rr 5.9.0 (the version in Fedora 43) aborts when recording any multi-threaded program on Linux 6.13+. The kernel added MADV_GUARD_INSTALL (advice value 102) and MADV_GUARD_REMOVE (103) in Linux 6.13, and modern glibc uses them for thread-stack guard installation. rr 5.9.0 doesn't know these advice values and trips a fatal assertion in its syscall recorder.

Impact: blocks all Fedora 43 users with kernel >= 6.13 from using rr on any threaded program, which is most real-world workloads (Python with threads, Rust async runtimes, anything using glibc pthread stack guards).

The fix is already merged upstream. rr master at commit 95b6860 (2026-04-21) has the handler at src/record_syscall.cc:5329-5332 emulating MADV_GUARD_INSTALL and MADV_GUARD_REMOVE as returning -EINVAL. Building rr from current master resolves the issue completely, verified on this box.

Request: either pull a snapshot build past commit 95b6860, or rebase on the next upstream release whenever it ships. Upstream has been quiet on the release front (5.9.0 was tagged 2025-02-13, nothing since), so a snapshot build may be the pragmatic move.

Reproducible: Always

Steps to Reproduce:
1. rr record python3 -c "import threading, time; t = threading.Thread(target=lambda: time.sleep(0.01)); t.start(); t.join()"
Actual Results:
[FATAL src/record_syscall.cc:6754:rec_process_syscall_arch()] Assertion `t->regs().syscall_result_signed() == -syscall_state.expect_errno' failed to hold. Expected EINVAL for 'madvise' but got result 0 (errno SUCCESS); unknown madvise(102)

rr then aborts and produces an emergency debug prompt. No trace is recorded.

Expected Results:
rr records the program successfully and produces a replayable trace.

Additional Information:
Version info:
- rr-5.9.0-4.fc43.x86_64
- kernel 6.19.11-200.fc43.x86_64
- glibc-2.42-10.fc43.x86_64

Upstream rr master (commit 95b6860 as of 2026-04-21) has the fix at src/record_syscall.cc:5329-5332:

  case MADV_GUARD_INSTALL:
  case MADV_GUARD_REMOVE:
    syscall_state.emulate_result(-EINVAL);
    break;

This emulates both advice values as returning -EINVAL at the rr layer, matching pre-6.13 kernel behavior. rr's deterministic replay doesn't need real guard-page installation, so no user-visible functionality is lost.

Verified a master build (cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -Ddisable32bit=ON .. && ninja) records and replays the previously-failing command. Happy to help test a package update once a new build is available.