Skip to contentSkip to conversation

ufs: setting "hba" private pointer too late -- oops in ufshcd_devfreq_get_dev_status()

DISCUSSION Alexey Dobriyan 2 participants 4 messages Nov 10, 2021 – Nov 17, 2021 SCSI

lore
Share & tools
lore
#0 lore raw

I've stumbled into a race while working on an earlier kernel, but it looks like mainline is affected as well.

        err = ufshcd_init(hba, mmio_base, irq);
		async_schedule(ufshcd_async_scan, hba);
		ufshcd_add_lus(hba);
		if (ufshcd_is_clkscaling_supported(hba)) {
			[enable devfreq]

        platform_set_drvdata(pdev, hba);

Device's private pointer is set too late, as devfreq hook get HBA pointer from private data and uses it:

	static int ufshcd_devfreq_get_dev_status(struct device *dev, struct devfreq_dev_status *stat)
	{
	        struct ufs_hba *hba = dev_get_drvdata(dev);
		if (!ufshcd_is_clkscaling_supported(hba))
			return -EINVAL;
Unable to handle kernel NULL pointer dereference at virtual address ...0f10

pc : ufshcd_devfreq_get_dev_status
lr : devfreq_simple_ondemand_func
update_devfreq
devfreq_monitor

I reproduced it by turning async LU scan into sync, so it is easier to trigger.

#1 lore raw
Quoted 21 lines · Alexey Dobriyan

I've stumbled into a race while working on an earlier kernel, but it looks like mainline is affected as well.

         err = ufshcd_init(hba, mmio_base, irq);
		async_schedule(ufshcd_async_scan, hba);
		ufshcd_add_lus(hba);
		if (ufshcd_is_clkscaling_supported(hba)) {
			[enable devfreq]

         platform_set_drvdata(pdev, hba);

Device's private pointer is set too late, as devfreq hook get HBA pointer from private data and uses it:

	static int ufshcd_devfreq_get_dev_status(struct device *dev, struct devfreq_dev_status *stat)
	{
	        struct ufs_hba *hba = dev_get_drvdata(dev);
		if (!ufshcd_is_clkscaling_supported(hba))
			return -EINVAL;
Unable to handle kernel NULL pointer dereference at virtual address ...0f10
Quoted 8 lines

pc : ufshcd_devfreq_get_dev_status
lr : devfreq_simple_ondemand_func
update_devfreq
devfreq_monitor

I reproduced it by turning async LU scan into sync, so it is easier to trigger.

Hi Alexey,

Thanks for having reported this. Do you perhaps plan to post a patch to fix this?

Thanks,

Bart.

#2 lore raw
Quoted 28 lines · Bart Van Assche
Quoted 27 lines · Alexey Dobriyan

I've stumbled into a race while working on an earlier kernel, but it looks like mainline is affected as well.

         err = ufshcd_init(hba, mmio_base, irq);
		async_schedule(ufshcd_async_scan, hba);
		ufshcd_add_lus(hba);
		if (ufshcd_is_clkscaling_supported(hba)) {
			[enable devfreq]

         platform_set_drvdata(pdev, hba);

Device's private pointer is set too late, as devfreq hook get HBA pointer from private data and uses it:

	static int ufshcd_devfreq_get_dev_status(struct device *dev, struct devfreq_dev_status *stat)
	{
	        struct ufs_hba *hba = dev_get_drvdata(dev);
		if (!ufshcd_is_clkscaling_supported(hba))
			return -EINVAL;
Unable to handle kernel NULL pointer dereference at virtual address ...0f10

pc : ufshcd_devfreq_get_dev_status
lr : devfreq_simple_ondemand_func
update_devfreq
devfreq_monitor

Quoted 7 lines
Quoted 2 lines

I reproduced it by turning async LU scan into sync, so it is easier to trigger.

Hi Alexey,

Thanks for having reported this. Do you perhaps plan to post a patch to fix this?

Not really, my workaround is

	if (!hba) {
		return -EINVAL;
	}

But it is likely incorrect.

#3 lore raw
Quoted 7 lines · Alexey Dobriyan

Not really, my workaround is

	if (!hba) {
		return -EINVAL;
	}

But it is likely incorrect.

How about the untested patch below?

Thanks,

Bart.

---
 drivers/scsi/ufs/tc-dwc-g210-pci.c | 1 -
 drivers/scsi/ufs/ufshcd-pci.c      | 2 --
 drivers/scsi/ufs/ufshcd-pltfrm.c   | 2 --
 drivers/scsi/ufs/ufshcd.c          | 7 +++++++
 4 files changed, 7 insertions(+), 5 deletions(-)
drivers/scsi/ufs/tc-dwc-g210-pci.c +0 -1
index 679289e1a78e..7b08e2e07cc5 100644
@@ -110,7 +110,6 @@ tc_dwc_g210_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 		return err;
 	}
 
-	pci_set_drvdata(pdev, hba);
 	pm_runtime_put_noidle(&pdev->dev);
 	pm_runtime_allow(&pdev->dev);
 
drivers/scsi/ufs/ufshcd-pci.c +0 -2
index 51424557810d..a673eedb2f05 100644
@@ -522,8 +522,6 @@ ufshcd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 		return err;
 	}
 
-	pci_set_drvdata(pdev, hba);
-
 	hba->vops = (struct ufs_hba_variant_ops *)id->driver_data;
 
 	err = ufshcd_init(hba, mmio_base, pdev->irq);
drivers/scsi/ufs/ufshcd-pltfrm.c +0 -2
index eaeae83b999f..8b16bbbcb806 100644
@@ -361,8 +361,6 @@ int ufshcd_pltfrm_init(struct platform_device *pdev,
 		goto dealloc_host;
 	}
 
-	platform_set_drvdata(pdev, hba);
-
 	pm_runtime_set_active(&pdev->dev);
 	pm_runtime_enable(&pdev->dev);
 
drivers/scsi/ufs/ufshcd.c +7 -0
index a7a8d3fbb89d..87b8bd837342 100644
@@ -9608,6 +9608,13 @@ int ufshcd_init(struct ufs_hba *hba, void __iomem *mmio_base, unsigned int irq)
 	 */
 	mb();
 
+	/*
+	 * dev_set_drvdata() must be called before
+	 * async_schedule(ufshcd_async_scan, hba) since the callbacks registered
+	 * with the devfreq framework use dev_get_drvdata().
+	 */
+	dev_set_drvdata(dev, hba);
+
 	/* IRQ registration */
 	err = devm_request_irq(dev, irq, ufshcd_intr, IRQF_SHARED, UFSHCD, hba);
 	if (err) {