fis-gtm/sr_port/t_retry.c

464 lines
19 KiB
C

/****************************************************************
* *
* Copyright 2001, 2002 Fidelity Information Services, Inc*
* *
* This source code contains the intellectual property *
* of its copyright holder(s), and is made available *
* under a license. If you do not know the terms of *
* the license, please stop and do not read further. *
* *
****************************************************************/
#include "mdef.h"
#if defined(VMS) && defined(DEBUG)
#include <descrip.h>
#endif
#include "gdsroot.h"
#include "gtm_facility.h"
#include "fileinfo.h"
#include "cdb_sc.h"
#include "gdsbt.h"
#include "gdsfhead.h"
#include "gdskill.h"
#include "gdscc.h"
#include "gdsbml.h"
#include "ccp.h"
#include "error.h"
#include "filestruct.h"
#include "jnl.h"
#include "buddy_list.h" /* needed for tp.h */
#include "hashtab_int4.h" /* needed for tp.h and cws_insert.h */
#include "tp.h"
#include "tp_frame.h"
#include "sleep_cnt.h"
#include "t_retry.h"
#include "format_targ_key.h"
#include "send_msg.h"
#include "longset.h" /* needed for cws_insert.h */
#include "cws_insert.h"
#include "wcs_mm_recover.h"
#include "wcs_sleep.h"
#include "have_crit.h"
#include "gdsbgtr.h" /* for the BG_TRACE_PRO macros */
#include "wcs_backoff.h"
#include "tp_restart.h"
#include "gtm_ctype.h" /* for ISALPHA_ASCII */
#include "anticipatory_freeze.h"
#include "wcs_recover.h"
#include "wbox_test_init.h"
#ifdef GTM_TRIGGER
#include "gtm_trigger_trc.h"
#endif
#ifdef UNIX
#include "gvcst_protos.h"
#include "gtmimagename.h"
#include "caller_id.h"
#endif
#ifdef DEBUG
#include "repl_msg.h"
#include "gtmsource.h"
#endif
/* In mu_reorg if we are in gvcst_bmp_mark_free, we actually have a valid gv_target. Find its root before the next iteration
* in mu_reorg.
*/
#define WANT_REDO_ROOT_SEARCH \
( (NULL != gv_target) \
&& (DIR_ROOT != gv_target->root) \
&& !redo_root_search_done \
&& !TREF(in_gvcst_redo_root_search) \
&& !mu_reorg_upgrd_dwngrd_in_prog \
&& (!TREF(in_gvcst_bmp_mark_free) || mu_reorg_process) \
)
GBLREF sgmnt_addrs *cs_addrs;
GBLREF sgmnt_data_ptr_t cs_data;
GBLREF short crash_count;
GBLREF uint4 dollar_tlevel;
GBLREF gd_region *gv_cur_region;
GBLREF gv_key *gv_currkey;
GBLREF gv_namehead *gv_target;
GBLREF tp_frame *tp_pointer;
GBLREF trans_num start_tn;
GBLREF unsigned char cw_set_depth, cw_map_depth, t_fail_hist[CDB_MAX_TRIES];
GBLREF boolean_t mu_reorg_process;
GBLREF boolean_t mu_reorg_upgrd_dwngrd_in_prog;
GBLREF unsigned int t_tries;
GBLREF uint4 t_err;
GBLREF jnl_gbls_t jgbl;
GBLREF boolean_t is_dollar_incr;
GBLREF uint4 update_trans;
#ifdef GTM_TRIGGER
GBLREF boolean_t skip_INVOKE_RESTART;
#endif
#ifdef DEBUG
GBLDEF unsigned char t_fail_hist_dbg[T_FAIL_HIST_DBG_SIZE];
GBLDEF unsigned int t_tries_dbg;
GBLREF sgm_info *sgm_info_ptr;
GBLREF boolean_t mupip_jnl_recover;
#endif
#ifdef UNIX
GBLREF boolean_t is_updproc;
GBLREF boolean_t need_kip_incr;
GBLREF sgmnt_addrs *kip_csa;
GBLREF jnlpool_addrs jnlpool;
#endif
#ifdef UNIX
error_def(ERR_DBROLLEDBACK);
error_def(ERR_GVFAILCORE);
error_def(ERR_REPLONLNRLBK);
#endif
error_def(ERR_GBLOFLOW);
error_def(ERR_GVINCRFAIL);
error_def(ERR_GVIS);
error_def(ERR_GVPUTFAIL);
error_def(ERR_TPRETRY);
void t_retry(enum cdb_sc failure)
{
tp_frame *tf;
unsigned char *end, buff[MAX_ZWR_KEY_SZ];
short tl;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
node_local_ptr_t cnl;
# ifdef DEBUG
unsigned int tries;
# endif
boolean_t skip_invoke_restart;
boolean_t redo_root_search_done = FALSE;
unsigned int local_t_tries;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
# ifdef GTM_TRIGGER
skip_invoke_restart = skip_INVOKE_RESTART; /* note down global value in local variable */
GTMTRIG_ONLY(DBGTRIGR((stderr, "t_retry: entered\n")));
# else
skip_invoke_restart = FALSE; /* no triggers so set local variable to default state */
# endif
/* We expect t_retry to be invoked with an abnormal failure code. mupip reorg is the only exception and can pass
* cdb_sc_normal failure code in case it finds a global variable existed at start of reorg, but not when it came
* into mu_reorg and did a gvcst_search. It cannot confirm this unless it holds crit for which it has to wait
* until the final retry which is why we accept this way of invoking t_retry. Assert accordingly.
*/
assert((cdb_sc_normal != failure) || mu_reorg_process);
t_fail_hist[t_tries] = (unsigned char)failure;
if (mu_reorg_process)
CWS_RESET;
DEBUG_ONLY(TREF(donot_commit) = FALSE;)
csa = cs_addrs;
cnl = csa ? csa->nl : NULL; /* making sure we do not try to dereference a NULL pointer */
if (!dollar_tlevel)
{
# ifdef DEBUG
if (0 == t_tries)
t_tries_dbg = 0;
assert(ARRAYSIZE(t_fail_hist_dbg) > t_tries_dbg);
t_fail_hist_dbg[t_tries_dbg++] = (unsigned char)failure;
TRACE_TRANS_RESTART(failure);
# endif
# ifdef UNIX
if (cdb_sc_instancefreeze == failure)
{
assert(REPL_ALLOWED(csa->hdr)); /* otherwise, a cdb_sc_instancefreeze retry would not have been signalled */
WAIT_FOR_REPL_INST_UNFREEZE(csa);
}
# endif
/* Even though rollback and recover operate standalone, there are certain kind of restarts that can still happen
* either due to whitebox test cases or stomping on our own buffers causing cdb_sc_lostcr/cdb_sc_rmisalign. Assert
* accordingly
*/
assert(!mupip_jnl_recover || WB_COMMIT_ERR_ENABLED || (CDB_STAGNATE > t_tries));
SET_WC_BLOCKED_FINAL_RETRY_IF_NEEDED(csa, cnl, failure); /* set wc_blocked if cache related status */
TREF(prev_t_tries) = t_tries;
TREF(rlbk_during_redo_root) = FALSE;
switch(t_tries)
{
case 0:
INCR_GVSTATS_COUNTER(csa, cnl, n_nontp_retries_0, 1);
break;
case 1:
INCR_GVSTATS_COUNTER(csa, cnl, n_nontp_retries_1, 1);
break;
case 2:
INCR_GVSTATS_COUNTER(csa, cnl, n_nontp_retries_2, 1);
break;
default:
assert(3 == t_tries);
INCR_GVSTATS_COUNTER(csa, cnl, n_nontp_retries_3, 1);
break;
}
if (csa->critical)
crash_count = csa->critical->crashcnt;
/* If the restart code is something that should not increment t_tries, handle that by decrementing t_tries
* for these special codes just before incrementing it unconditionally. Note that this should be done ONLY IF
* t_tries is CDB_STAGNATE or higher and not for lower values as otherwise it can cause livelocks (e.g.
* because cnl->wc_blocked is set to TRUE, it is possible we end up restarting with cdb_sc_helpedout
* without even doing a cache-recovery (due to the fast path in t_end that does not invoke grab_crit in case
* of read-only transactions). In this case, not incrementing t_tries causes us to eternally retry
* the transaction with no one eventually grabbing crit and doing the cache-recovery).
*/
assert(CDB_STAGNATE >= t_tries);
if (CDB_STAGNATE <= t_tries)
{
assert(cdb_sc_bkupss_statemod != failure); /* backup and snapshot state change cannot happen in
* final retry as they need crit which is held by us */
/* The following type of restarts can happen in the final retry.
* (a) cdb_sc_jnlstatemod : This is expected because csa->jnl_state is noted from csd->jnl_state only
* if they are different INSIDE crit. Therefore it is possible that in the final retry one might start
* with a stale value of csa->jnl_state which is noticed only in t_end just before commit as a
* result of which we would restart. Such a restart is okay (instead of the checking for jnl state
* change during the beginning of final retry) since jnl state changes are considered infrequent that
* too in the final retry.
* (b) cdb_sc_jnlclose : journaling might get turned off in the final retry INSIDE crit while trying to
* flush journal buffer or during extending the journal file (due to possible disk issues) in which
* case we come here with t_tries = CDB_STAGNATE.
* (c) cdb_sc_helpedout : cnl->wc_blocked being TRUE as well as file extension in MM (both of which is
* caused due to another process) can happen in final retry with failure status set to cdb_sc_helpedout
* (d) cdb_sc_needcrit : See GTM-7004 for how this is possible and why only a max of one such restart
* per non-TP transaction is possible.
* (e) cdb_sc_onln_rlbk[1,2] : See comment below as to why we allow online rollback related restarts even
* in the final retry.
* (f) cdb_sc_instancefreeze : Instance freeze detected while crit held.
* (g) cdb_sc_gvtrootmod2 : Similar to (e).
*/
if ((cdb_sc_jnlstatemod == failure) || (cdb_sc_jnlclose == failure) || (cdb_sc_helpedout == failure)
|| (cdb_sc_needcrit == failure) || (cdb_sc_onln_rlbk1 == failure)
|| (cdb_sc_onln_rlbk2 == failure) || (cdb_sc_instancefreeze == failure)
|| (cdb_sc_gvtrootmod2 == failure))
{
/* t_tries should never be greater than t_tries_dbg. The only exception is if this is DSE or online
* rollback operates with t_tries = CDB_STAGNATE and restarts if wc_blocked is set outside crit.
* But that's possible only if white box test cases to induce Phase 1 and Phase 2 errors are set.
* So, assert accordingly.
*/
assert((t_tries <= t_tries_dbg) UNIX_ONLY(|| (csa->hold_onto_crit && WB_COMMIT_ERR_ENABLED)));
/* Assert that the same kind of restart code can never occur more than once once we go to the
* final retry. The only exception is cdb_sc_helpedout which can happen due to other processes
* setting cnl->wc_blocked to TRUE without holding crit.
*/
assert(failure == t_fail_hist_dbg[t_tries_dbg - 1]);
DEBUG_ONLY(
for (tries = CDB_STAGNATE; tries < t_tries_dbg - 1; tries++)
assert((t_fail_hist_dbg[tries] != failure) || (cdb_sc_helpedout == failure));
)
t_tries = CDB_STAGNATE - 1;
}
}
if (CDB_STAGNATE <= ++t_tries)
{
DEBUG_ONLY(TREF(ok_to_call_wcs_recover) = TRUE;)
if (!csa->hold_onto_crit)
grab_crit(gv_cur_region);
else if (cnl->wc_blocked)
{ /* Possible ONLY for online rollback or DSE that grabs crit during startup and never grabs again.
* In such cases grab_crit (such as above) is skipped. As a result wcs_recover is also skipped.
* To avoid this, do wcs_recover if wc_blocked is TRUE. But, that's possible only if white box test
* cases to induce Phase 1 and Phase 2 errors are set. So, assert accordingly.
*/
assert(WB_COMMIT_ERR_ENABLED);
wcs_recover(gv_cur_region);
/* Note that if a concurrent Phase 2 error sets wc_blocked to TRUE anytime between now
* and the wc_blocked check at the top of t_end, we'll end up restarting and doing wcs_recover
* on the next restart.
*/
}
# ifdef UNIX
if (MISMATCH_ROOT_CYCLES(csa, cnl))
{ /* We came in to handle a different restart code in the penultimate retry and grab_crit before going
* to final retry. As part of grabbing crit, we detected an online rollback. Although we could treat
* this as just an online rollback restart and handle it by syncing cycles, but by doing so, we will
* loose the information that an online rollback happened when we go back to gvcst_{put,kill}. This
* is usually fine except when we are in implicit TP (due to triggers). In case of implicit TP,
* gvcst_{put,kill} has specific code to handle online rollback differently than other restart codes
* Because of this reason, we don't want to sync cycles but instead continue with the final retry.
* t_end/tp_tend/tp_hist will notice the cycle mismatch and will restart (once more) in final retry
* with the appropriate cdb_sc code which gvcst_put/gvcst_kill will intercept and act accordingly.
* Even if we are not syncing cycles, we need to do other basic cleanup to ensure the final retry
* proceeds smoothly.
*/
RESET_ALL_GVT_CLUES;
cw_set_depth = 0;
cw_map_depth = 0;
if (WANT_REDO_ROOT_SEARCH)
{
gvcst_redo_root_search();
redo_root_search_done = TRUE;
}
}
# endif
assert(csa->now_crit);
DEBUG_ONLY(TREF(ok_to_call_wcs_recover) = FALSE;)
csd = cs_data;
if (CDB_STAGNATE == t_tries)
{
if (csd->freeze && update_trans)
{ /* Final retry on an update transaction and region is frozen.
* Wait for it to be unfrozen and only then grab crit.
*/
GRAB_UNFROZEN_CRIT(gv_cur_region, csa, csd);
}
} else
{
assert((failure != cdb_sc_helpedout) && (failure != cdb_sc_jnlclose)
&& (failure != cdb_sc_jnlstatemod) && (failure != cdb_sc_bkupss_statemod)
&& (failure != cdb_sc_inhibitkills));
local_t_tries = t_tries;
if (!csa->hold_onto_crit)
{
rel_crit(gv_cur_region);
t_tries = 0;
}
if (NULL == (end = format_targ_key(buff, MAX_ZWR_KEY_SZ, gv_currkey, TRUE)))
end = &buff[MAX_ZWR_KEY_SZ - 1];
if (cdb_sc_gbloflow == failure)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_GBLOFLOW, 0, ERR_GVIS, 2, end - buff, buff);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_GBLOFLOW, 0, ERR_GVIS, 2, end - buff, buff);
}
if (IS_DOLLAR_INCREMENT)
{
assert(ERR_GVPUTFAIL == t_err);
t_err = ERR_GVINCRFAIL; /* print more specific error message */
}
UNIX_ONLY(send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) t_err, 2, local_t_tries, t_fail_hist,
ERR_GVIS, 2, end-buff, buff, ERR_GVFAILCORE));
#ifdef DEBUG
/* Core is not needed. We intentionally create this error. */
if (!gtm_white_box_test_case_enabled)
#endif
UNIX_ONLY(gtm_fork_n_core());
VMS_ONLY(send_msg_csa(CSA_ARG(csa) VARLSTCNT(8) t_err, 2, local_t_tries, t_fail_hist,
ERR_GVIS, 2, end-buff, buff));
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) t_err, 2, local_t_tries, t_fail_hist, ERR_GVIS, 2, end-buff,
buff);
}
}
CHECK_MM_DBFILEXT_REMAP_IF_NEEDED(csa, gv_cur_region);
if ((cdb_sc_blockflush == failure) && !CCP_SEGMENT_STATE(cnl, CCST_MASK_HAVE_DIRTY_BUFFERS))
{
assert(csa->hdr->clustered);
CCP_FID_MSG(gv_cur_region, CCTR_FLUSHLK);
ccp_userwait(gv_cur_region, CCST_MASK_HAVE_DIRTY_BUFFERS, 0, cnl->ccp_cycle);
}
cw_set_depth = 0;
cw_map_depth = 0;
/* In case triggers are supported, make sure we start with latest copy of file header's db_trigger_cycle
* to avoid unnecessary cdb_sc_triggermod type of restarts.
*/
GTMTRIG_ONLY(csa->db_trigger_cycle = csa->hdr->db_trigger_cycle);
GTMTRIG_ONLY(DBGTRIGR((stderr, "t_retry: csa->db_trigger_cycle updated to %d\n", csa->db_trigger_cycle)));
start_tn = csa->ti->curr_tn;
/* Note: If gv_target was NULL before the start of a transaction and the only operations done inside the transaction
* are trigger deletions causing bitmap free operations which got restarted due to a concurrent update, we can
* reach here with gv_target being NULL.
*/
if (NULL != gv_target)
gv_target->clue.end = 0;
# ifdef UNIX
if ((cdb_sc_onln_rlbk1 == failure) || (cdb_sc_onln_rlbk2 == failure))
{ /* restarted due to online rollback */
if (!redo_root_search_done)
RESET_ALL_GVT_CLUES;
if (!TREF(in_gvcst_bmp_mark_free) || mu_reorg_process)
{ /* Handle cleanup beyond just resetting clues */
if (cdb_sc_onln_rlbk2 == failure)
{
if (IS_MCODE_RUNNING || TREF(issue_DBROLLEDBACK_anyways))
{ /* We are in Non-TP and an online rollback too the database to a prior state. If we
* are in M code OR the caller has asked us to issue the DBROLLEDBACK rts_error
* unconditionally (MUPIP LOAD for eg.), then issue the DBROLLEDBACK. If this is M
* code we also increment $ZONLNRLBK ISV and do other necessary cleanup before
* issuing the rts_error. Instead of checking for M code, do the cleanup anyways
*/
assert(!is_updproc);
(TREF(dollar_zonlnrlbk))++;
/* Since "only_reset_clues_if_onln_rlbk" is FALSE, we are NOT in the second phase of
* KILL. So, assert that kip_csa is still NULL
*/
assert(NULL == kip_csa);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(1) ERR_DBROLLEDBACK);
}
}
assert(!redo_root_search_done);
if (WANT_REDO_ROOT_SEARCH)
gvcst_redo_root_search();
if (is_updproc)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(1) ERR_REPLONLNRLBK);
}
# ifdef DEBUG
else
{ /* Detected ONLINE ROLLBACK during second phase of KILLs in which case we don't want to do increment
* $ZONLNRLBK or SYNC cycles. Instead we will stop the second phase of the KILLs and return to the
* caller to continue with the next transaction at which point we will detect ONLINE ROLLBACK again
* and take the appropriate action.
* Note: as long as we are in Non-TP, kip_csa will be NULL in second phase of KILL. Only exception
* is if we started out as TP and did KILLs and after the commit, invoked gvcst_bmp_mark_free to
* complete the second phase of the KILL. So, assert accordingly.
*/
assert((NULL != kip_csa) || ((NULL != sgm_info_ptr) && (NULL != sgm_info_ptr->kip_csa)));
/* Note: DECR_KIP done by gvcst_kill (in case of Non-TP) or op_tcommit (in case of TP) takes care
* of resetting kip_csa and decrementing cs_data->kill_in_prog. So, we don't need to do it here
* explicitly.
*/
}
# endif
}
if (cdb_sc_gvtrootmod == failure) /* failure signaled by gvcst_kill */
{ /* If "gvcst_redo_root_search" has not yet been invoked in t_retry, do that now */
assert(NULL != gv_target);
if (!redo_root_search_done && (NULL != gv_target) && (DIR_ROOT != gv_target->root))
gvcst_redo_root_search();
}
if (cdb_sc_gvtrootmod2 == failure)
{
if (!redo_root_search_done)
RESET_ALL_GVT_CLUES;
/* It is possible for a read-only transaction to release crit after detecting gvtrootmod2, during which time
* yet another root block could have moved. In that case, the MISMATCH_ROOT_CYCLES check would have
* already done the redo_root_search.
*/
assert(!redo_root_search_done || !update_trans);
if (WANT_REDO_ROOT_SEARCH)
{ /* Note: An online rollback can occur DURING gvcst_redo_root_search, which can remove gbls from db,
* leading to gv_target->root being 0, even though failure code is not cdb_sc_onln_rlbk2
*/
gvcst_redo_root_search();
}
}
# endif
} else
{ /* for TP, do the minimum; most of the logic is in tp_retry, because it is also invoked directly from t_commit */
assert(failure == t_fail_hist[t_tries]);
assert((NULL == csa) || (NULL != csa->hdr)); /* both csa and csa->hdr should be NULL or non-NULL. */
if (NULL != csa)
{
SET_WC_BLOCKED_FINAL_RETRY_IF_NEEDED(csa, cnl, failure);
TP_RETRY_ACCOUNTING(csa, cnl);
} else /* csa can be NULL if cur_reg is not open yet (cdb_sc_needcrit) */
assert((CDB_STAGNATE == t_tries) && (cdb_sc_needcrit == failure));
if (NULL != gv_target)
{
if (cdb_sc_blkmod != failure)
TP_TRACE_HIST(CR_BLKEMPTY, gv_target);
gv_target->clue.end = 0;
} else /* only known case of gv_target being NULL is if t_retry is done from gvcst_init. assert this below */
assert((CDB_STAGNATE <= t_tries) && ((cdb_sc_needcrit == failure) || have_crit(CRIT_HAVE_ANY_REG)));
if (!skip_invoke_restart)
{
GTMTRIG_ONLY(DBGTRIGR((stderr, "t_retry: invoking restart logic (INVOKE_RESTART)\n")));
INVOKE_RESTART;
} else /* explicit trigger update caused implicit tp wrap so should return to caller without rts_error */
{
GTMTRIG_ONLY(DBGTRIGR((stderr, "t_retry: invoking tp_restart directly\n")));
tp_restart(1, !TP_RESTART_HANDLES_ERRORS);
}
}
}