[PATCH v2 00/21] Imagination BXS-4-64 MC1 GPU support

[PATCH v2 00/21] Imagination BXS-4-64 MC1 GPU support

[Matt Coster]

This GPU is found in the TI AM68 family of SoCs, with initial support
added to the k3-j721s2 devicetree and tested on a TI SK-AM68 board.

A suitable firmware binary can currently be found in the IMG
linux-firmware repository[1] as powervr/rogue_36.53.104.796_v1.fw. A
merge request will be sent out for this within the next few weeks.

No new UAPI will be necessary for this platform as it is sufficiently
similar to the already supported AXE-1-16M.

The firmware successfully boots with basic job submission appearing to
work correctly, as tested with a few of Sascha Willems' Vulkan demos[2].
Note these do not necessarily render correctly due to incomplete UMD
support.

UMD support is still a work in progress. The branch at [3] is nearly
feature complete from a Vulkan perspective. We're currently undertaking
a significant rework of the compiler to better accommodate this and
other Rogue GPUs which means we can't do more comprehensive driver
testing at this point. However, we expect to send a Mesa merge request
for the initial version of the compiler (capable of passing some of the
Vulkan CTS smoke tests) within the next few days, with a fully
functional version of the compiler in place by the end of the year.

There are several dt-bindings changes at the beginning of this series.
We expect the result to be versatile enough to handle all Imagination
Rogue GPUs while being a strong foundation to build bindings for the
newer Volcanic architecture (for which we're currently developing
support).

[1]: https://gitlab.freedesktop.org/imagination/linux-firmware/-/tree/powervr
[2]: https://github.com/SaschaWillems/Vulkan
[3]: https://gitlab.freedesktop.org/imagination/mesa/-/tree/dev/bxs

---
Changes in v2:
- Clarified justification for compatible strings (P1)
- Simplified clocks constraints (P2)
- Simplified power-domains constraints (P3/P4)
- Use normal reg syntax for 64-bit values (P8/P21)
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-0-4ed30e865892@imgtec.com

---
Matt Coster (20):
      dt-bindings: gpu: img: More explicit compatible strings
      dt-bindings: gpu: img: Further constrain clocks
      dt-bindings: gpu: img: Power domain details
      dt-bindings: gpu: img: Allow dma-coherent
      drm/imagination: Use more specific compatible strings
      drm/imagination: Add power domain control
      arm64: dts: ti: k3-am62: New GPU binding details
      dt-bindings: gpu: img: Add BXS-4-64 devicetree bindings
      drm/imagination: Revert to non-threaded IRQs
      drm/imagination: Remove firmware enable_reg
      drm/imagination: Rename event_mask -> status_mask
      drm/imagination: Make has_fixed_data_addr a value
      drm/imagination: Use a lookup table for fw defs
      drm/imagination: Use callbacks for fw irq handling
      drm/imagination: Add register required for RISC-V firmware
      drm/imagination: Move ELF fw utils to common file
      drm/imagination: Add platform overrides infrastructure
      drm/imagination: Add device_memory_force_cpu_cached override
      drm/imagination: Add support for TI AM68 GPU
      arm64: dts: ti: k3-j721s2: Add GPU node

Sarah Walker (1):
      drm/imagination: Add RISC-V firmware processor support

 .../devicetree/bindings/gpu/img,powervr-rogue.yaml | 107 ++++++++++++--
 arch/arm64/boot/dts/ti/k3-am62-main.dtsi           |   3 +-
 arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi         |  12 ++
 drivers/gpu/drm/imagination/Makefile               |   2 +
 drivers/gpu/drm/imagination/pvr_ccb.c              |  25 +++-
 drivers/gpu/drm/imagination/pvr_device.c           | 130 ++++++++--------
 drivers/gpu/drm/imagination/pvr_device.h           |  77 +++++++++-
 drivers/gpu/drm/imagination/pvr_drv.c              |  58 +++++++-
 drivers/gpu/drm/imagination/pvr_fw.c               |  26 +++-
 drivers/gpu/drm/imagination/pvr_fw.h               |  85 +++++------
 drivers/gpu/drm/imagination/pvr_fw_meta.c          |  23 +--
 drivers/gpu/drm/imagination/pvr_fw_mips.c          |  82 ++---------
 drivers/gpu/drm/imagination/pvr_fw_riscv.c         | 163 +++++++++++++++++++++
 drivers/gpu/drm/imagination/pvr_fw_startstop.c     |  21 +++
 drivers/gpu/drm/imagination/pvr_fw_util.c          |  67 +++++++++
 drivers/gpu/drm/imagination/pvr_gem.c              |   3 +
 drivers/gpu/drm/imagination/pvr_gem.h              |   7 +-
 drivers/gpu/drm/imagination/pvr_mmu.c              |   7 +-
 drivers/gpu/drm/imagination/pvr_power.c            | 118 ++++++++++++++-
 drivers/gpu/drm/imagination/pvr_power.h            |   3 +
 drivers/gpu/drm/imagination/pvr_queue.c            |  23 ++-
 drivers/gpu/drm/imagination/pvr_rogue_cr_defs.h    |  17 ++-
 drivers/gpu/drm/imagination/pvr_rogue_riscv.h      |  41 ++++++
 23 files changed, 864 insertions(+), 236 deletions(-)
---
base-commit: d78f0ee0406803cda8801fd5201746ccf89e5e4a
change-id: 20241021-sets-bxs-4-64-patch-v1-44cdf9cc555f

[PATCH v2 01/21] dt-bindings: gpu: img: More explicit compatible strings

[Matt Coster]

The current compatible strings are not specific enough to constrain the
hardware in devicetree. For example, the current "img,img-axe" string
refers to the entire family of Series AXE GPUs. The more specific
"img,img-axe-1-16m" string refers to the AXE-1-16M GPU which, unlike the
rest of its family, only uses a single power domain.

While in this instance there is already "ti,am62-gpu" for the more
specific case, the intent here is to draw a line between properties
inherent to the IP core and choices made by the silicon vendor at
integration time. For example, the number of power domains is a property
of the IP core, whereas the decision to use one or three clocks (see
next patch) is a vendor one.

Work is currently underway to add support for volcanic-based Imagination
GPUs, for which bindings will be added in "img,powervr-volcanic.yaml".
The split between rogue and volcanic cores is non-obvious at times, so
add a generic top-level "img,img-rogue" compatible string here to allow
for simpler differentiation in devicetrees without referring back to the
bindings.

Make these changes now before introducing more compatible strings to keep
the legacy versions to a minimum.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- Clarified justification for compatible strings
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-1-4ed30e865892@imgtec.com
---
 .../devicetree/bindings/gpu/img,powervr-rogue.yaml    | 19 ++++++++++++++-----
 1 file changed, 14 insertions(+), 5 deletions(-)

diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
index 256e252f8087fa0d6081f771a01601d34b66fe19..ef7070daf213277d0190fe319e202fdc597337d4 100644
--- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
+++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
@@ -12,10 +12,19 @@ maintainers:
 
 properties:
   compatible:
-    items:
-      - enum:
-          - ti,am62-gpu
-      - const: img,img-axe # IMG AXE GPU model/revision is fully discoverable
+    oneOf:
+      - items:
+          - enum:
+              - ti,am62-gpu
+          - const: img,img-axe-1-16m
+          - const: img,img-rogue
+
+      # This legacy combination of compatible strings was introduced early on before the more
+      # specific GPU identifiers were used. Keep it around here for compatibility, but never use
+      # "img,img-axe" in new devicetrees.
+      - items:
+          - const: ti,am62-gpu
+          - const: img,img-axe
 
   reg:
     maxItems: 1
@@ -64,7 +73,7 @@ examples:
     #include <dt-bindings/soc/ti,sci_pm_domain.h>
 
     gpu@fd00000 {
-        compatible = "ti,am62-gpu", "img,img-axe";
+        compatible = "ti,am62-gpu", "img,img-axe-1-16m", "img,img-rogue";
         reg = <0x0fd00000 0x20000>;
         clocks = <&k3_clks 187 0>;
         clock-names = "core";

-- 
2.47.0

Re: [PATCH v2 01/21] dt-bindings: gpu: img: More explicit compatible strings

[Krzysztof Kozlowski]

On Mon, Nov 18, 2024 at 01:01:53PM +0000, Matt Coster wrote:
> The current compatible strings are not specific enough to constrain the

No, they are specific enough.

> hardware in devicetree. For example, the current "img,img-axe" string
> refers to the entire family of Series AXE GPUs. The more specific
> "img,img-axe-1-16m" string refers to the AXE-1-16M GPU which, unlike the
> rest of its family, only uses a single power domain.
> 
> While in this instance there is already "ti,am62-gpu" for the more

That's the specific compatible.

> specific case, the intent here is to draw a line between properties
> inherent to the IP core and choices made by the silicon vendor at
> integration time. For example, the number of power domains is a property
> of the IP core, whereas the decision to use one or three clocks (see
> next patch) is a vendor one.
> 
> Work is currently underway to add support for volcanic-based Imagination
> GPUs, for which bindings will be added in "img,powervr-volcanic.yaml".
> The split between rogue and volcanic cores is non-obvious at times, so
> add a generic top-level "img,img-rogue" compatible string here to allow
> for simpler differentiation in devicetrees without referring back to the
> bindings.
> 
> Make these changes now before introducing more compatible strings to keep
> the legacy versions to a minimum.
> 
> Signed-off-by: Matt Coster <matt.coster@imgtec.com>
> ---
> Changes in v2:
> - Clarified justification for compatible strings
> - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-1-4ed30e865892@imgtec.com
> ---
>  .../devicetree/bindings/gpu/img,powervr-rogue.yaml    | 19 ++++++++++++++-----
>  1 file changed, 14 insertions(+), 5 deletions(-)
> 
> diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> index 256e252f8087fa0d6081f771a01601d34b66fe19..ef7070daf213277d0190fe319e202fdc597337d4 100644
> --- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> +++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> @@ -12,10 +12,19 @@ maintainers:
>  
>  properties:
>    compatible:
> -    items:
> -      - enum:
> -          - ti,am62-gpu
> -      - const: img,img-axe # IMG AXE GPU model/revision is fully discoverable
> +    oneOf:
> +      - items:
> +          - enum:
> +              - ti,am62-gpu
> +          - const: img,img-axe-1-16m
> +          - const: img,img-rogue

NAK, that's ABI break. You are stuck with whatever you sent earlier.


> +
> +      # This legacy combination of compatible strings was introduced early on before the more
> +      # specific GPU identifiers were used. Keep it around here for compatibility, but never use
> +      # "img,img-axe" in new devicetrees.

Wrap according to Linux coding style. But anyway this is not needed,
just deprecate things.

> +      - items:
> +          - const: ti,am62-gpu
> +          - const: img,img-axe

So you want to use deprecated here?

Anyway, entire change is not needed and without proper justification.
Your marketing terms are not proper justification.

Best regards,
Krzysztof

[PATCH v2 02/21] dt-bindings: gpu: img: Further constrain clocks

[Matt Coster]

All Imagination GPUs use three clocks: core, mem and sys. All reasonably
modern Imagination GPUs also support a single-clock mode where the SoC
only hooks up core and the other two are derived internally. On GPUs which
support this mode, it is the default and most commonly used integration.

Codify this "1 or 3" constraint in our bindings and hang the specifics off
the vendor compatible string to mirror the integration-time choice.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- Simplified clocks constraints (P2)
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-2-4ed30e865892@imgtec.com
---
 .../devicetree/bindings/gpu/img,powervr-rogue.yaml   | 20 +++++++++++---------
 1 file changed, 11 insertions(+), 9 deletions(-)

diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
index ef7070daf213277d0190fe319e202fdc597337d4..3b5a5b966585ac29ad104c7aef19881eca73ce80 100644
--- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
+++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
@@ -29,16 +29,16 @@ properties:
   reg:
     maxItems: 1
 
-  clocks:
-    minItems: 1
-    maxItems: 3
+  clocks: true
 
   clock-names:
-    items:
-      - const: core
-      - const: mem
-      - const: sys
-    minItems: 1
+    oneOf:
+      - items:
+          - const: core
+      - items:
+          - const: core
+          - const: mem
+          - const: sys
 
   interrupts:
     maxItems: 1
@@ -56,11 +56,13 @@ required:
 additionalProperties: false
 
 allOf:
+  # Vendor integrations using a single clock domain
   - if:
       properties:
         compatible:
           contains:
-            const: ti,am62-gpu
+            anyOf:
+              - const: ti,am62-gpu
     then:
       properties:
         clocks:

-- 
2.47.0

Re: [PATCH v2 02/21] dt-bindings: gpu: img: Further constrain clocks

[Krzysztof Kozlowski]

On Mon, Nov 18, 2024 at 01:01:54PM +0000, Matt Coster wrote:
> All Imagination GPUs use three clocks: core, mem and sys. All reasonably
> modern Imagination GPUs also support a single-clock mode where the SoC
> only hooks up core and the other two are derived internally. On GPUs which
> support this mode, it is the default and most commonly used integration.
> 
> Codify this "1 or 3" constraint in our bindings and hang the specifics off
> the vendor compatible string to mirror the integration-time choice.
> 
> Signed-off-by: Matt Coster <matt.coster@imgtec.com>
> ---
> Changes in v2:
> - Simplified clocks constraints (P2)
> - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-2-4ed30e865892@imgtec.com
> ---
>  .../devicetree/bindings/gpu/img,powervr-rogue.yaml   | 20 +++++++++++---------
>  1 file changed, 11 insertions(+), 9 deletions(-)
> 
> diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> index ef7070daf213277d0190fe319e202fdc597337d4..3b5a5b966585ac29ad104c7aef19881eca73ce80 100644
> --- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> +++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> @@ -29,16 +29,16 @@ properties:
>    reg:
>      maxItems: 1
>  
> -  clocks:
> -    minItems: 1
> -    maxItems: 3
> +  clocks: true

NAK. Widest constraints always stay here.

>  
>    clock-names:
> -    items:
> -      - const: core
> -      - const: mem
> -      - const: sys
> -    minItems: 1
> +    oneOf:
> +      - items:
> +          - const: core
> +      - items:
> +          - const: core
> +          - const: mem
> +          - const: sys

Why? Sorry, this is just wrong. This deviec has specific, one clock.

Anyway, this needs to use old syntax. You change here nothing.

NAK

>  
>    interrupts:
>      maxItems: 1
> @@ -56,11 +56,13 @@ required:
>  additionalProperties: false
>  
>  allOf:
> +  # Vendor integrations using a single clock domain
>    - if:
>        properties:
>          compatible:
>            contains:
> -            const: ti,am62-gpu
> +            anyOf:

Nope. That's not needed, not explained.

Best regards,
Krzysztof

Re: [PATCH v2 02/21] dt-bindings: gpu: img: Further constrain clocks

[Parthiban]

On 11/18/24 6:31 PM, Matt Coster wrote:
> All Imagination GPUs use three clocks: core, mem and sys. All reasonably
> modern Imagination GPUs also support a single-clock mode where the SoC
> only hooks up core and the other two are derived internally. On GPUs which
> support this mode, it is the default and most commonly used integration.
> 
> Codify this "1 or 3" constraint in our bindings and hang the specifics off
> the vendor compatible string to mirror the integration-time choice.
> 
> Signed-off-by: Matt Coster <matt.coster@imgtec.com>
> ---
> Changes in v2:
> - Simplified clocks constraints (P2)
> - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-2-4ed30e865892@imgtec.com
> ---
>  .../devicetree/bindings/gpu/img,powervr-rogue.yaml   | 20 +++++++++++---------
>  1 file changed, 11 insertions(+), 9 deletions(-)
> 
> diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> index ef7070daf213277d0190fe319e202fdc597337d4..3b5a5b966585ac29ad104c7aef19881eca73ce80 100644
> --- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> +++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> @@ -29,16 +29,16 @@ properties:
>    reg:
>      maxItems: 1
>  
> -  clocks:
> -    minItems: 1
> -    maxItems: 3
> +  clocks: true
>  
>    clock-names:
> -    items:
> -      - const: core
> -      - const: mem
> -      - const: sys
> -    minItems: 1
> +    oneOf:
> +      - items:
> +          - const: core
> +      - items:
> +          - const: core
> +          - const: mem
> +          - const: sys
Clock for GE8300 in Allwinner A133 is organized with core, bus and additional pll as
input. Where "bus" controlled as gate clock and optionally using pll.

If am not wrong, GE8300 also comes under rogue architecture without mem and sys clocks.
Does this needs to be considered into separate bindings?

Thanks,
Parthiban

>  
>    interrupts:
>      maxItems: 1
> @@ -56,11 +56,13 @@ required:
>  additionalProperties: false
>  
>  allOf:
> +  # Vendor integrations using a single clock domain
>    - if:
>        properties:
>          compatible:
>            contains:
> -            const: ti,am62-gpu
> +            anyOf:
> +              - const: ti,am62-gpu
>      then:
>        properties:
>          clocks:
> 

[PATCH v2 03/21] dt-bindings: gpu: img: Power domain details

[Matt Coster]

The single existing GPU (AXE-1-16M) only requires a single power domain.
Subsequent patches will add support for BXS-4-64 MC1, which has two power
domains. Add infrastructure now to allow for this.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- Simplified power-domains constraints P3
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-3-4ed30e865892@imgtec.com
---
 .../devicetree/bindings/gpu/img,powervr-rogue.yaml | 25 ++++++++++++++++++++--
 1 file changed, 23 insertions(+), 2 deletions(-)

diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
index 3b5a5b966585ac29ad104c7aef19881eca73ce80..c629f54c86c441b4cc9e57925f1d65129cbe285b 100644
--- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
+++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
@@ -43,8 +43,15 @@ properties:
   interrupts:
     maxItems: 1
 
-  power-domains:
-    maxItems: 1
+  power-domains: true
+
+  power-domain-names:
+    oneOf:
+      - items:
+          - const: a
+      - items:
+          - const: a
+          - const: b
 
 required:
   - compatible
@@ -52,10 +59,23 @@ required:
   - clocks
   - clock-names
   - interrupts
+  - power-domains
+  - power-domain-names
 
 additionalProperties: false
 
 allOf:
+  # Cores with a single power domain
+  - if:
+      properties:
+        compatible:
+          contains:
+            anyOf:
+              - const: img,img-axe-1-16m
+    then:
+      properties:
+        power-domains:
+          maxItems: 1
   # Vendor integrations using a single clock domain
   - if:
       properties:
@@ -81,4 +101,5 @@ examples:
         clock-names = "core";
         interrupts = <GIC_SPI 86 IRQ_TYPE_LEVEL_HIGH>;
         power-domains = <&k3_pds 187 TI_SCI_PD_EXCLUSIVE>;
+        power-domain-names = "a";
     };

-- 
2.47.0

Re: [PATCH v2 03/21] dt-bindings: gpu: img: Power domain details

[Krzysztof Kozlowski]

On Mon, Nov 18, 2024 at 01:01:55PM +0000, Matt Coster wrote:
> The single existing GPU (AXE-1-16M) only requires a single power domain.
> Subsequent patches will add support for BXS-4-64 MC1, which has two power
> domains. Add infrastructure now to allow for this.
> 
> Signed-off-by: Matt Coster <matt.coster@imgtec.com>
> ---
> Changes in v2:
> - Simplified power-domains constraints P3
> - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-3-4ed30e865892@imgtec.com
> ---
>  .../devicetree/bindings/gpu/img,powervr-rogue.yaml | 25 ++++++++++++++++++++--
>  1 file changed, 23 insertions(+), 2 deletions(-)
> 
> diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> index 3b5a5b966585ac29ad104c7aef19881eca73ce80..c629f54c86c441b4cc9e57925f1d65129cbe285b 100644
> --- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> +++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> @@ -43,8 +43,15 @@ properties:
>    interrupts:
>      maxItems: 1
>  
> -  power-domains:
> -    maxItems: 1
> +  power-domains: true

NAK for the same reasons.

> +
> +  power-domain-names:
> +    oneOf:
> +      - items:
> +          - const: a
> +      - items:
> +          - const: a
> +          - const: b

No, list items and use minItems.

But what's more: there is no user of it. Sorry, that's a no go. You
probably organized patches incorrectly, but this on its own is just
incorrect.


>  
>  required:
>    - compatible
> @@ -52,10 +59,23 @@ required:
>    - clocks
>    - clock-names
>    - interrupts
> +  - power-domains
> +  - power-domain-names

That's an ABI break. Explain in commit msg why and its impact on users.

>  
>  additionalProperties: false
>  
>  allOf:
> +  # Cores with a single power domain
> +  - if:
> +      properties:
> +        compatible:
> +          contains:
> +            anyOf:

No, drop.

Best regards,
Krzysztof

[PATCH v2 04/21] dt-bindings: gpu: img: Allow dma-coherent

[Matt Coster]

This attribute will be required for the BXS-4-64 MC1 and will be enabled in
the DTS for the TI k3-j721s2 in a subsequent patch; add it now so
dtbs_check doesn't complain later.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- Simplified power-domains constraints
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-4-4ed30e865892@imgtec.com
---
 Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
index c629f54c86c441b4cc9e57925f1d65129cbe285b..9dc55a6d0d4023983a3fc480340351f3fa974ce5 100644
--- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
+++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
@@ -53,6 +53,8 @@ properties:
           - const: a
           - const: b
 
+  dma-coherent: true
+
 required:
   - compatible
   - reg

-- 
2.47.0

Re: [PATCH v2 04/21] dt-bindings: gpu: img: Allow dma-coherent

[Krzysztof Kozlowski]

On Mon, Nov 18, 2024 at 01:01:56PM +0000, Matt Coster wrote:
> This attribute will be required for the BXS-4-64 MC1 and will be enabled in
> the DTS for the TI k3-j721s2 in a subsequent patch; add it now so
> dtbs_check doesn't complain later.

That's not a reason. You add it because device is DMA coherent, not for
some imaginary future checker errors.

Best regards,
Krzysztof

[PATCH v2 05/21] drm/imagination: Use more specific compatible strings

[Matt Coster]

Follow-on from the companion dt-bindings change ("dt-bindings: gpu: img:
More explicit compatible strings"), deprecating "img,img-axe" in favour of
the more specific "img,img-axe-1-16m".

Keep the previous compatible string around for backwards compatibility.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-5-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_drv.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/drivers/gpu/drm/imagination/pvr_drv.c b/drivers/gpu/drm/imagination/pvr_drv.c
index 684a9b9a2247b8f5b0f4a91ec984af2cff5a4c29..fbd8802abcf1271e260209957d95ea705dbe7f14 100644
--- a/drivers/gpu/drm/imagination/pvr_drv.c
+++ b/drivers/gpu/drm/imagination/pvr_drv.c
@@ -1471,6 +1471,16 @@ static void pvr_remove(struct platform_device *plat_dev)
 }
 
 static const struct of_device_id dt_match[] = {
+	{ .compatible = "img,img-rogue", .data = NULL },
+
+	/* All supported GPU models */
+	{ .compatible = "img,img-axe-1-16m", .data = NULL },
+
+	/*
+	 * This legacy compatible string was introduced early on before the more specific GPU
+	 * identifiers were used. Keep it around here for compatibility, but never use
+	 * "img,img-axe" in new devicetrees.
+	 */
 	{ .compatible = "img,img-axe", .data = NULL },
 	{}
 };

-- 
2.47.0

Re: [PATCH v2 05/21] drm/imagination: Use more specific compatible strings

[Krzysztof Kozlowski]

On 18/11/2024 14:01, Matt Coster wrote:
> Follow-on from the companion dt-bindings change ("dt-bindings: gpu: img:
> More explicit compatible strings"), deprecating "img,img-axe" in favour of
> the more specific "img,img-axe-1-16m".
> 
> Keep the previous compatible string around for backwards compatibility.
> 
> Signed-off-by: Matt Coster <matt.coster@imgtec.com>
> ---
> Changes in v2:
> - None
> - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-5-4ed30e865892@imgtec.com
> ---
>  drivers/gpu/drm/imagination/pvr_drv.c | 10 ++++++++++
>  1 file changed, 10 insertions(+)
> 
> diff --git a/drivers/gpu/drm/imagination/pvr_drv.c b/drivers/gpu/drm/imagination/pvr_drv.c
> index 684a9b9a2247b8f5b0f4a91ec984af2cff5a4c29..fbd8802abcf1271e260209957d95ea705dbe7f14 100644
> --- a/drivers/gpu/drm/imagination/pvr_drv.c
> +++ b/drivers/gpu/drm/imagination/pvr_drv.c
> @@ -1471,6 +1471,16 @@ static void pvr_remove(struct platform_device *plat_dev)
>  }
>  
>  static const struct of_device_id dt_match[] = {
> +	{ .compatible = "img,img-rogue", .data = NULL },
> +
> +	/* All supported GPU models */
> +	{ .compatible = "img,img-axe-1-16m", .data = NULL },

This makes no sense. It's the same as rogue. Use compatibility properly.

Best regards,
Krzysztof

[PATCH v2 06/21] drm/imagination: Add power domain control

[Matt Coster]

The first supported GPU only used a single power domain so this was
automatically handled by the device runtime.

In order to support multiple power domains, they must be enumerated from
devicetree and linked to both the GPU device and each other to ensure
correct power sequencing at start time.

For all Imagination Rogue GPUs, power domains are named "a", "b", etc. and
the sequence A->B->... is always valid for startup with the reverse true
for shutdown. Note this is not always the *only* valid sequence, but it's
simple and does not require special-casing for different GPU power
topologies.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-6-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_device.h |   8 +++
 drivers/gpu/drm/imagination/pvr_drv.c    |  10 ++-
 drivers/gpu/drm/imagination/pvr_power.c  | 114 +++++++++++++++++++++++++++++++
 drivers/gpu/drm/imagination/pvr_power.h  |   3 +
 4 files changed, 134 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/imagination/pvr_device.h b/drivers/gpu/drm/imagination/pvr_device.h
index b574e23d484ba80785a2220e046dbab3f91f6e15..470945ccfcac7ce91161aa6c70d33177fbb3533f 100644
--- a/drivers/gpu/drm/imagination/pvr_device.h
+++ b/drivers/gpu/drm/imagination/pvr_device.h
@@ -18,6 +18,7 @@
 #include <linux/bits.h>
 #include <linux/compiler_attributes.h>
 #include <linux/compiler_types.h>
+#include <linux/device.h>
 #include <linux/io.h>
 #include <linux/iopoll.h>
 #include <linux/kernel.h>
@@ -130,6 +131,13 @@ struct pvr_device {
 	 */
 	struct clk *mem_clk;
 
+	struct pvr_device_power {
+		struct device **domain_devs;
+		struct device_link **domain_links;
+
+		u32 domain_count;
+	} power;
+
 	/** @irq: IRQ number. */
 	int irq;
 
diff --git a/drivers/gpu/drm/imagination/pvr_drv.c b/drivers/gpu/drm/imagination/pvr_drv.c
index fbd8802abcf1271e260209957d95ea705dbe7f14..1ab97933e14f20ee3fbf603c23b8dde2d33572c2 100644
--- a/drivers/gpu/drm/imagination/pvr_drv.c
+++ b/drivers/gpu/drm/imagination/pvr_drv.c
@@ -1409,13 +1409,17 @@ pvr_probe(struct platform_device *plat_dev)
 
 	platform_set_drvdata(plat_dev, drm_dev);
 
+	err = pvr_power_domains_init(pvr_dev);
+	if (err)
+		goto err_context_fini;
+
 	init_rwsem(&pvr_dev->reset_sem);
 
 	pvr_context_device_init(pvr_dev);
 
 	err = pvr_queue_device_init(pvr_dev);
 	if (err)
-		goto err_context_fini;
+		goto err_power_domains_fini;
 
 	devm_pm_runtime_enable(&plat_dev->dev);
 	pm_runtime_mark_last_busy(&plat_dev->dev);
@@ -1448,6 +1452,9 @@ pvr_probe(struct platform_device *plat_dev)
 err_context_fini:
 	pvr_context_device_fini(pvr_dev);
 
+err_power_domains_fini:
+	pvr_power_domains_fini(pvr_dev);
+
 	return err;
 }
 
@@ -1468,6 +1475,7 @@ static void pvr_remove(struct platform_device *plat_dev)
 	pvr_watchdog_fini(pvr_dev);
 	pvr_queue_device_fini(pvr_dev);
 	pvr_context_device_fini(pvr_dev);
+	pvr_power_domains_fini(pvr_dev);
 }
 
 static const struct of_device_id dt_match[] = {
diff --git a/drivers/gpu/drm/imagination/pvr_power.c b/drivers/gpu/drm/imagination/pvr_power.c
index ba7816fd28ec77e6ca5ce408302a413ce1afeb6e..19b079b357df78e8bcdecfa377fc9c05b6e8e4b0 100644
--- a/drivers/gpu/drm/imagination/pvr_power.c
+++ b/drivers/gpu/drm/imagination/pvr_power.c
@@ -10,10 +10,13 @@
 
 #include <drm/drm_drv.h>
 #include <drm/drm_managed.h>
+#include <linux/cleanup.h>
 #include <linux/clk.h>
 #include <linux/interrupt.h>
 #include <linux/mutex.h>
+#include <linux/of.h>
 #include <linux/platform_device.h>
+#include <linux/pm_domain.h>
 #include <linux/pm_runtime.h>
 #include <linux/timer.h>
 #include <linux/types.h>
@@ -431,3 +434,114 @@ pvr_watchdog_fini(struct pvr_device *pvr_dev)
 {
 	cancel_delayed_work_sync(&pvr_dev->watchdog.work);
 }
+
+int pvr_power_domains_init(struct pvr_device *pvr_dev)
+{
+	struct device *dev = from_pvr_device(pvr_dev)->dev;
+
+	struct device_link **domain_links __free(kfree) = NULL;
+	struct device **domain_devs __free(kfree) = NULL;
+	int domain_count;
+	int link_count;
+
+	char dev_name[2] = "a";
+	int err;
+	int i;
+
+	domain_count = of_count_phandle_with_args(dev->of_node, "power-domains",
+						  "#power-domain-cells");
+	if (domain_count < 0)
+		return domain_count;
+
+	if (domain_count <= 1)
+		return 0;
+
+	link_count = domain_count + (domain_count - 1);
+
+	domain_devs = kcalloc(domain_count, sizeof(*domain_devs), GFP_KERNEL);
+	if (!domain_devs)
+		return -ENOMEM;
+
+	domain_links = kcalloc(link_count, sizeof(*domain_links), GFP_KERNEL);
+	if (!domain_links)
+		return -ENOMEM;
+
+	for (i = 0; i < domain_count; i++) {
+		struct device *domain_dev;
+
+		dev_name[0] = 'a' + i;
+		domain_dev = dev_pm_domain_attach_by_name(dev, dev_name);
+		if (IS_ERR_OR_NULL(domain_dev)) {
+			err = domain_dev ? PTR_ERR(domain_dev) : -ENODEV;
+			goto err_detach;
+		}
+
+		domain_devs[i] = domain_dev;
+	}
+
+	for (i = 0; i < domain_count; i++) {
+		struct device_link *link;
+
+		link = device_link_add(dev, domain_devs[i], DL_FLAG_STATELESS | DL_FLAG_PM_RUNTIME);
+		if (!link) {
+			err = -ENODEV;
+			goto err_unlink;
+		}
+
+		domain_links[i] = link;
+	}
+
+	for (i = domain_count; i < link_count; i++) {
+		struct device_link *link;
+
+		link = device_link_add(domain_devs[i - domain_count + 1],
+				       domain_devs[i - domain_count],
+				       DL_FLAG_STATELESS | DL_FLAG_PM_RUNTIME);
+		if (!link) {
+			err = -ENODEV;
+			goto err_unlink;
+		}
+
+		domain_links[i] = link;
+	}
+
+	pvr_dev->power = (struct pvr_device_power){
+		.domain_devs = no_free_ptr(domain_devs),
+		.domain_links = no_free_ptr(domain_links),
+		.domain_count = domain_count,
+	};
+
+	return 0;
+
+err_unlink:
+	while (--i >= 0)
+		device_link_del(domain_links[i]);
+
+	i = domain_count;
+
+err_detach:
+	while (--i >= 0)
+		dev_pm_domain_detach(domain_devs[i], true);
+
+	return err;
+}
+
+void pvr_power_domains_fini(struct pvr_device *pvr_dev)
+{
+	const int domain_count = pvr_dev->power.domain_count;
+
+	int i = domain_count + (domain_count - 1);
+
+	while (--i >= 0)
+		device_link_del(pvr_dev->power.domain_links[i]);
+
+	i = domain_count;
+
+	while (--i >= 0)
+		dev_pm_domain_detach(pvr_dev->power.domain_devs[i], true);
+
+	kfree(pvr_dev->power.domain_links);
+	kfree(pvr_dev->power.domain_devs);
+
+	pvr_dev->power = (struct pvr_device_power){ 0 };
+}
diff --git a/drivers/gpu/drm/imagination/pvr_power.h b/drivers/gpu/drm/imagination/pvr_power.h
index 9a9312dcb2dab7d36ee8ff7f69e69d126c5469a9..ada85674a7ca762dcf92df40424230e1c3910342 100644
--- a/drivers/gpu/drm/imagination/pvr_power.h
+++ b/drivers/gpu/drm/imagination/pvr_power.h
@@ -38,4 +38,7 @@ pvr_power_put(struct pvr_device *pvr_dev)
 	return pm_runtime_put(drm_dev->dev);
 }
 
+int pvr_power_domains_init(struct pvr_device *pvr_dev);
+void pvr_power_domains_fini(struct pvr_device *pvr_dev);
+
 #endif /* PVR_POWER_H */

-- 
2.47.0

[PATCH v2 07/21] arm64: dts: ti: k3-am62: New GPU binding details

[Matt Coster]

Use the new compatible string introduced earlier (in "dt-bindings: gpu:
img: More explicit compatible strings") and add a name to the single power
domain for this GPU (introduced in "dt-bindings: gpu: img: Power domain
details").

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-7-4ed30e865892@imgtec.com
---
 arch/arm64/boot/dts/ti/k3-am62-main.dtsi | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/boot/dts/ti/k3-am62-main.dtsi b/arch/arm64/boot/dts/ti/k3-am62-main.dtsi
index 5b92aef5b284b78749a01d44184b66e7776a124d..93dcc67c3138ad5b4a7ad6c9bcabb71a2b7e408d 100644
--- a/arch/arm64/boot/dts/ti/k3-am62-main.dtsi
+++ b/arch/arm64/boot/dts/ti/k3-am62-main.dtsi
@@ -693,12 +693,13 @@ ospi0: spi@fc40000 {
 	};
 
 	gpu: gpu@fd00000 {
-		compatible = "ti,am62-gpu", "img,img-axe";
+		compatible = "ti,am62-gpu", "img,img-axe-1-16m", "img,img-rogue";
 		reg = <0x00 0x0fd00000 0x00 0x20000>;
 		clocks = <&k3_clks 187 0>;
 		clock-names = "core";
 		interrupts = <GIC_SPI 86 IRQ_TYPE_LEVEL_HIGH>;
 		power-domains = <&k3_pds 187 TI_SCI_PD_EXCLUSIVE>;
+		power-domain-names = "a";
 	};
 
 	cpsw3g: ethernet@8000000 {

-- 
2.47.0

Re: [PATCH v2 07/21] arm64: dts: ti: k3-am62: New GPU binding details

[Krzysztof Kozlowski]

On 18/11/2024 14:01, Matt Coster wrote:
> Use the new compatible string introduced earlier (in "dt-bindings: gpu:
> img: More explicit compatible strings") and add a name to the single power
> domain for this GPU (introduced in "dt-bindings: gpu: img: Power domain
> details").
> 
> Signed-off-by: Matt Coster <matt.coster@imgtec.com>
> ---
> Changes in v2:
> - None
> - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-7-4ed30e865892@imgtec.com

Do not mix DTS with code. It always goes either to end of patchset or
entirely separate.

To be clear: this cannot be merged into DRM.


> ---
>  arch/arm64/boot/dts/ti/k3-am62-main.dtsi | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/boot/dts/ti/k3-am62-main.dtsi b/arch/arm64/boot/dts/ti/k3-am62-main.dtsi
> index 5b92aef5b284b78749a01d44184b66e7776a124d..93dcc67c3138ad5b4a7ad6c9bcabb71a2b7e408d 100644
> --- a/arch/arm64/boot/dts/ti/k3-am62-main.dtsi
> +++ b/arch/arm64/boot/dts/ti/k3-am62-main.dtsi
> @@ -693,12 +693,13 @@ ospi0: spi@fc40000 {
>  	};
>  
>  	gpu: gpu@fd00000 {
> -		compatible = "ti,am62-gpu", "img,img-axe";
> +		compatible = "ti,am62-gpu", "img,img-axe-1-16m", "img,img-rogue";

Definitely NAK, breaks users and nothing in commit msg provides any help
on that, including why or how it affects potential users.


Best regards,
Krzysztof

[PATCH v2 08/21] dt-bindings: gpu: img: Add BXS-4-64 devicetree bindings

[Matt Coster]

Like the existing AXE-1-16M integration, BXS-4-64 uses the single clock
integration in the TI k3-j721s2.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- Use normal reg syntax for 64-bit values
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-8-4ed30e865892@imgtec.com
---
 .../devicetree/bindings/gpu/img,powervr-rogue.yaml | 41 ++++++++++++++++++++++
 1 file changed, 41 insertions(+)

diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
index 9dc55a6d0d4023983a3fc480340351f3fa974ce5..b620baa56a4caa41246f7b53064d0e3309bdda8e 100644
--- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
+++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
@@ -18,6 +18,11 @@ properties:
               - ti,am62-gpu
           - const: img,img-axe-1-16m
           - const: img,img-rogue
+      - items:
+          - enum:
+              - ti,j721s2-gpu
+          - const: img,img-bxs-4-64
+          - const: img,img-rogue
 
       # This legacy combination of compatible strings was introduced early on before the more
       # specific GPU identifiers were used. Keep it around here for compatibility, but never use
@@ -78,6 +83,18 @@ allOf:
       properties:
         power-domains:
           maxItems: 1
+  # Cores with two power domains
+  - if:
+      properties:
+        compatible:
+          contains:
+            anyOf:
+              - const: img,img-bxs-4-64
+    then:
+      properties:
+        power-domains:
+          minItems: 2
+          maxItems: 2
   # Vendor integrations using a single clock domain
   - if:
       properties:
@@ -85,6 +102,7 @@ allOf:
           contains:
             anyOf:
               - const: ti,am62-gpu
+              - const: ti,j721s2-gpu
     then:
       properties:
         clocks:
@@ -105,3 +123,26 @@ examples:
         power-domains = <&k3_pds 187 TI_SCI_PD_EXCLUSIVE>;
         power-domain-names = "a";
     };
+  - |
+    #include <dt-bindings/interrupt-controller/irq.h>
+    #include <dt-bindings/interrupt-controller/arm-gic.h>
+    #include <dt-bindings/soc/ti,sci_pm_domain.h>
+
+    / {
+        #address-cells = <2>;
+        #size-cells = <2>;
+        interrupt-controller;
+        #interrupt-cells = <3>;
+
+        gpu@4e20000000 {
+            compatible = "ti,j721s2-gpu", "img,img-bxs-4-64", "img,img-rogue";
+            reg = <0x4e 0x20000000 0x00 0x80000>;
+            clocks = <&k3_clks 130 1>;
+            clock-names = "core";
+            interrupts = <GIC_SPI 24 IRQ_TYPE_LEVEL_HIGH>;
+            power-domains = <&k3_pds 130 TI_SCI_PD_EXCLUSIVE>,
+                            <&k3_pds 373 TI_SCI_PD_EXCLUSIVE>;
+            power-domain-names = "a", "b";
+            dma-coherent;
+        };
+    };

-- 
2.47.0

Re: [PATCH v2 08/21] dt-bindings: gpu: img: Add BXS-4-64 devicetree bindings

[Krzysztof Kozlowski]

On Mon, Nov 18, 2024 at 01:02:00PM +0000, Matt Coster wrote:
> Like the existing AXE-1-16M integration, BXS-4-64 uses the single clock
> integration in the TI k3-j721s2.
> 
> Signed-off-by: Matt Coster <matt.coster@imgtec.com>
> ---
> Changes in v2:
> - Use normal reg syntax for 64-bit values
> - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-8-4ed30e865892@imgtec.com
> ---
>  .../devicetree/bindings/gpu/img,powervr-rogue.yaml | 41 ++++++++++++++++++++++
>  1 file changed, 41 insertions(+)
> 
> diff --git a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> index 9dc55a6d0d4023983a3fc480340351f3fa974ce5..b620baa56a4caa41246f7b53064d0e3309bdda8e 100644
> --- a/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> +++ b/Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml
> @@ -18,6 +18,11 @@ properties:
>                - ti,am62-gpu
>            - const: img,img-axe-1-16m
>            - const: img,img-rogue
> +      - items:
> +          - enum:
> +              - ti,j721s2-gpu
> +          - const: img,img-bxs-4-64
> +          - const: img,img-rogue
>  
>        # This legacy combination of compatible strings was introduced early on before the more
>        # specific GPU identifiers were used. Keep it around here for compatibility, but never use
> @@ -78,6 +83,18 @@ allOf:
>        properties:
>          power-domains:
>            maxItems: 1
> +  # Cores with two power domains
> +  - if:
> +      properties:
> +        compatible:
> +          contains:
> +            anyOf:

Nope

> +              - const: img,img-bxs-4-64
> +    then:
> +      properties:
> +        power-domains:
> +          minItems: 2
> +          maxItems: 2

You need to constrain power-domain-names. Same for other variants.


>    # Vendor integrations using a single clock domain
>    - if:
>        properties:
> @@ -85,6 +102,7 @@ allOf:
>            contains:
>              anyOf:
>                - const: ti,am62-gpu
> +              - const: ti,j721s2-gpu
>      then:
>        properties:
>          clocks:
> @@ -105,3 +123,26 @@ examples:
>          power-domains = <&k3_pds 187 TI_SCI_PD_EXCLUSIVE>;
>          power-domain-names = "a";
>      };
> +  - |
> +    #include <dt-bindings/interrupt-controller/irq.h>
> +    #include <dt-bindings/interrupt-controller/arm-gic.h>
> +    #include <dt-bindings/soc/ti,sci_pm_domain.h>

Drop. No difference in this example.

Best regards,
Krzysztof

[PATCH v2 09/21] drm/imagination: Revert to non-threaded IRQs

[Matt Coster]

For some reason, threaded IRQs do not play nice with the RISC-V firmware
processor in BXS on our test platform (TI AM68).

Until we can resolve this issue, revert to a more traditional workqueue-
based IRQ implementation so the platform is at least functional.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-9-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_ccb.c    | 25 +++++++--
 drivers/gpu/drm/imagination/pvr_device.c | 94 ++++++++++----------------------
 drivers/gpu/drm/imagination/pvr_device.h | 12 ++++
 drivers/gpu/drm/imagination/pvr_fw.c     |  1 +
 drivers/gpu/drm/imagination/pvr_power.c  |  4 +-
 drivers/gpu/drm/imagination/pvr_queue.c  | 23 +++++++-
 6 files changed, 87 insertions(+), 72 deletions(-)

diff --git a/drivers/gpu/drm/imagination/pvr_ccb.c b/drivers/gpu/drm/imagination/pvr_ccb.c
index 2bbdc05a3b97791426fd1f2ffe8f306d15622901..3bc3dc1bcc30ace720a09a2af5cadc1c6ccfd93f 100644
--- a/drivers/gpu/drm/imagination/pvr_ccb.c
+++ b/drivers/gpu/drm/imagination/pvr_ccb.c
@@ -156,6 +156,13 @@ process_fwccb_command(struct pvr_device *pvr_dev, struct rogue_fwif_fwccb_cmd *c
 	}
 }
 
+static void pvr_fwccb_process_worker(struct work_struct *work)
+{
+	struct pvr_device *pvr_dev = container_of_const(work, struct pvr_device, fwccb_work);
+
+	pvr_fwccb_process(pvr_dev);
+}
+
 /**
  * pvr_fwccb_process() - Process any pending FWCCB commands
  * @pvr_dev: Target PowerVR device
@@ -463,19 +470,16 @@ struct pvr_kccb_fence {
 };
 
 /**
- * pvr_kccb_wake_up_waiters() - Check the KCCB waiters
+ * pvr_kccb_check_waiters() - Check the KCCB waiters
  * @pvr_dev: Target PowerVR device
  *
  * Signal as many KCCB fences as we have slots available.
  */
-void pvr_kccb_wake_up_waiters(struct pvr_device *pvr_dev)
+static void pvr_kccb_check_waiters(struct pvr_device *pvr_dev)
 {
 	struct pvr_kccb_fence *fence, *tmp_fence;
 	u32 used_count, available_count;
 
-	/* Wake up those waiting for KCCB slot execution. */
-	wake_up_all(&pvr_dev->kccb.rtn_q);
-
 	/* Then iterate over all KCCB fences and signal as many as we can. */
 	mutex_lock(&pvr_dev->kccb.ccb.lock);
 	used_count = pvr_kccb_used_slot_count_locked(pvr_dev);
@@ -499,12 +503,20 @@ void pvr_kccb_wake_up_waiters(struct pvr_device *pvr_dev)
 	mutex_unlock(&pvr_dev->kccb.ccb.lock);
 }
 
+static void pvr_kccb_process_worker(struct work_struct *work)
+{
+	struct pvr_device *pvr_dev = container_of_const(work, struct pvr_device, kccb.work);
+
+	pvr_kccb_check_waiters(pvr_dev);
+}
+
 /**
  * pvr_kccb_fini() - Cleanup device KCCB
  * @pvr_dev: Target PowerVR device
  */
 void pvr_kccb_fini(struct pvr_device *pvr_dev)
 {
+	cancel_work_sync(&pvr_dev->kccb.work);
 	pvr_ccb_fini(&pvr_dev->kccb.ccb);
 	WARN_ON(!list_empty(&pvr_dev->kccb.waiters));
 	WARN_ON(pvr_dev->kccb.reserved_count);
@@ -525,6 +537,7 @@ pvr_kccb_init(struct pvr_device *pvr_dev)
 	INIT_LIST_HEAD(&pvr_dev->kccb.waiters);
 	pvr_dev->kccb.fence_ctx.id = dma_fence_context_alloc(1);
 	spin_lock_init(&pvr_dev->kccb.fence_ctx.lock);
+	INIT_WORK(&pvr_dev->kccb.work, pvr_kccb_process_worker);
 
 	return pvr_ccb_init(pvr_dev, &pvr_dev->kccb.ccb,
 			    ROGUE_FWIF_KCCB_NUMCMDS_LOG2_DEFAULT,
@@ -639,6 +652,8 @@ void pvr_kccb_release_slot(struct pvr_device *pvr_dev)
 int
 pvr_fwccb_init(struct pvr_device *pvr_dev)
 {
+	INIT_WORK(&pvr_dev->fwccb_work, pvr_fwccb_process_worker);
+
 	return pvr_ccb_init(pvr_dev, &pvr_dev->fwccb,
 			    ROGUE_FWIF_FWCCB_NUMCMDS_LOG2,
 			    sizeof(struct rogue_fwif_fwccb_cmd));
diff --git a/drivers/gpu/drm/imagination/pvr_device.c b/drivers/gpu/drm/imagination/pvr_device.c
index 1704c0268589bdeb65fa6535f9ec63182b0a3e94..43411fe64fcecd8f84c0ceabb329f2901d63ed93 100644
--- a/drivers/gpu/drm/imagination/pvr_device.c
+++ b/drivers/gpu/drm/imagination/pvr_device.c
@@ -120,72 +120,24 @@ static int pvr_device_clk_init(struct pvr_device *pvr_dev)
 	return 0;
 }
 
-/**
- * pvr_device_process_active_queues() - Process all queue related events.
- * @pvr_dev: PowerVR device to check
- *
- * This is called any time we receive a FW event. It iterates over all
- * active queues and calls pvr_queue_process() on them.
- */
-static void pvr_device_process_active_queues(struct pvr_device *pvr_dev)
-{
-	struct pvr_queue *queue, *tmp_queue;
-	LIST_HEAD(active_queues);
-
-	mutex_lock(&pvr_dev->queues.lock);
-
-	/* Move all active queues to a temporary list. Queues that remain
-	 * active after we're done processing them are re-inserted to
-	 * the queues.active list by pvr_queue_process().
-	 */
-	list_splice_init(&pvr_dev->queues.active, &active_queues);
-
-	list_for_each_entry_safe(queue, tmp_queue, &active_queues, node)
-		pvr_queue_process(queue);
-
-	mutex_unlock(&pvr_dev->queues.lock);
-}
-
-static irqreturn_t pvr_device_irq_thread_handler(int irq, void *data)
+static irqreturn_t pvr_device_irq_handler(int irq, void *data)
 {
 	struct pvr_device *pvr_dev = data;
-	irqreturn_t ret = IRQ_NONE;
-
-	/* We are in the threaded handler, we can keep dequeuing events until we
-	 * don't see any. This should allow us to reduce the number of interrupts
-	 * when the GPU is receiving a massive amount of short jobs.
-	 */
-	while (pvr_fw_irq_pending(pvr_dev)) {
-		pvr_fw_irq_clear(pvr_dev);
 
-		if (pvr_dev->fw_dev.booted) {
-			pvr_fwccb_process(pvr_dev);
-			pvr_kccb_wake_up_waiters(pvr_dev);
-			pvr_device_process_active_queues(pvr_dev);
-		}
+	if (!pvr_fw_irq_pending(pvr_dev))
+		return IRQ_NONE; /* Spurious IRQ - ignore */
 
-		pm_runtime_mark_last_busy(from_pvr_device(pvr_dev)->dev);
+	pvr_fw_irq_clear(pvr_dev);
 
-		ret = IRQ_HANDLED;
+	/* Only process IRQ work if FW is currently running */
+	if (pvr_dev->fw_dev.booted) {
+		queue_work(pvr_dev->irq_wq, &pvr_dev->fwccb_work);
+		wake_up_all(&pvr_dev->kccb.rtn_q);
+		queue_work(pvr_dev->irq_wq, &pvr_dev->kccb.work);
+		queue_work(pvr_dev->irq_wq, &pvr_dev->queues.work);
 	}
 
-	/* Unmask FW irqs before returning, so new interrupts can be received. */
-	pvr_fw_irq_enable(pvr_dev);
-	return ret;
-}
-
-static irqreturn_t pvr_device_irq_handler(int irq, void *data)
-{
-	struct pvr_device *pvr_dev = data;
-
-	if (!pvr_fw_irq_pending(pvr_dev))
-		return IRQ_NONE; /* Spurious IRQ - ignore. */
-
-	/* Mask the FW interrupts before waking up the thread. Will be unmasked
-	 * when the thread handler is done processing events.
-	 */
-	pvr_fw_irq_disable(pvr_dev);
-	return IRQ_WAKE_THREAD;
+	return IRQ_HANDLED;
 }
 
 /**
@@ -202,20 +154,33 @@ pvr_device_irq_init(struct pvr_device *pvr_dev)
 {
 	struct drm_device *drm_dev = from_pvr_device(pvr_dev);
 	struct platform_device *plat_dev = to_platform_device(drm_dev->dev);
+	int err;
 
 	init_waitqueue_head(&pvr_dev->kccb.rtn_q);
 
+	pvr_dev->irq_wq = alloc_workqueue("powervr-irq", WQ_UNBOUND, 0);
+	if (!pvr_dev->irq_wq) {
+		err = -ENOMEM;
+		goto err_out;
+	}
+
 	pvr_dev->irq = platform_get_irq(plat_dev, 0);
-	if (pvr_dev->irq < 0)
-		return pvr_dev->irq;
+	if (pvr_dev->irq < 0) {
+		err = pvr_dev->irq;
+		goto err_destroy_wq;
+	}
 
 	/* Clear any pending events before requesting the IRQ line. */
 	pvr_fw_irq_clear(pvr_dev);
 	pvr_fw_irq_enable(pvr_dev);
 
-	return request_threaded_irq(pvr_dev->irq, pvr_device_irq_handler,
-				    pvr_device_irq_thread_handler,
-				    IRQF_SHARED, "gpu", pvr_dev);
+	return request_irq(pvr_dev->irq, pvr_device_irq_handler, 0, "gpu", pvr_dev);
+
+err_destroy_wq:
+	destroy_workqueue(pvr_dev->irq_wq);
+
+err_out:
+	return err;
 }
 
 /**
@@ -226,6 +191,7 @@ static void
 pvr_device_irq_fini(struct pvr_device *pvr_dev)
 {
 	free_irq(pvr_dev->irq, pvr_dev);
+	destroy_workqueue(pvr_dev->irq_wq);
 }
 
 /**
diff --git a/drivers/gpu/drm/imagination/pvr_device.h b/drivers/gpu/drm/imagination/pvr_device.h
index 470945ccfcac7ce91161aa6c70d33177fbb3533f..930129334807168cf11ee843f446eb5063464d55 100644
--- a/drivers/gpu/drm/imagination/pvr_device.h
+++ b/drivers/gpu/drm/imagination/pvr_device.h
@@ -141,9 +141,15 @@ struct pvr_device {
 	/** @irq: IRQ number. */
 	int irq;
 
+	/** @irq_wq: Workqueue for actions triggered off the IRQ handler. */
+	struct workqueue_struct *irq_wq;
+
 	/** @fwccb: Firmware CCB. */
 	struct pvr_ccb fwccb;
 
+	/** @fwccb_work: Work item for FWCCB processing. */
+	struct work_struct fwccb_work;
+
 	/**
 	 * @kernel_vm_ctx: Virtual memory context used for kernel mappings.
 	 *
@@ -210,6 +216,9 @@ struct pvr_device {
 		/** @queues.lock: Lock protecting access to the active/idle
 		 *  lists. */
 		struct mutex lock;
+
+		/** @queues.work: Work item for queue processing. */
+		struct work_struct work;
 	} queues;
 
 	/**
@@ -258,6 +267,9 @@ struct pvr_device {
 		 *  future use. */
 		u32 reserved_count;
 
+		/** @kccb.work: Work item for KCCB processing. */
+		struct work_struct work;
+
 		/**
 		 * @kccb.waiters: List of KCCB slot waiters.
 		 */
diff --git a/drivers/gpu/drm/imagination/pvr_fw.c b/drivers/gpu/drm/imagination/pvr_fw.c
index 3debc9870a82ae7de9b2dc173df84c466c137bb3..9c8929d8602ead3390aa5c1b2505845b961b1406 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.c
+++ b/drivers/gpu/drm/imagination/pvr_fw.c
@@ -1036,6 +1036,7 @@ pvr_fw_fini(struct pvr_device *pvr_dev)
 	 * Ensure FWCCB worker has finished executing before destroying FWCCB. The IRQ handler has
 	 * been unregistered at this point so no new work should be being submitted.
 	 */
+	flush_work(&pvr_dev->fwccb_work);
 	pvr_ccb_fini(&pvr_dev->fwccb);
 	pvr_kccb_fini(pvr_dev);
 	pvr_fw_cleanup(pvr_dev);
diff --git a/drivers/gpu/drm/imagination/pvr_power.c b/drivers/gpu/drm/imagination/pvr_power.c
index 19b079b357df78e8bcdecfa377fc9c05b6e8e4b0..000ded17f7ca4c5cdadc5aba4f4d76fac0bbcc0c 100644
--- a/drivers/gpu/drm/imagination/pvr_power.c
+++ b/drivers/gpu/drm/imagination/pvr_power.c
@@ -122,7 +122,7 @@ pvr_power_fw_enable(struct pvr_device *pvr_dev)
 		return err;
 	}
 
-	queue_delayed_work(pvr_dev->sched_wq, &pvr_dev->watchdog.work,
+	queue_delayed_work(pvr_dev->irq_wq, &pvr_dev->watchdog.work,
 			   msecs_to_jiffies(WATCHDOG_TIME_MS));
 
 	return 0;
@@ -212,7 +212,7 @@ pvr_watchdog_worker(struct work_struct *work)
 
 out_requeue:
 	if (!pvr_dev->lost) {
-		queue_delayed_work(pvr_dev->sched_wq, &pvr_dev->watchdog.work,
+		queue_delayed_work(pvr_dev->irq_wq, &pvr_dev->watchdog.work,
 				   msecs_to_jiffies(WATCHDOG_TIME_MS));
 	}
 }
diff --git a/drivers/gpu/drm/imagination/pvr_queue.c b/drivers/gpu/drm/imagination/pvr_queue.c
index c4f08432882b12f5cdfeb7fc991fd941f0946676..5f6a880987d9f35804da60e1258efa30beed56ab 100644
--- a/drivers/gpu/drm/imagination/pvr_queue.c
+++ b/drivers/gpu/drm/imagination/pvr_queue.c
@@ -838,7 +838,7 @@ pvr_queue_timedout_job(struct drm_sched_job *s_job)
 	} else {
 		atomic_set(&queue->in_flight_job_count, job_count);
 		list_move_tail(&queue->node, &pvr_dev->queues.active);
-		pvr_queue_process(queue);
+		queue_work(pvr_dev->irq_wq, &pvr_dev->queues.work);
 	}
 	mutex_unlock(&pvr_dev->queues.lock);
 
@@ -991,6 +991,26 @@ void pvr_queue_process(struct pvr_queue *queue)
 	pvr_queue_update_active_state_locked(queue);
 }
 
+static void pvr_queue_process_worker(struct work_struct *work)
+{
+	struct pvr_device *pvr_dev = container_of_const(work, struct pvr_device, queues.work);
+	struct pvr_queue *queue, *tmp_queue;
+	LIST_HEAD(active_queues);
+
+	mutex_lock(&pvr_dev->queues.lock);
+
+	list_splice_init(&pvr_dev->queues.active, &active_queues);
+
+	list_for_each_entry_safe(queue, tmp_queue, &active_queues, node) {
+		pvr_queue_check_job_waiting_for_cccb_space(queue);
+		pvr_queue_signal_done_fences(queue);
+
+		pvr_queue_update_active_state_locked(queue);
+	}
+
+	mutex_unlock(&pvr_dev->queues.lock);
+}
+
 static u32 get_dm_type(struct pvr_queue *queue)
 {
 	switch (queue->type) {
@@ -1407,6 +1427,7 @@ int pvr_queue_device_init(struct pvr_device *pvr_dev)
 {
 	int err;
 
+	INIT_WORK(&pvr_dev->queues.work, pvr_queue_process_worker);
 	INIT_LIST_HEAD(&pvr_dev->queues.active);
 	INIT_LIST_HEAD(&pvr_dev->queues.idle);
 	err = drmm_mutex_init(from_pvr_device(pvr_dev), &pvr_dev->queues.lock);

-- 
2.47.0

[PATCH v2 10/21] drm/imagination: Remove firmware enable_reg

[Matt Coster]

After the previous commit ("drm/imagination: Revert to non-threaded IRQs"),
this register is now only used to enable firmware interrupts at
start-of-day. This is, however, unnecessary since they are enabled by
default.

In addition, the soon-to-be-added RISC-V firmware processors do not have
an equivalent register.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-10-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_device.c  |  1 -
 drivers/gpu/drm/imagination/pvr_fw.h      | 11 +----------
 drivers/gpu/drm/imagination/pvr_fw_meta.c |  1 -
 drivers/gpu/drm/imagination/pvr_fw_mips.c |  1 -
 4 files changed, 1 insertion(+), 13 deletions(-)

diff --git a/drivers/gpu/drm/imagination/pvr_device.c b/drivers/gpu/drm/imagination/pvr_device.c
index 43411fe64fcecd8f84c0ceabb329f2901d63ed93..52d7641a1a0c62a9c4029092e846472d82950a61 100644
--- a/drivers/gpu/drm/imagination/pvr_device.c
+++ b/drivers/gpu/drm/imagination/pvr_device.c
@@ -172,7 +172,6 @@ pvr_device_irq_init(struct pvr_device *pvr_dev)
 
 	/* Clear any pending events before requesting the IRQ line. */
 	pvr_fw_irq_clear(pvr_dev);
-	pvr_fw_irq_enable(pvr_dev);
 
 	return request_irq(pvr_dev->irq, pvr_device_irq_handler, 0, "gpu", pvr_dev);
 
diff --git a/drivers/gpu/drm/imagination/pvr_fw.h b/drivers/gpu/drm/imagination/pvr_fw.h
index b7966bd574a924862b7877c175fa2b5d757d89db..29bae4bc244a243a6a95bcf838d924060cc043e2 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.h
+++ b/drivers/gpu/drm/imagination/pvr_fw.h
@@ -188,9 +188,6 @@ struct pvr_fw_defs {
 	 * processor backend in pvr_fw_funcs::init().
 	 */
 	struct {
-		/** @enable_reg: FW interrupt enable register. */
-		u32 enable_reg;
-
 		/** @status_reg: FW interrupt status register. */
 		u32 status_reg;
 
@@ -202,7 +199,7 @@ struct pvr_fw_defs {
 		 */
 		u32 clear_reg;
 
-		/** @event_mask: Bitmask of events to listen for. */
+		/** @event_mask: Bitmask of events to listen for in the status_reg. */
 		u32 event_mask;
 
 		/** @clear_mask: Value to write to the clear_reg in order to clear FW IRQs. */
@@ -412,12 +409,6 @@ struct pvr_fw_device {
 #define pvr_fw_irq_clear(pvr_dev) \
 	pvr_fw_irq_write_reg(pvr_dev, clear, (pvr_dev)->fw_dev.defs->irq.clear_mask)
 
-#define pvr_fw_irq_enable(pvr_dev) \
-	pvr_fw_irq_write_reg(pvr_dev, enable, (pvr_dev)->fw_dev.defs->irq.event_mask)
-
-#define pvr_fw_irq_disable(pvr_dev) \
-	pvr_fw_irq_write_reg(pvr_dev, enable, 0)
-
 extern const struct pvr_fw_defs pvr_fw_defs_meta;
 extern const struct pvr_fw_defs pvr_fw_defs_mips;
 
diff --git a/drivers/gpu/drm/imagination/pvr_fw_meta.c b/drivers/gpu/drm/imagination/pvr_fw_meta.c
index c39beb70c3173ebdab13b4e810ce5d9a3419f0ba..76b24ad9aa221b6a384dc7b55ed2e78d2e761550 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_meta.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_meta.c
@@ -546,7 +546,6 @@ const struct pvr_fw_defs pvr_fw_defs_meta = {
 	.wrapper_init = pvr_meta_wrapper_init,
 	.has_fixed_data_addr = pvr_meta_has_fixed_data_addr,
 	.irq = {
-		.enable_reg = ROGUE_CR_META_SP_MSLVIRQENABLE,
 		.status_reg = ROGUE_CR_META_SP_MSLVIRQSTATUS,
 		.clear_reg = ROGUE_CR_META_SP_MSLVIRQSTATUS,
 		.event_mask = ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_EN,
diff --git a/drivers/gpu/drm/imagination/pvr_fw_mips.c b/drivers/gpu/drm/imagination/pvr_fw_mips.c
index 0bed0257e2ab75f66d8b8966b2ceac6342396fb5..c810a67eeecf1016064e76baf534e31a44c859b5 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_mips.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_mips.c
@@ -243,7 +243,6 @@ const struct pvr_fw_defs pvr_fw_defs_mips = {
 	.wrapper_init = pvr_mips_wrapper_init,
 	.has_fixed_data_addr = pvr_mips_has_fixed_data_addr,
 	.irq = {
-		.enable_reg = ROGUE_CR_MIPS_WRAPPER_IRQ_ENABLE,
 		.status_reg = ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS,
 		.clear_reg = ROGUE_CR_MIPS_WRAPPER_IRQ_CLEAR,
 		.event_mask = ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS_EVENT_EN,

-- 
2.47.0

[PATCH v2 11/21] drm/imagination: Rename event_mask -> status_mask

[Matt Coster]

Now that enable_reg isn't used, rename the previously shared event_mask to
status_mask since it's only used with status_reg.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-11-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_fw.h      | 6 +++---
 drivers/gpu/drm/imagination/pvr_fw_meta.c | 2 +-
 drivers/gpu/drm/imagination/pvr_fw_mips.c | 2 +-
 3 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/imagination/pvr_fw.h b/drivers/gpu/drm/imagination/pvr_fw.h
index 29bae4bc244a243a6a95bcf838d924060cc043e2..eead744835726712622d5aba9b3480fe264a089f 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.h
+++ b/drivers/gpu/drm/imagination/pvr_fw.h
@@ -199,8 +199,8 @@ struct pvr_fw_defs {
 		 */
 		u32 clear_reg;
 
-		/** @event_mask: Bitmask of events to listen for in the status_reg. */
-		u32 event_mask;
+		/** @status_mask: Bitmask of events to listen for in the status_reg. */
+		u32 status_mask;
 
 		/** @clear_mask: Value to write to the clear_reg in order to clear FW IRQs. */
 		u32 clear_mask;
@@ -404,7 +404,7 @@ struct pvr_fw_device {
 	pvr_cr_write32((pvr_dev), (pvr_dev)->fw_dev.defs->irq.name ## _reg, value)
 
 #define pvr_fw_irq_pending(pvr_dev) \
-	(pvr_fw_irq_read_reg(pvr_dev, status) & (pvr_dev)->fw_dev.defs->irq.event_mask)
+	(pvr_fw_irq_read_reg(pvr_dev, status) & (pvr_dev)->fw_dev.defs->irq.status_mask)
 
 #define pvr_fw_irq_clear(pvr_dev) \
 	pvr_fw_irq_write_reg(pvr_dev, clear, (pvr_dev)->fw_dev.defs->irq.clear_mask)
diff --git a/drivers/gpu/drm/imagination/pvr_fw_meta.c b/drivers/gpu/drm/imagination/pvr_fw_meta.c
index 76b24ad9aa221b6a384dc7b55ed2e78d2e761550..cf86701ca8f14920329ccb4c2811424b0c394b14 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_meta.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_meta.c
@@ -548,7 +548,7 @@ const struct pvr_fw_defs pvr_fw_defs_meta = {
 	.irq = {
 		.status_reg = ROGUE_CR_META_SP_MSLVIRQSTATUS,
 		.clear_reg = ROGUE_CR_META_SP_MSLVIRQSTATUS,
-		.event_mask = ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_EN,
+		.status_mask = ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_EN,
 		.clear_mask = ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_CLRMSK,
 	},
 };
diff --git a/drivers/gpu/drm/imagination/pvr_fw_mips.c b/drivers/gpu/drm/imagination/pvr_fw_mips.c
index c810a67eeecf1016064e76baf534e31a44c859b5..f195c602bb112066e88210d0106cb5ffc0a9abc6 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_mips.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_mips.c
@@ -245,7 +245,7 @@ const struct pvr_fw_defs pvr_fw_defs_mips = {
 	.irq = {
 		.status_reg = ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS,
 		.clear_reg = ROGUE_CR_MIPS_WRAPPER_IRQ_CLEAR,
-		.event_mask = ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS_EVENT_EN,
+		.status_mask = ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS_EVENT_EN,
 		.clear_mask = ROGUE_CR_MIPS_WRAPPER_IRQ_CLEAR_EVENT_EN,
 	},
 };

-- 
2.47.0

[PATCH v2 12/21] drm/imagination: Make has_fixed_data_addr a value

[Matt Coster]

This is currently a callback function which takes no parameters; there's
no reason for this so let's make it a straightforward value in pvr_fw_defs.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-12-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_fw.c      |  2 +-
 drivers/gpu/drm/imagination/pvr_fw.h      | 23 ++++++++---------------
 drivers/gpu/drm/imagination/pvr_fw_meta.c |  8 +-------
 drivers/gpu/drm/imagination/pvr_fw_mips.c |  8 +-------
 4 files changed, 11 insertions(+), 30 deletions(-)

diff --git a/drivers/gpu/drm/imagination/pvr_fw.c b/drivers/gpu/drm/imagination/pvr_fw.c
index 9c8929d8602ead3390aa5c1b2505845b961b1406..808844eb10b5ccb29ed2b8e9bdfe3be829cc57d1 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.c
+++ b/drivers/gpu/drm/imagination/pvr_fw.c
@@ -663,7 +663,7 @@ pvr_fw_process(struct pvr_device *pvr_dev)
 		return PTR_ERR(fw_code_ptr);
 	}
 
-	if (pvr_dev->fw_dev.defs->has_fixed_data_addr()) {
+	if (pvr_dev->fw_dev.defs->has_fixed_data_addr) {
 		u32 base_addr = private_data->base_addr & pvr_dev->fw_dev.fw_heap_info.offset_mask;
 
 		fw_data_ptr =
diff --git a/drivers/gpu/drm/imagination/pvr_fw.h b/drivers/gpu/drm/imagination/pvr_fw.h
index eead744835726712622d5aba9b3480fe264a089f..180d310074e3585c641e540a9e2576b5ab2a5705 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.h
+++ b/drivers/gpu/drm/imagination/pvr_fw.h
@@ -166,21 +166,6 @@ struct pvr_fw_defs {
 	 */
 	int (*wrapper_init)(struct pvr_device *pvr_dev);
 
-	/**
-	 * @has_fixed_data_addr:
-	 *
-	 * Called to check if firmware fixed data must be loaded at the address given by the
-	 * firmware layout table.
-	 *
-	 * This function is mandatory.
-	 *
-	 * Returns:
-	 *  * %true if firmware fixed data must be loaded at the address given by the firmware
-	 *    layout table.
-	 *  * %false otherwise.
-	 */
-	bool (*has_fixed_data_addr)(void);
-
 	/**
 	 * @irq: FW Interrupt information.
 	 *
@@ -205,6 +190,14 @@ struct pvr_fw_defs {
 		/** @clear_mask: Value to write to the clear_reg in order to clear FW IRQs. */
 		u32 clear_mask;
 	} irq;
+
+	/**
+	 * @has_fixed_data_addr: Specify whether the firmware fixed data must be loaded at the
+	 *                       address given by the firmware layout table.
+	 *
+	 * This value is mandatory.
+	 */
+	bool has_fixed_data_addr;
 };
 
 /**
diff --git a/drivers/gpu/drm/imagination/pvr_fw_meta.c b/drivers/gpu/drm/imagination/pvr_fw_meta.c
index cf86701ca8f14920329ccb4c2811424b0c394b14..4433b04e0adb3684b86a4e90f63d670a81ecd826 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_meta.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_meta.c
@@ -531,12 +531,6 @@ pvr_meta_vm_unmap(struct pvr_device *pvr_dev, struct pvr_fw_object *fw_obj)
 		     fw_obj->fw_mm_node.size);
 }
 
-static bool
-pvr_meta_has_fixed_data_addr(void)
-{
-	return false;
-}
-
 const struct pvr_fw_defs pvr_fw_defs_meta = {
 	.init = pvr_meta_init,
 	.fw_process = pvr_meta_fw_process,
@@ -544,11 +538,11 @@ const struct pvr_fw_defs pvr_fw_defs_meta = {
 	.vm_unmap = pvr_meta_vm_unmap,
 	.get_fw_addr_with_offset = pvr_meta_get_fw_addr_with_offset,
 	.wrapper_init = pvr_meta_wrapper_init,
-	.has_fixed_data_addr = pvr_meta_has_fixed_data_addr,
 	.irq = {
 		.status_reg = ROGUE_CR_META_SP_MSLVIRQSTATUS,
 		.clear_reg = ROGUE_CR_META_SP_MSLVIRQSTATUS,
 		.status_mask = ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_EN,
 		.clear_mask = ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_CLRMSK,
 	},
+	.has_fixed_data_addr = false,
 };
diff --git a/drivers/gpu/drm/imagination/pvr_fw_mips.c b/drivers/gpu/drm/imagination/pvr_fw_mips.c
index f195c602bb112066e88210d0106cb5ffc0a9abc6..2c3172841886b70eb7a9992ec3851f18adcad8d5 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_mips.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_mips.c
@@ -227,12 +227,6 @@ pvr_mips_get_fw_addr_with_offset(struct pvr_fw_object *fw_obj, u32 offset)
 	       ROGUE_FW_HEAP_MIPS_BASE;
 }
 
-static bool
-pvr_mips_has_fixed_data_addr(void)
-{
-	return true;
-}
-
 const struct pvr_fw_defs pvr_fw_defs_mips = {
 	.init = pvr_mips_init,
 	.fini = pvr_mips_fini,
@@ -241,11 +235,11 @@ const struct pvr_fw_defs pvr_fw_defs_mips = {
 	.vm_unmap = pvr_vm_mips_unmap,
 	.get_fw_addr_with_offset = pvr_mips_get_fw_addr_with_offset,
 	.wrapper_init = pvr_mips_wrapper_init,
-	.has_fixed_data_addr = pvr_mips_has_fixed_data_addr,
 	.irq = {
 		.status_reg = ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS,
 		.clear_reg = ROGUE_CR_MIPS_WRAPPER_IRQ_CLEAR,
 		.status_mask = ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS_EVENT_EN,
 		.clear_mask = ROGUE_CR_MIPS_WRAPPER_IRQ_CLEAR_EVENT_EN,
 	},
+	.has_fixed_data_addr = true,
 };

-- 
2.47.0

[PATCH v2 13/21] drm/imagination: Use a lookup table for fw defs

[Matt Coster]

With more than two firmware processor types, the if/else chain in
pvr_fw_init() gets a bit ridiculous. Use a static array indexed on
pvr_fw_processor_type (which is now a proper enum instead of #defines)
instead.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-13-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_device.h |  4 ----
 drivers/gpu/drm/imagination/pvr_fw.c     | 21 ++++++++++++++++-----
 drivers/gpu/drm/imagination/pvr_fw.h     |  7 +++++++
 3 files changed, 23 insertions(+), 9 deletions(-)

diff --git a/drivers/gpu/drm/imagination/pvr_device.h b/drivers/gpu/drm/imagination/pvr_device.h
index 930129334807168cf11ee843f446eb5063464d55..ec5eb78af82a3f3c32d6c89b68b7bc0fcee0b9d2 100644
--- a/drivers/gpu/drm/imagination/pvr_device.h
+++ b/drivers/gpu/drm/imagination/pvr_device.h
@@ -738,8 +738,4 @@ pvr_ioctl_union_padding_check(void *instance, size_t union_offset,
 					      __union_size, __member_size);  \
 	})
 
-#define PVR_FW_PROCESSOR_TYPE_META  0
-#define PVR_FW_PROCESSOR_TYPE_MIPS  1
-#define PVR_FW_PROCESSOR_TYPE_RISCV 2
-
 #endif /* PVR_DEVICE_H */
diff --git a/drivers/gpu/drm/imagination/pvr_fw.c b/drivers/gpu/drm/imagination/pvr_fw.c
index 808844eb10b5ccb29ed2b8e9bdfe3be829cc57d1..b77c4a42caf2f208f6289160e5cec36c683b1e65 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.c
+++ b/drivers/gpu/drm/imagination/pvr_fw.c
@@ -926,16 +926,27 @@ pvr_fw_validate_init_device_info(struct pvr_device *pvr_dev)
 int
 pvr_fw_init(struct pvr_device *pvr_dev)
 {
+	static const struct pvr_fw_defs *fw_defs[PVR_FW_PROCESSOR_TYPE_COUNT] = {
+		[PVR_FW_PROCESSOR_TYPE_META] = &pvr_fw_defs_meta,
+		[PVR_FW_PROCESSOR_TYPE_MIPS] = &pvr_fw_defs_mips,
+		[PVR_FW_PROCESSOR_TYPE_RISCV] = NULL,
+	};
+
 	u32 kccb_size_log2 = ROGUE_FWIF_KCCB_NUMCMDS_LOG2_DEFAULT;
 	u32 kccb_rtn_size = (1 << kccb_size_log2) * sizeof(*pvr_dev->kccb.rtn);
 	struct pvr_fw_device *fw_dev = &pvr_dev->fw_dev;
 	int err;
 
-	if (fw_dev->processor_type == PVR_FW_PROCESSOR_TYPE_META)
-		fw_dev->defs = &pvr_fw_defs_meta;
-	else if (fw_dev->processor_type == PVR_FW_PROCESSOR_TYPE_MIPS)
-		fw_dev->defs = &pvr_fw_defs_mips;
-	else
+	if (fw_dev->processor_type >= PVR_FW_PROCESSOR_TYPE_COUNT)
+		return -EINVAL;
+
+	fw_dev->defs = fw_defs[fw_dev->processor_type];
+
+	/*
+	 * Not all firmware processor types are currently supported.
+	 * Once they are, this check can be removed.
+	 */
+	if (!fw_dev->defs)
 		return -EINVAL;
 
 	err = fw_dev->defs->init(pvr_dev);
diff --git a/drivers/gpu/drm/imagination/pvr_fw.h b/drivers/gpu/drm/imagination/pvr_fw.h
index 180d310074e3585c641e540a9e2576b5ab2a5705..88ad713468ce3a1ee459b04dde5363c24791a4f1 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.h
+++ b/drivers/gpu/drm/imagination/pvr_fw.h
@@ -402,6 +402,13 @@ struct pvr_fw_device {
 #define pvr_fw_irq_clear(pvr_dev) \
 	pvr_fw_irq_write_reg(pvr_dev, clear, (pvr_dev)->fw_dev.defs->irq.clear_mask)
 
+enum pvr_fw_processor_type {
+	PVR_FW_PROCESSOR_TYPE_META = 0,
+	PVR_FW_PROCESSOR_TYPE_MIPS,
+	PVR_FW_PROCESSOR_TYPE_RISCV,
+	PVR_FW_PROCESSOR_TYPE_COUNT,
+};
+
 extern const struct pvr_fw_defs pvr_fw_defs_meta;
 extern const struct pvr_fw_defs pvr_fw_defs_mips;
 

-- 
2.47.0

[PATCH v2 14/21] drm/imagination: Use callbacks for fw irq handling

[Matt Coster]

This allows for more versatility in checking and clearing firmware
registers used for interrupt handling.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-14-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_device.h  | 18 +++++++++++++
 drivers/gpu/drm/imagination/pvr_fw.h      | 45 +++++++++----------------------
 drivers/gpu/drm/imagination/pvr_fw_meta.c | 22 ++++++++++-----
 drivers/gpu/drm/imagination/pvr_fw_mips.c | 22 ++++++++++-----
 4 files changed, 63 insertions(+), 44 deletions(-)

diff --git a/drivers/gpu/drm/imagination/pvr_device.h b/drivers/gpu/drm/imagination/pvr_device.h
index ec5eb78af82a3f3c32d6c89b68b7bc0fcee0b9d2..76f79b18af354f0e0070530dfc5c8fe0f6a41ce1 100644
--- a/drivers/gpu/drm/imagination/pvr_device.h
+++ b/drivers/gpu/drm/imagination/pvr_device.h
@@ -738,4 +738,22 @@ pvr_ioctl_union_padding_check(void *instance, size_t union_offset,
 					      __union_size, __member_size);  \
 	})
 
+/*
+ * These utility functions should more properly be placed in pvr_fw.h, but that
+ * would cause a dependency cycle between that header and this one. Since
+ * they're primarily used in pvr_device.c, let's put them in here for now.
+ */
+
+static __always_inline bool
+pvr_fw_irq_pending(struct pvr_device *pvr_dev)
+{
+	return pvr_dev->fw_dev.defs->irq_pending(pvr_dev);
+}
+
+static __always_inline void
+pvr_fw_irq_clear(struct pvr_device *pvr_dev)
+{
+	pvr_dev->fw_dev.defs->irq_clear(pvr_dev);
+}
+
 #endif /* PVR_DEVICE_H */
diff --git a/drivers/gpu/drm/imagination/pvr_fw.h b/drivers/gpu/drm/imagination/pvr_fw.h
index 88ad713468ce3a1ee459b04dde5363c24791a4f1..ab69f40a7fbc6304171f16dd16d825a68b0362a5 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.h
+++ b/drivers/gpu/drm/imagination/pvr_fw.h
@@ -167,29 +167,22 @@ struct pvr_fw_defs {
 	int (*wrapper_init)(struct pvr_device *pvr_dev);
 
 	/**
-	 * @irq: FW Interrupt information.
+	 * @irq_pending: Check interrupt status register for pending interrupts.
 	 *
-	 * Those are processor dependent, and should be initialized by the
-	 * processor backend in pvr_fw_funcs::init().
+	 * @pvr_dev: Target PowerVR device.
+	 *
+	 * This function is mandatory.
 	 */
-	struct {
-		/** @status_reg: FW interrupt status register. */
-		u32 status_reg;
+	bool (*irq_pending)(struct pvr_device *pvr_dev);
 
-		/**
-		 * @clear_reg: FW interrupt clear register.
-		 *
-		 * If @status_reg == @clear_reg, we clear by write a bit to zero,
-		 * otherwise we clear by writing a bit to one.
-		 */
-		u32 clear_reg;
-
-		/** @status_mask: Bitmask of events to listen for in the status_reg. */
-		u32 status_mask;
-
-		/** @clear_mask: Value to write to the clear_reg in order to clear FW IRQs. */
-		u32 clear_mask;
-	} irq;
+	/**
+	 * @irq_clear: Clear pending interrupts.
+	 *
+	 * @pvr_dev: Target PowerVR device.
+	 *
+	 * This function is mandatory.
+	 */
+	void (*irq_clear)(struct pvr_device *pvr_dev);
 
 	/**
 	 * @has_fixed_data_addr: Specify whether the firmware fixed data must be loaded at the
@@ -390,18 +383,6 @@ struct pvr_fw_device {
 	} fw_objs;
 };
 
-#define pvr_fw_irq_read_reg(pvr_dev, name) \
-	pvr_cr_read32((pvr_dev), (pvr_dev)->fw_dev.defs->irq.name ## _reg)
-
-#define pvr_fw_irq_write_reg(pvr_dev, name, value) \
-	pvr_cr_write32((pvr_dev), (pvr_dev)->fw_dev.defs->irq.name ## _reg, value)
-
-#define pvr_fw_irq_pending(pvr_dev) \
-	(pvr_fw_irq_read_reg(pvr_dev, status) & (pvr_dev)->fw_dev.defs->irq.status_mask)
-
-#define pvr_fw_irq_clear(pvr_dev) \
-	pvr_fw_irq_write_reg(pvr_dev, clear, (pvr_dev)->fw_dev.defs->irq.clear_mask)
-
 enum pvr_fw_processor_type {
 	PVR_FW_PROCESSOR_TYPE_META = 0,
 	PVR_FW_PROCESSOR_TYPE_MIPS,
diff --git a/drivers/gpu/drm/imagination/pvr_fw_meta.c b/drivers/gpu/drm/imagination/pvr_fw_meta.c
index 4433b04e0adb3684b86a4e90f63d670a81ecd826..09de3a30b625013c190196e02074fe72d08629a6 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_meta.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_meta.c
@@ -531,6 +531,20 @@ pvr_meta_vm_unmap(struct pvr_device *pvr_dev, struct pvr_fw_object *fw_obj)
 		     fw_obj->fw_mm_node.size);
 }
 
+static bool
+pvr_meta_irq_pending(struct pvr_device *pvr_dev)
+{
+	return pvr_cr_read32(pvr_dev, ROGUE_CR_META_SP_MSLVIRQSTATUS) &
+	       ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_EN;
+}
+
+static void
+pvr_meta_irq_clear(struct pvr_device *pvr_dev)
+{
+	pvr_cr_write32(pvr_dev, ROGUE_CR_META_SP_MSLVIRQSTATUS,
+		       ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_CLRMSK);
+}
+
 const struct pvr_fw_defs pvr_fw_defs_meta = {
 	.init = pvr_meta_init,
 	.fw_process = pvr_meta_fw_process,
@@ -538,11 +552,7 @@ const struct pvr_fw_defs pvr_fw_defs_meta = {
 	.vm_unmap = pvr_meta_vm_unmap,
 	.get_fw_addr_with_offset = pvr_meta_get_fw_addr_with_offset,
 	.wrapper_init = pvr_meta_wrapper_init,
-	.irq = {
-		.status_reg = ROGUE_CR_META_SP_MSLVIRQSTATUS,
-		.clear_reg = ROGUE_CR_META_SP_MSLVIRQSTATUS,
-		.status_mask = ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_EN,
-		.clear_mask = ROGUE_CR_META_SP_MSLVIRQSTATUS_TRIGVECT2_CLRMSK,
-	},
+	.irq_pending = pvr_meta_irq_pending,
+	.irq_clear = pvr_meta_irq_clear,
 	.has_fixed_data_addr = false,
 };
diff --git a/drivers/gpu/drm/imagination/pvr_fw_mips.c b/drivers/gpu/drm/imagination/pvr_fw_mips.c
index 2c3172841886b70eb7a9992ec3851f18adcad8d5..524a9bd0a20b64c509f5708cc61d93b4c864b835 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_mips.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_mips.c
@@ -227,6 +227,20 @@ pvr_mips_get_fw_addr_with_offset(struct pvr_fw_object *fw_obj, u32 offset)
 	       ROGUE_FW_HEAP_MIPS_BASE;
 }
 
+static bool
+pvr_mips_irq_pending(struct pvr_device *pvr_dev)
+{
+	return pvr_cr_read32(pvr_dev, ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS) &
+	       ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS_EVENT_EN;
+}
+
+static void
+pvr_mips_irq_clear(struct pvr_device *pvr_dev)
+{
+	pvr_cr_write32(pvr_dev, ROGUE_CR_MIPS_WRAPPER_IRQ_CLEAR,
+		       ROGUE_CR_MIPS_WRAPPER_IRQ_CLEAR_EVENT_EN);
+}
+
 const struct pvr_fw_defs pvr_fw_defs_mips = {
 	.init = pvr_mips_init,
 	.fini = pvr_mips_fini,
@@ -235,11 +249,7 @@ const struct pvr_fw_defs pvr_fw_defs_mips = {
 	.vm_unmap = pvr_vm_mips_unmap,
 	.get_fw_addr_with_offset = pvr_mips_get_fw_addr_with_offset,
 	.wrapper_init = pvr_mips_wrapper_init,
-	.irq = {
-		.status_reg = ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS,
-		.clear_reg = ROGUE_CR_MIPS_WRAPPER_IRQ_CLEAR,
-		.status_mask = ROGUE_CR_MIPS_WRAPPER_IRQ_STATUS_EVENT_EN,
-		.clear_mask = ROGUE_CR_MIPS_WRAPPER_IRQ_CLEAR_EVENT_EN,
-	},
+	.irq_pending = pvr_mips_irq_pending,
+	.irq_clear = pvr_mips_irq_clear,
 	.has_fixed_data_addr = true,
 };

-- 
2.47.0

[PATCH v2 15/21] drm/imagination: Add register required for RISC-V firmware

[Matt Coster]

Add definitions for a register required for a subsequent patch adding
support for RISC-V firmware. ROGUE_CR_FWCORE_DMI_DMCONTROL is used to
control the debug module in the firmware processor.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-15-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_rogue_cr_defs.h | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/imagination/pvr_rogue_cr_defs.h b/drivers/gpu/drm/imagination/pvr_rogue_cr_defs.h
index 2a90d02796d3e071b18e18dead105e29798bcddc..4e99832a667a2d88ee21469595d3abdad3a07c06 100644
--- a/drivers/gpu/drm/imagination/pvr_rogue_cr_defs.h
+++ b/drivers/gpu/drm/imagination/pvr_rogue_cr_defs.h
@@ -3751,7 +3751,22 @@
 
 /* Register ROGUE_CR_FWCORE_DMI_DMCONTROL */
 #define ROGUE_CR_FWCORE_DMI_DMCONTROL 0x3480U
-#define ROGUE_CR_FWCORE_DMI_DMCONTROL_MASKFULL 0x0000000000000000ULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_MASKFULL 0x00000000D0000003ULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_HALTREQ_SHIFT 31U
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_HALTREQ_CLRMSK 0x7FFFFFFFULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_HALTREQ_EN 0x80000000ULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_RESUMEREQ_SHIFT 30U
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_RESUMEREQ_CLRMSK 0xBFFFFFFFULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_RESUMEREQ_EN 0x40000000ULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_ACKHAVERESET_SHIFT 28U
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_ACKHAVERESET_CLRMSK 0xEFFFFFFFULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_ACKHAVERESET_EN 0x10000000ULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_NDMRESET_SHIFT 1U
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_NDMRESET_CLRMSK 0xFFFFFFFDULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_NDMRESET_EN 0x00000002ULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_DMACTIVE_SHIFT 0U
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_DMACTIVE_CLRMSK 0xFFFFFFFEULL
+#define ROGUE_CR_FWCORE_DMI_DMCONTROL_DMACTIVE_EN 0x00000001ULL
 
 /* Register ROGUE_CR_FWCORE_DMI_DMSTATUS */
 #define ROGUE_CR_FWCORE_DMI_DMSTATUS 0x3488U

-- 
2.47.0

[PATCH v2 16/21] drm/imagination: Move ELF fw utils to common file

[Matt Coster]

Currently only MIPS firmware processors use ELF-formatted firmware. When
adding support for RISC-V firmware processors, it will be useful to have
ELF handling functions ready to go.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-16-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/Makefile      |  1 +
 drivers/gpu/drm/imagination/pvr_fw.h      |  5 +++
 drivers/gpu/drm/imagination/pvr_fw_mips.c | 59 +--------------------------
 drivers/gpu/drm/imagination/pvr_fw_util.c | 67 +++++++++++++++++++++++++++++++
 4 files changed, 75 insertions(+), 57 deletions(-)

diff --git a/drivers/gpu/drm/imagination/Makefile b/drivers/gpu/drm/imagination/Makefile
index 9bc6a3884c2239e44265f3cdebee149841b270de..077e4762c7c383b6e339da1584c3865d830ef8d6 100644
--- a/drivers/gpu/drm/imagination/Makefile
+++ b/drivers/gpu/drm/imagination/Makefile
@@ -16,6 +16,7 @@ powervr-y := \
 	pvr_fw_mips.o \
 	pvr_fw_startstop.o \
 	pvr_fw_trace.o \
+	pvr_fw_util.o \
 	pvr_gem.o \
 	pvr_hwrt.o \
 	pvr_job.o \
diff --git a/drivers/gpu/drm/imagination/pvr_fw.h b/drivers/gpu/drm/imagination/pvr_fw.h
index ab69f40a7fbc6304171f16dd16d825a68b0362a5..8d3f0ee0aac69373f15ac5919d5b0f27c67cb284 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.h
+++ b/drivers/gpu/drm/imagination/pvr_fw.h
@@ -478,4 +478,9 @@ pvr_fw_object_get_fw_addr(struct pvr_fw_object *fw_obj, u32 *fw_addr_out)
 	pvr_fw_object_get_fw_addr_offset(fw_obj, 0, fw_addr_out);
 }
 
+/* Util functions defined in pvr_util.c. These are intended for use in pvr_fw_<arch>.c files. */
+int
+pvr_fw_process_elf_command_stream(struct pvr_device *pvr_dev, const u8 *fw, u8 *fw_code_ptr,
+				  u8 *fw_data_ptr, u8 *fw_core_code_ptr, u8 *fw_core_data_ptr);
+
 #endif /* PVR_FW_H */
diff --git a/drivers/gpu/drm/imagination/pvr_fw_mips.c b/drivers/gpu/drm/imagination/pvr_fw_mips.c
index 524a9bd0a20b64c509f5708cc61d93b4c864b835..7f341ceb0661c61ac059654faeec91e149036467 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_mips.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_mips.c
@@ -8,7 +8,6 @@
 #include "pvr_rogue_mips.h"
 #include "pvr_vm_mips.h"
 
-#include <linux/elf.h>
 #include <linux/err.h>
 #include <linux/types.h>
 
@@ -16,60 +15,6 @@
 #define ROGUE_FW_HEAP_MIPS_SHIFT 24 /* 16 MB */
 #define ROGUE_FW_HEAP_MIPS_RESERVED_SIZE SZ_1M
 
-/**
- * process_elf_command_stream() - Process ELF firmware image and populate
- *                                firmware sections
- * @pvr_dev: Device pointer.
- * @fw: Pointer to firmware image.
- * @fw_code_ptr: Pointer to FW code section.
- * @fw_data_ptr: Pointer to FW data section.
- * @fw_core_code_ptr: Pointer to FW coremem code section.
- * @fw_core_data_ptr: Pointer to FW coremem data section.
- *
- * Returns :
- *  * 0 on success, or
- *  * -EINVAL on any error in ELF command stream.
- */
-static int
-process_elf_command_stream(struct pvr_device *pvr_dev, const u8 *fw, u8 *fw_code_ptr,
-			   u8 *fw_data_ptr, u8 *fw_core_code_ptr, u8 *fw_core_data_ptr)
-{
-	struct elf32_hdr *header = (struct elf32_hdr *)fw;
-	struct elf32_phdr *program_header = (struct elf32_phdr *)(fw + header->e_phoff);
-	struct drm_device *drm_dev = from_pvr_device(pvr_dev);
-	u32 entry;
-	int err;
-
-	for (entry = 0; entry < header->e_phnum; entry++, program_header++) {
-		void *write_addr;
-
-		/* Only consider loadable entries in the ELF segment table */
-		if (program_header->p_type != PT_LOAD)
-			continue;
-
-		err = pvr_fw_find_mmu_segment(pvr_dev, program_header->p_vaddr,
-					      program_header->p_memsz, fw_code_ptr, fw_data_ptr,
-					      fw_core_code_ptr, fw_core_data_ptr, &write_addr);
-		if (err) {
-			drm_err(drm_dev,
-				"Addr 0x%x (size: %d) not found in any firmware segment",
-				program_header->p_vaddr, program_header->p_memsz);
-			return err;
-		}
-
-		/* Write to FW allocation only if available */
-		if (write_addr) {
-			memcpy(write_addr, fw + program_header->p_offset,
-			       program_header->p_filesz);
-
-			memset((u8 *)write_addr + program_header->p_filesz, 0,
-			       program_header->p_memsz - program_header->p_filesz);
-		}
-	}
-
-	return 0;
-}
-
 static int
 pvr_mips_init(struct pvr_device *pvr_dev)
 {
@@ -100,8 +45,8 @@ pvr_mips_fw_process(struct pvr_device *pvr_dev, const u8 *fw,
 	u32 page_nr;
 	int err;
 
-	err = process_elf_command_stream(pvr_dev, fw, fw_code_ptr, fw_data_ptr, fw_core_code_ptr,
-					 fw_core_data_ptr);
+	err = pvr_fw_process_elf_command_stream(pvr_dev, fw, fw_code_ptr, fw_data_ptr,
+						fw_core_code_ptr, fw_core_data_ptr);
 	if (err)
 		return err;
 
diff --git a/drivers/gpu/drm/imagination/pvr_fw_util.c b/drivers/gpu/drm/imagination/pvr_fw_util.c
new file mode 100644
index 0000000000000000000000000000000000000000..7bc8a5c48e9b0eed2045607ab2cfed80a60a32b5
--- /dev/null
+++ b/drivers/gpu/drm/imagination/pvr_fw_util.c
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: GPL-2.0-only OR MIT
+/* Copyright (c) 2024 Imagination Technologies Ltd. */
+
+#include "pvr_device.h"
+#include "pvr_fw.h"
+
+#include <drm/drm_device.h>
+#include <drm/drm_print.h>
+
+#include <linux/elf.h>
+#include <linux/string.h>
+#include <linux/types.h>
+
+/**
+ * pvr_fw_process_elf_command_stream() - Process ELF firmware image and populate
+ *                                       firmware sections
+ * @pvr_dev: Device pointer.
+ * @fw: Pointer to firmware image.
+ * @fw_code_ptr: Pointer to FW code section.
+ * @fw_data_ptr: Pointer to FW data section.
+ * @fw_core_code_ptr: Pointer to FW coremem code section.
+ * @fw_core_data_ptr: Pointer to FW coremem data section.
+ *
+ * Returns :
+ *  * 0 on success, or
+ *  * -EINVAL on any error in ELF command stream.
+ */
+int
+pvr_fw_process_elf_command_stream(struct pvr_device *pvr_dev, const u8 *fw,
+				  u8 *fw_code_ptr, u8 *fw_data_ptr,
+				  u8 *fw_core_code_ptr, u8 *fw_core_data_ptr)
+{
+	struct elf32_hdr *header = (struct elf32_hdr *)fw;
+	struct elf32_phdr *program_header = (struct elf32_phdr *)(fw + header->e_phoff);
+	struct drm_device *drm_dev = from_pvr_device(pvr_dev);
+	u32 entry;
+	int err;
+
+	for (entry = 0; entry < header->e_phnum; entry++, program_header++) {
+		void *write_addr;
+
+		/* Only consider loadable entries in the ELF segment table */
+		if (program_header->p_type != PT_LOAD)
+			continue;
+
+		err = pvr_fw_find_mmu_segment(pvr_dev, program_header->p_vaddr,
+					      program_header->p_memsz, fw_code_ptr, fw_data_ptr,
+					      fw_core_code_ptr, fw_core_data_ptr, &write_addr);
+		if (err) {
+			drm_err(drm_dev,
+				"Addr 0x%x (size: %d) not found in any firmware segment",
+				program_header->p_vaddr, program_header->p_memsz);
+			return err;
+		}
+
+		/* Write to FW allocation only if available */
+		if (write_addr) {
+			memcpy(write_addr, fw + program_header->p_offset,
+			       program_header->p_filesz);
+
+			memset((u8 *)write_addr + program_header->p_filesz, 0,
+			       program_header->p_memsz - program_header->p_filesz);
+		}
+	}
+
+	return 0;
+}

-- 
2.47.0

[PATCH v2 17/21] drm/imagination: Add RISC-V firmware processor support

[Matt Coster]

From: Sarah Walker <sarah.walker@imgtec.com>

Newer PowerVR GPUs (such as the BXS-4-64 MC1) use a RISC-V firmware
processor instead of the previous MIPS or META.

Signed-off-by: Sarah Walker <sarah.walker@imgtec.com>
Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-17-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/Makefile           |   1 +
 drivers/gpu/drm/imagination/pvr_fw.c           |  18 +--
 drivers/gpu/drm/imagination/pvr_fw.h           |  10 ++
 drivers/gpu/drm/imagination/pvr_fw_riscv.c     | 163 +++++++++++++++++++++++++
 drivers/gpu/drm/imagination/pvr_fw_startstop.c |  21 ++++
 drivers/gpu/drm/imagination/pvr_rogue_riscv.h  |  41 +++++++
 6 files changed, 246 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/imagination/Makefile b/drivers/gpu/drm/imagination/Makefile
index 077e4762c7c383b6e339da1584c3865d830ef8d6..d748ad9d62e092ed7c3d772214ccd327818d507f 100644
--- a/drivers/gpu/drm/imagination/Makefile
+++ b/drivers/gpu/drm/imagination/Makefile
@@ -14,6 +14,7 @@ powervr-y := \
 	pvr_fw.o \
 	pvr_fw_meta.o \
 	pvr_fw_mips.o \
+	pvr_fw_riscv.o \
 	pvr_fw_startstop.o \
 	pvr_fw_trace.o \
 	pvr_fw_util.o \
diff --git a/drivers/gpu/drm/imagination/pvr_fw.c b/drivers/gpu/drm/imagination/pvr_fw.c
index b77c4a42caf2f208f6289160e5cec36c683b1e65..7373220fa4fc05f5beb239f1ceb92b3caba97765 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.c
+++ b/drivers/gpu/drm/imagination/pvr_fw.c
@@ -929,7 +929,7 @@ pvr_fw_init(struct pvr_device *pvr_dev)
 	static const struct pvr_fw_defs *fw_defs[PVR_FW_PROCESSOR_TYPE_COUNT] = {
 		[PVR_FW_PROCESSOR_TYPE_META] = &pvr_fw_defs_meta,
 		[PVR_FW_PROCESSOR_TYPE_MIPS] = &pvr_fw_defs_mips,
-		[PVR_FW_PROCESSOR_TYPE_RISCV] = NULL,
+		[PVR_FW_PROCESSOR_TYPE_RISCV] = &pvr_fw_defs_riscv,
 	};
 
 	u32 kccb_size_log2 = ROGUE_FWIF_KCCB_NUMCMDS_LOG2_DEFAULT;
@@ -942,13 +942,6 @@ pvr_fw_init(struct pvr_device *pvr_dev)
 
 	fw_dev->defs = fw_defs[fw_dev->processor_type];
 
-	/*
-	 * Not all firmware processor types are currently supported.
-	 * Once they are, this check can be removed.
-	 */
-	if (!fw_dev->defs)
-		return -EINVAL;
-
 	err = fw_dev->defs->init(pvr_dev);
 	if (err)
 		return err;
@@ -1455,6 +1448,15 @@ void pvr_fw_object_get_fw_addr_offset(struct pvr_fw_object *fw_obj, u32 offset,
 	*fw_addr_out = pvr_dev->fw_dev.defs->get_fw_addr_with_offset(fw_obj, offset);
 }
 
+u64
+pvr_fw_obj_get_gpu_addr(struct pvr_fw_object *fw_obj)
+{
+	struct pvr_device *pvr_dev = to_pvr_device(gem_from_pvr_gem(fw_obj->gem)->dev);
+	struct pvr_fw_device *fw_dev = &pvr_dev->fw_dev;
+
+	return fw_dev->fw_heap_info.gpu_addr + fw_obj->fw_addr_offset;
+}
+
 /*
  * pvr_fw_hard_reset() - Re-initialise the FW code and data segments, and reset all global FW
  *                       structures
diff --git a/drivers/gpu/drm/imagination/pvr_fw.h b/drivers/gpu/drm/imagination/pvr_fw.h
index 8d3f0ee0aac69373f15ac5919d5b0f27c67cb284..24b9852dbc18d6cec94a4efff2eb0925fb4b5374 100644
--- a/drivers/gpu/drm/imagination/pvr_fw.h
+++ b/drivers/gpu/drm/imagination/pvr_fw.h
@@ -392,6 +392,7 @@ enum pvr_fw_processor_type {
 
 extern const struct pvr_fw_defs pvr_fw_defs_meta;
 extern const struct pvr_fw_defs pvr_fw_defs_mips;
+extern const struct pvr_fw_defs pvr_fw_defs_riscv;
 
 int pvr_fw_validate_init_device_info(struct pvr_device *pvr_dev);
 int pvr_fw_init(struct pvr_device *pvr_dev);
@@ -478,6 +479,15 @@ pvr_fw_object_get_fw_addr(struct pvr_fw_object *fw_obj, u32 *fw_addr_out)
 	pvr_fw_object_get_fw_addr_offset(fw_obj, 0, fw_addr_out);
 }
 
+u64
+pvr_fw_obj_get_gpu_addr(struct pvr_fw_object *fw_obj);
+
+static __always_inline size_t
+pvr_fw_obj_get_object_size(struct pvr_fw_object *fw_obj)
+{
+	return pvr_gem_object_size(fw_obj->gem);
+}
+
 /* Util functions defined in pvr_util.c. These are intended for use in pvr_fw_<arch>.c files. */
 int
 pvr_fw_process_elf_command_stream(struct pvr_device *pvr_dev, const u8 *fw, u8 *fw_code_ptr,
diff --git a/drivers/gpu/drm/imagination/pvr_fw_riscv.c b/drivers/gpu/drm/imagination/pvr_fw_riscv.c
new file mode 100644
index 0000000000000000000000000000000000000000..14021e6ef7533f7185dc3c2378fc1d0c46c2ffed
--- /dev/null
+++ b/drivers/gpu/drm/imagination/pvr_fw_riscv.c
@@ -0,0 +1,163 @@
+// SPDX-License-Identifier: GPL-2.0 OR MIT
+/* Copyright (c) 2024 Imagination Technologies Ltd. */
+
+#include "pvr_device.h"
+#include "pvr_fw.h"
+#include "pvr_fw_info.h"
+#include "pvr_fw_mips.h"
+#include "pvr_gem.h"
+#include "pvr_rogue_cr_defs.h"
+#include "pvr_rogue_riscv.h"
+#include "pvr_vm.h"
+
+#include <linux/compiler.h>
+#include <linux/delay.h>
+#include <linux/firmware.h>
+#include <linux/ktime.h>
+#include <linux/types.h>
+
+#define ROGUE_FW_HEAP_RISCV_SHIFT 25 /* 32 MB */
+#define ROGUE_FW_HEAP_RISCV_SIZE (1u << ROGUE_FW_HEAP_RISCV_SHIFT)
+
+static int
+pvr_riscv_wrapper_init(struct pvr_device *pvr_dev)
+{
+	const u64 common_opts =
+		((u64)(ROGUE_FW_HEAP_RISCV_SIZE >> FWCORE_ADDR_REMAP_CONFIG0_SIZE_ALIGNSHIFT)
+		 << ROGUE_CR_FWCORE_ADDR_REMAP_CONFIG0_SIZE_SHIFT) |
+		((u64)MMU_CONTEXT_MAPPING_FWPRIV
+		 << FWCORE_ADDR_REMAP_CONFIG0_MMU_CONTEXT_SHIFT);
+
+	u64 code_addr = pvr_fw_obj_get_gpu_addr(pvr_dev->fw_dev.mem.code_obj);
+	u64 data_addr = pvr_fw_obj_get_gpu_addr(pvr_dev->fw_dev.mem.data_obj);
+
+	/* This condition allows us to OR the addresses into the register directly. */
+	static_assert(ROGUE_CR_FWCORE_ADDR_REMAP_CONFIG1_DEVVADDR_SHIFT ==
+		      ROGUE_CR_FWCORE_ADDR_REMAP_CONFIG1_DEVVADDR_ALIGNSHIFT);
+
+	WARN_ON(code_addr & ROGUE_CR_FWCORE_ADDR_REMAP_CONFIG1_DEVVADDR_CLRMSK);
+	WARN_ON(data_addr & ROGUE_CR_FWCORE_ADDR_REMAP_CONFIG1_DEVVADDR_CLRMSK);
+
+	pvr_cr_write64(pvr_dev, ROGUE_RISCVFW_REGION_REMAP_CR(BOOTLDR_CODE),
+		       code_addr | common_opts | ROGUE_CR_FWCORE_ADDR_REMAP_CONFIG0_FETCH_EN_EN);
+
+	pvr_cr_write64(pvr_dev, ROGUE_RISCVFW_REGION_REMAP_CR(BOOTLDR_DATA),
+		       data_addr | common_opts |
+			       ROGUE_CR_FWCORE_ADDR_REMAP_CONFIG0_LOAD_STORE_EN_EN);
+
+	/* Garten IDLE bit controlled by RISC-V. */
+	pvr_cr_write64(pvr_dev, ROGUE_CR_MTS_GARTEN_WRAPPER_CONFIG,
+		       ROGUE_CR_MTS_GARTEN_WRAPPER_CONFIG_IDLE_CTRL_META);
+
+	return 0;
+}
+
+struct rogue_riscv_fw_boot_data {
+	u64 coremem_code_dev_vaddr;
+	u64 coremem_data_dev_vaddr;
+	u32 coremem_code_fw_addr;
+	u32 coremem_data_fw_addr;
+	u32 coremem_code_size;
+	u32 coremem_data_size;
+	u32 flags;
+	u32 reserved;
+};
+
+static int
+pvr_riscv_fw_process(struct pvr_device *pvr_dev, const u8 *fw,
+		    u8 *fw_code_ptr, u8 *fw_data_ptr, u8 *fw_core_code_ptr, u8 *fw_core_data_ptr,
+		    u32 core_code_alloc_size)
+{
+	struct pvr_fw_device *fw_dev = &pvr_dev->fw_dev;
+	struct pvr_fw_mem *fw_mem = &fw_dev->mem;
+	struct rogue_riscv_fw_boot_data *boot_data;
+	int err;
+
+	err = pvr_fw_process_elf_command_stream(pvr_dev, fw, fw_code_ptr, fw_data_ptr,
+						fw_core_code_ptr, fw_core_data_ptr);
+	if (err)
+		goto err_out;
+
+	boot_data = (struct rogue_riscv_fw_boot_data *)fw_data_ptr;
+
+	if (fw_mem->core_code_obj) {
+		boot_data->coremem_code_dev_vaddr = pvr_fw_obj_get_gpu_addr(fw_mem->core_code_obj);
+		pvr_fw_object_get_fw_addr(fw_mem->core_code_obj, &boot_data->coremem_code_fw_addr);
+		boot_data->coremem_code_size = pvr_fw_obj_get_object_size(fw_mem->core_code_obj);
+	}
+
+	if (fw_mem->core_data_obj) {
+		boot_data->coremem_data_dev_vaddr = pvr_fw_obj_get_gpu_addr(fw_mem->core_data_obj);
+		pvr_fw_object_get_fw_addr(fw_mem->core_data_obj, &boot_data->coremem_data_fw_addr);
+		boot_data->coremem_data_size = pvr_fw_obj_get_object_size(fw_mem->core_data_obj);
+	}
+
+	return 0;
+
+err_out:
+	return err;
+}
+
+static int
+pvr_riscv_init(struct pvr_device *pvr_dev)
+{
+	pvr_fw_heap_info_init(pvr_dev, ROGUE_FW_HEAP_RISCV_SHIFT, 0);
+
+	return 0;
+}
+
+static u32
+pvr_riscv_get_fw_addr_with_offset(struct pvr_fw_object *fw_obj, u32 offset)
+{
+	u32 fw_addr = fw_obj->fw_addr_offset + offset;
+
+	/* RISC-V cacheability is determined by address. */
+	if (fw_obj->gem->flags & PVR_BO_FW_FLAGS_DEVICE_UNCACHED)
+		fw_addr |= ROGUE_RISCVFW_REGION_BASE(SHARED_UNCACHED_DATA);
+	else
+		fw_addr |= ROGUE_RISCVFW_REGION_BASE(SHARED_CACHED_DATA);
+
+	return fw_addr;
+}
+
+static int
+pvr_riscv_vm_map(struct pvr_device *pvr_dev, struct pvr_fw_object *fw_obj)
+{
+	struct pvr_gem_object *pvr_obj = fw_obj->gem;
+
+	return pvr_vm_map(pvr_dev->kernel_vm_ctx, pvr_obj, 0, fw_obj->fw_mm_node.start,
+			  pvr_gem_object_size(pvr_obj));
+}
+
+static void
+pvr_riscv_vm_unmap(struct pvr_device *pvr_dev, struct pvr_fw_object *fw_obj)
+{
+	pvr_vm_unmap(pvr_dev->kernel_vm_ctx, fw_obj->fw_mm_node.start,
+		     fw_obj->fw_mm_node.size);
+}
+
+static bool
+pvr_riscv_irq_pending(struct pvr_device *pvr_dev)
+{
+	return pvr_cr_read32(pvr_dev, ROGUE_CR_IRQ_OS0_EVENT_STATUS) &
+	       ROGUE_CR_IRQ_OS0_EVENT_STATUS_SOURCE_EN;
+}
+
+static void
+pvr_riscv_irq_clear(struct pvr_device *pvr_dev)
+{
+	pvr_cr_write32(pvr_dev, ROGUE_CR_IRQ_OS0_EVENT_CLEAR,
+		       ROGUE_CR_IRQ_OS0_EVENT_CLEAR_SOURCE_EN);
+}
+
+const struct pvr_fw_defs pvr_fw_defs_riscv = {
+	.init = pvr_riscv_init,
+	.fw_process = pvr_riscv_fw_process,
+	.vm_map = pvr_riscv_vm_map,
+	.vm_unmap = pvr_riscv_vm_unmap,
+	.get_fw_addr_with_offset = pvr_riscv_get_fw_addr_with_offset,
+	.wrapper_init = pvr_riscv_wrapper_init,
+	.irq_pending = pvr_riscv_irq_pending,
+	.irq_clear = pvr_riscv_irq_clear,
+	.has_fixed_data_addr = false,
+};
diff --git a/drivers/gpu/drm/imagination/pvr_fw_startstop.c b/drivers/gpu/drm/imagination/pvr_fw_startstop.c
index 36cec227cfe3cf5b1e293f48b164bb1be1b0ea54..31edf7e49a9c3f3d64022129c8bb7e2151aa9cdf 100644
--- a/drivers/gpu/drm/imagination/pvr_fw_startstop.c
+++ b/drivers/gpu/drm/imagination/pvr_fw_startstop.c
@@ -49,6 +49,14 @@ rogue_bif_init(struct pvr_device *pvr_dev)
 
 	pvr_cr_write64(pvr_dev, BIF_CAT_BASEX(MMU_CONTEXT_MAPPING_FWPRIV),
 		       pc_addr);
+
+	if (pvr_dev->fw_dev.processor_type == PVR_FW_PROCESSOR_TYPE_RISCV) {
+		pc_addr = (((u64)pc_dma_addr >> ROGUE_CR_FWCORE_MEM_CAT_BASE0_ADDR_ALIGNSHIFT)
+			   << ROGUE_CR_FWCORE_MEM_CAT_BASE0_ADDR_SHIFT) &
+			  ~ROGUE_CR_FWCORE_MEM_CAT_BASE0_ADDR_CLRMSK;
+
+		pvr_cr_write64(pvr_dev, FWCORE_MEM_CAT_BASEX(MMU_CONTEXT_MAPPING_FWPRIV), pc_addr);
+	}
 }
 
 static int
@@ -114,6 +122,9 @@ pvr_fw_start(struct pvr_device *pvr_dev)
 		(void)pvr_cr_read32(pvr_dev, ROGUE_CR_SYS_BUS_SECURE); /* Fence write */
 	}
 
+	if (pvr_dev->fw_dev.processor_type == PVR_FW_PROCESSOR_TYPE_RISCV)
+		pvr_cr_write32(pvr_dev, ROGUE_CR_FWCORE_BOOT, 0);
+
 	/* Set Rogue in soft-reset. */
 	pvr_cr_write64(pvr_dev, ROGUE_CR_SOFT_RESET, soft_reset_mask);
 	if (has_reset2)
@@ -167,6 +178,16 @@ pvr_fw_start(struct pvr_device *pvr_dev)
 	/* ... and afterwards. */
 	udelay(3);
 
+	if (pvr_dev->fw_dev.processor_type == PVR_FW_PROCESSOR_TYPE_RISCV) {
+		/* Bring Debug Module out of reset. */
+		pvr_cr_write32(pvr_dev, ROGUE_CR_FWCORE_DMI_DMCONTROL,
+			       ROGUE_CR_FWCORE_DMI_DMCONTROL_DMACTIVE_EN);
+
+		/* Boot the FW. */
+		pvr_cr_write32(pvr_dev, ROGUE_CR_FWCORE_BOOT, 1);
+		udelay(3);
+	}
+
 	return 0;
 
 err_reset:
diff --git a/drivers/gpu/drm/imagination/pvr_rogue_riscv.h b/drivers/gpu/drm/imagination/pvr_rogue_riscv.h
new file mode 100644
index 0000000000000000000000000000000000000000..9a070e24fa6a8bb44ff1e421ae6750cbf724d346
--- /dev/null
+++ b/drivers/gpu/drm/imagination/pvr_rogue_riscv.h
@@ -0,0 +1,41 @@
+/* SPDX-License-Identifier: GPL-2.0 OR MIT */
+/* Copyright (c) 2024 Imagination Technologies Ltd. */
+
+#ifndef PVR_ROGUE_RISCV_H
+#define PVR_ROGUE_RISCV_H
+
+#include "pvr_rogue_cr_defs.h"
+
+#include <linux/bitops.h>
+#include <linux/sizes.h>
+#include <linux/types.h>
+
+#define ROGUE_RISCVFW_REGION_SIZE SZ_256M
+#define ROGUE_RISCVFW_REGION_SHIFT __ffs(ROGUE_RISCVFW_REGION_SIZE)
+
+enum rogue_riscvfw_region {
+	ROGUE_RISCV_REGION__RESERVED_0 = 0,
+	ROGUE_RISCV_REGION__RESERVED_1,
+	ROGUE_RISCV_REGION_SOCIF,
+	ROGUE_RISCV_REGION__RESERVED_3,
+	ROGUE_RISCV_REGION__RESERVED_4,
+	ROGUE_RISCV_REGION_BOOTLDR_DATA,
+	ROGUE_RISCV_REGION_SHARED_CACHED_DATA,
+	ROGUE_RISCV_REGION__RESERVED_7,
+	ROGUE_RISCV_REGION_COREMEM,
+	ROGUE_RISCV_REGION__RESERVED_9,
+	ROGUE_RISCV_REGION__RESERVED_A,
+	ROGUE_RISCV_REGION__RESERVED_B,
+	ROGUE_RISCV_REGION_BOOTLDR_CODE,
+	ROGUE_RISCV_REGION_SHARED_UNCACHED_DATA,
+	ROGUE_RISCV_REGION__RESERVED_E,
+	ROGUE_RISCV_REGION__RESERVED_F,
+
+	ROGUE_RISCV_REGION__COUNT,
+};
+
+#define ROGUE_RISCVFW_REGION_BASE(r) ((u32)(ROGUE_RISCV_REGION_##r) << ROGUE_RISCVFW_REGION_SHIFT)
+#define ROGUE_RISCVFW_REGION_REMAP_CR(r) \
+	(ROGUE_CR_FWCORE_ADDR_REMAP_CONFIG0 + (u32)(ROGUE_RISCV_REGION_##r) * 8U)
+
+#endif /* PVR_ROGUE_RISCV_H */

-- 
2.47.0

[PATCH v2 18/21] drm/imagination: Add platform overrides infrastructure

[Matt Coster]

This infrastructure will be used in cases where a specific GPU integration
or implementation requires some special handling in the driver. The first
use case is the device cached memory override added in the next patch.

The infrastructure is built out in this separate commit to make it
clear which specific changes refer to the workaround added there.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-18-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_device.c | 26 ++++++++++++++++++++++++
 drivers/gpu/drm/imagination/pvr_device.h | 26 ++++++++++++++++++++++++
 drivers/gpu/drm/imagination/pvr_drv.c    | 34 ++++++++++++++++++++++++--------
 3 files changed, 78 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/imagination/pvr_device.c b/drivers/gpu/drm/imagination/pvr_device.c
index 52d7641a1a0c62a9c4029092e846472d82950a61..2ce46b9a8ab7609faebeeb4e7820751b00047806 100644
--- a/drivers/gpu/drm/imagination/pvr_device.c
+++ b/drivers/gpu/drm/imagination/pvr_device.c
@@ -621,3 +621,29 @@ pvr_device_has_feature(struct pvr_device *pvr_dev, u32 feature)
 		return false;
 	}
 }
+
+/**
+ * pvr_device_overrides_validate() - Ensure the overrides specified are compatible with a device.
+ * @pvr_dev: Target PowerVR device.
+ * @overrides: Device overrides to validate.
+ *
+ * Return:
+ *  * %true if every override specified in @overrides is compatible with the current device, or
+ *  * %false otherwise; as many incompatibilities as possible will be reported in the kernel log.
+ */
+bool
+pvr_device_overrides_validate(struct pvr_device *pvr_dev,
+			      const struct pvr_device_overrides *overrides)
+{
+	bool ret = true;
+
+	/*
+	 * Where possible, avoid early returns in this function. This allows for as
+	 * many errors to be reported at once as possible.
+	 *
+	 * Note that this function may be called early during device initialization
+	 * so it should not be assumed that @pvr_dev is ready for normal use yet.
+	 */
+
+	return ret;
+}
diff --git a/drivers/gpu/drm/imagination/pvr_device.h b/drivers/gpu/drm/imagination/pvr_device.h
index 76f79b18af354f0e0070530dfc5c8fe0f6a41ce1..ad0a02a37154099542247dfc62f411c10f4e41f4 100644
--- a/drivers/gpu/drm/imagination/pvr_device.h
+++ b/drivers/gpu/drm/imagination/pvr_device.h
@@ -16,6 +16,7 @@
 #include <drm/drm_mm.h>
 
 #include <linux/bits.h>
+#include <linux/compiler.h>
 #include <linux/compiler_attributes.h>
 #include <linux/compiler_types.h>
 #include <linux/device.h>
@@ -56,6 +57,13 @@ struct pvr_fw_version {
 	u16 major, minor;
 };
 
+/**
+ * struct pvr_device_overrides - Hardware-level overrides loaded from
+ * MODULE_DEVICE_TABLE() or similar.
+ */
+struct pvr_device_overrides {
+};
+
 /**
  * struct pvr_device - powervr-specific wrapper for &struct drm_device
  */
@@ -94,6 +102,13 @@ struct pvr_device {
 	 */
 	struct pvr_device_enhancements enhancements;
 
+	/**
+	 * @overrides: Platform-specific overrides required for this device.
+	 *
+	 * Do not access this member directly, instead use PVR_HAS_OVERRIDE().
+	 */
+	struct pvr_device_overrides overrides;
+
 	/** @fw_version: Firmware version detected at runtime. */
 	struct pvr_fw_version fw_version;
 
@@ -436,6 +451,13 @@ struct pvr_file {
  */
 #define PVR_HAS_ENHANCEMENT(pvr_dev, enhancement) ((pvr_dev)->enhancements.has_ern##enhancement)
 
+/**
+ * PVR_HAS_OVERRIDE() - Tests whether a physical device requires a given override
+ * @pvr_dev: [IN] Target PowerVR device.
+ * @override: [IN] Override name.
+ */
+#define PVR_HAS_OVERRIDE(pvr_dev, override) unlikely((pvr_dev)->overrides.override)
+
 #define from_pvr_device(pvr_dev) (&(pvr_dev)->base)
 
 #define to_pvr_device(drm_dev) container_of_const(drm_dev, struct pvr_device, base)
@@ -516,6 +538,10 @@ pvr_device_has_uapi_enhancement(struct pvr_device *pvr_dev, u32 enhancement);
 bool
 pvr_device_has_feature(struct pvr_device *pvr_dev, u32 feature);
 
+bool
+pvr_device_overrides_validate(struct pvr_device *pvr_dev,
+			      const struct pvr_device_overrides *overrides);
+
 /**
  * PVR_CR_FIELD_GET() - Extract a single field from a PowerVR control register
  * @val: Value of the target register.
diff --git a/drivers/gpu/drm/imagination/pvr_drv.c b/drivers/gpu/drm/imagination/pvr_drv.c
index 1ab97933e14f20ee3fbf603c23b8dde2d33572c2..b56ee2cda9b54c4388a6eef38b0ff81acdb05874 100644
--- a/drivers/gpu/drm/imagination/pvr_drv.c
+++ b/drivers/gpu/drm/imagination/pvr_drv.c
@@ -31,6 +31,7 @@
 #include <linux/mod_devicetable.h>
 #include <linux/module.h>
 #include <linux/moduleparam.h>
+#include <linux/of.h>
 #include <linux/of_device.h>
 #include <linux/of_platform.h>
 #include <linux/platform_device.h>
@@ -1396,11 +1397,13 @@ static struct drm_driver pvr_drm_driver = {
 static int
 pvr_probe(struct platform_device *plat_dev)
 {
+	const struct pvr_device_overrides *overrides;
+	struct device *dev = &plat_dev->dev;
 	struct pvr_device *pvr_dev;
 	struct drm_device *drm_dev;
 	int err;
 
-	pvr_dev = devm_drm_dev_alloc(&plat_dev->dev, &pvr_drm_driver,
+	pvr_dev = devm_drm_dev_alloc(dev, &pvr_drm_driver,
 				     struct pvr_device, base);
 	if (IS_ERR(pvr_dev))
 		return PTR_ERR(pvr_dev);
@@ -1409,6 +1412,15 @@ pvr_probe(struct platform_device *plat_dev)
 
 	platform_set_drvdata(plat_dev, drm_dev);
 
+	overrides = of_device_get_match_data(dev);
+	if (!overrides)
+		return -EINVAL;
+
+	if (!pvr_device_overrides_validate(pvr_dev, overrides))
+		return -EINVAL;
+
+	pvr_dev->overrides = *overrides;
+
 	err = pvr_power_domains_init(pvr_dev);
 	if (err)
 		goto err_context_fini;
@@ -1421,11 +1433,11 @@ pvr_probe(struct platform_device *plat_dev)
 	if (err)
 		goto err_power_domains_fini;
 
-	devm_pm_runtime_enable(&plat_dev->dev);
-	pm_runtime_mark_last_busy(&plat_dev->dev);
+	devm_pm_runtime_enable(dev);
+	pm_runtime_mark_last_busy(dev);
 
-	pm_runtime_set_autosuspend_delay(&plat_dev->dev, 50);
-	pm_runtime_use_autosuspend(&plat_dev->dev);
+	pm_runtime_set_autosuspend_delay(dev, 50);
+	pm_runtime_use_autosuspend(dev);
 	pvr_watchdog_init(pvr_dev);
 
 	err = pvr_device_init(pvr_dev);
@@ -1478,18 +1490,24 @@ static void pvr_remove(struct platform_device *plat_dev)
 	pvr_power_domains_fini(pvr_dev);
 }
 
+static const struct pvr_device_overrides pvr_device_overrides_default = {};
+
+/*
+ * Always specify &pvr_device_overrides_default instead of %NULL for &struct of_device_id->data so
+ * that we know of_device_get_match_data() returning %NULL is an error.
+ */
 static const struct of_device_id dt_match[] = {
-	{ .compatible = "img,img-rogue", .data = NULL },
+	{ .compatible = "img,img-rogue", .data = &pvr_device_overrides_default },
 
 	/* All supported GPU models */
-	{ .compatible = "img,img-axe-1-16m", .data = NULL },
+	{ .compatible = "img,img-axe-1-16m", .data = &pvr_device_overrides_default },
 
 	/*
 	 * This legacy compatible string was introduced early on before the more specific GPU
 	 * identifiers were used. Keep it around here for compatibility, but never use
 	 * "img,img-axe" in new devicetrees.
 	 */
-	{ .compatible = "img,img-axe", .data = NULL },
+	{ .compatible = "img,img-axe", .data = &pvr_device_overrides_default },
 	{}
 };
 MODULE_DEVICE_TABLE(of, dt_match);

-- 
2.47.0

[PATCH v2 19/21] drm/imagination: Add device_memory_force_cpu_cached override

[Matt Coster]

The TI k3-j721s2 platform has a bug relating to cache snooping on the AXI
ACE-Lite interface. Disabling cache snooping altogether would also resolve
the issue, but is considered more of a performance hit.

Given the platform is dma-coherent, forcing all device-accessible memory
allocations through the CPU cache is the preferred solution.

Implement this workaround so that it can later be enabled for the TI
k3-j721s2 platform.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-19-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_device.c | 11 ++++++++++-
 drivers/gpu/drm/imagination/pvr_device.h | 11 +++++++++++
 drivers/gpu/drm/imagination/pvr_drv.c    |  2 +-
 drivers/gpu/drm/imagination/pvr_gem.c    |  3 +++
 drivers/gpu/drm/imagination/pvr_gem.h    |  7 +++++--
 drivers/gpu/drm/imagination/pvr_mmu.c    |  7 ++++++-
 6 files changed, 36 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/imagination/pvr_device.c b/drivers/gpu/drm/imagination/pvr_device.c
index 2ce46b9a8ab7609faebeeb4e7820751b00047806..ffc177c383c1be16061eff0290c347918b0991f7 100644
--- a/drivers/gpu/drm/imagination/pvr_device.c
+++ b/drivers/gpu/drm/imagination/pvr_device.c
@@ -24,6 +24,7 @@
 #include <linux/gfp.h>
 #include <linux/interrupt.h>
 #include <linux/platform_device.h>
+#include <linux/property.h>
 #include <linux/pm_runtime.h>
 #include <linux/slab.h>
 #include <linux/stddef.h>
@@ -635,6 +636,7 @@ bool
 pvr_device_overrides_validate(struct pvr_device *pvr_dev,
 			      const struct pvr_device_overrides *overrides)
 {
+	struct drm_device *drm_dev = from_pvr_device(pvr_dev);
 	bool ret = true;
 
 	/*
@@ -643,7 +645,14 @@ pvr_device_overrides_validate(struct pvr_device *pvr_dev,
 	 *
 	 * Note that this function may be called early during device initialization
 	 * so it should not be assumed that @pvr_dev is ready for normal use yet.
-	 */
+	*/
+
+	if (overrides->device_memory_force_cpu_cached &&
+	    device_get_dma_attr(drm_dev->dev) != DEV_DMA_COHERENT) {
+		drm_err(drm_dev,
+			"Specifying device_memory_force_cpu_cached override without dma-coherent attribute is unsupported.");
+		ret = false;
+	}
 
 	return ret;
 }
diff --git a/drivers/gpu/drm/imagination/pvr_device.h b/drivers/gpu/drm/imagination/pvr_device.h
index ad0a02a37154099542247dfc62f411c10f4e41f4..7ae14899db24f4c747e8cf4d61d252eb403713f4 100644
--- a/drivers/gpu/drm/imagination/pvr_device.h
+++ b/drivers/gpu/drm/imagination/pvr_device.h
@@ -60,8 +60,19 @@ struct pvr_fw_version {
 /**
  * struct pvr_device_overrides - Hardware-level overrides loaded from
  * MODULE_DEVICE_TABLE() or similar.
+ *
+ * @device_memory_force_cpu_cached: By default, all device memory buffer objects
+ * are mapped write-combined on the CPU (see %PVR_BO_CPU_CACHED) including MMU
+ * page table backing pages which do not use the regular device memory objects.
+ * This override forces all CPU mappings to be mapped cached instead. Since this
+ * could require additional cache maintenance operations to be performed,
+ * pvr_device_overrides_validate() ensures that the dma-coherent attribute is
+ * set when this override is specified. Required on some TI platforms where a
+ * bug causes device-to-cpu cache snooping to behave incorrectly when
+ * interacting with cpu-uncached memory.
  */
 struct pvr_device_overrides {
+	bool device_memory_force_cpu_cached;
 };
 
 /**
diff --git a/drivers/gpu/drm/imagination/pvr_drv.c b/drivers/gpu/drm/imagination/pvr_drv.c
index b56ee2cda9b54c4388a6eef38b0ff81acdb05874..e074cfb0d2055b5387dbb142ca972108977f9854 100644
--- a/drivers/gpu/drm/imagination/pvr_drv.c
+++ b/drivers/gpu/drm/imagination/pvr_drv.c
@@ -1490,7 +1490,7 @@ static void pvr_remove(struct platform_device *plat_dev)
 	pvr_power_domains_fini(pvr_dev);
 }
 
-static const struct pvr_device_overrides pvr_device_overrides_default = {};
+static const struct pvr_device_overrides pvr_device_overrides_default = { 0 };
 
 /*
  * Always specify &pvr_device_overrides_default instead of %NULL for &struct of_device_id->data so
diff --git a/drivers/gpu/drm/imagination/pvr_gem.c b/drivers/gpu/drm/imagination/pvr_gem.c
index 6a8c81fe8c1e85c2130a4fe90fce35b6a2be35aa..c67c30518f89af3de2e617a9b65e5cd78870fa2c 100644
--- a/drivers/gpu/drm/imagination/pvr_gem.c
+++ b/drivers/gpu/drm/imagination/pvr_gem.c
@@ -345,6 +345,9 @@ pvr_gem_object_create(struct pvr_device *pvr_dev, size_t size, u64 flags)
 	if (size == 0 || !pvr_gem_object_flags_validate(flags))
 		return ERR_PTR(-EINVAL);
 
+	if (PVR_HAS_OVERRIDE(pvr_dev, device_memory_force_cpu_cached))
+		flags |= PVR_BO_CPU_CACHED;
+
 	shmem_obj = drm_gem_shmem_create(from_pvr_device(pvr_dev), size);
 	if (IS_ERR(shmem_obj))
 		return ERR_CAST(shmem_obj);
diff --git a/drivers/gpu/drm/imagination/pvr_gem.h b/drivers/gpu/drm/imagination/pvr_gem.h
index e0e5ea509a2e88a437b8d241ea13c7bab2220f56..9b3cbcbe48dfbbc8be211a8a409699a43452e178 100644
--- a/drivers/gpu/drm/imagination/pvr_gem.h
+++ b/drivers/gpu/drm/imagination/pvr_gem.h
@@ -44,8 +44,11 @@ struct pvr_file;
  * Bits not defined anywhere are "undefined".
  *
  * CPU mapping options
- *    :PVR_BO_CPU_CACHED: By default, all GEM objects are mapped write-combined on the CPU. Set this
- *       flag to override this behaviour and map the object cached.
+ *    :PVR_BO_CPU_CACHED: By default, all GEM objects are mapped write-combined on the CPU. Set
+ *       this flag to override this behaviour and map the object cached. If
+ *       &struct pvr_device_overrides->device_memory_force_cpu_cached is specified, all allocations
+ *       will be mapped as if this flag was set. This does not require any additional consideration
+ *       at allocation time since the override is only valid if the dma-coherent attribute is set.
  *
  * Firmware options
  *    :PVR_BO_FW_NO_CLEAR_ON_RESET: By default, all FW objects are cleared and reinitialised on hard
diff --git a/drivers/gpu/drm/imagination/pvr_mmu.c b/drivers/gpu/drm/imagination/pvr_mmu.c
index 4fe70610ed94cf707e631f8148af081a94f97327..7c7deb29b735308eaed26900f2f54a838382c255 100644
--- a/drivers/gpu/drm/imagination/pvr_mmu.c
+++ b/drivers/gpu/drm/imagination/pvr_mmu.c
@@ -259,6 +259,7 @@ pvr_mmu_backing_page_init(struct pvr_mmu_backing_page *page,
 	struct device *dev = from_pvr_device(pvr_dev)->dev;
 
 	struct page *raw_page;
+	pgprot_t prot;
 	int err;
 
 	dma_addr_t dma_addr;
@@ -268,7 +269,11 @@ pvr_mmu_backing_page_init(struct pvr_mmu_backing_page *page,
 	if (!raw_page)
 		return -ENOMEM;
 
-	host_ptr = vmap(&raw_page, 1, VM_MAP, pgprot_writecombine(PAGE_KERNEL));
+	prot = PAGE_KERNEL;
+	if (!PVR_HAS_OVERRIDE(pvr_dev, device_memory_force_cpu_cached))
+		prot = pgprot_writecombine(prot);
+
+	host_ptr = vmap(&raw_page, 1, VM_MAP, prot);
 	if (!host_ptr) {
 		err = -ENOMEM;
 		goto err_free_page;

-- 
2.47.0

Re: [PATCH v2 19/21] drm/imagination: Add device_memory_force_cpu_cached override

[Andrew Davis]

On 11/18/24 7:02 AM, Matt Coster wrote:
> The TI k3-j721s2 platform has a bug relating to cache snooping on the AXI

Well we don't really know it is the bug on our side until we root cause it..

Anyway, why do need such a complex work around here? GEM buffer objects
only need to be coherent, cached or uncached is not a concern for userspace.
For a device marked dma-coherent, both uncached and cached memory is coherent.

So, you could skip remapping these buffers as writecombine always when the
device is marked dma-coherent. Nothing J721s2 specific to that, and you
could drop all this PVR_HAS_OVERRIDE stuff here and in the previous patch.

Andrew

> ACE-Lite interface. Disabling cache snooping altogether would also resolve
> the issue, but is considered more of a performance hit.
> 
> Given the platform is dma-coherent, forcing all device-accessible memory
> allocations through the CPU cache is the preferred solution.
> 
> Implement this workaround so that it can later be enabled for the TI
> k3-j721s2 platform.
> 
> Signed-off-by: Matt Coster <matt.coster@imgtec.com>
> ---
> Changes in v2:
> - None
> - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-19-4ed30e865892@imgtec.com
> ---
>   drivers/gpu/drm/imagination/pvr_device.c | 11 ++++++++++-
>   drivers/gpu/drm/imagination/pvr_device.h | 11 +++++++++++
>   drivers/gpu/drm/imagination/pvr_drv.c    |  2 +-
>   drivers/gpu/drm/imagination/pvr_gem.c    |  3 +++
>   drivers/gpu/drm/imagination/pvr_gem.h    |  7 +++++--
>   drivers/gpu/drm/imagination/pvr_mmu.c    |  7 ++++++-
>   6 files changed, 36 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/gpu/drm/imagination/pvr_device.c b/drivers/gpu/drm/imagination/pvr_device.c
> index 2ce46b9a8ab7609faebeeb4e7820751b00047806..ffc177c383c1be16061eff0290c347918b0991f7 100644
> --- a/drivers/gpu/drm/imagination/pvr_device.c
> +++ b/drivers/gpu/drm/imagination/pvr_device.c
> @@ -24,6 +24,7 @@
>   #include <linux/gfp.h>
>   #include <linux/interrupt.h>
>   #include <linux/platform_device.h>
> +#include <linux/property.h>
>   #include <linux/pm_runtime.h>
>   #include <linux/slab.h>
>   #include <linux/stddef.h>
> @@ -635,6 +636,7 @@ bool
>   pvr_device_overrides_validate(struct pvr_device *pvr_dev,
>   			      const struct pvr_device_overrides *overrides)
>   {
> +	struct drm_device *drm_dev = from_pvr_device(pvr_dev);
>   	bool ret = true;
>   
>   	/*
> @@ -643,7 +645,14 @@ pvr_device_overrides_validate(struct pvr_device *pvr_dev,
>   	 *
>   	 * Note that this function may be called early during device initialization
>   	 * so it should not be assumed that @pvr_dev is ready for normal use yet.
> -	 */
> +	*/
> +
> +	if (overrides->device_memory_force_cpu_cached &&
> +	    device_get_dma_attr(drm_dev->dev) != DEV_DMA_COHERENT) {
> +		drm_err(drm_dev,
> +			"Specifying device_memory_force_cpu_cached override without dma-coherent attribute is unsupported.");
> +		ret = false;
> +	}
>   
>   	return ret;
>   }
> diff --git a/drivers/gpu/drm/imagination/pvr_device.h b/drivers/gpu/drm/imagination/pvr_device.h
> index ad0a02a37154099542247dfc62f411c10f4e41f4..7ae14899db24f4c747e8cf4d61d252eb403713f4 100644
> --- a/drivers/gpu/drm/imagination/pvr_device.h
> +++ b/drivers/gpu/drm/imagination/pvr_device.h
> @@ -60,8 +60,19 @@ struct pvr_fw_version {
>   /**
>    * struct pvr_device_overrides - Hardware-level overrides loaded from
>    * MODULE_DEVICE_TABLE() or similar.
> + *
> + * @device_memory_force_cpu_cached: By default, all device memory buffer objects
> + * are mapped write-combined on the CPU (see %PVR_BO_CPU_CACHED) including MMU
> + * page table backing pages which do not use the regular device memory objects.
> + * This override forces all CPU mappings to be mapped cached instead. Since this
> + * could require additional cache maintenance operations to be performed,
> + * pvr_device_overrides_validate() ensures that the dma-coherent attribute is
> + * set when this override is specified. Required on some TI platforms where a
> + * bug causes device-to-cpu cache snooping to behave incorrectly when
> + * interacting with cpu-uncached memory.
>    */
>   struct pvr_device_overrides {
> +	bool device_memory_force_cpu_cached;
>   };
>   
>   /**
> diff --git a/drivers/gpu/drm/imagination/pvr_drv.c b/drivers/gpu/drm/imagination/pvr_drv.c
> index b56ee2cda9b54c4388a6eef38b0ff81acdb05874..e074cfb0d2055b5387dbb142ca972108977f9854 100644
> --- a/drivers/gpu/drm/imagination/pvr_drv.c
> +++ b/drivers/gpu/drm/imagination/pvr_drv.c
> @@ -1490,7 +1490,7 @@ static void pvr_remove(struct platform_device *plat_dev)
>   	pvr_power_domains_fini(pvr_dev);
>   }
>   
> -static const struct pvr_device_overrides pvr_device_overrides_default = {};
> +static const struct pvr_device_overrides pvr_device_overrides_default = { 0 };
>   
>   /*
>    * Always specify &pvr_device_overrides_default instead of %NULL for &struct of_device_id->data so
> diff --git a/drivers/gpu/drm/imagination/pvr_gem.c b/drivers/gpu/drm/imagination/pvr_gem.c
> index 6a8c81fe8c1e85c2130a4fe90fce35b6a2be35aa..c67c30518f89af3de2e617a9b65e5cd78870fa2c 100644
> --- a/drivers/gpu/drm/imagination/pvr_gem.c
> +++ b/drivers/gpu/drm/imagination/pvr_gem.c
> @@ -345,6 +345,9 @@ pvr_gem_object_create(struct pvr_device *pvr_dev, size_t size, u64 flags)
>   	if (size == 0 || !pvr_gem_object_flags_validate(flags))
>   		return ERR_PTR(-EINVAL);
>   
> +	if (PVR_HAS_OVERRIDE(pvr_dev, device_memory_force_cpu_cached))
> +		flags |= PVR_BO_CPU_CACHED;
> +
>   	shmem_obj = drm_gem_shmem_create(from_pvr_device(pvr_dev), size);
>   	if (IS_ERR(shmem_obj))
>   		return ERR_CAST(shmem_obj);
> diff --git a/drivers/gpu/drm/imagination/pvr_gem.h b/drivers/gpu/drm/imagination/pvr_gem.h
> index e0e5ea509a2e88a437b8d241ea13c7bab2220f56..9b3cbcbe48dfbbc8be211a8a409699a43452e178 100644
> --- a/drivers/gpu/drm/imagination/pvr_gem.h
> +++ b/drivers/gpu/drm/imagination/pvr_gem.h
> @@ -44,8 +44,11 @@ struct pvr_file;
>    * Bits not defined anywhere are "undefined".
>    *
>    * CPU mapping options
> - *    :PVR_BO_CPU_CACHED: By default, all GEM objects are mapped write-combined on the CPU. Set this
> - *       flag to override this behaviour and map the object cached.
> + *    :PVR_BO_CPU_CACHED: By default, all GEM objects are mapped write-combined on the CPU. Set
> + *       this flag to override this behaviour and map the object cached. If
> + *       &struct pvr_device_overrides->device_memory_force_cpu_cached is specified, all allocations
> + *       will be mapped as if this flag was set. This does not require any additional consideration
> + *       at allocation time since the override is only valid if the dma-coherent attribute is set.
>    *
>    * Firmware options
>    *    :PVR_BO_FW_NO_CLEAR_ON_RESET: By default, all FW objects are cleared and reinitialised on hard
> diff --git a/drivers/gpu/drm/imagination/pvr_mmu.c b/drivers/gpu/drm/imagination/pvr_mmu.c
> index 4fe70610ed94cf707e631f8148af081a94f97327..7c7deb29b735308eaed26900f2f54a838382c255 100644
> --- a/drivers/gpu/drm/imagination/pvr_mmu.c
> +++ b/drivers/gpu/drm/imagination/pvr_mmu.c
> @@ -259,6 +259,7 @@ pvr_mmu_backing_page_init(struct pvr_mmu_backing_page *page,
>   	struct device *dev = from_pvr_device(pvr_dev)->dev;
>   
>   	struct page *raw_page;
> +	pgprot_t prot;
>   	int err;
>   
>   	dma_addr_t dma_addr;
> @@ -268,7 +269,11 @@ pvr_mmu_backing_page_init(struct pvr_mmu_backing_page *page,
>   	if (!raw_page)
>   		return -ENOMEM;
>   
> -	host_ptr = vmap(&raw_page, 1, VM_MAP, pgprot_writecombine(PAGE_KERNEL));
> +	prot = PAGE_KERNEL;
> +	if (!PVR_HAS_OVERRIDE(pvr_dev, device_memory_force_cpu_cached))
> +		prot = pgprot_writecombine(prot);
> +
> +	host_ptr = vmap(&raw_page, 1, VM_MAP, prot);
>   	if (!host_ptr) {
>   		err = -ENOMEM;
>   		goto err_free_page;
> 

[PATCH v2 20/21] drm/imagination: Add support for TI AM68 GPU

[Matt Coster]

Add the new compatible string recently introduced in the dt bindings
("img,img-bxs-4-64") to the dt_match table and link the appropriate
firmware.

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- None
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-20-4ed30e865892@imgtec.com
---
 drivers/gpu/drm/imagination/pvr_drv.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/gpu/drm/imagination/pvr_drv.c b/drivers/gpu/drm/imagination/pvr_drv.c
index e074cfb0d2055b5387dbb142ca972108977f9854..044321cab1114864aefa2ca0024f326ac52d4dd6 100644
--- a/drivers/gpu/drm/imagination/pvr_drv.c
+++ b/drivers/gpu/drm/imagination/pvr_drv.c
@@ -44,6 +44,7 @@
  * This driver supports the following PowerVR/IMG graphics cores from Imagination Technologies:
  *
  * * AXE-1-16M (found in Texas Instruments AM62)
+ * * BXS-4-64 MC1 (found in Texas Instruments J721S2/AM68)
  */
 
 /**
@@ -1501,6 +1502,12 @@ static const struct of_device_id dt_match[] = {
 
 	/* All supported GPU models */
 	{ .compatible = "img,img-axe-1-16m", .data = &pvr_device_overrides_default },
+	{ .compatible = "img,img-bxs-4-64", .data = &pvr_device_overrides_default },
+
+	/* Vendor integrations which require overrides */
+	{ .compatible = "ti,j721s2-gpu", .data = &(struct pvr_device_overrides){
+		.device_memory_force_cpu_cached = true,
+	} },
 
 	/*
 	 * This legacy compatible string was introduced early on before the more specific GPU
@@ -1532,3 +1539,4 @@ MODULE_DESCRIPTION(PVR_DRIVER_DESC);
 MODULE_LICENSE("Dual MIT/GPL");
 MODULE_IMPORT_NS(DMA_BUF);
 MODULE_FIRMWARE("powervr/rogue_33.15.11.3_v1.fw");
+MODULE_FIRMWARE("powervr/rogue_36.53.104.796_v1.fw");

-- 
2.47.0

[PATCH v2 21/21] arm64: dts: ti: k3-j721s2: Add GPU node

[Matt Coster]

The J721S2 binding is based on the TI downstream binding in 54b0f2a00d92
("arm64: dts: ti: k3-j721s2-main: add gpu node") from [1] but with updated
compatible strings.

The clock[2] and power[3] indices were verified from docs, but the
source of the interrupt index remains elusive.

References for indices: clocks[1], interrupts[2], power[3].

[1]: https://git.ti.com/cgit/ti-linux-kernel/ti-linux-kernel
[2]: https://downloads.ti.com/tisci/esd/latest/5_soc_doc/j721s2/clocks.html
[3]: https://downloads.ti.com/tisci/esd/latest/5_soc_doc/j721s2/devices.html

Signed-off-by: Matt Coster <matt.coster@imgtec.com>
---
Changes in v2:
- Use normal reg syntax for 64-bit values
- Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-21-4ed30e865892@imgtec.com
---
 arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi b/arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi
index 9ed6949b40e9dfafdaf6861944b0b128b053a44f..9adc2c704ba4b38d1a0e7c9ded035fe79630451d 100644
--- a/arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi
+++ b/arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi
@@ -2047,4 +2047,16 @@ watchdog8: watchdog@23f0000 {
 		/* reserved for MAIN_R5F1_1 */
 		status = "reserved";
 	};
+
+	gpu: gpu@4e20000000 {
+		compatible = "ti,j721s2-gpu", "img,img-bxs-4-64", "img,img-rogue";
+		reg = <0x4e 0x20000000 0x00 0x80000>;
+		clocks = <&k3_clks 130 1>;
+		clock-names = "core";
+		interrupts = <GIC_SPI 24 IRQ_TYPE_LEVEL_HIGH>;
+		power-domains = <&k3_pds 130 TI_SCI_PD_EXCLUSIVE>,
+				<&k3_pds 373 TI_SCI_PD_EXCLUSIVE>;
+		power-domain-names = "a", "b";
+		dma-coherent;
+	};
 };

-- 
2.47.0

Re: [PATCH v2 21/21] arm64: dts: ti: k3-j721s2: Add GPU node

[Krzysztof Kozlowski]

On 18/11/2024 14:02, Matt Coster wrote:
> The J721S2 binding is based on the TI downstream binding in 54b0f2a00d92
> ("arm64: dts: ti: k3-j721s2-main: add gpu node") from [1] but with updated
> compatible strings.
> 
> The clock[2] and power[3] indices were verified from docs, but the
> source of the interrupt index remains elusive.
> 
> References for indices: clocks[1], interrupts[2], power[3].
> 
> [1]: https://git.ti.com/cgit/ti-linux-kernel/ti-linux-kernel
> [2]: https://downloads.ti.com/tisci/esd/latest/5_soc_doc/j721s2/clocks.html
> [3]: https://downloads.ti.com/tisci/esd/latest/5_soc_doc/j721s2/devices.html
> 
> Signed-off-by: Matt Coster <matt.coster@imgtec.com>
> ---
> Changes in v2:
> - Use normal reg syntax for 64-bit values
> - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-21-4ed30e865892@imgtec.com

For DRM maintainers:
This patch cannot be merged into DRM tree.


> ---
>  arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi | 12 ++++++++++++
>  1 file changed, 12 insertions(+)
> 
> diff --git a/arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi b/arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi
> index 9ed6949b40e9dfafdaf6861944b0b128b053a44f..9adc2c704ba4b38d1a0e7c9ded035fe79630451d 100644
> --- a/arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi
> +++ b/arch/arm64/boot/dts/ti/k3-j721s2-main.dtsi
> @@ -2047,4 +2047,16 @@ watchdog8: watchdog@23f0000 {
>  		/* reserved for MAIN_R5F1_1 */
>  		status = "reserved";
>  	};
> +
> +	gpu: gpu@4e20000000 {
Drop unused label.

Best regards,
Krzysztof