In the first list_for_each_entry() macro of dma_async_device_register,
it gets the chan from list and calls __dma_async_device_channel_register
(..,chan). We can see that chan->local is allocated by alloc_percpu() and
it is freed chan->local by free_percpu(chan->local) when
__dma_async_device_channel_register() failed.
But after __dma_async_device_channel_register() failed, the caller will
goto err_out and freed the chan->local in the second time by free_percpu().
The cause of this problem is forget to set chan->local to NULL when
chan->local was freed in __dma_async_device_channel_register(). My
patch sets chan->local to NULL when the callee failed to avoid double free.
Fixes: d2fb0a0438384 ("dmaengine: break out channel registration")
Signed-off-by: Lv Yunlong <[email protected]>
---
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 fe6a460c4373..fef64b198c95 100644
--- a/drivers/dma/dmaengine.c
+++ b/drivers/dma/dmaengine.c
@@ -1249,8 +1249,10 @@ int dma_async_device_register(struct dma_device *device)
/* represent channels in sysfs. Probably want devs too */
list_for_each_entry(chan, &device->channels, device_node) {
rc = __dma_async_device_channel_register(device, chan);
- if (rc < 0)
+ if (rc < 0) {
+ chan->local = NULL;
goto err_out;
+ }
}
mutex_lock(&dma_list_mutex);
--
2.25.1
On 3/30/2021 2:01 AM, Lv Yunlong wrote:
> In the first list_for_each_entry() macro of dma_async_device_register,
> it gets the chan from list and calls __dma_async_device_channel_register
> (..,chan). We can see that chan->local is allocated by alloc_percpu() and
> it is freed chan->local by free_percpu(chan->local) when
> __dma_async_device_channel_register() failed.
>
> But after __dma_async_device_channel_register() failed, the caller will
> goto err_out and freed the chan->local in the second time by free_percpu().
>
> The cause of this problem is forget to set chan->local to NULL when
> chan->local was freed in __dma_async_device_channel_register(). My
> patch sets chan->local to NULL when the callee failed to avoid double free.
Thanks for the fix. I think it would make sense to set it to NULL in
__dma_async_device_channel_register() cleanup path after it calls
free_percpu(chan->local) right? That would address any other instances
of this issue happening else where.
>
> Fixes: d2fb0a0438384 ("dmaengine: break out channel registration")
> Signed-off-by: Lv Yunlong <[email protected]>
> ---
> 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 fe6a460c4373..fef64b198c95 100644
> --- a/drivers/dma/dmaengine.c
> +++ b/drivers/dma/dmaengine.c
> @@ -1249,8 +1249,10 @@ int dma_async_device_register(struct dma_device *device)
> /* represent channels in sysfs. Probably want devs too */
> list_for_each_entry(chan, &device->channels, device_node) {
> rc = __dma_async_device_channel_register(device, chan);
> - if (rc < 0)
> + if (rc < 0) {
> + chan->local = NULL;
> goto err_out;
> + }
> }
>
> mutex_lock(&dma_list_mutex);
> -----原始邮件-----
> 发件人: "Dave Jiang" <[email protected]>
> 发送时间: 2021-03-31 00:05:15 (星期三)
> 收件人: "Lv Yunlong" <[email protected]>, [email protected]
> 抄送: [email protected], [email protected]
> 主题: Re: [PATCH] dma: Fix a double free in dma_async_device_register
>
>
> On 3/30/2021 2:01 AM, Lv Yunlong wrote:
> > In the first list_for_each_entry() macro of dma_async_device_register,
> > it gets the chan from list and calls __dma_async_device_channel_register
> > (..,chan). We can see that chan->local is allocated by alloc_percpu() and
> > it is freed chan->local by free_percpu(chan->local) when
> > __dma_async_device_channel_register() failed.
> >
> > But after __dma_async_device_channel_register() failed, the caller will
> > goto err_out and freed the chan->local in the second time by free_percpu().
> >
> > The cause of this problem is forget to set chan->local to NULL when
> > chan->local was freed in __dma_async_device_channel_register(). My
> > patch sets chan->local to NULL when the callee failed to avoid double free.
>
> Thanks for the fix. I think it would make sense to set it to NULL in
> __dma_async_device_channel_register() cleanup path after it calls
> free_percpu(chan->local) right? That would address any other instances
> of this issue happening else where.
>
>
> >
> > Fixes: d2fb0a0438384 ("dmaengine: break out channel registration")
> > Signed-off-by: Lv Yunlong <[email protected]>
> > ---
> > 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 fe6a460c4373..fef64b198c95 100644
> > --- a/drivers/dma/dmaengine.c
> > +++ b/drivers/dma/dmaengine.c
> > @@ -1249,8 +1249,10 @@ int dma_async_device_register(struct dma_device *device)
> > /* represent channels in sysfs. Probably want devs too */
> > list_for_each_entry(chan, &device->channels, device_node) {
> > rc = __dma_async_device_channel_register(device, chan);
> > - if (rc < 0)
> > + if (rc < 0) {
> > + chan->local = NULL;
> > goto err_out;
> > + }
> > }
> >
> > mutex_lock(&dma_list_mutex);
Ok, that is a good idea. I have submitted the PATCH v2.
Thanks.