Re: [PATCH iwl-next 05/14] libeth: add control queue support

[Alexander Lobakin <aleksander.lobakin@intel.com>]

From: Leon Romanovsky <leon@kernel.org>
Date: Thu, 10 Apr 2025 16:27:06 +0300

> On Thu, Apr 10, 2025 at 02:58:28PM +0200, Larysa Zaremba wrote:
>> On Thu, Apr 10, 2025 at 02:23:49PM +0300, Leon Romanovsky wrote:
>>> On Thu, Apr 10, 2025 at 12:44:33PM +0200, Larysa Zaremba wrote:
>>>> On Thu, Apr 10, 2025 at 11:21:37AM +0300, Leon Romanovsky wrote:
>>>>> On Tue, Apr 08, 2025 at 02:47:51PM +0200, Larysa Zaremba wrote:
>>>>>> From: Phani R Burra <phani.r.burra@intel.com>
>>>>>>
>>>>>> Libeth will now support control queue setup and configuration APIs.
>>>>>> These are mainly used for mailbox communication between drivers and
>>>>>> control plane.
>>>>>>
>>>>>> Make use of the page pool support for managing controlq buffers.
>>>>>
>>>>> <...>
>>>>>
>>>>>>  libeth-y			:= rx.o
>>>>>>  
>>>>>> +obj-$(CONFIG_LIBETH_CP)		+= libeth_cp.o
>>>>>> +
>>>>>> +libeth_cp-y			:= controlq.o
>>>>>
>>>>> So why did you create separate module for it?
>>>>> Now you have pci -> libeth -> libeth_cp -> ixd, with the potential races between ixd and libeth, am I right?
>>>>>
>>>>
>>>> I am not sure what kind of races do you mean, all libeth modules themselves are 
>>>> stateless and will stay this way [0], all used data is owned by drivers.
>>>
>>> Somehow such separation doesn't truly work. There are multiple syzkaller
>>> reports per-cycle where module A tries to access module C, which already
>>> doesn't exist because it was proxied through module B.
>>
>> Are there similar reports for libeth and libie modules when iavf is enabled?
> 
> To get such report, syzkaller should run on physical iavf, it looks like it doesn't.
> Did I miss it here?
> https://syzkaller.appspot.com/upstream/s/net
> 
>> It is basically the same hierarchy. (iavf uses both libeth and libie, libie 
>> depends on libeth).
>>
>> I am just trying to understand, is this a regular situation or did I just mess 
>> smth up?
> 
> My review comment was general one. It is almost impossible to review
> this newly proposed architecture split for correctness.
> 
>>
>>>
>>>>
>>>> As for the module separation, I think there is no harm in keeping it modular. 
>>>
>>> Syzkaller reports disagree with you. 
>>>
>>
>> Could you please share them?
> 
> It is not an easy question to answer, because all these reports are complaining
> about some wrong locking order or NULL-pointer access. You will never know if
> it is because of programming or design error.
> 
> As an approximate example, see commits a27c6f46dcec ("RDMA/bnxt_re: Fix an issue in bnxt_re_async_notifier")
> and f0df225d12fc ("RDMA/bnxt_re: Add sanity checks on rdev validity").
> At the first glance, they look unrelated to our discussion, however
> they can serve as an example or races between deinit/disable paths in
> parent module vs. child.

Unrelated. At first, you were talking about module dependencies, now
you're talking about struct device etc dependencies, which is a
completely different thing.

As already said, libeth is stateless, so the latter one can't happen.
The former one is impossible at all. As long as at least 1 child module
is loaded, you can't unload the parent. And load/unload is serialized,
see module core code.

[...]

>> We did not think this would be a problem, intel has a tradition of calling the 
>> modules pretty ambiguously.
> 
> I know and it is worth to be changed.

Out of scope of this series.

Thanks,
Olek