Re: [PATCH V2] cgroup: Trace event cgroup id fields should be u64

From: William Kucharski
Date: Wed Dec 01 2021 - 11:28:30 EST


I had pondered that as a consequence for two of the three uses but wasn't sure
if it was worth reordering things; I can easily do so.

What do you suggest for cgroup_migrate as that will have a hole either way
as it's:

TP_STRUCT__entry(
__field( int, dst_root )
__field( u64, dst_id )
__field( int, dst_level )
__field( int, pid )
__string( dst_path, path )
__string( comm, task->comm )
),

if I put dst_level above dst_id, the int for pid field will leave a hole
anyway because the string pointer for dst_path will want to be 64-bit
aligned.

Thanks in advance.

> On Dec 1, 2021, at 9:14 AM, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> On Wed, 1 Dec 2021 09:07:46 -0700
> William Kucharski <william.kucharski@xxxxxxxxxx> wrote:
>
>> Various trace event fields that store cgroup IDs were declared as
>> ints, but cgroup_id(() returns a u64 and the structures and associated
>> TP_printk() calls were not updated to reflect this.
>>
>> Fixes: 743210386c03 ("cgroup: use cgrp->kn->id as the cgroup ID")
>> Signed-off-by: William Kucharski <william.kucharski@xxxxxxxxxx>
>> ---
>> V2: Do not remove spaces from macro arguments
>>
>> include/trace/events/cgroup.h | 12 ++++++------
>> 1 file changed, 6 insertions(+), 6 deletions(-)
>>
>> diff --git a/include/trace/events/cgroup.h b/include/trace/events/cgroup.h
>> index 7f42a3de59e6..1b68c842ac46 100644
>> --- a/include/trace/events/cgroup.h
>> +++ b/include/trace/events/cgroup.h
>> @@ -59,7 +59,7 @@ DECLARE_EVENT_CLASS(cgroup,
>>
>> TP_STRUCT__entry(
>> __field( int, root )
>> - __field( int, id )
>> + __field( u64, id )
>> __field( int, level )
>
> It's best to move them around to prevent holes. This is not a packed
> structure, and the above will create a structure on the ring buffer that
> looks like:
>
> int root;
> u64 id;
> int level;
> short path_offset;
> short path_size;
>
> (the string() macro is a 4 byte word where the half is the offset of the
> actual string and the other half is the size of the string).
>
> Having a 8 byte word between two 4 byte words, will add a 4 byte padding
> after the first 4 byte word (before the 8 byte word), and waste space on
> the ring buffer.
>
> Better to move it around to be:
>
> TP_STRUCT__entry(
> __field( int, root )
> __field( int, level )
> __field( u64, id )
>
> That way the two 4 byte words will be together followed directly by the 8
> byte word.
>
>> __string( path, path )
>> ),
>> @@ -71,7 +71,7 @@ DECLARE_EVENT_CLASS(cgroup,
>> __assign_str(path, path);
>> ),
>>
>> - TP_printk("root=%d id=%d level=%d path=%s",
>> + TP_printk("root=%d id=%llu level=%d path=%s",
>> __entry->root, __entry->id, __entry->level, __get_str(path))
>> );
>>
>> @@ -126,7 +126,7 @@ DECLARE_EVENT_CLASS(cgroup_migrate,
>>
>> TP_STRUCT__entry(
>> __field( int, dst_root )
>> - __field( int, dst_id )
>> + __field( u64, dst_id )
>
> Same here. Just move dst_level above dst_id.
>
>> __field( int, dst_level )
>> __field( int, pid )
>> __string( dst_path, path )
>> @@ -142,7 +142,7 @@ DECLARE_EVENT_CLASS(cgroup_migrate,
>> __assign_str(comm, task->comm);
>> ),
>>
>> - TP_printk("dst_root=%d dst_id=%d dst_level=%d dst_path=%s pid=%d comm=%s",
>> + TP_printk("dst_root=%d dst_id=%llu dst_level=%d dst_path=%s pid=%d comm=%s",
>> __entry->dst_root, __entry->dst_id, __entry->dst_level,
>> __get_str(dst_path), __entry->pid, __get_str(comm))
>> );
>> @@ -171,7 +171,7 @@ DECLARE_EVENT_CLASS(cgroup_event,
>>
>> TP_STRUCT__entry(
>> __field( int, root )
>> - __field( int, id )
>> + __field( u64, id )
>
> And here.
>
>> __field( int, level )
>> __string( path, path )
>> __field( int, val )
>> @@ -185,7 +185,7 @@ DECLARE_EVENT_CLASS(cgroup_event,
>> __entry->val = val;
>> ),
>>
>> - TP_printk("root=%d id=%d level=%d path=%s val=%d",
>> + TP_printk("root=%d id=%llu level=%d path=%s val=%d",
>> __entry->root, __entry->id, __entry->level, __get_str(path),
>> __entry->val)
>> );
>
>
>
> -- Steve