Re: [Linuxarm] [PATCH 1/2] hisi_acc_vfio_pci: Add debugfs to migration driver

[liulongfang <liulongfang@huawei.com>]

On 2022/10/18 14:55, liulongfang Wrote:
> On 2022/10/17 21:57, Shameerali Kolothum Thodi Wrote:
>>
>>
>>> -----Original Message-----
>>> From: liulongfang
>>> Sent: 17 October 2022 10:21
>>> To: John Garry <john.garry@huawei.com>; alex.williamson@redhat.com;
>>> jgg@nvidia.com; Shameerali Kolothum Thodi
>>> <shameerali.kolothum.thodi@huawei.com>
>>> Cc: cohuck@redhat.com; linux-kernel@vger.kernel.org;
>>> linuxarm@openeuler.org
>>> Subject: Re: [Linuxarm] [PATCH 1/2] hisi_acc_vfio_pci: Add debugfs to
>>> migration driver
>>>
>>> On 2022/10/14 17:20, John Garry wrote:
>>>> On 14/10/2022 03:57, Longfang Liu wrote:
>>>>> +static void hisi_acc_vf_debugfs_init(struct hisi_acc_vf_core_device
>>> *hisi_acc_vdev)
>>>>> +{
>>>>> +    struct pci_dev *vf_pdev = hisi_acc_vdev->vf_dev;
>>>>> +    struct device *dev = &vf_pdev->dev;
>>>>> +    int ret;
>>>>> +
>>>>> +    if (!atomic_read(&hisi_acc_root_ref))
>>>>> +        hisi_acc_debugfs_root = debugfs_create_dir("hisi_vfio_acc",
>>> NULL);
>>>>> +    atomic_inc(&hisi_acc_root_ref);
>>>>> +
>>>>
>>>> This looks totally racy, such that I wonder why even bother using an atomic
>>> for hisi_acc_root_ref.
>>>
>>>
>>> When enabling VF, it is possible for multiple VMs to enable VF at the same
>>> time. The atomic variable
>>> is used to ensure that only one "hisi_vfio_acc" is created. When other VFs
>>> are enabled,
>>> it will not be created again, but will be used directly.
>>>
>>>  Indeed, why is hisi_acc_debugfs_root not created in the driver module init?
>>>>
>>> Because the normal function of VF is to perform encryption and decryption
>>> services, the live migration
>>> function is an auxiliary function, which no need to be used in scenarios
>>> where only encryption and
>>> decryption services are performed.
>>>
>>> During module init, it can register ops(hisi_acc_vfio_pci_ops) that only
>>> perform encryption and
>>> decryption services, and can also register with live migration function
>>> ops(hisi_acc_vfio_pci_migrn_ops),
>>> and this debugfs only needs to register it when the the ops is
>>> hisi_acc_vfio_pci_migrn_ops.
>>
>> Isn't the ops registration happens at probe()? 

Both ops are registered in probe(), but if you move hisi_acc_debugfs_root() to probe(),
this debugfs will be created regardless of whether you need to perform live migration.

In any case, I think you can move
>> the hisi_acc_debugfs_root creation to module_init() as suggested above
>> and avoid the race.
>>
> 
> mig_root_ref uses atomic to prevent race, and another function is for reference counting.
> Even if this hisi_acc_debugfs_root() is moved to module_init(), this mig_root_ref is still
> needed, because the driver needs to ensure that the debugfs file is created when the first
> VF is enabled, destroy it when the last VF have unloaded.
> 
> 
>> Thanks,
>> Shameer
>>
> Thanks,
> Longfang.
> 
Thanks,
Longfang.