Skip to contentSkip to conversation

perf tools: Fix pre-existing bugs in machine, cs-etm, c2c, bpf, and dso

PATCH Superseded Arnaldo Carvalho de Melo 2 participants 17 messages Jun 16

Subsystems (3)

v4 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Arnaldo Carvalho de Melo

author maintainer
#0 lore raw

v4 cover perf tools: Fix pre-existing bugs in machine, cs-etm, c2c, bpf, and dso

Hi,

Nine more pre-existing bugs found by sashiko-bot during AI-assisted code review. All are independent of the perf-data-validation hardening series — they are latent bugs in surrounding code exposed during review.

The fixes are grouped by subsystem:

machine__init() error propagation (patches 1-2):
machine__init() always returns 0 on allocation failure because the error code is never propagated through the return statement. Callers (including machines__init() and __machine__new_host()) proceed with a partially initialized machine struct. The error cleanup also uses zfree() on refcounted kmaps instead of maps__zput(). Additionally, machines__findnew() and machines__create_guest_kernel_maps() use sprintf() with unsanitized guestmount paths that can overflow PATH_MAX stack buffers.

CoreSight ETM metadata validation (patches 3-5):
cs_etm__process_auxtrace_info_full() reads num_cpu from untrusted perf.data and uses it directly in a multiplication that can overflow to zero on 32-bit, producing a zero-sized allocation followed by OOB writes. The minimum size check in cs_etm__process_auxtrace_info() doesn't cover the global header fields actually accessed. cs_etm__get_queue() indexes queue_array[] without bounds checking the CPU value from untrusted trace payload, and several queue iteration loops dereference .priv without NULL checks after array growth zero-initializes new entries.

c2c hist entry leaks (patches 6-7):
When c2c_hists__init() fails, dynamically allocated format structures are leaked because the error path frees the container without unregistering them. During resort merges, c2c_he_free() only walks the output-sorted tree (empty before resort), leaking all inner hist_entry objects from entries_in_array[] and entries_collapsed.

BPF prog info pointer validation (patch 8):
Several functions cast bpf_prog_info u64 fields to pointers without checking whether bpil_offs_to_addr() actually converted the file offsets. A crafted perf.data with PERF_BPIL_* bits unset but non-zero counts causes raw file offsets to be dereferenced as pointers.

DSO decompression errno (patch 9):
dso__get_filename() sets errno to a negative custom DSO_LOAD_ERRNO value on decompression failure. __open_dso() computes fd = -errno, producing a large positive value that looks like a valid fd, causing close_data_fd() to close an unrelated file descriptor.

Build-tested with gcc and clang. Passes perf test on x86_64.

Changes in v4 (patch 2 only):
- Remove incorrect get_kernel_version() reference from commit message — that function already uses snprintf() in the baseline (sashiko-bot).

Changes in v3 (patch 1 only):
- Move perf_env__init() before machines__init() in __perf_session__new() so the goto out_delete error path doesn't call perf_env__exit() on uninitialized mutexes/rwlocks (sashiko-bot).

Changes in v2 (patch 1 only):
- Move dsos__init()/threads__init() before maps__new() so that machine__exit() is safe to call when machine__init() fails at the first allocation (sashiko-bot).
- Propagate machines__init() error in aslr_tool__init(), which was added by the ASLR patches after v1 was written (sashiko-bot).

Arnaldo Carvalho de Melo (9):
perf machine: Propagate machine__init() error to callers perf machine: Use snprintf() for guestmount path construction perf cs-etm: Validate num_cpu before metadata allocation perf cs-etm: Require full global header in auxtrace_info size check perf cs-etm: Bounds-check CPU in cs_etm__get_queue()
perf c2c: Free format list entries when c2c_hists__init() fails perf c2c: Fix hist entry and format list leaks in c2c_he_free() perf bpf: Validate array presence before casting BPF prog info pointers perf dso: Set standard errno on decompression failure

 tools/perf/builtin-c2c.c             |  3 ++-
 tools/perf/tests/hists_cumulate.c    |  3 ++-
 tools/perf/tests/hists_filter.c      |  3 ++-
 tools/perf/tests/hists_link.c        |  3 ++-
 tools/perf/tests/hists_output.c      |  3 ++-
 tools/perf/tests/thread-maps-share.c |  2 +-
 tools/perf/util/aslr.c               | 12 +++++++++---
 tools/perf/util/bpf-event.c          | 20 ++++++++++++++++---
 tools/perf/util/bpf-event.h          |  4 ++--
 tools/perf/util/cs-etm-base.c        |  4 +++-
 tools/perf/util/cs-etm.c             | 37 ++++++++++++++++++++++++++++++++++--
 tools/perf/util/dso.c                | 18 +++++++++++++++++-
 tools/perf/util/header.c             |  3 +--
 tools/perf/util/hist.c               |  2 +-
 tools/perf/util/hist.h               |  1 +
 tools/perf/util/machine.c            | 32 +++++++++++++++++--------------
 tools/perf/util/machine.h            |  2 +-
 tools/perf/util/session.c            |  7 ++++---
 18 files changed, 120 insertions(+), 39 deletions(-)

Developed with AI assistance (Claude/sashiko), tagged in commits.

Thanks,

- Arnaldo

author maintainer
#1 lore raw

PATCH v4 1/9 perf machine: Propagate machine__init() error to callers

From: Arnaldo Carvalho de Melo <[email protected]>

machine__init() always returns 0 even when memory allocation fails, because commit 81f981d7ec43ed93 ("perf machine: Free root_dir in machine__init() error path") introduced 'int err = -ENOMEM' and an error cleanup path but left the final 'return 0' instead of 'return err'.

Fix by returning err, check the return value in __machine__new_host() which was ignoring it, and change machines__init() from void to int so it too can propagate the error to perf_session__new(), aslr_tool__init() and test callers.

The error cleanup also used zfree(&machine->kmaps), but kmaps is a refcounted maps structure — use maps__zput() to properly drop the reference, matching machine__exit().

Move dsos__init() and threads__init() before the first fallible allocation (maps__new) so that machine__exit() is safe to call on any machine struct that machine__init() touched, even on early failure.

Fixes 81f981d7ec43ed93 ("perf machine: Free root_dir in machine__init() error path")
Cc Ian Rogers <[email protected]>
Cc Jiri Olsa <[email protected]>
Assisted-by Claude <[email protected]>
---
 tools/perf/tests/hists_cumulate.c    |  3 ++-
 tools/perf/tests/hists_filter.c      |  3 ++-
 tools/perf/tests/hists_link.c        |  3 ++-
 tools/perf/tests/hists_output.c      |  3 ++-
 tools/perf/tests/thread-maps-share.c |  2 +-
 tools/perf/util/aslr.c               | 12 +++++++++---
 tools/perf/util/machine.c            | 24 ++++++++++++++----------
 tools/perf/util/machine.h            |  2 +-
 tools/perf/util/session.c            |  7 ++++---
 9 files changed, 37 insertions(+), 22 deletions(-)
tools/perf/tests/hists_cumulate.c +2 -1
index 267cbc24691acd77..9f4d4e281e11edaa 100644
@@ -723,7 +723,8 @@ static int test__hists_cumulate(struct test_suite *test __maybe_unused, int subt
 		goto out;
 	err = TEST_FAIL;
 
-	machines__init(&machines);
+	if (machines__init(&machines))
+		goto out;
 
 	/* setup threads/dso/map/symbols also */
 	machine = setup_fake_machine(&machines);
tools/perf/tests/hists_filter.c +2 -1
index 002e3a4c1ca59b9d..fa47e995de82d427 100644
@@ -131,7 +131,8 @@ static int test__hists_filter(struct test_suite *test __maybe_unused, int subtes
 		goto out;
 	err = TEST_FAIL;
 
-	machines__init(&machines);
+	if (machines__init(&machines))
+		goto out;
 
 	/* setup threads/dso/map/symbols also */
 	machine = setup_fake_machine(&machines);
tools/perf/tests/hists_link.c +2 -1
index 996f5f0b3bd17fe5..05078dac93c4ec6b 100644
@@ -303,7 +303,8 @@ static int test__hists_link(struct test_suite *test __maybe_unused, int subtest
 		goto out;
 
 	err = TEST_FAIL;
-	machines__init(&machines);
+	if (machines__init(&machines))
+		goto out;
 
 	/* setup threads/dso/map/symbols also */
 	machine = setup_fake_machine(&machines);
tools/perf/tests/hists_output.c +2 -1
index fa683fd7b1e5ebb2..1ec51c15b1046eb4 100644
@@ -610,7 +610,8 @@ static int test__hists_output(struct test_suite *test __maybe_unused, int subtes
 		goto out;
 	err = TEST_FAIL;
 
-	machines__init(&machines);
+	if (machines__init(&machines))
+		goto out;
 
 	/* setup threads/dso/map/symbols also */
 	machine = setup_fake_machine(&machines);
tools/perf/tests/thread-maps-share.c +1 -1
index e9ecd30a5c058076..0431bff31b3a18c3 100644
@@ -27,7 +27,7 @@ static int test__thread_maps_share(struct test_suite *test __maybe_unused, int s
 	 * other  group (pid: 4, tids: 4, 5)
 	*/
 
-	machines__init(&machines);
+	TEST_ASSERT_VAL("failed to init machines", machines__init(&machines) == 0);
 	machine = &machines.host;
 
 	/* create process with 4 threads */
tools/perf/util/aslr.c +9 -3
index a946fff2ac4dd4b4..6a7542e7db827d1b 100644
@@ -1237,12 +1237,13 @@ void aslr_tool__strip_attr_event(union perf_event *event, struct evlist *evlist)
 	}
 }
 
-static void aslr_tool__init(struct aslr_tool *aslr, struct perf_tool *delegate)
+static int aslr_tool__init(struct aslr_tool *aslr, struct perf_tool *delegate)
 {
 	delegate_tool__init(&aslr->tool, delegate);
 	aslr->tool.tool.ordered_events = true;
 
-	machines__init(&aslr->machines);
+	if (machines__init(&aslr->machines))
+		return -ENOMEM;
 
 	hashmap__init(&aslr->remap_addresses,
 		      remap_addresses__hash, remap_addresses__equal,
@@ -1276,6 +1277,8 @@ static void aslr_tool__init(struct aslr_tool *aslr, struct perf_tool *delegate)
 	aslr->tool.tool.auxtrace = aslr_tool__process_auxtrace;
 	aslr->tool.tool.auxtrace_info = aslr_tool__process_auxtrace_info;
 	aslr->tool.tool.auxtrace_error = aslr_tool__process_auxtrace_error;
+
+	return 0;
 }
 
 struct perf_tool *aslr_tool__new(struct perf_tool *delegate)
@@ -1285,7 +1288,10 @@ struct perf_tool *aslr_tool__new(struct perf_tool *delegate)
 	if (!aslr)
 		return NULL;
 
-	aslr_tool__init(aslr, delegate);
+	if (aslr_tool__init(aslr, delegate)) {
+		free(aslr);
+		return NULL;
+	}
 	return &aslr->tool.tool;
 }
 
tools/perf/util/machine.c +14 -10
index 31715366e29ff704..9329d319bd033699 100644
@@ -79,15 +79,14 @@ int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
 	int err = -ENOMEM;
 
 	memset(machine, 0, sizeof(*machine));
-	machine->kmaps = maps__new(machine);
-	if (machine->kmaps == NULL)
-		return -ENOMEM;
-
 	RB_CLEAR_NODE(&machine->rb_node);
 	dsos__init(&machine->dsos);
-
 	threads__init(&machine->threads);
 
+	machine->kmaps = maps__new(machine);
+	if (machine->kmaps == NULL)
+		goto out;
+
 	machine->vdso_info = NULL;
 	machine->env = NULL;
 
@@ -124,11 +123,11 @@ int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
 
 out:
 	if (err) {
-		zfree(&machine->kmaps);
+		maps__zput(machine->kmaps);
 		zfree(&machine->root_dir);
 		zfree(&machine->mmap_name);
 	}
-	return 0;
+	return err;
 }
 
 static struct machine *__machine__new_host(struct perf_env *host_env, bool kernel_maps)
@@ -138,7 +137,10 @@ static struct machine *__machine__new_host(struct perf_env *host_env, bool kerne
 	if (!machine)
 		return NULL;
 
-	machine__init(machine, "", HOST_KERNEL_ID);
+	if (machine__init(machine, "", HOST_KERNEL_ID) != 0) {
+		free(machine);
+		return NULL;
+	}
 
 	if (kernel_maps && machine__create_kernel_maps(machine) < 0) {
 		free(machine);
@@ -231,10 +233,12 @@ void machine__delete(struct machine *machine)
 	}
 }
 
-void machines__init(struct machines *machines)
+int machines__init(struct machines *machines)
 {
-	machine__init(&machines->host, "", HOST_KERNEL_ID);
+	int err = machine__init(&machines->host, "", HOST_KERNEL_ID);
+
 	machines->guests = RB_ROOT_CACHED;
+	return err;
 }
 
 void machines__exit(struct machines *machines)
tools/perf/util/machine.h +1 -1
index aaddfb70ea665452..26f9827062f5eb5b 100644
@@ -152,7 +152,7 @@ struct machines {
 	struct rb_root_cached guests;
 };
 
-void machines__init(struct machines *machines);
+int machines__init(struct machines *machines);
 void machines__exit(struct machines *machines);
 
 void machines__process_guests(struct machines *machines,
tools/perf/util/session.c +4 -3
index 1a9a008ddda35120..f391a822480db001 100644
@@ -160,11 +160,12 @@ struct perf_session *__perf_session__new(struct perf_data *data,
 	session->decomp_data.zstd_decomp = &session->zstd_data;
 	session->active_decomp = &session->decomp_data;
 	INIT_LIST_HEAD(&session->auxtrace_index);
-	machines__init(&session->machines);
+	perf_env__init(&session->header.env);
+	if (machines__init(&session->machines))
+		goto out_delete;
+
 	ordered_events__init(&session->ordered_events,
 			     ordered_events__deliver_event, NULL);
-
-	perf_env__init(&session->header.env);
 	if (data) {
 		ret = perf_data__open(data);
 		if (ret < 0)
Signature · 1 line
2.54.0
author maintainer
#3 lore raw

PATCH v4 2/9 perf machine: Use snprintf() for guestmount path construction

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

machines__findnew() and machines__create_guest_kernel_maps() use sprintf() to build paths by prepending symbol_conf.guestmount. Both write into PATH_MAX stack buffers, but guestmount comes from user configuration and is not length-checked. A guestmount path at or near PATH_MAX causes a stack buffer overflow.

Switch to snprintf() with sizeof() to prevent overflow. The subsequent access()/fopen() calls will fail on a truncated path.

Fixes a1645ce12adb6c9c ("perf: 'perf kvm' tool for monitoring guest performance from host")
Cc Zhang, Yanmin <[email protected]>
Assisted-by Claude <[email protected]>
---
 tools/perf/util/machine.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
tools/perf/util/machine.c +4 -4
index 9329d319bd033699..0d2ebf6a84bcf880 100644
@@ -333,7 +333,7 @@ struct machine *machines__findnew(struct machines *machines, pid_t pid)
 	if ((pid != HOST_KERNEL_ID) &&
 	    (pid != DEFAULT_GUEST_KERNEL_ID) &&
 	    (symbol_conf.guestmount)) {
-		sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
+		snprintf(path, sizeof(path), "%s/%d", symbol_conf.guestmount, pid);
 		if (access(path, R_OK)) {
 			static struct strlist *seen;
 
@@ -1260,9 +1260,9 @@ int machines__create_guest_kernel_maps(struct machines *machines)
 					 namelist[i]->d_name);
 				continue;
 			}
-			sprintf(path, "%s/%s/proc/kallsyms",
-				symbol_conf.guestmount,
-				namelist[i]->d_name);
+			snprintf(path, sizeof(path), "%s/%s/proc/kallsyms",
+				 symbol_conf.guestmount,
+				 namelist[i]->d_name);
 			ret = access(path, R_OK);
 			if (ret) {
 				pr_debug("Can't access file %s\n", path);
Signature · 1 line
2.54.0
author maintainer
#5 lore raw

PATCH v4 3/9 perf cs-etm: Validate num_cpu before metadata allocation

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

cs_etm__process_auxtrace_info_full() reads num_cpu from untrusted perf.data and uses it to allocate the metadata pointer array:

  metadata = zalloc(sizeof(*metadata) * num_cpu);

On 32-bit, sizeof(*metadata) is 4, so num_cpu = 0x40000000 overflows the multiplication to 0, causing zalloc(0) to return a valid zero-sized allocation followed by out-of-bounds writes in the population loop.

Fix by computing priv_size early and using it to bound num_cpu: each CPU needs at least one u64 metadata entry, so num_cpu cannot exceed the total number of u64 entries in the event's private data area.

Fixes cd8bfd8c973eaff8 ("perf tools: Add processing of coresight metadata")
Cc Adrian Hunter <[email protected]>
Cc James Clark <[email protected]>
Cc Leo Yan <[email protected]>
Cc Tor Jeremiassen <[email protected]>
Assisted-by Claude <[email protected]>
---
 tools/perf/util/cs-etm.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
tools/perf/util/cs-etm.c +12 -0
index 0927b0b9c06b1504..d121c8f22028d5ba 100644
@@ -3431,6 +3431,18 @@ int cs_etm__process_auxtrace_info_full(union perf_event *event,
 	/* First the global part */
 	ptr = (u64 *) auxtrace_info->priv;
 	num_cpu = ptr[CS_PMU_TYPE_CPUS] & 0xffffffff;
+
+	/*
+	 * Bound num_cpu by the event size: the global header consumes
+	 * CS_ETM_HEADER_SIZE bytes, and each CPU needs at least one u64
+	 * metadata entry after that.
+	 */
+	priv_size = total_size - event_header_size - INFO_HEADER_SIZE -
+		    CS_ETM_HEADER_SIZE;
+	if (num_cpu <= 0 || priv_size <= 0 ||
+	    num_cpu > priv_size / (int)sizeof(u64))
+		return -EINVAL;
+
 	metadata = zalloc(sizeof(*metadata) * num_cpu);
 	if (!metadata)
 		return -ENOMEM;
Signature · 1 line
2.54.0
author maintainer
#7 lore raw

PATCH v4 4/9 perf cs-etm: Require full global header in auxtrace_info size check

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

cs_etm__process_auxtrace_info() checks that header.size covers event_header_size + INFO_HEADER_SIZE (16 bytes total), but then accesses ptr[CS_PMU_TYPE_CPUS] at offset 24 from the start of the event. A crafted 16-byte auxtrace_info event passes the size check but reads out-of-bounds.

Include CS_ETM_HEADER_SIZE in the minimum size check so that the global header entries (version, pmu_type_cpus, snapshot) are guaranteed to fit within the event.

Fixes 55c1de9973d66516 ("perf cs-etm: Print auxtrace info even if OpenCSD isn't linked")
Cc Adrian Hunter <[email protected]>
Cc James Clark <[email protected]>
Cc Leo Yan <[email protected]>
Assisted-by Claude <[email protected]>
---
 tools/perf/util/cs-etm-base.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
tools/perf/util/cs-etm-base.c +3 -1
index 4abe416e3febd2cc..aebef71d3a0a1d7b 100644
@@ -170,7 +170,9 @@ int cs_etm__process_auxtrace_info(union perf_event *event,
 	u64 *ptr = NULL;
 	u64 hdr_version;
 
-	if (auxtrace_info->header.size < (event_header_size + INFO_HEADER_SIZE))
+	/* Ensure priv[] is large enough for the global header entries */
+	if (auxtrace_info->header.size < (event_header_size + INFO_HEADER_SIZE +
+					  CS_ETM_HEADER_SIZE))
 		return -EINVAL;
 
 	/* First the global part */
Signature · 1 line
2.54.0
author maintainer
#9 lore raw

PATCH v4 5/9 perf cs-etm: Bounds-check CPU in cs_etm__get_queue()

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

cs_etm__get_queue() indexes etm->queues.queue_array[cpu] without validating that cpu is within nr_queues. When processing AUX_OUTPUT_HW_ID events, the cpu value comes from untrusted perf.data trace payload and flows through cs_etm__process_trace_id_v0_1() and cs_etm__queue_aux_fragment() without bounds checking, allowing an out-of-bounds read with a crafted file.

Add a bounds check in cs_etm__get_queue() and NULL checks in all callers.

Also add NULL checks for queue_array[i].priv in the queue iteration loops in cs_etm__map_trace_id_v0() and cs_etm__process_trace_id_v0_1() — after auxtrace_queues__grow() new entries are zero-initialized so .priv can be NULL. Add a get_cpu_data() NULL check in cs_etm__process_trace_id_v0_1(), matching the existing check in cs_etm__process_trace_id_v0().

Fixes 77c123f53e97ad4b ("perf: cs-etm: Move traceid_list to each queue")
Cc Adrian Hunter <[email protected]>
Cc James Clark <[email protected]>
Cc Leo Yan <[email protected]>
Assisted-by Claude <[email protected]>
---
 tools/perf/util/cs-etm.c | 25 +++++++++++++++++++++++--
 1 file changed, 23 insertions(+), 2 deletions(-)
tools/perf/util/cs-etm.c +23 -2
index d121c8f22028d5ba..5d0664ff73b79122 100644
@@ -292,8 +292,11 @@ static struct cs_etm_queue *cs_etm__get_queue(struct cs_etm_auxtrace *etm, int c
 {
 	if (etm->per_thread_decoding)
 		return etm->queues.queue_array[0].priv;
-	else
-		return etm->queues.queue_array[cpu].priv;
+
+	if (cpu < 0 || cpu >= (int)etm->queues.nr_queues)
+		return NULL;
+
+	return etm->queues.queue_array[cpu].priv;
 }
 
 static int cs_etm__map_trace_id_v0(struct cs_etm_auxtrace *etm, u8 trace_chan_id,
@@ -306,6 +309,9 @@ static int cs_etm__map_trace_id_v0(struct cs_etm_auxtrace *etm, u8 trace_chan_id
 	 * queue associated with that CPU so only one decoder is made.
 	 */
 	etmq = cs_etm__get_queue(etm, cpu_metadata[CS_ETM_CPU]);
+	if (!etmq)
+		return -EINVAL;
+
 	if (etmq->format == UNFORMATTED)
 		return cs_etm__insert_trace_id_node(etmq, trace_chan_id,
 						    cpu_metadata);
@@ -318,6 +324,9 @@ static int cs_etm__map_trace_id_v0(struct cs_etm_auxtrace *etm, u8 trace_chan_id
 		int ret;
 
 		etmq = etm->queues.queue_array[i].priv;
+		if (!etmq)
+			continue;
+
 		ret = cs_etm__insert_trace_id_node(etmq, trace_chan_id,
 						   cpu_metadata);
 		if (ret)
@@ -358,6 +367,9 @@ static int cs_etm__process_trace_id_v0_1(struct cs_etm_auxtrace *etm, int cpu,
 	u32 sink_id = FIELD_GET(CS_AUX_HW_ID_SINK_ID_MASK, hw_id);
 	u8 trace_id = FIELD_GET(CS_AUX_HW_ID_TRACE_ID_MASK, hw_id);
 
+	if (!etmq)
+		return -EINVAL;
+
 	/*
 	 * Check sink id hasn't changed in per-cpu mode. In per-thread mode,
 	 * let it pass for now until an actual overlapping trace ID is hit. In
@@ -375,6 +387,9 @@ static int cs_etm__process_trace_id_v0_1(struct cs_etm_auxtrace *etm, int cpu,
 	for (unsigned int i = 0; i < etm->queues.nr_queues; ++i) {
 		struct cs_etm_queue *other_etmq = etm->queues.queue_array[i].priv;
 
+		if (!other_etmq)
+			continue;
+
 		/* Different sinks, skip */
 		if (other_etmq->sink_id != etmq->sink_id)
 			continue;
@@ -396,6 +411,9 @@ static int cs_etm__process_trace_id_v0_1(struct cs_etm_auxtrace *etm, int cpu,
 	}
 
 	cpu_data = get_cpu_data(etm, cpu);
+	if (!cpu_data)
+		return -EINVAL;
+
 	ret = cs_etm__insert_trace_id_node(etmq, trace_id, cpu_data);
 	if (ret)
 		return ret;
@@ -3144,6 +3162,9 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o
 	    aux_offset + aux_size <= auxtrace_event->offset + auxtrace_event->size) {
 		struct cs_etm_queue *etmq = cs_etm__get_queue(etm, auxtrace_event->cpu);
 
+		if (!etmq)
+			return -EINVAL;
+
 		/*
 		 * If this AUX event was inside this buffer somewhere, create a new auxtrace event
 		 * based on the sizes of the aux event, and queue that fragment.
Signature · 1 line
2.54.0
author maintainer
#11 lore raw

PATCH v4 6/9 perf c2c: Free format list entries when c2c_hists__init() fails

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

When c2c_hists__init() fails partway through hpp_list__parse(), dynamically allocated format structures that were already added to hists->list are leaked because he__get_c2c_hists() frees the hists container without first unregistering the format entries.

Call perf_hpp__reset_output_field() before freeing the hists container on the error path, matching what c2c_he_free() already does on the normal destruction path.

Fixes 17a7c5946d79a12c ("perf c2c report: Decode c2c_stats for hist entries")
Cc Jiri Olsa <[email protected]>
Assisted-by Claude <[email protected]>
---
 tools/perf/builtin-c2c.c | 1 +
 1 file changed, 1 insertion(+)
tools/perf/builtin-c2c.c +1 -0
index 07c7e8fb315e6cf3..eabb922ef295ef86 100644
@@ -226,6 +226,7 @@ he__get_c2c_hists(struct hist_entry *he,
 
 	ret = c2c_hists__init(hists, sort, nr_header_lines, env);
 	if (ret) {
+		perf_hpp__reset_output_field(&hists->list);
 		c2c_he->hists = NULL;
 		free(hists);
 		return NULL;
Signature · 1 line
2.54.0
author maintainer
#12 lore raw

PATCH v4 7/9 perf c2c: Fix hist entry and format list leaks in c2c_he_free()

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

c2c_he_free() calls hists__delete_entries() which only walks the output-sorted entries tree. During c2c resort, when cacheline entries are merged and the redundant entry is freed, the inner hists have not been output-resorted yet, so hists->entries is empty. The actual inner hist_entry objects live in entries_in_array[] and entries_collapsed, which are never walked, leaking all inner hist_entry objects for every merged cacheline.

Additionally, the dynamically allocated format entries on hists->list are never unregistered or freed.

Fix both issues by switching to hists__delete_all_entries() which walks all rb_root trees, and calling perf_hpp__reset_output_field() to clean up format entries.

Fixes bf0e0d407ea09ce5 ("perf c2c report: Add sample processing")
Cc Jiri Olsa <[email protected]>
Assisted-by Claude <[email protected]>
---
 tools/perf/builtin-c2c.c | 2 +-
 tools/perf/util/hist.c   | 2 +-
 tools/perf/util/hist.h   | 1 +
 3 files changed, 3 insertions(+), 2 deletions(-)
tools/perf/builtin-c2c.c +1 -1
index eabb922ef295ef86..c9584dbedf77afe8 100644
@@ -184,7 +184,7 @@ static void c2c_he_free(void *he)
 
 	c2c_he = container_of(he, struct c2c_hist_entry, he);
 	if (c2c_he->hists) {
-		hists__delete_entries(&c2c_he->hists->hists);
+		hists__delete_all_entries(&c2c_he->hists->hists);
 		perf_hpp__reset_output_field(&c2c_he->hists->list);
 		zfree(&c2c_he->hists);
 	}
tools/perf/util/hist.c +1 -1
index df978c996b6c2262..c93915625ee75de1 100644
@@ -3041,7 +3041,7 @@ static void hists__delete_remaining_entries(struct rb_root_cached *root)
 	}
 }
 
-static void hists__delete_all_entries(struct hists *hists)
+void hists__delete_all_entries(struct hists *hists)
 {
 	hists__delete_entries(hists);
 	hists__delete_remaining_entries(&hists->entries_in_array[0]);
tools/perf/util/hist.h +1 -0
index 8fb89d81ef069d95..b830cbe7f95bf597 100644
@@ -391,6 +391,7 @@ int hists__collapse_resort(struct hists *hists, struct ui_progress *prog);
 
 void hists__decay_entries(struct hists *hists, bool zap_user, bool zap_kernel);
 void hists__delete_entries(struct hists *hists);
+void hists__delete_all_entries(struct hists *hists);
 void hists__output_recalc_col_len(struct hists *hists, int max_rows);
 
 struct hist_entry *hists__get_entry(struct hists *hists, int idx);
Signature · 1 line
2.54.0
author maintainer
#13 lore raw

PATCH v4 8/9 perf bpf: Validate array presence before casting BPF prog info pointers

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

Several functions cast bpf_prog_info fields (jited_ksyms, jited_func_lens, jited_prog_insns) from u64 to pointers and dereference them. These fields are only valid pointers if bpil_offs_to_addr() converted their file offsets to addresses, which only happens when the corresponding PERF_BPIL_* bits are set in info_linear->arrays.

A crafted perf.data can leave these bits unset while setting non-zero counts and offset values, causing the functions to dereference raw file offsets as pointers.

Add array bitmask validation to all perf.data processing paths:

- __bpf_event__print_bpf_prog_info(): check JITED_KSYMS and JITED_FUNC_LENS (changed to take struct perf_bpil *)
- machine__process_bpf_event_load(): check JITED_KSYMS
- bpf_read(): check JITED_INSNS before memcpy from jited_prog_insns
- dso__disassemble_filename(): check JITED_INSNS before returning jited_prog_insns pointer

Fixes f8dfeae009effc0b ("perf bpf: Show more BPF program info in print_bpf_prog_info()")
Cc Song Liu <[email protected]>
Assisted-by Claude <[email protected]>
---
 tools/perf/util/bpf-event.c | 20 +++++++++++++++++---
 tools/perf/util/bpf-event.h |  4 ++--
 tools/perf/util/dso.c       | 10 ++++++++++
 tools/perf/util/header.c    |  3 +--
 4 files changed, 30 insertions(+), 7 deletions(-)
tools/perf/util/bpf-event.c +17 -3
index 57d53ba848359e12..fa3ebc8ea7f09cdd 100644
@@ -59,6 +59,10 @@ static int machine__process_bpf_event_load(struct machine *machine,
 		return 0;
 	info_linear = info_node->info_linear;
 
+	/* jited_ksyms is only valid if bpil_offs_to_addr() converted it */
+	if (!(info_linear->arrays & (1UL << PERF_BPIL_JITED_KSYMS)))
+		return 0;
+
 	for (i = 0; i < info_linear->info.nr_jited_ksyms; i++) {
 		u64 *addrs = (u64 *)(uintptr_t)(info_linear->info.jited_ksyms);
 		u64 addr = addrs[i];
@@ -959,12 +963,15 @@ int evlist__add_bpf_sb_event(struct evlist *evlist, struct perf_env *env)
 	return evlist__add_sb_event(evlist, &attr, bpf_event__sb_cb, env);
 }
 
-void __bpf_event__print_bpf_prog_info(struct bpf_prog_info *info,
+void __bpf_event__print_bpf_prog_info(struct perf_bpil *info_linear,
 				      struct perf_env *env,
 				      FILE *fp)
 {
-	__u32 *prog_lens = (__u32 *)(uintptr_t)(info->jited_func_lens);
-	__u64 *prog_addrs = (__u64 *)(uintptr_t)(info->jited_ksyms);
+	struct bpf_prog_info *info = &info_linear->info;
+	__u64 required_arrays = (1UL << PERF_BPIL_JITED_KSYMS) |
+				(1UL << PERF_BPIL_JITED_FUNC_LENS);
+	__u32 *prog_lens;
+	__u64 *prog_addrs;
 	char name[KSYM_NAME_LEN];
 	struct btf *btf = NULL;
 	u32 sub_prog_cnt, i;
@@ -974,6 +981,13 @@ void __bpf_event__print_bpf_prog_info(struct bpf_prog_info *info,
 	    sub_prog_cnt != info->nr_jited_func_lens)
 		return;
 
+	/* Ensure the arrays were present and converted by bpil_offs_to_addr() */
+	if ((info_linear->arrays & required_arrays) != required_arrays)
+		return;
+
+	prog_lens = (__u32 *)(uintptr_t)(info->jited_func_lens);
+	prog_addrs = (__u64 *)(uintptr_t)(info->jited_ksyms);
+
 	if (info->btf_id) {
 		struct btf_node *node;
 
tools/perf/util/bpf-event.h +2 -2
index 60d2c6637af5d6eb..da4eeb4a1a73208c 100644
@@ -40,7 +40,7 @@ struct btf_node {
 int machine__process_bpf(struct machine *machine, union perf_event *event,
 			 struct perf_sample *sample);
 int evlist__add_bpf_sb_event(struct evlist *evlist, struct perf_env *env);
-void __bpf_event__print_bpf_prog_info(struct bpf_prog_info *info,
+void __bpf_event__print_bpf_prog_info(struct perf_bpil *info_linear,
 				      struct perf_env *env,
 				      FILE *fp);
 void bpf_metadata_free(struct bpf_metadata *metadata);
@@ -58,7 +58,7 @@ static inline int evlist__add_bpf_sb_event(struct evlist *evlist __maybe_unused,
 	return 0;
 }
 
-static inline void __bpf_event__print_bpf_prog_info(struct bpf_prog_info *info __maybe_unused,
+static inline void __bpf_event__print_bpf_prog_info(struct perf_bpil *info_linear __maybe_unused,
 						    struct perf_env *env __maybe_unused,
 						    FILE *fp __maybe_unused)
 {
tools/perf/util/dso.c +10 -0
index 1a2fc6d18da74d6c..79f1a30f3683d6b3 100644
@@ -880,6 +880,12 @@ static ssize_t bpf_read(struct dso *dso, u64 offset, char *data)
 		return -1;
 	}
 
+	/* jited_prog_insns is only valid if bpil_offs_to_addr() converted it */
+	if (!(node->info_linear->arrays & (1UL << PERF_BPIL_JITED_INSNS))) {
+		dso__data(dso)->status = DSO_DATA_STATUS_ERROR;
+		return -1;
+	}
+
 	len = node->info_linear->info.jited_prog_len;
 	buf = (u8 *)(uintptr_t)node->info_linear->info.jited_prog_insns;
 
@@ -1995,6 +2001,10 @@ const u8 *dso__read_symbol(struct dso *dso, const char *symfs_filename,
 			return NULL;
 		}
 		info_linear = info_node->info_linear;
+		if (!(info_linear->arrays & (1UL << PERF_BPIL_JITED_INSNS))) {
+			errno = SYMBOL_ANNOTATE_ERRNO__BPF_MISSING_BTF;
+			return NULL;
+		}
 		assert(len <= info_linear->info.jited_prog_len);
 		*out_buf_len = len;
 		return (const u8 *)(uintptr_t)(info_linear->info.jited_prog_insns);
tools/perf/util/header.c +1 -2
index d7f41db7322cbcb4..091d8f7f6bd2c9d5 100644
@@ -2107,8 +2107,7 @@ static void print_bpf_prog_info(struct feat_fd *ff __maybe_unused, FILE *fp)
 		node = rb_entry(next, struct bpf_prog_info_node, rb_node);
 		next = rb_next(&node->rb_node);
 
-		__bpf_event__print_bpf_prog_info(&node->info_linear->info,
-						 env, fp);
+		__bpf_event__print_bpf_prog_info(node->info_linear, env, fp);
 	}
 
 	up_read(&env->bpf_progs.lock);
Signature · 1 line
2.54.0
author maintainer
#15 lore raw

PATCH v4 9/9 perf dso: Set standard errno on decompression failure

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

dso__get_filename() sets errno to a negative custom DSO_LOAD_ERRNO value when kernel module decompression fails:

errno = *dso__load_errno(dso); /* e.g. -9996 */

The caller __open_dso() then computes fd = -errno, producing a large positive value (9996) that looks like a valid file descriptor. This can cause close_data_fd() to close an unrelated fd used by another subsystem.

Set errno to EIO instead. The detailed error code is already stored in dso__load_errno(dso) for diagnostic messages.

Fixes 1d6b3c9ba756a513 ("perf tools: Decompress kernel module when reading DSO data")
Cc Namhyung Kim <[email protected]>
Assisted-by Claude <[email protected]>
---
 tools/perf/util/dso.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
tools/perf/util/dso.c +7 -1
index 79f1a30f3683d6b3..2309196d8df3111c 100644
@@ -600,7 +600,13 @@ static char *dso__get_filename(struct dso *dso, const char *root_dir,
 		size_t len = sizeof(newpath);
 
 		if (dso__decompress_kmodule_path(dso, name, newpath, len) < 0) {
-			errno = *dso__load_errno(dso);
+			/*
+			 * Use a standard errno value, not the negative custom
+			 * DSO_LOAD_ERRNO stored in dso__load_errno(dso):
+			 * __open_dso() computes fd = -errno, so a negative
+			 * errno produces a positive fd that looks valid.
+			 */
+			errno = EIO;
 			goto out;
 		}
 
Signature · 1 line
2.54.0
#6 lore raw

Re: [PATCH 3/9] perf cs-etm: Validate num_cpu before metadata allocation

In reply to Arnaldo Carvalho de Melo
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] Insufficient bounds checking in the per-CPU metadata parsing loop allows arbitrary out-of-bounds reads based on untrusted event data.
- [Medium] Memory allocated for etm->queues array elements (etmq) is leaked when initialization fails on the error path. --

commit 3997f8cbed57aaf438aeb6594523932b36c0d9d5
Author: Arnaldo Carvalho de Melo <[email protected]>

perf cs-etm: Validate num_cpu before metadata allocation

This commit computes priv_size early to bound num_cpu against the private data area size. This prevents a 32-bit multiplication overflow and subsequent out-of-bounds writes when allocating the metadata array.

Quoted 10 lines
tools/perf/util/cs-etm.c +2 -0
index 0927b0b9c06b1..d121c8f22028d 100644
@@ -3431,6 +3431,18 @@ int cs_etm__process_auxtrace_info_full(union perf_event *event,
 	/* First the global part */
 	ptr = (u64 *) auxtrace_info->priv;
 	num_cpu = ptr[CS_PMU_TYPE_CPUS] & 0xffffffff;
+
+	/*

tools/perf/util/cs-etm.c int cs_etm__process_auxtrace_info_full(union perf_event *event, · line 3445

Quoted 10 lines
tools/perf/util/cs-etm.c +10 -0
+	 * Bound num_cpu by the event size: the global header consumes
+	 * CS_ETM_HEADER_SIZE bytes, and each CPU needs at least one u64
+	 * metadata entry after that.
+	 */
+	priv_size = total_size - event_header_size - INFO_HEADER_SIZE -
+		    CS_ETM_HEADER_SIZE;
+	if (num_cpu <= 0 || priv_size <= 0 ||
+	    num_cpu > priv_size / (int)sizeof(u64))
+		return -EINVAL;
+

[Severity: High]
This is a pre-existing issue, but does the new validation fully prevent out-of-bounds reads during metadata parsing?

While the check ensures a minimum of one 8-byte entry per CPU, it doesn't protect the inner parsing loop. When cs_etm__create_meta_blk() is called, it reads nr_in_params directly from the untrusted perf.data file payload:

tools/perf/util/cs-etm.c:cs_etm__create_meta_blk()

    ...
    nr_in_params = buff_in[i + CS_ETM_NR_TRC_PARAMS];
    ...
    for (k = 0; k < nr_in_params; k++)
        metadata[k] = buff_in[i + k];
    ...

Could a maliciously crafted perf.data file provide an arbitrarily large nr_in_params, causing an out-of-bounds read before the index i is finally validated against priv_size at the end of the function?

tools/perf/util/cs-etm.c int cs_etm__process_auxtrace_info_full(union perf_event *event, · line 3448

Quoted 3 lines
tools/perf/util/cs-etm.c +0 -0
 	metadata = zalloc(sizeof(*metadata) * num_cpu);
 	if (!metadata)
 		return -ENOMEM;

[Severity: Medium]
This is a pre-existing issue, but does the error handling in this function leak memory?

During initialization, cs_etm__setup_queue() allocates a struct cs_etm_queue and assigns it to the .priv pointer. If initialization fails later, the code branches here:

tools/perf/util/cs-etm.c:cs_etm__process_auxtrace_info_full()

    ...
err_free_queues:
    auxtrace_queues__free(&etm->queues);

Since auxtrace_queues__free() is a generic helper that doesn't clean up subsystem-specific .priv contexts, will these dynamically allocated queue structures be leaked when the error path is taken?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=3
#4 lore raw

Re: [PATCH 2/9] perf machine: Use snprintf() for guestmount path construction

In reply to Arnaldo Carvalho de Melo
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] Aborting processing of all remaining guest maps when a single directory lacks proc/kallsyms.
- [High] Uninitialized errno checked against ERANGE after strtol() causes valid directory names to be erroneously skipped. --

commit 33df4f15ea6c660ee8ff4b243d151622bd2d9271
Author: Arnaldo Carvalho de Melo <[email protected]>

perf machine: Use snprintf() for guestmount path construction

This commit prevents a potential stack buffer overflow by replacing sprintf() with snprintf() when building paths using the user-provided guestmount configuration.

Fixes a1645ce12adb6c9c ("perf: 'perf kvm' tool for monitoring guest performance from host")
Cc Zhang, Yanmin <[email protected]>
Assisted-by Claude <[email protected]>
Quoted 4 lines
tools/perf/util/machine.c +0 -0
index 9329d319bd033..0d2ebf6a84bcf 100644

[ ... ]

tools/perf/util/machine.c int machines__create_guest_kernel_maps(struct machines *machines) · line 1262

Quoted 4 lines
tools/perf/util/machine.c +0 -0
@@ -1260,9 +1260,9 @@ int machines__create_guest_kernel_maps(struct machines *machines)
 					 namelist[i]->d_name);
 				continue;
 			}

[Severity: High]
This isn't a bug introduced by this patch, but does the strtol() parsing in machines__create_guest_kernel_maps() handle errno correctly?

Since errno is not explicitly cleared to 0 before the strtol() call right above this continue statement, if a previous directory name caused an overflow, errno will remain set to ERANGE.

This would cause the (errno == ERANGE) check to evaluate to true for all subsequent valid directory PIDs, erroneously skipping them. Could we set errno = 0 before parsing?

tools/perf/util/machine.c int machines__create_guest_kernel_maps(struct machines *machines) · line 1268

Quoted 9 lines
tools/perf/util/machine.c +3 -3
-			sprintf(path, "%s/%s/proc/kallsyms",
-				symbol_conf.guestmount,
-				namelist[i]->d_name);
+			snprintf(path, sizeof(path), "%s/%s/proc/kallsyms",
+				 symbol_conf.guestmount,
+				 namelist[i]->d_name);
 			ret = access(path, R_OK);
 			if (ret) {
 				pr_debug("Can't access file %s\n", path);

[Severity: High]
This is also a pre-existing issue, but does this access failure in machines__create_guest_kernel_maps() intend to abort the entire initialization loop?

If access() fails here, the code executes a goto failure. This means one inaccessible guest directory or missing proc/kallsyms will abort the loop and skip all subsequent valid guest map directories.

Earlier in the loop, invalid directory names use continue to safely skip just that entry. Would it be better to continue instead of aborting the processing of all remaining maps?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2
#8 lore raw

Re: [PATCH 4/9] perf cs-etm: Require full global header in auxtrace_info size check

In reply to Arnaldo Carvalho de Melo
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] Unbounded memory accesses during per-CPU metadata parsing and printing allow out-of-bounds reads on maliciously crafted perf.data files. --

commit bac918b1be09ad8eb36712cc82561391a2657983
Author: Arnaldo Carvalho de Melo <[email protected]>

perf cs-etm: Require full global header in auxtrace_info size check

cs_etm__process_auxtrace_info() checks that header.size covers event_header_size + INFO_HEADER_SIZE (16 bytes total), but then accesses ptr[CS_PMU_TYPE_CPUS] at offset 24 from the start of the event. A crafted 16-byte auxtrace_info event passes the size check but reads out-of-bounds.

Include CS_ETM_HEADER_SIZE in the minimum size check so that the global header entries (version, pmu_type_cpus, snapshot) are guaranteed to fit within the event.

Quoted 13 lines
tools/perf/util/cs-etm-base.c +3 -1
index 4abe416e3febd..aebef71d3a0a1 100644
@@ -170,7 +170,9 @@ int cs_etm__process_auxtrace_info(union perf_event *event,
 	u64 *ptr = NULL;
 	u64 hdr_version;
 
-	if (auxtrace_info->header.size < (event_header_size + INFO_HEADER_SIZE))
+	/* Ensure priv[] is large enough for the global header entries */
+	if (auxtrace_info->header.size < (event_header_size + INFO_HEADER_SIZE +
+					  CS_ETM_HEADER_SIZE))
 		return -EINVAL;

[ ... ]

tools/perf/util/cs-etm-base.c

Quoted 7 lines
tools/perf/util/cs-etm-base.c +0 -0
 	if (dump_trace) {
 		num_cpu = ptr[CS_PMU_TYPE_CPUS] & 0xffffffff;
 		cs_etm__print_auxtrace_info(ptr, num_cpu);
 	}
 
 	return cs_etm__process_auxtrace_info_full(event, session);
 }

[Severity: Medium]
This is a pre-existing issue, but does this still leave an out-of-bounds read if the event size only covers the global header, and the parsed num_cpu value requires more data?

Looking at the dump_trace path above, cs_etm__print_auxtrace_info() uses num_cpu as a loop boundary:

tools/perf/util/cs-etm-base.c:cs_etm__print_auxtrace_info() {
        ...
        for (i = CS_HEADER_VERSION_MAX; cpu < num; cpu++) {
                ...
}

Could this read and print memory out of bounds if the remaining event size isn't checked to ensure it contains all the variable-length per-CPU parameters?

Additionally, in cs_etm__process_auxtrace_info_full(), num_cpu is checked against the total size:

tools/perf/util/cs-etm.c:cs_etm__process_auxtrace_info_full() {
        ...
        if (num_cpu > priv_size / (int)sizeof(u64))
                return -EINVAL;
        ...
}

Does this check only guarantee a single 64-bit word per CPU? If cs_etm__create_meta_blk() later consumes multiple words per CPU and dynamically advances the stream offset based on untrusted payload counts like nr_in_params, could this also result in reading past the end of the event payload before the final size match validation catches it?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=4
#16 lore raw

Re: [PATCH 9/9] perf dso: Set standard errno on decompression failure

In reply to Arnaldo Carvalho de Melo
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] If dso__get_filename() returns NULL without explicitly setting errno (e.g., when the target path is a directory and is_regular_file() returns false without altering errno, or when dso__has_build_id() is false), __open_dso() computes fd = -errno, resulting in fd = 0. The subsequent check if (fd >= 0) in open_dso() incorrectly treats this as a successful open and registers standard input (fd 0) as the DSO data file descriptor.
- [High] Under REFCNT_CHECKING, dso__list_del() removes a DSO from the open list and temporarily unlocks _dso__data_open_lock before decrementing dso__data_open_cnt. This breaks the invariant between the list length and the counter. If a concurrent thread calls close_first_dso(), it observes dso__data_open_cnt > 0 but reads from an empty list.
--

commit 6181e27c050bdb8f633dfde54842ef5d8d538b5a
Author: Arnaldo Carvalho de Melo <[email protected]>

perf dso: Set standard errno on decompression failure

dso__get_filename sets errno to a negative custom DSO_LOAD_ERRNO value when kernel module decompression fails. The caller __open_dso then computes fd = -errno, producing a large positive value that looks like a valid file descriptor. This commit sets errno to EIO instead to avoid incorrectly closing unrelated file descriptors.

Fixes 1d6b3c9ba756a513 ("perf tools: Decompress kernel module when reading DSO data")
Cc Namhyung Kim <[email protected]>
Assisted-by Claude <[email protected]>
Quoted 8 lines
tools/perf/util/dso.c +0 -0
index 79f1a30f3683d..2309196d8df31 100644
@@ -600,7 +600,13 @@ static char *dso__get_filename(struct dso *dso, const char *root_dir,
 		size_t len = sizeof(newpath);
 
 		if (dso__decompress_kmodule_path(dso, name, newpath, len) < 0) {

tools/perf/util/dso.c static char *dso__get_filename(struct dso *dso, const char *root_dir, · line 611

Quoted 10 lines
tools/perf/util/dso.c +7 -1
-			errno = *dso__load_errno(dso);
+			/*
+			 * Use a standard errno value, not the negative custom
+			 * DSO_LOAD_ERRNO stored in dso__load_errno(dso):
+			 * __open_dso() computes fd = -errno, so a negative
+			 * errno produces a positive fd that looks valid.
+			 */
+			errno = EIO;
 			goto out;
 		}

[Severity: High]
This isn't a bug introduced by this patch, but since we are fixing errno handling here, could a similar issue occur earlier in dso__get_filename when it fails without explicitly setting errno?

If stat succeeds on a directory, errno is 0 but S_ISREG fails. is_regular_file returns false and errno remains 0, jumping to out:

tools/perf/util/dso.c:dso__get_filename() {
    ...
	if (!is_regular_file(name)) {
		char *new_name;

		if (errno != ENOENT || dso__nsinfo(dso) == NULL)
			goto out;
    ...
}

In this case, __open_dso computes fd = -errno, resulting in fd = 0:

tools/perf/util/dso.c:__open_dso() {
    ...
	name = dso__get_filename(dso, machine ? machine->root_dir : "", &decomp);
	if (name)
		fd = do_open(name);
	else
		fd = -errno;
    ...
}

Then open_dso incorrectly treats this as a successful open and registers standard input as the DSO data file descriptor:

tools/perf/util/dso.c:open_dso() {
    ...
	fd = __open_dso(dso, machine);
	if (fd >= 0) {
		dso__list_add(dso);
    ...
}

Could this lead to resource hijacking or blocking on stdin?

[Severity: High]
This is another pre-existing issue, but I noticed a potential race condition with the DSO lists.

Under REFCNT_CHECKING, dso__list_del removes a DSO from the open list and temporarily unlocks dso__data_open_lock before decrementing the counter:

tools/perf/util/dso.c:dso__list_del() {
    ...
	list_del_init(&dso__data(dso)->open_entry);
#ifdef REFCNT_CHECKING
	mutex_unlock(dso__data_open_lock());
	dso__put(dso__data(dso)->dso);
	mutex_lock(dso__data_open_lock());
#endif
	WARN_ONCE(dso__data_open_cnt <= 0, ...);
	dso__data_open_cnt--;
    ...
}

If a concurrent thread calls close_first_dso, it checks the counter rather than checking if the list is empty:

tools/perf/util/dso.c:close_first_dso() {
    ...
	dso_data = list_first_entry(&dso__data_open, struct dso_data, open_entry);
#ifdef REFCNT_CHECKING
	dso = dso_data->dso;
#else
	...
	close_dso(dso);
    ...
}

Because dso__data_open_cnt is greater than zero even when the list is empty, list_first_entry computes an invalid pointer from the list head. This causes dso_data->dso to evaluate to the address of dso__data_open_cnt, treating the integer as a struct pointer and dereferencing it.

Can this result in a concurrent segmentation fault under REFCNT_CHECKING builds?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=9
#10 lore raw

Re: [PATCH 5/9] perf cs-etm: Bounds-check CPU in cs_etm__get_queue()

In reply to Arnaldo Carvalho de Melo
Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:

Pre-existing issues:
- [High] Integer overflow in auxtrace_queues__queue_buffer() leads to out-of-bounds memory access.
- [Medium] Null pointer dereference in cs_etm__dump_event() when processing a crafted perf.data file.
- [Medium] Out-of-bounds read in cs_etm__create_meta_blk() when parsing CPU metadata blocks. --

commit 6ff616c3e2c2577ec55eda8099842f63dcc43ed7
Author: Arnaldo Carvalho de Melo <[email protected]>

perf cs-etm: Bounds-check CPU in cs_etm__get_queue()

cs_etm__get_queue() indexes etm->queues.queue_array[cpu] without validating that cpu is within nr_queues. When processing AUX_OUTPUT_HW_ID events, the cpu value comes from untrusted perf.data trace payload and flows through cs_etm__process_trace_id_v0_1() and cs_etm__queue_aux_fragment() without bounds checking, allowing an out-of-bounds read with a crafted file.

Quoted 4 lines
tools/perf/util/cs-etm.c +0 -0
index d121c8f22028d..5d0664ff73b79 100644

[ ... ]

tools/perf/util/cs-etm.c static int cs_etm__map_trace_id_v0(struct cs_etm_auxtrace *etm, u8 trace_chan_id · line 329

Quoted 7 lines
tools/perf/util/cs-etm.c +3 -0
@@ -318,6 +324,9 @@ static int cs_etm__map_trace_id_v0(struct cs_etm_auxtrace *etm, u8 trace_chan_id
 		int ret;
 
 		etmq = etm->queues.queue_array[i].priv;
+		if (!etmq)
+			continue;
+

[Severity: Medium]
This is a pre-existing issue, but since we are adding NULL checks for zero-initialized queues here, could this also cause a null pointer dereference in cs_etm__dump_event()?

If dump_trace is enabled, cs_etm__process_auxtrace_event() and dump_queued_data() can pass the new queue's priv pointer directly:

tools/perf/util/cs-etm.c:cs_etm__process_auxtrace_event() {
    ...
    cs_etm__dump_event(etm->queues.queue_array[idx].priv, buffer);
    ...
}

And cs_etm__dump_event() unconditionally dereferences it:

tools/perf/util/cs-etm.c:cs_etm__dump_event() {
    ...
    cs_etm_decoder__get_name(etmq->decoder)
    ...
}

Should cs_etm__dump_event() also have a NULL check for etmq?

[ ... ]

tools/perf/util/cs-etm.c static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o · line 3167

Quoted 7 lines
tools/perf/util/cs-etm.c +3 -0
@@ -3144,6 +3162,9 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o
 	    aux_offset + aux_size <= auxtrace_event->offset + auxtrace_event->size) {
 		struct cs_etm_queue *etmq = cs_etm__get_queue(etm, auxtrace_event->cpu);
 
+		if (!etmq)
+			return -EINVAL;
+

[Severity: High]
This is a pre-existing issue, but I noticed a potential integer overflow in auxtrace_queues__queue_buffer() that could lead to out-of-bounds memory access.

When processing an untrusted perf.data file, if the idx field of a PERF_RECORD_AUXTRACE event is UINT_MAX, does the idx + 1 calculation overflow to 0?

tools/perf/util/auxtrace.c:auxtrace_queues__queue_buffer() {
    ...
	if (idx >= queues->nr_queues) {
		err = auxtrace_queues__grow(queues, idx + 1);
		if (err)
			return err;
	}

	queue = &queues->queue_array[idx];
    ...
}

Since auxtrace_queues__grow() evaluates nr_queues < new_nr_queues as false when new_nr_queues is 0, it returns without growing the array. Does this cause the subsequent array access at queues->queue_array[UINT_MAX] to write out of bounds?

[Severity: Medium]
This is a pre-existing issue, but there appears to be a potential out-of-bounds read in cs_etm__create_meta_blk() when parsing CPU metadata blocks.

When parsing a PERF_RECORD_AUXTRACE_INFO event, does cs_etm__process_auxtrace_info_full() validate that the payload contains enough bytes before calling cs_etm__create_meta_blk()?

tools/perf/util/cs-etm.c:cs_etm__process_auxtrace_info_full() {
    ...
		metadata[j] =
			cs_etm__create_meta_blk(ptr, &i,
						CS_ETMV4_PRIV_MAX,
						CS_ETMV4_NR_TRC_PARAMS_V0);
    ...
}

A crafted file with a large num_cpu and small auxtrace_info.header.size might bypass the num_cpu > priv_size / 8 check. cs_etm__create_meta_blk() then unconditionally reads up to CS_ETMV4_NR_TRC_PARAMS_V0 elements:

tools/perf/util/cs-etm.c:cs_etm__create_meta_blk() {
    ...
    metadata[k + 1] = buff_in[i + k];
    ...
}

Since the length validation (i * 8 != priv_size) only happens after the loop finishes, can this cause a heap buffer over-read of buff_in?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=5
#2 lore raw

Re: [PATCH 1/9] perf machine: Propagate machine__init() error to callers

In reply to Arnaldo Carvalho de Melo
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] The return value of the newly fallible `machine__init()` function is ignored in tests, leading…

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] The return value of the newly fallible machine__init() function is ignored in tests, leading to potential null pointer dereferences on initialization failure. --

commit 75ef0ba718fac4d2482673c59e4b0fde21e0bfe2
Author: Arnaldo Carvalho de Melo <[email protected]>

perf machine: Propagate machine__init() error to callers

machine__init() always returns 0 even when memory allocation fails, because commit 81f981d7ec43ed93 introduced an error cleanup path but left the final 'return 0' instead of 'return err'.

Fix by returning err, check the return value in __machine__new_host() which was ignoring it, and change machines__init() from void to int so it too can propagate the error.

Quoted 4 lines
tools/perf/util/machine.c +0 -0
index 31715366e29ff..9329d319bd033 100644

[ ... ]

tools/perf/util/machine.c int machine__init(struct machine *machine, const char *root_dir, pid_t pid) · line 131

Quoted 12 lines
tools/perf/util/machine.c +2 -2
@@ -124,11 +123,11 @@ int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
 
 out:
 	if (err) {
-		zfree(&machine->kmaps);
+		maps__zput(machine->kmaps);
 		zfree(&machine->root_dir);
 		zfree(&machine->mmap_name);
 	}
-	return 0;
+	return err;
 }

[Severity: Medium]
Since this patch changes machine__init() to return its internal error instead of returning 0 on failure, do we also need to update the direct callers in the tests to check this return value?

For example, test__kallsyms_split() ignores the return value:

tools/perf/tests/kallsyms-split.c:test__kallsyms_split() {
    ...
	machine__init(&m, root_dir, HOST_KERNEL_ID);
	if (machine__create_kernel_maps(&m) < 0) {
    ...
}

And test__vmlinux_matches_kallsyms() does as well:

tools/perf/tests/vmlinux-kallsyms.c:test__vmlinux_matches_kallsyms() {
    ...
	machine__init(&args.kallsyms, "", HOST_KERNEL_ID);
	machine__init(&vmlinux, "", HOST_KERNEL_ID);
    ...
}

If machine__init() fails, it cleans up and leaves machine.kmaps as NULL.

Could this lead to a NULL pointer dereference in functions like machine__create_kernel_maps() that expect kmaps to be valid?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#14 lore raw

Re: [PATCH 8/9] perf bpf: Validate array presence before casting BPF prog info pointers

In reply to Arnaldo Carvalho de Melo
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] The patch adds array bitmask validation for JITED_KSYMS and JITED_FUNC_LENS in __bpf_event__print_bpf_prog_info(), but fails to include PROG_TAGS and FUNC_INFO. This leaves the function vulnerable to arbitrary pointer dereferences when calling synthesize_bpf_prog_name().
- [High] The patch completely misses adding array bitmask validations to symbol__disassemble_bpf_libbfd() in libbfd.c, leaving it vulnerable to unconverted pointer dereferences.
--

commit 07052c4321d3973c606cfca5a84ae676ff46cd05
Author: Arnaldo Carvalho de Melo <[email protected]>

perf bpf: Validate array presence before casting BPF prog info pointers

This patch adds array bitmask validation to several perf.data processing paths to ensure that bpf_prog_info file offsets are properly converted to valid pointers before being dereferenced.

Quoted 8 lines

Add array bitmask validation to all perf.data processing paths:

- __bpf_event__print_bpf_prog_info(): check JITED_KSYMS and JITED_FUNC_LENS (changed to take struct perf_bpil *)
- machine__process_bpf_event_load(): check JITED_KSYMS
- bpf_read(): check JITED_INSNS before memcpy from jited_prog_insns
- dso__disassemble_filename(): check JITED_INSNS before returning jited_prog_insns pointer

[Severity: High]
This isn't a bug introduced by this patch, but does this miss the processing path in symbol__disassemble_bpf_libbfd() in tools/perf/util/libbfd.c?

Looking at symbol__disassemble_bpf_libbfd(), it also casts and dereferences jited_prog_insns and jited_ksyms without verifying their corresponding PERF_BPIL_ bits are set:

tools/perf/util/libbfd.c:symbol__disassemble_bpf_libbfd() {
    ...
    info.buffer = (void *)(uintptr_t)(info_linear->info.jited_prog_insns);
    ...
    addr = pc + ((u64 *)(uintptr_t)(info_linear->info.jited_ksyms))[sub_id];
    ...
}

Can a crafted perf.data missing these arrays still cause arbitrary memory reads or segmentation faults during offline disassembly via perf annotate?

Quoted 4 lines
tools/perf/util/bpf-event.c +0 -0
index 57d53ba848359..fa3ebc8ea7f09 100644

[ ... ]

Quoted 4 lines
tools/perf/util/bpf-event.c +0 -0
@@ -959,12 +963,15 @@ int evlist__add_bpf_sb_event(struct evlist *evlist, struct perf_env *env)
 	return evlist__add_sb_event(evlist, &attr, bpf_event__sb_cb, env);
 }
 

tools/perf/util/bpf-event.c int evlist__add_bpf_sb_event(struct evlist *evlist, struct perf_env *env) · line 972

Quoted 10 lines
tools/perf/util/bpf-event.c +4 -3
-void __bpf_event__print_bpf_prog_info(struct bpf_prog_info *info,
+void __bpf_event__print_bpf_prog_info(struct perf_bpil *info_linear,
 				      struct perf_env *env,
 				      FILE *fp)
 {
-	__u32 *prog_lens = (__u32 *)(uintptr_t)(info->jited_func_lens);
-	__u64 *prog_addrs = (__u64 *)(uintptr_t)(info->jited_ksyms);
+	struct bpf_prog_info *info = &info_linear->info;
+	__u64 required_arrays = (1UL << PERF_BPIL_JITED_KSYMS) |
+				(1UL << PERF_BPIL_JITED_FUNC_LENS);

[Severity: High]
This is a pre-existing issue, but does this bitmask also need to include PERF_BPIL_PROG_TAGS and PERF_BPIL_FUNC_INFO?

Later in this function, synthesize_bpf_prog_name() is called, which accesses info->prog_tags without checking if it was converted from a file offset:

tools/perf/util/bpf-event.c:synthesize_bpf_prog_name() {
    ...
    name_len += snprintf_hex(buf + name_len, size - name_len,
                             prog_tags[sub_id], BPF_TAG_SIZE);
    ...
}

If a crafted perf.data omits PERF_BPIL_PROG_TAGS from the arrays bitmask but provides matching counts (e.g., nr_prog_tags == nr_jited_ksyms), could this dereference unconverted file offsets and cause a crash during perf report?

Quoted 3 lines
tools/perf/util/bpf-event.c +0 -0
 	char name[KSYM_NAME_LEN];
 	struct btf *btf = NULL;
 	u32 sub_prog_cnt, i;
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=8