lustre/osd-ldiskfs/osd_iam_lvar.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * iam_lvar.c
  32  *
  33  * implementation of iam format for fixed size records, variable sized keys.
  34  *
  35  * Author: Nikita Danilov <nikita@clusterfs.com>
  36  */
  37
  38 #include <linux/types.h>
  39 #include "osd_internal.h"
  40
  41 /*
  42  * Leaf operations.
  43  */
  44
  45 enum {
  46         /* This is duplicated in lustre/utils/create_iam.c */
  47         IAM_LVAR_LEAF_MAGIC = 0x1973
  48 };
  49
  50 /* This is duplicated in lustre/utils/create_iam.c */
  51 struct lvar_leaf_header {
  52         __le16 vlh_magic; /* magic number IAM_LVAR_LEAF_MAGIC */
  53         __le16 vlh_used;  /* used bytes, including header */
  54 };
  55
  56 /*
  57  * Format of leaf entry:
  58  *
  59  * __le16 keysize
  60  *     u8 key[keysize]
  61  *     u8 record[rec_size]
  62  *
  63  * Entries are ordered in key order.
  64  */
  65
  66 /* This is duplicated in lustre/utils/create_iam.c */
  67 typedef u32 lvar_hash_t;
  68
  69 /* This is duplicated in lustre/utils/create_iam.c */
  70 struct lvar_leaf_entry {
  71         __le32 vle_hash;
  72         __le16 vle_keysize;
  73         u8 vle_key[0];
  74 };
  75
  76 #define PDIFF(ptr0, ptr1) (((char *)(ptr0)) - ((char *)(ptr1)))
  77
  78
  79 static inline int blocksize(const struct iam_leaf *leaf)
  80 {
  81         return iam_leaf_container(leaf)->ic_object->i_sb->s_blocksize;
  82 }
  83
  84 static inline const char *kchar(const struct iam_key *key)
  85 {
  86         return (void *)key;
  87 }
  88
  89 static inline struct iam_lentry *lvar_lentry(const struct lvar_leaf_entry *ent)
  90 {
  91         return (struct iam_lentry *)ent;
  92 }
  93
  94 static inline struct lvar_leaf_entry *lentry_lvar(const struct iam_lentry *lent)
  95 {
  96         return (struct lvar_leaf_entry *)lent;
  97 }
  98
  99
 100 static inline int e_keysize(const struct lvar_leaf_entry *ent)
 101 {
 102         return le16_to_cpu(ent->vle_keysize);
 103 }
 104
 105 /* This is duplicated in lustre/utils/create_iam.c */
 106 enum {
 107         LVAR_PAD   = 4,
 108         LVAR_ROUND = LVAR_PAD - 1
 109 };
 110
 111 static inline int getsize(const struct iam_leaf *leaf, int namelen, int recsize)
 112 {
 113         BUILD_BUG_ON((LVAR_PAD & (LVAR_PAD - 1)));
 114
 115         return (offsetof(struct lvar_leaf_entry, vle_key) +
 116                         namelen + recsize + LVAR_ROUND) & ~LVAR_ROUND;
 117 }
 118
 119 static inline int rec_size(const struct iam_rec *rec)
 120 {
 121         return *(const char *)rec;
 122 }
 123
 124 static inline struct iam_rec *e_rec(const struct lvar_leaf_entry *ent)
 125 {
 126         return ((void *)ent) +
 127                 offsetof(struct lvar_leaf_entry, vle_key) + e_keysize(ent);
 128 }
 129
 130 static inline int e_size(const struct iam_leaf *leaf,
 131                          const struct lvar_leaf_entry *ent)
 132 {
 133         return getsize(leaf, e_keysize(ent), rec_size(e_rec(ent)));
 134 }
 135
 136 static inline char *e_char(const struct lvar_leaf_entry *ent)
 137 {
 138         return (char *)&ent->vle_key;
 139 }
 140
 141 static inline struct iam_key *e_key(const struct lvar_leaf_entry *ent)
 142 {
 143         return (struct iam_key *)e_char(ent);
 144 }
 145
 146 static inline lvar_hash_t e_hash(const struct lvar_leaf_entry *ent)
 147 {
 148         return le32_to_cpu(ent->vle_hash);
 149 }
 150
 151 static void e_print(const struct lvar_leaf_entry *ent)
 152 {
 153         CERROR("        %p %8.8x \"%*.*s\"\n", ent, e_hash(ent),
 154                         e_keysize(ent), e_keysize(ent), e_char(ent));
 155 }
 156
 157 static inline struct lvar_leaf_entry *e_next(const struct iam_leaf *leaf,
 158                                              const struct lvar_leaf_entry *ent)
 159 {
 160         return ((void *)ent) + e_size(leaf, ent);
 161 }
 162
 163 #define LVAR_HASH_SANDWICH  (0)
 164 #define LVAR_HASH_TEA       (1)
 165 #define LVAR_HASH_R5        (0)
 166 #define LVAR_HASH_PREFIX    (0)
 167
 168 #ifdef HAVE_LDISKFSFS_DIRHASH_WITH_DIR
 169 #define e_ldiskfsfs_dirhash(dir, name, len, info) \
 170                 ldiskfsfs_dirhash((dir), (name), (len), (info))
 171 #else
 172 #define e_ldiskfsfs_dirhash(dir, name, len, info) \
 173                 ldiskfsfs_dirhash((name), (len), (info))
 174 #endif
 175
 176 static u32 hash_build0(const struct inode *dir, const char *name, int namelen)
 177 {
 178         u32 result;
 179
 180         if (namelen == 0)
 181                 return 0;
 182         if (strncmp(name, ".", 1) == 0 && namelen == 1)
 183                 return 1;
 184         if (strncmp(name, "..", 2) == 0 && namelen == 2)
 185                 return 2;
 186
 187         if (LVAR_HASH_PREFIX) {
 188                 result = 0;
 189                 strncpy((void *)&result,
 190                         name, min_t(int, namelen, sizeof(result)));
 191         } else {
 192                 struct ldiskfs_dx_hash_info hinfo;
 193
 194                 hinfo.hash_version = LDISKFS_DX_HASH_TEA;
 195                 hinfo.seed = NULL;
 196                 e_ldiskfsfs_dirhash(dir, name, namelen, &hinfo);
 197                 result = hinfo.hash;
 198                 if (LVAR_HASH_SANDWICH) {
 199                         u32 result2;
 200
 201                         hinfo.hash_version = LDISKFS_DX_HASH_TEA;
 202                         hinfo.seed = NULL;
 203                         e_ldiskfsfs_dirhash(dir, name, namelen, &hinfo);
 204                         result2 = hinfo.hash;
 205                         result = (0xfc000000 & result2) | (0x03ffffff & result);
 206                 }
 207         }
 208         return result;
 209 }
 210
 211 enum {
 212         HASH_GRAY_AREA = 1024,
 213         HASH_MAX_SIZE  = 0x7fffffffUL
 214 };
 215
 216 static u32 hash_build(const struct inode *dir, const char *name, int namelen)
 217 {
 218         u32 hash;
 219
 220         hash = (hash_build0(dir, name, namelen) << 1) & HASH_MAX_SIZE;
 221         if (hash > HASH_MAX_SIZE - HASH_GRAY_AREA)
 222                 hash &= HASH_GRAY_AREA - 1;
 223         return hash;
 224 }
 225
 226 static inline lvar_hash_t get_hash(const struct inode *dir,
 227                                    const char *name, int namelen)
 228 {
 229         return hash_build(dir, name, namelen);
 230 }
 231
 232 static inline lvar_hash_t iam_get_hash(const struct iam_leaf *leaf,
 233                                        const char *name, int namelen)
 234 {
 235         struct iam_path *iam_path = iam_leaf_path(leaf);
 236
 237         return get_hash(iam_path_obj(iam_path), name, namelen);
 238 }
 239
 240 static inline int e_eq(const struct lvar_leaf_entry *ent,
 241                        const char *name, int namelen)
 242 {
 243         return namelen == e_keysize(ent) && !memcmp(e_char(ent), name, namelen);
 244 }
 245
 246 static inline int e_cmp(const struct iam_leaf *leaf,
 247                         const struct lvar_leaf_entry *ent, lvar_hash_t hash)
 248 {
 249         lvar_hash_t ehash;
 250
 251         ehash = e_hash(ent);
 252         return ehash == hash ? 0 : (ehash < hash ? -1 : 1);
 253 }
 254
 255 static struct lvar_leaf_header *n_head(const struct iam_leaf *l)
 256 {
 257         return (struct lvar_leaf_header *)l->il_bh->b_data;
 258 }
 259
 260 static int h_used(const struct lvar_leaf_header *hdr)
 261 {
 262         return le16_to_cpu(hdr->vlh_used);
 263 }
 264
 265 static void h_used_adj(const struct iam_leaf *leaf,
 266                        struct lvar_leaf_header *hdr, int adj)
 267 {
 268         int used;
 269
 270         used = h_used(hdr) + adj;
 271         assert_corr(sizeof(*hdr) <= used && used <= blocksize(leaf));
 272         hdr->vlh_used = cpu_to_le16(used);
 273 }
 274
 275 static struct lvar_leaf_entry *n_start(const struct iam_leaf *leaf)
 276 {
 277         return (void *)leaf->il_bh->b_data + sizeof(struct lvar_leaf_header);
 278 }
 279
 280 static struct lvar_leaf_entry *n_end(const struct iam_leaf *l)
 281 {
 282         return (void *)l->il_bh->b_data + h_used(n_head(l));
 283 }
 284
 285 static struct lvar_leaf_entry *n_cur(const struct iam_leaf *l)
 286 {
 287         return lentry_lvar(l->il_at);
 288 }
 289
 290 void n_print(const struct iam_leaf *l)
 291 {
 292         struct lvar_leaf_entry *scan;
 293
 294         CERROR("used: %d\n", h_used(n_head(l)));
 295         for (scan = n_start(l); scan < n_end(l); scan = e_next(l, scan))
 296                 e_print(scan);
 297 }
 298
 299 #if LDISKFS_CORRECTNESS_ON
 300 static int n_at_rec(const struct iam_leaf *folio)
 301 {
 302         return n_start(folio) <= lentry_lvar(folio->il_at) &&
 303                 lentry_lvar(folio->il_at) < n_end(folio);
 304 }
 305
 306 #if LDISKFS_INVARIANT_ON
 307 static int n_invariant(const struct iam_leaf *leaf)
 308 {
 309         struct iam_path *path;
 310         struct inode *dir;
 311         struct lvar_leaf_entry *scan;
 312         struct lvar_leaf_entry *end;
 313         lvar_hash_t hash;
 314         lvar_hash_t nexthash;
 315         lvar_hash_t starthash;
 316
 317         end  = n_end(leaf);
 318         hash = 0;
 319         path = leaf->il_path;
 320
 321         if (h_used(n_head(leaf)) > blocksize(leaf))
 322                 return 0;
 323
 324         dir = iam_path_obj(iam_path);
 325         /*
 326          * Delimiting key in the parent index node. Clear least bit to account
 327          * for hash collision marker.
 328          */
 329         starthash = *(lvar_hash_t *)iam_ikey_at(path, path->ip_frame->at) & ~1;
 330         for (scan = n_start(leaf); scan < end; scan = e_next(leaf, scan)) {
 331                 nexthash = e_hash(scan);
 332                 if (nexthash != get_hash(dir, e_char(scan), e_keysize(scan))) {
 333                         BREAKPOINT();
 334                         return 0;
 335                 }
 336                 if (nexthash < hash) {
 337                         BREAKPOINT();
 338                         return 0;
 339                 }
 340                 hash = nexthash;
 341         }
 342         if (scan != end) {
 343                 BREAKPOINT();
 344                 return 0;
 345         }
 346         return 1;
 347 }
 348 /* LDISKFS_INVARIANT_ON */
 349 #endif
 350
 351 /* LDISKFS_CORRECTNESS_ON */
 352 #endif
 353
 354 static struct iam_ikey *lvar_ikey(const struct iam_leaf *l,
 355                                   struct iam_ikey *key)
 356 {
 357         lvar_hash_t *hash;
 358
 359         assert_corr(n_at_rec(l));
 360
 361         hash = (void *)key;
 362         *hash = e_hash(n_cur(l));
 363         return key;
 364 }
 365
 366 static struct iam_key *lvar_key(const struct iam_leaf *l)
 367 {
 368         return e_key(n_cur(l));
 369 }
 370
 371 static int lvar_key_size(const struct iam_leaf *l)
 372 {
 373         return e_keysize(n_cur(l));
 374 }
 375
 376 static void lvar_start(struct iam_leaf *l)
 377 {
 378         l->il_at = lvar_lentry(n_start(l));
 379 }
 380
 381 static int lvar_init(struct iam_leaf *l)
 382 {
 383         int result;
 384         int used;
 385         struct lvar_leaf_header *head;
 386
 387         assert_corr(l->il_bh != NULL);
 388
 389         head = n_head(l);
 390         used = h_used(head);
 391         if (le16_to_cpu(head->vlh_magic) == IAM_LVAR_LEAF_MAGIC &&
 392                         used <= blocksize(l)) {
 393                 l->il_at = l->il_entries = lvar_lentry(n_start(l));
 394                 result = 0;
 395         } else {
 396                 struct inode *obj;
 397
 398                 obj = iam_leaf_container(l)->ic_object;
 399                 CERROR(
 400                 "Bad magic in node %llu (#%lu): %#x != %#x or wrong used: %d\n",
 401                 (unsigned long long)l->il_bh->b_blocknr, obj->i_ino,
 402                 le16_to_cpu(head->vlh_magic), IAM_LVAR_LEAF_MAGIC,
 403                 used);
 404                 result = -EIO;
 405         }
 406         return result;
 407 }
 408
 409 static void lvar_fini(struct iam_leaf *l)
 410 {
 411         l->il_entries = l->il_at = NULL;
 412 }
 413
 414 static struct iam_rec *lvar_rec(const struct iam_leaf *l)
 415 {
 416         assert_corr(n_at_rec(l));
 417         return e_rec(n_cur(l));
 418 }
 419
 420 static void lvar_next(struct iam_leaf *l)
 421 {
 422         assert_corr(n_at_rec(l));
 423         assert_corr(iam_leaf_is_locked(l));
 424         l->il_at = lvar_lentry(e_next(l, n_cur(l)));
 425 }
 426
 427 static int lvar_lookup(struct iam_leaf *leaf, const struct iam_key *k)
 428 {
 429         struct lvar_leaf_entry *found;
 430         struct lvar_leaf_entry *scan;
 431         struct lvar_leaf_entry *end;
 432         int result;
 433         const char *name;
 434         int namelen;
 435         int found_equal;
 436         lvar_hash_t hash;
 437         int last;
 438
 439         assert_inv(n_invariant(leaf));
 440         end = n_end(leaf);
 441
 442         name = kchar(k);
 443         namelen = strlen(name);
 444         hash = iam_get_hash(leaf, name, namelen);
 445         found = NULL;
 446         found_equal = 0;
 447         last = 1;
 448
 449         for (scan = n_start(leaf); scan < end; scan = e_next(leaf, scan)) {
 450                 lvar_hash_t scan_hash;
 451
 452                 scan_hash = e_hash(scan);
 453                 if (scan_hash < hash)
 454                         found = scan;
 455                 else if (scan_hash == hash) {
 456                         if (e_eq(scan, name, namelen)) {
 457                                 /*
 458                                  * perfect match
 459                                  */
 460                                 leaf->il_at = lvar_lentry(scan);
 461                                 return IAM_LOOKUP_EXACT;
 462                         } else if (!found_equal) {
 463                                 found = scan;
 464                                 found_equal = 1;
 465                         }
 466                 } else {
 467                         last = 0;
 468                         break;
 469                 }
 470         }
 471         if (found == NULL) {
 472                 /*
 473                  * @k is less than all hashes in the leaf.
 474                  */
 475                 lvar_start(leaf);
 476                 result = IAM_LOOKUP_BEFORE;
 477         } else {
 478                 leaf->il_at = lvar_lentry(found);
 479                 result = IAM_LOOKUP_OK;
 480                 assert_corr(n_at_rec(leaf));
 481         }
 482         if (last)
 483                 result |= IAM_LOOKUP_LAST;
 484         assert_inv(n_invariant(leaf));
 485
 486         return result;
 487 }
 488
 489 static int lvar_ilookup(struct iam_leaf *leaf, const struct iam_ikey *ik)
 490 {
 491         struct lvar_leaf_entry *scan;
 492         struct lvar_leaf_entry *end;
 493         lvar_hash_t hash;
 494
 495         assert_inv(n_invariant(leaf));
 496         end  = n_end(leaf);
 497         hash = *(const lvar_hash_t *)ik;
 498
 499         lvar_start(leaf);
 500         for (scan = n_start(leaf); scan < end; scan = e_next(leaf, scan)) {
 501                 lvar_hash_t scan_hash;
 502
 503                 scan_hash = e_hash(scan);
 504                 if (scan_hash > hash)
 505                         return scan == n_start(leaf) ?
 506                                 IAM_LOOKUP_BEFORE : IAM_LOOKUP_OK;
 507                 leaf->il_at = lvar_lentry(scan);
 508                 if (scan_hash == hash)
 509                         return IAM_LOOKUP_EXACT;
 510         }
 511         assert_inv(n_invariant(leaf));
 512         /*
 513          * @ik is greater than any key in the node. Return last record in the
 514          * node.
 515          */
 516         return IAM_LOOKUP_OK;
 517 }
 518
 519 static void __lvar_key_set(struct iam_leaf *l, const struct iam_key *k)
 520 {
 521         memcpy(e_key(n_cur(l)), k, e_keysize(n_cur(l)));
 522 }
 523
 524 static void lvar_key_set(struct iam_leaf *l, const struct iam_key *k)
 525 {
 526         assert_corr(n_at_rec(l));
 527         assert_corr(strlen(kchar(k)) == e_keysize(n_cur(l)));
 528         assert_corr(iam_leaf_is_locked(l));
 529         __lvar_key_set(l, k);
 530         assert_inv(n_invariant(l));
 531 }
 532
 533 static int lvar_key_cmp(const struct iam_leaf *l, const struct iam_key *k)
 534 {
 535         lvar_hash_t hash;
 536         const char *name;
 537
 538         name = kchar(k);
 539
 540         hash = iam_get_hash(l, name, strlen(name));
 541         return e_cmp(l, n_cur(l), hash);
 542 }
 543
 544 static int lvar_key_eq(const struct iam_leaf *l, const struct iam_key *k)
 545 {
 546         const char *name;
 547
 548         name = kchar(k);
 549         return e_eq(n_cur(l), name, strlen(name));
 550 }
 551
 552 static void __lvar_rec_set(struct iam_leaf *l, const struct iam_rec *r)
 553 {
 554         memcpy(e_rec(n_cur(l)), r, rec_size(r));
 555 }
 556
 557 static void lvar_rec_set(struct iam_leaf *l, const struct iam_rec *r)
 558 {
 559         assert_corr(n_at_rec(l));
 560         assert_corr(iam_leaf_is_locked(l));
 561         __lvar_rec_set(l, r);
 562         assert_inv(n_invariant(l));
 563 }
 564
 565 static int lvar_rec_eq(const struct iam_leaf *l, const struct iam_rec *r)
 566 {
 567         struct iam_rec *rec = e_rec(n_cur(l));
 568
 569         if (rec_size(rec) != rec_size(r))
 570                 return 0;
 571         return !memcmp(rec, r, rec_size(r));
 572 }
 573
 574 static void lvar_rec_get(const struct iam_leaf *l, struct iam_rec *r)
 575 {
 576         struct iam_rec *rec;
 577
 578         rec = e_rec(n_cur(l));
 579         assert_corr(n_at_rec(l));
 580         assert_corr(iam_leaf_is_locked(l));
 581         memcpy(r, rec, rec_size(rec));
 582         assert_inv(n_invariant(l));
 583 }
 584
 585 static int lvar_can_add(const struct iam_leaf *l,
 586                         const struct iam_key *k, const struct iam_rec *r)
 587 {
 588         assert_corr(iam_leaf_is_locked(l));
 589         return h_used(n_head(l)) +
 590                 getsize(l, strlen(kchar(k)), rec_size(r)) <= blocksize(l);
 591 }
 592
 593 static int lvar_at_end(const struct iam_leaf *folio)
 594 {
 595         assert_corr(iam_leaf_is_locked(folio));
 596         return n_cur(folio) == n_end(folio);
 597 }
 598
 599 static void lvar_rec_add(struct iam_leaf *leaf,
 600                          const struct iam_key *k, const struct iam_rec *r)
 601 {
 602         const char *key;
 603         int ksize;
 604         int shift;
 605         void *end;
 606         void *start;
 607         ptrdiff_t diff;
 608
 609         assert_corr(lvar_can_add(leaf, k, r));
 610         assert_inv(n_invariant(leaf));
 611         assert_corr(iam_leaf_is_locked(leaf));
 612
 613         key   = kchar(k);
 614         ksize = strlen(key);
 615         shift = getsize(leaf, ksize, rec_size(r));
 616
 617         if (!lvar_at_end(leaf)) {
 618                 assert_corr(n_cur(leaf) < n_end(leaf));
 619                 end = n_end(leaf);
 620                 if (lvar_key_cmp(leaf, k) <= 0)
 621                         lvar_next(leaf);
 622                 else
 623                         /*
 624                          * Another exceptional case: insertion with the key
 625                          * less than least key in the leaf.
 626                          */
 627                         assert_corr(leaf->il_at == leaf->il_entries);
 628
 629                 start = leaf->il_at;
 630                 diff  = PDIFF(end, start);
 631                 assert_corr(diff >= 0);
 632                 memmove(start + shift, start, diff);
 633         }
 634         h_used_adj(leaf, n_head(leaf), shift);
 635         n_cur(leaf)->vle_keysize = cpu_to_le16(ksize);
 636         n_cur(leaf)->vle_hash = cpu_to_le32(iam_get_hash(leaf, key, ksize));
 637         __lvar_key_set(leaf, k);
 638         __lvar_rec_set(leaf, r);
 639         assert_corr(n_at_rec(leaf));
 640         assert_inv(n_invariant(leaf));
 641 }
 642
 643 static void lvar_rec_del(struct iam_leaf *leaf, int shift)
 644 {
 645         void *next;
 646         void *end;
 647         int nob;
 648
 649         assert_corr(n_at_rec(leaf));
 650         assert_inv(n_invariant(leaf));
 651         assert_corr(iam_leaf_is_locked(leaf));
 652
 653         end  = n_end(leaf);
 654         next = e_next(leaf, n_cur(leaf));
 655         nob  = e_size(leaf, n_cur(leaf));
 656         memmove(leaf->il_at, next, end - next);
 657         h_used_adj(leaf, n_head(leaf), -nob);
 658         assert_inv(n_invariant(leaf));
 659 }
 660
 661 static void lvar_init_new(struct iam_container *c, struct buffer_head *bh)
 662 {
 663         struct lvar_leaf_header *hdr;
 664
 665         hdr = (struct lvar_leaf_header *)bh->b_data;
 666         hdr->vlh_magic = cpu_to_le16(IAM_LVAR_LEAF_MAGIC);
 667         hdr->vlh_used  = sizeof(*hdr);
 668 }
 669
 670 static struct lvar_leaf_entry *find_pivot(const struct iam_leaf *leaf,
 671                                           struct lvar_leaf_entry **prev)
 672 {
 673         void *scan;
 674         void *start;
 675         int threshold;
 676
 677         *prev = NULL;
 678         threshold = blocksize(leaf) / 2;
 679         for (scan = start = n_start(leaf); scan - start <= threshold;
 680                         *prev = scan, scan = e_next(leaf, scan)) {
 681                 ;
 682         }
 683         return scan;
 684 }
 685
 686 static void lvar_split(struct iam_leaf *leaf, struct buffer_head **bh,
 687                        iam_ptr_t new_blknr)
 688 {
 689         struct lvar_leaf_entry *first_to_move;
 690         struct lvar_leaf_entry *last_to_stay;
 691         struct iam_path *path;
 692         struct lvar_leaf_header *hdr;
 693         struct buffer_head *new_leaf;
 694         ptrdiff_t tomove;
 695         lvar_hash_t hash;
 696
 697         assert_inv(n_invariant(leaf));
 698         assert_corr(iam_leaf_is_locked(leaf));
 699
 700         new_leaf = *bh;
 701         path = iam_leaf_path(leaf);
 702
 703         hdr = (void *)new_leaf->b_data;
 704
 705         first_to_move = find_pivot(leaf, &last_to_stay);
 706         assert_corr(last_to_stay != NULL);
 707         assert_corr(e_next(leaf, last_to_stay) == first_to_move);
 708
 709         hash = e_hash(first_to_move);
 710         if (hash == e_hash(last_to_stay))
 711                 /*
 712                  * Duplicate hash.
 713                  */
 714                 hash |= 1;
 715
 716         tomove = PDIFF(n_end(leaf), first_to_move);
 717         memmove(hdr + 1, first_to_move, tomove);
 718
 719         h_used_adj(leaf, hdr, tomove);
 720         h_used_adj(leaf, n_head(leaf), -tomove);
 721
 722         assert_corr(n_end(leaf) == first_to_move);
 723
 724         if (n_cur(leaf) >= first_to_move) {
 725                 /*
 726                  * insertion point moves into new leaf.
 727                  */
 728                 ptrdiff_t shift;
 729
 730                 shift = PDIFF(leaf->il_at, first_to_move);
 731                 *bh = leaf->il_bh;
 732                 leaf->il_bh = new_leaf;
 733                 leaf->il_curidx = new_blknr;
 734
 735                 assert_corr(iam_leaf_is_locked(leaf));
 736                 lvar_init(leaf);
 737                 /*
 738                  * init cannot fail, as node was just initialized.
 739                  */
 740                 assert_corr(result == 0);
 741                 leaf->il_at = ((void *)leaf->il_at) + shift;
 742         }
 743         /*
 744          * Insert pointer to the new node (together with the least key in
 745          * the node) into index node.
 746          */
 747         iam_insert_key_lock(path, path->ip_frame, (struct iam_ikey *)&hash,
 748                             new_blknr);
 749         assert_corr(n_cur(leaf) < n_end(leaf));
 750         assert_inv(n_invariant(leaf));
 751 }
 752
 753 static int lvar_leaf_empty(struct iam_leaf *leaf)
 754 {
 755         return h_used(n_head(leaf)) == sizeof(struct lvar_leaf_header);
 756 }
 757
 758 static const struct iam_leaf_operations lvar_leaf_ops = {
 759         .init           = lvar_init,
 760         .init_new       = lvar_init_new,
 761         .fini           = lvar_fini,
 762         .start          = lvar_start,
 763         .next           = lvar_next,
 764         .key            = lvar_key,
 765         .ikey           = lvar_ikey,
 766         .rec            = lvar_rec,
 767         .key_set        = lvar_key_set,
 768         .key_cmp        = lvar_key_cmp,
 769         .key_eq         = lvar_key_eq,
 770         .key_size       = lvar_key_size,
 771         .rec_set        = lvar_rec_set,
 772         .rec_eq         = lvar_rec_eq,
 773         .rec_get        = lvar_rec_get,
 774         .lookup         = lvar_lookup,
 775         .ilookup        = lvar_ilookup,
 776         .at_end         = lvar_at_end,
 777         .rec_add        = lvar_rec_add,
 778         .rec_del        = lvar_rec_del,
 779         .can_add        = lvar_can_add,
 780         .split          = lvar_split,
 781         .leaf_empty     = lvar_leaf_empty,
 782 };
 783
 784 /*
 785  * Index operations.
 786  */
 787
 788 enum {
 789         /* This is duplicated in lustre/utils/create_iam.c */
 790         /* egrep -i '^o?x?[olabcdef]*$' /usr/share/dict/words */
 791         IAM_LVAR_ROOT_MAGIC = 0xb01dface
 792 };
 793
 794 /* This is duplicated in lustre/utils/create_iam.c */
 795 struct lvar_root {
 796         __le32 vr_magic;
 797         __le16 vr_recsize;
 798         __le16 vr_ptrsize;
 799         u8 vr_indirect_levels;
 800         u8 vr_padding0;
 801         __le16 vr_padding1;
 802 };
 803
 804 static u32 lvar_root_ptr(struct iam_container *c)
 805 {
 806         return 0;
 807 }
 808
 809 static int lvar_node_init(struct iam_container *c, struct buffer_head *bh,
 810                           int root)
 811 {
 812         return 0;
 813 }
 814
 815 static struct iam_entry *lvar_root_inc(struct iam_container *c,
 816                                        struct iam_path *path,
 817                                        struct iam_frame *frame)
 818 {
 819         struct lvar_root *root;
 820         struct iam_entry *entries;
 821
 822         assert_corr(iam_frame_is_locked(path, frame));
 823         entries = frame->entries;
 824
 825         dx_set_count(entries, 2);
 826         assert_corr(dx_get_limit(entries) == dx_root_limit(path));
 827
 828         root = (void *)frame->bh->b_data;
 829         assert_corr(le64_to_cpu(root->vr_magic) == IAM_LVAR_ROOT_MAGIC);
 830         root->vr_indirect_levels++;
 831         frame->at = entries = iam_entry_shift(path, entries, 1);
 832         memset(iam_ikey_at(path, entries), 0,
 833                iam_path_descr(path)->id_ikey_size);
 834         return entries;
 835 }
 836
 837 static int lvar_node_check(struct iam_path *path, struct iam_frame *frame)
 838 {
 839         unsigned int count;
 840         unsigned int limit;
 841         unsigned int limit_correct;
 842         struct iam_entry *entries;
 843
 844         entries = dx_node_get_entries(path, frame);
 845
 846         if (frame == path->ip_frames) {
 847                 struct lvar_root *root;
 848
 849                 root = (void *)frame->bh->b_data;
 850                 if (le32_to_cpu(root->vr_magic) != IAM_LVAR_ROOT_MAGIC)
 851                         return -EIO;
 852                 limit_correct = dx_root_limit(path);
 853         } else
 854                 limit_correct = dx_node_limit(path);
 855         count = dx_get_count(entries);
 856         limit = dx_get_limit(entries);
 857         if (count > limit)
 858                 return -EIO;
 859         if (limit != limit_correct)
 860                 return -EIO;
 861         return 0;
 862 }
 863
 864 static int lvar_node_load(struct iam_path *path, struct iam_frame *frame)
 865 {
 866         struct iam_entry *entries;
 867         void *data;
 868
 869         entries = dx_node_get_entries(path, frame);
 870         data = frame->bh->b_data;
 871
 872         if (frame == path->ip_frames) {
 873                 struct lvar_root *root;
 874                 const char *name;
 875
 876                 root = data;
 877                 name = kchar(path->ip_key_target);
 878                 path->ip_indirect = root->vr_indirect_levels;
 879                 if (path->ip_ikey_target == NULL) {
 880                         path->ip_ikey_target = iam_path_ikey(path, 4);
 881                         *(lvar_hash_t *)path->ip_ikey_target =
 882                                 get_hash(iam_path_obj(path), name,
 883                                          strlen(name));
 884                 }
 885         }
 886         frame->entries = frame->at = entries;
 887         return 0;
 888 }
 889
 890 static int lvar_ikeycmp(const struct iam_container *c,
 891                         const struct iam_ikey *k1, const struct iam_ikey *k2)
 892 {
 893         lvar_hash_t p1 = le32_to_cpu(*(lvar_hash_t *)k1);
 894         lvar_hash_t p2 = le32_to_cpu(*(lvar_hash_t *)k2);
 895
 896         return p1 > p2 ? 1 : (p1 < p2 ? -1 : 0);
 897 }
 898
 899 static struct iam_path_descr *lvar_ipd_alloc(const struct iam_container *c,
 900                                              void *area)
 901 {
 902         return iam_ipd_alloc(area, c->ic_descr->id_ikey_size);
 903 }
 904
 905 static void lvar_root(void *buf,
 906                       int blocksize, int keysize, int ptrsize, int recsize)
 907 {
 908         struct lvar_root *root;
 909         struct dx_countlimit *limit;
 910         void *entry;
 911         int isize;
 912
 913         isize = sizeof(lvar_hash_t) + ptrsize;
 914         root = buf;
 915         *root = (typeof(*root)) {
 916                 .vr_magic            = cpu_to_le32(IAM_LVAR_ROOT_MAGIC),
 917                 .vr_recsize          = cpu_to_le16(recsize),
 918                 .vr_ptrsize          = cpu_to_le16(ptrsize),
 919                 .vr_indirect_levels  = 0
 920         };
 921
 922         limit = (void *)(root + 1);
 923         *limit = (typeof(*limit)){
 924                 /*
 925                  * limit itself + one pointer to the leaf.
 926                  */
 927                 .count = cpu_to_le16(2),
 928                 .limit = iam_root_limit(sizeof(struct lvar_root), blocksize,
 929                                         sizeof(lvar_hash_t) + ptrsize)
 930         };
 931
 932         /* To guarantee that the padding "keysize + ptrsize"
 933          * covers the "dx_countlimit" and the "idle_blocks". */
 934         LASSERT((keysize + ptrsize) >=
 935                 (sizeof(struct dx_countlimit) + sizeof(u32)));
 936
 937         entry = (void *)(limit + 1);
 938         /* Put "idle_blocks" just after the limit. There was padding after
 939          * the limit, the "idle_blocks" re-uses part of the padding, so no
 940          * compatibility issues with old layout.
 941          */
 942         *(u32 *)entry = 0;
 943
 944         /*
 945          * Skip over @limit.
 946          */
 947         entry = (void *)(root + 1) + isize;
 948
 949         /*
 950          * Entry format is <key> followed by <ptr>. In the minimal tree
 951          * consisting of a root and single node, <key> is a minimal possible
 952          * key.
 953          */
 954         *(lvar_hash_t *)entry = 0;
 955         entry += sizeof(lvar_hash_t);
 956         /* now @entry points to <ptr> */
 957         if (ptrsize == 4)
 958                 *(u_int32_t *)entry = cpu_to_le32(1);
 959         else
 960                 *(u_int64_t *)entry = cpu_to_le64(1);
 961 }
 962
 963 static int lvar_esize(int namelen, int recsize)
 964 {
 965         return (offsetof(struct lvar_leaf_entry, vle_key) +
 966                         namelen + recsize + LVAR_ROUND) & ~LVAR_ROUND;
 967 }
 968
 969 static void lvar_leaf(void *buf,
 970                       int blocksize, int keysize, int ptrsize, int recsize)
 971 {
 972         struct lvar_leaf_header *head;
 973         struct lvar_leaf_entry *entry;
 974
 975         /* form leaf */
 976         head = buf;
 977         *head = (typeof(*head)) {
 978                 .vlh_magic = cpu_to_le16(IAM_LVAR_LEAF_MAGIC),
 979                 .vlh_used  = cpu_to_le16(sizeof(*head) + lvar_esize(0, recsize))
 980         };
 981         entry = (void *)(head + 1);
 982         *entry = (typeof(*entry)) {
 983                 .vle_hash    = 0,
 984                 .vle_keysize = 0
 985         };
 986         memset(e_rec(entry), 0, recsize);
 987         *(char *)e_rec(entry) = recsize;
 988 }
 989
 990 int iam_lvar_create(struct inode *obj,
 991                     int keysize, int ptrsize, int recsize, handle_t *handle)
 992 {
 993         struct buffer_head *root_node;
 994         struct buffer_head *leaf_node;
 995         struct super_block *sb;
 996
 997         u32 blknr;
 998         int result = 0;
 999         unsigned long bsize;
1000
1001         assert_corr(obj->i_size == 0);
1002
1003         sb = obj->i_sb;
1004         bsize = sb->s_blocksize;
1005         root_node = osd_ldiskfs_append(handle, obj, &blknr);
1006         if (IS_ERR(root_node))
1007                 GOTO(out, result = PTR_ERR(root_node));
1008
1009         leaf_node = osd_ldiskfs_append(handle, obj, &blknr);
1010         if (IS_ERR(leaf_node))
1011                 GOTO(out_root, result = PTR_ERR(leaf_node));
1012
1013         lvar_root(root_node->b_data, bsize, keysize, ptrsize, recsize);
1014         lvar_leaf(leaf_node->b_data, bsize, keysize, ptrsize, recsize);
1015         ldiskfs_mark_inode_dirty(handle, obj);
1016         result = ldiskfs_handle_dirty_metadata(handle, NULL, root_node);
1017         if (result == 0)
1018                 result = ldiskfs_handle_dirty_metadata(handle, NULL, leaf_node);
1019         if (result != 0)
1020                 ldiskfs_std_error(sb, result);
1021
1022         brelse(leaf_node);
1023
1024         GOTO(out_root, result);
1025
1026 out_root:
1027         brelse(root_node);
1028 out:
1029         return result;
1030 }
1031
1032 static const struct iam_operations lvar_ops = {
1033         .id_root_ptr    = lvar_root_ptr,
1034         .id_node_read   = iam_node_read,
1035         .id_node_init   = lvar_node_init,
1036         .id_node_check  = lvar_node_check,
1037         .id_node_load   = lvar_node_load,
1038         .id_ikeycmp     = lvar_ikeycmp,
1039         .id_root_inc    = lvar_root_inc,
1040         .id_ipd_alloc   = lvar_ipd_alloc,
1041         .id_ipd_free    = iam_ipd_free,
1042         .id_name        = "lvar"
1043 };
1044
1045 int iam_lvar_guess(struct iam_container *c)
1046 {
1047         int result;
1048         struct buffer_head *bh;
1049         const struct lvar_root *root;
1050
1051         assert_corr(c->ic_object != NULL);
1052
1053         result = iam_node_read(c, lvar_root_ptr(c), NULL, &bh);
1054         if (result == 0) {
1055                 root = (void *)bh->b_data;
1056
1057                 if (le32_to_cpu(root->vr_magic) == IAM_LVAR_ROOT_MAGIC) {
1058                         struct iam_descr *descr;
1059
1060                         descr = c->ic_descr;
1061                         descr->id_key_size  = LDISKFS_NAME_LEN;
1062                         descr->id_ikey_size = sizeof(lvar_hash_t);
1063                         descr->id_rec_size  = le16_to_cpu(root->vr_recsize);
1064                         descr->id_ptr_size  = le16_to_cpu(root->vr_ptrsize);
1065                         descr->id_root_gap  = sizeof(*root);
1066                         descr->id_node_gap  = 0;
1067                         descr->id_ops       = &lvar_ops;
1068                         descr->id_leaf_ops  = &lvar_leaf_ops;
1069                         c->ic_root_bh = bh;
1070                 } else {
1071                         result = -EBADF;
1072                         brelse(bh);
1073                 }
1074         }
1075         return result;
1076 }