Skip to contentSkip to conversation

Adjust the dma-ranges for MTK IOMMU

PATCH Superseded Yong Wu 2 participants 19 messages Feb 14, 2023

Subsystems (7)

v3 of 7 Superseded Latest is v7 →

lore
Share & tools
lore

Yong Wu

author maintainer
#0 lore raw

v3 cover Adjust the dma-ranges for MTK IOMMU

After commit f1ad5338a4d5 ("of: Fix "dma-ranges" handling for bus controllers"), the dma-ranges is not allowed for dts leaf node. but we still would like to separate the different masters into different iova regions. Thus we adjust the internal flow, separate the 16GB iova range by the master HW larbid/portid and add the dma-ranges property in the parent "soc" node. This also could avoid the users forget/abuse the iova regions.

The commit f1ad5338a4d5 did affect the mt8195 venc, But it is not a fatal issue, it could also work well at 0-4GB iova. thus I don't add "Fixes:" tag.

In this series, I add functions for mt8192/mt8195/mt8186, mt8188 will be in its special patchset. and the previous mt8173/mt8183...support 0-4GB only, no need this function.

Change note:
v3: Add a new patch only for comment more in the code.

v2: https://lore.kernel.org/linux-mediatek/[email protected]/
a) Base on next-20230206 since mt8195 jpeg node is applied which affect this patch.
b) Reword the commit message [1/10][2/10] to explain effect.

v1: https://lore.kernel.org/linux-mediatek/[email protected]/
Base on v6.2-rc3.

Yong Wu (11):
dt-bindings: media: mediatek,vcodec: Remove dma-ranges property
dt-bindings: media: mediatek,jpeg: Remove dma-ranges property iommu/mediatek: Improve comment for the current region/bank iommu/mediatek: Get regionid from larb/port id iommu/mediatek: mt8195: Add iova_region_larb_msk iommu/mediatek: mt8186: Add iova_region_larb_msk iommu/mediatek: mt8192: Add iova_region_larb_msk iommu/mediatek: Add a gap for the iova regions
arm64: dts: mt8195: Add dma-ranges for the parent "soc" node
arm64: dts: mt8195: Remove the unnecessary dma-ranges
arm64: dts: mt8186: Add dma-ranges for the parent "soc" node

 .../media/mediatek,mt8195-jpegdec.yaml        |   7 -
 .../media/mediatek,mt8195-jpegenc.yaml        |   7 -
 .../media/mediatek,vcodec-decoder.yaml        |   5 -
 .../media/mediatek,vcodec-encoder.yaml        |   5 -
 .../media/mediatek,vcodec-subdev-decoder.yaml |   7 -
 .../bindings/media/mediatek-jpeg-encoder.yaml |   5 -
 arch/arm64/boot/dts/mediatek/mt8186.dtsi      |   1 +
 arch/arm64/boot/dts/mediatek/mt8195.dtsi      |   4 +-
 drivers/iommu/mtk_iommu.c                     | 137 ++++++++++++++----
 9 files changed, 108 insertions(+), 70 deletions(-)
Signature · 1 line
2.18.0
author maintainer
#1 lore raw

PATCH v3 1/11 dt-bindings: media: mediatek,vcodec: Remove dma-ranges property

After commit f1ad5338a4d5 ("of: Fix "dma-ranges" handling for bus controllers"), the dma-ranges of the leaf node doesn't work. Remove it for vcodec here.

1) For mediatek,vcodec-decoder.yaml and mediatek,vcodec-encoder.yaml, this property is in the leaf node, it is invalid as the above comment.

Currently there is only mt8195 VENC node has this property in upstream. Indeed, VENC is affected, but it is not a fatal issue. Originally it expects its iova range locate at 4GB-8GB. However after that commit, its expectation doesn't come true, it will fall back to 0-4GB iova and also could work well.

2) For mediatek,vcodec-subdev-decoder.yaml, It already uses parent-child node, this property works. Instead, MediaTek iommu will control the masters' iova ranges by the master's larb/port id internally, then the dma-ranges property is unnecessary for the master's node.

Cc Tiffany Lin <[email protected]>
Cc Andrew-CT Chen <[email protected]>
Cc Yunfei Dong <[email protected]>
Cc Mauro Carvalho Chehab <[email protected]>
Cc Matthias Brugger <[email protected]>
---
 .../devicetree/bindings/media/mediatek,vcodec-decoder.yaml | 5 -----
 .../devicetree/bindings/media/mediatek,vcodec-encoder.yaml | 5 -----
 .../bindings/media/mediatek,vcodec-subdev-decoder.yaml     | 7 -------
 3 files changed, 17 deletions(-)
Documentation/devicetree/bindings/media/mediatek,vcodec-decoder.yaml +0 -5
index aa55ca65d6ed..fad59b486d5d 100644
@@ -56,11 +56,6 @@ properties:
       List of the hardware port in respective IOMMU block for current Socs.
       Refer to bindings/iommu/mediatek,iommu.yaml.
 
-  dma-ranges:
-    maxItems: 1
-    description: |
-      Describes the physical address space of IOMMU maps to memory.
-
   mediatek,vpu:
     $ref: /schemas/types.yaml#/definitions/phandle
     description:
Documentation/devicetree/bindings/media/mediatek,vcodec-encoder.yaml +0 -5
index 0f2ea8d9a10c..a2051b31fa29 100644
@@ -49,11 +49,6 @@ properties:
       List of the hardware port in respective IOMMU block for current Socs.
       Refer to bindings/iommu/mediatek,iommu.yaml.
 
-  dma-ranges:
-    maxItems: 1
-    description: |
-      Describes the physical address space of IOMMU maps to memory.
-
   mediatek,vpu:
     $ref: /schemas/types.yaml#/definitions/phandle
     description:
Documentation/devicetree/bindings/media/mediatek,vcodec-subdev-decoder.yaml +0 -7
index c4f20acdc1f8..290594bc91cc 100644
@@ -76,11 +76,6 @@ properties:
       The node of system control processor (SCP), using
       the remoteproc & rpmsg framework.
 
-  dma-ranges:
-    maxItems: 1
-    description: |
-      Describes the physical address space of IOMMU maps to memory.
-
   "#address-cells":
     const: 2
 
@@ -203,7 +198,6 @@ required:
   - reg
   - iommus
   - mediatek,scp
-  - dma-ranges
   - ranges
 
 if:
@@ -236,7 +230,6 @@ examples:
             compatible = "mediatek,mt8192-vcodec-dec";
             mediatek,scp = <&scp>;
             iommus = <&iommu0 M4U_PORT_L4_VDEC_MC_EXT>;
-            dma-ranges = <0x1 0x0 0x0 0x40000000 0x0 0xfff00000>;
             #address-cells = <2>;
             #size-cells = <2>;
             ranges = <0 0 0 0x16000000 0 0x40000>;
Signature · 1 line
2.18.0
author maintainer
#2 lore raw

PATCH v3 2/11 dt-bindings: media: mediatek,jpeg: Remove dma-ranges property

In reply to Yong Wu

After commit f1ad5338a4d5 ("of: Fix "dma-ranges" handling for bus controllers"), the dma-ranges of the leaf node doesn't work. Remove it for jpeg here.

Currently there is only mt8195 jpeg node has this property in upstream, and it already uses parent-child node, this property did work. But instead, MediaTek iommu will control the masters' iova ranges by the master's larb/port id internally, then this property is unnecessary.

---
 .../devicetree/bindings/media/mediatek,mt8195-jpegdec.yaml | 7 -------
 .../devicetree/bindings/media/mediatek,mt8195-jpegenc.yaml | 7 -------
 .../devicetree/bindings/media/mediatek-jpeg-encoder.yaml   | 5 -----
 3 files changed, 19 deletions(-)
Documentation/devicetree/bindings/media/mediatek,mt8195-jpegdec.yaml +0 -7
index 71595c013dbb..e5448c60e3eb 100644
@@ -26,11 +26,6 @@ properties:
       Documentation/devicetree/bindings/iommu/mediatek,iommu.yaml for details.
       Ports are according to the HW.
 
-  dma-ranges:
-    maxItems: 1
-    description: |
-      Describes the physical address space of IOMMU maps to memory.
-
   "#address-cells":
     const: 2
 
@@ -89,7 +84,6 @@ required:
   - compatible
   - power-domains
   - iommus
-  - dma-ranges
   - ranges
 
 additionalProperties: false
@@ -115,7 +109,6 @@ examples:
                      <&iommu_vpp M4U_PORT_L19_JPGDEC_BSDMA1>,
                      <&iommu_vpp M4U_PORT_L19_JPGDEC_BUFF_OFFSET1>,
                      <&iommu_vpp M4U_PORT_L19_JPGDEC_BUFF_OFFSET0>;
-            dma-ranges = <0x1 0x0 0x0 0x40000000 0x0 0xfff00000>;
             #address-cells = <2>;
             #size-cells = <2>;
             ranges;
Documentation/devicetree/bindings/media/mediatek,mt8195-jpegenc.yaml +0 -7
index 95990539f7c0..596186497b68 100644
@@ -26,11 +26,6 @@ properties:
       Documentation/devicetree/bindings/iommu/mediatek,iommu.yaml for details.
       Ports are according to the HW.
 
-  dma-ranges:
-    maxItems: 1
-    description: |
-      Describes the physical address space of IOMMU maps to memory.
-
   "#address-cells":
     const: 2
 
@@ -89,7 +84,6 @@ required:
   - compatible
   - power-domains
   - iommus
-  - dma-ranges
   - ranges
 
 additionalProperties: false
@@ -113,7 +107,6 @@ examples:
                      <&iommu_vpp M4U_PORT_L20_JPGENC_C_RDMA>,
                      <&iommu_vpp M4U_PORT_L20_JPGENC_Q_TABLE>,
                      <&iommu_vpp M4U_PORT_L20_JPGENC_BSDMA>;
-            dma-ranges = <0x1 0x0 0x0 0x40000000 0x0 0xfff00000>;
             #address-cells = <2>;
             #size-cells = <2>;
             ranges;
Documentation/devicetree/bindings/media/mediatek-jpeg-encoder.yaml +0 -5
index c8412e8ab353..37800e1908cc 100644
@@ -44,11 +44,6 @@ properties:
       Documentation/devicetree/bindings/iommu/mediatek,iommu.yaml for details.
       Ports are according to the HW.
 
-  dma-ranges:
-    maxItems: 1
-    description: |
-      Describes the physical address space of IOMMU maps to memory.
-
 required:
   - compatible
   - reg
Signature · 1 line
2.18.0
author maintainer
#3 lore raw

PATCH v3 3/11 iommu/mediatek: Improve comment for the current region/bank

In reply to Yong Wu

No functional change. Just add more comment about the current region/bank in the code.

---
 drivers/iommu/mtk_iommu.c | 31 ++++++++++++++++++++++++++-----
 1 file changed, 26 insertions(+), 5 deletions(-)
drivers/iommu/mtk_iommu.c +26 -5
index d5a4955910ff..ab53edcb221f 100644
@@ -197,12 +197,33 @@ struct mtk_iommu_plat_data {
 
 	char			*pericfg_comp_str;
 	struct list_head	*hw_list;
-	unsigned int		iova_region_nr;
-	const struct mtk_iommu_iova_region	*iova_region;
 
-	u8                  banks_num;
-	bool                banks_enable[MTK_IOMMU_BANK_MAX];
-	unsigned int        banks_portmsk[MTK_IOMMU_BANK_MAX];
+	/*
+	 * The IOMMU HW may support 16GB iova. In order to balance the IOVA ranges,
+	 * different masters will be put in different iova ranges, for example vcodec
+	 * is in 4G-8G and cam is in 8G-12G. Meanwhile, some masters may have the
+	 * special IOVA range requirement, like CCU can only support the address
+	 * 0x40000000-0x44000000.
+	 * Here list the iova ranges this SoC supports and which larbs/ports are in
+	 * which region.
+	 *
+	 * 16GB iova all use one pgtable, but each a region is a iommu group.
+	 */
+	struct {
+		unsigned int	iova_region_nr;
+		const struct mtk_iommu_iova_region	*iova_region;
+	};
+
+	/*
+	 * The IOMMU HW may have 5 banks. Each bank has a indenpendent pgtable.
+	 * Here list how many banks this SoC supports/enables and which ports are in which bank.
+	 */
+	struct {
+		u8		banks_num;
+		bool		banks_enable[MTK_IOMMU_BANK_MAX];
+		unsigned int	banks_portmsk[MTK_IOMMU_BANK_MAX];
+	};
+
 	unsigned char       larbid_remap[MTK_LARB_COM_MAX][MTK_LARB_SUBCOM_MAX];
 };
 
Signature · 1 line
2.18.0
maintainer Reviewed-by on v3 3/11
#4 lore raw

Re: [PATCH v3 03/11] iommu/mediatek: Improve comment for the current region/bank

Il 14/02/23 04:11, Yong Wu ha scritto:

Quoted 30 lines

No functional change. Just add more comment about the current region/bank in the code.

---
  drivers/iommu/mtk_iommu.c | 31 ++++++++++++++++++++++++++-----
  1 file changed, 26 insertions(+), 5 deletions(-)
drivers/iommu/mtk_iommu.c +8 -5
index d5a4955910ff..ab53edcb221f 100644
@@ -197,12 +197,33 @@ struct mtk_iommu_plat_data {
 
  	char			*pericfg_comp_str;
  	struct list_head	*hw_list;
-	unsigned int		iova_region_nr;
-	const struct mtk_iommu_iova_region	*iova_region;
 
-	u8                  banks_num;
-	bool                banks_enable[MTK_IOMMU_BANK_MAX];
-	unsigned int        banks_portmsk[MTK_IOMMU_BANK_MAX];
+	/*
+	 * The IOMMU HW may support 16GB iova. In order to balance the IOVA ranges,
+	 * different masters will be put in different iova ranges, for example vcodec
+	 * is in 4G-8G and cam is in 8G-12G. Meanwhile, some masters may have the
+	 * special IOVA range requirement, like CCU can only support the address
+	 * 0x40000000-0x44000000.
+	 * Here list the iova ranges this SoC supports and which larbs/ports are in
+	 * which region.

drivers/iommu/mtk_iommu.c struct mtk_iommu_plat_data { · line 218

Quoted 10 lines
drivers/iommu/mtk_iommu.c +10 -0
+	 *
+	 * 16GB iova all use one pgtable, but each a region is a iommu group.
+	 */
+	struct {
+		unsigned int	iova_region_nr;
+		const struct mtk_iommu_iova_region	*iova_region;
+	};
+
+	/*
+	 * The IOMMU HW may have 5 banks. Each bank has a indenpendent pgtable.

You got a typo here: inde*n*pendent - after fixing that:

author maintainer
#5 lore raw

PATCH v3 4/11 iommu/mediatek: Get regionid from larb/port id

In reply to Yong Wu

After commit f1ad5338a4d5 ("of: Fix "dma-ranges" handling for bus controllers"), the dma-ranges is not allowed for dts leaf node. but we still would like to separate to different masters into different iova regions.

Thus we have to separate it by the HW larbid and portid. For example, larb1/2 are in region2 and larb3 is in region3. The problem is that some ports inside a larb are in region4 while some ports inside this larb are in region5. Therefore I define a "iova_region_larb_msk" to help record the information for each a port. Take a example for a larb:
[1] = ~0: means all ports in this larb are in region1;
[2] = BIT(3) | BIT(4): means port3/4 in this larb are region2;
[3] = ~(BIT(3) | BIT(4)): means all the other ports except port3/4 in this larb are region3.

This method also avoids the users forget/abuse the iova regions.

---
 drivers/iommu/mtk_iommu.c | 47 ++++++++++++++++++++++-----------------
 1 file changed, 27 insertions(+), 20 deletions(-)
drivers/iommu/mtk_iommu.c +27 -20
index ab53edcb221f..7e2cb3b8cac8 100644
@@ -8,7 +8,6 @@
 #include <linux/clk.h>
 #include <linux/component.h>
 #include <linux/device.h>
-#include <linux/dma-direct.h>
 #include <linux/err.h>
 #include <linux/interrupt.h>
 #include <linux/io.h>
@@ -212,6 +211,11 @@ struct mtk_iommu_plat_data {
 	struct {
 		unsigned int	iova_region_nr;
 		const struct mtk_iommu_iova_region	*iova_region;
+		/*
+		 * Indicate the correspondence between larbs/ports and regions.
+		 * The index is same with iova_region.
+		 */
+		const u32	(*iova_region_larb_msk)[MTK_LARB_NR_MAX];
 	};
 
 	/*
@@ -529,30 +533,33 @@ static unsigned int mtk_iommu_get_bank_id(struct device *dev,
 static int mtk_iommu_get_iova_region_id(struct device *dev,
 					const struct mtk_iommu_plat_data *plat_data)
 {
-	const struct mtk_iommu_iova_region *rgn = plat_data->iova_region;
-	const struct bus_dma_region *dma_rgn = dev->dma_range_map;
-	int i, candidate = -1;
-	dma_addr_t dma_end;
+	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
+	unsigned int portidmsk = 0, larbid;
+	const u32 *rgn_larb_msk;
+	int i, region_id = -1;
 
-	if (!dma_rgn || plat_data->iova_region_nr == 1)
+	if (plat_data->iova_region_nr == 1)
 		return 0;
 
-	dma_end = dma_rgn->dma_start + dma_rgn->size - 1;
-	for (i = 0; i < plat_data->iova_region_nr; i++, rgn++) {
-		/* Best fit. */
-		if (dma_rgn->dma_start == rgn->iova_base &&
-		    dma_end == rgn->iova_base + rgn->size - 1)
-			return i;
-		/* ok if it is inside this region. */
-		if (dma_rgn->dma_start >= rgn->iova_base &&
-		    dma_end < rgn->iova_base + rgn->size)
-			candidate = i;
+	larbid = MTK_M4U_TO_LARB(fwspec->ids[0]);
+	for (i = 0; i < fwspec->num_ids; i++)
+		portidmsk |= BIT(MTK_M4U_TO_PORT(fwspec->ids[i]));
+
+	for (i = 0; i < plat_data->iova_region_nr; i++) {
+		rgn_larb_msk = plat_data->iova_region_larb_msk[i];
+		if (!rgn_larb_msk)
+			continue;
+
+		if  ((rgn_larb_msk[larbid] & portidmsk) == portidmsk) {
+			region_id = i;
+			break;
+		}
 	}
 
-	if (candidate >= 0)
-		return candidate;
-	dev_err(dev, "Can NOT find the iommu domain id(%pad 0x%llx).\n",
-		&dma_rgn->dma_start, dma_rgn->size);
+	if (region_id >= 0)
+		return region_id;
+	dev_err(dev, "Can NOT find the region for larb(%d-%x).\n",
+		larbid, portidmsk);
 	return -EINVAL;
 }
 
Signature · 1 line
2.18.0
maintainer Reviewed-by on v3 4/11
#6 lore raw

Re: [PATCH v3 04/11] iommu/mediatek: Get regionid from larb/port id

Il 14/02/23 04:11, Yong Wu ha scritto:

Quoted 35 lines

After commit f1ad5338a4d5 ("of: Fix "dma-ranges" handling for bus controllers"), the dma-ranges is not allowed for dts leaf node. but we still would like to separate to different masters into different iova regions.

Thus we have to separate it by the HW larbid and portid. For example, larb1/2 are in region2 and larb3 is in region3. The problem is that some ports inside a larb are in region4 while some ports inside this larb are in region5. Therefore I define a "iova_region_larb_msk" to help record the information for each a port. Take a example for a larb:
[1] = ~0: means all ports in this larb are in region1;
[2] = BIT(3) | BIT(4): means port3/4 in this larb are region2;
[3] = ~(BIT(3) | BIT(4)): means all the other ports except port3/4 in this larb are region3.

This method also avoids the users forget/abuse the iova regions.

---
  drivers/iommu/mtk_iommu.c | 47 ++++++++++++++++++++++-----------------
  1 file changed, 27 insertions(+), 20 deletions(-)
drivers/iommu/mtk_iommu.c +0 -1
index ab53edcb221f..7e2cb3b8cac8 100644
@@ -8,7 +8,6 @@
  #include <linux/clk.h>
  #include <linux/component.h>
  #include <linux/device.h>
-#include <linux/dma-direct.h>
  #include <linux/err.h>
  #include <linux/interrupt.h>
  #include <linux/io.h>

drivers/iommu/mtk_iommu.c struct mtk_iommu_plat_data { · line 216

Quoted 6 lines
drivers/iommu/mtk_iommu.c +3 -0
@@ -212,6 +211,11 @@ struct mtk_iommu_plat_data {
  	struct {
  		unsigned int	iova_region_nr;
  		const struct mtk_iommu_iova_region	*iova_region;
+		/*
+		 * Indicate the correspondence between larbs/ports and regions.
+		 * The index is same with iova_region.

* The index is the same as iova_region and larb port numbers are
* described as bit positions.
* For example, storing BIT(0) at index 2 means "port 0 is in region 2".

That's most probably the best short explanation that we can give, so that nobody goes crazy with understanding this one.

After fixing that comment, you totally deserve my

Cheers!

#7 lore raw

Re: [PATCH v3 04/11] iommu/mediatek: Get regionid from larb/port id

Il 14/02/23 10:07, AngeloGioacchino Del Regno ha scritto:

Quoted 38 lines

Il 14/02/23 04:11, Yong Wu ha scritto:

Quoted 37 lines

After commit f1ad5338a4d5 ("of: Fix "dma-ranges" handling for bus controllers"), the dma-ranges is not allowed for dts leaf node. but we still would like to separate to different masters into different iova regions.

Thus we have to separate it by the HW larbid and portid. For example, larb1/2 are in region2 and larb3 is in region3. The problem is that some ports inside a larb are in region4 while some ports inside this larb are in region5. Therefore I define a "iova_region_larb_msk" to help record the information for each a port. Take a example for a larb:
  [1] = ~0: means all ports in this larb are in region1;
  [2] = BIT(3) | BIT(4): means port3/4 in this larb are region2;
  [3] = ~(BIT(3) | BIT(4)): means all the other ports except port3/4 in this larb are region3.

This method also avoids the users forget/abuse the iova regions.

---
  drivers/iommu/mtk_iommu.c | 47 ++++++++++++++++++++++-----------------

  1 file changed, 27 insertions(+), 20 deletions(-)
drivers/iommu/mtk_iommu.c +0 -1
index ab53edcb221f..7e2cb3b8cac8 100644
@@ -8,7 +8,6 @@
   #include <linux/clk.h>
   #include <linux/component.h>
   #include <linux/device.h>
-#include <linux/dma-direct.h>
   #include <linux/err.h>
   #include <linux/interrupt.h>
   #include <linux/io.h>
@@ -212,6 +211,11 @@ struct mtk_iommu_plat_data {
       struct {
           unsigned int    iova_region_nr;
Quoted 9 lines

drivers/iommu/mtk_iommu.c struct mtk_iommu_plat_data { · line 216

Quoted 4 lines
drivers/iommu/mtk_iommu.c +3 -0
           const struct mtk_iommu_iova_region    *iova_region;
+        /*
+         * Indicate the correspondence between larbs/ports and regions.
+         * The index is same with iova_region.

* The index is the same as iova_region and larb port numbers are
* described as bit positions.
* For example, storing BIT(0) at index 2 means "port 0 is in region 2".

Sorry, I'm reiterating because the proposed comment is wrong.

/*
* Indicates the correspondance between larbs, ports and regions.
*
* The index is the same as iova_region and larb port numbers are
* described as bit positions.
* For example, storing BIT(0) at index 2,1 means "larb 1, port0 is in region 2".
* [2] = { [1] = BIT(0) }
*/

Quoted 10 lines

That's most probably the best short explanation that we can give, so that nobody goes crazy with understanding this one.

After fixing that comment, you totally deserve my

Cheers!

author maintainer
#8 lore raw

PATCH v3 5/11 iommu/mediatek: mt8195: Add iova_region_larb_msk

In reply to Yong Wu

Add iova_region_larb_msk for mt8195. We separate the 16GB iova regions by each device's larbid/portid.
Refer to include/dt-bindings/memory/mt8195-memory-port.h

Define a new macro MT8192_MULTI_REGION_MAX_NR to indicate the index of mt8xxx_larb_region_msk and
"struct mtk_iommu_iova_region mt8192_multi_dom" are the same.

---
 drivers/iommu/mtk_iommu.c | 24 +++++++++++++++++++++++-
 1 file changed, 23 insertions(+), 1 deletion(-)
drivers/iommu/mtk_iommu.c +23 -1
index 7e2cb3b8cac8..3d9986fb49a5 100644
@@ -332,7 +332,12 @@ static const struct mtk_iommu_iova_region single_domain[] = {
 	{.iova_base = 0,		.size = SZ_4G},
 };
 
-static const struct mtk_iommu_iova_region mt8192_multi_dom[] = {
+#define MT8192_MULTI_REGION_MAX_NR	6
+
+#define MT8192_MULTI_REGION_NR	(IS_ENABLED(CONFIG_ARCH_DMA_ADDR_T_64BIT) ? \
+				 MT8192_MULTI_REGION_MAX_NR : 1)
+
+static const struct mtk_iommu_iova_region mt8192_multi_dom[MT8192_MULTI_REGION_NR] = {
 	{ .iova_base = 0x0,		.size = SZ_4G},		/* 0 ~ 4G */
 	#if IS_ENABLED(CONFIG_ARCH_DMA_ADDR_T_64BIT)
 	{ .iova_base = SZ_4G,		.size = SZ_4G},		/* 4G ~ 8G */
@@ -1562,6 +1567,21 @@ static const struct mtk_iommu_plat_data mt8195_data_infra = {
 	.iova_region_nr   = ARRAY_SIZE(single_domain),
 };
 
+static const unsigned int mt8195_larb_region_msk[MT8192_MULTI_REGION_MAX_NR][MTK_LARB_NR_MAX] = {
+	[0] = {~0, ~0, ~0, ~0},               /* Region0: all ports for larb0/1/2/3 */
+	[1] = {0, 0, 0, 0, 0, 0, 0, 0,
+	       0, 0, 0, 0, 0, 0, 0, 0,
+	       0, 0, 0, ~0, ~0, ~0, ~0, ~0,   /* Region1: larb19/20/21/22/23/24 */
+	       ~0},
+	[2] = {0, 0, 0, 0, ~0, ~0, ~0, ~0,    /* Region2: the other larbs. */
+	       ~0, ~0, ~0, ~0, ~0, ~0, ~0, ~0,
+	       ~0, ~0, 0, 0, 0, 0, 0, 0,
+	       0, ~0, ~0, ~0, ~0},
+	[3] = {0},
+	[4] = {[18] = BIT(0) | BIT(1)},       /* Only larb18 port0/1 */
+	[5] = {[18] = BIT(2) | BIT(3)},       /* Only larb18 port2/3 */
+};
+
 static const struct mtk_iommu_plat_data mt8195_data_vdo = {
 	.m4u_plat	= M4U_MT8195,
 	.flags          = HAS_BCLK | HAS_SUB_COMM_2BITS | OUT_ORDER_WR_EN |
@@ -1572,6 +1592,7 @@ static const struct mtk_iommu_plat_data mt8195_data_vdo = {
 	.banks_enable   = {true},
 	.iova_region	= mt8192_multi_dom,
 	.iova_region_nr	= ARRAY_SIZE(mt8192_multi_dom),
+	.iova_region_larb_msk = mt8195_larb_region_msk,
 	.larbid_remap   = {{2, 0}, {21}, {24}, {7}, {19}, {9, 10, 11},
 			   {13, 17, 15/* 17b */, 25}, {5}},
 };
@@ -1586,6 +1607,7 @@ static const struct mtk_iommu_plat_data mt8195_data_vpp = {
 	.banks_enable   = {true},
 	.iova_region	= mt8192_multi_dom,
 	.iova_region_nr	= ARRAY_SIZE(mt8192_multi_dom),
+	.iova_region_larb_msk = mt8195_larb_region_msk,
 	.larbid_remap   = {{1}, {3},
 			   {22, MTK_INVALID_LARBID, MTK_INVALID_LARBID, MTK_INVALID_LARBID, 23},
 			   {8}, {20}, {12},
Signature · 1 line
2.18.0
#9 lore raw

Re: [PATCH v3 05/11] iommu/mediatek: mt8195: Add iova_region_larb_msk

Il 14/02/23 04:11, Yong Wu ha scritto:

Quoted 10 lines

Add iova_region_larb_msk for mt8195. We separate the 16GB iova regions by each device's larbid/portid.
Refer to include/dt-bindings/memory/mt8195-memory-port.h

Define a new macro MT8192_MULTI_REGION_MAX_NR to indicate the index of mt8xxx_larb_region_msk and
"struct mtk_iommu_iova_region mt8192_multi_dom" are the same.

Sorry but... you *are* introducing a iova_region_larb_msk for MT8192 later, while you're introducing definitions containing the MT8192 name in a commit that introduces the same but for MT8195.

I strongly suggest to, at this point, introduce MT8192 region mask here, along with the MT8192 definitions, and MT8195 after that :-)

Regards,
Angelo

author maintainer
#10 lore raw

Re: [PATCH v3 05/11] iommu/mediatek: mt8195: Add iova_region_larb_msk

Quoted 19 lines · AngeloGioacchino Del Regno

Il 14/02/23 04:11, Yong Wu ha scritto:

Quoted 11 lines

Add iova_region_larb_msk for mt8195. We separate the 16GB iova regions
by each device's larbid/portid.
Refer to include/dt-bindings/memory/mt8195-memory-port.h

Define a new macro MT8192_MULTI_REGION_MAX_NR to indicate the index of mt8xxx_larb_region_msk and
"struct mtk_iommu_iova_region mt8192_multi_dom" are the same.

Sorry but... you *are* introducing a iova_region_larb_msk for MT8192 later,
while you're introducing definitions containing the MT8192 name in a commit
that introduces the same but for MT8195.

Quoted 3 lines

I strongly suggest to, at this point, introduce MT8192 region mask here, along
with the MT8192 definitions, and MT8195 after that :-)

Make sense.
Thanks very much. Will fix it.

Quoted 6 lines

Regards,
Angelo

author maintainer
#11 lore raw

PATCH v3 6/11 iommu/mediatek: mt8186: Add iova_region_larb_msk

In reply to Yong Wu

Add iova_region_larb_msk for mt8186. We separate the 16GB iova regions by each device's larbid/portid.
Note: larb5/6/10/12/14/15/18 connect nothing in this SoC. Refer to include/dt-bindings/memory/mt8186-memory-port.h

---
 drivers/iommu/mtk_iommu.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)
drivers/iommu/mtk_iommu.c +13 -0
index 3d9986fb49a5..49f018138d9a 100644
@@ -1525,6 +1525,18 @@ static const struct mtk_iommu_plat_data mt8183_data = {
 	.larbid_remap = {{0}, {4}, {5}, {6}, {7}, {2}, {3}, {1}},
 };
 
+static const unsigned int mt8186_larb_region_msk[MT8192_MULTI_REGION_MAX_NR][MTK_LARB_NR_MAX] = {
+	[0] = {~0, ~0, ~0},			/* Region0: all ports for larb0/1/2 */
+	[1] = {0, 0, 0, 0, ~0, 0, 0, ~0},	/* Region1: larb4/7 */
+	[2] = {0, 0, 0, 0, 0, 0, 0, 0,		/* Region2: larb8/9/11/13/16/17/19/20 */
+	       ~0, ~0, 0, ~0, 0, ~(u32)(BIT(9) | BIT(10)), 0, 0,
+						/* larb13: the other ports except port9/10 */
+	       ~0, ~0, 0, ~0, ~0},
+	[3] = {0},
+	[4] = {[13] = BIT(9) | BIT(10)},	/* larb13 port9/10 */
+	[5] = {[14] = ~0},			/* larb14 */
+};
+
 static const struct mtk_iommu_plat_data mt8186_data_mm = {
 	.m4u_plat       = M4U_MT8186,
 	.flags          = HAS_BCLK | HAS_SUB_COMM_2BITS | OUT_ORDER_WR_EN |
@@ -1537,6 +1549,7 @@ static const struct mtk_iommu_plat_data mt8186_data_mm = {
 	.banks_enable   = {true},
 	.iova_region    = mt8192_multi_dom,
 	.iova_region_nr = ARRAY_SIZE(mt8192_multi_dom),
+	.iova_region_larb_msk = mt8186_larb_region_msk,
 };
 
 static const struct mtk_iommu_plat_data mt8192_data = {
Signature · 1 line
2.18.0
maintainer Reviewed-by on v3 6/11
#12 lore raw

Re: [PATCH v3 06/11] iommu/mediatek: mt8186: Add iova_region_larb_msk

Il 14/02/23 04:11, Yong Wu ha scritto:

Quoted 6 lines

Add iova_region_larb_msk for mt8186. We separate the 16GB iova regions by each device's larbid/portid.
Note: larb5/6/10/12/14/15/18 connect nothing in this SoC. Refer to include/dt-bindings/memory/mt8186-memory-port.h

author maintainer
#13 lore raw

PATCH v3 7/11 iommu/mediatek: mt8192: Add iova_region_larb_msk

In reply to Yong Wu

Add iova_region_larb_msk for mt8192. We separate the 16GB iova regions by each device's larbid/portid.
Note: larb3/6/8/10/12/15 connect nothing in this SoC.
Refer to the comment in include/dt-bindings/memory/mt8192-larb-port.h

---
 drivers/iommu/mtk_iommu.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
drivers/iommu/mtk_iommu.c +12 -0
index 49f018138d9a..995e59cabf6c 100644
@@ -1552,6 +1552,17 @@ static const struct mtk_iommu_plat_data mt8186_data_mm = {
 	.iova_region_larb_msk = mt8186_larb_region_msk,
 };
 
+static const unsigned int mt8192_larb_region_msk[MT8192_MULTI_REGION_MAX_NR][MTK_LARB_NR_MAX] = {
+	[0] = {~0, ~0},				/* Region0: larb0/1 */
+	[1] = {0, 0, 0, 0, ~0, ~0, 0, ~0},	/* Region1: larb4/5/7 */
+	[2] = {0, 0, ~0, 0, 0, 0, 0, 0,		/* Region2: larb2/9/11/13/14/16/17/18/19/20 */
+	       0, ~0, 0, ~0, 0, ~(u32)(BIT(9) | BIT(10)), ~(u32)(BIT(4) | BIT(5)), 0,
+	       ~0, ~0, ~0, ~0, ~0},
+	[3] = {0},
+	[4] = {[13] = BIT(9) | BIT(10)},	/* larb13 port9/10 */
+	[5] = {[14] = BIT(4) | BIT(5)},		/* larb14 port4/5 */
+};
+
 static const struct mtk_iommu_plat_data mt8192_data = {
 	.m4u_plat       = M4U_MT8192,
 	.flags          = HAS_BCLK | HAS_SUB_COMM_2BITS | OUT_ORDER_WR_EN |
@@ -1561,6 +1572,7 @@ static const struct mtk_iommu_plat_data mt8192_data = {
 	.banks_enable   = {true},
 	.iova_region    = mt8192_multi_dom,
 	.iova_region_nr = ARRAY_SIZE(mt8192_multi_dom),
+	.iova_region_larb_msk = mt8192_larb_region_msk,
 	.larbid_remap   = {{0}, {1}, {4, 5}, {7}, {2}, {9, 11, 19, 20},
 			   {0, 14, 16}, {0, 13, 18, 17}},
 };
Signature · 1 line
2.18.0
author maintainer
#14 lore raw

PATCH v3 8/11 iommu/mediatek: Add a gap for the iova regions

In reply to Yong Wu

As the removed property in the vcodec dt-binding, the property is:
dma-ranges = <0x1 0x0 0x0 0x40000000 0x0 0xfff00000>;

The length is 0xfff0_0000 rather than 0x1_0000_0000, this means it requires 1M as a gap. This is because the end address for some vcodec HW is (address + size). If the size is 4G, the end address may be 0x2_0000_0000, and the width for vcodec register only is 32, then the HW may get the ZERO address.

Currently the consumer's dma-ranges property doesn't work, IOMMU has to consider this case. Add a bigger gap(8M) for all the regions to avoid it.

---
 drivers/iommu/mtk_iommu.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
drivers/iommu/mtk_iommu.c +5 -5
index 995e59cabf6c..4503fba72384 100644
@@ -329,7 +329,7 @@ static LIST_HEAD(m4ulist);	/* List all the M4U HWs */
 #define for_each_m4u(data, head)  list_for_each_entry(data, head, list)
 
 static const struct mtk_iommu_iova_region single_domain[] = {
-	{.iova_base = 0,		.size = SZ_4G},
+	{.iova_base = 0,		.size = SZ_4G - SZ_8M},
 };
 
 #define MT8192_MULTI_REGION_MAX_NR	6
@@ -338,11 +338,11 @@ static const struct mtk_iommu_iova_region single_domain[] = {
 				 MT8192_MULTI_REGION_MAX_NR : 1)
 
 static const struct mtk_iommu_iova_region mt8192_multi_dom[MT8192_MULTI_REGION_NR] = {
-	{ .iova_base = 0x0,		.size = SZ_4G},		/* 0 ~ 4G */
+	{ .iova_base = 0x0,		.size = SZ_4G - SZ_8M},	/* 0 ~ 4G, 8M as a gap. */
 	#if IS_ENABLED(CONFIG_ARCH_DMA_ADDR_T_64BIT)
-	{ .iova_base = SZ_4G,		.size = SZ_4G},		/* 4G ~ 8G */
-	{ .iova_base = SZ_4G * 2,	.size = SZ_4G},		/* 8G ~ 12G */
-	{ .iova_base = SZ_4G * 3,	.size = SZ_4G},		/* 12G ~ 16G */
+	{ .iova_base = SZ_4G,		.size = SZ_4G - SZ_8M},	/* 4G ~ 8G */
+	{ .iova_base = SZ_4G * 2,	.size = SZ_4G - SZ_8M},	/* 8G ~ 12G */
+	{ .iova_base = SZ_4G * 3,	.size = SZ_4G - SZ_8M},	/* 12G ~ 16G */
 
 	{ .iova_base = 0x240000000ULL,	.size = 0x4000000},	/* CCU0 */
 	{ .iova_base = 0x244000000ULL,	.size = 0x4000000},	/* CCU1 */
Signature · 1 line
2.18.0
maintainer Reviewed-by on v3 8/11
#15 lore raw

Re: [PATCH v3 08/11] iommu/mediatek: Add a gap for the iova regions

Il 14/02/23 04:11, Yong Wu ha scritto:

Quoted 9 lines

As the removed property in the vcodec dt-binding, the property is:
dma-ranges = <0x1 0x0 0x0 0x40000000 0x0 0xfff00000>;

The length is 0xfff0_0000 rather than 0x1_0000_0000, this means it requires 1M as a gap. This is because the end address for some vcodec HW is (address + size). If the size is 4G, the end address may be 0x2_0000_0000, and the width for vcodec register only is 32, then the HW may get the ZERO address.

Quoted 5 lines

Currently the consumer's dma-ranges property doesn't work, IOMMU has to consider this case. Add a bigger gap(8M) for all the regions to avoid it.

author maintainer
#16 lore raw

PATCH v3 9/11 arm64: dts: mt8195: Add dma-ranges for the parent "soc" node

In reply to Yong Wu

After commit f1ad5338a4d5 ("of: Fix "dma-ranges" handling for bus controllers"), the dma-ranges property is not allowed for the leaf node. But our iommu/dma-ranges is 16GB, we still expect separate the 16GB dma-range like:
a) display is in 0 - 4GB;
b) vcodec is in 4GB - 8GB;
c) camera is in 8GB - 12GB.
We can not expect all the masters add a parent node for them, especial for the existed drivers/nodes.
Thus, we add whole the 16GB dma-ranges in the parent "soc" node.

---
 arch/arm64/boot/dts/mediatek/mt8195.dtsi | 1 +
 1 file changed, 1 insertion(+)
arch/arm64/boot/dts/mediatek/mt8195.dtsi +1 -0
index 8f1264d5290b..0f5a31ca1b70 100644
@@ -364,6 +364,7 @@
 		#size-cells = <2>;
 		compatible = "simple-bus";
 		ranges;
+		dma-ranges = <0x0 0x0 0x0 0x0 0x4 0x0>;
 
 		gic: interrupt-controller@c000000 {
 			compatible = "arm,gic-v3";
Signature · 1 line
2.18.0
author maintainer
#17 lore raw

PATCH v3 10/11 arm64: dts: mt8195: Remove the unnecessary dma-ranges

In reply to Yong Wu

After we add the dma-ranges in the parent "soc" node, this property is unnecessary for the leaf node.

---
 arch/arm64/boot/dts/mediatek/mt8195.dtsi | 3 ---
 1 file changed, 3 deletions(-)
arch/arm64/boot/dts/mediatek/mt8195.dtsi +0 -3
index 0f5a31ca1b70..7d87cbabc9f1 100644
@@ -2300,7 +2300,6 @@
 			power-domains = <&spm MT8195_POWER_DOMAIN_VENC>;
 			#address-cells = <2>;
 			#size-cells = <2>;
-			dma-ranges = <0x1 0x0 0x0 0x40000000 0x0 0xfff00000>;
 		};
 
 		jpgdec-master {
@@ -2312,7 +2311,6 @@
 				 <&iommu_vdo M4U_PORT_L19_JPGDEC_BSDMA1>,
 				 <&iommu_vdo M4U_PORT_L19_JPGDEC_BUFF_OFFSET1>,
 				 <&iommu_vdo M4U_PORT_L19_JPGDEC_BUFF_OFFSET0>;
-			dma-ranges = <0x1 0x0 0x0 0x40000000 0x0 0xfff00000>;
 			#address-cells = <2>;
 			#size-cells = <2>;
 			ranges;
@@ -2384,7 +2382,6 @@
 					<&iommu_vpp M4U_PORT_L20_JPGENC_C_RDMA>,
 					<&iommu_vpp M4U_PORT_L20_JPGENC_Q_TABLE>,
 					<&iommu_vpp M4U_PORT_L20_JPGENC_BSDMA>;
-			dma-ranges = <0x1 0x0 0x0 0x40000000 0x0 0xfff00000>;
 			#address-cells = <2>;
 			#size-cells = <2>;
 			ranges;
Signature · 1 line
2.18.0
author maintainer
#18 lore raw

PATCH v3 11/11 arm64: dts: mt8186: Add dma-ranges for the parent "soc" node

In reply to Yong Wu

Prepare for the MM nodes whose dma-ranges(iova range) is 16GB.

---
 arch/arm64/boot/dts/mediatek/mt8186.dtsi | 1 +
 1 file changed, 1 insertion(+)
arch/arm64/boot/dts/mediatek/mt8186.dtsi +1 -0
index a0d3e1f731bd..251eace411c0 100644
@@ -324,6 +324,7 @@
 		#address-cells = <2>;
 		#size-cells = <2>;
 		compatible = "simple-bus";
+		dma-ranges = <0x0 0x0 0x0 0x0 0x4 0x0>;
 		ranges;
 
 		gic: interrupt-controller@c000000 {
Signature · 1 line
2.18.0