#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>
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. */
int autosync;
const char *path;
+ const char *qemu;
char *last_error;
str = getenv ("LIBGUESTFS_PATH");
g->path = str != NULL ? str : GUESTFS_DEFAULT_PATH;
- /* XXX We should probably make QEMU configurable as well. */
+
+ str = getenv ("LIBGUESTFS_QEMU");
+ g->qemu = str != NULL ? str : QEMU;
g->main_loop = guestfs_get_default_main_loop ();
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)
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;
char *kernel = NULL, *initrd = NULL;
char unixsock[256];
struct sockaddr_un addr;
+ struct stat statbuf;
/* Configured? */
if (!g->cmdline) {
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);
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,
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. */
-#if defined(__i386__) || defined(__x86_64)
+ add_cmdline (g, memsize_str);
+#if 0
add_cmdline (g, "-no-kqemu"); /* Avoids a warning. */
#endif
add_cmdline (g, "-kernel");
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");
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);
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;
fprintf (stderr, "sending SIGTERM to process %d\n", g->pid);
kill (g->pid, SIGTERM);
+ if (g->recoverypid > 0) kill (g->recoverypid, 9);
return 0;
}
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->fd[1] = -1;
g->sock = -1;
g->pid = 0;
+ g->recoverypid = 0;
g->start_t = 0;
g->stdout_watch = -1;
g->sock_watch = -1;
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;
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;
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
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;
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;
+ unsigned len;
int sent;
guestfs_chunk chunk;
XDR xdr;
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)
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;
- if (g->verbose)
- fprintf (stderr,
- "library sending chunk cancel = %d, len = %zu, buf = %p\n",
- cancel, len, 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;
git.annexia.org / libguestfs.git / blobdiff