Make backtraces work on Redox, copying Unix implementation

Author: ids1024

src/libstd/build.rs

@@ -21,7 +21,7 @@ fn main() {
     let target = env::var("TARGET").expect("TARGET was not set");
     let host = env::var("HOST").expect("HOST was not set");
     if cfg!(feature = "backtrace") && !target.contains("apple") && !target.contains("msvc") &&
-        !target.contains("emscripten") && !target.contains("fuchsia") && !target.contains("redox") {
+        !target.contains("emscripten") && !target.contains("fuchsia") {
         let _ = build_libbacktrace(&host, &target);
     }
 

src/libstd/sys/redox/backtrace.rs

@@ -0,0 +1,116 @@
+// Copyright 2015 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+/// Backtrace support built on libgcc with some extra OS-specific support
+///
+/// Some methods of getting a backtrace:
+///
+/// * The backtrace() functions on unix. It turns out this doesn't work very
+///   well for green threads on macOS, and the address to symbol portion of it
+///   suffers problems that are described below.
+///
+/// * Using libunwind. This is more difficult than it sounds because libunwind
+///   isn't installed everywhere by default. It's also a bit of a hefty library,
+///   so possibly not the best option. When testing, libunwind was excellent at
+///   getting both accurate backtraces and accurate symbols across platforms.
+///   This route was not chosen in favor of the next option, however.
+///
+/// * We're already using libgcc_s for exceptions in rust (triggering thread
+///   unwinding and running destructors on the stack), and it turns out that it
+///   conveniently comes with a function that also gives us a backtrace. All of
+///   these functions look like _Unwind_*, but it's not quite the full
+///   repertoire of the libunwind API. Due to it already being in use, this was
+///   the chosen route of getting a backtrace.
+///
+/// After choosing libgcc_s for backtraces, the sad part is that it will only
+/// give us a stack trace of instruction pointers. Thankfully these instruction
+/// pointers are accurate (they work for green and native threads), but it's
+/// then up to us again to figure out how to translate these addresses to
+/// symbols. As with before, we have a few options. Before, that, a little bit
+/// of an interlude about symbols. This is my very limited knowledge about
+/// symbol tables, and this information is likely slightly wrong, but the
+/// general idea should be correct.
+///
+/// When talking about symbols, it's helpful to know a few things about where
+/// symbols are located. Some symbols are located in the dynamic symbol table
+/// of the executable which in theory means that they're available for dynamic
+/// linking and lookup. Other symbols end up only in the local symbol table of
+/// the file. This loosely corresponds to pub and priv functions in Rust.
+///
+/// Armed with this knowledge, we know that our solution for address to symbol
+/// translation will need to consult both the local and dynamic symbol tables.
+/// With that in mind, here's our options of translating an address to
+/// a symbol.
+///
+/// * Use dladdr(). The original backtrace()-based idea actually uses dladdr()
+///   behind the scenes to translate, and this is why backtrace() was not used.
+///   Conveniently, this method works fantastically on macOS. It appears dladdr()
+///   uses magic to consult the local symbol table, or we're putting everything
+///   in the dynamic symbol table anyway. Regardless, for macOS, this is the
+///   method used for translation. It's provided by the system and easy to do.o
+///
+///   Sadly, all other systems have a dladdr() implementation that does not
+///   consult the local symbol table. This means that most functions are blank
+///   because they don't have symbols. This means that we need another solution.
+///
+/// * Use unw_get_proc_name(). This is part of the libunwind api (not the
+///   libgcc_s version of the libunwind api), but involves taking a dependency
+///   to libunwind. We may pursue this route in the future if we bundle
+///   libunwind, but libunwind was unwieldy enough that it was not chosen at
+///   this time to provide this functionality.
+///
+/// * Shell out to a utility like `readelf`. Crazy though it may sound, it's a
+///   semi-reasonable solution. The stdlib already knows how to spawn processes,
+///   so in theory it could invoke readelf, parse the output, and consult the
+///   local/dynamic symbol tables from there. This ended up not getting chosen
+///   due to the craziness of the idea plus the advent of the next option.
+///
+/// * Use `libbacktrace`. It turns out that this is a small library bundled in
+///   the gcc repository which provides backtrace and symbol translation
+///   functionality. All we really need from it is the backtrace functionality,
+///   and we only really need this on everything that's not macOS, so this is the
+///   chosen route for now.
+///
+/// In summary, the current situation uses libgcc_s to get a trace of stack
+/// pointers, and we use dladdr() or libbacktrace to translate these addresses
+/// to symbols. This is a bit of a hokey implementation as-is, but it works for
+/// all unix platforms we support right now, so it at least gets the job done.
+
+pub use self::tracing::unwind_backtrace;
+pub use self::printing::{foreach_symbol_fileline, resolve_symname};
+
+// tracing impls:
+mod tracing;
+// symbol resolvers:
+mod printing;
+
+#[cfg(not(any(target_os = "macos", target_os = "ios", target_os = "emscripten")))]
+pub mod gnu {
+    use io;
+    use fs;
+    use libc::c_char;
+    use vec::Vec;
+    use ffi::OsStr;
+    use os::unix::ffi::OsStrExt;
+    use io::Read;
+
+    pub fn get_executable_filename() -> io::Result<(Vec<c_char>, fs::File)> {
+        let mut exefile = fs::File::open("sys:exe")?;
+        let mut exename = Vec::new();
+        exefile.read_to_end(&mut exename)?;
+        if exename.last() == Some(&b'\n') {
+            exename.pop();
+        }
+        let file = fs::File::open(OsStr::from_bytes(&exename))?;
+        Ok((exename.into_iter().map(|c| c as c_char).collect(), file))
+    }
+}
+
+pub struct BacktraceContext;

src/libstd/sys/redox/backtrace/mod.rs

@@ -0,0 +1,53 @@
+// Copyright 2014-2015 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+use io;
+use intrinsics;
+use ffi::CStr;
+use libc;
+use sys::backtrace::BacktraceContext;
+use sys_common::backtrace::Frame;
+
+pub fn resolve_symname<F>(frame: Frame,
+                          callback: F,
+                          _: &BacktraceContext) -> io::Result<()>
+    where F: FnOnce(Option<&str>) -> io::Result<()>
+{
+    unsafe {
+        let mut info: Dl_info = intrinsics::init();
+        let symname = if dladdr(frame.exact_position, &mut info) == 0 {
+            None
+        } else {
+            CStr::from_ptr(info.dli_sname).to_str().ok()
+        };
+        callback(symname)
+    }
+}
+
+pub fn foreach_symbol_fileline<F>(_symbol_addr: Frame,
+                                  _f: F,
+                                  _: &BacktraceContext) -> io::Result<bool>
+    where F: FnMut(&[u8], libc::c_int) -> io::Result<()>
+{
+    Ok(false)
+}
+
+#[repr(C)]
+struct Dl_info {
+    dli_fname: *const libc::c_char,
+    dli_fbase: *mut libc::c_void,
+    dli_sname: *const libc::c_char,
+    dli_saddr: *mut libc::c_void,
+}
+
+extern {
+    fn dladdr(addr: *const libc::c_void,
+              info: *mut Dl_info) -> libc::c_int;
+}

src/libstd/sys/redox/backtrace/printing/dladdr.rs

@@ -0,0 +1,22 @@
+// Copyright 2014-2015 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+pub use self::imp::{foreach_symbol_fileline, resolve_symname};
+
+#[cfg(any(target_os = "macos", target_os = "ios",
+          target_os = "emscripten"))]
+#[path = "dladdr.rs"]
+mod imp;
+
+#[cfg(not(any(target_os = "macos", target_os = "ios",
+              target_os = "emscripten")))]
+mod imp {
+    pub use sys_common::gnu::libbacktrace::{foreach_symbol_fileline, resolve_symname};
+}

src/libstd/sys/redox/backtrace/printing/mod.rs

@@ -0,0 +1,48 @@
+// Copyright 2014-2015 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+/// As always - iOS on arm uses SjLj exceptions and
+/// _Unwind_Backtrace is even not available there. Still,
+/// backtraces could be extracted using a backtrace function,
+/// which thanks god is public
+///
+/// As mentioned in a huge comment block in `super::super`, backtrace
+/// doesn't play well with green threads, so while it is extremely nice and
+/// simple to use it should be used only on iOS devices as the only viable
+/// option.
+
+use io;
+use libc;
+use sys::backtrace::BacktraceContext;
+use sys_common::backtrace::Frame;
+
+#[inline(never)] // if we know this is a function call, we can skip it when
+                 // tracing
+pub fn unwind_backtrace(frames: &mut [Frame])
+    -> io::Result<(usize, BacktraceContext)>
+{
+    const FRAME_LEN: usize = 100;
+    assert!(FRAME_LEN >= frames.len());
+    let mut raw_frames = [::ptr::null_mut(); FRAME_LEN];
+    let nb_frames = unsafe {
+        backtrace(raw_frames.as_mut_ptr(), raw_frames.len() as libc::c_int)
+    } as usize;
+    for (from, to) in raw_frames.iter().zip(frames.iter_mut()).take(nb_frames) {
+        *to = Frame {
+            exact_position: *from,
+            symbol_addr: *from,
+        };
+    }
+    Ok((nb_frames as usize, BacktraceContext))
+}
+
+extern {
+    fn backtrace(buf: *mut *mut libc::c_void, sz: libc::c_int) -> libc::c_int;
+}

src/libstd/sys/redox/backtrace/tracing/backtrace_fn.rs

@@ -0,0 +1,106 @@
+// Copyright 2014-2015 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+use error::Error;
+use io;
+use libc;
+use sys::backtrace::BacktraceContext;
+use sys_common::backtrace::Frame;
+
+use unwind as uw;
+
+struct Context<'a> {
+    idx: usize,
+    frames: &'a mut [Frame],
+}
+
+#[derive(Debug)]
+struct UnwindError(uw::_Unwind_Reason_Code);
+
+impl Error for UnwindError {
+    fn description(&self) -> &'static str {
+        "unexpected return value while unwinding"
+    }
+}
+
+impl ::fmt::Display for UnwindError {
+    fn fmt(&self, f: &mut ::fmt::Formatter) -> ::fmt::Result {
+        write!(f, "{}: {:?}", self.description(), self.0)
+    }
+}
+
+#[inline(never)] // if we know this is a function call, we can skip it when
+                 // tracing
+pub fn unwind_backtrace(frames: &mut [Frame])
+    -> io::Result<(usize, BacktraceContext)>
+{
+    let mut cx = Context {
+        idx: 0,
+        frames: frames,
+    };
+    let result_unwind = unsafe {
+        uw::_Unwind_Backtrace(trace_fn,
+                              &mut cx as *mut Context
+                              as *mut libc::c_void)
+    };
+    // See libunwind:src/unwind/Backtrace.c for the return values.
+    // No, there is no doc.
+    match result_unwind {
+        // These return codes seem to be benign and need to be ignored for backtraces
+        // to show up properly on all tested platforms.
+        uw::_URC_END_OF_STACK | uw::_URC_FATAL_PHASE1_ERROR | uw::_URC_FAILURE => {
+            Ok((cx.idx, BacktraceContext))
+        }
+        _ => {
+            Err(io::Error::new(io::ErrorKind::Other,
+                               UnwindError(result_unwind)))
+        }
+    }
+}
+
+extern fn trace_fn(ctx: *mut uw::_Unwind_Context,
+                   arg: *mut libc::c_void) -> uw::_Unwind_Reason_Code {
+    let cx = unsafe { &mut *(arg as *mut Context) };
+    let mut ip_before_insn = 0;
+    let mut ip = unsafe {
+        uw::_Unwind_GetIPInfo(ctx, &mut ip_before_insn) as *mut libc::c_void
+    };
+    if !ip.is_null() && ip_before_insn == 0 {
+        // this is a non-signaling frame, so `ip` refers to the address
+        // after the calling instruction. account for that.
+        ip = (ip as usize - 1) as *mut _;
+    }
+
+    // dladdr() on osx gets whiny when we use FindEnclosingFunction, and
+    // it appears to work fine without it, so we only use
+    // FindEnclosingFunction on non-osx platforms. In doing so, we get a
+    // slightly more accurate stack trace in the process.
+    //
+    // This is often because panic involves the last instruction of a
+    // function being "call std::rt::begin_unwind", with no ret
+    // instructions after it. This means that the return instruction
+    // pointer points *outside* of the calling function, and by
+    // unwinding it we go back to the original function.
+    let symaddr = if cfg!(target_os = "macos") || cfg!(target_os = "ios") {
+        ip
+    } else {
+        unsafe { uw::_Unwind_FindEnclosingFunction(ip) }
+    };
+
+    if cx.idx < cx.frames.len() {
+        cx.frames[cx.idx] = Frame {
+            symbol_addr: symaddr,
+            exact_position: ip,
+        };
+        cx.idx += 1;
+    }
+
+    uw::_URC_NO_REASON
+}