* [chttp2] Tune initial window size/max frame size setting
- Only choose frame sizes and initial window sizes that are powers of two - as a more predictable function to avoid flapping.
- If window size drops below 1kb then snap to 0 to avoid frantically changing a value that needn't be
- Allow the initial window size to drop to zero (we have fuzzing in place that says this is safe)
* copy/paste fix
* fix build
* fix
* Automated change: Fix sanity tests
* fix announce bug
* cleanup
* put this change under an experiment
* more tests
* fix build
* Automated change: Fix sanity tests
Co-authored-by: ctiller <ctiller@users.noreply.github.com>
We have many tests that create 100 threads or more, and mounting evidence that
this is harmful to our CI environment.
When the original code for many of these tests was written we ran our tests
under run_tests, which had explicit handling for tracking the number of threads
each test needed and making sure that we weren't over subscribing the test
runner. Bazel has no such facility (and the facility in run_tests has since
been removed) and so we need to adjust.
This PR adjusts down a single test and is part of a series so that we can
review and roll back easily if required.
* Disable end2end_binder_transport_test on some platforms
The following test case is flaky on windows
End2EndBinderTransportTestWithDifferentDelayTimes/End2EndBinderTransportTest.UnaryCallServerTimeout/1,
where GetParam() = 10ns
Binder transport won't be run on platform other than Android so it
should be OK to disable the test on some platform.
* Regenerate projects.
* Reland: "Make GetDefaultEventEngine return a shared_ptr (#30280)"
This reverts commit 45959e7cc1.
* Attempted fix with NoDestruct
* Not a process-wide singleton for the type. Just a NonDestruct
* fix
This works around valgrind memory leaks by giving EventEngines a fixed
lifetime. We eventually want ref-counted EventEngines internally, so this is
a step in the right direction as well.
* Revert "Revert "Reland: Add SRV and TXT record lookup methods to the iomgr PAI (#30242)"
This reverts commit b5966f39eb.
* release lock before unreffing
* Revert "Reland: Add SRV and TXT record lookup methods to the iomgr API (#30206)"
This reverts commit c229703f9f.
* Automated change: Fix sanity tests
Co-authored-by: drfloob <drfloob@users.noreply.github.com>
* Revert "Revert "Add SRV and TXT record lookup methods to the iomgr API (#30078)" (#30176)"
This reverts commit 2c3acbb2b2.
* one way to fix the ares handle race. Another option: work_serializer
* replace mu with parent's work serializer
* add lock annotations
* Revert "replace mu with parent's work serializer"
This reverts commit 0fce0ae150.
* statusor -> optional
* Automated change: Fix sanity tests
* add missing dep
Co-authored-by: drfloob <drfloob@users.noreply.github.com>
* Rename ResolveName to LookupHostname (same as EventEngine)
* Add stubs and no-op impls for LookupTXT and LookupSRV
* add native resolver tests that assert unimplemented
* extract custom name_server-setting logic and remove goto
* Separate SRV queries from grpc_dns_lookup_ares
* add necessary fixits before merging
* Automated change: Fix sanity tests
* fix missing ExecCtx on resolver tests
* separate out TXT lookup from hostname lookup (now all 3 are separate)
* rm DNS and update docs
* fix the fixer (forgot to add deps to BUILD)
* remove unused SRV and TXT args from ares hostname lookup method
* rename hostname-only ares dns lookup method
* refactor AresRequest using template method pattern
* Add name_server to Ares LookupHostname internals (needs iomgr API change)
* fix resolver test, callback should not be called on cancellation
* implement Ares-iomgr SRV and TXT lookup methods (verified manually)
Used a custom bind server with some redacted tests from
`resolve_address_test` to ensure both are working as expected.
* cleanup cruft
* unify common ares request setup logic between A, AAAA, TXT, and SRV
* generate_projects
* comment out unused args
* DNSResolver iomgr API uses Duration; hostname has all args now
* rm stale TODOs
* windows fix - bad variable name
* windows fix
* Automated change: Fix sanity tests
* reviewer feedback
* make protected members private
* move common properties to AresRequest base class
* localhost TXT results are empty, not an error
* reviewer feedback
* fix
Co-authored-by: drfloob <drfloob@users.noreply.github.com>
* BinderTransport: Fix wire reader/writer issue with NDK
NDK sometimes call registered `AIBinder_Class_onTransact`
callback while we call its `AIBinder_transact`.
Sometimes this behavior happens consistently sometimes it
don't happen at all. Likely related to the short circuit
behavior happens when both gRPC server and gRPC client are in the same
process.
This commit also fixes some other existing issue in wire writer.
* Use atomic<int64_t> instead of atomic_int64_t
atomic_int64_t is not defined in older compilers.
* Lower log level to avoid excessive test output
* Add missing break
dropepd this while addressing reivew comments
* Use absl::make_unique
* Fix printf portability with PRId64
* clang-format
* generate projects.
* Move unnecessary absl includes from metadata_batch.h
metadata_batch.h is a widely used header, and so any includes it takes
that are for its internal use can negatively impact our build times.
* missing file
* build
* review feedback
* Eliminate post-init in channel stack builder
We've had a post init function on channel stack builder for a very long
time, an it serves to run some code after initialization completes.
We need the functionality for a few things, but the function passed in
is intimately tied to the filter in use - we never vary it between
multiple functions for the same filter... which means it makes more
sense to locate this functionality as part of the filter interface.
* fix
* Automated change: Fix sanity tests
* fix
Co-authored-by: ctiller <ctiller@users.noreply.github.com>
When single closure is scheduled multiple times before it is run, it
runs less times than the number of times it is scheduled.
As a result, when server receives multiple RPC calls in a very short
time frame, `accept_stream_locked` is not correctly called for every
RPC call received.
We are working on internal server side stress test to make sure this
kind of error won't happen again.
This commit also uses atomic int for issuing new stream id. Without this
there is 10% probability of stream id collision when 128 parallel RPC
calls are initiated at the same time.
Craig Tiller
Vignesh Babu
Cheng-Yu Chung
github-actions[bot]
Ming-Chuan
Yash Tibrewal
AJ Heller
Esun Kim
apolcyn