| Message ID | 1619034452-17334-1-git-send-email-wcheng@codeaurora.org (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | 43c4cab006f55b6ca549dd1214e22f5965a8675f |
| Headers | show |
| Series | [v2] usb: gadget: Fix double free of device descriptor pointers | expand |
Hi, Wesley Cheng <wcheng@codeaurora.org> writes: > From: Hemant Kumar <hemantk@codeaurora.org> > > Upon driver unbind usb_free_all_descriptors() function frees all > speed descriptor pointers without setting them to NULL. In case > gadget speed changes (i.e from super speed plus to super speed) > after driver unbind only upto super speed descriptor pointers get > populated. Super speed plus desc still holds the stale (already > freed) pointer. Fix this issue by setting all descriptor pointers > to NULL after freeing them in usb_free_all_descriptors(). could you describe this a little better? How can one trigger this case? Is the speed demotion happening after unbinding? It's not clear how to cause this bug.
On 4/22/2021 4:01 AM, Felipe Balbi wrote: > > Hi, > > Wesley Cheng <wcheng@codeaurora.org> writes: > >> From: Hemant Kumar <hemantk@codeaurora.org> >> >> Upon driver unbind usb_free_all_descriptors() function frees all >> speed descriptor pointers without setting them to NULL. In case >> gadget speed changes (i.e from super speed plus to super speed) >> after driver unbind only upto super speed descriptor pointers get >> populated. Super speed plus desc still holds the stale (already >> freed) pointer. Fix this issue by setting all descriptor pointers >> to NULL after freeing them in usb_free_all_descriptors(). > > could you describe this a little better? How can one trigger this case? > Is the speed demotion happening after unbinding? It's not clear how to > cause this bug. > Hi Felipe, Internally, we have a mechanism to switch the DWC3 core maximum speed parameter dynamically for displayport use cases. This issue happens whenever we have a maximum speed change occur on the USB gadget, which for DWC3 happens whenever we call gadget init. When we switch in and out of host mode, gadget init is being executed, leading to the change in the USB gadget max speed parameter: dwc->gadget->max_speed = dwc->maximum_speed; I know that configFS gadget has the max_speed sysfs file, which is a similar mechanism, but I haven't tried to see if we can reproduce the same issue with it. Let me see if we can reproduce this with that configfs speed setting. Thanks Wesley Cheng
On 4/23/2021 12:10 PM, Wesley Cheng wrote: > > > On 4/22/2021 4:01 AM, Felipe Balbi wrote: >> >> Hi, >> >> Wesley Cheng <wcheng@codeaurora.org> writes: >> >>> From: Hemant Kumar <hemantk@codeaurora.org> >>> >>> Upon driver unbind usb_free_all_descriptors() function frees all >>> speed descriptor pointers without setting them to NULL. In case >>> gadget speed changes (i.e from super speed plus to super speed) >>> after driver unbind only upto super speed descriptor pointers get >>> populated. Super speed plus desc still holds the stale (already >>> freed) pointer. Fix this issue by setting all descriptor pointers >>> to NULL after freeing them in usb_free_all_descriptors(). >> >> could you describe this a little better? How can one trigger this case? >> Is the speed demotion happening after unbinding? It's not clear how to >> cause this bug. >> > Hi Felipe, > > Internally, we have a mechanism to switch the DWC3 core maximum speed > parameter dynamically for displayport use cases. This issue happens > whenever we have a maximum speed change occur on the USB gadget, which > for DWC3 happens whenever we call gadget init. When we switch in and > out of host mode, gadget init is being executed, leading to the change > in the USB gadget max speed parameter: > > dwc->gadget->max_speed = dwc->maximum_speed; > > I know that configFS gadget has the max_speed sysfs file, which is a > similar mechanism, but I haven't tried to see if we can reproduce the > same issue with it. Let me see if we can reproduce this with that > configfs speed setting. > > Thanks > Wesley Cheng > Hi Felipe, So I tried with doing it through the configFS max_speed, but it doesn't have the same effect, as the setting done in dwc3_gadget_init() will still be assigning the composite/UDC device's maximum speed to SSP/SS. This is what the usb_assign_descriptor() uses to determine whether or not to copy the SSP and SS descriptors. So in summary, at least for a DWC3 based subsystem, the only way to reproduce it is if there is a way to dynamically switch the DWC3 core max speed parameter. Thanks Wesley Cheng
Hi, Wesley Cheng <wcheng@codeaurora.org> writes: >>>> From: Hemant Kumar <hemantk@codeaurora.org> >>>> >>>> Upon driver unbind usb_free_all_descriptors() function frees all >>>> speed descriptor pointers without setting them to NULL. In case >>>> gadget speed changes (i.e from super speed plus to super speed) >>>> after driver unbind only upto super speed descriptor pointers get >>>> populated. Super speed plus desc still holds the stale (already >>>> freed) pointer. Fix this issue by setting all descriptor pointers >>>> to NULL after freeing them in usb_free_all_descriptors(). >>> >>> could you describe this a little better? How can one trigger this case? >>> Is the speed demotion happening after unbinding? It's not clear how to >>> cause this bug. >>> >> Hi Felipe, >> >> Internally, we have a mechanism to switch the DWC3 core maximum speed >> parameter dynamically for displayport use cases. This issue happens >> whenever we have a maximum speed change occur on the USB gadget, which >> for DWC3 happens whenever we call gadget init. When we switch in and >> out of host mode, gadget init is being executed, leading to the change >> in the USB gadget max speed parameter: >> >> dwc->gadget->max_speed = dwc->maximum_speed; >> >> I know that configFS gadget has the max_speed sysfs file, which is a >> similar mechanism, but I haven't tried to see if we can reproduce the >> same issue with it. Let me see if we can reproduce this with that >> configfs speed setting. >> >> Thanks >> Wesley Cheng >> > > Hi Felipe, > > So I tried with doing it through the configFS max_speed, but it doesn't > have the same effect, as the setting done in dwc3_gadget_init() will > still be assigning the composite/UDC device's maximum speed to SSP/SS. > This is what the usb_assign_descriptor() uses to determine whether or > not to copy the SSP and SS descriptors. > > So in summary, at least for a DWC3 based subsystem, the only way to > reproduce it is if there is a way to dynamically switch the DWC3 core > max speed parameter. Could it be that you have a bug in your out-of-tree changes? Perhaps there's some assumption which your changes aren't guaranteeing.
On 4/24/2021 1:05 AM, Felipe Balbi wrote: > > Hi, > > Wesley Cheng <wcheng@codeaurora.org> writes: >>>>> From: Hemant Kumar <hemantk@codeaurora.org> >>>>> >>>>> Upon driver unbind usb_free_all_descriptors() function frees all >>>>> speed descriptor pointers without setting them to NULL. In case >>>>> gadget speed changes (i.e from super speed plus to super speed) >>>>> after driver unbind only upto super speed descriptor pointers get >>>>> populated. Super speed plus desc still holds the stale (already >>>>> freed) pointer. Fix this issue by setting all descriptor pointers >>>>> to NULL after freeing them in usb_free_all_descriptors(). >>>> >>>> could you describe this a little better? How can one trigger this case? >>>> Is the speed demotion happening after unbinding? It's not clear how to >>>> cause this bug. >>>> >>> Hi Felipe, >>> >>> Internally, we have a mechanism to switch the DWC3 core maximum speed >>> parameter dynamically for displayport use cases. This issue happens >>> whenever we have a maximum speed change occur on the USB gadget, which >>> for DWC3 happens whenever we call gadget init. When we switch in and >>> out of host mode, gadget init is being executed, leading to the change >>> in the USB gadget max speed parameter: >>> >>> dwc->gadget->max_speed = dwc->maximum_speed; >>> >>> I know that configFS gadget has the max_speed sysfs file, which is a >>> similar mechanism, but I haven't tried to see if we can reproduce the >>> same issue with it. Let me see if we can reproduce this with that >>> configfs speed setting. >>> >>> Thanks >>> Wesley Cheng >>> >> >> Hi Felipe, >> >> So I tried with doing it through the configFS max_speed, but it doesn't >> have the same effect, as the setting done in dwc3_gadget_init() will >> still be assigning the composite/UDC device's maximum speed to SSP/SS. >> This is what the usb_assign_descriptor() uses to determine whether or >> not to copy the SSP and SS descriptors. >> >> So in summary, at least for a DWC3 based subsystem, the only way to >> reproduce it is if there is a way to dynamically switch the DWC3 core >> max speed parameter. > > Could it be that you have a bug in your out-of-tree changes? Perhaps > there's some assumption which your changes aren't guaranteeing. > Hi Felipe, Unless there is a limitation on how the USB gadget max speed can be used, i.e. the USB gadget max speed MUST stay the same throughout a device's boot period, then our out of tree changes may have the wrong assumptions. However, I don't see that stated anywhere, and I still feel the current usb_assign_descriptors() and usb_free_all_descriptors() aren't aligned with one another. One API decides which descriptors to copy based on a parameter, whereas the other just frees all of them irrespective. Thanks Wesley Cheng
Hi, Wesley Cheng <wcheng@codeaurora.org> writes: > On 4/24/2021 1:05 AM, Felipe Balbi wrote: >> >> Hi, >> >> Wesley Cheng <wcheng@codeaurora.org> writes: >>>>>> From: Hemant Kumar <hemantk@codeaurora.org> >>>>>> >>>>>> Upon driver unbind usb_free_all_descriptors() function frees all >>>>>> speed descriptor pointers without setting them to NULL. In case >>>>>> gadget speed changes (i.e from super speed plus to super speed) >>>>>> after driver unbind only upto super speed descriptor pointers get >>>>>> populated. Super speed plus desc still holds the stale (already >>>>>> freed) pointer. Fix this issue by setting all descriptor pointers >>>>>> to NULL after freeing them in usb_free_all_descriptors(). >>>>> >>>>> could you describe this a little better? How can one trigger this case? >>>>> Is the speed demotion happening after unbinding? It's not clear how to >>>>> cause this bug. >>>>> >>>> Hi Felipe, >>>> >>>> Internally, we have a mechanism to switch the DWC3 core maximum speed >>>> parameter dynamically for displayport use cases. This issue happens >>>> whenever we have a maximum speed change occur on the USB gadget, which >>>> for DWC3 happens whenever we call gadget init. When we switch in and >>>> out of host mode, gadget init is being executed, leading to the change >>>> in the USB gadget max speed parameter: >>>> >>>> dwc->gadget->max_speed = dwc->maximum_speed; >>>> >>>> I know that configFS gadget has the max_speed sysfs file, which is a >>>> similar mechanism, but I haven't tried to see if we can reproduce the >>>> same issue with it. Let me see if we can reproduce this with that >>>> configfs speed setting. >>>> >>>> Thanks >>>> Wesley Cheng >>>> >>> >>> Hi Felipe, >>> >>> So I tried with doing it through the configFS max_speed, but it doesn't >>> have the same effect, as the setting done in dwc3_gadget_init() will >>> still be assigning the composite/UDC device's maximum speed to SSP/SS. >>> This is what the usb_assign_descriptor() uses to determine whether or >>> not to copy the SSP and SS descriptors. >>> >>> So in summary, at least for a DWC3 based subsystem, the only way to >>> reproduce it is if there is a way to dynamically switch the DWC3 core >>> max speed parameter. >> >> Could it be that you have a bug in your out-of-tree changes? Perhaps >> there's some assumption which your changes aren't guaranteeing. >> > Hi Felipe, > > Unless there is a limitation on how the USB gadget max speed can be > used, i.e. the USB gadget max speed MUST stay the same throughout a > device's boot period, then our out of tree changes may have the wrong no, not throughout boot, but it can't change while the device is enumerated. > assumptions. However, I don't see that stated anywhere, and I still > feel the current usb_assign_descriptors() and usb_free_all_descriptors() > aren't aligned with one another. One API decides which descriptors to > copy based on a parameter, whereas the other just frees all of them > irrespective. that's a valid point.
diff --git a/drivers/usb/gadget/config.c b/drivers/usb/gadget/config.c index 2d11535..8bb2577 100644 --- a/drivers/usb/gadget/config.c +++ b/drivers/usb/gadget/config.c @@ -194,9 +194,13 @@ EXPORT_SYMBOL_GPL(usb_assign_descriptors); void usb_free_all_descriptors(struct usb_function *f) { usb_free_descriptors(f->fs_descriptors); + f->fs_descriptors = NULL; usb_free_descriptors(f->hs_descriptors); + f->hs_descriptors = NULL; usb_free_descriptors(f->ss_descriptors); + f->ss_descriptors = NULL; usb_free_descriptors(f->ssp_descriptors); + f->ssp_descriptors = NULL; } EXPORT_SYMBOL_GPL(usb_free_all_descriptors);