arm: Generate privileged stacks

This patch adds the generation and incorporation of privileged stack
regions that are used by ARM user mode threads.  This patch adds the
infrastructure for privileged stacks.  Later patches will utilize the
generated stacks and helper functions.

Signed-off-by: Chunlin Han <chunlin.han@linaro.org>
Signed-off-by: Andy Gross <andy.gross@linaro.org>

CMakeLists.txt

@@ -520,6 +520,7 @@ if(CONFIG_CPU_HAS_MPU AND CONFIG_USERSPACE)
   if(CONFIG_MPU_REQUIRES_POWER_OF_TWO_ALIGNMENT)
     set(ALIGN_SIZING_DEP app_sizing_prebuilt linker_app_sizing_script)
   endif()
+  set(PRIV_STACK_DEP priv_stacks_prebuilt)
 endif()
 
 function(construct_add_custom_command_for_linker_pass linker_output_name output_variable)
@@ -578,7 +579,7 @@ add_custom_command(
 add_custom_target(
   linker_script
   DEPENDS
-  ${ALIGN_SIZING_DEP}
+  ${ALIGN_SIZING_DEP} ${PRIV_STACK_DEP}
   linker.cmd
   offsets_h
   )
@@ -638,6 +639,119 @@ if(CONFIG_GEN_ISR_TABLES)
   set_property(GLOBAL APPEND PROPERTY GENERATED_KERNEL_SOURCE_FILES isr_tables.c)
 endif()
 
+if(CONFIG_ARM AND CONFIG_USERSPACE)
+  set(GEN_PRIV_STACKS $ENV{ZEPHYR_BASE}/scripts/gen_priv_stacks.py)
+  set(PROCESS_PRIV_STACKS_GPERF $ENV{ZEPHYR_BASE}/scripts/process_gperf.py)
+
+  set(PRIV_STACKS                    priv_stacks_hash.gperf)
+  set(PRIV_STACKS_OUTPUT_SRC_PRE     priv_stacks_hash_preprocessed.c)
+  set(PRIV_STACKS_OUTPUT_SRC         priv_stacks_hash.c)
+  set(PRIV_STACKS_OUTPUT_OBJ         priv_stacks_hash.c.obj)
+  set(PRIV_STACKS_OUTPUT_OBJ_RENAMED priv_stacks_hash_renamed.o)
+
+  # Essentially what we are doing here is extracting some information
+  # out of the nearly finished elf file, generating the source code
+  # for a hash table based on that information, and then compiling and
+  # linking the hash table back into a now even more nearly finished
+  # elf file.
+
+  # Use the script GEN_PRIV_STACKS to scan the kernel binary's
+  # (zephyr_prebuilt) DWARF information to produce a table of kernel
+  # objects (PRIV_STACKS) which we will then pass to gperf
+  add_custom_command(
+    OUTPUT ${PRIV_STACKS}
+    COMMAND
+    ${PYTHON_EXECUTABLE}
+    ${GEN_PRIV_STACKS}
+    --kernel $<TARGET_FILE:priv_stacks_prebuilt>
+    --output ${PRIV_STACKS}
+    $<$<BOOL:${CMAKE_VERBOSE_MAKEFILE}>:--verbose>
+    DEPENDS priv_stacks_prebuilt
+    WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
+    )
+  add_custom_target(priv_stacks DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${PRIV_STACKS})
+
+  # Use gperf to generate C code (PRIV_STACKS_OUTPUT_SRC_PRE) which implements a
+  # perfect hashtable based on PRIV_STACKS
+  add_custom_command(
+    OUTPUT ${PRIV_STACKS_OUTPUT_SRC_PRE}
+    COMMAND
+    ${GPERF} -C
+    --output-file ${PRIV_STACKS_OUTPUT_SRC_PRE}
+    ${PRIV_STACKS}
+    DEPENDS priv_stacks
+    WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
+    )
+  add_custom_target(priv_stacks_output_src_pre DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${PRIV_STACKS_OUTPUT_SRC_PRE})
+
+  # For our purposes the code/data generated by gperf is not optimal.
+  #
+  # The script PROCESS_GPERF creates a new c file OUTPUT_SRC based on
+  # OUTPUT_SRC_PRE to greatly reduce the amount of code/data generated
+  # since we know we are always working with pointer values
+  add_custom_command(
+    OUTPUT ${PRIV_STACKS_OUTPUT_SRC}
+    COMMAND
+    ${PROCESS_PRIV_STACKS_GPERF}
+    -i ${PRIV_STACKS_OUTPUT_SRC_PRE}
+    -o ${PRIV_STACKS_OUTPUT_SRC}
+    -p "struct _k_priv_stack_map"
+    $<$<BOOL:${CMAKE_VERBOSE_MAKEFILE}>:--verbose>
+    DEPENDS priv_stacks_output_src_pre
+    WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
+    )
+  add_custom_target(priv_stacks_output_src DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${PRIV_STACKS_OUTPUT_SRC})
+
+  # We need precise control of where generated text/data ends up in the final
+  # kernel image. Disable function/data sections and use objcopy to move
+  # generated data into special section names
+  add_library(priv_stacks_output_lib STATIC
+    ${CMAKE_CURRENT_BINARY_DIR}/${PRIV_STACKS_OUTPUT_SRC}
+    )
+
+  target_link_libraries(priv_stacks_output_lib zephyr_interface)
+
+  # Turn off -ffunction-sections, etc.
+  # NB: Using a library instead of target_compile_options(priv_stacks_output_lib
+  # [...]) because a library's options have precedence
+  add_library(priv_stacks_output_lib_interface INTERFACE)
+  target_compile_options(priv_stacks_output_lib_interface INTERFACE
+    -fno-function-sections
+    -fno-data-sections
+    )
+  target_link_libraries(priv_stacks_output_lib priv_stacks_output_lib_interface)
+
+  set(PRIV_STACKS_OUTPUT_OBJ_PATH ${CMAKE_CURRENT_BINARY_DIR}/CMakeFiles/priv_stacks_output_lib.dir/${PRIV_STACKS_OUTPUT_OBJ})
+
+  add_custom_command(
+    OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/${PRIV_STACKS_OUTPUT_OBJ_RENAMED}
+    COMMAND
+    ${CMAKE_OBJCOPY}
+    --rename-section   .bss=.priv_stacks.noinit
+    --rename-section   .data=.priv_stacks.data
+    --rename-section   .text=.priv_stacks.text
+    --rename-section .rodata=.priv_stacks.rodata
+    ${PRIV_STACKS_OUTPUT_OBJ_PATH}
+    ${PRIV_STACKS_OUTPUT_OBJ_RENAMED}
+    DEPENDS priv_stacks_output_lib
+    WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
+    )
+  add_custom_target(priv_stacks_output_obj_renamed DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${PRIV_STACKS_OUTPUT_OBJ_RENAMED})
+
+  add_library(priv_stacks_output_obj_renamed_lib STATIC IMPORTED GLOBAL)
+  set_property(
+    TARGET priv_stacks_output_obj_renamed_lib
+    PROPERTY
+    IMPORTED_LOCATION ${CMAKE_CURRENT_BINARY_DIR}/${PRIV_STACKS_OUTPUT_OBJ_RENAMED}
+    )
+  add_dependencies(
+    priv_stacks_output_obj_renamed_lib
+    priv_stacks_output_obj_renamed
+    )
+
+  set_property(GLOBAL APPEND PROPERTY GENERATED_KERNEL_OBJECT_FILES priv_stacks_output_obj_renamed_lib)
+endif()
+
 if(CONFIG_USERSPACE)
   set(GEN_KOBJ_LIST ${ZEPHYR_BASE}/scripts/gen_kobject_list.py)
   set(PROCESS_GPERF ${ZEPHYR_BASE}/scripts/process_gperf.py)
@@ -694,6 +808,7 @@ if(CONFIG_USERSPACE)
     ${PROCESS_GPERF}
     -i ${OUTPUT_SRC_PRE}
     -o ${OUTPUT_SRC}
+    -p "struct _k_object"
     $<$<BOOL:${CMAKE_VERBOSE_MAKEFILE}>:--verbose>
     DEPENDS output_src_pre ${OUTPUT_SRC_PRE}
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
@@ -804,13 +919,39 @@ if(CONFIG_CPU_HAS_MPU AND CONFIG_USERSPACE)
       WORKING_DIRECTORY ${PROJECT_BINARY_DIR}/
    )
   endif()
+
+  construct_add_custom_command_for_linker_pass(linker_priv_stacks custom_command)
+  add_custom_command(
+    ${custom_command}
+    )
+
+  add_custom_target(
+    linker_priv_stacks_script
+    DEPENDS
+    ${ALIGN_SIZING_DEP}
+    linker_priv_stacks.cmd
+    offsets_h
+    )
+
+  set_property(TARGET
+    linker_priv_stacks_script
+    PROPERTY INCLUDE_DIRECTORIES
+    ${ZEPHYR_INCLUDE_DIRS}
+    )
+
+  set(PRIV_STACK_LIB priv_stacks_output_obj_renamed_lib)
+  add_executable(       priv_stacks_prebuilt misc/empty_file.c)
+  target_link_libraries(priv_stacks_prebuilt ${TOPT} ${PROJECT_BINARY_DIR}/linker_priv_stacks.cmd ${zephyr_lnk})
+  set_property(TARGET   priv_stacks_prebuilt PROPERTY LINK_DEPENDS ${PROJECT_BINARY_DIR}/linker_priv_stacks.cmd)
+  add_dependencies(     priv_stacks_prebuilt ${ALIGN_SIZING_DEP} linker_priv_stacks_script offsets)
+
 endif()
 
 # FIXME: Is there any way to get rid of empty_file.c?
 add_executable(       zephyr_prebuilt misc/empty_file.c)
-target_link_libraries(zephyr_prebuilt ${TOPT} ${PROJECT_BINARY_DIR}/linker.cmd ${zephyr_lnk})
+target_link_libraries(zephyr_prebuilt ${TOPT} ${PROJECT_BINARY_DIR}/linker.cmd ${PRIV_STACK_LIB} ${zephyr_lnk})
 set_property(TARGET   zephyr_prebuilt PROPERTY LINK_DEPENDS ${PROJECT_BINARY_DIR}/linker.cmd)
-add_dependencies(     zephyr_prebuilt ${ALIGN_SIZING_DEP} linker_script offsets)
+add_dependencies(     zephyr_prebuilt ${ALIGN_SIZING_DEP} ${PRIV_STACK_DEP} linker_script offsets)
 
 
 if(NOT CONFIG_NATIVE_APPLICATION)
@@ -831,7 +972,7 @@ if(GKOF OR GKSF)
   add_custom_target(
     linker_pass_final_script
     DEPENDS
-    ${ALIGN_SIZING_DEP}
+    ${ALIGN_SIZING_DEP} ${PRIV_STACK_DEP}
     zephyr_prebuilt
     linker_pass_final.cmd
     offsets_h
@@ -845,7 +986,7 @@ if(GKOF OR GKSF)
   add_executable(       kernel_elf misc/empty_file.c ${GKSF})
   target_link_libraries(kernel_elf ${GKOF} ${TOPT} ${PROJECT_BINARY_DIR}/linker_pass_final.cmd ${zephyr_lnk})
   set_property(TARGET   kernel_elf PROPERTY LINK_DEPENDS ${PROJECT_BINARY_DIR}/linker_pass_final.cmd)
-  add_dependencies(     kernel_elf ${ALIGN_SIZING_DEP} linker_pass_final_script)
+  add_dependencies(     kernel_elf ${ALIGN_SIZING_DEP} ${PRIV_STACK_DEP} linker_pass_final_script)
 else()
   set(logical_target_for_zephyr_elf zephyr_prebuilt)
   # Use the prebuilt elf as the final elf since we don't have a

arch/Kconfig

@@ -78,6 +78,16 @@ config USERSPACE
 	  This feature is under heavy development and APIs related to it are
 	  subject to change, even if declared non-private.
 
+config PRIVILEGED_STACK_SIZE
+	int "Size of privileged stack"
+	default 256
+	depends on ARCH_HAS_USERSPACE
+	help
+	This option sets the priviliged stack region size that will be used
+	in addition to the user mode thread stack.  During normal execution,
+	this region will be inaccessible from user mode.  During system calls,
+	this region will be utilized by the system call.
+
 config MAX_THREAD_BYTES
 	int "Bytes to use when tracking object thread permissions"
 	default 2

arch/arm/core/offsets/offsets.c

@@ -29,6 +29,10 @@
 GEN_OFFSET_SYM(_thread_arch_t, basepri);
 GEN_OFFSET_SYM(_thread_arch_t, swap_return_value);
 
+#ifdef CONFIG_USERSPACE
+GEN_OFFSET_SYM(_thread_arch_t, priv_stack_start);
+#endif
+
 #ifdef CONFIG_FLOAT
 GEN_OFFSET_SYM(_thread_arch_t, preempt_float);
 #endif

arch/arm/include/offsets_short_arch.h

@@ -26,6 +26,11 @@
 #define _thread_offset_to_preempt_float \
 	(___thread_t_arch_OFFSET + ___thread_arch_t_preempt_float_OFFSET)
 
+#ifdef CONFIG_USERSPACE
+#define _thread_offset_to_priv_stack_start \
+	(___thread_t_arch_OFFSET + ___thread_arch_t_priv_stack_start_OFFSET)
+#endif
+
 /* end - threads */
 
 #endif /* _offsets_short_arch__h_ */

include/arch/arm/cortex_m/scripts/linker.ld

@@ -126,6 +126,7 @@ SECTIONS
 	*(".text.*")
 	*(.gnu.linkonce.t.*)
 
+#include <linker/priv_stacks-text.ld>
 #include <linker/kobject-text.ld>
 
 	} GROUP_LINK_IN(ROMABLE_REGION)
@@ -162,6 +163,7 @@ SECTIONS
 #include <custom-rodata.ld>
 #endif
 
+#include <linker/priv_stacks-rom.ld>
 #include <linker/kobject-rom.ld>
 
 	/*
@@ -310,6 +312,7 @@ SECTIONS
     __data_rom_start = LOADADDR(_DATA_SECTION_NAME);
 
 #include <linker/common-ram.ld>
+#include <linker/priv_stacks.ld>
 #include <linker/kobject.ld>
 
     __data_ram_end = .;

include/linker/priv_stacks-rom.ld

@@ -0,0 +1,13 @@
+/*
+ * Copyright (c) 2017 Linaro Limited.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#ifdef CONFIG_USERSPACE
+        /* Kept in RAM on non-XIP */
+#ifdef CONFIG_XIP
+	*(".priv_stacks.rodata*")
+#endif
+#endif /* CONFIG_USERSPACE */
+

include/linker/priv_stacks-text.ld

@@ -0,0 +1,30 @@
+/*
+ * Copyright (c) 2017 Linaro Limited.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+#ifndef PRIV_STACKS_TEXT_AREA
+#if defined(CONFIG_DEBUG) || defined(CONFIG_STACK_CANARIES)
+#define PRIV_STACKS_TEXT_AREA	256
+#else
+#define PRIV_STACKS_TEXT_AREA	128
+#endif
+#endif
+
+#ifdef CONFIG_USERSPACE
+	/* We need to reserve room for the gperf generated hash functions.
+	 * Fortunately, unlike the data tables, the size of the code is
+	 * reasonably predictable.
+	 *
+	 * The linker will error out complaining that the location pointer
+	 * is moving backwards if the reserved room isn't large enough.
+	 */
+	_priv_stacks_text_area_start = .;
+	*(".priv_stacks.text*")
+	_priv_stacks_text_area_end = .;
+#ifndef LINKER_PASS2
+	PROVIDE(_k_priv_stack_find = .);
+#endif
+	. += PRIV_STACKS_TEXT_AREA - (_priv_stacks_text_area_end - _priv_stacks_text_area_start);
+#endif /* CONFIG_USERSPACE */
+

include/linker/priv_stacks.ld

@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2017 Linaro Limited
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#ifdef CONFIG_USERSPACE
+	/* Constraints:
+         *
+         * - changes to the size of this section between build phases
+         *   *must not* shift the memory address of any kernel obejcts,
+         *   since it contains a hashtable of the memory addresses of those
+         *   kernel objects
+         *
+         * - It is OK if this section itself is shifted in between builds; for
+         *   example some arches may precede this section with generated MMU
+         *    page tables which are also unpredictable in size.
+         *
+         * The size of the
+	 * gperf tables is both a function of the number of kernel objects,
+	 * *and* the specific memory addresses being hashed. It is not something
+	 * that can be predicted without actually building and compling it.
+         */
+	SECTION_DATA_PROLOGUE(priv_stacks, (OPTIONAL),)
+	{
+	*(".priv_stacks.data*")
+
+        /* This is also unpredictable in size, and has the same constraints.
+         * On XIP systems this will get put at the very end of ROM.
+         */
+#ifndef CONFIG_XIP
+	*(".priv_stacks.rodata*")
+#endif
+        } GROUP_DATA_LINK_IN(RAMABLE_REGION, ROMABLE_REGION)
+#endif /* CONFIG_USERSPACE */
+

scripts/gen_priv_stacks.py

@@ -0,0 +1,585 @@
+#!/usr/bin/env python3
+#
+# Copyright (c) 2017 Linaro Limited
+#
+# SPDX-License-Identifier: Apache-2.0
+
+import sys
+import argparse
+import pprint
+import os
+import struct
+from distutils.version import LooseVersion
+
+import elftools
+from elftools.elf.elffile import ELFFile
+from elftools.dwarf import descriptions
+from elftools.elf.sections import SymbolTableSection
+
+if LooseVersion(elftools.__version__) < LooseVersion('0.24'):
+    sys.stderr.write("pyelftools is out of date, need version 0.24 or later\n")
+    sys.exit(1)
+
+stack_objects = [
+        "k_stack",
+        "_k_thread_stack_element",
+        ]
+
+
+def subsystem_to_enum(subsys):
+    return "K_OBJ_DRIVER_" + subsys[:-11].upper()
+
+def kobject_to_enum(ko):
+    return "K_OBJ_" + ko[2:].upper()
+
+DW_OP_addr = 0x3
+DW_OP_fbreg = 0x91
+STACK_TYPE = "_k_thread_stack_element"
+thread_counter = 0
+
+# Global type environment. Populated by pass 1.
+type_env = {}
+
+# --- debug stuff ---
+
+scr = os.path.basename(sys.argv[0])
+
+def debug(text):
+    if not args.verbose:
+        return
+    sys.stdout.write(scr + ": " + text + "\n")
+
+def error(text):
+    sys.stderr.write("%s ERROR: %s\n" % (scr, text))
+    sys.exit(1)
+
+def debug_die(die, text):
+    fn, ln = get_filename_lineno(die)
+
+    debug(str(die))
+    debug("File '%s', line %d:" % (fn, ln))
+    debug("    %s" % text)
+
+# --- type classes ----
+
+class KobjectInstance:
+    def __init__(self, type_obj, addr):
+        global thread_counter
+
+        self.addr = addr
+        self.type_obj = type_obj
+
+        # Type name determined later since drivers needs to look at the
+        # API struct address
+        self.type_name = None
+
+        if self.type_obj.name == "k_thread":
+            # Assign an ID for this thread object, used to track its
+            # permissions to other kernel objects
+            self.data = thread_counter
+            thread_counter = thread_counter + 1
+        else:
+            self.data = 0
+
+
+class KobjectType:
+    def __init__(self, offset, name, size, api=False):
+        self.name = name
+        self.size = size
+        self.offset = offset
+        self.api = api
+
+    def __repr__(self):
+        return "<kobject %s>" % self.name
+
+    def has_kobject(self):
+        return True
+
+    def get_kobjects(self, addr):
+        return {addr: KobjectInstance(self, addr)}
+
+
+class ArrayType:
+    def __init__(self, offset, elements, member_type):
+        self.elements = elements
+        self.member_type = member_type
+        self.offset = offset
+
+    def __repr__(self):
+        return "<array of %d, size %d>" % (self.member_type, self.num_members)
+
+    def has_kobject(self):
+        if self.member_type not in type_env:
+            return False
+
+        return type_env[self.member_type].has_kobject()
+
+    def get_kobjects(self, addr):
+        mt = type_env[self.member_type]
+
+        # Stacks are arrays of _k_stack_element_t but we want to treat
+        # the whole array as one kernel object (a thread stack)
+        # Data value gets set to size of entire region
+        if isinstance(mt, KobjectType) and mt.name == STACK_TYPE:
+            # An array of stacks appears as a multi-dimensional array.
+            # The last size is the size of each stack. We need to track
+            # each stack within the array, not as one huge stack object.
+            *dimensions, stacksize = self.elements
+            num_members = 1
+            for e in dimensions:
+                num_members = num_members * e
+
+            ret = {}
+            for i in range(num_members):
+                a = addr + (i * stacksize)
+                o = mt.get_kobjects(a)
+                o[a].data = stacksize
+                ret.update(o)
+            return ret
+
+        objs = {}
+
+        # Multidimensional array flattened out
+        num_members = 1
+        for e in self.elements:
+            num_members = num_members * e
+
+        for i in range(num_members):
+            objs.update(mt.get_kobjects(addr + (i * mt.size)))
+        return objs
+
+
+class AggregateTypeMember:
+    def __init__(self, offset, member_name, member_type, member_offset):
+        self.member_name = member_name
+        self.member_type = member_type
+        self.member_offset = member_offset
+
+    def __repr__(self):
+        return "<member %s, type %d, offset %d>" % (self.member_name,
+                self.member_type, self.member_offset)
+
+    def has_kobject(self):
+        if self.member_type not in type_env:
+            return False
+
+        return type_env[self.member_type].has_kobject()
+
+    def get_kobjects(self, addr):
+        mt = type_env[self.member_type]
+        return mt.get_kobjects(addr + self.member_offset)
+
+
+class ConstType:
+    def __init__(self, child_type):
+        self.child_type = child_type
+
+    def __repr__(self):
+        return "<const %d>" % self.child_type
+
+    def has_kobject(self):
+        if self.child_type not in type_env:
+            return False
+
+        return type_env[self.child_type].has_kobject()
+
+    def get_kobjects(self, addr):
+        return type_env[self.child_type].get_kobjects(addr)
+
+
+class AggregateType:
+    def __init__(self, offset, name, size):
+        self.name = name
+        self.size = size
+        self.offset = offset
+        self.members = []
+
+    def add_member(self, member):
+        self.members.append(member)
+
+    def __repr__(self):
+        return "<struct %s, with %s>" % (self.name, self.members)
+
+    def has_kobject(self):
+        result = False
+
+        bad_members = []
+
+        for member in self.members:
+            if member.has_kobject():
+                result = True
+            else:
+                bad_members.append(member)
+                # Don't need to consider this again, just remove it
+
+        for bad_member in bad_members:
+            self.members.remove(bad_member)
+
+        return result
+
+    def get_kobjects(self, addr):
+        objs = {}
+        for member in self.members:
+            objs.update(member.get_kobjects(addr))
+        return objs
+
+
+# --- helper functions for getting data from DIEs ---
+
+def die_get_name(die):
+    if not 'DW_AT_name' in die.attributes:
+        return None
+    return die.attributes["DW_AT_name"].value.decode("utf-8")
+
+
+def die_get_type_offset(die):
+    if not 'DW_AT_type' in die.attributes:
+        return 0
+
+    return die.attributes["DW_AT_type"].value + die.cu.cu_offset
+
+
+def die_get_byte_size(die):
+    if not 'DW_AT_byte_size' in die.attributes:
+        return 0
+
+    return die.attributes["DW_AT_byte_size"].value
+
+def analyze_die_struct(die):
+    name = die_get_name(die) or "<anon>"
+    offset = die.offset
+    size = die_get_byte_size(die)
+
+    # Incomplete type
+    if not size:
+        return
+
+    if name in stack_objects:
+        type_env[offset] = KobjectType(offset, name, size)
+    else:
+        at = AggregateType(offset, name, size)
+        type_env[offset] = at
+
+        for child in die.iter_children():
+            if child.tag != "DW_TAG_member":
+                continue
+            child_type = die_get_type_offset(child)
+            member_offset = child.attributes["DW_AT_data_member_location"].value
+            cname = die_get_name(child) or "<anon>"
+            m = AggregateTypeMember(child.offset, cname, child_type,
+                    member_offset)
+            at.add_member(m)
+
+        return
+
+
+def analyze_die_const(die):
+    type_offset = die_get_type_offset(die)
+    if not type_offset:
+        return
+
+    type_env[die.offset] = ConstType(type_offset)
+
+
+def analyze_die_array(die):
+    type_offset = die_get_type_offset(die)
+    elements = []
+
+    for child in die.iter_children():
+        if child.tag != "DW_TAG_subrange_type":
+            continue
+        if "DW_AT_upper_bound" not in child.attributes:
+            continue
+
+        ub = child.attributes["DW_AT_upper_bound"]
+        if not ub.form.startswith("DW_FORM_data"):
+            continue
+
+        elements.append(ub.value + 1)
+
+    if not elements:
+        return
+
+    type_env[die.offset] = ArrayType(die.offset, elements, type_offset)
+
+
+def addr_deref(elf, addr):
+    for section in elf.iter_sections():
+        start = section['sh_addr']
+        end = start + section['sh_size']
+
+        if addr >= start and addr < end:
+            data = section.data()
+            offset = addr - start
+            return struct.unpack("<I" if args.little_endian else ">I",
+                    data[offset:offset+4])[0]
+
+    return 0
+
+
+def device_get_api_addr(elf, addr):
+    return addr_deref(elf, addr + 4)
+
+
+def get_filename_lineno(die):
+    lp_header = die.dwarfinfo.line_program_for_CU(die.cu).header
+    files = lp_header["file_entry"]
+    includes = lp_header["include_directory"]
+
+    fileinfo = files[die.attributes["DW_AT_decl_file"].value - 1]
+    filename = fileinfo.name.decode("utf-8")
+    filedir = includes[fileinfo.dir_index - 1].decode("utf-8")
+
+    path = os.path.join(filedir, filename)
+    lineno = die.attributes["DW_AT_decl_line"].value
+    return (path, lineno)
+
+
+def find_stack_objects(elf, syms):
+    if not elf.has_dwarf_info():
+        sys.stderr.write("ELF file has no DWARF information\n");
+        sys.exit(1)
+
+    kram_start = syms["__kernel_ram_start"]
+    kram_end = syms["__kernel_ram_end"]
+    krom_start = syms["_image_rom_start"]
+    krom_end = syms["_image_rom_end"]
+
+    di = elf.get_dwarf_info()
+
+    variables = []
+
+    # Step 1: collect all type information.
+    for CU in di.iter_CUs():
+        CU_path = CU.get_top_DIE().get_full_path()
+        lp = di.line_program_for_CU(CU)
+
+        for idx, die in enumerate(CU.iter_DIEs()):
+            # Unions are disregarded, kernel objects should never be union
+            # members since the memory is not dedicated to that object and
+            # could be something else
+            if die.tag == "DW_TAG_structure_type":
+                analyze_die_struct(die)
+            elif die.tag == "DW_TAG_const_type":
+                analyze_die_const(die)
+            elif die.tag == "DW_TAG_array_type":
+                analyze_die_array(die)
+            elif die.tag == "DW_TAG_variable":
+                variables.append(die)
+
+    # Step 2: filter type_env to only contain kernel objects, or structs and
+    # arrays of kernel objects
+    bad_offsets = []
+    for offset, type_object in type_env.items():
+        if not type_object.has_kobject():
+            bad_offsets.append(offset)
+
+    for offset in bad_offsets:
+        del type_env[offset]
+
+    # Step 3: Now that we know all the types we are looking for, examine
+    # all variables
+    all_objs = {}
+
+    # Gross hack, see below
+    work_q_found = False
+
+    for die in variables:
+        name = die_get_name(die)
+        if not name:
+            continue
+
+        type_offset = die_get_type_offset(die)
+
+        # Is this a kernel object, or a structure containing kernel objects?
+        if type_offset not in type_env:
+            continue
+
+        if "DW_AT_declaration" in die.attributes:
+            # FIXME: why does k_sys_work_q not resolve an address in the DWARF
+            # data??? Every single instance it finds is an extern definition
+            # but not the actual instance in system_work_q.c
+            # Is there something weird about how lib-y stuff is linked?
+            if name == "k_sys_work_q" and not work_q_found and name in syms:
+                addr = syms[name]
+                work_q_found = True
+            else:
+                continue
+        else:
+            if "DW_AT_location" not in die.attributes:
+                debug_die(die, "No location information for object '%s'; possibly stack allocated"
+                        % name)
+                continue
+
+            loc = die.attributes["DW_AT_location"]
+            if loc.form != "DW_FORM_exprloc":
+                debug_die(die, "kernel object '%s' unexpected location format" % name)
+                continue
+
+            opcode = loc.value[0]
+            if opcode != DW_OP_addr:
+
+                # Check if frame pointer offset DW_OP_fbreg
+                if opcode == DW_OP_fbreg:
+                    debug_die(die, "kernel object '%s' found on stack" % name)
+                else:
+                    debug_die(die, "kernel object '%s' unexpected exprloc opcode %s"
+                            % (name, hex(opcode)))
+                continue
+
+            addr = (loc.value[1] | (loc.value[2] << 8) | (loc.value[3] << 16) |
+                    (loc.value[4] << 24))
+
+        if addr == 0:
+            # Never linked; gc-sections deleted it
+            continue
+
+        if ((addr < kram_start or addr >= kram_end)
+                and (addr < krom_start or addr >= krom_end)):
+
+            debug_die(die, "object '%s' found in invalid location %s" %
+                    (name, hex(addr)));
+            continue
+
+        type_obj = type_env[type_offset]
+        objs = type_obj.get_kobjects(addr)
+        all_objs.update(objs)
+
+        debug("symbol '%s' at %s contains %d stack object(s)" % (name, hex(addr),
+              len(objs)))
+
+    # Step 4: objs is a dictionary mapping variable memory addresses to their
+    # associated type objects. Now that we have seen all variables and can
+    # properly look up API structs, convert this into a dictionary mapping
+    # variables to the C enumeration of what kernel object type it is.
+    ret = {}
+    for addr, ko in all_objs.items():
+        # API structs don't get into the gperf table
+        if ko.type_obj.api:
+            continue
+
+        if ko.type_obj.name != "device":
+            # Not a device struct so we immediately know its type
+            ko.type_name = kobject_to_enum(ko.type_obj.name)
+            ret[addr] = ko
+            continue
+
+        # Device struct. Need to get the address of its API struct, if it has
+        # one.
+        apiaddr = device_get_api_addr(elf, addr)
+        if apiaddr not in all_objs:
+            # API struct does not correspond to a known subsystem, skip it
+            continue
+
+        apiobj = all_objs[apiaddr]
+        ko.type_name = subsystem_to_enum(apiobj.type_obj.name)
+        ret[addr] = ko
+
+    debug("found %d stack object instances total" % len(ret))
+    return ret
+
+
+header = """%compare-lengths
+%define lookup-function-name _k_priv_stack_map_lookup
+%language=ANSI-C
+%global-table
+%struct-type
+"""
+
+
+priv_stack_decl_temp = "static u8_t __used __aligned(CONFIG_PRIVILEGED_STACK_SIZE) priv_stack_%x[CONFIG_PRIVILEGED_STACK_SIZE];\n"
+priv_stack_decl_size = "CONFIG_PRIVILEGED_STACK_SIZE"
+
+
+includes = """#include <kernel.h>
+#include <string.h>
+"""
+
+
+structure = """struct _k_priv_stack_map {
+    char *name;
+    u8_t *priv_stack_addr;
+};
+%%
+"""
+
+
+# Different versions of gperf have different prototypes for the lookup function,
+# best to implement the wrapper here. The pointer value itself is turned into
+# a string, we told gperf to expect binary strings that are not NULL-terminated.
+footer = """%%
+u8_t *_k_priv_stack_find(void *obj)
+{
+    const struct _k_priv_stack_map *map =
+        _k_priv_stack_map_lookup((const char *)obj, sizeof(void *));
+    return map->priv_stack_addr;
+}
+"""
+
+
+
+def write_gperf_table(fp, objs, static_begin, static_end):
+    fp.write(header)
+
+    # priv stack declarations
+    fp.write("%{\n")
+    fp.write(includes)
+    for obj_addr, ko in objs.items():
+        fp.write(priv_stack_decl_temp % (obj_addr))
+    fp.write("%}\n")
+
+    # structure declaration
+    fp.write(structure)
+
+    for obj_addr, ko in objs.items():
+        byte_str = struct.pack("<I" if args.little_endian else ">I", obj_addr)
+        fp.write("\"")
+        for byte in byte_str:
+            val = "\\x%02x" % byte
+            fp.write(val)
+
+        fp.write("\",priv_stack_%x\n" % obj_addr)
+
+    fp.write(footer)
+
+
+def get_symbols(obj):
+    for section in obj.iter_sections():
+        if isinstance(section, SymbolTableSection):
+            return {sym.name: sym.entry.st_value
+                    for sym in section.iter_symbols()}
+
+    raise LookupError("Could not find symbol table")
+
+
+def parse_args():
+    global args
+
+    parser = argparse.ArgumentParser(description = __doc__,
+            formatter_class = argparse.RawDescriptionHelpFormatter)
+
+    parser.add_argument("-k", "--kernel", required=True,
+            help="Input zephyr ELF binary")
+    parser.add_argument("-o", "--output", required=True,
+            help="Output list of kernel object addresses for gperf use")
+    parser.add_argument("-v", "--verbose", action="store_true",
+            help="Print extra debugging information")
+    args = parser.parse_args()
+
+
+def main():
+    parse_args()
+
+    with open(args.kernel, "rb") as fp:
+        elf = ELFFile(fp)
+        args.little_endian = elf.little_endian
+        syms = get_symbols(elf)
+        objs = find_stack_objects(elf, syms)
+
+    with open(args.output, "w") as fp:
+        write_gperf_table(fp, objs, syms["_static_kernel_objects_begin"],
+                syms["_static_kernel_objects_end"])
+
+if __name__ == "__main__":
+    main()
+

scripts/process_gperf.py

@@ -86,7 +86,7 @@ def process_line(line, fp):
 
     # Set the lookup function to static inline so it gets rolled into
     # _k_object_find(), nothing else will use it
-    if re.search("struct _k_object [*]$", line):
+    if re.search(args.pattern + " [*]$", line):
         fp.write("static inline " + line)
         return
 
@@ -142,6 +142,8 @@ def parse_args():
                         help="Input C file from gperf")
     parser.add_argument("-o", "--output", required=True,
                         help="Output C file with processing done")
+    parser.add_argument("-p", "--pattern", required=True,
+            help="Search pattern for objects")
     parser.add_argument("-v", "--verbose", action="store_true",
                         help="Print extra debugging information")
     args = parser.parse_args()