Signed-off-by: John Howard jhoward@microsoft.com

This PR adds container stats for Windows. It supports both Windows Server and Hyper-V containers.

Stats exposed from the platform are fundamentally quite different to those exposed by Linux. Where possible (and sensible), I have re-utilised Linux fields in the REST structure returned from the daemon to the client. However, in other cases, noticeably storage, the platforms are significantly different enough to warrant a different structure.

The REST structure has been enhanced, the same as for docker build and other calls, so that the client is aware of the platform of the daemon. In that way, the client can display appropriate information (eg PIDs are not appropriate on Windows). As the client only knows the daemon platform once there is at least one ContainerStat returned from the daemon, rather than just printing the headers immediately (which could be for the wrong platform), it will indicate that it is waiting for statistics, and only print the right header once at least one ContainerStat has been retrieved.

Memory management is sufficiently different between Windows and Linux, and indeed different enough between Windows Server containers and Hyper-V containers, that the only field which makes sense for client display purposes is the private working set. This is the same as task manager in Windows using the default settings.

CPU usage calculation is also sufficiently different, but generally the same fields can be utilised. The main difference is that for each interval to calculate percentage, the previous read time and number of processors are required from the daemon. In that way it can calculate the available number of 100ns ticks on the host between time samples, and from that calculate percentage utilisation of each container.

The final note worth mentioning on this PR is that all stats in the platform are returned via HCS. This is by design of the Windows API surface. HCS will indirectly get the networking stats from HNS is that same call. Hence there is no call into libnetwork to obtain networking stats as is done on Linux.

The net result of the change is that the client does "the right thing" regardless of whether it's Linux to Linux, Windows to Linux, Linux to Windows or any other platform combination you choose.

Here's an example of running Windows to Windows docker stats where there are a variety of Windows Server and Hyper-V containers running.

And a similar example on Linux to Windows. First before any stats are available (just running docker stats when there are no running containers)

And with a bunch of containers running on the target Windows machine:

And the corresponding examples using the updated client pointing to a Linux daemon to demonstrate that the familiar column set for Linux daemons is still displayed and hasn't regressed

Finally an example of the full REST output to query the stats of a single (Windows Server) container against a Windows daemon.

{"read":"2016-08-15T21:42:53.3872243Z","preread":"2016-08-15T21:42:52.3881218Z","pids_stats":{},"blkio_stats":{"io_service_bytes_recursive":null,"io_serviced_recursive":null,"io_queue_recursive":null,"io_service_time_recursive":null,"io_wait_time_recursive":null,"io_merged_recursive":null,"io_time_recursive":null,"sectors_recursive":null},"num_procs":8,"storage_stats":{"read_count_normalized":9309,"read_size_bytes":69064192,"write_count_normalized":1203177,"write_size_bytes":9854512640},"cpu_stats":{"cpu_usage":{"total_usage":158125000,"usage_in_kernelmode":102968750,"usage_in_usermode":102968750},"throttling_data":{"periods":0,"throttled_periods":0,"throttled_time":0}},"precpu_stats":{"cpu_usage":{"total_usage":158125000,"usage_in_kernelmode":102968750,"usage_in_usermode":102968750},"throttling_data":{"periods":0,"throttled_periods":0,"throttled_time":0}},"memory_stats":{"commitbytes":70619136,"commitpeakbytes":114601984,"privateworkingset":52281344},"networks":{"faed4256-741a-4492-ba5b-98da170c647c":{"rx_bytes":10670115968,"rx_packets":7120183,"rx_dropped":0,"tx_bytes":205687305,"tx_packets":3703891,"tx_dropped":0}}}

~~
This PR is in multiple commits to break down the changes into reasonable sized chunks to ease code review and make the logical parts easy to understand. Otherwise this would be a difficult PR to understand.

1. API and matching client changes. One change is for the fields in the REST structure, the other  for returning the daemon platform (`OSType`).
2. Daemon libcontainerd changes to call into HCS to retrieve the stats from the platform
3. Daemon refactoring of stats_collector to be present on Windows, and to populate the stats in a REST call
4. Client display changes depending on platform
5. Integration test changes

~~
Squashed 9/16/2016 for merge