[bpf] bpf, sockmap: Prevent lock inversion deadlock in map delete elem

Commit Message

Jakub Sitnicki April 2, 2024, 10:46 a.m. UTC

syzkaller started using corpuses where a BPF tracing program deletes
elements from a sockmap/sockhash map. Because BPF tracing programs can be
invoked from any interrupt context, locks taken during a map_delete_elem
operation must be hardirq-safe. Otherwise a deadlock due to lock inversion
is possible, as reported by lockdep:

       CPU0                    CPU1
       ----                    ----
  lock(&htab->buckets[i].lock);
                               local_irq_disable();
                               lock(&host->lock);
                               lock(&htab->buckets[i].lock);
  <Interrupt>
    lock(&host->lock);

Locks in sockmap are hardirq-unsafe by design. We expects elements to be
deleted from sockmap/sockhash only in task (normal) context with interrupts
enabled, or in softirq context.

Detect when map_delete_elem operation is invoked from a context which is
_not_ hardirq-unsafe, that is interrupts are disabled, and bail out with an
error.

Note that map updates are not affected by this issue. BPF verifier does not
allow updating sockmap/sockhash from a BPF tracing program today.

Reported-by: xingwei lee <[e-mail]>
Reported-by: yue sun <[e-mail]>
Reported-by: [e-mail]
Reported-and-tested-by: [e-mail]
Closes: https://syzkaller.appspot.com/bug?extid=d4066896495db380182e
Closes: https://syzkaller.appspot.com/bug?extid=bc922f476bd65abbd466
Fixes: 604326b41a6f ("bpf, sockmap: convert to generic sk_msg interface")
Signed-off-by: Jakub Sitnicki <[e-mail]>
---
Cc: John Fastabend <[e-mail]>
Cc: Edward Adam Davis <[e-mail]>
Cc: Shung-Hsi Yu <[e-mail]>
---
 net/core/sock_map.c | 6 ++++++
 1 file changed, 6 insertions(+)

Comments

John Fastabend April 2, 2024, 2:09 p.m. UTC | #1

Jakub Sitnicki wrote:
> syzkaller started using corpuses where a BPF tracing program deletes
> elements from a sockmap/sockhash map. Because BPF tracing programs can be
> invoked from any interrupt context, locks taken during a map_delete_elem
> operation must be hardirq-safe. Otherwise a deadlock due to lock inversion
> is possible, as reported by lockdep:
> 
>        CPU0                    CPU1
>        ----                    ----
>   lock(&htab->buckets[i].lock);
>                                local_irq_disable();
>                                lock(&host->lock);
>                                lock(&htab->buckets[i].lock);
>   <Interrupt>
>     lock(&host->lock);
> 
> Locks in sockmap are hardirq-unsafe by design. We expects elements to be
> deleted from sockmap/sockhash only in task (normal) context with interrupts
> enabled, or in softirq context.
> 
> Detect when map_delete_elem operation is invoked from a context which is
> _not_ hardirq-unsafe, that is interrupts are disabled, and bail out with an
> error.
> 
> Note that map updates are not affected by this issue. BPF verifier does not
> allow updating sockmap/sockhash from a BPF tracing program today.
> 
> Reported-by: xingwei lee <[e-mail]>
> Reported-by: yue sun <[e-mail]>
> Reported-by: [e-mail]
> Reported-and-tested-by: [e-mail]
> Closes: https://syzkaller.appspot.com/bug?extid=d4066896495db380182e
> Closes: https://syzkaller.appspot.com/bug?extid=bc922f476bd65abbd466
> Fixes: 604326b41a6f ("bpf, sockmap: convert to generic sk_msg interface")
> Signed-off-by: Jakub Sitnicki <[e-mail]>
> ---
> Cc: John Fastabend <[e-mail]>
> Cc: Edward Adam Davis <[e-mail]>
> Cc: Shung-Hsi Yu <[e-mail]>
> ---
>  net/core/sock_map.c | 6 ++++++
>  1 file changed, 6 insertions(+)

Agree.

Acked-by: John Fastabend <[e-mail]>

[e-mail] April 2, 2024, 2:40 p.m. UTC | #2

Hello:

This patch was applied to bpf/bpf.git (master)
by Daniel Borkmann <[e-mail]>:

On Tue,  2 Apr 2024 12:46:21 +0200 you wrote:
> syzkaller started using corpuses where a BPF tracing program deletes
> elements from a sockmap/sockhash map. Because BPF tracing programs can be
> invoked from any interrupt context, locks taken during a map_delete_elem
> operation must be hardirq-safe. Otherwise a deadlock due to lock inversion
> is possible, as reported by lockdep:
> 
>        CPU0                    CPU1
> 
> [...]

Here is the summary with links:
  - [bpf] bpf, sockmap: Prevent lock inversion deadlock in map delete elem
    https://git.kernel.org/bpf/bpf/c/ff9105993240

You are awesome, thank you!

Patch

diff --git a/net/core/sock_map.c b/net/core/sock_map.c
index 27d733c0f65e..8598466a3805 100644
--- a/net/core/sock_map.c
+++ b/net/core/sock_map.c
@@ -411,6 +411,9 @@  static int __sock_map_delete(struct bpf_stab *stab, struct sock *sk_test,
 	struct sock *sk;
 	int err = 0;
 
+	if (irqs_disabled())
+		return -EOPNOTSUPP; /* locks here are hardirq-unsafe */
+
 	spin_lock_bh(&stab->lock);
 	sk = *psk;
 	if (!sk_test || sk_test == sk)
@@ -933,6 +936,9 @@  static long sock_hash_delete_elem(struct bpf_map *map, void *key)
 	struct bpf_shtab_elem *elem;
 	int ret = -ENOENT;
 
+	if (irqs_disabled())
+		return -EOPNOTSUPP; /* locks here are hardirq-unsafe */
+
 	hash = sock_hash_bucket_hash(key, key_size);
 	bucket = sock_hash_select_bucket(htab, hash);