void heartbeat_timer(void)
{
gd_addr *addr_ptr;
sgmnt_addrs *csa;
jnl_private_control *jpc;
gd_region *r_local, *r_top;
int rc;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
/* It will take heartbeat_counter about 1014 years to overflow. */
heartbeat_counter++;
DEBUG_ONLY(set_enospc_if_needed());
/* Check every 1 minute if we have an older generation journal file open. If so, close it.
* The only exceptions are
* a) The source server can have older generations open and they should not be closed.
* b) If we are in the process of switching to a new journal file while we get interrupted
* by the heartbeat timer, we should not close the older generation journal file
* as it will anyways be closed by the mainline code. But identifying that we are in
* the midst of a journal file switch is tricky so we check if the process is in
* crit for this region and if so we skip the close this time and wait for the next heartbeat.
*/
if ((INTRPT_OK_TO_INTERRUPT == intrpt_ok_state) && !is_src_server
&& (0 == heartbeat_counter % NUM_HEARTBEATS_FOR_OLDERJNL_CHECK))
{
for (addr_ptr = get_next_gdr(NULL); addr_ptr; addr_ptr = get_next_gdr(addr_ptr))
{
for (r_local = addr_ptr->regions, r_top = r_local + addr_ptr->n_regions; r_local < r_top; r_local++)
{
if (!r_local->open || r_local->was_open)
continue;
if ((dba_bg != r_local->dyn.addr->acc_meth) && (dba_mm != r_local->dyn.addr->acc_meth))
continue;
csa = &FILE_INFO(r_local)->s_addrs;
if (csa->now_crit)
continue;
jpc = csa->jnl;
if ((NULL != jpc) && (NOJNL != jpc->channel) && JNL_FILE_SWITCHED(jpc))
{ /* The journal file we have as open is not the latest generation journal file. Close it */
/* Assert that we never have an active write on a previous generation journal file. */
assert(process_id != jpc->jnl_buff->io_in_prog_latch.u.parts.latch_pid);
JNL_FD_CLOSE(jpc->channel, rc); /* sets jpc->channel to NOJNL */
jpc->pini_addr = 0;
}
}
}
}
start_timer((TID)heartbeat_timer, HEARTBEAT_INTERVAL, heartbeat_timer, 0, NULL);
}
void set_enospc_if_needed()
{
gd_addr *addr_ptr;
char enospc_enable_list[MAX_REGIONS];
boolean_t ok_to_interrupt, is_time_to_act;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
if (TREF(gtm_test_fake_enospc) && is_jnlpool_creator && ANTICIPATORY_FREEZE_AVAILABLE)
{
ok_to_interrupt = (INTRPT_OK_TO_INTERRUPT == intrpt_ok_state) && (0 == gtmMallocDepth);
is_time_to_act = (next_heartbeat_counter == heartbeat_counter);
if (syslog_deferred && ok_to_interrupt)
{
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(3) ERR_FAKENOSPCLEARED, 1, (heartbeat_counter - syslog_deferred));
syslog_deferred = 0;
}
if (!is_time_to_act || syslog_deferred || (!ok_to_interrupt && !IS_REPL_INST_FROZEN))
{
/* We have to skip this because we have just fallen into deferred zone or we are currently in it */
if (is_time_to_act)
next_heartbeat_counter++; /* Try again in the next heartbeat */
return;
}
assert(0 == syslog_deferred);
srand(time(NULL));
addr_ptr = get_next_gdr(NULL);
if (NULL == addr_ptr) /* Ensure that there is a global directory to operate on. */
return;
assert(NULL == get_next_gdr(addr_ptr));
/* Randomly simulate ENOSPC or free space. NO more than 50 regions are allowed to avoid unnecessary
* malloc/frees in debug-only code
*/
assert(MAX_REGIONS >= addr_ptr->n_regions);
if (!IS_REPL_INST_FROZEN)
{ /* We are in an UNFROZEN state, and about to be FROZEN due to ENOSPC */
choose_random_reg_list(enospc_enable_list, addr_ptr->n_regions);
next_heartbeat_counter = heartbeat_counter + ENOSPC_FROZEN_DURATION;
} else
{ /* We are in a FROZEN state, and about to be UNFROZEN due to free space */
memset(enospc_enable_list, 0, MAX_REGIONS);
next_heartbeat_counter = heartbeat_counter + ENOSPC_UNFROZEN_DURATION;
if (!ok_to_interrupt)
syslog_deferred = heartbeat_counter;
}
set_enospc_flags(addr_ptr, enospc_enable_list, ok_to_interrupt);
}
}
void util_exit_handler()
{
int stat;
gd_region *r_top, *reg;
sgmnt_addrs *csa;
gd_addr *addr_ptr;
if (exit_handler_active) /* Don't recurse if exit handler exited */
return;
exit_handler_active = TRUE;
SET_PROCESS_EXITING_TRUE; /* set this BEFORE canceling timers as wcs_phase2_commit_wait relies on this */
if (IS_DSE_IMAGE)
{ /* Need to clear csa->hold_onto_crit in case it was set */
for (addr_ptr = get_next_gdr(NULL); addr_ptr; addr_ptr = get_next_gdr(addr_ptr))
{
for (reg = addr_ptr->regions, r_top = reg + addr_ptr->n_regions; reg < r_top; reg++)
{
if (reg->open && !reg->was_open)
{
csa = &FILE_INFO(reg)->s_addrs;
csa->hold_onto_crit = FALSE; /* need to do this before the rel_crit */
if (csa->now_crit)
rel_crit(reg);
}
}
}
}
CANCEL_TIMERS; /* Cancel all unsafe timers - No unpleasant surprises */
secshr_db_clnup(NORMAL_TERMINATION);
assert(!dollar_tlevel); /* MUPIP and GT.M are the only ones which can run TP and they have their own exit handlers.
* So no need to run op_trollback here like mupip_exit_handler and gtm_exit_handler. */
gv_rundown();
print_exit_stats();
util_out_close();
GTMCRYPT_CLOSE;
if (need_core && !created_core)
DUMP_CORE;
}
void gv_rundown(void)
{
gd_region *r_top, *r_save, *r_local;
gd_addr *addr_ptr;
sgm_info *si;
#ifdef VMS
vms_gds_info *gds_info;
#endif
error_def(ERR_TEXT);
r_save = gv_cur_region; /* Save for possible core dump */
gvcmy_rundown();
ENABLE_AST
if (pool_init)
rel_lock(jnlpool.jnlpool_dummy_reg);
for (addr_ptr = get_next_gdr(NULL); addr_ptr; addr_ptr = get_next_gdr(addr_ptr))
{
for (r_local = addr_ptr->regions, r_top = r_local + addr_ptr->n_regions; r_local < r_top; r_local++)
{
if (r_local->open && !r_local->was_open && dba_cm != r_local->dyn.addr->acc_meth)
{ /* Rundown has already occurred for GT.CM client regions through gvcmy_rundown() above.
* Hence the (dba_cm != ...) check in the if above. Note that for GT.CM client regions,
* region->open is TRUE although cs_addrs is NULL.
*/
gv_cur_region = r_local;
tp_change_reg();
gds_rundown();
/* Now that gds_rundown is done, free up the memory associated with the region.
* Ideally the following memory freeing code should go to gds_rundown, but
* GT.CM calls gds_rundown() and we want to reuse memory for GT.CM.
*/
if (NULL != cs_addrs)
{
if (NULL != cs_addrs->dir_tree)
FREE_CSA_DIR_TREE(cs_addrs);
if (cs_addrs->sgm_info_ptr)
{
si = cs_addrs->sgm_info_ptr;
/* It is possible we got interrupted before initializing all fields of "si"
* completely so account for NULL values while freeing/releasing those fields.
*/
assert((si->tp_csa == cs_addrs) || (NULL == si->tp_csa));
if (si->jnl_tail)
{
CAREFUL_FREEUP_BUDDY_LIST(si->format_buff_list);
CAREFUL_FREEUP_BUDDY_LIST(si->jnl_list);
}
CAREFUL_FREEUP_BUDDY_LIST(si->recompute_list);
CAREFUL_FREEUP_BUDDY_LIST(si->new_buff_list);
CAREFUL_FREEUP_BUDDY_LIST(si->tlvl_info_list);
CAREFUL_FREEUP_BUDDY_LIST(si->tlvl_cw_set_list);
CAREFUL_FREEUP_BUDDY_LIST(si->cw_set_list);
if (NULL != si->blks_in_use)
{
free_hashtab_int4(si->blks_in_use);
free(si->blks_in_use);
si->blks_in_use = NULL;
}
if (si->cr_array_size)
{
assert(NULL != si->cr_array);
if (NULL != si->cr_array)
free(si->cr_array);
}
if (NULL != si->first_tp_hist)
free(si->first_tp_hist);
free(si);
}
if (cs_addrs->jnl)
{
assert(&FILE_INFO(cs_addrs->jnl->region)->s_addrs == cs_addrs);
if (cs_addrs->jnl->jnllsb)
{
UNIX_ONLY(assert(FALSE));
free(cs_addrs->jnl->jnllsb);
}
free(cs_addrs->jnl);
}
GTMCRYPT_ONLY(
if (cs_addrs->encrypted_blk_contents)
free(cs_addrs->encrypted_blk_contents);
)
}
assert(gv_cur_region->dyn.addr->file_cntl->file_info);
VMS_ONLY(
gds_info = (vms_gds_info *)gv_cur_region->dyn.addr->file_cntl->file_info;
if (gds_info->xabpro)
free(gds_info->xabpro);
if (gds_info->xabfhc)
free(gds_info->xabfhc);
if (gds_info->nam)
{
free(gds_info->nam->nam$l_esa);
free(gds_info->nam);
}
if (gds_info->fab)
free(gds_info->fab);
)
free(gv_cur_region->dyn.addr->file_cntl->file_info);
free(gv_cur_region->dyn.addr->file_cntl);
}
r_local->open = r_local->was_open = FALSE;
}
void preemptive_db_clnup(int preemptive_severe)
{
sgmnt_addrs *csa;
sgm_info *si;
gd_region *r_top, *reg;
gd_addr *addr_ptr;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
/* Clear "inctn_opcode" global variable now that any in-progress transaction is aborted at this point.
* Not doing so would cause future calls to "t_end" to get confused and skip writing logical jnl recs
* and instead incorrectly write an INCTN record (GTM-8425).
*/
if (bml_save_dollar_tlevel)
{
assert(!dollar_tlevel);
dollar_tlevel = bml_save_dollar_tlevel;
bml_save_dollar_tlevel = 0;
}
assert(!dollar_tlevel || (inctn_invalid_op == inctn_opcode) || (inctn_bmp_mark_free_gtm == inctn_opcode));
assert(dollar_tlevel || update_trans || (inctn_invalid_op == inctn_opcode));
inctn_opcode = inctn_invalid_op;
if (!dollar_tlevel && update_trans)
{ /* It's possible we hit an error in the middle of an update, at which point we have
* a valid clue and non-NULL cse. However, this causes problems for subsequent
* transactions (see comment in t_begin). In particular we could end up pinning buffers
* unnecessarily. So clear the cse of any histories that may have been active during the update.
*/
CLEAR_CSE(gv_target);
if ((NULL != gv_target) && (NULL != gv_target->gd_csa))
{
CLEAR_CSE(gv_target->gd_csa->dir_tree);
GTMTRIG_ONLY(CLEAR_CSE(gv_target->gd_csa->hasht_tree));
}
/* Resetting this is necessary to avoid blowing an assert in t_begin that it is 0 at the start of a transaction. */
update_trans = 0;
}
if (INVALID_GV_TARGET != reset_gv_target)
{
if (SUCCESS != preemptive_severe && INFO != preemptive_severe)
{
/* We know of a few cases in Unix where gv_target and gv_currkey could be out of sync at this point.
* a) If we are inside trigger code which in turn does an update that does
* reads of ^#t global and ends up in a restart. This restart would
* in turn do a rts_error_csa(TPRETRY) which would invoke mdb_condition_handler
* that would in turn invoke preemptive_db_clnup which invokes this macro.
* In this tp restart case though, it is ok for gv_target and gv_currkey
* to be out of sync because they are going to be reset by tp_clean_up anyways.
* So skip the dbg-only in-sync check.
* b) If we are in gvtr_init reading the ^#t global and detect an error (e.g. TRIGINVCHSET)
* gv_target after the reset would be pointing to a regular global whereas gv_currkey
* would be pointing to ^#t. It is ok to be out-of-sync since in this case, we expect
* mdb_condition_handler to be calling us. That has code to reset gv_currkey (and
* cs_addrs/cs_data etc.) to reflect gv_target (i.e. get them back in sync).
* Therefore in Unix we pass SKIP_GVT_GVKEY_CHECK to skip the gvtarget/gvcurrkey out-of-sync check
* in RESET_GV_TARGET. In VMS we pass DO_GVT_GVKEY_CHECK as we dont yet know of an out-of-sync situation.
*/
RESET_GV_TARGET(UNIX_ONLY(SKIP_GVT_GVKEY_CHECK) VMS_ONLY(DO_GVT_GVKEY_CHECK));
}
}
need_kip_incr = FALSE; /* in case we got an error in t_end (e.g. GBLOFLOW), dont want this global variable to get
* carried over to the next non-TP transaction that this process does (e.g. inside an error trap).
*/
TREF(expand_prev_key) = FALSE; /* reset global (in case it is TRUE) so it does not get carried over to future operations */
if (dollar_tlevel)
{
for (si = first_sgm_info; si != NULL; si = si->next_sgm_info)
{
if (NULL != si->kip_csa)
{
csa = si->tp_csa;
assert(si->tp_csa == si->kip_csa);
PROBE_DECR_KIP(csa->hdr, csa, si->kip_csa);
}
}
} else if (NULL != kip_csa && (NULL != kip_csa->hdr) && (NULL != kip_csa->nl))
PROBE_DECR_KIP(kip_csa->hdr, kip_csa, kip_csa);
if (IS_DSE_IMAGE)
{ /* Release crit on any region that was obtained for the current erroring DSE operation.
* Take care NOT to release crits obtained by a previous CRIT -SEIZE command.
*/
for (addr_ptr = get_next_gdr(NULL); addr_ptr; addr_ptr = get_next_gdr(addr_ptr))
{
for (reg = addr_ptr->regions, r_top = reg + addr_ptr->n_regions; reg < r_top; reg++)
{
if (reg->open && !reg->was_open)
{
csa = &FILE_INFO(reg)->s_addrs;
assert(csa->hold_onto_crit || !csa->dse_crit_seize_done);
assert(!csa->hold_onto_crit || csa->now_crit);
if (csa->now_crit && (!csa->hold_onto_crit || !csa->dse_crit_seize_done))
{
rel_crit(reg);
csa->hold_onto_crit = FALSE;
t_abort(reg, csa); /* cancel mini-transaction if any in progress */
}
}
}
}
}
}
void gv_rundown(void)
{
gd_region *r_top, *r_save, *r_local;
gd_addr *addr_ptr;
sgm_info *si;
int4 rundown_status = EXIT_NRM; /* if gds_rundown went smoothly */
# ifdef VMS
vms_gds_info *gds_info;
# elif UNIX
unix_db_info *udi;
# endif
#if defined(DEBUG) && defined(UNIX)
sgmnt_addrs *csa;
# endif
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
r_save = gv_cur_region; /* Save for possible core dump */
gvcmy_rundown();
ENABLE_AST
if (pool_init)
rel_lock(jnlpool.jnlpool_dummy_reg);
for (addr_ptr = get_next_gdr(NULL); addr_ptr; addr_ptr = get_next_gdr(addr_ptr))
{
for (r_local = addr_ptr->regions, r_top = r_local + addr_ptr->n_regions; r_local < r_top; r_local++)
{
if (r_local->open && !r_local->was_open && dba_cm != r_local->dyn.addr->acc_meth)
{ /* Rundown has already occurred for GT.CM client regions through gvcmy_rundown() above.
* Hence the (dba_cm != ...) check in the if above. Note that for GT.CM client regions,
* region->open is TRUE although cs_addrs is NULL.
*/
# if defined(DEBUG) && defined(UNIX)
if (is_jnlpool_creator && ANTICIPATORY_FREEZE_AVAILABLE && TREF(gtm_test_fake_enospc))
{ /* Clear ENOSPC faking now that we are running down */
csa = REG2CSA(r_local);
if (csa->nl->fake_db_enospc || csa->nl->fake_jnl_enospc)
{
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_TEXT, 2, DB_LEN_STR(r_local), ERR_TEXT,
2, LEN_AND_LIT("Resetting fake_db_enospc and fake_jnl_enospc"));
csa->nl->fake_db_enospc = FALSE;
csa->nl->fake_jnl_enospc = FALSE;
}
}
# endif
gv_cur_region = r_local;
tp_change_reg();
UNIX_ONLY(rundown_status |=) gds_rundown();
/* Now that gds_rundown is done, free up the memory associated with the region.
* Ideally the following memory freeing code should go to gds_rundown, but
* GT.CM calls gds_rundown() and we want to reuse memory for GT.CM.
*/
if (NULL != cs_addrs)
{
if (NULL != cs_addrs->dir_tree)
FREE_CSA_DIR_TREE(cs_addrs);
if (cs_addrs->sgm_info_ptr)
{
si = cs_addrs->sgm_info_ptr;
/* It is possible we got interrupted before initializing all fields of "si"
* completely so account for NULL values while freeing/releasing those fields.
*/
assert((si->tp_csa == cs_addrs) || (NULL == si->tp_csa));
if (si->jnl_tail)
{
CAREFUL_FREEUP_BUDDY_LIST(si->format_buff_list);
CAREFUL_FREEUP_BUDDY_LIST(si->jnl_list);
}
CAREFUL_FREEUP_BUDDY_LIST(si->recompute_list);
CAREFUL_FREEUP_BUDDY_LIST(si->new_buff_list);
CAREFUL_FREEUP_BUDDY_LIST(si->tlvl_info_list);
CAREFUL_FREEUP_BUDDY_LIST(si->tlvl_cw_set_list);
CAREFUL_FREEUP_BUDDY_LIST(si->cw_set_list);
if (NULL != si->blks_in_use)
{
free_hashtab_int4(si->blks_in_use);
free(si->blks_in_use);
si->blks_in_use = NULL;
}
if (si->cr_array_size)
{
assert(NULL != si->cr_array);
if (NULL != si->cr_array)
free(si->cr_array);
}
if (NULL != si->first_tp_hist)
free(si->first_tp_hist);
free(si);
}
if (cs_addrs->jnl)
{
assert(&FILE_INFO(cs_addrs->jnl->region)->s_addrs == cs_addrs);
if (cs_addrs->jnl->jnllsb)
{
UNIX_ONLY(assert(FALSE));
free(cs_addrs->jnl->jnllsb);
}
free(cs_addrs->jnl);
}
GTMCRYPT_ONLY(
if (cs_addrs->encrypted_blk_contents)
free(cs_addrs->encrypted_blk_contents);
)
}
assert(gv_cur_region->dyn.addr->file_cntl->file_info);
VMS_ONLY(
gds_info = (vms_gds_info *)gv_cur_region->dyn.addr->file_cntl->file_info;
if (gds_info->xabpro)
free(gds_info->xabpro);
if (gds_info->xabfhc)
free(gds_info->xabfhc);
if (gds_info->nam)
{
free(gds_info->nam->nam$l_esa);
free(gds_info->nam);
}
if (gds_info->fab)
free(gds_info->fab);
)
free(gv_cur_region->dyn.addr->file_cntl->file_info);
free(gv_cur_region->dyn.addr->file_cntl);
}
r_local->open = r_local->was_open = FALSE;
}
}
/* Return number of regions (including jnlpool dummy region) if have or are aquiring crit or in_wtstart
* ** NOTE ** This routine is called from signal handlers and is thus called asynchronously.
* If CRIT_IN_COMMIT bit is set, we check if in middle of commit (PHASE1 inside crit or PHASE2 outside crit) on some region.
* If CRIT_RELEASE bit is set, we release crit on region(s) that:
* 1) we hold crit on (neither CRIT_IN_COMMIT NOR CRIT_TRANS_NO_REG is specified)
* 2) are part of the current transactions except those regions that are marked as being valid
* to have crit in by virtue of their crit_check_cycle value is the same as crit_deadlock_check_cycle.
* Note: CRIT_RELEASE implies CRIT_ALL_REGIONS
* If CRIT_ALL_REGIONS bit is set, go through the entire list of regions
*/
uint4 have_crit(uint4 crit_state)
{
gd_region *r_top, *r_local;
gd_addr *addr_ptr;
sgmnt_addrs *csa;
uint4 crit_reg_cnt = 0;
/* in order to proper release the necessary regions, CRIT_RELEASE implies going through all the regions */
if (crit_state & CRIT_RELEASE)
{
UNIX_ONLY(assert(!jgbl.onlnrlbk)); /* should not request crit to be released if online rollback */
crit_state |= CRIT_ALL_REGIONS;
}
if (0 != crit_count)
{
crit_reg_cnt++;
if (0 == (crit_state & CRIT_ALL_REGIONS))
return crit_reg_cnt;
}
for (addr_ptr = get_next_gdr(NULL); addr_ptr; addr_ptr = get_next_gdr(addr_ptr))
{
for (r_local = addr_ptr->regions, r_top = r_local + addr_ptr->n_regions; r_local < r_top; r_local++)
{
if (r_local->open && !r_local->was_open)
{
csa = &FILE_INFO(r_local)->s_addrs;
if (NULL != csa)
{
if (csa->now_crit)
{
crit_reg_cnt++;
/* It is possible that if DSE has done a CRIT REMOVE and stolen our crit, it
* could be given to someone else which would cause this test to fail. The
* current thinking is that the state DSE put this process is no longer viable
* and it should die at the earliest opportunity, there being no way to know if
* that is what happened anyway.
*/
if (csa->nl->in_crit != process_id)
GTMASSERT;
/* If we are releasing (all) regions with critical section or if special
* TP case, release if the cycle number doesn't match meaning this is a
* region we should not hold crit in (even if it is part of tp_reg_list).
*/
if ((0 != (crit_state & CRIT_RELEASE)) &&
(0 == (crit_state & CRIT_NOT_TRANS_REG) ||
crit_deadlock_check_cycle != csa->crit_check_cycle))
{
assert(FALSE);
assert(!csa->hold_onto_crit);
rel_crit(r_local);
send_msg(VARLSTCNT(8) ERR_MUTEXRELEASED, 6,
process_id, process_id, DB_LEN_STR(r_local),
dollar_tlevel, t_tries);
}
if (0 == (crit_state & CRIT_ALL_REGIONS))
return crit_reg_cnt;
}
/* In Commit-crit is defined as the time since when early_tn is 1 + curr_tn upto when
* t_commit_crit is set to FALSE. Note that the first check should be done only if we
* hold crit as otherwise we could see inconsistent values.
*/
if ((crit_state & CRIT_IN_COMMIT)
&& (csa->now_crit && (csa->ti->early_tn != csa->ti->curr_tn) || csa->t_commit_crit))
{
crit_reg_cnt++;
if (0 == (crit_state & CRIT_ALL_REGIONS))
return crit_reg_cnt;
}
if ((crit_state & CRIT_IN_WTSTART) && csa->in_wtstart)
{
crit_reg_cnt++;
if (0 == (crit_state & CRIT_ALL_REGIONS))
return crit_reg_cnt;
}
}
}
}
}
if (NULL != jnlpool.jnlpool_ctl)
{
csa = &FILE_INFO(jnlpool.jnlpool_dummy_reg)->s_addrs;
if (NULL != csa && csa->now_crit)
{
crit_reg_cnt++;
if (0 != (crit_state & CRIT_RELEASE))
{
assert(!csa->hold_onto_crit);
rel_lock(jnlpool.jnlpool_dummy_reg);
}
}
}
return crit_reg_cnt;
}
