As xgene_dma_cleanup_descriptors() is invoked by both tasklet
xgene_dma_tasklet_cb() under softirq context and
xgene_dma_free_chan_resources() callback that executed under process
context, the lock aquicision of &chan->lock inside
xgene_dma_cleanup_descriptors() should disable irq otherwise deadlock
could happen if the tasklet softirq preempts the execution of process
context code while the lock is held in process context on the same CPU.

Possible deadlock scenario:
xgene_dma_free_chan_resources()
    -> xgene_dma_cleanup_descriptors()
    -> spin_lock(&chan->lock)
        <tasklet softirq>
        -> xgene_dma_tasklet_cb()
        -> xgene_dma_cleanup_descriptors()
        -> spin_lock(&chan->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
plx_dma_process_desc() to disable irq while lock is held.

Signed-off-by: Chengfeng Ye <dg573847474@gmail.com>
---
 drivers/dma/xgene-dma.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/dma/xgene-dma.c b/drivers/dma/xgene-dma.c
index 3589b4ef50b8..e766511badcf 100644
--- a/drivers/dma/xgene-dma.c
+++ b/drivers/dma/xgene-dma.c
@@ -689,11 +689,12 @@ static void xgene_dma_cleanup_descriptors(struct xgene_dma_chan *chan)
 	struct xgene_dma_desc_sw *desc_sw, *_desc_sw;
 	struct xgene_dma_desc_hw *desc_hw;
 	struct list_head ld_completed;
+	unsigned long flags;
 	u8 status;
 
 	INIT_LIST_HEAD(&ld_completed);
 
-	spin_lock(&chan->lock);
+	spin_lock_irqsave(&chan->lock, flags);
 
 	/* Clean already completed and acked descriptors */
 	xgene_dma_clean_completed_descriptor(chan);
@@ -762,7 +763,7 @@ static void xgene_dma_cleanup_descriptors(struct xgene_dma_chan *chan)
 	 */
 	xgene_chan_xfer_ld_pending(chan);
 
-	spin_unlock(&chan->lock);
+	spin_unlock_irqrestore(&chan->lock, flags);
 
 	/* Run the callback for each descriptor, in order */
 	list_for_each_entry_safe(desc_sw, _desc_sw, &ld_completed, node) {
-- 
2.17.1

Le 26/07/2023 à 13:16, Chengfeng Ye a écrit :
> As xgene_dma_cleanup_descriptors() is invoked by both tasklet
> xgene_dma_tasklet_cb() under softirq context and
> xgene_dma_free_chan_resources() callback that executed under process
> context, the lock aquicision of &chan->lock inside
> xgene_dma_cleanup_descriptors() should disable irq otherwise deadlock
> could happen if the tasklet softirq preempts the execution of process
> context code while the lock is held in process context on the same CPU.
> 
> Possible deadlock scenario:
> xgene_dma_free_chan_resources()
>      -> xgene_dma_cleanup_descriptors()
>      -> spin_lock(&chan->lock)
>          <tasklet softirq>
>          -> xgene_dma_tasklet_cb()
>          -> xgene_dma_cleanup_descriptors()
>          -> spin_lock(&chan->lock) (deadlock here)
> 
> This flaw was found by an experimental static analysis tool I am developing
> for irq-related deadlock.

Hi,

first of all, for what I've seen from your numerous recent patches, all 
this look real great ! :)
And your experimental tool looks really promising.


Even if I'm not always confident with my understanding of locking and 
related subtilities, I wonder if in the cases of <tasklet softirq>, like 
above, using spin_lock_bh() would be enough?
It should be less agressive than spin_lock_irqsave() but still handle 
the use case you have spotted.


Just my 2c.

CJ


> 
> The tentative patch fixes the potential deadlock by spin_lock_irqsave() in
> plx_dma_process_desc() to disable irq while lock is held.
> 
> Signed-off-by: Chengfeng Ye <dg573847474@gmail.com>
> ---
>   drivers/dma/xgene-dma.c | 5 +++--
>   1 file changed, 3 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/dma/xgene-dma.c b/drivers/dma/xgene-dma.c
> index 3589b4ef50b8..e766511badcf 100644
> --- a/drivers/dma/xgene-dma.c
> +++ b/drivers/dma/xgene-dma.c
> @@ -689,11 +689,12 @@ static void xgene_dma_cleanup_descriptors(struct xgene_dma_chan *chan)
>   	struct xgene_dma_desc_sw *desc_sw, *_desc_sw;
>   	struct xgene_dma_desc_hw *desc_hw;
>   	struct list_head ld_completed;
> +	unsigned long flags;
>   	u8 status;
>   
>   	INIT_LIST_HEAD(&ld_completed);
>   
> -	spin_lock(&chan->lock);
> +	spin_lock_irqsave(&chan->lock, flags);
>   
>   	/* Clean already completed and acked descriptors */
>   	xgene_dma_clean_completed_descriptor(chan);
> @@ -762,7 +763,7 @@ static void xgene_dma_cleanup_descriptors(struct xgene_dma_chan *chan)
>   	 */
>   	xgene_chan_xfer_ld_pending(chan);
>   
> -	spin_unlock(&chan->lock);
> +	spin_unlock_irqrestore(&chan->lock, flags);
>   
>   	/* Run the callback for each descriptor, in order */
>   	list_for_each_entry_safe(desc_sw, _desc_sw, &ld_completed, node) {