vfs/zfs: remove per-file lock enforcement for ZFS read/write ops

OSv VFS layer enforces a per-file lock for file operations, including read and
write operations. This approach prevents ZFS from doing parallel reads on the
same file, or even parallel writes into different regions from the same file.
ZFS allows that parallelism by using r/w range locks from its own internal
nodes, mapped 1:1 to VFS vnodes. It properly satisfies POSIX requeriments.

Currently, every write/read to a specific file will always be serialized.
Workloads where many threads concurrently read/write from/into the same file
would perform terribly bad as compared to FreeBSD, for example.
FreeBSD also uses a r/w range lock within the VFS layer to protect concurrent
read/write ops, however, that implies double range locking. It's not necessary,
and it's also redundant.

Specifically to ZFS, this bottleneck - which prevents parallel ZFS reads/writes
on the same file from happening - could be pulverized by discarding the vnode
locks surrounding VOP_READ and VOP_WRITE within vfs_file::read and vfs_file::
write, respectively. For read/write ops from the virtual file systems, vnode
locking could be moved over their respective functions. Unlike ZFS, the virtual
file systems don't deal with per-file locks on their own. From there on, ZFS
r/w range locks would be working effectively. Before that, there was *no* r/w
parallelism when working on the same file.

vfs_file::get_arcbuf, which is entirely intended for ZFS, could also discard
the vnode lock. It calls zfs_arc that also uses the r/w range lock approach, so
properly protecting the underlying file. writeback() from the pagecache doesn't
need to take the vnode lock either, given that it calls zfs_write (protected by
r/w range lock).

Changes involved:
* For zfs_read, use zfs internal node data to avoid reading vnode.
* For zfs_write, when file has grown in size, take vnode lock for updating
vnode size.
* For virtual file systems, surround their read()/write() ops with vnode lock.

Signed-off-by: Raphael S. Carvalho <[email protected]>
Signed-off-by: Pekka Enberg <[email protected]>

bsd/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_vnops.c

@@ -565,15 +565,16 @@ zfs_read(vnode_t *vp, struct file* fp, uio_t *uio, int ioflag)
 	xuio_t		*xuio = NULL;
 #endif
 
-	// Return EISDIR when reading from a directory, as Linux does.
-	if (vp->v_type == VDIR) {
-	    return EISDIR;
-	}
-
 	ZFS_ENTER(zfsvfs);
 	ZFS_VERIFY_ZP(zp);
 	os = zfsvfs->z_os;
 
+	// Return EISDIR when reading from a directory, as Linux does.
+	if (S_ISDIR(zp->z_mode)) {
+		ZFS_EXIT(zfsvfs);
+		return EISDIR;
+	}
+
 	if (zp->z_pflags & ZFS_AV_QUARANTINED) {
 		ZFS_EXIT(zfsvfs);
 		return (EACCES);
@@ -958,8 +959,11 @@ zfs_write(vnode_t *vp, uio_t *uio, int ioflag)
 		 */
 		nbytes = MIN(n, max_blksz - P2PHASE(woff, max_blksz));
 
-		if (woff + nbytes > zp->z_size)
+		if (woff + nbytes > zp->z_size) {
+			vn_lock(vp);
 			vnode_pager_setsize(vp, woff + nbytes);
+			vn_unlock(vp);
+		}
 
 		if (abuf == NULL) {
 			tx_bytes = uio->uio_resid;

core/pagecache.cc

@@ -204,17 +204,12 @@ class cached_page_write : public cached_page {
     }
     int writeback()
     {
-        int error;
         struct iovec iov {_page, mmu::page_size};
         struct uio uio {&iov, 1, _key.offset, mmu::page_size, UIO_WRITE};
 
         _dirty = false;
 
-        vn_lock(_vp);
-        error = VOP_WRITE(_vp, &uio, 0);
-        vn_unlock(_vp);
-
-        return error;
+        return VOP_WRITE(_vp, &uio, 0);
     }
     void* release() { // called to demote a page from cache page to anonymous
         assert(boost::get<std::nullptr_t>(_ptes) == nullptr);

fs/devfs/devfs_vnops.cc

@@ -104,12 +104,14 @@ devfs_close(struct vnode *vp, struct file *fp)
 static int
 devfs_read(struct vnode *vp, struct file *fp, struct uio *uio, int ioflags)
 {
+	SCOPE_LOCK(vp->v_lock);
 	return device_read((device*)vp->v_data, uio, ioflags);
 }
 
 static int
 devfs_write(struct vnode *vp, struct uio *uio, int ioflags)
 {
+	SCOPE_LOCK(vp->v_lock);
 	return device_write((device*)vp->v_data, uio, ioflags);
 }
 

fs/procfs/procfs_vnops.cc

@@ -147,6 +147,7 @@ static int
 procfs_read(vnode* vp, file *fp, uio* uio, int ioflags)
 {
     auto* data = static_cast<string*>(fp->f_data);
+    SCOPE_LOCK(vp->v_lock);
 
     if (vp->v_type == VDIR)
         return EISDIR;

fs/ramfs/ramfs_vnops.cc

@@ -299,6 +299,7 @@ ramfs_read(struct vnode *vp, struct file *fp, struct uio *uio, int ioflag)
 {
 	struct ramfs_node *np = (ramfs_node*)vp->v_data;
 	size_t len;
+	SCOPE_LOCK(vp->v_lock);
 
 	if (vp->v_type == VDIR)
 		return EISDIR;
@@ -325,6 +326,7 @@ static int
 ramfs_write(struct vnode *vp, struct uio *uio, int ioflag)
 {
 	struct ramfs_node *np = (ramfs_node*)vp->v_data;
+	SCOPE_LOCK(vp->v_lock);
 
 	if (vp->v_type == VDIR)
 		return EISDIR;

fs/vfs/vfs_fops.cc

@@ -47,7 +47,6 @@ int vfs_file::read(struct uio *uio, int flags)
 
 	bytes = uio->uio_resid;
 
-	vn_lock(vp);
 	if ((flags & FOF_OFFSET) == 0)
 		uio->uio_offset = fp->f_offset;
 
@@ -57,7 +56,6 @@ int vfs_file::read(struct uio *uio, int flags)
 		if ((flags & FOF_OFFSET) == 0)
 			fp->f_offset += count;
 	}
-	vn_unlock(vp);
 
 	return error;
 }
@@ -74,8 +72,6 @@ int vfs_file::write(struct uio *uio, int flags)
 
 	bytes = uio->uio_resid;
 
-	vn_lock(vp);
-
 	if (fp->f_flags & O_APPEND)
 		ioflags |= IO_APPEND;
 	if (fp->f_flags & (O_DSYNC|O_SYNC))
@@ -91,7 +87,6 @@ int vfs_file::write(struct uio *uio, int flags)
 			fp->f_offset += count;
 	}
 
-	vn_unlock(vp);
 	return error;
 }
 
@@ -171,9 +166,7 @@ int vfs_file::get_arcbuf(void* key, off_t offset)
     data.uio_resid = mmu::page_size;
     data.uio_rw = UIO_READ;
 
-    vn_lock(vp);
     assert(VOP_CACHE(vp, this, &data) == 0);
-    vn_unlock(vp);
 
     return (data.uio_resid != 0) ? -1 : 0;
 }