This is missing in v3 vs v2
- in v2 we had Pipe setup so that multiple Pipe stages could be chained and only complete when the last stage had passed flow control, whereas in v3 the top stage will start accepting requests as soon as the first stage in the pipeline takes the message.
Closes#37868
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37868 from ctiller:drizzling 69209da8a7
PiperOrigin-RevId: 686652402
We'll probably disable some next week :)
But I want to watch a good selection and refine criteria for acceptance.
Closes#37903
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37903 from ctiller:ALL-the-things 5f829db870
PiperOrigin-RevId: 685010911
So far missing for HTTP/2 style flow control has been a primitive to query whether there's a receiver for flow control data at the other end of the message pipes.
Here I'm updating the state machine accessors to accommodate that functionality.
No new states were needed.
Whilst here, document the current member functions on `CallState`.
Closes#37867
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37867 from ctiller:like-the-river c9814c737d
PiperOrigin-RevId: 684972125
Allows use of the party <-> party wakeup batching stuff, which reduces threadhops drastically for this transport.
Closes#37078
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37078 from ctiller:chaotic-party-3 75c32e6a64
PiperOrigin-RevId: 679685211
The following files have been moved:
- src/core/lib/avl/*
- src/core/lib/backoff/*
- src/core/lib/debug/event_log*
- src/core/lib/iomgr/gethostname*
- src/core/lib/iomgr/grpc_if_nametoindex*
- src/core/lib/matchers/*
- src/core/lib/uri/* (renamed from uri_parser.* to uri.*)
- src/core/lib/gprpp/* (existing src/core/util/time.cc was renamed to gpr_time.cc to avoid conflict)
Closes#36792
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36792 from markdroth:reorg_util d4e8996f48
PiperOrigin-RevId: 676947640
Previously, if we pulled server trailing metadata *before* the call was added to the client transport then we'd never call `on_done_` on the spine and consequently never remove the call from the map. This change fixes that edge case.
In fixing it, I noticed a state in `CallState` that was both complicating the fix and completely irrelevant because we respecced earlier this year to say that ServerTrailingMetadata processing cannot be asynchronous, so I'm removing that state also.
Closes#37749
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37749 from ctiller:flake-fightas-11 847814a286
PiperOrigin-RevId: 676246259
[Gpr_To_Absl_Logging] Remove gpr logging header include from other headers
Some of the cc files are using gpr_log_verbosity_init() functions.
So I removed the #include <grpc/support/log.h> from all headers and put it selectively only in the cc files that used gpr_log_verbosity_init()
Closes#37513
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37513 from tanvi-jagtap:remove_gpr_headers_from_headers_01 612ca6d0f7
PiperOrigin-RevId: 663811895
[Gpr_To_Absl_Logging] Remove logging header from example and test/core/t folder
Closes#37489
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37489 from tanvi-jagtap:remove_header_test_core_t fd5422b1f5
PiperOrigin-RevId: 663597008
Two new benchmarks here-in.
Benchmark 1: `bm_picker`
------
Measures various load balancing policies pick performance. For now we cover `pick_first` and `weighted_round_robin` at 1, 10, 100, 1000, 10000, and 100000 backends.
Today's output:
```
------------------------------------------------------------------------------
Benchmark Time CPU Iterations
------------------------------------------------------------------------------
BM_Pick/pick_first/1 20.4 ns 20.4 ns 68285
BM_Pick/pick_first/10 20.6 ns 20.6 ns 68274
BM_Pick/pick_first/100 20.5 ns 20.5 ns 67817
BM_Pick/pick_first/1000 20.6 ns 20.6 ns 67347
BM_Pick/pick_first/10000 20.7 ns 20.7 ns 67317
BM_Pick/pick_first/100000 20.9 ns 20.9 ns 67385
BM_Pick/weighted_round_robin/1 54.7 ns 54.7 ns 26641
BM_Pick/weighted_round_robin/10 54.2 ns 54.2 ns 25828
BM_Pick/weighted_round_robin/100 55.2 ns 55.2 ns 26210
BM_Pick/weighted_round_robin/1000 54.1 ns 54.1 ns 25678
BM_Pick/weighted_round_robin/10000 77.3 ns 76.6 ns 15776
BM_Pick/weighted_round_robin/100000 148 ns 148 ns 9882
```
Benchmark 2: `bm_load_balanced_call_destination`
-----
This benchmark measures call performance when a call spine passes through a `LoadBalancedCallDestination`, and with `BM_LoadBalancedCallDestination` also the construction/destruction cost of this object.
We do not consider picker performance in this benchmark as it's separately covered by `bm_picker` above.
Today's output:
```
-----------------------------------------------------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations
-----------------------------------------------------------------------------------------------------------------------------------------
BM_UnaryWithSpawnPerEnd<UnstartedCallDestinationFixture<LoadBalancedCallDestinationTraits>> 1255 ns 1255 ns 1076
BM_UnaryWithSpawnPerOp<UnstartedCallDestinationFixture<LoadBalancedCallDestinationTraits>> 1459 ns 1459 ns 939
BM_ClientToServerStreaming<UnstartedCallDestinationFixture<LoadBalancedCallDestinationTraits>> 209 ns 209 ns 6775
BM_LoadBalancedCallDestination 92.8 ns 92.8 ns 15063
```
Notes
------
There's some duplicated code between the benchmarks & tests -- this is ok -- as the tests evolve we'll likely want to add more checks to the fixtures, whereas as the benchmarks evolve we may well want to optimize the fixtures so that performance of the systems under test dominate more. That is, the duplicated code is expected to have different evolutionary tracks.
Closes#37052
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37052 from ctiller:moar-benchy 30c7072d87
PiperOrigin-RevId: 658181731
- add a benchmark for various metadata creation styles
- add factory functions for status + message - these are 3-10x faster than going via absl::Status
- add a `MakePooledForOverwrite` function to Arena, use it everywhere -- this naming matches `std::make_unique_for_overwrite` in C++20, and avoids some language mandated initialization in `Table` (underlying `MetadataMap<>`) - speeding creation of metadata handles by 30%
For `bm_call_spine` we see before:
```
BM_UnaryWithSpawnPerEnd<CallSpineFixture>_median 745 ns 745 ns
```
and after:
```
BM_UnaryWithSpawnPerEnd<CallSpineFixture>_median 699 ns 699 ns
```
Closes#37111
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37111 from ctiller:meta-magic-2 100464fc7e
PiperOrigin-RevId: 652900726
Instead of passing the transport byte counts back up through the filter
stack to be reported to the `CallTracer`, we now have the transport
pass the transport byte counts directly to the `CallTracer` itself.
This will eventually allow us to avoid unnecessarily storing these byte
counts in cases where no `CallTracer` actually cares about the data, which
will reduce per-call memory. (In the short term, it actually increases
memory usage, but we can separately do some work to avoid the memory
usage in the transport by removing the `grpc_transport_stream_stats`
struct from the legacy filter API.)
This is a prereq for supporting `CallTracer` in the new call v3 stack,
which does not include the transport byte counts as part of the
receieve-trailing-metadata hook, unlike the legacy filter stack.
This change is controlled by the `call_tracer_in_transport` experiment,
which is enabled by default.
As part of this experiment, we also fix a couple of related bugs:
- On the client side, the chttp2 transport was incorrectly adding
annotations to the parent `ClientCallTracer` instead of the
`CallAttemptTracer`.
- The OpenCensus `ServerCallTracer` was incorrectly swapping the values
of sent and received bytes.
PiperOrigin-RevId: 650728181
Move event engine context to be an arena based context, and have party host arena directly.
Now `Party` can directly inject all context necessary without needing to call out to the derived type; this makes `Party` instantiable by itself (no need to inherit from it - though we'll still want to for `CallSpine` to simplify some arena lifetime management), meaning we can start to consider eliminating the base class `Activity` and merging these types.
Closes#37015
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37015 from ctiller:chaotic-party 7f44b37b5d
PiperOrigin-RevId: 646963003
Allow `CallFilters` to have more than one `CallFilters::Stack` attached to it, and execute all of the stacks when executing.
This permits a trivial implementation of `PassThrough` for `Interceptor`, and allows elimination of an inserted `ForwardCall` if an interception chain terminates in a `CallDestination` preceded by one or more filters.
Built on #36984 which should be merged first.
Closes#36985
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36985 from ctiller:step-by-step 07b14c9d91
PiperOrigin-RevId: 646586483
Originally when I was implementing call-v3 I added async handling for server trailing metadata because it was symmetrical with all the other handling we had. Turns out that we really don't need this, and further it's probably harmful to our ability to reason about the stack - so I'm removing that capability.
Closes#36984
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36984 from ctiller:no-infallible-seq f3b3548685
PiperOrigin-RevId: 646146666
There's some tooling floating around that looks for `BM_` prefixed strings and was failing with these benchmarks.
Re-jiggle some macros around so that's no longer the case.
Closes#37024
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37024 from ctiller:bm- 9b2fc76807
PiperOrigin-RevId: 646137939
As we've learned what configuration is needed for our benchmarks the settings have been growing more and more bespoke for each binary. Try to consolidate that into some useful defaults.
Also ensure we always `linkstatic=1`. `cc_binary` defaults to this, so it's reasonable to assume that's the performance our customers see. It also deeply impacts performance for small microbenchmarks, and so enabling it gives us more apples:apples, and saves chasing things that don't matter.
Closes#37012
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/37012 from ctiller:benchmark-bin b84cab58c6
PiperOrigin-RevId: 645483183
Previously, metadata mutations were made by the picker directly, which meant that they would be applied even if the channel winds up discarding the pick due to the returned subchannel having been disconnected by the time the pick result is returned. This changes the API such that pickers return metadata mutations along with the pick result, so that the mutations won't get applied unless the pick result is actually used.
Closes#36968
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36968 from markdroth:lb_metadata_api 2765da6121
PiperOrigin-RevId: 645451869
In the client fuzzer, some valid fuzzing scenarios would close the transport (thus deleting the endpoint), while the fuzzer mechanics still attempted to read/write to that endpoint. There was an inherent ownership problem, where both the transport and the fuzzer logic expected to own the endpoint lifetime.
This PR ensures that the transport owns the endpoint, and the fuzzer logic owns an object that can write to some shared endpoint state. This shared object can outlive the endpoint.
Closes#36966
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36966 from drfloob:fuzzer/4908841560506368 a9ea2e795d
PiperOrigin-RevId: 645081665
Specifically:
- use `OrphanablePtr<>` for `grpc_endpoint`
- use `absl::AnyInvocable<>` instead of `grpc_closure`
- use `EventEngine::Run()` instead of `ExecCtx::Run()`
- use `SliceBuffer` instead of `grpc_slice_buffer`
- use `absl::Status` instead of `grpc_error_handle`
- use `absl::string_view` instead of `const char*` for handshaker names
Also pass acceptor via `HandshakerArgs` instead of as a separate parameter.
Also changed chttp2 and httpcli to use `OrphanablePtr<>` for the endpoint.
PiperOrigin-RevId: 644551906
Preparation for switching away from `grpc_channel_filter*` to identify channel filters.
Closes#36907
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36907 from ctiller:type-name e7ad4c67a2
PiperOrigin-RevId: 644483948
Ensure arena accounting is working, and add a test that a constant call size results in a constant call size estimate.
Closes#36933
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36933 from ctiller:arena-accounts 116c805633
PiperOrigin-RevId: 644102412
This change brings up the direct channel, and inproc promise based transports.
This work exposed a bug that was very difficult to fix with the current call_filters.cc implementation, so I've substantially revamped that - instead of having a pipe-like object per call element, we now have a big ol' combined state machine for the entire call. It's a touch more code, but substantially easier to reason about individual cases, so I much prefer this form (it's also a slight memory improvement: 12 bytes total to track call state, and 10 of those are wakeup bitmasks...).
Closes#36734
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36734 from ctiller:transport-refs-9 3e2a80b40d
PiperOrigin-RevId: 644034593
Refines the test in the fuzzer to only check input == output if the frame was re-encodable.
Also tweaks some hpack encoder stuff to not crash but simply report errors where appropriate.
Closes#36821
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36821 from ctiller:f-frame 1b08875c4e
PiperOrigin-RevId: 642311442
[Gpr_To_Absl_Logging] Move function to test header form log.h
This is not really needed in log.h
Closes#36860
COPYBARA_INTEGRATE_REVIEW=https://github.com/grpc/grpc/pull/36860 from tanvi-jagtap:move_function_to_test_header e6494bd06f
PiperOrigin-RevId: 642080756
All TraceFlags are now configured in `src/core/lib/debug/trace_flags.yaml`. The format is:
```
my_flag:
default: false # the default value; default=false
description: Some Description
debug_only: false # debug_only flags only work in debug builds; default=false
internal: false # internal flags will not show up in documentation; default=false
```
To regenerate the trace flag source code, run `tools/codegen/core/gen_trace_flags.py` (requires mako). This script is also run when sanity checking.
This PR also adds two new features:
### Glob-based flag configuration
Trace flag configuration now supports `?` (single wildcard character) and `*` (one or more wildcard characters). For example, using `GRPC_TRACE='event_engine*'` will enable all flags that match that glob. It expands to:
* event_engine
* event_engine_client_channel_resolver
* event_engine_dns
* event_engine_endpoint
* event_engine_endpoint_data
* event_engine_poller
### A cleaner trace-logging macro in abseil logging format
If your goal is only to add log statements when the `fault_injection_filter` trace flag is enabled, you can use the macro:
```
GRPC_TRACE_LOG(fault_injection, INFO) << "Filtered:" << 42;
```
When the trace flag is enabled, the the log will show something like this:
```
I0000 00:00:1715733657.430042 16 file.cc:174] Filtered:42
```
----
Note: just like with the gpr_log to abseil logging conversion, the pre-existing trace logging usages can be replaced with the new tracing macro across multiple PRs.
Closes#36576
PiperOrigin-RevId: 641295215
This gives grpc_endpoint the same destruction-is-shutdown semantic as
EventEngine::Endpoint, which will make the migration easier.
PiperOrigin-RevId: 639867616
Craig Tiller
AJ Heller
Mark D. Roth
Yash Tibrewal
Esun Kim
Tanvi Jagtap
Vignesh Babu