On Monday 05 October 2009 15:13:22 you wrote:
> > > Hang on ... I looked at the bug report again: there's no actual kernel
> > > trace, just a theoretical function graph.
> > >
> > > Has this actually been seen or is it just the result of an analysis?
> > >
> > > If the latter (which I suspect), there's no actual problem. The
> > > explicit design of the calls is that device_initialize() and
> > > put_device() can be called from interrupt context. device_add() and
> > > device_del() must be called from user context.
> > >
> > > The path you seem to be showing is the put_device() path where there's
> > > been an error in the state model and the caller is doing last put on a
> > > visible device without having first called device_del().
> > >
> > > If you see the real kernel message about this, it means there's a bug
> > > in the device model handling somewhere in SCSI. If you haven't seen
> > > the message, it's just a bug in the static analysis tool.
> >
> > This bug report is the result of code inspection. I'm considering
> > functions which can call might_sleep macro and consequently which can not
> > be called from atomic context.
> > I choose function scsi_device_put. There are two paths to might_sleep
> > macro. First path was shown in the report, second is:
> > 1. scsi_device_put calls put_device at ./drivers/scsi/scsi.c:1111
> > 2. put_device calls kobject_put at ./drivers/base/core.c:1038
> > 3. kobject_put calls kref_put at ./lib/kobject.c
> > 4. kref_put may call callback function kobject_release at ./lib/kref.c if
> > refcount becomes zero
> > 5. kobject_cleanup calls kobject_del at ./lib/kobject.c:562
>
> only if state_in_sysfs is set.
>
> This is only set if the caller previously failed to call kobject_del
> (i.e. device_del).
>
> As long as devices follow the proper create->add->del->put paths, the
> final put may be called from interrupt context.
>
> Your analysis is wrong because you're basing it on the exception cleanup
> paths not the correct calling paths.
>
> James
>
> > 6. kobject_del calls sysfs_remove_dir at ./lib/kobject.c:516
> > 7. sysfs_remove_dir calls __sysfs_remove_dir at ./fs/sysfs/dir.c:821
> > 8. __sysfs_remove_dir calls sysfs_addrm_start at ./fs/sysfs/dir.c:789
> > 9. sysfs_addrm_start calls mutex_lock at ./fs/sysfs/dir.c:377, which
> > might_sleep because it calls might_sleep macro.
> >
> > As you wrote earlier, scsi_device_put was designed with the ability to
> > call last put from interrupt context, but as we can see from the paths
> > there might be situations where it is not true. Moreover, while analysing
> > different usage patterns of scsi_device_put, I found that people are
> > using scsi_device_put as if it can not be called from atomic context.
> > Because before calling scsi_device_put, spin_locks are always released
> > (i.e. spin_unlock is called before scsi_device_put and spin_lock is
> > called after it). Examples are: 1. drivers/scsi/dpt_i2o.c line 701
> > 2. drivers/ata/libata-scsi.c line 3626
> > 3. drivers/scsi/ipr.c line 2415
> >
> > >The path you seem to be showing is the put_device() path where there's
> > >been an error in the state model and the caller is doing last put on a
> > >visible device without having first called device_del().
> >
> > In scsi_lib.c prior to scsi_device_put we always do scsi_device_get. As
> > far as I understand, if we are sure that scsi_device_put is always not
> > last, then we can remove both calls to scsi_device_get and to
> > scsi_device_put from the code without introducing races.
> >
> > 347 list_for_each_entry_safe(sdev, tmp, &starget->devices,
> > 348 same_target_siblings) {
> > 349 if (sdev == current_sdev)
> > 350 continue;
> > 351 if (scsi_device_get(sdev))
> > 352 continue;
> > 353
> > 354 spin_unlock_irqrestore(shost->host_lock, flags);
> > 355 blk_run_queue(sdev->request_queue);
> > 356 spin_lock_irqsave(shost->host_lock, flags);
> > 357
> > 358 scsi_device_put(sdev);
> > 359 }
> >
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
> > the body of a message to majordomo@xxxxxxxxxxxxxxx
> > More majordomo info at http://vger.kernel.org/majordomo-info.html
James, what about code where spin_unlock is called before scsi_device_put,
especially for avoiding atomic context?
In code like
spin_unlock
scsi_device_put
spin_lock
Is spin_unlock/spin_lock redundant?
Why do we need scsi_device_get/scsi_device_put pair in scsi_lib.c at all? If
we are sure that scsi_device_put is always not last, for what purpose do we
call it together with scsi_device_get in the loop?
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
