lkml.org 
[lkml]   [2021]   [Jul]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v1 3/7] PCI: New Primary to Sideband (P2SB) bridge support library
On Thu, Apr 01, 2021 at 01:44:46PM -0500, Bjorn Helgaas wrote:
> On Thu, Apr 01, 2021 at 09:23:04PM +0300, Andy Shevchenko wrote:
> > On Thu, Apr 01, 2021 at 11:42:56AM -0500, Bjorn Helgaas wrote:
> > > On Thu, Apr 01, 2021 at 06:45:02PM +0300, Andy Shevchenko wrote:
> > > > On Tue, Mar 09, 2021 at 09:42:52AM +0100, Henning Schild wrote:
> > > > > Am Mon, 8 Mar 2021 19:42:21 -0600
> > > > > schrieb Bjorn Helgaas <helgaas@kernel.org>:
> > > > > > On Mon, Mar 08, 2021 at 09:16:50PM +0200, Andy Shevchenko wrote:
> > > > > > > On Mon, Mar 08, 2021 at 12:52:12PM -0600, Bjorn Helgaas wrote:
> > > > > > > > On Mon, Mar 08, 2021 at 02:20:16PM +0200, Andy Shevchenko wrote:
> > > >
> > > > ...
> > > >
> > > > > > > > > + /* Read the first BAR of the device in question */
> > > > > > > > > + __pci_bus_read_base(bus, devfn, pci_bar_unknown, mem,
> > > > > > > > > PCI_BASE_ADDRESS_0, true);
> > > > > > > >
> > > > > > > > I don't get this. Apparently this normally hidden device is
> > > > > > > > consuming PCI address space. The PCI core needs to know
> > > > > > > > about this. If it doesn't, the PCI core may assign this
> > > > > > > > space to another device.
> > > > > > >
> > > > > > > Right, it returns all 1:s to any request so PCI core *thinks*
> > > > > > > it's plugged off (like D3cold or so).
> > > > > >
> > > > > > I'm asking about the MMIO address space. The BAR is a register
> > > > > > in config space. AFAICT, clearing P2SBC_HIDE_BYTE makes that
> > > > > > BAR visible. The BAR describes a region of PCI address space.
> > > > > > It looks like setting P2SBC_HIDE_BIT makes the BAR disappear
> > > > > > from config space, but it sounds like the PCI address space
> > > > > > *described* by the BAR is still claimed by the device. If the
> > > > > > device didn't respond to that MMIO space, you would have no
> > > > > > reason to read the BAR at all.
> > > > > >
> > > > > > So what keeps the PCI core from assigning that MMIO space to
> > > > > > another device?
> > > > >
> > > > > The device will respond to MMIO while being hidden. I am afraid
> > > > > nothing stops a collision, except for the assumption that the BIOS
> > > > > is always right and PCI devices never get remapped. But just
> > > > > guessing here.
> > > > >
> > > > > I have seen devices with coreboot having the P2SB visible, and
> > > > > most likely relocatable. Making it visible in Linux and not hiding
> > > > > it again might work, but probably only as long as Linux will not
> > > > > relocate it. Which i am afraid might seriously upset the BIOS,
> > > > > depending on what a device does with those GPIOs and which parts
> > > > > are implemented in the BIOS.
> > > >
> > > > So the question is, do we have knobs in PCI core to mark device
> > > > fixes in terms of BARs, no relocation must be applied, no other
> > > > devices must have the region?
> > >
> > > I think the closest thing is the IORESOURCE_PCI_FIXED bit that we use
> > > for things that must not be moved. Generally PCI resources are
> > > associated with a pci_dev, and we set IORESOURCE_PCI_FIXED for BARs,
> > > e.g., dev->resource[n]. We do that for IDE legacy regions (see
> > > LEGACY_IO_RESOURCE), Langwell devices (pci_fixed_bar_fixup()),
> > > "enhanced allocation" (pci_ea_flags()), and some quirks (quirk_io()).
> > >
> > > In your case, the device is hidden so it doesn't respond to config
> > > accesses, so there is no pci_dev for it.
> >
> > Yes, and the idea is to unhide it on the early stage.
> > Would it be possible to quirk it to fix the IO resources?
>
> If I read your current patch right, it unhides the device, reads the
> BAR, then hides the device again. I didn't see that it would create a
> pci_dev for it.
>
> If you unhide it and then enumerate it normally (and mark the BAR as
> IORESOURCE_PCI_FIXED to make sure we never move it), that might work.
> Then there should be a pci_dev for it, and it would then show up in
> sysfs, lspci, etc. And we should insert the BAR in iomem_resource, so
> we should see it in /proc/iomem and we won't accidentally put
> something else on top of it.

If the PCI device is present and we have ACPI description for the one or more
devices (currently pin control), wouldn't be a conflicting resources issue?

When would be the suitable place to avoid that?

> > > resource, fills it in, sets IORESOURCE_PCI_FIXED, and does something
> > > similar to pci_claim_resource()?

--
With Best Regards,
Andy Shevchenko


\
 
 \ /
  Last update: 2021-07-12 14:16    [W:0.128 / U:0.688 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site