The grace period used internally within tracepoint.c:release_probes()
uses call_rcu() to batch waiting for quiescence of old probe arrays,
rather than using the tracepoint_synchronize_unregister() which blocks
while waiting for quiescence.

With the introduction of faultable syscall tracepoints, this causes
use-after-free issues reproduced with syzkaller.

Fix this by introducing tracepoint_call_rcu(), which uses the
appropriate call_rcu() or call_rcu_tasks_trace() before invoking the
rcu_free_old_probes callback.

Use tracepoint_call_rcu() in bpf_link_free() for raw tracepoints as
well, which has the same problem for syscall tracepoints.

Reported-by: syzbot+b390c8062d8387b6272a@syzkaller.appspotmail.com
Fixes: a363d27cdbc2 ("tracing: Allow system call tracepoints to handle page faults")
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Michael Jeanson <mjeanson@efficios.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Yonghong Song <yhs@fb.com>
Cc: Paul E. McKenney <paulmck@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Andrii Nakryiko <andrii.nakryiko@gmail.com>
Cc: bpf@vger.kernel.org
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Jordan Rife <jrife@google.com>
---
Changes since v0:
- Introduce tracepoint_call_rcu(),
- Fix bpf_link_free() use of call_rcu as well.
---
 include/linux/tracepoint.h |  9 +++++++++
 kernel/bpf/syscall.c       | 23 +++++++++++++++--------
 kernel/tracepoint.c        | 22 ++++++++++++++++++----
 3 files changed, 42 insertions(+), 12 deletions(-)

diff --git a/include/linux/tracepoint.h b/include/linux/tracepoint.h
index 0dc67fad706c..45025d6b2dd6 100644
--- a/include/linux/tracepoint.h
+++ b/include/linux/tracepoint.h
@@ -104,6 +104,8 @@ void for_each_tracepoint_in_module(struct module *mod,
  * tracepoint_synchronize_unregister must be called between the last tracepoint
  * probe unregistration and the end of module exit to make sure there is no
  * caller executing a probe when it is freed.
+ * An alternative to tracepoint_synchronize_unregister() is to use
+ * tracepoint_call_rcu() for batched reclaim.
  */
 #ifdef CONFIG_TRACEPOINTS
 static inline void tracepoint_synchronize_unregister(void)
@@ -111,9 +113,16 @@ static inline void tracepoint_synchronize_unregister(void)
 	synchronize_rcu_tasks_trace();
 	synchronize_rcu();
 }
+
+void tracepoint_call_rcu(struct tracepoint *tp, struct rcu_head *head,
+			 void (*callback)(struct rcu_head *head));
+
 #else
 static inline void tracepoint_synchronize_unregister(void)
 { }
+static inline void tracepoint_call_rcu(struct tracepoint *tp, struct rcu_head *head,
+				       void (*callback)(struct rcu_head *head))
+{ }
 #endif
 
 #ifdef CONFIG_HAVE_SYSCALL_TRACEPOINTS
diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
index 59de664e580d..1191dc1d4206 100644
--- a/kernel/bpf/syscall.c
+++ b/kernel/bpf/syscall.c
@@ -3006,14 +3006,21 @@ static void bpf_link_free(struct bpf_link *link)
 		bpf_prog_put(link->prog);
 	}
 	if (ops->dealloc_deferred) {
-		/* schedule BPF link deallocation; if underlying BPF program
-		 * is sleepable, we need to first wait for RCU tasks trace
-		 * sync, then go through "classic" RCU grace period
-		 */
-		if (sleepable)
-			call_rcu_tasks_trace(&link->rcu, bpf_link_defer_dealloc_mult_rcu_gp);
-		else
-			call_rcu(&link->rcu, bpf_link_defer_dealloc_rcu_gp);
+		if (link->type == BPF_LINK_TYPE_RAW_TRACEPOINT) {
+		        struct bpf_raw_tp_link *raw_tp =
+				container_of(link, struct bpf_raw_tp_link, link);
+
+			tracepoint_call_rcu(raw_tp->btp->tp, &link->rcu, bpf_link_defer_dealloc_rcu_gp);
+		} else {
+			/* schedule BPF link deallocation; if underlying BPF program
+			 * is sleepable, we need to first wait for RCU tasks trace
+			 * sync, then go through "classic" RCU grace period
+			 */
+			if (sleepable)
+				call_rcu_tasks_trace(&link->rcu, bpf_link_defer_dealloc_mult_rcu_gp);
+			else
+				call_rcu(&link->rcu, bpf_link_defer_dealloc_rcu_gp);
+		}
 	} else if (ops->dealloc)
 		ops->dealloc(link);
 }
diff --git a/kernel/tracepoint.c b/kernel/tracepoint.c
index 6474e2cf22c9..ef60c5484eda 100644
--- a/kernel/tracepoint.c
+++ b/kernel/tracepoint.c
@@ -106,13 +106,27 @@ static void rcu_free_old_probes(struct rcu_head *head)
 	kfree(container_of(head, struct tp_probes, rcu));
 }
 
-static inline void release_probes(struct tracepoint_func *old)
+static bool tracepoint_is_syscall(struct tracepoint *tp)
+{
+	return !strcmp(tp->name, "sys_enter") || !strcmp(tp->name, "sys_exit");
+}
+
+void tracepoint_call_rcu(struct tracepoint *tp, struct rcu_head *head,
+			 void (*callback)(struct rcu_head *head))
+{
+	if (tracepoint_is_syscall(tp))
+		call_rcu_tasks_trace(head, callback);
+	else
+		call_rcu(head, callback);
+}
+
+static inline void release_probes(struct tracepoint *tp, struct tracepoint_func *old)
 {
 	if (old) {
 		struct tp_probes *tp_probes = container_of(old,
 			struct tp_probes, probes[0]);
 
-		call_rcu(&tp_probes->rcu, rcu_free_old_probes);
+		tracepoint_call_rcu(tp, &tp_probes->rcu, rcu_free_old_probes);
 	}
 }
 
@@ -334,7 +348,7 @@ static int tracepoint_add_func(struct tracepoint *tp,
 		break;
 	}
 
-	release_probes(old);
+	release_probes(tp, old);
 	return 0;
 }
 
@@ -406,7 +420,7 @@ static int tracepoint_remove_func(struct tracepoint *tp,
 		WARN_ON_ONCE(1);
 		break;
 	}
-	release_probes(old);
+	release_probes(tp, old);
 	return 0;
 }
 
-- 
2.39.5

> diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> index 59de664e580d..1191dc1d4206 100644
> --- a/kernel/bpf/syscall.c
> +++ b/kernel/bpf/syscall.c
> @@ -3006,14 +3006,21 @@ static void bpf_link_free(struct bpf_link *link)
>                 bpf_prog_put(link->prog);

I think we would need the same treatment with bpf_prog_put here.
Something like,

tracepoint_call_rcu(raw_tp->btp->tp, &link->prog->aux->rcu,
		    bpf_link_defer_bpf_prog_put);

static void bpf_link_defer_bpf_prog_put(struct rcu_head *rcu)
{
	struct bpf_prog_aux *aux = container_of(rcu, struct bpf_prog_aux, rcu);
	bpf_prog_put(aux->prox);
}

Alternatively, some context would need to be passed down to
__bpf_prog_put_noref via the call to bpf_prog_put so it can choose
whether or not to use call_rcu or call_rcu_tasks_trace.

> -static inline void release_probes(struct tracepoint_func *old)
> +static bool tracepoint_is_syscall(struct tracepoint *tp)
> +{
> +       return !strcmp(tp->name, "sys_enter") || !strcmp(tp->name, "sys_exit");
> +}

I'm curious if it might be better to add some field to struct
tracepoint like "sleepable" rather than adding a special case here
based on the name? Of course, if it's only ever going to be these
two cases then maybe adding a new field doesn't make sense.

-Jordan

On 2024-10-25 15:08, Jordan Rife wrote:
>> diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
>> index 59de664e580d..1191dc1d4206 100644
>> --- a/kernel/bpf/syscall.c
>> +++ b/kernel/bpf/syscall.c
>> @@ -3006,14 +3006,21 @@ static void bpf_link_free(struct bpf_link *link)
>>                  bpf_prog_put(link->prog);
> 
> I think we would need the same treatment with bpf_prog_put here.
> Something like,
> 
> tracepoint_call_rcu(raw_tp->btp->tp, &link->prog->aux->rcu,
> 		    bpf_link_defer_bpf_prog_put);
> 
> static void bpf_link_defer_bpf_prog_put(struct rcu_head *rcu)
> {
> 	struct bpf_prog_aux *aux = container_of(rcu, struct bpf_prog_aux, rcu);
> 	bpf_prog_put(aux->prox);
> }

Sure, I'll add this in a v2.

> 
> Alternatively, some context would need to be passed down to
> __bpf_prog_put_noref via the call to bpf_prog_put so it can choose
> whether or not to use call_rcu or call_rcu_tasks_trace.

Also possible, but more cumbersome.

> 
>> -static inline void release_probes(struct tracepoint_func *old)
>> +static bool tracepoint_is_syscall(struct tracepoint *tp)
>> +{
>> +       return !strcmp(tp->name, "sys_enter") || !strcmp(tp->name, "sys_exit");
>> +}
> 
> I'm curious if it might be better to add some field to struct
> tracepoint like "sleepable" rather than adding a special case here
> based on the name? Of course, if it's only ever going to be these
> two cases then maybe adding a new field doesn't make sense.

I know Steven is reluctant to bloat the tracepoint struct because there
are lots of tracepoint instances (thousands). So for now I thought that
just comparing the name would be a good start.

We can eventually go a different route as well: introduce a section just
to put the syscall tracepoints, and compare the struct tracepoint
pointers to the section begin/end range. But it's rather complex
for what should remain a simple fix.

Thanks,

Mathieu


> 
> -Jordan

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

On Fri, 25 Oct 2024 15:38:48 -0400
Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:

> > I'm curious if it might be better to add some field to struct
> > tracepoint like "sleepable" rather than adding a special case here
> > based on the name? Of course, if it's only ever going to be these
> > two cases then maybe adding a new field doesn't make sense.  
> 
> I know Steven is reluctant to bloat the tracepoint struct because there
> are lots of tracepoint instances (thousands). So for now I thought that
> just comparing the name would be a good start.

You are correct. I really trying to keep the footprint of
tracepoints/events down.

> 
> We can eventually go a different route as well: introduce a section just
> to put the syscall tracepoints, and compare the struct tracepoint
> pointers to the section begin/end range. But it's rather complex
> for what should remain a simple fix.

A separate section could work.

-- Steve

On 2024-10-26 03:13, Steven Rostedt wrote:
> On Fri, 25 Oct 2024 15:38:48 -0400
> Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:
> 
>>> I'm curious if it might be better to add some field to struct
>>> tracepoint like "sleepable" rather than adding a special case here
>>> based on the name? Of course, if it's only ever going to be these
>>> two cases then maybe adding a new field doesn't make sense.
>>
>> I know Steven is reluctant to bloat the tracepoint struct because there
>> are lots of tracepoint instances (thousands). So for now I thought that
>> just comparing the name would be a good start.
> 
> You are correct. I really trying to keep the footprint of
> tracepoints/events down.
> 
>>
>> We can eventually go a different route as well: introduce a section just
>> to put the syscall tracepoints, and compare the struct tracepoint
>> pointers to the section begin/end range. But it's rather complex
>> for what should remain a simple fix.
> 
> A separate section could work.

I have another approach to suggest: it shrinks the
size of struct tracepoint from 80 bytes down to 72 bytes
on x86-64, we don't have to do any section/linker
script trickery, and it's extensible for future flags:

struct static_key {
         int enabled;
         void *p;
};

struct static_key_false {
         struct static_key key;
};

struct static_call_key {
         void *func;
         void *p;
};

struct tracepoint {
         const char *name;               /* Tracepoint name */
         struct static_key_false key;
         struct static_call_key *static_call_key;
         void *static_call_tramp;
         void *iterator;
         void *probestub;
         void *funcs;
         /* Flags. */
         unsigned int regfunc:1,
                      syscall:1;
};

struct tracepoint_regfunc {
         struct tracepoint tp;
         int (*regfunc)(void);
         void (*unregfunc)(void);
};

Basically, a tracepoint with regfunc would define a
struct tracepoint_regfunc rather than a struct tracepoint.
So we remove both regfunc and unregfunc NULL pointers in
the common case, which gives us plenty of room for flags.

When we want to access the regfunc/unregfunc from
a struct tracepoint, we check the regfunc flag, and
if set, we can use container_of() to get the struct
tracepoint_regfunc.

Thoughts ?

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

On 2024-10-26 10:25, Mathieu Desnoyers wrote:
> On 2024-10-26 03:13, Steven Rostedt wrote:
>> On Fri, 25 Oct 2024 15:38:48 -0400
>> Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:
>>
>>>> I'm curious if it might be better to add some field to struct
>>>> tracepoint like "sleepable" rather than adding a special case here
>>>> based on the name? Of course, if it's only ever going to be these
>>>> two cases then maybe adding a new field doesn't make sense.
>>>
>>> I know Steven is reluctant to bloat the tracepoint struct because there
>>> are lots of tracepoint instances (thousands). So for now I thought that
>>> just comparing the name would be a good start.
>>
>> You are correct. I really trying to keep the footprint of
>> tracepoints/events down.
>>
>>>
>>> We can eventually go a different route as well: introduce a section just
>>> to put the syscall tracepoints, and compare the struct tracepoint
>>> pointers to the section begin/end range. But it's rather complex
>>> for what should remain a simple fix.
>>
>> A separate section could work.
> 
> I have another approach to suggest: it shrinks the
> size of struct tracepoint from 80 bytes down to 72 bytes
> on x86-64, we don't have to do any section/linker
> script trickery, and it's extensible for future flags:
> 
> struct static_key {
>          int enabled;
>          void *p;
> };
> 
> struct static_key_false {
>          struct static_key key;
> };
> 
> struct static_call_key {
>          void *func;
>          void *p;
> };
> 
> struct tracepoint {
>          const char *name;               /* Tracepoint name */
>          struct static_key_false key;
>          struct static_call_key *static_call_key;
>          void *static_call_tramp;
>          void *iterator;
>          void *probestub;
>          void *funcs;
>          /* Flags. */
>          unsigned int regfunc:1,
>                       syscall:1;
> };
> 
> struct tracepoint_regfunc {
>          struct tracepoint tp;
>          int (*regfunc)(void);
>          void (*unregfunc)(void);
> };
> 
> Basically, a tracepoint with regfunc would define a
> struct tracepoint_regfunc rather than a struct tracepoint.
> So we remove both regfunc and unregfunc NULL pointers in
> the common case, which gives us plenty of room for flags.
> 
> When we want to access the regfunc/unregfunc from
> a struct tracepoint, we check the regfunc flag, and
> if set, we can use container_of() to get the struct
> tracepoint_regfunc.

Actually I can achieve the same space saving with fewer
changes like this:

struct tracepoint_ext {
     void *regfunc;
     void *unregfunc;
     /* Flags. */
     unsigned int syscall:1;
}

struct tracepoint {
         const char *name;               /* Tracepoint name */
         struct static_key_false key;
         struct static_call_key *static_call_key;
         void *static_call_tramp;
         void *iterator;
         void *probestub;
         void *funcs;
         struct tracepoint_ext *ext;
};

Thanks,

Mathieu

> 
> Thoughts ?
> 
> Thanks,
> 
> Mathieu
> 

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com