This makes us sure that all requests are completed before we unbind
gadget. There are assumptions in gadget API that all requests have to
be completed and leak of complete can break some usb function drivers.
For example unbind of ECM function can cause NULL pointer dereference:
[ 26.396595] configfs-gadget gadget: unbind function
'cdc_ethernet'/e79c4c00
[ 26.414999] Unable to handle kernel NULL pointer dereference at
virtual address 00000000
(...)
[ 26.452223] PC is at ecm_unbind+0x6c/0x9c
[ 26.456209] LR is at ecm_unbind+0x68/0x9c
(...)
[ 26.603696] [<c033fdb4>] (ecm_unbind) from [<c033661c>]
(purge_configs_funcs+0x94/0xd8)
[ 26.611674] [<c033661c>] (purge_configs_funcs) from [<c0336674>]
(configfs_composite_unbind+0x14/0x34)
[ 26.620961] [<c0336674>] (configfs_composite_unbind) from
[<c0337124>] (usb_gadget_remove_driver+0x68/0x9c)
[ 26.630683] [<c0337124>] (usb_gadget_remove_driver) from [<c03376c8>]
(usb_gadget_unregister_driver+0x64/0x94)
[ 26.640664] [<c03376c8>] (usb_gadget_unregister_driver) from
[<c0336be8>] (unregister_gadget+0x20/0x3c)
[ 26.650038] [<c0336be8>] (unregister_gadget) from [<c0336c84>]
(gadget_dev_desc_UDC_store+0x80/0xb8)
[ 26.659152] [<c0336c84>] (gadget_dev_desc_UDC_store) from
[<c0335120>] (gadget_info_attr_store+0x1c/0x28)
[ 26.668703] [<c0335120>] (gadget_info_attr_store) from [<c012135c>]
(configfs_write_file+0xe8/0x148)
[ 26.677818] [<c012135c>] (configfs_write_file) from [<c00c8dd4>]
(vfs_write+0xb0/0x1a0)
[ 26.685801] [<c00c8dd4>] (vfs_write) from [<c00c91b8>]
(SyS_write+0x44/0x84)
[ 26.692834] [<c00c91b8>] (SyS_write) from [<c000e560>]
(ret_fast_syscall+0x0/0x30)
[ 26.700381] Code: e30409f8 e34c0069 eb07b88d e59430a8 (e5930000)
[ 26.706485] ---[ end trace f62a082b323838a2 ]---
It's because in some cases request is still running on endpoint during
unbind and kill_all_requests() called from s3c_hsotg_udc_stop() function
doesn't cause call of complete() of request. Missing complete() call
causes ecm->notify_req equals NULL in ecm_unbind() function, and this
is reason of this bug.
Similar breaks can be observed in another usb function drivers.
This patch fixes this bug forcing usb request completion in when
s3c_hsotg_ep_disable() is called from s3c_hsotg_udc_stop().
Signed-off-by: Robert Baldyga <[email protected]>
---
drivers/usb/dwc2/gadget.c | 10 +++++++---
1 file changed, 7 insertions(+), 3 deletions(-)
diff --git a/drivers/usb/dwc2/gadget.c b/drivers/usb/dwc2/gadget.c
index 8b5c079..cb4c925 100644
--- a/drivers/usb/dwc2/gadget.c
+++ b/drivers/usb/dwc2/gadget.c
@@ -2590,7 +2590,7 @@ error:
* s3c_hsotg_ep_disable - disable given endpoint
* @ep: The endpoint to disable.
*/
-static int s3c_hsotg_ep_disable(struct usb_ep *ep)
+static int s3c_hsotg_ep_disable_force(struct usb_ep *ep, bool force)
{
struct s3c_hsotg_ep *hs_ep = our_ep(ep);
struct s3c_hsotg *hsotg = hs_ep->parent;
@@ -2611,7 +2611,7 @@ static int s3c_hsotg_ep_disable(struct usb_ep *ep)
spin_lock_irqsave(&hsotg->lock, flags);
/* terminate all requests with shutdown */
- kill_all_requests(hsotg, hs_ep, -ESHUTDOWN, false);
+ kill_all_requests(hsotg, hs_ep, -ESHUTDOWN, force);
hsotg->fifo_map &= ~(1<<hs_ep->fifo_index);
hs_ep->fifo_index = 0;
@@ -2632,6 +2632,10 @@ static int s3c_hsotg_ep_disable(struct usb_ep *ep)
return 0;
}
+static int s3c_hsotg_ep_disable(struct usb_ep *ep)
+{
+ return s3c_hsotg_ep_disable_force(ep, false);
+}
/**
* on_list - check request is on the given endpoint
* @ep: The endpoint to check.
@@ -2933,7 +2937,7 @@ static int s3c_hsotg_udc_stop(struct usb_gadget *gadget,
/* all endpoints should be shutdown */
for (ep = 1; ep < hsotg->num_of_eps; ep++)
- s3c_hsotg_ep_disable(&hsotg->eps[ep].ep);
+ s3c_hsotg_ep_disable_force(&hsotg->eps[ep].ep, true);
spin_lock_irqsave(&hsotg->lock, flags);
--
1.9.1
> From: Robert Baldyga [mailto:[email protected]]
> Sent: Tuesday, December 09, 2014 5:42 AM
>
> This makes us sure that all requests are completed before we unbind
> gadget. There are assumptions in gadget API that all requests have to
> be completed and leak of complete can break some usb function drivers.
>
> For example unbind of ECM function can cause NULL pointer dereference:
>
> [ 26.396595] configfs-gadget gadget: unbind function
> 'cdc_ethernet'/e79c4c00
> [ 26.414999] Unable to handle kernel NULL pointer dereference at
> virtual address 00000000
> (...)
> [ 26.452223] PC is at ecm_unbind+0x6c/0x9c
> [ 26.456209] LR is at ecm_unbind+0x68/0x9c
> (...)
> [ 26.603696] [<c033fdb4>] (ecm_unbind) from [<c033661c>]
> (purge_configs_funcs+0x94/0xd8)
> [ 26.611674] [<c033661c>] (purge_configs_funcs) from [<c0336674>]
> (configfs_composite_unbind+0x14/0x34)
> [ 26.620961] [<c0336674>] (configfs_composite_unbind) from
> [<c0337124>] (usb_gadget_remove_driver+0x68/0x9c)
> [ 26.630683] [<c0337124>] (usb_gadget_remove_driver) from [<c03376c8>]
> (usb_gadget_unregister_driver+0x64/0x94)
> [ 26.640664] [<c03376c8>] (usb_gadget_unregister_driver) from
> [<c0336be8>] (unregister_gadget+0x20/0x3c)
> [ 26.650038] [<c0336be8>] (unregister_gadget) from [<c0336c84>]
> (gadget_dev_desc_UDC_store+0x80/0xb8)
> [ 26.659152] [<c0336c84>] (gadget_dev_desc_UDC_store) from
> [<c0335120>] (gadget_info_attr_store+0x1c/0x28)
> [ 26.668703] [<c0335120>] (gadget_info_attr_store) from [<c012135c>]
> (configfs_write_file+0xe8/0x148)
> [ 26.677818] [<c012135c>] (configfs_write_file) from [<c00c8dd4>]
> (vfs_write+0xb0/0x1a0)
> [ 26.685801] [<c00c8dd4>] (vfs_write) from [<c00c91b8>]
> (SyS_write+0x44/0x84)
> [ 26.692834] [<c00c91b8>] (SyS_write) from [<c000e560>]
> (ret_fast_syscall+0x0/0x30)
> [ 26.700381] Code: e30409f8 e34c0069 eb07b88d e59430a8 (e5930000)
> [ 26.706485] ---[ end trace f62a082b323838a2 ]---
>
> It's because in some cases request is still running on endpoint during
> unbind and kill_all_requests() called from s3c_hsotg_udc_stop() function
> doesn't cause call of complete() of request. Missing complete() call
> causes ecm->notify_req equals NULL in ecm_unbind() function, and this
> is reason of this bug.
>
> Similar breaks can be observed in another usb function drivers.
>
> This patch fixes this bug forcing usb request completion in when
> s3c_hsotg_ep_disable() is called from s3c_hsotg_udc_stop().
>
> Signed-off-by: Robert Baldyga <[email protected]>
> ---
> drivers/usb/dwc2/gadget.c | 10 +++++++---
> 1 file changed, 7 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/usb/dwc2/gadget.c b/drivers/usb/dwc2/gadget.c
> index 8b5c079..cb4c925 100644
> --- a/drivers/usb/dwc2/gadget.c
> +++ b/drivers/usb/dwc2/gadget.c
> @@ -2590,7 +2590,7 @@ error:
> * s3c_hsotg_ep_disable - disable given endpoint
> * @ep: The endpoint to disable.
> */
> -static int s3c_hsotg_ep_disable(struct usb_ep *ep)
> +static int s3c_hsotg_ep_disable_force(struct usb_ep *ep, bool force)
> {
> struct s3c_hsotg_ep *hs_ep = our_ep(ep);
> struct s3c_hsotg *hsotg = hs_ep->parent;
> @@ -2611,7 +2611,7 @@ static int s3c_hsotg_ep_disable(struct usb_ep *ep)
>
> spin_lock_irqsave(&hsotg->lock, flags);
> /* terminate all requests with shutdown */
> - kill_all_requests(hsotg, hs_ep, -ESHUTDOWN, false);
> + kill_all_requests(hsotg, hs_ep, -ESHUTDOWN, force);
>
> hsotg->fifo_map &= ~(1<<hs_ep->fifo_index);
> hs_ep->fifo_index = 0;
> @@ -2632,6 +2632,10 @@ static int s3c_hsotg_ep_disable(struct usb_ep *ep)
> return 0;
> }
>
> +static int s3c_hsotg_ep_disable(struct usb_ep *ep)
> +{
> + return s3c_hsotg_ep_disable_force(ep, false);
> +}
> /**
> * on_list - check request is on the given endpoint
> * @ep: The endpoint to check.
> @@ -2933,7 +2937,7 @@ static int s3c_hsotg_udc_stop(struct usb_gadget *gadget,
>
> /* all endpoints should be shutdown */
> for (ep = 1; ep < hsotg->num_of_eps; ep++)
> - s3c_hsotg_ep_disable(&hsotg->eps[ep].ep);
> + s3c_hsotg_ep_disable_force(&hsotg->eps[ep].ep, true);
>
> spin_lock_irqsave(&hsotg->lock, flags);
>
Acked-by: Paul Zimmerman <[email protected]>
On 12/11/2014 08:34 PM, Paul Zimmerman wrote:
>> From: Robert Baldyga [mailto:[email protected]]
>> Sent: Tuesday, December 09, 2014 5:42 AM
>>
>> This makes us sure that all requests are completed before we unbind
>> gadget. There are assumptions in gadget API that all requests have to
>> be completed and leak of complete can break some usb function drivers.
>>
>> For example unbind of ECM function can cause NULL pointer dereference:
>>
>> [ 26.396595] configfs-gadget gadget: unbind function
>> 'cdc_ethernet'/e79c4c00
>> [ 26.414999] Unable to handle kernel NULL pointer dereference at
>> virtual address 00000000
>> (...)
>> [ 26.452223] PC is at ecm_unbind+0x6c/0x9c
>> [ 26.456209] LR is at ecm_unbind+0x68/0x9c
>> (...)
>> [ 26.603696] [<c033fdb4>] (ecm_unbind) from [<c033661c>]
>> (purge_configs_funcs+0x94/0xd8)
>> [ 26.611674] [<c033661c>] (purge_configs_funcs) from [<c0336674>]
>> (configfs_composite_unbind+0x14/0x34)
>> [ 26.620961] [<c0336674>] (configfs_composite_unbind) from
>> [<c0337124>] (usb_gadget_remove_driver+0x68/0x9c)
>> [ 26.630683] [<c0337124>] (usb_gadget_remove_driver) from [<c03376c8>]
>> (usb_gadget_unregister_driver+0x64/0x94)
>> [ 26.640664] [<c03376c8>] (usb_gadget_unregister_driver) from
>> [<c0336be8>] (unregister_gadget+0x20/0x3c)
>> [ 26.650038] [<c0336be8>] (unregister_gadget) from [<c0336c84>]
>> (gadget_dev_desc_UDC_store+0x80/0xb8)
>> [ 26.659152] [<c0336c84>] (gadget_dev_desc_UDC_store) from
>> [<c0335120>] (gadget_info_attr_store+0x1c/0x28)
>> [ 26.668703] [<c0335120>] (gadget_info_attr_store) from [<c012135c>]
>> (configfs_write_file+0xe8/0x148)
>> [ 26.677818] [<c012135c>] (configfs_write_file) from [<c00c8dd4>]
>> (vfs_write+0xb0/0x1a0)
>> [ 26.685801] [<c00c8dd4>] (vfs_write) from [<c00c91b8>]
>> (SyS_write+0x44/0x84)
>> [ 26.692834] [<c00c91b8>] (SyS_write) from [<c000e560>]
>> (ret_fast_syscall+0x0/0x30)
>> [ 26.700381] Code: e30409f8 e34c0069 eb07b88d e59430a8 (e5930000)
>> [ 26.706485] ---[ end trace f62a082b323838a2 ]---
>>
>> It's because in some cases request is still running on endpoint during
>> unbind and kill_all_requests() called from s3c_hsotg_udc_stop() function
>> doesn't cause call of complete() of request. Missing complete() call
>> causes ecm->notify_req equals NULL in ecm_unbind() function, and this
>> is reason of this bug.
>>
>> Similar breaks can be observed in another usb function drivers.
>>
>> This patch fixes this bug forcing usb request completion in when
>> s3c_hsotg_ep_disable() is called from s3c_hsotg_udc_stop().
>>
>> Signed-off-by: Robert Baldyga <[email protected]>
>> ---
>> drivers/usb/dwc2/gadget.c | 10 +++++++---
>> 1 file changed, 7 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/usb/dwc2/gadget.c b/drivers/usb/dwc2/gadget.c
>> index 8b5c079..cb4c925 100644
>> --- a/drivers/usb/dwc2/gadget.c
>> +++ b/drivers/usb/dwc2/gadget.c
>> @@ -2590,7 +2590,7 @@ error:
>> * s3c_hsotg_ep_disable - disable given endpoint
>> * @ep: The endpoint to disable.
>> */
>> -static int s3c_hsotg_ep_disable(struct usb_ep *ep)
>> +static int s3c_hsotg_ep_disable_force(struct usb_ep *ep, bool force)
>> {
>> struct s3c_hsotg_ep *hs_ep = our_ep(ep);
>> struct s3c_hsotg *hsotg = hs_ep->parent;
>> @@ -2611,7 +2611,7 @@ static int s3c_hsotg_ep_disable(struct usb_ep *ep)
>>
>> spin_lock_irqsave(&hsotg->lock, flags);
>> /* terminate all requests with shutdown */
>> - kill_all_requests(hsotg, hs_ep, -ESHUTDOWN, false);
>> + kill_all_requests(hsotg, hs_ep, -ESHUTDOWN, force);
>>
>> hsotg->fifo_map &= ~(1<<hs_ep->fifo_index);
>> hs_ep->fifo_index = 0;
>> @@ -2632,6 +2632,10 @@ static int s3c_hsotg_ep_disable(struct usb_ep *ep)
>> return 0;
>> }
>>
>> +static int s3c_hsotg_ep_disable(struct usb_ep *ep)
>> +{
>> + return s3c_hsotg_ep_disable_force(ep, false);
>> +}
>> /**
>> * on_list - check request is on the given endpoint
>> * @ep: The endpoint to check.
>> @@ -2933,7 +2937,7 @@ static int s3c_hsotg_udc_stop(struct usb_gadget *gadget,
>>
>> /* all endpoints should be shutdown */
>> for (ep = 1; ep < hsotg->num_of_eps; ep++)
>> - s3c_hsotg_ep_disable(&hsotg->eps[ep].ep);
>> + s3c_hsotg_ep_disable_force(&hsotg->eps[ep].ep, true);
>>
>> spin_lock_irqsave(&hsotg->lock, flags);
>>
>
> Acked-by: Paul Zimmerman <[email protected]>
>
One more thing which I have forgot to write: This patch makes sense only
with my patch to udc-core [1] which modifies order of calls in
usb_gadget_remove_driver() function. Hence this patch shouldn't be
applied before [1] will be accepted.
[1] https://lkml.org/lkml/2014/12/12/360
Best regards,
Robert Baldyga