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_merge_offset = lum->lum_stripe_count;
 248                 lmv->lmv_merge_hash = lum->lum_hash_type;
 249                 lmv->lmv_layout_version = cpu_to_le32(++version);
 250
 251                 buf->lb_buf = lmv;
 252                 buf->lb_len = sizeof(*lmv);
 253                 rc = mo_xattr_set(env, mdt_object_child(child), buf,
 254                                   XATTR_NAME_LMV, LU_XATTR_REPLACE);
 255                 if (rc)
 256                         RETURN(rc);
 257
 258                 *tfid = *mdt_object_fid(child);
 259                 ma->ma_need = MA_INODE;
 260                 ma->ma_valid = 0;
 261                 rc = mdt_attr_get_complex(info, child, ma);
 262         }
 263
 264         RETURN(rc);
 265 }
 266
 267 static int mdt_auto_split(struct mdt_thread_info *info)
 268 {
 269         const struct lu_env *env = info->mti_env;
 270         struct mdt_device *mdt = info->mti_mdt;
 271         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 272         struct md_attr *ma = &info->mti_attr;
 273         struct md_op_spec *spec = &info->mti_spec;
 274         struct lu_name *lname = &info->mti_name;
 275         struct lu_fid *fid = &info->mti_tmp_fid2;
 276         struct mdt_object *parent = NULL;
 277         struct mdt_object *child = NULL;
 278         struct mdt_object *stripe = NULL;
 279         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 280         struct mdt_lock_handle *lhp;
 281         struct mdt_lock_handle *lhc;
 282         u32 lmv_stripe_count = 0;
 283         u32 lum_stripe_count = 0;
 284         int rc;
 285
 286         ENTRY;
 287
 288         if (!atomic_read(&mdt->mdt_mds_mds_conns))
 289                 RETURN(-EINVAL);
 290
 291         spin_lock(&restriper->mdr_lock);
 292         if (!list_empty(&restriper->mdr_auto_splitting)) {
 293                 child = list_entry(restriper->mdr_auto_splitting.next,
 294                                    typeof(*child), mot_restripe_linkage);
 295                 list_del_init(&child->mot_restripe_linkage);
 296         }
 297         spin_unlock(&restriper->mdr_lock);
 298
 299         if (!child)
 300                 RETURN(0);
 301
 302         LASSERT(child->mot_restriping);
 303
 304         rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
 305         if (rc)
 306                 GOTO(out, rc);
 307
 308         if (ma->ma_valid & MA_LMV) {
 309                 /* stripe dirent exceeds threshold, find its master object */
 310                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
 311
 312                 /* auto-split won't be done on striped directory master object
 313                  * directly, because it's triggered when dirent count exceeds
 314                  * threshold, however dirent count of master object is its
 315                  * stripe count.
 316                  */
 317                 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_STRIPE)
 318                         GOTO(out, rc = -EINVAL);
 319
 320                 lmv_stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
 321
 322                 /* save stripe to clear 'restriping' flag in the end to avoid
 323                  * trigger auto-split multiple times.
 324                  */
 325                 stripe = child;
 326                 child = NULL;
 327
 328                 /* get master object FID from linkea */
 329                 rc = mdt_attr_get_pfid(info, stripe, &ma->ma_pfid);
 330                 if (rc)
 331                         GOTO(out, rc);
 332
 333                 child = mdt_object_find(env, mdt, &ma->ma_pfid);
 334                 if (IS_ERR(child))
 335                         GOTO(out, rc = PTR_ERR(child));
 336
 337                 spin_lock(&restriper->mdr_lock);
 338                 if (child->mot_restriping) {
 339                         /* race? */
 340                         spin_unlock(&restriper->mdr_lock);
 341                         GOTO(out, rc = -EBUSY);
 342                 }
 343                 child->mot_restriping = 1;
 344                 spin_unlock(&restriper->mdr_lock);
 345
 346                 /* skip if master object is remote, let the first stripe
 347                  * to start splitting because dir split needs to be done
 348                  * on where master object is.
 349                  */
 350                 if (mdt_object_remote(child))
 351                         GOTO(restriping_clear, rc = -EREMOTE);
 352         }
 353
 354         /* striped directory split adds mdr_auto_split_delta stripes */
 355         lum_stripe_count = min_t(unsigned int,
 356                                 lmv_stripe_count +
 357                                         mdt->mdt_restriper.mdr_dir_split_delta,
 358                                 atomic_read(&mdt->mdt_mds_mds_conns) + 1);
 359         if (lmv_stripe_count >= lum_stripe_count)
 360                 GOTO(restriping_clear, rc = -EALREADY);
 361
 362         /* get dir name and parent FID */
 363         rc = mdt_attr_get_pfid_name(info, child, fid, lname);
 364         if (rc)
 365                 GOTO(restriping_clear, rc);
 366
 367         /* copy name out because mti_linkea will be used later, and name should
 368          * end with '\0'
 369          */
 370         memcpy(info->mti_filename, lname->ln_name, lname->ln_namelen);
 371         info->mti_filename[lname->ln_namelen] = '\0';
 372         lname->ln_name = info->mti_filename;
 373         CDEBUG(D_INFO, "split "DFID"/"DNAME" to count %u (MDT count %d)\n",
 374                PFID(fid), PNAME(lname), lum_stripe_count,
 375                atomic_read(&mdt->mdt_mds_mds_conns) + 1);
 376
 377         parent = mdt_object_find(env, mdt, fid);
 378         if (IS_ERR(parent))
 379                 GOTO(restriping_clear, rc = PTR_ERR(parent));
 380
 381         rc = mdt_fid_alloc(env, mdt, fid, child, NULL);
 382         if (rc < 0)
 383                 GOTO(restriping_clear, rc);
 384
 385         lhp = &info->mti_lh[MDT_LH_PARENT];
 386         mdt_lock_pdo_init(lhp, LCK_PW, lname);
 387         rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
 388                                    true);
 389         if (rc)
 390                 GOTO(restriping_clear, rc);
 391
 392         lhc = &info->mti_lh[MDT_LH_CHILD];
 393         mdt_lock_reg_init(lhc, LCK_EX);
 394         if (mdt_object_remote(parent)) {
 395                 /* enqueue object remote LOOKUP lock */
 396                 rc = mdt_remote_object_lock(info, parent, mdt_object_fid(child),
 397                                             &lhc->mlh_rreg_lh,
 398                                             lhc->mlh_rreg_mode,
 399                                             MDS_INODELOCK_LOOKUP, false);
 400                 if (rc != ELDLM_OK)
 401                         GOTO(unlock_parent, rc);
 402         }
 403
 404         rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
 405                                     true);
 406         if (rc)
 407                 GOTO(unlock_child, rc);
 408
 409         mdt_auto_split_prep(info, spec, ma, lum_stripe_count);
 410
 411         rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
 412         EXIT;
 413
 414 unlock_child:
 415         mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 416 unlock_parent:
 417         mdt_object_unlock(info, parent, lhp, rc);
 418 restriping_clear:
 419         child->mot_restriping = 0;
 420         LASSERT(list_empty(&child->mot_restripe_linkage));
 421 out:
 422         /* -EALREADY:   dir is split already.
 423          * -EBUSY:      dir is opened, or is splitting by others.
 424          * -EREMOTE:    dir is remote.
 425          */
 426         if (rc && rc != -EALREADY && rc != -EBUSY && rc != -EREMOTE)
 427                 CERROR("%s: split "DFID"/"DNAME" to count %u failed: rc = %d\n",
 428                        mdt_obd_name(mdt), PFID(mdt_object_fid(child)),
 429                        PNAME(lname), lum_stripe_count, rc);
 430
 431         if (!IS_ERR_OR_NULL(child))
 432                 mdt_object_put(env, child);
 433
 434         if (stripe) {
 435                 LASSERT(stripe->mot_restriping);
 436                 LASSERT(list_empty(&stripe->mot_restripe_linkage));
 437                 stripe->mot_restriping = 0;
 438                 /* lock may not be taken, don't cache stripe LMV */
 439                 mo_invalidate(env, mdt_object_child(stripe));
 440                 mdt_object_put(env, stripe);
 441         }
 442
 443         if (!IS_ERR_OR_NULL(parent))
 444                 mdt_object_put(env, parent);
 445
 446         return rc;
 447 }
 448
 449 /* sub-files under one stripe are migrated, clear MIGRATION flag in its LMV */
 450 static int mdt_restripe_migrate_finish(struct mdt_thread_info *info,
 451                                        struct mdt_object *stripe,
 452                                        struct lmv_mds_md_v1 *lmv)
 453 {
 454         struct mdt_device *mdt = info->mti_mdt;
 455         struct lu_buf buf;
 456         struct mdt_lock_handle *lh;
 457         int rc;
 458
 459         ENTRY;
 460
 461         LASSERT(le32_to_cpu(lmv->lmv_magic) == LMV_MAGIC_STRIPE);
 462         LASSERT(lmv_is_restriping(lmv));
 463
 464         lmv->lmv_hash_type &= ~cpu_to_le32(LMV_HASH_FLAG_MIGRATION);
 465         buf.lb_buf = lmv;
 466         buf.lb_len = sizeof(*lmv);
 467
 468         lh = &info->mti_lh[MDT_LH_PARENT];
 469         mdt_lock_reg_init(lh, LCK_EX);
 470         rc = mdt_reint_object_lock(info, stripe, lh, MDS_INODELOCK_XATTR,
 471                                    false);
 472         if (!rc)
 473                 rc = mo_xattr_set(info->mti_env, mdt_object_child(stripe), &buf,
 474                                   XATTR_NAME_LMV, LU_XATTR_REPLACE);
 475         mdt_object_unlock(info, stripe, lh, rc);
 476         if (rc)
 477                 CERROR("%s: update "DFID" LMV failed: rc = %d\n",
 478                        mdt_obd_name(mdt), PFID(mdt_object_fid(stripe)), rc);
 479
 480         LASSERT(!list_empty(&stripe->mot_restripe_linkage));
 481         LASSERT(stripe->mot_restriping);
 482
 483         spin_lock(&mdt->mdt_lock);
 484         stripe->mot_restriping = 0;
 485         list_del_init(&stripe->mot_restripe_linkage);
 486         spin_unlock(&mdt->mdt_lock);
 487
 488         mdt_object_put(info->mti_env, stripe);
 489
 490         RETURN(rc);
 491 }
 492
 493 static void mdt_restripe_migrate_prep(struct mdt_thread_info *info,
 494                                       const struct lu_fid *fid1,
 495                                       const struct lu_fid *fid2,
 496                                       const struct lu_name *lname,
 497                                       __u16 type,
 498                                       const struct lmv_mds_md_v1 *lmv)
 499 {
 500         struct lu_attr *attr = &info->mti_attr.ma_attr;
 501         struct mdt_reint_record *rr = &info->mti_rr;
 502         struct md_op_spec *spec = &info->mti_spec;
 503         struct lmv_user_md_v1 *lum;
 504
 505         attr->la_ctime = attr->la_mtime = ktime_get_real_seconds();
 506         attr->la_valid = LA_CTIME | LA_MTIME;
 507
 508         rr->rr_fid1 = fid1;
 509         rr->rr_fid2 = fid2;
 510         rr->rr_name = *lname;
 511
 512         lum = &info->mti_mdt->mdt_restriper.mdr_lmv.lmv_user_md;
 513         lum->lum_magic = cpu_to_le32(LMV_USER_MAGIC);
 514         lum->lum_stripe_offset = cpu_to_le32(LMV_OFFSET_DEFAULT);
 515         if (lmv_is_splitting(lmv)) {
 516                 lum->lum_stripe_count = lmv->lmv_stripe_count;
 517                 lum->lum_hash_type =
 518                         lmv->lmv_hash_type & le32_to_cpu(LMV_HASH_TYPE_MASK);
 519         } else if (lmv_is_merging(lmv)) {
 520                 lum->lum_stripe_count = lmv->lmv_merge_offset;
 521                 lum->lum_hash_type = lmv->lmv_merge_hash;
 522         }
 523
 524         spec->u.sp_ea.eadatalen = sizeof(*lum);
 525         spec->u.sp_ea.eadata = lum;
 526         spec->sp_cr_flags = MDS_OPEN_HAS_EA;
 527         spec->no_create = 0;
 528         spec->sp_migrate_close = 0;
 529         /* if 'nsonly' is set, don't migrate inode */
 530         if (S_ISDIR(type))
 531                 spec->sp_migrate_nsonly = 1;
 532         else
 533                 spec->sp_migrate_nsonly =
 534                         info->mti_mdt->mdt_dir_restripe_nsonly;
 535 }
 536
 537 /* migrate sub-file from @mdr_restripe_offset */
 538 static int mdt_restripe_migrate(struct mdt_thread_info *info)
 539 {
 540         const struct lu_env *env = info->mti_env;
 541         struct mdt_device *mdt = info->mti_mdt;
 542         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 543         struct mdt_object *stripe = NULL;
 544         struct mdt_object *master = NULL;
 545         struct md_attr *ma = &info->mti_attr;
 546         struct lmv_mds_md_v1 *lmv;
 547         struct lu_name *lname = &info->mti_name;
 548         struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
 549         struct lu_fid fid1;
 550         struct lu_fid fid2;
 551         struct lu_dirpage *dp;
 552         struct lu_dirent *ent;
 553         const char *name = NULL;
 554         int namelen = 0;
 555         __u16 type;
 556         int idx = 0;
 557         int len;
 558         int rc;
 559
 560         ENTRY;
 561
 562         if (list_empty(&restriper->mdr_migrating))
 563                 RETURN(0);
 564
 565         stripe = list_entry(restriper->mdr_migrating.next, typeof(*stripe),
 566                             mot_restripe_linkage);
 567
 568         /* get master object FID and stripe name */
 569         rc = mdt_attr_get_pfid_name(info, stripe, &fid1, lname);
 570         if (rc)
 571                 GOTO(out, rc);
 572
 573         snprintf(info->mti_filename, sizeof(info->mti_filename), DFID,
 574                  PFID(mdt_object_fid(stripe)));
 575         len = strlen(info->mti_filename) + 1;
 576         if (len >= lname->ln_namelen)
 577                 GOTO(out, rc = -EBADF);
 578
 579         while (len < lname->ln_namelen) {
 580                 if (!isdigit(lname->ln_name[len]))
 581                         GOTO(out, rc = -EBADF);
 582
 583                 idx = idx * 10 + lname->ln_name[len++] - '0';
 584         };
 585
 586         /* check whether stripe is newly created in split */
 587         rc = mdt_stripe_get(info, stripe, ma, XATTR_NAME_LMV);
 588         if (rc)
 589                 GOTO(out, rc);
 590
 591         if (!(ma->ma_valid & MA_LMV))
 592                 GOTO(out, rc = -ENODATA);
 593
 594         lmv = &ma->ma_lmv->lmv_md_v1;
 595         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_STRIPE)
 596                 GOTO(out, rc = -EBADF);
 597
 598         if (!lmv_is_restriping(lmv))
 599                 GOTO(out, rc = -EINVAL);
 600
 601         if ((lmv_is_splitting(lmv) &&
 602              idx >= le32_to_cpu(lmv->lmv_split_offset)) ||
 603             (lmv_is_merging(lmv) &&
 604              le32_to_cpu(lmv->lmv_hash_type) == LMV_HASH_TYPE_CRUSH &&
 605              idx < le32_to_cpu(lmv->lmv_merge_offset))) {
 606                 /* new stripes doesn't need to migrate sub files in dir
 607                  * split, neither for target stripes in dir merge if hash type
 608                  * is CRUSH.
 609                  */
 610                 rc = mdt_restripe_migrate_finish(info, stripe, lmv);
 611                 RETURN(rc);
 612         }
 613
 614         /* get sub file name @mot_restripe_offset.
 615          * TODO: read one dirent instead of whole page.
 616          */
 617         rdpg->rp_hash = stripe->mot_restripe_offset;
 618         rdpg->rp_count = PAGE_SIZE;
 619         rdpg->rp_npages = 1;
 620         rdpg->rp_attrs = LUDA_64BITHASH | LUDA_FID | LUDA_TYPE;
 621         rdpg->rp_pages = &restriper->mdr_page;
 622         rc = mo_readpage(env, mdt_object_child(stripe), rdpg);
 623         if (rc < 0)
 624                 GOTO(out, rc);
 625
 626         dp = page_address(restriper->mdr_page);
 627         for (ent = lu_dirent_start(dp); ent; ent = lu_dirent_next(ent)) {
 628                 LASSERT(le64_to_cpu(ent->lde_hash) >= rdpg->rp_hash);
 629
 630                 if (unlikely(!(le32_to_cpu(ent->lde_attrs) & LUDA_TYPE)))
 631                         GOTO(out, rc = -EINVAL);
 632
 633                 namelen = le16_to_cpu(ent->lde_namelen);
 634                 if (!namelen)
 635                         continue;
 636
 637                 if (name_is_dot_or_dotdot(ent->lde_name, namelen))
 638                         continue;
 639
 640                 name = ent->lde_name;
 641                 type = lu_dirent_type_get(ent);
 642                 break;
 643         }
 644
 645         if (!name) {
 646                 if (le64_to_cpu(dp->ldp_hash_end) == MDS_DIR_END_OFF) {
 647                         rc = mdt_restripe_migrate_finish(info, stripe, lmv);
 648                         RETURN(rc);
 649                 }
 650
 651                 GOTO(out, rc = -EBADF);
 652         }
 653
 654         /* copy name out because it should end with '\0' */
 655         memcpy(info->mti_filename, name, namelen);
 656         info->mti_filename[namelen] = '\0';
 657         lname->ln_name = info->mti_filename;
 658         lname->ln_namelen = namelen;
 659
 660         CDEBUG(D_INFO, "migrate "DFID"/"DNAME" type %ho\n",
 661                PFID(&fid1), PNAME(lname), type);
 662
 663         master = mdt_object_find(env, mdt, &fid1);
 664         if (IS_ERR(master))
 665                 GOTO(out, rc = PTR_ERR(master));
 666
 667         rc = mdt_fid_alloc(env, mdt, &fid2, master, lname);
 668         mdt_object_put(env, master);
 669         if (rc < 0)
 670                 GOTO(out, rc);
 671
 672         mdt_restripe_migrate_prep(info, &fid1, &fid2, lname, type, lmv);
 673
 674         rc = mdt_reint_migrate(info, NULL);
 675         /* mti_big_buf is allocated in XATTR migration */
 676         if (unlikely(info->mti_big_buf.lb_buf))
 677                 lu_buf_free(&info->mti_big_buf);
 678         if (rc == -EALREADY)
 679                 rc = 0;
 680         if (rc)
 681                 GOTO(out, rc);
 682
 683         LASSERT(ent);
 684         do {
 685                 ent = lu_dirent_next(ent);
 686                 if (!ent)
 687                         break;
 688
 689                 namelen = le16_to_cpu(ent->lde_namelen);
 690         } while (namelen == 0); /* Skip dummy record */
 691
 692         if (ent)
 693                 stripe->mot_restripe_offset = le64_to_cpu(ent->lde_hash);
 694         else
 695                 stripe->mot_restripe_offset = le64_to_cpu(dp->ldp_hash_end);
 696
 697         EXIT;
 698 out:
 699         if (rc) {
 700                 /* -EBUSY: file is opened by others */
 701                 if (rc != -EBUSY)
 702                         CERROR("%s: migrate "DFID"/"DNAME" failed: rc = %d\n",
 703                                mdt_obd_name(mdt), PFID(&fid1), PNAME(lname),
 704                                rc);
 705
 706                 spin_lock(&mdt->mdt_lock);
 707                 stripe->mot_restriping = 0;
 708                 list_del_init(&stripe->mot_restripe_linkage);
 709                 spin_unlock(&mdt->mdt_lock);
 710
 711                 mdt_object_put(env, stripe);
 712         }
 713
 714         return rc;
 715 }
 716
 717 static inline bool mdt_restripe_update_pending(struct mdt_thread_info *info)
 718 {
 719         struct mdt_device *mdt = info->mti_mdt;
 720
 721         if (list_empty(&mdt->mdt_restriper.mdr_updating))
 722                 return false;
 723
 724         return mdt->mdt_restriper.mdr_update_time < ktime_get_real_seconds();
 725 }
 726
 727 static void mdt_restripe_layout_update_prep(struct mdt_thread_info *info,
 728                                             const struct lu_fid *fid,
 729                                             const struct lmv_mds_md_v1 *lmv)
 730 {
 731         struct lu_attr *attr = &info->mti_attr.ma_attr;
 732         struct mdt_reint_record *rr = &info->mti_rr;
 733         struct lmv_user_md_v1 *lum;
 734
 735         attr->la_ctime = attr->la_mtime = ktime_get_real_seconds();
 736         attr->la_valid = LA_CTIME | LA_MTIME;
 737
 738         strncpy(info->mti_filename, XATTR_NAME_LMV,
 739                 sizeof(info->mti_filename));
 740
 741         lum = &info->mti_mdt->mdt_restriper.mdr_lmv.lmv_user_md;
 742         lum->lum_magic = cpu_to_le32(LMV_USER_MAGIC);
 743         lum->lum_stripe_offset = cpu_to_le32(LMV_OFFSET_DEFAULT);
 744         if (lmv_is_splitting(lmv)) {
 745                 lum->lum_stripe_count = lmv->lmv_stripe_count;
 746                 lum->lum_hash_type =
 747                         lmv->lmv_hash_type & le32_to_cpu(LMV_HASH_TYPE_MASK);
 748         } else if (lmv_is_merging(lmv)) {
 749                 lum->lum_stripe_count = lmv->lmv_merge_offset;
 750                 lum->lum_hash_type = lmv->lmv_merge_hash;
 751         }
 752
 753         rr->rr_opcode = REINT_SETXATTR;
 754         rr->rr_fid1 = fid;
 755         rr->rr_name.ln_name = info->mti_filename;
 756         rr->rr_name.ln_namelen = strlen(info->mti_filename);
 757         rr->rr_eadata = lum;
 758         rr->rr_eadatalen = sizeof(*lum);
 759 }
 760
 761 static int mdt_restripe_layout_update(struct mdt_thread_info *info)
 762 {
 763         const struct lu_env *env = info->mti_env;
 764         struct mdt_device *mdt = info->mti_mdt;
 765         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 766         struct md_attr *ma = &info->mti_attr;
 767         struct lu_fid *fid = &info->mti_tmp_fid1;
 768         struct mdt_object *master;
 769         struct mdt_object *stripe;
 770         struct lmv_mds_md_v1 *lmv;
 771         int i;
 772         int rc;
 773
 774         ENTRY;
 775
 776         if (list_empty(&restriper->mdr_updating))
 777                 RETURN(0);
 778
 779         master = list_entry(restriper->mdr_updating.next, typeof(*master),
 780                             mot_restripe_linkage);
 781
 782         rc = mdt_stripe_get(info, master, ma, XATTR_NAME_LMV);
 783         if (rc)
 784                 GOTO(out, rc);
 785
 786         if (!(ma->ma_valid & MA_LMV))
 787                 GOTO(out, rc = -ENODATA);
 788
 789         lmv = &ma->ma_lmv->lmv_md_v1;
 790         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
 791                 GOTO(out, rc = -EBADF);
 792
 793         if (!lmv_is_restriping(lmv))
 794                 GOTO(out, rc = -EINVAL);
 795
 796         /* use different buffer to store stripe LMV */
 797         ma->ma_lmv = &restriper->mdr_lmv;
 798         ma->ma_lmv_size = sizeof(restriper->mdr_lmv);
 799         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
 800                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
 801                 stripe = mdt_object_find(env, mdt, fid);
 802                 if (IS_ERR(stripe))
 803                         GOTO(out, rc = PTR_ERR(stripe));
 804
 805                 ma->ma_valid = 0;
 806                 rc = __mdt_stripe_get(info, stripe, ma, XATTR_NAME_LMV);
 807                 /* LMV is checked without lock, don't cache it */
 808                 mo_invalidate(env, mdt_object_child(stripe));
 809                 mdt_object_put(env, stripe);
 810                 if (rc)
 811                         GOTO(out, rc);
 812
 813                 if (!(ma->ma_valid & MA_LMV))
 814                         GOTO(out, rc = -ENODATA);
 815
 816                 /* check MIGRATION flag cleared on all stripes */
 817                 if (lmv_is_restriping(&ma->ma_lmv->lmv_md_v1))
 818                         GOTO(out, rc = -EINPROGRESS);
 819         }
 820
 821         mdt_restripe_layout_update_prep(info, mdt_object_fid(master), lmv);
 822
 823         rc = mdt_dir_layout_update(info);
 824         if (rc) {
 825                 CERROR("update "DFID" layout failed: rc = %d\n",
 826                        PFID(mdt_object_fid(master)), rc);
 827                 GOTO(out, rc);
 828         }
 829
 830 out:
 831         LASSERT(!list_empty(&master->mot_restripe_linkage));
 832         if (rc == -EINPROGRESS) {
 833                 restriper->mdr_update_time = ktime_get_real_seconds() + 5;
 834         } else {
 835                 spin_lock(&restriper->mdr_lock);
 836                 master->mot_restriping = 0;
 837                 list_del_init(&master->mot_restripe_linkage);
 838                 spin_unlock(&restriper->mdr_lock);
 839
 840                 mdt_object_put(env, master);
 841         }
 842
 843         return rc;
 844 }
 845
 846 static int mdt_restriper_main(void *arg)
 847 {
 848         struct mdt_thread_info *info = arg;
 849         struct mdt_device *mdt = info->mti_mdt;
 850         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 851
 852         ENTRY;
 853
 854         while (({set_current_state(TASK_IDLE);
 855                  !kthread_should_stop(); })) {
 856                 if (!list_empty(&restriper->mdr_auto_splitting)) {
 857                         __set_current_state(TASK_RUNNING);
 858                         mdt_auto_split(info);
 859                         cond_resched();
 860                 } else if (mdt_restripe_update_pending(info)) {
 861                         __set_current_state(TASK_RUNNING);
 862                         mdt_restripe_layout_update(info);
 863                         cond_resched();
 864                 } else if (!list_empty(&restriper->mdr_migrating)) {
 865                         __set_current_state(TASK_RUNNING);
 866                         mdt_restripe_migrate(info);
 867                         cond_resched();
 868                 } else {
 869                         schedule();
 870                 }
 871         }
 872         __set_current_state(TASK_RUNNING);
 873
 874         RETURN(0);
 875 }
 876
 877 int mdt_restriper_start(struct mdt_device *mdt)
 878 {
 879         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 880         struct task_struct *task;
 881         struct mdt_thread_info *info;
 882         struct lu_ucred *uc;
 883         int rc;
 884
 885         ENTRY;
 886
 887         spin_lock_init(&restriper->mdr_lock);
 888         INIT_LIST_HEAD(&restriper->mdr_auto_splitting);
 889         INIT_LIST_HEAD(&restriper->mdr_migrating);
 890         INIT_LIST_HEAD(&restriper->mdr_updating);
 891         restriper->mdr_dir_split_count = DIR_SPLIT_COUNT_DEFAULT;
 892         restriper->mdr_dir_split_delta = DIR_SPLIT_DELTA_DEFAULT;
 893
 894         restriper->mdr_page = alloc_page(GFP_KERNEL);
 895         if (!restriper->mdr_page)
 896                 RETURN(-ENOMEM);
 897
 898         rc = lu_env_init(&restriper->mdr_env, LCT_MD_THREAD);
 899         if (rc)
 900                 GOTO(out_page, rc);
 901
 902         rc = lu_context_init(&restriper->mdr_session, LCT_SERVER_SESSION);
 903         if (rc)
 904                 GOTO(out_env, rc);
 905
 906         lu_context_enter(&restriper->mdr_session);
 907         restriper->mdr_env.le_ses = &restriper->mdr_session;
 908
 909         info = lu_context_key_get(&restriper->mdr_env.le_ctx, &mdt_thread_key);
 910         info->mti_env = &restriper->mdr_env;
 911         info->mti_mdt = mdt;
 912         info->mti_pill = NULL;
 913         info->mti_dlm_req = NULL;
 914
 915         uc = mdt_ucred(info);
 916         uc->uc_valid = UCRED_OLD;
 917         uc->uc_o_uid = 0;
 918         uc->uc_o_gid = 0;
 919         uc->uc_o_fsuid = 0;
 920         uc->uc_o_fsgid = 0;
 921         uc->uc_uid = 0;
 922         uc->uc_gid = 0;
 923         uc->uc_fsuid = 0;
 924         uc->uc_fsgid = 0;
 925         uc->uc_suppgids[0] = -1;
 926         uc->uc_suppgids[1] = -1;
 927         uc->uc_cap = CFS_CAP_FS_MASK;
 928         uc->uc_umask = 0644;
 929         uc->uc_ginfo = NULL;
 930         uc->uc_identity = NULL;
 931
 932         task = kthread_create(mdt_restriper_main, info, "mdt_restriper_%03d",
 933                               mdt_seq_site(mdt)->ss_node_id);
 934         if (IS_ERR(task)) {
 935                 rc = PTR_ERR(task);
 936                 CERROR("%s: Can't start directory restripe thread: rc %d\n",
 937                        mdt_obd_name(mdt), rc);
 938                 GOTO(out_ses, rc);
 939         }
 940         restriper->mdr_task = task;
 941         wake_up_process(task);
 942
 943         RETURN(0);
 944
 945 out_ses:
 946         lu_context_exit(restriper->mdr_env.le_ses);
 947         lu_context_fini(restriper->mdr_env.le_ses);
 948 out_env:
 949         lu_env_fini(&restriper->mdr_env);
 950 out_page:
 951         __free_page(restriper->mdr_page);
 952
 953         return rc;
 954 }
 955
 956 void mdt_restriper_stop(struct mdt_device *mdt)
 957 {
 958         struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
 959         struct lu_env *env = &restriper->mdr_env;
 960         struct mdt_object *mo, *next;
 961
 962         if (!restriper->mdr_task)
 963                 return;
 964
 965         kthread_stop(restriper->mdr_task);
 966         restriper->mdr_task = NULL;
 967
 968         list_for_each_entry_safe(mo, next, &restriper->mdr_auto_splitting,
 969                                  mot_restripe_linkage) {
 970                 list_del_init(&mo->mot_restripe_linkage);
 971                 mdt_object_put(env, mo);
 972         }
 973
 974         list_for_each_entry_safe(mo, next, &restriper->mdr_migrating,
 975                                  mot_restripe_linkage) {
 976                 list_del_init(&mo->mot_restripe_linkage);
 977                 mdt_object_put(env, mo);
 978         }
 979
 980         list_for_each_entry_safe(mo, next, &restriper->mdr_updating,
 981                                  mot_restripe_linkage) {
 982                 list_del_init(&mo->mot_restripe_linkage);
 983                 mdt_object_put(env, mo);
 984         }
 985
 986         __free_page(restriper->mdr_page);
 987
 988         lu_context_exit(env->le_ses);
 989         lu_context_fini(env->le_ses);
 990         lu_env_fini(env);
 991 }