Unwinder Support Per Android Release

This document describes the changes in the way the libunwindstack unwinder works on different Android versions. It does not describe every change in the code made between different versions, but is meant to allow an app developer to know what might be supported on different versions. It also describes the different way an unwind will display on different versions of Android.

Android P

libunwindstack was first introduced in Android P.

  • Supports up to and including Dwarf 4 unwinding information. See http://dwarfstd.org/ for Dwarf standards.
  • Supports Arm exidx unwinding.
  • Supports the gdb JIT unwinding interface, which is how ART creates unwinding information for the JIT'd Java frames.
  • Supports special frames added to represent an ART Java interpreter frame. ART has marked the dex pc using cfi information that the unwinder understands and handles by adding a new frame in the stacktrace.

Note

By default, lld creates two separate maps of the elf in memory, one read-only and one read/executable. The libunwindstack on P and the unwinder on older versions of Android will not unwind properly in this case. For apps that target Android P or older, make sure that -Wl,--no-rosegment is included in linker arguments when using lld.

Android Q

  • Fix bug (b/109824792) that handled load bias data incorrectly when FDEs use pc relative addressing in the eh_frame_hdr. Unfortunately, this wasn't fixed correctly in Q since it assumes that the bias is coming from the program header for the executable load. The real fix was to use the bias from the actual section data and is not completely fixed until Android R. For apps targeting Android Q, if it is being compiled with the llvm linker lld, it might be necessary to add the linker option -Wl,-zseparate-code to avoid creating an elf created this way.
  • Change the way the exidx section offset is found (b/110704153). Before the p_vaddr value from the program header minus the load bias was used to find the start of the exidx data. Changed to use the p_offset since it doesn't require any load bias manipulations.
  • Fix bug handling of dwarf sections without any header (b/110235461). Previously, the code assumed that FDEs are non-overlapping, and the FDEs are always in sorted order from low pc to high pc. Thus the code would read the entire set of CIEs/FDEs and then do a binary search to find the appropriate FDE for a given pc. Now the code does a sequential read and stops when it finds the FDE for a pc. It also understands the overlapping FDEs, so find the first FDE that matches a pc. In practice, elf files with this format only ever occurs if the file was generated without an eh_frame/eh_frame_hdr section and only a debug_frame. The other way this has been observed is when running simpleperf to unwind since sometimes there is not enough information in the eh_frame for all points in the executable. On Android P, this would result in some incorrect unwinds coming from simpleperf. Nearly all crashes from Android P should be correct since the eh_frame information was enough to do the unwind properly.
  • Be permissive of badly formed elf files. Previously, any detected error would result in unwinds stopping even if there is enough valid information to do an unwind.
    • The code now allows program header/section header offsets to point to unreadable memory. As long as the code can find the unwind tables, that is good enough.
    • The code allows program headers/section headers to be missing.
    • Allow a symbol table section header to point to invalid symbol table values.
  • Support for the linker read-only segment option (b/109657296). This is a feature of lld whereby there are two sections that contain elf data. The first is read-only and contains the elf header data, and the second is read-execute or execute only that contains the executable code from the elf. Before this, the unwinder always assumed that there was only a single read-execute section that contained the elf header data and the executable code.
  • Build ID information for elf objects added. This will display the NT_GNU_BUILD_ID note found in elf files. This information can be used to identify the exact version of a shared library to help get symbol information when looking at a crash.
  • Add support for displaying the soname from an apk frame. Previously, a frame map name would be only the apk, but now if the shared library in the apk has set a soname, the map name will be app.apk!libexample.so instead of only app.apk.
  • Minimal support for Dwarf 5. This merely treats a Dwarf 5 version elf file as Dwarf 4. It does not support the new dwarf ops in Dwarf 5. Since the new ops are not likely to be used very often, this allows continuing to unwind even when encountering Dwarf 5 elf files.
  • Fix bug in pc handling of signal frames (b/130302288). In the previous version, the pc would be wrong in the signal frame. The rest of the unwind was correct, only the frame in the signal handler was incorrect in Android P.
  • Detect when an elf file is not readable so that a message can be displayed indicating that. This can happen when an app puts the shared libraries in non-standard locations that are not readable due to security restrictions (selinux rules).

Android R

  • Display the offsets for Java interpreter frames. If this frame came from a non-zero offset map, no offset is printed. Previously, the line would look like:

    #17 pc 00500d7a GoogleCamera.apk (com.google.camera.AndroidPriorityThread.run+10)

    to:

    #17 pc 00500d7a GoogleCamera.apk (offset 0x11d0000) (com.google.camera.AndroidPriorityThread.run+10)

  • Fix bug where the load bias was set from the first PT_LOAD program header that has a zero p_offset value. Now it is set from the first executable PT_LOAD program header. This has only ever been a problem for host executables compiled for the x86_64 architecture.

  • Switched to the libc++ demangler for function names. Previously, the demangler used was not complete, so some less common demangled function names would not be properly demangled or the function name would not be demangled at all.

  • Fix bug in load bias handling. If the unwind information in the eh_frame or eh_frame_hdr does not have the same bias as the executable section, and uses pc relative FDEs, the unwind will be incorrect. This tends to truncate unwinds since the unwinder could not find the correct unwind information for a given pc.