| ## fdtrack | |
| [TOC] | |
| fdtrack is a file descriptor leak checker added to Android in API level 30. | |
| fdtrack consists of two parts: a set of hooks in bionic to register a callback | |
| that's invoked on file descriptor operations, and a library that implements a | |
| hook to perform and store backtraces for file descriptor creation. | |
| ### bionic hooks | |
| bionic provides a header in the `bionic_libc_platform_headers` header_lib at <[bionic/fdtrack.h](https://android.googlesource.com/platform/bionic/+/refs/heads/master/libc/platform/bionic/fdtrack.h)>. | |
| Register a callback with `android_fdtrack_compare_exchange_hook` to receive | |
| callbacks upon file descriptor creation and destruction. This function can be | |
| called at any point in order to start capturing events, but be sure to properly | |
| handle unbalanced closes. This callback may be called from an async signal safe | |
| context, but not vfork (bionic tracks whether a thread is vforked, and chooses | |
| not to call callbacks when this is the case). | |
| ### libfdtrack | |
| [libfdtrack](https://android.googlesource.com/platform/bionic/+/refs/heads/master/libfdtrack) | |
| implements a library that uses libunwindstack to unwind and store fd creation backtraces. | |
| #### Using libfdtrack | |
| libfdtrack registers its hook upon being loaded, so to start capturing | |
| backtraces, `dlopen("libfdtrack.so", RTLD_GLOBAL)` is all that's needed. To dump | |
| its output to logcat, either use `fdtrack_dump`, or send the signal | |
| `BIONIC_SIGNAL_FDTRACK` (available from `<bionic/reserved_signals.h>`) to the | |
| process. If you wish to iterate through the results programmatically, | |
| `fdtrack_iterate` can be used (warning: this interface is currently unstable, | |
| don't use it in code that can be used on multiple platform versions.) | |
| libfdtrack adds a significant amount of overhead, so for processes that are | |
| latency-critical like system_server, it's not feasible to always capture | |
| backtraces. Instead, if you can detect that an fd leak is ongoing, turning on | |
| backtraces for a while and then triggering a dump can be sufficient. | |
| system_server [implements this approach](https://android.googlesource.com/platform/frameworks/base/+/679f3e4242b8e018eb7df90ef433f81088a64fff%5E%21/), | |
| spawning a thread that regularly checks the count of fds in the process, turns | |
| on fdtrack when it hits a threshold, and then aborts after another threshold. | |
| This dumps the output to logcat, which will be available in both the tombstone | |
| and logcat from bugreports. | |
| #### Implementation details | |
| There are multiple methods to unwind in Android: | |
| * libunwindstack | |
| * Pros | |
| * Primary method on the platform | |
| * Able to unwind through ART | |
| * Cons | |
| * Uses malloc internally: unsafe unless a separate allocator is | |
| statically-linked and steps are taken to prevent the unwind from being | |
| interrupted by a signal | |
| * Slow - infeasible to be used always in latency-sensitive processes | |
| * `android_unsafe_frame_pointer_chase` | |
| * Pros | |
| * Definitely async signal safe | |
| * Fast | |
| * Cons | |
| * Unable to unwind through ART because it doesn't maintain the frame pointer | |
| * Requires -fno-omit-frame-pointer to be used on all code being unwound | |
| through, which currently isn't the case on Android | |
| * Frame layout is a mess on 32-bit ARM: the ARM standard, clang, and GCC | |
| [all disagree](https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92172) | |
| * Chasing the frame pointer will often result in multiple frames inside the | |
| same function | |
| libfdtrack chooses to use libunwindstack for now, since unwinding through ART | |
| is critical to being useful for the initial user, system_server. |