| Message ID | 20250601224231.24317-3-qasdev00@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | dmaengine: qcom_hidma: fix memory leak issues | expand |
On 6/2/25 01:42, Qasim Ijaz wrote: > hidma_ll_init() allocates a handoff FIFO, but the matching > hidma_ll_uninit() function (which is invoked in remove()) > never releases it, leaking memory. > > To fix this call kfifo_free in hidma_ll_uninit(). > > Fixes: d1615ca2e085 ("dmaengine: qcom_hidma: implement lower level hardware interface") > Cc: stable@vger.kernel.org > Signed-off-by: Qasim Ijaz <qasdev00@gmail.com> > > --- > drivers/dma/qcom/hidma_ll.c | 1 + > 1 file changed, 1 insertion(+) > > diff --git a/drivers/dma/qcom/hidma_ll.c b/drivers/dma/qcom/hidma_ll.c > index fee448499777..0c2bae46746c 100644 > --- a/drivers/dma/qcom/hidma_ll.c > +++ b/drivers/dma/qcom/hidma_ll.c > @@ -816,6 +816,7 @@ int hidma_ll_uninit(struct hidma_lldev *lldev) > > required_bytes = sizeof(struct hidma_tre) * lldev->nr_tres; > tasklet_kill(&lldev->task); > + kfifo_free(&lldev->handoff_fifo); > memset(lldev->trepool, 0, required_bytes); > lldev->trepool = NULL; > atomic_set(&lldev->pending_tre_count, 0); Is it possible that the handoff_fifo is freed, then we could observe reset complete interrupts before they are being cleared in hidma_ll_uninit later on, which would lead to the following call chain hidma_ll_inthandler - hidma_ll_int_handler_internal - hidma_handle_tre_completion - hidma_post_completed - tasklet_schedule(&lldev->task); - hidma_ll_tre_complete - kfifo_out ?
On 6/5/2025 9:04 AM, Eugen Hristev wrote: >> diff --git a/drivers/dma/qcom/hidma_ll.c b/drivers/dma/qcom/hidma_ll.c >> index fee448499777..0c2bae46746c 100644 >> --- a/drivers/dma/qcom/hidma_ll.c >> +++ b/drivers/dma/qcom/hidma_ll.c >> @@ -816,6 +816,7 @@ int hidma_ll_uninit(struct hidma_lldev *lldev) >> >> required_bytes = sizeof(struct hidma_tre) * lldev->nr_tres; >> tasklet_kill(&lldev->task); >> + kfifo_free(&lldev->handoff_fifo); >> memset(lldev->trepool, 0, required_bytes); >> lldev->trepool = NULL; >> atomic_set(&lldev->pending_tre_count, 0); > Is it possible that the handoff_fifo is freed, then we could observe > reset complete interrupts before they are being cleared in > hidma_ll_uninit later on, which would lead to the following call chain > > hidma_ll_inthandler - hidma_ll_int_handler_internal - > hidma_handle_tre_completion - hidma_post_completed - > tasklet_schedule(&lldev->task); - hidma_ll_tre_complete - kfifo_out According to the documentation, the way to guarantee this from not happening is to call tasklet_disable() to ensure that tasklet completes execution. Only after that data structures used by the tasklet can be freed. I think proper order is: 1. tasklet_disable 2. tasklet_kill 3. kfifo_free
On Fri, Jun 06, 2025 at 09:35:44AM -0400, Sinan Kaya wrote: > > On 6/5/2025 9:04 AM, Eugen Hristev wrote: > > > diff --git a/drivers/dma/qcom/hidma_ll.c b/drivers/dma/qcom/hidma_ll.c > > > index fee448499777..0c2bae46746c 100644 > > > --- a/drivers/dma/qcom/hidma_ll.c > > > +++ b/drivers/dma/qcom/hidma_ll.c > > > @@ -816,6 +816,7 @@ int hidma_ll_uninit(struct hidma_lldev *lldev) > > > required_bytes = sizeof(struct hidma_tre) * lldev->nr_tres; > > > tasklet_kill(&lldev->task); > > > + kfifo_free(&lldev->handoff_fifo); > > > memset(lldev->trepool, 0, required_bytes); > > > lldev->trepool = NULL; > > > atomic_set(&lldev->pending_tre_count, 0); > > Is it possible that the handoff_fifo is freed, then we could observe > > reset complete interrupts before they are being cleared in > > hidma_ll_uninit later on, which would lead to the following call chain > > > > hidma_ll_inthandler - hidma_ll_int_handler_internal - > > hidma_handle_tre_completion - hidma_post_completed - > > tasklet_schedule(&lldev->task); - hidma_ll_tre_complete - kfifo_out > > According to the documentation, the way to guarantee this from not happening > > is to call tasklet_disable() to ensure that tasklet completes execution. > Only after that > > data structures used by the tasklet can be freed. > > I think proper order is: > > 1. tasklet_disable > > 2. tasklet_kill > > 3. kfifo_free Hi Sinan, hi Eugen, Thanks for reviewing the patch and for pointing out the correct shutdown ordering. If you’re both happy with it, I’ll send a v2 that calls tasklet_disable() before tasklet_kill(), then frees the handoff_fifo. Just let me know and I’ll resend. Thanks Qasim > > >
On 6/15/2025 3:15 PM, Qasim Ijaz wrote: > Thanks for reviewing the patch and for pointing out the correct > shutdown ordering. > > If you’re both happy with it, I’ll send a v2 that calls > tasklet_disable() before tasklet_kill(), then frees the handoff_fifo. > > Just let me know and I’ll resend. Sure, please test it on a test kernel module before posting with a tasklet. This should be straightforward to verify.
diff --git a/drivers/dma/qcom/hidma_ll.c b/drivers/dma/qcom/hidma_ll.c index fee448499777..0c2bae46746c 100644 --- a/drivers/dma/qcom/hidma_ll.c +++ b/drivers/dma/qcom/hidma_ll.c @@ -816,6 +816,7 @@ int hidma_ll_uninit(struct hidma_lldev *lldev) required_bytes = sizeof(struct hidma_tre) * lldev->nr_tres; tasklet_kill(&lldev->task); + kfifo_free(&lldev->handoff_fifo); memset(lldev->trepool, 0, required_bytes); lldev->trepool = NULL; atomic_set(&lldev->pending_tre_count, 0);
hidma_ll_init() allocates a handoff FIFO, but the matching hidma_ll_uninit() function (which is invoked in remove()) never releases it, leaking memory. To fix this call kfifo_free in hidma_ll_uninit(). Fixes: d1615ca2e085 ("dmaengine: qcom_hidma: implement lower level hardware interface") Cc: stable@vger.kernel.org Signed-off-by: Qasim Ijaz <qasdev00@gmail.com> --- drivers/dma/qcom/hidma_ll.c | 1 + 1 file changed, 1 insertion(+)