[Stgt-devel] iscsi target works, minor fixes

[FUJITA Tomonori]

From: Pete Wyckoff <pw at osc.edu>
Subject: Re: [Stgt-devel] iscsi target works, minor fixes
Date: Sat, 4 Nov 2006 15:47:15 -0500

> fujita.tomonori at lab.ntt.co.jp wrote on Fri, 03 Nov 2006 23:28 +0900:
> > By the way, does the kernel crash when you kill the tgtd daemon (it's
> > due to the aiopoll problem)?
> 
> It did, but backing out the old aiopoll patch and putting in your
> new 2.6.19-rc4 one made that go away.  No clue why---there appears
> to be no significant difference in the patches.  I also updated my
> 2.6.19-rc4 tree with your git patches, but there was little changed
> there too.  The renaming of the kernel/user interface is all.
> 
> The old crash, in case you are curious, was via an NMI-detected
> lockup.  Tgtd was here:
> 
>     [<ffffffff8040263a>] _spin_lock_irqsave+0xa/0x20
>     [<ffffffff80242359>] remove_wait_queue+0x19/0x60
>     [<ffffffff802aa557>] ep_unregister_pollwait+0x37/0x70
>     [<ffffffff802aaee7>] ep_remove+0x27/0xf0
>     [<ffffffff802ab6f5>] eventpoll_release_file+0x65/0x90
>     [<ffffffff8027eb40>] __fput+0x90/0x1a0
>     [<ffffffff8027ec64>] fput+0x14/0x20
>     [<ffffffff8027c111>] filp_close+0x71/0x90
> 
> Presumably something else was holding the wait_queue spinlock.  But
> deleting the lun first, or deleting the lun and deleting the target,
> both resulted in the same lockup.  Maybe not the oops you were
> expecting.

I expected something like that, however, I didn't dig into the
details. I guess we have to drop the aiopoll patch rather than try to
fix it.