[PATCH v3 0/3] dmaengine: fix kref underflow and UAF in dma_chan_put()

[PATCH v3 0/3] dmaengine: fix kref underflow and UAF in dma_chan_put()

[Shivank Garg]

Fix bugs related to dma_chan_put(), found while testing with SDXI[1].

[1]: https://lore.kernel.org/dmaengine/20260605-sdxi-base-v3-0-4d38ca2bdffe@amd.com

Signed-off-by: Shivank Garg <shivankg@amd.com>
---
Changes in v3:
- Add patch 3: add synchronize_rcu() to wait for RCU readers to prevent
  use-after-free. (Sashiko)
- Link to v2: https://lore.kernel.org/r/20260526-dmaengine-kref-fix-v2-0-3df60afac01d@amd.com

Changes in v2:
- Add patch 2 fixing the dma_chan_put()/dma_release_channel() use-after-free (sashiko) 
- Link to v1: https://lore.kernel.org/r/20260518-dmaengine-kref-fix-v1-1-4d6125048fb7@amd.com

---
Shivank Garg (3):
      dmaengine: Fix device kref underflow in dma_chan_put()
      dmaengine: fix use-after-free in dma_chan_put() and dma_release_channel()
      dmaengine: wait for RCU readers before releasing dma_device

 drivers/dma/dmaengine.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
---
base-commit: 0d995da5fb97e8c312834575604d4423eb6225b7
change-id: 20260518-dmaengine-kref-fix-7b21acb09455

Best regards,
-- 
Shivank Garg <shivankg@amd.com>

[PATCH v3 1/3] dmaengine: Fix device kref underflow in dma_chan_put()

[Shivank Garg]

dma_chan_get() takes chan->device->ref only on the slow path:

	/* no kref on fast path */
	if (chan->client_count) {
		__module_get(owner);
		chan->client_count++;
		return 0;
	}
	if (!try_module_get(owner))
		return -ENODEV;
	ret = kref_get_unless_zero(&chan->device->ref);

dma_chan_put() drops the ref unconditionally, so every fast-path
get/put pair drops one extra device reference.

The bug fires when two conditions hold together: a non-private
provider has a persistent client holding chan->client_count > 0
and another client cycles dmaengine_get()/dmaengine_put().
When the kref hits zero, the subsequent dma_find_channel() returns
NULL even though the provider module is still loaded.

Fix this by dropping device->ref only on the last put, matching the
single slow-path get.

Fixes: 8ad342a86359 ("dmaengine: Add reference counting to dma_device struct")
Signed-off-by: Shivank Garg <shivankg@amd.com>
---
 drivers/dma/dmaengine.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/dma/dmaengine.c b/drivers/dma/dmaengine.c
index 6ffd8bd82154..516d6d933208 100644
--- a/drivers/dma/dmaengine.c
+++ b/drivers/dma/dmaengine.c
@@ -515,7 +515,9 @@ static void dma_chan_put(struct dma_chan *chan)
 		chan->route_data = NULL;
 	}
 
-	dma_device_put(chan->device);
+	/* This channel is not in use anymore, drop the device ref */
+	if (!chan->client_count)
+		dma_device_put(chan->device);
 	module_put(dma_chan_to_owner(chan));
 }
 

-- 
2.43.0

[PATCH v3 2/3] dmaengine: fix use-after-free in dma_chan_put() and dma_release_channel()

[Shivank Garg]

When dma_device_put() drops the last reference on chan->device->ref,
dma_device_release() runs and may free the dma_device along with its
channels.

Two paths still read that memory after the put:
 - dma_chan_put() reads chan->device->owner via dma_chan_to_owner()
   for the trailing module_put().
 - dma_release_channel() calls dma_chan_put() before reading chan->slave,
   chan->name, chan->dev and chan->dbg_client_name.

KASAN catches the first one:

	slab-use-after-free in dma_chan_put+0x3e6/0x4c0
	Read of size 8 by task insmod/6319
	Freed by task 6319:
	  kfree+0x225/0x470
	  dma_chan_put+0x395/0x4c0
	  dmaengine_put+0xf8/0x160

Cache the module owner in dma_chan_put() before the put so the trailing
module_put() does not need chan->device. In dma_release_channel(), move
dma_chan_put() to the end, after every chan/device read.

Fixes: 8ad342a86359 ("dmaengine: Add reference counting to dma_device struct")
Suggested-by: Sashiko <sashiko-bot@kernel.org>
Link: https://sashiko.dev/#/patchset/20260518-dmaengine-kref-fix-v1-1-4d6125048fb7@amd.com
Signed-off-by: Shivank Garg <shivankg@amd.com>
---
 drivers/dma/dmaengine.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/drivers/dma/dmaengine.c b/drivers/dma/dmaengine.c
index 516d6d933208..bf491eb10596 100644
--- a/drivers/dma/dmaengine.c
+++ b/drivers/dma/dmaengine.c
@@ -495,10 +495,13 @@ static int dma_chan_get(struct dma_chan *chan)
  */
 static void dma_chan_put(struct dma_chan *chan)
 {
+	struct module *owner;
+
 	/* This channel is not in use, bail out */
 	if (!chan->client_count)
 		return;
 
+	owner = dma_chan_to_owner(chan);
 	chan->client_count--;
 
 	/* This channel is not in use anymore, free it */
@@ -518,7 +521,7 @@ static void dma_chan_put(struct dma_chan *chan)
 	/* This channel is not in use anymore, drop the device ref */
 	if (!chan->client_count)
 		dma_device_put(chan->device);
-	module_put(dma_chan_to_owner(chan));
+	module_put(owner);
 }
 
 enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
@@ -916,8 +919,6 @@ void dma_release_channel(struct dma_chan *chan)
 	if (--chan->device->privatecnt == 0)
 		dma_cap_clear(DMA_PRIVATE, chan->device->cap_mask);
 
-	dma_chan_put(chan);
-
 	if (chan->slave) {
 		sysfs_remove_link(&chan->dev->device.kobj, DMA_SLAVE_NAME);
 		sysfs_remove_link(&chan->slave->kobj, chan->name);
@@ -930,6 +931,7 @@ void dma_release_channel(struct dma_chan *chan)
 	kfree(chan->dbg_client_name);
 	chan->dbg_client_name = NULL;
 #endif
+	dma_chan_put(chan);
 	mutex_unlock(&dma_list_mutex);
 }
 EXPORT_SYMBOL_GPL(dma_release_channel);

-- 
2.43.0

[PATCH v3 3/3] dmaengine: wait for RCU readers before releasing dma_device

[Shivank Garg]

dma_issue_pending_all() walks the dma_device_list with
list_for_each_entry_rcu() under rcu_read_lock(). dma_device_release()
unlinks the device with list_del_rcu() and then calls
device->device_release() (which in many drivers, such as plx_dma.c,
directly calls kfree()).

Because there is no grace period between unlinking the device and
freeing it, concurrent RCU readers in dma_issue_pending_all() can
access the device after it has been freed.

The lockless walk originally relied on clients holding a dmaengine
reference to pin the provider module, and therefore the device, for as
long as they might traverse the list. Commit 8ad342a86359 ("dmaengine:
Add reference counting to dma_device struct") decoupled the dma_device
lifetime from the module reference, so the device can now be released
while a reader is still walking the list.

Add synchronize_rcu() before the device is freed, so RCU readers are
guaranteed to have finished. Keep it unconditional: providers that do
not implement device_release() free the device themselves once
dma_async_device_unregister() returns, so they need the same grace
period.

Fixes: 2ba05622b8b1 ("dmaengine: provide a common 'issue_pending_all' implementation")
Suggested-by: Sashiko <sashiko-bot@kernel.org>
Link: https://sashiko.dev/#/patchset/20260526-dmaengine-kref-fix-v2-0-3df60afac01d@amd.com
Signed-off-by: Shivank Garg <shivankg@amd.com>
---
 drivers/dma/dmaengine.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/drivers/dma/dmaengine.c b/drivers/dma/dmaengine.c
index bf491eb10596..29b128aa0bea 100644
--- a/drivers/dma/dmaengine.c
+++ b/drivers/dma/dmaengine.c
@@ -429,6 +429,12 @@ static void dma_device_release(struct kref *ref)
 	list_del_rcu(&device->global_node);
 	dma_channel_rebalance();
 
+	/*
+	 * Wait for RCU readers (e.g. dma_issue_pending_all()) that may still
+	 * be traversing dma_device_list before the device is freed.
+	 */
+	synchronize_rcu();
+
 	if (device->device_release)
 		device->device_release(device);
 }

-- 
2.43.0

Re: [PATCH v3 1/3] dmaengine: Fix device kref underflow in dma_chan_put()

[Frank Li]

On Sun, Aug 16, 2026 at 03:49:25PM +0000, Shivank Garg wrote:
> [You don't often get email from shivankg@amd.com. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
>
> dma_chan_get() takes chan->device->ref only on the slow path:
>
>         /* no kref on fast path */
>         if (chan->client_count) {
>                 __module_get(owner);
>                 chan->client_count++;
>                 return 0;
>         }
>         if (!try_module_get(owner))
>                 return -ENODEV;
>         ret = kref_get_unless_zero(&chan->device->ref);
>
> dma_chan_put() drops the ref unconditionally, so every fast-path
> get/put pair drops one extra device reference.
>
> The bug fires when two conditions hold together: a non-private
> provider has a persistent client holding chan->client_count > 0
> and another client cycles dmaengine_get()/dmaengine_put().
> When the kref hits zero, the subsequent dma_find_channel() returns
> NULL even though the provider module is still loaded.
>
> Fix this by dropping device->ref only on the last put, matching the
> single slow-path get.
>
> Fixes: 8ad342a86359 ("dmaengine: Add reference counting to dma_device struct")
> Signed-off-by: Shivank Garg <shivankg@amd.com>
> ---

Nit: add helper dma_device_get() to match dma_device_put() to make code
more symmetry.

Reviewed-by: Frank Li <Frank.Li@nxp.com>

>  drivers/dma/dmaengine.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/dma/dmaengine.c b/drivers/dma/dmaengine.c
> index 6ffd8bd82154..516d6d933208 100644
> --- a/drivers/dma/dmaengine.c
> +++ b/drivers/dma/dmaengine.c
> @@ -515,7 +515,9 @@ static void dma_chan_put(struct dma_chan *chan)
>                 chan->route_data = NULL;
>         }
>
> -       dma_device_put(chan->device);
> +       /* This channel is not in use anymore, drop the device ref */
> +       if (!chan->client_count)
> +               dma_device_put(chan->device);
>         module_put(dma_chan_to_owner(chan));
>  }
>
>
> --
> 2.43.0
>

Re: [PATCH v3 2/3] dmaengine: fix use-after-free in dma_chan_put() and dma_release_channel()

[Frank Li]

On Sun, Aug 16, 2026 at 03:49:26PM +0000, Shivank Garg wrote:
> [You don't often get email from shivankg@amd.com. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
>
> When dma_device_put() drops the last reference on chan->device->ref,
> dma_device_release() runs and may free the dma_device along with its
> channels.
>
> Two paths still read that memory after the put:
>  - dma_chan_put() reads chan->device->owner via dma_chan_to_owner()
>    for the trailing module_put().
>  - dma_release_channel() calls dma_chan_put() before reading chan->slave,
>    chan->name, chan->dev and chan->dbg_client_name.
>
> KASAN catches the first one:
>
>         slab-use-after-free in dma_chan_put+0x3e6/0x4c0
>         Read of size 8 by task insmod/6319
>         Freed by task 6319:
>           kfree+0x225/0x470
>           dma_chan_put+0x395/0x4c0
>           dmaengine_put+0xf8/0x160
>
> Cache the module owner in dma_chan_put() before the put so the trailing
> module_put() does not need chan->device. In dma_release_channel(), move
> dma_chan_put() to the end, after every chan/device read.
>
> Fixes: 8ad342a86359 ("dmaengine: Add reference counting to dma_device struct")
> Suggested-by: Sashiko <sashiko-bot@kernel.org>
> Link: https://sashiko.dev/#/patchset/20260518-dmaengine-kref-fix-v1-1-4d6125048fb7@amd.com
> Signed-off-by: Shivank Garg <shivankg@amd.com>
> ---


Reviewed-by: Frank Li <Frank.Li@nxp.com>


>  drivers/dma/dmaengine.c | 8 +++++---
>  1 file changed, 5 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/dma/dmaengine.c b/drivers/dma/dmaengine.c
> index 516d6d933208..bf491eb10596 100644
> --- a/drivers/dma/dmaengine.c
> +++ b/drivers/dma/dmaengine.c
> @@ -495,10 +495,13 @@ static int dma_chan_get(struct dma_chan *chan)
>   */
>  static void dma_chan_put(struct dma_chan *chan)
>  {
> +       struct module *owner;
> +
>         /* This channel is not in use, bail out */
>         if (!chan->client_count)
>                 return;
>
> +       owner = dma_chan_to_owner(chan);
>         chan->client_count--;
>
>         /* This channel is not in use anymore, free it */
> @@ -518,7 +521,7 @@ static void dma_chan_put(struct dma_chan *chan)
>         /* This channel is not in use anymore, drop the device ref */
>         if (!chan->client_count)
>                 dma_device_put(chan->device);
> -       module_put(dma_chan_to_owner(chan));
> +       module_put(owner);
>  }
>
>  enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
> @@ -916,8 +919,6 @@ void dma_release_channel(struct dma_chan *chan)
>         if (--chan->device->privatecnt == 0)
>                 dma_cap_clear(DMA_PRIVATE, chan->device->cap_mask);
>
> -       dma_chan_put(chan);
> -
>         if (chan->slave) {
>                 sysfs_remove_link(&chan->dev->device.kobj, DMA_SLAVE_NAME);
>                 sysfs_remove_link(&chan->slave->kobj, chan->name);
> @@ -930,6 +931,7 @@ void dma_release_channel(struct dma_chan *chan)
>         kfree(chan->dbg_client_name);
>         chan->dbg_client_name = NULL;
>  #endif
> +       dma_chan_put(chan);
>         mutex_unlock(&dma_list_mutex);
>  }
>  EXPORT_SYMBOL_GPL(dma_release_channel);
>
> --
> 2.43.0
>

Re: [PATCH v3 3/3] dmaengine: wait for RCU readers before releasing dma_device

[Frank Li]

On Sun, Aug 16, 2026 at 03:49:27PM +0000, Shivank Garg wrote:
> [You don't often get email from shivankg@amd.com. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
>
> dma_issue_pending_all() walks the dma_device_list with
> list_for_each_entry_rcu() under rcu_read_lock(). dma_device_release()
> unlinks the device with list_del_rcu() and then calls
> device->device_release() (which in many drivers, such as plx_dma.c,
> directly calls kfree()).
>
> Because there is no grace period between unlinking the device and
> freeing it, concurrent RCU readers in dma_issue_pending_all() can
> access the device after it has been freed.
>
> The lockless walk originally relied on clients holding a dmaengine
> reference to pin the provider module, and therefore the device, for as
> long as they might traverse the list. Commit 8ad342a86359 ("dmaengine:
> Add reference counting to dma_device struct") decoupled the dma_device
> lifetime from the module reference, so the device can now be released
> while a reader is still walking the list.
>
> Add synchronize_rcu() before the device is freed, so RCU readers are
> guaranteed to have finished. Keep it unconditional: providers that do
> not implement device_release() free the device themselves once
> dma_async_device_unregister() returns, so they need the same grace
> period.
>
> Fixes: 2ba05622b8b1 ("dmaengine: provide a common 'issue_pending_all' implementation")
> Suggested-by: Sashiko <sashiko-bot@kernel.org>
> Link: https://sashiko.dev/#/patchset/20260526-dmaengine-kref-fix-v2-0-3df60afac01d@amd.com
> Signed-off-by: Shivank Garg <shivankg@amd.com>
> ---

Reviewed-by: Frank Li <Frank.Li@nxp.com>

>  drivers/dma/dmaengine.c | 6 ++++++
>  1 file changed, 6 insertions(+)
>
> diff --git a/drivers/dma/dmaengine.c b/drivers/dma/dmaengine.c
> index bf491eb10596..29b128aa0bea 100644
> --- a/drivers/dma/dmaengine.c
> +++ b/drivers/dma/dmaengine.c
> @@ -429,6 +429,12 @@ static void dma_device_release(struct kref *ref)
>         list_del_rcu(&device->global_node);
>         dma_channel_rebalance();
>
> +       /*
> +        * Wait for RCU readers (e.g. dma_issue_pending_all()) that may still
> +        * be traversing dma_device_list before the device is freed.
> +        */
> +       synchronize_rcu();
> +
>         if (device->device_release)
>                 device->device_release(device);
>  }
>
> --
> 2.43.0
>