Re: [PATCH v6 04/16] drm/bridge: synopsys: dw-dp: Add missing reinit_completion

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 5 potential issue(s) to consider:

Pre-existing issues:
- [High] The patch attempts to fix spurious completions from timed-out AUX transfers by adding `reinit_completion()`, but the fix is insufficient due to a TOCTOU race with the driver's threaded IRQ handler.
- [High] The driver fails to support DP AUX short reads, incorrectly rejecting them with -EBUSY instead of returning the number of bytes successfully transferred, which breaks I2C-over-AUX.
- [High] Missing hardware reset on AUX transfer timeout leaves the controller in an undefined/busy state.
- [High] The DP AUX channel and I2C adapter are registered and exposed before the hardware is initialized and interrupts are enabled, causing early access to fail with timeouts.
- [High] DP AUX transfers can execute with the PHY powered off, leading to hardware timeouts or crashes.
--

--- Patch [4]: [PATCH v6 04/16] drm/bridge: synopsys: dw-dp: Add missing reinit_completion ---
commit fc6159d600f530845f6f08962e449b720526ee08
Author: Sebastian Reichel <sebastian.reichel@collabora.com>
drm/bridge: synopsys: dw-dp: Add missing reinit_completion

This patch adds a reinit_completion call before sending an AUX command
to prevent subsequent transfers from immediately triggering if an earlier
one timed out and the interrupt fired later.

> diff --git a/drivers/gpu/drm/bridge/synopsys/dw-dp.c b/drivers/gpu/drm/bridge/synopsys/dw-dp.c
> index b28040d69fabf..2f0ee7ab7febd 100644
> --- a/drivers/gpu/drm/bridge/synopsys/dw-dp.c
> +++ b/drivers/gpu/drm/bridge/synopsys/dw-dp.c
> @@ -1465,6 +1465,8 @@ static ssize_t dw_dp_aux_transfer(struct drm_dp_aux *aux,
>  	if (WARN_ON(msg->size > 16))
>  		return -E2BIG;
>  
> +	reinit_completion(&dp->complete);
> +

[Severity: High]
This isn't a bug introduced by this patch, but the attempt to fix the
spurious completions seems incomplete due to a race condition. Because
reinit_completion is called locklessly here, could the threaded IRQ handler
execute concurrently?

If dw_dp_irq is processing a delayed interrupt from a previous timeout:

dw_dp_irq() {
    ...
    complete(&dp->complete);
}

Doesn't this mean it could call complete immediately after this
reinit_completion, instantly satisfying the new wait and causing the driver
to read garbage hardware status?

>  	switch (msg->request & ~DP_AUX_I2C_MOT) {
>  	case DP_AUX_NATIVE_WRITE:
>  	case DP_AUX_I2C_WRITE:

[Severity: High]
This is a pre-existing issue, but does this driver fail to support DP AUX
short reads?

Later in dw_dp_aux_transfer:

    if (count != msg->size)
        return -EBUSY;

If the hardware returns fewer bytes than requested, returning -EBUSY causes
the DRM core helper drm_dp_i2c_do_msg to endlessly retry the exact same
full-length transfer without advancing its buffer. Could this break
communication with I2C devices on the DP bus?

[Severity: High]
This is a pre-existing issue, but is a hardware reset missing when an AUX
transfer times out?

In the error path of dw_dp_aux_transfer:

    if (!status) {
        dev_err(dp->dev, "timeout waiting for AUX reply\n");
        return -ETIMEDOUT;
    }

If wait_for_completion_timeout expires, returning -ETIMEDOUT without
resetting the underlying AUX controller or clearing the pending state might
leave the hardware state machine permanently stuck. Will subsequent transfers
fail because the hardware is still busy?

[Severity: High]
This isn't a bug introduced by this patch, but is the AUX channel registered
before interrupts are enabled?

Looking at dw_dp_bind, the channel is registered early:

    ret = drm_dp_aux_register(&dp->aux);

But interrupts are enabled much later in the same function:

    enable_irq(dp->irq);

If a userspace tool like ddcutil or an I2C auto-prober attempts a transaction
in this window, won't dw_dp_aux_transfer fail because the hardware cannot
yet generate interrupts?

[Severity: High]
This is a pre-existing issue, but can DP AUX transfers execute with the PHY
powered off?

Direct accesses from userspace via the character device call directly into:

static ssize_t dw_dp_aux_transfer(struct drm_dp_aux *aux,
                                  struct drm_dp_aux_msg *msg)

Internal callers wrap this to ensure the PHY is powered on:

dw_dp_bridge_edid_read() {
    ...
    ret = phy_power_on(dp->phy);
    if (ret)
        return NULL;

    edid = drm_edid_read_ddc(connector, &dp->aux.ddc);

    phy_power_off(dp->phy);
    ...
}

Since dw_dp_aux_transfer itself doesn't check or manage the PHY power state,
could a direct userspace access while the display is suspended cause MMIO
timeouts or system crashes?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260724-synopsys-dw-dp-improvements-v6-0-041d99a19c4e@collabora.com?part=4