/* This called for TP and non-TP, but not for ZTP */ void jnl_write_logical(sgmnt_addrs *csa, jnl_format_buffer *jfb, uint4 com_csum, jnlpool_write_ctx_t *jplctx) { struct_jrec_upd *jrec; struct_jrec_null *jrec_null; struct_jrec_upd *jrec_alt; jnl_private_control *jpc; /* If REPL_WAS_ENABLED(csa) is TRUE, then we would not have gone through the code that initializes * jgbl.gbl_jrec_time or jpc->pini_addr. But in this case, we are not writing the journal record * to the journal buffer or journal file but write it only to the journal pool from where it gets * sent across to the update process that does not care about these fields so it is ok to leave them as is. */ jpc = csa->jnl; assert((0 != jpc->pini_addr) || REPL_WAS_ENABLED(csa)); assert(jgbl.gbl_jrec_time || REPL_WAS_ENABLED(csa)); assert(csa->now_crit); assert(IS_SET_KILL_ZKILL_ZTWORM_LGTRIG_ZTRIG(jfb->rectype) || (JRT_NULL == jfb->rectype)); assert(!IS_ZTP(jfb->rectype)); jrec = (struct_jrec_upd *)jfb->buff; assert(OFFSETOF(struct_jrec_null, prefix) == OFFSETOF(struct_jrec_upd, prefix)); assert(SIZEOF(jrec_null->prefix) == SIZEOF(jrec->prefix)); jrec->prefix.pini_addr = (0 == jpc->pini_addr) ? JNL_HDR_LEN : jpc->pini_addr; jrec->prefix.tn = csa->ti->curr_tn; jrec->prefix.time = jgbl.gbl_jrec_time; /* t_end/tp_tend/mur_output_record has already set token/jnl_seqno into jnl_fence_ctl.token */ assert((0 != jnl_fence_ctl.token) || (!dollar_tlevel && !jgbl.forw_phase_recovery && !REPL_ENABLED(csa)) || (!dollar_tlevel && jgbl.forw_phase_recovery && (repl_open != csa->hdr->intrpt_recov_repl_state))); assert(OFFSETOF(struct_jrec_null, jnl_seqno) == OFFSETOF(struct_jrec_upd, token_seq)); assert(SIZEOF(jrec_null->jnl_seqno) == SIZEOF(jrec->token_seq)); jrec->token_seq.token = jnl_fence_ctl.token; assert(OFFSETOF(struct_jrec_null, strm_seqno) == OFFSETOF(struct_jrec_upd, strm_seqno)); assert(SIZEOF(jrec_null->strm_seqno) == SIZEOF(jrec->strm_seqno)); jrec->strm_seqno = jnl_fence_ctl.strm_seqno; /* update checksum below */ if(JRT_NULL != jrec->prefix.jrec_type) { COMPUTE_LOGICAL_REC_CHECKSUM(jfb->checksum, jrec, com_csum, jrec->prefix.checksum); } else jrec->prefix.checksum = compute_checksum(INIT_CHECKSUM_SEED, (unsigned char *)jrec, SIZEOF(struct_jrec_null)); if (REPL_ALLOWED(csa) && USES_ANY_KEY(csa->hdr)) { jrec_alt = (struct_jrec_upd *)jfb->alt_buff; jrec_alt->prefix = jrec->prefix; jrec_alt->token_seq = jrec->token_seq; jrec_alt->strm_seqno = jrec->strm_seqno; jrec_alt->num_participants = jrec->num_participants; } JNL_WRITE_APPROPRIATE(csa, jpc, jfb->rectype, (jnl_record *)jrec, NULL, jfb, jplctx); }
void gtmsource_seqno_init(void) { /* Find the start_jnl_seqno */ gd_region *reg, *region_top; seq_num local_read_jsn, local_jsn; sgmnt_addrs *csa; sgmnt_data_ptr_t csd; sm_uc_ptr_t gld_fn; /* Unix and VMS have different field names for now, but will both be soon changed to instname instead of gtmgbldir */ gld_fn = (sm_uc_ptr_t)jnlpool.jnlpool_ctl->jnlpool_id.gtmgbldir; QWASSIGN(jnlpool.jnlpool_ctl->start_jnl_seqno, seq_num_zero); QWASSIGN(local_read_jsn, seq_num_zero); QWASSIGN(local_jsn, seq_num_zero); region_top = gd_header->regions + gd_header->n_regions; for (reg = gd_header->regions; reg < region_top; reg++) { assert(reg->open); csa = &FILE_INFO(reg)->s_addrs; csd = csa->hdr; if (REPL_ALLOWED(csd)) { if (QWLT(local_read_jsn, csd->resync_seqno)) QWASSIGN(local_read_jsn, csd->resync_seqno); if (QWLT(local_jsn, csd->reg_seqno)) QWASSIGN(local_jsn, csd->reg_seqno); /* Copy gtmgbldir into the database shared memory. * Used later to avoid updates from a different gtmgbldir to this database. */ assert(SIZEOF(csa->nl->replinstfilename) == SIZEOF(jnlpool.jnlpool_ctl->jnlpool_id.gtmgbldir)); memcpy(csa->nl->replinstfilename, gld_fn, SIZEOF(csa->nl->replinstfilename)); } } if (QWEQ(local_jsn, seq_num_zero)) { /* No replicated region, or databases created with older version of GTM */ gtm_putmsg(VARLSTCNT(5) ERR_NOREPLCTDREG, 3, LEN_AND_LIT("global directory"), gld_fn); /* Error, has to shutdown all regions 'cos mupip needs exclusive access to turn replication on */ gtmsource_autoshutdown(); } QWASSIGN(jnlpool.jnlpool_ctl->start_jnl_seqno, local_jsn); QWASSIGN(jnlpool.jnlpool_ctl->jnl_seqno, local_jsn); QWASSIGN(jnlpool.gtmsource_local->read_jnl_seqno, local_read_jsn); }
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; }
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);