usb: core: fix a double free in the usb driver

PATCH Chung-Geol Kim 2 participants 2 messages Jun 3, 2016

lore

Share & tools

lore

Chung-Geol Kim

author

2016-06-03 11:51 UTC #0 lore raw

Quoted 27 lines · Chung-Geol Kim

Quoted 26 lines

Quoted 5 lines · Chung-Geol Kim

Quoted 1 line
There is a double free problem in the usb driver.
Which driver?

When I using the USB OTG Storage, this issue happened. When remove the OTG Storage, it reproduced sometimes.

     cpu 0                                                   cpu 1
     ----------------------------------------------------------------------------------------
     (*Insert USB Storage)
     usb_create_shared_hcd()
     kmalloc(primary_hcd)
     kmalloc(primary_hcd->bandwidth_mutex)
      ->(primary_hcd->kref==1)
     usb_get_hcd()
      ->(primary_hcd->kref==2)
                                                     usb_create_shared_hcd()
                                                     kmalloc(hcd->shared_hcd)
                                                      ->hcd->shared_hcd->bandwidth_mutex=primary->bandwidth_mutex
                                                      ->primary_hcd->primary_hcd = primary_hcd
                                                      ->hcd->shared_hcd->primary_hcd = primary_hcd
                                                     	->(hcd->shared_hcd->kref==1)
                                                     usb_get_hcd()
                                                     	->(hcd->shared_hcd->kref==2)

Quoted 6 lines

Quoted 2 lines
                                                     usb_get_hcd()
                                                      ->(hcd->shared_hcd->kref==3)
I don't understand. Why do these actions take place on two different CPUs? Aren't the primary_hcd and the shared_hcd structures allocated by the same thread, on the same CPU?

Yes, you are right, The presentational errors in order to obtain an understanding of the process. Therefore, I will be happy to explain again the diagrammatic representation as shown below. If using usb 3.0 storage(OTG), you can see as below.

==============================================
At *Insert USB(3.0) Storage
sequence <1> --> <5>
==============================================
VOLD
=================================|============

                             (uevent)
                           ______|___________ 
                          |<5>               |
                          |      SCSI        |
                          |usb_get_hcd       |
                          |shared_hcd(kref=3)|
                          |__________________|
   ___________________     ________|_________ 
  |<2>                |   |<4>               |
  |dwc3_otg_sm_work   |   |dwc3_otg_sm_work  |
  |usb_get_hcd        |   |usb_get_hcd       |
  |primary_hcd(kref=2)|   |shared_hcd(kref=2)|
  |___________________|   |__________________|
   _________|_________     ________|_________ 
  |<1>                |   |<3>               |
  |New USB BUS #1     |   |New USB BUS #2    |
  |usb_create_hcd     |   |usb_create_hcd    |
  |primary_hcd(kref=1)|   |shared_hcd(kref=1)|
  |                   |   |                  |
  |bandXX_mutex(alloc)|<-(Link)-bandXX_mutex |
  |___________________|   |__________________|

==============================================
At *remove USB(3.0) Storage
sequence <1> --> <5> ((Normal Case))
==============================================
VOLD
=================================|============

                             (uevent)
                           ______|___________ 
                          |<1>               |
                          |      SCSI        |
                          |usb_put_hcd       |
                          |shared_hcd(kref=2)|
                          |__________________|
   ___________________     ________|_________ 
  |<4>                |   |<2>               |
  |dwc3_otg_sm_work   |   |dwc3_otg_sm_work  |
  |usb_put_hcd        |   |usb_put_hcd       |
  |primary_hcd(kref=1)|   |shared_hcd(kref=1)|
  |___________________|   |__________________|
   _________|_________     ________|_________ 
  |<5>                |   |<3>               |
  |New USB BUS #1     |   |New USB BUS #2    |
  |hcd_release        |   |hcd_release       |
  |primary_hcd(kref=0)|   |shared_hcd(kref=0)|
  |                   |   |                  |
  |bandXX_mutex(free) | -X-cut off)-bandXX_mutex|
  |___________________|   |__________________|

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

Quoted 22 lines

Quoted 13 lines

    ---------------------------------------------------------------------------------------
     (*remove USB Storage)	
                                                     usb_release_dev()
                                                      ->(hcd->shared_hcd-kref==2)
                                                     usb_release_dev()
                                                      ->(hcd->shared_hcd-kref==1)
     usb_release_dev()
      -> (primary_hcd-kref==1)
     usb_release_dev()
      -> (primary_hcd-kref==0)
     hcd_release()
      -> kfree(primary_hcd->bandwidth_mutex)

Quoted 8 lines

      -> hcd->shared_hcd->primary_hcd = NULL
      -> kfree(primary_hcd)
                                                     usb_release_dev()
                                                      -> (hcd->shared_hcd-kref==0)
                                                     hcd_release()
                                                      -> usb_hcd_is_primary_hcd(hcd->shared_hcd)
                                                          -> hcd->shared_hcd->primary_hcd already NULL, return 1
                                                      -> try to double kfree(primary_hcd->bandwidth_mutex)

Quoted 5 lines

The same question applies here. Aren't the shared_hcd and primary_hcd structures released by the same thread, on the same CPU?
The real bug here is that the shared_hcd is released after the primary_hcd. That's what you need to fix.

NO, It 's only depend on vold(scsi) release time. If the vold later released and is being released first hcd, Double free happened at <5> as below.

==============================================
At *remove USB(3.0) Storage
sequence <1> --> <5> ((Problem Case))
==============================================
VOLD
=================================|============

                             (uevent)
                           ______|___________ 
                          |<5>               |
                          |      SCSI        |
                          |usb_put_hcd       |
                          |shared_hcd(kref=0)|
                          |*hcd_release      |
                          |bandXX_mutex(free*)|<- double free
                          |__________________|
   ___________________     ________|_________ 
  |<3>                |   |<1>               |
  |dwc3_otg_sm_work   |   |dwc3_otg_sm_work  |
  |usb_put_hcd        |   |usb_put_hcd       |
  |primary_hcd(kref=1)|   |shared_hcd(kref=2)|
  |___________________|   |__________________|
   _________|_________     ________|_________ 
  |<4>                |   |<2>               |
  |New USB BUS #1     |   |New USB BUS #2    |
  |hcd_release        |   |                  |
  |primary_hcd(kref=0)|   |shared_hcd(kref=1)|
  |                   |   |                  |
  |bandXX_mutex(free) |<-(Link)-bandXX_mutex |
  |___________________|   |__________________|

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

Quoted 35 lines

Quoted 18 lines
Since hcd->shared_hcd->priary_hcd was Null it didn't reach (hcd == hcd->primary_hcd) in usb_hcd_is_primary_hcd(). It returned 1 at since condition !hcd->primary_hcd is met.
Quoted 11 lines
drivers/usb/core/hcd.c static void hcd_release(struct kref *kref) · line 2611
Quoted 8 lines
drivers/usb/core/hcd.c +1 -1
@@ -2608,7 +2608,7 @@ static void hcd_release(struct kref *kref)
 struct usb_hcd *hcd = container_of (kref, struct usb_hcd, kref);
 
 mutex_lock(&usb_port_peer_mutex);
- if (usb_hcd_is_primary_hcd(hcd)) {
+ if (hcd == hcd->primary_hcd) {
That doesn't make sense, usb_hcd_is_primary_hcd() is the same as this check, what are you changing here?
Since hcd->priary_hcd was Null it didn't reach (hcd == hcd->primary_hcd). It returned 1 at since condition !hcd->primary_hcd is met.
Quoted 6 lines
int usb_hcd_is_primary_hcd(struct usb_hcd *hcd)
{
	if (!hcd->primary_hcd)
		return 1;
	return hcd == hcd->primary_hcd;
}
That's just a symptom, not the real cause of the bug. You need to fix the real cause: the shared_hcd has to be released _before_ the primary_hcd.
The right way to do this is to make the shared_hcd take a reference to the primary_hcd. This reference should be dropped when hcd_release() is called for the shared_hcd.
Alan Stern

Alan Stern

2016-06-03 14:28 UTC #1 lore raw

Quoted 23 lines · Chung-Geol Kim

Yes, you are right, The presentational errors in order to obtain an understanding of the process. Therefore, I will be happy to explain again the diagrammatic representation as shown below. If using usb 3.0 storage(OTG), you can see as below.
==============================================
At *Insert USB(3.0) Storage
sequence <1> --> <5>
==============================================
VOLD
=================================|============
                             (uevent)
                           ______|___________ 
                          |<5>               |
                          |      SCSI        |
                          |usb_get_hcd       |
                          |shared_hcd(kref=3)|
                          |__________________|
   ___________________     ________|_________ 
  |<2>                |   |<4>               |
  |dwc3_otg_sm_work   |   |dwc3_otg_sm_work  |
  |usb_get_hcd        |   |usb_get_hcd       |
  |primary_hcd(kref=2)|   |shared_hcd(kref=2)|
  |___________________|   |__________________|

Quoted 9 lines

   _________|_________     ________|_________ 
  |<1>                |   |<3>               |
  |New USB BUS #1     |   |New USB BUS #2    |
  |usb_create_hcd     |   |usb_create_hcd    |
  |primary_hcd(kref=1)|   |shared_hcd(kref=1)|
  |                   |   |                  |
  |bandXX_mutex(alloc)|<-(Link)-bandXX_mutex |
  |___________________|   |__________________|

When people present diagrams like this, the universal convention is to show earlier times at the top and later times at the bottom. That way the order in which you read the diagram is the same as the order in which the events are supposed to occur.

Also, the convention is to put events next to each other if they happen at the same time. In your diagram, <1> and <3> are next to each other but they don't happen at the same time.

Quoted 67 lines

                             (uevent)
                           ______|___________ 
                          |<1>               |
                          |      SCSI        |
                          |usb_put_hcd       |
                          |shared_hcd(kref=2)|
                          |__________________|
   ___________________     ________|_________ 
  |<4>                |   |<2>               |
  |dwc3_otg_sm_work   |   |dwc3_otg_sm_work  |
  |usb_put_hcd        |   |usb_put_hcd       |
  |primary_hcd(kref=1)|   |shared_hcd(kref=1)|
  |___________________|   |__________________|
   _________|_________     ________|_________ 
  |<5>                |   |<3>               |
  |New USB BUS #1     |   |New USB BUS #2    |
  |hcd_release        |   |hcd_release       |
  |primary_hcd(kref=0)|   |shared_hcd(kref=0)|
  |                   |   |                  |
  |bandXX_mutex(free) | -X-cut off)-bandXX_mutex|
  |___________________|   |__________________|

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

Quoted 2 lines

The real bug here is that the shared_hcd is released after the primary_hcd. That's what you need to fix.

NO, It 's only depend on vold(scsi) release time. If the vold later released and is being released first hcd, Double free happened at <5> as below.

                             (uevent)
                           ______|___________ 
                          |<5>               |
                          |      SCSI        |
                          |usb_put_hcd       |
                          |shared_hcd(kref=0)|
                          |*hcd_release      |
                          |bandXX_mutex(free*)|<- double free
                          |__________________|
   ___________________     ________|_________ 
  |<3>                |   |<1>               |
  |dwc3_otg_sm_work   |   |dwc3_otg_sm_work  |
  |usb_put_hcd        |   |usb_put_hcd       |
  |primary_hcd(kref=1)|   |shared_hcd(kref=2)|
  |___________________|   |__________________|
   _________|_________     ________|_________ 
  |<4>                |   |<2>               |
  |New USB BUS #1     |   |New USB BUS #2    |
  |hcd_release        |   |                  |
  |primary_hcd(kref=0)|   |shared_hcd(kref=1)|
  |                   |   |                  |
  |bandXX_mutex(free) |<-(Link)-bandXX_mutex |
  |___________________|   |__________________|

Quoted 1 line

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

That's the same as what I said. In your diagram, vold releases the shared_hcd (in <5>) after dwc3_otg_sm_work releases the primary_hcd (in <4>).

If you change the code so that the shared_hcd takes a reference to the primary_hcd and drops that reference when the shared_hcd is released, this will never occur.

Alan Stern