wl.rs

// Copyright 2017 The ChromiumOS Authors
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

//! This module implements the virtio wayland used by the guest to access the host's wayland server.
//!
//! The virtio wayland protocol is done over two queues: `in` and `out`. The `in` queue is used for
//! sending commands to the guest that are generated by the host, usually messages from the wayland
//! server. The `out` queue is for commands from the guest, usually requests to allocate shared
//! memory, open a wayland server connection, or send data over an existing connection.
//!
//! Each `WlVfd` represents one virtual file descriptor created by either the guest or the host.
//! Virtual file descriptors contain actual file descriptors, either a shared memory file descriptor
//! or a unix domain socket to the wayland server. In the shared memory case, there is also an
//! associated slot that indicates which hypervisor memory slot the memory is installed into, as
//! well as a page frame number that the guest can access the memory from.
//!
//! The types starting with `Ctrl` are structures representing the virtio wayland protocol "on the
//! wire." They are decoded and executed in the `execute` function and encoded as some variant of
//! `WlResp` for responses.
//!
//! There is one `WlState` instance that contains every known vfd and the current state of `in`
//! queue. The `in` queue requires extra state to buffer messages to the guest in case the `in`
//! queue is already full. The `WlState` also has a control socket necessary to fulfill certain
//! requests, such as those registering guest memory.
//!
//! The `Worker` is responsible for the poll loop over all possible events, encoding/decoding from
//! the virtio queue, and routing messages in and out of `WlState`. Possible events include the kill
//! event, available descriptors on the `in` or `out` queue, and incoming data on any vfd's socket.

use std::cell::RefCell;
use std::collections::btree_map::Entry;
use std::collections::BTreeMap;
use std::collections::BTreeSet;
use std::collections::VecDeque;
use std::convert::From;
use std::error::Error as StdError;
use std::fmt;
use std::fs::File;
use std::io;
use std::io::Read;
use std::io::Seek;
use std::io::SeekFrom;
use std::io::Write;
use std::mem::size_of;
#[cfg(feature = "gbm")]
use std::os::raw::c_uint;
#[cfg(feature = "gbm")]
use std::os::raw::c_ulonglong;
use std::os::unix::net::UnixStream;
use std::path::Path;
use std::path::PathBuf;
use std::rc::Rc;
use std::result;
use std::time::Duration;

use anyhow::anyhow;
use anyhow::Context;
use base::error;
#[cfg(feature = "gbm")]
use base::ioctl_iow_nr;
use base::ioctl_iowr_nr;
use base::ioctl_with_ref;
use base::linux::SharedMemoryLinux;
use base::pagesize;
use base::pipe;
use base::round_up_to_page_size;
use base::unix::FileFlags;
use base::warn;
use base::AsRawDescriptor;
use base::Error;
use base::Event;
use base::EventToken;
use base::EventType;
#[cfg(feature = "gpu")]
use base::IntoRawDescriptor;
#[cfg(feature = "gbm")]
use base::MemoryMappingBuilder;
#[cfg(feature = "gbm")]
use base::MmapError;
use base::Protection;
use base::RawDescriptor;
use base::Result;
use base::SafeDescriptor;
use base::ScmSocket;
use base::SharedMemory;
use base::Tube;
use base::TubeError;
use base::VolatileMemoryError;
use base::WaitContext;
use base::WorkerThread;
use data_model::Le32;
use data_model::Le64;
use hypervisor::MemCacheType;
#[cfg(feature = "gbm")]
use libc::EBADF;
#[cfg(feature = "gbm")]
use libc::EINVAL;
#[cfg(feature = "gbm")]
use libc::ENOSYS;
use remain::sorted;
use resources::address_allocator::AddressAllocator;
use resources::AddressRange;
use resources::Alloc;
#[cfg(feature = "gbm")]
use rutabaga_gfx::DrmFormat;
#[cfg(feature = "gbm")]
use rutabaga_gfx::ImageAllocationInfo;
#[cfg(feature = "gbm")]
use rutabaga_gfx::ImageMemoryRequirements;
#[cfg(feature = "gbm")]
use rutabaga_gfx::RutabagaDescriptor;
#[cfg(feature = "gbm")]
use rutabaga_gfx::RutabagaError;
#[cfg(feature = "gbm")]
use rutabaga_gfx::RutabagaGralloc;
#[cfg(feature = "gbm")]
use rutabaga_gfx::RutabagaGrallocBackendFlags;
#[cfg(feature = "gbm")]
use rutabaga_gfx::RutabagaGrallocFlags;
#[cfg(feature = "gbm")]
use rutabaga_gfx::RutabagaIntoRawDescriptor;
#[cfg(feature = "gbm")]
use rutabaga_gfx::RUTABAGA_MAP_CACHE_CACHED;
#[cfg(feature = "gbm")]
use rutabaga_gfx::RUTABAGA_MAP_CACHE_MASK;
use static_assertions::const_assert_eq;
use thiserror::Error as ThisError;
use vm_control::VmMemorySource;
use vm_memory::GuestMemory;
use vm_memory::GuestMemoryError;
use zerocopy::FromBytes;
use zerocopy::Immutable;
use zerocopy::IntoBytes;
use zerocopy::KnownLayout;

#[cfg(feature = "gpu")]
use super::resource_bridge::get_resource_info;
#[cfg(feature = "gpu")]
use super::resource_bridge::BufferInfo;
#[cfg(feature = "gpu")]
use super::resource_bridge::ResourceBridgeError;
#[cfg(feature = "gpu")]
use super::resource_bridge::ResourceInfo;
#[cfg(feature = "gpu")]
use super::resource_bridge::ResourceRequest;
use super::DeviceType;
use super::Interrupt;
use super::Queue;
use super::Reader;
use super::SharedMemoryMapper;
use super::SharedMemoryRegion;
use super::VirtioDevice;
use super::Writer;
use crate::virtio::device_constants::wl::VIRTIO_WL_F_SEND_FENCES;
use crate::virtio::device_constants::wl::VIRTIO_WL_F_TRANS_FLAGS;
use crate::virtio::device_constants::wl::VIRTIO_WL_F_USE_SHMEM;

const QUEUE_SIZE: u16 = 256;
const QUEUE_SIZES: &[u16] = &[QUEUE_SIZE, QUEUE_SIZE];

const VIRTWL_SEND_MAX_ALLOCS: usize = 28;
const VIRTIO_WL_CMD_VFD_NEW: u32 = 256;
const VIRTIO_WL_CMD_VFD_CLOSE: u32 = 257;
const VIRTIO_WL_CMD_VFD_SEND: u32 = 258;
const VIRTIO_WL_CMD_VFD_RECV: u32 = 259;
const VIRTIO_WL_CMD_VFD_NEW_CTX: u32 = 260;
const VIRTIO_WL_CMD_VFD_NEW_PIPE: u32 = 261;
const VIRTIO_WL_CMD_VFD_HUP: u32 = 262;
#[cfg(feature = "gbm")]
const VIRTIO_WL_CMD_VFD_NEW_DMABUF: u32 = 263;
#[cfg(feature = "gbm")]
const VIRTIO_WL_CMD_VFD_DMABUF_SYNC: u32 = 264;
#[cfg(feature = "gpu")]
const VIRTIO_WL_CMD_VFD_SEND_FOREIGN_ID: u32 = 265;
const VIRTIO_WL_CMD_VFD_NEW_CTX_NAMED: u32 = 266;
const VIRTIO_WL_RESP_OK: u32 = 4096;
const VIRTIO_WL_RESP_VFD_NEW: u32 = 4097;
#[cfg(feature = "gbm")]
const VIRTIO_WL_RESP_VFD_NEW_DMABUF: u32 = 4098;
const VIRTIO_WL_RESP_ERR: u32 = 4352;
const VIRTIO_WL_RESP_OUT_OF_MEMORY: u32 = 4353;
const VIRTIO_WL_RESP_INVALID_ID: u32 = 4354;
const VIRTIO_WL_RESP_INVALID_TYPE: u32 = 4355;
const VIRTIO_WL_RESP_INVALID_FLAGS: u32 = 4356;
const VIRTIO_WL_RESP_INVALID_CMD: u32 = 4357;
const VIRTIO_WL_VFD_WRITE: u32 = 0x1;
const VIRTIO_WL_VFD_READ: u32 = 0x2;
const VIRTIO_WL_VFD_MAP: u32 = 0x2;
const VIRTIO_WL_VFD_CONTROL: u32 = 0x4;
const VIRTIO_WL_VFD_FENCE: u32 = 0x8;

const NEXT_VFD_ID_BASE: u32 = 0x40000000;
const VFD_ID_HOST_MASK: u32 = NEXT_VFD_ID_BASE;
// Each in-vq buffer is one page, so we need to leave space for the control header and the maximum
// number of allocs.
const IN_BUFFER_LEN: usize =
    0x1000 - size_of::<CtrlVfdRecv>() - VIRTWL_SEND_MAX_ALLOCS * size_of::<Le32>();

#[cfg(feature = "gbm")]
const VIRTIO_WL_VFD_DMABUF_SYNC_VALID_FLAG_MASK: u32 = 0x7;

#[cfg(feature = "gbm")]
const DMA_BUF_IOCTL_BASE: c_uint = 0x62;
#[cfg(feature = "gbm")]
const DMA_BUF_SYNC_WRITE: c_uint = 0x2;
#[cfg(feature = "gbm")]
const DMA_BUF_SYNC_END: c_uint = 0x4;

#[cfg(feature = "gbm")]
#[repr(C)]
#[derive(Copy, Clone)]
struct dma_buf_sync {
    flags: c_ulonglong,
}

#[cfg(feature = "gbm")]
ioctl_iow_nr!(DMA_BUF_IOCTL_SYNC, DMA_BUF_IOCTL_BASE, 0, dma_buf_sync);

#[repr(C)]
#[derive(Copy, Clone, Default)]
struct sync_file_info {
    name: [u8; 32],
    status: i32,
    flags: u32,
    num_fences: u32,
    pad: u32,
    sync_fence_info: u64,
}

ioctl_iowr_nr!(SYNC_IOC_FILE_INFO, 0x3e, 4, sync_file_info);

fn is_fence(f: &File) -> bool {
    let info = sync_file_info::default();
    // SAFETY:
    // Safe as f is a valid file
    unsafe { ioctl_with_ref(f, SYNC_IOC_FILE_INFO, &info) == 0 }
}

#[cfg(feature = "gbm")]
#[derive(Debug, Default)]
struct GpuMemoryPlaneDesc {
    stride: u32,
    offset: u32,
}

#[cfg(feature = "gbm")]
#[derive(Debug, Default)]
struct GpuMemoryDesc {
    planes: [GpuMemoryPlaneDesc; 3],
}

const VIRTIO_WL_CTRL_VFD_SEND_KIND_LOCAL: u32 = 0;
const VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU: u32 = 1;
const VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU_FENCE: u32 = 2;
const VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU_SIGNALED_FENCE: u32 = 3;

const VIRTIO_WL_PFN_SHIFT: u32 = 12;

fn encode_vfd_new(
    writer: &mut Writer,
    resp: bool,
    vfd_id: u32,
    flags: u32,
    pfn: u64,
    size: u32,
) -> WlResult<()> {
    let ctrl_vfd_new = CtrlVfdNew {
        hdr: CtrlHeader {
            type_: Le32::from(if resp {
                VIRTIO_WL_RESP_VFD_NEW
            } else {
                VIRTIO_WL_CMD_VFD_NEW
            }),
            flags: Le32::from(0),
        },
        id: Le32::from(vfd_id),
        flags: Le32::from(flags),
        pfn: Le64::from(pfn),
        size: Le32::from(size),
        padding: Default::default(),
    };

    writer
        .write_obj(ctrl_vfd_new)
        .map_err(WlError::WriteResponse)
}

#[cfg(feature = "gbm")]
fn encode_vfd_new_dmabuf(
    writer: &mut Writer,
    vfd_id: u32,
    flags: u32,
    pfn: u64,
    size: u32,
    desc: GpuMemoryDesc,
) -> WlResult<()> {
    let ctrl_vfd_new_dmabuf = CtrlVfdNewDmabuf {
        hdr: CtrlHeader {
            type_: Le32::from(VIRTIO_WL_RESP_VFD_NEW_DMABUF),
            flags: Le32::from(0),
        },
        id: Le32::from(vfd_id),
        flags: Le32::from(flags),
        pfn: Le64::from(pfn),
        size: Le32::from(size),
        width: Le32::from(0),
        height: Le32::from(0),
        format: Le32::from(0),
        stride0: Le32::from(desc.planes[0].stride),
        stride1: Le32::from(desc.planes[1].stride),
        stride2: Le32::from(desc.planes[2].stride),
        offset0: Le32::from(desc.planes[0].offset),
        offset1: Le32::from(desc.planes[1].offset),
        offset2: Le32::from(desc.planes[2].offset),
    };

    writer
        .write_obj(ctrl_vfd_new_dmabuf)
        .map_err(WlError::WriteResponse)
}

fn encode_vfd_recv(writer: &mut Writer, vfd_id: u32, data: &[u8], vfd_ids: &[u32]) -> WlResult<()> {
    let ctrl_vfd_recv = CtrlVfdRecv {
        hdr: CtrlHeader {
            type_: Le32::from(VIRTIO_WL_CMD_VFD_RECV),
            flags: Le32::from(0),
        },
        id: Le32::from(vfd_id),
        vfd_count: Le32::from(vfd_ids.len() as u32),
    };
    writer
        .write_obj(ctrl_vfd_recv)
        .map_err(WlError::WriteResponse)?;

    for &recv_vfd_id in vfd_ids.iter() {
        writer
            .write_obj(Le32::from(recv_vfd_id))
            .map_err(WlError::WriteResponse)?;
    }

    writer.write_all(data).map_err(WlError::WriteResponse)
}

fn encode_vfd_hup(writer: &mut Writer, vfd_id: u32) -> WlResult<()> {
    let ctrl_vfd_new = CtrlVfd {
        hdr: CtrlHeader {
            type_: Le32::from(VIRTIO_WL_CMD_VFD_HUP),
            flags: Le32::from(0),
        },
        id: Le32::from(vfd_id),
    };

    writer
        .write_obj(ctrl_vfd_new)
        .map_err(WlError::WriteResponse)
}

fn encode_resp(writer: &mut Writer, resp: WlResp) -> WlResult<()> {
    match resp {
        WlResp::VfdNew {
            id,
            flags,
            pfn,
            size,
            resp,
        } => encode_vfd_new(writer, resp, id, flags, pfn, size),
        #[cfg(feature = "gbm")]
        WlResp::VfdNewDmabuf {
            id,
            flags,
            pfn,
            size,
            desc,
        } => encode_vfd_new_dmabuf(writer, id, flags, pfn, size, desc),
        WlResp::VfdRecv { id, data, vfds } => encode_vfd_recv(writer, id, data, vfds),
        WlResp::VfdHup { id } => encode_vfd_hup(writer, id),
        r => writer
            .write_obj(Le32::from(r.get_code()))
            .map_err(WlError::WriteResponse),
    }
}

#[allow(dead_code)]
#[sorted]
#[derive(ThisError, Debug)]
enum WlError {
    #[error("overflow in calculation")]
    CheckedOffset,
    #[error("failed to synchronize DMABuf access: {0}")]
    DmabufSync(io::Error),
    #[error("failed to create shared memory from descriptor: {0}")]
    FromSharedMemory(Error),
    #[error("failed to get seals: {0}")]
    GetSeals(Error),
    #[error("gralloc error: {0}")]
    #[cfg(feature = "gbm")]
    GrallocError(#[from] RutabagaError),
    #[error("access violation in guest memory: {0}")]
    GuestMemory(#[from] GuestMemoryError),
    #[error("invalid string: {0}")]
    InvalidString(std::str::Utf8Error),
    #[error("failed to create shared memory allocation: {0}")]
    NewAlloc(Error),
    #[error("failed to create pipe: {0}")]
    NewPipe(Error),
    #[error("error parsing descriptor: {0}")]
    ParseDesc(io::Error),
    #[error("failed to read a pipe: {0}")]
    ReadPipe(io::Error),
    #[error("failed to recv on a socket: {0}")]
    RecvVfd(io::Error),
    #[error("failed to send on a socket: {0}")]
    SendVfd(io::Error),
    #[error("shmem mapper failure: {0}")]
    ShmemMapperError(anyhow::Error),
    #[error("failed to connect socket: {0}")]
    SocketConnect(io::Error),
    #[error("failed to set socket as non-blocking: {0}")]
    SocketNonBlock(io::Error),
    #[error("unknown socket name: {0}")]
    UnknownSocketName(String),
    #[error("invalid response from parent VM")]
    VmBadResponse,
    #[error("failed to control parent VM: {0}")]
    VmControl(TubeError),
    #[error("access violating in guest volatile memory: {0}")]
    VolatileMemory(#[from] VolatileMemoryError),
    #[error("failed to listen to descriptor on wait context: {0}")]
    WaitContextAdd(Error),
    #[error("failed to write to a pipe: {0}")]
    WritePipe(io::Error),
    #[error("failed to write response: {0}")]
    WriteResponse(io::Error),
}

type WlResult<T> = result::Result<T, WlError>;

pub const WL_SHMEM_ID: u8 = 0;
pub const WL_SHMEM_SIZE: u64 = 1 << 32;

struct VmRequesterState {
    mapper: Box<dyn SharedMemoryMapper>,
    #[cfg(feature = "gbm")]
    gralloc: RutabagaGralloc,

    // Allocator for shm address space
    address_allocator: AddressAllocator,

    // Map of existing mappings in the shm address space
    allocs: BTreeMap<u64 /* offset */, Alloc>,

    // The id for the next shmem allocation
    next_alloc: usize,
}

#[derive(Clone)]
struct VmRequester {
    state: Rc<RefCell<VmRequesterState>>,
}

// The following are wrappers to avoid base dependencies in the rutabaga crate
#[cfg(feature = "gbm")]
fn to_safe_descriptor(r: RutabagaDescriptor) -> SafeDescriptor {
    // SAFETY:
    // Safe because we own the SafeDescriptor at this point.
    unsafe { base::FromRawDescriptor::from_raw_descriptor(r.into_raw_descriptor()) }
}

impl VmRequester {
    fn new(
        mapper: Box<dyn SharedMemoryMapper>,
        #[cfg(feature = "gbm")] gralloc: RutabagaGralloc,
    ) -> VmRequester {
        VmRequester {
            state: Rc::new(RefCell::new(VmRequesterState {
                mapper,
                #[cfg(feature = "gbm")]
                gralloc,
                address_allocator: AddressAllocator::new(
                    AddressRange::from_start_and_size(0, WL_SHMEM_SIZE).unwrap(),
                    Some(pagesize() as u64),
                    None,
                )
                .expect("failed to create allocator"),
                allocs: BTreeMap::new(),
                next_alloc: 0,
            })),
        }
    }

    fn unregister_memory(&self, offset: u64) -> WlResult<()> {
        let mut state = self.state.borrow_mut();
        state
            .mapper
            .remove_mapping(offset)
            .map_err(WlError::ShmemMapperError)?;
        let alloc = state
            .allocs
            .remove(&offset)
            .context("unknown offset")
            .map_err(WlError::ShmemMapperError)?;
        state
            .address_allocator
            .release(alloc)
            .expect("corrupt address space");
        Ok(())
    }

    #[cfg(feature = "gbm")]
    fn allocate_and_register_gpu_memory(
        &self,
        width: u32,
        height: u32,
        format: u32,
    ) -> WlResult<(u64, SafeDescriptor, ImageMemoryRequirements)> {
        let mut state = self.state.borrow_mut();

        let img = ImageAllocationInfo {
            width,
            height,
            drm_format: DrmFormat::from(format),
            // Linear layout is a requirement as virtio wayland guest expects
            // this for CPU access to the buffer. Scanout and texturing are
            // optional as the consumer (wayland compositor) is expected to
            // fall-back to a less efficient mechanisms for presentation if
            // neccesary. In practice, linear buffers for commonly used formats
            // will also support scanout and texturing.
            flags: RutabagaGrallocFlags::empty().use_linear(true),
        };

        let reqs = state
            .gralloc
            .get_image_memory_requirements(img)
            .map_err(WlError::GrallocError)?;
        let handle = state
            .gralloc
            .allocate_memory(reqs)
            .map_err(WlError::GrallocError)?;
        drop(state);

        let safe_descriptor = to_safe_descriptor(handle.os_handle);
        self.register_memory(
            safe_descriptor
                .try_clone()
                .context("failed to dup gfx handle")
                .map_err(WlError::ShmemMapperError)?,
            reqs.size,
            Protection::read_write(),
        )
        .map(|info| (info, safe_descriptor, reqs))
    }

    fn register_shmem(&self, shm: &SharedMemory) -> WlResult<u64> {
        let prot = match FileFlags::from_file(shm) {
            Ok(FileFlags::Read) => Protection::read(),
            Ok(FileFlags::Write) => Protection::write(),
            Ok(FileFlags::ReadWrite) => {
                let seals = shm.get_seals().map_err(WlError::GetSeals)?;
                if seals.write_seal() {
                    Protection::read()
                } else {
                    Protection::read_write()
                }
            }
            Err(e) => {
                return Err(WlError::ShmemMapperError(anyhow!(
                    "failed to get file descriptor flags with error: {:?}",
                    e
                )))
            }
        };
        self.register_memory(
            SafeDescriptor::try_from(shm as &dyn AsRawDescriptor)
                .context("failed to create safe descriptor")
                .map_err(WlError::ShmemMapperError)?,
            shm.size(),
            prot,
        )
    }

    fn register_memory(
        &self,
        descriptor: SafeDescriptor,
        size: u64,
        prot: Protection,
    ) -> WlResult<u64> {
        let mut state = self.state.borrow_mut();
        let size = round_up_to_page_size(size as usize) as u64;

        let source = VmMemorySource::Descriptor {
            descriptor,
            offset: 0,
            size,
        };
        let alloc = Alloc::Anon(state.next_alloc);
        state.next_alloc += 1;
        let offset = state
            .address_allocator
            .allocate(size, alloc, "virtio-wl".to_owned())
            .context("failed to allocate offset")
            .map_err(WlError::ShmemMapperError)?;

        match state
            .mapper
            .add_mapping(source, offset, prot, MemCacheType::CacheCoherent)
        {
            Ok(()) => {
                state.allocs.insert(offset, alloc);
                Ok(offset)
            }
            Err(e) => {
                // We just allocated it ourselves, it must exist.
                state
                    .address_allocator
                    .release(alloc)
                    .expect("corrupt address space");
                Err(WlError::ShmemMapperError(e))
            }
        }
    }
}

#[repr(C)]
#[derive(Copy, Clone, Default, FromBytes, Immutable, IntoBytes, KnownLayout)]
struct CtrlHeader {
    type_: Le32,
    flags: Le32,
}

#[repr(C)]
#[derive(Copy, Clone, Default, FromBytes, Immutable, IntoBytes, KnownLayout)]
struct CtrlVfdNew {
    hdr: CtrlHeader,
    id: Le32,
    flags: Le32,
    pfn: Le64,
    size: Le32,
    padding: Le32,
}

#[repr(C)]
#[derive(Copy, Clone, Default, FromBytes, Immutable, IntoBytes, KnownLayout)]
struct CtrlVfdNewCtxNamed {
    hdr: CtrlHeader,
    id: Le32,
    flags: Le32, // Ignored.
    pfn: Le64,   // Ignored.
    size: Le32,  // Ignored.
    name: [u8; 32],
    _pad: u32,
}
const_assert_eq!(size_of::<CtrlVfdNewCtxNamed>(), 64);

#[repr(C)]
#[derive(Copy, Clone, Default, FromBytes, Immutable, IntoBytes, KnownLayout)]
#[cfg(feature = "gbm")]
struct CtrlVfdNewDmabuf {
    hdr: CtrlHeader,
    id: Le32,
    flags: Le32,
    pfn: Le64,
    size: Le32,
    width: Le32,
    height: Le32,
    format: Le32,
    stride0: Le32,
    stride1: Le32,
    stride2: Le32,
    offset0: Le32,
    offset1: Le32,
    offset2: Le32,
}

#[cfg(feature = "gbm")]
#[repr(C)]
#[derive(Copy, Clone, Default, FromBytes, Immutable, IntoBytes, KnownLayout)]
struct CtrlVfdDmabufSync {
    hdr: CtrlHeader,
    id: Le32,
    flags: Le32,
}

#[repr(C)]
#[derive(Copy, Clone, FromBytes, Immutable, IntoBytes, KnownLayout)]
struct CtrlVfdRecv {
    hdr: CtrlHeader,
    id: Le32,
    vfd_count: Le32,
}

#[repr(C)]
#[derive(Copy, Clone, Default, FromBytes, Immutable, IntoBytes, KnownLayout)]
struct CtrlVfd {
    hdr: CtrlHeader,
    id: Le32,
}

#[repr(C)]
#[derive(Copy, Clone, Default, FromBytes, Immutable, IntoBytes, KnownLayout)]
struct CtrlVfdSend {
    hdr: CtrlHeader,
    id: Le32,
    vfd_count: Le32,
    // Remainder is an array of vfd_count IDs followed by data.
}

#[repr(C)]
#[derive(Copy, Clone, Default, FromBytes, Immutable, IntoBytes, KnownLayout)]
struct CtrlVfdSendVfd {
    kind: Le32,
    id: Le32,
}

#[repr(C)]
#[derive(Copy, Clone, FromBytes, Immutable, KnownLayout)]
union CtrlVfdSendVfdV2Payload {
    id: Le32,
    seqno: Le64,
}

#[repr(C)]
#[derive(Copy, Clone, FromBytes, Immutable, KnownLayout)]
struct CtrlVfdSendVfdV2 {
    kind: Le32,
    payload: CtrlVfdSendVfdV2Payload,
}

impl CtrlVfdSendVfdV2 {
    fn id(&self) -> Le32 {
        assert!(
            self.kind == VIRTIO_WL_CTRL_VFD_SEND_KIND_LOCAL
                || self.kind == VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU
        );
        // SAFETY: trivially safe given we assert kind
        unsafe { self.payload.id }
    }
    #[cfg(feature = "gpu")]
    fn seqno(&self) -> Le64 {
        assert!(self.kind == VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU_FENCE);
        // SAFETY: trivially safe given we assert kind
        unsafe { self.payload.seqno }
    }
}

#[derive(Debug)]
#[allow(dead_code)]
enum WlResp<'a> {
    Ok,
    VfdNew {
        id: u32,
        flags: u32,
        pfn: u64,
        size: u32,
        // The VfdNew variant can be either a response or a command depending on this `resp`. This
        // is important for the `get_code` method.
        resp: bool,
    },
    #[cfg(feature = "gbm")]
    VfdNewDmabuf {
        id: u32,
        flags: u32,
        pfn: u64,
        size: u32,
        desc: GpuMemoryDesc,
    },
    VfdRecv {
        id: u32,
        data: &'a [u8],
        vfds: &'a [u32],
    },
    VfdHup {
        id: u32,
    },
    Err(Box<dyn StdError>),
    OutOfMemory,
    InvalidId,
    InvalidType,
    InvalidFlags,
    InvalidCommand,
}

impl WlResp<'_> {
    fn get_code(&self) -> u32 {
        match *self {
            WlResp::Ok => VIRTIO_WL_RESP_OK,
            WlResp::VfdNew { resp, .. } => {
                if resp {
                    VIRTIO_WL_RESP_VFD_NEW
                } else {
                    VIRTIO_WL_CMD_VFD_NEW
                }
            }
            #[cfg(feature = "gbm")]
            WlResp::VfdNewDmabuf { .. } => VIRTIO_WL_RESP_VFD_NEW_DMABUF,
            WlResp::VfdRecv { .. } => VIRTIO_WL_CMD_VFD_RECV,
            WlResp::VfdHup { .. } => VIRTIO_WL_CMD_VFD_HUP,
            WlResp::Err(_) => VIRTIO_WL_RESP_ERR,
            WlResp::OutOfMemory => VIRTIO_WL_RESP_OUT_OF_MEMORY,
            WlResp::InvalidId => VIRTIO_WL_RESP_INVALID_ID,
            WlResp::InvalidType => VIRTIO_WL_RESP_INVALID_TYPE,
            WlResp::InvalidFlags => VIRTIO_WL_RESP_INVALID_FLAGS,
            WlResp::InvalidCommand => VIRTIO_WL_RESP_INVALID_CMD,
        }
    }
}

#[derive(Default)]
struct WlVfd {
    socket: Option<ScmSocket<UnixStream>>,
    guest_shared_memory: Option<SharedMemory>,
    remote_pipe: Option<File>,
    local_pipe: Option<(u32 /* flags */, File)>,
    slot: Option<(u64 /* offset */, VmRequester)>,
    #[cfg(feature = "gbm")]
    is_dmabuf: bool,
    #[cfg(feature = "gbm")]
    map_info: u32,
    fence: Option<File>,
    is_fence: bool,
}

impl fmt::Debug for WlVfd {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        write!(f, "WlVfd {{")?;
        if let Some(s) = &self.socket {
            write!(f, " socket: {}", s.as_raw_descriptor())?;
        }
        if let Some((offset, _)) = &self.slot {
            write!(f, " offset: {offset}")?;
        }
        if let Some(s) = &self.remote_pipe {
            write!(f, " remote: {}", s.as_raw_descriptor())?;
        }
        if let Some((_, s)) = &self.local_pipe {
            write!(f, " local: {}", s.as_raw_descriptor())?;
        }
        write!(f, " }}")
    }
}

#[cfg(feature = "gbm")]
fn flush_shared_memory(shared_memory: &SharedMemory) -> Result<()> {
    let mmap = match MemoryMappingBuilder::new(shared_memory.size as usize)
        .from_shared_memory(shared_memory)
        .build()
    {
        Ok(v) => v,
        Err(_) => return Err(Error::new(EINVAL)),
    };
    if let Err(err) = mmap.flush_all() {
        base::error!("failed to flush shared memory: {}", err);
        return match err {
            MmapError::NotImplemented(_) => Err(Error::new(ENOSYS)),
            _ => Err(Error::new(EINVAL)),
        };
    }
    Ok(())
}

impl WlVfd {
    fn connect<P: AsRef<Path>>(path: P) -> WlResult<WlVfd> {
        let socket = UnixStream::connect(path).map_err(WlError::SocketConnect)?;
        let mut vfd = WlVfd::default();
        vfd.socket = Some(socket.try_into().map_err(WlError::SocketConnect)?);
        Ok(vfd)
    }

    fn allocate(vm: VmRequester, size: u64) -> WlResult<WlVfd> {
        let size_page_aligned = round_up_to_page_size(size as usize) as u64;
        let vfd_shm =
            SharedMemory::new("virtwl_alloc", size_page_aligned).map_err(WlError::NewAlloc)?;

        let offset = vm.register_shmem(&vfd_shm)?;

        let mut vfd = WlVfd::default();
        vfd.guest_shared_memory = Some(vfd_shm);
        vfd.slot = Some((offset, vm));
        Ok(vfd)
    }

    #[cfg(feature = "gbm")]
    fn dmabuf(
        vm: VmRequester,
        width: u32,
        height: u32,
        format: u32,
    ) -> WlResult<(WlVfd, GpuMemoryDesc)> {
        let (offset, desc, reqs) = vm.allocate_and_register_gpu_memory(width, height, format)?;
        let mut vfd = WlVfd::default();
        let vfd_shm =
            SharedMemory::from_safe_descriptor(desc, reqs.size).map_err(WlError::NewAlloc)?;

        let mut desc = GpuMemoryDesc::default();
        for i in 0..3 {
            desc.planes[i] = GpuMemoryPlaneDesc {
                stride: reqs.strides[i],
                offset: reqs.offsets[i],
            }
        }

        vfd.guest_shared_memory = Some(vfd_shm);
        vfd.slot = Some((offset, vm));
        vfd.is_dmabuf = true;
        vfd.map_info = reqs.map_info;
        Ok((vfd, desc))
    }

    #[cfg(feature = "gbm")]
    fn dmabuf_sync(&self, flags: u32) -> WlResult<()> {
        if !self.is_dmabuf {
            return Err(WlError::DmabufSync(io::Error::from_raw_os_error(EINVAL)));
        }

        match &self.guest_shared_memory {
            Some(descriptor) => {
                let sync = dma_buf_sync {
                    flags: flags as u64,
                };
                // SAFETY:
                // Safe as descriptor is a valid dmabuf and incorrect flags will return an error.
                if unsafe { ioctl_with_ref(descriptor, DMA_BUF_IOCTL_SYNC, &sync) } < 0 {
                    return Err(WlError::DmabufSync(io::Error::last_os_error()));
                }

                // virtio-wl kernel driver always maps dmabufs with WB memory type, regardless of
                // the host memory type (which is wrong). However, to avoid changing the protocol,
                // assume that all guest writes are cached and ensure clflush-like ops on all mapped
                // cachelines if the host mapping is not cached.
                const END_WRITE_MASK: u32 = DMA_BUF_SYNC_WRITE | DMA_BUF_SYNC_END;
                if (flags & END_WRITE_MASK) == END_WRITE_MASK
                    && (self.map_info & RUTABAGA_MAP_CACHE_MASK) != RUTABAGA_MAP_CACHE_CACHED
                {
                    if let Err(err) = flush_shared_memory(descriptor) {
                        base::warn!("failed to flush cached dmabuf mapping: {:?}", err);
                        return Err(WlError::DmabufSync(io::Error::from_raw_os_error(
                            err.errno(),
                        )));
                    }
                }
                Ok(())
            }
            None => Err(WlError::DmabufSync(io::Error::from_raw_os_error(EBADF))),
        }
    }

    fn pipe_remote_read_local_write() -> WlResult<WlVfd> {
        let (read_pipe, write_pipe) = pipe().map_err(WlError::NewPipe)?;
        let mut vfd = WlVfd::default();
        vfd.remote_pipe = Some(read_pipe);
        vfd.local_pipe = Some((VIRTIO_WL_VFD_WRITE, write_pipe));
        Ok(vfd)
    }

    fn pipe_remote_write_local_read() -> WlResult<WlVfd> {
        let (read_pipe, write_pipe) = pipe().map_err(WlError::NewPipe)?;
        let mut vfd = WlVfd::default();
        vfd.remote_pipe = Some(write_pipe);
        vfd.local_pipe = Some((VIRTIO_WL_VFD_READ, read_pipe));
        Ok(vfd)
    }

    fn from_file(vm: VmRequester, mut descriptor: File) -> WlResult<WlVfd> {
        // We need to determine if the given file is more like shared memory or a pipe/socket. A
        // quick and easy check is to seek to the end of the file. If it works we assume it's not a
        // pipe/socket because those have no end. We can even use that seek location as an indicator
        // for how big the shared memory chunk to map into guest memory is. If seeking to the end
        // fails, we assume it's a socket or pipe with read/write semantics.
        if descriptor.seek(SeekFrom::End(0)).is_ok() {
            let shm = SharedMemory::from_file(descriptor).map_err(WlError::FromSharedMemory)?;
            let offset = vm.register_shmem(&shm)?;

            let mut vfd = WlVfd::default();
            vfd.guest_shared_memory = Some(shm);
            vfd.slot = Some((offset, vm));
            Ok(vfd)
        } else if is_fence(&descriptor) {
            let mut vfd = WlVfd::default();
            vfd.is_fence = true;
            vfd.fence = Some(descriptor);
            Ok(vfd)
        } else {
            let flags = match FileFlags::from_file(&descriptor) {
                Ok(FileFlags::Read) => VIRTIO_WL_VFD_READ,
                Ok(FileFlags::Write) => VIRTIO_WL_VFD_WRITE,
                Ok(FileFlags::ReadWrite) => VIRTIO_WL_VFD_READ | VIRTIO_WL_VFD_WRITE,
                _ => 0,
            };
            let mut vfd = WlVfd::default();
            vfd.local_pipe = Some((flags, descriptor));
            Ok(vfd)
        }
    }

    fn flags(&self, use_transition_flags: bool) -> u32 {
        let mut flags = 0;
        if use_transition_flags {
            if self.socket.is_some() {
                flags |= VIRTIO_WL_VFD_WRITE | VIRTIO_WL_VFD_READ;
            }
            if let Some((f, _)) = self.local_pipe {
                flags |= f;
            }
            if self.is_fence {
                flags |= VIRTIO_WL_VFD_FENCE;
            }
        } else {
            if self.socket.is_some() {
                flags |= VIRTIO_WL_VFD_CONTROL;
            }
            if self.slot.is_some() {
                flags |= VIRTIO_WL_VFD_WRITE | VIRTIO_WL_VFD_MAP
            }
        }
        flags
    }

    // Offset within the shared memory region this VFD was mapped at.
    fn offset(&self) -> Option<u64> {
        self.slot.as_ref().map(|s| s.0)
    }

    // Size in bytes of the shared memory VFD.
    fn size(&self) -> Option<u64> {
        self.guest_shared_memory.as_ref().map(|shm| shm.size())
    }

    // The descriptor that gets sent if this VFD is sent over a socket.
    fn send_descriptor(&self) -> Option<RawDescriptor> {
        self.guest_shared_memory
            .as_ref()
            .map(|shm| shm.as_raw_descriptor())
            .or(self.socket.as_ref().map(|s| s.as_raw_descriptor()))
            .or(self.remote_pipe.as_ref().map(|p| p.as_raw_descriptor()))
            .or(self.fence.as_ref().map(|f| f.as_raw_descriptor()))
    }

    // The FD that is used for polling for events on this VFD.
    fn wait_descriptor(&self) -> Option<&dyn AsRawDescriptor> {
        self.socket
            .as_ref()
            .map(|s| s as &dyn AsRawDescriptor)
            .or_else(|| {
                self.local_pipe
                    .as_ref()
                    .map(|(_, p)| p as &dyn AsRawDescriptor)
            })
            .or_else(|| self.fence.as_ref().map(|f| f as &dyn AsRawDescriptor))
    }

    // Sends data/files from the guest to the host over this VFD.
    fn send(&mut self, rds: &[RawDescriptor], data: &mut Reader) -> WlResult<WlResp> {
        if let Some(socket) = &self.socket {
            socket
                .send_vectored_with_fds(&data.get_remaining(), rds)
                .map_err(WlError::SendVfd)?;
            // All remaining data in `data` is now considered consumed.
            data.consume(usize::MAX);
            Ok(WlResp::Ok)
        } else if let Some((_, local_pipe)) = &mut self.local_pipe {
            // Impossible to send descriptors over a simple pipe.
            if !rds.is_empty() {
                return Ok(WlResp::InvalidType);
            }
            data.read_to(local_pipe, usize::MAX)
                .map_err(WlError::WritePipe)?;
            Ok(WlResp::Ok)
        } else {
            Ok(WlResp::InvalidType)
        }
    }

    // Receives data/files from the host for this VFD and queues it for the guest.
    fn recv(&mut self, in_file_queue: &mut Vec<File>) -> WlResult<Vec<u8>> {
        if let Some(socket) = self.socket.take() {
            let mut buf = vec![0; IN_BUFFER_LEN];
            // If any errors happen, the socket will get dropped, preventing more reading.
            let (len, descriptors) = socket
                .recv_with_fds(&mut buf, VIRTWL_SEND_MAX_ALLOCS)
                .map_err(WlError::RecvVfd)?;
            // If any data gets read, the put the socket back for future recv operations.
            if len != 0 || !descriptors.is_empty() {
                buf.truncate(len);
                buf.shrink_to_fit();
                self.socket = Some(socket);
                in_file_queue.extend(descriptors.into_iter().map(File::from));
                return Ok(buf);
            }
            Ok(Vec::new())
        } else if let Some((flags, mut local_pipe)) = self.local_pipe.take() {
            let mut buf = vec![0; IN_BUFFER_LEN];
            let len = local_pipe.read(&mut buf[..]).map_err(WlError::ReadPipe)?;
            if len != 0 {
                buf.truncate(len);
                buf.shrink_to_fit();
                self.local_pipe = Some((flags, local_pipe));
                return Ok(buf);
            }
            Ok(Vec::new())
        } else {
            Ok(Vec::new())
        }
    }

    // Called after this VFD is sent over a socket to ensure the local end of the VFD receives hang
    // up events.
    fn close_remote(&mut self) {
        self.remote_pipe = None;
    }

    fn close(&mut self) -> WlResult<()> {
        if let Some((offset, vm)) = self.slot.take() {
            vm.unregister_memory(offset)?;
        }
        self.socket = None;
        self.remote_pipe = None;
        self.local_pipe = None;
        Ok(())
    }
}

impl Drop for WlVfd {
    fn drop(&mut self) {
        let _ = self.close();
    }
}

#[derive(Debug)]
enum WlRecv {
    Vfd { id: u32 },
    Data { buf: Vec<u8> },
    Hup,
}

pub struct WlState {
    wayland_paths: BTreeMap<String, PathBuf>,
    vm: VmRequester,
    resource_bridge: Option<Tube>,
    use_transition_flags: bool,
    wait_ctx: WaitContext<u32>,
    vfds: BTreeMap<u32, WlVfd>,
    next_vfd_id: u32,
    in_file_queue: Vec<File>,
    in_queue: VecDeque<(u32 /* vfd_id */, WlRecv)>,
    current_recv_vfd: Option<u32>,
    recv_vfds: Vec<u32>,
    #[cfg(feature = "gpu")]
    signaled_fence: Option<SafeDescriptor>,
    use_send_vfd_v2: bool,
    address_offset: Option<u64>,
}

impl WlState {
    /// Create a new `WlState` instance for running a virtio-wl device.
    pub fn new(
        wayland_paths: BTreeMap<String, PathBuf>,
        mapper: Box<dyn SharedMemoryMapper>,
        use_transition_flags: bool,
        use_send_vfd_v2: bool,
        resource_bridge: Option<Tube>,
        #[cfg(feature = "gbm")] gralloc: RutabagaGralloc,
        address_offset: Option<u64>,
    ) -> WlState {
        WlState {
            wayland_paths,
            vm: VmRequester::new(
                mapper,
                #[cfg(feature = "gbm")]
                gralloc,
            ),
            resource_bridge,
            wait_ctx: WaitContext::new().expect("failed to create WaitContext"),
            use_transition_flags,
            vfds: BTreeMap::new(),
            next_vfd_id: NEXT_VFD_ID_BASE,
            in_file_queue: Vec::new(),
            in_queue: VecDeque::new(),
            current_recv_vfd: None,
            recv_vfds: Vec::new(),
            #[cfg(feature = "gpu")]
            signaled_fence: None,
            use_send_vfd_v2,
            address_offset,
        }
    }

    /// This is a hack so that we can drive the inner WaitContext from an async fn. The proper
    /// long-term solution is to replace the WaitContext completely by spawning async workers
    /// instead.
    pub fn wait_ctx(&self) -> &WaitContext<u32> {
        &self.wait_ctx
    }

    fn new_pipe(&mut self, id: u32, flags: u32) -> WlResult<WlResp> {
        if id & VFD_ID_HOST_MASK != 0 {
            return Ok(WlResp::InvalidId);
        }

        if flags & !(VIRTIO_WL_VFD_WRITE | VIRTIO_WL_VFD_READ) != 0 {
            return Ok(WlResp::InvalidFlags);
        }

        if flags & VIRTIO_WL_VFD_WRITE != 0 && flags & VIRTIO_WL_VFD_READ != 0 {
            return Ok(WlResp::InvalidFlags);
        }

        match self.vfds.entry(id) {
            Entry::Vacant(entry) => {
                let vfd = if flags & VIRTIO_WL_VFD_WRITE != 0 {
                    WlVfd::pipe_remote_read_local_write()?
                } else if flags & VIRTIO_WL_VFD_READ != 0 {
                    WlVfd::pipe_remote_write_local_read()?
                } else {
                    return Ok(WlResp::InvalidFlags);
                };
                self.wait_ctx
                    .add(vfd.wait_descriptor().unwrap(), id)
                    .map_err(WlError::WaitContextAdd)?;
                let resp = WlResp::VfdNew {
                    id,
                    flags: 0,
                    pfn: 0,
                    size: 0,
                    resp: true,
                };
                entry.insert(vfd);
                Ok(resp)
            }
            Entry::Occupied(_) => Ok(WlResp::InvalidId),
        }
    }

    fn new_alloc(&mut self, id: u32, flags: u32, size: u32) -> WlResult<WlResp> {
        if id & VFD_ID_HOST_MASK != 0 {
            return Ok(WlResp::InvalidId);
        }

        if self.use_transition_flags {
            if flags != 0 {
                return Ok(WlResp::InvalidFlags);
            }
        } else if flags & !(VIRTIO_WL_VFD_WRITE | VIRTIO_WL_VFD_MAP) != 0 {
            return Ok(WlResp::Err(Box::from("invalid flags")));
        }

        if self.vfds.contains_key(&id) {
            return Ok(WlResp::InvalidId);
        }
        let vfd = WlVfd::allocate(self.vm.clone(), size as u64)?;
        let resp = WlResp::VfdNew {
            id,
            flags,
            pfn: self.compute_pfn(&vfd.offset()),
            size: vfd.size().unwrap_or_default() as u32,
            resp: true,
        };
        self.vfds.insert(id, vfd);
        Ok(resp)
    }

    #[cfg(feature = "gbm")]
    fn new_dmabuf(&mut self, id: u32, width: u32, height: u32, format: u32) -> WlResult<WlResp> {
        if id & VFD_ID_HOST_MASK != 0 {
            return Ok(WlResp::InvalidId);
        }

        if self.vfds.contains_key(&id) {
            return Ok(WlResp::InvalidId);
        }
        let (vfd, desc) = WlVfd::dmabuf(self.vm.clone(), width, height, format)?;
        let resp = WlResp::VfdNewDmabuf {
            id,
            flags: 0,
            pfn: self.compute_pfn(&vfd.offset()),
            size: vfd.size().unwrap_or_default() as u32,
            desc,
        };
        self.vfds.insert(id, vfd);
        Ok(resp)
    }

    #[cfg(feature = "gbm")]
    fn dmabuf_sync(&mut self, vfd_id: u32, flags: u32) -> WlResult<WlResp> {
        if flags & !(VIRTIO_WL_VFD_DMABUF_SYNC_VALID_FLAG_MASK) != 0 {
            return Ok(WlResp::InvalidFlags);
        }

        match self.vfds.get_mut(&vfd_id) {
            Some(vfd) => {
                vfd.dmabuf_sync(flags)?;
                Ok(WlResp::Ok)
            }
            None => Ok(WlResp::InvalidId),
        }
    }

    fn new_context(&mut self, id: u32, name: &str) -> WlResult<WlResp> {
        if id & VFD_ID_HOST_MASK != 0 {
            return Ok(WlResp::InvalidId);
        }

        let flags = if self.use_transition_flags {
            VIRTIO_WL_VFD_WRITE | VIRTIO_WL_VFD_READ
        } else {
            VIRTIO_WL_VFD_CONTROL
        };

        match self.vfds.entry(id) {
            Entry::Vacant(entry) => {
                let vfd = entry.insert(WlVfd::connect(
                    self.wayland_paths
                        .get(name)
                        .ok_or_else(|| WlError::UnknownSocketName(name.to_string()))?,
                )?);
                self.wait_ctx
                    .add(vfd.wait_descriptor().unwrap(), id)
                    .map_err(WlError::WaitContextAdd)?;
                Ok(WlResp::VfdNew {
                    id,
                    flags,
                    pfn: 0,
                    size: 0,
                    resp: true,
                })
            }
            Entry::Occupied(_) => Ok(WlResp::InvalidId),
        }
    }

    fn process_wait_context(&mut self) {
        let events = match self.wait_ctx.wait_timeout(Duration::from_secs(0)) {
            Ok(v) => v,
            Err(e) => {
                error!("failed waiting for vfd evens: {}", e);
                return;
            }
        };

        for event in events.iter().filter(|e| e.is_readable) {
            if let Err(e) = self.recv(event.token) {
                error!("failed to recv from vfd: {}", e)
            }
        }

        for event in events.iter().filter(|e| e.is_hungup) {
            if !event.is_readable {
                let vfd_id = event.token;
                if let Some(descriptor) =
                    self.vfds.get(&vfd_id).and_then(|vfd| vfd.wait_descriptor())
                {
                    if let Err(e) = self.wait_ctx.delete(descriptor) {
                        warn!("failed to remove hungup vfd from poll context: {}", e);
                    }
                }
                self.in_queue.push_back((vfd_id, WlRecv::Hup));
            }
        }
    }

    fn close(&mut self, vfd_id: u32) -> WlResult<WlResp> {
        let mut to_delete = BTreeSet::new();
        for (dest_vfd_id, q) in &self.in_queue {
            if *dest_vfd_id == vfd_id {
                if let WlRecv::Vfd { id } = q {
                    to_delete.insert(*id);
                }
            }
        }
        for vfd_id in to_delete {
            // Sorry sub-error, we can't have cascading errors leaving us in an inconsistent state.
            let _ = self.close(vfd_id);
        }
        match self.vfds.remove(&vfd_id) {
            Some(mut vfd) => {
                self.in_queue.retain(|&(id, _)| id != vfd_id);
                vfd.close()?;
                Ok(WlResp::Ok)
            }
            None => Ok(WlResp::InvalidId),
        }
    }

    #[cfg(feature = "gpu")]
    fn get_info(&mut self, request: ResourceRequest) -> Option<SafeDescriptor> {
        let sock = self.resource_bridge.as_ref().unwrap();
        match get_resource_info(sock, request) {
            Ok(ResourceInfo::Buffer(BufferInfo { handle, .. })) => Some(handle),
            Ok(ResourceInfo::Fence { handle }) => Some(handle),
            Err(ResourceBridgeError::InvalidResource(req)) => {
                warn!("attempt to send non-existent gpu resource {}", req);
                None
            }
            Err(e) => {
                error!("{}", e);
                // If there was an error with the resource bridge, it can no longer be
                // trusted to continue to function.
                self.resource_bridge = None;
                None
            }
        }
    }

    fn send(
        &mut self,
        vfd_id: u32,
        vfd_count: usize,
        foreign_id: bool,
        reader: &mut Reader,
    ) -> WlResult<WlResp> {
        // First stage gathers and normalizes all id information from guest memory.
        let mut send_vfd_ids = [CtrlVfdSendVfdV2 {
            kind: Le32::from(0),
            payload: CtrlVfdSendVfdV2Payload { id: Le32::from(0) },
        }; VIRTWL_SEND_MAX_ALLOCS];
        for vfd_id in send_vfd_ids.iter_mut().take(vfd_count) {
            *vfd_id = if foreign_id {
                if self.use_send_vfd_v2 {
                    reader.read_obj().map_err(WlError::ParseDesc)?
                } else {
                    let vfd: CtrlVfdSendVfd = reader.read_obj().map_err(WlError::ParseDesc)?;
                    CtrlVfdSendVfdV2 {
                        kind: vfd.kind,
                        payload: CtrlVfdSendVfdV2Payload { id: vfd.id },
                    }
                }
            } else {
                CtrlVfdSendVfdV2 {
                    kind: Le32::from(VIRTIO_WL_CTRL_VFD_SEND_KIND_LOCAL),
                    payload: CtrlVfdSendVfdV2Payload {
                        id: reader.read_obj().map_err(WlError::ParseDesc)?,
                    },
                }
            };
        }

        // Next stage collects corresponding file descriptors for each id.
        let mut rds = [0; VIRTWL_SEND_MAX_ALLOCS];
        #[cfg(feature = "gpu")]
        let mut bridged_files = Vec::new();
        for (&send_vfd_id, descriptor) in send_vfd_ids[..vfd_count].iter().zip(rds.iter_mut()) {
            match send_vfd_id.kind.to_native() {
                VIRTIO_WL_CTRL_VFD_SEND_KIND_LOCAL => {
                    match self.vfds.get(&send_vfd_id.id().to_native()) {
                        Some(vfd) => match vfd.send_descriptor() {
                            Some(vfd_fd) => *descriptor = vfd_fd,
                            None => return Ok(WlResp::InvalidType),
                        },
                        None => {
                            warn!(
                                "attempt to send non-existant vfd 0x{:08x}",
                                send_vfd_id.id().to_native()
                            );
                            return Ok(WlResp::InvalidId);
                        }
                    }
                }
                #[cfg(feature = "gpu")]
                VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU if self.resource_bridge.is_some() => {
                    match self.get_info(ResourceRequest::GetBuffer {
                        id: send_vfd_id.id().to_native(),
                    }) {
                        Some(handle) => {
                            *descriptor = handle.as_raw_descriptor();
                            bridged_files.push(handle.into());
                        }
                        None => return Ok(WlResp::InvalidId),
                    }
                }
                #[cfg(feature = "gpu")]
                VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU_FENCE if self.resource_bridge.is_some() => {
                    match self.get_info(ResourceRequest::GetFence {
                        seqno: send_vfd_id.seqno().to_native(),
                    }) {
                        Some(handle) => {
                            *descriptor = handle.as_raw_descriptor();
                            bridged_files.push(handle.into());
                        }
                        None => return Ok(WlResp::InvalidId),
                    }
                }
                #[cfg(feature = "gpu")]
                VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU_SIGNALED_FENCE
                    if self.resource_bridge.is_some() =>
                {
                    if self.signaled_fence.is_none() {
                        // If the guest is sending a signaled fence, we know a fence
                        // with seqno 0 must already be signaled.
                        match self.get_info(ResourceRequest::GetFence { seqno: 0 }) {
                            Some(handle) => self.signaled_fence = Some(handle),
                            None => return Ok(WlResp::InvalidId),
                        }
                    }
                    match self.signaled_fence.as_ref().unwrap().try_clone() {
                        Ok(dup) => {
                            *descriptor = dup.into_raw_descriptor();
                            // SAFETY:
                            // Safe because the fd comes from a valid SafeDescriptor.
                            let file: File = unsafe {
                                base::FromRawDescriptor::from_raw_descriptor(*descriptor)
                            };
                            bridged_files.push(file);
                        }
                        Err(_) => return Ok(WlResp::InvalidId),
                    }
                }
                VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU
                | VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU_FENCE
                | VIRTIO_WL_CTRL_VFD_SEND_KIND_VIRTGPU_SIGNALED_FENCE => {
                    let _ = self.resource_bridge.as_ref();
                    warn!("attempt to send foreign resource kind but feature is disabled");
                }
                kind => {
                    warn!("attempt to send unknown foreign resource kind: {}", kind);
                    return Ok(WlResp::InvalidId);
                }
            }
        }

        // Final stage sends file descriptors and data to the target vfd's socket.
        match self.vfds.get_mut(&vfd_id) {
            Some(vfd) => match vfd.send(&rds[..vfd_count], reader)? {
                WlResp::Ok => {}
                _ => return Ok(WlResp::InvalidType),
            },
            None => return Ok(WlResp::InvalidId),
        }
        // The vfds with remote FDs need to be closed so that the local side can receive
        // hangup events.
        for &send_vfd_id in &send_vfd_ids[..vfd_count] {
            if send_vfd_id.kind == VIRTIO_WL_CTRL_VFD_SEND_KIND_LOCAL {
                if let Some(vfd) = self.vfds.get_mut(&send_vfd_id.id().into()) {
                    vfd.close_remote();
                }
            }
        }
        Ok(WlResp::Ok)
    }

    fn recv(&mut self, vfd_id: u32) -> WlResult<()> {
        let buf = match self.vfds.get_mut(&vfd_id) {
            Some(vfd) => {
                if vfd.is_fence {
                    if let Err(e) = self.wait_ctx.delete(vfd.wait_descriptor().unwrap()) {
                        warn!("failed to remove hungup vfd from poll context: {}", e);
                    }
                    self.in_queue.push_back((vfd_id, WlRecv::Hup));
                    return Ok(());
                } else {
                    vfd.recv(&mut self.in_file_queue)?
                }
            }
            None => return Ok(()),
        };

        if self.in_file_queue.is_empty() && buf.is_empty() {
            self.in_queue.push_back((vfd_id, WlRecv::Hup));
            return Ok(());
        }
        for file in self.in_file_queue.drain(..) {
            let vfd = WlVfd::from_file(self.vm.clone(), file)?;
            if let Some(wait_descriptor) = vfd.wait_descriptor() {
                self.wait_ctx
                    .add(wait_descriptor, self.next_vfd_id)
                    .map_err(WlError::WaitContextAdd)?;
            }
            // Only necessary if we somehow wrap the id counter. The try_insert
            // API would be nicer, but that's currently experimental.
            while self.vfds.contains_key(&self.next_vfd_id) {
                self.next_vfd_id += 1;
            }
            self.vfds.insert(self.next_vfd_id, vfd);
            self.in_queue.push_back((
                vfd_id,
                WlRecv::Vfd {
                    id: self.next_vfd_id,
                },
            ));
            self.next_vfd_id += 1;
        }
        self.in_queue.push_back((vfd_id, WlRecv::Data { buf }));

        Ok(())
    }

    fn execute(&mut self, reader: &mut Reader) -> WlResult<WlResp> {
        let type_: Le32 = reader.peek_obj::<Le32>().map_err(WlError::ParseDesc)?;
        match type_.into() {
            VIRTIO_WL_CMD_VFD_NEW => {
                let ctrl = reader
                    .read_obj::<CtrlVfdNew>()
                    .map_err(WlError::ParseDesc)?;
                self.new_alloc(ctrl.id.into(), ctrl.flags.into(), ctrl.size.into())
            }
            VIRTIO_WL_CMD_VFD_CLOSE => {
                let ctrl = reader.read_obj::<CtrlVfd>().map_err(WlError::ParseDesc)?;
                self.close(ctrl.id.into())
            }
            VIRTIO_WL_CMD_VFD_SEND => {
                let ctrl = reader
                    .read_obj::<CtrlVfdSend>()
                    .map_err(WlError::ParseDesc)?;
                let foreign_id = false;
                self.send(
                    ctrl.id.into(),
                    ctrl.vfd_count.to_native() as usize,
                    foreign_id,
                    reader,
                )
            }
            #[cfg(feature = "gpu")]
            VIRTIO_WL_CMD_VFD_SEND_FOREIGN_ID => {
                let ctrl = reader
                    .read_obj::<CtrlVfdSend>()
                    .map_err(WlError::ParseDesc)?;
                let foreign_id = true;
                self.send(
                    ctrl.id.into(),
                    ctrl.vfd_count.to_native() as usize,
                    foreign_id,
                    reader,
                )
            }
            VIRTIO_WL_CMD_VFD_NEW_CTX => {
                let ctrl = reader.read_obj::<CtrlVfd>().map_err(WlError::ParseDesc)?;
                self.new_context(ctrl.id.into(), "")
            }
            VIRTIO_WL_CMD_VFD_NEW_PIPE => {
                let ctrl = reader
                    .read_obj::<CtrlVfdNew>()
                    .map_err(WlError::ParseDesc)?;
                self.new_pipe(ctrl.id.into(), ctrl.flags.into())
            }
            #[cfg(feature = "gbm")]
            VIRTIO_WL_CMD_VFD_NEW_DMABUF => {
                let ctrl = reader
                    .read_obj::<CtrlVfdNewDmabuf>()
                    .map_err(WlError::ParseDesc)?;
                self.new_dmabuf(
                    ctrl.id.into(),
                    ctrl.width.into(),
                    ctrl.height.into(),
                    ctrl.format.into(),
                )
            }
            #[cfg(feature = "gbm")]
            VIRTIO_WL_CMD_VFD_DMABUF_SYNC => {
                let ctrl = reader
                    .read_obj::<CtrlVfdDmabufSync>()
                    .map_err(WlError::ParseDesc)?;
                self.dmabuf_sync(ctrl.id.into(), ctrl.flags.into())
            }
            VIRTIO_WL_CMD_VFD_NEW_CTX_NAMED => {
                let ctrl = reader
                    .read_obj::<CtrlVfdNewCtxNamed>()
                    .map_err(WlError::ParseDesc)?;
                let name_len = ctrl
                    .name
                    .iter()
                    .position(|x| x == &0)
                    .unwrap_or(ctrl.name.len());
                let name =
                    std::str::from_utf8(&ctrl.name[..name_len]).map_err(WlError::InvalidString)?;
                self.new_context(ctrl.id.into(), name)
            }
            op_type => {
                warn!("unexpected command {}", op_type);
                Ok(WlResp::InvalidCommand)
            }
        }
    }

    fn next_recv(&self) -> Option<WlResp> {
        if let Some(q) = self.in_queue.front() {
            match *q {
                (vfd_id, WlRecv::Vfd { id }) => {
                    if self.current_recv_vfd.is_none() || self.current_recv_vfd == Some(vfd_id) {
                        match self.vfds.get(&id) {
                            Some(vfd) => Some(WlResp::VfdNew {
                                id,
                                flags: vfd.flags(self.use_transition_flags),
                                pfn: self.compute_pfn(&vfd.offset()),
                                size: vfd.size().unwrap_or_default() as u32,
                                resp: false,
                            }),
                            _ => Some(WlResp::VfdNew {
                                id,
                                flags: 0,
                                pfn: 0,
                                size: 0,
                                resp: false,
                            }),
                        }
                    } else {
                        Some(WlResp::VfdRecv {
                            id: self.current_recv_vfd.unwrap(),
                            data: &[],
                            vfds: &self.recv_vfds[..],
                        })
                    }
                }
                (vfd_id, WlRecv::Data { ref buf }) => {
                    if self.current_recv_vfd.is_none() || self.current_recv_vfd == Some(vfd_id) {
                        Some(WlResp::VfdRecv {
                            id: vfd_id,
                            data: &buf[..],
                            vfds: &self.recv_vfds[..],
                        })
                    } else {
                        Some(WlResp::VfdRecv {
                            id: self.current_recv_vfd.unwrap(),
                            data: &[],
                            vfds: &self.recv_vfds[..],
                        })
                    }
                }
                (vfd_id, WlRecv::Hup) => Some(WlResp::VfdHup { id: vfd_id }),
            }
        } else {
            None
        }
    }

    fn pop_recv(&mut self) {
        if let Some(q) = self.in_queue.front() {
            match *q {
                (vfd_id, WlRecv::Vfd { id }) => {
                    if self.current_recv_vfd.is_none() || self.current_recv_vfd == Some(vfd_id) {
                        self.recv_vfds.push(id);
                        self.current_recv_vfd = Some(vfd_id);
                    } else {
                        self.recv_vfds.clear();
                        self.current_recv_vfd = None;
                        return;
                    }
                }
                (vfd_id, WlRecv::Data { .. }) => {
                    self.recv_vfds.clear();
                    self.current_recv_vfd = None;
                    if !(self.current_recv_vfd.is_none() || self.current_recv_vfd == Some(vfd_id)) {
                        return;
                    }
                }
                (_, WlRecv::Hup) => {
                    self.recv_vfds.clear();
                    self.current_recv_vfd = None;
                }
            }
        }
        self.in_queue.pop_front();
    }

    fn compute_pfn(&self, offset: &Option<u64>) -> u64 {
        let addr = match (offset, self.address_offset) {
            (Some(o), Some(address_offset)) => o + address_offset,
            (Some(o), None) => *o,
            // without shmem, 0 is the special address for "no_pfn"
            (None, Some(_)) => 0,
            // with shmem, WL_SHMEM_SIZE is the special address for "no_pfn"
            (None, None) => WL_SHMEM_SIZE,
        };
        addr >> VIRTIO_WL_PFN_SHIFT
    }
}

#[derive(ThisError, Debug, PartialEq, Eq)]
#[error("no descriptors available in queue")]
pub struct DescriptorsExhausted;

/// Handle incoming events and forward them to the VM over the input queue.
pub fn process_in_queue(
    in_queue: &mut Queue,
    state: &mut WlState,
) -> ::std::result::Result<(), DescriptorsExhausted> {
    state.process_wait_context();

    let mut needs_interrupt = false;
    let mut exhausted_queue = false;
    loop {
        let mut desc = if let Some(d) = in_queue.peek() {
            d
        } else {
            exhausted_queue = true;
            break;
        };

        let mut should_pop = false;
        if let Some(in_resp) = state.next_recv() {
            match encode_resp(&mut desc.writer, in_resp) {
                Ok(()) => {
                    should_pop = true;
                }
                Err(e) => {
                    error!("failed to encode response to descriptor chain: {}", e);
                }
            }
            needs_interrupt = true;
            let desc = desc.pop();
            in_queue.add_used(desc);
        } else {
            break;
        }
        if should_pop {
            state.pop_recv();
        }
    }

    if needs_interrupt {
        in_queue.trigger_interrupt();
    }

    if exhausted_queue {
        Err(DescriptorsExhausted)
    } else {
        Ok(())
    }
}

/// Handle messages from the output queue and forward them to the display sever, if necessary.
pub fn process_out_queue(out_queue: &mut Queue, state: &mut WlState) {
    let mut needs_interrupt = false;
    while let Some(mut desc) = out_queue.pop() {
        let resp = match state.execute(&mut desc.reader) {
            Ok(r) => r,
            Err(e) => WlResp::Err(Box::new(e)),
        };

        match encode_resp(&mut desc.writer, resp) {
            Ok(()) => {}
            Err(e) => {
                error!("failed to encode response to descriptor chain: {}", e);
            }
        }

        out_queue.add_used(desc);
        needs_interrupt = true;
    }

    if needs_interrupt {
        out_queue.trigger_interrupt();
    }
}

struct Worker {
    in_queue: Queue,
    out_queue: Queue,
    state: WlState,
}

impl Worker {
    fn new(
        in_queue: Queue,
        out_queue: Queue,
        wayland_paths: BTreeMap<String, PathBuf>,
        mapper: Box<dyn SharedMemoryMapper>,
        use_transition_flags: bool,
        use_send_vfd_v2: bool,
        resource_bridge: Option<Tube>,
        #[cfg(feature = "gbm")] gralloc: RutabagaGralloc,
        address_offset: Option<u64>,
    ) -> Worker {
        Worker {
            in_queue,
            out_queue,
            state: WlState::new(
                wayland_paths,
                mapper,
                use_transition_flags,
                use_send_vfd_v2,
                resource_bridge,
                #[cfg(feature = "gbm")]
                gralloc,
                address_offset,
            ),
        }
    }

    fn run(&mut self, kill_evt: Event) -> anyhow::Result<()> {
        #[derive(EventToken)]
        enum Token {
            InQueue,
            OutQueue,
            Kill,
            State,
        }

        let wait_ctx: WaitContext<Token> = WaitContext::build_with(&[
            (self.in_queue.event(), Token::InQueue),
            (self.out_queue.event(), Token::OutQueue),
            (&kill_evt, Token::Kill),
            (&self.state.wait_ctx, Token::State),
        ])
        .context("failed creating WaitContext")?;

        let mut watching_state_ctx = true;
        'wait: loop {
            let events = match wait_ctx.wait() {
                Ok(v) => v,
                Err(e) => {
                    error!("failed waiting for events: {}", e);
                    break;
                }
            };

            for event in &events {
                match event.token {
                    Token::InQueue => {
                        let _ = self.in_queue.event().wait();
                        if !watching_state_ctx {
                            if let Err(e) =
                                wait_ctx.modify(&self.state.wait_ctx, EventType::Read, Token::State)
                            {
                                error!("Failed to modify wait_ctx descriptor for WlState: {}", e);
                                break;
                            }
                            watching_state_ctx = true;
                        }
                    }
                    Token::OutQueue => {
                        let _ = self.out_queue.event().wait();
                        process_out_queue(&mut self.out_queue, &mut self.state);
                    }
                    Token::Kill => break 'wait,
                    Token::State => {
                        if let Err(DescriptorsExhausted) =
                            process_in_queue(&mut self.in_queue, &mut self.state)
                        {
                            if let Err(e) =
                                wait_ctx.modify(&self.state.wait_ctx, EventType::None, Token::State)
                            {
                                error!(
                                    "Failed to stop watching wait_ctx descriptor for WlState: {}",
                                    e
                                );
                                break;
                            }
                            watching_state_ctx = false;
                        }
                    }
                }
            }
        }

        Ok(())
    }
}

pub struct Wl {
    worker_thread: Option<WorkerThread<BTreeMap<usize, Queue>>>,
    wayland_paths: BTreeMap<String, PathBuf>,
    mapper: Option<Box<dyn SharedMemoryMapper>>,
    resource_bridge: Option<Tube>,
    base_features: u64,
    acked_features: u64,
    #[cfg(feature = "gbm")]
    gralloc: Option<RutabagaGralloc>,
    address_offset: Option<u64>,
}

impl Wl {
    pub fn new(
        base_features: u64,
        wayland_paths: BTreeMap<String, PathBuf>,
        resource_bridge: Option<Tube>,
    ) -> Result<Wl> {
        Ok(Wl {
            worker_thread: None,
            wayland_paths,
            mapper: None,
            resource_bridge,
            base_features,
            acked_features: 0,
            #[cfg(feature = "gbm")]
            gralloc: None,
            address_offset: None,
        })
    }
}

impl VirtioDevice for Wl {
    fn keep_rds(&self) -> Vec<RawDescriptor> {
        let mut keep_rds = Vec::new();

        if let Some(mapper) = &self.mapper {
            if let Some(raw_descriptor) = mapper.as_raw_descriptor() {
                keep_rds.push(raw_descriptor);
            }
        }
        if let Some(resource_bridge) = &self.resource_bridge {
            keep_rds.push(resource_bridge.as_raw_descriptor());
        }
        keep_rds
    }

    #[cfg(feature = "gbm")]
    fn on_device_sandboxed(&mut self) {
        // Gralloc initialization can cause some GPU drivers to create their own threads
        // and that must be done after sandboxing.
        match RutabagaGralloc::new(RutabagaGrallocBackendFlags::new()) {
            Ok(g) => self.gralloc = Some(g),
            Err(e) => {
                error!("failed to initialize gralloc {:?}", e);
            }
        };
    }

    fn device_type(&self) -> DeviceType {
        DeviceType::Wl
    }

    fn queue_max_sizes(&self) -> &[u16] {
        QUEUE_SIZES
    }

    fn features(&self) -> u64 {
        self.base_features
            | 1 << VIRTIO_WL_F_TRANS_FLAGS
            | 1 << VIRTIO_WL_F_SEND_FENCES
            | 1 << VIRTIO_WL_F_USE_SHMEM
    }

    fn ack_features(&mut self, value: u64) {
        self.acked_features |= value;
    }

    fn activate(
        &mut self,
        _mem: GuestMemory,
        _interrupt: Interrupt,
        mut queues: BTreeMap<usize, Queue>,
    ) -> anyhow::Result<()> {
        if queues.len() != QUEUE_SIZES.len() {
            return Err(anyhow!(
                "expected {} queues, got {}",
                QUEUE_SIZES.len(),
                queues.len()
            ));
        }

        let mapper = self.mapper.take().context("missing mapper")?;

        let wayland_paths = self.wayland_paths.clone();
        let use_transition_flags = self.acked_features & (1 << VIRTIO_WL_F_TRANS_FLAGS) != 0;
        let use_send_vfd_v2 = self.acked_features & (1 << VIRTIO_WL_F_SEND_FENCES) != 0;
        let use_shmem = self.acked_features & (1 << VIRTIO_WL_F_USE_SHMEM) != 0;
        let resource_bridge = self.resource_bridge.take();
        #[cfg(feature = "gbm")]
        let gralloc = self
            .gralloc
            .take()
            .expect("gralloc already passed to worker");
        let address_offset = if !use_shmem {
            self.address_offset
        } else {
            None
        };

        self.worker_thread = Some(WorkerThread::start("v_wl", move |kill_evt| {
            let mut worker = Worker::new(
                queues.pop_first().unwrap().1,
                queues.pop_first().unwrap().1,
                wayland_paths,
                mapper,
                use_transition_flags,
                use_send_vfd_v2,
                resource_bridge,
                #[cfg(feature = "gbm")]
                gralloc,
                address_offset,
            );
            if let Err(e) = worker.run(kill_evt) {
                error!("wl worker failed: {e:#}");
            }
            BTreeMap::from_iter([worker.in_queue, worker.out_queue].into_iter().enumerate())
        }));

        Ok(())
    }

    fn get_shared_memory_region(&self) -> Option<SharedMemoryRegion> {
        Some(SharedMemoryRegion {
            id: WL_SHMEM_ID,
            length: WL_SHMEM_SIZE,
        })
    }

    fn set_shared_memory_region(&mut self, shmem_region: AddressRange) {
        self.address_offset = Some(shmem_region.start);
    }

    fn set_shared_memory_mapper(&mut self, mapper: Box<dyn SharedMemoryMapper>) {
        self.mapper = Some(mapper);
    }

    fn virtio_sleep(&mut self) -> anyhow::Result<Option<BTreeMap<usize, Queue>>> {
        if let Some(worker_thread) = self.worker_thread.take() {
            let queues = worker_thread.stop();
            return Ok(Some(queues));
        }
        Ok(None)
    }

    fn virtio_wake(
        &mut self,
        device_state: Option<(GuestMemory, Interrupt, BTreeMap<usize, Queue>)>,
    ) -> anyhow::Result<()> {
        match device_state {
            None => Ok(()),
            Some((mem, interrupt, queues)) => {
                // TODO: activate is just what we want at the moment, but we should probably move
                // it into a "start workers" function to make it obvious that it isn't strictly
                // used for activate events.
                self.activate(mem, interrupt, queues)?;
                Ok(())
            }
        }
    }
}