iommu: omap_iovmm: support non page-aligned buffers in iommu_vmap

From: Laurent Pinchart <[email protected]>

omap_iovmm requires page-aligned buffers, and that sometimes causes omap3isp failures (i.e. whenever the buffer passed from userspace is not page-aligned).

Remove this limitation by rounding the address of the first page entry down, and adding the offset back to the device address.

Signed-off-by Laurent Pinchart <[email protected]>

Acked-by Hiroshi DOYU <[email protected]>

[ohad at wizery.com: slightly edited the commit log]

Signed-off-by Ohad Ben-Cohen <[email protected]>

---
A fix by Laurent that was waiting until omap's iommu lands in drivers/.

Generally we're not extending omap-iovmm anymore, but this fixes a real breakage for omap3isp users, so there's no point in delaying it until the generic DMA API is ready and we've migrated.

Thanks Laurent!

?drivers/iommu/omap-iovmm.c | ? 32 ++++++++++++++++++++++++-------- ?1 files changed, 24 insertions(+), 8 deletions(-)

drivers/iommu/omap-iovmm.c +9 -0

index 5e7f97d..d28a256 100644
@@ -27,6 +27,15 @@
 
 ?static struct kmem_cache *iovm_area_cachep;
 
+/* return the offset of the first scatterlist entry in a sg table */
+static unsigned int sgtable_offset(const struct sg_table *sgt)
+{
+ ? ? ? if (!sgt || !sgt->nents)
+ ? ? ? ? ? ? ? return 0;
+
+ ? ? ? return sgt->sgl->offset;
+}
+

?/* return total bytes of sg buffers */ ?static size_t sgtable_len(const struct sg_table *sgt) ?{

drivers/iommu/omap-iovmm.c +14 -8

@@ -39,11 +48,17 @@ static size_t sgtable_len(const struct sg_table *sgt)
 ? ? ? ?for_each_sg(sgt->sgl, sg, sgt->nents, i) {
 ? ? ? ? ? ? ? ?size_t bytes;
 
- ? ? ? ? ? ? ? bytes = sg->length;
+ ? ? ? ? ? ? ? bytes = sg->length + sg->offset;
 
 ? ? ? ? ? ? ? ?if (!iopgsz_ok(bytes)) {
- ? ? ? ? ? ? ? ? ? ? ? pr_err("%s: sg[%d] not iommu pagesize(%x)\n",
- ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?__func__, i, bytes);
+ ? ? ? ? ? ? ? ? ? ? ? pr_err("%s: sg[%d] not iommu pagesize(%u %u)\n",
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?__func__, i, bytes, sg->offset);
+ ? ? ? ? ? ? ? ? ? ? ? return 0;
+ ? ? ? ? ? ? ? }
+
+ ? ? ? ? ? ? ? if (i && sg->offset) {
+ ? ? ? ? ? ? ? ? ? ? ? pr_err("%s: sg[%d] offset not allowed in internal "
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? "entries\n", __func__, i);
 ? ? ? ? ? ? ? ? ? ? ? ?return 0;
 ? ? ? ? ? ? ? ?}
 

@@ -164,8 +179,8 @@ static void *vmap_sg(const struct sg_table *sgt)
 ? ? ? ? ? ? ? ?u32 pa;
 ? ? ? ? ? ? ? ?int err;
 
- ? ? ? ? ? ? ? pa = sg_phys(sg);
- ? ? ? ? ? ? ? bytes = sg->length;
+ ? ? ? ? ? ? ? pa = sg_phys(sg) - sg->offset;
+ ? ? ? ? ? ? ? bytes = sg->length + sg->offset;
 
 ? ? ? ? ? ? ? ?BUG_ON(bytes != PAGE_SIZE);
 

@@ -405,8 +420,8 @@ static int map_iovm_area(struct iommu_domain *domain, struct iovm_struct *new,
 ? ? ? ? ? ? ? ?u32 pa;
 ? ? ? ? ? ? ? ?size_t bytes;
 
- ? ? ? ? ? ? ? pa = sg_phys(sg);
- ? ? ? ? ? ? ? bytes = sg->length;
+ ? ? ? ? ? ? ? pa = sg_phys(sg) - sg->offset;
+ ? ? ? ? ? ? ? bytes = sg->length + sg->offset;
 
 ? ? ? ? ? ? ? ?flags &= ~IOVMF_PGSZ_MASK;
 

@@ -600,7 +615,7 @@ u32 omap_iommu_vmap(struct iommu_domain *domain, struct omap_iommu *obj, u32 da,
 ? ? ? ?if (IS_ERR_VALUE(da))
 ? ? ? ? ? ? ? ?vunmap_sg(va);
 
- ? ? ? return da;
+ ? ? ? return da + sgtable_offset(sgt);
 ?}
 ?EXPORT_SYMBOL_GPL(omap_iommu_vmap);
 

@@ -620,6 +635,7 @@ omap_iommu_vunmap(struct iommu_domain *domain, struct omap_iommu *obj, u32 da)

? ? ? ? * 'sgt' is allocated before 'omap_iommu_vmalloc()' is called. ? ? ? ? * Just returns 'sgt' to the caller to free
? ? ? ? */
+ ? ? ? da &= PAGE_MASK;
? ? ? ?sgt = unmap_vm_area(domain, obj, da, vunmap_sg,
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?IOVMF_DISCONT | IOVMF_MMIO); ? ? ? ?if (!sgt)
--
1.7.4.1

Quoted 8 lines · Ohad Ben-Cohen

From: Laurent Pinchart <[email protected]>

omap_iovmm requires page-aligned buffers, and that sometimes causes omap3isp failures (i.e. whenever the buffer passed from userspace is not page-aligned).

Remove this limitation by rounding the address of the first page entry down, and adding the offset back to the device address.

I'm having second thoughts about this.

Obviously it works for omap3isp and its users because the buffer gets mapped and everyone is happy.

But I'm not sure this is a valid IOMMU interface that the kernel should have, because effectively we're now mapping physical memory which nobody asked us to, and which might contain sensitive stuff we don't want to give the device (e.g. a remote processor which might be running rogue code) access to.

Thoughts ?

Quoted 8 lines · Ohad Ben-Cohen

From: Laurent Pinchart <[email protected]>

omap_iovmm requires page-aligned buffers, and that sometimes causes omap3isp failures (i.e. whenever the buffer passed from userspace is not page-aligned).

Remove this limitation by rounding the address of the first page entry down, and adding the offset back to the device address.

I'm having second thoughts about this.

Obviously it works for omap3isp and its users because the buffer gets mapped and everyone is happy.

But I'm not sure this is a valid IOMMU interface that the kernel should have, because effectively we're now mapping physical memory which nobody asked us to, and which might contain sensitive stuff we don't want to give the device (e.g. a remote processor which might be running rogue code) access to.

Thoughts ?

Quoted 15 lines · Ohad Ben-Cohen

Quoted 8 lines · Ohad Ben-Cohen

From: Laurent Pinchart <[email protected]>

omap_iovmm requires page-aligned buffers, and that sometimes causes omap3isp failures (i.e. whenever the buffer passed from userspace is not page-aligned).

Remove this limitation by rounding the address of the first page entry down, and adding the offset back to the device address.

I'm having second thoughts about this.

Obviously it works for omap3isp and its users because the buffer gets mapped and everyone is happy.

Quoted 7 lines

But I'm not sure this is a valid IOMMU interface that the kernel should have, because effectively we're now mapping physical memory which nobody asked us to, and which might contain sensitive stuff we don't want to give the device (e.g. a remote processor which might be running rogue code) access to.

Thoughts ?

Do you mean the parts of the pages you map to the device that are not in the requested range (basically everything before offset and all after size)?
This issue exists in other iommu drivers as well. It is inherent to how the dma-api is defined and how the iommu hardware works.
The dma-api can work on byte granularity while the hardware usually only works on page granularity.

Quoted 7 lines · Joerg

Do you mean the parts of the pages you map to the device that are not in the requested range (basically everything before offset and all after size)?
This issue exists in other iommu drivers as well. It is inherent to how the dma-api is defined and how the iommu hardware works.
The dma-api can work on byte granularity while the hardware usually only works on page granularity.

True, but if we implement address rounding transparently in the IOMMU layer Ohad's concern can be valid, depending on whether the device is trusted. If we decide to push address rounding to the drivers that decision can be made on a per-device basis. However, drivers are usually not aware of what granularity the IOMMU works on, so that wouldn't be straightforward and clean.

True, but if we implement address rounding transparently in the IOMMU layer Ohad's concern can be valid, depending on whether the device is trusted. If we decide to push address rounding to the drivers that decision can be made on a per-device basis. However, drivers are usually not aware of what granularity the IOMMU works on, so that wouldn't be straightforward and clean.

From: Laurent Pinchart <[email protected]>

omap_iovmm requires page-aligned buffers, and that sometimes causes omap3isp failures (i.e. whenever the buffer passed from userspace is not page-aligned).

Remove this limitation by rounding the address of the first page entry down, and adding the offset back to the device address.

Quoted 6 lines · Ohad Ben-Cohen

From: Laurent Pinchart <[email protected]>

omap_iovmm requires page-aligned buffers, and that sometimes causes omap3isp failures (i.e. whenever the buffer passed from userspace is not page-aligned).

Quoted 2 lines

Remove this limitation by rounding the address of the first page entry down, and adding the offset back to the device address.

Seems like the unmap paths were skipped (need to adjust the sizes in the unmap path too).

Laurent, if it looks good to you, I'll just squash it to the original patch and repost:

drivers/iommu/omap-iovmm.c +2 -2

index d28a256..39bdb92 100644
@@ -447,7 +447,7 @@ err_out:
        for_each_sg(sgt->sgl, sg, i, j) {
                size_t bytes;
 
-               bytes = sg->length;
+               bytes = sg->length + sg->offset;
                order = get_order(bytes);
 
                /* ignore failures.. we're already handling one */

@@ -476,7 +476,7 @@ static void unmap_iovm_area(struct iommu_domain
 *domain, str size_t bytes;
                int order;
 
-               bytes = sg->length;
+               bytes = sg->length + sg->offset;
                order = get_order(bytes);
 
                err = iommu_unmap(domain, start, order);

Hi Ohad,

Quoted 10 lines · On Thursday 01 September 2011 13:47:26 Ohad Ben-Cohen wrote:

Quoted 8 lines · Ohad Ben-Cohen

From: Laurent Pinchart <[email protected]>

omap_iovmm requires page-aligned buffers, and that sometimes causes omap3isp failures (i.e. whenever the buffer passed from userspace is not page-aligned).

Remove this limitation by rounding the address of the first page entry down, and adding the offset back to the device address.

Quoted 5 lines

Seems like the unmap paths were skipped (need to adjust the sizes in the unmap path too).

Laurent, if it looks good to you, I'll just squash it to the original patch and repost:

Do you have a tree where the current code base can be found ?

Quoted 6 lines · Ohad Ben-Cohen

From: Laurent Pinchart <[email protected]>

omap_iovmm requires page-aligned buffers, and that sometimes causes omap3isp failures (i.e. whenever the buffer passed from userspace is not page-aligned).

Quoted 2 lines

Remove this limitation by rounding the address of the first page entry down, and adding the offset back to the device address.

Seems like the unmap paths were skipped (need to adjust the sizes in the unmap path too).

Laurent, if it looks good to you, I'll just squash it to the original patch and repost:

drivers/iommu/omap-iovmm.c +2 -2

index d28a256..39bdb92 100644
@@ -447,7 +447,7 @@ err_out:
        for_each_sg(sgt->sgl, sg, i, j) {
                size_t bytes;
 
-               bytes = sg->length;
+               bytes = sg->length + sg->offset;
                order = get_order(bytes);
 
                /* ignore failures.. we're already handling one */

@@ -476,7 +476,7 @@ static void unmap_iovm_area(struct iommu_domain
 *domain, str size_t bytes;
                int order;
 
-               bytes = sg->length;
+               bytes = sg->length + sg->offset;
                order = get_order(bytes);
 
                err = iommu_unmap(domain, start, order);

It looks good to me.

I haven't tested it though.

On Thu, Sep 1, 2011 at 4:59 PM, Laurent Pinchart

Quoted 1 line · <[email protected]> wrote:

It looks good to me.

Thanks.

Quoted 1 line

I haven't tested it though.

I did, but I always get aligned buffers so my testing is a bit moot. I can't see how unmap could work without this patch, though, so I'll squash this and re-post.