[Shaisolaris]

UVBaseTransport.__cinit__ sets the initial write buffer limits to

self._high_water = FLOW_CONTROL_HIGH_WATER * 1024
self._low_water = FLOW_CONTROL_HIGH_WATER // 4

FLOW_CONTROL_HIGH_WATER is in KiB (64), so the low-water mark starts at 16 bytes instead of 16 KiB. vanilla asyncio starts at (16384, 65536), and uvloop itself returns (16384, 65536) after set_write_buffer_limits() is called with no arguments, because add_flowcontrol_defaults() multiplies by 1024 before dividing.

Reproducer:

import asyncio, uvloop

async def main():
    loop = asyncio.get_running_loop()
    srv = await loop.create_server(asyncio.Protocol, '127.0.0.1', 0)
    port = srv.sockets[0].getsockname()[1]
    tr, _ = await loop.create_connection(asyncio.Protocol, '127.0.0.1', port)
    print(tr.get_write_buffer_limits())

uvloop.run(main())

On main this prints (16, 65536) for TCP, and the same for streams and write pipes; asyncio prints (16384, 65536). In practice a protocol that gets pause_writing() is only resumed once the buffer is almost fully drained, rather than at 16 KiB as with asyncio.

The change adds the missing * 1024, and adds an assertion on the initial limits to test_flowcontrol_mixin_set_write_limits.

Tested on Linux x86_64, CPython 3.12:

• tests/test_tcp.py -k flowcontrol_mixin_set_write_limits: fails before the change ((16, 65536) != (16384, 65536)), passes after.
• tests/test_tcp.py tests/test_pipes.py tests/test_unix.py: 141 passed, 21 skipped after the change.

I did not run the full test suite, macOS, or other Python versions.