void wcs_recover(gd_region *reg)
{
bt_rec_ptr_t bt;
cache_rec_ptr_t cr, cr_alt, cr_alt_new, cr_lo, cr_top, hash_hdr;
cache_que_head_ptr_t active_head, hq, wip_head, wq;
gd_region *save_reg;
que_ent_ptr_t back_link; /* should be crit & not need interlocked ops. */
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
node_local_ptr_t cnl;
int4 bml_full, dummy_errno, blk_size;
uint4 jnl_status, epid, r_epid;
int4 bt_buckets, bufindx; /* should be the same type as "csd->bt_buckets" */
inctn_opcode_t save_inctn_opcode;
unsigned int bplmap, lcnt, total_blks, wait_in_rip;
sm_uc_ptr_t buffptr;
blk_hdr_ptr_t blk_ptr;
INTPTR_T bp_lo, bp_top, old_block;
boolean_t backup_block_saved, change_bmm;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
sgm_info *si;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
/* If this is the source server, do not invoke cache recovery as that implies touching the database file header
* (incrementing the curr_tn etc.) and touching the journal file (writing INCTN records) both of which are better
* avoided by the source server; It is best to keep it as read-only to the db/jnl as possible. It is ok to do
* so because the source server anyways does not rely on the integrity of the database cache and so does not need
* to fix it right away. Any other process that does rely on the cache integrity will fix it when it gets crit next.
*/
if (is_src_server)
return;
save_reg = gv_cur_region; /* protect against [at least] M LOCK code which doesn't maintain cs_addrs and cs_data */
TP_CHANGE_REG(reg); /* which are needed by called routines such as wcs_wtstart and wcs_mm_recover */
if (dba_mm == reg->dyn.addr->acc_meth) /* MM uses wcs_recover to remap the database in case of a file extension */
{
wcs_mm_recover(reg);
TP_CHANGE_REG(save_reg);
TREF(wcs_recover_done) = TRUE;
return;
}
csa = &FILE_INFO(reg)->s_addrs;
csd = csa->hdr;
cnl = csa->nl;
si = csa->sgm_info_ptr;
/* If a mupip truncate operation was abruptly interrupted we have to correct any inconsistencies */
GTM_TRUNCATE_ONLY(recover_truncate(csa, csd, gv_cur_region);)
void wcs_recover(gd_region *reg)
{
bt_rec_ptr_t bt;
cache_rec_ptr_t cr, cr_alt, cr_lo, cr_top, hash_hdr;
cache_que_head_ptr_t active_head, hq, wip_head, wq;
gd_region *save_reg;
que_ent_ptr_t back_link; /* should be crit & not need interlocked ops. */
sgmnt_data_ptr_t csd;
sgmnt_addrs *csa;
bool blk_used, change_bmm;
int4 bml_full, dummy_errno, blk_size;
uint4 jnl_status, epid;
int bt_buckets;
inctn_opcode_t save_inctn_opcode;
unsigned int bplmap, lcnt, total_blks;
sm_uc_ptr_t buffptr;
error_def(ERR_BUFRDTIMEOUT);
error_def(ERR_DBCCERR);
error_def(ERR_DBCNTRLERR);
error_def(ERR_DBDANGER);
error_def(ERR_ERRCALL);
error_def(ERR_INVALIDRIP);
error_def(ERR_STOPTIMEOUT);
error_def(ERR_TEXT);
save_reg = gv_cur_region; /* protect against [at least] M LOCK code which doesn't maintain cs_addrs and cs_data */
TP_CHANGE_REG(reg); /* which are needed by called routines such as wcs_wtstart and wcs_mm_recover */
if (dba_mm == reg->dyn.addr->acc_meth) /* MM uses wcs_recover to remap the database in case of a file extension */
{
wcs_mm_recover(reg);
TP_CHANGE_REG(save_reg);
return;
}
csa = &FILE_INFO(reg)->s_addrs;
csd = csa->hdr;
assert(csa->now_crit || csd->clustered);
DEBUG_ONLY(in_wcs_recover = TRUE;) /* used by bt_put() called below */
void gtcmd_rundown(connection_struct *cnx, bool clean_exit)
{
int4 link;
cm_region_list *ptr, *last, *que_next, *que_last;
cm_region_head *region;
uint4 jnl_status;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
int refcnt;
boolean_t was_crit;
int4 rundown_status = EXIT_NRM; /* if gds_rundown went smoothly */
for (ptr = cnx->region_root; ptr;)
{
region = ptr->reghead;
TP_CHANGE_REG(region->reg);
jpc = cs_addrs->jnl;
if (ptr->pini_addr && clean_exit && JNL_ENABLED(cs_data) && (NOJNL != jpc->channel))
{
was_crit = cs_addrs->now_crit;
if (!was_crit)
grab_crit(gv_cur_region);
if (JNL_ENABLED(cs_data))
{
jpc->pini_addr = ptr->pini_addr;
SET_GBL_JREC_TIME; /* jnl_ensure_open/jnl_put_jrt_pfin needs this to be set */
jbp = jpc->jnl_buff;
/* Before writing to jnlfile, adjust jgbl.gbl_jrec_time if needed to maintain time order
* of jnl records. This needs to be done BEFORE the jnl_ensure_open as that could write
* journal records (if it decides to switch to a new journal file).
*/
ADJUST_GBL_JREC_TIME(jgbl, jbp);
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{
if (0 != jpc->pini_addr)
jnl_put_jrt_pfin(cs_addrs);
} else
send_msg(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(cs_data), DB_LEN_STR(gv_cur_region));
}
if (!was_crit)
rel_crit(gv_cur_region);
}
refcnt = --region->refcnt;
/* Dont know how refcnt can become negative but in pro handle it by bypassing this region. The reason is the
* following. refcnt should have originally been a positive value. Every time this function is invoked, it would
* be decremented by one. There should have been one invocation that saw refcnt to be zero. That would have
* done the rundown of the region or if it is still in the stack the rundown is still in progress. Therefore
* it is not a good idea to try running down this region when we see refcnt to be negative (as otherwise we
* will get confused and could potentially end up with SIG-11 or ACCVIO errors). The worst case is that we
* would not have rundown the region in which case an externally issued MUPIP RUNDOWN would be enough.
*/
assert(0 <= refcnt);
if (0 == refcnt)
{ /* free up only as little as needed to facilitate structure reuse when the region is opened again */
assert(region->head.fl == region->head.bl);
VMS_ONLY(gtcm_ast_avail++);
if (JNL_ALLOWED(cs_data))
jpc->pini_addr = 0;
UNIX_ONLY(rundown_status |=) gds_rundown();
gd_ht_kill(region->reg_hash, TRUE); /* TRUE to free up the table and the gv_targets it holds too */
FREE_CSA_DIR_TREE(cs_addrs);
cm_del_gdr_ptr(gv_cur_region);
}
que_next = (cm_region_list *)((unsigned char *)ptr + ptr->regque.fl);
que_last = (cm_region_list *)((unsigned char *)ptr + ptr->regque.bl);
link = (int4)((unsigned char *)que_next - (unsigned char *)que_last);
que_last->regque.fl = link;
que_next->regque.bl = -link;
last = ptr;
ptr = ptr->next;
free(last);
}
void mu_reorg_upgrd_dwngrd(void)
{
blk_hdr new_hdr;
blk_segment *bs1, *bs_ptr;
block_id *blkid_ptr, curblk, curbmp, start_blk, stop_blk, start_bmp, last_bmp;
block_id startblk_input, stopblk_input;
boolean_t upgrade, downgrade, safejnl, nosafejnl, region, first_reorg_in_this_db_fmt, reorg_entiredb;
boolean_t startblk_specified, stopblk_specified, set_fully_upgraded, db_got_to_v5_once, mark_blk_free;
cache_rec_ptr_t cr;
char *bml_lcl_buff = NULL, *command, *reorg_command;
sm_uc_ptr_t bptr = NULL;
cw_set_element *cse;
enum cdb_sc cdb_status;
enum db_ver new_db_format, ondsk_blkver;
gd_region *reg;
int cycle;
int4 blk_seg_cnt, blk_size; /* needed for BLK_INIT,BLK_SEG and BLK_FINI macros */
int4 blocks_left, expected_blks2upgrd, actual_blks2upgrd, total_blks, free_blks;
int4 status, status1, mapsize, lcnt, bml_status;
reorg_stats_t reorg_stats;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
sm_uc_ptr_t blkBase, bml_sm_buff; /* shared memory pointer to the bitmap global buffer */
srch_hist alt_hist;
srch_blk_status *blkhist, bmlhist;
tp_region *rptr;
trans_num curr_tn;
unsigned char save_cw_set_depth;
uint4 lcl_update_trans;
region = (CLI_PRESENT == cli_present("REGION"));
upgrade = (CLI_PRESENT == cli_present("UPGRADE"));
downgrade = (CLI_PRESENT == cli_present("DOWNGRADE"));
assert(upgrade && !downgrade || !upgrade && downgrade);
command = upgrade ? "UPGRADE" : "DOWNGRADE";
reorg_command = upgrade ? "MUPIP REORG UPGRADE" : "MUPIP REORG DOWNGRADE";
reorg_entiredb = TRUE; /* unless STARTBLK or STOPBLK is specified we are going to {up,down}grade the entire database */
startblk_specified = FALSE;
assert(SIZEOF(block_id) == SIZEOF(uint4));
if ((CLI_PRESENT == cli_present("STARTBLK")) && (cli_get_hex("STARTBLK", (uint4 *)&startblk_input)))
{
reorg_entiredb = FALSE;
startblk_specified = TRUE;
}
stopblk_specified = FALSE;
assert(SIZEOF(block_id) == SIZEOF(uint4));
if ((CLI_PRESENT == cli_present("STOPBLK")) && (cli_get_hex("STOPBLK", (uint4 *)&stopblk_input)))
{
reorg_entiredb = FALSE;
stopblk_specified = TRUE;
}
mu_reorg_upgrd_dwngrd_in_prog = TRUE;
mu_reorg_nosafejnl = (CLI_NEGATED == cli_present("SAFEJNL")) ? TRUE : FALSE;
assert(region);
status = SS_NORMAL;
error_mupip = FALSE;
gvinit(); /* initialize gd_header (needed by the later call to mu_getlst) */
mu_getlst("REG_NAME", SIZEOF(tp_region)); /* get the parameter corresponding to REGION qualifier */
if (error_mupip)
{
util_out_print("!/MUPIP REORG !AD cannot proceed with above errors!/", TRUE, LEN_AND_STR(command));
mupip_exit(ERR_MUNOACTION);
}
assert(DBKEYSIZE(MAX_KEY_SZ) == gv_keysize); /* no need to invoke GVKEYSIZE_INIT_IF_NEEDED macro */
gv_target = targ_alloc(gv_keysize, NULL, NULL); /* t_begin needs this initialized */
gv_target_list = NULL;
memset(&alt_hist, 0, SIZEOF(alt_hist)); /* null-initialize history */
blkhist = &alt_hist.h[0];
for (rptr = grlist; NULL != rptr; rptr = rptr->fPtr)
{
if (mu_ctrly_occurred || mu_ctrlc_occurred)
break;
reg = rptr->reg;
util_out_print("!/Region !AD : MUPIP REORG !AD started", TRUE, REG_LEN_STR(reg), LEN_AND_STR(command));
if (reg_cmcheck(reg))
{
util_out_print("Region !AD : MUPIP REORG !AD cannot run across network",
TRUE, REG_LEN_STR(reg), LEN_AND_STR(command));
status = ERR_MUNOFINISH;
continue;
}
mu_reorg_process = TRUE; /* gvcst_init will use this value to use gtm_poollimit settings. */
gvcst_init(reg);
mu_reorg_process = FALSE;
assert(update_array != NULL);
/* access method stored in global directory and database file header might be different in which case
* the database setting prevails. therefore, the access method check can be done only after opening
* the database (i.e. after the gvcst_init)
*/
if (dba_bg != REG_ACC_METH(reg))
{
util_out_print("Region !AD : MUPIP REORG !AD cannot continue as access method is not BG",
TRUE, REG_LEN_STR(reg), LEN_AND_STR(command));
status = ERR_MUNOFINISH;
continue;
}
/* The mu_getlst call above uses insert_region to create the grlist, which ensures that duplicate regions mapping to
* the same db file correspond to only one grlist entry.
*/
assert(FALSE == reg->was_open);
TP_CHANGE_REG(reg); /* sets gv_cur_region, cs_addrs, cs_data */
csa = cs_addrs;
csd = cs_data;
blk_size = csd->blk_size; /* "blk_size" is used by the BLK_FINI macro */
if (reg->read_only)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBRDONLY, 2, DB_LEN_STR(reg));
status = ERR_MUNOFINISH;
continue;
}
assert(GDSVCURR == GDSV6); /* so we trip this assert in case GDSVCURR changes without a change to this module */
new_db_format = (upgrade ? GDSV6 : GDSV4);
grab_crit(reg);
curr_tn = csd->trans_hist.curr_tn;
/* set the desired db format in the file header to the appropriate version, increment transaction number */
status1 = desired_db_format_set(reg, new_db_format, reorg_command);
assert(csa->now_crit); /* desired_db_format_set() should not have released crit */
first_reorg_in_this_db_fmt = TRUE; /* with the current desired_db_format, this is the first reorg */
if (SS_NORMAL != status1)
{ /* "desired_db_format_set" would have printed appropriate error messages */
if (ERR_MUNOACTION != status1)
{ /* real error occurred while setting the db format. skip to next region */
status = ERR_MUNOFINISH;
rel_crit(reg);
continue;
}
util_out_print("Region !AD : Desired DB Format remains at !AD after !AD", TRUE, REG_LEN_STR(reg),
LEN_AND_STR(gtm_dbversion_table[new_db_format]), LEN_AND_STR(reorg_command));
if (csd->reorg_db_fmt_start_tn == csd->desired_db_format_tn)
first_reorg_in_this_db_fmt = FALSE;
} else
util_out_print("Region !AD : Desired DB Format set to !AD by !AD", TRUE, REG_LEN_STR(reg),
LEN_AND_STR(gtm_dbversion_table[new_db_format]), LEN_AND_STR(reorg_command));
assert(dba_bg == csd->acc_meth);
/* Check blks_to_upgrd counter to see if upgrade/downgrade is complete */
total_blks = csd->trans_hist.total_blks;
free_blks = csd->trans_hist.free_blocks;
actual_blks2upgrd = csd->blks_to_upgrd;
/* If MUPIP REORG UPGRADE and there is no block to upgrade in the database as indicated by BOTH
* "csd->blks_to_upgrd" and "csd->fully_upgraded", then we can skip processing.
* If MUPIP REORG UPGRADE and all non-free blocks need to be upgraded then again we can skip processing.
*/
if ((upgrade && (0 == actual_blks2upgrd) && csd->fully_upgraded)
|| (!upgrade && ((total_blks - free_blks) == actual_blks2upgrd)))
{
util_out_print("Region !AD : Blocks to Upgrade counter indicates no action needed for MUPIP REORG !AD",
TRUE, REG_LEN_STR(reg), LEN_AND_STR(command));
util_out_print("Region !AD : Total Blocks = [0x!XL] : Free Blocks = [0x!XL] : "
"Blocks to upgrade = [0x!XL]",
TRUE, REG_LEN_STR(reg), total_blks, free_blks, actual_blks2upgrd);
util_out_print("Region !AD : MUPIP REORG !AD finished!/", TRUE, REG_LEN_STR(reg), LEN_AND_STR(command));
rel_crit(reg);
continue;
}
stop_blk = total_blks;
if (stopblk_specified && stopblk_input <= stop_blk)
stop_blk = stopblk_input;
if (first_reorg_in_this_db_fmt)
{ /* Note down reorg start tn (in case we are interrupted, future reorg will know to resume) */
csd->reorg_db_fmt_start_tn = csd->desired_db_format_tn;
csd->reorg_upgrd_dwngrd_restart_block = 0;
start_blk = (startblk_specified ? startblk_input : 0);
} else
{ /* Either a concurrent MUPIP REORG of the same type ({up,down}grade) is currently running
* or a previously running REORG of the same type was interrupted (Ctrl-Ced).
* In either case resume processing from whatever restart block number is stored in fileheader
* the only exception is if "STARTBLK" was specified in the input in which use that unconditionally.
*/
start_blk = (startblk_specified ? startblk_input : csd->reorg_upgrd_dwngrd_restart_block);
}
if (start_blk > stop_blk)
start_blk = stop_blk;
mu_reorg_upgrd_dwngrd_start_tn = csd->reorg_db_fmt_start_tn;
/* Before releasing crit, flush the file-header and dirty buffers in cache to disk. This is because we are now
* going to read each GDS block directly from disk to determine if it needs to be upgraded/downgraded or not.
*/
if (!wcs_flu(WCSFLU_FLUSH_HDR)) /* wcs_flu assumes gv_cur_region is set (which it is in this routine) */
{
rel_crit(reg);
gtm_putmsg_csa(CSA_ARG(csa)
VARLSTCNT(6) ERR_BUFFLUFAILED, 4, LEN_AND_LIT("MUPIP REORG UPGRADE/DOWNGRADE"), DB_LEN_STR(reg));
status = ERR_MUNOFINISH;
continue;
}
rel_crit(reg);
/* Loop through entire database one GDS block at a time and upgrade/downgrade each of them */
status1 = SS_NORMAL;
start_bmp = ROUND_DOWN2(start_blk, BLKS_PER_LMAP);
last_bmp = ROUND_DOWN2(stop_blk - 1, BLKS_PER_LMAP);
curblk = start_blk; /* curblk is the block to be upgraded/downgraded */
util_out_print("Region !AD : Started processing from block number [0x!XL]", TRUE, REG_LEN_STR(reg), curblk);
if (NULL != bptr)
{ /* malloc/free "bptr" for each region as GDS block-size can be different */
free(bptr);
bptr = NULL;
}
memset(&reorg_stats, 0, SIZEOF(reorg_stats)); /* initialize statistics for this region */
for (curbmp = start_bmp; curbmp <= last_bmp; curbmp += BLKS_PER_LMAP)
{
if (mu_ctrly_occurred || mu_ctrlc_occurred)
{
status1 = ERR_MUNOFINISH;
break;
}
/* --------------------------------------------------------------
* Read in current bitmap block
* --------------------------------------------------------------
*/
assert(!csa->now_crit);
bml_sm_buff = t_qread(curbmp, (sm_int_ptr_t)&cycle, &cr); /* bring block into the cache outside of crit */
reorg_stats.blks_read_from_disk_bmp++;
grab_crit_encr_cycle_sync(reg); /* needed so t_qread does not return NULL below */
if (mu_reorg_upgrd_dwngrd_start_tn != csd->desired_db_format_tn)
{ /* csd->desired_db_format changed since reorg started. discontinue the reorg */
/* see later comment on "csd->reorg_upgrd_dwngrd_restart_block" for why the assignment
* of this field should be done only if a db format change did not occur.
*/
rel_crit(reg);
status1 = ERR_MUNOFINISH;
/* This "start_tn" check is redone after the for-loop and an error message is printed there */
break;
} else if (reorg_entiredb)
{ /* Change "csd->reorg_upgrd_dwngrd_restart_block" only if STARTBLK or STOPBLK was NOT specified */
assert(csd->reorg_upgrd_dwngrd_restart_block <= MAX(start_blk, curbmp));
csd->reorg_upgrd_dwngrd_restart_block = curbmp; /* previous blocks have been upgraded/downgraded */
}
/* Check blks_to_upgrd counter to see if upgrade/downgrade is complete.
* Repeat check done a few steps earlier outside of this for loop.
*/
total_blks = csd->trans_hist.total_blks;
free_blks = csd->trans_hist.free_blocks;
actual_blks2upgrd = csd->blks_to_upgrd;
if ((upgrade && (0 == actual_blks2upgrd) && csd->fully_upgraded)
|| (!upgrade && ((total_blks - free_blks) == actual_blks2upgrd)))
{
rel_crit(reg);
break;
}
bml_sm_buff = t_qread(curbmp, (sm_int_ptr_t)&cycle, &cr); /* now that in crit, note down stable buffer */
if (NULL == bml_sm_buff)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(1) ERR_DSEBLKRDFAIL);
ondsk_blkver = cr->ondsk_blkver; /* note down db fmt on disk for bitmap block */
/* Take a copy of the shared memory bitmap buffer into process-private memory before releasing crit.
* We are interested in those blocks that are currently marked as USED in the bitmap.
* It is possible that once we release crit, concurrent updates change the bitmap state of those blocks.
* In that case, those updates will take care of doing the upgrade/downgrade of those blocks in the
* format currently set in csd->desired_db_format i.e. accomplishing MUPIP REORG UPGRADE/DOWNGRADE's job.
* If the desired_db_format changes concurrently, we will stop doing REORG UPGRADE/DOWNGRADE processing.
*/
if (NULL == bml_lcl_buff)
bml_lcl_buff = malloc(BM_SIZE(BLKS_PER_LMAP));
memcpy(bml_lcl_buff, (blk_hdr_ptr_t)bml_sm_buff, BM_SIZE(BLKS_PER_LMAP));
if (FALSE == cert_blk(reg, curbmp, (blk_hdr_ptr_t)bml_lcl_buff, 0, FALSE))
{ /* certify the block while holding crit as cert_blk uses fields from file-header (shared memory) */
assert(FALSE); /* in pro, skip ugprading/downgarding all blks in this unreliable local bitmap */
rel_crit(reg);
util_out_print("Region !AD : Bitmap Block [0x!XL] has integrity errors. Skipping this bitmap.",
TRUE, REG_LEN_STR(reg), curbmp);
status1 = ERR_MUNOFINISH;
continue;
}
rel_crit(reg);
/* ------------------------------------------------------------------------
* Upgrade/Downgrade all BUSY blocks in the current bitmap
* ------------------------------------------------------------------------
*/
curblk = (curbmp == start_bmp) ? start_blk : curbmp;
mapsize = (curbmp == last_bmp) ? (stop_blk - curbmp) : BLKS_PER_LMAP;
assert(0 != mapsize);
assert(mapsize <= BLKS_PER_LMAP);
db_got_to_v5_once = csd->db_got_to_v5_once;
for (lcnt = curblk - curbmp; lcnt < mapsize; lcnt++, curblk++)
{
if (mu_ctrly_occurred || mu_ctrlc_occurred)
{
status1 = ERR_MUNOFINISH;
goto stop_reorg_on_this_reg; /* goto needed because of nested FOR Loop */
}
GET_BM_STATUS(bml_lcl_buff, lcnt, bml_status);
assert(BLK_MAPINVALID != bml_status); /* cert_blk ran clean so we dont expect invalid entries */
if (BLK_FREE == bml_status)
{
reorg_stats.blks_skipped_free++;
continue;
}
/* MUPIP REORG UPGRADE/DOWNGRADE will convert USED & RECYCLED blocks */
if (db_got_to_v5_once || (BLK_RECYCLED != bml_status))
{ /* Do NOT read recycled V4 block from disk unless it is guaranteed NOT to be too full */
if (lcnt)
{ /* non-bitmap block */
/* read in block from disk into private buffer. dont pollute the cache yet */
if (NULL == bptr)
bptr = (sm_uc_ptr_t)malloc(blk_size);
status1 = dsk_read(curblk, bptr, &ondsk_blkver, FALSE);
/* dsk_read on curblk could return an error (DYNUPGRDFAIL) if curblk needs to be
* upgraded and if its block size was too big to allow the extra block-header space
* requirements for a dynamic upgrade. a MUPIP REORG DOWNGRADE should not error out
* in that case as the block is already in the downgraded format.
*/
if (SS_NORMAL != status1)
{
if (!upgrade && (ERR_DYNUPGRDFAIL == status1))
{
assert(GDSV4 == new_db_format);
ondsk_blkver = new_db_format;
} else
{
gtm_putmsg_csa(CSA_ARG(csa)
VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(reg), status1);
util_out_print("Region !AD : Error occurred while reading block "
"[0x!XL]", TRUE, REG_LEN_STR(reg), curblk);
status1 = ERR_MUNOFINISH;
goto stop_reorg_on_this_reg;/* goto needed due to nested FOR Loop */
}
}
reorg_stats.blks_read_from_disk_nonbmp++;
} /* else bitmap block has been read in crit earlier and ondsk_blkver appropriately set */
if (new_db_format == ondsk_blkver)
{
assert((SS_NORMAL == status1) || (!upgrade && (ERR_DYNUPGRDFAIL == status1)));
status1 = SS_NORMAL; /* treat DYNUPGRDFAIL as no error in case of downgrade */
reorg_stats.blks_skipped_newfmtindisk++;
continue; /* current disk version is identical to what is desired */
}
assert(SS_NORMAL == status1);
}
/* Begin non-TP transaction to upgrade/downgrade the block.
* The way we do that is by updating the block using a null update array.
* Any update to a block will trigger an automatic upgrade/downgrade of the block based on
* the current fileheader desired_db_format setting and we use that here.
*/
t_begin(ERR_MUREORGFAIL, UPDTRNS_DB_UPDATED_MASK);
for (; ;)
{
CHECK_AND_RESET_UPDATE_ARRAY; /* reset update_array_ptr to update_array */
curr_tn = csd->trans_hist.curr_tn;
db_got_to_v5_once = csd->db_got_to_v5_once;
if (db_got_to_v5_once || (BLK_RECYCLED != bml_status))
{
blkhist->cse = NULL; /* start afresh (do not use value from previous retry) */
blkBase = t_qread(curblk, (sm_int_ptr_t)&blkhist->cycle, &blkhist->cr);
if (NULL == blkBase)
{
t_retry((enum cdb_sc)rdfail_detail);
continue;
}
blkhist->blk_num = curblk;
blkhist->buffaddr = blkBase;
ondsk_blkver = blkhist->cr->ondsk_blkver;
new_hdr = *(blk_hdr_ptr_t)blkBase;
mu_reorg_upgrd_dwngrd_blktn = new_hdr.tn;
mark_blk_free = FALSE;
inctn_opcode = upgrade ? inctn_blkupgrd : inctn_blkdwngrd;
} else
{
mark_blk_free = TRUE;
inctn_opcode = inctn_blkmarkfree;
}
inctn_detail.blknum_struct.blknum = curblk;
/* t_end assumes that the history it is passed does not contain a bitmap block.
* for bitmap block, the history validation information is passed through cse instead.
* therefore we need to handle bitmap and non-bitmap cases separately.
*/
if (!lcnt)
{ /* Means a bitmap block.
* At this point we can do a "new_db_format != ondsk_blkver" check to determine
* if the block got converted since we did the dsk_read (see the non-bitmap case
* for a similar check done there), but in that case we will have a transaction
* which has read 1 bitmap block and is updating no block. "t_end" currently cannot
* handle this case as it expects any bitmap block that needs validation to also
* have a corresponding cse which will hold its history. Hence we avoid doing the
* new_db_format check. The only disadvantage of this is that we will end up
* modifying the bitmap block as part of this transaction (in an attempt to convert
* its ondsk_blkver) even though it is already in the right format. Since this
* overhead is going to be one per bitmap block and since the block is in the cache
* at this point, we should not lose much.
*/
assert(!mark_blk_free);
BLK_ADDR(blkid_ptr, SIZEOF(block_id), block_id);
*blkid_ptr = 0;
t_write_map(blkhist, (unsigned char *)blkid_ptr, curr_tn, 0);
assert(&alt_hist.h[0] == blkhist);
alt_hist.h[0].blk_num = 0; /* create empty history for bitmap block */
assert(update_trans);
} else
{ /* non-bitmap block. fill in history for validation in t_end */
assert(curblk); /* we should never come here for block 0 (bitmap) */
if (!mark_blk_free)
{
assert(blkhist->blk_num == curblk);
assert(blkhist->buffaddr == blkBase);
blkhist->tn = curr_tn;
alt_hist.h[1].blk_num = 0;
}
/* Also need to pass the bitmap as history to detect if any concurrent M-kill
* is freeing up the same USED block that we are trying to convert OR if any
* concurrent M-set is reusing the same RECYCLED block that we are trying to
* convert. Because of t_end currently not being able to validate a bitmap
* without that simultaneously having a cse, we need to create a cse for the
* bitmap that is used only for bitmap history validation, but should not be
* used to update the contents of the bitmap block in bg_update.
*/
bmlhist.buffaddr = t_qread(curbmp, (sm_int_ptr_t)&bmlhist.cycle, &bmlhist.cr);
if (NULL == bmlhist.buffaddr)
{
t_retry((enum cdb_sc)rdfail_detail);
continue;
}
bmlhist.blk_num = curbmp;
bmlhist.tn = curr_tn;
GET_BM_STATUS(bmlhist.buffaddr, lcnt, bml_status);
if (BLK_MAPINVALID == bml_status)
{
t_retry(cdb_sc_lostbmlcr);
continue;
}
if (!mark_blk_free)
{
if ((new_db_format != ondsk_blkver) && (BLK_FREE != bml_status))
{ /* block still needs to be converted. create cse */
BLK_INIT(bs_ptr, bs1);
BLK_SEG(bs_ptr, blkBase + SIZEOF(new_hdr),
new_hdr.bsiz - SIZEOF(new_hdr));
BLK_FINI(bs_ptr, bs1);
t_write(blkhist, (unsigned char *)bs1, 0, 0,
((blk_hdr_ptr_t)blkBase)->levl, FALSE,
FALSE, GDS_WRITE_PLAIN);
/* The directory tree status for now is only used to determine
* whether writing the block to snapshot file (see t_end_sysops.c).
* For reorg upgrade/downgrade process, the block is updated in a
* sequential way without changing the gv_target. In this case, we
* assume the block is in directory tree so as to have it written to
* the snapshot file.
*/
BIT_SET_DIR_TREE(cw_set[cw_set_depth-1].blk_prior_state);
/* reset update_trans in case previous retry had set it to 0 */
update_trans = UPDTRNS_DB_UPDATED_MASK;
if (BLK_RECYCLED == bml_status)
{ /* If block that we are upgarding is RECYCLED, indicate to
* bg_update that blks_to_upgrd counter should NOT be
* touched in this case by setting "mode" to a special value
*/
assert(cw_set[cw_set_depth-1].mode == gds_t_write);
cw_set[cw_set_depth-1].mode = gds_t_write_recycled;
/* we SET block as NOT RECYCLED, otherwise, the mm_update()
* or bg_update_phase2 may skip writing it to snapshot file
* when its level is 0
*/
BIT_CLEAR_RECYCLED(cw_set[cw_set_depth-1].blk_prior_state);
}
} else
{ /* Block got converted by another process since we did the dsk_read.
* or this block became marked free in the bitmap.
* No need to update this block. just call t_end for validation of
* both the non-bitmap block as well as the bitmap block.
* Note down that this transaction is no longer updating any blocks.
*/
update_trans = 0;
}
/* Need to put bit maps on the end of the cw set for concurrency checking.
* We want to simulate t_write_map, except we want to update "cw_map_depth"
* instead of "cw_set_depth". Hence the save and restore logic below.
* This part of the code is similar to the one in mu_swap_blk.c
*/
save_cw_set_depth = cw_set_depth;
assert(!cw_map_depth);
t_write_map(&bmlhist, NULL, curr_tn, 0); /* will increment cw_set_depth */
cw_map_depth = cw_set_depth; /* set cw_map_depth to latest cw_set_depth */
cw_set_depth = save_cw_set_depth;/* restore cw_set_depth */
/* t_write_map simulation end */
} else
{
if (BLK_RECYCLED != bml_status)
{ /* Block was RECYCLED at beginning but no longer so. Retry */
t_retry(cdb_sc_bmlmod);
continue;
}
/* Mark recycled block as FREE in bitmap */
assert(lcnt == (curblk - curbmp));
assert(update_array_ptr == update_array);
*((block_id *)update_array_ptr) = lcnt;
update_array_ptr += SIZEOF(block_id);
/* the following assumes SIZEOF(block_id) == SIZEOF(int) */
assert(SIZEOF(block_id) == SIZEOF(int));
*(int *)update_array_ptr = 0;
t_write_map(&bmlhist, (unsigned char *)update_array, curr_tn, 0);
update_trans = UPDTRNS_DB_UPDATED_MASK;
}
}
assert(SIZEOF(lcl_update_trans) == SIZEOF(update_trans));
lcl_update_trans = update_trans; /* take a copy before t_end modifies it */
if ((trans_num)0 != t_end(&alt_hist, NULL, TN_NOT_SPECIFIED))
{ /* In case this is MM and t_end() remapped an extended database, reset csd */
assert(csd == cs_data);
if (!lcl_update_trans)
{
assert(lcnt);
assert(!mark_blk_free);
assert((new_db_format == ondsk_blkver) || (BLK_BUSY != bml_status));
if (BLK_BUSY != bml_status)
reorg_stats.blks_skipped_free++;
else
reorg_stats.blks_skipped_newfmtincache++;
} else if (!lcnt)
reorg_stats.blks_converted_bmp++;
else
reorg_stats.blks_converted_nonbmp++;
break;
}
assert(csd == cs_data);
}
}
}
stop_reorg_on_this_reg:
/* even though ctrl-c occurred, update file-header fields to store reorg's progress before exiting */
grab_crit(reg);
blocks_left = 0;
assert(csd->trans_hist.total_blks >= csd->blks_to_upgrd);
actual_blks2upgrd = csd->blks_to_upgrd;
total_blks = csd->trans_hist.total_blks;
free_blks = csd->trans_hist.free_blocks;
/* Care should be taken not to set "csd->reorg_upgrd_dwngrd_restart_block" in case of a concurrent db fmt
* change. This is because let us say we are doing REORG UPGRADE. A concurrent REORG DOWNGRADE would
* have reset "csd->reorg_upgrd_dwngrd_restart_block" field to 0 and if that reorg is interrupted by a
* Ctrl-C (before this reorg came here) it would have updated "csd->reorg_upgrd_dwngrd_restart_block" to
* a non-zero value indicating how many blocks from 0 have been downgraded. We should not reset this
* field to "curblk" as it will be mis-interpreted as the number of blocks that have been DOWNgraded.
*/
set_fully_upgraded = FALSE;
if (mu_reorg_upgrd_dwngrd_start_tn != csd->desired_db_format_tn)
{ /* csd->desired_db_format changed since reorg started. discontinue the reorg */
util_out_print("Region !AD : Desired DB Format changed during REORG. Stopping REORG.",
TRUE, REG_LEN_STR(reg));
status1 = ERR_MUNOFINISH;
} else if (reorg_entiredb)
{ /* Change "csd->reorg_upgrd_dwngrd_restart_block" only if STARTBLK or STOPBLK was NOT specified */
assert(csd->reorg_upgrd_dwngrd_restart_block <= curblk);
csd->reorg_upgrd_dwngrd_restart_block = curblk; /* blocks lesser than this have been upgraded/downgraded */
expected_blks2upgrd = upgrade ? 0 : (total_blks - free_blks);
blocks_left = upgrade ? actual_blks2upgrd : (expected_blks2upgrd - actual_blks2upgrd);
/* If this reorg command went through all blocks in the database, then it should have
* correctly concluded at this point whether the reorg is complete or not.
* If this reorg command started from where a previous incomplete reorg left
* (i.e. first_reorg_in_this_db_fmt is FALSE), it cannot determine if the initial
* GDS blocks that it skipped are completely {up,down}graded or not.
*/
assert((0 == blocks_left) || (SS_NORMAL != status1) || !first_reorg_in_this_db_fmt);
/* If this is a MUPIP REORG UPGRADE that did go through every block in the database (indicated by
* "reorg_entiredb" && "first_reorg_in_this_db_fmt") and the current count of "blks_to_upgrd" is
* 0 in the file-header and the desired_db_format did not change since the start of the REORG,
* we can be sure that the entire database has been upgraded. Set "csd->fully_upgraded" to TRUE.
*/
if ((SS_NORMAL == status1) && first_reorg_in_this_db_fmt && upgrade && (0 == actual_blks2upgrd))
{
csd->fully_upgraded = TRUE;
csd->db_got_to_v5_once = TRUE;
set_fully_upgraded = TRUE;
}
/* flush all changes noted down in the file-header */
if (!wcs_flu(WCSFLU_FLUSH_HDR)) /* wcs_flu assumes gv_cur_region is set (which it is in this routine) */
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_BUFFLUFAILED, 4,
LEN_AND_LIT("MUPIP REORG UPGRADE/DOWNGRADE"), DB_LEN_STR(reg));
status = ERR_MUNOFINISH;
rel_crit(reg);
continue;
}
}
curr_tn = csd->trans_hist.curr_tn;
rel_crit(reg);
util_out_print("Region !AD : Stopped processing at block number [0x!XL]", TRUE, REG_LEN_STR(reg), curblk);
/* Print statistics */
util_out_print("Region !AD : Statistics : Blocks Read From Disk (Bitmap) : 0x!XL",
TRUE, REG_LEN_STR(reg), reorg_stats.blks_read_from_disk_bmp);
util_out_print("Region !AD : Statistics : Blocks Skipped (Free) : 0x!XL",
TRUE, REG_LEN_STR(reg), reorg_stats.blks_skipped_free);
util_out_print("Region !AD : Statistics : Blocks Read From Disk (Non-Bitmap) : 0x!XL",
TRUE, REG_LEN_STR(reg), reorg_stats.blks_read_from_disk_nonbmp);
util_out_print("Region !AD : Statistics : Blocks Skipped (new fmt in disk) : 0x!XL",
TRUE, REG_LEN_STR(reg), reorg_stats.blks_skipped_newfmtindisk);
util_out_print("Region !AD : Statistics : Blocks Skipped (new fmt in cache) : 0x!XL",
TRUE, REG_LEN_STR(reg), reorg_stats.blks_skipped_newfmtincache);
util_out_print("Region !AD : Statistics : Blocks Converted (Bitmap) : 0x!XL",
TRUE, REG_LEN_STR(reg), reorg_stats.blks_converted_bmp);
util_out_print("Region !AD : Statistics : Blocks Converted (Non-Bitmap) : 0x!XL",
TRUE, REG_LEN_STR(reg), reorg_stats.blks_converted_nonbmp);
if (reorg_entiredb && (SS_NORMAL == status1) && (0 != blocks_left))
{ /* file-header counter does not match what reorg on the entire database expected to see */
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBBTUWRNG, 2, expected_blks2upgrd, actual_blks2upgrd);
util_out_print("Region !AD : Run MUPIP INTEG (without FAST qualifier) to fix the counter",
TRUE, REG_LEN_STR(reg));
status1 = ERR_MUNOFINISH;
} else
util_out_print("Region !AD : Total Blocks = [0x!XL] : Free Blocks = [0x!XL] : "
"Blocks to upgrade = [0x!XL]",
TRUE, REG_LEN_STR(reg), total_blks, free_blks, actual_blks2upgrd);
/* Issue success or failure message for this region */
if (SS_NORMAL == status1)
{ /* issue success only if REORG did not encounter any error in its processing */
if (set_fully_upgraded)
util_out_print("Region !AD : Database is now FULLY UPGRADED", TRUE, REG_LEN_STR(reg));
util_out_print("Region !AD : MUPIP REORG !AD finished!/", TRUE, REG_LEN_STR(reg), LEN_AND_STR(command));
send_msg_csa(CSA_ARG(csa) VARLSTCNT(7) ERR_MUREUPDWNGRDEND, 5, REG_LEN_STR(reg),
process_id, process_id, &curr_tn);
} else
{
assert(ERR_MUNOFINISH == status1);
assert((SS_NORMAL == status) || (ERR_MUNOFINISH == status));
util_out_print("Region !AD : MUPIP REORG !AD incomplete. See above messages.!/",
TRUE, REG_LEN_STR(reg), LEN_AND_STR(command));
status = status1;
}
}
if (NULL != bptr)
free(bptr);
if (NULL != bml_lcl_buff)
free(bml_lcl_buff);
if (mu_ctrly_occurred || mu_ctrlc_occurred)
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_REORGCTRLY);
status = ERR_MUNOFINISH;
}
mupip_exit(status);
}
uint4 mur_process_intrpt_recov()
{
jnl_ctl_list *jctl, *last_jctl;
reg_ctl_list *rctl, *rctl_top;
int rename_fn_len, save_name_len, idx;
char prev_jnl_fn[MAX_FN_LEN + 1], rename_fn[MAX_FN_LEN + 1], save_name[MAX_FN_LEN + 1];
jnl_create_info jnl_info;
uint4 status, status2;
uint4 max_autoswitchlimit, max_jnl_alq, max_jnl_deq, freeblks;
sgmnt_data_ptr_t csd;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
boolean_t jfh_changed;
jnl_record *jnlrec;
jnl_file_header *jfh;
jnl_tm_t now;
for (rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++)
{
TP_CHANGE_REG(rctl->gd);
csd = cs_data; /* MM logic after wcs_flu call requires this to be set */
assert(csd == rctl->csa->hdr);
jctl = rctl->jctl_turn_around;
max_jnl_alq = max_jnl_deq = max_autoswitchlimit = 0;
for (last_jctl = NULL ; (NULL != jctl); last_jctl = jctl, jctl = jctl->next_gen)
{
jfh = jctl->jfh;
if (max_autoswitchlimit < jfh->autoswitchlimit)
{ /* Note that max_jnl_alq, max_jnl_deq are not the maximum journal allocation/extensions across
* generations, but rather the allocation/extension corresponding to the maximum autoswitchlimit.
*/
max_autoswitchlimit = jfh->autoswitchlimit;
max_jnl_alq = jfh->jnl_alq;
max_jnl_deq = jfh->jnl_deq;
}
/* Until now, "rctl->blks_to_upgrd_adjust" holds the number of V4 format newly created bitmap blocks
* seen in INCTN records in backward processing. It is possible that backward processing might have
* missed out on seeing those INCTN records which are part of virtually-truncated or completely-rolled-bak
* journal files. The journal file-header has a separate field "prev_recov_blks_to_upgrd_adjust" which
* maintains exactly this count. Therefore adjust the rctl counter accordingly.
*/
assert(!jfh->prev_recov_blks_to_upgrd_adjust || !jfh->recover_interrupted);
assert(!jfh->prev_recov_blks_to_upgrd_adjust || jfh->prev_recov_end_of_data);
rctl->blks_to_upgrd_adjust += jfh->prev_recov_blks_to_upgrd_adjust;
}
if (max_autoswitchlimit > last_jctl->jfh->autoswitchlimit)
{
csd->jnl_alq = max_jnl_alq;
csd->jnl_deq = max_jnl_deq;
csd->autoswitchlimit = max_autoswitchlimit;
} else
{
assert(csd->jnl_alq == last_jctl->jfh->jnl_alq);
assert(csd->jnl_deq == last_jctl->jfh->jnl_deq);
assert(csd->autoswitchlimit == last_jctl->jfh->autoswitchlimit);
}
jctl = rctl->jctl_turn_around;
/* Get a pointer to the turn around point EPOCH record */
jnlrec = rctl->mur_desc->jnlrec;
assert(JRT_EPOCH == jnlrec->prefix.jrec_type);
assert(jctl->turn_around_time == jnlrec->prefix.time);
assert(jctl->turn_around_seqno == jnlrec->jrec_epoch.jnl_seqno);
assert(jctl->turn_around_tn == jnlrec->prefix.tn);
assert(jctl->rec_offset == jctl->turn_around_offset);
/* Reset file-header "blks_to_upgrd" counter to the turn around point epoch value. Adjust this to include
* the number of new V4 format bitmaps created by post-turnaround-point db file extensions.
* The adjustment value is maintained in rctl->blks_to_upgrd_adjust.
*/
csd->blks_to_upgrd = jnlrec->jrec_epoch.blks_to_upgrd;
csd->blks_to_upgrd += rctl->blks_to_upgrd_adjust;
# ifdef GTM_TRIGGER
/* online rollback can potentially take the database to a point in the past where the triggers that were
* previously installed are no longer a part of the current database state and so any process that restarts
* AFTER online rollback completes SHOULD reload triggers and the only way to do that is by incrementing the
* db_trigger_cycle in the file header.
*/
if (jgbl.onlnrlbk && (0 < csd->db_trigger_cycle))
{ /* check for non-zero db_trigger_cycle is to prevent other processes (continuing after online rollback)
* to establish implicit TP (on seeing the trigger cycle mismatch) when there are actually no triggers
* installed in the database (because there were none at the start of online rollback).
*/
csd->db_trigger_cycle++;
if (0 == csd->db_trigger_cycle)
csd->db_trigger_cycle = 1; /* Don't allow cycle set to 0 which means uninitialized */
}
# endif
assert((WBTEST_ALLOW_ARBITRARY_FULLY_UPGRADED == gtm_white_box_test_case_number) ||
(FALSE == jctl->turn_around_fullyupgraded) || (TRUE == jctl->turn_around_fullyupgraded));
/* Set csd->fully_upgraded to FALSE if:
* a) The turn around EPOCH had the fully_upgraded field set to FALSE
* OR
* b) If csd->blks_to_upgrd counter is non-zero. This field can be non-zero even if the turnaround EPOCH's
* fully_upgraded field is TRUE. This is possible if the database was downgraded to V4 (post turnaround EPOCH)
* format and database extensions happened causing new V4 format bitmap blocks to be written. The count of V4
* format bitmap blocks is maintained ONLY as part of INCTN records (with INCTN opcode SET_JNL_FILE_CLOSE_EXTEND)
* noted down in rctl->blks_to_upgrd_adjust counter as part of BACKWARD processing which are finally added to
* csd->blks_to_upgrd.
*/
if (!jctl->turn_around_fullyupgraded || csd->blks_to_upgrd)
csd->fully_upgraded = FALSE;
csd->trans_hist.early_tn = jctl->turn_around_tn;
csd->trans_hist.curr_tn = csd->trans_hist.early_tn; /* INCREMENT_CURR_TN macro not used but noted in comment
* to identify all places that set curr_tn */
csd->jnl_eovtn = csd->trans_hist.curr_tn;
csd->turn_around_point = TRUE;
/* MUPIP REORG UPGRADE/DOWNGRADE stores its partially processed state in the database file header.
* It is difficult for recovery to restore those fields to a correct partial value.
* Hence reset the related fields as if the desired_db_format got set just ONE tn BEFORE the EPOCH record
* and that there was no more processing that happened.
* This might potentially mean some duplicate processing for MUPIP REORG UPGRADE/DOWNGRADE after the recovery.
* But that will only be the case as long as the database is in compatibility (mixed) mode (hopefully not long).
*/
if (csd->desired_db_format_tn >= jctl->turn_around_tn)
csd->desired_db_format_tn = jctl->turn_around_tn - 1;
if (csd->reorg_db_fmt_start_tn >= jctl->turn_around_tn)
csd->reorg_db_fmt_start_tn = jctl->turn_around_tn - 1;
if (csd->tn_upgrd_blks_0 > jctl->turn_around_tn)
csd->tn_upgrd_blks_0 = (trans_num)-1;
csd->reorg_upgrd_dwngrd_restart_block = 0;
/* Compute current value of "free_blocks" based on the value of "free_blocks" at the turnaround point epoch
* record and the change in "total_blks" since that epoch to the present form of the database. Any difference
* in "total_blks" implies database file extensions happened since the turnaround point. A backward rollback
* undoes everything (including all updates) except file extensions (it does not truncate the file size).
* Therefore every block that was newly allocated as part of those file extensions should be considered FREE
* for the current calculations except for the local bitmap blocks which are BUSY the moment they are created.
*/
assert(jnlrec->jrec_epoch.total_blks <= csd->trans_hist.total_blks);
csd->trans_hist.free_blocks = jnlrec->jrec_epoch.free_blocks
+ (csd->trans_hist.total_blks - jnlrec->jrec_epoch.total_blks)
- DIVIDE_ROUND_UP(csd->trans_hist.total_blks, BLKS_PER_LMAP)
+ DIVIDE_ROUND_UP(jnlrec->jrec_epoch.total_blks, BLKS_PER_LMAP);
assert(!csd->blks_to_upgrd || !csd->fully_upgraded);
assert((freeblks = mur_blocks_free(rctl)) == csd->trans_hist.free_blocks);
/* Update strm_reg_seqno[] in db file header to reflect the turn around point.
* Before updating "strm_reg_seqno", make sure value is saved into "save_strm_reg_seqno".
* This is relied upon by the function "mur_get_max_strm_reg_seqno" in case of interrupted rollback.
*/
for (idx = 0; idx < MAX_SUPPL_STRMS; idx++)
{
if (!csd->save_strm_reg_seqno[idx])
csd->save_strm_reg_seqno[idx] = csd->strm_reg_seqno[idx];
csd->strm_reg_seqno[idx] = jnlrec->jrec_epoch.strm_seqno[idx];
}
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_FSYNC_DB);
assert(cs_addrs->ti->curr_tn == jctl->turn_around_tn);
# ifdef UNIX
if (jgbl.onlnrlbk)
{
if (dba_bg == cs_addrs->hdr->acc_meth)
{ /* dryclean the cache (basically reset the cycle fields in all teh cache records) so as to make
* GT.M processes that only does 'reads' to require crit and hence realize that online rollback
* is in progress
*/
bt_refresh(cs_addrs, FALSE); /* sets earliest bt TN to be the turn around TN */
}
db_csh_ref(cs_addrs, FALSE);
assert(NULL != cs_addrs->jnl);
jpc = cs_addrs->jnl;
assert(NULL != jpc->jnl_buff);
jbp = jpc->jnl_buff;
/* Since Rollback simulates the journal record along with the timestamp at which the update was made, it
* sets jgbl.dont_reset_gbl_jrec_time to TRUE so that during forward processing t_end or tp_tend does not
* reset the gbl_jrec_time to reflect the current time. But, with Online Rollback, one can have the shared
* memory up and running and hence can have jbp->prev_jrec_time to be the time of the most recent journal
* update made. Later in t_end/tp_tend, ADJUST_GBL_JREC_TIME is invoked which ensures that if ever
* gbl_jrec_time (the time of the current update) is less than jbp->prev_jrec_time (time of the latest
* journal update), dont_reset_gbl_jrec_time better be FALSE. But, this assert will trip since Rollback
* sets the latter to TRUE. To fix this, set jbp->prev_jrec_time to the turn around time stamp. This way
* we are guaranteed that all the updates done in the forward processing will have a timestamp that is
* greater than the turn around timestamp
*/
SET_JNLBUFF_PREV_JREC_TIME(jbp, jctl->turn_around_time, DO_GBL_JREC_TIME_CHECK_FALSE);
} else if (dba_bg == csd->acc_meth)
{ /* set earliest bt TN to be the turn-around TN (taken from bt_refresh()) */
SET_OLDEST_HIST_TN(cs_addrs, cs_addrs->ti->curr_tn - 1);
}
# else
if (dba_bg == csd->acc_meth)
{ /* set earliest bt TN to be the turn-around TN (taken from bt_refresh()) */
SET_OLDEST_HIST_TN(cs_addrs, cs_addrs->ti->curr_tn - 1);
}
# endif
csd->turn_around_point = FALSE;
assert(OLDEST_HIST_TN(cs_addrs) == (cs_addrs->ti->curr_tn - 1));
/* In case this is MM and wcs_flu() remapped an extended database, reset rctl->csd */
assert((dba_mm == cs_data->acc_meth) || (rctl->csd == cs_data));
rctl->csd = cs_data;
}
JNL_SHORT_TIME(now);
for (rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++)
{
TP_CHANGE_REG_IF_NEEDED(rctl->gd);
if (!rctl->jfh_recov_interrupted)
jctl = rctl->jctl_turn_around;
else
{
DEBUG_ONLY(
for (jctl = rctl->jctl_turn_around; NULL != jctl->next_gen; jctl = jctl->next_gen)
;
/* check that latest gener file name does not match db header */
assert((rctl->csd->jnl_file_len != jctl->jnl_fn_len)
|| (0 != memcmp(rctl->csd->jnl_file_name, jctl->jnl_fn, jctl->jnl_fn_len)));
)
jctl = rctl->jctl_alt_head;
}
assert(NULL != jctl);
for ( ; NULL != jctl->next_gen; jctl = jctl->next_gen)
;
assert(rctl->csd->jnl_file_len == jctl->jnl_fn_len); /* latest gener file name */
assert(0 == memcmp(rctl->csd->jnl_file_name, jctl->jnl_fn, jctl->jnl_fn_len)); /* should match db header */
if (SS_NORMAL != (status = prepare_unique_name((char *)jctl->jnl_fn, jctl->jnl_fn_len, "", "",
rename_fn, &rename_fn_len, now, &status2)))
return status;
jctl->jnl_fn_len = rename_fn_len; /* change the name in memory to the proposed name */
memcpy(jctl->jnl_fn, rename_fn, rename_fn_len + 1);
/* Rename hasn't happened yet at the filesystem level. In case current recover command is interrupted,
* we need to update jfh->next_jnl_file_name before mur_forward(). Update jfh->next_jnl_file_name for
* all journal files from which PBLK records were applied. Create new journal files for forward play.
*/
assert(NULL != rctl->jctl_turn_around);
jctl = rctl->jctl_turn_around; /* points to journal file which has current recover's turn around point */
assert(0 != jctl->turn_around_offset);
jfh = jctl->jfh;
jfh->turn_around_offset = jctl->turn_around_offset; /* save progress in file header for */
jfh->turn_around_time = jctl->turn_around_time; /* possible re-issue of recover */
for (idx = 0; idx < MAX_SUPPL_STRMS; idx++)
jfh->strm_end_seqno[idx] = csd->strm_reg_seqno[idx];
jfh_changed = TRUE;
/* We are about to update the journal file header of the turnaround-point journal file to store the
* non-zero jfh->turn_around_offset. Ensure corresponding database is considered updated.
* This is needed in case journal recovery/rollback terminates abnormally and we go to mur_close_files.
* We need to ensure csd->recov_interrupted does not get reset to FALSE even if this region did not have
* have any updates to the corresponding database file otherwise. (GTM-8394)
*/
rctl->db_updated = TRUE;
for ( ; NULL != jctl; jctl = jctl->next_gen)
{ /* setup the next_jnl links. note that in the case of interrupted recovery, next_jnl links
* would have been already set starting from the turn-around point journal file of the
* interrupted recovery but the new recovery MIGHT have taken us to a still previous
* generation journal file that needs its next_jnl link set. this is why we do the next_jnl
* link setup even in the case of interrupted recovery although in most cases it is unnecessary.
*/
jfh = jctl->jfh;
if (NULL != jctl->next_gen)
{
jfh->next_jnl_file_name_length = jctl->next_gen->jnl_fn_len;
memcpy(jfh->next_jnl_file_name, jctl->next_gen->jnl_fn, jctl->next_gen->jnl_fn_len);
jfh_changed = TRUE;
} else
assert(0 == jfh->next_jnl_file_name_length); /* null link from latest generation */
if (jfh->turn_around_offset && (jctl != rctl->jctl_turn_around))
{ /* It is possible that the current recovery has a turn-around-point much before the
* previously interrupted recovery. If it happens to be a previous generation journal
* file then we have to reset the original turn-around-point to be zero in the journal
* file header in order to ensure if this recovery gets interrupted we do interrupted
* recovery processing until the new turn-around-point instead of stopping incorrectly
* at the original turn-around-point itself. Note that there could be more than one
* journal file with a non-zero turn_around_offset (depending on how many previous
* recoveries got interrupted in this loop) that need to be reset.
*/
assert(!jctl->turn_around_offset);
assert(rctl->recov_interrupted || rctl->jctl_apply_pblk); /* rctl->jfh_recov_interrupted can fail */
jfh->turn_around_offset = 0;
jfh->turn_around_time = 0;
jfh_changed = TRUE;
}
if (jfh_changed)
{
/* Since overwriting the journal file header (an already allocated block
* in the file) should not cause ENOSPC, we dont take the trouble of
* passing csa or jnl_fn (first two parameters). Instead we pass NULL.
*/
JNL_DO_FILE_WRITE(NULL, NULL, jctl->channel, 0, jfh,
REAL_JNL_HDR_LEN, jctl->status, jctl->status2);
if (SS_NORMAL != jctl->status)
{
assert(FALSE);
if (SS_NORMAL == jctl->status2)
gtm_putmsg_csa(CSA_ARG(rctl->csa) VARLSTCNT(5) ERR_JNLWRERR, 2, jctl->jnl_fn_len,
jctl->jnl_fn, jctl->status);
else
gtm_putmsg_csa(CSA_ARG(rctl->csa) VARLSTCNT1(6) ERR_JNLWRERR, 2, jctl->jnl_fn_len,
jctl->jnl_fn, jctl->status, PUT_SYS_ERRNO(jctl->status2));
return jctl->status;
}
GTM_JNL_FSYNC(rctl->csa, jctl->channel, jctl->status);
if (-1 == jctl->status)
{
jctl->status2 = errno;
assert(FALSE);
gtm_putmsg_csa(CSA_ARG(rctl->csa) VARLSTCNT(9) ERR_JNLFSYNCERR, 2,
jctl->jnl_fn_len, jctl->jnl_fn,
ERR_TEXT, 2, RTS_ERROR_TEXT("Error with fsync"), jctl->status2);
return ERR_JNLFSYNCERR;
}
}
jfh_changed = FALSE;
}
memset(&jnl_info, 0, SIZEOF(jnl_info));
jnl_info.status = jnl_info.status2 = SS_NORMAL;
jnl_info.prev_jnl = &prev_jnl_fn[0];
set_jnl_info(rctl->gd, &jnl_info);
jnl_info.prev_jnl_len = rctl->jctl_turn_around->jnl_fn_len;
memcpy(jnl_info.prev_jnl, rctl->jctl_turn_around->jnl_fn, rctl->jctl_turn_around->jnl_fn_len);
jnl_info.prev_jnl[jnl_info.prev_jnl_len] = 0;
jnl_info.jnl_len = rctl->csd->jnl_file_len;
memcpy(jnl_info.jnl, rctl->csd->jnl_file_name, jnl_info.jnl_len);
jnl_info.jnl[jnl_info.jnl_len] = 0;
assert(!mur_options.rollback || jgbl.mur_rollback);
jnl_info.reg_seqno = rctl->jctl_turn_around->turn_around_seqno;
jgbl.gbl_jrec_time = rctl->jctl_turn_around->turn_around_time; /* time needed for cre_jnl_file_common() */
if (EXIT_NRM != cre_jnl_file_common(&jnl_info, rename_fn, rename_fn_len))
{
gtm_putmsg_csa(CSA_ARG(rctl->csa) VARLSTCNT(4) ERR_JNLNOCREATE, 2, jnl_info.jnl_len, jnl_info.jnl);
return jnl_info.status;
}
# ifdef UNIX
if (jgbl.onlnrlbk)
{
cs_addrs = rctl->csa;
/* Mimic what jnl_file_close in case of cleanly a closed journal file */
jpc = cs_addrs->jnl; /* the previous loop makes sure cs_addrs->jnl->jnl_buff is valid*/
NULLIFY_JNL_FILE_ID(cs_addrs);
jpc->jnl_buff->cycle++; /* so that, all other processes knows to switch to newer journal file */
jpc->cycle--; /* decrement cycle so jnl_ensure_open() knows to reopen the journal */
}
# endif
if (NULL != rctl->jctl_alt_head) /* remove the journal files created by last interrupted recover process */
{
mur_rem_jctls(rctl);
rctl->jctl_alt_head = NULL;
}
/* From this point on, journal records are written into the newly created journal file. However, we still read
* from old journal files.
*/
}
