lkml.org 
[lkml]   [2020]   [Mar]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v2 5/5] exec: Add a exec_update_mutex to replace cred_guard_mutex
Kirill Tkhai <ktkhai@virtuozzo.com> writes:

> On 12.03.2020 15:24, Eric W. Biederman wrote:
>> Kirill Tkhai <ktkhai@virtuozzo.com> writes:
>>
>>> On 09.03.2020 00:38, Eric W. Biederman wrote:
>>>>
>>>> The cred_guard_mutex is problematic. The cred_guard_mutex is held
>>>> over the userspace accesses as the arguments from userspace are read.
>>>> The cred_guard_mutex is held of PTRACE_EVENT_EXIT as the the other
>>>> threads are killed. The cred_guard_mutex is held over
>>>> "put_user(0, tsk->clear_child_tid)" in exit_mm().
>>>>
>>>> Any of those can result in deadlock, as the cred_guard_mutex is held
>>>> over a possible indefinite userspace waits for userspace.
>>>>
>>>> Add exec_update_mutex that is only held over exec updating process
>>>> with the new contents of exec, so that code that needs not to be
>>>> confused by exec changing the mm and the cred in ways that can not
>>>> happen during ordinary execution of a process.
>>>>
>>>> The plan is to switch the users of cred_guard_mutex to
>>>> exec_udpate_mutex one by one. This lets us move forward while still
>>>> being careful and not introducing any regressions.
>>>>
>>>> Link: https://lore.kernel.org/lkml/20160921152946.GA24210@dhcp22.suse.cz/
>>>> Link: https://lore.kernel.org/lkml/AM6PR03MB5170B06F3A2B75EFB98D071AE4E60@AM6PR03MB5170.eurprd03.prod.outlook.com/
>>>> Link: https://lore.kernel.org/linux-fsdevel/20161102181806.GB1112@redhat.com/
>>>> Link: https://lore.kernel.org/lkml/20160923095031.GA14923@redhat.com/
>>>> Link: https://lore.kernel.org/lkml/20170213141452.GA30203@redhat.com/
>>>> Ref: 45c1a159b85b ("Add PTRACE_O_TRACEVFORKDONE and PTRACE_O_TRACEEXIT facilities.")
>>>> Ref: 456f17cd1a28 ("[PATCH] user-vm-unlock-2.5.31-A2")
>>>> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
>>>> ---
>>>> fs/exec.c | 9 +++++++++
>>>> include/linux/sched/signal.h | 9 ++++++++-
>>>> init/init_task.c | 1 +
>>>> kernel/fork.c | 1 +
>>>> 4 files changed, 19 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/fs/exec.c b/fs/exec.c
>>>> index d820a7272a76..ffeebb1f167b 100644
>>>> --- a/fs/exec.c
>>>> +++ b/fs/exec.c
>>>> @@ -1014,6 +1014,7 @@ static int exec_mmap(struct mm_struct *mm)
>>>> {
>>>> struct task_struct *tsk;
>>>> struct mm_struct *old_mm, *active_mm;
>>>> + int ret;
>>>>
>>>> /* Notify parent that we're no longer interested in the old VM */
>>>> tsk = current;
>>>> @@ -1034,6 +1035,11 @@ static int exec_mmap(struct mm_struct *mm)
>>>> return -EINTR;
>>>> }
>>>> }
>>>> +
>>>> + ret = mutex_lock_killable(&tsk->signal->exec_update_mutex);
>>>> + if (ret)
>>>> + return ret;
>>>
>>> You missed old_mm->mmap_sem unlock. See here:
>>
>> Duh. Thank you.
>>
>> I actually need to switch the lock ordering here, and I haven't yet
>> because my son was sick yesterday.
>
> There is some fundamental problem with your patch, since the below fires in 100% cases
> on current linux-next:

Thank you.

I have just backed this out of linux-next for now because it is clearly
flawed.

You make some good points about the recursion. I will go back to the
drawing board and see what I can work out.


> [ 22.838717] kernel BUG at fs/exec.c:1474!
>
> diff --git a/fs/exec.c b/fs/exec.c
> index 47582cd97f86..0f77f8c94905 100644
> --- a/fs/exec.c
> +++ b/fs/exec.c
> @@ -1470,8 +1470,10 @@ static void free_bprm(struct linux_binprm *bprm)
> {
> free_arg_pages(bprm);
> if (bprm->cred) {
> - if (!bprm->mm)
> + if (!bprm->mm) {
> + BUG_ON(!mutex_is_locked(&current->signal->exec_update_mutex));
> mutex_unlock(&current->signal->exec_update_mutex);
> + }
> mutex_unlock(&current->signal->cred_guard_mutex);
> abort_creds(bprm->cred);
> }
> @@ -1521,6 +1523,7 @@ void install_exec_creds(struct linux_binprm *bprm)
> * credentials; any time after this it may be unlocked.
> */
> security_bprm_committed_creds(bprm);
> + BUG_ON(!mutex_is_locked(&current->signal->exec_update_mutex));
> mutex_unlock(&current->signal->exec_update_mutex);
> mutex_unlock(&current->signal->cred_guard_mutex);
> }
>
> ---------------------------------------------------------------------------------------------
>
> First time the mutex is unlocked in:
>
> exec_binprm()->search_binary_handler()->.load_binary->install_exec_creds()
>
> Then exec_binprm()->search_binary_handler()->.load_binary->flush_old_exec() clears mm:
>
> bprm->mm = NULL;
>
> Second time the mutex is unlocked in free_bprm():
>
> if (bprm->cred) {
> if (!bprm->mm)
> mutex_unlock(&current->signal->exec_update_mutex);
>
> My opinion is we should not relay on side indicators like bprm->mm. Better you may
> introduce struct linux_binprm::exec_update_mutex_is_locked. So the next person dealing
> with this after you won't waste much time on diving into this. Also, if someone decides
> to change the place, where bprm->mm is set into NULL, this person will bump into hell
> of dependences between unrelated components like your newly introduced mutex.
>
> So, I'm strongly for *struct linux_binprm::exec_update_mutex_is_locked*, since this improves
> modularity.

Am I wrong or is that also a problem with cred_guard_mutex?

Eric

\
 
 \ /
  Last update: 2020-03-12 15:41    [W:0.209 / U:5.184 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site