Re: [PATCH] writeback: fix NULL dereference when device is gone

[Peter Wu <lekensteyn@gmail.com>]

Hi,

On Tuesday 20 August 2013 09:33:08 Tejun Heo wrote:
> On Tue, Aug 20, 2013 at 12:13:58PM +0200, Peter Wu wrote:
> ...
> 
> > > Hmmm... bdi->dev is cleared after bdi_wb_shutdown() so the work item
> > > should no longer be running.  It seems like something is queueing the
> > > work item after shutdown and the proper fix would be finding out which
> > > and fixing it.  Can you please verify whether adding
> > > WARN_ON(!bdi->dev) in bdi_wakeup_thread_delayed() trigger anything?
> >
> > 
> >
> > Initially I did not get any warnings, so I added more. The patch (on top
> > of
> 
> > v3.11-rc6-27-g94fc5d9 plus some unrelated r8169 patches):
> ...
> 
> > [  245.978170] WARNING: CPU: 1 PID: 2608 at
> > /home/pc/Linux-src/linux/mm/backing-dev.c:293
> > bdi_wakeup_thread_delayed+0x5e/0x60() [  245.978189] Modules linked in:
> > kvm_intel kvm dm_crypt binfmt_misc joydev snd_hda_codec_hdmi
> > snd_hda_codec_realtek hid_logitech_dj hid_generic mxm_wmi nls_iso8859_1
> > snd_hda_intel snd_hda_codec usbhid hid snd_hwdep psmouse usb_storage
> > snd_pcm serio_raw lpc_ich snd_seq_midi snd_seq_midi_event snd_rawmidi
> > snd_seq snd_seq_device snd_timer snd wmi it87 mac_hid hwmon_vid coretemp
> > soundcore snd_page_alloc r8169 eeprom_93cx6 mii pci_stub ahci libahci
> > i915 drm_kms_helper drm video i2c_algo_bit [  245.978191] CPU: 1 PID:
> > 2608 Comm: ata_id Tainted:
> > G        W    3.11.0-rc6-usbdbg-00030-g693d742-dirty #1 [  245.978192]
> > Hardware name: Gigabyte Technology Co., Ltd. To be filled by
> > O.E.M./Z68X-UD3H-B3, BIOS U1l 03/08/2013 [  245.978194]  0000000000000125
> > ffff8805d3a4fa98 ffffffff8165986e 00000000000060d0
> > [  245.978196]  0000000000000000 ffff8805d3a4fad8 ffffffff81047acc
> > ffff880602f6beb8 [  245.978197]  ffff8805fc024618 ffff880602f6be30
> > ffffffff81c58f80 ffff880602f6beb8 [  245.978198] Call Trace:
> > [  245.978202]  [<ffffffff8165986e>] dump_stack+0x55/0x76
> > [  245.978204]  [<ffffffff81047acc>] warn_slowpath_common+0x8c/0xc0
> > [  245.978206]  [<ffffffff81047b1a>] warn_slowpath_null+0x1a/0x20
> > [  245.978207]  [<ffffffff811424fe>] bdi_wakeup_thread_delayed+0x5e/0x60
> > [  245.978211]  [<ffffffff811bdc71>] bdev_inode_switch_bdi+0xf1/0x160
> > [  245.978212]  [<ffffffff811bedd2>] __blkdev_get+0x372/0x4d0
> > [  245.978216]  [<ffffffff811bf115>] blkdev_get+0x1e5/0x380
> > [  245.978221]  [<ffffffff811bf36f>] blkdev_open+0x5f/0x90
> > [  245.978223]  [<ffffffff81180cd6>] do_dentry_open+0x226/0x2a0
> > [  245.978225]  [<ffffffff81180fa5>] finish_open+0x35/0x50
> > [  245.978227]  [<ffffffff81192d9e>] do_last+0x48e/0x7a0
> > [  245.978229]  [<ffffffff81193174>] path_openat+0xc4/0x4e0
> > [  245.978230]  [<ffffffff81193d83>] do_filp_open+0x43/0xa0
> > [  245.978234]  [<ffffffff81182422>] do_sys_open+0x132/0x220
> > [  245.978236]  [<ffffffff8118252e>] SyS_open+0x1e/0x20
> > [  245.978238]  [<ffffffff8166e802>] system_call_fastpath+0x16/0x1b
> 
> Hmmm... looks like the udev event handling from hotunplugging ends up
> opening up the device which in turn schedules an already shutdown bdi.
> The root problem here seems that there is no proper synchronization
> around bdi shutdown.  Ideally, a bdi should be marked offline
> preventing further activities and then shut down but we instead shut
> it down first and then clear bdi->dev.  As bdi's lifetime is tied to
> the backing request_queue, it might be okay as unsynchronized access
> to bdi->dev should be safe as long as the bdi struct itself is
> accessible.  Still nasty tho.
> 
> Not sure what to do.  The quick fix would be doing the following from
> workfn().
> 
>         dev = bdi->dev;
>         if (!dev)               // bdi already shutdown
>                 return;
> 
>         use @dev;               // can't use bdi->dev, as it can be cleared
> anytime
>
> But it's nasty.  A better way to do it would be, from
> bdi_wb_shutdown(), marking the bdi as offline and ensure that
> bdi_wakeup_thread_delayed() sees that, flush the work item and then
> clear bdi->dev.

I applied your fix in the workfn, but now dd does not stop with EIO. If I run 
`sync`, then I see:

[11882.645618] quiet_error: 591905 callbacks suppressed
[11882.650589] Buffer I/O error on device sdd, logical block 129652880
[11882.656850] lost page write due to I/O error on sdd
[11882.661974] Buffer I/O error on device sdd, logical block 129652881
...

Any other ideas? For now I will stick to the original patch. That probably 
does not solve the root issues, but it is at least a workaround to prevent the 
machine from locking up.

Regards,
Peter