libfuse_fs/unionfs/
async_io.rs

1use super::utils;
2use super::{CachePolicy, HandleData, Inode, OverlayFs, RealHandle};
3use crate::util::open_options::OpenOptions;
4use rfuse3::raw::prelude::*;
5use rfuse3::*;
6use std::ffi::OsStr;
7use std::io::Error;
8use std::io::ErrorKind;
9use std::num::NonZeroU32;
10use std::sync::Arc;
11use std::sync::atomic::{AtomicU64, Ordering};
12use tokio::sync::Mutex;
13use tracing::info;
14use tracing::trace;
15
16impl Filesystem for OverlayFs {
17    /// initialize filesystem. Called before any other filesystem method.
18    async fn init(&self, req: Request) -> Result<ReplyInit> {
19        for layer in self.lower_layers.iter() {
20            layer.init(req).await?;
21        }
22        if let Some(upper) = &self.upper_layer {
23            upper.init(req).await?;
24        }
25        if self.config.do_import {
26            self.import().await?;
27        }
28        if !self.config.do_import || self.config.writeback {
29            self.writeback.store(true, Ordering::Relaxed);
30        }
31        if !self.config.do_import || self.config.no_open {
32            self.no_open.store(true, Ordering::Relaxed);
33        }
34        if !self.config.do_import || self.config.no_opendir {
35            self.no_opendir.store(true, Ordering::Relaxed);
36        }
37        if !self.config.do_import || self.config.killpriv_v2 {
38            self.killpriv_v2.store(true, Ordering::Relaxed);
39        }
40        if self.config.perfile_dax {
41            self.perfile_dax.store(true, Ordering::Relaxed);
42        }
43
44        Ok(ReplyInit {
45            max_write: NonZeroU32::new(128 * 1024).unwrap(),
46        })
47    }
48
49    /// clean up filesystem. Called on filesystem exit which is fuseblk, in normal fuse filesystem,
50    /// kernel may call forget for root. There is some discuss for this
51    /// <https://github.com/bazil/fuse/issues/82#issuecomment-88126886>,
52    /// <https://sourceforge.net/p/fuse/mailman/message/31995737/>
53    async fn destroy(&self, _req: Request) {}
54
55    /// look up a directory entry by name and get its attributes.
56    async fn lookup(&self, req: Request, parent: Inode, name: &OsStr) -> Result<ReplyEntry> {
57        let tmp = name.to_string_lossy().to_string();
58        let result = self.do_lookup(req, parent, tmp.as_str()).await;
59        match result {
60            Ok(e) => Ok(e),
61            Err(err) => Err(err.into()),
62        }
63    }
64
65    /// forget an inode. The nlookup parameter indicates the number of lookups previously
66    /// performed on this inode. If the filesystem implements inode lifetimes, it is recommended
67    /// that inodes acquire a single reference on each lookup, and lose nlookup references on each
68    /// forget. The filesystem may ignore forget calls, if the inodes don't need to have a limited
69    /// lifetime. On unmount it is not guaranteed, that all referenced inodes will receive a forget
70    /// message. When filesystem is normal(not fuseblk) and unmounting, kernel may send forget
71    /// request for root and this library will stop session after call forget. There is some
72    /// discussion for this <https://github.com/bazil/fuse/issues/82#issuecomment-88126886>,
73    /// <https://sourceforge.net/p/fuse/mailman/message/31995737/>
74    async fn forget(&self, _req: Request, inode: Inode, nlookup: u64) {
75        self.forget_one(inode, nlookup).await;
76    }
77
78    /// get file attributes. If `fh` is None, means `fh` is not set.
79    async fn getattr(
80        &self,
81        req: Request,
82        inode: Inode,
83        fh: Option<u64>,
84        flags: u32,
85    ) -> Result<ReplyAttr> {
86        if !self.no_open.load(Ordering::Relaxed)
87            && let Some(h) = fh
88        {
89            let handles = self.handles.lock().await;
90            if let Some(hd) = handles.get(&h)
91                && let Some(ref rh) = hd.real_handle
92            {
93                let mut rep: ReplyAttr = rh
94                    .layer
95                    .getattr(req, rh.inode, Some(rh.handle.load(Ordering::Relaxed)), 0)
96                    .await?;
97                rep.attr.ino = inode;
98                return Ok(rep);
99            }
100        }
101
102        let node: Arc<super::OverlayInode> = self.lookup_node(req, inode, "").await?;
103        let (layer, _, lower_inode) = node.first_layer_inode().await;
104        let mut re = layer.getattr(req, lower_inode, None, flags).await?;
105        re.attr.ino = inode;
106        Ok(re)
107    }
108
109    /// set file attributes. If `fh` is None, means `fh` is not set.
110    async fn setattr(
111        &self,
112        req: Request,
113        inode: Inode,
114        fh: Option<u64>,
115        set_attr: SetAttr,
116    ) -> Result<ReplyAttr> {
117        // Check if upper layer exists.
118        self.upper_layer
119            .as_ref()
120            .cloned()
121            .ok_or_else(|| Error::from_raw_os_error(libc::EROFS))?;
122
123        // deal with handle first
124        if !self.no_open.load(Ordering::Relaxed)
125            && let Some(h) = fh
126        {
127            let handles = self.handles.lock().await;
128            if let Some(hd) = handles.get(&h)
129                && let Some(ref rhd) = hd.real_handle
130            {
131                // handle opened in upper layer
132                if rhd.in_upper_layer {
133                    let mut rep = rhd
134                        .layer
135                        .setattr(
136                            req,
137                            rhd.inode,
138                            Some(rhd.handle.load(Ordering::Relaxed)),
139                            set_attr,
140                        )
141                        .await?;
142                    rep.attr.ino = inode;
143                    return Ok(rep);
144                }
145            }
146        }
147
148        let mut node = self.lookup_node(req, inode, "").await?;
149
150        if !node.in_upper_layer().await {
151            node = self.copy_node_up(req, node.clone()).await?
152        }
153
154        let (layer, _, real_inode) = node.first_layer_inode().await;
155        // layer.setattr(req, real_inode, None, set_attr).await
156        let mut rep = layer.setattr(req, real_inode, None, set_attr).await?;
157        rep.attr.ino = inode;
158        Ok(rep)
159    }
160
161    /// read symbolic link.
162    async fn readlink(&self, req: Request, inode: Inode) -> Result<ReplyData> {
163        trace!("READLINK: inode: {inode}\n");
164
165        let node = self.lookup_node(req, inode, "").await?;
166
167        if node.whiteout.load(Ordering::Relaxed) {
168            return Err(Error::from_raw_os_error(libc::ENOENT).into());
169        }
170
171        let (layer, _, inode) = node.first_layer_inode().await;
172        layer.readlink(req, inode).await
173    }
174
175    /// create a symbolic link.
176    async fn symlink(
177        &self,
178        req: Request,
179        parent: Inode,
180        name: &OsStr,
181        link: &OsStr,
182    ) -> Result<ReplyEntry> {
183        // soft link
184        let sname = name.to_string_lossy().into_owned().to_owned();
185        let slinkname = link.to_string_lossy().into_owned().to_owned();
186
187        let pnode = self.lookup_node(req, parent, "").await?;
188        self.do_symlink(req, slinkname.as_str(), &pnode, sname.as_str())
189            .await?;
190
191        self.do_lookup(req, parent, sname.as_str())
192            .await
193            .map_err(|e| e.into())
194    }
195
196    /// create file node. Create a regular file, character device, block device, fifo or socket
197    /// node. When creating file, most cases user only need to implement
198    /// [`create`][Filesystem::create].
199    async fn mknod(
200        &self,
201        req: Request,
202        parent: Inode,
203        name: &OsStr,
204        mode: u32,
205        rdev: u32,
206    ) -> Result<ReplyEntry> {
207        let sname = name.to_string_lossy().to_string();
208
209        // Check if parent exists.
210        let pnode = self.lookup_node(req, parent, "").await?;
211        if pnode.whiteout.load(Ordering::Relaxed) {
212            return Err(Error::from_raw_os_error(libc::ENOENT).into());
213        }
214
215        self.do_mknod(req, &pnode, sname.as_str(), mode, rdev, 0)
216            .await?;
217        self.do_lookup(req, parent, sname.as_str())
218            .await
219            .map_err(|e| e.into())
220    }
221
222    /// create a directory.
223    async fn mkdir(
224        &self,
225        req: Request,
226        parent: Inode,
227        name: &OsStr,
228        mode: u32,
229        umask: u32,
230    ) -> Result<ReplyEntry> {
231        let sname = name.to_string_lossy().to_string();
232
233        // no entry or whiteout
234        let pnode = self.lookup_node(req, parent, "").await?;
235        if pnode.whiteout.load(Ordering::Relaxed) {
236            return Err(Error::from_raw_os_error(libc::ENOENT).into());
237        }
238
239        self.do_mkdir(req, pnode, sname.as_str(), mode, umask)
240            .await?;
241        self.do_lookup(req, parent, sname.as_str())
242            .await
243            .map_err(|e| e.into())
244    }
245
246    /// remove a file.
247    async fn unlink(&self, req: Request, parent: Inode, name: &OsStr) -> Result<()> {
248        self.do_rm(req, parent, name, false)
249            .await
250            .map_err(|e| e.into())
251    }
252
253    /// remove a directory.
254    async fn rmdir(&self, req: Request, parent: Inode, name: &OsStr) -> Result<()> {
255        self.do_rm(req, parent, name, true)
256            .await
257            .map_err(|e| e.into())
258    }
259
260    /// rename a file or directory.
261    async fn rename(
262        &self,
263        req: Request,
264        parent: Inode,
265        name: &OsStr,
266        new_parent: Inode,
267        new_name: &OsStr,
268    ) -> Result<()> {
269        self.do_rename(req, parent, name, new_parent, new_name)
270            .await
271            .map_err(|e| e.into())
272    }
273
274    /// create a hard link.
275    async fn link(
276        &self,
277        req: Request,
278        inode: Inode,
279        new_parent: Inode,
280        new_name: &OsStr,
281    ) -> Result<ReplyEntry> {
282        let node = self.lookup_node(req, inode, "").await?;
283        if node.whiteout.load(Ordering::Relaxed) {
284            return Err(Error::from_raw_os_error(libc::ENOENT).into());
285        }
286
287        let newpnode = self.lookup_node(req, new_parent, "").await?;
288        if newpnode.whiteout.load(Ordering::Relaxed) {
289            return Err(Error::from_raw_os_error(libc::ENOENT).into());
290        }
291        let new_name = new_name.to_str().unwrap();
292        // trace!(
293        //     "LINK: inode: {}, new_parent: {}, trying to do_link: src_inode: {}, newpnode: {}",
294        //     inode, new_parent, node.inode, newpnode.inode
295        // );
296        self.do_link(req, &node, &newpnode, new_name).await?;
297        // trace!("LINK: done, looking up new entry");
298        self.do_lookup(req, new_parent, new_name)
299            .await
300            .map_err(|e| e.into())
301    }
302
303    /// open a file. Open flags (with the exception of `O_CREAT`, `O_EXCL` and `O_NOCTTY`) are
304    /// available in flags. Filesystem may store an arbitrary file handle (pointer, index, etc) in
305    /// fh, and use this in other all other file operations (read, write, flush, release, fsync).
306    /// Filesystem may also implement stateless file I/O and not store anything in fh. There are
307    /// also some flags (`direct_io`, `keep_cache`) which the filesystem may set, to change the way
308    /// the file is opened. A filesystem need not implement this method if it
309    /// sets [`MountOptions::no_open_support`][crate::MountOptions::no_open_support] and if the
310    /// kernel supports `FUSE_NO_OPEN_SUPPORT`.
311    ///
312    /// # Notes:
313    ///
314    /// See `fuse_file_info` structure in
315    /// [fuse_common.h](https://libfuse.github.io/doxygen/include_2fuse__common_8h_source.html) for
316    /// more details.
317    async fn open(&self, req: Request, inode: Inode, flags: u32) -> Result<ReplyOpen> {
318        if self.no_open.load(Ordering::Relaxed) {
319            info!("fuse: open is not supported.");
320            return Err(Error::from_raw_os_error(libc::ENOSYS).into());
321        }
322
323        let readonly: bool = flags
324            & (libc::O_APPEND | libc::O_CREAT | libc::O_TRUNC | libc::O_RDWR | libc::O_WRONLY)
325                as u32
326            == 0;
327        // toggle flags
328        let mut flags: i32 = flags as i32;
329
330        flags |= libc::O_NOFOLLOW;
331
332        if self.config.writeback {
333            if flags & libc::O_ACCMODE == libc::O_WRONLY {
334                flags &= !libc::O_ACCMODE;
335                flags |= libc::O_RDWR;
336            }
337
338            if flags & libc::O_APPEND != 0 {
339                flags &= !libc::O_APPEND;
340            }
341        }
342        // lookup node
343        let node = self.lookup_node(req, inode, "").await?;
344
345        // whiteout node
346        if node.whiteout.load(Ordering::Relaxed) {
347            return Err(Error::from_raw_os_error(libc::ENOENT).into());
348        }
349
350        if !readonly {
351            // copy up to upper layer
352            self.copy_node_up(req, node.clone()).await?;
353        }
354
355        // assign a handle in overlayfs and open it
356        let (_l, h) = node.open(req, flags as u32, 0).await?;
357
358        let hd = self.next_handle.fetch_add(1, Ordering::Relaxed);
359        let (layer, in_upper_layer, inode) = node.first_layer_inode().await;
360        let handle_data = HandleData {
361            node: node.clone(),
362            real_handle: Some(RealHandle {
363                layer,
364                in_upper_layer,
365                inode,
366                handle: AtomicU64::new(h.fh),
367            }),
368            dir_snapshot: Mutex::new(None),
369        };
370
371        self.handles.lock().await.insert(hd, Arc::new(handle_data));
372
373        let mut opts = OpenOptions::empty();
374        match self.config.cache_policy {
375            CachePolicy::Never => opts |= OpenOptions::DIRECT_IO,
376            CachePolicy::Always => opts |= OpenOptions::KEEP_CACHE,
377            _ => {}
378        }
379        // trace!("OPEN: returning handle: {hd}");
380
381        Ok(ReplyOpen {
382            fh: hd,
383            flags: opts.bits(),
384        })
385    }
386
387    /// read data. Read should send exactly the number of bytes requested except on EOF or error,
388    /// otherwise the rest of the data will be substituted with zeroes. An exception to this is
389    /// when the file has been opened in `direct_io` mode, in which case the return value of the
390    /// read system call will reflect the return value of this operation. `fh` will contain the
391    /// value set by the open method, or will be undefined if the open method didn't set any value.
392    async fn read(
393        &self,
394        req: Request,
395        inode: Inode,
396        fh: u64,
397        offset: u64,
398        size: u32,
399    ) -> Result<ReplyData> {
400        let data = self.get_data(req, Some(fh), inode, 0).await?;
401
402        match data.real_handle {
403            None => Err(Error::from_raw_os_error(libc::ENOENT).into()),
404            Some(ref hd) => {
405                hd.layer
406                    .read(
407                        req,
408                        hd.inode,
409                        hd.handle.load(Ordering::Relaxed),
410                        offset,
411                        size,
412                    )
413                    .await
414            }
415        }
416    }
417
418    /// write data. Write should return exactly the number of bytes requested except on error. An
419    /// exception to this is when the file has been opened in `direct_io` mode, in which case the
420    /// return value of the write system call will reflect the return value of this operation. `fh`
421    /// will contain the value set by the open method, or will be undefined if the open method
422    /// didn't set any value. When `write_flags` contains
423    /// [`FUSE_WRITE_CACHE`](crate::raw::flags::FUSE_WRITE_CACHE), means the write operation is a
424    /// delay write.
425    #[allow(clippy::too_many_arguments)]
426    async fn write(
427        &self,
428        req: Request,
429        inode: Inode,
430        fh: u64,
431        offset: u64,
432        data: &[u8],
433        write_flags: u32,
434        flags: u32,
435    ) -> Result<ReplyWrite> {
436        let handle_data: Arc<HandleData> = self.get_data(req, Some(fh), inode, flags).await?;
437
438        match handle_data.real_handle {
439            None => Err(Error::from_raw_os_error(libc::ENOENT).into()),
440            Some(ref hd) => {
441                hd.layer
442                    .write(
443                        req,
444                        hd.inode,
445                        hd.handle.load(Ordering::Relaxed),
446                        offset,
447                        data,
448                        write_flags,
449                        flags,
450                    )
451                    .await
452            }
453        }
454    }
455
456    /// Copy a range of data from one file to another. This can improve performance because it
457    /// reduces data copying: normally, data will be copied from FUSE server to kernel, then to
458    /// user-space, then to kernel, and finally sent back to FUSE server. By implementing this
459    /// method, data will only be copied internally within the FUSE server.
460    #[allow(clippy::too_many_arguments)]
461    async fn copy_file_range(
462        &self,
463        req: Request,
464        inode_in: Inode,
465        fh_in: u64,
466        offset_in: u64,
467        inode_out: Inode,
468        fh_out: u64,
469        offset_out: u64,
470        length: u64,
471        flags: u64,
472    ) -> Result<ReplyCopyFileRange> {
473        // Get handle data for source file
474        let data_in = self.get_data(req, Some(fh_in), inode_in, 0).await?;
475        let handle_in = match data_in.real_handle {
476            None => return Err(Error::from_raw_os_error(libc::ENOENT).into()),
477            Some(ref hd) => hd,
478        };
479
480        // Get handle data for destination file
481        let data_out = self.get_data(req, Some(fh_out), inode_out, 0).await?;
482        let handle_out = match data_out.real_handle {
483            None => return Err(Error::from_raw_os_error(libc::ENOENT).into()),
484            Some(ref hd) => hd,
485        };
486
487        // Both files must be on the same layer for copy_file_range to work
488        if !Arc::ptr_eq(&handle_in.layer, &handle_out.layer) {
489            // Different layers - return EXDEV to trigger fallback to read/write
490            return Err(Error::from_raw_os_error(libc::EXDEV).into());
491        }
492
493        // Delegate to the underlying PassthroughFs layer
494        handle_in
495            .layer
496            .copy_file_range(
497                req,
498                handle_in.inode,
499                handle_in.handle.load(Ordering::Relaxed),
500                offset_in,
501                handle_out.inode,
502                handle_out.handle.load(Ordering::Relaxed),
503                offset_out,
504                length,
505                flags,
506            )
507            .await
508    }
509
510    /// get filesystem statistics.
511    async fn statfs(&self, req: Request, inode: Inode) -> Result<ReplyStatFs> {
512        self.do_statvfs(req, inode).await.map_err(|e| e.into())
513    }
514
515    /// release an open file. Release is called when there are no more references to an open file:
516    /// all file descriptors are closed and all memory mappings are unmapped. For every open call
517    /// there will be exactly one release call. The filesystem may reply with an error, but error
518    /// values are not returned to `close()` or `munmap()` which triggered the release. `fh` will
519    /// contain the value set by the open method, or will be undefined if the open method didn't
520    /// set any value. `flags` will contain the same flags as for open. `flush` means flush the
521    /// data or not when closing file.
522    async fn release(
523        &self,
524        req: Request,
525        _inode: Inode,
526        fh: u64,
527        flags: u32,
528        lock_owner: u64,
529        flush: bool,
530    ) -> Result<()> {
531        if self.no_open.load(Ordering::Relaxed) {
532            info!("fuse: release is not supported.");
533            return Err(Error::from_raw_os_error(libc::ENOSYS).into());
534        }
535
536        if let Some(hd) = self.handles.lock().await.get(&fh) {
537            let rh = if let Some(ref h) = hd.real_handle {
538                h
539            } else {
540                return Err(
541                    Error::other(format!("no real handle found for file handle {fh}")).into(),
542                );
543            };
544            let real_handle = rh.handle.load(Ordering::Relaxed);
545            let real_inode = rh.inode;
546            rh.layer
547                .release(req, real_inode, real_handle, flags, lock_owner, flush)
548                .await?;
549        }
550
551        self.handles.lock().await.remove(&fh);
552
553        Ok(())
554    }
555
556    /// synchronize file contents. If the `datasync` is true, then only the user data should be
557    /// flushed, not the metadata.
558    async fn fsync(&self, req: Request, inode: Inode, fh: u64, datasync: bool) -> Result<()> {
559        self.do_fsync(req, inode, datasync, fh, false)
560            .await
561            .map_err(|e| e.into())
562    }
563
564    /// set an extended attribute.
565    async fn setxattr(
566        &self,
567        req: Request,
568        inode: Inode,
569        name: &OsStr,
570        value: &[u8],
571        flags: u32,
572        position: u32,
573    ) -> Result<()> {
574        let node = self.lookup_node(req, inode, "").await?;
575
576        if node.whiteout.load(Ordering::Relaxed) {
577            return Err(Error::from_raw_os_error(libc::ENOENT).into());
578        }
579
580        if !node.in_upper_layer().await {
581            // Copy node up.
582            self.copy_node_up(req, node.clone()).await?;
583        }
584
585        let (layer, _, real_inode) = node.first_layer_inode().await;
586
587        layer
588            .setxattr(req, real_inode, name, value, flags, position)
589            .await
590    }
591
592    /// Get an extended attribute. If `size` is too small, return `Err<ERANGE>`.
593    /// Otherwise, use [`ReplyXAttr::Data`] to send the attribute data, or
594    /// return an error.
595    async fn getxattr(
596        &self,
597        req: Request,
598        inode: Inode,
599        name: &OsStr,
600        size: u32,
601    ) -> Result<ReplyXAttr> {
602        let node = self.lookup_node(req, inode, "").await?;
603
604        if node.whiteout.load(Ordering::Relaxed) {
605            return Err(Error::from_raw_os_error(libc::ENOENT).into());
606        }
607
608        let (layer, real_inode) = self.find_real_inode(inode).await?;
609
610        layer.getxattr(req, real_inode, name, size).await
611    }
612
613    /// List extended attribute names.
614    ///
615    /// If `size` is too small, return `Err<ERANGE>`.  Otherwise, use
616    /// [`ReplyXAttr::Data`] to send the attribute list, or return an error.
617    async fn listxattr(&self, req: Request, inode: Inode, size: u32) -> Result<ReplyXAttr> {
618        let node = self.lookup_node(req, inode, "").await?;
619        if node.whiteout.load(Ordering::Relaxed) {
620            return Err(Error::from_raw_os_error(libc::ENOENT).into());
621        }
622        let (layer, real_inode) = self.find_real_inode(inode).await?;
623        layer.listxattr(req, real_inode, size).await
624    }
625
626    /// remove an extended attribute.
627    async fn removexattr(&self, req: Request, inode: Inode, name: &OsStr) -> Result<()> {
628        let node = self.lookup_node(req, inode, "").await?;
629
630        if node.whiteout.load(Ordering::Relaxed) {
631            return Err(Error::from_raw_os_error(libc::ENOENT).into());
632        }
633
634        if !node.in_upper_layer().await {
635            // copy node into upper layer
636            self.copy_node_up(req, node.clone()).await?;
637        }
638
639        let (layer, _, ino) = node.first_layer_inode().await;
640        layer.removexattr(req, ino, name).await
641
642        // TODO: recreate the node since removexattr may remove the opaque xattr.
643    }
644
645    /// flush method. This is called on each `close()` of the opened file. Since file descriptors
646    /// can be duplicated (`dup`, `dup2`, `fork`), for one open call there may be many flush calls.
647    /// Filesystems shouldn't assume that flush will always be called after some writes, or that if
648    /// will be called at all. `fh` will contain the value set by the open method, or will be
649    /// undefined if the open method didn't set any value.
650    ///
651    /// # Notes:
652    ///
653    /// the name of the method is misleading, since (unlike fsync) the filesystem is not forced to
654    /// flush pending writes. One reason to flush data, is if the filesystem wants to return write
655    /// errors. If the filesystem supports file locking operations ([`setlk`][Filesystem::setlk],
656    /// [`getlk`][Filesystem::getlk]) it should remove all locks belonging to `lock_owner`.
657    async fn flush(&self, req: Request, inode: Inode, fh: u64, lock_owner: u64) -> Result<()> {
658        if self.no_open.load(Ordering::Relaxed) {
659            return Err(Error::from_raw_os_error(libc::ENOSYS).into());
660        }
661
662        let node = self.lookup_node(req, inode, "").await;
663        match node {
664            Ok(n) => {
665                if n.whiteout.load(Ordering::Relaxed) {
666                    return Err(Error::from_raw_os_error(libc::ENOENT).into());
667                }
668            }
669            Err(e) => {
670                if e.raw_os_error() == Some(libc::ENOENT) {
671                    trace!("flush: inode {inode} is stale");
672                } else {
673                    return Err(e.into());
674                }
675            }
676        }
677
678        let (layer, real_inode, real_handle) = self.find_real_info_from_handle(fh).await?;
679
680        // FIXME: need to test if inode matches corresponding handle?
681        if inode
682            != self
683                .handles
684                .lock()
685                .await
686                .get(&fh)
687                .map(|h| h.node.inode)
688                .unwrap_or(0)
689        {
690            return Err(Error::other("inode does not match handle").into());
691        }
692
693        trace!("flushing, real_inode: {real_inode}, real_handle: {real_handle}");
694        layer.flush(req, real_inode, real_handle, lock_owner).await
695    }
696
697    /// open a directory. Filesystem may store an arbitrary file handle (pointer, index, etc) in
698    /// `fh`, and use this in other all other directory stream operations
699    /// ([`readdir`][Filesystem::readdir], [`releasedir`][Filesystem::releasedir],
700    /// [`fsyncdir`][Filesystem::fsyncdir]). Filesystem may also implement stateless directory
701    /// I/O and not store anything in `fh`.  A file system need not implement this method if it
702    /// sets [`MountOptions::no_open_dir_support`][crate::MountOptions::no_open_dir_support] and
703    /// if the kernel supports `FUSE_NO_OPENDIR_SUPPORT`.
704    async fn opendir(&self, req: Request, inode: Inode, flags: u32) -> Result<ReplyOpen> {
705        if self.no_opendir.load(Ordering::Relaxed) {
706            info!("fuse: opendir is not supported.");
707            return Err(Error::from_raw_os_error(libc::ENOSYS).into());
708        }
709
710        // lookup node
711        let node = self.lookup_node(req, inode, ".").await?;
712
713        if node.whiteout.load(Ordering::Relaxed) {
714            return Err(Error::from_raw_os_error(libc::ENOENT).into());
715        }
716
717        let st = node.stat64(req).await?;
718        if !utils::is_dir(&st.attr.kind) {
719            return Err(Error::from_raw_os_error(libc::ENOTDIR).into());
720        }
721
722        let handle = self.next_handle.fetch_add(1, Ordering::Relaxed);
723        // Get the layer information and open directory in the underlying layer
724        let (layer, in_upper_layer, real_inode) = node.first_layer_inode().await;
725        let reply = layer.opendir(req, real_inode, flags).await?;
726
727        self.handles.lock().await.insert(
728            handle,
729            Arc::new(HandleData {
730                node: Arc::clone(&node),
731                real_handle: Some(RealHandle {
732                    layer,
733                    in_upper_layer,
734                    inode: real_inode,
735                    handle: AtomicU64::new(reply.fh),
736                }),
737                dir_snapshot: Mutex::new(None),
738            }),
739        );
740
741        Ok(ReplyOpen { fh: handle, flags })
742    }
743
744    /// read directory. `offset` is used to track the offset of the directory entries. `fh` will
745    /// contain the value set by the [`opendir`][Filesystem::opendir] method, or will be
746    /// undefined if the [`opendir`][Filesystem::opendir] method didn't set any value.
747    async fn readdir<'a>(
748        &'a self,
749        req: Request,
750        parent: Inode,
751        fh: u64,
752        offset: i64,
753    ) -> Result<
754        ReplyDirectory<
755            impl futures_util::stream::Stream<Item = Result<DirectoryEntry>> + Send + 'a,
756        >,
757    > {
758        if self.config.no_readdir {
759            info!("fuse: readdir is not supported.");
760            return Err(Error::from_raw_os_error(libc::ENOTDIR).into());
761        }
762        let entries = self
763            .do_readdir(req, parent, fh, offset.try_into().unwrap())
764            .await?;
765        Ok(ReplyDirectory { entries })
766    }
767
768    /// read directory entries, but with their attribute, like [`readdir`][Filesystem::readdir]
769    /// + [`lookup`][Filesystem::lookup] at the same time.
770    async fn readdirplus<'a>(
771        &'a self,
772        req: Request,
773        parent: Inode,
774        fh: u64,
775        offset: u64,
776        _lock_owner: u64,
777    ) -> Result<
778        ReplyDirectoryPlus<
779            impl futures_util::stream::Stream<Item = Result<DirectoryEntryPlus>> + Send + 'a,
780        >,
781    > {
782        if self.config.no_readdir {
783            info!("fuse: readdir is not supported.");
784            return Err(Error::from_raw_os_error(libc::ENOTDIR).into());
785        }
786        trace!("readdirplus: parent: {parent}, fh: {fh}, offset: {offset}");
787        let entries = self.do_readdirplus(req, parent, fh, offset).await?;
788        match self.handles.lock().await.get(&fh) {
789            Some(h) => {
790                trace!(
791                    "after readdirplus: found handle, seeing real_handle: {}",
792                    h.real_handle.is_some()
793                );
794            }
795            None => trace!("after readdirplus: no handle found: {fh}"),
796        }
797        Ok(ReplyDirectoryPlus { entries })
798    }
799    /// release an open directory. For every [`opendir`][Filesystem::opendir] call there will
800    /// be exactly one `releasedir` call. `fh` will contain the value set by the
801    /// [`opendir`][Filesystem::opendir] method, or will be undefined if the
802    /// [`opendir`][Filesystem::opendir] method didn't set any value.
803    async fn releasedir(&self, req: Request, _inode: Inode, fh: u64, flags: u32) -> Result<()> {
804        if self.no_opendir.load(Ordering::Relaxed) {
805            info!("fuse: releasedir is not supported.");
806            return Err(Error::from_raw_os_error(libc::ENOSYS).into());
807        }
808
809        if let Some(hd) = self.handles.lock().await.get(&fh) {
810            let rh = if let Some(ref h) = hd.real_handle {
811                h
812            } else {
813                return Err(
814                    Error::other(format!("no real handle found for file handle {fh}")).into(),
815                );
816            };
817            let real_handle = rh.handle.load(Ordering::Relaxed);
818            let real_inode = rh.inode;
819            rh.layer
820                .releasedir(req, real_inode, real_handle, flags)
821                .await?;
822        }
823
824        self.handles.lock().await.remove(&fh);
825        Ok(())
826    }
827
828    /// synchronize directory contents. If the `datasync` is true, then only the directory contents
829    /// should be flushed, not the metadata. `fh` will contain the value set by the
830    /// [`opendir`][Filesystem::opendir] method, or will be undefined if the
831    /// [`opendir`][Filesystem::opendir] method didn't set any value.
832    async fn fsyncdir(&self, req: Request, inode: Inode, fh: u64, datasync: bool) -> Result<()> {
833        self.do_fsync(req, inode, datasync, fh, true)
834            .await
835            .map_err(|e| e.into())
836    }
837    /// check file access permissions. This will be called for the `access()` system call. If the
838    /// `default_permissions` mount option is given, this method is not be called. This method is
839    /// not called under Linux kernel versions 2.4.x.
840    async fn access(&self, req: Request, inode: Inode, mask: u32) -> Result<()> {
841        let node = self.lookup_node(req, inode, "").await?;
842
843        if node.whiteout.load(Ordering::Relaxed) {
844            return Err(Error::from_raw_os_error(libc::ENOENT).into());
845        }
846
847        let (layer, real_inode) = self.find_real_inode(inode).await?;
848        layer.access(req, real_inode, mask).await
849    }
850
851    /// create and open a file. If the file does not exist, first create it with the specified
852    /// mode, and then open it. Open flags (with the exception of `O_NOCTTY`) are available in
853    /// flags. Filesystem may store an arbitrary file handle (pointer, index, etc) in `fh`, and use
854    /// this in other all other file operations ([`read`][Filesystem::read],
855    /// [`write`][Filesystem::write], [`flush`][Filesystem::flush],
856    /// [`release`][Filesystem::release], [`fsync`][Filesystem::fsync]). There are also some flags
857    /// (`direct_io`, `keep_cache`) which the filesystem may set, to change the way the file is
858    /// opened. If this method is not implemented or under Linux kernel versions earlier than
859    /// 2.6.15, the [`mknod`][Filesystem::mknod] and [`open`][Filesystem::open] methods will be
860    /// called instead.
861    ///
862    /// # Notes:
863    ///
864    /// See `fuse_file_info` structure in
865    /// [fuse_common.h](https://libfuse.github.io/doxygen/include_2fuse__common_8h_source.html) for
866    /// more details.
867    async fn create(
868        &self,
869        req: Request,
870        parent: Inode,
871        name: &OsStr,
872        mode: u32,
873        flags: u32,
874    ) -> Result<ReplyCreated> {
875        // Parent doesn't exist.
876        let pnode = self.lookup_node(req, parent, "").await?;
877        if pnode.whiteout.load(Ordering::Relaxed) {
878            return Err(Error::from_raw_os_error(libc::ENOENT).into());
879        }
880
881        let mut flags: i32 = flags as i32;
882        flags |= libc::O_NOFOLLOW;
883        flags &= !libc::O_DIRECT;
884        if self.config.writeback {
885            if flags & libc::O_ACCMODE == libc::O_WRONLY {
886                flags &= !libc::O_ACCMODE;
887                flags |= libc::O_RDWR;
888            }
889
890            if flags & libc::O_APPEND != 0 {
891                flags &= !libc::O_APPEND;
892            }
893        }
894
895        let final_handle = self
896            .do_create(req, &pnode, name, mode, flags.try_into().unwrap())
897            .await?;
898        let entry = self.do_lookup(req, parent, name.to_str().unwrap()).await?;
899        let fh = final_handle
900            .ok_or_else(|| std::io::Error::new(ErrorKind::NotFound, "Handle not found"))?;
901
902        let mut opts = OpenOptions::empty();
903        match self.config.cache_policy {
904            CachePolicy::Never => opts |= OpenOptions::DIRECT_IO,
905            CachePolicy::Always => opts |= OpenOptions::KEEP_CACHE,
906            _ => {}
907        }
908
909        Ok(ReplyCreated {
910            ttl: entry.ttl,
911            attr: entry.attr,
912            generation: entry.generation,
913            fh,
914            flags: opts.bits(),
915        })
916    }
917
918    /// forget more than one inode. This is a batch version [`forget`][Filesystem::forget]
919    async fn batch_forget(&self, _req: Request, inodes: &[(Inode, u64)]) {
920        for inode in inodes {
921            self.forget_one(inode.0, inode.1).await;
922        }
923    }
924
925    /// allocate space for an open file. This function ensures that required space is allocated for
926    /// specified file.
927    ///
928    /// # Notes:
929    ///
930    /// more information about `fallocate`, please see **`man 2 fallocate`**
931    async fn fallocate(
932        &self,
933        req: Request,
934        inode: Inode,
935        fh: u64,
936        offset: u64,
937        length: u64,
938        mode: u32,
939    ) -> Result<()> {
940        // Use O_RDONLY flags which indicates no copy up.
941        let data = self
942            .get_data(req, Some(fh), inode, libc::O_RDONLY as u32)
943            .await?;
944
945        match data.real_handle {
946            None => Err(Error::from_raw_os_error(libc::ENOENT).into()),
947            Some(ref rhd) => {
948                if !rhd.in_upper_layer {
949                    // TODO: in lower layer, error out or just success?
950                    return Err(Error::from_raw_os_error(libc::EROFS).into());
951                }
952                rhd.layer
953                    .fallocate(
954                        req,
955                        rhd.inode,
956                        rhd.handle.load(Ordering::Relaxed),
957                        offset,
958                        length,
959                        mode,
960                    )
961                    .await
962            }
963        }
964    }
965
966    /// find next data or hole after the specified offset.
967    async fn lseek(
968        &self,
969        req: Request,
970        inode: Inode,
971        fh: u64,
972        offset: u64,
973        whence: u32,
974    ) -> Result<ReplyLSeek> {
975        let node = self.lookup_node(req, inode, "").await?;
976
977        if node.whiteout.load(Ordering::Relaxed) {
978            return Err(Error::from_raw_os_error(libc::ENOENT).into());
979        }
980
981        let st = node.stat64(req).await?;
982        if utils::is_dir(&st.attr.kind) {
983            // Special handling and security restrictions for directory operations.
984            // Use the common API to obtain the underlying layer and handle info.
985            let (layer, real_inode, real_handle) = self.find_real_info_from_handle(fh).await?;
986
987            // Verify that the underlying handle refers to a directory.
988            let handle_stat = match layer.getattr(req, real_inode, Some(real_handle), 0).await {
989                Ok(s) => s,
990                Err(_) => return Err(Error::from_raw_os_error(libc::EBADF).into()),
991            };
992
993            if !utils::is_dir(&handle_stat.attr.kind) {
994                return Err(Error::from_raw_os_error(libc::ENOTDIR).into());
995            }
996
997            // Handle directory lseek operations according to POSIX standard
998            // This enables seekdir/telldir functionality on directories
999            match whence {
1000                // SEEK_SET: Set the directory position to an absolute value
1001                x if x == libc::SEEK_SET as u32 => {
1002                    // Validate offset bounds to prevent overflow
1003                    // Directory offsets should not exceed i64::MAX
1004                    if offset > i64::MAX as u64 {
1005                        return Err(Error::from_raw_os_error(libc::EINVAL).into());
1006                    }
1007
1008                    // Perform the seek operation on the underlying layer
1009                    // Delegate to the lower layer implementation
1010                    layer
1011                        .lseek(req, real_inode, real_handle, offset, whence)
1012                        .await
1013                }
1014                // SEEK_CUR: Move relative to the current directory position
1015                x if x == libc::SEEK_CUR as u32 => {
1016                    // Get current position from underlying layer
1017                    // This is needed to calculate the new position
1018                    let current = match layer
1019                        .lseek(req, real_inode, real_handle, 0, libc::SEEK_CUR as u32)
1020                        .await
1021                    {
1022                        Ok(r) => r.offset,
1023                        Err(_) => return Err(Error::from_raw_os_error(libc::EINVAL).into()),
1024                    };
1025
1026                    // Check for potential overflow when adding the provided offset
1027                    // This prevents invalid position calculations
1028                    if let Some(new_offset) = current.checked_add(offset) {
1029                        // Ensure the new offset is within valid bounds
1030                        if new_offset > i64::MAX as u64 {
1031                            return Err(Error::from_raw_os_error(libc::EINVAL).into());
1032                        }
1033
1034                        // Actually set the underlying offset to the new value so behavior
1035                        // matches passthrough which uses libc::lseek64 to set the fd offset.
1036                        match layer
1037                            .lseek(
1038                                req,
1039                                real_inode,
1040                                real_handle,
1041                                new_offset,
1042                                libc::SEEK_SET as u32,
1043                            )
1044                            .await
1045                        {
1046                            Ok(_) => Ok(ReplyLSeek { offset: new_offset }),
1047                            Err(_) => Err(Error::from_raw_os_error(libc::EINVAL).into()),
1048                        }
1049                    } else {
1050                        Err(Error::from_raw_os_error(libc::EINVAL).into())
1051                    }
1052                }
1053                // Any other whence value is invalid for directories
1054                _ => Err(Error::from_raw_os_error(libc::EINVAL).into()),
1055            }
1056        } else {
1057            // Keep the original lseek behavior for regular files
1058            // Delegate directly to the underlying layer
1059            let (layer, real_inode, real_handle) = self.find_real_info_from_handle(fh).await?;
1060            layer
1061                .lseek(req, real_inode, real_handle, offset, whence)
1062                .await
1063        }
1064    }
1065
1066    async fn interrupt(&self, _req: Request, _unique: u64) -> Result<()> {
1067        Ok(())
1068    }
1069}
1070
1071#[cfg(test)]
1072mod tests {
1073    use std::{ffi::OsString, path::PathBuf, sync::Arc};
1074
1075    use rfuse3::{MountOptions, raw::Session};
1076    use tokio::signal;
1077    use tracing_subscriber::EnvFilter;
1078
1079    use crate::unionfs::BoxedLayer;
1080    use crate::{
1081        passthrough::{PassthroughArgs, new_passthroughfs_layer, newlogfs::LoggingFileSystem},
1082        unionfs::{OverlayFs, config::Config},
1083    };
1084
1085    #[tokio::test]
1086    #[ignore]
1087    async fn test_a_ovlfs() {
1088        let _ = tracing_subscriber::fmt()
1089            .with_env_filter(EnvFilter::from_default_env().add_directive("trace".parse().unwrap()))
1090            .try_init();
1091
1092        // Set up test environment
1093        let mountpoint = PathBuf::from("/home/luxian/megatest/true_temp");
1094        let lowerdir = vec![PathBuf::from("/home/luxian/github/buck2-rust-third-party")];
1095        let upperdir = PathBuf::from("/home/luxian/upper");
1096
1097        // Create lower layers
1098        let mut lower_layers: Vec<Arc<BoxedLayer>> = Vec::new();
1099        for lower in &lowerdir {
1100            let layer = new_passthroughfs_layer(PassthroughArgs {
1101                root_dir: lower.clone(),
1102                mapping: None::<&str>,
1103            })
1104            .await
1105            .unwrap();
1106            lower_layers.push(Arc::new(layer) as Arc<BoxedLayer>);
1107        }
1108        // Create upper layer
1109        let upper_layer: Arc<BoxedLayer> = Arc::new(
1110            new_passthroughfs_layer(PassthroughArgs {
1111                root_dir: upperdir,
1112                mapping: None::<&str>,
1113            })
1114            .await
1115            .unwrap(),
1116        );
1117        // Create overlayfs
1118        let config = Config {
1119            mountpoint: mountpoint.clone(),
1120            do_import: true,
1121            ..Default::default()
1122        };
1123
1124        let overlayfs = OverlayFs::new(Some(upper_layer), lower_layers, config, 1).unwrap();
1125
1126        let logfs = LoggingFileSystem::new(overlayfs);
1127
1128        let mount_path: OsString = OsString::from(mountpoint);
1129
1130        let uid = unsafe { libc::getuid() };
1131        let gid = unsafe { libc::getgid() };
1132
1133        let not_unprivileged = false;
1134
1135        let mut mount_options = MountOptions::default();
1136        // .allow_other(true)
1137        mount_options.force_readdir_plus(true).uid(uid).gid(gid);
1138
1139        let mut mount_handle: rfuse3::raw::MountHandle = if !not_unprivileged {
1140            Session::new(mount_options)
1141                .mount_with_unprivileged(logfs, mount_path)
1142                .await
1143                .unwrap()
1144        } else {
1145            Session::new(mount_options)
1146                .mount(logfs, mount_path)
1147                .await
1148                .unwrap()
1149        };
1150
1151        let handle = &mut mount_handle;
1152
1153        tokio::select! {
1154            res = handle => res.unwrap(),
1155            _ = signal::ctrl_c() => {
1156                mount_handle.unmount().await.unwrap()
1157            }
1158        }
1159    }
1160}
libfuse_fs/unionfs/async_io.rs

libfuse_fs/unionfs/
async_io.rs
