Re: [PATCH v2] PCI: Fix crash during pci_dev hot-unplug on pseries KVM guest

[Lizhi Hou <lizhi.hou@amd.com>]

On 7/23/24 12:54, Rob Herring wrote:
> On Tue, Jul 23, 2024 at 12:21 PM Lizhi Hou <lizhi.hou@amd.com> wrote:
>>
>> On 7/23/24 09:21, Rob Herring wrote:
>>> On Mon, Jul 15, 2024 at 01:52:30PM -0700, Lizhi Hou wrote:
>>>> On 7/15/24 11:55, Rob Herring wrote:
>>>>> On Mon, Jul 15, 2024 at 2:08 AM Amit Machhiwal <amachhiw@linux.ibm.com> wrote:
>>>>>> With CONFIG_PCI_DYNAMIC_OF_NODES [1], a hot-plug and hot-unplug sequence
>>>>>> of a PCI device attached to a PCI-bridge causes following kernel Oops on
>>>>>> a pseries KVM guest:
>>>>>>
>>>>>>     RTAS: event: 2, Type: Hotplug Event (229), Severity: 1
>>>>>>     Kernel attempted to read user page (10ec00000048) - exploit attempt? (uid: 0)
>>>>>>     BUG: Unable to handle kernel data access on read at 0x10ec00000048
>>>>>>     Faulting instruction address: 0xc0000000012d8728
>>>>>>     Oops: Kernel access of bad area, sig: 11 [#1]
>>>>>>     LE PAGE_SIZE=64K MMU=Radix SMP NR_CPUS=2048 NUMA pSeries
>>>>>> <snip>
>>>>>>     NIP [c0000000012d8728] __of_changeset_entry_invert+0x10/0x1ac
>>>>>>     LR [c0000000012da7f0] __of_changeset_revert_entries+0x98/0x180
>>>>>>     Call Trace:
>>>>>>     [c00000000bcc3970] [c0000000012daa60] of_changeset_revert+0x58/0xd8
>>>>>>     [c00000000bcc39c0] [c000000000d0ed78] of_pci_remove_node+0x74/0xb0
>>>>>>     [c00000000bcc39f0] [c000000000cdcfe0] pci_stop_bus_device+0xf4/0x138
>>>>>>     [c00000000bcc3a30] [c000000000cdd140] pci_stop_and_remove_bus_device_locked+0x34/0x64
>>>>>>     [c00000000bcc3a60] [c000000000cf3780] remove_store+0xf0/0x108
>>>>>>     [c00000000bcc3ab0] [c000000000e89e04] dev_attr_store+0x34/0x78
>>>>>>     [c00000000bcc3ad0] [c0000000007f8dd4] sysfs_kf_write+0x70/0xa4
>>>>>>     [c00000000bcc3af0] [c0000000007f7248] kernfs_fop_write_iter+0x1d0/0x2e0
>>>>>>     [c00000000bcc3b40] [c0000000006c9b08] vfs_write+0x27c/0x558
>>>>>>     [c00000000bcc3bf0] [c0000000006ca168] ksys_write+0x90/0x170
>>>>>>     [c00000000bcc3c40] [c000000000033248] system_call_exception+0xf8/0x290
>>>>>>     [c00000000bcc3e50] [c00000000000d05c] system_call_vectored_common+0x15c/0x2ec
>>>>>> <snip>
>>>>>>
>>>>>> A git bisect pointed this regression to be introduced via [1] that added
>>>>>> a mechanism to create device tree nodes for parent PCI bridges when a
>>>>>> PCI device is hot-plugged.
>>>>>>
>>>>>> The Oops is caused when `pci_stop_dev()` tries to remove a non-existing
>>>>>> device-tree node associated with the pci_dev that was earlier
>>>>>> hot-plugged and was attached under a pci-bridge. The PCI dev header
>>>>>> `dev->hdr_type` being 0, results a conditional check done with
>>>>>> `pci_is_bridge()` into false. Consequently, a call to
>>>>>> `of_pci_make_dev_node()` to create a device node is never made. When at
>>>>>> a later point in time, in the device node removal path, a memcpy is
>>>>>> attempted in `__of_changeset_entry_invert()`; since the device node was
>>>>>> never created, results in an Oops due to kernel read access to a bad
>>>>>> address.
>>>>>>
>>>>>> To fix this issue, the patch updates `of_changeset_create_node()` to
>>>>>> allocate a new node only when the device node doesn't exist and init it
>>>>>> in case it does already. Also, introduce `of_pci_free_node()` to be
>>>>>> called to only revert and destroy the changeset device node that was
>>>>>> created via a call to `of_changeset_create_node()`.
>>>>>>
>>>>>> [1] commit 407d1a51921e ("PCI: Create device tree node for bridge")
>>>>>>
>>>>>> Fixes: 407d1a51921e ("PCI: Create device tree node for bridge")
>>>>>> Reported-by: Kowshik Jois B S <kowsjois@linux.ibm.com>
>>>>>> Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>
>>>>>> Signed-off-by: Amit Machhiwal <amachhiw@linux.ibm.com>
>>>>>> ---
>>>>>> Changes since v1:
>>>>>>        * Included Lizhi's suggested changes on V1
>>>>>>        * Fixed below two warnings from Lizhi's changes and rearranged the cleanup
>>>>>>          part a bit in `of_pci_make_dev_node`
>>>>>>            drivers/pci/of.c:611:6: warning: no previous prototype for ‘of_pci_free_node’ [-Wmissing-prototypes]
>>>>>>              611 | void of_pci_free_node(struct device_node *np)
>>>>>>                  |      ^~~~~~~~~~~~~~~~
>>>>>>            drivers/pci/of.c: In function ‘of_pci_make_dev_node’:
>>>>>>            drivers/pci/of.c:696:1: warning: label ‘out_destroy_cset’ defined but not used [-Wunused-label]
>>>>>>              696 | out_destroy_cset:
>>>>>>                  | ^~~~~~~~~~~~~~~~
>>>>>>        * V1: https://lore.kernel.org/all/20240703141634.2974589-1-amachhiw@linux.ibm.com/
>>>>>>
>>>>>>     drivers/of/dynamic.c  | 16 ++++++++++++----
>>>>>>     drivers/of/unittest.c |  2 +-
>>>>>>     drivers/pci/bus.c     |  3 +--
>>>>>>     drivers/pci/of.c      | 39 ++++++++++++++++++++++++++-------------
>>>>>>     drivers/pci/pci.h     |  2 ++
>>>>>>     include/linux/of.h    |  1 +
>>>>>>     6 files changed, 43 insertions(+), 20 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/of/dynamic.c b/drivers/of/dynamic.c
>>>>>> index dda6092e6d3a..9bba5e82a384 100644
>>>>>> --- a/drivers/of/dynamic.c
>>>>>> +++ b/drivers/of/dynamic.c
>>>>>> @@ -492,21 +492,29 @@ struct device_node *__of_node_dup(const struct device_node *np,
>>>>>>      * a given changeset.
>>>>>>      *
>>>>>>      * @ocs: Pointer to changeset
>>>>>> + * @np: Pointer to device node. If null, allocate a new node. If not, init an
>>>>>> + *     existing one.
>>>>>>      * @parent: Pointer to parent device node
>>>>>>      * @full_name: Node full name
>>>>>>      *
>>>>>>      * Return: Pointer to the created device node or NULL in case of an error.
>>>>>>      */
>>>>>>     struct device_node *of_changeset_create_node(struct of_changeset *ocs,
>>>>>> +                                            struct device_node *np,
>>>>>>                                                 struct device_node *parent,
>>>>>>                                                 const char *full_name)
>>>>>>     {
>>>>>> -       struct device_node *np;
>>>>>>            int ret;
>>>>>>
>>>>>> -       np = __of_node_dup(NULL, full_name);
>>>>>> -       if (!np)
>>>>>> -               return NULL;
>>>>>> +       if (!np) {
>>>>>> +               np = __of_node_dup(NULL, full_name);
>>>>>> +               if (!np)
>>>>>> +                       return NULL;
>>>>>> +       } else {
>>>>>> +               of_node_set_flag(np, OF_DYNAMIC);
>>>>>> +               of_node_set_flag(np, OF_DETACHED);
>>>>> Are we going to rename the function to
>>>>> of_changeset_create_or_maybe_modify_node()? No. The functions here are
>>>>> very clear in that they allocate new objects and don't reuse what's
>>>>> passed in.
>>>> Ok. How about keeping of_changeset_create_node unchanged.
>>>>
>>>> Instead, call kzalloc(), of_node_init() and of_changeset_attach_node()
>>>>
>>>> in of_pci_make_dev_node() directly.
>>>>
>>>> A similar example is dlpar_parse_cc_node().
>>>>
>>>>
>>>> Does this sound better?
>>> No, because really that code should be re-written using of_changeset
>>> API.
>>>
>>> My suggestion is add a data pointer to struct of_changeset and then set
>>> that to something to know the data ptr is a changeset and is your
>>> changeset.
>> I do not fully understand the point. I think the issue is that we do not
>> know if a given of_node is created by of_pci_make_dev_node(), correct?
> Yes.
>
>> of_node->data can point to anything. And we do not know if it points a
>> cset or not.
> Right. But instead of checking "of_node->data == of_pci_free_node",
> you would just be checking "*(of_node->data) == of_pci_free_node"

if of_node->data is a char* pointer, it would be panic. So I used 
of_node->data == of_pci_free_node.

> (omitting a NULL check and cast for simplicity). I suppose in theory
> that could have a false match, but that could happen in this patch
> already.

I think if any other kernel code  put of_pci_free_node to of_node->data, 
it can be fixed over there.

>
>> Do you mean to add a flag (e.g. OF_PCI_DYNAMIC) to
>> indicate of_node->data points to cset?
> That would be another option, but OF_PCI_DYNAMIC would not be a good
> name because that would be a flag bit for every single caller needing
> similar functionality. Name it just what it indicates: of_node->data
> points to cset
>
> If we have that flag, then possibly the DT core can handle more
> clean-up itself like calling detach and freeing the changeset.
> Ideally, the flags should be internal to the DT code.

Sure. If you prefer this option, I will propose another fix.


Thanks,

Lizhi

>
> Rob