Re: [PATCH] PCI / ACPI: Do not set ACPI companions for host bridges with parents

From: Boris Ostrovsky
Date: Tue May 26 2015 - 22:16:24 EST




On 05/26/2015 08:07 PM, Rafael J. Wysocki wrote:
On Tuesday, May 26, 2015 10:32:03 AM Boris Ostrovsky wrote:
On 05/25/2015 10:17 PM, Rafael J. Wysocki wrote:
On Tuesday, May 26, 2015 03:08:17 AM Rafael J. Wysocki wrote:
On Tuesday, May 26, 2015 01:42:16 AM Rafael J. Wysocki wrote:
On Tuesday, May 26, 2015 01:22:12 AM Rafael J. Wysocki wrote:
On Friday, May 22, 2015 09:53:37 PM Boris Ostrovsky wrote:
On 05/22/2015 04:11 AM, Sander Eikelenboom wrote:
Hello Sander,

[cut]

(+Rafael again)

So the immediate cause of those errors is that pdev->evtchn is 0.
Backend is not notified and things not go well then.

And it is indeed caused by 97badf873ab60e841243b66133ff9eff2a46ef29:

We allocate pcifront_sd in pcifront_scan_root() and then pass it to
pci_scan_bus_parented() as sysdata. Eventually this sysdata is used in
pcibios_root_bridge_prepare() as pci_sysdata. It is dereferenced as
pci_sysdata->companion (which I believe is aliased to pcifront_sd->pdev)
Well, there is an int node field between them, so I'm not sure.

and then set_primary_fwnode() writes it, thus corrupting
pcifront_sd->pdev (and I think this is what sets evtchn to zero).
So the corruption happens when set_primary_fwnode() writes NULL to the
'secondary' field of object pointed to by 'fwnode'.

This isn't strictly necessary and we might avoid the crash by only
writing to fwnode->secondary if fn is not NULL.

So, Sander please test the patch below too if possible.

Of course, that doesn't solve a problem of passing an incorrect pointer
to ACPI_COMPANION_SET() in pcibios_root_bridge_prepare().
And here's one more thing to test.
And the below is how I'd fix it, so you can simply test this patch and skip the
previous ones.

---
From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
Subject: PCI / ACPI: Do not set ACPI companions for host bridges with parents

Commit 97badf873ab6 (device property: Make it possible to use
secondary firmware nodes) uncovered a bug in the x86 (and ia64) PCI
host bridge initialization code that assumes bridge->bus->sysdata
to always point to a struct pci_sysdata object which need not be
the case (in particular, the Xen PCI frontend driver sets it to point
to a different data type). If it is not the case, an incorrect
pointer (or a piece of data that is not a pointer at all) will be
passed to ACPI_COMPANION_SET() and that may cause interesting
breakage to happen going forward.

To work around this problem use the observation that the ACPI
host bridge initialization always passes NULL as parent to
pci_create_root_bus(), so if pcibios_root_bridge_prepare() sees
a non-NULL parent of the bridge, it should not attempt to set
an ACPI companion for it, because that means that
pci_create_root_bus() has been called by someone else.

I am wondering whether at some point we might want to use companion
information in Xen as well.

Another option might be to require that whoever creates their own
sysdata structure to have pci_sysdata as its first element, e.g.
Well, I was thinking about that, but it would be x86-specific and I believe
that Xen is supposed to work on other architectures too (or at least will be
in the future).

--- a/drivers/pci/xen-pcifront.c
+++ b/drivers/pci/xen-pcifront.c
@@ -52,7 +52,12 @@ struct pcifront_device {
};

struct pcifront_sd {
+ /*
+ * Must be the first element of this structure since pcifront_sd
+ * is assigned to bus' sysdata and may later be dereferenced as
+ * pci_sysdata.
+ */
+ struct pci_sysdata sd;
struct pcifront_device *pdev;
};
Also if you want to use an ACPI companion, it will be a good question *which* one.

My half-way educated guess is that will be the ACPI companion of the parent,
in which case it's better to modify pcibios_root_bridge_prepare(). In any
case, we need to talk about that beforehand, so please let me know when you
have more specific plans.

I don't have anything specific. And after looking at this code some more I am not sure pcifront will ever want to use companions since ACPI is not part of device initialization here (it's done via xenbus, which is a purely software construct).



As for 4.1 (which currently is broken for Sander), I think the $subject patch
is the cleanest and least intrusive thing we can do.

OK. Thank you for fixing this.

-boris

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