[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

[libvirt] PATCH: 0/2: Multi-thread support for libvirtd daemon



These two patches add support for making the libvirtd daemon multithreaded
in its dispatch of client requests. Any single client still has all its
requests serialized - since a virConnectPtr instance is only allowed to be
used from a single thread. Multiple clients though can now be processed
in parallel without blocking each other.

NB, these patches are not intended to be applied yet. They are *unsafe*
untill the stateful drivers (QEMU, LXC, OpenVZ, Test) get suitable 
mutex locking. I've nearly got those drivers done and will post for 
review in the near future.

The description that follows is also added to CVS as qemud/THREADING.txt
for reference:

 - the top level lock is on 'struct qemud_server'. This must be
   held before acquiring any other lock

 - Each 'struct qemud_client' object has a lock. The server lock
   must be held before acquiring it. Once the client lock is acquired
   the server lock can (optionally) be dropped.

 - The event loop has its own self-contained lock. You can ignore
   this as a caller of virEvent APIs.


The server lock is only needed / used once the daemon has entered
its main loop, which is the qemudRunLoop() . The initial thread
acquires the lock upon entering this method.

It immediatelty spawns 'n' worker threads, whose main loop is
the qemudWorker() method. The workers will immediately try to
acquire the server lock, and thus block since its held by the
initial thread.

When the initial thread enters the poll() call, it drops the
server lock. The worker locks now each wakeup, acquire the
server lock and go into a condition wait on the 'job' condition
variable. The workers are now all 'primed' for incoming RPC
calls.



A file descriptor event now occurrs, causing the initial thread
to exit poll(). It invokes the registered callback associated
with the file descriptors on which the event occurrs. The callbacks
are required to immediately acquire the server lock.

If the callback is dealing with a client event, it will then
acquire the client lock, and drop the server lock.

The callback will now handle the I/O event, reading or writing
a RPC message. Once a complete RPC message has been read the
client is marked as being in state QEMUD_MODE_WAIT_DISPATCH,
and the 'job' condition variable is signaled. The callback
now drops the client lock and goes back into the poll() loop
waiting for more I/O events.

Meanwhile one of the worker threads wakes up from its condition
variable sleep, holding the server lock. It now searches for a
client in state QEMUD_MODE_WAIT_DISPATCH. If it doesn't find
one, it goes back to sleep. If it does find one, then it calls
into the remoteDispatchClientRequest() method de-serialize the
incoming message into an XDR object and invoke the helper method
for the associated RPC call.

While the helper method is executing, no locks are held on either
the client or server, but the ref count on the 'struct qemud_client'
object is incremented to ensure its not deleted. The helper can
now safely invoke the neccessary libvirt API call.


Daniel
-- 
|: Red Hat, Engineering, London   -o-   http://people.redhat.com/berrange/ :|
|: http://libvirt.org  -o-  http://virt-manager.org  -o-  http://ovirt.org :|
|: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
|: GnuPG: 7D3B9505  -o-  F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]