On 2024/4/5 4:07, Alex Williamson wrote:
> On Tue, 2 Apr 2024 11:24:31 +0800
> Longfang Liu <liulongfang@xxxxxxxxxx> wrote:
>
>> 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
>> | +--attr
>> | +--data
>> | +--save
>> | +--cmd_state
>> |
>> +---<dev_name2>
>> +---migration
>> +--state
>> +--hisi_acc
>> +--attr
>> +--data
>> +--save
>> +--cmd_state
>>
>> data file: used to get the migration data from the driver
>> attr file: used to get device attributes parameters from the driver
>> save file: used to read the data of the live migration device and save
>> it to the driver.
>> cmd_state: used to get the cmd channel state for the device.
>>
>> +----------------+ +--------------+ +---------------+
>> | migration dev | | src dev | | dst dev |
>> +-------+--------+ +------+-------+ +-------+-------+
>> | | |
>> | | |
>> | | |
>> | | |
>> save | +------v-------+ +-------v-------+
>> | | saving_mif | | resuming_migf |
>> | | file | | file |
>> | +------+-------+ +-------+-------+
>> | | |
>> | mutex | |
>> +-------v--------+ | |
>> | | | |
>> | debug_migf file<---------------+-----------------------+
>> | | copy
>> +-------+--------+
>> |
>> cat |
>> |
>> +-------v--------+
>> | user |
>> +----------------+
>>
>> In debugfs scheme. The driver creates a separate debug_migf file.
>> It is completely separated from the two files of live migration,
>> thus preventing debugfs data from interfering with migration data.
>> Moreover, it only performs read operations on the device.
>>
>> For serialization of debugfs:
>> First, it only writes data when performing a debugfs save operation.
>
> This distinction between "writing" and "copying" is very confusing.
>
"Writing" means reading data from the device and writing it to debug_migf.
"Copying" is to copy the data that has been saved in other migf files to
debug_migf.
The destination of both operations is the same.
The data sources are different.
>> Second, it is only copied from the file on the migration device
>> when the live migration is complete.
>
> Why does it do this at all? If you're looking for a postmortem of the
> user generated buffer, that should be explicitly stated.
>
debug_migf is a data buffer. Used to cache debugfs data for users
>> These two operations are mutually exclusive through mutex.
>
> The mutual exclusion between debugfs operations is not the concern, the
> question is whether there's serialization that prevents the debugfs
> operations from interfering with the user migration flow. Nothing here
> seems to prevent concurrent use of the debugfs interface proposed here
> with a user migration.
>
Reading data directly from the device does not cause any problems.
The device supports concurrent requests for read operations.
Therefore, there is no mutual exclusion between the debugfs save
operation and the user migration operation.
>>
>> Signed-off-by: Longfang Liu <liulongfang@xxxxxxxxxx>
>> ---
>> .../vfio/pci/hisilicon/hisi_acc_vfio_pci.c | 204 ++++++++++++++++++
>> .../vfio/pci/hisilicon/hisi_acc_vfio_pci.h | 11 +
>> 2 files changed, 215 insertions(+)
>>
>> diff --git a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> index bf358ba94b5d..9f563a31a2a1 100644
>> --- a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> +++ b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c
>> @@ -15,6 +15,7 @@
>> #include <linux/anon_inodes.h>
>>
>> #include "hisi_acc_vfio_pci.h"
>> +#include "../../vfio.h"
>
> This include seems not to be required.
>
OK, let me modify it and verify it.
>>
>> /* Return 0 on VM acc device ready, -ETIMEDOUT hardware timeout */
>> static int qm_wait_dev_not_ready(struct hisi_qm *qm)
>> @@ -618,6 +619,22 @@ hisi_acc_check_int_state(struct hisi_acc_vf_core_device *hisi_acc_vdev)
>> }
>> }
>>
>> +static void hisi_acc_vf_migf_save(struct hisi_acc_vf_core_device *hisi_acc_vdev,
>> + struct hisi_acc_vf_migration_file *src_migf)
>
> Seems this should be named something relative to debug since it's only
> purpose is to copy the migration file to the debug migration file.
How about hisi_acc_debug_migf_copy?
>
>> +{
>> + struct hisi_acc_vf_migration_file *dst_migf = hisi_acc_vdev->debug_migf;
>> +
>> + if (!dst_migf)
>> + return;
>> +
>> + mutex_lock(&hisi_acc_vdev->enable_mutex);
>> + dst_migf->disabled = src_migf->disabled;
>
> In the cases where this is called, the caller is about to call
> hisi_acc_vf_disable_fd() which sets disabled = true and then
> hisi_acc_vf_debug_save() doesn't touch the value! What does it even
> mean to copy this value, let alone print it as part of the debugfs
> output later?
>
Yes, the disable assignment of debug_migf needs to be processed
in hisi_acc_vf_debug_save.
>
>> + dst_migf->total_length = src_migf->total_length;
>> + memcpy(&dst_migf->vf_data, &src_migf->vf_data,
>> + sizeof(struct acc_vf_data));
>> + mutex_unlock(&hisi_acc_vdev->enable_mutex);
>> +}
>> +
>> static void hisi_acc_vf_disable_fd(struct hisi_acc_vf_migration_file *migf)
>> {
>> mutex_lock(&migf->lock);
>> @@ -630,12 +647,14 @@ static void hisi_acc_vf_disable_fd(struct hisi_acc_vf_migration_file *migf)
>> static void hisi_acc_vf_disable_fds(struct hisi_acc_vf_core_device *hisi_acc_vdev)
>> {
>> if (hisi_acc_vdev->resuming_migf) {
>> + hisi_acc_vf_migf_save(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_vf_migf_save(hisi_acc_vdev, hisi_acc_vdev->saving_migf);
>
> Why are these buffers copied to the debug_migf in this case? This can
> happen asynchronous to accessing the debugfs migration file and there's
> no serialization.
>
We can try to copy when accessing debugfs.
>> hisi_acc_vf_disable_fd(hisi_acc_vdev->saving_migf);
>> fput(hisi_acc_vdev->saving_migf->filp);
>> hisi_acc_vdev->saving_migf = NULL;
>> @@ -1144,6 +1163,7 @@ static int hisi_acc_vf_qm_init(struct hisi_acc_vf_core_device *hisi_acc_vdev)
>> if (!vf_qm->io_base)
>> return -EIO;
>>
>> + mutex_init(&hisi_acc_vdev->enable_mutex);
>> vf_qm->fun_type = QM_HW_VF;
>> vf_qm->pdev = vf_dev;
>> mutex_init(&vf_qm->mailbox_lock);
>> @@ -1294,6 +1314,181 @@ 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_acc_vf_migration_file *migf = hisi_acc_vdev->debug_migf;
>> + struct hisi_qm *vf_qm = &hisi_acc_vdev->vf_qm;
>> + int ret;
>> +
>
> lockdep_assert_held(...)
>
>> + if (!vdev->mig_ops || !migf) {
>> + seq_printf(seq, "%s\n", "device does not support live migration!");
>> + return -EINVAL;
>
> Isn't the -EINVAL sufficient?
>
Which one do you think is better?
>> + }
>> +
>> + /**
>> + * 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 != DEV_OPEN) {
>
> Why are we assigning and testing a bool against and enum?!
>
OK, change to true and false assignment.
>> + seq_printf(seq, "%s\n", "device not opened!");
>> + return -EINVAL;
>> + }
>> +
>> + ret = qm_wait_dev_not_ready(vf_qm);
>> + if (ret) {
>> + seq_printf(seq, "%s\n", "VF device not ready!");
>> + return -EINVAL;
>
> -EBUSY? Again, not sure why we need the seq_printf() in addition to
> the error value.
>
OK, -EBUSY is better.
seq_printf() allows users to directly obtain the cause of the
error without checking dmesg.
>> + }
>> +
>> + 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->enable_mutex);
>> + ret = hisi_acc_vf_debug_check(seq, vdev);
>> + if (ret) {
>> + mutex_unlock(&hisi_acc_vdev->enable_mutex);
>> + return 0;
>
> Why do we squash the error return here and throughout?
>
seq_printf() gives the user a failure message.
This can be changed to "return ret;"
>> + }
>> +
>> + value = readl(vf_qm->io_base + QM_MB_CMD_SEND_BASE);
>> + mutex_unlock(&hisi_acc_vdev->enable_mutex);
>> + seq_printf(seq, "%s:0x%llx\n", "mailbox cmd channel state is OK", value);
>
> We didn't test the value, what makes the state OK? Can this readl() or
> those in qm_wait_dev_not_ready() interfere with the main device flow?
>
If the cmd channel is normal, it will return a non-all-F value.
Add exception checking in the next version.
cmd channel read operation will not affect the main migration process.
>> +
>> + return 0;
>> +}
>> +
>> +static int hisi_acc_vf_debug_save(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_acc_vf_migration_file *migf = hisi_acc_vdev->debug_migf;
>> + struct acc_vf_data *vf_data = &migf->vf_data;
>> + int ret;
>> +
>> + mutex_lock(&hisi_acc_vdev->enable_mutex);
>> + ret = hisi_acc_vf_debug_check(seq, vdev);
>> + if (ret) {
>> + mutex_unlock(&hisi_acc_vdev->enable_mutex);
>> + return 0;
>> + }
>> +
>> + vf_data->vf_qm_state = QM_READY;
>> + ret = vf_qm_read_data(&hisi_acc_vdev->vf_qm, vf_data);
>> + if (ret) {
>> + mutex_unlock(&hisi_acc_vdev->enable_mutex);
>> + seq_printf(seq, "%s\n", "failed to save device data!");
>> + return 0;
>> + }
>> +
>> + migf->total_length = sizeof(struct acc_vf_data);
>> + mutex_unlock(&hisi_acc_vdev->enable_mutex);
>> + seq_printf(seq, "%s\n", "successful to save device data!");
>> +
>> + return 0;
>> +}
>> +
>> +static int hisi_acc_vf_data_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);
>> + struct hisi_acc_vf_migration_file *debug_migf = hisi_acc_vdev->debug_migf;
>> + size_t vf_data_sz = offsetofend(struct acc_vf_data, padding);
>> +
>> + if (debug_migf && debug_migf->total_length)
>> + seq_hex_dump(seq, "Mig Data:", DUMP_PREFIX_OFFSET, 16, 1,
>> + (unsigned char *)&debug_migf->vf_data,
>> + vf_data_sz, false);
>> + else
>> + seq_printf(seq, "%s\n", "device not migrated!");
>
> "device state not saved"? Although I don't recall why this doesn't
> just return an errno.
>
OK, those who exit directly without migration will be processed according
to the error mode and an error code will be returned.
>> +
>> + return 0;
>> +}
>> +
>> +static int hisi_acc_vf_attr_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);
>> + struct hisi_acc_vf_migration_file *debug_migf = hisi_acc_vdev->debug_migf;
>> +
>> + if (debug_migf && debug_migf->total_length) {
>> + seq_printf(seq,
>> + "acc device:\n"
>> + "device state: %d\n"
>> + "device ready: %u\n"
>> + "data valid: %d\n"
>> + "data size: %lu\n",
>> + hisi_acc_vdev->mig_state,
>
> This is redundant to migration/state, however note
OK!
> hisi_acc_vfio_pci_get_device_state() protects the value with state
> mutex while reading it.
>
>> + hisi_acc_vdev->vf_qm_state,
>
> What's the purpose of this, it's ready or not, what does that mean?
>
If this status is not ready, the live migration process will exit directly.
It indicates that there are two possible exceptions:
1. The acc device driver in the VM does not have insmod.
2. The acc device driver in the VM is insmoded, but the cmd channel is abnormal.
>> + debug_migf->disabled,
>
> What's the purpose of this?
Get the enable status of migf file in the driver.
>
>> + debug_migf->total_length);
>
> Why not just have this printed or inferred via the above data_read
> function, this all seems unnecessary.
>
This is used to obtain some key status of the live migration driver.
It is more important than data in problem location.
So it is output in key-value pairs.
The above data is directly output in the form of hexadecimal data.
It is used for more detailed analysis when there are no abnormalities
in the key status.
>> + } else {
>> + seq_printf(seq, "%s\n", "device not migrated!");
>> + }
>> +
>> + return 0;
>> +}
>> +
>> +static int 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())
>> + return 0;
>> +
>> + migf = kzalloc(sizeof(struct hisi_acc_vf_migration_file), GFP_KERNEL);
>> + if (!migf)
>> + return -ENOMEM;
>> + hisi_acc_vdev->debug_migf = migf;
>> +
>> + vfio_dev_migration = debugfs_lookup("migration", vdev->debug_root);
>
> Fails to build without CONFIG_DEBUG_FS=y There should be a dependency
> on CONFIG_VFIO_DEBUGFS here somewhere.
>
Yes, the driver needs to add "!IS_ENABLED(CONFIG_VFIO_DEBUGFS)"
behind "debugfs_initialized()" above.
>> + if (!vfio_dev_migration) {
>> + kfree(migf);
>
> hisi_acc_vdev->debug_migf still points at this freed buffer, the return
> value of this function is not tested, allows a use-after-free in
> all of the below debugfs interfaces.
>
Yes, there needs to add "hisi_acc_vdev->debug_migf = NULL"
>> + dev_err(dev, "failed to lookup migration debugfs file!\n");
>> + return -ENODEV;
>> + }
>> +
>> + vfio_hisi_acc = debugfs_create_dir("hisi_acc", vfio_dev_migration);
>> + debugfs_create_devm_seqfile(dev, "data", vfio_hisi_acc,
>> + hisi_acc_vf_data_read);
>> + debugfs_create_devm_seqfile(dev, "attr", vfio_hisi_acc,
>> + hisi_acc_vf_attr_read);
>> + debugfs_create_devm_seqfile(dev, "cmd_state", vfio_hisi_acc,
>> + hisi_acc_vf_debug_cmd);
>> + debugfs_create_devm_seqfile(dev, "save", vfio_hisi_acc,
>> + hisi_acc_vf_debug_save);
>> +
>> + return 0;
>> +}
>> +
>> +static void hisi_acc_vf_debugfs_exit(struct hisi_acc_vf_core_device *hisi_acc_vdev)
>> +{
>> + if (!debugfs_initialized())
>> + return;
>> +
>> + kfree(hisi_acc_vdev->debug_migf);
>
> Double free if the lookup in init fails.
>
After adding "hisi_acc_vdev->debug_migf = NULL" above.
These processing codes need to be added here:
if (hisi_acc_vdev->debug_migf)
kfree(hisi_acc_vdev->debug_migf);
>> +}
>> +
>> 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);
>> @@ -1311,9 +1506,11 @@ static int hisi_acc_vfio_pci_open_device(struct vfio_device *core_vdev)
>> return ret;
>> }
>> hisi_acc_vdev->mig_state = VFIO_DEVICE_STATE_RUNNING;
>> + hisi_acc_vdev->dev_opened = DEV_OPEN;
>
> = true!
>
OK, the next version will not use enumeration values. will use true/false assignment.
>> }
>>
>> vfio_pci_core_finish_enable(vdev);
>> +
>> return 0;
>> }
>>
>> @@ -1322,7 +1519,10 @@ 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;
>>
>> + hisi_acc_vdev->dev_opened = DEV_CLOSE;
>> + mutex_lock(&hisi_acc_vdev->enable_mutex);
>> iounmap(vf_qm->io_base);
>> + mutex_unlock(&hisi_acc_vdev->enable_mutex);
>> vfio_pci_core_close_device(core_vdev);
>> }
>>
>> @@ -1413,6 +1613,9 @@ 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);
>> return 0;
>>
>> out_put_vdev:
>> @@ -1425,6 +1628,7 @@ static void hisi_acc_vfio_pci_remove(struct pci_dev *pdev)
>> struct hisi_acc_vf_core_device *hisi_acc_vdev = hisi_acc_drvdata(pdev);
>>
>> vfio_pci_core_unregister_device(&hisi_acc_vdev->core_device);
>> + hisi_acc_vf_debugfs_exit(hisi_acc_vdev);
>> vfio_put_device(&hisi_acc_vdev->core_device.vdev);
>> }
>>
>> diff --git a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.h b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.h
>> index 7a9dc87627cd..3a20d81d105c 100644
>> --- a/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.h
>> +++ b/drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.h
>> @@ -52,6 +52,11 @@
>> #define QM_EQC_DW0 0X8000
>> #define QM_AEQC_DW0 0X8020
>>
>> +enum acc_dev_state {
>> + DEV_CLOSE = 0x0,
>> + DEV_OPEN,
>> +};
>> +
>> struct acc_vf_data {
>> #define QM_MATCH_SIZE offsetofend(struct acc_vf_data, qm_rsv_state)
>> /* QM match information */
>> @@ -114,5 +119,11 @@ struct hisi_acc_vf_core_device {
>> int vf_id;
>> struct hisi_acc_vf_migration_file *resuming_migf;
>> struct hisi_acc_vf_migration_file *saving_migf;
>> +
>> + /* To make sure the device is enabled */
>> + struct mutex enable_mutex;
>> + bool dev_opened;
>> + /* For debugfs */
>> + struct hisi_acc_vf_migration_file *debug_migf;
>> };
>> #endif /* HISI_ACC_VFIO_PCI_H */
>
> .
>
Thank you very much for your careful inspection.
I will revise the inspection issues you mentioned above
one by one and publish them in the next version.
Thanks again!
Longfang.
