TCP_INQ is a socket option we added to Linux to report pending bytes
on the socket as a control message.
Using TCP_INQ we can accurately decide whether to continue read or not.
Add an urgent parameter, when we do not want to wait for EPOLLIN.
This commit improves the latency of 1 RPC unary (minimal benchmark)
significantly:
Before:
l_50: 61.3584984733
l_90: 94.8328711277
l_99: 126.211351174
l_999: 158.722406029
After:
l_50: 51.3546011488 (-16%)
l_90: 72.3420731581 (-23%)
l_99: 103.280218974 (-18%)
l_999: 130.905689996 (-17%)
grpc_handshake is renamed to GrpcHandshake, using C++ class definitions
instead of C-style vtable classes. Update callers to use new interfaces.
We use RefCountedPtr to simplify reference tracking.
This is to use `grpc_core::RefCount` to improve performnace.
This commit also replaces explicit C vtables, with C++ vtable
with its own compile time assertions and performance benefits.
It also makes use of `RefCountedPtr` wherever possible.
Added a flag-guarded feature that allows gRPC to load TLS/SSL
roots from the OS trust store. This is the Linux-specific
implementation of such feature.
Somehow some Dell servers we're trying to run a gRPC client on have an empty product name in the BIOS. When gRPC tries to creadte default credentials, it checks whether it's running on GCE by strcmp()ing the contents of /sys/class/dmi/id/product_name to some magic strings. When it reads that file, it gets only a newline; in trim() it skips over the newline in both directions, and since end < start it returns nullptr. This causes a segfault in the strcmp() call. Since a machine without a product name clearly isn't GCE, change it to return false instead.
Both the Windows and Linux tests use platform-specific macros to
determine whether they should execute, but they weren't including the
file where those macros were defined, so they were always running the
else case of that check, which meant they weren't testing anything.