Re: [PATCH] scsi: take module reference during async scan

From: Douglas Gilbert <dgilbert@interlog.com>
To: James Bottomley <James.Bottomley@HansenPartnership.com>,
	Tomas Henzl <thenzl@redhat.com>,
	linux-scsi@vger.kernel.org
Subject: Re: [PATCH] scsi: take module reference during async scan
Date: Mon, 7 Sep 2020 17:32:08 -0400	[thread overview]
Message-ID: <a31b8640-307f-a14f-7cf8-7673fa8a4ff1@interlog.com> (raw)
In-Reply-To: <1599500808.4232.19.camel@HansenPartnership.com>

On 2020-09-07 1:46 p.m., James Bottomley wrote:
> On Mon, 2020-09-07 at 17:47 +0200, Tomas Henzl wrote:
>> During an async scan the driver shost->hostt structures are used,
>> that may cause issues when the driver is removed at that time.
>> As protection take the module reference.
> 
> Can I just ask what issues?  Today, our module model is that
> scsi_device_get() bumps the module refcount and therefore makes the
> module ineligible to be removed.  scsi_host_get() doesn't do this
> because the way the host model is supposed to be coded, we can call
> remove at any time but the module won't get freed until the last put of
> the host.  I can see we have a potential problem with
> scsi_forget_host() racing with the async scan thread ... is that what
> you see? What's supposed to happen is that scsi_device_get() starts
> failing as soon as the module begins it's exit routine, so if a scan is
> in progress, it can't add any new devices ... in theory this means that
> the list is stable for scsi_forget_host(), so knowing how that
> assumption is breaking would be useful.

James,
If you think it is bullet-proof try using CONFIG_DEBUG_TEST_DRIVER_REMOVE=y .
John Garry reported that:

  # insmod scsi_debug.ko

Gave errors like this:

[  140.115244] debugfs: Directory 'sde' with parent 'block' already present!
[  140.376426] debugfs: Directory 'sde' with parent 'block' already present!
[  140.420613] sd 3:0:0:0: [sde] tag#40 access beyond end of device
[  140.426655] blk_update_request: I/O error, dev sde, sector 15984 op 
0x0:(READ) flags 0x80700 phys_seg 1 prio class 0
[  140.437319] sd 3:0:0:0: [sde] tag#41 access beyond end of device
[  140.443368] blk_update_request: I/O error, dev sde, sector 15984 op 
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
...

Which wasn't the scsi_debug driver directly as it doesn't use debugfs. So
I suspect something is rotten in the mid-level.

When I tried to replicate John's config I couldn't even boot my Ubuntu
20.04 based system (with a MKP kernel). Seemed to fail/lockup before any
kernel prints came out to the serial port (yes, still useful), perhaps in
initrd. I'm guessing another, non-SCSI module caused the lockup. So I
gave up and turned off that config setting.

Doug Gilbert