| Message ID | 20250219161417.GA20851@redhat.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | yama: don't abuse rcu_read_lock/get_task_struct in yama_task_prctl() | expand |
On Wed, Feb 19, 2025 at 05:14:17PM +0100, Oleg Nesterov wrote: > current->group_leader is stable, no need to take rcu_read_lock() and do > get/put_task_struct(). Can you explain why this is true? In trying to figure this out again, it seems that the only way current->group_leader can vanish is if the entire process vanishes (fork or thread exec), in which case the "current" in this prctl can't be happening; this appears to be locked behind tsk->sighand->siglock ? -Kees > > Signed-off-by: Oleg Nesterov <oleg@redhat.com> > --- > security/yama/yama_lsm.c | 9 ++------- > 1 file changed, 2 insertions(+), 7 deletions(-) > > diff --git a/security/yama/yama_lsm.c b/security/yama/yama_lsm.c > index 1971710620c1..3d064dd4e03f 100644 > --- a/security/yama/yama_lsm.c > +++ b/security/yama/yama_lsm.c > @@ -222,7 +222,7 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, > unsigned long arg4, unsigned long arg5) > { > int rc = -ENOSYS; > - struct task_struct *myself = current; > + struct task_struct *myself; > > switch (option) { > case PR_SET_PTRACER: > @@ -232,11 +232,7 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, > * leader checking is handled later when walking the ancestry > * at the time of PTRACE_ATTACH check. > */ > - rcu_read_lock(); > - if (!thread_group_leader(myself)) > - myself = rcu_dereference(myself->group_leader); > - get_task_struct(myself); > - rcu_read_unlock(); > + myself = current->group_leader; > > if (arg2 == 0) { > yama_ptracer_del(NULL, myself); > @@ -255,7 +251,6 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, > } > } > > - put_task_struct(myself); > break; > } > > -- > 2.25.1.362.g51ebf55 > >
On 02/19, Kees Cook wrote: > > On Wed, Feb 19, 2025 at 05:14:17PM +0100, Oleg Nesterov wrote: > > current->group_leader is stable, no need to take rcu_read_lock() and do > > get/put_task_struct(). > > Can you explain why this is true? In trying to figure this out again, > it seems that the only way current->group_leader can vanish is if > the entire process vanishes (fork or thread exec), in which case the > "current" in this prctl can't be happening; this appears to be locked > behind tsk->sighand->siglock ? Well, almost, but this has nothing to do with tsk->sighand->siglock... task->group_leader can only be changed by thread exec, when a non leader thread does exec, see de_thread(). But de_thread() can't succeed and change ->group_leader until all other threads exit, see the "Kill all other threads in the thread group" code in de_thread(). The "current" task can't exit, so current->group_leader is stable. Note also that we already have a lot of current->group_leader users which don't use rcu/get_task_struct. That said, we have a lot of buggy users of tsk->group_leader when same_thread_group(tsk, current) != true ;) For example, sys_prlimit64(). And note that rcu_read_lock/get_task_struct can't help in this case. I am going to send some fixes. Oleg. > > -Kees > > > > > Signed-off-by: Oleg Nesterov <oleg@redhat.com> > > --- > > security/yama/yama_lsm.c | 9 ++------- > > 1 file changed, 2 insertions(+), 7 deletions(-) > > > > diff --git a/security/yama/yama_lsm.c b/security/yama/yama_lsm.c > > index 1971710620c1..3d064dd4e03f 100644 > > --- a/security/yama/yama_lsm.c > > +++ b/security/yama/yama_lsm.c > > @@ -222,7 +222,7 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, > > unsigned long arg4, unsigned long arg5) > > { > > int rc = -ENOSYS; > > - struct task_struct *myself = current; > > + struct task_struct *myself; > > > > switch (option) { > > case PR_SET_PTRACER: > > @@ -232,11 +232,7 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, > > * leader checking is handled later when walking the ancestry > > * at the time of PTRACE_ATTACH check. > > */ > > - rcu_read_lock(); > > - if (!thread_group_leader(myself)) > > - myself = rcu_dereference(myself->group_leader); > > - get_task_struct(myself); > > - rcu_read_unlock(); > > + myself = current->group_leader; > > > > if (arg2 == 0) { > > yama_ptracer_del(NULL, myself); > > @@ -255,7 +251,6 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, > > } > > } > > > > - put_task_struct(myself); > > break; > > } > > > > -- > > 2.25.1.362.g51ebf55 > > > > > > -- > Kees Cook >
Forgot to say... with or without this patch the usage of ptrace_relation->tracer/tracee doesn't look right (safe) to me... but probably I missed something and this is another story. Currently I am trying to audit the usage of task->group_leader. On 02/19, Oleg Nesterov wrote: > > On 02/19, Kees Cook wrote: > > > > On Wed, Feb 19, 2025 at 05:14:17PM +0100, Oleg Nesterov wrote: > > > current->group_leader is stable, no need to take rcu_read_lock() and do > > > get/put_task_struct(). > > > > Can you explain why this is true? In trying to figure this out again, > > it seems that the only way current->group_leader can vanish is if > > the entire process vanishes (fork or thread exec), in which case the > > "current" in this prctl can't be happening; this appears to be locked > > behind tsk->sighand->siglock ? > > Well, almost, but this has nothing to do with tsk->sighand->siglock... > > task->group_leader can only be changed by thread exec, when a non leader > thread does exec, see de_thread(). But de_thread() can't succeed and change > ->group_leader until all other threads exit, see the "Kill all other threads > in the thread group" code in de_thread(). The "current" task can't exit, so > current->group_leader is stable. > > Note also that we already have a lot of current->group_leader users which > don't use rcu/get_task_struct. > > That said, we have a lot of buggy users of tsk->group_leader when > same_thread_group(tsk, current) != true ;) For example, sys_prlimit64(). > And note that rcu_read_lock/get_task_struct can't help in this case. > I am going to send some fixes. > > Oleg. > > > > > -Kees > > > > > > > > Signed-off-by: Oleg Nesterov <oleg@redhat.com> > > > --- > > > security/yama/yama_lsm.c | 9 ++------- > > > 1 file changed, 2 insertions(+), 7 deletions(-) > > > > > > diff --git a/security/yama/yama_lsm.c b/security/yama/yama_lsm.c > > > index 1971710620c1..3d064dd4e03f 100644 > > > --- a/security/yama/yama_lsm.c > > > +++ b/security/yama/yama_lsm.c > > > @@ -222,7 +222,7 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, > > > unsigned long arg4, unsigned long arg5) > > > { > > > int rc = -ENOSYS; > > > - struct task_struct *myself = current; > > > + struct task_struct *myself; > > > > > > switch (option) { > > > case PR_SET_PTRACER: > > > @@ -232,11 +232,7 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, > > > * leader checking is handled later when walking the ancestry > > > * at the time of PTRACE_ATTACH check. > > > */ > > > - rcu_read_lock(); > > > - if (!thread_group_leader(myself)) > > > - myself = rcu_dereference(myself->group_leader); > > > - get_task_struct(myself); > > > - rcu_read_unlock(); > > > + myself = current->group_leader; > > > > > > if (arg2 == 0) { > > > yama_ptracer_del(NULL, myself); > > > @@ -255,7 +251,6 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, > > > } > > > } > > > > > > - put_task_struct(myself); > > > break; > > > } > > > > > > -- > > > 2.25.1.362.g51ebf55 > > > > > > > > > > -- > > Kees Cook > >
Damn, sorry for the spam ;) On 02/19, Oleg Nesterov wrote: > > Forgot to say... > > with or without this patch the usage of ptrace_relation->tracer/tracee > doesn't look right (safe) to me... but probably I missed something ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Yes I did. I didn't realize that put_task_struct(tracer/tracee) calls security_task_free() -> yama_task_free(). Sorry fo the noise. Oleg.
diff --git a/security/yama/yama_lsm.c b/security/yama/yama_lsm.c index 1971710620c1..3d064dd4e03f 100644 --- a/security/yama/yama_lsm.c +++ b/security/yama/yama_lsm.c @@ -222,7 +222,7 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, unsigned long arg4, unsigned long arg5) { int rc = -ENOSYS; - struct task_struct *myself = current; + struct task_struct *myself; switch (option) { case PR_SET_PTRACER: @@ -232,11 +232,7 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, * leader checking is handled later when walking the ancestry * at the time of PTRACE_ATTACH check. */ - rcu_read_lock(); - if (!thread_group_leader(myself)) - myself = rcu_dereference(myself->group_leader); - get_task_struct(myself); - rcu_read_unlock(); + myself = current->group_leader; if (arg2 == 0) { yama_ptracer_del(NULL, myself); @@ -255,7 +251,6 @@ static int yama_task_prctl(int option, unsigned long arg2, unsigned long arg3, } } - put_task_struct(myself); break; }
current->group_leader is stable, no need to take rcu_read_lock() and do get/put_task_struct(). Signed-off-by: Oleg Nesterov <oleg@redhat.com> --- security/yama/yama_lsm.c | 9 ++------- 1 file changed, 2 insertions(+), 7 deletions(-)