libs/sys-util: implement reflink_copy()

Implement reflink_copy() to copy file by reflink, and fallback to normal
file copy.

Signed-off-by: Liu Jiang <gerry@linux.alibaba.com>
Signed-off-by: Eryu Guan <eguan@linux.alibaba.com>

src/libs/kata-sys-util/src/fs.rs

@@ -5,11 +5,13 @@
 //
 
 use std::ffi::OsString;
-use std::fs;
+use std::fs::{self, File};
-use std::io::Result;
+use std::io::{Error, Result};
+use std::os::unix::io::AsRawFd;
 use std::path::{Path, PathBuf};
+use std::process::Command;
 
-use crate::eother;
+use crate::{eother, sl};
 
 // from linux.git/fs/fuse/inode.c: #define FUSE_SUPER_MAGIC 0x65735546
 const FUSE_SUPER_MAGIC: u32 = 0x65735546;
@@ -58,12 +60,114 @@ pub fn is_symlink<P: AsRef<Path>>(path: P) -> std::io::Result<bool> {
     Ok(meta.file_type().is_symlink())
 }
 
+/// Reflink copy src to dst, and falls back to regular copy if reflink copy fails.
+///
+/// # Safety
+/// The `reflink_copy()` doesn't preserve permission/security context for the copied file,
+/// so caller needs to take care of it.
+pub fn reflink_copy<S: AsRef<Path>, D: AsRef<Path>>(src: S, dst: D) -> Result<()> {
+    let src_path = src.as_ref();
+    let dst_path = dst.as_ref();
+    let src = src_path.to_string_lossy();
+    let dst = dst_path.to_string_lossy();
+
+    if !src_path.is_file() {
+        return Err(eother!("reflink_copy src {} is not a regular file", src));
+    }
+
+    // Make sure dst's parent exist. If dst is a regular file, then unlink it for later copy.
+    if dst_path.exists() {
+        if !dst_path.is_file() {
+            return Err(eother!("reflink_copy dst {} is not a regular file", dst));
+        } else {
+            fs::remove_file(dst_path)?;
+        }
+    } else if let Some(dst_parent) = dst_path.parent() {
+        if !dst_parent.exists() {
+            if let Err(e) = fs::create_dir_all(dst_parent) {
+                return Err(eother!(
+                    "reflink_copy: create_dir_all {} failed: {:?}",
+                    dst_parent.to_str().unwrap(),
+                    e
+                ));
+            }
+        } else if !dst_parent.is_dir() {
+            return Err(eother!("reflink_copy parent of {} is not a directory", dst));
+        }
+    }
+
+    // Reflink copy, and fallback to regular copy if reflink fails.
+    let src_file = fs::File::open(src_path)?;
+    let dst_file = fs::File::create(dst_path)?;
+    if let Err(e) = do_reflink_copy(src_file, dst_file) {
+        match e.raw_os_error() {
+            // Cross dev copy or filesystem doesn't support reflink, do regular copy
+            Some(os_err)
+                if os_err == nix::Error::EXDEV as i32
+                    || os_err == nix::Error::EOPNOTSUPP as i32 =>
+            {
+                warn!(
+                    sl!(),
+                    "reflink_copy: reflink is not supported ({:?}), do regular copy instead", e,
+                );
+                if let Err(e) = do_regular_copy(src.as_ref(), dst.as_ref()) {
+                    return Err(eother!(
+                        "reflink_copy: regular copy {} to {} failed: {:?}",
+                        src,
+                        dst,
+                        e
+                    ));
+                }
+            }
+            // Reflink copy failed
+            _ => {
+                return Err(eother!(
+                    "reflink_copy: copy {} to {} failed: {:?}",
+                    src,
+                    dst,
+                    e,
+                ))
+            }
+        }
+    }
+
+    Ok(())
+}
+
+// Copy file using cp command, which handles sparse file copy.
+fn do_regular_copy(src: &str, dst: &str) -> Result<()> {
+    let mut cmd = Command::new("/bin/cp");
+    cmd.args(&["--sparse=auto", src, dst]);
+
+    match cmd.output() {
+        Ok(output) => match output.status.success() {
+            true => Ok(()),
+            false => Err(eother!("`{:?}` failed: {:?}", cmd, output)),
+        },
+        Err(e) => Err(eother!("`{:?}` failed: {:?}", cmd, e)),
+    }
+}
+
+/// Copy file by reflink
+fn do_reflink_copy(src: File, dst: File) -> Result<()> {
+    use nix::ioctl_write_int;
+    // FICLONE ioctl number definition, from include/linux/fs.h
+    const FS_IOC_MAGIC: u8 = 0x94;
+    const FS_IOC_FICLONE: u8 = 9;
+    // Define FICLONE ioctl using nix::ioctl_write_int! macro.
+    // The generated function has the following signature:
+    // pub unsafe fn ficlone(fd: libc::c_int, data: libc::c_ulang) -> Result<libc::c_int>
+    ioctl_write_int!(ficlone, FS_IOC_MAGIC, FS_IOC_FICLONE);
+
+    // Safe because the `src` and `dst` are valid file objects and we have checked the result.
+    unsafe { ficlone(dst.as_raw_fd(), src.as_raw_fd() as u64) }
+        .map(|_| ())
+        .map_err(|e| Error::from_raw_os_error(e as i32))
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::mount::umount_all;
-    use std::process::Command;
-    use thiserror::private::PathAsDisplay;
 
     #[test]
     fn test_get_base_name() {
@@ -84,28 +188,17 @@ mod tests {
     }
 
     #[test]
-    fn test_is_overlayfs() {
+    fn test_reflink_copy() {
-        let tmpdir1 = tempfile::tempdir().unwrap();
+        let tmpdir = tempfile::tempdir().unwrap();
-        let tmpdir2 = tempfile::tempdir().unwrap();
+        let path = tmpdir.path().join("mounts");
-        let tmpdir3 = tempfile::tempdir().unwrap();
+        reflink_copy("/proc/mounts", &path).unwrap();
-        let tmpdir4 = tempfile::tempdir().unwrap();
+        let content = fs::read_to_string(&path).unwrap();
+        assert!(!content.is_empty());
+        reflink_copy("/proc/mounts", &path).unwrap();
+        let content = fs::read_to_string(&path).unwrap();
+        assert!(!content.is_empty());
 
-        let option = format!(
+        reflink_copy("/proc/mounts", tmpdir.path()).unwrap_err();
-            "-o lowerdir={},upperdir={},workdir={}",
+        reflink_copy("/proc/mounts_not_exist", &path).unwrap_err();
-            tmpdir1.path().as_display(),
-            tmpdir2.path().display(),
-            tmpdir3.path().display()
-        );
-        let target = format!("{}", tmpdir4.path().display());
-
-        Command::new("/bin/mount")
-            .arg("-t overlay")
-            .arg(option)
-            .arg("overlay")
-            .arg(target)
-            .output()
-            .unwrap();
-        assert!(is_overlay_fs(tmpdir4.path()));
-        umount_all(tmpdir4.path(), false).unwrap();
     }
 }