Merge e759d2a^

3 files changed, 74 insertions, 13 deletions

diff --git a/doc/core/moving-to-c++.md b/doc/core/moving-to-c++.md
new file mode 100644
index 0000000000..4c745b38a9
--- /dev/null
+++ b/doc/core/moving-to-c++.md
@@ -0,0 +1,60 @@
+# Moving gRPC core to C++
+
+October 2017
+
+ctiller, markdroth, vjpai
+
+## Background and Goal
+
+gRPC core was originally written in C89 for several reasons
+(possibility of kernel integration, ease of wrapping, compiler
+support, etc). Over time, this was changed to C99 as all relevant
+compilers in active use came to support C99 effectively.
+[Now, gRPC core is C++](https://github.com/grpc/proposal/blob/master/L6-allow-c%2B%2B-in-grpc-core.md)
+(although the code is still idiomatically C code) with C linkage for
+public functions. Throughout all of these transitions, the public
+header files are committed to remain in C89.
+
+The goal now is to make the gRPC core implementation true idiomatic
+C++ compatible with
+[Google's C++ style guide](https://google.github.io/styleguide/cppguide.html).
+
+## Constraints
+
+- No use of standard library
+  - Standard library makes wrapping difficult/impossible and also reduces platform portability
+  - This takes precedence over using C++ style guide
+- But lambdas are ok
+- As are third-party libraries that meet our build requirements (such as many parts of abseil)
+- There will be some C++ features that don't work
+  - `new` and `delete`
+  - pure virtual functions are not allowed because the message that prints out "Pure Virtual Function called" is part of the standard library
+    - Make a `#define GRPC_ABSTRACT {GPR_ASSERT(false);}` instead of `= 0;`
+- The sanity for making sure that we don't depend on libstdc++ is that at least some tests should explicitly not include it
+  - Most tests can migrate to use gtest
+    - There are tremendous # of code paths that can now be exposed to unit tests because of the use of gtest and C++
+  - But at least some tests should not use gtest
+
+
+## Roadmap
+
+- What should be the phases of getting code converted to idiomatic C++
+  - Opportunistically do leaf code that other parts don't depend on
+  - Spend a little time deciding how to do non-leaf stuff that isn't central or polymorphic (e.g., timer, call combiner)
+  - For big central or polymorphic interfaces, actually do an API review (for things like transport, filter API, endpoint, closure, exec_ctx, ...) .
+    - Core internal changes don't need a gRFC, but core surface changes do
+    - But an API review should include at least a PR with the header change and tests to use it before it gets used more broadly
+  - iomgr polling for POSIX is a gray area whether it's a leaf or central
+- What is the schedule?
+  - In Q4 2017, if some stuff happens opportunistically, great; otherwise ¯\\\_(ツ)\_/¯
+  - More updates as team time becomes available and committed to this project
+
+## Implications for C++ API and wrapped languages
+
+- For C++ structs, switch to `using` when possible (e.g., Slice,
+ByteBuffer, ...)
+- The C++ API implementation might directly start using
+`grpc_transport_stream_op_batch` rather than the core surface `grpc_op`.
+- Can we get wrapped languages to a point where we can statically link C++? This will take a year in probability but that would allow the use of `std::`
+  - Are there other environments that don't support std library, like maybe Android NDK?
+    - Probably, that might push things out to 18 months
diff --git a/doc/environment_variables.md b/doc/environment_variables.md
index f775de1664..40af758f69 100644
--- a/doc/environment_variables.md
+++ b/doc/environment_variables.md
@@ -49,6 +49,7 @@ some configuration as environment variables that can be set.
   - connectivity_state - traces connectivity state changes to channels
   - channel_stack_builder - traces information about channel stacks being built
   - executor - traces grpc's internal thread pool ('the executor')
+  - glb - traces the grpclb load balancer
   - http - traces state in the http2 transport engine
   - http2_stream_state - traces all http2 stream state mutations.
   - http1 - traces HTTP/1.x operations performed by gRPC
@@ -56,11 +57,12 @@ some configuration as environment variables that can be set.
   - flowctl - traces http2 flow control
   - op_failure - traces error information when failure is pushed onto a
     completion queue
-  - round_robin - traces the round_robin load balancing policy
   - pick_first - traces the pick first load balancing policy
   - plugin_credentials - traces plugin credentials
+  - pollable_refcount - traces reference counting of 'pollable' objects (only 
+    in DEBUG)
   - resource_quota - trace resource quota objects internals
-  - glb - traces the grpclb load balancer
+  - round_robin - traces the round_robin load balancing policy
   - queue_pluck
   - queue_timeout
   - server_channel - lightweight trace of significant server channel events
@@ -118,10 +120,10 @@ some configuration as environment variables that can be set.
     perform name resolution
   - ares - a DNS resolver based around the c-ares library
 
-* GRPC_DISABLE_CHANNEL_CONNECTIVITY_WATCHER
-  The channel connectivity watcher uses one extra thread to check the channel
-  state every 500 ms on the client side. It can help reconnect disconnected
-  client channels (mostly due to idleness), so that the next RPC on this channel
-  won't fail. Set to 1 to turn off this watcher and save a thread. Please note
-  this is a temporary work-around, it will be removed in the future once we have
-  support for automatically reestablishing failed connections.
+* GRPC_CLIENT_CHANNEL_BACKUP_POLL_INTERVAL_MS
+  Default: 5000
+  Declares the interval between two backup polls on client channels. These polls
+  are run in the timer thread so that gRPC can process connection failures while
+  there is no active polling thread. They help reconnect disconnected client
+  channels (mostly due to idleness), so that the next RPC on this channel won't
+  fail. Set to 0 to turn off the backup polls.
diff --git a/doc/load-balancing.md b/doc/load-balancing.md
index 88ff35496f..8ff94075b5 100644
--- a/doc/load-balancing.md
+++ b/doc/load-balancing.md
@@ -129,10 +129,9 @@ works:
         by the resolver. It asks the balancer for the server addresses to
         use for the server name originally requested by the client (i.e.,
         the same one originally passed to the name resolver).
-        - Note: The `grpclb` policy currently ignores any non-balancer
-          addresses returned by the resolver. However, in the future, it
-          may be changed to use these addresses as a fallback in case no
-          balancers can be contacted.
+        - Note: In the `grpclb` policy, the non-balancer addresses returned
+          by the resolver are used as a fallback in case no balancers can be
+          contacted when the LB policy is started.
      2. The gRPC servers to which the load balancer is directing the client
         may report load to the load balancers, if that information is needed
         by the load balancer's configuration.