Hi Rafael,
Thanks for your comments! And please refer to inlined comments below.
On 11/19/2012 04:43 PM, Rafael J. Wysocki wrote:
> Hi,
>
> On Wednesday, November 07, 2012 12:18:07 AM Jiang Liu wrote:
>> Hi Bjorn,
>> Thanks for your review and please refer to inlined comments below.
>>
>> On 11/06/2012 05:05 AM, Bjorn Helgaas wrote:
>>> On Sat, Nov 3, 2012 at 10:07 AM, Jiang Liu <liuj97@xxxxxxxxx> wrote:
>>>> Modern high-end servers may support advanced RAS features, such as
>>>> system device dynamic reconfiguration. On x86 and IA64 platforms,
>>>> system device means processor(CPU), memory device, PCI host bridge
>>>> and even computer node.
>>>>
>>>> The ACPI specifications have provided standard interfaces between
>>>> firmware and OS to support device dynamic reconfiguraiton at runtime.
>>>> This patch series introduces a new framework for system device
>>>> dynamic reconfiguration based on ACPI specification, which will
>>>> replace current existing system device hotplug logic embedded in
>>>> ACPI processor/memory/container device drivers.
>>>>
>>>> The new ACPI based hotplug framework is modelled after the PCI hotplug
>>>> architecture and target to achieve following goals:
>>>> 1) Optimize device configuration order to achieve best performance for
>>>> hot-added system devices. For best perforamnce, system device should
>>>> be configured in order of memory -> CPU -> IOAPIC/IOMMU -> PCI HB.
>>>> 2) Resolve dependencies among hotplug slots. You need first to remove
>>>> the memory device before removing a physical processor if a
>>>> hotpluggable memory device is connected to a hotpluggable physical
>>>> processor.
>>>
>>> Doesn't the namespace already have a way to communicate these dependencies?
>
>> The namespace could could resolve most dependency issues, but there are still
>> several corner cases need special care.
>> 1) On a typical Intel Nehalem/Westmere platform, an IOH will be connected to
>> two physical processors through QPI. The IOH depends on the two processors.
>> And the ACPI namespace is something like:
>> /_SB
>> |_SCK0
>> |_SCK1
>> |_PCI1
>> 2) For a large system composed up of multiple computer nodes, nodes may have
>> dependency on neighbors due to interconnect topology constraints.
>>
>> So we need to resolve dependency by both evaluating _EDL and analyze ACPI
>> namespace topology.
>
> Well, this doesn't explain why we need a new framework.
Currently hotplug capable system devices are managed by different ACPI drivers,
such as processor, container and acpi_memhotplug. And these ACPI drivers only
know hotplug-capable devices managed by itself and don't know hotplug-capable
devices managed by other drivers. For example, ACPI processor driver doesn't
know hotplug memory devices. So existing ACPI system device hotplug drivers
don't resolve dependencies at all.
With the new framework, it knows all hotplug domains within the system, and
it will automatically resolves dependencies when changing state of a hotplug
domain. This is very important because hotplug requests may be triggered
concurrently through different mechanism, such pressing the hotplug button,
issued from OS or issued from the OOB system management center. The new
framework will guarantee that all requests are correctly serialized.
>>>> 3) Provide interface to cancel ongoing hotplug operations. It may take
>>>> a very long time to remove a memory device, so provide interface to
>>>> cancel the inprogress hotplug operations.
>>>> 4) Support new advanced RAS features, such as socket/memory migration.
>>>> 5) Provide better user interfaces to access the hotplug functionalities.
>>>> 6) Provide a mechanism to detect hotplug slots by checking existence
>>>> of ACPI _EJ0 method or by other hardware platform specific methods.
>>>
>>> I don't know what "hotplug slot" means for ACPI. ACPI allows hotplug
>>> of arbitrary devices in the namespace, whether they have EJ0 or not.
>
>> Here "hotplug" slot is an abstraction of receptacles where a group of
>> system devices could be attached to, or where we could control a group
>> of system devices. It's totally conceptual, may or may not has
>> corresponding physical slots.
>
> Can that be called something different from "slot", then, to avoid confusion?
On some other OSes, it's called "attachment point" or "AP". I use slot
because its used by PCI hotplug. If that causes confusion, we may call
it something else, any suggestion?
>> For example,
>> 1) a hotplug slot for a hotpluggable memory board has a physical slot.
>
> So let's call that a "slot" and the abstraction above a "hotplug domain" or
> something similar. Because in fact we're talking about hotplug domains,
> aren't we?
Yeah, maybe "hotplug domain" is a good candidate for "hotplug slot".
>> 2) a hotplug slot for a non-hotpluggable processor with power control
>> capability has no physical slot. (That means you may power on/off a
>> physical processor but can't hotplug it at runtime). This case is useful
>> for hardware partitioning.
>
> People have been working on this particular thing for years, so I wonder
> why you think that your apprach is going to be better here?
To support hardware partitioning or power management, OS needs to provide
interfaces to control those power-control-capable devices. But currently
we have only interface to power off a device and without interface to
power on a device.
Another benefit of the new framework is to support memory migration. A memory
device may support memory migration but don't support memory hotplug. The
new framework will abstract the migration-capable memory device as a hotplug
domain, so user can manage the memory device.
>> Detecting hotplug slots by checking existence of _EJ0 is the default
>> but unreliable way. For a real high-end server with system device
>> hotplug capabilities should provide some static ACPI table to describe
>> hotplug slots/capabilities. There are some ongoing efforts for that from
>> Intel, but not in the public domain yet. So the hotplug slot enumeration
>> driver is designed to extensible:)
>>
>>>> 7) Unify the way to enumerate ACPI based hotplug slots. All hotplug
>>>> slots will be enumerated by the enumeration driver (acpihp_slot),
>>>> instead of by individual ACPI device drivers.
>>>
>>> Why do we need to enumerate these "slots" specifically?
>>>
>>> I think this patch adds things in /sys. It might help if you
>>> described what they are.
>
>> There's no standard way in ACPI5.0 to describe system device hotplug slots yet.
>> And we want to show user the system device hotplug capabilities even when there
>> is no device attached to a slot. In other word, user could now how much
>> devices they could connect to the system by hotplugging.
>
> Bjorn probably meant "provide documentation describing the user space interfaces
> being introduced". Which in fact is a requirement.
Thanks for reminder. Documentation for sysfs interface is a must, and we have
plan for that. How about providing sysfs interface docs after we have settled
down about the overall idea and design?
>>>> 8) Unify the way to handle ACPI hotplug events. All ACPI hotplug events
>>>> for system devices will be handled by a generic ACPI hotplug driver
>>>> (acpihp_drv) instead of by individual ACPI device drivers.
>>>> 9) Provide better error handling and error recovery.
>>>> 10) Trigger hotplug events/operations by software. This feature is useful
>>>> for hardware fault management and/or power saving.
>>>>
>>>> The new framework is composed up of three major components:
>>>> 1) A system device hotplug slot enumerator driver, which enumerates
>>>> hotplug slots in the system and provides platform specific methods
>>>> to control those slots.
>>>> 2) A system device hotplug driver, which is a platform independent
>>>> driver to manage all hotplug slots created by the slot enumerator.
>>>> The hotplug driver implements a state machine for hotplug slots and
>>>> provides user interfaces to manage hotplug slots.
>>>> 3) Several ACPI device drivers to configure/unconfigure system devices
>>>> at runtime.
>>>>
>>>> To get rid of inter dependengcy between the slot enumerator and hotplug
>>>> driver, common code shared by them will be built into the kernel. The
>>>> shared code provides some helper routines and a device class named
>>>> acpihp_slot_class with following default sysfs properties:
>>>> capabilities: RAS capabilities of the hotplug slot
>>>> state: current state of the hotplug slot state machine
>>>> status: current health status of the hotplug slot
>>>> object: ACPI object corresponding to the hotplug slot
>>>>
>>>> Signed-off-by: Jiang Liu <jiang.liu@xxxxxxxxxx>
>>>> Signed-off-by: Gaohuai Han <hangaohuai@xxxxxxxxxx>
>>>
>>> ...
>>>> +static char *acpihp_dev_mem_ids[] = {
>>>> + "PNP0C80",
>>>> + NULL
>>>> +};
>>>> +
>>>> +static char *acpihp_dev_pcihb_ids[] = {
>>>> + "PNP0A03",
>>>> + NULL
>>>> +};
>>>
>>> Why should this driver need to know about these PNP IDs? We ought to
>>> be able to support hotplug of any device in the namespace, no matter
>>> what its ID.
>
>> We need PNP IDs for:
>> 1) Give a meaningful name for each slot.
>> lrwxrwxrwx CPU00 -> ../../../devices/LNXSYSTM:00/acpihp/CPU00
>> lrwxrwxrwx CPU01 -> ../../../devices/LNXSYSTM:00/acpihp/CPU01
>> lrwxrwxrwx CPU02 -> ../../../devices/LNXSYSTM:00/acpihp/CPU02
>> lrwxrwxrwx CPU03 -> ../../../devices/LNXSYSTM:00/acpihp/CPU03
>> lrwxrwxrwx IOX01 -> ../../../devices/LNXSYSTM:00/acpihp/IOX01
>> lrwxrwxrwx MEM00 -> ../../../devices/LNXSYSTM:00/acpihp/CPU00/MEM00
>> lrwxrwxrwx MEM01 -> ../../../devices/LNXSYSTM:00/acpihp/CPU00/MEM01
>> lrwxrwxrwx MEM02 -> ../../../devices/LNXSYSTM:00/acpihp/CPU01/MEM02
>> lrwxrwxrwx MEM03 -> ../../../devices/LNXSYSTM:00/acpihp/CPU01/MEM03
>> lrwxrwxrwx MEM04 -> ../../../devices/LNXSYSTM:00/acpihp/CPU02/MEM04
>> lrwxrwxrwx MEM05 -> ../../../devices/LNXSYSTM:00/acpihp/CPU02/MEM05
>> lrwxrwxrwx MEM06 -> ../../../devices/LNXSYSTM:00/acpihp/CPU03/MEM06
>> lrwxrwxrwx MEM07 -> ../../../devices/LNXSYSTM:00/acpihp/CPU03/MEM07
>>
>> 2) Classify system device into groups according to device types, so we could
>> configure/unconfigure them in optimal order for performance as:
>> memory -> CPU -> IOAPIC -> PCI host bridge
>>
>> 3) The new hotplug framework are designed to handle system device hotplug,
>> and it won't hand IO device hotplug such as PCI etc. So it need to stop
>> scanning subtree of PCI host bridges.
>
> Well, we probably need a hotplug domains framework, which the thing you're
> proposing seems to be. However, the question is: Why should it cover "system
> devices" only?
>
> To me, it looks like such a framework should cover all hotplug devices in the
> system, or at least all ACPI-based hotplug devices.
It will be great if we could provide a framework for all types of devices,
including both system devices and IO devices. But when starting the ACPI
based system device hotplug project, we found almost all hotplug capable
IO buses/devices have already provided hotplug support, so we focused
on system device hotplug because it's still incomplete.
According to my experience, it's really convenient if OS could provide a
unified tool/interface to support both system device and io device hotplug,
with consistent commands and outputs. But it's hard to provide a framework
to support both system device and io device hotplug at kernel level.
And an example system's hotplug framework is composed up with:
1) a hotplug driver for each type of devices, such as USB, SCSI, PCI and
system device.
2) a unified userspace application/interface to support all types of hotplug
drivers
So with that in mind, we have designed the ACPI based system device hotplug
framework as a hotplug driver for system devices. And we still lack of a
unified userspace application/interface to support all types of device hotplug.
Thanks
Gerry
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
