2 * Written by Josh Dybnis and released to the public domain, as explained at
3 * http://creativecommons.org/licenses/publicdomain
10 #define UNDETERMINED_VERSION 0
11 #define ABORTED_VERSION TAG_VALUE(0, TAG1)
12 #define INITIAL_WRITES_SIZE 4
14 typedef struct update_rec update_t;
17 update_t *next; // an earlier update
22 typedef struct write_rec {
33 uint32_t writes_count;
38 static uint64_t version_ = 1;
40 static txn_state_e txn_validate (txn_t *txn);
42 static skiplist_t *active_ = NULL;
44 void txn_init (void) {
45 active_ = sl_alloc(NULL);
48 // Validate the updates for <key>. Validation fails if there is a write-write conflict. That is if after our
49 // read version another transaction committed a change to an entry we are also trying to change.
51 // If we encounter a potential conflict with a transaction that is in the process of validating, we help it
52 // complete validating. It must be finished before we can decide to rollback or commit.
54 static txn_state_e validate_key (txn_t *txn, void *key) {
55 assert(txn->state != TXN_RUNNING);
57 update_t *update = (update_t *) map_get(txn->map, key);
58 for (; update != NULL; update = update->next) {
60 // If the update or its version is not tagged it means the update is committed.
62 // We can stop at the first committed record we find that is at least as old as our read version. All
63 // the other committed records following it will be older. And all the uncommitted records following it
64 // will eventually conflict with it and abort.
65 if (!IS_TAGGED(update, TAG2))
67 update = (update_t *)STRIP_TAG(update, TAG2);
68 if (!IS_TAGGED(update->version, TAG1))
69 return (update->version <= txn->rv) ? TXN_VALIDATED : TXN_ABORTED;
71 // If the update's version is tagged then either the update was aborted or the the version number is
72 // actually a pointer to a running transaction's txn_t.
74 // Skip aborted transactions.
75 if (EXPECT_FALSE(update->version == ABORTED_VERSION))
78 // The update's transaction is still in progress. Access its txn_t.
79 txn_t *writer = (txn_t *)STRIP_TAG(update->version, TAG1);
81 continue; // Skip our own updates.
82 txn_state_e writer_state = writer->state;
84 // Any running transaction will only be able to acquire a wv greater than ours. A transaction changes its
85 // state to validating before aquiring a wv. We can ignore an unvalidated transaction if its version is
86 // greater than ours. See next comment below for why.
87 if (writer_state == TXN_RUNNING)
90 // If <writer> has a later version than us we can safely ignore its updates. It will not commit until
91 // we have completed validation (in order to remain non-blocking it will help us validate if necessary).
92 // This protocol ensures a deterministic resolution to every conflict and avoids infinite ping-ponging
93 // between validating two conflicting transactions.
94 if (writer_state == TXN_VALIDATING) {
95 if (writer->wv > txn->wv)
97 // Help <writer> commit. We need to know if <writer> aborts or commits before we can decide what to
98 // do. But we don't want to block, so we assist.
99 writer_state = txn_validate(writer);
102 // Skip updates from aborted transactions.
103 if (writer_state == TXN_ABORTED)
106 assert(writer_state == TXN_VALIDATED);
107 return (writer->wv <= txn->rv) ? TXN_VALIDATED : TXN_ABORTED;
110 return TXN_VALIDATED;
113 static txn_state_e txn_validate (txn_t *txn) {
114 assert(txn->state != TXN_RUNNING);
116 switch (txn->state) {
119 if (txn->wv == UNDETERMINED_VERSION) {
120 uint64_t wv = SYNC_ADD(&version_, 1);
121 SYNC_CAS(&txn->wv, UNDETERMINED_VERSION, wv);
124 for (i = 0; i < txn->writes_count; ++i) {
125 txn_state_e s = validate_key(txn, txn->writes[i].key);
126 if (s == TXN_ABORTED) {
127 txn->state = TXN_ABORTED;
130 assert(s == TXN_VALIDATED);
132 if (txn->state == TXN_VALIDATING) {
133 txn->state = TXN_VALIDATED;
148 static update_t *alloc_update_rec (void) {
149 update_t *u = (update_t *)nbd_malloc(sizeof(update_t));
150 memset(u, 0, sizeof(update_t));
154 txn_t *txn_begin (map_t *map) {
155 txn_t *txn = (txn_t *)nbd_malloc(sizeof(txn_t));
156 memset(txn, 0, sizeof(txn_t));
157 txn->wv = UNDETERMINED_VERSION;
158 txn->state = TXN_RUNNING;
160 txn->writes = nbd_malloc(sizeof(*txn->writes) * INITIAL_WRITES_SIZE);
161 txn->writes_size = INITIAL_WRITES_SIZE;
163 // acquire the read version for txn. must be careful to avoid a race
171 temp = (uint64_t)sl_cas(active_, (void *)txn->rv, old_count, old_count + 1);
172 } while (temp != old_count);
174 if (txn->rv == version_)
180 temp = sl_cas(active_, (void *)txn->rv, old_count, old_count - 1);
181 } while (temp != old_count);
187 void txn_abort (txn_t *txn) {
188 if (txn->state != TXN_RUNNING)
189 return; // TODO: return some sort of error code
192 for (i = 0; i < txn->writes_count; ++i) {
193 update_t *update = (update_t *)txn->writes[i].rec;
194 update->version = ABORTED_VERSION;
197 nbd_defer_free(txn->writes);
201 txn_state_e txn_commit (txn_t *txn) {
202 if (txn->state != TXN_RUNNING)
203 return txn->state; // TODO: return some sort of error code
205 assert(txn->state == TXN_RUNNING);
206 txn->state = TXN_VALIDATING;
207 txn_state_e state = txn_validate(txn);
209 // Detach <txn> from its updates.
210 uint64_t wv = (txn->state == TXN_ABORTED) ? ABORTED_VERSION : txn->wv;
212 for (i = 0; i < txn->writes_count; ++i) {
213 update_t *update = (update_t *)txn->writes[i].rec;
214 update->version = wv;
217 // Lower the reference count for <txn>'s read version
222 temp = sl_cas(active_, (void *)txn->rv, old_count, old_count - 1);
223 if (temp == 1 && txn->rv != version_) {
224 sl_remove(active_, (void *)txn->rv);
227 } while (old_count != temp);
229 nbd_defer_free(txn->writes);
235 // Get most recent committed version prior to our read version.
236 uint64_t txn_map_get (txn_t *txn, void *key) {
237 if (txn->state != TXN_RUNNING)
238 return ERROR_TXN_NOT_RUNNING;
240 update_t *newest_update = (update_t *) map_get(txn->map, key);
241 if (!IS_TAGGED(newest_update, TAG2))
242 return (uint64_t)newest_update;
244 // Iterate through the update records to find the latest committed version prior to our read version.
246 for (update = newest_update; ; update = update->next) {
248 if (!IS_TAGGED(update, TAG2))
249 return (uint64_t)update;
251 update = (update_t *)STRIP_TAG(update, TAG2);
252 assert(update != NULL);
254 // If the update's version is not tagged it means the update is committed.
255 if (!IS_TAGGED(update->version, TAG1)) {
256 if (update->version <= txn->rv)
261 // If the update's version is tagged then either the update was aborted or the the version number is
262 // actually a pointer to a running transaction's txn_t.
264 // Skip updates from aborted transactions.
265 if (EXPECT_FALSE(update->version == ABORTED_VERSION))
268 // The update's transaction is still in progress. Access its txn_t.
269 txn_t *writer = (txn_t *)STRIP_TAG(update->version, TAG1);
270 if (writer == txn) // found our own update
273 txn_state_e writer_state = writer->state;
274 if (writer_state == TXN_RUNNING)
277 if (writer_state == TXN_VALIDATING) {
278 if (writer->wv > txn->rv)
280 writer_state = txn_validate(writer);
283 // Skip updates from aborted transactions.
284 if (writer_state == TXN_ABORTED)
287 assert(writer_state == TXN_VALIDATED);
288 if (writer->wv > txn->rv)
293 uint64_t value = update->value;
295 // collect some garbage
296 update_t *last = update;
297 update_t *next = update->next;
298 uint64_t min_active = 0;
299 if (IS_TAGGED(next, TAG2)) {
300 next = (update_t *)STRIP_TAG(next, TAG2);
301 min_active = (uint64_t)sl_min_key(active_);
302 if (next->version < min_active) {
304 // Skip over aborted versions to verify the chain of updates is old enough for collection
305 update_t *temp = next;
306 while (temp->version == ABORTED_VERSION) {
307 assert(!IS_TAGGED(temp->version, TAG1));
308 update_t *temp = next->next;
309 if (!IS_TAGGED(temp, TAG2))
311 temp = (update_t *)STRIP_TAG(temp, TAG2);
312 if (temp->version >= min_active)
317 // collect <next> and all the update records following it
319 next = SYNC_SWAP(&update->next, NULL);
321 // if we find ourself in a race just back off and let the other thread take care of it
328 } while (IS_TAGGED(next, TAG2));
332 // If there is one item left and it is visible by all active transactions we can merge it into the map itself.
333 // There is no need for an update record.
334 if (next == NULL && last == (update_t *)STRIP_TAG(newest_update, TAG2)) {
335 if (min_active == UNDETERMINED_VERSION) {
336 min_active = (uint64_t)sl_min_key(active_);
338 if (last->version <= min_active) {
339 if (map_cas(txn->map, key, TAG_VALUE(last, TAG2), value) == TAG_VALUE(last, TAG2)) {
340 nbd_defer_free(last);
348 void txn_map_set (txn_t *txn, void *key, uint64_t value) {
349 if (txn->state != TXN_RUNNING)
350 return; // TODO: return some sort of error code
352 // create a new update record
353 update_t *update = alloc_update_rec();
354 update->value = value;
355 update->version = TAG_VALUE(txn, TAG1);
357 // push the new update record onto <key>'s update list
360 old_update = map_get(txn->map, key);
361 update->next = (update_t *)old_update;
362 } while (map_cas(txn->map, key, old_update, TAG_VALUE(update, TAG2)) != old_update);
364 // add <key> to the write set for commit-time validation
365 if (txn->writes_count == txn->writes_size) {
366 write_rec_t *w = nbd_malloc(sizeof(write_rec_t) * txn->writes_size * 2);
367 memcpy(w, txn->writes, txn->writes_size * sizeof(write_rec_t));
368 txn->writes_size *= 2;
369 nbd_free(txn->writes);
372 int i = txn->writes_count++;
373 txn->writes[i].key = key;
374 txn->writes[i].rec = update;