slow boot with 7fef431be9c9 ("mm/page_alloc: place pages to tail in __free_pages_core()")

[David Hildenbrand]

On 13.03.21 14:48, Mike Rapoport wrote:
> Hi,
> 
> On Sat, Mar 13, 2021 at 10:05:23AM +0100, David Hildenbrand wrote:
>>> Am 13.03.2021 um 05:04 schrieb Liang, Liang (Leo) <Liang.Liang at amd.com>:
>>>
>>> Hi David,
>>>
>>> Which benchmark tool you prefer? Memtest86+ or else?
>>
>> Hi Leo,
>>
>> I think you want something that runs under Linux natively.
>>
>> I‘m planning on coding up a kernel module to walk all 4MB pages in the
>> freelists and perform a stream benchmark individually. Then we might be
>> able to identify the problematic range - if there is a problematic range :)
> 
> My wild guess would be that the pages that are now at the head of free
> lists have wrong caching enabled. Might be worth checking in your test
> module.

I hacked something up real quick:

https://github.com/davidhildenbrand/kstream

Only briefly tested inside a VM. The output looks something like

[...]
[ 8396.432225] [0x0000000045800000 - 0x0000000045bfffff] 25322 MB/s / 
38948 MB/s
[ 8396.448749] [0x0000000045c00000 - 0x0000000045ffffff] 24481 MB/s / 
38946 MB/s
[ 8396.465197] [0x0000000046000000 - 0x00000000463fffff] 24892 MB/s / 
39170 MB/s
[ 8396.481552] [0x0000000046400000 - 0x00000000467fffff] 25222 MB/s / 
39156 MB/s
[ 8396.498012] [0x0000000046800000 - 0x0000000046bfffff] 24416 MB/s / 
39159 MB/s
[ 8396.514397] [0x0000000046c00000 - 0x0000000046ffffff] 25469 MB/s / 
38940 MB/s
[ 8396.530849] [0x0000000047000000 - 0x00000000473fffff] 24885 MB/s / 
38734 MB/s
[ 8396.547195] [0x0000000047400000 - 0x00000000477fffff] 25458 MB/s / 
38941 MB/s
[...]

The benchmark allocates one 4 MiB chunk at a time and runs a simplified 
STREAM benchmark a) without flushing caches b) flushing caches before 
every memory access.

It would be great if you could run that with the *old behavior* kernel 
(IOW, without 7fef431be9c9), so we might still be lucky to catch the 
problematic area in the freelist.

Let's see if that will indicate anything.

-- 
Thanks,

David / dhildenb