Created attachment 166011 [details]
Initialize the inode's i_flag to 0 during allocation

fsck.ext3 on a cleanly unmounted EXT3 filesystem that went through heavy file creation (OpenOffice build, about 80000 new files) gives a dozen or so of these errors during "Pass 1: Checking inodes, blocks, and sizes":

Inode 157938 has INDEX_FL flag set but is not a directory.
Clear HTree index?

This error is benign to our ext2fs driver, as it requires the inode's mode to be directory for this flag to have any effect. However it's not benign to other EXT3 implementations - it breaks at least fsck.ext3 itself if the "-n" option to it is used or "no" is answered to that question, since it misinterprets the inode as a directory, giving further false errors because that pseudo-directory will appear corrupt.

Adding this hack:

  if (!S_ISDIR(ip->i_mode) && (ip->i_flag & IN_E4INDEX) != 0)
    panic("non-directory has index!?\n");

to ext2_i2ei() to catch wrong use of this flag just before the inode is written to disk, produces this revealing stack trace:

KDB: stack backtrace:
db_trace_self_wrapper()
vpanic()
panic()
ext2_i2ei()
ext2_update()
ext2_makeinode()
ext2_create()
VOP_CREATE_APV()
vn_open_cred()
kern_openat()
amd64_syscall()
Xfast_syscall()

Reading through those functions shows ext2_makeinode() calls ext2_valloc() which apparently reuses inodes from disk without initializing their i_flag field, hence if a previously deleted directory's inode is reused for a file, the IN_E4INDEX flag from it will still be set, and wrongly written to the file's inode!

I am attaching a patch that initializes i_flag to 0. With it, fsck.ext3 reports a clean scan after the same test.