Re: [PATCH] mm/cma: add an API to enable/disable concurrent memory allocation for the CMA

[Ge Yang <yangge1116@126.com>]

在 2025/1/28 7:04, Andrew Morton 写道:
> On Fri, 24 Jan 2025 19:21:27 +0800 yangge1116@126.com wrote:
> 
>> From: yangge <yangge1116@126.com>
>>
>> Commit 60a60e32cf91 ("Revert "mm/cma.c: remove redundant cma_mutex lock"")
>> simply reverts to the original method of using the cma_mutex to ensure
>> that alloc_contig_range() runs sequentially. This change was made to avoid
>> concurrency allocation failures. However, it can negatively impact
>> performance when concurrent allocation of CMA memory is required.
>>
>> To address this issue, we could introduce an API for concurrency settings,
>> allowing users to decide whether their CMA can perform concurrent memory
>> allocations or not.
> 
> The term "users" tends to refer to userspace code.  Here I'm thinking
> you mean in-kernel code, so a better term to use is "callers".
Ok, thank you. I will change it in the next version.

> 
> This new interface has no callers.  We prefer not to merge unused code!
> Please send along the patch which calls cma_set_concurrency() so we
> can better understand this proposal and so that the new code is
> testable.
Ok, thank you. I will add the caller in the next version.

   In fact the patch has cc:stable, which makes things
> stranger.  Why should the -stable maintainers merge a patch which
> doesn't do anything?
> 
> And please quantify the benefit.  "negatively impact" is too vague.
> How much benefit can we expect our users to see from this?  Some
> runtime testing results would be good.
> 
> And please describe in more detail why this particular caller doesn't
> require concurrency protection.  And help other developers understand
> when it is safe for them to use concurr_alloc==false.
Ok, thank you.