drm/asahi: dispatch G15 compute through extended T8122 firmware

MAINTAINERS

@@ -2723,7 +2723,9 @@ F:	drivers/pwm/pwm-apple.c
 F:	drivers/reset/reset-apple-cio.c
 F:	drivers/rtc/rtc-macsmc.c
 F:	drivers/soc/apple/*
+F:	tools/asahi/m3-uapi-*
 F:	tools/asahi/pmp-bootargs-tests.rs
+F:	tools/asahi/test-m3-uapi.py
 F:	drivers/spi/spi-apple.c
 F:	drivers/spmi/spmi-apple-controller.c
 F:	drivers/thunderbolt/.kunitconfig.apple-dpin

drivers/gpu/drm/apple/j613_dcp_activate.c

@@ -40,6 +40,13 @@ bool apple_t8122_25g83_board(void)
 	return ok;
 }
 
+/* Match the PMGR provider: automatic domains can be clock-gated while active. */
+static bool j613_power_state_active(u32 state)
+{
+	return FIELD_GET(GENMASK(7, 4), state) == 15 ||
+	       ((state & BIT(28)) && FIELD_GET(GENMASK(3, 0), state) == 15);
+}
+
 /* These are DT domain nodes, carrying the PMGR offsets, not copied registers. */
 static int j613_power_ready(void)
 {
@@ -82,7 +89,7 @@ static int j613_power_ready(void)
 			return -ENOMEM;
 		state = readl(pmgr + offset);
 		iounmap(pmgr);
-		if (FIELD_GET(GENMASK(7, 4), state) != 15)
+		if (!j613_power_state_active(state))
 			return -EBUSY;
 	}
 	return 0;

drivers/gpu/drm/asahi/m3/m3_adt_config.rs

@@ -1,3 +1,3 @@
 // SPDX-License-Identifier: GPL-2.0-only OR MIT
 
 
@@ -260,7 +260,14 @@
 
 /// The runtime backend's InitData contents: the generated object images and the
 /// performance-state policy.
+pub(crate) struct ExtendedInputs {
+    pub(crate) power: hw::PwrConfig,
+    pub(crate) sram_uv: KVec<u32>,
+    pub(crate) bases: super::m3_mmio_26::RegisterBases,
+}
+
 pub(crate) struct Contents {
+    pub(crate) extended: Option<ExtendedInputs>,
     /// Generated images, per owner; None for owners that stay zero. Only the three image
     /// owners' images are uploaded; the records' are checked.
     images: KVec<Option<KVVec<u8>>>,
@@ -304,6 +311,10 @@
         experiment: Option<&Experiment>,
     ) -> Result<Self> {
         let dev = pdev.as_ref();
+        if firmware.profile() == super::m3_profile::Profile::Extended {
+            if experiment.is_some() { return Err(EINVAL); }
+            return Self::extended(dev, soc);
+        }
         let param = m3_params::initdata_param();
         check_version(dev, firmware, experiment)?;
         let iomaps = match experiment {
@@ -370,12 +381,54 @@
         );
         let read_only_slots = experiment.map(|e| e.read_only_slots()).unwrap_or(0);
         let hwdata_object = soc.hwdata_object;
-        Ok(Contents { images, pstates, iomaps, table, read_only_slots, hwdata_object })
+        Ok(Contents { extended: None, images, pstates, iomaps, table, read_only_slots, hwdata_object })
+    }
+
+    fn extended(dev: &device::Device, soc: &Soc) -> Result<Self> {
+        let cfg = soc.hwcfg.ok_or(ENODEV)?;
+        if soc.chip_id != 0x8122 || cfg.chip_id != 0x8122 { return Err(ENODEV); }
+        let power = hw::PwrConfig::load(dev, cfg)?;
+        let tables = initdata::PStateTables::new(&power, true, None, false)?;
+        if tables.len != 9 || tables.max != 8 || tables.max_frequency_khz(&power) != 1_338_000 {
+            return Err(EINVAL);
+        }
+        let node = dev.of_node().ok_or(ENODEV)?;
+        let opps = node.parse_phandle(c_str!("operating-points-v2"), 0).ok_or(EINVAL)?;
+        let mut sram_uv = KVec::new();
+        for opp in opps.children() {
+            let values: KVec<u32> = opp.get_property(c_str!("opp-microvolt-sram"))?;
+            if values.len() != 1 { return Err(EINVAL); }
+            sram_uv.push(values[0], GFP_KERNEL)?;
+        }
+        if sram_uv.len() != power.perf_states.len() { return Err(EINVAL); }
+        let block = |index: usize| -> Result<u64> {
+            Ok(cfg.io_mappings.get(index).and_then(|value| value.as_ref()).ok_or(EINVAL)?.base as u64)
+        };
+        let bases = super::m3_mmio_26::RegisterBases {
+            gpu: soc.windows.sgx, interrupt: block(2)?,
+            memory: block(11)?.checked_add(0x1c4000).ok_or(EOVERFLOW)?,
+            telemetry_write: block(18)?, telemetry_read: block(19)?,
+            neural: block(25)?, metrology: block(26)?,
+        };
+        let mut freqs = [0; 16];
+        for (index, value) in freqs.iter_mut().enumerate().take(tables.len) {
+            *value = tables.primary_state(&power, index).freq_hz / 1_000_000;
+        }
+        // This profile enables the native firmware's thermal/PID controllers.
+        // The legacy host-only thermal governor must not replace those inputs.
+        let pstates = PstatePolicy { table_max: 8, table_max_mhz: 1338,
+            max: 8, safe: 8, boot: 1, max_mhz: 1338, safe_mhz: 1338,
+            boot_mhz: freqs[1], freqs, thermal: ThermalMode::Off };
+        let mut images = KVec::new();
+        for _ in 0..super::m3_init_26::COUNT { images.push(None, GFP_KERNEL)?; }
+        Ok(Self { extended: Some(ExtendedInputs { power, sram_uv, bases }),
+            images, pstates, iomaps: &[], table: (8, 1338), read_only_slots: 0, hwdata_object: None })
     }
 
     /// Check the contents before they are uploaded.
     pub(crate) fn check_for_upload(&self) -> Result {
-        if self.images.len() == storage::COUNT {
+        let count = if self.extended.is_some() { super::m3_init_26::COUNT } else { storage::COUNT };
+        if self.images.len() == count {
             Ok(())
         } else {
             Err(EINVAL)

drivers/gpu/drm/asahi/m3/m3_channel_format.rs

@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: GPL-2.0-only OR MIT
+
+//! Bounded firmware channel records. Publication and retirement belong to Config.
+
+use super::m3_init_layout::Error;
+
+pub(crate) const CONTROL_BYTES: usize = 64;
+pub(crate) const RECEIVE_BYTES: usize = 72;
+pub(crate) const SLOTS: u32 = 256;
+
+#[derive(Clone, Copy, Debug)]
+pub(crate) enum Action {
+    Initialize,
+    Register {
+        pipe: u32,
+        engine: u32,
+    },
+    Idle {
+        enabled: bool,
+    },
+    Release {
+        group: u8,
+        index: u8,
+        weight: u8,
+        address: u64,
+    },
+}
+
+impl Action {
+    /// The first two actions precede the first generation-bearing publication.
+    pub(crate) fn generation(self, published: u32, consumed: u32) -> Result<u32, Error> {
+        match self {
+            Self::Initialize | Self::Register { .. } => {
+                if published != 0 || consumed != 0 {
+                    return Err(Error::Address);
+                }
+                Ok(0)
+            }
+            _ => published.max(consumed).checked_add(1).ok_or(Error::Address),
+        }
+    }
+
+    pub(crate) fn encode(self, output: &mut [u8]) -> Result<(), Error> {
+        let mut words = [0u32; 3];
+        let mut release = None;
+        match self {
+            Self::Initialize => words[0] = 0x16,
+            Self::Register { pipe, engine } => {
+                if pipe >= 3 || !(1..=2).contains(&engine) {
+                    return Err(Error::Address);
+                }
+                words = [0x1b, pipe, engine];
+            }
+            Self::Idle { enabled } => words = [0x0a, enabled.into(), 0],
+            Self::Release {
+                group,
+                index,
+                weight,
+                address,
+            } => {
+                if index == 255
+                    || weight == 255
+                    || address < 0xffff_fc00_0000_0000
+                    || address & 7 != 0
+                {
+                    return Err(Error::Address);
+                }
+                words[0] = 0x14;
+                words[2] = u32::from_le_bytes([group, index, weight, 4]);
+                release = Some(address);
+            }
+        }
+        let bytes = output.get_mut(..CONTROL_BYTES).ok_or(Error::Size)?;
+        bytes.fill(0);
+        for (index, word) in words.iter().enumerate() {
+            bytes[index * 4..index * 4 + 4].copy_from_slice(&word.to_le_bytes());
+        }
+        if let Some(address) = release {
+            bytes[12..20].copy_from_slice(&address.to_le_bytes());
+        }
+        Ok(())
+    }
+}
+
+/// An empty ring is required before publishing a synchronous control transaction.
+/// Both consumers must retire it before the next control action may be published.
+pub(crate) fn idle_cursor(cursors: [u32; 3]) -> Result<u32, Error> {
+    if cursors.iter().any(|value| *value >= SLOTS) {
+        return Err(Error::Bounds);
+    }
+    if cursors[0] != cursors[2] || cursors[1] != cursors[2] {
+        return Err(Error::Address);
+    }
+    Ok(cursors[2])
+}
+
+pub(crate) fn payload_offset(thread: usize, slot: u32) -> Result<usize, Error> {
+    if thread >= 9 {
+        return Err(Error::Bounds);
+    }
+    Ok((thread * 256 + (slot & 255) as usize) * 0xd8)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn initialization_records_have_only_their_defined_words() {
+        for action in [Action::Initialize, Action::Register { pipe: 2, engine: 1 }] {
+            let mut output = [0xa5; CONTROL_BYTES + 8];
+            action.encode(&mut output).unwrap();
+            assert_eq!(&output[12..CONTROL_BYTES], &[0; CONTROL_BYTES - 12]);
+            assert_eq!(&output[CONTROL_BYTES..], &[0xa5; 8]);
+            assert_eq!(action.generation(0, 0), Ok(0));
+            assert!(action.generation(1, 0).is_err());
+        }
+    }
+
+    #[test]
+    fn invalid_actions_preserve_the_entire_output() {
+        for action in [
+            Action::Register { pipe: 3, engine: 1 },
+            Action::Register { pipe: 0, engine: 0 },
+            Action::Release {
+                group: 0,
+                index: 255,
+                weight: 1,
+                address: 0xffff_fc21_0000_0000,
+            },
+            Action::Release {
+                group: 0,
+                index: 0,
+                weight: 1,
+                address: 8,
+            },
+        ] {
+            let mut output = [0xa5; CONTROL_BYTES];
+            assert!(action.encode(&mut output).is_err());
+            assert_eq!(output, [0xa5; CONTROL_BYTES]);
+        }
+    }
+
+    #[test]
+    fn generations_follow_consumption_and_never_wrap() {
+        let action = Action::Idle { enabled: true };
+        assert_eq!(action.generation(3, 5), Ok(6));
+        assert_eq!(action.generation(5, 3), Ok(6));
+        assert!(action.generation(u32::MAX, 0).is_err());
+        assert!(action.generation(0, u32::MAX).is_err());
+        let mut output = [0; CONTROL_BYTES];
+        action.encode(&mut output).unwrap();
+        assert_eq!(&output[..8], &[10, 0, 0, 0, 1, 0, 0, 0]);
+    }
+
+    #[test]
+    fn both_consumers_and_all_indices_are_checked() {
+        assert_eq!(idle_cursor([255, 255, 255]), Ok(255));
+        for cursors in [[0, 1, 1], [1, 0, 1], [256, 256, 256]] {
+            assert!(idle_cursor(cursors).is_err());
+        }
+        assert_eq!(payload_offset(8, 0x101), Ok((8 * 256 + 1) * 0xd8));
+        assert!(payload_offset(9, 0).is_err());
+    }
+
+    #[test]
+    fn release_uses_the_packed_owned_pointer() {
+        let action = Action::Release {
+            group: 1,
+            index: 2,
+            weight: 3,
+            address: 0xffff_fc21_0400_0000,
+        };
+        let mut output = [0; CONTROL_BYTES];
+        action.encode(&mut output).unwrap();
+        assert_eq!(&output[8..12], &[1, 2, 3, 4]);
+        assert_eq!(
+            u64::from_le_bytes(output[12..20].try_into().unwrap()),
+            0xffff_fc21_0400_0000
+        );
+    }
+}

drivers/gpu/drm/asahi/m3/m3_client.rs

@@ -1,3 +1,3 @@
 // SPDX-License-Identifier: GPL-2.0-only OR MIT
 
 //! Safety checks for clients of the M3 (G15) render node.
@@ -35,7 +35,7 @@
 const AGX3_PARAMS_SIZE: u64 =
     (core::mem::offset_of!(uapi::drm_asahi_params_global, gpu_hal_generation)
         + core::mem::size_of::<u32>()) as u64;
-const _: () = assert!(AGX3_PARAMS_SIZE == 600);
+const _: () = assert!(AGX3_PARAMS_SIZE == super::m3_profile::PARAMS_REQUIRED);
 
 /// Refused clients that are logged per boot; later refusals only count.
 const REFUSALS_LOGGED: u32 = 16;
@@ -45,10 +45,19 @@
 /// [`AGX3_PARAMS_SIZE`] bytes (so it knows this GPU runs AGX3 command streams), and pass
 /// VM_CREATE flags 0 (legacy HAL). `size` is the largest GET_PARAMS size the file asked for.
 pub(crate) fn admit_vm(device: &kernel::device::Device, size: u64, flags: u32) -> Result {
+    admit_vm_for(device, super::m3_profile::Profile::Legacy, size, flags)
+}
+
+pub(crate) fn admit_vm_for(
+    device: &kernel::device::Device,
+    profile: super::m3_profile::Profile,
+    size: u64,
+    flags: u32,
+) -> Result {
     {
         // The positive acknowledgement of the command ABI (7.1.12 `g16_profile::client_admitted`):
         // GET_PARAMS coverage, and VM_CREATE flags 0 on the legacy-HAL M3 runtime.
-        if size >= AGX3_PARAMS_SIZE && flags == 0 {
+        if profile.vm_acknowledged(size, flags) {
             return Ok(());
         }
         let refusals = REFUSALS.fetch_add(1, Ordering::Relaxed).saturating_add(1);
@@ -64,11 +73,13 @@
             };
             dev_info!(
                 device,
-                "M3: refusing a VM for {}[{}]: it asked GET_PARAMS for {} of the {} parameter bytes, so it cannot know this GPU runs AGX3 command streams; M3-aware userspace is needed{}\n",
+                "M3: refusing a VM for {}[{}]: GET_PARAMS bytes={}/{} HAL={} flags={:#x}{}\n",
                 name,
                 task.group_leader().pid(),
                 size,
                 AGX3_PARAMS_SIZE,
+                profile.hal(),
+                flags,
                 last
             );
         }