Add `jl_print_task_backtraces()`

[kpamnany]

Iterates through jl_all_tls_states and through all live_tasks in ptls->heap, printing backtraces.

The purpose is to help find deadlocks.

Replaces https://github.com/JuliaLang/julia/pull/44990. Closes #46177.

[kpamnany]

For this error:

/cache/build/default-amdci5-6/julialang/julia-master/src/stackwalk.c:1124:9: error: Passing non-rooted value as argument to function that may GC [julia.GCChecker]
--
  | jlbacktracet(ptls2->root_task);
  | ^            ~~~~~~~~~~~~~~~~

Do I have to JL_GC_PUSH1(&ptls2->root_task) and JL_GC_POP() after?

[kpamnany]

Also, I don't understand this?

/cache/build/default-amdci5-6/julialang/julia-master/src/stackwalk.c:1119:54: note: Argument value was derived global with untracked type. You may want to update the checker's type list
--
  | jl_safe_printf("     ---- Root task (%p)\n", ptls2->root_task);
  | ^~~~~~~~~~~~~~~~

[kpamnany]

We're experimenting with this to see if it's useful. As Nathan said, this is basically for use in gdb, i.e. all threads will be stopped, so poking into other threads' local storage is safe.

I find that I cannot get a backtrace (Linux glibc x86-64) with the default JL_HAVE_ASM and have to turn on JL_HAVE_UNW_CONTEXT. Does https://github.com/JuliaLang/julia/pull/45110 fix that @vtjnash?

[vtjnash]

This should be useful eventually for figuring out why the Sockets test stales on CI, such as https://buildkite.com/julialang/julia-master/builds/16200#01838101-13bb-48fd-8aff-7705c619cd66

[NHDaly]

I find that I cannot get a backtrace (Linux glibc x86-64) with the default JL_HAVE_ASM and have to turn on JL_HAVE_UNW_CONTEXT. Does #45110 fix that @vtjnash?

Oh, interesting! I wonder if that's why I originally was thinking this didn't work. Thanks for tracking all of this down, @kpamnany!

@vtjnash / @kpamnany: is this resolved? Is there anything I/we can do to help move this along? Thanks! 😊

[kpamnany]

I added a comment warning that this is only intended for use when all threads are stopped (i.e. in gdb). I also removed it from exported functions for that reason.

We've verified that this can be useful. So, apart from the GC checker errors, I think this is good to go.

[Keno]

This is failing analyzegc on master. Not sure why it didn't fail on this PR:

/cache/build/default-amdci5-5/julialang/julia-master/src/stackwalk.c:1131:28: error: Implicit Atomic seq_cst synchronization [concurrency-implicit-atomics,-warnings-as-errors]
--
  | for (size_t i = 0; i < jl_n_threads; i++) {
  | ^
  | /cache/build/default-amdci5-5/julialang/julia-master/src/stackwalk.c:1132:27: error: Implicit Atomic seq_cst synchronization [concurrency-implicit-atomics,-warnings-as-errors]
  | jl_ptls_t ptls2 = jl_all_tls_states[i];
  | ^

[giordano]

This PR was reverted in #47182 because of the problem with analyzegc reported above.