Re: [PATCH] dmaengine: stm32-dmamux: don't resume runtime PM under a spinlock

From: Amelie Delaunay

Date: Wed Nov 18 2026 - 10:25:40 EST


Hi,

On 11/18/26 09:28, Hyunjung Ko wrote:
stm32_dmamux_route_allocate() calls pm_runtime_resume_and_get() while
holding dmamux->lock with interrupts disabled. The runtime-resume
callback stm32_dmamux_runtime_resume() calls clk_prepare_enable(), which
can sleep, and the device is not marked pm_runtime_irq_safe(), so this
can sleep in atomic context.

Indeed.


The lock does not protect anything here: the dma_inuse bitmap is already
updated under the lock earlier in the function, and this second critical
section only wraps the pm_runtime call. Drop the lock around it.

The lock is misplaced and completely useless here. Your patch is correct but incomplete.
The lock should instead protect the CCR register write, as in stm32_dmamux_free(), which also needs to be fixed.


Fixes: 4f3ceca254e0 ("dmaengine: stm32-dmamux: Add PM Runtime support")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Hyunjung Ko <hj351016@xxxxxxxxx>
---
drivers/dma/stm32/stm32-dmamux.c | 6 +-----
1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/drivers/dma/stm32/stm32-dmamux.c b/drivers/dma/stm32/stm32-dmamux.c
index 57318715d9f3..890328b3d302 100644
--- a/drivers/dma/stm32/stm32-dmamux.c
+++ b/drivers/dma/stm32/stm32-dmamux.c
@@ -139,13 +139,9 @@ static void *stm32_dmamux_route_allocate(struct of_phandle_args *dma_spec,
}
/* Set dma request */
- spin_lock_irqsave(&dmamux->lock, flags);
ret = pm_runtime_resume_and_get(&pdev->dev);
- if (ret < 0) {
- spin_unlock_irqrestore(&dmamux->lock, flags);
+ if (ret < 0)
goto err_put_dma_spec_np;
- }
- spin_unlock_irqrestore(&dmamux->lock, flags);
mux->request = dma_spec->args[0];

In addition, I suggest:
@@
- stm32_dmamux_write(dmamux->iomem, STM32_DMAMUX_CCR(mux->chan_id),
- mux->request);
+ spin_lock_irqsave(&dmamux->lock, flags);
+ stm32_dmamux_write(dmamux->iomem, STM32_DMAMUX_CCR(mux->chan_id),
+ mux->request);
+ spin_unlock_irqrestore(&dmamux->lock, flags);
+

And moving pm_runtime_put_sync() in stm32_dmamux_free():
@@ static void stm32_dmamux_free(struct device *dev, void *route_data)
/* Clear dma request */
spin_lock_irqsave(&dmamux->lock, flags);

stm32_dmamux_write(dmamux->iomem, STM32_DMAMUX_CCR(mux->chan_id), 0);
clear_bit(mux->chan_id, dmamux->dma_inuse);

- pm_runtime_put_sync(dev);
-
spin_unlock_irqrestore(&dmamux->lock, flags);

+ pm_runtime_put_sync(dev);
+
dev_dbg(dev, "Unmapping DMAMUX(%u) to DMA%u(%u)\n",
mux->request, mux->master, mux->chan_id);


Runtime PM calls are made outside the lock, which is held only for the CCR write and the dma_inuse bitmap.

Regards,
Amelie