APS5_HANG_DUMP_SECS on Linux

[fisherjoey]

Problem

#406 and #521 added the all-threads dump and the APS5_HANG_DUMP_SECS watchdog in core/libs/prx/libc/src/specifics/windows/CrashReport.cpp. Linux has neither: there is no Linux crash reporter in libc/src/specifics/, and APS5_HANG_DUMP_SECS is read only in that Windows file. When a title stops presenting on Linux, the stacks come from attaching gdb or eu-stack by hand, which assumes the person is at the machine when it happens and has the tools installed. Bug reports from Linux users then have no stacks.

Proposal

A Linux counterpart with the same variable and the same behaviour: with APS5_HANG_DUMP_SECS=N, a watcher thread sleeps N seconds, writes every thread's stack to stderr in the same format as the Windows report (thread id, name, registers, module+offset frames) and aborts.

• Threads are listed from /proc/self/task. The watcher sends each one a real-time signal with tgkill; the handler records its context and frames into a preallocated slot and returns. The watcher waits for each slot with a timeout, so a thread that can't run the handler is reported as such instead of hanging the dump.
• Addresses are described after collection, with dladdr, outside any signal handler.
• Thread names come from /proc/self/task/<tid>/comm.
• New file core/libs/prx/libc/src/specifics/linux/CrashReport.cpp, built only on Linux.

Not in scope: an all-threads dump on fatal signals on Linux. That is the natural follow-up and can reuse the same collection code, but it interacts with the guest signal delivery discussed in #727, so it would be a separate pull request.

What we have

An external script that does the same from outside the process (eu-stack over every task, frames mapped to the title's ELF as app+offset). We used it to diagnose a title that hangs at its menu. The in-process version is not written yet.

Questions

• Would you take this as a pull request?
• Which real-time signal number should it use, given the other host signal users in the project?
• Is a fixed timer (as on Windows) enough, or would you also want a variant that fires only when no frame has been presented for N seconds? That would need a hook in the presenter, so we would propose it separately.

AI-assisted: yes