/* Waits for a concurrently running read (from disk into a global buffer) to complete.
*
* Returns TRUE if read completes within timeout of approx. 1 minute.
* Returns FALSE otherwise.
*
* Similar logic is also present in t_qread and wcs_recover but they are different enough that
* they have not been folded into this routine yet.
*/
boolean_t wcs_read_in_progress_wait(cache_rec_ptr_t cr, wbtest_code_t wbox_test_code)
{
uint4 lcnt, r_epid;
int4 n;
for (lcnt = 1; -1 != cr->read_in_progress; lcnt++)
{
if (-1 > cr->read_in_progress)
{ /* outside of design; clear to known state */
INTERLOCK_INIT(cr);
assert(0 == cr->r_epid);
cr->r_epid = 0;
break;
}
wcs_sleep(lcnt);
GTM_WHITE_BOX_TEST(wbox_test_code, lcnt, (2 * BUF_OWNER_STUCK));
if (BUF_OWNER_STUCK < lcnt)
{ /* sick of waiting */
/* Since cr->r_epid can be changing concurrently, take a local copy before using it below,
* particularly before calling is_proc_alive as we dont want to call it with a 0 r_epid.
*/
r_epid = cr->r_epid;
if (0 != r_epid)
{
if (FALSE == is_proc_alive(r_epid, cr->image_count))
{ /* process gone; release its lock */
RELEASE_BUFF_READ_LOCK(cr);
} else
{
assert(gtm_white_box_test_case_enabled);
return FALSE;
}
} else
{ /* process stopped before could set r_epid */
RELEASE_BUFF_READ_LOCK(cr);
if (-1 > cr->read_in_progress)
{ /* process released since if (cr->r_epid); rectify semaphore */
LOCK_BUFF_FOR_READ(cr, n);
}
}
} /* sick of waiting */
}
return 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 */
{
uint4 status, duint4, blocking_pid;
cache_rec_ptr_t cr;
bt_rec_ptr_t bt;
bool clustered, was_crit;
int dummy, lcnt, ocnt;
cw_set_element *cse;
off_chain chain1;
register sgmnt_addrs *csa;
register sgmnt_data_ptr_t csd;
int4 dummy_errno;
boolean_t already_built, is_mm, reset_first_tp_srch_status, set_wc_blocked;
error_def(ERR_DBFILERR);
error_def(ERR_BUFOWNERSTUCK);
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);
assert((t_tries < CDB_STAGNATE) || csa->now_crit);
if (0 < 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
{
first_tp_srch_status = (srch_blk_status *)lookup_hashtab_ent(sgm_info_ptr->blks_in_use,
(void *)blk, &duint4);
ASSERT_IS_WITHIN_TP_HIST_ARRAY_BOUNDS(first_tp_srch_status, sgm_info_ptr);
cse = first_tp_srch_status ? first_tp_srch_status->ptr : NULL;
}
assert(!cse || !cse->high_tlevel);
if (cse)
{ /* transaction has modified the sought after block */
assert(gds_t_writemap != cse->mode);
if (FALSE == cse->done)
{ /* out of date, so make it current */
already_built = (NULL != cse->new_buff);
gvcst_blk_build(cse, (uchar_ptr_t)cse->new_buff, 0);
assert(cse->blk_target);
if (!already_built && !chain1.flag)
{
assert(first_tp_srch_status && (is_mm || first_tp_srch_status->cr)
&& first_tp_srch_status->buffaddr);
if (first_tp_srch_status->tn <=
((blk_hdr_ptr_t)(first_tp_srch_status->buffaddr))->tn)
{
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,
((blk_hdr_ptr_t)(first_tp_srch_status->buffaddr))->tn,
((blk_hdr_ptr_t)(first_tp_srch_status->buffaddr))->levl);
return (sm_uc_ptr_t)NULL;
}
if ((!is_mm) && (first_tp_srch_status->cycle != first_tp_srch_status->cr->cycle
|| first_tp_srch_status->blk_num != first_tp_srch_status->cr->blk))
{
assert(CDB_STAGNATE > t_tries);
rdfail_detail = cdb_sc_lostcr; /* should this be something else */
return (sm_uc_ptr_t)NULL;
}
if (certify_all_blocks &&
FALSE == cert_blk(gv_cur_region, blk, (blk_hdr_ptr_t)cse->new_buff,
cse->blk_target->root))
GTMASSERT;
}
cse->done = TRUE;
}
*cycle = CYCLE_PVT_COPY;
*cr_out = 0;
return (sm_uc_ptr_t)cse->new_buff;
}
assert(!chain1.flag);
} else
first_tp_srch_status =
(srch_blk_status *)lookup_hashtab_ent(sgm_info_ptr->blks_in_use, (void *)blk, &duint4);
ASSERT_IS_WITHIN_TP_HIST_ARRAY_BOUNDS(first_tp_srch_status, sgm_info_ptr);
if (!is_mm && first_tp_srch_status)
{
assert(first_tp_srch_status->cr && !first_tp_srch_status->ptr);
if (first_tp_srch_status->cycle == first_tp_srch_status->cr->cycle)
{
*cycle = first_tp_srch_status->cycle;
*cr_out = first_tp_srch_status->cr;
first_tp_srch_status->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. Allow for
* recycling. But update the first_tp_srch_status (for this blk) in the si->first_tp_hist
* array to reflect the new buffer, cycle and cache-record. Since we know those only at the end of
* t_qread, set a variable here that will enable the updation before returning from t_qread().
*/
reset_first_tp_srch_status = TRUE;
}
}
}
if ((blk >= csa->ti->total_blks) || (blk < 0))
{ /* requested block out of range; could occur because of a concurrency conflict */
if ((&FILE_INFO(gv_cur_region)->s_addrs != csa) || (csd != cs_data))
GTMASSERT;
assert(FALSE == 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));
}
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;
was_crit = csa->now_crit;
ocnt = 0;
set_wc_blocked = FALSE; /* to indicate whether csd->wc_blocked was set to TRUE by us */
do
{
if (NULL == (cr = db_csh_get(blk)))
{ /* not in memory */
if (clustered && (NULL != (bt = bt_get(blk))) && (FALSE == bt->flushing))
bt = NULL;
if (FALSE == csa->now_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 */
}
if (csd->flush_trigger <= csa->nl->wcs_active_lvl && FALSE == gv_cur_region->read_only)
JNL_ENSURE_OPEN_WCS_WTSTART(csa, gv_cur_region, 0, dummy_errno);
/* a macro that dclast's wcs_wtstart() and checks for errors etc. */
grab_crit(gv_cur_region);
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)
{
SET_TRACEABLE_VAR(cs_data->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 = csa->ti->curr_tn;
if (FALSE == was_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);
INCR_DB_CSH_COUNTER(csa, n_dsk_reads, 1);
if (SS_NORMAL != (status = dsk_read(blk, GDS_REL2ABS(cr->buffaddr))))
{
RELEASE_BUFF_READ_LOCK(cr);
assert(was_crit == csa->now_crit);
if (FUTURE_READ == status)
{ /* in cluster, block can be in the "future" with respect to the local history */
assert(TRUE == clustered);
assert(FALSE == csa->now_crit);
rdfail_detail = cdb_sc_future_read; /* t_retry forces the history up to date */
return (sm_uc_ptr_t)NULL;
}
rts_error(VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), status);
}
assert(0 <= cr->read_in_progress);
assert(0 == cr->dirty);
cr->r_epid = 0;
RELEASE_BUFF_READ_LOCK(cr);
assert(-1 <= cr->read_in_progress);
*cr_out = cr;
assert(was_crit == csa->now_crit);
if (reset_first_tp_srch_status)
{ /* keep the parantheses for the if (although single line) since the following is a macro */
RESET_FIRST_TP_SRCH_STATUS(first_tp_srch_status, cr, *cycle);
}
return (sm_uc_ptr_t)GDS_REL2ABS(cr->buffaddr);
} else if ((FALSE == was_crit) && (BAD_LUCK_ABOUNDS > ocnt))
{
assert(TRUE == csa->now_crit);
assert(csa->nl->in_crit == process_id);
rel_crit(gv_cur_region);
}
}
if (CR_NOTVALID == (sm_long_t)cr)
{
SET_TRACEABLE_VAR(cs_data->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wc_blocked_t_qread_db_csh_get_invalid_blk);
set_wc_blocked = TRUE;
break;
}
for (lcnt = 1; ; lcnt++)
{
if (0 > cr->read_in_progress)
{ /* it's not being read */
if (clustered && (0 == cr->bt_index) &&
(cr->tn < ((th_rec *)((uchar_ptr_t)csa->th_base + csa->th_base->tnque.fl))->tn))
{ /* can't rely on the buffer */
cr->cycle++; /* increment cycle whenever blk number changes (tp_hist depends on this) */
cr->blk = CR_BLKEMPTY;
break;
}
*cycle = cr->cycle;
*cr_out = cr;
VMS_ONLY(
/* If we were doing the db_csh_get() above (in t_qread itself) and located the cache-record
* which, before coming here and taking a copy of cr->cycle a few lines above, was made an
* older twin by another process in bg_update (note this can happen in VMS only) which has
* already incremented the cycle, we will end up having a copy of the old cache-record with
* its incremented cycle number and hence will succeed in tp_hist validation if we return
* this <cr,cycle> combination although we don't want to since this "cr" is not current for
* the given block as of now. Note that the "indexmod" optimization in tp_tend() relies on
* an accurate intermediate validation by tp_hist() which in turn relies on the <cr,cycle>
* value returned by t_qread() to be accurate for a given blk at the current point in time.
* We detect the older-twin case by the following check. Note that here we depend on the
* the fact that bg_update() sets cr->bt_index to 0 before incrementing cr->cycle.
* Given that order, cr->bt_index can be guaranteed to be 0 if we read the incremented cycle
*/
if (cr->twin && (0 == cr->bt_index))
break;
)
if (cr->blk != blk)
break;
if (was_crit != csa->now_crit)
rel_crit(gv_cur_region);
assert(was_crit == csa->now_crit);
if (reset_first_tp_srch_status)
{ /* keep the parantheses for the if (although single line) since the following is a macro */
RESET_FIRST_TP_SRCH_STATUS(first_tp_srch_status, cr, *cycle);
}
/* Note that at this point we expect t_qread() to return a <cr,cycle> combination that
* corresponds to "blk" passed in. It is crucial to get an accurate value for both the fields
* since tp_hist() relies on this for its intermediate validation.
*/
return (sm_uc_ptr_t)GDS_ANY_REL2ABS(csa, cr->buffaddr);
}
if (blk != cr->blk)
break;
if (lcnt >= BUF_OWNER_STUCK && (0 == (lcnt % BUF_OWNER_STUCK)))
{
if (FALSE == csa->now_crit)
grab_crit(gv_cur_region);
if (cr->read_in_progress < -1)
{ /* outside of design; clear to known state */
BG_TRACE_PRO(t_qread_out_of_design);
INTERLOCK_INIT(cr);
assert(0 == cr->r_epid);
cr->r_epid = 0;
} else if (cr->read_in_progress >= 0)
{
BG_TRACE_PRO(t_qread_buf_owner_stuck);
if (0 != (blocking_pid = cr->r_epid))
{
if (FALSE == is_proc_alive(blocking_pid, cr->image_count))
{ /* process gone: release that process's lock */
assert(0 == cr->bt_index);
if (cr->bt_index)
{
SET_TRACEABLE_VAR(csd->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wc_blocked_t_qread_bad_bt_index1);
set_wc_blocked = TRUE;
break;
}
cr->cycle++; /* increment cycle for blk number changes (for tp_hist) */
cr->blk = CR_BLKEMPTY;
RELEASE_BUFF_READ_LOCK(cr);
} else
{
rel_crit(gv_cur_region);
send_msg(VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region));
send_msg(VARLSTCNT(9) ERR_BUFOWNERSTUCK, 7, process_id, blocking_pid,
cr->blk, cr->blk, (lcnt / BUF_OWNER_STUCK),
cr->read_in_progress, cr->rip_latch.latch_pid);
if ((4 * BUF_OWNER_STUCK) <= lcnt)
GTMASSERT;
/* Kickstart the process taking a long time in case it was suspended */
UNIX_ONLY(continue_proc(blocking_pid));
}
} else
{ /* process stopped before could set r_epid */
assert(0 == cr->bt_index);
if (cr->bt_index)
{
SET_TRACEABLE_VAR(csd->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wc_blocked_t_qread_bad_bt_index2);
set_wc_blocked = TRUE;
break;
}
cr->cycle++; /* increment cycle for blk number changes (for tp_hist) */
cr->blk = CR_BLKEMPTY;
RELEASE_BUFF_READ_LOCK(cr);
if (cr->read_in_progress < -1) /* race: process released since if r_epid */
LOCK_BUFF_FOR_READ(cr, dummy);
}
}
if (was_crit != csa->now_crit)
rel_crit(gv_cur_region);
} else
wcs_sleep(lcnt);
}
if (set_wc_blocked) /* cannot use csd->wc_blocked here as we might not necessarily have crit */
break;
ocnt++;
if (BAD_LUCK_ABOUNDS <= ocnt)
{
if (BAD_LUCK_ABOUNDS < ocnt || csa->now_crit)
{
rel_crit(gv_cur_region);
GTMASSERT;
}
if (FALSE == csa->now_crit)
grab_crit(gv_cur_region);
}
} while (TRUE);