/*
* Lease the property causing others to wait before leasing or purchasing.
*/
FusionResult
fusion_property_lease (FusionProperty *property)
{
FusionResult ret = FUSION_SUCCESS;
struct sembuf op[3];
/* lock */
op[0].sem_num = 0;
op[0].sem_op = -1;
op[0].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 1);
while (property->state == FUSION_PROPERTY_LEASED) {
/* increment wait counter */
op[0].sem_num = 1;
op[0].sem_op = 1;
op[0].sem_flg = SEM_UNDO;
/* unlock */
op[1].sem_num = 0;
op[1].sem_op = 1;
op[1].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 2);
/* lock */
op[0].sem_num = 0;
op[0].sem_op = -1;
op[0].sem_flg = SEM_UNDO;
/* wait */
op[1].sem_num = 2;
op[1].sem_op = -1;
op[1].sem_flg = 0;
/* decrement wait counter */
op[2].sem_num = 1;
op[2].sem_op = -1;
op[2].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 3);
}
if (property->state == FUSION_PROPERTY_AVAILABLE)
property->state = FUSION_PROPERTY_LEASED;
else
ret = FUSION_INUSE;
/* unlock */
op[0].sem_num = 0;
op[0].sem_op = 1;
op[0].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 1);
return ret;
}
/*
* Cede the property allowing others to lease or purchase it.
*/
FusionResult
fusion_property_cede (FusionProperty *property)
{
struct sembuf op[2];
union semun semopts;
int waiters;
/* lock */
op[0].sem_num = 0;
op[0].sem_op = -1;
op[0].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 1);
/* debug check */
if (property->state == FUSION_PROPERTY_AVAILABLE)
FDEBUG("BUG! property not leased/purchased!");
/* make available */
property->state = FUSION_PROPERTY_AVAILABLE;
/* get number of waiters */
waiters = semctl (property->sem_id, 1, GETVAL, semopts);
if (waiters < 0) {
FPERROR ("semctl");
/* unlock */
op[0].sem_num = 0;
op[0].sem_op = 1;
op[0].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 1);
return FUSION_FAILURE;
}
/* unlock */
op[0].sem_num = 0;
op[0].sem_op = 1;
op[0].sem_flg = SEM_UNDO;
/* awake waiters */
op[1].sem_num = 2;
op[1].sem_op = waiters;
op[1].sem_flg = 0;
SEMOP (property->sem_id, op, waiters ? 2 : 1);
return FUSION_SUCCESS;
}
int grab_sem_all_receive()
{
int rc;
assert(!holds_sem[RECV][RECV_POOL_ACCESS_SEM]);
sop[0].sem_op = 0; /* Wait for 0 */
sop[0].sem_num = RECV_POOL_ACCESS_SEM;
sop[1].sem_op = 1; /* Increment it */
sop[1].sem_num = RECV_POOL_ACCESS_SEM;
sop[2].sem_op = 1; /* Increment it */
sop[2].sem_num = RECV_SERV_COUNT_SEM;
sop[3].sem_op = 1; /* Increment it */
sop[3].sem_num = UPD_PROC_COUNT_SEM;
sop[4].sem_op = 1; /* Increment it */
sop[4].sem_num = RECV_SERV_OPTIONS_SEM;
sop[0].sem_flg = sop[1].sem_flg = sop[2].sem_flg = sop[3].sem_flg = sop[4].sem_flg = SEM_UNDO;
SEMOP(sem_set_id[RECV], sop, 5, rc, FORCED_WAIT);
if (0 == rc)
{
holds_sem[RECV][RECV_POOL_ACCESS_SEM] = TRUE;
holds_sem[RECV][RECV_SERV_COUNT_SEM] = TRUE;
holds_sem[RECV][UPD_PROC_COUNT_SEM] = TRUE;
holds_sem[RECV][RECV_SERV_OPTIONS_SEM] = TRUE;
}
return rc;
}
int incr_sem(int set_index, int sem_num)
{
int rc;
ASSERT_SET_INDEX;
sop[0].sem_op = 1; /* Increment it */
sop[0].sem_num = sem_num;
sop[0].sem_flg = SEM_UNDO;
SEMOP(sem_set_id[set_index], sop, 1, rc);
return rc;
}
int grab_sem_immediate(int set_index, int sem_num)
{
int rc;
ASSERT_SET_INDEX;
assert(!holds_sem[set_index][sem_num]);
sop[0].sem_op = 0; /* Wait for 0 */
sop[0].sem_num = sem_num;
sop[1].sem_op = 1; /* Increment it */
sop[1].sem_num = sem_num;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO | IPC_NOWAIT;
SEMOP(sem_set_id[set_index], sop, 2, rc, NO_WAIT);
if (0 == rc)
holds_sem[set_index][sem_num] = TRUE;
return rc;
}
int grab_sem(int set_index, int sem_num)
{
int rc;
assert((int)NUM_SRC_SEMS == (int)NUM_RECV_SEMS); /* holds_sem[][] relies on this as it uses NUM_SRC_SEMS for array bounds */
assert(!holds_sem[set_index][sem_num]);
ASSERT_SET_INDEX;
sop[0].sem_op = 0; /* Wait for 0 */
sop[0].sem_num = sem_num;
sop[1].sem_op = 1; /* lock it */
sop[1].sem_num = sem_num;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO;
SEMOP(sem_set_id[set_index], sop, 2, rc, FORCED_WAIT);
if (0 == rc)
holds_sem[set_index][sem_num] = TRUE;
return rc;
}
int grab_sem(int set_index, int sem_num)
{
int rc;
assert(NUM_SRC_SEMS == NUM_RECV_SEMS); /* holds_sem[][] relies on this as it uses NUM_SRC_SEMS in the array definition */
assert(!holds_sem[set_index][sem_num]);
ASSERT_SET_INDEX;
sop[0].sem_op = 0; /* Wait for 0 */
sop[0].sem_num = sem_num;
sop[1].sem_op = 1; /* Increment it */
sop[1].sem_num = sem_num;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO;
SEMOP(sem_set_id[set_index], sop, 2, rc);
if (0 == rc)
holds_sem[set_index][sem_num] = TRUE;
return rc;
}
int incr_sem(int set_index, int sem_num)
{
int rc;
ASSERT_SET_INDEX;
sop[0].sem_op = 1; /* Increment it */
sop[0].sem_num = sem_num;
sop[0].sem_flg = SEM_UNDO;
SEMOP(sem_set_id[set_index], sop, 1, rc, NO_WAIT);
if (0 == rc)
{
/* decr_sem internally calls rel_sem directly which expects that the entry for hold_sem[SOURCE][SRC_SERV_COUNT_SEM]
* is set to TRUE. But, incr_sem doesn't set this entry. This causes decr_sem (done below) to assert fail. to avoid
* the assert, set the hold_sem array to TRUE even if this is an increment operation
*/
holds_sem[set_index][sem_num] = TRUE;
}
return rc;
}
int grab_sem_all_source()
{
int rc;
assert(!holds_sem[SOURCE][JNL_POOL_ACCESS_SEM]);
sop[0].sem_op = 0; /* Wait for 0 */
sop[0].sem_num = JNL_POOL_ACCESS_SEM;
sop[1].sem_op = 1; /* Increment it */
sop[1].sem_num = JNL_POOL_ACCESS_SEM;
sop[2].sem_op = 1; /* Increment it */
sop[2].sem_num = SRC_SERV_COUNT_SEM;
sop[0].sem_flg = sop[1].sem_flg = sop[2].sem_flg = SEM_UNDO;
SEMOP(sem_set_id[SOURCE], sop, 3, rc, FORCED_WAIT);
if (0 == rc)
{
holds_sem[SOURCE][JNL_POOL_ACCESS_SEM] = TRUE;
holds_sem[SOURCE][SRC_SERV_COUNT_SEM] = TRUE;
}
return rc;
}
void semInit() {
union semun arg;
int semkey = getSemKey();
unsigned short initial_values[] = { 0, 0, 0 };
struct sembuf signal_shm = SEMOP(OP_SIGNAL, SHM_MUTEX);
arg.array = initial_values;
do {
sem.id = semget(semkey, SEM_NUM, IPC_CREAT | IPC_EXCL | 0666);
if (sem.id != -1) {
/* Só entra aqui o que conseguiu criar os semáforos. */
printf("[INFO] Semáforos inicializados com chave 0x%x\n", semkey);
semctl(sem.id, 0, SETALL, arg);
semop(sem.id, &signal_shm, 1);
break;
}
else sem.id = semget(semkey, SEM_NUM, 0666);
} while ((sem.id == -1 && errno == ENOENT) || !waitFirstOp());
sem.nops = 0;
}
void gds_rundown(void)
{
bool is_mm, we_are_last_user, we_are_last_writer;
boolean_t ipc_deleted, remove_shm, cancelled_timer, cancelled_dbsync_timer, vermismatch;
now_t now; /* for GET_CUR_TIME macro */
char *time_ptr, time_str[CTIME_BEFORE_NL + 2]; /* for GET_CUR_TIME macro */
gd_region *reg;
int save_errno, status;
int4 semval, ftok_semval, sopcnt, ftok_sopcnt;
short crash_count;
sm_long_t munmap_len;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
struct shmid_ds shm_buf;
struct sembuf sop[2], ftok_sop[2];
uint4 jnl_status;
unix_db_info *udi;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
error_def(ERR_CRITSEMFAIL);
error_def(ERR_DBCCERR);
error_def(ERR_DBFILERR);
error_def(ERR_DBRNDWNWRN);
error_def(ERR_ERRCALL);
error_def(ERR_GBLOFLOW);
error_def(ERR_GTMASSERT);
error_def(ERR_IPCNOTDEL);
error_def(ERR_JNLFLUSH);
error_def(ERR_RNDWNSEMFAIL);
error_def(ERR_TEXT);
error_def(ERR_WCBLOCKED);
forced_exit = FALSE; /* Okay, we're dying already -- let rel_crit live in peace now.
* If coming through a DAL, not necessarily dying. what to do then? -- nars -- 8/15/2001
*/
grabbed_access_sem = FALSE;
jnl_status = 0;
reg = gv_cur_region; /* Local copy */
/*
* early out for cluster regions
* to avoid tripping the assert below.
* Note:
* This early out is consistent with VMS. It has been
* noted that all of the gtcm assignments
* to gv_cur_region should use the TP_CHANGE_REG
* macro. This would also avoid the assert problem
* and should be done eventually.
*/
if (dba_cm == reg->dyn.addr->acc_meth)
return;
udi = FILE_INFO(reg);
csa = &udi->s_addrs;
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
if ((reg->open) && (dba_usr == csd->acc_meth))
{
change_reg();
gvusr_rundown();
return;
}
ESTABLISH(gds_rundown_ch);
if (!reg->open) /* Not open, no point to rundown */
{
if (reg->opening) /* Died partway open, kill rest of way */
{
rel_crit(reg);
mutex_cleanup(reg);
/* revist this to handle MM properly SMW 98/12/16
if (NULL != csa->nl)
{
status = shmdt((caddr_t)csa->nl);
if (-1 == status)
send_msg(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during shmdt"), errno);
}
*/
shmdt((caddr_t)csa->nl);
csa->nl = NULL;
}
REVERT;
return;
}
switch(csd->acc_meth)
{ /* Pass mm and bg through */
case dba_bg:
is_mm = FALSE;
break;
case dba_mm:
is_mm = TRUE;
break;
case dba_usr:
assert(FALSE);
default:
REVERT;
return;
}
/* Cancel any pending flush timer for this region by this task */
CANCEL_DB_TIMERS(reg, cancelled_timer, cancelled_dbsync_timer);
we_are_last_user = FALSE;
if (!csa->persistent_freeze)
region_freeze(reg, FALSE, FALSE, FALSE);
assert(!csa->read_lock);
rel_crit(reg); /* get locks to known state */
mutex_cleanup(reg);
/*
* We need to guarantee that none else access database file header when semid/shmid fields are reset.
* We already have created ftok semaphore in db_init or, mu_rndwn_file and did not remove it.
* So just lock it. We do it in blocking mode.
*/
if (!ftok_sem_lock(reg, FALSE, FALSE))
rts_error(VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
/*
* For mupip_jnl_recover we already have database access control semaphore.
* We do not release it. We release it from mur_close_files.
*/
if (!mupip_jnl_recover)
{
sop[0].sem_num = 0; sop[0].sem_op = 0; /* Wait for 0 */
sop[1].sem_num = 0; sop[1].sem_op = 1; /* Lock */
sopcnt = 2;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO | IPC_NOWAIT; /* Don't wait the first time thru */
SEMOP(udi->semid, sop, sopcnt, status);
if (-1 == status) /* We couldn't get it in one shot -- see if we already have it */
{
save_errno = errno;
/* see comment about Linux specific difference in behaviour of semctl() with GETPID in gds_rundown_ch() */
if (semctl(udi->semid, 0, GETPID) == process_id)
{
send_msg(VARLSTCNT(5) MAKE_MSG_INFO(ERR_CRITSEMFAIL), 2,
DB_LEN_STR(reg),
ERR_RNDWNSEMFAIL);
REVERT;
return; /* Already in rundown for this region */
}
if (EAGAIN != save_errno)
{
assert(FALSE);
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown first semop/semctl"), save_errno);
}
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO; /* Try again - blocking this time */
SEMOP(udi->semid, sop, 2, status);
if (-1 == status) /* We couldn't get it at all.. */
rts_error(VARLSTCNT(5) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), errno);
}
}
grabbed_access_sem = TRUE;
/*
* We now have the dbinit/rundown lock, so we are alone in this code for this region
* and nobody else can attach.
* See if we are all alone in accessing this database shared memory.
*/
assert(csa->ref_cnt); /* decrement private ref_cnt before shared ref_cnt decrement. */
csa->ref_cnt--; /* Currently journaling logic in gds_rundown() in VMS relies on this order to detect last writer */
assert(!csa->ref_cnt);
--csa->nl->ref_cnt;
if (memcmp(csa->nl->now_running, gtm_release_name, gtm_release_name_len + 1))
{ /* VERMISMATCH condition. Possible only if DSE */
assert(dse_running);
vermismatch = TRUE;
} else
vermismatch = FALSE;
if (-1 == shmctl(udi->shmid, IPC_STAT, &shm_buf))
{
save_errno = errno;
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown shmctl"), save_errno);
} else
we_are_last_user = (1 == shm_buf.shm_nattch) && !vermismatch;
assert(!mupip_jnl_recover || we_are_last_user); /* recover => one user */
if (-1 == (semval = semctl(udi->semid, 1, GETVAL)))
rts_error(VARLSTCNT(5) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), errno);
we_are_last_writer = (1 == semval) && (FALSE == reg->read_only) && !vermismatch;/* There's one writer left and I am it */
assert(!(mupip_jnl_recover && !reg->read_only) || we_are_last_writer); /* recover + R/W region => one writer */
if (-1 == (ftok_semval = semctl(udi->ftok_semid, 1, GETVAL)))
rts_error(VARLSTCNT(5) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), errno);
/* If csa->nl->donotflush_dbjnl is set, it means mupip recover/rollback was interrupted and therefore we should
* not flush shared memory contents to disk as they might be in an inconsistent state.
* In this case, we will go ahead and remove shared memory (without flushing the contents) in this routine.
* A reissue of the recover/rollback command will restore the database to a consistent state.
* Otherwise, if we have write access to this region, let us perform a few writing tasks.
*/
if (csa->nl->donotflush_dbjnl)
csa->wbuf_dqd = 0; /* ignore csa->wbuf_dqd status as we do not care about the cache contents */
else if (!reg->read_only && !vermismatch)
{ /* If we had an orphaned block and were interrupted, set wc_blocked so we can invoke wcs_recover */
if (csa->wbuf_dqd)
{
grab_crit(reg);
SET_TRACEABLE_VAR(csd->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wcb_gds_rundown);
send_msg(VARLSTCNT(8) ERR_WCBLOCKED, 6, LEN_AND_LIT("wcb_gds_rundown"),
process_id, &csa->ti->curr_tn, DB_LEN_STR(reg));
csa->wbuf_dqd = 0;
wcs_recover(reg);
if (is_mm)
{
assert(FALSE);
csd = csa->hdr;
}
BG_TRACE_PRO_ANY(csa, lost_block_recovery);
rel_crit(reg);
}
if (JNL_ENABLED(csd) && (GTCM_GNP_SERVER_IMAGE == image_type))
originator_prc_vec = NULL;
/* If we are the last writing user, then everything must be flushed */
if (we_are_last_writer)
{ /* Time to flush out all of our buffers */
if (is_mm)
{
if (csa->total_blks != csa->ti->total_blks) /* do remap if file had been extended */
{
grab_crit(reg);
wcs_mm_recover(reg);
csd = csa->hdr;
rel_crit(reg);
}
csa->nl->remove_shm = TRUE;
}
/* Note WCSFLU_SYNC_EPOCH ensures the epoch is synced to the journal and indirectly
* also ensures that the db is fsynced. We don't want to use it in the calls to
* wcs_flu() from t_end() and tp_tend() since we can defer it to out-of-crit there.
* In this case, since we are running down, we don't have any such option.
*/
csa->nl->remove_shm = wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
/* Since we_are_last_writer, we should be guaranteed that wcs_flu() did not change csd, (in
* case of MM for potential file extension), even if it did a grab_crit(). Therefore, make
* sure that's true.
*/
assert(csd == csa->hdr);
assert(0 == memcmp(csd->label, GDS_LABEL, GDS_LABEL_SZ - 1));
csd->trans_hist.header_open_tn = csd->trans_hist.curr_tn;
} else if ((cancelled_timer && (0 > csa->nl->wcs_timers)) || cancelled_dbsync_timer)
{ /* cancelled pending db or jnl flush timers - flush database and journal buffers to disk */
grab_crit(reg);
/* we need to sync the epoch as the fact that there is no active pending flush timer implies
* there will be noone else who will flush the dirty buffers and EPOCH to disk in a timely fashion
*/
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
rel_crit(reg);
assert((dba_mm == cs_data->acc_meth) || (csd == cs_data));
csd = cs_data; /* In case this is MM and wcs_flu() remapped an extended database, reset csd */
}
/* Do rundown journal processing after buffer flushes since they require jnl to be open */
if (JNL_ENABLED(csd))
{ /* the following tp_change_reg() is not needed due to the assert csa == cs_addrs at the beginning
* of gds_rundown(), but just to be safe. To be removed by 2002!! --- nars -- 2001/04/25.
*/
tp_change_reg(); /* call this because jnl_ensure_open checks cs_addrs rather than gv_cur_region */
jpc = csa->jnl;
jbp = jpc->jnl_buff;
if (jbp->fsync_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->fsync_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (jbp->io_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->io_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (((NOJNL != jpc->channel) && !JNL_FILE_SWITCHED(jpc))
|| we_are_last_writer && (0 != csa->nl->jnl_file.u.inode))
{ /* We need to close the journal file cleanly if we have the latest generation journal file open
* or if we are the last writer and the journal file is open in shared memory (not necessarily
* by ourselves e.g. the only process that opened the journal got shot abnormally)
* Note: we should not infer anything from the shared memory value of csa->nl->jnl_file.u.inode
* if we are not the last writer as it can be concurrently updated.
*/
grab_crit(reg);
if (JNL_ENABLED(csd))
{
SET_GBL_JREC_TIME; /* jnl_ensure_open/jnl_put_jrt_pini/pfin/jnl_file_close all need it */
/* Before writing to jnlfile, adjust jgbl.gbl_jrec_time if needed to maintain time order
* of jnl records. This needs to be done BEFORE the jnl_ensure_open as that could write
* journal records (if it decides to switch to a new journal file).
*/
ADJUST_GBL_JREC_TIME(jgbl, jbp);
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{ /* If we_are_last_writer, we would have already done a wcs_flu() which would
* have written an epoch record and we are guaranteed no further updates
* since we are the last writer. So, just close the journal.
* Although we assert pini_addr should be non-zero for last_writer, we
* play it safe in PRO and write a PINI record if not written already.
*/
assert(!jbp->before_images || is_mm
|| !we_are_last_writer || 0 != jpc->pini_addr);
if (we_are_last_writer && 0 == jpc->pini_addr)
jnl_put_jrt_pini(csa);
if (0 != jpc->pini_addr)
jnl_put_jrt_pfin(csa);
/* If not the last writer and no pending flush timer left, do jnl flush now */
if (!we_are_last_writer && (0 > csa->nl->wcs_timers))
{
if (SS_NORMAL == (jnl_status = jnl_flush(reg)))
{
assert(jbp->freeaddr == jbp->dskaddr);
jnl_fsync(reg, jbp->dskaddr);
assert(jbp->fsync_dskaddr == jbp->dskaddr);
} else
{
send_msg(VARLSTCNT(9) ERR_JNLFLUSH, 2, JNL_LEN_STR(csd),
ERR_TEXT, 2,
RTS_ERROR_TEXT("Error with journal flush in gds_rundown"),
jnl_status);
assert(NOJNL == jpc->channel);/* jnl file lost has been triggered */
/* In this routine, all code that follows from here on does not
* assume anything about the journaling characteristics of this
* database so it is safe to continue execution even though
* journaling got closed in the middle.
*/
}
}
jnl_file_close(reg, we_are_last_writer, FALSE);
} else
send_msg(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(reg));
}
rel_crit(reg);
}
}
if (we_are_last_writer) /* Flush the fileheader last and harden the file to disk */
{
grab_crit(reg); /* To satisfy crit requirement in fileheader_sync() */
memset(csd->machine_name, 0, MAX_MCNAMELEN); /* clear the machine_name field */
if (!mupip_jnl_recover && we_are_last_user)
{ /* mupip_jnl_recover will do this after mur_close_file */
csd->semid = INVALID_SEMID;
csd->shmid = INVALID_SHMID;
csd->gt_sem_ctime.ctime = 0;
csd->gt_shm_ctime.ctime = 0;
}
fileheader_sync(reg);
rel_crit(reg);
if (FALSE == is_mm)
{
if (-1 == fsync(udi->fd)) /* Sync it all */
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
} else
{ /* Now do final MM file sync before exit */
#if !defined(TARGETED_MSYNC) && !defined(NO_MSYNC)
if (-1 == fsync(udi->fd)) /* Sync it all */
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
#else
if (-1 == msync((caddr_t)csa->db_addrs[0], (size_t)(csa->db_addrs[1] - csa->db_addrs[0]), MS_SYNC))
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file msync at close"), errno);
}
#endif
}
}
} /* end if (!reg->read_only && !csa->nl->donotflush_dbjnl) */
if (reg->read_only && we_are_last_user && !mupip_jnl_recover)
{ /* mupip_jnl_recover will do this after mur_close_file */
db_ipcs.semid = INVALID_SEMID;
db_ipcs.shmid = INVALID_SHMID;
db_ipcs.gt_sem_ctime = 0;
db_ipcs.gt_shm_ctime = 0;
db_ipcs.fn_len = reg->dyn.addr->fname_len;
memcpy(db_ipcs.fn, reg->dyn.addr->fname, reg->dyn.addr->fname_len);
db_ipcs.fn[reg->dyn.addr->fname_len] = 0;
/* request gtmsecshr to flush. read_only cannot flush itself */
if (0 != send_mesg2gtmsecshr(FLUSH_DB_IPCS_INFO, 0, (char *)NULL, 0))
rts_error(VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gtmsecshr failed to update database file header"));
}
/* Done with file now, close it */
if (-1 == close(udi->fd))
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during file close"), errno);
}
/* Unmap storage if mm mode but only the part that is not the fileheader (so shows up in dumps) */
if (is_mm)
{
munmap_len = (sm_long_t)((csa->db_addrs[1] - csa->db_addrs[0]) - ROUND_UP(SIZEOF_FILE_HDR(csa->hdr),
MSYNC_ADDR_INCS));
if (munmap_len > 0)
{
munmap((caddr_t)(csa->db_addrs[0] + ROUND_UP(SIZEOF_FILE_HDR(csa->hdr), MSYNC_ADDR_INCS)),
(size_t)(munmap_len));
#ifdef DEBUG_DB64
rel_mmseg((caddr_t)csa->db_addrs[0]);
#endif
}
}
/* Detach our shared memory while still under lock so reference counts will be
* correct for the next process to run down this region.
* In the process also get the remove_shm status from node_local before detaching.
* If csa->nl->donotflush_dbjnl is TRUE, it means we can safely remove shared memory without compromising data
* integrity as a reissue of recover will restore the database to a consistent state.
*/
remove_shm = !vermismatch && (csa->nl->remove_shm || csa->nl->donotflush_dbjnl);
status = shmdt((caddr_t)csa->nl);
csa->nl = NULL; /* dereferencing nl after detach is not right, so we set it to NULL so that we can test before dereference*/
if (-1 == status)
send_msg(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg), ERR_TEXT, 2, LEN_AND_LIT("Error during shmdt"), errno);
reg->open = FALSE;
/* If file is still not in good shape, die here and now before we get rid of our storage */
if (csa->wbuf_dqd)
GTMASSERT;
ipc_deleted = FALSE;
/* If we are the very last user, remove shared storage id and the semaphores */
if (we_are_last_user)
{ /* remove shared storage, only if last writer to rundown did a successful wcs_flu() */
assert(!vermismatch);
if (remove_shm)
{
ipc_deleted = TRUE;
if (0 != shm_rmid(udi->shmid))
rts_error(VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove shared memory"));
} else if (is_src_server || is_updproc)
{
gtm_putmsg(VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
send_msg(VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
send_msg(VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
/*
* Don't release semaphore in case of mupip recover/rollback; since it has standalone access.
* It will release the semaphore in mur_close_files.
*/
if (!mupip_jnl_recover)
{
if (0 != sem_rmid(udi->semid))
rts_error(VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove semaphore"));
grabbed_access_sem = FALSE;
}
} else
{
assert(!mupip_jnl_recover);
/* If we were writing, get rid of our writer access count semaphore */
if (!reg->read_only)
if (0 != (save_errno = do_semop(udi->semid, 1, -1, SEM_UNDO)))
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown write semaphore release"), save_errno);
/* Now remove the rundown lock */
if (0 != (save_errno = do_semop(udi->semid, 0, -1, SEM_UNDO)))
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown rundown semaphore release"), save_errno);
grabbed_access_sem = FALSE;
}
if (!ftok_sem_release(reg, !mupip_jnl_recover, FALSE))
rts_error(VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
if (!ipc_deleted)
{
GET_CUR_TIME;
if (is_src_server)
gtm_putmsg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Source server"), REG_LEN_STR(reg));
if (is_updproc)
gtm_putmsg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Update process"), REG_LEN_STR(reg));
if (mupip_jnl_recover)
{
gtm_putmsg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
send_msg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
}
}
REVERT;
}
void semSync(int semaph) {
struct sembuf sync = SEMOP(OP_SYNC, 0);
sync.sem_num = semaph;
semop(sem.id, &sync, 1);
}
void semSignal(int semaph) {
struct sembuf signal = SEMOP(OP_SIGNAL, 0);
signal.sem_num = semaph;
semop(sem.id, &signal, 1);
}
void semSafeWait(int semaph) {
struct sembuf wait = SEMOP(OP_WAIT, 0);
wait.sem_num = semaph;
semSafeOp(&wait, 1);
}
void semWait(int semaph) {
struct sembuf wait = SEMOP(OP_WAIT, 0);
wait.sem_num = semaph;
semop(sem.id, &wait, 1);
}
int4 gds_rundown(void)
{
boolean_t canceled_dbsync_timer, canceled_flush_timer, ok_to_write_pfin;
boolean_t have_standalone_access, ipc_deleted, err_caught;
boolean_t is_cur_process_ss_initiator, remove_shm, vermismatch, we_are_last_user, we_are_last_writer, is_mm;
boolean_t unsafe_last_writer;
char time_str[CTIME_BEFORE_NL + 2]; /* for GET_CUR_TIME macro */
gd_region *reg;
int save_errno, status, rc;
int4 semval, ftok_semval, sopcnt, ftok_sopcnt;
short crash_count;
sm_long_t munmap_len;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
node_local_ptr_t cnl;
struct shmid_ds shm_buf;
struct sembuf sop[2], ftok_sop[2];
uint4 jnl_status;
unix_db_info *udi;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
shm_snapshot_t *ss_shm_ptr;
uint4 ss_pid, onln_rlbk_pid, holder_pid;
boolean_t was_crit;
boolean_t safe_mode; /* Do not flush or take down shared memory. */
boolean_t bypassed_ftok = FALSE, bypassed_access = FALSE, may_bypass_ftok, inst_is_frozen,
ftok_counter_halted,
access_counter_halted;
int secshrstat;
intrpt_state_t prev_intrpt_state;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
jnl_status = 0;
reg = gv_cur_region; /* Local copy */
/* early out for cluster regions
* to avoid tripping the assert below.
* Note:
* This early out is consistent with VMS. It has been
* noted that all of the gtcm assignments
* to gv_cur_region should use the TP_CHANGE_REG
* macro. This would also avoid the assert problem
* and should be done eventually.
*/
if (dba_cm == reg->dyn.addr->acc_meth)
return EXIT_NRM;
udi = FILE_INFO(reg);
csa = &udi->s_addrs;
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
if ((reg->open) && (dba_usr == csd->acc_meth))
{
change_reg();
gvusr_rundown();
return EXIT_NRM;
}
/* If the process has standalone access, it has udi->grabbed_access_sem set to TRUE at this point. Note that down in a local
* variable as the udi->grabbed_access_sem is set to TRUE even for non-standalone access below and hence we can't rely on
* that later to determine if the process had standalone access or not when it entered this function. We need to guarantee
* that none else access database file header when semid/shmid fields are reset. We already have created ftok semaphore in
* db_init or, mu_rndwn_file and did not remove it. So just lock it. We do it in blocking mode.
*/
have_standalone_access = udi->grabbed_access_sem; /* process holds standalone access */
DEFER_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
ESTABLISH_NORET(gds_rundown_ch, err_caught);
if (err_caught)
{
REVERT;
WITH_CH(gds_rundown_ch, gds_rundown_err_cleanup(have_standalone_access), 0);
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
DEBUG_ONLY(ok_to_UNWIND_in_exit_handling = FALSE);
return EXIT_ERR;
}
assert(reg->open); /* if we failed to open, dbinit_ch should have taken care of proper clean up */
assert(!reg->opening); /* see comment above */
assert((dba_bg == csd->acc_meth) || (dba_mm == csd->acc_meth));
is_mm = (dba_bg != csd->acc_meth);
assert(!csa->hold_onto_crit || (csa->now_crit && jgbl.onlnrlbk));
/* If we are online rollback, we should already be holding crit and should release it only at the end of this module. This
* is usually done by noting down csa->now_crit in a local variable (was_crit) and using it whenever we are about to
* grab_crit. But, there are instances (like mupip_set_journal.c) where we grab_crit but invoke gds_rundown without any
* preceeding rel_crit. Such code relies on the fact that gds_rundown does rel_crit unconditionally (to get locks to a known
* state). So, augment csa->now_crit with jgbl.onlnrlbk to track if we can rel_crit unconditionally or not in gds_rundown.
*/
was_crit = (csa->now_crit && jgbl.onlnrlbk);
/* Cancel any pending flush timer for this region by this task */
canceled_flush_timer = FALSE;
canceled_dbsync_timer = FALSE;
CANCEL_DB_TIMERS(reg, csa, canceled_flush_timer, canceled_dbsync_timer);
we_are_last_user = FALSE;
inst_is_frozen = IS_REPL_INST_FROZEN && REPL_ALLOWED(csa->hdr);
if (!csa->persistent_freeze)
region_freeze(reg, FALSE, FALSE, FALSE);
if (!was_crit)
{
rel_crit(reg); /* get locks to known state */
mutex_cleanup(reg);
}
/* The only process that can invoke gds_rundown while holding access control semaphore is RECOVER/ROLLBACK. All the others
* (like MUPIP SET -FILE/MUPIP EXTEND would have invoked db_ipcs_reset() before invoking gds_rundown (from
* mupip_exit_handler). The only exception is when these processes encounter a terminate signal and they reach
* mupip_exit_handler while holding access control semaphore. Assert accordingly.
*/
assert(!have_standalone_access || mupip_jnl_recover || process_exiting);
/* If we have standalone access, then ensure that a concurrent online rollback cannot be running at the same time as it
* needs the access control lock as well. The only expection is we are online rollback and currently running down.
*/
cnl = csa->nl;
onln_rlbk_pid = cnl->onln_rlbk_pid;
assert(!have_standalone_access || mupip_jnl_recover || !onln_rlbk_pid || !is_proc_alive(onln_rlbk_pid, 0));
if (!have_standalone_access)
{
if (-1 == (ftok_semval = semctl(udi->ftok_semid, DB_COUNTER_SEM, GETVAL))) /* Check # of procs counted on FTOK */
{
save_errno = errno;
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get ftok_semval"), CALLFROM, errno);
}
may_bypass_ftok = CAN_BYPASS(ftok_semval, csd, inst_is_frozen); /* Do we need a blocking wait? */
/* We need to guarantee that no one else access database file header when semid/shmid fields are reset.
* We already have created ftok semaphore in db_init or mu_rndwn_file and did not remove it. So just lock it.
*/
if (!ftok_sem_lock(reg, may_bypass_ftok))
{
if (may_bypass_ftok)
{ /* We did a non-blocking wait. It's ok to proceed without locking */
bypassed_ftok = TRUE;
holder_pid = semctl(udi->ftok_semid, DB_CONTROL_SEM, GETPID);
if ((uint4)-1 == holder_pid)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get holder_pid"),
CALLFROM, errno);
if (!IS_GTM_IMAGE) /* MUMPS processes should not flood syslog with bypass messages. */
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_RESRCINTRLCKBYPAS, 10,
LEN_AND_STR(gtmImageNames[image_type].imageName), process_id, LEN_AND_LIT("FTOK"),
REG_LEN_STR(reg), DB_LEN_STR(reg), holder_pid);
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_TEXT, 2,
LEN_AND_LIT("FTOK bypassed at rundown"));
}
} else
{ /* We did a blocking wait but something bad happened. */
FTOK_TRACE(csa, csa->ti->curr_tn, ftok_ops_lock, process_id);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
}
}
sop[0].sem_num = DB_CONTROL_SEM; sop[0].sem_op = 0; /* Wait for 0 */
sop[1].sem_num = DB_CONTROL_SEM; sop[1].sem_op = 1; /* Lock */
sopcnt = 2;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO | IPC_NOWAIT; /* Don't wait the first time thru */
SEMOP(udi->semid, sop, sopcnt, status, NO_WAIT);
if (0 != status)
{
save_errno = errno;
/* Check # of processes counted on access sem. */
if (-1 == (semval = semctl(udi->semid, DB_COUNTER_SEM, GETVAL)))
{
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get semval"), CALLFROM, errno);
}
bypassed_access = CAN_BYPASS(semval, csd, inst_is_frozen) || onln_rlbk_pid || csd->file_corrupt;
/* Before attempting again in the blocking mode, see if the holding process is an online rollback.
* If so, it is likely we won't get the access control semaphore anytime soon. In that case, we
* are better off skipping rundown and continuing with sanity cleanup and exit.
*/
holder_pid = semctl(udi->semid, DB_CONTROL_SEM, GETPID);
if ((uint4)-1 == holder_pid)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get holder_pid"), CALLFROM, errno);
if (!bypassed_access)
{ /* We couldn't get it in one shot-- see if we already have it */
if (holder_pid == process_id)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(5) MAKE_MSG_INFO(ERR_CRITSEMFAIL), 2, DB_LEN_STR(reg),
ERR_RNDWNSEMFAIL);
REVERT;
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
assert(FALSE);
return EXIT_ERR;
}
if (EAGAIN != save_errno)
{
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMOP on access control semaphore"),
CALLFROM, save_errno);
}
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO; /* Try again - blocking this time */
SEMOP(udi->semid, sop, 2, status, FORCED_WAIT);
if (-1 == status) /* We couldn't get it at all.. */
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMOP on access control semaphore"),
CALLFROM, errno);
} else if (!IS_GTM_IMAGE)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_RESRCINTRLCKBYPAS, 10,
LEN_AND_STR(gtmImageNames[image_type].imageName), process_id,
LEN_AND_LIT("access control"), REG_LEN_STR(reg), DB_LEN_STR(reg), holder_pid);
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_TEXT, 2,
LEN_AND_LIT("Access control bypassed at rundown"));
}
udi->grabbed_access_sem = !bypassed_access;
}
} /* else we we hold the access control semaphore and therefore have standalone access. We do not release it now - we
* release it later in mupip_exit_handler.c. Since we already hold the access control semaphore, we don't need the
* ftok semaphore and trying it could cause deadlock
*/
/* Note that in the case of online rollback, "udi->grabbed_access_sem" (and in turn "have_standalone_access") is TRUE.
* But there could be other processes still having the database open so we cannot safely reset the halted fields.
*/
if (have_standalone_access && !jgbl.onlnrlbk)
csd->ftok_counter_halted = csd->access_counter_halted = FALSE;
ftok_counter_halted = csd->ftok_counter_halted;
access_counter_halted = csd->access_counter_halted;
/* If we bypassed any of the semaphores, activate safe mode.
* Also, if the replication instance is frozen and this db has replication turned on (which means
* no flushes of dirty buffers to this db can happen while the instance is frozen) activate safe mode.
*/
ok_to_write_pfin = !(bypassed_access || bypassed_ftok || inst_is_frozen);
safe_mode = !ok_to_write_pfin || ftok_counter_halted || access_counter_halted;
/* At this point we are guaranteed no one else is doing a db_init/rundown as we hold the access control semaphore */
assert(csa->ref_cnt); /* decrement private ref_cnt before shared ref_cnt decrement. */
csa->ref_cnt--; /* Currently journaling logic in gds_rundown() in VMS relies on this order to detect last writer */
assert(!csa->ref_cnt);
--cnl->ref_cnt;
if (memcmp(cnl->now_running, gtm_release_name, gtm_release_name_len + 1))
{ /* VERMISMATCH condition. Possible only if DSE */
assert(dse_running);
vermismatch = TRUE;
} else
vermismatch = FALSE;
if (-1 == shmctl(udi->shmid, IPC_STAT, &shm_buf))
{
save_errno = errno;
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown shmctl"), CALLFROM, save_errno);
} else
we_are_last_user = (1 == shm_buf.shm_nattch) && !vermismatch && !safe_mode;
/* recover => one user except ONLINE ROLLBACK, or standalone with frozen instance */
assert(!have_standalone_access || we_are_last_user || jgbl.onlnrlbk || inst_is_frozen);
if (-1 == (semval = semctl(udi->semid, DB_COUNTER_SEM, GETVAL)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get semval"), CALLFROM, errno);
/* There's one writer left and I am it */
assert(reg->read_only || semval >= 0);
unsafe_last_writer = (DB_COUNTER_SEM_INCR == semval) && (FALSE == reg->read_only) && !vermismatch;
we_are_last_writer = unsafe_last_writer && !safe_mode;
assert(!we_are_last_writer || !safe_mode);
assert(!we_are_last_user || !safe_mode);
/* recover + R/W region => one writer except ONLINE ROLLBACK, or standalone with frozen instance, leading to safe_mode */
assert(!(have_standalone_access && !reg->read_only) || we_are_last_writer || jgbl.onlnrlbk || inst_is_frozen);
GTM_WHITE_BOX_TEST(WBTEST_ANTIFREEZE_JNLCLOSE, we_are_last_writer, 1); /* Assume we are the last writer to invoke wcs_flu */
if (!have_standalone_access && (-1 == (ftok_semval = semctl(udi->ftok_semid, DB_COUNTER_SEM, GETVAL))))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get ftok_semval"), CALLFROM, errno);
if (NULL != csa->ss_ctx)
ss_destroy_context(csa->ss_ctx);
/* SS_MULTI: If multiple snapshots are supported, then we have to run through each of the snapshots */
assert(1 == MAX_SNAPSHOTS);
ss_shm_ptr = (shm_snapshot_ptr_t)SS_GETSTARTPTR(csa);
ss_pid = ss_shm_ptr->ss_info.ss_pid;
is_cur_process_ss_initiator = (process_id == ss_pid);
if (ss_pid && (is_cur_process_ss_initiator || we_are_last_user))
{
/* Try getting snapshot crit latch. If we don't get latch, we won't hang for eternity and will skip
* doing the orphaned snapshot cleanup. It will be cleaned up eventually either by subsequent MUPIP
* INTEG or by a MUPIP RUNDOWN.
*/
if (ss_get_lock_nowait(reg) && (ss_pid == ss_shm_ptr->ss_info.ss_pid)
&& (is_cur_process_ss_initiator || !is_proc_alive(ss_pid, 0)))
{
ss_release(NULL);
ss_release_lock(reg);
}
}
/* If cnl->donotflush_dbjnl is set, it means mupip recover/rollback was interrupted and therefore we need not flush
* shared memory contents to disk as they might be in an inconsistent state. Moreover, any more flushing will only cause
* future rollback to undo more journal records (PBLKs). In this case, we will go ahead and remove shared memory (without
* flushing the contents) in this routine. A reissue of the recover/rollback command will restore the database to a
* consistent state.
*/
if (!cnl->donotflush_dbjnl && !reg->read_only && !vermismatch)
{ /* If we had an orphaned block and were interrupted, set wc_blocked so we can invoke wcs_recover. Do it ONLY
* if there is NO concurrent online rollback running (as we need crit to set wc_blocked)
*/
if (csa->wbuf_dqd && !is_mm)
{ /* If we had an orphaned block and were interrupted, mupip_exit_handler will invoke secshr_db_clnup which
* will clear this field and so we should never come to gds_rundown with a non-zero wbuf_dqd. The only
* exception is if we are recover/rollback in which case gds_rundown (from mur_close_files) is invoked
* BEFORE secshr_db_clnup in mur_close_files.
* Note: It is NOT possible for online rollback to reach here with wbuf_dqd being non-zero. This is because
* the moment we apply the first PBLK, we stop all interrupts and hence can never be interrupted in
* wcs_wtstart or wcs_get_space. Assert accordingly.
*/
assert(mupip_jnl_recover && !jgbl.onlnrlbk && !safe_mode);
if (!was_crit)
grab_crit(reg);
SET_TRACEABLE_VAR(cnl->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wcb_gds_rundown);
send_msg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_WCBLOCKED, 6, LEN_AND_LIT("wcb_gds_rundown"),
process_id, &csa->ti->curr_tn, DB_LEN_STR(reg));
csa->wbuf_dqd = 0;
wcs_recover(reg);
BG_TRACE_PRO_ANY(csa, lost_block_recovery);
if (!was_crit)
rel_crit(reg);
}
if (JNL_ENABLED(csd) && IS_GTCM_GNP_SERVER_IMAGE)
originator_prc_vec = NULL;
/* If we are the last writing user, then everything must be flushed */
if (we_are_last_writer)
{ /* Time to flush out all of our buffers */
assert(!safe_mode);
if (is_mm)
{
MM_DBFILEXT_REMAP_IF_NEEDED(csa, reg);
cnl->remove_shm = TRUE;
}
if (cnl->wc_blocked && jgbl.onlnrlbk)
{ /* if the last update done by online rollback was not committed in the normal code-path but was
* completed by secshr_db_clnup, wc_blocked will be set to TRUE. But, since online rollback never
* invokes grab_crit (since csa->hold_onto_crit is set to TRUE), wcs_recover is never invoked. This
* could result in the last update never getting flushed to the disk and if online rollback happened
* to be the last writer then the shared memory will be flushed and removed and the last update will
* be lost. So, force wcs_recover if we find ourselves in such a situation. But, wc_blocked is
* possible only if phase1 or phase2 errors are induced using white box test cases
*/
assert(WB_COMMIT_ERR_ENABLED);
wcs_recover(reg);
}
/* Note WCSFLU_SYNC_EPOCH ensures the epoch is synced to the journal and indirectly
* also ensures that the db is fsynced. We don't want to use it in the calls to
* wcs_flu() from t_end() and tp_tend() since we can defer it to out-of-crit there.
* In this case, since we are running down, we don't have any such option.
*/
cnl->remove_shm = wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
/* Since we_are_last_writer, we should be guaranteed that wcs_flu() did not change csd, (in
* case of MM for potential file extension), even if it did a grab_crit(). Therefore, make
* sure that's true.
*/
assert(csd == csa->hdr);
assert(0 == memcmp(csd->label, GDS_LABEL, GDS_LABEL_SZ - 1));
} else if (((canceled_flush_timer && (0 > cnl->wcs_timers)) || canceled_dbsync_timer) && !inst_is_frozen)
{ /* canceled pending db or jnl flush timers - flush database and journal buffers to disk */
if (!was_crit)
grab_crit(reg);
/* we need to sync the epoch as the fact that there is no active pending flush timer implies
* there will be noone else who will flush the dirty buffers and EPOCH to disk in a timely fashion
*/
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
if (!was_crit)
rel_crit(reg);
assert((dba_mm == cs_data->acc_meth) || (csd == cs_data));
csd = cs_data; /* In case this is MM and wcs_flu() remapped an extended database, reset csd */
}
/* Do rundown journal processing after buffer flushes since they require jnl to be open */
if (JNL_ENABLED(csd))
{ /* the following tp_change_reg() is not needed due to the assert csa == cs_addrs at the beginning
* of gds_rundown(), but just to be safe. To be removed by 2002!! --- nars -- 2001/04/25.
*/
tp_change_reg(); /* call this because jnl_ensure_open checks cs_addrs rather than gv_cur_region */
jpc = csa->jnl;
jbp = jpc->jnl_buff;
if (jbp->fsync_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->fsync_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (jbp->io_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->io_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if ((((NOJNL != jpc->channel) && !JNL_FILE_SWITCHED(jpc))
|| we_are_last_writer && (0 != cnl->jnl_file.u.inode)) && ok_to_write_pfin)
{ /* We need to close the journal file cleanly if we have the latest generation journal file open
* or if we are the last writer and the journal file is open in shared memory (not necessarily
* by ourselves e.g. the only process that opened the journal got shot abnormally)
* Note: we should not infer anything from the shared memory value of cnl->jnl_file.u.inode
* if we are not the last writer as it can be concurrently updated.
*/
if (!was_crit)
grab_crit(reg);
if (JNL_ENABLED(csd))
{
SET_GBL_JREC_TIME; /* jnl_ensure_open/jnl_put_jrt_pini/pfin/jnl_file_close all need it */
/* Before writing to jnlfile, adjust jgbl.gbl_jrec_time if needed to maintain time order
* of jnl records. This needs to be done BEFORE the jnl_ensure_open as that could write
* journal records (if it decides to switch to a new journal file).
*/
ADJUST_GBL_JREC_TIME(jgbl, jbp);
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{ /* If we_are_last_writer, we would have already done a wcs_flu() which would
* have written an epoch record and we are guaranteed no further updates
* since we are the last writer. So, just close the journal.
* If the freeaddr == post_epoch_freeaddr, wcs_flu may have skipped writing
* a pini, so allow for that.
*/
assert(!jbp->before_images || is_mm
|| !we_are_last_writer || (0 != jpc->pini_addr) || jgbl.mur_extract
|| (jpc->jnl_buff->freeaddr == jpc->jnl_buff->post_epoch_freeaddr));
/* If we haven't written a pini, let jnl_file_close write the pini/pfin. */
if (!jgbl.mur_extract && (0 != jpc->pini_addr))
jnl_put_jrt_pfin(csa);
/* If not the last writer and no pending flush timer left, do jnl flush now */
if (!we_are_last_writer && (0 > cnl->wcs_timers))
{
if (SS_NORMAL == (jnl_status = jnl_flush(reg)))
{
assert(jbp->freeaddr == jbp->dskaddr);
jnl_fsync(reg, jbp->dskaddr);
assert(jbp->fsync_dskaddr == jbp->dskaddr);
} else
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_JNLFLUSH, 2,
JNL_LEN_STR(csd), ERR_TEXT, 2,
RTS_ERROR_TEXT("Error with journal flush in gds_rundown"),
jnl_status);
assert(NOJNL == jpc->channel);/* jnl file lost has been triggered */
/* In this routine, all code that follows from here on does not
* assume anything about the journaling characteristics of this
* database so it is safe to continue execution even though
* journaling got closed in the middle.
*/
}
}
jnl_file_close(reg, we_are_last_writer, FALSE);
} else
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd),
DB_LEN_STR(reg));
}
if (!was_crit)
rel_crit(reg);
}
}
if (we_are_last_writer) /* Flush the fileheader last and harden the file to disk */
{
if (!was_crit)
grab_crit(reg); /* To satisfy crit requirement in fileheader_sync() */
memset(csd->machine_name, 0, MAX_MCNAMELEN); /* clear the machine_name field */
if (!have_standalone_access && we_are_last_user)
{ /* mupip_exit_handler will do this after mur_close_file */
csd->semid = INVALID_SEMID;
csd->shmid = INVALID_SHMID;
csd->gt_sem_ctime.ctime = 0;
csd->gt_shm_ctime.ctime = 0;
}
fileheader_sync(reg);
if (!was_crit)
rel_crit(reg);
if (!is_mm)
{
GTM_DB_FSYNC(csa, udi->fd, rc); /* Sync it all */
if (-1 == rc)
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
} else
{ /* Now do final MM file sync before exit */
assert(csa->ti->total_blks == csa->total_blks);
#ifdef _AIX
GTM_DB_FSYNC(csa, udi->fd, rc);
if (-1 == rc)
#else
if (-1 == MSYNC((caddr_t)csa->db_addrs[0], (caddr_t)csa->db_addrs[1]))
#endif
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
}
} else if (unsafe_last_writer && !cnl->lastwriterbypas_msg_issued)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_LASTWRITERBYPAS, 2, DB_LEN_STR(reg));
cnl->lastwriterbypas_msg_issued = TRUE;
}
} /* end if (!reg->read_only && !cnl->donotflush_dbjnl) */
/* We had canceled all db timers at start of rundown. In case as part of rundown (wcs_flu above), we had started
* any timers, cancel them BEFORE setting reg->open to FALSE (assert in wcs_clean_dbsync relies on this).
*/
CANCEL_DB_TIMERS(reg, csa, canceled_flush_timer, canceled_dbsync_timer);
if (reg->read_only && we_are_last_user && !have_standalone_access && cnl->remove_shm)
{ /* mupip_exit_handler will do this after mur_close_file */
db_ipcs.semid = INVALID_SEMID;
db_ipcs.shmid = INVALID_SHMID;
db_ipcs.gt_sem_ctime = 0;
db_ipcs.gt_shm_ctime = 0;
db_ipcs.fn_len = reg->dyn.addr->fname_len;
memcpy(db_ipcs.fn, reg->dyn.addr->fname, reg->dyn.addr->fname_len);
db_ipcs.fn[reg->dyn.addr->fname_len] = 0;
/* request gtmsecshr to flush. read_only cannot flush itself */
WAIT_FOR_REPL_INST_UNFREEZE_SAFE(csa);
if (!csa->read_only_fs)
{
secshrstat = send_mesg2gtmsecshr(FLUSH_DB_IPCS_INFO, 0, (char *)NULL, 0);
if (0 != secshrstat)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gtmsecshr failed to update database file header"));
}
}
/* Done with file now, close it */
CLOSEFILE_RESET(udi->fd, rc); /* resets "udi->fd" to FD_INVALID */
if (-1 == rc)
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during file close"), errno);
}
/* Unmap storage if mm mode but only the part that is not the fileheader (so shows up in dumps) */
# if !defined(_AIX)
if (is_mm && (NULL != csa->db_addrs[0]))
{
assert(csa->db_addrs[1] > csa->db_addrs[0]);
munmap_len = (sm_long_t)(csa->db_addrs[1] - csa->db_addrs[0]);
if (0 < munmap_len)
munmap((caddr_t)(csa->db_addrs[0]), (size_t)(munmap_len));
}
# endif
/* Detach our shared memory while still under lock so reference counts will be correct for the next process to run down
* this region. In the process also get the remove_shm status from node_local before detaching.
* If cnl->donotflush_dbjnl is TRUE, it means we can safely remove shared memory without compromising data
* integrity as a reissue of recover will restore the database to a consistent state.
*/
remove_shm = !vermismatch && (cnl->remove_shm || cnl->donotflush_dbjnl);
/* We are done with online rollback on this region. Indicate to other processes by setting the onln_rlbk_pid to 0.
* Do it before releasing crit (t_end relies on this ordering when accessing cnl->onln_rlbk_pid).
*/
if (jgbl.onlnrlbk)
cnl->onln_rlbk_pid = 0;
rel_crit(reg); /* Since we are about to detach from the shared memory, release crit and reset onln_rlbk_pid */
/* If we had skipped flushing journal and database buffers due to a concurrent online rollback, increment the counter
* indicating that in the shared memory so that online rollback can report the # of such processes when it shuts down.
* The same thing is done for both FTOK and access control semaphores when there are too many MUMPS processes.
*/
if (safe_mode) /* indicates flushing was skipped */
{
if (bypassed_access)
cnl->dbrndwn_access_skip++; /* Access semaphore can be bypassed during online rollback */
if (bypassed_ftok)
cnl->dbrndwn_ftok_skip++;
}
if (jgbl.onlnrlbk)
csa->hold_onto_crit = FALSE;
GTM_WHITE_BOX_TEST(WBTEST_HOLD_SEM_BYPASS, cnl->wbox_test_seq_num, 0);
status = shmdt((caddr_t)cnl);
csa->nl = NULL; /* dereferencing nl after detach is not right, so we set it to NULL so that we can test before dereference*/
/* Note that although csa->nl is NULL, we use CSA_ARG(csa) below (not CSA_ARG(NULL)) to be consistent with similar
* usages before csa->nl became NULL. The "is_anticipatory_freeze_needed" function (which is in turn called by the
* CHECK_IF_FREEZE_ON_ERROR_NEEDED macro) does a check of csa->nl before dereferencing shared memory contents so
* we are safe passing "csa".
*/
if (-1 == status)
send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg), ERR_TEXT, 2,
LEN_AND_LIT("Error during shmdt"), errno);
REMOVE_CSA_FROM_CSADDRSLIST(csa); /* remove "csa" from list of open regions (cs_addrs_list) */
reg->open = FALSE;
/* If file is still not in good shape, die here and now before we get rid of our storage */
assertpro(0 == csa->wbuf_dqd);
ipc_deleted = FALSE;
/* If we are the very last user, remove shared storage id and the semaphores */
if (we_are_last_user)
{ /* remove shared storage, only if last writer to rundown did a successful wcs_flu() */
assert(!vermismatch);
if (remove_shm)
{
ipc_deleted = TRUE;
if (0 != shm_rmid(udi->shmid))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove shared memory"));
/* Note that we no longer have a new shared memory. Currently only used/usable for standalone rollback. */
udi->new_shm = FALSE;
/* mupip recover/rollback don't release the semaphore here, but do it later in db_ipcs_reset (invoked from
* mur_close_files())
*/
if (!have_standalone_access)
{
if (0 != sem_rmid(udi->semid))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove semaphore"));
udi->new_sem = FALSE; /* Note that we no longer have a new semaphore */
udi->grabbed_access_sem = FALSE;
udi->counter_acc_incremented = FALSE;
}
} else if (is_src_server || is_updproc)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
{
assert(!have_standalone_access || jgbl.onlnrlbk || safe_mode);
if (!jgbl.onlnrlbk && !have_standalone_access)
{ /* If we were writing, get rid of our writer access count semaphore */
if (!reg->read_only)
{
if (!access_counter_halted)
{
save_errno = do_semop(udi->semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO);
if (0 != save_errno)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown access control semaphore decrement"),
CALLFROM, save_errno);
}
udi->counter_acc_incremented = FALSE;
}
assert(safe_mode || !bypassed_access);
/* Now remove the rundown lock */
if (!bypassed_access)
{
if (0 != (save_errno = do_semop(udi->semid, DB_CONTROL_SEM, -1, SEM_UNDO)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown access control semaphore release"),
CALLFROM, save_errno);
udi->grabbed_access_sem = FALSE;
}
} /* else access control semaphore will be released in db_ipcs_reset */
}
if (!have_standalone_access)
{
if (bypassed_ftok)
{
if (!ftok_counter_halted)
if (0 != (save_errno = do_semop(udi->ftok_semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
} else if (!ftok_sem_release(reg, !ftok_counter_halted, FALSE))
{
FTOK_TRACE(csa, csa->ti->curr_tn, ftok_ops_release, process_id);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
}
udi->grabbed_ftok_sem = FALSE;
udi->counter_ftok_incremented = FALSE;
}
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
if (!ipc_deleted)
{
GET_CUR_TIME(time_str);
if (is_src_server)
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Source server"), REG_LEN_STR(reg));
if (is_updproc)
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Update process"), REG_LEN_STR(reg));
if (mupip_jnl_recover && (!jgbl.onlnrlbk || !we_are_last_user))
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
send_msg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
}
}
REVERT;
return EXIT_NRM;
}
void db_init(gd_region *reg, sgmnt_data_ptr_t tsd)
{
static boolean_t mutex_init_done = FALSE;
boolean_t is_bg, read_only;
char machine_name[MAX_MCNAMELEN];
file_control *fc;
int gethostname_res, stat_res, mm_prot;
int4 status, semval, dblksize, fbwsize;
sm_long_t status_l;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
struct sembuf sop[3];
struct stat stat_buf;
union semun semarg;
struct semid_ds semstat;
struct shmid_ds shmstat;
struct statvfs dbvfs;
uint4 sopcnt;
unix_db_info *udi;
#ifdef periodic_timer_removed
void periodic_flush_check();
#endif
error_def(ERR_CLSTCONFLICT);
error_def(ERR_CRITSEMFAIL);
error_def(ERR_DBNAMEMISMATCH);
error_def(ERR_DBIDMISMATCH);
error_def(ERR_NLMISMATCHCALC);
error_def(ERR_REQRUNDOWN);
error_def(ERR_SYSCALL);
assert(tsd->acc_meth == dba_bg || tsd->acc_meth == dba_mm);
is_bg = (dba_bg == tsd->acc_meth);
read_only = reg->read_only;
new_dbinit_ipc = FALSE; /* we did not create a new ipc resource */
udi = FILE_INFO(reg);
memset(machine_name, 0, sizeof(machine_name));
if (GETHOSTNAME(machine_name, MAX_MCNAMELEN, gethostname_res))
rts_error(VARLSTCNT(5) ERR_TEXT, 2, LEN_AND_LIT("Unable to get the hostname"), errno);
assert(strlen(machine_name) < MAX_MCNAMELEN);
csa = &udi->s_addrs;
csa->db_addrs[0] = csa->db_addrs[1] = csa->lock_addrs[0] = NULL; /* to help in dbinit_ch and gds_rundown */
reg->opening = TRUE;
/*
* Create ftok semaphore for this region.
* We do not want to make ftok counter semaphore to be 2 for on mupip journal recover process.
*/
if (!ftok_sem_get(reg, !mupip_jnl_recover, GTM_ID, FALSE))
rts_error(VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
/*
* At this point we have ftok_semid sempahore based on ftok key.
* Any ftok conflicted region will block at this point.
* Say, a.dat and b.dat both has same ftok and we have process A to access a.dat and
* process B to access b.dat. In this case only one can continue to do db_init()
*/
fc = reg->dyn.addr->file_cntl;
fc->file_type = reg->dyn.addr->acc_meth;
fc->op = FC_READ;
fc->op_buff = (sm_uc_ptr_t)tsd;
fc->op_len = sizeof(*tsd);
fc->op_pos = 1;
dbfilop(fc); /* Read file header */
udi->shmid = tsd->shmid;
udi->semid = tsd->semid;
udi->sem_ctime = tsd->sem_ctime.ctime;
udi->shm_ctime = tsd->shm_ctime.ctime;
dbsecspc(reg, tsd); /* Find db segment size */
if (!mupip_jnl_recover)
{
if (INVALID_SEMID == udi->semid)
{
if (0 != udi->sem_ctime || INVALID_SHMID != udi->shmid || 0 != udi->shm_ctime)
/* We must have somthing wrong in protocol or, code, if this happens */
GTMASSERT;
/*
* Create new semaphore using IPC_PRIVATE. System guarantees a unique id.
*/
if (-1 == (udi->semid = semget(IPC_PRIVATE, FTOK_SEM_PER_ID, RWDALL | IPC_CREAT)))
{
udi->semid = INVALID_SEMID;
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error with database control semget"), errno);
}
udi->shmid = INVALID_SHMID; /* reset shmid so dbinit_ch does not get confused in case we go there */
new_dbinit_ipc = TRUE;
tsd->semid = udi->semid;
semarg.val = GTM_ID;
/*
* Following will set semaphore number 2 (=FTOK_SEM_PER_ID - 1) value as GTM_ID.
* In case we have orphaned semaphore for some reason, mupip rundown will be
* able to identify GTM semaphores from the value and can remove.
*/
if (-1 == semctl(udi->semid, FTOK_SEM_PER_ID - 1, SETVAL, semarg))
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error with database control semctl SETVAL"), errno);
/*
* Warning: We must read the sem_ctime using IPC_STAT after SETVAL, which changes it.
* We must NOT do any more SETVAL after this. Our design is to use
* sem_ctime as creation time of semaphore.
*/
semarg.buf = &semstat;
if (-1 == semctl(udi->semid, FTOK_SEM_PER_ID - 1, IPC_STAT, semarg))
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error with database control semctl IPC_STAT"), errno);
tsd->sem_ctime.ctime = udi->sem_ctime = semarg.buf->sem_ctime;
} else
{
if (INVALID_SHMID == udi->shmid)
/* if mu_rndwn_file gets standalone access of this region and
* somehow mupip process crashes, we can have semid != -1 but shmid == -1
*/
rts_error(VARLSTCNT(10) ERR_REQRUNDOWN, 4, DB_LEN_STR(reg), LEN_AND_STR(tsd->machine_name),
ERR_TEXT, 2, LEN_AND_LIT("semid is valid but shmid is invalid"));
semarg.buf = &semstat;
if (-1 == semctl(udi->semid, 0, IPC_STAT, semarg))
/* file header has valid semid but semaphore does not exists */
rts_error(VARLSTCNT(6) ERR_REQRUNDOWN, 4, DB_LEN_STR(reg), LEN_AND_STR(tsd->machine_name));
else if (semarg.buf->sem_ctime != tsd->sem_ctime.ctime)
rts_error(VARLSTCNT(10) ERR_REQRUNDOWN, 4, DB_LEN_STR(reg), LEN_AND_STR(tsd->machine_name),
ERR_TEXT, 2, LEN_AND_LIT("sem_ctime does not match"));
if (-1 == shmctl(udi->shmid, IPC_STAT, &shmstat))
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error with database control shmctl"), errno);
else if (shmstat.shm_ctime != tsd->shm_ctime.ctime)
rts_error(VARLSTCNT(10) ERR_REQRUNDOWN, 4, DB_LEN_STR(reg), LEN_AND_STR(tsd->machine_name),
ERR_TEXT, 2, LEN_AND_LIT("shm_ctime does not match"));
}
/* We already have ftok semaphore of this region, so just plainly do semaphore operation */
/* This is the database access control semaphore for any region */
sop[0].sem_num = 0; sop[0].sem_op = 0; /* Wait for 0 */
sop[1].sem_num = 0; sop[1].sem_op = 1; /* Lock */
sopcnt = 2;
if (!read_only)
{
sop[2].sem_num = 1; sop[2].sem_op = 1; /* increment r/w access counter */
sopcnt = 3;
}
sop[0].sem_flg = sop[1].sem_flg = sop[2].sem_flg = SEM_UNDO | IPC_NOWAIT;
SEMOP(udi->semid, sop, sopcnt, status);
if (-1 == status)
{
errno_save = errno;
gtm_putmsg(VARLSTCNT(4) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg));
rts_error(VARLSTCNT(8) ERR_SYSCALL, 5, RTS_ERROR_LITERAL("semop()"), CALLFROM, errno_save);
}
} else /* for mupip_jnl_recover we were already in mu_rndwn_file and got "semid" semaphore */
{
if (INVALID_SEMID == udi->semid || 0 == udi->sem_ctime)
/* make sure mu_rndwn_file() has reset created semaphore for standalone access */
GTMASSERT;
if (INVALID_SHMID != udi->shmid || 0 != udi->shm_ctime)
/* make sure mu_rndwn_file() has reset shared memory */
GTMASSERT;
udi->shmid = INVALID_SHMID; /* reset shmid so dbinit_ch does not get confused in case we go there */
new_dbinit_ipc = TRUE;
}
sem_incremented = TRUE;
if (new_dbinit_ipc)
{
/* Create new shared memory using IPC_PRIVATE. System guarantees a unique id */
#ifdef __MVS__
if (-1 == (status_l = udi->shmid = shmget(IPC_PRIVATE, ROUND_UP(reg->sec_size, MEGA_BOUND),
__IPC_MEGA | IPC_CREAT | RWDALL)))
#else
if (-1 == (status_l = udi->shmid = shmget(IPC_PRIVATE, reg->sec_size, RWDALL | IPC_CREAT)))
#endif
{
udi->shmid = status_l = INVALID_SHMID;
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error with database shmget"), errno);
}
tsd->shmid = udi->shmid;
if (-1 == shmctl(udi->shmid, IPC_STAT, &shmstat))
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error with database control shmctl"), errno);
tsd->shm_ctime.ctime = udi->shm_ctime = shmstat.shm_ctime;
}
#ifdef DEBUG_DB64
status_l = (sm_long_t)(csa->db_addrs[0] = (sm_uc_ptr_t)do_shmat(udi->shmid, next_smseg, SHM_RND));
next_smseg = (sm_uc_ptr_t)ROUND_UP((sm_long_t)(next_smseg + reg->sec_size), SHMAT_ADDR_INCS);
#else
status_l = (sm_long_t)(csa->db_addrs[0] = (sm_uc_ptr_t)do_shmat(udi->shmid, 0, SHM_RND));
#endif
if (-1 == status_l)
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error attaching to database shared memory"), errno);
}
csa->nl = (node_local_ptr_t)csa->db_addrs[0];
csa->critical = (mutex_struct_ptr_t)(csa->db_addrs[0] + NODE_LOCAL_SIZE);
assert(((int)csa->critical & 0xf) == 0); /* critical should be 16-byte aligned */
#ifdef CACHELINE_SIZE
assert(0 == ((int)csa->critical & (CACHELINE_SIZE - 1)));
#endif
/* Note: Here we check jnl_sate from database file and its value cannot change without standalone access.
* The jnl_buff buffer should be initialized irrespective of read/write process */
JNL_INIT(csa, reg, tsd);
csa->backup_buffer = (backup_buff_ptr_t)(csa->db_addrs[0] + NODE_LOCAL_SPACE + JNL_SHARE_SIZE(tsd));
csa->lock_addrs[0] = (sm_uc_ptr_t)csa->backup_buffer + BACKUP_BUFFER_SIZE + 1;
csa->lock_addrs[1] = csa->lock_addrs[0] + LOCK_SPACE_SIZE(tsd) - 1;
csa->total_blks = tsd->trans_hist.total_blks; /* For test to see if file has extended */
if (new_dbinit_ipc)
{
memset(csa->nl, 0, sizeof(*csa->nl)); /* We allocated shared storage -- we have to init it */
if (JNL_ALLOWED(csa))
{ /* initialize jb->cycle to a value different from initial value of jpc->cycle (0). although this is not
* necessary right now, in the future, the plan is to change jnl_ensure_open() to only do a cycle mismatch
* check in order to determine whether to call jnl_file_open() or not. this is in preparation for that.
*/
csa->jnl->jnl_buff->cycle = 1;
}
}
if (is_bg)
csd = csa->hdr = (sgmnt_data_ptr_t)(csa->lock_addrs[1] + 1 + CACHE_CONTROL_SIZE(tsd));
else
{
csa->acc_meth.mm.mmblk_state = (mmblk_que_heads_ptr_t)(csa->lock_addrs[1] + 1);
FSTAT_FILE(udi->fd, &stat_buf, stat_res);
if (-1 == stat_res)
rts_error(VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(reg), errno);
mm_prot = read_only ? PROT_READ : (PROT_READ | PROT_WRITE);
#ifdef DEBUG_DB64
if (-1 == (sm_long_t)(csa->db_addrs[0] = (sm_uc_ptr_t)mmap((caddr_t)get_mmseg((size_t)stat_buf.st_size),
(size_t)stat_buf.st_size,
mm_prot,
GTM_MM_FLAGS, udi->fd, (off_t)0)))
rts_error(VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(reg), errno);
put_mmseg((caddr_t)(csa->db_addrs[0]), (size_t)stat_buf.st_size);
#else
if (-1 == (sm_long_t)(csa->db_addrs[0] = (sm_uc_ptr_t)mmap((caddr_t)NULL,
(size_t)stat_buf.st_size,
mm_prot,
GTM_MM_FLAGS, udi->fd, (off_t)0)))
rts_error(VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(reg), errno);
#endif
csa->db_addrs[1] = csa->db_addrs[0] + stat_buf.st_size - 1;
csd = csa->hdr = (sgmnt_data_ptr_t)csa->db_addrs[0];
}
if (!csa->nl->glob_sec_init)
{
assert(new_dbinit_ipc);
if (is_bg)
*csd = *tsd;
if (csd->machine_name[0]) /* crash occured */
{
if (0 != memcmp(csd->machine_name, machine_name, MAX_MCNAMELEN)) /* crashed on some other node */
rts_error(VARLSTCNT(6) ERR_CLSTCONFLICT, 4, DB_LEN_STR(reg), LEN_AND_STR(csd->machine_name));
else
rts_error(VARLSTCNT(6) ERR_REQRUNDOWN, 4, DB_LEN_STR(reg), LEN_AND_STR(csd->machine_name));
}
if (is_bg)
{
bt_malloc(csa);
csa->nl->cache_off = -CACHE_CONTROL_SIZE(tsd);
db_csh_ini(csa);
}
db_csh_ref(csa);
strcpy(csa->nl->machine_name, machine_name); /* machine name */
assert(MAX_REL_NAME > gtm_release_name_len);
memcpy(csa->nl->now_running, gtm_release_name, gtm_release_name_len + 1); /* GT.M release name */
memcpy(csa->nl->label, GDS_LABEL, GDS_LABEL_SZ - 1); /* GDS label */
memcpy(csa->nl->fname, reg->dyn.addr->fname, reg->dyn.addr->fname_len); /* database filename */
csa->nl->creation_date_time = csd->creation.date_time;
csa->nl->highest_lbm_blk_changed = -1;
csa->nl->wcs_timers = -1;
csa->nl->nbb = BACKUP_NOT_IN_PROGRESS;
csa->nl->unique_id.uid = FILE_INFO(reg)->fileid; /* save what file we initialized this storage for */
/* save pointers in csa to access shared memory */
csa->nl->critical = (sm_off_t)((sm_uc_ptr_t)csa->critical - (sm_uc_ptr_t)csa->nl);
if (JNL_ALLOWED(csa))
csa->nl->jnl_buff = (sm_off_t)((sm_uc_ptr_t)csa->jnl->jnl_buff - (sm_uc_ptr_t)csa->nl);
csa->nl->backup_buffer = (sm_off_t)((sm_uc_ptr_t)csa->backup_buffer - (sm_uc_ptr_t)csa->nl);
csa->nl->hdr = (sm_off_t)((sm_uc_ptr_t)csd - (sm_uc_ptr_t)csa->nl);
csa->nl->lock_addrs = (sm_off_t)((sm_uc_ptr_t)csa->lock_addrs[0] - (sm_uc_ptr_t)csa->nl);
if (!read_only || is_bg)
{
csd->trans_hist.early_tn = csd->trans_hist.curr_tn;
csd->max_update_array_size = csd->max_non_bm_update_array_size
= ROUND_UP2(MAX_NON_BITMAP_UPDATE_ARRAY_SIZE(csd), UPDATE_ARRAY_ALIGN_SIZE);
csd->max_update_array_size += ROUND_UP2(MAX_BITMAP_UPDATE_ARRAY_SIZE, UPDATE_ARRAY_ALIGN_SIZE);
/* add current db_csh counters into the cumulative counters and reset the current counters */
#define TAB_DB_CSH_ACCT_REC(COUNTER, DUMMY1, DUMMY2) \
csd->COUNTER.cumul_count += csd->COUNTER.curr_count; \
csd->COUNTER.curr_count = 0;
#include "tab_db_csh_acct_rec.h"
#undef TAB_DB_CSH_ACCT_REC
}
if (!read_only)
{
if (is_bg)
{
assert(memcmp(csd, GDS_LABEL, GDS_LABEL_SZ - 1) == 0);
LSEEKWRITE(udi->fd, (off_t)0, (sm_uc_ptr_t)csd, sizeof(sgmnt_data), errno_save);
if (0 != errno_save)
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error with database write"), errno_save);
}
}
}
reg->dyn.addr->ext_blk_count = csd->extension_size;
mlk_shr_init(csa->lock_addrs[0], csd->lock_space_size, csa, (FALSE == read_only));
DEBUG_ONLY(locknl = csa->nl;) /* for DEBUG_ONLY LOCK_HIST macro */
/*
* Purchase the property disallowing others to lease or purchase it.
*/
FusionResult
fusion_property_purchase (FusionProperty *property)
{
struct sembuf op[3];
/* lock */
op[0].sem_num = 0;
op[0].sem_op = -1;
op[0].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 1);
while (property->state == FUSION_PROPERTY_LEASED) {
/* increment wait counter */
op[0].sem_num = 1;
op[0].sem_op = 1;
op[0].sem_flg = SEM_UNDO;
/* unlock */
op[1].sem_num = 0;
op[1].sem_op = 1;
op[1].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 2);
/* lock */
op[0].sem_num = 0;
op[0].sem_op = -1;
op[0].sem_flg = SEM_UNDO;
/* wait */
op[1].sem_num = 2;
op[1].sem_op = -1;
op[1].sem_flg = 0;
/* decrement wait counter */
op[2].sem_num = 1;
op[2].sem_op = -1;
op[2].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 3);
}
if (property->state == FUSION_PROPERTY_AVAILABLE) {
union semun semopts;
int waiters;
property->state = FUSION_PROPERTY_PURCHASED;
/* get number of waiters */
waiters = semctl (property->sem_id, 1, GETVAL, semopts);
if (waiters < 0) {
FPERROR ("semctl");
/* unlock */
op[0].sem_num = 0;
op[0].sem_op = 1;
op[0].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 1);
return FUSION_FAILURE;
}
/* unlock */
op[0].sem_num = 0;
op[0].sem_op = 1;
op[0].sem_flg = SEM_UNDO;
/* awake waiters */
op[1].sem_num = 2;
op[1].sem_op = waiters;
op[1].sem_flg = 0;
SEMOP (property->sem_id, op, waiters ? 2 : 1);
return FUSION_SUCCESS;
}
/* unlock */
op[0].sem_num = 0;
op[0].sem_op = 1;
op[0].sem_flg = SEM_UNDO;
SEMOP (property->sem_id, op, 1);
return FUSION_INUSE;
}
