LTP TIMEOUT-class audit

2026-11-07 · P0-class rule: a TIMEOUT is a "cannot-run" verdict, prioritized over optimization work · source doc docs/test/ltp-timeout-audit.md
Scope: every 13 TIMEOUT + 1 WEDGE row of the r27 census (1,869-case list, main@e23d3c3f, x86_64+KVM). Method: each row re-run one-test-per-boot on current main (scripts/test/ltpscan/scan-x86_64.py --chunk 1; the in-guest runner gives each test a 30 s budget, then pid-1 SIGTERMs it and records rc=143), with the LTP 20240524 sources open beside the serials.

§1Result: where the 14 rows belong

ClassRowsNotes
Kernel defect (fixed)1ptrace06's lost exec-trap stop + errno mapping for bad PEEK/POKE/REGS/SIGINFO parameters (f558181b)
Test/list defect119 by-design resident helpers + shmat1 (the 64-bit signal-recovery block is excluded by #ifdef __i386__) + mmap1 (274,152 verify iterations pass inside the budget, then the LTP-internal timeout TBROKs it)
Known lost-wake family1openfile — owned by the lost-wake producer hunt (produced on the TraceRing::push side)
Slow, not hung1ht_affinity — zero TFAIL, roughly one affinity mask per 30 s budget; needs a timing dig

After the fix, the ptrace family 01/02/03/05/06/11 all PASS one-per-boot.

§2Row-by-row disposition

CaseClass on mainVerdictEvidence
ptrace06TIMEOUTfixed → PASSthe tracer blocked in its first wait4: a TRACEME'd vfork+exec child never took the exec SIGTRAP stop. The fix queues the trap at the exec commit point; do_signal's traced arm converts it to the tracer-reported stop. Plus EIO errno mapping for bad parameters
shmat1CRASH 5/5test defect on x86_64the SIGSEGV recovery block is #ifdef __i386__; the 64-bit body is fprintf("caught signal %d -- exiting"); exit(-1). The test's designed shm-detach race fires within ~200 iterations on our timing; all 5 crashes fault at exactly map_address (kernel USEGV-FOR forensics). The kernel side is verified correct by probes: thread-directed fault delivery, SA_SIGINFO handler replacement and siglongjmp recovery all work. The test also runs alarm(exec_time*3600) = 24 h by design
openfileTIMEOUTknown lost-wake familymain parks in futex (join) after "Creating Reading Threads"; the reader threads are gone. STRAGGLER-RESCUE fired in-boot and recovered one lost mid-exit thread, but the join's wake was already lost. Owned by the lost-wake producer hunt
ht_affinityTIMEOUTslow progress, not a wedgezero TFAIL; ~one affinity mask per 30 s budget (loop = cpu_count masks × child spawn + sleep(1) sync, twice)
mmap1FAIL(int)slow by design274,152 verify iterations PASS inside the 30 s budget, then the LTP-internal timeout TBROKs it (the r26 census already documents this budget/racy class)
cgroup_fj_procTIMEOUTby-design helperdo {...} while (test_switch) — runs until its parent suite signals it
cgroup_regression_fork_processesTIMEOUTby-design helperwhile (1) fork loop
cpuctl_fj_cpu-hogTIMEOUTby-design helpersigsuspend waiting for the parent's SIGUSR1
fork_exec_loopTIMEOUTby-design helperwhile (1) fork+exec loop (the freezer suite's load generator)
memcg_test_2TIMEOUTby-design helperwhile (1)
memcg_test_4TIMEOUTby-design helperwhile (1) sleep(1)
pids_task2TIMEOUTby-design helperpause()
pthservTIMEOUTby-design helperfor (;;) accept loop (a server until killed)
timed_forkbombTIMEOUTby-design helperreads its bomb program from stdin (the parent pipes it); standalone it blocks on scanf
mmap1 (WEDGE row)—as abovethe r27 WEDGE row is the same test under chunked load

§3Follow-ups

  1. Census hygiene: the 9 by-design resident helpers should carry a BY-DESIGN marker in future censuses (runner-side list or verdict post-processing) so they stop reading as kernel TIMEOUT regressions.
  2. The r28 census: will pick up the ptrace06 PASS and these reclassifications.
  3. ht_affinity: a per-step timing dig (affinity syscall + signal handshake cost).
  4. openfile: rides the lost-wake producer hunt; the rescue machinery bounds its damage to sub-second.