passthroughfs: support per-file DAX

Guest Fuse FS can enable per-file DAX by adding the "-o dax=inode"
parameter at mount time. This also enable per-file of virtiofs.

Whether the file enables DAX or not is determined by virtiofs. We
use a relatively simple strategy to control which file enables DAX.
Now the strategy is when the file is larger than a value, such
as 2MB, it will enable DAX. Now this value is configured by
Config.dax_file_size.

* If not set dax_file_size, DAX will disable to all files.
* If set dax_file_size=0, DAX will enable all files.
* If set dax_file_size=N, DAX will enable only when the file size is
  greater than or equal to N Bytes.

Signed-off-by: hugh.gxy <[email protected]>

Author: useeSol

src/abi/linux_abi.rs

@@ -182,6 +182,20 @@ const SUBMOUNTS: u32 = 0x800_0000;
 // Filesystem responsible for clearing security.capability xattr and setuid/setgid bits.
 const HANDLE_KILLPRIV_V2: u32 = 0x1000_0000;
 
+// This flag indicates whether the guest kernel enable per-file dax
+const PERFILE_DAX: u32 = 0x4000_0000;
+
+/**
+ *
+ * fuse_attr flags
+ *
+ * upstream kernel use (1 << 0) as FUSE_ATTR_SUBMOUNT,
+ * so FUSE_ATTR_DAX will use (1 << 1)
+ *
+ */
+/// This attribute indicates whether the file supports dax in per-file DAX mode
+pub const FUSE_ATTR_DAX: u32 = 1 << 1;
+
 bitflags! {
     /// A bitfield passed in as a parameter to and returned from the `init` method of the
     /// `FileSystem` trait.
@@ -415,6 +429,12 @@ bitflags! {
         ///  -. create has O_TRUNC and FOPEN_IN_KILL_SUIDGID flag set.
         ///  -. write has WRITE_KILL_PRIV
         const HANDLE_KILLPRIV_V2 = HANDLE_KILLPRIV_V2;
+
+        /// Indicates whether the guest kernel enable per-file dax
+        ///
+        /// If this feature is enabled, filesystem will notify guest kernel whether file
+        /// enable DAX by EntryOut.Attr.flags of inode when lookup
+        const PERFILE_DAX = PERFILE_DAX;
     }
 }
 

src/api/filesystem/mod.rs

@@ -51,6 +51,9 @@ pub struct Entry {
     /// If this value is not correct, incorrect data will be returned.
     pub attr: libc::stat64,
 
+    /// Flags for 'fuse::Attr.flags'
+    pub attr_flags: u32,
+
     /// How long the values in `attr` should be considered valid. If the attributes of the `Entry`
     /// are only modified by the FUSE client, then this should be set to a very large value.
     pub attr_timeout: Duration,
@@ -70,7 +73,7 @@ impl From<Entry> for fuse::EntryOut {
             attr_valid: entry.attr_timeout.as_secs(),
             entry_valid_nsec: entry.entry_timeout.subsec_nanos(),
             attr_valid_nsec: entry.attr_timeout.subsec_nanos(),
-            attr: entry.attr.into(),
+            attr: fuse::Attr::with_flags(entry.attr, entry.attr_flags),
         }
     }
 }
@@ -429,6 +432,7 @@ mod tests {
             inode: 1,
             generation: 2,
             attr: attr.into(),
+            attr_flags: 0,
             attr_timeout: Default::default(),
             entry_timeout: Default::default(),
         };

src/api/pseudo_fs.rs

@@ -275,6 +275,7 @@ impl<S: BitmapSlice> PseudoFs<S> {
             inode: ino,
             generation: 0,
             attr: attr.into(),
+            attr_flags: 0,
             attr_timeout: Duration::from_secs(PSEUDOFS_DEFAULT_ATTR_TIMEOUT),
             entry_timeout: Duration::from_secs(PSEUDOFS_DEFAULT_ENTRY_TIMEOUT),
         }

src/api/server/sync_io.rs

@@ -607,7 +607,8 @@ impl<F: FileSystem<S> + Sync, D: AsyncDrive, S: BitmapSlice> Server<F, D, S> {
             | FsOptions::ASYNC_DIO
             | FsOptions::HAS_IOCTL_DIR
             | FsOptions::MAX_PAGES
-            | FsOptions::EXPLICIT_INVAL_DATA;
+            | FsOptions::EXPLICIT_INVAL_DATA
+            | FsOptions::PERFILE_DAX;
 
         let capable = FsOptions::from_bits_truncate(flags);
 

src/api/vfs/mod.rs

@@ -214,7 +214,8 @@ impl Default for VfsOptions {
                 | FsOptions::DO_READDIRPLUS
                 | FsOptions::READDIRPLUS_AUTO
                 | FsOptions::ZERO_MESSAGE_OPENDIR
-                | FsOptions::HANDLE_KILLPRIV_V2,
+                | FsOptions::HANDLE_KILLPRIV_V2
+                | FsOptions::PERFILE_DAX,
         }
     }
 }
@@ -513,6 +514,7 @@ mod tests {
                 inode: 0,
                 generation: 0,
                 attr: Attr::default().into(),
+                attr_flags: 0,
                 attr_timeout: Duration::new(0, 0),
                 entry_timeout: Duration::new(0, 0),
             })
@@ -528,6 +530,7 @@ mod tests {
                 inode: 1,
                 generation: 0,
                 attr: Attr::default().into(),
+                attr_flags: 0,
                 attr_timeout: Duration::new(0, 0),
                 entry_timeout: Duration::new(0, 0),
             })
@@ -822,6 +825,7 @@ mod tests {
                     inode: 1,
                     generation: 0,
                     attr: Attr::default().into(),
+                    attr_flags: 0,
                     attr_timeout: Duration::new(0, 0),
                     entry_timeout: Duration::new(0, 0),
                 },
@@ -842,6 +846,7 @@ mod tests {
                     inode: 1,
                     generation: 0,
                     attr: Attr::default().into(),
+                    attr_flags: 0,
                     attr_timeout: Duration::new(0, 0),
                     entry_timeout: Duration::new(0, 0),
                 },
@@ -870,7 +875,8 @@ mod tests {
             | FsOptions::DO_READDIRPLUS
             | FsOptions::READDIRPLUS_AUTO
             | FsOptions::ZERO_MESSAGE_OPENDIR
-            | FsOptions::HANDLE_KILLPRIV_V2;
+            | FsOptions::HANDLE_KILLPRIV_V2
+            | FsOptions::PERFILE_DAX;
         let opts = vfs.opts.load();
 
         assert_eq!(opts.no_open, true);

src/passthrough/async_io.rs

@@ -314,6 +314,17 @@ impl<D: AsyncDrive + Sync, S: BitmapSlice + Send + Sync> AsyncFileSystem<D, S>
             })
             .await?;
 
+        let mut attr_flags: u32 = 0;
+        if let Some(dax_file_size) = self.cfg.dax_file_size {
+            // st.stat.st_size is i64
+            if self.perfile_dax.load(Ordering::Relaxed)
+                && st.stat.st_size >= 0x0
+                && st.stat.st_size as u64 >= dax_file_size
+            {
+                attr_flags |= fuse::FUSE_ATTR_DAX;
+            }
+        }
+
         let mut found = None;
         'search: loop {
             match self.inode_map.get_alt(&ids_altkey, handle_altkey.as_ref()) {
@@ -394,6 +405,7 @@ impl<D: AsyncDrive + Sync, S: BitmapSlice + Send + Sync> AsyncFileSystem<D, S>
             inode,
             generation: 0,
             attr: st.get_stat(),
+            attr_flags,
             attr_timeout: self.cfg.attr_timeout,
             entry_timeout: self.cfg.entry_timeout,
         })

src/passthrough/mod.rs

@@ -455,6 +455,13 @@ pub struct Config {
     /// Control whether readdir/readdirplus requests return zero dirent to client, as if the
     /// directory is empty even if it has children.
     pub no_readdir: bool,
+
+    /// What size file supports dax
+    /// * If dax_file_size == None, DAX will disable to all files.
+    /// * If dax_file_size == 0, DAX will enable all files.
+    /// * If dax_file_size == N, DAX will enable only when the file size is greater than or equal
+    /// to N Bytes.
+    pub dax_file_size: Option<u64>,
 }
 
 impl Default for Config {
@@ -472,6 +479,7 @@ impl Default for Config {
             killpriv_v2: false,
             inode_file_handles: false,
             no_readdir: false,
+            dax_file_size: None,
         }
     }
 }
@@ -521,6 +529,10 @@ pub struct PassthroughFs<D: AsyncDrive = AsyncDriver, S: BitmapSlice + Send + Sy
     // Whether no_readdir is enabled.
     no_readdir: AtomicBool,
 
+    // Whether per-file DAX feature is enabled.
+    // Init from guest kernel Init cmd of fuse fs.
+    perfile_dax: AtomicBool,
+
     cfg: Config,
 
     phantom: PhantomData<D>,
@@ -554,6 +566,7 @@ impl<D: AsyncDrive, S: BitmapSlice + Send + Sync> PassthroughFs<D, S> {
             no_opendir: AtomicBool::new(false),
             killpriv_v2: AtomicBool::new(false),
             no_readdir: AtomicBool::new(cfg.no_readdir),
+            perfile_dax: AtomicBool::new(false),
             cfg,
 
             phantom: PhantomData,
@@ -756,6 +769,18 @@ impl<D: AsyncDrive, S: BitmapSlice + Send + Sync> PassthroughFs<D, S> {
             |fd, flags| Self::open_proc_file(&self.proc, fd, flags),
         )?;
 
+        // Whether to enable file DAX according to the value of dax_file_size
+        let mut attr_flags: u32 = 0;
+        if let Some(dax_file_size) = self.cfg.dax_file_size {
+            // st.stat.st_size is i64
+            if self.perfile_dax.load(Ordering::Relaxed)
+                && st.stat.st_size >= 0x0
+                && st.stat.st_size as u64 >= dax_file_size
+            {
+                attr_flags |= fuse::FUSE_ATTR_DAX;
+            }
+        }
+
         let mut found = None;
         'search: loop {
             match self.inode_map.get_alt(&ids_altkey, handle_altkey.as_ref()) {
@@ -836,6 +861,7 @@ impl<D: AsyncDrive, S: BitmapSlice + Send + Sync> PassthroughFs<D, S> {
             inode,
             generation: 0,
             attr: st.get_stat(),
+            attr_flags,
             attr_timeout: self.cfg.attr_timeout,
             entry_timeout: self.cfg.entry_timeout,
         })

src/passthrough/sync_io.rs

@@ -333,6 +333,11 @@ impl<D: AsyncDrive, S: BitmapSlice + Send + Sync> FileSystem<S> for PassthroughF
             self.killpriv_v2.store(true, Ordering::Relaxed);
         }
 
+        if capable.contains(FsOptions::PERFILE_DAX) {
+            opts |= FsOptions::PERFILE_DAX;
+            self.perfile_dax.store(true, Ordering::Relaxed);
+        }
+
         Ok(opts)
     }