PIM/MS Windows/SQLite Folder Indices

From KDE Community Wiki

There are issues with locking index files for KMail folders and mmap()/munmap() operations on Windows. Therefore, SQLite-based indices are in development. This page presents detailed development notes for this task.

Started: jstaniek 11:35, 23 April 2008 (CEST)

Introduction

  • we call the new implementation SQLite mode for short
  • SQLite 3.5.4 is used, as provided by emerge sqlite module; we should not allow using much older versions of sqlite, e.g. 3.1 because of file format differences
  • we are using one .index.db file per account, not folder (NOTE: currently that's per folder...)
  • kmailprivate links to sqlite library for SQLite mode, and KMAIL_SQLITE_INDEX is defined to enable #ifdef'd code
  • kmfolderindex_sqlite.cpp is created and edited as a copy of kmfolderindex.cpp; kmfolderindex.cpp #includes kmfolderindex_sqlite.cpp for SQLite mode and then skips its own code completely
  • kmfolderindex_common.cpp is always included by kmfolderindex.cpp; implements KMFolderIndex::openInternal() and KMFolderIndex::createInternal()
  • kmfolderindex.h is a common header for both kmfolderindex*.cpp implementations

KMFolderIndex

api docs

  • 2008-04-23..25
    • mIndexId unused - removed as well as serialIndexId()
    • indexLocation(): added .db suffix to indicate the index is sqlite-based, implementation moved to FolderStorage (before FolderStorage only had it as abstract method)
    • INDEX_VERSION is written and checked using 'PRAGMA user_version = <integer>' command [1]
    • we do not use temporary filenames, e.g. in writeIndex(): SQLite takes care about safe storage
    • updateIndex(): no changes, we're changing implementation of KMMsgBase::syncIndexString() and writeIndex() instead
  • added openDatabase( int mode ) for SQLite mode
  • readIndex() implemented for SQLite mode - uses SELECT command on messages table
  • readIndexHeader() uses "PRAGMA user_version" sqlite command
  • writeIndex() implemented for SQLite mode - uses INSERT command on messages table, within transaction
  • common code from {KMFolderMbox|KMFolderMaildir}::open( const char * ) moved to KMFolderIndex::openInternal()
  • common code from {KMFolderMbox|KMFolderMaildir}::create() moved to KMFolderIndex::createInternal()

FolderStorage

  • as implementation of KMFolderIndex::indexLocation() is moved to FolderStorage (before FolderStorage only had it as abstract method); FolderStorage::idsLocation() and FolderStorage::sortedLocation() are added to avoid performing the math like mFolder->indexLocation() + ".sorted"
  • added QString location(const QString& suffix) - returns full path to .index, .ids or .sorted file (depending on the suffix)

KMMsgBase

api docs

  • 2008-04-23
    • fixed memory leak in KMMsgBase::asIndexString() by adding K_GLOBAL_STATIC for uchar* ret

Status of porting to SQLite

TOPIC PORTED TESTED NOTES
QString KMFolderIndex::indexLocation() yes added .db suffix to indicate the index is sqlite-based
int KMFolderIndex::updateIndex() yes no changes
int KMFolderIndex::writeIndex( bool createEmptyIndex ) yes creates db; creates tables messages table, insert messages to messages table, encoded as blobs using KMMsgBase::asIndexString(); header is not needed, but INDEX_VERSION is saved using PRAGMA user_version; byte order info is not saved: every integer is written using network order or as string
bool KMFolderIndex::readIndex()
int KMFolderIndex::count(bool cache) yes no changes
bool KMFolderIndex::readIndexHeader(int *gv)
bool KMFolderIndex::updateIndexStreamPtr(bool)
KMFolderIndex::IndexStatus KMFolderIndex::indexStatus()
void KMFolderIndex::truncateIndex() use "DELETE FROM..."
void KMFolderIndex::fillMessageDict() yes no change as it just inserts messages from KMFolderIndex::mMsgList into KMMsgDict
KMMsgInfo* KMFolderIndex::setIndexEntry( int idx, KMMessage *msg ) yes no change as it just sets creates a new KMMsgInfo object and inserts it into KMFolderIndex::mMsgList
bool KMFolderIndex::recreateIndex() yes no changes as it just calls createIndexFromContents() and readIndex()
off_t KMFolderIndex::mHeaderOffset replace its public use (e.g. in KMFolderIndex::truncateIndex()) with additional bool indexOpened()
FILE* KMFolderIndex::mIndexStream, uchar* mIndexStreamPtr, size_t mIndexStreamPtrLength, bool mIndexSwapByteOrder, int mIndexSizeOfLong these members are unused for SQLite mode because are related to file strorage; moreover byte order and size of long is handled by SQLite in a portable way
bool KMMsgBase::syncIndexString() const use 'UPDATE' SQL command to store the message serialized to a BLOB; KMMsgBase::getStringPart() will read it
QString KMMsgBase::getStringPart(MsgPartType t) const use 'SELECT' SQL command, harmonize with syncIndexString() implementation

Open Questions

  • should we port .sorted files to sqlite too? (possibly to the same .db file)