Skip to contentSkip to conversation

xen/virtio: Handle cases when page offset > PAGE_SIZE properly

PATCH Superseded Oleksandr Tyshchenko 4 participants 6 messages Oct 7, 2022 – Oct 8, 2022 Xen hypervisor interface

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xen/virtio: Handle cases when page offset > PAGE_SIZE properly

From: Oleksandr Tyshchenko <[email protected]>

Passed to xen_grant_dma_map_page() offset in the page can be > PAGE_SIZE even if the guest uses the same page granularity as Xen (4KB).

Before current patch, if such case happened we ended up providing grants for the whole region in xen_grant_dma_map_page() which was really unnecessary. The more, we ended up not releasing all grants which represented that region in xen_grant_dma_unmap_page().

Current patch updates the code to be able to deal with such cases.

---
Cc: Juergen Gross <[email protected]>
Cc: Xenia Ragiadakou <[email protected]>

Depens on:
https://lore.kernel.org/xen-devel/[email protected]/

Should go in only after that series.

---
 drivers/xen/grant-dma-ops.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/xen/grant-dma-ops.c +5 -3
index c66f56d24013..1385f0e686fe 100644
@@ -168,7 +168,9 @@ 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 long dma_offset = offset_in_page(offset),
+			gfn_offset = PFN_DOWN(offset);
+	unsigned int i, n_pages = PFN_UP(dma_offset + size);
 	grant_ref_t grant;
 	dma_addr_t dma_handle;
 
@@ -187,10 +189,10 @@ static dma_addr_t xen_grant_dma_map_page(struct device *dev, struct page *page,
 
 	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);
+				xen_page_to_gfn(page) + i + gfn_offset, dir == DMA_TO_DEVICE);
 	}
 
-	dma_handle = grant_to_dma(grant) + offset;
+	dma_handle = grant_to_dma(grant) + dma_offset;
 
 	return dma_handle;
 }
Signature · 1 line
2.25.1
maintainer Reviewed-by on v1 1/1
#1 lore raw
Quoted 6 lines · Oleksandr Tyshchenko

From: Oleksandr Tyshchenko <[email protected]>

Passed to xen_grant_dma_map_page() offset in the page can be > PAGE_SIZE even if the guest uses the same page granularity as Xen (4KB).

Quoted 8 lines

Before current patch, if such case happened we ended up providing grants for the whole region in xen_grant_dma_map_page() which was really unnecessary. The more, we ended up not releasing all grants which represented that region in xen_grant_dma_unmap_page().

Current patch updates the code to be able to deal with such cases.

Juergen

#2 lore raw
In reply to Oleksandr Tyshchenko

On 10/7/22 16:27, Oleksandr Tyshchenko wrote:

Hi Oleksandr

Quoted 31 lines

From: Oleksandr Tyshchenko <[email protected]>

Passed to xen_grant_dma_map_page() offset in the page can be > PAGE_SIZE even if the guest uses the same page granularity as Xen (4KB).

Before current patch, if such case happened we ended up providing grants for the whole region in xen_grant_dma_map_page() which was really unnecessary. The more, we ended up not releasing all grants which represented that region in xen_grant_dma_unmap_page().

Current patch updates the code to be able to deal with such cases.

---
Cc: Juergen Gross <[email protected]>
Cc: Xenia Ragiadakou <[email protected]>

Depens on:
https://lore.kernel.org/xen-devel/[email protected]/

Should go in only after that series.

---
  drivers/xen/grant-dma-ops.c | 8 +++++---
  1 file changed, 5 insertions(+), 3 deletions(-)
drivers/xen/grant-dma-ops.c +0 -0
index c66f56d24013..1385f0e686fe 100644

drivers/xen/grant-dma-ops.c static dma_addr_t xen_grant_dma_map_page(struct device *dev, struct page *page, · line 173

Quoted 7 lines
drivers/xen/grant-dma-ops.c +3 -1
@@ -168,7 +168,9 @@ 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 long dma_offset = offset_in_page(offset),
+			gfn_offset = PFN_DOWN(offset);
+	unsigned int i, n_pages = PFN_UP(dma_offset + size);

IIUC, the above with a later patch will become:

dma_offset = xen_offset_in_page(offset)
gfn_offset = XEN_PFN_DOWN(offset)
n_pages = XEN_PFN_UP(dma_offset + size)

drivers/xen/grant-dma-ops.c static dma_addr_t xen_grant_dma_map_page(struct device *dev, struct page *page, · line 192

Quoted 9 lines
drivers/xen/grant-dma-ops.c +1 -1
  	grant_ref_t grant;
  	dma_addr_t dma_handle;
 
@@ -187,10 +189,10 @@ static dma_addr_t xen_grant_dma_map_page(struct device *dev, struct page *page,
 
  	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);
+				xen_page_to_gfn(page) + i + gfn_offset, dir == DMA_TO_DEVICE);

Here, why the pfn is not calculated before passing it to pfn_to_gfn()? I mean sth like pfn_to_gfn(page_to_xen_pfn(page) + gfn_offset + i)

Quoted 7 lines
drivers/xen/grant-dma-ops.c +1 -1
  	}
 
-	dma_handle = grant_to_dma(grant) + offset;
+	dma_handle = grant_to_dma(grant) + dma_offset;
 
  	return dma_handle;
  }
Signature · 1 line
Xenia
#3 lore raw

On 08.10.22 14:08, Xenia Ragiadakou wrote:

Hello Xenia

Quoted 48 lines · Oleksandr Tyshchenko

Hi Oleksandr

Quoted 32 lines

From: Oleksandr Tyshchenko <[email protected]>

Passed to xen_grant_dma_map_page() offset in the page can be > PAGE_SIZE even if the guest uses the same page granularity as Xen (4KB).

Before current patch, if such case happened we ended up providing grants for the whole region in xen_grant_dma_map_page() which was really unnecessary. The more, we ended up not releasing all grants which represented that region in xen_grant_dma_unmap_page().

Current patch updates the code to be able to deal with such cases.

---
Cc: Juergen Gross <[email protected]>
Cc: Xenia Ragiadakou <[email protected]>

Depens on:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!xnkNaKpfZ4LssQJcJs_J91KERZKMP2Rd-xEdBqXNXJ8GyCXJ0gkRer1elVYfxOWtwN_FOl9tVieDWlfN-UZaHQsyLMhA$
[lore[.]kernel[.]org]

Should go in only after that series.
---
  drivers/xen/grant-dma-ops.c | 8 +++++---

  1 file changed, 5 insertions(+), 3 deletions(-)
drivers/xen/grant-dma-ops.c +0 -0
index c66f56d24013..1385f0e686fe 100644

drivers/xen/grant-dma-ops.c static dma_addr_t xen_grant_dma_map_page(struct · line 174

Quoted 8 lines
drivers/xen/grant-dma-ops.c +3 -1
@@ -168,7 +168,9 @@ 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 long dma_offset = offset_in_page(offset),
+            gfn_offset = PFN_DOWN(offset);
+    unsigned int i, n_pages = PFN_UP(dma_offset + size);

IIUC, the above with a later patch will become:

Quoted 3 lines

dma_offset = xen_offset_in_page(offset)
gfn_offset = XEN_PFN_DOWN(offset)
n_pages = XEN_PFN_UP(dma_offset + size)

If saying "later" patch you meant "xen/virtio: Convert PAGE_SIZE/PAGE_SHIFT/PFN_UP to Xen counterparts" then yes, exactly.

Quoted 13 lines
Quoted 4 lines
drivers/xen/grant-dma-ops.c +0 -0
       grant_ref_t grant;
       dma_addr_t dma_handle;
Quoted 8 lines

  @@ -187,10 +189,10 @@ static dma_addr_t xen_grant_dma_map_page(struct device *dev, struct page *page, 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); +                xen_page_to_gfn(page) + i + gfn_offset, dir == DMA_TO_DEVICE);

Quoted 2 lines

Here, why the pfn is not calculated before passing it to pfn_to_gfn()? I mean sth like pfn_to_gfn(page_to_xen_pfn(page) + gfn_offset + i)

The gfn_offset is just a const value here, which just means how many gfns we should skip. But ...

... I think, I get your point. So, if the region which is contiguous in pfn might be non-contiguous in gfn (which seems to be the case for x86's PV, but I may mistake) we should indeed use open-coded

construction "pfn_to_gfn(page_to_xen_pfn(page) + gfn_offset + i)".  And the gfn_offset should be renamed to pfn_offset then.

Correct?

Thanks.

Quoted 7 lines
Quoted 6 lines

      }
  -    dma_handle = grant_to_dma(grant) + offset; +    dma_handle = grant_to_dma(grant) + dma_offset; return dma_handle;
  }

Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#4 lore raw
Quoted 53 lines · Oleksandr Tyshchenko

On 08.10.22 14:08, Xenia Ragiadakou wrote:

Hello Xenia

Quoted 48 lines · Oleksandr Tyshchenko

Hi Oleksandr

Quoted 32 lines

From: Oleksandr Tyshchenko <[email protected]>

Passed to xen_grant_dma_map_page() offset in the page can be > PAGE_SIZE even if the guest uses the same page granularity as Xen (4KB).

Before current patch, if such case happened we ended up providing grants for the whole region in xen_grant_dma_map_page() which was really unnecessary. The more, we ended up not releasing all grants which represented that region in xen_grant_dma_unmap_page().

Current patch updates the code to be able to deal with such cases.

---
Cc: Juergen Gross <[email protected]>
Cc: Xenia Ragiadakou <[email protected]>

Depens on:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!xnkNaKpfZ4LssQJcJs_J91KERZKMP2Rd-xEdBqXNXJ8GyCXJ0gkRer1elVYfxOWtwN_FOl9tVieDWlfN-UZaHQsyLMhA$
[lore[.]kernel[.]org]

Should go in only after that series.

---
   drivers/xen/grant-dma-ops.c | 8 +++++---
   1 file changed, 5 insertions(+), 3 deletions(-)
drivers/xen/grant-dma-ops.c +0 -0
index c66f56d24013..1385f0e686fe 100644

drivers/xen/grant-dma-ops.c static dma_addr_t xen_grant_dma_map_page(struct · line 174

Quoted 8 lines
drivers/xen/grant-dma-ops.c +3 -1
@@ -168,7 +168,9 @@ 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 long dma_offset = offset_in_page(offset),
+            gfn_offset = PFN_DOWN(offset);
+    unsigned int i, n_pages = PFN_UP(dma_offset + size);

IIUC, the above with a later patch will become:

Quoted 7 lines
Quoted 3 lines

dma_offset = xen_offset_in_page(offset)
gfn_offset = XEN_PFN_DOWN(offset)
n_pages = XEN_PFN_UP(dma_offset + size)

If saying "later" patch you meant "xen/virtio: Convert PAGE_SIZE/PAGE_SHIFT/PFN_UP to Xen counterparts" then yes, exactly.

Ah ok, I see.

Quoted 23 lines
Quoted 13 lines
Quoted 4 lines
drivers/xen/grant-dma-ops.c +0 -0
       grant_ref_t grant;
       dma_addr_t dma_handle;
Quoted 8 lines

  @@ -187,10 +189,10 @@ static dma_addr_t
xen_grant_dma_map_page(struct device *dev, struct page *page,
        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); +                xen_page_to_gfn(page) + i + gfn_offset, dir == DMA_TO_DEVICE);

Quoted 2 lines

Here, why the pfn is not calculated before passing it to pfn_to_gfn()? I mean sth like pfn_to_gfn(page_to_xen_pfn(page) + gfn_offset + i)

The gfn_offset is just a const value here, which just means how many gfns we should skip. But ...

... I think, I get your point. So, if the region which is contiguous in pfn might be non-contiguous in gfn (which seems to be the case for x86's PV, but I may mistake) we should indeed use open-coded

Quoted 5 lines

construction "pfn_to_gfn(page_to_xen_pfn(page) + gfn_offset + i)".  And the gfn_offset should be renamed to pfn_offset then.

Correct?

Yes, that 's what I had in mind unless I 'm missing sth.

Quoted 7 lines
Quoted 7 lines
Quoted 6 lines

      }
  -    dma_handle = grant_to_dma(grant) + offset;
+    dma_handle = grant_to_dma(grant) + dma_offset;
        return dma_handle;
  }

Signature · 1 line
Xenia
#5 lore raw

On 08.10.22 15:59, Xenia Ragiadakou wrote:

Hello Xenia

Quoted 89 lines · Oleksandr Tyshchenko
Quoted 54 lines · Xenia Ragiadakou

Hello Xenia

Quoted 49 lines · Oleksandr Tyshchenko

Hi Oleksandr

Quoted 33 lines

From: Oleksandr Tyshchenko <[email protected]>

Passed to xen_grant_dma_map_page() offset in the page can be > PAGE_SIZE even if the guest uses the same page granularity as Xen (4KB).

Before current patch, if such case happened we ended up providing grants for the whole region in xen_grant_dma_map_page() which was really unnecessary. The more, we ended up not releasing all grants which represented that region in xen_grant_dma_unmap_page().

Current patch updates the code to be able to deal with such cases.

---
Cc: Juergen Gross <[email protected]>
Cc: Xenia Ragiadakou <[email protected]>

Depens on:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!xnkNaKpfZ4LssQJcJs_J91KERZKMP2Rd-xEdBqXNXJ8GyCXJ0gkRer1elVYfxOWtwN_FOl9tVieDWlfN-UZaHQsyLMhA$

[lore[.]kernel[.]org]

Should go in only after that series.
---
   drivers/xen/grant-dma-ops.c | 8 +++++---

   1 file changed, 5 insertions(+), 3 deletions(-)
drivers/xen/grant-dma-ops.c +0 -0
index c66f56d24013..1385f0e686fe 100644

drivers/xen/grant-dma-ops.c static dma_addr_t xen_grant_dma_map_page(struct · line 174

Quoted 8 lines
drivers/xen/grant-dma-ops.c +3 -1
@@ -168,7 +168,9 @@ 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 long dma_offset = offset_in_page(offset),
+            gfn_offset = PFN_DOWN(offset);
+    unsigned int i, n_pages = PFN_UP(dma_offset + size);

IIUC, the above with a later patch will become:

Quoted 7 lines
Quoted 3 lines

dma_offset = xen_offset_in_page(offset)
gfn_offset = XEN_PFN_DOWN(offset)
n_pages = XEN_PFN_UP(dma_offset + size)

If saying "later" patch you meant "xen/virtio: Convert PAGE_SIZE/PAGE_SHIFT/PFN_UP to Xen counterparts" then yes, exactly.

Ah ok, I see.

Quoted 23 lines
Quoted 13 lines
Quoted 4 lines
drivers/xen/grant-dma-ops.c +0 -0
        grant_ref_t grant;
        dma_addr_t dma_handle;
Quoted 8 lines

   @@ -187,10 +189,10 @@ static dma_addr_t xen_grant_dma_map_page(struct device *dev, struct page *page, 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); +                xen_page_to_gfn(page) + i + gfn_offset, dir == DMA_TO_DEVICE);

Quoted 2 lines

Here, why the pfn is not calculated before passing it to pfn_to_gfn()? I mean sth like pfn_to_gfn(page_to_xen_pfn(page) + gfn_offset + i)

The gfn_offset is just a const value here, which just means how many gfns we should skip. But ...

... I think, I get your point. So, if the region which is contiguous in pfn might be non-contiguous in gfn (which seems to be the case for x86's PV, but I may mistake) we should indeed use open-coded

Quoted 7 lines
Quoted 5 lines

construction "pfn_to_gfn(page_to_xen_pfn(page) + gfn_offset + i)".  And the gfn_offset should be renamed to pfn_offset then.

Correct?

Yes, that 's what I had in mind unless I 'm missing sth.

ok, thanks for confirming. So I will create V2 then.

Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

       }
   -    dma_handle = grant_to_dma(grant) + offset; +    dma_handle = grant_to_dma(grant) + dma_offset; return dma_handle;
   }

Signature · 3 lines
Regards,

Oleksandr Tyshchenko