2010-01-18 16:23:52

by Jiri Pirko

[permalink] [raw]
Subject: [PATCH] list.h: add list_for_each_struct_entry macro

When preparing to upcoming migration of mc_list (list of multicast MACs)
to list_head, the need of traversing the list over one structure member
appeared. I thought I'll do it locally but I decided that an intruduction
the macro in list.h would be much clearer. Please kindly for a review.

Jirka

Signed-off-by: Jiri Pirko <[email protected]>

diff --git a/include/linux/list.h b/include/linux/list.h
index 969f6e9..8350a94 100644
--- a/include/linux/list.h
+++ b/include/linux/list.h
@@ -420,6 +420,22 @@ static inline void list_splice_tail_init(struct list_head *list,
pos = list_entry(pos->member.prev, typeof(*pos), member))

/**
+ * list_for_each_struct_entry - iterate over list of given type using
+ * the struct member.
+ * @pos: the type * to use as a loop cursor.
+ * @head: the head for your list.
+ * @type: the type of the struct.
+ * @posmember: the name ot the loop cursor within the struct.
+ * @member: the name of the list_struct within the struct.
+ */
+#define list_for_each_struct_entry(pos, head, type, posmember, member) \
+ for (pos = list_entry((head)->next, type, member)->posmember; \
+ prefetch(container_of(pos, type, posmember)->member.next), \
+ &container_of(pos, type, posmember)->member != (head); \
+ pos = list_entry(container_of(pos, type, posmember)->member.next, \
+ type, member)->posmember)
+
+/**
* list_prepare_entry - prepare a pos entry for use in list_for_each_entry_continue()
* @pos: the type * to use as a start point
* @head: the head of the list


2010-01-18 17:17:27

by Ben Hutchings

[permalink] [raw]
Subject: Re: [PATCH] list.h: add list_for_each_struct_entry macro

On Mon, 2010-01-18 at 17:23 +0100, Jiri Pirko wrote:
> When preparing to upcoming migration of mc_list (list of multicast MACs)
> to list_head, the need of traversing the list over one structure member
> appeared. I thought I'll do it locally but I decided that an intruduction
> the macro in list.h would be much clearer. Please kindly for a review.
>
> Jirka
>
> Signed-off-by: Jiri Pirko <[email protected]>
>
> diff --git a/include/linux/list.h b/include/linux/list.h
> index 969f6e9..8350a94 100644
> --- a/include/linux/list.h
> +++ b/include/linux/list.h
> @@ -420,6 +420,22 @@ static inline void list_splice_tail_init(struct list_head *list,
> pos = list_entry(pos->member.prev, typeof(*pos), member))
>
> /**
> + * list_for_each_struct_entry - iterate over list of given type using
> + * the struct member.
> + * @pos: the type * to use as a loop cursor.
> + * @head: the head for your list.
> + * @type: the type of the struct.
> + * @posmember: the name ot the loop cursor within the struct.
> + * @member: the name of the list_struct within the struct.
> + */
> +#define list_for_each_struct_entry(pos, head, type, posmember, member) \
> + for (pos = list_entry((head)->next, type, member)->posmember; \

Surely &list_entry(...)->posmember ?

> + prefetch(container_of(pos, type, posmember)->member.next), \
> + &container_of(pos, type, posmember)->member != (head); \

It seems a bit dodgy to do this pointer arithmetic on a list node which
might be the list head.

> + pos = list_entry(container_of(pos, type, posmember)->member.next, \
> + type, member)->posmember)

My version:

#define list_for_each_struct_entry(pos, head, type, posmember, member) \
for (pos = list_empty(head) ? NULL : \
&list_first_entry(head, type, member)->posmember; \
prefetch(container_of(pos, type, posmember)->member.next), pos; \
pos = list_is_last(&container_of(pos, type, posmember)->member, \
head) ? NULL : \
&list_entry(container_of(pos, type, posmember)->member.next, \
type, member)->posmember)

Ben.

--
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

2010-01-18 21:07:32

by Jiri Pirko

[permalink] [raw]
Subject: Re: [PATCH] list.h: add list_for_each_struct_entry macro

Mon, Jan 18, 2010 at 06:17:13PM CET, [email protected] wrote:
>On Mon, 2010-01-18 at 17:23 +0100, Jiri Pirko wrote:
>> When preparing to upcoming migration of mc_list (list of multicast MACs)
>> to list_head, the need of traversing the list over one structure member
>> appeared. I thought I'll do it locally but I decided that an intruduction
>> the macro in list.h would be much clearer. Please kindly for a review.
>>
>> Jirka
>>
>> Signed-off-by: Jiri Pirko <[email protected]>
>>
>> diff --git a/include/linux/list.h b/include/linux/list.h
>> index 969f6e9..8350a94 100644
>> --- a/include/linux/list.h
>> +++ b/include/linux/list.h
>> @@ -420,6 +420,22 @@ static inline void list_splice_tail_init(struct list_head *list,
>> pos = list_entry(pos->member.prev, typeof(*pos), member))
>>
>> /**
>> + * list_for_each_struct_entry - iterate over list of given type using
>> + * the struct member.
>> + * @pos: the type * to use as a loop cursor.
>> + * @head: the head for your list.
>> + * @type: the type of the struct.
>> + * @posmember: the name ot the loop cursor within the struct.
>> + * @member: the name of the list_struct within the struct.
>> + */
>> +#define list_for_each_struct_entry(pos, head, type, posmember, member) \
>> + for (pos = list_entry((head)->next, type, member)->posmember; \
>
>Surely &list_entry(...)->posmember ?

Ugh, right, I tested this with char * so it worked well. Got i corrected now so
I can repost.

>
>> + prefetch(container_of(pos, type, posmember)->member.next), \
>> + &container_of(pos, type, posmember)->member != (head); \
>
>It seems a bit dodgy to do this pointer arithmetic on a list node which
>might be the list head.

Well I tried to be as much close to other traverse macros as it could be.

>
>> + pos = list_entry(container_of(pos, type, posmember)->member.next, \
>> + type, member)->posmember)
>
>My version:
>
>#define list_for_each_struct_entry(pos, head, type, posmember, member) \
> for (pos = list_empty(head) ? NULL : \
> &list_first_entry(head, type, member)->posmember; \
> prefetch(container_of(pos, type, posmember)->member.next), pos; \
> pos = list_is_last(&container_of(pos, type, posmember)->member, \
> head) ? NULL : \
> &list_entry(container_of(pos, type, posmember)->member.next, \
> type, member)->posmember)
>

At the first glance, this would take even more cputime for lists longer
than 2 or so, wouldn't it?


Jirka

>Ben.
>
>--
>Ben Hutchings, Senior Software Engineer, Solarflare Communications
>Not speaking for my employer; that's the marketing department's job.
>They asked us to note that Solarflare product names are trademarked.
>

2010-01-18 21:15:05

by Ben Hutchings

[permalink] [raw]
Subject: Re: [PATCH] list.h: add list_for_each_struct_entry macro

On Mon, 2010-01-18 at 22:07 +0100, Jiri Pirko wrote:
> Mon, Jan 18, 2010 at 06:17:13PM CET, [email protected] wrote:
[...]
> >#define list_for_each_struct_entry(pos, head, type, posmember, member) \
> > for (pos = list_empty(head) ? NULL : \
> > &list_first_entry(head, type, member)->posmember; \
> > prefetch(container_of(pos, type, posmember)->member.next), pos; \
> > pos = list_is_last(&container_of(pos, type, posmember)->member, \
> > head) ? NULL : \
> > &list_entry(container_of(pos, type, posmember)->member.next, \
> > type, member)->posmember)
> >
>
> At the first glance, this would take even more cputime for lists longer
> than 2 or so, wouldn't it?

If you're concerned about speed, measure it, don't guess.

Ben.

--
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

2010-01-18 21:30:52

by Jiri Pirko

[permalink] [raw]
Subject: Re: [PATCH] list.h: add list_for_each_struct_entry macro

Mon, Jan 18, 2010 at 10:14:18PM CET, [email protected] wrote:
>On Mon, 2010-01-18 at 22:07 +0100, Jiri Pirko wrote:
>> Mon, Jan 18, 2010 at 06:17:13PM CET, [email protected] wrote:
>[...]
>> >#define list_for_each_struct_entry(pos, head, type, posmember, member) \
>> > for (pos = list_empty(head) ? NULL : \
>> > &list_first_entry(head, type, member)->posmember; \
>> > prefetch(container_of(pos, type, posmember)->member.next), pos; \
>> > pos = list_is_last(&container_of(pos, type, posmember)->member, \
>> > head) ? NULL : \
>> > &list_entry(container_of(pos, type, posmember)->member.next, \
>> > type, member)->posmember)
>> >
>>
>> At the first glance, this would take even more cputime for lists longer
>> than 2 or so, wouldn't it?
>
>If you're concerned about speed, measure it, don't guess.

Well I just see extra code
"list_is_last(&container_of(pos, type, posmember)->member, head)"
to be done in each iteration. Also I do not see additional value in doing this.
(Unlike in checking list_empty(head)).

Anyway, if you want to use this optimization, I guess more code in list.h could
use this.

Jirka
>
>Ben.
>
>--
>Ben Hutchings, Senior Software Engineer, Solarflare Communications
>Not speaking for my employer; that's the marketing department's job.
>They asked us to note that Solarflare product names are trademarked.
>