X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fovsdb-idl.c;h=7556b7f390122ed6b6eab4611572be6025d2d1ad;hb=ec988646afe6aee6a63d6894a3e9b50f715d5941;hp=b264591dac537da204eb6517eb989240510ca1d1;hpb=0f0b102dcf5c067259788e54a0a5352851ed1565;p=sliver-openvswitch.git diff --git a/lib/ovsdb-idl.c b/lib/ovsdb-idl.c index b264591da..7556b7f39 100644 --- a/lib/ovsdb-idl.c +++ b/lib/ovsdb-idl.c @@ -1,4 +1,4 @@ -/* Copyright (c) 2009, 2010, 2011 Nicira Networks. +/* Copyright (c) 2009, 2010, 2011, 2012, 2013 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -17,7 +17,6 @@ #include "ovsdb-idl.h" -#include #include #include #include @@ -70,6 +69,13 @@ struct ovsdb_idl { struct json *monitor_request_id; unsigned int last_monitor_request_seqno; unsigned int change_seqno; + bool verify_write_only; + + /* Database locking. */ + char *lock_name; /* Name of lock we need, NULL if none. */ + bool has_lock; /* Has db server told us we have the lock? */ + bool is_lock_contended; /* Has db server told us we can't get lock? */ + struct json *lock_request_id; /* JSON-RPC ID of in-flight lock request. */ /* Transaction support. */ struct ovsdb_idl_txn *txn; @@ -87,14 +93,14 @@ struct ovsdb_idl_txn { struct ds comment; /* Increments. */ - char *inc_table; - char *inc_column; - struct json *inc_where; + const char *inc_table; + const char *inc_column; + struct uuid inc_row; unsigned int inc_index; int64_t inc_new_value; /* Inserted rows. */ - struct hmap inserted_rows; + struct hmap inserted_rows; /* Contains "struct ovsdb_idl_txn_insert"s. */ }; struct ovsdb_idl_txn_insert { @@ -136,12 +142,23 @@ static void ovsdb_idl_txn_abort_all(struct ovsdb_idl *); static bool ovsdb_idl_txn_process_reply(struct ovsdb_idl *, const struct jsonrpc_msg *msg); +static void ovsdb_idl_send_lock_request(struct ovsdb_idl *); +static void ovsdb_idl_send_unlock_request(struct ovsdb_idl *); +static void ovsdb_idl_parse_lock_reply(struct ovsdb_idl *, + const struct json *); +static void ovsdb_idl_parse_lock_notify(struct ovsdb_idl *, + const struct json *params, + bool new_has_lock); + /* Creates and returns a connection to database 'remote', which should be in a * form acceptable to jsonrpc_session_open(). The connection will maintain an * in-memory replica of the remote database whose schema is described by * 'class'. (Ordinarily 'class' is compiled from an OVSDB schema automatically * by ovsdb-idlc.) * + * Passes 'retry' to jsonrpc_session_open(). See that function for + * documentation. + * * If 'monitor_everything_by_default' is true, then everything in the remote * database will be replicated by default. ovsdb_idl_omit() and * ovsdb_idl_omit_alert() may be used to selectively drop some columns from @@ -153,7 +170,7 @@ static bool ovsdb_idl_txn_process_reply(struct ovsdb_idl *, */ struct ovsdb_idl * ovsdb_idl_create(const char *remote, const struct ovsdb_idl_class *class, - bool monitor_everything_by_default) + bool monitor_everything_by_default, bool retry) { struct ovsdb_idl *idl; uint8_t default_mode; @@ -165,7 +182,7 @@ ovsdb_idl_create(const char *remote, const struct ovsdb_idl_class *class, idl = xzalloc(sizeof *idl); idl->class = class; - idl->session = jsonrpc_session_open(remote); + idl->session = jsonrpc_session_open(remote, retry); shash_init(&idl->table_by_name); idl->tables = xmalloc(class->n_tables * sizeof *idl->tables); for (i = 0; i < class->n_tables; i++) { @@ -200,7 +217,7 @@ ovsdb_idl_destroy(struct ovsdb_idl *idl) if (idl) { size_t i; - assert(!idl->txn); + ovs_assert(!idl->txn); ovsdb_idl_clear(idl); jsonrpc_session_close(idl->session); @@ -213,6 +230,9 @@ ovsdb_idl_destroy(struct ovsdb_idl *idl) shash_destroy(&idl->table_by_name); free(idl->tables); json_destroy(idl->monitor_request_id); + free(idl->lock_name); + json_destroy(idl->lock_request_id); + hmap_destroy(&idl->outstanding_txns); free(idl); } } @@ -253,33 +273,15 @@ ovsdb_idl_clear(struct ovsdb_idl *idl) } } -/* Processes a batch of messages from the database server on 'idl'. Returns - * true if the database as seen through 'idl' changed, false if it did not - * change. The initial fetch of the entire contents of the remote database is - * considered to be one kind of change. - * - * When this function returns false, the client may continue to use any data - * structures it obtained from 'idl' in the past. But when it returns true, - * the client must not access any of these data structures again, because they - * could have freed or reused for other purposes. - * - * This function can return occasional false positives, that is, report that - * the database changed even though it didn't. This happens if the connection - * to the database drops and reconnects, which causes the database contents to - * be reloaded even if they didn't change. (It could also happen if the - * database server sends out a "change" that reflects what we already thought - * was in the database, but the database server is not supposed to do that.) - * - * As an alternative to checking the return value, the client may check for - * changes in the value returned by ovsdb_idl_get_seqno(). - */ -bool +/* Processes a batch of messages from the database server on 'idl'. This may + * cause the IDL's contents to change. The client may check for that with + * ovsdb_idl_get_seqno(). */ +void ovsdb_idl_run(struct ovsdb_idl *idl) { - unsigned int initial_change_seqno = idl->change_seqno; int i; - assert(!idl->txn); + ovs_assert(!idl->txn); jsonrpc_session_run(idl->session); for (i = 0; jsonrpc_session_is_connected(idl->session) && i < 50; i++) { struct jsonrpc_msg *msg; @@ -290,6 +292,9 @@ ovsdb_idl_run(struct ovsdb_idl *idl) idl->last_monitor_request_seqno = seqno; ovsdb_idl_txn_abort_all(idl); ovsdb_idl_send_monitor_request(idl); + if (idl->lock_name) { + ovsdb_idl_send_lock_request(idl); + } break; } @@ -299,22 +304,34 @@ ovsdb_idl_run(struct ovsdb_idl *idl) } if (msg->type == JSONRPC_NOTIFY - && !strcmp(msg->method, "update") - && msg->params->type == JSON_ARRAY - && msg->params->u.array.n == 2 - && msg->params->u.array.elems[0]->type == JSON_NULL) { + && !strcmp(msg->method, "update") + && msg->params->type == JSON_ARRAY + && msg->params->u.array.n == 2 + && msg->params->u.array.elems[0]->type == JSON_NULL) { + /* Database contents changed. */ ovsdb_idl_parse_update(idl, msg->params->u.array.elems[1]); } else if (msg->type == JSONRPC_REPLY && idl->monitor_request_id && json_equal(idl->monitor_request_id, msg->id)) { + /* Reply to our "monitor" request. */ idl->change_seqno++; json_destroy(idl->monitor_request_id); idl->monitor_request_id = NULL; ovsdb_idl_clear(idl); ovsdb_idl_parse_update(idl, msg->result); - } else if (msg->type == JSONRPC_REPLY && msg->id->type == JSON_STRING - && !strcmp(msg->id->u.string, "echo")) { - /* It's a reply to our echo request. Ignore it. */ + } else if (msg->type == JSONRPC_REPLY + && idl->lock_request_id + && json_equal(idl->lock_request_id, msg->id)) { + /* Reply to our "lock" request. */ + ovsdb_idl_parse_lock_reply(idl, msg->result); + } else if (msg->type == JSONRPC_NOTIFY + && !strcmp(msg->method, "locked")) { + /* We got our lock. */ + ovsdb_idl_parse_lock_notify(idl, msg->params, true); + } else if (msg->type == JSONRPC_NOTIFY + && !strcmp(msg->method, "stolen")) { + /* Someone else stole our lock. */ + ovsdb_idl_parse_lock_notify(idl, msg->params, false); } else if ((msg->type == JSONRPC_ERROR || msg->type == JSONRPC_REPLY) && ovsdb_idl_txn_process_reply(idl, msg)) { @@ -329,8 +346,6 @@ ovsdb_idl_run(struct ovsdb_idl *idl) } jsonrpc_msg_destroy(msg); } - - return initial_change_seqno != idl->change_seqno; } /* Arranges for poll_block() to wake up when ovsdb_idl_run() has something to @@ -342,8 +357,23 @@ ovsdb_idl_wait(struct ovsdb_idl *idl) jsonrpc_session_recv_wait(idl->session); } -/* Returns a number that represents the state of 'idl'. When 'idl' is updated - * (by ovsdb_idl_run()), the return value changes. */ +/* Returns a "sequence number" that represents the state of 'idl'. When + * ovsdb_idl_run() changes the database, the sequence number changes. The + * initial fetch of the entire contents of the remote database is considered to + * be one kind of change. Successfully acquiring a lock, if one has been + * configured with ovsdb_idl_set_lock(), is also considered to be a change. + * + * As long as the sequence number does not change, the client may continue to + * use any data structures it obtains from 'idl'. But when it changes, the + * client must not access any of these data structures again, because they + * could have freed or reused for other purposes. + * + * The sequence number can occasionally change even if the database does not. + * This happens if the connection to the database drops and reconnects, which + * causes the database contents to be reloaded even if they didn't change. (It + * could also happen if the database server sends out a "change" that reflects + * what the IDL already thought was in the database. The database server is + * not supposed to do that, but bugs could in theory cause it to do so.) */ unsigned int ovsdb_idl_get_seqno(const struct ovsdb_idl *idl) { @@ -364,6 +394,14 @@ ovsdb_idl_has_ever_connected(const struct ovsdb_idl *idl) return ovsdb_idl_get_seqno(idl) != 0; } +/* Reconfigures 'idl' so that it would reconnect to the database, if + * connection was dropped. */ +void +ovsdb_idl_enable_reconnect(struct ovsdb_idl *idl) +{ + jsonrpc_session_enable_reconnect(idl->session); +} + /* Forces 'idl' to drop its connection to the database and reconnect. In the * meantime, the contents of 'idl' will not change. */ void @@ -371,6 +409,28 @@ ovsdb_idl_force_reconnect(struct ovsdb_idl *idl) { jsonrpc_session_force_reconnect(idl->session); } + +/* Some IDL users should only write to write-only columns. Furthermore, + * writing to a column which is not write-only can cause serious performance + * degradations for these users. This function causes 'idl' to reject writes + * to columns which are not marked write only using ovsdb_idl_omit_alert(). */ +void +ovsdb_idl_verify_write_only(struct ovsdb_idl *idl) +{ + idl->verify_write_only = true; +} + +bool +ovsdb_idl_is_alive(const struct ovsdb_idl *idl) +{ + return jsonrpc_session_is_alive(idl->session); +} + +int +ovsdb_idl_get_last_error(const struct ovsdb_idl *idl) +{ + return jsonrpc_session_get_last_error(idl->session); +} static unsigned char * ovsdb_idl_get_mode(struct ovsdb_idl *idl, @@ -378,7 +438,7 @@ ovsdb_idl_get_mode(struct ovsdb_idl *idl, { size_t i; - assert(!idl->change_seqno); + ovs_assert(!idl->change_seqno); for (i = 0; i < idl->class->n_tables; i++) { const struct ovsdb_idl_table *table = &idl->tables[i]; @@ -389,7 +449,7 @@ ovsdb_idl_get_mode(struct ovsdb_idl *idl, } } - NOT_REACHED(); + OVS_NOT_REACHED(); } static void @@ -453,7 +513,7 @@ ovsdb_idl_add_table(struct ovsdb_idl *idl, } } - NOT_REACHED(); + OVS_NOT_REACHED(); } /* Turns off OVSDB_IDL_ALERT for 'column' in 'idl'. @@ -794,7 +854,7 @@ ovsdb_idl_row_unparse(struct ovsdb_idl_row *row) static void ovsdb_idl_row_clear_old(struct ovsdb_idl_row *row) { - assert(row->old == row->new); + ovs_assert(row->old == row->new); if (!ovsdb_idl_row_is_orphan(row)) { const struct ovsdb_idl_table_class *class = row->table->class; size_t i; @@ -876,6 +936,7 @@ static struct ovsdb_idl_row * ovsdb_idl_row_create__(const struct ovsdb_idl_table_class *class) { struct ovsdb_idl_row *row = xzalloc(class->allocation_size); + class->row_init(row); list_init(&row->src_arcs); list_init(&row->dst_arcs); hmap_node_nullify(&row->txn_node); @@ -908,7 +969,7 @@ ovsdb_idl_insert_row(struct ovsdb_idl_row *row, const struct json *row_json) const struct ovsdb_idl_table_class *class = row->table->class; size_t i; - assert(!row->old && !row->new); + ovs_assert(!row->old && !row->new); row->old = row->new = xmalloc(class->n_columns * sizeof *row->old); for (i = 0; i < class->n_columns; i++) { ovsdb_datum_init_default(&row->old[i], &class->columns[i].type); @@ -977,6 +1038,7 @@ ovsdb_idl_table_from_class(const struct ovsdb_idl *idl, return &idl->tables[table_class - idl->class->tables]; } +/* Called by ovsdb-idlc generated code. */ struct ovsdb_idl_row * ovsdb_idl_get_row_arc(struct ovsdb_idl_row *src, struct ovsdb_idl_table_class *dst_table_class, @@ -1021,6 +1083,8 @@ ovsdb_idl_get_row_arc(struct ovsdb_idl_row *src, } } +/* Searches 'tc''s table in 'idl' for a row with UUID 'uuid'. Returns a + * pointer to the row if there is one, otherwise a null pointer. */ const struct ovsdb_idl_row * ovsdb_idl_get_row_for_uuid(const struct ovsdb_idl *idl, const struct ovsdb_idl_table_class *tc, @@ -1043,6 +1107,12 @@ next_real_row(struct ovsdb_idl_table *table, struct hmap_node *node) return NULL; } +/* Returns a row in 'table_class''s table in 'idl', or a null pointer if that + * table is empty. + * + * Database tables are internally maintained as hash tables, so adding or + * removing rows while traversing the same table can cause some rows to be + * visited twice or not at apply. */ const struct ovsdb_idl_row * ovsdb_idl_first_row(const struct ovsdb_idl *idl, const struct ovsdb_idl_table_class *table_class) @@ -1052,6 +1122,8 @@ ovsdb_idl_first_row(const struct ovsdb_idl *idl, return next_real_row(table, hmap_first(&table->rows)); } +/* Returns a row following 'row' within its table, or a null pointer if 'row' + * is the last row in its table. */ const struct ovsdb_idl_row * ovsdb_idl_next_row(const struct ovsdb_idl_row *row) { @@ -1075,11 +1147,16 @@ const struct ovsdb_datum * ovsdb_idl_read(const struct ovsdb_idl_row *row, const struct ovsdb_idl_column *column) { - const struct ovsdb_idl_table_class *class = row->table->class; - size_t column_idx = column - class->columns; + const struct ovsdb_idl_table_class *class; + size_t column_idx; + + ovs_assert(!ovsdb_idl_row_is_synthetic(row)); - assert(row->new != NULL); - assert(column_idx < class->n_columns); + class = row->table->class; + column_idx = column - class->columns; + + ovs_assert(row->new != NULL); + ovs_assert(column_idx < class->n_columns); if (row->written && bitmap_is_set(row->written, column_idx)) { return &row->new[column_idx]; @@ -1103,21 +1180,37 @@ ovsdb_idl_get(const struct ovsdb_idl_row *row, enum ovsdb_atomic_type key_type OVS_UNUSED, enum ovsdb_atomic_type value_type OVS_UNUSED) { - assert(column->type.key.type == key_type); - assert(column->type.value.type == value_type); + ovs_assert(column->type.key.type == key_type); + ovs_assert(column->type.value.type == value_type); return ovsdb_idl_read(row, column); } + +/* Returns false if 'row' was obtained from the IDL, true if it was initialized + * to all-zero-bits by some other entity. If 'row' was set up some other way + * then the return value is indeterminate. */ +bool +ovsdb_idl_row_is_synthetic(const struct ovsdb_idl_row *row) +{ + return row->table == NULL; +} /* Transactions. */ static void ovsdb_idl_txn_complete(struct ovsdb_idl_txn *txn, enum ovsdb_idl_txn_status); +/* Returns a string representation of 'status'. The caller must not modify or + * free the returned string. + * + * The return value is probably useful only for debug log messages and unit + * tests. */ const char * ovsdb_idl_txn_status_to_string(enum ovsdb_idl_txn_status status) { switch (status) { + case TXN_UNCOMMITTED: + return "uncommitted"; case TXN_UNCHANGED: return "unchanged"; case TXN_INCOMPLETE: @@ -1128,30 +1221,34 @@ ovsdb_idl_txn_status_to_string(enum ovsdb_idl_txn_status status) return "success"; case TXN_TRY_AGAIN: return "try again"; + case TXN_NOT_LOCKED: + return "not locked"; case TXN_ERROR: return "error"; } return ""; } +/* Starts a new transaction on 'idl'. A given ovsdb_idl may only have a single + * active transaction at a time. See the large comment in ovsdb-idl.h for + * general information on transactions. */ struct ovsdb_idl_txn * ovsdb_idl_txn_create(struct ovsdb_idl *idl) { struct ovsdb_idl_txn *txn; - assert(!idl->txn); + ovs_assert(!idl->txn); idl->txn = txn = xmalloc(sizeof *txn); txn->request_id = NULL; txn->idl = idl; hmap_init(&txn->txn_rows); - txn->status = TXN_INCOMPLETE; + txn->status = TXN_UNCOMMITTED; txn->error = NULL; txn->dry_run = false; ds_init(&txn->comment); txn->inc_table = NULL; txn->inc_column = NULL; - txn->inc_where = NULL; hmap_init(&txn->inserted_rows); @@ -1176,22 +1273,50 @@ ovsdb_idl_txn_add_comment(struct ovsdb_idl_txn *txn, const char *s, ...) va_end(args); } +/* Marks 'txn' as a transaction that will not actually modify the database. In + * almost every way, the transaction is treated like other transactions. It + * must be committed or aborted like other transactions, it will be sent to the + * database server like other transactions, and so on. The only difference is + * that the operations sent to the database server will include, as the last + * step, an "abort" operation, so that any changes made by the transaction will + * not actually take effect. */ void ovsdb_idl_txn_set_dry_run(struct ovsdb_idl_txn *txn) { txn->dry_run = true; } +/* Causes 'txn', when committed, to increment the value of 'column' within + * 'row' by 1. 'column' must have an integer type. After 'txn' commits + * successfully, the client may retrieve the final (incremented) value of + * 'column' with ovsdb_idl_txn_get_increment_new_value(). + * + * The client could accomplish something similar with ovsdb_idl_read(), + * ovsdb_idl_txn_verify() and ovsdb_idl_txn_write(), or with ovsdb-idlc + * generated wrappers for these functions. However, ovsdb_idl_txn_increment() + * will never (by itself) fail because of a verify error. + * + * The intended use is for incrementing the "next_cfg" column in the + * Open_vSwitch table. */ void -ovsdb_idl_txn_increment(struct ovsdb_idl_txn *txn, const char *table, - const char *column, const struct json *where) +ovsdb_idl_txn_increment(struct ovsdb_idl_txn *txn, + const struct ovsdb_idl_row *row, + const struct ovsdb_idl_column *column) { - assert(!txn->inc_table); - txn->inc_table = xstrdup(table); - txn->inc_column = xstrdup(column); - txn->inc_where = where ? json_clone(where) : json_array_create_empty(); + ovs_assert(!txn->inc_table); + ovs_assert(column->type.key.type == OVSDB_TYPE_INTEGER); + ovs_assert(column->type.value.type == OVSDB_TYPE_VOID); + + txn->inc_table = row->table->class->name; + txn->inc_column = column->name; + txn->inc_row = row->uuid; } +/* Destroys 'txn' and frees all associated memory. If ovsdb_idl_txn_commit() + * has been called for 'txn' but the commit is still incomplete (that is, the + * last call returned TXN_INCOMPLETE) then the transaction may or may not still + * end up committing at the database server, but the client will not be able to + * get any further status information back. */ void ovsdb_idl_txn_destroy(struct ovsdb_idl_txn *txn) { @@ -1204,9 +1329,6 @@ ovsdb_idl_txn_destroy(struct ovsdb_idl_txn *txn) ovsdb_idl_txn_abort(txn); ds_destroy(&txn->comment); free(txn->error); - free(txn->inc_table); - free(txn->inc_column); - json_destroy(txn->inc_where); HMAP_FOR_EACH_SAFE (insert, next, hmap_node, &txn->inserted_rows) { free(insert); } @@ -1214,10 +1336,11 @@ ovsdb_idl_txn_destroy(struct ovsdb_idl_txn *txn) free(txn); } +/* Causes poll_block() to wake up if 'txn' has completed committing. */ void ovsdb_idl_txn_wait(const struct ovsdb_idl_txn *txn) { - if (txn->status != TXN_INCOMPLETE) { + if (txn->status != TXN_UNCOMMITTED && txn->status != TXN_INCOMPLETE) { poll_immediate_wake(); } } @@ -1344,6 +1467,55 @@ ovsdb_idl_txn_disassemble(struct ovsdb_idl_txn *txn) hmap_init(&txn->txn_rows); } +/* Attempts to commit 'txn'. Returns the status of the commit operation, one + * of the following TXN_* constants: + * + * TXN_INCOMPLETE: + * + * The transaction is in progress, but not yet complete. The caller + * should call again later, after calling ovsdb_idl_run() to let the IDL + * do OVSDB protocol processing. + * + * TXN_UNCHANGED: + * + * The transaction is complete. (It didn't actually change the database, + * so the IDL didn't send any request to the database server.) + * + * TXN_ABORTED: + * + * The caller previously called ovsdb_idl_txn_abort(). + * + * TXN_SUCCESS: + * + * The transaction was successful. The update made by the transaction + * (and possibly other changes made by other database clients) should + * already be visible in the IDL. + * + * TXN_TRY_AGAIN: + * + * The transaction failed for some transient reason, e.g. because a + * "verify" operation reported an inconsistency or due to a network + * problem. The caller should wait for a change to the database, then + * compose a new transaction, and commit the new transaction. + * + * Use the return value of ovsdb_idl_get_seqno() to wait for a change in + * the database. It is important to use its return value *before* the + * initial call to ovsdb_idl_txn_commit() as the baseline for this + * purpose, because the change that one should wait for can happen after + * the initial call but before the call that returns TXN_TRY_AGAIN, and + * using some other baseline value in that situation could cause an + * indefinite wait if the database rarely changes. + * + * TXN_NOT_LOCKED: + * + * The transaction failed because the IDL has been configured to require + * a database lock (with ovsdb_idl_set_lock()) but didn't get it yet or + * has already lost it. + * + * Committing a transaction rolls back all of the changes that it made to the + * IDL's copy of the database. If the transaction commits successfully, then + * the database server will send an update and, thus, the IDL will be updated + * with the committed changes. */ enum ovsdb_idl_txn_status ovsdb_idl_txn_commit(struct ovsdb_idl_txn *txn) { @@ -1355,9 +1527,24 @@ ovsdb_idl_txn_commit(struct ovsdb_idl_txn *txn) return txn->status; } + /* If we need a lock but don't have it, give up quickly. */ + if (txn->idl->lock_name && !ovsdb_idl_has_lock(txn->idl)) { + txn->status = TXN_NOT_LOCKED; + ovsdb_idl_txn_disassemble(txn); + return txn->status; + } + operations = json_array_create_1( json_string_create(txn->idl->class->database)); + /* Assert that we have the required lock (avoiding a race). */ + if (txn->idl->lock_name) { + struct json *op = json_object_create(); + json_array_add(operations, op); + json_object_put_string(op, "op", "assert"); + json_object_put_string(op, "lock", txn->idl->lock_name); + } + /* Add prerequisites and declarations of new rows. */ HMAP_FOR_EACH (row, txn_node, &txn->txn_rows) { /* XXX check that deleted rows exist even if no prereqs? */ @@ -1394,16 +1581,18 @@ ovsdb_idl_txn_commit(struct ovsdb_idl_txn *txn) HMAP_FOR_EACH (row, txn_node, &txn->txn_rows) { const struct ovsdb_idl_table_class *class = row->table->class; - if (row->old == row->new) { - continue; - } else if (!row->new) { - struct json *op = json_object_create(); - json_object_put_string(op, "op", "delete"); - json_object_put_string(op, "table", class->name); - json_object_put(op, "where", where_uuid_equals(&row->uuid)); - json_array_add(operations, op); - any_updates = true; - } else { + if (!row->new) { + if (class->is_root) { + struct json *op = json_object_create(); + json_object_put_string(op, "op", "delete"); + json_object_put_string(op, "table", class->name); + json_object_put(op, "where", where_uuid_equals(&row->uuid)); + json_array_add(operations, op); + any_updates = true; + } else { + /* Let ovsdb-server decide whether to really delete it. */ + } + } else if (row->old != row->new) { struct json *row_json; struct json *op; size_t idx; @@ -1416,6 +1605,8 @@ ovsdb_idl_txn_commit(struct ovsdb_idl_txn *txn) } else { struct ovsdb_idl_txn_insert *insert; + any_updates = true; + json_object_put(op, "uuid-name", json_string_create_nocopy( uuid_name_from_uuid(&row->uuid))); @@ -1436,22 +1627,29 @@ ovsdb_idl_txn_commit(struct ovsdb_idl_txn *txn) &class->columns[idx]; if (row->old - ? !ovsdb_datum_equals(&row->old[idx], &row->new[idx], - &column->type) - : !ovsdb_datum_is_default(&row->new[idx], + || !ovsdb_datum_is_default(&row->new[idx], &column->type)) { json_object_put(row_json, column->name, substitute_uuids( ovsdb_datum_to_json(&row->new[idx], &column->type), txn)); + + /* If anything really changed, consider it an update. + * We can't suppress not-really-changed values earlier + * or transactions would become nonatomic (see the big + * comment inside ovsdb_idl_txn_write()). */ + if (!any_updates && row->old && + !ovsdb_datum_equals(&row->old[idx], &row->new[idx], + &column->type)) { + any_updates = true; + } } } } if (!row->old || !shash_is_empty(json_object(row_json))) { json_array_add(operations, op); - any_updates = true; } else { json_destroy(op); } @@ -1468,7 +1666,8 @@ ovsdb_idl_txn_commit(struct ovsdb_idl_txn *txn) json_object_put_string(op, "op", "mutate"); json_object_put_string(op, "table", txn->inc_table); json_object_put(op, "where", - substitute_uuids(json_clone(txn->inc_where), txn)); + substitute_uuids(where_uuid_equals(&txn->inc_row), + txn)); json_object_put(op, "mutations", json_array_create_1( json_array_create_3( @@ -1481,7 +1680,8 @@ ovsdb_idl_txn_commit(struct ovsdb_idl_txn *txn) json_object_put_string(op, "op", "select"); json_object_put_string(op, "table", txn->inc_table); json_object_put(op, "where", - substitute_uuids(json_clone(txn->inc_where), txn)); + substitute_uuids(where_uuid_equals(&txn->inc_row), + txn)); json_object_put(op, "columns", json_array_create_1(json_string_create( txn->inc_column))); @@ -1510,6 +1710,7 @@ ovsdb_idl_txn_commit(struct ovsdb_idl_txn *txn) "transact", operations, &txn->request_id))) { hmap_insert(&txn->idl->outstanding_txns, &txn->hmap_node, json_hash(txn->request_id, 0)); + txn->status = TXN_INCOMPLETE; } else { txn->status = TXN_TRY_AGAIN; } @@ -1520,7 +1721,10 @@ ovsdb_idl_txn_commit(struct ovsdb_idl_txn *txn) /* Attempts to commit 'txn', blocking until the commit either succeeds or * fails. Returns the final commit status, which may be any TXN_* value other - * than TXN_INCOMPLETE. */ + * than TXN_INCOMPLETE. + * + * This function calls ovsdb_idl_run() on 'txn''s IDL, so it may cause the + * return value of ovsdb_idl_get_seqno() to change. */ enum ovsdb_idl_txn_status ovsdb_idl_txn_commit_block(struct ovsdb_idl_txn *txn) { @@ -1536,22 +1740,39 @@ ovsdb_idl_txn_commit_block(struct ovsdb_idl_txn *txn) return status; } +/* Returns the final (incremented) value of the column in 'txn' that was set to + * be incremented by ovsdb_idl_txn_increment(). 'txn' must have committed + * successfully. */ int64_t ovsdb_idl_txn_get_increment_new_value(const struct ovsdb_idl_txn *txn) { - assert(txn->status == TXN_SUCCESS); + ovs_assert(txn->status == TXN_SUCCESS); return txn->inc_new_value; } +/* Aborts 'txn' without sending it to the database server. This is effective + * only if ovsdb_idl_txn_commit() has not yet been called for 'txn'. + * Otherwise, it has no effect. + * + * Aborting a transaction doesn't free its memory. Use + * ovsdb_idl_txn_destroy() to do that. */ void ovsdb_idl_txn_abort(struct ovsdb_idl_txn *txn) { ovsdb_idl_txn_disassemble(txn); - if (txn->status == TXN_INCOMPLETE) { + if (txn->status == TXN_UNCOMMITTED || txn->status == TXN_INCOMPLETE) { txn->status = TXN_ABORTED; } } +/* Returns a string that reports the error status for 'txn'. The caller must + * not modify or free the returned string. A call to ovsdb_idl_txn_destroy() + * for 'txn' may free the returned string. + * + * The return value is ordinarily one of the strings that + * ovsdb_idl_txn_status_to_string() would return, but if the transaction failed + * due to an error reported by the database server, the return value is that + * error. */ const char * ovsdb_idl_txn_get_error(const struct ovsdb_idl_txn *txn) { @@ -1588,7 +1809,7 @@ ovsdb_idl_txn_get_insert_uuid(const struct ovsdb_idl_txn *txn, { const struct ovsdb_idl_txn_insert *insert; - assert(txn->status == TXN_SUCCESS || txn->status == TXN_UNCHANGED); + ovs_assert(txn->status == TXN_SUCCESS || txn->status == TXN_UNCHANGED); HMAP_FOR_EACH_IN_BUCKET (insert, hmap_node, uuid_hash(uuid), &txn->inserted_rows) { if (uuid_equals(uuid, &insert->dummy)) { @@ -1618,20 +1839,55 @@ ovsdb_idl_txn_complete(struct ovsdb_idl_txn *txn, * been disabled (by calling ovsdb_idl_omit()). * * Usually this function is used indirectly through one of the "set" functions - * generated by ovsdb-idlc. */ -void -ovsdb_idl_txn_write(const struct ovsdb_idl_row *row_, - const struct ovsdb_idl_column *column, - struct ovsdb_datum *datum) + * generated by ovsdb-idlc. + * + * Takes ownership of what 'datum' points to (and in some cases destroys that + * data before returning) but makes a copy of 'datum' itself. (Commonly + * 'datum' is on the caller's stack.) */ +static void +ovsdb_idl_txn_write__(const struct ovsdb_idl_row *row_, + const struct ovsdb_idl_column *column, + struct ovsdb_datum *datum, bool owns_datum) { - struct ovsdb_idl_row *row = (struct ovsdb_idl_row *) row_; - const struct ovsdb_idl_table_class *class = row->table->class; - size_t column_idx = column - class->columns; + struct ovsdb_idl_row *row = CONST_CAST(struct ovsdb_idl_row *, row_); + const struct ovsdb_idl_table_class *class; + size_t column_idx; + bool write_only; - assert(row->new != NULL); - assert(column_idx < class->n_columns); - assert(row->old == NULL || - row->table->modes[column_idx] & OVSDB_IDL_MONITOR); + if (ovsdb_idl_row_is_synthetic(row)) { + goto discard_datum; + } + + class = row->table->class; + column_idx = column - class->columns; + write_only = row->table->modes[column_idx] == OVSDB_IDL_MONITOR; + + ovs_assert(row->new != NULL); + ovs_assert(column_idx < class->n_columns); + ovs_assert(row->old == NULL || + row->table->modes[column_idx] & OVSDB_IDL_MONITOR); + + if (row->table->idl->verify_write_only && !write_only) { + VLOG_ERR("Bug: Attempt to write to a read/write column (%s:%s) when" + " explicitly configured not to.", class->name, column->name); + goto discard_datum; + } + + /* If this is a write-only column and the datum being written is the same + * as the one already there, just skip the update entirely. This is worth + * optimizing because we have a lot of columns that get periodically + * refreshed into the database but don't actually change that often. + * + * We don't do this for read/write columns because that would break + * atomicity of transactions--some other client might have written a + * different value in that column since we read it. (But if a whole + * transaction only does writes of existing values, without making any real + * changes, we will drop the whole transaction later in + * ovsdb_idl_txn_commit().) */ + if (write_only && ovsdb_datum_equals(ovsdb_idl_read(row, column), + datum, &column->type)) { + goto discard_datum; + } if (hmap_node_is_null(&row->txn_node)) { hmap_insert(&row->table->idl->txn->txn_rows, &row->txn_node, @@ -1648,16 +1904,45 @@ ovsdb_idl_txn_write(const struct ovsdb_idl_row *row_, } else { bitmap_set1(row->written, column_idx); } - row->new[column_idx] = *datum; + if (owns_datum) { + row->new[column_idx] = *datum; + } else { + ovsdb_datum_clone(&row->new[column_idx], datum, &column->type); + } (column->unparse)(row); (column->parse)(row, &row->new[column_idx]); + return; + +discard_datum: + if (owns_datum) { + ovsdb_datum_destroy(datum, &column->type); + } +} + +void +ovsdb_idl_txn_write(const struct ovsdb_idl_row *row, + const struct ovsdb_idl_column *column, + struct ovsdb_datum *datum) +{ + ovsdb_idl_txn_write__(row, column, datum, true); +} + +void +ovsdb_idl_txn_write_clone(const struct ovsdb_idl_row *row, + const struct ovsdb_idl_column *column, + const struct ovsdb_datum *datum) +{ + ovsdb_idl_txn_write__(row, column, + CONST_CAST(struct ovsdb_datum *, datum), false); } /* Causes the original contents of 'column' in 'row_' to be verified as a * prerequisite to completing the transaction. That is, if 'column' in 'row_' * changed (or if 'row_' was deleted) between the time that the IDL originally * read its contents and the time that the transaction commits, then the - * transaction aborts and ovsdb_idl_txn_commit() returns TXN_TRY_AGAIN. + * transaction aborts and ovsdb_idl_txn_commit() returns TXN_AGAIN_WAIT or + * TXN_AGAIN_NOW (depending on whether the database change has already been + * received). * * The intention is that, to ensure that no transaction commits based on dirty * reads, an application should call ovsdb_idl_txn_verify() on each data item @@ -1683,13 +1968,20 @@ void ovsdb_idl_txn_verify(const struct ovsdb_idl_row *row_, const struct ovsdb_idl_column *column) { - struct ovsdb_idl_row *row = (struct ovsdb_idl_row *) row_; - const struct ovsdb_idl_table_class *class = row->table->class; - size_t column_idx = column - class->columns; + struct ovsdb_idl_row *row = CONST_CAST(struct ovsdb_idl_row *, row_); + const struct ovsdb_idl_table_class *class; + size_t column_idx; - assert(row->new != NULL); - assert(row->old == NULL || - row->table->modes[column_idx] & OVSDB_IDL_MONITOR); + if (ovsdb_idl_row_is_synthetic(row)) { + return; + } + + class = row->table->class; + column_idx = column - class->columns; + + ovs_assert(row->new != NULL); + ovs_assert(row->old == NULL || + row->table->modes[column_idx] & OVSDB_IDL_MONITOR); if (!row->old || (row->written && bitmap_is_set(row->written, column_idx))) { return; @@ -1715,13 +2007,17 @@ ovsdb_idl_txn_verify(const struct ovsdb_idl_row *row_, void ovsdb_idl_txn_delete(const struct ovsdb_idl_row *row_) { - struct ovsdb_idl_row *row = (struct ovsdb_idl_row *) row_; + struct ovsdb_idl_row *row = CONST_CAST(struct ovsdb_idl_row *, row_); - assert(row->new != NULL); + if (ovsdb_idl_row_is_synthetic(row)) { + return; + } + + ovs_assert(row->new != NULL); if (!row->old) { ovsdb_idl_row_unparse(row); ovsdb_idl_row_clear_new(row); - assert(!row->prereqs); + ovs_assert(!row->prereqs); hmap_remove(&row->table->rows, &row->hmap_node); hmap_remove(&row->table->idl->txn->txn_rows, &row->txn_node); free(row); @@ -1755,7 +2051,7 @@ ovsdb_idl_txn_insert(struct ovsdb_idl_txn *txn, struct ovsdb_idl_row *row = ovsdb_idl_row_create__(class); if (uuid) { - assert(!ovsdb_idl_txn_get_row(txn, uuid)); + ovs_assert(!ovsdb_idl_txn_get_row(txn, uuid)); row->uuid = *uuid; } else { uuid_generate(&row->uuid); @@ -1817,7 +2113,7 @@ ovsdb_idl_txn_process_inc_reply(struct ovsdb_idl_txn *txn, if (txn->inc_index + 2 > results->n) { VLOG_WARN_RL(&syntax_rl, "reply does not contain enough operations " - "for increment (has %zu, needs %u)", + "for increment (has %"PRIuSIZE", needs %u)", results->n, txn->inc_index + 2); return false; } @@ -1842,7 +2138,7 @@ ovsdb_idl_txn_process_inc_reply(struct ovsdb_idl_txn *txn, return false; } if (rows->u.array.n != 1) { - VLOG_WARN_RL(&syntax_rl, "\"select\" reply \"rows\" has %zu elements " + VLOG_WARN_RL(&syntax_rl, "\"select\" reply \"rows\" has %"PRIuSIZE" elements " "instead of 1", rows->u.array.n); return false; @@ -1872,7 +2168,7 @@ ovsdb_idl_txn_process_insert_reply(struct ovsdb_idl_txn_insert *insert, if (insert->op_index >= results->n) { VLOG_WARN_RL(&syntax_rl, "reply does not contain enough operations " - "for insert (has %zu, needs %u)", + "for insert (has %"PRIuSIZE", needs %u)", results->n, insert->op_index); return false; } @@ -1891,6 +2187,7 @@ ovsdb_idl_txn_process_insert_reply(struct ovsdb_idl_txn_insert *insert, VLOG_WARN_RL(&syntax_rl, "\"insert\" reply \"uuid\" is not a JSON " "UUID: %s", s); free(s); + ovsdb_error_destroy(error); return false; } @@ -1920,6 +2217,7 @@ ovsdb_idl_txn_process_reply(struct ovsdb_idl *idl, struct json_array *ops = &msg->result->u.array; int hard_errors = 0; int soft_errors = 0; + int lock_errors = 0; size_t i; for (i = 0; i < ops->n; i++) { @@ -1937,6 +2235,8 @@ ovsdb_idl_txn_process_reply(struct ovsdb_idl *idl, if (error->type == JSON_STRING) { if (!strcmp(error->u.string, "timed out")) { soft_errors++; + } else if (!strcmp(error->u.string, "not owner")) { + lock_errors++; } else if (strcmp(error->u.string, "aborted")) { hard_errors++; ovsdb_idl_txn_set_error_json(txn, op); @@ -1956,7 +2256,7 @@ ovsdb_idl_txn_process_reply(struct ovsdb_idl *idl, } } - if (!soft_errors && !hard_errors) { + if (!soft_errors && !hard_errors && !lock_errors) { struct ovsdb_idl_txn_insert *insert; if (txn->inc_table && !ovsdb_idl_txn_process_inc_reply(txn, ops)) { @@ -1971,6 +2271,7 @@ ovsdb_idl_txn_process_reply(struct ovsdb_idl *idl, } status = (hard_errors ? TXN_ERROR + : lock_errors ? TXN_NOT_LOCKED : soft_errors ? TXN_TRY_AGAIN : TXN_SUCCESS); } @@ -1979,17 +2280,155 @@ ovsdb_idl_txn_process_reply(struct ovsdb_idl *idl, return true; } +/* Returns the transaction currently active for 'row''s IDL. A transaction + * must currently be active. */ struct ovsdb_idl_txn * ovsdb_idl_txn_get(const struct ovsdb_idl_row *row) { struct ovsdb_idl_txn *txn = row->table->idl->txn; - assert(txn != NULL); + ovs_assert(txn != NULL); return txn; } +/* Returns the IDL on which 'txn' acts. */ struct ovsdb_idl * ovsdb_idl_txn_get_idl (struct ovsdb_idl_txn *txn) { return txn->idl; } + +/* If 'lock_name' is nonnull, configures 'idl' to obtain the named lock from + * the database server and to avoid modifying the database when the lock cannot + * be acquired (that is, when another client has the same lock). + * + * If 'lock_name' is NULL, drops the locking requirement and releases the + * lock. */ +void +ovsdb_idl_set_lock(struct ovsdb_idl *idl, const char *lock_name) +{ + ovs_assert(!idl->txn); + ovs_assert(hmap_is_empty(&idl->outstanding_txns)); + + if (idl->lock_name && (!lock_name || strcmp(lock_name, idl->lock_name))) { + /* Release previous lock. */ + ovsdb_idl_send_unlock_request(idl); + free(idl->lock_name); + idl->lock_name = NULL; + idl->is_lock_contended = false; + } + + if (lock_name && !idl->lock_name) { + /* Acquire new lock. */ + idl->lock_name = xstrdup(lock_name); + ovsdb_idl_send_lock_request(idl); + } +} + +/* Returns true if 'idl' is configured to obtain a lock and owns that lock. + * + * Locking and unlocking happens asynchronously from the database client's + * point of view, so the information is only useful for optimization (e.g. if + * the client doesn't have the lock then there's no point in trying to write to + * the database). */ +bool +ovsdb_idl_has_lock(const struct ovsdb_idl *idl) +{ + return idl->has_lock; +} +/* Returns true if 'idl' is configured to obtain a lock but the database server + * has indicated that some other client already owns the requested lock. */ +bool +ovsdb_idl_is_lock_contended(const struct ovsdb_idl *idl) +{ + return idl->is_lock_contended; +} + +static void +ovsdb_idl_update_has_lock(struct ovsdb_idl *idl, bool new_has_lock) +{ + if (new_has_lock && !idl->has_lock) { + if (!idl->monitor_request_id) { + idl->change_seqno++; + } else { + /* We're waiting for a monitor reply, so don't signal that the + * database changed. The monitor reply will increment change_seqno + * anyhow. */ + } + idl->is_lock_contended = false; + } + idl->has_lock = new_has_lock; +} + +static void +ovsdb_idl_send_lock_request__(struct ovsdb_idl *idl, const char *method, + struct json **idp) +{ + ovsdb_idl_update_has_lock(idl, false); + + json_destroy(idl->lock_request_id); + idl->lock_request_id = NULL; + + if (jsonrpc_session_is_connected(idl->session)) { + struct json *params; + + params = json_array_create_1(json_string_create(idl->lock_name)); + jsonrpc_session_send(idl->session, + jsonrpc_create_request(method, params, idp)); + } +} + +static void +ovsdb_idl_send_lock_request(struct ovsdb_idl *idl) +{ + ovsdb_idl_send_lock_request__(idl, "lock", &idl->lock_request_id); +} + +static void +ovsdb_idl_send_unlock_request(struct ovsdb_idl *idl) +{ + ovsdb_idl_send_lock_request__(idl, "unlock", NULL); +} + +static void +ovsdb_idl_parse_lock_reply(struct ovsdb_idl *idl, const struct json *result) +{ + bool got_lock; + + json_destroy(idl->lock_request_id); + idl->lock_request_id = NULL; + + if (result->type == JSON_OBJECT) { + const struct json *locked; + + locked = shash_find_data(json_object(result), "locked"); + got_lock = locked && locked->type == JSON_TRUE; + } else { + got_lock = false; + } + + ovsdb_idl_update_has_lock(idl, got_lock); + if (!got_lock) { + idl->is_lock_contended = true; + } +} + +static void +ovsdb_idl_parse_lock_notify(struct ovsdb_idl *idl, + const struct json *params, + bool new_has_lock) +{ + if (idl->lock_name + && params->type == JSON_ARRAY + && json_array(params)->n > 0 + && json_array(params)->elems[0]->type == JSON_STRING) { + const char *lock_name = json_string(json_array(params)->elems[0]); + + if (!strcmp(idl->lock_name, lock_name)) { + ovsdb_idl_update_has_lock(idl, new_has_lock); + if (!new_has_lock) { + idl->is_lock_contended = true; + } + } + } +}