#include "common.h"
#include "txn.h"
#include "mem.h"
+#include "rcu.h"
#include "skiplist.h"
#define UNDETERMINED_VERSION 0
#define INITIAL_WRITES_SIZE 4
typedef struct update_rec update_t;
+typedef map_key_t version_t;
struct update_rec {
- uint64_t version;
+ version_t version;
map_val_t value;
map_val_t next; // an earlier update
};
} write_rec_t;
struct txn {
- uint64_t rv;
- uint64_t wv;
+ version_t rv;
+ version_t wv;
map_t *map;
write_rec_t *writes;
- uint64_t writes_size;
- uint64_t writes_count;
- uint64_t writes_scan;
+ size_t writes_size;
+ size_t writes_count;
+ size_t writes_scan;
txn_state_e state;
};
static txn_state_e txn_validate (txn_t *txn);
-static uint64_t version_ = 1;
+static version_t version_ = 1;
static skiplist_t *active_ = NULL;
// will eventually conflict with it and abort.
if (!IS_TAGGED(val, TAG2))
return TXN_VALIDATED;
- update = (update_t *)(size_t)STRIP_TAG(val, TAG2);
+ update = (update_t *)STRIP_TAG(val, TAG2);
if (!IS_TAGGED(update->version, TAG1))
return (update->version <= txn->rv) ? TXN_VALIDATED : TXN_ABORTED;
continue;
// The update's transaction is still in progress. Access its txn_t.
- txn_t *writer = (txn_t *)(size_t)STRIP_TAG(update->version, TAG1);
+ 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;
case TXN_VALIDATING:
if (txn->wv == UNDETERMINED_VERSION) {
- uint64_t wv = SYNC_ADD(&version_, 1);
+ version_t wv = SYNC_ADD(&version_, 1);
SYNC_CAS(&txn->wv, UNDETERMINED_VERSION, wv);
}
do {
txn->rv = version_;
- uint64_t old_count;
- uint64_t temp = 0;
+ unsigned old_count;
+ unsigned temp = 0;
do {
old_count = temp;
- temp = (uint64_t)sl_cas(active_, (map_key_t)txn->rv, old_count, old_count + 1);
+ temp = sl_cas(active_, txn->rv, old_count, old_count + 1);
} while (temp != old_count);
if (txn->rv == version_)
update->version = ABORTED_VERSION;
}
- nbd_defer_free(txn->writes);
- nbd_defer_free(txn);
+ rcu_defer_free(txn->writes);
+ rcu_defer_free(txn);
}
txn_state_e txn_commit (txn_t *txn) {
txn_state_e state = txn_validate(txn);
// Detach <txn> from its updates.
- uint64_t wv = (txn->state == TXN_ABORTED) ? ABORTED_VERSION : txn->wv;
+ version_t wv = (txn->state == TXN_ABORTED) ? ABORTED_VERSION : txn->wv;
int i;
for (i = 0; i < txn->writes_count; ++i) {
update_t *update = (update_t *)txn->writes[i].rec;
}
// Lower the reference count for <txn>'s read version
- uint64_t temp = 2;
- uint64_t old_count;
+ unsigned temp = 2;
+ unsigned old_count;
do {
old_count = temp;
temp = sl_cas(active_, (map_key_t)txn->rv, old_count, old_count - 1);
}
} while (old_count != temp);
- nbd_defer_free(txn->writes);
- nbd_defer_free(txn);
+ rcu_defer_free(txn->writes);
+ rcu_defer_free(txn);
return state;
}
// Get most recent committed version prior to our read version.
-uint64_t txn_map_get (txn_t *txn, map_key_t key) {
+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.
- uint64_t newest_val = map_get(txn->map, key);
- uint64_t val = newest_val;
+ 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 *)(size_t)STRIP_TAG(val, TAG2);
+ update = (update_t *)STRIP_TAG(val, TAG2);
assert(update != NULL);
// If the update's version is not tagged it means the update is committed.
continue;
// The update's transaction is still in progress. Access its txn_t.
- txn_t *writer = (txn_t *)(size_t)STRIP_TAG(update->version, TAG1);
+ txn_t *writer = (txn_t *)STRIP_TAG(update->version, TAG1);
if (writer == txn) // found our own update
break; // success
break; // success
}
- uint64_t value = update->value;
+ map_val_t value = update->value;
// collect some garbage
- uint64_t min_active_version = UNDETERMINED_VERSION;
+ version_t min_active_version = UNDETERMINED_VERSION;
update_t *next_update = NULL;
if (IS_TAGGED(update->next, TAG2)) {
- next_update = (update_t *)(size_t)STRIP_TAG(update->next, TAG2);
- min_active_version = (uint64_t)sl_min_key(active_);
+ next_update = (update_t *)STRIP_TAG(update->next, TAG2);
+ min_active_version = (version_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.
update_t *temp = next_update;
while (temp->version == ABORTED_VERSION) {
assert(!IS_TAGGED(temp->version, TAG1));
- uint64_t next = next_update->next;
+ map_val_t next = next_update->next;
if (!IS_TAGGED(next, TAG2))
break;
- temp = (update_t *)(size_t)STRIP_TAG(next, TAG2);
+ temp = (update_t *)STRIP_TAG(next, TAG2);
if (temp->version >= min_active_version)
return 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);
+ map_val_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)
if (!IS_TAGGED(next, TAG2))
break;
- temp = (update_t *)(size_t)STRIP_TAG(next, TAG2);
+ temp = (update_t *)STRIP_TAG(next, TAG2);
nbd_free(update);
}
}
// 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_);
+ min_active_version = (version_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);
+ rcu_defer_free(update);
}
}
}
return value;
}
-void txn_map_set (txn_t *txn, map_key_t 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_t *update = alloc_update_rec();
update->value = value;
- update->version = TAG_VALUE((uint64_t)(size_t)txn, TAG1);
+ update->version = TAG_VALUE((version_t)txn, TAG1);
// push the new update record onto <key>'s update list
- uint64_t old_update;
+ map_val_t old_update;
do {
old_update = map_get(txn->map, key);
update->next = old_update;
- } while (map_cas(txn->map, key, old_update, TAG_VALUE((uint64_t)(size_t)update, TAG2)) != old_update);
+ } while (map_cas(txn->map, key, old_update, TAG_VALUE((map_val_t)update, TAG2)) != old_update);
// add <key> to the write set for commit-time validation
if (txn->writes_count == txn->writes_size) {
projects / nbds / blobdiff