On 10.09.20 13:35, Laurent Dufour wrote:
> Le 10/09/2020 à 13:12, Michal Hocko a écrit :
>> On Thu 10-09-20 09:51:39, Laurent Dufour wrote:
>>> Le 10/09/2020 à 09:23, Michal Hocko a écrit :
>>>> On Wed 09-09-20 18:07:15, Laurent Dufour wrote:
>>>>> Le 09/09/2020 à 12:59, Michal Hocko a écrit :
>>>>>> On Wed 09-09-20 11:21:58, Laurent Dufour wrote:
>>>> [...]
>>>>>>> For the point a, using the enum allows to know in
>>>>>>> register_mem_sect_under_node() if the link operation is due to a hotplug
>>>>>>> operation or done at boot time.
>>>>>>
>>>>>> Yes, but let me repeat. We have a mess here and different paths check
>>>>>> for the very same condition by different ways. We need to unify those.
>>>>>
>>>>> What are you suggesting to unify these checks (using a MP_* enum as
>>>>> suggested by David, something else)?
>>>>
>>>> We do have system_state check spread at different places. I would use
>>>> this one and wrap it behind a helper. Or have I missed any reason why
>>>> that wouldn't work for this case?
>>>
>>> That would not work in that case because memory can be hot-added at the
>>> SYSTEM_SCHEDULING system state and the regular memory is also registered at
>>> that system state too. So system state is not enough to discriminate between
>>> the both.
>>
>> If that is really the case all other places need a fix as well.
>> Btw. could you be more specific about memory hotplug during early boot?
>> How that happens? I am only aware of https://lkml.kernel.org/r/[email protected]
>> and that doesn't happen as early as SYSTEM_SCHEDULING.
>
> That points has been raised by David, quoting him here:
>
>> IIRC, ACPI can hotadd memory while SCHEDULING, this patch would break that.
>>
>> Ccing Oscar, I think he mentioned recently that this is the case with ACPI.
>
> Oscar told that he need to investigate further on that.
>
> On my side I can't get these ACPI "early" hot-plug operations to happen so I
> can't check that.
>
> If this is clear that ACPI memory hotplug doesn't happen at SYSTEM_SCHEDULING,
> the patch I proposed at first is enough to fix the issue.
>
Booting a qemu guest with 4 coldplugged DIMMs gives me:
:/root# dmesg | grep link_mem
[ 0.302247] link_mem_sections() during 1
[ 0.445086] link_mem_sections() during 1
[ 0.445766] link_mem_sections() during 1
[ 0.446749] link_mem_sections() during 1
[ 0.447746] link_mem_sections() during 1
So AFAICs everything happens during SYSTEM_SCHEDULING - boot memory and
ACPI (cold)plug.
To make forward progress with this, relying on the system_state is
obviously not sufficient.
1. We have to fix this instance and the instance directly in
get_nid_for_pfn() by passing in the context (I once had a patch to clean
that up, to not have two state checks, but it got lost somewhere).
2. The "system_state < SYSTEM_RUNNING" check in
register_memory_resource() is correct. Actual memory hotplug after boot
is not impacted. (I remember we discussed this exact behavior back then)
3. build_all_zonelists() should work as expected, called from
start_kernel() before sched_init().
--
Thanks,
David / dhildenb
Le 10/09/2020 à 14:00, David Hildenbrand a écrit :
> On 10.09.20 13:35, Laurent Dufour wrote:
>> Le 10/09/2020 à 13:12, Michal Hocko a écrit :
>>> On Thu 10-09-20 09:51:39, Laurent Dufour wrote:
>>>> Le 10/09/2020 à 09:23, Michal Hocko a écrit :
>>>>> On Wed 09-09-20 18:07:15, Laurent Dufour wrote:
>>>>>> Le 09/09/2020 à 12:59, Michal Hocko a écrit :
>>>>>>> On Wed 09-09-20 11:21:58, Laurent Dufour wrote:
>>>>> [...]
>>>>>>>> For the point a, using the enum allows to know in
>>>>>>>> register_mem_sect_under_node() if the link operation is due to a hotplug
>>>>>>>> operation or done at boot time.
>>>>>>>
>>>>>>> Yes, but let me repeat. We have a mess here and different paths check
>>>>>>> for the very same condition by different ways. We need to unify those.
>>>>>>
>>>>>> What are you suggesting to unify these checks (using a MP_* enum as
>>>>>> suggested by David, something else)?
>>>>>
>>>>> We do have system_state check spread at different places. I would use
>>>>> this one and wrap it behind a helper. Or have I missed any reason why
>>>>> that wouldn't work for this case?
>>>>
>>>> That would not work in that case because memory can be hot-added at the
>>>> SYSTEM_SCHEDULING system state and the regular memory is also registered at
>>>> that system state too. So system state is not enough to discriminate between
>>>> the both.
>>>
>>> If that is really the case all other places need a fix as well.
>>> Btw. could you be more specific about memory hotplug during early boot?
>>> How that happens? I am only aware of https://lkml.kernel.org/r/[email protected]
>>> and that doesn't happen as early as SYSTEM_SCHEDULING.
>>
>> That points has been raised by David, quoting him here:
>>
>>> IIRC, ACPI can hotadd memory while SCHEDULING, this patch would break that.
>>>
>>> Ccing Oscar, I think he mentioned recently that this is the case with ACPI.
>>
>> Oscar told that he need to investigate further on that.
>>
>> On my side I can't get these ACPI "early" hot-plug operations to happen so I
>> can't check that.
>>
>> If this is clear that ACPI memory hotplug doesn't happen at SYSTEM_SCHEDULING,
>> the patch I proposed at first is enough to fix the issue.
>>
>
> Booting a qemu guest with 4 coldplugged DIMMs gives me:
>
> :/root# dmesg | grep link_mem
> [ 0.302247] link_mem_sections() during 1
> [ 0.445086] link_mem_sections() during 1
> [ 0.445766] link_mem_sections() during 1
> [ 0.446749] link_mem_sections() during 1
> [ 0.447746] link_mem_sections() during 1
>
> So AFAICs everything happens during SYSTEM_SCHEDULING - boot memory and
> ACPI (cold)plug.
>
> To make forward progress with this, relying on the system_state is
> obviously not sufficient.
>
> 1. We have to fix this instance and the instance directly in
> get_nid_for_pfn() by passing in the context (I once had a patch to clean
> that up, to not have two state checks, but it got lost somewhere).
>
> 2. The "system_state < SYSTEM_RUNNING" check in
> register_memory_resource() is correct. Actual memory hotplug after boot
> is not impacted. (I remember we discussed this exact behavior back then)
>
> 3. build_all_zonelists() should work as expected, called from
> start_kernel() before sched_init().
I'm bit confused now.
Since hotplug operation is happening at SYSTEM_SCHEDULING like the regular
memory registration, would it be enough to add a parameter to
register_mem_sect_under_node() (reworking the memmap_context enum)?
That way the check is not based on the system state but on the calling path.