Placing coroutines on the global pool should be preferrable, because it
can help all threads. But if the global pool is full, we can still
try to save some allocations by stashing completed coroutines on the
local pool. This is quite cheap too, because it does not require
atomic operations, and provides a gain of 15% in the best case.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 1417518350-6167-8-git-send-email-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
This is not needed anymore. The new TLS-based algorithm is adaptive.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 1417518350-6167-7-git-send-email-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
This patch removes the mutex by using fancy lock-free manipulation of
the pool. Lock-free stacks and queues are not hard, but they can suffer
from the ABA problem so they are better avoided unless you have some
deferred reclamation scheme like RCU. Otherwise you have to stick
with adding to a list, and emptying it completely. This is what this
patch does, by coupling a lock-free global list of available coroutines
with per-CPU lists that are actually used on coroutine creation.
Whenever the destruction pool is big enough, the next thread that runs
out of coroutines will steal the whole destruction pool. This is positive
in two ways:
1) the allocation does not have to do any atomic operation in the fast
path, it's entirely using thread-local storage. Once every POOL_BATCH_SIZE
allocations it will do a single atomic_xchg. Release does an atomic_cmpxchg
loop, that hopefully doesn't cause any starvation, and an atomic_inc.
A later patch will also remove atomic operations from the release path,
and try to avoid the atomic_xchg altogether---succeeding in doing so if
all devices either use ioeventfd or are not submitting requests actively.
2) in theory this should be completely adaptive. The number of coroutines
around should be a little more than POOL_BATCH_SIZE * number of allocating
threads; so this also empties qemu_coroutine_adjust_pool_size. (The previous
pool size was POOL_BATCH_SIZE * number of block backends, so it was a bit
more generous. But if you actually have many high-iodepth disks, it's better
to put them in different iothreads, which will also use separate thread
pools and aio=native file descriptors).
This speeds up perf/cost (in tests/test-coroutine) by a factor of ~1.33.
No matter if we end with some kind of coroutine bypass scheme or not,
it cannot hurt to optimize hot code.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 1417518350-6167-6-git-send-email-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Allow coroutine users to adjust the pool size. For example, if the
guest has multiple emulated disk drives we should keep around more
coroutines.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
The 'gthread' coroutine backend was written before the freelist (aka
pool) existed in qemu-coroutine.c.
This means that every thread is expected to exit when its coroutine
terminates. It is not possible to reuse threads from a pool.
This patch automatically disables the pool when 'gthread' is used. This
allows the 'gthread' backend to work again (for example,
tests/test-coroutine completes successfully instead of hanging).
I considered implementing thread reuse but I don't want quirks like CPU
affinity differences due to coroutine threads being recycled. The
'gthread' backend is a reference backend and it's therefore okay to skip
the pool optimization.
Note this patch also makes it easy to toggle the pool for benchmarking
purposes:
./configure --with-coroutine-backend=ucontext \
--disable-coroutine-pool
Reported-by: Gabriel Kerneis <gabriel@kerneis.info>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Gabriel Kerneis <gabriel@kerneis.info>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
qemu_co_queue_next(&queue) arranges that the next queued coroutine is
run at a later point in time. This deferred restart is useful because
the caller may not want to transfer control yet.
This behavior was implemented using QEMUBH in the past, which meant that
CoQueue (and hence CoMutex and CoRwlock) had a dependency on the
AioContext event loop. This hidden dependency causes trouble when we
move to a world with multiple event loops - now qemu_co_queue_next()
needs to know which event loop to schedule the QEMUBH in.
After pondering how to stash AioContext I realized the best solution is
to not use AioContext at all. This patch implements the deferred
restart behavior purely in terms of coroutines and no longer uses
QEMUBH.
Here is how it works:
Each Coroutine has a wakeup queue that starts out empty. When
qemu_co_queue_next() is called, the next coroutine is added to our
wakeup queue. The wakeup queue is processed when we yield or terminate.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The coroutine freelist is a global pool of unused coroutines. It avoids
the setup/teardown overhead associated with the coroutine lifecycle.
Since the pool is global, we need to synchronize access so that
coroutines can be used outside the BQL.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The coroutine pool code is duplicated between the ucontext and
sigaltstack backends, and absent from the win32 backend. But the
code can be shared easily by moving it to qemu-coroutine.c.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Asynchronous code is becoming very complex. At the same time
synchronous code is growing because it is convenient to write.
Sometimes duplicate code paths are even added, one synchronous and the
other asynchronous. This patch introduces coroutines which allow code
that looks synchronous but is asynchronous under the covers.
A coroutine has its own stack and is therefore able to preserve state
across blocking operations, which traditionally require callback
functions and manual marshalling of parameters.
Creating and starting a coroutine is easy:
coroutine = qemu_coroutine_create(my_coroutine);
qemu_coroutine_enter(coroutine, my_data);
The coroutine then executes until it returns or yields:
void coroutine_fn my_coroutine(void *opaque) {
MyData *my_data = opaque;
/* do some work */
qemu_coroutine_yield();
/* do some more work */
}
Yielding switches control back to the caller of qemu_coroutine_enter().
This is typically used to switch back to the main thread's event loop
after issuing an asynchronous I/O request. The request callback will
then invoke qemu_coroutine_enter() once more to switch back to the
coroutine.
Note that if coroutines are used only from threads which hold the global
mutex they will never execute concurrently. This makes programming with
coroutines easier than with threads. Race conditions cannot occur since
only one coroutine may be active at any time. Other coroutines can only
run across yield.
This coroutines implementation is based on the gtk-vnc implementation
written by Anthony Liguori <anthony@codemonkey.ws> but it has been
significantly rewritten by Kevin Wolf <kwolf@redhat.com> to use
setjmp()/longjmp() instead of the more expensive swapcontext() and by
Paolo Bonzini <pbonzini@redhat.com> for Windows Fibers support.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Peter Lieven