Re: [3.12-rc] sg_open: leaving the kernel with locks still held!

[Douglas Gilbert <dgilbert@interlog.com>]

On 13-10-23 03:44 AM, James Bottomley wrote:
> On Tue, 2013-10-22 at 20:41 -0400, Douglas Gilbert wrote:
>> On 13-10-22 04:56 PM, Simon Kirby wrote:
>>> Hello!
>>>
>>> While trying to figure out why the request queue to sda (ext4) was
>>> clogging up on one of our btrfs backup boxes, I noticed a megarc process
>>> in D state, so enabled locking debugging, and got this (on 3.12-rc6):
>>>
>>> [  205.372823] ================================================
>>> [  205.372901] [ BUG: lock held when returning to user space! ]
>>> [  205.372979] 3.12.0-rc6-hw-debug-pagealloc+ #67 Not tainted
>>> [  205.373055] ------------------------------------------------
>>> [  205.373132] megarc.bin/5283 is leaving the kernel with locks still held!
>>> [  205.373212] 1 lock held by megarc.bin/5283:
>>> [  205.373285]  #0:  (&sdp->o_sem){.+.+..}, at: [<ffffffff8161e650>] sg_open+0x3a0/0x4d0
>>>
>>> Vaughan, it seems you touched this area last in 15b06f9a02406e, and git
>>> tag --contains says this went in for 3.12-rc. We didn't see this on 3.11,
>>> though I haven't tried with lockdep.
>>>
>>> This is caused by some of our internal RAID monitoring scripts that run
>>> "megarc.bin -dispCfg -a0" (even though that controller isn't present on
>>> this server -- a PowerEdge 2950 w/Perc 5).
>>>
>>> strace output of the program execution that causes the above message is
>>> here: http://0x.ca/sim/ref/3.12-rc6/megarc_strace.txt
>>
>> This has been reported. That patch will be reverted or,
>> if there is enough time, a fix will (or at least should)
>> go in before the release of lk 3.12 .
>
> I think you've got about a week to prove you can fix it (before 3.12
> goes final).  I'll send my current set of fixes to Linus without doing
> anything about sg.

"prove" is a big ask, especially coming from a
mathematician. I consider it more hacking (in the
golf sense) on my part to tweak well-meaning patches
to the sg driver that cause collateral damage. Further,
I suspect Vaughan's patch was an attempt to fix
damage left be a previous sg_open() hacker.

I have asked Simon Kirby to apply the patch:
   http://marc.info/?l=linux-scsi&m=138237283432010&w=2
and report if it fixes his problems. Further I have
written three test programs to test O_EXCL handling on
SCSI devices, two of which are in the examples directory
of sg3_utils version 1.37 . The latest one (single
exclusive writer, multiple readers) can be found in
the News section of:
    http://sg.danny.cz/sg/
These tests don't check all possibilities (e.g. random
signals, ml error processing and detached devices) but
they are better than nothing. And, as a side issue, they
break bsg (cause it ignores O_EXCL) and break the block
layer (e.g. /dev/sdb) so perhaps it should be reverted :-)

Perhaps the original bug reporter (Madper Xie) might also
test the proposed patch and report if it fixes what he saw.

Doug Gilbert