Re: [PATCH] usb: xhci: Fix xhci_free_virt_devices_depth_first()

[Mathias Nyman <mathias.nyman@linux.intel.com>]

On 2.9.2025 13.55, Michał Pecio wrote:
> On Tue, 2 Sep 2025 13:13:12 +0300, Mathias Nyman wrote:
>> On 2.9.2025 12.07, Michał Pecio wrote:
>>> On Tue, 2 Sep 2025 10:46:30 +0200, Michał Pecio wrote:
>>>> On Tue, 2 Sep 2025 16:30:48 +0800 (CST), David Wang wrote:
>>>>> About the change from "<" to "<=", I did not observe any difference on my system. Is it because my system does not use up all slots?
>>>>
>>>> This too, you would need to fiddle with devices (or connect enough
>>>> of them) to reach Slot ID 255 (probably the highest on most systems),
>>>> depending on the xHCI controller and its ID allocation policy.
>>>
>>> This made me wonder what those policies are. I'm too lazy for thorough
>>> testing, but I plugged and unplugged the same device a few times.
>>>
>>> Most HCs kept assigning ID 1, so they likely always pick the lowest.
>>>
>>> My AMD chipset, two ASMedia USB 3.1 controllers and a Fresco FL1100
>>> kept assigning sequentially increasing IDs, so I suppose I could pump
>>> it up near the top, connect two high speed hubs and trigger this bug.
>>>    
>>>> But also as explained, this bug doesn't make things go boom just yet.
>>>>
>>>> Except if combined with your bug in an obscure edge case:
>>>>
>>>> 1. A high speed hub has slot ID HCS_MAX_SLOTS-1 and some TT children.
>>>> 2. Another high speed hub has slot ID HCS_MAX_SLOTS.
>>>> 3. We start with freeing the second hub.
>>>> 4. The loop is entered and leaves vdev pointing at the first hub.
>>>> 5. The first hub is freed instead of the second one.
>>>> 6. Then its children are freed and UAF its tt_info.
>>
>> I'm not sure I follow the above.
>>
>> I agree that changing the "<" to "<=" makes sense, but fortunately for us there shouldn't be any
>> issue with current implementation as xhci_free_virt_devices_depth_first() is called with highest possible
>> slot_id value first:
>>
>> in xhci-memm.c:
>>    for (i = HCS_MAX_SLOTS(xhci->hcs_params1); i > 0; i--)
>>                   xhci_free_virt_devices_depth_first(xhci, i);
>>
>> if HCS_MAX_SLOTS slot_id is a hs-hub then all its children have slot_id < HCS_MAX_SLOTS,
>> and loop works fine.
> 
> The loop works fine, but it exists with vdev pointing at MAX_SLOTS-1
> due to off by one and then this happens:
> 
> 	xhci_free_virt_device(xhci, vdev, slot_id);
> 
> which means:
> 
> 	xhci_free_virt_device(xhci, xhci->devs[MAX_SLOTS-1], MAX_SLOTS);
> 
> If MAX_SLOTS-1 is a high speed hub, it will be freed right now, without
> freeing its children first.
> 
> And whatever this device is, it will be freed without nulling
> xhci->devs[MAX_SLOTS-1], which might cause other UAF later (not sure).
> 
> I think it's possible, though I haven't tried actually triggering it.
> The problem didn't exist before this recent patch.
> 

I see, yes with the slot_id regression the incorrect vdev will cause issues.

Sent the patch to fix it forward

Thanks
Mathias