Re: [PATCH v4 3/5] usb: xhci: sideband: allocate sideband ring segments from a dedicated pool

[Wesley Cheng <wesley.cheng@oss.qualcomm.com>]

On 10/14/2026 5:44 PM, Wesley Cheng wrote:
> 
> 
> On 10/14/2026 7:26 AM, Mathias Nyman wrote:
>> On 10/13/26 12:11, Michal Pecio wrote:
>>> On Mon, 12 Oct 2026 18:01:50 -0700, Wesley Cheng wrote:
>>>> On 10/5/2026 4:03 AM, Mathias Nyman wrote:
>>>>> xhci is the one allocating the ring when adding an endpoint, and
>>>>> telling xHC hardware to write to that ring. It seems more
>>>>> approriate that xhci should be the one freeing the ring after it
>>>>> told xHC hardware to drop the endpoint, and there no longer is a
>>>>> risk on hardware accessing it.
>>>>
>>>> Hi Mathias,
>>>>
>>>> Sorry for the late response.  Was out last week.
>>>>
>>>> Specifically in the disconnect path (while audio offload is active),
>>>> when we utilize the usb_set_interface() path, it'll return early due
>>>> to the iface->unregistering check.  This would lead to the
>>>> drop_endpoint not to be called, and rely on the xhci_free_dev() to
>>>> handle the transfer ring cleanup.  This wouldn't work in the offload
>>>> situation, because qc_audio_offload is the only entity managing the
>>>> DMA pool, and that would be freed during the .disconnect callback,
>>>> which happens BEFORE xhci_free_dev().
>>
>> You are right, audio class driver .disconnect is called, driver unbound
>> and interface device gone before xhci_free_dev() is called.
>>
>> If sideband is unregistered as part of .disconnect callback it also means
>> ring/pool would be destroyed before xHC has dropped the endpoint/disabled
>> the slot. Would be better to free the ring after we know xHC controller
>> won't access it.
>>
>>>>
>>>> Would you be more inclined to prefer a solution where we can defer
>>>> the DMA pool free'ing to the xHCI core during xhci_free_dev, or maybe
>>>> adding another sideband notifier (sb->notify_client) enum to tell the
>>>> client driver that it can release the pool?
>>
>> sb->notify_client callback has the risk that this sideband entity is
>> unregistered and gone by the time we want to call it
>>
>> Deferring freeing the DMA poolby some kind of reference counting again adds
>> complexity, and maybe not a suitable solution for this issue.
>> (taking a reference at sideband register and every dma_pool_zalloc(), 
>> putting
>> when freeing or sideband unregister, finally freeing pool when reaching 
>> zero.)
>>
>> I'm trying to figure out a simpler solution.
>>
>> How about skipping dma pools completely for endpoints in these cases?
>> (page size over 4k with offloaded endpoints)
>>
>> Instead of sideband client creating a custom dma pool it could just set a
>> 'no dma pool' flag for the endpoint. xhci driver would allocate the
>> ring segments using dma_alloc_coherent() instead of dma_pool_zalloc() for
>> these, possibly even allocate a full page for each segment?
>> Freeing would be done similarly.
>>
>> Would this be a viable simple solution?
>>

Hi Mathias,

There is one difficulty I ran into, and that's w/ dma_free_coherent() not 
being able to be done in the current xhci_ring_free() context, as it is 
under atomic context.  I could move the dma memory ownership from xhci core 
to xhci sideband, which could probably get around this.

xHCI sideband will allocate the DMA backed memory then populate the 
new_ring reference.  Then when xhci_add_endpoint() --> xhci_endpoint_init() 
is run it'll check if new_ring is present, if not, follow the non offload 
path and allocate from xHCI segment pool.  Then on the sideband unregister 
path, we can handle the freeing of this DMA memory as well.  This kind of 
moves the memory operations from xHCI to xHCI sideband.  We would need to 
have a set of sideband specific ring alloc/free though.

Thanks
Wesley Cheng

> 
> I think so, this was something similar I proposed on v0:
> https://lore.kernel.org/all/20260824-16k_offload_v1_b4- 
> v1-0-49a6be60ca30@oss.qualcomm.com/
> 
> In that case, the client offload driver wasn't responsible for freeing 
> anything and that was all taken care of by xHCI core.  The sideband just 
> set the alignment_req.  Might need to make some adjustments after the 
> further discussions we've had.
> 
> Thanks
> Wesley Cheng
> 
>>>
>>> I'm not Mathias but my opinion is the same as always: none of these
>>> problems would exist if the sideband client allocated and freed all
>>> transfer/event ring memory, and xhci-sideband only owned a mapping.
>>>
>>> The mapping would be freed by xhci_sideband_unregister() or similar
>>> calls, which should ensure that the xHC stops using the rings, and
>>> memory would only be freed by the client driver after unregistering
>>> the sideband and stopping the coprocessor. Nothing is ever UAF or
>>> leaked, and ensuring this is *easy*. And if xhci-hcd randomly wants
>>> to get rid of some endpoint, it simply stops using the mapping and
>>> lets the sideband client clean it up two milliseconds later.
>>>
>>> One thing I'm not entirely sure about is how we could create this
>>> mapping. Using bare iommu_map() creates a hard dependency on IOMMU.
>>> Documentation/core-api/dma-attributes.rst mentions possibility of
>>> using generic dma_map_() helpers on multiple devices, so maybe this
>>> can work, but I haven't seen existing examples of such use.
>>>
>> Could be an option, but maybe not for this specific issue.
>> It's a larger, complex change without any real need (yet).
>>
>> I'm hoping vendors would move away from this sideband implementation
>> instead of evolving it, adapting the 'xHCI Audio Sideband Capability'
>> defined in xHCI  specification section 7.9 instead.
>>
>> Thanks
>> Mathias
>