SQLite

Check-in [95eaa49f4e]
Login

Many hyperlinks are disabled.
Use anonymous login to enable hyperlinks.

Overview
Comment:A different approach to preventing buffer overreads when comparing a vector of values with a corrupt index record that spans at least one overflow page.
Downloads: Tarball | ZIP archive
Timelines: family | ancestors | descendants | both | trunk
Files: files | file ages | folders
SHA1: 95eaa49f4ee071c7a0f690a695f1f2d1ebca68a8
User & Date: drh 2015-05-27 15:42:53.434
Context
2015-05-27
18:13
All bytes of a page are initialized, even bytes that follow a 2-byte cell on a malformed page. (check-in: 7373a004ed user: drh tags: trunk)
15:42
A different approach to preventing buffer overreads when comparing a vector of values with a corrupt index record that spans at least one overflow page. (check-in: 95eaa49f4e user: drh tags: trunk)
15:10
Disallow the use of "rowid" in CTEs - it has never worked correctly and it makes no sense, so we might as well make it an explicit error. Also: add the PRAGMA cell_size_check=ON command. (check-in: 19e2cebc12 user: drh tags: trunk)
14:21
A different approach to preventing buffer overreads when comparing a vector of values with a corrupt index record that spans at least one overflow page. (Closed-Leaf check-in: 7e9e1b6123 user: dan tags: avoid-buffer-overread)
Changes
Unified Diff Ignore Whitespace Patch
Changes to src/btree.c.
4938
4939
4940
4941
4942
4943
4944
4945





4946
4947
4948
4949





4950
4951
4952
4953
4954
4955
4956
4957
          ** fits entirely on the main b-tree page.  */
          testcase( pCell+nCell+2==pPage->aDataEnd );
          c = xRecordCompare(nCell, (void*)&pCell[2], pIdxKey);
        }else{
          /* The record flows over onto one or more overflow pages. In
          ** this case the whole cell needs to be parsed, a buffer allocated
          ** and accessPayload() used to retrieve the record into the
          ** buffer before VdbeRecordCompare() can be called. */





          void *pCellKey;
          u8 * const pCellBody = pCell - pPage->childPtrSize;
          btreeParseCellPtr(pPage, pCellBody, &pCur->info);
          nCell = (int)pCur->info.nKey;





          pCellKey = sqlite3Malloc( nCell );
          if( pCellKey==0 ){
            rc = SQLITE_NOMEM;
            goto moveto_finish;
          }
          pCur->aiIdx[pCur->iPage] = (u16)idx;
          rc = accessPayload(pCur, 0, nCell, (unsigned char*)pCellKey, 2);
          if( rc ){







|
>
>
>
>
>




>
>
>
>
>
|







4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
          ** fits entirely on the main b-tree page.  */
          testcase( pCell+nCell+2==pPage->aDataEnd );
          c = xRecordCompare(nCell, (void*)&pCell[2], pIdxKey);
        }else{
          /* The record flows over onto one or more overflow pages. In
          ** this case the whole cell needs to be parsed, a buffer allocated
          ** and accessPayload() used to retrieve the record into the
          ** buffer before VdbeRecordCompare() can be called. 
          **
          ** If the record is corrupt, the xRecordCompare routine may read
          ** up to two varints past the end of the buffer. An extra 18 
          ** bytes of padding is allocated at the end of the buffer in
          ** case this happens.  */
          void *pCellKey;
          u8 * const pCellBody = pCell - pPage->childPtrSize;
          btreeParseCellPtr(pPage, pCellBody, &pCur->info);
          nCell = (int)pCur->info.nKey;
          testcase( nCell<0 );
          if( nCell<2 ){
            rc = SQLITE_CORRUPT_BKPT;
            goto moveto_finish;
          }
          pCellKey = sqlite3Malloc( nCell+18 );
          if( pCellKey==0 ){
            rc = SQLITE_NOMEM;
            goto moveto_finish;
          }
          pCur->aiIdx[pCur->iPage] = (u16)idx;
          rc = accessPayload(pCur, 0, nCell, (unsigned char*)pCellKey, 2);
          if( rc ){