Re: [PATCH 1/4] hisi_acc_vfio_pci: fix VF reset timeout issue

[liulongfang <liulongfang@huawei.com>]

On 2026/1/17 0:47, Alex Williamson wrote:
> On Sun, 4 Jan 2026 15:07:03 +0800
> Longfang Liu <liulongfang@huawei.com> wrote:
> 
>> From: Weili Qian <qianweili@huawei.com>
>>
>> If device error occurs during live migration, qemu will
>> reset the VF. At this time, VF reset and device reset are performed
>> simultaneously. The VF reset will timeout. Therefore, the QM_RESETTING
>> flag is used to ensure that VF reset and device reset are performed
>> serially.
>>
>> Fixes: b0eed085903e ("hisi_acc_vfio_pci: Add support for VFIO live migration")
>> Signed-off-by: Weili Qian <qianweili@huawei.com>
>> ---
>>  .../vfio/pci/hisilicon/hisi_acc_vfio_pci.c    | 24 +++++++++++++++++++
>>  .../vfio/pci/hisilicon/hisi_acc_vfio_pci.h    |  2 ++
>>  2 files changed, 26 insertions(+)
>>
>> diff --git a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> index fe2ffcd00d6e..d55365b21f78 100644
>> --- a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> +++ b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> @@ -1188,14 +1188,37 @@ hisi_acc_vfio_pci_get_device_state(struct vfio_device *vdev,
>>  	return 0;
>>  }
>>  
>> +static void hisi_acc_vf_pci_reset_prepare(struct pci_dev *pdev)
>> +{
>> +	struct hisi_acc_vf_core_device *hisi_acc_vdev = hisi_acc_drvdata(pdev);
>> +	struct hisi_qm *qm = hisi_acc_vdev->pf_qm;
>> +	struct device *dev = &qm->pdev->dev;
>> +	u32 delay = 0;
>> +
>> +	/* All reset requests need to be queued for processing */
>> +	while (test_and_set_bit(QM_RESETTING, &qm->misc_ctl)) {
>> +		msleep(1);
>> +		if (++delay > QM_RESET_WAIT_TIMEOUT) {
>> +			dev_err(dev, "reset prepare failed\n");
>> +			return;
>> +		}
>> +	}
>> +
>> +	hisi_acc_vdev->set_reset_flag = true;
>> +}
>> +
>>  static void hisi_acc_vf_pci_aer_reset_done(struct pci_dev *pdev)
>>  {
>>  	struct hisi_acc_vf_core_device *hisi_acc_vdev = hisi_acc_drvdata(pdev);
>> +	struct hisi_qm *qm = hisi_acc_vdev->pf_qm;
>>  
>>  	if (hisi_acc_vdev->core_device.vdev.migration_flags !=
>>  				VFIO_MIGRATION_STOP_COPY)
>>  		return;
>>  
>> +	if (hisi_acc_vdev->set_reset_flag)
>> +		clear_bit(QM_RESETTING, &qm->misc_ctl);
> 
> 
> .reset_prepare sets QM_RESETTING unconditionally, .reset_done clears
> QM_RESETTING conditionally based on the migration state.  In 2/ this
> becomes conditional on the device supporting migration ops.  Doesn't
> this enable a scenario where a device that does not support migration
> puts QM_RESETTING into an inconsistent state that is never cleared?
> Should the clear_bit() occur before the migration state/capability
> check?
>

Yes,  it makes more sense to move clear_bit() before the migration state
or capability check.

Thanks,
Longfang.

> Thanks,
> Alex
> 
>> +
>>  	mutex_lock(&hisi_acc_vdev->state_mutex);
>>  	hisi_acc_vf_reset(hisi_acc_vdev);
>>  	mutex_unlock(&hisi_acc_vdev->state_mutex);
>> @@ -1746,6 +1769,7 @@ static const struct pci_device_id hisi_acc_vfio_pci_table[] = {
>>  MODULE_DEVICE_TABLE(pci, hisi_acc_vfio_pci_table);
>>  
>>  static const struct pci_error_handlers hisi_acc_vf_err_handlers = {
>> +	.reset_prepare = hisi_acc_vf_pci_reset_prepare,
>>  	.reset_done = hisi_acc_vf_pci_aer_reset_done,
>>  	.error_detected = vfio_pci_core_aer_err_detected,
>>  };
>> diff --git a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.h b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.h
>> index cd55eba64dfb..a3d91a31e3d8 100644
>> --- a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.h
>> +++ b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.h
>> @@ -27,6 +27,7 @@
>>  
>>  #define ERROR_CHECK_TIMEOUT		100
>>  #define CHECK_DELAY_TIME		100
>> +#define QM_RESET_WAIT_TIMEOUT  60000
>>  
>>  #define QM_SQC_VFT_BASE_SHIFT_V2	28
>>  #define QM_SQC_VFT_BASE_MASK_V2		GENMASK(15, 0)
>> @@ -128,6 +129,7 @@ struct hisi_acc_vf_migration_file {
>>  struct hisi_acc_vf_core_device {
>>  	struct vfio_pci_core_device core_device;
>>  	u8 match_done;
>> +	bool set_reset_flag;
>>  	/*
>>  	 * io_base is only valid when dev_opened is true,
>>  	 * which is protected by open_mutex.
> 
> .
>