[Gluster-devel] Stale state->fd->inode and race condition with fd_destroy()

[Emmanuel Dreyfus]

Hi

I get a reprodcutbile crash of glusterfsd, running 3.2.1 code. I get it
by running multiple tar -xzf on a client, and after a while, a
glusterfsd on a brick crashes:

Program terminated with signal 11, Segmentation fault.
#0  0xba0d652e in server_rchecksum_cbk (frame=0xbad007d0,
    cookie=0xbaf00300, this=0xba810000, op_ret=-1, op_errno=9,
    weak_checksum=0, strong_checksum=0xb91ffc74 "") at
    server3_1-fops.c:1305

Here is the offending code

        if (op_ret == -1)
                gf_log (this->name, GF_LOG_INFO,
                        "%"PRId64": RCHECKSUM %"PRId64" (%"PRId64") ==>
%"PRId32" (%s)",
                        frame->root->unique, state->resolve.fd_no,
                        state->fd ? state->fd->inode->ino : 0, op_ret,
                        strerror (op_errno));

The problem is state->fd->inode value:

(gdb) print *((server_state_t *)frame->root->state)->fd 
$7 = {pid = 2610, flags = 2, refcount = 2, inode_list =
        {next = 0xb9801088, prev = 0xb9801088}, inode = 0xaaaaaaaa,
        lock = {pts_magic = 3735879687, pts_spin = 0 '\0', pts_flags =
        0}, _ctx = 0xbb96b080, xl_count = 8}

inode = 0xaaaaaaaa is set in fd_destroy() to denote a stale object (It
is less fun than using 0xdeadbeef :-)

That suggests a race condition where a thread uses a fd that another
thread destroyed. Of course, the value could be checked at the beginning
of server_rchecksum_cbk(), but I suspect the problem is more widespread
that this. There are many other places in server3_1-fops.c where
state->fd->inode->ino is used.

And should the value be checked at the beginning of
server_rchecksum_cbk() and its friends, or in any gf_log() call, like
this:
        if (op_ret == -1)
                gf_log (this->name, GF_LOG_INFO,
                        "%"PRId64": RCHECKSUM %"PRId64" (%"PRId64") "
                        "==> %"PRId32" (%s)",
                        frame->root->unique, state->resolve.fd_no,
                        state->fd && (state->fd->inode != 0xaaaaaaaa) ?
                        state->fd->inode->ino : 0, op_ret,
                        strerror (op_errno));

FWIW this is a 2x2 replicated and distributed setup.

-- 
Emmanuel Dreyfus
http://hcpnet.free.fr/pubz
manu at netbsd.org