node_pointer* complete_new_btree(couchfile_modify_result* mr, couchstore_error_t *errcode)
{
*errcode = flush_mr(mr);
if(*errcode != COUCHSTORE_SUCCESS) {
return NULL;
}
couchfile_modify_result* targ_mr = make_modres(mr->arena, mr->rq);
if (!targ_mr) {
*errcode = COUCHSTORE_ERROR_ALLOC_FAIL;
return NULL;
}
targ_mr->modified = 1;
targ_mr->node_type = KP_NODE;
*errcode = mr_move_pointers(mr, targ_mr);
if(*errcode != COUCHSTORE_SUCCESS) {
return NULL;
}
node_pointer* ret_ptr;
if(targ_mr->count > 1 || targ_mr->pointers != targ_mr->pointers_end) {
ret_ptr = finish_root(mr->rq, targ_mr, errcode);
} else {
ret_ptr = targ_mr->values_end->pointer;
}
return copy_node_pointer(ret_ptr);
}
couchfile_pointer_info* modify_btree(couchfile_modify_request *rq,
couchfile_pointer_info *root, int *errcode)
{
couchfile_pointer_info* ret_ptr = root;
couchfile_modify_result* root_result = make_modres(rq);
root_result->node_type = KP_NODE;
*errcode = modify_node(rq, root, 0, rq->num_actions, root_result);
if(*errcode < 0)
{
free_modres(root_result);
return NULL;
}
if(root_result->values_end->value.pointer == root)
{
//If we got the root pointer back, remove it from the list
//so we don't try to free it.
root_result->values_end->value.mem = NULL;
}
if(!root_result->modified)
{
free_modres(root_result);
}
else
{
if(root_result->count > 1 || root_result->pointers != root_result->pointers_end)
{
//The root was split
//Write it to disk and return the pointer to it.
ret_ptr = finish_root(rq, root_result, errcode);
if(*errcode < 0)
{
ret_ptr = NULL;
}
}
else
{
ret_ptr = root_result->values_end->value.pointer;
root_result->values_end->value.mem = NULL;
free_modres(root_result);
}
}
return ret_ptr;
}
couchfile_pointer_info* finish_root(couchfile_modify_request* rq,
couchfile_modify_result *root_result, int *errcode)
{
couchfile_pointer_info *ret_ptr = NULL;
couchfile_modify_result *collector = make_modres(rq);
collector->modified = 1;
collector->node_type = KP_NODE;
flush_mr(root_result);
while(1)
{
if(root_result->pointers_end == root_result->pointers->next)
{
//The root result split into exactly one kp_node.
//Return the pointer to it.
ret_ptr = root_result->pointers_end->value.pointer;
root_result->pointers_end->value.mem = NULL;
break;
}
else
{
//The root result split into more than one kp_node.
//Move the pointer list to the value list and write out the new node.
*errcode = mr_move_pointers(root_result, collector);
if(*errcode < 0)
{
goto cleanup;
}
*errcode = flush_mr(collector);
if(*errcode < 0)
{
goto cleanup;
}
//Swap root_result and collector mr's.
couchfile_modify_result *tmp = root_result;
root_result = collector;
collector = tmp;
}
}
cleanup:
free_modres(root_result);
free_modres(collector);
return ret_ptr;
}
node_pointer *modify_btree(couchfile_modify_request *rq,
node_pointer *root,
couchstore_error_t *errcode)
{
arena* a = new_arena(0);
node_pointer *ret_ptr = root;
couchfile_modify_result *root_result = make_modres(a, rq);
if (!root_result) {
delete_arena(a);
*errcode = COUCHSTORE_ERROR_ALLOC_FAIL;
return root;
}
root_result->node_type = KP_NODE;
*errcode = modify_node(rq, root, 0, rq->num_actions, root_result);
if (*errcode < 0) {
delete_arena(a);
return NULL;
}
if (root_result->values_end->pointer == root) {
//If we got the root pointer back, remove it from the list
//so we don't try to free it.
root_result->values_end->pointer = NULL;
}
if (root_result->modified) {
if (root_result->count > 1 || root_result->pointers != root_result->pointers_end) {
//The root was split
//Write it to disk and return the pointer to it.
ret_ptr = finish_root(rq, root_result, errcode);
if (*errcode < 0) {
ret_ptr = NULL;
}
} else {
ret_ptr = root_result->values_end->pointer;
}
}
if (ret_ptr != root) {
ret_ptr = copy_node_pointer(ret_ptr);
}
delete_arena(a);
return ret_ptr;
}
static node_pointer *finish_root(couchfile_modify_request *rq,
couchfile_modify_result *root_result,
couchstore_error_t *errcode)
{
node_pointer *ret_ptr = NULL;
couchfile_modify_result *collector = make_modres(root_result->arena, rq);
if (!collector) {
*errcode = COUCHSTORE_ERROR_ALLOC_FAIL;
return NULL;
}
collector->modified = 1;
collector->node_type = KP_NODE;
flush_mr(root_result);
while (1) {
if (root_result->pointers_end == root_result->pointers->next) {
//The root result split into exactly one kp_node.
//Return the pointer to it.
ret_ptr = root_result->pointers_end->pointer;
break;
} else {
//The root result split into more than one kp_node.
//Move the pointer list to the value list and write out the new node.
*errcode = mr_move_pointers(root_result, collector);
if (*errcode < 0) {
goto cleanup;
}
*errcode = flush_mr(collector);
if (*errcode < 0) {
goto cleanup;
}
//Swap root_result and collector mr's.
couchfile_modify_result *tmp = root_result;
root_result = collector;
collector = tmp;
}
}
cleanup:
return ret_ptr;
}
couchfile_modify_result *new_btree_modres(arena *a, arena *transient_arena, Db* db, compare_info* cmp, reduce_fn reduce,
reduce_fn rereduce)
{
couchfile_modify_request* rq = arena_alloc(a, sizeof(couchfile_modify_request));
rq->cmp = *cmp;
rq->db = db;
rq->num_actions = 0;
rq->fetch_callback = NULL;
rq->reduce = reduce;
rq->rereduce = rereduce;
rq->compacting = 1;
couchfile_modify_result* mr = make_modres(a, rq);
if (!mr)
return NULL;
mr->arena_transient = transient_arena;
mr->modified = 1;
mr->node_type = KV_NODE;
return mr;
}
static couchstore_error_t modify_node(couchfile_modify_request *rq,
node_pointer *nptr,
int start, int end,
couchfile_modify_result *dst)
{
char *nodebuf = NULL; // FYI, nodebuf is a malloced block, not in the arena
int bufpos = 1;
int nodebuflen = 0;
int errcode = 0;
couchfile_modify_result *local_result = NULL;
if (start == end) {
return 0;
}
if (nptr) {
if ((nodebuflen = pread_compressed(rq->db, nptr->pointer, (char **) &nodebuf)) < 0) {
error_pass(COUCHSTORE_ERROR_READ);
}
}
local_result = make_modres(dst->arena, rq);
error_unless(local_result, COUCHSTORE_ERROR_ALLOC_FAIL);
if (nptr == NULL || nodebuf[0] == 1) { //KV Node
local_result->node_type = KV_NODE;
while (bufpos < nodebuflen) {
sized_buf cmp_key, val_buf;
bufpos += read_kv(nodebuf + bufpos, &cmp_key, &val_buf);
int advance = 0;
while (!advance && start < end) {
advance = 1;
int cmp_val = rq->cmp.compare(&cmp_key, rq->actions[start].key);
if (cmp_val < 0) { //Key less than action key
mr_push_item(&cmp_key, &val_buf, local_result);
} else if (cmp_val > 0) { //Key greater than action key
switch (rq->actions[start].type) {
case ACTION_INSERT:
local_result->modified = 1;
mr_push_item(rq->actions[start].key, rq->actions[start].value.data, local_result);
break;
case ACTION_REMOVE:
local_result->modified = 1;
break;
case ACTION_FETCH:
if (rq->fetch_callback) {
//not found
(*rq->fetch_callback)(rq, rq->actions[start].key, NULL, rq->actions[start].value.arg);
}
}
start++;
//Do next action on same item in the node, as our action was
//not >= it.
advance = 0;
} else if (cmp_val == 0) { //Node key is equal to action key
switch (rq->actions[start].type) {
case ACTION_INSERT:
local_result->modified = 1;
mr_push_item(rq->actions[start].key, rq->actions[start].value.data, local_result);
break;
case ACTION_REMOVE:
local_result->modified = 1;
break;
case ACTION_FETCH:
if (rq->fetch_callback) {
(*rq->fetch_callback)(rq, rq->actions[start].key, &val_buf, rq->actions[start].value.arg);
}
//Do next action on same item in the node, as our action was a fetch
//and there may be an equivalent insert or remove
//following.
advance = 0;
}
start++;
}
}
if (start == end && !advance) {
//If we've exhausted actions then just keep this key
mr_push_item(&cmp_key, &val_buf, local_result);
}
}
while (start < end) {
//We're at the end of a leaf node.
switch (rq->actions[start].type) {
case ACTION_INSERT:
local_result->modified = 1;
mr_push_item(rq->actions[start].key, rq->actions[start].value.data, local_result);
break;
case ACTION_REMOVE:
local_result->modified = 1;
break;
case ACTION_FETCH:
if (rq->fetch_callback) {
//not found
(*rq->fetch_callback)(rq, rq->actions[start].key, NULL, rq->actions[start].value.arg);
}
break;
}
start++;
}
} else if (nodebuf[0] == 0) { //KP Node
local_result->node_type = KP_NODE;
while (bufpos < nodebuflen && start < end) {
sized_buf cmp_key, val_buf;
bufpos += read_kv(nodebuf + bufpos, &cmp_key, &val_buf);
int cmp_val = rq->cmp.compare(&cmp_key, rq->actions[start].key);
if (bufpos == nodebuflen) {
//We're at the last item in the kpnode, must apply all our
//actions here.
node_pointer *desc = read_pointer(dst->arena, &cmp_key, val_buf.buf);
if (!desc) {
errcode = COUCHSTORE_ERROR_ALLOC_FAIL;
goto cleanup;
}
errcode = modify_node(rq, desc, start, end, local_result);
if (errcode != COUCHSTORE_SUCCESS) {
goto cleanup;
}
break;
}
if (cmp_val < 0) {
//Key in node item less than action item and not at end
//position, so just add it and continue.
node_pointer *add = read_pointer(dst->arena, &cmp_key, val_buf.buf);
if (!add) {
errcode = COUCHSTORE_ERROR_ALLOC_FAIL;
goto cleanup;
}
errcode = mr_push_pointerinfo(add, local_result);
if (errcode != COUCHSTORE_SUCCESS) {
goto cleanup;
}
} else if (cmp_val >= 0) {
//Found a key in the node greater than the one in the current
//action. Descend into the pointed node with as many actions as
//are less than the key here.
int range_end = start;
while (range_end < end &&
rq->cmp.compare(rq->actions[range_end].key, &cmp_key) <= 0) {
range_end++;
}
node_pointer *desc = read_pointer(dst->arena, &cmp_key, val_buf.buf);
if (!desc) {
errcode = COUCHSTORE_ERROR_ALLOC_FAIL;
goto cleanup;
}
errcode = modify_node(rq, desc, start, range_end, local_result);
start = range_end;
if (errcode != COUCHSTORE_SUCCESS) {
goto cleanup;
}
}
}
while (bufpos < nodebuflen) {
sized_buf cmp_key, val_buf;
bufpos += read_kv(nodebuf + bufpos, &cmp_key, &val_buf);
node_pointer *add = read_pointer(dst->arena, &cmp_key, val_buf.buf);
if (!add) {
errcode = COUCHSTORE_ERROR_ALLOC_FAIL;
goto cleanup;
}
errcode = mr_push_pointerinfo(add, local_result);
if (errcode != COUCHSTORE_SUCCESS) {
goto cleanup;
}
}
} else {
errcode = COUCHSTORE_ERROR_CORRUPT;
goto cleanup;
}
//If we've done modifications, write out the last leaf node.
error_pass(flush_mr(local_result));
if (!local_result->modified && nptr != NULL) {
//If we didn't do anything, give back the pointer to the original
mr_push_pointerinfo(nptr, dst);
} else {
//Otherwise, give back the pointers to the nodes we've created.
dst->modified = 1;
error_pass(mr_move_pointers(local_result, dst));
}
cleanup:
if (nodebuf) {
free(nodebuf);
}
return errcode;
}
int modify_node(couchfile_modify_request *rq, couchfile_pointer_info *nptr,
int start, int end, couchfile_modify_result *dst)
{
eterm_buf current_node;
int curnode_pos = 0;
int read_size = 0;
int list_start_pos = 0;
int node_len = 0;
int node_bound = 0;
int errcode = 0;
int kpos = 0;
char node_type[MAXATOMLEN + 1];
node_type[0] = 0;
DBG("Enter modify_node. %d - %d\r\n", start, end);
if(start == end)
{
return 0;
}
if(nptr == NULL)
{
current_node = empty_root;
}
else
{
if((read_size = pread_bin(rq->fd, nptr->pointer, ¤t_node.buf)) < 0)
{
return ERROR_READ_FILE;
}
current_node.size = read_size;
DBG("... read node from %d\r\n", nptr->pointer);
curnode_pos++; //Skip over 131.
}
couchfile_modify_result *local_result = make_modres(rq);
ei_decode_tuple_header(current_node.buf, &curnode_pos, NULL);
if(ei_decode_atom(current_node.buf, &curnode_pos, node_type) < 0)
{
errcode = ERROR_PARSE;
goto cleanup;
}
list_start_pos = curnode_pos;
if(ei_decode_list_header(current_node.buf, &curnode_pos, &node_len) < 0)
{
errcode = ERROR_PARSE;
goto cleanup;
}
if(strcmp("kv_node", node_type) == 0)
{
local_result->node_type = KV_NODE;
while(start < end)
{
DBG("act on kvnode item\r\n");
if(node_bound >= node_len)
{ //We're at the end of a leaf node.
DBG(" ... exec action at end!\r\n");
switch(rq->actions[start].type)
{
case ACTION_INSERT:
local_result->modified = 1;
mr_push_action(&rq->actions[start], local_result);
break;
case ACTION_REMOVE:
local_result->modified = 1;
break;
case ACTION_FETCH:
if(rq->fetch_callback)
{
//not found
(*rq->fetch_callback)(rq, rq->actions[start].key, NULL);
}
break;
}
start++;
}
else
{
kpos = find_first_gteq(current_node.buf, list_start_pos,
rq->actions[start].cmp_key,
&rq->cmp, node_bound);
if(kpos < 0)
{
errcode = ERROR_PARSE;
goto cleanup;
}
//Add items from node_bound up to but not including the current
mr_push_kv_range(current_node.buf, list_start_pos, node_bound,
rq->cmp.list_pos, local_result);
if(rq->cmp.last_cmp_val > 0) // Node key > action key
{
DBG(" Inserting action before\r\n");
switch(rq->actions[start].type)
{
case ACTION_INSERT:
local_result->modified = 1;
mr_push_action(&rq->actions[start], local_result);
break;
case ACTION_REMOVE:
local_result->modified = 1;
break;
case ACTION_FETCH:
if(rq->fetch_callback)
{
//not found
(*rq->fetch_callback)(rq, rq->actions[start].key, NULL);
}
break;
}
start++;
node_bound = rq->cmp.list_pos;
}
else if(rq->cmp.last_cmp_val < 0) // Node key < action key
{
DBG(" -- Continue with this action\r\n");
node_bound = rq->cmp.list_pos + 1;
mr_push_kv_range(current_node.buf, list_start_pos, node_bound - 1,
node_bound, local_result);
}
else //Node key == action key
{
DBG(" Replacing value with action\r\n");
switch(rq->actions[start].type)
{
case ACTION_INSERT:
local_result->modified = 1;
mr_push_action(&rq->actions[start], local_result);
node_bound = rq->cmp.list_pos + 1;
break;
case ACTION_REMOVE:
local_result->modified = 1;
node_bound = rq->cmp.list_pos + 1;
break;
case ACTION_FETCH:
if(rq->fetch_callback)
{
eterm_buf cb_tmp;
int cb_vpos = kpos;
ei_decode_tuple_header(current_node.buf, &cb_vpos, NULL);
ei_skip_term(current_node.buf, &cb_vpos);
cb_tmp.buf = current_node.buf + cb_vpos;
cb_tmp.size = cb_vpos;
ei_skip_term(current_node.buf, &cb_vpos);
cb_tmp.size = cb_vpos - cb_tmp.size;
(*rq->fetch_callback)(rq, rq->actions[start].key, &cb_tmp);
}
node_bound = rq->cmp.list_pos;
break;
}
start++;
}
}
}
//Push any items past the end of what we dealt with onto result.
if(node_bound < node_len)
{
mr_push_kv_range(current_node.buf, list_start_pos, node_bound,
node_len, local_result);
}
}
else if(strcmp("kp_node", node_type) == 0)
{
local_result->node_type = KP_NODE;
while(start < end)
{
kpos = find_first_gteq(current_node.buf, list_start_pos,
rq->actions[start].cmp_key,
&rq->cmp, node_bound);
if(kpos < 0)
{
errcode = ERROR_PARSE;
goto cleanup;
}
if(rq->cmp.list_pos == (node_len - 1)) //got last item in kp_node
{
//Push all items in node onto mr
mr_push_kp_range(current_node.buf, list_start_pos, node_bound,
rq->cmp.list_pos, local_result);
DBG(" ...descending into final item of kpnode\r\n");
couchfile_pointer_info *desc = read_pointer(current_node.buf, kpos);
errcode = modify_node(rq, desc, start, end, local_result);
if(local_result->values_end->value.pointer != desc)
{
free(desc);
}
if(errcode < 0)
{
goto cleanup;
}
node_bound = node_len;
break;
}
else
{
//Get all actions with key <= the key of the current item in the
//kp_node
//Push items in node up to but not including current onto mr
mr_push_kp_range(current_node.buf, list_start_pos, node_bound,
rq->cmp.list_pos - 1, local_result);
int range_end = start;
while(range_end < end &&
((*rq->cmp.compare)(rq->actions[range_end].cmp_key, rq->cmp.last_cmp_key) <= 0))
{
range_end++;
}
DBG(" ...descending into item %d of kpnode\r\n", rq->cmp.list_pos);
node_bound = rq->cmp.list_pos + 1;
couchfile_pointer_info *desc = read_pointer(current_node.buf, kpos);
errcode = modify_node(rq, desc, start, range_end, local_result);
if(local_result->values_end->value.pointer != desc)
{
free(desc);
}
if(errcode < 0)
{
goto cleanup;
}
start = range_end;
}
}
DBG(".. Finished kp node, up to %d\r\n", node_bound);
if(node_bound < node_len)
{
//Processed all the actions but haven't exhausted this kpnode.
//push the rest of it onto the mr.
mr_push_kp_range(current_node.buf, list_start_pos, node_bound, node_len,
local_result);
}
}
else
{
errcode = ERROR_PARSE;
goto cleanup;
}
//If we've done modifications, write out the last leaf node.
errcode = flush_mr(local_result);
if(errcode == 0)
{
if(!local_result->modified && nptr != NULL)
{
//If we didn't do anything, give back the pointer to the original
mr_push_pointerinfo(nptr, dst);
}
else
{
//Otherwise, give back the pointers to the nodes we've created.
dst->modified = 1;
errcode = mr_move_pointers(local_result, dst);
}
}
cleanup:
free_modres(local_result);
if(current_node.buf != empty_root.buf)
{
free(current_node.buf);
}
return errcode;
}