1# Unwinder Support Per Android Release 2This document describes the changes in the way the libunwindstack 3unwinder works on different Android versions. It does not describe 4every change in the code made between different versions, but is 5meant to allow an app developer to know what might be supported 6on different versions. It also describes the different way an unwind 7will display on different versions of Android. 8 9## Android 9 ("Pie", API level 28) 10libunwindstack was first introduced in API level 28. 11 12* Supports up to and including Dwarf 4 unwinding information. 13 See http://dwarfstd.org/ for Dwarf standards. 14* Supports Arm exidx unwinding. 15* Supports the gdb JIT unwinding interface, which is how ART creates unwinding 16 information for the JIT'd Java frames. 17* Supports special frames added to represent an ART Java interpreter frame. 18 ART has marked the dex pc using cfi information that the unwinder 19 understands and handles by adding a new frame in the stacktrace. 20 21## Note 22By default, lld creates two separate maps of the elf in memory, one read-only 23and one read/executable. The libunwindstack on P and the unwinder on older 24versions of Android will not unwind properly in this case. For apps that 25target API level 28 or older, make sure that `-Wl,--no-rosegment` is 26included in linker arguments when using lld. 27 28## Android 10 ("Q", API level 29) 29* Fix bug (b/109824792) that handled load bias data incorrectly when 30 FDEs use pc relative addressing in the eh\_frame\_hdr. 31 Unfortunately, this wasn't fixed correctly in Q since it assumes 32 that the bias is coming from the program header for the executable 33 load. The real fix was to use the bias from the actual section data and 34 is not completely fixed until API level 30. For apps targeting API level 29, 35 if it is being compiled with the llvm linker lld, it might be necessary 36 to add the linker option `-Wl,-zseparate-code` to avoid creating an elf 37 created this way. 38* Change the way the exidx section offset is found (b/110704153). Before 39 the p\_vaddr value from the program header minus the load bias was used 40 to find the start of the exidx data. Changed to use the p\_offset since 41 it doesn't require any load bias manipulations. 42* Fix bug handling of dwarf sections without any header (b/110235461). 43 Previously, the code assumed that FDEs are non-overlapping, and the FDEs 44 are always in sorted order from low pc to high pc. Thus the code would 45 read the entire set of CIEs/FDEs and then do a binary search to find 46 the appropriate FDE for a given pc. Now the code does a sequential read 47 and stops when it finds the FDE for a pc. It also understands the 48 overlapping FDEs, so find the first FDE that matches a pc. In practice, 49 elf files with this format only ever occurs if the file was generated 50 without an eh\_frame/eh\_frame\_hdr section and only a debug\_frame. The 51 other way this has been observed is when running simpleperf to unwind since 52 sometimes there is not enough information in the eh\_frame for all points 53 in the executable. On API level 28, this would result in some incorrect 54 unwinds coming from simpleperf. Nearly all crashes from API level 28 should 55 be correct since the eh\_frame information was enough to do the unwind 56 properly. 57* Be permissive of badly formed elf files. Previously, any detected error 58 would result in unwinds stopping even if there is enough valid information 59 to do an unwind. 60 * The code now allows program header/section header offsets to point 61 to unreadable memory. As long as the code can find the unwind tables, 62 that is good enough. 63 * The code allows program headers/section headers to be missing. 64 * Allow a symbol table section header to point to invalid symbol table 65 values. 66* Support for the linker read-only segment option (b/109657296). 67 This is a feature of lld whereby there are two sections that 68 contain elf data. The first is read-only and contains the elf header data, 69 and the second is read-execute or execute only that 70 contains the executable code from the elf. Before this, the unwinder 71 always assumed that there was only a single read-execute section that 72 contained the elf header data and the executable code. 73* Build ID information for elf objects added. This will display the 74 NT\_GNU\_BUILD\_ID note found in elf files. This information can be used 75 to identify the exact version of a shared library to help get symbol 76 information when looking at a crash. 77* Add support for displaying the soname from an apk frame. Previously, 78 a frame map name would be only the apk, but now if the shared library 79 in the apk has set a soname, the map name will be `app.apk!libexample.so` 80 instead of only `app.apk`. 81* Minimal support for Dwarf 5. This merely treats a Dwarf 5 version 82 elf file as Dwarf 4. It does not support the new dwarf ops in Dwarf 5. 83 Since the new ops are not likely to be used very often, this allows 84 continuing to unwind even when encountering Dwarf 5 elf files. 85* Fix bug in pc handling of signal frames (b/130302288). In the previous 86 version, the pc would be wrong in the signal frame. The rest of the 87 unwind was correct, only the frame in the signal handler was incorrect 88 in API level 28. 89* Detect when an elf file is not readable so that a message can be 90 displayed indicating that. This can happen when an app puts the shared 91 libraries in non-standard locations that are not readable due to 92 security restrictions (selinux rules). 93 94## Android 11 ("R", API level 30) 95* Display the offsets for Java interpreter frames. If this frame came 96 from a non-zero offset map, no offset is printed. Previously, the 97 line would look like: 98 99 #17 pc 00500d7a GoogleCamera.apk (com.google.camera.AndroidPriorityThread.run+10) 100 101 to: 102 103 #17 pc 00500d7a GoogleCamera.apk (offset 0x11d0000) (com.google.camera.AndroidPriorityThread.run+10) 104* Fix bug where the load bias was set from the first PT\_LOAD program 105 header that has a zero p\_offset value. Now it is set from the first 106 executable PT\_LOAD program header. This has only ever been a problem 107 for host executables compiled for the x86\_64 architecture. 108* Switched to the libc++ demangler for function names. Previously, the 109 demangler used was not complete, so some less common demangled function 110 names would not be properly demangled or the function name would not be 111 demangled at all. 112* Fix bug in load bias handling. If the unwind information in the eh\_frame 113 or eh\_frame\_hdr does not have the same bias as the executable section, 114 and uses pc relative FDEs, the unwind will be incorrect. This tends 115 to truncate unwinds since the unwinder could not find the correct unwind 116 information for a given pc. 117