Message-ID: <CAC1AzdfwMxLJk9G5x7WSDO9QdaFDpVhCWdNSys5hDho9h1DhEA@mail.gmail.com>
Date:	Thu, 23 Jul 2015 21:31:23 -0500
From:	Jordan Hargrave <jharg93@...il.com>
To:	Bjorn Helgaas <bhelgaas@...gle.com>
Cc:	Jean Delvare <jdelvare@...e.de>, linux-kernel@...r.kernel.org,
	linux-pci@...r.kernel.org,
	Jordan Hargrave <Jordan_Hargrave@...l.com>,
	Matthew Wilcox <matthew.r.wilcox@...el.com>
Subject: Re: [PATCH] Add support for reading SMBIOS Slot number for PCI devices

On Thu, Jul 23, 2015 at 12:24 PM, Bjorn Helgaas <bhelgaas@...gle.com> wrote:
> [+cc Matthew for PCIe-SSD perspective]
>
> On Wed, Jul 22, 2015 at 03:07:46PM -0500, Jordan Hargrave wrote:
>> On Tue, Jul 21, 2015 at 8:09 PM, Bjorn Helgaas <bhelgaas@...gle.com> wrote:
>> > On Tue, Jul 21, 2015 at 12:31:35PM -0500, Jordan Hargrave wrote:
>> >> On Tue, Jul 21, 2015 at 11:57 AM, Bjorn Helgaas <bhelgaas@...gle.com> wrote:
>> >> > On Mon, Jul 13, 2015 at 09:57:32AM -0500, Jordan Hargrave wrote:
>> >> >> On Mon, Jul 13, 2015 at 2:35 AM, Jean Delvare <jdelvare@...e.de> wrote:
>> >> >>
>> >> >> > Hi Jordan,
>> >> >> >
>> >> >> > On Fri, 10 Jul 2015 17:02:46 -0500, Jordan Hargrave wrote:
>> >> >> > > From: Jordan Hargrave <Jordan_Hargrave@...l.com>
>> >> >> > >
>> >> >> > > There currently isn't an easy way to determine which PCI devices belong
>> >> >> > to
>> >> >> > > system slots.  This patch adds support to read SMBIOS Type 9 (System
>> >> >> > Slots).
>> >> >> >
>> >> >> > I'm wondering, can't you use dmidecode or libsmbios to retrieve the
>> >> >> > same information?
>> >> >> >
>> >> >> > --
>> >> >> > Jean Delvare
>> >> >> > SUSE L3 Support
>> >> >> >
>> >> >>
>> >> >> You can but it's as not easy to determine the slot number for leaf devices
>> >> >> on bridges.  Eventually planning on using this for pulling slot number for
>> >> >> identifying network cards and disk numbering for systemd
>> >> >
>> >> > Can you outline the problems with using dmidecode or libsmbios?
>> >>
>> >> Neither dmidecode nor libsmbios report the slot number for devices
>> >> behind bridges in a slot.
>> >
>> > True, but it's straightforward to walk up the PCI tree in sysfs, e.g.,
>> > given a path like /sys/devices/pci0000:00/0000:00:1c.5/0000:03:00.0/, it's
>> > easy to see what the upstream bridges are.
>> >
>> It makes it more complicated I think. You have to check all functions
>> on all devices as well on the walk to the root.
>>
>> Would it look something like this?
>> while (pdev) {
>>   if (pdev->bus->number == dslot->bus && PCI_SLOT(pdev->devfn) ==
>> PCI_SLOT(dslot->devfn) &&
>>      (pci_pcie_type(pdev) != PCI_ROOT_PORT || PCI_FUNC(pdev->devfn) ==
>> PCI_FUNC(dslot->devfn))
>>     return dslot->instance;
>>   if (!pdev->bus->parent)
>>     break;
>>   pdev = pdev->bus->parent->self;
>> }
>>
>> >> I'm wanting to use this sysfs variable to
>> >> get slot numbers for systemd, so using libsmbios and dmidecode aren't
>> >> very useful.
>> >
>> > If you want this in systemd, I see why you wouldn't want a command like
>> > dmidecode.  Help me understand the problem with libsmbios.  Is it not
>> > useful because (a) systemd doesn't want to link with it, or (b) libsmbios
>> > doesn't have the right information, or (c) something else?
>> >
>> Linking with libsmbios would be a problem, and libsmbios doesn't have
>> this info anyway.
>>
>> > It doesn't *look* like this is using any information that is only available
>> > in the kernel, so in principle it seems like this could be done in
>> > user-space.
>> >
>>
>> I actually just got an email from someone who needs to determine the
>> card slot number in their driver... so I've added an external callable
>> 'pci_get_smbios_slot' function to enable this.
>>
>> >> We already report the index for embedded devices in
>> >> pci-label.c, this code should have gone in at the same time.
>> >>
>> >> For example.  The SMBIOS entry for slot 3 is 40:00.0 There is a
>> >> quad-port NIC in the slot with a bridge at 40:00.0
>> >>
>> >> 42:00.0 Bridge (sec=43, sub=45)
>> >> 43:02.0 Bridge (sec=44, sub=44)
>> >> 43:04.0 Bridge (sec=45, sub=45)
>> >> 44:00.0 Ethernet
>> >> 44:00.1 Ethernet
>> >> 45:00.0 Ethernet
>> >> 45:00.1 Ethernet
>> >>
>> >> So dmidecode only returns the slot number for 42:00.0 but not any
>> >> child devices.  This code will provide a 'slot' sysfs variable that
>> >> reports '3' for all devices under and including the bridge.
>> >
>> > What if the card in slot 3 is an adapter leading to an external PCI
>> > chassis?  Wouldn't we then have a 'slot' file for every card in that
>> > chassis, all containing '3'?  This sounds confusing, although it is true
>> > that they all would be connected via the system board slot 3.
>> >
>> Yes that is correct.  Unless SMBIOS had a table of the second chassis.
>>
>> > Also, we do have the /sys/bus/pci/slots/ hierarchy already.  If we do put
>> > something like this in the kernel, how would it relate to that hierarchy?
>> > Could this SMBIOS stuff be integrated into that somehow?
>> >
>>
>> /sys/bus/pci/slots only map a slot to a single PCI device.  systemd
>> does use /sys/bus/pci/slots but it can't see the slot number on cards
>> with bridges as the actual slot number is a parent.  And that's not
>> easily to determine a parent device from the slots interface.  I'd
>> really like something generic here. I'm also looking for some method
>> for reporting bay/enclosure ID for PCIe-SSD devices.
>>
>> > We have a bit of a hodge-podge of slot names already, and I'd like to
>> > simplify things if we can.
>
> We aren't really converging here yet.
>
> We need to figure out exactly what has to be done in the kernel because it
> can't be done in user-space.  So far, your patch *looks* like it could (in
> principle) be done in user-space, given a sysfs PCI hierarchy and the
> SMBIOS information.

It's a pain to do in user space as you have to read DMI information,
then traverse PCI hierarchy until you find a match.  Every utility or
code that wants to match a PCI device to a SMBIOS slot must reinvent
the wheel.  biosdevname has its own code to read the entire SMBIOS
table. It then reads in all pci devices and attempts to do a mapping.
Why do this when the kernel already has this info easily available.

>
> The goal ("determine which PCI devices belong to system slots") is
> definitely generic and useful.  We have quite a bit of slot stuff already
> in the kernel, and some is already exposed via sysfs, so if we're still
> missing what you need, it seems like the current code is off the rails
> somewhere and should be fixed.
>
> Can we explore what you need in a little more detail, with some concrete
> examples?  I heard:
>
>   - Systemd uses /sys/bus/pci/slots but gets the wrong slot information for
>     cards with bridges on them.  Can you give an example including all the
>     PCI devices involved and the related /sys/bus/pci/slots info so we can
>     see exactly what's wrong?

I created a quick and dirty module that populates /sys/bus/pci/slots:

#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/pci.h>
#include <linux/dmi.h>

static int __init kslot_init(void)
{
        const struct dmi_device *dmi;
        struct dmi_dev_onboard *dslot;
        struct pci_dev *sdev;

        dmi = NULL;
        while ((dmi = dmi_find_device(DMI_DEV_TYPE_SYSTEM_SLOT, NULL,
dmi)) != NULL) {
                dslot = dmi->device_data;
                sdev = pci_get_domain_bus_and_slot(dslot->segment, dslot->bus,
                                             dslot->devfn);
                if (!sdev)
                        continue;
                pci_create_slot(sdev->bus, dslot->instance, dslot->dev.name,
                                NULL);
                pci_dev_put(sdev);
        }
        return 0;
}

static void __exit kslot_exit(void)
{
}

module_init(kslot_init);
module_exit(kslot_exit);

MODULE_AUTHOR("jordan_hargrave@...l.com");
MODULE_LICENSE("GPL");
===

# ls -l /sys/bus/pci/slots
total 0
drwxr-xr-x 2 root root 0 Jul 23 21:43 PCI1
drwxr-xr-x 2 root root 0 Jul 23 21:43 PCI2
drwxr-xr-x 2 root root 0 Jul 23 21:43 PCI3

# for X in /sys/bus/pci/slots/* ; do echo -n "$X = "; cat $X/address; done
/sys/bus/pci/slots/PCI1 = 0000:41:01
/sys/bus/pci/slots/PCI2 = 0000:42:02
/sys/bus/pci/slots/PCI3 = 0000:04:03

So one problem with this already is the string that gets output. It's
<domain>:<bus>:<slot>   so you still have the problem of what about
bus 43-45 that are children of bus 42.

So wrote a script:
#!/usr/bin/bash
for X in /sys/bus/pci/devices/* ; do
    RX=$(readlink -f $X)
    for Z in /sys/bus/pci/slots/* ; do
        NAME=$(basename $Z)
        ADDR=$(cat $Z/address | cut -f1-2 -d:)
        if [[ $RX =~ $ADDR ]] ; then
            echo $RX,$NAME
        fi
    done
done

for each PCI device you have to iterate all devices in
/sys/bus/pci/slots, and munge the 'address' to actually find a match

output is:
/sys/devices/pci0000:00/0000:00:03.0/0000:04:00.0,PCI3 [SMBIOS entry]
/sys/devices/pci0000:00/0000:00:03.0/0000:04:00.1,PCI3
/sys/devices/pci0000:00/0000:00:03.0/0000:04:00.2,PCI3
/sys/devices/pci0000:00/0000:00:03.0/0000:04:00.3,PCI3
/sys/devices/pci0000:00/0000:00:03.0/0000:04:00.4,PCI3
/sys/devices/pci0000:00/0000:00:03.0/0000:04:00.5,PCI3
/sys/devices/pci0000:00/0000:00:03.0/0000:04:00.6,PCI3
/sys/devices/pci0000:00/0000:00:03.0/0000:04:00.7,PCI3
/sys/devices/pci0000:40/0000:40:01.0/0000:41:00.0,PCI1 [SMBIOS entry]
/sys/devices/pci0000:40/0000:40:03.0/0000:42:00.0,PCI2 [SMBIOS entry]
/sys/devices/pci0000:40/0000:40:03.0/0000:42:00.0/0000:43:02.0,PCI2
/sys/devices/pci0000:40/0000:40:03.0/0000:42:00.0/0000:43:04.0,PCI2
/sys/devices/pci0000:40/0000:40:03.0/0000:42:00.0/0000:43:02.0/0000:44:00.0,PCI2
/sys/devices/pci0000:40/0000:40:03.0/0000:42:00.0/0000:43:02.0/0000:44:00.1,PCI2
/sys/devices/pci0000:40/0000:40:03.0/0000:42:00.0/0000:43:04.0/0000:45:00.0,PCI2
/sys/devices/pci0000:40/0000:40:03.0/0000:42:00.0/0000:43:04.0/0000:45:00.1,PCI2

And I'm not totally convinced this will work on some weird
configurations I've seen of SMBIOS entries.  as in the B:D:F must
match, not just the bus number.

>
>   - A driver needs the card slot number.  A slot number seems like a user
>     interface thing, so I'm surprised that a driver would be concerned with
>     it.  And of course, SMBIOS is an arch-specific thing, so the driver
>     would have to be able to get along without the slot number anyway.
>
pci_get_smbios_slot would just return -ENODEV or something on systems
without SMBIOS.

>   - PCIe-SSD bay/enclosure IDs.  Where would these IDs come from, and what
>     sort of reporting mechanism are you looking for?  How are these
>     structured?  Is there a separate PCIe device for every bay/enclosure
>     ID, so there would be a 1:1 mapping from PCIe device to ID?
>
On our system the enclosure IDs actually come from IPMI commands.
However you must query every PCI Bus number in the system to get the
correct enclosure mapping.  Currently yes there is a 1:1 mapping.

> Bjorn
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists