| Message ID | 1403911380-27787-10-git-send-email-keescook@chromium.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 06/27, Kees Cook wrote: > > static u32 seccomp_run_filters(int syscall) > { > - struct seccomp_filter *f; > + struct seccomp_filter *f = ACCESS_ONCE(current->seccomp.filter); I am not sure... This is fine if this ->filter is the 1st (and only) one, in this case we can rely on rmb() in the caller. But the new filter can be installed at any moment. Say, right after that rmb() although this doesn't matter. Either we need smp_read_barrier_depends() after that, or smp_load_acquire() like the previous version did? Oleg.
On 07/09, Oleg Nesterov wrote: > > On 06/27, Kees Cook wrote: > > > > static u32 seccomp_run_filters(int syscall) > > { > > - struct seccomp_filter *f; > > + struct seccomp_filter *f = ACCESS_ONCE(current->seccomp.filter); > > I am not sure... > > This is fine if this ->filter is the 1st (and only) one, in this case > we can rely on rmb() in the caller. > > But the new filter can be installed at any moment. Say, right after that > rmb() although this doesn't matter. Either we need smp_read_barrier_depends() > after that, or smp_load_acquire() like the previous version did? Wait... and it seems that seccomp_sync_threads() needs smp_store_release() when it sets thread->filter = current->filter by the same reason? OTOH. smp_store_release() in seccomp_attach_filter() can die, "current" doesn't need a barrier to serialize with itself. Oleg.
On 06/27, Kees Cook wrote: > > void put_seccomp_filter(struct task_struct *tsk) > { > - struct seccomp_filter *orig = tsk->seccomp.filter; > + struct seccomp_filter *orig = smp_load_acquire(&tsk->seccomp.filter); > /* Clean up single-reference branches iteratively. */ > while (orig && atomic_dec_and_test(&orig->usage)) { And this smp_load_acquire() looks unnecessary. atomic_dec_and_test() is a barrier. Oleg.
On Wed, Jul 9, 2014 at 11:55 AM, Oleg Nesterov <oleg@redhat.com> wrote: > On 07/09, Oleg Nesterov wrote: >> >> On 06/27, Kees Cook wrote: >> > >> > static u32 seccomp_run_filters(int syscall) >> > { >> > - struct seccomp_filter *f; >> > + struct seccomp_filter *f = ACCESS_ONCE(current->seccomp.filter); >> >> I am not sure... >> >> This is fine if this ->filter is the 1st (and only) one, in this case >> we can rely on rmb() in the caller. >> >> But the new filter can be installed at any moment. Say, right after that >> rmb() although this doesn't matter. Either we need smp_read_barrier_depends() >> after that, or smp_load_acquire() like the previous version did? > > Wait... and it seems that seccomp_sync_threads() needs smp_store_release() > when it sets thread->filter = current->filter by the same reason? > > OTOH. smp_store_release() in seccomp_attach_filter() can die, "current" > doesn't need a barrier to serialize with itself. I have lost track of what you're suggesting to change. :) Since rmb() happens before run_filters, isn't the ACCESS_ONCE sufficient? We only care that TIF_SECCOMP, mode, and some filter is valid. In a tsync thread race, it's okay to use not use the deepest filter node in the list, it just needs A filter. -Kees
On 07/10, Kees Cook wrote: > > On Wed, Jul 9, 2014 at 11:55 AM, Oleg Nesterov <oleg@redhat.com> wrote: > > On 07/09, Oleg Nesterov wrote: > >> > >> On 06/27, Kees Cook wrote: > >> > > >> > static u32 seccomp_run_filters(int syscall) > >> > { > >> > - struct seccomp_filter *f; > >> > + struct seccomp_filter *f = ACCESS_ONCE(current->seccomp.filter); > >> > >> I am not sure... > >> > >> This is fine if this ->filter is the 1st (and only) one, in this case > >> we can rely on rmb() in the caller. > >> > >> But the new filter can be installed at any moment. Say, right after that > >> rmb() although this doesn't matter. Either we need smp_read_barrier_depends() > >> after that, or smp_load_acquire() like the previous version did? > > > > Wait... and it seems that seccomp_sync_threads() needs smp_store_release() > > when it sets thread->filter = current->filter by the same reason? > > > > OTOH. smp_store_release() in seccomp_attach_filter() can die, "current" > > doesn't need a barrier to serialize with itself. > > I have lost track of what you're suggesting to change. :) Perhaps I am just trying to confuse you and myself ;) But, > Since rmb() happens before run_filters, isn't the ACCESS_ONCE > sufficient? Yes. But see above. ACCESS_ONCE is sufficient if we read the first filter installed by another thread, in this case rmb() pairs with mb_before_atomic() before set_bit(TIF_SECCOMP). IOW, if this threads sees TIF_SECCOMP, it should also see all modifications which were done before set_bit, including the data in ->filter points to. > We only care that TIF_SECCOMP, mode, and some filter is > valid. In a tsync thread race, it's okay to use not use the deepest > filter node in the list, Yes, it is fine if we miss yet another filter which was just installed by another thread. But, unless I missed something, the problem is that we can get this new filter. Just to simplify. Suppose TIF_SECCOMP was set a long ago. This thread has a single filter F1 and it enters seccomp_run_filters(). Right before it does ACCESS_ONCE() to read the pointer, another thread does seccomp_sync_threads() and sets .filter = F2. If ACCESS_ONCE() returns F1 - everything is fine. But it can see the new pointer F2, and in this case we need a barrier to ensure that, say, LOAD(F2->prog) will see all the preceding changes in this memory. Oleg.
On Thu, Jul 10, 2014 at 8:24 AM, Oleg Nesterov <oleg@redhat.com> wrote: > On 07/10, Kees Cook wrote: >> >> On Wed, Jul 9, 2014 at 11:55 AM, Oleg Nesterov <oleg@redhat.com> wrote: >> > On 07/09, Oleg Nesterov wrote: >> >> >> >> On 06/27, Kees Cook wrote: >> >> > >> >> > static u32 seccomp_run_filters(int syscall) >> >> > { >> >> > - struct seccomp_filter *f; >> >> > + struct seccomp_filter *f = ACCESS_ONCE(current->seccomp.filter); >> >> >> >> I am not sure... >> >> >> >> This is fine if this ->filter is the 1st (and only) one, in this case >> >> we can rely on rmb() in the caller. >> >> >> >> But the new filter can be installed at any moment. Say, right after that >> >> rmb() although this doesn't matter. Either we need smp_read_barrier_depends() >> >> after that, or smp_load_acquire() like the previous version did? >> > >> > Wait... and it seems that seccomp_sync_threads() needs smp_store_release() >> > when it sets thread->filter = current->filter by the same reason? >> > >> > OTOH. smp_store_release() in seccomp_attach_filter() can die, "current" >> > doesn't need a barrier to serialize with itself. >> >> I have lost track of what you're suggesting to change. :) > > Perhaps I am just trying to confuse you and myself ;) > > But, > >> Since rmb() happens before run_filters, isn't the ACCESS_ONCE >> sufficient? > > Yes. But see above. ACCESS_ONCE is sufficient if we read the first filter > installed by another thread, in this case rmb() pairs with mb_before_atomic() > before set_bit(TIF_SECCOMP). > > IOW, if this threads sees TIF_SECCOMP, it should also see all modifications > which were done before set_bit, including the data in ->filter points to. > >> We only care that TIF_SECCOMP, mode, and some filter is >> valid. In a tsync thread race, it's okay to use not use the deepest >> filter node in the list, > > Yes, it is fine if we miss yet another filter which was just installed by > another thread. > > But, unless I missed something, the problem is that we can get this new > filter. > > Just to simplify. Suppose TIF_SECCOMP was set a long ago. This thread > has a single filter F1 and it enters seccomp_run_filters(). > > Right before it does ACCESS_ONCE() to read the pointer, another thread > does seccomp_sync_threads() and sets .filter = F2. > > If ACCESS_ONCE() returns F1 - everything is fine. But it can see the new > pointer F2, and in this case we need a barrier to ensure that, say, > LOAD(F2->prog) will see all the preceding changes in this memory. And the rmb() isn't sufficient for that? Is another barrier needed before assigning the filter pointer to make sure the contents it points to are flushed? What's the least time-consuming operation I can use in run_filters? -Kees
On 07/10, Kees Cook wrote: > > On Thu, Jul 10, 2014 at 8:24 AM, Oleg Nesterov <oleg@redhat.com> wrote: > > > > Just to simplify. Suppose TIF_SECCOMP was set a long ago. This thread > > has a single filter F1 and it enters seccomp_run_filters(). > > > > Right before it does ACCESS_ONCE() to read the pointer, another thread > > does seccomp_sync_threads() and sets .filter = F2. > > > > If ACCESS_ONCE() returns F1 - everything is fine. But it can see the new > > pointer F2, and in this case we need a barrier to ensure that, say, > > LOAD(F2->prog) will see all the preceding changes in this memory. > > And the rmb() isn't sufficient for that? But it has no effect if the pointer was changed _after_ rmb() was already called. And, you need a barrier _after_ ACCESS_ONCE(). (Unless, again, we know that this is the first filter, but this is only by accident). > Is another barrier needed > before assigning the filter pointer to make sure the contents it > points to are flushed? I think smp_store_release() should be moved from seccomp_attach_filter() to seccomp_sync_threads(). Although probably it _should_ work either way, but at least this looks confusing because a) "current" doesn't need a barrier to serialize wuth itself, and b) it is not clear why it is safe to change the pointer dereferenced by another thread without a barrier. > What's the least time-consuming operation I can use in run_filters? As I said smp_read_barrier_depends() (nop unless alpha) or smp_load_acquire() which you used in the previous version. And to remind, afaics smp_load_acquire() in put_filter() should die ;) Oleg.
diff --git a/include/linux/seccomp.h b/include/linux/seccomp.h index 4054b0994071..9ff98b4bfe2e 100644 --- a/include/linux/seccomp.h +++ b/include/linux/seccomp.h @@ -14,11 +14,11 @@ struct seccomp_filter; * * @mode: indicates one of the valid values above for controlled * system calls available to a process. - * @filter: The metadata and ruleset for determining what system calls - * are allowed for a task. + * @filter: must always point to a valid seccomp-filter or NULL as it is + * accessed without locking during system call entry. * * @filter must only be accessed from the context of current as there - * is no locking. + * is no read locking. */ struct seccomp { int mode; diff --git a/kernel/fork.c b/kernel/fork.c index 6a13c46cd87d..ffc1b43e351f 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -315,6 +315,15 @@ static struct task_struct *dup_task_struct(struct task_struct *orig) goto free_ti; tsk->stack = ti; +#ifdef CONFIG_SECCOMP + /* + * We must handle setting up seccomp filters once we're under + * the sighand lock in case orig has changed between now and + * then. Until then, filter must be NULL to avoid messing up + * the usage counts on the error path calling free_task. + */ + tsk->seccomp.filter = NULL; +#endif setup_thread_stack(tsk, orig); clear_user_return_notifier(tsk); @@ -1081,6 +1090,35 @@ static int copy_signal(unsigned long clone_flags, struct task_struct *tsk) return 0; } +static void copy_seccomp(struct task_struct *p) +{ +#ifdef CONFIG_SECCOMP + /* + * Must be called with sighand->lock held, which is common to + * all threads in the group. Holding cred_guard_mutex is not + * needed because this new task is not yet running and cannot + * be racing exec. + */ + BUG_ON(!spin_is_locked(¤t->sighand->siglock)); + + /* Ref-count the new filter user, and assign it. */ + get_seccomp_filter(current); + p->seccomp = current->seccomp; + + /* + * Explicitly enable no_new_privs here in case it got set + * between the task_struct being duplicated and holding the + * sighand lock. The seccomp state and nnp must be in sync. + */ + if (task_no_new_privs(current)) + task_set_no_new_privs(p); + + /* If we have a seccomp mode, enable the thread flag. */ + if (p->seccomp.mode != SECCOMP_MODE_DISABLED) + set_tsk_thread_flag(p, TIF_SECCOMP); +#endif +} + SYSCALL_DEFINE1(set_tid_address, int __user *, tidptr) { current->clear_child_tid = tidptr; @@ -1196,7 +1234,6 @@ static struct task_struct *copy_process(unsigned long clone_flags, goto fork_out; ftrace_graph_init_task(p); - get_seccomp_filter(p); rt_mutex_init_task(p); @@ -1437,6 +1474,12 @@ static struct task_struct *copy_process(unsigned long clone_flags, spin_lock(¤t->sighand->siglock); /* + * Copy seccomp details explicitly here, in case they were changed + * before holding sighand lock. + */ + copy_seccomp(p); + + /* * Process group and session signals need to be delivered to just the * parent before the fork or both the parent and the child after the * fork. Restart if a signal comes in before we add the new process to diff --git a/kernel/seccomp.c b/kernel/seccomp.c index 502e54d7f86d..e1ff2c193190 100644 --- a/kernel/seccomp.c +++ b/kernel/seccomp.c @@ -173,12 +173,12 @@ static int seccomp_check_filter(struct sock_filter *filter, unsigned int flen) */ static u32 seccomp_run_filters(int syscall) { - struct seccomp_filter *f; + struct seccomp_filter *f = ACCESS_ONCE(current->seccomp.filter); struct seccomp_data sd; u32 ret = SECCOMP_RET_ALLOW; /* Ensure unexpected behavior doesn't result in failing open. */ - if (WARN_ON(current->seccomp.filter == NULL)) + if (unlikely(WARN_ON(f == NULL))) return SECCOMP_RET_KILL; populate_seccomp_data(&sd); @@ -187,7 +187,7 @@ static u32 seccomp_run_filters(int syscall) * All filters in the list are evaluated and the lowest BPF return * value always takes priority (ignoring the DATA). */ - for (f = current->seccomp.filter; f; f = f->prev) { + for (; f; f = f->prev) { u32 cur_ret = SK_RUN_FILTER(f->prog, (void *)&sd); if ((cur_ret & SECCOMP_RET_ACTION) < (ret & SECCOMP_RET_ACTION)) @@ -199,6 +199,8 @@ static u32 seccomp_run_filters(int syscall) static inline bool seccomp_check_mode(unsigned long seccomp_mode) { + BUG_ON(!spin_is_locked(¤t->sighand->siglock)); + if (current->seccomp.mode && current->seccomp.mode != seccomp_mode) return false; @@ -207,6 +209,8 @@ static inline bool seccomp_check_mode(unsigned long seccomp_mode) static inline void seccomp_assign_mode(unsigned long seccomp_mode) { + BUG_ON(!spin_is_locked(¤t->sighand->siglock)); + current->seccomp.mode = seccomp_mode; set_tsk_thread_flag(current, TIF_SECCOMP); } @@ -330,6 +334,8 @@ out: * @flags: flags to change filter behavior * @filter: seccomp filter to add to the current process * + * Caller must be holding current->sighand->siglock lock. + * * Returns 0 on success, -ve on error. */ static long seccomp_attach_filter(unsigned int flags, @@ -338,6 +344,8 @@ static long seccomp_attach_filter(unsigned int flags, unsigned long total_insns; struct seccomp_filter *walker; + BUG_ON(!spin_is_locked(¤t->sighand->siglock)); + /* Validate resulting filter length. */ total_insns = filter->prog->len; for (walker = current->seccomp.filter; walker; walker = walker->prev) @@ -350,7 +358,7 @@ static long seccomp_attach_filter(unsigned int flags, * task reference. */ filter->prev = current->seccomp.filter; - current->seccomp.filter = filter; + smp_store_release(¤t->seccomp.filter, filter); return 0; } @@ -376,7 +384,7 @@ static inline void seccomp_filter_free(struct seccomp_filter *filter) /* put_seccomp_filter - decrements the ref count of tsk->seccomp.filter */ void put_seccomp_filter(struct task_struct *tsk) { - struct seccomp_filter *orig = tsk->seccomp.filter; + struct seccomp_filter *orig = smp_load_acquire(&tsk->seccomp.filter); /* Clean up single-reference branches iteratively. */ while (orig && atomic_dec_and_test(&orig->usage)) { struct seccomp_filter *freeme = orig; @@ -527,6 +535,8 @@ static long seccomp_set_mode_strict(void) const unsigned long seccomp_mode = SECCOMP_MODE_STRICT; long ret = -EINVAL; + spin_lock_irq(¤t->sighand->siglock); + if (!seccomp_check_mode(seccomp_mode)) goto out; @@ -537,6 +547,7 @@ static long seccomp_set_mode_strict(void) ret = 0; out: + spin_unlock_irq(¤t->sighand->siglock); return ret; } @@ -564,13 +575,15 @@ static long seccomp_set_mode_filter(unsigned int flags, /* Validate flags. */ if (flags != 0) - goto out; + return -EINVAL; /* Prepare the new filter before holding any locks. */ prepared = seccomp_prepare_user_filter(filter); if (IS_ERR(prepared)) return PTR_ERR(prepared); + spin_lock_irq(¤t->sighand->siglock); + if (!seccomp_check_mode(seccomp_mode)) goto out; @@ -582,6 +595,7 @@ static long seccomp_set_mode_filter(unsigned int flags, seccomp_assign_mode(seccomp_mode); out: + spin_unlock_irq(¤t->sighand->siglock); seccomp_filter_free(prepared); return ret; }
Normally, task_struct.seccomp.filter is only ever read or modified by the task that owns it (current). This property aids in fast access during system call filtering as read access is lockless. Updating the pointer from another task, however, opens up race conditions. To allow cross-thread filter pointer updates, writes to the seccomp fields are now protected by the sighand spinlock (which is unique to the thread group). Read access remains lockless because pointer updates themselves are atomic. However, writes (or cloning) often entail additional checking (like maximum instruction counts) which require locking to perform safely. In the case of cloning threads, the child is invisible to the system until it enters the task list. To make sure a child can't be cloned from a thread and left in a prior state, seccomp duplication is additionally moved under the sighand lock. Then parent and child are certain have the same seccomp state when they exit the lock. Based on patches by Will Drewry and David Drysdale. Signed-off-by: Kees Cook <keescook@chromium.org> --- include/linux/seccomp.h | 6 +++--- kernel/fork.c | 45 ++++++++++++++++++++++++++++++++++++++++++++- kernel/seccomp.c | 26 ++++++++++++++++++++------ 3 files changed, 67 insertions(+), 10 deletions(-)