Re: The process hangs during memory-hotplug.

[David Hildenbrand <david@redhat.com>]

On 13.02.22 12:07, Yang Yanchao wrote:
> Hello,
> 

Hi,

> I find a hanging issue during memory-hotplug on kernel-4.18.

you actually mean memory hotunplug / memory offlinig IIUC.

> Repetition steps:
> 1. malloc for all system memory, write 'x', then free
> 2. for each removable memory block:

Note that "removable=yes" was always racy and upstream Linux nowadays
only keeps that property around to not break older user space --
upstream Linux always reports "removable=yes" if memory offlining is
supported.

> echo offline > /sys/devices/system/memory/memoryXXX/state
> Then during the offline process, there is a high probability of being stuck for more than 20 minutes to five hours.
> cat /sys/ Device/system/Memory/memoryXXX/state
> The status is "going-offline"
> I try to understand it by adding some print to the kernel.The discovery process can't exit in this loop:
> __offline_pages
>     do_migrate_range
>         migrate_pages
>             unmap_and_move
>                 move_to_new_page
>                     fallback_migrate_page   --> return EAGAIN
> I try to clear the cache, but it don't seems to solve the problem.
> echo 3 > /proc/sys/vm/drop_caches
> Can I fix this problem with other Settings? Or can I see why it's stuck?

There are no real guarantees what will happen when trying offlinig a
memory block that's not onlined to ZONE_MOVABLE.

You can observe the zone e.g., via

$ cat /sys/devices/system/memory/memory40/valid_zones
Normal

Even with ZONE_MOVABLE, it can take quite a while (and in corner cases
eventually forever) until offlining succeeds.

Now, 20 minutes are a bit extreme. User space can always cancel
offlining -- in your example, by killing the "echo offline >
/sys/devices/system/memory/memoryXXX/state" process.


Having that said, as raised by Matthew, a lot changed since 4.18, so you
should try reproducing upstream. But even there, you can just cancel
offlining if it takes too long. If you observe similar behavior on
ZONE_MOVABLE, it would be interesting to find out how to better handle
that to make offlining succeed faster.

-- 
Thanks,

David / dhildenb