Handle compressed debug sections in ELF files

ELF files allow debug info sections to be compressed. The libbacktrace
backed supported these compressed sections, but the Gimli backend did
not. This commit adds that support to the Gimli backend.

In my tests these debug info sections do not obey the alignment
requirements that the object crate expects for the gABI compression
header (nor can I find a source documenting any alignment requirements),
so this commit additionally enables the "unaligned" feature in the
upcoming version of the object crate.

There is a bit of unsafe to ensure the lifetime of the decompressed
sections matches the lifetime of the mmap'd file. I don't think there is
a way around this unsafe code, unless we are willing to ditch Gimli's
EndianSlice for an (apparently slower) EndianReader backed by a
Cow<[u8]>.

Fix #342.

Author: benesch

.github/workflows/main.yml

@@ -78,6 +78,10 @@ jobs:
     - run: cargo test --features gimli-symbolize --manifest-path crates/without_debuginfo/Cargo.toml
     - run: cargo test --manifest-path crates/line-tables-only/Cargo.toml --features libbacktrace
     - run: cargo test --manifest-path crates/line-tables-only/Cargo.toml --features gimli-symbolize
+    - run: RUSTFLAGS="-C link-arg=-Wl,--compress-debug-sections=zlib-gabi" cargo test --features gimli-symbolize
+      if: contains(matrix.os, 'ubuntu')
+    - run: RUSTFLAGS="-C link-arg=-Wl,--compress-debug-sections=zlib-gnu" cargo test --features gimli-symbolize
+      if: contains(matrix.os, 'ubuntu')
 
   windows_arm64:
     name: Windows AArch64

Cargo.toml

@@ -39,11 +39,12 @@ compiler_builtins = { version = '0.1.2', optional = true }
 # Optional dependencies enabled through the `gimli-symbolize` feature, do not
 # use these features directly.
 addr2line = { version = "0.12.0", optional = true, default-features = false }
+flate2 = { version = "1.0.14", optional = true }
 [dependencies.object]
-version = "0.19"
+git = "https://github.com/gimli-rs/object.git"
 optional = true
 default-features = false
-features = ['read_core', 'elf', 'macho', 'pe']
+features = ['read_core', 'elf', 'macho', 'pe', 'unaligned']
 
 [target.'cfg(windows)'.dependencies]
 winapi = { version = "0.3.3", optional = true }
@@ -71,7 +72,7 @@ std = []
 # be affected by feature selection here. Also note that it's highly unlikely you
 # want to configure this. If you're having trouble getting backtraces it's
 # likely best to open an issue.
-gimli-symbolize = ["addr2line", "object", "std"]
+gimli-symbolize = ["addr2line", "flate2", "object", "std"]
 libbacktrace = ["backtrace-sys/backtrace-sys"]
 
 #=======================================

src/symbolize/gimli/elf.rs

@@ -1,7 +1,8 @@
-use super::{Mapping, Path, Vec};
-use object::read::elf::{FileHeader, SectionHeader, SectionTable, Sym};
+use super::{Mapping, Mmap, Path, Vec};
+use object::elf::{ELFCOMPRESS_ZLIB, SHF_COMPRESSED};
+use object::read::elf::{CompressionHeader, FileHeader, SectionHeader, SectionTable, Sym};
 use object::read::StringTable;
-use object::{Bytes, NativeEndian};
+use object::{BigEndian, Bytes, NativeEndian};
 
 #[cfg(target_pointer_width = "32")]
 type Elf = object::elf::FileHeader32<NativeEndian>;
@@ -28,16 +29,16 @@ pub struct Object<'a> {
     /// We could use a literal instead, but this helps ensure correctness.
     endian: NativeEndian,
     /// The entire file data.
-    data: Bytes<'a>,
+    mmap: &'a Mmap,
     sections: SectionTable<'a, Elf>,
     strings: StringTable<'a>,
     /// List of pre-parsed and sorted symbols by base address.
     syms: Vec<ParsedSym>,
 }
 
 impl<'a> Object<'a> {
-    fn parse(data: &'a [u8]) -> Option<Object<'a>> {
-        let data = object::Bytes(data);
+    fn parse(mmap: &'a Mmap) -> Option<Object<'a>> {
+        let data = object::Bytes(mmap);
         let elf = Elf::parse(data).ok()?;
         let endian = elf.endian().ok()?;
         let sections = elf.sections(endian, data).ok()?;
@@ -80,22 +81,59 @@ impl<'a> Object<'a> {
         syms.sort_unstable_by_key(|s| s.address);
         Some(Object {
             endian,
-            data,
+            mmap,
             sections,
             strings,
             syms,
         })
     }
 
     pub fn section(&self, name: &str) -> Option<&'a [u8]> {
-        Some(
-            self.sections
-                .section_by_name(self.endian, name.as_bytes())?
-                .1
-                .data(self.endian, self.data)
-                .ok()?
-                .0,
-        )
+        if let Some(section) = self.section_header(name) {
+            let mut data = section.data(self.endian, self.data()).ok()?;
+
+            // Check for DWARF-standard (gABI) compression, i.e., as generated
+            // by ld's `--compress-debug-sections=zlib-gabi` flag.
+            let flags: u64 = section.sh_flags(self.endian).into();
+            if (flags & u64::from(SHF_COMPRESSED)) == 0 {
+                // Not compressed.
+                return Some(data.0);
+            }
+
+            let header = data.read::<<Elf as FileHeader>::CompressionHeader>().ok()?;
+            if header.ch_type(self.endian) != ELFCOMPRESS_ZLIB {
+                // Zlib compression is the only known type.
+                return None;
+            }
+            let size = header.ch_size(self.endian) as usize;
+            let buf = decompress_zlib(data.0, size)?;
+            return Some(self.mmap.stash(buf));
+        }
+
+        // Check for the nonstandard GNU compression format, i.e., as generated
+        // by ld's `--compress-debug-sections=zlib-gnu` flag.
+        let zdebug_name = format!(".zdebug_{}", &name[7..]);
+        if let Some(section) = self.section_header(&zdebug_name) {
+            let mut data = section.data(self.endian, self.data()).ok()?;
+            if data.read_bytes(8).ok()?.0 != b"ZLIB\0\0\0\0" {
+                return None;
+            }
+            let size = data.read::<object::U32Bytes<_>>().ok()?.get(BigEndian);
+            let buf = decompress_zlib(data.0, size as usize)?;
+            return Some(self.mmap.stash(buf));
+        }
+
+        None
+    }
+
+    fn section_header(&self, name: &str) -> Option<&<Elf as FileHeader>::SectionHeader> {
+        self.sections
+            .section_by_name(self.endian, name.as_bytes())
+            .map(|(_index, section)| section)
+    }
+
+    fn data(&self) -> Bytes<'a> {
+        object::Bytes(&self.mmap)
     }
 
     pub fn search_symtab<'b>(&'b self, addr: u64) -> Option<&'b [u8]> {
@@ -112,3 +150,12 @@ impl<'a> Object<'a> {
         }
     }
 }
+
+fn decompress_zlib(data: &[u8], size: usize) -> Option<Vec<u8>> {
+    let mut buf = Vec::with_capacity(size);
+    let header_expected = true;
+    flate2::Decompress::new(header_expected)
+        .decompress_vec(data, &mut buf, flate2::FlushDecompress::Finish)
+        .ok()?;
+    Some(buf)
+}

src/symbolize/gimli/mmap_unix.rs

@@ -1,12 +1,16 @@
+use std::cell::UnsafeCell;
 use std::fs::File;
 use std::ops::Deref;
 use std::os::unix::prelude::*;
 use std::ptr;
 use std::slice;
+use std::vec::Vec;
 
 pub struct Mmap {
     ptr: *mut libc::c_void,
     len: usize,
+    /// Additional byte vectors that need to live as long as the mmap.
+    buffers: UnsafeCell<Vec<Vec<u8>>>,
 }
 
 impl Mmap {
@@ -22,7 +26,24 @@ impl Mmap {
         if ptr == libc::MAP_FAILED {
             return None;
         }
-        Some(Mmap { ptr, len })
+        Some(Mmap {
+            ptr,
+            len,
+            buffers: UnsafeCell::new(vec![]),
+        })
+    }
+
+    /// Takes ownership of `buf` and returns a reference to its contents that
+    /// lives as long as this `Mmap` does.
+    pub fn stash(&self, buf: Vec<u8>) -> &[u8] {
+        // SAFETY: this is the only function that ever constructs a mutable
+        // reference to `self.buffers`.
+        let buffers = unsafe { &mut *self.buffers.get() };
+        let i = buffers.len();
+        buffers.push(buf);
+        // SAFETY: we never remove elements from `self.buffers`, so a reference
+        // to the data inside any buffer will live as long as `Mmap` does.
+        &buffers[i]
     }
 }