starnix_modules_kgsl/

file.rs

// Copyright 2025 The Fuchsia Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

use crate::util::maur::{self, TaskWritable};
use crate::util::{KgslCmdBatchFlags, KgslContextFlags, KgslMemFlags};
use fdio::service_connect;
use kgsl_libmagma::{
    Buffer, Connection, Context, Device, QueryOutput, Semaphore, initialize_logging,
};
use kgsl_magma_params::{AdrenoKgslParams, MAGMA_QCOM_ADRENO_QUERY_KGSL_PARAMS};
use kgsl_strings::{ioctl_kgsl, kgsl_prop};
use magma::{
    MAGMA_MAP_FLAG_READ, MAGMA_MAP_FLAG_WRITE, MAGMA_QUERY_DEVICE_ID, MAGMA_QUERY_VENDOR_ID,
};
use range_alloc::RangeAllocator;
use starnix_core::mm::memory::MemoryObject;
use starnix_core::mm::{MappingName, MemoryAccessorExt, PAGE_SIZE};
use starnix_core::task::CurrentTask;
use starnix_core::vfs::{FileObject, FileOps, FsNode};
use starnix_core::{fileops_impl_dataless, fileops_impl_nonseekable, fileops_impl_noop_sync};
use starnix_logging::{log_error, log_info, log_warn, track_stub};
use starnix_sync::{
    KgslAllocatorLock, KgslContextsLock, KgslGpuObjsLock, KgslSyncSourcesLock, LockDepMutex,
};
use starnix_syscalls::{SUCCESS, SyscallArg, SyscallResult};
use starnix_uapi::device_id::DeviceId;
use starnix_uapi::errors::Errno;
use starnix_uapi::open_flags::OpenFlags;
use starnix_uapi::user_address::{UserAddress, UserRef};
use starnix_uapi::{errno, error, kgsl_command_object, kgsl_command_syncpoint, uapi};
use std::collections::HashMap;
use std::collections::hash_map::Entry;
use std::sync::atomic::{AtomicU32, Ordering};
use std::sync::{Arc, Once};

#[cfg(feature = "starnix-kgsl-debug")]
#[macro_export]
macro_rules! kgsl_debug {
    ($fmt:expr $(, $arg:expr)*) => {
        log_info!("kgsl: {}:{}: {}", file!(), line!(), format_args!($fmt $(, $arg)*));
    };
}

#[cfg(not(feature = "starnix-kgsl-debug"))]
#[macro_export]
macro_rules! kgsl_debug {
    ($($arg:tt)*) => {};
}

/// The maximum alignment of any buffer that can be allocated. The additional address space used by
/// many small buffers that only need page alignment is not a concern, as the total amount of
/// address space is very large.
const BUFFER_ALIGNMENT: u64 = 65536;

/// Helper trait to manage allocations in units of BUFFER_ALIGNMENT.
trait RangeAllocatorExt {
    fn create(size: u64) -> Self;
    fn allocate(&mut self, size: u64) -> Option<u64>;
    fn free(&mut self, gpuaddr: u64, size: u64);
}

impl RangeAllocatorExt for RangeAllocator<u64> {
    fn create(size: u64) -> Self {
        RangeAllocator::new(0..(size / BUFFER_ALIGNMENT))
    }

    fn allocate(&mut self, size: u64) -> Option<u64> {
        self.allocate_range(size.div_ceil(BUFFER_ALIGNMENT))
            .ok()
            .map(|r| r.start * BUFFER_ALIGNMENT)
    }

    fn free(&mut self, gpuaddr: u64, size: u64) {
        let start_unit = gpuaddr / BUFFER_ALIGNMENT;
        let units = size.div_ceil(BUFFER_ALIGNMENT);
        self.free_range(start_unit..(start_unit + units));
    }
}

pub struct KgslFile {
    // This member will be read in a future change. A linter attribute is used
    // instead of an underscore prefix so that field init shorthand still works.
    #[expect(dead_code)]
    device: Device,
    connection: Connection,
    adreno_kgsl_params: AdrenoKgslParams,
    // TODO(b/481419355): transition to id-map container once available
    syncsources: LockDepMutex<HashMap<u32, Semaphore>, KgslSyncSourcesLock>,
    next_syncsource_id: AtomicU32,
    // TODO(b/481419355): transition to id-map container once available
    gpuobjs: LockDepMutex<HashMap<u32, GpuObject>, KgslGpuObjsLock>,
    next_gpuobj_id: AtomicU32,
    allocator: LockDepMutex<RangeAllocator<u64>, KgslAllocatorLock>,
    shadow_properties: uapi::kgsl_shadowprop,
    // TODO(b/481419355): transition to id-map container once available
    contexts: LockDepMutex<HashMap<u32, Context>, KgslContextsLock>,
    next_context_id: AtomicU32,
}

struct GpuObject {
    buffer: Buffer,
    flags: u64,
    size: u64,
    mmapsize: u64,
    gpuaddr: u64,
}

fn map_flags(flags: KgslMemFlags) -> Result<u64, Errno> {
    match (flags.gpu_read_only(), flags.gpu_write_only()) {
        (true, false) => Ok(MAGMA_MAP_FLAG_READ),
        (false, true) => Ok(MAGMA_MAP_FLAG_WRITE),
        (false, false) => Ok(MAGMA_MAP_FLAG_READ | MAGMA_MAP_FLAG_WRITE),
        (true, true) => Err(errno!(EINVAL)),
    }
}

impl KgslFile {
    pub fn init() {
        match Self::init_magma_logging() {
            Ok(()) => log_info!("kgsl: magma logging enabled"),
            Err(()) => log_warn!("kgsl: magma logging failed to initialize"),
        };
    }

    fn init_magma_logging() -> Result<(), ()> {
        let (client, server) = zx::Channel::create();
        service_connect("/svc/fuchsia.logger.LogSink", server).map_err(|_| ())?;
        return initialize_logging(client);
    }

    fn import_device(path: &str) -> Result<Device, zx::Status> {
        let (client, server) = zx::Channel::create();
        service_connect(&path, server)?;
        let device = Device::from_channel(client).map_err(|_| zx::Status::INTERNAL)?;
        let QueryOutput::Value(vendor_id) =
            device.query(MAGMA_QUERY_VENDOR_ID).map_err(|_| zx::Status::INTERNAL)?
        else {
            return Err(zx::Status::INTERNAL);
        };
        let QueryOutput::Value(device_id) =
            device.query(MAGMA_QUERY_DEVICE_ID).map_err(|_| zx::Status::INTERNAL)?
        else {
            return Err(zx::Status::INTERNAL);
        };

        log_info!(
            "kgsl: magma device at {} is vendor {:#04x} device {:#04x}",
            path,
            vendor_id,
            device_id
        );
        Ok(device)
    }

    pub fn new_file(
        _current_task: &CurrentTask,
        _dev: DeviceId,
        _node: &FsNode,
        _flags: OpenFlags,
    ) -> Result<Box<dyn FileOps>, Errno> {
        static INIT: Once = Once::new();
        INIT.call_once(|| {
            Self::init();
        });
        let mut devices = std::fs::read_dir("/svc/fuchsia.gpu.magma.Service")
            .map_err(|_| errno!(ENXIO))?
            .filter_map(|x| x.ok())
            .filter_map(|entry| entry.path().join("device").into_os_string().into_string().ok())
            .filter_map(|path| Self::import_device(&path).ok());
        let device = devices.next().ok_or_else(|| errno!(ENXIO))?;
        let QueryOutput::Buffer(adreno_kgsl_params_vmo) =
            device.query(MAGMA_QCOM_ADRENO_QUERY_KGSL_PARAMS).map_err(|_| errno!(ENXIO))?
        else {
            return Err(errno!(ENXIO));
        };

        let adreno_kgsl_params = adreno_kgsl_params_vmo
            .read_to_object::<AdrenoKgslParams>(0)
            .map_err(|_| errno!(ENXIO))?;

        let connection = device.create_connection().map_err(|_| errno!(ENXIO))?;

        let mut allocator = RangeAllocator::create(adreno_kgsl_params.gpu_va64_size);

        // Reserve GPU secure VA space.
        allocator.allocate(adreno_kgsl_params.gpu_secure_va_size).ok_or_else(|| errno!(ENOMEM))?;

        // Shadow buffer must immediately follow the secure VA space, which spans 0..gpu_secure_va_size.
        let shadow_size = adreno_kgsl_params.device_shadow_size;
        let shadow_buffer = connection.create_buffer(shadow_size).map_err(|_| errno!(ENOMEM))?;
        // Reserve space for the shadow buffer, but assign it the end of the secure VA space.
        // This is necessary as the shadow must immediately follow the secure VA space which may
        // not be aligned to the overly-conservative BUFFER_ALIGNMENT.
        allocator.allocate(shadow_size).ok_or_else(|| errno!(ENOMEM))?;
        let shadow_gpuaddr = adreno_kgsl_params.gpu_secure_va_size;
        shadow_buffer
            .map(shadow_gpuaddr, 0, shadow_size, MAGMA_MAP_FLAG_READ | MAGMA_MAP_FLAG_WRITE)
            .map_err(|_| errno!(ENOMEM))?;
        let shadow = GpuObject {
            buffer: shadow_buffer,
            flags: adreno_kgsl_params.device_shadow_flags.into(),
            size: shadow_size,
            mmapsize: shadow_size,
            gpuaddr: shadow_gpuaddr,
        };
        let shadow_properties = uapi::kgsl_shadowprop {
            gpuaddr: shadow_gpuaddr.try_into().map_err(|_| errno!(ENXIO))?,
            size: shadow_size.try_into().map_err(|_| errno!(ENXIO))?,
            flags: adreno_kgsl_params.device_shadow_flags,
            ..Default::default()
        };

        let shadow_id = 1;
        let mut gpuobjs = HashMap::new();
        gpuobjs.insert(shadow_id, shadow);

        Ok(Box::new(Self {
            device,
            connection,
            adreno_kgsl_params,
            syncsources: Default::default(),
            next_syncsource_id: AtomicU32::new(1),
            gpuobjs: gpuobjs.into(),
            next_gpuobj_id: AtomicU32::new(shadow_id + 1),
            allocator: allocator.into(),
            shadow_properties,
            contexts: Default::default(),
            next_context_id: AtomicU32::new(1),
        }))
    }

    fn kgsl_device_getproperty(
        &self,
        current_task: &CurrentTask,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        let params_ref = maur::kgsl_device_getproperty::new(current_task, arg);
        let params = current_task.read_multi_arch_object(params_ref)?;
        kgsl_debug!("kgsl_device_getproperty {:?}", params);

        let params_size: usize = params.sizebytes.try_into().map_err(|_| errno!(EINVAL))?;
        // TODO(b/393160668): check params_size against all property types

        match params.type_ {
            uapi::KGSL_PROP_DEVICE_INFO => {
                let prop_value = uapi::kgsl_devinfo {
                    device_id: self.adreno_kgsl_params.device_id,
                    chip_id: self.adreno_kgsl_params.chip_id,
                    mmu_enabled: self.adreno_kgsl_params.mmu_enabled,
                    gmem_gpubaseaddr: 0, // This field is unused by the driver.
                    gpu_id: self.adreno_kgsl_params.gpu_id,
                    gmem_sizebytes: self.adreno_kgsl_params.gmem_sizebytes,
                    ..Default::default()
                };
                kgsl_debug!("KGSL_PROP_DEVICE_INFO: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_DEVICE_SHADOW => {
                let prop_value = self.shadow_properties;
                kgsl_debug!("KGSL_PROP_DEVICE_SHADOW: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_UCHE_GMEM_VADDR => {
                let prop_value = 0u32; // Unified cache unsupported.
                kgsl_debug!("KGSL_PROP_UCHE_GMEM_VADDR: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_UCODE_VERSION => {
                let prop_value = uapi::kgsl_ucode_version {
                    pfp: self.adreno_kgsl_params.ucode_version_pfp,
                    pm4: self.adreno_kgsl_params.ucode_version_pm4,
                    ..Default::default()
                };
                kgsl_debug!("KGSL_PROP_UCODE_VERSION: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_HIGHEST_BANK_BIT => {
                let prop_value = self.adreno_kgsl_params.highest_bank_bit;
                kgsl_debug!("KGSL_PROP_HIGHEST_BANK_BIT: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_DEVICE_BITNESS => {
                let prop_value = self.adreno_kgsl_params.device_bitness;
                kgsl_debug!("KGSL_PROP_DEVICE_BITNESS: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_DEVICE_QDSS_STM => {
                // This is intentionally zero-sized to indicate lack of support.
                let prop_value =
                    uapi::kgsl_qdss_stm_prop { gpuaddr: 0, size: 0, ..Default::default() };
                kgsl_debug!("KGSL_PROP_DEVICE_QDSS_STM: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_MIN_ACCESS_LENGTH => {
                let prop_value = self.adreno_kgsl_params.min_access_length;
                kgsl_debug!("KGSL_PROP_MIN_ACCESS_LENGTH: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_UBWC_MODE => {
                let prop_value = self.adreno_kgsl_params.ubwc_mode;
                kgsl_debug!("KGSL_PROP_UBWC_MODE: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_DEVICE_QTIMER => {
                // This is intentionally zero-sized to indicate lack of support.
                let prop_value =
                    uapi::kgsl_qtimer_prop { gpuaddr: 0, size: 0, ..Default::default() };
                kgsl_debug!("KGSL_PROP_DEVICE_QTIMER: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_SECURE_BUFFER_ALIGNMENT => {
                let prop_value = self.adreno_kgsl_params.secure_buf_alignment;
                kgsl_debug!("KGSL_PROP_SECURE_BUFFER_ALIGNMENT: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_SECURE_CTXT_SUPPORT => {
                let prop_value = self.adreno_kgsl_params.secure_ctxt_support;
                kgsl_debug!("KGSL_PROP_SECURE_CTXT_SUPPORT: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_SPEED_BIN => {
                let prop_value = 0u64; // Default speed bin.
                kgsl_debug!("KGSL_PROP_SPEED_BIN: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_GAMING_BIN => {
                kgsl_debug!("KGSL_PROP_GAMING_BIN returning EINVAL");
                error!(EINVAL, "gaming bin unsupported")
            }
            uapi::KGSL_PROP_GPU_MODEL => {
                if params_size < self.adreno_kgsl_params.gpu_model.len() {
                    return error!(EINVAL);
                }
                let prop_value = self.adreno_kgsl_params.gpu_model;
                kgsl_debug!("KGSL_PROP_GPU_MODEL: {:?}", prop_value);
                let result_ref = UserRef::from(UserAddress::from(params.value));
                current_task.write_object(result_ref, &prop_value)?;
                Ok(SUCCESS)
            }
            uapi::KGSL_PROP_VK_DEVICE_ID => {
                let prop_value = self.adreno_kgsl_params.vk_device_id;
                kgsl_debug!("KGSL_PROP_VK_DEVICE_ID: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_IS_LPAC_ENABLED => {
                let prop_value = 0u32; // Asynchronous compute unsupported.
                kgsl_debug!("KGSL_PROP_IS_LPAC_ENABLED: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_GPU_VA64_SIZE => {
                let prop_value = self.adreno_kgsl_params.gpu_va64_size;
                kgsl_debug!("KGSL_PROP_GPU_VA64_SIZE: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_IS_RAYTRACING_ENABLED => {
                let prop_value = 0u32; // Raytracing unsupported.
                kgsl_debug!("KGSL_PROP_IS_RAYTRACING_ENABLED: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_IS_FASTBLEND_ENABLED => {
                let prop_value = 0u32; // Fast blend unsupported.
                kgsl_debug!("KGSL_PROP_IS_FASTBLEND_ENABLED: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            uapi::KGSL_PROP_UCHE_TRAP_BASE => {
                kgsl_debug!("KGSL_PROP_UCHE_TRAP_BASE returning EINVAL");
                error!(EINVAL, "uche_trap_base unset")
            }
            uapi::KGSL_PROP_GPU_SECURE_VA_SIZE => {
                let prop_value = self.adreno_kgsl_params.gpu_secure_va_size;
                kgsl_debug!("KGSL_PROP_GPU_SECURE_VA_SIZE: {:?}", prop_value);
                prop_value.write(&current_task, params.value)
            }
            _ => {
                track_stub!(TODO("https://fxbug.dev/393160668"), "kgsl property", params.type_);
                log_error!("kgsl: unimplemented GetProperty type {}", kgsl_prop(params.type_));
                error!(ENOTSUP)
            }
        }
    }

    fn kgsl_gpuobj_alloc(
        &self,
        current_task: &CurrentTask,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        let params_ref = maur::kgsl_gpuobj_alloc::new(current_task, arg);
        let mut params = current_task.read_multi_arch_object(params_ref)?;
        kgsl_debug!("kgsl_gpuobj_alloc {:?}", params);

        let flags = KgslMemFlags::try_from(params.flags).map_err(|bits| {
            log_error!("kgsl: unknown memory flags {:#x}", bits);
            errno!(EINVAL)
        })?;
        if BUFFER_ALIGNMENT % (1 << flags.align_bits()) != 0 {
            log_error!("kgsl: unsupported alignment {}", flags.align_bits());
            return error!(ENOTSUP);
        }

        let buffer = self.connection.create_buffer(params.size).map_err(|_| errno!(ENOMEM))?;
        let size = buffer.size();

        let gpuaddr = self.allocator.lock().allocate(size).ok_or_else(|| errno!(ENOMEM))?;
        buffer.map(gpuaddr, 0, size, map_flags(flags)?).map_err(|_| errno!(ENOMEM))?;

        let id = self.next_gpuobj_id.fetch_add(1, Ordering::Relaxed);
        if id == 0 {
            log_error!("kgsl: gpuobj ids exhausted");
            return error!(ENOMEM);
        }

        let gpuobj = GpuObject { buffer, flags: params.flags, size, mmapsize: size, gpuaddr };

        self.gpuobjs.lock().insert(id, gpuobj);

        params.size = size;
        params.mmapsize = size;
        params.id = id;

        current_task.write_multi_arch_object(params_ref, params)?;
        Ok(SUCCESS)
    }

    fn kgsl_gpuobj_free(
        &self,
        current_task: &CurrentTask,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        let params_ref = maur::kgsl_gpuobj_free::new(current_task, arg);
        let params = current_task.read_multi_arch_object(params_ref)?;
        kgsl_debug!("kgsl_gpuobj_free {:?}", params);

        if let Entry::Occupied(entry) = self.gpuobjs.lock().entry(params.id) {
            self.allocator.lock().free(entry.get().gpuaddr, entry.get().size);
            entry.remove();
            Ok(SUCCESS)
        } else {
            error!(EINVAL)
        }
    }

    fn kgsl_gpuobj_info(
        &self,
        current_task: &CurrentTask,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        let params_ref = maur::kgsl_gpuobj_info::new(current_task, arg);
        let mut params = current_task.read_multi_arch_object(params_ref)?;
        kgsl_debug!("kgsl_gpuobj_info {:?}", params);

        let gpuobjs = self.gpuobjs.lock();
        let gpuobj = gpuobjs.get(&params.id).ok_or_else(|| errno!(EINVAL))?;

        params.gpuaddr = gpuobj.gpuaddr;
        params.size = gpuobj.size;
        params.flags = gpuobj.flags;
        params.va_len = gpuobj.size;
        params.va_addr = 0;

        current_task.write_multi_arch_object(params_ref, params)?;
        Ok(SUCCESS)
    }

    fn kgsl_syncsource_create(
        &self,
        current_task: &CurrentTask,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        let params_ref = maur::kgsl_syncsource_create::new(current_task, arg);
        let mut params = current_task.read_multi_arch_object(params_ref)?;
        kgsl_debug!("kgsl_syncsource_create {:?}", params);

        let semaphore = self.connection.create_semaphore().map_err(|_| errno!(ENOMEM))?;
        let id = self.next_syncsource_id.fetch_add(1, Ordering::Relaxed);
        if id == 0 {
            // fetch_add wraps to zero on overflow. Transitioning to id-map container
            // will avoid this issue.
            log_error!("kgsl: ids exhausted");
            return error!(ENOMEM);
        }
        self.syncsources.lock().insert(id, semaphore);

        params.id = id;

        current_task.write_multi_arch_object(params_ref, params)?;
        Ok(SUCCESS)
    }

    fn kgsl_syncsource_destroy(
        &self,
        current_task: &CurrentTask,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        let params_ref = maur::kgsl_syncsource_destroy::new(current_task, arg);
        let params = current_task.read_multi_arch_object(params_ref)?;
        kgsl_debug!("kgsl_syncsource_destroy {:?}", params);

        if self.syncsources.lock().remove(&params.id).is_some() {
            Ok(SUCCESS)
        } else {
            error!(EINVAL)
        }
    }

    fn kgsl_drawctxt_create(
        &self,
        current_task: &CurrentTask,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        let params_ref = maur::kgsl_drawctxt_create::new(current_task, arg);
        let mut params = current_task.read_multi_arch_object(params_ref)?;
        kgsl_debug!("kgsl_drawctxt_create {:?}", params);
        let flags = KgslContextFlags::try_from(params.flags).map_err(|bits| {
            log_error!("kgsl: unknown context flags {:#x}", bits);
            errno!(EINVAL)
        })?;
        let context =
            self.connection.create_context(flags.priority().into()).map_err(|_| errno!(ENXIO))?;
        let id = self.next_context_id.fetch_add(1, Ordering::Relaxed);
        if id == 0 {
            // fetch_add wraps to zero on overflow. Transitioning to id-map container
            // will avoid this issue.
            log_error!("kgsl: ids exhausted");
            return error!(ENOMEM);
        }
        self.contexts.lock().insert(id, context);
        params.drawctxt_id = id;
        current_task.write_multi_arch_object(params_ref, params)?;
        Ok(SUCCESS)
    }

    fn kgsl_drawctxt_destroy(
        &self,
        current_task: &CurrentTask,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        let params_ref = maur::kgsl_drawctxt_destroy::new(current_task, arg);
        let params = current_task.read_multi_arch_object(params_ref)?;
        kgsl_debug!("kgsl_drawctxt_destroy {:?}", params);
        if self.contexts.lock().remove(&params.drawctxt_id).is_some() {
            Ok(SUCCESS)
        } else {
            error!(EINVAL)
        }
    }

    fn kgsl_gpu_command(
        &self,
        current_task: &CurrentTask,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        let params_ref = maur::kgsl_gpu_command::new(current_task, arg);
        let params = current_task.read_multi_arch_object(params_ref)?;
        kgsl_debug!("kgsl_gpu_command {:?}", params);
        let contexts = self.contexts.lock();
        let context = contexts.get(&params.context_id).ok_or_else(|| errno!(EINVAL))?;

        let cmds = current_task.read_objects_to_vec::<kgsl_command_object>(
            UserRef::from(UserAddress::from(params.cmdlist)),
            params.numcmds as usize,
        )?;
        kgsl_debug!("kgsl_gpu_command cmds {:?}", cmds);
        let objs = current_task.read_objects_to_vec::<kgsl_command_object>(
            UserRef::from(UserAddress::from(params.objlist)),
            params.numobjs as usize,
        )?;
        kgsl_debug!("kgsl_gpu_command objs {:?}", objs);
        let syncs = current_task.read_objects_to_vec::<kgsl_command_syncpoint>(
            UserRef::from(UserAddress::from(params.synclist)),
            params.numsyncs as usize,
        )?;
        kgsl_debug!("kgsl_gpu_command syncs {:?}", syncs);
        if syncs.len() > 0 {
            // TODO(b/359831679): implement syncpoints
            log_error!("kgsl: syncs not supported yet");
            return error!(ENOTSUP);
        }

        if params.flags != 0 {
            let flags = KgslCmdBatchFlags::try_from(params.flags).map_err(|bits| {
                log_error!("kgsl: unknown command flags {:#x}", bits);
                errno!(EINVAL)
            })?;
            log_warn!("kgsl: unsupported flags {:?}", flags);
        }

        let gpuobjs = self.gpuobjs.lock();

        let to_exec_resources = |objects: &[kgsl_command_object],
                                 kind: &str|
         -> Result<Vec<kgsl_libmagma::ExecResource>, Errno> {
            // The caller does not always set the id field of the command object, and the gpuaddr
            // may not be the start of the buffer, so we have to search for the object manually.
            // If this turns out to be a bottleneck, we may want to try an initial base address
            // search before falling back to a full search and/or cache the results of the search.
            // TODO(b/393160668): profile and optimize as necessary
            objects
                .iter()
                .map(|obj| {
                    let gpuobj = gpuobjs
                        .values()
                        .find(|o| obj.gpuaddr >= o.gpuaddr && obj.gpuaddr < o.gpuaddr + o.size)
                        .ok_or_else(|| {
                            log_error!("kgsl: {} gpuaddr {:#x} not found", kind, obj.gpuaddr);
                            errno!(EINVAL)
                        })?;
                    Ok(kgsl_libmagma::ExecResource {
                        buffer: gpuobj.buffer.clone(),
                        offset: (obj.gpuaddr - gpuobj.gpuaddr) + obj.offset,
                        length: obj.size,
                    })
                })
                .collect()
        };

        let magma_resources = to_exec_resources(&objs, "resource")?;
        let magma_command_buffers = to_exec_resources(&cmds, "command")?;

        context
            .execute_command(magma_command_buffers, magma_resources, vec![], vec![], 0)
            .map_err(|_| errno!(EINVAL))?;

        Ok(SUCCESS)
    }
}

impl Drop for KgslFile {
    fn drop(&mut self) {}
}

impl FileOps for KgslFile {
    fileops_impl_dataless!();
    fileops_impl_nonseekable!();
    fileops_impl_noop_sync!();

    fn ioctl(
        &self,
        _file: &FileObject,
        current_task: &CurrentTask,
        request: u32,
        arg: SyscallArg,
    ) -> Result<SyscallResult, Errno> {
        // Special ioctl to signal container to use kgsl.
        // TODO(b/429239527): remove after transitioned
        const IOCTL_KGSL_ENABLE: u32 = 42;
        if request == IOCTL_KGSL_ENABLE {
            if cfg!(not(feature = "starnix-kgsl-enable")) {
                log_info!("kgsl: suppressing further use of kgsl");
                return error!(ENXIO);
            }
            return Ok(SUCCESS);
        }
        match crate::util::canonicalize_ioctl_request(current_task, request) {
            uapi::IOCTL_KGSL_DEVICE_GETPROPERTY => self.kgsl_device_getproperty(current_task, arg),
            uapi::IOCTL_KGSL_GPUOBJ_ALLOC => self.kgsl_gpuobj_alloc(current_task, arg),
            uapi::IOCTL_KGSL_GPUOBJ_FREE => self.kgsl_gpuobj_free(current_task, arg),
            uapi::IOCTL_KGSL_GPUOBJ_INFO => self.kgsl_gpuobj_info(current_task, arg),
            uapi::IOCTL_KGSL_SYNCSOURCE_CREATE => self.kgsl_syncsource_create(current_task, arg),
            uapi::IOCTL_KGSL_SYNCSOURCE_DESTROY => self.kgsl_syncsource_destroy(current_task, arg),
            uapi::IOCTL_KGSL_DRAWCTXT_CREATE => self.kgsl_drawctxt_create(current_task, arg),
            uapi::IOCTL_KGSL_DRAWCTXT_DESTROY => self.kgsl_drawctxt_destroy(current_task, arg),
            uapi::IOCTL_KGSL_GPU_COMMAND => self.kgsl_gpu_command(current_task, arg),
            _ => {
                track_stub!(TODO("https://fxbug.dev/393160668"), "kgsl ioctl", request);
                log_error!("kgsl: unimplemented ioctl {}", ioctl_kgsl(request));
                error!(ENOTSUP)
            }
        }
    }

    fn mmap(
        &self,
        file: &FileObject,
        current_task: &CurrentTask,
        addr: starnix_core::mm::DesiredAddress,
        memory_offset: u64, // Callers encode the buffer ID using this field.
        length: usize,
        prot_flags: starnix_core::mm::ProtectionFlags,
        mapping_options: starnix_core::mm::MappingOptions,
        _filename: starnix_core::vfs::NamespaceNode,
    ) -> Result<UserAddress, Errno> {
        kgsl_debug!("mmap {:?} {:?} {:?} {:?}", addr, memory_offset, length, prot_flags);
        let id = (memory_offset / *PAGE_SIZE) as u32;
        let gpuobjs = self.gpuobjs.lock();
        let gpuobj = gpuobjs.get(&id).ok_or_else(|| errno!(EINVAL))?;
        if length as u64 > gpuobj.mmapsize {
            return error!(EINVAL);
        }
        let handle = gpuobj.buffer.get_handle().map_err(|_| errno!(ENXIO))?;
        let vmo = zx::Vmo::from(handle);
        let memory = Arc::new(MemoryObject::from(vmo).with_zx_name(b"starnix:kgsl"));
        // The memory manager persists the memory object until the client exits or calls munmap.
        current_task.mm()?.map_memory(
            addr,
            memory,
            0,
            length,
            prot_flags,
            file.max_access_for_memory_mapping(),
            mapping_options,
            MappingName::None,
        )
    }
}