diff mbox series

[v5,10/19] gendwarfksyms: Limit structure expansion

Message ID 20241030170106.1501763-31-samitolvanen@google.com (mailing list archive)
State Handled Elsewhere
Headers show
Series Implement DWARF modversions | expand

Commit Message

Sami Tolvanen Oct. 30, 2024, 5:01 p.m. UTC
Expand each structure type only once per exported symbol. This
is necessary to support self-referential structures, which would
otherwise result in infinite recursion, but is still sufficient for
catching ABI changes.

For pointers, limit structure expansion after the first pointer
in the symbol type. This should be plenty for detecting ABI
differences, but it stops us from pulling in half the kernel for
types that contain pointers to large kernel data structures, like
task_struct, for example.

Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
Acked-by: Neal Gompa <neal@gompa.dev>
---
 scripts/gendwarfksyms/Makefile        |   1 +
 scripts/gendwarfksyms/cache.c         |  44 +++++++++++
 scripts/gendwarfksyms/dwarf.c         | 109 ++++++++++++++++++++++++--
 scripts/gendwarfksyms/gendwarfksyms.h |  37 +++++++++
 4 files changed, 183 insertions(+), 8 deletions(-)
 create mode 100644 scripts/gendwarfksyms/cache.c

Comments

Sami Tolvanen Nov. 20, 2024, 9:54 p.m. UTC | #1
On Wed, Oct 30, 2024 at 05:01:17PM +0000, Sami Tolvanen wrote:
> 
> For pointers, limit structure expansion after the first pointer
> in the symbol type. This should be plenty for detecting ABI
> differences, but it stops us from pulling in half the kernel for
> types that contain pointers to large kernel data structures, like
> task_struct, for example.

We spent some time backtesting gendwarfksyms with android15-6.6
releases to validate the kABI stability options that are currently
implemented [1]. While we discovered a few missing knobs that are
needed, we also realized that pointer expansion limits don't work
when symbols are removed (which is possible without breaking the
existing ABI when TRIM_UNUSED_KSYMS is used), as the processing
cut-off can change. Petr also had some concerns about this earlier,
so instead of coming up with an alternative solution, I'm going to
drop the pointer expansion limit in v6.

I still think that it would be great to come up with a better way to
limit how deep into internal data structures we go when calculating
versions, but I suspect that's going to be challenging unless C
grows an annotation for public / private struct members.

While looking into another version stability issue compared to
genksyms, we also came to the conclusion that private structure
definitions in .c files should be excluded from processing as they
are opaque outside the CU and cannot change the ABI. I'll include
this in v6 as well.

[1] https://github.com/samitolvanen/linux/commits/android15-6.6-2024-11-gendwarfksyms/

Sami
diff mbox series

Patch

diff --git a/scripts/gendwarfksyms/Makefile b/scripts/gendwarfksyms/Makefile
index c0d4ce50fc27..c06145d84df8 100644
--- a/scripts/gendwarfksyms/Makefile
+++ b/scripts/gendwarfksyms/Makefile
@@ -2,6 +2,7 @@ 
 hostprogs-always-y += gendwarfksyms
 
 gendwarfksyms-objs += gendwarfksyms.o
+gendwarfksyms-objs += cache.o
 gendwarfksyms-objs += die.o
 gendwarfksyms-objs += dwarf.o
 gendwarfksyms-objs += symbols.o
diff --git a/scripts/gendwarfksyms/cache.c b/scripts/gendwarfksyms/cache.c
new file mode 100644
index 000000000000..2f1517133a20
--- /dev/null
+++ b/scripts/gendwarfksyms/cache.c
@@ -0,0 +1,44 @@ 
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2024 Google LLC
+ */
+
+#include "gendwarfksyms.h"
+
+struct expanded {
+	uintptr_t addr;
+	struct hlist_node hash;
+};
+
+void __cache_mark_expanded(struct expansion_cache *ec, uintptr_t addr)
+{
+	struct expanded *es;
+
+	es = xmalloc(sizeof(struct expanded));
+	es->addr = addr;
+	hash_add(ec->cache, &es->hash, addr_hash(addr));
+}
+
+bool __cache_was_expanded(struct expansion_cache *ec, uintptr_t addr)
+{
+	struct expanded *es;
+
+	hash_for_each_possible(ec->cache, es, hash, addr_hash(addr)) {
+		if (es->addr == addr)
+			return true;
+	}
+
+	return false;
+}
+
+void cache_clear_expanded(struct expansion_cache *ec)
+{
+	struct hlist_node *tmp;
+	struct expanded *es;
+
+	hash_for_each_safe(ec->cache, es, tmp, hash) {
+		free(es);
+	}
+
+	hash_init(ec->cache);
+}
diff --git a/scripts/gendwarfksyms/dwarf.c b/scripts/gendwarfksyms/dwarf.c
index 98191ca7cef0..cf929d6808be 100644
--- a/scripts/gendwarfksyms/dwarf.c
+++ b/scripts/gendwarfksyms/dwarf.c
@@ -26,6 +26,7 @@  static void process_linebreak(struct die *cache, int n)
 		       !dwarf_form##attr(&da, value);                  \
 	}
 
+DEFINE_GET_ATTR(flag, bool)
 DEFINE_GET_ATTR(udata, Dwarf_Word)
 
 static bool get_ref_die_attr(Dwarf_Die *die, unsigned int id, Dwarf_Die *value)
@@ -79,6 +80,13 @@  static bool match_export_symbol(struct state *state, Dwarf_Die *die)
 	return !!state->sym;
 }
 
+static bool is_declaration(Dwarf_Die *die)
+{
+	bool value;
+
+	return get_flag_attr(die, DW_AT_declaration, &value) && value;
+}
+
 /*
  * Type string processing
  */
@@ -455,19 +463,27 @@  static void __process_structure_type(struct state *state, struct die *cache,
 				     die_callback_t process_func,
 				     die_match_callback_t match_func)
 {
+	bool is_decl;
+
 	process(cache, type);
 	process_fqn(cache, die);
 	process(cache, " {");
 	process_linebreak(cache, 1);
 
-	check(process_die_container(state, cache, die, process_func,
-				    match_func));
+	is_decl = is_declaration(die);
+
+	if (!is_decl && state->expand.expand) {
+		check(process_die_container(state, cache, die, process_func,
+					    match_func));
+	}
 
 	process_linebreak(cache, -1);
 	process(cache, "}");
 
-	process_byte_size_attr(cache, die);
-	process_alignment_attr(cache, die);
+	if (!is_decl && state->expand.expand) {
+		process_byte_size_attr(cache, die);
+		process_alignment_attr(cache, die);
+	}
 }
 
 #define DEFINE_PROCESS_STRUCTURE_TYPE(structure)                        \
@@ -552,6 +568,42 @@  static void process_cached(struct state *state, struct die *cache,
 	}
 }
 
+static void state_init(struct state *state)
+{
+	state->expand.expand = true;
+	state->expand.ptr_depth = 0;
+	state->expand.ptr_expansion_depth = 0;
+	hash_init(state->expansion_cache.cache);
+}
+
+static void expansion_state_restore(struct expansion_state *state,
+				    struct expansion_state *saved)
+{
+	state->expand = saved->expand;
+	state->ptr_depth = saved->ptr_depth;
+	state->ptr_expansion_depth = saved->ptr_expansion_depth;
+}
+
+static void expansion_state_save(struct expansion_state *state,
+				 struct expansion_state *saved)
+{
+	expansion_state_restore(saved, state);
+}
+
+static bool is_pointer_type(int tag)
+{
+	return tag == DW_TAG_pointer_type || tag == DW_TAG_reference_type;
+}
+
+static bool is_expanded_type(int tag)
+{
+	return tag == DW_TAG_class_type || tag == DW_TAG_structure_type ||
+	       tag == DW_TAG_union_type || tag == DW_TAG_enumeration_type;
+}
+
+/* The maximum depth for expanding structures in pointers */
+#define MAX_POINTER_EXPANSION_DEPTH 2
+
 #define PROCESS_TYPE(type)                                \
 	case DW_TAG_##type##_type:                        \
 		process_##type##_type(state, cache, die); \
@@ -559,18 +611,55 @@  static void process_cached(struct state *state, struct die *cache,
 
 static int process_type(struct state *state, struct die *parent, Dwarf_Die *die)
 {
+	enum die_state want_state = DIE_COMPLETE;
 	struct die *cache;
+	struct expansion_state saved;
 	int tag = dwarf_tag(die);
 
+	expansion_state_save(&state->expand, &saved);
+
 	/*
-	 * If we have the DIE already cached, use it instead of walking
+	 * Structures and enumeration types are expanded only once per
+	 * exported symbol. This is sufficient for detecting ABI changes
+	 * within the structure.
+	 *
+	 * We fully expand the first pointer reference in the exported
+	 * symbol, but limit the expansion of further pointer references
+	 * to at most MAX_POINTER_EXPANSION_DEPTH levels.
+	 */
+	if (is_pointer_type(tag))
+		state->expand.ptr_depth++;
+
+	if (is_expanded_type(tag)) {
+		if (cache_was_expanded(&state->expansion_cache, die->addr))
+			state->expand.expand = false;
+
+		if (state->expand.ptr_depth > 0) {
+			if (state->expand.ptr_expansion_depth >=
+			    MAX_POINTER_EXPANSION_DEPTH)
+				state->expand.expand = false;
+
+			if (state->expand.expand)
+				state->expand.ptr_expansion_depth++;
+		}
+
+		if (state->expand.expand)
+			cache_mark_expanded(&state->expansion_cache, die->addr);
+		else
+			want_state = DIE_UNEXPANDED;
+	}
+
+	/*
+	 * If we have want_state already cached, use it instead of walking
 	 * through DWARF.
 	 */
-	cache = die_map_get(die, DIE_COMPLETE);
+	cache = die_map_get(die, want_state);
 
-	if (cache->state == DIE_COMPLETE) {
+	if (cache->state == want_state) {
 		process_cached(state, cache, die);
 		die_map_add_die(parent, cache);
+
+		expansion_state_restore(&state->expand, &saved);
 		return 0;
 	}
 
@@ -611,9 +700,10 @@  static int process_type(struct state *state, struct die *parent, Dwarf_Die *die)
 
 	/* Update cache state and append to the parent (if any) */
 	cache->tag = tag;
-	cache->state = DIE_COMPLETE;
+	cache->state = want_state;
 	die_map_add_die(parent, cache);
 
+	expansion_state_restore(&state->expand, &saved);
 	return 0;
 }
 
@@ -675,11 +765,14 @@  static int process_exported_symbols(struct state *unused, struct die *cache,
 		if (!match_export_symbol(&state, die))
 			return 0;
 
+		state_init(&state);
+
 		if (tag == DW_TAG_subprogram)
 			process_subprogram(&state, &state.die);
 		else
 			process_variable(&state, &state.die);
 
+		cache_clear_expanded(&state.expansion_cache);
 		return 0;
 	}
 	default:
diff --git a/scripts/gendwarfksyms/gendwarfksyms.h b/scripts/gendwarfksyms/gendwarfksyms.h
index 8e087a5d58e2..da28a35d19b2 100644
--- a/scripts/gendwarfksyms/gendwarfksyms.h
+++ b/scripts/gendwarfksyms/gendwarfksyms.h
@@ -105,6 +105,7 @@  void symbol_free(void);
 
 enum die_state {
 	DIE_INCOMPLETE,
+	DIE_UNEXPANDED,
 	DIE_COMPLETE,
 	DIE_LAST = DIE_COMPLETE
 };
@@ -134,6 +135,7 @@  static inline const char *die_state_name(enum die_state state)
 {
 	switch (state) {
 	CASE_CONST_TO_STR(DIE_INCOMPLETE)
+	CASE_CONST_TO_STR(DIE_UNEXPANDED)
 	CASE_CONST_TO_STR(DIE_COMPLETE)
 	}
 
@@ -156,9 +158,40 @@  void die_map_add_linebreak(struct die *pd, int linebreak);
 void die_map_add_die(struct die *pd, struct die *child);
 void die_map_free(void);
 
+/*
+ * cache.c
+ */
+
+#define EXPANSION_CACHE_HASH_BITS 11
+
+/* A cache for addresses we've already seen. */
+struct expansion_cache {
+	HASHTABLE_DECLARE(cache, 1 << EXPANSION_CACHE_HASH_BITS);
+};
+
+void __cache_mark_expanded(struct expansion_cache *ec, uintptr_t addr);
+bool __cache_was_expanded(struct expansion_cache *ec, uintptr_t addr);
+
+static inline void cache_mark_expanded(struct expansion_cache *ec, void *addr)
+{
+	__cache_mark_expanded(ec, (uintptr_t)addr);
+}
+
+static inline bool cache_was_expanded(struct expansion_cache *ec, void *addr)
+{
+	return __cache_was_expanded(ec, (uintptr_t)addr);
+}
+
+void cache_clear_expanded(struct expansion_cache *ec);
+
 /*
  * dwarf.c
  */
+struct expansion_state {
+	bool expand;
+	unsigned int ptr_depth;
+	unsigned int ptr_expansion_depth;
+};
 
 struct state {
 	struct symbol *sym;
@@ -166,6 +199,10 @@  struct state {
 
 	/* List expansion */
 	bool first_list_item;
+
+	/* Structure expansion */
+	struct expansion_state expand;
+	struct expansion_cache expansion_cache;
 };
 
 typedef int (*die_callback_t)(struct state *state, struct die *cache,