On 11/28/2013 12:40 PM, Paolo Bonzini wrote:
Il 28/11/2013 11:16, Avi Kivity ha scritto:
The QRCU I linked would work great latency-wise (it has roughly the same
latency of an rwsem but readers are lock-free). However, the locked
operations in the read path would hurt because of cache misses, so it's
not good either.
I guess srcu would work. Do you know what's the typical write-side
latency there? (in terms of what it waits for, not nanoseconds).
If there's no concurrent reader, it's zero if I read the code right.
Otherwise it depends:
- if there are many callbacks, only 10 of them are processed per
millisecond. But unless there are concurrent synchronize_srcu calls
there should not be any callback at all. If all VCPUs were to furiously
change the MSIs, the latency could go up to #vcpu/10 milliseconds.
- if there are no callbacks, but there are readers, synchronize_srcu
busy-loops for some time checking if the readers complete. After a
while (20 us for synchronize_srcu, 120 us for
synchronize_srcu_expedited) it gives up and starts using a workqueue to
poll every millisecond. This should never happen unless
So, given the very restricted usage this SRCU would have, we probably
can expect synchronize_srcu_expedited to finish its job in the
busy-looping phase, and 120 us should be the expected maximum
latency---more likely to be an order of magnitude smaller, and in very
rare cases higher.
Looks like it's a good fit then.
