Re: [PATCH v11 05/21] drm/bridge: synopsys: dw-dp: Add missing mutex cleanups on module removal

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:

Pre-existing issues:
- [High] The bridge's atomic enable and disable callbacks execute sleeping functions, violating atomic context rules for display drivers.
- [High] Missing devm-based cancellation of `hpd_work` and unregistration of `dp->aux` leads to a Use-After-Free of devres-managed resources on physical device unbind.
- [High] Out-of-bounds reads and writes in link training arrays due to unbounded lane count derived from untrusted DPCD.
- [High] Missing `reinit_completion()` in DP AUX transfer causes premature completion and data corruption.
--

--- Patch [5]: [PATCH v11 05/21] drm/bridge: synopsys: dw-dp: Add missing mutex cleanups on module removal ---
commit 1e2ea3ee5b8e223fa29b59cba2907fd3cca590e6
Author: Sebastian Reichel <sebastian.reichel@collabora.com>

drm/bridge: synopsys: dw-dp: Add missing mutex cleanups on module removal

The driver is currently missing to fully clean up after itself. Ensure
that the mutex is cleaned up.

> diff --git a/drivers/gpu/drm/bridge/synopsys/dw-dp.c b/drivers/gpu/drm/bridge/synopsys/dw-dp.c
> index 52e3b37259a69..b1856418f4419 100644
> --- a/drivers/gpu/drm/bridge/synopsys/dw-dp.c
> +++ b/drivers/gpu/drm/bridge/synopsys/dw-dp.c
> @@ -2040,10 +2040,13 @@ struct dw_dp *dw_dp_probe(struct platform_device *pdev, const struct dw_dp_plat_
>  	dp->pixel_mode = plat_data->pixel_mode;
>  	dp->plat_data.max_link_rate = plat_data->max_link_rate;
>  
> -	mutex_init(&dp->irq_lock);
>  	INIT_WORK(&dp->hpd_work, dw_dp_hpd_work);

[Severity: High]
This is a pre-existing issue, but by making the irq_lock managed by devres,
does this workqueue now trigger a use-after-free on device unbind?

If the physical device is unbound (e.g., via sysfs) while the DRM master is
still active, devres will destroy dp->irq_lock and free the dp structure.
However, hpd_work is only cancelled in dw_dp_bridge_detach(), which is tied
to the DRM master's lifecycle. 

If the worker executes:

drivers/gpu/drm/bridge/synopsys/dw-dp.c:dw_dp_hpd_work() {
    ...
    mutex_lock(&dp->irq_lock);
    ...
}

Will it attempt to lock the destroyed mutex and access the freed dp structure?
Should we use devm_add_action_or_reset() to ensure the work is cancelled
during physical device unbind?

>  	init_completion(&dp->complete);

[Severity: High]
This isn't a bug introduced by this patch, but is it possible for DP AUX
transfers to prematurely complete due to a missing reinit_completion()?

If dw_dp_aux_transfer() times out, but the hardware eventually signals the
interrupt later, dw_dp_irq() calls complete(&dp->complete). 

Because dw_dp_aux_transfer() fails to call reinit_completion() before starting
the next transfer, will the subsequent call to wait_for_completion_timeout()
return immediately and incorrectly read the status registers?

>  
> +	ret = devm_mutex_init(dev, &dp->irq_lock);
> +	if (ret)
> +		return ERR_PTR(ret);
> +
>  	res = devm_platform_ioremap_resource(pdev, 0);

[Severity: High]
This is a pre-existing issue, but do the atomic enable and disable callbacks
execute sleeping functions?

The DRM bridge's atomic disable callback (dw_dp_bridge_atomic_disable) is
reached from the atomic commit tail, but it invokes dw_dp_reset(), which
uses blocking functions:

drivers/gpu/drm/bridge/synopsys/dw-dp.c:dw_dp_reset() {
    ...
    disable_irq(dp->irq);
    ...
    usleep_range(10, 20);
    ...
    regmap_read_poll_timeout(...);
    ...
}

Similarly, dw_dp_bridge_atomic_enable() reaches dw_dp_link_train_fast()
which calls usleep_range().  Since functions reached from
drm_atomic_helper_commit_tail() run in atomic context, will this cause
scheduling-while-atomic bugs or kernel warnings?

[Severity: High]
This is a pre-existing issue, but is there a risk of out-of-bounds reads and
writes during link training due to an unbounded lane count?

In dw_dp_link_parse(), link->lanes is determined by reading the capability
from the DP sink:

    link->lanes = min_t(u8, phy_get_bus_width(dp->phy),
                        drm_dp_max_lane_count(link->dpcd));

If the PHY driver returns a bus width greater than 4, and the untrusted DPCD
reports more than 4 lanes, link->lanes will exceed 4.  

The link training functions then iterate up to link->lanes:

drivers/gpu/drm/bridge/synopsys/dw-dp.c:dw_dp_link_train_update_vs_emph() {
    ...
    for (i = 0; i < lanes; i++) {
        phy_cfg.dp.voltage[i] = vs[i];
        phy_cfg.dp.pre[i] = pe[i];
    }
    ...
}

Will this cause out-of-bounds writes to arrays like phy_cfg.dp.voltage and
phy_cfg.dp.pre (which are sized for 4 elements) when a non-compliant or
malicious DP sink is connected?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260806-synopsys-dw-dp-improvements-v11-0-0d508505f383@collabora.com?part=5