Re: [PATCH v2] ice: ice_adapter: release xa entry on adapter allocation failure

[Jacob Keller <jacob.e.keller@intel.com>]

[-- Attachment #1.1: Type: text/plain, Size: 3249 bytes --]



On 9/30/2025 10:29 PM, Przemek Kitszel wrote:
> On 9/30/25 03:51, Haotian Zhang wrote:
>> When ice_adapter_new() fails, the reserved XArray entry created by
>> xa_insert() is not released. This causes subsequent insertions at
>> the same index to return -EBUSY, potentially leading to
>> NULL pointer dereferences.
>>
>> Release the reserved entry with xa_release() when adapter allocation
>> fails to prevent this issue.
>>
>> Fixes: 0f0023c649c7 ("ice: do not init struct ice_adapter more times than needed")
>> Suggested-by: Jacob Keller <jacob.e.keller@intel.com>
>> Signed-off-by: Haotian Zhang <vulab@iscas.ac.cn>
>>
>> ---
>> Changes in v2:
>>    - Instead of checking the return value of xa_store(), fix the real bug
>>      where a failed ice_adapter_new() would leave a stale entry in the
>>      XArray.
>>    - Use xa_release() to clean up the reserved entry, as suggested by
>>      Jacob Keller.
> 
> this is a correct improvement, but please let me propose other options,
> with 2. being my favorite:
> 
> 1. (just ice changes)
> change the call order to be:
> (note that it is under a mutex)
> xa_load() // return early if another adapter exists
> xa_reserve() // return early if no mem
> ice_adapter_new() // return early if err

You still have to xa_release() here if we return early, but adding the
call to xa_reserve might be more expressive of the intended behavior vs
using xa_insert was.

> xa_store()
> 
> 
> 2. (xarray changes)
> (perhaps I'm biased as the one introducing the error on error path):
> change xa_insert() to return 0 or -EEXIST when used as a reserving call
> (IOW: caller wanted to reserve, entry is reserved, so return should be 0
> (or -EEXIST if we really want to differentiate in the callers)).
> 

If we go this route, I think -EEXIST is the right answer, as it should
only return 0 if *this* call reserved the entry. -EEXIST instead of
-EBUSY could differentiate between "slot is reserved" and "slot is
filled" though.

That would let us fix the issue by having xa_insert differentiate and go
ahead if it fines a reserved entry that was unused. Thats safe for *our*
use case because we know we were under lock and the only way we'd have a
stale reserved entry is if we failed to release it...

I am not certain how other users or maintainer of xarray would feel
about such a change, which makes me think the ice side change is the
best at least initially.

> 
>> ---
>>   drivers/net/ethernet/intel/ice/ice_adapter.c | 4 +++-
>>   1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/net/ethernet/intel/ice/ice_adapter.c b/drivers/net/ethernet/intel/ice/ice_adapter.c
>> index b53561c34708..9eb100b11439 100644
>> --- a/drivers/net/ethernet/intel/ice/ice_adapter.c
>> +++ b/drivers/net/ethernet/intel/ice/ice_adapter.c
>> @@ -110,8 +110,10 @@ struct ice_adapter *ice_adapter_get(struct pci_dev *pdev)
>>   			return ERR_PTR(err);
>>   
>>   		adapter = ice_adapter_new(pdev);
>> -		if (!adapter)
>> +		if (!adapter) {
>> +			xa_release(&ice_adapters, index);
>>   			return ERR_PTR(-ENOMEM);
>> +		}
>>   		xa_store(&ice_adapters, index, adapter, GFP_KERNEL);
>>   	}
>>   	return adapter;
> 


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 236 bytes --]