On 13.03.2020 04:05, Bernd Edlinger wrote: > On 3/12/20 3:38 PM, Eric W. Biederman wrote: >> Kirill Tkhai <ktkhai@xxxxxxxxxxxxx> writes: >> >>> On 12.03.2020 15:24, Eric W. Biederman wrote: >>>> >>>> I actually need to switch the lock ordering here, and I haven't yet >>>> because my son was sick yesterday. > > All the best wishes to you and your son. I hope he will get well soon. > > And sorry for not missing the issue in the review. The reason turns > out that bprm_mm_init is called after prepare_bprm_creds, but there > are error pathes between those where free_bprm is called up with > cred != NULL and mm == NULL, but the mutex not locked. > > I figured out a possible fix for the problem that was pointed out: > > > From ceb6f65b52b3a7f0280f4f20509a1564a439edf6 Mon Sep 17 00:00:00 2001 > From: Bernd Edlinger <bernd.edlinger@xxxxxxxxxx> > Date: Wed, 11 Mar 2020 15:31:07 +0100 > Subject: [PATCH] Fix issues with exec_update_mutex > > Signed-off-by: Bernd Edlinger <bernd.edlinger@xxxxxxxxxx> > --- > fs/exec.c | 17 ++++++++++------- > 1 file changed, 10 insertions(+), 7 deletions(-) > > diff --git a/fs/exec.c b/fs/exec.c > index ffeebb1..cde4937 100644 > --- a/fs/exec.c > +++ b/fs/exec.c > @@ -1021,8 +1021,14 @@ static int exec_mmap(struct mm_struct *mm) > old_mm = current->mm; > exec_mm_release(tsk, old_mm); > > - if (old_mm) { > + if (old_mm) > sync_mm_rss(old_mm); > + > + ret = mutex_lock_killable(&tsk->signal->exec_update_mutex); > + if (ret) > + return ret; > + > + if (old_mm) { > /* > * Make sure that if there is a core dump in progress > * for the old mm, we get out and die instead of going > @@ -1032,14 +1038,11 @@ static int exec_mmap(struct mm_struct *mm) > down_read(&old_mm->mmap_sem); > if (unlikely(old_mm->core_state)) { > up_read(&old_mm->mmap_sem); > + mutex_unlock(&tsk->signal->exec_update_mutex); > return -EINTR; > } > } > > - ret = mutex_lock_killable(&tsk->signal->exec_update_mutex); > - if (ret) > - return ret; > - > task_lock(tsk); > active_mm = tsk->active_mm; > membarrier_exec_mmap(mm); > @@ -1444,8 +1447,6 @@ static void free_bprm(struct linux_binprm *bprm) > { > free_arg_pages(bprm); > if (bprm->cred) { > - if (!bprm->mm) > - mutex_unlock(¤t->signal->exec_update_mutex); > mutex_unlock(¤t->signal->cred_guard_mutex); > abort_creds(bprm->cred); > } > @@ -1846,6 +1847,8 @@ static int __do_execve_file(int fd, struct filename *filename, > would_dump(bprm, bprm->file); > > retval = exec_binprm(bprm); > + if (bprm->cred && !bprm->mm) > + mutex_unlock(¤t->signal->exec_update_mutex); Despite this should fix the problem, this looks like a broken puzzle. We can't use bprm->cred as an identifier whether the mutex was locked or not. We can check for bprm->cred in regard to cred_guard_mutex, because of there is strong rule: "cred_guard_mutex is becomes locked together with bprm->cred assignment (see prepare_bprm_creds()), and it becomes unlocked together with bprm->cred zeroing". Take attention on modularity of all this: there is no dependencies between anything else. In regard to newly introduced exec_update_mutex, your fix and source patch way look like an obfuscation. The mutex becomes deadly glued to unrelated bprm->cred and bprm->mm, and this introduces the problems in the future modifications and support of all involved entities. If someone wants to move some functions in relation to each other, there will be a pain, and this person will have to go again the same dependencies and bug way, Eric stepped on in the original patch.