On Tue, 2020-06-30 at 11:33 -0500, Bjorn Helgaas wrote:
> On Tue, Jun 30, 2020 at 12:39:08PM +0300, Andy Shevchenko wrote:
> > On Mon, Jun 29, 2020 at 06:20:11PM -0500, Bjorn Helgaas wrote:
> > > On Thu, Jun 25, 2020 at 02:58:23PM -0500, Bjorn Helgaas wrote:
> > > > [+cc Thomas]
> > > >
> > > > On Thu, Jun 25, 2020 at 12:24:49PM -0400, Jon Derrick wrote:
> > > > > From: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> > > > >
> > > > > The VMD domain does not subscribe to ACPI, and so does not operate on
> > > > > it's irqdomain fwnode. It was freeing the handle after allocation of the
> > > > > domain. As of 181e9d4efaf6a (irqdomain: Make __irq_domain_add() less
> > > > > OF-dependent), the fwnode is put during irq_domain_remove causing a page
> > > > > fault. This patch keeps VMD's fwnode allocated through the lifetime of
> > > > > the VMD irqdomain.
> > > > >
> > > > > Fixes: ae904cafd59d ("PCI/vmd: Create named irq domain")
> > > > > Signed-off-by: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> > > > > Co-developed-by: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> > > > > Signed-off-by: Jon Derrick <jonathan.derrick@xxxxxxxxx>
> > > > > ---
> > > > > Hi Lorenzo, Bjorn,
> > > > >
> > > > > Please take this patch for v5.8 fixes. It fixes an issue during VMD
> > > > > unload.
> > > >
> > > > I tentatively applied this to for-linus for v5.8.
> > > >
> > > > But I would like to clarify the commit log. It says this fixes
> > > > Thomas' ae904cafd59d ("PCI/vmd: Create named irq domain"). That
> > > > appeared in v4.13, which suggests that this patch should be backported
> > > > to v4.13 and later.
> > >
> > > I didn't word this correctly. What I meant was "I will consider
> > > applying this after the commit log is clarified." Just FYI that this
> > > isn't resolved yet.
> > >
> > > Since this is proposed for v5.8, I really want to understand this
> > > better before asking Linus to pull it as a fix.
> >
> > The problem here is in the original patch which relies on the
> > knowledge that fwnode is (was) not used anyhow specifically for ACPI
> > case. That said, it makes fwnode a dangling pointer which I
> > personally consider as a mine left for others. That's why the Fixes
> > refers to the initial commit. The latter just has been blasted on
> > that mine.
>
> IIUC, you're saying this pattern:
>
> fwnode = irq_domain_alloc_named_id_fwnode(...)
> irq_domain = pci_msi_create_irq_domain(fwnode, ...)
> irq_domain_free_fwnode(fwnode)
>
> leaves a dangling fwnode pointer. That does look suspicious because
> __irq_domain_add() saves fwnode:
>
> irq_domain = pci_msi_create_irq_domain(fwnode, ...)
> msi_create_irq_domain(fwnode, ...)
> irq_domain_create_hierarchy(..., fwnode, ...)
> irq_domain_create_linear(fwnode, ...)
> __irq_domain_add(fwnode, ...)
> domain->fwnode = fwnode
>
> and irq_domain_free_fwnode() frees it. But I'm confused because there
> are several other instances of this pattern:
>
> bridge_probe() # arch/mips/pci/pci-xtalk-bridge.c
> mp_irqdomain_create()
> arch_init_msi_domain()
> arch_create_remap_msi_irq_domain()
> dmar_get_irq_domain()
> hpet_create_irq_domain()
> ...
>
> Are they all wrong? I definitely think it's a bad idea to keep a copy
> of a pointer after we free the data it points to. But if they're all
> wrong, I don't want to fix just one and leave all the others.
>
> Thomas, can you enlighten us?
>
I see that struct irqchip_fwid contains the actual fwnode structure and
when that is freed, it's causes the issue.
I'm noticing in each caller of irq_domain_free_fwnode, we have the
domain itself available. It seems like it should be up to the caller to
deal with the fwnode pointer, but we could just do:
diff --git a/kernel/irq/irqdomain.c b/kernel/irq/irqdomain.c
index a4c2c915511d..61f0070285ff 100644
--- a/kernel/irq/irqdomain.c
+++ b/kernel/irq/irqdomain.c
@@ -101,7 +101,7 @@ EXPORT_SYMBOL_GPL(__irq_domain_alloc_fwnode);
*
* Free a fwnode_handle allocated with irq_domain_alloc_fwnode.
*/
-void irq_domain_free_fwnode(struct fwnode_handle *fwnode)
+void irq_domain_free_fwnode(struct irq_domain *domain, struct
fwnode_handle *fwnode)
{
struct irqchip_fwid *fwid;
@@ -109,6 +109,7 @@ void irq_domain_free_fwnode(struct fwnode_handle
*fwnode)
return;
fwid = container_of(fwnode, struct irqchip_fwid, fwnode);
+ domain->fwnode = NULL;
kfree(fwid->name);
kfree(fwid);
}
> > If you think that's fine to have such trick, feel free to update Fixes tag.
> >
> > > > But it's not clear to me that ae904cafd59d is actually broken, since
> > > > the log also says the problem appeared after 181e9d4efaf6 ("irqdomain:
> > > > Make __irq_domain_add() less OF-dependent"), which we just merged for
> > > > v5.8-rc1.
> > > >
> > > > And obviously, freeing the fwnode doesn't *cause* a page fault. A
> > > > use-after-free might cause a fault, but it's not clear where that
> > > > happens. I guess fwnode is used in the interval between:
> > > >
> > > > vmd_enable_domain
> > > > pci_msi_create_irq_domain
> > > >
> > > > ... <-- fwnode used here somewhere
> > > >
> > > > vmd_remove
> > > > vmd_cleanup_srcu
> > > > irq_domain_free_fwnode
> > > >
> > > > But I can't tell how 181e9d4efaf6a and/or ae904cafd59d are related to
> > > > that.
> > > >
> > > > > drivers/pci/controller/vmd.c | 8 ++++++--
> > > > > 1 file changed, 6 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/drivers/pci/controller/vmd.c b/drivers/pci/controller/vmd.c
> > > > > index e386d4eac407..ebec0a6e77ed 100644
> > > > > --- a/drivers/pci/controller/vmd.c
> > > > > +++ b/drivers/pci/controller/vmd.c
> > > > > @@ -546,9 +546,10 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
> > > > >
> > > > > vmd->irq_domain = pci_msi_create_irq_domain(fn, &vmd_msi_domain_info,
> > > > > x86_vector_domain);
> > > > > - irq_domain_free_fwnode(fn);
> > > > > - if (!vmd->irq_domain)
> > > > > + if (!vmd->irq_domain) {
> > > > > + irq_domain_free_fwnode(fn);
> > > > > return -ENODEV;
> > > > > + }
> > > > >
> > > > > pci_add_resource(&resources, &vmd->resources[0]);
> > > > > pci_add_resource_offset(&resources, &vmd->resources[1], offset[0]);
> > > > > @@ -559,6 +560,7 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
> > > > > if (!vmd->bus) {
> > > > > pci_free_resource_list(&resources);
> > > > > irq_domain_remove(vmd->irq_domain);
> > > > > + irq_domain_free_fwnode(fn);
> > > > > return -ENODEV;
> > > > > }
> > > > >
> > > > > @@ -672,6 +674,7 @@ static void vmd_cleanup_srcu(struct vmd_dev *vmd)
> > > > > static void vmd_remove(struct pci_dev *dev)
> > > > > {
> > > > > struct vmd_dev *vmd = pci_get_drvdata(dev);
> > > > > + struct fwnode_handle *fn = vmd->irq_domain->fwnode;
> > > > >
> > > > > sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
> > > > > pci_stop_root_bus(vmd->bus);
> > > > > @@ -679,6 +682,7 @@ static void vmd_remove(struct pci_dev *dev)
> > > > > vmd_cleanup_srcu(vmd);
> > > > > vmd_detach_resources(vmd);
> > > > > irq_domain_remove(vmd->irq_domain);
> > > > > + irq_domain_free_fwnode(fn);
> > > > > }
> > > > >
> > > > > #ifdef CONFIG_PM_SLEEP
> > > > > --
> > > > > 2.18.1
> > > > >
> >
> > --
> > With Best Regards,
> > Andy Shevchenko
> >
> >
