On Fri, 16 Jan 2009 14:58:32 -0500
Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> On Thu, 1 Jan 2009 19:36:51 -0500
> "J. Bruce Fields" <bfields@xxxxxxxxxxxxxx> wrote:
>
> > I've got a slight fear that the problem Marc tripped across won't be the
> > last that's caused by this hack of faking up a struct file with only
> > some fields initialized. We may as well just do an open, just as we do
> > with reads and writes in the v2/v3 case, and get a proper struct file.
> >
> > Signed-off-by: J. Bruce Fields <bfields@xxxxxxxxxxxxxx>
> > ---
>
> I was working on backporting this patch, but it seems to be causing a
> reliable oops:
>
> ----------------[snip]------------------
>
> BUG: unable to handle kernel NULL pointer dereference at (null)
> IP: [<ffffffff811928d9>] kref_get+0xc/0x2f
> PGD 1685f067 PUD 16860067 PMD 0
> Oops: 0000 [#1] SMP DEBUG_PAGEALLOC
> last sysfs file: /sys/fs/gfs2/jtltest:v1/lock_module/recover_status
> CPU 0
> Modules linked in: nfsd lockd nfs_acl exportfs ipt_MASQUERADE iptable_nat nf_nat bridge stp llc lock_dlm gfs2 dlm configfs autofs4 rpcsec_gss_krb5 auth_rpcgss des_generic sunrpc ipv6 dm_multipath uinput i2c_piix4 8139cp pcspkr i2c_core 8139too mii ata_generic pata_acpi [last unloaded: freq_table]
> Pid: 4088, comm: nfsd Tainted: G W 2.6.29-0.35.rc1.git4.fc11.x86_64 #1
> RIP: 0010:[<ffffffff811928d9>] [<ffffffff811928d9>] kref_get+0xc/0x2f
> RSP: 0018:ffff88000dadfc40 EFLAGS: 00010282
> RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000009
> RDX: 0000000000000000 RSI: ffff88000fec42e0 RDI: 0000000000000000
> RBP: ffff88000dadfc50 R08: ffff8800154b03d8 R09: ffffffff810dfc9f
> R10: ffff88000f496d80 R11: ffff88000dadfc20 R12: ffff88000fec42e0
> R13: ffff88000dadfcc0 R14: ffff88000dadfcc0 R15: ffffffffa04938d0
> FS: 0000000000000000(0000) GS:ffffffff81934000(0000) knlGS:0000000000000000
> CS: 0010 DS: 0018 ES: 0018 CR0: 000000008005003b
> CR2: 0000000000000000 CR3: 000000001685e000 CR4: 00000000000006e0
> DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
> Process nfsd (pid: 4088, threadinfo ffff88000dade000, task ffff88000dad2350)
> Stack:
> ffff88000fec42a8 0000000000000000 ffff88000dadfc80 ffffffffa04628c3
> ffff88000dadfc70 0000000000000000 ffff88000fec42a8 000000001a270000
> ffff88000dadfde0 ffffffffa0464db0 0000000000000004 ffff88000f58b360
> Call Trace:
> [<ffffffffa04628c3>] nfs4_set_lock_denied+0x2c/0xa9 [nfsd]
> [<ffffffffa0464db0>] nfsd4_lockt+0x323/0x38d [nfsd]
> [<ffffffffa045b6a4>] nfsd4_proc_compound+0x1d0/0x30c [nfsd]
> [<ffffffffa044d256>] nfsd_dispatch+0xe9/0x1ca [nfsd]
> [<ffffffffa01d4327>] svc_process+0x3fc/0x63f [sunrpc]
> [<ffffffff813828a2>] ? down_read+0x77/0x7f
> [<ffffffffa044d862>] nfsd+0x149/0x1a9 [nfsd]
> [<ffffffffa044d719>] ? nfsd+0x0/0x1a9 [nfsd]
> [<ffffffffa044d719>] ? nfsd+0x0/0x1a9 [nfsd]
> [<ffffffff8105e954>] kthread+0x49/0x76
> [<ffffffff8101262a>] child_rip+0xa/0x20
> [<ffffffff8138397a>] ? _spin_unlock_irq+0x2b/0x37
> [<ffffffff81011f3e>] ? restore_args+0x0/0x30
> [<ffffffff8105e8e6>] ? kthreadd+0x176/0x19b
> [<ffffffff8105e90b>] ? kthread+0x0/0x76
> [<ffffffff81012620>] ? child_rip+0x0/0x20
> Code: ff f0 ff 0b 0f 94 c0 31 d2 84 c0 74 0b 48 89 df 41 ff d4 ba 01 00 00 00 5b 89 d0 41 5c c9 c3 55 48 89 e5 53 48 89 fb 48 83 ec 08 <8b> 07 85 c0 75 13 31 d2 be 2b 00 00 00 48 c7 c7 a9 5d 4c 81 e8
> RIP [<ffffffff811928d9>] kref_get+0xc/0x2f
> RSP <ffff88000dadfc40>
> CR2: 0000000000000000
> ---[ end trace 4eaa2a86a8e2da24 ]---
>
> ----------------[snip]------------------
>
> Basically I have a GFS2 filesystem and have one process on the server
> taking a lock on a file. When a NFSv4 client tries to do a GETLK
> against the same file I get the above stack trace.
>
> The problem is that nfsd4_lockt does this:
>
> lockt->lt_stateowner = find_lockstateowner_str(inode,
> &lockt->lt_clientid, &lockt->lt_owner);
> if (lockt->lt_stateowner)
> file_lock.fl_owner = (fl_owner_t)lockt->lt_stateowner;
>
> ...but it's not finding anything in the lockowner hash since no NFSv4
> client is holding a lock. So fl_owner ends up being NULL, but fl_type
> is F_UNLCK. We then call into nfs4_set_lock_denied(), which does this:
>
> if (fl->fl_lmops == &nfsd_posix_mng_ops) {
> sop = (struct nfs4_stateowner *) fl->fl_owner;
> hval = lockownerid_hashval(sop->so_id);
> kref_get(&sop->so_ref);
>
> ...and that then oops when dereferencing sop->so_ref.
>
> So in any case, I see the cause here but the fix is not immediately
> clear to me. Does nfs4_set_lock_denied() need to take into account the
> possibility of a NULL fl_owner, or should we be allocating a new
> lockstateowner in this case?
>
A little bit more info (and this is a little strange). Both with and
without the patch, nfsd4_lockt does this:
file_lock.fl_lmops = &nfsd_posix_mng_ops;
...but pre-patch, the fl_lmops ends up getting nulled out before calling
into nfs4_set_lock_denied(), but after the patch it remains set to
nfsd_posix_mng_ops. This accounts for why we didn't see this before the
patch was applied.
--
Jeff Layton <jlayton@xxxxxxxxxx>
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
