diff mbox series

iommu: Improve the performance for direct_mapping

Message ID 20201120090628.6566-1-yong.wu@mediatek.com (mailing list archive)
State New
Headers show
Series iommu: Improve the performance for direct_mapping | expand

Commit Message

Yong Wu Nov. 20, 2020, 9:06 a.m. UTC
Currently direct_mapping always use the smallest pgsize which is SZ_4K
normally to mapping. This is unnecessary. we could gather the size, and
call iommu_map then, iommu_map could decide how to map better with the
just right pgsize.

From the original comment, we should take care overlap, otherwise,
iommu_map may return -EEXIST. In this overlap case, we should map the
previous region before overlap firstly. then map the left part.

Each a iommu device will call this direct_mapping when its iommu
initialize, This patch is effective to improve the boot/initialization
time especially while it only needs level 1 mapping.

Signed-off-by: Anan Sun <anan.sun@mediatek.com>
Signed-off-by: Yong Wu <yong.wu@mediatek.com>
---
 drivers/iommu/iommu.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)

Comments

Will Deacon Nov. 23, 2020, 12:32 p.m. UTC | #1
On Fri, Nov 20, 2020 at 05:06:28PM +0800, Yong Wu wrote:
> Currently direct_mapping always use the smallest pgsize which is SZ_4K
> normally to mapping. This is unnecessary. we could gather the size, and
> call iommu_map then, iommu_map could decide how to map better with the
> just right pgsize.
> 
> From the original comment, we should take care overlap, otherwise,
> iommu_map may return -EEXIST. In this overlap case, we should map the
> previous region before overlap firstly. then map the left part.
> 
> Each a iommu device will call this direct_mapping when its iommu
> initialize, This patch is effective to improve the boot/initialization
> time especially while it only needs level 1 mapping.
> 
> Signed-off-by: Anan Sun <anan.sun@mediatek.com>
> Signed-off-by: Yong Wu <yong.wu@mediatek.com>
> ---
>  drivers/iommu/iommu.c | 20 ++++++++++++++++++--
>  1 file changed, 18 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> index df87c8e825f7..854a8fcb928d 100644
> --- a/drivers/iommu/iommu.c
> +++ b/drivers/iommu/iommu.c
> @@ -737,6 +737,7 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
>  	/* We need to consider overlapping regions for different devices */
>  	list_for_each_entry(entry, &mappings, list) {
>  		dma_addr_t start, end, addr;
> +		size_t unmapped_sz = 0;

I think "unmapped" is the wrong word here, as this variable actually
represents the amount we want to map! I suggest "map_size" instead.

>  		if (domain->ops->apply_resv_region)
>  			domain->ops->apply_resv_region(dev, domain, entry);
> @@ -752,10 +753,25 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
>  			phys_addr_t phys_addr;
>  
>  			phys_addr = iommu_iova_to_phys(domain, addr);
> -			if (phys_addr)
> +			if (phys_addr == 0) {
> +				unmapped_sz += pg_size; /* Gather the size. */
>  				continue;
> +			}
>  
> -			ret = iommu_map(domain, addr, addr, pg_size, entry->prot);
> +			if (unmapped_sz) {
> +				/* Map the region before the overlap. */
> +				ret = iommu_map(domain, start, start,
> +						unmapped_sz, entry->prot);
> +				if (ret)
> +					goto out;
> +				start += unmapped_sz;

I think it's a bit confusing to update start like this. Can we call
iommu_map(domain, addr - map_size, addr - map_size, map_size, entry->prot)
instead?

> +				unmapped_sz = 0;
> +			}
> +			start += pg_size;
> +		}
> +		if (unmapped_sz) {
> +			ret = iommu_map(domain, start, start, unmapped_sz,
> +					entry->prot);

Can you avoid this hunk by changing your loop check to something like:

	if (!phys_addr) {
		map_size += pg_size;
		if (addr + pg_size < end)
			continue;
	}

Will
Yong Wu Nov. 24, 2020, 9:24 a.m. UTC | #2
On Mon, 2020-11-23 at 12:32 +0000, Will Deacon wrote:
> On Fri, Nov 20, 2020 at 05:06:28PM +0800, Yong Wu wrote:
> > Currently direct_mapping always use the smallest pgsize which is SZ_4K
> > normally to mapping. This is unnecessary. we could gather the size, and
> > call iommu_map then, iommu_map could decide how to map better with the
> > just right pgsize.
> > 
> > From the original comment, we should take care overlap, otherwise,
> > iommu_map may return -EEXIST. In this overlap case, we should map the
> > previous region before overlap firstly. then map the left part.
> > 
> > Each a iommu device will call this direct_mapping when its iommu
> > initialize, This patch is effective to improve the boot/initialization
> > time especially while it only needs level 1 mapping.
> > 
> > Signed-off-by: Anan Sun <anan.sun@mediatek.com>
> > Signed-off-by: Yong Wu <yong.wu@mediatek.com>
> > ---
> >  drivers/iommu/iommu.c | 20 ++++++++++++++++++--
> >  1 file changed, 18 insertions(+), 2 deletions(-)
> > 
> > diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> > index df87c8e825f7..854a8fcb928d 100644
> > --- a/drivers/iommu/iommu.c
> > +++ b/drivers/iommu/iommu.c
> > @@ -737,6 +737,7 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
> >  	/* We need to consider overlapping regions for different devices */
> >  	list_for_each_entry(entry, &mappings, list) {
> >  		dma_addr_t start, end, addr;
> > +		size_t unmapped_sz = 0;
> 
> I think "unmapped" is the wrong word here, as this variable actually
> represents the amount we want to map! I suggest "map_size" instead.
> 
> >  		if (domain->ops->apply_resv_region)
> >  			domain->ops->apply_resv_region(dev, domain, entry);
> > @@ -752,10 +753,25 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
> >  			phys_addr_t phys_addr;
> >  
> >  			phys_addr = iommu_iova_to_phys(domain, addr);
> > -			if (phys_addr)
> > +			if (phys_addr == 0) {
> > +				unmapped_sz += pg_size; /* Gather the size. */
> >  				continue;
> > +			}
> >  
> > -			ret = iommu_map(domain, addr, addr, pg_size, entry->prot);
> > +			if (unmapped_sz) {
> > +				/* Map the region before the overlap. */
> > +				ret = iommu_map(domain, start, start,
> > +						unmapped_sz, entry->prot);
> > +				if (ret)
> > +					goto out;
> > +				start += unmapped_sz;
> 
> I think it's a bit confusing to update start like this. Can we call
> iommu_map(domain, addr - map_size, addr - map_size, map_size, entry->prot)
> instead?
> 
> > +				unmapped_sz = 0;
> > +			}
> > +			start += pg_size;
> > +		}
> > +		if (unmapped_sz) {
> > +			ret = iommu_map(domain, start, start, unmapped_sz,
> > +					entry->prot);
> 
> Can you avoid this hunk by changing your loop check to something like:
> 
> 	if (!phys_addr) {
> 		map_size += pg_size;
> 		if (addr + pg_size < end)
> 			continue;
> 	}

Thanks for your quick review. I have fixed and tested it. the patch is
simple. I copy it here. Is this readable for you now?


--- a/drivers/iommu/iommu.c
+++ b/drivers/iommu/iommu.c
@@ -737,6 +737,7 @@ static int
iommu_create_device_direct_mappings(struct iommu_group *group,
 	/* We need to consider overlapping regions for different devices */
 	list_for_each_entry(entry, &mappings, list) {
 		dma_addr_t start, end, addr;
+		size_t map_size = 0;
 
 		if (domain->ops->apply_resv_region)
 			domain->ops->apply_resv_region(dev, domain, entry);
@@ -752,12 +753,21 @@ static int
iommu_create_device_direct_mappings(struct iommu_group *group,
 			phys_addr_t phys_addr;
 
 			phys_addr = iommu_iova_to_phys(domain, addr);
-			if (phys_addr)
-				continue;
+			if (!phys_addr) {
+				map_size += pg_size;
+				if (addr + pg_size < end)
+					continue;
+				else
+					addr += pg_size; /*Point to End */
+			}
 
-			ret = iommu_map(domain, addr, addr, pg_size, entry->prot);
-			if (ret)
-				goto out;
+			if (map_size) {
+				ret = iommu_map(domain, addr - map_size, addr - map_size,
+						map_size, entry->prot);
+				if (ret)
+					goto out;
+				map_size = 0;
+			}
 		}

> 
> Will
Will Deacon Nov. 24, 2020, 11:05 a.m. UTC | #3
On Tue, Nov 24, 2020 at 05:24:44PM +0800, Yong Wu wrote:
> On Mon, 2020-11-23 at 12:32 +0000, Will Deacon wrote:
> > On Fri, Nov 20, 2020 at 05:06:28PM +0800, Yong Wu wrote:
> > > +				unmapped_sz = 0;
> > > +			}
> > > +			start += pg_size;
> > > +		}
> > > +		if (unmapped_sz) {
> > > +			ret = iommu_map(domain, start, start, unmapped_sz,
> > > +					entry->prot);
> > 
> > Can you avoid this hunk by changing your loop check to something like:
> > 
> > 	if (!phys_addr) {
> > 		map_size += pg_size;
> > 		if (addr + pg_size < end)
> > 			continue;
> > 	}
> 
> Thanks for your quick review. I have fixed and tested it. the patch is
> simple. I copy it here. Is this readable for you now?
> 
> 
> --- a/drivers/iommu/iommu.c
> +++ b/drivers/iommu/iommu.c
> @@ -737,6 +737,7 @@ static int
> iommu_create_device_direct_mappings(struct iommu_group *group,
>  	/* We need to consider overlapping regions for different devices */
>  	list_for_each_entry(entry, &mappings, list) {
>  		dma_addr_t start, end, addr;
> +		size_t map_size = 0;
>  
>  		if (domain->ops->apply_resv_region)
>  			domain->ops->apply_resv_region(dev, domain, entry);
> @@ -752,12 +753,21 @@ static int
> iommu_create_device_direct_mappings(struct iommu_group *group,
>  			phys_addr_t phys_addr;
>  
>  			phys_addr = iommu_iova_to_phys(domain, addr);
> -			if (phys_addr)
> -				continue;
> +			if (!phys_addr) {
> +				map_size += pg_size;
> +				if (addr + pg_size < end)
> +					continue;
> +				else

You don't need the 'else' here  ^^^

> +					addr += pg_size; /*Point to End */

addr = end ?

That said, maybe we could simplify this further by changing the loop bounds
to be:

	for (addr = start; addr <= end; addr += pg_size)

and checking:

	if (!phys_addr && addr != end) {
		map_size += pg_size;
		continue;
	}

does that work?

Will
Yong Wu Nov. 25, 2020, 11:03 a.m. UTC | #4
On Tue, 2020-11-24 at 11:05 +0000, Will Deacon wrote:
> On Tue, Nov 24, 2020 at 05:24:44PM +0800, Yong Wu wrote:
> > On Mon, 2020-11-23 at 12:32 +0000, Will Deacon wrote:
> > > On Fri, Nov 20, 2020 at 05:06:28PM +0800, Yong Wu wrote:
> > > > +				unmapped_sz = 0;
> > > > +			}
> > > > +			start += pg_size;
> > > > +		}
> > > > +		if (unmapped_sz) {
> > > > +			ret = iommu_map(domain, start, start, unmapped_sz,
> > > > +					entry->prot);
> > > 
> > > Can you avoid this hunk by changing your loop check to something like:
> > > 
> > > 	if (!phys_addr) {
> > > 		map_size += pg_size;
> > > 		if (addr + pg_size < end)
> > > 			continue;
> > > 	}
> > 
> > Thanks for your quick review. I have fixed and tested it. the patch is
> > simple. I copy it here. Is this readable for you now?
> > 
> > 
> > --- a/drivers/iommu/iommu.c
> > +++ b/drivers/iommu/iommu.c
> > @@ -737,6 +737,7 @@ static int
> > iommu_create_device_direct_mappings(struct iommu_group *group,
> >  	/* We need to consider overlapping regions for different devices */
> >  	list_for_each_entry(entry, &mappings, list) {
> >  		dma_addr_t start, end, addr;
> > +		size_t map_size = 0;
> >  
> >  		if (domain->ops->apply_resv_region)
> >  			domain->ops->apply_resv_region(dev, domain, entry);
> > @@ -752,12 +753,21 @@ static int
> > iommu_create_device_direct_mappings(struct iommu_group *group,
> >  			phys_addr_t phys_addr;
> >  
> >  			phys_addr = iommu_iova_to_phys(domain, addr);
> > -			if (phys_addr)
> > -				continue;
> > +			if (!phys_addr) {
> > +				map_size += pg_size;
> > +				if (addr + pg_size < end)
> > +					continue;
> > +				else
> 
> You don't need the 'else' here  ^^^
> 
> > +					addr += pg_size; /*Point to End */
> 
> addr = end ?
> 
> That said, maybe we could simplify this further by changing the loop bounds
> to be:
> 
> 	for (addr = start; addr <= end; addr += pg_size)
> 
> and checking:
> 
> 	if (!phys_addr && addr != end) {
> 		map_size += pg_size;
> 		continue;
> 	}
> 
> does that work?

It works but I think we can not check iommu_iova_to_phys(domain, end).
We should add a "if", like:

for (addr = start; addr <= end; addr += pg_size) {
...
	if (addr < end) {
		phys_addr = iommu_iova_to_phys(domain, addr);
		if (!phys_addr) {
			map_size += pg_size;
			continue;
		}
	}
...


If you don't like this "if (addr < end)", then we have to add a "goto".
like this:


for (addr = start; addr <= end; addr += pg_size) {
 	phys_addr_t phys_addr;
 
	if (addr == end)
		goto map_last;

	phys_addr = iommu_iova_to_phys(domain, addr);
	if (!phys_addr) {
		map_size += pg_size;
		continue;
	}

map_last:
	if (!map_size)
		continue;
	ret = iommu_map(domain, addr - map_size,
			addr - map_size, map_size, entry->prot);
	if (ret)
		goto out;
}

Which one is better?

> 
> Will
> 
> _______________________________________________
> Linux-mediatek mailing list
> Linux-mediatek@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-mediatek
Will Deacon Nov. 25, 2020, 11:13 a.m. UTC | #5
On Wed, Nov 25, 2020 at 07:03:34PM +0800, Yong Wu wrote:
> On Tue, 2020-11-24 at 11:05 +0000, Will Deacon wrote:
> > On Tue, Nov 24, 2020 at 05:24:44PM +0800, Yong Wu wrote:
> > > On Mon, 2020-11-23 at 12:32 +0000, Will Deacon wrote:
> > That said, maybe we could simplify this further by changing the loop bounds
> > to be:
> > 
> > 	for (addr = start; addr <= end; addr += pg_size)
> > 
> > and checking:
> > 
> > 	if (!phys_addr && addr != end) {
> > 		map_size += pg_size;
> > 		continue;
> > 	}
> > 
> > does that work?
> 
> It works but I think we can not check iommu_iova_to_phys(domain, end).
> We should add a "if", like:
> 
> for (addr = start; addr <= end; addr += pg_size) {
> ...
> 	if (addr < end) {
> 		phys_addr = iommu_iova_to_phys(domain, addr);
> 		if (!phys_addr) {
> 			map_size += pg_size;
> 			continue;
> 		}
> 	}
> ...
> 

Oh yes, you're right.

> If you don't like this "if (addr < end)", then we have to add a "goto".
> like this:
> 
> 
> for (addr = start; addr <= end; addr += pg_size) {
>  	phys_addr_t phys_addr;
>  
> 	if (addr == end)
> 		goto map_last;
> 
> 	phys_addr = iommu_iova_to_phys(domain, addr);
> 	if (!phys_addr) {
> 		map_size += pg_size;
> 		continue;
> 	}
> 
> map_last:
> 	if (!map_size)
> 		continue;
> 	ret = iommu_map(domain, addr - map_size,
> 			addr - map_size, map_size, entry->prot);

I think it's cleared to invert this as you had before:

	if (map_size)
		ret = iommu_map(...);

> Which one is better?

The second one is easier to read. I'll stop making suggestions now, thanks.

Will
Robin Murphy Nov. 26, 2020, 3:19 p.m. UTC | #6
On 2020-11-20 09:06, Yong Wu wrote:
> Currently direct_mapping always use the smallest pgsize which is SZ_4K
> normally to mapping. This is unnecessary. we could gather the size, and
> call iommu_map then, iommu_map could decide how to map better with the
> just right pgsize.
> 
>  From the original comment, we should take care overlap, otherwise,
> iommu_map may return -EEXIST. In this overlap case, we should map the
> previous region before overlap firstly. then map the left part.
> 
> Each a iommu device will call this direct_mapping when its iommu
> initialize, This patch is effective to improve the boot/initialization
> time especially while it only needs level 1 mapping.
> 
> Signed-off-by: Anan Sun <anan.sun@mediatek.com>
> Signed-off-by: Yong Wu <yong.wu@mediatek.com>
> ---
>   drivers/iommu/iommu.c | 20 ++++++++++++++++++--
>   1 file changed, 18 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> index df87c8e825f7..854a8fcb928d 100644
> --- a/drivers/iommu/iommu.c
> +++ b/drivers/iommu/iommu.c
> @@ -737,6 +737,7 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
>   	/* We need to consider overlapping regions for different devices */
>   	list_for_each_entry(entry, &mappings, list) {
>   		dma_addr_t start, end, addr;
> +		size_t unmapped_sz = 0;
>   
>   		if (domain->ops->apply_resv_region)
>   			domain->ops->apply_resv_region(dev, domain, entry);
> @@ -752,10 +753,25 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
>   			phys_addr_t phys_addr;
>   
>   			phys_addr = iommu_iova_to_phys(domain, addr);
> -			if (phys_addr)
> +			if (phys_addr == 0) {
> +				unmapped_sz += pg_size; /* Gather the size. */
>   				continue;
> +			}

I guess the reason we need to validate every page is because they may 
already have been legitimately mapped if someone else's reserved region 
overlaps - is it worth explicitly validating that, i.e. bail out if 
something's gone wrong enough that phys_addr != addr?

Other than the naming issue (I agree that map_size is a far, far better 
choice), I don't have any strong opinions about the rest of the 
implementation - I've written enough variations of this pattern to know 
that there's just no "nice" way to do it in C; all you can do is shuffle 
the clunkiness around :)

Robin.

>   
> -			ret = iommu_map(domain, addr, addr, pg_size, entry->prot);
> +			if (unmapped_sz) {
> +				/* Map the region before the overlap. */
> +				ret = iommu_map(domain, start, start,
> +						unmapped_sz, entry->prot);
> +				if (ret)
> +					goto out;
> +				start += unmapped_sz;
> +				unmapped_sz = 0;
> +			}
> +			start += pg_size;
> +		}
> +		if (unmapped_sz) {
> +			ret = iommu_map(domain, start, start, unmapped_sz,
> +					entry->prot);
>   			if (ret)
>   				goto out;
>   		}
>
Yong Wu Nov. 27, 2020, 6:21 a.m. UTC | #7
On Thu, 2020-11-26 at 15:19 +0000, Robin Murphy wrote:
> On 2020-11-20 09:06, Yong Wu wrote:
> > Currently direct_mapping always use the smallest pgsize which is SZ_4K
> > normally to mapping. This is unnecessary. we could gather the size, and
> > call iommu_map then, iommu_map could decide how to map better with the
> > just right pgsize.
> > 
> >  From the original comment, we should take care overlap, otherwise,
> > iommu_map may return -EEXIST. In this overlap case, we should map the
> > previous region before overlap firstly. then map the left part.
> > 
> > Each a iommu device will call this direct_mapping when its iommu
> > initialize, This patch is effective to improve the boot/initialization
> > time especially while it only needs level 1 mapping.
> > 
> > Signed-off-by: Anan Sun <anan.sun@mediatek.com>
> > Signed-off-by: Yong Wu <yong.wu@mediatek.com>
> > ---
> >   drivers/iommu/iommu.c | 20 ++++++++++++++++++--
> >   1 file changed, 18 insertions(+), 2 deletions(-)
> > 
> > diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> > index df87c8e825f7..854a8fcb928d 100644
> > --- a/drivers/iommu/iommu.c
> > +++ b/drivers/iommu/iommu.c
> > @@ -737,6 +737,7 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
> >   	/* We need to consider overlapping regions for different devices */
> >   	list_for_each_entry(entry, &mappings, list) {
> >   		dma_addr_t start, end, addr;
> > +		size_t unmapped_sz = 0;
> >   
> >   		if (domain->ops->apply_resv_region)
> >   			domain->ops->apply_resv_region(dev, domain, entry);
> > @@ -752,10 +753,25 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
> >   			phys_addr_t phys_addr;
> >   
> >   			phys_addr = iommu_iova_to_phys(domain, addr);
> > -			if (phys_addr)
> > +			if (phys_addr == 0) {
> > +				unmapped_sz += pg_size; /* Gather the size. */
> >   				continue;
> > +			}
> 
> I guess the reason we need to validate every page is because they may 
> already have been legitimately mapped if someone else's reserved region 
> overlaps - is it worth explicitly validating that, i.e. bail out if 
> something's gone wrong enough that phys_addr != addr?

I'm not sure the history about why to validate every page. this
direct_mapping is called very early, normally after alloc_default_domain
and _attach_device. the "phys_addr != addr" looks impossible.

If there is a normal flow that may cause "phys_addr != addr", then
something go wrong, Could we give a warning like adding a
WARN_ON_ONCE(phys_addr != addr)? and it should be in a another patch.

> 
> Other than the naming issue (I agree that map_size is a far, far better 
> choice), I don't have any strong opinions about the rest of the 
> implementation - I've written enough variations of this pattern to know 
> that there's just no "nice" way to do it in C; all you can do is shuffle 
> the clunkiness around :)

:). I will send a v2.
Thanks.

> 
> Robin.
> 
> >   
> > -			ret = iommu_map(domain, addr, addr, pg_size, entry->prot);
> > +			if (unmapped_sz) {
> > +				/* Map the region before the overlap. */
> > +				ret = iommu_map(domain, start, start,
> > +						unmapped_sz, entry->prot);
> > +				if (ret)
> > +					goto out;
> > +				start += unmapped_sz;
> > +				unmapped_sz = 0;
> > +			}
> > +			start += pg_size;
> > +		}
> > +		if (unmapped_sz) {
> > +			ret = iommu_map(domain, start, start, unmapped_sz,
> > +					entry->prot);
> >   			if (ret)
> >   				goto out;
> >   		}
> > 
> 
> _______________________________________________
> Linux-mediatek mailing list
> Linux-mediatek@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-mediatek
diff mbox series

Patch

diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
index df87c8e825f7..854a8fcb928d 100644
--- a/drivers/iommu/iommu.c
+++ b/drivers/iommu/iommu.c
@@ -737,6 +737,7 @@  static int iommu_create_device_direct_mappings(struct iommu_group *group,
 	/* We need to consider overlapping regions for different devices */
 	list_for_each_entry(entry, &mappings, list) {
 		dma_addr_t start, end, addr;
+		size_t unmapped_sz = 0;
 
 		if (domain->ops->apply_resv_region)
 			domain->ops->apply_resv_region(dev, domain, entry);
@@ -752,10 +753,25 @@  static int iommu_create_device_direct_mappings(struct iommu_group *group,
 			phys_addr_t phys_addr;
 
 			phys_addr = iommu_iova_to_phys(domain, addr);
-			if (phys_addr)
+			if (phys_addr == 0) {
+				unmapped_sz += pg_size; /* Gather the size. */
 				continue;
+			}
 
-			ret = iommu_map(domain, addr, addr, pg_size, entry->prot);
+			if (unmapped_sz) {
+				/* Map the region before the overlap. */
+				ret = iommu_map(domain, start, start,
+						unmapped_sz, entry->prot);
+				if (ret)
+					goto out;
+				start += unmapped_sz;
+				unmapped_sz = 0;
+			}
+			start += pg_size;
+		}
+		if (unmapped_sz) {
+			ret = iommu_map(domain, start, start, unmapped_sz,
+					entry->prot);
 			if (ret)
 				goto out;
 		}