lustre/include/lustre_lmv.h

   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,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2014, 2016, Intel Corporation.
  24  */
  25 /*
  26  * lustre/include/lustre_lmv.h
  27  *
  28  * Lustre LMV structures and functions.
  29  *
  30  * Author: Di Wang <di.wang@intel.com>
  31  */
  32
  33 #ifndef _LUSTRE_LMV_H
  34 #define _LUSTRE_LMV_H
  35 #include <uapi/linux/lustre/lustre_idl.h>
  36
  37 struct lmv_oinfo {
  38         struct lu_fid   lmo_fid;
  39         u32             lmo_mds;
  40         struct inode    *lmo_root;
  41 };
  42
  43 struct lmv_stripe_md {
  44         __u32   lsm_md_magic;
  45         __u32   lsm_md_stripe_count;
  46         __u32   lsm_md_master_mdt_index;
  47         __u32   lsm_md_hash_type;
  48         __u8    lsm_md_max_inherit;
  49         __u8    lsm_md_max_inherit_rr;
  50         __u32   lsm_md_layout_version;
  51         __u32   lsm_md_migrate_offset;
  52         __u32   lsm_md_migrate_hash;
  53         __u32   lsm_md_default_count;
  54         __u32   lsm_md_default_index;
  55         char    lsm_md_pool_name[LOV_MAXPOOLNAME + 1];
  56         struct lmv_oinfo lsm_md_oinfo[0];
  57 };
  58
  59 static inline bool lmv_dir_striped(const struct lmv_stripe_md *lsm)
  60 {
  61         return lsm && lsm->lsm_md_magic == LMV_MAGIC;
  62 }
  63
  64 static inline bool lmv_dir_foreign(const struct lmv_stripe_md *lsm)
  65 {
  66         return lsm && lsm->lsm_md_magic == LMV_MAGIC_FOREIGN;
  67 }
  68
  69 static inline bool lmv_dir_layout_changing(const struct lmv_stripe_md *lsm)
  70 {
  71         return lmv_dir_striped(lsm) &&
  72                lmv_hash_is_layout_changing(lsm->lsm_md_hash_type);
  73 }
  74
  75 static inline bool lmv_dir_bad_hash(const struct lmv_stripe_md *lsm)
  76 {
  77         if (!lmv_dir_striped(lsm))
  78                 return false;
  79
  80         if (lsm->lsm_md_hash_type & LMV_HASH_FLAG_BAD_TYPE)
  81                 return true;
  82
  83         return !lmv_is_known_hash_type(lsm->lsm_md_hash_type);
  84 }
  85
  86 static inline bool
  87 lsm_md_eq(const struct lmv_stripe_md *lsm1, const struct lmv_stripe_md *lsm2)
  88 {
  89         __u32 idx;
  90
  91         if (lsm1->lsm_md_magic != lsm2->lsm_md_magic ||
  92             lsm1->lsm_md_stripe_count != lsm2->lsm_md_stripe_count ||
  93             lsm1->lsm_md_master_mdt_index !=
  94                                 lsm2->lsm_md_master_mdt_index ||
  95             lsm1->lsm_md_hash_type != lsm2->lsm_md_hash_type ||
  96             lsm1->lsm_md_max_inherit != lsm2->lsm_md_max_inherit ||
  97             lsm1->lsm_md_max_inherit_rr != lsm2->lsm_md_max_inherit_rr ||
  98             lsm1->lsm_md_layout_version !=
  99                                 lsm2->lsm_md_layout_version ||
 100             lsm1->lsm_md_migrate_offset !=
 101                                 lsm2->lsm_md_migrate_offset ||
 102             lsm1->lsm_md_migrate_hash !=
 103                                 lsm2->lsm_md_migrate_hash ||
 104             strncmp(lsm1->lsm_md_pool_name, lsm2->lsm_md_pool_name,
 105                     sizeof(lsm1->lsm_md_pool_name)) != 0)
 106                 return false;
 107
 108         if (lmv_dir_striped(lsm1)) {
 109                 for (idx = 0; idx < lsm1->lsm_md_stripe_count; idx++) {
 110                         if (!lu_fid_eq(&lsm1->lsm_md_oinfo[idx].lmo_fid,
 111                                        &lsm2->lsm_md_oinfo[idx].lmo_fid))
 112                                 return false;
 113                 }
 114         } else if (lsm1->lsm_md_magic == LMV_USER_MAGIC_SPECIFIC) {
 115                 for (idx = 0; idx < lsm1->lsm_md_stripe_count; idx++) {
 116                         if (lsm1->lsm_md_oinfo[idx].lmo_mds !=
 117                             lsm2->lsm_md_oinfo[idx].lmo_mds)
 118                                 return false;
 119                 }
 120         }
 121
 122         return true;
 123 }
 124
 125 static inline void lsm_md_dump(int mask, const struct lmv_stripe_md *lsm)
 126 {
 127         int i;
 128
 129         CDEBUG_LIMIT(mask,
 130                "dump LMV: magic=%#x count=%u index=%u hash=%s:%#x max_inherit=%hhu max_inherit_rr=%hhu version=%u migrate_offset=%u migrate_hash=%s:%x pool=%.*s\n",
 131                lsm->lsm_md_magic, lsm->lsm_md_stripe_count,
 132                lsm->lsm_md_master_mdt_index,
 133                lmv_is_known_hash_type(lsm->lsm_md_hash_type) ?
 134                 mdt_hash_name[lsm->lsm_md_hash_type & LMV_HASH_TYPE_MASK] :
 135                 "invalid", lsm->lsm_md_hash_type,
 136                lsm->lsm_md_max_inherit, lsm->lsm_md_max_inherit_rr,
 137                lsm->lsm_md_layout_version, lsm->lsm_md_migrate_offset,
 138                lmv_is_known_hash_type(lsm->lsm_md_migrate_hash) ?
 139                 mdt_hash_name[lsm->lsm_md_migrate_hash & LMV_HASH_TYPE_MASK] :
 140                 "invalid", lsm->lsm_md_migrate_hash,
 141                LOV_MAXPOOLNAME, lsm->lsm_md_pool_name);
 142
 143         if (!lmv_dir_striped(lsm))
 144                 return;
 145
 146         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
 147                 CDEBUG(mask, "stripe[%d] "DFID"\n",
 148                        i, PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
 149 }
 150
 151 union lmv_mds_md;
 152
 153 void lmv_free_memmd(struct lmv_stripe_md *lsm);
 154
 155 static inline void lmv1_le_to_cpu(struct lmv_mds_md_v1 *lmv_dst,
 156                                   const struct lmv_mds_md_v1 *lmv_src)
 157 {
 158         __u32 i;
 159
 160         lmv_dst->lmv_magic = le32_to_cpu(lmv_src->lmv_magic);
 161         lmv_dst->lmv_stripe_count = le32_to_cpu(lmv_src->lmv_stripe_count);
 162         lmv_dst->lmv_master_mdt_index =
 163                                 le32_to_cpu(lmv_src->lmv_master_mdt_index);
 164         lmv_dst->lmv_hash_type = le32_to_cpu(lmv_src->lmv_hash_type);
 165         lmv_dst->lmv_layout_version = le32_to_cpu(lmv_src->lmv_layout_version);
 166         if (lmv_src->lmv_stripe_count > LMV_MAX_STRIPE_COUNT)
 167                 return;
 168         for (i = 0; i < lmv_src->lmv_stripe_count; i++)
 169                 fid_le_to_cpu(&lmv_dst->lmv_stripe_fids[i],
 170                               &lmv_src->lmv_stripe_fids[i]);
 171 }
 172
 173 static inline void lmv_le_to_cpu(union lmv_mds_md *lmv_dst,
 174                                  const union lmv_mds_md *lmv_src)
 175 {
 176         switch (le32_to_cpu(lmv_src->lmv_magic)) {
 177         case LMV_MAGIC_V1:
 178                 lmv1_le_to_cpu(&lmv_dst->lmv_md_v1, &lmv_src->lmv_md_v1);
 179                 break;
 180         default:
 181                 break;
 182         }
 183 }
 184
 185 /* This hash is only for testing purpose */
 186 static inline unsigned int
 187 lmv_hash_all_chars(unsigned int count, const char *name, int namelen)
 188 {
 189         unsigned int c = 0;
 190         const unsigned char *p = (const unsigned char *)name;
 191
 192         while (--namelen >= 0)
 193                 c += p[namelen];
 194
 195         c = c % count;
 196
 197         return c;
 198 }
 199
 200 static inline unsigned int
 201 lmv_hash_fnv1a(unsigned int count, const char *name, int namelen)
 202 {
 203         __u64 hash;
 204
 205         hash = lustre_hash_fnv_1a_64(name, namelen);
 206
 207         return do_div(hash, count);
 208 }
 209
 210 /*
 211  * Robert Jenkins' function for mixing 32-bit values
 212  * http://burtleburtle.net/bob/hash/evahash.html
 213  * a, b = random bits, c = input and output
 214  *
 215  * Mixing inputs to generate an evenly distributed hash.
 216  */
 217 #define crush_hashmix(a, b, c)                          \
 218 do {                                                    \
 219         a = a - b;  a = a - c;  a = a ^ (c >> 13);      \
 220         b = b - c;  b = b - a;  b = b ^ (a << 8);       \
 221         c = c - a;  c = c - b;  c = c ^ (b >> 13);      \
 222         a = a - b;  a = a - c;  a = a ^ (c >> 12);      \
 223         b = b - c;  b = b - a;  b = b ^ (a << 16);      \
 224         c = c - a;  c = c - b;  c = c ^ (b >> 5);       \
 225         a = a - b;  a = a - c;  a = a ^ (c >> 3);       \
 226         b = b - c;  b = b - a;  b = b ^ (a << 10);      \
 227         c = c - a;  c = c - b;  c = c ^ (b >> 15);      \
 228 } while (0)
 229
 230 #define crush_hash_seed 1315423911
 231
 232 static inline __u32 crush_hash(__u32 a, __u32 b)
 233 {
 234         __u32 hash = crush_hash_seed ^ a ^ b;
 235         __u32 x = 231232;
 236         __u32 y = 1232;
 237
 238         crush_hashmix(a, b, hash);
 239         crush_hashmix(x, a, hash);
 240         crush_hashmix(b, y, hash);
 241
 242         return hash;
 243 }
 244
 245 /* refer to https://github.com/ceph/ceph/blob/master/src/crush/hash.c and
 246  * https://www.ssrc.ucsc.edu/Papers/weil-sc06.pdf for details of CRUSH
 247  * algorithm.
 248  */
 249 static inline unsigned int
 250 lmv_hash_crush(unsigned int count, const char *name, int namelen, bool crush2)
 251 {
 252         unsigned long long straw;
 253         unsigned long long highest_straw = 0;
 254         unsigned int pg_id;
 255         unsigned int idx = 0;
 256         int i;
 257
 258         /* put temp and backup file on the same MDT where target is located.
 259          * temporary file naming rule:
 260          * 1. rsync: .<target>.XXXXXX
 261          * 2. dstripe: <target>.XXXXXXXX
 262          */
 263         if (lu_name_is_temp_file(name, namelen, true, 6, crush2)) {
 264                 name++;
 265                 namelen -= 8;
 266         } else if (lu_name_is_temp_file(name, namelen, false, 8, crush2)) {
 267                 namelen -= 9;
 268         } else if (lu_name_is_backup_file(name, namelen, &i)) {
 269                 LASSERT(i < namelen);
 270                 namelen -= i;
 271         }
 272
 273         pg_id = lmv_hash_fnv1a(LMV_CRUSH_PG_COUNT, name, namelen);
 274
 275         /* distribute PG among all stripes pseudo-randomly, so they are almost
 276          * evenly distributed, and when stripe count changes, only (delta /
 277          * total) sub files need to be moved, herein 'delta' is added or removed
 278          * stripe count, 'total' is total stripe count before change for
 279          * removal, or count after change for addition.
 280          */
 281         for (i = 0; i < count; i++) {
 282                 straw = crush_hash(pg_id, i);
 283                 if (straw > highest_straw) {
 284                         highest_straw = straw;
 285                         idx = i;
 286                 }
 287         }
 288         LASSERT(idx < count);
 289
 290         return idx;
 291 }
 292
 293 /* directory layout may change in three ways:
 294  * 1. directory migration, in its LMV source stripes are appended after
 295  *    target stripes, \a migrate_hash is source hash type, \a migrate_offset is
 296  *    target stripe count,
 297  * 2. directory split, \a migrate_hash is hash type before split,
 298  *    \a migrate_offset is stripe count before split.
 299  * 3. directory merge, \a migrate_hash is hash type after merge,
 300  *    \a migrate_offset is stripe count after merge.
 301  */
 302 static inline int
 303 __lmv_name_to_stripe_index(__u32 hash_type, __u32 stripe_count,
 304                            __u32 migrate_hash, __u32 migrate_offset,
 305                            const char *name, int namelen, bool new_layout)
 306 {
 307         __u32 saved_hash = hash_type;
 308         __u32 saved_count = stripe_count;
 309         int stripe_index = 0;
 310
 311         LASSERT(namelen > 0);
 312         LASSERT(stripe_count > 0);
 313
 314         if (lmv_hash_is_splitting(hash_type)) {
 315                 if (!new_layout) {
 316                         hash_type = migrate_hash;
 317                         stripe_count = migrate_offset;
 318                 }
 319         } else if (lmv_hash_is_merging(hash_type)) {
 320                 if (new_layout) {
 321                         hash_type = migrate_hash;
 322                         stripe_count = migrate_offset;
 323                 }
 324         } else if (lmv_hash_is_migrating(hash_type)) {
 325                 if (new_layout) {
 326                         stripe_count = migrate_offset;
 327                 } else {
 328                         hash_type = migrate_hash;
 329                         stripe_count -= migrate_offset;
 330                 }
 331         }
 332
 333         if (stripe_count > 1) {
 334                 switch (hash_type & LMV_HASH_TYPE_MASK) {
 335                 case LMV_HASH_TYPE_ALL_CHARS:
 336                         stripe_index = lmv_hash_all_chars(stripe_count, name,
 337                                                           namelen);
 338                         break;
 339                 case LMV_HASH_TYPE_FNV_1A_64:
 340                         stripe_index = lmv_hash_fnv1a(stripe_count, name,
 341                                                       namelen);
 342                         break;
 343                 case LMV_HASH_TYPE_CRUSH:
 344                         stripe_index = lmv_hash_crush(stripe_count, name,
 345                                                       namelen, false);
 346                         break;
 347                 case LMV_HASH_TYPE_CRUSH2:
 348                         stripe_index = lmv_hash_crush(stripe_count, name,
 349                                                       namelen, true);
 350                         break;
 351                 default:
 352                         return -EBADFD;
 353                 }
 354         }
 355
 356         LASSERT(stripe_index < stripe_count);
 357
 358         if (!new_layout && lmv_hash_is_migrating(saved_hash))
 359                 stripe_index += migrate_offset;
 360
 361         LASSERT(stripe_index < saved_count);
 362
 363         CDEBUG(D_INFO, "name %.*s hash=%#x/%#x idx=%d/%u/%u under %s layout\n",
 364                namelen, name, saved_hash, migrate_hash, stripe_index,
 365                saved_count, migrate_offset, new_layout ? "new" : "old");
 366
 367         return stripe_index;
 368 }
 369
 370 static inline int lmv_name_to_stripe_index(struct lmv_mds_md_v1 *lmv,
 371                                            const char *name, int namelen)
 372 {
 373         if (lmv->lmv_magic == LMV_MAGIC_V1)
 374                 return __lmv_name_to_stripe_index(lmv->lmv_hash_type,
 375                                                   lmv->lmv_stripe_count,
 376                                                   lmv->lmv_migrate_hash,
 377                                                   lmv->lmv_migrate_offset,
 378                                                   name, namelen, true);
 379
 380         if (lmv->lmv_magic == cpu_to_le32(LMV_MAGIC_V1))
 381                 return __lmv_name_to_stripe_index(
 382                                         le32_to_cpu(lmv->lmv_hash_type),
 383                                         le32_to_cpu(lmv->lmv_stripe_count),
 384                                         le32_to_cpu(lmv->lmv_migrate_hash),
 385                                         le32_to_cpu(lmv->lmv_migrate_offset),
 386                                         name, namelen, true);
 387
 388         return -EINVAL;
 389 }
 390
 391 static inline int lmv_name_to_stripe_index_old(struct lmv_mds_md_v1 *lmv,
 392                                                const char *name, int namelen)
 393 {
 394         if (lmv->lmv_magic == LMV_MAGIC_V1 ||
 395             lmv->lmv_magic == LMV_MAGIC_STRIPE)
 396                 return __lmv_name_to_stripe_index(lmv->lmv_hash_type,
 397                                                   lmv->lmv_stripe_count,
 398                                                   lmv->lmv_migrate_hash,
 399                                                   lmv->lmv_migrate_offset,
 400                                                   name, namelen, false);
 401
 402         if (lmv->lmv_magic == cpu_to_le32(LMV_MAGIC_V1) ||
 403             lmv->lmv_magic == cpu_to_le32(LMV_MAGIC_STRIPE))
 404                 return __lmv_name_to_stripe_index(
 405                                         le32_to_cpu(lmv->lmv_hash_type),
 406                                         le32_to_cpu(lmv->lmv_stripe_count),
 407                                         le32_to_cpu(lmv->lmv_migrate_hash),
 408                                         le32_to_cpu(lmv->lmv_migrate_offset),
 409                                         name, namelen, false);
 410
 411         return -EINVAL;
 412 }
 413
 414 static inline bool lmv_user_magic_supported(__u32 lum_magic)
 415 {
 416         return lum_magic == LMV_USER_MAGIC ||
 417                lum_magic == LMV_USER_MAGIC_SPECIFIC ||
 418                lum_magic == LMV_MAGIC_FOREIGN;
 419 }
 420
 421 #define LMV_DEBUG(mask, lmv, msg)                                             \
 422         CDEBUG_LIMIT(mask,                                                    \
 423                "%s LMV: magic=%#x count=%u index=%u hash=%s:%#x version=%u migrate_offset=%u migrate_hash=%s:%x pool=%.*s\n",\
 424                msg, (lmv)->lmv_magic, (lmv)->lmv_stripe_count,                \
 425                (lmv)->lmv_master_mdt_index,                                   \
 426                lmv_is_known_hash_type((lmv)->lmv_hash_type) ?                 \
 427                 mdt_hash_name[(lmv)->lmv_hash_type & LMV_HASH_TYPE_MASK] :    \
 428                 "invalid", (lmv)->lmv_hash_type,                              \
 429                (lmv)->lmv_layout_version, (lmv)->lmv_migrate_offset,          \
 430                lmv_is_known_hash_type((lmv)->lmv_migrate_hash) ?              \
 431                 mdt_hash_name[(lmv)->lmv_migrate_hash & LMV_HASH_TYPE_MASK] : \
 432                 "invalid", (lmv)->lmv_migrate_hash,                           \
 433                LOV_MAXPOOLNAME, lmv->lmv_pool_name)
 434
 435 /* master LMV is sane */
 436 static inline bool lmv_is_sane(const struct lmv_mds_md_v1 *lmv)
 437 {
 438         if (!lmv)
 439                 return false;
 440
 441         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
 442                 goto insane;
 443
 444         if (le32_to_cpu(lmv->lmv_stripe_count) == 0)
 445                 goto insane;
 446
 447         if (!lmv_is_known_hash_type(le32_to_cpu(lmv->lmv_hash_type)))
 448                 goto insane;
 449
 450         return true;
 451 insane:
 452         LMV_DEBUG(D_ERROR, lmv, "unknown layout");
 453         return false;
 454 }
 455
 456 /* LMV can be either master or stripe LMV */
 457 static inline bool lmv_is_sane2(const struct lmv_mds_md_v1 *lmv)
 458 {
 459         if (!lmv)
 460                 return false;
 461
 462         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1 &&
 463             le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_STRIPE)
 464                 goto insane;
 465
 466         if (le32_to_cpu(lmv->lmv_stripe_count) == 0)
 467                 goto insane;
 468
 469         if (!lmv_is_known_hash_type(le32_to_cpu(lmv->lmv_hash_type)))
 470                 goto insane;
 471
 472         return true;
 473 insane:
 474         LMV_DEBUG(D_ERROR, lmv, "unknown layout");
 475         return false;
 476 }
 477
 478 static inline bool lmv_is_splitting(const struct lmv_mds_md_v1 *lmv)
 479 {
 480         if (!lmv_is_sane2(lmv))
 481                 return false;
 482
 483         return lmv_hash_is_splitting(cpu_to_le32(lmv->lmv_hash_type));
 484 }
 485
 486 static inline bool lmv_is_merging(const struct lmv_mds_md_v1 *lmv)
 487 {
 488         if (!lmv_is_sane2(lmv))
 489                 return false;
 490
 491         return lmv_hash_is_merging(cpu_to_le32(lmv->lmv_hash_type));
 492 }
 493
 494 static inline bool lmv_is_migrating(const struct lmv_mds_md_v1 *lmv)
 495 {
 496         if (!lmv_is_sane(lmv))
 497                 return false;
 498
 499         return lmv_hash_is_migrating(cpu_to_le32(lmv->lmv_hash_type));
 500 }
 501
 502 static inline bool lmv_is_restriping(const struct lmv_mds_md_v1 *lmv)
 503 {
 504         if (!lmv_is_sane2(lmv))
 505                 return false;
 506
 507         return lmv_hash_is_splitting(cpu_to_le32(lmv->lmv_hash_type)) ||
 508                lmv_hash_is_merging(cpu_to_le32(lmv->lmv_hash_type));
 509 }
 510
 511 static inline bool lmv_is_layout_changing(const struct lmv_mds_md_v1 *lmv)
 512 {
 513         if (!lmv_is_sane2(lmv))
 514                 return false;
 515
 516         return lmv_hash_is_splitting(cpu_to_le32(lmv->lmv_hash_type)) ||
 517                lmv_hash_is_merging(cpu_to_le32(lmv->lmv_hash_type)) ||
 518                lmv_hash_is_migrating(cpu_to_le32(lmv->lmv_hash_type));
 519 }
 520
 521 static inline bool lmv_is_fixed(const struct lmv_mds_md_v1 *lmv)
 522 {
 523         return cpu_to_le32(lmv->lmv_hash_type) & LMV_HASH_FLAG_FIXED;
 524 }
 525
 526 #endif