Re: [PATCH v9 3/4] hisi_acc_vfio_pci: register debugfs for hisilicon migration driver

[liulongfang]

On 2024/10/11 16:09, Shameerali Kolothum Thodi wrote:
> 
> 
>> -----Original Message-----
>> From: liulongfang <liulongfang@xxxxxxxxxx>
>> Sent: Friday, September 13, 2024 10:55 AM
>> To: alex.williamson@xxxxxxxxxx; jgg@xxxxxxxxxx; Shameerali Kolothum
>> Thodi <shameerali.kolothum.thodi@xxxxxxxxxx>; Jonathan Cameron
>> <jonathan.cameron@xxxxxxxxxx>
>> Cc: kvm@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
>> linuxarm@xxxxxxxxxxxxx; liulongfang <liulongfang@xxxxxxxxxx>
>> Subject: [PATCH v9 3/4] hisi_acc_vfio_pci: register debugfs for hisilicon
>> migration driver
>>
>> On the debugfs framework of VFIO, if the CONFIG_VFIO_DEBUGFS macro is
>> enabled, the debug function is registered for the live migration driver
>> of the HiSilicon accelerator device.
>>
>> After registering the HiSilicon accelerator device on the debugfs
>> framework of live migration of vfio, a directory file "hisi_acc"
>> of debugfs is created, and then three debug function files are
>> created in this directory:
>>
>>    vfio
>>     |
>>     +---<dev_name1>
>>     |    +---migration
>>     |        +--state
>>     |        +--hisi_acc
>>     |            +--dev_data
>>     |            +--migf_data
>>     |            +--cmd_state
>>     |
>>     +---<dev_name2>
>>          +---migration
>>              +--state
>>              +--hisi_acc
>>                  +--dev_data
>>                  +--migf_data
>>                  +--cmd_state
>>
>> dev_data file: read device data that needs to be migrated from the
>> current device in real time
>> migf_data file: read the migration data of the last live migration
>> from the current driver.
>> cmd_state: used to get the cmd channel state for the device.
>>
>> +----------------+        +--------------+       +---------------+
>> | migration dev  |        |   src  dev   |       |   dst  dev    |
>> +-------+--------+        +------+-------+       +-------+-------+
>>         |                        |                       |
>>         |                 +------v-------+       +-------v-------+
>>         |                 |  saving_migf |       | resuming_migf |
>>   read  |                 |     file     |       |     file      |
>>         |                 +------+-------+       +-------+-------+
>>         |                        |          copy         |
>>         |                        +------------+----------+
>>         |                                     |
>> +-------v--------+                    +-------v--------+
>> |   data buffer  |                    |   debug_migf   |
>> +-------+--------+                    +-------+--------+
>>         |                                     |
>>    cat  |                                 cat |
>> +-------v--------+                    +-------v--------+
>> |   dev_data     |                    |   migf_data    |
>> +----------------+                    +----------------+
>>
>> When accessing debugfs, user can obtain the most recent status data
>> of the device through the "dev_data" file. It can read recent
>> complete status data of the device. If the current device is being
>> migrated, it will wait for it to complete.
>> The data for the last completed migration function will be stored
>> in debug_migf. Users can read it via "migf_data".
>>
>> Signed-off-by: Longfang Liu <liulongfang@xxxxxxxxxx>
>> ---
>>  .../vfio/pci/hisilicon/hisi_acc_vfio_pci.c    | 204 ++++++++++++++++++
>>  .../vfio/pci/hisilicon/hisi_acc_vfio_pci.h    |   7 +
>>  2 files changed, 211 insertions(+)
>>
>> diff --git a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> index a8c53952d82e..da9f5b9e6c5b 100644
>> --- a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> +++ b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> @@ -627,15 +627,30 @@ static void hisi_acc_vf_disable_fd(struct
>> hisi_acc_vf_migration_file *migf)
>>  	mutex_unlock(&migf->lock);
>>  }
>>
>> +static void hisi_acc_debug_migf_copy(struct hisi_acc_vf_core_device
>> *hisi_acc_vdev,
>> +	struct hisi_acc_vf_migration_file *src_migf)
>> +{
>> +	struct hisi_acc_vf_migration_file *dst_migf = hisi_acc_vdev-
>>> debug_migf;
>> +
>> +	if (!dst_migf)
>> +		return;
>> +
>> +	dst_migf->total_length = src_migf->total_length;
>> +	memcpy(&dst_migf->vf_data, &src_migf->vf_data,
>> +		sizeof(struct acc_vf_data));
>> +}
>> +
>>  static void hisi_acc_vf_disable_fds(struct hisi_acc_vf_core_device
>> *hisi_acc_vdev)
>>  {
>>  	if (hisi_acc_vdev->resuming_migf) {
>> +		hisi_acc_debug_migf_copy(hisi_acc_vdev, hisi_acc_vdev-
>>> resuming_migf);
>>  		hisi_acc_vf_disable_fd(hisi_acc_vdev->resuming_migf);
>>  		fput(hisi_acc_vdev->resuming_migf->filp);
>>  		hisi_acc_vdev->resuming_migf = NULL;
>>  	}
>>
>>  	if (hisi_acc_vdev->saving_migf) {
>> +		hisi_acc_debug_migf_copy(hisi_acc_vdev, hisi_acc_vdev-
>>> saving_migf);
>>  		hisi_acc_vf_disable_fd(hisi_acc_vdev->saving_migf);
>>  		fput(hisi_acc_vdev->saving_migf->filp);
>>  		hisi_acc_vdev->saving_migf = NULL;
>> @@ -1294,6 +1309,181 @@ static long hisi_acc_vfio_pci_ioctl(struct
>> vfio_device *core_vdev, unsigned int
>>  	return vfio_pci_core_ioctl(core_vdev, cmd, arg);
>>  }
>>
>> +static int hisi_acc_vf_debug_check(struct seq_file *seq, struct vfio_device
>> *vdev)
>> +{
>> +	struct hisi_acc_vf_core_device *hisi_acc_vdev =
>> hisi_acc_get_vf_dev(vdev);
>> +	struct hisi_qm *vf_qm = &hisi_acc_vdev->vf_qm;
>> +	int ret;
>> +
>> +	lockdep_assert_held(&hisi_acc_vdev->open_mutex);
>> +	/*
>> +	 * When the device is not opened, the io_base is not mapped.
>> +	 * The driver cannot perform device read and write operations.
>> +	 */
>> +	if (!hisi_acc_vdev->dev_opened) {
>> +		seq_printf(seq, "device not opened!\n");
>> +		return -EINVAL;
>> +	}
>> +
>> +	ret = qm_wait_dev_not_ready(vf_qm);
>> +	if (ret) {
>> +		seq_printf(seq, "VF device not ready!\n");
>> +		return -EBUSY;
>> +	}
> 
> Still not very sure this vf_qm ready() check actually helps or not? What guarantee
> is there that the qm will stay Ready after this call?  Any read/write afterwards
> will eventually fail if it is not ready  later for some reason, right? 
> Perhaps helps in early detection and bails out.
>

The purpose here is to intercept invalid debugfs operations.
When VF does not add a driver in Guest OS. The internal data reading and recovery
operations of the live migration operation will be skipped.
The data read is all empty data.

>> +	return 0;
>> +}
>> +
>> +static int hisi_acc_vf_debug_cmd(struct seq_file *seq, void *data)
>> +{
>> +	struct device *vf_dev = seq->private;
>> +	struct vfio_pci_core_device *core_device =
>> dev_get_drvdata(vf_dev);
>> +	struct vfio_device *vdev = &core_device->vdev;
>> +	struct hisi_acc_vf_core_device *hisi_acc_vdev =
>> hisi_acc_get_vf_dev(vdev);
>> +	struct hisi_qm *vf_qm = &hisi_acc_vdev->vf_qm;
>> +	u64 value;
>> +	int ret;
>> +
>> +	mutex_lock(&hisi_acc_vdev->open_mutex);
>> +	ret = hisi_acc_vf_debug_check(seq, vdev);
>> +	if (ret) {
>> +		mutex_unlock(&hisi_acc_vdev->open_mutex);
>> +		return ret;
>> +	}
>> +
>> +	value = readl(vf_qm->io_base + QM_MB_CMD_SEND_BASE);
>> +	if (value == QM_MB_CMD_NOT_READY) {
>> +		mutex_unlock(&hisi_acc_vdev->open_mutex);
>> +		seq_printf(seq, "mailbox cmd channel not ready!\n");
>> +		return -EINVAL;
>> +	}
>> +	mutex_unlock(&hisi_acc_vdev->open_mutex);
>> +	seq_printf(seq, "mailbox cmd channel ready!\n");
>> +
>> +	return 0;
>> +}
>> +
>> +static int hisi_acc_vf_dev_read(struct seq_file *seq, void *data)
>> +{
>> +	struct device *vf_dev = seq->private;
>> +	struct vfio_pci_core_device *core_device =
>> dev_get_drvdata(vf_dev);
>> +	struct vfio_device *vdev = &core_device->vdev;
>> +	struct hisi_acc_vf_core_device *hisi_acc_vdev =
>> hisi_acc_get_vf_dev(vdev);
>> +	size_t vf_data_sz = offsetofend(struct acc_vf_data, padding);
>> +	struct acc_vf_data *vf_data = NULL;
>> +	int ret;
>> +
>> +	vf_data = kzalloc(sizeof(struct acc_vf_data), GFP_KERNEL);
>> +	if (!vf_data)
>> +		return -ENOMEM;\
> 
> You could move the allocation after the below checks and to just before
> vf_qm_read_data().
>

OK.

>> +
>> +	mutex_lock(&hisi_acc_vdev->open_mutex);
>> +	ret = hisi_acc_vf_debug_check(seq, vdev);
>> +	if (ret) {
>> +		mutex_unlock(&hisi_acc_vdev->open_mutex);
>> +		goto migf_err;
>> +	}
>> +
>> +	mutex_lock(&hisi_acc_vdev->state_mutex);
>> +	vf_data->vf_qm_state = hisi_acc_vdev->vf_qm_state;
>> +	ret = vf_qm_read_data(&hisi_acc_vdev->vf_qm, vf_data);
>> +	if (ret) {
>> +		mutex_unlock(&hisi_acc_vdev->open_mutex);
>> +		mutex_unlock(&hisi_acc_vdev->state_mutex);
> 
> I think it is better to unlock in the reverse order. Also probably you can move
> the unlocks to a  goto area.
>

My consideration is to release the mutex before outputting the log.
it can be optimized.

>> +		goto migf_err;
>> +	}
>> +
>> +	mutex_unlock(&hisi_acc_vdev->open_mutex);
>> +	mutex_unlock(&hisi_acc_vdev->state_mutex);
> 
> Same as above.

OK.

> 
>> +	seq_hex_dump(seq, "Dev Data:", DUMP_PREFIX_OFFSET, 16, 1,
>> +			(unsigned char *)vf_data,
>> +			vf_data_sz, false);
>> +
>> +	seq_printf(seq,
>> +		 "acc device:\n"
>> +		 "guest driver load: %u\n"
>> +		 "data size: %lu\n",
>> +		 hisi_acc_vdev->vf_qm_state,
>> +		 sizeof(struct acc_vf_data));
>> +
>> +migf_err:
>> +	kfree(vf_data);
>> +
>> +	return ret;
>> +}
>> +
>> +static int hisi_acc_vf_migf_read(struct seq_file *seq, void *data)
>> +{
>> +	struct device *vf_dev = seq->private;
>> +	struct vfio_pci_core_device *core_device =
>> dev_get_drvdata(vf_dev);
>> +	struct vfio_device *vdev = &core_device->vdev;
>> +	struct hisi_acc_vf_core_device *hisi_acc_vdev =
>> hisi_acc_get_vf_dev(vdev);
>> +	size_t vf_data_sz = offsetofend(struct acc_vf_data, padding);
>> +	struct hisi_acc_vf_migration_file *debug_migf = hisi_acc_vdev-
>>> debug_migf;
>> +
>> +	/* Check whether the live migration operation has been performed
>> */
>> +	if (debug_migf->total_length < QM_MATCH_SIZE) {
>> +		seq_printf(seq, "device not migrated!\n");
>> +		return -EAGAIN;
>> +	}
>> +
>> +	seq_hex_dump(seq, "Mig Data:", DUMP_PREFIX_OFFSET, 16, 1,
>> +			(unsigned char *)&debug_migf->vf_data,
>> +			vf_data_sz, false);
>> +
>> +	seq_printf(seq,
>> +		 "acc device:\n"
>> +		 "guest driver load: %u\n"
>> +		 "device opened: %d\n"
>> +		 "migrate data length: %lu\n",
>> +		 hisi_acc_vdev->vf_qm_state,
>> +		 hisi_acc_vdev->dev_opened,
>> +		 debug_migf->total_length);
>> +
>> +	return 0;
>> +}
>> +
>> +static void hisi_acc_vfio_debug_init(struct hisi_acc_vf_core_device
>> *hisi_acc_vdev)
>> +{
>> +	struct vfio_device *vdev = &hisi_acc_vdev->core_device.vdev;
>> +	struct dentry *vfio_dev_migration = NULL;
>> +	struct dentry *vfio_hisi_acc = NULL;
>> +	struct device *dev = vdev->dev;
>> +	void *migf = NULL;
>> +
>> +	if (!debugfs_initialized() ||
>> +	    !IS_ENABLED(CONFIG_VFIO_DEBUGFS))
>> +		return;
>> +
>> +	vfio_dev_migration = debugfs_lookup("migration", vdev-
>>> debug_root);
>> +	if (!vfio_dev_migration) {
>> +		dev_err(dev, "failed to lookup migration debugfs file!\n");
>> +		return;
>> +	}
>> +
>> +	migf = kzalloc(sizeof(struct hisi_acc_vf_migration_file),
>> GFP_KERNEL);
>> +	if (!migf)
>> +		return;
>> +	hisi_acc_vdev->debug_migf = migf;
>> +
>> +	vfio_hisi_acc = debugfs_create_dir("hisi_acc", vfio_dev_migration);
>> +	debugfs_create_devm_seqfile(dev, "dev_data", vfio_hisi_acc,
>> +				  hisi_acc_vf_dev_read);
>> +	debugfs_create_devm_seqfile(dev, "migf_data", vfio_hisi_acc,
>> +				  hisi_acc_vf_migf_read);
>> +	debugfs_create_devm_seqfile(dev, "cmd_state", vfio_hisi_acc,
>> +				  hisi_acc_vf_debug_cmd);
>> +}
>> +
>> +static void hisi_acc_vf_debugfs_exit(struct hisi_acc_vf_core_device
>> *hisi_acc_vdev)
>> +{
>> +	if (hisi_acc_vdev->debug_migf) {
>> +		kfree(hisi_acc_vdev->debug_migf);
>> +		hisi_acc_vdev->debug_migf = NULL;
>> +	}
>> +}
>> +
>>  static int hisi_acc_vfio_pci_open_device(struct vfio_device *core_vdev)
>>  {
>>  	struct hisi_acc_vf_core_device *hisi_acc_vdev =
>> hisi_acc_get_vf_dev(core_vdev);
>> @@ -1305,12 +1495,16 @@ static int hisi_acc_vfio_pci_open_device(struct
>> vfio_device *core_vdev)
>>  		return ret;
>>
>>  	if (core_vdev->mig_ops) {
>> +		mutex_lock(&hisi_acc_vdev->open_mutex);
>>  		ret = hisi_acc_vf_qm_init(hisi_acc_vdev);
>>  		if (ret) {
>> +			mutex_unlock(&hisi_acc_vdev->open_mutex);
>>  			vfio_pci_core_disable(vdev);
>>  			return ret;
>>  		}
>>  		hisi_acc_vdev->mig_state = VFIO_DEVICE_STATE_RUNNING;
>> +		hisi_acc_vdev->dev_opened = true;
>> +		mutex_unlock(&hisi_acc_vdev->open_mutex);
>>  	}
>>
>>  	vfio_pci_core_finish_enable(vdev);
>> @@ -1322,7 +1516,10 @@ static void hisi_acc_vfio_pci_close_device(struct
>> vfio_device *core_vdev)
>>  	struct hisi_acc_vf_core_device *hisi_acc_vdev =
>> hisi_acc_get_vf_dev(core_vdev);
>>  	struct hisi_qm *vf_qm = &hisi_acc_vdev->vf_qm;
>>
>> +	mutex_lock(&hisi_acc_vdev->open_mutex);
>> +	hisi_acc_vdev->dev_opened = false;
>>  	iounmap(vf_qm->io_base);
>> +	mutex_unlock(&hisi_acc_vdev->open_mutex);
>>  	vfio_pci_core_close_device(core_vdev);
>>  }
>>
>> @@ -1342,6 +1539,7 @@ static int hisi_acc_vfio_pci_migrn_init_dev(struct
>> vfio_device *core_vdev)
>>  	hisi_acc_vdev->pf_qm = pf_qm;
>>  	hisi_acc_vdev->vf_dev = pdev;
>>  	mutex_init(&hisi_acc_vdev->state_mutex);
>> +	mutex_init(&hisi_acc_vdev->open_mutex);
>>
>>  	core_vdev->migration_flags = VFIO_MIGRATION_STOP_COPY |
>> VFIO_MIGRATION_PRE_COPY;
>>  	core_vdev->mig_ops = &hisi_acc_vfio_pci_migrn_state_ops;
>> @@ -1413,6 +1611,9 @@ static int hisi_acc_vfio_pci_probe(struct pci_dev
>> *pdev, const struct pci_device
>>  	ret = vfio_pci_core_register_device(&hisi_acc_vdev->core_device);
>>  	if (ret)
>>  		goto out_put_vdev;
>> +
>> +	if (ops == &hisi_acc_vfio_pci_migrn_ops)
>> +		hisi_acc_vfio_debug_init(hisi_acc_vdev);
> 
> I think there was a comment earlier on this. Still not sure why it is not possible to
> move ops == &hisi_acc_vfio_pci_migrn_ops check inside hisi_acc_vfio_debug_init().
>

The variable hisi_acc_vfio_pci_migrn_ops is declared after the function hisi_acc_vfio_debug_init.
If placed inside the function, a compilation error will occur.

>>  	return 0;
>>
>>  out_put_vdev:
>> @@ -1423,8 +1624,11 @@ static int hisi_acc_vfio_pci_probe(struct pci_dev
>> *pdev, const struct pci_device
>>  static void hisi_acc_vfio_pci_remove(struct pci_dev *pdev)
>>  {
>>  	struct hisi_acc_vf_core_device *hisi_acc_vdev =
>> hisi_acc_drvdata(pdev);
>> +	struct vfio_device *vdev = &hisi_acc_vdev->core_device.vdev;
>>
>>  	vfio_pci_core_unregister_device(&hisi_acc_vdev->core_device);
>> +	if (vdev->ops == &hisi_acc_vfio_pci_migrn_ops)
>> +		hisi_acc_vf_debugfs_exit(hisi_acc_vdev);
> 
> Do we need to do vdev->ops == &hisi_acc_vfio_pci_migrn_ops check here?
> 
> Since we are checking
>    hisi_acc_vdev->debug_migf inside the exit function, which I think is only
> set when the ops == migrn_ops. Right?
>

"vdev->ops == &hisi_acc_vfio_pci_migrn_ops" is added here to maintain symmetry
with the code processing in the probe.

I added it also because of another reviewer’s suggestion.

Thanks.
Longfang.

> Thanks,
> Shameer
> 
> .
>