Re: [PATCH] pci: check bus pointer before dereference

[Philippe Mathieu-Daudé]

+Igor

On 9/15/20 3:51 PM, Li Qiang wrote:
> P J P <ppandit@redhat.com> 于2020年8月27日周四 下午7:52写道：
>>
>> From: Prasad J Pandit <pjp@fedoraproject.org>
>>
>> While mapping IRQ level in pci_change_irq_level() routine,
>> it does not check if pci_get_bus() returned a valid pointer.
>> It may lead to a NULL pointer dereference issue. Add check to
>> avoid it.
>>
>>  -> https://ruhr-uni-bochum.sciebo.de/s/NNWP2GfwzYKeKwE?path=%2Flsi_nullptr1
>>     ==1183858==Hint: address points to the zero page.
>>     #0 pci_change_irq_level hw/pci/pci.c:259
>>     #1 pci_irq_handler hw/pci/pci.c:1445
>>     #2 pci_set_irq hw/pci/pci.c:1463
>>     #3 lsi_set_irq hw/scsi/lsi53c895a.c:488
>>     #4 lsi_update_irq hw/scsi/lsi53c895a.c:523
>>     #5 lsi_script_scsi_interrupt hw/scsi/lsi53c895a.c:554
>>     #6 lsi_execute_script hw/scsi/lsi53c895a.c:1149
>>     #7 lsi_reg_writeb hw/scsi/lsi53c895a.c:1984
>>     #8 lsi_io_write hw/scsi/lsi53c895a.c:2146
>>     ...
>>
>> Reported-by: Ruhr-University <bugs-syssec@rub.de>
>> Signed-off-by: Prasad J Pandit <pjp@fedoraproject.org>
>> ---
>>  hw/pci/pci.c | 3 +++
>>  1 file changed, 3 insertions(+)
>>
>> diff --git a/hw/pci/pci.c b/hw/pci/pci.c
>> index de0fae10ab..df5a2c3294 100644
>> --- a/hw/pci/pci.c
>> +++ b/hw/pci/pci.c
>> @@ -253,6 +253,9 @@ static void pci_change_irq_level(PCIDevice *pci_dev, int 
>> irq_num, int change)
>>      PCIBus *bus;
>>      for (;;) {
>>          bus = pci_get_bus(pci_dev);
>> +        if (!bus) {
> 
> Hi Prasad,
> 
> I think in normal this 'bus' will be not NULL.
> I have look at the link in the commit msg.
> I find it is another DMA to MMIO issue which we have discussed a lot
> but didn't come up with an
> satisfying solution.
> 
> Maybe we can try to the DMA to MMIO issue direction.
> CC: Peter, Jason and Alex
> 
> Thanks,
> Li Qiang
> 
> 
>> +            return;

Nack, this should be an abort().

As usual, question is how we got here.

As Li said, it is another DMA to MMIO bug class.

lsi_execute_script
 -> address_space_write
    -> acpi_pcihp_eject_slot
       -> bus_remove_child

So at this point the PCI device is still MMIO-mapped
but eject from the bus... ???
Then IRQ is triggered, which the device wants to
propagate via its PCI bus but it doesn't have any more
and b00m.

If a device is hotpluggable, who is responsible to
unmap its regions?

>> +        }
>>          irq_num = bus->map_irq(pci_dev, irq_num);
>>          if (bus->set_irq)
>>              break;
>> --
>> 2.26.2
>>
>>
>