X-Git-Url: http://git.annexia.org/?p=libguestfs.git;a=blobdiff_plain;f=src%2Fguestfs.c;h=b30ade06e3aed24ecbca3764eb1e8a2b177603ea;hp=1b7d61ca96bea0bedf088f6491188f43b09a6788;hb=d901cc916102f1aaccfb73396b48aa303e5b8cd7;hpb=170f262f0413de843af62b968f6d12c1c476ae7f

diff --git a/src/guestfs.c b/src/guestfs.c
index 1b7d61c..b30ade0 100644
--- a/src/guestfs.c
+++ b/src/guestfs.c
@@ -29,7 +29,9 @@
 #include <string.h>
 #include <fcntl.h>
 #include <time.h>
+#include <sys/stat.h>
 #include <sys/select.h>
+
 #include <rpc/types.h>
 #include <rpc/xdr.h>
 
@@ -141,7 +143,8 @@ struct guestfs_h
 
   int fd[2];			/* Stdin/stdout of qemu. */
   int sock;			/* Daemon communications socket. */
-  int pid;			/* Qemu PID. */
+  pid_t pid;			/* Qemu PID. */
+  pid_t recoverypid;		/* Recovery process PID. */
   time_t start_t;		/* The time when we started qemu. */
 
   int stdout_watch;		/* Watches qemu stdout for log messages. */
@@ -155,7 +158,9 @@ struct guestfs_h
   int verbose;
   int autosync;
 
-  const char *path;
+  char *path;			/* Path to kernel, initrd. */
+  char *qemu;			/* Qemu binary. */
+  char *append;			/* Append to kernel command line. */
 
   char *last_error;
 
@@ -216,8 +221,18 @@ guestfs_create (void)
   g->verbose = str != NULL && strcmp (str, "1") == 0;
 
   str = getenv ("LIBGUESTFS_PATH");
-  g->path = str != NULL ? str : GUESTFS_DEFAULT_PATH;
-  /* XXX We should probably make QEMU configurable as well. */
+  g->path = str != NULL ? strdup (str) : strdup (GUESTFS_DEFAULT_PATH);
+  if (!g->path) goto error;
+
+  str = getenv ("LIBGUESTFS_QEMU");
+  g->qemu = str != NULL ? strdup (str) : strdup (QEMU);
+  if (!g->qemu) goto error;
+
+  str = getenv ("LIBGUESTFS_APPEND");
+  if (str) {
+    g->append = strdup (str);
+    if (!g->append) goto error;
+  }
 
   g->main_loop = guestfs_get_default_main_loop ();
 
@@ -242,6 +257,13 @@ guestfs_create (void)
     fprintf (stderr, "new guestfs handle %p\n", g);
 
   return g;
+
+ error:
+  free (g->path);
+  free (g->qemu);
+  free (g->append);
+  free (g);
+  return NULL;
 }
 
 void
@@ -261,8 +283,10 @@ guestfs_close (guestfs_h *g)
     fprintf (stderr, "closing guestfs handle %p (state %d)\n", g, g->state);
 
   /* Try to sync if autosync flag is set. */
-  if (g->autosync && g->state == READY)
+  if (g->autosync && g->state == READY) {
+    guestfs_umount_all (g);
     guestfs_sync (g);
+  }
 
   /* Remove any handlers that might be called back before we kill the
    * subprocess.
@@ -300,7 +324,12 @@ guestfs_close (guestfs_h *g)
   }
   /* release mutex (XXX) */
 
+  free (g->msg_in);
+  free (g->msg_out);
   free (g->last_error);
+  free (g->path);
+  free (g->qemu);
+  free (g->append);
   free (g);
 }
 
@@ -433,8 +462,11 @@ xread (int fd, void *buf, size_t len)
 
   while (len > 0) {
     r = read (fd, buf, len);
-    if (r == -1)
+    if (r == -1) {
+      if (errno == EINTR || errno == EAGAIN)
+	continue;
       return -1;
+    }
 
     buf += r;
     len -= r;
@@ -498,10 +530,12 @@ guestfs_get_autosync (guestfs_h *g)
 int
 guestfs_set_path (guestfs_h *g, const char *path)
 {
-  if (path == NULL)
-    g->path = GUESTFS_DEFAULT_PATH;
-  else
-    g->path = path;
+  free (g->path);
+  g->path = NULL;
+
+  g->path =
+    path == NULL ?
+    safe_strdup (g, GUESTFS_DEFAULT_PATH) : safe_strdup (g, path);
   return 0;
 }
 
@@ -511,6 +545,38 @@ guestfs_get_path (guestfs_h *g)
   return g->path;
 }
 
+int
+guestfs_set_qemu (guestfs_h *g, const char *qemu)
+{
+  free (g->qemu);
+  g->qemu = NULL;
+
+  g->qemu = qemu == NULL ? safe_strdup (g, QEMU) : safe_strdup (g, qemu);
+  return 0;
+}
+
+const char *
+guestfs_get_qemu (guestfs_h *g)
+{
+  return g->qemu;
+}
+
+int
+guestfs_set_append (guestfs_h *g, const char *append)
+{
+  free (g->append);
+  g->append = NULL;
+
+  g->append = append ? safe_strdup (g, append) : NULL;
+  return 0;
+}
+
+const char *
+guestfs_get_append (guestfs_h *g)
+{
+  return g->append;
+}
+
 /* Add a string to the current command line. */
 static void
 incr_cmdline_size (guestfs_h *g)
@@ -575,7 +641,7 @@ guestfs_config (guestfs_h *g,
 int
 guestfs_add_drive (guestfs_h *g, const char *filename)
 {
-  int len = strlen (filename) + 64;
+  size_t len = strlen (filename) + 64;
   char buf[len];
 
   if (strchr (filename, ',') != NULL) {
@@ -613,7 +679,8 @@ int
 guestfs_launch (guestfs_h *g)
 {
   static const char *dir_template = "/tmp/libguestfsXXXXXX";
-  int r, i, len, pmore;
+  int r, i, pmore, memsize;
+  size_t len;
   int wfd[2], rfd[2];
   int tries;
   const char *kernel_name = "vmlinuz." REPO "." host_cpu;
@@ -681,6 +748,15 @@ guestfs_launch (guestfs_h *g)
     goto cleanup0;
   }
 
+  /* Choose a suitable memory size.  Previously we tried to choose
+   * a minimal memory size, but this isn't really necessary since
+   * recent QEMU and KVM don't do anything nasty like locking
+   * memory into core any more.  This we can safely choose a
+   * large, generous amount of memory, and it'll just get swapped
+   * on smaller systems.
+   */
+  memsize = 384;
+
   /* Make the temporary directory containing the socket. */
   if (!g->tmpdir) {
     g->tmpdir = safe_strdup (g, dir_template);
@@ -711,11 +787,12 @@ guestfs_launch (guestfs_h *g)
   if (r == 0) {			/* Child (qemu). */
     char vmchannel[256];
     char append[256];
+    char memsize_str[256];
 
     /* Set up the full command line.  Do this in the subprocess so we
      * don't need to worry about cleaning up.
      */
-    g->cmdline[0] = (char *) QEMU;
+    g->cmdline[0] = g->qemu;
 
     /* Construct the -net channel parameter for qemu. */
     snprintf (vmchannel, sizeof vmchannel,
@@ -724,13 +801,19 @@ guestfs_launch (guestfs_h *g)
 
     /* Linux kernel command line. */
     snprintf (append, sizeof append,
-	      "console=ttyS0 guestfs=%s:%d%s",
+	      "panic=1 console=ttyS0 guestfs=%s:%d%s%s%s",
 	      VMCHANNEL_ADDR, VMCHANNEL_PORT,
-	      g->verbose ? " guestfs_verbose=1" : "");
+	      g->verbose ? " guestfs_verbose=1" : "",
+	      g->append ? " " : "", g->append ? g->append : "");
+
+    snprintf (memsize_str, sizeof memsize_str, "%d", memsize);
 
     add_cmdline (g, "-m");
-    add_cmdline (g, "384");	  /* XXX Choose best size. */
+    add_cmdline (g, memsize_str);
+#if 0
     add_cmdline (g, "-no-kqemu"); /* Avoids a warning. */
+#endif
+    add_cmdline (g, "-no-reboot"); /* Force exit instead of reboot on panic */
     add_cmdline (g, "-kernel");
     add_cmdline (g, (char *) kernel);
     add_cmdline (g, "-initrd");
@@ -745,12 +828,12 @@ guestfs_launch (guestfs_h *g)
     add_cmdline (g, "-net");
     add_cmdline (g, "user,vlan=0");
     add_cmdline (g, "-net");
-    add_cmdline (g, "nic,vlan=0");
+    add_cmdline (g, "nic,model=virtio,vlan=0");
     incr_cmdline_size (g);
     g->cmdline[g->cmdline_size-1] = NULL;
 
     if (g->verbose) {
-      fprintf (stderr, "%s", QEMU);
+      fprintf (stderr, "%s", g->qemu);
       for (i = 0; g->cmdline[i]; ++i)
 	fprintf (stderr, " %s", g->cmdline[i]);
       fprintf (stderr, "\n");
@@ -773,14 +856,53 @@ guestfs_launch (guestfs_h *g)
     setpgid (0, 0);
 #endif
 
-    execv (QEMU, g->cmdline);	/* Run qemu. */
-    perror (QEMU);
+    execv (g->qemu, g->cmdline); /* Run qemu. */
+    perror (g->qemu);
     _exit (1);
   }
 
   /* Parent (library). */
   g->pid = r;
 
+  free (kernel);
+  kernel = NULL;
+  free (initrd);
+  initrd = NULL;
+
+  /* Fork the recovery process off which will kill qemu if the parent
+   * process fails to do so (eg. if the parent segfaults).
+   */
+  r = fork ();
+  if (r == 0) {
+    pid_t qemu_pid = g->pid;
+    pid_t parent_pid = getppid ();
+
+    /* Writing to argv is hideously complicated and error prone.  See:
+     * http://anoncvs.postgresql.org/cvsweb.cgi/pgsql/src/backend/utils/misc/ps_status.c?rev=1.33.2.1;content-type=text%2Fplain
+     */
+
+    /* Loop around waiting for one or both of the other processes to
+     * disappear.  It's fair to say this is very hairy.  The PIDs that
+     * we are looking at might be reused by another process.  We are
+     * effectively polling.  Is the cure worse than the disease?
+     */
+    for (;;) {
+      if (kill (qemu_pid, 0) == -1) /* qemu's gone away, we aren't needed */
+	_exit (0);
+      if (kill (parent_pid, 0) == -1) {
+	/* Parent's gone away, qemu still around, so kill qemu. */
+	kill (qemu_pid, 9);
+	_exit (0);
+      }
+      sleep (2);
+    }
+  }
+
+  /* Don't worry, if the fork failed, this will be -1.  The recovery
+   * process isn't essential.
+   */
+  g->recoverypid = r;
+
   /* Start the clock ... */
   time (&g->start_t);
 
@@ -822,10 +944,9 @@ guestfs_launch (guestfs_h *g)
   addr.sun_path[UNIX_PATH_MAX-1] = '\0';
 
   tries = 100;
+  /* Always sleep at least once to give qemu a small chance to start up. */
+  usleep (10000);
   while (tries > 0) {
-    /* Always sleep at least once to give qemu a small chance to start up. */
-    usleep (10000);
-
     r = connect (g->sock, (struct sockaddr *) &addr, sizeof addr);
     if ((r == -1 && errno == EINPROGRESS) || r == 0)
       goto connected;
@@ -833,6 +954,7 @@ guestfs_launch (guestfs_h *g)
     if (errno != ENOENT)
       perrorf (g, "connect");
     tries--;
+    usleep (100000);
   }
 
   error (g, "failed to connect to vmchannel socket");
@@ -877,11 +999,14 @@ guestfs_launch (guestfs_h *g)
   close (wfd[1]);
   close (rfd[0]);
   kill (g->pid, 9);
+  if (g->recoverypid > 0) kill (g->recoverypid, 9);
   waitpid (g->pid, NULL, 0);
+  if (g->recoverypid > 0) waitpid (g->recoverypid, NULL, 0);
   g->fd[0] = -1;
   g->fd[1] = -1;
   g->sock = -1;
   g->pid = 0;
+  g->recoverypid = 0;
   g->start_t = 0;
   g->stdout_watch = -1;
   g->sock_watch = -1;
@@ -957,6 +1082,7 @@ guestfs_kill_subprocess (guestfs_h *g)
     fprintf (stderr, "sending SIGTERM to process %d\n", g->pid);
 
   kill (g->pid, SIGTERM);
+  if (g->recoverypid > 0) kill (g->recoverypid, 9);
 
   return 0;
 }
@@ -1014,6 +1140,25 @@ guestfs_set_busy (guestfs_h *g)
   return 0;
 }
 
+int
+guestfs_end_busy (guestfs_h *g)
+{
+  switch (g->state)
+    {
+    case BUSY:
+      g->state = READY;
+      break;
+    case CONFIG:
+    case READY:
+      break;
+    case LAUNCHING:
+    case NO_HANDLE:
+      error (g, "guestfs_end_busy: called when in state %d", g->state);
+      return -1;
+    }
+  return 0;
+}
+
 /* Structure-freeing functions.  These rely on the fact that the
  * structure format is identical to the XDR format.  See note in
  * generator.ml.
@@ -1045,6 +1190,39 @@ guestfs_free_lvm_lv_list (struct guestfs_lvm_lv_list *x)
   free (x);
 }
 
+/* We don't know if stdout_event or sock_read_event will be the
+ * first to receive EOF if the qemu process dies.  This function
+ * has the common cleanup code for both.
+ */
+static void
+child_cleanup (guestfs_h *g)
+{
+  if (g->verbose)
+    fprintf (stderr, "stdout_event: %p: child process died\n", g);
+  /*kill (g->pid, SIGTERM);*/
+  if (g->recoverypid > 0) kill (g->recoverypid, 9);
+  waitpid (g->pid, NULL, 0);
+  if (g->recoverypid > 0) waitpid (g->recoverypid, NULL, 0);
+  if (g->stdout_watch >= 0)
+    g->main_loop->remove_handle (g->main_loop, g, g->stdout_watch);
+  if (g->sock_watch >= 0)
+    g->main_loop->remove_handle (g->main_loop, g, g->sock_watch);
+  close (g->fd[0]);
+  close (g->fd[1]);
+  close (g->sock);
+  g->fd[0] = -1;
+  g->fd[1] = -1;
+  g->sock = -1;
+  g->pid = 0;
+  g->recoverypid = 0;
+  g->start_t = 0;
+  g->stdout_watch = -1;
+  g->sock_watch = -1;
+  g->state = CONFIG;
+  if (g->subprocess_quit_cb)
+    g->subprocess_quit_cb (g, g->subprocess_quit_cb_data);
+}
+
 /* This function is called whenever qemu prints something on stdout.
  * Qemu's stdout is also connected to the guest's serial console, so
  * we see kernel messages here too.
@@ -1071,32 +1249,12 @@ stdout_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
   n = read (fd, buf, sizeof buf);
   if (n == 0) {
     /* Hopefully this indicates the qemu child process has died. */
-    if (g->verbose)
-      fprintf (stderr, "stdout_event: %p: child process died\n", g);
-    /*kill (g->pid, SIGTERM);*/
-    waitpid (g->pid, NULL, 0);
-    if (g->stdout_watch >= 0)
-      g->main_loop->remove_handle (g->main_loop, g, g->stdout_watch);
-    if (g->sock_watch >= 0)
-      g->main_loop->remove_handle (g->main_loop, g, g->sock_watch);
-    close (g->fd[0]);
-    close (g->fd[1]);
-    close (g->sock);
-    g->fd[0] = -1;
-    g->fd[1] = -1;
-    g->sock = -1;
-    g->pid = 0;
-    g->start_t = 0;
-    g->stdout_watch = -1;
-    g->sock_watch = -1;
-    g->state = CONFIG;
-    if (g->subprocess_quit_cb)
-      g->subprocess_quit_cb (g, g->subprocess_quit_cb_data);
+    child_cleanup (g);
     return;
   }
 
   if (n == -1) {
-    if (errno != EAGAIN)
+    if (errno != EINTR && errno != EAGAIN)
       perrorf (g, "read");
     return;
   }
@@ -1118,7 +1276,7 @@ sock_read_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
 		 int watch, int fd, int events)
 {
   XDR xdr;
-  unsigned len;
+  u_int32_t len;
   int n;
 
   if (g->verbose)
@@ -1137,14 +1295,14 @@ sock_read_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
   }
   n = read (g->sock, g->msg_in + g->msg_in_size,
 	    g->msg_in_allocated - g->msg_in_size);
-  if (n == 0)
-    /* Disconnected?  Ignore it because stdout_watch will get called
-     * and will do the cleanup.
-     */
+  if (n == 0) {
+    /* Disconnected. */
+    child_cleanup (g);
     return;
+  }
 
   if (n == -1) {
-    if (errno != EAGAIN)
+    if (errno != EINTR && errno != EAGAIN)
       perrorf (g, "read");
     return;
   }
@@ -1203,6 +1361,7 @@ sock_read_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
   if (g->msg_in_size-4 < len) return; /* Need more of this message. */
 
   /* Got the full message, begin processing it. */
+#if 0
   if (g->verbose) {
     int i, j;
 
@@ -1223,6 +1382,7 @@ sock_read_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
       printf ("|\n");
     }
   }
+#endif
 
   /* Not in the expected state. */
   if (g->state != BUSY)
@@ -1231,6 +1391,9 @@ sock_read_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
   /* Push the message up to the higher layer. */
   if (g->reply_cb)
     g->reply_cb (g, g->reply_cb_data, &xdr);
+  else
+    /* This message (probably) should never be printed. */
+    fprintf (stderr, "libguesfs: sock_read_event: !!! dropped message !!!\n");
 
   g->msg_in_size -= len + 4;
   memmove (g->msg_in, g->msg_in+len+4, g->msg_in_size);
@@ -1431,10 +1594,9 @@ int
 guestfs__send_sync (guestfs_h *g, int proc_nr,
 		    xdrproc_t xdrp, char *args)
 {
-  char buffer[GUESTFS_MESSAGE_MAX];
   struct guestfs_message_header hdr;
   XDR xdr;
-  unsigned len;
+  u_int32_t len;
   int serial = g->msg_next_serial++;
   int sent;
   guestfs_main_loop *ml = guestfs_get_main_loop (g);
@@ -1444,6 +1606,23 @@ guestfs__send_sync (guestfs_h *g, int proc_nr,
     return -1;
   }
 
+  /* This is probably an internal error.  Or perhaps we should just
+   * free the buffer anyway?
+   */
+  if (g->msg_out != NULL) {
+    error (g, "guestfs__send_sync: msg_out should be NULL");
+    return -1;
+  }
+
+  /* We have to allocate this message buffer on the heap because
+   * it is quite large (although will be mostly unused).  We
+   * can't allocate it on the stack because in some environments
+   * we have quite limited stack space available, notably when
+   * running in the JVM.
+   */
+  g->msg_out = safe_malloc (g, GUESTFS_MESSAGE_MAX + 4);
+  xdrmem_create (&xdr, g->msg_out + 4, GUESTFS_MESSAGE_MAX, XDR_ENCODE);
+
   /* Serialize the header. */
   hdr.prog = GUESTFS_PROGRAM;
   hdr.vers = GUESTFS_PROTOCOL_VERSION;
@@ -1452,10 +1631,9 @@ guestfs__send_sync (guestfs_h *g, int proc_nr,
   hdr.serial = serial;
   hdr.status = GUESTFS_STATUS_OK;
 
-  xdrmem_create (&xdr, buffer, sizeof buffer, XDR_ENCODE);
   if (!xdr_guestfs_message_header (&xdr, &hdr)) {
     error (g, "xdr_guestfs_message_header failed");
-    return -1;
+    goto cleanup1;
   }
 
   /* Serialize the args.  If any, because some message types
@@ -1464,24 +1642,23 @@ guestfs__send_sync (guestfs_h *g, int proc_nr,
   if (xdrp) {
     if (!(*xdrp) (&xdr, args)) {
       error (g, "dispatch failed to marshal args");
-      return -1;
+      goto cleanup1;
     }
   }
 
+  /* Get the actual length of the message, resize the buffer to match
+   * the actual length, and write the length word at the beginning.
+   */
   len = xdr_getpos (&xdr);
   xdr_destroy (&xdr);
 
-  /* Allocate the outgoing message buffer. */
-  g->msg_out = safe_malloc (g, len + 4);
-
+  g->msg_out = safe_realloc (g, g->msg_out, len + 4);
   g->msg_out_size = len + 4;
   g->msg_out_pos = 0;
 
   xdrmem_create (&xdr, g->msg_out, 4, XDR_ENCODE);
   xdr_uint32_t (&xdr, &len);
 
-  memcpy (g->msg_out + 4, buffer, len);
-
   if (guestfs__switch_to_sending (g) == -1)
     goto cleanup1;
 
@@ -1532,13 +1709,20 @@ guestfs__send_file_sync (guestfs_h *g, const char *filename)
   }
 
   /* Send file in chunked encoding. */
-  while (!cancel && (r = read (fd, buf, sizeof buf)) > 0) {
+  while (!cancel) {
+    r = read (fd, buf, sizeof buf);
+    if (r == -1 && (errno == EINTR || errno == EAGAIN))
+      continue;
+    if (r <= 0) break;
     err = send_file_data_sync (g, buf, r);
-    if (err < 0)
+    if (err < 0) {
+      if (err == -2)		/* daemon sent cancellation */
+	send_file_cancellation_sync (g);
       return err;
+    }
   }
 
-  if (cancel) {
+  if (cancel) {			/* cancel from either end */
     send_file_cancellation_sync (g);
     return -1;
   }
@@ -1589,10 +1773,9 @@ send_file_complete_sync (guestfs_h *g)
 static int check_for_daemon_cancellation (guestfs_h *g);
 
 static int
-send_file_chunk_sync (guestfs_h *g, int cancel, const char *buf, size_t len)
+send_file_chunk_sync (guestfs_h *g, int cancel, const char *buf, size_t buflen)
 {
-  char data[GUESTFS_MAX_CHUNK_SIZE + 48];
-  unsigned datalen;
+  u_int32_t len;
   int sent;
   guestfs_chunk chunk;
   XDR xdr;
@@ -1603,34 +1786,49 @@ send_file_chunk_sync (guestfs_h *g, int cancel, const char *buf, size_t len)
     return -1;
   }
 
+  /* This is probably an internal error.  Or perhaps we should just
+   * free the buffer anyway?
+   */
+  if (g->msg_out != NULL) {
+    error (g, "guestfs__send_sync: msg_out should be NULL");
+    return -1;
+  }
+
   /* Did the daemon send a cancellation message? */
-  if (check_for_daemon_cancellation (g))
+  if (check_for_daemon_cancellation (g)) {
+    if (g->verbose)
+      fprintf (stderr, "got daemon cancellation\n");
     return -2;
+  }
+
+  /* Allocate the chunk buffer.  Don't use the stack to avoid
+   * excessive stack usage and unnecessary copies.
+   */
+  g->msg_out = safe_malloc (g, GUESTFS_MAX_CHUNK_SIZE + 4 + 48);
+  xdrmem_create (&xdr, g->msg_out + 4, GUESTFS_MAX_CHUNK_SIZE + 48, XDR_ENCODE);
 
   /* Serialize the chunk. */
   chunk.cancel = cancel;
-  chunk.data.data_len = len;
+  chunk.data.data_len = buflen;
   chunk.data.data_val = (char *) buf;
 
-  xdrmem_create (&xdr, data, sizeof data, XDR_ENCODE);
   if (!xdr_guestfs_chunk (&xdr, &chunk)) {
-    error (g, "xdr_guestfs_chunk failed (buf = %p, len = %zu)", buf, len);
+    error (g, "xdr_guestfs_chunk failed (buf = %p, buflen = %zu)",
+	   buf, buflen);
     xdr_destroy (&xdr);
-    return -1;
+    goto cleanup1;
   }
 
-  datalen = xdr_getpos (&xdr);
+  len = xdr_getpos (&xdr);
   xdr_destroy (&xdr);
 
-  /* Allocate outgoing message buffer. */
-  g->msg_out = safe_malloc (g, datalen + 4);
-  g->msg_out_size = datalen + 4;
+  /* Reduce the size of the outgoing message buffer to the real length. */
+  g->msg_out = safe_realloc (g, g->msg_out, len + 4);
+  g->msg_out_size = len + 4;
   g->msg_out_pos = 0;
 
   xdrmem_create (&xdr, g->msg_out, 4, XDR_ENCODE);
-  xdr_uint32_t (&xdr, &datalen);
-
-  memcpy (g->msg_out + 4, data, datalen);
+  xdr_uint32_t (&xdr, &len);
 
   if (guestfs__switch_to_sending (g) == -1)
     goto cleanup1;
@@ -1701,13 +1899,15 @@ check_for_daemon_cancellation (guestfs_h *g)
 
 /* Synchronously receive a file. */
 
-static int receive_file_data_sync (guestfs_h *g, void **buf);
+/* Returns -1 = error, 0 = EOF, 1 = more data */
+static int receive_file_data_sync (guestfs_h *g, void **buf, size_t *len);
 
 int
 guestfs__receive_file_sync (guestfs_h *g, const char *filename)
 {
   void *buf;
   int fd, r;
+  size_t len;
 
   fd = open (filename, O_WRONLY|O_CREAT|O_TRUNC|O_NOCTTY, 0666);
   if (fd == -1) {
@@ -1716,13 +1916,14 @@ guestfs__receive_file_sync (guestfs_h *g, const char *filename)
   }
 
   /* Receive the file in chunked encoding. */
-  while ((r = receive_file_data_sync (g, &buf)) > 0) {
-    if (xwrite (fd, buf, r) == -1) {
+  while ((r = receive_file_data_sync (g, &buf, &len)) >= 0) {
+    if (xwrite (fd, buf, len) == -1) {
       perrorf (g, "%s: write", filename);
       free (buf);
       goto cancel;
     }
     free (buf);
+    if (r == 0) break; /* End of file. */
   }
 
   if (r == -1) {
@@ -1754,75 +1955,124 @@ guestfs__receive_file_sync (guestfs_h *g, const char *filename)
     return -1;
   }
 
-  while ((r = receive_file_data_sync (g, &buf)) > 0)
-    free (buf);			/* just discard it */
+  while ((r = receive_file_data_sync (g, NULL, NULL)) > 0)
+    ;				/* just discard it */
 
   return -1;
 }
 
+/* Note that the reply callback can be called multiple times before
+ * the main loop quits and we get back to the synchronous code.  So
+ * we have to be prepared to save multiple chunks on a list here.
+ */
 struct receive_file_ctx {
-  int code;
-  void **buf;
+  int count;			/* 0 if receive_file_cb not called, or
+				 * else count number of chunks.
+				 */
+  guestfs_chunk *chunks;	/* Array of chunks. */
 };
 
 static void
+free_chunks (struct receive_file_ctx *ctx)
+{
+  int i;
+
+  for (i = 0; i < ctx->count; ++i)
+    free (ctx->chunks[i].data.data_val);
+
+  free (ctx->chunks);
+}
+
+static void
 receive_file_cb (guestfs_h *g, void *data, XDR *xdr)
 {
   guestfs_main_loop *ml = guestfs_get_main_loop (g);
   struct receive_file_ctx *ctx = (struct receive_file_ctx *) data;
   guestfs_chunk chunk;
 
+  if (ctx->count == -1)		/* Parse error occurred previously. */
+    return;
+
   ml->main_loop_quit (ml, g);
 
   memset (&chunk, 0, sizeof chunk);
 
   if (!xdr_guestfs_chunk (xdr, &chunk)) {
     error (g, "failed to parse file chunk");
-    ctx->code = -1;
-    return;
-  }
-  if (chunk.cancel) {
-    error (g, "file receive cancelled by daemon");
-    ctx->code = -2;
-    return;
-  }
-  if (chunk.data.data_len == 0) { /* end of transfer */
-    ctx->code = 0;
+    free_chunks (ctx);
+    ctx->chunks = NULL;
+    ctx->count = -1;
     return;
   }
 
-  ctx->code = chunk.data.data_len;
-  *ctx->buf = chunk.data.data_val; /* caller frees */
+  /* Copy the chunk to the list. */
+  ctx->chunks = safe_realloc (g, ctx->chunks,
+			      sizeof (guestfs_chunk) * (ctx->count+1));
+  ctx->chunks[ctx->count] = chunk;
+  ctx->count++;
 }
 
 /* Receive a chunk of file data. */
+/* Returns -1 = error, 0 = EOF, 1 = more data */
 static int
-receive_file_data_sync (guestfs_h *g, void **buf)
+receive_file_data_sync (guestfs_h *g, void **buf, size_t *len_r)
 {
   struct receive_file_ctx ctx;
   guestfs_main_loop *ml = guestfs_get_main_loop (g);
+  int i;
+  size_t len;
 
-  ctx.code = -3;
-  ctx.buf = buf;
+  ctx.count = 0;
+  ctx.chunks = NULL;
 
   guestfs_set_reply_callback (g, receive_file_cb, &ctx);
   (void) ml->main_loop_run (ml, g);
   guestfs_set_reply_callback (g, NULL, NULL);
 
-  if (g->verbose)
-    fprintf (stderr, "receive_file_data_sync: code %d\n", ctx.code);
-
-  switch (ctx.code) {
-  case 0:			/* end of file */
-    return 0;
-  case -1: case -2:
+  if (ctx.count == 0) {
+    error (g, "receive_file_data_sync: reply callback not called\n");
     return -1;
-  case -3:
-    error (g, "failed to call receive_file_cb");
+  }
+
+  if (ctx.count == -1) {
+    error (g, "receive_file_data_sync: parse error in reply callback\n");
+    /* callback already freed the chunks */
     return -1;
-  default:			/* received n bytes of data */
-    return ctx.code;
   }
+
+  if (g->verbose)
+    fprintf (stderr, "receive_file_data_sync: got %d chunks\n", ctx.count);
+
+  /* Process each chunk in the list. */
+  if (buf) *buf = NULL;		/* Accumulate data in this buffer. */
+  len = 0;
+
+  for (i = 0; i < ctx.count; ++i) {
+    if (ctx.chunks[i].cancel) {
+      error (g, "file receive cancelled by daemon");
+      free_chunks (&ctx);
+      if (buf) free (*buf);
+      if (len_r) *len_r = 0;
+      return -1;
+    }
+
+    if (ctx.chunks[i].data.data_len == 0) { /* end of transfer */
+      free_chunks (&ctx);
+      if (len_r) *len_r = len;
+      return 0;
+    }
+
+    if (buf) {
+      *buf = safe_realloc (g, *buf, len + ctx.chunks[i].data.data_len);
+      memcpy (*buf+len, ctx.chunks[i].data.data_val,
+	      ctx.chunks[i].data.data_len);
+    }
+    len += ctx.chunks[i].data.data_len;
+  }
+
+  if (len_r) *len_r = len;
+  free_chunks (&ctx);
+  return 1;
 }
 
 /* This is the default main loop implementation, using select(2). */
@@ -1963,6 +2213,8 @@ select_main_loop_run (guestfs_main_loop *mlv, guestfs_h *g)
     xset2 = ml->xset;
     r = select (ml->max_fd+1, &rset2, &wset2, &xset2, NULL);
     if (r == -1) {
+      if (errno == EINTR || errno == EAGAIN)
+	continue;
       perrorf (g, "select");
       ml->is_running = 0;
       return -1;