Re: [Qemu-devel] [RFC/BUG] xen-mapcache: buggy invalidate map cache?

["Herongguang (Stephen)" <herongguang.he@huawei.com>]

On 2017/4/12 6:32, Stefano Stabellini wrote:
> On Tue, 11 Apr 2017, hrg wrote:
>> On Tue, Apr 11, 2017 at 3:50 AM, Stefano Stabellini
>> <sstabellini@kernel.org> wrote:
>>> On Mon, 10 Apr 2017, Stefano Stabellini wrote:
>>>> On Mon, 10 Apr 2017, hrg wrote:
>>>>> On Sun, Apr 9, 2017 at 11:55 PM, hrg <hrgstephen@gmail.com> wrote:
>>>>>> On Sun, Apr 9, 2017 at 11:52 PM, hrg <hrgstephen@gmail.com> wrote:
>>>>>>> Hi,
>>>>>>>
>>>>>>> In xen_map_cache_unlocked(), map to guest memory maybe in entry->next
>>>>>>> instead of first level entry (if map to rom other than guest memory
>>>>>>> comes first), while in xen_invalidate_map_cache(), when VM ballooned
>>>>>>> out memory, qemu did not invalidate cache entries in linked
>>>>>>> list(entry->next), so when VM balloon back in memory, gfns probably
>>>>>>> mapped to different mfns, thus if guest asks device to DMA to these
>>>>>>> GPA, qemu may DMA to stale MFNs.
>>>>>>>
>>>>>>> So I think in xen_invalidate_map_cache() linked lists should also be
>>>>>>> checked and invalidated.
>>>>>>>
>>>>>>> What’s your opinion? Is this a bug? Is my analyze correct?
>>>> Yes, you are right. We need to go through the list for each element of
>>>> the array in xen_invalidate_map_cache. Can you come up with a patch?
>>> I spoke too soon. In the regular case there should be no locked mappings
>>> when xen_invalidate_map_cache is called (see the DPRINTF warning at the
>>> beginning of the functions). Without locked mappings, there should never
>>> be more than one element in each list (see xen_map_cache_unlocked:
>>> entry->lock == true is a necessary condition to append a new entry to
>>> the list, otherwise it is just remapped).
>>>
>>> Can you confirm that what you are seeing are locked mappings
>>> when xen_invalidate_map_cache is called? To find out, enable the DPRINTK
>>> by turning it into a printf or by defininig MAPCACHE_DEBUG.
>> In fact, I think the DPRINTF above is incorrect too. In
>> pci_add_option_rom(), rtl8139 rom is locked mapped in
>> pci_add_option_rom->memory_region_get_ram_ptr (after
>> memory_region_init_ram). So actually I think we should remove the
>> DPRINTF warning as it is normal.
> Let me explain why the DPRINTF warning is there: emulated dma operations
> can involve locked mappings. Once a dma operation completes, the related
> mapping is unlocked and can be safely destroyed. But if we destroy a
> locked mapping in xen_invalidate_map_cache, while a dma is still
> ongoing, QEMU will crash. We cannot handle that case.
>
> However, the scenario you described is different. It has nothing to do
> with DMA. It looks like pci_add_option_rom calls
> memory_region_get_ram_ptr to map the rtl8139 rom. The mapping is a
> locked mapping and it is never unlocked or destroyed.
>
> It looks like "ptr" is not used after pci_add_option_rom returns. Does
> the append patch fix the problem you are seeing? For the proper fix, I
> think we probably need some sort of memory_region_unmap wrapper or maybe
> a call to address_space_unmap.

Yes, I think so, maybe this is the proper way to fix this.

>
>
> diff --git a/hw/pci/pci.c b/hw/pci/pci.c
> index e6b08e1..04f98b7 100644
> --- a/hw/pci/pci.c
> +++ b/hw/pci/pci.c
> @@ -2242,6 +2242,7 @@ static void pci_add_option_rom(PCIDevice *pdev, bool is_default_rom,
>       }
>   
>       pci_register_bar(pdev, PCI_ROM_SLOT, 0, &pdev->rom);
> +    xen_invalidate_map_cache_entry(ptr);
>   }
>   
>   static void pci_del_option_rom(PCIDevice *pdev)