metal: delay rendering until shortly before page flip

release-notes.md

@@ -1,6 +1,10 @@
 # Unreleased
 
 - Add remaining layer-shell features.
+- Add JAY_MAX_RENDER_TIME_NSEC environment variable.
+  This can be used to delay rendering until shortly before a page flip, reducing input
+  delay.
+  This is an unstable feature that might change in the future.
 
 # 1.2.0 (2024-05-05)
 

src/backends/metal/video.rs

@@ -18,6 +18,7 @@ use {
         renderer::RenderResult,
         state::State,
         theme::Color,
+        time::now_nsec,
         tree::OutputNode,
         udev::UdevDevice,
         utils::{
@@ -45,6 +46,7 @@ use {
     indexmap::{indexset, IndexSet},
     isnt::std_1::collections::IsntHashMap2Ext,
     jay_config::video::GfxApi,
+    once_cell::sync::Lazy,
     std::{
         any::Any,
         cell::{Cell, RefCell},
@@ -416,6 +418,7 @@ pub struct MetalConnector {
     pub can_present: Cell<bool>,
     pub has_damage: Cell<bool>,
     pub cursor_changed: Cell<bool>,
+    pub next_flip_nsec: Cell<u64>,
 
     pub display: RefCell<ConnectorDisplayData>,
 
@@ -578,6 +581,20 @@ impl MetalConnector {
     async fn present_loop(self: Rc<Self>) {
         loop {
             self.present_trigger.triggered().await;
+            static DELTA: Lazy<Option<u64>> = Lazy::new(|| {
+                if let Ok(max_render_time) = std::env::var("JAY_MAX_RENDER_TIME_NSEC") {
+                    if let Ok(max_render_time) = max_render_time.parse() {
+                        return Some(max_render_time);
+                    }
+                }
+                None
+            });
+            if let Some(delta) = *DELTA {
+                let next_present = self.next_flip_nsec.get().saturating_sub(delta);
+                if now_nsec() < next_present {
+                    self.state.ring.timeout(next_present).await.unwrap();
+                }
+            }
             match self.present(true) {
                 Ok(_) => self.state.set_backend_idle(false),
                 Err(e) => {
@@ -1397,6 +1414,7 @@ fn create_connector(
         active_framebuffer: Default::default(),
         next_framebuffer: Default::default(),
         direct_scanout_active: Cell::new(false),
+        next_flip_nsec: Cell::new(0),
     });
     let futures = ConnectorFutures {
         present: backend
@@ -2161,6 +2179,9 @@ impl MetalBackend {
             connector.schedule_present();
         }
         let dd = connector.display.borrow_mut();
+        connector
+            .next_flip_nsec
+            .set(tv_sec as u64 * 1_000_000_000 + tv_usec as u64 * 1000 + dd.refresh as u64);
         {
             let global = self.state.root.outputs.get(&connector.connector_id);
             let mut rr = connector.render_result.borrow_mut();

src/io_uring.rs

@@ -6,7 +6,7 @@ use {
             ops::{
                 accept::AcceptTask, async_cancel::AsyncCancelTask, connect::ConnectTask,
                 poll::PollTask, read_write::ReadWriteTask, recvmsg::RecvmsgTask,
-                sendmsg::SendmsgTask, timeout::TimeoutTask,
+                sendmsg::SendmsgTask, timeout::TimeoutTask, timeout_link::TimeoutLinkTask,
             },
             pending_result::PendingResults,
             sys::{
@@ -211,6 +211,7 @@ impl IoUring {
             cached_sendmsg: Default::default(),
             cached_recvmsg: Default::default(),
             cached_timeouts: Default::default(),
+            cached_timeout_links: Default::default(),
             cached_cmsg_bufs: Default::default(),
             cached_connects: Default::default(),
             cached_accepts: Default::default(),
@@ -266,6 +267,7 @@ struct IoUringData {
     cached_sendmsg: Stack<Box<SendmsgTask>>,
     cached_recvmsg: Stack<Box<RecvmsgTask>>,
     cached_timeouts: Stack<Box<TimeoutTask>>,
+    cached_timeout_links: Stack<Box<TimeoutLinkTask>>,
     cached_cmsg_bufs: Stack<Buf>,
     cached_connects: Stack<Box<ConnectTask>>,
     cached_accepts: Stack<Box<AcceptTask>>,

src/io_uring/ops.rs

@@ -8,6 +8,7 @@ pub mod read_write;
 pub mod recvmsg;
 pub mod sendmsg;
 pub mod timeout;
+pub mod timeout_link;
 
 pub type TaskResult<T> = Result<Result<T, OsError>, IoUringError>;
 

src/io_uring/ops/read_write.rs

@@ -51,7 +51,7 @@ impl IoUring {
             });
             self.ring.schedule(pw);
             if let Some(time) = timeout {
-                self.schedule_timeout(time);
+                self.schedule_timeout_link(time);
             }
         }
         Ok(pr.await.map(|v| v as usize)).merge()

src/io_uring/ops/sendmsg.rs

@@ -78,7 +78,7 @@ impl IoUring {
             st.has_timeout = timeout.is_some();
             self.ring.schedule(st);
             if let Some(timeout) = timeout {
-                self.schedule_timeout(timeout);
+                self.schedule_timeout_link(timeout);
             }
         }
         Ok(pr.await? as _)

src/io_uring/ops/timeout.rs

@@ -1,38 +1,44 @@
 use {
-    crate::{
-        io_uring::{
-            sys::{io_uring_sqe, IORING_OP_LINK_TIMEOUT, IORING_TIMEOUT_ABS},
-            IoUring, IoUringData, Task,
-        },
-        time::Time,
+    crate::io_uring::{
+        pending_result::PendingResult,
+        sys::{io_uring_sqe, IORING_OP_TIMEOUT, IORING_TIMEOUT_ABS},
+        IoUring, IoUringData, IoUringError, Task,
     },
     uapi::c,
 };
 
 #[allow(non_camel_case_types)]
 #[repr(C)]
 #[derive(Default)]
-struct timespec64 {
-    tv_sec: i64,
-    tv_nsec: c::c_long,
+pub(super) struct timespec64 {
+    pub tv_sec: i64,
+    pub tv_nsec: c::c_long,
 }
 
 #[derive(Default)]
 pub struct TimeoutTask {
     id: u64,
     timespec: timespec64,
+    pr: Option<PendingResult>,
 }
 
 impl IoUring {
-    pub(super) fn schedule_timeout(&self, timeout: Time) {
-        let id = self.ring.id_raw();
+    pub async fn timeout(&self, timeout_nsec: u64) -> Result<(), IoUringError> {
+        self.ring.check_destroyed()?;
+        let id = self.ring.id();
+        let pr = self.ring.pending_results.acquire();
         {
-            let mut to = self.ring.cached_timeouts.pop().unwrap_or_default();
-            to.id = id;
-            to.timespec.tv_sec = timeout.0.tv_sec as _;
-            to.timespec.tv_nsec = timeout.0.tv_nsec as _;
-            self.ring.schedule(to);
+            let mut pw = self.ring.cached_timeouts.pop().unwrap_or_default();
+            pw.id = id.id;
+            pw.timespec = timespec64 {
+                tv_sec: (timeout_nsec / 1_000_000_000) as _,
+                tv_nsec: (timeout_nsec % 1_000_000_000) as _,
+            };
+            pw.pr = Some(pr.clone());
+            self.ring.schedule(pw);
         }
+        let _ = pr.await;
+        Ok(())
     }
 }
 
@@ -41,14 +47,18 @@ unsafe impl Task for TimeoutTask {
         self.id
     }
 
-    fn complete(self: Box<Self>, ring: &IoUringData, _res: i32) {
+    fn complete(mut self: Box<Self>, ring: &IoUringData, res: i32) {
+        if let Some(pr) = self.pr.take() {
+            pr.complete(res);
+        }
         ring.cached_timeouts.push(self);
     }
 
     fn encode(&self, sqe: &mut io_uring_sqe) {
-        sqe.opcode = IORING_OP_LINK_TIMEOUT;
+        sqe.opcode = IORING_OP_TIMEOUT;
         sqe.u2.addr = &self.timespec as *const _ as _;
         sqe.len = 1;
         sqe.u3.timeout_flags = IORING_TIMEOUT_ABS;
+        sqe.u1.off = 0;
     }
 }

src/io_uring/ops/timeout_link.rs

@@ -0,0 +1,44 @@
+use crate::{
+    io_uring::{
+        ops::timeout::timespec64,
+        sys::{io_uring_sqe, IORING_OP_LINK_TIMEOUT, IORING_TIMEOUT_ABS},
+        IoUring, IoUringData, Task,
+    },
+    time::Time,
+};
+
+#[derive(Default)]
+pub struct TimeoutLinkTask {
+    id: u64,
+    timespec: timespec64,
+}
+
+impl IoUring {
+    pub(super) fn schedule_timeout_link(&self, timeout: Time) {
+        let id = self.ring.id_raw();
+        {
+            let mut to = self.ring.cached_timeout_links.pop().unwrap_or_default();
+            to.id = id;
+            to.timespec.tv_sec = timeout.0.tv_sec as _;
+            to.timespec.tv_nsec = timeout.0.tv_nsec as _;
+            self.ring.schedule(to);
+        }
+    }
+}
+
+unsafe impl Task for TimeoutLinkTask {
+    fn id(&self) -> u64 {
+        self.id
+    }
+
+    fn complete(self: Box<Self>, ring: &IoUringData, _res: i32) {
+        ring.cached_timeout_links.push(self);
+    }
+
+    fn encode(&self, sqe: &mut io_uring_sqe) {
+        sqe.opcode = IORING_OP_LINK_TIMEOUT;
+        sqe.u2.addr = &self.timespec as *const _ as _;
+        sqe.len = 1;
+        sqe.u3.timeout_flags = IORING_TIMEOUT_ABS;
+    }
+}

src/time.rs

@@ -62,7 +62,12 @@ impl Time {
         }
     }
 
-    #[allow(dead_code)]
+    pub fn nsec(self) -> u64 {
+        let sec = self.0.tv_sec as u64 * 1_000_000_000;
+        let nsec = self.0.tv_nsec as u64;
+        sec + nsec
+    }
+
     pub fn usec(self) -> u64 {
         let sec = self.0.tv_sec as u64 * 1_000_000;
         let nsec = self.0.tv_nsec as u64 / 1_000;
@@ -119,6 +124,10 @@ impl Add<Duration> for Time {
     }
 }
 
+pub fn now_nsec() -> u64 {
+    Time::now_unchecked().nsec()
+}
+
 pub fn now_usec() -> u64 {
     Time::now_unchecked().usec()
 }