std.debug: fix some corner cases #23927
Conversation
| @@ -959,8 +985,8 @@ pub fn writeCurrentStackTrace( | |||
| start_addr: ?usize, | |||
| ) !void { | |||
| if (native_os == .windows) { | |||
| var context: ThreadContext = undefined; | |||
| assert(getContext(&context)); | |||
| var context: windows.CONTEXT = std.mem.zeroes(windows.CONTEXT); | |||
There was a problem hiding this comment.
I'm not sure. I'm manually inlining std.debug.getContext here, and it already does the zeroing. I don't see any documentation that suggests this is necessary or not, though.
| // Getting the backtrace inside the signal handler (with the ucontext_t) | ||
| // gets stuck in a loop on some systems: | ||
| const expect_signal_frame_overflow = | ||
| (native_arch == .arm and link_libc); // loops above main() |
There was a problem hiding this comment.
glibc vs musl? armeb, thumb, thumbeb?
There was a problem hiding this comment.
I haven't investigated the failure cases too closely yet. I'm just trying to get the test to compile and not blow up on every architecture. And I'm trying to avoid watering the test down too much on the platforms where it works reliably. That said, the failure I see for this one is when statically linking musl to the test case.
I haven't been building the other ARM variants, so I'll try and mix those in too.
| native_arch == .mips or | ||
| native_arch == .mipsel or | ||
| native_arch == .mips64 or | ||
| native_arch == .mips64el or | ||
| native_arch == .powerpc64 or | ||
| native_arch == .powerpc64le; |
There was a problem hiding this comment.
Generally they don't seem to generate traces at all (either through dumpCurrentStackTrace or with StackIterator). Zig doesn't have a ucontext_t on MIPS. I'm not sure what's up with the PowerPC ones.
On a related note, Zig CI failed on aarch64-linux (no libc) because the stack trace gets stuck in a loop above main (see https://github.com/ziglang/zig/actions/runs/15104670128/job/42451228605?pr=23927). So I've added .aarch64 to this list of ignorable failures for now. Its flaky though. The test sometimes builds traces without looping for me locally, and sometimes not. (From what I can tell a specific build of test test is deterministic, but across multiple builds its not.)
This test creates three nested stack frames and then tests stack trace creation. Add some additional tests of stack traces by invoking "dumpCurrentStackTrace()" and by using a signal handler's "context" parameter to feed backtrace construction. Make the test case at least runnable on a wide variety of systems (including Windows, and WASI). Because `ucontext_t` and `getcontext` are not evenly supported everywhere, some systems are expected only get through parts of the test.
rootbeer
force-pushed
the
debug-mcontext-stage1
branch
from
b95af2b to
37ebc96
Compare
Add infinite loop detection to the
std.debugbacktraces. Make the backtrace and stacktrace code more robust on corner-case architectures.Expand the "unwind.zig" test case to exercise
std.debug.dumpCurrentStackTrace(). And trigger a signal handler so the test can exercisestd.debug.dumpStackTraceFromBase()andstd.debug.StackIterator.initWithContext()using a kernel-constructed context.This is preparation for moving
std.debugaway fromgetContext()(#23801).