Re: [PATCH] PCI/ASPM: Update ASPM sysfs on MFD function removal to avoid use-after-free

[Bjorn Helgaas <helgaas@kernel.org>]

On Thu, Aug 01, 2024 at 08:05:23PM +0800, Jay Fang wrote:
> On 2024/8/1 5:46, Bjorn Helgaas wrote:
> > On Tue, Jul 30, 2024 at 05:57:43PM +0800, Ding Hui wrote:
> >> On 2024/7/30 9:16, Jay Fang wrote:
> >>>  From 'commit 456d8aa37d0f ("PCI/ASPM: Disable ASPM on MFD function removal
> >>> to avoid use-after-free")' we know that PCIe spec r6.0, sec 7.5.3.7,
> >>> recommends that software program the same ASPM Control(pcie_link_state)
> >>> value in all functions of multi-function devices, and free the
> >>> pcie_link_state when any child function is removed.
> >>>
> >>> However, ASPM Control sysfs is still visible to other children even if it
> >>> has been removed by any child function, and careless use it will
> >>> trigger use-after-free error, e.g.:
> >>>
> >>>    # lspci -tv
> >>>      -[0000:16]---00.0-[17]--+-00.0  Device 19e5:0222
> >>>                              \-00.1  Device 19e5:0222
> >>>    # echo 1 > /sys/bus/pci/devices/0000:17:00.0/remove       // pcie_link_state will be released
> >>>    # echo 1 > /sys/bus/pci/devices/0000:17:00.1/link/l1_aspm // will trigger error
> >>>
> >>>    Unable to handle kernel NULL pointer dereference at virtual address 0000000000000030
> >>>    Call trace:
> >>>     aspm_attr_store_common.constprop.0+0x10c/0x154
> >>>     l1_aspm_store+0x24/0x30
> >>>     dev_attr_store+0x20/0x34
> >>>     sysfs_kf_write+0x4c/0x5c
> >>>
> >>> We can solve this problem by updating the ASPM Control sysfs of all
> >>> children immediately after ASPM Control have been freed.
> >>>
> >>> Fixes: 456d8aa37d0f ("PCI/ASPM: Disable ASPM on MFD function removal to avoid use-after-free")
> >>> Signed-off-by: Jay Fang <f.fangjian@huawei.com>
> >>> ---
> >>>   drivers/pci/pcie/aspm.c | 2 ++
> >>>   1 file changed, 2 insertions(+)
> >>>
> >>> diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
> >>> index cee2365e54b8..eee9e6739924 100644
> >>> --- a/drivers/pci/pcie/aspm.c
> >>> +++ b/drivers/pci/pcie/aspm.c
> >>> @@ -1262,6 +1262,8 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)
> >>>   		pcie_config_aspm_path(parent_link);
> >>>   	}
> >>> +	pcie_aspm_update_sysfs_visibility(parent);
> >>> +
> >>
> >> To be more rigorous, is there still a race window in
> >> aspm_attr_{show,store}_common or clkpm_{show,store} before updating
> >> the visibility, we can get an old or NULL pointer by
> >> pcie_aspm_get_link()?
> > 
> > Yeah, I think we still have a problem even with this patch.
>
> If so, maybe we need a new solution to completely sovle this problem.

I think so.  The pcie_link_state struct is kind of problematic to
begin with.  It basically encodes the PCI hierarchy again, even though
the hierarchy is already completely described via struct pci_dev.

IMO only the ASPM and clock PM state is really new information.  I'm
not convinced that we even need all of that (how can
supported/enabled/capable/default/disabled *all* be useful and
understandable?).  But even if we *do* need all of that, it's only 39
bits of information per device.

> > For one thing, aspm_attr_store_common() captures the pointer from
> > pcie_aspm_get_link() before the critical section, so by the time it
> > *uses* the pointer, pcie_aspm_exit_link_state() may have freed the
> > link state.
> > 
> > And there are several other callers of pcie_aspm_get_link() that
> > either call it before a critical section or don't have a critical
> > section at all.
> > 
> > I think it may be a mistake to alloc/free the link state separately
> > from the pci_dev itself.
> > 
> >>>   	mutex_unlock(&aspm_lock);
> >>>   	up_read(&pci_bus_sem);
> >>>   }
> >>
> >> -- 
> >> Thanks,
> >> - Ding Hui
> >>
> > 
> > .
> > 
>