/*
* Function: Close storage record if it open and also set flags
*
* Parameters:
* urec : UDF record
*
* Return value : 0 in case storage was open
* 1 in case storage was not open
*
* Callers:
* udf_record_close
*
* Side effect:
* flag will be reset
* bins will be closed
*/
int
udf_storage_record_close(udf_record *urecord)
{
if (urecord->flag & UDF_RECORD_FLAG_STORAGE_OPEN) {
as_index_ref *r_ref = urecord->r_ref;
as_storage_rd *rd = urecord->rd;
// In case allow update is not set .. the record has been opened for
// the aggregation. Do not do any rec property update.
// Pick info from index and put it in storage record.
size_t rec_props_data_size = as_storage_record_rec_props_size(rd);
uint8_t rec_props_data[rec_props_data_size];
if (urecord->flag & UDF_RECORD_FLAG_ALLOW_UPDATES) {
if (rec_props_data_size > 0) {
cf_detail(AS_LDT, "LDT_INDEXBITS Setting Property at close time parent=%d, esr=%d, sub=%d",
as_ldt_record_is_parent(rd->r),
as_ldt_record_is_esr(rd->r),
as_ldt_record_is_subrec(rd->r));
as_storage_record_set_rec_props(rd, rec_props_data);
}
}
if (!(urecord->flag & UDF_RECORD_FLAG_IS_SUBRECORD)) {
if (as_ldt_record_is_parent(rd->r)) {
cf_detail_digest(AS_LDT, &rd->keyd, "LDT_INDEXBIT Parent @ write: Digest:");
}
} else {
as_ldt_subrec_storage_validate(rd, "Writing");
}
if (r_ref) {
if (urecord->flag & UDF_RECORD_FLAG_HAS_UPDATES) {
as_storage_record_write(r_ref->r, rd);
urecord->flag &= ~UDF_RECORD_FLAG_HAS_UPDATES; // TODO - necessary?
}
as_storage_record_close(r_ref->r, rd);
} else {
// Should never happen.
cf_warning(AS_UDF, "Unexpected Internal Error (null r_ref)");
}
urecord->flag &= ~UDF_RECORD_FLAG_STORAGE_OPEN;
cf_detail_digest(AS_UDF, &urecord->tr->keyd, "Storage Close:: Rec(%p) Flag(%x) Digest:",
urecord, urecord->flag );
return 0;
} else {
return 1;
}
}
/* Internal Function: Does the post processing for the UDF record after the
* UDF execution. Does the following:
* 1. Record is closed
* 2. urecord_op is updated to delete in case there is no bin left in it.
* 3. record->pickled_buf is populated before the record is close in case
* it was write operation
* 4. UDF updates cache is cleared
*
* Returns: Nothing
*
* Parameters: urecord - UDF record to operate on
* urecord_op (out) - Populated with the optype
*/
void
udf_rw_post_processing(udf_record *urecord, udf_optype *urecord_op, uint16_t set_id)
{
as_storage_rd *rd = urecord->rd;
as_transaction *tr = urecord->tr;
as_index_ref *r_ref = urecord->r_ref;
// INIT
urecord->pickled_buf = NULL;
urecord->pickled_sz = 0;
urecord->pickled_void_time = 0;
as_rec_props_clear(&urecord->pickled_rec_props);
bool udf_xdr_ship_op = false;
// TODO: optimize not to allocate buffer if it is single
// node cluster. No remote to send data to
// Check if UDF has updates.
if (urecord->flag & UDF_RECORD_FLAG_HAS_UPDATES) {
// Check if the record is not deleted after an update
if ( urecord->flag & UDF_RECORD_FLAG_OPEN) {
*urecord_op = UDF_OPTYPE_WRITE;
udf_xdr_ship_op = true;
}
else {
// If the record has updates and it is not open,
// and if it pre-existed it's an update followed by a delete.
if ( urecord->flag & UDF_RECORD_FLAG_PREEXISTS) {
*urecord_op = UDF_OPTYPE_DELETE;
udf_xdr_ship_op = true;
}
// If the record did not pre-exist and is updated
// and it is not open, then it is create followed by
// delete essentially no_op.
else {
*urecord_op = UDF_OPTYPE_NONE;
}
}
} else if ((urecord->flag & UDF_RECORD_FLAG_PREEXISTS)
&& !(urecord->flag & UDF_RECORD_FLAG_OPEN)) {
*urecord_op = UDF_OPTYPE_DELETE;
udf_xdr_ship_op = true;
} else {
*urecord_op = UDF_OPTYPE_READ;
}
cf_detail(AS_UDF, "FINISH working with LDT Record %p %p %p %p %d", &urecord,
urecord->tr, urecord->r_ref, urecord->rd,
(urecord->flag & UDF_RECORD_FLAG_STORAGE_OPEN));
// If there exists a record reference but no bin of the record is in use,
// delete the record. remove from the tree. Only LDT_RECORD here not needed
// for LDT_SUBRECORD (only do it if requested by UDF). All the SUBRECORD of
// removed LDT_RECORD will be lazily cleaned up by defrag.
if (!(urecord->flag & UDF_RECORD_FLAG_IS_SUBRECORD)
&& urecord->flag & UDF_RECORD_FLAG_OPEN
&& !as_bin_inuse_has(rd)) {
as_index_delete(tr->rsv.tree, &tr->keyd);
urecord->starting_memory_bytes = 0;
*urecord_op = UDF_OPTYPE_DELETE;
udf_xdr_ship_op = true;
} else if (*urecord_op == UDF_OPTYPE_WRITE) {
cf_detail(AS_UDF, "Committing Changes %"PRIx64" n_bins %d", rd->keyd, as_bin_get_n_bins(r_ref->r, rd));
size_t rec_props_data_size = as_storage_record_rec_props_size(rd);
uint8_t rec_props_data[rec_props_data_size];
if (rec_props_data_size > 0) {
as_storage_record_set_rec_props(rd, rec_props_data);
}
write_local_post_processing(tr, tr->rsv.ns, NULL, &urecord->pickled_buf,
&urecord->pickled_sz, &urecord->pickled_void_time,
&urecord->pickled_rec_props, true/*increment_generation*/,
NULL, r_ref->r, rd, urecord->starting_memory_bytes);
// Now ok to accommodate a new stored key...
if (! as_index_is_flag_set(r_ref->r, AS_INDEX_FLAG_KEY_STORED) && rd->key) {
if (rd->ns->storage_data_in_memory) {
as_record_allocate_key(r_ref->r, rd->key, rd->key_size);
}
as_index_set_flags(r_ref->r, AS_INDEX_FLAG_KEY_STORED);
}
// ... or drop a stored key.
else if (as_index_is_flag_set(r_ref->r, AS_INDEX_FLAG_KEY_STORED) && ! rd->key) {
if (rd->ns->storage_data_in_memory) {
as_record_remove_key(r_ref->r);
}
as_index_clear_flags(r_ref->r, AS_INDEX_FLAG_KEY_STORED);
}
}
// Collect the record information (for XDR) before closing the record
as_generation generation = 0;
if (urecord->flag & UDF_RECORD_FLAG_OPEN) {
generation = r_ref->r->generation;
set_id = as_index_get_set_id(r_ref->r);
}
// Close the record for all the cases
udf_record_close(urecord, false);
// Write to XDR pipe after closing the record, in order to release the record lock as
// early as possible.
if (udf_xdr_ship_op == true) {
if (UDF_OP_IS_WRITE(*urecord_op)) {
cf_detail(AS_UDF, "UDF write shipping for key %" PRIx64, tr->keyd);
xdr_write(tr->rsv.ns, tr->keyd, generation, 0, false, set_id);
} else if (UDF_OP_IS_DELETE(*urecord_op)) {
cf_detail(AS_UDF, "UDF delete shipping for key %" PRIx64, tr->keyd);
xdr_write(tr->rsv.ns, tr->keyd, generation, 0, true, set_id);
}
}
// Replication happens when the main record replicates
if (urecord->particle_data) {
cf_free(urecord->particle_data);
urecord->particle_data = 0;
}
udf_record_cache_free(urecord);
}
/*
* Internal function: udf_aerospike__apply_update_atomic
*
* Parameters:
* rec -- udf_record to be updated
*
* Return Values:
* 0 success
* -1 failure
*
* Description:
* This function applies all the updates atomically. That is,
* if one of the bin update/delete/create fails, the entire function
* will fail. If the nth update fails, all the n-1 updates are rolled
* back to their initial values
*
* Special Notes:
* i. The basic checks of bin name being too long or if there is enough space
* on the disk for the bin values is done before allocating space for any
* of the bins.
*
* ii. If one of the updates to be rolled back is a bin creation,
* udf_aerospike_delbin is called. This will not free up the bin metadata.
* So there will be a small memory mismatch b/w replica (which did not get the
* record at all and hence no memory is accounted) and the master will be seen.
* To avoid such cases, we are doing checks upfront.
*
* Callers:
* udf_aerospike__execute_updates
* In this function, if udf_aerospike__apply_update_atomic fails, the record
* is not committed to the storage. On success, record is closed which commits to
* the storage and reopened for the next set of udf updates.
* The return value from udf_aerospike__apply_update_atomic is passed on to the
* callers of this function.
*/
int
udf_aerospike__apply_update_atomic(udf_record *urecord)
{
int rc = 0;
int failmax = 0;
int new_bins = 0; // How many new bins have to be created in this update
as_storage_rd * rd = urecord->rd;
as_namespace * ns = rd->ns;
bool has_sindex = as_sindex_ns_has_sindex(ns);
bool is_record_dirty = false;
bool is_record_flag_dirty = false;
uint8_t old_index_flags = as_index_get_flags(rd->r);
uint8_t new_index_flags = 0;
// This will iterate over all the updates and apply them to storage.
// The items will remain, and be used as cache values. If an error
// occurred during setbin(), we rollback all the operation which
// is and return failure
cf_detail(AS_UDF, "execute updates: %d updates", urecord->nupdates);
// loop twice to make sure the updates are performed first so in case
// something wrong it can be rolled back. The deletes will go through
// successfully generally.
// In first iteration, just calculate how many new bins need to be created
for(uint32_t i = 0; i < urecord->nupdates; i++ ) {
if ( urecord->updates[i].dirty ) {
char * k = urecord->updates[i].name;
if ( k != NULL ) {
if ( !as_bin_get(rd, k) ) {
new_bins++;
}
}
}
}
// Free bins - total bins not in use in the record
// Delta bins - new bins that need to be created
int inuse_bins = as_bin_inuse_count(rd);
int free_bins = rd->n_bins - inuse_bins;
int delta_bins = new_bins - free_bins;
cf_detail(AS_UDF, "Total bins %d, In use bins %d, Free bins %d , New bins %d, Delta bins %d",
rd->n_bins, as_bin_inuse_count(urecord->rd), free_bins, new_bins, delta_bins);
// Check bin usage limit.
if ((inuse_bins + new_bins > UDF_RECORD_BIN_ULIMIT) ||
(urecord->flag & UDF_RECORD_FLAG_TOO_MANY_BINS)) {
cf_warning(AS_UDF, "bin limit of %d for UDF exceeded: %d bins in use, %d bins free, %s%d new bins needed",
(int)UDF_RECORD_BIN_ULIMIT, inuse_bins, free_bins,
(urecord->flag & UDF_RECORD_FLAG_TOO_MANY_BINS) ? ">" : "", new_bins);
goto Rollback;
}
// Allocate space for all the new bins that need to be created beforehand
if (delta_bins > 0 && rd->ns->storage_data_in_memory && ! rd->ns->single_bin) {
as_bin_allocate_bin_space(urecord->r_ref->r, rd, delta_bins);
}
if (!rd->ns->storage_data_in_memory && !urecord->particle_data) {
// 256 as upper bound on the LDT control bin, we may write version below
// leave it at the end for its use
urecord->particle_data = cf_malloc(rd->ns->storage_write_block_size + 256);
urecord->cur_particle_data = urecord->particle_data;
urecord->end_particle_data = urecord->particle_data + rd->ns->storage_write_block_size;
}
if (has_sindex) {
SINDEX_GRLOCK();
}
// In second iteration apply updates.
for(uint32_t i = 0; i < urecord->nupdates; i++ ) {
urecord->updates[i].oldvalue = NULL;
urecord->updates[i].washidden = false;
if ( urecord->updates[i].dirty && rc == 0) {
char * k = urecord->updates[i].name;
as_val * v = urecord->updates[i].value;
bool h = urecord->updates[i].ishidden;
if ( k != NULL ) {
if ( v == NULL || v->type == AS_NIL ) {
// if the value is NIL, then do a delete
cf_detail(AS_UDF, "execute update: position %d deletes bin %s", i, k);
urecord->updates[i].oldvalue = udf_record_storage_get(urecord, k);
urecord->updates[i].washidden = udf_record_bin_ishidden(urecord, k);
// Only case delete fails if bin is not found that is
// as good as delete. Ignore return code !!
udf_aerospike_delbin(urecord, k);
if (urecord->dirty != NULL) {
xdr_fill_dirty_bins(urecord->dirty);
}
}
else {
// otherwise, it is a set
cf_detail(AS_UDF, "execute update: position %d sets bin %s", i, k);
urecord->updates[i].oldvalue = udf_record_storage_get(urecord, k);
urecord->updates[i].washidden = udf_record_bin_ishidden(urecord, k);
rc = udf_aerospike_setbin(urecord, i, k, v, h);
if (rc) {
if (urecord->updates[i].oldvalue) {
as_val_destroy(urecord->updates[i].oldvalue);
urecord->updates[i].oldvalue = NULL;
}
failmax = i;
goto Rollback;
}
if (urecord->dirty != NULL) {
xdr_add_dirty_bin(ns, urecord->dirty, k, strlen(k));
}
}
}
is_record_dirty = true;
}
}
if (urecord->ldt_rectype_bit_update) {
if (urecord->ldt_rectype_bit_update < 0) {
// ldt_rectype_bit_update is negative in case we want to reset the bits
uint8_t rectype_bits = urecord->ldt_rectype_bit_update * -1;
new_index_flags = old_index_flags & ~rectype_bits;
} else {
new_index_flags = old_index_flags | urecord->ldt_rectype_bit_update;
}
if (new_index_flags != old_index_flags) {
as_index_clear_flags(rd->r, old_index_flags);
as_index_set_flags(rd->r, new_index_flags);
is_record_flag_dirty = true;
cf_detail_digest(AS_RW, &urecord->tr->keyd, "Setting index flags from %d to %d new flag %d", old_index_flags, new_index_flags, as_index_get_flags(rd->r));
}
}
{
// This is _NOT_ for writing to the storage but for simply performing sizing
// calculation. If we know the upper bounds of size of rec_props.. we could
// avoid this work and check with that much correction ...
//
// See
// - udf_rw_post_processing for building rec_props for replication
// - udf_record_close for building rec_props for writing it to storage
size_t rec_props_data_size = as_storage_record_rec_props_size(rd);
uint8_t rec_props_data[rec_props_data_size];
if (rec_props_data_size > 0) {
as_storage_record_set_rec_props(rd, rec_props_data);
}
// Version is set in the end after record size check. Setting version won't change the size of
// the record. And if it were before size check then this setting of version as well needs to
// be backed out.
// TODO: Add backout logic would work till very first create call of LDT end up crossing over
// record boundary
if (rd->ns->ldt_enabled && as_ldt_record_is_parent(rd->r)) {
int rv = as_ldt_parent_storage_set_version(rd, urecord->lrecord->version, urecord->end_particle_data, __FILE__, __LINE__);
if (rv < 0) {
cf_warning(AS_LDT, "udf_aerospike__apply_update_atomic: Internal Error "
" [Failed to set the version on storage rv=%d]... Fail",rv);
goto Rollback;
}
// TODO - if size check below fails, won't write to device -
// different behavior than write_to_device flag - OK?
is_record_dirty = true;
}
if (! as_storage_record_size_and_check(rd)) {
cf_warning(AS_UDF, "record failed storage size check, will not be updated");
failmax = (int)urecord->nupdates;
goto Rollback;
}
if (cf_atomic32_get(rd->ns->stop_writes) == 1) {
cf_warning(AS_UDF, "UDF failed by stop-writes, record will not be updated");
failmax = (int)urecord->nupdates;
goto Rollback;
}
if (! as_storage_has_space(rd->ns)) {
cf_warning(AS_UDF, "drives full, record will not be updated");
failmax = (int)urecord->nupdates;
goto Rollback;
}
if (! is_valid_ttl(rd->ns, urecord->tr->msgp->msg.record_ttl)) {
cf_warning(AS_UDF, "invalid ttl %u", urecord->tr->msgp->msg.record_ttl);
failmax = (int)urecord->nupdates;
goto Rollback;
}
}
if (has_sindex) {
SINDEX_GUNLOCK();
}
// If there were updates do miscellaneous successful commit
// tasks
if (is_record_dirty
|| is_record_flag_dirty
|| (urecord->flag & UDF_RECORD_FLAG_METADATA_UPDATED)) {
urecord->flag |= UDF_RECORD_FLAG_HAS_UPDATES; // will write to storage
}
urecord->ldt_rectype_bit_update = 0;
// Clean up oldvalue cache and reset dirty. All the changes made
// here has made to the particle buffer. Nothing will now be backed out.
for (uint32_t i = 0; i < urecord->nupdates; i++) {
udf_record_bin * bin = &urecord->updates[i];
if (bin->oldvalue != NULL ) {
as_val_destroy(bin->oldvalue);
bin->oldvalue = NULL;
}
bin->dirty = false;
}
return rc;
Rollback:
cf_debug(AS_UDF, "Rollback Called: failmax %d", failmax);
for (int i = 0; i < failmax; i++) {
if (urecord->updates[i].dirty) {
char * k = urecord->updates[i].name;
// Pick the oldvalue for rollback
as_val * v = urecord->updates[i].oldvalue;
bool h = urecord->updates[i].washidden;
if ( k != NULL ) {
if ( v == NULL || v->type == AS_NIL ) {
// if the value is NIL, then do a delete
cf_detail(AS_UDF, "execute rollback: position %d deletes bin %s", i, k);
rc = udf_aerospike_delbin(urecord, k);
}
else {
// otherwise, it is a set
cf_detail(AS_UDF, "execute rollback: position %d sets bin %s", i, k);
rc = udf_aerospike_setbin(urecord, i, k, v, h);
if (rc) {
cf_warning(AS_UDF, "Rollback failed .. not good ... !!");
}
}
}
if (v) {
as_val_destroy(v);
cf_debug(AS_UDF, "ROLLBACK as_val_destroy()");
}
}
}
if (is_record_dirty && urecord->dirty != NULL) {
xdr_clear_dirty_bins(urecord->dirty);
}
if (is_record_flag_dirty) {
as_index_clear_flags(rd->r, new_index_flags);
as_index_set_flags(rd->r, old_index_flags);
is_record_flag_dirty = false;
}
urecord->ldt_rectype_bit_update = 0;
if (has_sindex) {
SINDEX_GUNLOCK();
}
// Reset the flat size in case the stuff is backedout !!! it should not
// fail in the backout code ...
if (! as_storage_record_size_and_check(rd)) {
cf_warning(AS_LDT, "Does not fit even after rollback... it is trouble");
}
// Do not clean up the cache in case of failure
return -1;
}
/* Internal Function: Does the post processing for the UDF record after the
* UDF execution. Does the following:
* 1. Record is closed
* 2. urecord_op is updated to delete in case there is no bin left in it.
* 3. record->pickled_buf is populated before the record is close in case
* it was write operation
* 4. UDF updates cache is cleared
*
* Returns: Nothing
*
* Parameters: urecord - UDF record to operate on
* urecord_op (out) - Populated with the optype
*/
static void
post_processing(udf_record *urecord, udf_optype *urecord_op, uint16_t set_id)
{
as_storage_rd *rd = urecord->rd;
as_transaction *tr = urecord->tr;
as_index_ref *r_ref = urecord->r_ref;
// INIT
urecord->pickled_buf = NULL;
urecord->pickled_sz = 0;
as_rec_props_clear(&urecord->pickled_rec_props);
bool udf_xdr_ship_op = false;
getop(urecord, urecord_op);
if (UDF_OP_IS_DELETE(*urecord_op)
|| UDF_OP_IS_WRITE(*urecord_op)) {
udf_xdr_ship_op = true;
}
cf_detail(AS_UDF, "FINISH working with LDT Record %p %p %p %p %d", &urecord,
urecord->tr, urecord->r_ref, urecord->rd,
(urecord->flag & UDF_RECORD_FLAG_STORAGE_OPEN));
// If there exists a record reference but no bin of the record is in use,
// delete the record. remove from the tree. Only LDT_RECORD here not needed
// for LDT_SUBRECORD (only do it if requested by UDF). All the SUBRECORD of
// removed LDT_RECORD will be lazily cleaned up by defrag.
if (udf_zero_bins_left(urecord)) {
as_transaction *tr = urecord->tr;
as_index_delete(tr->rsv.tree, &tr->keyd);
urecord->starting_memory_bytes = 0;
*urecord_op = UDF_OPTYPE_DELETE;
}
else if (*urecord_op == UDF_OPTYPE_WRITE) {
cf_detail_digest(AS_UDF, &rd->keyd, "Committing Changes n_bins %d", as_bin_get_n_bins(r_ref->r, rd));
size_t rec_props_data_size = as_storage_record_rec_props_size(rd);
uint8_t rec_props_data[rec_props_data_size];
if (rec_props_data_size > 0) {
as_storage_record_set_rec_props(rd, rec_props_data);
}
write_udf_post_processing(tr, rd, &urecord->pickled_buf,
&urecord->pickled_sz, &urecord->pickled_rec_props,
urecord->starting_memory_bytes);
// Now ok to accommodate a new stored key...
if (! as_index_is_flag_set(r_ref->r, AS_INDEX_FLAG_KEY_STORED) && rd->key) {
if (rd->ns->storage_data_in_memory) {
as_record_allocate_key(r_ref->r, rd->key, rd->key_size);
}
as_index_set_flags(r_ref->r, AS_INDEX_FLAG_KEY_STORED);
}
// ... or drop a stored key.
else if (as_index_is_flag_set(r_ref->r, AS_INDEX_FLAG_KEY_STORED) && ! rd->key) {
if (rd->ns->storage_data_in_memory) {
as_record_remove_key(r_ref->r);
}
as_index_clear_flags(r_ref->r, AS_INDEX_FLAG_KEY_STORED);
}
}
// Collect the record information (for XDR) before closing the record
as_generation generation = 0;
if (urecord->flag & UDF_RECORD_FLAG_OPEN) {
generation = r_ref->r->generation;
set_id = as_index_get_set_id(r_ref->r);
}
urecord->op = *urecord_op;
// Close the record for all the cases
udf_record_close(urecord);
// Write to XDR pipe after closing the record, in order to release the record lock as
// early as possible.
if (udf_xdr_ship_op == true) {
if (UDF_OP_IS_WRITE(*urecord_op)) {
cf_detail(AS_UDF, "UDF write shipping for key %" PRIx64, tr->keyd);
xdr_write(tr->rsv.ns, tr->keyd, generation, 0, false, set_id);
} else if (UDF_OP_IS_DELETE(*urecord_op)) {
cf_detail(AS_UDF, "UDF delete shipping for key %" PRIx64, tr->keyd);
xdr_write(tr->rsv.ns, tr->keyd, generation, 0, true, set_id);
}
}
}
