* [PATCH v6 1/7] drm/tyr: add resources to RegistrationData
2026-07-09 21:36 [PATCH v6 0/7] drm/tyr: firmware loading and MCU boot support Deborah Brouwer
@ 2026-07-09 21:36 ` Deborah Brouwer
2026-07-09 21:36 ` [PATCH v6 2/7] drm/tyr: add a generic slot manager Deborah Brouwer
` (5 subsequent siblings)
6 siblings, 0 replies; 13+ messages in thread
From: Deborah Brouwer @ 2026-07-09 21:36 UTC (permalink / raw)
To: Daniel Almeida, Alice Ryhl, Danilo Krummrich, David Airlie,
Simona Vetter, Benno Lossin, Gary Guo
Cc: dri-devel, linux-kernel, rust-for-linux, Deborah Brouwer,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
Currently Tyr is not storing any resources in its drm::Driver
RegistrationData.
Move Tyr's device-private resources and gpu information from
drm::Driver::Data to drm::Driver::RegistrationData. This allows Tyr to
access this data safely within the lifetime of its binding to its parent
platform device and while registered with userspace.
Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
---
drivers/gpu/drm/tyr/driver.rs | 42 +++++++++++++++++++++---------------------
drivers/gpu/drm/tyr/file.rs | 11 ++++++-----
2 files changed, 27 insertions(+), 26 deletions(-)
diff --git a/drivers/gpu/drm/tyr/driver.rs b/drivers/gpu/drm/tyr/driver.rs
index 8348c6cd3929..46ce5c41e310 100644
--- a/drivers/gpu/drm/tyr/driver.rs
+++ b/drivers/gpu/drm/tyr/driver.rs
@@ -6,6 +6,7 @@
OptionalClk, //
},
device::{
+ Bound,
Core,
Device,
DeviceContext, //
@@ -27,10 +28,7 @@
regulator,
regulator::Regulator,
sizes::SZ_2M,
- sync::{
- aref::ARef,
- Mutex, //
- },
+ sync::Mutex,
time, //
};
@@ -53,13 +51,17 @@
#[pin_data(PinnedDrop)]
pub(crate) struct TyrPlatformDriverData<'bound> {
- _device: ARef<TyrDrmDevice>,
_reg: drm::Registration<'bound, TyrDrmDriver>,
}
+/// Data owned by the DRM [`Registration`].
+///
+/// This data can have references tied to the parent platform device binding scope
+/// and is accessible only while the DRM device is registered with userspace.
#[pin_data]
-pub(crate) struct TyrDrmDeviceData {
- pub(crate) pdev: ARef<platform::Device>,
+pub(crate) struct TyrDrmRegistrationData<'bound> {
+ /// Parent platform device.
+ pub(crate) pdev: &'bound platform::Device<Bound>,
#[pin]
clks: Mutex<Clocks>,
@@ -67,9 +69,10 @@ pub(crate) struct TyrDrmDeviceData {
#[pin]
regulators: Mutex<Regulators>,
- /// Some information on the GPU.
- ///
- /// This is mainly queried by userspace, i.e.: Mesa.
+ /// GPU MMIO register mapping.
+ pub(crate) iomem: IoMem<'bound>,
+
+ /// GPU information read from hardware during probe.
pub(crate) gpu_info: GpuInfo,
}
@@ -134,10 +137,10 @@ fn probe<'bound>(
// other threads of execution.
unsafe { pdev.dma_set_mask_and_coherent(DmaMask::try_new(pa_bits)?)? };
- let platform: ARef<platform::Device> = pdev.into();
+ let unreg_dev = drm::UnregisteredDevice::<TyrDrmDriver>::new(pdev, Ok(()))?;
- let data = try_pin_init!(TyrDrmDeviceData {
- pdev: platform.clone(),
+ let reg_data = try_pin_init!(TyrDrmRegistrationData {
+ pdev,
clks <- new_mutex!(Clocks {
core: core_clk,
stacks: stacks_clk,
@@ -147,18 +150,15 @@ fn probe<'bound>(
_mali: mali_regulator,
_sram: sram_regulator,
}),
+ iomem,
gpu_info,
});
- let tdev = drm::UnregisteredDevice::<TyrDrmDriver>::new(pdev, data)?;
// SAFETY: `reg` is stored in `TyrPlatformDriverData` and dropped when the driver is
// unbound; it is never forgotten.
- let reg = unsafe { drm::Registration::new(pdev.as_ref(), tdev, (), 0)? };
+ let reg = unsafe { drm::Registration::new(pdev.as_ref(), unreg_dev, reg_data, 0)? };
- let driver = TyrPlatformDriverData {
- _device: reg.device().into(),
- _reg: reg,
- };
+ let driver = TyrPlatformDriverData { _reg: reg };
// We need this to be dev_info!() because dev_dbg!() does not work at
// all in Rust for now, and we need to see whether probe succeeded.
@@ -184,8 +184,8 @@ fn drop(self: Pin<&mut Self>) {}
#[vtable]
impl drm::Driver for TyrDrmDriver {
- type Data = TyrDrmDeviceData;
- type RegistrationData<'a> = ();
+ type Data = ();
+ type RegistrationData<'bound> = TyrDrmRegistrationData<'bound>;
type File = TyrDrmFileData;
type Object = drm::gem::shmem::Object<BoData>;
type ParentDevice<Ctx: DeviceContext> = platform::Device<Ctx>;
diff --git a/drivers/gpu/drm/tyr/file.rs b/drivers/gpu/drm/tyr/file.rs
index b686041d5d6b..9f60a90d4948 100644
--- a/drivers/gpu/drm/tyr/file.rs
+++ b/drivers/gpu/drm/tyr/file.rs
@@ -12,7 +12,8 @@
use crate::driver::{
TyrDrmDevice,
- TyrDrmDriver, //
+ TyrDrmDriver,
+ TyrDrmRegistrationData, //
};
#[pin_data]
@@ -31,15 +32,15 @@ fn open(_dev: &drm::Device<Self::Driver>) -> Result<Pin<KBox<Self>>> {
impl TyrDrmFileData {
pub(crate) fn dev_query(
- ddev: &TyrDrmDevice<Registered>,
- _reg_data: &(),
+ _ddev: &TyrDrmDevice<Registered>,
+ reg_data: &TyrDrmRegistrationData<'_>,
devquery: &mut uapi::drm_panthor_dev_query,
_file: &TyrDrmFile,
) -> Result<u32> {
if devquery.pointer == 0 {
match devquery.type_ {
uapi::drm_panthor_dev_query_type_DRM_PANTHOR_DEV_QUERY_GPU_INFO => {
- devquery.size = core::mem::size_of_val(&ddev.gpu_info) as u32;
+ devquery.size = core::mem::size_of_val(®_data.gpu_info) as u32;
Ok(0)
}
_ => Err(EINVAL),
@@ -53,7 +54,7 @@ pub(crate) fn dev_query(
)
.writer();
- writer.write(&ddev.gpu_info)?;
+ writer.write(®_data.gpu_info)?;
Ok(0)
}
--
2.54.0
^ permalink raw reply related [flat|nested] 13+ messages in thread* [PATCH v6 2/7] drm/tyr: add a generic slot manager
2026-07-09 21:36 [PATCH v6 0/7] drm/tyr: firmware loading and MCU boot support Deborah Brouwer
2026-07-09 21:36 ` [PATCH v6 1/7] drm/tyr: add resources to RegistrationData Deborah Brouwer
@ 2026-07-09 21:36 ` Deborah Brouwer
2026-07-10 13:23 ` Alice Ryhl
2026-07-09 21:36 ` [PATCH v6 3/7] drm/tyr: add Memory Management Unit (MMU) support Deborah Brouwer
` (4 subsequent siblings)
6 siblings, 1 reply; 13+ messages in thread
From: Deborah Brouwer @ 2026-07-09 21:36 UTC (permalink / raw)
To: Daniel Almeida, Alice Ryhl, Danilo Krummrich, David Airlie,
Simona Vetter, Benno Lossin, Gary Guo
Cc: dri-devel, linux-kernel, rust-for-linux, Deborah Brouwer,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
From: Boris Brezillon <boris.brezillon@collabora.com>
Introduce a generic slot manager to dynamically allocate limited hardware
slots to software "seats". It can be used for both address space (AS) and
command stream group (CSG) slots.
The slot manager initially assigns seats to its free slots. It will
continue to reuse the same slot for a seat, as long as another seat does
not start to use the slot in the interim.
When contention arises because all of the slots are allocated, the slot
manager will lazily evict and reuse slots that have become idle (if any).
The seat state is protected using the LockedBy pattern with the same lock
that guards the SlotManager. This ensures the seat state stays consistent
across slot operations.
Hardware specific behaviour is controlled through the SlotManager's
specific manager type that implements the `SlotOperations` trait.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Co-developed-by: Deborah Brouwer <deborah.brouwer@collabora.com>
Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
---
drivers/gpu/drm/tyr/slot.rs | 385 ++++++++++++++++++++++++++++++++++++++++++++
drivers/gpu/drm/tyr/tyr.rs | 1 +
2 files changed, 386 insertions(+)
diff --git a/drivers/gpu/drm/tyr/slot.rs b/drivers/gpu/drm/tyr/slot.rs
new file mode 100644
index 000000000000..3f58b23238ef
--- /dev/null
+++ b/drivers/gpu/drm/tyr/slot.rs
@@ -0,0 +1,385 @@
+// SPDX-License-Identifier: GPL-2.0 or MIT
+
+//! Slot management abstraction for limited hardware resources.
+//!
+//! This module provides a generic [`SlotManager`] that assigns limited hardware
+//! slots to logical "seats". A seat represents an entity (such as a virtual memory
+//! (VM) address space) that needs access to a hardware slot.
+//!
+//! The [`SlotManager`] tracks slot allocation using sequence numbers (seqno) to detect
+//! when a seat's binding has been invalidated. When a seat requests activation,
+//! the manager will either reuse the seat's existing slot (if still valid),
+//! allocate a free slot (if any are available), or evict the oldest idle slot if any
+//! slots are idle.
+//!
+//! Hardware-specific behavior is customized by implementing the [`SlotOperations`]
+//! trait, which allows callbacks when slots are activated or evicted.
+//!
+//! This is currently used for managing address space slots in the GPU, and it will
+//! also be used to manage Command Stream Group (CSG) interface slots in the future.
+//!
+//! [SlotOperations]: crate::slot::SlotOperations
+//! [SlotManager]: crate::slot::SlotManager
+#![allow(dead_code)]
+
+use core::{
+ mem::take,
+ ops::{
+ Deref,
+ DerefMut, //
+ }, //
+};
+
+use kernel::{
+ prelude::*,
+ sync::LockedBy, //
+};
+
+/// Seat information.
+///
+/// This can't be accessed directly by the element embedding a `Seat`,
+/// but is used by the generic slot manager logic to control residency
+/// of a certain object on a hardware slot.
+pub(crate) struct SeatInfo {
+ /// Slot used by this seat.
+ ///
+ /// This index is only valid if the slot pointed to by this index
+ /// has its `SlotInfo::seqno` match `SeatInfo::seqno`. Otherwise,
+ /// it means the object has been evicted from the hardware slot,
+ /// and a new slot needs to be acquired to make this object
+ /// resident again.
+ slot: u8,
+
+ /// Sequence number encoding the last time this seat was active.
+ /// We also use it to check if a slot is still bound to a seat.
+ seqno: u64,
+}
+
+/// Seat state.
+///
+/// This is meant to be embedded in the object that wants to acquire
+/// hardware slots. It also starts in the `Seat::NoSeat` state, and
+/// the slot manager will change the object value when an active/evict
+/// request is issued.
+#[derive(Default)]
+pub(crate) enum Seat {
+ #[expect(clippy::enum_variant_names)]
+ /// Resource is not resident.
+ ///
+ /// All objects start with a seat in the `Seat::NoSeat` state. The seat also
+ /// gets back to that state if the user requests eviction. It
+ /// can also end up in that state next time an operation is done
+ /// on a `Seat::Idle` seat and the slot manager finds out this
+ /// object has been evicted from the slot.
+ #[default]
+ NoSeat,
+
+ /// Resource is actively used and resident.
+ ///
+ /// When a seat is in the `Seat::Active` state, it can't be evicted, and the
+ /// slot pointed to by `SeatInfo::slot` is guaranteed to be reserved
+ /// for this object as long as the seat stays active.
+ Active(SeatInfo),
+
+ /// Resource is idle and might or might not be resident.
+ ///
+ /// When a seat is in the`Seat::Idle` state, we can't know for sure if the
+ /// object is resident or evicted until the next request we issue
+ /// to the slot manager. This tells the slot manager it can
+ /// reclaim the underlying slot if needed.
+ /// In order for the hardware to use this object again, the seat
+ /// needs to be turned into an `Seat::Active` state again
+ /// with a `SlotManager::activate()` call.
+ Idle(SeatInfo),
+}
+
+impl Seat {
+ /// Get the slot index this seat is pointing to.
+ ///
+ /// If the seat is not `Seat::Active` we can't trust the
+ /// `SeatInfo`. In that case `None` is returned, otherwise
+ /// `Some(SeatInfo::slot)` is returned.
+ pub(super) fn slot(&self) -> Option<u8> {
+ match self {
+ Self::Active(info) => Some(info.slot),
+ _ => None,
+ }
+ }
+}
+
+/// Information related to a slot.
+struct SlotInfo<T> {
+ /// Type specific data attached to a slot.
+ slot_data: T,
+
+ /// Sequence number from when this slot was last activated.
+ seqno: u64,
+}
+
+/// Slot state.
+#[derive(Default)]
+enum SlotState<T> {
+ /// Slot is free.
+ #[default]
+ Free,
+
+ /// Slot is active.
+ Active(SlotInfo<T>),
+
+ /// Slot is idle.
+ Idle(SlotInfo<T>),
+}
+
+/// Trait describing the slot-related operations.
+pub(crate) trait SlotOperations {
+ /// Implementation-specific data associated with each slot.
+ type SlotData;
+
+ /// Called when a slot is being activated for a seat.
+ fn activate(&mut self, _slot_idx: usize, _slot_data: &Self::SlotData) -> Result {
+ Ok(())
+ }
+
+ /// Called when a slot is being evicted and freed.
+ fn evict(&mut self, _slot_idx: usize, _slot_data: &Self::SlotData) -> Result {
+ Ok(())
+ }
+}
+
+/// A generic slot manager that provides access to a limited number of hardware slots.
+pub(crate) struct SlotManager<T: SlotOperations, const MAX_SLOTS: usize> {
+ /// A specific implementation of the generic slot manager.
+ manager: T,
+
+ /// Number of slots actually available.
+ slot_count: usize,
+
+ /// Slot array used to track the state of each slot.
+ slots: [SlotState<T::SlotData>; MAX_SLOTS],
+
+ /// Sequence number incremented each time a Seat is successfully activated
+ use_seqno: u64,
+}
+
+/// A `Seat` that can only be accessed when holding a `SlotManager` lock guard.
+type LockedSeat<T, const MAX_SLOTS: usize> = LockedBy<Seat, SlotManager<T, MAX_SLOTS>>;
+
+impl<T: SlotOperations, const MAX_SLOTS: usize> SlotManager<T, MAX_SLOTS> {
+ /// Creates a specific instance of a slot manager.
+ pub(crate) fn new(manager: T, slot_count: usize) -> Result<Self> {
+ if slot_count == 0 {
+ pr_err!("Invalid slot count: 0");
+ return Err(EINVAL);
+ }
+ if slot_count > MAX_SLOTS {
+ pr_err!(
+ "Slot count: {} cannot exceed MAX_SLOTS: {}",
+ slot_count,
+ MAX_SLOTS
+ );
+ return Err(EINVAL);
+ }
+ Ok(Self {
+ manager,
+ slot_count,
+ slots: [const { SlotState::Free }; MAX_SLOTS],
+ use_seqno: 1,
+ })
+ }
+
+ /// Records a slot as active for the given seat.
+ fn record_active_slot(
+ &mut self,
+ slot_idx: usize,
+ locked_seat: &LockedSeat<T, MAX_SLOTS>,
+ slot_data: T::SlotData,
+ ) -> Result {
+ let cur_seqno = self.use_seqno;
+
+ *locked_seat.access_mut(self) = Seat::Active(SeatInfo {
+ slot: slot_idx as u8,
+ seqno: cur_seqno,
+ });
+
+ self.slots[slot_idx] = SlotState::Active(SlotInfo {
+ slot_data,
+ seqno: cur_seqno,
+ });
+
+ self.use_seqno += 1;
+ Ok(())
+ }
+
+ /// Activates a slot for the given seat.
+ fn activate_slot(
+ &mut self,
+ slot_idx: usize,
+ locked_seat: &LockedSeat<T, MAX_SLOTS>,
+ slot_data: T::SlotData,
+ ) -> Result {
+ self.manager.activate(slot_idx, &slot_data)?;
+ self.record_active_slot(slot_idx, locked_seat, slot_data)
+ }
+
+ /// Finds a slot for the given seat. A free slot is preferred, but if none
+ /// are available, the oldest idle slot is evicted and reused. Otherwise, if
+ /// there are no free or idle slots, return [`EBUSY`].
+ fn allocate_slot(
+ &mut self,
+ locked_seat: &LockedSeat<T, MAX_SLOTS>,
+ slot_data: T::SlotData,
+ ) -> Result {
+ let slots = &self.slots[..self.slot_count];
+
+ let mut idle_slot_idx = None;
+ let mut idle_slot_seqno: u64 = 0;
+
+ for (slot_idx, slot) in slots.iter().enumerate() {
+ match slot {
+ SlotState::Free => {
+ return self.activate_slot(slot_idx, locked_seat, slot_data);
+ }
+ SlotState::Idle(slot_info) => {
+ if idle_slot_idx.is_none() || slot_info.seqno < idle_slot_seqno {
+ idle_slot_idx = Some(slot_idx);
+ idle_slot_seqno = slot_info.seqno;
+ }
+ }
+ SlotState::Active(_) => (),
+ }
+ }
+
+ match idle_slot_idx {
+ Some(slot_idx) => {
+ // Lazily evict idle slot just before it is reused.
+ if let SlotState::Idle(slot_info) = &self.slots[slot_idx] {
+ self.manager.evict(slot_idx, &slot_info.slot_data)?;
+ }
+ self.activate_slot(slot_idx, locked_seat, slot_data)
+ }
+ None => {
+ pr_err!(
+ "Slot allocation failed: all {} slots in use\n",
+ self.slot_count
+ );
+ Err(EBUSY)
+ }
+ }
+ }
+
+ /// Converts an active slot and its seat to idle state.
+ fn idle_slot(&mut self, slot_idx: usize, locked_seat: &LockedSeat<T, MAX_SLOTS>) -> Result {
+ let slot = take(&mut self.slots[slot_idx]);
+
+ // If the slot was active, make it idle.
+ if let SlotState::Active(slot_info) = slot {
+ self.slots[slot_idx] = SlotState::Idle(slot_info);
+ }
+
+ // If the seat was active, make it idle, or keep it idle if it was already idle.
+ *locked_seat.access_mut(self) = match locked_seat.access(self) {
+ Seat::Active(seat_info) | Seat::Idle(seat_info) => Seat::Idle(SeatInfo {
+ slot: seat_info.slot,
+ seqno: seat_info.seqno,
+ }),
+ Seat::NoSeat => Seat::NoSeat,
+ };
+ Ok(())
+ }
+
+ /// Evicts an active or idle slot: calls the eviction callback and marks the slot as free
+ /// and the seat as NoSeat.
+ fn evict_slot(&mut self, slot_idx: usize, locked_seat: &LockedSeat<T, MAX_SLOTS>) -> Result {
+ match &self.slots[slot_idx] {
+ SlotState::Active(slot_info) | SlotState::Idle(slot_info) => {
+ self.manager.evict(slot_idx, &slot_info.slot_data)?;
+ take(&mut self.slots[slot_idx]);
+ }
+ _ => (),
+ }
+
+ *locked_seat.access_mut(self) = Seat::NoSeat;
+ Ok(())
+ }
+
+ /// Checks that the seat state matches the slot's state.
+ /// If they don't match, the seat is stale and is reset to `NoSeat`.
+ fn check_seat(&mut self, locked_seat: &LockedSeat<T, MAX_SLOTS>) {
+ let (slot_idx, seat_seqno, is_active) = match locked_seat.access(self) {
+ Seat::Active(seat_info) => (seat_info.slot as usize, seat_info.seqno, true),
+ Seat::Idle(seat_info) => (seat_info.slot as usize, seat_info.seqno, false),
+ _ => return,
+ };
+
+ let valid = if is_active {
+ !kernel::warn_on!(!matches!(
+ &self.slots[slot_idx],
+ SlotState::Active(slot_info) if slot_info.seqno == seat_seqno
+ ))
+ } else {
+ matches!(
+ &self.slots[slot_idx],
+ SlotState::Idle(slot_info) if slot_info.seqno == seat_seqno
+ )
+ };
+
+ if !valid {
+ *locked_seat.access_mut(self) = Seat::NoSeat;
+ }
+ }
+
+ /// Activate a resource on any available/reclaimable slot.
+ pub(crate) fn activate(
+ &mut self,
+ locked_seat: &LockedSeat<T, MAX_SLOTS>,
+ slot_data: T::SlotData,
+ ) -> Result {
+ self.check_seat(locked_seat);
+ match locked_seat.access(self) {
+ // If a seat still has a valid slot, just reuse the slot and refresh the bookkeeping.
+ Seat::Active(seat_info) | Seat::Idle(seat_info) => {
+ self.record_active_slot(seat_info.slot as usize, locked_seat, slot_data)
+ }
+ _ => self.allocate_slot(locked_seat, slot_data),
+ }
+ }
+
+ /// Flag a resource as idle. This method will be used for user VM support.
+ #[expect(dead_code)]
+ pub(crate) fn idle(&mut self, locked_seat: &LockedSeat<T, MAX_SLOTS>) -> Result {
+ self.check_seat(locked_seat);
+ if let Seat::Active(seat_info) = locked_seat.access(self) {
+ self.idle_slot(seat_info.slot as usize, locked_seat)?;
+ }
+ Ok(())
+ }
+
+ /// Evict a resource from its slot.
+ pub(crate) fn evict(&mut self, locked_seat: &LockedSeat<T, MAX_SLOTS>) -> Result {
+ self.check_seat(locked_seat);
+
+ match locked_seat.access(self) {
+ Seat::Active(seat_info) | Seat::Idle(seat_info) => {
+ let slot_idx = seat_info.slot as usize;
+ self.evict_slot(slot_idx, locked_seat)?;
+ }
+ _ => (),
+ }
+
+ Ok(())
+ }
+}
+
+impl<T: SlotOperations, const MAX_SLOTS: usize> Deref for SlotManager<T, MAX_SLOTS> {
+ type Target = T;
+
+ fn deref(&self) -> &Self::Target {
+ &self.manager
+ }
+}
+
+impl<T: SlotOperations, const MAX_SLOTS: usize> DerefMut for SlotManager<T, MAX_SLOTS> {
+ fn deref_mut(&mut self) -> &mut Self::Target {
+ &mut self.manager
+ }
+}
diff --git a/drivers/gpu/drm/tyr/tyr.rs b/drivers/gpu/drm/tyr/tyr.rs
index 95cda7b0962f..7c9a8063b3b9 100644
--- a/drivers/gpu/drm/tyr/tyr.rs
+++ b/drivers/gpu/drm/tyr/tyr.rs
@@ -12,6 +12,7 @@
mod gem;
mod gpu;
mod regs;
+mod slot;
kernel::module_platform_driver! {
type: TyrPlatformDriver,
--
2.54.0
^ permalink raw reply related [flat|nested] 13+ messages in thread* Re: [PATCH v6 2/7] drm/tyr: add a generic slot manager
2026-07-09 21:36 ` [PATCH v6 2/7] drm/tyr: add a generic slot manager Deborah Brouwer
@ 2026-07-10 13:23 ` Alice Ryhl
0 siblings, 0 replies; 13+ messages in thread
From: Alice Ryhl @ 2026-07-10 13:23 UTC (permalink / raw)
To: Deborah Brouwer
Cc: Daniel Almeida, Danilo Krummrich, David Airlie, Simona Vetter,
Benno Lossin, Gary Guo, dri-devel, linux-kernel, rust-for-linux,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
On Thu, Jul 09, 2026 at 02:36:42PM -0700, Deborah Brouwer wrote:
> + None => {
> + pr_err!(
> + "Slot allocation failed: all {} slots in use\n",
> + self.slot_count
> + );
> + Err(EBUSY)
This branch seems like something that can happen in normal execution, so
I don't believe we should print anything here.
Alice
^ permalink raw reply [flat|nested] 13+ messages in thread
* [PATCH v6 3/7] drm/tyr: add Memory Management Unit (MMU) support
2026-07-09 21:36 [PATCH v6 0/7] drm/tyr: firmware loading and MCU boot support Deborah Brouwer
2026-07-09 21:36 ` [PATCH v6 1/7] drm/tyr: add resources to RegistrationData Deborah Brouwer
2026-07-09 21:36 ` [PATCH v6 2/7] drm/tyr: add a generic slot manager Deborah Brouwer
@ 2026-07-09 21:36 ` Deborah Brouwer
2026-07-10 13:45 ` Alice Ryhl
2026-07-09 21:36 ` [PATCH v6 4/7] drm/tyr: add GPU virtual memory (VM) support Deborah Brouwer
` (3 subsequent siblings)
6 siblings, 1 reply; 13+ messages in thread
From: Deborah Brouwer @ 2026-07-09 21:36 UTC (permalink / raw)
To: Daniel Almeida, Alice Ryhl, Danilo Krummrich, David Airlie,
Simona Vetter, Benno Lossin, Gary Guo
Cc: dri-devel, linux-kernel, rust-for-linux, Deborah Brouwer,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
From: Boris Brezillon <boris.brezillon@collabora.com>
Add Memory Management Unit (MMU) support in Tyr. The MMU module wraps a
SlotManager instance to allocate MMU address-space slots for use by
virtual memory (VM) address spaces. The MMU's SlotManager uses an
AddressSpaceManager to handle the hardware-specific callbacks. For
example, the AddressSpaceManager activates and evicts VMs from slots by
writing commands to the MMU registers.
Add an implementation block for the MMU's MEMATTR register to provide
a method for translating the Memory Attribute Indirection Register (MAIR)
format from the pagetable configuration to a format understood by the MMU.
Create an mmu instance during probe, it will be used by subsequent patches
in this series.
Wrap the iomem stored in TyrDrmRegistrationData in an Arc. The iomem
is stored in the mmu through its AddressSpaceManager. In anticipation
of the iomem also being stored in the firmware object, set up shared
ownership of the iomem now.
Update Kconfig to add the new MMU and IOMMU dependencies required
by this MMU module.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Co-developed-by: Deborah Brouwer <deborah.brouwer@collabora.com>
Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
---
drivers/gpu/drm/tyr/Kconfig | 3 +
drivers/gpu/drm/tyr/driver.rs | 13 +-
drivers/gpu/drm/tyr/mmu.rs | 103 +++++++
drivers/gpu/drm/tyr/mmu/address_space.rs | 484 +++++++++++++++++++++++++++++++
drivers/gpu/drm/tyr/regs.rs | 121 ++++++++
drivers/gpu/drm/tyr/tyr.rs | 1 +
6 files changed, 722 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/tyr/Kconfig b/drivers/gpu/drm/tyr/Kconfig
index 51a68ef8212c..61a2fd6f961a 100644
--- a/drivers/gpu/drm/tyr/Kconfig
+++ b/drivers/gpu/drm/tyr/Kconfig
@@ -5,9 +5,12 @@ config DRM_TYR
depends on DRM=y
depends on RUST
depends on ARM || ARM64 || COMPILE_TEST
+ depends on MMU
depends on !GENERIC_ATOMIC64 # for IOMMU_IO_PGTABLE_LPAE
depends on COMMON_CLK
+ depends on IOMMU_SUPPORT
default n
+ select IOMMU_IO_PGTABLE_LPAE
select RUST_DRM_GEM_SHMEM_HELPER
help
Rust DRM driver for ARM Mali CSF-based GPUs.
diff --git a/drivers/gpu/drm/tyr/driver.rs b/drivers/gpu/drm/tyr/driver.rs
index 46ce5c41e310..74b55a389754 100644
--- a/drivers/gpu/drm/tyr/driver.rs
+++ b/drivers/gpu/drm/tyr/driver.rs
@@ -28,7 +28,10 @@
regulator,
regulator::Regulator,
sizes::SZ_2M,
- sync::Mutex,
+ sync::{
+ Arc,
+ Mutex, //
+ },
time, //
};
@@ -37,6 +40,7 @@
gem::BoData,
gpu,
gpu::GpuInfo,
+ mmu::Mmu,
regs::gpu_control::*, //
};
@@ -70,7 +74,7 @@ pub(crate) struct TyrDrmRegistrationData<'bound> {
regulators: Mutex<Regulators>,
/// GPU MMIO register mapping.
- pub(crate) iomem: IoMem<'bound>,
+ pub(crate) iomem: Arc<IoMem<'bound>>,
/// GPU information read from hardware during probe.
pub(crate) gpu_info: GpuInfo,
@@ -121,7 +125,8 @@ fn probe<'bound>(
let sram_regulator = Regulator::<regulator::Enabled>::get(pdev.as_ref(), c"sram")?;
let request = pdev.io_request_by_index(0).ok_or(ENODEV)?;
- let iomem = request.iomap_sized::<SZ_2M>()?;
+
+ let iomem = Arc::new(request.iomap_sized::<SZ_2M>()?, GFP_KERNEL)?;
issue_soft_reset(pdev.as_ref(), &iomem)?;
gpu::l2_power_on(pdev.as_ref(), &iomem)?;
@@ -139,6 +144,8 @@ fn probe<'bound>(
let unreg_dev = drm::UnregisteredDevice::<TyrDrmDriver>::new(pdev, Ok(()))?;
+ let _mmu = Mmu::new(iomem.as_arc_borrow(), &gpu_info)?;
+
let reg_data = try_pin_init!(TyrDrmRegistrationData {
pdev,
clks <- new_mutex!(Clocks {
diff --git a/drivers/gpu/drm/tyr/mmu.rs b/drivers/gpu/drm/tyr/mmu.rs
new file mode 100644
index 000000000000..c0341557d730
--- /dev/null
+++ b/drivers/gpu/drm/tyr/mmu.rs
@@ -0,0 +1,103 @@
+// SPDX-License-Identifier: GPL-2.0 or MIT
+
+//! Memory Management Unit (MMU) module.
+//!
+//! The GPU MMU provides a limited number of memory address spaces for use by command streams.
+//! The MMU translates virtual addresses to physical addresses and manages memory configuration
+//! and access permissions.
+//!
+//! This MMU module is essentially a locked wrapper around a [`SlotManager`] instance.
+//! The [`SlotManager`] manages the assignment of virtual address spaces to hardware address-space
+//! (AS) slots. MMU commands such as updates and flushes are carried out by the
+//! [`AddressSpaceManager`] which actually writes to the MMU registers.
+#![allow(dead_code)]
+
+use core::ops::Range;
+
+use kernel::{
+ new_mutex,
+ prelude::*,
+ sync::{
+ Arc,
+ ArcBorrow,
+ Mutex, //
+ }, //
+};
+
+use crate::{
+ driver::IoMem,
+ gpu::GpuInfo,
+ mmu::address_space::{
+ AddressSpaceManager,
+ VmAsData, //
+ },
+ regs::{
+ gpu_control::AS_PRESENT,
+ MAX_AS, //
+ },
+ slot::SlotManager, //
+};
+
+pub(crate) mod address_space;
+
+pub(crate) type AsSlotManager<'bound> = SlotManager<AddressSpaceManager<'bound>, MAX_AS>;
+
+/// Locked wrapper for carrying out virtual memory (VM) operations on the MMU.
+#[pin_data]
+pub(crate) struct Mmu<'bound> {
+ /// Slot Manager instance used to allocate hardware slots and write to MMU registers.
+ #[pin]
+ pub(crate) as_manager: Mutex<AsSlotManager<'bound>>,
+}
+
+impl<'bound> Mmu<'bound> {
+ /// Create an MMU component for this device.
+ pub(crate) fn new(
+ iomem: ArcBorrow<'_, IoMem<'bound>>,
+ gpu_info: &GpuInfo,
+ ) -> Result<Arc<Mmu<'bound>>> {
+ let present = AS_PRESENT::from_raw(gpu_info.as_present).present().get();
+ let slot_count = present.count_ones().try_into()?;
+
+ let as_manager = AddressSpaceManager::new(iomem.into(), present)?;
+ let mmu_init = try_pin_init!(Self{
+ as_manager <- new_mutex!(SlotManager::new(as_manager, slot_count)?),
+ });
+ Arc::pin_init(mmu_init, GFP_KERNEL)
+ }
+
+ /// Assign a VM to an AS slot, provide a translation table,
+ /// and update the MMU to make the VM resident.
+ pub(crate) fn activate_vm(&self, vm: ArcBorrow<'_, VmAsData<'bound>>) -> Result {
+ self.as_manager.lock().activate_vm(vm)
+ }
+
+ /// Evict a VM from its AS slot and flush the MMU.
+ pub(crate) fn deactivate_vm(&self, vm: &VmAsData<'bound>) -> Result {
+ self.as_manager.lock().deactivate_vm(vm)
+ }
+
+ /// Flush MMU translation caches after a VM update.
+ pub(crate) fn flush_vm(&self, vm: &VmAsData<'bound>) -> Result {
+ self.as_manager.lock().flush_vm(vm)
+ }
+
+ /// Flags the start of a VM update.
+ ///
+ /// If the VM is resident, any GPU access on the memory range being
+ /// updated will be blocked until `Mmu::end_vm_update()` is called.
+ /// This guarantees the atomicity of a VM update.
+ /// If the VM is not resident, this is a NOP.
+ pub(crate) fn start_vm_update(&self, vm: &VmAsData<'bound>, region: &Range<u64>) -> Result {
+ self.as_manager.lock().start_vm_update(vm, region)
+ }
+
+ /// Flags the end of a VM update.
+ ///
+ /// If the VM is resident, this will let GPU accesses on the updated
+ /// range go through, in case any of them were blocked.
+ /// If the VM is not resident, this is a NOP.
+ pub(crate) fn end_vm_update(&self, vm: &VmAsData<'bound>) -> Result {
+ self.as_manager.lock().end_vm_update(vm)
+ }
+}
diff --git a/drivers/gpu/drm/tyr/mmu/address_space.rs b/drivers/gpu/drm/tyr/mmu/address_space.rs
new file mode 100644
index 000000000000..a97f44774576
--- /dev/null
+++ b/drivers/gpu/drm/tyr/mmu/address_space.rs
@@ -0,0 +1,484 @@
+// SPDX-License-Identifier: GPL-2.0 or MIT
+
+//! Address space module.
+//!
+//! This module handles the hardware interaction for MMU operations through
+//! MMIO register access.
+//!
+
+use core::ops::Range;
+
+use kernel::{
+ device::{
+ Bound,
+ Device, //
+ }, //
+ error::Result,
+ io::{
+ poll,
+ register::Array,
+ Io, //
+ },
+ iommu::pgtable::{
+ Config,
+ IoPageTable,
+ ARM64LPAES1, //
+ },
+ prelude::*,
+ sizes::{
+ SZ_2M,
+ SZ_4K, //
+ },
+ sync::{
+ Arc,
+ ArcBorrow,
+ LockedBy, //
+ },
+ time::Delta, //
+};
+
+use crate::{
+ driver::IoMem,
+ mmu::{
+ AsSlotManager,
+ Mmu, //
+ },
+ regs::{
+ mmu_control::mmu_as_control,
+ mmu_control::mmu_as_control::*,
+ MAX_AS, //
+ },
+ slot::{
+ Seat,
+ SlotOperations, //
+ }, //
+};
+
+/// Address space configuration values to be written to MMU registers.
+#[derive(Clone, Copy)]
+struct AddressSpaceConfig {
+ /// Translation configuration. Configures how the MMU walks the page table for this
+ /// address space.
+ transcfg: u64,
+
+ /// Translation table base address. The address of the page table.
+ transtab: u64,
+
+ /// Memory attributes such as cacheability.
+ memattr: u64,
+}
+
+/// Virtual memory (VM) address space data for use in MMU operations.
+#[pin_data]
+pub(crate) struct VmAsData<'bound> {
+ /// The address space seat tracks this VM's binding to a hardware address space slot.
+ /// Uses [`LockedBy`] to ensure safe concurrent access to the slot assignment state,
+ /// protected by the [`AsSlotManager`] lock.
+ as_seat: LockedBy<Seat, AsSlotManager<'bound>>,
+
+ /// Virtual address bits for this address space.
+ va_bits: u8,
+
+ /// The page table which maps GPU virtual addresses to physical addresses for this VM.
+ #[pin]
+ pub(crate) page_table: IoPageTable<'bound, ARM64LPAES1>,
+}
+
+impl<'bound> VmAsData<'bound> {
+ /// Creates VM address space data by initializing all of its fields.
+ pub(crate) fn new<'a>(
+ mmu: &'a Mmu<'bound>,
+ pdev: &'bound Device<Bound>,
+ va_bits: u32,
+ pa_bits: u32,
+ ) -> impl pin_init::PinInit<VmAsData<'bound>, Error> + 'a {
+ let pt_config = Config {
+ quirks: 0,
+ pgsize_bitmap: SZ_4K | SZ_2M,
+ ias: va_bits,
+ oas: pa_bits,
+ coherent_walk: false,
+ };
+
+ let page_table_init = IoPageTable::new(pdev, pt_config);
+
+ try_pin_init!(Self {
+ as_seat: LockedBy::new(&mmu.as_manager, Seat::NoSeat),
+ va_bits: va_bits as u8,
+ page_table <- page_table_init,
+ }? Error)
+ }
+
+ /// Computes the hardware configuration for this address space.
+ fn as_config(&self) -> Result<AddressSpaceConfig> {
+ let pt = &self.page_table;
+ // The hardware computes the valid input address range as:
+ // INA_BITS_VALID = min(HW_INA_BITS, 55 - INA_BITS)
+ // To configure our desired va_bits, we solve for INA_BITS:
+ // INA_BITS = 55 - va_bits
+ // This assumes HW_INA_BITS (hardware capability) >= va_bits.
+ let ina_bits_field_value = 55 - self.va_bits;
+ let ina_bits = match ina_bits_field_value {
+ 7 => mmu_as_control::InaBits::Bits48,
+ 8 => mmu_as_control::InaBits::Bits47,
+ 9 => mmu_as_control::InaBits::Bits46,
+ 10 => mmu_as_control::InaBits::Bits45,
+ 11 => mmu_as_control::InaBits::Bits44,
+ 12 => mmu_as_control::InaBits::Bits43,
+ 13 => mmu_as_control::InaBits::Bits42,
+ 14 => mmu_as_control::InaBits::Bits41,
+ 15 => mmu_as_control::InaBits::Bits40,
+ 16 => mmu_as_control::InaBits::Bits39,
+ 17 => mmu_as_control::InaBits::Bits38,
+ 18 => mmu_as_control::InaBits::Bits37,
+ 19 => mmu_as_control::InaBits::Bits36,
+ 20 => mmu_as_control::InaBits::Bits35,
+ 21 => mmu_as_control::InaBits::Bits34,
+ 22 => mmu_as_control::InaBits::Bits33,
+ 23 => mmu_as_control::InaBits::Bits32,
+ 24 => mmu_as_control::InaBits::Bits31,
+ 25 => mmu_as_control::InaBits::Bits30,
+ 26 => mmu_as_control::InaBits::Bits29,
+ 27 => mmu_as_control::InaBits::Bits28,
+ 28 => mmu_as_control::InaBits::Bits27,
+ 29 => mmu_as_control::InaBits::Bits26,
+ 30 => mmu_as_control::InaBits::Bits25,
+ _ => return Err(EINVAL),
+ };
+
+ let transcfg = mmu_as_control::TRANSCFG::zeroed()
+ .with_ptw_memattr(mmu_as_control::PtwMemattr::WriteBack)
+ .with_r_allocate(true)
+ .with_mode(mmu_as_control::AddressSpaceMode::Aarch64_4K)
+ .with_ina_bits(ina_bits)
+ .into_raw();
+
+ Ok(AddressSpaceConfig {
+ transcfg,
+ // SAFETY: The caller must ensure that the address space is evicted
+ // and cleaned up before the `VmAsData` is dropped.
+ transtab: unsafe { pt.ttbr() },
+ memattr: MEMATTR::from_mair(pt.mair()).into_raw(),
+ })
+ }
+}
+
+/// Coordinates all hardware-level address space operations through MMIO register
+/// operations including enabling, disabling, flushing, and updating address spaces.
+pub(crate) struct AddressSpaceManager<'bound> {
+ /// Memory-mapped I/O region for GPU register access.
+ iomem: Arc<IoMem<'bound>>,
+
+ /// Bitmask of present address space slots from GPU_AS_PRESENT register.
+ as_present: u32,
+}
+
+impl<'bound> AddressSpaceManager<'bound> {
+ /// Creates a new address space manager.
+ ///
+ /// Initializes the manager with references to the platform device and
+ /// I/O memory region, along with the bitmask of available AS slots.
+ pub(super) fn new(
+ iomem: Arc<IoMem<'bound>>,
+ as_present: u32,
+ ) -> Result<AddressSpaceManager<'bound>> {
+ Ok(Self { iomem, as_present })
+ }
+
+ /// Validates that an AS slot number is within range and present in hardware.
+ ///
+ /// Checks that the slot index is less than [`MAX_AS`] and that
+ /// the corresponding bit is set in the `as_present` mask read from the GPU.
+ ///
+ /// Returns [`EINVAL`] if the slot is out of range or not present in hardware.
+ fn validate_as_slot(&self, as_nr: usize) -> Result {
+ if as_nr >= MAX_AS {
+ pr_err!("AS slot {} out of valid range (max {})\n", as_nr, MAX_AS);
+ return Err(EINVAL);
+ }
+
+ if (self.as_present & (1 << as_nr)) == 0 {
+ pr_err!(
+ "AS slot {} not present in hardware (AS_PRESENT={:#x})\n",
+ as_nr,
+ self.as_present
+ );
+ return Err(EINVAL);
+ }
+
+ Ok(())
+ }
+
+ /// Waits for an AS slot to become ready (not active).
+ ///
+ /// Returns an error if polling times out after 10ms or if register access fails.
+ fn as_wait_ready(&self, as_nr: usize) -> Result {
+ let io = &*self.iomem;
+ let op = || {
+ let status_reg = STATUS::try_at(as_nr).ok_or(EINVAL)?;
+ Ok(io.read(status_reg))
+ };
+ let cond = |status: &STATUS| -> bool { !status.active_ext() };
+ poll::read_poll_timeout(op, cond, Delta::from_millis(0), Delta::from_millis(10))?;
+
+ Ok(())
+ }
+
+ /// Sends a command to an AS slot.
+ ///
+ /// Returns an error if waiting for ready times out or if register write fails.
+ fn as_send_cmd(&mut self, as_nr: usize, cmd: MmuCommand) -> Result {
+ self.as_wait_ready(as_nr)?;
+ let io = &*self.iomem;
+ let command_reg = COMMAND::try_at(as_nr).ok_or(EINVAL)?;
+ io.write(command_reg, COMMAND::zeroed().with_command(cmd));
+ Ok(())
+ }
+
+ /// Sends a command to an AS slot and waits for completion.
+ ///
+ /// Returns an error if sending the command fails or if waiting for completion times out.
+ fn as_send_cmd_and_wait(&mut self, as_nr: usize, cmd: MmuCommand) -> Result {
+ self.as_send_cmd(as_nr, cmd)?;
+ self.as_wait_ready(as_nr)?;
+ Ok(())
+ }
+
+ /// Enables an AS slot with the provided configuration.
+ ///
+ /// Returns an error if the slot is invalid or if register writes/commands fail.
+ fn as_enable(&mut self, as_nr: usize, as_config: &AddressSpaceConfig) -> Result {
+ self.validate_as_slot(as_nr)?;
+
+ let io = &*self.iomem;
+
+ let transtab = as_config.transtab;
+ io.write(
+ TRANSTAB_LO::try_at(as_nr).ok_or(EINVAL)?,
+ TRANSTAB_LO::from_raw(transtab as u32),
+ );
+ io.write(
+ TRANSTAB_HI::try_at(as_nr).ok_or(EINVAL)?,
+ TRANSTAB_HI::from_raw((transtab >> 32) as u32),
+ );
+
+ let transcfg = as_config.transcfg;
+ io.write(
+ TRANSCFG_LO::try_at(as_nr).ok_or(EINVAL)?,
+ TRANSCFG_LO::from_raw(transcfg as u32),
+ );
+ io.write(
+ TRANSCFG_HI::try_at(as_nr).ok_or(EINVAL)?,
+ TRANSCFG_HI::from_raw((transcfg >> 32) as u32),
+ );
+
+ let memattr = as_config.memattr;
+ io.write(
+ MEMATTR_LO::try_at(as_nr).ok_or(EINVAL)?,
+ MEMATTR_LO::from_raw(memattr as u32),
+ );
+ io.write(
+ MEMATTR_HI::try_at(as_nr).ok_or(EINVAL)?,
+ MEMATTR_HI::from_raw((memattr >> 32) as u32),
+ );
+
+ self.as_send_cmd_and_wait(as_nr, MmuCommand::Update)?;
+
+ Ok(())
+ }
+
+ /// Disables an AS slot and clears its configuration.
+ ///
+ /// Returns an error if the slot is invalid or if register writes/commands fail.
+ fn as_disable(&mut self, as_nr: usize) -> Result {
+ self.validate_as_slot(as_nr)?;
+
+ // Flush AS before disabling
+ self.as_send_cmd_and_wait(as_nr, MmuCommand::FlushMem)?;
+
+ let io = &*self.iomem;
+
+ io.write(
+ TRANSTAB_LO::try_at(as_nr).ok_or(EINVAL)?,
+ TRANSTAB_LO::from_raw(0),
+ );
+ io.write(
+ TRANSTAB_HI::try_at(as_nr).ok_or(EINVAL)?,
+ TRANSTAB_HI::from_raw(0),
+ );
+
+ io.write(
+ MEMATTR_LO::try_at(as_nr).ok_or(EINVAL)?,
+ MEMATTR_LO::from_raw(0),
+ );
+ io.write(
+ MEMATTR_HI::try_at(as_nr).ok_or(EINVAL)?,
+ MEMATTR_HI::from_raw(0),
+ );
+
+ let transcfg = TRANSCFG::zeroed()
+ .with_mode(AddressSpaceMode::Unmapped)
+ .into_raw();
+
+ io.write(
+ TRANSCFG_LO::try_at(as_nr).ok_or(EINVAL)?,
+ TRANSCFG_LO::from_raw(transcfg as u32),
+ );
+ io.write(
+ TRANSCFG_HI::try_at(as_nr).ok_or(EINVAL)?,
+ TRANSCFG_HI::from_raw((transcfg >> 32) as u32),
+ );
+
+ self.as_send_cmd_and_wait(as_nr, MmuCommand::Update)?;
+
+ Ok(())
+ }
+
+ /// Locks a region of the translation tables for an atomic update.
+ ///
+ /// Programs the MMU LOCKADDR register for the given address space and issues
+ /// the lock command. The hardware rounds the requested range up to a
+ /// power-of-two region aligned to its size.
+ ///
+ /// Returns an error if the slot is invalid or if register writes/commands fail.
+ fn as_start_update(&mut self, as_nr: usize, region: &Range<u64>) -> Result {
+ self.validate_as_slot(as_nr)?;
+
+ // The lock operates on full 64-byte cache lines of translation table entries.
+ // Since each translation table entry (TTE) is 8 bytes, a cache line has 8 TTEs.
+ // Since each TTE maps one page, the minimum locked region size will be 8 pages.
+ //
+ // With 4KiB pages (Aarch64_4K mode), the minimum locked region is 32KiB.
+ let lock_region_min_size: u64 = 32 * 1024;
+
+ // Count the number of trailing zero bits (zeros at the right/least-significant
+ // end of the binary representation). For a power-of-two value, this equals the
+ // base-2 exponent (e.g., 32 KiB = 2^15 → 15).
+ let lock_region_min_size_log2 = lock_region_min_size.trailing_zeros() as u8;
+
+ // XOR the first and last addresses to identify which bits differ between them.
+ // The highest set bit in the result determines the exponent of the smallest
+ // power-of-two region that can contain both addresses.
+ //
+ // Example:
+ // addr_xor = 0x1000 ^ 0x2FFF = 0x3FFF
+ // highest set bit in 0x3FFF is bit 13
+ // minimum region size = 2^(13 + 1) = 16 KiB
+ let addr_xor = region.start ^ (region.end - 1);
+ let region_size_log2 = 64 - addr_xor.leading_zeros() as u8;
+
+ let lock_region_log2 = core::cmp::max(region_size_log2, lock_region_min_size_log2);
+
+ // Align the LOCKADDR base address down to the lock region size (1 << lock_region_log2).
+ //
+ // The MMU ignores the low lock_region_log2 bits of LOCKADDR base, so ensure
+ // they are cleared in software to avoid ambiguity.
+ let lockaddr_base = region.start & !((1u64 << lock_region_log2) - 1);
+
+ // The LOCKADDR size field encodes the lock region size as log2(size) - 1,
+ // per the hardware definition. For example, a 32 KiB region is encoded as 14
+ // because log2(32 KiB) = 15.
+ let lockaddr_size = lock_region_log2 - 1;
+
+ let io = &*self.iomem;
+
+ let lockaddr_val = LOCKADDR::zeroed()
+ .try_with_size(lockaddr_size)?
+ .try_with_base(lockaddr_base)?
+ .into_raw();
+
+ io.write(
+ LOCKADDR_LO::try_at(as_nr).ok_or(EINVAL)?,
+ LOCKADDR_LO::from_raw(lockaddr_val as u32),
+ );
+ io.write(
+ LOCKADDR_HI::try_at(as_nr).ok_or(EINVAL)?,
+ LOCKADDR_HI::from_raw((lockaddr_val >> 32) as u32),
+ );
+
+ self.as_send_cmd(as_nr, MmuCommand::Lock)
+ }
+
+ /// Completes an atomic translation table update.
+ ///
+ /// Returns an error if the slot is invalid or if the flush command fails.
+ fn as_end_update(&mut self, as_nr: usize) -> Result {
+ self.validate_as_slot(as_nr)?;
+ self.as_send_cmd_and_wait(as_nr, MmuCommand::FlushPt)?;
+ Ok(())
+ }
+
+ /// Flushes the translation table cache for an AS slot.
+ ///
+ /// Returns an error if the slot is invalid or if the flush command fails.
+ fn as_flush(&mut self, as_nr: usize) -> Result {
+ self.validate_as_slot(as_nr)?;
+ self.as_send_cmd(as_nr, MmuCommand::FlushPt)
+ }
+}
+
+impl<'bound> SlotOperations for AddressSpaceManager<'bound> {
+ /// VM address space data associated with a hardware slot.
+ type SlotData = Arc<VmAsData<'bound>>;
+
+ /// Activates an address space in a hardware slot.
+ fn activate(&mut self, slot_idx: usize, slot_data: &Self::SlotData) -> Result {
+ let as_config = slot_data.as_config()?;
+ self.as_enable(slot_idx, &as_config)
+ }
+
+ /// Evicts an address space from a hardware slot.
+ fn evict(&mut self, slot_idx: usize, _slot_data: &Self::SlotData) -> Result {
+ self.as_flush(slot_idx)?;
+ self.as_disable(slot_idx)?;
+ Ok(())
+ }
+}
+
+impl<'bound> AsSlotManager<'bound> {
+ /// Locks a region for translation table updates if the VM has an active slot.
+ pub(super) fn start_vm_update(&mut self, vm: &VmAsData<'bound>, region: &Range<u64>) -> Result {
+ let seat = vm.as_seat.access(self);
+ match seat.slot() {
+ Some(slot) => {
+ let as_nr = slot as usize;
+ self.as_start_update(as_nr, region)
+ }
+ _ => Ok(()),
+ }
+ }
+
+ /// Completes translation table updates and unlocks the region.
+ pub(super) fn end_vm_update(&mut self, vm: &VmAsData<'bound>) -> Result {
+ let seat = vm.as_seat.access(self);
+ match seat.slot() {
+ Some(slot) => {
+ let as_nr = slot as usize;
+ self.as_end_update(as_nr)
+ }
+ _ => Ok(()),
+ }
+ }
+
+ /// Flushes the translation table cache if the VM has an active slot.
+ pub(super) fn flush_vm(&mut self, vm: &VmAsData<'bound>) -> Result {
+ let seat = vm.as_seat.access(self);
+ match seat.slot() {
+ Some(slot) => {
+ let as_nr = slot as usize;
+ self.as_flush(as_nr)
+ }
+ _ => Ok(()),
+ }
+ }
+
+ /// Activates a VM by assigning it to a hardware slot.
+ pub(super) fn activate_vm(&mut self, vm: ArcBorrow<'_, VmAsData<'bound>>) -> Result {
+ self.activate(&vm.as_seat, vm.into())
+ }
+
+ /// Deactivates a VM by evicting it from its hardware slot.
+ pub(super) fn deactivate_vm(&mut self, vm: &VmAsData<'bound>) -> Result {
+ self.evict(&vm.as_seat)
+ }
+}
diff --git a/drivers/gpu/drm/tyr/regs.rs b/drivers/gpu/drm/tyr/regs.rs
index 831357a8ef87..c4f6b1596b31 100644
--- a/drivers/gpu/drm/tyr/regs.rs
+++ b/drivers/gpu/drm/tyr/regs.rs
@@ -45,6 +45,8 @@ pub(crate) fn read_u64_no_tearing(lo_read: impl Fn() -> u32, hi_read: impl Fn()
}
}
+pub(crate) use mmu_control::mmu_as_control::MAX_AS;
+
/// These registers correspond to the GPU_CONTROL register page.
/// They are involved in GPU configuration and control.
pub(crate) mod gpu_control {
@@ -965,6 +967,8 @@ pub(crate) mod mmu_as_control {
register, //
};
+ use pin_init::Zeroable;
+
/// Maximum number of hardware address space slots.
/// The actual number of slots available is usually lower.
pub(crate) const MAX_AS: usize = 16;
@@ -1158,7 +1162,124 @@ fn from(val: MMU_MEMATTR_STAGE1) -> Self {
pub(crate) MEMATTR_HI(u32)[MAX_AS, stride = STRIDE] @ 0x240c {
31:0 value;
}
+ }
+
+ impl MEMATTR {
+ ///
+ /// In the ARM Architecture Reference Manual, the MAIR encoding for Normal memory
+ /// uses the format `0bxxRW` where:
+ /// - `W` (bit 0) = Write-Allocate policy
+ /// - `R` (bit 1) = Read-Allocate policy
+ /// E.g., `0b0011` would allow both read and write allocation on a cache miss.
+ ///
+ /// ARM MAIR Write-Allocate bit (bit 0 of a cache policy nibble).
+ const ARM_MAIR_WRITE_ALLOCATE: u8 = 0x1;
+ /// ARM MAIR Read-Allocate bit (bit 1 of a cache policy nibble).
+ const ARM_MAIR_READ_ALLOCATE: u8 = 0x2;
+ /// ARM MAIR Write-back bit (bit 2 of a cache policy nibble).
+ const ARM_MAIR_WRITE_BACK: u8 = 0x4;
+ /// Mask for the inner cache policy nibble in MAIR attribute bytes.
+ const ARM_MAIR_INNER_MASK: u8 = 0x0f;
+
+ /// Check if a MAIR attribute byte represents device memory.
+ ///
+ /// Device memory (memory-mapped I/O, registers) cannot be cached and must
+ /// be mapped as GPU `NonCacheable`.
+ fn is_device_memory(mair_attr: u8) -> bool {
+ // In AArch64 MAIR, device memory has bits [1:0] of outer nibble = 0.
+ let outer = mair_attr >> 4;
+ (outer & 0x3) == 0
+ }
+
+ /// Check if normal memory is fully write-back cacheable.
+ ///
+ /// ARM MAIR has two cache policy levels (outer [7:4] and inner [3:0]).
+ /// For memory to be truly write-back, BOTH levels must have the write-back bit set.
+ /// If only one level is write-back, treat it as non-cacheable for GPU purposes.
+ fn is_writeback_cacheable(mair_attr: u8) -> bool {
+ let outer = mair_attr >> 4;
+ let inner = mair_attr & Self::ARM_MAIR_INNER_MASK;
+
+ (outer & Self::ARM_MAIR_WRITE_BACK) != 0 && (inner & Self::ARM_MAIR_WRITE_BACK) != 0
+ }
+
+ // Helper to encode a MEMATTR attribute from its individual fields.
+ fn encode_attribute(
+ alloc_w: bool,
+ alloc_r: bool,
+ alloc_sel: AllocPolicySelect,
+ coherency: Coherency,
+ memory_type: MemoryType,
+ ) -> MMU_MEMATTR_STAGE1 {
+ MMU_MEMATTR_STAGE1::zeroed()
+ .with_alloc_w(alloc_w)
+ .with_alloc_r(alloc_r)
+ .with_alloc_sel(alloc_sel)
+ .with_coherency(coherency)
+ .with_memory_type(memory_type)
+ }
+
+ /// Convert one MAIR attribute byte into a MEMATTR attribute.
+ // TODO: Add a `coherent` parameter like panthor's mair_to_memattr().
+ // For now, assume a non-coherent system and always encode write-back
+ // memory with MidgardInnerDomain coherency.
+ fn attribute_from_mair(mair_attr: u8) -> MMU_MEMATTR_STAGE1 {
+ // Device memory or non-write-back normal memory
+ if Self::is_device_memory(mair_attr) || !Self::is_writeback_cacheable(mair_attr) {
+ return Self::encode_attribute(
+ false,
+ false,
+ AllocPolicySelect::Alloc,
+ Coherency::MidgardInnerDomain,
+ MemoryType::NonCacheable,
+ );
+ }
+
+ // Write-back cacheable normal memory
+ let inner: u8 = mair_attr & Self::ARM_MAIR_INNER_MASK;
+ Self::encode_attribute(
+ (inner & Self::ARM_MAIR_WRITE_ALLOCATE) != 0,
+ (inner & Self::ARM_MAIR_READ_ALLOCATE) != 0,
+ AllocPolicySelect::Alloc,
+ Coherency::MidgardInnerDomain,
+ MemoryType::WriteBack,
+ )
+ }
+
+ /// Write one converted MAIR attribute into a corresponding MEMATTR slot.
+ fn with_encoded_attribute(self, index: usize, attr: MMU_MEMATTR_STAGE1) -> Self {
+ debug_assert!(index < 8);
+
+ let shift = index * 8;
+ let mask = !(0xffu64 << shift);
+ let raw = (self.into_raw() & mask) | ((u64::from(attr.into_raw())) << shift);
+
+ Self::from_raw(raw)
+ }
+ /// Convert an AArch64 MAIR value into the GPU MEMATTR register encoding.
+ ///
+ /// Both MAIR and MEMATTR are 64-bit values with eight 8-bit memory
+ /// attribute entries, but the bits do not map directly. The GPU MEMATTR encoding
+ /// is less detailed than the MAIR encoding, so MAIR is converted to MEMATTR
+ /// conservatively as follows:
+ ///
+ /// 1. Device memory, or Normal Memory that is not write-back cacheable, is encoded
+ /// as GPU `NonCacheable`
+ ///
+ /// 2. Normal memory that is write-back cacheable is encoded as GPU `WriteBack`,
+ /// and the inner allocation hints are preserved.
+ pub(crate) fn from_mair(mair: u64) -> Self {
+ mair.to_le_bytes()
+ .into_iter()
+ .enumerate()
+ .fold(Self::zeroed(), |acc, (i, attr)| {
+ acc.with_encoded_attribute(i, Self::attribute_from_mair(attr))
+ })
+ }
+ }
+
+ register! {
/// Lock region address for each address space.
pub(crate) LOCKADDR(u64)[MAX_AS, stride = STRIDE] @ 0x2410 {
/// Lock region size.
diff --git a/drivers/gpu/drm/tyr/tyr.rs b/drivers/gpu/drm/tyr/tyr.rs
index 7c9a8063b3b9..79045d0135a8 100644
--- a/drivers/gpu/drm/tyr/tyr.rs
+++ b/drivers/gpu/drm/tyr/tyr.rs
@@ -11,6 +11,7 @@
mod file;
mod gem;
mod gpu;
+mod mmu;
mod regs;
mod slot;
--
2.54.0
^ permalink raw reply related [flat|nested] 13+ messages in thread* Re: [PATCH v6 3/7] drm/tyr: add Memory Management Unit (MMU) support
2026-07-09 21:36 ` [PATCH v6 3/7] drm/tyr: add Memory Management Unit (MMU) support Deborah Brouwer
@ 2026-07-10 13:45 ` Alice Ryhl
0 siblings, 0 replies; 13+ messages in thread
From: Alice Ryhl @ 2026-07-10 13:45 UTC (permalink / raw)
To: Deborah Brouwer
Cc: Daniel Almeida, Danilo Krummrich, David Airlie, Simona Vetter,
Benno Lossin, Gary Guo, dri-devel, linux-kernel, rust-for-linux,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
On Thu, Jul 09, 2026 at 02:36:43PM -0700, Deborah Brouwer wrote:
> From: Boris Brezillon <boris.brezillon@collabora.com>
>
> Add Memory Management Unit (MMU) support in Tyr. The MMU module wraps a
> SlotManager instance to allocate MMU address-space slots for use by
> virtual memory (VM) address spaces. The MMU's SlotManager uses an
> AddressSpaceManager to handle the hardware-specific callbacks. For
> example, the AddressSpaceManager activates and evicts VMs from slots by
> writing commands to the MMU registers.
>
> Add an implementation block for the MMU's MEMATTR register to provide
> a method for translating the Memory Attribute Indirection Register (MAIR)
> format from the pagetable configuration to a format understood by the MMU.
>
> Create an mmu instance during probe, it will be used by subsequent patches
> in this series.
>
> Wrap the iomem stored in TyrDrmRegistrationData in an Arc. The iomem
> is stored in the mmu through its AddressSpaceManager. In anticipation
> of the iomem also being stored in the firmware object, set up shared
> ownership of the iomem now.
>
> Update Kconfig to add the new MMU and IOMMU dependencies required
> by this MMU module.
>
> Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
> Co-developed-by: Deborah Brouwer <deborah.brouwer@collabora.com>
> Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
> +//! Memory Management Unit (MMU) module.
> +//!
> +//! The GPU MMU provides a limited number of memory address spaces for use by command streams.
> +//! The MMU translates virtual addresses to physical addresses and manages memory configuration
> +//! and access permissions.
> +//!
> +//! This MMU module is essentially a locked wrapper around a [`SlotManager`] instance.
> +//! The [`SlotManager`] manages the assignment of virtual address spaces to hardware address-space
> +//! (AS) slots. MMU commands such as updates and flushes are carried out by the
> +//! [`AddressSpaceManager`] which actually writes to the MMU registers.
> +#![allow(dead_code)]
This occurs in a few different commits, but I'd make sure to use
expect(dead_code) here so we are sure to remove it when everything is
used.
> +/// Virtual memory (VM) address space data for use in MMU operations.
> +#[pin_data]
> +pub(crate) struct VmAsData<'bound> {
> + /// The address space seat tracks this VM's binding to a hardware address space slot.
> + /// Uses [`LockedBy`] to ensure safe concurrent access to the slot assignment state,
> + /// protected by the [`AsSlotManager`] lock.
> + as_seat: LockedBy<Seat, AsSlotManager<'bound>>,
Why not use the LockedSeat type alias here?
> + let lockaddr_val = LOCKADDR::zeroed()
> + .try_with_size(lockaddr_size)?
> + .try_with_base(lockaddr_base)?
> + .into_raw();
Should this be:
let lockaddr_val = LOCKADDR::zeroed()
.try_with_size(lockaddr_size)?
.try_with_base(lockaddr_base >> 12)?
.into_raw();
or even just
let lockaddr_val = lockaddr_size | lockaddr_base;
?
Alice
^ permalink raw reply [flat|nested] 13+ messages in thread
* [PATCH v6 4/7] drm/tyr: add GPU virtual memory (VM) support
2026-07-09 21:36 [PATCH v6 0/7] drm/tyr: firmware loading and MCU boot support Deborah Brouwer
` (2 preceding siblings ...)
2026-07-09 21:36 ` [PATCH v6 3/7] drm/tyr: add Memory Management Unit (MMU) support Deborah Brouwer
@ 2026-07-09 21:36 ` Deborah Brouwer
2026-07-10 14:15 ` Alice Ryhl
2026-07-10 14:27 ` Alice Ryhl
2026-07-09 21:36 ` [PATCH v6 5/7] drm/tyr: add a kernel buffer object Deborah Brouwer
` (2 subsequent siblings)
6 siblings, 2 replies; 13+ messages in thread
From: Deborah Brouwer @ 2026-07-09 21:36 UTC (permalink / raw)
To: Daniel Almeida, Alice Ryhl, Danilo Krummrich, David Airlie,
Simona Vetter, Benno Lossin, Gary Guo
Cc: dri-devel, linux-kernel, rust-for-linux, Deborah Brouwer,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
From: Boris Brezillon <boris.brezillon@collabora.com>
Add GPU virtual address space management using the DRM GPUVM framework.
Each virtual memory (VM) space is backed by ARM64 LPAE Stage 1 page tables
and can be mapped into hardware address space (AS) slots for GPU execution.
The implementation provides memory isolation and virtual address
allocation. VMs support mapping GEM buffer objects with configurable
protection flags (readonly, noexec, uncached) and handle both 4KB and 2MB
page sizes. A new_dummy_object() helper is provided to create a dummy GEM
object for use as a GPUVM root.
The vm module integrates with the MMU for address space activation and
provides map/unmap/remap operations with page table synchronization.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Co-developed-by: Daniel Almeida <daniel.almeida@collabora.com>
Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>
Co-developed-by: Deborah Brouwer <deborah.brouwer@collabora.com>
Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
---
drivers/gpu/drm/tyr/Kconfig | 1 +
drivers/gpu/drm/tyr/driver.rs | 4 +-
drivers/gpu/drm/tyr/gem.rs | 26 +-
drivers/gpu/drm/tyr/tyr.rs | 1 +
drivers/gpu/drm/tyr/vm.rs | 807 ++++++++++++++++++++++++++++++++++++++++++
5 files changed, 835 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/tyr/Kconfig b/drivers/gpu/drm/tyr/Kconfig
index 61a2fd6f961a..79ea4bb214de 100644
--- a/drivers/gpu/drm/tyr/Kconfig
+++ b/drivers/gpu/drm/tyr/Kconfig
@@ -12,6 +12,7 @@ config DRM_TYR
default n
select IOMMU_IO_PGTABLE_LPAE
select RUST_DRM_GEM_SHMEM_HELPER
+ select RUST_DRM_GPUVM
help
Rust DRM driver for ARM Mali CSF-based GPUs.
diff --git a/drivers/gpu/drm/tyr/driver.rs b/drivers/gpu/drm/tyr/driver.rs
index 74b55a389754..9195c8be5203 100644
--- a/drivers/gpu/drm/tyr/driver.rs
+++ b/drivers/gpu/drm/tyr/driver.rs
@@ -37,7 +37,7 @@
use crate::{
file::TyrDrmFileData,
- gem::BoData,
+ gem::Bo,
gpu,
gpu::GpuInfo,
mmu::Mmu,
@@ -194,7 +194,7 @@ impl drm::Driver for TyrDrmDriver {
type Data = ();
type RegistrationData<'bound> = TyrDrmRegistrationData<'bound>;
type File = TyrDrmFileData;
- type Object = drm::gem::shmem::Object<BoData>;
+ type Object = Bo;
type ParentDevice<Ctx: DeviceContext> = platform::Device<Ctx>;
const INFO: drm::DriverInfo = INFO;
diff --git a/drivers/gpu/drm/tyr/gem.rs b/drivers/gpu/drm/tyr/gem.rs
index 1640a161754b..c28be61a01bb 100644
--- a/drivers/gpu/drm/tyr/gem.rs
+++ b/drivers/gpu/drm/tyr/gem.rs
@@ -5,8 +5,12 @@
//! DRM's GEM subsystem with shmem backing.
use kernel::{
- drm::gem,
- prelude::*, //
+ drm::gem::{
+ self,
+ shmem, //
+ },
+ prelude::*,
+ sync::aref::ARef, //
};
use crate::driver::{
@@ -34,3 +38,21 @@ fn new(_dev: &TyrDrmDevice, _size: usize, args: BoCreateArgs) -> impl PinInit<Se
try_pin_init!(Self { flags: args.flags })
}
}
+
+/// Type alias for Tyr GEM buffer objects.
+pub(crate) type Bo = gem::shmem::Object<BoData>;
+
+/// Creates a dummy GEM object to serve as the root of a GPUVM.
+pub(crate) fn new_dummy_object(ddev: &TyrDrmDevice) -> Result<ARef<Bo>> {
+ let bo = Bo::new(
+ ddev,
+ 4096,
+ shmem::ObjectConfig {
+ map_wc: true,
+ parent_resv_obj: None,
+ },
+ BoCreateArgs { flags: 0 },
+ )?;
+
+ Ok(bo)
+}
diff --git a/drivers/gpu/drm/tyr/tyr.rs b/drivers/gpu/drm/tyr/tyr.rs
index 79045d0135a8..92f6885cdaae 100644
--- a/drivers/gpu/drm/tyr/tyr.rs
+++ b/drivers/gpu/drm/tyr/tyr.rs
@@ -14,6 +14,7 @@
mod mmu;
mod regs;
mod slot;
+mod vm;
kernel::module_platform_driver! {
type: TyrPlatformDriver,
diff --git a/drivers/gpu/drm/tyr/vm.rs b/drivers/gpu/drm/tyr/vm.rs
new file mode 100644
index 000000000000..57929c2d3e94
--- /dev/null
+++ b/drivers/gpu/drm/tyr/vm.rs
@@ -0,0 +1,807 @@
+// SPDX-License-Identifier: GPL-2.0 or MIT
+
+//! GPU virtual memory management using the DRM GPUVM framework.
+//!
+//! This module manages GPU virtual address spaces, providing memory isolation and
+//! the illusion of owning the entire virtual address (VA) range, similar to CPU virtual memory.
+//! Each virtual memory (VM) area is backed by ARM64 LPAE Stage 1 page tables and can be
+//! mapped into hardware address space (AS) slots for GPU execution.
+#![allow(dead_code)]
+
+use core::marker::PhantomData;
+use core::ops::Range;
+
+use kernel::{
+ c_str,
+ device::{
+ Bound,
+ Device, //
+ },
+ drm::{
+ gpuvm::{
+ DriverGpuVm,
+ GpuVaAlloc,
+ GpuVm,
+ GpuVmBo,
+ OpMap,
+ OpMapRequest,
+ OpMapped,
+ OpRemap,
+ OpRemapped,
+ OpUnmap,
+ OpUnmapped,
+ UniqueRefGpuVm, //
+ }, //
+ },
+ impl_flags,
+ iommu::pgtable::{
+ prot,
+ IoPageTable,
+ ARM64LPAES1, //
+ },
+ new_mutex,
+ prelude::*,
+ sizes::{
+ SZ_1G,
+ SZ_2M,
+ SZ_4K, //
+ },
+ sync::{
+ aref::ARef,
+ Arc,
+ ArcBorrow,
+ Mutex, //
+ },
+ uapi, //
+};
+
+use crate::{
+ driver::{
+ TyrDrmDevice,
+ TyrDrmDriver, //
+ },
+ gem,
+ gem::Bo,
+ gpu::GpuInfo,
+ mmu::{
+ address_space::VmAsData,
+ Mmu, //
+ },
+ regs::gpu_control::MMU_FEATURES,
+};
+
+impl_flags!(
+ /// Flags controlling virtual memory mapping behavior.
+ ///
+ /// These flags control access permissions and caching behavior for GPU virtual
+ /// memory mappings.
+ #[derive(Debug, Clone, Default, Copy, PartialEq, Eq)]
+ pub(crate) struct VmMapFlags(u32);
+
+ /// Individual flags that can be combined in [`VmMapFlags`].
+ #[derive(Debug, Clone, Copy, PartialEq, Eq)]
+ pub(crate) enum VmFlag {
+ /// Map as read-only.
+ Readonly = uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_READONLY as u32,
+ /// Map as non-executable.
+ Noexec = uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_NOEXEC as u32,
+ /// Map as uncached.
+ Uncached = uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_UNCACHED as u32,
+ }
+);
+
+impl VmMapFlags {
+ /// Convert the flags to `pgtable::prot`.
+ fn to_prot(self) -> u32 {
+ let mut prot = 0;
+
+ if self.contains(VmFlag::Readonly) {
+ prot |= prot::READ;
+ } else {
+ prot |= prot::READ | prot::WRITE;
+ }
+
+ if self.contains(VmFlag::Noexec) {
+ prot |= prot::NOEXEC;
+ }
+
+ if !self.contains(VmFlag::Uncached) {
+ prot |= prot::CACHE;
+ }
+
+ prot
+ }
+}
+
+impl core::fmt::Display for VmMapFlags {
+ fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
+ let mut first = true;
+
+ if self.contains(VmFlag::Readonly) {
+ write!(f, "READONLY")?;
+ first = false;
+ }
+ if self.contains(VmFlag::Noexec) {
+ if !first {
+ write!(f, " | ")?;
+ }
+ write!(f, "NOEXEC")?;
+ first = false;
+ }
+
+ if self.contains(VmFlag::Uncached) {
+ if !first {
+ write!(f, " | ")?;
+ }
+ write!(f, "UNCACHED")?;
+ }
+
+ Ok(())
+ }
+}
+
+impl TryFrom<u32> for VmMapFlags {
+ type Error = Error;
+
+ fn try_from(value: u32) -> core::result::Result<Self, Self::Error> {
+ let valid = (kernel::uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_READONLY
+ | kernel::uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_NOEXEC
+ | kernel::uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_UNCACHED)
+ as u32;
+
+ if value & !valid != 0 {
+ pr_err!("Invalid VM map flags: {:#x}\n", value);
+ return Err(EINVAL);
+ }
+ Ok(Self(value))
+ }
+}
+
+/// Arguments for a virtual memory map operation.
+struct VmMapArgs<'bound> {
+ /// Access permissions and caching behavior for the mapping.
+ flags: VmMapFlags,
+ /// GEM buffer object registered with the GPUVM framework.
+ vm_bo: ARef<GpuVmBo<GpuVmData<'bound>>>,
+ /// Offset in bytes from the start of the buffer object.
+ bo_offset: u64,
+}
+
+/// Type of virtual memory operation.
+enum VmOpType<'bound> {
+ /// Map a GEM buffer object into the virtual address space.
+ Map(VmMapArgs<'bound>),
+ /// Unmap a region from the virtual address space.
+ Unmap,
+}
+
+/// Preallocated resources needed to execute a VM operation.
+///
+/// VM operations may require allocating new GPUVA objects to track mappings.
+/// To avoid allocation failures during the operation, preallocate the
+/// maximum number of GPUVAs that might be needed.
+struct VmOpResources<'bound> {
+ /// Preallocated GPUVA objects for remap operations.
+ ///
+ /// Partial unmap requests or map requests overlapping existing mappings
+ /// will trigger a remap call, which needs to register up to three VA
+ /// objects (one for the new mapping, and two for the previous and next
+ /// mappings).
+ preallocated_gpuvas: [Option<GpuVaAlloc<GpuVmData<'bound>>>; 3],
+}
+
+/// Request to execute a virtual memory operation.
+struct VmOpRequest<'bound> {
+ /// Request type.
+ op_type: VmOpType<'bound>,
+
+ /// Region of the virtual address space covered by this request.
+ region: Range<u64>,
+}
+
+/// Arguments for a page table map operation.
+struct PtMapArgs {
+ /// Memory protection flags describing allowed accesses for this mapping.
+ ///
+ /// This is directly derived from [`VmMapFlags`] via [`VmMapFlags::to_prot`].
+ prot: u32,
+}
+
+/// Type of page table operation.
+enum PtOpType {
+ /// Map pages into the page table.
+ Map(PtMapArgs),
+ /// Unmap pages from the page table.
+ Unmap,
+}
+
+/// Context for updating the GPU page table.
+///
+/// This context is created when beginning a page table update operation and
+/// automatically flushes changes when dropped. It ensures that the
+/// Memory Management Unit (MMU) state is properly managed and Translation
+/// Lookaside Buffer (TLB) entries are flushed.
+pub(crate) struct PtUpdateContext<'ctx, 'bound> {
+ /// Device used for DMA-mapping GEM shmem SG tables.
+ dev: &'ctx Device<Bound>,
+
+ /// Page table.
+ pt: &'ctx IoPageTable<'bound, ARM64LPAES1>,
+
+ /// MMU manager.
+ mmu: &'ctx Mmu<'bound>,
+
+ /// Reference to the address space data to pass to the MMU functions.
+ as_data: &'ctx VmAsData<'bound>,
+
+ /// Region of the virtual address space covered by this request.
+ region: Range<u64>,
+
+ /// Operation type.
+ op_type: PtOpType,
+
+ /// Preallocated resources that can be used when executing the request.
+ resources: &'ctx mut VmOpResources<'bound>,
+}
+
+impl<'ctx, 'bound> PtUpdateContext<'ctx, 'bound> {
+ /// Creates a new page table update context.
+ ///
+ /// This prepares the MMU for a page table update.
+ /// The context will automatically flush the TLB and
+ /// complete the update when dropped.
+ fn new(
+ dev: &'ctx Device<Bound>,
+ pt: &'ctx IoPageTable<'bound, ARM64LPAES1>,
+ mmu: &'ctx Mmu<'bound>,
+ as_data: &'ctx VmAsData<'bound>,
+ region: Range<u64>,
+ op_type: PtOpType,
+ resources: &'ctx mut VmOpResources<'bound>,
+ ) -> Result<PtUpdateContext<'ctx, 'bound>> {
+ mmu.start_vm_update(as_data, ®ion)?;
+
+ Ok(Self {
+ dev,
+ pt,
+ mmu,
+ as_data,
+ region,
+ op_type,
+ resources,
+ })
+ }
+
+ /// Finds one of our pre-allocated VAs.
+ ///
+ /// It is a logic error to call this more than three times for a given
+ /// PtUpdateContext.
+ fn preallocated_gpuva(&mut self) -> Result<GpuVaAlloc<GpuVmData<'bound>>> {
+ self.resources
+ .preallocated_gpuvas
+ .iter_mut()
+ .find_map(|f| f.take())
+ .ok_or(EINVAL)
+ }
+}
+
+impl Drop for PtUpdateContext<'_, '_> {
+ fn drop(&mut self) {
+ if let Err(e) = self.mmu.end_vm_update(self.as_data) {
+ pr_err!("Failed to end VM update {:?}\n", e);
+ }
+
+ if let Err(e) = self.mmu.flush_vm(self.as_data) {
+ pr_err!("Failed to flush VM {:?}\n", e);
+ }
+ }
+}
+
+/// Driver implementation for the GPUVM framework.
+///
+/// Implements [`DriverGpuVm`] to provide VM operation callbacks (map, unmap, remap)
+/// and associated types for buffer objects, virtual addresses, and contexts.
+pub(crate) struct GpuVmData<'bound> {
+ _phantom: PhantomData<&'bound ()>,
+}
+
+/// GPU virtual address space.
+///
+/// Each VM can be mapped into a hardware address space slot.
+#[pin_data]
+pub(crate) struct Vm<'bound> {
+ /// Data referenced by an AS when the VM is active
+ as_data: Arc<VmAsData<'bound>>,
+ /// MMU manager.
+ mmu: Arc<Mmu<'bound>>,
+ /// Parent device used for DMA mapping and page-table operations.
+ dev: &'bound Device<Bound>,
+ /// DRM GPUVM core for managing virtual address space.
+ #[pin]
+ gpuvm_unique: Mutex<UniqueRefGpuVm<GpuVmData<'bound>>>,
+ /// Non-core part of the GPUVM. Can be used for stuff that doesn't modify the
+ /// internal mapping tree, like GpuVm::obtain()
+ gpuvm: ARef<GpuVm<GpuVmData<'bound>>>,
+ /// VA range for this VM.
+ va_range: Range<u64>,
+}
+
+impl<'bound> Vm<'bound> {
+ /// Creates a new GPU virtual address space.
+ ///
+ /// The VM is initialized with a page table configured according to the GPU's
+ /// address translation capabilities and registered with the GPUVM framework.
+ pub(crate) fn new(
+ dev: &'bound Device<Bound>,
+ ddev: &TyrDrmDevice,
+ mmu: ArcBorrow<'_, Mmu<'bound>>,
+ gpu_info: &GpuInfo,
+ ) -> Result<Arc<Vm<'bound>>> {
+ let mmu_features = MMU_FEATURES::from_raw(gpu_info.mmu_features);
+ let va_bits = mmu_features.va_bits().get();
+ let pa_bits = mmu_features.pa_bits().get();
+
+ let range = 0..(1u64 << va_bits);
+ let reserve_range = 0..0u64;
+
+ // dummy_obj is used to initialize the GPUVM tree.
+ let dummy_obj = gem::new_dummy_object(ddev).inspect_err(|e| {
+ pr_err!("Failed to create dummy GEM object: {:?}\n", e);
+ })?;
+
+ let gpuvm_unique = kernel::drm::gpuvm::GpuVm::new::<Error, _>(
+ c_str!("Tyr::GpuVm"),
+ ddev,
+ &*dummy_obj,
+ range.clone(),
+ reserve_range,
+ GpuVmData::<'bound> {
+ _phantom: PhantomData::<&()>,
+ },
+ )
+ .inspect_err(|e| {
+ pr_err!("Failed to create GpuVm: {:?}\n", e);
+ })?;
+ let gpuvm = ARef::from(&*gpuvm_unique);
+
+ let as_data = Arc::pin_init(VmAsData::new(&mmu, dev, va_bits, pa_bits), GFP_KERNEL)?;
+
+ let vm = Arc::pin_init(
+ pin_init!(Self{
+ as_data,
+ dev,
+ mmu: mmu.into(),
+ gpuvm,
+ gpuvm_unique <- new_mutex!(gpuvm_unique),
+ va_range: range,
+ }),
+ GFP_KERNEL,
+ )?;
+
+ Ok(vm)
+ }
+
+ /// Activate the VM in a hardware address space slot.
+ pub(crate) fn activate(&self) -> Result {
+ self.mmu
+ .activate_vm(self.as_data.as_arc_borrow())
+ .inspect_err(|e| {
+ pr_err!("Failed to activate VM: {:?}\n", e);
+ })
+ }
+
+ /// Deactivate the VM by evicting it from its address space slot.
+ fn deactivate(&self) -> Result {
+ self.mmu.deactivate_vm(&self.as_data).inspect_err(|e| {
+ pr_err!("Failed to deactivate VM: {:?}\n", e);
+ })
+ }
+
+ /// Kills the VM by deactivating it and unmapping all regions.
+ pub(crate) fn kill(&self) {
+ // TODO: Turn the VM into a state where it can't be used.
+ let _ = self.deactivate().inspect_err(|e| {
+ pr_err!("Failed to deactivate VM: {:?}\n", e);
+ });
+ let _ = self
+ .unmap_range(self.va_range.start, self.va_range.end - self.va_range.start)
+ .inspect_err(|e| {
+ pr_err!("Failed to unmap range during deactivate: {:?}\n", e);
+ });
+ }
+
+ /// Executes a virtual memory operation.
+ ///
+ /// This handles both map and unmap operations by coordinating between the
+ /// GPUVM framework and the hardware page table.
+ fn exec_op<'a>(
+ &self,
+ gpuvm_unique: &mut UniqueRefGpuVm<GpuVmData<'bound>>,
+ req: VmOpRequest<'bound>,
+ resources: &'a mut VmOpResources<'bound>,
+ ) -> Result {
+ let pt = &self.as_data.page_table;
+
+ match req.op_type {
+ VmOpType::Map(args) => {
+ let mut pt_upd = PtUpdateContext::new(
+ self.dev,
+ pt,
+ &self.mmu,
+ &self.as_data,
+ req.region,
+ PtOpType::Map(PtMapArgs {
+ prot: args.flags.to_prot(),
+ }),
+ resources,
+ )?;
+
+ gpuvm_unique.sm_map(OpMapRequest {
+ addr: pt_upd.region.start,
+ range: pt_upd.region.end - pt_upd.region.start,
+ gem_offset: args.bo_offset,
+ vm_bo: &args.vm_bo,
+ context: &mut pt_upd,
+ })
+ //PtUpdateContext drops here flushing the page table
+ }
+ VmOpType::Unmap => {
+ let mut pt_upd = PtUpdateContext::new(
+ self.dev,
+ pt,
+ &self.mmu,
+ &self.as_data,
+ req.region,
+ PtOpType::Unmap,
+ resources,
+ )?;
+
+ gpuvm_unique.sm_unmap(
+ pt_upd.region.start,
+ pt_upd.region.end - pt_upd.region.start,
+ &mut pt_upd,
+ )
+ //PtUpdateContext drops here flushing the page table
+ }
+ }
+ }
+
+ /// Maps a GEM buffer object range into the VM at the specified virtual address.
+ ///
+ /// This creates a mapping from GPU virtual address `va` to the physical pages
+ /// backing the GEM object, starting at `bo_offset` bytes into the object and
+ /// spanning `size` bytes. The mapping respects the access permissions and
+ /// caching behavior specified in `flags`.
+ pub(crate) fn map_bo_range(
+ &self,
+ bo: &Bo,
+ bo_offset: u64,
+ size: u64,
+ va: u64,
+ flags: VmMapFlags,
+ ) -> Result {
+ let req = VmOpRequest {
+ op_type: VmOpType::Map(VmMapArgs {
+ vm_bo: self.gpuvm.obtain(bo, ())?,
+ flags,
+ bo_offset,
+ }),
+ region: va..(va + size),
+ };
+ let mut resources = VmOpResources {
+ preallocated_gpuvas: [
+ Some(GpuVaAlloc::<GpuVmData<'bound>>::new(GFP_KERNEL)?),
+ Some(GpuVaAlloc::<GpuVmData<'bound>>::new(GFP_KERNEL)?),
+ Some(GpuVaAlloc::<GpuVmData<'bound>>::new(GFP_KERNEL)?),
+ ],
+ };
+ let mut gpuvm_unique = self.gpuvm_unique.lock();
+
+ self.exec_op(gpuvm_unique.as_mut().get_mut(), req, &mut resources)?;
+
+ // We flush the defer cleanup list now. Things will be different in
+ // the asynchronous VM_BIND path, where we want the cleanup to
+ // happen outside the DMA signalling path.
+ self.gpuvm.deferred_cleanup();
+ Ok(())
+ }
+
+ /// Unmaps a virtual address range from the VM.
+ ///
+ /// This removes any existing mappings in the specified range, freeing the
+ /// virtual address space for reuse.
+ pub(crate) fn unmap_range(&self, va: u64, size: u64) -> Result {
+ let req = VmOpRequest {
+ op_type: VmOpType::Unmap,
+ region: va..(va + size),
+ };
+ let mut resources = VmOpResources {
+ preallocated_gpuvas: [
+ Some(GpuVaAlloc::<GpuVmData<'bound>>::new(GFP_KERNEL)?),
+ Some(GpuVaAlloc::<GpuVmData<'bound>>::new(GFP_KERNEL)?),
+ None,
+ ],
+ };
+ let mut gpuvm_unique = self.gpuvm_unique.lock();
+
+ self.exec_op(gpuvm_unique.as_mut().get_mut(), req, &mut resources)?;
+
+ // We flush the defer cleanup list now. Things will be different in
+ // the asynchronous VM_BIND path, where we want the cleanup to
+ // happen outside the DMA signalling path.
+ self.gpuvm.deferred_cleanup();
+ Ok(())
+ }
+}
+
+impl<'bound> DriverGpuVm for GpuVmData<'bound> {
+ type Driver = TyrDrmDriver;
+ type Object = Bo;
+ type VmBoData = ();
+ type VaData = ();
+ type SmContext<'ctx>
+ = PtUpdateContext<'ctx, 'bound>
+ where
+ Self: 'ctx;
+
+ /// Indicates that a new mapping should be created.
+ fn sm_step_map<'op>(
+ &mut self,
+ op: OpMap<'op, Self>,
+ context: &mut Self::SmContext<'_>,
+ ) -> Result<OpMapped<'op, Self>, Error> {
+ let start_iova = op.addr();
+ let mut iova = start_iova;
+ let mut bytes_left_to_map = op.length();
+ let mut gem_offset = op.gem_offset();
+ let sgt = op.obj().sg_table(context.dev).inspect_err(|e| {
+ pr_err!("Failed to get sg_table: {:?}\n", e);
+ })?;
+ let prot = match &context.op_type {
+ PtOpType::Map(args) => args.prot,
+ _ => {
+ return Err(EINVAL);
+ }
+ };
+
+ for sgt_entry in sgt.iter() {
+ let mut paddr = sgt_entry.dma_address();
+ let mut sgt_entry_length: u64 = sgt_entry.dma_len();
+
+ if bytes_left_to_map == 0 {
+ break;
+ }
+
+ if gem_offset > 0 {
+ // Skip the entire SGT entry if the gem_offset exceeds its length
+ let skip = sgt_entry_length.min(gem_offset);
+ paddr += skip;
+ sgt_entry_length -= skip;
+ gem_offset -= skip;
+ }
+
+ if sgt_entry_length == 0 {
+ continue;
+ }
+
+ if gem_offset != 0 {
+ pr_err!("Invalid gem_offset {} in page table mapping.\n", gem_offset);
+ return Err(EINVAL);
+ }
+ let len = sgt_entry_length.min(bytes_left_to_map);
+
+ let segment_mapped = match pt_map(context.pt, iova, paddr, len, prot) {
+ Ok(segment_mapped) => segment_mapped,
+ Err(e) => {
+ // clean up any successful mappings from previous SGT entries.
+ let total_mapped = iova - start_iova;
+ if total_mapped > 0 {
+ pt_unmap(context.pt, start_iova..(start_iova + total_mapped)).ok();
+ }
+ return Err(e);
+ }
+ };
+
+ // Since there could be a partial mapping, only advance by the actual amount mapped
+ bytes_left_to_map -= segment_mapped;
+ iova += segment_mapped;
+ }
+
+ let gpuva = context.preallocated_gpuva()?;
+ let op = op.insert(gpuva, pin_init::init_zeroed());
+
+ Ok(op)
+ }
+
+ /// Indicates that an existing mapping should be removed.
+ fn sm_step_unmap<'op>(
+ &mut self,
+ op: OpUnmap<'op, Self>,
+ context: &mut Self::SmContext<'_>,
+ ) -> Result<OpUnmapped<'op, Self>, Error> {
+ let start_iova = op.va().addr();
+ let length = op.va().length();
+
+ let region = start_iova..(start_iova + length);
+ pt_unmap(context.pt, region.clone()).inspect_err(|e| {
+ pr_err!(
+ "Failed to unmap region {:#x}..{:#x}: {:?}\n",
+ region.start,
+ region.end,
+ e
+ );
+ })?;
+
+ let (op_unmapped, _va_removed) = op.remove();
+
+ Ok(op_unmapped)
+ }
+
+ /// Indicates that an existing mapping should be split up.
+ fn sm_step_remap<'op>(
+ &mut self,
+ op: OpRemap<'op, Self>,
+ context: &mut Self::SmContext<'_>,
+ ) -> Result<OpRemapped<'op, Self>, Error> {
+ let unmap_start = if let Some(prev) = op.prev() {
+ prev.addr() + prev.length()
+ } else {
+ op.va_to_unmap().addr()
+ };
+
+ let unmap_end = if let Some(next) = op.next() {
+ next.addr()
+ } else {
+ op.va_to_unmap().addr() + op.va_to_unmap().length()
+ };
+
+ let unmap_length = unmap_end - unmap_start;
+
+ if unmap_length > 0 {
+ let region = unmap_start..(unmap_start + unmap_length);
+ pt_unmap(context.pt, region.clone()).inspect_err(|e| {
+ pr_err!(
+ "Failed to unmap remap region {:#x}..{:#x}: {:?}\n",
+ region.start,
+ region.end,
+ e
+ );
+ })?;
+ }
+
+ let prev_va = context.preallocated_gpuva()?;
+ let next_va = context.preallocated_gpuva()?;
+
+ let (op_remapped, _remap_ret) = op.remap(
+ [prev_va, next_va],
+ pin_init::init_zeroed(),
+ pin_init::init_zeroed(),
+ );
+
+ Ok(op_remapped)
+ }
+}
+
+/// This function selects the largest supported block size (currently 4KB or 2MB)
+/// that can be used for a mapping at the given address and size, respecting alignment constraints.
+///
+/// We can map multiple pages at once but we can't exceed the size of the
+// table entry itself. So, if mapping 4KB pages, figure out how many pages
+// can be mapped before we hit the 2MB boundary. Or, if mapping 2MB pages,
+// figure out how many pages can be mapped before hitting the 1GB boundary
+// Returns the page size (4KB or 2MB) and the number of pages that can be mapped at that size.
+fn get_pgsize(addr: u64, size: u64) -> (u64, u64) {
+ // Get the distance to the next boundary of 2MB block
+ let blk_offset_2m = addr.wrapping_neg() % (SZ_2M as u64);
+
+ // Use 4K blocks if the address is not 2MB aligned, or we have less than 2MB to map
+ if blk_offset_2m != 0 || size < SZ_2M as u64 {
+ let pgcount = if blk_offset_2m == 0 {
+ size / SZ_4K as u64
+ } else {
+ blk_offset_2m.min(size) / SZ_4K as u64
+ };
+ return (SZ_4K as u64, pgcount);
+ }
+
+ let blk_offset_1g = addr.wrapping_neg() % (SZ_1G as u64);
+ let blk_offset = if blk_offset_1g == 0 {
+ SZ_1G as u64
+ } else {
+ blk_offset_1g
+ };
+ let pgcount = blk_offset.min(size) / SZ_2M as u64;
+
+ (SZ_2M as u64, pgcount)
+}
+
+/// Maps a physical address range into the page table at the specified virtual address.
+///
+/// This function maps `len` bytes of physical memory starting at `paddr` to the
+/// virtual address `iova`, using the protection flags specified in `prot`. It
+/// automatically selects optimal page sizes to minimize page table overhead.
+///
+/// If the mapping fails partway through, all successfully mapped pages are
+/// unmapped before returning an error.
+///
+/// Returns the number of bytes successfully mapped.
+fn pt_map(
+ pt: &IoPageTable<'_, ARM64LPAES1>,
+ iova: u64,
+ paddr: u64,
+ len: u64,
+ prot: u32,
+) -> Result<u64> {
+ let mut segment_mapped = 0u64;
+ while segment_mapped < len {
+ let remaining = len - segment_mapped;
+ let curr_iova = iova + segment_mapped;
+ let curr_paddr = paddr + segment_mapped;
+
+ let (pgsize, pgcount) = get_pgsize(curr_iova | curr_paddr, remaining);
+
+ // SAFETY: Exclusive access to the page table is ensured because
+ // the pt reference comes from PtUpdateContext, which is created
+ // during a VM update operation, ensuring the driver does not concurrently
+ // modify the page table.
+ let (mapped, result) = unsafe {
+ pt.map_pages(
+ curr_iova as usize,
+ (curr_paddr as usize).try_into().unwrap(),
+ pgsize as usize,
+ pgcount as usize,
+ prot,
+ GFP_KERNEL,
+ )
+ };
+
+ if let Err(e) = result {
+ pr_err!("pt.map_pages failed at iova {:#x}: {:?}\n", curr_iova, e);
+ if segment_mapped > 0 {
+ pt_unmap(pt, iova..(iova + segment_mapped)).ok();
+ }
+ return Err(e);
+ }
+
+ if mapped == 0 {
+ pr_err!("Failed to map any pages at iova {:#x}\n", curr_iova);
+ if segment_mapped > 0 {
+ pt_unmap(pt, iova..(iova + segment_mapped)).ok();
+ }
+ return Err(ENOMEM);
+ }
+
+ segment_mapped += mapped as u64;
+ }
+
+ Ok(segment_mapped)
+}
+
+/// Unmaps a virtual address range from the page table.
+///
+/// This function removes all page table entries in the specified range,
+/// automatically handling different page sizes that may be present.
+fn pt_unmap(pt: &IoPageTable<'_, ARM64LPAES1>, range: Range<u64>) -> Result {
+ let mut iova = range.start;
+ let mut bytes_left_to_unmap = range.end - range.start;
+
+ while bytes_left_to_unmap > 0 {
+ let (pgsize, pgcount) = get_pgsize(iova, bytes_left_to_unmap);
+
+ // SAFETY: Exclusive access to the page table is ensured because
+ // the pt reference comes from PtUpdateContext, which was
+ // created while holding &mut Vm, preventing any other access to the
+ // page table for the duration of this operation.
+ let unmapped = unsafe { pt.unmap_pages(iova as usize, pgsize as usize, pgcount as usize) };
+
+ if unmapped == 0 {
+ pr_err!("Failed to unmap any bytes at iova {:#x}\n", iova);
+ return Err(EINVAL);
+ }
+
+ bytes_left_to_unmap -= unmapped as u64;
+ iova += unmapped as u64;
+ }
+
+ Ok(())
+}
--
2.54.0
^ permalink raw reply related [flat|nested] 13+ messages in thread* Re: [PATCH v6 4/7] drm/tyr: add GPU virtual memory (VM) support
2026-07-09 21:36 ` [PATCH v6 4/7] drm/tyr: add GPU virtual memory (VM) support Deborah Brouwer
@ 2026-07-10 14:15 ` Alice Ryhl
2026-07-14 3:12 ` Deborah Brouwer
2026-07-10 14:27 ` Alice Ryhl
1 sibling, 1 reply; 13+ messages in thread
From: Alice Ryhl @ 2026-07-10 14:15 UTC (permalink / raw)
To: Deborah Brouwer
Cc: Daniel Almeida, Danilo Krummrich, David Airlie, Simona Vetter,
Benno Lossin, Gary Guo, dri-devel, linux-kernel, rust-for-linux,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
On Thu, Jul 09, 2026 at 02:36:44PM -0700, Deborah Brouwer wrote:
> From: Boris Brezillon <boris.brezillon@collabora.com>
>
> Add GPU virtual address space management using the DRM GPUVM framework.
> Each virtual memory (VM) space is backed by ARM64 LPAE Stage 1 page tables
> and can be mapped into hardware address space (AS) slots for GPU execution.
>
> The implementation provides memory isolation and virtual address
> allocation. VMs support mapping GEM buffer objects with configurable
> protection flags (readonly, noexec, uncached) and handle both 4KB and 2MB
> page sizes. A new_dummy_object() helper is provided to create a dummy GEM
> object for use as a GPUVM root.
>
> The vm module integrates with the MMU for address space activation and
> provides map/unmap/remap operations with page table synchronization.
>
> Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
> Co-developed-by: Daniel Almeida <daniel.almeida@collabora.com>
> Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>
> Co-developed-by: Deborah Brouwer <deborah.brouwer@collabora.com>
> Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
> +impl core::fmt::Display for VmMapFlags {
Use kernel::fmt instead of core::fmt.
> +impl TryFrom<u32> for VmMapFlags {
> + type Error = Error;
> +
> + fn try_from(value: u32) -> core::result::Result<Self, Self::Error> {
Simplifies to Result<Self, Self::Error> without full path.
> + let valid = (kernel::uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_READONLY
> + | kernel::uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_NOEXEC
> + | kernel::uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_UNCACHED)
> + as u32;
Simplifies to
let valid = Self::Readonly as u32 | Self::Noexec as u32 | Self::Uncached as u32;
> +impl Drop for PtUpdateContext<'_, '_> {
> + fn drop(&mut self) {
> + if let Err(e) = self.mmu.end_vm_update(self.as_data) {
> + pr_err!("Failed to end VM update {:?}\n", e);
> + }
> +
> + if let Err(e) = self.mmu.flush_vm(self.as_data) {
> + pr_err!("Failed to flush VM {:?}\n", e);
> + }
I assume it's not possible for this to run during a VM eviction, right?
> + let gpuvm_unique = kernel::drm::gpuvm::GpuVm::new::<Error, _>(
> + c_str!("Tyr::GpuVm"),
This can just be c"Tyr::GpuVm". The macro is only needed in macro
contexts where this rewrite is not possible e.g. because the macro's
caller provides a non-c str literal.
> + let mut gpuvm_unique = self.gpuvm_unique.lock();
> +
> + self.exec_op(gpuvm_unique.as_mut().get_mut(), req, &mut resources)?;
> +
> + // We flush the defer cleanup list now. Things will be different in
> + // the asynchronous VM_BIND path, where we want the cleanup to
> + // happen outside the DMA signalling path.
> + self.gpuvm.deferred_cleanup();
You should probably release the gpuvm_unique mutex before invoking
deferred_cleanup().
(Multiple occurances in this patch.)
> + if gem_offset > 0 {
> + // Skip the entire SGT entry if the gem_offset exceeds its length
> + let skip = sgt_entry_length.min(gem_offset);
Nit: this may just be my idiosyncrasy, but I think this is more readable as:
usize::min(sgt_entry_length, gem_offset)
(Multiple occurances in this file.)
> + let segment_mapped = match pt_map(context.pt, iova, paddr, len, prot) {
> + Ok(segment_mapped) => segment_mapped,
> + Err(e) => {
> + // clean up any successful mappings from previous SGT entries.
> + let total_mapped = iova - start_iova;
> + if total_mapped > 0 {
> + pt_unmap(context.pt, start_iova..(start_iova + total_mapped)).ok();
Nit: the idiomatic way to ignore errors is to write:
let _ = pt_unmap(context.pt, start_iova..(start_iova + total_mapped));
instead of converting the Result into an Option with .ok() and abusing
the fact that Option is not #[must_use].
> +/// This function selects the largest supported block size (currently 4KB or 2MB)
> +/// that can be used for a mapping at the given address and size, respecting alignment constraints.
> +///
> +/// We can map multiple pages at once but we can't exceed the size of the
> +// table entry itself. So, if mapping 4KB pages, figure out how many pages
> +// can be mapped before we hit the 2MB boundary. Or, if mapping 2MB pages,
> +// figure out how many pages can be mapped before hitting the 1GB boundary
> +// Returns the page size (4KB or 2MB) and the number of pages that can be mapped at that size.
> +fn get_pgsize(addr: u64, size: u64) -> (u64, u64) {
Some of these comments are missing a / character.
> + // SAFETY: Exclusive access to the page table is ensured because
> + // the pt reference comes from PtUpdateContext, which is created
> + // during a VM update operation, ensuring the driver does not concurrently
> + // modify the page table.
> + let (mapped, result) = unsafe {
> + pt.map_pages(
> + curr_iova as usize,
> + (curr_paddr as usize).try_into().unwrap(),
> + pgsize as usize,
> + pgcount as usize,
> + prot,
> + GFP_KERNEL,
> + )
> + };
It would be ideal if this SAFETY comment explains why this safety
requirement is satisfied:
This page table must not contain any mapping that overlaps with
the mapping created by this call.
(It's because GPUVM tells you to unmap pages before it tells you to map
them.)
> +/// Unmaps a virtual address range from the page table.
> +///
> +/// This function removes all page table entries in the specified range,
> +/// automatically handling different page sizes that may be present.
> +fn pt_unmap(pt: &IoPageTable<'_, ARM64LPAES1>, range: Range<u64>) -> Result {
> + let mut iova = range.start;
> + let mut bytes_left_to_unmap = range.end - range.start;
> +
> + while bytes_left_to_unmap > 0 {
> + let (pgsize, pgcount) = get_pgsize(iova, bytes_left_to_unmap);
> +
> + // SAFETY: Exclusive access to the page table is ensured because
> + // the pt reference comes from PtUpdateContext, which was
> + // created while holding &mut Vm, preventing any other access to the
> + // page table for the duration of this operation.
> + let unmapped = unsafe { pt.unmap_pages(iova as usize, pgsize as usize, pgcount as usize) };
It would be ideal if this SAFETY comment explains why this safety
requirement is satisfied:
This page table must contain one or more consecutive mappings
starting at `iova` whose total size is `pgcount * pgsize`.
I.e. why can we be sure that this call isn't trying to unmap a 4KB
sub-page in something that exists as a 2MB mapping in the page table?
(It's because GPUVM tells you to unmap exactly the range you were
previously told to map.)
Alice
^ permalink raw reply [flat|nested] 13+ messages in thread* Re: [PATCH v6 4/7] drm/tyr: add GPU virtual memory (VM) support
2026-07-10 14:15 ` Alice Ryhl
@ 2026-07-14 3:12 ` Deborah Brouwer
0 siblings, 0 replies; 13+ messages in thread
From: Deborah Brouwer @ 2026-07-14 3:12 UTC (permalink / raw)
To: Alice Ryhl
Cc: Daniel Almeida, Danilo Krummrich, David Airlie, Simona Vetter,
Benno Lossin, Gary Guo, dri-devel, linux-kernel, rust-for-linux,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
On Fri, Jul 10, 2026 at 02:15:09PM +0000, Alice Ryhl wrote:
> On Thu, Jul 09, 2026 at 02:36:44PM -0700, Deborah Brouwer wrote:
> > From: Boris Brezillon <boris.brezillon@collabora.com>
> >
> > Add GPU virtual address space management using the DRM GPUVM framework.
> > Each virtual memory (VM) space is backed by ARM64 LPAE Stage 1 page tables
> > and can be mapped into hardware address space (AS) slots for GPU execution.
> >
> > The implementation provides memory isolation and virtual address
> > allocation. VMs support mapping GEM buffer objects with configurable
> > protection flags (readonly, noexec, uncached) and handle both 4KB and 2MB
> > page sizes. A new_dummy_object() helper is provided to create a dummy GEM
> > object for use as a GPUVM root.
> >
> > The vm module integrates with the MMU for address space activation and
> > provides map/unmap/remap operations with page table synchronization.
> >
> > Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
> > Co-developed-by: Daniel Almeida <daniel.almeida@collabora.com>
> > Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>
> > Co-developed-by: Deborah Brouwer <deborah.brouwer@collabora.com>
> > Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
>
> > +impl core::fmt::Display for VmMapFlags {
>
> Use kernel::fmt instead of core::fmt.
>
> > +impl TryFrom<u32> for VmMapFlags {
> > + type Error = Error;
> > +
> > + fn try_from(value: u32) -> core::result::Result<Self, Self::Error> {
>
> Simplifies to Result<Self, Self::Error> without full path.
>
> > + let valid = (kernel::uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_READONLY
> > + | kernel::uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_NOEXEC
> > + | kernel::uapi::drm_panthor_vm_bind_op_flags_DRM_PANTHOR_VM_BIND_OP_MAP_UNCACHED)
> > + as u32;
>
> Simplifies to
>
> let valid = Self::Readonly as u32 | Self::Noexec as u32 | Self::Uncached as u32;
>
> > +impl Drop for PtUpdateContext<'_, '_> {
> > + fn drop(&mut self) {
> > + if let Err(e) = self.mmu.end_vm_update(self.as_data) {
> > + pr_err!("Failed to end VM update {:?}\n", e);
> > + }
> > +
> > + if let Err(e) = self.mmu.flush_vm(self.as_data) {
> > + pr_err!("Failed to flush VM {:?}\n", e);
> > + }
>
> I assume it's not possible for this to run during a VM eviction, right?
In this series I don't think it's possible.
VM eviction only happens on firmware drop which happens on device unbind.
So, since firmware is held in RegistrationData, my understanding is that
it won't be dropped while a device-bound VM operation is in progress.
The slot manager does have a function to evict an idle vm to allocate an address
space for a new vm during a mapping/unmapping, but for this fw boot series there
is only one vm and it’s never marked as idle.
>
> > + let gpuvm_unique = kernel::drm::gpuvm::GpuVm::new::<Error, _>(
> > + c_str!("Tyr::GpuVm"),
>
> This can just be c"Tyr::GpuVm". The macro is only needed in macro
> contexts where this rewrite is not possible e.g. because the macro's
> caller provides a non-c str literal.
>
> > + let mut gpuvm_unique = self.gpuvm_unique.lock();
> > +
> > + self.exec_op(gpuvm_unique.as_mut().get_mut(), req, &mut resources)?;
> > +
> > + // We flush the defer cleanup list now. Things will be different in
> > + // the asynchronous VM_BIND path, where we want the cleanup to
> > + // happen outside the DMA signalling path.
> > + self.gpuvm.deferred_cleanup();
>
> You should probably release the gpuvm_unique mutex before invoking
> deferred_cleanup().
>
> (Multiple occurances in this patch.)
>
> > + if gem_offset > 0 {
> > + // Skip the entire SGT entry if the gem_offset exceeds its length
> > + let skip = sgt_entry_length.min(gem_offset);
>
> Nit: this may just be my idiosyncrasy, but I think this is more readable as:
>
> usize::min(sgt_entry_length, gem_offset)
>
> (Multiple occurances in this file.)
>
> > + let segment_mapped = match pt_map(context.pt, iova, paddr, len, prot) {
> > + Ok(segment_mapped) => segment_mapped,
> > + Err(e) => {
> > + // clean up any successful mappings from previous SGT entries.
> > + let total_mapped = iova - start_iova;
> > + if total_mapped > 0 {
> > + pt_unmap(context.pt, start_iova..(start_iova + total_mapped)).ok();
>
> Nit: the idiomatic way to ignore errors is to write:
>
> let _ = pt_unmap(context.pt, start_iova..(start_iova + total_mapped));
>
> instead of converting the Result into an Option with .ok() and abusing
> the fact that Option is not #[must_use].
>
> > +/// This function selects the largest supported block size (currently 4KB or 2MB)
> > +/// that can be used for a mapping at the given address and size, respecting alignment constraints.
> > +///
> > +/// We can map multiple pages at once but we can't exceed the size of the
> > +// table entry itself. So, if mapping 4KB pages, figure out how many pages
> > +// can be mapped before we hit the 2MB boundary. Or, if mapping 2MB pages,
> > +// figure out how many pages can be mapped before hitting the 1GB boundary
> > +// Returns the page size (4KB or 2MB) and the number of pages that can be mapped at that size.
> > +fn get_pgsize(addr: u64, size: u64) -> (u64, u64) {
>
> Some of these comments are missing a / character.
>
> > + // SAFETY: Exclusive access to the page table is ensured because
> > + // the pt reference comes from PtUpdateContext, which is created
> > + // during a VM update operation, ensuring the driver does not concurrently
> > + // modify the page table.
> > + let (mapped, result) = unsafe {
> > + pt.map_pages(
> > + curr_iova as usize,
> > + (curr_paddr as usize).try_into().unwrap(),
> > + pgsize as usize,
> > + pgcount as usize,
> > + prot,
> > + GFP_KERNEL,
> > + )
> > + };
>
> It would be ideal if this SAFETY comment explains why this safety
> requirement is satisfied:
>
> This page table must not contain any mapping that overlaps with
> the mapping created by this call.
>
> (It's because GPUVM tells you to unmap pages before it tells you to map
> them.)
>
> > +/// Unmaps a virtual address range from the page table.
> > +///
> > +/// This function removes all page table entries in the specified range,
> > +/// automatically handling different page sizes that may be present.
> > +fn pt_unmap(pt: &IoPageTable<'_, ARM64LPAES1>, range: Range<u64>) -> Result {
> > + let mut iova = range.start;
> > + let mut bytes_left_to_unmap = range.end - range.start;
> > +
> > + while bytes_left_to_unmap > 0 {
> > + let (pgsize, pgcount) = get_pgsize(iova, bytes_left_to_unmap);
> > +
> > + // SAFETY: Exclusive access to the page table is ensured because
> > + // the pt reference comes from PtUpdateContext, which was
> > + // created while holding &mut Vm, preventing any other access to the
> > + // page table for the duration of this operation.
> > + let unmapped = unsafe { pt.unmap_pages(iova as usize, pgsize as usize, pgcount as usize) };
>
> It would be ideal if this SAFETY comment explains why this safety
> requirement is satisfied:
>
> This page table must contain one or more consecutive mappings
> starting at `iova` whose total size is `pgcount * pgsize`.
>
> I.e. why can we be sure that this call isn't trying to unmap a 4KB
> sub-page in something that exists as a 2MB mapping in the page table?
>
> (It's because GPUVM tells you to unmap exactly the range you were
> previously told to map.)
>
> Alice
^ permalink raw reply [flat|nested] 13+ messages in thread
* Re: [PATCH v6 4/7] drm/tyr: add GPU virtual memory (VM) support
2026-07-09 21:36 ` [PATCH v6 4/7] drm/tyr: add GPU virtual memory (VM) support Deborah Brouwer
2026-07-10 14:15 ` Alice Ryhl
@ 2026-07-10 14:27 ` Alice Ryhl
1 sibling, 0 replies; 13+ messages in thread
From: Alice Ryhl @ 2026-07-10 14:27 UTC (permalink / raw)
To: Deborah Brouwer
Cc: Daniel Almeida, Danilo Krummrich, David Airlie, Simona Vetter,
Benno Lossin, Gary Guo, dri-devel, linux-kernel, rust-for-linux,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
On Thu, Jul 09, 2026 at 02:36:44PM -0700, Deborah Brouwer wrote:
> From: Boris Brezillon <boris.brezillon@collabora.com>
>
> Add GPU virtual address space management using the DRM GPUVM framework.
> Each virtual memory (VM) space is backed by ARM64 LPAE Stage 1 page tables
> and can be mapped into hardware address space (AS) slots for GPU execution.
>
> The implementation provides memory isolation and virtual address
> allocation. VMs support mapping GEM buffer objects with configurable
> protection flags (readonly, noexec, uncached) and handle both 4KB and 2MB
> page sizes. A new_dummy_object() helper is provided to create a dummy GEM
> object for use as a GPUVM root.
>
> The vm module integrates with the MMU for address space activation and
> provides map/unmap/remap operations with page table synchronization.
>
> Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
> Co-developed-by: Daniel Almeida <daniel.almeida@collabora.com>
> Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>
> Co-developed-by: Deborah Brouwer <deborah.brouwer@collabora.com>
> Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
I'm seeing some build failures on 32-bit arm:
error[E0308]: mismatched types
--> /home/runner/work/linux/linux/linux/drivers/gpu/drm/tyr/vm.rs:572:45
|
572 | let mut sgt_entry_length: u64 = sgt_entry.dma_len();
| --- ^^^^^^^^^^^^^^^^^^^ expected `u64`, found `u32`
| |
| expected due to this
|
help: you can convert a `u32` to a `u64`
|
572 | let mut sgt_entry_length: u64 = sgt_entry.dma_len().into();
| +++++++
error[E0308]: mismatched types
--> /home/runner/work/linux/linux/linux/drivers/gpu/drm/tyr/vm.rs:581:26
|
581 | paddr += skip;
| ^^^^ expected `u32`, found `u64`
error[E0277]: cannot add-assign `u64` to `u32`
--> /home/runner/work/linux/linux/linux/drivers/gpu/drm/tyr/vm.rs:581:23
|
581 | paddr += skip;
| ^^ no implementation for `u32 += u64`
|
= help: the trait `core::ops::AddAssign<u64>` is not implemented for `u32`
help: `u32` implements trait `core::ops::AddAssign<Rhs>`
--> /home/runner/.rustup/toolchains/stable-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/core/src/internal_macros.rs:61:9
|
61 | impl const $imp<&$u> for $t {
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^ `core::ops::AddAssign<&u32>`
|
::: /home/runner/work/linux/linux/linux/rust/zerocopy/src/byteorder.rs:316:9
|
316 | impl<O: ByteOrder> core::ops::$trait_assign<$name<O>> for $native {
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ `core::ops::AddAssign<zerocopy::byteorder::U32<O>>`
...
720 | / define_type!(
721 | | A,
722 | | "A 32-bit unsigned integer",
723 | | U32,
... |
735 | | [U64, U128]
736 | | );
| |_- in this macro invocation
|
::: /home/runner/.rustup/toolchains/stable-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/core/src/ops/arith.rs:786:9
|
786 | impl const AddAssign for $t {
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^ `core::ops::AddAssign`
CC drivers/base/firmware_loader/builtin/main.o
...
799 | add_assign_impl! { usize u8 u16 u32 u64 u128 isize i8 i16 i32 i64 i128 f16 f32 f64 f128 }
| ----------------------------------------------------------------------------------------- in this macro invocation
= note: this error originates in the macro `impl_ops_traits` which comes from the expansion of the macro `add_assign_impl` (in Nightly builds, run with -Z macro-backtrace for more info)
error[E0308]: mismatched types
--> /home/runner/work/linux/linux/linux/drivers/gpu/drm/tyr/vm.rs:596:65
|
596 | let segment_mapped = match pt_map(context.pt, iova, paddr, len, prot) {
| ------ ^^^^^ expected `u64`, found `u32`
| |
| arguments to this function are incorrect
|
note: function defined here
--> /home/runner/work/linux/linux/linux/drivers/gpu/drm/tyr/vm.rs:731:4
|
731 | fn pt_map(
| ^^^^^^
...
734 | paddr: u64,
| ----------
help: you can convert a `u32` to a `u64`
|
596 | let segment_mapped = match pt_map(context.pt, iova, paddr.into(), len, prot) {
| +++++++
CC net/core/tso.o
error: aborting due to 4 previous errors
^ permalink raw reply [flat|nested] 13+ messages in thread
* [PATCH v6 5/7] drm/tyr: add a kernel buffer object
2026-07-09 21:36 [PATCH v6 0/7] drm/tyr: firmware loading and MCU boot support Deborah Brouwer
` (3 preceding siblings ...)
2026-07-09 21:36 ` [PATCH v6 4/7] drm/tyr: add GPU virtual memory (VM) support Deborah Brouwer
@ 2026-07-09 21:36 ` Deborah Brouwer
2026-07-09 21:36 ` [PATCH v6 6/7] drm/tyr: add parser for firmware binary Deborah Brouwer
2026-07-09 21:36 ` [PATCH v6 7/7] drm/tyr: add Microcontroller Unit (MCU) booting Deborah Brouwer
6 siblings, 0 replies; 13+ messages in thread
From: Deborah Brouwer @ 2026-07-09 21:36 UTC (permalink / raw)
To: Daniel Almeida, Alice Ryhl, Danilo Krummrich, David Airlie,
Simona Vetter, Benno Lossin, Gary Guo
Cc: dri-devel, linux-kernel, rust-for-linux, Deborah Brouwer,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
Introduce a buffer object type (KernelBo) for internal driver allocations
that are managed by the kernel rather than userspace.
KernelBo wraps a GEM shmem object and automatically handles GPU virtual
address space mapping during creation and unmapping on drop. This provides
a safe and convenient way for the driver to both allocate and clean up
internal buffers for kernel-managed resources.
Co-developed-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
---
drivers/gpu/drm/tyr/gem.rs | 101 +++++++++++++++++++++++++++++++++++++++++++--
1 file changed, 97 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/tyr/gem.rs b/drivers/gpu/drm/tyr/gem.rs
index c28be61a01bb..47a05a33388e 100644
--- a/drivers/gpu/drm/tyr/gem.rs
+++ b/drivers/gpu/drm/tyr/gem.rs
@@ -4,18 +4,29 @@
//! This module provides buffer object (BO) management functionality using
//! DRM's GEM subsystem with shmem backing.
+use core::ops::Range;
+
use kernel::{
drm::gem::{
self,
shmem, //
},
prelude::*,
- sync::aref::ARef, //
+ sync::{
+ aref::ARef,
+ Arc, //
+ }, //
};
-use crate::driver::{
- TyrDrmDevice,
- TyrDrmDriver, //
+use crate::{
+ driver::{
+ TyrDrmDevice,
+ TyrDrmDriver, //
+ },
+ vm::{
+ Vm,
+ VmMapFlags, //
+ },
};
/// Tyr's DriverObject type for GEM objects.
@@ -56,3 +67,85 @@ pub(crate) fn new_dummy_object(ddev: &TyrDrmDevice) -> Result<ARef<Bo>> {
Ok(bo)
}
+
+/// Specifies how to choose a GPU virtual address for a [`KernelBo`].
+/// An automatic VA allocation strategy will be added in the future.
+pub(crate) enum KernelBoVaAlloc {
+ /// Explicit VA address specified by the caller.
+ #[expect(dead_code)]
+ Explicit(u64),
+}
+
+/// A kernel-owned buffer object with automatic GPU virtual address mapping.
+///
+/// This structure represents a buffer object that is created and managed entirely
+/// by the kernel driver, as opposed to userspace-created GEM objects. It combines
+/// a GEM object with automatic GPU virtual address (VA) space mapping and cleanup.
+///
+/// When dropped, the buffer is automatically unmapped from the GPU VA space.
+pub(crate) struct KernelBo<'bound> {
+ /// The underlying GEM buffer object.
+ #[expect(dead_code)]
+ pub(crate) bo: ARef<Bo>,
+ /// The GPU VM this buffer is mapped into.
+ vm: Arc<Vm<'bound>>,
+ /// The GPU VA range occupied by this buffer.
+ va_range: Range<u64>,
+}
+
+impl<'bound> KernelBo<'bound> {
+ /// Creates a new kernel-owned buffer object and maps it into GPU VA space.
+ ///
+ /// This function allocates a new shmem-backed GEM object and immediately maps
+ /// it into the specified GPU virtual memory space. The mapping is automatically
+ /// cleaned up when the [`KernelBo`] is dropped.
+ #[expect(dead_code)]
+ pub(crate) fn new(
+ ddev: &TyrDrmDevice,
+ vm: Arc<Vm<'bound>>,
+ size: u64,
+ va_alloc: KernelBoVaAlloc,
+ flags: VmMapFlags,
+ ) -> Result<Self> {
+ if size == 0 {
+ pr_err!("Cannot create KernelBo with size 0\n");
+ return Err(EINVAL);
+ }
+
+ let KernelBoVaAlloc::Explicit(va) = va_alloc;
+
+ let bo = Bo::new(
+ ddev,
+ size as usize,
+ shmem::ObjectConfig {
+ map_wc: true,
+ parent_resv_obj: None,
+ },
+ BoCreateArgs { flags: 0 },
+ )?;
+
+ vm.map_bo_range(&bo, 0, size, va, flags)?;
+
+ Ok(KernelBo {
+ bo,
+ vm,
+ va_range: va..(va + size),
+ })
+ }
+}
+
+impl Drop for KernelBo<'_> {
+ fn drop(&mut self) {
+ let va = self.va_range.start;
+ let size = self.va_range.end - self.va_range.start;
+
+ if let Err(e) = self.vm.unmap_range(va, size) {
+ pr_err!(
+ "Failed to unmap KernelBo range {:#x}..{:#x}: {:?}\n",
+ self.va_range.start,
+ self.va_range.end,
+ e
+ );
+ }
+ }
+}
--
2.54.0
^ permalink raw reply related [flat|nested] 13+ messages in thread* [PATCH v6 6/7] drm/tyr: add parser for firmware binary
2026-07-09 21:36 [PATCH v6 0/7] drm/tyr: firmware loading and MCU boot support Deborah Brouwer
` (4 preceding siblings ...)
2026-07-09 21:36 ` [PATCH v6 5/7] drm/tyr: add a kernel buffer object Deborah Brouwer
@ 2026-07-09 21:36 ` Deborah Brouwer
2026-07-09 21:36 ` [PATCH v6 7/7] drm/tyr: add Microcontroller Unit (MCU) booting Deborah Brouwer
6 siblings, 0 replies; 13+ messages in thread
From: Deborah Brouwer @ 2026-07-09 21:36 UTC (permalink / raw)
To: Daniel Almeida, Alice Ryhl, Danilo Krummrich, David Airlie,
Simona Vetter, Benno Lossin, Gary Guo
Cc: dri-devel, linux-kernel, rust-for-linux, Deborah Brouwer,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
From: Daniel Almeida <daniel.almeida@collabora.com>
Add a parser for the Mali CSF GPU firmware binary format. The firmware
consists of a header followed by entries describing how to load firmware
sections into the MCU's memory.
The parser extracts section metadata including virtual address ranges,
data byte offsets within the binary, and section flags controlling
permissions and cache modes. It validates the basic firmware structure
and alignment and ignores protected-mode sections for now.
Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>
Co-developed-by: Beata Michalska <beata.michalska@arm.com>
Signed-off-by: Beata Michalska <beata.michalska@arm.com>
Co-developed-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Co-developed-by: Deborah Brouwer <deborah.brouwer@collabora.com>
Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
---
drivers/gpu/drm/tyr/fw/parser.rs | 521 +++++++++++++++++++++++++++++++++++++++
1 file changed, 521 insertions(+)
diff --git a/drivers/gpu/drm/tyr/fw/parser.rs b/drivers/gpu/drm/tyr/fw/parser.rs
new file mode 100644
index 000000000000..198a754b294f
--- /dev/null
+++ b/drivers/gpu/drm/tyr/fw/parser.rs
@@ -0,0 +1,521 @@
+// SPDX-License-Identifier: GPL-2.0 or MIT
+
+//! Firmware binary parser for Mali CSF (Command Stream Frontend) GPU.
+//!
+//! This module implements a parser for the Mali GPU firmware binary format. The firmware
+//! file contains a header followed by a sequence of entries, each describing how to load
+//! firmware sections into the MCU (Microcontroller Unit) memory. The parser extracts section
+//! metadata including:
+//! - Virtual address ranges where sections should be mapped
+//! - Data ranges (byte offsets) within the firmware binary
+//! - Section flags (permissions, cache modes)
+
+use core::{
+ mem::size_of,
+ ops::Range, //
+};
+
+use kernel::{
+ bits::bit_u32,
+ prelude::*,
+ str::CString, //
+};
+
+use crate::{
+ fw::{
+ SectionFlag,
+ SectionFlags,
+ CSF_MCU_SHARED_REGION_START, //
+ },
+ vm::{
+ VmFlag,
+ VmMapFlags, //
+ }, //
+};
+
+/// A parsed firmware section ready for loading into MCU memory.
+///
+/// Represents a single firmware section extracted from the firmware binary, containing
+/// all information needed to map the section's data into the MCU's virtual address space.
+pub(super) struct ParsedSection {
+ /// Byte offset range within the firmware binary where this section's data resides.
+ pub(super) data_range: Range<u32>,
+ /// MCU virtual address range where this section should be mapped.
+ pub(super) va: Range<u32>,
+ /// Memory protection and caching flags for the mapping.
+ pub(super) vm_map_flags: VmMapFlags,
+}
+
+/// A bare-bones `std::io::Cursor<[u8]>` clone to keep track of the current position in the
+/// firmware binary.
+///
+/// Provides methods to sequentially read primitive types and byte arrays from the firmware
+/// binary while maintaining the current read position.
+struct Cursor<'a> {
+ data: &'a [u8],
+ pos: usize,
+}
+
+impl<'a> Cursor<'a> {
+ fn new(data: &'a [u8]) -> Self {
+ Self { data, pos: 0 }
+ }
+
+ fn len(&self) -> usize {
+ self.data.len()
+ }
+
+ fn pos(&self) -> usize {
+ self.pos
+ }
+
+ /// Returns a view into the cursor's data.
+ ///
+ /// This spawns a new cursor, leaving the current cursor unchanged.
+ fn view(&self, range: Range<usize>) -> Result<Cursor<'_>> {
+ if range.start < self.pos || range.end > self.data.len() {
+ pr_err!(
+ "Invalid cursor range {:?} for data of length {}",
+ range,
+ self.data.len()
+ );
+
+ Err(EINVAL)
+ } else {
+ Ok(Self {
+ data: &self.data[range],
+ pos: 0,
+ })
+ }
+ }
+
+ /// Reads a slice of bytes from the current position and advances the cursor.
+ ///
+ /// Returns an error if the read would exceed the data bounds.
+ fn read(&mut self, nbytes: usize) -> Result<&[u8]> {
+ let start = self.pos;
+ let end = start + nbytes;
+
+ if end > self.data.len() {
+ pr_err!(
+ "Invalid firmware file: read of size {} at position {} is out of bounds",
+ nbytes,
+ start,
+ );
+ return Err(EINVAL);
+ }
+
+ self.pos += nbytes;
+ Ok(&self.data[start..end])
+ }
+
+ /// Reads a little-endian `u8` from the current position and advances the cursor.
+ fn read_u8(&mut self) -> Result<u8> {
+ let bytes = self.read(size_of::<u8>())?;
+ Ok(bytes[0])
+ }
+
+ /// Reads a little-endian `u16` from the current position and advances the cursor.
+ fn read_u16(&mut self) -> Result<u16> {
+ let bytes = self.read(size_of::<u16>())?;
+ Ok(u16::from_le_bytes(bytes.try_into().unwrap()))
+ }
+
+ /// Reads a little-endian `u32` from the current position and advances the cursor.
+ fn read_u32(&mut self) -> Result<u32> {
+ let bytes = self.read(size_of::<u32>())?;
+ Ok(u32::from_le_bytes(bytes.try_into().unwrap()))
+ }
+
+ /// Advances the cursor position by the specified number of bytes.
+ ///
+ /// Returns an error if the advance would exceed the data bounds.
+ fn advance(&mut self, nbytes: usize) -> Result {
+ if self.pos + nbytes > self.data.len() {
+ pr_err!(
+ "Invalid firmware file: advance of size {} at position {} is out of bounds",
+ nbytes,
+ self.pos,
+ );
+ return Err(EINVAL);
+ }
+ self.pos += nbytes;
+ Ok(())
+ }
+}
+
+/// Parser for Mali CSF GPU firmware binaries.
+///
+/// Parses the firmware binary format, extracting section metadata including virtual
+/// address ranges, data offsets, and memory protection flags needed to load firmware
+/// into the MCU's memory.
+pub(super) struct FwParser<'a> {
+ cursor: Cursor<'a>,
+}
+
+impl<'a> FwParser<'a> {
+ /// Creates a new firmware parser for the given firmware binary data.
+ pub(super) fn new(data: &'a [u8]) -> Self {
+ Self {
+ cursor: Cursor::new(data),
+ }
+ }
+
+ /// Parses the firmware binary and returns a collection of parsed sections.
+ ///
+ /// This method validates the firmware header and iterates through all entries
+ /// in the binary, extracting section information needed for loading.
+ pub(super) fn parse(&mut self) -> Result<KVec<ParsedSection>> {
+ let fw_header = self.parse_fw_header()?;
+
+ let mut parsed_sections = KVec::new();
+ while (self.cursor.pos() as u32) < fw_header.size {
+ let entry_section = self.parse_entry()?;
+
+ if let Some(inner) = entry_section.inner {
+ parsed_sections.push(inner, GFP_KERNEL)?;
+ }
+ }
+
+ Ok(parsed_sections)
+ }
+
+ fn parse_fw_header(&mut self) -> Result<FirmwareHeader> {
+ let fw_header: FirmwareHeader = match FirmwareHeader::new(&mut self.cursor) {
+ Ok(fw_header) => fw_header,
+ Err(e) => {
+ pr_err!("Invalid firmware file: {}", e.to_errno());
+ return Err(e);
+ }
+ };
+
+ if fw_header.size > self.cursor.len() as u32 {
+ pr_err!("Firmware image is truncated");
+ return Err(EINVAL);
+ }
+ Ok(fw_header)
+ }
+
+ fn parse_entry(&mut self) -> Result<EntrySection> {
+ let entry_section = EntrySection {
+ entry_hdr: EntryHeader(self.cursor.read_u32()?),
+ inner: None,
+ };
+
+ if self.cursor.pos() % size_of::<u32>() != 0
+ || entry_section.entry_hdr.size() as usize % size_of::<u32>() != 0
+ {
+ pr_err!(
+ "Firmware entry isn't 32 bit aligned, offset={:#x} size={:#x}\n",
+ self.cursor.pos() - size_of::<u32>(),
+ entry_section.entry_hdr.size()
+ );
+ return Err(EINVAL);
+ }
+
+ let section_hdr_size = entry_section.entry_hdr.size() as usize - size_of::<EntryHeader>();
+
+ let entry_section = {
+ let mut entry_cursor = self
+ .cursor
+ .view(self.cursor.pos()..self.cursor.pos() + section_hdr_size)?;
+
+ match entry_section.entry_hdr.entry_type() {
+ Ok(EntryType::Iface) => Ok(EntrySection {
+ entry_hdr: entry_section.entry_hdr,
+ inner: Self::parse_section_entry(&mut entry_cursor)?,
+ }),
+ Ok(
+ EntryType::Config
+ | EntryType::FutfTest
+ | EntryType::TraceBuffer
+ | EntryType::TimelineMetadata
+ | EntryType::BuildInfoMetadata,
+ ) => Ok(entry_section),
+
+ entry_type => {
+ if entry_type.is_err() || !entry_section.entry_hdr.optional() {
+ if !entry_section.entry_hdr.optional() {
+ pr_err!(
+ "Failed to handle firmware entry type: {}\n",
+ entry_type
+ .map_or(entry_section.entry_hdr.entry_type_raw(), |e| e as u8)
+ );
+ Err(EINVAL)
+ } else {
+ Ok(entry_section)
+ }
+ } else {
+ Ok(entry_section)
+ }
+ }
+ }
+ };
+
+ if entry_section.is_ok() {
+ self.cursor.advance(section_hdr_size)?;
+ }
+
+ entry_section
+ }
+
+ fn parse_section_entry(entry_cursor: &mut Cursor<'_>) -> Result<Option<ParsedSection>> {
+ let section_hdr: SectionHeader = SectionHeader::new(entry_cursor)?;
+
+ if section_hdr.data.end < section_hdr.data.start {
+ pr_err!(
+ "Firmware corrupted, data.end < data.start (0x{:x} < 0x{:x})\n",
+ section_hdr.data.end,
+ section_hdr.data.start
+ );
+ return Err(EINVAL);
+ }
+
+ if section_hdr.va.end < section_hdr.va.start {
+ pr_err!(
+ "Firmware corrupted, section_hdr.va.end < section_hdr.va.start (0x{:x} < 0x{:x})\n",
+ section_hdr.va.end,
+ section_hdr.va.start
+ );
+ return Err(EINVAL);
+ }
+
+ if section_hdr.section_flags.contains(SectionFlag::Prot) {
+ pr_info!("Firmware protected mode entry not supported, ignoring");
+ return Ok(None);
+ }
+
+ if section_hdr.va.start == CSF_MCU_SHARED_REGION_START
+ && !section_hdr.section_flags.contains(SectionFlag::Shared)
+ {
+ pr_err!(
+ "Interface at 0x{:x} must be shared",
+ CSF_MCU_SHARED_REGION_START
+ );
+ return Err(EINVAL);
+ }
+
+ let name_len = entry_cursor.len() - entry_cursor.pos();
+ let name_bytes = entry_cursor.read(name_len)?;
+
+ let mut name = KVec::with_capacity(name_bytes.len() + 1, GFP_KERNEL)?;
+ name.extend_from_slice(name_bytes, GFP_KERNEL)?;
+ name.push(0, GFP_KERNEL)?;
+
+ let _name = CStr::from_bytes_with_nul(&name)
+ .ok()
+ .and_then(|name| CString::try_from(name).ok());
+
+ let cache_mode = section_hdr.section_flags.cache_mode();
+ let mut vm_map_flags = VmMapFlags::empty();
+
+ if !section_hdr.section_flags.contains(SectionFlag::Write) {
+ vm_map_flags |= VmFlag::Readonly;
+ }
+ if !section_hdr.section_flags.contains(SectionFlag::Exec) {
+ vm_map_flags |= VmFlag::Noexec;
+ }
+ if cache_mode != SectionFlag::CacheModeCached.into() {
+ vm_map_flags |= VmFlag::Uncached;
+ }
+
+ Ok(Some(ParsedSection {
+ data_range: section_hdr.data.clone(),
+ va: section_hdr.va,
+ vm_map_flags,
+ }))
+ }
+}
+
+/// Firmware binary header containing version and size information.
+///
+/// The header is located at the beginning of the firmware binary and contains
+/// a magic value for validation, version information, and the total size of
+/// all structured headers that follow.
+#[expect(dead_code)]
+struct FirmwareHeader {
+ /// Magic value to check binary validity.
+ magic: u32,
+
+ /// Minor firmware version.
+ minor: u8,
+
+ /// Major firmware version.
+ major: u8,
+
+ /// Padding. Must be set to zero.
+ _padding1: u16,
+
+ /// Firmware version hash.
+ version_hash: u32,
+
+ /// Padding. Must be set to zero.
+ _padding2: u32,
+
+ /// Total size of all the structured data headers at beginning of firmware binary.
+ size: u32,
+}
+
+impl FirmwareHeader {
+ const FW_BINARY_MAGIC: u32 = 0xc3f13a6e;
+ const FW_BINARY_MAJOR_MAX: u8 = 0;
+
+ /// Reads and validates a firmware header from the cursor.
+ ///
+ /// Verifies the magic value, version compatibility, and padding fields.
+ fn new(cursor: &mut Cursor<'_>) -> Result<Self> {
+ let magic = cursor.read_u32()?;
+ if magic != Self::FW_BINARY_MAGIC {
+ pr_err!("Invalid firmware magic");
+ return Err(EINVAL);
+ }
+
+ let minor = cursor.read_u8()?;
+ let major = cursor.read_u8()?;
+
+ if major > Self::FW_BINARY_MAJOR_MAX {
+ pr_err!(
+ "Unsupported firmware binary header version {}.{} (expected {}.x)\n",
+ major,
+ minor,
+ Self::FW_BINARY_MAJOR_MAX
+ );
+ return Err(EINVAL);
+ }
+
+ let padding1 = cursor.read_u16()?;
+ let version_hash = cursor.read_u32()?;
+ let padding2 = cursor.read_u32()?;
+ let size = cursor.read_u32()?;
+
+ if padding1 != 0 || padding2 != 0 {
+ pr_err!("Invalid firmware file: header padding is not zero");
+ return Err(EINVAL);
+ }
+
+ let fw_header = Self {
+ magic,
+ minor,
+ major,
+ _padding1: padding1,
+ version_hash,
+ _padding2: padding2,
+ size,
+ };
+
+ Ok(fw_header)
+ }
+}
+
+/// Firmware section header for loading binary sections into MCU memory.
+#[derive(Debug)]
+struct SectionHeader {
+ section_flags: SectionFlags,
+ /// MCU virtual range to map this binary section to.
+ va: Range<u32>,
+ /// References the data in the FW binary.
+ data: Range<u32>,
+}
+
+impl SectionHeader {
+ /// Reads and validates a section header from the cursor.
+ ///
+ /// Parses section flags, virtual address range, and data range from the firmware binary.
+ fn new(cursor: &mut Cursor<'_>) -> Result<Self> {
+ let section_flags = cursor.read_u32()?;
+ let section_flags = SectionFlags::try_from(section_flags)?;
+
+ let va_start = cursor.read_u32()?;
+ let va_end = cursor.read_u32()?;
+
+ let va = va_start..va_end;
+
+ if va.is_empty() {
+ pr_err!(
+ "Invalid firmware file: empty VA range at pos {}\n",
+ cursor.pos(),
+ );
+ return Err(EINVAL);
+ }
+
+ let data_start = cursor.read_u32()?;
+ let data_end = cursor.read_u32()?;
+ let data = data_start..data_end;
+
+ Ok(Self {
+ section_flags,
+ va,
+ data,
+ })
+ }
+}
+
+/// A firmware entry containing a header and optional parsed section data.
+///
+/// Represents a single entry in the firmware binary, which may contain loadable
+/// section data or metadata that doesn't require loading.
+struct EntrySection {
+ entry_hdr: EntryHeader,
+ inner: Option<ParsedSection>,
+}
+
+/// Header for a firmware entry, packed into a single u32.
+///
+/// The entry header encodes the entry type, size, and optional flag in a
+/// 32-bit value with the following layout:
+/// - Bits 0-7: Entry type
+/// - Bits 8-15: Size in bytes
+/// - Bit 31: Optional flag
+struct EntryHeader(u32);
+
+impl EntryHeader {
+ fn entry_type_raw(&self) -> u8 {
+ (self.0 & 0xff) as u8
+ }
+
+ fn entry_type(&self) -> Result<EntryType> {
+ let v = self.entry_type_raw();
+ EntryType::try_from(v)
+ }
+
+ fn optional(&self) -> bool {
+ self.0 & bit_u32(31) != 0
+ }
+
+ fn size(&self) -> u32 {
+ self.0 >> 8 & 0xff
+ }
+}
+
+#[derive(Clone, Copy, Debug)]
+#[repr(u8)]
+enum EntryType {
+ /// Host <-> FW interface.
+ Iface = 0,
+ /// FW config.
+ Config = 1,
+ /// Unit tests.
+ FutfTest = 2,
+ /// Trace buffer interface.
+ TraceBuffer = 3,
+ /// Timeline metadata interface.
+ TimelineMetadata = 4,
+ /// Metadata about how the FW binary was built.
+ BuildInfoMetadata = 6,
+}
+
+impl TryFrom<u8> for EntryType {
+ type Error = Error;
+
+ fn try_from(value: u8) -> Result<Self, Self::Error> {
+ match value {
+ 0 => Ok(EntryType::Iface),
+ 1 => Ok(EntryType::Config),
+ 2 => Ok(EntryType::FutfTest),
+ 3 => Ok(EntryType::TraceBuffer),
+ 4 => Ok(EntryType::TimelineMetadata),
+ 6 => Ok(EntryType::BuildInfoMetadata),
+ _ => Err(EINVAL),
+ }
+ }
+}
--
2.54.0
^ permalink raw reply related [flat|nested] 13+ messages in thread* [PATCH v6 7/7] drm/tyr: add Microcontroller Unit (MCU) booting
2026-07-09 21:36 [PATCH v6 0/7] drm/tyr: firmware loading and MCU boot support Deborah Brouwer
` (5 preceding siblings ...)
2026-07-09 21:36 ` [PATCH v6 6/7] drm/tyr: add parser for firmware binary Deborah Brouwer
@ 2026-07-09 21:36 ` Deborah Brouwer
6 siblings, 0 replies; 13+ messages in thread
From: Deborah Brouwer @ 2026-07-09 21:36 UTC (permalink / raw)
To: Daniel Almeida, Alice Ryhl, Danilo Krummrich, David Airlie,
Simona Vetter, Benno Lossin, Gary Guo
Cc: dri-devel, linux-kernel, rust-for-linux, Deborah Brouwer,
boris.brezillon, samitolvanen, acourbot, alvin.sun, laura.nao,
work, beata.michalska, steven.price, lyude
Add a firmware module to load, parse, and map the MCU firmware sections
into shared GEM memory at the required virtual addresses accessible by the
GPU.
Create a firmware instance during probe and store it inside the
TyrDrmRegistrationData to keep it alive after probe. Use the firmware
instance to boot the MCU.
Remove the dead-code annotations from the MMU, VM, slot manager, and
kernel BO code now that these paths are used by the firmware module.
Update Kconfig to add the RUST_FW_LOADER_ABSTRACTIONS dependency
required by this module.
Co-developed-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
---
drivers/gpu/drm/tyr/Kconfig | 1 +
drivers/gpu/drm/tyr/driver.rs | 17 ++-
drivers/gpu/drm/tyr/fw.rs | 263 ++++++++++++++++++++++++++++++++++++++++++
drivers/gpu/drm/tyr/gem.rs | 3 -
drivers/gpu/drm/tyr/mmu.rs | 1 -
drivers/gpu/drm/tyr/slot.rs | 1 -
drivers/gpu/drm/tyr/tyr.rs | 1 +
drivers/gpu/drm/tyr/vm.rs | 1 -
8 files changed, 281 insertions(+), 7 deletions(-)
diff --git a/drivers/gpu/drm/tyr/Kconfig b/drivers/gpu/drm/tyr/Kconfig
index 79ea4bb214de..8f13e49f11f9 100644
--- a/drivers/gpu/drm/tyr/Kconfig
+++ b/drivers/gpu/drm/tyr/Kconfig
@@ -13,6 +13,7 @@ config DRM_TYR
select IOMMU_IO_PGTABLE_LPAE
select RUST_DRM_GEM_SHMEM_HELPER
select RUST_DRM_GPUVM
+ select RUST_FW_LOADER_ABSTRACTIONS
help
Rust DRM driver for ARM Mali CSF-based GPUs.
diff --git a/drivers/gpu/drm/tyr/driver.rs b/drivers/gpu/drm/tyr/driver.rs
index 9195c8be5203..46ba91af9aca 100644
--- a/drivers/gpu/drm/tyr/driver.rs
+++ b/drivers/gpu/drm/tyr/driver.rs
@@ -37,6 +37,7 @@
use crate::{
file::TyrDrmFileData,
+ fw::Firmware,
gem::Bo,
gpu,
gpu::GpuInfo,
@@ -67,6 +68,9 @@ pub(crate) struct TyrDrmRegistrationData<'bound> {
/// Parent platform device.
pub(crate) pdev: &'bound platform::Device<Bound>,
+ /// Firmware sections.
+ pub(crate) fw: Firmware<'bound>,
+
#[pin]
clks: Mutex<Clocks>,
@@ -144,10 +148,21 @@ fn probe<'bound>(
let unreg_dev = drm::UnregisteredDevice::<TyrDrmDriver>::new(pdev, Ok(()))?;
- let _mmu = Mmu::new(iomem.as_arc_borrow(), &gpu_info)?;
+ let mmu = Mmu::new(iomem.as_arc_borrow(), &gpu_info)?;
+
+ let firmware = Firmware::new(
+ pdev.as_ref(),
+ iomem.clone(),
+ &unreg_dev,
+ mmu.as_arc_borrow(),
+ &gpu_info,
+ )?;
+
+ firmware.boot()?;
let reg_data = try_pin_init!(TyrDrmRegistrationData {
pdev,
+ fw: firmware,
clks <- new_mutex!(Clocks {
core: core_clk,
stacks: stacks_clk,
diff --git a/drivers/gpu/drm/tyr/fw.rs b/drivers/gpu/drm/tyr/fw.rs
new file mode 100644
index 000000000000..554808a792aa
--- /dev/null
+++ b/drivers/gpu/drm/tyr/fw.rs
@@ -0,0 +1,263 @@
+// SPDX-License-Identifier: GPL-2.0 or MIT
+
+//! Firmware loading and management for Mali CSF GPU.
+//!
+//! This module handles loading the Mali GPU firmware binary, parsing it into sections,
+//! and mapping those sections into the MCU's virtual address space. Each firmware section
+//! has specific properties (read/write/execute permissions, cache modes) and must be loaded
+//! at specific virtual addresses expected by the MCU.
+//!
+//! See [`Firmware`] for the main firmware management interface and [`Section`] for
+//! individual firmware sections.
+//!
+//! [`Firmware`]: crate::fw::Firmware
+//! [`Section`]: crate::fw::Section
+
+use kernel::{
+ bits::genmask_u32,
+ device::{
+ Bound,
+ Device, //
+ },
+ drm::{
+ gem::BaseObject, //
+ },
+ impl_flags,
+ io::{
+ poll,
+ Io, //
+ },
+ prelude::*,
+ str::CString,
+ sync::{
+ Arc,
+ ArcBorrow, //
+ },
+ time, //
+};
+
+use crate::{
+ driver::{
+ IoMem,
+ TyrDrmDevice, //
+ },
+ fw::parser::{
+ FwParser,
+ ParsedSection, //
+ },
+ gem,
+ gem::{
+ KernelBo,
+ KernelBoVaAlloc, //
+ },
+ gpu::GpuInfo,
+
+ mmu::Mmu,
+ regs::{
+ gpu_control::{
+ McuControlMode,
+ McuStatus,
+ GPU_ID,
+ MCU_CONTROL,
+ MCU_STATUS, //
+ }, //
+ job_control::JOB_IRQ_RAWSTAT, //
+ },
+ vm::Vm, //
+};
+
+mod parser;
+
+impl_flags!(
+ #[derive(Debug, Clone, Default, Copy, PartialEq, Eq)]
+ pub(super) struct SectionFlags(u32);
+
+ #[derive(Debug, Clone, Copy, PartialEq, Eq)]
+ pub(super) enum SectionFlag {
+ Read = 1 << 0,
+ Write = 1 << 1,
+ Exec = 1 << 2,
+ CacheModeNone = 0 << 3,
+ CacheModeCached = 1 << 3,
+ CacheModeUncachedCoherent = 2 << 3,
+ CacheModeCachedCoherent = 3 << 3,
+ Prot = 1 << 5,
+ Shared = 1 << 30,
+ Zero = 1 << 31,
+ }
+);
+
+pub(super) const CACHE_MODE_MASK: SectionFlags = SectionFlags(genmask_u32(3..=4));
+
+pub(super) const CSF_MCU_SHARED_REGION_START: u32 = 0x04000000;
+
+impl SectionFlags {
+ fn cache_mode(&self) -> SectionFlags {
+ *self & CACHE_MODE_MASK
+ }
+}
+
+impl TryFrom<u32> for SectionFlags {
+ type Error = Error;
+
+ fn try_from(value: u32) -> Result<Self, Self::Error> {
+ let valid_flags = SectionFlags::from(SectionFlag::Read)
+ | SectionFlags::from(SectionFlag::Write)
+ | SectionFlags::from(SectionFlag::Exec)
+ | CACHE_MODE_MASK
+ | SectionFlags::from(SectionFlag::Prot)
+ | SectionFlags::from(SectionFlag::Shared)
+ | SectionFlags::from(SectionFlag::Zero);
+
+ if value & valid_flags.0 != value {
+ Err(EINVAL)
+ } else {
+ Ok(Self(value))
+ }
+ }
+}
+
+/// A parsed section of the firmware binary.
+struct Section<'bound> {
+ // Raw firmware section data for reset purposes
+ #[expect(dead_code)]
+ data: KVec<u8>,
+
+ // Keep the BO backing this firmware section so that both the
+ // GPU mapping and CPU mapping remain valid until the Section is dropped.
+ #[expect(dead_code)]
+ mem: gem::KernelBo<'bound>,
+}
+
+/// Loaded firmware with sections mapped into MCU VM.
+pub(crate) struct Firmware<'bound> {
+ /// Iomem need to access registers.
+ iomem: Arc<IoMem<'bound>>,
+
+ /// MCU VM.
+ vm: Arc<Vm<'bound>>,
+
+ /// List of firmware sections.
+ #[expect(dead_code)]
+ sections: KVec<Section<'bound>>,
+}
+
+impl<'bound> Drop for Firmware<'bound> {
+ fn drop(&mut self) {
+ // AS slots retain a VM ref, we need to kill the circular ref manually.
+ self.vm.kill();
+ }
+}
+
+impl<'bound> Firmware<'bound> {
+ fn init_section_mem(mem: &mut KernelBo<'bound>, data: &KVec<u8>) -> Result {
+ if data.is_empty() {
+ return Ok(());
+ }
+
+ let vmap = mem.bo.vmap::<0>()?;
+ let size = mem.bo.size();
+
+ if data.len() > size {
+ pr_err!("fw section {} bigger than BO {}\n", data.len(), size);
+ return Err(EINVAL);
+ }
+
+ for (i, &byte) in data.iter().enumerate() {
+ vmap.try_write8(byte, i)?;
+ }
+
+ Ok(())
+ }
+
+ fn request(ddev: &TyrDrmDevice, gpu_info: &GpuInfo) -> Result<kernel::firmware::Firmware> {
+ let gpu_id = GPU_ID::from_raw(gpu_info.gpu_id);
+
+ let path = CString::try_from_fmt(fmt!(
+ "arm/mali/arch{}.{}/mali_csffw.bin",
+ gpu_id.arch_major().get(),
+ gpu_id.arch_minor().get()
+ ))?;
+
+ kernel::firmware::Firmware::request(&path, ddev.as_ref().as_ref())
+ }
+
+ fn load(
+ ddev: &TyrDrmDevice,
+ gpu_info: &GpuInfo,
+ ) -> Result<(kernel::firmware::Firmware, KVec<ParsedSection>)> {
+ let fw = Self::request(ddev, gpu_info)?;
+ let mut parser = FwParser::new(fw.data());
+
+ let parsed_sections = parser.parse()?;
+
+ Ok((fw, parsed_sections))
+ }
+
+ /// Load firmware and map sections into MCU VM.
+ pub(crate) fn new(
+ dev: &'bound Device<Bound>,
+ iomem: Arc<IoMem<'bound>>,
+ ddev: &TyrDrmDevice,
+ mmu: ArcBorrow<'_, Mmu<'bound>>,
+ gpu_info: &GpuInfo,
+ ) -> Result<Firmware<'bound>> {
+ let vm = Vm::new(dev, ddev, mmu, gpu_info)?;
+ vm.activate()?;
+
+ let (fw, parsed_sections) = Self::load(ddev, gpu_info)?;
+ let mut sections = KVec::new();
+ for parsed in parsed_sections {
+ let size = (parsed.va.end - parsed.va.start) as usize;
+ let va = u64::from(parsed.va.start);
+
+ let mut mem = KernelBo::new(
+ ddev,
+ vm.clone(),
+ size.try_into().unwrap(),
+ KernelBoVaAlloc::Explicit(va),
+ parsed.vm_map_flags,
+ )?;
+
+ let section_start = parsed.data_range.start as usize;
+ let section_end = parsed.data_range.end as usize;
+ let mut data = KVec::new();
+
+ // Ensure that the firmware slice is not out of bounds.
+ let fw_data = fw.data();
+ let bytes = fw_data.get(section_start..section_end).ok_or(EINVAL)?;
+ data.extend_from_slice(bytes, GFP_KERNEL)?;
+
+ Self::init_section_mem(&mut mem, &data)?;
+
+ sections.push(Section { data, mem }, GFP_KERNEL)?;
+ }
+
+ let firmware = Firmware {
+ iomem,
+ vm,
+ sections,
+ };
+
+ Ok(firmware)
+ }
+
+ pub(crate) fn boot(&self) -> Result {
+ let io = &self.iomem;
+ io.write_reg(MCU_CONTROL::zeroed().with_req(McuControlMode::Auto));
+
+ if let Err(e) = poll::read_poll_timeout(
+ || Ok((io.read(MCU_STATUS), io.read(JOB_IRQ_RAWSTAT))),
+ |(mcu_status, irq_rawstat)| {
+ mcu_status.value() == McuStatus::Enabled && irq_rawstat.glb()
+ },
+ time::Delta::from_millis(1),
+ time::Delta::from_millis(100),
+ ) {
+ let status = io.read(MCU_STATUS);
+ pr_err!("MCU failed to boot, status: {:?}", status.value());
+ return Err(e);
+ }
+ Ok(())
+ }
+}
diff --git a/drivers/gpu/drm/tyr/gem.rs b/drivers/gpu/drm/tyr/gem.rs
index 47a05a33388e..e9bf35682a47 100644
--- a/drivers/gpu/drm/tyr/gem.rs
+++ b/drivers/gpu/drm/tyr/gem.rs
@@ -72,7 +72,6 @@ pub(crate) fn new_dummy_object(ddev: &TyrDrmDevice) -> Result<ARef<Bo>> {
/// An automatic VA allocation strategy will be added in the future.
pub(crate) enum KernelBoVaAlloc {
/// Explicit VA address specified by the caller.
- #[expect(dead_code)]
Explicit(u64),
}
@@ -85,7 +84,6 @@ pub(crate) enum KernelBoVaAlloc {
/// When dropped, the buffer is automatically unmapped from the GPU VA space.
pub(crate) struct KernelBo<'bound> {
/// The underlying GEM buffer object.
- #[expect(dead_code)]
pub(crate) bo: ARef<Bo>,
/// The GPU VM this buffer is mapped into.
vm: Arc<Vm<'bound>>,
@@ -99,7 +97,6 @@ impl<'bound> KernelBo<'bound> {
/// This function allocates a new shmem-backed GEM object and immediately maps
/// it into the specified GPU virtual memory space. The mapping is automatically
/// cleaned up when the [`KernelBo`] is dropped.
- #[expect(dead_code)]
pub(crate) fn new(
ddev: &TyrDrmDevice,
vm: Arc<Vm<'bound>>,
diff --git a/drivers/gpu/drm/tyr/mmu.rs b/drivers/gpu/drm/tyr/mmu.rs
index c0341557d730..0ed20e03569c 100644
--- a/drivers/gpu/drm/tyr/mmu.rs
+++ b/drivers/gpu/drm/tyr/mmu.rs
@@ -10,7 +10,6 @@
//! The [`SlotManager`] manages the assignment of virtual address spaces to hardware address-space
//! (AS) slots. MMU commands such as updates and flushes are carried out by the
//! [`AddressSpaceManager`] which actually writes to the MMU registers.
-#![allow(dead_code)]
use core::ops::Range;
diff --git a/drivers/gpu/drm/tyr/slot.rs b/drivers/gpu/drm/tyr/slot.rs
index 3f58b23238ef..bc70ea49f9f5 100644
--- a/drivers/gpu/drm/tyr/slot.rs
+++ b/drivers/gpu/drm/tyr/slot.rs
@@ -20,7 +20,6 @@
//!
//! [SlotOperations]: crate::slot::SlotOperations
//! [SlotManager]: crate::slot::SlotManager
-#![allow(dead_code)]
use core::{
mem::take,
diff --git a/drivers/gpu/drm/tyr/tyr.rs b/drivers/gpu/drm/tyr/tyr.rs
index 92f6885cdaae..e7ec450bdc9c 100644
--- a/drivers/gpu/drm/tyr/tyr.rs
+++ b/drivers/gpu/drm/tyr/tyr.rs
@@ -9,6 +9,7 @@
mod driver;
mod file;
+mod fw;
mod gem;
mod gpu;
mod mmu;
diff --git a/drivers/gpu/drm/tyr/vm.rs b/drivers/gpu/drm/tyr/vm.rs
index 57929c2d3e94..54dfd5bbeab4 100644
--- a/drivers/gpu/drm/tyr/vm.rs
+++ b/drivers/gpu/drm/tyr/vm.rs
@@ -6,7 +6,6 @@
//! the illusion of owning the entire virtual address (VA) range, similar to CPU virtual memory.
//! Each virtual memory (VM) area is backed by ARM64 LPAE Stage 1 page tables and can be
//! mapped into hardware address space (AS) slots for GPU execution.
-#![allow(dead_code)]
use core::marker::PhantomData;
use core::ops::Range;
--
2.54.0
^ permalink raw reply related [flat|nested] 13+ messages in thread