[EXTERNAL] Re: [PATCH v3 1/3] PCI: Introduce pcibios_ignore_alignment_request

[Shawn Anastasio]

On 5/30/19 1:55 AM, Sam Bobroff wrote:
> On Tue, May 28, 2019 at 03:36:34PM +1000, Oliver wrote:
>> On Tue, May 28, 2019 at 2:03 PM Shawn Anastasio <shawn at anastas.io> wrote:
>>>
>>> Introduce a new pcibios function pcibios_ignore_alignment_request
>>> which allows the PCI core to defer to platform-specific code to
>>> determine whether or not to ignore alignment requests for PCI resources.
>>>
>>> The existing behavior is to simply ignore alignment requests when
>>> PCI_PROBE_ONLY is set. This is behavior is maintained by the
>>> default implementation of pcibios_ignore_alignment_request.
>>>
>>> Signed-off-by: Shawn Anastasio <shawn at anastas.io>
>>> ---
>>>   drivers/pci/pci.c   | 9 +++++++--
>>>   include/linux/pci.h | 1 +
>>>   2 files changed, 8 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
>>> index 8abc843b1615..8207a09085d1 100644
>>> --- a/drivers/pci/pci.c
>>> +++ b/drivers/pci/pci.c
>>> @@ -5882,6 +5882,11 @@ resource_size_t __weak pcibios_default_alignment(void)
>>>          return 0;
>>>   }
>>>
>>> +int __weak pcibios_ignore_alignment_request(void)
>>> +{
>>> +       return pci_has_flag(PCI_PROBE_ONLY);
>>> +}
>>> +
>>>   #define RESOURCE_ALIGNMENT_PARAM_SIZE COMMAND_LINE_SIZE
>>>   static char resource_alignment_param[RESOURCE_ALIGNMENT_PARAM_SIZE] = {0};
>>>   static DEFINE_SPINLOCK(resource_alignment_lock);
>>> @@ -5906,9 +5911,9 @@ static resource_size_t pci_specified_resource_alignment(struct pci_dev *dev,
>>>          p = resource_alignment_param;
>>>          if (!*p && !align)
>>>                  goto out;
>>> -       if (pci_has_flag(PCI_PROBE_ONLY)) {
>>> +       if (pcibios_ignore_alignment_request()) {
>>>                  align = 0;
>>> -               pr_info_once("PCI: Ignoring requested alignments (PCI_PROBE_ONLY)\n");
>>> +               pr_info_once("PCI: Ignoring requested alignments\n");
>>>                  goto out;
>>>          }
>>
>> I think the logic here is questionable to begin with. If the user has
>> explicitly requested re-aligning a resource via the command line then
>> we should probably do it even if PCI_PROBE_ONLY is set. When it breaks
>> they get to keep the pieces.
>>
>> That said, the real issue here is that PCI_PROBE_ONLY probably
>> shouldn't be set under qemu/kvm. Under the other hypervisor (PowerVM)
>> hotplugged devices are configured by firmware before it's passed to
>> the guest and we need to keep the FW assignments otherwise things
>> break. QEMU however doesn't do any BAR assignments and relies on that
>> being handled by the guest. At boot time this is done by SLOF, but
>> Linux only keeps SLOF around until it's extracted the device-tree.
>> Once that's done SLOF gets blown away and the kernel needs to do it's
>> own BAR assignments. I'm guessing there's a hack in there to make it
>> work today, but it's a little surprising that it works at all...
>>
>> IIRC Sam Bobroff was looking at hotplug under pseries recently so he
>> might have something to add. He's sick at the moment, but I'll ask him
>> to take a look at this once he's back among the living
> 
> There seems to be some code already in the kernel that will disable
> PCI_PROBE_ONLY based on a device tree property, so I did a quick test
> today and it seems to work. Only a trivial tweak is needed in QEMU to
> do it (have spapr_dt_chosen() add a node called "linux,pci-probe-only"
> with a value of 0), and that would allow us to set it only for QEMU (and
> not PowerVM) if that's what we want to do. Is that useful?
> 
> (I haven't done any real testing yet but the guest booted up OK.)

It was my understanding that PCI_PROBE_ONLY should actually be set
initially so that Linux uses SLOF's BAR assignments. The issue here
is that PCI_PROBE_ONLY shouldn't be honored after initial bringup
on KVM so that hotplugged PCI devices can have custom BAR alignments.

Of course, if there's no need to honor SLOF's initial assignments,
I assume disabling PCI_PROBE_ONLY would work fine. In fact, I'm
not entirely sure why it's done in the first place. Does anybody
know?

If there is actually a valid reason for preserving SLOF's initial
assignments, then it seems like the correct solution is to disable
PCI_PROBE_ONLY after initial PCI bringup or ignore it in
pci_specified_resource_alignment() like I do in this patch set.

Bjorn Helgaas also suggested marking individual resources provided
by SLOF/PHYP with IORESOURCE_PCI_FIXED which would remove the need
to use PCI_PROBE_ONLY altogether.

Any thoughts?

- Shawn