lustre/mdt/mdt_restripe.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  * This file is part of Lustre, http://www.lustre.org/
  24  *
  25  * lustre/mdt/mdt_restriper.c
  26  *
  27  * Lustre directory restripe and auto-split
  28  */
  29
  30 #define DEBUG_SUBSYSTEM S_MDS
  31
  32 #include <linux/sched.h>
  33 #include <linux/kthread.h>
  34 #include "mdt_internal.h"
  35
  36 /* add directory into splitting list and wake up restripe thread */
  37 void mdt_auto_split_add(struct mdt_thread_info *info, struct mdt_object *o)
  38 {
  39         struct mdt_device *mdt = info->mti_mdt;
  40         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
  41
  42         spin_lock(&restriper->mdr_lock);
  43         if (mdt->mdt_enable_dir_auto_split && !o->mot_restriping) {
  44                 o->mot_restriping = 1;
  45                 mdt_object_get(NULL, o);
  46                 LASSERT(list_empty(&o->mot_restripe_linkage));
  47                 list_add_tail(&o->mot_restripe_linkage,
  48                               &restriper->mdr_auto_splitting);
  49
  50                 CDEBUG(D_INFO, "add "DFID" into auto split list.\n",
  51                        PFID(mdt_object_fid(o)));
  52         }
  53         spin_unlock(&restriper->mdr_lock);
  54
  55         wake_up_process(restriper->mdr_task);
  56 }
  57
  58 void mdt_restripe_migrate_add(struct mdt_thread_info *info,
  59                               struct mdt_object *o)
  60 {
  61         struct mdt_device *mdt = info->mti_mdt;
  62         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
  63
  64         spin_lock(&restriper->mdr_lock);
  65         if (!o->mot_restriping) {
  66                 o->mot_restriping = 1;
  67                 o->mot_restripe_offset = 0;
  68                 mdt_object_get(NULL, o);
  69                 LASSERT(list_empty(&o->mot_restripe_linkage));
  70                 list_add_tail(&o->mot_restripe_linkage,
  71                               &restriper->mdr_migrating);
  72
  73                 CDEBUG(D_INFO, "add "DFID" into migrate list.\n",
  74                        PFID(mdt_object_fid(o)));
  75         }
  76         spin_unlock(&restriper->mdr_lock);
  77
  78         wake_up_process(restriper->mdr_task);
  79 }
  80
  81 void mdt_restripe_update_add(struct mdt_thread_info *info,
  82                              struct mdt_object *o)
  83 {
  84         struct mdt_device *mdt = info->mti_mdt;
  85         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
  86
  87         spin_lock(&restriper->mdr_lock);
  88         if (!o->mot_restriping) {
  89                 /* update LMV */
  90                 o->mot_restriping = 1;
  91                 mdt_object_get(NULL, o);
  92                 if (list_empty(&restriper->mdr_updating))
  93                         restriper->mdr_update_time = ktime_get_real_seconds();
  94                 LASSERT(list_empty(&o->mot_restripe_linkage));
  95                 list_add_tail(&o->mot_restripe_linkage,
  96                               &restriper->mdr_updating);
  97
  98                 CDEBUG(D_INFO, "add "DFID" into update list.\n",
  99                        PFID(mdt_object_fid(o)));
 100         }
 101         spin_unlock(&restriper->mdr_lock);
 102
 103         wake_up_process(restriper->mdr_task);
 104 }
 105
 106 static inline int mdt_fid_alloc(const struct lu_env *env,
 107                                 struct mdt_device *mdt,
 108                                 struct lu_fid *fid,
 109                                 struct mdt_object *parent,
 110                                 const struct lu_name *name)
 111 {
 112         struct lu_device *next = &mdt->mdt_child->md_lu_dev;
 113         struct lu_object *o = lu_object_next(&parent->mot_obj);
 114
 115         return next->ld_ops->ldo_fid_alloc(env, next, fid, o, name);
 116 }
 117
 118 static void mdt_auto_split_prep(struct mdt_thread_info *info,
 119                                 struct md_op_spec *spec,
 120                                 struct md_attr *ma,
 121                                 u32 lum_stripe_count)
 122 {
 123         struct lu_attr *attr = &ma->ma_attr;
 124         struct lmv_user_md_v1 *lum;
 125
 126         attr->la_ctime = attr->la_mtime = ktime_get_real_seconds();
 127         attr->la_valid = LA_CTIME | LA_MTIME;
 128
 129         lum = &info->mti_mdt->mdt_restriper.mdr_lmv.lmv_user_md;
 130         lum->lum_magic = cpu_to_le32(LMV_USER_MAGIC);
 131         lum->lum_stripe_count = cpu_to_le32(lum_stripe_count);
 132         lum->lum_stripe_offset = cpu_to_le32(LMV_OFFSET_DEFAULT);
 133         lum->lum_hash_type = 0;
 134
 135         spec->u.sp_ea.eadatalen = sizeof(*lum);
 136         spec->u.sp_ea.eadata = lum;
 137         spec->sp_cr_flags = MDS_OPEN_HAS_EA;
 138         spec->no_create = 0;
 139         spec->sp_migrate_close = 0;
 140 }
 141
 142 /* restripe directory: split or merge stripes */
 143 int mdt_restripe_internal(struct mdt_thread_info *info,
 144                           struct mdt_object *parent,
 145                           struct mdt_object *child,
 146                           const struct lu_name *lname,
 147                           struct lu_fid *tfid,
 148                           struct md_op_spec *spec,
 149                           struct md_attr *ma)
 150 {
 151         const struct lu_env *env = info->mti_env;
 152         struct mdt_device *mdt = info->mti_mdt;
 153         struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 154         struct lmv_mds_md_v1 *lmv;
 155         u32 lmv_stripe_count = 0;
 156         int rc;
 157
 158         ENTRY;
 159
 160         rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
 161         if (rc)
 162                 RETURN(rc);
 163
 164         if (ma->ma_valid & MA_LMV) {
 165                 lmv = &ma->ma_lmv->lmv_md_v1;
 166                 if (!lmv_is_sane(lmv))
 167                         RETURN(-EBADF);
 168
 169                 /* don't allow restripe if dir layout is changing */
 170                 if (lmv_is_layout_changing(lmv))
 171                         RETURN(-EBUSY);
 172
 173                 /* check whether stripe count and hash unchanged */
 174                 if (lum->lum_stripe_count == lmv->lmv_stripe_count &&
 175                     lum->lum_hash_type == lmv->lmv_hash_type)
 176                         RETURN(-EALREADY);
 177
 178                 lmv_stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
 179         } else if (le32_to_cpu(lum->lum_stripe_count) < 2) {
 180                 /* stripe count unchanged for plain directory */
 181                 RETURN(-EALREADY);
 182         }
 183
 184         if (le32_to_cpu(lum->lum_stripe_count) > lmv_stripe_count) {
 185                 /* split */
 186                 struct md_layout_change *mlc = &info->mti_mlc;
 187                 struct mdt_object *tobj = NULL;
 188                 s64 mtime = ma->ma_attr.la_mtime;
 189
 190                 ma->ma_need = MA_INODE;
 191                 ma->ma_valid = 0;
 192                 rc = mdt_attr_get_complex(info, child, ma);
 193                 if (rc)
 194                         RETURN(rc);
 195
 196                 if (!(ma->ma_valid & MA_INODE))
 197                         RETURN(-EBADF);
 198
 199                 /* mtime is from from client or set outside */
 200                 ma->ma_attr.la_mtime = mtime;
 201
 202                 if (!lmv_stripe_count) {
 203                         /* if child is plain directory, allocate @tobj as the
 204                          * master object, and make child the first stripe of
 205                          * @tobj.
 206                          */
 207                         tobj = mdt_object_new(env, mdt, tfid);
 208                         if (unlikely(IS_ERR(tobj)))
 209                                 RETURN(PTR_ERR(tobj));
 210                 }
 211
 212                 mlc->mlc_opc = MD_LAYOUT_SPLIT;
 213                 mlc->mlc_parent = mdt_object_child(parent);
 214                 mlc->mlc_target = tobj ? mdt_object_child(tobj) : NULL;
 215                 mlc->mlc_attr = &ma->ma_attr;
 216                 mlc->mlc_name = lname;
 217                 mlc->mlc_spec = spec;
 218                 rc = mo_layout_change(env, mdt_object_child(child), mlc);
 219                 if (!rc) {
 220                         /* FID and attr need to be replied to client for manual
 221                          * restripe.
 222                          */
 223                         ma->ma_need = MA_INODE;
 224                         ma->ma_valid = 0;
 225                         rc = mdt_attr_get_complex(info,
 226                                         lmv_stripe_count ? child : tobj, ma);
 227                 }
 228                 if (tobj)
 229                         mdt_object_put(env, tobj);
 230                 else
 231                         *tfid = *mdt_object_fid(child);
 232         } else {
 233                 /* merge only needs to override LMV */
 234                 struct lu_buf *buf = &info->mti_buf;
 235                 __u32 version;
 236
 237                 LASSERT(ma->ma_valid & MA_LMV);
 238                 lmv = &ma->ma_lmv->lmv_md_v1;
 239                 version = cpu_to_le32(lmv->lmv_layout_version);
 240
 241                 /* adjust 0 to 1 */
 242                 if (lum->lum_stripe_count == 0)
 243                         lum->lum_stripe_count = cpu_to_le32(1);
 244
 245                 lmv->lmv_hash_type |= cpu_to_le32(LMV_HASH_FLAG_MERGE |
 246                                                   LMV_HASH_FLAG_MIGRATION);
 247                 lmv->lmv_hash_type |= lum->lum_hash_type &
 248                                       cpu_to_le32(LMV_HASH_FLAG_FIXED);
 249                 lmv->lmv_merge_offset = lum->lum_stripe_count;
 250                 lmv->lmv_merge_hash = lum->lum_hash_type;
 251                 lmv->lmv_layout_version = cpu_to_le32(++version);
 252
 253                 buf->lb_buf = lmv;
 254                 buf->lb_len = sizeof(*lmv);
 255                 rc = mo_xattr_set(env, mdt_object_child(child), buf,
 256                                   XATTR_NAME_LMV, LU_XATTR_REPLACE);
 257                 if (rc)
 258                         RETURN(rc);
 259
 260                 *tfid = *mdt_object_fid(child);
 261                 ma->ma_need = MA_INODE;
 262                 ma->ma_valid = 0;
 263                 rc = mdt_attr_get_complex(info, child, ma);
 264         }
 265
 266         RETURN(rc);
 267 }
 268
 269 static int mdt_auto_split(struct mdt_thread_info *info)
 270 {
 271         const struct lu_env *env = info->mti_env;
 272         struct mdt_device *mdt = info->mti_mdt;
 273         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 274         struct md_attr *ma = &info->mti_attr;
 275         struct md_op_spec *spec = &info->mti_spec;
 276         struct lu_name *lname = &info->mti_name;
 277         struct lu_fid *fid = &info->mti_tmp_fid2;
 278         struct mdt_object *parent = NULL;
 279         struct mdt_object *child = NULL;
 280         struct mdt_object *stripe = NULL;
 281         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 282         struct mdt_lock_handle *lhp;
 283         struct mdt_lock_handle *lhc;
 284         u32 lmv_stripe_count = 0;
 285         u32 lum_stripe_count = 0;
 286         int rc;
 287
 288         ENTRY;
 289
 290         if (!atomic_read(&mdt->mdt_mds_mds_conns))
 291                 RETURN(-EINVAL);
 292
 293         spin_lock(&restriper->mdr_lock);
 294         if (!list_empty(&restriper->mdr_auto_splitting)) {
 295                 child = list_entry(restriper->mdr_auto_splitting.next,
 296                                    typeof(*child), mot_restripe_linkage);
 297                 list_del_init(&child->mot_restripe_linkage);
 298         }
 299         spin_unlock(&restriper->mdr_lock);
 300
 301         if (!child)
 302                 RETURN(0);
 303
 304         LASSERT(child->mot_restriping);
 305
 306         rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
 307         if (rc)
 308                 GOTO(out, rc);
 309
 310         if (ma->ma_valid & MA_LMV) {
 311                 /* stripe dirent exceeds threshold, find its master object */
 312                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
 313
 314                 /* auto-split won't be done on striped directory master object
 315                  * directly, because it's triggered when dirent count exceeds
 316                  * threshold, however dirent count of master object is its
 317                  * stripe count.
 318                  */
 319                 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_STRIPE)
 320                         GOTO(out, rc = -EINVAL);
 321
 322                 /* race with migrate? */
 323                 if (lmv_hash_is_migrating(cpu_to_le32(lmv->lmv_hash_type)))
 324                         GOTO(out, rc = -EBUSY);
 325
 326                 lmv_stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
 327
 328                 /* save stripe to clear 'restriping' flag in the end to avoid
 329                  * trigger auto-split multiple times.
 330                  */
 331                 stripe = child;
 332                 child = NULL;
 333
 334                 /* get master object FID from linkea */
 335                 rc = mdt_attr_get_pfid(info, stripe, &ma->ma_pfid);
 336                 if (rc)
 337                         GOTO(out, rc);
 338
 339                 child = mdt_object_find(env, mdt, &ma->ma_pfid);
 340                 if (IS_ERR(child))
 341                         GOTO(out, rc = PTR_ERR(child));
 342
 343                 spin_lock(&restriper->mdr_lock);
 344                 if (child->mot_restriping) {
 345                         /* race? */
 346                         spin_unlock(&restriper->mdr_lock);
 347                         GOTO(out, rc = -EBUSY);
 348                 }
 349                 child->mot_restriping = 1;
 350                 spin_unlock(&restriper->mdr_lock);
 351
 352                 /* skip if master object is remote, let the first stripe
 353                  * to start splitting because dir split needs to be done
 354                  * on where master object is.
 355                  */
 356                 if (mdt_object_remote(child))
 357                         GOTO(restriping_clear, rc = -EREMOTE);
 358         }
 359
 360         /* striped directory split adds mdr_auto_split_delta stripes */
 361         lum_stripe_count = min_t(unsigned int,
 362                                 lmv_stripe_count +
 363                                         mdt->mdt_restriper.mdr_dir_split_delta,
 364                                 atomic_read(&mdt->mdt_mds_mds_conns) + 1);
 365         if (lmv_stripe_count >= lum_stripe_count)
 366                 GOTO(restriping_clear, rc = -EALREADY);
 367
 368         /* get dir name and parent FID */
 369         rc = mdt_attr_get_pfid_name(info, child, fid, lname);
 370         if (rc)
 371                 GOTO(restriping_clear, rc);
 372
 373         /* copy name out because mti_linkea will be used later, and name should
 374          * end with '\0'
 375          */
 376         memcpy(info->mti_filename, lname->ln_name, lname->ln_namelen);
 377         info->mti_filename[lname->ln_namelen] = '\0';
 378         lname->ln_name = info->mti_filename;
 379         CDEBUG(D_INFO, "split "DFID"/"DNAME" to count %u (MDT count %d)\n",
 380                PFID(fid), PNAME(lname), lum_stripe_count,
 381                atomic_read(&mdt->mdt_mds_mds_conns) + 1);
 382
 383         parent = mdt_object_find(env, mdt, fid);
 384         if (IS_ERR(parent))
 385                 GOTO(restriping_clear, rc = PTR_ERR(parent));
 386
 387         rc = mdt_fid_alloc(env, mdt, fid, child, NULL);
 388         if (rc < 0)
 389                 GOTO(restriping_clear, rc);
 390
 391         lhp = &info->mti_lh[MDT_LH_PARENT];
 392         rc = mdt_parent_lock(info, parent, lhp, lname, LCK_PW, true);
 393         if (rc)
 394                 GOTO(restriping_clear, rc);
 395
 396         lhc = &info->mti_lh[MDT_LH_CHILD];
 397         rc = mdt_object_stripes_lock(info, parent, child, lhc, einfo,
 398                                      MDS_INODELOCK_FULL, LCK_EX, true);
 399         if (rc)
 400                 GOTO(unlock_parent, rc);
 401
 402         mdt_auto_split_prep(info, spec, ma, lum_stripe_count);
 403
 404         rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
 405         EXIT;
 406
 407         mdt_object_stripes_unlock(info, child, lhc, einfo, rc);
 408 unlock_parent:
 409         mdt_object_unlock(info, parent, lhp, rc);
 410 restriping_clear:
 411         child->mot_restriping = 0;
 412         LASSERT(list_empty(&child->mot_restripe_linkage));
 413 out:
 414         /* -EALREADY:   dir is split already.
 415          * -EBUSY:      dir is opened, or is splitting by others.
 416          * -EREMOTE:    dir is remote.
 417          */
 418         if (rc && rc != -EALREADY && rc != -EBUSY && rc != -EREMOTE)
 419                 CERROR("%s: split "DFID"/"DNAME" to count %u failed: rc = %d\n",
 420                        mdt_obd_name(mdt), PFID(mdt_object_fid(child)),
 421                        PNAME(lname), lum_stripe_count, rc);
 422
 423         if (!IS_ERR_OR_NULL(child))
 424                 mdt_object_put(env, child);
 425
 426         if (stripe) {
 427                 LASSERT(stripe->mot_restriping);
 428                 LASSERT(list_empty(&stripe->mot_restripe_linkage));
 429                 stripe->mot_restriping = 0;
 430                 /* lock may not be taken, don't cache stripe LMV */
 431                 mo_invalidate(env, mdt_object_child(stripe));
 432                 mdt_object_put(env, stripe);
 433         }
 434
 435         if (!IS_ERR_OR_NULL(parent))
 436                 mdt_object_put(env, parent);
 437
 438         return rc;
 439 }
 440
 441 /* sub-files under one stripe are migrated, clear MIGRATION flag in its LMV */
 442 static int mdt_restripe_migrate_finish(struct mdt_thread_info *info,
 443                                        struct mdt_object *stripe,
 444                                        struct lmv_mds_md_v1 *lmv)
 445 {
 446         struct mdt_device *mdt = info->mti_mdt;
 447         struct lu_buf buf;
 448         struct mdt_lock_handle *lh;
 449         int rc;
 450
 451         ENTRY;
 452
 453         LASSERT(le32_to_cpu(lmv->lmv_magic) == LMV_MAGIC_STRIPE);
 454         LASSERT(lmv_is_restriping(lmv));
 455
 456         lmv->lmv_hash_type &= ~cpu_to_le32(LMV_HASH_FLAG_MIGRATION);
 457         buf.lb_buf = lmv;
 458         buf.lb_len = sizeof(*lmv);
 459
 460         lh = &info->mti_lh[MDT_LH_PARENT];
 461         rc = mdt_object_lock(info, stripe, lh, MDS_INODELOCK_XATTR, LCK_EX,
 462                              false);
 463         if (!rc)
 464                 rc = mo_xattr_set(info->mti_env, mdt_object_child(stripe), &buf,
 465                                   XATTR_NAME_LMV, LU_XATTR_REPLACE);
 466         mdt_object_unlock(info, stripe, lh, rc);
 467         if (rc)
 468                 CERROR("%s: update "DFID" LMV failed: rc = %d\n",
 469                        mdt_obd_name(mdt), PFID(mdt_object_fid(stripe)), rc);
 470
 471         LASSERT(!list_empty(&stripe->mot_restripe_linkage));
 472         LASSERT(stripe->mot_restriping);
 473
 474         spin_lock(&mdt->mdt_lock);
 475         stripe->mot_restriping = 0;
 476         list_del_init(&stripe->mot_restripe_linkage);
 477         spin_unlock(&mdt->mdt_lock);
 478
 479         mdt_object_put(info->mti_env, stripe);
 480
 481         RETURN(rc);
 482 }
 483
 484 static void mdt_restripe_migrate_prep(struct mdt_thread_info *info,
 485                                       const struct lu_fid *fid1,
 486                                       const struct lu_fid *fid2,
 487                                       const struct lu_name *lname,
 488                                       __u16 type,
 489                                       const struct lmv_mds_md_v1 *lmv)
 490 {
 491         struct lu_attr *attr = &info->mti_attr.ma_attr;
 492         struct mdt_reint_record *rr = &info->mti_rr;
 493         struct md_op_spec *spec = &info->mti_spec;
 494         struct lmv_user_md_v1 *lum;
 495
 496         attr->la_ctime = attr->la_mtime = ktime_get_real_seconds();
 497         attr->la_valid = LA_CTIME | LA_MTIME;
 498
 499         rr->rr_fid1 = fid1;
 500         rr->rr_fid2 = fid2;
 501         rr->rr_name = *lname;
 502
 503         lum = &info->mti_mdt->mdt_restriper.mdr_lmv.lmv_user_md;
 504         lum->lum_magic = cpu_to_le32(LMV_USER_MAGIC);
 505         lum->lum_stripe_offset = cpu_to_le32(LMV_OFFSET_DEFAULT);
 506         if (lmv_is_splitting(lmv)) {
 507                 lum->lum_stripe_count = lmv->lmv_stripe_count;
 508                 lum->lum_hash_type =
 509                         lmv->lmv_hash_type & le32_to_cpu(LMV_HASH_TYPE_MASK);
 510         } else if (lmv_is_merging(lmv)) {
 511                 lum->lum_stripe_count = lmv->lmv_merge_offset;
 512                 lum->lum_hash_type = lmv->lmv_merge_hash;
 513         }
 514
 515         spec->u.sp_ea.eadatalen = sizeof(*lum);
 516         spec->u.sp_ea.eadata = lum;
 517         spec->sp_cr_flags = MDS_OPEN_HAS_EA;
 518         spec->no_create = 0;
 519         spec->sp_migrate_close = 0;
 520         /* if 'nsonly' is set, don't migrate inode */
 521         if (S_ISDIR(type))
 522                 spec->sp_migrate_nsonly = 1;
 523         else
 524                 spec->sp_migrate_nsonly =
 525                         info->mti_mdt->mdt_dir_restripe_nsonly;
 526 }
 527
 528 /* migrate sub-file from @mdr_restripe_offset */
 529 static int mdt_restripe_migrate(struct mdt_thread_info *info)
 530 {
 531         const struct lu_env *env = info->mti_env;
 532         struct mdt_device *mdt = info->mti_mdt;
 533         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 534         struct mdt_object *stripe = NULL;
 535         struct mdt_object *master = NULL;
 536         struct md_attr *ma = &info->mti_attr;
 537         struct lmv_mds_md_v1 *lmv;
 538         struct lu_name *lname = &info->mti_name;
 539         struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
 540         struct lu_fid fid1;
 541         struct lu_fid fid2;
 542         struct lu_dirpage *dp;
 543         struct lu_dirent *ent;
 544         const char *name = NULL;
 545         int namelen = 0;
 546         __u16 type;
 547         int idx = 0;
 548         int len;
 549         int rc;
 550
 551         ENTRY;
 552
 553         if (list_empty(&restriper->mdr_migrating))
 554                 RETURN(0);
 555
 556         stripe = list_entry(restriper->mdr_migrating.next, typeof(*stripe),
 557                             mot_restripe_linkage);
 558
 559         /* get master object FID and stripe name */
 560         rc = mdt_attr_get_pfid_name(info, stripe, &fid1, lname);
 561         if (rc)
 562                 GOTO(out, rc);
 563
 564         snprintf(info->mti_filename, sizeof(info->mti_filename), DFID,
 565                  PFID(mdt_object_fid(stripe)));
 566         len = strlen(info->mti_filename) + 1;
 567         if (len >= lname->ln_namelen)
 568                 GOTO(out, rc = -EBADF);
 569
 570         while (len < lname->ln_namelen) {
 571                 if (!isdigit(lname->ln_name[len]))
 572                         GOTO(out, rc = -EBADF);
 573
 574                 idx = idx * 10 + lname->ln_name[len++] - '0';
 575         };
 576
 577         /* check whether stripe is newly created in split */
 578         rc = mdt_stripe_get(info, stripe, ma, XATTR_NAME_LMV);
 579         if (rc)
 580                 GOTO(out, rc);
 581
 582         if (!(ma->ma_valid & MA_LMV))
 583                 GOTO(out, rc = -ENODATA);
 584
 585         lmv = &ma->ma_lmv->lmv_md_v1;
 586         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_STRIPE)
 587                 GOTO(out, rc = -EBADF);
 588
 589         if (!lmv_is_restriping(lmv))
 590                 GOTO(out, rc = -EINVAL);
 591
 592         if ((lmv_is_splitting(lmv) &&
 593              idx >= le32_to_cpu(lmv->lmv_split_offset)) ||
 594             (lmv_is_merging(lmv) &&
 595              ((le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_TYPE_MASK) ==
 596                                                  LMV_HASH_TYPE_CRUSH ||
 597               (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_TYPE_MASK) ==
 598                                                  LMV_HASH_TYPE_CRUSH2) &&
 599              idx < le32_to_cpu(lmv->lmv_merge_offset))) {
 600                 /* new stripes doesn't need to migrate sub files in dir
 601                  * split, neither for target stripes in dir merge if hash type
 602                  * is CRUSH or CRUSH2.
 603                  */
 604                 rc = mdt_restripe_migrate_finish(info, stripe, lmv);
 605                 RETURN(rc);
 606         }
 607
 608         /* get sub file name @mot_restripe_offset.
 609          * TODO: read one dirent instead of whole page.
 610          */
 611         rdpg->rp_hash = stripe->mot_restripe_offset;
 612         rdpg->rp_count = PAGE_SIZE;
 613         rdpg->rp_npages = 1;
 614         rdpg->rp_attrs = LUDA_64BITHASH | LUDA_FID | LUDA_TYPE;
 615         rdpg->rp_pages = &restriper->mdr_page;
 616         rc = mo_readpage(env, mdt_object_child(stripe), rdpg);
 617         if (rc < 0)
 618                 GOTO(out, rc);
 619
 620         dp = page_address(restriper->mdr_page);
 621         for (ent = lu_dirent_start(dp); ent; ent = lu_dirent_next(ent)) {
 622                 LASSERT(le64_to_cpu(ent->lde_hash) >= rdpg->rp_hash);
 623
 624                 if (unlikely(!(le32_to_cpu(ent->lde_attrs) & LUDA_TYPE)))
 625                         GOTO(out, rc = -EINVAL);
 626
 627                 namelen = le16_to_cpu(ent->lde_namelen);
 628                 if (!namelen)
 629                         continue;
 630
 631                 if (name_is_dot_or_dotdot(ent->lde_name, namelen))
 632                         continue;
 633
 634                 name = ent->lde_name;
 635                 type = lu_dirent_type_get(ent);
 636                 break;
 637         }
 638
 639         if (!name) {
 640                 if (le64_to_cpu(dp->ldp_hash_end) == MDS_DIR_END_OFF) {
 641                         rc = mdt_restripe_migrate_finish(info, stripe, lmv);
 642                         RETURN(rc);
 643                 }
 644
 645                 GOTO(out, rc = -EBADF);
 646         }
 647
 648         /* copy name out because it should end with '\0' */
 649         memcpy(info->mti_filename, name, namelen);
 650         info->mti_filename[namelen] = '\0';
 651         lname->ln_name = info->mti_filename;
 652         lname->ln_namelen = namelen;
 653
 654         CDEBUG(D_INFO, "migrate "DFID"/"DNAME" type %ho\n",
 655                PFID(&fid1), PNAME(lname), type);
 656
 657         master = mdt_object_find(env, mdt, &fid1);
 658         if (IS_ERR(master))
 659                 GOTO(out, rc = PTR_ERR(master));
 660
 661         rc = mdt_fid_alloc(env, mdt, &fid2, master, lname);
 662         mdt_object_put(env, master);
 663         if (rc < 0)
 664                 GOTO(out, rc);
 665
 666         mdt_restripe_migrate_prep(info, &fid1, &fid2, lname, type, lmv);
 667
 668         rc = mdt_reint_migrate(info, NULL);
 669         /* mti_big_buf is allocated in XATTR migration */
 670         if (unlikely(info->mti_big_buf.lb_buf))
 671                 lu_buf_free(&info->mti_big_buf);
 672         if (rc == -EALREADY)
 673                 rc = 0;
 674         if (rc)
 675                 GOTO(out, rc);
 676
 677         LASSERT(ent);
 678         do {
 679                 ent = lu_dirent_next(ent);
 680                 if (!ent)
 681                         break;
 682
 683                 namelen = le16_to_cpu(ent->lde_namelen);
 684         } while (namelen == 0); /* Skip dummy record */
 685
 686         if (ent)
 687                 stripe->mot_restripe_offset = le64_to_cpu(ent->lde_hash);
 688         else
 689                 stripe->mot_restripe_offset = le64_to_cpu(dp->ldp_hash_end);
 690
 691         EXIT;
 692 out:
 693         if (rc) {
 694                 /* -EBUSY: file is opened by others */
 695                 if (rc != -EBUSY)
 696                         CERROR("%s: migrate "DFID"/"DNAME" failed: rc = %d\n",
 697                                mdt_obd_name(mdt), PFID(&fid1), PNAME(lname),
 698                                rc);
 699
 700                 spin_lock(&mdt->mdt_lock);
 701                 stripe->mot_restriping = 0;
 702                 list_del_init(&stripe->mot_restripe_linkage);
 703                 spin_unlock(&mdt->mdt_lock);
 704
 705                 mdt_object_put(env, stripe);
 706         }
 707
 708         return rc;
 709 }
 710
 711 static inline bool mdt_restripe_update_pending(struct mdt_thread_info *info)
 712 {
 713         struct mdt_device *mdt = info->mti_mdt;
 714
 715         if (list_empty(&mdt->mdt_restriper.mdr_updating))
 716                 return false;
 717
 718         return mdt->mdt_restriper.mdr_update_time < ktime_get_real_seconds();
 719 }
 720
 721 static void mdt_restripe_layout_update_prep(struct mdt_thread_info *info,
 722                                             const struct lu_fid *fid,
 723                                             const struct lmv_mds_md_v1 *lmv)
 724 {
 725         struct lu_attr *attr = &info->mti_attr.ma_attr;
 726         struct mdt_reint_record *rr = &info->mti_rr;
 727         struct lmv_user_md_v1 *lum;
 728
 729         attr->la_ctime = attr->la_mtime = ktime_get_real_seconds();
 730         attr->la_valid = LA_CTIME | LA_MTIME;
 731
 732         strncpy(info->mti_filename, XATTR_NAME_LMV,
 733                 sizeof(info->mti_filename));
 734
 735         lum = &info->mti_mdt->mdt_restriper.mdr_lmv.lmv_user_md;
 736         lum->lum_magic = cpu_to_le32(LMV_USER_MAGIC);
 737         lum->lum_stripe_offset = cpu_to_le32(LMV_OFFSET_DEFAULT);
 738         if (lmv_is_splitting(lmv)) {
 739                 lum->lum_stripe_count = lmv->lmv_stripe_count;
 740                 lum->lum_hash_type =
 741                         lmv->lmv_hash_type & le32_to_cpu(LMV_HASH_TYPE_MASK);
 742         } else if (lmv_is_merging(lmv)) {
 743                 lum->lum_stripe_count = lmv->lmv_merge_offset;
 744                 lum->lum_hash_type = lmv->lmv_merge_hash;
 745         }
 746
 747         rr->rr_opcode = REINT_SETXATTR;
 748         rr->rr_fid1 = fid;
 749         rr->rr_name.ln_name = info->mti_filename;
 750         rr->rr_name.ln_namelen = strlen(info->mti_filename);
 751         rr->rr_eadata = lum;
 752         rr->rr_eadatalen = sizeof(*lum);
 753 }
 754
 755 static int mdt_restripe_layout_update(struct mdt_thread_info *info)
 756 {
 757         const struct lu_env *env = info->mti_env;
 758         struct mdt_device *mdt = info->mti_mdt;
 759         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 760         struct md_attr *ma = &info->mti_attr;
 761         struct lu_fid *fid = &info->mti_tmp_fid1;
 762         struct mdt_object *master;
 763         struct mdt_object *stripe;
 764         struct lmv_mds_md_v1 *lmv;
 765         int i;
 766         int rc;
 767
 768         ENTRY;
 769
 770         if (list_empty(&restriper->mdr_updating))
 771                 RETURN(0);
 772
 773         master = list_entry(restriper->mdr_updating.next, typeof(*master),
 774                             mot_restripe_linkage);
 775
 776         rc = mdt_stripe_get(info, master, ma, XATTR_NAME_LMV);
 777         if (rc)
 778                 GOTO(out, rc);
 779
 780         if (!(ma->ma_valid & MA_LMV))
 781                 GOTO(out, rc = -ENODATA);
 782
 783         lmv = &ma->ma_lmv->lmv_md_v1;
 784         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
 785                 GOTO(out, rc = -EBADF);
 786
 787         if (!lmv_is_restriping(lmv))
 788                 GOTO(out, rc = -EINVAL);
 789
 790         /* use different buffer to store stripe LMV */
 791         ma->ma_lmv = &restriper->mdr_lmv;
 792         ma->ma_lmv_size = sizeof(restriper->mdr_lmv);
 793         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
 794                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
 795                 stripe = mdt_object_find(env, mdt, fid);
 796                 if (IS_ERR(stripe))
 797                         GOTO(out, rc = PTR_ERR(stripe));
 798
 799                 ma->ma_valid = 0;
 800                 rc = __mdt_stripe_get(info, stripe, ma, XATTR_NAME_LMV);
 801                 /* LMV is checked without lock, don't cache it */
 802                 mo_invalidate(env, mdt_object_child(stripe));
 803                 mdt_object_put(env, stripe);
 804                 if (rc)
 805                         GOTO(out, rc);
 806
 807                 if (!(ma->ma_valid & MA_LMV))
 808                         GOTO(out, rc = -ENODATA);
 809
 810                 /* check MIGRATION flag cleared on all stripes */
 811                 if (lmv_is_restriping(&ma->ma_lmv->lmv_md_v1))
 812                         GOTO(out, rc = -EINPROGRESS);
 813         }
 814
 815         mdt_restripe_layout_update_prep(info, mdt_object_fid(master), lmv);
 816
 817         rc = mdt_dir_layout_update(info);
 818         if (rc) {
 819                 CERROR("update "DFID" layout failed: rc = %d\n",
 820                        PFID(mdt_object_fid(master)), rc);
 821                 GOTO(out, rc);
 822         }
 823
 824 out:
 825         LASSERT(!list_empty(&master->mot_restripe_linkage));
 826         if (rc == -EINPROGRESS) {
 827                 restriper->mdr_update_time = ktime_get_real_seconds() + 5;
 828         } else {
 829                 spin_lock(&restriper->mdr_lock);
 830                 master->mot_restriping = 0;
 831                 list_del_init(&master->mot_restripe_linkage);
 832                 spin_unlock(&restriper->mdr_lock);
 833
 834                 mdt_object_put(env, master);
 835         }
 836
 837         return rc;
 838 }
 839
 840 static int mdt_restriper_main(void *arg)
 841 {
 842         struct mdt_thread_info *info = arg;
 843         struct mdt_device *mdt = info->mti_mdt;
 844         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 845
 846         ENTRY;
 847
 848         while (({set_current_state(TASK_IDLE);
 849                  !kthread_should_stop(); })) {
 850                 if (!list_empty(&restriper->mdr_auto_splitting)) {
 851                         __set_current_state(TASK_RUNNING);
 852                         mdt_auto_split(info);
 853                         cond_resched();
 854                 } else if (mdt_restripe_update_pending(info)) {
 855                         __set_current_state(TASK_RUNNING);
 856                         mdt_restripe_layout_update(info);
 857                         cond_resched();
 858                 } else if (!list_empty(&restriper->mdr_migrating)) {
 859                         __set_current_state(TASK_RUNNING);
 860                         mdt_restripe_migrate(info);
 861                         cond_resched();
 862                 } else {
 863                         schedule();
 864                 }
 865         }
 866         __set_current_state(TASK_RUNNING);
 867
 868         RETURN(0);
 869 }
 870
 871 int mdt_restriper_start(struct mdt_device *mdt)
 872 {
 873         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 874         struct task_struct *task;
 875         struct mdt_thread_info *info;
 876         struct lu_ucred *uc;
 877         int rc;
 878
 879         ENTRY;
 880
 881         spin_lock_init(&restriper->mdr_lock);
 882         INIT_LIST_HEAD(&restriper->mdr_auto_splitting);
 883         INIT_LIST_HEAD(&restriper->mdr_migrating);
 884         INIT_LIST_HEAD(&restriper->mdr_updating);
 885         restriper->mdr_dir_split_count = DIR_SPLIT_COUNT_DEFAULT;
 886         restriper->mdr_dir_split_delta = DIR_SPLIT_DELTA_DEFAULT;
 887
 888         restriper->mdr_page = alloc_page(GFP_KERNEL);
 889         if (!restriper->mdr_page)
 890                 RETURN(-ENOMEM);
 891
 892         rc = lu_env_init(&restriper->mdr_env, LCT_MD_THREAD);
 893         if (rc)
 894                 GOTO(out_page, rc);
 895
 896         rc = lu_context_init(&restriper->mdr_session, LCT_SERVER_SESSION);
 897         if (rc)
 898                 GOTO(out_env, rc);
 899
 900         lu_context_enter(&restriper->mdr_session);
 901         restriper->mdr_env.le_ses = &restriper->mdr_session;
 902
 903         info = lu_context_key_get(&restriper->mdr_env.le_ctx, &mdt_thread_key);
 904         info->mti_env = &restriper->mdr_env;
 905         info->mti_mdt = mdt;
 906         info->mti_pill = NULL;
 907         info->mti_dlm_req = NULL;
 908
 909         uc = mdt_ucred(info);
 910         uc->uc_valid = UCRED_OLD;
 911         uc->uc_o_uid = 0;
 912         uc->uc_o_gid = 0;
 913         uc->uc_o_fsuid = 0;
 914         uc->uc_o_fsgid = 0;
 915         uc->uc_uid = 0;
 916         uc->uc_gid = 0;
 917         uc->uc_fsuid = 0;
 918         uc->uc_fsgid = 0;
 919         uc->uc_suppgids[0] = -1;
 920         uc->uc_suppgids[1] = -1;
 921         uc->uc_cap = cap_combine(CAP_FS_SET, CAP_NFSD_SET);
 922         uc->uc_umask = 0644;
 923         uc->uc_ginfo = NULL;
 924         uc->uc_identity = NULL;
 925         /* do not let rbac interfere with restriper internal processing */
 926         uc->uc_rbac_file_perms = 1;
 927         uc->uc_rbac_dne_ops = 1;
 928         uc->uc_rbac_quota_ops = 1;
 929         uc->uc_rbac_byfid_ops = 1;
 930         uc->uc_rbac_chlg_ops = 1;
 931         uc->uc_rbac_fscrypt_admin = 1;
 932
 933         task = kthread_create(mdt_restriper_main, info, "mdt_restriper_%03d",
 934                               mdt_seq_site(mdt)->ss_node_id);
 935         if (IS_ERR(task)) {
 936                 rc = PTR_ERR(task);
 937                 CERROR("%s: Can't start directory restripe thread: rc %d\n",
 938                        mdt_obd_name(mdt), rc);
 939                 GOTO(out_ses, rc);
 940         }
 941         restriper->mdr_task = task;
 942         wake_up_process(task);
 943
 944         RETURN(0);
 945
 946 out_ses:
 947         lu_context_exit(restriper->mdr_env.le_ses);
 948         lu_context_fini(restriper->mdr_env.le_ses);
 949 out_env:
 950         lu_env_fini(&restriper->mdr_env);
 951 out_page:
 952         __free_page(restriper->mdr_page);
 953
 954         return rc;
 955 }
 956
 957 void mdt_restriper_stop(struct mdt_device *mdt)
 958 {
 959         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 960         struct lu_env *env = &restriper->mdr_env;
 961         struct mdt_object *mo, *next;
 962
 963         if (!restriper->mdr_task)
 964                 return;
 965
 966         kthread_stop(restriper->mdr_task);
 967         restriper->mdr_task = NULL;
 968
 969         list_for_each_entry_safe(mo, next, &restriper->mdr_auto_splitting,
 970                                  mot_restripe_linkage) {
 971                 list_del_init(&mo->mot_restripe_linkage);
 972                 mdt_object_put(env, mo);
 973         }
 974
 975         list_for_each_entry_safe(mo, next, &restriper->mdr_migrating,
 976                                  mot_restripe_linkage) {
 977                 list_del_init(&mo->mot_restripe_linkage);
 978                 mdt_object_put(env, mo);
 979         }
 980
 981         list_for_each_entry_safe(mo, next, &restriper->mdr_updating,
 982                                  mot_restripe_linkage) {
 983                 list_del_init(&mo->mot_restripe_linkage);
 984                 mdt_object_put(env, mo);
 985         }
 986
 987         __free_page(restriper->mdr_page);
 988
 989         lu_context_exit(env->le_ses);
 990         lu_context_fini(env->le_ses);
 991         lu_env_fini(env);
 992 }