On 02.04.2015 16:33, Martin Kletzander wrote: > On Thu, Apr 02, 2015 at 03:07:11PM +0200, Michal Privoznik wrote: >> https://bugzilla.redhat.com/show_bug.cgi?id=1200149 >> >> Even though we have a mutex mechanism so that two clients don't spawn >> two daemons, it's not strong enough. It can happen that while one >> client is spawning the daemon, the other one fails to connect. >> Basically two possible errors can happen: >> >> error: Failed to connect socket to >> '/home/mprivozn/.cache/libvirt/libvirt-sock': Connection refused >> >> or: >> >> error: Failed to connect socket to >> '/home/mprivozn/.cache/libvirt/libvirt-sock': No such file or directory >> >> The problem in both cases is, the daemon is only starting up, while we >> are trying to connect (and fail). We should postpone the connecting >> phase until the daemon is started (by the other thread that is >> spawning it). In order to do that, create a file lock 'libvirt-lock' >> in the directory where session daemon would create its socket. So even >> when called from multiple processes, spawning a daemon will serialize >> on the file lock. So only the first to come will spawn the daemon. >> >> Signed-off-by: Michal Privoznik <mprivozn@xxxxxxxxxx> >> --- >> src/rpc/virnetsocket.c | 46 >> ++++++++++++++++++++++++++++++---------------- >> 1 file changed, 30 insertions(+), 16 deletions(-) >> >> diff --git a/src/rpc/virnetsocket.c b/src/rpc/virnetsocket.c >> index 6b019cc..8d8b6e0 100644 >> --- a/src/rpc/virnetsocket.c >> +++ b/src/rpc/virnetsocket.c >> @@ -545,8 +545,10 @@ int virNetSocketNewConnectUNIX(const char *path, >> { >> char *binname = NULL; >> char *pidpath = NULL; >> + char *lockpath = NULL; >> int fd, passfd = -1; >> int pidfd = -1; >> + int lockfd = -1; >> virSocketAddr localAddr; >> virSocketAddr remoteAddr; >> >> @@ -561,6 +563,22 @@ int virNetSocketNewConnectUNIX(const char *path, >> return -1; >> } >> >> + if (spawnDaemon) { >> + if (virPidFileConstructPath(false, NULL, "libvirt-lock", >> &lockpath) < 0) >> + goto error; >> + >> + if ((lockfd = open(lockpath, O_RDWR | O_CREAT, 0600)) < 0 || >> + virSetCloseExec(lockfd) < 0) { >> + virReportSystemError(errno, _("Unable to create lock >> '%s'"), lockpath); >> + goto error; >> + } >> + >> + if (virFileLock(lockfd, false, 0, 1, true) < 0) { >> + virReportSystemError(errno, _("Unable to lock '%s'"), >> lockpath); >> + goto error; >> + } >> + } >> + >> if ((fd = socket(PF_UNIX, SOCK_STREAM, 0)) < 0) { >> virReportSystemError(errno, "%s", _("Failed to create socket")); >> goto error; >> @@ -595,17 +613,8 @@ int virNetSocketNewConnectUNIX(const char *path, >> if (virPidFileConstructPath(false, NULL, binname, &pidpath) < 0) >> goto error; >> >> - pidfd = virPidFileAcquirePath(pidpath, false, getpid()); >> - if (pidfd < 0) { >> - /* >> - * This can happen in a very rare case of two clients >> spawning two >> - * daemons at the same time, and the error in the logs >> that gets >> - * reset here can be a clue to some future debugging. >> - */ >> - virResetLastError(); >> - spawnDaemon = false; >> - goto retry; >> - } >> + if ((pidfd = virPidFileAcquirePath(pidpath, false, getpid())) >> < 0) >> + goto error; >> > > What if someone starts the daemon by hand in the meantime? There > should be 'goto retry;', I guess. Okay. > >> if ((passfd = socket(PF_UNIX, SOCK_STREAM, 0)) < 0) { >> virReportSystemError(errno, "%s", _("Failed to create >> socket")); >> @@ -668,16 +677,17 @@ int virNetSocketNewConnectUNIX(const char *path, >> goto error; >> } >> >> - /* >> - * Do we need to eliminate the super-rare race here any >> more? It would >> - * need incorporating the following VIR_FORCE_CLOSE() into a >> - * virCommandHook inside a virNetSocketForkDaemon(). >> - */ >> VIR_FORCE_CLOSE(pidfd); >> if (virNetSocketForkDaemon(binary, passfd) < 0) >> goto error; > > There's still a problem here. If someone starts the daemon between > the VIR_FORCE_CLOSE() and us starting the daemon, > virNetSocketForkDaemon() won't fail as the process will start, but it > will die after a while, realizing that there is another daemon > running. And this thread will continue with socket connected nowhere; > the socket that was passed to the daemon being started was unlinked by > the daemon that was started manually. > > I know this is very, very rare case, but I thought the same about > starting multiple virshes (what the heck is the plural of 'virsh' > anyway) and look where that got us! > > I'm not against this approach, it's definitely an improvement for the > current situation, but it's not a complete solution. What are your > thoughts on that? I think the whole daemon spanning approach is fubar-ed. But let me see if we can rely here purely on the daemon acquiring the pidfile itself. It's strange that we allocate the PID file anyway. However, it's not achievable I'd push this as improvement and let rare cases for the future. Michal -- libvir-list mailing list libvir-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/libvir-list