diff --git a/docs/32-bit-abi.md b/docs/32-bit-abi.md
index 5e3ae45..607b5d2 100644
--- a/docs/32-bit-abi.md
+++ b/docs/32-bit-abi.md
@@ -1,8 +1,6 @@
-32-bit ABI bugs
-===============
+# 32-bit ABI bugs
 
-`off_t` is 32-bit
------------------
+## 32-bit `off_t` and `_FILE_OFFSET_BITS=64`
 
 On 32-bit Android, `off_t` is a signed 32-bit integer. This limits functions
 that use `off_t` to working on files no larger than 2GiB.
@@ -34,13 +32,19 @@
 increase your target API level, you'll have more and more of the functions
 available. API 12 adds some of the `<unistd.h>` functions, API 21 adds `mmap`,
 and by API 24 you have everything including `<stdio.h>`. See the
-[linker map](libc/libc.map.txt) for full details.
+[linker map](libc/libc.map.txt) for full details. Note also that in NDK r16 and
+later, we inline an mmap64 implementation in the headers when you target an API
+before 21 because it's an easy special case that's often needed. This means
+that code using `_FILE_OFFSET_BITS=64` and `mmap` will always compile.
+
+If your code stops compiling when you move to NDK r15 or later, removing any
+definition of `_FILE_OFFSET_BITS=64` will restore the behavior you used to have:
+you'll have a 32-bit `off_t` and use the 32-bit functions.
 
 In the 64-bit ABI, `off_t` is always 64-bit.
 
 
-`sigset_t` is too small for real-time signals
----------------------------------------------
+## `sigset_t` is too small for real-time signals
 
 On 32-bit Android, `sigset_t` is too small for ARM and x86 (but correct for
 MIPS). This means that there is no support for real-time signals in 32-bit
@@ -49,8 +53,7 @@
 In the 64-bit ABI, `sigset_t` is the correct size for every architecture.
 
 
-`time_t` is 32-bit
-------------------
+## `time_t` is 32-bit
 
 On 32-bit Android, `time_t` is 32-bit. The header `<time64.h>` and type
 `time64_t` exist as a workaround, but the kernel interfaces exposed on 32-bit
@@ -58,8 +61,7 @@
 
 In the 64-bit ABI, `time_t` is 64-bit.
 
-`pthread_mutex_t` is too small for large pids
----------------------------------------------
+## `pthread_mutex_t` is too small for large pids
 
 This doesn't generally affect Android devices, because on devices
 `/proc/sys/kernel/pid_max` is usually too small to hit the 16-bit limit,
diff --git a/libc/include/stdio.h b/libc/include/stdio.h
index 2d45fc5..e02a371 100644
--- a/libc/include/stdio.h
+++ b/libc/include/stdio.h
@@ -173,6 +173,7 @@
 int fseek(FILE* __fp, long __offset, int __whence);
 long ftell(FILE* __fp);
 
+/* See https://android.googlesource.com/platform/bionic/+/master/docs/32-bit-abi.md */
 #if defined(__USE_FILE_OFFSET64)
 int fgetpos(FILE* __fp, fpos_t* __pos) __RENAME(fgetpos64) __INTRODUCED_IN(24);
 int fsetpos(FILE* __fp, const fpos_t* __pos) __RENAME(fsetpos64) __INTRODUCED_IN(24);
diff --git a/libc/include/sys/cdefs.h b/libc/include/sys/cdefs.h
index 833ced0..3cf6723 100644
--- a/libc/include/sys/cdefs.h
+++ b/libc/include/sys/cdefs.h
@@ -205,25 +205,28 @@
  * In our header files we test against __USE_BSD and __USE_GNU.
  */
 #if defined(_GNU_SOURCE)
-# define __USE_BSD 1
-# define __USE_GNU 1
+#  define __USE_BSD 1
+#  define __USE_GNU 1
 #endif
 
 #if defined(_BSD_SOURCE)
-# define __USE_BSD 1
+#  define __USE_BSD 1
 #endif
 
-/* _FILE_OFFSET_BITS 64 support. */
+/*
+ * _FILE_OFFSET_BITS 64 support.
+ * See https://android.googlesource.com/platform/bionic/+/master/docs/32-bit-abi.md
+ */
 #if !defined(__LP64__) && defined(_FILE_OFFSET_BITS) && _FILE_OFFSET_BITS == 64
-#define __USE_FILE_OFFSET64 1
+#  define __USE_FILE_OFFSET64 1
 /*
  * Note that __RENAME_IF_FILE_OFFSET64 is only valid if the off_t and off64_t
  * functions were both added at the same API level because if you use this,
  * you only have one declaration to attach __INTRODUCED_IN to.
  */
-#define __RENAME_IF_FILE_OFFSET64(func) __RENAME(func)
+#  define __RENAME_IF_FILE_OFFSET64(func) __RENAME(func)
 #else
-#define __RENAME_IF_FILE_OFFSET64(func)
+#  define __RENAME_IF_FILE_OFFSET64(func)
 #endif
 
 /*
diff --git a/libc/include/sys/mman.h b/libc/include/sys/mman.h
index 028b024..5a7e3c0 100644
--- a/libc/include/sys/mman.h
+++ b/libc/include/sys/mman.h
@@ -44,8 +44,9 @@
 #define MREMAP_MAYMOVE  1
 #define MREMAP_FIXED    2
 
-#if defined(__USE_FILE_OFFSET64)
 /*
+ * See https://android.googlesource.com/platform/bionic/+/master/docs/32-bit-abi.md
+ *
  * mmap64 wasn't really around until L, but we added an inline for it since it
  * allows a lot more code to compile with _FILE_OFFSET_BITS=64.
  *
@@ -54,17 +55,18 @@
  * mmap64 to every translation unit that includes this header. Instead, just
  * preserve the old behavior for GCC and emit a useful diagnostic.
  */
+#if defined(__USE_FILE_OFFSET64)
 void* mmap(void* __addr, size_t __size, int __prot, int __flags, int __fd, off_t __offset)
-#if !defined(__clang__) && __ANDROID_API__ < __ANDROID_API_L__
-    __attribute__((error("mmap is not available with _FILE_OFFSET_BITS=64 when using GCC until "
-                         "android-21. Either raise your minSdkVersion, disable "
-                         "_FILE_OFFSET_BITS=64, or switch to Clang.")));
-#else
+#  if !defined(__clang__) && __ANDROID_API__ < __ANDROID_API_L__
+      __attribute__((error("mmap is not available with _FILE_OFFSET_BITS=64 when using GCC until "
+                           "android-21. Either raise your minSdkVersion, disable "
+                           "_FILE_OFFSET_BITS=64, or switch to Clang.")));
+#  else
     __RENAME(mmap64);
-#endif  /* defined(__clang__) */
+#  endif
 #else
 void* mmap(void* __addr, size_t __size, int __prot, int __flags, int __fd, off_t __offset);
-#endif  /* defined(__USE_FILE_OFFSET64) */
+#endif
 
 #if __ANDROID_API__ >= __ANDROID_API_L__
 void* mmap64(void* __addr, size_t __size, int __prot, int __flags, int __fd, off64_t __offset) __INTRODUCED_IN(21);
diff --git a/libc/include/sys/sendfile.h b/libc/include/sys/sendfile.h
index ecdb76c..97432c0 100644
--- a/libc/include/sys/sendfile.h
+++ b/libc/include/sys/sendfile.h
@@ -34,6 +34,7 @@
 
 __BEGIN_DECLS
 
+/* See https://android.googlesource.com/platform/bionic/+/master/docs/32-bit-abi.md */
 #if defined(__USE_FILE_OFFSET64)
 ssize_t sendfile(int __out_fd, int __in_fd, off_t* __offset, size_t __count) __RENAME(sendfile64) __INTRODUCED_IN(21);
 #else
diff --git a/libc/include/sys/types.h b/libc/include/sys/types.h
index 26ad6a5..f07c8fd 100644
--- a/libc/include/sys/types.h
+++ b/libc/include/sys/types.h
@@ -95,12 +95,13 @@
 typedef __kernel_time_t __time_t;
 typedef __time_t time_t;
 
+/* This historical accident means that we had a 32-bit off_t on 32-bit architectures. */
+/* See https://android.googlesource.com/platform/bionic/+/master/docs/32-bit-abi.md */
 #if defined(__USE_FILE_OFFSET64) || defined(__LP64__)
 typedef int64_t off_t;
 typedef off_t loff_t;
 typedef loff_t off64_t;
 #else
-/* This historical accident means that we had a 32-bit off_t on 32-bit architectures. */
 typedef __kernel_off_t off_t;
 typedef __kernel_loff_t loff_t;
 typedef loff_t off64_t;
diff --git a/libc/include/unistd.h b/libc/include/unistd.h
index c60cf80..ef75c84 100644
--- a/libc/include/unistd.h
+++ b/libc/include/unistd.h
@@ -168,16 +168,10 @@
 int fsync(int __fd);
 int fdatasync(int __fd) __INTRODUCED_IN(9);
 
-#if defined(__USE_FILE_OFFSET64)
-off_t lseek(int __fd, off_t __offset, int __whence) __RENAME(lseek64);
-#else
-off_t lseek(int __fd, off_t __offset, int __whence);
-#endif
-
-off64_t lseek64(int __fd, off64_t __offset, int __whence);
-
+/* See https://android.googlesource.com/platform/bionic/+/master/docs/32-bit-abi.md */
 #if defined(__USE_FILE_OFFSET64)
 int truncate(const char* __path, off_t __length) __RENAME(truncate64) __INTRODUCED_IN(21);
+off_t lseek(int __fd, off_t __offset, int __whence) __RENAME(lseek64);
 ssize_t pread(int __fd, void* __buf, size_t __count, off_t __offset)
   __overloadable __RENAME(pread64) __INTRODUCED_IN(12);
 ssize_t pwrite(int __fd, const void* __buf, size_t __count, off_t __offset)
@@ -185,6 +179,7 @@
 int ftruncate(int __fd, off_t __length) __RENAME(ftruncate64) __INTRODUCED_IN(12);
 #else
 int truncate(const char* __path, off_t __length);
+off_t lseek(int __fd, off_t __offset, int __whence);
 ssize_t pread(int __fd, void* __buf, size_t __count, off_t __offset)
     __overloadable __RENAME_CLANG(pread);
 ssize_t pwrite(int __fd, const void* __buf, size_t __count, off_t __offset)
@@ -193,6 +188,7 @@
 #endif
 
 int truncate64(const char* __path, off64_t __length) __INTRODUCED_IN(21);
+off64_t lseek64(int __fd, off64_t __offset, int __whence);
 ssize_t pread64(int __fd, void* __buf, size_t __count, off64_t __offset)
     __INTRODUCED_IN(12) __overloadable __RENAME_CLANG(pread64);
 ssize_t pwrite64(int __fd, const void* __buf, size_t __count, off64_t __offset)