Re: [PATCH] net/tls: Fix slab-use-after-free in tls_encrypt_done

[Juntong Deng]

On 2023/10/17 18:31, Paolo Abeni wrote:
> On Thu, 2023-10-12 at 19:02 +0800, Juntong Deng wrote:
> > In the current implementation, ctx->async_wait.completion is completed
> > after spin_lock_bh, which causes tls_sw_release_resources_tx to
> > continue executing and return to tls_sk_proto_cleanup, then return
> > to tls_sk_proto_close, and after that enter tls_sw_free_ctx_tx to kfree
> > the entire struct tls_context (including ctx->encrypt_compl_lock).
> >
> > Since ctx->encrypt_compl_lock has been freed, subsequent spin_unlock_bh
> > will result in slab-use-after-free error. Due to SMP, even using
> > spin_lock_bh does not prevent tls_sw_release_resources_tx from continuing
> > on other CPUs. After tls_sw_release_resources_tx is woken up, there is no
> > attempt to hold ctx->encrypt_compl_lock again, therefore everything
> > described above is possible.
> >
> > The fix is to put complete(&ctx->async_wait.completion) after
> > spin_unlock_bh, making the release after the unlock. Since complete is
> > only executed if pending is 0, which means this is the last record, there
> > is no need to worry about race condition causing duplicate completes.
> >
> > Reported-by: syzbot+29c22ea2d6b2c5fd2eae@xxxxxxxxxxxxxxxxxxxxxxxxx
> > Closes: https://syzkaller.appspot.com/bug?extid=29c22ea2d6b2c5fd2eae
> > Signed-off-by: Juntong Deng <juntong.deng@xxxxxxxxxxx>
>
> Have you tested this patch vs the syzbot reproducer?
>
> I think the following race is still present:
>
> CPU0                            CPU1
> tls_sw_release_resources_tx     tls_encrypt_done
> spin_lock_bh
> spin_unlock_bh
>                                  spin_lock_bh
>                                  spin_unlock_bh
>                                  complete
>
> wait
> // ...
> tls_sk_proto_close
>
>                                  test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask
> 				// UaF
>
> regardless of 'complete()' being invoked before or after the
> 'spin_unlock_bh()'.
>
> Paolo

Yes, I think you are right.

My previous thought was that test_and_set_bit() is only called if
'ready' is true, but 'ready' will only be true on the first record,
and complete() is only called when processing the last record.

I simply thought before that the first record would not be the last
record, so I thought before that the test_and_set_bit() would not be
called when complete() was called.

But your reply inspired me and I thought about it carefully and the
situation with only one record is possible.

I will make version 2 patch to solve this problem.

Thanks.