-
unknown authored
* Recovery of the table's live checksum (CREATE TABLE ... CHECKSUM=1) is achieved in this patch. The table's live checksum (info->s->state.state.checksum) is updated in inwrite_rec_hook's under the log mutex when writing UNDO_ROW_INSERT|UPDATE|DELETE and REDO_DELETE_ALL. The checksum variation caused by the operation is stored in these UNDOs, so that the REDO phase, when it sees such UNDOs, can update the live checksum if it is older (state.is_of_lsn is lower) than the record. It is also used, as a nice add-on with no cost, to do less row checksum computation during the UNDO phase (as we have it in the record already). Doing this work, it became pressing to move in-write hooks (write_hook_for_redo() et al) to ma_blockrec.c. The 'parts' argument of inwrite_rec_hook is unpredictable (it comes mangled at this stage, for example by LSN compression) so it is replaced by a 'void* hook_arg', which is used to pass down information, currently only to write_hook_for_clr_end() (previous undo_lsn and type of undone record). * If from ha_maria, we print to stderr how many seconds (with one fractional digit) the REDO phase took, same for UNDO phase and for final table close. Just to give an indication for debugging and maybe also for Support. storage/maria/ha_maria.cc: question for Monty storage/maria/ma_blockrec.c: * log in-write hooks (write_hook_for_redo() etc) move from ma_loghandler.c to here; this is natural: the hooks are coupled to their callers (functions in ma_blockrec.c). * translog_write_record() now has a new argument "hook_arg"; using it to pass down to write_hook_for_clr_end() the transaction's previous_undo_lsn and the type of the being undone record, and also to pass down to all UNDOs the live checksum variation caused by the operation. * If table has live checksum, store in UNDO_ROW_INSERT|UPDATE|DELETE and in CLR_END the checksum variation ("delta") caused by the operation. For example if a DELETE caused the table's live checksum to change from 123 to 456, we store in the UNDO_ROW_DELETE, in 4 bytes, the value 333 (456-123). * Instead of hard-coded "1" as length of the place where we store the undone record's type in CLR_END, use a symbol CLR_TYPE_STORE_SIZE; use macros clr_type_store and clr_type_korr. * write_block_record() has a new parameter 'old_record_checksum' which is the pre-computed checksum of old_record; that value is used to update the table's live checksum when writing UNDO_ROW_UPDATE|CLR_END. * In allocate_write_block_record(), if we are executing UNDO_ROW_DELETE the row's checksum is already computed. * _ma_update_block_record2() now expect the new row's checksum into cur_row.checksum (was already true) and the old row's checksum into new_row.checksum (that's new). Its two callers, maria_update() and _ma_apply_undo_row_update(), honour this. * When executing an UNDO_ROW_INSERT|UPDATE|DELETE in UNDO phase, pick up the checksum delta from the log record. It is then used to update the table's live checksum when writing CLR_END, and saves us a computation of record. storage/maria/ma_blockrec.h: in-write hooks move from ma_loghandler.c storage/maria/ma_check.c: more straightforward size of buffer storage/maria/ma_checkpoint.c: <= is enough storage/maria/ma_commit.c: new prototype of translog_write_record() storage/maria/ma_create.c: new prototype of translog_write_record() storage/maria/ma_delete.c: The row's checksum must be computed before calling(*delete_record)(), not after, because it must be known inside _ma_delete_block_record() (to update the table's live checksum when writing UNDO_ROW_DELETE). If deleting from a transactional table, live checksum was already updated when writing UNDO_ROW_DELETE. storage/maria/ma_delete_all.c: @todo is now done (in ma_loghandler.c) storage/maria/ma_delete_table.c: new prototype of translog_write_record() storage/maria/ma_loghandler.c: * in-write hooks move to ma_blockrec.c. * translog_write_record() gets a new argument 'hook_arg' which is passed down to pre|inwrite_rec_hook. It is more useful that 'parts' for those hooks, because when those hooks are called, 'parts' has possibly been mangled (like with LSN compression) and is so unpredictable. * fix for compiler warning (unused buffer_start when compiling without debug support) * Because checksum delta is stored into UNDO_ROW_INSERT|UPDATE|DELETE and CLR_END, but only if the table has live checksum, these records are not PSEUDOFIXEDLENGTH anymore, they are now VARIABLE_LENGTH (their length is X if no live checksum and X+4 otherwise). * add an inwrite_rec_hook for UNDO_ROW_UPDATE, which updates the table's live checksum. Update it also in hooks of UNDO_ROW_INSERT| DELETE and REDO_DELETE_ALL and CLR_END. * Bugfix: when reading a record in translog_read_record(), it happened that "length" became negative, because the function assumed that the record extended beyond the page's end, whereas it may be shorter. storage/maria/ma_loghandler.h: * Instead of hard-coded "1" and "4", use symbols and macros to store/retrieve the type of record which the CLR_END corresponds to, and the checksum variation caused by the operation which logs the record * translog_write_record() gets a new argument 'hook_arg' which is passed down to pre|inwrite_rec_hook. It is more useful that 'parts' for those hooks, because when those hooks are called, 'parts' has possibly been mangled (like with LSN compression) and is so unpredictable. storage/maria/ma_open.c: fix for "empty body in if() statement" (when compiling without safemutex) storage/maria/ma_pagecache.c: <= is enough storage/maria/ma_recovery.c: * print the time that each recovery phase (REDO/UNDO/flush) took; this is enabled only when recovering from ha_maria. Is it printed n seconds with a fractional part of one digit (like 123.4 seconds). * In the REDO phase, update the table's live checksum by using the checksum delta stored in UNDO_ROW_INSERT|DELETE|UPDATE and CLR_END. Update it too when seeing REDO_DELETE_ALL. * In the UNDO phase, when executing UNDO_ROW_INSERT, if the table does not have live checksum then reading the record's header (as done by the master loop of run_undo_phase()) is enough; otherwise we do a translog_read_record() to have the checksum delta ready for _ma_apply_undo_row_insert(). * When at the end of the REDO phase we notice that there is an unfinished group of REDOs, don't assert in debug binaries, as I verified that it can happen in real life (with kill -9) * removing ' in #error as it confuses gcc3 storage/maria/ma_rename.c: new prototype of translog_write_record() storage/maria/ma_test_recovery.expected: Change in output of ma_test_recovery: now all live checksums of original tables equal those of tables recreated by the REDO phase and those of tables fixed by the UNDO phase. I.e. recovery of the live checksum looks like working (which was after all the only goal of this changeset). I checked by hand that it's not just all live checksums which are now 0 and that's why they match. They are the old values like 3757530372. maria.test has hard-coded checksum values in its result file so checks this too. storage/maria/ma_update.c: * It's useless to put up HA_STATE_CHANGED in 'key_changed', as we put up HA_STATE_CHANGED in info->update anyway. * We need to compute the old and new rows' checksum before calling (*update_record)(), as checksum delta must be known when logging UNDO_ROW_UPDATE which is done by _ma_update_block_record(). Note that some functions change the 'newrec' record (at least _ma_check_unique() does) so we cannot move the checksum computation too early in the function. storage/maria/ma_write.c: If inserting into a transactional table, live's checksum was already updated when writing UNDO_ROW_INSERT. The multiplication is a trick to save an if(). storage/maria/unittest/ma_test_loghandler-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_first_lsn-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_max_lsn-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_multigroup-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_multithread-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_noflush-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_pagecache-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_purge-t.c: new prototype of translog_write_record() storage/myisam/sort.c: fix for compiler warnings in pushbuild (write_merge_key* functions didn't have their declaration match MARIA_HA::write_key).
198a1b6f