// If remote record is better than local record, replace local with remote.
int
as_record_replace_if_better(as_remote_record *rr, bool is_repl_write,
bool skip_sindex, bool do_xdr_write)
{
as_namespace *ns = rr->rsv->ns;
if (! as_storage_has_space(ns)) {
cf_warning(AS_RECORD, "{%s} record replace: drives full", ns->name);
return AS_ERR_OUT_OF_SPACE;
}
CF_ALLOC_SET_NS_ARENA(ns);
as_index_tree *tree = rr->rsv->tree;
as_index_ref r_ref;
int rv = as_record_get_create(tree, rr->keyd, &r_ref, ns);
if (rv < 0) {
return AS_ERR_OUT_OF_SPACE;
}
bool is_create = rv == 1;
as_index *r = r_ref.r;
int result;
conflict_resolution_pol policy = ns->conflict_resolution_policy;
if (is_repl_write) {
bool from_replica;
if ((result = as_partition_check_source(ns, rr->rsv->p, rr->src,
&from_replica)) != AS_OK) {
record_replace_failed(rr, &r_ref, NULL, is_create);
return result;
}
repl_write_init_repl_state(rr, from_replica);
policy = repl_write_conflict_resolution_policy(ns);
}
if (! is_create && record_replace_check(r, ns) < 0) {
record_replace_failed(rr, &r_ref, NULL, is_create);
return AS_ERR_FORBIDDEN;
}
// If local record is better, no-op or fail.
if (! is_create && (result = as_record_resolve_conflict(policy,
r->generation, r->last_update_time, (uint16_t)rr->generation,
rr->last_update_time)) <= 0) {
record_replace_failed(rr, &r_ref, NULL, is_create);
return result == 0 ? AS_ERR_RECORD_EXISTS : AS_ERR_GENERATION;
}
// else - remote winner - apply it.
// If creating record, write set-ID into index.
if (is_create) {
if (rr->set_name && (result = as_index_set_set_w_len(r, ns,
rr->set_name, rr->set_name_len, false)) < 0) {
record_replace_failed(rr, &r_ref, NULL, is_create);
return -result;
}
r->last_update_time = rr->last_update_time;
// Don't write record if it would be truncated.
if (as_truncate_record_is_truncated(r, ns)) {
record_replace_failed(rr, &r_ref, NULL, is_create);
return AS_OK;
}
}
// else - not bothering to check that sets match.
as_storage_rd rd;
if (is_create) {
as_storage_record_create(ns, r, &rd);
}
else {
as_storage_record_open(ns, r, &rd);
}
// TODO - old pickle - remove condition in "six months".
if (rr->is_old_pickle) {
// Prepare to store set name, if there is one.
rd.set_name = rr->set_name;
rd.set_name_len = rr->set_name_len;
}
else {
rd.pickle = rr->pickle;
rd.pickle_sz = rr->pickle_sz;
rd.orig_pickle_sz = as_flat_orig_pickle_size(rr, rd.pickle_sz);
}
// Note - deal with key after reading existing record (if such), in case
// we're dropping the key.
// Split according to configuration to replace local record.
bool is_delete = false;
if (ns->storage_data_in_memory) {
if (ns->single_bin) {
result = record_apply_dim_single_bin(rr, &rd, &is_delete);
}
else {
result = record_apply_dim(rr, &rd, skip_sindex, &is_delete);
}
}
else {
if (ns->single_bin) {
result = record_apply_ssd_single_bin(rr, &rd, &is_delete);
}
else {
result = record_apply_ssd(rr, &rd, skip_sindex, &is_delete);
}
}
if (result != 0) {
record_replace_failed(rr, &r_ref, &rd, is_create);
return result;
}
uint16_t set_id = as_index_get_set_id(r); // save for XDR write
record_replaced(r, rr);
as_storage_record_close(&rd);
as_record_done(&r_ref, ns);
if (do_xdr_write) {
xdr_write_replica(rr, is_delete, set_id);
}
return AS_OK;
}
/*
* 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;
}
int
write_replica(as_partition_reservation* rsv, cf_digest* keyd,
uint8_t* pickled_buf, size_t pickled_sz,
const as_rec_props* p_rec_props, as_generation generation,
uint32_t void_time, uint64_t last_update_time, cf_node master,
uint32_t info, ldt_prole_info* linfo)
{
as_namespace* ns = rsv->ns;
if (! as_storage_has_space(rsv->ns)) {
cf_warning(AS_RW, "{%s} write_replica: drives full", ns->name);
return AS_PROTO_RESULT_FAIL_PARTITION_OUT_OF_SPACE;
}
as_index_tree* tree = rsv->tree;
bool is_subrec = false;
bool is_ldt_parent = false;
if (ns->ldt_enabled) {
if ((info & RW_INFO_LDT_SUBREC) != 0 || (info & RW_INFO_LDT_ESR) != 0) {
tree = rsv->sub_tree;
is_subrec = true;
}
else if ((info & RW_INFO_LDT_PARENTREC) != 0) {
is_ldt_parent = true;
}
}
as_index_ref r_ref;
r_ref.skip_lock = false;
int rv = as_record_get_create(tree, keyd, &r_ref, ns, is_subrec);
if (rv < 0) {
cf_warning_digest(AS_RW, keyd, "{%s} write_replica: fail as_record_get_create() ", ns->name);
return AS_PROTO_RESULT_FAIL_UNKNOWN;
}
as_record* r = r_ref.r;
as_storage_rd rd;
bool is_create = false;
if (rv == 1) {
as_storage_record_create(ns, r, &rd, keyd);
is_create = true;
}
else {
as_storage_record_open(ns, r, &rd, keyd);
}
bool has_sindex = (info & RW_INFO_SINDEX_TOUCHED) != 0;
rd.ignore_record_on_device = ! has_sindex && ! is_ldt_parent;
rd.n_bins = as_bin_get_n_bins(r, &rd);
// TODO - we really need an inline utility for this!
uint16_t newbins = ntohs(*(uint16_t*)pickled_buf);
if (! rd.ns->storage_data_in_memory && ! rd.ns->single_bin &&
newbins > rd.n_bins) {
rd.n_bins = newbins;
}
as_bin stack_bins[rd.ns->storage_data_in_memory ? 0 : rd.n_bins];
rd.bins = as_bin_get_all(r, &rd, stack_bins);
uint32_t stack_particles_sz = rd.ns->storage_data_in_memory ?
0 : as_record_buf_get_stack_particles_sz(pickled_buf);
uint8_t stack_particles[stack_particles_sz + 256];
uint8_t* p_stack_particles = stack_particles;
// + 256 for LDT control bin, to hold version.
if (! ldt_get_prole_version(rsv, keyd, linfo, info, &rd, is_create)) {
if (is_create) {
as_index_delete(tree, keyd);
}
as_storage_record_close(r, &rd);
as_record_done(&r_ref, ns);
return AS_PROTO_RESULT_FAIL_UNKNOWN;
}
uint64_t memory_bytes = 0;
if (! is_create) {
memory_bytes = as_storage_record_get_n_bytes_memory(&rd);
}
as_record_set_properties(&rd, p_rec_props);
if (as_record_unpickle_replace(r, &rd, pickled_buf, pickled_sz,
&p_stack_particles, has_sindex) != 0) {
if (is_create) {
as_index_delete(tree, keyd);
}
as_storage_record_close(r, &rd);
as_record_done(&r_ref, ns);
return AS_PROTO_RESULT_FAIL_UNKNOWN; // TODO - better granularity?
}
r->generation = generation;
r->void_time = void_time;
r->last_update_time = last_update_time;
as_storage_record_adjust_mem_stats(&rd, memory_bytes);
uint64_t version_to_set = 0;
bool set_version = false;
if (is_ldt_parent) {
if (linfo->replication_partition_version_match &&
linfo->ldt_prole_version_set) {
version_to_set = linfo->ldt_prole_version;
set_version = true;
}
else if (! linfo->replication_partition_version_match) {
version_to_set = linfo->ldt_source_version;
set_version = true;
}
}
if (set_version) {
int ldt_rv = as_ldt_parent_storage_set_version(&rd, version_to_set,
p_stack_particles, __FILE__, __LINE__);
if (ldt_rv < 0) {
cf_warning(AS_LDT, "write_replica: LDT parent storage version set failed %d", ldt_rv);
// TODO - roll back.
}
}
bool is_delete = false;
if (! as_bin_inuse_has(&rd)) {
// A master write that deletes a record by deleting (all) bins sends a
// binless pickle that ends up here.
is_delete = true;
as_index_delete(tree, keyd);
}
as_storage_record_write(r, &rd);
as_storage_record_close(r, &rd);
uint16_t set_id = as_index_get_set_id(r);
as_record_done(&r_ref, ns);
// Don't send an XDR delete if it's disallowed.
if (is_delete && ! is_xdr_delete_shipping_enabled()) {
// TODO - should we also not ship if there was no record here before?
return AS_PROTO_RESULT_OK;
}
// Do XDR write if the write is a non-XDR write or forwarding is enabled.
if ((info & RW_INFO_XDR) == 0 ||
is_xdr_forwarding_enabled() || ns->ns_forward_xdr_writes) {
xdr_write(ns, *keyd, generation, master, is_delete, set_id, NULL);
}
return AS_PROTO_RESULT_OK;
}