Revert "ipc,sem: remove uneeded sem_undo_list lock usage in exit_sem()"

This reverts commit a979558.

Commit a979558 ("ipc,sem: remove uneeded sem_undo_list lock usage
in exit_sem()") removes a lock that is needed.  This leads to a process
looping infinitely in exit_sem() and can also lead to a crash.  There is
a reproducer available in [1] and with the commit reverted the issue
does not reproduce anymore.

Using the reproducer found in [1] is fairly easy to reach a point where
one of the child processes is looping infinitely in exit_sem between
for(;;) and if (semid == -1) block, while it's trying to free its last
sem_undo structure which has already been freed by freeary().

Each sem_undo struct is on two lists: one per semaphore set (list_id)
and one per process (list_proc).  The list_id list tracks undos by
semaphore set, and the list_proc by process.

Undo structures are removed either by freeary() or by exit_sem().  The
freeary function is invoked when the user invokes a syscall to remove a
semaphore set.  During this operation freeary() traverses the list_id
associated with the semaphore set and removes the undo structures from
both the list_id and list_proc lists.

For this case, exit_sem() is called at process exit.  Each process
contains a struct sem_undo_list (referred to as "ulp") which contains
the head for the list_proc list.  When the process exits, exit_sem()
traverses this list to remove each sem_undo struct.  As in freeary(),
whenever a sem_undo struct is removed from list_proc, it is also removed
from the list_id list.

Removing elements from list_id is safe for both exit_sem() and freeary()
due to sem_lock().  Removing elements from list_proc is not safe;
freeary() locks &un->ulp->lock when it performs
list_del_rcu(&un->list_proc) but exit_sem() does not (locking was
removed by commit a979558 ("ipc,sem: remove uneeded sem_undo_list
lock usage in exit_sem()").

This can result in the following situation while executing the
reproducer [1] : Consider a child process in exit_sem() and the parent
in freeary() (because of semctl(sid[i], NSEM, IPC_RMID)).

 - The list_proc for the child contains the last two undo structs A and
   B (the rest have been removed either by exit_sem() or freeary()).

 - The semid for A is 1 and semid for B is 2.

 - exit_sem() removes A and at the same time freeary() removes B.

 - Since A and B have different semid sem_lock() will acquire different
   locks for each process and both can proceed.

The bug is that they remove A and B from the same list_proc at the same
time because only freeary() acquires the ulp lock. When exit_sem()
removes A it makes ulp->list_proc.next to point at B and at the same
time freeary() removes B setting B->semid=-1.

At the next iteration of for(;;) loop exit_sem() will try to remove B.

The only way to break from for(;;) is for (&un->list_proc ==
&ulp->list_proc) to be true which is not. Then exit_sem() will check if
B->semid=-1 which is and will continue looping in for(;;) until the
memory for B is reallocated and the value at B->semid is changed.

At that point, exit_sem() will crash attempting to unlink B from the
lists (this can be easily triggered by running the reproducer [1] a
second time).

To prove this scenario instrumentation was added to keep information
about each sem_undo (un) struct that is removed per process and per
semaphore set (sma).

          CPU0                                CPU1
  [caller holds sem_lock(sma for A)]      ...
  freeary()                               exit_sem()
  ...                                     ...
  ...                                     sem_lock(sma for B)
  spin_lock(A->ulp->lock)                 ...
  list_del_rcu(un_A->list_proc)           list_del_rcu(un_B->list_proc)

Undo structures A and B have different semid and sem_lock() operations
proceed.  However they belong to the same list_proc list and they are
removed at the same time.  This results into ulp->list_proc.next
pointing to the address of B which is already removed.

After reverting commit a979558 ("ipc,sem: remove uneeded
sem_undo_list lock usage in exit_sem()") the issue was no longer
reproducible.

[1] https://bugzilla.redhat.com/show_bug.cgi?id=1694779

Link: http://lkml.kernel.org/r/[email protected]
Fixes: a979558 ("ipc,sem: remove uneeded sem_undo_list lock usage in exit_sem()")
Signed-off-by: Ioanna Alifieraki <[email protected]>
Acked-by: Manfred Spraul <[email protected]>
Acked-by: Herton R. Krzesinski <[email protected]>
Cc: Arnd Bergmann <[email protected]>
Cc: Catalin Marinas <[email protected]>
Cc: <[email protected]>
Cc: Joel Fernandes (Google) <[email protected]>
Cc: Davidlohr Bueso <[email protected]>
Cc: Jay Vosburgh <[email protected]>
Cc: <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

ipc/sem.c

@@ -2384,11 +2384,9 @@ void exit_sem(struct task_struct *tsk)
 		ipc_assert_locked_object(&sma->sem_perm);
 		list_del(&un->list_id);
 
-		/* we are the last process using this ulp, acquiring ulp->lock
-		 * isn't required. Besides that, we are also protected against
-		 * IPC_RMID as we hold sma->sem_perm lock now
-		 */
+		spin_lock(&ulp->lock);
 		list_del_rcu(&un->list_proc);
+		spin_unlock(&ulp->lock);
 
 		/* perform adjustments registered in un */
 		for (i = 0; i < sma->sem_nsems; i++) {