On 02/25/2013 07:15 AM, Ján Tomko wrote:
Use virURIParse in qemuMigrationPrepareDirect to allow parsing IPv6 addresses, which would cause an 'incorrect :port' error message before.
To be able to migrate over IPv6, QEMU needs to listen on [::] instead of 0.0.0.0. This patch adds a call to getaddrinfo and sets the listen address based on the result.
It also uses the same listen address for the NBD server.
This will break migration if a hostname does not resolve to the same address family on both sides.
Bug: https://bugzilla.redhat.com/show_bug.cgi?id=846013 ---
Diff to V1: * initialize uri_str * reuse STRSKIP("tcp:") result instead of doing strlen on it * print a warning instead of failing when the hostname can't be resolved
Diff to V2: * freeaddrinfo * separate the listen address to allow reuse in qemuMigrationStartNBDServer
src/qemu/qemu_migration.c | 77 ++++++++++++++++++++++++++++++++++++----------- 1 file changed, 59 insertions(+), 18 deletions(-)
diff --git a/src/qemu/qemu_migration.c b/src/qemu/qemu_migration.c index cae58fa..ff9b959 100644 --- a/src/qemu/qemu_migration.c +++ b/src/qemu/qemu_migration.c @@ -22,7 +22,10 @@
#include <config.h>
+#include <netdb.h> +#include <sys/socket.h> #include <sys/time.h> +#include <sys/types.h>
You rarely ever need <sys/types.h> - most types in that header are guaranteed to be provided by any other standard header that has an interface using that type; or put another way, POSIX says that inclusion of other headers may drag in <sys/types.h>, and we have already been relying on that reliably happening. The other two additions are important, though.
@@ -2292,9 +2296,14 @@ qemuMigrationPrepareDirect(virQEMUDriverPtr driver, static int port = 0; int this_port; char *hostname = NULL; + char listenAddr[8]; char migrateFrom [64];
As long as you are touching this, remove the space before [64]. Then again, 64 was a magic number; a much safer approach would have been char migrateFrom[sizeof("tcp:0.0.0.0:") + INT_BUFSIZE_BOUND(int)]; (using gnulib's "intprops.h"), or even getting rid of the stack-alloc temporary (see below for thoughts about using virBufferPtr instead)
+ if (getaddrinfo(hostname, NULL, NULL, &info)) { + VIR_WARN("unable to get address info for %s, defaulting to IPv4", + hostname);
Would it be nice to include the gai_strerror() reason for why getaddrinfo() failed?
+ } else { + ipv6 = info->ai_family == AF_INET6; + freeaddrinfo(info); + } + if (*uri_out) VIR_DEBUG("Generated uri_out=%s", *uri_out);
- /* QEMU will be started with -incoming tcp:0.0.0.0:port */ - snprintf(migrateFrom, sizeof(migrateFrom), "tcp:0.0.0.0:%d", this_port); + /* QEMU will be started with -incoming tcp:0.0.0.0:port + * or -incoming tcp:[::]:port for IPv6 */ + if (ipv6) + snprintf(listenAddr, sizeof(listenAddr), "[::]"); + else + snprintf(listenAddr, sizeof(listenAddr), "0.0.0.0");
Eww. Even though this is safe (because you sized listenAddr big enough), I'd rather see: const char *listenAddr; if (ipv6) listenAddr = "[::]"; else listenAddr = "0.0.0.0";
+ + snprintf(migrateFrom, sizeof(migrateFrom), + "tcp:%s:%d", listenAddr, this_port);
And while this snprintf was pre-existing and also happens to be large enough, I can't help but think it would be nicer to pass a virBufferPtr instead of a const char * to qemuMigrationPrepareAny, although that would mean more refactoring done in an independent patch. -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org