On 2/3/20 3:50 AM, Ondrej Mosnacek wrote:
Avoiding taking a lock in an IRQ context is not enough to prevent
deadlocks, as discovered by syzbot:
===
WARNING: SOFTIRQ-safe -> SOFTIRQ-unsafe lock order detected
5.5.0-syzkaller #0 Not tainted
-----------------------------------------------------
syz-executor.0/8927 [HC0[0]:SC0[2]:HE1:SE0] is trying to acquire:
ffff888027c94098 (&(&s->cache_lock)->rlock){+.+.}, at: spin_lock include/linux/spinlock.h:338 [inline]
ffff888027c94098 (&(&s->cache_lock)->rlock){+.+.}, at: sidtab_sid2str_put.part.0+0x36/0x880 security/selinux/ss/sidtab.c:533
and this task is already holding:
ffffffff898639b0 (&(&nf_conntrack_locks[i])->rlock){+.-.}, at: spin_lock include/linux/spinlock.h:338 [inline]
ffffffff898639b0 (&(&nf_conntrack_locks[i])->rlock){+.-.}, at: nf_conntrack_lock+0x17/0x70 net/netfilter/nf_conntrack_core.c:91
which would create a new lock dependency:
(&(&nf_conntrack_locks[i])->rlock){+.-.} -> (&(&s->cache_lock)->rlock){+.+.}
but this new dependency connects a SOFTIRQ-irq-safe lock:
(&(&nf_conntrack_locks[i])->rlock){+.-.}
[...]
other info that might help us debug this:
Possible interrupt unsafe locking scenario:
CPU0 CPU1
---- ----
lock(&(&s->cache_lock)->rlock);
local_irq_disable();
lock(&(&nf_conntrack_locks[i])->rlock);
lock(&(&s->cache_lock)->rlock);
<Interrupt>
lock(&(&nf_conntrack_locks[i])->rlock);
*** DEADLOCK ***
[...]
===
Fix this by simply locking with irqsave/irqrestore and stop giving up on
!in_task(). It makes the locking a bit slower, but it shouldn't make a
big difference in real workloads. Under the scenario from [1] (only
cache hits) it only increased the runtime overhead from the
security_secid_to_secctx() function from ~2% to ~3% (it was ~5-65%
before introducing the cache).
[1] https://bugzilla.redhat.com/show_bug.cgi?id=1733259
Fixes: d97bd23c2d7d ("selinux: cache the SID -> context string translation")
Reported-by: syzbot+61cba5033e2072d61806@xxxxxxxxxxxxxxxxxxxxxxxxx
Signed-off-by: Ondrej Mosnacek <omosnace@xxxxxxxxxx>
Acked-by: Stephen Smalley <sds@xxxxxxxxxxxxx>
[...]
