applyparallelworker.c

Go to the documentation of this file.

/*-------------------------------------------------------------------------
 * applyparallelworker.c
 *     Support routines for applying xact by parallel apply worker
 *
 * Copyright (c) 2023, PostgreSQL Global Development Group
 *
 * IDENTIFICATION
 *    src/backend/replication/logical/applyparallelworker.c
 *
 * This file contains the code to launch, set up, and teardown a parallel apply
 * worker which receives the changes from the leader worker and invokes routines
 * to apply those on the subscriber database. Additionally, this file contains
 * routines that are intended to support setting up, using, and tearing down a
 * ParallelApplyWorkerInfo which is required so the leader worker and parallel
 * apply workers can communicate with each other.
 *
 * The parallel apply workers are assigned (if available) as soon as xact's
 * first stream is received for subscriptions that have set their 'streaming'
 * option as parallel. The leader apply worker will send changes to this new
 * worker via shared memory. We keep this worker assigned till the transaction
 * commit is received and also wait for the worker to finish at commit. This
 * preserves commit ordering and avoid file I/O in most cases, although we
 * still need to spill to a file if there is no worker available. See comments
 * atop logical/worker to know more about streamed xacts whose changes are
 * spilled to disk. It is important to maintain commit order to avoid failures
 * due to: (a) transaction dependencies - say if we insert a row in the first
 * transaction and update it in the second transaction on publisher then
 * allowing the subscriber to apply both in parallel can lead to failure in the
 * update; (b) deadlocks - allowing transactions that update the same set of
 * rows/tables in the opposite order to be applied in parallel can lead to
 * deadlocks.
 *
 * A worker pool is used to avoid restarting workers for each streaming
 * transaction. We maintain each worker's information (ParallelApplyWorkerInfo)
 * in the ParallelApplyWorkerPool. After successfully launching a new worker,
 * its information is added to the ParallelApplyWorkerPool. Once the worker
 * finishes applying the transaction, it is marked as available for re-use.
 * Now, before starting a new worker to apply the streaming transaction, we
 * check the list for any available worker. Note that we retain a maximum of
 * half the max_parallel_apply_workers_per_subscription workers in the pool and
 * after that, we simply exit the worker after applying the transaction.
 *
 * XXX This worker pool threshold is arbitrary and we can provide a GUC
 * variable for this in the future if required.
 *
 * The leader apply worker will create a separate dynamic shared memory segment
 * when each parallel apply worker starts. The reason for this design is that
 * we cannot predict how many workers will be needed. It may be possible to
 * allocate enough shared memory in one segment based on the maximum number of
 * parallel apply workers (max_parallel_apply_workers_per_subscription), but
 * this would waste memory if no process is actually started.
 *
 * The dynamic shared memory segment contains: (a) a shm_mq that is used to
 * send changes in the transaction from leader apply worker to parallel apply
 * worker; (b) another shm_mq that is used to send errors (and other messages
 * reported via elog/ereport) from the parallel apply worker to leader apply
 * worker; (c) necessary information to be shared among parallel apply workers
 * and the leader apply worker (i.e. members of ParallelApplyWorkerShared).
 *
 * Locking Considerations
 * ----------------------
 * We have a risk of deadlock due to concurrently applying the transactions in
 * parallel mode that were independent on the publisher side but became
 * dependent on the subscriber side due to the different database structures
 * (like schema of subscription tables, constraints, etc.) on each side. This
 * can happen even without parallel mode when there are concurrent operations
 * on the subscriber. In order to detect the deadlocks among leader (LA) and
 * parallel apply (PA) workers, we used lmgr locks when the PA waits for the
 * next stream (set of changes) and LA waits for PA to finish the transaction.
 * An alternative approach could be to not allow parallelism when the schema of
 * tables is different between the publisher and subscriber but that would be
 * too restrictive and would require the publisher to send much more
 * information than it is currently sending.
 *
 * Consider a case where the subscribed table does not have a unique key on the
 * publisher and has a unique key on the subscriber. The deadlock can happen in
 * the following ways:
 *
 * 1) Deadlock between the leader apply worker and a parallel apply worker
 *
 * Consider that the parallel apply worker (PA) is executing TX-1 and the
 * leader apply worker (LA) is executing TX-2 concurrently on the subscriber.
 * Now, LA is waiting for PA because of the unique key constraint of the
 * subscribed table while PA is waiting for LA to send the next stream of
 * changes or transaction finish command message.
 *
 * In order for lmgr to detect this, we have LA acquire a session lock on the
 * remote transaction (by pa_lock_stream()) and have PA wait on the lock before
 * trying to receive the next stream of changes. Specifically, LA will acquire
 * the lock in AccessExclusive mode before sending the STREAM_STOP and will
 * release it if already acquired after sending the STREAM_START, STREAM_ABORT
 * (for toplevel transaction), STREAM_PREPARE, and STREAM_COMMIT. The PA will
 * acquire the lock in AccessShare mode after processing STREAM_STOP and
 * STREAM_ABORT (for subtransaction) and then release the lock immediately
 * after acquiring it.
 *
 * The lock graph for the above example will look as follows:
 * LA (waiting to acquire the lock on the unique index) -> PA (waiting to
 * acquire the stream lock) -> LA
 *
 * This way, when PA is waiting for LA for the next stream of changes, we can
 * have a wait-edge from PA to LA in lmgr, which will make us detect the
 * deadlock between LA and PA.
 *
 * 2) Deadlock between the leader apply worker and parallel apply workers
 *
 * This scenario is similar to the first case but TX-1 and TX-2 are executed by
 * two parallel apply workers (PA-1 and PA-2 respectively). In this scenario,
 * PA-2 is waiting for PA-1 to complete its transaction while PA-1 is waiting
 * for subsequent input from LA. Also, LA is waiting for PA-2 to complete its
 * transaction in order to preserve the commit order. There is a deadlock among
 * the three processes.
 *
 * In order for lmgr to detect this, we have PA acquire a session lock (this is
 * a different lock than referred in the previous case, see
 * pa_lock_transaction()) on the transaction being applied and have LA wait on
 * the lock before proceeding in the transaction finish commands. Specifically,
 * PA will acquire this lock in AccessExclusive mode before executing the first
 * message of the transaction and release it at the xact end. LA will acquire
 * this lock in AccessShare mode at transaction finish commands and release it
 * immediately.
 *
 * The lock graph for the above example will look as follows:
 * LA (waiting to acquire the transaction lock) -> PA-2 (waiting to acquire the
 * lock due to unique index constraint) -> PA-1 (waiting to acquire the stream
 * lock) -> LA
 *
 * This way when LA is waiting to finish the transaction end command to preserve
 * the commit order, we will be able to detect deadlock, if any.
 *
 * One might think we can use XactLockTableWait(), but XactLockTableWait()
 * considers PREPARED TRANSACTION as still in progress which means the lock
 * won't be released even after the parallel apply worker has prepared the
 * transaction.
 *
 * 3) Deadlock when the shm_mq buffer is full
 *
 * In the previous scenario (ie. PA-1 and PA-2 are executing transactions
 * concurrently), if the shm_mq buffer between LA and PA-2 is full, LA has to
 * wait to send messages, and this wait doesn't appear in lmgr.
 *
 * To avoid this wait, we use a non-blocking write and wait with a timeout. If
 * the timeout is exceeded, the LA will serialize all the pending messages to
 * a file and indicate PA-2 that it needs to read that file for the remaining
 * messages. Then LA will start waiting for commit as in the previous case
 * which will detect deadlock if any. See pa_send_data() and
 * enum TransApplyAction.
 *
 * Lock types
 * ----------
 * Both the stream lock and the transaction lock mentioned above are
 * session-level locks because both locks could be acquired outside the
 * transaction, and the stream lock in the leader needs to persist across
 * transaction boundaries i.e. until the end of the streaming transaction.
 *-------------------------------------------------------------------------
 */
 
#include "postgres.h"
 
#include "libpq/pqformat.h"
#include "libpq/pqmq.h"
#include "pgstat.h"
#include "postmaster/interrupt.h"
#include "replication/logicallauncher.h"
#include "replication/logicalworker.h"
#include "replication/origin.h"
#include "replication/worker_internal.h"
#include "storage/ipc.h"
#include "storage/lmgr.h"
#include "tcop/tcopprot.h"
#include "utils/inval.h"
#include "utils/memutils.h"
#include "utils/syscache.h"
 
#define PG_LOGICAL_APPLY_SHM_MAGIC 0x787ca067
 
/*
 * DSM keys for parallel apply worker. Unlike other parallel execution code,
 * since we don't need to worry about DSM keys conflicting with plan_node_id we
 * can use small integers.
 */
#define PARALLEL_APPLY_KEY_SHARED       1
#define PARALLEL_APPLY_KEY_MQ           2
#define PARALLEL_APPLY_KEY_ERROR_QUEUE  3
 
/* Queue size of DSM, 16 MB for now. */
#define DSM_QUEUE_SIZE  (16 * 1024 * 1024)
 
/*
 * Error queue size of DSM. It is desirable to make it large enough that a
 * typical ErrorResponse can be sent without blocking. That way, a worker that
 * errors out can write the whole message into the queue and terminate without
 * waiting for the user backend.
 */
#define DSM_ERROR_QUEUE_SIZE            (16 * 1024)
 
/*
 * There are three fields in each message received by the parallel apply
 * worker: start_lsn, end_lsn and send_time. Because we have updated these
 * statistics in the leader apply worker, we can ignore these fields in the
 * parallel apply worker (see function LogicalRepApplyLoop).
 */
#define SIZE_STATS_MESSAGE (2 * sizeof(XLogRecPtr) + sizeof(TimestampTz))
 
/*
 * The type of session-level lock on a transaction being applied on a logical
 * replication subscriber.
 */
#define PARALLEL_APPLY_LOCK_STREAM  0
#define PARALLEL_APPLY_LOCK_XACT    1
 
/*
 * Hash table entry to map xid to the parallel apply worker state.
 */
typedef struct ParallelApplyWorkerEntry
{
    TransactionId xid;          /* Hash key -- must be first */
    ParallelApplyWorkerInfo *winfo;
} ParallelApplyWorkerEntry;
 
/*
 * A hash table used to cache the state of streaming transactions being applied
 * by the parallel apply workers.
 */
static HTAB *ParallelApplyTxnHash = NULL;
 
/*
* A list (pool) of active parallel apply workers. The information for
* the new worker is added to the list after successfully launching it. The
* list entry is removed if there are already enough workers in the worker
* pool at the end of the transaction. For more information about the worker
* pool, see comments atop this file.
 */
static List *ParallelApplyWorkerPool = NIL;
 
/*
 * Information shared between leader apply worker and parallel apply worker.
 */
ParallelApplyWorkerShared *MyParallelShared = NULL;
 
/*
 * Is there a message sent by a parallel apply worker that the leader apply
 * worker needs to receive?
 */
volatile sig_atomic_t ParallelApplyMessagePending = false;
 
/*
 * Cache the parallel apply worker information required for applying the
 * current streaming transaction. It is used to save the cost of searching the
 * hash table when applying the changes between STREAM_START and STREAM_STOP.
 */
static ParallelApplyWorkerInfo *stream_apply_worker = NULL;
 
/* A list to maintain subtransactions, if any. */
static List *subxactlist = NIL;
 
static void pa_free_worker_info(ParallelApplyWorkerInfo *winfo);
static ParallelTransState pa_get_xact_state(ParallelApplyWorkerShared *wshared);
static PartialFileSetState pa_get_fileset_state(void);
 
/*
 * Returns true if it is OK to start a parallel apply worker, false otherwise.
 */
static bool
pa_can_start(void)
{
    /* Only leader apply workers can start parallel apply workers. */
    if (!am_leader_apply_worker())
        return false;
 
    /*
     * It is good to check for any change in the subscription parameter to
     * avoid the case where for a very long time the change doesn't get
     * reflected. This can happen when there is a constant flow of streaming
     * transactions that are handled by parallel apply workers.
     *
     * It is better to do it before the below checks so that the latest values
     * of subscription can be used for the checks.
     */
    maybe_reread_subscription();
 
    /*
     * Don't start a new parallel apply worker if the subscription is not
     * using parallel streaming mode, or if the publisher does not support
     * parallel apply.
     */
    if (!MyLogicalRepWorker->parallel_apply)
        return false;
 
    /*
     * Don't start a new parallel worker if user has set skiplsn as it's
     * possible that they want to skip the streaming transaction. For
     * streaming transactions, we need to serialize the transaction to a file
     * so that we can get the last LSN of the transaction to judge whether to
     * skip before starting to apply the change.
     *
     * One might think that we could allow parallelism if the first lsn of the
     * transaction is greater than skiplsn, but we don't send it with the
     * STREAM START message, and it doesn't seem worth sending the extra eight
     * bytes with the STREAM START to enable parallelism for this case.
     */
    if (!XLogRecPtrIsInvalid(MySubscription->skiplsn))
        return false;
 
    /*
     * For streaming transactions that are being applied using a parallel
     * apply worker, we cannot decide whether to apply the change for a
     * relation that is not in the READY state (see
     * should_apply_changes_for_rel) as we won't know remote_final_lsn by that
     * time. So, we don't start the new parallel apply worker in this case.
     */
    if (!AllTablesyncsReady())
        return false;
 
    return true;
}
 
/*
 * Set up a dynamic shared memory segment.
 *
 * We set up a control region that contains a fixed-size worker info
 * (ParallelApplyWorkerShared), a message queue, and an error queue.
 *
 * Returns true on success, false on failure.
 */
static bool
pa_setup_dsm(ParallelApplyWorkerInfo *winfo)
{
    shm_toc_estimator e;
    Size        segsize;
    dsm_segment *seg;
    shm_toc    *toc;
    ParallelApplyWorkerShared *shared;
    shm_mq     *mq;
    Size        queue_size = DSM_QUEUE_SIZE;
    Size        error_queue_size = DSM_ERROR_QUEUE_SIZE;
 
    /*
     * Estimate how much shared memory we need.
     *
     * Because the TOC machinery may choose to insert padding of oddly-sized
     * requests, we must estimate each chunk separately.
     *
     * We need one key to register the location of the header, and two other
     * keys to track the locations of the message queue and the error message
     * queue.
     */
    shm_toc_initialize_estimator(&e);
    shm_toc_estimate_chunk(&e, sizeof(ParallelApplyWorkerShared));
    shm_toc_estimate_chunk(&e, queue_size);
    shm_toc_estimate_chunk(&e, error_queue_size);
 
    shm_toc_estimate_keys(&e, 3);
    segsize = shm_toc_estimate(&e);
 
    /* Create the shared memory segment and establish a table of contents. */
    seg = dsm_create(shm_toc_estimate(&e), 0);
    if (!seg)
        return false;
 
    toc = shm_toc_create(PG_LOGICAL_APPLY_SHM_MAGIC, dsm_segment_address(seg),
                         segsize);
 
    /* Set up the header region. */
    shared = shm_toc_allocate(toc, sizeof(ParallelApplyWorkerShared));
    SpinLockInit(&shared->mutex);
 
    shared->xact_state = PARALLEL_TRANS_UNKNOWN;
    pg_atomic_init_u32(&(shared->pending_stream_count), 0);
    shared->last_commit_end = InvalidXLogRecPtr;
    shared->fileset_state = FS_EMPTY;
 
    shm_toc_insert(toc, PARALLEL_APPLY_KEY_SHARED, shared);
 
    /* Set up message queue for the worker. */
    mq = shm_mq_create(shm_toc_allocate(toc, queue_size), queue_size);
    shm_toc_insert(toc, PARALLEL_APPLY_KEY_MQ, mq);
    shm_mq_set_sender(mq, MyProc);
 
    /* Attach the queue. */
    winfo->mq_handle = shm_mq_attach(mq, seg, NULL);
 
    /* Set up error queue for the worker. */
    mq = shm_mq_create(shm_toc_allocate(toc, error_queue_size),
                       error_queue_size);
    shm_toc_insert(toc, PARALLEL_APPLY_KEY_ERROR_QUEUE, mq);
    shm_mq_set_receiver(mq, MyProc);
 
    /* Attach the queue. */
    winfo->error_mq_handle = shm_mq_attach(mq, seg, NULL);
 
    /* Return results to caller. */
    winfo->dsm_seg = seg;
    winfo->shared = shared;
 
    return true;
}
 
/*
 * Try to get a parallel apply worker from the pool. If none is available then
 * start a new one.
 */
static ParallelApplyWorkerInfo *
pa_launch_parallel_worker(void)
{
    MemoryContext oldcontext;
    bool        launched;
    ParallelApplyWorkerInfo *winfo;
    ListCell   *lc;
 
    /* Try to get an available parallel apply worker from the worker pool. */
    foreach(lc, ParallelApplyWorkerPool)
    {
        winfo = (ParallelApplyWorkerInfo *) lfirst(lc);
 
        if (!winfo->in_use)
            return winfo;
    }
 
    /*
     * Start a new parallel apply worker.
     *
     * The worker info can be used for the lifetime of the worker process, so
     * create it in a permanent context.
     */
    oldcontext = MemoryContextSwitchTo(ApplyContext);
 
    winfo = (ParallelApplyWorkerInfo *) palloc0(sizeof(ParallelApplyWorkerInfo));
 
    /* Setup shared memory. */
    if (!pa_setup_dsm(winfo))
    {
        MemoryContextSwitchTo(oldcontext);
        pfree(winfo);
        return NULL;
    }
 
    launched = logicalrep_worker_launch(WORKERTYPE_PARALLEL_APPLY,
                                        MyLogicalRepWorker->dbid,
                                        MySubscription->oid,
                                        MySubscription->name,
                                        MyLogicalRepWorker->userid,
                                        InvalidOid,
                                        dsm_segment_handle(winfo->dsm_seg));
 
    if (launched)
    {
        ParallelApplyWorkerPool = lappend(ParallelApplyWorkerPool, winfo);
    }
    else
    {
        pa_free_worker_info(winfo);
        winfo = NULL;
    }
 
    MemoryContextSwitchTo(oldcontext);
 
    return winfo;
}
 
/*
 * Allocate a parallel apply worker that will be used for the specified xid.
 *
 * We first try to get an available worker from the pool, if any and then try
 * to launch a new worker. On successful allocation, remember the worker
 * information in the hash table so that we can get it later for processing the
 * streaming changes.
 */
void
pa_allocate_worker(TransactionId xid)
{
    bool        found;
    ParallelApplyWorkerInfo *winfo = NULL;
    ParallelApplyWorkerEntry *entry;
 
    if (!pa_can_start())
        return;
 
    winfo = pa_launch_parallel_worker();
    if (!winfo)
        return;
 
    /* First time through, initialize parallel apply worker state hashtable. */
    if (!ParallelApplyTxnHash)
    {
        HASHCTL     ctl;
 
        MemSet(&ctl, 0, sizeof(ctl));
        ctl.keysize = sizeof(TransactionId);
        ctl.entrysize = sizeof(ParallelApplyWorkerEntry);
        ctl.hcxt = ApplyContext;
 
        ParallelApplyTxnHash = hash_create("logical replication parallel apply workers hash",
                                           16, &ctl,
                                           HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
    }
 
    /* Create an entry for the requested transaction. */
    entry = hash_search(ParallelApplyTxnHash, &xid, HASH_ENTER, &found);
    if (found)
        elog(ERROR, "hash table corrupted");
 
    /* Update the transaction information in shared memory. */
    SpinLockAcquire(&winfo->shared->mutex);
    winfo->shared->xact_state = PARALLEL_TRANS_UNKNOWN;
    winfo->shared->xid = xid;
    SpinLockRelease(&winfo->shared->mutex);
 
    winfo->in_use = true;
    winfo->serialize_changes = false;
    entry->winfo = winfo;
    entry->xid = xid;
}
 
/*
 * Find the assigned worker for the given transaction, if any.
 */
ParallelApplyWorkerInfo *
pa_find_worker(TransactionId xid)
{
    bool        found;
    ParallelApplyWorkerEntry *entry;
 
    if (!TransactionIdIsValid(xid))
        return NULL;
 
    if (!ParallelApplyTxnHash)
        return NULL;
 
    /* Return the cached parallel apply worker if valid. */
    if (stream_apply_worker)
        return stream_apply_worker;
 
    /* Find an entry for the requested transaction. */
    entry = hash_search(ParallelApplyTxnHash, &xid, HASH_FIND, &found);
    if (found)
    {
        /* The worker must not have exited.  */
        Assert(entry->winfo->in_use);
        return entry->winfo;
    }
 
    return NULL;
}
 
/*
 * Makes the worker available for reuse.
 *
 * This removes the parallel apply worker entry from the hash table so that it
 * can't be used. If there are enough workers in the pool, it stops the worker
 * and frees the corresponding info. Otherwise it just marks the worker as
 * available for reuse.
 *
 * For more information about the worker pool, see comments atop this file.
 */
static void
pa_free_worker(ParallelApplyWorkerInfo *winfo)
{
    Assert(!am_parallel_apply_worker());
    Assert(winfo->in_use);
    Assert(pa_get_xact_state(winfo->shared) == PARALLEL_TRANS_FINISHED);
 
    if (!hash_search(ParallelApplyTxnHash, &winfo->shared->xid, HASH_REMOVE, NULL))
        elog(ERROR, "hash table corrupted");
 
    /*
     * Stop the worker if there are enough workers in the pool.
     *
     * XXX Additionally, we also stop the worker if the leader apply worker
     * serialize part of the transaction data due to a send timeout. This is
     * because the message could be partially written to the queue and there
     * is no way to clean the queue other than resending the message until it
     * succeeds. Instead of trying to send the data which anyway would have
     * been serialized and then letting the parallel apply worker deal with
     * the spurious message, we stop the worker.
     */
    if (winfo->serialize_changes ||
        list_length(ParallelApplyWorkerPool) >
        (max_parallel_apply_workers_per_subscription / 2))
    {
        logicalrep_pa_worker_stop(winfo);
        pa_free_worker_info(winfo);
 
        return;
    }
 
    winfo->in_use = false;
    winfo->serialize_changes = false;
}
 
/*
 * Free the parallel apply worker information and unlink the files with
 * serialized changes if any.
 */
static void
pa_free_worker_info(ParallelApplyWorkerInfo *winfo)
{
    Assert(winfo);
 
    if (winfo->mq_handle)
        shm_mq_detach(winfo->mq_handle);
 
    if (winfo->error_mq_handle)
        shm_mq_detach(winfo->error_mq_handle);
 
    /* Unlink the files with serialized changes. */
    if (winfo->serialize_changes)
        stream_cleanup_files(MyLogicalRepWorker->subid, winfo->shared->xid);
 
    if (winfo->dsm_seg)
        dsm_detach(winfo->dsm_seg);
 
    /* Remove from the worker pool. */
    ParallelApplyWorkerPool = list_delete_ptr(ParallelApplyWorkerPool, winfo);
 
    pfree(winfo);
}
 
/*
 * Detach the error queue for all parallel apply workers.
 */
void
pa_detach_all_error_mq(void)
{
    ListCell   *lc;
 
    foreach(lc, ParallelApplyWorkerPool)
    {
        ParallelApplyWorkerInfo *winfo = (ParallelApplyWorkerInfo *) lfirst(lc);
 
        if (winfo->error_mq_handle)
        {
            shm_mq_detach(winfo->error_mq_handle);
            winfo->error_mq_handle = NULL;
        }
    }
}
 
/*
 * Check if there are any pending spooled messages.
 */
static bool
pa_has_spooled_message_pending()
{
    PartialFileSetState fileset_state;
 
    fileset_state = pa_get_fileset_state();
 
    return (fileset_state != FS_EMPTY);
}
 
/*
 * Replay the spooled messages once the leader apply worker has finished
 * serializing changes to the file.
 *
 * Returns false if there aren't any pending spooled messages, true otherwise.
 */
static bool
pa_process_spooled_messages_if_required(void)
{
    PartialFileSetState fileset_state;
 
    fileset_state = pa_get_fileset_state();
 
    if (fileset_state == FS_EMPTY)
        return false;
 
    /*
     * If the leader apply worker is busy serializing the partial changes then
     * acquire the stream lock now and wait for the leader worker to finish
     * serializing the changes. Otherwise, the parallel apply worker won't get
     * a chance to receive a STREAM_STOP (and acquire the stream lock) until
     * the leader had serialized all changes which can lead to undetected
     * deadlock.
     *
     * Note that the fileset state can be FS_SERIALIZE_DONE once the leader
     * worker has finished serializing the changes.
     */
    if (fileset_state == FS_SERIALIZE_IN_PROGRESS)
    {
        pa_lock_stream(MyParallelShared->xid, AccessShareLock);
        pa_unlock_stream(MyParallelShared->xid, AccessShareLock);
 
        fileset_state = pa_get_fileset_state();
    }
 
    /*
     * We cannot read the file immediately after the leader has serialized all
     * changes to the file because there may still be messages in the memory
     * queue. We will apply all spooled messages the next time we call this
     * function and that will ensure there are no messages left in the memory
     * queue.
     */
    if (fileset_state == FS_SERIALIZE_DONE)
    {
        pa_set_fileset_state(MyParallelShared, FS_READY);
    }
    else if (fileset_state == FS_READY)
    {
        apply_spooled_messages(&MyParallelShared->fileset,
                               MyParallelShared->xid,
                               InvalidXLogRecPtr);
        pa_set_fileset_state(MyParallelShared, FS_EMPTY);
    }
 
    return true;
}
 
/*
 * Interrupt handler for main loop of parallel apply worker.
 */
static void
ProcessParallelApplyInterrupts(void)
{
    CHECK_FOR_INTERRUPTS();
 
    if (ShutdownRequestPending)
    {
        ereport(LOG,
                (errmsg("logical replication parallel apply worker for subscription \"%s\" has finished",
                        MySubscription->name)));
 
        proc_exit(0);
    }
 
    if (ConfigReloadPending)
    {
        ConfigReloadPending = false;
        ProcessConfigFile(PGC_SIGHUP);
    }
}
 
/* Parallel apply worker main loop. */
static void
LogicalParallelApplyLoop(shm_mq_handle *mqh)
{
    shm_mq_result shmq_res;
    ErrorContextCallback errcallback;
    MemoryContext oldcxt = CurrentMemoryContext;
 
    /*
     * Init the ApplyMessageContext which we clean up after each replication
     * protocol message.
     */
    ApplyMessageContext = AllocSetContextCreate(ApplyContext,
                                                "ApplyMessageContext",
                                                ALLOCSET_DEFAULT_SIZES);
 
    /*
     * Push apply error context callback. Fields will be filled while applying
     * a change.
     */
    errcallback.callback = apply_error_callback;
    errcallback.previous = error_context_stack;
    error_context_stack = &errcallback;
 
    for (;;)
    {
        void       *data;
        Size        len;
 
        ProcessParallelApplyInterrupts();
 
        /* Ensure we are reading the data into our memory context. */
        MemoryContextSwitchTo(ApplyMessageContext);
 
        shmq_res = shm_mq_receive(mqh, &len, &data, true);
 
        if (shmq_res == SHM_MQ_SUCCESS)
        {
            StringInfoData s;
            int         c;
 
            if (len == 0)
                elog(ERROR, "invalid message length");
 
            s.cursor = 0;
            s.maxlen = -1;
            s.data = (char *) data;
            s.len = len;
 
            /*
             * The first byte of messages sent from leader apply worker to
             * parallel apply workers can only be 'w'.
             */
            c = pq_getmsgbyte(&s);
            if (c != 'w')
                elog(ERROR, "unexpected message \"%c\"", c);
 
            /*
             * Ignore statistics fields that have been updated by the leader
             * apply worker.
             *
             * XXX We can avoid sending the statistics fields from the leader
             * apply worker but for that, it needs to rebuild the entire
             * message by removing these fields which could be more work than
             * simply ignoring these fields in the parallel apply worker.
             */
            s.cursor += SIZE_STATS_MESSAGE;
 
            apply_dispatch(&s);
        }
        else if (shmq_res == SHM_MQ_WOULD_BLOCK)
        {
            /* Replay the changes from the file, if any. */
            if (!pa_process_spooled_messages_if_required())
            {
                int         rc;
 
                /* Wait for more work. */
                rc = WaitLatch(MyLatch,
                               WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
                               1000L,
                               WAIT_EVENT_LOGICAL_PARALLEL_APPLY_MAIN);
 
                if (rc & WL_LATCH_SET)
                    ResetLatch(MyLatch);
            }
        }
        else
        {
            Assert(shmq_res == SHM_MQ_DETACHED);
 
            ereport(ERROR,
                    (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                     errmsg("lost connection to the logical replication apply worker")));
        }
 
        MemoryContextReset(ApplyMessageContext);
        MemoryContextSwitchTo(oldcxt);
    }
 
    /* Pop the error context stack. */
    error_context_stack = errcallback.previous;
 
    MemoryContextSwitchTo(oldcxt);
}
 
/*
 * Make sure the leader apply worker tries to read from our error queue one more
 * time. This guards against the case where we exit uncleanly without sending
 * an ErrorResponse, for example because some code calls proc_exit directly.
 *
 * Also explicitly detach from dsm segment to invoke on_dsm_detach callbacks,
 * if any. See ParallelWorkerShutdown for details.
 */
static void
pa_shutdown(int code, Datum arg)
{
    SendProcSignal(MyLogicalRepWorker->leader_pid,
                   PROCSIG_PARALLEL_APPLY_MESSAGE,
                   InvalidBackendId);
 
    dsm_detach((dsm_segment *) DatumGetPointer(arg));
}
 
/*
 * Parallel apply worker entry point.
 */
void
ParallelApplyWorkerMain(Datum main_arg)
{
    ParallelApplyWorkerShared *shared;
    dsm_handle  handle;
    dsm_segment *seg;
    shm_toc    *toc;
    shm_mq     *mq;
    shm_mq_handle *mqh;
    shm_mq_handle *error_mqh;
    RepOriginId originid;
    int         worker_slot = DatumGetInt32(main_arg);
    char        originname[NAMEDATALEN];
 
    InitializingApplyWorker = true;
 
    /* Setup signal handling. */
    pqsignal(SIGHUP, SignalHandlerForConfigReload);
    pqsignal(SIGINT, SignalHandlerForShutdownRequest);
    pqsignal(SIGTERM, die);
    BackgroundWorkerUnblockSignals();
 
    /*
     * Attach to the dynamic shared memory segment for the parallel apply, and
     * find its table of contents.
     *
     * Like parallel query, we don't need resource owner by this time. See
     * ParallelWorkerMain.
     */
    memcpy(&handle, MyBgworkerEntry->bgw_extra, sizeof(dsm_handle));
    seg = dsm_attach(handle);
    if (!seg)
        ereport(ERROR,
                (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                 errmsg("could not map dynamic shared memory segment")));
 
    toc = shm_toc_attach(PG_LOGICAL_APPLY_SHM_MAGIC, dsm_segment_address(seg));
    if (!toc)
        ereport(ERROR,
                (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                 errmsg("invalid magic number in dynamic shared memory segment")));
 
    /* Look up the shared information. */
    shared = shm_toc_lookup(toc, PARALLEL_APPLY_KEY_SHARED, false);
    MyParallelShared = shared;
 
    /*
     * Attach to the message queue.
     */
    mq = shm_toc_lookup(toc, PARALLEL_APPLY_KEY_MQ, false);
    shm_mq_set_receiver(mq, MyProc);
    mqh = shm_mq_attach(mq, seg, NULL);
 
    /*
     * Primary initialization is complete. Now, we can attach to our slot.
     * This is to ensure that the leader apply worker does not write data to
     * the uninitialized memory queue.
     */
    logicalrep_worker_attach(worker_slot);
 
    /*
     * Register the shutdown callback after we are attached to the worker
     * slot. This is to ensure that MyLogicalRepWorker remains valid when this
     * callback is invoked.
     */
    before_shmem_exit(pa_shutdown, PointerGetDatum(seg));
 
    SpinLockAcquire(&MyParallelShared->mutex);
    MyParallelShared->logicalrep_worker_generation = MyLogicalRepWorker->generation;
    MyParallelShared->logicalrep_worker_slot_no = worker_slot;
    SpinLockRelease(&MyParallelShared->mutex);
 
    /*
     * Attach to the error queue.
     */
    mq = shm_toc_lookup(toc, PARALLEL_APPLY_KEY_ERROR_QUEUE, false);
    shm_mq_set_sender(mq, MyProc);
    error_mqh = shm_mq_attach(mq, seg, NULL);
 
    pq_redirect_to_shm_mq(seg, error_mqh);
    pq_set_parallel_leader(MyLogicalRepWorker->leader_pid,
                           InvalidBackendId);
 
    MyLogicalRepWorker->last_send_time = MyLogicalRepWorker->last_recv_time =
        MyLogicalRepWorker->reply_time = 0;
 
    InitializeLogRepWorker();
 
    InitializingApplyWorker = false;
 
    /* Setup replication origin tracking. */
    StartTransactionCommand();
    ReplicationOriginNameForLogicalRep(MySubscription->oid, InvalidOid,
                                       originname, sizeof(originname));
    originid = replorigin_by_name(originname, false);
 
    /*
     * The parallel apply worker doesn't need to monopolize this replication
     * origin which was already acquired by its leader process.
     */
    replorigin_session_setup(originid, MyLogicalRepWorker->leader_pid);
    replorigin_session_origin = originid;
    CommitTransactionCommand();
 
    /*
     * Setup callback for syscache so that we know when something changes in
     * the subscription relation state.
     */
    CacheRegisterSyscacheCallback(SUBSCRIPTIONRELMAP,
                                  invalidate_syncing_table_states,
                                  (Datum) 0);
 
    set_apply_error_context_origin(originname);
 
    LogicalParallelApplyLoop(mqh);
 
    /*
     * The parallel apply worker must not get here because the parallel apply
     * worker will only stop when it receives a SIGTERM or SIGINT from the
     * leader, or when there is an error. None of these cases will allow the
     * code to reach here.
     */
    Assert(false);
}
 
/*
 * Handle receipt of an interrupt indicating a parallel apply worker message.
 *
 * Note: this is called within a signal handler! All we can do is set a flag
 * that will cause the next CHECK_FOR_INTERRUPTS() to invoke
 * HandleParallelApplyMessages().
 */
void
HandleParallelApplyMessageInterrupt(void)
{
    InterruptPending = true;
    ParallelApplyMessagePending = true;
    SetLatch(MyLatch);
}
 
/*
 * Handle a single protocol message received from a single parallel apply
 * worker.
 */
static void
HandleParallelApplyMessage(StringInfo msg)
{
    char        msgtype;
 
    msgtype = pq_getmsgbyte(msg);
 
    switch (msgtype)
    {
        case 'E':               /* ErrorResponse */
            {
                ErrorData   edata;
 
                /* Parse ErrorResponse. */
                pq_parse_errornotice(msg, &edata);
 
                /*
                 * If desired, add a context line to show that this is a
                 * message propagated from a parallel apply worker. Otherwise,
                 * it can sometimes be confusing to understand what actually
                 * happened.
                 */
                if (edata.context)
                    edata.context = psprintf("%s\n%s", edata.context,
                                             _("logical replication parallel apply worker"));
                else
                    edata.context = pstrdup(_("logical replication parallel apply worker"));
 
                /*
                 * Context beyond that should use the error context callbacks
                 * that were in effect in LogicalRepApplyLoop().
                 */
                error_context_stack = apply_error_context_stack;
 
                /*
                 * The actual error must have been reported by the parallel
                 * apply worker.
                 */
                ereport(ERROR,
                        (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                         errmsg("logical replication parallel apply worker exited due to error"),
                         errcontext("%s", edata.context)));
            }
 
            /*
             * Don't need to do anything about NoticeResponse and
             * NotifyResponse as the logical replication worker doesn't need
             * to send messages to the client.
             */
        case 'N':
        case 'A':
            break;
 
        default:
            elog(ERROR, "unrecognized message type received from logical replication parallel apply worker: %c (message length %d bytes)",
                 msgtype, msg->len);
    }
}
 
/*
 * Handle any queued protocol messages received from parallel apply workers.
 */
void
HandleParallelApplyMessages(void)
{
    ListCell   *lc;
    MemoryContext oldcontext;
 
    static MemoryContext hpam_context = NULL;
 
    /*
     * This is invoked from ProcessInterrupts(), and since some of the
     * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
     * for recursive calls if more signals are received while this runs. It's
     * unclear that recursive entry would be safe, and it doesn't seem useful
     * even if it is safe, so let's block interrupts until done.
     */
    HOLD_INTERRUPTS();
 
    /*
     * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
     * don't want to risk leaking data into long-lived contexts, so let's do
     * our work here in a private context that we can reset on each use.
     */
    if (!hpam_context)          /* first time through? */
        hpam_context = AllocSetContextCreate(TopMemoryContext,
                                             "HandleParallelApplyMessages",
                                             ALLOCSET_DEFAULT_SIZES);
    else
        MemoryContextReset(hpam_context);
 
    oldcontext = MemoryContextSwitchTo(hpam_context);
 
    ParallelApplyMessagePending = false;
 
    foreach(lc, ParallelApplyWorkerPool)
    {
        shm_mq_result res;
        Size        nbytes;
        void       *data;
        ParallelApplyWorkerInfo *winfo = (ParallelApplyWorkerInfo *) lfirst(lc);
 
        /*
         * The leader will detach from the error queue and set it to NULL
         * before preparing to stop all parallel apply workers, so we don't
         * need to handle error messages anymore. See
         * logicalrep_worker_detach.
         */
        if (!winfo->error_mq_handle)
            continue;
 
        res = shm_mq_receive(winfo->error_mq_handle, &nbytes, &data, true);
 
        if (res == SHM_MQ_WOULD_BLOCK)
            continue;
        else if (res == SHM_MQ_SUCCESS)
        {
            StringInfoData msg;
 
            initStringInfo(&msg);
            appendBinaryStringInfo(&msg, data, nbytes);
            HandleParallelApplyMessage(&msg);
            pfree(msg.data);
        }
        else
            ereport(ERROR,
                    (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                     errmsg("lost connection to the logical replication parallel apply worker")));
    }
 
    MemoryContextSwitchTo(oldcontext);
 
    /* Might as well clear the context on our way out */
    MemoryContextReset(hpam_context);
 
    RESUME_INTERRUPTS();
}
 
/*
 * Send the data to the specified parallel apply worker via shared-memory
 * queue.
 *
 * Returns false if the attempt to send data via shared memory times out, true
 * otherwise.
 */
bool
pa_send_data(ParallelApplyWorkerInfo *winfo, Size nbytes, const void *data)
{
    int         rc;
    shm_mq_result result;
    TimestampTz startTime = 0;
 
    Assert(!IsTransactionState());
    Assert(!winfo->serialize_changes);
 
    /*
     * We don't try to send data to parallel worker for 'immediate' mode. This
     * is primarily used for testing purposes.
     */
    if (unlikely(debug_logical_replication_streaming == DEBUG_LOGICAL_REP_STREAMING_IMMEDIATE))
        return false;
 
/*
 * This timeout is a bit arbitrary but testing revealed that it is sufficient
 * to send the message unless the parallel apply worker is waiting on some
 * lock or there is a serious resource crunch. See the comments atop this file
 * to know why we are using a non-blocking way to send the message.
 */
#define SHM_SEND_RETRY_INTERVAL_MS 1000
#define SHM_SEND_TIMEOUT_MS     (10000 - SHM_SEND_RETRY_INTERVAL_MS)
 
    for (;;)
    {
        result = shm_mq_send(winfo->mq_handle, nbytes, data, true, true);
 
        if (result == SHM_MQ_SUCCESS)
            return true;
        else if (result == SHM_MQ_DETACHED)
            ereport(ERROR,
                    (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                     errmsg("could not send data to shared-memory queue")));
 
        Assert(result == SHM_MQ_WOULD_BLOCK);
 
        /* Wait before retrying. */
        rc = WaitLatch(MyLatch,
                       WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
                       SHM_SEND_RETRY_INTERVAL_MS,
                       WAIT_EVENT_LOGICAL_APPLY_SEND_DATA);
 
        if (rc & WL_LATCH_SET)
        {
            ResetLatch(MyLatch);
            CHECK_FOR_INTERRUPTS();
        }
 
        if (startTime == 0)
            startTime = GetCurrentTimestamp();
        else if (TimestampDifferenceExceeds(startTime, GetCurrentTimestamp(),
                                            SHM_SEND_TIMEOUT_MS))
            return false;
    }
}
 
/*
 * Switch to PARTIAL_SERIALIZE mode for the current transaction -- this means
 * that the current data and any subsequent data for this transaction will be
 * serialized to a file. This is done to prevent possible deadlocks with
 * another parallel apply worker (refer to the comments atop this file).
 */
void
pa_switch_to_partial_serialize(ParallelApplyWorkerInfo *winfo,
                               bool stream_locked)
{
    ereport(LOG,
            (errmsg("logical replication apply worker will serialize the remaining changes of remote transaction %u to a file",
                    winfo->shared->xid)));
 
    /*
     * The parallel apply worker could be stuck for some reason (say waiting
     * on some lock by other backend), so stop trying to send data directly to
     * it and start serializing data to the file instead.
     */
    winfo->serialize_changes = true;
 
    /* Initialize the stream fileset. */
    stream_start_internal(winfo->shared->xid, true);
 
    /*
     * Acquires the stream lock if not already to make sure that the parallel
     * apply worker will wait for the leader to release the stream lock until
     * the end of the transaction.
     */
    if (!stream_locked)
        pa_lock_stream(winfo->shared->xid, AccessExclusiveLock);
 
    pa_set_fileset_state(winfo->shared, FS_SERIALIZE_IN_PROGRESS);
}
 
/*
 * Wait until the parallel apply worker's transaction state has reached or
 * exceeded the given xact_state.
 */
static void
pa_wait_for_xact_state(ParallelApplyWorkerInfo *winfo,
                       ParallelTransState xact_state)
{
    for (;;)
    {
        /*
         * Stop if the transaction state has reached or exceeded the given
         * xact_state.
         */
        if (pa_get_xact_state(winfo->shared) >= xact_state)
            break;
 
        /* Wait to be signalled. */
        (void) WaitLatch(MyLatch,
                         WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
                         10L,
                         WAIT_EVENT_LOGICAL_PARALLEL_APPLY_STATE_CHANGE);
 
        /* Reset the latch so we don't spin. */
        ResetLatch(MyLatch);
 
        /* An interrupt may have occurred while we were waiting. */
        CHECK_FOR_INTERRUPTS();
    }
}
 
/*
 * Wait until the parallel apply worker's transaction finishes.
 */
static void
pa_wait_for_xact_finish(ParallelApplyWorkerInfo *winfo)
{
    /*
     * Wait until the parallel apply worker set the state to
     * PARALLEL_TRANS_STARTED which means it has acquired the transaction
     * lock. This is to prevent leader apply worker from acquiring the
     * transaction lock earlier than the parallel apply worker.
     */
    pa_wait_for_xact_state(winfo, PARALLEL_TRANS_STARTED);
 
    /*
     * Wait for the transaction lock to be released. This is required to
     * detect deadlock among leader and parallel apply workers. Refer to the
     * comments atop this file.
     */
    pa_lock_transaction(winfo->shared->xid, AccessShareLock);
    pa_unlock_transaction(winfo->shared->xid, AccessShareLock);
 
    /*
     * Check if the state becomes PARALLEL_TRANS_FINISHED in case the parallel
     * apply worker failed while applying changes causing the lock to be
     * released.
     */
    if (pa_get_xact_state(winfo->shared) != PARALLEL_TRANS_FINISHED)
        ereport(ERROR,
                (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                 errmsg("lost connection to the logical replication parallel apply worker")));
}
 
/*
 * Set the transaction state for a given parallel apply worker.
 */
void
pa_set_xact_state(ParallelApplyWorkerShared *wshared,
                  ParallelTransState xact_state)
{
    SpinLockAcquire(&wshared->mutex);
    wshared->xact_state = xact_state;
    SpinLockRelease(&wshared->mutex);
}
 
/*
 * Get the transaction state for a given parallel apply worker.
 */
static ParallelTransState
pa_get_xact_state(ParallelApplyWorkerShared *wshared)
{
    ParallelTransState xact_state;
 
    SpinLockAcquire(&wshared->mutex);
    xact_state = wshared->xact_state;
    SpinLockRelease(&wshared->mutex);
 
    return xact_state;
}
 
/*
 * Cache the parallel apply worker information.
 */
void
pa_set_stream_apply_worker(ParallelApplyWorkerInfo *winfo)
{
    stream_apply_worker = winfo;
}
 
/*
 * Form a unique savepoint name for the streaming transaction.
 *
 * Note that different subscriptions for publications on different nodes can
 * receive same remote xid, so we need to use subscription id along with it.
 *
 * Returns the name in the supplied buffer.
 */
static void
pa_savepoint_name(Oid suboid, TransactionId xid, char *spname, Size szsp)
{
    snprintf(spname, szsp, "pg_sp_%u_%u", suboid, xid);
}
 
/*
 * Define a savepoint for a subxact in parallel apply worker if needed.
 *
 * The parallel apply worker can figure out if a new subtransaction was
 * started by checking if the new change arrived with a different xid. In that
 * case define a named savepoint, so that we are able to rollback to it
 * if required.
 */
void
pa_start_subtrans(TransactionId current_xid, TransactionId top_xid)
{
    if (current_xid != top_xid &&
        !list_member_xid(subxactlist, current_xid))
    {
        MemoryContext oldctx;
        char        spname[NAMEDATALEN];
 
        pa_savepoint_name(MySubscription->oid, current_xid,
                          spname, sizeof(spname));
 
        elog(DEBUG1, "defining savepoint %s in logical replication parallel apply worker", spname);
 
        /* We must be in transaction block to define the SAVEPOINT. */
        if (!IsTransactionBlock())
        {
            if (!IsTransactionState())
                StartTransactionCommand();
 
            BeginTransactionBlock();
            CommitTransactionCommand();
        }
 
        DefineSavepoint(spname);
 
        /*
         * CommitTransactionCommand is needed to start a subtransaction after
         * issuing a SAVEPOINT inside a transaction block (see
         * StartSubTransaction()).
         */
        CommitTransactionCommand();
 
        oldctx = MemoryContextSwitchTo(TopTransactionContext);
        subxactlist = lappend_xid(subxactlist, current_xid);
        MemoryContextSwitchTo(oldctx);
    }
}
 
/* Reset the list that maintains subtransactions. */
void
pa_reset_subtrans(void)
{
    /*
     * We don't need to free this explicitly as the allocated memory will be
     * freed at the transaction end.
     */
    subxactlist = NIL;
}
 
/*
 * Handle STREAM ABORT message when the transaction was applied in a parallel
 * apply worker.
 */
void
pa_stream_abort(LogicalRepStreamAbortData *abort_data)
{
    TransactionId xid = abort_data->xid;
    TransactionId subxid = abort_data->subxid;
 
    /*
     * Update origin state so we can restart streaming from correct position
     * in case of crash.
     */
    replorigin_session_origin_lsn = abort_data->abort_lsn;
    replorigin_session_origin_timestamp = abort_data->abort_time;
 
    /*
     * If the two XIDs are the same, it's in fact abort of toplevel xact, so
     * just free the subxactlist.
     */
    if (subxid == xid)
    {
        pa_set_xact_state(MyParallelShared, PARALLEL_TRANS_FINISHED);
 
        /*
         * Release the lock as we might be processing an empty streaming
         * transaction in which case the lock won't be released during
         * transaction rollback.
         *
         * Note that it's ok to release the transaction lock before aborting
         * the transaction because even if the parallel apply worker dies due
         * to crash or some other reason, such a transaction would still be
         * considered aborted.
         */
        pa_unlock_transaction(xid, AccessExclusiveLock);
 
        AbortCurrentTransaction();
 
        if (IsTransactionBlock())
        {
            EndTransactionBlock(false);
            CommitTransactionCommand();
        }
 
        pa_reset_subtrans();
 
        pgstat_report_activity(STATE_IDLE, NULL);
    }
    else
    {
        /* OK, so it's a subxact. Rollback to the savepoint. */
        int         i;
        char        spname[NAMEDATALEN];
 
        pa_savepoint_name(MySubscription->oid, subxid, spname, sizeof(spname));
 
        elog(DEBUG1, "rolling back to savepoint %s in logical replication parallel apply worker", spname);
 
        /*
         * Search the subxactlist, determine the offset tracked for the
         * subxact, and truncate the list.
         *
         * Note that for an empty sub-transaction we won't find the subxid
         * here.
         */
        for (i = list_length(subxactlist) - 1; i >= 0; i--)
        {
            TransactionId xid_tmp = lfirst_xid(list_nth_cell(subxactlist, i));
 
            if (xid_tmp == subxid)
            {
                RollbackToSavepoint(spname);
                CommitTransactionCommand();
                subxactlist = list_truncate(subxactlist, i);
                break;
            }
        }
    }
}
 
/*
 * Set the fileset state for a particular parallel apply worker. The fileset
 * will be set once the leader worker serialized all changes to the file
 * so that it can be used by parallel apply worker.
 */
void
pa_set_fileset_state(ParallelApplyWorkerShared *wshared,
                     PartialFileSetState fileset_state)
{
    SpinLockAcquire(&wshared->mutex);
    wshared->fileset_state = fileset_state;
 
    if (fileset_state == FS_SERIALIZE_DONE)
    {
        Assert(am_leader_apply_worker());
        Assert(MyLogicalRepWorker->stream_fileset);
        wshared->fileset = *MyLogicalRepWorker->stream_fileset;
    }
 
    SpinLockRelease(&wshared->mutex);
}
 
/*
 * Get the fileset state for the current parallel apply worker.
 */
static PartialFileSetState
pa_get_fileset_state(void)
{
    PartialFileSetState fileset_state;
 
    Assert(am_parallel_apply_worker());
 
    SpinLockAcquire(&MyParallelShared->mutex);
    fileset_state = MyParallelShared->fileset_state;
    SpinLockRelease(&MyParallelShared->mutex);
 
    return fileset_state;
}
 
/*
 * Helper functions to acquire and release a lock for each stream block.
 *
 * Set locktag_field4 to PARALLEL_APPLY_LOCK_STREAM to indicate that it's a
 * stream lock.
 *
 * Refer to the comments atop this file to see how the stream lock is used.
 */
void
pa_lock_stream(TransactionId xid, LOCKMODE lockmode)
{
    LockApplyTransactionForSession(MyLogicalRepWorker->subid, xid,
                                   PARALLEL_APPLY_LOCK_STREAM, lockmode);
}
 
void
pa_unlock_stream(TransactionId xid, LOCKMODE lockmode)
{
    UnlockApplyTransactionForSession(MyLogicalRepWorker->subid, xid,
                                     PARALLEL_APPLY_LOCK_STREAM, lockmode);
}
 
/*
 * Helper functions to acquire and release a lock for each local transaction
 * apply.
 *
 * Set locktag_field4 to PARALLEL_APPLY_LOCK_XACT to indicate that it's a
 * transaction lock.
 *
 * Note that all the callers must pass a remote transaction ID instead of a
 * local transaction ID as xid. This is because the local transaction ID will
 * only be assigned while applying the first change in the parallel apply but
 * it's possible that the first change in the parallel apply worker is blocked
 * by a concurrently executing transaction in another parallel apply worker. We
 * can only communicate the local transaction id to the leader after applying
 * the first change so it won't be able to wait after sending the xact finish
 * command using this lock.
 *
 * Refer to the comments atop this file to see how the transaction lock is
 * used.
 */
void
pa_lock_transaction(TransactionId xid, LOCKMODE lockmode)
{
    LockApplyTransactionForSession(MyLogicalRepWorker->subid, xid,
                                   PARALLEL_APPLY_LOCK_XACT, lockmode);
}
 
void
pa_unlock_transaction(TransactionId xid, LOCKMODE lockmode)
{
    UnlockApplyTransactionForSession(MyLogicalRepWorker->subid, xid,
                                     PARALLEL_APPLY_LOCK_XACT, lockmode);
}
 
/*
 * Decrement the number of pending streaming blocks and wait on the stream lock
 * if there is no pending block available.
 */
void
pa_decr_and_wait_stream_block(void)
{
    Assert(am_parallel_apply_worker());
 
    /*
     * It is only possible to not have any pending stream chunks when we are
     * applying spooled messages.
     */
    if (pg_atomic_read_u32(&MyParallelShared->pending_stream_count) == 0)
    {
        if (pa_has_spooled_message_pending())
            return;
 
        elog(ERROR, "invalid pending streaming chunk 0");
    }
 
    if (pg_atomic_sub_fetch_u32(&MyParallelShared->pending_stream_count, 1) == 0)
    {
        pa_lock_stream(MyParallelShared->xid, AccessShareLock);
        pa_unlock_stream(MyParallelShared->xid, AccessShareLock);
    }
}
 
/*
 * Finish processing the streaming transaction in the leader apply worker.
 */
void
pa_xact_finish(ParallelApplyWorkerInfo *winfo, XLogRecPtr remote_lsn)
{
    Assert(am_leader_apply_worker());
 
    /*
     * Unlock the shared object lock so that parallel apply worker can
     * continue to receive and apply changes.
     */
    pa_unlock_stream(winfo->shared->xid, AccessExclusiveLock);
 
    /*
     * Wait for that worker to finish. This is necessary to maintain commit
     * order which avoids failures due to transaction dependencies and
     * deadlocks.
     */
    pa_wait_for_xact_finish(winfo);
 
    if (!XLogRecPtrIsInvalid(remote_lsn))
        store_flush_position(remote_lsn, winfo->shared->last_commit_end);
 
    pa_free_worker(winfo);
}