Re: [PATCH v3] PCI: vmd: Create domain symlink before pci_bus_add_devices()

[Bjorn Helgaas <helgaas@kernel.org>]

[+cc Pawel, Alexey, Tomasz for mdadm history]

On Wed, Jun 05, 2024 at 08:48:44PM +0800, Jiwei Sun wrote:
> From: Jiwei Sun <sunjw10@lenovo.com>
> 
> During booting into the kernel, the following error message appears:
> 
>   (udev-worker)[2149]: nvme1n1: '/sbin/mdadm -I /dev/nvme1n1'(err) 'mdadm: Unable to get real path for '/sys/bus/pci/drivers/vmd/0000:c7:00.5/domain/device''
>   (udev-worker)[2149]: nvme1n1: '/sbin/mdadm -I /dev/nvme1n1'(err) 'mdadm: /dev/nvme1n1 is not attached to Intel(R) RAID controller.'
>   (udev-worker)[2149]: nvme1n1: '/sbin/mdadm -I /dev/nvme1n1'(err) 'mdadm: No OROM/EFI properties for /dev/nvme1n1'
>   (udev-worker)[2149]: nvme1n1: '/sbin/mdadm -I /dev/nvme1n1'(err) 'mdadm: no RAID superblock on /dev/nvme1n1.'
>   (udev-worker)[2149]: nvme1n1: Process '/sbin/mdadm -I /dev/nvme1n1' failed with exit code 1.
> 
> This symptom prevents the OS from booting successfully.

I guess the root filesystem must be on a RAID device, and it's the
failure to assemble that RAID device that prevents OS boot?  The
messages are just details about why the assembly failed?

> After a NVMe disk is probed/added by the nvme driver, the udevd executes
> some rule scripts by invoking mdadm command to detect if there is a
> mdraid associated with this NVMe disk. The mdadm determines if one
> NVMe devce is connected to a particular VMD domain by checking the
> domain symlink. Here is the root cause:

Can you tell us something about what makes this a vmd-specific issue?

I guess vmd is the only driver that creates a "domain" symlink, so
*that* part is vmd-specific.  But I guess there's something in mdadm
or its configuration that looks for that symlink?

I suppose it has to do with the mdadm code at [1] and the commit at
[2]?

[1] https://github.com/md-raid-utilities/mdadm/blob/96b8035a09b6449ea99f2eb91f9ba4f6912e5bd6/platform-intel.c#L199
[2] https://github.com/md-raid-utilities/mdadm/commit/60f0f54d6f5227f229e7131d34f93f76688b085f

I assume this is a race between vmd_enable_domain() and mdadm?  And
vmd_enable_domain() only loses the race sometimes?  Trying to figure
out why this hasn't been reported before or on non-VMD configurations.
Now that I found [2], the non-VMD part is obvious, but I'm still
curious about why we haven't seen it before.

The VMD device is sort of like another host bridge, and I wouldn't
think mdadm would normally care about a host bridge, but it looks like
mdadm does need to know about VMD for some reason.

> Thread A                   Thread B             Thread mdadm
> vmd_enable_domain
>   pci_bus_add_devices
>     __driver_probe_device
>      ...
>      work_on_cpu
>        schedule_work_on
>        : wakeup Thread B
>                            nvme_probe
>                            : wakeup scan_work
>                              to scan nvme disk
>                              and add nvme disk
>                              then wakeup udevd
>                                                 : udevd executes
>                                                   mdadm command
>        flush_work                               main
>        : wait for nvme_probe done                ...
>     __driver_probe_device                        find_driver_devices
>     : probe next nvme device                     : 1) Detect the domain
>     ...                                            symlink; 2) Find the
>     ...                                            domain symlink from
>     ...                                            vmd sysfs; 3) The
>     ...                                            domain symlink is not
>     ...                                            created yet, failed
>   sysfs_create_link
>   : create domain symlink
> 
> sysfs_create_link() is invoked at the end of vmd_enable_domain().
> However, this implementation introduces a timing issue, where mdadm
> might fail to retrieve the vmd symlink path because the symlink has not
> been created yet.
> 
> Fix the issue by creating VMD domain symlinks before invoking
> pci_bus_add_devices().
> 
> Signed-off-by: Jiwei Sun <sunjw10@lenovo.com>
> Suggested-by: Adrian Huang <ahuang12@lenovo.com>
> ---
> v3 changes:
>  - Per Paul's comment, move sysfs_remove_link() after
>    pci_stop_root_bus()
> 
> v2 changes:
>  - Add "()" after function names in subject and commit log
>  - Move sysfs_create_link() after vmd_attach_resources()
> 
>  drivers/pci/controller/vmd.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/pci/controller/vmd.c b/drivers/pci/controller/vmd.c
> index 87b7856f375a..4e7fe2e13cac 100644
> --- a/drivers/pci/controller/vmd.c
> +++ b/drivers/pci/controller/vmd.c
> @@ -925,6 +925,9 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
>  		dev_set_msi_domain(&vmd->bus->dev,
>  				   dev_get_msi_domain(&vmd->dev->dev));
>  
> +	WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
> +			       "domain"), "Can't create symlink to domain\n");
> +
>  	vmd_acpi_begin();
>  
>  	pci_scan_child_bus(vmd->bus);
> @@ -964,9 +967,6 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
>  	pci_bus_add_devices(vmd->bus);
>  
>  	vmd_acpi_end();
> -
> -	WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
> -			       "domain"), "Can't create symlink to domain\n");
>  	return 0;
>  }
>  
> @@ -1042,8 +1042,8 @@ static void vmd_remove(struct pci_dev *dev)
>  {
>  	struct vmd_dev *vmd = pci_get_drvdata(dev);
>  
> -	sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
>  	pci_stop_root_bus(vmd->bus);
> +	sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
>  	pci_remove_root_bus(vmd->bus);
>  	vmd_cleanup_srcu(vmd);
>  	vmd_detach_resources(vmd);
> -- 
> 2.27.0
>