[PATCH]: Fix deadlock in USBIP driver (staging), linux-2.6.34-rc5

[PATCH]: Fix deadlock in USBIP driver (staging), linux-2.6.34-rc5

[Eric Lescouet]

Resent with hopefully everything ...

When detaching a port from the client side (usbip --detach 0),
the event thread, on the server side, is going to deadlock.
The "eh" server thread is getting USBIP_EH_RESET event and calls:
  -> stub_device_reset() -> usb_reset_device()
the USB framework is then calling back _in the same "eh" thread_ :
  -> stub_disconnect() -> usbip_stop_eh() -> wait_for_completion()
the "eh" thread is being asleep forever, waiting for its own completion.
This patch checks if "eh" is the current thread, in usbip_stop_eh().

Signed-off-by: Eric Lescouet <eric@lescouet.org>
---

diff -Nur linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c
--- linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c	2010-04-20 01:29:56.000000000 +0200
+++ linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c	2010-04-22 19:19:52.997889126 +0200
@@ -117,6 +117,9 @@
 {
 	struct usbip_task *eh = &ud->eh;
 
+	if (eh->thread == current)
+		return; /* do not wait for myself */
+
 	wait_for_completion(&eh->thread_done);
 	usbip_dbg_eh("usbip_eh has finished\n");
 }

Re: [PATCH]: Fix deadlock in USBIP driver (staging), linux-2.6.34-rc5

[Eric Lescouet]

> Odd, the patch has all leading tabs and spaces stripped out of it, is
> linewrapped, and the tabs are converted to spaces :(
>
> Care to try again, third time's a charm...

OK. It looked better on my box and on the list as well ...
... but let's try another client.

    Eric.

Signed-off-by: Eric Lescouet <eric@lescouet.org>
-------------

diff -Nur linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c
--- linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c	2010-04-20 01:29:56.000000000 +0200
+++ linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c	2010-04-22 19:19:52.997889126 +0200
@@ -117,6 +117,9 @@
 {
 	struct usbip_task *eh = &ud->eh;
 
+	if (eh->thread == current)
+		return; /* do not wait for myself */
+
 	wait_for_completion(&eh->thread_done);
 	usbip_dbg_eh("usbip_eh has finished\n");
 }

Re: [PATCH]: Fix deadlock in USBIP driver (staging), linux-2.6.34-rc5

[Greg KH]

On Thu, Apr 22, 2010 at 08:12:05PM +0200, Eric Lescouet wrote:
> Greg KH wrote:
> [...]
> >Ok, but does the lock_kernel() call you just made actually prevent this
> >from happening?  Isn't there some other lock you can use instead?
> >
> 
> Right. lock_kernel() is not needed (was copy/paste).
> The deadlock is avoided by not calling wait_for_completion()
> when the code is executed by the "eh" thread itself.
> 
> >Also, to have patches that we can apply, we need a "Signed-off-by:"
> >line.  See the file, Documentation/SubmittingPatches for what this
> >means.
> 
> Understood.
> 
> >Your email client wrapped the patch and made it unappliable :(
> >
> >thanks,
> >
> >greg k-h
> 
> Sorry about that. Please find below an hopefully better one.
> Thanks,
> 	Eric.
> 
> 
> Signed-off-by: Eric Lescouet <eric@lescouet.org>
> -------------
> 
> diff -Nur linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c
> linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c
> --- linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c
> 2010-04-20 01:29:56.000000000 +0200
> +++ linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c
> 2010-04-22 19:19:52.997889126 +0200
> @@ -117,6 +117,9 @@
> {
> struct usbip_task *eh = &ud->eh;
> +       if (eh->thread == current)
> +               return; /* do not wait for myself */
> +
> wait_for_completion(&eh->thread_done);
> usbip_dbg_eh("usbip_eh has finished\n");

Odd, the patch has all leading tabs and spaces stripped out of it, is
linewrapped, and the tabs are converted to spaces :(

Care to try again, third time's a charm...

thanks,

greg k-h

Re: [PATCH]: Fix deadlock in USBIP driver (staging), linux-2.6.34-rc5

[Eric Lescouet]

Greg KH wrote:
[...]
> Ok, but does the lock_kernel() call you just made actually prevent this
> from happening?  Isn't there some other lock you can use instead?
> 

Right. lock_kernel() is not needed (was copy/paste).
The deadlock is avoided by not calling wait_for_completion()
when the code is executed by the "eh" thread itself.

> Also, to have patches that we can apply, we need a "Signed-off-by:"
> line.  See the file, Documentation/SubmittingPatches for what this
> means.

Understood.

> Your email client wrapped the patch and made it unappliable :(
> 
> thanks,
> 
> greg k-h

Sorry about that. Please find below an hopefully better one.
Thanks,
	Eric.


Signed-off-by: Eric Lescouet <eric@lescouet.org>
-------------

diff -Nur linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c                   
--- linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c        2010-04-20 01:29:56.000000000 +0200                                       
+++ linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c    2010-04-22 19:19:52.997889126 +0200                                       
@@ -117,6 +117,9 @@
 {                                                                                                                                        
        struct usbip_task *eh = &ud->eh;                                                                                                  
                                                                                                                                          
+       if (eh->thread == current)                                                                                                        
+               return; /* do not wait for myself */                                                                                      
+                                                                                                                                         
        wait_for_completion(&eh->thread_done);                                                                                            
        usbip_dbg_eh("usbip_eh has finished\n");                                                                                          
 }                                                                                                                                        

Re: [PATCH]: Fix deadlock in USBIP driver (staging), linux-2.6.34-rc5

[Greg KH]

On Thu, Apr 22, 2010 at 06:54:02PM +0200, Eric Lescouet wrote:
> Hi,
> When detaching a port from the client side (usbip --detach 0),
> the event thread, on the server side, is going to deadlock.
> The "eh" server thread is getting USBIP_EH_RESET event and calls:
>  -> stub_device_reset() -> usb_reset_device()
> the USB framework is then calling back _in the same "eh" thread_ :
>  -> stub_disconnect() -> usbip_stop_eh() -> wait_for_completion()
> the "eh" thread is being asleep forever, waiting for its own completion.
> The patch checks if "eh" is the current thread, in usbip_stop_eh().

Ok, but does the lock_kernel() call you just made actually prevent this
from happening?  Isn't there some other lock you can use instead?

> Please Cc me in reply, I'm not in the list.

Also, to have patches that we can apply, we need a "Signed-off-by:"
line.  See the file, Documentation/SubmittingPatches for what this
means.
> 
> b.r.
> 
> ------------------
> 
> diff -Nur linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c
> linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c
> --- linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c
> 2010-04-20 01:29:56.000000000 +0200

Your email client wrapped the patch and made it unappliable :(

thanks,

greg k-h

[PATCH]: Fix deadlock in USBIP driver (staging), linux-2.6.34-rc5

[Eric Lescouet]

Hi,
When detaching a port from the client side (usbip --detach 0),
the event thread, on the server side, is going to deadlock.
The "eh" server thread is getting USBIP_EH_RESET event and calls:
  -> stub_device_reset() -> usb_reset_device()
the USB framework is then calling back _in the same "eh" thread_ :
  -> stub_disconnect() -> usbip_stop_eh() -> wait_for_completion()
the "eh" thread is being asleep forever, waiting for its own completion.
The patch checks if "eh" is the current thread, in usbip_stop_eh().

Please Cc me in reply, I'm not in the list.

b.r.

------------------

diff -Nur linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c 
linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c
--- linux-2.6.34-rc5/drivers/staging/usbip/usbip_event.c    2010-04-20 
01:29:56.000000000 +0200
+++ linux-2.6.34-rc5.new/drivers/staging/usbip/usbip_event.c    
2010-04-22 17:07:36.249588273 +0200
@@ -116,6 +116,13 @@
  void usbip_stop_eh(struct usbip_device *ud)
  {
      struct usbip_task *eh = &ud->eh;
+    int                i_am_eh;
+
+    lock_kernel();
+    i_am_eh = (eh->thread == current);
+    unlock_kernel();
+    if (i_am_eh)
+        return; /* do not wait for myself */

      wait_for_completion(&eh->thread_done);
      usbip_dbg_eh("usbip_eh has finished\n");