Support collecting gRPC core metrics

[overvenus]

Collect gRPC core metrics:

• Counters:
  • client_calls_created
  • server_calls_created
  • client_channels_created
  • client_subchannels_created
  • server_channels_created
  • insecure_connections_created
  • syscall_write
  • syscall_read
  • tcp_read_alloc_8k
  • tcp_read_alloc_64k
  • http2_settings_writes
  • http2_pings_sent
  • http2_writes_begun
  • http2_transport_stalls
  • http2_stream_stalls
  • cq_pluck_creates
  • cq_next_creates
  • cq_callback_creates
• Histograms
  • call_initial_size
  • tcp_write_size
  • tcp_write_iov_size
  • tcp_read_size
  • tcp_read_offer
  • tcp_read_offer_iov_size
  • http2_send_message_size
  • http2_metadata_size

Details: https://github.com/grpc/grpc/blob/v1.56.0/src/core/lib/debug/stats_data.yaml

This PR add a new cc file called grpc_wrap_stats.cc and generates bindings
into a new file called binding_stats.rs.
Collecting metrics needs to include headers under in grpc/src/.
I think it's better to be an optional feature and codes are in separated files.

[BusyJay]

If the stats are always collected in C/C++ code, then there seems to be no benefits to make it an optional feature.

[overvenus]

If the stats are always collected in C/C++ code, then there seems to be no benefits to make it an optional feature.

If so, GRPCIO_SYS_USE_PKG_CONFIG needs to be removed because stats need to include files under grpc/src.

[BusyJay]

If so, GRPCIO_SYS_USE_PKG_CONFIG needs to be removed because stats need to include files under grpc/src.

You can depend on GRPCIO_SYS_USE_PKG_CONFIG to decide whether implement the feature.

By the way, there should be also standard way to expose metrics like open census and

[overvenus]

You can depend on GRPCIO_SYS_USE_PKG_CONFIG to decide whether implement the feature.

Do you mean functions in stats become no-op when GRPCIO_SYS_USE_PKG_CONFIG=1? Could you elaborate?

By the way, there should be also standard way to expose metrics like open census and

open census and channelz are included in grpc/include, while stats is not. I'm not sure what's the standard way.

[BusyJay]

Do you mean functions in stats become no-op when GRPCIO_SYS_USE_PKG_CONFIG=1

Maybe you should use a feature name internals to indicate the APIs are available only when they are compiled from sources.

open census and channelz should be the standard ways as they are public APIs.

[BusyJay]

How about moving c changes, for example, exposing a header, to tikv/grpc instead? So that users can choose to patch grpc or just ignore unknown symbols.

[overvenus]

tikv/grpc#50

Not sure if this is what you mean, but please take a look.

[nolouch]

Can we merge it?

[overvenus]

I think it's ready. I am fine with the current implementation and the implementation suggested by jay in the previous comment #625 (comment).

[Connor1996]

LGTM

[BusyJay]

Does it help collecting metrics? grpc/grpc#35348 And you need to update your branch before getting it merged.