2014-04-22 05:28:48

by Jianyu Zhan

[permalink] [raw]
Subject: [PATCH] perf-event/cgroup: explicitly init the early_init field

For a cgroup subsystem who should init early, then it should carefully
take care of the implementation of css_alloc, because it will be called
before mm_init() setup the world.

Luckily we don't, and we better explicitly assign the early_init field
to 0, for document reason.

Signed-off-by: Jianyu Zhan <[email protected]>
---
kernel/events/core.c | 1 +
1 file changed, 1 insertion(+)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index f83a71a..b3a8916 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -8062,5 +8062,6 @@ struct cgroup_subsys perf_event_cgrp_subsys = {
.css_free = perf_cgroup_css_free,
.exit = perf_cgroup_exit,
.attach = perf_cgroup_attach,
+ .early_init = 0,
};
#endif /* CONFIG_CGROUP_PERF */
--
2.0.0-rc0


2014-04-22 06:07:04

by Ingo Molnar

[permalink] [raw]
Subject: Re: [PATCH] perf-event/cgroup: explicitly init the early_init field


* Jianyu Zhan <[email protected]> wrote:

> For a cgroup subsystem who should init early, then it should carefully
> take care of the implementation of css_alloc, because it will be called
> before mm_init() setup the world.
>
> Luckily we don't, and we better explicitly assign the early_init field
> to 0, for document reason.
>
> Signed-off-by: Jianyu Zhan <[email protected]>
> ---
> kernel/events/core.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index f83a71a..b3a8916 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -8062,5 +8062,6 @@ struct cgroup_subsys perf_event_cgrp_subsys = {
> .css_free = perf_cgroup_css_free,
> .exit = perf_cgroup_exit,
> .attach = perf_cgroup_attach,
> + .early_init = 0,

How can that field ever be nonzero?

I.e. under what exact circumstances does this patch make sense?

Thanks,

Ingo

2014-04-22 06:42:58

by Jianyu Zhan

[permalink] [raw]
Subject: Re: [PATCH] perf-event/cgroup: explicitly init the early_init field

On Tue, Apr 22, 2014 at 2:06 PM, Ingo Molnar <[email protected]> wrote:
> How can that field ever be nonzero?
>
> I.e. under what exact circumstances does this patch make sense?

Hi, Ingo,

For cpuset subsystem, this filed is nonzero; for other subsystems, this is zero.
Actually none of these subsystem will change this field, I just
thought that this is
a little subtle. as the comment in patches says, this is mainly for
document reason
to inform other maintainers that css_alloc implementation should be
taken care of ,
depending on this field is zero or not.

Thanks,
Jianyu Zhan

2014-04-22 07:13:31

by Jianyu Zhan

[permalink] [raw]
Subject: Re: [PATCH] perf-event/cgroup: explicitly init the early_init field

On Tue, Apr 22, 2014 at 2:06 PM, Ingo Molnar <[email protected]> wrote:
> How can that field ever be nonzero?
>
> I.e. under what exact circumstances does this patch make sense?

Hi, Ingo,

More explanation.

Sure, for this global variable struct, if not initailized, its all
fields will be initialized
to 0 or null(depending on its type). The point here is no to deprive
the rights of compiler/linker of doing this initialization, it is mainly for
documentation reason. Actually this field's value would affect how ->css_alloc
should implemented.

Concretely, if early_init is nonzero, then ->css_alloc *must not* call kzalloc,
because in cgroup implementation, ->css_alloc will be called earlier before
mm_init().

I don't think that the value of one field(early_init) has a so subtle
restrition on the another field(css_alloc) is a good thing, but since
it is there,
docment it should be needed.

I could resend the patch with more comment.

Thanks,
Jianyu Zhan

2014-04-22 09:33:10

by Zefan Li

[permalink] [raw]
Subject: Re: [PATCH] perf-event/cgroup: explicitly init the early_init field

On 2014/4/22 15:12, Jianyu Zhan wrote:
> On Tue, Apr 22, 2014 at 2:06 PM, Ingo Molnar <[email protected]> wrote:
>> How can that field ever be nonzero?
>>
>> I.e. under what exact circumstances does this patch make sense?
>
> Hi, Ingo,
>
> More explanation.
>
> Sure, for this global variable struct, if not initailized, its all
> fields will be initialized
> to 0 or null(depending on its type). The point here is no to deprive
> the rights of compiler/linker of doing this initialization, it is mainly for
> documentation reason. Actually this field's value would affect how ->css_alloc
> should implemented.
>
> Concretely, if early_init is nonzero, then ->css_alloc *must not* call kzalloc,
> because in cgroup implementation, ->css_alloc will be called earlier before
> mm_init().
>
> I don't think that the value of one field(early_init) has a so subtle
> restrition on the another field(css_alloc) is a good thing,
> but since
> it is there,
> docment it should be needed.
>
> I could resend the patch with more comment.
>

nack

As I said in another mail thread, this change makes no sense.