[PATCH bpf] bpf: Take module reference for ip in module code

Subsystems: bpf [core], bpf [general] (safe dynamic programs and tools), the rest

STALE1960d

4 messages, 3 authors, 2021-03-24 · open the first message on its own page

[PATCH bpf] bpf: Take module reference for ip in module code

From: Jiri Olsa <jolsa@kernel.org>
Date: 2021-03-23 21:16:48

Currently module can be unloaded even if there's a trampoline
register in it. It's easily reproduced by running in parallel:

  # while :; do ./test_progs -t module_attach; done
  # while :; do ./test_progs -t fentry_test; done

Taking the module reference in case the trampoline's ip is
within the module code. Releasing it when the trampoline's
ip is unregistered.

Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
 kernel/bpf/trampoline.c | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)

diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
index 1f3a4be4b175..f6cb179842b2 100644
--- a/kernel/bpf/trampoline.c
+++ b/kernel/bpf/trampoline.c

@@ -87,6 +87,27 @@ static struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
 	return tr;
 }
 
+static struct module *ip_module_get(unsigned long ip)
+{
+	struct module *mod;
+	int err = 0;
+
+	preempt_disable();
+	mod = __module_text_address(ip);
+	if (mod && !try_module_get(mod))
+		err = -ENOENT;
+	preempt_enable();
+	return err ? ERR_PTR(err) : mod;
+}
+
+static void ip_module_put(unsigned long ip)
+{
+	struct module *mod = __module_text_address(ip);
+
+	if (mod)
+		module_put(mod);
+}
+
 static int is_ftrace_location(void *ip)
 {
 	long addr;

@@ -108,6 +129,9 @@ static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr)
 		ret = unregister_ftrace_direct((long)ip, (long)old_addr);
 	else
 		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL);
+
+	if (!ret)
+		ip_module_put((unsigned long) ip);
 	return ret;
 }

@@ -126,6 +150,7 @@ static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_ad
 /* first time registering */
 static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
 {
+	struct module *mod;
 	void *ip = tr->func.addr;
 	int ret;

@@ -134,10 +159,17 @@ static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
 		return ret;
 	tr->func.ftrace_managed = ret;
 
+	mod = ip_module_get((unsigned long) ip);
+	if (IS_ERR(mod))
+		return -ENOENT;
+
 	if (tr->func.ftrace_managed)
 		ret = register_ftrace_direct((long)ip, (long)new_addr);
 	else
 		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr);
+
+	if (ret)
+		module_put(mod);
 	return ret;
 }

-- 
2.30.2

Re: [PATCH bpf] bpf: Take module reference for ip in module code

From: Alexei Starovoitov <hidden>
Date: 2021-03-24 01:23:21

On Tue, Mar 23, 2021 at 10:15:33PM +0100, Jiri Olsa wrote:

quoted hunk

Currently module can be unloaded even if there's a trampoline
register in it. It's easily reproduced by running in parallel:

  # while :; do ./test_progs -t module_attach; done
  # while :; do ./test_progs -t fentry_test; done

Taking the module reference in case the trampoline's ip is
within the module code. Releasing it when the trampoline's
ip is unregistered.

Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
 kernel/bpf/trampoline.c | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)

diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
index 1f3a4be4b175..f6cb179842b2 100644
--- a/kernel/bpf/trampoline.c
+++ b/kernel/bpf/trampoline.c

@@ -87,6 +87,27 @@ static struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
 	return tr;
 }
 
+static struct module *ip_module_get(unsigned long ip)
+{
+	struct module *mod;
+	int err = 0;
+
+	preempt_disable();
+	mod = __module_text_address(ip);
+	if (mod && !try_module_get(mod))
+		err = -ENOENT;
+	preempt_enable();
+	return err ? ERR_PTR(err) : mod;
+}
+
+static void ip_module_put(unsigned long ip)
+{
+	struct module *mod = __module_text_address(ip);

Conceptually looks correct, but how did you test it?!
Just doing your reproducer:
while :; do ./test_progs -t module_attach; done & while :; do ./test_progs -t fentry_test; done

I immediately hit:
[   19.461162] WARNING: CPU: 1 PID: 232 at kernel/module.c:264 module_assert_mutex_or_preempt+0x2e/0x40
[   19.477126] Call Trace:
[   19.477464]  __module_address+0x28/0xf0
[   19.477865]  ? __bpf_trace_bpf_testmod_test_write_bare+0x10/0x10 [bpf_testmod]
[   19.478711]  __module_text_address+0xe/0x60
[   19.479156]  bpf_trampoline_update+0x2ff/0x470

Which points to an obvious bug above.

How did you debug it to this module going away issue?
Why does test_progs -t fentry_test help to repro?
Or does it?
It doesn't touch anything in modules.

+
+	if (mod)
+		module_put(mod);

Re: [PATCH bpf] bpf: Take module reference for ip in module code

From: Jiri Olsa <hidden>
Date: 2021-03-24 11:32:30

On Tue, Mar 23, 2021 at 06:22:37PM -0700, Alexei Starovoitov wrote:

On Tue, Mar 23, 2021 at 10:15:33PM +0100, Jiri Olsa wrote:

quoted

Currently module can be unloaded even if there's a trampoline
register in it. It's easily reproduced by running in parallel:

  # while :; do ./test_progs -t module_attach; done
  # while :; do ./test_progs -t fentry_test; done

Taking the module reference in case the trampoline's ip is
within the module code. Releasing it when the trampoline's
ip is unregistered.

Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
 kernel/bpf/trampoline.c | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)

diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
index 1f3a4be4b175..f6cb179842b2 100644
--- a/kernel/bpf/trampoline.c
+++ b/kernel/bpf/trampoline.c

@@ -87,6 +87,27 @@ static struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
 	return tr;
 }
 
+static struct module *ip_module_get(unsigned long ip)
+{
+	struct module *mod;
+	int err = 0;
+
+	preempt_disable();
+	mod = __module_text_address(ip);
+	if (mod && !try_module_get(mod))
+		err = -ENOENT;
+	preempt_enable();
+	return err ? ERR_PTR(err) : mod;
+}
+
+static void ip_module_put(unsigned long ip)
+{
+	struct module *mod = __module_text_address(ip);

Conceptually looks correct, but how did you test it?!
Just doing your reproducer:
while :; do ./test_progs -t module_attach; done & while :; do ./test_progs -t fentry_test; done

I immediately hit:
[   19.461162] WARNING: CPU: 1 PID: 232 at kernel/module.c:264 module_assert_mutex_or_preempt+0x2e/0x40
[   19.477126] Call Trace:
[   19.477464]  __module_address+0x28/0xf0
[   19.477865]  ? __bpf_trace_bpf_testmod_test_write_bare+0x10/0x10 [bpf_testmod]
[   19.478711]  __module_text_address+0xe/0x60
[   19.479156]  bpf_trampoline_update+0x2ff/0x470

I don't have lockdep enabled.. ah the module_mutex is held
during module init, that's why all the code I was using as
a reference did not take it.. sorry, will fix

Which points to an obvious bug above.

How did you debug it to this module going away issue?
Why does test_progs -t fentry_test help to repro?
Or does it?
It doesn't touch anything in modules.

test_prog also loads/unloads that module, but it could be
just insmod/rmmod instead, will change

jirka

Re: [PATCH bpf] bpf: Take module reference for ip in module code

From: Jiri Olsa <hidden>
Date: 2021-03-24 13:48:58

On Wed, Mar 24, 2021 at 12:31:42PM +0100, Jiri Olsa wrote:

On Tue, Mar 23, 2021 at 06:22:37PM -0700, Alexei Starovoitov wrote:

quoted

On Tue, Mar 23, 2021 at 10:15:33PM +0100, Jiri Olsa wrote:

quoted

Currently module can be unloaded even if there's a trampoline
register in it. It's easily reproduced by running in parallel:

  # while :; do ./test_progs -t module_attach; done
  # while :; do ./test_progs -t fentry_test; done

Taking the module reference in case the trampoline's ip is
within the module code. Releasing it when the trampoline's
ip is unregistered.

Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
 kernel/bpf/trampoline.c | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)

diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
index 1f3a4be4b175..f6cb179842b2 100644
--- a/kernel/bpf/trampoline.c
+++ b/kernel/bpf/trampoline.c

@@ -87,6 +87,27 @@ static struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
 	return tr;
 }
 
+static struct module *ip_module_get(unsigned long ip)
+{
+	struct module *mod;
+	int err = 0;
+
+	preempt_disable();
+	mod = __module_text_address(ip);
+	if (mod && !try_module_get(mod))
+		err = -ENOENT;
+	preempt_enable();
+	return err ? ERR_PTR(err) : mod;
+}
+
+static void ip_module_put(unsigned long ip)
+{
+	struct module *mod = __module_text_address(ip);

Conceptually looks correct, but how did you test it?!
Just doing your reproducer:
while :; do ./test_progs -t module_attach; done & while :; do ./test_progs -t fentry_test; done

I immediately hit:
[   19.461162] WARNING: CPU: 1 PID: 232 at kernel/module.c:264 module_assert_mutex_or_preempt+0x2e/0x40
[   19.477126] Call Trace:
[   19.477464]  __module_address+0x28/0xf0
[   19.477865]  ? __bpf_trace_bpf_testmod_test_write_bare+0x10/0x10 [bpf_testmod]
[   19.478711]  __module_text_address+0xe/0x60
[   19.479156]  bpf_trampoline_update+0x2ff/0x470

I don't have lockdep enabled.. ah the module_mutex is held
during module init, that's why all the code I was using as
a reference did not take it.. sorry, will fix

ah it's the missing preempt_disable ;-) ok

jirka

quoted

Which points to an obvious bug above.

How did you debug it to this module going away issue?
Why does test_progs -t fentry_test help to repro?
Or does it?
It doesn't touch anything in modules.

test_prog also loads/unloads that module, but it could be
just insmod/rmmod instead, will change

jirka

`h`	back out one level
`j`	next message in thread
`k`	previous message in thread
`l`	drill in
`Esc`	close help / fold thread tree
`?`	toggle this help