On Tue 07-03-17 08:10:29, James Bottomley wrote: > On Tue, 2017-03-07 at 15:41 +0100, Jan Kara wrote: > > On Mon 06-03-17 09:25:42, James Bottomley wrote: > > > On Mon, 2017-03-06 at 17:13 +0100, Jan Kara wrote: > > > > On Mon 06-03-17 07:44:55, James Bottomley wrote: > > ... > > > > > > Sure. The call trace is: > > > > > > > > > > > > [ 41.919244] ------------[ cut here ]------------ > > > > > > [ 41.919263] WARNING: CPU: 4 PID: 2335 at > > > > > > drivers/scsi/sd.c:3332 > > > > > > sd_shutdown+0x2f/0x100 > > > > > > [ 41.919268] Modules linked in: scsi_debug(+) netconsole > > > > > > loop btrfs raid6_pq zlib_deflate lzo_compress xor > > > > > > [ 41.919319] CPU: 4 PID: 2335 Comm: modprobe Not tainted > > > > > > 4.11.0-rc1-xen+ #49 > > > > > > [ 41.919325] Hardware name: Bochs Bochs, BIOS Bochs > > > > > > 01/01/2011 > > > > > > [ 41.919331] Call Trace: > > > > > > [ 41.919343] dump_stack+0x8e/0xf0 > > > > > > [ 41.919354] __warn+0x116/0x120 > > > > > > [ 41.919361] warn_slowpath_null+0x1d/0x20 > > > > > > [ 41.919368] sd_shutdown+0x2f/0x100 > > > > > > [ 41.919374] sd_remove+0x70/0xd0 > > > > > > > > > > > > *** Here is the unexpected step I guess... > > > > > > > > > > > > [ 41.919383] driver_probe_device+0xe0/0x4c0 > > > > > > > > > > Exactly. It's this, I think > > > > > > > > > > bool test_remove = > > > > > IS_ENABLED(CONFIG_DEBUG_TEST_DRIVER_REMOVE) > > > > > && > > > > > !drv->suppress_bind_attrs; > > > > > > > > > > You have that config option set. > > > > > > > > Yes - or better said 0-day testing has it set. Maybe that is not > > > > a sane default for 0-day tests? The option is explicitely marked > > > > as "unstable"... Fengguang? > > > > > > > > > So the drivers base layer is calling ->remove after probe and > > > > > triggering the destruction of the queue. > > > > > > > > > > What to do about this (apart from nuke such a stupid option) is > > > > > somewhat more problematic. > > > > > > > > I guess this is between you and Greg :). > > > > > > Nice try, sport ... you qualify just behind Dan in the "not my > > > problem" olympic hurdles event. I'm annoyed because there's no > > > indication in the log that the driver add behaviour is radically > > > altered, so I've been staring at the wrong code for weeks. > > > However, the unbind/rebind should work, so the real issue is that > > > your bdi changes (or perhaps something else in block) have induced > > > a regression in the unbinding of upper layer drivers. If you're > > > going to release the bdi in del_gendisk, you have to have some > > > mechanism for re-acquiring it on re-probe (likely with the same > > > name) otherwise rebind is broken for every block driver. > > > > So my patch does not release bdi in del_gendisk(). Bdi has two > > initialization / destruction phases (similarly to request queue). You > > allocate and initialize bdi through bdi_init(), then you call > > bdi_register() to register it (which happens in device_add_disk()). > > On shutdown you have to first call bdi_unregister() (used to be > > called from blk_cleanup_queue(), my patch moved it to del_gendisk()). > > After that the last reference to bdi may be dropped which does final > > bdi destruction. > > > > So do I understand correctly that SCSI may call device_add_disk(), > > del_gendisk() repeatedly for the same request queue? > > Yes. The upper drivers (sd, sr, st and sg) follow a device model. > They can thus be bound and unbound many times during the lifetime of a > SCSI device. > > > If yes, then indeed I have a bug to fix... But gendisk seems to get > > allocated from scratch on each probe so we don't call > > device_add_disk(), del_gendisk() more times on the same disk, right? > > Right, gendisk, being a generic representation of a disk is a property > of the upper layer drivers. We actually cheat and use it in all of > them (including the apparent character ones, they use alloc_disk, > put_disk but not add_disk). So it has to be freed when the driver is > unbound and reallocated when it is bound. It's the fundamental entity > which embeds the SCSI upper layer driver lifetime. > > > > The fact that the second rebind tried with a different name > > > indicates that sd_devt_release wasn't called, so some vestige of > > > the devt remains on the subsequent rebind. > > > > Yep, I guess that's caused by Dan's patch (commit 0dba1314d4f8 now) > > which calls put_disk_devt() only in blk_cleanup_queue() which, if I > > understood you correctly, does not get called during unbind-bind > > cycle? In fact Dan's patch would end up leaking devt's because of > > repeated device_add_disk() calls for the same request queue... > > That's a bit unfortunate. > > > > Here's the problem: the queue belongs to SCSI (the lower layer), so > > > it's not going to change because it doesn't see the release. The > > > gendisk and its allied stuff belongs to sd so it gets freed and re > > > -created for the same queue. Something in block is very confused > > > when this happens. > > > > Yep, I think the binding of request queue to different gendisks is > > something I or Dan did not expect. > > OK, so I think we now understand why this is happening ... the question > is what we do about it? Is it fixable or do we need to go back to > basics? The fundamental property which got broken by these patches is > the idea of the separation of the lifecycles of the queue and the > gendisk. The only requirement is that the queue must exist before the > gendisk is created and must be destroyed after it. This doesn't forbid > there being many gendisk lifetimes over one queue lifetime and this is > what we rely on for the SCSI driver model. We rarely get called on it > because unbind/rebind of a ULD is rarely done in production (it's > mostly a devtest thing). For now I think it's fixable. Let's see what I'll be able to come up with. Let me also note that before my patch moving bdi_unregister() to del_gendisk() (commit 165a5e22fafb), the SCSI driver model was buggy in a different way - because bdi_unregister() was called only from blk_cleanup_queue(), we were leaking bdi->owner reference to gendisk by repeated calls to bdi_register_owner() from device_add_disk() and thus gendisks were never freed after unbind/bind (I've actually verified this bug exists with debug messages). Also as a result request queue use count was steadily increasing with each new gendisk and request queue could never get freed after unbind / rebind. My patch silently fixed this leakage but introduced a different set of problems which I'll try to fix. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
