RE: Boot regression (was "Re: [PATCH] genhd: Do not hold event lock when scheduling workqueue elements")

From: hch@lst.de [mailto:hch@lst.de]
Sent: Tuesday, February 14, 2017 22:51
To: Dexuan Cui <decui@microsoft.com>
Cc: hch@lst.de; Jens Axboe <axboe@kernel.dk>; Bart Van Assche
[off-list ref]; hare@suse.com; hare@suse.de; Martin K.
Petersen [off-list ref]; linux-kernel@vger.kernel.org;
linux-block@vger.kernel.org; jth@kernel.org; Nick Meier
[off-list ref]; Alex Ng (LIS) [off-list ref]; Long Li
[off-list ref]; Adrian Suhov (Cloudbase Solutions SRL) <v-
adsuho@microsoft.com>; Chris Valean (Cloudbase Solutions SRL) <v-
chvale@microsoft.com>
Subject: Re: Boot regression (was "Re: [PATCH] genhd: Do not hold event lock
when scheduling workqueue elements")

On Tue, Feb 14, 2017 at 02:46:41PM +0000, Dexuan Cui wrote:

quoted

quoted

From: hch@lst.de [mailto:hch@lst.de]
Sent: Tuesday, February 14, 2017 22:29
To: Dexuan Cui <decui@microsoft.com>
Subject: Re: Boot regression (was "Re: [PATCH] genhd: Do not hold event

lock

quoted

quoted

when scheduling workqueue elements")

Ok, thanks for testing.  Can you try the patch below?  It fixes a
clear problem which was partially papered over before the commit
you bisected to, although it can't explain why blk-mq still works.

Still bad luck. :-(

BTW, I'm using the first "bad" commit (scsi: allocate scsi_cmnd structures

as

quoted

part of struct request) + the 2 patches you provided today.

I suppose I don't need to test the 2 patches on the latest linux-next repo.

I'd love a test on that repo actually.  We had a few other for sense
handling since then I think.

I tested today's linux-next (next-20170214) + the 2 patches just now and got
a weird result: 
sometimes the VM stills hung with a new calltrace (BUG: spinlock bad
magic) , but sometimes the VM did boot up despite the new calltrace!

Attached is the log of a "good" boot.

It looks we have a memory corruption issue somewhere...

Actually previously I saw the "BUG: spinlock bad magic" message once, but I
couldn't repro it later, so I didn't mention it to you.

The good news is that now I can repro the "spinlock bad magic" message
every time. 
I tried to dig into this by enabling Kernel hacking -> Memory debugging,
but didn't find anything abnormal. 
Is it possible that the SCSI layer passes a wrong memory address?

Thanks,
-- Dexuan