The IOMMU's IOTLB invalidation is a costly process.  When iommu mode
is not set to "strict", it is done asynchronously. Current code
amortizes the cost of invalidating IOTLB entries by batching all the
invalidations in the system and performing a single global invalidation
instead. The code queues pending invalidations in a global queue that
is accessed under the global "async_umap_flush_lock" spinlock, which
can result is significant spinlock contention.

This patch splits this deferred queue into multiple per-cpu deferred
queues, and thus gets rid of the "async_umap_flush_lock" and its
contention.  To keep existing deferred invalidation behavior, it still
invalidates the pending invalidations of all CPUs whenever a CPU
reaches its watermark or a timeout occurs.
Signed-off-by: Omer Peleg <omer@cs.technion.ac.il>
[mad@cs.technion.ac.il: rebased, cleaned up and reworded the commit message]
Signed-off-by: Adam Morrison <mad@cs.technion.ac.il>
Reviewed-by: Shaohua Li <shli@fb.com>
Reviewed-by: Ben Serebrin <serebrin@google.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>