parallel.c

Go to the documentation of this file.

/*-------------------------------------------------------------------------
 *
 * parallel.c
 *
 *  Parallel support for pg_dump and pg_restore
 *
 * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
 * Portions Copyright (c) 1994, Regents of the University of California
 *
 * IDENTIFICATION
 *      src/bin/pg_dump/parallel.c
 *
 *-------------------------------------------------------------------------
 */

/*
 * Parallel operation works like this:
 *
 * The original, leader process calls ParallelBackupStart(), which forks off
 * the desired number of worker processes, which each enter WaitForCommands().
 *
 * The leader process dispatches an individual work item to one of the worker
 * processes in DispatchJobForTocEntry().  We send a command string such as
 * "DUMP 1234" or "RESTORE 1234", where 1234 is the TocEntry ID.
 * The worker process receives and decodes the command and passes it to the
 * routine pointed to by AH->WorkerJobDumpPtr or AH->WorkerJobRestorePtr,
 * which are routines of the current archive format.  That routine performs
 * the required action (dump or restore) and returns an integer status code.
 * This is passed back to the leader where we pass it to the
 * ParallelCompletionPtr callback function that was passed to
 * DispatchJobForTocEntry().  The callback function does state updating
 * for the leader control logic in pg_backup_archiver.c.
 *
 * In principle additional archive-format-specific information might be needed
 * in commands or worker status responses, but so far that hasn't proved
 * necessary, since workers have full copies of the ArchiveHandle/TocEntry
 * data structures.  Remember that we have forked off the workers only after
 * we have read in the catalog.  That's why our worker processes can also
 * access the catalog information.  (In the Windows case, the workers are
 * threads in the same process.  To avoid problems, they work with cloned
 * copies of the Archive data structure; see RunWorker().)
 *
 * In the leader process, the workerStatus field for each worker has one of
 * the following values:
 *      WRKR_NOT_STARTED: we've not yet forked this worker
 *      WRKR_IDLE: it's waiting for a command
 *      WRKR_WORKING: it's working on a command
 *      WRKR_TERMINATED: process ended
 * The pstate->te[] entry for each worker is valid when it's in WRKR_WORKING
 * state, and must be NULL in other states.
 */

#include "postgres_fe.h"

#ifndef WIN32
#include <sys/wait.h>
#include <signal.h>
#include <unistd.h>
#include <fcntl.h>
#endif
#ifdef HAVE_SYS_SELECT_H
#include <sys/select.h>
#endif

#include "fe_utils/string_utils.h"
#include "parallel.h"
#include "pg_backup_utils.h"
#include "port/pg_bswap.h"

/* Mnemonic macros for indexing the fd array returned by pipe(2) */
#define PIPE_READ                           0
#define PIPE_WRITE                          1

#define NO_SLOT (-1)            /* Failure result for GetIdleWorker() */

/* Worker process statuses */
typedef enum
{
    WRKR_NOT_STARTED = 0,
    WRKR_IDLE,
    WRKR_WORKING,
    WRKR_TERMINATED
} T_WorkerStatus;

#define WORKER_IS_RUNNING(workerStatus) \
    ((workerStatus) == WRKR_IDLE || (workerStatus) == WRKR_WORKING)

/*
 * Private per-parallel-worker state (typedef for this is in parallel.h).
 *
 * Much of this is valid only in the leader process (or, on Windows, should
 * be touched only by the leader thread).  But the AH field should be touched
 * only by workers.  The pipe descriptors are valid everywhere.
 */
struct ParallelSlot
{
    T_WorkerStatus workerStatus;    /* see enum above */

    /* These fields are valid if workerStatus == WRKR_WORKING: */
    ParallelCompletionPtr callback; /* function to call on completion */
    void       *callback_data;  /* passthrough data for it */

    ArchiveHandle *AH;          /* Archive data worker is using */

    int         pipeRead;       /* leader's end of the pipes */
    int         pipeWrite;
    int         pipeRevRead;    /* child's end of the pipes */
    int         pipeRevWrite;

    /* Child process/thread identity info: */
#ifdef WIN32
    uintptr_t   hThread;
    unsigned int threadId;
#else
    pid_t       pid;
#endif
};

#ifdef WIN32

/*
 * Structure to hold info passed by _beginthreadex() to the function it calls
 * via its single allowed argument.
 */
typedef struct
{
    ArchiveHandle *AH;          /* leader database connection */
    ParallelSlot *slot;         /* this worker's parallel slot */
} WorkerInfo;

/* Windows implementation of pipe access */
static int  pgpipe(int handles[2]);
static int  piperead(int s, char *buf, int len);
#define pipewrite(a,b,c)    send(a,b,c,0)

#else                           /* !WIN32 */

/* Non-Windows implementation of pipe access */
#define pgpipe(a)           pipe(a)
#define piperead(a,b,c)     read(a,b,c)
#define pipewrite(a,b,c)    write(a,b,c)

#endif                          /* WIN32 */

/*
 * State info for archive_close_connection() shutdown callback.
 */
typedef struct ShutdownInformation
{
    ParallelState *pstate;
    Archive    *AHX;
} ShutdownInformation;

static ShutdownInformation shutdown_info;

/*
 * State info for signal handling.
 * We assume signal_info initializes to zeroes.
 *
 * On Unix, myAH is the leader DB connection in the leader process, and the
 * worker's own connection in worker processes.  On Windows, we have only one
 * instance of signal_info, so myAH is the leader connection and the worker
 * connections must be dug out of pstate->parallelSlot[].
 */
typedef struct DumpSignalInformation
{
    ArchiveHandle *myAH;        /* database connection to issue cancel for */
    ParallelState *pstate;      /* parallel state, if any */
    bool        handler_set;    /* signal handler set up in this process? */
#ifndef WIN32
    bool        am_worker;      /* am I a worker process? */
#endif
} DumpSignalInformation;

static volatile DumpSignalInformation signal_info;

#ifdef WIN32
static CRITICAL_SECTION signal_info_lock;
#endif

/*
 * Write a simple string to stderr --- must be safe in a signal handler.
 * We ignore the write() result since there's not much we could do about it.
 * Certain compilers make that harder than it ought to be.
 */
#define write_stderr(str) \
    do { \
        const char *str_ = (str); \
        int     rc_; \
        rc_ = write(fileno(stderr), str_, strlen(str_)); \
        (void) rc_; \
    } while (0)


#ifdef WIN32
/* file-scope variables */
static DWORD tls_index;

/* globally visible variables (needed by exit_nicely) */
bool        parallel_init_done = false;
DWORD       mainThreadId;
#endif                          /* WIN32 */

/* Local function prototypes */
static ParallelSlot *GetMyPSlot(ParallelState *pstate);
static void archive_close_connection(int code, void *arg);
static void ShutdownWorkersHard(ParallelState *pstate);
static void WaitForTerminatingWorkers(ParallelState *pstate);
static void setup_cancel_handler(void);
static void set_cancel_pstate(ParallelState *pstate);
static void set_cancel_slot_archive(ParallelSlot *slot, ArchiveHandle *AH);
static void RunWorker(ArchiveHandle *AH, ParallelSlot *slot);
static int  GetIdleWorker(ParallelState *pstate);
static bool HasEveryWorkerTerminated(ParallelState *pstate);
static void lockTableForWorker(ArchiveHandle *AH, TocEntry *te);
static void WaitForCommands(ArchiveHandle *AH, int pipefd[2]);
static bool ListenToWorkers(ArchiveHandle *AH, ParallelState *pstate,
                            bool do_wait);
static char *getMessageFromLeader(int pipefd[2]);
static void sendMessageToLeader(int pipefd[2], const char *str);
static int  select_loop(int maxFd, fd_set *workerset);
static char *getMessageFromWorker(ParallelState *pstate,
                                  bool do_wait, int *worker);
static void sendMessageToWorker(ParallelState *pstate,
                                int worker, const char *str);
static char *readMessageFromPipe(int fd);

#define messageStartsWith(msg, prefix) \
    (strncmp(msg, prefix, strlen(prefix)) == 0)


/*
 * Shutdown callback to clean up socket access
 */
#ifdef WIN32
static void
shutdown_parallel_dump_utils(int code, void *unused)
{
    /* Call the cleanup function only from the main thread */
    if (mainThreadId == GetCurrentThreadId())
        WSACleanup();
}
#endif

/*
 * Initialize parallel dump support --- should be called early in process
 * startup.  (Currently, this is called whether or not we intend parallel
 * activity.)
 */
void
init_parallel_dump_utils(void)
{
#ifdef WIN32
    if (!parallel_init_done)
    {
        WSADATA     wsaData;
        int         err;

        /* Prepare for threaded operation */
        tls_index = TlsAlloc();
        mainThreadId = GetCurrentThreadId();

        /* Initialize socket access */
        err = WSAStartup(MAKEWORD(2, 2), &wsaData);
        if (err != 0)
        {
            pg_log_error("WSAStartup failed: %d", err);
            exit_nicely(1);
        }
        /* ... and arrange to shut it down at exit */
        on_exit_nicely(shutdown_parallel_dump_utils, NULL);
        parallel_init_done = true;
    }
#endif
}

/*
 * Find the ParallelSlot for the current worker process or thread.
 *
 * Returns NULL if no matching slot is found (this implies we're the leader).
 */
static ParallelSlot *
GetMyPSlot(ParallelState *pstate)
{
    int         i;

    for (i = 0; i < pstate->numWorkers; i++)
    {
#ifdef WIN32
        if (pstate->parallelSlot[i].threadId == GetCurrentThreadId())
#else
        if (pstate->parallelSlot[i].pid == getpid())
#endif
            return &(pstate->parallelSlot[i]);
    }

    return NULL;
}

/*
 * A thread-local version of getLocalPQExpBuffer().
 *
 * Non-reentrant but reduces memory leakage: we'll consume one buffer per
 * thread, which is much better than one per fmtId/fmtQualifiedId call.
 */
#ifdef WIN32
static PQExpBuffer
getThreadLocalPQExpBuffer(void)
{
    /*
     * The Tls code goes awry if we use a static var, so we provide for both
     * static and auto, and omit any use of the static var when using Tls. We
     * rely on TlsGetValue() to return 0 if the value is not yet set.
     */
    static PQExpBuffer s_id_return = NULL;
    PQExpBuffer id_return;

    if (parallel_init_done)
        id_return = (PQExpBuffer) TlsGetValue(tls_index);
    else
        id_return = s_id_return;

    if (id_return)              /* first time through? */
    {
        /* same buffer, just wipe contents */
        resetPQExpBuffer(id_return);
    }
    else
    {
        /* new buffer */
        id_return = createPQExpBuffer();
        if (parallel_init_done)
            TlsSetValue(tls_index, id_return);
        else
            s_id_return = id_return;
    }

    return id_return;
}
#endif                          /* WIN32 */

/*
 * pg_dump and pg_restore call this to register the cleanup handler
 * as soon as they've created the ArchiveHandle.
 */
void
on_exit_close_archive(Archive *AHX)
{
    shutdown_info.AHX = AHX;
    on_exit_nicely(archive_close_connection, &shutdown_info);
}

/*
 * on_exit_nicely handler for shutting down database connections and
 * worker processes cleanly.
 */
static void
archive_close_connection(int code, void *arg)
{
    ShutdownInformation *si = (ShutdownInformation *) arg;

    if (si->pstate)
    {
        /* In parallel mode, must figure out who we are */
        ParallelSlot *slot = GetMyPSlot(si->pstate);

        if (!slot)
        {
            /*
             * We're the leader.  Forcibly shut down workers, then close our
             * own database connection, if any.
             */
            ShutdownWorkersHard(si->pstate);

            if (si->AHX)
                DisconnectDatabase(si->AHX);
        }
        else
        {
            /*
             * We're a worker.  Shut down our own DB connection if any.  On
             * Windows, we also have to close our communication sockets, to
             * emulate what will happen on Unix when the worker process exits.
             * (Without this, if this is a premature exit, the leader would
             * fail to detect it because there would be no EOF condition on
             * the other end of the pipe.)
             */
            if (slot->AH)
                DisconnectDatabase(&(slot->AH->public));

#ifdef WIN32
            closesocket(slot->pipeRevRead);
            closesocket(slot->pipeRevWrite);
#endif
        }
    }
    else
    {
        /* Non-parallel operation: just kill the leader DB connection */
        if (si->AHX)
            DisconnectDatabase(si->AHX);
    }
}

/*
 * Forcibly shut down any remaining workers, waiting for them to finish.
 *
 * Note that we don't expect to come here during normal exit (the workers
 * should be long gone, and the ParallelState too).  We're only here in a
 * fatal() situation, so intervening to cancel active commands is
 * appropriate.
 */
static void
ShutdownWorkersHard(ParallelState *pstate)
{
    int         i;

    /*
     * Close our write end of the sockets so that any workers waiting for
     * commands know they can exit.  (Note: some of the pipeWrite fields might
     * still be zero, if we failed to initialize all the workers.  Hence, just
     * ignore errors here.)
     */
    for (i = 0; i < pstate->numWorkers; i++)
        closesocket(pstate->parallelSlot[i].pipeWrite);

    /*
     * Force early termination of any commands currently in progress.
     */
#ifndef WIN32
    /* On non-Windows, send SIGTERM to each worker process. */
    for (i = 0; i < pstate->numWorkers; i++)
    {
        pid_t       pid = pstate->parallelSlot[i].pid;

        if (pid != 0)
            kill(pid, SIGTERM);
    }
#else

    /*
     * On Windows, send query cancels directly to the workers' backends.  Use
     * a critical section to ensure worker threads don't change state.
     */
    EnterCriticalSection(&signal_info_lock);
    for (i = 0; i < pstate->numWorkers; i++)
    {
        ArchiveHandle *AH = pstate->parallelSlot[i].AH;
        char        errbuf[1];

        if (AH != NULL && AH->connCancel != NULL)
            (void) PQcancel(AH->connCancel, errbuf, sizeof(errbuf));
    }
    LeaveCriticalSection(&signal_info_lock);
#endif

    /* Now wait for them to terminate. */
    WaitForTerminatingWorkers(pstate);
}

/*
 * Wait for all workers to terminate.
 */
static void
WaitForTerminatingWorkers(ParallelState *pstate)
{
    while (!HasEveryWorkerTerminated(pstate))
    {
        ParallelSlot *slot = NULL;
        int         j;

#ifndef WIN32
        /* On non-Windows, use wait() to wait for next worker to end */
        int         status;
        pid_t       pid = wait(&status);

        /* Find dead worker's slot, and clear the PID field */
        for (j = 0; j < pstate->numWorkers; j++)
        {
            slot = &(pstate->parallelSlot[j]);
            if (slot->pid == pid)
            {
                slot->pid = 0;
                break;
            }
        }
#else                           /* WIN32 */
        /* On Windows, we must use WaitForMultipleObjects() */
        HANDLE     *lpHandles = pg_malloc(sizeof(HANDLE) * pstate->numWorkers);
        int         nrun = 0;
        DWORD       ret;
        uintptr_t   hThread;

        for (j = 0; j < pstate->numWorkers; j++)
        {
            if (WORKER_IS_RUNNING(pstate->parallelSlot[j].workerStatus))
            {
                lpHandles[nrun] = (HANDLE) pstate->parallelSlot[j].hThread;
                nrun++;
            }
        }
        ret = WaitForMultipleObjects(nrun, lpHandles, false, INFINITE);
        Assert(ret != WAIT_FAILED);
        hThread = (uintptr_t) lpHandles[ret - WAIT_OBJECT_0];
        free(lpHandles);

        /* Find dead worker's slot, and clear the hThread field */
        for (j = 0; j < pstate->numWorkers; j++)
        {
            slot = &(pstate->parallelSlot[j]);
            if (slot->hThread == hThread)
            {
                /* For cleanliness, close handles for dead threads */
                CloseHandle((HANDLE) slot->hThread);
                slot->hThread = (uintptr_t) INVALID_HANDLE_VALUE;
                break;
            }
        }
#endif                          /* WIN32 */

        /* On all platforms, update workerStatus and te[] as well */
        Assert(j < pstate->numWorkers);
        slot->workerStatus = WRKR_TERMINATED;
        pstate->te[j] = NULL;
    }
}


/*
 * Code for responding to cancel interrupts (SIGINT, control-C, etc)
 *
 * This doesn't quite belong in this module, but it needs access to the
 * ParallelState data, so there's not really a better place either.
 *
 * When we get a cancel interrupt, we could just die, but in pg_restore that
 * could leave a SQL command (e.g., CREATE INDEX on a large table) running
 * for a long time.  Instead, we try to send a cancel request and then die.
 * pg_dump probably doesn't really need this, but we might as well use it
 * there too.  Note that sending the cancel directly from the signal handler
 * is safe because PQcancel() is written to make it so.
 *
 * In parallel operation on Unix, each process is responsible for canceling
 * its own connection (this must be so because nobody else has access to it).
 * Furthermore, the leader process should attempt to forward its signal to
 * each child.  In simple manual use of pg_dump/pg_restore, forwarding isn't
 * needed because typing control-C at the console would deliver SIGINT to
 * every member of the terminal process group --- but in other scenarios it
 * might be that only the leader gets signaled.
 *
 * On Windows, the cancel handler runs in a separate thread, because that's
 * how SetConsoleCtrlHandler works.  We make it stop worker threads, send
 * cancels on all active connections, and then return FALSE, which will allow
 * the process to die.  For safety's sake, we use a critical section to
 * protect the PGcancel structures against being changed while the signal
 * thread runs.
 */

#ifndef WIN32

/*
 * Signal handler (Unix only)
 */
static void
sigTermHandler(SIGNAL_ARGS)
{
    int         i;
    char        errbuf[1];

    /*
     * Some platforms allow delivery of new signals to interrupt an active
     * signal handler.  That could muck up our attempt to send PQcancel, so
     * disable the signals that setup_cancel_handler enabled.
     */
    pqsignal(SIGINT, SIG_IGN);
    pqsignal(SIGTERM, SIG_IGN);
    pqsignal(SIGQUIT, SIG_IGN);

    /*
     * If we're in the leader, forward signal to all workers.  (It seems best
     * to do this before PQcancel; killing the leader transaction will result
     * in invalid-snapshot errors from active workers, which maybe we can
     * quiet by killing workers first.)  Ignore any errors.
     */
    if (signal_info.pstate != NULL)
    {
        for (i = 0; i < signal_info.pstate->numWorkers; i++)
        {
            pid_t       pid = signal_info.pstate->parallelSlot[i].pid;

            if (pid != 0)
                kill(pid, SIGTERM);
        }
    }

    /*
     * Send QueryCancel if we have a connection to send to.  Ignore errors,
     * there's not much we can do about them anyway.
     */
    if (signal_info.myAH != NULL && signal_info.myAH->connCancel != NULL)
        (void) PQcancel(signal_info.myAH->connCancel, errbuf, sizeof(errbuf));

    /*
     * Report we're quitting, using nothing more complicated than write(2).
     * When in parallel operation, only the leader process should do this.
     */
    if (!signal_info.am_worker)
    {
        if (progname)
        {
            write_stderr(progname);
            write_stderr(": ");
        }
        write_stderr("terminated by user\n");
    }

    /*
     * And die, using _exit() not exit() because the latter will invoke atexit
     * handlers that can fail if we interrupted related code.
     */
    _exit(1);
}

/*
 * Enable cancel interrupt handler, if not already done.
 */
static void
setup_cancel_handler(void)
{
    /*
     * When forking, signal_info.handler_set will propagate into the new
     * process, but that's fine because the signal handler state does too.
     */
    if (!signal_info.handler_set)
    {
        signal_info.handler_set = true;

        pqsignal(SIGINT, sigTermHandler);
        pqsignal(SIGTERM, sigTermHandler);
        pqsignal(SIGQUIT, sigTermHandler);
    }
}

#else                           /* WIN32 */

/*
 * Console interrupt handler --- runs in a newly-started thread.
 *
 * After stopping other threads and sending cancel requests on all open
 * connections, we return FALSE which will allow the default ExitProcess()
 * action to be taken.
 */
static BOOL WINAPI
consoleHandler(DWORD dwCtrlType)
{
    int         i;
    char        errbuf[1];

    if (dwCtrlType == CTRL_C_EVENT ||
        dwCtrlType == CTRL_BREAK_EVENT)
    {
        /* Critical section prevents changing data we look at here */
        EnterCriticalSection(&signal_info_lock);

        /*
         * If in parallel mode, stop worker threads and send QueryCancel to
         * their connected backends.  The main point of stopping the worker
         * threads is to keep them from reporting the query cancels as errors,
         * which would clutter the user's screen.  We needn't stop the leader
         * thread since it won't be doing much anyway.  Do this before
         * canceling the main transaction, else we might get invalid-snapshot
         * errors reported before we can stop the workers.  Ignore errors,
         * there's not much we can do about them anyway.
         */
        if (signal_info.pstate != NULL)
        {
            for (i = 0; i < signal_info.pstate->numWorkers; i++)
            {
                ParallelSlot *slot = &(signal_info.pstate->parallelSlot[i]);
                ArchiveHandle *AH = slot->AH;
                HANDLE      hThread = (HANDLE) slot->hThread;

                /*
                 * Using TerminateThread here may leave some resources leaked,
                 * but it doesn't matter since we're about to end the whole
                 * process.
                 */
                if (hThread != INVALID_HANDLE_VALUE)
                    TerminateThread(hThread, 0);

                if (AH != NULL && AH->connCancel != NULL)
                    (void) PQcancel(AH->connCancel, errbuf, sizeof(errbuf));
            }
        }

        /*
         * Send QueryCancel to leader connection, if enabled.  Ignore errors,
         * there's not much we can do about them anyway.
         */
        if (signal_info.myAH != NULL && signal_info.myAH->connCancel != NULL)
            (void) PQcancel(signal_info.myAH->connCancel,
                            errbuf, sizeof(errbuf));

        LeaveCriticalSection(&signal_info_lock);

        /*
         * Report we're quitting, using nothing more complicated than
         * write(2).  (We might be able to get away with using pg_log_*()
         * here, but since we terminated other threads uncleanly above, it
         * seems better to assume as little as possible.)
         */
        if (progname)
        {
            write_stderr(progname);
            write_stderr(": ");
        }
        write_stderr("terminated by user\n");
    }

    /* Always return FALSE to allow signal handling to continue */
    return FALSE;
}

/*
 * Enable cancel interrupt handler, if not already done.
 */
static void
setup_cancel_handler(void)
{
    if (!signal_info.handler_set)
    {
        signal_info.handler_set = true;

        InitializeCriticalSection(&signal_info_lock);

        SetConsoleCtrlHandler(consoleHandler, TRUE);
    }
}

#endif                          /* WIN32 */


/*
 * set_archive_cancel_info
 *
 * Fill AH->connCancel with cancellation info for the specified database
 * connection; or clear it if conn is NULL.
 */
void
set_archive_cancel_info(ArchiveHandle *AH, PGconn *conn)
{
    PGcancel   *oldConnCancel;

    /*
     * Activate the interrupt handler if we didn't yet in this process.  On
     * Windows, this also initializes signal_info_lock; therefore it's
     * important that this happen at least once before we fork off any
     * threads.
     */
    setup_cancel_handler();

    /*
     * On Unix, we assume that storing a pointer value is atomic with respect
     * to any possible signal interrupt.  On Windows, use a critical section.
     */

#ifdef WIN32
    EnterCriticalSection(&signal_info_lock);
#endif

    /* Free the old one if we have one */
    oldConnCancel = AH->connCancel;
    /* be sure interrupt handler doesn't use pointer while freeing */
    AH->connCancel = NULL;

    if (oldConnCancel != NULL)
        PQfreeCancel(oldConnCancel);

    /* Set the new one if specified */
    if (conn)
        AH->connCancel = PQgetCancel(conn);

    /*
     * On Unix, there's only ever one active ArchiveHandle per process, so we
     * can just set signal_info.myAH unconditionally.  On Windows, do that
     * only in the main thread; worker threads have to make sure their
     * ArchiveHandle appears in the pstate data, which is dealt with in
     * RunWorker().
     */
#ifndef WIN32
    signal_info.myAH = AH;
#else
    if (mainThreadId == GetCurrentThreadId())
        signal_info.myAH = AH;
#endif

#ifdef WIN32
    LeaveCriticalSection(&signal_info_lock);
#endif
}

/*
 * set_cancel_pstate
 *
 * Set signal_info.pstate to point to the specified ParallelState, if any.
 * We need this mainly to have an interlock against Windows signal thread.
 */
static void
set_cancel_pstate(ParallelState *pstate)
{
#ifdef WIN32
    EnterCriticalSection(&signal_info_lock);
#endif

    signal_info.pstate = pstate;

#ifdef WIN32
    LeaveCriticalSection(&signal_info_lock);
#endif
}

/*
 * set_cancel_slot_archive
 *
 * Set ParallelSlot's AH field to point to the specified archive, if any.
 * We need this mainly to have an interlock against Windows signal thread.
 */
static void
set_cancel_slot_archive(ParallelSlot *slot, ArchiveHandle *AH)
{
#ifdef WIN32
    EnterCriticalSection(&signal_info_lock);
#endif

    slot->AH = AH;

#ifdef WIN32
    LeaveCriticalSection(&signal_info_lock);
#endif
}


/*
 * This function is called by both Unix and Windows variants to set up
 * and run a worker process.  Caller should exit the process (or thread)
 * upon return.
 */
static void
RunWorker(ArchiveHandle *AH, ParallelSlot *slot)
{
    int         pipefd[2];

    /* fetch child ends of pipes */
    pipefd[PIPE_READ] = slot->pipeRevRead;
    pipefd[PIPE_WRITE] = slot->pipeRevWrite;

    /*
     * Clone the archive so that we have our own state to work with, and in
     * particular our own database connection.
     *
     * We clone on Unix as well as Windows, even though technically we don't
     * need to because fork() gives us a copy in our own address space
     * already.  But CloneArchive resets the state information and also clones
     * the database connection which both seem kinda helpful.
     */
    AH = CloneArchive(AH);

    /* Remember cloned archive where signal handler can find it */
    set_cancel_slot_archive(slot, AH);

    /*
     * Call the setup worker function that's defined in the ArchiveHandle.
     */
    (AH->SetupWorkerPtr) ((Archive *) AH);

    /*
     * Execute commands until done.
     */
    WaitForCommands(AH, pipefd);

    /*
     * Disconnect from database and clean up.
     */
    set_cancel_slot_archive(slot, NULL);
    DisconnectDatabase(&(AH->public));
    DeCloneArchive(AH);
}

/*
 * Thread base function for Windows
 */
#ifdef WIN32
static unsigned __stdcall
init_spawned_worker_win32(WorkerInfo *wi)
{
    ArchiveHandle *AH = wi->AH;
    ParallelSlot *slot = wi->slot;

    /* Don't need WorkerInfo anymore */
    free(wi);

    /* Run the worker ... */
    RunWorker(AH, slot);

    /* Exit the thread */
    _endthreadex(0);
    return 0;
}
#endif                          /* WIN32 */

/*
 * This function starts a parallel dump or restore by spawning off the worker
 * processes.  For Windows, it creates a number of threads; on Unix the
 * workers are created with fork().
 */
ParallelState *
ParallelBackupStart(ArchiveHandle *AH)
{
    ParallelState *pstate;
    int         i;

    Assert(AH->public.numWorkers > 0);

    pstate = (ParallelState *) pg_malloc(sizeof(ParallelState));

    pstate->numWorkers = AH->public.numWorkers;
    pstate->te = NULL;
    pstate->parallelSlot = NULL;

    if (AH->public.numWorkers == 1)
        return pstate;

    /* Create status arrays, being sure to initialize all fields to 0 */
    pstate->te = (TocEntry **)
        pg_malloc0(pstate->numWorkers * sizeof(TocEntry *));
    pstate->parallelSlot = (ParallelSlot *)
        pg_malloc0(pstate->numWorkers * sizeof(ParallelSlot));

#ifdef WIN32
    /* Make fmtId() and fmtQualifiedId() use thread-local storage */
    getLocalPQExpBuffer = getThreadLocalPQExpBuffer;
#endif

    /*
     * Set the pstate in shutdown_info, to tell the exit handler that it must
     * clean up workers as well as the main database connection.  But we don't
     * set this in signal_info yet, because we don't want child processes to
     * inherit non-NULL signal_info.pstate.
     */
    shutdown_info.pstate = pstate;

    /*
     * Temporarily disable query cancellation on the leader connection.  This
     * ensures that child processes won't inherit valid AH->connCancel
     * settings and thus won't try to issue cancels against the leader's
     * connection.  No harm is done if we fail while it's disabled, because
     * the leader connection is idle at this point anyway.
     */
    set_archive_cancel_info(AH, NULL);

    /* Ensure stdio state is quiesced before forking */
    fflush(NULL);

    /* Create desired number of workers */
    for (i = 0; i < pstate->numWorkers; i++)
    {
#ifdef WIN32
        WorkerInfo *wi;
        uintptr_t   handle;
#else
        pid_t       pid;
#endif
        ParallelSlot *slot = &(pstate->parallelSlot[i]);
        int         pipeMW[2],
                    pipeWM[2];

        /* Create communication pipes for this worker */
        if (pgpipe(pipeMW) < 0 || pgpipe(pipeWM) < 0)
            fatal("could not create communication channels: %m");

        /* leader's ends of the pipes */
        slot->pipeRead = pipeWM[PIPE_READ];
        slot->pipeWrite = pipeMW[PIPE_WRITE];
        /* child's ends of the pipes */
        slot->pipeRevRead = pipeMW[PIPE_READ];
        slot->pipeRevWrite = pipeWM[PIPE_WRITE];

#ifdef WIN32
        /* Create transient structure to pass args to worker function */
        wi = (WorkerInfo *) pg_malloc(sizeof(WorkerInfo));

        wi->AH = AH;
        wi->slot = slot;

        handle = _beginthreadex(NULL, 0, (void *) &init_spawned_worker_win32,
                                wi, 0, &(slot->threadId));
        slot->hThread = handle;
        slot->workerStatus = WRKR_IDLE;
#else                           /* !WIN32 */
        pid = fork();
        if (pid == 0)
        {
            /* we are the worker */
            int         j;

            /* this is needed for GetMyPSlot() */
            slot->pid = getpid();

            /* instruct signal handler that we're in a worker now */
            signal_info.am_worker = true;

            /* close read end of Worker -> Leader */
            closesocket(pipeWM[PIPE_READ]);
            /* close write end of Leader -> Worker */
            closesocket(pipeMW[PIPE_WRITE]);

            /*
             * Close all inherited fds for communication of the leader with
             * previously-forked workers.
             */
            for (j = 0; j < i; j++)
            {
                closesocket(pstate->parallelSlot[j].pipeRead);
                closesocket(pstate->parallelSlot[j].pipeWrite);
            }

            /* Run the worker ... */
            RunWorker(AH, slot);

            /* We can just exit(0) when done */
            exit(0);
        }
        else if (pid < 0)
        {
            /* fork failed */
            fatal("could not create worker process: %m");
        }

        /* In Leader after successful fork */
        slot->pid = pid;
        slot->workerStatus = WRKR_IDLE;

        /* close read end of Leader -> Worker */
        closesocket(pipeMW[PIPE_READ]);
        /* close write end of Worker -> Leader */
        closesocket(pipeWM[PIPE_WRITE]);
#endif                          /* WIN32 */
    }

    /*
     * Having forked off the workers, disable SIGPIPE so that leader isn't
     * killed if it tries to send a command to a dead worker.  We don't want
     * the workers to inherit this setting, though.
     */
#ifndef WIN32
    pqsignal(SIGPIPE, SIG_IGN);
#endif

    /*
     * Re-establish query cancellation on the leader connection.
     */
    set_archive_cancel_info(AH, AH->connection);

    /*
     * Tell the cancel signal handler to forward signals to worker processes,
     * too.  (As with query cancel, we did not need this earlier because the
     * workers have not yet been given anything to do; if we die before this
     * point, any already-started workers will see EOF and quit promptly.)
     */
    set_cancel_pstate(pstate);

    return pstate;
}

/*
 * Close down a parallel dump or restore.
 */
void
ParallelBackupEnd(ArchiveHandle *AH, ParallelState *pstate)
{
    int         i;

    /* No work if non-parallel */
    if (pstate->numWorkers == 1)
        return;

    /* There should not be any unfinished jobs */
    Assert(IsEveryWorkerIdle(pstate));

    /* Close the sockets so that the workers know they can exit */
    for (i = 0; i < pstate->numWorkers; i++)
    {
        closesocket(pstate->parallelSlot[i].pipeRead);
        closesocket(pstate->parallelSlot[i].pipeWrite);
    }

    /* Wait for them to exit */
    WaitForTerminatingWorkers(pstate);

    /*
     * Unlink pstate from shutdown_info, so the exit handler will not try to
     * use it; and likewise unlink from signal_info.
     */
    shutdown_info.pstate = NULL;
    set_cancel_pstate(NULL);

    /* Release state (mere neatnik-ism, since we're about to terminate) */
    free(pstate->te);
    free(pstate->parallelSlot);
    free(pstate);
}

/*
 * These next four functions handle construction and parsing of the command
 * strings and response strings for parallel workers.
 *
 * Currently, these can be the same regardless of which archive format we are
 * processing.  In future, we might want to let format modules override these
 * functions to add format-specific data to a command or response.
 */

/*
 * buildWorkerCommand: format a command string to send to a worker.
 *
 * The string is built in the caller-supplied buffer of size buflen.
 */
static void
buildWorkerCommand(ArchiveHandle *AH, TocEntry *te, T_Action act,
                   char *buf, int buflen)
{
    if (act == ACT_DUMP)
        snprintf(buf, buflen, "DUMP %d", te->dumpId);
    else if (act == ACT_RESTORE)
        snprintf(buf, buflen, "RESTORE %d", te->dumpId);
    else
        Assert(false);
}

/*
 * parseWorkerCommand: interpret a command string in a worker.
 */
static void
parseWorkerCommand(ArchiveHandle *AH, TocEntry **te, T_Action *act,
                   const char *msg)
{
    DumpId      dumpId;
    int         nBytes;

    if (messageStartsWith(msg, "DUMP "))
    {
        *act = ACT_DUMP;
        sscanf(msg, "DUMP %d%n", &dumpId, &nBytes);
        Assert(nBytes == strlen(msg));
        *te = getTocEntryByDumpId(AH, dumpId);
        Assert(*te != NULL);
    }
    else if (messageStartsWith(msg, "RESTORE "))
    {
        *act = ACT_RESTORE;
        sscanf(msg, "RESTORE %d%n", &dumpId, &nBytes);
        Assert(nBytes == strlen(msg));
        *te = getTocEntryByDumpId(AH, dumpId);
        Assert(*te != NULL);
    }
    else
        fatal("unrecognized command received from leader: \"%s\"",
              msg);
}

/*
 * buildWorkerResponse: format a response string to send to the leader.
 *
 * The string is built in the caller-supplied buffer of size buflen.
 */
static void
buildWorkerResponse(ArchiveHandle *AH, TocEntry *te, T_Action act, int status,
                    char *buf, int buflen)
{
    snprintf(buf, buflen, "OK %d %d %d",
             te->dumpId,
             status,
             status == WORKER_IGNORED_ERRORS ? AH->public.n_errors : 0);
}

/*
 * parseWorkerResponse: parse the status message returned by a worker.
 *
 * Returns the integer status code, and may update fields of AH and/or te.
 */
static int
parseWorkerResponse(ArchiveHandle *AH, TocEntry *te,
                    const char *msg)
{
    DumpId      dumpId;
    int         nBytes,
                n_errors;
    int         status = 0;

    if (messageStartsWith(msg, "OK "))
    {
        sscanf(msg, "OK %d %d %d%n", &dumpId, &status, &n_errors, &nBytes);

        Assert(dumpId == te->dumpId);
        Assert(nBytes == strlen(msg));

        AH->public.n_errors += n_errors;
    }
    else
        fatal("invalid message received from worker: \"%s\"",
              msg);

    return status;
}

/*
 * Dispatch a job to some free worker.
 *
 * te is the TocEntry to be processed, act is the action to be taken on it.
 * callback is the function to call on completion of the job.
 *
 * If no worker is currently available, this will block, and previously
 * registered callback functions may be called.
 */
void
DispatchJobForTocEntry(ArchiveHandle *AH,
                       ParallelState *pstate,
                       TocEntry *te,
                       T_Action act,
                       ParallelCompletionPtr callback,
                       void *callback_data)
{
    int         worker;
    char        buf[256];

    /* Get a worker, waiting if none are idle */
    while ((worker = GetIdleWorker(pstate)) == NO_SLOT)
        WaitForWorkers(AH, pstate, WFW_ONE_IDLE);

    /* Construct and send command string */
    buildWorkerCommand(AH, te, act, buf, sizeof(buf));

    sendMessageToWorker(pstate, worker, buf);

    /* Remember worker is busy, and which TocEntry it's working on */
    pstate->parallelSlot[worker].workerStatus = WRKR_WORKING;
    pstate->parallelSlot[worker].callback = callback;
    pstate->parallelSlot[worker].callback_data = callback_data;
    pstate->te[worker] = te;
}

/*
 * Find an idle worker and return its slot number.
 * Return NO_SLOT if none are idle.
 */
static int
GetIdleWorker(ParallelState *pstate)
{
    int         i;

    for (i = 0; i < pstate->numWorkers; i++)
    {
        if (pstate->parallelSlot[i].workerStatus == WRKR_IDLE)
            return i;
    }
    return NO_SLOT;
}

/*
 * Return true iff no worker is running.
 */
static bool
HasEveryWorkerTerminated(ParallelState *pstate)
{
    int         i;

    for (i = 0; i < pstate->numWorkers; i++)
    {
        if (WORKER_IS_RUNNING(pstate->parallelSlot[i].workerStatus))
            return false;
    }
    return true;
}

/*
 * Return true iff every worker is in the WRKR_IDLE state.
 */
bool
IsEveryWorkerIdle(ParallelState *pstate)
{
    int         i;

    for (i = 0; i < pstate->numWorkers; i++)
    {
        if (pstate->parallelSlot[i].workerStatus != WRKR_IDLE)
            return false;
    }
    return true;
}

/*
 * Acquire lock on a table to be dumped by a worker process.
 *
 * The leader process is already holding an ACCESS SHARE lock.  Ordinarily
 * it's no problem for a worker to get one too, but if anything else besides
 * pg_dump is running, there's a possible deadlock:
 *
 * 1) Leader dumps the schema and locks all tables in ACCESS SHARE mode.
 * 2) Another process requests an ACCESS EXCLUSIVE lock (which is not granted
 *    because the leader holds a conflicting ACCESS SHARE lock).
 * 3) A worker process also requests an ACCESS SHARE lock to read the table.
 *    The worker is enqueued behind the ACCESS EXCLUSIVE lock request.
 * 4) Now we have a deadlock, since the leader is effectively waiting for
 *    the worker.  The server cannot detect that, however.
 *
 * To prevent an infinite wait, prior to touching a table in a worker, request
 * a lock in ACCESS SHARE mode but with NOWAIT.  If we don't get the lock,
 * then we know that somebody else has requested an ACCESS EXCLUSIVE lock and
 * so we have a deadlock.  We must fail the backup in that case.
 */
static void
lockTableForWorker(ArchiveHandle *AH, TocEntry *te)
{
    const char *qualId;
    PQExpBuffer query;
    PGresult   *res;

    /* Nothing to do for BLOBS */
    if (strcmp(te->desc, "BLOBS") == 0)
        return;

    query = createPQExpBuffer();

    qualId = fmtQualifiedId(te->namespace, te->tag);

    appendPQExpBuffer(query, "LOCK TABLE %s IN ACCESS SHARE MODE NOWAIT",
                      qualId);

    res = PQexec(AH->connection, query->data);

    if (!res || PQresultStatus(res) != PGRES_COMMAND_OK)
        fatal("could not obtain lock on relation \"%s\"\n"
              "This usually means that someone requested an ACCESS EXCLUSIVE lock "
              "on the table after the pg_dump parent process had gotten the "
              "initial ACCESS SHARE lock on the table.", qualId);

    PQclear(res);
    destroyPQExpBuffer(query);
}

/*
 * WaitForCommands: main routine for a worker process.
 *
 * Read and execute commands from the leader until we see EOF on the pipe.
 */
static void
WaitForCommands(ArchiveHandle *AH, int pipefd[2])
{
    char       *command;
    TocEntry   *te;
    T_Action    act;
    int         status = 0;
    char        buf[256];

    for (;;)
    {
        if (!(command = getMessageFromLeader(pipefd)))
        {
            /* EOF, so done */
            return;
        }

        /* Decode the command */
        parseWorkerCommand(AH, &te, &act, command);

        if (act == ACT_DUMP)
        {
            /* Acquire lock on this table within the worker's session */
            lockTableForWorker(AH, te);

            /* Perform the dump command */
            status = (AH->WorkerJobDumpPtr) (AH, te);
        }
        else if (act == ACT_RESTORE)
        {
            /* Perform the restore command */
            status = (AH->WorkerJobRestorePtr) (AH, te);
        }
        else
            Assert(false);

        /* Return status to leader */
        buildWorkerResponse(AH, te, act, status, buf, sizeof(buf));

        sendMessageToLeader(pipefd, buf);

        /* command was pg_malloc'd and we are responsible for free()ing it. */
        free(command);
    }
}

/*
 * Check for status messages from workers.
 *
 * If do_wait is true, wait to get a status message; otherwise, just return
 * immediately if there is none available.
 *
 * When we get a status message, we pass the status code to the callback
 * function that was specified to DispatchJobForTocEntry, then reset the
 * worker status to IDLE.
 *
 * Returns true if we collected a status message, else false.
 *
 * XXX is it worth checking for more than one status message per call?
 * It seems somewhat unlikely that multiple workers would finish at exactly
 * the same time.
 */
static bool
ListenToWorkers(ArchiveHandle *AH, ParallelState *pstate, bool do_wait)
{
    int         worker;
    char       *msg;

    /* Try to collect a status message */
    msg = getMessageFromWorker(pstate, do_wait, &worker);

    if (!msg)
    {
        /* If do_wait is true, we must have detected EOF on some socket */
        if (do_wait)
            fatal("a worker process died unexpectedly");
        return false;
    }

    /* Process it and update our idea of the worker's status */
    if (messageStartsWith(msg, "OK "))
    {
        ParallelSlot *slot = &pstate->parallelSlot[worker];
        TocEntry   *te = pstate->te[worker];
        int         status;

        status = parseWorkerResponse(AH, te, msg);
        slot->callback(AH, te, status, slot->callback_data);
        slot->workerStatus = WRKR_IDLE;
        pstate->te[worker] = NULL;
    }
    else
        fatal("invalid message received from worker: \"%s\"",
              msg);

    /* Free the string returned from getMessageFromWorker */
    free(msg);

    return true;
}

/*
 * Check for status results from workers, waiting if necessary.
 *
 * Available wait modes are:
 * WFW_NO_WAIT: reap any available status, but don't block
 * WFW_GOT_STATUS: wait for at least one more worker to finish
 * WFW_ONE_IDLE: wait for at least one worker to be idle
 * WFW_ALL_IDLE: wait for all workers to be idle
 *
 * Any received results are passed to the callback specified to
 * DispatchJobForTocEntry.
 *
 * This function is executed in the leader process.
 */
void
WaitForWorkers(ArchiveHandle *AH, ParallelState *pstate, WFW_WaitOption mode)
{
    bool        do_wait = false;

    /*
     * In GOT_STATUS mode, always block waiting for a message, since we can't
     * return till we get something.  In other modes, we don't block the first
     * time through the loop.
     */
    if (mode == WFW_GOT_STATUS)
    {
        /* Assert that caller knows what it's doing */
        Assert(!IsEveryWorkerIdle(pstate));
        do_wait = true;
    }

    for (;;)
    {
        /*
         * Check for status messages, even if we don't need to block.  We do
         * not try very hard to reap all available messages, though, since
         * there's unlikely to be more than one.
         */
        if (ListenToWorkers(AH, pstate, do_wait))
        {
            /*
             * If we got a message, we are done by definition for GOT_STATUS
             * mode, and we can also be certain that there's at least one idle
             * worker.  So we're done in all but ALL_IDLE mode.
             */
            if (mode != WFW_ALL_IDLE)
                return;
        }

        /* Check whether we must wait for new status messages */
        switch (mode)
        {
            case WFW_NO_WAIT:
                return;         /* never wait */
            case WFW_GOT_STATUS:
                Assert(false);  /* can't get here, because we waited */
                break;
            case WFW_ONE_IDLE:
                if (GetIdleWorker(pstate) != NO_SLOT)
                    return;
                break;
            case WFW_ALL_IDLE:
                if (IsEveryWorkerIdle(pstate))
                    return;
                break;
        }

        /* Loop back, and this time wait for something to happen */
        do_wait = true;
    }
}

/*
 * Read one command message from the leader, blocking if necessary
 * until one is available, and return it as a malloc'd string.
 * On EOF, return NULL.
 *
 * This function is executed in worker processes.
 */
static char *
getMessageFromLeader(int pipefd[2])
{
    return readMessageFromPipe(pipefd[PIPE_READ]);
}

/*
 * Send a status message to the leader.
 *
 * This function is executed in worker processes.
 */
static void
sendMessageToLeader(int pipefd[2], const char *str)
{
    int         len = strlen(str) + 1;

    if (pipewrite(pipefd[PIPE_WRITE], str, len) != len)
        fatal("could not write to the communication channel: %m");
}

/*
 * Wait until some descriptor in "workerset" becomes readable.
 * Returns -1 on error, else the number of readable descriptors.
 */
static int
select_loop(int maxFd, fd_set *workerset)
{
    int         i;
    fd_set      saveSet = *workerset;

    for (;;)
    {
        *workerset = saveSet;
        i = select(maxFd + 1, workerset, NULL, NULL, NULL);

#ifndef WIN32
        if (i < 0 && errno == EINTR)
            continue;
#else
        if (i == SOCKET_ERROR && WSAGetLastError() == WSAEINTR)
            continue;
#endif
        break;
    }

    return i;
}


/*
 * Check for messages from worker processes.
 *
 * If a message is available, return it as a malloc'd string, and put the
 * index of the sending worker in *worker.
 *
 * If nothing is available, wait if "do_wait" is true, else return NULL.
 *
 * If we detect EOF on any socket, we'll return NULL.  It's not great that
 * that's hard to distinguish from the no-data-available case, but for now
 * our one caller is okay with that.
 *
 * This function is executed in the leader process.
 */
static char *
getMessageFromWorker(ParallelState *pstate, bool do_wait, int *worker)
{
    int         i;
    fd_set      workerset;
    int         maxFd = -1;
    struct timeval nowait = {0, 0};

    /* construct bitmap of socket descriptors for select() */
    FD_ZERO(&workerset);
    for (i = 0; i < pstate->numWorkers; i++)
    {
        if (!WORKER_IS_RUNNING(pstate->parallelSlot[i].workerStatus))
            continue;
        FD_SET(pstate->parallelSlot[i].pipeRead, &workerset);
        if (pstate->parallelSlot[i].pipeRead > maxFd)
            maxFd = pstate->parallelSlot[i].pipeRead;
    }

    if (do_wait)
    {
        i = select_loop(maxFd, &workerset);
        Assert(i != 0);
    }
    else
    {
        if ((i = select(maxFd + 1, &workerset, NULL, NULL, &nowait)) == 0)
            return NULL;
    }

    if (i < 0)
        fatal("select() failed: %m");

    for (i = 0; i < pstate->numWorkers; i++)
    {
        char       *msg;

        if (!WORKER_IS_RUNNING(pstate->parallelSlot[i].workerStatus))
            continue;
        if (!FD_ISSET(pstate->parallelSlot[i].pipeRead, &workerset))
            continue;

        /*
         * Read the message if any.  If the socket is ready because of EOF,
         * we'll return NULL instead (and the socket will stay ready, so the
         * condition will persist).
         *
         * Note: because this is a blocking read, we'll wait if only part of
         * the message is available.  Waiting a long time would be bad, but
         * since worker status messages are short and are always sent in one
         * operation, it shouldn't be a problem in practice.
         */
        msg = readMessageFromPipe(pstate->parallelSlot[i].pipeRead);
        *worker = i;
        return msg;
    }
    Assert(false);
    return NULL;
}

/*
 * Send a command message to the specified worker process.
 *
 * This function is executed in the leader process.
 */
static void
sendMessageToWorker(ParallelState *pstate, int worker, const char *str)
{
    int         len = strlen(str) + 1;

    if (pipewrite(pstate->parallelSlot[worker].pipeWrite, str, len) != len)
    {
        fatal("could not write to the communication channel: %m");
    }
}

/*
 * Read one message from the specified pipe (fd), blocking if necessary
 * until one is available, and return it as a malloc'd string.
 * On EOF, return NULL.
 *
 * A "message" on the channel is just a null-terminated string.
 */
static char *
readMessageFromPipe(int fd)
{
    char       *msg;
    int         msgsize,
                bufsize;
    int         ret;

    /*
     * In theory, if we let piperead() read multiple bytes, it might give us
     * back fragments of multiple messages.  (That can't actually occur, since
     * neither leader nor workers send more than one message without waiting
     * for a reply, but we don't wish to assume that here.)  For simplicity,
     * read a byte at a time until we get the terminating '\0'.  This method
     * is a bit inefficient, but since this is only used for relatively short
     * command and status strings, it shouldn't matter.
     */
    bufsize = 64;               /* could be any number */
    msg = (char *) pg_malloc(bufsize);
    msgsize = 0;
    for (;;)
    {
        Assert(msgsize < bufsize);
        ret = piperead(fd, msg + msgsize, 1);
        if (ret <= 0)
            break;              /* error or connection closure */

        Assert(ret == 1);

        if (msg[msgsize] == '\0')
            return msg;         /* collected whole message */

        msgsize++;
        if (msgsize == bufsize) /* enlarge buffer if needed */
        {
            bufsize += 16;      /* could be any number */
            msg = (char *) pg_realloc(msg, bufsize);
        }
    }

    /* Other end has closed the connection */
    pg_free(msg);
    return NULL;
}

#ifdef WIN32

/*
 * This is a replacement version of pipe(2) for Windows which allows the pipe
 * handles to be used in select().
 *
 * Reads and writes on the pipe must go through piperead()/pipewrite().
 *
 * For consistency with Unix we declare the returned handles as "int".
 * This is okay even on WIN64 because system handles are not more than
 * 32 bits wide, but we do have to do some casting.
 */
static int
pgpipe(int handles[2])
{
    pgsocket    s,
                tmp_sock;
    struct sockaddr_in serv_addr;
    int         len = sizeof(serv_addr);

    /* We have to use the Unix socket invalid file descriptor value here. */
    handles[0] = handles[1] = -1;

    /*
     * setup listen socket
     */
    if ((s = socket(AF_INET, SOCK_STREAM, 0)) == PGINVALID_SOCKET)
    {
        pg_log_error("pgpipe: could not create socket: error code %d",
                     WSAGetLastError());
        return -1;
    }

    memset((void *) &serv_addr, 0, sizeof(serv_addr));
    serv_addr.sin_family = AF_INET;
    serv_addr.sin_port = pg_hton16(0);
    serv_addr.sin_addr.s_addr = pg_hton32(INADDR_LOOPBACK);
    if (bind(s, (SOCKADDR *) &serv_addr, len) == SOCKET_ERROR)
    {
        pg_log_error("pgpipe: could not bind: error code %d",
                     WSAGetLastError());
        closesocket(s);
        return -1;
    }
    if (listen(s, 1) == SOCKET_ERROR)
    {
        pg_log_error("pgpipe: could not listen: error code %d",
                     WSAGetLastError());
        closesocket(s);
        return -1;
    }
    if (getsockname(s, (SOCKADDR *) &serv_addr, &len) == SOCKET_ERROR)
    {
        pg_log_error("pgpipe: getsockname() failed: error code %d",
                     WSAGetLastError());
        closesocket(s);
        return -1;
    }

    /*
     * setup pipe handles
     */
    if ((tmp_sock = socket(AF_INET, SOCK_STREAM, 0)) == PGINVALID_SOCKET)
    {
        pg_log_error("pgpipe: could not create second socket: error code %d",
                     WSAGetLastError());
        closesocket(s);
        return -1;
    }
    handles[1] = (int) tmp_sock;

    if (connect(handles[1], (SOCKADDR *) &serv_addr, len) == SOCKET_ERROR)
    {
        pg_log_error("pgpipe: could not connect socket: error code %d",
                     WSAGetLastError());
        closesocket(handles[1]);
        handles[1] = -1;
        closesocket(s);
        return -1;
    }
    if ((tmp_sock = accept(s, (SOCKADDR *) &serv_addr, &len)) == PGINVALID_SOCKET)
    {
        pg_log_error("pgpipe: could not accept connection: error code %d",
                     WSAGetLastError());
        closesocket(handles[1]);
        handles[1] = -1;
        closesocket(s);
        return -1;
    }
    handles[0] = (int) tmp_sock;

    closesocket(s);
    return 0;
}

/*
 * Windows implementation of reading from a pipe.
 */
static int
piperead(int s, char *buf, int len)
{
    int         ret = recv(s, buf, len, 0);

    if (ret < 0 && WSAGetLastError() == WSAECONNRESET)
    {
        /* EOF on the pipe! */
        ret = 0;
    }
    return ret;
}

#endif                          /* WIN32 */