qemu-ga: persist tracking of fsfreeze state via filesystem

Currently, qemu-ga may die/get killed/go away for whatever reason after
guest-fsfreeze-freeze has been issued, and before guest-fsfreeze-thaw
has been issued. This means the only way to unfreeze the guest is via
VNC/network/console access, but obtaining that access after-the-fact can
often be very difficult when filesystems are frozen. Logins will almost
always hang, for instance. In many cases the only recourse would be to
reboot the guest without any quiescing of volatile state, which makes
this a corner-case worth giving some attention to.

A likely failsafe for this situation would be to use a watchdog to
restart qemu-ga if it goes away. There are some precautions qemu-ga
needs to take in order to avoid immediately hanging itself on I/O,
however, namely, we must disable logging and defer to processing/creation
of user-specific logfiles, along with creation of the pid file if we're
running as a daemon. We also need to disable non-fsfreeze-safe commands,
as we normally would when processing the guest-fsfreeze-freeze command.

To track when we need to do this in a way that persists between multiple
invocations of qemu-ga, we create a file on the guest filesystem before
issuing the fsfreeze, and delete it when doing the thaw. On qemu-ga
startup, we check for the existance of this file to determine
the need to take the above precautions.

We're forced to do it this way since a more traditional approach such as
reading/writing state to a dedicated state file will cause
access/modification time updates, respectively, both of which will hang
if the file resides on a frozen filesystem. Both can occur even if
relatime is enabled. Checking for file existence will not update the
access time, however, so it's a safe way to check for fsfreeze state.

An actual watchdog-based restart of qemu-ga can itself cause an access
time update that would thus hang the invocation of qemu-ga, but the
logic to workaround that can be handled via the watchdog, so we don't
address that here (for relatime we'd periodically touch the qemu-ga
binary if the file $qga_statedir/qga.state.isfrozen is not present, this
avoids qemu-ga updates or the 1 day relatime threshold causing an
access-time update if we try to respawn qemu-ga shortly after it goes
away)

Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>

2 files changed, 174 insertions, 47 deletions

diff --git a/qapi-schema-guest.json b/qapi-schema-guest.json
index 0eedb98765..d7a073ee70 100644
--- a/qapi-schema-guest.json
+++ b/qapi-schema-guest.json
@@ -309,8 +309,7 @@
 # Returns: GuestFsfreezeStatus ("thawed", "frozen", etc., as defined below)
 #
 # Note: This may fail to properly report the current state as a result of
-# qemu-ga having been restarted, or other guest processes having issued
-# an fs freeze/thaw.
+# some other guest processes having issued an fs freeze/thaw.
 #
 # Since: 0.15.0
 ##
diff --git a/qemu-ga.c b/qemu-ga.c
index ac29b733ef..216be39072 100644
--- a/qemu-ga.c
+++ b/qemu-ga.c
@@ -18,6 +18,7 @@
 #ifndef _WIN32
 #include <syslog.h>
 #include <sys/wait.h>
+#include <sys/stat.h>
 #endif
 #include "json-streamer.h"
 #include "json-parser.h"
@@ -41,6 +42,7 @@
 #define QGA_VIRTIO_PATH_DEFAULT "\\\\.\\Global\\org.qemu.guest_agent.0"
 #endif
 #define QGA_PIDFILE_DEFAULT "/var/run/qemu-ga.pid"
+#define QGA_STATEDIR_DEFAULT "/tmp"
 #define QGA_SENTINEL_BYTE 0xFF
 
 struct GAState {
@@ -58,6 +60,11 @@ struct GAState {
     bool delimit_response;
     bool frozen;
     GList *blacklist;
+    const char *state_filepath_isfrozen;
+    struct {
+        const char *log_filepath;
+        const char *pid_filepath;
+    } deferred_options;
 };
 
 struct GAState *ga_state;
@@ -147,6 +154,8 @@ static void usage(const char *cmd)
 "                    %s)\n"
 "  -l, --logfile     set logfile path, logs to stderr by default\n"
 "  -f, --pidfile     specify pidfile (default is %s)\n"
+"  -t, --statedir    specify dir to store state information (absolute paths\n"
+"                    only, default is %s)\n"
 "  -v, --verbose     log extra debugging information\n"
 "  -V, --version     print version information and exit\n"
 "  -d, --daemonize   become a daemon\n"
@@ -158,7 +167,8 @@ static void usage(const char *cmd)
 "  -h, --help        display this help and exit\n"
 "\n"
 "Report bugs to <mdroth@linux.vnet.ibm.com>\n"
-    , cmd, QGA_VERSION, QGA_VIRTIO_PATH_DEFAULT, QGA_PIDFILE_DEFAULT);
+    , cmd, QGA_VERSION, QGA_VIRTIO_PATH_DEFAULT, QGA_PIDFILE_DEFAULT,
+    QGA_STATEDIR_DEFAULT);
 }
 
 static const char *ga_log_level_str(GLogLevelFlags level)
@@ -227,6 +237,41 @@ void ga_set_response_delimited(GAState *s)
     s->delimit_response = true;
 }
 
+#ifndef _WIN32
+static bool ga_open_pidfile(const char *pidfile)
+{
+    int pidfd;
+    char pidstr[32];
+
+    pidfd = open(pidfile, O_CREAT|O_WRONLY, S_IRUSR|S_IWUSR);
+    if (pidfd == -1 || lockf(pidfd, F_TLOCK, 0)) {
+        g_critical("Cannot lock pid file, %s", strerror(errno));
+        return false;
+    }
+
+    if (ftruncate(pidfd, 0) || lseek(pidfd, 0, SEEK_SET)) {
+        g_critical("Failed to truncate pid file");
+        goto fail;
+    }
+    sprintf(pidstr, "%d", getpid());
+    if (write(pidfd, pidstr, strlen(pidstr)) != strlen(pidstr)) {
+        g_critical("Failed to write pid file");
+        goto fail;
+    }
+
+    return true;
+
+fail:
+    unlink(pidfile);
+    return false;
+}
+#else /* _WIN32 */
+static bool ga_open_pidfile(const char *pidfile)
+{
+    return true;
+}
+#endif
+
 static gint ga_strcmp(gconstpointer str1, gconstpointer str2)
 {
     return strcmp(str1, str2);
@@ -277,6 +322,28 @@ static void ga_enable_non_blacklisted(GList *blacklist)
     g_free(list_head);
 }
 
+static bool ga_create_file(const char *path)
+{
+    int fd = open(path, O_CREAT | O_WRONLY, S_IWUSR | S_IRUSR);
+    if (fd == -1) {
+        g_warning("unable to open/create file %s: %s", path, strerror(errno));
+        return false;
+    }
+    close(fd);
+    return true;
+}
+
+static bool ga_delete_file(const char *path)
+{
+    int ret = unlink(path);
+    if (ret == -1) {
+        g_warning("unable to delete file: %s: %s", path, strerror(errno));
+        return false;
+    }
+
+    return true;
+}
+
 bool ga_is_frozen(GAState *s)
 {
     return s->frozen;
@@ -292,6 +359,10 @@ void ga_set_frozen(GAState *s)
     g_warning("disabling logging due to filesystem freeze");
     ga_disable_logging(s);
     s->frozen = true;
+    if (!ga_create_file(s->state_filepath_isfrozen)) {
+        g_warning("unable to create %s, fsfreeze may not function properly",
+                  s->state_filepath_isfrozen);
+    }
 }
 
 void ga_unset_frozen(GAState *s)
@@ -300,20 +371,38 @@ void ga_unset_frozen(GAState *s)
         return;
     }
 
+    /* if we delayed creation/opening of pid/log files due to being
+     * in a frozen state at start up, do it now
+     */
+    if (s->deferred_options.log_filepath) {
+        s->log_file = fopen(s->deferred_options.log_filepath, "a");
+        if (!s->log_file) {
+            s->log_file = stderr;
+        }
+        s->deferred_options.log_filepath = NULL;
+    }
     ga_enable_logging(s);
-    g_warning("logging re-enabled");
+    g_warning("logging re-enabled due to filesystem unfreeze");
+    if (s->deferred_options.pid_filepath) {
+        if (!ga_open_pidfile(s->deferred_options.pid_filepath)) {
+            g_warning("failed to create/open pid file");
+        }
+        s->deferred_options.pid_filepath = NULL;
+    }
 
     /* enable all disabled, non-blacklisted commands */
     ga_enable_non_blacklisted(s->blacklist);
     s->frozen = false;
+    if (!ga_delete_file(s->state_filepath_isfrozen)) {
+        g_warning("unable to delete %s, fsfreeze may not function properly",
+                  s->state_filepath_isfrozen);
+    }
 }
 
-#ifndef _WIN32
 static void become_daemon(const char *pidfile)
 {
+#ifndef _WIN32
     pid_t pid, sid;
-    int pidfd;
-    char *pidstr = NULL;
 
     pid = fork();
     if (pid < 0) {
@@ -323,20 +412,11 @@ static void become_daemon(const char *pidfile)
         exit(EXIT_SUCCESS);
     }
 
-    pidfd = open(pidfile, O_CREAT|O_WRONLY|O_EXCL, S_IRUSR|S_IWUSR);
-    if (pidfd == -1) {
-        g_critical("Cannot create pid file, %s", strerror(errno));
-        exit(EXIT_FAILURE);
-    }
-
-    if (asprintf(&pidstr, "%d", getpid()) == -1) {
-        g_critical("Cannot allocate memory");
-        goto fail;
-    }
-    if (write(pidfd, pidstr, strlen(pidstr)) != strlen(pidstr)) {
-        free(pidstr);
-        g_critical("Failed to write pid file");
-        goto fail;
+    if (pidfile) {
+        if (!ga_open_pidfile(pidfile)) {
+            g_critical("failed to create pidfile");
+            exit(EXIT_FAILURE);
+        }
     }
 
     umask(0);
@@ -351,15 +431,14 @@ static void become_daemon(const char *pidfile)
     close(STDIN_FILENO);
     close(STDOUT_FILENO);
     close(STDERR_FILENO);
-    free(pidstr);
     return;
 
 fail:
     unlink(pidfile);
     g_critical("failed to daemonize");
     exit(EXIT_FAILURE);
-}
 #endif
+}
 
 static int send_response(GAState *s, QObject *payload)
 {
@@ -597,9 +676,11 @@ VOID WINAPI service_main(DWORD argc, TCHAR *argv[])
 
 int main(int argc, char **argv)
 {
-    const char *sopt = "hVvdm:p:l:f:b:s:";
-    const char *method = NULL, *path = NULL, *pidfile = QGA_PIDFILE_DEFAULT;
-    const char *log_file_name = NULL;
+    const char *sopt = "hVvdm:p:l:f:b:s:t:";
+    const char *method = NULL, *path = NULL;
+    const char *log_filepath = NULL;
+    const char *pid_filepath = QGA_PIDFILE_DEFAULT;
+    const char *state_dir = QGA_STATEDIR_DEFAULT;
 #ifdef _WIN32
     const char *service = NULL;
 #endif
@@ -616,11 +697,11 @@ int main(int argc, char **argv)
 #ifdef _WIN32
         { "service", 1, NULL, 's' },
 #endif
+        { "statedir", 1, NULL, 't' },
         { NULL, 0, NULL, 0 }
     };
     int opt_ind = 0, ch, daemonize = 0, i, j, len;
     GLogLevelFlags log_level = G_LOG_LEVEL_ERROR | G_LOG_LEVEL_CRITICAL;
-    FILE *log_file = stderr;
     GList *blacklist = NULL;
     GAState *s;
 
@@ -635,17 +716,14 @@ int main(int argc, char **argv)
             path = optarg;
             break;
         case 'l':
-            log_file_name = optarg;
-            log_file = fopen(log_file_name, "a");
-            if (!log_file) {
-                g_critical("unable to open specified log file: %s",
-                           strerror(errno));
-                return EXIT_FAILURE;
-            }
+            log_filepath = optarg;
             break;
         case 'f':
-            pidfile = optarg;
+            pid_filepath = optarg;
             break;
+        case 't':
+             state_dir = optarg;
+             break;
         case 'v':
             /* enable all log levels */
             log_level = G_LOG_LEVEL_MASK;
@@ -684,7 +762,7 @@ int main(int argc, char **argv)
         case 's':
             service = optarg;
             if (strcmp(service, "install") == 0) {
-                return ga_install_service(path, log_file_name);
+                return ga_install_service(path, log_filepath);
             } else if (strcmp(service, "uninstall") == 0) {
                 return ga_uninstall_service();
             } else {
@@ -703,20 +781,70 @@ int main(int argc, char **argv)
         }
     }
 
-#ifndef _WIN32
-    if (daemonize) {
-        g_debug("starting daemon");
-        become_daemon(pidfile);
-    }
-#endif
-
     s = g_malloc0(sizeof(GAState));
-    s->log_file = log_file;
     s->log_level = log_level;
+    s->log_file = stderr;
     g_log_set_default_handler(ga_log, s);
     g_log_set_fatal_mask(NULL, G_LOG_LEVEL_ERROR);
-    s->logging_enabled = true;
+    ga_enable_logging(s);
+    s->state_filepath_isfrozen = g_strdup_printf("%s/qga.state.isfrozen",
+                                                 state_dir);
     s->frozen = false;
+#ifndef _WIN32
+    /* check if a previous instance of qemu-ga exited with filesystems' state
+     * marked as frozen. this could be a stale value (a non-qemu-ga process
+     * or reboot may have since unfrozen them), but better to require an
+     * uneeded unfreeze than to risk hanging on start-up
+     */
+    struct stat st;
+    if (stat(s->state_filepath_isfrozen, &st) == -1) {
+        /* it's okay if the file doesn't exist, but if we can't access for
+         * some other reason, such as permissions, there's a configuration
+         * that needs to be addressed. so just bail now before we get into
+         * more trouble later
+         */
+        if (errno != ENOENT) {
+            g_critical("unable to access state file at path %s: %s",
+                       s->state_filepath_isfrozen, strerror(errno));
+            return EXIT_FAILURE;
+        }
+    } else {
+        g_warning("previous instance appears to have exited with frozen"
+                  " filesystems. deferring logging/pidfile creation and"
+                  " disabling non-fsfreeze-safe commands until"
+                  " guest-fsfreeze-thaw is issued, or filesystems are"
+                  " manually unfrozen and the file %s is removed",
+                  s->state_filepath_isfrozen);
+        s->frozen = true;
+    }
+#endif
+
+    if (ga_is_frozen(s)) {
+        if (daemonize) {
+            /* delay opening/locking of pidfile till filesystem are unfrozen */
+            s->deferred_options.pid_filepath = pid_filepath;
+            become_daemon(NULL);
+        }
+        if (log_filepath) {
+            /* delay opening the log file till filesystems are unfrozen */
+            s->deferred_options.log_filepath = log_filepath;
+        }
+        ga_disable_logging(s);
+        ga_disable_non_whitelisted();
+    } else {
+        if (daemonize) {
+            become_daemon(pid_filepath);
+        }
+        if (log_filepath) {
+            s->log_file = fopen(log_filepath, "a");
+            if (!s->log_file) {
+                g_critical("unable to open specified log file: %s",
+                           strerror(errno));
+                goto out_bad;
+            }
+        }
+    }
+
     if (blacklist) {
         s->blacklist = blacklist;
         do {
@@ -758,13 +886,13 @@ int main(int argc, char **argv)
     ga_channel_free(ga_state->channel);
 
     if (daemonize) {
-        unlink(pidfile);
+        unlink(pid_filepath);
     }
     return 0;
 
 out_bad:
     if (daemonize) {
-        unlink(pidfile);
+        unlink(pid_filepath);
     }
     return EXIT_FAILURE;
 }