uint4 set_jnl_file_close(set_jnl_file_close_opcode_t set_jnl_file_close_opcode)
{
uint4 jnl_status = 0;
cs_addrs = &FILE_INFO(gv_cur_region)->s_addrs;
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{
if (0 == cs_addrs->jnl->pini_addr)
jnl_put_jrt_pini(cs_addrs);
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH);
jnl_put_jrt_pfin(cs_addrs);
jnl_file_close(gv_cur_region, TRUE, TRUE);
} else
gtm_putmsg(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(cs_addrs->hdr),
DB_LEN_STR(gv_cur_region));
return jnl_status;
}
boolean_t mu_truncate(int4 truncate_percent)
{
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
int num_local_maps;
int lmap_num, lmap_blk_num;
int bml_status, sigkill;
int save_errno;
int ftrunc_status;
uint4 jnl_status;
uint4 old_total, new_total;
uint4 old_free, new_free;
uint4 end_blocks;
int4 blks_in_lmap, blk;
gtm_uint64_t before_trunc_file_size;
off_t trunc_file_size;
off_t padding;
uchar_ptr_t lmap_addr;
boolean_t was_crit;
uint4 found_busy_blk;
srch_blk_status bmphist;
srch_blk_status *blkhist;
srch_hist alt_hist;
trans_num curr_tn;
blk_hdr_ptr_t lmap_blk_hdr;
block_id *blkid_ptr;
unix_db_info *udi;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
char *err_msg;
intrpt_state_t prev_intrpt_state;
off_t offset;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
csa = cs_addrs;
csd = cs_data;
if (dba_mm == csd->acc_meth)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_MUTRUNCNOTBG, 2, REG_LEN_STR(gv_cur_region));
return TRUE;
}
if ((GDSVCURR != csd->desired_db_format) || (csd->blks_to_upgrd != 0))
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_MUTRUNCNOV4, 2, REG_LEN_STR(gv_cur_region));
return TRUE;
}
if (csa->ti->free_blocks < (truncate_percent * csa->ti->total_blks / 100))
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(5) ERR_MUTRUNCNOSPACE, 3, REG_LEN_STR(gv_cur_region), truncate_percent);
return TRUE;
}
/* already checked for parallel truncates on this region --- see mupip_reorg.c */
gv_target = NULL;
assert(csa->nl->trunc_pid == process_id);
assert(dba_mm != csd->acc_meth);
old_total = csa->ti->total_blks;
old_free = csa->ti->free_blocks;
sigkill = 0;
found_busy_blk = 0;
memset(&alt_hist, 0, SIZEOF(alt_hist)); /* null-initialize history */
assert(csd->bplmap == BLKS_PER_LMAP);
end_blocks = old_total % BLKS_PER_LMAP; /* blocks in the last lmap (first one we start scanning) */
if (0 == end_blocks)
end_blocks = BLKS_PER_LMAP;
num_local_maps = DIVIDE_ROUND_UP(old_total, BLKS_PER_LMAP);
/* ======================================== PHASE 1 ======================================== */
for (lmap_num = num_local_maps - 1; (lmap_num > 0 && !found_busy_blk); lmap_num--)
{
if (mu_ctrly_occurred || mu_ctrlc_occurred)
return TRUE;
assert(csa->ti->total_blks >= old_total); /* otherwise, a concurrent truncate happened... */
if (csa->ti->total_blks != old_total) /* Extend (likely called by mupip extend) -- don't truncate */
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(5) ERR_MUTRUNCNOSPACE, 3, REG_LEN_STR(gv_cur_region),
truncate_percent);
return TRUE;
}
lmap_blk_num = lmap_num * BLKS_PER_LMAP;
if (csa->nl->highest_lbm_with_busy_blk >= lmap_blk_num)
{
found_busy_blk = lmap_blk_num;
break;
}
blks_in_lmap = (lmap_num == num_local_maps - 1) ? end_blocks : BLKS_PER_LMAP;
/* Loop through non-bitmap blocks of this lmap, do recycled2free */
DBGEHND((stdout, "DBG:: lmap_num = [%lu], lmap_blk_num = [%lu], blks_in_lmap = [%lu]\n",
lmap_num, lmap_blk_num, blks_in_lmap));
for (blk = 1; blk < blks_in_lmap && blk != -1 && !found_busy_blk;)
{
t_begin(ERR_MUTRUNCFAIL, UPDTRNS_DB_UPDATED_MASK);
for (;;) /* retry loop for recycled to free transactions */
{
curr_tn = csd->trans_hist.curr_tn;
/* Read the nth local bitmap into memory */
bmphist.blk_num = lmap_blk_num;
bmphist.buffaddr = t_qread(bmphist.blk_num, &bmphist.cycle, &bmphist.cr);
lmap_blk_hdr = (blk_hdr_ptr_t)bmphist.buffaddr;
if (!(bmphist.buffaddr) || (BM_SIZE(BLKS_PER_LMAP) != lmap_blk_hdr->bsiz))
{ /* Could not read the block successfully. Retry. */
t_retry((enum cdb_sc)rdfail_detail);
continue;
}
lmap_addr = bmphist.buffaddr + SIZEOF(blk_hdr);
/* starting from the hint (blk itself), find the first busy or recycled block */
blk = bml_find_busy_recycled(blk, lmap_addr, blks_in_lmap, &bml_status);
assert(blk < BLKS_PER_LMAP);
if (blk == -1 || blk >= blks_in_lmap)
{ /* done with this lmap, continue to next */
t_abort(gv_cur_region, csa);
break;
}
else if (BLK_BUSY == bml_status || csa->nl->highest_lbm_with_busy_blk >= lmap_blk_num)
{ /* stop processing blocks... skip ahead to phase 2 */
found_busy_blk = lmap_blk_num;
t_abort(gv_cur_region, csa);
break;
}
else if (BLK_RECYCLED == bml_status)
{ /* Write PBLK records for recycled blocks only if before_image journaling is
* enabled. t_end() takes care of checking if journaling is enabled and
* writing PBLK record. We have to at least mark the recycled block as free.
*/
RESET_UPDATE_ARRAY;
update_trans = UPDTRNS_DB_UPDATED_MASK;
*((block_id *)update_array_ptr) = blk;
update_array_ptr += SIZEOF(block_id);
*(int *)update_array_ptr = 0;
alt_hist.h[1].blk_num = 0;
alt_hist.h[0].level = 0;
alt_hist.h[0].cse = NULL;
alt_hist.h[0].tn = curr_tn;
alt_hist.h[0].blk_num = lmap_blk_num + blk;
alt_hist.h[0].buffaddr = t_qread(alt_hist.h[0].blk_num,
&alt_hist.h[0].cycle, &alt_hist.h[0].cr);
if (!alt_hist.h[0].buffaddr)
{
t_retry((enum cdb_sc)rdfail_detail);
continue;
}
if (!t_recycled2free(&alt_hist.h[0]))
{
t_retry(cdb_sc_lostbmlcr);
continue;
}
t_write_map(&bmphist, (unsigned char *)update_array, curr_tn, 0);
/* Set the opcode for INCTN record written by t_end() */
inctn_opcode = inctn_blkmarkfree;
if ((trans_num)0 == t_end(&alt_hist, NULL, TN_NOT_SPECIFIED))
continue;
/* block processed, scan from the next one */
blk++;
break;
} else
{
assert(t_tries < CDB_STAGNATE);
t_retry(cdb_sc_badbitmap);
continue;
}
} /* END recycled2free retry loop */
} /* END scanning blocks of this particular lmap */
/* Write PBLK for the bitmap block, in case it hasn't been written i.e. t_end() was never called above */
/* Do a transaction that just increments the bitmap block's tn so that t_end() can do its thing */
DBGEHND((stdout, "DBG:: bitmap block inctn -- lmap_blk_num = [%lu]\n", lmap_blk_num));
t_begin(ERR_MUTRUNCFAIL, UPDTRNS_DB_UPDATED_MASK);
for (;;)
{
RESET_UPDATE_ARRAY;
BLK_ADDR(blkid_ptr, SIZEOF(block_id), block_id);
*blkid_ptr = 0;
update_trans = UPDTRNS_DB_UPDATED_MASK;
inctn_opcode = inctn_mu_reorg; /* inctn_mu_truncate */
curr_tn = csd->trans_hist.curr_tn;
blkhist = &alt_hist.h[0];
blkhist->blk_num = lmap_blk_num;
blkhist->tn = curr_tn;
blkhist->cse = NULL; /* start afresh (do not use value from previous retry) */
/* Read the nth local bitmap into memory */
blkhist->buffaddr = t_qread(lmap_blk_num, (sm_int_ptr_t)&blkhist->cycle, &blkhist->cr);
lmap_blk_hdr = (blk_hdr_ptr_t)blkhist->buffaddr;
if (!(blkhist->buffaddr) || (BM_SIZE(BLKS_PER_LMAP) != lmap_blk_hdr->bsiz))
{ /* Could not read the block successfully. Retry. */
t_retry((enum cdb_sc)rdfail_detail);
continue;
}
t_write_map(blkhist, (unsigned char *)blkid_ptr, curr_tn, 0);
blkhist->blk_num = 0; /* create empty history for bitmap block */
if ((trans_num)0 == t_end(&alt_hist, NULL, TN_NOT_SPECIFIED))
continue;
break;
}
} /* END scanning lmaps */
/* ======================================== PHASE 2 ======================================== */
assert(!csa->now_crit);
for (;;)
{ /* wait for FREEZE, we don't want to truncate a frozen database */
grab_crit(gv_cur_region);
if (FROZEN_CHILLED(cs_data))
DO_CHILLED_AUTORELEASE(csa, cs_data);
if (!FROZEN(cs_data) && !IS_REPL_INST_FROZEN)
break;
rel_crit(gv_cur_region);
while (FROZEN(cs_data) || IS_REPL_INST_FROZEN)
{
hiber_start(1000);
if (FROZEN_CHILLED(cs_data) && CHILLED_AUTORELEASE(cs_data))
break;
}
}
assert(csa->nl->trunc_pid == process_id);
/* Flush pending updates to disk. If this is not done, old updates can be flushed AFTER ftruncate, extending the file. */
if (!wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_MSYNC_DB))
{
assert(FALSE);
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_BUFFLUFAILED, 4, LEN_AND_LIT("MUPIP REORG TRUNCATE"),
DB_LEN_STR(gv_cur_region));
rel_crit(gv_cur_region);
return FALSE;
}
csa->nl->highest_lbm_with_busy_blk = MAX(found_busy_blk, csa->nl->highest_lbm_with_busy_blk);
assert(IS_BITMAP_BLK(csa->nl->highest_lbm_with_busy_blk));
new_total = MIN(old_total, csa->nl->highest_lbm_with_busy_blk + BLKS_PER_LMAP);
if (mu_ctrly_occurred || mu_ctrlc_occurred)
{
rel_crit(gv_cur_region);
return TRUE;
} else if (csa->ti->total_blks != old_total || new_total == old_total)
{
assert(csa->ti->total_blks >= old_total); /* Better have been an extend, not a truncate... */
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(5) ERR_MUTRUNCNOSPACE, 3, REG_LEN_STR(gv_cur_region), truncate_percent);
rel_crit(gv_cur_region);
return TRUE;
} else if (GDSVCURR != csd->desired_db_format || csd->blks_to_upgrd != 0 || !csd->fully_upgraded)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_MUTRUNCNOV4, 2, REG_LEN_STR(gv_cur_region));
rel_crit(gv_cur_region);
return TRUE;
} else if (SNAPSHOTS_IN_PROG(csa->nl))
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_MUTRUNCSSINPROG, 2, REG_LEN_STR(gv_cur_region));
rel_crit(gv_cur_region);
return TRUE;
} else if (BACKUP_NOT_IN_PROGRESS != cs_addrs->nl->nbb)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_MUTRUNCBACKINPROG, 2, REG_LEN_STR(gv_cur_region));
rel_crit(gv_cur_region);
return TRUE;
}
DEFER_INTERRUPTS(INTRPT_IN_TRUNC, prev_intrpt_state);
if (JNL_ENABLED(csa))
{ /* Write JRT_TRUNC and INCTN records */
if (!jgbl.dont_reset_gbl_jrec_time)
SET_GBL_JREC_TIME; /* needed before jnl_ensure_open as that can write jnl records */
jpc = csa->jnl;
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(gv_cur_region, csa);
if (SS_NORMAL != jnl_status)
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region));
else
{
if (0 == jpc->pini_addr)
jnl_put_jrt_pini(csa);
jnl_write_trunc_rec(csa, old_total, csa->ti->free_blocks, new_total);
inctn_opcode = inctn_mu_reorg;
jnl_write_inctn_rec(csa);
jnl_status = jnl_flush(gv_cur_region);
if (SS_NORMAL != jnl_status)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_JNLFLUSH, 2, JNL_LEN_STR(csd),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error with journal flush during mu_truncate"),
jnl_status);
assert(NOJNL == jpc->channel); /* jnl file lost has been triggered */
}
}
}
/* Good to go ahead and REALLY truncate (reduce total_blks, clear cache_array, FTRUNCATE) */
curr_tn = csa->ti->curr_tn;
CHECK_TN(csa, csd, curr_tn);
udi = FILE_INFO(gv_cur_region);
/* Information used by recover_truncate to check if the file size and csa->ti->total_blks are INCONSISTENT */
trunc_file_size = BLK_ZERO_OFF(csd->start_vbn) + ((off_t)csd->blk_size * (new_total + 1));
csd->after_trunc_total_blks = new_total;
csd->before_trunc_free_blocks = csa->ti->free_blocks;
csd->before_trunc_total_blks = old_total; /* Flags interrupted truncate for recover_truncate */
/* file size and total blocks: INCONSISTENT */
csa->ti->total_blks = new_total;
/* past the point of no return -- shared memory intact */
assert(csa->ti->free_blocks >= DELTA_FREE_BLOCKS(old_total, new_total));
csa->ti->free_blocks -= DELTA_FREE_BLOCKS(old_total, new_total);
new_free = csa->ti->free_blocks;
KILL_TRUNC_TEST(WBTEST_CRASH_TRUNCATE_1); /* 55 : Issue a kill -9 before 1st fsync */
fileheader_sync(gv_cur_region);
DB_FSYNC(gv_cur_region, udi, csa, db_fsync_in_prog, save_errno);
CHECK_DBSYNC(gv_cur_region, save_errno);
/* past the point of no return -- shared memory deleted */
KILL_TRUNC_TEST(WBTEST_CRASH_TRUNCATE_2); /* 56 : Issue a kill -9 after 1st fsync */
clear_cache_array(csa, csd, gv_cur_region, new_total, old_total);
offset = (off_t)BLK_ZERO_OFF(csd->start_vbn) + (off_t)new_total * csd->blk_size;
save_errno = db_write_eof_block(udi, udi->fd, csd->blk_size, offset, &(TREF(dio_buff)));
if (0 != save_errno)
{
err_msg = (char *)STRERROR(errno);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_MUTRUNCERROR, 4, REG_LEN_STR(gv_cur_region), LEN_AND_STR(err_msg));
return FALSE;
}
KILL_TRUNC_TEST(WBTEST_CRASH_TRUNCATE_3); /* 57 : Issue a kill -9 after reducing csa->ti->total_blks, before FTRUNCATE */
/* Execute an ftruncate() and truncate the DB file
* ftruncate() is a SYSTEM CALL on almost all platforms (except SunOS)
* It ignores kill -9 signal till its operation is completed.
* So we can safely assume that the result of ftruncate() will be complete.
*/
FTRUNCATE(FILE_INFO(gv_cur_region)->fd, trunc_file_size, ftrunc_status);
if (0 != ftrunc_status)
{
err_msg = (char *)STRERROR(errno);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_MUTRUNCERROR, 4, REG_LEN_STR(gv_cur_region), LEN_AND_STR(err_msg));
/* should go through recover_truncate now, which will again try to FTRUNCATE */
return FALSE;
}
/* file size and total blocks: CONSISTENT (shrunk) */
KILL_TRUNC_TEST(WBTEST_CRASH_TRUNCATE_4); /* 58 : Issue a kill -9 after FTRUNCATE, before 2nd fsync */
csa->nl->root_search_cycle++; /* Force concurrent processes to restart in t_end/tp_tend to make sure no one
* tries to commit updates past the end of the file. Bitmap validations together
* with highest_lbm_with_busy_blk should actually be sufficient, so this is
* just to be safe.
*/
csd->before_trunc_total_blks = 0; /* indicate CONSISTENT */
/* Increment TN */
assert(csa->ti->early_tn == csa->ti->curr_tn);
csd->trans_hist.early_tn = csd->trans_hist.curr_tn + 1;
INCREMENT_CURR_TN(csd);
fileheader_sync(gv_cur_region);
DB_FSYNC(gv_cur_region, udi, csa, db_fsync_in_prog, save_errno);
KILL_TRUNC_TEST(WBTEST_CRASH_TRUNCATE_5); /* 58 : Issue a kill -9 after after 2nd fsync */
CHECK_DBSYNC(gv_cur_region, save_errno);
ENABLE_INTERRUPTS(INTRPT_IN_TRUNC, prev_intrpt_state);
curr_tn = csa->ti->curr_tn;
rel_crit(gv_cur_region);
send_msg_csa(CSA_ARG(csa) VARLSTCNT(7) ERR_MUTRUNCSUCCESS, 5, DB_LEN_STR(gv_cur_region), old_total, new_total, &curr_tn);
util_out_print("Truncated region: !AD. Reduced total blocks from [!UL] to [!UL]. Reduced free blocks from [!UL] to [!UL].",
FLUSH, REG_LEN_STR(gv_cur_region), old_total, new_total, old_free, new_free);
return TRUE;
} /* END of mu_truncate() */
int jnl_file_extend(jnl_private_control *jpc, uint4 total_jnl_rec_size)
{
file_control *fc;
boolean_t need_extend;
jnl_buffer_ptr_t jb;
jnl_create_info jnl_info;
jnl_file_header header;
uint4 new_alq;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
char prev_jnl_fn[JNL_NAME_SIZE];
uint4 jnl_status = 0, status;
int new_blocks, result;
GTM_BAVAIL_TYPE avail_blocks;
uint4 aligned_tot_jrec_size, count;
switch(jpc->region->dyn.addr->acc_meth)
{
case dba_mm:
case dba_bg:
csa = &FILE_INFO(jpc->region)->s_addrs;
break;
default:
GTMASSERT;
}
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
assert(csa->now_crit || (csd->clustered && (CCST_CLOSED == csa->nl->ccp_state)));
assert(jpc->region == gv_cur_region);
assert(csa->jnl_state == csd->jnl_state);
if (!JNL_ENABLED(csa) || (NOJNL == jpc->channel) || (JNL_FILE_SWITCHED(jpc)))
GTMASSERT; /* crit and messing with the journal file - how could it have vanished? */
if (!csd->jnl_deq)
{
assert(DIVIDE_ROUND_UP(total_jnl_rec_size, DISK_BLOCK_SIZE) <= csd->jnl_alq);
assert(csd->jnl_alq == csd->autoswitchlimit);
new_blocks = csd->jnl_alq;
} else
/* May cause extension of csd->jnl_deq * n blocks where n > 0 */
new_blocks = ROUND_UP(DIVIDE_ROUND_UP(total_jnl_rec_size, DISK_BLOCK_SIZE), csd->jnl_deq);
jpc->status = SS_NORMAL;
jb = jpc->jnl_buff;
assert(0 <= new_blocks);
DEBUG_ONLY(count = 0);
for (need_extend = (0 != new_blocks); need_extend; )
{
DEBUG_ONLY(count++);
/* usually we will do the loop just once where we do the file extension.
* rarely we might need to do an autoswitch instead after which again rarely
* we might need to do an extension on the new journal to fit in the transaction's journal requirements.
* therefore we should do this loop a maximum of twice. hence the assert below.
*/
assert(count <= 2);
need_extend = FALSE;
if (SS_NORMAL == (status = disk_block_available(jpc->channel, &avail_blocks, TRUE)))
{
if ((new_blocks * EXTEND_WARNING_FACTOR) > avail_blocks)
{
if (new_blocks > avail_blocks)
{ /* if we cannot satisfy the request, it is an error */
send_msg(VARLSTCNT(6) ERR_NOSPACEEXT, 4, JNL_LEN_STR(csd),
new_blocks, avail_blocks);
new_blocks = 0;
jpc->status = SS_NORMAL;
break;
} else
send_msg(VARLSTCNT(5) ERR_DSKSPACEFLOW, 3, JNL_LEN_STR(csd),
(avail_blocks - new_blocks));
}
} else
send_msg(VARLSTCNT(5) ERR_JNLFILEXTERR, 2, JNL_LEN_STR(csd), status);
new_alq = jb->filesize + new_blocks;
/* ensure current journal file size is well within autoswitchlimit --> design constraint */
assert(csd->autoswitchlimit >= jb->filesize);
if (csd->autoswitchlimit < (jb->filesize + (EXTEND_WARNING_FACTOR * new_blocks))) /* close to max */
send_msg(VARLSTCNT(5) ERR_JNLSPACELOW, 3, JNL_LEN_STR(csd), csd->autoswitchlimit - jb->filesize);
if (csd->autoswitchlimit < new_alq)
{ /* Reached max, need to autoswitch */
/* Ensure new journal file can hold the entire current transaction's journal record requirements */
assert(csd->autoswitchlimit >= MAX_REQD_JNL_FILE_SIZE(total_jnl_rec_size));
memset(&jnl_info, 0, sizeof(jnl_info));
jnl_info.prev_jnl = &prev_jnl_fn[0];
set_jnl_info(gv_cur_region, &jnl_info);
assert(JNL_ENABLED(csa) && (NOJNL != jpc->channel) && !(JNL_FILE_SWITCHED(jpc)));
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{ /* flush the cache and jnl-buffer-contents to current journal file before
* switching to a new journal. */
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH);
jnl_file_close(gv_cur_region, TRUE, TRUE);
} else
rts_error(VARLSTCNT(7) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region), jpc->status);
assert(!jgbl.forw_phase_recovery || (NULL != jgbl.mur_pini_addr_reset_fnptr));
if (jgbl.forw_phase_recovery && (NULL != jgbl.mur_pini_addr_reset_fnptr))
(*jgbl.mur_pini_addr_reset_fnptr)();
assert(!jnl_info.no_rename);
assert(!jnl_info.no_prev_link);
if (EXIT_NRM == cre_jnl_file(&jnl_info))
{
assert(0 == memcmp(csd->jnl_file_name, jnl_info.jnl, jnl_info.jnl_len));
assert(csd->jnl_file_name[jnl_info.jnl_len] == '\0');
assert(csd->jnl_file_len == jnl_info.jnl_len);
assert(csd->jnl_buffer_size == jnl_info.buffer);
assert(csd->jnl_alq == jnl_info.alloc);
assert(csd->jnl_deq == jnl_info.extend);
assert(csd->jnl_before_image == jnl_info.before_images);
csd->trans_hist.header_open_tn = jnl_info.tn; /* needed for successful jnl_file_open() */
send_msg(VARLSTCNT(4) ERR_NEWJNLFILECREATE, 2, JNL_LEN_STR(csd));
fc = gv_cur_region->dyn.addr->file_cntl;
fc->op = FC_WRITE;
fc->op_buff = (sm_uc_ptr_t)csd;
status = dbfilop(fc);
if (SS_NORMAL != status)
send_msg(VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), status);
assert(JNL_ENABLED(csa));
/* call jnl_ensure_open instead of jnl_file_open to make sure jpc->pini_addr is set to 0 */
jnl_status = jnl_ensure_open(); /* sets jpc->status */
if (0 != jnl_status)
rts_error(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region));
assert(jb->filesize == csd->jnl_alq);
aligned_tot_jrec_size = ALIGNED_ROUND_UP(MAX_REQD_JNL_FILE_SIZE(total_jnl_rec_size),
csd->jnl_alq, csd->jnl_deq);
if (aligned_tot_jrec_size > csd->jnl_alq)
{ /* need to extend more than initial allocation in the new journal file
* to accommodate the current transaction.
*/
new_blocks = aligned_tot_jrec_size - csd->jnl_alq;
assert(new_blocks);
assert(0 == new_blocks % csd->jnl_deq);
need_extend = TRUE;
}
} else
{
send_msg(VARLSTCNT(4) ERR_JNLCREATERR, 2, JNL_LEN_STR(csd));
jpc->status = ERR_JNLNOCREATE;
new_blocks = -1;
}
} else
{
assert(!need_extend); /* ensure we won't go through the for loop again */
/* Virtually extend currently used journal file */
jb->filesize = new_alq; /* Actually this is virtual file size blocks */
DO_FILE_READ(jpc->channel, 0, &header, JNL_HDR_LEN, jpc->status, jpc->status2);
if (SS_NORMAL != jpc->status)
rts_error(VARLSTCNT(5) ERR_JNLRDERR, 2, JNL_LEN_STR(csd), jpc->status);
assert((header.virtual_size + new_blocks) == new_alq);
header.virtual_size = new_alq;
DO_FILE_WRITE(jpc->channel, 0, &header, JNL_HDR_LEN, jpc->status, jpc->status2);
if (SS_NORMAL != jpc->status)
rts_error(VARLSTCNT(5) ERR_JNLWRERR, 2, JNL_LEN_STR(csd), jpc->status);
}
if (0 >= new_blocks)
break;
}
if (0 >= new_blocks)
{
jpc->status = ERR_JNLREADEOF;
jnl_file_lost(jpc, ERR_JNLEXTEND);
new_blocks = -1;
}
return new_blocks;
}
int4 mupip_set_file(int db_fn_len, char *db_fn)
{
bool got_standalone;
boolean_t bypass_partial_recov, need_standalone = FALSE;
char acc_spec[MAX_ACC_METH_LEN], ver_spec[MAX_DB_VER_LEN], exit_stat, *fn;
unsigned short acc_spec_len = MAX_ACC_METH_LEN, ver_spec_len = MAX_DB_VER_LEN;
int fd, fn_len;
int4 status;
int4 status1;
int glbl_buff_status, defer_status, rsrvd_bytes_status,
extn_count_status, lock_space_status, disk_wait_status;
int4 new_disk_wait, new_cache_size, new_extn_count, new_lock_space, reserved_bytes, defer_time;
sgmnt_data_ptr_t csd;
tp_region *rptr, single;
enum db_acc_method access, access_new;
enum db_ver desired_dbver;
gd_region *temp_cur_region;
char *errptr, *command = "MUPIP SET VERSION";
int save_errno;
error_def(ERR_DBPREMATEOF);
error_def(ERR_DBRDERR);
error_def(ERR_DBRDONLY);
error_def(ERR_INVACCMETHOD);
error_def(ERR_MUNOACTION);
error_def(ERR_RBWRNNOTCHG);
error_def(ERR_WCERRNOTCHG);
error_def(ERR_WCWRNNOTCHG);
error_def(ERR_MMNODYNDWNGRD);
exit_stat = EXIT_NRM;
defer_status = cli_present("DEFER_TIME");
if (defer_status)
need_standalone = TRUE;
bypass_partial_recov = cli_present("PARTIAL_RECOV_BYPASS") == CLI_PRESENT;
if (bypass_partial_recov)
need_standalone = TRUE;
if (disk_wait_status = cli_present("WAIT_DISK"))
{
if (cli_get_int("WAIT_DISK", &new_disk_wait))
{
if (new_disk_wait < 0)
{
util_out_print("!UL negative, minimum WAIT_DISK allowed is 0.", TRUE, new_disk_wait);
return (int4)ERR_WCWRNNOTCHG;
}
need_standalone = TRUE;
} else
{
util_out_print("Error getting WAIT_DISK qualifier value", TRUE);
return (int4)ERR_WCWRNNOTCHG;
}
}
if (glbl_buff_status = cli_present("GLOBAL_BUFFERS"))
{
if (cli_get_int("GLOBAL_BUFFERS", &new_cache_size))
{
if (new_cache_size > WC_MAX_BUFFS)
{
util_out_print("!UL too large, maximum write cache buffers allowed is !UL", TRUE, new_cache_size,
WC_MAX_BUFFS);
return (int4)ERR_WCWRNNOTCHG;
}
if (new_cache_size < WC_MIN_BUFFS)
{
util_out_print("!UL too small, minimum cache buffers allowed is !UL", TRUE, new_cache_size,
WC_MIN_BUFFS);
return (int4)ERR_WCWRNNOTCHG;
}
} else
{
util_out_print("Error getting GLOBAL BUFFER qualifier value", TRUE);
return (int4)ERR_WCWRNNOTCHG;
}
need_standalone = TRUE;
}
/* EXTENSION_COUNT does not require standalone access and hence need_standalone will not be set to TRUE for this. */
if (extn_count_status = cli_present("EXTENSION_COUNT"))
{
if (cli_get_int("EXTENSION_COUNT", &new_extn_count))
{
if (new_extn_count > MAX_EXTN_COUNT)
{
util_out_print("!UL too large, maximum extension count allowed is !UL", TRUE, new_extn_count,
MAX_EXTN_COUNT);
return (int4)ERR_WCWRNNOTCHG;
}
if (new_extn_count < MIN_EXTN_COUNT)
{
util_out_print("!UL too small, minimum extension count allowed is !UL", TRUE, new_extn_count,
MIN_EXTN_COUNT);
return (int4)ERR_WCWRNNOTCHG;
}
} else
{
util_out_print("Error getting EXTENSION COUNT qualifier value", TRUE);
return (int4)ERR_WCWRNNOTCHG;
}
}
if (lock_space_status = cli_present("LOCK_SPACE"))
{
if (cli_get_int("LOCK_SPACE", &new_lock_space))
{
if (new_lock_space > MAX_LOCK_SPACE)
{
util_out_print("!UL too large, maximum lock space allowed is !UL", TRUE,
new_lock_space, MAX_LOCK_SPACE);
return (int4)ERR_WCWRNNOTCHG;
}
else if (new_lock_space < MIN_LOCK_SPACE)
{
util_out_print("!UL too small, minimum lock space allowed is !UL", TRUE,
new_lock_space, MIN_LOCK_SPACE);
return (int4)ERR_WCWRNNOTCHG;
}
} else
{
util_out_print("Error getting LOCK_SPACE qualifier value", TRUE);
return (int4)ERR_WCWRNNOTCHG;
}
need_standalone = TRUE;
}
if (rsrvd_bytes_status = cli_present("RESERVED_BYTES"))
{
if (!cli_get_int("RESERVED_BYTES", &reserved_bytes))
{
util_out_print("Error getting RESERVED BYTES qualifier value", TRUE);
return (int4)ERR_RBWRNNOTCHG;
}
need_standalone = TRUE;
}
if (cli_present("ACCESS_METHOD"))
{
cli_get_str("ACCESS_METHOD", acc_spec, &acc_spec_len);
cli_strupper(acc_spec);
if (0 == memcmp(acc_spec, "MM", acc_spec_len))
access = dba_mm;
else if (0 == memcmp(acc_spec, "BG", acc_spec_len))
access = dba_bg;
else
mupip_exit(ERR_INVACCMETHOD);
need_standalone = TRUE;
} else
access = n_dba; /* really want to keep current method,
which has not yet been read */
if (cli_present("VERSION"))
{
assert(!need_standalone);
cli_get_str("VERSION", ver_spec, &ver_spec_len);
cli_strupper(ver_spec);
if (0 == memcmp(ver_spec, "V4", ver_spec_len))
desired_dbver = GDSV4;
else if (0 == memcmp(ver_spec, "V5", ver_spec_len))
desired_dbver = GDSV5;
else
GTMASSERT; /* CLI should prevent us ever getting here */
} else
desired_dbver = GDSVLAST; /* really want to keep version, which has not yet been read */
if (region)
rptr = grlist;
else
{
rptr = &single;
memset(&single, 0, sizeof(single));
}
csd = (sgmnt_data *)malloc(ROUND_UP(sizeof(sgmnt_data), DISK_BLOCK_SIZE));
in_backup = FALSE; /* Only want yes/no from mupfndfil, not an address */
for (; rptr != NULL; rptr = rptr->fPtr)
{
if (region)
{
if (dba_usr == rptr->reg->dyn.addr->acc_meth)
{
util_out_print("!/Region !AD is not a GDS access type", TRUE, REG_LEN_STR(rptr->reg));
exit_stat |= EXIT_WRN;
continue;
}
if (!mupfndfil(rptr->reg, NULL))
continue;
fn = (char *)rptr->reg->dyn.addr->fname;
fn_len = rptr->reg->dyn.addr->fname_len;
} else
{
fn = db_fn;
fn_len = db_fn_len;
}
mu_gv_cur_reg_init();
strcpy((char *)gv_cur_region->dyn.addr->fname, fn);
gv_cur_region->dyn.addr->fname_len = fn_len;
if (!need_standalone)
{
gvcst_init(gv_cur_region);
change_reg(); /* sets cs_addrs and cs_data */
if (gv_cur_region->read_only)
{
gtm_putmsg(VARLSTCNT(4) ERR_DBRDONLY, 2, DB_LEN_STR(gv_cur_region));
exit_stat |= EXIT_ERR;
gds_rundown();
mu_gv_cur_reg_free();
continue;
}
grab_crit(gv_cur_region);
status = EXIT_NRM;
access_new = (n_dba == access ? cs_data->acc_meth : access);
/* recalculate; n_dba is a proxy for no change */
change_fhead_timer("FLUSH_TIME", cs_data->flush_time,
(dba_bg == access_new ? TIM_FLU_MOD_BG : TIM_FLU_MOD_MM),
FALSE);
if (GDSVLAST != desired_dbver)
{
if ((dba_mm != access_new) || (GDSV4 != desired_dbver))
status1 = desired_db_format_set(gv_cur_region, desired_dbver, command);
else
{
status1 = ERR_MMNODYNDWNGRD;
gtm_putmsg(VARLSTCNT(4) status1, 2, REG_LEN_STR(gv_cur_region));
}
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 = EXIT_ERR;
}
}
}
if (EXIT_NRM == status)
{
if (extn_count_status)
cs_data->extension_size = (uint4)new_extn_count;
wcs_flu(WCSFLU_FLUSH_HDR);
if (extn_count_status)
util_out_print("Database file !AD now has extension count !UL",
TRUE, fn_len, fn, cs_data->extension_size);
if (GDSVLAST != desired_dbver)
util_out_print("Database file !AD now has desired DB format !AD", TRUE,
fn_len, fn, LEN_AND_STR(gtm_dbversion_table[cs_data->desired_db_format]));
} else
exit_stat |= status;
rel_crit(gv_cur_region);
gds_rundown();
} else
{ /* Following part needs standalone access */
assert(GDSVLAST == desired_dbver);
got_standalone = mu_rndwn_file(gv_cur_region, TRUE);
if (FALSE == got_standalone)
return (int4)ERR_WCERRNOTCHG;
/* we should open it (for changing) after mu_rndwn_file, since mu_rndwn_file changes the file header too */
if (-1 == (fd = OPEN(fn, O_RDWR)))
{
save_errno = errno;
errptr = (char *)STRERROR(save_errno);
util_out_print("open : !AZ", TRUE, errptr);
exit_stat |= EXIT_ERR;
db_ipcs_reset(gv_cur_region, FALSE);
mu_gv_cur_reg_free();
continue;
}
LSEEKREAD(fd, 0, csd, sizeof(sgmnt_data), status);
if (0 != status)
{
save_errno = errno;
PERROR("Error reading header of file");
errptr = (char *)STRERROR(save_errno);
util_out_print("read : !AZ", TRUE, errptr);
util_out_print("Error reading header of file", TRUE);
util_out_print("Database file !AD not changed: ", TRUE, fn_len, fn);
if (-1 != status)
rts_error(VARLSTCNT(4) ERR_DBRDERR, 2, fn_len, fn);
else
rts_error(VARLSTCNT(4) ERR_DBPREMATEOF, 2, fn_len, fn);
}
if (rsrvd_bytes_status)
{
if (reserved_bytes > MAX_RESERVE_B(csd))
{
util_out_print("!UL too large, maximum reserved bytes allowed is !UL for database file !AD",
TRUE, reserved_bytes, MAX_RESERVE_B(csd), fn_len, fn);
close(fd);
db_ipcs_reset(gv_cur_region, FALSE);
return (int4)ERR_RBWRNNOTCHG;
}
csd->reserved_bytes = reserved_bytes;
}
access_new = (n_dba == access ? csd->acc_meth : access);
/* recalculate; n_dba is a proxy for no change */
change_fhead_timer("FLUSH_TIME", csd->flush_time,
(dba_bg == access_new ? TIM_FLU_MOD_BG : TIM_FLU_MOD_MM),
FALSE);
if ((n_dba != access) && (csd->acc_meth != access)) /* n_dba is a proxy for no change */
{
if (dba_mm == access)
csd->defer_time = 1; /* defer defaults to 1 */
csd->acc_meth = access;
if (0 == csd->n_bts)
{
csd->n_bts = WC_DEF_BUFFS;
csd->bt_buckets = getprime(csd->n_bts);
}
}
if (glbl_buff_status)
{
csd->n_bts = BT_FACTOR(new_cache_size);
csd->bt_buckets = getprime(csd->n_bts);
csd->n_wrt_per_flu = 7;
csd->flush_trigger = FLUSH_FACTOR(csd->n_bts);
}
if (disk_wait_status)
csd->wait_disk_space = new_disk_wait;
if (extn_count_status)
csd->extension_size = (uint4)new_extn_count;
if (lock_space_status)
csd->lock_space_size = (uint4)new_lock_space * OS_PAGELET_SIZE;
if (bypass_partial_recov)
{
csd->file_corrupt = FALSE;
util_out_print("Database file !AD now has partial recovery flag set to !UL(FALSE) ",
TRUE, fn_len, fn, csd->file_corrupt);
}
if (dba_mm == access_new)
{
if (CLI_NEGATED == defer_status)
csd->defer_time = 0;
else if (CLI_PRESENT == defer_status)
{
if (!cli_get_num("DEFER_TIME", &defer_time))
{
util_out_print("Error getting DEFER_TIME qualifier value", TRUE);
db_ipcs_reset(gv_cur_region, FALSE);
return (int4)ERR_RBWRNNOTCHG;
}
if (-1 > defer_time)
{
util_out_print("DEFER_TIME cannot take negative values less than -1", TRUE);
util_out_print("Database file !AD not changed", TRUE, fn_len, fn);
exit_stat |= EXIT_WRN;
db_ipcs_reset(gv_cur_region, FALSE);
mu_gv_cur_reg_free();
continue;
}
csd->defer_time = defer_time;
}
if (csd->blks_to_upgrd)
{
util_out_print("MM access method cannot be set if there are blocks to upgrade", TRUE);
util_out_print("Database file !AD not changed", TRUE, fn_len, fn);
exit_stat |= EXIT_WRN;
db_ipcs_reset(gv_cur_region, FALSE);
mu_gv_cur_reg_free();
continue;
}
if (GDSVCURR != csd->desired_db_format)
{
util_out_print("MM access method cannot be set in DB compatibility mode",
TRUE);
util_out_print("Database file !AD not changed", TRUE, fn_len, fn);
exit_stat |= EXIT_WRN;
db_ipcs_reset(gv_cur_region, FALSE);
mu_gv_cur_reg_free();
continue;
}
if (JNL_ENABLED(csd) && csd->jnl_before_image)
{
util_out_print("MM access method cannot be set with BEFORE image journaling", TRUE);
util_out_print("Database file !AD not changed", TRUE, fn_len, fn);
exit_stat |= EXIT_WRN;
db_ipcs_reset(gv_cur_region, FALSE);
mu_gv_cur_reg_free();
continue;
}
csd->jnl_before_image = FALSE;
} else
{
if (defer_status)
{
util_out_print("DEFER cannot be specified with BG access method.", TRUE);
util_out_print("Database file !AD not changed", TRUE, fn_len, fn);
exit_stat |= EXIT_WRN;
db_ipcs_reset(gv_cur_region, FALSE);
mu_gv_cur_reg_free();
continue;
}
}
LSEEKWRITE(fd, 0, csd, sizeof(sgmnt_data), status);
if (0 != status)
{
save_errno = errno;
errptr = (char *)STRERROR(save_errno);
util_out_print("write : !AZ", TRUE, errptr);
util_out_print("Error writing header of file", TRUE);
util_out_print("Database file !AD not changed: ", TRUE, fn_len, fn);
rts_error(VARLSTCNT(4) ERR_DBRDERR, 2, fn_len, fn);
}
close(fd);
/* --------------------- report results ------------------------- */
if (glbl_buff_status)
util_out_print("Database file !AD now has !UL global buffers",
TRUE, fn_len, fn, csd->n_bts);
if (defer_status && (dba_mm == csd->acc_meth))
util_out_print("Database file !AD now has defer_time set to !SL",
TRUE, fn_len, fn, csd->defer_time);
if (rsrvd_bytes_status)
util_out_print("Database file !AD now has !UL reserved bytes",
TRUE, fn_len, fn, csd->reserved_bytes);
if (extn_count_status)
util_out_print("Database file !AD now has extension count !UL",
TRUE, fn_len, fn, csd->extension_size);
if (lock_space_status)
util_out_print("Database file !AD now has lock space !UL pages",
TRUE, fn_len, fn, csd->lock_space_size/OS_PAGELET_SIZE);
if (disk_wait_status)
util_out_print("Database file !AD now has wait disk set to !UL seconds",
TRUE, fn_len, fn, csd->wait_disk_space);
db_ipcs_reset(gv_cur_region, FALSE);
} /* end of else part if (!need_standalone) */
mu_gv_cur_reg_free();
}
free(csd);
assert(!(exit_stat & EXIT_INF));
return (exit_stat & EXIT_ERR ? (int4)ERR_WCERRNOTCHG :
(exit_stat & EXIT_WRN ? (int4)ERR_WCWRNNOTCHG : SS_NORMAL));
}
int4 gds_rundown(void)
{
boolean_t canceled_dbsync_timer, canceled_flush_timer, ok_to_write_pfin;
boolean_t have_standalone_access, ipc_deleted, err_caught;
boolean_t is_cur_process_ss_initiator, remove_shm, vermismatch, we_are_last_user, we_are_last_writer, is_mm;
boolean_t unsafe_last_writer;
char time_str[CTIME_BEFORE_NL + 2]; /* for GET_CUR_TIME macro */
gd_region *reg;
int save_errno, status, rc;
int4 semval, ftok_semval, sopcnt, ftok_sopcnt;
short crash_count;
sm_long_t munmap_len;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
node_local_ptr_t cnl;
struct shmid_ds shm_buf;
struct sembuf sop[2], ftok_sop[2];
uint4 jnl_status;
unix_db_info *udi;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
shm_snapshot_t *ss_shm_ptr;
uint4 ss_pid, onln_rlbk_pid, holder_pid;
boolean_t was_crit;
boolean_t safe_mode; /* Do not flush or take down shared memory. */
boolean_t bypassed_ftok = FALSE, bypassed_access = FALSE, may_bypass_ftok, inst_is_frozen,
ftok_counter_halted,
access_counter_halted;
int secshrstat;
intrpt_state_t prev_intrpt_state;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
jnl_status = 0;
reg = gv_cur_region; /* Local copy */
/* early out for cluster regions
* to avoid tripping the assert below.
* Note:
* This early out is consistent with VMS. It has been
* noted that all of the gtcm assignments
* to gv_cur_region should use the TP_CHANGE_REG
* macro. This would also avoid the assert problem
* and should be done eventually.
*/
if (dba_cm == reg->dyn.addr->acc_meth)
return EXIT_NRM;
udi = FILE_INFO(reg);
csa = &udi->s_addrs;
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
if ((reg->open) && (dba_usr == csd->acc_meth))
{
change_reg();
gvusr_rundown();
return EXIT_NRM;
}
/* If the process has standalone access, it has udi->grabbed_access_sem set to TRUE at this point. Note that down in a local
* variable as the udi->grabbed_access_sem is set to TRUE even for non-standalone access below and hence we can't rely on
* that later to determine if the process had standalone access or not when it entered this function. We need to guarantee
* that none else access database file header when semid/shmid fields are reset. We already have created ftok semaphore in
* db_init or, mu_rndwn_file and did not remove it. So just lock it. We do it in blocking mode.
*/
have_standalone_access = udi->grabbed_access_sem; /* process holds standalone access */
DEFER_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
ESTABLISH_NORET(gds_rundown_ch, err_caught);
if (err_caught)
{
REVERT;
WITH_CH(gds_rundown_ch, gds_rundown_err_cleanup(have_standalone_access), 0);
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
DEBUG_ONLY(ok_to_UNWIND_in_exit_handling = FALSE);
return EXIT_ERR;
}
assert(reg->open); /* if we failed to open, dbinit_ch should have taken care of proper clean up */
assert(!reg->opening); /* see comment above */
assert((dba_bg == csd->acc_meth) || (dba_mm == csd->acc_meth));
is_mm = (dba_bg != csd->acc_meth);
assert(!csa->hold_onto_crit || (csa->now_crit && jgbl.onlnrlbk));
/* If we are online rollback, we should already be holding crit and should release it only at the end of this module. This
* is usually done by noting down csa->now_crit in a local variable (was_crit) and using it whenever we are about to
* grab_crit. But, there are instances (like mupip_set_journal.c) where we grab_crit but invoke gds_rundown without any
* preceeding rel_crit. Such code relies on the fact that gds_rundown does rel_crit unconditionally (to get locks to a known
* state). So, augment csa->now_crit with jgbl.onlnrlbk to track if we can rel_crit unconditionally or not in gds_rundown.
*/
was_crit = (csa->now_crit && jgbl.onlnrlbk);
/* Cancel any pending flush timer for this region by this task */
canceled_flush_timer = FALSE;
canceled_dbsync_timer = FALSE;
CANCEL_DB_TIMERS(reg, csa, canceled_flush_timer, canceled_dbsync_timer);
we_are_last_user = FALSE;
inst_is_frozen = IS_REPL_INST_FROZEN && REPL_ALLOWED(csa->hdr);
if (!csa->persistent_freeze)
region_freeze(reg, FALSE, FALSE, FALSE);
if (!was_crit)
{
rel_crit(reg); /* get locks to known state */
mutex_cleanup(reg);
}
/* The only process that can invoke gds_rundown while holding access control semaphore is RECOVER/ROLLBACK. All the others
* (like MUPIP SET -FILE/MUPIP EXTEND would have invoked db_ipcs_reset() before invoking gds_rundown (from
* mupip_exit_handler). The only exception is when these processes encounter a terminate signal and they reach
* mupip_exit_handler while holding access control semaphore. Assert accordingly.
*/
assert(!have_standalone_access || mupip_jnl_recover || process_exiting);
/* If we have standalone access, then ensure that a concurrent online rollback cannot be running at the same time as it
* needs the access control lock as well. The only expection is we are online rollback and currently running down.
*/
cnl = csa->nl;
onln_rlbk_pid = cnl->onln_rlbk_pid;
assert(!have_standalone_access || mupip_jnl_recover || !onln_rlbk_pid || !is_proc_alive(onln_rlbk_pid, 0));
if (!have_standalone_access)
{
if (-1 == (ftok_semval = semctl(udi->ftok_semid, DB_COUNTER_SEM, GETVAL))) /* Check # of procs counted on FTOK */
{
save_errno = errno;
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get ftok_semval"), CALLFROM, errno);
}
may_bypass_ftok = CAN_BYPASS(ftok_semval, csd, inst_is_frozen); /* Do we need a blocking wait? */
/* We need to guarantee that no one else access database file header when semid/shmid fields are reset.
* We already have created ftok semaphore in db_init or mu_rndwn_file and did not remove it. So just lock it.
*/
if (!ftok_sem_lock(reg, may_bypass_ftok))
{
if (may_bypass_ftok)
{ /* We did a non-blocking wait. It's ok to proceed without locking */
bypassed_ftok = TRUE;
holder_pid = semctl(udi->ftok_semid, DB_CONTROL_SEM, GETPID);
if ((uint4)-1 == holder_pid)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get holder_pid"),
CALLFROM, errno);
if (!IS_GTM_IMAGE) /* MUMPS processes should not flood syslog with bypass messages. */
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_RESRCINTRLCKBYPAS, 10,
LEN_AND_STR(gtmImageNames[image_type].imageName), process_id, LEN_AND_LIT("FTOK"),
REG_LEN_STR(reg), DB_LEN_STR(reg), holder_pid);
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_TEXT, 2,
LEN_AND_LIT("FTOK bypassed at rundown"));
}
} else
{ /* We did a blocking wait but something bad happened. */
FTOK_TRACE(csa, csa->ti->curr_tn, ftok_ops_lock, process_id);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
}
}
sop[0].sem_num = DB_CONTROL_SEM; sop[0].sem_op = 0; /* Wait for 0 */
sop[1].sem_num = DB_CONTROL_SEM; sop[1].sem_op = 1; /* Lock */
sopcnt = 2;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO | IPC_NOWAIT; /* Don't wait the first time thru */
SEMOP(udi->semid, sop, sopcnt, status, NO_WAIT);
if (0 != status)
{
save_errno = errno;
/* Check # of processes counted on access sem. */
if (-1 == (semval = semctl(udi->semid, DB_COUNTER_SEM, GETVAL)))
{
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get semval"), CALLFROM, errno);
}
bypassed_access = CAN_BYPASS(semval, csd, inst_is_frozen) || onln_rlbk_pid || csd->file_corrupt;
/* Before attempting again in the blocking mode, see if the holding process is an online rollback.
* If so, it is likely we won't get the access control semaphore anytime soon. In that case, we
* are better off skipping rundown and continuing with sanity cleanup and exit.
*/
holder_pid = semctl(udi->semid, DB_CONTROL_SEM, GETPID);
if ((uint4)-1 == holder_pid)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get holder_pid"), CALLFROM, errno);
if (!bypassed_access)
{ /* We couldn't get it in one shot-- see if we already have it */
if (holder_pid == process_id)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(5) MAKE_MSG_INFO(ERR_CRITSEMFAIL), 2, DB_LEN_STR(reg),
ERR_RNDWNSEMFAIL);
REVERT;
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
assert(FALSE);
return EXIT_ERR;
}
if (EAGAIN != save_errno)
{
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMOP on access control semaphore"),
CALLFROM, save_errno);
}
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO; /* Try again - blocking this time */
SEMOP(udi->semid, sop, 2, status, FORCED_WAIT);
if (-1 == status) /* We couldn't get it at all.. */
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMOP on access control semaphore"),
CALLFROM, errno);
} else if (!IS_GTM_IMAGE)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_RESRCINTRLCKBYPAS, 10,
LEN_AND_STR(gtmImageNames[image_type].imageName), process_id,
LEN_AND_LIT("access control"), REG_LEN_STR(reg), DB_LEN_STR(reg), holder_pid);
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_TEXT, 2,
LEN_AND_LIT("Access control bypassed at rundown"));
}
udi->grabbed_access_sem = !bypassed_access;
}
} /* else we we hold the access control semaphore and therefore have standalone access. We do not release it now - we
* release it later in mupip_exit_handler.c. Since we already hold the access control semaphore, we don't need the
* ftok semaphore and trying it could cause deadlock
*/
/* Note that in the case of online rollback, "udi->grabbed_access_sem" (and in turn "have_standalone_access") is TRUE.
* But there could be other processes still having the database open so we cannot safely reset the halted fields.
*/
if (have_standalone_access && !jgbl.onlnrlbk)
csd->ftok_counter_halted = csd->access_counter_halted = FALSE;
ftok_counter_halted = csd->ftok_counter_halted;
access_counter_halted = csd->access_counter_halted;
/* If we bypassed any of the semaphores, activate safe mode.
* Also, if the replication instance is frozen and this db has replication turned on (which means
* no flushes of dirty buffers to this db can happen while the instance is frozen) activate safe mode.
*/
ok_to_write_pfin = !(bypassed_access || bypassed_ftok || inst_is_frozen);
safe_mode = !ok_to_write_pfin || ftok_counter_halted || access_counter_halted;
/* At this point we are guaranteed no one else is doing a db_init/rundown as we hold the access control semaphore */
assert(csa->ref_cnt); /* decrement private ref_cnt before shared ref_cnt decrement. */
csa->ref_cnt--; /* Currently journaling logic in gds_rundown() in VMS relies on this order to detect last writer */
assert(!csa->ref_cnt);
--cnl->ref_cnt;
if (memcmp(cnl->now_running, gtm_release_name, gtm_release_name_len + 1))
{ /* VERMISMATCH condition. Possible only if DSE */
assert(dse_running);
vermismatch = TRUE;
} else
vermismatch = FALSE;
if (-1 == shmctl(udi->shmid, IPC_STAT, &shm_buf))
{
save_errno = errno;
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown shmctl"), CALLFROM, save_errno);
} else
we_are_last_user = (1 == shm_buf.shm_nattch) && !vermismatch && !safe_mode;
/* recover => one user except ONLINE ROLLBACK, or standalone with frozen instance */
assert(!have_standalone_access || we_are_last_user || jgbl.onlnrlbk || inst_is_frozen);
if (-1 == (semval = semctl(udi->semid, DB_COUNTER_SEM, GETVAL)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get semval"), CALLFROM, errno);
/* There's one writer left and I am it */
assert(reg->read_only || semval >= 0);
unsafe_last_writer = (DB_COUNTER_SEM_INCR == semval) && (FALSE == reg->read_only) && !vermismatch;
we_are_last_writer = unsafe_last_writer && !safe_mode;
assert(!we_are_last_writer || !safe_mode);
assert(!we_are_last_user || !safe_mode);
/* recover + R/W region => one writer except ONLINE ROLLBACK, or standalone with frozen instance, leading to safe_mode */
assert(!(have_standalone_access && !reg->read_only) || we_are_last_writer || jgbl.onlnrlbk || inst_is_frozen);
GTM_WHITE_BOX_TEST(WBTEST_ANTIFREEZE_JNLCLOSE, we_are_last_writer, 1); /* Assume we are the last writer to invoke wcs_flu */
if (!have_standalone_access && (-1 == (ftok_semval = semctl(udi->ftok_semid, DB_COUNTER_SEM, GETVAL))))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get ftok_semval"), CALLFROM, errno);
if (NULL != csa->ss_ctx)
ss_destroy_context(csa->ss_ctx);
/* SS_MULTI: If multiple snapshots are supported, then we have to run through each of the snapshots */
assert(1 == MAX_SNAPSHOTS);
ss_shm_ptr = (shm_snapshot_ptr_t)SS_GETSTARTPTR(csa);
ss_pid = ss_shm_ptr->ss_info.ss_pid;
is_cur_process_ss_initiator = (process_id == ss_pid);
if (ss_pid && (is_cur_process_ss_initiator || we_are_last_user))
{
/* Try getting snapshot crit latch. If we don't get latch, we won't hang for eternity and will skip
* doing the orphaned snapshot cleanup. It will be cleaned up eventually either by subsequent MUPIP
* INTEG or by a MUPIP RUNDOWN.
*/
if (ss_get_lock_nowait(reg) && (ss_pid == ss_shm_ptr->ss_info.ss_pid)
&& (is_cur_process_ss_initiator || !is_proc_alive(ss_pid, 0)))
{
ss_release(NULL);
ss_release_lock(reg);
}
}
/* If cnl->donotflush_dbjnl is set, it means mupip recover/rollback was interrupted and therefore we need not flush
* shared memory contents to disk as they might be in an inconsistent state. Moreover, any more flushing will only cause
* future rollback to undo more journal records (PBLKs). In this case, we will go ahead and remove shared memory (without
* flushing the contents) in this routine. A reissue of the recover/rollback command will restore the database to a
* consistent state.
*/
if (!cnl->donotflush_dbjnl && !reg->read_only && !vermismatch)
{ /* If we had an orphaned block and were interrupted, set wc_blocked so we can invoke wcs_recover. Do it ONLY
* if there is NO concurrent online rollback running (as we need crit to set wc_blocked)
*/
if (csa->wbuf_dqd && !is_mm)
{ /* If we had an orphaned block and were interrupted, mupip_exit_handler will invoke secshr_db_clnup which
* will clear this field and so we should never come to gds_rundown with a non-zero wbuf_dqd. The only
* exception is if we are recover/rollback in which case gds_rundown (from mur_close_files) is invoked
* BEFORE secshr_db_clnup in mur_close_files.
* Note: It is NOT possible for online rollback to reach here with wbuf_dqd being non-zero. This is because
* the moment we apply the first PBLK, we stop all interrupts and hence can never be interrupted in
* wcs_wtstart or wcs_get_space. Assert accordingly.
*/
assert(mupip_jnl_recover && !jgbl.onlnrlbk && !safe_mode);
if (!was_crit)
grab_crit(reg);
SET_TRACEABLE_VAR(cnl->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wcb_gds_rundown);
send_msg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_WCBLOCKED, 6, LEN_AND_LIT("wcb_gds_rundown"),
process_id, &csa->ti->curr_tn, DB_LEN_STR(reg));
csa->wbuf_dqd = 0;
wcs_recover(reg);
BG_TRACE_PRO_ANY(csa, lost_block_recovery);
if (!was_crit)
rel_crit(reg);
}
if (JNL_ENABLED(csd) && IS_GTCM_GNP_SERVER_IMAGE)
originator_prc_vec = NULL;
/* If we are the last writing user, then everything must be flushed */
if (we_are_last_writer)
{ /* Time to flush out all of our buffers */
assert(!safe_mode);
if (is_mm)
{
MM_DBFILEXT_REMAP_IF_NEEDED(csa, reg);
cnl->remove_shm = TRUE;
}
if (cnl->wc_blocked && jgbl.onlnrlbk)
{ /* if the last update done by online rollback was not committed in the normal code-path but was
* completed by secshr_db_clnup, wc_blocked will be set to TRUE. But, since online rollback never
* invokes grab_crit (since csa->hold_onto_crit is set to TRUE), wcs_recover is never invoked. This
* could result in the last update never getting flushed to the disk and if online rollback happened
* to be the last writer then the shared memory will be flushed and removed and the last update will
* be lost. So, force wcs_recover if we find ourselves in such a situation. But, wc_blocked is
* possible only if phase1 or phase2 errors are induced using white box test cases
*/
assert(WB_COMMIT_ERR_ENABLED);
wcs_recover(reg);
}
/* Note WCSFLU_SYNC_EPOCH ensures the epoch is synced to the journal and indirectly
* also ensures that the db is fsynced. We don't want to use it in the calls to
* wcs_flu() from t_end() and tp_tend() since we can defer it to out-of-crit there.
* In this case, since we are running down, we don't have any such option.
*/
cnl->remove_shm = wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
/* Since we_are_last_writer, we should be guaranteed that wcs_flu() did not change csd, (in
* case of MM for potential file extension), even if it did a grab_crit(). Therefore, make
* sure that's true.
*/
assert(csd == csa->hdr);
assert(0 == memcmp(csd->label, GDS_LABEL, GDS_LABEL_SZ - 1));
} else if (((canceled_flush_timer && (0 > cnl->wcs_timers)) || canceled_dbsync_timer) && !inst_is_frozen)
{ /* canceled pending db or jnl flush timers - flush database and journal buffers to disk */
if (!was_crit)
grab_crit(reg);
/* we need to sync the epoch as the fact that there is no active pending flush timer implies
* there will be noone else who will flush the dirty buffers and EPOCH to disk in a timely fashion
*/
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
if (!was_crit)
rel_crit(reg);
assert((dba_mm == cs_data->acc_meth) || (csd == cs_data));
csd = cs_data; /* In case this is MM and wcs_flu() remapped an extended database, reset csd */
}
/* Do rundown journal processing after buffer flushes since they require jnl to be open */
if (JNL_ENABLED(csd))
{ /* the following tp_change_reg() is not needed due to the assert csa == cs_addrs at the beginning
* of gds_rundown(), but just to be safe. To be removed by 2002!! --- nars -- 2001/04/25.
*/
tp_change_reg(); /* call this because jnl_ensure_open checks cs_addrs rather than gv_cur_region */
jpc = csa->jnl;
jbp = jpc->jnl_buff;
if (jbp->fsync_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->fsync_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (jbp->io_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->io_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if ((((NOJNL != jpc->channel) && !JNL_FILE_SWITCHED(jpc))
|| we_are_last_writer && (0 != cnl->jnl_file.u.inode)) && ok_to_write_pfin)
{ /* We need to close the journal file cleanly if we have the latest generation journal file open
* or if we are the last writer and the journal file is open in shared memory (not necessarily
* by ourselves e.g. the only process that opened the journal got shot abnormally)
* Note: we should not infer anything from the shared memory value of cnl->jnl_file.u.inode
* if we are not the last writer as it can be concurrently updated.
*/
if (!was_crit)
grab_crit(reg);
if (JNL_ENABLED(csd))
{
SET_GBL_JREC_TIME; /* jnl_ensure_open/jnl_put_jrt_pini/pfin/jnl_file_close all need it */
/* 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 we_are_last_writer, we would have already done a wcs_flu() which would
* have written an epoch record and we are guaranteed no further updates
* since we are the last writer. So, just close the journal.
* If the freeaddr == post_epoch_freeaddr, wcs_flu may have skipped writing
* a pini, so allow for that.
*/
assert(!jbp->before_images || is_mm
|| !we_are_last_writer || (0 != jpc->pini_addr) || jgbl.mur_extract
|| (jpc->jnl_buff->freeaddr == jpc->jnl_buff->post_epoch_freeaddr));
/* If we haven't written a pini, let jnl_file_close write the pini/pfin. */
if (!jgbl.mur_extract && (0 != jpc->pini_addr))
jnl_put_jrt_pfin(csa);
/* If not the last writer and no pending flush timer left, do jnl flush now */
if (!we_are_last_writer && (0 > cnl->wcs_timers))
{
if (SS_NORMAL == (jnl_status = jnl_flush(reg)))
{
assert(jbp->freeaddr == jbp->dskaddr);
jnl_fsync(reg, jbp->dskaddr);
assert(jbp->fsync_dskaddr == jbp->dskaddr);
} else
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_JNLFLUSH, 2,
JNL_LEN_STR(csd), ERR_TEXT, 2,
RTS_ERROR_TEXT("Error with journal flush in gds_rundown"),
jnl_status);
assert(NOJNL == jpc->channel);/* jnl file lost has been triggered */
/* In this routine, all code that follows from here on does not
* assume anything about the journaling characteristics of this
* database so it is safe to continue execution even though
* journaling got closed in the middle.
*/
}
}
jnl_file_close(reg, we_are_last_writer, FALSE);
} else
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd),
DB_LEN_STR(reg));
}
if (!was_crit)
rel_crit(reg);
}
}
if (we_are_last_writer) /* Flush the fileheader last and harden the file to disk */
{
if (!was_crit)
grab_crit(reg); /* To satisfy crit requirement in fileheader_sync() */
memset(csd->machine_name, 0, MAX_MCNAMELEN); /* clear the machine_name field */
if (!have_standalone_access && we_are_last_user)
{ /* mupip_exit_handler will do this after mur_close_file */
csd->semid = INVALID_SEMID;
csd->shmid = INVALID_SHMID;
csd->gt_sem_ctime.ctime = 0;
csd->gt_shm_ctime.ctime = 0;
}
fileheader_sync(reg);
if (!was_crit)
rel_crit(reg);
if (!is_mm)
{
GTM_DB_FSYNC(csa, udi->fd, rc); /* Sync it all */
if (-1 == rc)
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
} else
{ /* Now do final MM file sync before exit */
assert(csa->ti->total_blks == csa->total_blks);
#ifdef _AIX
GTM_DB_FSYNC(csa, udi->fd, rc);
if (-1 == rc)
#else
if (-1 == MSYNC((caddr_t)csa->db_addrs[0], (caddr_t)csa->db_addrs[1]))
#endif
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
}
} else if (unsafe_last_writer && !cnl->lastwriterbypas_msg_issued)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_LASTWRITERBYPAS, 2, DB_LEN_STR(reg));
cnl->lastwriterbypas_msg_issued = TRUE;
}
} /* end if (!reg->read_only && !cnl->donotflush_dbjnl) */
/* We had canceled all db timers at start of rundown. In case as part of rundown (wcs_flu above), we had started
* any timers, cancel them BEFORE setting reg->open to FALSE (assert in wcs_clean_dbsync relies on this).
*/
CANCEL_DB_TIMERS(reg, csa, canceled_flush_timer, canceled_dbsync_timer);
if (reg->read_only && we_are_last_user && !have_standalone_access && cnl->remove_shm)
{ /* mupip_exit_handler will do this after mur_close_file */
db_ipcs.semid = INVALID_SEMID;
db_ipcs.shmid = INVALID_SHMID;
db_ipcs.gt_sem_ctime = 0;
db_ipcs.gt_shm_ctime = 0;
db_ipcs.fn_len = reg->dyn.addr->fname_len;
memcpy(db_ipcs.fn, reg->dyn.addr->fname, reg->dyn.addr->fname_len);
db_ipcs.fn[reg->dyn.addr->fname_len] = 0;
/* request gtmsecshr to flush. read_only cannot flush itself */
WAIT_FOR_REPL_INST_UNFREEZE_SAFE(csa);
if (!csa->read_only_fs)
{
secshrstat = send_mesg2gtmsecshr(FLUSH_DB_IPCS_INFO, 0, (char *)NULL, 0);
if (0 != secshrstat)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gtmsecshr failed to update database file header"));
}
}
/* Done with file now, close it */
CLOSEFILE_RESET(udi->fd, rc); /* resets "udi->fd" to FD_INVALID */
if (-1 == rc)
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during file close"), errno);
}
/* Unmap storage if mm mode but only the part that is not the fileheader (so shows up in dumps) */
# if !defined(_AIX)
if (is_mm && (NULL != csa->db_addrs[0]))
{
assert(csa->db_addrs[1] > csa->db_addrs[0]);
munmap_len = (sm_long_t)(csa->db_addrs[1] - csa->db_addrs[0]);
if (0 < munmap_len)
munmap((caddr_t)(csa->db_addrs[0]), (size_t)(munmap_len));
}
# endif
/* Detach our shared memory while still under lock so reference counts will be correct for the next process to run down
* this region. In the process also get the remove_shm status from node_local before detaching.
* If cnl->donotflush_dbjnl is TRUE, it means we can safely remove shared memory without compromising data
* integrity as a reissue of recover will restore the database to a consistent state.
*/
remove_shm = !vermismatch && (cnl->remove_shm || cnl->donotflush_dbjnl);
/* We are done with online rollback on this region. Indicate to other processes by setting the onln_rlbk_pid to 0.
* Do it before releasing crit (t_end relies on this ordering when accessing cnl->onln_rlbk_pid).
*/
if (jgbl.onlnrlbk)
cnl->onln_rlbk_pid = 0;
rel_crit(reg); /* Since we are about to detach from the shared memory, release crit and reset onln_rlbk_pid */
/* If we had skipped flushing journal and database buffers due to a concurrent online rollback, increment the counter
* indicating that in the shared memory so that online rollback can report the # of such processes when it shuts down.
* The same thing is done for both FTOK and access control semaphores when there are too many MUMPS processes.
*/
if (safe_mode) /* indicates flushing was skipped */
{
if (bypassed_access)
cnl->dbrndwn_access_skip++; /* Access semaphore can be bypassed during online rollback */
if (bypassed_ftok)
cnl->dbrndwn_ftok_skip++;
}
if (jgbl.onlnrlbk)
csa->hold_onto_crit = FALSE;
GTM_WHITE_BOX_TEST(WBTEST_HOLD_SEM_BYPASS, cnl->wbox_test_seq_num, 0);
status = shmdt((caddr_t)cnl);
csa->nl = NULL; /* dereferencing nl after detach is not right, so we set it to NULL so that we can test before dereference*/
/* Note that although csa->nl is NULL, we use CSA_ARG(csa) below (not CSA_ARG(NULL)) to be consistent with similar
* usages before csa->nl became NULL. The "is_anticipatory_freeze_needed" function (which is in turn called by the
* CHECK_IF_FREEZE_ON_ERROR_NEEDED macro) does a check of csa->nl before dereferencing shared memory contents so
* we are safe passing "csa".
*/
if (-1 == status)
send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg), ERR_TEXT, 2,
LEN_AND_LIT("Error during shmdt"), errno);
REMOVE_CSA_FROM_CSADDRSLIST(csa); /* remove "csa" from list of open regions (cs_addrs_list) */
reg->open = FALSE;
/* If file is still not in good shape, die here and now before we get rid of our storage */
assertpro(0 == csa->wbuf_dqd);
ipc_deleted = FALSE;
/* If we are the very last user, remove shared storage id and the semaphores */
if (we_are_last_user)
{ /* remove shared storage, only if last writer to rundown did a successful wcs_flu() */
assert(!vermismatch);
if (remove_shm)
{
ipc_deleted = TRUE;
if (0 != shm_rmid(udi->shmid))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove shared memory"));
/* Note that we no longer have a new shared memory. Currently only used/usable for standalone rollback. */
udi->new_shm = FALSE;
/* mupip recover/rollback don't release the semaphore here, but do it later in db_ipcs_reset (invoked from
* mur_close_files())
*/
if (!have_standalone_access)
{
if (0 != sem_rmid(udi->semid))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove semaphore"));
udi->new_sem = FALSE; /* Note that we no longer have a new semaphore */
udi->grabbed_access_sem = FALSE;
udi->counter_acc_incremented = FALSE;
}
} else if (is_src_server || is_updproc)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
{
assert(!have_standalone_access || jgbl.onlnrlbk || safe_mode);
if (!jgbl.onlnrlbk && !have_standalone_access)
{ /* If we were writing, get rid of our writer access count semaphore */
if (!reg->read_only)
{
if (!access_counter_halted)
{
save_errno = do_semop(udi->semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO);
if (0 != save_errno)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown access control semaphore decrement"),
CALLFROM, save_errno);
}
udi->counter_acc_incremented = FALSE;
}
assert(safe_mode || !bypassed_access);
/* Now remove the rundown lock */
if (!bypassed_access)
{
if (0 != (save_errno = do_semop(udi->semid, DB_CONTROL_SEM, -1, SEM_UNDO)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown access control semaphore release"),
CALLFROM, save_errno);
udi->grabbed_access_sem = FALSE;
}
} /* else access control semaphore will be released in db_ipcs_reset */
}
if (!have_standalone_access)
{
if (bypassed_ftok)
{
if (!ftok_counter_halted)
if (0 != (save_errno = do_semop(udi->ftok_semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
} else if (!ftok_sem_release(reg, !ftok_counter_halted, FALSE))
{
FTOK_TRACE(csa, csa->ti->curr_tn, ftok_ops_release, process_id);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
}
udi->grabbed_ftok_sem = FALSE;
udi->counter_ftok_incremented = FALSE;
}
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
if (!ipc_deleted)
{
GET_CUR_TIME(time_str);
if (is_src_server)
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Source server"), REG_LEN_STR(reg));
if (is_updproc)
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Update process"), REG_LEN_STR(reg));
if (mupip_jnl_recover && (!jgbl.onlnrlbk || !we_are_last_user))
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
send_msg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
}
}
REVERT;
return EXIT_NRM;
}
void gds_rundown(void)
{
bool is_mm, we_are_last_user, we_are_last_writer;
boolean_t ipc_deleted, remove_shm, cancelled_timer, cancelled_dbsync_timer, vermismatch;
now_t now; /* for GET_CUR_TIME macro */
char *time_ptr, time_str[CTIME_BEFORE_NL + 2]; /* for GET_CUR_TIME macro */
gd_region *reg;
int save_errno, status;
int4 semval, ftok_semval, sopcnt, ftok_sopcnt;
short crash_count;
sm_long_t munmap_len;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
struct shmid_ds shm_buf;
struct sembuf sop[2], ftok_sop[2];
uint4 jnl_status;
unix_db_info *udi;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
error_def(ERR_CRITSEMFAIL);
error_def(ERR_DBCCERR);
error_def(ERR_DBFILERR);
error_def(ERR_DBRNDWNWRN);
error_def(ERR_ERRCALL);
error_def(ERR_GBLOFLOW);
error_def(ERR_GTMASSERT);
error_def(ERR_IPCNOTDEL);
error_def(ERR_JNLFLUSH);
error_def(ERR_RNDWNSEMFAIL);
error_def(ERR_TEXT);
error_def(ERR_WCBLOCKED);
forced_exit = FALSE; /* Okay, we're dying already -- let rel_crit live in peace now.
* If coming through a DAL, not necessarily dying. what to do then? -- nars -- 8/15/2001
*/
grabbed_access_sem = FALSE;
jnl_status = 0;
reg = gv_cur_region; /* Local copy */
/*
* early out for cluster regions
* to avoid tripping the assert below.
* Note:
* This early out is consistent with VMS. It has been
* noted that all of the gtcm assignments
* to gv_cur_region should use the TP_CHANGE_REG
* macro. This would also avoid the assert problem
* and should be done eventually.
*/
if (dba_cm == reg->dyn.addr->acc_meth)
return;
udi = FILE_INFO(reg);
csa = &udi->s_addrs;
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
if ((reg->open) && (dba_usr == csd->acc_meth))
{
change_reg();
gvusr_rundown();
return;
}
ESTABLISH(gds_rundown_ch);
if (!reg->open) /* Not open, no point to rundown */
{
if (reg->opening) /* Died partway open, kill rest of way */
{
rel_crit(reg);
mutex_cleanup(reg);
/* revist this to handle MM properly SMW 98/12/16
if (NULL != csa->nl)
{
status = shmdt((caddr_t)csa->nl);
if (-1 == status)
send_msg(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during shmdt"), errno);
}
*/
shmdt((caddr_t)csa->nl);
csa->nl = NULL;
}
REVERT;
return;
}
switch(csd->acc_meth)
{ /* Pass mm and bg through */
case dba_bg:
is_mm = FALSE;
break;
case dba_mm:
is_mm = TRUE;
break;
case dba_usr:
assert(FALSE);
default:
REVERT;
return;
}
/* Cancel any pending flush timer for this region by this task */
CANCEL_DB_TIMERS(reg, cancelled_timer, cancelled_dbsync_timer);
we_are_last_user = FALSE;
if (!csa->persistent_freeze)
region_freeze(reg, FALSE, FALSE, FALSE);
assert(!csa->read_lock);
rel_crit(reg); /* get locks to known state */
mutex_cleanup(reg);
/*
* We need to guarantee that none else access database file header when semid/shmid fields are reset.
* We already have created ftok semaphore in db_init or, mu_rndwn_file and did not remove it.
* So just lock it. We do it in blocking mode.
*/
if (!ftok_sem_lock(reg, FALSE, FALSE))
rts_error(VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
/*
* For mupip_jnl_recover we already have database access control semaphore.
* We do not release it. We release it from mur_close_files.
*/
if (!mupip_jnl_recover)
{
sop[0].sem_num = 0; sop[0].sem_op = 0; /* Wait for 0 */
sop[1].sem_num = 0; sop[1].sem_op = 1; /* Lock */
sopcnt = 2;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO | IPC_NOWAIT; /* Don't wait the first time thru */
SEMOP(udi->semid, sop, sopcnt, status);
if (-1 == status) /* We couldn't get it in one shot -- see if we already have it */
{
save_errno = errno;
/* see comment about Linux specific difference in behaviour of semctl() with GETPID in gds_rundown_ch() */
if (semctl(udi->semid, 0, GETPID) == process_id)
{
send_msg(VARLSTCNT(5) MAKE_MSG_INFO(ERR_CRITSEMFAIL), 2,
DB_LEN_STR(reg),
ERR_RNDWNSEMFAIL);
REVERT;
return; /* Already in rundown for this region */
}
if (EAGAIN != save_errno)
{
assert(FALSE);
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown first semop/semctl"), save_errno);
}
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO; /* Try again - blocking this time */
SEMOP(udi->semid, sop, 2, status);
if (-1 == status) /* We couldn't get it at all.. */
rts_error(VARLSTCNT(5) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), errno);
}
}
grabbed_access_sem = TRUE;
/*
* We now have the dbinit/rundown lock, so we are alone in this code for this region
* and nobody else can attach.
* See if we are all alone in accessing this database shared memory.
*/
assert(csa->ref_cnt); /* decrement private ref_cnt before shared ref_cnt decrement. */
csa->ref_cnt--; /* Currently journaling logic in gds_rundown() in VMS relies on this order to detect last writer */
assert(!csa->ref_cnt);
--csa->nl->ref_cnt;
if (memcmp(csa->nl->now_running, gtm_release_name, gtm_release_name_len + 1))
{ /* VERMISMATCH condition. Possible only if DSE */
assert(dse_running);
vermismatch = TRUE;
} else
vermismatch = FALSE;
if (-1 == shmctl(udi->shmid, IPC_STAT, &shm_buf))
{
save_errno = errno;
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown shmctl"), save_errno);
} else
we_are_last_user = (1 == shm_buf.shm_nattch) && !vermismatch;
assert(!mupip_jnl_recover || we_are_last_user); /* recover => one user */
if (-1 == (semval = semctl(udi->semid, 1, GETVAL)))
rts_error(VARLSTCNT(5) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), errno);
we_are_last_writer = (1 == semval) && (FALSE == reg->read_only) && !vermismatch;/* There's one writer left and I am it */
assert(!(mupip_jnl_recover && !reg->read_only) || we_are_last_writer); /* recover + R/W region => one writer */
if (-1 == (ftok_semval = semctl(udi->ftok_semid, 1, GETVAL)))
rts_error(VARLSTCNT(5) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), errno);
/* If csa->nl->donotflush_dbjnl is set, it means mupip recover/rollback was interrupted and therefore we should
* not flush shared memory contents to disk as they might be in an inconsistent state.
* In this case, we will go ahead and remove shared memory (without flushing the contents) in this routine.
* A reissue of the recover/rollback command will restore the database to a consistent state.
* Otherwise, if we have write access to this region, let us perform a few writing tasks.
*/
if (csa->nl->donotflush_dbjnl)
csa->wbuf_dqd = 0; /* ignore csa->wbuf_dqd status as we do not care about the cache contents */
else if (!reg->read_only && !vermismatch)
{ /* If we had an orphaned block and were interrupted, set wc_blocked so we can invoke wcs_recover */
if (csa->wbuf_dqd)
{
grab_crit(reg);
SET_TRACEABLE_VAR(csd->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wcb_gds_rundown);
send_msg(VARLSTCNT(8) ERR_WCBLOCKED, 6, LEN_AND_LIT("wcb_gds_rundown"),
process_id, &csa->ti->curr_tn, DB_LEN_STR(reg));
csa->wbuf_dqd = 0;
wcs_recover(reg);
if (is_mm)
{
assert(FALSE);
csd = csa->hdr;
}
BG_TRACE_PRO_ANY(csa, lost_block_recovery);
rel_crit(reg);
}
if (JNL_ENABLED(csd) && (GTCM_GNP_SERVER_IMAGE == image_type))
originator_prc_vec = NULL;
/* If we are the last writing user, then everything must be flushed */
if (we_are_last_writer)
{ /* Time to flush out all of our buffers */
if (is_mm)
{
if (csa->total_blks != csa->ti->total_blks) /* do remap if file had been extended */
{
grab_crit(reg);
wcs_mm_recover(reg);
csd = csa->hdr;
rel_crit(reg);
}
csa->nl->remove_shm = TRUE;
}
/* Note WCSFLU_SYNC_EPOCH ensures the epoch is synced to the journal and indirectly
* also ensures that the db is fsynced. We don't want to use it in the calls to
* wcs_flu() from t_end() and tp_tend() since we can defer it to out-of-crit there.
* In this case, since we are running down, we don't have any such option.
*/
csa->nl->remove_shm = wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
/* Since we_are_last_writer, we should be guaranteed that wcs_flu() did not change csd, (in
* case of MM for potential file extension), even if it did a grab_crit(). Therefore, make
* sure that's true.
*/
assert(csd == csa->hdr);
assert(0 == memcmp(csd->label, GDS_LABEL, GDS_LABEL_SZ - 1));
csd->trans_hist.header_open_tn = csd->trans_hist.curr_tn;
} else if ((cancelled_timer && (0 > csa->nl->wcs_timers)) || cancelled_dbsync_timer)
{ /* cancelled pending db or jnl flush timers - flush database and journal buffers to disk */
grab_crit(reg);
/* we need to sync the epoch as the fact that there is no active pending flush timer implies
* there will be noone else who will flush the dirty buffers and EPOCH to disk in a timely fashion
*/
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
rel_crit(reg);
assert((dba_mm == cs_data->acc_meth) || (csd == cs_data));
csd = cs_data; /* In case this is MM and wcs_flu() remapped an extended database, reset csd */
}
/* Do rundown journal processing after buffer flushes since they require jnl to be open */
if (JNL_ENABLED(csd))
{ /* the following tp_change_reg() is not needed due to the assert csa == cs_addrs at the beginning
* of gds_rundown(), but just to be safe. To be removed by 2002!! --- nars -- 2001/04/25.
*/
tp_change_reg(); /* call this because jnl_ensure_open checks cs_addrs rather than gv_cur_region */
jpc = csa->jnl;
jbp = jpc->jnl_buff;
if (jbp->fsync_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->fsync_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (jbp->io_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->io_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (((NOJNL != jpc->channel) && !JNL_FILE_SWITCHED(jpc))
|| we_are_last_writer && (0 != csa->nl->jnl_file.u.inode))
{ /* We need to close the journal file cleanly if we have the latest generation journal file open
* or if we are the last writer and the journal file is open in shared memory (not necessarily
* by ourselves e.g. the only process that opened the journal got shot abnormally)
* Note: we should not infer anything from the shared memory value of csa->nl->jnl_file.u.inode
* if we are not the last writer as it can be concurrently updated.
*/
grab_crit(reg);
if (JNL_ENABLED(csd))
{
SET_GBL_JREC_TIME; /* jnl_ensure_open/jnl_put_jrt_pini/pfin/jnl_file_close all need it */
/* 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 we_are_last_writer, we would have already done a wcs_flu() which would
* have written an epoch record and we are guaranteed no further updates
* since we are the last writer. So, just close the journal.
* Although we assert pini_addr should be non-zero for last_writer, we
* play it safe in PRO and write a PINI record if not written already.
*/
assert(!jbp->before_images || is_mm
|| !we_are_last_writer || 0 != jpc->pini_addr);
if (we_are_last_writer && 0 == jpc->pini_addr)
jnl_put_jrt_pini(csa);
if (0 != jpc->pini_addr)
jnl_put_jrt_pfin(csa);
/* If not the last writer and no pending flush timer left, do jnl flush now */
if (!we_are_last_writer && (0 > csa->nl->wcs_timers))
{
if (SS_NORMAL == (jnl_status = jnl_flush(reg)))
{
assert(jbp->freeaddr == jbp->dskaddr);
jnl_fsync(reg, jbp->dskaddr);
assert(jbp->fsync_dskaddr == jbp->dskaddr);
} else
{
send_msg(VARLSTCNT(9) ERR_JNLFLUSH, 2, JNL_LEN_STR(csd),
ERR_TEXT, 2,
RTS_ERROR_TEXT("Error with journal flush in gds_rundown"),
jnl_status);
assert(NOJNL == jpc->channel);/* jnl file lost has been triggered */
/* In this routine, all code that follows from here on does not
* assume anything about the journaling characteristics of this
* database so it is safe to continue execution even though
* journaling got closed in the middle.
*/
}
}
jnl_file_close(reg, we_are_last_writer, FALSE);
} else
send_msg(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(reg));
}
rel_crit(reg);
}
}
if (we_are_last_writer) /* Flush the fileheader last and harden the file to disk */
{
grab_crit(reg); /* To satisfy crit requirement in fileheader_sync() */
memset(csd->machine_name, 0, MAX_MCNAMELEN); /* clear the machine_name field */
if (!mupip_jnl_recover && we_are_last_user)
{ /* mupip_jnl_recover will do this after mur_close_file */
csd->semid = INVALID_SEMID;
csd->shmid = INVALID_SHMID;
csd->gt_sem_ctime.ctime = 0;
csd->gt_shm_ctime.ctime = 0;
}
fileheader_sync(reg);
rel_crit(reg);
if (FALSE == is_mm)
{
if (-1 == fsync(udi->fd)) /* Sync it all */
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
} else
{ /* Now do final MM file sync before exit */
#if !defined(TARGETED_MSYNC) && !defined(NO_MSYNC)
if (-1 == fsync(udi->fd)) /* Sync it all */
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
#else
if (-1 == msync((caddr_t)csa->db_addrs[0], (size_t)(csa->db_addrs[1] - csa->db_addrs[0]), MS_SYNC))
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file msync at close"), errno);
}
#endif
}
}
} /* end if (!reg->read_only && !csa->nl->donotflush_dbjnl) */
if (reg->read_only && we_are_last_user && !mupip_jnl_recover)
{ /* mupip_jnl_recover will do this after mur_close_file */
db_ipcs.semid = INVALID_SEMID;
db_ipcs.shmid = INVALID_SHMID;
db_ipcs.gt_sem_ctime = 0;
db_ipcs.gt_shm_ctime = 0;
db_ipcs.fn_len = reg->dyn.addr->fname_len;
memcpy(db_ipcs.fn, reg->dyn.addr->fname, reg->dyn.addr->fname_len);
db_ipcs.fn[reg->dyn.addr->fname_len] = 0;
/* request gtmsecshr to flush. read_only cannot flush itself */
if (0 != send_mesg2gtmsecshr(FLUSH_DB_IPCS_INFO, 0, (char *)NULL, 0))
rts_error(VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gtmsecshr failed to update database file header"));
}
/* Done with file now, close it */
if (-1 == close(udi->fd))
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during file close"), errno);
}
/* Unmap storage if mm mode but only the part that is not the fileheader (so shows up in dumps) */
if (is_mm)
{
munmap_len = (sm_long_t)((csa->db_addrs[1] - csa->db_addrs[0]) - ROUND_UP(SIZEOF_FILE_HDR(csa->hdr),
MSYNC_ADDR_INCS));
if (munmap_len > 0)
{
munmap((caddr_t)(csa->db_addrs[0] + ROUND_UP(SIZEOF_FILE_HDR(csa->hdr), MSYNC_ADDR_INCS)),
(size_t)(munmap_len));
#ifdef DEBUG_DB64
rel_mmseg((caddr_t)csa->db_addrs[0]);
#endif
}
}
/* Detach our shared memory while still under lock so reference counts will be
* correct for the next process to run down this region.
* In the process also get the remove_shm status from node_local before detaching.
* If csa->nl->donotflush_dbjnl is TRUE, it means we can safely remove shared memory without compromising data
* integrity as a reissue of recover will restore the database to a consistent state.
*/
remove_shm = !vermismatch && (csa->nl->remove_shm || csa->nl->donotflush_dbjnl);
status = shmdt((caddr_t)csa->nl);
csa->nl = NULL; /* dereferencing nl after detach is not right, so we set it to NULL so that we can test before dereference*/
if (-1 == status)
send_msg(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg), ERR_TEXT, 2, LEN_AND_LIT("Error during shmdt"), errno);
reg->open = FALSE;
/* If file is still not in good shape, die here and now before we get rid of our storage */
if (csa->wbuf_dqd)
GTMASSERT;
ipc_deleted = FALSE;
/* If we are the very last user, remove shared storage id and the semaphores */
if (we_are_last_user)
{ /* remove shared storage, only if last writer to rundown did a successful wcs_flu() */
assert(!vermismatch);
if (remove_shm)
{
ipc_deleted = TRUE;
if (0 != shm_rmid(udi->shmid))
rts_error(VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove shared memory"));
} else if (is_src_server || is_updproc)
{
gtm_putmsg(VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
send_msg(VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
send_msg(VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
/*
* Don't release semaphore in case of mupip recover/rollback; since it has standalone access.
* It will release the semaphore in mur_close_files.
*/
if (!mupip_jnl_recover)
{
if (0 != sem_rmid(udi->semid))
rts_error(VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove semaphore"));
grabbed_access_sem = FALSE;
}
} else
{
assert(!mupip_jnl_recover);
/* If we were writing, get rid of our writer access count semaphore */
if (!reg->read_only)
if (0 != (save_errno = do_semop(udi->semid, 1, -1, SEM_UNDO)))
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown write semaphore release"), save_errno);
/* Now remove the rundown lock */
if (0 != (save_errno = do_semop(udi->semid, 0, -1, SEM_UNDO)))
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown rundown semaphore release"), save_errno);
grabbed_access_sem = FALSE;
}
if (!ftok_sem_release(reg, !mupip_jnl_recover, FALSE))
rts_error(VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
if (!ipc_deleted)
{
GET_CUR_TIME;
if (is_src_server)
gtm_putmsg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Source server"), REG_LEN_STR(reg));
if (is_updproc)
gtm_putmsg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Update process"), REG_LEN_STR(reg));
if (mupip_jnl_recover)
{
gtm_putmsg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
send_msg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
}
}
REVERT;
}
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 jnl_file_extend(jnl_private_control *jpc, uint4 total_jnl_rec_size)
{
file_control *fc;
boolean_t need_extend;
jnl_buffer_ptr_t jb;
jnl_create_info jnl_info;
jnl_file_header *header;
unsigned char hdr_buff[REAL_JNL_HDR_LEN + MAX_IO_BLOCK_SIZE];
uint4 new_alq;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
char prev_jnl_fn[JNL_NAME_SIZE];
uint4 jnl_status = 0, status;
int new_blocks, warn_blocks, result;
gtm_uint64_t avail_blocks;
uint4 aligned_tot_jrec_size, count;
uint4 jnl_fs_block_size, read_write_size;
DCL_THREADGBL_ACCESS;
switch(jpc->region->dyn.addr->acc_meth)
{
case dba_mm:
case dba_bg:
csa = &FILE_INFO(jpc->region)->s_addrs;
break;
default:
GTMASSERT;
}
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
assert(csa->now_crit || (csd->clustered && (CCST_CLOSED == csa->nl->ccp_state)));
assert(&FILE_INFO(jpc->region)->s_addrs == csa);
assert(csa->jnl_state == csd->jnl_state);
assertpro(JNL_ENABLED(csa) && (NOJNL != jpc->channel) && (!JNL_FILE_SWITCHED(jpc)));
/* crit and messing with the journal file - how could it have vanished? */
if (!csd->jnl_deq || (csd->jnl_alq + csd->jnl_deq > csd->autoswitchlimit))
{
assert(DIVIDE_ROUND_UP(total_jnl_rec_size, DISK_BLOCK_SIZE) <= csd->jnl_alq);
assert(csd->jnl_alq == csd->autoswitchlimit);
new_blocks = csd->jnl_alq;
} else
/* May cause extension of csd->jnl_deq * n blocks where n > 0 */
new_blocks = ROUND_UP(DIVIDE_ROUND_UP(total_jnl_rec_size, DISK_BLOCK_SIZE), csd->jnl_deq);
jpc->status = SS_NORMAL;
jb = jpc->jnl_buff;
assert(0 <= new_blocks);
DEBUG_ONLY(count = 0);
for (need_extend = (0 != new_blocks); need_extend; )
{
DEBUG_ONLY(count++);
/* usually we will do the loop just once where we do the file extension.
* rarely we might need to do an autoswitch instead after which again rarely
* we might need to do an extension on the new journal to fit in the transaction's journal requirements.
* therefore we should do this loop a maximum of twice. hence the assert below.
*/
assert(count <= 2);
need_extend = FALSE;
if (SS_NORMAL == (status = disk_block_available(jpc->channel, &avail_blocks, TRUE)))
{
warn_blocks = (csd->jnl_alq + csd->jnl_deq > csd->autoswitchlimit)
? ((csd->jnl_deq > csd->autoswitchlimit) ? csd->jnl_deq : csd->autoswitchlimit)
: new_blocks;
if ((warn_blocks * EXTEND_WARNING_FACTOR) > avail_blocks)
{
if (new_blocks > avail_blocks)
{ /* If we cannot satisfy the request, it is an error, unless the anticipatory freeze
* scheme is in effect in which case, we will assume space is available even if
* it is not and go ahead with writes to the disk. If the writes fail with ENOSPC
* we will freeze the instance and wait for space to become available and keep
* retrying the writes. Therefore, we make the NOSPACEEXT a warning in this case.
*/
SETUP_THREADGBL_ACCESS;
if (!ANTICIPATORY_FREEZE_ENABLED(csa))
{
send_msg(VARLSTCNT(6) ERR_NOSPACEEXT, 4,
JNL_LEN_STR(csd), new_blocks, avail_blocks);
new_blocks = 0;
jpc->status = SS_NORMAL;
break;
} else
send_msg(VARLSTCNT(6) MAKE_MSG_WARNING(ERR_NOSPACEEXT), 4,
JNL_LEN_STR(csd), new_blocks, avail_blocks);
} else
send_msg(VARLSTCNT(5) ERR_DSKSPACEFLOW, 3, JNL_LEN_STR(csd), (avail_blocks - warn_blocks));
}
} else
send_msg(VARLSTCNT(5) ERR_JNLFILEXTERR, 2, JNL_LEN_STR(csd), status);
new_alq = jb->filesize + new_blocks;
/* ensure current journal file size is well within autoswitchlimit --> design constraint */
assert(csd->autoswitchlimit >= jb->filesize);
if (csd->autoswitchlimit < (jb->filesize + (EXTEND_WARNING_FACTOR * new_blocks))) /* close to max */
send_msg(VARLSTCNT(5) ERR_JNLSPACELOW, 3, JNL_LEN_STR(csd), csd->autoswitchlimit - jb->filesize);
if (csd->autoswitchlimit < new_alq)
{ /* Reached max, need to autoswitch */
/* Ensure new journal file can hold the entire current transaction's journal record requirements */
assert(csd->autoswitchlimit >= MAX_REQD_JNL_FILE_SIZE(total_jnl_rec_size));
memset(&jnl_info, 0, SIZEOF(jnl_info));
jnl_info.prev_jnl = &prev_jnl_fn[0];
set_jnl_info(gv_cur_region, &jnl_info);
assert(JNL_ENABLED(csa) && (NOJNL != jpc->channel) && !(JNL_FILE_SWITCHED(jpc)));
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{ /* flush the cache and jnl-buffer-contents to current journal file before
* switching to a new journal. Set a global variable in_jnl_file_autoswitch
* so jnl_write can know not to do the padding check. But because this is a global
* variable, we also need to make sure it is reset in case of errors during the
* autoswitch (or else calls to jnl_write after we are out of the autoswitch logic
* will continue to incorrectly not do the padding check. Hence a condition handler.
*/
assert(!in_jnl_file_autoswitch);
in_jnl_file_autoswitch = TRUE;
/* Also make sure time is not changed. This way if "jnl_write" as part of writing a
* journal record invokes jnl_file_extend, when the autoswitch is done and writing
* of the parent jnl_write resumes, we want it to continue with the same timestamp
* and not have to reset its time (non-trivial task) to reflect any changes since then.
*/
assert(!jgbl.save_dont_reset_gbl_jrec_time);
jgbl.save_dont_reset_gbl_jrec_time = jgbl.dont_reset_gbl_jrec_time;
jgbl.dont_reset_gbl_jrec_time = TRUE;
/* Establish a condition handler so we reset a few global variables that have
* temporarily been modified in case of errors inside wcs_flu/jnl_file_close.
*/
ESTABLISH_RET(jnl_file_autoswitch_ch, EXIT_ERR);
/* It is possible we still have not written a PINI record in this journal file
* (e.g. mupip extend saw the need to do jnl_file_extend inside jnl_write while
* trying to write a PINI record). Write a PINI record in that case before closing
* the journal file that way the EOF record will have a non-zero pini_addr.
*/
if (0 == jpc->pini_addr)
jnl_put_jrt_pini(csa);
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SPEEDUP_NOBEFORE);
jnl_file_close(gv_cur_region, TRUE, TRUE);
REVERT;
in_jnl_file_autoswitch = FALSE;
jgbl.dont_reset_gbl_jrec_time = jgbl.save_dont_reset_gbl_jrec_time;
DEBUG_ONLY(jgbl.save_dont_reset_gbl_jrec_time = FALSE);
assert((dba_mm == cs_data->acc_meth) || (csd == cs_data));
csd = cs_data; /* In MM, wcs_flu() can remap an extended DB, so reset csd to be sure */
} else
{
if (SS_NORMAL != jpc->status)
rts_error(VARLSTCNT(7) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region),
jpc->status);
else
rts_error(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region));
}
assert(!jgbl.forw_phase_recovery || (NULL != jgbl.mur_pini_addr_reset_fnptr));
assert(jgbl.forw_phase_recovery || (NULL == jgbl.mur_pini_addr_reset_fnptr));
if (NULL != jgbl.mur_pini_addr_reset_fnptr)
(*jgbl.mur_pini_addr_reset_fnptr)(csa);
assert(!jnl_info.no_rename);
assert(!jnl_info.no_prev_link);
if (EXIT_NRM == cre_jnl_file(&jnl_info))
{
assert(0 == memcmp(csd->jnl_file_name, jnl_info.jnl, jnl_info.jnl_len));
assert(csd->jnl_file_name[jnl_info.jnl_len] == '\0');
assert(csd->jnl_file_len == jnl_info.jnl_len);
assert(csd->jnl_buffer_size == jnl_info.buffer);
assert(csd->jnl_alq == jnl_info.alloc);
assert(csd->jnl_deq == jnl_info.extend);
assert(csd->jnl_before_image == jnl_info.before_images);
csd->jnl_checksum = jnl_info.checksum;
csd->jnl_eovtn = csd->trans_hist.curr_tn;
send_msg(VARLSTCNT(4) ERR_NEWJNLFILECREAT, 2, JNL_LEN_STR(csd));
fc = gv_cur_region->dyn.addr->file_cntl;
fc->op = FC_WRITE;
fc->op_buff = (sm_uc_ptr_t)csd;
fc->op_len = SGMNT_HDR_LEN;
fc->op_pos = 1;
status = dbfilop(fc);
if (SS_NORMAL != status)
send_msg(VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), status);
assert(JNL_ENABLED(csa));
/* call jnl_ensure_open instead of jnl_file_open to make sure jpc->pini_addr is set to 0 */
jnl_status = jnl_ensure_open(); /* sets jpc->status */
if (0 != jnl_status)
{
if (jpc->status)
rts_error(VARLSTCNT(7) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region),
jpc->status);
else
rts_error(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region));
}
assert(jb->filesize == csd->jnl_alq);
if (csd->jnl_alq + csd->jnl_deq <= csd->autoswitchlimit)
{
aligned_tot_jrec_size = ALIGNED_ROUND_UP(MAX_REQD_JNL_FILE_SIZE(total_jnl_rec_size),
csd->jnl_alq, csd->jnl_deq);
if (aligned_tot_jrec_size > csd->jnl_alq)
{ /* need to extend more than initial allocation in the new journal file
* to accommodate the current transaction.
*/
new_blocks = aligned_tot_jrec_size - csd->jnl_alq;
assert(new_blocks);
assert(0 == new_blocks % csd->jnl_deq);
need_extend = TRUE;
}
}
} else
{
send_msg(VARLSTCNT(4) ERR_JNLNOCREATE, 2, JNL_LEN_STR(csd));
jpc->status = ERR_JNLNOCREATE;
new_blocks = -1;
}
} else
{
assert(!need_extend); /* ensure we won't go through the for loop again */
/* Virtually extend currently used journal file */
jnl_fs_block_size = jb->fs_block_size;
header = (jnl_file_header *)(ROUND_UP2((uintszofptr_t)hdr_buff, jnl_fs_block_size));
read_write_size = ROUND_UP2(REAL_JNL_HDR_LEN, jnl_fs_block_size);
assert((unsigned char *)header + read_write_size <= ARRAYTOP(hdr_buff));
DO_FILE_READ(jpc->channel, 0, header, read_write_size, jpc->status, jpc->status2);
if (SS_NORMAL != jpc->status)
{
assert(FALSE);
rts_error(VARLSTCNT(5) ERR_JNLRDERR, 2, JNL_LEN_STR(csd), jpc->status);
}
assert((header->virtual_size + new_blocks) == new_alq);
jb->filesize = new_alq; /* Actually this is virtual file size blocks */
header->virtual_size = new_alq;
JNL_DO_FILE_WRITE(csa, csd->jnl_file_name, jpc->channel, 0,
header, read_write_size, jpc->status, jpc->status2);
if (SS_NORMAL != jpc->status)
{
assert(FALSE);
rts_error(VARLSTCNT(5) ERR_JNLWRERR, 2, JNL_LEN_STR(csd), jpc->status);
}
}
if (0 >= new_blocks)
break;
}
if (0 < new_blocks)
{
INCR_GVSTATS_COUNTER(csa, csa->nl, n_jnl_extends, 1);
return EXIT_NRM;
}
jpc->status = ERR_JNLREADEOF;
jnl_file_lost(jpc, ERR_JNLEXTEND);
return EXIT_ERR;
}
uint4 mur_process_intrpt_recov()
{
jnl_ctl_list *jctl, *last_jctl;
reg_ctl_list *rctl, *rctl_top;
int rename_fn_len, save_name_len;
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;
#if defined(VMS)
io_status_block_disk iosb;
#endif
boolean_t jfh_changed;
for (rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++)
{
gv_cur_region = rctl->gd; /* wcs_flu requires this to be set */
cs_addrs = rctl->csa;
csd = cs_data = rctl->csd; /* 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)
{
if (max_autoswitchlimit < jctl->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 = jctl->jfh->autoswitchlimit;
max_jnl_alq = jctl->jfh->jnl_alq;
max_jnl_deq = jctl->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(!jctl->jfh->prev_recov_blks_to_upgrd_adjust || !jctl->jfh->recover_interrupted);
assert(!jctl->jfh->prev_recov_blks_to_upgrd_adjust || jctl->jfh->prev_recov_end_of_data);
rctl->blks_to_upgrd_adjust += jctl->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);
}
/* now that rctl->blks_to_upgrd_adjust is completely computed, use that to increment filehdr blks_to_upgrd. */
csd->blks_to_upgrd += rctl->blks_to_upgrd_adjust;
if (csd->blks_to_upgrd)
csd->fully_upgraded = FALSE;
jctl = rctl->jctl_turn_around;
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(rctl->trnarnd_total_blks <= csd->trans_hist.total_blks);
csd->trans_hist.free_blocks = rctl->trnarnd_free_blocks + (csd->trans_hist.total_blks - rctl->trnarnd_total_blks)
- DIVIDE_ROUND_UP(csd->trans_hist.total_blks, BLKS_PER_LMAP)
+ DIVIDE_ROUND_UP(rctl->trnarnd_total_blks, BLKS_PER_LMAP);
assert((freeblks = mur_blocks_free(rctl)) == csd->trans_hist.free_blocks);
if (dba_bg == csd->acc_meth)
/* This is taken from bt_refresh() */
((th_rec *)((uchar_ptr_t)cs_addrs->th_base + cs_addrs->th_base->tnque.fl))->tn = jctl->turn_around_tn - 1;
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_FSYNC_DB);
csd->turn_around_point = FALSE;
/* 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;
}
for (rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++)
{
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, &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);
jctl->jfh->turn_around_offset = jctl->turn_around_offset; /* save progress in file header for */
jctl->jfh->turn_around_time = jctl->turn_around_time; /* possible re-issue of recover */
jfh_changed = 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.
*/
if (NULL != jctl->next_gen)
{
jctl->jfh->next_jnl_file_name_length = jctl->next_gen->jnl_fn_len;
memcpy(jctl->jfh->next_jnl_file_name, jctl->next_gen->jnl_fn, jctl->next_gen->jnl_fn_len);
jfh_changed = TRUE;
} else
assert(0 == jctl->jfh->next_jnl_file_name_length); /* null link from latest generation */
if (jctl->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->jfh_recov_interrupted can fail */
jctl->jfh->turn_around_offset = 0;
jctl->jfh->turn_around_time = 0;
jfh_changed = TRUE;
}
if (jfh_changed)
{
DO_FILE_WRITE(jctl->channel, 0, jctl->jfh, REAL_JNL_HDR_LEN, jctl->status, jctl->status2);
if (SS_NORMAL != jctl->status)
{
assert(FALSE);
if (SS_NORMAL == jctl->status2)
gtm_putmsg(VARLSTCNT(5) ERR_JNLWRERR, 2, jctl->jnl_fn_len,
jctl->jnl_fn, jctl->status);
else
gtm_putmsg(VARLSTCNT1(6) ERR_JNLWRERR, 2, jctl->jnl_fn_len,
jctl->jnl_fn, jctl->status, PUT_SYS_ERRNO(jctl->status2));
return jctl->status;
}
UNIX_ONLY(
GTM_FSYNC(jctl->channel, jctl->status);
if (-1 == jctl->status)
{
jctl->status2 = errno;
assert(FALSE);
gtm_putmsg(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(VARLSTCNT(4) ERR_JNLNOCREATE, 2, jnl_info.jnl_len, jnl_info.jnl);
return jnl_info.status;
}
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.
*/
}
uint4 mur_process_intrpt_recov()
{
jnl_ctl_list *jctl;
reg_ctl_list *rctl, *rctl_top;
int rename_fn_len, save_name_len;
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;
#if defined(VMS)
io_status_block_disk iosb;
#endif
boolean_t jfh_changed, first_time;
error_def (ERR_PREMATEOF); /* for DO_FILE_WRITE */
error_def (ERR_JNLCREATERR);
error_def (ERR_JNLWRERR);
error_def (ERR_JNLFSYNCERR);
error_def (ERR_TEXT);
for (mur_regno = 0, rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++, mur_regno++)
{
cs_addrs = rctl->csa;
cs_data = rctl->csd;
assert(rctl->csd == rctl->csa->hdr);
mur_jctl = jctl = rctl->jctl_turn_around;
assert(rctl->csd == rctl->csa->hdr);
mur_master_map();
cs_data->trans_hist.header_open_tn =
cs_data->trans_hist.early_tn =
cs_data->trans_hist.curr_tn = jctl->turn_around_tn;
cs_data->trans_hist.free_blocks = mur_blocks_free();
if (dba_bg == cs_data->acc_meth)
/* This is taken from bt_refresh() */
((th_rec *)((uchar_ptr_t)cs_addrs->th_base + cs_addrs->th_base->tnque.fl))->tn =
jctl->turn_around_tn - 1;
max_jnl_alq = max_jnl_deq = max_autoswitchlimit = 0;
for ( ; (NULL != jctl->next_gen); jctl = jctl->next_gen)
{
if (max_autoswitchlimit < jctl->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 = jctl->jfh->autoswitchlimit;
max_jnl_alq = jctl->jfh->jnl_alq;
max_jnl_deq = jctl->jfh->jnl_deq;
}
}
if (max_autoswitchlimit > jctl->jfh->autoswitchlimit)
{
cs_data->jnl_alq = max_jnl_alq;
cs_data->jnl_deq = max_jnl_deq;
cs_data->autoswitchlimit = max_autoswitchlimit;
} else
{
assert(cs_data->jnl_alq == jctl->jfh->jnl_alq);
assert(cs_data->jnl_deq == jctl->jfh->jnl_deq);
assert(cs_data->autoswitchlimit == jctl->jfh->autoswitchlimit);
}
gv_cur_region = rctl->gd;
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_FSYNC_DB);
}
for (mur_regno = 0, rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++, mur_regno++)
{
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, &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);
jctl->jfh->turn_around_offset = jctl->turn_around_offset; /* save progress in file header for */
jctl->jfh->turn_around_time = jctl->turn_around_time; /* possible re-issue of recover */
jfh_changed = TRUE;
DEBUG_ONLY(first_time = TRUE;)
for ( ; NULL != jctl; jctl = jctl->next_gen)
void gv_select(char *cli_buff, int n_len, boolean_t freeze, char opname[], glist *gl_head,
int *reg_max_rec, int *reg_max_key, int *reg_max_blk)
{
bool stashed = FALSE;
int num_quote, len;
char *ptr, *ptr1, *c;
mstr gmap[256], *gmap_ptr, gmap_beg, gmap_end;
mval val, curr_gbl_name;
glist *gl_tail, *gl_ptr;
htab_desc *ext_hash;
mname ht_mname;
ht_entry *h;
error_def(ERR_FREEZE);
error_def(ERR_DBRDONLY);
error_def(ERR_SELECTSYNTAX);
error_def(ERR_MUNOFINISH);
error_def(ERR_MUNOACTION);
memset(gmap, 0, sizeof(gmap));
for (ptr = cli_buff; *ptr; ptr = ptr1)
{
for (ptr1 = ptr; ; ptr1++)
{
if (',' == *ptr1)
{
len = ptr1 - ptr;
ptr1++;
break;
}
else if (!*ptr1)
{
len = ptr1 - ptr;
break;
}
}
gmap_beg.addr = ptr;
c = gmap_beg.addr + len - 1;
num_quote = 0;
while ('"' == *c)
{
len--;
c--;
num_quote++;
}
if (0 >= len)
{
gtm_putmsg(VARLSTCNT(4) ERR_SELECTSYNTAX, 2, LEN_AND_STR(opname));
mupip_exit(ERR_MUNOACTION);
}
c = gmap_beg.addr;
while (0 < num_quote)
{
if ('"' == *c)
{
c++;
len--;
}
else
{
gtm_putmsg(VARLSTCNT(4) ERR_SELECTSYNTAX, 2, LEN_AND_STR(opname));
mupip_exit(ERR_MUNOACTION);
}
num_quote--;
}
gmap_beg.addr = c;
if ('^' == *c)
{
gmap_beg.addr++;
len--;
}
gmap_beg.len = len;
c = mu_extr_ident(&gmap_beg);
len -= (c - gmap_beg.addr);
assert(len >= 0);
if (0 == len)
gmap_end = gmap_beg;
else if (gmap_beg.len == 1 && '*' == *c)
{
gmap_beg.addr = (char*)&percent_lit;
gmap_beg.len = sizeof(percent_lit);
gmap_end.addr = (char*)&tilde_lit;
gmap_end.len = sizeof(tilde_lit);
}
else if (1 == len && '*' == *c)
{
gmap_end = gmap_beg;
gmap_beg.len--;
*c = '~';
}
else if (':' != *c)
{
gtm_putmsg(VARLSTCNT(4) ERR_SELECTSYNTAX, 2, LEN_AND_STR(opname));
mupip_exit(ERR_MUNOACTION);
}
else
{
gmap_beg.len = c - gmap_beg.addr;
c++;
gmap_end.addr = c;
gmap_end.len = len - 1;
if ('^' == *c)
{
gmap_end.addr++;
gmap_end.len--;
}
c = mu_extr_ident(&gmap_end);
if (c - gmap_end.addr != gmap_end.len || mstrcmp(&gmap_beg, &gmap_end) > 0)
{
gtm_putmsg(VARLSTCNT(4) ERR_SELECTSYNTAX, 2, LEN_AND_STR(opname));
mupip_exit(ERR_MUNOACTION);
}
}
global_map(gmap, &gmap_beg, &gmap_end);
}
if (freeze)
{
ext_hash = (htab_desc *)malloc(sizeof(htab_desc));
ht_init(ext_hash, 0);
memset(&ht_mname.txt[0], 0, sizeof(mname));
}
gl_head->next = NULL;
gl_tail = gl_head;
*reg_max_rec = 0;
*reg_max_key = 0;
*reg_max_blk = 0;
for (gmap_ptr = gmap; gmap_ptr->addr ; gmap_ptr++)
{
curr_gbl_name.mvtype = MV_STR;
curr_gbl_name.str = *gmap_ptr++;
op_gvname(VARLSTCNT(1) &curr_gbl_name);
if (dba_cm == gv_cur_region->dyn.addr->acc_meth)
{ util_out_print("Can not select globals from region !AD across network",TRUE,gv_cur_region->rname_len,
gv_cur_region->rname);
mupip_exit(ERR_MUNOFINISH);
}
if (dba_bg != gv_cur_region->dyn.addr->acc_meth && dba_mm != gv_cur_region->dyn.addr->acc_meth)
{
assert(gv_cur_region->dyn.addr->acc_meth == dba_usr);
util_out_print("Can not select globals from non-GTC region !AD",TRUE,gv_cur_region->rname_len,
gv_cur_region->rname);
mupip_exit(ERR_MUNOFINISH);
}
op_gvdata(&val);
if (0 == val.m[1])
{
op_gvname(VARLSTCNT(1) &curr_gbl_name);
op_gvorder(&curr_gbl_name);
if (!curr_gbl_name.str.len)
break;
assert('^' == *curr_gbl_name.str.addr);
curr_gbl_name.str.addr++;
curr_gbl_name.str.len--;
}
for (;;)
{
if (mstrcmp(&curr_gbl_name.str, gmap_ptr) > 0)
break;
if (freeze)
{
memcpy(&ht_mname.txt[0], &gv_cur_region, sizeof(int4));
h = ht_put(ext_hash, &ht_mname, &stashed);
if (stashed)
{
if (cs_addrs->hdr->freeze)
{
gtm_putmsg(VARLSTCNT(4) ERR_FREEZE, 2, gv_cur_region->rname_len,
gv_cur_region->rname);
mupip_exit(ERR_MUNOFINISH);
}
/* Cannot proceed for read-only data files */
if (gv_cur_region->read_only)
{
util_out_print("Cannot freeze the database",TRUE);
gtm_putmsg(VARLSTCNT(4) ERR_DBRDONLY, 2,
DB_LEN_STR(gv_cur_region));
mupip_exit(ERR_MUNOFINISH);
}
while (FALSE == region_freeze(gv_cur_region, TRUE, FALSE))
hiber_start(1000);
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
}
}
assert(curr_gbl_name.str.len > 0);
gl_ptr = (glist*)malloc(sizeof(glist) - 1 + curr_gbl_name.str.len);
gl_ptr->name.mvtype = MV_STR;
gl_ptr->name.str.addr = (char*)gl_ptr->nbuf;
gl_ptr->name.str.len = curr_gbl_name.str.len;
memcpy(gl_ptr->nbuf, curr_gbl_name.str.addr, curr_gbl_name.str.len);
gl_ptr->next = 0;
gl_tail->next = gl_ptr;
gl_tail = gl_ptr;
if (*reg_max_rec < cs_data->max_rec_size) *reg_max_rec = cs_data->max_rec_size;
if (*reg_max_key < cs_data->max_key_size) *reg_max_key = cs_data->max_key_size;
if (*reg_max_blk < cs_data->blk_size) *reg_max_blk = cs_data->blk_size;
op_gvname(VARLSTCNT(1) &gl_tail->name);
op_gvorder(&curr_gbl_name);
if (0 == curr_gbl_name.str.len)
{
(gmap_ptr + 1)->addr = 0;
break;
}
assert('^' == *curr_gbl_name.str.addr);
curr_gbl_name.str.addr++;
curr_gbl_name.str.len--;
}
}
}
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.
*/
}
