commit 7828169ef0ad9f2620ae2bfb9aeeb2d8535a6d9e tree f7815ecad59d4724887d6072c061ab24937dc8d2 parent 2130323875a8b0382ae6c67c8d8dcbad25c3fcca author Adhemerval Zanella 2026-09-10 11:19:38 -0300 committer Adhemerval Zanella 2026-09-10 11:19:38 -0300 round-robin.sh (probe_glibc_test_env): Record container capabilities The container the glibc testsuite runs in differs between the postcommit and precommit node pools, and nothing says so. On tcwg-g3-01 (tcwg-postcommit) every test that calls support_become_root() fails with error: xwrite.c:32: write of N bytes failed after 0: Operation not permitted -- the write to /proc/self/uid_map -- and every test-container test fails with "could not create a private mount namespace". On tcwg-g3-08 (tcwg-precommit) they pass. Both jobs pass the same --cap-add=SYS_ADMIN and --security-opt apparmor=unconfined, so the difference is in the host, not in ci-job-configs. That accounts for 100 of the 105 failures in the aarch64 baseline of tcwg_glibc_check and 115 on arm: all of resolv/, most of nss/, and every test-container test. They land in the baseline xfail list, where they hide any real regression in those tests. Probe for both capabilities before the glibc check and record the result in artifacts, so the next person to look at a 100-failure baseline can see why in one file. This only reports; it does not change the build result. :100644 100644 8fd33d87fab51fd22b73124a4f2e283b9c7f172b ef9ea4786eb4427bdb61feca520b07fd0b5c11de M round-robin.sh commit 2130323875a8b0382ae6c67c8d8dcbad25c3fcca tree 91c47dc75cf62609c2686418979a13f7f0840150 parent 56c86ab1c5179adf897872e3187ea72ca904fafd author Adhemerval Zanella 2026-09-10 11:19:38 -0300 committer Adhemerval Zanella 2026-09-10 11:19:38 -0300 round-robin.sh (check_regression): Don't call an XPASS a precommit regression An XPASS means the baseline and this build disagree about whether a test is expected to fail -- or whether it is built at all. When we track a branch that is worth reporting: the xfail entry has gone stale. In precommit the baseline comes from a different build, and since the postcommit and precommit pools were split it also comes from a different machine, so an XPASS describes the two environments rather than the patch under test. Builds 10630 and 10631 of tcwg_glibc_check--master-aarch64-precommit blamed a RISC-V patch for "XPASS: elf/tst-protected1a" and "XPASS: elf/tst-protected1b" on aarch64. Those tests do not appear in the baseline results at all, and their flaky entry had expired the day before. Pass --ignore_xpass to validate_failures.py in precommit mode. round-robin-notify.sh evaluates the same condition when it generates regressions.sum, so the report and this verdict cannot disagree; the two must be deployed together. Bisects are excluded: they have a meaningful baseline and want to see every difference. Requires the matching gcc-compare-results change adding --ignore_xpass. :100644 100644 fbf46c77ebb7b03e9929ae69ced61e7ab61f9583 8fd33d87fab51fd22b73124a4f2e283b9c7f172b M round-robin.sh commit 56c86ab1c5179adf897872e3187ea72ca904fafd tree f6299675e281a306bdb1bac1fd9f32cda8a9a558 parent c8070b9ee8e4298de94cebf724db927465a20a62 author Adhemerval Zanella 2026-09-09 13:52:07 -0300 committer Adhemerval Zanella 2026-09-09 13:52:07 -0300 round-robin.sh (build_abe): Scale glibc test timeouts glibc's test driver kills a test after TIMEOUT seconds, 20 by default (DEFAULT_TIMEOUT in support/test-driver.h), scaled by $TIMEOUTFACTOR. We run "make check" with -j $cpus, so on a 64-core machine the testsuite saturates every CPU and tests whose runtime depends on scheduling latency time out. posix/tst-affinity, tst-affinity-pid and tst-affinity-static are the usual victims: they pin the process to each CPU in turn and wait to be migrated, ~450 times over the various CPU set sizes. They were reported as regressions in 3 of the 9 failing tcwg_glibc_check--master-aarch64-precommit builds between 10563 and 10632, every time to a contributor whose patch had nothing to do with affinity. ABE already raises $TIMEOUTFACTOR by one on each re-run, so the base it scales is what we can usefully set from here. Its ramp alone does not help: a test needing more than twice the default fails on try #0 and try #1 alike, is classified as a stable fail rather than a flaky one, and is never re-run again. In build 10574 the affinity tests timed out at both 20s and 40s and were reported as regressions. Set TIMEOUTFACTOR_BASE=10 for the glibc check, giving 200s on the first try. It costs nothing for tests that pass; a test that really hangs now takes 200s rather than 20s to be killed. Requires the matching abe change adding $TIMEOUTFACTOR_BASE. :100644 100644 7bcab54a42d55dc8e484c888148546794faabc27 fbf46c77ebb7b03e9929ae69ced61e7ab61f9583 M round-robin.sh