There is a path that may lead to freed memory being referenced,
and causing kernel panics.
The kernel panic has the following stack trace:
Workqueue: uvcgadget uvcg_video_pump.c51fb85fece46625450f86adbf92c56c.cfi_jt
pstate: 60c00085 (nZCv daIf +PAN +UAO -TCO BTYPE=--)
pc : __list_del_entry_valid+0xc0/0xd4
lr : __list_del_entry_valid+0xc0/0xd4
Call trace:
__list_del_entry_valid+0xc0/0xd4
uvc_video_free_request+0x60/0x98
uvcg_video_pump+0x1cc/0x204
process_one_work+0x21c/0x4b8
worker_thread+0x29c/0x574
kthread+0x158/0x1b0
ret_from_fork+0x10/0x30
The root cause is that uvcg_video_usb_req_queue frees the uvc_request
if is_enabled is false and returns an error status. video_pump also
frees the associated request if uvcg_video_usb_req_queue returns an e
rror status, leading to double free and accessing garbage memory.
To fix the issue, this patch removes freeing logic from
uvcg_video_usb_req_queue, and lets the callers to the function handle
queueing errors as they see fit.
Signed-off-by: Avichal Rakesh <[email protected]>
---
drivers/usb/gadget/function/uvc_video.c | 12 +++++++++---
1 file changed, 9 insertions(+), 3 deletions(-)
diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c
index 98ba524c27f5..e5db1be14ca3 100644
--- a/drivers/usb/gadget/function/uvc_video.c
+++ b/drivers/usb/gadget/function/uvc_video.c
@@ -277,8 +277,7 @@ static int uvcg_video_usb_req_queue(struct uvc_video *video,
struct list_head *list = NULL;
if (!video->is_enabled) {
- uvc_video_free_request(req->context, video->ep);
- return -ENODEV;
+ return -EINVAL;
}
if (queue_to_ep) {
struct uvc_request *ureq = req->context;
@@ -464,8 +463,15 @@ uvc_video_complete(struct usb_ep *ep, struct usb_request *req)
* and this thread for isoc endpoints.
*/
ret = uvcg_video_usb_req_queue(video, to_queue, !is_bulk);
- if (ret < 0)
+ if (ret < 0) {
+ /*
+ * Endpoint error, but the stream is still enabled.
+ * Put request back in req_free for it to be cleaned
+ * up later.
+ */
uvcg_queue_cancel(queue, 0);
+ list_add_tail(&to_queue->list, &video->req_free);
+ }
} else {
uvc_video_free_request(ureq, ep);
}
--
2.43.0.472.g3155946c3a-goog
When handling error status from uvcg_video_usb_req_queue,
uvc_video_complete currently calls uvcg_queue_cancel with
video->req_lock held. uvcg_queue_cancel internally locks
queue->irqlock, which nests queue->irqlock inside
video->req_lock. This isn't a functional bug at the
moment, but does open up possibilities for ABBA
deadlocks in the future.
This patch fixes the accidental nesting by dropping
video->req_lock before calling uvcg_queue_cancel.
Signed-off-by: Avichal Rakesh <[email protected]>
---
drivers/usb/gadget/function/uvc_video.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c
index e5db1be14ca3..d15b9b4c4fcc 100644
--- a/drivers/usb/gadget/function/uvc_video.c
+++ b/drivers/usb/gadget/function/uvc_video.c
@@ -469,13 +469,15 @@ uvc_video_complete(struct usb_ep *ep, struct usb_request *req)
* Put request back in req_free for it to be cleaned
* up later.
*/
- uvcg_queue_cancel(queue, 0);
list_add_tail(&to_queue->list, &video->req_free);
}
} else {
uvc_video_free_request(ureq, ep);
+ ret = 0;
}
spin_unlock_irqrestore(&video->req_lock, flags);
+ if (ret < 0)
+ uvcg_queue_cancel(queue, 0);
}
static int
--
2.43.0.472.g3155946c3a-goog
On Fri, Dec 15, 2023 at 01:07:44PM -0800, Avichal Rakesh wrote:
> There is a path that may lead to freed memory being referenced,
> and causing kernel panics.
>
> The kernel panic has the following stack trace:
>
> Workqueue: uvcgadget uvcg_video_pump.c51fb85fece46625450f86adbf92c56c.cfi_jt
> pstate: 60c00085 (nZCv daIf +PAN +UAO -TCO BTYPE=--)
> pc : __list_del_entry_valid+0xc0/0xd4
> lr : __list_del_entry_valid+0xc0/0xd4
> Call trace:
> __list_del_entry_valid+0xc0/0xd4
> uvc_video_free_request+0x60/0x98
> uvcg_video_pump+0x1cc/0x204
> process_one_work+0x21c/0x4b8
> worker_thread+0x29c/0x574
> kthread+0x158/0x1b0
> ret_from_fork+0x10/0x30
>
> The root cause is that uvcg_video_usb_req_queue frees the uvc_request
> if is_enabled is false and returns an error status. video_pump also
> frees the associated request if uvcg_video_usb_req_queue returns an e
> rror status, leading to double free and accessing garbage memory.
Odd line wrapping :(
>
> To fix the issue, this patch removes freeing logic from
> uvcg_video_usb_req_queue, and lets the callers to the function handle
> queueing errors as they see fit.
>
> Signed-off-by: Avichal Rakesh <[email protected]>
> ---
> drivers/usb/gadget/function/uvc_video.c | 12 +++++++++---
> 1 file changed, 9 insertions(+), 3 deletions(-)
What commit id does this fix?
>
> diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c
> index 98ba524c27f5..e5db1be14ca3 100644
> --- a/drivers/usb/gadget/function/uvc_video.c
> +++ b/drivers/usb/gadget/function/uvc_video.c
> @@ -277,8 +277,7 @@ static int uvcg_video_usb_req_queue(struct uvc_video *video,
> struct list_head *list = NULL;
>
> if (!video->is_enabled) {
> - uvc_video_free_request(req->context, video->ep);
> - return -ENODEV;
> + return -EINVAL;
Isn't this a separate change? And does it actually matter?
thanks,
greg k-h
On Fri, Dec 15, 2023 at 01:07:45PM -0800, Avichal Rakesh wrote:
> When handling error status from uvcg_video_usb_req_queue,
> uvc_video_complete currently calls uvcg_queue_cancel with
> video->req_lock held. uvcg_queue_cancel internally locks
> queue->irqlock, which nests queue->irqlock inside
> video->req_lock. This isn't a functional bug at the
> moment, but does open up possibilities for ABBA
> deadlocks in the future.
>
> This patch fixes the accidental nesting by dropping
> video->req_lock before calling uvcg_queue_cancel.
>
> Signed-off-by: Avichal Rakesh <[email protected]>
> ---
> drivers/usb/gadget/function/uvc_video.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
Again, what commit does this fix?
thanks,
greg k-h
There is a path that may lead to freed memory being referenced,
causing kernel panics.
The kernel panic has the following stack trace:
Workqueue: uvcgadget uvcg_video_pump.c51fb85fece46625450f86adbf92c56c.cfi_jt
pstate: 60c00085 (nZCv daIf +PAN +UAO -TCO BTYPE=--)
pc : __list_del_entry_valid+0xc0/0xd4
lr : __list_del_entry_valid+0xc0/0xd4
Call trace:
__list_del_entry_valid+0xc0/0xd4
uvc_video_free_request+0x60/0x98
uvcg_video_pump+0x1cc/0x204
process_one_work+0x21c/0x4b8
worker_thread+0x29c/0x574
kthread+0x158/0x1b0
ret_from_fork+0x10/0x30
The root cause is that uvcg_video_usb_req_queue frees the uvc_request
if is_enabled is false and returns an error status. video_pump also
frees the associated request if uvcg_video_usb_req_queue returns an
error status, leading to double free and accessing garbage memory.
To fix the issue, this patch removes freeing logic from
uvcg_video_usb_req_queue, and lets the callers to the function handle
queueing errors as they see fit.
Fixes: 6acba0345b68 ("usb:gadget:uvc Do not use worker thread to pump isoc usb requests")
Tested-by: Avichal Rakesh <[email protected]>
Signed-off-by: Avichal Rakesh <[email protected]>
---
v1 -> v2: Address review comment; add "Fixes" tag; fix goofy line wrap
drivers/usb/gadget/function/uvc_video.c | 14 ++++++++++----
1 file changed, 10 insertions(+), 4 deletions(-)
diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c
index 98ba524c27f5..7f18dc471be3 100644
--- a/drivers/usb/gadget/function/uvc_video.c
+++ b/drivers/usb/gadget/function/uvc_video.c
@@ -276,10 +276,9 @@ static int uvcg_video_usb_req_queue(struct uvc_video *video,
bool is_bulk = video->max_payload_size;
struct list_head *list = NULL;
- if (!video->is_enabled) {
- uvc_video_free_request(req->context, video->ep);
+ if (!video->is_enabled)
return -ENODEV;
- }
+
if (queue_to_ep) {
struct uvc_request *ureq = req->context;
/*
@@ -464,8 +463,15 @@ uvc_video_complete(struct usb_ep *ep, struct usb_request *req)
* and this thread for isoc endpoints.
*/
ret = uvcg_video_usb_req_queue(video, to_queue, !is_bulk);
- if (ret < 0)
+ if (ret < 0) {
+ /*
+ * Endpoint error, but the stream is still enabled.
+ * Put request back in req_free for it to be cleaned
+ * up later.
+ */
uvcg_queue_cancel(queue, 0);
+ list_add_tail(&to_queue->list, &video->req_free);
+ }
} else {
uvc_video_free_request(ureq, ep);
}
--
2.43.0.472.g3155946c3a-goog
When handling error status from uvcg_video_usb_req_queue,
uvc_video_complete currently calls uvcg_queue_cancel with
video->req_lock held. uvcg_queue_cancel internally locks
queue->irqlock, which nests queue->irqlock inside
video->req_lock. This isn't a functional bug at the
moment, but does open up possibilities for ABBA
deadlocks in the future.
This patch fixes the accidental nesting by dropping
video->req_lock before calling uvcg_queue_cancel.
Fixes: 6acba0345b68 ("usb:gadget:uvc Do not use worker thread to pump isoc usb requests")
Signed-off-by: Avichal Rakesh <[email protected]>
---
v1 -> v2: Add "Fixes" tag.
drivers/usb/gadget/function/uvc_video.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c
index 7f18dc471be3..dd3241fc6939 100644
--- a/drivers/usb/gadget/function/uvc_video.c
+++ b/drivers/usb/gadget/function/uvc_video.c
@@ -469,13 +469,15 @@ uvc_video_complete(struct usb_ep *ep, struct usb_request *req)
* Put request back in req_free for it to be cleaned
* up later.
*/
- uvcg_queue_cancel(queue, 0);
list_add_tail(&to_queue->list, &video->req_free);
}
} else {
uvc_video_free_request(ureq, ep);
+ ret = 0;
}
spin_unlock_irqrestore(&video->req_lock, flags);
+ if (ret < 0)
+ uvcg_queue_cancel(queue, 0);
}
static int
--
2.43.0.472.g3155946c3a-goog
Thank you for the review, Greg!
Sent out V2 with the comments addressed.
On 1/4/24 06:19, Greg KH wrote:
> On Fri, Dec 15, 2023 at 01:07:44PM -0800, Avichal Rakesh wrote:
>> There is a path that may lead to freed memory being referenced,
>> and causing kernel panics.
>>
>> The kernel panic has the following stack trace:
>>
>> Workqueue: uvcgadget uvcg_video_pump.c51fb85fece46625450f86adbf92c56c.cfi_jt
>> pstate: 60c00085 (nZCv daIf +PAN +UAO -TCO BTYPE=--)
>> pc : __list_del_entry_valid+0xc0/0xd4
>> lr : __list_del_entry_valid+0xc0/0xd4
>> Call trace:
>> __list_del_entry_valid+0xc0/0xd4
>> uvc_video_free_request+0x60/0x98
>> uvcg_video_pump+0x1cc/0x204
>> process_one_work+0x21c/0x4b8
>> worker_thread+0x29c/0x574
>> kthread+0x158/0x1b0
>> ret_from_fork+0x10/0x30
>>
>> The root cause is that uvcg_video_usb_req_queue frees the uvc_request
>> if is_enabled is false and returns an error status. video_pump also
>> frees the associated request if uvcg_video_usb_req_queue returns an e
>> rror status, leading to double free and accessing garbage memory.
>
> Odd line wrapping :(
Sigh! Fixed. Apologies.
>
>>
>> To fix the issue, this patch removes freeing logic from
>> uvcg_video_usb_req_queue, and lets the callers to the function handle
>> queueing errors as they see fit.
>>
>> Signed-off-by: Avichal Rakesh <[email protected]>
>> ---
>> drivers/usb/gadget/function/uvc_video.c | 12 +++++++++---
>> 1 file changed, 9 insertions(+), 3 deletions(-)
>
> What commit id does this fix?
The commit that caused the bug is not in linus' branch yet. Added
the "Fixes" tag with the commit SHA from usb-next branch. Let me know
if I should be using some other SHA.
>
>
>>
>> diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c
>> index 98ba524c27f5..e5db1be14ca3 100644
>> --- a/drivers/usb/gadget/function/uvc_video.c
>> +++ b/drivers/usb/gadget/function/uvc_video.c
>> @@ -277,8 +277,7 @@ static int uvcg_video_usb_req_queue(struct uvc_video *video,
>> struct list_head *list = NULL;
>>
>> if (!video->is_enabled) {
>> - uvc_video_free_request(req->context, video->ep);
>> - return -ENODEV;
>> + return -EINVAL;
>
> Isn't this a separate change? And does it actually matter?
It was meant to differentiate between usb_ep_queue failing
and the video stream being disabled, but we don't really
care about that. Reverted.
Thank you!
- Avi.