X-Git-Url: http://git.annexia.org/?p=libguestfs.git;a=blobdiff_plain;f=src%2Fguestfs.c;h=05c02f3f42ce71777c2e3d50a55c350ee5cb2d31;hp=6c49e9091887e6da643079be0fb556cb265040aa;hb=41b959fd9b057354f642d84398b875d02b88b864;hpb=1765330e07a48dc6f7bdef7007f69ebe606fa731

diff --git a/src/guestfs.c b/src/guestfs.c
index 6c49e90..05c02f3 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>
 
@@ -56,12 +58,12 @@
 #include "guestfs.h"
 #include "guestfs_protocol.h"
 
-static void error (guestfs_h *g, const char *fs, ...);
-static void perrorf (guestfs_h *g, const char *fs, ...);
-static void *safe_malloc (guestfs_h *g, size_t nbytes);
-static void *safe_realloc (guestfs_h *g, void *ptr, int nbytes);
-static char *safe_strdup (guestfs_h *g, const char *str);
-static void *safe_memdup (guestfs_h *g, void *ptr, size_t size);
+#define error guestfs_error
+#define perrorf guestfs_perrorf
+#define safe_malloc guestfs_safe_malloc
+#define safe_realloc guestfs_safe_realloc
+#define safe_strdup guestfs_safe_strdup
+#define safe_memdup guestfs_safe_memdup
 
 static void default_error_cb (guestfs_h *g, void *data, const char *msg);
 static void stdout_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data, int watch, int fd, int events);
@@ -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. */
@@ -156,6 +159,7 @@ struct guestfs_h
   int autosync;
 
   const char *path;
+  const char *qemu;
 
   char *last_error;
 
@@ -174,18 +178,6 @@ struct guestfs_h
   guestfs_launch_done_cb     launch_done_cb;
   void *                     launch_done_cb_data;
 
-  /* These callbacks are called before send_cb, reply_cb and
-   * launch_done_cb, and are used to implement the high-level
-   * API without needing to interfere with callbacks that the
-   * user might have set.
-   */
-  guestfs_send_cb            send_cb_internal;
-  void *                     send_cb_internal_data;
-  guestfs_reply_cb           reply_cb_internal;
-  void *                     reply_cb_internal_data;
-  guestfs_launch_done_cb     launch_done_cb_internal;
-  void *                     launch_done_cb_internal_data;
-
   /* Main loop used by this handle. */
   guestfs_main_loop *main_loop;
 
@@ -229,9 +221,11 @@ guestfs_create (void)
 
   str = getenv ("LIBGUESTFS_PATH");
   g->path = str != NULL ? str : GUESTFS_DEFAULT_PATH;
-  /* XXX We should probably make QEMU configurable as well. */
 
-  g->main_loop = (guestfs_main_loop *) &default_main_loop;
+  str = getenv ("LIBGUESTFS_QEMU");
+  g->qemu = str != NULL ? str : QEMU;
+
+  g->main_loop = guestfs_get_default_main_loop ();
 
   /* Start with large serial numbers so they are easy to spot
    * inside the protocol.
@@ -342,8 +336,8 @@ default_error_cb (guestfs_h *g, void *data, const char *msg)
   fprintf (stderr, "libguestfs: error: %s\n", msg);
 }
 
-static void
-error (guestfs_h *g, const char *fs, ...)
+void
+guestfs_error (guestfs_h *g, const char *fs, ...)
 {
   va_list args;
   char *msg;
@@ -358,8 +352,8 @@ error (guestfs_h *g, const char *fs, ...)
   free (msg);
 }
 
-static void
-perrorf (guestfs_h *g, const char *fs, ...)
+void
+guestfs_perrorf (guestfs_h *g, const char *fs, ...)
 {
   va_list args;
   char *msg;
@@ -388,32 +382,32 @@ perrorf (guestfs_h *g, const char *fs, ...)
   free (msg);
 }
 
-static void *
-safe_malloc (guestfs_h *g, size_t nbytes)
+void *
+guestfs_safe_malloc (guestfs_h *g, size_t nbytes)
 {
   void *ptr = malloc (nbytes);
   if (!ptr) g->abort_cb ();
   return ptr;
 }
 
-static void *
-safe_realloc (guestfs_h *g, void *ptr, int nbytes)
+void *
+guestfs_safe_realloc (guestfs_h *g, void *ptr, int nbytes)
 {
   void *p = realloc (ptr, nbytes);
   if (!p) g->abort_cb ();
   return p;
 }
 
-static char *
-safe_strdup (guestfs_h *g, const char *str)
+char *
+guestfs_safe_strdup (guestfs_h *g, const char *str)
 {
   char *s = strdup (str);
   if (!s) g->abort_cb ();
   return s;
 }
 
-static void *
-safe_memdup (guestfs_h *g, void *ptr, size_t size)
+void *
+guestfs_safe_memdup (guestfs_h *g, void *ptr, size_t size)
 {
   void *p = malloc (size);
   if (!p) g->abort_cb ();
@@ -421,6 +415,40 @@ safe_memdup (guestfs_h *g, void *ptr, size_t size)
   return p;
 }
 
+static int
+xwrite (int fd, const void *buf, size_t len)
+{
+  int r;
+
+  while (len > 0) {
+    r = write (fd, buf, len);
+    if (r == -1)
+      return -1;
+
+    buf += r;
+    len -= r;
+  }
+
+  return 0;
+}
+
+static int
+xread (int fd, void *buf, size_t len)
+{
+  int r;
+
+  while (len > 0) {
+    r = read (fd, buf, len);
+    if (r == -1)
+      return -1;
+
+    buf += r;
+    len -= r;
+  }
+
+  return 0;
+}
+
 void
 guestfs_set_out_of_memory_handler (guestfs_h *g, guestfs_abort_cb cb)
 {
@@ -489,6 +517,22 @@ guestfs_get_path (guestfs_h *g)
   return g->path;
 }
 
+int
+guestfs_set_qemu (guestfs_h *g, const char *qemu)
+{
+  if (qemu == NULL)
+    g->qemu = QEMU;
+  else
+    g->qemu = qemu;
+  return 0;
+}
+
+const char *
+guestfs_get_qemu (guestfs_h *g)
+{
+  return g->qemu;
+}
+
 /* Add a string to the current command line. */
 static void
 incr_cmdline_size (guestfs_h *g)
@@ -591,7 +635,7 @@ int
 guestfs_launch (guestfs_h *g)
 {
   static const char *dir_template = "/tmp/libguestfsXXXXXX";
-  int r, i, len, pmore;
+  int r, i, len, pmore, memsize;
   int wfd[2], rfd[2];
   int tries;
   const char *kernel_name = "vmlinuz." REPO "." host_cpu;
@@ -600,6 +644,7 @@ guestfs_launch (guestfs_h *g)
   char *kernel = NULL, *initrd = NULL;
   char unixsock[256];
   struct sockaddr_un addr;
+  struct stat statbuf;
 
   /* Configured? */
   if (!g->cmdline) {
@@ -659,6 +704,40 @@ guestfs_launch (guestfs_h *g)
     goto cleanup0;
   }
 
+  /* Choose a suitable memory size (in MB).  This is more art
+   * than science, but you can help by doing
+   *   ./configure --enable-debug-command
+   * and then running:
+   *   debug sh free
+   *   debug mem ''
+   * and seeing how much free memory is left for particular
+   * configurations.
+   *
+   * It's also helpful to report both the compressed and uncompressed
+   * size of the initramfs (ls -lh initramfs*.img; du -sh initramfs).
+   *
+   * XXX KVM virtio balloon driver?
+   */
+  if (stat (initrd, &statbuf) != -1) {
+    /* Approximate size of the initramfs after it is decompressed
+     * in kernel memory.  The compression factor is ~2.5-3.
+     */
+    memsize = 3 * statbuf.st_size / 1024 / 1024;
+
+    /* Approximate size used by the kernel. */
+    memsize += 10;
+
+    /* Want to give userspace some room, so: */
+    memsize += 128;
+
+#if SIZEOF_LONG == 8
+    /* On 64 bit, assume some overhead. */
+    memsize += 64;
+#endif
+  } else
+    memsize = 512;
+  
+
   /* Make the temporary directory containing the socket. */
   if (!g->tmpdir) {
     g->tmpdir = safe_strdup (g, dir_template);
@@ -689,11 +768,12 @@ 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] = (char *) g->qemu;
 
     /* Construct the -net channel parameter for qemu. */
     snprintf (vmchannel, sizeof vmchannel,
@@ -702,11 +782,17 @@ guestfs_launch (guestfs_h *g)
 
     /* Linux kernel command line. */
     snprintf (append, sizeof append,
-	      "console=ttyS0 guestfs=%s:%d", VMCHANNEL_ADDR, VMCHANNEL_PORT);
+	      "console=ttyS0 guestfs=%s:%d%s",
+	      VMCHANNEL_ADDR, VMCHANNEL_PORT,
+	      g->verbose ? " guestfs_verbose=1" : "");
+
+    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, "-kernel");
     add_cmdline (g, (char *) kernel);
     add_cmdline (g, "-initrd");
@@ -721,12 +807,12 @@ 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=rtl8139,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");
@@ -749,14 +835,48 @@ 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;
 
+  /* 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);
 
@@ -834,14 +954,8 @@ guestfs_launch (guestfs_h *g)
     goto cleanup3;
   }
 
-  g->sock_watch =
-    g->main_loop->add_handle (g->main_loop, g, g->sock,
-			      GUESTFS_HANDLE_READABLE,
-			      sock_read_event, NULL);
-  if (g->sock_watch == -1) {
-    error (g, "could not watch daemon communications socket");
+  if (guestfs__switch_to_receiving (g) == -1)
     goto cleanup3;
-  }
 
   g->state = LAUNCHING;
   return 0;
@@ -859,11 +973,14 @@ 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;
@@ -901,11 +1018,11 @@ guestfs_wait_ready (guestfs_h *g)
     return -1;
   }
 
-  g->launch_done_cb_internal = finish_wait_ready;
-  g->launch_done_cb_internal_data = &finished;
+  g->launch_done_cb = finish_wait_ready;
+  g->launch_done_cb_data = &finished;
   r = g->main_loop->main_loop_run (g->main_loop, g);
-  g->launch_done_cb_internal = NULL;
-  g->launch_done_cb_internal_data = NULL;
+  g->launch_done_cb = NULL;
+  g->launch_done_cb_data = NULL;
 
   if (r == -1) return -1;
 
@@ -939,10 +1056,95 @@ 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;
+}
+
+/* Access current state. */
+int
+guestfs_is_config (guestfs_h *g)
+{
+  return g->state == CONFIG;
+}
+
+int
+guestfs_is_launching (guestfs_h *g)
+{
+  return g->state == LAUNCHING;
+}
 
+int
+guestfs_is_ready (guestfs_h *g)
+{
+  return g->state == READY;
+}
+
+int
+guestfs_is_busy (guestfs_h *g)
+{
+  return g->state == BUSY;
+}
+
+int
+guestfs_get_state (guestfs_h *g)
+{
+  return g->state;
+}
+
+int
+guestfs_set_ready (guestfs_h *g)
+{
+  if (g->state != BUSY) {
+    error (g, "guestfs_set_ready: called when in state %d != BUSY", g->state);
+    return -1;
+  }
+  g->state = READY;
   return 0;
 }
 
+int
+guestfs_set_busy (guestfs_h *g)
+{
+  if (g->state != READY) {
+    error (g, "guestfs_set_busy: called when in state %d != READY", g->state);
+    return -1;
+  }
+  g->state = BUSY;
+  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.
+ */
+void
+guestfs_free_int_bool (struct guestfs_int_bool *x)
+{
+  free (x);
+}
+
+void
+guestfs_free_lvm_pv_list (struct guestfs_lvm_pv_list *x)
+{
+  xdr_free ((xdrproc_t) xdr_guestfs_lvm_int_pv_list, (char *) x);
+  free (x);
+}
+
+void
+guestfs_free_lvm_vg_list (struct guestfs_lvm_vg_list *x)
+{
+  xdr_free ((xdrproc_t) xdr_guestfs_lvm_int_vg_list, (char *) x);
+  free (x);
+}
+
+void
+guestfs_free_lvm_lv_list (struct guestfs_lvm_lv_list *x)
+{
+  xdr_free ((xdrproc_t) xdr_guestfs_lvm_int_lv_list, (char *) x);
+  free (x);
+}
+
 /* 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.
@@ -972,7 +1174,9 @@ stdout_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
     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)
@@ -984,6 +1188,7 @@ stdout_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
     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;
@@ -1050,6 +1255,7 @@ sock_read_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
   g->msg_in_size += n;
 
   /* Have we got enough of a message to be able to process it yet? */
+ again:
   if (g->msg_in_size < 4) return;
 
   xdrmem_create (&xdr, g->msg_in, g->msg_in_size, XDR_DECODE);
@@ -1062,7 +1268,7 @@ sock_read_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
    * starts up it sends a "magic" value (longer than any possible
    * message).  Check for this.
    */
-  if (len == 0xf5f55ff5) {
+  if (len == GUESTFS_LAUNCH_FLAG) {
     if (g->state != LAUNCHING)
       error (g, "received magic signature from guestfsd, but in state %d",
 	     g->state);
@@ -1071,8 +1277,6 @@ sock_read_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
 	     g->msg_in_size);
     else {
       g->state = READY;
-      if (g->launch_done_cb_internal)
-	g->launch_done_cb_internal (g, g->launch_done_cb_internal_data);
       if (g->launch_done_cb)
 	g->launch_done_cb (g, g->launch_done_cb_data);
     }
@@ -1080,7 +1284,19 @@ sock_read_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
     goto cleanup;
   }
 
-  /* If this happens, it's pretty bad and we've probably lost synchronization.*/
+  /* This can happen if a cancellation happens right at the end
+   * of us sending a FileIn parameter to the daemon.  Discard.  The
+   * daemon should send us an error message next.
+   */
+  if (len == GUESTFS_CANCEL_FLAG) {
+    g->msg_in_size -= 4;
+    memmove (g->msg_in, g->msg_in+4, g->msg_in_size);
+    goto again;
+  }
+
+  /* If this happens, it's pretty bad and we've probably lost
+   * synchronization.
+   */
   if (len > GUESTFS_MESSAGE_MAX) {
     error (g, "message length (%u) > maximum possible size (%d)",
 	   len, GUESTFS_MESSAGE_MAX);
@@ -1089,15 +1305,8 @@ 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. */
 
-  /* This should not happen, and if it does it probably means we've
-   * lost all hope of synchronization.
-   */
-  if (g->msg_in_size-4 > len) {
-    error (g, "len = %d, but msg_in_size-4 = %d", len, g->msg_in_size-4);
-    goto cleanup;
-  }
-
   /* Got the full message, begin processing it. */
+#if 0
   if (g->verbose) {
     int i, j;
 
@@ -1118,20 +1327,22 @@ 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)
     error (g, "state %d != BUSY", g->state);
 
-  /* Push the message up to the higher layer.  Note that unlike
-   * launch_done_cb / launch_done_cb_internal, we only call at
-   * most one of the callback functions here.
-   */
-  g->state = READY;
-  if (g->reply_cb_internal)
-    g->reply_cb_internal (g, g->reply_cb_internal_data, &xdr);
-  else if (g->reply_cb)
-    g->reply_cb (g, g->reply_cb, &xdr);
+  /* 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);
+  if (g->msg_in_size > 0) goto again;
 
  cleanup:
   /* Free the message buffer if it's grown excessively large. */
@@ -1191,43 +1402,172 @@ sock_write_event (struct guestfs_main_loop *ml, guestfs_h *g, void *data,
     return;
 
   if (g->verbose)
-    fprintf (stderr, "sock_write_event: done writing, switching back to reading events\n");
+    fprintf (stderr, "sock_write_event: done writing, calling send_cb\n");
 
   free (g->msg_out);
+  g->msg_out = NULL;
   g->msg_out_pos = g->msg_out_size = 0;
 
-  if (g->main_loop->remove_handle (g->main_loop, g, g->sock_watch) == -1) {
-    error (g, "remove_handle failed in sock_write_event");
-    return;
+  /* Done writing, call the higher layer. */
+  if (g->send_cb)
+    g->send_cb (g, g->send_cb_data);
+}
+
+void
+guestfs_set_send_callback (guestfs_h *g,
+			   guestfs_send_cb cb, void *opaque)
+{
+  g->send_cb = cb;
+  g->send_cb_data = opaque;
+}
+
+void
+guestfs_set_reply_callback (guestfs_h *g,
+			    guestfs_reply_cb cb, void *opaque)
+{
+  g->reply_cb = cb;
+  g->reply_cb_data = opaque;
+}
+
+void
+guestfs_set_log_message_callback (guestfs_h *g,
+				  guestfs_log_message_cb cb, void *opaque)
+{
+  g->log_message_cb = cb;
+  g->log_message_cb_data = opaque;
+}
+
+void
+guestfs_set_subprocess_quit_callback (guestfs_h *g,
+				      guestfs_subprocess_quit_cb cb, void *opaque)
+{
+  g->subprocess_quit_cb = cb;
+  g->subprocess_quit_cb_data = opaque;
+}
+
+void
+guestfs_set_launch_done_callback (guestfs_h *g,
+				  guestfs_launch_done_cb cb, void *opaque)
+{
+  g->launch_done_cb = cb;
+  g->launch_done_cb_data = opaque;
+}
+
+/* Access to the handle's main loop and the default main loop. */
+void
+guestfs_set_main_loop (guestfs_h *g, guestfs_main_loop *main_loop)
+{
+  g->main_loop = main_loop;
+}
+
+guestfs_main_loop *
+guestfs_get_main_loop (guestfs_h *g)
+{
+  return g->main_loop;
+}
+
+guestfs_main_loop *
+guestfs_get_default_main_loop (void)
+{
+  return (guestfs_main_loop *) &default_main_loop;
+}
+
+/* Change the daemon socket handler so that we are now writing.
+ * This sets the handle to sock_write_event.
+ */
+int
+guestfs__switch_to_sending (guestfs_h *g)
+{
+  if (g->sock_watch >= 0) {
+    if (g->main_loop->remove_handle (g->main_loop, g, g->sock_watch) == -1) {
+      error (g, "remove_handle failed");
+      g->sock_watch = -1;
+      return -1;
+    }
+  }
+
+  g->sock_watch =
+    g->main_loop->add_handle (g->main_loop, g, g->sock,
+			      GUESTFS_HANDLE_WRITABLE,
+			      sock_write_event, NULL);
+  if (g->sock_watch == -1) {
+    error (g, "add_handle failed");
+    return -1;
+  }
+
+  return 0;
+}
+
+int
+guestfs__switch_to_receiving (guestfs_h *g)
+{
+  if (g->sock_watch >= 0) {
+    if (g->main_loop->remove_handle (g->main_loop, g, g->sock_watch) == -1) {
+      error (g, "remove_handle failed");
+      g->sock_watch = -1;
+      return -1;
+    }
   }
+
   g->sock_watch =
     g->main_loop->add_handle (g->main_loop, g, g->sock,
 			      GUESTFS_HANDLE_READABLE,
 			      sock_read_event, NULL);
   if (g->sock_watch == -1) {
-    error (g, "add_handle failed in sock_write_event");
-    return;
+    error (g, "add_handle failed");
+    return -1;
   }
+
+  return 0;
 }
 
-/* Dispatch a call to the remote daemon.  This function just queues
- * the call in msg_out, to be sent when we next enter the main loop.
- * Returns -1 for error, or the message serial number.
+/* Dispatch a call (len + header + args) to the remote daemon,
+ * synchronously (ie. using the guest's main loop to wait until
+ * it has been sent).  Returns -1 for error, or the serial
+ * number of the message.
  */
-static int
-dispatch (guestfs_h *g, int proc_nr, xdrproc_t xdrp, char *args)
+static void
+send_cb (guestfs_h *g, void *data)
+{
+  guestfs_main_loop *ml = guestfs_get_main_loop (g);
+
+  *((int *)data) = 1;
+  ml->main_loop_quit (ml, g);
+}
+
+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;
   int serial = g->msg_next_serial++;
+  int sent;
+  guestfs_main_loop *ml = guestfs_get_main_loop (g);
 
-  if (g->state != READY) {
-    error (g, "dispatch: state %d != READY", g->state);
+  if (g->state != BUSY) {
+    error (g, "guestfs__send_sync: state %d != BUSY", g->state);
     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;
@@ -1236,10 +1576,9 @@ dispatch (guestfs_h *g, int proc_nr, xdrproc_t xdrp, char *args)
   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
@@ -1248,39 +1587,32 @@ dispatch (guestfs_h *g, int proc_nr, xdrproc_t xdrp, char *args)
   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;
-  g->state = BUSY;
 
   xdrmem_create (&xdr, g->msg_out, 4, XDR_ENCODE);
-  if (!xdr_uint32_t (&xdr, &len)) {
-    error (g, "xdr_uint32_t failed in dispatch");
-    goto cleanup1;
-  }
+  xdr_uint32_t (&xdr, &len);
 
-  memcpy (g->msg_out + 4, buffer, len);
+  if (guestfs__switch_to_sending (g) == -1)
+    goto cleanup1;
 
-  /* Change the handle to sock_write_event. */
-  if (g->main_loop->remove_handle (g->main_loop, g, g->sock_watch) == -1) {
-    error (g, "remove_handle failed in dispatch");
+  sent = 0;
+  guestfs_set_send_callback (g, send_cb, &sent);
+  if (ml->main_loop_run (ml, g) == -1)
     goto cleanup1;
-  }
-  g->sock_watch =
-    g->main_loop->add_handle (g->main_loop, g, g->sock,
-			      GUESTFS_HANDLE_WRITABLE,
-			      sock_write_event, NULL);
-  if (g->sock_watch == -1) {
-    error (g, "add_handle failed in dispatch");
+  if (sent != 1) {
+    error (g, "send failed, see earlier error messages");
     goto cleanup1;
   }
 
@@ -1290,23 +1622,31 @@ dispatch (guestfs_h *g, int proc_nr, xdrproc_t xdrp, char *args)
   free (g->msg_out);
   g->msg_out = NULL;
   g->msg_out_size = 0;
-  g->state = READY;
   return -1;
 }
 
-#if 0
 static int cancel = 0; /* XXX Implement file cancellation. */
-
-static int
-send_file (guestfs_h *g, const char *filename)
+static int send_file_chunk_sync (guestfs_h *g, int cancel, const char *buf, size_t len);
+static int send_file_data_sync (guestfs_h *g, const char *buf, size_t len);
+static int send_file_cancellation_sync (guestfs_h *g);
+static int send_file_complete_sync (guestfs_h *g);
+
+/* Synchronously send a file.
+ * Returns:
+ *   0 OK
+ *   -1 error
+ *   -2 daemon cancelled (we must read the error message)
+ */
+int
+guestfs__send_file_sync (guestfs_h *g, const char *filename)
 {
   char buf[GUESTFS_MAX_CHUNK_SIZE];
-  int fd, r;
+  int fd, r, err;
 
   fd = open (filename, O_RDONLY);
   if (fd == -1) {
     perrorf (g, "open: %s", filename);
-    send_file_cancellation (g);
+    send_file_cancellation_sync (g);
     /* Daemon sees cancellation and won't reply, so caller can
      * just return here.
      */
@@ -1315,18 +1655,22 @@ send_file (guestfs_h *g, const char *filename)
 
   /* Send file in chunked encoding. */
   while (!cancel && (r = read (fd, buf, sizeof buf)) > 0) {
-    if (send_file_data (g, buf, r) == -1)
-      return -1;
+    err = send_file_data_sync (g, buf, r);
+    if (err < 0) {
+      if (err == -2)		/* daemon sent cancellation */
+	send_file_cancellation_sync (g);
+      return err;
+    }
   }
 
-  if (cancel) {
-    send_file_cancellation (g);
+  if (cancel) {			/* cancel from either end */
+    send_file_cancellation_sync (g);
     return -1;
   }
 
   if (r == -1) {
     perrorf (g, "read: %s", filename);
-    send_file_cancellation (g);
+    send_file_cancellation_sync (g);
     return -1;
   }
 
@@ -1335,61 +1679,107 @@ send_file (guestfs_h *g, const char *filename)
    */
   if (close (fd) == -1) {
     perrorf (g, "close: %s", filename);
-    send_file_cancellation (g);
+    send_file_cancellation_sync (g);
     return -1;
   }
 
-  return send_file_complete (g);
+  return send_file_complete_sync (g);
 }
 
-/* Send a chunk, cancellation or end of file, wait for it to go. */
+/* Send a chunk of file data. */
 static int
-send_file_chunk (guestfs_h *g, int cancel, const char *buf, size_t len)
+send_file_data_sync (guestfs_h *g, const char *buf, size_t len)
 {
-  void *data;
+  return send_file_chunk_sync (g, 0, buf, len);
+}
+
+/* Send a cancellation message. */
+static int
+send_file_cancellation_sync (guestfs_h *g)
+{
+  return send_file_chunk_sync (g, 1, NULL, 0);
+}
+
+/* Send a file complete chunk. */
+static int
+send_file_complete_sync (guestfs_h *g)
+{
+  char buf[1];
+  return send_file_chunk_sync (g, 0, buf, 0);
+}
+
+/* Send a chunk, cancellation or end of file, synchronously (ie. wait
+ * for it to go).
+ */
+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 buflen)
+{
+  unsigned len;
+  int sent;
   guestfs_chunk chunk;
   XDR xdr;
+  guestfs_main_loop *ml = guestfs_get_main_loop (g);
 
   if (g->state != BUSY) {
-    error (g, "send_file_chunk: state %d != READY", g->state);
+    error (g, "send_file_chunk_sync: state %d != READY", g->state);
+    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 (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;
 
-  data = safe_malloc (g, GUESTFS_MAX_CHUNK_SIZE + 48);
-  xdrmem_create (&xdr, data, GUESTFS_MAX_CHUNK_SIZE + 48, XDR_ENCODE);
-  if (xdr_guestfs_chunk (&xdr, &chunk)) {
-    error (g, "xdr_guestfs_chunk failed");
-    free (data);
-    return -1;
+  if (!xdr_guestfs_chunk (&xdr, &chunk)) {
+    error (g, "xdr_guestfs_chunk failed (buf = %p, buflen = %zu)",
+	   buf, buflen);
+    xdr_destroy (&xdr);
+    goto cleanup1;
   }
 
-  chunkdatalen = xdr_getpos (&xdr);
-  xdr_destroy (&xdr);
-
   len = xdr_getpos (&xdr);
   xdr_destroy (&xdr);
 
-  data = safe_realloc (g, data, len);
-  g->msg_out = data;
-  g->msg_out_size = len;
+  /* 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;
 
-  /* Change the handle to sock_write_event. */
-  if (g->main_loop->remove_handle (g->main_loop, g, g->sock_watch) == -1) {
-    error (g, "remove_handle failed in dispatch");
+  xdrmem_create (&xdr, g->msg_out, 4, XDR_ENCODE);
+  xdr_uint32_t (&xdr, &len);
+
+  if (guestfs__switch_to_sending (g) == -1)
     goto cleanup1;
-  }
-  g->sock_watch =
-    g->main_loop->add_handle (g->main_loop, g, g->sock,
-			      GUESTFS_HANDLE_WRITABLE,
-			      sock_write_event, NULL);
-  if (g->sock_watch == -1) {
-    error (g, "add_handle failed in dispatch");
+
+  sent = 0;
+  guestfs_set_send_callback (g, send_cb, &sent);
+  if (ml->main_loop_run (ml, g) == -1)
+    goto cleanup1;
+  if (sent != 1) {
+    error (g, "send file chunk failed, see earlier error messages");
     goto cleanup1;
   }
 
@@ -1399,100 +1789,229 @@ send_file_chunk (guestfs_h *g, int cancel, const char *buf, size_t len)
   free (g->msg_out);
   g->msg_out = NULL;
   g->msg_out_size = 0;
-  g->state = READY;
   return -1;
 }
 
-/* Send a chunk of file data. */
+/* At this point we are sending FileIn file(s) to the guest, and not
+ * expecting to read anything, so if we do read anything, it must be
+ * a cancellation message.  This checks for this case without blocking.
+ */
 static int
-send_file_data (guestfs_h *g, const char *buf, size_t len)
+check_for_daemon_cancellation (guestfs_h *g)
 {
-  return send_file_chunk (g, 0, buf, len);
-}
+  fd_set rset;
+  struct timeval tv;
+  int r;
+  char buf[4];
+  uint32_t flag;
+  XDR xdr;
 
-/* Send a cancellation message. */
-static int
-send_file_cancellation (guestfs_h *g)
-{
-  char buf[1];
-  return send_file_chunk (g, 1, buf, 0);
-}
+  FD_ZERO (&rset);
+  FD_SET (g->sock, &rset);
+  tv.tv_sec = 0;
+  tv.tv_usec = 0;
+  r = select (g->sock+1, &rset, NULL, NULL, &tv);
+  if (r == -1) {
+    perrorf (g, "select");
+    return 0;
+  }
+  if (r == 0)
+    return 0;
 
-/* Send a file complete chunk. */
-static int
-send_file_complete (guestfs_h *g)
-{
-  char buf[0];
-  return send_file_chunk (g, 0, buf, 0);
+  /* Read the message from the daemon. */
+  r = xread (g->sock, buf, sizeof buf);
+  if (r == -1) {
+    perrorf (g, "read");
+    return 0;
+  }
+
+  xdrmem_create (&xdr, buf, sizeof buf, XDR_DECODE);
+  xdr_uint32_t (&xdr, &flag);
+  xdr_destroy (&xdr);
+
+  if (flag != GUESTFS_CANCEL_FLAG) {
+    error (g, "check_for_daemon_cancellation: read 0x%x from daemon, expected 0x%x\n",
+	   flag, GUESTFS_CANCEL_FLAG);
+    return 0;
+  }
+
+  return 1;
 }
-#endif
 
-/* Check the return message from a call for validity. */
-static int
-check_reply_header (guestfs_h *g,
-		    const struct guestfs_message_header *hdr,
-		    int proc_nr, int serial)
+/* Synchronously receive a file. */
+
+/* Returns -1 = error, 0 = EOF, 1 = more data */
+static int receive_file_data_sync (guestfs_h *g, void **buf, int *len);
+
+int
+guestfs__receive_file_sync (guestfs_h *g, const char *filename)
 {
-  if (hdr->prog != GUESTFS_PROGRAM) {
-    error (g, "wrong program (%d/%d)", hdr->prog, GUESTFS_PROGRAM);
-    return -1;
+  void *buf;
+  int fd, r, len;
+
+  fd = open (filename, O_WRONLY|O_CREAT|O_TRUNC|O_NOCTTY, 0666);
+  if (fd == -1) {
+    perrorf (g, "open: %s", filename);
+    goto cancel;
   }
-  if (hdr->vers != GUESTFS_PROTOCOL_VERSION) {
-    error (g, "wrong protocol version (%d/%d)",
-	   hdr->vers, GUESTFS_PROTOCOL_VERSION);
-    return -1;
+
+  /* Receive the file in chunked encoding. */
+  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 (hdr->direction != GUESTFS_DIRECTION_REPLY) {
-    error (g, "unexpected message direction (%d/%d)",
-	   hdr->direction, GUESTFS_DIRECTION_REPLY);
+
+  if (r == -1) {
+    error (g, "%s: error in chunked encoding", filename);
     return -1;
   }
-  if (hdr->proc != proc_nr) {
-    error (g, "unexpected procedure number (%d/%d)", hdr->proc, proc_nr);
+
+  if (close (fd) == -1) {
+    perrorf (g, "close: %s", filename);
     return -1;
   }
-  if (hdr->serial != serial) {
-    error (g, "unexpected serial (%d/%d)", hdr->serial, serial);
+
+  return 0;
+
+ cancel: ;
+  /* Send cancellation message to daemon, then wait until it
+   * cancels (just throwing away data).
+   */
+  XDR xdr;
+  char fbuf[4];
+  uint32_t flag = GUESTFS_CANCEL_FLAG;
+
+  xdrmem_create (&xdr, fbuf, sizeof fbuf, XDR_ENCODE);
+  xdr_uint32_t (&xdr, &flag);
+  xdr_destroy (&xdr);
+
+  if (xwrite (g->sock, fbuf, sizeof fbuf) == -1) {
+    perrorf (g, "write to daemon socket");
     return -1;
   }
 
-  return 0;
+  while ((r = receive_file_data_sync (g, NULL, NULL)) > 0)
+    ;				/* just discard it */
+
+  return -1;
 }
 
-/* The high-level actions are autogenerated by generator.ml.  Include
- * them here.
+/* 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.
  */
-#include "guestfs-actions.c"
+struct receive_file_ctx {
+  int count;			/* 0 if receive_file_cb not called, or
+				 * else count number of chunks.
+				 */
+  guestfs_chunk *chunks;	/* Array of chunks. */
+};
 
-/* Structure-freeing functions.  These rely on the fact that the
- * structure format is identical to the XDR format.  See note in
- * generator.ml.
- */
-void
-guestfs_free_int_bool (struct guestfs_int_bool *x)
+static void
+free_chunks (struct receive_file_ctx *ctx)
 {
-  free (x);
-}
+  int i;
 
-void
-guestfs_free_lvm_pv_list (struct guestfs_lvm_pv_list *x)
-{
-  xdr_free ((xdrproc_t) xdr_guestfs_lvm_int_pv_list, (char *) x);
-  free (x);
+  for (i = 0; i < ctx->count; ++i)
+    free (ctx->chunks[i].data.data_val);
+
+  free (ctx->chunks);
 }
 
-void
-guestfs_free_lvm_vg_list (struct guestfs_lvm_vg_list *x)
+static void
+receive_file_cb (guestfs_h *g, void *data, XDR *xdr)
 {
-  xdr_free ((xdrproc_t) xdr_guestfs_lvm_int_vg_list, (char *) x);
-  free (x);
+  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");
+    free_chunks (ctx);
+    ctx->chunks = NULL;
+    ctx->count = -1;
+    return;
+  }
+
+  /* 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++;
 }
 
-void
-guestfs_free_lvm_lv_list (struct guestfs_lvm_lv_list *x)
+/* 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, int *len_r)
 {
-  xdr_free ((xdrproc_t) xdr_guestfs_lvm_int_lv_list, (char *) x);
-  free (x);
+  struct receive_file_ctx ctx;
+  guestfs_main_loop *ml = guestfs_get_main_loop (g);
+  int i, len;
+
+  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 (ctx.count == 0) {
+    error (g, "receive_file_data_sync: reply callback not called\n");
+    return -1;
+  }
+
+  if (ctx.count == -1) {
+    error (g, "receive_file_data_sync: parse error in reply callback\n");
+    /* callback already freed the chunks */
+    return -1;
+  }
+
+  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). */
@@ -1665,10 +2184,9 @@ select_main_loop_quit (guestfs_main_loop *mlv, guestfs_h *g)
 {
   struct select_main_loop *ml = (struct select_main_loop *) mlv;
 
-  if (!ml->is_running) {
-    error (g, "cannot quit, we are not running in a main loop");
-    return -1;
-  }
+  /* Note that legitimately ml->is_running can be zero when
+   * this function is called.
+   */
 
   ml->is_running = 0;
   return 0;