On Thu, Feb 21, 2013 at 10:47:05AM +0100, Kevin Wolf wrote:
> On Mon, Feb 18, 2013 at 04:46:49PM +0100, Stefan Hajnoczi wrote:
> > On Wed, Feb 13, 2013 at 02:22:10PM +0100, Kevin Wolf wrote:
> > > @@ -707,6 +710,16 @@ int qcow2_alloc_cluster_link_l2(BlockDriverState
> > > *bs, QCowL2Meta *m)
> > >
On Mon, Feb 18, 2013 at 04:46:49PM +0100, Stefan Hajnoczi wrote:
> On Wed, Feb 13, 2013 at 02:22:10PM +0100, Kevin Wolf wrote:
> > @@ -707,6 +710,16 @@ int qcow2_alloc_cluster_link_l2(BlockDriverState *bs,
> > QCowL2Meta *m)
> > }
> >
> > /* Update L2 table. */
> > +qemu_co_mutex_u
On Wed, Feb 13, 2013 at 02:22:10PM +0100, Kevin Wolf wrote:
> @@ -707,6 +710,16 @@ int qcow2_alloc_cluster_link_l2(BlockDriverState *bs,
> QCowL2Meta *m)
> }
>
> /* Update L2 table. */
> +qemu_co_mutex_unlock(&s->lock);
> +qemu_co_rwlock_wrlock(&m->l2_writeback_lock);
> +ha
On Wed, Feb 13, 2013 at 02:22:10PM +0100, Kevin Wolf wrote:
> @@ -689,6 +691,7 @@ int qcow2_alloc_cluster_link_l2(BlockDriverState *bs,
> QCowL2Meta *m)
> int i, j = 0, l2_index, ret;
> uint64_t *old_cluster, *l2_table;
> uint64_t cluster_offset = m->alloc_offset;
> +bool has_wr
This is the first part of an optimisation to improve the performance of
sequential cluster allocations.
Typically, write requests aren't aligned to cluster boundaries, so
sequential allocation means that every other request has to wait for the
COW of the previous request to complete. We can do bet
