linux-kernel - Re: [PATCH] drivers/base/memory: Avoid overhead from for_each_present_section

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <5a37ec06-ff98-475a-8998-0330dc8073c0@redhat.com>
Date: Fri, 11 Apr 2025 10:15:42 +0200
From: David Hildenbrand <david@...hat.com>
To: Gavin Shan <gshan@...hat.com>, Oscar Salvador <osalvador@...e.de>
Cc: linux-mm@...ck.org, linux-kernel@...r.kernel.org, adityag@...ux.ibm.com,
 donettom@...ux.ibm.com, gregkh@...uxfoundation.org, rafael@...nel.org,
 dakr@...nel.org, akpm@...ux-foundation.org, shan.gavin@...il.com
Subject: Re: [PATCH] drivers/base/memory: Avoid overhead from
 for_each_present_section_nr()

On 11.04.25 07:04, Gavin Shan wrote:
> On 4/11/25 12:25 AM, David Hildenbrand wrote:
>> On 10.04.25 16:12, Oscar Salvador wrote:
>>> On Thu, Apr 10, 2025 at 03:55:19PM +0200, Oscar Salvador wrote:
>>>> All in all, I think we are better, and the code is slightly simpler?
>>>
>>> One thing to notice is that maybe we could further improve and leap 'nr'
>>> by the number of sections_per_block, so in those scenarios where
>>> a memory-block spans multiple sections this could be faster?
>>
>> Essentially, when we created a block we could always start with the next section that starts after the block.
>>
> 
> I think it's a good point. Tried a quick test on a ARM64 machine whose memory
> capacity is 1TB. Leaping 'nr' by 'sections_per_block' improves the performance a bit,
> even it's not too much. The time taken by memory_dev_init() drops from 110ms to 100ms.
> For the IBM Power9 machine (64GB memory) I have, there are not too much space to be
> improved because the time taken by memory_dev_init() is only 10ms. I will post a patch
> for review after this patch gets merged, if you agree.
> 
>           for_each_present_section_nr(0, nr) {
> -               if (block_id != ULONG_MAX && memory_block_id(nr) == block_id)
> -                       continue;
> -
> -               block_id = memory_block_id(nr);
> -               ret = add_memory_block(block_id, MEM_ONLINE, NULL, NULL);
> +               ret = add_memory_block(memory_block_id(nr), MEM_ONLINE, NULL, NULL);
>                   if (ret) {
>                           panic("%s() failed to add memory block: %d\n",
>                                 __func__, ret);
>                   }
> +
> +               /* Align to next block, minus one section */

/*
  * Forward to the last section in this block, so we'll process the first
  * section of the next block in the next iteration.
  */

> +               nr = ALIGN(nr + 1, sections_per_block) - 1;

Yeah, that should work.

-- 
Cheers,

David / dhildenb