/**************************************************************** * * * Copyright 2001, 2011 Fidelity Information Services, Inc * * * * This source code contains the intellectual property * * of its copyright holder(s), and is made available * * under a license. If you do not know the terms of * * the license, please stop and do not read further. * * * ****************************************************************/ #ifndef __GDSCC_H__ #define __GDSCC_H__ /* this requires gdsroot.h gtm_facilit.h fileinfo.h gdsbt.h gdsfhead.h gdir.h gdskey.h */ #include /* BIG_UA is the maximum size of a single update array specified as an unsigned quantity (usages rely on this). It is 16MB. */ #define BIG_UA (uint4)16777216 #define CDB_R_SET_SIZE 32 #define CDB_CW_SET_SIZE (MAX_BT_DEPTH * 3 + 1 + 2) #define CDB_W_SET_SIZE 16 /* CDB_CW_SET_SIZE = 24 * 3 for all the levels (including updated block, newly created sibling and possible bitmap update) * 1 extra for the root level (to take care of gds_t_write_root case) * 2 in the case of creation of a new global variable (1 index block with a * key and 1 data block * containing the key) */ #define CDB_T_CREATE 0 #define CDB_T_WRITE 1 #define CDB_T_WRITE_ROOT 2 /* The following defines the write_type for a block that is going to be updated. * GDS_WRITE_PLAIN is the default type for most updates. * GDS_WRITE_BLOCK_SPLIT is set in case of a block update due to a block split. It is currently not used anywhere in the code. * GDS_WRITE_KILLTN requires a little more explanation. * * The TP commit logic ("tp_tend") makes use of an optimization referred to as the "indexmod" optimization. * This optimization tries to avoid a restart in the case where a TP transaction does a SET to a data block and later finds * at TCOMMIT time that the index block which was part of the SET had been updated by a concurrent SET (or a REORG split * operation) to a different data block (that also had the same index block as an ancestor) which resulted in a block split * causing the index block to be updated. In this case there is no reason to restart. The index block could have been * modified by other operations as well (e.g. M-kill, REORG coalesce or swap operations or any DSE command or a block split * operation that caused the height of the global variable tree to increase [C9B11-001813]). In these cases, we dont want * this optimization to take effect as we cant be sure everything that was relied upon for the TP transaction was still valid. * These disallowed operations are generically referred to as "kill" type of operations. This optimization is implemented by * having a field "killtn" (name derived from "kill" type of operations) in the bt (block-table) structure for each block. * This field is assigned the same value as the "tn" whenever an index block gets updated due to one of the disallowed operations. * Otherwise it stays untouched (i.e. killtn <= tn at all times). It is the "killtn" (and not "tn") that is used in the * cdb_sc_blkmod validation check in tp_tend if we are validating a index block. Since KILLs, MUPIP REORG and/or DSE operations * are usually rare compared to SET activity, most of the cases we expect the indexmod optimization to be in effect and * therefore help reduce the # of TP restarts due to index block changes. Note that each SET/GET/KILL operation in TP goes * through an intermediate validation routine tp_hist which does the cdb_sc_blkmod validation using "tn" (not "killtn"). * Only if that passed, do we relax the commit time validation for index blocks. * * The operations that are not allowed to use this optimization (M-kill, REORG or DSE) are supposed to make sure they * set the write_type of the cw-set-element to GDS_WRITE_KILLTN. Failing to do so cause "killtn" in the bt to NOT be uptodate * which in turn can cause false validation passes (in the cdb_sc_blkmod check) causing GT.M processes to incorrectly commit * when they should not. This can lead to GT.M/application level data integrity errors. */ #define GDS_WRITE_PLAIN 0 #define GDS_WRITE_KILLTN 1 #define GDS_WRITE_BLOCK_SPLIT 2 /* macro to traverse to the end of an horizontal cw_set_element list */ #define TRAVERSE_TO_LATEST_CSE(x) \ { \ GBLREF uint4 dollar_tlevel; \ \ assert(dollar_tlevel); \ if (x) \ for ( ; (x)->high_tlevel; x = (x)->high_tlevel) \ ; \ } typedef uint4 block_offset; typedef int4 block_index; /* If a new mode is added to the table below, make sure pre-existing mode usages in the current codebase are examined to see * if the new mode needs to be added there as well. For example, there is code in tp_incr_commit.c and tp_incr_clean_up.c * where gds_t_create and kill_t_create are used explicitly. If the new mode is yet another *create* type, then it might need * to be added in those places as well. */ enum gds_t_mode { gds_t_noop = 0, /* there is code that initializes stuff to 0 relying on it being equal to gds_t_noop */ gds_t_create, gds_t_write, gds_t_write_recycled, /* modify a recycled block (currently only done by MUPIP REORG UPGRADE/DOWNGRADE) */ gds_t_acquired, gds_t_writemap, gds_t_committed, /* t_end relies on this particular placement */ gds_t_write_root, /* t_end relies on this being AFTER gds_t_committed */ gds_t_busy2free, /* t_end relies on this being AFTER gds_t_committed */ n_gds_t_op, /* tp_tend and other routines rely on this being BEFORE kill_t* modes and AFTER all gds_t_* modes */ kill_t_create, /* tp_tend relies on this being AFTER n_gds_t_op */ kill_t_write, /* tp_tend relies on this being AFTER n_gds_t_op */ }; typedef struct key_value_struct { gv_key key; /* note that the following array holds the actual key contents */ char key_contents[DBKEYSIZE(MAX_KEY_SZ)]; mstr value; struct key_value_struct *next; } key_cum_value; /* Create/write set element. This is used to describe modification of a database block */ typedef struct cw_set_element_struct { trans_num tn; /* transaction number for bit maps */ sm_uc_ptr_t old_block; /* Address of 'before-image' of block to be over-written */ cache_rec_ptr_t cr; struct cw_set_element_struct *next_cw_set; struct cw_set_element_struct *prev_cw_set; /* linked list (vertical) of cw_set_elements with one link per block */ struct cw_set_element_struct *high_tlevel; struct cw_set_element_struct *low_tlevel; /* linked list (horizontal) of cw_set elements for a given block with * different transaction levels. Latest cw_set_elements (for a given block) * are inserted at the beginning of the horizontal list */ off_jnl_t jnl_freeaddr; /* journal update address */ uint4 write_type; /* can be GDS_WRITE_PLAIN or GDS_WRITE_KILLTN or GDS_WRITE_BLOCK_SPLIT * or bit-wise-or of both */ key_cum_value *recompute_list_head; /* pointer to a list of keys (with values) that need to be recomputed */ key_cum_value *recompute_list_tail; /* pointer to a list of keys (with values) that need to be recomputed */ enum gds_t_mode mode; /* Create, write, or write root */ block_id blk; /* Block number or a hint block number for creates */ unsigned char *upd_addr; /* Address of the block segment array containing update info * for this block */ unsigned char *new_buff; /* Address of a buffer created for each global mentioned inside of a * transaction more then once (for tp) */ gv_namehead *blk_target; /* address of the "gv_target" associated with a new_buff * used to invalidate clues that point to malloc'ed copies */ int4 cycle; /* When a block splits a new block must be created and the parent must be updated to * to have a record pointing to the new block. The created block number will not be * known until the last possible moment. Thus it is not possible to completely modify * the parent. The following 2 fields are used in such a case. "ins_off" tells where * the created block's number should be put in the parent block. "index" tells which * element of the create/write set is being created. */ block_offset first_off; block_offset ins_off; /* Insert block number offset */ block_offset next_off; block_index index; /* Insert block number index */ int4 reference_cnt; /* Relevant only for a bitmap block. * > 0 => # of non-bitmap blocks to be allocated in this bitmap; * < 0 => # of non-bitmap blocks to be freed up in this bitmap; * == 0 => change to bitmap block without any non-bitmap block change * Used to update csd->free_blocks when the bitmap block is built */ int4 level; /* Block level for newly created blocks */ boolean_t done; /* Has this update been done already? */ boolean_t first_copy; /* If overlaying same buffer, set if first copy needed */ /* just an optimisation - avoids copying first few bytes, if anyway * we are just overlaying the new_buff in the same transaction */ boolean_t forward_process; /* Need to process update array from front when doing kills */ uint4 t_level; /* transaction level associated with cw element, for incremental rollback */ enum db_ver ondsk_blkver; /* Actual block version from block header as it exists on disk. * If "cse->mode" is gds_t_write_root, this is uninitialized. * If "cse->mode" is gds_t_create/gds_t_acquired, this is GDSVCURR. * Otherwise, this is set to cr->ondsk_blkver (cr is got from the history). * Whenever "cse->old_block" is reset, this needs to be reset too (except * in the case of gds_t_create/gds_t_acquired). */ int4 old_mode; /* Saved copy of "cse->mode" before being reset to gds_t_committed. * Is negated at end of bg_update_phase1 to indicate (to secshr_db_clnup) * that phase1 is complete. Is negated back to the postive value at end * of bg_update_phase2. Since this can take on negative values, its type * is int4 (signed) and not enum gds_t_mode (which is unsigned). */ /* The following two fields aid in rolling back the transactions. 'undo_next_off' holds the * original next_off in the blk buffer that would be if another nested transaction was not * started. 'undo_offset' holds the offset at which 'undo_next_off' should be applied in case * of an undo due to trollback. * A 'kill' might change the next_off field at most in two places in the blk buffer. So, is * an array of size two. */ block_offset undo_next_off[2]; block_offset undo_offset[2]; uint4 blk_checksum; uint4 was_free; } cw_set_element; #endif