From c955d1d1d769f8864a43cf98dc9dbacf5e6c7f9c Mon Sep 17 00:00:00 2001 From: Nathan Bossart Date: Fri, 28 Jun 2024 11:02:44 -0500 Subject: [PATCH v5 03/13] introduce framework for parallelizing pg_upgrade tasks --- src/bin/pg_upgrade/Makefile | 1 + src/bin/pg_upgrade/async.c | 327 +++++++++++++++++++++++++++++++ src/bin/pg_upgrade/meson.build | 1 + src/bin/pg_upgrade/pg_upgrade.h | 16 ++ src/tools/pgindent/typedefs.list | 4 + 5 files changed, 349 insertions(+) create mode 100644 src/bin/pg_upgrade/async.c diff --git a/src/bin/pg_upgrade/Makefile b/src/bin/pg_upgrade/Makefile index bde91e2beb..3bc4f5d740 100644 --- a/src/bin/pg_upgrade/Makefile +++ b/src/bin/pg_upgrade/Makefile @@ -12,6 +12,7 @@ include $(top_builddir)/src/Makefile.global OBJS = \ $(WIN32RES) \ + async.o \ check.o \ controldata.o \ dump.o \ diff --git a/src/bin/pg_upgrade/async.c b/src/bin/pg_upgrade/async.c new file mode 100644 index 0000000000..742e9df37c --- /dev/null +++ b/src/bin/pg_upgrade/async.c @@ -0,0 +1,327 @@ +/* + * async.c + * + * parallelization via libpq's async APIs + * + * Copyright (c) 2024, PostgreSQL Global Development Group + * src/bin/pg_upgrade/async.c + */ + +#include "postgres_fe.h" + +#include "common/connect.h" +#include "fe_utils/string_utils.h" +#include "pg_upgrade.h" + +static int dbs_complete; +static int dbs_processing; + +typedef struct AsyncTaskCallbacks +{ + AsyncTaskGetQueryCB query_cb; + AsyncTaskProcessCB process_cb; + char *query; + bool free_result; + void *arg; +} AsyncTaskCallbacks; + +typedef struct AsyncTask +{ + AsyncTaskCallbacks *cbs; + int num_cb_sets; +} AsyncTask; + +typedef enum +{ + FREE, + CONNECTING, + SETTING_SEARCH_PATH, + RUNNING_QUERY, +} AsyncSlotState; + +typedef struct +{ + AsyncSlotState state; + int db; + int query; + PGconn *conn; +} AsyncSlot; + +AsyncTask * +async_task_create(void) +{ + return pg_malloc0(sizeof(AsyncTask)); +} + +void +async_task_free(AsyncTask *task) +{ + for (int i = 0; i < task->num_cb_sets; i++) + { + if (task->cbs[i].query) + pg_free(task->cbs[i].query); + } + + if (task->cbs) + pg_free(task->cbs); + + pg_free(task); +} + +void +async_task_add_step(AsyncTask *task, + AsyncTaskGetQueryCB query_cb, + AsyncTaskProcessCB process_cb, bool free_result, + void *arg) +{ + AsyncTaskCallbacks *new_cbs; + + task->cbs = pg_realloc(task->cbs, + ++task->num_cb_sets * sizeof(AsyncTaskCallbacks)); + + new_cbs = &task->cbs[task->num_cb_sets - 1]; + new_cbs->query_cb = query_cb; + new_cbs->process_cb = process_cb; + new_cbs->query = NULL; + new_cbs->free_result = free_result; + new_cbs->arg = arg; +} + +static void +conn_failure(PGconn *conn) +{ + pg_fatal("connection failure: %s", PQerrorMessage(conn)); +} + +static void +start_conn(const ClusterInfo *cluster, AsyncSlot *slot) +{ + PQExpBufferData conn_opts; + + /* Build connection string with proper quoting */ + initPQExpBuffer(&conn_opts); + appendPQExpBufferStr(&conn_opts, "dbname="); + appendConnStrVal(&conn_opts, cluster->dbarr.dbs[slot->db].db_name); + appendPQExpBufferStr(&conn_opts, " user="); + appendConnStrVal(&conn_opts, os_info.user); + appendPQExpBuffer(&conn_opts, " port=%d", cluster->port); + if (cluster->sockdir) + { + appendPQExpBufferStr(&conn_opts, " host="); + appendConnStrVal(&conn_opts, cluster->sockdir); + } + + slot->conn = PQconnectStart(conn_opts.data); + termPQExpBuffer(&conn_opts); + + if (!slot->conn) + conn_failure(slot->conn); +} + +static void +dispatch_query(const ClusterInfo *cluster, AsyncSlot *slot, + const AsyncTask *task) +{ + AsyncTaskCallbacks *cbs = &task->cbs[slot->query]; + + if (!cbs->query) + cbs->query = cbs->query_cb(cbs->arg); + + if (!PQsendQuery(slot->conn, cbs->query)) + conn_failure(slot->conn); +} + +static PGresult * +get_last_result(PGconn *conn) +{ + PGresult *tmp; + PGresult *res = NULL; + + while ((tmp = PQgetResult(conn)) != NULL) + { + PQclear(res); + res = tmp; + if (PQstatus(conn) == CONNECTION_BAD) + conn_failure(conn); + } + + if (PQresultStatus(res) != PGRES_COMMAND_OK && + PQresultStatus(res) != PGRES_TUPLES_OK) + conn_failure(conn); + + return res; +} + +static void +process_query_result(const ClusterInfo *cluster, AsyncSlot *slot, + const AsyncTask *task) +{ + AsyncTaskCallbacks *cbs = &task->cbs[slot->query]; + AsyncTaskProcessCB process_cb = cbs->process_cb; + DbInfo *dbinfo = &cluster->dbarr.dbs[slot->db]; + PGresult *res = get_last_result(slot->conn); + + (*process_cb) (dbinfo, res, cbs->arg); + + if (cbs->free_result) + PQclear(res); +} + +static void +process_slot(const ClusterInfo *cluster, AsyncSlot *slot, const AsyncTask *task) +{ + switch (slot->state) + { + case FREE: + if (dbs_processing >= cluster->dbarr.ndbs) + return; + slot->db = dbs_processing++; + slot->state = CONNECTING; + start_conn(cluster, slot); + return; + + case CONNECTING: + if (PQconnectPoll(slot->conn) == PGRES_POLLING_FAILED) + conn_failure(slot->conn); + if (PQconnectPoll(slot->conn) != PGRES_POLLING_OK) + return; + slot->state = SETTING_SEARCH_PATH; + if (!PQsendQuery(slot->conn, ALWAYS_SECURE_SEARCH_PATH_SQL)) + conn_failure(slot->conn); + return; + + case SETTING_SEARCH_PATH: + if (!PQconsumeInput(slot->conn)) + conn_failure(slot->conn); + if (PQisBusy(slot->conn)) + return; + PQclear(get_last_result(slot->conn)); + slot->state = RUNNING_QUERY; + dispatch_query(cluster, slot, task); + return; + + case RUNNING_QUERY: + if (!PQconsumeInput(slot->conn)) + conn_failure(slot->conn); + if (PQisBusy(slot->conn)) + return; + process_query_result(cluster, slot, task); + if (++slot->query >= task->num_cb_sets) + { + dbs_complete++; + PQfinish(slot->conn); + memset(slot, 0, sizeof(AsyncSlot)); + return; + } + dispatch_query(cluster, slot, task); + return; + } +} + +/* + * Wait on the slots to either finish connecting or to receive query results if + * possible. This avoids a tight loop in async_task_run(). + */ +static void +wait_on_slots(AsyncSlot *slots, int numslots) +{ + fd_set input_mask; + fd_set output_mask; + fd_set except_mask; + int maxFd = 0; + + FD_ZERO(&input_mask); + FD_ZERO(&output_mask); + FD_ZERO(&except_mask); + + for (int i = 0; i < numslots; i++) + { + int sock; + bool read = false; + + switch (slots[i].state) + { + case FREE: + + /* + * If we see a free slot, return right away so that it can be + * reused immediately for the next database. This might cause + * us to spin more than necessary as we finish processing the + * last few databases, but that shouldn't cause too much harm. + */ + return; + + case CONNECTING: + + /* + * If we are waiting for the connection to establish, choose + * whether to wait for reading or for writing on the socket as + * appropriate. If neither apply, just return immediately so + * that we can handle the slot. + */ + { + PostgresPollingStatusType status; + + status = PQconnectPoll(slots[i].conn); + if (status == PGRES_POLLING_READING) + read = true; + else if (status != PGRES_POLLING_WRITING) + return; + } + break; + + case SETTING_SEARCH_PATH: + case RUNNING_QUERY: + + /* + * If we've sent a query, we must wait for the socket to be + * read-ready. Note that process_slot() handles calling + * PQconsumeInput() as required. + */ + read = true; + break; + } + + /* + * If there's some problem retrieving the socket, just pretend this + * slot doesn't exist. We don't expect this to happen regularly in + * practice, so it seems unlikely to cause too much harm. + */ + sock = PQsocket(slots[i].conn); + if (sock < 0) + continue; + + /* + * Add the socket to the set. + */ + FD_SET(sock, read ? &input_mask : &output_mask); + FD_SET(sock, &except_mask); + maxFd = Max(maxFd, sock); + } + + /* + * If we found socket(s) to wait on, wait. + */ + if (maxFd != 0) + (void) select(maxFd + 1, &input_mask, &output_mask, &except_mask, NULL); +} + +void +async_task_run(const AsyncTask *task, const ClusterInfo *cluster) +{ + int jobs = Max(1, user_opts.jobs); + AsyncSlot *slots = pg_malloc0(sizeof(AsyncSlot) * jobs); + + dbs_complete = 0; + dbs_processing = 0; + + while (dbs_complete < cluster->dbarr.ndbs) + { + for (int i = 0; i < jobs; i++) + process_slot(cluster, &slots[i], task); + + wait_on_slots(slots, jobs); + } + + pg_free(slots); +} diff --git a/src/bin/pg_upgrade/meson.build b/src/bin/pg_upgrade/meson.build index 9825fa3305..9eb48e176c 100644 --- a/src/bin/pg_upgrade/meson.build +++ b/src/bin/pg_upgrade/meson.build @@ -1,6 +1,7 @@ # Copyright (c) 2022-2024, PostgreSQL Global Development Group pg_upgrade_sources = files( + 'async.c', 'check.c', 'controldata.c', 'dump.c', diff --git a/src/bin/pg_upgrade/pg_upgrade.h b/src/bin/pg_upgrade/pg_upgrade.h index 08b2ad26d7..dea418ec3f 100644 --- a/src/bin/pg_upgrade/pg_upgrade.h +++ b/src/bin/pg_upgrade/pg_upgrade.h @@ -494,3 +494,19 @@ void parallel_transfer_all_new_dbs(DbInfoArr *old_db_arr, DbInfoArr *new_db_arr char *old_pgdata, char *new_pgdata, char *old_tablespace); bool reap_child(bool wait_for_child); + +/* async.c */ + +typedef char *(*AsyncTaskGetQueryCB) (void *arg); +typedef void (*AsyncTaskProcessCB) (DbInfo *dbinfo, PGresult *res, void *arg); + +/* struct definition is private to async.c */ +typedef struct AsyncTask AsyncTask; + +AsyncTask *async_task_create(void); +void async_task_add_step(AsyncTask *task, + AsyncTaskGetQueryCB query_cb, + AsyncTaskProcessCB process_cb, bool free_result, + void *arg); +void async_task_run(const AsyncTask *task, const ClusterInfo *cluster); +void async_task_free(AsyncTask *task); diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list index b4d7f9217c..a2330f4346 100644 --- a/src/tools/pgindent/typedefs.list +++ b/src/tools/pgindent/typedefs.list @@ -153,6 +153,10 @@ ArrayMetaState ArraySubWorkspace ArrayToken ArrayType +AsyncSlot +AsyncSlotState +AsyncTask +AsyncTaskCallbacks AsyncQueueControl AsyncQueueEntry AsyncRequest -- 2.39.3 (Apple Git-146)