| 1 | use crate::device::Shared; |
| 2 | use crate::error::{WAIT, observe, waited}; |
| 3 | use ash::vk; |
| 4 | use rustix::event::{PollFd, PollFlags, poll}; |
| 5 | use smithay::backend::renderer::sync::{Fence, Interrupted}; |
| 6 | use std::fmt; |
| 7 | use std::os::fd::{FromRawFd, OwnedFd}; |
| 8 | use std::sync::{Arc, Mutex}; |
| 9 | |
| 10 | enum State { |
| 11 | Pending, |
| 12 | /// Exporting a sync_file resets the Vulkan fence, so from then on the |
| 13 | /// file is the only record of completion. `None` means it had already |
| 14 | /// signalled when exported. |
| 15 | Exported(Option<OwnedFd>), |
| 16 | } |
| 17 | |
| 18 | pub(crate) struct FenceInner { |
| 19 | shared: Arc<Shared>, |
| 20 | fence: vk::Fence, |
| 21 | exportable: bool, |
| 22 | state: Mutex<State>, |
| 23 | } |
| 24 | |
| 25 | impl FenceInner { |
| 26 | pub fn new(shared: &Arc<Shared>) -> Result<Arc<Self>, vk::Result> { |
| 27 | let exportable = shared.device.dmabuf.is_some(); |
| 28 | let mut export = vk::ExportFenceCreateInfo::default() |
| 29 | .handle_types(vk::ExternalFenceHandleTypeFlags::SYNC_FD); |
| 30 | let mut info = vk::FenceCreateInfo::default(); |
| 31 | if exportable { |
| 32 | info = info.push_next(&mut export); |
| 33 | } |
| 34 | let spare = shared |
| 35 | .device |
| 36 | .spare_fences |
| 37 | .lock() |
| 38 | .unwrap_or_else(|e| e.into_inner()) |
| 39 | .pop(); |
| 40 | let fence = match spare { |
| 41 | Some(f) => { |
| 42 | unsafe { shared.device.device.reset_fences(&[f])? }; |
| 43 | f |
| 44 | } |
| 45 | None => unsafe { shared.device.device.create_fence(&info, None)? }, |
| 46 | }; |
| 47 | Ok(Arc::new(Self { |
| 48 | shared: shared.clone(), |
| 49 | fence, |
| 50 | exportable, |
| 51 | state: Mutex::new(State::Pending), |
| 52 | })) |
| 53 | } |
| 54 | |
| 55 | pub fn fence(&self) -> vk::Fence { |
| 56 | self.fence |
| 57 | } |
| 58 | |
| 59 | /// Whether the GPU has signalled, or the error the query failed with. |
| 60 | pub fn status(&self) -> Result<bool, vk::Result> { |
| 61 | match &*self.state.lock().unwrap_or_else(|e| e.into_inner()) { |
| 62 | State::Pending => unsafe { self.shared.device.device.get_fence_status(self.fence) }, |
| 63 | State::Exported(None) => Ok(true), |
| 64 | State::Exported(Some(fd)) => Ok(readable(fd, Some(&Default::default()))), |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | /// Like `status` for callers that cannot report a failure. A failed query |
| 69 | /// counts as signalled so nobody waits on a fence that cannot signal. |
| 70 | pub fn signaled(&self) -> bool { |
| 71 | self.status().unwrap_or_else(|e| { |
| 72 | observe(e); |
| 73 | true |
| 74 | }) |
| 75 | } |
| 76 | |
| 77 | /// Blocks until the GPU signals. A pending fence is first turned into its |
| 78 | /// sync_file and polled, because waiting on the Vulkan fence itself sleeps |
| 79 | /// in 10 ms steps on NVIDIA even when the work is done in microseconds. |
| 80 | /// Gives up with `TIMEOUT` after [`WAIT`], so a hung GPU cannot freeze the |
| 81 | /// caller. |
| 82 | pub fn wait(&self) -> Result<(), vk::Result> { |
| 83 | if self.exportable { |
| 84 | drop(self.export()); |
| 85 | } |
| 86 | let waited_for = match &*self.state.lock().unwrap_or_else(|e| e.into_inner()) { |
| 87 | State::Pending => unsafe { |
| 88 | self.shared.device.device.wait_for_fences( |
| 89 | &[self.fence], |
| 90 | true, |
| 91 | WAIT.as_nanos() as u64, |
| 92 | ) |
| 93 | }, |
| 94 | State::Exported(None) => Ok(()), |
| 95 | State::Exported(Some(fd)) => { |
| 96 | let timeout = rustix::event::Timespec { |
| 97 | tv_sec: WAIT.as_secs() as _, |
| 98 | tv_nsec: 0, |
| 99 | }; |
| 100 | if readable(fd, Some(&timeout)) { |
| 101 | Ok(()) |
| 102 | } else { |
| 103 | Err(vk::Result::TIMEOUT) |
| 104 | } |
| 105 | } |
| 106 | }; |
| 107 | if waited_for.is_ok() { |
| 108 | waited(); |
| 109 | } |
| 110 | waited_for |
| 111 | } |
| 112 | |
| 113 | #[profiling::function] |
| 114 | fn export(&self) -> Option<OwnedFd> { |
| 115 | if !self.exportable { |
| 116 | return None; |
| 117 | } |
| 118 | let ext = self.shared.device.dmabuf.as_ref()?; |
| 119 | let mut state = self.state.lock().unwrap_or_else(|e| e.into_inner()); |
| 120 | if let State::Pending = *state { |
| 121 | let info = vk::FenceGetFdInfoKHR::default() |
| 122 | .fence(self.fence) |
| 123 | .handle_type(vk::ExternalFenceHandleTypeFlags::SYNC_FD); |
| 124 | let raw = unsafe { ext.fence_fd.get_fence_fd(&info) }.ok()?; |
| 125 | *state = State::Exported((raw >= 0).then(|| unsafe { OwnedFd::from_raw_fd(raw) })); |
| 126 | } |
| 127 | match &*state { |
| 128 | State::Exported(Some(fd)) => fd.try_clone().ok(), |
| 129 | _ => None, |
| 130 | } |
| 131 | } |
| 132 | } |
| 133 | |
| 134 | fn readable(fd: &OwnedFd, timeout: Option<&rustix::event::Timespec>) -> bool { |
| 135 | let mut fds = [PollFd::new(fd, PollFlags::IN)]; |
| 136 | loop { |
| 137 | match poll(&mut fds, timeout) { |
| 138 | Ok(n) => return n > 0, |
| 139 | Err(rustix::io::Errno::INTR) => continue, |
| 140 | Err(_) => return true, |
| 141 | } |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | /// Fences go back to a pool instead of being destroyed, which cut the median |
| 146 | /// queue submit on NVIDIA from 87 to 52 us. One still pending is destroyed. |
| 147 | impl Drop for FenceInner { |
| 148 | fn drop(&mut self) { |
| 149 | let d = &self.shared.device.device; |
| 150 | let state = self.state.get_mut().unwrap_or_else(|e| e.into_inner()); |
| 151 | let idle = match state { |
| 152 | State::Pending => unsafe { d.get_fence_status(self.fence) }.unwrap_or(false), |
| 153 | State::Exported(_) => true, |
| 154 | }; |
| 155 | if idle { |
| 156 | self.shared |
| 157 | .device |
| 158 | .spare_fences |
| 159 | .lock() |
| 160 | .unwrap_or_else(|e| e.into_inner()) |
| 161 | .push(self.fence); |
| 162 | } else { |
| 163 | unsafe { d.destroy_fence(self.fence, None) }; |
| 164 | } |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | impl fmt::Debug for FenceInner { |
| 169 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
| 170 | f.debug_tuple("FenceInner").field(&self.fence).finish() |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | #[derive(Debug)] |
| 175 | pub(crate) struct VkFence(pub Arc<FenceInner>); |
| 176 | |
| 177 | impl Fence for VkFence { |
| 178 | fn is_signaled(&self) -> bool { |
| 179 | self.0.signaled() |
| 180 | } |
| 181 | |
| 182 | fn wait(&self) -> Result<(), Interrupted> { |
| 183 | self.0.wait().map_err(|e| { |
| 184 | observe(e); |
| 185 | Interrupted |
| 186 | }) |
| 187 | } |
| 188 | |
| 189 | fn is_exportable(&self) -> bool { |
| 190 | self.0.exportable |
| 191 | } |
| 192 | |
| 193 | fn export(&self) -> Option<OwnedFd> { |
| 194 | self.0.export() |
| 195 | } |
| 196 | } |