On Thu, 2014-04-24 at 12:02 +0200, Rafael J. Wysocki wrote:
On 4/24/2014 10:59 AM, Li Zhong wrote:I can understand that hot-addition needs the device_hotplug lock, but
On Wed, 2014-04-23 at 18:12 +0200, Rafael J. Wysocki wrote:No and no.
On 4/23/2014 4:23 PM, Tejun Heo wrote:Thank you for the education. It's more clear to me now why we need this
Hello, Rafael.Hi,
On Wed, Apr 23, 2014 at 12:21:33AM +0200, Rafael J. Wysocki wrote:Well, I do my best not to add unnecessary locks if that's what you mean.
Can you please elaborate a bit?Because it can get involved in larger locking dependency issues by
joining dependency graphs of two otherwise largely disjoint
subsystems. It has potential to create possible deadlocks which don't
need to exist.
That locking is not sufficient.It is there to protect hotplug operations involving multiple devicesBut why would different subsystems, say cpu and memory, use the same
(in different subsystems) from racing with each other. Why exactly
is it bad?
lock? Wouldn't those subsystems already have proper locking inside
their own subsystems?
Why add this additional global lock across multiple subsystems?That basically is because of how eject works when it is triggered via ACPI.
It is signaled for a device at the top of a subtree. It may be a
container of some sort and the eject involves everything below that
device in the ACPI namespace. That may involve multiple subsystem
(CPUs, memory, PCI host bridge, etc.).
We do that in two steps, offline (which can fail) and eject proper
(which can't fail and makes all of the involved devices go away). All
that has to be done in one go with respect to the sysfs-triggered
offline/online and that's why the lock is there.
lock.
I still have some small questions about when this lock is needed:
I could understand sysfs-triggered online is not acceptable when
removing devices in multiple subsystems. But maybe concurrent offline
and remove(with proper per subsystem locks) seems not harmful?
And if we just want to remove some devices in a specific subsystem, e.g.
like writing /cpu/release, if it just wants to offline and remove some
cpus, then maybe we don't require the device_hotplug_lock to be taken?
If the offline phase fails for any device in the subtree, we roll back
the operation
and online devices that have already been offlined by it. Also the ACPI
hot-addition
needs to acquire device_hotplug_lock so that it doesn't race with ejects
and so
that lock needs to be taken by sysfs-triggered offline too.
still not very clear about the offline.
I guess your are describing following scenario:
user A: (trying remove cpu 1 and memory 1-10)
lock_device_hotplug
offline cpu with cpu locks -- successful
offline memories with memory locks -- failed, e.g. for memory8
online cpu and memory with their locks
unlock_device_hotplug
user B: (trying offline cpu 1)
offline cpu with cpu locks
But I don't see any problem for user B not taking the device_hotplug
lock. The result may be different for user B to take or not take the
lock. But I think it could be seen as concurrent online/offline for cpu1
under cpu hotplug locks, which just depends on which is executed last?
Or did I miss something here?