#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>
#include "guestfs.h"
#include "guestfs_protocol.h"
-void guestfs_error (guestfs_h *g, const char *fs, ...);
-void guestfs_perrorf (guestfs_h *g, const char *fs, ...);
-void *guestfs_safe_malloc (guestfs_h *g, size_t nbytes);
-void *guestfs_safe_realloc (guestfs_h *g, void *ptr, int nbytes);
-char *guestfs_safe_strdup (guestfs_h *g, const char *str);
-void *guestfs_safe_memdup (guestfs_h *g, void *ptr, size_t size);
-
#define error guestfs_error
#define perrorf guestfs_perrorf
#define safe_malloc guestfs_safe_malloc
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 ();
while (len > 0) {
r = write (fd, buf, len);
- if (r == -1) {
- perror ("write");
+ 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 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,
/* 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");
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;
}
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.
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;
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);
* 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);
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);
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 (g->verbose) {
int i, j;
error (g, "state %d != BUSY", g->state);
/* Push the message up to the higher layer. */
- g->state = READY;
if (g->reply_cb)
g->reply_cb (g, g->reply_cb_data, &xdr);
+ 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. */
if (g->msg_in_allocated > 65536) {
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;
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 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;
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;
- 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;
- }
-
- memcpy (g->msg_out + 4, buffer, len);
+ xdr_uint32_t (&xdr, &len);
if (guestfs__switch_to_sending (g) == -1)
goto cleanup1;
free (g->msg_out);
g->msg_out = NULL;
g->msg_out_size = 0;
- g->state = READY;
return -1;
}
static int send_file_cancellation_sync (guestfs_h *g);
static int send_file_complete_sync (guestfs_h *g);
-/* Synchronously send a file. */
+/* 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) {
/* Send file in chunked encoding. */
while (!cancel && (r = read (fd, buf, sizeof buf)) > 0) {
- if (send_file_data_sync (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) {
+ if (cancel) { /* cancel from either end */
send_file_cancellation_sync (g);
return -1;
}
static int
send_file_cancellation_sync (guestfs_h *g)
{
- char buf[1];
- return send_file_chunk_sync (g, 1, buf, 0);
+ return send_file_chunk_sync (g, 1, NULL, 0);
}
/* Send a file complete chunk. */
/* 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 len)
+send_file_chunk_sync (guestfs_h *g, int cancel, const char *buf, size_t buflen)
{
- void *data;
- unsigned datalen;
+ 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;
}
- datalen = xdr_getpos (&xdr);
+ len = xdr_getpos (&xdr);
xdr_destroy (&xdr);
- data = safe_realloc (g, data, datalen);
- g->msg_out = data;
- g->msg_out_size = datalen;
+ /* 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, &len);
+
if (guestfs__switch_to_sending (g) == -1)
goto cleanup1;
free (g->msg_out);
g->msg_out = NULL;
g->msg_out_size = 0;
- g->state = READY;
return -1;
}
-/* Synchronously receive a file.
- * XXX No way to cancel file receives. We would need to send an
- * error to the daemon and have it see this and stop sending.
+/* 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
+check_for_daemon_cancellation (guestfs_h *g)
+{
+ fd_set rset;
+ struct timeval tv;
+ int r;
+ char buf[4];
+ uint32_t flag;
+ XDR xdr;
+
+ 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;
+
+ /* 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;
+}
+
+/* Synchronously receive a file. */
+
static int receive_file_data_sync (guestfs_h *g, void **buf);
int
void *buf;
int fd, r;
- fd = open (filename, O_WRONLY|O_CREAT|O_TRUNC|O_NOCTTY);
+ fd = open (filename, O_WRONLY|O_CREAT|O_TRUNC|O_NOCTTY, 0666);
if (fd == -1) {
perrorf (g, "open: %s", filename);
- return -1;
+ goto cancel;
}
/* Receive the file in chunked encoding. */
while ((r = receive_file_data_sync (g, &buf)) > 0) {
if (xwrite (fd, buf, r) == -1) {
+ perrorf (g, "%s: write", filename);
free (buf);
- return -1;
+ goto cancel;
}
free (buf);
}
}
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;
+ }
+
+ while ((r = receive_file_data_sync (g, &buf)) > 0)
+ free (buf); /* just discard it */
+
+ return -1;
}
struct receive_file_ctx {
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;
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;
{
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;
git.annexia.org / libguestfs.git / blobdiff