tdbBtree.c 21.2 KB
Newer Older
H
Hongze Cheng 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
/*
 * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
 *
 * This program is free software: you can use, redistribute, and/or modify
 * it under the terms of the GNU Affero General Public License, version 3
 * or later ("AGPL"), as published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 * FITNESS FOR A PARTICULAR PURPOSE.
 *
 * You should have received a copy of the GNU Affero General Public License
 * along with this program. If not, see <http://www.gnu.org/licenses/>.
H
refact  
Hongze Cheng 已提交
14 15
 */

H
Hongze Cheng 已提交
16 17
#include "tdbInt.h"

H
Hongze Cheng 已提交
18 19 20 21 22
#define TDB_BTREE_ROOT 0x1
#define TDB_BTREE_LEAF 0x2

#define TDB_BTREE_PAGE_IS_ROOT(flags) TDB_FLAG_HAS(flags, TDB_BTREE_ROOT)
#define TDB_BTREE_PAGE_IS_LEAF(flags) TDB_FLAG_HAS(flags, TDB_BTREE_LEAF)
H
Hongze Cheng 已提交
23 24 25
#define TDB_BTREE_ASSERT_FLAG(flags)                                                 \
  ASSERT(TDB_FLAG_IS(flags, TDB_BTREE_ROOT) || TDB_FLAG_IS(flags, TDB_BTREE_LEAF) || \
         TDB_FLAG_IS(flags, TDB_BTREE_ROOT | TDB_BTREE_LEAF) || TDB_FLAG_IS(flags, 0))
H
Hongze Cheng 已提交
26

H
Hongze Cheng 已提交
27
struct SBTree {
H
more  
Hongze Cheng 已提交
28 29 30
  SPgno          root;
  int            keyLen;
  int            valLen;
H
Hongze Cheng 已提交
31
  SPager        *pPager;
H
more  
Hongze Cheng 已提交
32
  FKeyComparator kcmpr;
H
more  
Hongze Cheng 已提交
33
  u8             fanout;
H
Hongze Cheng 已提交
34
  int            pageSize;
H
more  
Hongze Cheng 已提交
35 36 37 38
  int            maxLocal;
  int            minLocal;
  int            maxLeaf;
  int            minLeaf;
H
Hongze Cheng 已提交
39
  u8            *pTmp;
H
Hongze Cheng 已提交
40 41
};

H
Hongze Cheng 已提交
42 43 44
typedef struct __attribute__((__packed__)) {
  SPgno rChild;
} SBtPageHdr;
H
more  
Hongze Cheng 已提交
45

H
Hongze Cheng 已提交
46 47 48 49 50
typedef struct {
  u16     flags;
  SBTree *pBt;
} SBtreeZeroPageArg;

H
Hongze Cheng 已提交
51 52 53 54 55 56 57 58 59
typedef struct {
  int   kLen;
  u8   *pKey;
  int   vLen;
  u8   *pVal;
  SPgno pgno;
  u8   *pTmpSpace;
} SCellDecoder;

H
Hongze Cheng 已提交
60
static int tdbBtCursorMoveTo(SBtCursor *pCur, const void *pKey, int kLen, int *pCRst);
H
Hongze Cheng 已提交
61
static int tdbDefaultKeyCmprFn(const void *pKey1, int keyLen1, const void *pKey2, int keyLen2);
H
Hongze Cheng 已提交
62
static int tdbBtreeOpenImpl(SBTree *pBt);
H
Hongze Cheng 已提交
63 64
static int tdbBtreeZeroPage(SPage *pPage, void *arg);
static int tdbBtreeInitPage(SPage *pPage, void *arg);
H
Hongze Cheng 已提交
65 66
static int tdbBtreeEncodeCell(SPage *pPage, const void *pKey, int kLen, const void *pVal, int vLen, SCell *pCell,
                              int *szCell);
H
Hongze Cheng 已提交
67
static int tdbBtreeDecodeCell(SPage *pPage, const SCell *pCell, SCellDecoder *pDecoder);
H
Hongze Cheng 已提交
68
static int tdbBtreeBalance(SBtCursor *pCur);
H
Hongze Cheng 已提交
69

H
Hongze Cheng 已提交
70
int tdbBtreeOpen(int keyLen, int valLen, SPager *pPager, FKeyComparator kcmpr, SBTree **ppBt) {
H
Hongze Cheng 已提交
71
  SBTree *pBt;
H
Hongze Cheng 已提交
72
  int     ret;
H
more  
Hongze Cheng 已提交
73

H
Hongze Cheng 已提交
74
  *ppBt = NULL;
H
Hongze Cheng 已提交
75 76 77 78 79 80 81 82 83 84

  pBt = (SBTree *)calloc(1, sizeof(*pBt));
  if (pBt == NULL) {
    return -1;
  }

  // pBt->keyLen
  pBt->keyLen = keyLen;
  // pBt->valLen
  pBt->valLen = valLen;
H
refact  
Hongze Cheng 已提交
85 86
  // pBt->pPager
  pBt->pPager = pPager;
H
Hongze Cheng 已提交
87 88
  // pBt->kcmpr
  pBt->kcmpr = kcmpr ? kcmpr : tdbDefaultKeyCmprFn;
H
more  
Hongze Cheng 已提交
89 90 91 92 93 94 95
  // pBt->fanout
  if (keyLen == TDB_VARIANT_LEN) {
    pBt->fanout = TDB_DEFAULT_FANOUT;
  } else {
    ASSERT(0);
    // TODO: pBt->fanout = 0;
  }
H
Hongze Cheng 已提交
96 97 98
  // pBt->pageSize
  pBt->pageSize = tdbPagerGetPageSize(pPager);
  // pBt->maxLocal
H
Hongze Cheng 已提交
99
  pBt->maxLocal = (pBt->pageSize - 14) / pBt->fanout;
H
Hongze Cheng 已提交
100
  // pBt->minLocal: Should not be allowed smaller than 15, which is [nPayload][nKey][nData]
H
Hongze Cheng 已提交
101
  pBt->minLocal = (pBt->pageSize - 14) / pBt->fanout / 2;
H
Hongze Cheng 已提交
102
  // pBt->maxLeaf
H
Hongze Cheng 已提交
103
  pBt->maxLeaf = pBt->pageSize - 14;
H
Hongze Cheng 已提交
104 105
  // pBt->minLeaf
  pBt->minLeaf = pBt->minLocal;
H
Hongze Cheng 已提交
106

H
Hongze Cheng 已提交
107 108 109 110 111 112 113
  // TODO: pBt->root
  ret = tdbBtreeOpenImpl(pBt);
  if (ret < 0) {
    free(pBt);
    return -1;
  }

H
Hongze Cheng 已提交
114
  *ppBt = pBt;
H
Hongze Cheng 已提交
115 116 117 118 119 120 121 122
  return 0;
}

int tdbBtreeClose(SBTree *pBt) {
  // TODO
  return 0;
}

H
more  
Hongze Cheng 已提交
123
int tdbBtreeCursor(SBtCursor *pCur, SBTree *pBt) {
H
more  
Hongze Cheng 已提交
124 125
  pCur->pBt = pBt;
  pCur->iPage = -1;
H
more  
Hongze Cheng 已提交
126
  pCur->pPage = NULL;
H
Hongze Cheng 已提交
127
  pCur->idx = -1;
H
more  
Hongze Cheng 已提交
128

H
more  
Hongze Cheng 已提交
129 130 131
  return 0;
}

H
Hongze Cheng 已提交
132
int tdbBtCursorInsert(SBtCursor *pCur, const void *pKey, int kLen, const void *pVal, int vLen) {
H
Hongze Cheng 已提交
133
  int     ret;
H
Hongze Cheng 已提交
134
  int     idx;
H
Hongze Cheng 已提交
135
  SPager *pPager;
H
Hongze Cheng 已提交
136 137 138
  SCell  *pCell;
  int     szCell;
  int     cret;
H
Hongze Cheng 已提交
139
  SBTree *pBt;
H
Hongze Cheng 已提交
140

H
Hongze Cheng 已提交
141
  ret = tdbBtCursorMoveTo(pCur, pKey, kLen, &cret);
H
Hongze Cheng 已提交
142 143 144 145 146
  if (ret < 0) {
    // TODO: handle error
    return -1;
  }

H
Hongze Cheng 已提交
147
  if (pCur->idx == -1) {
H
Hongze Cheng 已提交
148
    ASSERT(TDB_PAGE_TOTAL_CELLS(pCur->pPage) == 0);
H
Hongze Cheng 已提交
149 150 151
    idx = 0;
  } else {
    if (cret > 0) {
H
Hongze Cheng 已提交
152
      idx = pCur->idx + 1;
H
Hongze Cheng 已提交
153
    } else if (cret < 0) {
H
Hongze Cheng 已提交
154
      idx = pCur->idx;
H
Hongze Cheng 已提交
155 156 157 158 159 160
    } else {
      /* TODO */
      ASSERT(0);
    }
  }

H
Hongze Cheng 已提交
161 162 163 164 165 166 167 168 169 170 171
  // TODO: refact code here
  pBt = pCur->pBt;
  if (!pBt->pTmp) {
    pBt->pTmp = (u8 *)malloc(pBt->pageSize);
    if (pBt->pTmp == NULL) {
      return -1;
    }
  }

  pCell = pBt->pTmp;

H
Hongze Cheng 已提交
172
  // Encode the cell
H
Hongze Cheng 已提交
173 174 175 176
  ret = tdbBtreeEncodeCell(pCur->pPage, pKey, kLen, pVal, vLen, pCell, &szCell);
  if (ret < 0) {
    return -1;
  }
H
Hongze Cheng 已提交
177 178 179 180 181

  // Insert the cell to the index
  ret = tdbPageInsertCell(pCur->pPage, idx, pCell, szCell);
  if (ret < 0) {
    return -1;
H
Hongze Cheng 已提交
182 183
  }

H
Hongze Cheng 已提交
184 185 186 187 188 189 190 191
  // If page is overflow, balance the tree
  if (pCur->pPage->nOverflow > 0) {
    ret = tdbBtreeBalance(pCur);
    if (ret < 0) {
      return -1;
    }
  }

H
more  
Hongze Cheng 已提交
192 193 194
  return 0;
}

H
Hongze Cheng 已提交
195 196 197 198 199
static int tdbBtCursorMoveToChild(SBtCursor *pCur, SPgno pgno) {
  // TODO
  return 0;
}

H
Hongze Cheng 已提交
200
static int tdbBtCursorMoveTo(SBtCursor *pCur, const void *pKey, int kLen, int *pCRst) {
H
Hongze Cheng 已提交
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220
  int     ret;
  SBTree *pBt;
  SPager *pPager;

  pBt = pCur->pBt;
  pPager = pBt->pPager;

  if (pCur->iPage < 0) {
    ASSERT(pCur->iPage == -1);
    ASSERT(pCur->idx == -1);

    // Move from the root
    ret = tdbPagerFetchPage(pPager, pBt->root, &(pCur->pPage), tdbBtreeInitPage, pBt);
    if (ret < 0) {
      ASSERT(0);
      return -1;
    }

    pCur->iPage = 0;

H
Hongze Cheng 已提交
221
    if (TDB_PAGE_TOTAL_CELLS(pCur->pPage) == 0) {
H
Hongze Cheng 已提交
222
      // Current page is empty
H
Hongze Cheng 已提交
223
      ASSERT(TDB_FLAG_IS(TDB_PAGE_FLAGS(pCur->pPage), TDB_BTREE_ROOT | TDB_BTREE_LEAF));
H
Hongze Cheng 已提交
224 225 226
      return 0;
    }

H
Hongze Cheng 已提交
227
    for (;;) {
H
Hongze Cheng 已提交
228
      int          lidx, ridx, midx, c, nCells;
H
Hongze Cheng 已提交
229 230 231 232 233
      SCell       *pCell;
      SPage       *pPage;
      SCellDecoder cd = {0};

      pPage = pCur->pPage;
H
Hongze Cheng 已提交
234
      nCells = TDB_PAGE_TOTAL_CELLS(pPage);
H
Hongze Cheng 已提交
235
      lidx = 0;
H
Hongze Cheng 已提交
236 237 238 239
      ridx = nCells - 1;

      ASSERT(nCells > 0);

H
Hongze Cheng 已提交
240
      for (;;) {
H
Hongze Cheng 已提交
241 242 243 244
        if (lidx > ridx) break;

        midx = (lidx + ridx) >> 1;

H
Hongze Cheng 已提交
245
        pCell = tdbPageGetCell(pPage, midx);
H
Hongze Cheng 已提交
246 247 248 249 250 251
        ret = tdbBtreeDecodeCell(pPage, pCell, &cd);
        if (ret < 0) {
          // TODO: handle error
          ASSERT(0);
          return -1;
        }
H
Hongze Cheng 已提交
252

H
Hongze Cheng 已提交
253 254 255
        // Compare the key values
        c = pBt->kcmpr(pKey, kLen, cd.pKey, cd.kLen);
        if (c < 0) {
H
Hongze Cheng 已提交
256 257
          /* input-key < cell-key */
          ridx = midx - 1;
H
Hongze Cheng 已提交
258
        } else if (c > 0) {
H
Hongze Cheng 已提交
259 260
          /* input-key > cell-key */
          lidx = midx + 1;
H
Hongze Cheng 已提交
261
        } else {
H
Hongze Cheng 已提交
262 263 264 265 266
          /* input-key == cell-key */
          break;
        }
      }

H
Hongze Cheng 已提交
267
      // Move downward or break
H
Hongze Cheng 已提交
268 269 270 271
      u16 flags = TDB_PAGE_FLAGS(pPage);
      u8  leaf = TDB_BTREE_PAGE_IS_LEAF(flags);
      if (leaf) {
        pCur->idx = midx;
H
Hongze Cheng 已提交
272
        *pCRst = c;
H
Hongze Cheng 已提交
273 274 275 276 277 278 279 280 281 282 283 284 285
        break;
      } else {
        if (c <= 0) {
          pCur->idx = midx;
          tdbBtCursorMoveToChild(pCur, cd.pgno);
        } else {
          if (midx == nCells - 1) {
            /* Move to right-most child */
            pCur->idx = midx + 1;
            tdbBtCursorMoveToChild(pCur, ((SBtPageHdr *)(pPage->pAmHdr))->rChild);
          } else {
            // TODO: reset cd as uninitialized
            pCur->idx = midx + 1;
H
Hongze Cheng 已提交
286
            pCell = tdbPageGetCell(pPage, midx + 1);
H
Hongze Cheng 已提交
287 288 289
            tdbBtreeDecodeCell(pPage, pCell, &cd);
            tdbBtCursorMoveToChild(pCur, cd.pgno);
          }
H
Hongze Cheng 已提交
290 291
        }
      }
H
Hongze Cheng 已提交
292 293 294 295
    }

  } else {
    // TODO: Move the cursor from a some position instead of a clear state
H
Hongze Cheng 已提交
296
    ASSERT(0);
H
Hongze Cheng 已提交
297
  }
H
more  
Hongze Cheng 已提交
298

H
Hongze Cheng 已提交
299 300
  return 0;
}
H
more  
Hongze Cheng 已提交
301

H
more  
Hongze Cheng 已提交
302
static int tdbBtCursorMoveToRoot(SBtCursor *pCur) {
H
refact  
Hongze Cheng 已提交
303 304
  SBTree *pBt;
  SPager *pPager;
H
Hongze Cheng 已提交
305
  SPage  *pPage;
H
refact  
Hongze Cheng 已提交
306
  int     ret;
H
more  
Hongze Cheng 已提交
307 308

  pBt = pCur->pBt;
H
refact  
Hongze Cheng 已提交
309
  pPager = pBt->pPager;
H
more  
Hongze Cheng 已提交
310

H
Hongze Cheng 已提交
311 312 313 314
  // pPage = tdbPagerGet(pPager, pBt->root, true);
  // if (pPage == NULL) {
  //   // TODO: handle error
  // }
H
more  
Hongze Cheng 已提交
315

H
Hongze Cheng 已提交
316 317 318 319 320
  // ret = tdbInitBtPage(pPage, &pBtPage);
  // if (ret < 0) {
  //   // TODO
  //   return 0;
  // }
H
more  
Hongze Cheng 已提交
321

H
Hongze Cheng 已提交
322 323
  // pCur->pPage = pBtPage;
  // pCur->iPage = 0;
H
more  
Hongze Cheng 已提交
324 325 326 327

  return 0;
}

H
Hongze Cheng 已提交
328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345
static int tdbDefaultKeyCmprFn(const void *pKey1, int keyLen1, const void *pKey2, int keyLen2) {
  int mlen;
  int cret;

  ASSERT(keyLen1 > 0 && keyLen2 > 0 && pKey1 != NULL && pKey2 != NULL);

  mlen = keyLen1 < keyLen2 ? keyLen1 : keyLen2;
  cret = memcmp(pKey1, pKey2, mlen);
  if (cret == 0) {
    if (keyLen1 < keyLen2) {
      cret = -1;
    } else if (keyLen1 > keyLen2) {
      cret = 1;
    } else {
      cret = 0;
    }
  }
  return cret;
H
Hongze Cheng 已提交
346 347 348 349 350 351 352 353 354 355
}

static int tdbBtreeOpenImpl(SBTree *pBt) {
  // Try to get the root page of the an existing btree

  SPgno  pgno;
  SPage *pPage;
  int    ret;

  {
H
Hongze Cheng 已提交
356
    // 1. TODO: Search the main DB to check if the DB exists
H
Hongze Cheng 已提交
357 358 359 360 361 362 363 364 365
    pgno = 0;
  }

  if (pgno != 0) {
    pBt->root = pgno;
    return 0;
  }

  // Try to create a new database
H
Hongze Cheng 已提交
366 367 368 369 370 371 372
  SBtreeZeroPageArg zArg = {.flags = TDB_BTREE_ROOT | TDB_BTREE_LEAF, .pBt = pBt};
  ret = tdbPagerNewPage(pBt->pPager, &pgno, &pPage, tdbBtreeZeroPage, &zArg);
  if (ret < 0) {
    return -1;
  }

  // TODO: Unref the page
H
Hongze Cheng 已提交
373

H
Hongze Cheng 已提交
374 375
  ASSERT(pgno != 0);
  pBt->root = pgno;
H
Hongze Cheng 已提交
376

H
Hongze Cheng 已提交
377 378 379
  return 0;
}

H
Hongze Cheng 已提交
380
static int tdbBtreeInitPage(SPage *pPage, void *arg) {
H
Hongze Cheng 已提交
381
  SBTree *pBt;
H
Hongze Cheng 已提交
382
  u16     flags;
H
Hongze Cheng 已提交
383
  u8      isLeaf;
H
Hongze Cheng 已提交
384
  u8      szAmHdr;
H
Hongze Cheng 已提交
385

H
Hongze Cheng 已提交
386
  pBt = (SBTree *)arg;
H
Hongze Cheng 已提交
387

H
Hongze Cheng 已提交
388 389 390
  ASSERT(0);

  // TODO: here has problem
H
Hongze Cheng 已提交
391
  flags = TDB_PAGE_FLAGS(pPage);
H
Hongze Cheng 已提交
392 393
  isLeaf = TDB_BTREE_PAGE_IS_LEAF(flags);
  if (isLeaf) {
H
Hongze Cheng 已提交
394
    szAmHdr = 0;
H
Hongze Cheng 已提交
395
  } else {
H
Hongze Cheng 已提交
396
    szAmHdr = sizeof(SBtPageHdr);
H
Hongze Cheng 已提交
397
  }
H
Hongze Cheng 已提交
398 399
  pPage->xCellSize = NULL;  // TODO

H
Hongze Cheng 已提交
400
  tdbPageInit(pPage, szAmHdr);
H
Hongze Cheng 已提交
401

H
Hongze Cheng 已提交
402 403
  TDB_BTREE_ASSERT_FLAG(flags);

H
Hongze Cheng 已提交
404
  if (isLeaf) {
H
Hongze Cheng 已提交
405 406 407 408 409 410 411 412 413 414
    pPage->kLen = pBt->keyLen;
    pPage->vLen = pBt->valLen;
    pPage->maxLocal = pBt->maxLeaf;
    pPage->minLocal = pBt->minLeaf;
  } else {
    pPage->kLen = pBt->keyLen;
    pPage->vLen = sizeof(SPgno);
    pPage->maxLocal = pBt->maxLocal;
    pPage->minLocal = pBt->minLocal;
  }
H
Hongze Cheng 已提交
415 416 417 418

  return 0;
}

H
Hongze Cheng 已提交
419
static int tdbBtreeZeroPage(SPage *pPage, void *arg) {
H
Hongze Cheng 已提交
420
  u16     flags;
H
Hongze Cheng 已提交
421
  SBTree *pBt;
H
Hongze Cheng 已提交
422
  u8      isLeaf;
H
Hongze Cheng 已提交
423
  u8      szAmHdr;
H
Hongze Cheng 已提交
424

H
Hongze Cheng 已提交
425 426
  flags = ((SBtreeZeroPageArg *)arg)->flags;
  pBt = ((SBtreeZeroPageArg *)arg)->pBt;
H
Hongze Cheng 已提交
427
  isLeaf = TDB_BTREE_PAGE_IS_LEAF(flags);
H
Hongze Cheng 已提交
428

H
Hongze Cheng 已提交
429
  if (isLeaf) {
H
Hongze Cheng 已提交
430
    szAmHdr = 0;
H
Hongze Cheng 已提交
431
  } else {
H
Hongze Cheng 已提交
432
    szAmHdr = sizeof(SBtPageHdr);
H
Hongze Cheng 已提交
433 434
  }
  pPage->xCellSize = NULL;  // TODO
H
Hongze Cheng 已提交
435

H
Hongze Cheng 已提交
436
  tdbPageZero(pPage, szAmHdr);
H
Hongze Cheng 已提交
437
  TDB_PAGE_FLAGS_SET(pPage, flags);
H
Hongze Cheng 已提交
438 439 440 441 442 443 444 445 446 447 448 449

  if (isLeaf) {
    pPage->kLen = pBt->keyLen;
    pPage->vLen = pBt->valLen;
    pPage->maxLocal = pBt->maxLeaf;
    pPage->minLocal = pBt->minLeaf;
  } else {
    pPage->kLen = pBt->keyLen;
    pPage->vLen = sizeof(SPgno);
    pPage->maxLocal = pBt->maxLocal;
    pPage->minLocal = pBt->minLocal;
  }
H
Hongze Cheng 已提交
450

H
Hongze Cheng 已提交
451
  return 0;
H
Hongze Cheng 已提交
452 453 454
}

#ifndef TDB_BTREE_BALANCE
H
Hongze Cheng 已提交
455 456 457 458
typedef struct {
  SBTree *pBt;
  SPage  *pParent;
  int     idx;
H
Hongze Cheng 已提交
459
  i8      nOld;
H
Hongze Cheng 已提交
460 461 462 463 464
  SPage  *pOldPages[3];
  i8      nNewPages;
  SPage  *pNewPages[5];
} SBtreeBalanceHelper;

H
Hongze Cheng 已提交
465
static int tdbBtreeCopyPageContent(SPage *pFrom, SPage *pTo) {
H
Hongze Cheng 已提交
466 467
#if 0
  int nCells = TDB_PAGE_TOTAL_CELLS(pFrom);
H
Hongze Cheng 已提交
468 469 470 471
  int cCells = TDB_PAGE_CCELLS(pFrom);
  int fCell = TDB_PAGE_FCELL(pFrom);
  int nFree = TDB_PAGE_NFREE(pFrom);

H
Hongze Cheng 已提交
472 473
  pTo->pFreeStart = pTo->pCellIdx + nCells * pFrom->pPageMethods->szOffset;
  memcpy(pTo->pCellIdx, pFrom->pCellIdx, nCells * pFrom->pPageMethods->szOffset);
H
Hongze Cheng 已提交
474 475
  pTo->pFreeEnd = (u8 *)pTo->pPageFtr - (u8 *)(pFrom->pPageFtr) + pFrom->pFreeEnd;
  memcpy(pTo->pFreeEnd, pFrom->pFreeEnd, (u8 *)pFrom->pPageFtr - pFrom->pFreeEnd);
H
Hongze Cheng 已提交
476 477 478 479 480 481 482

  TDB_PAGE_NCELLS_SET(pTo, nCells);
  TDB_PAGE_CCELLS_SET(pTo, cCells);
  TDB_PAGE_FCELL_SET(pTo, fCell);
  TDB_PAGE_NFREE_SET(pTo, nFree);

  // TODO: update other fields
H
Hongze Cheng 已提交
483
#endif
H
Hongze Cheng 已提交
484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510

  return 0;
}

static int tdbBtreeBalanceDeeper(SBTree *pBt, SPage *pRoot, SPage **ppChild) {
  SPager           *pPager;
  SPage            *pChild;
  SPgno             pgnoChild;
  int               ret;
  SBtreeZeroPageArg zArg;

  pPager = pRoot->pPager;

  // Allocate a new child page
  zArg.flags = TDB_BTREE_LEAF;
  zArg.pBt = pBt;
  ret = tdbPagerNewPage(pPager, &pgnoChild, &pChild, tdbBtreeZeroPage, &zArg);
  if (ret < 0) {
    return -1;
  }

  // Copy the root page content to the child page
  ret = tdbBtreeCopyPageContent(pRoot, pChild);
  if (ret < 0) {
    return -1;
  }

H
Hongze Cheng 已提交
511 512 513 514
  pChild->nOverflow = pRoot->nOverflow;
  for (int i = 0; i < pChild->nOverflow; i++) {
    pChild->apOvfl[i] = pRoot->apOvfl[i];
    pChild->aiOvfl[i] = pRoot->aiOvfl[i];
H
Hongze Cheng 已提交
515 516 517 518 519 520 521 522 523 524
  }

  // Reinitialize the root page
  zArg.flags = TDB_BTREE_ROOT;
  zArg.pBt = pBt;
  ret = tdbBtreeZeroPage(pRoot, &zArg);
  if (ret < 0) {
    return -1;
  }

H
Hongze Cheng 已提交
525 526
  ((SBtPageHdr *)pRoot->pAmHdr)[0].rChild = pgnoChild;

H
Hongze Cheng 已提交
527 528 529 530
  *ppChild = pChild;
  return 0;
}

H
Hongze Cheng 已提交
531
static int tdbBtreeBalanceStep1(SBtreeBalanceHelper *pBlh) {
H
Hongze Cheng 已提交
532 533 534 535 536 537 538 539 540 541
  int          nCells;
  int          i;
  int          idxStart;
  int          nChild;
  int          ret;
  SPage       *pParent;
  SPgno        pgno;
  SCell       *pCell;
  SCellDecoder cd;
  SBTree      *pBt;
H
Hongze Cheng 已提交
542 543

  pParent = pBlh->pParent;
H
Hongze Cheng 已提交
544
  nCells = TDB_PAGE_TOTAL_CELLS(pParent);
H
Hongze Cheng 已提交
545 546
  nChild = nCells + 1;
  pBt = pBlh->pBt;
H
Hongze Cheng 已提交
547

H
Hongze Cheng 已提交
548 549 550 551 552 553 554
  // TODO:  ASSERT below needs to be removed
  ASSERT(pParent->nOverflow == 0);
  ASSERT(pBlh->idx <= nCells);

  if (nChild < 3) {
    idxStart = 0;
    pBlh->nOld = nChild;
H
Hongze Cheng 已提交
555
  } else {
H
Hongze Cheng 已提交
556
    if (pBlh->idx == 0) {
H
Hongze Cheng 已提交
557 558 559
      idxStart = 0;
    } else if (pBlh->idx == nCells) {
      idxStart = pBlh->idx - 2;
H
Hongze Cheng 已提交
560
    } else {
H
Hongze Cheng 已提交
561
      idxStart = pBlh->idx - 1;
H
Hongze Cheng 已提交
562
    }
H
Hongze Cheng 已提交
563
    pBlh->nOld = 3;
H
Hongze Cheng 已提交
564
  }
H
Hongze Cheng 已提交
565

H
Hongze Cheng 已提交
566 567 568 569
  i = pBlh->nOld - 1;

  if (idxStart + i == nCells) {
    pgno = ((SBtPageHdr *)(pParent->pAmHdr))[0].rChild;
H
Hongze Cheng 已提交
570
  } else {
H
Hongze Cheng 已提交
571
    pCell = tdbPageGetCell(pParent, idxStart + i);
H
Hongze Cheng 已提交
572 573 574 575 576 577 578 579
    // TODO: no need to decode the payload part, and even the kLen, vLen part
    // we only need the pgno part
    ret = tdbBtreeDecodeCell(pParent, pCell, &cd);
    if (ret < 0) {
      ASSERT(0);
      return -1;
    }
    pgno = cd.pgno;
H
Hongze Cheng 已提交
580 581
  }
  for (;;) {
H
Hongze Cheng 已提交
582
    ret = tdbPagerFetchPage(pBt->pPager, pgno, &(pBlh->pOldPages[i]), tdbBtreeInitPage, pBt);
H
Hongze Cheng 已提交
583
    if (ret < 0) {
H
Hongze Cheng 已提交
584
      ASSERT(0);
H
Hongze Cheng 已提交
585
      return -1;
H
Hongze Cheng 已提交
586
    }
H
Hongze Cheng 已提交
587

H
Hongze Cheng 已提交
588
    // Loop over
H
Hongze Cheng 已提交
589 590
    if ((i--) == 0) break;

H
Hongze Cheng 已提交
591 592 593
    {
      // TODO
      // ASSERT(0);
H
Hongze Cheng 已提交
594 595 596
    }
  }

H
Hongze Cheng 已提交
597 598 599 600
  return 0;
}

static int tdbBtreeBalanceStep2(SBtreeBalanceHelper *pBlh) {
H
Hongze Cheng 已提交
601
#if 0
H
Hongze Cheng 已提交
602
  SPage *pPage;
H
Hongze Cheng 已提交
603 604 605 606
  int    oidx;
  int    cidx;
  int    limit;
  SCell *pCell;
H
Hongze Cheng 已提交
607

H
Hongze Cheng 已提交
608
  for (int i = 0; i < pBlh->nOld; i++) {
H
Hongze Cheng 已提交
609
    pPage = pBlh->pOldPages[i];
H
Hongze Cheng 已提交
610 611 612 613 614 615 616 617 618 619 620 621 622 623 624
    oidx = 0;
    cidx = 0;

    if (oidx < pPage->nOverflow) {
      limit = pPage->aiOvfl[oidx];
    } else {
      limit = pPage->pPageHdr->nCells;
    }

    // Loop to copy each cell pointer out
    for (;;) {
      if (oidx >= pPage->nOverflow && cidx >= pPage->pPageHdr->nCells) break;

      if (cidx < limit) {
        // Get local cells
H
Hongze Cheng 已提交
625
        pCell = tdbPageGetCell(pPage, cidx);
H
Hongze Cheng 已提交
626 627 628 629 630 631 632 633 634 635 636 637 638
      } else if (cidx == limit) {
        // Get overflow cells
        pCell = pPage->apOvfl[oidx++];

        if (oidx < pPage->nOverflow) {
          limit = pPage->aiOvfl[oidx];
        } else {
          limit = pPage->pPageHdr->nCells;
        }
      } else {
        ASSERT(0);
      }
    }
H
Hongze Cheng 已提交
639 640 641 642

    {
      // TODO: Copy divider cells here
    }
H
Hongze Cheng 已提交
643 644 645 646
  }

  /* TODO */

H
Hongze Cheng 已提交
647
#endif
H
Hongze Cheng 已提交
648 649 650 651
  return 0;
}

static int tdbBtreeBalanceStep3(SBtreeBalanceHelper *pBlh) {
H
Hongze Cheng 已提交
652
  // Figure out number of pages needed after balance
H
Hongze Cheng 已提交
653
  for (int i = 0; i < pBlh->nOld; i++) {
H
Hongze Cheng 已提交
654
    /* TODO */
H
Hongze Cheng 已提交
655 656
  }

H
Hongze Cheng 已提交
657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678
  return 0;
}

static int tdbBtreeBalanceStep4(SBtreeBalanceHelper *pBlh) {
  // TODO
  return 0;
}

static int tdbBtreeBalanceStep5(SBtreeBalanceHelper *pBlh) {
  // TODO
  return 0;
}

static int tdbBtreeBalanceStep6(SBtreeBalanceHelper *pBlh) {
  // TODO
  return 0;
}

static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx) {
  int                 ret;
  SBtreeBalanceHelper blh;

H
Hongze Cheng 已提交
679 680
  ASSERT(!TDB_BTREE_PAGE_IS_LEAF(TDB_PAGE_FLAGS(pParent)));

H
Hongze Cheng 已提交
681 682 683 684 685 686 687 688 689
  blh.pBt = pBt;
  blh.pParent = pParent;
  blh.idx = idx;

  // Step 1: find two sibling pages and get engough info about the old pages
  ret = tdbBtreeBalanceStep1(&blh);
  if (ret < 0) {
    ASSERT(0);
    return -1;
H
Hongze Cheng 已提交
690 691
  }

H
Hongze Cheng 已提交
692 693 694 695 696
  // Step 2: Load all cells on the old page and the divider cells
  ret = tdbBtreeBalanceStep2(&blh);
  if (ret < 0) {
    ASSERT(0);
    return -1;
H
Hongze Cheng 已提交
697 698
  }

H
Hongze Cheng 已提交
699 700 701 702 703
  // Step 3: Get the number of pages needed to hold all cells
  ret = tdbBtreeBalanceStep3(&blh);
  if (ret < 0) {
    ASSERT(0);
    return -1;
H
Hongze Cheng 已提交
704 705
  }

H
Hongze Cheng 已提交
706 707 708 709 710
  // Step 4: Allocate enough new pages. Reuse old pages as much as possible
  ret = tdbBtreeBalanceStep4(&blh);
  if (ret < 0) {
    ASSERT(0);
    return -1;
H
Hongze Cheng 已提交
711 712
  }

H
Hongze Cheng 已提交
713 714 715 716 717 718 719 720 721 722 723 724
  // Step 5: Insert new divider cells into pParent
  ret = tdbBtreeBalanceStep5(&blh);
  if (ret < 0) {
    ASSERT(0);
    return -1;
  }

  // Step 6: Update the sibling pages
  ret = tdbBtreeBalanceStep6(&blh);
  if (ret < 0) {
    ASSERT(0);
    return -1;
H
Hongze Cheng 已提交
725 726 727 728 729 730 731 732 733
  }

  {
      // TODO: Reset states
  }

  {
    // TODO: Clear resources
  }
H
Hongze Cheng 已提交
734

H
Hongze Cheng 已提交
735 736 737 738 739
  return 0;
}

static int tdbBtreeBalance(SBtCursor *pCur) {
  int    iPage;
H
Hongze Cheng 已提交
740
  SPage *pParent;
H
Hongze Cheng 已提交
741
  SPage *pPage;
H
Hongze Cheng 已提交
742
  int    ret;
H
Hongze Cheng 已提交
743 744 745
  u16    flags;
  u8     leaf;
  u8     root;
H
Hongze Cheng 已提交
746 747 748 749

  // Main loop to balance the BTree
  for (;;) {
    iPage = pCur->iPage;
H
Hongze Cheng 已提交
750 751 752 753
    pPage = pCur->pPage;
    flags = TDB_PAGE_FLAGS(pPage);
    leaf = TDB_BTREE_PAGE_IS_LEAF(flags);
    root = TDB_BTREE_PAGE_IS_ROOT(flags);
H
Hongze Cheng 已提交
754 755 756 757 758 759 760 761

    // TODO: Get the page free space if not get yet
    // if (pPage->nFree < 0) {
    //   if (tdbBtreeComputeFreeSpace(pPage) < 0) {
    //     return -1;
    //   }
    // }

H
Hongze Cheng 已提交
762 763
    // when the page is not overflow and not too empty, the balance work
    // is finished. Just break out the balance loop.
H
Hongze Cheng 已提交
764
    if (pPage->nOverflow == 0 /* TODO: && pPage->nFree <= */) {
H
Hongze Cheng 已提交
765 766 767 768
      break;
    }

    if (iPage == 0) {
H
Hongze Cheng 已提交
769 770 771
      // For the root page, only balance when the page is overfull,
      // ignore the case of empty
      if (pPage->nOverflow == 0) break;
H
Hongze Cheng 已提交
772 773 774 775 776 777 778 779 780 781 782

      ret = tdbBtreeBalanceDeeper(pCur->pBt, pCur->pPage, &(pCur->pgStack[1]));
      if (ret < 0) {
        return -1;
      }

      pCur->idx = 0;
      pCur->idxStack[0] = 0;
      pCur->pgStack[0] = pCur->pPage;
      pCur->iPage = 1;
      pCur->pPage = pCur->pgStack[1];
H
Hongze Cheng 已提交
783
    } else {
H
Hongze Cheng 已提交
784
      // Generalized balance step
H
Hongze Cheng 已提交
785
      pParent = pCur->pgStack[iPage - 1];
H
Hongze Cheng 已提交
786

H
Hongze Cheng 已提交
787
      ret = tdbBtreeBalanceNonRoot(pCur->pBt, pParent, pCur->idxStack[pCur->iPage - 1]);
H
Hongze Cheng 已提交
788 789 790 791 792 793
      if (ret < 0) {
        return -1;
      }

      pCur->iPage--;
      pCur->pPage = pCur->pgStack[pCur->iPage];
H
Hongze Cheng 已提交
794 795 796 797 798
    }
  }

  return 0;
}
H
Hongze Cheng 已提交
799 800
#endif

H
Hongze Cheng 已提交
801
#ifndef TDB_BTREE_CELL  // =========================================================
H
Hongze Cheng 已提交
802
static int tdbBtreeEncodePayload(SPage *pPage, u8 *pPayload, const void *pKey, int kLen, const void *pVal, int vLen,
H
Hongze Cheng 已提交
803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831
                                 int *szPayload) {
  int nPayload;

  ASSERT(pKey != NULL);

  if (pVal == NULL) {
    vLen = 0;
  }

  nPayload = kLen + vLen;
  if (nPayload <= pPage->maxLocal) {
    // General case without overflow
    memcpy(pPayload, pKey, kLen);
    if (pVal) {
      memcpy(pPayload + kLen, pVal, vLen);
    }

    *szPayload = nPayload;
    return 0;
  }

  {
    // TODO: handle overflow case
    ASSERT(0);
  }

  return 0;
}

H
Hongze Cheng 已提交
832 833
static int tdbBtreeEncodeCell(SPage *pPage, const void *pKey, int kLen, const void *pVal, int vLen, SCell *pCell,
                              int *szCell) {
H
Hongze Cheng 已提交
834
  u16 flags;
H
Hongze Cheng 已提交
835 836 837 838
  u8  leaf;
  int nHeader;
  int nPayload;
  int ret;
H
Hongze Cheng 已提交
839 840 841 842

  ASSERT(pPage->kLen == TDB_VARIANT_LEN || pPage->kLen == kLen);
  ASSERT(pPage->vLen == TDB_VARIANT_LEN || pPage->vLen == vLen);

H
Hongze Cheng 已提交
843 844 845 846
  nPayload = 0;
  nHeader = 0;
  flags = TDB_PAGE_FLAGS(pPage);
  leaf = TDB_BTREE_PAGE_IS_LEAF(flags);
H
Hongze Cheng 已提交
847

H
Hongze Cheng 已提交
848 849
  // 1. Encode Header part
  /* Encode kLen if need */
H
Hongze Cheng 已提交
850
  if (pPage->kLen == TDB_VARIANT_LEN) {
H
Hongze Cheng 已提交
851
    nHeader += tdbPutVarInt(pCell + nHeader, kLen);
H
Hongze Cheng 已提交
852 853
  }

H
Hongze Cheng 已提交
854
  /* Encode vLen if need */
H
Hongze Cheng 已提交
855
  if (pPage->vLen == TDB_VARIANT_LEN) {
H
Hongze Cheng 已提交
856
    nHeader += tdbPutVarInt(pCell + nHeader, vLen);
H
Hongze Cheng 已提交
857 858
  }

H
Hongze Cheng 已提交
859 860
  /* Encode SPgno if interior page */
  if (!leaf) {
H
Hongze Cheng 已提交
861 862
    ASSERT(pPage->vLen == sizeof(SPgno));

H
Hongze Cheng 已提交
863 864 865 866 867 868 869 870 871 872 873 874 875
    ((SPgno *)(pCell + nHeader))[0] = ((SPgno *)pVal)[0];
    nHeader = nHeader + sizeof(SPgno);
  }

  // 2. Encode payload part
  if (leaf) {
    ret = tdbBtreeEncodePayload(pPage, pCell + nHeader, pKey, kLen, pVal, vLen, &nPayload);
  } else {
    ret = tdbBtreeEncodePayload(pPage, pCell + nHeader, pKey, kLen, NULL, 0, &nPayload);
  }
  if (ret < 0) {
    // TODO: handle error
    return -1;
H
Hongze Cheng 已提交
876 877
  }

H
Hongze Cheng 已提交
878
  *szCell = nHeader + nPayload;
H
Hongze Cheng 已提交
879 880 881
  return 0;
}

H
Hongze Cheng 已提交
882 883 884 885
static int tdbBtreeDecodePayload(SPage *pPage, const u8 *pPayload, SCellDecoder *pDecoder) {
  int nPayload;

  ASSERT(pDecoder->pKey == NULL);
H
Hongze Cheng 已提交
886

H
Hongze Cheng 已提交
887 888
  if (pDecoder->pVal) {
    nPayload = pDecoder->kLen + pDecoder->vLen;
H
Hongze Cheng 已提交
889
  } else {
H
Hongze Cheng 已提交
890
    nPayload = pDecoder->kLen;
H
Hongze Cheng 已提交
891 892
  }

H
Hongze Cheng 已提交
893 894 895
  if (nPayload <= pPage->maxLocal) {
    // General case without overflow
    pDecoder->pKey = (void *)pPayload;
H
Hongze Cheng 已提交
896 897
    if (!pDecoder->pVal) {
      pDecoder->pVal = (void *)(pPayload + pDecoder->kLen);
H
Hongze Cheng 已提交
898
    }
H
Hongze Cheng 已提交
899
  } else {
H
Hongze Cheng 已提交
900 901
    // TODO: handle overflow case
    ASSERT(0);
H
Hongze Cheng 已提交
902 903
  }

H
Hongze Cheng 已提交
904 905 906 907 908 909 910 911 912 913
  return 0;
}

static int tdbBtreeDecodeCell(SPage *pPage, const SCell *pCell, SCellDecoder *pDecoder) {
  u16 flags;
  u8  leaf;
  int nHeader;
  int ret;

  nHeader = 0;
H
Hongze Cheng 已提交
914
  flags = TDB_PAGE_FLAGS(pPage);
H
Hongze Cheng 已提交
915 916
  leaf = TDB_BTREE_PAGE_IS_LEAF(flags);

H
Hongze Cheng 已提交
917 918 919 920 921 922 923
  // Clear the state of decoder
  pDecoder->kLen = -1;
  pDecoder->pKey = NULL;
  pDecoder->vLen = -1;
  pDecoder->pVal = NULL;
  pDecoder->pgno = 0;

H
Hongze Cheng 已提交
924 925 926 927 928 929 930 931 932
  // 1. Decode header part
  if (pPage->kLen == TDB_VARIANT_LEN) {
    nHeader += tdbGetVarInt(pCell + nHeader, &(pDecoder->kLen));
  } else {
    pDecoder->kLen = pPage->kLen;
  }

  if (pPage->vLen == TDB_VARIANT_LEN) {
    nHeader += tdbGetVarInt(pCell + nHeader, &(pDecoder->vLen));
H
Hongze Cheng 已提交
933
  } else {
H
Hongze Cheng 已提交
934 935
    pDecoder->vLen = pPage->vLen;
  }
H
Hongze Cheng 已提交
936

H
Hongze Cheng 已提交
937 938
  if (!leaf) {
    ASSERT(pPage->vLen == sizeof(SPgno));
H
Hongze Cheng 已提交
939

H
Hongze Cheng 已提交
940 941 942 943 944 945 946 947 948
    pDecoder->pgno = ((SPgno *)(pCell + nHeader))[0];
    pDecoder->pVal = (u8 *)(&(pDecoder->pgno));
    nHeader = nHeader + sizeof(SPgno);
  }

  // 2. Decode payload part
  ret = tdbBtreeDecodePayload(pPage, pCell + nHeader, pDecoder);
  if (ret < 0) {
    return -1;
H
Hongze Cheng 已提交
949 950 951 952 953
  }

  return 0;
}

H
Hongze Cheng 已提交
954
#endif