#include "skiplist.h"
#define UNDETERMINED_VERSION 0
+#define ABORTED_VERSION TAG_VALUE(0, TAG1)
#define INITIAL_WRITES_SIZE 4
-typedef enum { UPDATE_TYPE_PUT, UPDATE_TYPE_DELETE } update_type_t;
-
-typedef struct update_rec update_rec_t;
+typedef struct update_rec update_t;
struct update_rec {
- update_type_t type;
- uint64_t value;
uint64_t version;
- update_rec_t *prev; // a previous update
+ map_val_t value;
+ map_val_t next; // an earlier update
};
typedef struct write_rec {
- void *key;
- update_rec_t *rec;
+ map_key_t key;
+ update_t *rec;
} write_rec_t;
struct txn {
uint64_t wv;
map_t *map;
write_rec_t *writes;
- uint32_t writes_size;
- uint32_t writes_count;
- uint32_t writes_scan;
- txn_type_e type;
+ size_t writes_size;
+ size_t writes_count;
+ size_t writes_scan;
txn_state_e state;
};
-static uint64_t version_ = 1;
-
static txn_state_e txn_validate (txn_t *txn);
-static map_t *active_ = NULL;
+static uint64_t version_ = 1;
+
+static skiplist_t *active_ = NULL;
void txn_init (void) {
- active_ = map_alloc(&sl_map_impl, NULL);
+ active_ = sl_alloc(NULL);
}
-// Validate the updates for <key>. Validation fails for a key we have written to if there is a
-// write committed newer than our read version.
-static txn_state_e tm_validate_key (txn_t *txn, void *key) {
+// Validate the updates for <key>. Validation fails if there is a write-write conflict. That is if after our
+// read version another transaction committed a change to an entry we are also trying to change.
+//
+// If we encounter a potential conflict with a transaction that is in the process of validating, we help it
+// complete validating. It must be finished before we can decide to rollback or commit.
+//
+static txn_state_e validate_key (txn_t *txn, map_key_t key) {
+ assert(txn->state != TXN_RUNNING);
- update_rec_t *update = (update_rec_t *) map_get(txn->map, key);
- for (; update != NULL; update = update->prev) {
- uint64_t writer_version = update->version;
- if (writer_version <= txn->rv)
+ map_val_t val = map_get(txn->map, key);
+ update_t *update = NULL;
+ for (; val != DOES_NOT_EXIST; val = update->next) {
+
+ // If the update or its version is not tagged it means the update is committed.
+ //
+ // We can stop at the first committed record we find that is at least as old as our read version. All
+ // the other committed records following it will be older. And all the uncommitted records following it
+ // will eventually conflict with it and abort.
+ if (!IS_TAGGED(val, TAG2))
return TXN_VALIDATED;
-
- // If the version is tagged, it means it is a pointer to a transaction in progress.
- if (IS_TAGGED(writer_version)) {
-
- // Skip aborted transactions.
- if (EXPECT_FALSE(writer_version == TAG_VALUE(0)))
- continue;
-
- // Skip our own updates.
- txn_t *writer = (txn_t *)STRIP_TAG(writer_version);
- if (writer == txn)
- continue;
-
- writer_version = writer->wv;
- if (writer_version <= txn->rv && writer_version != UNDETERMINED_VERSION)
- return TXN_VALIDATED;
-
- txn_state_e writer_state = writer->state;
- if (EXPECT_FALSE(writer_state == TXN_ABORTED))
+ update = (update_t *)STRIP_TAG(val, TAG2);
+ if (!IS_TAGGED(update->version, TAG1))
+ return (update->version <= txn->rv) ? TXN_VALIDATED : TXN_ABORTED;
+
+ // If the update's version is tagged then either the update was aborted or the the version number is
+ // actually a pointer to a running transaction's txn_t.
+
+ // Skip aborted transactions.
+ if (EXPECT_FALSE(update->version == ABORTED_VERSION))
+ continue;
+
+ // The update's transaction is still in progress. Access its txn_t.
+ txn_t *writer = (txn_t *)STRIP_TAG(update->version, TAG1);
+ if (writer == txn)
+ continue; // Skip our own updates.
+ txn_state_e writer_state = writer->state;
+
+ // Any running transaction will only be able to acquire a wv greater than ours. A transaction changes its
+ // state to validating before aquiring a wv. We can ignore an unvalidated transaction if its version is
+ // greater than ours. See next comment below for why.
+ if (writer_state == TXN_RUNNING)
+ continue;
+
+ // If <writer> has a later version than us we can safely ignore its updates. It will not commit until
+ // we have completed validation (in order to remain non-blocking it will help us validate if necessary).
+ // This protocol ensures a deterministic resolution to every conflict and avoids infinite ping-ponging
+ // between validating two conflicting transactions.
+ if (writer_state == TXN_VALIDATING) {
+ if (writer->wv > txn->wv)
continue;
-
- // Help validate <writer> if it is a committing transaction that might cause us to
- // abort. However, if the <writer> has a later version than us we can safely ignore its
- // updates. This protocol ensures a deterministic resolution to every conflict, and
- // avoids infinite ping-ponging between validating two conflicting transactions.
- if (writer_state == TXN_VALIDATING && (writer_version < txn->wv ||
- writer_version == UNDETERMINED_VERSION)) {
- writer_state = txn_validate(writer);
- }
-
- if (writer_state == TXN_VALIDATED)
- return TXN_ABORTED;
+ // Help <writer> commit. We need to know if <writer> aborts or commits before we can decide what to
+ // do. But we don't want to block, so we assist.
+ writer_state = txn_validate(writer);
}
- return TXN_ABORTED;
+ // Skip updates from aborted transactions.
+ if (writer_state == TXN_ABORTED)
+ continue;
+
+ assert(writer_state == TXN_VALIDATED);
+ return (writer->wv <= txn->rv) ? TXN_VALIDATED : TXN_ABORTED;
}
return TXN_VALIDATED;
}
static txn_state_e txn_validate (txn_t *txn) {
+ assert(txn->state != TXN_RUNNING);
int i;
switch (txn->state) {
}
for (i = 0; i < txn->writes_count; ++i) {
- txn_state_e s = tm_validate_key(txn, txn->writes[i].key);
+ txn_state_e s = validate_key(txn, txn->writes[i].key);
if (s == TXN_ABORTED) {
txn->state = TXN_ABORTED;
break;
}
+ assert(s == TXN_VALIDATED);
}
if (txn->state == TXN_VALIDATING) {
txn->state = TXN_VALIDATED;
return txn->state;
}
-static update_rec_t *alloc_update_rec (void) {
- update_rec_t *u = (update_rec_t *)nbd_malloc(sizeof(update_rec_t));
- memset(u, 0, sizeof(update_rec_t));
+static update_t *alloc_update_rec (void) {
+ update_t *u = (update_t *)nbd_malloc(sizeof(update_t));
+ memset(u, 0, sizeof(update_t));
return u;
}
-txn_t *txn_begin (txn_type_e type, map_t *map) {
+txn_t *txn_begin (map_t *map) {
txn_t *txn = (txn_t *)nbd_malloc(sizeof(txn_t));
memset(txn, 0, sizeof(txn_t));
- txn->type = type;
txn->wv = UNDETERMINED_VERSION;
txn->state = TXN_RUNNING;
txn->map = map;
- if (type != TXN_READ_ONLY) {
- txn->writes = nbd_malloc(sizeof(*txn->writes) * INITIAL_WRITES_SIZE);
- txn->writes_size = INITIAL_WRITES_SIZE;
- }
+ txn->writes = nbd_malloc(sizeof(*txn->writes) * INITIAL_WRITES_SIZE);
+ txn->writes_size = INITIAL_WRITES_SIZE;
- // aquire the read version for txn.
+ // acquire the read version for txn. must be careful to avoid a race
do {
txn->rv = version_;
uint64_t temp = 0;
do {
old_count = temp;
- temp = (uint64_t)map_cas(active_, (void *)txn->rv, old_count, old_count + 1);
+ temp = (uint64_t)sl_cas(active_, (map_key_t)txn->rv, old_count, old_count + 1);
} while (temp != old_count);
if (txn->rv == version_)
temp = 1;
do {
old_count = temp;
- temp = map_cas(active_, (void *)txn->rv, old_count, old_count - 1);
+ temp = sl_cas(active_, (map_key_t)txn->rv, old_count, old_count - 1);
} while (temp != old_count);
} while (1);
}
void txn_abort (txn_t *txn) {
+ if (txn->state != TXN_RUNNING)
+ return; // TODO: return some sort of error code
int i;
for (i = 0; i < txn->writes_count; ++i) {
- update_rec_t *update = (update_rec_t *)txn->writes[i].rec;
- update->version = TAG_VALUE(0);
+ update_t *update = (update_t *)txn->writes[i].rec;
+ update->version = ABORTED_VERSION;
}
nbd_defer_free(txn->writes);
}
txn_state_e txn_commit (txn_t *txn) {
+ if (txn->state != TXN_RUNNING)
+ return txn->state; // TODO: return some sort of error code
assert(txn->state == TXN_RUNNING);
txn->state = TXN_VALIDATING;
txn_state_e state = txn_validate(txn);
// Detach <txn> from its updates.
- uint64_t wv = (txn->state == TXN_ABORTED) ? TAG_VALUE(0) : txn->wv;
+ uint64_t wv = (txn->state == TXN_ABORTED) ? ABORTED_VERSION : txn->wv;
int i;
for (i = 0; i < txn->writes_count; ++i) {
- update_rec_t *update = (update_rec_t *)txn->writes[i].rec;
+ update_t *update = (update_t *)txn->writes[i].rec;
update->version = wv;
}
+ // Lower the reference count for <txn>'s read version
+ uint64_t temp = 2;
+ uint64_t old_count;
+ do {
+ old_count = temp;
+ temp = sl_cas(active_, (map_key_t)txn->rv, old_count, old_count - 1);
+ if (temp == 1 && txn->rv != version_) {
+ sl_remove(active_, (map_key_t)txn->rv);
+ break;
+ }
+ } while (old_count != temp);
+
nbd_defer_free(txn->writes);
nbd_defer_free(txn);
}
// Get most recent committed version prior to our read version.
-uint64_t tm_get (txn_t *txn, void *key) {
-
- // Iterate through update records associated with <key> to find the latest committed version.
- // We can use the first matching version. Older updates always come later in the list.
- update_rec_t *update = (update_rec_t *) map_get(txn->map, key);
- for (; update != NULL; update = update->prev) {
- uint64_t writer_version = update->version;
- if (writer_version < txn->rv)
- return update->value;
-
- // If the version is tagged, it means that it is not a version number, but a pointer to an
- // in progress transaction.
- if (IS_TAGGED(update->version)) {
- txn_t *writer = (txn_t *)STRIP_TAG(writer_version);
-
- if (writer == txn)
- return update->type == UPDATE_TYPE_DELETE ? DOES_NOT_EXIST : update->value;
-
- // Skip updates from aborted transactions.
- txn_state_e writer_state = writer->state;
- if (EXPECT_FALSE(writer_state == TXN_ABORTED))
+map_val_t txn_map_get (txn_t *txn, map_key_t key) {
+ if (txn->state != TXN_RUNNING)
+ return ERROR_TXN_NOT_RUNNING;
+
+ // Iterate through the update records to find the latest committed version prior to our read version.
+ map_val_t newest_val = map_get(txn->map, key);
+ map_val_t val = newest_val;
+ update_t *update = NULL;
+ for ( ; ; val = update->next) {
+
+ if (!IS_TAGGED(val, TAG2))
+ return val;
+
+ update = (update_t *)STRIP_TAG(val, TAG2);
+ assert(update != NULL);
+
+ // If the update's version is not tagged it means the update is committed.
+ if (!IS_TAGGED(update->version, TAG1)) {
+ if (update->version <= txn->rv)
+ break; // success
+ continue;
+ }
+
+ // If the update's version is tagged then either the update was aborted or the the version number is
+ // actually a pointer to a running transaction's txn_t.
+
+ // Skip updates from aborted transactions.
+ if (EXPECT_FALSE(update->version == ABORTED_VERSION))
+ continue;
+
+ // The update's transaction is still in progress. Access its txn_t.
+ txn_t *writer = (txn_t *)STRIP_TAG(update->version, TAG1);
+ if (writer == txn) // found our own update
+ break; // success
+
+ txn_state_e writer_state = writer->state;
+ if (writer_state == TXN_RUNNING)
+ continue;
+
+ if (writer_state == TXN_VALIDATING) {
+ if (writer->wv > txn->rv)
continue;
+ writer_state = txn_validate(writer);
+ }
+
+ // Skip updates from aborted transactions.
+ if (writer_state == TXN_ABORTED)
+ continue;
+
+ assert(writer_state == TXN_VALIDATED);
+ if (writer->wv > txn->rv)
+ continue;
+ break; // success
+ }
+
+ map_val_t value = update->value;
+
+ // collect some garbage
+ uint64_t min_active_version = UNDETERMINED_VERSION;
+ update_t *next_update = NULL;
+ if (IS_TAGGED(update->next, TAG2)) {
+ next_update = (update_t *)STRIP_TAG(update->next, TAG2);
+ min_active_version = (uint64_t)sl_min_key(active_);
+ if (next_update->version < min_active_version) {
+ // <next_update> (and all update records following it [execpt if it is aborted]) is old enough that it is
+ // not visible to any active transaction. We can safely free it.
+
+ // Skip over aborted versions to look for more recent updates
+ update_t *temp = next_update;
+ while (temp->version == ABORTED_VERSION) {
+ assert(!IS_TAGGED(temp->version, TAG1));
+ uint64_t next = next_update->next;
+ if (!IS_TAGGED(next, TAG2))
+ break;
- if (writer_state == TXN_VALIDATING) {
- writer_state = txn_validate(writer);
+ temp = (update_t *)STRIP_TAG(next, TAG2);
+ if (temp->version >= min_active_version)
+ return value;
}
- if (writer_state == TXN_VALIDATED && writer->wv <= txn->rv && writer->wv != UNDETERMINED_VERSION)
- return update->type == UPDATE_TYPE_DELETE ? DOES_NOT_EXIST : update->value;
+ // free <next> and all the update records following it
+ temp = next_update;
+ while (1) {
+ uint64_t next = SYNC_SWAP(&temp->next, DOES_NOT_EXIST);
+
+ // if we find ourself in a race just back off and let the other thread take care of it
+ if (next == DOES_NOT_EXIST)
+ return value;
+
+ if (!IS_TAGGED(next, TAG2))
+ break;
+
+ temp = (update_t *)STRIP_TAG(next, TAG2);
+ nbd_free(update);
+ }
}
}
- return DOES_NOT_EXIST;
+
+ // If there is one item left and it is visible by all active transactions we can merge it into the map itself.
+ // There is no need for an update record.
+ if (next_update == NULL && val == newest_val) {
+ if (min_active_version == UNDETERMINED_VERSION) {
+ min_active_version = (uint64_t)sl_min_key(active_);
+ }
+ if (update->version <= min_active_version) {
+ if (map_cas(txn->map, key, TAG_VALUE(val, TAG2), value) == TAG_VALUE(val, TAG2)) {
+ nbd_defer_free(update);
+ }
+ }
+ }
+
+ return value;
}
-void tm_set (txn_t *txn, void *key, uint64_t value) {
+void txn_map_set (txn_t *txn, map_key_t key, map_val_t value) {
+ if (txn->state != TXN_RUNNING)
+ return; // TODO: return some sort of error code
// create a new update record
- update_rec_t *update = alloc_update_rec();
- update->type = UPDATE_TYPE_PUT;
+ update_t *update = alloc_update_rec();
update->value = value;
- update->version = TAG_VALUE((uint64_t)txn);
+ update->version = TAG_VALUE((uint64_t)txn, TAG1);
// push the new update record onto <key>'s update list
- uint64_t update_prev;
+ uint64_t old_update;
do {
- update->prev = (update_rec_t *) map_get(txn->map, key);
- update_prev = (uint64_t)update->prev;
- } while (map_cas(txn->map, key, update_prev, (uint64_t)update) != update_prev);
+ old_update = map_get(txn->map, key);
+ update->next = old_update;
+ } while (map_cas(txn->map, key, old_update, TAG_VALUE((uint64_t)update, TAG2)) != old_update);
// add <key> to the write set for commit-time validation
if (txn->writes_count == txn->writes_size) {
write_rec_t *w = nbd_malloc(sizeof(write_rec_t) * txn->writes_size * 2);
memcpy(w, txn->writes, txn->writes_size * sizeof(write_rec_t));
txn->writes_size *= 2;
+ nbd_free(txn->writes);
+ txn->writes = w;
}
int i = txn->writes_count++;
txn->writes[i].key = key;