diff mbox series

dmaengine: sf-pdma: Add multithread support for a DMA channel

Message ID 20220512091327.349563-1-v.v.mitrofanov@yadro.com (mailing list archive)
State Superseded
Headers show
Series dmaengine: sf-pdma: Add multithread support for a DMA channel | expand

Commit Message

Viacheslav Mitrofanov May 12, 2022, 9:13 a.m. UTC
When we get a DMA channel and try to use it in multiple threads it
will cause oops and hanging the system.

% echo 64 > /sys/module/dmatest/parameters/threads_per_chan
% echo 10000 > /sys/module/dmatest/parameters/iterations
% echo 1 > /sys/module/dmatest/parameters/run
[   89.480664] Unable to handle kernel NULL pointer dereference at virtual
               address 00000000000000a0
[   89.488725] Oops [#1]
[   89.494708] CPU: 2 PID: 1008 Comm: dma0chan0-copy0 Not tainted
               5.17.0-rc5
[   89.509385] epc : vchan_find_desc+0x32/0x46
[   89.513553]  ra : sf_pdma_tx_status+0xca/0xd6

This happens because of data race. Each thread rewrite channels's
descriptor as soon as device_prep_dma_memcpy() is called. It leads to the
situation when the driver thinks that it uses right descriptor that
actually is freed or substituted for other one.

With current fixes a descriptor changes it's value only when it has
been used. A new descriptor is acquired from vc->desc_issued queue that
is already filled with descriptors that are ready to be sent. Threads
have no direct access to DMA channel descriptor, now it is just possible
to queue a descriptor for further processing.

Fixes: 6973886ad58e ("dmaengine: sf-pdma: add platform DMA support for HiFive Unleashed A00")
Signed-off-by: Viacheslav Mitrofanov <v.v.mitrofanov@yadro.com>
---
 drivers/dma/sf-pdma/sf-pdma.c | 44 ++++++++++++++++++++++++-----------
 1 file changed, 30 insertions(+), 14 deletions(-)

Comments

Dave Jiang May 12, 2022, 5:01 p.m. UTC | #1
On 5/12/2022 2:13 AM, Viacheslav Mitrofanov wrote:
> When we get a DMA channel and try to use it in multiple threads it
> will cause oops and hanging the system.
>
> % echo 64 > /sys/module/dmatest/parameters/threads_per_chan
> % echo 10000 > /sys/module/dmatest/parameters/iterations
> % echo 1 > /sys/module/dmatest/parameters/run
> [   89.480664] Unable to handle kernel NULL pointer dereference at virtual
>                 address 00000000000000a0
> [   89.488725] Oops [#1]
> [   89.494708] CPU: 2 PID: 1008 Comm: dma0chan0-copy0 Not tainted
>                 5.17.0-rc5
> [   89.509385] epc : vchan_find_desc+0x32/0x46
> [   89.513553]  ra : sf_pdma_tx_status+0xca/0xd6
>
> This happens because of data race. Each thread rewrite channels's
> descriptor as soon as device_prep_dma_memcpy() is called. It leads to the
> situation when the driver thinks that it uses right descriptor that
> actually is freed or substituted for other one.
>
> With current fixes a descriptor changes it's value only when it has

With the current fix a descriptor ...

s/it's/its/


> been used. A new descriptor is acquired from vc->desc_issued queue that
> is already filled with descriptors that are ready to be sent. Threads
> have no direct access to DMA channel descriptor, now it is just possible

I suggest a '.' here instead of ','

DJ

> to queue a descriptor for further processing.
>
> Fixes: 6973886ad58e ("dmaengine: sf-pdma: add platform DMA support for HiFive Unleashed A00")
> Signed-off-by: Viacheslav Mitrofanov <v.v.mitrofanov@yadro.com>
> ---
>   drivers/dma/sf-pdma/sf-pdma.c | 44 ++++++++++++++++++++++++-----------
>   1 file changed, 30 insertions(+), 14 deletions(-)
>
> diff --git a/drivers/dma/sf-pdma/sf-pdma.c b/drivers/dma/sf-pdma/sf-pdma.c
> index f12606aeff87..70bb032c59c2 100644
> --- a/drivers/dma/sf-pdma/sf-pdma.c
> +++ b/drivers/dma/sf-pdma/sf-pdma.c
> @@ -52,16 +52,6 @@ static inline struct sf_pdma_desc *to_sf_pdma_desc(struct virt_dma_desc *vd)
>   static struct sf_pdma_desc *sf_pdma_alloc_desc(struct sf_pdma_chan *chan)
>   {
>   	struct sf_pdma_desc *desc;
> -	unsigned long flags;
> -
> -	spin_lock_irqsave(&chan->lock, flags);
> -
> -	if (chan->desc && !chan->desc->in_use) {
> -		spin_unlock_irqrestore(&chan->lock, flags);
> -		return chan->desc;
> -	}
> -
> -	spin_unlock_irqrestore(&chan->lock, flags);
>   
>   	desc = kzalloc(sizeof(*desc), GFP_NOWAIT);
>   	if (!desc)
> @@ -111,7 +101,6 @@ sf_pdma_prep_dma_memcpy(struct dma_chan *dchan,	dma_addr_t dest, dma_addr_t src,
>   	desc->async_tx = vchan_tx_prep(&chan->vchan, &desc->vdesc, flags);
>   
>   	spin_lock_irqsave(&chan->vchan.lock, iflags);
> -	chan->desc = desc;
>   	sf_pdma_fill_desc(desc, dest, src, len);
>   	spin_unlock_irqrestore(&chan->vchan.lock, iflags);
>   
> @@ -170,11 +159,17 @@ static size_t sf_pdma_desc_residue(struct sf_pdma_chan *chan,
>   	unsigned long flags;
>   	u64 residue = 0;
>   	struct sf_pdma_desc *desc;
> -	struct dma_async_tx_descriptor *tx;
> +	struct dma_async_tx_descriptor *tx = NULL;
>   
>   	spin_lock_irqsave(&chan->vchan.lock, flags);
>   
> -	tx = &chan->desc->vdesc.tx;
> +	list_for_each_entry(vd, &chan->vchan.desc_submitted, node)
> +		if (vd->tx.cookie == cookie)
> +			tx = &vd->tx;
> +
> +	if (!tx)
> +		goto out;
> +
>   	if (cookie == tx->chan->completed_cookie)
>   		goto out;
>   
> @@ -241,6 +236,19 @@ static void sf_pdma_enable_request(struct sf_pdma_chan *chan)
>   	writel(v, regs->ctrl);
>   }
>   
> +static struct sf_pdma_desc *sf_pdma_get_first_pending_desc(struct sf_pdma_chan *chan)
> +{
> +	struct virt_dma_chan *vchan = &chan->vchan;
> +	struct virt_dma_desc *vdesc;
> +
> +	if (list_empty(&vchan->desc_issued))
> +		return NULL;
> +
> +	vdesc = list_first_entry(&vchan->desc_issued, struct virt_dma_desc, node);
> +
> +	return container_of(vdesc, struct sf_pdma_desc, vdesc);
> +}
> +
>   static void sf_pdma_xfer_desc(struct sf_pdma_chan *chan)
>   {
>   	struct sf_pdma_desc *desc = chan->desc;
> @@ -268,8 +276,11 @@ static void sf_pdma_issue_pending(struct dma_chan *dchan)
>   
>   	spin_lock_irqsave(&chan->vchan.lock, flags);
>   
> -	if (vchan_issue_pending(&chan->vchan) && chan->desc)
> +	if ((chan->desc == NULL) && vchan_issue_pending(&chan->vchan)) {
> +		/* vchan_issue_pending has made a check that desc in not NULL */
> +		chan->desc = sf_pdma_get_first_pending_desc(chan);
>   		sf_pdma_xfer_desc(chan);
> +	}
>   
>   	spin_unlock_irqrestore(&chan->vchan.lock, flags);
>   }
> @@ -298,6 +309,11 @@ static void sf_pdma_donebh_tasklet(struct tasklet_struct *t)
>   	spin_lock_irqsave(&chan->vchan.lock, flags);
>   	list_del(&chan->desc->vdesc.node);
>   	vchan_cookie_complete(&chan->desc->vdesc);
> +
> +	chan->desc = sf_pdma_get_first_pending_desc(chan);
> +	if (chan->desc)
> +		sf_pdma_xfer_desc(chan);
> +
>   	spin_unlock_irqrestore(&chan->vchan.lock, flags);
>   }
>
diff mbox series

Patch

diff --git a/drivers/dma/sf-pdma/sf-pdma.c b/drivers/dma/sf-pdma/sf-pdma.c
index f12606aeff87..70bb032c59c2 100644
--- a/drivers/dma/sf-pdma/sf-pdma.c
+++ b/drivers/dma/sf-pdma/sf-pdma.c
@@ -52,16 +52,6 @@  static inline struct sf_pdma_desc *to_sf_pdma_desc(struct virt_dma_desc *vd)
 static struct sf_pdma_desc *sf_pdma_alloc_desc(struct sf_pdma_chan *chan)
 {
 	struct sf_pdma_desc *desc;
-	unsigned long flags;
-
-	spin_lock_irqsave(&chan->lock, flags);
-
-	if (chan->desc && !chan->desc->in_use) {
-		spin_unlock_irqrestore(&chan->lock, flags);
-		return chan->desc;
-	}
-
-	spin_unlock_irqrestore(&chan->lock, flags);
 
 	desc = kzalloc(sizeof(*desc), GFP_NOWAIT);
 	if (!desc)
@@ -111,7 +101,6 @@  sf_pdma_prep_dma_memcpy(struct dma_chan *dchan,	dma_addr_t dest, dma_addr_t src,
 	desc->async_tx = vchan_tx_prep(&chan->vchan, &desc->vdesc, flags);
 
 	spin_lock_irqsave(&chan->vchan.lock, iflags);
-	chan->desc = desc;
 	sf_pdma_fill_desc(desc, dest, src, len);
 	spin_unlock_irqrestore(&chan->vchan.lock, iflags);
 
@@ -170,11 +159,17 @@  static size_t sf_pdma_desc_residue(struct sf_pdma_chan *chan,
 	unsigned long flags;
 	u64 residue = 0;
 	struct sf_pdma_desc *desc;
-	struct dma_async_tx_descriptor *tx;
+	struct dma_async_tx_descriptor *tx = NULL;
 
 	spin_lock_irqsave(&chan->vchan.lock, flags);
 
-	tx = &chan->desc->vdesc.tx;
+	list_for_each_entry(vd, &chan->vchan.desc_submitted, node)
+		if (vd->tx.cookie == cookie)
+			tx = &vd->tx;
+
+	if (!tx)
+		goto out;
+
 	if (cookie == tx->chan->completed_cookie)
 		goto out;
 
@@ -241,6 +236,19 @@  static void sf_pdma_enable_request(struct sf_pdma_chan *chan)
 	writel(v, regs->ctrl);
 }
 
+static struct sf_pdma_desc *sf_pdma_get_first_pending_desc(struct sf_pdma_chan *chan)
+{
+	struct virt_dma_chan *vchan = &chan->vchan;
+	struct virt_dma_desc *vdesc;
+
+	if (list_empty(&vchan->desc_issued))
+		return NULL;
+
+	vdesc = list_first_entry(&vchan->desc_issued, struct virt_dma_desc, node);
+
+	return container_of(vdesc, struct sf_pdma_desc, vdesc);
+}
+
 static void sf_pdma_xfer_desc(struct sf_pdma_chan *chan)
 {
 	struct sf_pdma_desc *desc = chan->desc;
@@ -268,8 +276,11 @@  static void sf_pdma_issue_pending(struct dma_chan *dchan)
 
 	spin_lock_irqsave(&chan->vchan.lock, flags);
 
-	if (vchan_issue_pending(&chan->vchan) && chan->desc)
+	if ((chan->desc == NULL) && vchan_issue_pending(&chan->vchan)) {
+		/* vchan_issue_pending has made a check that desc in not NULL */
+		chan->desc = sf_pdma_get_first_pending_desc(chan);
 		sf_pdma_xfer_desc(chan);
+	}
 
 	spin_unlock_irqrestore(&chan->vchan.lock, flags);
 }
@@ -298,6 +309,11 @@  static void sf_pdma_donebh_tasklet(struct tasklet_struct *t)
 	spin_lock_irqsave(&chan->vchan.lock, flags);
 	list_del(&chan->desc->vdesc.node);
 	vchan_cookie_complete(&chan->desc->vdesc);
+
+	chan->desc = sf_pdma_get_first_pending_desc(chan);
+	if (chan->desc)
+		sf_pdma_xfer_desc(chan);
+
 	spin_unlock_irqrestore(&chan->vchan.lock, flags);
 }