@@ -1094,12 +1094,15 @@ static irqreturn_t venus_isr(struct venus_core *core)
{
struct venus_hfi_device *hdev = to_hfi_priv(core);
u32 status;
- void __iomem *cpu_cs_base = hdev->core->cpu_cs_base;
- void __iomem *wrapper_base = hdev->core->wrapper_base;
+ void __iomem *cpu_cs_base;
+ void __iomem *wrapper_base;
if (!hdev)
return IRQ_NONE;
+ cpu_cs_base = hdev->core->cpu_cs_base;
+ wrapper_base = hdev->core->wrapper_base;
+
status = readl(wrapper_base + WRAPPER_INTR_STATUS);
if (IS_V6(core)) {
if (status & WRAPPER_INTR_STATUS_A2H_MASK ||
@@ -1650,10 +1653,10 @@ void venus_hfi_destroy(struct venus_core *core)
{
struct venus_hfi_device *hdev = to_hfi_priv(core);
+ core->priv = NULL;
venus_interface_queues_release(hdev);
mutex_destroy(&hdev->lock);
kfree(hdev);
- core->priv = NULL;
core->ops = NULL;
}
Smatch is warning that: drivers/media/platform/qcom/venus/hfi_venus.c:1100 venus_isr() warn: variable dereferenced before check 'hdev' (see line 1097) The logic basically does: hdev = to_hfi_priv(core); with is translated to: hdev = core->priv; If the IRQ code can receive a NULL pointer for hdev, there's a bug there, as it will first try to de-reference the pointer, and then check if it is null. After looking at the code, it seems that this indeed can happen: Basically, the venus IRQ thread is started with: devm_request_threaded_irq() So, it will only be freed after the driver unbinds. In order to prevent the IRQ code to work with freed data, the logic at venus_hfi_destroy() sets core->priv to NULL, which would make the IRQ code to ignore any pending IRQs. There is, however a race condition, as core->priv is set to NULL only after being freed. So, we need also to move the core->priv = NULL to happen earlier. Signed-off-by: Mauro Carvalho Chehab <mchehab+huawei@kernel.org> --- drivers/media/platform/qcom/venus/hfi_venus.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)