Message ID | 73570a5dfe7b3411d256367d4a2a02169aa9b900.1617867599.git.mchehab+huawei@kernel.org |
---|---|
State | Accepted |
Commit | 686ee9b6253f9b6d7f1151e73114698940cc0894 |
Headers | show |
Series | None | expand |
Hi Mauro, On 4/8/21 10:40 AM, Mauro Carvalho Chehab wrote: > 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(-) Acked-by: Stanimir Varbanov <stanimir.varbanov@linaro.org> > > diff --git a/drivers/media/platform/qcom/venus/hfi_venus.c b/drivers/media/platform/qcom/venus/hfi_venus.c > index cebb20cf371f..ce98c523b3c6 100644 > --- a/drivers/media/platform/qcom/venus/hfi_venus.c > +++ b/drivers/media/platform/qcom/venus/hfi_venus.c > @@ -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; > } > >
diff --git a/drivers/media/platform/qcom/venus/hfi_venus.c b/drivers/media/platform/qcom/venus/hfi_venus.c index cebb20cf371f..ce98c523b3c6 100644 --- a/drivers/media/platform/qcom/venus/hfi_venus.c +++ b/drivers/media/platform/qcom/venus/hfi_venus.c @@ -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(-)