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 [sun.com URL with a
  18  * copy of GPLv2].
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * iam_lvar.c
  37  *
  38  * implementation of iam format for fixed size records, variable sized keys.
  39  *
  40  * Author: Nikita Danilov <nikita@clusterfs.com>
  41  */
  42
  43 #include <linux/types.h>
  44 #include "osd_internal.h"
  45
  46 /*
  47  * Leaf operations.
  48  */
  49
  50 enum {
  51         IAM_LVAR_LEAF_MAGIC = 0x1973 /* This is duplicated in
  52                                       * lustre/utils/create_iam.c */
  53 };
  54
  55 /* This is duplicated in lustre/utils/create_iam.c */
  56 struct lvar_leaf_header {
  57         __le16 vlh_magic; /* magic number IAM_LVAR_LEAF_MAGIC */
  58         __le16 vlh_used;  /* used bytes, including header */
  59 };
  60
  61 /*
  62  * Format of leaf entry:
  63  *
  64  * __le16 keysize
  65  *     u8 key[keysize]
  66  *     u8 record[rec_size]
  67  *
  68  * Entries are ordered in key order.
  69  */
  70
  71 /* This is duplicated in lustre/utils/create_iam.c */
  72 typedef __u32 lvar_hash_t;
  73
  74 /* This is duplicated in lustre/utils/create_iam.c */
  75 struct lvar_leaf_entry {
  76         __le32 vle_hash;
  77         __le16 vle_keysize;
  78         u8     vle_key[0];
  79 };
  80
  81 #define PDIFF(ptr0, ptr1) (((char *)(ptr0)) - ((char *)(ptr1)))
  82
  83
  84 static inline int blocksize(const struct iam_leaf *leaf)
  85 {
  86         return iam_leaf_container(leaf)->ic_object->i_sb->s_blocksize;
  87 }
  88
  89 static inline const char *kchar(const struct iam_key *key)
  90 {
  91         return (void *)key;
  92 }
  93
  94 static inline struct iam_lentry *lvar_lentry(const struct lvar_leaf_entry *ent)
  95 {
  96         return (struct iam_lentry *)ent;
  97 }
  98
  99 static inline struct lvar_leaf_entry *lentry_lvar(const struct iam_lentry *lent)
 100 {
 101         return (struct lvar_leaf_entry *)lent;
 102 }
 103
 104
 105 static inline int e_keysize(const struct lvar_leaf_entry *ent)
 106 {
 107         return le16_to_cpu(ent->vle_keysize);
 108 }
 109
 110 /* This is duplicated in lustre/utils/create_iam.c */
 111 enum {
 112         LVAR_PAD   = 4,
 113         LVAR_ROUND = LVAR_PAD - 1
 114 };
 115
 116 static inline int getsize(const struct iam_leaf *leaf, int namelen, int recsize)
 117 {
 118         CLASSERT(!(LVAR_PAD & (LVAR_PAD - 1)));
 119
 120         return (offsetof(struct lvar_leaf_entry, vle_key) +
 121                 namelen + recsize + LVAR_ROUND) & ~LVAR_ROUND;
 122 }
 123
 124 static inline int rec_size(const struct iam_rec *rec)
 125 {
 126         return *(const char *)rec;
 127 }
 128
 129 static inline struct iam_rec *e_rec(const struct lvar_leaf_entry *ent)
 130 {
 131         return ((void *)ent) +
 132                 offsetof(struct lvar_leaf_entry, vle_key) + e_keysize(ent);
 133 }
 134
 135 static inline int e_size(const struct iam_leaf *leaf,
 136                          const struct lvar_leaf_entry *ent)
 137 {
 138         return getsize(leaf, e_keysize(ent), rec_size(e_rec(ent)));
 139 }
 140
 141 static inline char *e_char(const struct lvar_leaf_entry *ent)
 142 {
 143         return (char *)&ent->vle_key;
 144 }
 145
 146 static inline struct iam_key *e_key(const struct lvar_leaf_entry *ent)
 147 {
 148         return (struct iam_key *)e_char(ent);
 149 }
 150
 151 static inline lvar_hash_t e_hash(const struct lvar_leaf_entry *ent)
 152 {
 153         return le32_to_cpu(ent->vle_hash);
 154 }
 155
 156 static void e_print(const struct lvar_leaf_entry *ent)
 157 {
 158         printk("        %p %8.8x \"%*.*s\"\n", ent, e_hash(ent),
 159                e_keysize(ent), e_keysize(ent), e_char(ent));
 160 }
 161
 162 static inline struct lvar_leaf_entry *e_next(const struct iam_leaf *leaf,
 163                                              const struct lvar_leaf_entry *ent)
 164 {
 165         return ((void *)ent) + e_size(leaf, ent);
 166 }
 167
 168 #define LVAR_HASH_SANDWICH  (0)
 169 #define LVAR_HASH_TEA       (1)
 170 #define LVAR_HASH_R5        (0)
 171 #define LVAR_HASH_PREFIX    (0)
 172
 173 static __u32 hash_build0(const char *name, int namelen)
 174 {
 175         __u32 result;
 176
 177         if (namelen == 0)
 178                 return 0;
 179         if (strncmp(name, ".", 1) == 0 && namelen == 1)
 180                 return 1;
 181         if (strncmp(name, "..", 2) == 0 && namelen == 2)
 182                 return 2;
 183
 184         if (LVAR_HASH_PREFIX) {
 185                 result = 0;
 186                 strncpy((void *)&result,
 187                         name, min(namelen, (int)sizeof result));
 188         } else {
 189                 struct ldiskfs_dx_hash_info hinfo;
 190
 191                 hinfo.hash_version = LDISKFS_DX_HASH_TEA;
 192                 hinfo.seed = NULL;
 193                 ldiskfsfs_dirhash(name, namelen, &hinfo);
 194                 result = hinfo.hash;
 195                 if (LVAR_HASH_SANDWICH) {
 196                         __u32 result2;
 197
 198                         hinfo.hash_version = LDISKFS_DX_HASH_TEA;
 199                         hinfo.seed = NULL;
 200                         ldiskfsfs_dirhash(name, namelen, &hinfo);
 201                         result2 = hinfo.hash;
 202                         result = (0xfc000000 & result2) | (0x03ffffff & result);
 203                 }
 204         }
 205         return result;
 206 }
 207
 208 enum {
 209         HASH_GRAY_AREA = 1024,
 210         HASH_MAX_SIZE  = 0x7fffffffUL
 211 };
 212
 213 static __u32 hash_build(const char *name, int namelen)
 214 {
 215         __u32 hash;
 216
 217         hash = (hash_build0(name, namelen) << 1) & HASH_MAX_SIZE;
 218         if (hash > HASH_MAX_SIZE - HASH_GRAY_AREA)
 219                 hash &= HASH_GRAY_AREA - 1;
 220         return hash;
 221 }
 222
 223 static inline lvar_hash_t get_hash(const struct iam_container *bag,
 224                                    const char *name, int namelen)
 225 {
 226         return hash_build(name, namelen);
 227 }
 228
 229 static inline int e_eq(const struct lvar_leaf_entry *ent,
 230                        const char *name, int namelen)
 231 {
 232         return namelen == e_keysize(ent) && !memcmp(e_char(ent), name, namelen);
 233 }
 234
 235 static inline int e_cmp(const struct iam_leaf *leaf,
 236                         const struct lvar_leaf_entry *ent, lvar_hash_t hash)
 237 {
 238         lvar_hash_t ehash;
 239
 240         ehash = e_hash(ent);
 241         return ehash == hash ? 0 : (ehash < hash ? -1 : +1);
 242 }
 243
 244 static struct lvar_leaf_header *n_head(const struct iam_leaf *l)
 245 {
 246         return (struct lvar_leaf_header *)l->il_bh->b_data;
 247 }
 248
 249 static int h_used(const struct lvar_leaf_header *hdr)
 250 {
 251         return le16_to_cpu(hdr->vlh_used);
 252 }
 253
 254 static void h_used_adj(const struct iam_leaf *leaf,
 255                        struct lvar_leaf_header *hdr, int adj)
 256 {
 257         int used;
 258
 259         used = h_used(hdr) + adj;
 260         assert_corr(sizeof *hdr <= used && used <= blocksize(leaf));
 261         hdr->vlh_used = cpu_to_le16(used);
 262 }
 263
 264 static struct lvar_leaf_entry *n_start(const struct iam_leaf *leaf)
 265 {
 266         return (void *)leaf->il_bh->b_data + sizeof(struct lvar_leaf_header);
 267 }
 268
 269 static struct lvar_leaf_entry *n_end(const struct iam_leaf *l)
 270 {
 271         return (void *)l->il_bh->b_data + h_used(n_head(l));
 272 }
 273
 274 static struct lvar_leaf_entry *n_cur(const struct iam_leaf *l)
 275 {
 276         return lentry_lvar(l->il_at);
 277 }
 278
 279 void n_print(const struct iam_leaf *l)
 280 {
 281         struct lvar_leaf_entry *scan;
 282
 283         printk(KERN_EMERG "used: %d\n", h_used(n_head(l)));
 284         for (scan = n_start(l); scan < n_end(l); scan = e_next(l, scan))
 285                 e_print(scan);
 286 }
 287
 288 #if LDISKFS_CORRECTNESS_ON
 289 static int n_at_rec(const struct iam_leaf *folio)
 290 {
 291         return
 292                 n_start(folio) <= lentry_lvar(folio->il_at) &&
 293                 lentry_lvar(folio->il_at) < n_end(folio);
 294 }
 295
 296 #if LDISKFS_INVARIANT_ON
 297 static int n_invariant(const struct iam_leaf *leaf)
 298 {
 299         struct iam_path        *path;
 300         struct lvar_leaf_entry *scan;
 301         struct lvar_leaf_entry *end;
 302         lvar_hash_t             hash;
 303         lvar_hash_t             nexthash;
 304         lvar_hash_t             starthash;
 305
 306         end  = n_end(leaf);
 307         hash = 0;
 308         path = leaf->il_path;
 309
 310         if (h_used(n_head(leaf)) > blocksize(leaf))
 311                 return 0;
 312
 313         /*
 314          * Delimiting key in the parent index node. Clear least bit to account
 315          * for hash collision marker.
 316          */
 317         starthash = *(lvar_hash_t *)iam_ikey_at(path, path->ip_frame->at) & ~1;
 318         for (scan = n_start(leaf); scan < end; scan = e_next(leaf, scan)) {
 319                 nexthash = e_hash(scan);
 320                 if (nexthash != get_hash(iam_leaf_container(leaf),
 321                                          e_char(scan), e_keysize(scan))) {
 322                         BREAKPOINT();
 323                         return 0;
 324                 }
 325                 if (0 && nexthash < starthash) {
 326                         /*
 327                          * Unfortunately this useful invariant cannot be
 328                          * reliably checked as parent node is nor necessarily
 329                          * locked.
 330                          */
 331                         n_print(leaf);
 332                         printk("%#x < %#x\n", nexthash, starthash);
 333                         dump_stack();
 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("Wrong magic in node %llu (#%lu): %#x != %#x or "
 400                        "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 = get_hash(iam_leaf_container(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 = get_hash(iam_leaf_container(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
 590                 h_used(n_head(l)) +
 591                 getsize(l, strlen(kchar(k)), rec_size(r)) <= blocksize(l);
 592 }
 593
 594 static int lvar_at_end(const struct iam_leaf *folio)
 595 {
 596         assert_corr(iam_leaf_is_locked(folio));
 597         return n_cur(folio) == n_end(folio);
 598 }
 599
 600 static void lvar_rec_add(struct iam_leaf *leaf,
 601                          const struct iam_key *k, const struct iam_rec *r)
 602 {
 603         const char *key;
 604         int   ksize;
 605         int   shift;
 606         void *end;
 607         void *start;
 608         ptrdiff_t diff;
 609
 610         assert_corr(lvar_can_add(leaf, k, r));
 611         assert_inv(n_invariant(leaf));
 612         assert_corr(iam_leaf_is_locked(leaf));
 613
 614         key   = kchar(k);
 615         ksize = strlen(key);
 616         shift = getsize(leaf, ksize, rec_size(r));
 617
 618         if (!lvar_at_end(leaf)) {
 619                 assert_corr(n_cur(leaf) < n_end(leaf));
 620                 end = n_end(leaf);
 621                 if (lvar_key_cmp(leaf, k) <= 0)
 622                         lvar_next(leaf);
 623                 else
 624                         /*
 625                          * Another exceptional case: insertion with the key
 626                          * less than least key in the leaf.
 627                          */
 628                         assert_corr(leaf->il_at == leaf->il_entries);
 629
 630                 start = leaf->il_at;
 631                 diff  = PDIFF(end, start);
 632                 assert_corr(diff >= 0);
 633                 memmove(start + shift, start, diff);
 634         }
 635         h_used_adj(leaf, n_head(leaf), shift);
 636         n_cur(leaf)->vle_keysize = cpu_to_le16(ksize);
 637         n_cur(leaf)->vle_hash = cpu_to_le32(get_hash(iam_leaf_container(leaf),
 638                                                      key, ksize));
 639         __lvar_key_set(leaf, k);
 640         __lvar_rec_set(leaf, r);
 641         assert_corr(n_at_rec(leaf));
 642         assert_inv(n_invariant(leaf));
 643 }
 644
 645 static void lvar_rec_del(struct iam_leaf *leaf, int shift)
 646 {
 647         void *next;
 648         void *end;
 649         int nob;
 650
 651         assert_corr(n_at_rec(leaf));
 652         assert_inv(n_invariant(leaf));
 653         assert_corr(iam_leaf_is_locked(leaf));
 654
 655         end  = n_end(leaf);
 656         next = e_next(leaf, n_cur(leaf));
 657         nob  = e_size(leaf, n_cur(leaf));
 658         memmove(leaf->il_at, next, end - next);
 659         h_used_adj(leaf, n_head(leaf), -nob);
 660         assert_inv(n_invariant(leaf));
 661 }
 662
 663 static void lvar_init_new(struct iam_container *c, struct buffer_head *bh)
 664 {
 665         struct lvar_leaf_header *hdr;
 666
 667         hdr = (struct lvar_leaf_header *)bh->b_data;
 668         hdr->vlh_magic = cpu_to_le16(IAM_LVAR_LEAF_MAGIC);
 669         hdr->vlh_used  = sizeof *hdr;
 670 }
 671
 672 static struct lvar_leaf_entry *find_pivot(const struct iam_leaf *leaf,
 673                                           struct lvar_leaf_entry **prev)
 674 {
 675         void *scan;
 676         void *start;
 677         int threshold;
 678
 679         *prev = NULL;
 680         threshold = blocksize(leaf) / 2;
 681         for (scan = start = n_start(leaf); scan - start <= threshold;
 682              *prev = scan, scan = e_next(leaf, scan)) {
 683                 ;
 684         }
 685         return scan;
 686 }
 687
 688 static void lvar_split(struct iam_leaf *leaf, struct buffer_head **bh,
 689                        iam_ptr_t new_blknr)
 690 {
 691         struct lvar_leaf_entry  *first_to_move;
 692         struct lvar_leaf_entry  *last_to_stay;
 693         struct iam_path         *path;
 694         struct lvar_leaf_header *hdr;
 695         struct buffer_head      *new_leaf;
 696
 697         ptrdiff_t   tomove;
 698         lvar_hash_t hash;
 699
 700         assert_inv(n_invariant(leaf));
 701         assert_corr(iam_leaf_is_locked(leaf));
 702
 703         new_leaf = *bh;
 704         path = iam_leaf_path(leaf);
 705
 706         hdr = (void *)new_leaf->b_data;
 707
 708         first_to_move = find_pivot(leaf, &last_to_stay);
 709         assert_corr(last_to_stay != NULL);
 710         assert_corr(e_next(leaf, last_to_stay) == first_to_move);
 711
 712         hash = e_hash(first_to_move);
 713         if (hash == e_hash(last_to_stay))
 714                 /*
 715                  * Duplicate hash.
 716                  */
 717                 hash |= 1;
 718
 719         tomove = PDIFF(n_end(leaf), first_to_move);
 720         memmove(hdr + 1, first_to_move, tomove);
 721
 722         h_used_adj(leaf, hdr, tomove);
 723         h_used_adj(leaf, n_head(leaf), -tomove);
 724
 725         assert_corr(n_end(leaf) == first_to_move);
 726
 727         if (n_cur(leaf) >= first_to_move) {
 728                 /*
 729                  * insertion point moves into new leaf.
 730                  */
 731                 ptrdiff_t shift;
 732                 int result;
 733
 734                 shift = PDIFF(leaf->il_at, first_to_move);
 735                 *bh = leaf->il_bh;
 736                 leaf->il_bh = new_leaf;
 737                 leaf->il_curidx = new_blknr;
 738
 739                 assert_corr(iam_leaf_is_locked(leaf));
 740                 result = lvar_init(leaf);
 741                 /*
 742                  * init cannot fail, as node was just initialized.
 743                  */
 744                 assert_corr(result == 0);
 745                 leaf->il_at = ((void *)leaf->il_at) + shift;
 746         }
 747         /*
 748          * Insert pointer to the new node (together with the least key in
 749          * the node) into index node.
 750          */
 751         iam_insert_key_lock(path, path->ip_frame, (struct iam_ikey *)&hash,
 752                             new_blknr);
 753         assert_corr(n_cur(leaf) < n_end(leaf));
 754         assert_inv(n_invariant(leaf));
 755 }
 756
 757 static int lvar_leaf_empty(struct iam_leaf *leaf)
 758 {
 759         return h_used(n_head(leaf)) == sizeof(struct lvar_leaf_header);
 760 }
 761
 762 static struct iam_leaf_operations lvar_leaf_ops = {
 763         .init           = lvar_init,
 764         .init_new       = lvar_init_new,
 765         .fini           = lvar_fini,
 766         .start          = lvar_start,
 767         .next           = lvar_next,
 768         .key            = lvar_key,
 769         .ikey           = lvar_ikey,
 770         .rec            = lvar_rec,
 771         .key_set        = lvar_key_set,
 772         .key_cmp        = lvar_key_cmp,
 773         .key_eq         = lvar_key_eq,
 774         .key_size       = lvar_key_size,
 775         .rec_set        = lvar_rec_set,
 776         .rec_eq         = lvar_rec_eq,
 777         .rec_get        = lvar_rec_get,
 778         .lookup         = lvar_lookup,
 779         .ilookup        = lvar_ilookup,
 780         .at_end         = lvar_at_end,
 781         .rec_add        = lvar_rec_add,
 782         .rec_del        = lvar_rec_del,
 783         .can_add        = lvar_can_add,
 784         .split          = lvar_split,
 785         .leaf_empty     = lvar_leaf_empty,
 786 };
 787
 788 /*
 789  * Index operations.
 790  */
 791
 792 enum {
 793         /* This is duplicated in lustre/utils/create_iam.c */
 794         /* egrep -i '^o?x?[olabcdef]*$' /usr/share/dict/words */
 795         IAM_LVAR_ROOT_MAGIC = 0xb01dface
 796 };
 797
 798 /* This is duplicated in lustre/utils/create_iam.c */
 799 struct lvar_root {
 800         __le32 vr_magic;
 801         __le16 vr_recsize;
 802         __le16 vr_ptrsize;
 803         u8     vr_indirect_levels;
 804         u8     vr_padding0;
 805         __le16 vr_padding1;
 806 };
 807
 808 static __u32 lvar_root_ptr(struct iam_container *c)
 809 {
 810         return 0;
 811 }
 812
 813 static int lvar_node_init(struct iam_container *c, struct buffer_head *bh,
 814                           int root)
 815 {
 816         return 0;
 817 }
 818
 819 static struct iam_entry *lvar_root_inc(struct iam_container *c,
 820                                        struct iam_path *path,
 821                                        struct iam_frame *frame)
 822 {
 823         struct lvar_root *root;
 824         struct iam_entry *entries;
 825
 826         assert_corr(iam_frame_is_locked(path, frame));
 827         entries = frame->entries;
 828
 829         dx_set_count(entries, 2);
 830         assert_corr(dx_get_limit(entries) == dx_root_limit(path));
 831
 832         root = (void *)frame->bh->b_data;
 833         assert_corr(le64_to_cpu(root->vr_magic) == IAM_LVAR_ROOT_MAGIC);
 834         root->vr_indirect_levels ++;
 835         frame->at = entries = iam_entry_shift(path, entries, 1);
 836         memset(iam_ikey_at(path, entries), 0,
 837                iam_path_descr(path)->id_ikey_size);
 838         return entries;
 839 }
 840
 841 static int lvar_node_check(struct iam_path *path, struct iam_frame *frame)
 842 {
 843         unsigned count;
 844         unsigned limit;
 845         unsigned limit_correct;
 846         struct iam_entry *entries;
 847
 848         entries = dx_node_get_entries(path, frame);
 849
 850         if (frame == path->ip_frames) {
 851                 struct lvar_root *root;
 852
 853                 root = (void *)frame->bh->b_data;
 854                 if (le32_to_cpu(root->vr_magic) != IAM_LVAR_ROOT_MAGIC)
 855                         return -EIO;
 856                 limit_correct = dx_root_limit(path);
 857         } else
 858                 limit_correct = dx_node_limit(path);
 859         count = dx_get_count(entries);
 860         limit = dx_get_limit(entries);
 861         if (count > limit)
 862                 return -EIO;
 863         if (limit != limit_correct)
 864                 return -EIO;
 865         return 0;
 866 }
 867
 868 static int lvar_node_load(struct iam_path *path, struct iam_frame *frame)
 869 {
 870         struct iam_entry *entries;
 871         void *data;
 872         entries = dx_node_get_entries(path, frame);
 873
 874         data = frame->bh->b_data;
 875
 876         if (frame == path->ip_frames) {
 877                 struct lvar_root *root;
 878                 const char *name;
 879
 880                 root = data;
 881                 name = kchar(path->ip_key_target);
 882                 path->ip_indirect = root->vr_indirect_levels;
 883                 if (path->ip_ikey_target == NULL) {
 884                         path->ip_ikey_target = iam_path_ikey(path, 4);
 885                         *(lvar_hash_t *)path->ip_ikey_target =
 886                                 get_hash(path->ip_container, name,
 887                                          strlen(name));
 888                 }
 889         }
 890         frame->entries = frame->at = entries;
 891         return 0;
 892 }
 893
 894 static int lvar_ikeycmp(const struct iam_container *c,
 895                         const struct iam_ikey *k1, const struct iam_ikey *k2)
 896 {
 897         lvar_hash_t p1 = le32_to_cpu(*(lvar_hash_t *)k1);
 898         lvar_hash_t p2 = le32_to_cpu(*(lvar_hash_t *)k2);
 899
 900         return p1 > p2 ? +1 : (p1 < p2 ? -1 : 0);
 901 }
 902
 903 static struct iam_path_descr *lvar_ipd_alloc(const struct iam_container *c,
 904                                              void *area)
 905 {
 906         return iam_ipd_alloc(area, c->ic_descr->id_ikey_size);
 907 }
 908
 909 static void lvar_root(void *buf,
 910                       int blocksize, int keysize, int ptrsize, int recsize)
 911 {
 912         struct lvar_root *root;
 913         struct dx_countlimit *limit;
 914         void                 *entry;
 915         int isize;
 916
 917         isize = sizeof(lvar_hash_t) + ptrsize;
 918         root = buf;
 919         *root = (typeof(*root)) {
 920                 .vr_magic            = cpu_to_le32(IAM_LVAR_ROOT_MAGIC),
 921                 .vr_recsize          = cpu_to_le16(recsize),
 922                 .vr_ptrsize          = cpu_to_le16(ptrsize),
 923                 .vr_indirect_levels  = 0
 924         };
 925
 926         limit = (void *)(root + 1);
 927         *limit = (typeof(*limit)){
 928                 /*
 929                  * limit itself + one pointer to the leaf.
 930                  */
 931                 .count = cpu_to_le16(2),
 932                 .limit = iam_root_limit(sizeof(struct lvar_root), blocksize,
 933                                         sizeof (lvar_hash_t) + ptrsize)
 934         };
 935
 936         /* To guarantee that the padding "keysize + ptrsize"
 937          * covers the "dx_countlimit" and the "idle_blocks". */
 938         LASSERT((keysize + ptrsize) >=
 939                 (sizeof(struct dx_countlimit) + sizeof(__u32)));
 940
 941         entry = (void *)(limit + 1);
 942         /* Put "idle_blocks" just after the limit. There was padding after
 943          * the limit, the "idle_blocks" re-uses part of the padding, so no
 944          * compatibility issues with old layout.
 945          */
 946         *(__u32 *)entry = 0;
 947
 948         /*
 949          * Skip over @limit.
 950          */
 951         entry = (void *)(root + 1) + isize;
 952
 953         /*
 954          * Entry format is <key> followed by <ptr>. In the minimal tree
 955          * consisting of a root and single node, <key> is a minimal possible
 956          * key.
 957          */
 958         *(lvar_hash_t *)entry = 0;
 959         entry += sizeof(lvar_hash_t);
 960         /* now @entry points to <ptr> */
 961         if (ptrsize == 4)
 962                 *(u_int32_t *)entry = cpu_to_le32(1);
 963         else
 964                 *(u_int64_t *)entry = cpu_to_le64(1);
 965 }
 966
 967 static int lvar_esize(int namelen, int recsize)
 968 {
 969         return (offsetof(struct lvar_leaf_entry, vle_key) +
 970                 namelen + recsize + LVAR_ROUND) & ~LVAR_ROUND;
 971 }
 972
 973 static void lvar_leaf(void *buf,
 974                       int blocksize, int keysize, int ptrsize, int recsize)
 975 {
 976         struct lvar_leaf_header *head;
 977         struct lvar_leaf_entry  *entry;
 978
 979         /* form leaf */
 980         head = buf;
 981         *head = (typeof(*head)) {
 982                 .vlh_magic = cpu_to_le16(IAM_LVAR_LEAF_MAGIC),
 983                 .vlh_used  = cpu_to_le16(sizeof *head + lvar_esize(0, recsize))
 984         };
 985         entry = (void *)(head + 1);
 986         *entry = (typeof(*entry)) {
 987                 .vle_hash    = 0,
 988                 .vle_keysize = 0
 989         };
 990         memset(e_rec(entry), 0, recsize);
 991         *(char *)e_rec(entry) = recsize;
 992 }
 993
 994 int iam_lvar_create(struct inode *obj,
 995                     int keysize, int ptrsize, int recsize, handle_t *handle)
 996 {
 997         struct buffer_head *root_node;
 998         struct buffer_head *leaf_node;
 999         struct super_block *sb;
1000
1001         u32 blknr;
1002         int result = 0;
1003         unsigned long bsize;
1004
1005         assert_corr(obj->i_size == 0);
1006
1007         sb = obj->i_sb;
1008         bsize = sb->s_blocksize;
1009         root_node = osd_ldiskfs_append(handle, obj, &blknr);
1010         if (IS_ERR(root_node))
1011                 GOTO(out, result = PTR_ERR(root_node));
1012
1013         leaf_node = osd_ldiskfs_append(handle, obj, &blknr);
1014         if (IS_ERR(leaf_node))
1015                 GOTO(out_root, result = PTR_ERR(leaf_node));
1016
1017         lvar_root(root_node->b_data, bsize, keysize, ptrsize, recsize);
1018         lvar_leaf(leaf_node->b_data, bsize, keysize, ptrsize, recsize);
1019         ldiskfs_mark_inode_dirty(handle, obj);
1020         result = ldiskfs_handle_dirty_metadata(handle, NULL, root_node);
1021         if (result == 0)
1022                 result = ldiskfs_handle_dirty_metadata(handle, NULL, leaf_node);
1023         if (result != 0)
1024                 ldiskfs_std_error(sb, result);
1025
1026         brelse(leaf_node);
1027
1028         GOTO(out_root, result);
1029
1030 out_root:
1031         brelse(root_node);
1032 out:
1033         return result;
1034 }
1035
1036 static struct iam_operations lvar_ops = {
1037         .id_root_ptr    = lvar_root_ptr,
1038         .id_node_read   = iam_node_read,
1039         .id_node_init   = lvar_node_init,
1040         .id_node_check  = lvar_node_check,
1041         .id_node_load   = lvar_node_load,
1042         .id_ikeycmp     = lvar_ikeycmp,
1043         .id_root_inc    = lvar_root_inc,
1044         .id_ipd_alloc   = lvar_ipd_alloc,
1045         .id_ipd_free    = iam_ipd_free,
1046         .id_name        = "lvar"
1047 };
1048
1049 static int lvar_guess(struct iam_container *c)
1050 {
1051         int result;
1052         struct buffer_head *bh;
1053         const struct lvar_root *root;
1054
1055         assert_corr(c->ic_object != NULL);
1056
1057         result = iam_node_read(c, lvar_root_ptr(c), NULL, &bh);
1058         if (result == 0) {
1059                 root = (void *)bh->b_data;
1060                 if (le32_to_cpu(root->vr_magic) == IAM_LVAR_ROOT_MAGIC) {
1061                         struct iam_descr *descr;
1062
1063                         descr = c->ic_descr;
1064                         descr->id_key_size  = LDISKFS_NAME_LEN;
1065                         descr->id_ikey_size = sizeof (lvar_hash_t);
1066                         descr->id_rec_size  = le16_to_cpu(root->vr_recsize);
1067                         descr->id_ptr_size  = le16_to_cpu(root->vr_ptrsize);
1068                         descr->id_root_gap  = sizeof *root;
1069                         descr->id_node_gap  = 0;
1070                         descr->id_ops       = &lvar_ops;
1071                         descr->id_leaf_ops  = &lvar_leaf_ops;
1072
1073                         c->ic_root_bh = bh;
1074                 } else {
1075                         result = -EBADF;
1076                         brelse(bh);
1077                 }
1078         }
1079         return result;
1080 }
1081
1082 static struct iam_format lvar_format = {
1083         .if_guess = lvar_guess
1084 };
1085
1086 void iam_lvar_format_init(void)
1087 {
1088         iam_format_register(&lvar_format);
1089 }
1090