[PATCH v2] powerpc/powernv: Override dma_get_required_mask()

When using bypass window on IODA2, the incorrect DMA operations
"dma_iommu_ops" is used by devices. The device driver calls
dma_get_required_mask() to determine using 32-bits or 64-bits
bypass DMA window. Unfortunately, the returned DMA mask always
forces the driver to use 32-bits DMA window. The problem was
reported on the device as follows:

0004:03:00.0 0107: 1000:0087 (rev 05)
0004:03:00.0 Serial Attached SCSI controller: LSI Logic / Symbios \
             Logic SAS2308 PCI-Express Fusion-MPT SAS-2 (rev 05)

The patch fixes above issue by overriding dma_get_required_mask(),
which returns mask corresponding to bypass window base. Otherwise,
"dma_iommu_ops::get_required_mask" will be called to return mask
corresponding to 32-bits DMA window.

Reported-by: Murali N. Iyer <redacted>
Signed-off-by: Gavin Shan <redacted>
---
 arch/powerpc/include/asm/dma-mapping.h    |  1 +
 arch/powerpc/kernel/dma.c                 | 14 ++++++++++----
 arch/powerpc/platforms/powernv/pci-ioda.c | 23 +++++++++++++++++++++++
 arch/powerpc/platforms/powernv/pci.c      | 11 +++++++++++
 arch/powerpc/platforms/powernv/pci.h      |  2 ++
 arch/powerpc/platforms/powernv/powernv.h  |  6 ++++++
 arch/powerpc/platforms/powernv/setup.c    |  9 +++++++++
 7 files changed, 62 insertions(+), 4 deletions(-)

diff --git a/arch/powerpc/include/asm/dma-mapping.h b/arch/powerpc/include/asm/dma-mapping.h
index 150866b..894d538 100644
--- a/arch/powerpc/include/asm/dma-mapping.h
+++ b/arch/powerpc/include/asm/dma-mapping.h

@@ -135,6 +135,7 @@ static inline int dma_supported(struct device *dev, u64 mask)
 
 extern int dma_set_mask(struct device *dev, u64 dma_mask);
 extern int __dma_set_mask(struct device *dev, u64 dma_mask);
+extern u64 __dma_get_required_mask(struct device *dev);
 
 #define dma_alloc_coherent(d,s,h,f)	dma_alloc_attrs(d,s,h,f,NULL)
 

diff --git a/arch/powerpc/kernel/dma.c b/arch/powerpc/kernel/dma.c
index ee78f6e..210ff9d 100644
--- a/arch/powerpc/kernel/dma.c
+++ b/arch/powerpc/kernel/dma.c

@@ -202,6 +202,7 @@ int __dma_set_mask(struct device *dev, u64 dma_mask)
 	*dev->dma_mask = dma_mask;
 	return 0;
 }
+
 int dma_set_mask(struct device *dev, u64 dma_mask)
 {
 	if (ppc_md.dma_set_mask)

@@ -210,13 +211,10 @@ int dma_set_mask(struct device *dev, u64 dma_mask)
 }
 EXPORT_SYMBOL(dma_set_mask);
 
-u64 dma_get_required_mask(struct device *dev)
+u64 __dma_get_required_mask(struct device *dev)
 {
 	struct dma_map_ops *dma_ops = get_dma_ops(dev);
 
-	if (ppc_md.dma_get_required_mask)
-		return ppc_md.dma_get_required_mask(dev);
-
 	if (unlikely(dma_ops == NULL))
 		return 0;
 

@@ -225,6 +223,14 @@ u64 dma_get_required_mask(struct device *dev)
 
 	return DMA_BIT_MASK(8 * sizeof(dma_addr_t));
 }
+
+u64 dma_get_required_mask(struct device *dev)
+{
+	if (ppc_md.dma_get_required_mask)
+		return ppc_md.dma_get_required_mask(dev);
+
+	return __dma_get_required_mask(dev);
+}
 EXPORT_SYMBOL_GPL(dma_get_required_mask);
 
 static int __init dma_init(void)

diff --git a/arch/powerpc/platforms/powernv/pci-ioda.c b/arch/powerpc/platforms/powernv/pci-ioda.c
index 36b1a7a..380ebc9 100644
--- a/arch/powerpc/platforms/powernv/pci-ioda.c
+++ b/arch/powerpc/platforms/powernv/pci-ioda.c

@@ -890,6 +890,28 @@ static int pnv_pci_ioda_dma_set_mask(struct pnv_phb *phb,
 	return 0;
 }
 
+static u64 pnv_pci_ioda_dma_get_required_mask(struct pnv_phb *phb,
+					      struct pci_dev *pdev)
+{
+	struct pci_dn *pdn = pci_get_pdn(pdev);
+	struct pnv_ioda_pe *pe;
+	u64 end, mask;
+
+	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
+		return 0;
+
+	pe = &phb->ioda.pe_array[pdn->pe_number];
+	if (!pe->tce_bypass_enabled)
+		return __dma_get_required_mask(&pdev->dev);
+
+
+	end = pe->tce_bypass_base + memblock_end_of_DRAM();
+	mask = 1ULL << (fls64(end) - 1);
+	mask += mask - 1;
+
+	return mask;
+}
+
 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
 				   struct pci_bus *bus,
 				   bool add_to_iommu_group)

@@ -1782,6 +1804,7 @@ void __init pnv_pci_init_ioda_phb(struct device_node *np,
 	/* Setup TCEs */
 	phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
 	phb->dma_set_mask = pnv_pci_ioda_dma_set_mask;
+	phb->dma_get_required_mask = pnv_pci_ioda_dma_get_required_mask;
 
 	/* Setup shutdown function for kexec */
 	phb->shutdown = pnv_pci_ioda_shutdown;

diff --git a/arch/powerpc/platforms/powernv/pci.c b/arch/powerpc/platforms/powernv/pci.c
index b854b57..e9f509b 100644
--- a/arch/powerpc/platforms/powernv/pci.c
+++ b/arch/powerpc/platforms/powernv/pci.c

@@ -761,6 +761,17 @@ int pnv_pci_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
 	return __dma_set_mask(&pdev->dev, dma_mask);
 }
 
+u64 pnv_pci_dma_get_required_mask(struct pci_dev *pdev)
+{
+	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
+	struct pnv_phb *phb = hose->private_data;
+
+	if (phb && phb->dma_get_required_mask)
+		return phb->dma_get_required_mask(phb, pdev);
+
+	return __dma_get_required_mask(&pdev->dev);
+}
+
 void pnv_pci_shutdown(void)
 {
 	struct pci_controller *hose;

diff --git a/arch/powerpc/platforms/powernv/pci.h b/arch/powerpc/platforms/powernv/pci.h
index b54af34..aa95fcb 100644
--- a/arch/powerpc/platforms/powernv/pci.h
+++ b/arch/powerpc/platforms/powernv/pci.h

@@ -124,6 +124,8 @@ struct pnv_phb {
 	void (*dma_dev_setup)(struct pnv_phb *phb, struct pci_dev *pdev);
 	int (*dma_set_mask)(struct pnv_phb *phb, struct pci_dev *pdev,
 			    u64 dma_mask);
+	u64 (*dma_get_required_mask)(struct pnv_phb *phb,
+				     struct pci_dev *pdev);
 	void (*fixup_phb)(struct pci_controller *hose);
 	u32 (*bdfn_to_pe)(struct pnv_phb *phb, struct pci_bus *bus, u32 devfn);
 	void (*shutdown)(struct pnv_phb *phb);

diff --git a/arch/powerpc/platforms/powernv/powernv.h b/arch/powerpc/platforms/powernv/powernv.h
index 75501bf..6c8e2d1 100644
--- a/arch/powerpc/platforms/powernv/powernv.h
+++ b/arch/powerpc/platforms/powernv/powernv.h

@@ -13,6 +13,7 @@ struct pci_dev;
 extern void pnv_pci_init(void);
 extern void pnv_pci_shutdown(void);
 extern int pnv_pci_dma_set_mask(struct pci_dev *pdev, u64 dma_mask);
+extern u64 pnv_pci_dma_get_required_mask(struct pci_dev *pdev);
 #else
 static inline void pnv_pci_init(void) { }
 static inline void pnv_pci_shutdown(void) { }

@@ -21,6 +22,11 @@ static inline int pnv_pci_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
 {
 	return -ENODEV;
 }
+
+static inline u64 pnv_pci_dma_get_required_mask(struct pci_dev *pdev)
+{
+	return 0;
+}
 #endif
 
 extern void pnv_lpc_init(void);

diff --git a/arch/powerpc/platforms/powernv/setup.c b/arch/powerpc/platforms/powernv/setup.c
index 5a0e2dc..0cb3a07 100644
--- a/arch/powerpc/platforms/powernv/setup.c
+++ b/arch/powerpc/platforms/powernv/setup.c

@@ -173,6 +173,14 @@ static int pnv_dma_set_mask(struct device *dev, u64 dma_mask)
 	return __dma_set_mask(dev, dma_mask);
 }
 
+static u64 pnv_dma_get_required_mask(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return pnv_pci_dma_get_required_mask(to_pci_dev(dev));
+
+	return __dma_get_required_mask(dev);
+}
+
 static void pnv_shutdown(void)
 {
 	/* Let the PCI code clear up IODA tables */

@@ -335,6 +343,7 @@ define_machine(powernv) {
 	.power_save             = power7_idle,
 	.calibrate_decr		= generic_calibrate_decr,
 	.dma_set_mask		= pnv_dma_set_mask,
+	.dma_get_required_mask	= pnv_dma_get_required_mask,
 #ifdef CONFIG_KEXEC
 	.kexec_cpu_down		= pnv_kexec_cpu_down,
 #endif

-- 
1.8.3.2

On Thu, 2014-09-25 at 17:24 +1000, Gavin Shan wrote:

When using bypass window on IODA2, the incorrect DMA operations
"dma_iommu_ops" is used by devices. The device driver calls
dma_get_required_mask() to determine using 32-bits or 64-bits
bypass DMA window. Unfortunately, the returned DMA mask always
forces the driver to use 32-bits DMA window. The problem was
reported on the device as follows:

I would write the above comment a bit differently:

The dma_get_required_mask() function is used by some drivers to
query the platform about what DMA mask is needed to cover all of memory.
This is a bit of a strange semantic when we have to chose between iommu
translation or bypass, but essentially what it means is "what DMA mask
will give best performances".

Currently, our iommu backend always returns a 32-bit mask here, we don't
do anything special to it when we have bypass available. This causes
some drivers to chose a 32-bit mask, thus losing the ability to use the
bypass window, thinking this is more efficient.

This patch adds an override of that function in order to, instead,
return a 64-bit mask whenever a bypass window is available in order for
drivers to prefer this configuration.

 ... or something along those lines.

On Fri, Sep 26, 2014 at 07:51:56AM +1000, Benjamin Herrenschmidt wrote:

On Thu, 2014-09-25 at 17:24 +1000, Gavin Shan wrote:

quoted

When using bypass window on IODA2, the incorrect DMA operations
"dma_iommu_ops" is used by devices. The device driver calls
dma_get_required_mask() to determine using 32-bits or 64-bits
bypass DMA window. Unfortunately, the returned DMA mask always
forces the driver to use 32-bits DMA window. The problem was
reported on the device as follows:

I would write the above comment a bit differently:

The dma_get_required_mask() function is used by some drivers to
query the platform about what DMA mask is needed to cover all of memory.
This is a bit of a strange semantic when we have to chose between iommu
translation or bypass, but essentially what it means is "what DMA mask
will give best performances".

Currently, our iommu backend always returns a 32-bit mask here, we don't
do anything special to it when we have bypass available. This causes
some drivers to chose a 32-bit mask, thus losing the ability to use the
bypass window, thinking this is more efficient.

This patch adds an override of that function in order to, instead,
return a 64-bit mask whenever a bypass window is available in order for
drivers to prefer this configuration.

... or something along those lines.

Thanks, Ben. I'll fold your comments into next revision.

Thanks,
Gavin

quoted

0004:03:00.0 0107: 1000:0087 (rev 05)
0004:03:00.0 Serial Attached SCSI controller: LSI Logic / Symbios \
             Logic SAS2308 PCI-Express Fusion-MPT SAS-2 (rev 05)

The patch fixes above issue by overriding dma_get_required_mask(),
which returns mask corresponding to bypass window base. Otherwise,
"dma_iommu_ops::get_required_mask" will be called to return mask
corresponding to 32-bits DMA window.

Reported-by: Murali N. Iyer <redacted>
Signed-off-by: Gavin Shan <redacted>
---
 arch/powerpc/include/asm/dma-mapping.h    |  1 +
 arch/powerpc/kernel/dma.c                 | 14 ++++++++++----
 arch/powerpc/platforms/powernv/pci-ioda.c | 23 +++++++++++++++++++++++
 arch/powerpc/platforms/powernv/pci.c      | 11 +++++++++++
 arch/powerpc/platforms/powernv/pci.h      |  2 ++
 arch/powerpc/platforms/powernv/powernv.h  |  6 ++++++
 arch/powerpc/platforms/powernv/setup.c    |  9 +++++++++
 7 files changed, 62 insertions(+), 4 deletions(-)

diff --git a/arch/powerpc/include/asm/dma-mapping.h b/arch/powerpc/include/asm/dma-mapping.h
index 150866b..894d538 100644
--- a/arch/powerpc/include/asm/dma-mapping.h
+++ b/arch/powerpc/include/asm/dma-mapping.h

@@ -135,6 +135,7 @@ static inline int dma_supported(struct device *dev, u64 mask)
 
 extern int dma_set_mask(struct device *dev, u64 dma_mask);
 extern int __dma_set_mask(struct device *dev, u64 dma_mask);
+extern u64 __dma_get_required_mask(struct device *dev);
 
 #define dma_alloc_coherent(d,s,h,f)	dma_alloc_attrs(d,s,h,f,NULL)
 

diff --git a/arch/powerpc/kernel/dma.c b/arch/powerpc/kernel/dma.c
index ee78f6e..210ff9d 100644
--- a/arch/powerpc/kernel/dma.c
+++ b/arch/powerpc/kernel/dma.c

@@ -202,6 +202,7 @@ int __dma_set_mask(struct device *dev, u64 dma_mask)
 	*dev->dma_mask = dma_mask;
 	return 0;
 }
+
 int dma_set_mask(struct device *dev, u64 dma_mask)
 {
 	if (ppc_md.dma_set_mask)

@@ -210,13 +211,10 @@ int dma_set_mask(struct device *dev, u64 dma_mask)
 }
 EXPORT_SYMBOL(dma_set_mask);
 
-u64 dma_get_required_mask(struct device *dev)
+u64 __dma_get_required_mask(struct device *dev)
 {
 	struct dma_map_ops *dma_ops = get_dma_ops(dev);
 
-	if (ppc_md.dma_get_required_mask)
-		return ppc_md.dma_get_required_mask(dev);
-
 	if (unlikely(dma_ops == NULL))
 		return 0;
 

@@ -225,6 +223,14 @@ u64 dma_get_required_mask(struct device *dev)
 
 	return DMA_BIT_MASK(8 * sizeof(dma_addr_t));
 }
+
+u64 dma_get_required_mask(struct device *dev)
+{
+	if (ppc_md.dma_get_required_mask)
+		return ppc_md.dma_get_required_mask(dev);
+
+	return __dma_get_required_mask(dev);
+}
 EXPORT_SYMBOL_GPL(dma_get_required_mask);
 
 static int __init dma_init(void)

diff --git a/arch/powerpc/platforms/powernv/pci-ioda.c b/arch/powerpc/platforms/powernv/pci-ioda.c
index 36b1a7a..380ebc9 100644
--- a/arch/powerpc/platforms/powernv/pci-ioda.c
+++ b/arch/powerpc/platforms/powernv/pci-ioda.c

@@ -890,6 +890,28 @@ static int pnv_pci_ioda_dma_set_mask(struct pnv_phb *phb,
 	return 0;
 }
 
+static u64 pnv_pci_ioda_dma_get_required_mask(struct pnv_phb *phb,
+					      struct pci_dev *pdev)
+{
+	struct pci_dn *pdn = pci_get_pdn(pdev);
+	struct pnv_ioda_pe *pe;
+	u64 end, mask;
+
+	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
+		return 0;
+
+	pe = &phb->ioda.pe_array[pdn->pe_number];
+	if (!pe->tce_bypass_enabled)
+		return __dma_get_required_mask(&pdev->dev);
+
+
+	end = pe->tce_bypass_base + memblock_end_of_DRAM();
+	mask = 1ULL << (fls64(end) - 1);
+	mask += mask - 1;
+
+	return mask;
+}
+
 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
 				   struct pci_bus *bus,
 				   bool add_to_iommu_group)

@@ -1782,6 +1804,7 @@ void __init pnv_pci_init_ioda_phb(struct device_node *np,
 	/* Setup TCEs */
 	phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
 	phb->dma_set_mask = pnv_pci_ioda_dma_set_mask;
+	phb->dma_get_required_mask = pnv_pci_ioda_dma_get_required_mask;
 
 	/* Setup shutdown function for kexec */
 	phb->shutdown = pnv_pci_ioda_shutdown;

diff --git a/arch/powerpc/platforms/powernv/pci.c b/arch/powerpc/platforms/powernv/pci.c
index b854b57..e9f509b 100644
--- a/arch/powerpc/platforms/powernv/pci.c
+++ b/arch/powerpc/platforms/powernv/pci.c

@@ -761,6 +761,17 @@ int pnv_pci_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
 	return __dma_set_mask(&pdev->dev, dma_mask);
 }
 
+u64 pnv_pci_dma_get_required_mask(struct pci_dev *pdev)
+{
+	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
+	struct pnv_phb *phb = hose->private_data;
+
+	if (phb && phb->dma_get_required_mask)
+		return phb->dma_get_required_mask(phb, pdev);
+
+	return __dma_get_required_mask(&pdev->dev);
+}
+
 void pnv_pci_shutdown(void)
 {
 	struct pci_controller *hose;

diff --git a/arch/powerpc/platforms/powernv/pci.h b/arch/powerpc/platforms/powernv/pci.h
index b54af34..aa95fcb 100644
--- a/arch/powerpc/platforms/powernv/pci.h
+++ b/arch/powerpc/platforms/powernv/pci.h

@@ -124,6 +124,8 @@ struct pnv_phb {
 	void (*dma_dev_setup)(struct pnv_phb *phb, struct pci_dev *pdev);
 	int (*dma_set_mask)(struct pnv_phb *phb, struct pci_dev *pdev,
 			    u64 dma_mask);
+	u64 (*dma_get_required_mask)(struct pnv_phb *phb,
+				     struct pci_dev *pdev);
 	void (*fixup_phb)(struct pci_controller *hose);
 	u32 (*bdfn_to_pe)(struct pnv_phb *phb, struct pci_bus *bus, u32 devfn);
 	void (*shutdown)(struct pnv_phb *phb);

diff --git a/arch/powerpc/platforms/powernv/powernv.h b/arch/powerpc/platforms/powernv/powernv.h
index 75501bf..6c8e2d1 100644
--- a/arch/powerpc/platforms/powernv/powernv.h
+++ b/arch/powerpc/platforms/powernv/powernv.h

@@ -13,6 +13,7 @@ struct pci_dev;
 extern void pnv_pci_init(void);
 extern void pnv_pci_shutdown(void);
 extern int pnv_pci_dma_set_mask(struct pci_dev *pdev, u64 dma_mask);
+extern u64 pnv_pci_dma_get_required_mask(struct pci_dev *pdev);
 #else
 static inline void pnv_pci_init(void) { }
 static inline void pnv_pci_shutdown(void) { }

@@ -21,6 +22,11 @@ static inline int pnv_pci_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
 {
 	return -ENODEV;
 }
+
+static inline u64 pnv_pci_dma_get_required_mask(struct pci_dev *pdev)
+{
+	return 0;
+}
 #endif
 
 extern void pnv_lpc_init(void);

diff --git a/arch/powerpc/platforms/powernv/setup.c b/arch/powerpc/platforms/powernv/setup.c
index 5a0e2dc..0cb3a07 100644
--- a/arch/powerpc/platforms/powernv/setup.c
+++ b/arch/powerpc/platforms/powernv/setup.c

@@ -173,6 +173,14 @@ static int pnv_dma_set_mask(struct device *dev, u64 dma_mask)
 	return __dma_set_mask(dev, dma_mask);
 }
 
+static u64 pnv_dma_get_required_mask(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return pnv_pci_dma_get_required_mask(to_pci_dev(dev));
+
+	return __dma_get_required_mask(dev);
+}
+
 static void pnv_shutdown(void)
 {
 	/* Let the PCI code clear up IODA tables */

@@ -335,6 +343,7 @@ define_machine(powernv) {
 	.power_save             = power7_idle,
 	.calibrate_decr		= generic_calibrate_decr,
 	.dma_set_mask		= pnv_dma_set_mask,
+	.dma_get_required_mask	= pnv_dma_get_required_mask,
 #ifdef CONFIG_KEXEC
 	.kexec_cpu_down		= pnv_kexec_cpu_down,
 #endif