[2/3] PCI: Allow specifying devices using a base bus and path of devfns

When specifying PCI devices on the kernel command line using a
BDF, the bus numbers can change when adding or replacing a device,
changing motherboard firmware, or applying kernel parameters like
pci=assign-buses. When this happens, it is usually undesirable to
apply whatever command line tweak to the wrong device.

Therefore, it is useful to be able to specify devices with a base
bus number and the path of devfns needed to get to it. (Similar to
the "device scope" structure in the Intel VT-d spec, Section 8.3.1.)

Thus, we add an option to specify devices in the following format:

path:[<domain>:]<bus>:<slot>.<func>/<slot>.<func>[/ ...]

The path can be any segment within the PCI hierarchy of any length and
determined through the use of 'lspci -t'. When specified this way, it is
less likely that a renumbered bus will result in a valid device specification
and the tweak won't be applied to the wrong device.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Stephen Bates <sbates@raithlin.com>
---
 Documentation/admin-guide/kernel-parameters.txt |  12 ++-
 drivers/pci/pci.c                               | 106 +++++++++++++++++++++++-
 2 files changed, 112 insertions(+), 6 deletions(-)

On Thu, 24 May 2018 15:48:15 -0600
Logan Gunthorpe <logang@deltatee.com> wrote:

> When specifying PCI devices on the kernel command line using a
> BDF, the bus numbers can change when adding or replacing a device,
> changing motherboard firmware, or applying kernel parameters like
> pci=assign-buses. When this happens, it is usually undesirable to
> apply whatever command line tweak to the wrong device.
> 
> Therefore, it is useful to be able to specify devices with a base
> bus number and the path of devfns needed to get to it. (Similar to
> the "device scope" structure in the Intel VT-d spec, Section 8.3.1.)
> 
> Thus, we add an option to specify devices in the following format:
> 
> path:[<domain>:]<bus>:<slot>.<func>/<slot>.<func>[/ ...]
> 
> The path can be any segment within the PCI hierarchy of any length and
> determined through the use of 'lspci -t'. When specified this way, it is
> less likely that a renumbered bus will result in a valid device specification
> and the tweak won't be applied to the wrong device.
> 
> Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
> Reviewed-by: Stephen Bates <sbates@raithlin.com>
> ---
>  Documentation/admin-guide/kernel-parameters.txt |  12 ++-
>  drivers/pci/pci.c                               | 106 +++++++++++++++++++++++-
>  2 files changed, 112 insertions(+), 6 deletions(-)
> 
> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
> index 894aa516ceab..519ab95bb418 100644
> --- a/Documentation/admin-guide/kernel-parameters.txt
> +++ b/Documentation/admin-guide/kernel-parameters.txt
> @@ -2986,9 +2986,10 @@
>  
>  				Some options herein operate on a specific device
>  				or a set of devices (<pci_dev>). These are
> -				specified in one of two formats:
> +				specified in one of three formats:
>  
>  				[<domain>:]<bus>:<slot>.<func>
> +				path:[<domain>:]<bus>:<slot>.<func>/<slot>.<func>[/ ...]
>  				pci:<vendor>:<device>[:<subvendor>:<subdevice>]
>  
>  				Note: the first format specifies a PCI
> @@ -2996,9 +2997,12 @@
>  				if new hardware is inserted, if motherboard
>  				firmware changes, or due to changes caused
>  				by other kernel parameters. The second format
> -				selects devices using IDs from the
> -				configuration space which may match multiple
> -				devices in the system.
> +				specifies a path from a device through
> +				a path of multiple slot/function addresses
> +				(this is more robust against renumbering
> +				issues). The third format selects devices using
> +				IDs from the configuration space which may match
> +				multiple devices in the system.
>  
>  		earlydump	[X86] dump PCI config space before the kernel
>  			        changes anything
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index 85fec5e2640b..53ea0d7b02ce 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -184,22 +184,116 @@ EXPORT_SYMBOL_GPL(pci_ioremap_wc_bar);
>  #endif
>  
>  /**
> + * pci_dev_str_match_path - test if a path string matches a device
> + * @dev:    the PCI device to test
> + * @p:      string to match the device against
> + * @endptr: pointer to the string after the match
> + *
> + * Test if a string (typically from a kernel parameter) formated as a
> + * path of slot/function addresses matches a PCI device. The string must
> + * be of the form:
> + *
> + *   [<domain>:]<bus>:<slot>.<func>/<slot>.<func>[/ ...]
> + *
> + * A path for a device can be obtained using 'lspci -t'. Using a path
> + * is more robust against renumbering of devices than using only
> + * a single bus, slot and function address.
> + *
> + * Returns 1 if the string matches the device, 0 if it does not and
> + * a negative error code if it fails to parse the string.
> + */
> +static int pci_dev_str_match_path(struct pci_dev *dev, const char *p,
> +				  const char **endptr)
> +{
> +	int ret;
> +	int seg, bus, slot, func, count;
> +	u8 *devfn_path;
> +	int num_devfn = 0;
> +	struct pci_dev *tmp;
> +
> +	ret = sscanf(p, "%x:%x:%x.%x%n", &seg, &bus, &slot,
> +		     &func, &count);
> +	if (ret != 4) {
> +		seg = 0;
> +		ret = sscanf(p, "%x:%x.%x%n", &bus, &slot,
> +			     &func, &count);
> +		if (ret != 3)
> +			return -EINVAL;
> +	}
> +
> +	p += count;
> +
> +	devfn_path = kmalloc(PAGE_SIZE, GFP_KERNEL);
> +	devfn_path[num_devfn++] = PCI_DEVFN(slot, func);
> +
> +	while (*p && *p != ',' && *p != ';') {
> +		ret = sscanf(p, "/%x.%x%n", &slot, &func, &count);
> +		if (ret != 2) {
> +			ret = -EINVAL;
> +			goto free_and_exit;
> +		}
> +
> +		p += count;
> +		devfn_path[num_devfn++] = PCI_DEVFN(slot, func);
> +		if (num_devfn >= PAGE_SIZE) {
> +			ret = -EINVAL;
> +			goto free_and_exit;
> +		}
> +	}
> +
> +	*endptr = p;
> +	ret = 0;
> +
> +	if (seg != pci_domain_nr(dev->bus))
> +		goto free_and_exit;
> +
> +	pci_dev_get(dev);
> +	while (num_devfn > 0 && dev) {
> +		num_devfn--;
> +
> +		if (devfn_path[num_devfn] != dev->devfn)
> +			goto put_and_exit;
> +
> +		if (num_devfn == 0 && bus == dev->bus->number) {
> +			ret = 1;
> +			goto put_and_exit;
> +		}
> +
> +		tmp = pci_dev_get(pci_upstream_bridge(dev));
> +		pci_dev_put(dev);
> +		dev = tmp;
> +	}
> +
> +put_and_exit:
> +	pci_dev_put(dev);
> +free_and_exit:
> +	kfree(devfn_path);
> +	return ret;
> +}

I like the idea, but can't we improve the implementation?  It seems
that we shouldn't need to allocate more than a working copy of the
original path string.  We can use strrchr() to find the last path
divider ('/'), match the slot.fn after that to the current devfn, set
that path divider to null, step to the next upstream device and
repeat.  Also, since we're working from a downstream device up, I
suspect we don't need to get and put references at each step, the
downstream device probably already holds a reference to the upstream
device for each step along the way.  It would be interesting to adapt
this to allow specifying a driver_override for a specific device as
well, I can think of a bunch of vfio users that would prefer that to
initrd scripts, but I know how much Bjorn loves more commandline
bloat ;)  Thanks,

Alex

> +
> +/**
>   * pci_dev_str_match - test if a string matches a device
>   * @dev:    the PCI device to test
>   * @p:      string to match the device against
>   * @endptr: pointer to the string after the match
>   *
>   * Test if a string (typically from a kernel parameter) matches a
> - * specified. The string may be of one of two forms formats:
> + * specified. The string may be of one of three formats:
>   *
>   *   [<domain>:]<bus>:<slot>.<func>
> + *   path:[<domain>:]<bus>:<slot>.<func>/<slot>.<func>[/ ...]
>   *   pci:<vendor>:<device>[:<subvendor>:<subdevice>]
>   *
>   * The first format specifies a PCI bus/slot/function address which
>   * may change if new hardware is inserted, if motherboard firmware changes,
>   * or due to changes caused in kernel parameters.
>   *
> - * The second format matches devices using IDs in the configuration
> + * The second format specifies a PCI bus/slot/function root address and
> + * a path of slot/function addresses to the specific device from the root.
> + * The path for a device can be determined through the use of 'lspci -t'.
> + * This format is more robust against renumbering issues than the first format.
> +
> + * The third format matches devices using IDs in the configuration
>   * space which may match multiple devices in the system. A value of 0
>   * for any field will match all devices.
>   *
> @@ -236,7 +330,15 @@ static int pci_dev_str_match(struct pci_dev *dev, const char *p,
>  		    (!subsystem_device ||
>  			    subsystem_device == dev->subsystem_device))
>  			goto found;
> +	} else if (strncmp(p, "path:", 5) == 0) {
> +		/* PCI Root Bus and a path of Slot,Function IDs */
> +		p += 5;
>  
> +		ret = pci_dev_str_match_path(dev, p, &p);
> +		if (ret < 0)
> +			return ret;
> +		else if (ret)
> +			goto found;
>  	} else {
>  		/* PCI Bus,Slot,Function ids are specified */
>  		ret = sscanf(p, "%x:%x:%x.%x%n", &seg, &bus, &slot,

On 30/05/18 10:23 AM, Alex Williamson wrote:
> I like the idea, but can't we improve the implementation?  It seems
> that we shouldn't need to allocate more than a working copy of the
> original path string.  We can use strrchr() to find the last path
> divider ('/'), match the slot.fn after that to the current devfn, set
> that path divider to null, step to the next upstream device and
> repeat.  

Ok, I'll give it a shot. I thought this would be a bit more tricky, but
perhaps not.

> Also, since we're working from a downstream device up, I
> suspect we don't need to get and put references at each step, the
> downstream device probably already holds a reference to the upstream
> device for each step along the way.  

That makes sense to me. I think it's something I added in similar p2pdma
code from somebody's review. But sounds like I can probably strip it out
there too and just put a comment noting this.

Thanks,

Logan