| Message ID | 20171221155905.3793-3-berrange@redhat.com |
|---|---|
| State | New |
| Headers | show |
| Series | [v2,1/2] io: move fd_is_socket() into common sockets code | expand |
On 12/21/2017 09:59 AM, Daniel P. Berrange wrote: > When starting QEMU management apps will usually setup a monitor socket, and > then open it immediately after startup. If not using QEMU's own -daemonize > arg, this process can be troublesome to handle correctly. The mgmt app will > need to repeatedly call connect() until it succeeds, because it does not > know when QEMU has created the listener socket. If can't retry connect() > forever though, because an error might have caused QEMU to exit before it > even creates the monitor. > > The obvious way to fix this kind of problem is to just pass in a pre-opened > socket file descriptor for the QEMU monitor to listen on. The management app > can now immediately call connect() just once. If connect() fails it knows > that QEMU has exited with an error. > > Signed-off-by: Daniel P. Berrange <berrange@redhat.com> > --- > chardev/char-socket.c | 32 ++++++++-- > chardev/char.c | 3 + > tests/test-char.c | 173 +++++++++++++++++++++++++++++++++++++++++++++++++- > util/qemu-sockets.c | 42 +++++++++++- > 4 files changed, 238 insertions(+), 12 deletions(-) > > @@ -983,25 +984,36 @@ static void qemu_chr_parse_socket(QemuOpts *opts, ChardevBackend *backend, > const char *path = qemu_opt_get(opts, "path"); > const char *host = qemu_opt_get(opts, "host"); > const char *port = qemu_opt_get(opts, "port"); > + const char *fd = qemu_opt_get(opts, "fd"); > const char *tls_creds = qemu_opt_get(opts, "tls-creds"); > SocketAddressLegacy *addr; > ChardevSocket *sock; > > + if ((!!path + !!fd + !!host) != 1) { I see you liked my suggestion for a compact rendering; the resulting line has more symbols than it does alphanumerics ;) > +/* Syms in libqemustub.a are discarded at .o file granularity. > + * To replace monitor_get_fd() we must ensure everything in > + * stubs/monitor.c is defined, to make sure monitor.o is discarded > + * otherwise we get duplicate syms at link time. > + */ > +Monitor *cur_mon = NULL; Patchew is correct that you can omit the ' = NULL'. With that tweaked, Reviewed-by: Eric Blake <eblake@redhat.com>
"Daniel P. Berrange" <berrange@redhat.com> writes: > When starting QEMU management apps will usually setup a monitor socket, and > then open it immediately after startup. If not using QEMU's own -daemonize > arg, this process can be troublesome to handle correctly. The mgmt app will > need to repeatedly call connect() until it succeeds, because it does not > know when QEMU has created the listener socket. If can't retry connect() > forever though, because an error might have caused QEMU to exit before it > even creates the monitor. > > The obvious way to fix this kind of problem is to just pass in a pre-opened > socket file descriptor for the QEMU monitor to listen on. The management app > can now immediately call connect() just once. If connect() fails it knows > that QEMU has exited with an error. > > The SocketAddress(Legacy) structs allow for FD passing via the monitor, using > the 'getfd' command, but only when using QMP JSON syntax. The HMP & CLI syntax > for chardevs, however, has no way to initialize the SocketAddress(Legacy) 'fd' > variant. So this patch wires up the 'fd' parameter to refer to a refer to a > passed in file descriptor. It can refer either to a named FD passed via the > 'getfd' command, as already supported with QMP, or if numeric it refers to an > FD passed to QEMU at startup. This allows both HMP usage: Missing piece of the argument: getfd rejects names starting with a digit. Without that, we'd be looking at incompatible change. > > (pass any FD number with SCM_RIGHTS) > getfd myfd > chardev-add socket,fd=myfd > > And also CLI usage > > (leak FD 3 from parent by clearing O_CLOEXEC) Took me a few seconds to get this is prescription, not description. Drop the parenthesis? Or make it clearer, say (parent process must set up and pass FD 3 to QEMU)? > -chardev socket,fd=3,id=mon > -mon chardev=mon,mode=control Not 100% clear whether a monitor command can use numeric syntax. Peeking ahead at qemu-sockets.c: it can't. Please spell it out. > > Note that we do not wire this up in the legacy chardev syntax, so you cannot > use FD passing with '-qmp', you must use the modern '-mon' + '-chardev' pair > > An illustrative example of usage is: > > #!/usr/bin/perl > > use IO::Socket::UNIX; > use Fcntl; > > unlink "/tmp/qmp"; > my $srv = IO::Socket::UNIX->new( > Type => SOCK_STREAM(), > Local => "/tmp/qmp", > Listen => 1, > ); > > my $flags = fcntl $srv, F_GETFD, 0; > fcntl $srv, F_SETFD, $flags & ~FD_CLOEXEC; > > my $fd = $srv->fileno(); > > exec "qemu-system-x86_64", \ > "-chardev", "socket,fd=$fd,server,nowait,id=mon", \ > "-mon", "chardev=mon,mode=control"; > > Signed-off-by: Daniel P. Berrange <berrange@redhat.com> Despite the issues I noted: this is a nice commit message, in particular the example. However, please wrap your lines between column 70 and 75. Two reasons: 1. it should fit within 80 columns even with the four spaces git-log adds, and more importantly 2. humans tend to have trouble following long lines with our eyes (I sure do). Typographic manuals suggest to limit columns to roughly 60 characters for exactly that reason[*]. > --- > chardev/char-socket.c | 32 ++++++++-- > chardev/char.c | 3 + > tests/test-char.c | 173 +++++++++++++++++++++++++++++++++++++++++++++++++- > util/qemu-sockets.c | 42 +++++++++++- > 4 files changed, 238 insertions(+), 12 deletions(-) > > diff --git a/chardev/char-socket.c b/chardev/char-socket.c > index 6013972f72..e162db7542 100644 > --- a/chardev/char-socket.c > +++ b/chardev/char-socket.c > @@ -28,6 +28,7 @@ > #include "qemu/error-report.h" > #include "qapi/error.h" > #include "qapi/clone-visitor.h" > +#include "qemu/cutils.h" > > #include "chardev/char-io.h" > > @@ -983,25 +984,36 @@ static void qemu_chr_parse_socket(QemuOpts *opts, ChardevBackend *backend, > const char *path = qemu_opt_get(opts, "path"); > const char *host = qemu_opt_get(opts, "host"); > const char *port = qemu_opt_get(opts, "port"); > + const char *fd = qemu_opt_get(opts, "fd"); > const char *tls_creds = qemu_opt_get(opts, "tls-creds"); > SocketAddressLegacy *addr; > ChardevSocket *sock; > > + if ((!!path + !!fd + !!host) != 1) { > + error_setg(errp, > + "Exactly one of 'path', 'fd' or 'host' required"); > + return; > + } > + > backend->type = CHARDEV_BACKEND_KIND_SOCKET; > - if (!path) { > - if (!host) { > - error_setg(errp, "chardev: socket: no host given"); > + if (path) { > + if (tls_creds) { > + error_setg(errp, "TLS can only be used over TCP socket"); > return; > } > + } else if (host) { > if (!port) { > error_setg(errp, "chardev: socket: no port given"); > return; > } > - } else { > - if (tls_creds) { > - error_setg(errp, "TLS can only be used over TCP socket"); > + } else if (fd) { > + /* We don't know what host to validate against when in client mode */ > + if (tls_creds && !is_listen) { > + error_setg(errp, "TLS can not be used with pre-opened client FD"); > return; Worth mentioning this restriction in the commit message? > } > + } else { > + g_assert_not_reached(); > } > > sock = backend->u.socket.data = g_new0(ChardevSocket, 1); This is cleanup and feature squashed together. Separate the two please. > @@ -1027,7 +1039,7 @@ static void qemu_chr_parse_socket(QemuOpts *opts, ChardevBackend *backend, > addr->type = SOCKET_ADDRESS_LEGACY_KIND_UNIX; > q_unix = addr->u.q_unix.data = g_new0(UnixSocketAddress, 1); > q_unix->path = g_strdup(path); > - } else { > + } else if (host) { > addr->type = SOCKET_ADDRESS_LEGACY_KIND_INET; > addr->u.inet.data = g_new(InetSocketAddress, 1); > *addr->u.inet.data = (InetSocketAddress) { > @@ -1040,6 +1052,12 @@ static void qemu_chr_parse_socket(QemuOpts *opts, ChardevBackend *backend, > .has_ipv6 = qemu_opt_get(opts, "ipv6"), > .ipv6 = qemu_opt_get_bool(opts, "ipv6", 0), > }; > + } else if (fd) { > + addr->type = SOCKET_ADDRESS_LEGACY_KIND_FD; > + addr->u.fd.data = g_new(String, 1); > + addr->u.fd.data->str = g_strdup(fd); > + } else { > + g_assert_not_reached(); > } > sock->addr = addr; > } > diff --git a/chardev/char.c b/chardev/char.c > index 2ae4f465ec..9d674855ae 100644 > --- a/chardev/char.c > +++ b/chardev/char.c > @@ -798,6 +798,9 @@ QemuOptsList qemu_chardev_opts = { > },{ > .name = "port", > .type = QEMU_OPT_STRING, > + },{ > + .name = "fd", > + .type = QEMU_OPT_STRING, > },{ > .name = "localaddr", > .type = QEMU_OPT_STRING, > diff --git a/tests/test-char.c b/tests/test-char.c > index 7ac25ff73f..e72a20f0c2 100644 > --- a/tests/test-char.c > +++ b/tests/test-char.c > @@ -9,6 +9,7 @@ > #include "qapi/error.h" > #include "qom/qom-qobject.h" > #include "qmp-commands.h" > +#include "monitor/monitor.h" > > static bool quit; > > @@ -284,9 +285,8 @@ static int socket_can_read_hello(void *opaque) > return 10; > } > > -static void char_socket_test(void) > +static void char_socket_test_common(Chardev *chr) > { > - Chardev *chr = qemu_chr_new("server", "tcp:127.0.0.1:0,server,nowait"); > Chardev *chr_client; > QObject *addr; > QDict *qdict; > @@ -341,6 +341,169 @@ static void char_socket_test(void) > object_unparent(OBJECT(chr)); > } > > + > +static void char_socket_basic_test(void) > +{ > + Chardev *chr = qemu_chr_new("server", "tcp:127.0.0.1:0,server,nowait"); > + > + char_socket_test_common(chr); > +} > + > + > +static int char_socket_listener(void) > +{ > + SocketAddress *addr = g_new0(SocketAddress, 1); > + int srv; > + > + addr->type = SOCKET_ADDRESS_TYPE_INET; > + addr->u.inet.host = g_strdup("127.0.0.1"); > + addr->u.inet.port = g_strdup("0"); > + > + srv = socket_listen(addr, &error_abort); > + g_assert(srv >= 0); > + > + qapi_free_SocketAddress(addr); > + return srv; > +} > + > + > +/* If a monitor is not active (ie cur_mon == NULL), then > + * we should be able to use fd=<NUMBER> syntax > + */ > +static void char_socket_fdpass_cli_test(void) > +{ > + Chardev *chr; > + char *optstr; > + QemuOpts *opts; > + int fd; > + > + g_assert_null(cur_mon); > + > + fd = char_socket_listener(); > + > + optstr = g_strdup_printf("socket,id=cdev,fd=%d,server,nowait", fd); > + I was asking myself where the fdset with ID @fd is set up, and only then discovered that your monitor_get_fd() ignores @fdname, so no further setup is necessary. Might be worth a comment somewhere. > + opts = qemu_opts_parse_noisily(qemu_find_opts("chardev"), > + optstr, true); > + g_assert_nonnull(opts); Leaks @optstr? > + > + chr = qemu_chr_new_from_opts(opts, &error_abort); > + > + qemu_opts_del(opts); > + > + char_socket_test_common(chr); > +} > + > + > +static int mon_fd = -1; > + > +int monitor_get_fd(Monitor *mon, const char *fdname, Error **errp) > +{ > + if (mon_fd == -1) { > + error_setg(errp, "No fd named %s", fdname); > + return -1; > + } > + return mon_fd; > +} Maps all names to @mon_fd. I guess that's good enough for this test. > + > +/* Syms in libqemustub.a are discarded at .o file granularity. > + * To replace monitor_get_fd() we must ensure everything in > + * stubs/monitor.c is defined, to make sure monitor.o is discarded > + * otherwise we get duplicate syms at link time. > + */ > +Monitor *cur_mon = NULL; > +void monitor_init(Chardev *chr, int flags) {} > + > +/* If a monitor is active (ie cur_mon != NULL), then > + * we should be able to use fd=<NAME> syntax > + */ > +static void char_socket_fdpass_mon_test(void) > +{ > + Chardev *chr; > + const char *optstr; > + QemuOpts *opts; > + int fd; > + > + fd = char_socket_listener(); > + mon_fd = fd; > + cur_mon = g_malloc(1); /* Pretend we have a mon available */ Feels unnecessarily dirty. Suggest to define cur_mon like this: static Monitor dummy_mon; Monitor *cur_mon = &dummy_mon; /* Pretend we have a mon available */ Or in case cur_mon must remain null outside this function, set it like this: Monitor dummy_mon = {0}; cur_mon = &dummy_mon; /* Pretend we have a mon available */ More of the same below. > + > + optstr = "socket,id=cdev,fd=myfd,server,nowait"; > + Similar to above, no further setup needed for @myfd. > + opts = qemu_opts_parse_noisily(qemu_find_opts("chardev"), > + optstr, true); > + g_assert_nonnull(opts); > + > + chr = qemu_chr_new_from_opts(opts, &error_abort); > + > + qemu_opts_del(opts); > + > + char_socket_test_common(chr); > + mon_fd = -1; > + g_free(cur_mon); > + cur_mon = NULL; > +} > + > + > +/* If a monitor is active (ie cur_mon != NULL), then > + * we should not allow using fd=<NUMBER> syntax > + */ > +static void char_socket_fdpass_nocli_test(void) > +{ > + Chardev *chr; > + char *optstr; > + QemuOpts *opts; > + int fd; > + Error *local_err = NULL; > + > + fd = char_socket_listener(); > + cur_mon = g_malloc(1); /* Pretend we have a mon available */ > + > + optstr = g_strdup_printf("socket,id=cdev,fd=%d,server,nowait", fd); > + > + opts = qemu_opts_parse_noisily(qemu_find_opts("chardev"), > + optstr, true); > + g_assert_nonnull(opts); Leaks @optstr? > + > + chr = qemu_chr_new_from_opts(opts, &local_err); > + > + qemu_opts_del(opts); > + > + g_assert_nonnull(local_err); Please use error_free_or_abort() and ... > + g_assert_null(chr); > + error_free(local_err); ... drop the error_free(). > + g_free(cur_mon); > + cur_mon = NULL; > +} > + > + > +/* If a monitor is not active (ie cur_mon == NULL), then > + * we should not allow using fd=<NAME> syntax > + */ > +static void char_socket_fdpass_nomon_test(void) > +{ > + Chardev *chr; > + const char *optstr; > + QemuOpts *opts; > + Error *local_err = NULL; > + > + g_assert_null(cur_mon); > + optstr = "socket,id=cdev,fd=myfd,server,nowait"; > + > + opts = qemu_opts_parse_noisily(qemu_find_opts("chardev"), > + optstr, true); > + g_assert_nonnull(opts); > + > + chr = qemu_chr_new_from_opts(opts, &local_err); > + > + qemu_opts_del(opts); > + > + g_assert_nonnull(local_err); > + g_assert_null(chr); > + error_free(local_err); Likewise. > +} > + > + > #ifndef _WIN32 > static void char_pipe_test(void) > { > @@ -757,7 +920,11 @@ int main(int argc, char **argv) > #ifndef _WIN32 > g_test_add_func("/char/file-fifo", char_file_fifo_test); > #endif > - g_test_add_func("/char/socket", char_socket_test); > + g_test_add_func("/char/socket/basic", char_socket_basic_test); > + g_test_add_func("/char/socket/fdpass/cli", char_socket_fdpass_cli_test); > + g_test_add_func("/char/socket/fdpass/mon", char_socket_fdpass_mon_test); > + g_test_add_func("/char/socket/fdpass/nocli", char_socket_fdpass_nocli_test); > + g_test_add_func("/char/socket/fdpass/nomon", char_socket_fdpass_nomon_test); > g_test_add_func("/char/udp", char_udp_test); > #ifdef HAVE_CHARDEV_SERIAL > g_test_add_func("/char/serial", char_serial_test); > diff --git a/util/qemu-sockets.c b/util/qemu-sockets.c > index 1d23f0b742..9400f9a940 100644 > --- a/util/qemu-sockets.c > +++ b/util/qemu-sockets.c > @@ -1046,7 +1046,26 @@ int socket_connect(SocketAddress *addr, Error **errp) > break; > > case SOCKET_ADDRESS_TYPE_FD: > - fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); > + if (cur_mon) { > + fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); > + if (fd < 0) { > + return -1; > + } > + } else { > + unsigned long i; Naming a long @i is bad taste. Let's rename to @ul. > + if (qemu_strtoul(addr->u.fd.str, NULL, 10, &i) < 0) { > + error_setg_errno(errp, errno, > + "Unable to parse FD number %s", > + addr->u.fd.str); > + return -1; > + } > + fd = i; Truncates silently. Shouldn't you check for range? If the parent process screws up passing the file descriptor, fd can hijack some random internal file. I'd ask you to catch that if I had any idea how to do that easily. In monitor context, you can only use named fds, as before. Okay. Outside monitor context, you can now use numeric fds, and only numeric fds. Makes sense, because named fds are associated with a monitor. Note that before the patch, we crashed in monitor_get_fd() dereferencing cur_mon. > + } > + if (!fd_is_socket(fd)) { > + error_setg(errp, "Expected a socket FD %s", addr->u.fd.str); Would "File descriptor '%s' is not a socket" be clearer? > + close(fd); > + return -1; > + } > break; > > case SOCKET_ADDRESS_TYPE_VSOCK: > @@ -1073,7 +1092,26 @@ int socket_listen(SocketAddress *addr, Error **errp) > break; > > case SOCKET_ADDRESS_TYPE_FD: > - fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); > + if (cur_mon) { > + fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); > + if (fd < 0) { > + return -1; > + } > + } else { > + unsigned long i; > + if (qemu_strtoul(addr->u.fd.str, NULL, 10, &i) < 0) { > + error_setg_errno(errp, errno, > + "Unable to parse FD number %s", > + addr->u.fd.str); > + return -1; > + } > + fd = i; > + } > + if (!fd_is_socket(fd)) { > + error_setg(errp, "Expected a socket FD %s", addr->u.fd.str); > + close(fd); > + return -1; > + } Code duplication. Please factor it out. > break; > > case SOCKET_ADDRESS_TYPE_VSOCK: [*] https://en.wikipedia.org/wiki/Column_(typography)#Typographic_style
On Fri, Dec 22, 2017 at 11:06:12AM +0100, Markus Armbruster wrote: > "Daniel P. Berrange" <berrange@redhat.com> writes: > > + > > +/* Syms in libqemustub.a are discarded at .o file granularity. > > + * To replace monitor_get_fd() we must ensure everything in > > + * stubs/monitor.c is defined, to make sure monitor.o is discarded > > + * otherwise we get duplicate syms at link time. > > + */ > > +Monitor *cur_mon = NULL; > > +void monitor_init(Chardev *chr, int flags) {} > > + > > +/* If a monitor is active (ie cur_mon != NULL), then > > + * we should be able to use fd=<NAME> syntax > > + */ > > +static void char_socket_fdpass_mon_test(void) > > +{ > > + Chardev *chr; > > + const char *optstr; > > + QemuOpts *opts; > > + int fd; > > + > > + fd = char_socket_listener(); > > + mon_fd = fd; > > + cur_mon = g_malloc(1); /* Pretend we have a mon available */ > > Feels unnecessarily dirty. Suggest to define cur_mon like this: > > static Monitor dummy_mon; > Monitor *cur_mon = &dummy_mon; /* Pretend we have a mon available */ > > Or in case cur_mon must remain null outside this function, set it like > this: > > Monitor dummy_mon = {0}; > cur_mon = &dummy_mon; /* Pretend we have a mon available */ > > More of the same below. FYI, I didn't do that because 'struct Monitor' is defined inside monitor.c, not exposed in header files. I felt it would be worse to pollute the header file with what's supposed to be a private struct definition, just for sake of tests, particularly since we don't actually need any of the Monitor object contents. We could create a monitor-internal.h for the "struct Monitor" definition, if you feel strongly we should take this approach in the tests instead of my hack here ? > > diff --git a/util/qemu-sockets.c b/util/qemu-sockets.c > > index 1d23f0b742..9400f9a940 100644 > > --- a/util/qemu-sockets.c > > +++ b/util/qemu-sockets.c > > @@ -1046,7 +1046,26 @@ int socket_connect(SocketAddress *addr, Error **errp) > > break; > > > > case SOCKET_ADDRESS_TYPE_FD: > > - fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); > > + if (cur_mon) { > > + fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); > > + if (fd < 0) { > > + return -1; > > + } > > + } else { > > + unsigned long i; > > Naming a long @i is bad taste. Let's rename to @ul. > > > + if (qemu_strtoul(addr->u.fd.str, NULL, 10, &i) < 0) { > > + error_setg_errno(errp, errno, > > + "Unable to parse FD number %s", > > + addr->u.fd.str); > > + return -1; > > + } > > + fd = i; > > Truncates silently. Shouldn't you check for range? > > If the parent process screws up passing the file descriptor, fd can > hijack some random internal file. I'd ask you to catch that if I had > any idea how to do that easily. I guess it is just a matter of defining yet another qemu_strtoNN variant that takes an "int" parameter instead of "long", and does range checking. > Outside monitor context, you can now use numeric fds, and only numeric > fds. Makes sense, because named fds are associated with a monitor. > Note that before the patch, we crashed in monitor_get_fd() dereferencing > cur_mon. Yeah, that is fun, but I don't think there's any code path that could trigger it. Regards, Daniel
"Daniel P. Berrange" <berrange@redhat.com> writes: > On Fri, Dec 22, 2017 at 11:06:12AM +0100, Markus Armbruster wrote: >> "Daniel P. Berrange" <berrange@redhat.com> writes: > >> > + >> > +/* Syms in libqemustub.a are discarded at .o file granularity. >> > + * To replace monitor_get_fd() we must ensure everything in >> > + * stubs/monitor.c is defined, to make sure monitor.o is discarded >> > + * otherwise we get duplicate syms at link time. >> > + */ >> > +Monitor *cur_mon = NULL; >> > +void monitor_init(Chardev *chr, int flags) {} >> > + >> > +/* If a monitor is active (ie cur_mon != NULL), then >> > + * we should be able to use fd=<NAME> syntax >> > + */ >> > +static void char_socket_fdpass_mon_test(void) >> > +{ >> > + Chardev *chr; >> > + const char *optstr; >> > + QemuOpts *opts; >> > + int fd; >> > + >> > + fd = char_socket_listener(); >> > + mon_fd = fd; >> > + cur_mon = g_malloc(1); /* Pretend we have a mon available */ >> >> Feels unnecessarily dirty. Suggest to define cur_mon like this: >> >> static Monitor dummy_mon; >> Monitor *cur_mon = &dummy_mon; /* Pretend we have a mon available */ >> >> Or in case cur_mon must remain null outside this function, set it like >> this: >> >> Monitor dummy_mon = {0}; >> cur_mon = &dummy_mon; /* Pretend we have a mon available */ >> >> More of the same below. > > FYI, I didn't do that because 'struct Monitor' is defined inside > monitor.c, not exposed in header files. I felt it would be worse I see. > to pollute the header file with what's supposed to be a private > struct definition, just for sake of tests, particularly since > we don't actually need any of the Monitor object contents. > > We could create a monitor-internal.h for the "struct Monitor" > definition, if you feel strongly we should take this approach > in the tests instead of my hack here ? Might be more trouble than it's worth. I can tolerate the hack in a test. >> > diff --git a/util/qemu-sockets.c b/util/qemu-sockets.c >> > index 1d23f0b742..9400f9a940 100644 >> > --- a/util/qemu-sockets.c >> > +++ b/util/qemu-sockets.c >> > @@ -1046,7 +1046,26 @@ int socket_connect(SocketAddress *addr, Error **errp) >> > break; >> > >> > case SOCKET_ADDRESS_TYPE_FD: >> > - fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); >> > + if (cur_mon) { >> > + fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); >> > + if (fd < 0) { >> > + return -1; >> > + } >> > + } else { >> > + unsigned long i; >> >> Naming a long @i is bad taste. Let's rename to @ul. >> >> > + if (qemu_strtoul(addr->u.fd.str, NULL, 10, &i) < 0) { >> > + error_setg_errno(errp, errno, >> > + "Unable to parse FD number %s", >> > + addr->u.fd.str); >> > + return -1; >> > + } >> > + fd = i; >> >> Truncates silently. Shouldn't you check for range? >> >> If the parent process screws up passing the file descriptor, fd can >> hijack some random internal file. I'd ask you to catch that if I had >> any idea how to do that easily. > > I guess it is just a matter of defining yet another qemu_strtoNN variant > that takes an "int" parameter instead of "long", and does range checking. Should do here. Won't do for cases where the valid range doesn't match an integer type. I'm not asking you to address the general problem, though. >> Outside monitor context, you can now use numeric fds, and only numeric >> fds. Makes sense, because named fds are associated with a monitor. >> Note that before the patch, we crashed in monitor_get_fd() dereferencing >> cur_mon. > > Yeah, that is fun, but I don't think there's any code path that could > trigger it. One more thing: semantics of SocketAddress member @fd needs to be documented in the QAPI schema. I'd specify regardless of context, i.e. decimal number is the file descriptor, anything else is a file descriptor name. Only then say which context accepts which.
diff --git a/chardev/char-socket.c b/chardev/char-socket.c index 6013972f72..e162db7542 100644 --- a/chardev/char-socket.c +++ b/chardev/char-socket.c @@ -28,6 +28,7 @@ #include "qemu/error-report.h" #include "qapi/error.h" #include "qapi/clone-visitor.h" +#include "qemu/cutils.h" #include "chardev/char-io.h" @@ -983,25 +984,36 @@ static void qemu_chr_parse_socket(QemuOpts *opts, ChardevBackend *backend, const char *path = qemu_opt_get(opts, "path"); const char *host = qemu_opt_get(opts, "host"); const char *port = qemu_opt_get(opts, "port"); + const char *fd = qemu_opt_get(opts, "fd"); const char *tls_creds = qemu_opt_get(opts, "tls-creds"); SocketAddressLegacy *addr; ChardevSocket *sock; + if ((!!path + !!fd + !!host) != 1) { + error_setg(errp, + "Exactly one of 'path', 'fd' or 'host' required"); + return; + } + backend->type = CHARDEV_BACKEND_KIND_SOCKET; - if (!path) { - if (!host) { - error_setg(errp, "chardev: socket: no host given"); + if (path) { + if (tls_creds) { + error_setg(errp, "TLS can only be used over TCP socket"); return; } + } else if (host) { if (!port) { error_setg(errp, "chardev: socket: no port given"); return; } - } else { - if (tls_creds) { - error_setg(errp, "TLS can only be used over TCP socket"); + } else if (fd) { + /* We don't know what host to validate against when in client mode */ + if (tls_creds && !is_listen) { + error_setg(errp, "TLS can not be used with pre-opened client FD"); return; } + } else { + g_assert_not_reached(); } sock = backend->u.socket.data = g_new0(ChardevSocket, 1); @@ -1027,7 +1039,7 @@ static void qemu_chr_parse_socket(QemuOpts *opts, ChardevBackend *backend, addr->type = SOCKET_ADDRESS_LEGACY_KIND_UNIX; q_unix = addr->u.q_unix.data = g_new0(UnixSocketAddress, 1); q_unix->path = g_strdup(path); - } else { + } else if (host) { addr->type = SOCKET_ADDRESS_LEGACY_KIND_INET; addr->u.inet.data = g_new(InetSocketAddress, 1); *addr->u.inet.data = (InetSocketAddress) { @@ -1040,6 +1052,12 @@ static void qemu_chr_parse_socket(QemuOpts *opts, ChardevBackend *backend, .has_ipv6 = qemu_opt_get(opts, "ipv6"), .ipv6 = qemu_opt_get_bool(opts, "ipv6", 0), }; + } else if (fd) { + addr->type = SOCKET_ADDRESS_LEGACY_KIND_FD; + addr->u.fd.data = g_new(String, 1); + addr->u.fd.data->str = g_strdup(fd); + } else { + g_assert_not_reached(); } sock->addr = addr; } diff --git a/chardev/char.c b/chardev/char.c index 2ae4f465ec..9d674855ae 100644 --- a/chardev/char.c +++ b/chardev/char.c @@ -798,6 +798,9 @@ QemuOptsList qemu_chardev_opts = { },{ .name = "port", .type = QEMU_OPT_STRING, + },{ + .name = "fd", + .type = QEMU_OPT_STRING, },{ .name = "localaddr", .type = QEMU_OPT_STRING, diff --git a/tests/test-char.c b/tests/test-char.c index 7ac25ff73f..e72a20f0c2 100644 --- a/tests/test-char.c +++ b/tests/test-char.c @@ -9,6 +9,7 @@ #include "qapi/error.h" #include "qom/qom-qobject.h" #include "qmp-commands.h" +#include "monitor/monitor.h" static bool quit; @@ -284,9 +285,8 @@ static int socket_can_read_hello(void *opaque) return 10; } -static void char_socket_test(void) +static void char_socket_test_common(Chardev *chr) { - Chardev *chr = qemu_chr_new("server", "tcp:127.0.0.1:0,server,nowait"); Chardev *chr_client; QObject *addr; QDict *qdict; @@ -341,6 +341,169 @@ static void char_socket_test(void) object_unparent(OBJECT(chr)); } + +static void char_socket_basic_test(void) +{ + Chardev *chr = qemu_chr_new("server", "tcp:127.0.0.1:0,server,nowait"); + + char_socket_test_common(chr); +} + + +static int char_socket_listener(void) +{ + SocketAddress *addr = g_new0(SocketAddress, 1); + int srv; + + addr->type = SOCKET_ADDRESS_TYPE_INET; + addr->u.inet.host = g_strdup("127.0.0.1"); + addr->u.inet.port = g_strdup("0"); + + srv = socket_listen(addr, &error_abort); + g_assert(srv >= 0); + + qapi_free_SocketAddress(addr); + return srv; +} + + +/* If a monitor is not active (ie cur_mon == NULL), then + * we should be able to use fd=<NUMBER> syntax + */ +static void char_socket_fdpass_cli_test(void) +{ + Chardev *chr; + char *optstr; + QemuOpts *opts; + int fd; + + g_assert_null(cur_mon); + + fd = char_socket_listener(); + + optstr = g_strdup_printf("socket,id=cdev,fd=%d,server,nowait", fd); + + opts = qemu_opts_parse_noisily(qemu_find_opts("chardev"), + optstr, true); + g_assert_nonnull(opts); + + chr = qemu_chr_new_from_opts(opts, &error_abort); + + qemu_opts_del(opts); + + char_socket_test_common(chr); +} + + +static int mon_fd = -1; + +int monitor_get_fd(Monitor *mon, const char *fdname, Error **errp) +{ + if (mon_fd == -1) { + error_setg(errp, "No fd named %s", fdname); + return -1; + } + return mon_fd; +} + +/* Syms in libqemustub.a are discarded at .o file granularity. + * To replace monitor_get_fd() we must ensure everything in + * stubs/monitor.c is defined, to make sure monitor.o is discarded + * otherwise we get duplicate syms at link time. + */ +Monitor *cur_mon = NULL; +void monitor_init(Chardev *chr, int flags) {} + +/* If a monitor is active (ie cur_mon != NULL), then + * we should be able to use fd=<NAME> syntax + */ +static void char_socket_fdpass_mon_test(void) +{ + Chardev *chr; + const char *optstr; + QemuOpts *opts; + int fd; + + fd = char_socket_listener(); + mon_fd = fd; + cur_mon = g_malloc(1); /* Pretend we have a mon available */ + + optstr = "socket,id=cdev,fd=myfd,server,nowait"; + + opts = qemu_opts_parse_noisily(qemu_find_opts("chardev"), + optstr, true); + g_assert_nonnull(opts); + + chr = qemu_chr_new_from_opts(opts, &error_abort); + + qemu_opts_del(opts); + + char_socket_test_common(chr); + mon_fd = -1; + g_free(cur_mon); + cur_mon = NULL; +} + + +/* If a monitor is active (ie cur_mon != NULL), then + * we should not allow using fd=<NUMBER> syntax + */ +static void char_socket_fdpass_nocli_test(void) +{ + Chardev *chr; + char *optstr; + QemuOpts *opts; + int fd; + Error *local_err = NULL; + + fd = char_socket_listener(); + cur_mon = g_malloc(1); /* Pretend we have a mon available */ + + optstr = g_strdup_printf("socket,id=cdev,fd=%d,server,nowait", fd); + + opts = qemu_opts_parse_noisily(qemu_find_opts("chardev"), + optstr, true); + g_assert_nonnull(opts); + + chr = qemu_chr_new_from_opts(opts, &local_err); + + qemu_opts_del(opts); + + g_assert_nonnull(local_err); + g_assert_null(chr); + error_free(local_err); + g_free(cur_mon); + cur_mon = NULL; +} + + +/* If a monitor is not active (ie cur_mon == NULL), then + * we should not allow using fd=<NAME> syntax + */ +static void char_socket_fdpass_nomon_test(void) +{ + Chardev *chr; + const char *optstr; + QemuOpts *opts; + Error *local_err = NULL; + + g_assert_null(cur_mon); + optstr = "socket,id=cdev,fd=myfd,server,nowait"; + + opts = qemu_opts_parse_noisily(qemu_find_opts("chardev"), + optstr, true); + g_assert_nonnull(opts); + + chr = qemu_chr_new_from_opts(opts, &local_err); + + qemu_opts_del(opts); + + g_assert_nonnull(local_err); + g_assert_null(chr); + error_free(local_err); +} + + #ifndef _WIN32 static void char_pipe_test(void) { @@ -757,7 +920,11 @@ int main(int argc, char **argv) #ifndef _WIN32 g_test_add_func("/char/file-fifo", char_file_fifo_test); #endif - g_test_add_func("/char/socket", char_socket_test); + g_test_add_func("/char/socket/basic", char_socket_basic_test); + g_test_add_func("/char/socket/fdpass/cli", char_socket_fdpass_cli_test); + g_test_add_func("/char/socket/fdpass/mon", char_socket_fdpass_mon_test); + g_test_add_func("/char/socket/fdpass/nocli", char_socket_fdpass_nocli_test); + g_test_add_func("/char/socket/fdpass/nomon", char_socket_fdpass_nomon_test); g_test_add_func("/char/udp", char_udp_test); #ifdef HAVE_CHARDEV_SERIAL g_test_add_func("/char/serial", char_serial_test); diff --git a/util/qemu-sockets.c b/util/qemu-sockets.c index 1d23f0b742..9400f9a940 100644 --- a/util/qemu-sockets.c +++ b/util/qemu-sockets.c @@ -1046,7 +1046,26 @@ int socket_connect(SocketAddress *addr, Error **errp) break; case SOCKET_ADDRESS_TYPE_FD: - fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); + if (cur_mon) { + fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); + if (fd < 0) { + return -1; + } + } else { + unsigned long i; + if (qemu_strtoul(addr->u.fd.str, NULL, 10, &i) < 0) { + error_setg_errno(errp, errno, + "Unable to parse FD number %s", + addr->u.fd.str); + return -1; + } + fd = i; + } + if (!fd_is_socket(fd)) { + error_setg(errp, "Expected a socket FD %s", addr->u.fd.str); + close(fd); + return -1; + } break; case SOCKET_ADDRESS_TYPE_VSOCK: @@ -1073,7 +1092,26 @@ int socket_listen(SocketAddress *addr, Error **errp) break; case SOCKET_ADDRESS_TYPE_FD: - fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); + if (cur_mon) { + fd = monitor_get_fd(cur_mon, addr->u.fd.str, errp); + if (fd < 0) { + return -1; + } + } else { + unsigned long i; + if (qemu_strtoul(addr->u.fd.str, NULL, 10, &i) < 0) { + error_setg_errno(errp, errno, + "Unable to parse FD number %s", + addr->u.fd.str); + return -1; + } + fd = i; + } + if (!fd_is_socket(fd)) { + error_setg(errp, "Expected a socket FD %s", addr->u.fd.str); + close(fd); + return -1; + } break; case SOCKET_ADDRESS_TYPE_VSOCK:
When starting QEMU management apps will usually setup a monitor socket, and then open it immediately after startup. If not using QEMU's own -daemonize arg, this process can be troublesome to handle correctly. The mgmt app will need to repeatedly call connect() until it succeeds, because it does not know when QEMU has created the listener socket. If can't retry connect() forever though, because an error might have caused QEMU to exit before it even creates the monitor. The obvious way to fix this kind of problem is to just pass in a pre-opened socket file descriptor for the QEMU monitor to listen on. The management app can now immediately call connect() just once. If connect() fails it knows that QEMU has exited with an error. The SocketAddress(Legacy) structs allow for FD passing via the monitor, using the 'getfd' command, but only when using QMP JSON syntax. The HMP & CLI syntax for chardevs, however, has no way to initialize the SocketAddress(Legacy) 'fd' variant. So this patch wires up the 'fd' parameter to refer to a refer to a passed in file descriptor. It can refer either to a named FD passed via the 'getfd' command, as already supported with QMP, or if numeric it refers to an FD passed to QEMU at startup. This allows both HMP usage: (pass any FD number with SCM_RIGHTS) getfd myfd chardev-add socket,fd=myfd And also CLI usage (leak FD 3 from parent by clearing O_CLOEXEC) -chardev socket,fd=3,id=mon -mon chardev=mon,mode=control Note that we do not wire this up in the legacy chardev syntax, so you cannot use FD passing with '-qmp', you must use the modern '-mon' + '-chardev' pair An illustrative example of usage is: #!/usr/bin/perl use IO::Socket::UNIX; use Fcntl; unlink "/tmp/qmp"; my $srv = IO::Socket::UNIX->new( Type => SOCK_STREAM(), Local => "/tmp/qmp", Listen => 1, ); my $flags = fcntl $srv, F_GETFD, 0; fcntl $srv, F_SETFD, $flags & ~FD_CLOEXEC; my $fd = $srv->fileno(); exec "qemu-system-x86_64", \ "-chardev", "socket,fd=$fd,server,nowait,id=mon", \ "-mon", "chardev=mon,mode=control"; Signed-off-by: Daniel P. Berrange <berrange@redhat.com> --- chardev/char-socket.c | 32 ++++++++-- chardev/char.c | 3 + tests/test-char.c | 173 +++++++++++++++++++++++++++++++++++++++++++++++++- util/qemu-sockets.c | 42 +++++++++++- 4 files changed, 238 insertions(+), 12 deletions(-)