Re: [PATCH] PCI/MSI: fix page fault when msi_populate_sysfs() failed

["wanghai (M)" <wanghai38@huawei.com>]

在 2021/10/12 1:11, Bjorn Helgaas 写道:
> For v2, please note "git log --oneline drivers/pci/msi.c" and make
> your patch follow the style, including capitalization.
>
> On Mon, Oct 11, 2021 at 05:15:28PM +0800, wanghai (M) wrote:
>> 在 2021/10/11 16:52, Barry Song 写道:
>>> On Mon, Oct 11, 2021 at 9:24 PM Wang Hai <wanghai38@huawei.com> wrote:
>>>> I got a page fault report when doing fault injection test:
> When you send v2, can you include information about how you injected
> the fault?  If it's easy, others can reproduce the failure that way.
Sorry, the reproduction needs to be based on the fault injection framework
provided by Hulk Robot. I don't know how the framework is implemented.

The way to reproduce this is to do a fault injection to make
'msi_attrs = kcalloc() in msi_populate_sysfs()' fail when insmod
9pnet_virtio.ko.

I sent v2 yesterday, can you help review it?
https://lore.kernel.org/linux-pci/20211011130837.766323-1-wanghai38@huawei.com/
>>>> BUG: unable to handle page fault for address: fffffffffffffff4
>>>> ...
>>>> RIP: 0010:sysfs_remove_groups+0x25/0x60
>>>> ...
>>>> Call Trace:
>>>>    msi_destroy_sysfs+0x30/0xa0
>>>>    free_msi_irqs+0x11d/0x1b0
>>>>    __pci_enable_msix_range+0x67f/0x760
>>>>    pci_alloc_irq_vectors_affinity+0xe7/0x170
>>>>    vp_find_vqs_msix+0x129/0x560
>>>>    vp_find_vqs+0x52/0x230
>>>>    vp_modern_find_vqs+0x47/0xb0
>>>>    p9_virtio_probe+0xa1/0x460 [9pnet_virtio]
>>>>    virtio_dev_probe+0x1ed/0x2e0
>>>>    really_probe+0x1c7/0x400
>>>>    __driver_probe_device+0xa4/0x120
>>>>    driver_probe_device+0x32/0xe0
>>>>    __driver_attach+0xbf/0x130
>>>>    bus_for_each_dev+0xbb/0x110
>>>>    driver_attach+0x27/0x30
>>>>    bus_add_driver+0x1d9/0x270
>>>>    driver_register+0xa9/0x180
>>>>    register_virtio_driver+0x31/0x50
>>>>    p9_virtio_init+0x3c/0x1000 [9pnet_virtio]
>>>>    do_one_initcall+0x7b/0x380
>>>>    do_init_module+0x5f/0x21e
>>>>    load_module+0x265c/0x2c60
>>>>    __do_sys_finit_module+0xb0/0xf0
>>>>    __x64_sys_finit_module+0x1a/0x20
>>>>    do_syscall_64+0x34/0xb0
>>>>    entry_SYSCALL_64_after_hwframe+0x44/0xae
>>>>
>>>> When populating msi_irqs sysfs failed in msi_capability_init() or
>>>> msix_capability_init(), dev->msi_irq_groups will point to ERR_PTR(...).
>>>> This will cause a page fault when destroying the wrong
>>>> dev->msi_irq_groups in free_msi_irqs().
>>>>
>>>> Fix this by setting dev->msi_irq_groups to NULL when msi_populate_sysfs()
>>>> failed.
>>>>
>>>> Fixes: 2f170814bdd2 ("genirq/msi: Move MSI sysfs handling from PCI to MSI core")
>>>> Reported-by: Hulk Robot <hulkci@huawei.com>
> What exactly was reported by the Hulk Robot?  Did it really do the
> fault injection and report the page fault?
Yes, it reported the error and provided a way to reproduce it
>>>> Signed-off-by: Wang Hai <wanghai38@huawei.com>
>>> Acked-by: Barry Song <song.bao.hua@hisilicon.com>
>>>
>>>> ---
>>>>    drivers/pci/msi.c | 2 ++
>>>>    1 file changed, 2 insertions(+)
>>>>
>>>> diff --git a/drivers/pci/msi.c b/drivers/pci/msi.c
>>>> index 0099a00af361..6f75db9f3be7 100644
>>>> --- a/drivers/pci/msi.c
>>>> +++ b/drivers/pci/msi.c
>>>> @@ -561,6 +561,7 @@ static int msi_capability_init(struct pci_dev *dev, int nvec,
>>>>           dev->msi_irq_groups = msi_populate_sysfs(&dev->dev);
>>>>           if (IS_ERR(dev->msi_irq_groups)) {
>>>>                   ret = PTR_ERR(dev->msi_irq_groups);
>>>> +               dev->msi_irq_groups = NULL;
>>>>                   goto err;
>>>>           }
>>>>
>>>> @@ -733,6 +734,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
>>>>           dev->msi_irq_groups = msi_populate_sysfs(&dev->dev);
>>>>           if (IS_ERR(dev->msi_irq_groups)) {
>>>>                   ret = PTR_ERR(dev->msi_irq_groups);
>>>> +               dev->msi_irq_groups = NULL;
>>> Can you define a temp variable and assign it to dev->msi_irq_groups if
>>> the temp variable
>>> is not PTR_ERR?
>> Of course, I will send a v2 patch.
>>>>                   goto out_free;
>>>>           }
>>>>
>>>> --
>>>> 2.17.1
>>>>
>>> Thanks
>>> Barry
>>> .
>>>
>> -- 
>> Wang Hai
>>
> .
>
-- 
Wang Hai