]> granicus.if.org Git - postgresql/commitdiff
Redesign handling of SIGTERM/control-C in parallel pg_dump/pg_restore.
authorTom Lane <tgl@sss.pgh.pa.us>
Thu, 2 Jun 2016 17:27:53 +0000 (13:27 -0400)
committerTom Lane <tgl@sss.pgh.pa.us>
Thu, 2 Jun 2016 17:28:17 +0000 (13:28 -0400)
Formerly, Unix builds of pg_dump/pg_restore would trap SIGINT and similar
signals and set a flag that was tested in various data-transfer loops.
This was prone to errors of omission (cf commit 3c8aa6654); and even if
the client-side response was prompt, we did nothing that would cause
long-running SQL commands (e.g. CREATE INDEX) to terminate early.
Also, the master process would effectively do nothing at all upon receipt
of SIGINT; the only reason it seemed to work was that in typical scenarios
the signal would also be delivered to the child processes.  We should
support termination when a signal is delivered only to the master process,
though.

Windows builds had no console interrupt handler, so they would just fall
over immediately at control-C, again leaving long-running SQL commands to
finish unmolested.

To fix, remove the flag-checking approach altogether.  Instead, allow the
Unix signal handler to send a cancel request directly and then exit(1).
In the master process, also have it forward the signal to the children.
On Windows, add a console interrupt handler that behaves approximately
the same.  The main difference is that a single execution of the Windows
handler can send all the cancel requests since all the info is available
in one process, whereas on Unix each process sends a cancel only for its
own database connection.

In passing, fix an old problem that DisconnectDatabase tends to send a
cancel request before exiting a parallel worker, even if nothing went
wrong.  This is at least a waste of cycles, and could lead to unexpected
log messages, or maybe even data loss if it happened in pg_restore (though
in the current code the problem seems to affect only pg_dump).  The cause
was that after a COPY step, pg_dump was leaving libpq in PGASYNC_BUSY
state, causing PQtransactionStatus() to report PQTRANS_ACTIVE.  That's
normally harmless because the next PQexec() will silently clear the
PGASYNC_BUSY state; but in a parallel worker we might exit without any
additional SQL commands after a COPY step.  So add an extra PQgetResult()
call after a COPY to allow libpq to return to PGASYNC_IDLE state.

This is a bug fix, IMO, so back-patch to 9.3 where parallel dump/restore
were introduced.

Thanks to Kyotaro Horiguchi for Windows testing and code suggestions.

Original-Patch: <7005.1464657274@sss.pgh.pa.us>
Discussion: <20160602.174941.256342236.horiguchi.kyotaro@lab.ntt.co.jp>

src/bin/pg_dump/compress_io.c
src/bin/pg_dump/parallel.c
src/bin/pg_dump/parallel.h
src/bin/pg_dump/pg_backup_archiver.c
src/bin/pg_dump/pg_backup_archiver.h
src/bin/pg_dump/pg_backup_db.c
src/bin/pg_dump/pg_backup_directory.c
src/bin/pg_dump/pg_dump.c

index 4cf0935663e14f2f414d4eb4e157d76555020581..e0ef9cd60ac896ccfc7cce7518fe7a9cc4c2bd80 100644 (file)
@@ -54,7 +54,6 @@
 #include "postgres_fe.h"
 
 #include "compress_io.h"
-#include "parallel.h"
 #include "pg_backup_utils.h"
 
 /*----------------------
@@ -184,9 +183,6 @@ void
 WriteDataToArchive(ArchiveHandle *AH, CompressorState *cs,
                                   const void *data, size_t dLen)
 {
-       /* Are we aborting? */
-       checkAborting(AH);
-
        switch (cs->comprAlg)
        {
                case COMPR_ALG_LIBZ:
@@ -351,9 +347,6 @@ ReadDataFromArchiveZlib(ArchiveHandle *AH, ReadFunc readF)
        /* no minimal chunk size for zlib */
        while ((cnt = readF(AH, &buf, &buflen)))
        {
-               /* Are we aborting? */
-               checkAborting(AH);
-
                zp->next_in = (void *) buf;
                zp->avail_in = cnt;
 
@@ -414,9 +407,6 @@ ReadDataFromArchiveNone(ArchiveHandle *AH, ReadFunc readF)
 
        while ((cnt = readF(AH, &buf, &buflen)))
        {
-               /* Are we aborting? */
-               checkAborting(AH);
-
                ahwrite(buf, 1, cnt, AH);
        }
 
index 6a808dcbf25568ad988fac78ca2ba3a29f06e5af..7647271c48be45842a8ffbf2665440709626fc6e 100644 (file)
@@ -43,7 +43,7 @@
  * 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 init_spawned_worker_win32().)
+ * data structure; see RunWorker().)
  *
  * In the master process, the workerStatus field for each worker has one of
  * the following values:
@@ -83,9 +83,8 @@
  */
 typedef struct
 {
-       ArchiveHandle *AH;
-       int                     pipeRead;
-       int                     pipeWrite;
+       ArchiveHandle *AH;                      /* master database connection */
+       ParallelSlot *slot;                     /* this worker's parallel slot */
 } WorkerInfo;
 
 /* Windows implementation of pipe access */
@@ -95,9 +94,6 @@ static int    piperead(int s, char *buf, int len);
 
 #else                                                  /* !WIN32 */
 
-/* Signal handler flag */
-static volatile sig_atomic_t wantAbort = 0;
-
 /* Non-Windows implementation of pipe access */
 #define pgpipe(a)                      pipe(a)
 #define piperead(a,b,c)                read(a,b,c)
@@ -116,10 +112,37 @@ typedef struct ShutdownInformation
 
 static ShutdownInformation shutdown_info;
 
+/*
+ * State info for signal handling.
+ * We assume signal_info initializes to zeroes.
+ *
+ * On Unix, myAH is the master DB connection in the master process, and the
+ * worker's own connection in worker processes.  On Windows, we have only one
+ * instance of signal_info, so myAH is the master 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
+
+/* Used from signal handlers, no buffering */
+#define write_stderr(str)      write(fileno(stderr), str, strlen(str))
+
+
 #ifdef WIN32
 /* file-scope variables */
-static unsigned int tMasterThreadId = 0;
-static HANDLE termEvent = INVALID_HANDLE_VALUE;
 static DWORD tls_index;
 
 /* globally visible variables (needed by exit_nicely) */
@@ -134,7 +157,10 @@ 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 RunWorker(ArchiveHandle *AH, int pipefd[2]);
+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 bool HasEveryWorkerTerminated(ParallelState *pstate);
 static void lockTableForWorker(ArchiveHandle *AH, TocEntry *te);
 static void WaitForCommands(ArchiveHandle *AH, int pipefd[2]);
@@ -291,13 +317,13 @@ archive_close_connection(int code, void *arg)
                if (!slot)
                {
                        /*
-                        * We're the master.  Close our own database connection, if any,
-                        * and then forcibly shut down workers.
+                        * We're the master.  Forcibly shut down workers, then close our
+                        * own database connection, if any.
                         */
+                       ShutdownWorkersHard(si->pstate);
+
                        if (si->AHX)
                                DisconnectDatabase(si->AHX);
-
-                       ShutdownWorkersHard(si->pstate);
                }
                else
                {
@@ -326,34 +352,13 @@ archive_close_connection(int code, void *arg)
        }
 }
 
-/*
- * Check to see if we've been told to abort, and exit the process/thread if
- * so.  We don't print any error message; that would just clutter the screen.
- *
- * 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...).  In Unix, the master sends SIGTERM and the worker's signal
- * handler sets wantAbort to 1.  In Windows we set a termEvent and this serves
- * as the signal for worker threads to exit.  Note that while we check this
- * fairly frequently during data transfers, an idle worker doesn't come here
- * at all, so additional measures are needed to force shutdown.
- *
- * XXX in parallel restore, slow server-side operations like CREATE INDEX
- * are not interrupted by anything we do here.  This needs more work.
- */
-void
-checkAborting(ArchiveHandle *AH)
-{
-#ifdef WIN32
-       if (WaitForSingleObject(termEvent, 0) == WAIT_OBJECT_0)
-#else
-       if (wantAbort)
-#endif
-               exit_nicely(1);
-}
-
 /*
  * 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 an
+ * exit_horribly() situation, so intervening to cancel active commands is
+ * appropriate.
  */
 static void
 ShutdownWorkersHard(ParallelState *pstate)
@@ -367,15 +372,37 @@ ShutdownWorkersHard(ParallelState *pstate)
        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 abort commands-in-progress. */
+       /* On non-Windows, send SIGTERM to each worker process. */
        for (i = 0; i < pstate->numWorkers; i++)
-               kill(pstate->parallelSlot[i].pid, SIGTERM);
+       {
+               pid_t           pid = pstate->parallelSlot[i].pid;
+
+               if (pid != 0)
+                       kill(pid, SIGTERM);
+       }
 #else
-       /* Non-idle workers monitor this event via checkAborting(). */
-       SetEvent(termEvent);
+
+       /*
+        * 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].args->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);
 }
 
@@ -445,16 +472,316 @@ WaitForTerminatingWorkers(ParallelState *pstate)
        }
 }
 
+
 /*
- * Signal handler (Unix only)
+ * 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 master 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 master 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)
 {
-       wantAbort = 1;
+       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 master, forward signal to all workers.  (It seems best
+        * to do this before PQcancel; killing the master 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 master process should do this.
+        */
+       if (!signal_info.am_worker)
+       {
+               if (progname)
+               {
+                       write_stderr(progname);
+                       write_stderr(": ");
+               }
+               write_stderr("terminated by user\n");
+       }
+
+       /* And die. */
+       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 master
+                * 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->args->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 master 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 write_msg()
+                * 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->args->AH = AH;
+
+#ifdef WIN32
+       LeaveCriticalSection(&signal_info_lock);
+#endif
+}
+
 
 /*
  * This function is called by both Unix and Windows variants to set up
@@ -462,19 +789,44 @@ sigTermHandler(SIGNAL_ARGS)
  * upon return.
  */
 static void
-RunWorker(ArchiveHandle *AH, int pipefd[2])
+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);
 
-       Assert(AH->connection != NULL);
-
        /*
         * Execute commands until done.
         */
        WaitForCommands(AH, pipefd);
+
+       /*
+        * Disconnect from database and clean up.
+        */
+       set_cancel_slot_archive(slot, NULL);
+       DisconnectDatabase(&(AH->public));
+       DeCloneArchive(AH);
 }
 
 /*
@@ -484,22 +836,16 @@ RunWorker(ArchiveHandle *AH, int pipefd[2])
 static unsigned __stdcall
 init_spawned_worker_win32(WorkerInfo *wi)
 {
-       ArchiveHandle *AH;
-       int                     pipefd[2] = {wi->pipeRead, wi->pipeWrite};
-
-       /*
-        * Clone the archive so that we have our own state to work with, and in
-        * particular our own database connection.
-        */
-       AH = CloneArchive(wi->AH);
+       ArchiveHandle *AH = wi->AH;
+       ParallelSlot *slot = wi->slot;
 
+       /* Don't need WorkerInfo anymore */
        free(wi);
 
        /* Run the worker ... */
-       RunWorker(AH, pipefd);
+       RunWorker(AH, slot);
 
-       /* Clean up and exit the thread */
-       DeCloneArchive(AH);
+       /* Exit the thread */
        _endthreadex(0);
        return 0;
 }
@@ -519,9 +865,6 @@ ParallelBackupStart(ArchiveHandle *AH)
 
        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;
@@ -533,25 +876,31 @@ ParallelBackupStart(ArchiveHandle *AH)
        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;
-
 #ifdef WIN32
-       /* Set up thread management state */
-       tMasterThreadId = GetCurrentThreadId();
-       termEvent = CreateEvent(NULL, true, false, "Terminate");
        /* Make fmtId() and fmtQualifiedId() use thread-local storage */
        getLocalPQExpBuffer = getThreadLocalPQExpBuffer;
-#else
-       /* Set up signal handling state */
-       signal(SIGTERM, sigTermHandler);
-       signal(SIGINT, sigTermHandler);
-       signal(SIGQUIT, sigTermHandler);
 #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 master connection.  This
+        * ensures that child processes won't inherit valid AH->connCancel
+        * settings and thus won't try to issue cancels against the master's
+        * connection.  No harm is done if we fail while it's disabled, because
+        * the master 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++)
        {
@@ -561,6 +910,7 @@ ParallelBackupStart(ArchiveHandle *AH)
 #else
                pid_t           pid;
 #endif
+               ParallelSlot *slot = &(pstate->parallelSlot[i]);
                int                     pipeMW[2],
                                        pipeWM[2];
 
@@ -570,41 +920,40 @@ ParallelBackupStart(ArchiveHandle *AH)
                                                  "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;
+               slot->workerStatus = WRKR_IDLE;
+               slot->args = (ParallelArgs *) pg_malloc(sizeof(ParallelArgs));
+               slot->args->AH = NULL;
+               slot->args->te = NULL;
 
                /* master's ends of the pipes */
-               pstate->parallelSlot[i].pipeRead = pipeWM[PIPE_READ];
-               pstate->parallelSlot[i].pipeWrite = pipeMW[PIPE_WRITE];
+               slot->pipeRead = pipeWM[PIPE_READ];
+               slot->pipeWrite = pipeMW[PIPE_WRITE];
                /* child's ends of the pipes */
-               pstate->parallelSlot[i].pipeRevRead = pipeMW[PIPE_READ];
-               pstate->parallelSlot[i].pipeRevWrite = pipeWM[PIPE_WRITE];
+               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->pipeRead = pipeMW[PIPE_READ];
-               wi->pipeWrite = pipeWM[PIPE_WRITE];
+               wi->slot = slot;
 
                handle = _beginthreadex(NULL, 0, (void *) &init_spawned_worker_win32,
-                                                               wi, 0, &(pstate->parallelSlot[i].threadId));
-               pstate->parallelSlot[i].hThread = handle;
+                                                               wi, 0, &(slot->threadId));
+               slot->hThread = handle;
 #else                                                  /* !WIN32 */
                pid = fork();
                if (pid == 0)
                {
                        /* we are the worker */
                        int                     j;
-                       int                     pipefd[2];
 
-                       pipefd[0] = pipeMW[PIPE_READ];
-                       pipefd[1] = pipeWM[PIPE_WRITE];
+                       /* this is needed for GetMyPSlot() */
+                       slot->pid = getpid();
 
-                       pstate->parallelSlot[i].pid = getpid();
+                       /* instruct signal handler that we're in a worker now */
+                       signal_info.am_worker = true;
 
                        /* close read end of Worker -> Master */
                        closesocket(pipeWM[PIPE_READ]);
@@ -621,17 +970,8 @@ ParallelBackupStart(ArchiveHandle *AH)
                                closesocket(pstate->parallelSlot[j].pipeWrite);
                        }
 
-                       /*
-                        * Call CloneArchive 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.
-                        */
-                       pstate->parallelSlot[i].args->AH = CloneArchive(AH);
-
                        /* Run the worker ... */
-                       RunWorker(pstate->parallelSlot[i].args->AH, pipefd);
+                       RunWorker(AH, slot);
 
                        /* We can just exit(0) when done */
                        exit(0);
@@ -645,7 +985,7 @@ ParallelBackupStart(ArchiveHandle *AH)
                }
 
                /* In Master after successful fork */
-               pstate->parallelSlot[i].pid = pid;
+               slot->pid = pid;
 
                /* close read end of Master -> Worker */
                closesocket(pipeMW[PIPE_READ]);
@@ -660,9 +1000,22 @@ ParallelBackupStart(ArchiveHandle *AH)
         * the workers to inherit this setting, though.
         */
 #ifndef WIN32
-       signal(SIGPIPE, SIG_IGN);
+       pqsignal(SIGPIPE, SIG_IGN);
 #endif
 
+       /*
+        * Re-establish query cancellation on the master 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;
 }
 
@@ -692,10 +1045,11 @@ ParallelBackupEnd(ArchiveHandle *AH, ParallelState *pstate)
        WaitForTerminatingWorkers(pstate);
 
        /*
-        * Unlink pstate from shutdown_info, so the exit handler will again fall
-        * back to closing AH->connection (if connected).
+        * 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->parallelSlot);
@@ -848,9 +1202,7 @@ WaitForCommands(ArchiveHandle *AH, int pipefd[2])
        {
                if (!(command = getMessageFromMaster(pipefd)))
                {
-                       /* EOF ... clean up */
-                       PQfinish(AH->connection);
-                       AH->connection = NULL;
+                       /* EOF, so done */
                        return;
                }
 
@@ -1114,44 +1466,20 @@ select_loop(int maxFd, fd_set *workerset)
        int                     i;
        fd_set          saveSet = *workerset;
 
-#ifdef WIN32
-       for (;;)
-       {
-               /*
-                * Sleep a quarter of a second before checking if we should terminate.
-                *
-                * XXX we're not actually checking for a cancel interrupt ... but we
-                * should be.
-                */
-               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                                                  /* !WIN32 */
        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.
-                */
-               if (wantAbort)
-                       exit_horribly(modulename, "terminated by user\n");
-
+#ifndef WIN32
                if (i < 0 && errno == EINTR)
                        continue;
+#else
+               if (i == SOCKET_ERROR && WSAGetLastError() == WSAEINTR)
+                       continue;
+#endif
                break;
        }
-#endif   /* WIN32 */
 
        return i;
 }
index 8d704283e1873a91261e29979d06261408906a66..21739ca87c19208a763d7485c76c8da9142d6ba1 100644 (file)
@@ -82,6 +82,6 @@ extern void DispatchJobForTocEntry(ArchiveHandle *AH,
                                           TocEntry *te, T_Action act);
 extern void ParallelBackupEnd(ArchiveHandle *AH, ParallelState *pstate);
 
-extern void checkAborting(ArchiveHandle *AH);
+extern void set_archive_cancel_info(ArchiveHandle *AH, PGconn *conn);
 
 #endif   /* PG_DUMP_PARALLEL_H */
index 32e1c829c497236711ee39dee186ee8c1723c9fe..eadc874f9526b0cb96b2004446b9f6d1291535bf 100644 (file)
@@ -4420,6 +4420,7 @@ CloneArchive(ArchiveHandle *AH)
 
        /* The clone will have its own connection, so disregard connection state */
        clone->connection = NULL;
+       clone->connCancel = NULL;
        clone->currUser = NULL;
        clone->currSchema = NULL;
        clone->currTablespace = NULL;
@@ -4493,6 +4494,9 @@ CloneArchive(ArchiveHandle *AH)
 void
 DeCloneArchive(ArchiveHandle *AH)
 {
+       /* Should not have an open database connection */
+       Assert(AH->connection == NULL);
+
        /* Clear format-specific state */
        (AH->DeClonePtr) (AH);
 
index 4aa7190c40edbc8a13654ba2c9efb00071f2b5dd..0376f2bff760c8b27ac23b5f3b6a3df43d65082e 100644 (file)
@@ -285,6 +285,9 @@ struct _archiveHandle
        char       *savedPassword;      /* password for ropt->username, if known */
        char       *use_role;
        PGconn     *connection;
+       /* If connCancel isn't NULL, SIGINT handler will send a cancel */
+       PGcancel   *volatile connCancel;
+
        int                     connectToDB;    /* Flag to indicate if direct DB connection is
                                                                 * required */
        ArchiverOutput outputKind;      /* Flag for what we're currently writing */
index 818bc9efe14a0c04b4678e5fa65b291ae7812c2d..352595e49faa16015266b071d400e9d73aafd65f 100644 (file)
@@ -12,6 +12,7 @@
 #include "postgres_fe.h"
 
 #include "dumputils.h"
+#include "parallel.h"
 #include "pg_backup_archiver.h"
 #include "pg_backup_db.h"
 #include "pg_backup_utils.h"
@@ -106,6 +107,9 @@ ReconnectToServer(ArchiveHandle *AH, const char *dbname, const char *username)
 
        newConn = _connectDB(AH, newdbname, newusername);
 
+       /* Update ArchiveHandle's connCancel before closing old connection */
+       set_archive_cancel_info(AH, newConn);
+
        PQfinish(AH->connection);
        AH->connection = newConn;
 
@@ -327,6 +331,9 @@ ConnectDatabase(Archive *AHX,
        _check_database_version(AH);
 
        PQsetNoticeProcessor(AH->connection, notice_processor, NULL);
+
+       /* arrange for SIGINT to issue a query cancel on this connection */
+       set_archive_cancel_info(AH, AH->connection);
 }
 
 /*
@@ -337,19 +344,25 @@ void
 DisconnectDatabase(Archive *AHX)
 {
        ArchiveHandle *AH = (ArchiveHandle *) AHX;
-       PGcancel   *cancel;
        char            errbuf[1];
 
        if (!AH->connection)
                return;
 
-       if (PQtransactionStatus(AH->connection) == PQTRANS_ACTIVE)
+       if (AH->connCancel)
        {
-               if ((cancel = PQgetCancel(AH->connection)))
-               {
-                       PQcancel(cancel, errbuf, sizeof(errbuf));
-                       PQfreeCancel(cancel);
-               }
+               /*
+                * If we have an active query, send a cancel before closing.  This is
+                * of no use for a normal exit, but might be helpful during
+                * exit_horribly().
+                */
+               if (PQtransactionStatus(AH->connection) == PQTRANS_ACTIVE)
+                       PQcancel(AH->connCancel, errbuf, sizeof(errbuf));
+
+               /*
+                * Prevent signal handler from sending a cancel after this.
+                */
+               set_archive_cancel_info(AH, NULL);
        }
 
        PQfinish(AH->connection);
@@ -631,6 +644,11 @@ EndDBCopyMode(Archive *AHX, const char *tocEntryTag)
                                                                tocEntryTag, PQerrorMessage(AH->connection));
                PQclear(res);
 
+               /* Do this to ensure we've pumped libpq back to idle state */
+               if (PQgetResult(AH->connection) != NULL)
+                       write_msg(NULL, "WARNING: unexpected extra results during COPY of table \"%s\"\n",
+                                         tocEntryTag);
+
                AH->pgCopyIn = false;
        }
 }
index 27c619054618445cbece42dc5acfd12d969a2db3..e52f12258a72514b5ee0006103596d61310e8f2b 100644 (file)
@@ -356,9 +356,6 @@ _WriteData(ArchiveHandle *AH, const void *data, size_t dLen)
 {
        lclContext *ctx = (lclContext *) AH->formatData;
 
-       /* Are we aborting? */
-       checkAborting(AH);
-
        if (dLen > 0 && cfwrite(data, dLen, ctx->dataFH) != dLen)
                WRITE_ERROR_EXIT;
 
@@ -407,9 +404,6 @@ _PrintFileData(ArchiveHandle *AH, char *filename)
 
        while ((cnt = cfread(buf, buflen, cfp)))
        {
-               /* Are we aborting? */
-               checkAborting(AH);
-
                ahwrite(buf, 1, cnt, AH);
        }
 
@@ -529,9 +523,6 @@ _WriteBuf(ArchiveHandle *AH, const void *buf, size_t len)
 {
        lclContext *ctx = (lclContext *) AH->formatData;
 
-       /* Are we aborting? */
-       checkAborting(AH);
-
        if (cfwrite(buf, len, ctx->dataFH) != len)
                WRITE_ERROR_EXIT;
 
@@ -548,9 +539,6 @@ _ReadBuf(ArchiveHandle *AH, void *buf, size_t len)
 {
        lclContext *ctx = (lclContext *) AH->formatData;
 
-       /* Are we aborting? */
-       checkAborting(AH);
-
        /*
         * If there was an I/O error, we already exited in cfread(), so here we
         * exit on short reads.
index 9a0eb24e3eb43785dc95219d37fb9522376f8214..296273ba4abfcc0110caf10b604a4da9df911dee 100644 (file)
@@ -1780,6 +1780,11 @@ dumpTableData_copy(Archive *fout, void *dcontext)
        }
        PQclear(res);
 
+       /* Do this to ensure we've pumped libpq back to idle state */
+       if (PQgetResult(conn) != NULL)
+               write_msg(NULL, "WARNING: unexpected extra results during COPY of table \"%s\"\n",
+                                 classname);
+
        destroyPQExpBuffer(q);
        return 1;
 }