lkml.org 
[lkml]   [2022]   [Apr]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH 07/10] crypto: Use ARCH_DMA_MINALIGN instead of ARCH_KMALLOC_MINALIGN
On Fri, 15 Apr 2022 at 00:26, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> On Thu, Apr 14, 2022 at 12:49 PM Catalin Marinas
> <catalin.marinas@arm.com> wrote:
> >
> > It's a lot worse, ARCH_KMALLOC_MINALIGN is currently 128 bytes on arm64.
> > I want to at least get it down to 64 with this series while preserving
> > the current kmalloc() semantics.
>
> So here's a thought - maybe we could do the reverse of GFP_DMA, and
> add a flag to the places that want small allocations and know they
> don't need DMA?
>
> Because even 64 bytes is _huge_ for those small allocations. And 128
> bytes is just insane.
>
> Maybe we don't have a ton of them, but I just checked my desktop, and
> if my laptop had 17k 8-byte allocation, on my desktop it's currently
> sitting at 43k of them. And I've got a lot of 16- and 32-byte ones
> too:
>
> kmalloc-32 51529 51712 32 128 1 :
> kmalloc-16 45056 45056 16 256 1 :
> kmalloc-8 43008 43008 8 512 1 :
>
> Don't ask me what they are. It's probably fairly easy to find out, and
> it's probably something silly like sysfs private pointer data or
> whatever.
>
> If I did my math right, with a 128-byte minimum allocation, that is
> 16MB of wasted memory.
>
> Yeah, yeah, I've got 64GB or RAM in this thing, and there are things
> that take a lot more memory than the above (mem_map etc), and 64MB is
> peanuts at just 0.1% of RAM.
>

Actually, I think the impact on D-cache efficiency is a much bigger
concern, although I don't have any data to back that up. On arm64,
every 8 byte allocation takes up an entire cacheline, pushing out
other data that we could meaningfully keep there. And the doubling to
128 that is unnecessary on most arm64 systems probably has an
additional negative effect, as it means those allocations are heavily
skewed to use even indexes into the cache sets, causing extra
contention.

> Maybe nobody cares. But I really feel bad about that kind of egregious
> waste. The mem_map[] array may be big, it may use 1.5% of the memory I
> have, but at least it *does* something.
>
> And it could be that if I have 150k of those smallish allocations, a
> server with lots of active users might have millions. Not having
> looked at where they come from, maybe that isn't the case, but it
> *might* be.
>
> Maybe adding something like a
>
> static int warn_every_1k = 0;
> WARN_ON(size < 32 && (1023 & ++warn_every_1k));
>
> to kmalloc() would give us a statistical view of "lots of these small
> allocations" thing, and we could add GFP_NODMA to them. There probably
> aren't that many places that have those small allocations, and it's
> certainly safer to annotate "this is not for DMA" than have the
> requirement that all DMA allocations must be marked.
>
> Then just teach kmalloc() something like
>
> align = (gfp_flags & __GFP_NODMA) ? 8 : 128;
>
> on arm.
>

Sounds like a lot of churn as well, tbh. But perhaps there are a few
hot spots that we can fix that would alleviate the bulk of the issue,
who knows?

\
 
 \ /
  Last update: 2022-04-15 08:04    [W:0.295 / U:1.052 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site