]> git.ipfire.org Git - thirdparty/kernel/stable.git/commitdiff
posix-timers: Ensure timer ID search-loop limit is valid
authorThomas Gleixner <tglx@linutronix.de>
Thu, 1 Jun 2023 18:58:47 +0000 (20:58 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 1 Mar 2024 12:21:55 +0000 (13:21 +0100)
[ Upstream commit 8ce8849dd1e78dadcee0ec9acbd259d239b7069f ]

posix_timer_add() tries to allocate a posix timer ID by starting from the
cached ID which was stored by the last successful allocation.

This is done in a loop searching the ID space for a free slot one by
one. The loop has to terminate when the search wrapped around to the
starting point.

But that's racy vs. establishing the starting point. That is read out
lockless, which leads to the following problem:

CPU0                    CPU1
posix_timer_add()
  start = sig->posix_timer_id;
  lock(hash_lock);
  ...    posix_timer_add()
  if (++sig->posix_timer_id < 0)
                    start = sig->posix_timer_id;
     sig->posix_timer_id = 0;

So CPU1 can observe a negative start value, i.e. -1, and the loop break
never happens because the condition can never be true:

  if (sig->posix_timer_id == start)
     break;

While this is unlikely to ever turn into an endless loop as the ID space is
huge (INT_MAX), the racy read of the start value caught the attention of
KCSAN and Dmitry unearthed that incorrectness.

Rewrite it so that all id operations are under the hash lock.

Reported-by: syzbot+5c54bd3eb218bb595aa9@syzkaller.appspotmail.com
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Link: https://lore.kernel.org/r/87bkhzdn6g.ffs@tglx
Signed-off-by: Sasha Levin <sashal@kernel.org>
include/linux/sched/signal.h
kernel/time/posix-timers.c

index 5f0e8403e8cebedaf799c98a8c8285277c7452f1..9743f7d173a0bc2ede1c15801c249f93d9ac27db 100644 (file)
@@ -125,7 +125,7 @@ struct signal_struct {
 #ifdef CONFIG_POSIX_TIMERS
 
        /* POSIX.1b Interval Timers */
-       int                     posix_timer_id;
+       unsigned int            next_posix_timer_id;
        struct list_head        posix_timers;
 
        /* ITIMER_REAL timer for the process */
index ed3c4a9543982402f7564fbbea699a567da1c89c..2d6cf93ca370af86bc7a91814b4a9689cfb7b717 100644 (file)
@@ -140,25 +140,30 @@ static struct k_itimer *posix_timer_by_id(timer_t id)
 static int posix_timer_add(struct k_itimer *timer)
 {
        struct signal_struct *sig = current->signal;
-       int first_free_id = sig->posix_timer_id;
        struct hlist_head *head;
-       int ret = -ENOENT;
+       unsigned int cnt, id;
 
-       do {
+       /*
+        * FIXME: Replace this by a per signal struct xarray once there is
+        * a plan to handle the resulting CRIU regression gracefully.
+        */
+       for (cnt = 0; cnt <= INT_MAX; cnt++) {
                spin_lock(&hash_lock);
-               head = &posix_timers_hashtable[hash(sig, sig->posix_timer_id)];
-               if (!__posix_timers_find(head, sig, sig->posix_timer_id)) {
+               id = sig->next_posix_timer_id;
+
+               /* Write the next ID back. Clamp it to the positive space */
+               sig->next_posix_timer_id = (id + 1) & INT_MAX;
+
+               head = &posix_timers_hashtable[hash(sig, id)];
+               if (!__posix_timers_find(head, sig, id)) {
                        hlist_add_head_rcu(&timer->t_hash, head);
-                       ret = sig->posix_timer_id;
+                       spin_unlock(&hash_lock);
+                       return id;
                }
-               if (++sig->posix_timer_id < 0)
-                       sig->posix_timer_id = 0;
-               if ((sig->posix_timer_id == first_free_id) && (ret == -ENOENT))
-                       /* Loop over all possible ids completed */
-                       ret = -EAGAIN;
                spin_unlock(&hash_lock);
-       } while (ret == -ENOENT);
-       return ret;
+       }
+       /* POSIX return code when no timer ID could be allocated */
+       return -EAGAIN;
 }
 
 static inline void unlock_timer(struct k_itimer *timr, unsigned long flags)