[JackieTien97]

Summary

On dev/1.3, the IT framework never detects an occupied port. EnvUtils.checkPortsAvailable() passed a shell pipeline (lsof -iTCP -sTCP:LISTEN,TIME_WAIT -P -n | awk '{print $9}' | grep -E ... on Unix, netstat -aon -p tcp | findStr ... on Windows) to Runtime.exec(String), which does not use a shell. lsof received |, awk, ... as file names and exited with 1, which the method treats as "available". searchAvailablePorts() therefore accepted every port range even when one of its ports was in use, and a node could then fail with BindException.

This PR ports the master implementation, which runs lsof / netstat without a shell and parses the output in Java:

• the EnvUtils part of Enhance it for cluster-restart tests #15794 (source commit 88abd3e3097f9f1c16c51945c0ab0cf74307263b): checkPortsAvailable() delegates to listPortOccupation(...);
• [IT] Read the address column of lsof when checking port occupation #19070 (source commit 2fd7d838cbab3bd8645681b8020c5597e340378d): read the lsof address from column 8 instead of the state column 9.

Target base: dev/1.3 at c0811e54d01c1f4480d10506930c44d06cfb12e6. After this change, the port-check block of EnvUtils (checkPortsAvailable through listPortOccupationUnix) is identical to master at 2fd7d838cb.

Compatibility adaptations

• Explicit java.io imports instead of java.io.*, because the Checkstyle AvoidStarImport rule is error level.
• The lsof state filter stays LISTEN,TIME_WAIT, as already on dev/1.3 ([To dev/1.3] Clean up dead RPC thread config and use node-specific selectorNum #17618) and on master (Clean up dead RPC thread config and use node-specific selectorNum #17551).
• getSearchAvailablePortCmd, getWindowsSearchPortCmd and getUnixSearchPortCmd become unused and are removed, as master did in Clean up dead RPC thread config and use node-specific selectorNum #17551.

Only the EnvUtils part of #15794 is included. Its cluster-restart handling (AbstractEnv, AbstractNodeWrapper) and server-side changes are not part of this backport; on dev/1.3 only checkPortsAvailable() uses the (port, pid) map returned by listPortOccupation(...).

Validation

Local validation with JDK 17.0.5 and Maven 3.9.9:

• mvn clean package -DskipTests -pl integration-test -am -P with-integration-tests: build success, Spotless clean, 0 Checkstyle violations.
• Before the change (macOS): ranges containing a port held by python3 -m http.server, or ports held by another local application, were all reported as available; lsof printed status error on |: No such file or directory and exited with 1.
• After the change (macOS, compiled EnvUtils): occupied ranges are reported with the owning pid and checkPortsAvailable() returns false (IPv4 loopback, wildcard, a listener in the same JVM, the last port of the range). Free ranges and a port just past the range are reported as free, and searchAvailablePorts() returns a free range.
• Linux (ubuntu:24.04 container, lsof 4.95.0 as on ubuntu-latest): LISTEN lines have 10 columns with the address at index 8 and the pid at index 1, also for a multi-threaded process. IPv4 loopback, IPv4 wildcard and IPv6 loopback listeners are detected; the old index 9 matches nothing.
• IoTDBSimpleQueryIT (34 tests, a fresh cluster per test) and IoTDBFlushQueryIT (7 tests) with with-integration-tests (SimpleIT): 0 failures/errors; the 3 skipped tests are @Ignored in source.
• The Windows path (netstat -aon -p tcp) could not be run locally. It is the same code master runs, and the Windows job of Cluster IT - 1C1D covers it on this PR.

🤖 Generated with Claude Code