lustre/osd-ldiskfs/osd_iam_lfix.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, 2014, 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_lfix.c
  37  * implementation of iam format for fixed size records.
  38  *
  39  * Author: Wang Di <wangdi@clusterfs.com>
  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_LEAF_HEADER_MAGIC = 0x1976 /* This is duplicated in
  52                                         * lustre/utils/create_iam.c */
  53 };
  54
  55 /* This is duplicated in lustre/utils/create_iam.c */
  56 struct iam_leaf_head {
  57         __le16 ill_magic;
  58         __le16 ill_count;
  59 };
  60
  61 static inline int iam_lfix_entry_size(const struct iam_leaf *l)
  62 {
  63         return iam_leaf_descr(l)->id_key_size + iam_leaf_descr(l)->id_rec_size;
  64 }
  65
  66 static inline struct iam_lentry *
  67 iam_lfix_shift(const struct iam_leaf *l, struct iam_lentry *entry, int shift)
  68 {
  69         return (void *)entry + shift * iam_lfix_entry_size(l);
  70 }
  71
  72 static inline struct iam_key *iam_leaf_key_at(struct iam_lentry *entry)
  73 {
  74         return (struct iam_key *)entry;
  75 }
  76
  77 static inline int lfix_keycmp(const struct iam_container *c,
  78                               const struct iam_key *k1,
  79                               const struct iam_key *k2)
  80 {
  81         return memcmp(k1, k2, c->ic_descr->id_key_size);
  82 }
  83
  84 static struct iam_leaf_head *iam_get_head(const struct iam_leaf *l)
  85 {
  86         return (struct iam_leaf_head *)l->il_bh->b_data;
  87 }
  88
  89 static struct iam_lentry *iam_entries(const struct buffer_head *bh)
  90 {
  91         return (void *)bh->b_data + sizeof(struct iam_leaf_head);
  92 }
  93
  94 static struct iam_lentry *iam_get_lentries(const struct iam_leaf *l)
  95 {
  96         return iam_entries(l->il_bh);
  97 }
  98
  99 static int leaf_count_limit(const struct iam_leaf *leaf)
 100 {
 101         int free_space;
 102
 103         free_space = iam_leaf_container(leaf)->ic_object->i_sb->s_blocksize;
 104         free_space -= sizeof(struct iam_leaf_head);
 105         return free_space / iam_lfix_entry_size(leaf);
 106 }
 107
 108 static int lentry_count_get(const struct iam_leaf *leaf)
 109 {
 110         return le16_to_cpu(iam_get_head(leaf)->ill_count);
 111 }
 112
 113 static void lentry_count_set(struct iam_leaf *leaf, unsigned count)
 114 {
 115         assert_corr(0 <= count && count <= leaf_count_limit(leaf));
 116         iam_get_head(leaf)->ill_count = cpu_to_le16(count);
 117 }
 118
 119 static struct iam_lentry *iam_lfix_get_end(const struct iam_leaf *l);
 120
 121 #if LDISKFS_CORRECTNESS_ON || LDISKFS_INVARIANT_ON
 122 static int iam_leaf_at_rec(const struct iam_leaf *folio)
 123 {
 124         return
 125                 iam_get_lentries(folio) <= folio->il_at &&
 126                 folio->il_at < iam_lfix_get_end(folio);
 127 }
 128 #endif
 129
 130 static struct iam_ikey *iam_lfix_ikey(const struct iam_leaf *l,
 131                                       struct iam_ikey *key)
 132 {
 133         void *ie = l->il_at;
 134         assert_corr(iam_leaf_at_rec(l));
 135         return (struct iam_ikey*)ie;
 136 }
 137
 138 static struct iam_key *iam_lfix_key(const struct iam_leaf *l)
 139 {
 140         void *ie = l->il_at;
 141         assert_corr(iam_leaf_at_rec(l));
 142         return (struct iam_key*)ie;
 143 }
 144
 145 static int iam_lfix_key_size(const struct iam_leaf *l)
 146 {
 147         return iam_leaf_descr(l)->id_key_size;
 148 }
 149
 150 static void iam_lfix_start(struct iam_leaf *l)
 151 {
 152         l->il_at = iam_get_lentries(l);
 153 }
 154
 155 static inline ptrdiff_t iam_lfix_diff(const struct iam_leaf *l,
 156                                       const struct iam_lentry *e1,
 157                                       const struct iam_lentry *e2)
 158 {
 159         ptrdiff_t diff;
 160         int esize;
 161
 162         esize = iam_lfix_entry_size(l);
 163         diff = (void *)e1 - (void *)e2;
 164         assert_corr(diff / esize * esize == diff);
 165         return diff / esize;
 166 }
 167
 168 static int iam_lfix_init(struct iam_leaf *l)
 169 {
 170         int result;
 171         struct iam_leaf_head *ill;
 172         int count;
 173
 174         assert_corr(l->il_bh != NULL);
 175
 176         ill = iam_get_head(l);
 177         count = le16_to_cpu(ill->ill_count);
 178         if (le16_to_cpu(ill->ill_magic) == IAM_LEAF_HEADER_MAGIC &&
 179             0 <= count && count <= leaf_count_limit(l)) {
 180                 l->il_at = l->il_entries = iam_get_lentries(l);
 181                 result = 0;
 182         } else {
 183                 struct inode *obj;
 184
 185                 obj = iam_leaf_container(l)->ic_object;
 186                 CERROR("Wrong magic in node %llu (#%lu): %#x != %#x or "
 187                        "wrong count: %d (%d)\n",
 188                        (unsigned long long)l->il_bh->b_blocknr, obj->i_ino,
 189                        le16_to_cpu(ill->ill_magic), IAM_LEAF_HEADER_MAGIC,
 190                        count, leaf_count_limit(l));
 191                 result = -EIO;
 192         }
 193         return result;
 194 }
 195
 196 static void iam_lfix_fini(struct iam_leaf *l)
 197 {
 198         l->il_entries = l->il_at = NULL;
 199 }
 200
 201 static struct iam_lentry *iam_lfix_get_end(const struct iam_leaf *l)
 202 {
 203         int count = lentry_count_get(l);
 204         struct iam_lentry *ile = iam_lfix_shift(l, l->il_entries, count);
 205
 206         return ile;
 207 }
 208
 209 static struct iam_rec *iam_lfix_rec(const struct iam_leaf *l)
 210 {
 211         void *e = l->il_at;
 212         assert_corr(iam_leaf_at_rec(l));
 213         return e + iam_leaf_descr(l)->id_key_size;
 214 }
 215
 216 static void iam_lfix_next(struct iam_leaf *l)
 217 {
 218         assert_corr(iam_leaf_at_rec(l));
 219         l->il_at = iam_lfix_shift(l, l->il_at, 1);
 220 }
 221
 222 /*
 223  * Bug chasing.
 224  */
 225 int lfix_dump = 0;
 226
 227 static char hdigit(char ch)
 228 {
 229         static char d[] = "0123456789abcdef";
 230         return d[ch & 0xf];
 231 }
 232
 233 static char *hex(char ch, char *area)
 234 {
 235         area[0] = hdigit(ch >> 4);
 236         area[1] = hdigit(ch);
 237         area[2] = 0;
 238         return area;
 239 }
 240
 241 static void l_print(struct iam_leaf *leaf, struct iam_lentry *entry)
 242 {
 243         int i;
 244         char *area;
 245         char h[3];
 246
 247         area = (char *)entry;
 248         printk(KERN_EMERG "[");
 249         for (i = iam_lfix_key_size(leaf); i > 0; --i, ++area)
 250                 printk("%s", hex(*area, h));
 251         printk("]-(");
 252         for (i = iam_leaf_descr(leaf)->id_rec_size; i > 0; --i, ++area)
 253                 printk("%s", hex(*area, h));
 254         printk(")\n");
 255 }
 256
 257 static void lfix_print(struct iam_leaf *leaf)
 258 {
 259         struct iam_lentry *entry;
 260         int count;
 261         int i;
 262
 263         entry = leaf->il_entries;
 264         count = lentry_count_get(leaf);
 265         printk(KERN_EMERG "lfix: %p %p %d\n", leaf, leaf->il_at, count);
 266         for (i = 0; i < count; ++i, entry = iam_lfix_shift(leaf, entry, 1))
 267                 l_print(leaf, entry);
 268 }
 269
 270 static int iam_lfix_lookup(struct iam_leaf *l, const struct iam_key *k)
 271 {
 272         struct iam_lentry *p, *q, *m, *t;
 273         struct iam_container *c;
 274         int count;
 275         int result;
 276
 277         count = lentry_count_get(l);
 278         if (count == 0)
 279                 return IAM_LOOKUP_EMPTY;
 280
 281         result = IAM_LOOKUP_OK;
 282         c = iam_leaf_container(l);
 283
 284         p = l->il_entries;
 285         q = iam_lfix_shift(l, p, count - 1);
 286         if (lfix_keycmp(c, k, iam_leaf_key_at(p)) < 0) {
 287                 /*
 288                  * @k is less than the least key in the leaf
 289                  */
 290                 l->il_at = p;
 291                 result = IAM_LOOKUP_BEFORE;
 292         } else if (lfix_keycmp(c, iam_leaf_key_at(q), k) <= 0) {
 293                 l->il_at = q;
 294         } else {
 295                 /*
 296                  * EWD1293
 297                  */
 298                 while (iam_lfix_shift(l, p, 1) != q) {
 299                         m = iam_lfix_shift(l, p, iam_lfix_diff(l, q, p) / 2);
 300                         assert_corr(p < m && m < q);
 301                         if (lfix_keycmp(c, iam_leaf_key_at(m), k) <= 0)
 302                                 p = m;
 303                         else
 304                                 q = m;
 305                 }
 306                 assert_corr(lfix_keycmp(c, iam_leaf_key_at(p), k) <= 0 &&
 307                             lfix_keycmp(c, k, iam_leaf_key_at(q)) < 0);
 308                 /*
 309                  * skip over records with duplicate keys.
 310                  */
 311                 while (p > l->il_entries) {
 312                         t = iam_lfix_shift(l, p, -1);
 313                         if (lfix_keycmp(c, iam_leaf_key_at(t), k) == 0)
 314                                 p = t;
 315                         else
 316                                 break;
 317                 }
 318                 l->il_at = p;
 319         }
 320         assert_corr(iam_leaf_at_rec(l));
 321
 322         if (lfix_keycmp(c, iam_leaf_key_at(l->il_at), k) == 0)
 323                 result = IAM_LOOKUP_EXACT;
 324
 325         if (lfix_dump)
 326                 lfix_print(l);
 327
 328         return result;
 329 }
 330
 331 static int iam_lfix_ilookup(struct iam_leaf *l, const struct iam_ikey *ik)
 332 {
 333         return iam_lfix_lookup(l, (const struct iam_key *)ik);
 334 }
 335
 336 static void iam_lfix_key_set(struct iam_leaf *l, const struct iam_key *k)
 337 {
 338         assert_corr(iam_leaf_at_rec(l));
 339         memcpy(iam_leaf_key_at(l->il_at), k, iam_leaf_descr(l)->id_key_size);
 340 }
 341
 342 static int iam_lfix_key_cmp(const struct iam_leaf *l, const struct iam_key *k)
 343 {
 344         return lfix_keycmp(iam_leaf_container(l), iam_leaf_key_at(l->il_at), k);
 345 }
 346
 347 static int iam_lfix_key_eq(const struct iam_leaf *l, const struct iam_key *k)
 348 {
 349         return !lfix_keycmp(iam_leaf_container(l),
 350                             iam_leaf_key_at(l->il_at), k);
 351 }
 352
 353 static void iam_lfix_rec_set(struct iam_leaf *l, const struct iam_rec *r)
 354 {
 355         assert_corr(iam_leaf_at_rec(l));
 356         memcpy(iam_lfix_rec(l), r, iam_leaf_descr(l)->id_rec_size);
 357 }
 358
 359 static inline int lfix_reccmp(const struct iam_container *c,
 360                               const struct iam_rec *r1,
 361                               const struct iam_rec *r2)
 362 {
 363         return memcmp(r1, r2, c->ic_descr->id_rec_size);
 364 }
 365
 366 static int iam_lfix_rec_eq(const struct iam_leaf *l, const struct iam_rec *r)
 367 {
 368         return !lfix_reccmp(iam_leaf_container(l), iam_lfix_rec(l), r);
 369 }
 370
 371 static void iam_lfix_rec_get(const struct iam_leaf *l, struct iam_rec *r)
 372 {
 373         assert_corr(iam_leaf_at_rec(l));
 374         memcpy(r, iam_lfix_rec(l), iam_leaf_descr(l)->id_rec_size);
 375 }
 376
 377 static void iam_lfix_rec_add(struct iam_leaf *leaf,
 378                              const struct iam_key *k, const struct iam_rec *r)
 379 {
 380         struct iam_lentry *end;
 381         struct iam_lentry *cur;
 382         struct iam_lentry *start;
 383         ptrdiff_t diff;
 384         int count;
 385
 386         assert_corr(iam_leaf_can_add(leaf, k, r));
 387
 388         count = lentry_count_get(leaf);
 389         /*
 390          * This branch handles two exceptional cases:
 391          *
 392          *   - leaf positioned beyond last record, and
 393          *
 394          *   - empty leaf.
 395          */
 396         if (!iam_leaf_at_end(leaf)) {
 397                 end   = iam_lfix_get_end(leaf);
 398                 cur   = leaf->il_at;
 399                 if (lfix_keycmp(iam_leaf_container(leaf),
 400                                k, iam_leaf_key_at(cur)) >= 0)
 401                         iam_lfix_next(leaf);
 402                 else
 403                         /*
 404                          * Another exceptional case: insertion with the key
 405                          * less than least key in the leaf.
 406                          */
 407                         assert_corr(cur == leaf->il_entries);
 408
 409                 start = leaf->il_at;
 410                 diff  = (void *)end - (void *)start;
 411                 assert_corr(diff >= 0);
 412                 memmove(iam_lfix_shift(leaf, start, 1), start, diff);
 413         }
 414         lentry_count_set(leaf, count + 1);
 415         iam_lfix_key_set(leaf, k);
 416         iam_lfix_rec_set(leaf, r);
 417         assert_corr(iam_leaf_at_rec(leaf));
 418 }
 419
 420 static void iam_lfix_rec_del(struct iam_leaf *leaf, int shift)
 421 {
 422         struct iam_lentry *next, *end;
 423         int count;
 424         ptrdiff_t diff;
 425
 426         assert_corr(iam_leaf_at_rec(leaf));
 427
 428         count = lentry_count_get(leaf);
 429         end = iam_lfix_get_end(leaf);
 430         next = iam_lfix_shift(leaf, leaf->il_at, 1);
 431         diff = (void *)end - (void *)next;
 432         memmove(leaf->il_at, next, diff);
 433
 434         lentry_count_set(leaf, count - 1);
 435 }
 436
 437 static int iam_lfix_can_add(const struct iam_leaf *l,
 438                             const struct iam_key *k, const struct iam_rec *r)
 439 {
 440         return lentry_count_get(l) < leaf_count_limit(l);
 441 }
 442
 443 static int iam_lfix_at_end(const struct iam_leaf *folio)
 444 {
 445         return folio->il_at == iam_lfix_get_end(folio);
 446 }
 447
 448 static void iam_lfix_init_new(struct iam_container *c, struct buffer_head *bh)
 449 {
 450         struct iam_leaf_head *hdr;
 451
 452         hdr = (struct iam_leaf_head*)bh->b_data;
 453         hdr->ill_magic = cpu_to_le16(IAM_LEAF_HEADER_MAGIC);
 454         hdr->ill_count = cpu_to_le16(0);
 455 }
 456
 457 static void iam_lfix_split(struct iam_leaf *l, struct buffer_head **bh,
 458                            iam_ptr_t new_blknr)
 459 {
 460         struct iam_path       *path;
 461         struct iam_leaf_head  *hdr;
 462         const struct iam_ikey *pivot;
 463         struct buffer_head    *new_leaf;
 464
 465         unsigned count;
 466         unsigned split;
 467
 468         void *start;
 469         void *finis;
 470
 471         new_leaf = *bh;
 472         path = iam_leaf_path(l);
 473
 474         hdr = (void *)new_leaf->b_data;
 475
 476         count = lentry_count_get(l);
 477         split = count / 2;
 478
 479         start = iam_lfix_shift(l, iam_get_lentries(l), split);
 480         finis = iam_lfix_shift(l, iam_get_lentries(l), count);
 481
 482         pivot = (const struct iam_ikey *)iam_leaf_key_at(start);
 483
 484         memmove(iam_entries(new_leaf), start, finis - start);
 485         hdr->ill_count = cpu_to_le16(count - split);
 486         lentry_count_set(l, split);
 487         if ((void *)l->il_at >= start) {
 488                 /*
 489                  * insertion point moves into new leaf.
 490                  */
 491                 int shift;
 492                 int result;
 493
 494                 shift = iam_lfix_diff(l, l->il_at, start);
 495                 *bh = l->il_bh;
 496                 l->il_bh = new_leaf;
 497                 l->il_curidx = new_blknr;
 498                 result = iam_lfix_init(l);
 499                 /*
 500                  * init cannot fail, as node was just initialized.
 501                  */
 502                 assert_corr(result == 0);
 503                 l->il_at = iam_lfix_shift(l, iam_get_lentries(l), shift);
 504         }
 505         /*
 506          * Insert pointer to the new node (together with the least key in
 507          * the node) into index node.
 508          */
 509         iam_insert_key_lock(path, path->ip_frame, pivot, new_blknr);
 510 }
 511
 512 static int iam_lfix_leaf_empty(struct iam_leaf *leaf)
 513 {
 514         return lentry_count_get(leaf) == 0;
 515 }
 516
 517 static struct iam_leaf_operations iam_lfix_leaf_ops = {
 518         .init           = iam_lfix_init,
 519         .init_new       = iam_lfix_init_new,
 520         .fini           = iam_lfix_fini,
 521         .start          = iam_lfix_start,
 522         .next           = iam_lfix_next,
 523         .key            = iam_lfix_key,
 524         .ikey           = iam_lfix_ikey,
 525         .rec            = iam_lfix_rec,
 526         .key_set        = iam_lfix_key_set,
 527         .key_cmp        = iam_lfix_key_cmp,
 528         .key_eq         = iam_lfix_key_eq,
 529         .key_size       = iam_lfix_key_size,
 530         .rec_set        = iam_lfix_rec_set,
 531         .rec_eq         = iam_lfix_rec_eq,
 532         .rec_get        = iam_lfix_rec_get,
 533         .lookup         = iam_lfix_lookup,
 534         .ilookup        = iam_lfix_ilookup,
 535         .at_end         = iam_lfix_at_end,
 536         .rec_add        = iam_lfix_rec_add,
 537         .rec_del        = iam_lfix_rec_del,
 538         .can_add        = iam_lfix_can_add,
 539         .split          = iam_lfix_split,
 540         .leaf_empty     = iam_lfix_leaf_empty,
 541 };
 542
 543 /*
 544  * Index operations.
 545  */
 546
 547 enum {
 548         /* This is duplicated in lustre/utils/create_iam.c */
 549         /*
 550          * Then shalt thou see the dew-BEDABBLED wretch
 551          * Turn, and return, indenting with the way;
 552          * Each envious brier his weary legs doth scratch,
 553          * Each shadow makes him stop, each murmur stay:
 554          * For misery is trodden on by many,
 555          * And being low never relieved by any.
 556          */
 557         IAM_LFIX_ROOT_MAGIC = 0xbedabb1edULL // d01efull
 558 };
 559
 560 /* This is duplicated in lustre/utils/create_iam.c */
 561 struct iam_lfix_root {
 562         __le64  ilr_magic;
 563         __le16  ilr_keysize;
 564         __le16  ilr_recsize;
 565         __le16  ilr_ptrsize;
 566         u8      ilr_indirect_levels;
 567         u8      ilr_padding;
 568 };
 569
 570 static __u32 iam_lfix_root_ptr(struct iam_container *c)
 571 {
 572         return 0;
 573 }
 574
 575 static int iam_lfix_node_init(struct iam_container *c, struct buffer_head *bh,
 576                               int root)
 577 {
 578         return 0;
 579 }
 580
 581 static struct iam_entry *iam_lfix_root_inc(struct iam_container *c,
 582                                            struct iam_path *path,
 583                                            struct iam_frame *frame)
 584 {
 585         struct iam_lfix_root *root;
 586         struct iam_entry     *entries;
 587
 588         entries = frame->entries;
 589
 590         dx_set_count(entries, 2);
 591         assert_corr(dx_get_limit(entries) == dx_root_limit(path));
 592
 593         root = (void *)frame->bh->b_data;
 594         assert_corr(le64_to_cpu(root->ilr_magic) == IAM_LFIX_ROOT_MAGIC);
 595         root->ilr_indirect_levels ++;
 596         frame->at = entries = iam_entry_shift(path, entries, 1);
 597         memset(iam_ikey_at(path, entries), 0,
 598                iam_path_descr(path)->id_ikey_size);
 599         return entries;
 600 }
 601
 602 static int iam_lfix_node_check(struct iam_path *path, struct iam_frame *frame)
 603 {
 604         unsigned count;
 605         unsigned limit;
 606         unsigned limit_correct;
 607         struct iam_entry *entries;
 608
 609         entries = dx_node_get_entries(path, frame);
 610
 611         if (frame == path->ip_frames) {
 612                 struct iam_lfix_root *root;
 613
 614                 root = (void *)frame->bh->b_data;
 615                 if (le64_to_cpu(root->ilr_magic) != IAM_LFIX_ROOT_MAGIC) {
 616                         return -EIO;
 617                 }
 618                 limit_correct = dx_root_limit(path);
 619         } else
 620                 limit_correct = dx_node_limit(path);
 621         count = dx_get_count(entries);
 622         limit = dx_get_limit(entries);
 623         if (count > limit) {
 624                 return -EIO;
 625         }
 626         if (limit != limit_correct) {
 627                 return -EIO;
 628         }
 629         return 0;
 630 }
 631
 632 static int iam_lfix_node_load(struct iam_path *path, struct iam_frame *frame)
 633 {
 634         struct iam_entry *entries;
 635         void *data;
 636         entries = dx_node_get_entries(path, frame);
 637
 638         data = frame->bh->b_data;
 639
 640         if (frame == path->ip_frames) {
 641                 struct iam_lfix_root *root;
 642
 643                 root = data;
 644                 path->ip_indirect = root->ilr_indirect_levels;
 645                 if (path->ip_ikey_target == NULL)
 646                         path->ip_ikey_target =
 647                                 (struct iam_ikey *)path->ip_key_target;
 648         }
 649         frame->entries = frame->at = entries;
 650         return 0;
 651 }
 652
 653 static int iam_lfix_ikeycmp(const struct iam_container *c,
 654                             const struct iam_ikey *k1,
 655                             const struct iam_ikey *k2)
 656 {
 657         return memcmp(k1, k2, c->ic_descr->id_ikey_size);
 658 }
 659
 660 static struct iam_path_descr *iam_lfix_ipd_alloc(const struct iam_container *c,
 661                                                  void *area)
 662 {
 663         return iam_ipd_alloc(area, c->ic_descr->id_ikey_size);
 664 }
 665
 666 static struct iam_operations iam_lfix_ops = {
 667         .id_root_ptr    = iam_lfix_root_ptr,
 668         .id_node_read   = iam_node_read,
 669         .id_node_init   = iam_lfix_node_init,
 670         .id_node_check  = iam_lfix_node_check,
 671         .id_node_load   = iam_lfix_node_load,
 672         .id_ikeycmp     = iam_lfix_ikeycmp,
 673         .id_root_inc    = iam_lfix_root_inc,
 674         .id_ipd_alloc   = iam_lfix_ipd_alloc,
 675         .id_ipd_free    = iam_ipd_free,
 676         .id_name        = "lfix"
 677 };
 678
 679 static int iam_lfix_guess(struct iam_container *c)
 680 {
 681         int result;
 682         struct buffer_head *bh;
 683         const struct iam_lfix_root *root;
 684
 685         assert_corr(c->ic_object != NULL);
 686
 687         result = iam_node_read(c, iam_lfix_root_ptr(c), NULL, &bh);
 688         if (result == 0) {
 689                 root = (void *)bh->b_data;
 690                 if (le64_to_cpu(root->ilr_magic) == IAM_LFIX_ROOT_MAGIC) {
 691                         struct iam_descr *descr;
 692
 693                         descr = c->ic_descr;
 694                         descr->id_key_size  = le16_to_cpu(root->ilr_keysize);
 695                         descr->id_ikey_size = le16_to_cpu(root->ilr_keysize);
 696                         descr->id_rec_size  = le16_to_cpu(root->ilr_recsize);
 697                         descr->id_ptr_size  = le16_to_cpu(root->ilr_ptrsize);
 698                         descr->id_root_gap  = sizeof(struct iam_lfix_root);
 699                         descr->id_node_gap  = 0;
 700                         descr->id_ops       = &iam_lfix_ops;
 701                         descr->id_leaf_ops  = &iam_lfix_leaf_ops;
 702
 703                         c->ic_root_bh = bh;
 704                 } else {
 705                         result = -EBADF;
 706                         brelse(bh);
 707                 }
 708         }
 709         return result;
 710 }
 711
 712 static struct iam_format iam_lfix_format = {
 713         .if_guess = iam_lfix_guess
 714 };
 715
 716 void iam_lfix_format_init(void)
 717 {
 718         iam_format_register(&iam_lfix_format);
 719 }
 720
 721 /*
 722  * Debugging aid.
 723  */
 724
 725 #define KEYSIZE (8)
 726 #define RECSIZE (8)
 727 #define PTRSIZE (4)
 728
 729 #define LFIX_ROOT_RECNO \
 730         ((4096 - sizeof(struct iam_lfix_root)) / (KEYSIZE + PTRSIZE))
 731
 732 #define LFIX_INDEX_RECNO (4096 / (KEYSIZE + PTRSIZE))
 733
 734 #define LFIX_LEAF_RECNO \
 735         ((4096 - sizeof(struct iam_leaf_head)) / (KEYSIZE + RECSIZE))
 736
 737 struct lfix_root {
 738         struct iam_lfix_root lr_root;
 739         struct {
 740                 char key[KEYSIZE];
 741                 char ptr[PTRSIZE];
 742         } lr_entry[LFIX_ROOT_RECNO];
 743 };
 744
 745 struct lfix_index {
 746         struct dx_countlimit li_cl;
 747         char   li_padding[KEYSIZE + PTRSIZE - sizeof(struct dx_countlimit)];
 748         struct {
 749                 char key[KEYSIZE];
 750                 char ptr[PTRSIZE];
 751         } li_entry[LFIX_INDEX_RECNO - 1];
 752 };
 753
 754 struct lfix_leaf {
 755         struct iam_leaf_head ll_head;
 756         struct {
 757                 char key[KEYSIZE];
 758                 char rec[RECSIZE];
 759         } ll_entry[LFIX_LEAF_RECNO];
 760 };
 761
 762 #define STORE_UNALIGNED(val, dst)                       \
 763 ({                                                      \
 764         typeof(val) __val = (val);                      \
 765         CLASSERT(sizeof(val) == sizeof(*(dst)));        \
 766         memcpy(dst, &__val, sizeof(*(dst)));            \
 767 })
 768
 769 static void lfix_root(void *buf,
 770                       int blocksize, int keysize, int ptrsize, int recsize)
 771 {
 772         struct iam_lfix_root *root;
 773         struct dx_countlimit *limit;
 774         void                 *entry;
 775
 776         root = buf;
 777         *root = (typeof(*root)) {
 778                 .ilr_magic           = cpu_to_le64(IAM_LFIX_ROOT_MAGIC),
 779                 .ilr_keysize         = cpu_to_le16(keysize),
 780                 .ilr_recsize         = cpu_to_le16(recsize),
 781                 .ilr_ptrsize         = cpu_to_le16(ptrsize),
 782                 .ilr_indirect_levels = 0
 783         };
 784
 785         limit = (void *)(root + 1);
 786         *limit = (typeof(*limit)){
 787                 /*
 788                  * limit itself + one pointer to the leaf.
 789                  */
 790                 .count = cpu_to_le16(2),
 791                 .limit = iam_root_limit(sizeof(struct iam_lfix_root),
 792                                         blocksize, keysize + ptrsize)
 793         };
 794
 795         /* To guarantee that the padding "keysize + ptrsize"
 796          * covers the "dx_countlimit" and the "idle_blocks". */
 797         LASSERT((keysize + ptrsize) >=
 798                 (sizeof(struct dx_countlimit) + sizeof(__u32)));
 799
 800         entry = (void *)(limit + 1);
 801         /* Put "idle_blocks" just after the limit. There was padding after
 802          * the limit, the "idle_blocks" re-uses part of the padding, so no
 803          * compatibility issues with old layout.
 804          */
 805         *(__u32 *)entry = 0;
 806
 807         /*
 808          * Skip over @limit.
 809          */
 810         entry = (void *)(root + 1) + keysize + ptrsize;
 811
 812         /*
 813          * Entry format is <key> followed by <ptr>. In the minimal tree
 814          * consisting of a root and single node, <key> is a minimal possible
 815          * key.
 816          *
 817          * XXX: this key is hard-coded to be a sequence of 0's.
 818          */
 819
 820         memset(entry, 0, keysize);
 821         entry += keysize;
 822         /* now @entry points to <ptr> */
 823         if (ptrsize == 4)
 824                 STORE_UNALIGNED(cpu_to_le32(1), (u_int32_t *)entry);
 825         else
 826                 STORE_UNALIGNED(cpu_to_le64(1), (u_int64_t *)entry);
 827 }
 828
 829 static void lfix_leaf(void *buf,
 830                       int blocksize, int keysize, int ptrsize, int recsize)
 831 {
 832         struct iam_leaf_head *head;
 833         void *entry;
 834
 835         /* form leaf */
 836         head = buf;
 837         *head = (struct iam_leaf_head) {
 838                 .ill_magic = cpu_to_le16(IAM_LEAF_HEADER_MAGIC),
 839                 /*
 840                  * Leaf contains an entry with the smallest possible key
 841                  * (created by zeroing).
 842                  */
 843                 .ill_count = cpu_to_le16(1),
 844         };
 845
 846         entry = (void *)(head + 1);
 847         memset(entry, 0, keysize + recsize);
 848 }
 849
 850 int iam_lfix_create(struct inode *obj,
 851                     int keysize, int ptrsize, int recsize, handle_t *handle)
 852 {
 853         struct buffer_head *root_node;
 854         struct buffer_head *leaf_node;
 855         struct super_block *sb;
 856
 857         u32 blknr;
 858         int result = 0;
 859         unsigned long bsize;
 860
 861         assert_corr(obj->i_size == 0);
 862
 863         sb = obj->i_sb;
 864         bsize = sb->s_blocksize;
 865         root_node = osd_ldiskfs_append(handle, obj, &blknr, &result);
 866         leaf_node = osd_ldiskfs_append(handle, obj, &blknr, &result);
 867         if (root_node != NULL && leaf_node != NULL) {
 868                 lfix_root(root_node->b_data, bsize, keysize, ptrsize, recsize);
 869                 lfix_leaf(leaf_node->b_data, bsize, keysize, ptrsize, recsize);
 870                 ldiskfs_mark_inode_dirty(handle, obj);
 871                 result = ldiskfs_handle_dirty_metadata(handle, NULL, root_node);
 872                 if (result == 0)
 873                         result = ldiskfs_handle_dirty_metadata(handle, NULL,
 874                                                                leaf_node);
 875                 if (result != 0)
 876                         ldiskfs_std_error(sb, result);
 877         }
 878         brelse(leaf_node);
 879         brelse(root_node);
 880         return result;
 881 }