postgresql/src/bin/pg_dump/parallel.c

1363 lines
35 KiB
C
Raw Normal View History

/*-------------------------------------------------------------------------
*
* parallel.c
*
* Parallel support for the pg_dump archiver
*
* Portions Copyright (c) 1996-2016, PostgreSQL Global Development Group
* Portions Copyright (c) 1994, Regents of the University of California
*
* The author is not responsible for loss or damages that may
* result from its use.
*
* IDENTIFICATION
* src/bin/pg_dump/parallel.c
*
*-------------------------------------------------------------------------
*/
#include "postgres_fe.h"
#include "parallel.h"
#include "pg_backup_utils.h"
#include "fe_utils/string_utils.h"
#ifndef WIN32
#include <sys/types.h>
#include <sys/wait.h>
#include "signal.h"
#include <unistd.h>
#include <fcntl.h>
#endif
#define PIPE_READ 0
#define PIPE_WRITE 1
/* file-scope variables */
#ifdef WIN32
static unsigned int tMasterThreadId = 0;
static HANDLE termEvent = INVALID_HANDLE_VALUE;
static int pgpipe(int handles[2]);
static int piperead(int s, char *buf, int len);
/*
* Structure to hold info passed by _beginthreadex() to the function it calls
* via its single allowed argument.
*/
typedef struct
{
ArchiveHandle *AH;
int worker;
int pipeRead;
int pipeWrite;
} WorkerInfo;
#define pipewrite(a,b,c) send(a,b,c,0)
#else
/*
* aborting is only ever used in the master, the workers are fine with just
* wantAbort.
*/
static bool aborting = false;
static volatile sig_atomic_t wantAbort = 0;
#define pgpipe(a) pipe(a)
#define piperead(a,b,c) read(a,b,c)
#define pipewrite(a,b,c) write(a,b,c)
#endif
typedef struct ShutdownInformation
{
ParallelState *pstate;
Archive *AHX;
} ShutdownInformation;
static ShutdownInformation shutdown_info;
static const char *modulename = gettext_noop("parallel archiver");
static ParallelSlot *GetMyPSlot(ParallelState *pstate);
static void archive_close_connection(int code, void *arg);
static void ShutdownWorkersHard(ParallelState *pstate);
static void WaitForTerminatingWorkers(ParallelState *pstate);
#ifndef WIN32
static void sigTermHandler(int signum);
#endif
static void SetupWorker(ArchiveHandle *AH, int pipefd[2], int worker);
static bool HasEveryWorkerTerminated(ParallelState *pstate);
static void lockTableNoWait(ArchiveHandle *AH, TocEntry *te);
static void WaitForCommands(ArchiveHandle *AH, int pipefd[2]);
static char *getMessageFromMaster(int pipefd[2]);
static void sendMessageToMaster(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)
#define messageEquals(msg, pattern) \
(strcmp(msg, pattern) == 0)
#ifdef WIN32
static void shutdown_parallel_dump_utils(int code, void *unused);
bool parallel_init_done = false;
static DWORD tls_index;
DWORD mainThreadId;
#endif
#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
void
init_parallel_dump_utils(void)
{
#ifdef WIN32
if (!parallel_init_done)
{
WSADATA wsaData;
int err;
tls_index = TlsAlloc();
mainThreadId = GetCurrentThreadId();
err = WSAStartup(MAKEWORD(2, 2), &wsaData);
if (err != 0)
{
fprintf(stderr, _("%s: WSAStartup failed: %d\n"), progname, err);
exit_nicely(1);
}
on_exit_nicely(shutdown_parallel_dump_utils, NULL);
parallel_init_done = true;
}
#endif
}
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. (On Windows the memory leakage
* will be one buffer per thread, which is at least better than one per call).
*/
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.
*/
static PQExpBuffer s_id_return = NULL;
PQExpBuffer id_return;
#ifdef WIN32
if (parallel_init_done)
id_return = (PQExpBuffer) TlsGetValue(tls_index); /* 0 when not set */
else
id_return = s_id_return;
#else
id_return = s_id_return;
#endif
if (id_return) /* first time through? */
{
/* same buffer, just wipe contents */
resetPQExpBuffer(id_return);
}
else
{
/* new buffer */
id_return = createPQExpBuffer();
#ifdef WIN32
if (parallel_init_done)
TlsSetValue(tls_index, id_return);
else
s_id_return = id_return;
#else
s_id_return = id_return;
#endif
}
return id_return;
}
/*
* pg_dump and pg_restore register the Archive pointer for the exit handler
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
* (called from exit_nicely). This function mainly exists so that we can
* keep shutdown_info in file scope only.
*/
void
on_exit_close_archive(Archive *AHX)
{
shutdown_info.AHX = AHX;
on_exit_nicely(archive_close_connection, &shutdown_info);
}
/*
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
* 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)
{
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
/* In parallel mode, must figure out who we are */
ParallelSlot *slot = GetMyPSlot(si->pstate);
if (!slot)
{
/*
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
* We're the master. Close our own database connection, if any,
* and then forcibly shut down workers.
*/
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
if (si->AHX)
DisconnectDatabase(si->AHX);
#ifndef WIN32
/*
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
* Setting aborting to true shuts off error/warning messages that
* are no longer useful once we start killing workers.
*/
aborting = true;
#endif
ShutdownWorkersHard(si->pstate);
}
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
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 master would
* fail to detect it because there would be no EOF condition on
* the other end of the pipe.)
*/
if (slot->args->AH)
DisconnectDatabase(&(slot->args->AH->public));
#ifdef WIN32
closesocket(slot->pipeRevRead);
closesocket(slot->pipeRevWrite);
#endif
}
}
else
{
/* Non-parallel operation: just kill the master DB connection */
if (si->AHX)
DisconnectDatabase(si->AHX);
}
}
/*
* If we have one worker that terminates for some reason, we'd like the other
* threads to terminate as well (and not finish with their 70 GB table dump
* first...). Now in UNIX we can just kill these processes, and let the signal
* handler set wantAbort to 1. In Windows we set a termEvent and this serves
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
* as the signal for everyone to terminate. We don't print any error message,
* that would just clutter the screen.
*/
void
checkAborting(ArchiveHandle *AH)
{
#ifdef WIN32
if (WaitForSingleObject(termEvent, 0) == WAIT_OBJECT_0)
#else
if (wantAbort)
#endif
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
exit_nicely(1);
}
/*
* Shut down any remaining workers, waiting for them to finish.
*/
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.
*/
for (i = 0; i < pstate->numWorkers; i++)
closesocket(pstate->parallelSlot[i].pipeWrite);
#ifndef WIN32
/* On non-Windows, send SIGTERM to abort commands-in-progress. */
for (i = 0; i < pstate->numWorkers; i++)
kill(pstate->parallelSlot[i].pid, SIGTERM);
#else
/* Non-idle workers monitor this event via checkAborting(). */
SetEvent(termEvent);
#endif
WaitForTerminatingWorkers(pstate);
}
/*
* Wait for the termination of the processes using the OS-specific method.
*/
static void
WaitForTerminatingWorkers(ParallelState *pstate)
{
while (!HasEveryWorkerTerminated(pstate))
{
ParallelSlot *slot = NULL;
int j;
#ifndef WIN32
int status;
pid_t pid = wait(&status);
for (j = 0; j < pstate->numWorkers; j++)
if (pstate->parallelSlot[j].pid == pid)
slot = &(pstate->parallelSlot[j]);
#else
uintptr_t hThread;
DWORD ret;
uintptr_t *lpHandles = pg_malloc(sizeof(HANDLE) * pstate->numWorkers);
int nrun = 0;
for (j = 0; j < pstate->numWorkers; j++)
if (pstate->parallelSlot[j].workerStatus != WRKR_TERMINATED)
{
lpHandles[nrun] = pstate->parallelSlot[j].hThread;
nrun++;
}
ret = WaitForMultipleObjects(nrun, (HANDLE *) lpHandles, false, INFINITE);
Assert(ret != WAIT_FAILED);
hThread = lpHandles[ret - WAIT_OBJECT_0];
for (j = 0; j < pstate->numWorkers; j++)
if (pstate->parallelSlot[j].hThread == hThread)
slot = &(pstate->parallelSlot[j]);
free(lpHandles);
#endif
Assert(slot);
slot->workerStatus = WRKR_TERMINATED;
}
Assert(HasEveryWorkerTerminated(pstate));
}
#ifndef WIN32
/* Signal handling (UNIX only) */
static void
sigTermHandler(int signum)
{
wantAbort = 1;
}
#endif
/*
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
* 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
SetupWorker(ArchiveHandle *AH, int pipefd[2], int worker)
{
/*
* Call the setup worker function that's defined in the ArchiveHandle.
*/
(AH->SetupWorkerPtr) ((Archive *) AH);
Assert(AH->connection != NULL);
WaitForCommands(AH, pipefd);
}
#ifdef WIN32
static unsigned __stdcall
init_spawned_worker_win32(WorkerInfo *wi)
{
ArchiveHandle *AH;
int pipefd[2] = {wi->pipeRead, wi->pipeWrite};
int worker = wi->worker;
AH = CloneArchive(wi->AH);
free(wi);
SetupWorker(AH, pipefd, worker);
DeCloneArchive(AH);
_endthreadex(0);
return 0;
}
#endif
/*
* This function starts the parallel dump or restore by spawning off the
* worker processes in both Unix and Windows. For Windows, it creates a number
* of threads while it does a fork() on Unix.
*/
ParallelState *
ParallelBackupStart(ArchiveHandle *AH)
{
ParallelState *pstate;
int i;
const size_t slotSize = AH->public.numWorkers * sizeof(ParallelSlot);
Assert(AH->public.numWorkers > 0);
/* Ensure stdio state is quiesced before forking */
fflush(NULL);
pstate = (ParallelState *) pg_malloc(sizeof(ParallelState));
pstate->numWorkers = AH->public.numWorkers;
pstate->parallelSlot = NULL;
if (AH->public.numWorkers == 1)
return pstate;
pstate->parallelSlot = (ParallelSlot *) pg_malloc(slotSize);
memset((void *) pstate->parallelSlot, 0, slotSize);
/*
* Set the pstate in the shutdown_info. The exit handler uses pstate if
* set and falls back to AHX otherwise.
*/
shutdown_info.pstate = pstate;
getLocalPQExpBuffer = getThreadLocalPQExpBuffer;
#ifdef WIN32
tMasterThreadId = GetCurrentThreadId();
termEvent = CreateEvent(NULL, true, false, "Terminate");
#else
signal(SIGTERM, sigTermHandler);
signal(SIGINT, sigTermHandler);
signal(SIGQUIT, sigTermHandler);
#endif
for (i = 0; i < pstate->numWorkers; i++)
{
#ifdef WIN32
WorkerInfo *wi;
uintptr_t handle;
#else
pid_t pid;
#endif
int pipeMW[2],
pipeWM[2];
if (pgpipe(pipeMW) < 0 || pgpipe(pipeWM) < 0)
exit_horribly(modulename,
"could not create communication channels: %s\n",
strerror(errno));
pstate->parallelSlot[i].workerStatus = WRKR_IDLE;
pstate->parallelSlot[i].args = (ParallelArgs *) pg_malloc(sizeof(ParallelArgs));
pstate->parallelSlot[i].args->AH = NULL;
pstate->parallelSlot[i].args->te = NULL;
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
/* master's ends of the pipes */
pstate->parallelSlot[i].pipeRead = pipeWM[PIPE_READ];
pstate->parallelSlot[i].pipeWrite = pipeMW[PIPE_WRITE];
/* child's ends of the pipes */
pstate->parallelSlot[i].pipeRevRead = pipeMW[PIPE_READ];
pstate->parallelSlot[i].pipeRevWrite = pipeWM[PIPE_WRITE];
#ifdef WIN32
/* Allocate a new structure for every worker */
wi = (WorkerInfo *) pg_malloc(sizeof(WorkerInfo));
wi->worker = i;
wi->AH = AH;
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
wi->pipeRead = pipeMW[PIPE_READ];
wi->pipeWrite = pipeWM[PIPE_WRITE];
handle = _beginthreadex(NULL, 0, (void *) &init_spawned_worker_win32,
wi, 0, &(pstate->parallelSlot[i].threadId));
pstate->parallelSlot[i].hThread = handle;
#else
pid = fork();
if (pid == 0)
{
/* we are the worker */
int j;
int pipefd[2];
pipefd[0] = pipeMW[PIPE_READ];
pipefd[1] = pipeWM[PIPE_WRITE];
pstate->parallelSlot[i].pid = getpid();
/*
* Call CloneArchive on Unix as well 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 (for parallel dump)
* which both seem kinda helpful.
*/
pstate->parallelSlot[i].args->AH = CloneArchive(AH);
/* close read end of Worker -> Master */
closesocket(pipeWM[PIPE_READ]);
/* close write end of Master -> Worker */
closesocket(pipeMW[PIPE_WRITE]);
/*
* Close all inherited fds for communication of the master with
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
* previously-forked workers.
*/
for (j = 0; j < i; j++)
{
closesocket(pstate->parallelSlot[j].pipeRead);
closesocket(pstate->parallelSlot[j].pipeWrite);
}
SetupWorker(pstate->parallelSlot[i].args->AH, pipefd, i);
exit(0);
}
else if (pid < 0)
/* fork failed */
exit_horribly(modulename,
"could not create worker process: %s\n",
strerror(errno));
/* we are the Master, pid > 0 here */
Assert(pid > 0);
/* close read end of Master -> Worker */
closesocket(pipeMW[PIPE_READ]);
/* close write end of Worker -> Master */
closesocket(pipeWM[PIPE_WRITE]);
pstate->parallelSlot[i].pid = pid;
#endif
}
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
/*
* Having forked off the workers, disable SIGPIPE so that master isn't
* killed if it tries to send a command to a dead worker.
*/
#ifndef WIN32
signal(SIGPIPE, SIG_IGN);
#endif
return pstate;
}
/*
* Tell all of our workers to terminate.
*
* Pretty straightforward routine, first we tell everyone to terminate, then
* we listen to the workers' replies and finally close the sockets that we
* have used for communication.
*/
void
ParallelBackupEnd(ArchiveHandle *AH, ParallelState *pstate)
{
int i;
if (pstate->numWorkers == 1)
return;
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);
}
WaitForTerminatingWorkers(pstate);
/*
* Remove the pstate again, so the exit handler in the parent will now
* again fall back to closing AH->connection (if connected).
*/
shutdown_info.pstate = NULL;
free(pstate->parallelSlot);
free(pstate);
}
/*
* The sequence is the following (for dump, similar for restore):
*
* The master process starts the parallel backup in ParllelBackupStart, this
* forks the worker processes which enter WaitForCommand().
*
* The master process dispatches an individual work item to one of the worker
* processes in DispatchJobForTocEntry(). It calls
* AH->MasterStartParallelItemPtr, a routine of the output format. This
* function's arguments are the parents archive handle AH (containing the full
* catalog information), the TocEntry that the worker should work on and a
* T_Action act indicating whether this is a backup or a restore item. The
* function then converts the TocEntry assignment into a string that is then
* sent over to the worker process. In the simplest case that would be
* something like "DUMP 1234", with 1234 being the TocEntry id.
*
* The worker receives the message in the routine pointed to by
* WorkerJobDumpPtr or WorkerJobRestorePtr. These are also pointers to
* corresponding routines of the respective output format, e.g.
* _WorkerJobDumpDirectory().
*
* 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. Now they re-translate the textual representation to a TocEntry
* on their side and do the required action (restore or dump).
*
* The result is again a textual string that is sent back to the master and is
* interpreted by AH->MasterEndParallelItemPtr. This function can update state
* or catalog information on the master's side, depending on the reply from
* the worker process. In the end it returns status which is 0 for successful
* execution.
*
* ---------------------------------------------------------------------
* Master Worker
*
* enters WaitForCommands()
* DispatchJobForTocEntry(...te...)
*
* [ Worker is IDLE ]
*
* arg = (MasterStartParallelItemPtr)()
* send: DUMP arg
* receive: DUMP arg
* str = (WorkerJobDumpPtr)(arg)
* [ Worker is WORKING ] ... gets te from arg ...
* ... dump te ...
* send: OK DUMP info
*
* In ListenToWorkers():
*
* [ Worker is FINISHED ]
* receive: OK DUMP info
* status = (MasterEndParallelItemPtr)(info)
*
* In ReapWorkerStatus(&ptr):
* *ptr = status;
* [ Worker is IDLE ]
* ---------------------------------------------------------------------
*/
void
DispatchJobForTocEntry(ArchiveHandle *AH, ParallelState *pstate, TocEntry *te,
T_Action act)
{
int worker;
char *arg;
/* our caller makes sure that at least one worker is idle */
Assert(GetIdleWorker(pstate) != NO_SLOT);
worker = GetIdleWorker(pstate);
Assert(worker != NO_SLOT);
arg = (AH->MasterStartParallelItemPtr) (AH, te, act);
sendMessageToWorker(pstate, worker, arg);
pstate->parallelSlot[worker].workerStatus = WRKR_WORKING;
pstate->parallelSlot[worker].args->te = te;
}
/*
* Find the first free parallel slot (if any).
*/
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 every worker process is in the WRKR_TERMINATED state.
*/
static bool
HasEveryWorkerTerminated(ParallelState *pstate)
{
int i;
for (i = 0; i < pstate->numWorkers; i++)
if (pstate->parallelSlot[i].workerStatus != WRKR_TERMINATED)
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;
}
/*
* ---------------------------------------------------------------------
* One danger of the parallel backup is a possible deadlock:
*
* 1) Master 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 master holds a conflicting ACCESS SHARE lock).
* 3) The worker process also requests an ACCESS SHARE lock to read the table.
* The worker's not granted that lock but is enqueued behind the ACCESS
* EXCLUSIVE lock request.
* ---------------------------------------------------------------------
*
* Now what we do here is to just request a lock in ACCESS SHARE but with
* NOWAIT in the worker prior to touching the table. If we don't get the lock,
* then we know that somebody else has requested an ACCESS EXCLUSIVE lock and
* are good to just fail the whole backup because we have detected a deadlock.
*/
static void
lockTableNoWait(ArchiveHandle *AH, TocEntry *te)
{
Archive *AHX = (Archive *) AH;
const char *qualId;
PQExpBuffer query = createPQExpBuffer();
PGresult *res;
Assert(AH->format == archDirectory);
Assert(strcmp(te->desc, "BLOBS") != 0);
appendPQExpBuffer(query,
"SELECT pg_namespace.nspname,"
" pg_class.relname "
" FROM pg_class "
" JOIN pg_namespace on pg_namespace.oid = relnamespace "
" WHERE pg_class.oid = %u", te->catalogId.oid);
res = PQexec(AH->connection, query->data);
if (!res || PQresultStatus(res) != PGRES_TUPLES_OK)
exit_horribly(modulename,
"could not get relation name for OID %u: %s\n",
te->catalogId.oid, PQerrorMessage(AH->connection));
resetPQExpBuffer(query);
qualId = fmtQualifiedId(AHX->remoteVersion,
PQgetvalue(res, 0, 0),
PQgetvalue(res, 0, 1));
appendPQExpBuffer(query, "LOCK TABLE %s IN ACCESS SHARE MODE NOWAIT",
qualId);
PQclear(res);
res = PQexec(AH->connection, query->data);
if (!res || PQresultStatus(res) != PGRES_COMMAND_OK)
exit_horribly(modulename,
"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.\n", qualId);
PQclear(res);
destroyPQExpBuffer(query);
}
/*
* That's the main routine for the worker.
* When it starts up it enters this routine and waits for commands from the
* master process. After having processed a command it comes back to here to
* wait for the next command. Finally it will receive a TERMINATE command and
* exit.
*/
static void
WaitForCommands(ArchiveHandle *AH, int pipefd[2])
{
char *command;
DumpId dumpId;
int nBytes;
char *str = NULL;
TocEntry *te;
for (;;)
{
if (!(command = getMessageFromMaster(pipefd)))
{
PQfinish(AH->connection);
AH->connection = NULL;
return;
}
if (messageStartsWith(command, "DUMP "))
{
Assert(AH->format == archDirectory);
sscanf(command + strlen("DUMP "), "%d%n", &dumpId, &nBytes);
Assert(nBytes == strlen(command) - strlen("DUMP "));
te = getTocEntryByDumpId(AH, dumpId);
Assert(te != NULL);
/*
* Lock the table but with NOWAIT. Note that the parent is already
* holding a lock. If we cannot acquire another ACCESS SHARE MODE
* lock, then somebody else has requested an exclusive lock in the
* meantime. lockTableNoWait dies in this case to prevent a
* deadlock.
*/
if (strcmp(te->desc, "BLOBS") != 0)
lockTableNoWait(AH, te);
/*
* The message we return here has been pg_malloc()ed and we are
* responsible for free()ing it.
*/
str = (AH->WorkerJobDumpPtr) (AH, te);
Assert(AH->connection != NULL);
sendMessageToMaster(pipefd, str);
free(str);
}
else if (messageStartsWith(command, "RESTORE "))
{
Assert(AH->format == archDirectory || AH->format == archCustom);
Assert(AH->connection != NULL);
sscanf(command + strlen("RESTORE "), "%d%n", &dumpId, &nBytes);
Assert(nBytes == strlen(command) - strlen("RESTORE "));
te = getTocEntryByDumpId(AH, dumpId);
Assert(te != NULL);
/*
* The message we return here has been pg_malloc()ed and we are
* responsible for free()ing it.
*/
str = (AH->WorkerJobRestorePtr) (AH, te);
Assert(AH->connection != NULL);
sendMessageToMaster(pipefd, str);
free(str);
}
else
exit_horribly(modulename,
"unrecognized command on communication channel: %s\n",
command);
/* command was pg_malloc'd and we are responsible for free()ing it. */
free(command);
}
}
/*
* ---------------------------------------------------------------------
* Note the status change:
*
* DispatchJobForTocEntry WRKR_IDLE -> WRKR_WORKING
* ListenToWorkers WRKR_WORKING -> WRKR_FINISHED / WRKR_TERMINATED
* ReapWorkerStatus WRKR_FINISHED -> WRKR_IDLE
* ---------------------------------------------------------------------
*
* Just calling ReapWorkerStatus() when all workers are working might or might
* not give you an idle worker because you need to call ListenToWorkers() in
* between and only thereafter ReapWorkerStatus(). This is necessary in order
* to get and deal with the status (=result) of the worker's execution.
*/
void
ListenToWorkers(ArchiveHandle *AH, ParallelState *pstate, bool do_wait)
{
int worker;
char *msg;
msg = getMessageFromWorker(pstate, do_wait, &worker);
if (!msg)
{
if (do_wait)
exit_horribly(modulename, "a worker process died unexpectedly\n");
return;
}
if (messageStartsWith(msg, "OK "))
{
char *statusString;
TocEntry *te;
pstate->parallelSlot[worker].workerStatus = WRKR_FINISHED;
te = pstate->parallelSlot[worker].args->te;
if (messageStartsWith(msg, "OK RESTORE "))
{
statusString = msg + strlen("OK RESTORE ");
pstate->parallelSlot[worker].status =
(AH->MasterEndParallelItemPtr)
(AH, te, statusString, ACT_RESTORE);
}
else if (messageStartsWith(msg, "OK DUMP "))
{
statusString = msg + strlen("OK DUMP ");
pstate->parallelSlot[worker].status =
(AH->MasterEndParallelItemPtr)
(AH, te, statusString, ACT_DUMP);
}
else
exit_horribly(modulename,
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
"invalid message received from worker: \"%s\"\n",
msg);
}
else
Fix broken error handling in parallel pg_dump/pg_restore. In the original design for parallel dump, worker processes reported errors by sending them up to the master process, which would print the messages. This is unworkably fragile for a couple of reasons: it risks deadlock if a worker sends an error at an unexpected time, and if the master has already died for some reason, the user will never get to see the error at all. Revert that idea and go back to just always printing messages to stderr. This approach means that if all the workers fail for similar reasons (eg, bad password or server shutdown), the user will see N copies of that message, not only one as before. While that's slightly annoying, it's certainly better than not seeing any message; not to mention that we shouldn't assume that only the first failure is interesting. An additional problem in the same area was that the master failed to disable SIGPIPE (at least until much too late), which meant that sending a command to an already-dead worker would cause the master to crash silently. That was bad enough in itself but was made worse by the total reliance on the master to print errors: even if the worker had reported an error, you would probably not see it, depending on timing. Instead disable SIGPIPE right after we've forked the workers, before attempting to send them anything. Additionally, the master relies on seeing socket EOF to realize that a worker has exited prematurely --- but on Windows, there would be no EOF since the socket is attached to the process that includes both the master and worker threads, so it remains open. Make archive_close_connection() close the worker end of the sockets so that this acts more like the Unix case. It's not perfect, because if a worker thread exits without going through exit_nicely() the closures won't happen; but that's not really supposed to happen. This has been wrong all along, so back-patch to 9.3 where parallel dump was introduced. Report: <2458.1450894615@sss.pgh.pa.us>
2016-05-25 18:39:57 +02:00
exit_horribly(modulename,
"invalid message received from worker: \"%s\"\n",
msg);
/* both Unix and Win32 return pg_malloc()ed space, so we free it */
free(msg);
}
/*
* This function is executed in the master process.
*
* This function is used to get the return value of a terminated worker
* process. If a process has terminated, its status is stored in *status and
* the id of the worker is returned.
*/
int
ReapWorkerStatus(ParallelState *pstate, int *status)
{
int i;
for (i = 0; i < pstate->numWorkers; i++)
{
if (pstate->parallelSlot[i].workerStatus == WRKR_FINISHED)
{
*status = pstate->parallelSlot[i].status;
pstate->parallelSlot[i].status = 0;
pstate->parallelSlot[i].workerStatus = WRKR_IDLE;
return i;
}
}
return NO_SLOT;
}
/*
* This function is executed in the master process.
*
* It looks for an idle worker process and only returns if there is one.
*/
void
EnsureIdleWorker(ArchiveHandle *AH, ParallelState *pstate)
{
int ret_worker;
int work_status;
for (;;)
{
int nTerm = 0;
while ((ret_worker = ReapWorkerStatus(pstate, &work_status)) != NO_SLOT)
{
if (work_status != 0)
exit_horribly(modulename, "error processing a parallel work item\n");
nTerm++;
}
/*
* We need to make sure that we have an idle worker before dispatching
* the next item. If nTerm > 0 we already have that (quick check).
*/
if (nTerm > 0)
return;
/* explicit check for an idle worker */
if (GetIdleWorker(pstate) != NO_SLOT)
return;
/*
* If we have no idle worker, read the result of one or more workers
* and loop the loop to call ReapWorkerStatus() on them
*/
ListenToWorkers(AH, pstate, true);
}
}
/*
* This function is executed in the master process.
*
* It waits for all workers to terminate.
*/
void
EnsureWorkersFinished(ArchiveHandle *AH, ParallelState *pstate)
{
int work_status;
if (!pstate || pstate->numWorkers == 1)
return;
/* Waiting for the remaining worker processes to finish */
while (!IsEveryWorkerIdle(pstate))
{
if (ReapWorkerStatus(pstate, &work_status) == NO_SLOT)
ListenToWorkers(AH, pstate, true);
else if (work_status != 0)
exit_horribly(modulename,
"error processing a parallel work item\n");
}
}
/*
* This function is executed in the worker process.
*
* It returns the next message on the communication channel, blocking until it
* becomes available.
*/
static char *
getMessageFromMaster(int pipefd[2])
{
return readMessageFromPipe(pipefd[PIPE_READ]);
}
/*
* This function is executed in the worker process.
*
* It sends a message to the master on the communication channel.
*/
static void
sendMessageToMaster(int pipefd[2], const char *str)
{
int len = strlen(str) + 1;
if (pipewrite(pipefd[PIPE_WRITE], str, len) != len)
exit_horribly(modulename,
"could not write to the communication channel: %s\n",
strerror(errno));
}
/*
* A select loop that repeats calling select until a descriptor in the read
* set becomes readable. On Windows we have to check for the termination event
* from time to time, on Unix we can just block forever.
*/
static int
select_loop(int maxFd, fd_set *workerset)
{
int i;
fd_set saveSet = *workerset;
#ifdef WIN32
/* should always be the master */
Assert(tMasterThreadId == GetCurrentThreadId());
for (;;)
{
/*
* sleep a quarter of a second before checking if we should terminate.
*/
struct timeval tv = {0, 250000};
*workerset = saveSet;
i = select(maxFd + 1, workerset, NULL, NULL, &tv);
if (i == SOCKET_ERROR && WSAGetLastError() == WSAEINTR)
continue;
if (i)
break;
}
#else /* UNIX */
for (;;)
{
*workerset = saveSet;
i = select(maxFd + 1, workerset, NULL, NULL, NULL);
/*
* If we Ctrl-C the master process, it's likely that we interrupt
* select() here. The signal handler will set wantAbort == true and
* the shutdown journey starts from here. Note that we'll come back
* here later when we tell all workers to terminate and read their
* responses. But then we have aborting set to true.
*/
if (wantAbort && !aborting)
exit_horribly(modulename, "terminated by user\n");
if (i < 0 && errno == EINTR)
continue;
break;
}
#endif
return i;
}
/*
* This function is executed in the master process.
*
* It returns the next message from the worker on the communication channel,
* optionally blocking (do_wait) until it becomes available.
*
* The id of the worker is returned in *worker.
*/
static char *
getMessageFromWorker(ParallelState *pstate, bool do_wait, int *worker)
{
int i;
fd_set workerset;
int maxFd = -1;
struct timeval nowait = {0, 0};
FD_ZERO(&workerset);
for (i = 0; i < pstate->numWorkers; i++)
{
if (pstate->parallelSlot[i].workerStatus == WRKR_TERMINATED)
continue;
FD_SET(pstate->parallelSlot[i].pipeRead, &workerset);
/* actually WIN32 ignores the first parameter to select()... */
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)
exit_horribly(modulename, "error in ListenToWorkers(): %s\n", strerror(errno));
for (i = 0; i < pstate->numWorkers; i++)
{
char *msg;
if (!FD_ISSET(pstate->parallelSlot[i].pipeRead, &workerset))
continue;
msg = readMessageFromPipe(pstate->parallelSlot[i].pipeRead);
*worker = i;
return msg;
}
Assert(false);
return NULL;
}
/*
* This function is executed in the master process.
*
* It sends a message to a certain worker on the communication channel.
*/
static void
sendMessageToWorker(ParallelState *pstate, int worker, const char *str)
{
int len = strlen(str) + 1;
if (pipewrite(pstate->parallelSlot[worker].pipeWrite, str, len) != len)
{
/*
* If we're already aborting anyway, don't care if we succeed or not.
* The child might have gone already.
*/
#ifndef WIN32
if (!aborting)
#endif
exit_horribly(modulename,
"could not write to the communication channel: %s\n",
strerror(errno));
}
}
/*
* The underlying function to read a message from the communication channel
* (fd) with optional blocking (do_wait).
*/
static char *
readMessageFromPipe(int fd)
{
char *msg;
int msgsize,
bufsize;
int ret;
/*
2016-03-15 23:06:11 +01:00
* The problem here is that we need to deal with several possibilities: we
* could receive only a partial message or several messages at once. The
* caller expects us to return exactly one message however.
*
* We could either read in as much as we can and keep track of what we
* delivered back to the caller or we just read byte by byte. Once we see
* (char) 0, we know that it's the message's end. This would be quite
* inefficient for more data but since we are reading only on the command
* channel, the performance loss does not seem worth the trouble of
* keeping internal states for different file descriptors.
*/
bufsize = 64; /* could be any number */
msg = (char *) pg_malloc(bufsize);
msgsize = 0;
for (;;)
{
Assert(msgsize <= bufsize);
ret = piperead(fd, msg + msgsize, 1);
/* worker has closed the connection or another error happened */
if (ret <= 0)
break;
Assert(ret == 1);
if (msg[msgsize] == '\0')
return msg;
msgsize++;
if (msgsize == bufsize)
{
/* could be any number */
bufsize += 16;
msg = (char *) pg_realloc(msg, bufsize);
}
}
/*
* Worker has closed the connection, make sure to clean up before return
* since we are not returning msg (but did allocate it).
*/
pg_free(msg);
return NULL;
}
#ifdef WIN32
/*
* This is a replacement version of pipe for Win32 which allows returned
* handles to be used in select(). Note that read/write calls must be replaced
* with recv/send. "handles" have to be integers so we check for errors then
* cast to integers.
*/
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)
{
write_msg(modulename, "pgpipe: could not create socket: error code %d\n",
WSAGetLastError());
return -1;
}
memset((void *) &serv_addr, 0, sizeof(serv_addr));
serv_addr.sin_family = AF_INET;
serv_addr.sin_port = htons(0);
serv_addr.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
if (bind(s, (SOCKADDR *) &serv_addr, len) == SOCKET_ERROR)
{
write_msg(modulename, "pgpipe: could not bind: error code %d\n",
WSAGetLastError());
closesocket(s);
return -1;
}
if (listen(s, 1) == SOCKET_ERROR)
{
write_msg(modulename, "pgpipe: could not listen: error code %d\n",
WSAGetLastError());
closesocket(s);
return -1;
}
if (getsockname(s, (SOCKADDR *) &serv_addr, &len) == SOCKET_ERROR)
{
write_msg(modulename, "pgpipe: getsockname() failed: error code %d\n",
WSAGetLastError());
closesocket(s);
return -1;
}
/*
* setup pipe handles
*/
if ((tmp_sock = socket(AF_INET, SOCK_STREAM, 0)) == PGINVALID_SOCKET)
{
write_msg(modulename, "pgpipe: could not create second socket: error code %d\n",
WSAGetLastError());
closesocket(s);
return -1;
}
handles[1] = (int) tmp_sock;
if (connect(handles[1], (SOCKADDR *) &serv_addr, len) == SOCKET_ERROR)
{
write_msg(modulename, "pgpipe: could not connect socket: error code %d\n",
WSAGetLastError());
closesocket(s);
return -1;
}
if ((tmp_sock = accept(s, (SOCKADDR *) &serv_addr, &len)) == PGINVALID_SOCKET)
{
write_msg(modulename, "pgpipe: could not accept connection: error code %d\n",
WSAGetLastError());
closesocket(handles[1]);
handles[1] = -1;
closesocket(s);
return -1;
}
handles[0] = (int) tmp_sock;
closesocket(s);
return 0;
}
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! (win32 socket based implementation) */
ret = 0;
return ret;
}
#endif