As &mc->vc.lock is acquired by milbeaut_hdmac_interrupt() under irq
context, other acquisition of the same lock under process context should
disable irq, otherwise deadlock could happen if the irq preempts the
execution of process context code while the lock is held in process context
on the same CPU.

milbeaut_hdmac_chan_config(), milbeaut_hdmac_chan_resume() and
milbeaut_hdmac_chan_pause() are such callback functions not disable irq by
default.

Possible deadlock scenario:
milbeaut_hdmac_chan_config()
    -> spin_lock(&mc->vc.lock)
        <hard interruption>
        -> milbeaut_hdmac_interrupt()
        -> spin_lock(&mc->vc.lock); (deadlock here)

This flaw was found by an experimental static analysis tool I am developing
for irq-related deadlock.

The tentative patch fixes the potential deadlock by spin_lock_irqsave() in
the three callback functions to disable irq while lock is held.

Signed-off-by: Chengfeng Ye <dg573847474@gmail.com>
---
 drivers/dma/milbeaut-hdmac.c | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/drivers/dma/milbeaut-hdmac.c b/drivers/dma/milbeaut-hdmac.c
index 1b0a95892627..6151c830ff6e 100644
--- a/drivers/dma/milbeaut-hdmac.c
+++ b/drivers/dma/milbeaut-hdmac.c
@@ -214,10 +214,11 @@ milbeaut_hdmac_chan_config(struct dma_chan *chan, struct dma_slave_config *cfg)
 {
 	struct virt_dma_chan *vc = to_virt_chan(chan);
 	struct milbeaut_hdmac_chan *mc = to_milbeaut_hdmac_chan(vc);
+	unsigned long flags;
 
-	spin_lock(&mc->vc.lock);
+	spin_lock_irqsave(&mc->vc.lock, flags);
 	mc->cfg = *cfg;
-	spin_unlock(&mc->vc.lock);
+	spin_unlock_irqrestore(&mc->vc.lock, flags);
 
 	return 0;
 }
@@ -226,13 +227,14 @@ static int milbeaut_hdmac_chan_pause(struct dma_chan *chan)
 {
 	struct virt_dma_chan *vc = to_virt_chan(chan);
 	struct milbeaut_hdmac_chan *mc = to_milbeaut_hdmac_chan(vc);
+	unsigned long flags;
 	u32 val;
 
-	spin_lock(&mc->vc.lock);
+	spin_lock_irqsave(&mc->vc.lock, flags);
 	val = readl_relaxed(mc->reg_ch_base + MLB_HDMAC_DMACA);
 	val |= MLB_HDMAC_PB;
 	writel_relaxed(val, mc->reg_ch_base + MLB_HDMAC_DMACA);
-	spin_unlock(&mc->vc.lock);
+	spin_unlock_irqrestore(&mc->vc.lock, flags);
 
 	return 0;
 }
@@ -241,13 +243,14 @@ static int milbeaut_hdmac_chan_resume(struct dma_chan *chan)
 {
 	struct virt_dma_chan *vc = to_virt_chan(chan);
 	struct milbeaut_hdmac_chan *mc = to_milbeaut_hdmac_chan(vc);
+	unsigned long flags;
 	u32 val;
 
-	spin_lock(&mc->vc.lock);
+	spin_lock_irqsave(&mc->vc.lock, flags);
 	val = readl_relaxed(mc->reg_ch_base + MLB_HDMAC_DMACA);
 	val &= ~MLB_HDMAC_PB;
 	writel_relaxed(val, mc->reg_ch_base + MLB_HDMAC_DMACA);
-	spin_unlock(&mc->vc.lock);
+	spin_unlock_irqrestore(&mc->vc.lock, flags);
 
 	return 0;
 }
-- 
2.17.1

On Wed, 26 Jul 2023 at 04:06, Chengfeng Ye <dg573847474@gmail.com> wrote:
>
> As &mc->vc.lock is acquired by milbeaut_hdmac_interrupt() under irq
> context, other acquisition of the same lock under process context should
> disable irq, otherwise deadlock could happen if the irq preempts the
> execution of process context code while the lock is held in process context
> on the same CPU.
>
> milbeaut_hdmac_chan_config(), milbeaut_hdmac_chan_resume() and
> milbeaut_hdmac_chan_pause() are such callback functions not disable irq by
> default.
>
> Possible deadlock scenario:
> milbeaut_hdmac_chan_config()
>     -> spin_lock(&mc->vc.lock)
>         <hard interruption>
>         -> milbeaut_hdmac_interrupt()
>         -> spin_lock(&mc->vc.lock); (deadlock here)
>
> This flaw was found by an experimental static analysis tool I am developing
> for irq-related deadlock.
>
> The tentative patch fixes the potential deadlock by spin_lock_irqsave() in
> the three callback functions to disable irq while lock is held.
>
> Signed-off-by: Chengfeng Ye <dg573847474@gmail.com>
> ---
>  drivers/dma/milbeaut-hdmac.c | 15 +++++++++------
>  1 file changed, 9 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/dma/milbeaut-hdmac.c b/drivers/dma/milbeaut-hdmac.c
> index 1b0a95892627..6151c830ff6e 100644
> --- a/drivers/dma/milbeaut-hdmac.c
> +++ b/drivers/dma/milbeaut-hdmac.c
> @@ -214,10 +214,11 @@ milbeaut_hdmac_chan_config(struct dma_chan *chan, struct dma_slave_config *cfg)
>  {
>         struct virt_dma_chan *vc = to_virt_chan(chan);
>         struct milbeaut_hdmac_chan *mc = to_milbeaut_hdmac_chan(vc);
> +       unsigned long flags;
>
> -       spin_lock(&mc->vc.lock);
> +       spin_lock_irqsave(&mc->vc.lock, flags);
>
while at it, maybe also use vc->lock, instead of mc->vc.lock here and
in other two places, just like the rest of driver.

Acked-by: Jassi Brar <jaswinder.singh@linaro.org>

thanks.

> while at it, maybe also use vc->lock, instead of mc->vc.lock here and
> in other two places, just like the rest of driver.

Thanks for the review, v2 patch is just sent to address the problem.

Thanks,
Chengfeng

Hi maintainers,

May I ask if anyone would like to review the patch?

Thanks,
Chengfeng