Re: [PATCH v2 1/2] nfc: llcp: avoid double release/put on LLCP_CLOSED in nfc_llcp_recv_disc()

[くさあさ]

> I ask why do you think your code is correct.

Because the current code can execute release_sock()/nfc_llcp_sock_put() twice
for the single ref acquired by nfc_llcp_sock_get(), and it can keep operating
on llcp_sock/sk after the first put/unlock. The fix ensures exactly one cleanup
(recv_disc) or exits immediately after the CLOSED cleanup (recv_hdlc), while
keeping the existing DM_DISC send behavior.

> The refcnt has imbalance only if you assume initial refcnt was 0.

No. Let baseline refcnt be N before nfc_llcp_sock_get(); sock_hold() makes N+1.
CLOSED path put -> N, common exit put -> N-1. So it drops one extra ref
regardless of N (whether it immediately frees depends on N). git blame shows
both cleanup sites trace back to d646960f7986.

I will not send a new revision until this discussion is resolved.

Best regards,
Qianchang

On Thu, Dec 18, 2025 at 7:22 PM Krzysztof Kozlowski <krzk@xxxxxxxxxx> wrote:
>
> On 17/12/2025 14:05, くさあさ wrote:
> > Hi Krzysztof,
> >
> > Sorry about that — my previous response might not have made it to the
> > list/thread.
> > Replying here to address your concerns before sending v3.
> >
> > 1) DM_DISC reply after LLCP_CLOSED
> > This is not a new behavior introduced by my change. In the old code, the
> > LLCP_CLOSED branch did release_sock() and nfc_llcp_sock_put(), but it did not
> > return/goto, so execution continued and still reached nfc_llcp_send_dm(...,
> > LLCP_DM_DISC) afterwards. The disc patch only removes the redundant
> > CLOSED-branch
> > cleanup so release_sock()/nfc_llcp_sock_put() are performed exactly once via the
> > common exit path, while keeping the existing DM_DISC reply behavior.
>
> I understand that you did not change the flow. I did not claim you did.
> I ask why do you think your code is correct.
>
> Do not top post and do not send new versions while the discussion is
> still going.
> >
> > 2) Initial refcount / double free concern
> > nfc_llcp_recv_disc()/recv_hdlc() take an extra reference via nfc_llcp_sock_get()
> > (sock_hold()). The issue is the mismatched put/unlock: the CLOSED branch drops
> > the reference and releases the lock, and then the common exit path does the same
> > again. This is a refcount/locking imbalance regardless of whether it immediately
> > frees the object, and it may become a UAF depending on timing/refcounting.
>
> You did not really address the problem. The refcnt has imbalance only if
> you assume initial refcnt was 0.
>
>
>
> Best regards,
> Krzysztof