2351 static inline bool lmv_topdir_qos_migrate(const struct md_op_data *op_data)
2352 {
2353         const struct lmv_user_md *lum = op_data->op_data;
2354
2355         if (!lmv_op_topdir_migrate(op_data))
2356                 return false;
2357
2358         return le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
2359 }
2360
2361 static inline bool lmv_subdir_specific_migrate(const struct md_op_data *op_data)
2362 {
2363         const struct lmv_user_md *lum = op_data->op_data;
2364
2365         if (!S_ISDIR(op_data->op_mode))
2366                 return false;
2367
2368         if (!lmv_dir_layout_changing(op_data->op_lso1))
2369                 return false;
2370
2371         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2372 }
2373
2374 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2375                         const char *name, size_t namelen,
2376                         struct ptlrpc_request **request)
2377 {
2378         struct obd_device *obd = exp->exp_obd;
2379         struct lmv_obd *lmv = &obd->u.lmv;
2380         struct lmv_stripe_object *lso = op_data->op_lso1;
2381         struct lmv_tgt_desc *parent_tgt;
2382         struct lmv_tgt_desc *sp_tgt;
2383         struct lmv_tgt_desc *tp_tgt = NULL;
2384         struct lmv_tgt_desc *child_tgt;
2385         struct lmv_tgt_desc *tgt;
2386         struct lu_fid target_fid = { 0 };
2387         int rc;
2388
2389         ENTRY;
2390
2391         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2392
2393         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2394                PFID(&op_data->op_fid1), (int)namelen, name);
2395
2396         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2397         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2398         op_data->op_cap = current_cap();
2399
2400         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2401         if (IS_ERR(parent_tgt))
2402                 RETURN(PTR_ERR(parent_tgt));
2403
2404         if (lmv_dir_striped(lso)) {
2405                 const struct lmv_oinfo *oinfo;
2406
2407                 oinfo = lsm_name_to_stripe_info(lso, name, namelen, false);
2408                 if (IS_ERR(oinfo))
2409                         RETURN(PTR_ERR(oinfo));
2410
2411                 /* save source stripe FID in fid4 temporarily for ELC */
2412                 op_data->op_fid4 = oinfo->lmo_fid;
2413                 sp_tgt = lmv_tgt_retry(lmv, oinfo->lmo_mds);
2414                 if (!sp_tgt)
2415                         RETURN(-ENODEV);
2416
2417                 /*
2418                  * if parent is being migrated too, fill op_fid2 with target
2419                  * stripe fid, otherwise the target stripe is not created yet.
2420                  */
2421                 if (lmv_dir_layout_changing(lso)) {
2422                         oinfo = lsm_name_to_stripe_info(lso, name, namelen,
2423                                                         true);
2424                         if (IS_ERR(oinfo))
2425                                 RETURN(PTR_ERR(oinfo));
2426
2427                         op_data->op_fid2 = oinfo->lmo_fid;
2428                         tp_tgt = lmv_tgt_retry(lmv, oinfo->lmo_mds);
2429                         if (!tp_tgt)
2430                                 RETURN(-ENODEV);
2431
2432                         /* parent unchanged and update namespace only */
2433                         if (lu_fid_eq(&op_data->op_fid4, &op_data->op_fid2) &&
2434                             op_data->op_bias & MDS_MIGRATE_NSONLY)
2435                                 RETURN(-EALREADY);
2436                 }
2437         } else {
2438                 sp_tgt = parent_tgt;
2439         }
2440
2441         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2442         if (IS_ERR(child_tgt))
2443                 RETURN(PTR_ERR(child_tgt));
2444
2445         if (lmv_topdir_specific_migrate(op_data)) {
2446                 struct lmv_user_md *lum = op_data->op_data;
2447
2448                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2449         } else if (lmv_topdir_qos_migrate(op_data)) {
2450                 tgt = lmv_locate_tgt_lf(lmv);
2451                 if (tgt == ERR_PTR(-EAGAIN))
2452                         tgt = lmv_locate_tgt_rr(lmv);
2453                 if (IS_ERR(tgt))
2454                         RETURN(PTR_ERR(tgt));
2455
2456                 op_data->op_mds = tgt->ltd_index;
2457         } else if (lmv_subdir_specific_migrate(op_data)) {
2458                 struct lmv_user_md *lum = op_data->op_data;
2459                 __u32 i;
2460
2461                 LASSERT(tp_tgt);
2462                 if (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) {
2463                         /* adjust MDTs in lum, since subdir is located on where
2464                          * its parent stripe is, not the first specified MDT.
2465                          */
2466                         for (i = 0; i < le32_to_cpu(lum->lum_stripe_count);
2467                              i++) {
2468                                 if (le32_to_cpu(lum->lum_objects[i].lum_mds) ==
2469                                     tp_tgt->ltd_index)
2470                                         break;
2471                         }
2472
2473                         if (i == le32_to_cpu(lum->lum_stripe_count))
2474                                 RETURN(-ENODEV);
2475
2476                         lum->lum_objects[i].lum_mds =
2477                                 lum->lum_objects[0].lum_mds;
2478                         lum->lum_objects[0].lum_mds =
2479                                 cpu_to_le32(tp_tgt->ltd_index);
2480                 }
2481                 /* NB, the above adjusts subdir migration for command like
2482                  * "lfs migrate -m 0,1,2 ...", but for migration like
2483                  * "lfs migrate -m 0 -c 2 ...", the top dir is migrated to MDT0
2484                  * and MDT1, however its subdir may be migrated to MDT1 and MDT2
2485                  */
2486
2487                 lum->lum_stripe_offset = cpu_to_le32(tp_tgt->ltd_index);
2488                 op_data->op_mds = tp_tgt->ltd_index;
2489         } else if (tp_tgt) {
2490                 op_data->op_mds = tp_tgt->ltd_index;
2491         } else {
2492                 op_data->op_mds = sp_tgt->ltd_index;
2493         }
2494
2495         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2496         if (rc)
2497                 RETURN(rc);
2498
2499         /*
2500          * for directory, send migrate request to the MDT where the object will
2501          * be migrated to, because we can't create a striped directory remotely.
2502          *
2503          * otherwise, send to the MDT where source is located because regular
2504          * file may open lease.
2505          *
2506          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2507          * backward compatibility.
2508          */
2509         if (S_ISDIR(op_data->op_mode) &&
2510             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2511                 tgt = lmv_fid2tgt(lmv, &target_fid);
2512                 if (IS_ERR(tgt))
2513                         RETURN(PTR_ERR(tgt));
2514         } else {
2515                 tgt = child_tgt;
2516         }
2517
2518         /* cancel UPDATE lock of parent master object */
2519         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2520                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2521         if (rc)
2522                 RETURN(rc);
2523
2524         /* cancel UPDATE lock of source parent */
2525         if (sp_tgt != parent_tgt) {
2526                 /*
2527                  * migrate RPC packs master object FID, because we can only pack
2528                  * two FIDs in reint RPC, but MDS needs to know both source
2529                  * parent and target parent, and it will obtain them from master
2530                  * FID and LMV, the other FID in RPC is kept for target.
2531                  *
2532                  * since this FID is not passed to MDC, cancel it anyway.
2533                  */
2534                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2535                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2536                 if (rc)
2537                         RETURN(rc);
2538
2539                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2540         }
2541         op_data->op_fid4 = target_fid;
2542
2543         /* cancel UPDATE locks of target parent */
2544         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2545                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2546         if (rc)
2547                 RETURN(rc);
2548
2549         /* cancel LOOKUP lock of source if source is remote object */
2550         if (child_tgt != sp_tgt) {
2551                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2552                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2553                                       MF_MDC_CANCEL_FID3);
2554                 if (rc)
2555                         RETURN(rc);
2556         }
2557
2558         /* cancel ELC locks of source */
2559         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2560                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2561         if (rc)
2562                 RETURN(rc);
2563
2564         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2565
2566         RETURN(rc);
2567 }
2568
2569 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2570                       const char *old, size_t oldlen,
2571                       const char *new, size_t newlen,
2572                       struct ptlrpc_request **request)
2573 {
2574         struct obd_device *obd = exp->exp_obd;
2575         struct lmv_obd *lmv = &obd->u.lmv;
2576         struct lmv_tgt_desc *sp_tgt;
2577         struct lmv_tgt_desc *tp_tgt = NULL;
2578         struct lmv_tgt_desc *src_tgt = NULL;
2579         struct lmv_tgt_desc *tgt;
2580         struct mdt_body *body;
2581         int rc;
2582
2583         ENTRY;
2584
2585         LASSERT(oldlen != 0);
2586
2587         if (op_data->op_cli_flags & CLI_MIGRATE) {
2588                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2589                 RETURN(rc);
2590         }
2591
2592         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2593         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2594         op_data->op_cap = current_cap();
2595
2596         op_data->op_name = new;
2597         op_data->op_namelen = newlen;
2598
2599         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2600         if (IS_ERR(tp_tgt))
2601                 RETURN(PTR_ERR(tp_tgt));
2602
2603         /* Since the target child might be destroyed, and it might become
2604          * orphan, and we can only check orphan on the local MDT right now, so
2605          * we send rename request to the MDT where target child is located. If
2606          * target child does not exist, then it will send the request to the
2607          * target parent */
2608         if (fid_is_sane(&op_data->op_fid4)) {
2609                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2610                 if (IS_ERR(tgt))
2611                         RETURN(PTR_ERR(tgt));
2612         } else {
2613                 tgt = tp_tgt;
2614         }
2615
2616         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2617
2618         /* cancel UPDATE locks of target parent */
2619         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2620                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2621         if (rc != 0)
2622                 RETURN(rc);
2623
2624         if (fid_is_sane(&op_data->op_fid4)) {
2625                 /* cancel LOOKUP lock of target on target parent */
2626                 if (tgt != tp_tgt) {
2627                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2628                                               tgt->ltd_index, LCK_EX,
2629                                               MDS_INODELOCK_LOOKUP,
2630                                               MF_MDC_CANCEL_FID4);
2631                         if (rc != 0)
2632                                 RETURN(rc);
2633                 }
2634         }
2635
2636         if (fid_is_sane(&op_data->op_fid3)) {
2637                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2638                 if (IS_ERR(src_tgt))
2639                         RETURN(PTR_ERR(src_tgt));
2640
2641                 /* cancel ELC locks of source */
2642                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2643                                       LCK_EX, MDS_INODELOCK_ELC,
2644                                       MF_MDC_CANCEL_FID3);
2645                 if (rc != 0)
2646                         RETURN(rc);
2647         }
2648
2649         op_data->op_name = old;
2650         op_data->op_namelen = oldlen;
2651 retry:
2652         sp_tgt = lmv_locate_tgt(lmv, op_data);
2653         if (IS_ERR(sp_tgt))
2654                 RETURN(PTR_ERR(sp_tgt));
2655
2656         /* cancel UPDATE locks of source parent */
2657         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2658                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2659         if (rc != 0)
2660                 RETURN(rc);
2661
2662         if (fid_is_sane(&op_data->op_fid3)) {
2663                 /* cancel LOOKUP lock of source on source parent */
2664                 if (src_tgt != sp_tgt) {
2665                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2666                                               tgt->ltd_index, LCK_EX,
2667                                               MDS_INODELOCK_LOOKUP,
2668                                               MF_MDC_CANCEL_FID3);
2669                         if (rc != 0)
2670                                 RETURN(rc);
2671                 }
2672         }
2673
2674 rename:
2675         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2676                 PFID(&op_data->op_fid1), (int)oldlen, old,
2677                 PFID(&op_data->op_fid2), (int)newlen, new);
2678
2679         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2680                         request);
2681         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2682                 ptlrpc_req_finished(*request);
2683                 *request = NULL;
2684                 goto retry;
2685         }
2686
2687         if (rc && rc != -EXDEV)
2688                 RETURN(rc);
2689
2690         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2691         if (body == NULL)
2692                 RETURN(-EPROTO);
2693
2694         /* Not cross-ref case, just get out of here. */
2695         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2696                 RETURN(rc);
2697
2698         op_data->op_fid4 = body->mbo_fid1;
2699
2700         ptlrpc_req_finished(*request);
2701         *request = NULL;
2702
2703         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2704         if (IS_ERR(tgt))
2705                 RETURN(PTR_ERR(tgt));
2706
2707         if (fid_is_sane(&op_data->op_fid4)) {
2708                 /* cancel LOOKUP lock of target on target parent */
2709                 if (tgt != tp_tgt) {
2710                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2711                                               tgt->ltd_index, LCK_EX,
2712                                               MDS_INODELOCK_LOOKUP,
2713                                               MF_MDC_CANCEL_FID4);
2714                         if (rc != 0)
2715                                 RETURN(rc);
2716                 }
2717         }
2718
2719         goto rename;
2720 }
2721
2722 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2723                        void *ea, size_t ealen, struct ptlrpc_request **request)
2724 {
2725         struct obd_device *obd = exp->exp_obd;
2726         struct lmv_obd *lmv = &obd->u.lmv;
2727         struct lmv_tgt_desc *tgt;
2728         int rc = 0;
2729
2730         ENTRY;
2731
2732         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2733                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2734                op_data->op_xvalid);
2735
2736         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2737         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2738         if (IS_ERR(tgt))
2739                 RETURN(PTR_ERR(tgt));
2740
2741         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2742
2743         RETURN(rc);
2744 }
2745
2746 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2747                      struct ptlrpc_request **request)
2748 {
2749         struct obd_device *obd = exp->exp_obd;
2750         struct lmv_obd *lmv = &obd->u.lmv;
2751         struct lmv_tgt_desc *tgt;
2752         int rc;
2753
2754         ENTRY;
2755
2756         tgt = lmv_fid2tgt(lmv, fid);
2757         if (IS_ERR(tgt))
2758                 RETURN(PTR_ERR(tgt));
2759
2760         rc = md_fsync(tgt->ltd_exp, fid, request);
2761         RETURN(rc);
2762 }
2763
2764 struct stripe_dirent {
2765         struct page             *sd_page;
2766         struct lu_dirpage       *sd_dp;
2767         struct lu_dirent        *sd_ent;
2768         bool                     sd_eof;
2769 };
2770
2771 struct lmv_dir_ctxt {
2772         struct lmv_obd          *ldc_lmv;
2773         struct md_op_data       *ldc_op_data;
2774         struct md_readdir_info  *ldc_mrinfo;
2775         __u64                    ldc_hash;
2776         int                      ldc_count;
2777         struct stripe_dirent     ldc_stripes[0];
2778 };
2779
2780 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2781 {
2782         if (stripe->sd_page) {
2783                 kunmap(stripe->sd_page);
2784                 put_page(stripe->sd_page);
2785                 stripe->sd_page = NULL;
2786                 stripe->sd_ent = NULL;
2787         }
2788 }
2789
2790 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2791 {
2792         int i;
2793
2794         for (i = 0; i < ctxt->ldc_count; i++)
2795                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2796 }
2797
2798 /* if @ent is dummy, or . .., get next */
2799 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2800                                            struct lu_dirent *ent,
2801                                            int stripe_index)
2802 {
2803         for (; ent; ent = lu_dirent_next(ent)) {
2804                 /* Skip dummy entry */
2805                 if (le16_to_cpu(ent->lde_namelen) == 0)
2806                         continue;
2807
2808                 /* skip . and .. for other stripes */
2809                 if (stripe_index &&
2810                     (strncmp(ent->lde_name, ".",
2811                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2812                      strncmp(ent->lde_name, "..",
2813                              le16_to_cpu(ent->lde_namelen)) == 0))
2814                         continue;
2815
2816                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2817                         break;
2818         }
2819
2820         return ent;
2821 }
2822
2823 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2824                                             struct stripe_dirent *stripe,
2825                                             int stripe_index)
2826 {
2827         struct md_op_data *op_data = ctxt->ldc_op_data;
2828         struct lmv_oinfo *oinfo;
2829         struct lu_fid fid = op_data->op_fid1;
2830         struct inode *inode = op_data->op_data;
2831         struct lmv_tgt_desc *tgt;
2832         struct lu_dirent *ent = stripe->sd_ent;
2833         __u64 hash = ctxt->ldc_hash;
2834         int rc = 0;
2835
2836         ENTRY;
2837
2838         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2839         LASSERT(!ent);
2840
2841         do {
2842                 if (stripe->sd_page) {
2843                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2844
2845                         /* @hash should be the last dirent hash */
2846                         LASSERTF(hash <= end,
2847                                  "ctxt@%px stripe@%px hash %llx end %llx\n",
2848                                  ctxt, stripe, hash, end);
2849                         /* unload last page */
2850                         stripe_dirent_unload(stripe);
2851                         /* eof */
2852                         if (end == MDS_DIR_END_OFF) {
2853                                 stripe->sd_eof = true;
2854                                 break;
2855                         }
2856                         hash = end;
2857                 }
2858
2859                 oinfo = &op_data->op_lso1->lso_lsm.lsm_md_oinfo[stripe_index];
2860                 if (!oinfo->lmo_root) {
2861                         rc = -ENOENT;
2862                         break;
2863                 }
2864
2865                 tgt = lmv_tgt_retry(ctxt->ldc_lmv, oinfo->lmo_mds);
2866                 if (!tgt) {
2867                         rc = -ENODEV;
2868                         break;
2869                 }
2870
2871                 /* op_data is shared by stripes, reset after use */
2872                 op_data->op_fid1 = oinfo->lmo_fid;
2873                 op_data->op_fid2 = oinfo->lmo_fid;
2874                 op_data->op_data = oinfo->lmo_root;
2875
2876                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_mrinfo, hash,
2877                                   &stripe->sd_page);
2878
2879                 op_data->op_fid1 = fid;
2880                 op_data->op_fid2 = fid;
2881                 op_data->op_data = inode;
2882
2883                 if (rc)
2884                         break;
2885
2886                 stripe->sd_dp = page_address(stripe->sd_page);
2887                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2888                                         stripe_index);
2889                 /* in case a page filled with ., .. and dummy, read next */
2890         } while (!ent);
2891
2892         stripe->sd_ent = ent;
2893         if (rc) {
2894                 LASSERT(!ent);
2895                 /* treat error as eof, so dir can be partially accessed */
2896                 stripe->sd_eof = true;
2897                 ctxt->ldc_mrinfo->mr_partial_readdir_rc = rc;
2898                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2899                               "directory is partially accessed!\n",
2900                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2901                               rc);
2902         }
2903
2904         RETURN(ent);
2905 }
2906
2907 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2908 {
2909         struct obd_device *obd = exp->exp_obd;
2910         struct lmv_obd *lmv = &obd->u.lmv;
2911         struct lmv_tgt_desc *tgt;
2912         int rc;
2913
2914         ENTRY;
2915
2916         rc = lmv_check_connect(obd);
2917         if (rc != 0)
2918                 RETURN(rc);
2919
2920         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2921         if (IS_ERR(tgt))
2922                 RETURN(PTR_ERR(tgt));
2923
2924         data->op_flags |= MF_MDC_CANCEL_FID1;
2925         rc = md_file_resync(tgt->ltd_exp, data);
2926         RETURN(rc);
2927 }
2928
2929 /**
2930  * Get dirent with the closest hash for striped directory
2931  *
2932  * This function will search the dir entry, whose hash value is the
2933  * closest(>=) to hash from all of sub-stripes, and it is only being called
2934  * for striped directory.
2935  *
2936  * \param[in] ctxt              dir read context
2937  *
2938  * \retval                      dirent get the entry successfully
2939  *                              NULL does not get the entry, normally it means
2940  *                              it reaches the end of the directory, while read
2941  *                              stripe dirent error is ignored to allow partial
2942  *                              access.
2943  */
2944 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2945 {
2946         struct stripe_dirent *stripe;
2947         struct lu_dirent *ent = NULL;
2948         int i;
2949         int min = -1;
2950
2951         /* TODO: optimize with k-way merge sort */
2952         for (i = 0; i < ctxt->ldc_count; i++) {
2953                 stripe = &ctxt->ldc_stripes[i];
2954                 if (stripe->sd_eof)
2955                         continue;
2956
2957                 if (!stripe->sd_ent) {
2958                         stripe_dirent_load(ctxt, stripe, i);
2959                         if (!stripe->sd_ent) {
2960                                 LASSERT(stripe->sd_eof);
2961                                 continue;
2962                         }
2963                 }
2964
2965                 if (min == -1 ||
2966                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2967                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2968                         min = i;
2969                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2970                             ctxt->ldc_hash)
2971                                 break;
2972                 }
2973         }
2974
2975         if (min != -1) {
2976                 stripe = &ctxt->ldc_stripes[min];
2977                 ent = stripe->sd_ent;
2978                 /* pop found dirent */
2979                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2980                                                    min);
2981         }
2982
2983         return ent;
2984 }
2985
2986 /**
2987  * Build dir entry page for striped directory
2988  *
2989  * This function gets one entry by @offset from a striped directory. It will
2990  * read entries from all of stripes, and choose one closest to the required
2991  * offset(&offset). A few notes
2992  * 1. skip . and .. for non-zero stripes, because there can only have one .
2993  * and .. in a directory.
2994  * 2. op_data will be shared by all of stripes, instead of allocating new
2995  * one, so need to restore before reusing.
2996  *
2997  * \param[in] exp       obd export refer to LMV
2998  * \param[in] op_data   hold those MD parameters of read_entry
2999  * \param[in] mrinfo    ldlm callback being used in enqueue in mdc_read_entry,
3000  *                      and partial readdir result will be stored in it.
3001  * \param[in] offset    starting hash offset
3002  * \param[out] ppage    the page holding the entry. Note: because the entry
3003  *                      will be accessed in upper layer, so we need hold the
3004  *                      page until the usages of entry is finished, see
3005  *                      ll_dir_entry_next.
3006  *
3007  * retval               =0 if get entry successfully
3008  *                      <0 cannot get entry
3009  */
3010 static int lmv_striped_read_page(struct obd_export *exp,
3011                                  struct md_op_data *op_data,
3012                                  struct md_readdir_info *mrinfo, __u64 offset,
3013                                  struct page **ppage)
3014 {
3015         struct page *page = NULL;
3016         struct lu_dirpage *dp;
3017         void *start;
3018         struct lu_dirent *ent;
3019         struct lu_dirent *last_ent;
3020         int stripe_count;
3021         struct lmv_dir_ctxt *ctxt;
3022         struct lu_dirent *next = NULL;
3023         __u16 ent_size;
3024         size_t left_bytes;
3025         int rc = 0;
3026         ENTRY;
3027
3028         /* Allocate a page and read entries from all of stripes and fill
3029          * the page by hash order */
3030         page = alloc_page(GFP_KERNEL);
3031         if (!page)
3032                 RETURN(-ENOMEM);
3033
3034         /* Initialize the entry page */
3035         dp = kmap(page);
3036         memset(dp, 0, sizeof(*dp));
3037         dp->ldp_hash_start = cpu_to_le64(offset);
3038
3039         start = dp + 1;
3040         left_bytes = PAGE_SIZE - sizeof(*dp);
3041         ent = start;
3042         last_ent = ent;
3043
3044         /* initalize dir read context */
3045         stripe_count = op_data->op_lso1->lso_lsm.lsm_md_stripe_count;
3046         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
3047         if (!ctxt)
3048                 GOTO(free_page, rc = -ENOMEM);
3049         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
3050         ctxt->ldc_op_data = op_data;
3051         ctxt->ldc_mrinfo = mrinfo;
3052         ctxt->ldc_hash = offset;
3053         ctxt->ldc_count = stripe_count;
3054
3055         while (1) {
3056                 next = lmv_dirent_next(ctxt);
3057
3058                 /* end of directory */
3059                 if (!next) {
3060                         ctxt->ldc_hash = MDS_DIR_END_OFF;
3061                         break;
3062                 }
3063                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
3064
3065                 ent_size = le16_to_cpu(next->lde_reclen);
3066
3067                 /* the last entry lde_reclen is 0, but it might not be the last
3068                  * one of this temporay dir page */
3069                 if (!ent_size)
3070                         ent_size = lu_dirent_calc_size(
3071                                         le16_to_cpu(next->lde_namelen),
3072                                         le32_to_cpu(next->lde_attrs));
3073                 /* page full */
3074                 if (ent_size > left_bytes)
3075                         break;
3076
3077                 memcpy(ent, next, ent_size);
3078
3079                 /* Replace . with master FID and Replace .. with the parent FID
3080                  * of master object */
3081                 if (strncmp(ent->lde_name, ".",
3082                             le16_to_cpu(ent->lde_namelen)) == 0 &&
3083                     le16_to_cpu(ent->lde_namelen) == 1)
3084                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
3085                 else if (strncmp(ent->lde_name, "..",
3086                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
3087                            le16_to_cpu(ent->lde_namelen) == 2)
3088                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
3089
3090                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
3091                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
3092                        le64_to_cpu(ent->lde_hash));
3093
3094                 left_bytes -= ent_size;
3095                 ent->lde_reclen = cpu_to_le16(ent_size);
3096                 last_ent = ent;
3097                 ent = (void *)ent + ent_size;
3098         };
3099
3100         last_ent->lde_reclen = 0;
3101
3102         if (ent == start)
3103                 dp->ldp_flags |= LDF_EMPTY;
3104         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
3105                 dp->ldp_flags |= LDF_COLLIDE;
3106         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
3107         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
3108
3109         put_lmv_dir_ctxt(ctxt);
3110         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
3111
3112         *ppage = page;
3113
3114         RETURN(0);
3115
3116 free_page:
3117         kunmap(page);
3118         __free_page(page);
3119
3120         return rc;
3121 }
3122
3123 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
3124                          struct md_readdir_info *mrinfo, __u64 offset,
3125                          struct page **ppage)
3126 {
3127         struct obd_device *obd = exp->exp_obd;
3128         struct lmv_obd *lmv = &obd->u.lmv;
3129         struct lmv_tgt_desc *tgt;
3130         int rc;
3131
3132         ENTRY;
3133
3134         if (unlikely(lmv_dir_foreign(op_data->op_lso1)))
3135                 RETURN(-ENODATA);
3136
3137         if (unlikely(lmv_dir_striped(op_data->op_lso1))) {
3138                 rc = lmv_striped_read_page(exp, op_data, mrinfo, offset, ppage);
3139                 RETURN(rc);
3140         }
3141
3142         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
3143         if (IS_ERR(tgt))
3144                 RETURN(PTR_ERR(tgt));
3145
3146         rc = md_read_page(tgt->ltd_exp, op_data, mrinfo, offset, ppage);
3147
3148         RETURN(rc);
3149 }
3150
3151 /**
3152  * Unlink a file/directory
3153  *
3154  * Unlink a file or directory under the parent dir. The unlink request
3155  * usually will be sent to the MDT where the child is located, but if
3156  * the client does not have the child FID then request will be sent to the
3157  * MDT where the parent is located.
3158  *
3159  * If the parent is a striped directory then it also needs to locate which
3160  * stripe the name of the child is located, and replace the parent FID
3161  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
3162  * it will walk through all of sub-stripes until the child is being
3163  * unlinked finally.
3164  *
3165  * \param[in] exp       export refer to LMV
3166  * \param[in] op_data   different parameters transferred beween client
3167  *                      MD stacks, name, namelen, FIDs etc.
3168  *                      op_fid1 is the parent FID, op_fid2 is the child
3169  *                      FID.
3170  * \param[out] request  point to the request of unlink.
3171  *
3172  * retval               0 if succeed
3173  *                      negative errno if failed.
3174  */
3175 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
3176                       struct ptlrpc_request **request)
3177 {
3178         struct obd_device *obd = exp->exp_obd;
3179         struct lmv_obd *lmv = &obd->u.lmv;
3180         struct lmv_tgt_desc *tgt;
3181         struct lmv_tgt_desc *parent_tgt;
3182         struct mdt_body *body;
3183         int rc;
3184
3185         ENTRY;
3186
3187         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
3188         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
3189         op_data->op_cap = current_cap();
3190
3191 retry:
3192         parent_tgt = lmv_locate_tgt(lmv, op_data);
3193         if (IS_ERR(parent_tgt))
3194                 RETURN(PTR_ERR(parent_tgt));
3195
3196         if (likely(!fid_is_zero(&op_data->op_fid2))) {
3197                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3198                 if (IS_ERR(tgt))
3199                         RETURN(PTR_ERR(tgt));
3200         } else {
3201                 tgt = parent_tgt;
3202         }
3203
3204         /*
3205          * If child's fid is given, cancel unused locks for it if it is from
3206          * another export than parent.
3207          *
3208          * LOOKUP lock for child (fid3) should also be cancelled on parent
3209          * tgt_tgt in mdc_unlink().
3210          */
3211         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
3212
3213         if (parent_tgt != tgt)
3214                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
3215                                       LCK_EX, MDS_INODELOCK_LOOKUP,
3216                                       MF_MDC_CANCEL_FID3);
3217
3218         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
3219                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
3220         if (rc)
3221                 RETURN(rc);
3222
3223         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
3224                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
3225                tgt->ltd_index);
3226
3227         rc = md_unlink(tgt->ltd_exp, op_data, request);
3228         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
3229                 ptlrpc_req_finished(*request);
3230                 *request = NULL;
3231                 goto retry;
3232         }
3233
3234         if (rc != -EREMOTE)
3235                 RETURN(rc);
3236
3237         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
3238         if (body == NULL)
3239                 RETURN(-EPROTO);
3240
3241         /* Not cross-ref case, just get out of here. */
3242         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
3243                 RETURN(rc);
3244
3245         /* This is a remote object, try remote MDT. */
3246         op_data->op_fid2 = body->mbo_fid1;
3247         ptlrpc_req_finished(*request);
3248         *request = NULL;
3249
3250         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3251         if (IS_ERR(tgt))
3252                 RETURN(PTR_ERR(tgt));
3253
3254         goto retry;
3255 }
3256
3257 static int lmv_precleanup(struct obd_device *obd)
3258 {
3259         ENTRY;
3260         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
3261         RETURN(0);
3262 }
3263
3264 /**
3265  * Get by key a value associated with a LMV device.
3266  *
3267  * Dispatch request to lower-layer devices as needed.
3268  *
3269  * \param[in] env               execution environment for this thread
3270  * \param[in] exp               export for the LMV device
3271  * \param[in] keylen            length of key identifier
3272  * \param[in] key               identifier of key to get value for
3273  * \param[in] vallen            size of \a val
3274  * \param[out] val              pointer to storage location for value
3275  * \param[in] lsm               optional striping metadata of object
3276  *
3277  * \retval 0            on success
3278  * \retval negative     negated errno on failure
3279  */
3280 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
3281                         __u32 keylen, void *key, __u32 *vallen, void *val)
3282 {
3283         struct obd_device *obd;
3284         struct lmv_obd *lmv;
3285         struct lu_tgt_desc *tgt;
3286         int rc = 0;
3287
3288         ENTRY;
3289
3290         obd = class_exp2obd(exp);
3291         if (obd == NULL) {
3292                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3293                        exp->exp_handle.h_cookie);
3294                 RETURN(-EINVAL);
3295         }
3296
3297         lmv = &obd->u.lmv;
3298         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
3299                 LASSERT(*vallen == sizeof(__u32));
3300                 lmv_foreach_connected_tgt(lmv, tgt) {
3301                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
3302                                           vallen, val))
3303                                 RETURN(0);
3304                 }
3305                 RETURN(-EINVAL);
3306         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3307                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3308                    KEY_IS(KEY_CONN_DATA)) {
3309                 /*
3310                  * Forwarding this request to first MDS, it should know LOV
3311                  * desc.
3312                  */
3313                 tgt = lmv_tgt(lmv, 0);
3314                 if (!tgt)
3315                         RETURN(-ENODEV);
3316
3317                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3318                 if (!rc && KEY_IS(KEY_CONN_DATA))
3319                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3320                 RETURN(rc);
3321         } else if (KEY_IS(KEY_TGT_COUNT)) {
3322                 *((int *)val) = lmv->lmv_mdt_descs.ltd_tgts_size;
3323                 RETURN(0);
3324         }
3325
3326         CDEBUG(D_IOCTL, "Invalid key\n");
3327         RETURN(-EINVAL);
3328 }
3329
3330 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3331                      int *__rcs, struct ptlrpc_request_set *_set)
3332 {
3333         struct obd_device *obd = class_exp2obd(exp);
3334         struct ptlrpc_request_set *set = _set;
3335         struct lmv_obd *lmv = &obd->u.lmv;
3336         int tgt_count = lmv->lmv_mdt_count;
3337         struct lu_tgt_desc *tgt;
3338         struct fid_array *fat, **fas = NULL;
3339         int i, rc, **rcs = NULL;
3340
3341         if (!set) {
3342                 set = ptlrpc_prep_set();
3343                 if (!set)
3344                         RETURN(-ENOMEM);
3345         }
3346
3347         /* split FIDs by targets */
3348         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3349         if (fas == NULL)
3350                 GOTO(out, rc = -ENOMEM);
3351         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3352         if (rcs == NULL)
3353                 GOTO(out_fas, rc = -ENOMEM);
3354
3355         for (i = 0; i < fa->fa_nr; i++) {
3356                 unsigned int idx;
3357
3358                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3359                 if (rc) {
3360                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3361                                PFID(&fa->fa_fids[i]), rc);
3362                         continue;
3363                 }
3364                 LASSERT(idx < tgt_count);
3365                 if (!fas[idx])
3366                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3367                                   fa_fids[fa->fa_nr]));
3368                 if (!fas[idx])
3369                         GOTO(out, rc = -ENOMEM);
3370                 if (!rcs[idx])
3371                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3372                 if (!rcs[idx])
3373                         GOTO(out, rc = -ENOMEM);
3374
3375                 fat = fas[idx];
3376                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3377         }
3378
3379         lmv_foreach_connected_tgt(lmv, tgt) {
3380                 fat = fas[tgt->ltd_index];
3381                 if (!fat || fat->fa_nr == 0)
3382                         continue;
3383                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3384         }
3385
3386         rc = ptlrpc_set_wait(NULL, set);
3387         if (rc == 0) {
3388                 int j = 0;
3389                 for (i = 0; i < tgt_count; i++) {
3390                         fat = fas[i];
3391                         if (!fat || fat->fa_nr == 0)
3392                                 continue;
3393                         /* copy FIDs back */
3394                         memcpy(fa->fa_fids + j, fat->fa_fids,
3395                                fat->fa_nr * sizeof(struct lu_fid));
3396                         /* copy rcs back */
3397                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3398                         j += fat->fa_nr;
3399                 }
3400         }
3401         if (set != _set)
3402                 ptlrpc_set_destroy(set);
3403
3404 out:
3405         for (i = 0; i < tgt_count; i++) {
3406                 if (fas && fas[i])
3407                         OBD_FREE(fas[i], offsetof(struct fid_array,
3408                                                 fa_fids[fa->fa_nr]));
3409                 if (rcs && rcs[i])
3410                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3411         }
3412         if (rcs)
3413                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3414 out_fas:
3415         if (fas)
3416                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3417
3418         RETURN(rc);
3419 }
3420
3421 /**
3422  * Asynchronously set by key a value associated with a LMV device.
3423  *
3424  * Dispatch request to lower-layer devices as needed.
3425  *
3426  * \param[in] env       execution environment for this thread
3427  * \param[in] exp       export for the LMV device
3428  * \param[in] keylen    length of key identifier
3429  * \param[in] key       identifier of key to store value for
3430  * \param[in] vallen    size of value to store
3431  * \param[in] val       pointer to data to be stored
3432  * \param[in] set       optional list of related ptlrpc requests
3433  *
3434  * \retval 0            on success
3435  * \retval negative     negated errno on failure
3436  */
3437 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3438                               __u32 keylen, void *key, __u32 vallen, void *val,
3439                               struct ptlrpc_request_set *set)
3440 {
3441         struct lmv_tgt_desc *tgt;
3442         struct obd_device *obd;
3443         struct lmv_obd *lmv;
3444         int rc = 0;
3445         ENTRY;
3446
3447         obd = class_exp2obd(exp);
3448         if (obd == NULL) {
3449                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3450                        exp->exp_handle.h_cookie);
3451                 RETURN(-EINVAL);
3452         }
3453         lmv = &obd->u.lmv;
3454
3455         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3456             KEY_IS(KEY_DEFAULT_EASIZE)) {
3457                 int err = 0;
3458
3459                 lmv_foreach_connected_tgt(lmv, tgt) {
3460                         err = obd_set_info_async(env, tgt->ltd_exp,
3461                                                  keylen, key, vallen, val, set);
3462                         if (err && rc == 0)
3463                                 rc = err;
3464                 }
3465
3466                 RETURN(rc);
3467         }
3468
3469         RETURN(-EINVAL);
3470 }
3471
3472 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3473                             const struct lmv_mds_md_v1 *lmm1)
3474 {
3475         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3476         int             stripe_count;
3477         int             cplen;
3478         int             i;
3479         int             rc = 0;
3480         ENTRY;
3481
3482         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3483         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3484         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3485         if (CFS_FAIL_CHECK(OBD_FAIL_LMV_UNKNOWN_STRIPE))
3486                 lsm->lsm_md_hash_type = cfs_fail_val ?: LMV_HASH_TYPE_UNKNOWN;
3487         else
3488                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3489         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3490         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3491         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3492         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3493                         sizeof(lsm->lsm_md_pool_name));
3494
3495         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3496                 RETURN(-E2BIG);
3497
3498         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3499                "layout_version %d\n", lsm->lsm_md_stripe_count,
3500                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3501                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3502                lsm->lsm_md_layout_version);
3503
3504         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3505         for (i = 0; i < stripe_count; i++) {
3506                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3507                               &lmm1->lmv_stripe_fids[i]);
3508                 /*
3509                  * set default value -1, so lmv_locate_tgt() knows this stripe
3510                  * target is not initialized.
3511                  */
3512                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3513                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3514                         continue;
3515
3516                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3517                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3518                 if (rc == -ENOENT)
3519                         continue;
3520
3521                 if (rc)
3522                         RETURN(rc);
3523
3524                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3525                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3526         }
3527
3528         RETURN(rc);
3529 }
3530
3531 static inline int lmv_unpack_user_md(struct obd_export *exp,
3532                                      struct lmv_stripe_md *lsm,
3533                                      const struct lmv_user_md *lmu)
3534 {
3535         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3536         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3537         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3538         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3539         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3540         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3541         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3542
3543         return 0;
3544 }
3545
3546 struct lmv_stripe_object *lmv_stripe_object_alloc(__u32 magic,
3547                                                   const union lmv_mds_md *lmm,
3548                                                   size_t lmm_size)
3549 {
3550         struct lmv_stripe_object *lsm_obj;
3551         int size;
3552
3553         if (magic == LMV_MAGIC_FOREIGN) {
3554                 struct lmv_foreign_md *lfm;
3555
3556                 size = offsetof(typeof(*lfm), lfm_value[0]);
3557                 if (lmm_size < size)
3558                         RETURN(ERR_PTR(-EPROTO));
3559
3560                 size += le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3561                 if (lmm_size < size)
3562                         RETURN(ERR_PTR(-EPROTO));
3563
3564                 OBD_ALLOC_LARGE(lsm_obj, lmm_size +
3565                                 offsetof(typeof(*lsm_obj), lso_lfm));
3566         } else {
3567                 if (magic == LMV_MAGIC_V1) {
3568                         int count;
3569
3570                         size = offsetof(struct lmv_mds_md_v1,
3571                                         lmv_stripe_fids[0]);
3572                         if (lmm_size < size)
3573                                 RETURN(ERR_PTR(-EPROTO));
3574
3575                         count = lmv_mds_md_stripe_count_get(lmm);
3576                         size += count * sizeof(struct lu_fid);
3577                         if (lmm_size < size)
3578                                 RETURN(ERR_PTR(-EPROTO));
3579
3580                         size = lmv_stripe_md_size(count);
3581                 } else {
3582                         if (lmm && lmm_size < sizeof(struct lmv_user_md))
3583                                 RETURN(ERR_PTR(-EPROTO));
3584
3585                         /**
3586                          * Unpack default dirstripe(lmv_user_md) to
3587                          * lmv_stripe_md, stripecount should be 0 then.
3588                          */
3589                         size = lmv_stripe_md_size(0);
3590                 }
3591                 size += offsetof(typeof(*lsm_obj), lso_lsm);
3592                 OBD_ALLOC(lsm_obj, size);
3593         }
3594
3595         if (lsm_obj) {
3596                 atomic_set(&lsm_obj->lso_refs, 1);
3597                 RETURN(lsm_obj);
3598         }
3599
3600         RETURN(ERR_PTR(-ENOMEM));
3601 }
3602 EXPORT_SYMBOL(lmv_stripe_object_alloc);
3603
3604 static int lmv_stripe_object_create(struct obd_export *exp,
3605                                     struct lmv_stripe_object **lsop,
3606                                     const union lmv_mds_md *lmm,
3607                                     size_t lmm_size)
3608 {
3609         struct lmv_stripe_object *lsm_obj;
3610         __u32 magic;
3611         int rc;
3612         ENTRY;
3613
3614         LASSERT(lsop != NULL && *lsop == NULL);
3615
3616         if (lmm_size == 0)
3617                 RETURN(-EPROTO);
3618
3619         magic = le32_to_cpu(lmm->lmv_magic);
3620         if (magic == LMV_MAGIC_STRIPE)
3621                 RETURN(-EPERM);
3622
3623         if (magic != LMV_MAGIC_V1 && magic != LMV_USER_MAGIC &&
3624             magic != LMV_MAGIC_FOREIGN) {
3625                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3626                        exp->exp_obd->obd_name, magic, -EIO);
3627                 RETURN(-EIO);
3628         }
3629
3630         /* foreign lmv case */
3631         if (magic == LMV_MAGIC_FOREIGN) {
3632                 struct lmv_foreign_md *lfm;
3633
3634                 lsm_obj = lmv_stripe_object_alloc(magic, lmm, lmm_size);
3635                 if (IS_ERR(lsm_obj))
3636                         RETURN(PTR_ERR(lsm_obj));
3637
3638                 *lsop = lsm_obj;
3639                 lfm = &lsm_obj->lso_lfm;
3640                 lfm->lfm_magic = magic;
3641                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3642                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3643                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3644                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3645                        lfm->lfm_length);
3646                 RETURN(0);
3647         }
3648
3649         /* Unpack memmd */
3650         lsm_obj = lmv_stripe_object_alloc(magic, lmm, lmm_size);
3651         if (IS_ERR(lsm_obj))
3652                 RETURN(PTR_ERR(lsm_obj));
3653
3654         switch (magic) {
3655         case LMV_MAGIC_V1:
3656                 rc = lmv_unpack_md_v1(exp, &lsm_obj->lso_lsm, &lmm->lmv_md_v1);
3657                 break;
3658         case LMV_USER_MAGIC:
3659                 rc = lmv_unpack_user_md(exp, &lsm_obj->lso_lsm,
3660                                         &lmm->lmv_user_md);
3661                 break;
3662         default:
3663                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3664                        magic);
3665                 rc = -EINVAL;
3666                 break;
3667         }
3668
3669         if (rc != 0)
3670                 lmv_stripe_object_put(&lsm_obj);
3671
3672         *lsop = lsm_obj;
3673         RETURN(rc);
3674 }
3675
3676 struct lmv_stripe_object *
3677 lmv_stripe_object_get(struct lmv_stripe_object *lsm_obj)
3678 {
3679         if (lsm_obj == NULL)
3680                 return NULL;
3681
3682         atomic_inc(&lsm_obj->lso_refs);
3683         CDEBUG(D_INODE, "get %p %u\n", lsm_obj,
3684                atomic_read(&lsm_obj->lso_refs));
3685         return lsm_obj;
3686 }
3687 EXPORT_SYMBOL(lmv_stripe_object_get);
3688
3689 void lmv_stripe_object_put(struct lmv_stripe_object **lsop)
3690 {
3691         struct lmv_stripe_object *lsm_obj;
3692         size_t size;
3693         int i;
3694
3695         LASSERT(lsop != NULL);
3696
3697         lsm_obj = *lsop;
3698         if (lsm_obj == NULL)
3699                 return;
3700
3701         *lsop = NULL;
3702         CDEBUG(D_INODE, "put %p %u\n", lsm_obj,
3703                atomic_read(&lsm_obj->lso_refs) - 1);
3704
3705         if (!atomic_dec_and_test(&lsm_obj->lso_refs))
3706                 return;
3707
3708         if (lmv_dir_foreign(lsm_obj)) {
3709                 size = lsm_obj->lso_lfm.lfm_length +
3710                         offsetof(typeof(lsm_obj->lso_lfm), lfm_value[0]) +
3711                         offsetof(typeof(*lsm_obj), lso_lsm);
3712                 OBD_FREE_LARGE(lsm_obj, size);
3713                 return;
3714         }
3715
3716         if (lmv_dir_striped(lsm_obj)) {
3717                 struct lmv_stripe_md *lsm = &lsm_obj->lso_lsm;
3718
3719                 for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3720                         iput(lsm->lsm_md_oinfo[i].lmo_root);
3721                 size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3722         } else {
3723                 size = lmv_stripe_md_size(0);
3724         }
3725         OBD_FREE(lsm_obj, size + offsetof(typeof(*lsm_obj), lso_lsm));
3726 }
3727 EXPORT_SYMBOL(lmv_stripe_object_put);
3728
3729 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3730                              union ldlm_policy_data *policy,
3731                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3732                              void *opaque)
3733 {
3734         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3735         struct lu_tgt_desc *tgt;
3736         int err;
3737         int rc = 0;
3738
3739         ENTRY;
3740
3741         LASSERT(fid != NULL);
3742
3743         lmv_foreach_connected_tgt(lmv, tgt) {
3744                 if (!tgt->ltd_active)
3745                         continue;
3746
3747                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3748                                        opaque);
3749                 if (!rc)
3750                         rc = err;
3751         }
3752         RETURN(rc);
3753 }
3754
3755 static int lmv_set_lock_data(struct obd_export *exp,
3756                              const struct lustre_handle *lockh,
3757                              void *data, __u64 *bits)
3758 {
3759         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3760         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3761         int rc;
3762
3763         ENTRY;
3764
3765         if (tgt == NULL || tgt->ltd_exp == NULL)
3766                 RETURN(-EINVAL);
3767         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3768         RETURN(rc);
3769 }
3770
3771 static enum ldlm_mode
3772 lmv_lock_match(struct obd_export *exp, __u64 flags,
3773                const struct lu_fid *fid, enum ldlm_type type,
3774                union ldlm_policy_data *policy,
3775                enum ldlm_mode mode, struct lustre_handle *lockh)
3776 {
3777         struct obd_device *obd = exp->exp_obd;
3778         struct lmv_obd *lmv = &obd->u.lmv;
3779         struct lu_tgt_desc *tgt;
3780         __u64 bits = policy->l_inodebits.bits;
3781         enum ldlm_mode rc = LCK_MINMODE;
3782         int index;
3783         int i;
3784
3785         /* only one bit is set */
3786         LASSERT(bits && !(bits & (bits - 1)));
3787         /* With DNE every object can have two locks in different namespaces:
3788          * lookup lock in space of MDT storing direntry and update/open lock in
3789          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3790          * since this can be easily found, and only try others if that fails.
3791          */
3792         if (bits == MDS_INODELOCK_LOOKUP) {
3793                 for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3794                      i < lmv->lmv_mdt_descs.ltd_tgts_size; i++,
3795                      index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3796                         if (index < 0) {
3797                                 CDEBUG(D_HA,
3798                                        "%s: "DFID" is inaccessible: rc = %d\n",
3799                                        obd->obd_name, PFID(fid), index);
3800                                 index = 0;
3801                         }
3802                         tgt = lmv_tgt(lmv, index);
3803                         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3804                                 continue;
3805                         rc = md_lock_match(tgt->ltd_exp, flags, fid, type,
3806                                            policy, mode, lockh);
3807                         if (rc)
3808                                 break;
3809                 }
3810         } else {
3811                 tgt = lmv_fid2tgt(lmv, fid);
3812                 if (!IS_ERR(tgt) && tgt->ltd_exp && tgt->ltd_active)
3813                         rc = md_lock_match(tgt->ltd_exp, flags, fid, type,
3814                                            policy, mode, lockh);
3815         }
3816
3817         CDEBUG(D_INODE, "Lock match for "DFID": %d\n", PFID(fid), rc);
3818
3819         return rc;
3820 }
3821
3822 static int
3823 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3824                   struct obd_export *dt_exp, struct obd_export *md_exp,
3825                   struct lustre_md *md)
3826 {
3827         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3828         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3829
3830         if (!tgt || !tgt->ltd_exp)
3831                 return -EINVAL;
3832
3833         return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3834 }
3835
3836 static int lmv_put_lustre_md(struct obd_export *exp, struct lustre_md *md)
3837 {
3838         struct obd_device *obd = exp->exp_obd;
3839         struct lmv_obd *lmv = &obd->u.lmv;
3840         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3841         ENTRY;
3842
3843         lmv_stripe_object_put(&md->def_lsm_obj);
3844         lmv_stripe_object_put(&md->lsm_obj);
3845
3846         if (!tgt || !tgt->ltd_exp)
3847                 RETURN(-EINVAL);
3848         RETURN(0);
3849 }
3850
3851 static int lmv_set_open_replay_data(struct obd_export *exp,
3852                                     struct obd_client_handle *och,
3853                                     struct lookup_intent *it)
3854 {
3855         struct obd_device *obd = exp->exp_obd;
3856         struct lmv_obd *lmv = &obd->u.lmv;
3857         struct lmv_tgt_desc *tgt;
3858
3859         ENTRY;
3860
3861         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3862         if (IS_ERR(tgt))
3863                 RETURN(PTR_ERR(tgt));
3864
3865         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3866 }
3867
3868 static int lmv_clear_open_replay_data(struct obd_export *exp,
3869                                       struct obd_client_handle *och)
3870 {
3871         struct obd_device *obd = exp->exp_obd;
3872         struct lmv_obd *lmv = &obd->u.lmv;
3873         struct lmv_tgt_desc *tgt;
3874
3875         ENTRY;
3876
3877         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3878         if (IS_ERR(tgt))
3879                 RETURN(PTR_ERR(tgt));
3880
3881         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3882 }
3883
3884 static int lmv_intent_getattr_async(struct obd_export *exp,
3885                                     struct md_op_item *item)
3886 {
3887         struct md_op_data *op_data = &item->mop_data;
3888         struct obd_device *obd = exp->exp_obd;
3889         struct lmv_obd *lmv = &obd->u.lmv;
3890         struct lmv_tgt_desc *ptgt;
3891         struct lmv_tgt_desc *ctgt;
3892         int rc;
3893
3894         ENTRY;
3895
3896         if (!(fid_is_sane(&op_data->op_fid2) ||
3897               fid_is_zero(&op_data->op_fid2)))
3898                 RETURN(-EINVAL);
3899
3900         ptgt = lmv_locate_tgt(lmv, op_data);
3901         if (IS_ERR(ptgt))
3902                 RETURN(PTR_ERR(ptgt));
3903
3904         /*
3905          * Zeroed FID @op_fid2 means that the intent getattr() comes from
3906          * statahead by regularized file names. Currently only do statahead
3907          * for the children files located same as the parent directory.
3908          */
3909         if (!fid_is_zero(&op_data->op_fid2)) {
3910                 ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3911                 if (IS_ERR(ctgt))
3912                         RETURN(PTR_ERR(ctgt));
3913
3914                 /*
3915                  * remote object needs two RPCs to lookup and getattr,
3916                  * considering the complexity don't support statahead for now.
3917                  */
3918                 if (ctgt != ptgt)
3919                         RETURN(-EREMOTE);
3920         }
3921
3922         rc = md_intent_getattr_async(ptgt->ltd_exp, item);
3923
3924         RETURN(rc);
3925 }
3926
3927 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3928                                struct lu_fid *fid, __u64 *bits)
3929 {
3930         struct obd_device *obd = exp->exp_obd;
3931         struct lmv_obd *lmv = &obd->u.lmv;
3932         struct lmv_tgt_desc *tgt;
3933         int rc;
3934
3935         ENTRY;
3936
3937         tgt = lmv_fid2tgt(lmv, fid);
3938         if (IS_ERR(tgt))
3939                 RETURN(PTR_ERR(tgt));
3940
3941         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3942         RETURN(rc);
3943 }
3944
3945 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3946                                 const struct lmv_stripe_object *lso,
3947                                 const char *name, int namelen,
3948                                 struct lu_fid *fid)
3949 {
3950         const struct lmv_oinfo *oinfo;
3951
3952         LASSERT(lmv_dir_striped(lso));
3953         oinfo = lsm_name_to_stripe_info(lso, name, namelen, false);
3954         if (IS_ERR(oinfo))
3955                 return PTR_ERR(oinfo);
3956
3957         *fid = oinfo->lmo_fid;
3958
3959         RETURN(0);
3960 }
3961
3962 /**
3963  * For lmv, only need to send request to master MDT, and the master MDT will
3964  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3965  * we directly fetch data from the slave MDTs.
3966  */
3967 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3968                         struct obd_quotactl *oqctl)
3969 {
3970         struct obd_device *obd = class_exp2obd(exp);
3971         struct lmv_obd *lmv = &obd->u.lmv;
3972         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3973         __u64 curspace, curinodes;
3974         int rc = 0;
3975
3976         ENTRY;
3977
3978         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3979                 CERROR("master lmv inactive\n");
3980                 RETURN(-EIO);
3981         }
3982
3983         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3984                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3985                 RETURN(rc);
3986         }
3987
3988         curspace = curinodes = 0;
3989         lmv_foreach_connected_tgt(lmv, tgt) {
3990                 int err;
3991
3992                 if (!tgt->ltd_active)
3993                         continue;
3994
3995                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3996                 if (err) {
3997                         CERROR("getquota on mdt %d failed. %d\n",
3998                                tgt->ltd_index, err);
3999                         if (!rc)
4000                                 rc = err;
4001                 } else {
4002                         curspace += oqctl->qc_dqblk.dqb_curspace;
4003                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
4004                 }
4005         }
4006         oqctl->qc_dqblk.dqb_curspace = curspace;
4007         oqctl->qc_dqblk.dqb_curinodes = curinodes;
4008
4009         RETURN(rc);
4010 }
4011
4012 static int lmv_merge_attr(struct obd_export *exp,
4013                           const struct lmv_stripe_object *lso,
4014                           struct cl_attr *attr,
4015                           ldlm_blocking_callback cb_blocking)
4016 {
4017         const struct lmv_stripe_md *lsm = &lso->lso_lsm;
4018         int rc;
4019         int i;
4020
4021         if (!lmv_dir_striped(lso))
4022                 return 0;
4023
4024         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
4025         if (rc < 0)
4026                 return rc;
4027
4028         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
4029                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
4030
4031                 if (!inode)
4032                         continue;
4033
4034                 CDEBUG(D_INFO,
4035                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
4036                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
4037                        i_size_read(inode), (unsigned long long)inode->i_blocks,
4038                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
4039                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
4040
4041                 /* for slave stripe, it needs to subtract nlink for . and .. */
4042                 if (i != 0)
4043                         attr->cat_nlink += inode->i_nlink - 2;
4044                 else
4045                         attr->cat_nlink = inode->i_nlink;
4046
4047                 attr->cat_size += i_size_read(inode);
4048                 attr->cat_blocks += inode->i_blocks;
4049
4050                 if (attr->cat_atime < inode->i_atime.tv_sec)
4051                         attr->cat_atime = inode->i_atime.tv_sec;
4052
4053                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
4054                         attr->cat_ctime = inode->i_ctime.tv_sec;
4055
4056                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
4057                         attr->cat_mtime = inode->i_mtime.tv_sec;
4058         }
4059         return 0;
4060 }
4061
4062 static struct lu_batch *lmv_batch_create(struct obd_export *exp,
4063                                          enum lu_batch_flags flags,
4064                                          __u32 max_count)
4065 {
4066         struct lu_batch *bh;
4067         struct lmv_batch *lbh;
4068
4069         ENTRY;
4070         OBD_ALLOC_PTR(lbh);
4071         if (!lbh)
4072                 RETURN(ERR_PTR(-ENOMEM));
4073
4074         bh = &lbh->lbh_super;
4075         bh->lbt_flags = flags;
4076         bh->lbt_max_count = max_count;
4077
4078         if (flags & BATCH_FL_RQSET) {
4079                 bh->lbt_rqset = ptlrpc_prep_set();
4080                 if (bh->lbt_rqset == NULL) {
4081                         OBD_FREE_PTR(lbh);
4082                         RETURN(ERR_PTR(-ENOMEM));
4083                 }
4084         }
4085
4086         INIT_LIST_HEAD(&lbh->lbh_sub_batch_list);
4087         RETURN(bh);
4088 }
4089
4090 static int lmv_batch_stop(struct obd_export *exp, struct lu_batch *bh)
4091 {
4092         struct lmv_batch *lbh;
4093         struct lmvsub_batch *sub;
4094         struct lmvsub_batch *tmp;
4095         int rc = 0;
4096
4097         ENTRY;
4098
4099         lbh = container_of(bh, struct lmv_batch, lbh_super);
4100         list_for_each_entry_safe(sub, tmp, &lbh->lbh_sub_batch_list,
4101                                  sbh_sub_item) {
4102                 list_del(&sub->sbh_sub_item);
4103                 rc = md_batch_stop(sub->sbh_tgt->ltd_exp, sub->sbh_sub);
4104                 if (rc < 0) {
4105                         CERROR("%s: stop batch processing failed: rc = %d\n",
4106                                exp->exp_obd->obd_name, rc);
4107                         if (bh->lbt_result == 0)
4108                                 bh->lbt_result = rc;
4109                 }
4110                 OBD_FREE_PTR(sub);
4111         }
4112
4113         if (bh->lbt_flags & BATCH_FL_RQSET) {
4114                 rc = ptlrpc_set_wait(NULL, bh->lbt_rqset);
4115                 ptlrpc_set_destroy(bh->lbt_rqset);
4116         }
4117
4118         OBD_FREE_PTR(lbh);
4119         RETURN(rc);
4120 }
4121
4122 static int lmv_batch_flush(struct obd_export *exp, struct lu_batch *bh,
4123                            bool wait)
4124 {
4125         struct lmv_batch *lbh;
4126         struct lmvsub_batch *sub;
4127         int rc = 0;
4128         int rc1;
4129
4130         ENTRY;
4131
4132         lbh = container_of(bh, struct lmv_batch, lbh_super);
4133         list_for_each_entry(sub, &lbh->lbh_sub_batch_list, sbh_sub_item) {
4134                 rc1 = md_batch_flush(sub->sbh_tgt->ltd_exp, sub->sbh_sub, wait);
4135                 if (rc1 < 0) {
4136                         CERROR("%s: stop batch processing failed: rc = %d\n",
4137                                exp->exp_obd->obd_name, rc);
4138                         if (bh->lbt_result == 0)
4139                                 bh->lbt_result = rc;
4140
4141                         if (rc == 0)
4142                                 rc = rc1;
4143                 }
4144         }
4145
4146         if (wait && bh->lbt_flags & BATCH_FL_RQSET) {
4147                 rc1 = ptlrpc_set_wait(NULL, bh->lbt_rqset);
4148                 if (rc == 0)
4149                         rc = rc1;
4150         }
4151
4152         RETURN(rc);
4153 }
4154
4155 static inline struct lmv_tgt_desc *
4156 lmv_batch_locate_tgt(struct lmv_obd *lmv, struct md_op_item *item)
4157 {
4158         struct md_op_data *op_data = &item->mop_data;
4159         struct lmv_tgt_desc *tgt;
4160
4161         switch (item->mop_opc) {
4162         case MD_OP_GETATTR: {
4163                 struct lmv_tgt_desc *ptgt;
4164
4165                 if (!(fid_is_sane(&op_data->op_fid2) ||
4166                       fid_is_zero(&op_data->op_fid2)))
4167                         RETURN(ERR_PTR(-EINVAL));
4168
4169                 ptgt = lmv_locate_tgt(lmv, op_data);
4170                 if (IS_ERR(ptgt))
4171                         RETURN(ptgt);
4172
4173                 /*
4174                  * Zeroed @op_fid2 means that it is a statahead populating call
4175                  * in the file name pattern which is using file name format to
4176                  * prefetch the attributes. Thus it has no idea about the FID of
4177                  * the children file. The children file is considered to be
4178                  * located on the same storage target with the parent directory
4179                  * or the stripped directory.
4180                  */
4181                 if (fid_is_zero(&op_data->op_fid2)) {
4182                         tgt = ptgt;
4183                         break;
4184                 }
4185
4186                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
4187                 if (IS_ERR(tgt))
4188                         RETURN(tgt);
4189
4190                 /*
4191                  * Remote object needs two RPCs to lookup and getattr,
4192                  * considering the complexity don't support statahead for now.
4193                  */
4194                 if (tgt != ptgt)
4195                         RETURN(ERR_PTR(-EREMOTE));
4196
4197                 break;
4198         }
4199         default:
4200                 tgt = ERR_PTR(-ENOTSUPP);
4201         }
4202
4203         return tgt;
4204 }
4205
4206 static struct lu_batch *lmv_batch_lookup_sub(struct lmv_batch *lbh,
4207                                              struct lmv_tgt_desc *tgt)
4208 {
4209         struct lmvsub_batch *sub;
4210
4211         list_for_each_entry(sub, &lbh->lbh_sub_batch_list, sbh_sub_item) {
4212                 if (sub->sbh_tgt == tgt)
4213                         return sub->sbh_sub;
4214         }
4215
4216         return NULL;
4217 }
4218
4219 static struct lu_batch *lmv_batch_get_sub(struct lmv_batch *lbh,
4220                                           struct lmv_tgt_desc *tgt)
4221 {
4222         struct lmvsub_batch *sbh;
4223         struct lu_batch *child_bh;
4224         struct lu_batch *bh;
4225
4226         ENTRY;
4227
4228         child_bh = lmv_batch_lookup_sub(lbh, tgt);
4229         if (child_bh != NULL)
4230                 RETURN(child_bh);
4231
4232         OBD_ALLOC_PTR(sbh);
4233         if (sbh == NULL)
4234                 RETURN(ERR_PTR(-ENOMEM));
4235
4236         INIT_LIST_HEAD(&sbh->sbh_sub_item);
4237         sbh->sbh_tgt = tgt;
4238
4239         bh = &lbh->lbh_super;
4240         child_bh = md_batch_create(tgt->ltd_exp, bh->lbt_flags,
4241                                    bh->lbt_max_count);
4242         if (IS_ERR(child_bh)) {
4243                 OBD_FREE_PTR(sbh);
4244                 RETURN(child_bh);
4245         }
4246
4247         child_bh->lbt_rqset = bh->lbt_rqset;
4248         sbh->sbh_sub = child_bh;
4249         list_add(&sbh->sbh_sub_item, &lbh->lbh_sub_batch_list);
4250         RETURN(child_bh);
4251 }
4252
4253 static int lmv_batch_add(struct obd_export *exp, struct lu_batch *bh,
4254                          struct md_op_item *item)
4255 {
4256         struct obd_device *obd = exp->exp_obd;
4257         struct lmv_obd *lmv = &obd->u.lmv;
4258         struct lmv_tgt_desc *tgt;
4259         struct lmv_batch *lbh;
4260         struct lu_batch *child_bh;
4261         int rc;
4262
4263         ENTRY;
4264
4265         tgt = lmv_batch_locate_tgt(lmv, item);
4266         if (IS_ERR(tgt))
4267                 RETURN(PTR_ERR(tgt));
4268
4269         lbh = container_of(bh, struct lmv_batch, lbh_super);
4270         child_bh = lmv_batch_get_sub(lbh, tgt);
4271         if (IS_ERR(child_bh))
4272                 RETURN(PTR_ERR(child_bh));
4273
4274         rc = md_batch_add(tgt->ltd_exp, child_bh, item);
4275         RETURN(rc);
4276 }
4277
4278 static const struct obd_ops lmv_obd_ops = {
4279         .o_owner                = THIS_MODULE,
4280         .o_setup                = lmv_setup,
4281         .o_cleanup              = lmv_cleanup,
4282         .o_precleanup           = lmv_precleanup,
4283         .o_process_config       = lmv_process_config,
4284         .o_connect              = lmv_connect,
4285         .o_disconnect           = lmv_disconnect,
4286         .o_statfs               = lmv_statfs,
4287         .o_get_info             = lmv_get_info,
4288         .o_set_info_async       = lmv_set_info_async,
4289         .o_notify               = lmv_notify,
4290         .o_get_uuid             = lmv_get_uuid,
4291         .o_fid_alloc            = lmv_fid_alloc,
4292         .o_iocontrol            = lmv_iocontrol,
4293         .o_quotactl             = lmv_quotactl
4294 };
4295
4296 static const struct md_ops lmv_md_ops = {
4297         .m_get_root             = lmv_get_root,
4298         .m_null_inode           = lmv_null_inode,
4299         .m_close                = lmv_close,
4300         .m_create               = lmv_create,
4301         .m_enqueue              = lmv_enqueue,
4302         .m_getattr              = lmv_getattr,
4303         .m_getxattr             = lmv_getxattr,
4304         .m_getattr_name         = lmv_getattr_name,
4305         .m_intent_lock          = lmv_intent_lock,
4306         .m_link                 = lmv_link,
4307         .m_rename               = lmv_rename,
4308         .m_setattr              = lmv_setattr,
4309         .m_setxattr             = lmv_setxattr,
4310         .m_fsync                = lmv_fsync,
4311         .m_file_resync          = lmv_file_resync,
4312         .m_read_page            = lmv_read_page,
4313         .m_unlink               = lmv_unlink,
4314         .m_init_ea_size         = lmv_init_ea_size,
4315         .m_cancel_unused        = lmv_cancel_unused,
4316         .m_set_lock_data        = lmv_set_lock_data,
4317         .m_lock_match           = lmv_lock_match,
4318         .m_get_lustre_md        = lmv_get_lustre_md,
4319         .m_put_lustre_md        = lmv_put_lustre_md,
4320         .m_merge_attr           = lmv_merge_attr,
4321         .m_set_open_replay_data = lmv_set_open_replay_data,
4322         .m_clear_open_replay_data = lmv_clear_open_replay_data,
4323         .m_intent_getattr_async = lmv_intent_getattr_async,
4324         .m_revalidate_lock      = lmv_revalidate_lock,
4325         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
4326         .m_stripe_object_create = lmv_stripe_object_create,
4327         .m_rmfid                = lmv_rmfid,
4328         .m_batch_create         = lmv_batch_create,
4329         .m_batch_add            = lmv_batch_add,
4330         .m_batch_stop           = lmv_batch_stop,
4331         .m_batch_flush          = lmv_batch_flush,
4332 };
4333
4334 static int __init lmv_init(void)
4335 {
4336         int rc;
4337
4338         rc  = libcfs_setup();
4339         if (rc)
4340                 return rc;
4341
4342         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
4343                                    LUSTRE_LMV_NAME, NULL);
4344 }
4345
4346 static void __exit lmv_exit(void)
4347 {
4348         class_unregister_type(LUSTRE_LMV_NAME);
4349 }
4350
4351 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
4352 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
4353 MODULE_VERSION(LUSTRE_VERSION_STRING);
4354 MODULE_LICENSE("GPL");
4355
4356 module_init(lmv_init);
4357 module_exit(lmv_exit);