2022-05-17 00:31:43

by Roman Gushchin

Subject: Re: [PATCH] tracing: add ACCOUNT flag for allocations from marked slab caches

On Mon, May 16, 2022 at 11:41:27PM +0200, Vlastimil Babka wrote:
> On 5/16/22 21:10, Shakeel Butt wrote:
> > On Mon, May 16, 2022 at 11:53 AM Vasily Averin <[email protected]> wrote:
> >>
> >> Slab caches marked with SLAB_ACCOUNT force accounting for every
> >> allocation from this cache even if __GFP_ACCOUNT flag is not passed.
> >> Unfortunately, at the moment this flag is not visible in ftrace output,
> >> and this makes it difficult to analyze the accounted allocations.
> >>
> >> This patch adds the __GFP_ACCOUNT flag for allocations from slab caches
> >> marked with SLAB_ACCOUNT to the ftrace output.
> >>
> >> Signed-off-by: Vasily Averin <[email protected]>
> >> ---
> >> mm/slab.c | 3 +++
> >> mm/slub.c | 3 +++
> >> 2 files changed, 6 insertions(+)
> >>
> >> diff --git a/mm/slab.c b/mm/slab.c
> >> index 0edb474edef1..4c3da8dfcbdb 100644
> >> --- a/mm/slab.c
> >> +++ b/mm/slab.c
> >> @@ -3492,6 +3492,9 @@ void *__kmem_cache_alloc_lru(struct kmem_cache *cachep, struct list_lru *lru,
> >
> > What about kmem_cache_alloc_node()?
> >
> >> {
> >> void *ret = slab_alloc(cachep, lru, flags, cachep->object_size, _RET_IP_);
> >>
> >> + if (cachep->flags & SLAB_ACCOUNT)
> >
> > Should this 'if' be unlikely() or should we trace cachep->flags
> > explicitly to avoid this branch altogether?
>
> Hm I think ideally the tracepoint accepts cachep instead of current
> cachep->*size parameters and does the necessary extraction and
> modification in its fast_assign.

+1 for fast_assign

Changing flags just for tracing looks a bit excessive.

2022-05-17 04:16:45

by Vasily Averin

Subject: Re: [PATCH] tracing: add ACCOUNT flag for allocations from marked slab caches

On 5/17/22 01:08, Roman Gushchin wrote:
> On Mon, May 16, 2022 at 11:41:27PM +0200, Vlastimil Babka wrote:
>> On 5/16/22 21:10, Shakeel Butt wrote:
>>> On Mon, May 16, 2022 at 11:53 AM Vasily Averin <[email protected]> wrote:
>>>> {
>>>> void *ret = slab_alloc(cachep, lru, flags, cachep->object_size, _RET_IP_);
>>>>
>>>> + if (cachep->flags & SLAB_ACCOUNT)
>>>
>>> Should this 'if' be unlikely() or should we trace cachep->flags
>>> explicitly to avoid this branch altogether?
>>
>> Hm I think ideally the tracepoint accepts cachep instead of current
>> cachep->*size parameters and does the necessary extraction and
>> modification in its fast_assign.
>
> +1 for fast_assign
>
> Changing flags just for tracing looks a bit excessive.

At the kmem_cache_alloc and kmem_alloc use the same tracing template.
Ok, I'll try to redesign this.