- Generated by
1.12.0
|
mUPnP for C
|
Investigated PR #46 at 3d84d76f3bebc3017601ca29eac870ffe6fe7c6a. The existing StateVariable test starts a network device; the new StateVariableDefaultValueAndDataType test only parses SCPD XML. Neither the SCPD accessors nor either existing test is changed here.
A C probe using the actual test/TestDevice.c and library reproduced the startup stall in the current restricted Linux environment:
The original library made over 14 million HTTP-open attempts in three seconds; the external timeout returned 124. Linker wrapping recorded calls to the real mupnp_http_serverlist_open() without modifying its behavior. strace could not run because this environment prohibits ptrace.
The causal path is:
upnp_test_device_new() parses an in-memory device description and the root service's SCPD, then installs listeners. It does not start network services.mupnp_device_start() first stops the device, then populates dev->ifCache.mupnp_net_gethostinterfaces() returns an empty list when getifaddrs() fails. It also excludes down and loopback interfaces; an isolated host with no eligible addresses can reach the same result.mupnp_http_serverlist_open() enumerates interfaces separately. With no interfaces, it returns false without opening a listening socket.This establishes a reproducible cause of the observed type of stall. There is no syscall trace of the October 7 run, so its exact runtime cause cannot be proved retrospectively. A different host could still block elsewhere.
Extend the existing ESP empty-interface check to all platforms. If the cached interface list is empty, startup returns false before attempting HTTP ports. The normal startup and shutdown paths are otherwise unchanged.
test/unix/device-start-no-interfaces.sh injects an empty interface list using Linux linker wrapping. It uses the real TestDevice fixture and device lifecycle. An HTTP-open attempt fails the test immediately, so the old implementation produces a clear failure instead of hanging the test runner. The test checks failed startup, unchanged port, inactive advertiser, empty server lists, two successive attempts, stop, and delete. Automake and its source generator both include this test. Non-Linux hosts skip this GNU-linker-specific check.
Validated locally:
device.c: exit 1, "HTTP port retry reached
with no usable interfaces".device.c instrumented for coverage and CODE_COVERAGE_LIBS=--coverage.git diff --check passed.The current environment lacks Boost.Test, Autoconf/Automake, and clang-format; the package installation attempt could not locate those packages. The old Boost test executable cannot load libboost_unit_test_framework.so.1.83.0. Consequently, this investigation does not claim a fresh full Boost suite, full coverage run, or formatter run. The C build used the checked-in configure script, an explicit ./config.status config.h, and suppressed regeneration of Autotools outputs. It is not a substitute for the normal CI build below.
The original StateVariable test will report a startup assertion failure in an environment with no usable interfaces after this fix; that environment still cannot run network integration tests successfully. It will no longer spin in this particular retry loop. Persistent bind failures with a nonempty interface list, or interfaces disappearing after the cache check, remain separate cases; this change does not redesign the HTTP retry policy.
PR #46's build run 37639510483, job 112854401345, passed make check on Ubuntu 24.04.5. Its runner image was ubuntu24/20261004.327. The test executable ran for approximately 37 seconds after linking, and ran successfully again during coverage generation. The published job log does not record interface inventory, so a particular CI NIC configuration is not established.
The ESP-IDF success is additional, separate evidence: its host regressions and ESP32 cross-build do not run the ordinary Boost StateVariable case.
On a normal development host and on a disposable CI runner, collect:
Record actual interface enumeration failures as well as interface inventory. An ip or getifaddrs permission error is evidence, not evidence that the machine has no physical NIC. Do not remove host security restrictions to make the tests pass; use an ordinary permitted developer host or CI runner.
Build with CI's settings (requires its documented build dependencies):
Use GNU timeout (gtimeout on macOS). Record each exit code and preserve test/unix/test-suite.log and individual test logs even after failure. The Automake summary counts test programs/scripts, not Boost assertions. make -j4 versus CI's -j20 controls build/test-driver concurrency; it does not itself make individual Boost cases run in parallel. A 20-second whole-suite timeout is not sufficient evidence of deadlock when CI takes about 37 seconds.
Run just the deterministic regression from a configured build tree:
If the focused StateVariable case still stalls on a permitted host, use:
In GDB, run, interrupt once stalled, then collect thread apply all bt full. Break on mupnp_device_start, mupnp_http_serverlist_open, mupnp_device_advertiser_start, mupnp_device_advertiser_stop, mupnp_http_serverlist_stop, and mupnp_ssdp_serverlist_stop to distinguish startup, advertising, and shutdown phases. Remember that start itself calls stop before opening any services. At an HTTP-open breakpoint inspect port and the caller's dev->ifCache; repeated enumeration failures with rising ports identify this issue. Repeated bind failures with eligible interfaces identify a different retry problem. A blocked main thread in condition/join waiting requires all worker stacks (HTTP accept, SSDP receive, advertiser), not only the main-thread stack.
Where ptrace is unavailable, use flushed stderr markers before and after device creation, start, state update, stop, and delete in a temporary probe, and linker wrapping as used by the regression. Avoid committed timing sleeps or disabling the StateVariable test merely to make the suite appear green.
1.12.0