Re: [PATCH v10 11/26] scsi: ufs: mediatek: Remove undocumented downstream reset cruft

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] The kobject reference for `phy_dev` is dropped too early during probe, leaving a dangling pointer stored in `host->phy_dev` that leads to a use-after-free in runtime PM callbacks.
- [High] The runtime PM usage count and enablement state for `phy_dev` are permanently leaked if the probe fails, and the PM state is never restored on driver removal.
--

--- Patch [11]: [PATCH v10 11/26] scsi: ufs: mediatek: Remove undocumented downstream reset cruft ---
commit 2e47ec1b9626d31e1288701fe0c4a25c195e38d4
Author: Nicolas Frattaroli <[email protected]>

scsi: ufs: mediatek: Remove undocumented downstream reset cruft

The MediaTek UFS host driver's probe function allows using a ti,syscon-reset
as a reset, without going through the appropriate abstractions or documenting
this in the binding. This commit removes this downstream code as it does not
belong here.

> diff --git a/drivers/ufs/host/ufs-mediatek.c b/drivers/ufs/host/ufs-mediatek.c
> index 65036c8390318..a2749e79514c4 100644
> --- a/drivers/ufs/host/ufs-mediatek.c
> +++ b/drivers/ufs/host/ufs-mediatek.c
> @@ -2402,38 +2402,12 @@ MODULE_DEVICE_TABLE(of, ufs_mtk_of_match);
>  static int ufs_mtk_probe(struct platform_device *pdev)
>  {
>  	int err;
> -	struct device *dev = &pdev->dev, *phy_dev = NULL;
> -	struct device_node *reset_node, *phy_node = NULL;
> -	struct platform_device *reset_pdev, *phy_pdev = NULL;
> -	struct device_link *link;
>  	struct ufs_hba *hba;
> +	struct platform_device *phy_pdev = NULL;
> +	struct device *dev = &pdev->dev, *phy_dev = NULL;
> +	struct device_node *phy_node = NULL;
>  	struct ufs_mtk_host *host;
>  
> -	reset_node = of_find_compatible_node(NULL, NULL,
> -					     "ti,syscon-reset");
> -	if (!reset_node) {
> -		dev_notice(dev, "find ti,syscon-reset fail\n");
> -		goto skip_reset;
> -	}
> -	reset_pdev = of_find_device_by_node(reset_node);
> -	if (!reset_pdev) {
> -		dev_notice(dev, "find reset_pdev fail\n");
> -		goto skip_reset;
> -	}
> -	link = device_link_add(dev, &reset_pdev->dev,
> -		DL_FLAG_AUTOPROBE_CONSUMER);
> -	put_device(&reset_pdev->dev);
> -	if (!link) {
> -		dev_notice(dev, "add reset device_link fail\n");
> -		goto skip_reset;
> -	}
> -	/* supplier is not probed */
> -	if (link->status == DL_STATE_DORMANT) {
> -		err = -EPROBE_DEFER;
> -		goto out;
> -	}
> -
> -skip_reset:
>  	/* find phy node */
>  	phy_node = of_parse_phandle(dev->of_node, "phys", 0);
>  
> @@ -2479,7 +2453,6 @@ static int ufs_mtk_probe(struct platform_device *pdev)
>  
>  out:
>  	of_node_put(phy_node);
> -	of_node_put(reset_node);
>  	return err;
>  }

[Severity: High]
This isn't a bug introduced by this patch, but could this lead to a
use-after-free of phy_dev?

In ufs_mtk_probe(), after looking up the PHY device, the kobject reference
is dropped almost immediately:

drivers/ufs/host/ufs-mediatek.c:ufs_mtk_probe() {
	phy_pdev = of_find_device_by_node(phy_node);
	if (!phy_pdev)
		goto skip_phy;
	phy_dev = &phy_pdev->dev;
	...
	put_device(phy_dev);
	...
}

However, the pointer is later saved into the host structure for later use:

drivers/ufs/host/ufs-mediatek.c:ufs_mtk_probe() {
	...
	if (phy_node && phy_dev) {
		host = ufshcd_get_variant(hba);
		host->phy_dev = phy_dev;
	}
	...
}

If the PHY device is later unbound via sysfs and its memory freed, would the
UFS host's runtime PM callbacks, such as ufs_mtk_runtime_suspend(),
dereference this dangling host->phy_dev pointer when calling
pm_runtime_put_sync()?

[Severity: High]
This is also a pre-existing issue, but is there a PM runtime usage count
leak in ufs_mtk_probe() on error paths?

If a failure occurs during initialization (e.g., ufshcd_pltfrm_init() fails),
the function jumps to the out label without calling pm_runtime_put_sync() or
pm_runtime_disable() for phy_dev:

drivers/ufs/host/ufs-mediatek.c:ufs_mtk_probe() {
	if (phy_node) {
		...
		pm_runtime_enable(phy_dev);
		pm_runtime_get_sync(phy_dev);
		...
	}
	...
	err = ufshcd_pltfrm_init(pdev, &ufs_hba_mtk_vops);
	if (err)
		goto out;
	...
out:
	of_node_put(phy_node);
	return err;
}

Additionally, ufs_mtk_remove() doesn't appear to tear down this state.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=11