int mur_forward_multi_proc(reg_ctl_list *rctl) { boolean_t multi_proc, this_reg_stuck, release_latch, ok_to_play; boolean_t cancelled_dbsync_timer, cancelled_timer; reg_ctl_list *rctl_top, *prev_rctl; jnl_ctl_list *jctl; gd_region *reg; sgmnt_addrs *csa; seq_num rec_token_seq; jnl_tm_t rec_time; enum broken_type recstat; jnl_record *rec; enum jnl_record_type rectype; char errstr[256]; int i, rctl_index, save_errno, num_procs_stuck, num_reg_stuck; uint4 status, regcnt_stuck, num_partners, start_hrtbt_cntr; forw_multi_struct *forw_multi; shm_forw_multi_t *sfm; multi_struct *multi; jnl_tm_t adjusted_resolve_time; shm_reg_ctl_t *shm_rctl_start, *shm_rctl, *first_shm_rctl; size_t shm_size, reccnt, copy_size; int4 *size_ptr; char *shmPtr; /* not using "shm_ptr" since it is already used in an AIX include file */ int shmid; multi_proc_shm_hdr_t *mp_hdr; /* Pointer to "multi_proc_shm_hdr_t" structure in shared memory */ status = 0; /* Although we made sure the # of tasks is the same as the # of processes forked off (in the "gtm_multi_proc" * invocation in "mur_forward"), it is possible one of the forked process finishes one invocation of * "mur_forward_multi_proc" before even another forked process gets assigned one task in "gtm_multi_proc_helper". * In this case, we would be invoked more than once. But the first invocation would have done all the needed stuff * so return for later invocations. */ if (mur_forward_multi_proc_done) return 0; mur_forward_multi_proc_done = TRUE; /* Note: "rctl" is unused. But cannot avoid passing it since "gtm_multi_proc" expects something */ prev_rctl = NULL; rctl_start = NULL; adjusted_resolve_time = murgbl.adjusted_resolve_time; assert(0 == murgbl.regcnt_remaining); multi_proc = multi_proc_in_use; /* cache value in "local" to speed up access inside loops below */ if (multi_proc) { mp_hdr = multi_proc_shm_hdr; shm_rctl_start = mur_shm_hdr->shm_rctl_start; if (jgbl.onlnrlbk) { for (rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++) { assert(rctl->csa->hold_onto_crit); /* would have been set in parent process */ rctl->csa->hold_onto_crit = FALSE; /* reset since we dont own this region */ assert(rctl->csa->now_crit); /* would have been set in parent process */ rctl->csa->now_crit = FALSE; /* reset since we dont own this region */ } } START_HEARTBEAT_IF_NEEDED; /* heartbeat timer needed later (in case not already started by "gtm_multi_proc") */ } first_shm_rctl = NULL; /* Phase1 of forward recovery starts */ for (rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++) { /* Check if "rctl" is available for us or if some other concurrent process has taken it */ if (multi_proc) { rctl_index = rctl - &mur_ctl[0]; shm_rctl = &shm_rctl_start[rctl_index]; if (shm_rctl->owning_pid) { assert(process_id != shm_rctl->owning_pid); continue; } GRAB_MULTI_PROC_LATCH_IF_NEEDED(release_latch); assert(release_latch); for ( ; rctl < rctl_top; rctl++, shm_rctl++) { if (shm_rctl->owning_pid) { assert(process_id != shm_rctl->owning_pid); continue; } shm_rctl->owning_pid = process_id; /* Declare ownership */ rctl->this_pid_is_owner = TRUE; if (jgbl.onlnrlbk) { /* This is an online rollback and crit was grabbed on all regions by the parent rollback * process. But this child process now owns this region and does the actual rollback on * this region so borrow crit for the duration of this child process. */ csa = rctl->csa; csa->hold_onto_crit = TRUE; csa->now_crit = TRUE; assert(csa->nl->in_crit == mp_hdr->parent_pid); csa->nl->in_crit = process_id; assert(csa->nl->onln_rlbk_pid == mp_hdr->parent_pid); csa->nl->onln_rlbk_pid = process_id; } if (NULL == first_shm_rctl) first_shm_rctl = shm_rctl; break; } REL_MULTI_PROC_LATCH_IF_NEEDED(release_latch); if (rctl >= rctl_top) { assert(rctl == rctl_top); break; } /* Set key to print this rctl'ss region-name as prefix in case this forked off process prints any output */ MUR_SET_MULTI_PROC_KEY(rctl, multi_proc_key); # ifdef MUR_DEBUG fprintf(stderr, "pid = %d : Owns region %s\n", process_id, multi_proc_key); # endif } else rctl->this_pid_is_owner = TRUE; if (mur_options.forward) { assert(NULL == rctl->jctl_turn_around); jctl = rctl->jctl = rctl->jctl_head; assert(jctl->reg_ctl == rctl); jctl->rec_offset = JNL_HDR_LEN; jnl_fence_ctl.fence_list = JNL_FENCE_LIST_END; /* initialized to reflect journaling is not enabled */ if (mur_options.rollback) jgbl.mur_jrec_seqno = jctl->jfh->start_seqno; } else { jctl = rctl->jctl = (NULL == rctl->jctl_turn_around) ? rctl->jctl_head : rctl->jctl_turn_around; assert(jctl->reg_ctl == rctl); jctl->rec_offset = jctl->turn_around_offset; jgbl.mur_jrec_seqno = jctl->turn_around_seqno; assert((NULL != rctl->jctl_turn_around) || (0 == jctl->rec_offset)); } if (mur_options.rollback) { if (murgbl.consist_jnl_seqno < jgbl.mur_jrec_seqno) { /* Assert that murgbl.losttn_seqno is never lesser than jgbl.mur_jrec_seqno (the turnaround * point seqno) as this is what murgbl.consist_jnl_seqno is going to be set to and will * eventually be the post-rollback seqno. If this condition is violated, the result of the * recovery is a compromised database (the file header will indicate a Region Seqno which * is not necessarily correct since seqnos prior to it might be absent in the database). * Therefore, this is an out-of-design situation with respect to rollback and so stop it. */ assert(murgbl.losttn_seqno >= jgbl.mur_jrec_seqno); murgbl.consist_jnl_seqno = jgbl.mur_jrec_seqno; } assert(murgbl.consist_jnl_seqno <= murgbl.losttn_seqno); } if (mur_options.update || mur_options.extr[GOOD_TN]) { reg = rctl->gd; gv_cur_region = reg; tp_change_reg(); /* note : sets cs_addrs to non-NULL value even if gv_cur_region->open is FALSE * (cs_data could still be NULL). */ rctl->csa = cs_addrs; cs_addrs->miscptr = (void *)rctl; rctl->csd = cs_data; rctl->sgm_info_ptr = cs_addrs->sgm_info_ptr; assert(!reg->open || (NULL != cs_addrs->dir_tree)); gv_target = cs_addrs->dir_tree; } jctl->after_end_of_data = FALSE; status = mur_next(jctl, jctl->rec_offset); assert(ERR_JNLREADEOF != status); /* cannot get EOF at start of forward processing */ if (SS_NORMAL != status) goto finish; PRINT_VERBOSE_STAT(jctl, "mur_forward:at the start"); rctl->process_losttn = FALSE; /* Any multi-region TP transaction will be processed as multiple single-region TP transactions up * until the tp-resolve-time is reached. From then on, they will be treated as one multi-region TP * transaction. This is needed for proper lost-tn determination (any multi-region transaction that * gets played in a region AFTER it has already encountered a broken tn should treat this as a lost tn). */ do { if (multi_proc && IS_FORCED_MULTI_PROC_EXIT(mp_hdr)) { /* We are at a logical point. So exit if signaled by parent */ status = ERR_FORCEDHALT; goto finish; } assert(jctl == rctl->jctl); rec = rctl->mur_desc->jnlrec; rec_time = rec->prefix.time; if (rec_time > mur_options.before_time) break; /* Records after -BEFORE_TIME do not go to extract or losttrans or brkntrans files */ if (rec_time < mur_options.after_time) { status = mur_next_rec(&jctl); continue; /* Records before -AFTER_TIME do not go to extract or losttrans or brkntrans files */ } if (rec_time >= adjusted_resolve_time) break; /* Records after this adjusted resolve_time will be processed below in phase2 */ /* Note: Since we do hashtable token processing only for records from tp_resolve_time onwards, * it is possible that if we encounter any broken transactions here we wont know they are broken * but will play them as is. That is unavoidable. Specify -SINCE_TIME (for -BACKWARD rollback/recover) * and -VERIFY (for -FORWARD rollback/recover) to control tp_resolve_time (and in turn more * effective broken tn determination). */ status = mur_forward_play_cur_jrec(rctl); if (SS_NORMAL != status) break; status = mur_next_rec(&jctl); } while (SS_NORMAL == status); CHECK_IF_EOF_REACHED(rctl, status); /* sets rctl->forw_eof_seen if needed; resets "status" to SS_NORMAL */ if (SS_NORMAL != status) { /* ERR_FILENOTCREATE is possible from "mur_cre_file_extfmt" OR ERR_FORCEDHALT is possible * from "mur_forward_play_cur_jrec". No other errors are known to occur here. Assert accordingly. */ assert((ERR_FILENOTCREATE == status) || (ERR_FORCEDHALT == status)); goto finish; } if (rctl->forw_eof_seen) { PRINT_VERBOSE_STAT(jctl, "mur_forward:Reached EOF before tp_resolve_time"); continue; /* Reached EOF before even getting to tp_resolve_time. * Do not even consider region for next processing loop */ } rctl->last_tn = 0; murgbl.regcnt_remaining++; /* # of regions participating in recovery at this point */ if (NULL == rctl_start) rctl_start = rctl; if (NULL != prev_rctl) { prev_rctl->next_rctl = rctl; rctl->prev_rctl = prev_rctl; } prev_rctl = rctl; assert(murgbl.ok_to_update_db || !rctl->db_updated); PRINT_VERBOSE_STAT(jctl, "mur_forward:at tp_resolve_time"); } if (multi_proc) multi_proc_key = NULL; /* reset key until it can be set to rctl's region-name again */ /* Note that it is possible for rctl_start to be NULL at this point. That is there is no journal record in any region * AFTER the calculated tp-resolve-time. This is possible if for example -AFTER_TIME was used and has a time later * than any journal record in all journal files. If rctl_start is NULL, prev_rctl should also be NULL and vice versa. */ if (NULL != rctl_start) { assert(NULL != prev_rctl); prev_rctl->next_rctl = rctl_start; rctl_start->prev_rctl = prev_rctl; } rctl = rctl_start; regcnt_stuck = 0; /* # of regions we are stuck in waiting for other regions to resolve a multi-region TP transaction */ assert((NULL == rctl) || (NULL == rctl->forw_multi)); gv_cur_region = NULL; /* clear out any previous value to ensure gv_cur_region/cs_addrs/cs_data * all get set in sync by the MUR_CHANGE_REG macro below. */ /* Phase2 of forward recovery starts */ while (NULL != rctl) { /* while there is at least one region remaining with unprocessed journal records */ assert(NULL != rctl_start); assert(0 < murgbl.regcnt_remaining); if (NULL != rctl->forw_multi) { /* This region's current journal record is part of a TP transaction waiting for other regions */ regcnt_stuck++; assert(regcnt_stuck <= murgbl.regcnt_remaining); if (regcnt_stuck == murgbl.regcnt_remaining) { assertpro(multi_proc_in_use); /* Else : Out-of-design situation. Stuck in ALL regions. */ /* Check one last time if all regions are stuck waiting for another process to resolve the * multi-region TP transaction. If so, wait in a sleep loop. If not, we can proceed. */ rctl = rctl_start; start_hrtbt_cntr = heartbeat_counter; do { if (IS_FORCED_MULTI_PROC_EXIT(mp_hdr)) { /* We are at a logical point. So exit if signaled by parent */ status = ERR_FORCEDHALT; goto finish; } forw_multi = rctl->forw_multi; assert(NULL != forw_multi); sfm = forw_multi->shm_forw_multi; assert(NULL != sfm); assert(sfm->num_reg_seen_forward <= sfm->num_reg_seen_backward); # ifdef MUR_DEBUG fprintf(stderr, "Pid = %d : Line %d : token = %llu : forward = %d : backward = %d\n", process_id, __LINE__, (long long int)sfm->token, sfm->num_reg_seen_forward, sfm->num_reg_seen_backward); # endif if (sfm->num_reg_seen_forward == sfm->num_reg_seen_backward) { /* We are no longer stuck in this region */ assert(!forw_multi->no_longer_stuck); forw_multi->no_longer_stuck = TRUE; break; } rctl = rctl->next_rctl; /* Move on to the next available region */ assert(NULL != rctl); if (rctl == rctl_start) { /* We went through all regions once and are still stuck. * Sleep until at leat TWO heartbeats have elapsed after which check for deadlock. * Do this only in the child process that owns the FIRST region in the region list. * This way we dont have contention for the GRAB_MULTI_PROC_LATCH from * all children at more or less the same time. */ if ((rctl == mur_ctl) && (heartbeat_counter > (start_hrtbt_cntr + 2))) { /* Check if all processes are stuck for a while. If so assertpro */ GRAB_MULTI_PROC_LATCH_IF_NEEDED(release_latch); assert(release_latch); shm_rctl_start = mur_shm_hdr->shm_rctl_start; num_reg_stuck = 0; for (i = 0; i < murgbl.reg_total; i++) { shm_rctl = &shm_rctl_start[i]; sfm = shm_rctl->shm_forw_multi; if (NULL != sfm) { if (sfm->num_reg_seen_forward != sfm->num_reg_seen_backward) num_reg_stuck++; } } REL_MULTI_PROC_LATCH_IF_NEEDED(release_latch); /* If everyone is stuck at this point, it is an out-of-design situation */ assertpro(num_reg_stuck < murgbl.reg_total); start_hrtbt_cntr = heartbeat_counter; } else { /* Sleep and recheck if any region we are stuck in got resolved. * To minimize time spent sleeping, we just yield our timeslice. */ rel_quant(); continue; } } } while (TRUE); } else { rctl = rctl->next_rctl; /* Move on to the next available region */ assert(NULL != rctl); continue; } } regcnt_stuck = 0; /* restart the counter now that we found at least one non-stuck region */ MUR_CHANGE_REG(rctl); jctl = rctl->jctl; this_reg_stuck = FALSE; for ( status = SS_NORMAL; SS_NORMAL == status; ) { if (multi_proc && IS_FORCED_MULTI_PROC_EXIT(mp_hdr)) { /* We are at a logical point. So exit if signaled by parent */ status = ERR_FORCEDHALT; goto finish; } assert(jctl == rctl->jctl); rec = rctl->mur_desc->jnlrec; rec_time = rec->prefix.time; if (rec_time > mur_options.before_time) break; /* Records after -BEFORE_TIME do not go to extract or losttrans or brkntrans files */ assert((rec_time >= adjusted_resolve_time) || (mur_options.notncheck && !mur_options.verify)); assert((0 == mur_options.after_time) || (mur_options.forward && !rctl->db_updated)); if (rec_time < mur_options.after_time) { status = mur_next_rec(&jctl); continue; /* Records before -AFTER_TIME do not go to extract or losttrans or brkntrans files */ } /* Check if current journal record can be played right away or need to wait for corresponding journal * records from other participating TP regions to be reached. A non-TP or ZTP transaction can be played * without issues (i.e. has no dependencies with any other regions). A single-region TP transaction too * falls in the same category. A multi-region TP transaction needs to wait until all participating regions * have played all journal records BEFORE this TP in order to ensure recover plays records in the exact * same order that GT.M performed them in. */ /* If FENCE_NONE is specified, we would not have maintained any multi hashtable in mur_back_process for * broken transaction processing. So we process multi-region TP transactions as multiple single-region * TP transactions in forward phase. */ if (FENCE_NONE != mur_options.fences) { rectype = (enum jnl_record_type)rec->prefix.jrec_type; if (IS_TP(rectype) && IS_TUPD(rectype)) { assert(IS_SET_KILL_ZKILL_ZTWORM_LGTRIG_ZTRIG(rectype)); assert(&rec->jrec_set_kill.num_participants == &rec->jrec_ztworm.num_participants); assert(&rec->jrec_set_kill.num_participants == &rec->jrec_lgtrig.num_participants); num_partners = rec->jrec_set_kill.num_participants; assert(0 < num_partners); if (1 < num_partners) { this_reg_stuck = TRUE; assert(&rec->jrec_set_kill.update_num == &rec->jrec_ztworm.update_num); assert(&rec->jrec_set_kill.update_num == &rec->jrec_lgtrig.update_num); } } } if (this_reg_stuck) { rec_token_seq = GET_JNL_SEQNO(rec); MUR_FORW_TOKEN_LOOKUP(forw_multi, rec_token_seq, rec_time); if (NULL != forw_multi) { /* This token has already been seen in another region in forward processing. * Add current region as well. If all regions have been resolved, then play * the entire transaction maintaining the exact same order of updates within. */ if (!forw_multi->no_longer_stuck) MUR_FORW_TOKEN_ONE_MORE_REG(forw_multi, rctl); } else { /* First time we are seeing this token in forward processing. Check if this * has already been determined to be a broken transaction. */ recstat = GOOD_TN; multi = NULL; if (IS_REC_POSSIBLY_BROKEN(rec_time, rec_token_seq)) { multi = MUR_TOKEN_LOOKUP(rec_token_seq, rec_time, TPFENCE); if ((NULL != multi) && (0 < multi->partner)) recstat = BROKEN_TN; } MUR_FORW_TOKEN_ADD(forw_multi, rec_token_seq, rec_time, rctl, num_partners, recstat, multi); } /* Check that "tabent" field has been initialized above (by either the MUR_FORW_TOKEN_LOOKUP * or MUR_FORW_TOKEN_ADD macros). This is relied upon by "mur_forward_play_multireg_tp" below. */ assert(NULL != forw_multi->u.tabent); assert(forw_multi->num_reg_seen_forward <= forw_multi->num_reg_seen_backward); if (multi_proc) { sfm = forw_multi->shm_forw_multi; ok_to_play = (NULL == sfm) || (sfm->num_reg_seen_forward == sfm->num_reg_seen_backward); } else ok_to_play = (forw_multi->num_reg_seen_forward == forw_multi->num_reg_seen_backward); assert(ok_to_play || !forw_multi->no_longer_stuck); if (ok_to_play ) { /* We have enough information to proceed with playing this multi-region TP in * forward processing (even if we might not have seen all needed regions). Now play it. * Note that the TP could be BROKEN_TN or GOOD_TN. The callee handles it. */ assert(forw_multi == rctl->forw_multi); status = mur_forward_play_multireg_tp(forw_multi, rctl); this_reg_stuck = FALSE; /* Note that as part of playing the TP transaction, we could have reached * the EOF of rctl. In this case, we need to break out of the loop. */ if ((SS_NORMAL != status) || rctl->forw_eof_seen) break; assert(NULL == rctl->forw_multi); assert(!dollar_tlevel); jctl = rctl->jctl; /* In case the first record after the most recently processed * TP transaction is in the next generation journal file */ continue; } break; } else { status = mur_forward_play_cur_jrec(rctl); if (SS_NORMAL != status) break; } assert(!this_reg_stuck); status = mur_next_rec(&jctl); } assert((NULL == rctl->forw_multi) || this_reg_stuck); assert((NULL != rctl->forw_multi) || !this_reg_stuck); if (!this_reg_stuck) { /* We are not stuck in this region (to resolve a multi-region TP). * This means we are done processing all the records of this region. */ assert(NULL == rctl->forw_multi); if (!rctl->forw_eof_seen) { CHECK_IF_EOF_REACHED(rctl, status); /* sets rctl->forw_eof_seen if needed; resets "status" to SS_NORMAL */ if (SS_NORMAL != status) { assert(ERR_FILENOTCREATE == status); goto finish; } assert(!dollar_tlevel); DELETE_RCTL_FROM_UNPROCESSED_LIST(rctl); /* since all of its records should have been processed */ } else { /* EOF was seen in rctl inside "mur_forward_play_multireg_tp" and it was removed * from the unprocessed list of rctls. At the time rctl was removed, its "next_rctl" * field could have been pointing to another <rctl> that has since then also been * removed inside the same function. Therefore the "next_rctl" field is not reliable * in this case but instead we should rely on the global variable "rctl_start" which * points to the list of unprocessed rctls. Set "next_rctl" accordingly. */ rctl->next_rctl = rctl_start; if (ERR_JNLREADEOF == status) status = SS_NORMAL; } assert(rctl->deleted_from_unprocessed_list); } assert(SS_NORMAL == status); assert(!this_reg_stuck || !rctl->forw_eof_seen); assert((NULL == rctl->next_rctl) || (NULL != rctl_start)); assert((NULL == rctl->next_rctl) || (0 < murgbl.regcnt_remaining)); rctl = rctl->next_rctl; /* Note : even though "rctl" could have been deleted from the doubly linked list above, * rctl->next_rctl is not touched so we can still use it to get to the next element. */ } assert(0 == murgbl.regcnt_remaining); jgbl.mur_pini_addr_reset_fnptr = NULL; /* No more simulation of GT.M activity for any region */ prc_vec = murgbl.prc_vec; /* Use process-vector of MUPIP RECOVER (not any simulating GT.M process) now onwards */ assert(0 == dollar_tlevel); for (rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++) { if (!rctl->this_pid_is_owner) { assert(multi_proc_in_use); continue; /* in a parallel processing environment, process only regions we own */ } if (multi_proc) { /* Set key to print this rctl's region-name as prefix in case this forked off process prints any output */ MUR_SET_MULTI_PROC_KEY(rctl, multi_proc_key); } PRINT_VERBOSE_STAT(rctl->jctl, "mur_forward:at the end"); assert(!mur_options.rollback || (0 != murgbl.consist_jnl_seqno)); assert(mur_options.rollback || (0 == murgbl.consist_jnl_seqno)); assert(!dollar_tlevel); /* In case it applied a broken TUPD */ assert(murgbl.ok_to_update_db || !rctl->db_updated); rctl->mur_plst = NULL; /* reset now that simulation of GT.M updates is done */ /* Ensure mur_block_count_correct is called if updates allowed */ if (murgbl.ok_to_update_db && (SS_NORMAL != mur_block_count_correct(rctl))) { gtm_putmsg_csa(CSA_ARG(rctl->csa) VARLSTCNT(4) ERR_BLKCNTEDITFAIL, 2, DB_LEN_STR(rctl->gd)); murgbl.wrn_count++; } } finish: if (multi_proc) multi_proc_key = NULL; /* reset key until it can be set to rctl's region-name again */ if ((SS_NORMAL == status) && mur_options.show) mur_output_show(); if (NULL != first_shm_rctl) { /* Transfer needed process-private information to shared memory so parent process can later inherit this. */ first_shm_rctl->err_cnt = murgbl.err_cnt; first_shm_rctl->wrn_count = murgbl.wrn_count; first_shm_rctl->consist_jnl_seqno = murgbl.consist_jnl_seqno; /* If extract files were created by this process for one or more regions, then copy that information to * shared memory so parent process can use this information to do a merge sort. */ shm_rctl = mur_shm_hdr->shm_rctl_start; for (rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++, shm_rctl++) { assert(multi_proc_in_use); if (!rctl->this_pid_is_owner) continue; /* in a parallel processing environment, process only regions we own */ /* Cancel any flush/dbsync timers by this child process for this region. This is because the * child is not going to go through exit handling code (no gds_rundown etc.). And we need to * clear up csa->nl->wcs_timers. (normally done by gds_rundown). */ if (NULL != rctl->csa) /* rctl->csa can be NULL in case of "mupip journal -extract" etc. */ CANCEL_DB_TIMERS(rctl->gd, rctl->csa, cancelled_timer, cancelled_dbsync_timer); reccnt = 0; for (size_ptr = &rctl->jnlext_multi_list_size[0], recstat = 0; recstat < TOT_EXTR_TYPES; recstat++, size_ptr++) { /* Assert "extr_file_created" information is in sync between rctl and shm_rctl. * This was done at the end of "mur_cre_file_extfmt". */ assert(shm_rctl->extr_file_created[recstat] == rctl->extr_file_created[recstat]); /* Assert that if *size_ptr is non-zero, then we better have created an extract file. * Note that the converse is not true. It is possible we created a file for example to * write an INCTN record but decided to not write anything because it was not a -detail * type of extract. So *sizeptr could be 0 even though we created the extract file. */ assert(!*size_ptr || rctl->extr_file_created[recstat]); shm_rctl->jnlext_list_size[recstat] = *size_ptr; reccnt += *size_ptr; } assert(INVALID_SHMID == shm_rctl->jnlext_shmid); shm_size = reccnt * SIZEOF(jnlext_multi_t); /* If we are quitting because of an abnormal status OR a forced signal to terminate * OR if the parent is dead (kill -9) dont bother creating shmid to communicate back with parent. */ if (mp_hdr->parent_pid != getppid()) { SET_FORCED_MULTI_PROC_EXIT; /* Also signal sibling children to stop processing */ if (SS_NORMAL != status) status = ERR_FORCEDHALT; } if ((SS_NORMAL == status) && shm_size) { shmid = shmget(IPC_PRIVATE, shm_size, 0600 | IPC_CREAT); if (-1 == shmid) { save_errno = errno; SNPRINTF(errstr, SIZEOF(errstr), "shmget() : shmsize=0x%llx", shm_size); MUR_SET_MULTI_PROC_KEY(rctl, multi_proc_key); /* to print region name prefix */ rts_error_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_SYSCALL, 5, LEN_AND_STR(errstr), CALLFROM, save_errno); } shmPtr = (char *)do_shmat(shmid, 0, 0); if (-1 == (sm_long_t)shmPtr) { save_errno = errno; SNPRINTF(errstr, SIZEOF(errstr), "shmat() : shmid=%d shmsize=0x%llx", shmid, shm_size); MUR_SET_MULTI_PROC_KEY(rctl, multi_proc_key); /* to print region name prefix */ rts_error_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_SYSCALL, 5, LEN_AND_STR(errstr), CALLFROM, save_errno); } shm_rctl->jnlext_shmid = shmid; shm_rctl->jnlext_shm_size = shm_size; for (size_ptr = &rctl->jnlext_multi_list_size[0], recstat = 0; recstat < TOT_EXTR_TYPES; recstat++, size_ptr++) { shm_size = *size_ptr; if (shm_size) { copy_size = copy_list_to_buf(rctl->jnlext_multi_list[recstat], (int4)shm_size, shmPtr); assert(copy_size == (shm_size * SIZEOF(jnlext_multi_t))); shmPtr += copy_size; } } } } } mur_close_file_extfmt(IN_MUR_CLOSE_FILES_FALSE); /* Need to flush buffered extract/losttrans/brokentrans files */ return (int)status; }
uint4 mur_forward(jnl_tm_t min_broken_time, seq_num min_broken_seqno, seq_num losttn_seqno) { boolean_t added, this_reg_stuck; boolean_t is_set_kill_zkill_ztworm, is_set_kill_zkill; jnl_record *rec; enum jnl_record_type rectype; enum rec_fence_type rec_fence; enum broken_type recstat; jnl_tm_t rec_time; int4 rec_image_count = 0; /* This is a dummy variable for UNIX */ uint4 status, regcnt_stuck, num_partners; mval mv; reg_ctl_list *rctl, *rctl_top, *prev_rctl; jnl_ctl_list *jctl; gd_region *reg; sgmnt_addrs *csa; seq_num rec_token_seq; forw_multi_struct *forw_multi; multi_struct *multi; error_def(ERR_JNLREADEOF); error_def(ERR_BLKCNTEDITFAIL); skip_dbtriggers = TRUE; /* do not want to invoke any triggers for updates done by journal recovery */ murgbl.extr_buff = (char *)malloc(murgbl.max_extr_record_length); for (recstat = (enum broken_type)0; recstat < TOT_EXTR_TYPES; recstat++) murgbl.extr_file_create[recstat] = TRUE; jgbl.dont_reset_gbl_jrec_time = jgbl.forw_phase_recovery = TRUE; assert(NULL == jgbl.mur_pini_addr_reset_fnptr); jgbl.mur_pini_addr_reset_fnptr = (pini_addr_reset_fnptr)mur_pini_addr_reset; gv_keysize = DBKEYSIZE(MAX_KEY_SZ); mu_gv_stack_init(); murgbl.consist_jnl_seqno = 0; /* Note down passed in values in murgbl global so "mur_forward_play_cur_jrec" function can see it as well */ murgbl.min_broken_time = min_broken_time; murgbl.min_broken_seqno = min_broken_seqno; murgbl.losttn_seqno = losttn_seqno; assert(!mur_options.rollback || (murgbl.losttn_seqno <= murgbl.min_broken_seqno)); prev_rctl = NULL; rctl_start = NULL; assert(0 == murgbl.regcnt_remaining); for (rctl = mur_ctl, rctl_top = mur_ctl + murgbl.reg_total; rctl < rctl_top; rctl++) { if (mur_options.forward) { assert(NULL == rctl->jctl_turn_around); jctl = rctl->jctl = rctl->jctl_head; assert(jctl->reg_ctl == rctl); jctl->rec_offset = JNL_HDR_LEN; jnl_fence_ctl.fence_list = JNL_FENCE_LIST_END; /* initialized to reflect journaling is not enabled */ } else { jctl = rctl->jctl = (NULL == rctl->jctl_turn_around) ? rctl->jctl_head : rctl->jctl_turn_around; assert(jctl->reg_ctl == rctl); jctl->rec_offset = jctl->turn_around_offset; jgbl.mur_jrec_seqno = jctl->turn_around_seqno; if (mur_options.rollback && murgbl.consist_jnl_seqno < jgbl.mur_jrec_seqno) murgbl.consist_jnl_seqno = jgbl.mur_jrec_seqno; assert(murgbl.consist_jnl_seqno <= murgbl.losttn_seqno); assert((NULL != rctl->jctl_turn_around) || (0 == jctl->rec_offset)); } if (mur_options.update || mur_options.extr[GOOD_TN]) { reg = rctl->gd; gv_cur_region = reg; tp_change_reg(); /* note : sets cs_addrs to non-NULL value even if gv_cur_region->open is FALSE * (cs_data could still be NULL). */ rctl->csa = cs_addrs; cs_addrs->rctl = rctl; rctl->csd = cs_data; rctl->sgm_info_ptr = cs_addrs->sgm_info_ptr; SET_CSA_DIR_TREE(cs_addrs, MAX_KEY_SZ, reg); gv_target = cs_addrs->dir_tree; } jctl->after_end_of_data = FALSE; status = mur_next(jctl, jctl->rec_offset); assert(ERR_JNLREADEOF != status); /* cannot get EOF at start of forward processing */ if (SS_NORMAL != status) return status; PRINT_VERBOSE_STAT(jctl, "mur_forward:at the start"); /* Any multi-region TP transaction will be processed as multiple single-region TP transactions up * until the tp-resolve-time is reached. From then on, they will be treated as one multi-region TP * transaction. This is needed for proper lost-tn determination (any multi-region transaction that * gets played in a region AFTER it has already encountered a broken tn should treat this as a lost tn). */ do { assert(jctl == rctl->jctl); rec = rctl->mur_desc->jnlrec; rec_time = rec->prefix.time; if (rec_time > mur_options.before_time) break; /* Records after -BEFORE_TIME do not go to extract or losttrans or brkntrans files */ if (rec_time < mur_options.after_time) { status = mur_next_rec(&jctl); continue; /* Records before -AFTER_TIME do not go to extract or losttrans or brkntrans files */ } if (rec_time >= jgbl.mur_tp_resolve_time) break; /* Records after tp-resolve-time will be processed below */ /* TODO: what do we do if we find a BROKEN tn here? */ status = mur_forward_play_cur_jrec(rctl); if (SS_NORMAL != status) break; status = mur_next_rec(&jctl); } while (SS_NORMAL == status); CHECK_IF_EOF_REACHED(rctl, status); /* sets rctl->forw_eof_seen if needed; resets "status" to SS_NORMAL */ if (SS_NORMAL != status) return status; if (rctl->forw_eof_seen) { PRINT_VERBOSE_STAT(jctl, "mur_forward:Reached EOF before tp_resolve_time"); continue; /* Reached EOF before even getting to tp_resolve_time. * Do not even consider region for next processing loop */ } rctl->last_tn = 0; rctl->process_losttn = FALSE; murgbl.regcnt_remaining++; /* # of regions participating in recovery at this point */ if (NULL == rctl_start) rctl_start = rctl; if (NULL != prev_rctl) { prev_rctl->next_rctl = rctl; rctl->prev_rctl = prev_rctl; } prev_rctl = rctl; assert(murgbl.ok_to_update_db || !rctl->db_updated); PRINT_VERBOSE_STAT(jctl, "mur_forward:at tp_resolve_time"); } /* Note that it is possible for rctl_start to be NULL at this point. That is there is no journal record in any region * AFTER the calculated tp-resolve-time. This is possible if for example -AFTER_TIME was used and has a time later * than any journal record in all journal files. If rctl_start is NULL, prev_rctl should also be NULL and vice versa. */ if (prev_rctl != rctl_start) { assert(NULL != prev_rctl); assert(NULL != rctl_start); prev_rctl->next_rctl = rctl_start; rctl_start->prev_rctl = prev_rctl; } else { /* prev_rctl & rctl_start are identical. They both should be NULL or should point to a single element linked list */ assert((NULL == rctl_start) || (NULL == rctl_start->next_rctl) && (NULL == rctl_start->prev_rctl)); } rctl = rctl_start; regcnt_stuck = 0; /* # of regions we are stuck in waiting for other regions to resolve a multi-region TP transaction */ assert((NULL == rctl) || (NULL == rctl->forw_multi)); gv_cur_region = NULL; /* clear out any previous value to ensure gv_cur_region/cs_addrs/cs_data * all get set in sync by the MUR_CHANGE_REG macro below. */ while (NULL != rctl) { /* while there is at least one region remaining with unprocessed journal records */ assert(NULL != rctl_start); assert(0 < murgbl.regcnt_remaining); if (NULL != rctl->forw_multi) { /* This region's current journal record is part of a TP transaction waiting for other regions */ regcnt_stuck++; if (regcnt_stuck >= murgbl.regcnt_remaining) GTMASSERT; /* Out-of-design situation. Stuck in ALL regions. */ rctl = rctl->next_rctl; /* Move on to the next available region */ assert(NULL != rctl); continue; } regcnt_stuck = 0; /* restart the counter now that we found at least one non-stuck region */ MUR_CHANGE_REG(rctl); jctl = rctl->jctl; this_reg_stuck = FALSE; for ( status = SS_NORMAL; SS_NORMAL == status; ) { assert(jctl == rctl->jctl); rec = rctl->mur_desc->jnlrec; rec_time = rec->prefix.time; assert(rec_time >= jgbl.mur_tp_resolve_time); if (rec_time > mur_options.before_time) break; /* Records after -BEFORE_TIME do not go to extract or losttrans or brkntrans files */ assert((0 == mur_options.after_time) || mur_options.forward && !rctl->db_updated); if (rec_time < mur_options.after_time) { status = mur_next_rec(&jctl); continue; /* Records before -AFTER_TIME do not go to extract or losttrans or brkntrans files */ } /* Check if current journal record can be played right away or need to wait for corresponding journal * records from other participating TP regions to be reached. A non-TP or ZTP transaction can be played * without issues (i.e. has no dependencies with any other regions). A single-region TP transaction too * falls in the same category. A multi-region TP transaction needs to wait until all participating regions * have played all journal records BEFORE this TP in order to ensure recover plays records in the exact * same order that GT.M performed them in. */ /* If FENCE_NONE is specified, we would not have maintained any multi hashtable in mur_back_process for * broken transaction processing. So we process multi-region TP transactions as multiple single-region * TP transactions in forward phase. */ if (FENCE_NONE != mur_options.fences) { rectype = (enum jnl_record_type)rec->prefix.jrec_type; if (IS_TP(rectype) && IS_TUPD(rectype)) { assert(IS_SET_KILL_ZKILL_ZTRIG_ZTWORM(rectype)); assert(&rec->jrec_set_kill.num_participants == &rec->jrec_ztworm.num_participants); num_partners = rec->jrec_set_kill.num_participants; assert(0 < num_partners); if (1 < num_partners) { this_reg_stuck = TRUE; assert(&rec->jrec_set_kill.update_num == &rec->jrec_ztworm.update_num); } } } if (this_reg_stuck) { rec_token_seq = GET_JNL_SEQNO(rec); VMS_ONLY( /* In VMS, pid is not unique. We need "image_count" as well. But this is not needed * in case of rollback as the token is guaranteed to be unique in that case. */ if (!mur_options.rollback) { MUR_GET_IMAGE_COUNT(jctl, rec, rec_image_count, status); if (SS_NORMAL != status) { this_reg_stuck = FALSE; /* so abnormal "status" gets checked below */ break; } } ) /* In Unix, "rec_image_count" is ignored by the MUR_FORW* macros */ MUR_FORW_TOKEN_LOOKUP(forw_multi, rec_token_seq, rec_time, rec_image_count); if (NULL != forw_multi) { /* This token has already been seen in another region in forward processing. * Add current region as well. If all regions have been resolved, then play * the entire transaction maintaining the exact same order of updates within. */ MUR_FORW_TOKEN_ONE_MORE_REG(forw_multi, rctl); } else { /* First time we are seeing this token in forward processing. Check if this * has already been determined to be a broken transaction. */ recstat = GOOD_TN; multi = NULL; if (IS_REC_POSSIBLY_BROKEN(rec_time, rec_token_seq)) { multi = MUR_TOKEN_LOOKUP(rec_token_seq, rec_image_count, rec_time, TPFENCE); if ((NULL != multi) && (0 < multi->partner)) recstat = BROKEN_TN; } MUR_FORW_TOKEN_ADD(forw_multi, rec_token_seq, rec_time, rctl, num_partners, recstat, multi, rec_image_count); } /* Check that "tabent" field has been initialized above (by either the MUR_FORW_TOKEN_LOOKUP * or MUR_FORW_TOKEN_ADD macros). This is relied upon by "mur_forward_play_multireg_tp" below. */ assert(NULL != forw_multi->u.tabent); assert(forw_multi->num_reg_seen_forward <= forw_multi->num_reg_seen_backward); if (forw_multi->num_reg_seen_forward == forw_multi->num_reg_seen_backward) { /* All regions have been seen in forward processing. Now play it. * Note that the TP could be BROKEN_TN or GOOD_TN. The callee handles it. */ assert(forw_multi == rctl->forw_multi); status = mur_forward_play_multireg_tp(forw_multi, rctl); this_reg_stuck = FALSE; /* Note that as part of playing the TP transaction, we could have reached * the EOF of rctl. In this case, we need to break out of the loop. */ if ((SS_NORMAL != status) || rctl->forw_eof_seen) break; assert(NULL == rctl->forw_multi); assert(!dollar_tlevel); jctl = rctl->jctl; /* In case the first record after the most recently processed * TP transaction is in the next generation journal file */ continue; } break; } else { status = mur_forward_play_cur_jrec(rctl); if (SS_NORMAL != status) break; } assert(!this_reg_stuck); status = mur_next_rec(&jctl); }