int continue_proc(pid_t pid)
{
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
DEBUG_ONLY(if (!TREF(gtm_usesecshr))) /* Cause debug builds to talk to gtmsecshr more often */
{
if (WBTEST_ENABLED(WBTEST_HOLD_GTMSOURCE_SRV_LATCH))
{
/* Simulate the kill below, but ignore its return status so that we end up invoking gtmsecshr */
kill(pid, SIGCONT);
/* Wait until the target quits so that kill() call by gtmsecshr fails with ESRCH */
while (is_proc_alive(pid, 0))
LONG_SLEEP(1);
}
else if (0 == kill(pid, SIGCONT))
return 0;
else if (ESRCH == errno)
{
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(5) ERR_NOSUCHPROC, 3, pid, RTS_ERROR_LITERAL("continue"));
return ESRCH;
} else
assert(EINVAL != errno);
}
return send_mesg2gtmsecshr(CONTINUE_PROCESS, pid, NULL, 0);
}
int main(int argc, char *argv[])
{
DCL_THREADGBL_ACCESS;
GTM_THREADGBL_INIT;
set_blocksig();
gtm_imagetype_init(DSE_IMAGE);
gtm_wcswidth_fnptr = gtm_wcswidth;
gtm_env_init(); /* read in all environment variables */
licensed = TRUE;
TREF(transform) = TRUE;
op_open_ptr = op_open;
patch_curr_blk = get_dir_root();
err_init(util_base_ch);
GTM_ICU_INIT_IF_NEEDED; /* Note: should be invoked after err_init (since it may error out) and before CLI parsing */
sig_init(generic_signal_handler, dse_ctrlc_handler, suspsigs_handler);
atexit(util_exit_handler);
SET_LATCH_GLOBAL(&defer_latch, LOCK_AVAILABLE);
get_page_size();
stp_init(STP_INITSIZE);
rts_stringpool = stringpool;
getjobname();
INVOKE_INIT_SECSHR_ADDRS;
getzdir();
prealloc_gt_timers();
initialize_pattern_table();
gvinit();
region_init(FALSE);
INIT_GBL_ROOT(); /* Needed for GVT initialization */
getjobnum();
util_out_print("!/File !_!AD", TRUE, DB_LEN_STR(gv_cur_region));
util_out_print("Region!_!AD!/", TRUE, REG_LEN_STR(gv_cur_region));
cli_lex_setup(argc, argv);
CREATE_DUMMY_GBLDIR(gd_header, original_header, gv_cur_region, gd_map, gd_map_top);
gtm_chk_dist(argv[0]);
# ifdef DEBUG
if ((gtm_white_box_test_case_enabled && (WBTEST_SEMTOOLONG_STACK_TRACE == gtm_white_box_test_case_number) ))
{
sgmnt_addrs * csa;
node_local_ptr_t cnl;
csa = &FILE_INFO(gv_cur_region)->s_addrs;
cnl = csa->nl;
cnl->wbox_test_seq_num = 1; /*Signal the first step and wait here*/
while (2 != cnl->wbox_test_seq_num) /*Wait for another process to get hold of the semaphore and signal next step*/
LONG_SLEEP(10);
}
# endif
if (argc < 2)
display_prompt();
io_init(TRUE);
while (1)
{
if (!dse_process(argc))
break;
display_prompt();
}
dse_exit();
REVERT;
}
void deferred_signal_handler(void)
{
void (*signal_routine)();
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
/* To avoid nested calls to this routine, we set forced_exit to FALSE at the very beginning */
forced_exit = FALSE;
if (exit_handler_active)
{
assert(FALSE); /* at this point in time (June 2003) there is no way we know of to get here, hence the assert */
return; /* since anyway we are exiting currently, resume exit handling instead of reissuing another one */
}
/* For signals that get a delayed response so we can get out of crit, we also delay the messages.
* This routine will output those delayed messages from the appropriate structures to both the
* user and the system console.
*/
/* note can't use switch here because ERR_xxx are not defined as constants */
if (ERR_KILLBYSIG == forced_exit_err)
{
send_msg(VARLSTCNT(6) ERR_KILLBYSIG, 4, GTMIMAGENAMETXT(image_type), process_id, signal_info.signal);
gtm_putmsg(VARLSTCNT(6) ERR_KILLBYSIG, 4, GTMIMAGENAMETXT(image_type), process_id, signal_info.signal);
} else if (ERR_KILLBYSIGUINFO == forced_exit_err)
{
send_msg(VARLSTCNT(8) ERR_KILLBYSIGUINFO, 6, GTMIMAGENAMETXT(image_type), process_id,
signal_info.signal, signal_info.send_pid, signal_info.send_uid);
gtm_putmsg(VARLSTCNT(8) ERR_KILLBYSIGUINFO, 6, GTMIMAGENAMETXT(image_type), process_id,
signal_info.signal, signal_info.send_pid, signal_info.send_uid);
} else if (ERR_KILLBYSIGSINFO1 == forced_exit_err)
{
send_msg(VARLSTCNT(8) ERR_KILLBYSIGSINFO1, 6, GTMIMAGENAMETXT(image_type),
process_id, signal_info.signal, signal_info.int_iadr, signal_info.bad_vadr);
gtm_putmsg(VARLSTCNT(8) ERR_KILLBYSIGSINFO1, 6, GTMIMAGENAMETXT(image_type),
process_id, signal_info.signal, signal_info.int_iadr, signal_info.bad_vadr);
} else if (ERR_KILLBYSIGSINFO2 == forced_exit_err)
{
send_msg(VARLSTCNT(7) ERR_KILLBYSIGSINFO2, 5, GTMIMAGENAMETXT(image_type),
process_id, signal_info.signal, signal_info.int_iadr);
gtm_putmsg(VARLSTCNT(7) ERR_KILLBYSIGSINFO2, 5, GTMIMAGENAMETXT(image_type),
process_id, signal_info.signal, signal_info.int_iadr);
} else if (ERR_KILLBYSIGSINFO3 == forced_exit_err)
{
send_msg(VARLSTCNT(7) ERR_KILLBYSIGSINFO3, 5, GTMIMAGENAMETXT(image_type),
process_id, signal_info.signal, signal_info.bad_vadr);
gtm_putmsg(VARLSTCNT(7) ERR_KILLBYSIGSINFO3, 5, GTMIMAGENAMETXT(image_type),
process_id, signal_info.signal, signal_info.bad_vadr);
} else if (ERR_FORCEDHALT != forced_exit_err || !gtm_quiet_halt)
{ /* No HALT messages if quiet halt is requested */
send_msg(VARLSTCNT(1) forced_exit_err);
gtm_putmsg(VARLSTCNT(1) forced_exit_err);
}
assert(OK_TO_INTERRUPT);
/* Signal intent to exit BEFORE driving condition handlers. This avoids checks that will otherwise fail (for example
* if mdb_condition_handler/preemptive_ch gets called below, that could invoke the RESET_GV_TARGET macro which in turn
* would assert that gv_target->gd_csa is equal to cs_addrs. This could not be true in case we were in mainline code
* that was interrupted by the flush timer for a different region which in turn was interrupted by an external signal
* that would drive us to exit. Setting the "process_exiting" variable causes those csa checks to pass.
*/
SET_PROCESS_EXITING_TRUE;
# ifdef DEBUG
if (gtm_white_box_test_case_enabled && (WBTEST_DEFERRED_TIMERS == gtm_white_box_test_case_number)
&& (2 == gtm_white_box_test_case_count))
{
DEFER_INTERRUPTS(INTRPT_NO_TIMER_EVENTS);
DBGFPF((stderr, "DEFERRED_SIGNAL_HANDLER: will sleep for 20 seconds\n"));
LONG_SLEEP(20);
DBGFPF((stderr, "DEFERRED_SIGNAL_HANDLER: done sleeping\n"));
ENABLE_INTERRUPTS(INTRPT_NO_TIMER_EVENTS);
}
# endif
/* If any special routines are registered to be driven on a signal, drive them now */
if ((0 != exi_condition) && (NULL != call_on_signal))
{
signal_routine = call_on_signal;
call_on_signal = NULL; /* So we don't recursively call ourselves */
(*signal_routine)();
}
/* Note, we do not drive create_fatal_error zshow_dmp() in this routine since any deferrable signals are
* by definition not fatal.
*/
exit(-exi_condition);
}
uint4 gdsfilext(uint4 blocks, uint4 filesize, boolean_t trans_in_prog)
{
sm_uc_ptr_t old_base[2], mmap_retaddr;
boolean_t was_crit, is_mm;
int result, save_errno, status;
DEBUG_ONLY(int first_save_errno);
uint4 new_bit_maps, bplmap, map, new_blocks, new_total, max_tot_blks, old_total;
uint4 jnl_status;
gtm_uint64_t avail_blocks, mmap_sz;
off_t new_eof, new_size;
trans_num curr_tn;
unix_db_info *udi;
inctn_opcode_t save_inctn_opcode;
int4 prev_extend_blks_to_upgrd;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
cache_rec_ptr_t cr;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
assert(!IS_DSE_IMAGE);
assert((cs_addrs->nl == NULL) || (process_id != cs_addrs->nl->trunc_pid)); /* mu_truncate shouldn't extend file... */
assert(!process_exiting);
DEBUG_ONLY(old_base[0] = old_base[1] = NULL);
assert(!gv_cur_region->read_only);
udi = FILE_INFO(gv_cur_region);
is_mm = (dba_mm == cs_addrs->hdr->acc_meth);
# if !defined(MM_FILE_EXT_OK)
if (!udi->grabbed_access_sem && is_mm)
return (uint4)(NO_FREE_SPACE); /* should this be changed to show extension not allowed ? */
# endif
/* Both blocks and total blocks are unsigned ints so make sure we aren't asking for huge numbers that will
overflow and end up doing silly things.
*/
assert((blocks <= (MAXTOTALBLKS(cs_data) - cs_data->trans_hist.total_blks)) || WBTEST_ENABLED(WBTEST_FILE_EXTEND_ERROR));
# if defined(__sun) || defined(__hpux)
cs_data->defer_allocate = TRUE;
# endif
if (!blocks && (cs_data->defer_allocate || (TRANS_IN_PROG_TRUE == trans_in_prog)))
return (uint4)(NO_FREE_SPACE); /* should this be changed to show extension not enabled ? */
bplmap = cs_data->bplmap;
/* New total of non-bitmap blocks will be number of current, non-bitmap blocks, plus new blocks desired
* There are (bplmap - 1) non-bitmap blocks per bitmap, so add (bplmap - 2) to number of non-bitmap blocks
* and divide by (bplmap - 1) to get total number of bitmaps for expanded database. (must round up in this
* manner as every non-bitmap block must have an associated bitmap)
* Current number of bitmaps is (total number of current blocks + bplmap - 1) / bplmap.
* Subtract current number of bitmaps from number needed for expanded database to get number of new bitmaps needed.
*/
new_bit_maps = DIVIDE_ROUND_UP(cs_data->trans_hist.total_blks
- DIVIDE_ROUND_UP(cs_data->trans_hist.total_blks, bplmap) + blocks, bplmap - 1)
- DIVIDE_ROUND_UP(cs_data->trans_hist.total_blks, bplmap);
new_blocks = blocks + new_bit_maps;
assert((0 < (int)new_blocks) || (!cs_data->defer_allocate && (0 == new_blocks)));
if (new_blocks + cs_data->trans_hist.total_blks > MAXTOTALBLKS(cs_data))
{
assert(WBTEST_ENABLED(WBTEST_FILE_EXTEND_ERROR));
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(1) ERR_TOTALBLKMAX);
return (uint4)(NO_FREE_SPACE);
}
if (0 != (save_errno = disk_block_available(udi->fd, &avail_blocks, FALSE)))
{
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), save_errno);
rts_error_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), save_errno);
} else
{
if (!(gtmDebugLevel & GDL_IgnoreAvailSpace))
{ /* Bypass this space check if debug flag above is on. Allows us to create a large sparce DB
* in space it could never fit it if wasn't sparse. Needed for some tests.
*/
avail_blocks = avail_blocks / (cs_data->blk_size / DISK_BLOCK_SIZE);
if ((blocks * EXTEND_WARNING_FACTOR) > avail_blocks)
{
if (blocks > (uint4)avail_blocks)
{
if (!INST_FREEZE_ON_NOSPC_ENABLED(cs_addrs))
return (uint4)(NO_FREE_SPACE);
else
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(6) MAKE_MSG_WARNING(ERR_NOSPACEEXT), 4,
DB_LEN_STR(gv_cur_region), new_blocks, (uint4)avail_blocks);
} else
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DSKSPACEFLOW, 3, DB_LEN_STR(gv_cur_region),
(uint4)(avail_blocks - ((new_blocks <= avail_blocks) ? new_blocks : 0)));
}
}
}
# ifdef DEBUG
if (WBTEST_ENABLED(WBTEST_MM_CONCURRENT_FILE_EXTEND) && dollar_tlevel && !MEMCMP_LIT(gv_cur_region->rname, "DEFAULT"))
{
SYSTEM("$gtm_dist/mumps -run $gtm_wbox_mrtn");
assert(1 == cs_addrs->nl->wbox_test_seq_num); /* should have been set by mubfilcpy */
cs_addrs->nl->wbox_test_seq_num = 2; /* signal mupip backup to stop sleeping in mubfilcpy */
}
# endif
/* From here on, we need to use GDSFILEXT_CLNUP before returning to the caller */
was_crit = cs_addrs->now_crit;
assert(!cs_addrs->hold_onto_crit || was_crit);
/* If we are coming from mupip_extend (which gets crit itself) we better have waited for any unfreezes to occur.
* If we are coming from online rollback (when that feature is available), we will come in holding crit and in
* the final retry. In that case too, we expect to have waited for unfreezes to occur in the caller itself.
* Therefore if we are coming in holding crit from MUPIP, we expect the db to be unfrozen so no need to wait for
* freeze.
* If we are coming from GT.M and final retry (in which case we come in holding crit) we expect to have waited
* for any unfreezes (by invoking tp_crit_all_regions) to occur (TP or non-TP) before coming into this
* function. However, there is one exception. In the final retry, if tp_crit_all_regions notices that
* at least one of the participating regions did ONLY READs, it will not wait for any freeze on THAT region
* to complete before grabbing crit. Later, in the final retry, if THAT region did an update which caused
* op_tcommit to invoke bm_getfree->gdsfilext, then we would have come here with a frozen region on which
* we hold crit.
*/
assert(!was_crit || !FROZEN_HARD(cs_data) || (dollar_tlevel && (CDB_STAGNATE <= t_tries)));
/*
* If we are in the final retry and already hold crit, it is possible that csa->nl->wc_blocked is also set to TRUE
* (by a concurrent process in phase2 which encountered an error in the midst of commit and secshr_db_clnup
* finished the job for it). In this case we do NOT want to invoke wcs_recover as that will update the "bt"
* transaction numbers without correspondingly updating the history transaction numbers (effectively causing
* a cdb_sc_blkmod type of restart). Therefore do NOT call grab_crit (which unconditionally invokes wcs_recover)
* if we already hold crit.
*/
if (!was_crit)
{
for ( ; ; )
{
grab_crit(gv_cur_region);
if (FROZEN_CHILLED(cs_data))
DO_CHILLED_AUTORELEASE(cs_addrs, cs_data);
if (!FROZEN(cs_data) && !IS_REPL_INST_FROZEN)
break;
rel_crit(gv_cur_region);
while (FROZEN(cs_data) || IS_REPL_INST_FROZEN)
{
hiber_start(1000);
if (FROZEN_CHILLED(cs_data) && CHILLED_AUTORELEASE(cs_data))
break;
}
}
} else if (FROZEN_HARD(cs_data) && dollar_tlevel)
{ /* We don't want to continue with file extension as explained above. Hence return with an error code which
* op_tcommit will recognize (as a cdb_sc_needcrit/cdb_sc_instancefreeze type of restart) and restart accordingly.
*/
assert(CDB_STAGNATE <= t_tries);
GDSFILEXT_CLNUP;
return (uint4)FINAL_RETRY_FREEZE_PROG;
} else
WAIT_FOR_REGION_TO_UNCHILL(cs_addrs, cs_data);
if (IS_REPL_INST_FROZEN && trans_in_prog)
{
assert(CDB_STAGNATE <= t_tries);
GDSFILEXT_CLNUP;
return (uint4)FINAL_RETRY_INST_FREEZE;
}
assert(cs_addrs->ti->total_blks == cs_data->trans_hist.total_blks);
old_total = cs_data->trans_hist.total_blks;
if (old_total != filesize)
{ /* Somebody else has already extended it, since we are in crit, this is trust-worthy. However, in case of MM,
* we still need to remap the database
*/
assert((old_total > filesize) || !is_mm);
/* For BG, someone else could have truncated or extended - we have no idea */
GDSFILEXT_CLNUP;
return (SS_NORMAL);
}
if (trans_in_prog && SUSPICIOUS_EXTEND)
{
if (!was_crit)
{
GDSFILEXT_CLNUP;
return (uint4)(EXTEND_SUSPECT);
}
/* If free_blocks counter is not ok, then correct it. Do the check again. If still fails, then it means we held
* crit through bm_getfree into gdsfilext and still didn't get it right.
*/
assertpro(!is_free_blks_ctr_ok() && !SUSPICIOUS_EXTEND);
}
if (JNL_ENABLED(cs_data))
{
if (!jgbl.dont_reset_gbl_jrec_time)
SET_GBL_JREC_TIME; /* needed before jnl_ensure_open as that can write jnl records */
jpc = cs_addrs->jnl;
jbp = jpc->jnl_buff;
/* 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(gv_cur_region, cs_addrs);
if (jnl_status)
{
GDSFILEXT_CLNUP;
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(cs_data), DB_LEN_STR(gv_cur_region));
return (uint4)(NO_FREE_SPACE); /* should have better return status */
}
}
if (is_mm)
{
cs_addrs->nl->mm_extender_pid = process_id;
status = wcs_wtstart(gv_cur_region, 0, NULL, NULL);
cs_addrs->nl->mm_extender_pid = 0;
assertpro(SS_NORMAL == status);
old_base[0] = cs_addrs->db_addrs[0];
old_base[1] = cs_addrs->db_addrs[1];
cs_addrs->db_addrs[0] = NULL; /* don't rely on it until the mmap below */
# ifdef _AIX
status = shmdt(old_base[0] - BLK_ZERO_OFF(cs_data->start_vbn));
# else
status = munmap((caddr_t)old_base[0], (size_t)(old_base[1] - old_base[0]));
# endif
if (0 != status)
{
save_errno = errno;
GDSFILEXT_CLNUP;
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(12) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region),
ERR_SYSCALL, 5, LEN_AND_STR(MEM_UNMAP_SYSCALL), CALLFROM, save_errno);
return (uint4)(NO_FREE_SPACE);
}
} else
{ /* Due to concurrency issues, it is possible some process had issued a disk read of the GDS block# corresponding
* to "old_total" right after a truncate wrote a GDS-block of zeros on disk (to signal end of the db file).
* If so, the global buffer containing this block needs to be invalidated now as part of the extend. If not, it is
* possible the EOF block on disk is now going to be overwritten by a properly initialized bitmap block (as part
* of the gdsfilext below) while the global buffer continues to have an incorrect copy of that bitmap block and
* this in turn would cause XXXX failures due to a bad bitmap block in shared memory. (GTM-7519)
*/
cr = db_csh_get((block_id)old_total);
if ((NULL != cr) && ((cache_rec_ptr_t)CR_NOTVALID != cr))
{
assert((0 == cr->dirty) && (0 == cr->bt_index) && !cr->stopped);
cr->cycle++;
cr->blk = CR_BLKEMPTY;
}
}
CHECK_TN(cs_addrs, cs_data, cs_data->trans_hist.curr_tn); /* can issue rts_error TNTOOLARGE */
new_total = old_total + new_blocks;
new_eof = BLK_ZERO_OFF(cs_data->start_vbn) + ((off_t)new_total * cs_data->blk_size);
# if !defined(__sun) && !defined(__hpux)
if (!cs_data->defer_allocate)
{
new_size = new_eof + cs_data->blk_size;
save_errno = posix_fallocate(udi->fd, 0, new_size);
DEBUG_ONLY(first_save_errno = save_errno);
if ((ENOSPC == save_errno) && IS_GTM_IMAGE)
save_errno = extend_wait_for_fallocate(udi, new_size);
if (0 != save_errno)
{
GDSFILEXT_CLNUP;
assert(ENOSPC == save_errno);
if (ENOSPC != save_errno)
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_PREALLOCATEFAIL, 2, DB_LEN_STR(gv_cur_region),
save_errno);
return (uint4)(NO_FREE_SPACE);
}
}
# endif
save_errno = db_write_eof_block(udi, udi->fd, cs_data->blk_size, new_eof, &(TREF(dio_buff)));
if ((ENOSPC == save_errno) && IS_GTM_IMAGE)
save_errno = extend_wait_for_write(udi, cs_data->blk_size, new_eof);
if (0 != save_errno)
{
GDSFILEXT_CLNUP;
if (ENOSPC != save_errno)
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), save_errno);
return (uint4)(NO_FREE_SPACE);
}
if (WBTEST_ENABLED(WBTEST_FILE_EXTEND_INTERRUPT_1))
{
LONG_SLEEP(600);
assert(FALSE);
}
/* Ensure the EOF and metadata get to disk BEFORE any bitmap writes. Otherwise, the file size could no longer reflect
* a proper extent and subsequent invocations of gdsfilext could corrupt the database.
*/
if (!IS_STATSDB_CSA(cs_addrs))
{
GTM_DB_FSYNC(cs_addrs, udi->fd, status);
assert(0 == status);
if (0 != status)
{
GDSFILEXT_CLNUP;
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(8) ERR_DBFILERR, 5,
RTS_ERROR_LITERAL("fsync1()"), CALLFROM, status);
return (uint4)(NO_FREE_SPACE);
}
}
if (WBTEST_ENABLED(WBTEST_FILE_EXTEND_INTERRUPT_2))
{
LONG_SLEEP(600);
assert(FALSE); /* Should be killed before that */
}
DEBUG_ONLY(prev_extend_blks_to_upgrd = cs_data->blks_to_upgrd;)
void mu_int_reg(gd_region *reg, boolean_t *return_value, boolean_t return_after_open)
{
boolean_t read_only, was_crit;
freeze_status status;
node_local_ptr_t cnl;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
sgmnt_data *csd_copy_ptr;
gd_segment *seg;
int gtmcrypt_errno;
# ifdef DEBUG
boolean_t need_to_wait = FALSE;
int trynum;
uint4 curr_wbox_seq_num;
# endif
*return_value = FALSE;
jnlpool_init_needed = TRUE;
ESTABLISH(mu_int_reg_ch);
if (dba_usr == reg->dyn.addr->acc_meth)
{
util_out_print("!/Can't integ region !AD; not GDS format", TRUE, REG_LEN_STR(reg));
mu_int_skipreg_cnt++;
return;
}
gv_cur_region = reg;
if (reg_cmcheck(reg))
{
util_out_print("!/Can't integ region across network", TRUE);
mu_int_skipreg_cnt++;
return;
}
gvcst_init(gv_cur_region);
if (gv_cur_region->was_open)
{ /* already open under another name */
gv_cur_region->open = FALSE;
return;
}
if (return_after_open)
{
*return_value = TRUE;
return;
}
change_reg();
csa = &FILE_INFO(gv_cur_region)->s_addrs;
cnl = csa->nl;
csd = csa->hdr;
read_only = gv_cur_region->read_only;
assert(NULL != mu_int_master);
/* Ensure that we don't see an increase in the file header and master map size compared to it's maximum values */
assert(SGMNT_HDR_LEN >= SIZEOF(sgmnt_data) && (MASTER_MAP_SIZE_MAX >= MASTER_MAP_SIZE(csd)));
/* ONLINE INTEG if asked for explicitly by specifying -ONLINE is an error if the db has partial V4 blocks.
* However, if -ONLINE is not explicitly specified but rather assumed implicitly (as default for -REG)
* then turn off ONLINE INTEG for this region and continue as if -NOONLINE was specified
*/
if (!csd->fully_upgraded)
{
ointeg_this_reg = FALSE; /* Turn off ONLINE INTEG for this region */
if (online_specified)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_SSV4NOALLOW, 2, DB_LEN_STR(gv_cur_region));
util_out_print(NO_ONLINE_ERR_MSG, TRUE);
mu_int_skipreg_cnt++;
return;
}
}
if (!ointeg_this_reg || read_only)
{
status = region_freeze(gv_cur_region, TRUE, FALSE, TRUE, FALSE, !read_only);
switch (status)
{
case REG_ALREADY_FROZEN:
if (csa->read_only_fs)
break;
util_out_print("!/Database for region !AD is already frozen, not integing",
TRUE, REG_LEN_STR(gv_cur_region));
mu_int_skipreg_cnt++;
return;
case REG_FLUSH_ERROR:
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_BUFFLUFAILED, 4, LEN_AND_LIT(MUPIP_INTEG),
DB_LEN_STR(gv_cur_region));
mu_int_skipreg_cnt++;
return;
case REG_HAS_KIP:
/* We have already waited for KIP to reset. This time do not wait for KIP */
status = region_freeze(gv_cur_region, TRUE, FALSE, FALSE, FALSE, !read_only);
if (REG_ALREADY_FROZEN == status)
{
if (csa->read_only_fs)
break;
util_out_print("!/Database for region !AD is already frozen, not integing",
TRUE, REG_LEN_STR(gv_cur_region));
mu_int_skipreg_cnt++;
return;
} else if (REG_FLUSH_ERROR == status)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_BUFFLUFAILED, 4, LEN_AND_LIT(MUPIP_INTEG),
DB_LEN_STR(gv_cur_region));
mu_int_skipreg_cnt++;
return;
}
assert(REG_FREEZE_SUCCESS == status);
/* no break */
case REG_FREEZE_SUCCESS:
break;
default:
assert(FALSE);
/* no break */
}
if (read_only && (dba_bg == csa->hdr->acc_meth) && !mu_int_wait_rdonly(csa, MUPIP_INTEG))
{
mu_int_skipreg_cnt++;
return;
}
}
if (!ointeg_this_reg)
{ /* Take a copy of the file-header. To ensure it is consistent, do it while holding crit. */
was_crit = csa->now_crit;
if (!was_crit)
grab_crit(gv_cur_region);
memcpy((uchar_ptr_t)&mu_int_data, (uchar_ptr_t)csd, SIZEOF(sgmnt_data));
if (!was_crit)
rel_crit(gv_cur_region);
memcpy(mu_int_master, MM_ADDR(csd), MASTER_MAP_SIZE(csd));
csd_copy_ptr = &mu_int_data;
} else
{
if (!ss_initiate(gv_cur_region, util_ss_ptr, &csa->ss_ctx, preserve_snapshot, MUPIP_INTEG))
{
mu_int_skipreg_cnt++;
assert(NULL != csa->ss_ctx);
ss_release(&csa->ss_ctx);
ointeg_this_reg = FALSE; /* Turn off ONLINE INTEG for this region */
assert(process_id != cnl->in_crit); /* Ensure ss_initiate released the crit before returning */
assert(!FROZEN_HARD(csd)); /* Ensure region is unfrozen before returning from ss_initiate */
assert(INTRPT_IN_SS_INITIATE != intrpt_ok_state); /* Ensure ss_initiate released intrpt_ok_state */
return;
}
assert(process_id != cnl->in_crit); /* Ensure ss_initiate released the crit before returning */
assert(INTRPT_IN_SS_INITIATE != intrpt_ok_state); /* Ensure ss_initiate released intrpt_ok_state */
csd_copy_ptr = &csa->ss_ctx->ss_shm_ptr->shadow_file_header;
# if defined(DEBUG)
curr_wbox_seq_num = 1;
cnl->wbox_test_seq_num = curr_wbox_seq_num; /* indicate we took the next step */
GTM_WHITE_BOX_TEST(WBTEST_OINTEG_WAIT_ON_START, need_to_wait, TRUE);
if (need_to_wait) /* wait for them to take next step */
{
trynum = 30; /* given 30 cycles to tell you to go */
while ((curr_wbox_seq_num == cnl->wbox_test_seq_num) && trynum--)
LONG_SLEEP(1);
cnl->wbox_test_seq_num++; /* let them know we took the next step */
assert(trynum);
}
# endif
}
if (USES_ANY_KEY(csd_copy_ptr))
{ /* Initialize mu_int_encrypt_key_handle to be used in mu_int_read */
seg = gv_cur_region->dyn.addr;
INIT_DB_OR_JNL_ENCRYPTION(&mu_int_encr_handles, csd_copy_ptr, seg->fname_len, (char *)seg->fname, gtmcrypt_errno);
if (0 != gtmcrypt_errno)
{
GTMCRYPT_REPORT_ERROR(gtmcrypt_errno, gtm_putmsg, seg->fname_len, seg->fname);
mu_int_skipreg_cnt++;
return;
}
}
*return_value = mu_int_fhead();
REVERT;
return;
}
int gtmrecv_poll_actions1(int *pending_data_len, int *buff_unprocessed, unsigned char *buffp)
{
static int report_cnt = 1;
static int next_report_at = 1;
static boolean_t send_xoff = FALSE;
static boolean_t xoff_sent = FALSE;
static seq_num send_seqno;
static boolean_t log_draining_msg = FALSE;
static boolean_t send_badtrans = FALSE;
static boolean_t send_cmp2uncmp = FALSE;
static boolean_t upd_shut_too_early_logged = FALSE;
static time_t last_reap_time = 0;
repl_msg_t xoff_msg;
repl_badtrans_msg_t bad_trans_msg;
boolean_t alert = FALSE, info = FALSE;
int return_status;
gd_region *region_top;
unsigned char *msg_ptr; /* needed for REPL_{SEND,RECV}_LOOP */
int tosend_len, sent_len, sent_this_iter; /* needed for REPL_SEND_LOOP */
int torecv_len, recvd_len, recvd_this_iter; /* needed for REPL_RECV_LOOP */
int status, poll_dir; /* needed for REPL_{SEND,RECV}_LOOP */
int temp_len, pending_msg_size;
int upd_start_status, upd_start_attempts;
int buffered_data_len;
int upd_exit_status;
seq_num temp_send_seqno;
boolean_t bad_trans_detected = FALSE, onln_rlbk_flg_set = FALSE;
recvpool_ctl_ptr_t recvpool_ctl;
upd_proc_local_ptr_t upd_proc_local;
gtmrecv_local_ptr_t gtmrecv_local;
upd_helper_ctl_ptr_t upd_helper_ctl;
pid_t waitpid_res;
int4 msg_type, msg_len;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
recvpool_ctl = recvpool.recvpool_ctl;
upd_proc_local = recvpool.upd_proc_local;
gtmrecv_local = recvpool.gtmrecv_local;
upd_helper_ctl = recvpool.upd_helper_ctl;
if (SHUTDOWN == gtmrecv_local->shutdown)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Shutdown signalled\n");
gtmrecv_end(); /* Won't return */
}
# ifdef GTM_TLS
/* If we sent a REPL_RENEG_ACK, then we cannot afford to send anymore asynchronous messages (like XOFF_ACK_ME) until we
* receive a REPL_RENEG_COMPLETE from the source server. This ensures that while the source server attempts to do a SSL/TLS
* renegotiation, it doesn't have any application data (like XOFF_ACK_ME) sitting in the pipe.
*/
if (REPLTLS_WAITING_FOR_RENEG_COMPLETE == repl_tls.renegotiate_state)
return STOP_POLL;
# endif
/* Reset report_cnt and next_report_at to 1 when a new upd proc is forked */
if ((1 == report_cnt) || (report_cnt == next_report_at))
{
/* A comment on the usage of NO_SHUTDOWN below for the alert variable. Since upd_proc_local->upd_proc_shutdown is
* a shared memory field (and could be concurrently changed by either the receiver server or the update process),
* we want to make sure it is the same value BEFORE and AFTER checking whether the update process is alive or not.
* If it is not NO_SHUTDOWN (i.e. is SHUTDOWN or NORMAL_SHUTDOWN or ABNORMAL_SHUTDOWN) it has shut down due to
* an external request so we do want to send out a false update-process-is-not-alive alert.
*/
if ((alert = ((NO_SHUTDOWN == upd_proc_local->upd_proc_shutdown) && (SRV_DEAD == is_updproc_alive())
&& (NO_SHUTDOWN == upd_proc_local->upd_proc_shutdown)))
|| (info = (((NORMAL_SHUTDOWN == upd_proc_local->upd_proc_shutdown)
|| (ABNORMAL_SHUTDOWN == upd_proc_local->upd_proc_shutdown)) && (SRV_DEAD == is_updproc_alive()))))
{
if (alert)
repl_log(gtmrecv_log_fp, TRUE, TRUE,
"ALERT : Receiver Server detected that Update Process is not ALIVE\n");
else
repl_log(gtmrecv_log_fp, TRUE, TRUE,
"INFO : Update process not running due to user initiated shutdown\n");
if (1 == report_cnt)
{
send_xoff = TRUE;
recvpool_ctl->old_jnl_seqno = recvpool_ctl->jnl_seqno;
recvpool_ctl->jnl_seqno = 0;
/* Even though we have identified that the update process is NOT alive, a waitpid on the update
* process PID is necessary so that the system doesn't leave any zombie process lying around.
* This is possible since any child process that dies without the parent doing a waitpid on it
* will be defunct unless the parent dies at which point the "init" process takes the role of
* the parent and invokes waitpid to remove the zombies.
* NOTE: It is possible that the update process was killed before the receiver server got a
* chance to record it's PID in the recvpool.upd_proc_local structure. In such a case, don't
* invoke waitpid as that will block us (receiver server) if this instance of the receiver
* server was started with helper processes.
*/
if (0 < upd_proc_local->upd_proc_pid)
{
WAITPID(upd_proc_local->upd_proc_pid, &upd_exit_status, 0, waitpid_res);
/* Since the update process as part of its shutdown does NOT reset the upd_proc_pid, reset
* it here ONLY if the update process was NOT kill -9ed. This is needed because receiver
* server as part of its shutdown relies on this field (upd_proc_pid) to determine if the
* update process was cleanly shutdown or was kill -9ed.
*/
if (!alert)
upd_proc_local->upd_proc_pid = 0;
}
upd_proc_local->bad_trans = FALSE; /* No point in doing bad transaction processing */
upd_proc_local->onln_rlbk_flg = FALSE; /* No point handling online rollback */
}
gtmrecv_wait_for_jnl_seqno = TRUE;
REPL_DPRINT1(
"gtmrecv_poll_actions : Setting gtmrecv_wait_for_jnl_seqno to TRUE because of upd crash/shutdown\n");
next_report_at *= GTMRECV_NEXT_REPORT_FACTOR;
report_cnt++;
}
} else
report_cnt++;
/* Check if REPL_CMP2UNCMP or REPL_BADTRANS message needs to be sent */
if (upd_proc_local->onln_rlbk_flg)
{ /* Update process detected an online rollback and is requesting us to restart the connection. But before that, send
* REPL_XOFF source side and drain the replication pipe
*/
onln_rlbk_flg_set = TRUE;
send_xoff = TRUE;
} else if (!send_cmp2uncmp && gtmrecv_send_cmp2uncmp)
{
send_xoff = TRUE;
send_seqno = recvpool_ctl->jnl_seqno;
send_cmp2uncmp = TRUE;
} else if (!send_badtrans && upd_proc_local->bad_trans)
{
send_xoff = TRUE;
send_seqno = upd_proc_local->read_jnl_seqno;
send_badtrans = TRUE;
bad_trans_detected = TRUE;
} else if (!upd_proc_local->bad_trans && send_badtrans && 1 != report_cnt)
{
send_badtrans = FALSE;
bad_trans_detected = FALSE;
}
if (send_xoff && !xoff_sent)
{ /* Send XOFF_ACK_ME if the receiver has a connection to the source. Do not attempt to send it if we dont even
* know the endianness of the remote side. In that case, we are guaranteed no initial handshake occurred and
* so no point sending the XOFF too. This saves us lots of trouble in case of cross-endian replication connections.
*/
assert((FD_INVALID != gtmrecv_sock_fd) || repl_connection_reset);
if ((FD_INVALID != gtmrecv_sock_fd) && remote_side->endianness_known)
{
send_seqno = upd_proc_local->read_jnl_seqno;
if (!remote_side->cross_endian)
{
xoff_msg.type = REPL_XOFF_ACK_ME;
xoff_msg.len = MIN_REPL_MSGLEN;
memcpy((uchar_ptr_t)&xoff_msg.msg[0], (uchar_ptr_t)&send_seqno, SIZEOF(seq_num));
} else
{
xoff_msg.type = GTM_BYTESWAP_32(REPL_XOFF_ACK_ME);
xoff_msg.len = GTM_BYTESWAP_32(MIN_REPL_MSGLEN);
temp_send_seqno = GTM_BYTESWAP_64(send_seqno);
memcpy((uchar_ptr_t)&xoff_msg.msg[0], (uchar_ptr_t)&temp_send_seqno, SIZEOF(seq_num));
}
REPL_SEND_LOOP(gtmrecv_sock_fd, &xoff_msg, MIN_REPL_MSGLEN, REPL_POLL_NOWAIT)
; /* Empty Body */
if (SS_NORMAL != status)
{
if (REPL_CONN_RESET(status) && EREPL_SEND == repl_errno)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Connection reset while sending XOFF_ACK_ME. "
"Status = %d ; %s\n", status, STRERROR(status));
repl_close(>mrecv_sock_fd);
repl_connection_reset = TRUE;
xoff_sent = FALSE;
send_badtrans = FALSE;
} else if (EREPL_SEND == repl_errno)
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error sending XOFF msg due to BAD_TRANS or UPD crash/shutdown. "
"Error in send"), status);
else
{
assert(EREPL_SELECT == repl_errno);
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error sending XOFF msg due to BAD_TRANS or UPD crash/shutdown. "
"Error in select"), status);
}
} else
{
xoff_sent = TRUE;
log_draining_msg = TRUE;
}
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL_XOFF_ACK_ME sent due to upd shutdown/crash or bad trans "
"or ONLINE_ROLLBACK\n");
send_xoff = FALSE;
} else
{ /* Connection has been lost OR initial handshake needs to happen again, so no point sending XOFF/BADTRANS */
send_xoff = FALSE;
send_badtrans = FALSE;
}
}
/* Drain pipe */
if (xoff_sent)
{
if (log_draining_msg)
{ /* avoid multiple logs per instance */
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL INFO - Draining replication pipe due to %s\n",
send_cmp2uncmp ? "CMP2UNCMP" : (send_badtrans ? "BAD_TRANS" :
(onln_rlbk_flg_set ? "ONLINE_ROLLBACK" : "UPD shutdown/crash")));
log_draining_msg = FALSE;
}
if (0 != *buff_unprocessed)
{ /* Throw away the current contents of the buffer */
buffered_data_len = ((*pending_data_len <= *buff_unprocessed) ? *pending_data_len : *buff_unprocessed);
*buff_unprocessed -= buffered_data_len;
buffp += buffered_data_len;
*pending_data_len -= buffered_data_len;
REPL_DPRINT2("gtmrecv_poll_actions : (1) Throwing away %d bytes from old buffer while draining\n",
buffered_data_len);
assert(remote_side->endianness_known); /* only then is remote_side->cross_endian reliable */
while (REPL_MSG_HDRLEN <= *buff_unprocessed)
{
assert(0 == (((unsigned long)buffp) % REPL_MSG_ALIGN));
msg_len = ((repl_msg_ptr_t)buffp)->len;
msg_type = ((repl_msg_ptr_t)buffp)->type;
if (remote_side->cross_endian)
{
msg_len = GTM_BYTESWAP_32(msg_len);
msg_type = GTM_BYTESWAP_32(msg_type);
}
msg_type = (msg_type & REPL_TR_CMP_MSG_TYPE_MASK);
assert((REPL_TR_CMP_JNL_RECS == msg_type) || (0 == (msg_len % REPL_MSG_ALIGN)));
*pending_data_len = ROUND_UP2(msg_len, REPL_MSG_ALIGN);
buffered_data_len = ((*pending_data_len <= *buff_unprocessed) ?
*pending_data_len : *buff_unprocessed);
*buff_unprocessed -= buffered_data_len;
buffp += buffered_data_len;
*pending_data_len -= buffered_data_len;
REPL_DPRINT3("gtmrecv_poll_actions : (1) Throwing away message of "
"type %d and length %d from old buffer while draining\n", msg_type, buffered_data_len);
}
if (0 < *buff_unprocessed)
{
memmove((unsigned char *)gtmrecv_msgp, buffp, *buff_unprocessed);
REPL_DPRINT2("gtmrecv_poll_actions : Incomplete header of length %d while draining\n",
*buff_unprocessed);
}
}
status = SS_NORMAL;
if (0 != *buff_unprocessed || 0 == *pending_data_len)
{ /* Receive the header of a message */
assert(REPL_MSG_HDRLEN > *buff_unprocessed); /* so we dont pass negative length in REPL_RECV_LOOP */
REPL_RECV_LOOP(gtmrecv_sock_fd, ((unsigned char *)gtmrecv_msgp) + *buff_unprocessed,
(REPL_MSG_HDRLEN - *buff_unprocessed), REPL_POLL_WAIT)
; /* Empty Body */
if (SS_NORMAL == status)
{
assert(remote_side->endianness_known); /* only then is remote_side->cross_endian reliable */
if (!remote_side->cross_endian)
{
msg_len = gtmrecv_msgp->len;
msg_type = gtmrecv_msgp->type;
} else
{
msg_len = GTM_BYTESWAP_32(gtmrecv_msgp->len);
msg_type = GTM_BYTESWAP_32(gtmrecv_msgp->type);
}
msg_type = (msg_type & REPL_TR_CMP_MSG_TYPE_MASK);
assert((REPL_TR_CMP_JNL_RECS == msg_type) || (0 == (msg_len % REPL_MSG_ALIGN)));
msg_len = ROUND_UP2(msg_len, REPL_MSG_ALIGN);
REPL_DPRINT3("gtmrecv_poll_actions : Received message of type %d and length %d while draining\n",
msg_type, msg_len);
}
}
if ((SS_NORMAL == status) && (0 != *buff_unprocessed || 0 == *pending_data_len) && (REPL_XOFF_ACK == msg_type))
{ /* Receive the rest of the XOFF_ACK msg and signal the drain as complete */
REPL_RECV_LOOP(gtmrecv_sock_fd, gtmrecv_msgp, (MIN_REPL_MSGLEN - REPL_MSG_HDRLEN), REPL_POLL_WAIT)
; /* Empty Body */
if (SS_NORMAL == status)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE,
"REPL INFO - XOFF_ACK received. Drained replication pipe completely\n");
upd_shut_too_early_logged = FALSE;
xoff_sent = FALSE;
return_status = STOP_POLL;
}
} else if (SS_NORMAL == status)
{ /* Drain the rest of the message */
if (0 < *pending_data_len)
{
pending_msg_size = *pending_data_len;
REPL_DPRINT2("gtmrecv_poll_actions : (2) Throwing away %d bytes from pipe\n", pending_msg_size);
} else
{
pending_msg_size = msg_len - REPL_MSG_HDRLEN;
REPL_DPRINT3("gtmrecv_poll_actions : (2) Throwing away message of "
"type %d and length %d from pipe\n", msg_type, msg_len);
}
for ( ; SS_NORMAL == status && 0 < pending_msg_size; pending_msg_size -= gtmrecv_max_repl_msglen)
{
temp_len = (pending_msg_size < gtmrecv_max_repl_msglen)? pending_msg_size : gtmrecv_max_repl_msglen;
REPL_RECV_LOOP(gtmrecv_sock_fd, gtmrecv_msgp, temp_len, REPL_POLL_WAIT)
; /* Empty Body */
}
*buff_unprocessed = 0; *pending_data_len = 0;
if (SS_NORMAL == status && info && !upd_shut_too_early_logged)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "ALERT : User initiated shutdown of Update Process done "
"when there was data in the replication pipe\n");
upd_shut_too_early_logged = TRUE;
}
return_status = CONTINUE_POLL;
}
if (SS_NORMAL != status)
{
if (EREPL_RECV == repl_errno)
{
if (REPL_CONN_RESET(status))
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Connection reset while receiving XOFF_ACK. "
"Status = %d ; %s\n", status, STRERROR(status));
repl_close(>mrecv_sock_fd);
repl_connection_reset = TRUE;
xoff_sent = FALSE;
send_badtrans = FALSE;
return_status = STOP_POLL;
} else
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error while draining replication pipe. Error in recv"), status);
} else
{
assert(EREPL_SELECT == repl_errno);
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error while draining replication pipe. Error in select"), status);
}
}
} else
return_status = STOP_POLL;
/* Like was done before for the XOFF_ACK_ME message, send a BADTRANS/CMP2UNCMP message only if we know
* the endianness of the other side. If not, no point in sending one anyways and saves us trouble in
* case of cross-endian replication connections.
*/
if ((STOP_POLL == return_status) && (send_badtrans || send_cmp2uncmp)
&& (FD_INVALID != gtmrecv_sock_fd) && remote_side->endianness_known)
{ /* Send REPL_BADTRANS or REPL_CMP2UNCMP message */
if (!remote_side->cross_endian)
{
bad_trans_msg.type = send_cmp2uncmp ? REPL_CMP2UNCMP : REPL_BADTRANS;
bad_trans_msg.len = MIN_REPL_MSGLEN;
bad_trans_msg.start_seqno = send_seqno;
} else
{
bad_trans_msg.type = send_cmp2uncmp ? GTM_BYTESWAP_32(REPL_CMP2UNCMP) : GTM_BYTESWAP_32(REPL_BADTRANS);
bad_trans_msg.len = GTM_BYTESWAP_32(MIN_REPL_MSGLEN);
bad_trans_msg.start_seqno = GTM_BYTESWAP_64(send_seqno);
}
REPL_SEND_LOOP(gtmrecv_sock_fd, &bad_trans_msg, bad_trans_msg.len, REPL_POLL_NOWAIT)
; /* Empty Body */
if (SS_NORMAL == status)
{
if (send_cmp2uncmp)
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL_CMP2UNCMP message sent with seqno %llu\n", send_seqno);
else
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL_BADTRANS message sent with seqno %llu\n", send_seqno);
} else
{
if (REPL_CONN_RESET(status) && EREPL_SEND == repl_errno)
{
if (send_cmp2uncmp)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Connection reset while sending REPL_CMP2UNCMP. "
"Status = %d ; %s\n", status, STRERROR(status));
} else
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Connection reset while sending REPL_BADTRANS. "
"Status = %d ; %s\n", status, STRERROR(status));
}
repl_close(>mrecv_sock_fd);
repl_connection_reset = TRUE;
return_status = STOP_POLL;
} else if (EREPL_SEND == repl_errno)
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error sending REPL_BADTRANS/REPL_CMP2UNCMP. Error in send"), status);
else
{
assert(EREPL_SELECT == repl_errno);
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
LEN_AND_LIT("Error sending REPL_BADTRANS/REPL_CMP2UNCMP. Error in select"), status);
}
}
send_badtrans = FALSE;
if (send_cmp2uncmp)
{
REPL_DPRINT1("gtmrecv_poll_actions : Setting gtmrecv_wait_for_jnl_seqno to TRUE because this receiver"
"server requested a fall-back from compressed to uncompressed operation\n");
gtmrecv_wait_for_jnl_seqno = TRUE;/* set this to TRUE to break out and go back to a fresh "do_main_loop" */
gtmrecv_bad_trans_sent = TRUE;
gtmrecv_send_cmp2uncmp = FALSE;
send_cmp2uncmp = FALSE;
}
}
if ((upd_proc_local->bad_trans && bad_trans_detected) || onln_rlbk_flg_set
|| (UPDPROC_START == upd_proc_local->start_upd) && (1 != report_cnt))
{
if (UPDPROC_START == upd_proc_local->start_upd)
{
assert(is_updproc_alive() != SRV_ALIVE);
upd_proc_local->upd_proc_shutdown = NO_SHUTDOWN;
}
recvpool_ctl->wrapped = FALSE;
recvpool_ctl->write_wrap = recvpool_ctl->recvpool_size;
recvpool_ctl->write = 0;
/* Reset last_rcvd_histinfo, last_valid_histinfo etc. as they reflect context from unprocessed data
* in the receive pool and those are no longer valid because we have drained the receive pool.
*/
GTMRECV_CLEAR_CACHED_HISTINFO(recvpool.recvpool_ctl, jnlpool, jnlpool_ctl, INSERT_STRM_HISTINFO_FALSE);
if (UPDPROC_START == upd_proc_local->start_upd)
{
/* Attempt starting the update process */
for (upd_start_attempts = 0;
UPDPROC_START_ERR == (upd_start_status = gtmrecv_upd_proc_init(FALSE)) &&
GTMRECV_MAX_UPDSTART_ATTEMPTS > upd_start_attempts;
upd_start_attempts++)
{
if (EREPL_UPDSTART_SEMCTL == repl_errno || EREPL_UPDSTART_BADPATH == repl_errno)
{
gtmrecv_autoshutdown();
} else if (EREPL_UPDSTART_FORK == repl_errno)
{
/* Couldn't start up update now, can try later */
LONG_SLEEP(GTMRECV_WAIT_FOR_PROC_SLOTS);
continue;
} else if (EREPL_UPDSTART_EXEC == repl_errno)
{
/* In forked child, could not exec, should exit */
gtmrecv_exit(ABNORMAL_SHUTDOWN);
}
}
if (UPDPROC_STARTED == (upd_proc_local->start_upd = upd_start_status))
{
REPL_DPRINT1("gtmrecv_poll_actions : Setting gtmrecv_wait_for_jnl_seqno to TRUE because of "
"upd restart\n");
gtmrecv_wait_for_jnl_seqno = TRUE;
report_cnt = next_report_at = 1;
if (send_xoff && (FD_INVALID == gtmrecv_sock_fd))
{
/* Update start command was issued before connection was established,
* no point in sending XOFF. */
send_xoff = FALSE;
}
} else
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "%d failed attempts to fork update process. Try later\n",
upd_start_attempts);
}
} else
{
gtmrecv_wait_for_jnl_seqno = TRUE;/* set this to TRUE to break out and go back to a fresh "do_main_loop" */
if (onln_rlbk_flg_set)
{
assert(NULL != jnlpool_ctl);
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Closing connection due to ONLINE ROLLBACK\n");
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL INFO - Current Jnlpool Seqno : %llu\n",
jnlpool_ctl->jnl_seqno);
repl_log(gtmrecv_log_fp, TRUE, TRUE, "REPL INFO - Current Receive Pool Seqno : %llu\n",
recvpool_ctl->jnl_seqno);
repl_close(>mrecv_sock_fd);
repl_connection_reset = TRUE;
xoff_sent = FALSE;
send_badtrans = FALSE;
upd_proc_local->onln_rlbk_flg = FALSE;
/* Before restarting afresh, sync the online rollback cycles. This way any future grab_lock that
* we do after restarting should not realize an unhandled online rollback. For receiver, it is
* just syncing the journal pool cycles as the databases are not opened. But, to be safe, grab
* the lock and sync the cycles.
*/
grab_lock(jnlpool.jnlpool_dummy_reg, TRUE, GRAB_LOCK_ONLY);
SYNC_ONLN_RLBK_CYCLES;
rel_lock(jnlpool.jnlpool_dummy_reg);
return_status = STOP_POLL;
recvpool_ctl->jnl_seqno = 0;
} else
{
REPL_DPRINT1("gtmrecv_poll_actions : Setting gtmrecv_wait_for_jnl_seqno to TRUE because bad trans"
"sent\n");
gtmrecv_bad_trans_sent = TRUE;
upd_proc_local->bad_trans = FALSE;
recvpool_ctl->jnl_seqno = upd_proc_local->read_jnl_seqno;
}
}
}
if ((0 == *pending_data_len) && (0 != gtmrecv_local->changelog))
{
if (gtmrecv_local->changelog & REPLIC_CHANGE_LOGINTERVAL)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Changing log interval from %u to %u\n",
log_interval, gtmrecv_local->log_interval);
log_interval = gtmrecv_local->log_interval;
gtmrecv_reinit_logseqno(); /* will force a LOG on the first recv following the interval change */
}
if (gtmrecv_local->changelog & REPLIC_CHANGE_LOGFILE)
{
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Changing log file to %s\n", gtmrecv_local->log_file);
repl_log_init(REPL_GENERAL_LOG, >mrecv_log_fd, gtmrecv_local->log_file);
repl_log_fd2fp(>mrecv_log_fp, gtmrecv_log_fd);
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Change log to %s successful\n",gtmrecv_local->log_file);
}
/* NOTE: update process and receiver each ignore any setting specific to the other (REPLIC_CHANGE_UPD_LOGINTERVAL,
* REPLIC_CHANGE_LOGINTERVAL) */
if (REPLIC_CHANGE_LOGINTERVAL == gtmrecv_local->changelog)
upd_proc_local->changelog = 0;
else
upd_proc_local->changelog = gtmrecv_local->changelog; /* Pass changelog request to the update process */
gtmrecv_local->changelog = 0;
}
if (0 == *pending_data_len && !gtmrecv_logstats && gtmrecv_local->statslog)
{
gtmrecv_logstats = TRUE;
repl_log(gtmrecv_log_fp, TRUE, TRUE, "Begin statistics logging\n");
} else if (0 == *pending_data_len && gtmrecv_logstats && !gtmrecv_local->statslog)
{
gtmrecv_logstats = FALSE;
/* Force all data out to the file before closing the file */
repl_log(gtmrecv_log_fp, TRUE, TRUE, "End statistics logging\n");
}
if (0 == *pending_data_len)
{
if (upd_helper_ctl->start_helpers)
{
gtmrecv_helpers_init(upd_helper_ctl->start_n_readers, upd_helper_ctl->start_n_writers);
upd_helper_ctl->start_helpers = FALSE;
}
if (HELPER_REAP_NONE != (status = upd_helper_ctl->reap_helpers) ||
(double)GTMRECV_REAP_HELPERS_INTERVAL <= difftime(gtmrecv_now, last_reap_time))
{
gtmrecv_reap_helpers(HELPER_REAP_WAIT == status);
last_reap_time = gtmrecv_now;
}
}
return (return_status);
}
/*
* ------------------------------------------
* Hang the process for a specified time.
*
* Goes to sleep for a positive value.
* Any caught signal will terminate the sleep
* following the execution of that signal's catching routine.
*
* Arguments:
* num - time to sleep
*
* Return:
* none
* ------------------------------------------
*/
void op_hang(mval* num)
{
int ms;
mv_stent *mv_zintcmd;
ABS_TIME cur_time, end_time;
# ifdef VMS
uint4 time[2];
int4 efn_mask, status;
# endif
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
ms = 0;
MV_FORCE_NUM(num);
if (num->mvtype & MV_INT)
{
if (0 < num->m[1])
{
assert(MV_BIAS >= 1000); /* if formats change overflow may need attention */
ms = num->m[1] * (1000 / MV_BIAS);
}
} else if (0 == num->sgn) /* if sign is not 0 it means num is negative */
ms = mval2i(num) * 1000; /* too big to care about fractional amounts */
if (ms)
{
if (TREF(tpnotacidtime) * 1000 < ms)
TPNOTACID_CHECK(HANGSTR);
# if defined(DEBUG) && defined(UNIX)
if (gtm_white_box_test_case_enabled
&& (WBTEST_DEFERRED_TIMERS == gtm_white_box_test_case_number)
&& (3 > gtm_white_box_test_case_count)
&& (123000 == ms))
{
DEFER_INTERRUPTS(INTRPT_NO_TIMER_EVENTS);
DBGFPF((stderr, "OP_HANG: will sleep for 20 seconds\n"));
LONG_SLEEP(20);
DBGFPF((stderr, "OP_HANG: done sleeping\n"));
ENABLE_INTERRUPTS(INTRPT_NO_TIMER_EVENTS);
return;
}
if (gtm_white_box_test_case_enabled
&& (WBTEST_BREAKMPC == gtm_white_box_test_case_number)
&& (0 == gtm_white_box_test_case_count)
&& (999 == ms))
{
frame_pointer->old_frame_pointer->mpc = (unsigned char *)GTM64_ONLY(0xdeadbeef12345678)
NON_GTM64_ONLY(0xdead1234);
return;
}
/* Upon seeing a .999s hang this white-box test launches a timer that pops with a period of UTIL_OUT_SYSLOG_INTERVAL
* and prints a long message via util_out_ptr.
*/
if (gtm_white_box_test_case_enabled
&& (WBTEST_UTIL_OUT_BUFFER_PROTECTION == gtm_white_box_test_case_number)
&& (0 == gtm_white_box_test_case_count)
&& (999 == ms))
{
start_timer((TID)&util_out_syslog_dump, UTIL_OUT_SYSLOG_INTERVAL, util_out_syslog_dump, 0, NULL);
return;
}
# endif
sys_get_curr_time(&cur_time);
mv_zintcmd = find_mvstent_cmd(ZINTCMD_HANG, restart_pc, restart_ctxt, FALSE);
if (!mv_zintcmd)
add_int_to_abs_time(&cur_time, ms, &end_time);
else
{
end_time = mv_zintcmd->mv_st_cont.mvs_zintcmd.end_or_remain;
cur_time = sub_abs_time(&end_time, &cur_time); /* get remaing time to sleep */
if (0 <= cur_time.at_sec)
ms = (int4)(cur_time.at_sec * 1000 + cur_time.at_usec / 1000);
else
ms = 0; /* all done */
/* restore/pop previous zintcmd_active[ZINTCMD_HANG] hints */
TAREF1(zintcmd_active, ZINTCMD_HANG).restart_pc_last = mv_zintcmd->mv_st_cont.mvs_zintcmd.restart_pc_prior;
TAREF1(zintcmd_active, ZINTCMD_HANG).restart_ctxt_last
= mv_zintcmd->mv_st_cont.mvs_zintcmd.restart_ctxt_prior;
TAREF1(zintcmd_active, ZINTCMD_HANG).count--;
assert(0 <= TAREF1(zintcmd_active, ZINTCMD_HANG).count);
if (mv_chain == mv_zintcmd)
POP_MV_STENT(); /* just pop if top of stack */
else
{ /* flag as not active */
mv_zintcmd->mv_st_cont.mvs_zintcmd.command = ZINTCMD_NOOP;
mv_zintcmd->mv_st_cont.mvs_zintcmd.restart_pc_check = NULL;
}
if (0 == ms)
return; /* done HANGing */
}
UNIX_ONLY(hiber_start(ms);)
VMS_ONLY(
time[0] = -time_low_ms(ms);
time[1] = -time_high_ms(ms) - 1;
efn_mask = (1 << efn_outofband | 1 << efn_timer);
if (SS$_NORMAL != (status = sys$setimr(efn_timer, &time, NULL, &time, 0)))
rts_error(VARLSTCNT(8) ERR_SYSCALL, 5, RTS_ERROR_LITERAL("$setimr"), CALLFROM, status);
if (SS$_NORMAL != (status = sys$wflor(efn_outofband, efn_mask)))
rts_error(VARLSTCNT(8) ERR_SYSCALL, 5, RTS_ERROR_LITERAL("$wflor"), CALLFROM, status);
)
if (outofband)
/*
* ------------------------------------------
* Hang the process for a specified time.
*
* Goes to sleep for a positive value.
* Any caught signal will terminate the sleep
* following the execution of that signal's catching routine.
*
* The actual hang duration should be NO LESS than the specified
* duration for specified durations greater than .001 seconds.
* Certain applications depend on this assumption.
*
* Arguments:
* num - time to sleep
*
* Return:
* none
* ------------------------------------------
*/
void op_hang(mval* num)
{
int ms;
double tmp;
mv_stent *mv_zintcmd;
ABS_TIME cur_time, end_time;
# ifdef VMS
uint4 time[2];
int4 efn_mask, status;
# endif
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
ms = 0;
MV_FORCE_NUM(num);
if (num->mvtype & MV_INT)
{
if (0 < num->m[1])
{
assert(MV_BIAS >= 1000); /* if formats change overflow may need attention */
ms = num->m[1] * (1000 / MV_BIAS);
}
} else if (0 == num->sgn) /* if sign is not 0 it means num is negative */
{
tmp = mval2double(num) * (double)1000;
ms = ((double)MAXPOSINT4 >= tmp) ? (int)tmp : (int)MAXPOSINT4;
}
if (ms)
{
if (TREF(tpnotacidtime) * 1000 < ms)
TPNOTACID_CHECK(HANGSTR);
# if defined(DEBUG) && defined(UNIX)
if (WBTEST_ENABLED(WBTEST_DEFERRED_TIMERS) && (3 > gtm_white_box_test_case_count) && (123000 == ms))
{
DEFER_INTERRUPTS(INTRPT_NO_TIMER_EVENTS);
DBGFPF((stderr, "OP_HANG: will sleep for 20 seconds\n"));
LONG_SLEEP(20);
DBGFPF((stderr, "OP_HANG: done sleeping\n"));
ENABLE_INTERRUPTS(INTRPT_NO_TIMER_EVENTS);
return;
}
if (WBTEST_ENABLED(WBTEST_BREAKMPC)&& (0 == gtm_white_box_test_case_count) && (999 == ms))
{
frame_pointer->old_frame_pointer->mpc = (unsigned char *)GTM64_ONLY(0xdeadbeef12345678)
NON_GTM64_ONLY(0xdead1234);
return;
}
if (WBTEST_ENABLED(WBTEST_UTIL_OUT_BUFFER_PROTECTION) && (0 == gtm_white_box_test_case_count) && (999 == ms))
{ /* Upon seeing a .999s hang this white-box test launches a timer that pops with a period of
* UTIL_OUT_SYSLOG_INTERVAL and prints a long message via util_out_ptr.
*/
start_timer((TID)&util_out_syslog_dump, UTIL_OUT_SYSLOG_INTERVAL, util_out_syslog_dump, 0, NULL);
return;
}
# endif
sys_get_curr_time(&cur_time);
mv_zintcmd = find_mvstent_cmd(ZINTCMD_HANG, restart_pc, restart_ctxt, FALSE);
if (!mv_zintcmd)
add_int_to_abs_time(&cur_time, ms, &end_time);
else
{
end_time = mv_zintcmd->mv_st_cont.mvs_zintcmd.end_or_remain;
cur_time = sub_abs_time(&end_time, &cur_time); /* get remaing time to sleep */
if (0 <= cur_time.at_sec)
ms = (int4)(cur_time.at_sec * 1000 + cur_time.at_usec / 1000);
else
ms = 0; /* all done */
/* restore/pop previous zintcmd_active[ZINTCMD_HANG] hints */
TAREF1(zintcmd_active, ZINTCMD_HANG).restart_pc_last = mv_zintcmd->mv_st_cont.mvs_zintcmd.restart_pc_prior;
TAREF1(zintcmd_active, ZINTCMD_HANG).restart_ctxt_last
= mv_zintcmd->mv_st_cont.mvs_zintcmd.restart_ctxt_prior;
TAREF1(zintcmd_active, ZINTCMD_HANG).count--;
assert(0 <= TAREF1(zintcmd_active, ZINTCMD_HANG).count);
if (mv_chain == mv_zintcmd)
POP_MV_STENT(); /* just pop if top of stack */
else
{ /* flag as not active */
mv_zintcmd->mv_st_cont.mvs_zintcmd.command = ZINTCMD_NOOP;
mv_zintcmd->mv_st_cont.mvs_zintcmd.restart_pc_check = NULL;
}
if (0 == ms)
return; /* done HANGing */
}
# ifdef UNIX
if (ms < 10)
SLEEP_USEC(ms * 1000, TRUE); /* Finish the sleep if it is less than 10ms. */
else
hiber_start(ms);
# elif defined(VMS)
time[0] = -time_low_ms(ms);
time[1] = -time_high_ms(ms) - 1;
efn_mask = (1 << efn_outofband | 1 << efn_timer);
if (SS$_NORMAL != (status = sys$setimr(efn_timer, &time, NULL, &time, 0)))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_SYSCALL, 5, RTS_ERROR_LITERAL("$setimr"), CALLFROM, status);
if (SS$_NORMAL != (status = sys$wflor(efn_outofband, efn_mask)))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_SYSCALL, 5, RTS_ERROR_LITERAL("$wflor"), CALLFROM, status);
if (outofband)
{
if (SS$_WASCLR == (status = sys$readef(efn_timer, &efn_mask)))
{
if (SS$_NORMAL != (status = sys$cantim(&time, 0)))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_SYSCALL, 5, RTS_ERROR_LITERAL("$cantim"),
CALLFROM, status);
} else
assertpro(SS$_WASSET == status);
}
# endif
} else
rel_quant();
if (outofband)
{
PUSH_MV_STENT(MVST_ZINTCMD);
mv_chain->mv_st_cont.mvs_zintcmd.end_or_remain = end_time;
mv_chain->mv_st_cont.mvs_zintcmd.restart_ctxt_check = restart_ctxt;
mv_chain->mv_st_cont.mvs_zintcmd.restart_pc_check = restart_pc;
/* save current information from zintcmd_active */
mv_chain->mv_st_cont.mvs_zintcmd.restart_ctxt_prior = TAREF1(zintcmd_active, ZINTCMD_HANG).restart_ctxt_last;
mv_chain->mv_st_cont.mvs_zintcmd.restart_pc_prior = TAREF1(zintcmd_active, ZINTCMD_HANG).restart_pc_last;
TAREF1(zintcmd_active, ZINTCMD_HANG).restart_pc_last = restart_pc;
TAREF1(zintcmd_active, ZINTCMD_HANG).restart_ctxt_last = restart_ctxt;
TAREF1(zintcmd_active, ZINTCMD_HANG).count++;
mv_chain->mv_st_cont.mvs_zintcmd.command = ZINTCMD_HANG;
outofband_action(FALSE);
}
return;
}
int gtmsource_changelog(void)
{
uint4 changelog_accepted = 0;
int log_fd = 0; /*used to indicate whether the new specified log file is writable*/
int close_status = 0; /*used to indicate if log file is successfully closed*/
char* err_code;
int save_errno = 0;
int retry_count = 5;
assert(holds_sem[SOURCE][JNL_POOL_ACCESS_SEM]);
repl_log(stderr, TRUE, TRUE, "Initiating CHANGELOG operation on source server pid [%d] for secondary instance [%s]\n",
jnlpool.gtmsource_local->gtmsource_pid, jnlpool.gtmsource_local->secondary_instname);
if (0 != jnlpool.gtmsource_local->changelog)
{
retry_count = 5;
while (0 != retry_count--)
{
LONG_SLEEP(5);
if (!jnlpool.gtmsource_local->changelog)
break;
}
}
if (0 != jnlpool.gtmsource_local->changelog)
{
util_out_print("Change log is already in progress. Not initiating change in log file or log interval", TRUE);
return (ABNORMAL_SHUTDOWN);
}
if ('\0' != gtmsource_options.log_file[0]) /* trigger change in log file */
{
if (0 != STRCMP(jnlpool.gtmsource_local->log_file, gtmsource_options.log_file))
{ /*check if the new log file is writable*/
OPENFILE3_CLOEXEC(gtmsource_options.log_file,
O_RDWR | O_CREAT | O_APPEND, S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP | S_IROTH, log_fd);
if (log_fd < 0)
{
save_errno = ERRNO;
err_code = STRERROR(save_errno);
if ('\0' != jnlpool.gtmsource_local->log_file[0])
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_REPLLOGOPN, 6,
LEN_AND_STR(gtmsource_options.log_file),
LEN_AND_STR(err_code),
LEN_AND_STR(jnlpool.gtmsource_local->log_file));
else
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_REPLLOGOPN, 6,
LEN_AND_STR(gtmsource_options.log_file),
LEN_AND_STR(err_code),
LEN_AND_STR(NULL_DEVICE));
} else {
CLOSEFILE_IF_OPEN(log_fd, close_status);
assert(close_status==0);
changelog_accepted |= REPLIC_CHANGE_LOGFILE;
STRCPY(jnlpool.gtmsource_local->log_file, gtmsource_options.log_file);
util_out_print("Change log initiated with file !AD", TRUE, LEN_AND_STR(gtmsource_options.log_file));
}
} else
util_out_print("Log file is already !AD. Not initiating change in log file", TRUE,
LEN_AND_STR(gtmsource_options.log_file));
}
if (0 != gtmsource_options.src_log_interval) /* trigger change in log interval */
{
if (gtmsource_options.src_log_interval != jnlpool.gtmsource_local->log_interval)
{
changelog_accepted |= REPLIC_CHANGE_LOGINTERVAL;
jnlpool.gtmsource_local->log_interval = gtmsource_options.src_log_interval;
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_CHANGELOGINTERVAL, 5,
LEN_AND_LIT("Source"),
LEN_AND_STR(jnlpool.gtmsource_local->log_file),
gtmsource_options.src_log_interval);
} else
util_out_print("Log interval is already !UL. Not initiating change in log interval", TRUE,
gtmsource_options.src_log_interval);
}
if (0 != changelog_accepted)
jnlpool.gtmsource_local->changelog = changelog_accepted;
else
util_out_print("No change to log file or log interval", TRUE);
return (0 != save_errno) ? ABNORMAL_SHUTDOWN : NORMAL_SHUTDOWN;
}
int main(int argc, char *argv[])
{
DCL_THREADGBL_ACCESS;
GTM_THREADGBL_INIT;
common_startup_init(DSE_IMAGE);
licensed = TRUE;
TREF(transform) = TRUE;
TREF(no_spangbls) = TRUE; /* dse operates on a per-region basis irrespective of global mapping in gld */
TREF(skip_file_corrupt_check) = TRUE; /* do not let csd->file_corrupt flag cause errors in dse */
op_open_ptr = op_open;
patch_curr_blk = get_dir_root();
err_init(util_base_ch);
UNICODE_ONLY(gtm_strToTitle_ptr = >m_strToTitle);
GTM_ICU_INIT_IF_NEEDED; /* Note: should be invoked after err_init (since it may error out) and before CLI parsing */
sig_init(generic_signal_handler, dse_ctrlc_handler, suspsigs_handler, continue_handler);
atexit(util_exit_handler);
SET_LATCH_GLOBAL(&defer_latch, LOCK_AVAILABLE);
stp_init(STP_INITSIZE);
rts_stringpool = stringpool;
getjobname();
INVOKE_INIT_SECSHR_ADDRS;
io_init(TRUE);
getzdir();
gtm_chk_dist(argv[0]);
prealloc_gt_timers();
gt_timers_add_safe_hndlrs();
initialize_pattern_table();
gvinit();
region_init(FALSE);
util_out_print("!/File !_!AD", TRUE, DB_LEN_STR(gv_cur_region));
util_out_print("Region!_!AD!/", TRUE, REG_LEN_STR(gv_cur_region));
cli_lex_setup(argc, argv);
/* Since DSE operates on a region-by-region basis (for the most part), do not use a global directory at all from now on */
original_header = gd_header;
gd_header = NULL;
OPERATOR_LOG_MSG;
# ifdef DEBUG
if ((gtm_white_box_test_case_enabled && (WBTEST_SEMTOOLONG_STACK_TRACE == gtm_white_box_test_case_number) ))
{
sgmnt_addrs * csa;
node_local_ptr_t cnl;
csa = &FILE_INFO(gv_cur_region)->s_addrs;
cnl = csa->nl;
cnl->wbox_test_seq_num = 1; /*Signal the first step and wait here*/
/* The signal to the shell. MUPIP must not start BEFORE DSE */
util_out_print("DSE is ready. MUPIP can start. Note: This message is a part of WBTEST_SEMTOOLONG_STACK_TRACE test. "
"It will not appear in PRO version.", TRUE);
while (2 != cnl->wbox_test_seq_num) /*Wait for another process to get hold of the semaphore and signal next step*/
LONG_SLEEP(1);
}
# endif
if (argc < 2)
display_prompt();
while (1)
{
if (!dse_process(argc))
break;
display_prompt();
}
dse_exit();
REVERT;
return 0;
}
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);
sm_uc_ptr_t t_qread(block_id blk, sm_int_ptr_t cycle, cache_rec_ptr_ptr_t cr_out)
/* cycle is used in t_end to detect if the buffer has been refreshed since the t_qread */
{
int4 status;
uint4 blocking_pid;
cache_rec_ptr_t cr;
bt_rec_ptr_t bt;
boolean_t clustered, hold_onto_crit, was_crit, issued_db_init_crypt_warning, sync_needed;
int dummy, lcnt, ocnt;
cw_set_element *cse;
off_chain chain1;
register sgmnt_addrs *csa;
register sgmnt_data_ptr_t csd;
enum db_ver ondsk_blkver;
int4 dummy_errno, gtmcrypt_errno;
boolean_t already_built, is_mm, reset_first_tp_srch_status, set_wc_blocked, sleep_invoked;
ht_ent_int4 *tabent;
srch_blk_status *blkhist;
trans_num dirty, blkhdrtn;
sm_uc_ptr_t buffaddr;
uint4 stuck_cnt = 0;
boolean_t lcl_blk_free;
node_local_ptr_t cnl;
gd_segment *seg;
uint4 buffs_per_flush, flush_target;
enc_info_t *encr_ptr;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
lcl_blk_free = block_is_free;
block_is_free = FALSE; /* Reset to FALSE so that if t_qread fails below, we don't have an incorrect state of this var */
first_tp_srch_status = NULL;
reset_first_tp_srch_status = FALSE;
csa = cs_addrs;
csd = csa->hdr;
INCR_DB_CSH_COUNTER(csa, n_t_qreads, 1);
is_mm = (dba_mm == csd->acc_meth);
/* We better hold crit in the final retry (TP & non-TP). Only exception is journal recovery */
assert((t_tries < CDB_STAGNATE) || csa->now_crit || mupip_jnl_recover);
if (dollar_tlevel)
{
assert(sgm_info_ptr);
if (0 != sgm_info_ptr->cw_set_depth)
{
chain1 = *(off_chain *)&blk;
if (1 == chain1.flag)
{
assert(sgm_info_ptr->cw_set_depth);
if ((int)chain1.cw_index < sgm_info_ptr->cw_set_depth)
tp_get_cw(sgm_info_ptr->first_cw_set, (int)chain1.cw_index, &cse);
else
{
assert(FALSE == csa->now_crit);
rdfail_detail = cdb_sc_blknumerr;
return (sm_uc_ptr_t)NULL;
}
} else
{
if (NULL != (tabent = lookup_hashtab_int4(sgm_info_ptr->blks_in_use, (uint4 *)&blk)))
first_tp_srch_status = tabent->value;
else
first_tp_srch_status = NULL;
ASSERT_IS_WITHIN_TP_HIST_ARRAY_BOUNDS(first_tp_srch_status, sgm_info_ptr);
cse = first_tp_srch_status ? first_tp_srch_status->cse : NULL;
}
assert(!cse || !cse->high_tlevel);
assert(!chain1.flag || cse);
if (cse)
{ /* transaction has modified the sought after block */
if ((gds_t_committed != cse->mode) || (n_gds_t_op < cse->old_mode))
{ /* Changes have not been committed to shared memory, i.e. still in private memory.
* Build block in private buffer if not already done and return the same.
*/
assert(gds_t_writemap != cse->mode);
if (FALSE == cse->done)
{ /* out of date, so make it current */
assert(gds_t_committed != cse->mode);
already_built = (NULL != cse->new_buff);
/* Validate the block's search history right after building a private copy.
* This is not needed in case gvcst_search is going to reuse the clue's search
* history and return (because tp_hist will do the validation of this block).
* But if gvcst_search decides to do a fresh traversal (because the clue does not
* cover the path of the current input key etc.) the block build that happened now
* will not get validated in tp_hist since it will instead be given the current
* key's search history path (a totally new path) for validation. Since a private
* copy of the block has been built, tp_tend would also skip validating this block
* so it is necessary that we validate the block right here. Since it is tricky to
* accurately differentiate between the two cases, we do the validation
* unconditionally here (besides it is only a few if checks done per block build
* so it is considered okay performance-wise).
*/
gvcst_blk_build(cse, (uchar_ptr_t)cse->new_buff, 0);
assert(NULL != cse->blk_target);
if (!already_built && !chain1.flag)
{
buffaddr = first_tp_srch_status->buffaddr;
cr = first_tp_srch_status->cr;
assert((is_mm || cr) && buffaddr);
blkhdrtn = ((blk_hdr_ptr_t)buffaddr)->tn;
if (TP_IS_CDB_SC_BLKMOD3(cr, first_tp_srch_status, blkhdrtn))
{
assert(CDB_STAGNATE > t_tries);
rdfail_detail = cdb_sc_blkmod; /* should this be something else */
TP_TRACE_HIST_MOD(blk, gv_target, tp_blkmod_t_qread, cs_data,
first_tp_srch_status->tn, blkhdrtn,
((blk_hdr_ptr_t)buffaddr)->levl);
return (sm_uc_ptr_t)NULL;
}
if (!is_mm && ((first_tp_srch_status->cycle != cr->cycle)
|| (first_tp_srch_status->blk_num != cr->blk)))
{
assert(CDB_STAGNATE > t_tries);
rdfail_detail = cdb_sc_lostcr; /* should this be something else */
return (sm_uc_ptr_t)NULL;
}
}
cse->done = TRUE;
}
*cycle = CYCLE_PVT_COPY;
*cr_out = 0;
return (sm_uc_ptr_t)cse->new_buff;
} else
{ /* Block changes are already committed to shared memory (possible if we are in TP
* in the 2nd phase of M-Kill in gvcst_expand_free_subtree.c). In this case, read
* block from shared memory; do not look at private memory (i.e. cse) as that might
* not be as uptodate as shared memory.
*/
assert(csa->now_crit); /* gvcst_expand_free_subtree does t_qread in crit */
/* If this block was newly created as part of the TP transaction, it should not be killed
* as part of the 2nd phase of M-kill. This is because otherwise the block's cse would
* have had an old_mode of kill_t_create in which case we would not have come into this
* else block. Assert accordingly.
*/
assert(!chain1.flag);
first_tp_srch_status = NULL; /* do not use any previous srch_hist information */
}
}
} else
{
if (NULL != (tabent = lookup_hashtab_int4(sgm_info_ptr->blks_in_use, (uint4 *)&blk)))
first_tp_srch_status = tabent->value;
else
first_tp_srch_status = NULL;
}
ASSERT_IS_WITHIN_TP_HIST_ARRAY_BOUNDS(first_tp_srch_status, sgm_info_ptr);
if (!is_mm && first_tp_srch_status)
{
cr = first_tp_srch_status->cr;
assert(cr && !first_tp_srch_status->cse);
if (first_tp_srch_status->cycle == cr->cycle)
{
*cycle = first_tp_srch_status->cycle;
*cr_out = cr;
cr->refer = TRUE;
if (CDB_STAGNATE <= t_tries) /* mu_reorg doesn't use TP else should have an || for that */
CWS_INSERT(blk);
return (sm_uc_ptr_t)first_tp_srch_status->buffaddr;
} else
{ /* Block was already part of the read-set of this transaction, but got recycled in the cache.
* Allow block recycling by resetting first_tp_srch_status for this blk to reflect the new
* buffer, cycle and cache-record. tp_hist (invoked much later) has validation checks to detect
* if block recycling happened within the same mini-action and restart in that case.
* Updating first_tp_srch_status has to wait until the end of t_qread since only then do we know
* the values to update to. Set a variable that will enable the updation before returning.
* Also assert that if we are in the final retry, we are never in a situation where we have a
* block that got recycled since the start of the current mini-action. This is easily detected since
* as part of the final retry we maintain a hash-table "cw_stagnate" that holds the blocks that
* have been read as part of the current mini-action until now.
*/
assert(CDB_STAGNATE > t_tries || (NULL == lookup_hashtab_int4(&cw_stagnate, (uint4 *)&blk)));
reset_first_tp_srch_status = TRUE;
}
}
}
if ((uint4)blk >= (uint4)csa->ti->total_blks)
{ /* Requested block out of range; could occur because of a concurrency conflict. mm_read and dsk_read assume blk is
* never negative or greater than the maximum possible file size. If a concurrent REORG truncates the file, t_qread
* can proceed despite blk being greater than total_blks. But dsk_read handles this fine; see comments below.
*/
assert((&FILE_INFO(gv_cur_region)->s_addrs == csa) && (csd == cs_data));
assert(!csa->now_crit);
rdfail_detail = cdb_sc_blknumerr;
return (sm_uc_ptr_t)NULL;
}
if (is_mm)
{
*cycle = CYCLE_SHRD_COPY;
*cr_out = 0;
return (sm_uc_ptr_t)(mm_read(blk));
}
was_crit = csa->now_crit;
cnl = csa->nl;
encr_ptr = csa->encr_ptr;
if (NULL != encr_ptr)
{
/* If this is an encrypted database and we hold crit, make sure our private cycle matches the shared cycle.
* Or else we would need to call "process_reorg_encrypt_restart" below (a heavyweight operation) holding crit.
*/
assert(!was_crit || (cnl->reorg_encrypt_cycle == encr_ptr->reorg_encrypt_cycle));
seg = gv_cur_region->dyn.addr;
issued_db_init_crypt_warning = encr_ptr->issued_db_init_crypt_warning;
if (!IS_BITMAP_BLK(blk) && issued_db_init_crypt_warning)
{ /* A non-GT.M process is attempting to read a non-bitmap block, yet it has previously encountered an error
* during db_init (because it did not have access to the encryption keys) and reported it with a -W-
* severity. Since the block it is attempting to read can be in the unencrypted shared memory (read from
* disk by another process with access to the encryption keys), we cannot let it access it without a valid
* handle, so issue an rts_error.
*
* TODO: DSE and LKE could bypass getting the ftok semaphore. LKE is not an issue, but DSE does care about
* the csa->reorg_encrypt_cycle. So it means DSE could get an inconsistent copy of reorg_encrypt_cycle
* and associated hashes if it had done a bypass and a concurrent REORG -ENCRYPT is holding the ftok
* semaphore and changing these values at the same time.
*/
assert(!IS_GTM_IMAGE); /* GT.M would have error'ed out in db_init */
gtmcrypt_errno = SET_REPEAT_MSG_MASK(SET_CRYPTERR_MASK(ERR_CRYPTBADCONFIG));
GTMCRYPT_REPORT_ERROR(gtmcrypt_errno, rts_error, seg->fname_len, seg->fname);
} else if (cnl->reorg_encrypt_cycle != encr_ptr->reorg_encrypt_cycle)
{ /* A concurrent MUPIP REORG ENCRYPT occurred. Cannot proceed with the read even if the block is
* already loaded from disk into the unencrypted global buffers (security issue). Need to load the
* new encryption keys and only let those processes which are able to successfully do this proceed
* with the read. First, copy the key hashes from csd into csa->encr_ptr. That needs crit
* to ensure a concurrent MUPIP REORG ENCRYPT does not sneak in.
*
* Note: Even though we asserted a few lines above that if "was_crit" is TRUE, then we expect
* the encryption cycles to be in sync, we handle this out-of-design situation in "pro" by fixing
* the cycles while holding crit (hopefully rare case so it is okay to hold crit for a heavyweight call).
*/
if (!was_crit)
grab_crit(gv_cur_region);
/* Now that we have crit, sync them up by copying the new keys inside crit and opening the key handles
* outside crit (a potentially long running operation).
*/
SIGNAL_REORG_ENCRYPT_RESTART(mu_reorg_encrypt_in_prog, reorg_encrypt_restart_csa,
cnl, csa, csd, rdfail_detail, process_id);
assert(csa == reorg_encrypt_restart_csa);
if (!was_crit)
rel_crit(gv_cur_region);
/* If we are inside a TP read-write transaction, it is possible we already used the old keys for
* prior calls to "jnl_format" so we have to restart (cannot sync up cycles). Do the same for
* TP read-only transaction as well as NON-TP read-write transaction. In all these cases we know
* the caller is capable of restarting. All other cases we dont know if the caller is capable so
* sync up the cycles and proceed using the new keys for the read.
*
* But since it is possible the caller does not call t_retry right away (e.g. mupip reorg which can
* choose to abandone this tree path and move on to another block without aborting this transaction)
* it is better we finish the pending call to "process_reorg_encrypt_restart" right here before returning.
*/
process_reorg_encrypt_restart();
assert(NULL == reorg_encrypt_restart_csa);
if (IS_NOT_SAFE_TO_SYNC_NEW_KEYS(dollar_tlevel, update_trans))
{
assert(cdb_sc_reorg_encrypt == rdfail_detail); /* set by SIGNAL_REORG_ENCRYPT_RESTART macro */
return (sm_uc_ptr_t)NULL;
}
}
}
assert(dba_bg == csd->acc_meth);
assert(!first_tp_srch_status || !first_tp_srch_status->cr
|| first_tp_srch_status->cycle != first_tp_srch_status->cr->cycle);
if (FALSE == (clustered = csd->clustered))
bt = NULL;
ocnt = 0;
set_wc_blocked = FALSE; /* to indicate whether cnl->wc_blocked was set to TRUE by us */
hold_onto_crit = csa->hold_onto_crit; /* note down in local to avoid csa-> dereference in multiple usages below */
do
{
if (NULL == (cr = db_csh_get(blk)))
{ /* not in memory */
if (clustered && (NULL != (bt = bt_get(blk))) && (FALSE == bt->flushing))
bt = NULL;
if (!csa->now_crit)
{
assert(!hold_onto_crit);
if (NULL != bt)
{ /* at this point, bt is not NULL only if clustered and flushing - wait no crit */
assert(clustered);
wait_for_block_flush(bt, blk); /* try for no other node currently writing the block */
}
/* assume defaults for flush_target and buffs_per_flush */
flush_target = csd->flush_trigger;
buffs_per_flush = 0;
if ((0 != csd->epoch_taper) && (FALSE == gv_cur_region->read_only) && JNL_ENABLED(csd) &&
(0 != cnl->wcs_active_lvl) && (NOJNL != csa->jnl->channel) &&
(0 != cnl->jnl_file.u.inode) && csd->jnl_before_image)
{
EPOCH_TAPER_IF_NEEDED(csa, csd, cnl, (gd_region *) 0, FALSE, buffs_per_flush, flush_target);
}
if ((flush_target <= cnl->wcs_active_lvl) && (FALSE == gv_cur_region->read_only))
JNL_ENSURE_OPEN_WCS_WTSTART(csa, gv_cur_region, buffs_per_flush, dummy_errno);
/* a macro that dclast's "wcs_wtstart" and checks for errors etc. */
/* Get crit but also ensure encryption cycles are in sync ("dsk_read" relies on this).
* Note: "sync_needed" should be TRUE very rarely since we synced the cycles just a few lines
* above. But in case a MUPIP REORG ENCRYPT concurrently sneaked in between these lines we
* need to resync.
*/
sync_needed = grab_crit_encr_cycle_sync(gv_cur_region);
assert(NULL == reorg_encrypt_restart_csa);
assert(!sync_needed || (NULL != encr_ptr));
if (sync_needed && IS_NOT_SAFE_TO_SYNC_NEW_KEYS(dollar_tlevel, update_trans))
{
assert(cnl->reorg_encrypt_cycle == encr_ptr->reorg_encrypt_cycle);
rel_crit(gv_cur_region);
rdfail_detail = cdb_sc_reorg_encrypt; /* set by SIGNAL_REORG_ENCRYPT_RESTART macro */
return (sm_uc_ptr_t)NULL;
}
cr = db_csh_get(blk); /* in case blk arrived before crit */
}
if (clustered && (NULL != (bt = bt_get(blk))) && (TRUE == bt->flushing))
{ /* Once crit, need to assure that if clustered, that flushing is [still] complete
* If it isn't, we missed an entire WM cycle and have to wait for another node to finish */
wait_for_block_flush(bt, blk); /* ensure no other node currently writing the block */
}
if (NULL == cr)
{ /* really not in memory - must get a new buffer */
assert(csa->now_crit);
cr = db_csh_getn(blk);
if (CR_NOTVALID == (sm_long_t)cr)
{
assert(cnl->wc_blocked); /* only reason we currently know wcs_get_space could fail */
assert(gtm_white_box_test_case_enabled);
SET_TRACEABLE_VAR(cnl->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wc_blocked_t_qread_db_csh_getn_invalid_blk);
set_wc_blocked = TRUE;
break;
}
assert(0 <= cr->read_in_progress);
*cycle = cr->cycle;
cr->tn = csd->trans_hist.curr_tn;
/* Record history of most recent disk reads only in dbg builds for now. Although the macro
* is just a couple dozen instructions, it is done while holding crit so we want to avoid
* delaying crit unless really necessary. Whoever wants this information can enable it
* by a build change to remove the DEBUG_ONLY part below.
*/
DEBUG_ONLY(DSKREAD_TRACE(csa, GDS_ANY_ABS2REL(csa,cr), cr->tn, process_id, blk, cr->cycle);)
if (!was_crit && !hold_onto_crit)
rel_crit(gv_cur_region);
/* read outside of crit may be of a stale block but should be detected by t_end or tp_tend */
assert(0 == cr->dirty);
assert(cr->read_in_progress >= 0);
CR_BUFFER_CHECK(gv_cur_region, csa, csd, cr);
buffaddr = (sm_uc_ptr_t)GDS_REL2ABS(cr->buffaddr);
# ifdef DEBUG
/* stop self to test sechshr_db_clnup clears the read state */
if (gtm_white_box_test_case_enabled
&& (WBTEST_SIGTSTP_IN_T_QREAD == gtm_white_box_test_case_number))
{ /* this should never fail, but because of the way we developed the test we got paranoid */
dummy = kill(process_id, SIGTERM);
assert(0 == dummy);
for (dummy = 10; dummy; dummy--)
LONG_SLEEP(10); /* time for sigterm to take hit before we clear block_now_locked */
}
# endif
if (SS_NORMAL != (status = dsk_read(blk, buffaddr, &ondsk_blkver, lcl_blk_free)))
{
/* buffer does not contain valid data, so reset blk to be empty */
cr->cycle++; /* increment cycle for blk number changes (for tp_hist and others) */
cr->blk = CR_BLKEMPTY;
cr->r_epid = 0;
RELEASE_BUFF_READ_LOCK(cr);
TREF(block_now_locked) = NULL;
assert(-1 <= cr->read_in_progress);
assert(was_crit == csa->now_crit);
if (ERR_DYNUPGRDFAIL == status)
{ /* if we dont hold crit on the region, it is possible due to concurrency conflicts
* that this block is unused (i.e. marked free/recycled in bitmap, see comments in
* gds_blk_upgrade.h). in this case we should not error out but instead restart.
*/
if (was_crit)
{
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(5) status, 3, blk,
DB_LEN_STR(gv_cur_region));
} else
{
rdfail_detail = cdb_sc_lostcr;
return (sm_uc_ptr_t)NULL;
}
}
if ((-1 == status) && !was_crit)
{ /* LSEEKREAD and, consequently, dsk_read return -1 in case pread is unable to fetch
* a full database block's length of data. This can happen if the requested read is
* past the end of the file, which can happen if a concurrent truncate occurred
* after the blk >= csa->ti->total_blks comparison above. Allow for this scenario
* by restarting. However, if we've had crit the whole time, no truncate could have
* happened. -1 indicates a problem with the file, so fall through to DBFILERR.
*/
rdfail_detail = cdb_sc_truncate;
return (sm_uc_ptr_t)NULL;
} else if (IS_CRYPTERR_MASK(status))
{
seg = gv_cur_region->dyn.addr;
GTMCRYPT_REPORT_ERROR(status, rts_error, seg->fname_len, seg->fname);
}
else
{ /* A DBFILERR can be thrown for two possible reasons:
* (1) LSEEKREAD returned an unexpected error due to a filesystem problem; or
* (2) csa/cs_addrs/csd/cs_data are out of sync, and we're trying to read a block
* number for one region from another region with fewer total_blks.
* We suspect the former is what happened in GTM-7623. Apparently the latter
* has been an issue before, too. If either occurs again in pro, this assertpro
* distinguishes the two possibilities.
*/
assertpro((&FILE_INFO(gv_cur_region)->s_addrs == csa) && (csd == cs_data));
rts_error_csa(CSA_ARG(csa) VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region),
status);
}
}
disk_blk_read = TRUE;
assert(0 <= cr->read_in_progress);
assert(0 == cr->dirty);
/* Only set in cache if read was success */
cr->ondsk_blkver = (lcl_blk_free ? GDSVCURR : ondsk_blkver);
cr->r_epid = 0;
RELEASE_BUFF_READ_LOCK(cr);
TREF(block_now_locked) = NULL;
assert(-1 <= cr->read_in_progress);
*cr_out = cr;
assert(was_crit == csa->now_crit);
if (reset_first_tp_srch_status)
RESET_FIRST_TP_SRCH_STATUS(first_tp_srch_status, cr, *cycle);
return buffaddr;
} else if (!was_crit && (BAD_LUCK_ABOUNDS > ocnt))
{
assert(!hold_onto_crit);
assert(TRUE == csa->now_crit);
assert(cnl->in_crit == process_id);
rel_crit(gv_cur_region);
}
}