On 11.09.2013 17:37, Doug Goldstein wrote:
On Wed, Sep 11, 2013 at 10:06 AM, Laine Stump <laine@laine.org> wrote:
This resolves: https://bugzilla.redhat.com/show_bug.cgi?id=983026
The netcf interface driver previously had no state driver associated with it - as a connection was opened, it would create a new netcf instance just for that connection, and close it when it was finished. the problem with this is that each connection to libvirt used up a netlink socket, and there is a per process maximum of ~1000 netlink sockets.
The solution is to create a state driver to go along with the netcf driver. The state driver will opens a netcf instance, then all connections share that same netcf instance, thus only a single netlink socket will be used no matter how many connections are mde to libvirtd.
This was rather simple to do - a new virObjectLockable class is created for the single driverState object, which is created in netcfStateInitialize and contains the single netcf handle; instead of creating a new object for each client connection, netcfInterfaceOpen now just increments the driverState object's reference count and puts a pointer to it into the connection's privateData. Similarly, netcfInterfaceClose() just un-refs the driverState object (as does netcfStateCleanup()), and virNetcfInterfaceDriverStateDispose() handles closing the netcf instance. Since all the functions already have locking around them, the static lock functions used by all functions just needed to be changed to call virObjectLock() and virObjectUnlock() instead of directly calling the virMutex* functions. --- Changes from V1:
* make driverState a static.
* switch to using a virObjectLockable for driverState, at Eric's suggestion.
* add a simple error message if ncf_init() fails.
Again, I've tried this with a small number of simultaneous connections (including virt-manager), but I don't have a ready-made stress test.
src/interface/interface_backend_netcf.c | 173 +++++++++++++++++++++++--------- 1 file changed, 125 insertions(+), 48 deletions(-)
diff --git a/src/interface/interface_backend_netcf.c b/src/interface/interface_backend_netcf.c index f47669e..627c225 100644 --- a/src/interface/interface_backend_netcf.c +++ b/src/interface/interface_backend_netcf.c @@ -41,19 +41,119 @@ /* Main driver state */ typedef struct { - virMutex lock; + virObjectLockable parent; struct netcf *netcf; } virNetcfDriverState, *virNetcfDriverStatePtr;
+static virClassPtr virNetcfDriverStateClass; +static void virNetcfDriverStateDispose(void *obj);
-static void interfaceDriverLock(virNetcfDriverStatePtr driver) +static int +virNetcfDriverStateOnceInit(void) +{ + if (!(virNetcfDriverStateClass = virClassNew(virClassForObjectLockable(), + "virNetcfDriverState", + sizeof(virNetcfDriverState), + virNetcfDriverStateDispose))) + return -1; + return 0;
Bad spacing?
+} + +VIR_ONCE_GLOBAL_INIT(virNetcfDriverState) + +static virNetcfDriverStatePtr driverState = NULL;
So my understanding of libvirt's object and locking code isn't too good so I hope someone else will review. But I don't see virNetcfDriverStateOnceInit() called anywhere and I don't see virNetcfDriverStateInitialize() defined anywhere? Are those suppose to be one in the same or is there some magical macro expansion going on somewhere that I just don't know about.
The VIR_ONCE_GLOBAL_INIT(virDummy) macro creates a function called virDummyInitialize() which when called for the 1st time calls virDummyOnceInit(). Any subsequent call to virDummyInitialize() results in immediate return. Michal