Re: [PATCH v3] tracing: add 'accounted' entry into output of allocation tracepoints

[Vasily Averin <vvs@openvz.org>]

On 5/19/22 17:03, Steven Rostedt wrote:
> On Thu, 19 May 2022 14:35:46 +0300
> Vasily Averin <vvs@openvz.org> wrote:
> 
>>>> @@ -33,42 +35,46 @@ DECLARE_EVENT_CLASS(kmem_alloc,
>>>>  		__entry->bytes_req	= bytes_req;
>>>>  		__entry->bytes_alloc	= bytes_alloc;
>>>>  		__entry->gfp_flags	= (__force unsigned long)gfp_flags;
>>>> +		__entry->accounted	= (gfp_flags & __GFP_ACCOUNT) ||
>>>> +					  (s && s->flags & SLAB_ACCOUNT);  
>>>
>>> Now you could make this even faster in the fast path and save just the
>>> s->flags.
>>>
>>> 	__entry->sflags = s ? s->flags : 0;
>>>   
>>>>  	),
>>>>  
>>>> -	TP_printk("call_site=%pS ptr=%p bytes_req=%zu bytes_alloc=%zu gfp_flags=%s",
>>>> +	TP_printk("call_site=%pS ptr=%p bytes_req=%zu bytes_alloc=%zu gfp_flags=%s accounted=%s",
>>>>  		(void *)__entry->call_site,
>>>>  		__entry->ptr,
>>>>  		__entry->bytes_req,
>>>>  		__entry->bytes_alloc,
>>>> -		show_gfp_flags(__entry->gfp_flags))
>>>> +		show_gfp_flags(__entry->gfp_flags),
>>>> +		__entry->accounted ? "true" : "false")  
>>>
>>> And then have: "accounted=%s":
>>>
>>> 		(__entry->gfp_flags & __GFP_ACCOUNT) ||
>>> 		(__entry->sflags & SLAB_ACCOUNT) ? "true" : "false"  
>>
>> Unfortunately this returns back sparse warnings about bitwise gfp_t and slab_flags_t casts.
>> Could you please explain why your variant is faster?
> 
> Micro-optimization, grant you, but it is faster because it moves some of
> the logic into the slow path (the read side), and takes it out of the fast
> path (the write side).
> 
> The idea of tracing is to squeeze out every cycle we can to keep the
> tracing overhead down.
> 
> But it's really up to you if you need that. I'm not going to let this be a
> blocker. This is more of an FYI than anything else.

Frankly speaking I vote for performance with both hands.
However I'm still would like to avoid new sparse warnings.
Christoph Hellwig just recently taught me, "never add '__force' before
thinking hard about them", but in this case I would need to use it three times.

I found that bitwise typecasts can be avoided by using translation unions. 

What do you think about following trick?

diff --git a/mm/slab.h b/mm/slab.h
index 95eb34174c1b..f676612ca40f 100644
--- a/mm/slab.h
+++ b/mm/slab.h
@@ -882,4 +882,14 @@ void __check_heap_object(const void *ptr, unsigned long n,
 }
 #endif
 
+union gfp_flags_u {
+	unsigned long ulong;
+	gfp_t flags;
+};
+
+union slab_flags_u {
+	unsigned int uint;
+	slab_flags_t sflags;
+};
+
 #endif /* MM_SLAB_H */

diff --git a/include/trace/events/kmem.h b/include/trace/events/kmem.h
index 71c141804222..91632a61e16d 100644
--- a/include/trace/events/kmem.h
+++ b/include/trace/events/kmem.h
@@ -13,18 +13,20 @@ DECLARE_EVENT_CLASS(kmem_alloc,
 
 	TP_PROTO(unsigned long call_site,
 		 const void *ptr,
+		 struct kmem_cache *s,
 		 size_t bytes_req,
 		 size_t bytes_alloc,
 		 gfp_t gfp_flags),
 
-	TP_ARGS(call_site, ptr, bytes_req, bytes_alloc, gfp_flags),
+	TP_ARGS(call_site, ptr, s, bytes_req, bytes_alloc, gfp_flags),
 
 	TP_STRUCT__entry(
 		__field(	unsigned long,	call_site	)
 		__field(	const void *,	ptr		)
 		__field(	size_t,		bytes_req	)
 		__field(	size_t,		bytes_alloc	)
-		__field(	unsigned long,	gfp_flags	)
+		__field_struct(	union gfp_flags_u,	gfp	)
+		__field_struct(	union slab_flags_u,	s	)
 	),
 
 	TP_fast_assign(
@@ -32,51 +34,57 @@ DECLARE_EVENT_CLASS(kmem_alloc,
 		__entry->ptr		= ptr;
 		__entry->bytes_req	= bytes_req;
 		__entry->bytes_alloc	= bytes_alloc;
-		__entry->gfp_flags	= (__force unsigned long)gfp_flags;
+		__entry->gfp.flags	= gfp_flags;
+		__entry->s.sflags	= s ? s->flags : 0;
 	),
 
-	TP_printk("call_site=%pS ptr=%p bytes_req=%zu bytes_alloc=%zu gfp_flags=%s",
+	TP_printk("call_site=%pS ptr=%p bytes_req=%zu bytes_alloc=%zu gfp_flags=%s accounted=%s",
 		(void *)__entry->call_site,
 		__entry->ptr,
 		__entry->bytes_req,
 		__entry->bytes_alloc,
-		show_gfp_flags(__entry->gfp_flags))
+		show_gfp_flags(__entry->gfp.ulong),
+		((__entry->gfp.flags & __GFP_ACCOUNT) ||
+		 (__entry->s.sflags & SLAB_ACCOUNT)) ? "true" : "false")
 );
 
Thank you,
	Vasily Averin