Add syscall ppoll

docs/src/kernel/linux-compatibility.md

@@ -291,7 +291,7 @@ provided by Linux on x86-64 architecture.
 | 268     | fchmodat         | ✅              |
 | 269     | faccessat        | ✅              |
 | 270     | pselect6         | ✅              |
-| 271     | ppoll            | ❌              |
+| 271     | ppoll            | ✅              |
 | 272     | unshare          | ❌              |
 | 273     | set_robust_list  | ✅              |
 | 274     | get_robust_list  | ❌              |

kernel/src/process/signal/mod.rs

@@ -20,7 +20,7 @@ use c_types::{siginfo_t, ucontext_t};
 use constants::SIGSEGV;
 pub use events::{SigEvents, SigEventsFilter};
 use ostd::{cpu::context::UserContext, user::UserContextApi};
-pub use pause::{with_signal_blocked, Pause};
+pub use pause::{with_sigmask_changed, Pause};
 pub use poll::{PollAdaptor, PollHandle, Pollable, Pollee, Poller};
 use sig_action::{SigAction, SigActionFlags, SigDefaultAction};
 use sig_mask::SigMask;

kernel/src/process/signal/pause.rs

@@ -206,16 +206,22 @@ impl Pause for WaitQueue {
     }
 }
 
-/// Executes a closure while temporarily blocking some signals for the current POSIX thread.
-pub fn with_signal_blocked<R>(ctx: &Context, mask: SigMask, operate: impl FnOnce() -> R) -> R {
-    let posix_thread = ctx.posix_thread;
-    let sig_mask = posix_thread.sig_mask();
+/// Executes a closure after temporarily adjusting the signal mask of the current POSIX thread.
+pub fn with_sigmask_changed<R>(
+    ctx: &Context,
+    mask_op: impl FnOnce(SigMask) -> SigMask,
+    operate: impl FnOnce() -> R,
+) -> R {
+    let sig_mask = ctx.posix_thread.sig_mask();
 
+    // Save the original signal mask and apply the mask updates.
     let old_mask = sig_mask.load(Ordering::Relaxed);
-    sig_mask.store(old_mask + mask, Ordering::Relaxed);
+    sig_mask.store(mask_op(old_mask), Ordering::Relaxed);
 
+    // Perform the operation.
     let res = operate();
 
+    // Restore the original signal mask.
     sig_mask.store(old_mask, Ordering::Relaxed);
 
     res

kernel/src/process/wait.rs

@@ -2,13 +2,16 @@
 
 #![expect(dead_code)]
 
-use super::{process_filter::ProcessFilter, signal::constants::SIGCHLD, ExitCode, Pid, Process};
+use super::{
+    process_filter::ProcessFilter,
+    signal::{constants::SIGCHLD, with_sigmask_changed},
+    ExitCode, Pid, Process,
+};
 use crate::{
     prelude::*,
     process::{
         posix_thread::{thread_table, AsPosixThread},
         process_table,
-        signal::with_signal_blocked,
     },
 };
 
@@ -40,51 +43,55 @@ pub fn wait_child_exit(
     ctx: &Context,
 ) -> Result<Option<Arc<Process>>> {
     let current = ctx.process;
-    let zombie_child = with_signal_blocked(ctx, SIGCHLD.into(), || {
-        current.children_wait_queue().pause_until(|| {
-            let unwaited_children = current
-                .children()
-                .lock()
-                .values()
-                .filter(|child| match child_filter {
-                    ProcessFilter::Any => true,
-                    ProcessFilter::WithPid(pid) => child.pid() == pid,
-                    ProcessFilter::WithPgid(pgid) => child.pgid() == pgid,
-                })
-                .cloned()
-                .collect::<Vec<_>>();
+    let zombie_child = with_sigmask_changed(
+        ctx,
+        |sigmask| sigmask + SIGCHLD,
+        || {
+            current.children_wait_queue().pause_until(|| {
+                let unwaited_children = current
+                    .children()
+                    .lock()
+                    .values()
+                    .filter(|child| match child_filter {
+                        ProcessFilter::Any => true,
+                        ProcessFilter::WithPid(pid) => child.pid() == pid,
+                        ProcessFilter::WithPgid(pgid) => child.pgid() == pgid,
+                    })
+                    .cloned()
+                    .collect::<Vec<_>>();
 
-            if unwaited_children.is_empty() {
-                return Some(Err(Error::with_message(
-                    Errno::ECHILD,
-                    "the process has no child to wait",
-                )));
-            }
-
-            // return immediately if we find a zombie child
-            let zombie_child = unwaited_children
-                .iter()
-                .find(|child| child.status().is_zombie());
-
-            if let Some(zombie_child) = zombie_child {
-                let zombie_pid = zombie_child.pid();
-                if wait_options.contains(WaitOptions::WNOWAIT) {
-                    // does not reap child, directly return
-                    return Some(Ok(Some(zombie_child.clone())));
-                } else {
-                    reap_zombie_child(current, zombie_pid);
-                    return Some(Ok(Some(zombie_child.clone())));
+                if unwaited_children.is_empty() {
+                    return Some(Err(Error::with_message(
+                        Errno::ECHILD,
+                        "the process has no child to wait",
+                    )));
                 }
-            }
 
-            if wait_options.contains(WaitOptions::WNOHANG) {
-                return Some(Ok(None));
-            }
+                // return immediately if we find a zombie child
+                let zombie_child = unwaited_children
+                    .iter()
+                    .find(|child| child.status().is_zombie());
 
-            // wait
-            None
-        })
-    })??;
+                if let Some(zombie_child) = zombie_child {
+                    let zombie_pid = zombie_child.pid();
+                    if wait_options.contains(WaitOptions::WNOWAIT) {
+                        // does not reap child, directly return
+                        return Some(Ok(Some(zombie_child.clone())));
+                    } else {
+                        reap_zombie_child(current, zombie_pid);
+                        return Some(Ok(Some(zombie_child.clone())));
+                    }
+                }
+
+                if wait_options.contains(WaitOptions::WNOHANG) {
+                    return Some(Ok(None));
+                }
+
+                // wait
+                None
+            })
+        },
+    )??;
 
     Ok(zombie_child)
 }

kernel/src/syscall/arch/x86.rs

@@ -73,6 +73,7 @@ use crate::syscall::{
     pause::sys_pause,
     pipe::{sys_pipe, sys_pipe2},
     poll::sys_poll,
+    ppoll::sys_ppoll,
     prctl::sys_prctl,
     pread64::sys_pread64,
     preadv::{sys_preadv, sys_preadv2, sys_readv},
@@ -336,6 +337,7 @@ impl_syscall_nums_and_dispatch_fn! {
     SYS_FCHMODAT = 268         => sys_fchmodat(args[..3]);
     SYS_FACCESSAT = 269        => sys_faccessat(args[..3]);
     SYS_PSELECT6 = 270         => sys_pselect6(args[..6]);
+    SYS_PPOLL = 271            => sys_ppoll(args[..5]);
     SYS_SET_ROBUST_LIST = 273  => sys_set_robust_list(args[..2]);
     SYS_UTIMENSAT = 280        => sys_utimensat(args[..4]);
     SYS_EPOLL_PWAIT = 281      => sys_epoll_pwait(args[..6]);

kernel/src/syscall/mod.rs

@@ -80,6 +80,7 @@ mod open;
 mod pause;
 mod pipe;
 mod poll;
+mod ppoll;
 mod prctl;
 mod pread64;
 mod preadv;

kernel/src/syscall/poll.rs

@@ -10,10 +10,38 @@ use crate::{
         file_table::{FileDesc, FileTable},
     },
     prelude::*,
-    process::signal::Poller,
+    process::{signal::Poller, ResourceType},
 };
 
-pub fn sys_poll(fds: Vaddr, nfds: u64, timeout: i32, ctx: &Context) -> Result<SyscallReturn> {
+pub fn sys_poll(fds: Vaddr, nfds: u32, timeout: i32, ctx: &Context) -> Result<SyscallReturn> {
+    let timeout = if timeout >= 0 {
+        Some(Duration::from_millis(timeout as _))
+    } else {
+        None
+    };
+
+    do_sys_poll(fds, nfds, timeout, ctx)
+}
+
+pub fn do_sys_poll(
+    fds: Vaddr,
+    nfds: u32,
+    timeout: Option<Duration>,
+    ctx: &Context,
+) -> Result<SyscallReturn> {
+    if nfds as u64
+        > ctx
+            .process
+            .resource_limits()
+            .get_rlimit(ResourceType::RLIMIT_NOFILE)
+            .get_cur()
+    {
+        return_errno_with_message!(
+            Errno::EINVAL,
+            "the `nfds` value exceeds the `RLIMIT_NOFILE` value"
+        )
+    }
+
     let user_space = ctx.user_space();
 
     let poll_fds = {
@@ -33,12 +61,6 @@ pub fn sys_poll(fds: Vaddr, nfds: u64, timeout: i32, ctx: &Context) -> Result<Sy
         poll_fds
     };
 
-    let timeout = if timeout >= 0 {
-        Some(Duration::from_millis(timeout as _))
-    } else {
-        None
-    };
-
     debug!(
         "poll_fds = {:?}, nfds = {}, timeout = {:?}",
         poll_fds, nfds, timeout

kernel/src/syscall/ppoll.rs

@@ -0,0 +1,48 @@
+// SPDX-License-Identifier: MPL-2.0
+
+use core::time::Duration;
+
+use super::{poll::do_sys_poll, SyscallReturn};
+use crate::{
+    prelude::*,
+    process::signal::{sig_mask::SigMask, with_sigmask_changed},
+    time::timespec_t,
+};
+
+pub fn sys_ppoll(
+    fds: Vaddr,
+    nfds: u32,
+    timespec_addr: Vaddr,
+    sigmask_addr: Vaddr,
+    sigmask_size: usize,
+    ctx: &Context,
+) -> Result<SyscallReturn> {
+    let user_space = ctx.user_space();
+
+    let timeout = if timespec_addr != 0 {
+        let time_spec = user_space.read_val::<timespec_t>(timespec_addr)?;
+        Some(Duration::try_from(time_spec)?)
+    } else {
+        None
+    };
+
+    if sigmask_addr != 0 {
+        if sigmask_size != size_of::<SigMask>() {
+            return_errno_with_message!(Errno::EINVAL, "invalid sigmask size");
+        }
+
+        let sigmask = user_space.read_val::<SigMask>(sigmask_addr)?;
+        with_sigmask_changed(ctx, |_| sigmask, || do_sys_poll(fds, nfds, timeout, ctx))
+    } else {
+        do_sys_poll(fds, nfds, timeout, ctx)
+    }
+
+    // TODO: Write back the remaining time to `timespec_addr`.
+    //
+    // The ppoll system call should write back the remaining time,
+    // yet the function counterpart in libc hides this behavior to
+    // make the API more portable across different UNIX-like OSes.
+    // For the maximized Linux compatibility, we should follow Linux's behavior.
+    // But this cannot be readily achieved given how our internal synchronization primitives
+    // such as `Pause` and `WaitTimeout` work.
+}

kernel/src/syscall/pselect6.rs

@@ -1,10 +1,13 @@
 // SPDX-License-Identifier: MPL-2.0
 
-use core::{sync::atomic::Ordering, time::Duration};
+use core::time::Duration;
 
 use super::{select::do_sys_select, SyscallReturn};
 use crate::{
-    fs::file_table::FileDesc, prelude::*, process::signal::sig_mask::SigMask, time::timespec_t,
+    fs::file_table::FileDesc,
+    prelude::*,
+    process::signal::{sig_mask::SigMask, with_sigmask_changed},
+    time::timespec_t,
 };
 
 pub fn sys_pselect6(
@@ -17,45 +20,34 @@ pub fn sys_pselect6(
     ctx: &Context,
 ) -> Result<SyscallReturn> {
     let user_space = ctx.user_space();
-    let old_simask = if sigmask_addr != 0 {
-        let sigmask_with_size: SigMaskWithSize = user_space.read_val(sigmask_addr)?;
-
-        if !sigmask_with_size.is_valid() {
-            return_errno_with_message!(Errno::EINVAL, "sigmask size is invalid")
-        }
-        let old_sigmask = ctx
-            .posix_thread
-            .sig_mask()
-            .swap(sigmask_with_size.sigmask, Ordering::Relaxed);
-
-        Some(old_sigmask)
-    } else {
-        None
-    };
 
     let timeout = if timespec_addr != 0 {
-        let time_spec: timespec_t = user_space.read_val(timespec_addr)?;
+        let time_spec = user_space.read_val::<timespec_t>(timespec_addr)?;
         Some(Duration::try_from(time_spec)?)
     } else {
         None
     };
 
-    let res = do_sys_select(
-        nfds,
-        readfds_addr,
-        writefds_addr,
-        exceptfds_addr,
-        timeout,
-        ctx,
-    );
+    let operate = || {
+        do_sys_select(
+            nfds,
+            readfds_addr,
+            writefds_addr,
+            exceptfds_addr,
+            timeout,
+            ctx,
+        )
+    };
 
-    if let Some(old_mask) = old_simask {
-        ctx.posix_thread
-            .sig_mask()
-            .store(old_mask, Ordering::Relaxed);
+    if sigmask_addr != 0 {
+        let sigmask_with_size = user_space.read_val::<SigMaskWithSize>(sigmask_addr)?;
+        if !sigmask_with_size.is_valid() {
+            return_errno_with_message!(Errno::EINVAL, "sigmask size is invalid")
+        }
+        with_sigmask_changed(ctx, |_: SigMask| sigmask_with_size.sigmask, operate)
+    } else {
+        operate()
     }
-
-    res
 }
 
 #[repr(C)]

kernel/src/syscall/rt_sigsuspend.rs

@@ -8,7 +8,7 @@ use crate::{
     process::signal::{
         constants::{SIGKILL, SIGSTOP},
         sig_mask::SigMask,
-        with_signal_blocked,
+        with_sigmask_changed,
     },
 };
 
@@ -37,7 +37,11 @@ pub fn sys_rt_sigsuspend(
 
     // Wait until receiving any signal
     let waiter = Waiter::new_pair().0;
-    with_signal_blocked(ctx, sigmask, || waiter.pause_until(|| None::<()>))?;
+    with_sigmask_changed(
+        ctx,
+        |old_mask| old_mask + sigmask,
+        || waiter.pause_until(|| None::<()>),
+    )?;
 
     // This syscall should always return `Err(EINTR)`. This path should never be reached.
     unreachable!("rt_sigsuspend always return EINTR");

test/syscall_test/Makefile

@@ -27,6 +27,7 @@ TESTS ?= \
 	mount_test \
 	open_create_test \
 	open_test \
+	ppoll_test \
 	prctl_setuid_test \
 	pread64_test \
 	preadv2_test \

test/syscall_test/blocklists/ppoll_test

@@ -0,0 +1,3 @@
+PpollTest.NullFds
+PpollTest.SignalMaskBlocksSignal
+PpollTest.SignalMaskAllowsSignal