Merge pull request #57 from pythonbpf/vmlinux-debug-info

Add debug info handling to vmlinux
* Does not add support for recursive ctypes pointer based resolution
* Still does not support unions and function pointers.
* Has the mechanism to build for function pointers added.

Author: varun-r-mallya

pythonbpf/debuginfo/debug_info_generator.py

@@ -81,6 +81,20 @@ def create_array_type(self, base_type: Any, count: int) -> Any:
             },
         )
 
+    def create_array_type_vmlinux(self, type_info: Any, count: int) -> Any:
+        """Create an array type of the given base type with specified count"""
+        base_type, type_sizing = type_info
+        subrange = self.module.add_debug_info("DISubrange", {"count": count})
+        return self.module.add_debug_info(
+            "DICompositeType",
+            {
+                "tag": dc.DW_TAG_array_type,
+                "baseType": base_type,
+                "size": type_sizing,
+                "elements": [subrange],
+            },
+        )
+
     @staticmethod
     def _compute_array_size(base_type: Any, count: int) -> int:
         # Extract size from base_type if possible
@@ -101,6 +115,23 @@ def create_struct_member(self, name: str, base_type: Any, offset: int) -> Any:
             },
         )
 
+    def create_struct_member_vmlinux(
+        self, name: str, base_type_with_size: Any, offset: int
+    ) -> Any:
+        """Create a struct member with the given name, type, and offset"""
+        base_type, type_size = base_type_with_size
+        return self.module.add_debug_info(
+            "DIDerivedType",
+            {
+                "tag": dc.DW_TAG_member,
+                "name": name,
+                "file": self.module._file_metadata,
+                "baseType": base_type,
+                "size": type_size,
+                "offset": offset,
+            },
+        )
+
     def create_struct_type(
         self, members: List[Any], size: int, is_distinct: bool
     ) -> Any:
@@ -116,6 +147,22 @@ def create_struct_type(
             is_distinct=is_distinct,
         )
 
+    def create_struct_type_with_name(
+        self, name: str, members: List[Any], size: int, is_distinct: bool
+    ) -> Any:
+        """Create a struct type with the given members and size"""
+        return self.module.add_debug_info(
+            "DICompositeType",
+            {
+                "name": name,
+                "tag": dc.DW_TAG_structure_type,
+                "file": self.module._file_metadata,
+                "size": size,
+                "elements": members,
+            },
+            is_distinct=is_distinct,
+        )
+
     def create_global_var_debug_info(
         self, name: str, var_type: Any, is_local: bool = False
     ) -> Any:

pythonbpf/vmlinux_parser/class_handler.py

@@ -99,11 +99,45 @@ def process_vmlinux_post_ast(
                 local_module_name = getattr(elem_type, "__module__", None)
                 new_dep_node.add_field(elem_name, elem_type, ready=False)
                 if local_module_name == ctypes.__name__:
+                    # TODO: need to process pointer to ctype and also CFUNCTYPES here recursively. Current processing is a single dereference
                     new_dep_node.set_field_bitfield_size(elem_name, elem_bitfield_size)
-                    new_dep_node.set_field_ready(elem_name, is_ready=True)
-                    logger.debug(
-                        f"Field {elem_name} is direct ctypes type: {elem_type}"
-                    )
+
+                    # Process pointer to ctype
+                    if isinstance(elem_type, type) and issubclass(
+                        elem_type, ctypes._Pointer
+                    ):
+                        # Get the pointed-to type
+                        pointed_type = elem_type._type_
+                        logger.debug(f"Found pointer to type: {pointed_type}")
+                        new_dep_node.set_field_containing_type(elem_name, pointed_type)
+                        new_dep_node.set_field_ctype_complex_type(
+                            elem_name, ctypes._Pointer
+                        )
+                        new_dep_node.set_field_ready(elem_name, is_ready=True)
+
+                    # Process function pointers (CFUNCTYPE)
+                    elif hasattr(elem_type, "_restype_") and hasattr(
+                        elem_type, "_argtypes_"
+                    ):
+                        # This is a CFUNCTYPE or similar
+                        logger.info(
+                            f"Function pointer detected for {elem_name} with return type {elem_type._restype_} and arguments {elem_type._argtypes_}"
+                        )
+                        # Set the field as ready but mark it with special handling
+                        new_dep_node.set_field_ctype_complex_type(
+                            elem_name, ctypes.CFUNCTYPE
+                        )
+                        new_dep_node.set_field_ready(elem_name, is_ready=True)
+                        logger.warning(
+                            "Blindly processing CFUNCTYPE ctypes to ensure compilation. Unsupported"
+                        )
+
+                    else:
+                        # Regular ctype
+                        new_dep_node.set_field_ready(elem_name, is_ready=True)
+                        logger.debug(
+                            f"Field {elem_name} is direct ctypes type: {elem_type}"
+                        )
                 elif local_module_name == "vmlinux":
                     new_dep_node.set_field_bitfield_size(elem_name, elem_bitfield_size)
                     logger.debug(
@@ -127,6 +161,10 @@ def process_vmlinux_post_ast(
                                     ctype_complex_type = ctypes.Array
                                 elif issubclass(elem_type, ctypes._Pointer):
                                     ctype_complex_type = ctypes._Pointer
+                                else:
+                                    raise ImportError(
+                                        "Non Array and Pointer type ctype imports not supported in current version"
+                                    )
                             else:
                                 raise TypeError("Unsupported ctypes subclass")
                         else:

pythonbpf/vmlinux_parser/ir_gen/debug_info_gen.py

@@ -1,15 +1,161 @@
-from pythonbpf.debuginfo import DebugInfoGenerator
+from pythonbpf.debuginfo import DebugInfoGenerator, dwarf_constants as dc
+from ..dependency_node import DependencyNode
+import ctypes
+import logging
+from typing import List, Any, Tuple
 
+logger = logging.getLogger(__name__)
 
-def debug_info_generation(struct, llvm_module):
+
+def debug_info_generation(
+    struct: DependencyNode,
+    llvm_module,
+    generated_debug_info: List[Tuple[DependencyNode, Any]],
+) -> Any:
+    """
+    Generate DWARF debug information for a struct defined in a DependencyNode.
+
+    Args:
+        struct: The dependency node containing struct information
+        llvm_module: The LLVM module to add debug info to
+        generated_debug_info: List of tuples (struct, debug_info) to track generated debug info
+
+    Returns:
+        The generated global variable debug info
+    """
+    # Set up debug info generator
     generator = DebugInfoGenerator(llvm_module)
-    # this is sample debug info generation
-    # i64type = generator.get_uint64_type()
 
-    struct_type = generator.create_struct_type([], 64 * 4, is_distinct=True)
+    # Check if debug info for this struct has already been generated
+    for existing_struct, debug_info in generated_debug_info:
+        if existing_struct.name == struct.name:
+            return debug_info
+
+    # Process all fields and create members for the struct
+    members = []
+    for field_name, field in struct.fields.items():
+        # Get appropriate debug type for this field
+        field_type = _get_field_debug_type(
+            field_name, field, generator, struct, generated_debug_info
+        )
+        # Create struct member with proper offset
+        member = generator.create_struct_member_vmlinux(
+            field_name, field_type, field.offset * 8
+        )
+        members.append(member)
 
-    global_var = generator.create_global_var_debug_info(
-        struct.name, struct_type, is_local=False
+    if struct.name.startswith("struct_"):
+        struct_name = struct.name.removeprefix("struct_")
+    else:
+        raise ValueError("Unions are not supported in the current version")
+    # Create struct type with all members
+    struct_type = generator.create_struct_type_with_name(
+        struct_name, members, struct.__sizeof__() * 8, is_distinct=True
     )
 
-    return global_var
+    return struct_type
+
+
+def _get_field_debug_type(
+    field_name: str,
+    field,
+    generator: DebugInfoGenerator,
+    parent_struct: DependencyNode,
+    generated_debug_info: List[Tuple[DependencyNode, Any]],
+) -> tuple[Any, int]:
+    """
+    Determine the appropriate debug type for a field based on its Python/ctypes type.
+
+    Args:
+        field_name: Name of the field
+        field: Field object containing type information
+        generator: DebugInfoGenerator instance
+        parent_struct: The parent struct containing this field
+        generated_debug_info: List of already generated debug info
+
+    Returns:
+        The debug info type for this field
+    """
+    # Handle complex types (arrays, pointers)
+    if field.ctype_complex_type is not None:
+        if issubclass(field.ctype_complex_type, ctypes.Array):
+            # Handle array types
+            element_type, base_type_size = _get_basic_debug_type(
+                field.containing_type, generator
+            )
+            return generator.create_array_type_vmlinux(
+                (element_type, base_type_size * field.type_size), field.type_size
+            ), field.type_size * base_type_size
+        elif issubclass(field.ctype_complex_type, ctypes._Pointer):
+            # Handle pointer types
+            pointee_type, _ = _get_basic_debug_type(field.containing_type, generator)
+            return generator.create_pointer_type(pointee_type), 64
+
+    # Handle other vmlinux types (nested structs)
+    if field.type.__module__ == "vmlinux":
+        # If it's a struct from vmlinux, check if we've already generated debug info for it
+        struct_name = field.type.__name__
+
+        # Look for existing debug info in the list
+        for existing_struct, debug_info in generated_debug_info:
+            if existing_struct.name == struct_name:
+                # Use existing debug info
+                return debug_info, existing_struct.__sizeof__()
+
+        # If not found, create a forward declaration
+        # This will be completed when the actual struct is processed
+        logger.warning("Forward declaration in struct created")
+        forward_type = generator.create_struct_type([], 0, is_distinct=True)
+        return forward_type, 0
+
+    # Handle basic C types
+    return _get_basic_debug_type(field.type, generator)
+
+
+def _get_basic_debug_type(ctype, generator: DebugInfoGenerator) -> Any:
+    """
+    Map a ctypes type to a DWARF debug type.
+
+    Args:
+        ctype: A ctypes type or Python type
+        generator: DebugInfoGenerator instance
+
+    Returns:
+        The corresponding debug type
+    """
+    # Map ctypes to debug info types
+    if ctype == ctypes.c_char or ctype == ctypes.c_byte:
+        return generator.get_basic_type("char", 8, dc.DW_ATE_signed_char), 8
+    elif ctype == ctypes.c_ubyte or ctype == ctypes.c_uint8:
+        return generator.get_basic_type("unsigned char", 8, dc.DW_ATE_unsigned_char), 8
+    elif ctype == ctypes.c_short or ctype == ctypes.c_int16:
+        return generator.get_basic_type("short", 16, dc.DW_ATE_signed), 16
+    elif ctype == ctypes.c_ushort or ctype == ctypes.c_uint16:
+        return generator.get_basic_type("unsigned short", 16, dc.DW_ATE_unsigned), 16
+    elif ctype == ctypes.c_int or ctype == ctypes.c_int32:
+        return generator.get_basic_type("int", 32, dc.DW_ATE_signed), 32
+    elif ctype == ctypes.c_uint or ctype == ctypes.c_uint32:
+        return generator.get_basic_type("unsigned int", 32, dc.DW_ATE_unsigned), 32
+    elif ctype == ctypes.c_long:
+        return generator.get_basic_type("long", 64, dc.DW_ATE_signed), 64
+    elif ctype == ctypes.c_ulong:
+        return generator.get_basic_type("unsigned long", 64, dc.DW_ATE_unsigned), 64
+    elif ctype == ctypes.c_longlong or ctype == ctypes.c_int64:
+        return generator.get_basic_type("long long", 64, dc.DW_ATE_signed), 64
+    elif ctype == ctypes.c_ulonglong or ctype == ctypes.c_uint64:
+        return generator.get_basic_type(
+            "unsigned long long", 64, dc.DW_ATE_unsigned
+        ), 64
+    elif ctype == ctypes.c_float:
+        return generator.get_basic_type("float", 32, dc.DW_ATE_float), 32
+    elif ctype == ctypes.c_double:
+        return generator.get_basic_type("double", 64, dc.DW_ATE_float), 64
+    elif ctype == ctypes.c_bool:
+        return generator.get_basic_type("bool", 8, dc.DW_ATE_boolean), 8
+    elif ctype == ctypes.c_char_p:
+        char_type = generator.get_basic_type("char", 8, dc.DW_ATE_signed_char), 8
+        return generator.create_pointer_type(char_type)
+    elif ctype == ctypes.c_void_p:
+        return generator.create_pointer_type(None), 64
+    else:
+        return generator.get_uint64_type(), 64

pythonbpf/vmlinux_parser/ir_gen/ir_generation.py

@@ -14,6 +14,7 @@ def __init__(self, llvm_module, handler: DependencyHandler, assignment=None):
         self.llvm_module = llvm_module
         self.handler: DependencyHandler = handler
         self.generated: list[str] = []
+        self.generated_debug_info: list = []
         if not handler.is_ready:
             raise ImportError(
                 "Semantic analysis of vmlinux imports failed. Cannot generate IR"
@@ -67,18 +68,22 @@ def struct_processor(self, struct, processing_stack=None):
                             )
 
             # Actual processor logic here after dependencies are resolved
-            self.gen_ir(struct)
+            self.generated_debug_info.append(
+                (struct, self.gen_ir(struct, self.generated_debug_info))
+            )
             self.generated.append(struct.name)
 
         finally:
             # Remove from processing stack after we're done
             processing_stack.discard(struct.name)
 
-    def gen_ir(self, struct):
+    def gen_ir(self, struct, generated_debug_info):
         # TODO: we add the btf_ama attribute by monkey patching in the end of compilation, but once llvmlite
         #  accepts our issue, we will resort to normal accessed attribute based attribute addition
         # currently we generate all possible field accesses for CO-RE and put into the assignment table
-        debug_info = debug_info_generation(struct, self.llvm_module)
+        debug_info = debug_info_generation(
+            struct, self.llvm_module, generated_debug_info
+        )
         field_index = 0
         for field_name, field in struct.fields.items():
             # does not take arrays and similar types into consideration yet.
@@ -126,6 +131,7 @@ def gen_ir(self, struct):
                 )
                 globvar.linkage = "external"
                 globvar.set_metadata("llvm.preserve.access.index", debug_info)
+        return debug_info
 
     def _struct_name_generator(
         self,

tests/c-form/ex7.bpf.c

@@ -19,7 +19,7 @@ struct {
 SEC("tp/syscalls/sys_enter_setuid")
 int handle_setuid_entry(struct trace_event_raw_sys_enter *ctx) {
   struct event data = {};
-
+  struct blk_integrity_iter it = {};
   // Extract UID from the syscall arguments
   data.uid = (unsigned int)ctx->args[0];
   data.ts = bpf_ktime_get_ns();

tests/passing_tests/vmlinux/simple_struct_test.py

@@ -0,0 +1,30 @@
+from pythonbpf import bpf, section, bpfglobal, compile_to_ir, compile
+from vmlinux import TASK_COMM_LEN  # noqa: F401
+from vmlinux import struct_trace_event_raw_sys_enter  # noqa: F401
+
+# from vmlinux import struct_uinput_device
+# from vmlinux import struct_blk_integrity_iter
+from ctypes import c_int64
+
+
+# Instructions to how to run this program
+# 1. Install PythonBPF: pip install pythonbpf
+# 2. Run the program: python examples/simple_struct_test.py
+# 3. Run the program with sudo: sudo tools/check.sh run examples/simple_struct_test.o
+# 4. Attach object file to any network device with something like ./check.sh run examples/simple_struct_test.o tailscale0
+# 5. send traffic through the device and observe effects
+@bpf
+@section("tracepoint/syscalls/sys_enter_execve")
+def hello_world(ctx: struct_trace_event_raw_sys_enter) -> c_int64:
+    print("Hello, World!")
+    return c_int64(0)
+
+
+@bpf
+@bpfglobal
+def LICENSE() -> str:
+    return "GPL"
+
+
+compile_to_ir("simple_struct_test.py", "simple_struct_test.ll")
+compile()