[libos] Rewrite the getdents

deps/sefs

@@ -1 +1 @@
-Subproject commit c2a1fe48b06cf9cedbad7a51d9cb846706630403
+Subproject commit 3314355400614c799e237ed1eae5682d27ca1400

src/libos/src/fs/file.rs

@@ -67,7 +67,7 @@ pub trait File: Debug + Sync + Send + Any {
         return_op_unsupported_error!("set_len")
     }
 
-    fn iterate_entries(&self, writer: &mut dyn DirentWriter) -> Result<usize> {
+    fn iterate_entries(&self, visitor: &mut dyn DirentVisitor) -> Result<usize> {
         return_op_unsupported_error!("iterate_entries")
     }
 

src/libos/src/fs/file_ops/getdents.rs

@@ -9,7 +9,7 @@ pub fn do_getdents(fd: FileDesc, buf: &mut [u8]) -> Result<usize> {
     getdents_common::<LinuxDirent>(fd, buf)
 }
 
-fn getdents_common<T: Dirent>(fd: FileDesc, buf: &mut [u8]) -> Result<usize> {
+fn getdents_common<T: DirentSerializer>(fd: FileDesc, buf: &mut [u8]) -> Result<usize> {
     debug!(
         "getdents: fd: {}, buf: {:?}, buf_size: {}",
         fd,
@@ -23,17 +23,17 @@ fn getdents_common<T: Dirent>(fd: FileDesc, buf: &mut [u8]) -> Result<usize> {
         return_errno!(ENOTDIR, "");
     }
     let mut writer = DirentBufWriter::<T>::new(buf);
-    let written_len = file_ref.iterate_entries(&mut writer)?;
+    let _ = file_ref.iterate_entries(&mut writer)?;
-    Ok(written_len)
+    Ok(writer.written_len())
 }
 
-struct DirentBufWriter<'a, T: Dirent> {
+struct DirentBufWriter<'a, T: DirentSerializer> {
     buf: &'a mut [u8],
     written_len: usize,
     phantom: PhantomData<T>,
 }
 
-impl<'a, T: Dirent> DirentBufWriter<'a, T> {
+impl<'a, T: DirentSerializer> DirentBufWriter<'a, T> {
     fn new(buf: &'a mut [u8]) -> Self {
         Self {
             buf,
@@ -41,29 +41,38 @@ impl<'a, T: Dirent> DirentBufWriter<'a, T> {
             phantom: PhantomData,
         }
     }
-}
 
-impl<'a, T: Dirent> DirentWriter for DirentBufWriter<'a, T> {
+    pub fn written_len(&self) -> usize {
-    fn write_entry(&mut self, name: &str, ino: u64, type_: FileType) -> rcore_fs::vfs::Result<()> {
-        let dirent: T = Dirent::new(name, ino, type_);
-        if self.buf.len() - self.written_len < dirent.rec_len() {
-            return Err(FsError::InvalidParam);
-        }
-        dirent
-            .serialize(&mut self.buf[self.written_len..], name, type_)
-            .map_err(|_| FsError::InvalidParam)?;
-        self.written_len += dirent.rec_len();
-        Ok(())
-    }
-
-    fn written_len(&self) -> usize {
         self.written_len
     }
 }
 
-trait Dirent: Sync + Send {
+impl<'a, T: DirentSerializer> DirentVisitor for DirentBufWriter<'a, T> {
-    fn new(name: &str, ino: u64, type_: FileType) -> Self;
+    fn visit_entry(
+        &mut self,
+        name: &str,
+        ino: u64,
+        type_: FileType,
+        offset: usize,
+    ) -> rcore_fs::vfs::Result<()> {
+        let dirent_serializer = T::new(name, ino, type_, offset);
+        if self.written_len >= self.buf.len() {
+            return Err(FsError::InvalidParam);
+        }
+        dirent_serializer
+            .serialize(&mut self.buf[self.written_len..], name, type_)
+            .map_err(|_| FsError::InvalidParam)?;
+        self.written_len += dirent_serializer.rec_len();
+        Ok(())
+    }
+}
+
+trait DirentSerializer: Sync + Send {
+    /// Creates a new DirentSerializer.
+    fn new(name: &str, ino: u64, type_: FileType, offset: usize) -> Self;
+    /// Returns the length.
     fn rec_len(&self) -> usize;
+    /// Tries to serialize a directory entry into buffer.
     fn serialize(&self, buf: &mut [u8], name: &str, type_: FileType) -> Result<()>;
 }
 
@@ -83,13 +92,13 @@ struct LinuxDirent64 {
     pub name: [u8; 0],
 }
 
-impl Dirent for LinuxDirent64 {
+impl DirentSerializer for LinuxDirent64 {
-    fn new(name: &str, ino: u64, type_: FileType) -> Self {
+    fn new(name: &str, ino: u64, type_: FileType, offset: usize) -> Self {
         let ori_len = core::mem::size_of::<Self>() + name.len() + 1;
         let len = align_up(ori_len, 8); // align up to 8 bytes
         Self {
             ino,
-            offset: 0,
+            offset: offset as u64,
             rec_len: len as u16,
             type_: DirentType::from_file_type(type_),
             name: [],
@@ -101,6 +110,10 @@ impl Dirent for LinuxDirent64 {
     }
 
     fn serialize(&self, buf: &mut [u8], name: &str, _type_: FileType) -> Result<()> {
+        if self.rec_len() > buf.len() {
+            return_errno!(EINVAL, "not enough buf");
+        }
+
         unsafe {
             let ptr = buf.as_mut_ptr() as *mut Self;
             ptr.write(*self);
@@ -131,14 +144,14 @@ struct LinuxDirent {
     */
 }
 
-impl Dirent for LinuxDirent {
+impl DirentSerializer for LinuxDirent {
-    fn new(name: &str, ino: u64, type_: FileType) -> Self {
+    fn new(name: &str, ino: u64, type_: FileType, offset: usize) -> Self {
         let ori_len =
             core::mem::size_of::<Self>() + name.len() + 1 + core::mem::size_of::<FileType>();
         let len = align_up(ori_len, 8); // align up to 8 bytes
         Self {
             ino,
-            offset: 0,
+            offset: offset as u64,
             rec_len: len as u16,
             name: [],
         }
@@ -149,6 +162,10 @@ impl Dirent for LinuxDirent {
     }
 
     fn serialize(&self, buf: &mut [u8], name: &str, type_: FileType) -> Result<()> {
+        if self.rec_len() > buf.len() {
+            return_errno!(EINVAL, "not enough buf");
+        }
+
         unsafe {
             let ptr = buf.as_mut_ptr() as *mut Self;
             ptr.write(*self);

src/libos/src/fs/hostfs.rs

@@ -250,32 +250,41 @@ impl INode for HNode {
         }
     }
 
-    fn iterate_entries(&self, ctx: &mut DirentWriterContext) -> Result<usize> {
+    fn iterate_entries(&self, offset: usize, visitor: &mut dyn DirentVisitor) -> Result<usize> {
         if !self.is_dir() {
             return Err(FsError::NotDir);
         }
-        let idx = ctx.pos();
+
-        for entry in try_std!(self.path.read_dir()).skip(idx) {
+        let try_iterate = |offset: &mut usize, visitor: &mut dyn DirentVisitor| -> Result<()> {
+            let start_offset = *offset;
+            for (idx, entry) in try_std!(self.path.read_dir())
+                .enumerate()
+                .skip_while(|(idx, _)| idx < &start_offset)
+            {
                 let entry = try_std!(entry);
-            if let Err(e) = ctx.write_entry(
+                visitor.visit_entry(
                     &entry
                         .file_name()
                         .into_string()
                         .map_err(|_| FsError::InvalidParam)?,
-                entry.ino(),
+                    entry.ino() as u64,
                     entry
                         .file_type()
                         .map_err(|_| FsError::InvalidParam)?
                         .into_fs_filetype(),
-            ) {
+                    idx,
-                if ctx.written_len() == 0 {
+                )?;
-                    return Err(e);
+                *offset = idx + 1;
-                } else {
-                    break;
             }
+
+            Ok(())
         };
+
+        let mut iterate_offset = offset;
+        match try_iterate(&mut iterate_offset, visitor) {
+            Err(e) if iterate_offset == offset => Err(e),
+            _ => Ok(iterate_offset - offset),
         }
-        Ok(ctx.written_len())
     }
 
     fn io_control(&self, cmd: u32, data: usize) -> Result<()> {

src/libos/src/fs/inode_file.rs

@@ -193,15 +193,14 @@ impl File for INodeFile {
         Ok(())
     }
 
-    fn iterate_entries(&self, writer: &mut dyn DirentWriter) -> Result<usize> {
+    fn iterate_entries(&self, visitor: &mut dyn DirentVisitor) -> Result<usize> {
         if !self.access_mode.readable() {
             return_errno!(EBADF, "File not readable. Can't read entry.");
         }
         let mut offset = self.offset.lock().unwrap();
-        let mut dir_ctx = DirentWriterContext::new(*offset, writer);
+        let visited_len = self.inode.iterate_entries(*offset, visitor)?;
-        let written_size = self.inode.iterate_entries(&mut dir_ctx)?;
+        *offset += visited_len;
-        *offset = dir_ctx.pos();
+        Ok(visited_len)
-        Ok(written_size)
     }
 
     fn access_mode(&self) -> Result<AccessMode> {

src/libos/src/fs/mod.rs

@@ -2,8 +2,7 @@ use super::*;
 
 use process;
 use rcore_fs::vfs::{
-    DirentWriter, DirentWriterContext, FileSystem, FileType, FsError, INode, Metadata, Timespec,
+    DirentVisitor, FileSystem, FileType, FsError, INode, Metadata, Timespec, PATH_MAX,
-    PATH_MAX,
 };
 use std;
 use std::any::Any;

src/libos/src/fs/procfs/mod.rs

@@ -157,44 +157,51 @@ impl DirProcINode for LockedProcRootINode {
         }
     }
 
-    fn iterate_entries(&self, mut ctx: &mut DirentWriterContext) -> vfs::Result<usize> {
+    fn iterate_entries(
+        &self,
+        offset: usize,
+        visitor: &mut dyn DirentVisitor,
+    ) -> vfs::Result<usize> {
         let file = self.0.read().unwrap();
-        let idx = ctx.pos();
 
-        // Write first two special entries
+        let try_iterate =
-        if idx == 0 {
+            |mut offset: &mut usize, mut visitor: &mut dyn DirentVisitor| -> vfs::Result<()> {
+                // The two special entries
+                if *offset == 0 {
                     let this_inode = file.this.upgrade().unwrap();
-            write_inode_entry!(&mut ctx, ".", &this_inode);
+                    rcore_fs::visit_inode_entry!(&mut visitor, ".", &this_inode, &mut offset);
                 }
-        if idx <= 1 {
+                if *offset == 1 {
                     let parent_inode = file.this.upgrade().unwrap();
-            write_inode_entry!(&mut ctx, "..", &parent_inode);
+                    rcore_fs::visit_inode_entry!(&mut visitor, "..", &parent_inode, &mut offset);
                 }
 
-        // Write the non volatile entries
+                // The non volatile entries
-        let skipped = if idx < 2 { 0 } else { idx - 2 };
+                let start_offset = *offset;
-        for (name, inode) in file.non_volatile_entries.iter().skip(skipped) {
+                for (name, child) in file.non_volatile_entries.iter().skip(start_offset - 2) {
-            write_inode_entry!(&mut ctx, name, inode);
+                    rcore_fs::visit_inode_entry!(&mut visitor, name, child, &mut offset);
                 }
 
-        // Write the pid entries
+                // The pid entries
-        let skipped = {
+                let start_offset = *offset;
-            let prior_entries_len = 2 + file.non_volatile_entries.len();
+                let skipped_len = 2 + file.non_volatile_entries.len();
-            if idx < prior_entries_len {
+                for process in get_all_processes().iter().skip(start_offset - skipped_len) {
-                0
+                    rcore_fs::visit_entry!(
-            } else {
+                        &mut visitor,
-                idx - prior_entries_len
-            }
-        };
-        for process in get_all_processes().iter().skip(skipped) {
-            write_entry!(
-                &mut ctx,
                         &process.pid().to_string(),
                         PROC_INO,
-                vfs::FileType::Dir
+                        vfs::FileType::Dir,
+                        &mut offset
                     );
                 }
 
-        Ok(ctx.written_len())
+                Ok(())
+            };
+
+        let mut iterate_offset = offset;
+        match try_iterate(&mut iterate_offset, visitor) {
+            Err(e) if iterate_offset == offset => Err(e),
+            _ => Ok(iterate_offset - offset),
+        }
     }
 }

src/libos/src/fs/procfs/pid/fd.rs

@@ -53,26 +53,45 @@ impl DirProcINode for LockedProcFdDirINode {
         }
     }
 
-    fn iterate_entries(&self, mut ctx: &mut DirentWriterContext) -> vfs::Result<usize> {
+    fn iterate_entries(
+        &self,
+        offset: usize,
+        visitor: &mut dyn DirentVisitor,
+    ) -> vfs::Result<usize> {
         let file = self.0.read().unwrap();
-        let idx = ctx.pos();
 
-        // Write first two special entries
+        let try_iterate =
-        write_first_two_entries!(idx, &mut ctx, &file);
+            |mut offset: &mut usize, mut visitor: &mut dyn DirentVisitor| -> vfs::Result<()> {
+                // The two special entries
+                visit_first_two_entries!(&mut visitor, &file, &mut offset);
 
-        // Write the fd entries
+                // The fd entries
-        let skipped = if idx < 2 { 0 } else { idx - 2 };
                 let main_thread = match file.process_ref.main_thread() {
                     Some(main_thread) => main_thread,
                     None => {
-                return Ok(ctx.written_len());
+                        return Ok(());
                     }
                 };
                 let fds = main_thread.files().lock().unwrap().fds();
-        for fd in fds.iter().skip(skipped) {
+                let start_offset = *offset;
-            write_entry!(&mut ctx, &fd.to_string(), PROC_INO, vfs::FileType::SymLink);
+                for fd in fds.iter().skip(start_offset - 2) {
+                    rcore_fs::visit_entry!(
+                        &mut visitor,
+                        &fd.to_string(),
+                        PROC_INO,
+                        vfs::FileType::SymLink,
+                        &mut offset
+                    );
+                }
+
+                Ok(())
+            };
+
+        let mut iterate_offset = offset;
+        match try_iterate(&mut iterate_offset, visitor) {
+            Err(e) if iterate_offset == offset => Err(e),
+            _ => Ok(iterate_offset - offset),
         }
-        Ok(ctx.written_len())
     }
 }
 

src/libos/src/fs/procfs/pid/mod.rs

@@ -113,23 +113,42 @@ impl DirProcINode for LockedPidDirINode {
         }
     }
 
-    fn iterate_entries(&self, mut ctx: &mut DirentWriterContext) -> vfs::Result<usize> {
+    fn iterate_entries(
+        &self,
+        offset: usize,
+        visitor: &mut dyn DirentVisitor,
+    ) -> vfs::Result<usize> {
         let file = self.0.read().unwrap();
-        let idx = ctx.pos();
 
-        // Write first two special entries
+        let try_iterate =
-        write_first_two_entries!(idx, &mut ctx, &file);
+            |mut offset: &mut usize, mut visitor: &mut dyn DirentVisitor| -> vfs::Result<()> {
+                // The two special entries
+                visit_first_two_entries!(&mut visitor, &file, &mut offset);
 
-        // Write the normal entries
+                // The normal entries
-        let skipped = if idx < 2 { 0 } else { idx - 2 };
+                let start_offset = *offset;
-        for (name, inode) in file.entries.iter().skip(skipped) {
+                for (name, child) in file.entries.iter().skip(start_offset - 2) {
-            write_inode_entry!(&mut ctx, name, inode);
+                    rcore_fs::visit_inode_entry!(&mut visitor, name, child, &mut offset);
                 }
 
-        // Write the fd entry
+                // The fd entry
-        if idx <= 2 + file.entries.len() {
+                if *offset == 2 + file.entries.len() {
-            write_entry!(&mut ctx, "fd", PROC_INO, vfs::FileType::Dir);
+                    rcore_fs::visit_entry!(
+                        &mut visitor,
+                        "fd",
+                        PROC_INO as u64,
+                        vfs::FileType::Dir,
+                        &mut offset
+                    );
+                }
+
+                Ok(())
+            };
+
+        let mut iterate_offset = offset;
+        match try_iterate(&mut iterate_offset, visitor) {
+            Err(e) if iterate_offset == offset => Err(e),
+            _ => Ok(iterate_offset - offset),
         }
-        Ok(ctx.written_len())
     }
 }

src/libos/src/fs/procfs/proc_inode/dir.rs

@@ -65,8 +65,12 @@ where
         self.inner().get_entry(id)
     }
 
-    fn iterate_entries(&self, ctx: &mut DirentWriterContext) -> vfs::Result<usize> {
+    fn iterate_entries(
-        self.inner().iterate_entries(ctx)
+        &self,
+        offset: usize,
+        visitor: &mut dyn DirentVisitor,
+    ) -> vfs::Result<usize> {
+        self.inner().iterate_entries(offset, visitor)
     }
 
     fn as_any_ref(&self) -> &dyn Any {

src/libos/src/fs/procfs/proc_inode/mod.rs

@@ -15,7 +15,8 @@ pub trait ProcINode {
 pub trait DirProcINode {
     fn find(&self, name: &str) -> vfs::Result<Arc<dyn INode>>;
     fn get_entry(&self, id: usize) -> vfs::Result<String>;
-    fn iterate_entries(&self, ctx: &mut DirentWriterContext) -> vfs::Result<usize>;
+    fn iterate_entries(&self, offset: usize, visitor: &mut dyn DirentVisitor)
+        -> vfs::Result<usize>;
 }
 
 #[macro_export]
@@ -54,7 +55,11 @@ macro_rules! impl_inode_for_file_or_symlink {
             Err(FsError::NotDir)
         }
 
-        fn iterate_entries(&self, ctx: &mut DirentWriterContext) -> vfs::Result<usize> {
+        fn iterate_entries(
+            &self,
+            offset: usize,
+            visitor: &mut dyn DirentVisitor,
+        ) -> vfs::Result<usize> {
             Err(FsError::NotDir)
         }
 
@@ -65,47 +70,19 @@ macro_rules! impl_inode_for_file_or_symlink {
 }
 
 #[macro_export]
-macro_rules! write_first_two_entries {
+macro_rules! visit_first_two_entries {
-    ($idx: expr, $ctx:expr, $file:expr) => {
+    ($visitor:expr, $file:expr, $offset: expr) => {
-        let idx = $idx;
+        use rcore_fs::visit_inode_entry;
         let file = $file;
 
-        if idx == 0 {
+        let offset = **$offset;
+        if offset == 0 {
             let this_inode = file.this.upgrade().unwrap();
-            write_inode_entry!($ctx, ".", &this_inode);
+            visit_inode_entry!($visitor, ".", &this_inode, $offset);
-        }
-        if idx <= 1 {
-            write_inode_entry!($ctx, "..", &file.parent);
-        }
-    };
-}
-
-#[macro_export]
-macro_rules! write_inode_entry {
-    ($ctx:expr, $name:expr, $inode:expr) => {
-        let ctx = $ctx;
-        let name = $name;
-        let ino = $inode.metadata()?.inode;
-        let type_ = $inode.metadata()?.type_;
-
-        write_entry!(ctx, name, ino, type_);
-    };
-}
-
-#[macro_export]
-macro_rules! write_entry {
-    ($ctx:expr, $name:expr, $ino:expr, $type_:expr) => {
-        let ctx = $ctx;
-        let name = $name;
-        let ino = $ino;
-        let type_ = $type_;
-
-        if let Err(e) = ctx.write_entry(name, ino as u64, type_) {
-            if ctx.written_len() == 0 {
-                return Err(e);
-            } else {
-                return Ok(ctx.written_len());
         }
+        let offset = **$offset;
+        if offset == 1 {
+            visit_inode_entry!($visitor, "..", &file.parent, $offset);
         }
     };
 }