[yosh-matsuda]

Related to #1396. That issue is closed, but a comment of 2026-09-01 reports that Client::stop() still does not end a request on Windows with the latest version.

Problem

stop() ends a request in flight with shutdown() on its socket. On Linux and macOS this wakes the poll() that waits for the response. On Windows, shutdown() does not wake a WSAPoll() on the same socket. The request then waits until the server answers, the read timeout passes (300 s by default), or the OS drops the connection.

801bf2e fixed the same problem for SSEClient::stop(). This PR uses the same mechanism for ClientImpl::stop().

Change

• stop() also sets the atomic flag stop_requested_ when a request is in flight.
• send_() gives the stream of the request an interrupt check that reads the flag, with set_read_interrupt() from 801bf2e. A wait for data then ends within 100 ms on every platform.
• The flag is cleared when a request starts while no other request is in flight. A redirect or a digest retry that is sent inside a request keeps the flag.
• shutdown_socket() stays, so Linux and macOS still wake at once.
• select_read_interruptible() now rounds its last wait up to a whole millisecond. poll() takes whole milliseconds, and the old code dropped the rest, so the wait could end up to 1 ms before the deadline. The SSE path had the same error. When every client request waits through this function, MaxTimeoutTest.ContentStream and MaxTimeoutTest.ContentStreamSSL fail with "Timeout exceeded by -1ms" on the Mbed TLS, wolfSSL, and BoringSSL jobs.

stop() sets the flag under socket_mutex_, and send_() counts the request as in flight under the same lock. Thus a stop() that comes after that point, but before the request starts to wait, is not lost. A stop() before that point closes the idle socket, as before.

Notes:

• Each wait for data of a client request now wakes every 100 ms to read the flag. SSEClient uses the same slice.
• Only the wait for data is interrupted. A wait to write does not change.
• ClientImpl gets one new member, so abidiff reports an ABI change against v0.64.0.

Test

ClientStopTest.StopEndsWaitForResponse sends a request to a handler that does not answer. Another thread calls stop() once, and the test expects the request to end with Error::Read within 2 s.

• The test alone, without the fix: "windows without SSL" and "windows with SSL" fail.
• The fix without the rounding: the Windows jobs pass, but the Mbed TLS, wolfSSL, and BoringSSL jobs fail on MaxTimeoutTest.
• Linux, local, clang 22 with ASan: the test passes in approximately 0.1 s, 30 of 30 runs. With shutdown_socket() removed from stop() to imitate Windows, the test fails without the fix and passes in approximately 0.2 s with the fix.

[yhirose]

@yosh-matsuda thanks for the thorough investigation and fix!