quoted

quoted

last_offset=-1, last_fpos=-1, f_pos=4024

-1 means we hit the bug in the first iteration in the insde while in
ext3_readdir().

I've checked how ext3_readdir() works and how f_pos, f_version and i_version
get initialized and modified. Now I'm lost. I really can't see how f_pos got
corrupted. :(

  Hum, it looks really curious. So f_pos has been 4024 when we entered
ext3_readdir()?

dunno. but what else can be

quoted

Do you know what it was when we last left ext3_readdir()
for that filp? You can store that value in some debug entry added to struct
file... Also any chance we ever hit:
                                if (version != filp->f_version)
                                        goto revalidate;
I don't think it can ever happen since we hold i_mutex and
generic_file_llseek() takes i_mutex as well. But better be sure.

Yesterday I've added more debug aids, which convers all the above information
mentioned. Actually the code tracks all the places that change f_pos, and
I think only lseek() and readdir() can change it.

Now I'm waiting for the bug to happen again, can be several days...

The bug was triggered again:

EXT3-fs error (device sda7): ext3_readdir: bad entry in directory #9372013: rec_len is smaller than minimal - offset=4028, inode=0, rec_len=0, name_len=0

And I've confirmed f_pos=4028 when we entered ext3_readdir(), while it should be 4096.

I wrote a simple ring buffer to track operations on log dirs, and from the
ring buffer, we can see that there were no lseek, unlink, rename, etc.

This is correct:

dir=9372013, seq=1549, spot=readdir_1, f_pos=0, f_pos_delta=0
dir=9372013, seq=1550, spot=readdir_3, f_pos=0, f_pos_delta=0
dir=9372013, seq=1551, spot=readdir_5, f_pos=12, f_pos_delta=12
dir=9372013, seq=1552, spot=readdir_5, f_pos=24, f_pos_delta=12
...
dir=9372013, seq=1595, spot=readdir_5, f_pos=1488, f_pos_delta=28
dir=9372013, seq=1596, spot=readdir_5, f_pos=1516, f_pos_delta=28
dir=9372013, seq=1597, spot=readdir_5, f_pos=1556, f_pos_delta=40
dir=9372013, seq=1598, spot=readdir_5, f_pos=1584, f_pos_delta=28
...
dir=9372013, seq=1627, spot=readdir_5, f_pos=2392, f_pos_delta=28
dir=9372013, seq=1628, spot=readdir_5, f_pos=4096, f_pos_delta=1704
dir=9372013, seq=1629, spot=readdir_1, f_pos=4096, f_pos_delta=0

(readir_1 is the entry of readdir(), and readdir_3 is when we enter (f_version != i_version),
and readdir_5 is we iterate the dir block)

Then f_pos went wrong suddenly:

dir=9372013, seq=1676, spot=readdir_5, f_pos=1488, f_pos_delta=28
dir=9372013, seq=1677, spot=readdir_5, f_pos=1516, f_pos_delta=28
dir=9372013, seq=1678, spot=readdir_5, f_pos=1556, f_pos_delta=40
dir=9372013, seq=1679, spot=readdir_5, f_pos=1516, f_pos_delta=28   <-- !!!!!!!!
dir=9372013, seq=1680, spot=readdir_5, f_pos=1540, f_pos_delta=24
...
dir=9372013, seq=1708, spot=readdir_5, f_pos=2324, f_pos_delta=28
dir=9372013, seq=1709, spot=readdir_5, f_pos=4028, f_pos_delta=1704
dir=9372013, seq=1710, spot=readdir_1, f_pos=4028, f_pos_delta=0

This is odd...

While f_pos was wrong, offset is always correct, and this is not some
single-bit error in memory, so someone else changed f_pos? but we were
holding i_mutex, and we see nothing else except readdir in the ring
buffer...