On 10/03/2018 04:31 PM, Bjorn Helgaas wrote:

[EXTERNAL EMAIL]
Please report any suspicious attachments, links, or requests for sensitive information.


[+cc Stephen, Martin (for possible lspci changes)]

Hi Alexandru,

On Mon, Sep 03, 2018 at 01:02:28PM -0500, Alexandru Gagniuc wrote:

quoted

For certain bandwidth-critical devices (e.g. multi-port network cards)
it is useful to know the available bandwidth to the root complex. This
information is only available via the system log, which doesn't
account for link degradation after probing.

With a sysfs attribute, we can computes the bandwidth on-demand, and
will detect degraded links.

Signed-off-by: Alexandru Gagniuc <redacted>
---
  drivers/pci/pci-sysfs.c | 13 +++++++++++++
  1 file changed, 13 insertions(+)

diff --git a/drivers/pci/pci-sysfs.c b/drivers/pci/pci-sysfs.c
index 9ecfe13157c0..6658e927b1f5 100644
--- a/drivers/pci/pci-sysfs.c
+++ b/drivers/pci/pci-sysfs.c

@@ -218,6 +218,18 @@ static ssize_t current_link_width_show(struct device *dev,
  }
  static DEVICE_ATTR_RO(current_link_width);
  
+static ssize_t available_bandwidth_show(struct device *dev,
+				       struct device_attribute *attr, char *buf)
+{
+	struct pci_dev *pci_dev = to_pci_dev(dev);
+	u32 bw_avail;
+
+	bw_avail = pcie_bandwidth_available(pci_dev, NULL, NULL, NULL);
+
+	return sprintf(buf, "%u.%03u Gb/s\n", bw_avail / 1000, bw_avail % 1000);
+}
+static DEVICE_ATTR_RO(available_bandwidth);

Help me understand this.  We already have these sysfs attributes:

   max_link_speed          # eg, 16 GT/s
   max_link_width          # eg, 8
   current_link_speed      # eg, 16 GT/s
   current_link_width      # eg, 8

so I think the raw materials are already exposed.

quoted

The benefits I see for this new file are that

   - pcie_bandwidth_available() does the work of traversing up the
     tree, doing the computations (link width * speed, reduced by
     encoding overhead), and finding the minimum, and

   - it re-traverses the path every time we look at it, while the
     boot-time check is a one-time event.

In principle this could all be done in user space with the attributes
that are already exported.  There's some precedent for things like
this in lspci, e.g., "NUMA node" [1], and lspci might even be a more
user-friendly place for users to look for this, as opposed to
searching through sysfs.

Parsing the endpoint to root port bandwidth is, in principle, possible 
from userspace. It's just that in practice it's very clumsy to do, and, 
as you pointed out, not that reliable.

I understand it's not information that all users would jump in the air 
to know. However, it was important enough for certain use cases, that 
the kernel already has a very reliable way to calculate it.

It seems to me that the most elegant way is to let the kernel tell us, 
since the kernel already has this facility. To quote one of the texts 
under Documentation/, it is an elegant way to "avoid reinventing kernel 
wheels in userspace".

Alex

[1] https://git.kernel.org/pub/scm/utils/pciutils/pciutils.git/commit/?id=90ec4a6d0ae8

quoted

  static ssize_t secondary_bus_number_show(struct device *dev,
  					 struct device_attribute *attr,
  					 char *buf)

@@ -786,6 +798,7 @@ static struct attribute *pcie_dev_attrs[] = {
  	&dev_attr_current_link_width.attr,
  	&dev_attr_max_link_width.attr,
  	&dev_attr_max_link_speed.attr,
+	&dev_attr_available_bandwidth.attr,
  	NULL,
  };
  

-- 
2.17.1