Skip to contentSkip to conversation

perf stat: events inheritance can break task targets

BUG Alexander Yarygin 3 participants 4 messages Jul 7, 2014 – Jul 8, 2014 Tracing

lore
Share & tools
lore
#0 lore raw

perf stat can block pthread_create() for a multithreaded userspace process (i.e. qemu) when:
- process is running with non-root privileges
- perf stat is running as root with trace events in -e option
- it is attached to the process's pid.

Here is a simple test scenario:

~$ cat test.c
#include <stdio.h>
#include <pthread.h>
#include <unistd.h>
#include <string.h>

#define THREADS 50

static pthread_t threads[THREADS];

void *loop()

{
    while(1);
}

int main()

{
    for (int i = 0; i < THREADS; i++) {

        int err = pthread_create(&threads[i], NULL, loop, 0);

        if (!err)
            printf("thread created: %i\n", i);
        else
            printf("couldn't create thread %i: %s\n", i, strerror(err));

        sleep(1);
    }

    return 0;
}

~$ gcc test.c -lpthread -std=c99 -o test
~$ ./test
thread created: 0
thread created: 1

# now perf is running:
# ~$ sudo perf stat -e "kvm:*" -p `pidof test`
couldn't create thread 2: Operation not permitted
couldn't create thread 3: Operation not permitted
couldn't create thread 4: Operation not permitted
# here is perf was stopped
thread created: 5
thread created: 6
^C

When perf is running, every invoke of pthread_create() returns -EPERM.

On the kernel side, copy_process() creates a task, scheduled it, than perf_event_init_task() (kernel/events/core.c) returns an error, and the kernel cleans task's resources.

It looks like child process doesn't have access to trace events, so perf_trace_event_perm() (kernel/trace/trace_event_perf.c) returns -EPERM:

static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
struct perf_event *p_event)
{
...

    /*
     * ...otherwise raw tracepoint data can be a severe data leak,
     * only allow root to have these.
     */
    if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
        return -EPERM;

...
}

If we explicitly use the --no-inherit option, payload wouldn't die.

#1 lore raw

On Mon, Jul 07, 2014 at 08:41:57PM +0400, Alexander Yarygin wrote:

SNIP

Quoted 16 lines

^C

When perf is running, every invoke of pthread_create() returns -EPERM.

On the kernel side, copy_process() creates a task, scheduled it, than perf_event_init_task() (kernel/events/core.c) returns an error, and the kernel cleans task's resources.

It looks like child process doesn't have access to trace events, so perf_trace_event_perm() (kernel/trace/trace_event_perf.c) returns -EPERM:

static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
struct perf_event *p_event)
{
...

Quoted 8 lines
    /*
     * ...otherwise raw tracepoint data can be a severe data leak,
     * only allow root to have these.
     */
    if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
        return -EPERM;

...
}

right, so normaly it's done the check is done in current context via syscall and the current task is the tracer

while in inheriting we are checking the tracee here, we want to check the owner instead like in attached patch.. totaly untested, just to ilustrate the point

we might have same issue for other event types

jirka

---

kernel/trace/trace_event_perf.c +5 -2
index 5d12bb4..b44184b 100644
@@ -24,6 +24,9 @@ static int	total_ref_count;
 static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
 				 struct perf_event *p_event)
 {
+	struct task_struct owner = p_event->parent ? p_event->parent->owner :
+						     p_event->owner;
+
 	if (tp_event->perf_perm) {
 		int ret = tp_event->perf_perm(tp_event, p_event);
 		if (ret)
@@ -32,7 +35,7 @@ static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
 
 	/* The ftrace function trace is allowed only for root. */
 	if (ftrace_event_is_function(tp_event)) {
-		if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
+		if (perf_paranoid_tracepoint_raw() && !has_capability(owner, CAP_SYS_ADMIN))
 			return -EPERM;
 
 		/*
@@ -65,7 +68,7 @@ static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
 	 * ...otherwise raw tracepoint data can be a severe data leak,
 	 * only allow root to have these.
 	 */
-	if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
+	if (perf_paranoid_tracepoint_raw() && !has_capability(owner, CAP_SYS_ADMIN))
 		return -EPERM;
 
 	return 0;
#2 lore raw
Quoted 24 lines · Jiri Olsa
kernel/trace/trace_event_perf.c +4 -1
index 5d12bb4..b44184b 100644
@@ -24,6 +24,9 @@ static int	total_ref_count;
 static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
 				 struct perf_event *p_event)
 {
+	struct task_struct owner = p_event->parent ? p_event->parent->owner :
+						     p_event->owner;
+
 	if (tp_event->perf_perm) {
 		int ret = tp_event->perf_perm(tp_event, p_event);
 		if (ret)
@@ -32,7 +35,7 @@ static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
 
 	/* The ftrace function trace is allowed only for root. */
 	if (ftrace_event_is_function(tp_event)) {
-		if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
+		if (perf_paranoid_tracepoint_raw() && !has_capability(owner, CAP_SYS_ADMIN))
 			return -EPERM;
 
 		/*

kernel/trace/trace_event_perf.c static int perf_trace_event_perm(struct ftrace_event_call *tp_event, · line 74

Quoted 8 lines
kernel/trace/trace_event_perf.c +1 -1
@@ -65,7 +68,7 @@ static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
 	 * ...otherwise raw tracepoint data can be a severe data leak,
 	 * only allow root to have these.
 	 */
-	if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
+	if (perf_paranoid_tracepoint_raw() && !has_capability(owner, CAP_SYS_ADMIN))
 		return -EPERM;
 
 	return 0;

You need to either hold rcu_read_lock() or otherwise ensure the owner is still valid.

#3 lore raw
Quoted 32 lines · Peter Zijlstra
Quoted 31 lines · Jiri Olsa
kernel/trace/trace_event_perf.c +5 -2
index 5d12bb4..b44184b 100644
@@ -24,6 +24,9 @@ static int	total_ref_count;
 static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
 				 struct perf_event *p_event)
 {
+	struct task_struct owner = p_event->parent ? p_event->parent->owner :
+						     p_event->owner;
+
 	if (tp_event->perf_perm) {
 		int ret = tp_event->perf_perm(tp_event, p_event);
 		if (ret)
@@ -32,7 +35,7 @@ static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
 
 	/* The ftrace function trace is allowed only for root. */
 	if (ftrace_event_is_function(tp_event)) {
-		if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
+		if (perf_paranoid_tracepoint_raw() && !has_capability(owner, CAP_SYS_ADMIN))
 			return -EPERM;
 
 		/*
@@ -65,7 +68,7 @@ static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
 	 * ...otherwise raw tracepoint data can be a severe data leak,
 	 * only allow root to have these.
 	 */
-	if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
+	if (perf_paranoid_tracepoint_raw() && !has_capability(owner, CAP_SYS_ADMIN))
 		return -EPERM;
 
Quoted 4 lines

kernel/trace/trace_event_perf.c static int perf_trace_event_perm(struct ftrace_event_call *tp_event, · line 74

Quoted 1 line
kernel/trace/trace_event_perf.c +0 -0
 	return 0;

You need to either hold rcu_read_lock() or otherwise ensure the owner is still valid.

ok, I'll resend

thanks,
jirka