[igt-dev] [PATCH i-g-t] lib/igt_device_scan: Rescan pci properties if PCI_SLOT_NAME wasn't found

[Zbigniew Kempczyński]

On Wed, Aug 10, 2022 at 03:38:40PM +0300, Petri Latvala wrote:
> On Wed, Aug 10, 2022 at 10:19:49AM +0200, Zbigniew Kempczyński wrote:
> > References: https://gitlab.freedesktop.org/drm/intel/-/issues/6543
> > 
> > Due to lack of reproduction path of the above issue more data about
> > missing PCI_SLOT_NAME property is required.
> > 
> > What is extremely weird when PCI_SLOT_NAME is missing udev returns some
> > properties, like:
> > 
> > [properties]
> > DEVPATH                         : /devices/pci0000:00/0000:00:02.0
> > DRIVER                          : i915
> > PCI_CLASS                       : 30000
> > PCI_ID                          : 8086:191E
> > SUBSYSTEM                       : pci
> > 
> > To narrow the problematic code when PCI_SLOT_NAME is missing lets dump
> > kernel uevent file and retry scanning properties from udev. Retry path
> > allows to detect if udev returns same list of properties.
> > 
> > Above doesn't fix the issue. It is providing additional information about
> > devices and their properties provided from udev, especially when missing
> > PCI_SLOT_NAME will strike again. It also contains warning which might
> > detect situation where two scanning of properties will provide different
> > results.
> > 
> > Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
> > Cc: Petri Latvala <petri.latvala at intel.com>
> > ---
> >  lib/igt_device_scan.c | 48 +++++++++++++++++++++++++++++++++++++------
> >  1 file changed, 42 insertions(+), 6 deletions(-)
> > 
> > diff --git a/lib/igt_device_scan.c b/lib/igt_device_scan.c
> > index d6fae0650c..c2ecfcd3d4 100644
> > --- a/lib/igt_device_scan.c
> > +++ b/lib/igt_device_scan.c
> > @@ -568,28 +568,54 @@ static void dump_props_and_attrs(const struct igt_device *dev)
> >  	printf("\n");
> >  }
> >  
> > +static void dump_uevent_file(struct igt_device *dev)
> > +{
> > +	char filename[FILENAME_MAX];
> > +	const char *devpath = get_prop(dev, "DEVPATH");
> > +	char *line = NULL;
> > +	FILE *in;
> > +	size_t n;
> > +
> > +	igt_assert_f(devpath, "DEVPATH property doesn't exist\n");
> > +	snprintf(filename, FILENAME_MAX, "/sys%s/uevent", devpath);
> > +
> > +	in = fopen(filename, "r");
> > +	igt_assert(in);
> > +
> > +	printf("[uevent: %s]\n", filename);
> > +	while (getline(&line, &n, in) >= 0)
> > +		printf("%s", line);
> 
> Bleh, I was going to ask you to use igt_info instead so we get the log
> buffer out in the correct order, but the file already has a lot of
> printfs for lsgpu's sake. Let's consider that a cleanup TODO for
> later...
> 
> Reviewed-by: Petri Latvala <petri.latvala at intel.com>
> 
> > +
> > +	free(line);
> > +	fclose(in);
> > +}
> > +
> >  /*
> >   * Get PCI_SLOT_NAME property, it should be in format of
> >   * xxxx:yy:zz.z
> >   */
> > -static void set_pci_slot_name(struct igt_device *dev)
> > +static bool set_pci_slot_name(struct igt_device *dev)
> >  {
> >  	const char *pci_slot_name = get_prop(dev, "PCI_SLOT_NAME");
> >  	int len;
> >  
> >  	if (!pci_slot_name) {
> >  		dump_props_and_attrs(dev);
> > -		igt_assert_f(pci_slot_name, "PCI_SLOT_NAME property == NULL\n");
> > +		igt_warn("PCI_SLOT_NAME property == NULL\n");
> > +		dump_uevent_file(dev);
> > +		return false;
> >  	}
> >  
> >  	len = strlen(pci_slot_name);
> >  	if (len != PCI_SLOT_NAME_SIZE) {
> >  		dump_props_and_attrs(dev);
> > -		igt_assert_f(len != PCI_SLOT_NAME_SIZE,
> > -			     "PCI_SLOT_NAME length != %d [%s]\n", len, pci_slot_name);
> > +		igt_warn("PCI_SLOT_NAME length != %d [%s]\n", len, pci_slot_name);
> > +		dump_uevent_file(dev);
> > +		return false;
> >  	}
> >  
> >  	dev->pci_slot_name = strdup(pci_slot_name);
> > +	return true;
> >  }
> >  
> >  /*
> > @@ -649,7 +675,16 @@ static struct igt_device *igt_device_new_from_udev(struct udev_device *dev)
> >  		uint16_t vendor, device;
> >  
> >  		set_vendor_device(idev);
> > -		set_pci_slot_name(idev);
> > +
> > +		/*
> > +		 * Very rare we observe there's no PCI_SLOT_NAME property.
> > +		 * We depend on it so retry acquiring properties from udev.
> > +		 */
> > +		if (!set_pci_slot_name(idev)) {
> > +			g_hash_table_remove_all(idev->props_ht);
> > +			get_props(dev, idev);
> > +			igt_assert(set_pci_slot_name(idev));
> > +		}
> >  		get_pci_vendor_device(idev, &vendor, &device);
> >  		idev->codename = __pci_codename(vendor, device);
> >  		idev->dev_type = __pci_devtype(vendor, device, idev->pci_slot_name);
> > @@ -1270,7 +1305,8 @@ igt_devs_print_detail(struct igt_list_head *view,
> >  			_print_key_value("codename", dev->codename);
> >  		}
> >  
> > -		dump_props_and_attrs(dev);
> > +		if (is_pci_subsystem(dev))
> > +			dump_props_and_attrs(dev);

This incidentally left in patch I've sent (I just narrowed debugging
to pci subsystem). Do you want to resend or I can just go back with 
previous code and merge?

--
Zbigniew

> >  	}
> >  }
> >  
> > -- 
> > 2.34.1
> >