Re: [PATCH 06/12] mm, slub: don't create kmalloc-rcl caches with CONFIG_SLUB_TINY

From: Vlastimil Babka
Date: Thu Nov 24 2022 - 07:12:19 EST


On 11/24/22 13:06, Hyeonggon Yoo wrote:
> On Wed, Nov 23, 2022 at 02:53:43PM +0100, Vlastimil Babka wrote:
>> On 11/21/22 18:11, Vlastimil Babka wrote:
>> > Distinguishing kmalloc(__GFP_RECLAIMABLE) can help against fragmentation
>> > by grouping pages by mobility, but on tiny systems the extra memory
>> > overhead of separate set of kmalloc-rcl caches will probably be worse,
>> > and mobility grouping likely disabled anyway.
>> >
>> > Thus with CONFIG_SLUB_TINY, don't create kmalloc-rcl caches and use the
>> > regular ones.
>> >
>> > Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
>>
>> Fixed up in response to lkp report for a MEMCG_KMEM+SLUB_TINY combo:
>> ---8<---
>> From c1ec0b924850a2863d061f316615d596176f15bb Mon Sep 17 00:00:00 2001
>> From: Vlastimil Babka <vbabka@xxxxxxx>
>> Date: Tue, 15 Nov 2022 18:19:28 +0100
>> Subject: [PATCH 06/12] mm, slub: don't create kmalloc-rcl caches with
>> CONFIG_SLUB_TINY
>>
>> Distinguishing kmalloc(__GFP_RECLAIMABLE) can help against fragmentation
>> by grouping pages by mobility, but on tiny systems the extra memory
>> overhead of separate set of kmalloc-rcl caches will probably be worse,
>> and mobility grouping likely disabled anyway.
>>
>> Thus with CONFIG_SLUB_TINY, don't create kmalloc-rcl caches and use the
>> regular ones.
>>
>> Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
>> ---
>> include/linux/slab.h | 9 +++++++--
>> mm/slab_common.c | 10 ++++++++--
>> 2 files changed, 15 insertions(+), 4 deletions(-)
>>
>> diff --git a/include/linux/slab.h b/include/linux/slab.h
>> index 45efc6c553b8..ae2d19ec8467 100644
>> --- a/include/linux/slab.h
>> +++ b/include/linux/slab.h
>> @@ -336,12 +336,17 @@ enum kmalloc_cache_type {
>> #endif
>> #ifndef CONFIG_MEMCG_KMEM
>> KMALLOC_CGROUP = KMALLOC_NORMAL,
>> -#else
>> - KMALLOC_CGROUP,
>> #endif
>> +#ifdef CONFIG_SLUB_TINY
>> + KMALLOC_RECLAIM = KMALLOC_NORMAL,
>> +#else
>> KMALLOC_RECLAIM,
>> +#endif
>> #ifdef CONFIG_ZONE_DMA
>> KMALLOC_DMA,
>> +#endif
>> +#ifdef CONFIG_MEMCG_KMEM
>> + KMALLOC_CGROUP,
>> #endif
>> NR_KMALLOC_TYPES
>> };
>
> Can you please elaborate what the lkp report was about
> and how you fixed it? I'm not getting what the problem of previous
> version is.

Report here:
https://lore.kernel.org/all/202211231949.nIyAWKam-lkp@xxxxxxxxx/

Problem is that if the preprocessing results in e.g.
KMALLOC_NORMAL = 0,
KMALLOC_DMA = KMALLOC_NORMAL
KMALLOC_CGROUP,
KMALLOC_RECLAIM = KMALLOC_NORMAL,
NR_KMALLOC_TYPES

then NR_KMALLOC_TYPES is not 2, but 1, because the enum's internal counter
got reset to 0 by KMALLOC_RECLAIM = KMALLOC_NORMAL. A common gotcha :/