Skip to contentSkip to conversation

perf: Fix kwork memory sanitization and libunwind test builds

PATCH Applied Ian Rogers 3 participants 7 messages May 21 – May 22 perf events

lore
Share & tools
lore

Ian Rogers

author reviewer
#0 lore raw

cover perf: Fix kwork memory sanitization and libunwind test builds

The combination of the libunwind refactoring and build system changes exposed a build-test regression for the libunwind build using a cached feature dump. Re-send kwork fixes [1] moving a definition to kwork.h to avoid breaking no libbpf builds.

1. Build Fix for libunwind:
"make feature-dump" runs without LIBUNWIND=1, skipping the setup of libunwind LDFLAGS. When Makefile.feature is included to generate the cached FEATURE-DUMP, the local test-libunwind.bin compiles without arch-specific link flags (-lunwind-x86_64) and fails, caching a false negative ("feature-libunwind=0"). Later builds (such as make_libunwind_O reusing the dump) compile the unwinder based on CONFIG_LIBUNWIND (remote) but disable HAVE_LIBUNWIND_SUPPORT (local), breaking compilation due to missing maps__e_machine declarations in maps.h. Fixed by unconditionally setting up libunwind feature build flags in Makefile.config.

2. kwork ASAN Double-Free & Leaks:
- Solves double-free in the record command argument array due to parse_options mutation, by preserving original pointers in a calloc'ed to_free buffer.
- Ensures kwork_usage string is freed on all exit paths.

3. kwork Unified Memory Lifecycle:
- Establishes a unified memory ownership model for kwork_work by ensuring work names are dynamically duplicated via strdup() and consistently freed using work_exit() and a central perf_kwork__exit() teardown path.

[1] https://lore.kernel.org/linux-perf-users/[email protected]/

Ian Rogers (3):
perf build: Unconditionally set up libunwind feature build flags perf kwork: Fix address sanitizer issues
perf kwork: Fix memory management of kwork_work

 tools/perf/Makefile.config  | 103 ++++++++++++------------
 tools/perf/builtin-kwork.c  | 151 ++++++++++++++++++++++++++++--------
 tools/perf/util/bpf_kwork.c |  14 +++-
 tools/perf/util/kwork.h     |   9 +++
 4 files changed, 188 insertions(+), 89 deletions(-)
Signature · 1 line
2.54.0.746.g67dd491aae-goog
author reviewer
#1 lore raw

PATCH 1/3 perf build: Unconditionally set up libunwind feature build flags

A "make feature-dump" build does not specify LIBUNWIND=1 because it is run with the default configuration to detect system-wide capabilities. This sets NO_LIBUNWIND := 1, causing Makefile.config to skip setting LIBUNWIND_LIBS and FEATURE_CHECK_LDFLAGS-libunwind.

Consequently, when Makefile.feature is included and attempts to run all feature checks (via FEATURE_TESTS := all), the local feature test test-libunwind.bin compiles without the required architecture-specific library flags (-lunwind-x86_64) and fails to link on x86_64. This results in a corrupted cached BUILD_TEST_FEATURE_DUMP showing feature-libunwind=0 even when the host supports it.

Subsequent test builds (like make_libunwind_O in the build-test suite) which reuse the feature dump and specify LIBUNWIND=1 will fail to compile due to a mismatch where CONFIG_LIBUNWIND is set (via remote architecture checks which are self-contained) but HAVE_LIBUNWIND_SUPPORT is disabled, causing compiler errors due to missing maps__e_machine definitions in maps.h.

Fix this by unconditionally setting up the libunwind library lists and feature check LDFLAGS in Makefile.config so they are always populated and available to the feature detection engine regardless of whether LIBUNWIND=1 is opted-in for the current run.

Fixes 444508cd7c7b ("perf build: Be more programmatic when setting up libunwind variables")
---
 tools/perf/Makefile.config | 103 +++++++++++++++++++------------------
 1 file changed, 52 insertions(+), 51 deletions(-)
tools/perf/Makefile.config +52 -51
index b56fa8419f7d..c531b9315609 100644
@@ -79,42 +79,43 @@ ifeq ($(ARCH),s390)
   CFLAGS += -fPIC
 endif
 
+# Unconditionally set up the libunwind feature build flags as a
+# feature-dump build doesn't specify LIBUNWIND=1. This means that
+# dumping the libunwind features will be broken that can impact later
+# builds that use the feature dump.
+ifeq ($(SRCARCH),arm)
+  LIBUNWIND_LIBS = -lunwind -lunwind-arm
+endif
+ifeq ($(SRCARCH),arm64)
+  LIBUNWIND_LIBS = -lunwind -lunwind-aarch64
+endif
+ifeq ($(SRCARCH),loongarch)
+  LIBUNWIND_LIBS = -lunwind -lunwind-loongarch64
+endif
+ifeq ($(ARCH),mips)
+  LIBUNWIND_LIBS = -lunwind -lunwind-mips
+endif
+ifeq ($(SRCARCH),powerpc)
+  LIBUNWIND_LIBS := -lunwind -lunwind-ppc64
+endif
+ifeq ($(SRCARCH),riscv)
+  LIBUNWIND_LIBS := -lunwind -lunwind-riscv
+endif
+ifeq ($(SRCARCH),s390)
+  LIBUNWIND_LIBS := -lunwind -lunwind-s390x
+endif
+ifeq ($(SRCARCH),x86)
+  ifeq (${IS_64_BIT}, 1)
+    LIBUNWIND_LIBS = -lunwind-x86_64 -lunwind -llzma
+  else
+    LIBUNWIND_LIBS = -lunwind-x86 -lunwind -llzma
+  endif
+endif
 ifneq ($(LIBUNWIND),1)
   NO_LIBUNWIND := 1
 endif
-
-ifndef NO_LIBUNWIND
-  ifeq ($(SRCARCH),arm)
-    LIBUNWIND_LIBS = -lunwind -lunwind-arm
-  endif
-  ifeq ($(SRCARCH),arm64)
-    LIBUNWIND_LIBS = -lunwind -lunwind-aarch64
-  endif
-  ifeq ($(SRCARCH),loongarch)
-    LIBUNWIND_LIBS = -lunwind -lunwind-loongarch64
-  endif
-  ifeq ($(ARCH),mips)
-    LIBUNWIND_LIBS = -lunwind -lunwind-mips
-  endif
-  ifeq ($(SRCARCH),powerpc)
-    LIBUNWIND_LIBS := -lunwind -lunwind-ppc64
-  endif
-  ifeq ($(SRCARCH),riscv)
-    LIBUNWIND_LIBS := -lunwind -lunwind-riscv
-  endif
-  ifeq ($(SRCARCH),s390)
-    LIBUNWIND_LIBS := -lunwind -lunwind-s390x
-  endif
-  ifeq ($(SRCARCH),x86)
-    ifeq (${IS_64_BIT}, 1)
-      LIBUNWIND_LIBS = -lunwind-x86_64 -lunwind -llzma
-    else
-      LIBUNWIND_LIBS = -lunwind-x86 -lunwind -llzma
-    endif
-  endif
-  ifeq ($(LIBUNWIND_LIBS),)
-    NO_LIBUNWIND := 1
-  endif
+ifeq ($(LIBUNWIND_LIBS),)
+  NO_LIBUNWIND := 1
 endif
 
 #
@@ -124,24 +125,24 @@ endif
 #
 LIBUNWIND_ARCHS:=aarch64 arm loongarch64 mips ppc32 ppc64 riscv s390x x86 x86_64
 
-ifndef NO_LIBUNWIND
-  FEATURE_CHECK_CFLAGS-libunwind = $(LIBUNWIND_CFLAGS)
-  FEATURE_CHECK_LDFLAGS-libunwind = $(LIBUNWIND_LDFLAGS) $(LIBUNWIND_LIBS)
-  FEATURE_CHECK_CFLAGS-libunwind-debug-frame = $(LIBUNWIND_CFLAGS)
-  FEATURE_CHECK_LDFLAGS-libunwind-debug-frame = $(LIBUNWIND_LDFLAGS) $(LIBUNWIND_LIBS)
-
-  ifdef LIBUNWIND_DIR
-    LIBUNWIND_CFLAGS  = -I$(LIBUNWIND_DIR)/include
-    LIBUNWIND_LDFLAGS = -L$(LIBUNWIND_DIR)/lib
-
-    define libunwind_arch_set_flags
-      FEATURE_CHECK_CFLAGS-libunwind-$(1)  = -I$(LIBUNWIND_DIR)/include
-      FEATURE_CHECK_LDFLAGS-libunwind-$(1) = -L$(LIBUNWIND_DIR)/lib -lunwind -lunwind-$(1)
-    endef
-    $(foreach arch,$(LIBUNWIND_ARCHS), \
-      $(eval $(call libunwind_arch_set_flags,$(arch))) \
-    )
-  endif
+# "Local" (no arch specified) feature test flags.
+FEATURE_CHECK_CFLAGS-libunwind = $(LIBUNWIND_CFLAGS)
+FEATURE_CHECK_LDFLAGS-libunwind = $(LIBUNWIND_LDFLAGS) $(LIBUNWIND_LIBS)
+FEATURE_CHECK_CFLAGS-libunwind-debug-frame = $(LIBUNWIND_CFLAGS)
+FEATURE_CHECK_LDFLAGS-libunwind-debug-frame = $(LIBUNWIND_LDFLAGS) $(LIBUNWIND_LIBS)
+
+# Add directory into the "remote" (build for a a specific arch) feature tests.
+ifdef LIBUNWIND_DIR
+  LIBUNWIND_CFLAGS  = -I$(LIBUNWIND_DIR)/include
+  LIBUNWIND_LDFLAGS = -L$(LIBUNWIND_DIR)/lib
+
+  define libunwind_arch_set_flags
+    FEATURE_CHECK_CFLAGS-libunwind-$(1)  = -I$(LIBUNWIND_DIR)/include
+    FEATURE_CHECK_LDFLAGS-libunwind-$(1) = -L$(LIBUNWIND_DIR)/lib -lunwind -lunwind-$(1)
+  endef
+  $(foreach arch,$(LIBUNWIND_ARCHS), \
+    $(eval $(call libunwind_arch_set_flags,$(arch))) \
+  )
 endif
 
 ifdef CSINCLUDES
Signature · 1 line
2.54.0.746.g67dd491aae-goog
author reviewer
#2 lore raw

PATCH 2/3 perf kwork: Fix address sanitizer issues

In reply to Ian Rogers

There is a double free in the record array due to how parse_options will mutate the array. Fix by keeping an array that isn't mutated. Ensure kwork_usage is freed on all paths.

---
 tools/perf/builtin-kwork.c | 30 +++++++++++++++++++++---------
 1 file changed, 21 insertions(+), 9 deletions(-)
tools/perf/builtin-kwork.c +21 -9
index 59a54d11f7fa..a4604e152002 100644
@@ -2258,7 +2258,7 @@ static void setup_event_list(struct perf_kwork *kwork,
 static int perf_kwork__record(struct perf_kwork *kwork,
 			      int argc, const char **argv)
 {
-	const char **rec_argv;
+	const char **rec_argv, **to_free = NULL;
 	unsigned int rec_argc, i, j;
 	struct kwork_class *class;
 	int ret;
@@ -2295,16 +2295,27 @@ static int perf_kwork__record(struct perf_kwork *kwork,
 
 	BUG_ON(i != rec_argc);
 
+	/* Save the pointers as the array will be mutated by cmd_record. */
+	to_free = calloc(rec_argc + 1, sizeof(char *));
+	if (to_free == NULL) {
+		ret = -ENOMEM;
+		goto EXIT;
+	}
+
 	pr_debug("record comm: ");
-	for (j = 0; j < rec_argc; j++)
+	for (j = 0; j < rec_argc; j++) {
 		pr_debug("%s ", rec_argv[j]);
+		to_free[j] = rec_argv[j];
+	}
 	pr_debug("\n");
 
 	ret = cmd_record(i, rec_argv);
 
 EXIT:
 	for (i = 0; i < rec_argc; i++)
-		free((void *)rec_argv[i]);
+		free((void *)(to_free ? to_free[i] : rec_argv[i]));
+
+	free(to_free);
 	free(rec_argv);
 	return ret;
 }
@@ -2447,6 +2458,7 @@ int cmd_kwork(int argc, const char **argv)
 	const char *const kwork_subcommands[] = {
 		"record", "report", "latency", "timehist", "top", NULL
 	};
+	int ret = 0;
 
 	perf_tool__init(&kwork.tool, /*ordered_events=*/true);
 	kwork.tool.mmap	  = perf_event__process_mmap;
@@ -2463,7 +2475,7 @@ int cmd_kwork(int argc, const char **argv)
 
 	if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
 		setup_event_list(&kwork, kwork_options, kwork_usage);
-		return perf_kwork__record(&kwork, argc, argv);
+		ret = perf_kwork__record(&kwork, argc, argv);
 	} else if (strlen(argv[0]) > 2 && strstarts("report", argv[0])) {
 		kwork.sort_order = default_report_sort_order;
 		if (argc > 1) {
@@ -2474,7 +2486,7 @@ int cmd_kwork(int argc, const char **argv)
 		kwork.report = KWORK_REPORT_RUNTIME;
 		setup_sorting(&kwork, report_options, report_usage);
 		setup_event_list(&kwork, kwork_options, kwork_usage);
-		return perf_kwork__report(&kwork);
+		ret = perf_kwork__report(&kwork);
 	} else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
 		kwork.sort_order = default_latency_sort_order;
 		if (argc > 1) {
@@ -2485,7 +2497,7 @@ int cmd_kwork(int argc, const char **argv)
 		kwork.report = KWORK_REPORT_LATENCY;
 		setup_sorting(&kwork, latency_options, latency_usage);
 		setup_event_list(&kwork, kwork_options, kwork_usage);
-		return perf_kwork__report(&kwork);
+		ret = perf_kwork__report(&kwork);
 	} else if (strlen(argv[0]) > 2 && strstarts("timehist", argv[0])) {
 		if (argc > 1) {
 			argc = parse_options(argc, argv, timehist_options, timehist_usage, 0);
@@ -2494,7 +2506,7 @@ int cmd_kwork(int argc, const char **argv)
 		}
 		kwork.report = KWORK_REPORT_TIMEHIST;
 		setup_event_list(&kwork, kwork_options, kwork_usage);
-		return perf_kwork__timehist(&kwork);
+		ret = perf_kwork__timehist(&kwork);
 	} else if (strlen(argv[0]) > 2 && strstarts("top", argv[0])) {
 		kwork.sort_order = default_top_sort_order;
 		if (argc > 1) {
@@ -2507,12 +2519,12 @@ int cmd_kwork(int argc, const char **argv)
 			kwork.event_list_str = "sched, irq, softirq";
 		setup_event_list(&kwork, kwork_options, kwork_usage);
 		setup_sorting(&kwork, top_options, top_usage);
-		return perf_kwork__top(&kwork);
+		ret = perf_kwork__top(&kwork);
 	} else
 		usage_with_options(kwork_usage, kwork_options);
 
 	/* free usage string allocated by parse_options_subcommand */
 	free((void *)kwork_usage[0]);
 
-	return 0;
+	return ret;
 }
Signature · 1 line
2.54.0.746.g67dd491aae-goog
author reviewer
#3 lore raw

PATCH 3/3 perf kwork: Fix memory management of kwork_work

In reply to Ian Rogers

This commit addresses several memory management issues in builtin-kwork.c:
1. Implements a global cleanup function perf_kwork__exit to free all kwork_work and kwork_atom_page objects at the end of the command.
2. Ensures all 'name' fields in struct kwork_work are malloc-ed (or NULL) and properly freed by using strdup and zfree.
3. Fixes memory leaks in top_merge_tasks where kwork_work objects were dropped without being freed.
4. Adds robustness with NULL checks for name fields.
5. Fixes workqueue_work_init to correctly resolve and strdup kernel function names, preventing bad-free errors.

---
 tools/perf/builtin-kwork.c  | 121 ++++++++++++++++++++++++++++--------
 tools/perf/util/bpf_kwork.c |  14 +++--
 tools/perf/util/kwork.h     |   9 +++
 3 files changed, 115 insertions(+), 29 deletions(-)
tools/perf/builtin-kwork.c +96 -25
index a4604e152002..f793ea578515 100644
@@ -323,8 +323,8 @@ static struct kwork_work *work_search(struct rb_root_cached *root,
 		else if (cmp < 0)
 			node = node->rb_right;
 		else {
-			if (work->name == NULL)
-				work->name = key->name;
+			if (work->name == NULL && key->name != NULL)
+				work->name = strdup(key->name);
 			return work;
 		}
 	}
@@ -371,11 +371,54 @@ static struct kwork_work *work_new(struct kwork_work *key)
 
 	work->id = key->id;
 	work->cpu = key->cpu;
-	work->name = key->name;
+	work->name = key->name ? strdup(key->name) : NULL;
 	work->class = key->class;
 	return work;
 }
 
+
+static void work_delete(struct kwork_work *work)
+{
+	if (work) {
+		work_exit(work);
+		free(work);
+	}
+}
+
+static void kwork_work__free_root(struct rb_root_cached *root)
+{
+	struct rb_node *next;
+	struct kwork_work *work;
+
+	while ((next = rb_first_cached(root))) {
+		work = rb_entry(next, struct kwork_work, node);
+		rb_erase_cached(next, root);
+		work_delete(work);
+	}
+}
+
+static void perf_kwork__exit(struct perf_kwork *kwork)
+{
+	struct kwork_class *class;
+	struct kwork_atom_page *page, *tmp_page;
+
+	list_for_each_entry(class, &kwork->class_list, list) {
+		kwork_work__free_root(&class->work_root);
+	}
+
+	kwork_work__free_root(&kwork->sorted_work_root);
+
+	list_for_each_entry_safe(page, tmp_page, &kwork->atom_page_list, list) {
+		list_del_init(&page->list);
+		free(page);
+	}
+
+	INIT_LIST_HEAD(&kwork->class_list);
+	INIT_LIST_HEAD(&kwork->atom_page_list);
+	INIT_LIST_HEAD(&kwork->sort_list);
+	INIT_LIST_HEAD(&kwork->cmp_id);
+}
+
 static struct kwork_work *work_findnew(struct rb_root_cached *root,
 				       struct kwork_work *key,
 				       struct list_head *sort_list)
@@ -453,25 +496,29 @@ static int work_push_atom(struct perf_kwork *kwork,
 			  struct kwork_work **ret_work,
 			  bool overwrite)
 {
-	struct kwork_atom *atom, *dst_atom, *last_atom;
+	struct kwork_atom *atom = NULL, *dst_atom, *last_atom;
 	struct kwork_work *work, key;
+	int ret = 0;
 
 	BUG_ON(class->work_init == NULL);
 	class->work_init(kwork, class, &key, src_type, sample, machine);
 
 	atom = atom_new(kwork, sample);
-	if (atom == NULL)
-		return -1;
+	if (atom == NULL) {
+		ret = -1;
+		goto out;
+	}
 
 	work = work_findnew(&class->work_root, &key, &kwork->cmp_id);
 	if (work == NULL) {
 		atom_free(atom);
-		return -1;
+		ret = -1;
+		goto out;
 	}
 
 	if (!profile_event_match(kwork, work, sample)) {
 		atom_free(atom);
-		return 0;
+		goto out;
 	}
 
 	if (dst_type < KWORK_TRACE_MAX) {
@@ -498,8 +545,9 @@ static int work_push_atom(struct perf_kwork *kwork,
 	}
 
 	list_add_tail(&atom->list, &work->atom_list[src_type]);
-
-	return 0;
+out:
+	work_exit(&key);
+	return ret;
 }
 
 static struct kwork_atom *work_pop_atom(struct perf_kwork *kwork,
@@ -510,7 +558,7 @@ static struct kwork_atom *work_pop_atom(struct perf_kwork *kwork,
 					struct machine *machine,
 					struct kwork_work **ret_work)
 {
-	struct kwork_atom *atom, *src_atom;
+	struct kwork_atom *atom = NULL, *src_atom;
 	struct kwork_work *work, key;
 
 	BUG_ON(class->work_init == NULL);
@@ -521,15 +569,15 @@ static struct kwork_atom *work_pop_atom(struct perf_kwork *kwork,
 		*ret_work = work;
 
 	if (work == NULL)
-		return NULL;
+		goto out;
 
 	if (!profile_event_match(kwork, work, sample))
-		return NULL;
+		goto out;
 
 	atom = list_last_entry_or_null(&work->atom_list[dst_type],
 				       struct kwork_atom, list);
 	if (atom != NULL)
-		return atom;
+		goto out;
 
 	src_atom = atom_new(kwork, sample);
 	if (src_atom != NULL)
@@ -538,8 +586,9 @@ static struct kwork_atom *work_pop_atom(struct perf_kwork *kwork,
 		if (ret_work != NULL)
 			*ret_work = NULL;
 	}
-
-	return NULL;
+out:
+	work_exit(&key);
+	return atom;
 }
 
 static struct kwork_work *find_work_by_id(struct rb_root_cached *root,
@@ -1002,13 +1051,16 @@ static void irq_work_init(struct perf_kwork *kwork,
 		work->name = NULL;
 	} else {
 		work->id = perf_sample__intval(sample, "irq");
-		work->name = perf_sample__strval(sample, "name");
+		work->name = strdup(perf_sample__strval(sample, "name") ?: "<unknown>");
 	}
 }
 
 static void irq_work_name(struct kwork_work *work, char *buf, int len)
 {
-	snprintf(buf, len, "%s:%" PRIu64 "", work->name, work->id);
+	if (work->name != NULL)
+		snprintf(buf, len, "%s:%" PRIu64 "", work->name, work->id);
+	else
+		snprintf(buf, len, "%" PRIu64 "", work->id);
 }
 
 static struct kwork_class kwork_irq = {
@@ -1135,7 +1187,10 @@ static void softirq_work_init(struct perf_kwork *kwork,
 
 static void softirq_work_name(struct kwork_work *work, char *buf, int len)
 {
-	snprintf(buf, len, "(s)%s:%" PRIu64 "", work->name, work->id);
+	if (work->name != NULL)
+		snprintf(buf, len, "(s)%s:%" PRIu64 "", work->name, work->id);
+	else
+		snprintf(buf, len, "(s)%" PRIu64 "", work->id);
 }
 
 static struct kwork_class kwork_softirq = {
@@ -1220,8 +1275,14 @@ static void workqueue_work_init(struct perf_kwork *kwork __maybe_unused,
 	work->class = class;
 	work->cpu = sample->cpu;
 	work->id = perf_sample__intval(sample, "work");
-	work->name = function_addr == 0 ? NULL :
-		machine__resolve_kernel_addr(machine, &function_addr, &modp);
+	work->name = NULL;
+
+	if (function_addr != 0) {
+		const char *name = machine__resolve_kernel_addr(machine, &function_addr, &modp);
+
+		if (name)
+			work->name = strdup(name);
+	}
 }
 
 static void workqueue_work_name(struct kwork_work *work, char *buf, int len)
@@ -1284,16 +1345,16 @@ static void sched_work_init(struct perf_kwork *kwork __maybe_unused,
 
 	if (src_type == KWORK_TRACE_EXIT) {
 		work->id = perf_sample__intval(sample, "prev_pid");
-		work->name = strdup(perf_sample__strval(sample, "prev_comm"));
+		work->name = strdup(perf_sample__strval(sample, "prev_comm") ?: "<unknown>");
 	} else if (src_type == KWORK_TRACE_ENTRY) {
 		work->id = perf_sample__intval(sample, "next_pid");
-		work->name = strdup(perf_sample__strval(sample, "next_comm"));
+		work->name = strdup(perf_sample__strval(sample, "next_comm") ?: "<unknown>");
 	}
 }
 
 static void sched_work_name(struct kwork_work *work, char *buf, int len)
 {
-	snprintf(buf, len, "%s", work->name);
+	snprintf(buf, len, "%s", work->name ?: "");
 }
 
 static struct kwork_class kwork_sched = {
@@ -2100,8 +2161,10 @@ static void top_merge_tasks(struct perf_kwork *kwork)
 		rb_erase_cached(node, &class->work_root);
 		data = rb_entry(node, struct kwork_work, node);
 
-		if (!profile_name_match(kwork, data))
+		if (!profile_name_match(kwork, data)) {
+			work_delete(data);
 			continue;
+		}
 
 		cpu = data->cpu;
 		merged_work = find_work_by_id(&merged_root, data->id,
@@ -2109,11 +2172,17 @@ static void top_merge_tasks(struct perf_kwork *kwork)
 		if (!merged_work) {
 			work_insert(&merged_root, data, &kwork->cmp_id);
 		} else {
+			if (merged_work->name == NULL && data->name != NULL)
+				merged_work->name = strdup(data->name);
+
 			merged_work->total_runtime += data->total_runtime;
 			merged_work->cpu_usage += data->cpu_usage;
 		}
 
 		top_calc_load_runtime(kwork, data);
+
+		if (merged_work)
+			work_delete(data);
 	}
 
 	work_sort(kwork, class, &merged_root);
@@ -2523,6 +2592,8 @@ int cmd_kwork(int argc, const char **argv)
 	} else
 		usage_with_options(kwork_usage, kwork_options);
 
+	perf_kwork__exit(&kwork);
+
 	/* free usage string allocated by parse_options_subcommand */
 	free((void *)kwork_usage[0]);
 
tools/perf/util/bpf_kwork.c +10 -4
index d3a2e548f2b6..2248f462a847 100644
@@ -273,6 +273,7 @@ static int add_work(struct perf_kwork *kwork,
 		.cpu = key->cpu,
 	};
 	enum kwork_class_type type = key->type;
+	int ret = 0;
 
 	if (!valid_kwork_class_type(type)) {
 		pr_debug("Invalid class type %d to add work\n", type);
@@ -287,8 +288,10 @@ static int add_work(struct perf_kwork *kwork,
 		return -1;
 
 	work = kwork->add_work(kwork, tmp.class, &tmp);
-	if (work == NULL)
-		return -1;
+	if (work == NULL) {
+		ret = -1;
+		goto out;
+	}
 
 	if (kwork->report == KWORK_REPORT_RUNTIME) {
 		work->nr_atoms = data->nr;
@@ -304,13 +307,16 @@ static int add_work(struct perf_kwork *kwork,
 		work->max_latency_end = data->max_time_end;
 	} else {
 		pr_debug("Invalid bpf report type %d\n", kwork->report);
-		return -1;
+		ret = -1;
+		goto out;
 	}
 
 	kwork->timestart = (u64)ts_start.tv_sec * NSEC_PER_SEC + ts_start.tv_nsec;
 	kwork->timeend = (u64)ts_end.tv_sec * NSEC_PER_SEC + ts_end.tv_nsec;
 
-	return 0;
+out:
+	work_exit(&tmp);
+	return ret;
 }
 
 int perf_kwork__report_read_bpf(struct perf_kwork *kwork)
tools/perf/util/kwork.h +9 -0
index abf637d44794..81d39a7f78c8 100644
@@ -5,6 +5,7 @@
 #include "util/tool.h"
 #include "util/time-utils.h"
 
+#include <stdlib.h>
 #include <linux/bitmap.h>
 #include <linux/list.h>
 #include <linux/rbtree.h>
@@ -164,6 +165,14 @@ struct kwork_class {
 			  char *buf, int len);
 };
 
+static inline void work_exit(struct kwork_work *work)
+{
+	if (work) {
+		free(work->name);
+		work->name = NULL;
+	}
+}
+
 struct trace_kwork_handler {
 	int (*raise_event)(struct perf_kwork *kwork,
 			   struct kwork_class *class,
Signature · 1 line
2.54.0.746.g67dd491aae-goog
#4 lore raw

Re: [PATCH v1 3/3] perf kwork: Fix memory management of kwork_work

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [Medium] The top_print_work function passes work->name to printf without verifying it is non-NULL, which can cause a crash. --

commit 03666bc88a2d8e65da244923eb1ef63e0a473153
Author: Ian Rogers <[email protected]>

perf kwork: Fix memory management of kwork_work

This commit implements a global cleanup function to free all kwork_work and kwork_atom_page objects at exit. It also fixes memory leaks during task merging and ensures all name fields in struct kwork_work are properly allocated and freed, while adding robustness checks for NULL names.

Quoted 4 lines
tools/perf/builtin-kwork.c +0 -0
index a4604e1520023..f793ea578515d 100644

[ ... ]

Quoted 7 lines
tools/perf/builtin-kwork.c +1 -1
@@ -1002,13 +1051,16 @@ static void irq_work_init(struct perf_kwork *kwork,
 		work->name = NULL;
 	} else {
 		work->id = perf_sample__intval(sample, "irq");
-		work->name = perf_sample__strval(sample, "name");
+		work->name = strdup(perf_sample__strval(sample, "name") ?: "<unknown>");
 	}

tools/perf/builtin-kwork.c static void irq_work_init(struct perf_kwork *kwork, · line 1064

Quoted 10 lines
tools/perf/builtin-kwork.c +4 -1
 }
 
 static void irq_work_name(struct kwork_work *work, char *buf, int len)
 {
-	snprintf(buf, len, "%s:%" PRIu64 "", work->name, work->id);
+	if (work->name != NULL)
+		snprintf(buf, len, "%s:%" PRIu64 "", work->name, work->id);
+	else
+		snprintf(buf, len, "%" PRIu64 "", work->id);
 }

[Severity: Medium]
This is a pre-existing issue, but while you are adding NULL checks for work->name here, should top_print_work() also be updated?

If strdup() fails due to memory pressure, or if bpf_map_lookup_elem() fails in BPF mode, work->name can be NULL. In top_print_work(), work->name is passed directly to printf() as a string, which can cause a segmentation fault on strict libc implementations:

tools/perf/builtin-kwork.c:top_print_work() {
    ...
	if (kwork->use_bpf)
		ret += printf(" %s%s%s",
			      work->is_kthread ? "[" : "",
			      work->name,
			      work->is_kthread ? "]" : "");
	else
		ret += printf(" %-*s", PRINT_TASK_NAME_WIDTH, work->name);
    ...
}
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=3
author reviewer
#5 lore raw
In reply to Ian Rogers
Quoted 23 lines · Ian Rogers

The combination of the libunwind refactoring and build system changes exposed a build-test regression for the libunwind build using a cached feature dump. Re-send kwork fixes [1] moving a definition to kwork.h to avoid breaking no libbpf builds.

1. Build Fix for libunwind:
"make feature-dump" runs without LIBUNWIND=1, skipping the setup of libunwind LDFLAGS. When Makefile.feature is included to generate the cached FEATURE-DUMP, the local test-libunwind.bin compiles without arch-specific link flags (-lunwind-x86_64) and fails, caching a false negative ("feature-libunwind=0"). Later builds (such as make_libunwind_O reusing the dump) compile the unwinder based on CONFIG_LIBUNWIND (remote) but disable HAVE_LIBUNWIND_SUPPORT (local), breaking compilation due to missing maps__e_machine declarations in maps.h. Fixed by unconditionally setting up libunwind feature build flags in Makefile.config.

2. kwork ASAN Double-Free & Leaks:
- Solves double-free in the record command argument array due to parse_options mutation, by preserving original pointers in a calloc'ed to_free buffer.
- Ensures kwork_usage string is freed on all exit paths.

Quoted 5 lines

3. kwork Unified Memory Lifecycle:
- Establishes a unified memory ownership model for kwork_work by ensuring work names are dynamically duplicated via strdup() and consistently freed using work_exit() and a central perf_kwork__exit() teardown path.

It would be nice to land these fixes. Thomas reported the kwork crashes in:
https://lore.kernel.org/linux-perf-users/[email protected]/

Thanks,
Ian

Quoted 16 lines

[1] https://lore.kernel.org/linux-perf-users/[email protected]/

Ian Rogers (3):
perf build: Unconditionally set up libunwind feature build flags perf kwork: Fix address sanitizer issues
perf kwork: Fix memory management of kwork_work

 tools/perf/Makefile.config  | 103 ++++++++++++------------
 tools/perf/builtin-kwork.c  | 151 ++++++++++++++++++++++++++++--------
 tools/perf/util/bpf_kwork.c |  14 +++-
 tools/perf/util/kwork.h     |   9 +++
 4 files changed, 188 insertions(+), 89 deletions(-)

--
2.54.0.746.g67dd491aae-goog

maintainer Applied
#6 lore raw
Quoted 24 lines · Ian Rogers
Quoted 23 lines · Ian Rogers

The combination of the libunwind refactoring and build system changes exposed a build-test regression for the libunwind build using a cached feature dump. Re-send kwork fixes [1] moving a definition to kwork.h to avoid breaking no libbpf builds.

1. Build Fix for libunwind:
"make feature-dump" runs without LIBUNWIND=1, skipping the setup of libunwind LDFLAGS. When Makefile.feature is included to generate the cached FEATURE-DUMP, the local test-libunwind.bin compiles without arch-specific link flags (-lunwind-x86_64) and fails, caching a false negative ("feature-libunwind=0"). Later builds (such as make_libunwind_O reusing the dump) compile the unwinder based on CONFIG_LIBUNWIND (remote) but disable HAVE_LIBUNWIND_SUPPORT (local), breaking compilation due to missing maps__e_machine declarations in maps.h. Fixed by unconditionally setting up libunwind feature build flags in Makefile.config.

2. kwork ASAN Double-Free & Leaks:
- Solves double-free in the record command argument array due to parse_options mutation, by preserving original pointers in a calloc'ed to_free buffer.
- Ensures kwork_usage string is freed on all exit paths.

Quoted 8 lines
Quoted 5 lines

3. kwork Unified Memory Lifecycle:
- Establishes a unified memory ownership model for kwork_work by ensuring work names are dynamically duplicated via strdup() and consistently freed using work_exit() and a central perf_kwork__exit() teardown path.

It would be nice to land these fixes. Thomas reported the kwork crashes in:
https://lore.kernel.org/linux-perf-users/[email protected]/

Thanks, applied to perf-tools-next, for v7.2.

- Arnaldo