Skip to contentSkip to conversation

ppc/cell: beat dma ops cleanup

PATCH In review Nishanth Aravamudan 8 participants 32 messages Sep 15, 2010 – Dec 1, 2010 2 messages not in the archive

Subsystems (2)
lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw

PATCH 1/15 ppc: fix return type of BUID_{HI,LO} macros

BUID_HI and BUID_LO are used to pass data to call_rtas, which expects ints or u32s. But the macro doesn't cast the return, so the result is still u64. Use the upper_32_bits and lower_32_bits macros that have been added to kernel.h.

Found by getting printf format errors trying to debug print the args, no actual code change for 64 bit kernels where the macros are actually used.

---
 arch/powerpc/include/asm/ppc-pci.h |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
arch/powerpc/include/asm/ppc-pci.h +2 -2
index 42fdff0..43268f1 100644
@@ -28,8 +28,8 @@ extern void find_and_init_phbs(void);
 extern struct pci_dev *isa_bridge_pcidev;	/* may be NULL if no ISA bus */
 
 /** Bus Unit ID macros; get low and hi 32-bits of the 64-bit BUID */
-#define BUID_HI(buid) ((buid) >> 32)
-#define BUID_LO(buid) ((buid) & 0xffffffff)
+#define BUID_HI(buid) upper_32_bits(buid)
+#define BUID_LO(buid) lower_32_bits(buid)
 
 /* PCI device_node operations */
 struct device_node;
Signature · 1 line
1.7.0.4
Acked-by on v1 1/15
#2 lore raw

Re: [PATCH 01/15] ppc: fix return type of BUID_{HI,LO} macros

I'm guessing this worked up til now because the rtas_call function prototyp= e
was telling compiler to cast these to 32-bit before passing them as args. (and since these would still get passed as one arg per 64-bit reg, it still wouldn't go wrong.)

What I'm wondering about is why there was no compiler warning about an implicit cast of a 64-bit int to a 32-bit int? Surely, this is something t= hat
should be warned about!

-- Linas

arch/powerpc/include/as=

Quoted 16 lines · Nishanth Aravamudan

BUID_HI and BUID_LO are used to pass data to call_rtas, which expects ints or u32s. But the macro doesn't cast the return, so the result is still u64. Use the upper_32_bits and lower_32_bits macros that have been added to kernel.h.

Found by getting printf format errors trying to debug print the args, no actual code change for 64 bit kernels where the macros are actually used.

---
=C2=A0arch/powerpc/include/asm/ppc-pci.h | =C2=A0 =C2=A04 ++-- =C2=A01 files changed, 2 insertions(+), 2 deletions(-)

arch/powerpc/include/as= renamed from arch/powerpc/include/asm/ppc-pci.h +0 -0

m/ppc-pci.h

Quoted 5 lines

index 42fdff0..43268f1 100644

arch/powerpc/include/asm/ppc-pci.h +0 -0
@@ -28,8 +28,8 @@ extern void find_and_init_phbs(void);

=C2=A0extern struct pci_dev *isa_bridge_pcidev; =C2=A0 =C2=A0 =C2=A0/* ma=

y be NULL if no ISA bus */

Quoted 2 lines

=C2=A0/** Bus Unit ID macros; get low and hi 32-bits of the 64-bit BUID *=

/

Quoted 11 lines
arch/powerpc/include/asm/ppc-pci.h +2 -2
-#define BUID_HI(buid) ((buid) >> 32)
-#define BUID_LO(buid) ((buid) & 0xffffffff)
+#define BUID_HI(buid) upper_32_bits(buid)
+#define BUID_LO(buid) lower_32_bits(buid)

=C2=A0/* PCI device_node operations */
=C2=A0struct device_node;
--
1.7.0.4

#3 lore raw

Re: [PATCH 01/15] ppc: fix return type of BUID_{HI,LO} macros

Quoted 10 lines · On Thu, 16 Sep 2010 17:54:46 -0500 Linas Vepstas <[email protected]> wrote:

I'm guessing this worked up til now because the rtas_call function prototype was telling compiler to cast these to 32-bit before passing them as args. (and since these would still get passed as one arg per 64-bit reg, it still wouldn't go wrong.)

What I'm wondering about is why there was no compiler warning about an implicit cast of a 64-bit int to a 32-bit int? Surely, this is something that should be warned about!

-Wconversion enables this.

There'd be a lot of false positives to squash with that enabled -- and it might not be worth the resulting explosion in casts.

-Scott

#4 lore raw

PATCH 2/15 ppc64: fix dma_iommu_dma_supported compare

In reply to a message not in the archive

The table offset is in entries, each of which imply a dma address of an IOMMU page.

Also, we should check the device can reach the whole IOMMU table.

---
 arch/powerpc/kernel/dma-iommu.c |   21 +++++++++++----------
 1 files changed, 11 insertions(+), 10 deletions(-)
arch/powerpc/kernel/dma-iommu.c +11 -10
index 37771a5..6e54a0f 100644
@@ -74,16 +74,17 @@ static int dma_iommu_dma_supported(struct device *dev, u64 mask)
 {
 	struct iommu_table *tbl = get_iommu_table_base(dev);
 
-	if (!tbl || tbl->it_offset > mask) {
-		printk(KERN_INFO
-		       "Warning: IOMMU offset too big for device mask\n");
-		if (tbl)
-			printk(KERN_INFO
-			       "mask: 0x%08llx, table offset: 0x%08lx\n",
-				mask, tbl->it_offset);
-		else
-			printk(KERN_INFO "mask: 0x%08llx, table unavailable\n",
-				mask);
+	if (!tbl) {
+		dev_info(dev, "Warning: IOMMU dma not supported: mask 0x%08llx"
+			", table unavailable\n", mask);
+		return 0;
+	}
+
+	if ((tbl->it_offset + tbl->it_size) > (mask >> IOMMU_PAGE_SHIFT)) {
+		dev_info(dev, "Warning: IOMMU window too big for device mask\n");
+		dev_info(dev, "mask: 0x%08llx, table end: 0x%08lx\n",
+				mask, (tbl->it_offset + tbl->it_size) <<
+				IOMMU_PAGE_SHIFT);
 		return 0;
 	} else
 		return 1;
Signature · 1 line
1.7.0.4
#5 lore raw

PATCH 3/15 ppc64 iommu: fix check for direct DMA support

In reply to a message not in the archive

The current check is wrong because it does not take the DMA offset intot account, and in the case of a driver which doesn't actually support 64bits would falsely report that device as working.

---
 arch/powerpc/kernel/dma.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/powerpc/kernel/dma.c +1 -1
index 84d6367..494ab12 100644
@@ -89,7 +89,7 @@ static int dma_direct_dma_supported(struct device *dev, u64 mask)
 	/* Could be improved so platforms can set the limit in case
 	 * they have limited DMA windows
 	 */
-	return mask >= (memblock_end_of_DRAM() - 1);
+	return mask >= get_dma_offset(dev) + (memblock_end_of_DRAM() - 1);
 #else
 	return 1;
 #endif
Signature · 1 line
1.7.0.4
#6 lore raw

PATCH 4/15 vio: put device on device_register failure

In reply to a message not in the archive

The kernel doc for device_register (and device_initialize) very clearly state to call put_device not kfree after calling, even on error.

---
 arch/powerpc/kernel/vio.c |    3 +--
 1 files changed, 1 insertions(+), 2 deletions(-)
arch/powerpc/kernel/vio.c +1 -2
index fa3469d..72db4b0 100644
@@ -1254,8 +1254,7 @@ struct vio_dev *vio_register_device_node(struct device_node *of_node)
 	if (device_register(&viodev->dev)) {
 		printk(KERN_ERR "%s: failed to register device %s\n",
 				__func__, dev_name(&viodev->dev));
-		/* XXX free TCE table */
-		kfree(viodev);
+		put_device(&viodev->dev);
 		return NULL;
 	}
 
Signature · 1 line
1.7.0.4
Acked-by on v1 4/15
#7 lore raw

Re: [PATCH 04/15] vio: put device on device_register failure

On Wed, Sep 15, 2010 at 12:05 PM, Nishanth Aravamudan <[email protected]> wro= te:

Quoted 5 lines

The kernel doc for device_register (and device_initialize) very clearly state to call put_device not kfree after calling, even on error.

arch/powerpc/kernel/vio.c struct vio_dev *vio_register_device_node(struct dev=

Quoted 9 lines

---
=A0arch/powerpc/kernel/vio.c | =A0 =A03 +-- =A01 files changed, 1 insertions(+), 2 deletions(-)

arch/powerpc/kernel/vio.c +0 -0
index fa3469d..72db4b0 100644
@@ -1254,8 +1254,7 @@ struct vio_dev *vio_register_device_node(struct dev=

ice_node *of_node)

Quoted 2 lines
arch/powerpc/kernel/vio.c +0 -0
 =A0 =A0 =A0 =A0if (device_register(&viodev->dev)) {

=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0printk(KERN_ERR "%s: failed to register de=

vice %s\n",

Quoted 1 line

=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0__func__, =

dev_name(&viodev->dev));

Quoted 10 lines

- =A0 =A0 =A0 =A0 =A0 =A0 =A0 /* XXX free TCE table */
- =A0 =A0 =A0 =A0 =A0 =A0 =A0 kfree(viodev);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 put_device(&viodev->dev); =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0return NULL; =A0 =A0 =A0 =A0}

--
1.7.0.4

--=20
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.

#8 lore raw

PATCH 5/15 viobus: free TCE table on device release

In reply to a message not in the archive

Release the TCE table as the XXX suggests, except on FW_FEATURE_ISERIES, where the tables are allocated globally and reused.

---
 arch/powerpc/kernel/vio.c |    7 ++++++-
 1 files changed, 6 insertions(+), 1 deletions(-)
arch/powerpc/kernel/vio.c +6 -1
index 72db4b0..d692989 100644
@@ -1184,7 +1184,12 @@ EXPORT_SYMBOL(vio_unregister_driver);
 /* vio_dev refcount hit 0 */
 static void __devinit vio_dev_release(struct device *dev)
 {
-	/* XXX should free TCE table */
+	struct iommu_table *tbl = get_iommu_table_base(dev);
+
+	/* iSeries uses a common table for all vio devices */
+	if (!firmware_has_feature(FW_FEATURE_ISERIES) && tbl)
+		iommu_free_table(tbl, dev->of_node ?
+			dev->of_node->full_name : dev_name(dev));
 	of_node_put(dev->of_node);
 	kfree(to_vio_dev(dev));
 }
Signature · 1 line
1.7.0.4
#9 lore raw

PATCH 6/15 pseries/dlpar: use kmemdup

In reply to a message not in the archive

While looking at some code paths I came across this code that zeros memory then copies over the entire length.

---
 arch/powerpc/platforms/pseries/dlpar.c |    3 +--
 1 files changed, 1 insertions(+), 2 deletions(-)
arch/powerpc/platforms/pseries/dlpar.c +1 -2
index 72d8054..6004c81 100644
@@ -55,13 +55,12 @@ static struct property *dlpar_parse_cc_property(struct cc_workarea *ccwa)
 
 	prop->length = ccwa->prop_length;
 	value = (char *)ccwa + ccwa->prop_offset;
-	prop->value = kzalloc(prop->length, GFP_KERNEL);
+	prop->value = kmemdup(value, prop->length, GFP_KERNEL);
 	if (!prop->value) {
 		dlpar_free_cc_property(prop);
 		return NULL;
 	}
 
-	memcpy(prop->value, value, prop->length);
 	return prop;
 }
 
Signature · 1 line
1.7.0.4
#10 lore raw

PATCH 7/15 ppc: pci-common cleanup

In reply to a message not in the archive

Use set_dma_ops and remove unused oddly-named temp pointer sd.

---
 arch/powerpc/kernel/pci-common.c |    4 +---
 1 files changed, 1 insertions(+), 3 deletions(-)
arch/powerpc/kernel/pci-common.c +1 -3
index 9021c4a..10a44e6 100644
@@ -1090,8 +1090,6 @@ void __devinit pcibios_setup_bus_devices(struct pci_bus *bus)
 		 bus->number, bus->self ? pci_name(bus->self) : "PHB");
 
 	list_for_each_entry(dev, &bus->devices, bus_list) {
-		struct dev_archdata *sd = &dev->dev.archdata;
-
 		/* Cardbus can call us to add new devices to a bus, so ignore
 		 * those who are already fully discovered
 		 */
@@ -1107,7 +1105,7 @@ void __devinit pcibios_setup_bus_devices(struct pci_bus *bus)
 		set_dev_node(&dev->dev, pcibus_to_node(dev->bus));
 
 		/* Hook up default DMA ops */
-		sd->dma_ops = pci_dma_ops;
+		set_dma_ops(&dev->dev, pci_dma_ops);
 		set_dma_offset(&dev->dev, PCI_DRAM_OFFSET);
 
 		/* Additional platform DMA/iommu setup */
Signature · 1 line
1.7.0.4
Acked-by on v1 7/15
#11 lore raw

Re: [PATCH 07/15] ppc: pci-common cleanup

On Wed, Sep 15, 2010 at 12:05 PM, Nishanth Aravamudan <[email protected]> wro= te:

Quoted 4 lines

Use set_dma_ops and remove unused oddly-named temp pointer sd.

arch/powerpc/kernel/pci-c=

Quoted 5 lines

---
=A0arch/powerpc/kernel/pci-common.c | =A0 =A04 +--- =A01 files changed, 1 insertions(+), 3 deletions(-)

arch/powerpc/kernel/pci-c= renamed from arch/powerpc/kernel/pci-common.c +0 -0

ommon.c

arch/powerpc/kernel/pci-common.c void __devinit pcibios_setup_bus_devices(struct pci=

Quoted 4 lines

index 9021c4a..10a44e6 100644

arch/powerpc/kernel/pci-common.c +0 -0
@@ -1090,8 +1090,6 @@ void __devinit pcibios_setup_bus_devices(struct pci=

_bus *bus)

Quoted 1 line

=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 bus->number, bus->self ? pci_name(bus->se=

lf) : "PHB");

Quoted 3 lines

=A0 =A0 =A0 =A0list_for_each_entry(dev, &bus->devices, bus_list) {
- =A0 =A0 =A0 =A0 =A0 =A0 =A0 struct dev_archdata *sd =3D &dev->dev.archd=

ata;

Quoted 2 lines

-
=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0/* Cardbus can call us to add new devices =

to a bus, so ignore

arch/powerpc/kernel/pci-common.c void __devinit pcibios_setup_bus_devices(struct pci=

Quoted 3 lines

=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 * those who are already fully discovered =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 */

arch/powerpc/kernel/pci-common.c +0 -0
@@ -1107,7 +1105,7 @@ void __devinit pcibios_setup_bus_devices(struct pci=

_bus *bus)

Quoted 1 line

=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0set_dev_node(&dev->dev, pcibus_to_node(dev=

->bus));

Quoted 5 lines

=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0/* Hook up default DMA ops */
- =A0 =A0 =A0 =A0 =A0 =A0 =A0 sd->dma_ops =3D pci_dma_ops;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 set_dma_ops(&dev->dev, pci_dma_ops); =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0set_dma_offset(&dev->dev, PCI_DRAM_OFFSET)=

;

Quoted 6 lines

=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0/* Additional platform DMA/iommu setup */ --
1.7.0.4

--=20
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.

#12 lore raw

PATCH 9/15 ppc/vio: use dma ops helpers

In reply to a message not in the archive

Use the set_dma_ops helper. Instead of modifying vio_dma_mapping_ops, just create a trivial wrapper for dma_supported.

---
 arch/powerpc/kernel/vio.c |   11 ++++++++---
 1 files changed, 8 insertions(+), 3 deletions(-)
arch/powerpc/kernel/vio.c +8 -3
index d692989..3c3083f 100644
@@ -602,6 +602,11 @@ static void vio_dma_iommu_unmap_sg(struct device *dev,
 	vio_cmo_dealloc(viodev, alloc_size);
 }
 
+static int vio_dma_iommu_dma_supported(struct device *dev, u64 mask)
+{
+        return dma_iommu_ops.dma_supported(dev, mask);
+}
+
 struct dma_map_ops vio_dma_mapping_ops = {
 	.alloc_coherent = vio_dma_iommu_alloc_coherent,
 	.free_coherent  = vio_dma_iommu_free_coherent,
@@ -609,6 +614,7 @@ struct dma_map_ops vio_dma_mapping_ops = {
 	.unmap_sg       = vio_dma_iommu_unmap_sg,
 	.map_page       = vio_dma_iommu_map_page,
 	.unmap_page     = vio_dma_iommu_unmap_page,
+	.dma_supported  = vio_dma_iommu_dma_supported,
 
 };
 
@@ -860,8 +866,7 @@ static void vio_cmo_bus_remove(struct vio_dev *viodev)
 
 static void vio_cmo_set_dma_ops(struct vio_dev *viodev)
 {
-	vio_dma_mapping_ops.dma_supported = dma_iommu_ops.dma_supported;
-	viodev->dev.archdata.dma_ops = &vio_dma_mapping_ops;
+	set_dma_ops(&viodev->dev, &vio_dma_mapping_ops);
 }
 
 /**
@@ -1246,7 +1251,7 @@ struct vio_dev *vio_register_device_node(struct device_node *of_node)
 	if (firmware_has_feature(FW_FEATURE_CMO))
 		vio_cmo_set_dma_ops(viodev);
 	else
-		viodev->dev.archdata.dma_ops = &dma_iommu_ops;
+		set_dma_ops(&viodev->dev, &dma_iommu_ops);
 	set_iommu_table_base(&viodev->dev, vio_build_iommu_table(viodev));
 	set_dev_node(&viodev->dev, of_node_to_nid(of_node));
 
Signature · 1 line
1.7.0.4
#13 lore raw

Re: [PATCH 09/15] ppc/vio: use dma ops helpers

On Wed, Sep 15, 2010 at 12:05 PM, Nishanth Aravamudan <[email protected]> wro= te:

Quoted 5 lines

Use the set_dma_ops helper. Instead of modifying vio_dma_mapping_ops, just create a trivial wrapper for dma_supported.

Looks right to me.

arch/powerpc/kernel/vio.c static void vio_dma_iommu_unmap_sg(struct device *de=

Quoted 9 lines

---
=A0arch/powerpc/kernel/vio.c | =A0 11 ++++++++--- =A01 files changed, 8 insertions(+), 3 deletions(-)

arch/powerpc/kernel/vio.c +0 -0
index d692989..3c3083f 100644
@@ -602,6 +602,11 @@ static void vio_dma_iommu_unmap_sg(struct device *de=

v,

arch/powerpc/kernel/vio.c static void vio_cmo_bus_remove(struct vio_dev *viodev=

Quoted 20 lines
arch/powerpc/kernel/vio.c +6 -0
 =A0 =A0 =A0 =A0vio_cmo_dealloc(viodev, alloc_size);
 =A0}
 
+static int vio_dma_iommu_dma_supported(struct device *dev, u64 mask)
+{
+ =A0 =A0 =A0 =A0return dma_iommu_ops.dma_supported(dev, mask);
+}
+
 =A0struct dma_map_ops vio_dma_mapping_ops =3D {
 =A0 =A0 =A0 =A0.alloc_coherent =3D vio_dma_iommu_alloc_coherent,
 =A0 =A0 =A0 =A0.free_coherent =A0=3D vio_dma_iommu_free_coherent,
@@ -609,6 +614,7 @@ struct dma_map_ops vio_dma_mapping_ops =3D {
 =A0 =A0 =A0 =A0.unmap_sg =A0 =A0 =A0 =3D vio_dma_iommu_unmap_sg,
 =A0 =A0 =A0 =A0.map_page =A0 =A0 =A0 =3D vio_dma_iommu_map_page,
 =A0 =A0 =A0 =A0.unmap_page =A0 =A0 =3D vio_dma_iommu_unmap_page,
+ =A0 =A0 =A0 .dma_supported =A0=3D vio_dma_iommu_dma_supported,
 
 =A0};
 
@@ -860,8 +866,7 @@ static void vio_cmo_bus_remove(struct vio_dev *viodev=

)

arch/powerpc/kernel/vio.c static void vio_cmo_bus_remove(struct vio_dev *viodev= · line 862

Quoted 4 lines
arch/powerpc/kernel/vio.c +0 -1
 =A0static void vio_cmo_set_dma_ops(struct vio_dev *viodev)
 =A0{
- =A0 =A0 =A0 vio_dma_mapping_ops.dma_supported =3D dma_iommu_ops.dma_sup=

ported;

arch/powerpc/kernel/vio.c struct vio_dev *vio_register_device_node(struct dev=

Quoted 6 lines
arch/powerpc/kernel/vio.c +1 -1
- =A0 =A0 =A0 viodev->dev.archdata.dma_ops =3D &vio_dma_mapping_ops;
+ =A0 =A0 =A0 set_dma_ops(&viodev->dev, &vio_dma_mapping_ops);
 =A0}
 

=A0/**

arch/powerpc/kernel/vio.c +0 -0
@@ -1246,7 +1251,7 @@ struct vio_dev *vio_register_device_node(struct dev=

ice_node *of_node)

Quoted 4 lines
arch/powerpc/kernel/vio.c +0 -0
 =A0 =A0 =A0 =A0if (firmware_has_feature(FW_FEATURE_CMO))
 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0vio_cmo_set_dma_ops(viodev);

=A0 =A0 =A0 =A0else
- =A0 =A0 =A0 =A0 =A0 =A0 =A0 viodev->dev.archdata.dma_ops =3D &dma_iommu=

_ops;

Quoted 2 lines

+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 set_dma_ops(&viodev->dev, &dma_iommu_ops); =A0 =A0 =A0 =A0set_iommu_table_base(&viodev->dev, vio_build_iommu_table(v=

iodev));

Quoted 6 lines

=A0 =A0 =A0 =A0set_dev_node(&viodev->dev, of_node_to_nid(of_node));

--
1.7.0.4

--=20
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.

#14 lore raw

Re: [PATCH 09/15] ppc/vio: use dma ops helpers

Quoted 8 lines · Grant Likely
Quoted 5 lines · Nishanth Aravamudan

Use the set_dma_ops helper. Instead of modifying vio_dma_mapping_ops, just create a trivial wrapper for dma_supported.

Looks right to me.

I never saw 8/15 btw ... I'm applying up to 7, please resend the rest.

Cheers,
Ben.

Quoted 54 lines
Quoted 51 lines
---
 arch/powerpc/kernel/vio.c |   11 ++++++++---
 1 files changed, 8 insertions(+), 3 deletions(-)
arch/powerpc/kernel/vio.c +8 -4
index d692989..3c3083f 100644
@@ -602,6 +602,11 @@ static void vio_dma_iommu_unmap_sg(struct device *dev,
       vio_cmo_dealloc(viodev, alloc_size);
 }
 
+static int vio_dma_iommu_dma_supported(struct device *dev, u64 mask)
+{
+        return dma_iommu_ops.dma_supported(dev, mask);
+}
+
 struct dma_map_ops vio_dma_mapping_ops = {
       .alloc_coherent = vio_dma_iommu_alloc_coherent,
       .free_coherent  = vio_dma_iommu_free_coherent,
@@ -609,6 +614,7 @@ struct dma_map_ops vio_dma_mapping_ops = {
       .unmap_sg       = vio_dma_iommu_unmap_sg,
       .map_page       = vio_dma_iommu_map_page,
       .unmap_page     = vio_dma_iommu_unmap_page,
+       .dma_supported  = vio_dma_iommu_dma_supported,
 
 };
 
@@ -860,8 +866,7 @@ static void vio_cmo_bus_remove(struct vio_dev *viodev)
 
 static void vio_cmo_set_dma_ops(struct vio_dev *viodev)
 {
-       vio_dma_mapping_ops.dma_supported = dma_iommu_ops.dma_supported;
-       viodev->dev.archdata.dma_ops = &vio_dma_mapping_ops;
+       set_dma_ops(&viodev->dev, &vio_dma_mapping_ops);
 }
 
 /**
@@ -1246,7 +1251,7 @@ struct vio_dev *vio_register_device_node(struct device_node *of_node)
       if (firmware_has_feature(FW_FEATURE_CMO))
               vio_cmo_set_dma_ops(viodev);
       else
-               viodev->dev.archdata.dma_ops = &dma_iommu_ops;
+               set_dma_ops(&viodev->dev, &dma_iommu_ops);
       set_iommu_table_base(&viodev->dev, vio_build_iommu_table(viodev));
       set_dev_node(&viodev->dev, of_node_to_nid(of_node));
 
--

1.7.0.4

#15 lore raw

Re: [PATCH 09/15] ppc/vio: use dma ops helpers

Quoted 11 lines · Benjamin Herrenschmidt
Quoted 8 lines · Grant Likely
Quoted 5 lines · Nishanth Aravamudan

Use the set_dma_ops helper. Instead of modifying vio_dma_mapping_ops, just create a trivial wrapper for dma_supported.

Looks right to me.

I never saw 8/15 btw ... I'm applying up to 7, please resend the rest.

Ah apologies, 8/15 was for microblaze and get_maintainer.pl didn't cc many of the same folks. Will resend 8-15 and make sure ppc-dev gets cc'd.

Thanks,
Nish

Signature · 2 lines
Nishanth Aravamudan <[email protected]>
IBM Linux Technology Center
#16 lore raw

PATCH 10/15 ppc/pasemi: clean up pasemi iommu table initializations

In reply to a message not in the archive

No need for empty helpers with iommu off, the ppc_md hooks are optional.

The direct_dma_ops are the default pci_dma_ops, so no need to set in the them iommu off case.

No need to set the device tree device_node pci node iommu pointer, its only used for dlpar remove.

---
 arch/powerpc/platforms/pasemi/iommu.c |   19 +------------------
 1 files changed, 1 insertions(+), 18 deletions(-)
arch/powerpc/platforms/pasemi/iommu.c +1 -18
index 1f9fb2c..14943ef 100644
@@ -156,20 +156,12 @@ static void iommu_table_iobmap_setup(void)
 
 static void pci_dma_bus_setup_pasemi(struct pci_bus *bus)
 {
-	struct device_node *dn;
-
 	pr_debug("pci_dma_bus_setup, bus %p, bus->self %p\n", bus, bus->self);
 
 	if (!iommu_table_iobmap_inited) {
 		iommu_table_iobmap_inited = 1;
 		iommu_table_iobmap_setup();
 	}
-
-	dn = pci_bus_to_OF_node(bus);
-
-	if (dn)
-		PCI_DN(dn)->iommu_table = &iommu_table_iobmap;
-
 }
 
 
@@ -192,9 +184,6 @@ static void pci_dma_dev_setup_pasemi(struct pci_dev *dev)
 	set_iommu_table_base(&dev->dev, &iommu_table_iobmap);
 }
 
-static void pci_dma_bus_setup_null(struct pci_bus *b) { }
-static void pci_dma_dev_setup_null(struct pci_dev *d) { }
-
 int __init iob_init(struct device_node *dn)
 {
 	unsigned long tmp;
@@ -251,14 +240,8 @@ void __init iommu_init_early_pasemi(void)
 	iommu_off = of_chosen &&
 			of_get_property(of_chosen, "linux,iommu-off", NULL);
 #endif
-	if (iommu_off) {
-		/* Direct I/O, IOMMU off */
-		ppc_md.pci_dma_dev_setup = pci_dma_dev_setup_null;
-		ppc_md.pci_dma_bus_setup = pci_dma_bus_setup_null;
-		set_pci_dma_ops(&dma_direct_ops);
-
+	if (iommu_off)
 		return;
-	}
 
 	iob_init(NULL);
 
Signature · 1 line
1.7.0.4
Acked-by on v1 10/15
#17 lore raw

Re: [PATCH 10/15] ppc/pasemi: clean up pasemi iommu table initializations

Quoted 1 line · Nishanth Aravamudan

No need for empty helpers with iommu off, the ppc_md hooks are optional.

Not any more, they used to be needed. :-)

Quoted 8 lines

The direct_dma_ops are the default pci_dma_ops, so no need to set in the them iommu off case.

No need to set the device tree device_node pci node iommu pointer, its only used for dlpar remove.

-Olof

#18 lore raw

PATCH 11/15 ppc/cell: beat dma ops cleanup

In reply to a message not in the archive

direct_dma_ops is the default pci dma ops.

No need to call a function to get the pci dma ops, we know they are the dma_direct_ops.

---
 arch/powerpc/platforms/cell/beat_iommu.c |    3 +--
 1 files changed, 1 insertions(+), 2 deletions(-)
arch/powerpc/platforms/cell/beat_iommu.c +1 -2
index beec405..3ce6855 100644
@@ -76,7 +76,7 @@ static void __init celleb_init_direct_mapping(void)
 
 static void celleb_dma_dev_setup(struct device *dev)
 {
-	dev->archdata.dma_ops = get_pci_dma_ops();
+	set_dma_ops(dev, &dma_direct_ops);
 	set_dma_offset(dev, celleb_dma_direct_offset);
 }
 
@@ -106,7 +106,6 @@ static struct notifier_block celleb_of_bus_notifier = {
 static int __init celleb_init_iommu(void)
 {
 	celleb_init_direct_mapping();
-	set_pci_dma_ops(&dma_direct_ops);
 	ppc_md.pci_dma_dev_setup = celleb_pci_dma_dev_setup;
 	bus_register_notifier(&platform_bus_type, &celleb_of_bus_notifier);
 
Signature · 1 line
1.7.0.4
#20 lore raw

PATCH 12/15 ppc/dart: iommu table cleanup

In reply to a message not in the archive

No need to set the device tree device_node pci node iommu pointer, its only used for dlpar remove.

direct_dma_ops are the default, no need to restore them or dma bus ops if we haven't set them in the error paths.

---
 arch/powerpc/sysdev/dart_iommu.c |   18 +-----------------
 1 files changed, 1 insertions(+), 17 deletions(-)
arch/powerpc/sysdev/dart_iommu.c +1 -17
index 559db2b..64fe9c7 100644
@@ -302,17 +302,10 @@ static void pci_dma_dev_setup_dart(struct pci_dev *dev)
 
 static void pci_dma_bus_setup_dart(struct pci_bus *bus)
 {
-	struct device_node *dn;
-
 	if (!iommu_table_dart_inited) {
 		iommu_table_dart_inited = 1;
 		iommu_table_dart_setup();
 	}
-
-	dn = pci_bus_to_OF_node(bus);
-
-	if (dn)
-		PCI_DN(dn)->iommu_table = &iommu_table_dart;
 }
 
 void __init iommu_init_early_dart(void)
@@ -324,7 +317,7 @@ void __init iommu_init_early_dart(void)
 	if (dn == NULL) {
 		dn = of_find_compatible_node(NULL, "dart", "u4-dart");
 		if (dn == NULL)
-			goto bail;
+			return;	/* use default direct_dma_ops */
 		dart_is_u4 = 1;
 	}
 
@@ -340,16 +333,7 @@ void __init iommu_init_early_dart(void)
 
 		/* Setup pci_dma ops */
 		set_pci_dma_ops(&dma_iommu_ops);
-		return;
 	}
-
- bail:
-	/* If init failed, use direct iommu and null setup functions */
-	ppc_md.pci_dma_dev_setup = NULL;
-	ppc_md.pci_dma_bus_setup = NULL;
-
-	/* Setup pci_dma ops */
-	set_pci_dma_ops(&dma_direct_ops);
 }
 
 #ifdef CONFIG_PM
Signature · 1 line
1.7.0.4
#21 lore raw

PATCH 13/15 ppc/pseries: iommu cleanup

In reply to a message not in the archive

No need to initialize per-cpu pointer to NULL, it is the default.

Direct dma ops and no setup are the defaults, no need to set for iommu-off.

---
 arch/powerpc/platforms/pseries/iommu.c |    9 ++-------
 1 files changed, 2 insertions(+), 7 deletions(-)
arch/powerpc/platforms/pseries/iommu.c +2 -7
index a77bcae..9184db3 100644
@@ -140,7 +140,7 @@ static int tce_build_pSeriesLP(struct iommu_table *tbl, long tcenum,
 	return ret;
 }
 
-static DEFINE_PER_CPU(u64 *, tce_page) = NULL;
+static DEFINE_PER_CPU(u64 *, tce_page);
 
 static int tce_buildmulti_pSeriesLP(struct iommu_table *tbl, long tcenum,
 				     long npages, unsigned long uaddr,
@@ -589,13 +589,8 @@ static struct notifier_block iommu_reconfig_nb = {
 /* These are called very early. */
 void iommu_init_early_pSeries(void)
 {
-	if (of_chosen && of_get_property(of_chosen, "linux,iommu-off", NULL)) {
-		/* Direct I/O, IOMMU off */
-		ppc_md.pci_dma_dev_setup = NULL;
-		ppc_md.pci_dma_bus_setup = NULL;
-		set_pci_dma_ops(&dma_direct_ops);
+	if (of_chosen && of_get_property(of_chosen, "linux,iommu-off", NULL))
 		return;
-	}
 
 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
 		if (firmware_has_feature(FW_FEATURE_MULTITCE)) {
Signature · 1 line
1.7.0.4

A message that is not in the archive (sent off-list or not yet received): 1284573958-8397-14-git-send-email-nacc-r/[email protected]

Reviewed-by
#23 lore raw

Re: [PATCH 13/15] ppc/pseries: iommu cleanup

Quoted 7 lines · Nishanth Aravamudan

No need to initialize per-cpu pointer to NULL, it is the default.

Direct dma ops and no setup are the defaults, no need to set for iommu-off.

Also looks correct.

g.

Quoted 36 lines

---
 arch/powerpc/platforms/pseries/iommu.c |    9 ++-------

 1 files changed, 2 insertions(+), 7 deletions(-)
arch/powerpc/platforms/pseries/iommu.c +2 -8
index a77bcae..9184db3 100644
@@ -140,7 +140,7 @@ static int tce_build_pSeriesLP(struct iommu_table *tbl, long tcenum,
        return ret;
  }
 
-static DEFINE_PER_CPU(u64 *, tce_page) = NULL;
+static DEFINE_PER_CPU(u64 *, tce_page);
 
  static int tce_buildmulti_pSeriesLP(struct iommu_table *tbl, long tcenum,
                                     long npages, unsigned long uaddr,
@@ -589,13 +589,8 @@ static struct notifier_block iommu_reconfig_nb = {
  /* These are called very early. */
  void iommu_init_early_pSeries(void)
  {
-       if (of_chosen && of_get_property(of_chosen, "linux,iommu-off", NULL)) {
-               /* Direct I/O, IOMMU off */
-               ppc_md.pci_dma_dev_setup = NULL;
-               ppc_md.pci_dma_bus_setup = NULL;
-               set_pci_dma_ops(&dma_direct_ops);
+       if (of_chosen && of_get_property(of_chosen, "linux,iommu-off", NULL))
                return;
-       }
 
        if (firmware_has_feature(FW_FEATURE_LPAR)) {
                if (firmware_has_feature(FW_FEATURE_MULTITCE)) {
--

1.7.0.4

Signature · 2 lines
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.
#24 lore raw

PATCH 14/15 ppc64 iommu: use coherent_dma_mask for alloc_coherent

In reply to a message not in the archive

The IOMMU code has been passing the dma-mask instead of the coherent_dma_mask to the iommu allocator. Coherent allocations should be made using the coherent_dma_mask.

---
We currently don't check the mask other than to warn when its being set, so I don't think this is stable material.

---
 arch/powerpc/kernel/dma-iommu.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/powerpc/kernel/dma-iommu.c +1 -1
index 6e54a0f..e755415 100644
@@ -19,7 +19,7 @@ static void *dma_iommu_alloc_coherent(struct device *dev, size_t size,
 				      dma_addr_t *dma_handle, gfp_t flag)
 {
 	return iommu_alloc_coherent(dev, get_iommu_table_base(dev), size,
-				    dma_handle, device_to_mask(dev), flag,
+				    dma_handle, dev->coherent_dma_mask, flag,
 				    dev_to_node(dev));
 }
 
Signature · 1 line
1.7.0.4
#25 lore raw

Re: [PATCH 14/15] ppc64 iommu: use coherent_dma_mask for alloc_coherent

Quoted 3 lines · Nishanth Aravamudan

The IOMMU code has been passing the dma-mask instead of the coherent_dma_mask to the iommu allocator. Coherent allocations should be made using the coherent_dma_mask.

Won't that break macio devices too ? afaik, they don't set coherent_dma_mask. Have you tried booting on a G5 with iommu enabled ?

(It may not be broken, just asking...)

Cheers,
Ben.

Quoted 22 lines

---
We currently don't check the mask other than to warn when its being set, so I don't think this is stable material.

---
 arch/powerpc/kernel/dma-iommu.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/powerpc/kernel/dma-iommu.c +1 -1
index 6e54a0f..e755415 100644
@@ -19,7 +19,7 @@ static void *dma_iommu_alloc_coherent(struct device *dev, size_t size,
 				      dma_addr_t *dma_handle, gfp_t flag)
 {
 	return iommu_alloc_coherent(dev, get_iommu_table_base(dev), size,
-				    dma_handle, device_to_mask(dev), flag,
+				    dma_handle, dev->coherent_dma_mask, flag,
 				    dev_to_node(dev));
 }
 
#26 lore raw

Re: [PATCH 14/15] ppc64 iommu: use coherent_dma_mask for alloc_coherent

Quoted 9 lines · Benjamin Herrenschmidt
Quoted 3 lines · Nishanth Aravamudan

The IOMMU code has been passing the dma-mask instead of the coherent_dma_mask to the iommu allocator. Coherent allocations should be made using the coherent_dma_mask.

Won't that break macio devices too ? afaik, they don't set coherent_dma_mask. Have you tried booting on a G5 with iommu enabled ?

(It may not be broken, just asking...)

I have not tried this. I unfortunately do not have immediate access to a G5 to test on, but will ask around.

Thanks,
Nish

Signature · 2 lines
Nishanth Aravamudan <[email protected]>
IBM Linux Technology Center
#27 lore raw

PATCH 15/15 ppc/vio: ensure dma_coherent_mask is set

In reply to a message not in the archive

Without this change drivers, such as ibmvscsi, fail to load with the previous change.

---
 arch/powerpc/kernel/vio.c |    3 +++
 1 files changed, 3 insertions(+), 0 deletions(-)
arch/powerpc/kernel/vio.c +3 -0
index 3c3083f..e8d73de 100644
@@ -1259,6 +1259,9 @@ struct vio_dev *vio_register_device_node(struct device_node *of_node)
 	viodev->dev.parent = &vio_bus_device.dev;
 	viodev->dev.bus = &vio_bus_type;
 	viodev->dev.release = vio_dev_release;
+        /* needed to ensure proper operation of coherent allocations
+         * later, in case driver doesn't set it explicitly */
+        dma_set_coherent_mask(&viodev->dev, DMA_BIT_MASK(64));
 
 	/* register with generic device framework */
 	if (device_register(&viodev->dev)) {
Signature · 1 line
1.7.0.4
#28 lore raw

Re: [PATCH 15/15] ppc/vio: ensure dma_coherent_mask is set

On Wed, Sep 15, 2010 at 12:05 PM, Nishanth Aravamudan <[email protected]> wro= te:

Quoted 2 lines

Without this change drivers, such as ibmvscsi, fail to load with the previous change.

Shouldn't this patch be ordered before the previous change then to preserve bisectability?

Also, patch descriptions should be explicit about what the "previous change" refers to. Once this is committed, git log may very well insert other changes from other branches between this commit and whatever "previous change" refers to.

g.

arch/powerpc/kernel/vio.c struct vio_dev *vio_register_device_node(struct dev=

Quoted 9 lines

---
=A0arch/powerpc/kernel/vio.c | =A0 =A03 +++ =A01 files changed, 3 insertions(+), 0 deletions(-)

arch/powerpc/kernel/vio.c +0 -0
index 3c3083f..e8d73de 100644
@@ -1259,6 +1259,9 @@ struct vio_dev *vio_register_device_node(struct dev=

ice_node *of_node)

arch/powerpc/kernel/vio.c struct vio_dev *vio_register_device_node(struct dev= · line 1262

Quoted 4 lines
arch/powerpc/kernel/vio.c +1 -0
 =A0 =A0 =A0 =A0viodev->dev.parent =3D &vio_bus_device.dev;
 =A0 =A0 =A0 =A0viodev->dev.bus =3D &vio_bus_type;
 =A0 =A0 =A0 =A0viodev->dev.release =3D vio_dev_release;
+ =A0 =A0 =A0 =A0/* needed to ensure proper operation of coherent allocat=

ions

Quoted 9 lines
arch/powerpc/kernel/vio.c +2 -0
+ =A0 =A0 =A0 =A0 * later, in case driver doesn't set it explicitly */
+ =A0 =A0 =A0 =A0dma_set_coherent_mask(&viodev->dev, DMA_BIT_MASK(64));
 

=A0 =A0 =A0 =A0/* register with generic device framework */ =A0 =A0 =A0 =A0if (device_register(&viodev->dev)) { --
1.7.0.4

--=20
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.

#29 lore raw

Re: [PATCH 15/15] ppc/vio: ensure dma_coherent_mask is set

Quoted 6 lines · Grant Likely
Quoted 2 lines · Nishanth Aravamudan

Without this change drivers, such as ibmvscsi, fail to load with the previous change.

Shouldn't this patch be ordered before the previous change then to preserve bisectability?

You are probably right. I wasn't sure if I should fold it in or keep it separate. I should have changed the order, though. Sorry about that!

Quoted 4 lines

Also, patch descriptions should be explicit about what the "previous change" refers to. Once this is committed, git log may very well insert other changes from other branches between this commit and whatever "previous change" refers to.

Yep -- what's the best way to make the reference? By subject from the patch? Obviously I don't have the SHA1 with which the commit will go upstream.

Thanks,
Nish

Signature · 2 lines
Nishanth Aravamudan <[email protected]>
IBM Linux Technology Center
#30 lore raw

Re: [PATCH 15/15] ppc/vio: ensure dma_coherent_mask is set

Quoted 10 lines · Nishanth Aravamudan
Quoted 6 lines · Grant Likely
Quoted 2 lines · Nishanth Aravamudan

Without this change drivers, such as ibmvscsi, fail to load with the previous change.

Shouldn't this patch be ordered before the previous change then to preserve bisectability?

You are probably right. I wasn't sure if I should fold it in or keep it separate. I should have changed the order, though. Sorry about that!

It's such a small patch I would just fold it in.

Quoted 9 lines
Quoted 5 lines

Also, patch descriptions should be explicit about what the "previous change" refers to. Once this is committed, git log may very well insert other changes from other branches between this commit and whatever "previous change" refers to.

Yep -- what's the best way to make the reference? By subject from the patch? Obviously I don't have the SHA1 with which the commit will go upstream.

By name should be good. As long as a reader doesn't need background information from your head to figure out why the change was made then it should be okay.

g.

#31 lore raw

Re: [PATCH 15/15] ppc/vio: ensure dma_coherent_mask is set

In reply to Nishanth Aravamudan
Quoted 3 lines · Nishanth Aravamudan

Without this change drivers, such as ibmvscsi, fail to load with the previous change.
---

So you broke bisection... fold the patch instead or invert them

Cheers,
Ben.

Quoted 17 lines
 arch/powerpc/kernel/vio.c |    3 +++
 1 files changed, 3 insertions(+), 0 deletions(-)
arch/powerpc/kernel/vio.c +3 -0
index 3c3083f..e8d73de 100644
@@ -1259,6 +1259,9 @@ struct vio_dev *vio_register_device_node(struct device_node *of_node)
 	viodev->dev.parent = &vio_bus_device.dev;
 	viodev->dev.bus = &vio_bus_type;
 	viodev->dev.release = vio_dev_release;
+        /* needed to ensure proper operation of coherent allocations
+         * later, in case driver doesn't set it explicitly */
+        dma_set_coherent_mask(&viodev->dev, DMA_BIT_MASK(64));
 
 	/* register with generic device framework */
 	if (device_register(&viodev->dev)) {
#32 lore raw

Re: [PATCH 15/15] ppc/vio: ensure dma_coherent_mask is set

Hi Ben,

Quoted 6 lines · Benjamin Herrenschmidt
Quoted 3 lines · Nishanth Aravamudan

Without this change drivers, such as ibmvscsi, fail to load with the previous change.
---

So you broke bisection... fold the patch instead or invert them

Thanks for the review. I resent this series on 10/18 and it had the patches folded together.

Thanks,
Nish

Signature · 2 lines
Nishanth Aravamudan <[email protected]>
IBM Linux Technology Center
#33 lore raw

Re: [PATCH 15/15] ppc/vio: ensure dma_coherent_mask is set

Quoted 4 lines · Nishanth Aravamudan
Quoted 1 line

So you broke bisection... fold the patch instead or invert them

Thanks for the review. I resent this series on 10/18 and it had the patches folded together.

Right, I know :-) I just mistakenly had that old one still tagged in patchwork, my bad.

I've merged some of the new ones already.

Cheers,
Ben.