Bug 1175711

We need to verify if the 'struct dirent'->d_type is retrieved correctly over the fuse filesystem. In case it is not, this would be a bug in fuse.

I did check that (using following script) on latest master branch code. It does fill that.

#!/usr/bin/env python

# Return status indicates if d_type returned

import ctypes
import sys

(DT_UNKNOWN, DT_DIR,) = (0, 4,)

class dirent(ctypes.Structure):
  _fields_ = [
    ("d_ino", ctypes.c_long),
    ("d_off", ctypes.c_long),
    ("d_reclen", ctypes.c_ushort),
    ("d_type", ctypes.c_ubyte),
    ("d_name", ctypes.c_char*256)]

direntp = ctypes.POINTER(dirent)

libc = ctypes.cdll.LoadLibrary("libc.so.6")
libc.readdir.restype = direntp

dirp = libc.opendir(".")
if dirp:
  ep = libc.readdir(dirp)
else:
  sys.exit(1)

print ep.contents.d_type

I was wrong. The above script failed to detect it because d_type is always set correctly for "." and ".." entries. GlusterFS correctly propagates d_type from posix xlator up the stack till FUSE.

It turns out that XFS does't fill correct d_type until recently (Linux>=3.15 and xfsprogs>=3.2.0). If one formats his/her filesystem with XFS's newer version 5 on-disk format, d_type is rightly set.

Example: mkfs.xfs -m crc=1 /srv/disk1

However, GlusterFS can support filling the right d_type in readdirp() responses even if XFS doesn't using the pre-fetched stat information.

REVIEW: http://review.gluster.org/14095 (posix: Set correct d_type for readdirp() calls) posted (#1) for review on master by Prashanth Pai (ppai)

Created a nested fs tree of depth = 4 on glusterfs mountpoint.

In the below example: ls command from coreutils is capable of avoiding additional lstat() if it finds d_type to be set correctly.

BEFORE http://review.gluster.org/14095:

root# strace -fc -e getdents,lstat ls -fR /mnt/gluster-object/gsmetadata >> /dev/null
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 55.95    0.672307          30     22226           getdents
 44.05    0.529388          24     22224           lstat
------ ----------- ----------- --------- --------- ----------------
100.00    1.201695                 44450           total


AFTER http://review.gluster.org/14095:
root# strace -fc -e getdents,lstat ls -fR /mnt/gluster-object/gsmetadata >> /dev/null
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
100.00    0.595680          27     22226           getdents
------ ----------- ----------- --------- --------- ----------------
100.00    0.595680                 22226           total

REVIEW: http://review.gluster.org/14095 (posix: Set correct d_type for readdirp() calls) posted (#2) for review on master by Prashanth Pai (ppai)

COMMIT: http://review.gluster.org/14095 committed in master by Jeff Darcy (jdarcy) 
------
commit 77def44d497d090ef3f393b6d9403c1a29dcf993
Author: Prashanth Pai <ppai>
Date:   Wed Apr 27 13:37:07 2016 +0530

    posix: Set correct d_type for readdirp() calls
    
    dirent.d_type can contain the type of the directory entry. The 'd_type'
    struct member in dirent is present in Linux and many BSD flavours.
    However, filling d_type with correct value requires support from the
    underlying filesystem. If not, d_type is set to DT_UNKNOWN. XFS added
    support for d_type as part of their newer version 5 on-disk format.
    However, this requires Linux >= 3.15, xfsprogs >= 3.2.0 and the bricks
    to be formatted using the new format.
    
    This patch enables posix xlator to set d_type to the right value even
    when the underlying filesystem does not support it. d_type can be set
    using information previously fetched by stat() on the dir entry.
    This will aid FUSE applications to leverage d_type to avoid the expense
    of calling lstat() if further actions depend on the type of the file.
    
    Refer `man 3 readdir` and `man 2 getdents`
    
    BUG: 1175711
    Change-Id: Ic5a262fe4c64122726b4fae2d1bea375c559ca04
    Signed-off-by: Prashanth Pai <ppai>
    Reviewed-on: http://review.gluster.org/14095
    Smoke: Gluster Build System <jenkins.com>
    NetBSD-regression: NetBSD Build System <jenkins.org>
    CentOS-regression: Gluster Build System <jenkins.com>
    Reviewed-by: Jeff Darcy <jdarcy>