usb: gadget: uvc: Fix use are free during STREAMOFF
Subsystems (2)
PATCH 1/2 usb: gadget: uvc: Fix use are free during STREAMOFF
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.
--- drivers/usb/gadget/function/uvc_video.c | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-)
drivers/usb/gadget/function/uvc_video.c +9 -3
@@ -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); }
Signature · 1 line
2.43.0.472.g3155946c3a-goog
PATCH 2/2 usb: gadget: uvc: Remove nested locking
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.
--- drivers/usb/gadget/function/uvc_video.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/usb/gadget/function/uvc_video.c +3 -1
@@ -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
Signature · 1 line
2.43.0.472.g3155946c3a-goog
Greg KH
maintainerRe: [PATCH 2/2] usb: gadget: uvc: Remove nested locking
Quoted 8 lines · Avichal Rakesh
When handling error status from
uvcg_video_usb_req_queue,uvc_video_completecurrently callsuvcg_queue_cancelwithvideo->req_lockheld.uvcg_queue_cancelinternally locksqueue->irqlock, which nestsqueue->irqlockinsidevideo->req_lock. This isn't a functional bug at the moment, but does open up possibilities for ABBA deadlocks in the future.
Quoted 7 lines
This patch fixes the accidental nesting by dropping
video->req_lockbefore callinguvcg_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
Greg KH
maintainerQuoted 18 lines · Avichal Rakesh
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
Quoted 4 lines
The root cause is that
uvcg_video_usb_req_queuefrees theuvc_requestifis_enabledis false and returns an error status.video_pumpalso frees the associated request ifuvcg_video_usb_req_queuereturns an e rror status, leading to double free and accessing garbage memory.
Odd line wrapping :(
Quoted 9 lines
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?
drivers/usb/gadget/function/uvc_video.c static int uvcg_video_usb_req_queue(struct uvc_video *video, · line 280
Quoted 12 lines
drivers/usb/gadget/function/uvc_video.c +1
-2@@ -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
Thank you for the review, Greg!
Sent out V2 with the comments addressed.
Quoted 19 lines · Greg KH
Quoted 18 lines · Avichal Rakesh
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
Quoted 6 lines
Quoted 4 lines
The root cause is that
uvcg_video_usb_req_queuefrees theuvc_requestifis_enabledis false and returns an error status.video_pumpalso frees the associated request ifuvcg_video_usb_req_queuereturns an e rror status, leading to double free and accessing garbage memory.Odd line wrapping :(
Sigh! Fixed. Apologies.
Quoted 6 lines
Quoted 6 lines
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.
Quoted 6 lines
Quoted 4 lines
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.
Quoted 10 lines
Quoted 10 lines
drivers/usb/gadget/function/uvc_video.c +0
-0@@ -277,8 +277,7 @@ static int uvcg_video_usb_req_queue(struct uvc_video *video,struct list_head *list =NULL ;
Quoted 6 lines
drivers/usb/gadget/function/uvc_video.c static int uvcg_video_usb_req_queue(struct uvc_video *video, · line 280
Quoted 4 lines
drivers/usb/gadget/function/uvc_video.c +1
-2if (!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.
PATCH v2 1/2 usb: gadget: uvc: Fix use are free during STREAMOFF
In reply to Avichal RakeshThere 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.
---
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(-)
drivers/usb/gadget/function/uvc_video.c +10 -4
@@ -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
PATCH v2 2/2 usb: gadget: uvc: Remove nested locking
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.
---
v1 -> v2: Add "Fixes" tag.
drivers/usb/gadget/function/uvc_video.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/usb/gadget/function/uvc_video.c +3 -1
@@ -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