On Mon, Nov 07 2011 at 10:36am -0500,
Mike Snitzer <snitzer@xxxxxxxxxx> wrote:
> On Mon, Nov 07 2011 at 6:30am -0500,
> Jun'ichi Nomura <j-nomura@xxxxxxxxxxxxx> wrote:
>
> > On 11/04/11 18:19, Heiko Carstens wrote:
> > > It's the s390 only zfcp device driver.
> > >
> > > FWIW, yet another use-after-free crash, this time however in multipath_end_io:
> > >
> > > [96875.870593] Unable to handle kernel pointer dereference at virtual kernel address 6b6b6b6b6b6b6000
> > > [96875.870602] Oops: 0038 [#1]
> > > [96875.870674] PREEMPT SMP DEBUG_PAGEALLOC
> > > [96875.870683] Modules linked in: dm_round_robin sunrpc ipv6 qeth_l2 binfmt_misc dm_multipath scsi_dh dm_mod qeth ccwgroup [la\
> > > st unloaded: scsi_wait_scan]
> > > [96875.870722] CPU: 2 Tainted: G W 3.0.7-50.x.20111024-s390xdefault #1
> > > [96875.870728] Process udevd (pid: 36697, task: 0000000072c8a3a8, ksp: 0000000057c43868)
> > > [96875.870732] Krnl PSW : 0704200180000000 000003e001347138 (multipath_end_io+0x50/0x140 [dm_multipath])
> > > [96875.870746] R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:0 CC:2 PM:0 EA:3
> > > [96875.870751] Krnl GPRS: 0000000000000000 000003e000000000 6b6b6b6b6b6b6b6b 00000000717ab940
> > > [96875.870755] 0000000000000000 00000000717abab0 0000000000000002 0700000000000008
> > > [96875.870759] 0000000000000002 0000000000000000 0000000058dd37a8 000000006f845478
> > > [96875.870764] 000003e0012e1000 000000005613d1f0 000000007a737bf0 000000007a737ba0
> > > [96875.870768] Krnl Code: 000003e00134712a: b90200dd ltgr %r13,%r13
> > > [96875.870793] 000003e00134712e: a7840017 brc 8,3e00134715c
> > > [96875.870800] 000003e001347132: e320d0100004 lg %r2,16(%r13)
> > > [96875.870809] >000003e001347138: e31020180004 lg %r1,24(%r2)
> > > [96875.870818] 000003e00134713e: e31010580004 lg %r1,88(%r1)
> > > [96875.870827] 000003e001347144: b9020011 ltgr %r1,%r1
> > > [96875.870835] 000003e001347148: a784000a brc 8,3e00134715c
> > > [96875.870841] 000003e00134714c: 41202018 la %r2,24(%r2)
> > > [96875.870889] Call Trace:
> > > [96875.870892] ([<0700000000000008>] 0x700000000000008)
> > > [96875.870897] [<000003e0012e3662>] dm_softirq_done+0x9a/0x140 [dm_mod]
> > > [96875.870915] [<000000000040d29c>] blk_done_softirq+0xd4/0xf0
> > > [96875.870925] [<00000000001587c2>] __do_softirq+0xda/0x398
> > > [96875.870932] [<000000000010f47e>] do_softirq+0xe2/0xe8
> > > [96875.870940] [<0000000000158e2c>] irq_exit+0xc8/0xcc
> > > [96875.870945] [<00000000004ceb48>] do_IRQ+0x910/0x1bfc
> > > [96875.870953] [<000000000061a164>] io_return+0x0/0x16
> > > [96875.870961] [<000000000019c84e>] lock_acquire+0xd2/0x204
> > > [96875.870969] ([<000000000019c836>] lock_acquire+0xba/0x204)
> > > [96875.870974] [<0000000000615f8e>] mutex_lock_killable_nested+0x92/0x520
> > > [96875.870983] [<0000000000292796>] vfs_readdir+0x8a/0xe4
> > > [96875.870992] [<00000000002928e0>] SyS_getdents+0x60/0xe8
> > > [96875.870999] [<0000000000619af2>] sysc_noemu+0x16/0x1c
> > > [96875.871024] [<000003fffd1ec83e>] 0x3fffd1ec83e
> > > [96875.871028] INFO: lockdep is turned off.
> > > [96875.871031] Last Breaking-Event-Address:
> > > [96875.871037] [<000003e0012e3660>] dm_softirq_done+0x98/0x140 [dm_mod]
> > >
> > > static int multipath_end_io(struct dm_target *ti, struct request *clone,
> > > int error, union map_info *map_context)
> > > {
> > > struct multipath *m = ti->private;
> > > struct dm_mpath_io *mpio = map_context->ptr;
> > > struct pgpath *pgpath = mpio->pgpath;
> > > struct path_selector *ps;
> > > int r;
> > >
> > > r = do_end_io(m, clone, error, mpio);
> > > if (pgpath) {
> > > ps = &pgpath->pg->ps; <--- crashes here
> > > if (ps->type->end_io)
> > > ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
> > > }
> > > mempool_free(mpio, m->mpio_pool);
> > >
> > > return r;
> > > }
> > >
> > > It crashes when trying to derefence pgpath, which was freed. Since we have
> > > SLUB debugging turned on the freed object tells us that it was allocated
> > > via a call to multipath_ctr() and freed via a call to free_priority_group().
> >
> > struct pgpath is freed before dm_target when tearing down dm table.
> > So if the problematic completion was being done after freeing pgpath
> > but before freeing dm_target, crash would look like that
> > and what's happening seems the same for these dm crashes:
> > dm table was somehow destroyed while I/O was in-flight.
>
> Could be the block layer's onstack plugging changes are at the heart of
> this.
>
> I voiced onstack plugging concerns relative to DM some time ago
> (https://lkml.org/lkml/2011/3/9/450) but somehow convinced myself DM was
> fine to no longer need dm_table_unplug_all() etc. Unfortunately I
> cannot recall _why_ I felt that was the case.
>
> So DM needs further review relative to block's onstack plugging changes
> and DM IO completion.
dm_suspend is performed as part of the DM table reload that is being
done my multipathd during path failure. Seems DM no longer insures
inflight requests have finished during dm_suspend().
Before onstack plugging (< 2.6.39):
dm_suspend() -> dm_wait_for_completion() -> dm_unplug_all() -> dm_table_unplug_all()
After onstack plugging (>= 2.6.39, commit 7eaceaccab5f40bb):
dm_suspend's call to dm_wait_for_completion() no longer unplugs IO
(dm_unplug_all and dm_table_unplug_all were removed without introducing
a clear equivalent).
Mike
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
