Re: [PATCH v3] nfc: nci: ignore unexpected CORE_RESET_NTF and CORE_RESET_RSP

[Simon Horman <horms@kernel.org>]

On Sun, Sep 27, 2026 at 09:29:14PM +0800, Yuchao Zhang wrote:
> Commit bcd684aace34 ("net/nfc/nci: Support NCI 2.x initial sequence")
> added handling of CORE_RESET_NTF in nci_core_reset_ntf_packet(). When
> received, it updates ndev->nci_ver, ndev->manufact_id, and
> ndev->manufact_specific_info, and calls nci_req_complete(ndev,
> NCI_STATUS_OK) to finish the pending reset request.
> 
> However, unlike other notification handlers in ntf.c (which validate
> ndev->state before completing requests), nci_core_reset_ntf_packet()
> does not check whether a core reset request is actually pending.
> If an unsolicited or delayed CORE_RESET_NTF arrives (e.g. after a reset
> command times out or from a misbehaving NFCC), it unconditionally:
> 1. Completes whatever request is currently in-flight (such as CORE_INIT,
>    RF_DISCOVER, or CONN_CREATE) with NCI_STATUS_OK, leading to kernel
>    state desynchronization.
> 2. Overwrites ndev->nci_ver and manufacturer info. Because
>    ndev->nci_ver is used as a selector for subsequent packet formats
>    and parsers (e.g., in nci_open_device() and nci_core_init_rsp_packet()),
>    unexpectedly modifying it can cause protocol format confusion.
> 
> A similar issue exists in nci_core_reset_rsp_packet(): an unexpected or
> delayed response packet can prematurely complete an unrelated in-flight
> request.
> 
> Fix this by ensuring CORE_RESET_NTF and CORE_RESET_RSP are only processed
> by the core layer when a reset command is actively awaiting them:
> - Set NCI_RESET_PENDING in nci_reset_req() when sending CORE_RESET_CMD.
> - In nci_core_reset_rsp_packet(), ignore the response if NCI_RESET_PENDING
>   is not set. If set, atomically clear the flag and complete the request
>   on failure or for NCI 1.x (checking skb->len >= sizeof(*rsp)).
> - In __nci_request(), ensure NCI_RESET_PENDING is cleared upon request
>   completion, cancellation, or timeout.
> - In nci_core_reset_ntf_packet(), fail any in-flight reset request if the
>   controller reports an unrecoverable error (0x00), and only process
>   command-triggered notifications (0x02). Non-command triggers (e.g.
>   power-on 0x01 or proprietary triggers) and unexpected notifications
>   return 0 without modifying core state, ensuring driver-specific
>   notification handlers (such as fdp firmware patch handling) still
>   receive them.
> - In nci_core_reset_ntf_packet(), support notifications shorter than 9
>   bytes (e.g. NCI 2.0 frames with manufacturer_specific_len == 0),
>   validating the fixed header length and safely bounding
>   manufact_specific_info reads.
> 
> Fixes: bcd684aace34 ("net/nfc/nci: Support NCI 2.x initial sequence")
> Cc: stable@vger.kernel.org
> Signed-off-by: Yuchao Zhang <ndaugoing@gmail.com>
> ---
> v3:
>  - In nci_core_reset_ntf_packet(), do not reject notifications shorter than
>    9 bytes with -EINVAL; instead allow valid short frames (e.g. 2-byte NCI 1.x
>    or 5-byte NCI 2.0 with manufacturer_specific_len == 0), safely cap field
>    reads, and return 0 so driver-registered hooks (such as fdp firmware
>    patching) are not skipped (per Sashiko AI).
>  - Fix trigger semantics: in NCI, CORE_RESET_CMD trigger is 0x02
>    (NCI_RESET_TRIGGER_RESET_CMD). If controller reports an unrecoverable error
>    (0x00, NCI_RESET_TRIGGER_ERROR) while reset is pending, fail the request
>    instead of completing with OK; ignore power-on (0x01) and proprietary
>    triggers at core layer (per Sashiko AI).
>  - In nci_core_reset_rsp_packet(), use test_and_clear_bit(NCI_RESET_PENDING)
>    upon completing the request to consume the flag atomically (per Sashiko AI).
>  - Link to v2: https://lore.kernel.org/netdev/20260922102144.39689-1-ndaugoing@gmail.com/
> 
> v2:
>  - Do not abort the notification pipeline on unexpected CORE_RESET_NTF (return 0
>    instead of -EINVAL), preserving driver-level hooks (e.g. fdp firmware
>    patching) per Simon Horman.
>  - Skip power-on notifications (trigger 0x01, identical in NCI 1.0 and 2.0)
>    while gating both NCI 1.0 (0x00) and NCI 2.0 (0x02) command-triggered resets
>    on NCI_RESET_PENDING alone.
>  - Check NCI_RESET_PENDING in nci_core_reset_rsp_packet() to avoid completing
>    unrelated requests on unexpected responses.
>  - Explicitly check skb->len >= sizeof(*rsp) in nci_core_reset_rsp_packet()
>    for NCI 1.x handling.
>  - Link to v1: https://lore.kernel.org/netdev/20260918013337.82214-1-ndaugoing@gmail.com/

Reviewed-by: Simon Horman <horms@kernel.org>