platform/posix: Port fuzzer to upstream "native_sim" board

The older native_posix board is being deprecated, use native_sim,
which is the future-proof API.  In theory this should be as simple as
just swapping the board name at the west level, but there are a few
changes:

The C API is broadly identical between the two, modulo some prefix
renaming.

Unfortunately linkage is more of a hassle, as the fuzzing framework
inverts the sense of "entry point" and causes some trouble with the
way native_sim does its two-stage link.  We have to add some hackery:

  1. Make sure the fuzz entry point doesn't get dropped during the
     initial zephyr.elf link, as it calls OS/sim layer and not the
     reverse.

  2. Force it to be a global symbol in the final stage, so it can be
     seen by the code in libfuzzer that needs to call it (normally
     all Zephyr-side symbols are forced to be library-private to
     prevent collisions with the global Linux/glibc namespace
     environment)

Signed-off-by: Andy Ross <[email protected]>

scripts/fuzz.sh

@@ -34,7 +34,7 @@ with the -b option.
 
 Simple wrapper around a libfuzzer test run, as much for
 documentation as direct use. The idea here is really simple: build
-for the Zephyr "native_posix" board (which is a just a x86
+for the Zephyr "native_sim" board (which is a just a x86
 executable for the build host, not an emulated device) and run the
 resulting zephyr.exe file. This specifies a "fuzz_corpus" directory
 to save the seeds that produce useful coverage output for use in
@@ -124,7 +124,7 @@ main()
  (set -x
  # When passing conflicting -DVAR='VAL UE1' -DVAR='VAL UE2' to CMake,
  # the last 'VAL UE2' wins. Previous ones are silently ignored.
- west build -d build-fuzz -b native_posix "$SOF_TOP"/app/ -- \
+ west build -d build-fuzz -b native_sim "$SOF_TOP"/app/ -- \
  "${fuzz_configs[@]}" "$@"
  )
 

src/platform/posix/fuzz.c

@@ -9,9 +9,9 @@
 #include <irq_ctrl.h>
 #include <zephyr/sys/time_units.h>
 
-/* Zephyr arch APIs, not in a header (native_sim has them though) */
-void posix_init(int argc, char *argv[]);
-void posix_exec_for(uint64_t us);
+/* Zephyr arch APIs, not in a header */
+void nsi_init(int argc, char *argv[]);
+void nsi_exec_for(uint64_t us);
 
 const uint8_t *posix_fuzz_buf;
 size_t posix_fuzz_sz;
@@ -28,7 +28,7 @@ int LLVMFuzzerTestOneInput(const uint8_t *data, size_t sz)
  static bool runner_initialized;
 
  if (!runner_initialized) {
- posix_init(0, NULL);
+ nsi_init(0, NULL);
  runner_initialized = true;
  }
 
@@ -42,6 +42,6 @@ int LLVMFuzzerTestOneInput(const uint8_t *data, size_t sz)
  /* Give the OS time to process whatever happened in that
  * interrupt and reach an idle state.
  */
- posix_exec_for(k_ticks_to_us_ceil64(CONFIG_ZEPHYR_POSIX_FUZZ_TICKS));
+ nsi_exec_for(k_ticks_to_us_ceil64(CONFIG_ZEPHYR_POSIX_FUZZ_TICKS));
  return 0;
 }

src/platform/posix/posix.c

@@ -33,6 +33,17 @@ __asm__(".globl _trace_ctx_start\n"
  ".globl _trace_ctx_end\n"
  "_trace_ctx_end:\n");
 
+/* Sort of a kluge: this is an app-provided symbol, Zephyr doesn't
+ * know about it. But it's the ENTRY POINT of the binary, so SOF
+ * doesn't ever reference or call it either. It therefore gets
+ * dropped by the initial link of zephyr.elf, so isn't present anymore
+ * in the final link vs. the native_sim runner to link vs. the
+ * framework code in libfuzzer. Stash its address meaninglessly just
+ * to be sure the symbol gets referenced.
+ */
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t sz);
+void *native_sim_ptr_save;
+
 struct ipc_data_host_buffer *ipc_platform_get_host_buffer(struct ipc *ipc)
 {
  return NULL;
@@ -81,6 +92,8 @@ int platform_init(struct sof *sof)
  posix_dma_init(sof);
  ipc_init(sof);
 
+ native_sim_ptr_save = LLVMFuzzerTestOneInput;
+
  return 0;
 }
 

zephyr/CMakeLists.txt

@@ -401,6 +401,14 @@ zephyr_library_sources_ifdef(CONFIG_ZEPHYR_POSIX
  ${SOF_PLATFORM_PATH}/posix/fuzz.c
 )
 
+# When building the fuzzer, the LLVMFuzzerTestOneInput() entry point
+# (which is defined here in the app) must be exposed to the secondary
+# native_sim "runner" linkage step, which normally hides all
+# Zephyr-side symbols to protect against collisions. Somewhat opaque
+# cmake interface to this, alas.
+set_property(TARGET native_simulator APPEND PROPERTY
+ LOCALIZE_EXTRA_OPTIONS "--globalize-symbol=LLVMFuzzerTestOneInput")
+
 zephyr_library_sources_ifdef(CONFIG_LIBRARY
  ${SOF_PLATFORM_PATH}/library/platform.c
  ${SOF_PLATFORM_PATH}/library/lib/dai.c