int gtmsource_end1(boolean_t auto_shutdown)
{
int exit_status;
seq_num log_seqno, log_seqno1, diff_seqno;
int fclose_res;
unsigned char seq_num_str[32], *seq_num_ptr;
#ifdef VMS
int4 status;
#endif
gtmsource_ctl_close();
rel_lock(jnlpool.jnlpool_dummy_reg);
UNIX_ONLY(mutex_cleanup(jnlpool.jnlpool_dummy_reg);)
void gtmsource_onln_rlbk_clnup()
{
gtmsource_local_ptr_t gtmsource_local;
boolean_t was_crit;
sgmnt_addrs *repl_csa;
gtmsource_local = jnlpool.gtmsource_local;
repl_csa = &FILE_INFO(jnlpool.jnlpool_dummy_reg)->s_addrs;
was_crit = repl_csa->now_crit;
assert(!repl_csa->hold_onto_crit);
assert(was_crit || (process_id == gtmsource_local->gtmsource_srv_latch.u.parts.latch_pid)
|| (0 != have_crit(CRIT_HAVE_ANY_REG)));
/* Reset source server context to indicate a fresh connection that is about to take place */
assert(NULL != gtmsource_local);
if (NULL != gtmsource_local)
{
/* If ROLLBACK has not taken the instance past the source server's read_jnl_seqno, then the source server should
* just continue from where it currently is and start sending the journal records from that point onwards. But, this
* is non-trivial. The reason is because, when the source server detected the online rollback, it could be in the
* READ_POOL state. But, since the instance has been rolled back, the journal pool cannot be relied upon in its
* entirety. To illustrate this -- consider that the journal pool contains the data from 1-100 and the source server
* is currently sending sequence number 30 and is reading from the pool. Assume an online rollback happens that
* takes the instance from sequence number 100 to sequence number 80 and leaves the journal pool write_addr and
* early_write_addr untouched. Now, lets say GT.M process comes in after this and does a few more updates. All of
* these updates will be written in the journal pool right after the "old-rolled-back" sequence number 100. If the
* source server continues to read from the pool, it will send the valid data until sequence number 80. After that,
* it will start sending the "old-rolled-back" sequence numbers 81-100 which is not right. To avoid this, rollback
* should set the write_addr and early_write_addr by searching in the journal pool for sequence number 81. This is
* currently not done, but is something that we can think about when it comes to optimization. Until then, force
* rollback to reset jnlpool's write_addr, write and early_write_addr to 0 and let source server be forced into
* READ_FILE mode.
*/
gtmsource_local->read_state = READ_FILE;
/* Set the state which gets bubbled up the call chain to gtmsource_process at which point we will close and
* re-establish the connection with the other end.
*/
gtmsource_local->gtmsource_state = gtmsource_state = GTMSOURCE_HANDLE_ONLN_RLBK;
if (!was_crit)
grab_lock(jnlpool.jnlpool_dummy_reg, TRUE, ASSERT_NO_ONLINE_ROLLBACK);
/* We have to let the read files logic know that until we have sent data "upto" the current journal sequence number
* at this point, we cannot rely on the journal pool. Indicate this through the gtmsource_save_read_jnl_seqno global
* variable
*/
gtmsource_save_read_jnl_seqno = jnlpool.jnlpool_ctl->jnl_seqno;
gtmsource_local->read = jnlpool.jnlpool_ctl->write;
gtmsource_local->read_addr = jnlpool.jnlpool_ctl->write_addr;
if (!was_crit)
rel_lock(jnlpool.jnlpool_dummy_reg);
}
return;
}
int gtmsource_secnd_update(boolean_t print_message)
{
if (grab_sem(SOURCE, SRC_SERV_OPTIONS_SEM) < 0)
{
util_out_print("Error grabbing jnlpool option write lock. Could not initiate change log", TRUE);
return(ABNORMAL_SHUTDOWN);
}
grab_lock(jnlpool.jnlpool_dummy_reg, ASSERT_NO_ONLINE_ROLLBACK);
jnlpool.jnlpool_ctl->upd_disabled = update_disable;
rel_lock(jnlpool.jnlpool_dummy_reg);
rel_sem(SOURCE, SRC_SERV_OPTIONS_SEM);
if (print_message)
util_out_print("Updates are now !AZ", TRUE, update_disable ? "disabled" : "enabled");
return(NORMAL_SHUTDOWN);
}
static void updproc_stop(boolean_t exit)
{
int4 status;
int fclose_res;
call_on_signal = NULL; /* Don't reenter on error */
if (pool_init)
{
rel_lock(jnlpool.jnlpool_dummy_reg);
/* nullify jnlpool_ctl before detaching from jnlpool since if it is the other way, we might be interrupted
* by the periodic timer routines and end up in jnl_write_epoch_rec() routine that dereferences jnlpool_ctl
* since it is non-NULL although it has been detached from and is no longer valid memory.
*/
jnlpool_ctl = NULL;
#ifdef UNIX
mutex_cleanup(jnlpool.jnlpool_dummy_reg);
SHMDT(jnlpool.jnlpool_ctl);
#elif defined(VMS)
if (SS$_NORMAL != (status = detach_shm(jnlpool.shm_range)))
repl_log(stderr, TRUE, TRUE, "Error detaching from jnlpool : %s\n", REPL_STR_ERROR);
if (SS$_NORMAL != (status = signoff_from_gsec(jnlpool.shm_lockid)))
repl_log(stderr, TRUE, TRUE, "Error dequeueing lock on jnlpool global section : %s\n", REPL_STR_ERROR);
#else
#error Unsupported Platform
#endif
jnlpool.jnlpool_ctl = NULL;
pool_init = FALSE;
}
recvpool.upd_proc_local->upd_proc_shutdown = NORMAL_SHUTDOWN;
recvpool.upd_proc_local->upd_proc_pid = 0;
#ifdef UNIX
SHMDT(recvpool.recvpool_ctl);
#elif defined(VMS)
if(SS$_NORMAL != (status = detach_shm(recvpool.shm_range)))
repl_log(stderr, TRUE, TRUE, "Update process could not detach from recvpool : %s\n", REPL_STR_ERROR);
if (SS$_NORMAL != (status = signoff_from_gsec(recvpool.shm_lockid)))
repl_log(stderr, TRUE, TRUE, "Error dequeueing lock on recvpool global section : %s\n", REPL_STR_ERROR);
#else
#error Unsupported Platform
#endif
recvpool.recvpool_ctl = NULL;
gtm_event_log_close();
if (exit)
mupip_exit(SS_NORMAL);
return;
}
/* This function sets the "ftok_counter_halted" field to TRUE in the instance file header and flushes it to disk.
* Caller could be attached to the journal pool or not. If not, update file header directly. If yes, go through locks.
*/
void repl_inst_ftok_counter_halted(unix_db_info *udi, char *file_type, repl_inst_hdr *repl_instance)
{
assert(udi->grabbed_ftok_sem); /* this ensures we have a lock before we modify the instance file header */
if (NULL != jnlpool.repl_inst_filehdr)
{
assert(!jnlpool.repl_inst_filehdr->ftok_counter_halted);
jnlpool.repl_inst_filehdr->ftok_counter_halted = TRUE;
grab_lock(jnlpool.jnlpool_dummy_reg, TRUE, ASSERT_NO_ONLINE_ROLLBACK);
repl_inst_flush_filehdr();
rel_lock(jnlpool.jnlpool_dummy_reg);
} else
{
assert(!repl_instance->ftok_counter_halted);
repl_instance->ftok_counter_halted = TRUE;
repl_inst_write(udi->fn, (off_t)0, (sm_uc_ptr_t)repl_instance, SIZEOF(repl_inst_hdr));
}
/* Ignore any errors while flushing the "halted" value to the file header. The only consequence is other processes
* will incur a performance overhead trying to unnecessarily bump the semaphore counter when it is already ERANGE.
*/
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_NOMORESEMCNT, 5, LEN_AND_LIT("ftok"), file_type, LEN_AND_STR(udi->fn));
}
int gtmsource_losttncomplete(void)
{
int idx;
gtmsource_local_ptr_t gtmsourcelocal_ptr;
error_def(ERR_MUPCLIERR);
error_def(ERR_TEXT);
assert(holds_sem[SOURCE][JNL_POOL_ACCESS_SEM]);
/* We dont need the access control semaphore here. So release it first and avoid any potential deadlocks. */
if (0 != rel_sem(SOURCE, JNL_POOL_ACCESS_SEM))
rts_error(VARLSTCNT(5) ERR_TEXT, 2, RTS_ERROR_LITERAL("Error in source server losttncomplete rel_sem"),
REPL_SEM_ERRNO);
assert(NULL == jnlpool.gtmsource_local);
repl_log(stderr, TRUE, TRUE, "Initiating LOSTTNCOMPLETE operation on instance [%s]\n",
jnlpool.repl_inst_filehdr->this_instname);
/* If this is a root primary instance, propagate this information to secondaries as well so they reset zqgblmod_seqno to 0.
* If propagating primary, no need to send this to tertiaries as the receiver on the tertiary cannot have started with
* non-zero "zqgblmod_seqno" to begin with (PRIMARYNOTROOT error would have been issued).
*/
if (!jnlpool.jnlpool_ctl->upd_disabled)
{
grab_lock(jnlpool.jnlpool_dummy_reg);
jnlpool.jnlpool_ctl->send_losttn_complete = TRUE;
gtmsourcelocal_ptr = jnlpool.gtmsource_local_array;
for (idx = 0; idx < NUM_GTMSRC_LCL; idx++, gtmsourcelocal_ptr++)
{
if (('\0' == gtmsourcelocal_ptr->secondary_instname[0])
&& (0 == gtmsourcelocal_ptr->read_jnl_seqno)
&& (0 == gtmsourcelocal_ptr->connect_jnl_seqno))
continue;
gtmsourcelocal_ptr->send_losttn_complete = TRUE;
}
rel_lock(jnlpool.jnlpool_dummy_reg);
}
/* Reset zqgblmod_seqno and zqgblmod_tn to 0 in this instance as well */
repl_inst_reset_zqgblmod_seqno_and_tn();
return (NORMAL_SHUTDOWN);
}
/* This function is called primarily to append a new histinfo record to the replication instance file by one of the following
* 1) MUPIP REPLIC -SOURCE -START -ROOTPRIMARY command (after forking the child source server) if it created the journal pool.
* 2) MUPIP REPLIC -SOURCE -ACTIVATE -ROOTPRIMARY command if this is a propagating primary to root primary transition.
* In addition, this function also initializes the "lms_group_info" field in the instance file (from the "inst_info" field)
* if the current value is NULL.
*/
void gtmsource_rootprimary_init(seq_num start_seqno)
{
unix_db_info *udi;
repl_histinfo histinfo;
boolean_t was_crit, switch_jnl;
gd_region *reg, *region_top;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
uint4 jnl_status;
udi = FILE_INFO(jnlpool.jnlpool_dummy_reg);
assert(NULL != jnlpool.repl_inst_filehdr);
/* Update journal pool fields to reflect this is a root primary startup and updates are enabled */
assert(!udi->s_addrs.hold_onto_crit || jgbl.onlnrlbk);
was_crit = udi->s_addrs.now_crit;
if (!was_crit)
grab_lock(jnlpool.jnlpool_dummy_reg, TRUE, ASSERT_NO_ONLINE_ROLLBACK);
jnlpool.repl_inst_filehdr->root_primary_cycle++;
/* If this instance is transitioning from a non-rootprimary to rootprimary, switch journal files.
* This helps with maintaining accurate value of csd->zqgblmod_tn when the former primary connects
* to the current primary through a fetchresync-rollback or receiver-server-autorollback..
*/
switch_jnl = (!jnlpool.repl_inst_filehdr->was_rootprimary && (0 < jnlpool.repl_inst_filehdr->num_histinfo));
jnlpool.repl_inst_filehdr->was_rootprimary = TRUE;
assert(start_seqno >= jnlpool.jnlpool_ctl->start_jnl_seqno);
assert(start_seqno == jnlpool.jnlpool_ctl->jnl_seqno);
jnlpool.repl_inst_filehdr->jnl_seqno = start_seqno;
assert(jgbl.onlnrlbk || jnlpool.jnlpool_ctl->upd_disabled);
if (!jgbl.onlnrlbk)
jnlpool.jnlpool_ctl->upd_disabled = FALSE;
if (IS_REPL_INST_UUID_NULL(jnlpool.repl_inst_filehdr->lms_group_info))
{ /* This is the first time this instance is being brought up either as a root primary or as a propagating
* primary. Initialize the "lms_group_info" fields in the instance file header in journal pool shared memory.
* They will be flushed to the instance file as part of the "repl_inst_histinfo_add -> repl_inst_flush_filehdr"
* function invocation below.
*/
assert('\0' == jnlpool.repl_inst_filehdr->lms_group_info.created_nodename[0]);
assert('\0' == jnlpool.repl_inst_filehdr->lms_group_info.this_instname[0]);
assert(!jnlpool.repl_inst_filehdr->lms_group_info.creator_pid);
jnlpool.repl_inst_filehdr->lms_group_info = jnlpool.repl_inst_filehdr->inst_info;
assert('\0' != jnlpool.repl_inst_filehdr->lms_group_info.created_nodename[0]);
DBG_CHECK_CREATED_NODENAME(jnlpool.repl_inst_filehdr->lms_group_info.created_nodename);
assert('\0' != jnlpool.repl_inst_filehdr->lms_group_info.this_instname[0]);
assert(jnlpool.repl_inst_filehdr->lms_group_info.created_time);
assert(jnlpool.repl_inst_filehdr->lms_group_info.creator_pid);
}
/* Initialize histinfo fields */
memcpy(histinfo.root_primary_instname, jnlpool.repl_inst_filehdr->inst_info.this_instname, MAX_INSTNAME_LEN - 1);
histinfo.root_primary_instname[MAX_INSTNAME_LEN - 1] = '\0';
assert('\0' != histinfo.root_primary_instname[0]);
histinfo.start_seqno = start_seqno;
assert(jnlpool.jnlpool_ctl->strm_seqno[0] == jnlpool.repl_inst_filehdr->strm_seqno[0]);
assert(jnlpool.repl_inst_filehdr->is_supplementary || (0 == jnlpool.jnlpool_ctl->strm_seqno[0]));
histinfo.strm_seqno = (!jnlpool.repl_inst_filehdr->is_supplementary) ? 0 : jnlpool.jnlpool_ctl->strm_seqno[0];
histinfo.root_primary_cycle = jnlpool.repl_inst_filehdr->root_primary_cycle;
assert(process_id == getpid());
histinfo.creator_pid = process_id;
JNL_SHORT_TIME(histinfo.created_time);
histinfo.strm_index = 0;
histinfo.history_type = HISTINFO_TYPE_NORMAL;
NULL_INITIALIZE_REPL_INST_UUID(histinfo.lms_group);
/* The following fields will be initialized in the "repl_inst_histinfo_add" function call below.
* histinfo.histinfo_num
* histinfo.prev_histinfo_num
* histinfo.last_histinfo_num[]
*/
/* Add the histinfo record to the instance file and flush the changes in the journal pool to the file header */
repl_inst_histinfo_add(&histinfo);
if (!was_crit)
rel_lock(jnlpool.jnlpool_dummy_reg);
if (switch_jnl)
{
SET_GBL_JREC_TIME; /* jnl_ensure_open/jnl_file_extend and its callees assume jgbl.gbl_jrec_time is set */
for (reg = gd_header->regions, region_top = gd_header->regions + gd_header->n_regions; reg < region_top; reg++)
{
gv_cur_region = reg;
change_reg(); /* sets cs_addrs/cs_data (needed by jnl_ensure_open) */
if (!JNL_ENABLED(cs_addrs))
continue;
grab_crit(gv_cur_region);
jpc = cs_addrs->jnl;
/* 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)
*/
jbp = jpc->jnl_buff;
ADJUST_GBL_JREC_TIME(jgbl, jbp);
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{
if (EXIT_ERR == SWITCH_JNL_FILE(jpc))
rts_error_csa(CSA_ARG(cs_addrs) VARLSTCNT(4) ERR_JNLEXTEND, 2, JNL_LEN_STR(cs_data));
} else
{
if (SS_NORMAL != jpc->status)
rts_error_csa(CSA_ARG(cs_addrs) VARLSTCNT(7) jnl_status, 4, JNL_LEN_STR(cs_data),
DB_LEN_STR(gv_cur_region), jpc->status);
else
rts_error_csa(CSA_ARG(cs_addrs) VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(cs_data),
DB_LEN_STR(gv_cur_region));
}
rel_crit(gv_cur_region);
}
}
}
/*
* This will rundown a replication instance journal (and receiver) pool.
* Input Parameter:
* replpool_id of the instance. Instance file name must be null terminated in replpool_id.
* Returns :
* TRUE, if successful.
* FALSE, otherwise.
*/
boolean_t mu_rndwn_repl_instance(replpool_identifier *replpool_id, boolean_t immediate, boolean_t rndwn_both_pools,
boolean_t *jnlpool_sem_created)
{
boolean_t jnlpool_stat = SS_NORMAL, recvpool_stat = SS_NORMAL, decr_cnt, sem_created = FALSE, ipc_rmvd;
char *instfilename;
unsigned char ipcs_buff[MAX_IPCS_ID_BUF], *ipcs_ptr;
gd_region *r_save;
repl_inst_hdr repl_instance;
static gd_region *reg = NULL;
struct semid_ds semstat;
struct shmid_ds shmstat;
unix_db_info *udi;
int save_errno, sem_id, shm_id, status;
sgmnt_addrs *repl_csa;
boolean_t was_crit;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
if (NULL == reg)
{
r_save = gv_cur_region;
mu_gv_cur_reg_init();
reg = gv_cur_region;
gv_cur_region = r_save;
}
*jnlpool_sem_created = FALSE;
/* Assert that the layout of replpool_identifier is identical for all versions going forward as the function
* "validate_replpool_shm_entry" (used by the argumentless mupip rundown aka "mupip rundown") relies on this.
* This assert is placed here (instead of there) because the automated tests exercise this logic much more
* than the argumentless code. If any of these asserts fail, "validate_replpool_shm_entry" needs to change
* to handle the old and new layouts.
*
* Structure ----> replpool_identifier <---- size 312 [0x0138]
*
* offset = 0000 [0x0000] size = 0012 [0x000c] ----> replpool_identifier.label
* offset = 0012 [0x000c] size = 0001 [0x0001] ----> replpool_identifier.pool_type
* offset = 0013 [0x000d] size = 0036 [0x0024] ----> replpool_identifier.now_running
* offset = 0052 [0x0034] size = 0004 [0x0004] ----> replpool_identifier.repl_pool_key_filler
* offset = 0056 [0x0038] size = 0256 [0x0100] ----> replpool_identifier.instfilename
*/
assert(0 == OFFSETOF(replpool_identifier, label[0]));
assert(12 == SIZEOF(((replpool_identifier *)NULL)->label));
assert(12 == OFFSETOF(replpool_identifier, pool_type));
assert(1 == SIZEOF(((replpool_identifier *)NULL)->pool_type));
assert(13 == OFFSETOF(replpool_identifier, now_running[0]));
assert(36 == SIZEOF(((replpool_identifier *)NULL)->now_running));
assert(56 == OFFSETOF(replpool_identifier, instfilename[0]));
assert(256 == SIZEOF(((replpool_identifier *)NULL)->instfilename));
/* End asserts */
jnlpool.jnlpool_dummy_reg = reg;
recvpool.recvpool_dummy_reg = reg;
instfilename = replpool_id->instfilename;
reg->dyn.addr->fname_len = strlen(instfilename);
assert(0 == instfilename[reg->dyn.addr->fname_len]);
memcpy((char *)reg->dyn.addr->fname, instfilename, reg->dyn.addr->fname_len + 1);
udi = FILE_INFO(reg);
udi->fn = (char *)reg->dyn.addr->fname;
/* Lock replication instance using ftok semaphore so that no other replication process can startup until we are done with
* rundown
*/
if (!ftok_sem_get(reg, TRUE, REPLPOOL_ID, immediate))
return FALSE;
ESTABLISH_RET(mu_rndwn_repl_instance_ch, FALSE);
repl_inst_read(instfilename, (off_t)0, (sm_uc_ptr_t)&repl_instance, SIZEOF(repl_inst_hdr));
assert(rndwn_both_pools || JNLPOOL_SEGMENT == replpool_id->pool_type || RECVPOOL_SEGMENT == replpool_id->pool_type);
if (rndwn_both_pools || (JNLPOOL_SEGMENT == replpool_id->pool_type))
{ /* --------------------------
* First rundown Journal pool
* --------------------------
*/
shm_id = repl_instance.jnlpool_shmid;
if (SS_NORMAL == (jnlpool_stat = mu_replpool_grab_sem(&repl_instance, JNLPOOL_SEGMENT, &sem_created, immediate)))
{
/* Got JNL_POOL_ACCESS_SEM and incremented SRC_SRV_COUNT_SEM */
assert(holds_sem[SOURCE][JNL_POOL_ACCESS_SEM]);
assert(holds_sem[SOURCE][SRC_SERV_COUNT_SEM]);
sem_id = repl_instance.jnlpool_semid;
if ((INVALID_SHMID == shm_id) || (-1 == shmctl(shm_id, IPC_STAT, &shmstat))
|| (shmstat.shm_ctime != repl_instance.jnlpool_shmid_ctime))
{
repl_instance.jnlpool_shmid = shm_id = INVALID_SHMID;
repl_instance.jnlpool_shmid_ctime = 0;
}
assert((INVALID_SHMID != shm_id) || ((NULL == jnlpool.jnlpool_ctl) && (NULL == jnlpool_ctl)));
ipc_rmvd = TRUE;
if (INVALID_SHMID != shm_id)
{
replpool_id->pool_type = JNLPOOL_SEGMENT;
jnlpool_stat = mu_rndwn_replpool(replpool_id, &repl_instance, shm_id, &ipc_rmvd);
ipcs_ptr = i2asc((uchar_ptr_t)ipcs_buff, shm_id);
*ipcs_ptr = '\0';
if (rndwn_both_pools && ((SS_NORMAL != jnlpool_stat) || ipc_rmvd))
gtm_putmsg(VARLSTCNT(6) (jnlpool_stat ? ERR_MUJPOOLRNDWNFL : ERR_MUJPOOLRNDWNSUC),
4, LEN_AND_STR(ipcs_buff), LEN_AND_STR(instfilename));
}
assert(ipc_rmvd || (NULL != jnlpool_ctl));
assert((NULL == jnlpool.jnlpool_ctl) || (SS_NORMAL == jnlpool_stat) || jgbl.onlnrlbk);
assert((INVALID_SHMID != repl_instance.jnlpool_shmid) || (0 == repl_instance.jnlpool_shmid_ctime));
assert((INVALID_SHMID == repl_instance.jnlpool_shmid) || (0 != repl_instance.jnlpool_shmid_ctime));
assert(INVALID_SEMID != sem_id);
if (!mur_options.rollback)
{ /* Invoked by MUPIP RUNDOWN in which case the semaphores needs to be removed. But, remove the
* semaphore ONLY if we created it here OR the journal pool was successfully removed.
*/
if (NULL == jnlpool_ctl)
{
if (((sem_created || (SS_NORMAL == jnlpool_stat))
&& (SS_NORMAL == mu_replpool_release_sem(&repl_instance, JNLPOOL_SEGMENT, TRUE))))
{ /* Now that semaphores are removed, reset fields in file header */
if (!sem_created)
{ /* If sem_id was created by mu_replpool_grab_sem then do NOT report the
* MURPOOLRNDWNSUC message as it indicates that the semaphore was orphaned
* and we removed it when in fact there was no orphaned semaphore and we
* created it as part of mu_replpool_grab_sem to get standalone access to
* rundown the receiver pool (which may or may not exist)
*/
ipcs_ptr = i2asc((uchar_ptr_t)ipcs_buff, sem_id);
*ipcs_ptr = '\0';
gtm_putmsg(VARLSTCNT(9) ERR_MUJPOOLRNDWNSUC, 4, LEN_AND_STR(ipcs_buff),
LEN_AND_STR(instfilename), ERR_SEMREMOVED, 1, sem_id);
}
repl_inst_jnlpool_reset();
}
} else
{ /* Anticipatory Freeze scheme is turned ON. So, release just the JNL_POOL_ACCESS_SEM. The
* semaphore will be released/removed in the caller (mupip_rundown)
*/
assert(ANTICIPATORY_FREEZE_AVAILABLE);
assertpro(SS_NORMAL == (status = rel_sem(SOURCE, JNL_POOL_ACCESS_SEM)));
assert(!holds_sem[SOURCE][JNL_POOL_ACCESS_SEM]);
/* Since we are not resetting the semaphore IDs in the file header, we need to write out
* the semaphore IDs in the instance file (if we created them).
*/
if (sem_created)
repl_inst_write(instfilename, (off_t)0, (sm_uc_ptr_t)&repl_instance,
SIZEOF(repl_inst_hdr));
}
/* If semaphore is not created and the journal pool rundown failed (due to attached processes),
* rundown process continues to holds the journal pool access control semaphore. This way, we hold
* the semaphore on behalf of the source server (now no longer alive) to prevent mu_rndwn_sem_all
* (invoked later) from cleaning up this orphaned semaphore (which causes REPLREQROLLBACK if the
* source server is restarted). But, since the semaphore is not released (until the rundown process
* dies), holds_sem[SOURCE][JNL_POOL_ACCESS_SEM] continues to remain TRUE. This causes asserts in
* ftok_sem_get if mu_rndwn_repl_instance is invoked for a different journal/receive pool. To
* workaround it, set holds_sem[SOURCE][JNL_POOL_ACCESS_SEM] to FALSE. This is an interim solution
* until we record such semaphores in an ignore-list (or some such) and change mu_rndwn_sem_all to
* skip the ones that are present in the ignore list.
*/
holds_sem[SOURCE][JNL_POOL_ACCESS_SEM] = FALSE;
}
} else if (rndwn_both_pools && (INVALID_SHMID != shm_id))
{
ipcs_ptr = i2asc((uchar_ptr_t)ipcs_buff, shm_id);
*ipcs_ptr = '\0';
if (rndwn_both_pools)
gtm_putmsg(VARLSTCNT(6) ERR_MUJPOOLRNDWNFL, 4, LEN_AND_STR(ipcs_buff),
LEN_AND_STR(instfilename));
}
*jnlpool_sem_created = sem_created;
}
if (((SS_NORMAL == jnlpool_stat) || !jgbl.mur_rollback) &&
(rndwn_both_pools || (RECVPOOL_SEGMENT == replpool_id->pool_type)))
{ /* --------------------------
* Now rundown Receivpool
* --------------------------
* Note: RECVPOOL is rundown ONLY if the JNLPOOL rundown was successful. This way, we don't end up
* creating new semaphores for the RECVPOOL if ROLLBACK is not going to start anyways because of the failed
* JNLPOOL rundown. The only exception is MUPIP RUNDOWN command in which case we try running down the
* RECVPOOL even if the JNLPOOL rundown failed.
*/
shm_id = repl_instance.recvpool_shmid;
if (SS_NORMAL == (recvpool_stat = mu_replpool_grab_sem(&repl_instance, RECVPOOL_SEGMENT, &sem_created, immediate)))
{
sem_id = repl_instance.recvpool_semid;
if ((INVALID_SHMID == shm_id) || (-1 == shmctl(shm_id, IPC_STAT, &shmstat))
|| (shmstat.shm_ctime != repl_instance.recvpool_shmid_ctime))
{
repl_instance.recvpool_shmid = shm_id = INVALID_SHMID;
repl_instance.recvpool_shmid_ctime = 0;
}
ipc_rmvd = TRUE;
if (INVALID_SHMID != shm_id)
{
replpool_id->pool_type = RECVPOOL_SEGMENT;
recvpool_stat = mu_rndwn_replpool(replpool_id, &repl_instance, shm_id, &ipc_rmvd);
ipcs_ptr = i2asc((uchar_ptr_t)ipcs_buff, shm_id);
*ipcs_ptr = '\0';
if (rndwn_both_pools && ((SS_NORMAL != recvpool_stat) || ipc_rmvd))
gtm_putmsg(VARLSTCNT(6) (recvpool_stat ? ERR_MURPOOLRNDWNFL : ERR_MURPOOLRNDWNSUC),
4, LEN_AND_STR(ipcs_buff), LEN_AND_STR(instfilename));
}
assert((TRUE == ipc_rmvd) || (SS_NORMAL != recvpool_stat) || jgbl.onlnrlbk);
assert((INVALID_SHMID != repl_instance.recvpool_shmid) || (0 == repl_instance.recvpool_shmid_ctime));
assert((INVALID_SHMID == repl_instance.recvpool_shmid) || (0 != repl_instance.recvpool_shmid_ctime));
assert(INVALID_SEMID != sem_id);
if (!mur_options.rollback)
{ /* Invoked by MUPIP RUNDOWN in which case the semaphores needs to be removed. But, remove the
* semaphore ONLY if we created it here OR the receive pool was successfully removed.
*/
if ((sem_created || (SS_NORMAL == recvpool_stat))
&& (SS_NORMAL == mu_replpool_release_sem(&repl_instance, RECVPOOL_SEGMENT, TRUE)))
{ /* Now that semaphores are removed, reset fields in file header */
if (!sem_created)
{ /* if sem_id was "created" by mu_replpool_grab_sem then do NOT report the
* MURPOOLRNDWNSUC message as it indicates that the semaphore was orphaned and we
* removed it when in fact there was no orphaned semaphore and we "created" it as
* part of mu_replpool_grab_sem to get standalone access to rundown the receiver
* pool (which may or may not exist)
*/
ipcs_ptr = i2asc((uchar_ptr_t)ipcs_buff, sem_id);
*ipcs_ptr = '\0';
gtm_putmsg(VARLSTCNT(9) ERR_MURPOOLRNDWNSUC, 4, LEN_AND_STR(ipcs_buff),
LEN_AND_STR(instfilename), ERR_SEMREMOVED, 1, sem_id);
}
if (NULL != jnlpool_ctl)
{ /* Journal pool is not yet removed. So, grab lock before resetting semid/shmid
* fields in the file header as the function expects the caller to hold crit
* if the journal pool is available
*/
repl_csa = &FILE_INFO(jnlpool.jnlpool_dummy_reg)->s_addrs;
assert(!repl_csa->now_crit);
assert(!repl_csa->hold_onto_crit);
was_crit = repl_csa->now_crit;
/* Since we do grab_lock, below, we need to do a per-process initialization. Also,
* start heartbeat so that grab_lock can issue MUTEXLCKALERT and get C-stacks if
* waiting for crit
*/
START_HEARTBEAT_IF_NEEDED;
mutex_per_process_init();
if (!was_crit)
grab_lock(jnlpool.jnlpool_dummy_reg, TRUE, GRAB_LOCK_ONLY);
}
repl_inst_recvpool_reset();
if ((NULL != jnlpool_ctl) && !was_crit)
rel_lock(jnlpool.jnlpool_dummy_reg);
}
/* If semaphore is not created and the receive pool rundown failed (due to attached processes),
* rundown process continues to holds the receive pool access control semaphore. This way, we hold
* the semaphore on behalf of the receiver server (now no longer alive) to prevent mu_rndwn_sem_all
* (invoked later) from cleaning up this orphaned semaphore (which causes REPLREQROLLBACK if the
* receiver is restarted). But, since the semaphore is not released (until the rundown process
* dies), holds_sem[RECV][RECV_POOL_ACCESS_SEM] continues to remain TRUE. This causes asserts in
* ftok_sem_get if mu_rndwn_repl_instance is invoked for a different journal/receive pool. To
* workaround it, set holds_sem[SOURCE][RECV_POOL_ACCESS_SEM] to FALSE. This is an interim solution
* until we record such semaphores in an ignore-list (or some such) and change mu_rndwn_sem_all to
* skip the ones that are present in the ignore list.
*/
assert((sem_created || (SS_NORMAL == recvpool_stat)) || holds_sem[RECV][RECV_POOL_ACCESS_SEM]);
DEBUG_ONLY(set_sem_set_recvr(sem_id));
}
} else if (rndwn_both_pools && (INVALID_SHMID != shm_id))
{
ipcs_ptr = i2asc((uchar_ptr_t)ipcs_buff, shm_id);
*ipcs_ptr = '\0';
if (rndwn_both_pools)
gtm_putmsg(VARLSTCNT(6) ERR_MURPOOLRNDWNFL, 4, LEN_AND_STR(ipcs_buff),
LEN_AND_STR(instfilename));
}
}
assert(jgbl.onlnrlbk || ANTICIPATORY_FREEZE_AVAILABLE || (NULL == jnlpool.repl_inst_filehdr));
if (mur_options.rollback && (SS_NORMAL == jnlpool_stat) && (SS_NORMAL == recvpool_stat))
{
assert(jgbl.onlnrlbk || ANTICIPATORY_FREEZE_AVAILABLE || ((INVALID_SHMID == repl_instance.jnlpool_shmid)
&& (INVALID_SHMID == repl_instance.recvpool_shmid)));
/* Initialize jnlpool.repl_inst_filehdr as it is used later by gtmrecv_fetchresync() */
decr_cnt = FALSE;
if (NULL == jnlpool.repl_inst_filehdr)
{ /* Possible if there is NO journal pool in the first place. In this case, malloc the structure here and
* copy the file header from repl_instance structure.
*/
jnlpool.repl_inst_filehdr = (repl_inst_hdr_ptr_t)malloc(SIZEOF(repl_inst_hdr));
memcpy(jnlpool.repl_inst_filehdr, &repl_instance, SIZEOF(repl_inst_hdr));
} else
{
assert(repl_instance.jnlpool_semid == jnlpool.repl_inst_filehdr->jnlpool_semid);
assert(repl_instance.jnlpool_semid_ctime == jnlpool.repl_inst_filehdr->jnlpool_semid_ctime);
assert(repl_instance.jnlpool_shmid == jnlpool.repl_inst_filehdr->jnlpool_shmid);
assert(repl_instance.jnlpool_shmid_ctime == jnlpool.repl_inst_filehdr->jnlpool_shmid_ctime);
/* If the ONLINE ROLLBACK command is run on the primary when the source server is up and running,
* jnlpool.repl_inst_filehdr->recvpool_semid will be INVALID because there is NO receiver server
* running. However, ROLLBACK creates semaphores for both journal pool and receive pool and writes
* it to the instance file header. Copy this information to the file header copy in the jnlpool
* as well
*/
jnlpool.repl_inst_filehdr->recvpool_semid = repl_instance.recvpool_semid;
jnlpool.repl_inst_filehdr->recvpool_semid_ctime = repl_instance.recvpool_semid_ctime;
}
/* Flush changes to the replication instance file header to disk */
repl_inst_write(instfilename, (off_t)0, (sm_uc_ptr_t)&repl_instance, SIZEOF(repl_inst_hdr));
} else /* for MUPIP RUNDOWN, semid fields in the file header are reset and is written in mu_replpool_release_sem() above */
decr_cnt = (NULL == jnlpool_ctl); /* for anticipatory freeze, mupip_rundown releases the semaphore */
REVERT;
/* Release replication instance ftok semaphore lock */
if (!ftok_sem_release(reg, decr_cnt, immediate)) /* Do not decrement the counter if ROLLBACK */
return FALSE;
return ((SS_NORMAL == jnlpool_stat) && (SS_NORMAL == recvpool_stat));
}
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;
}
int gtmrecv_poll_actions1(int *pending_data_len, int *buff_unprocessed, unsigned char *buffp)
{
static int report_cnt = 1;
static int next_report_at = 1;
static boolean_t send_xoff = FALSE;
static boolean_t xoff_sent = FALSE;
static seq_num send_seqno;
static boolean_t log_draining_msg = FALSE;
static boolean_t send_badtrans = FALSE;
static boolean_t send_cmp2uncmp = FALSE;
static boolean_t upd_shut_too_early_logged = FALSE;
static time_t last_reap_time = 0;
repl_msg_t xoff_msg;
repl_badtrans_msg_t bad_trans_msg;
boolean_t alert = FALSE, info = FALSE;
int return_status;
gd_region *region_top;
unsigned char *msg_ptr; /* needed for REPL_{SEND,RECV}_LOOP */
int tosend_len, sent_len, sent_this_iter; /* needed for REPL_SEND_LOOP */
int torecv_len, recvd_len, recvd_this_iter; /* needed for REPL_RECV_LOOP */
int status, poll_dir; /* needed for REPL_{SEND,RECV}_LOOP */
int temp_len, pending_msg_size;
int upd_start_status, upd_start_attempts;
int buffered_data_len;
int upd_exit_status;
seq_num temp_send_seqno;
boolean_t bad_trans_detected = FALSE, onln_rlbk_flg_set = FALSE;
recvpool_ctl_ptr_t recvpool_ctl;
upd_proc_local_ptr_t upd_proc_local;
gtmrecv_local_ptr_t gtmrecv_local;
upd_helper_ctl_ptr_t upd_helper_ctl;
pid_t waitpid_res;
int4 msg_type, msg_len;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
recvpool_ctl = recvpool.recvpool_ctl;
upd_proc_local = recvpool.upd_proc_local;
gtmrecv_local = recvpool.gtmrecv_local;
upd_helper_ctl = recvpool.upd_helper_ctl;
if (SHUTDOWN == gtmrecv_local->shutdown)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Shutdown signalled\n");
gtmrecv_end(); /* Won't return */
}
# ifdef GTM_TLS
/* If we sent a REPL_RENEG_ACK, then we cannot afford to send anymore asynchronous messages (like XOFF_ACK_ME) until we
* receive a REPL_RENEG_COMPLETE from the source server. This ensures that while the source server attempts to do a SSL/TLS
* renegotiation, it doesn't have any application data (like XOFF_ACK_ME) sitting in the pipe.
*/
if (REPLTLS_WAITING_FOR_RENEG_COMPLETE == repl_tls.renegotiate_state)
return STOP_POLL;
# endif
/* Reset report_cnt and next_report_at to 1 when a new upd proc is forked */
if ((1 == report_cnt) || (report_cnt == next_report_at))
{
/* A comment on the usage of NO_SHUTDOWN below for the alert variable. Since upd_proc_local->upd_proc_shutdown is
* a shared memory field (and could be concurrently changed by either the receiver server or the update process),
* we want to make sure it is the same value BEFORE and AFTER checking whether the update process is alive or not.
* If it is not NO_SHUTDOWN (i.e. is SHUTDOWN or NORMAL_SHUTDOWN or ABNORMAL_SHUTDOWN) it has shut down due to
* an external request so we do want to send out a false update-process-is-not-alive alert.
*/
if ((alert = ((NO_SHUTDOWN == upd_proc_local->upd_proc_shutdown) && (SRV_DEAD == is_updproc_alive())
&& (NO_SHUTDOWN == upd_proc_local->upd_proc_shutdown)))
|| (info = (((NORMAL_SHUTDOWN == upd_proc_local->upd_proc_shutdown)
|| (ABNORMAL_SHUTDOWN == upd_proc_local->upd_proc_shutdown)) && (SRV_DEAD == is_updproc_alive()))))
{
if (alert)
repl_log(gtmrecv_log_fp, TRUE, TRUE,
"ALERT : Receiver Server detected that Update Process is not ALIVE\n");
else
repl_log(gtmrecv_log_fp, TRUE, TRUE,
"INFO : Update process not running due to user initiated shutdown\n");
if (1 == report_cnt)
{
send_xoff = TRUE;
recvpool_ctl->old_jnl_seqno = recvpool_ctl->jnl_seqno;
recvpool_ctl->jnl_seqno = 0;
/* Even though we have identified that the update process is NOT alive, a waitpid on the update
* process PID is necessary so that the system doesn't leave any zombie process lying around.
* This is possible since any child process that dies without the parent doing a waitpid on it
* will be defunct unless the parent dies at which point the "init" process takes the role of
* the parent and invokes waitpid to remove the zombies.
* NOTE: It is possible that the update process was killed before the receiver server got a
* chance to record it's PID in the recvpool.upd_proc_local structure. In such a case, don't
* invoke waitpid as that will block us (receiver server) if this instance of the receiver
* server was started with helper processes.
*/
if (0 < upd_proc_local->upd_proc_pid)
{
WAITPID(upd_proc_local->upd_proc_pid, &upd_exit_status, 0, waitpid_res);
/* Since the update process as part of its shutdown does NOT reset the upd_proc_pid, reset
* it here ONLY if the update process was NOT kill -9ed. This is needed because receiver
* server as part of its shutdown relies on this field (upd_proc_pid) to determine if the
* update process was cleanly shutdown or was kill -9ed.
*/
if (!alert)
upd_proc_local->upd_proc_pid = 0;
}
upd_proc_local->bad_trans = FALSE; /* No point in doing bad transaction processing */
upd_proc_local->onln_rlbk_flg = FALSE; /* No point handling online rollback */
}
gtmrecv_wait_for_jnl_seqno = TRUE;
REPL_DPRINT1(
"gtmrecv_poll_actions : Setting gtmrecv_wait_for_jnl_seqno to TRUE because of upd crash/shutdown\n");
next_report_at *= GTMRECV_NEXT_REPORT_FACTOR;
report_cnt++;
}
} else
report_cnt++;
/* Check if REPL_CMP2UNCMP or REPL_BADTRANS message needs to be sent */
if (upd_proc_local->onln_rlbk_flg)
{ /* Update process detected an online rollback and is requesting us to restart the connection. But before that, send
* REPL_XOFF source side and drain the replication pipe
*/
onln_rlbk_flg_set = TRUE;
send_xoff = TRUE;
} else if (!send_cmp2uncmp && gtmrecv_send_cmp2uncmp)
{
send_xoff = TRUE;
send_seqno = recvpool_ctl->jnl_seqno;
send_cmp2uncmp = TRUE;
} else if (!send_badtrans && upd_proc_local->bad_trans)
{
send_xoff = TRUE;
send_seqno = upd_proc_local->read_jnl_seqno;
send_badtrans = TRUE;
bad_trans_detected = TRUE;
} else if (!upd_proc_local->bad_trans && send_badtrans && 1 != report_cnt)
{
send_badtrans = FALSE;
bad_trans_detected = FALSE;
}
if (send_xoff && !xoff_sent)
{ /* Send XOFF_ACK_ME if the receiver has a connection to the source. Do not attempt to send it if we dont even
* know the endianness of the remote side. In that case, we are guaranteed no initial handshake occurred and
* so no point sending the XOFF too. This saves us lots of trouble in case of cross-endian replication connections.
*/
assert((FD_INVALID != gtmrecv_sock_fd) || repl_connection_reset);
if ((FD_INVALID != gtmrecv_sock_fd) && remote_side->endianness_known)
{
send_seqno = upd_proc_local->read_jnl_seqno;
if (!remote_side->cross_endian)
{
xoff_msg.type = REPL_XOFF_ACK_ME;
xoff_msg.len = MIN_REPL_MSGLEN;
memcpy((uchar_ptr_t)&xoff_msg.msg[0], (uchar_ptr_t)&send_seqno, SIZEOF(seq_num));
} else
{
xoff_msg.type = GTM_BYTESWAP_32(REPL_XOFF_ACK_ME);
xoff_msg.len = GTM_BYTESWAP_32(MIN_REPL_MSGLEN);
temp_send_seqno = GTM_BYTESWAP_64(send_seqno);
memcpy((uchar_ptr_t)&xoff_msg.msg[0], (uchar_ptr_t)&temp_send_seqno, SIZEOF(seq_num));
}
REPL_SEND_LOOP(gtmrecv_sock_fd, &xoff_msg, MIN_REPL_MSGLEN, REPL_POLL_NOWAIT)
; /* Empty Body */
if (SS_NORMAL != status)
{
if (REPL_CONN_RESET(status) && EREPL_SEND == repl_errno)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Connection reset while sending XOFF_ACK_ME. "
"Status = %d ; %s\n", status, STRERROR(status));
repl_close(>mrecv_sock_fd);
repl_connection_reset = TRUE;
xoff_sent = FALSE;
send_badtrans = FALSE;
} else if (EREPL_SEND == repl_errno)
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error sending XOFF msg due to BAD_TRANS or UPD crash/shutdown. "
"Error in send"), status);
else
{
assert(EREPL_SELECT == repl_errno);
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error sending XOFF msg due to BAD_TRANS or UPD crash/shutdown. "
"Error in select"), status);
}
} else
{
xoff_sent = TRUE;
log_draining_msg = TRUE;
}
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL_XOFF_ACK_ME sent due to upd shutdown/crash or bad trans "
"or ONLINE_ROLLBACK\n");
send_xoff = FALSE;
} else
{ /* Connection has been lost OR initial handshake needs to happen again, so no point sending XOFF/BADTRANS */
send_xoff = FALSE;
send_badtrans = FALSE;
}
}
/* Drain pipe */
if (xoff_sent)
{
if (log_draining_msg)
{ /* avoid multiple logs per instance */
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL INFO - Draining replication pipe due to %s\n",
send_cmp2uncmp ? "CMP2UNCMP" : (send_badtrans ? "BAD_TRANS" :
(onln_rlbk_flg_set ? "ONLINE_ROLLBACK" : "UPD shutdown/crash")));
log_draining_msg = FALSE;
}
if (0 != *buff_unprocessed)
{ /* Throw away the current contents of the buffer */
buffered_data_len = ((*pending_data_len <= *buff_unprocessed) ? *pending_data_len : *buff_unprocessed);
*buff_unprocessed -= buffered_data_len;
buffp += buffered_data_len;
*pending_data_len -= buffered_data_len;
REPL_DPRINT2("gtmrecv_poll_actions : (1) Throwing away %d bytes from old buffer while draining\n",
buffered_data_len);
assert(remote_side->endianness_known); /* only then is remote_side->cross_endian reliable */
while (REPL_MSG_HDRLEN <= *buff_unprocessed)
{
assert(0 == (((unsigned long)buffp) % REPL_MSG_ALIGN));
msg_len = ((repl_msg_ptr_t)buffp)->len;
msg_type = ((repl_msg_ptr_t)buffp)->type;
if (remote_side->cross_endian)
{
msg_len = GTM_BYTESWAP_32(msg_len);
msg_type = GTM_BYTESWAP_32(msg_type);
}
msg_type = (msg_type & REPL_TR_CMP_MSG_TYPE_MASK);
assert((REPL_TR_CMP_JNL_RECS == msg_type) || (0 == (msg_len % REPL_MSG_ALIGN)));
*pending_data_len = ROUND_UP2(msg_len, REPL_MSG_ALIGN);
buffered_data_len = ((*pending_data_len <= *buff_unprocessed) ?
*pending_data_len : *buff_unprocessed);
*buff_unprocessed -= buffered_data_len;
buffp += buffered_data_len;
*pending_data_len -= buffered_data_len;
REPL_DPRINT3("gtmrecv_poll_actions : (1) Throwing away message of "
"type %d and length %d from old buffer while draining\n", msg_type, buffered_data_len);
}
if (0 < *buff_unprocessed)
{
memmove((unsigned char *)gtmrecv_msgp, buffp, *buff_unprocessed);
REPL_DPRINT2("gtmrecv_poll_actions : Incomplete header of length %d while draining\n",
*buff_unprocessed);
}
}
status = SS_NORMAL;
if (0 != *buff_unprocessed || 0 == *pending_data_len)
{ /* Receive the header of a message */
assert(REPL_MSG_HDRLEN > *buff_unprocessed); /* so we dont pass negative length in REPL_RECV_LOOP */
REPL_RECV_LOOP(gtmrecv_sock_fd, ((unsigned char *)gtmrecv_msgp) + *buff_unprocessed,
(REPL_MSG_HDRLEN - *buff_unprocessed), REPL_POLL_WAIT)
; /* Empty Body */
if (SS_NORMAL == status)
{
assert(remote_side->endianness_known); /* only then is remote_side->cross_endian reliable */
if (!remote_side->cross_endian)
{
msg_len = gtmrecv_msgp->len;
msg_type = gtmrecv_msgp->type;
} else
{
msg_len = GTM_BYTESWAP_32(gtmrecv_msgp->len);
msg_type = GTM_BYTESWAP_32(gtmrecv_msgp->type);
}
msg_type = (msg_type & REPL_TR_CMP_MSG_TYPE_MASK);
assert((REPL_TR_CMP_JNL_RECS == msg_type) || (0 == (msg_len % REPL_MSG_ALIGN)));
msg_len = ROUND_UP2(msg_len, REPL_MSG_ALIGN);
REPL_DPRINT3("gtmrecv_poll_actions : Received message of type %d and length %d while draining\n",
msg_type, msg_len);
}
}
if ((SS_NORMAL == status) && (0 != *buff_unprocessed || 0 == *pending_data_len) && (REPL_XOFF_ACK == msg_type))
{ /* Receive the rest of the XOFF_ACK msg and signal the drain as complete */
REPL_RECV_LOOP(gtmrecv_sock_fd, gtmrecv_msgp, (MIN_REPL_MSGLEN - REPL_MSG_HDRLEN), REPL_POLL_WAIT)
; /* Empty Body */
if (SS_NORMAL == status)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE,
"REPL INFO - XOFF_ACK received. Drained replication pipe completely\n");
upd_shut_too_early_logged = FALSE;
xoff_sent = FALSE;
return_status = STOP_POLL;
}
} else if (SS_NORMAL == status)
{ /* Drain the rest of the message */
if (0 < *pending_data_len)
{
pending_msg_size = *pending_data_len;
REPL_DPRINT2("gtmrecv_poll_actions : (2) Throwing away %d bytes from pipe\n", pending_msg_size);
} else
{
pending_msg_size = msg_len - REPL_MSG_HDRLEN;
REPL_DPRINT3("gtmrecv_poll_actions : (2) Throwing away message of "
"type %d and length %d from pipe\n", msg_type, msg_len);
}
for ( ; SS_NORMAL == status && 0 < pending_msg_size; pending_msg_size -= gtmrecv_max_repl_msglen)
{
temp_len = (pending_msg_size < gtmrecv_max_repl_msglen)? pending_msg_size : gtmrecv_max_repl_msglen;
REPL_RECV_LOOP(gtmrecv_sock_fd, gtmrecv_msgp, temp_len, REPL_POLL_WAIT)
; /* Empty Body */
}
*buff_unprocessed = 0; *pending_data_len = 0;
if (SS_NORMAL == status && info && !upd_shut_too_early_logged)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "ALERT : User initiated shutdown of Update Process done "
"when there was data in the replication pipe\n");
upd_shut_too_early_logged = TRUE;
}
return_status = CONTINUE_POLL;
}
if (SS_NORMAL != status)
{
if (EREPL_RECV == repl_errno)
{
if (REPL_CONN_RESET(status))
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Connection reset while receiving XOFF_ACK. "
"Status = %d ; %s\n", status, STRERROR(status));
repl_close(>mrecv_sock_fd);
repl_connection_reset = TRUE;
xoff_sent = FALSE;
send_badtrans = FALSE;
return_status = STOP_POLL;
} else
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error while draining replication pipe. Error in recv"), status);
} else
{
assert(EREPL_SELECT == repl_errno);
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error while draining replication pipe. Error in select"), status);
}
}
} else
return_status = STOP_POLL;
/* Like was done before for the XOFF_ACK_ME message, send a BADTRANS/CMP2UNCMP message only if we know
* the endianness of the other side. If not, no point in sending one anyways and saves us trouble in
* case of cross-endian replication connections.
*/
if ((STOP_POLL == return_status) && (send_badtrans || send_cmp2uncmp)
&& (FD_INVALID != gtmrecv_sock_fd) && remote_side->endianness_known)
{ /* Send REPL_BADTRANS or REPL_CMP2UNCMP message */
if (!remote_side->cross_endian)
{
bad_trans_msg.type = send_cmp2uncmp ? REPL_CMP2UNCMP : REPL_BADTRANS;
bad_trans_msg.len = MIN_REPL_MSGLEN;
bad_trans_msg.start_seqno = send_seqno;
} else
{
bad_trans_msg.type = send_cmp2uncmp ? GTM_BYTESWAP_32(REPL_CMP2UNCMP) : GTM_BYTESWAP_32(REPL_BADTRANS);
bad_trans_msg.len = GTM_BYTESWAP_32(MIN_REPL_MSGLEN);
bad_trans_msg.start_seqno = GTM_BYTESWAP_64(send_seqno);
}
REPL_SEND_LOOP(gtmrecv_sock_fd, &bad_trans_msg, bad_trans_msg.len, REPL_POLL_NOWAIT)
; /* Empty Body */
if (SS_NORMAL == status)
{
if (send_cmp2uncmp)
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL_CMP2UNCMP message sent with seqno %llu\n", send_seqno);
else
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL_BADTRANS message sent with seqno %llu\n", send_seqno);
} else
{
if (REPL_CONN_RESET(status) && EREPL_SEND == repl_errno)
{
if (send_cmp2uncmp)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Connection reset while sending REPL_CMP2UNCMP. "
"Status = %d ; %s\n", status, STRERROR(status));
} else
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Connection reset while sending REPL_BADTRANS. "
"Status = %d ; %s\n", status, STRERROR(status));
}
repl_close(>mrecv_sock_fd);
repl_connection_reset = TRUE;
return_status = STOP_POLL;
} else if (EREPL_SEND == repl_errno)
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error sending REPL_BADTRANS/REPL_CMP2UNCMP. Error in send"), status);
else
{
assert(EREPL_SELECT == repl_errno);
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error sending REPL_BADTRANS/REPL_CMP2UNCMP. Error in select"), status);
}
}
send_badtrans = FALSE;
if (send_cmp2uncmp)
{
REPL_DPRINT1("gtmrecv_poll_actions : Setting gtmrecv_wait_for_jnl_seqno to TRUE because this receiver"
"server requested a fall-back from compressed to uncompressed operation\n");
gtmrecv_wait_for_jnl_seqno = TRUE;/* set this to TRUE to break out and go back to a fresh "do_main_loop" */
gtmrecv_bad_trans_sent = TRUE;
gtmrecv_send_cmp2uncmp = FALSE;
send_cmp2uncmp = FALSE;
}
}
if ((upd_proc_local->bad_trans && bad_trans_detected) || onln_rlbk_flg_set
|| (UPDPROC_START == upd_proc_local->start_upd) && (1 != report_cnt))
{
if (UPDPROC_START == upd_proc_local->start_upd)
{
assert(is_updproc_alive() != SRV_ALIVE);
upd_proc_local->upd_proc_shutdown = NO_SHUTDOWN;
}
recvpool_ctl->wrapped = FALSE;
recvpool_ctl->write_wrap = recvpool_ctl->recvpool_size;
recvpool_ctl->write = 0;
/* Reset last_rcvd_histinfo, last_valid_histinfo etc. as they reflect context from unprocessed data
* in the receive pool and those are no longer valid because we have drained the receive pool.
*/
GTMRECV_CLEAR_CACHED_HISTINFO(recvpool.recvpool_ctl, jnlpool, jnlpool_ctl, INSERT_STRM_HISTINFO_FALSE);
if (UPDPROC_START == upd_proc_local->start_upd)
{
/* Attempt starting the update process */
for (upd_start_attempts = 0;
UPDPROC_START_ERR == (upd_start_status = gtmrecv_upd_proc_init(FALSE)) &&
GTMRECV_MAX_UPDSTART_ATTEMPTS > upd_start_attempts;
upd_start_attempts++)
{
if (EREPL_UPDSTART_SEMCTL == repl_errno || EREPL_UPDSTART_BADPATH == repl_errno)
{
gtmrecv_autoshutdown();
} else if (EREPL_UPDSTART_FORK == repl_errno)
{
/* Couldn't start up update now, can try later */
LONG_SLEEP(GTMRECV_WAIT_FOR_PROC_SLOTS);
continue;
} else if (EREPL_UPDSTART_EXEC == repl_errno)
{
/* In forked child, could not exec, should exit */
gtmrecv_exit(ABNORMAL_SHUTDOWN);
}
}
if (UPDPROC_STARTED == (upd_proc_local->start_upd = upd_start_status))
{
REPL_DPRINT1("gtmrecv_poll_actions : Setting gtmrecv_wait_for_jnl_seqno to TRUE because of "
"upd restart\n");
gtmrecv_wait_for_jnl_seqno = TRUE;
report_cnt = next_report_at = 1;
if (send_xoff && (FD_INVALID == gtmrecv_sock_fd))
{
/* Update start command was issued before connection was established,
* no point in sending XOFF. */
send_xoff = FALSE;
}
} else
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "%d failed attempts to fork update process. Try later\n",
upd_start_attempts);
}
} else
{
gtmrecv_wait_for_jnl_seqno = TRUE;/* set this to TRUE to break out and go back to a fresh "do_main_loop" */
if (onln_rlbk_flg_set)
{
assert(NULL != jnlpool_ctl);
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Closing connection due to ONLINE ROLLBACK\n");
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL INFO - Current Jnlpool Seqno : %llu\n",
jnlpool_ctl->jnl_seqno);
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL INFO - Current Receive Pool Seqno : %llu\n",
recvpool_ctl->jnl_seqno);
repl_close(>mrecv_sock_fd);
repl_connection_reset = TRUE;
xoff_sent = FALSE;
send_badtrans = FALSE;
upd_proc_local->onln_rlbk_flg = FALSE;
/* Before restarting afresh, sync the online rollback cycles. This way any future grab_lock that
* we do after restarting should not realize an unhandled online rollback. For receiver, it is
* just syncing the journal pool cycles as the databases are not opened. But, to be safe, grab
* the lock and sync the cycles.
*/
grab_lock(jnlpool.jnlpool_dummy_reg, TRUE, GRAB_LOCK_ONLY);
SYNC_ONLN_RLBK_CYCLES;
rel_lock(jnlpool.jnlpool_dummy_reg);
return_status = STOP_POLL;
recvpool_ctl->jnl_seqno = 0;
} else
{
REPL_DPRINT1("gtmrecv_poll_actions : Setting gtmrecv_wait_for_jnl_seqno to TRUE because bad trans"
"sent\n");
gtmrecv_bad_trans_sent = TRUE;
upd_proc_local->bad_trans = FALSE;
recvpool_ctl->jnl_seqno = upd_proc_local->read_jnl_seqno;
}
}
}
if ((0 == *pending_data_len) && (0 != gtmrecv_local->changelog))
{
if (gtmrecv_local->changelog & REPLIC_CHANGE_LOGINTERVAL)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Changing log interval from %u to %u\n",
log_interval, gtmrecv_local->log_interval);
log_interval = gtmrecv_local->log_interval;
gtmrecv_reinit_logseqno(); /* will force a LOG on the first recv following the interval change */
}
if (gtmrecv_local->changelog & REPLIC_CHANGE_LOGFILE)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Changing log file to %s\n", gtmrecv_local->log_file);
repl_log_init(REPL_GENERAL_LOG, >mrecv_log_fd, gtmrecv_local->log_file);
repl_log_fd2fp(>mrecv_log_fp, gtmrecv_log_fd);
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Change log to %s successful\n",gtmrecv_local->log_file);
}
/* NOTE: update process and receiver each ignore any setting specific to the other (REPLIC_CHANGE_UPD_LOGINTERVAL,
* REPLIC_CHANGE_LOGINTERVAL) */
if (REPLIC_CHANGE_LOGINTERVAL == gtmrecv_local->changelog)
upd_proc_local->changelog = 0;
else
upd_proc_local->changelog = gtmrecv_local->changelog; /* Pass changelog request to the update process */
gtmrecv_local->changelog = 0;
}
if (0 == *pending_data_len && !gtmrecv_logstats && gtmrecv_local->statslog)
{
gtmrecv_logstats = TRUE;
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Begin statistics logging\n");
} else if (0 == *pending_data_len && gtmrecv_logstats && !gtmrecv_local->statslog)
{
gtmrecv_logstats = FALSE;
/* Force all data out to the file before closing the file */
repl_log(gtmrecv_log_fp, TRUE, TRUE, "End statistics logging\n");
}
if (0 == *pending_data_len)
{
if (upd_helper_ctl->start_helpers)
{
gtmrecv_helpers_init(upd_helper_ctl->start_n_readers, upd_helper_ctl->start_n_writers);
upd_helper_ctl->start_helpers = FALSE;
}
if (HELPER_REAP_NONE != (status = upd_helper_ctl->reap_helpers) ||
(double)GTMRECV_REAP_HELPERS_INTERVAL <= difftime(gtmrecv_now, last_reap_time))
{
gtmrecv_reap_helpers(HELPER_REAP_WAIT == status);
last_reap_time = gtmrecv_now;
}
}
return (return_status);
}
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;
}
int gtmsource()
{
int status, log_init_status, waitpid_res, save_errno;
char print_msg[1024], tmpmsg[1024];
gd_region *reg, *region_top;
sgmnt_addrs *csa, *repl_csa;
boolean_t all_files_open, isalive;
pid_t pid, ppid, procgp;
seq_num read_jnl_seqno, jnl_seqno;
unix_db_info *udi;
gtmsource_local_ptr_t gtmsource_local;
boolean_t this_side_std_null_coll;
int null_fd, rc;
memset((uchar_ptr_t)&jnlpool, 0, SIZEOF(jnlpool_addrs));
call_on_signal = gtmsource_sigstop;
ESTABLISH_RET(gtmsource_ch, SS_NORMAL);
if (-1 == gtmsource_get_opt())
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_MUPCLIERR);
if (gtmsource_options.shut_down)
{ /* Wait till shutdown time nears even before going to "jnlpool_init". This is because the latter will return
* with the ftok semaphore and access semaphore held and we do not want to be holding those locks (while
* waiting for the user specified timeout to expire) as that will affect new GTM processes and/or other
* MUPIP REPLIC commands that need these locks for their function.
*/
if (0 < gtmsource_options.shutdown_time)
{
repl_log(stdout, TRUE, TRUE, "Waiting for %d seconds before signalling shutdown\n",
gtmsource_options.shutdown_time);
LONG_SLEEP(gtmsource_options.shutdown_time);
} else
repl_log(stdout, TRUE, TRUE, "Signalling shutdown immediate\n");
} else if (gtmsource_options.start)
{
repl_log(stdout, TRUE, TRUE, "Initiating START of source server for secondary instance [%s]\n",
gtmsource_options.secondary_instname);
}
if (gtmsource_options.activate && (ROOTPRIMARY_SPECIFIED == gtmsource_options.rootprimary))
{ /* MUPIP REPLIC -SOURCE -ACTIVATE -UPDOK has been specified. We need to open the gld and db regions now
* in case this is a secondary -> primary transition. This is so we can later switch journal files in all
* journaled regions when the transition actually happens inside "gtmsource_rootprimary_init". But since
* we have not yet done a "jnlpool_init", we dont know if updates are disabled in it or not. Although we
* need to do the gld/db open only if updates are currently disabled in the jnlpool, we do this always
* because once we do a jnlpool_init, we will come back with the ftok on the jnlpool held and that has
* issues with later db open since we will try to hold the db ftok as part of db open and the ftok logic
* currently has assumptions that a process holds only one ftok at any point in time.
*/
assert(NULL == gd_header);
gvinit();
all_files_open = region_init(FALSE);
if (!all_files_open)
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_NOTALLDBOPN);
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
}
jnlpool_init(GTMSOURCE, gtmsource_options.start, &is_jnlpool_creator);
/* is_jnlpool_creator == TRUE ==> this process created the journal pool
* is_jnlpool_creator == FALSE ==> journal pool already existed and this process simply attached to it.
*/
if (gtmsource_options.shut_down)
gtmsource_exit(gtmsource_shutdown(FALSE, NORMAL_SHUTDOWN) - NORMAL_SHUTDOWN);
else if (gtmsource_options.activate)
gtmsource_exit(gtmsource_mode_change(GTMSOURCE_MODE_ACTIVE_REQUESTED) - NORMAL_SHUTDOWN);
else if (gtmsource_options.deactivate)
gtmsource_exit(gtmsource_mode_change(GTMSOURCE_MODE_PASSIVE_REQUESTED) - NORMAL_SHUTDOWN);
else if (gtmsource_options.checkhealth)
gtmsource_exit(gtmsource_checkhealth() - NORMAL_SHUTDOWN);
else if (gtmsource_options.changelog)
gtmsource_exit(gtmsource_changelog() - NORMAL_SHUTDOWN);
else if (gtmsource_options.showbacklog)
gtmsource_exit(gtmsource_showbacklog() - NORMAL_SHUTDOWN);
else if (gtmsource_options.stopsourcefilter)
gtmsource_exit(gtmsource_stopfilter() - NORMAL_SHUTDOWN);
else if (gtmsource_options.jnlpool)
gtmsource_exit(gtmsource_jnlpool() - NORMAL_SHUTDOWN);
else if (gtmsource_options.losttncomplete)
gtmsource_exit(gtmsource_losttncomplete() - NORMAL_SHUTDOWN);
else if (gtmsource_options.needrestart)
gtmsource_exit(gtmsource_needrestart() - NORMAL_SHUTDOWN);
else if (gtmsource_options.showfreeze)
gtmsource_exit(gtmsource_showfreeze() - NORMAL_SHUTDOWN);
else if (gtmsource_options.setfreeze)
gtmsource_exit(gtmsource_setfreeze() - NORMAL_SHUTDOWN);
else if (!gtmsource_options.start)
{
assert(CLI_PRESENT == cli_present("STATSLOG"));
gtmsource_exit(gtmsource_statslog() - NORMAL_SHUTDOWN);
}
assert(gtmsource_options.start);
# ifndef REPL_DEBUG_NOBACKGROUND
/* Set "child_server_running" to FALSE before forking off child. Wait for it to be set to TRUE by the child. */
gtmsource_local = jnlpool.gtmsource_local;
gtmsource_local->child_server_running = FALSE;
FORK(pid);
if (0 > pid)
{
save_errno = errno;
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0,
ERR_TEXT, 2, RTS_ERROR_LITERAL("Could not fork source server"), save_errno);
} else if (0 < pid)
{ /* Parent. Wait until child sets "child_server_running" to FALSE. That is an indication that the child
* source server has completed its initialization phase and is all set so the parent command can return.
*/
while (isalive = is_proc_alive(pid, 0)) /* note : intended assignment */
{
if (gtmsource_local->child_server_running)
break;
/* To take care of reassignment of PIDs, the while condition should be && with the condition
* (PPID of pid == process_id)
*/
SHORT_SLEEP(GTMSOURCE_WAIT_FOR_SRV_START);
WAITPID(pid, &status, WNOHANG, waitpid_res); /* Release defunct child if dead */
}
if (isalive)
{ /* Child process is alive and started with no issues */
if (0 != (save_errno = rel_sem(SOURCE, JNL_POOL_ACCESS_SEM)))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0,
ERR_TEXT, 2, RTS_ERROR_LITERAL("Error in rel_sem"), save_errno);
ftok_sem_release(jnlpool.jnlpool_dummy_reg, TRUE, TRUE);
} else
{ /* Child source server process errored out at startup and is no longer alive.
* If we were the one who created the journal pool, let us clean it up.
*/
repl_log(stdout, TRUE, TRUE, "Source server startup failed. See source server log file\n");
if (is_jnlpool_creator)
status = gtmsource_shutdown(TRUE, NORMAL_SHUTDOWN);
}
/* If the parent is killed (or crashes) between the fork and exit, checkhealth may not detect that startup
* is in progress - parent forks and dies, the system will release sem 0 and 1, checkhealth might test the
* value of sem 1 before the child grabs sem 1.
*/
gtmsource_exit(isalive ? SRV_ALIVE : SRV_ERR);
}
/* Point stdin to /dev/null */
OPENFILE("/dev/null", O_RDONLY, null_fd);
if (0 > null_fd)
rts_error_csa(CSA_ARG(NULL) ERR_REPLERR, RTS_ERROR_LITERAL("Failed to open /dev/null for read"), errno, 0);
FCNTL3(null_fd, F_DUPFD, 0, rc);
if (0 > rc)
rts_error_csa(CSA_ARG(NULL) ERR_REPLERR, RTS_ERROR_LITERAL("Failed to set stdin to /dev/null"), errno, 0);
CLOSEFILE(null_fd, rc);
if (0 > rc)
rts_error_csa(CSA_ARG(NULL) ERR_REPLERR, RTS_ERROR_LITERAL("Failed to close /dev/null"), errno, 0);
/* The parent process (source server startup command) will be holding the ftok semaphore and jnlpool access semaphore
* at this point. The variables that indicate this would have been copied over to the child during the fork. This will
* make the child think it is actually holding them as well when actually it is not. Reset those variables in the child
* to ensure they do not misrepresent the holder of those semaphores.
*/
ftok_sem_reg = NULL;
udi = FILE_INFO(jnlpool.jnlpool_dummy_reg);
assert(udi->grabbed_ftok_sem);
udi->grabbed_ftok_sem = FALSE;
assert(holds_sem[SOURCE][JNL_POOL_ACCESS_SEM]);
holds_sem[SOURCE][JNL_POOL_ACCESS_SEM] = FALSE;
assert(!holds_sem[SOURCE][SRC_SERV_COUNT_SEM]);
/* Start child source server initialization */
is_src_server = TRUE;
OPERATOR_LOG_MSG;
process_id = getpid();
/* Reinvoke secshr related initialization with the child's pid */
INVOKE_INIT_SECSHR_ADDRS;
/* Initialize mutex socket, memory semaphore etc. before any "grab_lock" is done by this process on the journal pool.
* Note that the initialization would already have been done by the parent receiver startup command but we need to
* redo the initialization with the child process id.
*/
assert(mutex_per_process_init_pid && (mutex_per_process_init_pid != process_id));
mutex_per_process_init();
START_HEARTBEAT_IF_NEEDED;
ppid = getppid();
log_init_status = repl_log_init(REPL_GENERAL_LOG, >msource_log_fd, gtmsource_options.log_file);
assert(SS_NORMAL == log_init_status);
repl_log_fd2fp(>msource_log_fp, gtmsource_log_fd);
if (-1 == (procgp = setsid()))
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Source server error in setsid"), errno);
# endif /* REPL_DEBUG_NOBACKGROUND */
if (ZLIB_CMPLVL_NONE != gtm_zlib_cmp_level)
gtm_zlib_init(); /* Open zlib shared library for compression/decompression */
REPL_DPRINT1("Setting up regions\n");
gvinit();
/* We use the same code dse uses to open all regions but we must make sure they are all open before proceeding. */
all_files_open = region_init(FALSE);
if (!all_files_open)
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_NOTALLDBOPN);
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
/* Determine primary side null subscripts collation order */
/* Also check whether all regions have same null collation order */
this_side_std_null_coll = -1;
for (reg = gd_header->regions, region_top = gd_header->regions + gd_header->n_regions; reg < region_top; reg++)
{
csa = &FILE_INFO(reg)->s_addrs;
if (this_side_std_null_coll != csa->hdr->std_null_coll)
{
if (-1 == this_side_std_null_coll)
this_side_std_null_coll = csa->hdr->std_null_coll;
else
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_NULLCOLLDIFF);
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
}
if (!REPL_ALLOWED(csa) && JNL_ALLOWED(csa))
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_REPLOFFJNLON, 2, DB_LEN_STR(reg));
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
if (reg->read_only && REPL_ALLOWED(csa))
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(6) ERR_JNLPOOLSETUP, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Source Server does not have write permissions to one or "
"more database files that are replicated"));
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
}
/* Initialize source server alive/dead state related fields in "gtmsource_local" before the ftok semaphore is released */
gtmsource_local->gtmsource_pid = process_id;
gtmsource_local->gtmsource_state = GTMSOURCE_START;
if (is_jnlpool_creator)
{
DEBUG_ONLY(jnlpool.jnlpool_ctl->jnlpool_creator_pid = process_id);
gtmsource_seqno_init(this_side_std_null_coll);
if (ROOTPRIMARY_SPECIFIED == gtmsource_options.rootprimary)
{ /* Created the journal pool as a root primary. Append a history record to the replication instance file.
* Invoke the function "gtmsource_rootprimary_init" to do that.
*/
gtmsource_rootprimary_init(jnlpool.jnlpool_ctl->jnl_seqno);
}
}
/* after this point we can no longer have the case where all the regions are unreplicated/non-journaled. */
# ifndef REPL_DEBUG_NOBACKGROUND
/* It is necessary for every process that is using the ftok semaphore to increment the counter by 1. This is used
* by the last process that shuts down to delete the ftok semaphore when it notices the counter to be 0.
* Note that the parent source server startup command would have done an increment of the ftok counter semaphore
* for the replication instance file. But the source server process (the child) that comes here would not have done
* that. Do that while the parent is still holding on to the ftok semaphore waiting for our okay.
*/
if (!ftok_sem_incrcnt(jnlpool.jnlpool_dummy_reg))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_JNLPOOLSETUP);
/* Increment the source server count semaphore */
status = incr_sem(SOURCE, SRC_SERV_COUNT_SEM);
if (0 != status)
{
save_errno = errno;
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Counter semaphore increment failure in child source server"), save_errno);
}
# else
if (0 != (save_errno = rel_sem_immediate(SOURCE, JNL_POOL_ACCESS_SEM)))
{
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Error in rel_sem_immediate"), save_errno);
}
# endif /* REPL_DEBUG_NOBACKGROUND */
gtmsource_srv_count++;
gtmsource_local->child_server_running = TRUE; /* At this point, the parent startup command will stop waiting for child */
gtm_event_log_init();
/* Log source server startup command line first */
SPRINTF(tmpmsg, "%s %s\n", cli_lex_in_ptr->argv[0], cli_lex_in_ptr->in_str);
repl_log(gtmsource_log_fp, TRUE, TRUE, tmpmsg);
SPRINTF(tmpmsg, "GTM Replication Source Server with Pid [%d] started for Secondary Instance [%s]",
process_id, gtmsource_local->secondary_instname);
sgtm_putmsg(print_msg, VARLSTCNT(4) ERR_REPLINFO, 2, LEN_AND_STR(tmpmsg));
repl_log(gtmsource_log_fp, TRUE, TRUE, print_msg);
if (is_jnlpool_creator)
{
repl_log(gtmsource_log_fp, TRUE, TRUE, "Created jnlpool with shmid = [%d] and semid = [%d]\n",
jnlpool.repl_inst_filehdr->jnlpool_shmid, jnlpool.repl_inst_filehdr->jnlpool_semid);
} else
repl_log(gtmsource_log_fp, TRUE, TRUE, "Attached to existing jnlpool with shmid = [%d] and semid = [%d]\n",
jnlpool.repl_inst_filehdr->jnlpool_shmid, jnlpool.repl_inst_filehdr->jnlpool_semid);
gtm_event_log(GTM_EVENT_LOG_ARGC, "MUPIP", "REPLINFO", print_msg);
# ifdef GTM_TLS
if (REPL_TLS_REQUESTED)
{
repl_do_tls_init(gtmsource_log_fp);
assert(REPL_TLS_REQUESTED || PLAINTEXT_FALLBACK);
}
# endif
if (jnlpool.jnlpool_ctl->freeze)
{
last_seen_freeze_flag = jnlpool.jnlpool_ctl->freeze;
sgtm_putmsg(print_msg, VARLSTCNT(3) ERR_REPLINSTFROZEN, 1, jnlpool.repl_inst_filehdr->inst_info.this_instname);
repl_log(gtmsource_log_fp, TRUE, FALSE, print_msg);
sgtm_putmsg(print_msg, VARLSTCNT(3) ERR_REPLINSTFREEZECOMMENT, 1, jnlpool.jnlpool_ctl->freeze_comment);
repl_log(gtmsource_log_fp, TRUE, TRUE, print_msg);
}
gtmsource_local->jnlfileonly = gtmsource_options.jnlfileonly;
do
{ /* If mode is passive, go to sleep. Wakeup every now and then and check to see if I have to become active. */
gtmsource_state = gtmsource_local->gtmsource_state = GTMSOURCE_START;
if ((gtmsource_local->mode == GTMSOURCE_MODE_PASSIVE) && (gtmsource_local->shutdown == NO_SHUTDOWN))
{
gtmsource_poll_actions(FALSE);
SHORT_SLEEP(GTMSOURCE_WAIT_FOR_MODE_CHANGE);
continue;
}
if (GTMSOURCE_MODE_PASSIVE == gtmsource_local->mode)
{ /* Shutdown initiated */
assert(gtmsource_local->shutdown == SHUTDOWN);
sgtm_putmsg(print_msg, VARLSTCNT(4) ERR_REPLINFO, 2,
RTS_ERROR_LITERAL("GTM Replication Source Server Shutdown signalled"));
repl_log(gtmsource_log_fp, TRUE, TRUE, print_msg);
gtm_event_log(GTM_EVENT_LOG_ARGC, "MUPIP", "REPLINFO", print_msg);
break;
}
gtmsource_poll_actions(FALSE);
if (GTMSOURCE_CHANGING_MODE == gtmsource_state)
continue;
if (GTMSOURCE_MODE_ACTIVE_REQUESTED == gtmsource_local->mode)
gtmsource_local->mode = GTMSOURCE_MODE_ACTIVE;
SPRINTF(tmpmsg, "GTM Replication Source Server now in ACTIVE mode using port %d", gtmsource_local->secondary_port);
sgtm_putmsg(print_msg, VARLSTCNT(4) ERR_REPLINFO, 2, LEN_AND_STR(tmpmsg));
repl_log(gtmsource_log_fp, TRUE, TRUE, print_msg);
gtm_event_log(GTM_EVENT_LOG_ARGC, "MUPIP", "REPLINFO", print_msg);
DEBUG_ONLY(repl_csa = &FILE_INFO(jnlpool.jnlpool_dummy_reg)->s_addrs;)
assert(!repl_csa->hold_onto_crit); /* so it is ok to invoke "grab_lock" and "rel_lock" unconditionally */
grab_lock(jnlpool.jnlpool_dummy_reg, TRUE, HANDLE_CONCUR_ONLINE_ROLLBACK);
if (GTMSOURCE_HANDLE_ONLN_RLBK == gtmsource_state)
{
repl_log(gtmsource_log_fp, TRUE, TRUE, "Starting afresh due to ONLINE ROLLBACK\n");
repl_log(gtmsource_log_fp, TRUE, TRUE, "REPL INFO - Current Jnlpool Seqno : %llu\n",
jnlpool.jnlpool_ctl->jnl_seqno);
continue;
}
QWASSIGN(gtmsource_local->read_addr, jnlpool.jnlpool_ctl->write_addr);
gtmsource_local->read = jnlpool.jnlpool_ctl->write;
gtmsource_local->read_state = gtmsource_local->jnlfileonly ? READ_FILE : READ_POOL;
read_jnl_seqno = gtmsource_local->read_jnl_seqno;
assert(read_jnl_seqno <= jnlpool.jnlpool_ctl->jnl_seqno);
if (read_jnl_seqno < jnlpool.jnlpool_ctl->jnl_seqno)
{
gtmsource_local->read_state = READ_FILE;
QWASSIGN(gtmsource_save_read_jnl_seqno, jnlpool.jnlpool_ctl->jnl_seqno);
gtmsource_pool2file_transition = TRUE; /* so that we read the latest gener jnl files */
}
rel_lock(jnlpool.jnlpool_dummy_reg);
if (SS_NORMAL != (status = gtmsource_alloc_tcombuff()))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Error allocating initial tcom buffer space. Malloc error"), status);
gtmsource_filter = NO_FILTER;
if ('\0' != gtmsource_local->filter_cmd[0])
{
if (SS_NORMAL == (status = repl_filter_init(gtmsource_local->filter_cmd)))
gtmsource_filter |= EXTERNAL_FILTER;
else
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
gtmsource_process();
/* gtmsource_process returns only when mode needs to be changed to PASSIVE */
assert(gtmsource_state == GTMSOURCE_CHANGING_MODE);
gtmsource_ctl_close();
gtmsource_free_msgbuff();
gtmsource_free_tcombuff();
gtmsource_free_filter_buff();
gtmsource_stop_heartbeat();
if (FD_INVALID != gtmsource_sock_fd)
repl_close(>msource_sock_fd);
if (gtmsource_filter & EXTERNAL_FILTER)
repl_stop_filter();
} while (TRUE);
/* Note about usage of this function : Create dummy gd_region, gd_segment, file_control,
* unix_db_info, sgmnt_addrs, and allocate mutex_struct (and NUM_CRIT_ENTRY * mutex_que_entry),
* mutex_spin_parms_struct, and node_local in shared memory. Initialize the fields as in
* jnlpool_init(). Pass the address of the dummy region as argument to this function.
*/
boolean_t grab_lock(gd_region *reg, boolean_t is_blocking_wait, uint4 onln_rlbk_action)
{
unix_db_info *udi;
sgmnt_addrs *csa;
enum cdb_sc status;
mutex_spin_parms_ptr_t mutex_spin_parms;
char scndry_msg[OUT_BUFF_SIZE];
# ifdef DEBUG
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
# endif
udi = FILE_INFO(reg);
csa = &udi->s_addrs;
assert(!csa->hold_onto_crit);
assert(!csa->now_crit);
if (!csa->now_crit)
{
assert(0 == crit_count);
crit_count++; /* prevent interrupts */
DEBUG_ONLY(locknl = csa->nl); /* for DEBUG_ONLY LOCK_HIST macro */
mutex_spin_parms = (mutex_spin_parms_ptr_t)((sm_uc_ptr_t)csa->critical + JNLPOOL_CRIT_SPACE);
/* This assumes that mutex_spin_parms_t is located immediately after the crit structures */
/* As of 10/07/98, crashcnt field in mutex_struct is not changed by any function for the dummy region */
if (is_blocking_wait)
status = mutex_lockw(reg, mutex_spin_parms, 0);
else
status = mutex_lockwim(reg, mutex_spin_parms, 0);
DEBUG_ONLY(locknl = NULL); /* restore "locknl" to default value */
if (status != cdb_sc_normal)
{
crit_count = 0;
switch(status)
{
case cdb_sc_critreset: /* As of 10/07/98, this return value is not possible */
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_CRITRESET, 2, REG_LEN_STR(reg));
case cdb_sc_dbccerr:
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_DBCCERR, 2, REG_LEN_STR(reg));
case cdb_sc_nolock:
return FALSE;
default:
assertpro(FALSE && status);
}
return FALSE;
}
/* There is only one case we know of when csa->nl->in_crit can be non-zero and that is when a process holding
* crit gets kill -9ed and another process ends up invoking "secshr_db_clnup" which in turn clears the
* crit semaphore (making it available for waiters) but does not also clear csa->nl->in_crit since it does not
* hold crit at that point. But in that case, the pid reported in csa->nl->in_crit should be dead. Check that.
*/
assert((0 == csa->nl->in_crit) || (FALSE == is_proc_alive(csa->nl->in_crit, 0)));
csa->nl->in_crit = process_id;
CRIT_TRACE(crit_ops_gw); /* see gdsbt.h for comment on placement */
crit_count = 0;
if (jnlpool.repl_inst_filehdr->file_corrupt && !jgbl.onlnrlbk)
{ /* Journal pool indicates an abnormally terminated online rollback. Cannot continue until the rollback
* command is re-run to bring the journal pool/file and instance file to a consistent state.
*/
SNPRINTF(scndry_msg, OUT_BUFF_SIZE, "Instance file header has file_corrupt field set to TRUE");
/* No need to do rel_lock before rts_error (mupip_exit_handler will do it for us) - BYPASSOK rts_error */
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_REPLREQROLLBACK, 2, LEN_AND_STR(udi->fn),
ERR_TEXT, 2, LEN_AND_STR(scndry_msg));
}
/* If ASSERT_NO_ONLINE_ROLLBACK, then no concurrent online rollbacks can happen at this point. So, the jnlpool
* should be in in sync. There are two exceptions. If this is GT.CM GNP Server and the last client disconnected, the
* server invokes gtcmd_rundown which in-turn invokes gds_rundown thereby running down all active databases at this
* point but leaves the journal pool up and running. Now, if an online rollback is attempted, it increments the
* onln_rlbk_cycle in the journal pool, but csa->onln_rlbk_cycle is not synced yet. So, the grab_crit done in t_end
* will NOT detect a concurrent online rollback and it doesn't need to because the rollback happened AFTER the
* rundown. Assert that this is the only case we know of for the cycles to be out-of-sync. In PRO
* jnlpool_ctl->onln_rlbk_cycle is used only by the replication servers (which GT.CM is not) and so even if it
* continues with an out-of-sync csa->onln_rlbk_cycle, t_end logic does the right thing. The other exception is if
* GT.M initialized journal pool while opening database (belonging to a different instance) in gvcst_init (for
* anticipatory freeze) followed by an online rollback which increments the jnlpool_ctl->onln_rlbk_cycle but leaves
* the repl_csa->onln_rlbk_cycle out-of-sync. At this point, if a replicated database is open for the first time,
* we'll reach t_end to commit the update but will end up failing the below assert due to the out-of-sync
* onln_rlbk_cycle. So, assert accordingly. Note : even though the cycles are out-of-sync they are not an issue for
* GT.M because it always relies on the onln_rlbk_cycle from csa->nl and not from repl_csa. But, we don't remove the
* assert as it is valuable for replication servers (Source, Receiver and Update Process).
*/
assert((ASSERT_NO_ONLINE_ROLLBACK != onln_rlbk_action)
|| (csa->onln_rlbk_cycle == jnlpool.jnlpool_ctl->onln_rlbk_cycle) || IS_GTCM_GNP_SERVER_IMAGE
|| (jnlpool_init_needed && INST_FREEZE_ON_ERROR_POLICY));
if ((HANDLE_CONCUR_ONLINE_ROLLBACK == onln_rlbk_action)
&& (csa->onln_rlbk_cycle != jnlpool.jnlpool_ctl->onln_rlbk_cycle))
{
assert(is_src_server);
SYNC_ONLN_RLBK_CYCLES;
gtmsource_onln_rlbk_clnup(); /* side-effect : sets gtmsource_state */
rel_lock(reg); /* caller knows to disconnect and re-establish the connection */
}
}
return TRUE;
}
