Add QUIC_PERF_COUNTER_CONN_LOAD_REJECT counter.

[rzikm]

Description

This PR adds a counter which reports the number of connections which were automatically rejected on MsQuic layer because the the worker is already overloaded and is barely able to service current connections.

Testing

Tested by running .NET System.Net.Quic tests suite in parallel which exhibited this problem in the past.

Documentation

Documentation was added.

[nibanks]

I'm not opposed to this, but I'm curious why the Listener statistics isn't enough, since this is all in-proc for .NET.

[nibanks]

Looks like we're getting a compile error:

D:\a\msquic\msquic\src\core\library.c(268,1): warning C4310: cast truncates constant value [D:\a\msquic\msquic\src\core\core.kernel.vcxproj]

This is coming from:

    QuicTraceEvent(
        PerfCountersRundown,
        "[ lib] Perf counters Rundown, Counters=%!CID!",
        CASTED_CLOG_BYTEARRAY(sizeof(PerfCounterSamples), PerfCounterSamples));

Which uses:

#define CLOG_BYTEARRAY(Len, Data) (uint8_t)(Len), (uint8_t*)(Data)
#define CASTED_CLOG_BYTEARRAY(Len, Data) CLOG_BYTEARRAY((unsigned char)(Len), (const unsigned char*)(Data))

The issue is that you just pushed past 256 bytes. I need to look at what needs to be changed to address this.

[nibanks]

Ok, luckily the ETW event actually takes a 16-bit in, not 8-bit, so it's just a few macros we need to update I believe. Let's define

#define CLOG_BYTEARRAY16(Len, Data) (uint16_t)(Len), (uint8_t*)(Data)
#define CASTED_CLOG_BYTEARRAY16(Len, Data) CLOG_BYTEARRAY16((unsigned short)(Len), (const unsigned char*)(Data))

And use that in the two places you're getting errors.

Note - You will also need to define the version for all the other platforms. I think you can just copy/paste and make a similar update.

[rzikm]

I'm not opposed to this, but I'm curious why the Listener statistics isn't enough, since this is all in-proc for .NET.

I am working on introducing some global diagnostic counters which can be observed using external tools (e.g. dotnet counters) and these diagnostics are process-wide. It is difficult to aggregate statistics from instance objects. We would probably have to maintain a list of all listeners and periodically pull and aggreagate them. And then listener statistics don't distinguish between load-rejected and app-rejected so we would have to add more tracking on our layer.

That aside, global perf counter for load-rejected connections seemed to me a good one to have in MsQuic anyway so I figured it would be better overall to add it here.

[nibanks]

Looks like you need to update the CLOG files too now.

[codecov]

Codecov Report

Attention: 1 lines in your changes are missing coverage. Please review.

Comparison is base (86e08e5) 87.38% compared to head (992d553) 64.97%.
Report is 27 commits behind head on main.

❗ Current head 992d553 differs from pull request most recent head 884b6d5. Consider uploading reports for the commit 884b6d5 to get more accurate results

Files	Patch %	Lines
src/core/listener.c	0.00%	1 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #4120       +/-   ##
===========================================
- Coverage   87.38%   64.97%   -22.41%     
===========================================
  Files          56       56               
  Lines       16984    15187     -1797     
===========================================
- Hits        14841     9868     -4973     
- Misses       2143     5319     +3176     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.