Re: [PATCH] xen/virtio: Convert PAGE_SIZE/PAGE_SHIFT/PFN_UP to Xen counterparts

From: Oleksandr Tyshchenko
Date: Fri Oct 07 2022 - 12:16:50 EST



On 07.10.22 18:50, Xenia Ragiadakou wrote:


Hello Xenia

>
> On 10/7/22 16:43, Oleksandr Tyshchenko wrote:
>>
>> On 07.10.22 10:15, Xenia Ragiadakou wrote:
>>>
>>> On 10/7/22 00:13, Oleksandr Tyshchenko wrote:
>>>
>>> Hi Oleksandr
>>
>>
>> Hello Xenia
>>
>>
>>>
>>>>
>>>> On 06.10.22 20:59, Xenia Ragiadakou wrote:
>>>>
>>>> Hello Xenia
>>>>
>>>>>
>>>>> On 10/6/22 15:09, Oleksandr Tyshchenko wrote:
>>>>>> From: Oleksandr Tyshchenko <oleksandr_tyshchenko@xxxxxxxx>
>>>>>>
>>>>>> Although XEN_PAGE_SIZE is equal to PAGE_SIZE (4KB) for now, it would
>>>>>> be more correct to use Xen specific #define-s as XEN_PAGE_SIZE can
>>>>>> be changed at some point in the future.
>>>>>>
>>>>>> Signed-off-by: Oleksandr Tyshchenko <oleksandr_tyshchenko@xxxxxxxx>
>>>>>> ---
>>>>>> Cc: Juergen Gross <jgross@xxxxxxxx>
>>>>>> Cc: Xenia Ragiadakou <burzalodowa@xxxxxxxxx>
>>>>>>
>>>>>> As it was proposed at:
>>>>>> https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/20221005174823.1800761-1-olekstysh@xxxxxxxxx/__;!!GF_29dbcQIUBPA!zHt-xZ_7tZc_EM6zva21E_YgwIiEeimFWfsJIpPwAu-TBcnzQhXHqlKzmXmwIcI6uIx_arHNZiaZeHt_428_8p-DyMpd$
>>>>>>
>>>>>>
>>>>>> [lore[.]kernel[.]org]
>>>>>>
>>>>>> Should go in only after that series.
>>>>>> ---
>>>>>>     drivers/xen/grant-dma-ops.c | 20 ++++++++++----------
>>>>>>     1 file changed, 10 insertions(+), 10 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/xen/grant-dma-ops.c
>>>>>> b/drivers/xen/grant-dma-ops.c
>>>>>> index c66f56d24013..5392fdc25dca 100644
>>>>>> --- a/drivers/xen/grant-dma-ops.c
>>>>>> +++ b/drivers/xen/grant-dma-ops.c
>>>>>> @@ -31,12 +31,12 @@ static
>>>>>> DEFINE_XARRAY_FLAGS(xen_grant_dma_devices,
>>>>>> XA_FLAGS_LOCK_IRQ);
>>>>>>       static inline dma_addr_t grant_to_dma(grant_ref_t grant)
>>>>>>     {
>>>>>> -    return XEN_GRANT_DMA_ADDR_OFF | ((dma_addr_t)grant <<
>>>>>> PAGE_SHIFT);
>>>>>> +    return XEN_GRANT_DMA_ADDR_OFF | ((dma_addr_t)grant <<
>>>>>> XEN_PAGE_SHIFT);
>>>>>>     }
>>>>>
>>>>> With this change, can the offset added to the dma handle,
>>>>> generated by
>>>>> grant_to_dma(), be the offset in the page? Couldn't it corrupt the
>>>>> grant ref?
>>>>
>>>>
>>>> Good point, indeed, I think it could corrupt if guest uses a different
>>>> than Xen page granularity (i.e 64KB).
>>>>
>>>>
>>>>>
>>>>>>       static inline grant_ref_t dma_to_grant(dma_addr_t dma)
>>>>>>     {
>>>>>> -    return (grant_ref_t)((dma & ~XEN_GRANT_DMA_ADDR_OFF) >>
>>>>>> PAGE_SHIFT);
>>>>>> +    return (grant_ref_t)((dma & ~XEN_GRANT_DMA_ADDR_OFF) >>
>>>>>> XEN_PAGE_SHIFT);
>>>>>>     }
>>>>>>       static struct xen_grant_dma_data
>>>>>> *find_xen_grant_dma_data(struct
>>>>>> device *dev)
>>>>>> @@ -79,7 +79,7 @@ static void *xen_grant_dma_alloc(struct device
>>>>>> *dev, size_t size,
>>>>>>                      unsigned long attrs)
>>>>>>     {
>>>>>>         struct xen_grant_dma_data *data;
>>>>>> -    unsigned int i, n_pages = PFN_UP(size);
>>>>>> +    unsigned int i, n_pages = XEN_PFN_UP(size);
>>>>>>         unsigned long pfn;
>>>>>>         grant_ref_t grant;
>>>>>>         void *ret;
>>>>>> @@ -91,14 +91,14 @@ static void *xen_grant_dma_alloc(struct device
>>>>>> *dev, size_t size,
>>>>>>         if (unlikely(data->broken))
>>>>>>             return NULL;
>>>>>>     -    ret = alloc_pages_exact(n_pages * PAGE_SIZE, gfp);
>>>>>> +    ret = alloc_pages_exact(n_pages * XEN_PAGE_SIZE, gfp);
>>>>>>         if (!ret)
>>>>>>             return NULL;
>>>>>>           pfn = virt_to_pfn(ret);
>>>>>>           if (gnttab_alloc_grant_reference_seq(n_pages, &grant)) {
>>>>>> -        free_pages_exact(ret, n_pages * PAGE_SIZE);
>>>>>> +        free_pages_exact(ret, n_pages * XEN_PAGE_SIZE);
>>>>>>             return NULL;
>>>>>>         }
>>>>>>     @@ -116,7 +116,7 @@ static void xen_grant_dma_free(struct device
>>>>>> *dev, size_t size, void *vaddr,
>>>>>>                        dma_addr_t dma_handle, unsigned long attrs)
>>>>>>     {
>>>>>>         struct xen_grant_dma_data *data;
>>>>>> -    unsigned int i, n_pages = PFN_UP(size);
>>>>>> +    unsigned int i, n_pages = XEN_PFN_UP(size);
>>>>>>         grant_ref_t grant;
>>>>>>           data = find_xen_grant_dma_data(dev);
>>>>>> @@ -138,7 +138,7 @@ static void xen_grant_dma_free(struct device
>>>>>> *dev, size_t size, void *vaddr,
>>>>>>           gnttab_free_grant_reference_seq(grant, n_pages);
>>>>>>     -    free_pages_exact(vaddr, n_pages * PAGE_SIZE);
>>>>>> +    free_pages_exact(vaddr, n_pages * XEN_PAGE_SIZE);
>>>>>>     }
>>>>>>       static struct page *xen_grant_dma_alloc_pages(struct device
>>>>>> *dev,
>>>>>> size_t size,
>>>>>> @@ -168,7 +168,7 @@ static dma_addr_t xen_grant_dma_map_page(struct
>>>>>> device *dev, struct page *page,
>>>>>>                          unsigned long attrs)
>>>>>>     {
>>>>>>         struct xen_grant_dma_data *data;
>>>>>> -    unsigned int i, n_pages = PFN_UP(offset + size);
>>>>>> +    unsigned int i, n_pages = XEN_PFN_UP(offset + size);
>>>>>
>>>>> The offset, here, refers to the offset in the page ...
>>>>>
>>>>>>         grant_ref_t grant;
>>>>>>         dma_addr_t dma_handle;
>>>>>>     @@ -200,8 +200,8 @@ static void xen_grant_dma_unmap_page(struct
>>>>>> device *dev, dma_addr_t dma_handle,
>>>>>>                          unsigned long attrs)
>>>>>>     {
>>>>>>         struct xen_grant_dma_data *data;
>>>>>> -    unsigned long offset = dma_handle & (PAGE_SIZE - 1);
>>>>>> -    unsigned int i, n_pages = PFN_UP(offset + size);
>>>>>> +    unsigned long offset = dma_handle & ~XEN_PAGE_MASK;
>>>>>
>>>>> ... while, here, it refers to the offset in the grant.
>>>>> So, the calculated number of grants may differ.
>>>>
>>>> Good point, I think you are right, so we need to additionally use
>>>> xen_offset_in_page() macro in xen_grant_dma_map_page(),
>>>>
>>>> something like that to be squashed with current patch:
>>>>
>>>>
>>>> diff --git a/drivers/xen/grant-dma-ops.c b/drivers/xen/grant-dma-ops.c
>>>> index 9d5eca6d638a..bb984dc05deb 100644
>>>> --- a/drivers/xen/grant-dma-ops.c
>>>> +++ b/drivers/xen/grant-dma-ops.c
>>>> @@ -169,7 +169,7 @@ static dma_addr_t xen_grant_dma_map_page(struct
>>>> device *dev, struct page *page,
>>>>                                             unsigned long attrs)
>>>>     {
>>>>            struct xen_grant_dma_data *data;
>>>> -       unsigned int i, n_pages = XEN_PFN_UP(offset + size);
>>>> +       unsigned int i, n_pages =
>>>> XEN_PFN_UP(xen_offset_in_page(offset)
>>>> + size);
>>>>            grant_ref_t grant;
>>>>            dma_addr_t dma_handle;
>>>>
>>>> @@ -191,7 +191,7 @@ static dma_addr_t xen_grant_dma_map_page(struct
>>>> device *dev, struct page *page,
>>>>                                    xen_page_to_gfn(page) + i, dir ==
>>>> DMA_TO_DEVICE);
>>>>            }
>>>>
>>>> -       dma_handle = grant_to_dma(grant) + offset;
>>>> +       dma_handle = grant_to_dma(grant) + xen_offset_in_page(offset);
>>>>
>>>>            return dma_handle;
>>>>     }
>>>>
>>>> Did I get your point right?
>>>>
>>>
>>> I think it 's more complicated than that.
>>> Let's say that the offset in page is > XEN_PAGE_SIZE, then the
>>> calculation of the number of grants won't take into account the part
>>> of the offset that is multiple of the XEN_PAGE_SIZE i.e it will
>>> calculate only the strictly necessary number of grants.
>>> But xen_grant_dma_map_page() grants access to the whole page because,
>>> as it can be observed in the code snippet below, it does not take into
>>> account the page offset.
>>>
>>> for (i = 0; i < n_pages; i++) {
>>>    gnttab_grant_foreign_access_ref(grant + i, data->backend_domid,
>>> xen_page_to_gfn(page) + i, dir == DMA_TO_DEVICE);
>>> }
>>
>>
>> Thanks, valid point. Agree it's indeed more complicated. I will comment
>> on that later. I have just pushed another fix, it is not related to
>> XEN_PAGE_SIZE directly, but also about page offset > PAGE_SIZE
>>
>
> I got a little bit confused with the order that the patches will be
> applied :)

This series should go in the first [1]:

Current patch depends on it and new patch [2] also depends on it. I
think (but might mistake) that current patch we need/want to postpone
(because it doesn't fix/improve something immediately, but more for
future use-cases),

but the new patch is a fix for the real situation. Once we decide with
new patch [2] we will be able to return to the current patch and rebase it.


[1]
https://lore.kernel.org/xen-devel/20221005174823.1800761-1-olekstysh@xxxxxxxxx/

[2]
https://lore.kernel.org/xen-devel/20221007132736.2275574-1-olekstysh@xxxxxxxxx/



> IIUC the above scenario cannot happen, i.e the offset to be >
> PAGE_SIZE, because this callback is used to map for transfer a portion
> of a single page.

It happens. I have rechecked that. And can provide some debug prints if
needed.


>
>
>> so touches the same code and should be prereq:
>>
>> https://urldefense.com/v3/__https://lore.kernel.org/all/20221007132736.2275574-1-olekstysh@xxxxxxxxx/__;!!GF_29dbcQIUBPA!36CJ5S5T47HMh9Nq-0WVUPSlnHjvlUon-ooFGowbvGA8BdWSCC9niF0f_btvite1G6LIRwyw6XsU6PAnCjMy92KiNkMG$
>> [lore[.]kernel[.]org]
>>
>>
>>>
>>>>>
>>>>>
>>>>>> +    unsigned int i, n_pages = XEN_PFN_UP(offset + size);
>>>>>>         grant_ref_t grant;
>>>>>>           if (WARN_ON(dir == DMA_NONE))
>>>>>
>>>>
>>>> Thank you.
>>>>
>>>>
>>>
>
--
Regards,

Oleksandr Tyshchenko