boolean_t gvcst_query(void)
{ /* Similar to gvcst_order and gvcst_zprevious. In each case we skip over hidden subscripts as needed.
*
* 1 2 3 NULL <--- order/zprev...
* 1 2 3 NULL NULL
* 1 2 3 NULL NULL NULL <--- query from here...
* 1 2 3 NULL NULL NULL hidden
* 1 2 3 NULL NULL hidden
* 1 2 3 NULL hidden
* 1 2 3 hidden <--- ... skip this guy and go to bottom/top, respectively
* 1 2 3 7 <--- ... needs to end up here
*/
boolean_t found, is_hidden, sn_tpwrapped;
boolean_t est_first_pass;
gv_key save_currkey[DBKEYALLOC(MAX_KEY_SZ)];
int i;
int save_dollar_tlevel;
DEBUG_ONLY(save_dollar_tlevel = dollar_tlevel);
found = gvcst_query2();
# ifdef UNIX
assert(save_dollar_tlevel == dollar_tlevel);
CHECK_HIDDEN_SUBSCRIPT_AND_RETURN(found, gv_altkey, is_hidden);
IF_SN_DISALLOWED_AND_NO_SPAN_IN_DB(return found);
assert(found && is_hidden);
SAVE_GV_CURRKEY(save_currkey);
if (!dollar_tlevel)
{
sn_tpwrapped = TRUE;
op_tstart((IMPLICIT_TSTART), TRUE, &literal_batch, 0);
ESTABLISH_NORET(gvcst_query_ch, est_first_pass);
GVCST_ROOT_SEARCH_AND_PREP(est_first_pass);
} else
sn_tpwrapped = FALSE;
for (i = 0; i <= MAX_GVSUBSCRIPTS; i++)
{
INCR_GVSTATS_COUNTER(cs_addrs, cs_addrs->nl, n_query, (gtm_uint64_t) -1);
found = gvcst_query2();
CHECK_HIDDEN_SUBSCRIPT_AND_BREAK(found, gv_altkey, is_hidden);
assert(found && is_hidden);
/* Replace last subscript to be the highest possible hidden subscript so another
* gvcst_query2 will give us the next non-hidden subscript.
*/
REPLACE_HIDDEN_SUB_TO_HIGHEST(gv_altkey, gv_currkey); /* uses gv_altkey to modify gv_currkey */
}
if (sn_tpwrapped)
{
op_tcommit();
REVERT; /* remove our condition handler */
}
RESTORE_GV_CURRKEY(save_currkey);
assert(save_dollar_tlevel == dollar_tlevel);
# endif
return found;
}
boolean_t gvcst_order(void)
{ /* See gvcst_query.c */
boolean_t found, is_hidden, sn_tpwrapped;
boolean_t est_first_pass;
gv_key save_currkey[DBKEYALLOC(MAX_KEY_SZ)];
int end, prev, oldend;
int save_dollar_tlevel;
DEBUG_ONLY(save_dollar_tlevel = dollar_tlevel);
found = gvcst_order2();
# ifdef UNIX
assert(save_dollar_tlevel == dollar_tlevel);
CHECK_HIDDEN_SUBSCRIPT_AND_RETURN(found, gv_altkey, is_hidden);
assert(found && is_hidden);
IF_SN_DISALLOWED_AND_NO_SPAN_IN_DB(return found);
SAVE_GV_CURRKEY_LAST_SUBSCRIPT(save_currkey, prev, oldend);
if (!dollar_tlevel)
{
sn_tpwrapped = TRUE;
op_tstart((IMPLICIT_TSTART), TRUE, &literal_batch, 0);
ESTABLISH_NORET(gvcst_order_ch, est_first_pass);
GVCST_ROOT_SEARCH_AND_PREP(est_first_pass);
INCR_GVSTATS_COUNTER(cs_addrs, cs_addrs->nl, n_order, (gtm_uint64_t) -1);
found = gvcst_order2();
} else
sn_tpwrapped = FALSE;
if (found)
{
CHECK_HIDDEN_SUBSCRIPT(gv_altkey, is_hidden);
if (is_hidden)
{ /* Replace last subscript to be the highest possible hidden subscript so another
* gvcst_order2 will give us the next non-hidden subscript.
*/
REPLACE_HIDDEN_SUB_TO_HIGHEST(gv_altkey, gv_currkey); /* uses gv_altkey to modify gv_currkey */
/* fix up since it should only be externally counted as one $order */
INCR_GVSTATS_COUNTER(cs_addrs, cs_addrs->nl, n_order, (gtm_uint64_t) -1);
found = gvcst_order2();
}
}
if (sn_tpwrapped)
{
op_tcommit();
REVERT; /* remove our condition handler */
}
RESTORE_GV_CURRKEY_LAST_SUBSCRIPT(save_currkey, prev, oldend);
assert(save_dollar_tlevel == dollar_tlevel);
# endif
return found;
}
boolean_t gvcst_queryget(mval *val)
{
bool found, is_hidden, is_dummy = FALSE, sn_tpwrapped;
boolean_t est_first_pass;
char save_currkey[SIZEOF(gv_key) + DBKEYSIZE(MAX_KEY_SZ)];
gv_key *save_gv_currkey;
int save_dollar_tlevel;
DEBUG_ONLY(save_dollar_tlevel = dollar_tlevel);
found = gvcst_queryget2(val, NULL);
# ifdef UNIX
assert(save_dollar_tlevel == dollar_tlevel);
CHECK_HIDDEN_SUBSCRIPT(gv_altkey, is_hidden);
if (found && IS_SN_DUMMY(val->str.len, val->str.addr))
is_dummy = TRUE;
if (!found || (!is_dummy && !is_hidden))
return found;
IF_SN_DISALLOWED_AND_NO_SPAN_IN_DB(return found);
SAVE_GV_CURRKEY;
if (!dollar_tlevel)
{
sn_tpwrapped = TRUE;
op_tstart((IMPLICIT_TSTART), TRUE, &literal_batch, 0);
ESTABLISH_NORET(gvcst_queryget_ch, est_first_pass);
GVCST_ROOT_SEARCH_AND_PREP(est_first_pass);
} else
sn_tpwrapped = FALSE;
found = gvcst_query();
COPY_KEY(gv_currkey, gv_altkey); /* set gv_currkey to gv_altkey */
found = gvcst_get(val);
INCR_GVSTATS_COUNTER(cs_addrs, cs_addrs->nl, n_get, (gtm_uint64_t) -1); /* only counted externally as one get */
INCR_GVSTATS_COUNTER(cs_addrs, cs_addrs->nl, n_query, (gtm_uint64_t) -1);
if (sn_tpwrapped)
{
op_tcommit();
REVERT; /* remove our condition handler */
}
RESTORE_GV_CURRKEY;
assert(save_dollar_tlevel == dollar_tlevel);
# endif
return found;
}
/* Upgrade ^#t global in "reg" region */
void trigger_upgrade(gd_region *reg)
{
boolean_t est_first_pass, do_upgrade, is_defined;
boolean_t was_null = FALSE, is_null = FALSE;
int seq_num, trig_seq_num;
int currlabel;
mval tmpmval, xecuteimval, *gvname, *tmpmv, *tmpmv2;
int4 result, tmpint4;
uint4 curend, gvname_prev, xecute_curend;
uint4 hash_code, kill_hash_code;
int count, i, xecutei, tncount;
char *trigname, *trigindex, *ptr;
char name_and_index[MAX_MIDENT_LEN + 1 + MAX_DIGITS_IN_INT];
char trigvn[MAX_MIDENT_LEN + 1 + MAX_DIGITS_IN_INT], nullbyte[1];
uint4 trigname_len, name_index_len;
int ilen;
sgmnt_addrs *csa;
jnl_private_control *jpc;
uint4 sts;
int close_res;
hash128_state_t hash_state, kill_hash_state;
uint4 hash_totlen, kill_hash_totlen;
int trig_protected_mval_push_count;
# ifdef DEBUG
int save_dollar_tlevel;
# endif
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
assert(gv_cur_region == reg);
assert(!dollar_tlevel); /* caller should have ensured this. this is needed as otherwise things get complicated. */
assert(!is_replicator); /* caller should have ensured this. this is needed so we dont bump jnl_seqno (if replicating) */
csa = &FILE_INFO(reg)->s_addrs;
assert(csa->hdr->hasht_upgrade_needed);
/* If before-image journaling is turned on in this region (does not matter if replication is turned on or not),
* once this transaction is done, we need to switch to new journal file and cut the back link because
* otherwise it is possible for backward journal recovery (or rollback) or source server to encounter
* the journal records generated in this ^#t-upgrade-transaction in which case they dont know to handle
* it properly (e.g. rollback or backward recovery does not know to restore csa->hdr->hasht_upgrade_needed
* if it rolls back this transaction). To achieve this, we set hold_onto_crit to TRUE and do the jnl link
* cut AFTER the transaction commits but before anyone else can sneak in to do any more updates.
* Since most often we expect databases to be journaled, we do this hold_onto_crit even for the non-journaled case.
*/
grab_crit(reg);
csa->hold_onto_crit = TRUE;
DEBUG_ONLY(save_dollar_tlevel = dollar_tlevel);
assert(!donot_INVOKE_MUMTSTART);
DEBUG_ONLY(donot_INVOKE_MUMTSTART = TRUE);
op_tstart(IMPLICIT_TSTART, TRUE, &literal_batch, 0); /* 0 ==> save no locals but RESTART OK */
ESTABLISH_NORET(trigger_upgrade_ch, est_first_pass);
/* On a TP restart anywhere down below, this line is where the restart resumes execution from */
assert(donot_INVOKE_MUMTSTART); /* Make sure still set for every try/retry of TP transaction */
change_reg(); /* TP_CHANGE_REG wont work as we need to set sgm_info_ptr */
assert(NULL != cs_addrs);
assert(csa == cs_addrs);
SET_GVTARGET_TO_HASHT_GBL(csa); /* sets up gv_target */
assert(NULL != gv_target);
INITIAL_HASHT_ROOT_SEARCH_IF_NEEDED; /* Needed to do every retry in case restart was due to an online rollback.
* This also sets up gv_currkey */
/* Do actual upgrade of ^#t global.
*
* Below is a sample layout of the label 2 ^#t global
* -------------------------------------------------------
* ^#t("#TNAME","x")="a"_$C(0)_"1" (present in DEFAULT only)
* ^#t("#TRHASH",89771515,1)="a"_$C(0)_"1" (present in DEFAULT only)
* ^#t("#TRHASH",106937755,1)="a"_$C(0)_"1" (present in DEFAULT only)
* ^#t("a",1,"BHASH")="106937755"
* ^#t("a",1,"CHSET")="M"
* ^#t("a",1,"CMD")="S"
* ^#t("a",1,"LHASH")="89771515"
* ^#t("a",1,"TRIGNAME")="x#"
* ^#t("a",1,"XECUTE")=" do ^twork"
* ^#t("a","#COUNT")="1"
* ^#t("a","#CYCLE")="1"
* ^#t("a","#LABEL")="2"
*
* Below is a sample layout of the label 3 ^#t global
* -------------------------------------------------------
* ^#t("#LABEL")="3" (present only after upgrade, not regular trigger load)
* ^#t("#TNAME","x")="a"_$C(0)_"1" (present in CURRENT region)
* ^#t("a",1,"BHASH")="71945627"
* ^#t("a",1,"CHSET")="M"
* ^#t("a",1,"CMD")="S"
* ^#t("a",1,"LHASH")="71945627"
* ^#t("a",1,"TRIGNAME")="x#"
* ^#t("a",1,"XECUTE")=" do ^twork"
* ^#t("a","#COUNT")="1"
* ^#t("a","#CYCLE")="2"
* ^#t("a","#LABEL")="3"
* ^#t("a","#TRHASH",71945627,1)="a"_$C(0)_"1"
*
* Key aspects of the format change
* ----------------------------------
* 1) New ^#t("#LABEL")="3" to indicate the format of the ^#t global. This is in addition to
* ^#t("a","#LABEL") etc. which is already there. This way we have a #LABEL for not just the installed
* triggers but also for the name information stored in the #TNAME nodes.
* 2) In the BHASH and LHASH fields. The hash computation is different so there are more chances of BHASH and LHASH
* matching in which case we store only one #TRHASH entry (instead of two). So thre is fewer ^#t records in the new
* format in most cases.
* 3) ^#t("a","#LABEL") bumps from 2 to 3. Similarly ^#t("a","#CYCLE") bumps by one (to make sure triggers for this
* global get re-read if and when we implement an -ONLINE upgrade).
* 4) DEFAULT used to have ^#t("#TNAME",...) nodes corresponding to triggers across ALL regions in the gbldir and
* other regions used to have NO ^#t("#TNAME",...) nodes whereas after the upgrade every region have
* ^#t("#TNAME",...) nodes corresponding to triggers installed in that region. So it is safer to kill ^#t("#TNAME")
* nodes and add them as needed.
* 5) #TRHASH has moved from ^#t() to ^#t(<gbl>). So it is safer to kill ^#t("#TRHASH") nodes and add them as needed.
*
* Below is a sample layout of the label 4 ^#t global
* -------------------------------------------------------
* ^#t("#TNAME","x")="a"_$C(0)_"1" (present in CURRENT region)
* ^#t("a",1,"BHASH")="71945627"
* ^#t("a",1,"CHSET")="M"
* ^#t("a",1,"CMD")="S"
* ^#t("a",1,"LHASH")="71945627"
* ^#t("a",1,"TRIGNAME")="x#"
* ^#t("a",1,"XECUTE")=" do ^twork"
* ^#t("a","#COUNT")="1"
* ^#t("a","#CYCLE")="2"
* ^#t("a","#LABEL")="4"
* ^#t("a","#TRHASH",71945627,1)="a"_$C(0)_"1"
*
* Key aspects of the format change
* ----------------------------------
* 1) Removed ^#t("#LABEL") as it is redundant information and trigger load does not include it
* 2) Multiline triggers were incorrectly processed resulting in incorrect BHASH and LHASH values. Upgrade fixes this
* 3) ^#t("a","#LABEL") bumps from 3 to 4. Similarly ^#t("a","#CYCLE") bumps by one (to make sure
* triggers for this global get re-read if and when we implement an -ONLINE upgrade).
*/
tmpmv = &tmpmval; /* At all points maintain this relationship. The two are used interchangeably below */
if (gv_target->root)
do_upgrade = TRUE;
/* The below logic assumes ^#t global does not have any integrity errors */
assert(do_upgrade); /* caller should have not invoked us otherwise */
if (do_upgrade)
{ /* kill ^#t("#TRHASH"), ^#t("#TNAME") and ^#t("#LABEL") first. Regenerate each again as we process ^#t(<gbl>,...) */
csa->incr_db_trigger_cycle = TRUE; /* so that we increment csd->db_trigger_cycle at commit time.
* this forces concurrent processes to read upgraded triggers.
*/
if (JNL_WRITE_LOGICAL_RECS(csa))
{ /* Note that the ^#t upgrade is a physical layout change. But it has no logical change (i.e. users
* see the same MUPIP TRIGGER -SELECT output as before). So write only a dummy LGTRIG journal
* record for this operation. Hence write a string that starts with a trigger comment character ";".
*/
assert(!gv_cur_region->read_only);
jnl_format(JNL_LGTRIG, NULL, (mval *)&literal_trigjnlrec, 0);
}
/* KILL ^#t("#LABEL") unconditionally */
BUILD_HASHT_SUB_CURRKEY(LITERAL_HASHLABEL, STRLEN(LITERAL_HASHLABEL));
if (0 != gvcst_data())
gvcst_kill(TRUE);
/* KILL ^#t("#TNAME") unconditionally and regenerate */
BUILD_HASHT_SUB_CURRKEY(LITERAL_HASHTNAME, STRLEN(LITERAL_HASHTNAME));
if (0 != gvcst_data())
gvcst_kill(TRUE);
/* KILL ^#t("#TRHASH") unconditionally and regenerate */
BUILD_HASHT_SUB_CURRKEY(LITERAL_HASHTRHASH, STRLEN(LITERAL_HASHTRHASH));
if (0 != gvcst_data())
gvcst_kill(TRUE);
/* Loop through all global names for which ^#t(<gvn>) exists. The only first-level subscripts of ^#t starting
* with # are #TNAME and #TRHASH in collation order. So after #TRHASH we expect to find subscripts that are
* global names. Hence the HASHTRHASH code is placed AFTER the HASHTNAME code above.
*/
TREF(gd_targ_gvnh_reg) = NULL; /* needed so op_gvorder below goes through gvcst_order (i.e. focuses only
* on the current region) and NOT through gvcst_spr_order (which does not
* apply anyways in the case of ^#t).
*/
nullbyte[0] = '\0';
trig_protected_mval_push_count = 0;
INCR_AND_PUSH_MV_STENT(gvname); /* Protect gvname from garbage collection */
do
{
op_gvorder(gvname);
if (0 == gvname->str.len)
break;
assert(ARRAYSIZE(trigvn) > gvname->str.len);
memcpy(&trigvn[0], gvname->str.addr, gvname->str.len);
gvname->str.addr = &trigvn[0]; /* point away from stringpool to avoid stp_gcol issues */
/* Save gv_currkey->prev so it is restored before next call to op_gvorder (which cares about this field).
* gv_currkey->prev gets tampered with in the for loop below (e.g. BUILD_HASHT_SUB_CURRKEY macro).
* No need to do this for gv_currkey->end since the body of the for loop takes care of restoring it.
*/
gvname_prev = gv_currkey->prev;
BUILD_HASHT_SUB_CURRKEY(gvname->str.addr, gvname->str.len);
/* At this point, gv_currkey is ^#t(<gvn>) */
/* Increment ^#t(<gvn>,"#CYCLE") */
is_defined = gvtr_get_hasht_gblsubs((mval *)&literal_hashcycle, tmpmv);
assert(is_defined);
tmpint4 = mval2i(tmpmv);
tmpint4++;
i2mval(tmpmv, tmpint4);
gvtr_set_hasht_gblsubs((mval *)&literal_hashcycle, tmpmv);
/* Read ^#t(<gvn>,"#COUNT") */
is_defined = gvtr_get_hasht_gblsubs((mval *)&literal_hashcount, tmpmv);
if (is_defined)
{
tmpint4 = mval2i(tmpmv);
count = tmpint4;
/* Get ^#t(<gvn>,"#LABEL"), error out for invalid values. Upgrade disallowed for label 1 triggers */
is_defined = gvtr_get_hasht_gblsubs((mval *)&literal_hashlabel, tmpmv);
assert(is_defined);
currlabel = mval2i(tmpmv);
if ((V19_HASHT_GBL_LABEL_INT >= currlabel) || (HASHT_GBL_CURLABEL_INT <= currlabel))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_TRIGUPBADLABEL, 6, currlabel,
HASHT_GBL_CURLABEL_INT, gvname->str.len, gvname->str.addr,
REG_LEN_STR(reg));
/* Set ^#t(<gvn>,"#LABEL")=HASHT_GBL_CURLABEL */
gvtr_set_hasht_gblsubs((mval *)&literal_hashlabel, (mval *)&literal_curlabel);
} else
count = 0;
/* Kill ^#t(<gvn>,"#TRHASH") unconditionally and regenerate */
gvtr_kill_hasht_gblsubs((mval *)&literal_hashtrhash, TRUE);
/* At this point, gv_currkey is ^#t(<gvn>) */
for (i = 1; i <= count; i++)
{
/* At this point, gv_currkey is ^#t(<gvn>) */
curend = gv_currkey->end; /* note gv_currkey->end before changing it so we can restore it later */
assert(KEY_DELIMITER == gv_currkey->base[curend]);
assert(gv_target->gd_csa == cs_addrs);
i2mval(tmpmv, i);
COPY_SUBS_TO_GVCURRKEY(tmpmv, gv_cur_region, gv_currkey, was_null, is_null);
/* At this point, gv_currkey is ^#t(<gvn>,i) */
/* Compute new LHASH and BHASH hash values.
* LHASH uses : GVSUBS, XECUTE
* BHASH uses : GVSUBS, DELIM, ZDELIM, PIECES, XECUTE
* So reach each of these pieces and compute hash along the way.
*/
STR_PHASH_INIT(hash_state, hash_totlen);
STR_PHASH_PROCESS(hash_state, hash_totlen, gvname->str.addr, gvname->str.len);
STR_PHASH_PROCESS(hash_state, hash_totlen, nullbyte, 1);
/* Read in ^#t(<gvn>,i,"GVSUBS") */
is_defined = gvtr_get_hasht_gblsubs((mval *)&literal_gvsubs, tmpmv);
if (is_defined)
{
STR_PHASH_PROCESS(hash_state, hash_totlen, tmpmval.str.addr, tmpmval.str.len);
STR_PHASH_PROCESS(hash_state, hash_totlen, nullbyte, 1);
}
/* Copy over SET hash state (2-tuple <state,totlen>) to KILL hash state before adding
* the PIECES, DELIM, ZDELIM portions (those are only part of the SET hash).
*/
kill_hash_state = hash_state;
kill_hash_totlen = hash_totlen;
/* Read in ^#t(<gvn>,i,"PIECES") */
is_defined = gvtr_get_hasht_gblsubs((mval *)&literal_pieces, tmpmv);
if (is_defined)
{
STR_PHASH_PROCESS(hash_state, hash_totlen, tmpmval.str.addr, tmpmval.str.len);
STR_PHASH_PROCESS(hash_state, hash_totlen, nullbyte, 1);
}
/* Read in ^#t(<gvn>,i,"DELIM") */
is_defined = gvtr_get_hasht_gblsubs((mval *)&literal_delim, tmpmv);
if (is_defined)
{
STR_PHASH_PROCESS(hash_state, hash_totlen, tmpmval.str.addr, tmpmval.str.len);
STR_PHASH_PROCESS(hash_state, hash_totlen, nullbyte, 1);
}
/* Read in ^#t(<gvn>,i,"ZDELIM") */
is_defined = gvtr_get_hasht_gblsubs((mval *)&literal_zdelim, tmpmv);
if (is_defined)
{
STR_PHASH_PROCESS(hash_state, hash_totlen, tmpmval.str.addr, tmpmval.str.len);
STR_PHASH_PROCESS(hash_state, hash_totlen, nullbyte, 1);
}
/* Read in ^#t(<gvn>,i,"XECUTE").
* Note: The XECUTE portion of the trigger definition is used in SET and KILL hash.
* But since we have started maintaining "hash_state" and "kill_hash_state" separately
* (due to PIECES, DELIM, ZDELIM) we need to update the hash for both using same input string.
*/
is_defined = gvtr_get_hasht_gblsubs((mval *)&literal_xecute, tmpmv);
if (is_defined)
{
STR_PHASH_PROCESS(hash_state, hash_totlen, tmpmval.str.addr, tmpmval.str.len);
STR_PHASH_PROCESS(kill_hash_state, kill_hash_totlen, tmpmval.str.addr, tmpmval.str.len);
} else
{ /* Multi-record XECUTE string */
/* At this point, gv_currkey is ^#t(<gvn>,i) */
xecute_curend = gv_currkey->end; /* note gv_currkey->end so we can restore it later */
assert(KEY_DELIMITER == gv_currkey->base[xecute_curend]);
tmpmv2 = (mval *)&literal_xecute;
COPY_SUBS_TO_GVCURRKEY(tmpmv2, gv_cur_region, gv_currkey, was_null, is_null);
xecutei = 1;
do
{
i2mval(&xecuteimval, xecutei);
is_defined = gvtr_get_hasht_gblsubs(&xecuteimval, tmpmv);
if (!is_defined)
break;
STR_PHASH_PROCESS(hash_state, hash_totlen, tmpmval.str.addr, tmpmval.str.len);
STR_PHASH_PROCESS(kill_hash_state, kill_hash_totlen,
tmpmval.str.addr, tmpmval.str.len);
xecutei++;
} while (TRUE);
/* Restore gv_currkey to ^#t(<gvn>,i) */
gv_currkey->end = xecute_curend;
gv_currkey->base[xecute_curend] = KEY_DELIMITER;
}
STR_PHASH_RESULT(hash_state, hash_totlen, hash_code);
STR_PHASH_RESULT(kill_hash_state, kill_hash_totlen, kill_hash_code);
/* Set ^#t(<gvn>,i,"LHASH") */
MV_FORCE_UMVAL(tmpmv, kill_hash_code);
gvtr_set_hasht_gblsubs((mval *)&literal_lhash, tmpmv);
/* Set ^#t(<gvn>,i,"BHASH") */
MV_FORCE_UMVAL(tmpmv, hash_code);
gvtr_set_hasht_gblsubs((mval *)&literal_bhash, tmpmv);
/* Read in ^#t(<gvn>,i,"TRIGNAME") to determine if #SEQNUM/#TNCOUNT needs to be maintained */
is_defined = gvtr_get_hasht_gblsubs((mval *)&literal_trigname, tmpmv);
assert(is_defined);
assert('#' == tmpmval.str.addr[tmpmval.str.len - 1]);
tmpmval.str.len--;
if ((tmpmval.str.len <= ARRAYSIZE(name_and_index)) &&
(NULL != (ptr = memchr(tmpmval.str.addr, '#', tmpmval.str.len))))
{ /* Auto-generated name. Need to maintain #SEQNUM/#TNCOUNT */
/* Take copy of trigger name into non-stringpool location to avoid stp_gcol issues */
trigname_len = ptr - tmpmval.str.addr;
ptr++;
name_index_len = (tmpmval.str.addr + tmpmval.str.len) - ptr;
assert(ARRAYSIZE(name_and_index) >= (trigname_len + 1 + name_index_len));
trigname = &name_and_index[0];
trigindex = ptr;
memcpy(trigname, tmpmval.str.addr, tmpmval.str.len);
A2I(ptr, ptr + name_index_len, trig_seq_num);
/* At this point, gv_currkey is ^#t(<gvn>,i) */
/* $get(^#t("#TNAME",<trigger name>,"#SEQNUM")) */
BUILD_HASHT_SUB_SUB_SUB_CURRKEY(LITERAL_HASHTNAME, STR_LIT_LEN(LITERAL_HASHTNAME),
trigname, trigname_len, LITERAL_HASHSEQNUM, STR_LIT_LEN(LITERAL_HASHSEQNUM));
seq_num = gvcst_get(tmpmv) ? mval2i(tmpmv) : 0;
if (trig_seq_num > seq_num)
{ /* Set ^#t("#TNAME",<trigger name>,"#SEQNUM") = trig_seq_num */
SET_TRIGGER_GLOBAL_SUB_SUB_SUB_STR(LITERAL_HASHTNAME,
STR_LIT_LEN(LITERAL_HASHTNAME), trigname, trigname_len,
LITERAL_HASHSEQNUM, STR_LIT_LEN(LITERAL_HASHSEQNUM),
trigindex, name_index_len, result);
assert(PUT_SUCCESS == result);
}
/* set ^#t("#TNAME",<trigger name>,"#TNCOUNT")++ */
BUILD_HASHT_SUB_SUB_SUB_CURRKEY(LITERAL_HASHTNAME, STR_LIT_LEN(LITERAL_HASHTNAME),
trigname, trigname_len, LITERAL_HASHTNCOUNT, STR_LIT_LEN(LITERAL_HASHTNCOUNT));
tncount = gvcst_get(tmpmv) ? mval2i(tmpmv) + 1 : 1;
i2mval(tmpmv, tncount);
SET_TRIGGER_GLOBAL_SUB_SUB_SUB_MVAL(LITERAL_HASHTNAME, STR_LIT_LEN(LITERAL_HASHTNAME),
trigname, trigname_len, LITERAL_HASHTNCOUNT, STR_LIT_LEN(LITERAL_HASHTNCOUNT),
tmpmval, result);
trigname_len += 1 + name_index_len; /* in preparation for ^#t("#TNAME") set below */
assert(PUT_SUCCESS == result);
BUILD_HASHT_SUB_CURRKEY(gvname->str.addr, gvname->str.len);
/* At this point, gv_currkey is ^#t(<gvn>) */
} else
{
/* Take copy of trigger name into non-stringpool location to avoid stp_gcol issues */
trigname = &name_and_index[0]; /* in preparation for ^#t("#TNAME") set below */
trigname_len = MIN(tmpmval.str.len, ARRAYSIZE(name_and_index));
assert(ARRAYSIZE(name_and_index) >= trigname_len);
memcpy(trigname, tmpmval.str.addr, trigname_len);
/* Restore gv_currkey to what it was at beginning of for loop iteration */
gv_currkey->end = curend;
gv_currkey->base[curend] = KEY_DELIMITER;
}
/* At this point, gv_currkey is ^#t(<gvn>) */
if (kill_hash_code != hash_code)
gvtr_set_hashtrhash(gvname->str.addr, gvname->str.len, kill_hash_code, i);
/* Set ^#t(<gvn>,"#TRHASH",hash_code,i) */
gvtr_set_hashtrhash(gvname->str.addr, gvname->str.len, hash_code, i);
/* Set ^#t("#TNAME",<trigname>)=<gvn>_$c(0)_<trigindx> */
/* The upgrade assumes that the region does not contain two triggers with the same name.
* V62000 and before could potentially have this out of design case. Once implemented
* the trigger integrity check will warn users of this edge case */
ptr = &trigvn[gvname->str.len];
*ptr++ = '\0';
ilen = 0;
I2A(ptr, ilen, i);
ptr += ilen;
assert(ptr <= ARRAYTOP(trigvn));
SET_TRIGGER_GLOBAL_SUB_SUB_STR(LITERAL_HASHTNAME, STR_LIT_LEN(LITERAL_HASHTNAME),
trigname, trigname_len, trigvn, ptr - gvname->str.addr, result);
assert(PUT_SUCCESS == result);
BUILD_HASHT_SUB_CURRKEY(gvname->str.addr, gvname->str.len);
/* At this point, gv_currkey is ^#t(<gvn>) */
}
/* At this point, gv_currkey is ^#t(<gvn>) i.e. gv_currkey->end is correct but gv_currkey->prev
* might have been tampered with. Restore it to proper value first.
*/
gv_currkey->prev = gvname_prev;
gvname->mvtype = 0; /* can now be garbage collected in the next iteration */
} while (TRUE);
}
op_tcommit();
REVERT; /* remove our condition handler */
DEBUG_ONLY(donot_INVOKE_MUMTSTART = FALSE;)
if (csa->hold_onto_crit)
int4 gds_rundown(void)
{
boolean_t canceled_dbsync_timer, canceled_flush_timer, ok_to_write_pfin;
boolean_t have_standalone_access, ipc_deleted, err_caught;
boolean_t is_cur_process_ss_initiator, remove_shm, vermismatch, we_are_last_user, we_are_last_writer, is_mm;
boolean_t unsafe_last_writer;
char time_str[CTIME_BEFORE_NL + 2]; /* for GET_CUR_TIME macro */
gd_region *reg;
int save_errno, status, rc;
int4 semval, ftok_semval, sopcnt, ftok_sopcnt;
short crash_count;
sm_long_t munmap_len;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
node_local_ptr_t cnl;
struct shmid_ds shm_buf;
struct sembuf sop[2], ftok_sop[2];
uint4 jnl_status;
unix_db_info *udi;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
shm_snapshot_t *ss_shm_ptr;
uint4 ss_pid, onln_rlbk_pid, holder_pid;
boolean_t was_crit;
boolean_t safe_mode; /* Do not flush or take down shared memory. */
boolean_t bypassed_ftok = FALSE, bypassed_access = FALSE, may_bypass_ftok, inst_is_frozen,
ftok_counter_halted,
access_counter_halted;
int secshrstat;
intrpt_state_t prev_intrpt_state;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
jnl_status = 0;
reg = gv_cur_region; /* Local copy */
/* early out for cluster regions
* to avoid tripping the assert below.
* Note:
* This early out is consistent with VMS. It has been
* noted that all of the gtcm assignments
* to gv_cur_region should use the TP_CHANGE_REG
* macro. This would also avoid the assert problem
* and should be done eventually.
*/
if (dba_cm == reg->dyn.addr->acc_meth)
return EXIT_NRM;
udi = FILE_INFO(reg);
csa = &udi->s_addrs;
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
if ((reg->open) && (dba_usr == csd->acc_meth))
{
change_reg();
gvusr_rundown();
return EXIT_NRM;
}
/* If the process has standalone access, it has udi->grabbed_access_sem set to TRUE at this point. Note that down in a local
* variable as the udi->grabbed_access_sem is set to TRUE even for non-standalone access below and hence we can't rely on
* that later to determine if the process had standalone access or not when it entered this function. We need to guarantee
* that none else access database file header when semid/shmid fields are reset. We already have created ftok semaphore in
* db_init or, mu_rndwn_file and did not remove it. So just lock it. We do it in blocking mode.
*/
have_standalone_access = udi->grabbed_access_sem; /* process holds standalone access */
DEFER_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
ESTABLISH_NORET(gds_rundown_ch, err_caught);
if (err_caught)
{
REVERT;
WITH_CH(gds_rundown_ch, gds_rundown_err_cleanup(have_standalone_access), 0);
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
DEBUG_ONLY(ok_to_UNWIND_in_exit_handling = FALSE);
return EXIT_ERR;
}
assert(reg->open); /* if we failed to open, dbinit_ch should have taken care of proper clean up */
assert(!reg->opening); /* see comment above */
assert((dba_bg == csd->acc_meth) || (dba_mm == csd->acc_meth));
is_mm = (dba_bg != csd->acc_meth);
assert(!csa->hold_onto_crit || (csa->now_crit && jgbl.onlnrlbk));
/* If we are online rollback, we should already be holding crit and should release it only at the end of this module. This
* is usually done by noting down csa->now_crit in a local variable (was_crit) and using it whenever we are about to
* grab_crit. But, there are instances (like mupip_set_journal.c) where we grab_crit but invoke gds_rundown without any
* preceeding rel_crit. Such code relies on the fact that gds_rundown does rel_crit unconditionally (to get locks to a known
* state). So, augment csa->now_crit with jgbl.onlnrlbk to track if we can rel_crit unconditionally or not in gds_rundown.
*/
was_crit = (csa->now_crit && jgbl.onlnrlbk);
/* Cancel any pending flush timer for this region by this task */
canceled_flush_timer = FALSE;
canceled_dbsync_timer = FALSE;
CANCEL_DB_TIMERS(reg, csa, canceled_flush_timer, canceled_dbsync_timer);
we_are_last_user = FALSE;
inst_is_frozen = IS_REPL_INST_FROZEN && REPL_ALLOWED(csa->hdr);
if (!csa->persistent_freeze)
region_freeze(reg, FALSE, FALSE, FALSE);
if (!was_crit)
{
rel_crit(reg); /* get locks to known state */
mutex_cleanup(reg);
}
/* The only process that can invoke gds_rundown while holding access control semaphore is RECOVER/ROLLBACK. All the others
* (like MUPIP SET -FILE/MUPIP EXTEND would have invoked db_ipcs_reset() before invoking gds_rundown (from
* mupip_exit_handler). The only exception is when these processes encounter a terminate signal and they reach
* mupip_exit_handler while holding access control semaphore. Assert accordingly.
*/
assert(!have_standalone_access || mupip_jnl_recover || process_exiting);
/* If we have standalone access, then ensure that a concurrent online rollback cannot be running at the same time as it
* needs the access control lock as well. The only expection is we are online rollback and currently running down.
*/
cnl = csa->nl;
onln_rlbk_pid = cnl->onln_rlbk_pid;
assert(!have_standalone_access || mupip_jnl_recover || !onln_rlbk_pid || !is_proc_alive(onln_rlbk_pid, 0));
if (!have_standalone_access)
{
if (-1 == (ftok_semval = semctl(udi->ftok_semid, DB_COUNTER_SEM, GETVAL))) /* Check # of procs counted on FTOK */
{
save_errno = errno;
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get ftok_semval"), CALLFROM, errno);
}
may_bypass_ftok = CAN_BYPASS(ftok_semval, csd, inst_is_frozen); /* Do we need a blocking wait? */
/* We need to guarantee that no one else access database file header when semid/shmid fields are reset.
* We already have created ftok semaphore in db_init or mu_rndwn_file and did not remove it. So just lock it.
*/
if (!ftok_sem_lock(reg, may_bypass_ftok))
{
if (may_bypass_ftok)
{ /* We did a non-blocking wait. It's ok to proceed without locking */
bypassed_ftok = TRUE;
holder_pid = semctl(udi->ftok_semid, DB_CONTROL_SEM, GETPID);
if ((uint4)-1 == holder_pid)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get holder_pid"),
CALLFROM, errno);
if (!IS_GTM_IMAGE) /* MUMPS processes should not flood syslog with bypass messages. */
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_RESRCINTRLCKBYPAS, 10,
LEN_AND_STR(gtmImageNames[image_type].imageName), process_id, LEN_AND_LIT("FTOK"),
REG_LEN_STR(reg), DB_LEN_STR(reg), holder_pid);
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_TEXT, 2,
LEN_AND_LIT("FTOK bypassed at rundown"));
}
} else
{ /* We did a blocking wait but something bad happened. */
FTOK_TRACE(csa, csa->ti->curr_tn, ftok_ops_lock, process_id);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
}
}
sop[0].sem_num = DB_CONTROL_SEM; sop[0].sem_op = 0; /* Wait for 0 */
sop[1].sem_num = DB_CONTROL_SEM; sop[1].sem_op = 1; /* Lock */
sopcnt = 2;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO | IPC_NOWAIT; /* Don't wait the first time thru */
SEMOP(udi->semid, sop, sopcnt, status, NO_WAIT);
if (0 != status)
{
save_errno = errno;
/* Check # of processes counted on access sem. */
if (-1 == (semval = semctl(udi->semid, DB_COUNTER_SEM, GETVAL)))
{
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get semval"), CALLFROM, errno);
}
bypassed_access = CAN_BYPASS(semval, csd, inst_is_frozen) || onln_rlbk_pid || csd->file_corrupt;
/* Before attempting again in the blocking mode, see if the holding process is an online rollback.
* If so, it is likely we won't get the access control semaphore anytime soon. In that case, we
* are better off skipping rundown and continuing with sanity cleanup and exit.
*/
holder_pid = semctl(udi->semid, DB_CONTROL_SEM, GETPID);
if ((uint4)-1 == holder_pid)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get holder_pid"), CALLFROM, errno);
if (!bypassed_access)
{ /* We couldn't get it in one shot-- see if we already have it */
if (holder_pid == process_id)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(5) MAKE_MSG_INFO(ERR_CRITSEMFAIL), 2, DB_LEN_STR(reg),
ERR_RNDWNSEMFAIL);
REVERT;
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
assert(FALSE);
return EXIT_ERR;
}
if (EAGAIN != save_errno)
{
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMOP on access control semaphore"),
CALLFROM, save_errno);
}
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO; /* Try again - blocking this time */
SEMOP(udi->semid, sop, 2, status, FORCED_WAIT);
if (-1 == status) /* We couldn't get it at all.. */
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMOP on access control semaphore"),
CALLFROM, errno);
} else if (!IS_GTM_IMAGE)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_RESRCINTRLCKBYPAS, 10,
LEN_AND_STR(gtmImageNames[image_type].imageName), process_id,
LEN_AND_LIT("access control"), REG_LEN_STR(reg), DB_LEN_STR(reg), holder_pid);
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_TEXT, 2,
LEN_AND_LIT("Access control bypassed at rundown"));
}
udi->grabbed_access_sem = !bypassed_access;
}
} /* else we we hold the access control semaphore and therefore have standalone access. We do not release it now - we
* release it later in mupip_exit_handler.c. Since we already hold the access control semaphore, we don't need the
* ftok semaphore and trying it could cause deadlock
*/
/* Note that in the case of online rollback, "udi->grabbed_access_sem" (and in turn "have_standalone_access") is TRUE.
* But there could be other processes still having the database open so we cannot safely reset the halted fields.
*/
if (have_standalone_access && !jgbl.onlnrlbk)
csd->ftok_counter_halted = csd->access_counter_halted = FALSE;
ftok_counter_halted = csd->ftok_counter_halted;
access_counter_halted = csd->access_counter_halted;
/* If we bypassed any of the semaphores, activate safe mode.
* Also, if the replication instance is frozen and this db has replication turned on (which means
* no flushes of dirty buffers to this db can happen while the instance is frozen) activate safe mode.
*/
ok_to_write_pfin = !(bypassed_access || bypassed_ftok || inst_is_frozen);
safe_mode = !ok_to_write_pfin || ftok_counter_halted || access_counter_halted;
/* At this point we are guaranteed no one else is doing a db_init/rundown as we hold the access control semaphore */
assert(csa->ref_cnt); /* decrement private ref_cnt before shared ref_cnt decrement. */
csa->ref_cnt--; /* Currently journaling logic in gds_rundown() in VMS relies on this order to detect last writer */
assert(!csa->ref_cnt);
--cnl->ref_cnt;
if (memcmp(cnl->now_running, gtm_release_name, gtm_release_name_len + 1))
{ /* VERMISMATCH condition. Possible only if DSE */
assert(dse_running);
vermismatch = TRUE;
} else
vermismatch = FALSE;
if (-1 == shmctl(udi->shmid, IPC_STAT, &shm_buf))
{
save_errno = errno;
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown shmctl"), CALLFROM, save_errno);
} else
we_are_last_user = (1 == shm_buf.shm_nattch) && !vermismatch && !safe_mode;
/* recover => one user except ONLINE ROLLBACK, or standalone with frozen instance */
assert(!have_standalone_access || we_are_last_user || jgbl.onlnrlbk || inst_is_frozen);
if (-1 == (semval = semctl(udi->semid, DB_COUNTER_SEM, GETVAL)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get semval"), CALLFROM, errno);
/* There's one writer left and I am it */
assert(reg->read_only || semval >= 0);
unsafe_last_writer = (DB_COUNTER_SEM_INCR == semval) && (FALSE == reg->read_only) && !vermismatch;
we_are_last_writer = unsafe_last_writer && !safe_mode;
assert(!we_are_last_writer || !safe_mode);
assert(!we_are_last_user || !safe_mode);
/* recover + R/W region => one writer except ONLINE ROLLBACK, or standalone with frozen instance, leading to safe_mode */
assert(!(have_standalone_access && !reg->read_only) || we_are_last_writer || jgbl.onlnrlbk || inst_is_frozen);
GTM_WHITE_BOX_TEST(WBTEST_ANTIFREEZE_JNLCLOSE, we_are_last_writer, 1); /* Assume we are the last writer to invoke wcs_flu */
if (!have_standalone_access && (-1 == (ftok_semval = semctl(udi->ftok_semid, DB_COUNTER_SEM, GETVAL))))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get ftok_semval"), CALLFROM, errno);
if (NULL != csa->ss_ctx)
ss_destroy_context(csa->ss_ctx);
/* SS_MULTI: If multiple snapshots are supported, then we have to run through each of the snapshots */
assert(1 == MAX_SNAPSHOTS);
ss_shm_ptr = (shm_snapshot_ptr_t)SS_GETSTARTPTR(csa);
ss_pid = ss_shm_ptr->ss_info.ss_pid;
is_cur_process_ss_initiator = (process_id == ss_pid);
if (ss_pid && (is_cur_process_ss_initiator || we_are_last_user))
{
/* Try getting snapshot crit latch. If we don't get latch, we won't hang for eternity and will skip
* doing the orphaned snapshot cleanup. It will be cleaned up eventually either by subsequent MUPIP
* INTEG or by a MUPIP RUNDOWN.
*/
if (ss_get_lock_nowait(reg) && (ss_pid == ss_shm_ptr->ss_info.ss_pid)
&& (is_cur_process_ss_initiator || !is_proc_alive(ss_pid, 0)))
{
ss_release(NULL);
ss_release_lock(reg);
}
}
/* If cnl->donotflush_dbjnl is set, it means mupip recover/rollback was interrupted and therefore we need not flush
* shared memory contents to disk as they might be in an inconsistent state. Moreover, any more flushing will only cause
* future rollback to undo more journal records (PBLKs). In this case, we will go ahead and remove shared memory (without
* flushing the contents) in this routine. A reissue of the recover/rollback command will restore the database to a
* consistent state.
*/
if (!cnl->donotflush_dbjnl && !reg->read_only && !vermismatch)
{ /* If we had an orphaned block and were interrupted, set wc_blocked so we can invoke wcs_recover. Do it ONLY
* if there is NO concurrent online rollback running (as we need crit to set wc_blocked)
*/
if (csa->wbuf_dqd && !is_mm)
{ /* If we had an orphaned block and were interrupted, mupip_exit_handler will invoke secshr_db_clnup which
* will clear this field and so we should never come to gds_rundown with a non-zero wbuf_dqd. The only
* exception is if we are recover/rollback in which case gds_rundown (from mur_close_files) is invoked
* BEFORE secshr_db_clnup in mur_close_files.
* Note: It is NOT possible for online rollback to reach here with wbuf_dqd being non-zero. This is because
* the moment we apply the first PBLK, we stop all interrupts and hence can never be interrupted in
* wcs_wtstart or wcs_get_space. Assert accordingly.
*/
assert(mupip_jnl_recover && !jgbl.onlnrlbk && !safe_mode);
if (!was_crit)
grab_crit(reg);
SET_TRACEABLE_VAR(cnl->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wcb_gds_rundown);
send_msg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_WCBLOCKED, 6, LEN_AND_LIT("wcb_gds_rundown"),
process_id, &csa->ti->curr_tn, DB_LEN_STR(reg));
csa->wbuf_dqd = 0;
wcs_recover(reg);
BG_TRACE_PRO_ANY(csa, lost_block_recovery);
if (!was_crit)
rel_crit(reg);
}
if (JNL_ENABLED(csd) && IS_GTCM_GNP_SERVER_IMAGE)
originator_prc_vec = NULL;
/* If we are the last writing user, then everything must be flushed */
if (we_are_last_writer)
{ /* Time to flush out all of our buffers */
assert(!safe_mode);
if (is_mm)
{
MM_DBFILEXT_REMAP_IF_NEEDED(csa, reg);
cnl->remove_shm = TRUE;
}
if (cnl->wc_blocked && jgbl.onlnrlbk)
{ /* if the last update done by online rollback was not committed in the normal code-path but was
* completed by secshr_db_clnup, wc_blocked will be set to TRUE. But, since online rollback never
* invokes grab_crit (since csa->hold_onto_crit is set to TRUE), wcs_recover is never invoked. This
* could result in the last update never getting flushed to the disk and if online rollback happened
* to be the last writer then the shared memory will be flushed and removed and the last update will
* be lost. So, force wcs_recover if we find ourselves in such a situation. But, wc_blocked is
* possible only if phase1 or phase2 errors are induced using white box test cases
*/
assert(WB_COMMIT_ERR_ENABLED);
wcs_recover(reg);
}
/* Note WCSFLU_SYNC_EPOCH ensures the epoch is synced to the journal and indirectly
* also ensures that the db is fsynced. We don't want to use it in the calls to
* wcs_flu() from t_end() and tp_tend() since we can defer it to out-of-crit there.
* In this case, since we are running down, we don't have any such option.
*/
cnl->remove_shm = wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
/* Since we_are_last_writer, we should be guaranteed that wcs_flu() did not change csd, (in
* case of MM for potential file extension), even if it did a grab_crit(). Therefore, make
* sure that's true.
*/
assert(csd == csa->hdr);
assert(0 == memcmp(csd->label, GDS_LABEL, GDS_LABEL_SZ - 1));
} else if (((canceled_flush_timer && (0 > cnl->wcs_timers)) || canceled_dbsync_timer) && !inst_is_frozen)
{ /* canceled pending db or jnl flush timers - flush database and journal buffers to disk */
if (!was_crit)
grab_crit(reg);
/* we need to sync the epoch as the fact that there is no active pending flush timer implies
* there will be noone else who will flush the dirty buffers and EPOCH to disk in a timely fashion
*/
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
if (!was_crit)
rel_crit(reg);
assert((dba_mm == cs_data->acc_meth) || (csd == cs_data));
csd = cs_data; /* In case this is MM and wcs_flu() remapped an extended database, reset csd */
}
/* Do rundown journal processing after buffer flushes since they require jnl to be open */
if (JNL_ENABLED(csd))
{ /* the following tp_change_reg() is not needed due to the assert csa == cs_addrs at the beginning
* of gds_rundown(), but just to be safe. To be removed by 2002!! --- nars -- 2001/04/25.
*/
tp_change_reg(); /* call this because jnl_ensure_open checks cs_addrs rather than gv_cur_region */
jpc = csa->jnl;
jbp = jpc->jnl_buff;
if (jbp->fsync_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->fsync_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (jbp->io_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->io_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if ((((NOJNL != jpc->channel) && !JNL_FILE_SWITCHED(jpc))
|| we_are_last_writer && (0 != cnl->jnl_file.u.inode)) && ok_to_write_pfin)
{ /* We need to close the journal file cleanly if we have the latest generation journal file open
* or if we are the last writer and the journal file is open in shared memory (not necessarily
* by ourselves e.g. the only process that opened the journal got shot abnormally)
* Note: we should not infer anything from the shared memory value of cnl->jnl_file.u.inode
* if we are not the last writer as it can be concurrently updated.
*/
if (!was_crit)
grab_crit(reg);
if (JNL_ENABLED(csd))
{
SET_GBL_JREC_TIME; /* jnl_ensure_open/jnl_put_jrt_pini/pfin/jnl_file_close all need it */
/* Before writing to jnlfile, adjust jgbl.gbl_jrec_time if needed to maintain time order
* of jnl records. This needs to be done BEFORE the jnl_ensure_open as that could write
* journal records (if it decides to switch to a new journal file).
*/
ADJUST_GBL_JREC_TIME(jgbl, jbp);
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{ /* If we_are_last_writer, we would have already done a wcs_flu() which would
* have written an epoch record and we are guaranteed no further updates
* since we are the last writer. So, just close the journal.
* If the freeaddr == post_epoch_freeaddr, wcs_flu may have skipped writing
* a pini, so allow for that.
*/
assert(!jbp->before_images || is_mm
|| !we_are_last_writer || (0 != jpc->pini_addr) || jgbl.mur_extract
|| (jpc->jnl_buff->freeaddr == jpc->jnl_buff->post_epoch_freeaddr));
/* If we haven't written a pini, let jnl_file_close write the pini/pfin. */
if (!jgbl.mur_extract && (0 != jpc->pini_addr))
jnl_put_jrt_pfin(csa);
/* If not the last writer and no pending flush timer left, do jnl flush now */
if (!we_are_last_writer && (0 > cnl->wcs_timers))
{
if (SS_NORMAL == (jnl_status = jnl_flush(reg)))
{
assert(jbp->freeaddr == jbp->dskaddr);
jnl_fsync(reg, jbp->dskaddr);
assert(jbp->fsync_dskaddr == jbp->dskaddr);
} else
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_JNLFLUSH, 2,
JNL_LEN_STR(csd), ERR_TEXT, 2,
RTS_ERROR_TEXT("Error with journal flush in gds_rundown"),
jnl_status);
assert(NOJNL == jpc->channel);/* jnl file lost has been triggered */
/* In this routine, all code that follows from here on does not
* assume anything about the journaling characteristics of this
* database so it is safe to continue execution even though
* journaling got closed in the middle.
*/
}
}
jnl_file_close(reg, we_are_last_writer, FALSE);
} else
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd),
DB_LEN_STR(reg));
}
if (!was_crit)
rel_crit(reg);
}
}
if (we_are_last_writer) /* Flush the fileheader last and harden the file to disk */
{
if (!was_crit)
grab_crit(reg); /* To satisfy crit requirement in fileheader_sync() */
memset(csd->machine_name, 0, MAX_MCNAMELEN); /* clear the machine_name field */
if (!have_standalone_access && we_are_last_user)
{ /* mupip_exit_handler will do this after mur_close_file */
csd->semid = INVALID_SEMID;
csd->shmid = INVALID_SHMID;
csd->gt_sem_ctime.ctime = 0;
csd->gt_shm_ctime.ctime = 0;
}
fileheader_sync(reg);
if (!was_crit)
rel_crit(reg);
if (!is_mm)
{
GTM_DB_FSYNC(csa, udi->fd, rc); /* Sync it all */
if (-1 == rc)
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
} else
{ /* Now do final MM file sync before exit */
assert(csa->ti->total_blks == csa->total_blks);
#ifdef _AIX
GTM_DB_FSYNC(csa, udi->fd, rc);
if (-1 == rc)
#else
if (-1 == MSYNC((caddr_t)csa->db_addrs[0], (caddr_t)csa->db_addrs[1]))
#endif
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
}
} else if (unsafe_last_writer && !cnl->lastwriterbypas_msg_issued)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_LASTWRITERBYPAS, 2, DB_LEN_STR(reg));
cnl->lastwriterbypas_msg_issued = TRUE;
}
} /* end if (!reg->read_only && !cnl->donotflush_dbjnl) */
/* We had canceled all db timers at start of rundown. In case as part of rundown (wcs_flu above), we had started
* any timers, cancel them BEFORE setting reg->open to FALSE (assert in wcs_clean_dbsync relies on this).
*/
CANCEL_DB_TIMERS(reg, csa, canceled_flush_timer, canceled_dbsync_timer);
if (reg->read_only && we_are_last_user && !have_standalone_access && cnl->remove_shm)
{ /* mupip_exit_handler will do this after mur_close_file */
db_ipcs.semid = INVALID_SEMID;
db_ipcs.shmid = INVALID_SHMID;
db_ipcs.gt_sem_ctime = 0;
db_ipcs.gt_shm_ctime = 0;
db_ipcs.fn_len = reg->dyn.addr->fname_len;
memcpy(db_ipcs.fn, reg->dyn.addr->fname, reg->dyn.addr->fname_len);
db_ipcs.fn[reg->dyn.addr->fname_len] = 0;
/* request gtmsecshr to flush. read_only cannot flush itself */
WAIT_FOR_REPL_INST_UNFREEZE_SAFE(csa);
if (!csa->read_only_fs)
{
secshrstat = send_mesg2gtmsecshr(FLUSH_DB_IPCS_INFO, 0, (char *)NULL, 0);
if (0 != secshrstat)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gtmsecshr failed to update database file header"));
}
}
/* Done with file now, close it */
CLOSEFILE_RESET(udi->fd, rc); /* resets "udi->fd" to FD_INVALID */
if (-1 == rc)
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during file close"), errno);
}
/* Unmap storage if mm mode but only the part that is not the fileheader (so shows up in dumps) */
# if !defined(_AIX)
if (is_mm && (NULL != csa->db_addrs[0]))
{
assert(csa->db_addrs[1] > csa->db_addrs[0]);
munmap_len = (sm_long_t)(csa->db_addrs[1] - csa->db_addrs[0]);
if (0 < munmap_len)
munmap((caddr_t)(csa->db_addrs[0]), (size_t)(munmap_len));
}
# endif
/* Detach our shared memory while still under lock so reference counts will be correct for the next process to run down
* this region. In the process also get the remove_shm status from node_local before detaching.
* If cnl->donotflush_dbjnl is TRUE, it means we can safely remove shared memory without compromising data
* integrity as a reissue of recover will restore the database to a consistent state.
*/
remove_shm = !vermismatch && (cnl->remove_shm || cnl->donotflush_dbjnl);
/* We are done with online rollback on this region. Indicate to other processes by setting the onln_rlbk_pid to 0.
* Do it before releasing crit (t_end relies on this ordering when accessing cnl->onln_rlbk_pid).
*/
if (jgbl.onlnrlbk)
cnl->onln_rlbk_pid = 0;
rel_crit(reg); /* Since we are about to detach from the shared memory, release crit and reset onln_rlbk_pid */
/* If we had skipped flushing journal and database buffers due to a concurrent online rollback, increment the counter
* indicating that in the shared memory so that online rollback can report the # of such processes when it shuts down.
* The same thing is done for both FTOK and access control semaphores when there are too many MUMPS processes.
*/
if (safe_mode) /* indicates flushing was skipped */
{
if (bypassed_access)
cnl->dbrndwn_access_skip++; /* Access semaphore can be bypassed during online rollback */
if (bypassed_ftok)
cnl->dbrndwn_ftok_skip++;
}
if (jgbl.onlnrlbk)
csa->hold_onto_crit = FALSE;
GTM_WHITE_BOX_TEST(WBTEST_HOLD_SEM_BYPASS, cnl->wbox_test_seq_num, 0);
status = shmdt((caddr_t)cnl);
csa->nl = NULL; /* dereferencing nl after detach is not right, so we set it to NULL so that we can test before dereference*/
/* Note that although csa->nl is NULL, we use CSA_ARG(csa) below (not CSA_ARG(NULL)) to be consistent with similar
* usages before csa->nl became NULL. The "is_anticipatory_freeze_needed" function (which is in turn called by the
* CHECK_IF_FREEZE_ON_ERROR_NEEDED macro) does a check of csa->nl before dereferencing shared memory contents so
* we are safe passing "csa".
*/
if (-1 == status)
send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg), ERR_TEXT, 2,
LEN_AND_LIT("Error during shmdt"), errno);
REMOVE_CSA_FROM_CSADDRSLIST(csa); /* remove "csa" from list of open regions (cs_addrs_list) */
reg->open = FALSE;
/* If file is still not in good shape, die here and now before we get rid of our storage */
assertpro(0 == csa->wbuf_dqd);
ipc_deleted = FALSE;
/* If we are the very last user, remove shared storage id and the semaphores */
if (we_are_last_user)
{ /* remove shared storage, only if last writer to rundown did a successful wcs_flu() */
assert(!vermismatch);
if (remove_shm)
{
ipc_deleted = TRUE;
if (0 != shm_rmid(udi->shmid))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove shared memory"));
/* Note that we no longer have a new shared memory. Currently only used/usable for standalone rollback. */
udi->new_shm = FALSE;
/* mupip recover/rollback don't release the semaphore here, but do it later in db_ipcs_reset (invoked from
* mur_close_files())
*/
if (!have_standalone_access)
{
if (0 != sem_rmid(udi->semid))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove semaphore"));
udi->new_sem = FALSE; /* Note that we no longer have a new semaphore */
udi->grabbed_access_sem = FALSE;
udi->counter_acc_incremented = FALSE;
}
} else if (is_src_server || is_updproc)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
{
assert(!have_standalone_access || jgbl.onlnrlbk || safe_mode);
if (!jgbl.onlnrlbk && !have_standalone_access)
{ /* If we were writing, get rid of our writer access count semaphore */
if (!reg->read_only)
{
if (!access_counter_halted)
{
save_errno = do_semop(udi->semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO);
if (0 != save_errno)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown access control semaphore decrement"),
CALLFROM, save_errno);
}
udi->counter_acc_incremented = FALSE;
}
assert(safe_mode || !bypassed_access);
/* Now remove the rundown lock */
if (!bypassed_access)
{
if (0 != (save_errno = do_semop(udi->semid, DB_CONTROL_SEM, -1, SEM_UNDO)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown access control semaphore release"),
CALLFROM, save_errno);
udi->grabbed_access_sem = FALSE;
}
} /* else access control semaphore will be released in db_ipcs_reset */
}
if (!have_standalone_access)
{
if (bypassed_ftok)
{
if (!ftok_counter_halted)
if (0 != (save_errno = do_semop(udi->ftok_semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
} else if (!ftok_sem_release(reg, !ftok_counter_halted, FALSE))
{
FTOK_TRACE(csa, csa->ti->curr_tn, ftok_ops_release, process_id);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
}
udi->grabbed_ftok_sem = FALSE;
udi->counter_ftok_incremented = FALSE;
}
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
if (!ipc_deleted)
{
GET_CUR_TIME(time_str);
if (is_src_server)
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Source server"), REG_LEN_STR(reg));
if (is_updproc)
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Update process"), REG_LEN_STR(reg));
if (mupip_jnl_recover && (!jgbl.onlnrlbk || !we_are_last_user))
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
send_msg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
}
}
REVERT;
return EXIT_NRM;
}
/*
* -----------------------------------------------
* gvn2gds()
* Converts a global variable name (GVN) into its internal database repesentation (GDS)
*
* Arguments:
* gvn - Pointer to Source Name string mval. Must be in GVN form.
* buf - Pointer to a buffer large enough to fit the whole GDS of the passed in GVN.
* col - Collation number.
* Return:
* unsigned char - Pointer to the end of the GDS written to gvkey.
* -----------------------------------------------
*/
unsigned char *gvn2gds(mval *gvn, gv_key *gvkey, int act)
{
boolean_t est_first_pass, retn;
collseq *csp;
gd_region tmpreg;
gv_namehead temp_gv_target;
unsigned char *key, *key_top, *key_start;
int subscript, i, contains_env;
int *start, *stop;
gv_name_and_subscripts start_buff, stop_buff;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
/* determine which buffer to use */
DETERMINE_BUFFER(gvn, start_buff, stop_buff, start, stop);
if (0 != act)
{
csp = ready_collseq(act);
if (NULL == csp)
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(3) ERR_COLLATIONUNDEF, 1, act);
} else
csp = NULL; /* Do not issue COLLATIONUNDEF for 0 collation */
retn = TRUE;
assert(MV_IS_STRING(gvn));
key_start = &gvkey->base[0];
key = key_start;
gvkey->prev = 0;
gvkey->top = DBKEYSIZE(MAX_KEY_SZ);
key_top = key_start + gvkey->top;
/* We will parse all of the components up front. */
if (!parse_gv_name_and_subscripts(gvn, &subscript, start, stop, &contains_env))
NOCANONICNAME_ERROR(gvn);
if (stop[contains_env] - start[contains_env] > gvkey->top)
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_GVSUBOFLOW);
memcpy(key, gvn->str.addr + start[contains_env], stop[contains_env] - start[contains_env]);
key += stop[contains_env] - start[contains_env];
*key++ = KEY_DELIMITER;
gvkey->end = key - key_start;
/* Temporarily repoint global variables "gv_cur_region", "gv_target" and "transform".
* They are needed by mval2subsc for the following
* "transform", "gv_target->nct", "gv_target->collseq" and "gv_cur_region->std_null_coll"
* Note that transform is usually ON, specifying that collation transformation is "enabled",
* and is only shut off for minor periods when something is being critically formatted (like
* we're doing here). Note that mval2subsc could issue an rts_error, so we establish a
* condition handler to restore the above.
*/
save_transform = TREF(transform);
assert(save_transform);
TREF(transform) = TRUE;
reset_gv_target = gv_target;
gv_target = &temp_gv_target;
memset(gv_target, 0, SIZEOF(gv_namehead));
gv_target->collseq = csp;
memset(&tmpreg, 0, SIZEOF(gd_region));
/* Assign "gv_cur_region" only after tmpreg has been fully initialized or timer interrupts can look at inconsistent copy */
save_gv_cur_region = gv_cur_region;
gv_cur_region = &tmpreg;
gv_cur_region->std_null_coll = TRUE;
ESTABLISH_NORET(gvn2gds_ch, est_first_pass);
/* we know the number of subscripts, so we convert them all */
for (i = 1 + contains_env; i <= contains_env + subscript; ++i)
{
if (!(retn = convert_key_to_db(gvn, start[i], stop[i], gvkey, &key)))
break;
}
REVERT;
RESTORE_GBL_VARS_BEFORE_FUN_RETURN;
if (!retn || !CAN_APPEND_HIDDEN_SUBS(gvkey))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_GVSUBOFLOW);
*key++ = KEY_DELIMITER; /* add double terminating null byte */
assert(key <= key_top);
return key;
}
void gvcst_spr_kill(void)
{
boolean_t spr_tpwrapped;
boolean_t est_first_pass;
int reg_index;
gd_binding *start_map, *end_map, *map;
gd_region *reg, *gd_reg_start;
gd_addr *addr;
gv_namehead *start_map_gvt;
gvnh_reg_t *gvnh_reg;
trans_num gd_targ_tn, *tn_array;
# ifdef DEBUG
int save_dollar_tlevel;
# endif
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
start_map = TREF(gd_targ_map); /* set up by op_gvname/op_gvnaked/op_gvextnam done just before invoking op_gvkill */
start_map_gvt = gv_target; /* save gv_target corresponding to start_map so we can restore at end */
/* Find out if the next (in terms of $order) key maps to same map as currkey. If so, no spanning activity needed */
GVKEY_INCREMENT_ORDER(gv_currkey);
end_map = gv_srch_map_linear(start_map, (char *)&gv_currkey->base[0], gv_currkey->end - 1);
BACK_OFF_ONE_MAP_ENTRY_IF_EDGECASE(gv_currkey->base, gv_currkey->end - 1, end_map);
GVKEY_UNDO_INCREMENT_ORDER(gv_currkey);
if (start_map == end_map)
{
assert(gv_target == start_map_gvt);
if (IS_OK_TO_INVOKE_GVCST_KILL(start_map_gvt))
gvcst_kill(TRUE);
return;
}
/* Do any initialization that is independent of retries BEFORE the op_tstart */
addr = TREF(gd_targ_addr);
assert(NULL != addr);
gd_reg_start = &addr->regions[0];
tn_array = TREF(gd_targ_reg_array);
gvnh_reg = TREF(gd_targ_gvnh_reg);
assert(NULL != gvnh_reg);
assert(NULL != gvnh_reg->gvspan);
/* Now that we know the keyrange maps to more than one region, go through each of them and do the kill.
* Since multiple regions are potentially involved, need a TP fence.
*/
DEBUG_ONLY(save_dollar_tlevel = dollar_tlevel);
if (!dollar_tlevel)
{
spr_tpwrapped = TRUE;
op_tstart((IMPLICIT_TSTART), TRUE, &literal_batch, 0);
ESTABLISH_NORET(gvcst_spr_kill_ch, est_first_pass);
GVCST_ROOT_SEARCH_AND_PREP(est_first_pass);
} else
spr_tpwrapped = FALSE;
assert(gv_cur_region == start_map->reg.addr);
DBG_CHECK_GVTARGET_GVCURRKEY_IN_SYNC(CHECK_CSA_TRUE);
/* Do any initialization that is dependent on retries AFTER the op_tstart */
map = start_map;
INCREMENT_GD_TARG_TN(gd_targ_tn); /* takes a copy of incremented "TREF(gd_targ_tn)" into local variable "gd_targ_tn" */
/* Verify that initializations that happened before op_tstart are still unchanged */
assert(addr == TREF(gd_targ_addr));
assert(tn_array == TREF(gd_targ_reg_array));
assert(gvnh_reg == TREF(gd_targ_gvnh_reg));
for ( ; map <= end_map; map++)
{
reg = map->reg.addr;
GET_REG_INDEX(addr, gd_reg_start, reg, reg_index); /* sets "reg_index" */
assert((map != start_map) || (tn_array[reg_index] != gd_targ_tn));
assert(TREF(gd_targ_reg_array_size) > reg_index);
if (tn_array[reg_index] == gd_targ_tn)
continue;
if (map != start_map)
GV_BIND_SUBSREG(addr, reg, gvnh_reg); /* sets gv_target/gv_cur_region/cs_addrs */
assert(reg->open);
if (IS_OK_TO_INVOKE_GVCST_KILL(gv_target))
gvcst_kill(TRUE);
tn_array[reg_index] = gd_targ_tn;
}
if (gv_target != start_map_gvt)
{ /* Restore gv_cur_region/gv_target etc. */
gv_target = start_map_gvt;
gv_cur_region = start_map->reg.addr;
change_reg();
}
DBG_CHECK_GVTARGET_GVCURRKEY_IN_SYNC(CHECK_CSA_TRUE);
if (spr_tpwrapped)
{
op_tcommit();
REVERT; /* remove our condition handler */
}
assert(save_dollar_tlevel == dollar_tlevel);
return;
}
