bpf: Fix a potential deadlock with bpf_map_do_batch

Commit 05799638 ("bpf: Add batch ops to all htab bpf map")
added lookup_and_delete batch operation for hash table.
The current implementation has bpf_lru_push_free() inside
the bucket lock, which may cause a deadlock.

syzbot reports:
   -> #2 (&htab->buckets[i].lock#2){....}:
       __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline]
       _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159
       htab_lru_map_delete_node+0xce/0x2f0 kernel/bpf/hashtab.c:593
       __bpf_lru_list_shrink_inactive kernel/bpf/bpf_lru_list.c:220 [inline]
       __bpf_lru_list_shrink+0xf9/0x470 kernel/bpf/bpf_lru_list.c:266
       bpf_lru_list_pop_free_to_local kernel/bpf/bpf_lru_list.c:340 [inline]
       bpf_common_lru_pop_free kernel/bpf/bpf_lru_list.c:447 [inline]
       bpf_lru_pop_free+0x87c/0x1670 kernel/bpf/bpf_lru_list.c:499
       prealloc_lru_pop+0x2c/0xa0 kernel/bpf/hashtab.c:132
       __htab_lru_percpu_map_update_elem+0x67e/0xa90 kernel/bpf/hashtab.c:1069
       bpf_percpu_hash_update+0x16e/0x210 kernel/bpf/hashtab.c:1585
       bpf_map_update_value.isra.0+0x2d7/0x8e0 kernel/bpf/syscall.c:181
       generic_map_update_batch+0x41f/0x610 kernel/bpf/syscall.c:1319
       bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348
       __do_sys_bpf+0x9b7/0x41e0 kernel/bpf/syscall.c:3460
       __se_sys_bpf kernel/bpf/syscall.c:3355 [inline]
       __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355
       do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294
       entry_SYSCALL_64_after_hwframe+0x49/0xbe

   -> #0 (&loc_l->lock){....}:
       check_prev_add kernel/locking/lockdep.c:2475 [inline]
       check_prevs_add kernel/locking/lockdep.c:2580 [inline]
       validate_chain kernel/locking/lockdep.c:2970 [inline]
       __lock_acquire+0x2596/0x4a00 kernel/locking/lockdep.c:3954
       lock_acquire+0x190/0x410 kernel/locking/lockdep.c:4484
       __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline]
       _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159
       bpf_common_lru_push_free kernel/bpf/bpf_lru_list.c:516 [inline]
       bpf_lru_push_free+0x250/0x5b0 kernel/bpf/bpf_lru_list.c:555
       __htab_map_lookup_and_delete_batch+0x8d4/0x1540 kernel/bpf/hashtab.c:1374
       htab_lru_map_lookup_and_delete_batch+0x34/0x40 kernel/bpf/hashtab.c:1491
       bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348
       __do_sys_bpf+0x1f7d/0x41e0 kernel/bpf/syscall.c:3456
       __se_sys_bpf kernel/bpf/syscall.c:3355 [inline]
       __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355
       do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294
       entry_SYSCALL_64_after_hwframe+0x49/0xbe

    Possible unsafe locking scenario:

          CPU0                    CPU2
          ----                    ----
     lock(&htab->buckets[i].lock#2);
                                  lock(&l->lock);
                                  lock(&htab->buckets[i].lock#2);
     lock(&loc_l->lock);

    *** DEADLOCK ***

To fix the issue, for htab_lru_map_lookup_and_delete_batch() in CPU0,
let us do bpf_lru_push_free() out of the htab bucket lock. This can
avoid the above deadlock scenario.

Fixes: 05799638 ("bpf: Add batch ops to all htab bpf map")
Reported-by: syzbot+a38ff3d9356388f2fb83@syzkaller.appspotmail.com
Reported-by: syzbot+122b5421d14e68f29cd1@syzkaller.appspotmail.com
Suggested-by: Hillf Danton <hdanton@sina.com>
Suggested-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com>
Acked-by: Brian Vazquez <brianvv@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Link: https://lore.kernel.org/bpf/20200219234757.3544014-1-yhs@fb.com

kernel/bpf/hashtab.c

@@ -56,6 +56,7 @@ struct htab_elem {
 			union {
 				struct bpf_htab *htab;
 				struct pcpu_freelist_node fnode;
+				struct htab_elem *batch_flink;
 			};
 		};
 	};
@@ -126,6 +127,17 @@ static void htab_free_elems(struct bpf_htab *htab)
 	bpf_map_area_free(htab->elems);
 }
 
+/* The LRU list has a lock (lru_lock). Each htab bucket has a lock
+ * (bucket_lock). If both locks need to be acquired together, the lock
+ * order is always lru_lock -> bucket_lock and this only happens in
+ * bpf_lru_list.c logic. For example, certain code path of
+ * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
+ * will acquire lru_lock first followed by acquiring bucket_lock.
+ *
+ * In hashtab.c, to avoid deadlock, lock acquisition of
+ * bucket_lock followed by lru_lock is not allowed. In such cases,
+ * bucket_lock needs to be released first before acquiring lru_lock.
+ */
 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 					  u32 hash)
 {
@@ -1256,6 +1268,7 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
 	void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
 	void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
 	u32 batch, max_count, size, bucket_size;
+	struct htab_elem *node_to_free = NULL;
 	u64 elem_map_flags, map_flags;
 	struct hlist_nulls_head *head;
 	struct hlist_nulls_node *n;
@@ -1388,10 +1401,18 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
 		}
 		if (do_delete) {
 			hlist_nulls_del_rcu(&l->hash_node);
-			if (is_lru_map)
-				bpf_lru_push_free(&htab->lru, &l->lru_node);
-			else
+
+			/* bpf_lru_push_free() will acquire lru_lock, which
+			 * may cause deadlock. See comments in function
+			 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
+			 * after releasing the bucket lock.
+			 */
+			if (is_lru_map) {
+				l->batch_flink = node_to_free;
+				node_to_free = l;
+			} else {
 				free_htab_elem(htab, l);
+			}
 		}
 		dst_key += key_size;
 		dst_val += value_size;
@@ -1399,6 +1420,13 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
 
 	raw_spin_unlock_irqrestore(&b->lock, flags);
 	locked = false;
+
+	while (node_to_free) {
+		l = node_to_free;
+		node_to_free = node_to_free->batch_flink;
+		bpf_lru_push_free(&htab->lru, &l->lru_node);
+	}
+
 next_batch:
 	/* If we are not copying data, we can go to next bucket and avoid
 	 * unlocking the rcu.