/*
* Internal function which adds digests to the defrag_list
* Mallocs the nodes of defrag_list
* Returns :
* -1 : Error
* number of digests found : success
*
*/
static long
build_defrag_list_from_nbtr(as_namespace *ns, ai_obj *acol, bt *nbtr, ulong nofst, ulong *limit, uint64_t * tot_found, cf_ll *gc_list)
{
int error = -1;
btEntry *nbe;
// STEP 1: go thru a portion of the nbtr and find to-be-deleted-PKs
// TODO: a range query may be smarter then using the Xth Iterator
btSIter *nbi = (nofst ? btGetFullXthIter(nbtr, nofst, 1, NULL, 0) :
btGetFullRangeIter(nbtr, 1, NULL));
if (!nbi) {
return error;
}
long found = 0;
long processed = 0;
while ((nbe = btRangeNext(nbi, 1))) {
ai_obj *akey = nbe->key;
int ret = as_sindex_can_defrag_record(ns, (cf_digest *) (&akey->y));
if (ret == AS_SINDEX_GC_SKIP_ITERATION) {
*limit = 0;
break;
} else if (ret == AS_SINDEX_GC_OK) {
bool create = (cf_ll_size(gc_list) == 0) ? true : false;
objs_to_defrag_arr *dt;
if (!create) {
cf_ll_element * ele = cf_ll_get_tail(gc_list);
dt = ((ll_sindex_gc_element*)ele)->objs_to_defrag;
if (dt->num == SINDEX_GC_NUM_OBJS_PER_ARR) {
create = true;
}
}
if (create) {
dt = as_sindex_gc_get_defrag_arr();
if (!dt) {
*tot_found += found;
return -1;
}
ll_sindex_gc_element * node;
node = cf_malloc(sizeof(ll_sindex_gc_element));
node->objs_to_defrag = dt;
cf_ll_append(gc_list, (cf_ll_element *)node);
}
cloneDigestFromai_obj(&(dt->acol_digs[dt->num].dig), akey);
ai_objClone(&(dt->acol_digs[dt->num].acol), acol);
dt->num += 1;
found++;
}
processed++;
(*limit)--;
if (*limit == 0) break;
}
btReleaseRangeIterator(nbi);
*tot_found += found;
return processed;
}
static long
build_defrag_list_from_arr(as_namespace *ns, ai_obj *acol, ai_arr *arr, long nofst, long *limit, uint64_t * tot_found, cf_ll *gc_list)
{
long found = 0;
long processed = 0;
uint64_t validation_time_ns = 0;
for (int i = nofst; i < arr->used; i++) {
SET_TIME_FOR_SINDEX_GC_HIST(validation_time_ns);
int ret = as_sindex_can_defrag_record(ns, (cf_digest *) &arr->data[i * CF_DIGEST_KEY_SZ]);
SINDEX_GC_HIST_INSERT_DATA_POINT(sindex_gc_validate_obj_hist, validation_time_ns);
validation_time_ns = 0;
if (ret == AS_SINDEX_GC_SKIP_ITERATION) {
*limit = 0;
break;
} else if (ret == AS_SINDEX_GC_OK) {
bool create = (cf_ll_size(gc_list) == 0) ? true : false;
objs_to_defrag_arr *dt;
if (!create) {
cf_ll_element * ele = cf_ll_get_tail(gc_list);
dt = ((ll_sindex_gc_element*)ele)->objs_to_defrag;
if (dt->num == SINDEX_GC_NUM_OBJS_PER_ARR) {
create = true;
}
}
if (create) {
dt = as_sindex_gc_get_defrag_arr();
if (!dt) {
*tot_found += found;
return -1;
}
ll_sindex_gc_element * node;
node = cf_malloc(sizeof(ll_sindex_gc_element));
node->objs_to_defrag = dt;
cf_ll_append(gc_list, (cf_ll_element *)node);
}
memcpy(&(dt->acol_digs[dt->num].dig), (cf_digest *) &arr->data[i * CF_DIGEST_KEY_SZ], CF_DIGEST_KEY_SZ);
ai_objClone(&(dt->acol_digs[dt->num].acol), acol);
dt->num += 1;
found++;
}
processed++;
(*limit)--;
if (*limit == 0) {
break;
}
}
*tot_found += found;
return processed;
}
/*
* Return 0 in case of success
* -1 in case of failure
*/
static int
add_recs_from_nbtr(as_sindex_metadata *imd, ai_obj *ikey, bt *nbtr, as_sindex_qctx *qctx, bool fullrng)
{
int ret = 0;
ai_obj sfk, efk;
init_ai_obj(&sfk);
init_ai_obj(&efk);
btSIter *nbi;
btEntry *nbe;
btSIter stack_nbi;
if (fullrng) {
nbi = btSetFullRangeIter(&stack_nbi, nbtr, 1, NULL);
} else { // search from LAST batches end-point
init_ai_objFromDigest(&sfk, &qctx->bdig);
assignMaxKey(nbtr, &efk);
nbi = btSetRangeIter(&stack_nbi, nbtr, &sfk, &efk, 1);
}
if (nbi) {
while ((nbe = btRangeNext(nbi, 1))) {
ai_obj *akey = nbe->key;
// FIRST can be REPEAT (last batch)
if (!fullrng && ai_objEQ(&sfk, akey)) {
continue;
}
if (btree_addsinglerec(imd, ikey, (cf_digest *)&akey->y, qctx->recl, &qctx->n_bdigs,
qctx->can_partition_query, qctx->partitions_pre_reserved)) {
ret = -1;
break;
}
if (qctx->n_bdigs == qctx->bsize) {
if (ikey) {
ai_objClone(qctx->bkey, ikey);
}
cloneDigestFromai_obj(&qctx->bdig, akey);
break;
}
}
btReleaseRangeIterator(nbi);
} else {
cf_warning(AS_QUERY, "Could not find nbtr iterator.. skipping !!");
}
return ret;
}
static int
add_recs_from_arr(as_sindex_metadata *imd, ai_obj *ikey, ai_arr *arr, as_sindex_qctx *qctx)
{
bool ret = 0;
for (int i = 0; i < arr->used; i++) {
if (btree_addsinglerec(imd, (cf_digest *)&arr->data[i * CF_DIGEST_KEY_SZ], qctx->recl, &qctx->n_bdigs)) {
ret = -1;
break;
}
// do not break on hitting batch limit, if the tree converts to
// bt from arr, there is no way to know which digest were already
// returned when attempting subsequent batch. Return the entire
// thing.
}
// mark nbtr as finished and copy the offset
qctx->nbtr_done = true;
if (ikey) {
ai_objClone(qctx->bkey, ikey);
}
return ret;
}
/*
* Aerospike Index interface to build a defrag_list.
*
* Returns :
* AS_SINDEX_DONE ---> The current pimd has been scanned completely for defragging
* AS_SINDEX_CONTINUE ---> Current pimd sill may have some candidate digest to be defragged
* AS_SINDEX_ERR ---> Error. Abort this pimd.
*
* Notes : Caller has the responsibility to free the iterators.
* Requires a proper offset value from the caller.
*/
int
ai_btree_build_defrag_list(as_sindex_metadata *imd, as_sindex_pmetadata *pimd, ai_obj *icol,
long *nofst, long limit, uint64_t * tot_processed, uint64_t * tot_found, cf_ll *gc_list)
{
int ret = AS_SINDEX_ERR;
if (!pimd || !imd) {
return ret;
}
as_namespace *ns = imd->si->ns;
if (!ns) {
ns = as_namespace_get_byname((char *)imd->ns_name);
}
char *iname = get_iname_from_imd(imd);
if (!iname) {
ret = AS_SINDEX_ERR_NO_MEMORY;
return ret;
}
if (!pimd || !pimd->ibtr || !pimd->ibtr->numkeys) {
goto END;
}
//Entry is range query, FROM previous icol TO maxKey(ibtr)
if (icol->empty) {
assignMinKey(pimd->ibtr, icol); // init first call
}
ai_obj iH;
assignMaxKey(pimd->ibtr, &iH);
btEntry *be = NULL;
btSIter *bi = btGetRangeIter(pimd->ibtr, icol, &iH, 1);
if (!bi) {
goto END;
}
while ( true ) {
be = btRangeNext(bi, 1);
if (!be) {
ret = AS_SINDEX_DONE;
break;
}
ai_obj *acol = be->key;
ai_nbtr *anbtr = be->val;
long processed = 0;
if (!anbtr) {
break;
}
if (anbtr->is_btree) {
processed = build_defrag_list_from_nbtr(ns, acol, anbtr->u.nbtr, *nofst, &limit, tot_found, gc_list);
} else {
processed = build_defrag_list_from_arr(ns, acol, anbtr->u.arr, *nofst, &limit, tot_found, gc_list);
}
if (processed < 0) { // error .. abort everything.
cf_detail(AS_SINDEX, "build_defrag_list returns an error. Aborting defrag on current pimd");
ret = AS_SINDEX_ERR;
break;
}
*tot_processed += processed;
// This tree may have some more digest to defrag
if (limit == 0) {
*nofst = *nofst + processed;
ai_objClone(icol, acol);
cf_detail(AS_SINDEX, "Current pimd may need more iteration of defragging.");
ret = AS_SINDEX_CONTINUE;
break;
}
// We have finished this tree. Yet we have not reached our limit to defrag.
// Goes to next iteration
*nofst = 0;
ai_objClone(icol, acol);
};
btReleaseRangeIterator(bi);
END:
cf_free(iname);
return ret;
}