On Thu, Aug 31, 2017 at 11:50:23AM +0200, Maxime Coquelin wrote:
> Prior to this patch, iotlb cache's read/write lock was
> read-locked at every guest IOVA to app VA translation,
> i.e. at least once per packet with indirect off and twice
> with indirect on.
>
> The problem is that rte_rwlock_read_lock() makes use of atomic
> operation, which is costly.
>
> This patch introduces iotlb lock helpers, so that a full burst
> can be protected with taking the lock once, which reduces the
> number of atomic operations by up to 64 with indirect
> descriptors.
You were assuming there is no single miss during a burst. If a miss
happens, it requries 2 locks: one for _pending_miss and another one
for _pending_insert. From this point of view, it's actually more
expensive.
However, I won't call it's a bad assumption (for the case of virtio
PMD). And if you take this assumption, why not just deleting the
pending list and moving the lock outside the _iotlb_find function()
like what you did in this patch?
I don't really see the point of introducing the pending list.
--yliu
