On Wed, May 15, 2019 at 11:12 AM Pavel Tatashin
<pasha.tatashin@xxxxxxxxxx> wrote:
>
> > Hi Pavel,
> >
> > I am working on adding this sort of a workflow into a new daxctl command
> > (daxctl-reconfigure-device)- this will allow changing the 'mode' of a
> > dax device to kmem, online the resulting memory, and with your patches,
> > also attempt to offline the memory, and change back to device-dax.
> >
> > In running with these patches, and testing the offlining part, I ran
> > into the following lockdep below.
> >
> > This is with just these three patches on top of -rc7.
> >
> >
> > [ +0.004886] ======================================================
> > [ +0.001576] WARNING: possible circular locking dependency detected
> > [ +0.001506] 5.1.0-rc7+ #13 Tainted: G O
> > [ +0.000929] ------------------------------------------------------
> > [ +0.000708] daxctl/22950 is trying to acquire lock:
> > [ +0.000548] 00000000f4d397f7 (kn->count#424){++++}, at: kernfs_remove_by_name_ns+0x40/0x80
> > [ +0.000922]
> > but task is already holding lock:
> > [ +0.000657] 000000002aa52a9f (mem_sysfs_mutex){+.+.}, at: unregister_memory_section+0x22/0xa0
>
> I have studied this issue, and now have a clear understanding why it
> happens, I am not yet sure how to fix it, so suggestions are welcomed
> :)
I would think that ACPI hotplug would have a similar problem, but it does this:
acpi_unbind_memory_blocks(info);
__remove_memory(nid, info->start_addr, info->length);
I wonder if that ordering prevents going too deep into the
device_unregister() call stack that you highlighted below.
>
> Here is the problem:
>
> When we offline pages we have the following call stack:
>
> # echo offline > /sys/devices/system/memory/memory8/state
> ksys_write
> vfs_write
> __vfs_write
> kernfs_fop_write
> kernfs_get_active
> lock_acquire kn->count#122 (lock for
> "memory8/state" kn)
> sysfs_kf_write
> dev_attr_store
> state_store
> device_offline
> memory_subsys_offline
> memory_block_action
> offline_pages
> __offline_pages
> percpu_down_write
> down_write
> lock_acquire mem_hotplug_lock.rw_sem
>
> When we unbind dax0.0 we have the following stack:
> # echo dax0.0 > /sys/bus/dax/drivers/kmem/unbind
> drv_attr_store
> unbind_store
> device_driver_detach
> device_release_driver_internal
> dev_dax_kmem_remove
> remove_memory device_hotplug_lock
> try_remove_memory mem_hotplug_lock.rw_sem
> arch_remove_memory
> __remove_pages
> __remove_section
> unregister_memory_section
> remove_memory_section mem_sysfs_mutex
> unregister_memory
> device_unregister
> device_del
> device_remove_attrs
> sysfs_remove_groups
> sysfs_remove_group
> remove_files
> kernfs_remove_by_name
> kernfs_remove_by_name_ns
> __kernfs_remove kn->count#122
>
> So, lockdep found the ordering issue with the above two stacks:
>
> 1. kn->count#122 -> mem_hotplug_lock.rw_sem
> 2. mem_hotplug_lock.rw_sem -> kn->count#122
