Add and refactor read-write syscalls

docs/src/kernel/linux-compatibility.md

@@ -38,8 +38,8 @@ provided by Linux on x86-64 architecture.
 | 15      | rt_sigreturn     | ✅              |
 | 16      | ioctl            | ✅              |
 | 17      | pread64          | ✅              |
-| 18      | pwrite64         | ❌              |
-| 19      | readv            | ❌              |
+| 18      | pwrite64         | ✅              |
+| 19      | readv            | ✅              |
 | 20      | writev           | ✅              |
 | 21      | access           | ✅              |
 | 22      | pipe             | ✅              |
@@ -315,8 +315,8 @@ provided by Linux on x86-64 architecture.
 | 292     | dup3             | ✅              |
 | 293     | pipe2            | ✅              |
 | 294     | inotify_init1    | ❌              |
-| 295     | preadv           | ❌              |
-| 296     | pwritev          | ❌              |
+| 295     | preadv           | ✅              |
+| 296     | pwritev          | ✅              |
 | 297     | rt_tgsigqueueinfo | ❌             |
 | 298     | perf_event_open  | ❌              |
 | 299     | recvmmsg         | ❌              |
@@ -336,6 +336,8 @@ provided by Linux on x86-64 architecture.
 | 313	  | finit_module     | ❌              |
 | 318	  | getrandom        | ✅              |
 | 322	  | execveat         | ✅              |
+| 327	  | preadv2          | ✅              |
+| 328	  | pwritev2         | ✅              |
 | 435	  | clone3           | ✅              |
 
 ## File Systems

kernel/aster-nix/src/fs/file_handle.rs

@@ -32,7 +32,7 @@ pub trait FileLike: Send + Sync + Any {
     ///
     /// [`read`]: FileLike::read
     fn read_at(&self, offset: usize, buf: &mut [u8]) -> Result<usize> {
-        return_errno_with_message!(Errno::EINVAL, "read_at is not supported");
+        return_errno_with_message!(Errno::ESPIPE, "read_at is not supported");
     }
 
     /// Write at the given file offset.
@@ -43,7 +43,7 @@ pub trait FileLike: Send + Sync + Any {
     ///
     /// [`write`]: FileLike::write
     fn write_at(&self, offset: usize, buf: &[u8]) -> Result<usize> {
-        return_errno_with_message!(Errno::EINVAL, "write_at is not supported");
+        return_errno_with_message!(Errno::ESPIPE, "write_at is not supported");
     }
 
     fn ioctl(&self, cmd: IoctlCmd, arg: usize) -> Result<i32> {

kernel/aster-nix/src/syscall/arch/x86.rs

@@ -66,7 +66,10 @@ use crate::syscall::{
     poll::sys_poll,
     prctl::sys_prctl,
     pread64::sys_pread64,
+    preadv::{sys_preadv, sys_preadv2, sys_readv},
     prlimit64::sys_prlimit64,
+    pwrite64::sys_pwrite64,
+    pwritev::{sys_pwritev, sys_pwritev2, sys_writev},
     read::sys_read,
     readlink::{sys_readlink, sys_readlinkat},
     recvfrom::sys_recvfrom,
@@ -121,7 +124,6 @@ use crate::syscall::{
     wait4::sys_wait4,
     waitid::sys_waitid,
     write::sys_write,
-    writev::sys_writev,
 };
 
 impl_syscall_nums_and_dispatch_fn! {
@@ -143,6 +145,8 @@ impl_syscall_nums_and_dispatch_fn! {
     SYS_RT_SIGRETURN = 15      => sys_rt_sigreturn(args[..0], &mut context);
     SYS_IOCTL = 16             => sys_ioctl(args[..3]);
     SYS_PREAD64 = 17           => sys_pread64(args[..4]);
+    SYS_PWRITE64 = 18          => sys_pwrite64(args[..4]);
+    SYS_READV = 19             => sys_readv(args[..3]);
     SYS_WRITEV = 20            => sys_writev(args[..3]);
     SYS_ACCESS = 21            => sys_access(args[..2]);
     SYS_PIPE = 22              => sys_pipe(args[..1]);
@@ -282,8 +286,12 @@ impl_syscall_nums_and_dispatch_fn! {
     SYS_EPOLL_CREATE1 = 291    => sys_epoll_create1(args[..1]);
     SYS_DUP3 = 292             => sys_dup3(args[..3]);
     SYS_PIPE2 = 293            => sys_pipe2(args[..2]);
+    SYS_PREADV = 295           => sys_preadv(args[..4]);
+    SYS_PWRITEV = 296          => sys_pwritev(args[..4]);
     SYS_PRLIMIT64 = 302        => sys_prlimit64(args[..4]);
     SYS_GETRANDOM = 318        => sys_getrandom(args[..3]);
     SYS_EXECVEAT = 322         => sys_execveat(args[..5], &mut context);
+    SYS_PREADV2 = 327          => sys_preadv2(args[..5]);
+    SYS_PWRITEV2 = 328         => sys_pwritev2(args[..5]);
     SYS_CLONE3 = 435           => sys_clone3(args[..2], &context);
 }

kernel/aster-nix/src/syscall/mod.rs

@@ -73,7 +73,10 @@ mod pipe;
 mod poll;
 mod prctl;
 mod pread64;
+mod preadv;
 mod prlimit64;
+mod pwrite64;
+mod pwritev;
 mod read;
 mod readlink;
 mod recvfrom;
@@ -128,7 +131,6 @@ mod utimens;
 mod wait4;
 mod waitid;
 mod write;
-mod writev;
 
 /// This macro is used to define syscall handler.
 /// The first param is ths number of parameters,

kernel/aster-nix/src/syscall/pread64.rs

@@ -1,29 +1,39 @@
 // SPDX-License-Identifier: MPL-2.0
 
 use super::SyscallReturn;
-use crate::{
-    fs::{file_table::FileDesc, utils::SeekFrom},
-    prelude::*,
-    util::write_bytes_to_user,
-};
+use crate::{fs::file_table::FileDesc, prelude::*, util::write_bytes_to_user};
 
-pub fn sys_pread64(fd: FileDesc, buf_ptr: Vaddr, count: usize, pos: i64) -> Result<SyscallReturn> {
+pub fn sys_pread64(
+    fd: FileDesc,
+    user_buf_ptr: Vaddr,
+    user_buf_len: usize,
+    offset: i64,
+) -> Result<SyscallReturn> {
     debug!(
-        "fd = {}, buf = 0x{:x}, count = 0x{:x}, pos = 0x{:x}",
-        fd, buf_ptr, count, pos
+        "fd = {}, buf = 0x{:x}, user_buf_len = 0x{:x}, offset = 0x{:x}",
+        fd, user_buf_ptr, user_buf_len, offset
     );
 
-    let current = current!();
-    let file_table = current.file_table().lock();
-    let file = file_table.get_file(fd)?;
-
-    let seek_from = SeekFrom::Start(pos as usize);
-    file.seek(seek_from)?;
+    if offset < 0 {
+        return_errno_with_message!(Errno::EINVAL, "offset cannot be negative");
+    }
+    let file = {
+        let current = current!();
+        let filetable = current.file_table().lock();
+        filetable.get_file(fd)?.clone()
+    };
+    // TODO: Check (f.file->f_mode & FMODE_PREAD); We don't have f_mode in our FileLike trait
+    if user_buf_len == 0 {
+        return Ok(SyscallReturn::Return(0));
+    }
+    if offset.checked_add(user_buf_len as i64).is_none() {
+        return_errno_with_message!(Errno::EINVAL, "offset + user_buf_len overflow");
+    }
 
     let read_len = {
-        let mut buffer = vec![0u8; count];
-        let read_len = file.read(&mut buffer)?;
-        write_bytes_to_user(buf_ptr, &buffer)?;
+        let mut buffer = vec![0u8; user_buf_len];
+        let read_len = file.read_at(offset as usize, &mut buffer)?;
+        write_bytes_to_user(user_buf_ptr, &buffer)?;
         read_len
     };
 

kernel/aster-nix/src/syscall/preadv.rs

@@ -0,0 +1,171 @@
+// SPDX-License-Identifier: MPL-2.0
+
+use super::SyscallReturn;
+use crate::{
+    fs::file_table::FileDesc,
+    prelude::*,
+    util::{copy_iovs_from_user, IoVec},
+};
+
+pub fn sys_readv(fd: FileDesc, io_vec_ptr: Vaddr, io_vec_count: usize) -> Result<SyscallReturn> {
+    let res = do_sys_readv(fd, io_vec_ptr, io_vec_count)?;
+    Ok(SyscallReturn::Return(res as _))
+}
+
+pub fn sys_preadv(
+    fd: FileDesc,
+    io_vec_ptr: Vaddr,
+    io_vec_count: usize,
+    offset: i64,
+) -> Result<SyscallReturn> {
+    let res = do_sys_preadv(fd, io_vec_ptr, io_vec_count, offset, RWFFlag::empty())?;
+    Ok(SyscallReturn::Return(res as _))
+}
+
+pub fn sys_preadv2(
+    fd: FileDesc,
+    io_vec_ptr: Vaddr,
+    io_vec_count: usize,
+    offset: i64,
+    flags: u32,
+) -> Result<SyscallReturn> {
+    let flags = match RWFFlag::from_bits(flags) {
+        Some(flags) => flags,
+        None => return_errno_with_message!(Errno::EINVAL, "invalid flags"),
+    };
+    let res = if offset == -1 {
+        do_sys_readv(fd, io_vec_ptr, io_vec_count)?
+    } else {
+        do_sys_preadv(fd, io_vec_ptr, io_vec_count, offset, flags)?
+    };
+    Ok(SyscallReturn::Return(res as _))
+}
+
+fn do_sys_preadv(
+    fd: FileDesc,
+    io_vec_ptr: Vaddr,
+    io_vec_count: usize,
+    offset: i64,
+    _flags: RWFFlag,
+) -> Result<usize> {
+    debug!(
+        "preadv: fd = {}, io_vec_ptr = 0x{:x}, io_vec_counter = 0x{:x}, offset = 0x{:x}",
+        fd, io_vec_ptr, io_vec_count, offset
+    );
+
+    if offset < 0 {
+        return_errno_with_message!(Errno::EINVAL, "offset cannot be negative");
+    }
+
+    let file = {
+        let current = current!();
+        let filetable = current.file_table().lock();
+        filetable.get_file(fd)?.clone()
+    };
+
+    if io_vec_count == 0 {
+        return Ok(0);
+    }
+
+    // Calculate the total buffer length and check for overflow
+    let total_len = io_vec_count
+        .checked_mul(core::mem::size_of::<IoVec>())
+        .and_then(|val| val.checked_add(offset as usize));
+    if total_len.is_none() {
+        return_errno_with_message!(Errno::EINVAL, "offset + io_vec_count overflow");
+    }
+
+    let mut total_len: usize = 0;
+    let mut cur_offset = offset as usize;
+
+    let io_vecs = copy_iovs_from_user(io_vec_ptr, io_vec_count)?;
+    for io_vec in io_vecs.as_ref() {
+        if io_vec.is_empty() {
+            continue;
+        }
+        if total_len.checked_add(io_vec.len()).is_none()
+            || total_len
+                .checked_add(io_vec.len())
+                .and_then(|sum| sum.checked_add(cur_offset))
+                .is_none()
+            || total_len
+                .checked_add(io_vec.len())
+                .and_then(|sum| sum.checked_add(cur_offset))
+                .map(|sum| sum > isize::MAX as usize)
+                .unwrap_or(false)
+        {
+            return_errno_with_message!(Errno::EINVAL, "Total length overflow");
+        }
+
+        let mut buffer = vec![0u8; io_vec.len()];
+
+        // TODO: According to the man page
+        // at <https://man7.org/linux/man-pages/man2/readv.2.html>,
+        // readv must be atomic,
+        // but the current implementation does not ensure atomicity.
+        // A suitable fix would be to add a `readv` method for the `FileLike` trait,
+        // allowing each subsystem to implement atomicity.
+        let read_len = file.read_at(cur_offset, &mut buffer)?;
+        io_vec.write_exact_to_user(&buffer)?;
+        total_len += read_len;
+        cur_offset += read_len;
+        if read_len == 0 || read_len < buffer.len() {
+            // End of file reached or no more data to read
+            break;
+        }
+    }
+
+    Ok(total_len)
+}
+
+fn do_sys_readv(fd: FileDesc, io_vec_ptr: Vaddr, io_vec_count: usize) -> Result<usize> {
+    debug!(
+        "fd = {}, io_vec_ptr = 0x{:x}, io_vec_counter = 0x{:x}",
+        fd, io_vec_ptr, io_vec_count
+    );
+
+    let file = {
+        let current = current!();
+        let filetable = current.file_table().lock();
+        filetable.get_file(fd)?.clone()
+    };
+
+    if io_vec_count == 0 {
+        return Ok(0);
+    }
+
+    let mut total_len = 0;
+
+    let io_vecs = copy_iovs_from_user(io_vec_ptr, io_vec_count)?;
+    for io_vec in io_vecs.as_ref() {
+        if io_vec.is_empty() {
+            continue;
+        }
+
+        let mut buffer = vec![0u8; io_vec.len()];
+        // TODO: According to the man page
+        // at <https://man7.org/linux/man-pages/man2/readv.2.html>,
+        // readv must be atomic,
+        // but the current implementation does not ensure atomicity.
+        // A suitable fix would be to add a `readv` method for the `FileLike` trait,
+        // allowing each subsystem to implement atomicity.
+        let read_len = file.read(&mut buffer)?;
+        io_vec.write_exact_to_user(&buffer)?;
+        total_len += read_len;
+        if read_len == 0 || read_len < buffer.len() {
+            // End of file reached or no more data to read
+            break;
+        }
+    }
+
+    Ok(total_len)
+}
+
+bitflags! {
+    struct RWFFlag: u32 {
+        const RWF_DSYNC = 0x00000001;
+        const RWF_HIPRI = 0x00000002;
+        const RWF_SYNC = 0x00000004;
+        const RWF_NOWAIT = 0x00000008;
+    }
+}

kernel/aster-nix/src/syscall/pwrite64.rs

@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MPL-2.0
+
+use super::SyscallReturn;
+use crate::{fs::file_table::FileDesc, prelude::*, util::read_bytes_from_user};
+
+pub fn sys_pwrite64(
+    fd: FileDesc,
+    user_buf_ptr: Vaddr,
+    user_buf_len: usize,
+    offset: i64,
+) -> Result<SyscallReturn> {
+    debug!(
+        "fd = {}, user_buf_ptr = 0x{:x}, user_buf_len = 0x{:x}, offset = 0x{:x}",
+        fd, user_buf_ptr, user_buf_len, offset
+    );
+    if offset < 0 {
+        return_errno_with_message!(Errno::EINVAL, "offset cannot be negative");
+    }
+    let file = {
+        let current = current!();
+        let filetable = current.file_table().lock();
+        filetable.get_file(fd)?.clone()
+    };
+    // TODO: Check (f.file->f_mode & FMODE_PWRITE); We don't have f_mode in our FileLike trait
+    if user_buf_len == 0 {
+        return Ok(SyscallReturn::Return(0));
+    }
+    if offset.checked_add(user_buf_len as i64).is_none() {
+        return_errno_with_message!(Errno::EINVAL, "offset + user_buf_len overflow");
+    }
+
+    let mut buffer = vec![0u8; user_buf_len];
+    read_bytes_from_user(user_buf_ptr, &mut buffer)?;
+    let write_len = file.write_at(offset as _, &buffer)?;
+    Ok(SyscallReturn::Return(write_len as _))
+}

kernel/aster-nix/src/syscall/pwritev.rs

@@ -0,0 +1,149 @@
+// SPDX-License-Identifier: MPL-2.0
+
+use super::SyscallReturn;
+use crate::{fs::file_table::FileDesc, prelude::*, util::copy_iovs_from_user};
+
+pub fn sys_writev(fd: FileDesc, io_vec_ptr: Vaddr, io_vec_count: usize) -> Result<SyscallReturn> {
+    let res = do_sys_writev(fd, io_vec_ptr, io_vec_count)?;
+    Ok(SyscallReturn::Return(res as _))
+}
+
+pub fn sys_pwritev(
+    fd: FileDesc,
+    io_vec_ptr: Vaddr,
+    io_vec_count: usize,
+    offset: i64,
+) -> Result<SyscallReturn> {
+    let res = do_sys_pwritev(fd, io_vec_ptr, io_vec_count, offset, RWFFlag::empty())?;
+    Ok(SyscallReturn::Return(res as _))
+}
+
+pub fn sys_pwritev2(
+    fd: FileDesc,
+    io_vec_ptr: Vaddr,
+    io_vec_count: usize,
+    offset: i64,
+    flags: u32,
+) -> Result<SyscallReturn> {
+    let flags = match RWFFlag::from_bits(flags) {
+        Some(flags) => flags,
+        None => return_errno_with_message!(Errno::EINVAL, "invalid flags"),
+    };
+    let res = if offset == -1 {
+        do_sys_writev(fd, io_vec_ptr, io_vec_count)?
+    } else {
+        do_sys_pwritev(fd, io_vec_ptr, io_vec_count, offset, flags)?
+    };
+    Ok(SyscallReturn::Return(res as _))
+}
+
+fn do_sys_pwritev(
+    fd: FileDesc,
+    io_vec_ptr: Vaddr,
+    io_vec_count: usize,
+    offset: i64,
+    _flags: RWFFlag,
+) -> Result<usize> {
+    // TODO: Implement flags support
+    debug!(
+        "fd = {}, io_vec_ptr = 0x{:x}, io_vec_counter = 0x{:x}, offset = 0x{:x}",
+        fd, io_vec_ptr, io_vec_count, offset
+    );
+    if offset < 0 {
+        return_errno_with_message!(Errno::EINVAL, "offset cannot be negative");
+    }
+    let file = {
+        let current = current!();
+        let filetable = current.file_table().lock();
+        filetable.get_file(fd)?.clone()
+    };
+    // TODO: Check (f.file->f_mode & FMODE_PREAD); We don't have f_mode in our FileLike trait
+    if io_vec_count == 0 {
+        return Ok(0);
+    }
+
+    let mut total_len: usize = 0;
+    let mut cur_offset = offset as usize;
+
+    let io_vecs = copy_iovs_from_user(io_vec_ptr, io_vec_count)?;
+    for io_vec in io_vecs.as_ref() {
+        if io_vec.is_empty() {
+            continue;
+        }
+        if total_len.checked_add(io_vec.len()).is_none()
+            || total_len
+                .checked_add(io_vec.len())
+                .and_then(|sum| sum.checked_add(cur_offset))
+                .is_none()
+            || total_len
+                .checked_add(io_vec.len())
+                .and_then(|sum| sum.checked_add(cur_offset))
+                .map(|sum| sum > isize::MAX as usize)
+                .unwrap_or(false)
+        {
+            return_errno_with_message!(Errno::EINVAL, "Total length overflow");
+        }
+
+        let buffer = {
+            let mut buffer = vec![0u8; io_vec.len()];
+            io_vec.read_exact_from_user(&mut buffer)?;
+            buffer
+        };
+
+        // TODO: According to the man page
+        // at <https://man7.org/linux/man-pages/man2/readv.2.html>,
+        // writev must be atomic,
+        // but the current implementation does not ensure atomicity.
+        // A suitable fix would be to add a `writev` method for the `FileLike` trait,
+        // allowing each subsystem to implement atomicity.
+        let write_len = file.write_at(cur_offset, &buffer)?;
+        total_len += write_len;
+        cur_offset += write_len;
+    }
+    Ok(total_len)
+}
+
+fn do_sys_writev(fd: FileDesc, io_vec_ptr: Vaddr, io_vec_count: usize) -> Result<usize> {
+    debug!(
+        "fd = {}, io_vec_ptr = 0x{:x}, io_vec_counter = 0x{:x}",
+        fd, io_vec_ptr, io_vec_count
+    );
+    let file = {
+        let current = current!();
+        let filetable = current.file_table().lock();
+        filetable.get_file(fd)?.clone()
+    };
+    let mut total_len = 0;
+
+    let io_vecs = copy_iovs_from_user(io_vec_ptr, io_vec_count)?;
+    for io_vec in io_vecs.as_ref() {
+        if io_vec.is_empty() {
+            continue;
+        }
+
+        let buffer = {
+            let mut buffer = vec![0u8; io_vec.len()];
+            io_vec.read_exact_from_user(&mut buffer)?;
+            buffer
+        };
+
+        // TODO: According to the man page
+        // at <https://man7.org/linux/man-pages/man2/readv.2.html>,
+        // writev must be atomic,
+        // but the current implementation does not ensure atomicity.
+        // A suitable fix would be to add a `writev` method for the `FileLike` trait,
+        // allowing each subsystem to implement atomicity.
+        let write_len = file.write(&buffer)?;
+        total_len += write_len;
+    }
+    Ok(total_len)
+}
+
+bitflags! {
+    struct RWFFlag: u32 {
+        const RWF_DSYNC = 0x00000001;
+        const RWF_HIPRI = 0x00000002;
+        const RWF_SYNC = 0x00000004;
+        const RWF_NOWAIT = 0x00000008;
+    }
+}

kernel/aster-nix/src/syscall/writev.rs

@@ -1,48 +0,0 @@
-// SPDX-License-Identifier: MPL-2.0
-
-#![allow(dead_code)]
-
-use super::SyscallReturn;
-use crate::{fs::file_table::FileDesc, prelude::*, util::copy_iovs_from_user};
-
-pub fn sys_writev(fd: FileDesc, io_vec_ptr: Vaddr, io_vec_count: usize) -> Result<SyscallReturn> {
-    let res = do_sys_writev(fd, io_vec_ptr, io_vec_count)?;
-    Ok(SyscallReturn::Return(res as _))
-}
-
-fn do_sys_writev(fd: FileDesc, io_vec_ptr: Vaddr, io_vec_count: usize) -> Result<usize> {
-    debug!(
-        "fd = {}, io_vec_ptr = 0x{:x}, io_vec_counter = 0x{:x}",
-        fd, io_vec_ptr, io_vec_count
-    );
-    let file = {
-        let current = current!();
-        let filetable = current.file_table().lock();
-        filetable.get_file(fd)?.clone()
-    };
-
-    let mut total_len = 0;
-
-    let io_vecs = copy_iovs_from_user(io_vec_ptr, io_vec_count)?;
-    for io_vec in io_vecs.as_ref() {
-        if io_vec.is_empty() {
-            continue;
-        }
-
-        let buffer = {
-            let mut buffer = vec![0u8; io_vec.len()];
-            io_vec.read_exact_from_user(&mut buffer)?;
-            buffer
-        };
-
-        // FIXME: According to the man page
-        // at <https://man7.org/linux/man-pages/man2/readv.2.html>,
-        // writev must be atomic,
-        // but the current implementation does not ensure atomicity.
-        // A suitable fix would be to add a `writev` method for the `FileLike` trait,
-        // allowing each subsystem to implement atomicity.
-        let write_len = file.write(&buffer)?;
-        total_len += write_len;
-    }
-    Ok(total_len)
-}

regression/syscall_test/Makefile

@@ -24,19 +24,13 @@ TESTS ?= \
 	mount_test \
 	open_create_test \
 	open_test \
+	pread64_test \
+	preadv2_test \
+	pwrite64_test \
+	pwritev2_test \
 	pty_test \
 	read_test \
-	rename_test \
-	sendfile_test \
-	stat_test \
-	statfs_test \
-	symlink_test \
-	sync_test \
-	timers_test \
-	truncate_test \
-	uidgid_test \
-	unlink_test \
-	vdso_clock_gettime_test \
+	readv_test \
 	write_test \
 	utimes_test \
 	# The end of the list

regression/syscall_test/blocklists/eventfd_test

@@ -1,3 +1,2 @@
-EventfdTest.IllegalPwrite
 EventfdTest.SpliceFromPipePartialSucceeds
 EventfdTest.NotifyNonZero_NoRandomSave

regression/syscall_test/blocklists/pread64_test

@@ -0,0 +1 @@
+Pread64Test.Overflow

regression/syscall_test/blocklists/preadv2_test

@@ -0,0 +1,3 @@
+Preadv2Test.TestInvalidFlag
+Preadv2Test.TestUnreadableFile
+Preadv2Test.TestUnseekableFileInvalid

regression/syscall_test/blocklists/pwritev2_test

@@ -0,0 +1 @@
+Pwritev2Test.InvalidFlag

regression/syscall_test/blocklists/readv_test

@@ -0,0 +1,5 @@
+ReadvTest.BadIovecBase_File
+ReadvTest.BadIovecBase_Pipe
+ReadvTest.NotReadable_Pipe
+ReadvTest.IovecOutsideTaskAddressRangeInNonemptyArray
+ReadvTestNoFixture.TruncatedAtMax_NoRandomSave