[PATCHv5 0/4] dma, x86, xen: reduce SWIOTLB usage in Xen guests

[PATCHv5 0/4] dma, x86, xen: reduce SWIOTLB usage in Xen guests

[David Vrabel]

On systems where DMA addresses and physical addresses are not 1:1
(such as Xen PV guests), the generic dma_get_required_mask() will not
return the correct mask (since it uses max_pfn).

Some device drivers (such as mptsas, mpt2sas) use
dma_get_required_mask() to set the device's DMA mask to allow them to use
only 32-bit DMA addresses in hardware structures.  This results in
unnecessary use of the SWIOTLB if DMA addresses are more than 32-bits,
impacting performance significantly.

This series allows Xen PV guests to override the default
dma_get_required_mask() with a more suitable one.

Changes in v5:
- xen_swiotlb_get_required_mask() is x86 only.

Changes in v4:
- Assume 64-bit mask is required.

Changes in v3:
- fix off-by-one in xen_dma_get_required_mask()
- split ia64 changes into separate patch.

Changes in v2:
- split x86 and xen changes into separate patches

David

[PATCH 1/4] dma: add dma_get_required_mask_from_max_pfn()

[David Vrabel]

A generic dma_get_required_mask() is useful even for architectures (such
as ia64) that define ARCH_HAS_GET_REQUIRED_MASK.

Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
---
 drivers/base/platform.c     |   10 ++++++++--
 include/linux/dma-mapping.h |    1 +
 2 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/drivers/base/platform.c b/drivers/base/platform.c
index b2afc29..f9f3930 100644
--- a/drivers/base/platform.c
+++ b/drivers/base/platform.c
@@ -1009,8 +1009,7 @@ int __init platform_bus_init(void)
 	return error;
 }
 
-#ifndef ARCH_HAS_DMA_GET_REQUIRED_MASK
-u64 dma_get_required_mask(struct device *dev)
+u64 dma_get_required_mask_from_max_pfn(struct device *dev)
 {
 	u32 low_totalram = ((max_pfn - 1) << PAGE_SHIFT);
 	u32 high_totalram = ((max_pfn - 1) >> (32 - PAGE_SHIFT));
@@ -1028,6 +1027,13 @@ u64 dma_get_required_mask(struct device *dev)
 	}
 	return mask;
 }
+EXPORT_SYMBOL_GPL(dma_get_required_mask_from_max_pfn);
+
+#ifndef ARCH_HAS_DMA_GET_REQUIRED_MASK
+u64 dma_get_required_mask(struct device *dev)
+{
+	return dma_get_required_mask_from_max_pfn(dev);
+}
 EXPORT_SYMBOL_GPL(dma_get_required_mask);
 #endif
 
diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
index d5d3881..6e2fdfc 100644
--- a/include/linux/dma-mapping.h
+++ b/include/linux/dma-mapping.h
@@ -127,6 +127,7 @@ static inline int dma_coerce_mask_and_coherent(struct device *dev, u64 mask)
 	return dma_set_mask_and_coherent(dev, mask);
 }
 
+extern u64 dma_get_required_mask_from_max_pfn(struct device *dev);
 extern u64 dma_get_required_mask(struct device *dev);
 
 #ifndef set_arch_dma_coherent_ops
-- 
1.7.10.4

Re: [PATCH 1/4] dma: add dma_get_required_mask_from_max_pfn()

[Greg Kroah-Hartman]

On Fri, Dec 05, 2014 at 02:08:00PM +0000, David Vrabel wrote:
> A generic dma_get_required_mask() is useful even for architectures (such
> as ia64) that define ARCH_HAS_GET_REQUIRED_MASK.
> 
> Signed-off-by: David Vrabel <david.vrabel@citrix.com>
> Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
> ---
>  drivers/base/platform.c     |   10 ++++++++--

Is this why you sent this to me?  The x86 maintainers should handle this
patch set, not me for a tiny 8 lines in just one of the files, sorry.

greg k-h

Re: [PATCH 1/4] dma: add dma_get_required_mask_from_max_pfn()

[David Vrabel]

On 05/12/14 21:31, Greg Kroah-Hartman wrote:
> On Fri, Dec 05, 2014 at 02:08:00PM +0000, David Vrabel wrote:
>> A generic dma_get_required_mask() is useful even for architectures (such
>> as ia64) that define ARCH_HAS_GET_REQUIRED_MASK.
>>
>> Signed-off-by: David Vrabel <david.vrabel@citrix.com>
>> Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
>> ---
>>  drivers/base/platform.c     |   10 ++++++++--
> 
> Is this why you sent this to me?  The x86 maintainers should handle this
> patch set, not me for a tiny 8 lines in just one of the files, sorry.

This series will be merged via the Xen tree, but this patch still needs
your review or ack.

David

Re: [PATCH 1/4] dma: add dma_get_required_mask_from_max_pfn()

[Greg Kroah-Hartman]

On Mon, Dec 08, 2014 at 10:36:14AM +0000, David Vrabel wrote:
> On 05/12/14 21:31, Greg Kroah-Hartman wrote:
> > On Fri, Dec 05, 2014 at 02:08:00PM +0000, David Vrabel wrote:
> >> A generic dma_get_required_mask() is useful even for architectures (such
> >> as ia64) that define ARCH_HAS_GET_REQUIRED_MASK.
> >>
> >> Signed-off-by: David Vrabel <david.vrabel@citrix.com>
> >> Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
> >> ---
> >>  drivers/base/platform.c     |   10 ++++++++--
> > 
> > Is this why you sent this to me?  The x86 maintainers should handle this
> > patch set, not me for a tiny 8 lines in just one of the files, sorry.
> 
> This series will be merged via the Xen tree, but this patch still needs
> your review or ack.

How about waiting until after the merge window and resending it, asking
for that, instead of making me guess :)

greg k-h

[PATCH 2/4] ia64: use common dma_get_required_mask_from_pfn()

[David Vrabel]

Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Fenghua Yu <fenghua.yu@intel.com>
Cc: linux-ia64@vger.kernel.org
---
 arch/ia64/include/asm/machvec.h      |    2 +-
 arch/ia64/include/asm/machvec_init.h |    1 -
 arch/ia64/pci/pci.c                  |   20 --------------------
 3 files changed, 1 insertion(+), 22 deletions(-)

diff --git a/arch/ia64/include/asm/machvec.h b/arch/ia64/include/asm/machvec.h
index 9c39bdf..beaa47d 100644
--- a/arch/ia64/include/asm/machvec.h
+++ b/arch/ia64/include/asm/machvec.h
@@ -287,7 +287,7 @@ extern struct dma_map_ops *dma_get_ops(struct device *);
 # define platform_dma_get_ops		dma_get_ops
 #endif
 #ifndef platform_dma_get_required_mask
-# define  platform_dma_get_required_mask	ia64_dma_get_required_mask
+# define  platform_dma_get_required_mask	dma_get_required_mask_from_max_pfn
 #endif
 #ifndef platform_irq_to_vector
 # define platform_irq_to_vector		__ia64_irq_to_vector
diff --git a/arch/ia64/include/asm/machvec_init.h b/arch/ia64/include/asm/machvec_init.h
index 37a4698..ef964b2 100644
--- a/arch/ia64/include/asm/machvec_init.h
+++ b/arch/ia64/include/asm/machvec_init.h
@@ -3,7 +3,6 @@
 
 extern ia64_mv_send_ipi_t ia64_send_ipi;
 extern ia64_mv_global_tlb_purge_t ia64_global_tlb_purge;
-extern ia64_mv_dma_get_required_mask ia64_dma_get_required_mask;
 extern ia64_mv_irq_to_vector __ia64_irq_to_vector;
 extern ia64_mv_local_vector_to_irq __ia64_local_vector_to_irq;
 extern ia64_mv_pci_get_legacy_mem_t ia64_pci_get_legacy_mem;
diff --git a/arch/ia64/pci/pci.c b/arch/ia64/pci/pci.c
index 291a582..79da21b 100644
--- a/arch/ia64/pci/pci.c
+++ b/arch/ia64/pci/pci.c
@@ -791,26 +791,6 @@ static void __init set_pci_dfl_cacheline_size(void)
 	pci_dfl_cache_line_size = (1 << cci.pcci_line_size) / 4;
 }
 
-u64 ia64_dma_get_required_mask(struct device *dev)
-{
-	u32 low_totalram = ((max_pfn - 1) << PAGE_SHIFT);
-	u32 high_totalram = ((max_pfn - 1) >> (32 - PAGE_SHIFT));
-	u64 mask;
-
-	if (!high_totalram) {
-		/* convert to mask just covering totalram */
-		low_totalram = (1 << (fls(low_totalram) - 1));
-		low_totalram += low_totalram - 1;
-		mask = low_totalram;
-	} else {
-		high_totalram = (1 << (fls(high_totalram) - 1));
-		high_totalram += high_totalram - 1;
-		mask = (((u64)high_totalram) << 32) + 0xffffffff;
-	}
-	return mask;
-}
-EXPORT_SYMBOL_GPL(ia64_dma_get_required_mask);
-
 u64 dma_get_required_mask(struct device *dev)
 {
 	return platform_dma_get_required_mask(dev);
-- 
1.7.10.4

[PATCH 3/4] x86: allow dma_get_required_mask() to be overridden

[David Vrabel]

Use dma_ops->get_required_mask() if provided, defaulting to
dma_get_requried_mask_from_max_pfn().

This is needed on systems (such as Xen PV guests) where the DMA
address and the physical address are not equal.

ARCH_HAS_DMA_GET_REQUIRED_MASK is defined in asm/device.h instead of
asm/dma-mapping.h because linux/dma-mapping.h uses the define before
including asm/dma-mapping.h

Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
---
 arch/x86/include/asm/device.h |    2 ++
 arch/x86/kernel/pci-dma.c     |    8 ++++++++
 2 files changed, 10 insertions(+)

diff --git a/arch/x86/include/asm/device.h b/arch/x86/include/asm/device.h
index 03dd729..10bc628 100644
--- a/arch/x86/include/asm/device.h
+++ b/arch/x86/include/asm/device.h
@@ -13,4 +13,6 @@ struct dev_archdata {
 struct pdev_archdata {
 };
 
+#define ARCH_HAS_DMA_GET_REQUIRED_MASK
+
 #endif /* _ASM_X86_DEVICE_H */
diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
index a25e202..5154400 100644
--- a/arch/x86/kernel/pci-dma.c
+++ b/arch/x86/kernel/pci-dma.c
@@ -140,6 +140,14 @@ void dma_generic_free_coherent(struct device *dev, size_t size, void *vaddr,
 		free_pages((unsigned long)vaddr, get_order(size));
 }
 
+u64 dma_get_required_mask(struct device *dev)
+{
+	if (dma_ops->get_required_mask)
+		return dma_ops->get_required_mask(dev);
+	return dma_get_required_mask_from_max_pfn(dev);
+}
+EXPORT_SYMBOL_GPL(dma_get_required_mask);
+
 /*
  * See <Documentation/x86/x86_64/boot-options.txt> for the iommu kernel
  * parameter documentation.
-- 
1.7.10.4

[PATCH 4/4] x86/xen: assume a 64-bit DMA mask is required

[David Vrabel]

On a Xen PV guest the DMA addresses and physical addresses are not 1:1
(such as Xen PV guests) and the generic dma_get_required_mask() does
not return the correct mask (since it uses max_pfn).

Some device drivers (such as mptsas, mpt2sas) use
dma_get_required_mask() to set the device's DMA mask to allow them to
use only 32-bit DMA addresses in hardware structures.  This results in
unnecessary use of the SWIOTLB if DMA addresses are more than 32-bits,
impacting performance significantly.

We could base the DMA mask on the maximum MFN but:

a) The hypercall op to get the maximum MFN (XENMEM_maximum_ram_page)
will truncate the result to an int in 32-bit guests.

b) Future uses of the IOMMU in Xen may map frames at bus addresses
above the end of RAM.

So, just assume a 64-bit DMA mask is always required.

Signed-off-by: David Vrabel <david.vrabel@citrix.com>
---
 arch/x86/xen/pci-swiotlb-xen.c |    6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/arch/x86/xen/pci-swiotlb-xen.c b/arch/x86/xen/pci-swiotlb-xen.c
index 0e98e5d..35774f8 100644
--- a/arch/x86/xen/pci-swiotlb-xen.c
+++ b/arch/x86/xen/pci-swiotlb-xen.c
@@ -18,6 +18,11 @@
 
 int xen_swiotlb __read_mostly;
 
+static u64 xen_swiotlb_get_required_mask(struct device *dev)
+{
+	return DMA_BIT_MASK(64);
+}
+
 static struct dma_map_ops xen_swiotlb_dma_ops = {
 	.mapping_error = xen_swiotlb_dma_mapping_error,
 	.alloc = xen_swiotlb_alloc_coherent,
@@ -31,6 +36,7 @@ static struct dma_map_ops xen_swiotlb_dma_ops = {
 	.map_page = xen_swiotlb_map_page,
 	.unmap_page = xen_swiotlb_unmap_page,
 	.dma_supported = xen_swiotlb_dma_supported,
+	.get_required_mask = xen_swiotlb_get_required_mask,
 };
 
 /*
-- 
1.7.10.4

Re: [PATCHv5 0/4] dma, x86, xen: reduce SWIOTLB usage in Xen guests

[Greg Kroah-Hartman]

On Fri, Dec 05, 2014 at 02:07:59PM +0000, David Vrabel wrote:
> On systems where DMA addresses and physical addresses are not 1:1
> (such as Xen PV guests), the generic dma_get_required_mask() will not
> return the correct mask (since it uses max_pfn).
> 
> Some device drivers (such as mptsas, mpt2sas) use
> dma_get_required_mask() to set the device's DMA mask to allow them to use
> only 32-bit DMA addresses in hardware structures.  This results in
> unnecessary use of the SWIOTLB if DMA addresses are more than 32-bits,
> impacting performance significantly.
> 
> This series allows Xen PV guests to override the default
> dma_get_required_mask() with a more suitable one.
> 
> Changes in v5:
> - xen_swiotlb_get_required_mask() is x86 only.
> 
> Changes in v4:
> - Assume 64-bit mask is required.
> 
> Changes in v3:
> - fix off-by-one in xen_dma_get_required_mask()
> - split ia64 changes into separate patch.
> 
> Changes in v2:
> - split x86 and xen changes into separate patches
> 
> David

Why are you sending these to me?  Am I the DMA maintainer and forgot
about it?

/me digs in MAINTAINERS...

Nope, not me!  Patches are now deleted from my queue, go use
scripts/get_maintainer.pl like you should have done...

greg k-h