Skip to contentSkip to conversation

Q: (stupid) can't we "fix" hlist_for_each_entry() ?

DISCUSSION Oleg Nesterov 4 participants 4 messages Mar 12, 2008

lore
Share & tools
lore
#0 lore raw

hlist_for_each_entry/hlist_for_each_entry_rcu doesn't actually need 4 arguments, it could be

#define hlist_for_each_entry_rcu(pos, head, member)		 			\
	for (pos = hlist_entry((head)->first, typeof(*(pos)), member);			\
	     rcu_dereference(pos) != hlist_entry(NULL, typeof(*(pos)), member) &&	\
		({ prefetch((pos)->member.next); 1; });					\
	     (pos) = hlist_entry((pos)->member.next, typeof(*(pos)), member))

Or,

#define hlist_for_each_entry_rcu(pos, head, member)		 			\
	for (pos = (void*)(head)->first;						\
	     rcu_dereference(pos) && ({ prefetch(((hlist_node*)pos)->next); 1; }) &&	\
		({ (pos) = hlist_entry((void*)(pos), typeof(*(pos)), member)); 1; });	\
	     (pos) = (void*)(pos)->member.next)

Q: is it worth "fixing" ?

If yes, what is the "right" way to do this? These macros are spread all over the kernel...

Oleg.

#1 lore raw
Quoted 11 lines · Oleg Nesterov

hlist_for_each_entry/hlist_for_each_entry_rcu doesn't actually need 4 arguments, it could be

#define hlist_for_each_entry_rcu(pos, head, member)		 			\
	for (pos = hlist_entry((head)->first, typeof(*(pos)), member);			\
	     rcu_dereference(pos) != hlist_entry(NULL, typeof(*(pos)), member) &&	\
		({ prefetch((pos)->member.next); 1; });					\
	     (pos) = hlist_entry((pos)->member.next, typeof(*(pos)), member))

Or,

Quoted 7 lines
#define hlist_for_each_entry_rcu(pos, head, member)		 			\
	for (pos = (void*)(head)->first;						\
	     rcu_dereference(pos) && ({ prefetch(((hlist_node*)pos)->next); 1; }) &&	\
		({ (pos) = hlist_entry((void*)(pos), typeof(*(pos)), member)); 1; });	\
	     (pos) = (void*)(pos)->member.next)

Q: is it worth "fixing" ?

I'm in favour.

Quoted 2 lines

If yes, what is the "right" way to do this? These macros are spread all over the kernel...

The usual way would be to prepare a git tree for Linus to pull right after -rc1 I think was the best point, and at the same time supply Andrew with a bunch of patches fixing up the various users in his tree.

#2 lore raw
Quoted 20 lines · On Wed, 12 Mar 2008 10:48:50 +0100 Peter Zijlstra <[email protected]> wrote:
Quoted 11 lines · Oleg Nesterov

hlist_for_each_entry/hlist_for_each_entry_rcu doesn't actually need 4 arguments, it could be

#define hlist_for_each_entry_rcu(pos, head, member)		 			\
	for (pos = hlist_entry((head)->first, typeof(*(pos)), member);			\
	     rcu_dereference(pos) != hlist_entry(NULL, typeof(*(pos)), member) &&	\
		({ prefetch((pos)->member.next); 1; });					\
	     (pos) = hlist_entry((pos)->member.next, typeof(*(pos)), member))

Or,

Quoted 7 lines
#define hlist_for_each_entry_rcu(pos, head, member)		 			\
	for (pos = (void*)(head)->first;						\
	     rcu_dereference(pos) && ({ prefetch(((hlist_node*)pos)->next); 1; }) &&	\
		({ (pos) = hlist_entry((void*)(pos), typeof(*(pos)), member)); 1; });	\
	     (pos) = (void*)(pos)->member.next)

Q: is it worth "fixing" ?

Quoted 8 lines

I'm in favour.

Quoted 2 lines

If yes, what is the "right" way to do this? These macros are spread all over the kernel...

The usual way would be to prepare a git tree for Linus to pull right after -rc1 I think was the best point, and at the same time supply Andrew with a bunch of patches fixing up the various users in his tree.

That, or create new functions with new names, migrate over to them piecemeal and later remove the old functions.

It's a bit of a problem that there is no alternative name.

eh, send over the jumbopatch after 2.6.25 is released - I'll take care of it.

#3 lore raw
In reply to Oleg Nesterov
Quoted 11 lines · Oleg Nesterov

hlist_for_each_entry/hlist_for_each_entry_rcu doesn't actually need 4 arguments, it could be

#define hlist_for_each_entry_rcu(pos, head, member)		 			\
	for (pos = hlist_entry((head)->first, typeof(*(pos)), member);			\
	     rcu_dereference(pos) != hlist_entry(NULL, typeof(*(pos)), member) &&	\
		({ prefetch((pos)->member.next); 1; });					\
	     (pos) = hlist_entry((pos)->member.next, typeof(*(pos)), member))

Or,

Quoted 7 lines
#define hlist_for_each_entry_rcu(pos, head, member)		 			\
	for (pos = (void*)(head)->first;						\
	     rcu_dereference(pos) && ({ prefetch(((hlist_node*)pos)->next); 1; }) &&	\
		({ (pos) = hlist_entry((void*)(pos), typeof(*(pos)), member)); 1; });	\
	     (pos) = (void*)(pos)->member.next)

Q: is it worth "fixing" ?

I have already come out in favor: http://lwn.net/Articles/262464/
answer to Quick Quiz 3. ;-)

The first option above looks more straightforward to me.

Quoted 2 lines

If yes, what is the "right" way to do this? These macros are spread all over the kernel...

Peter's approach looked reasonable to me.

Thanx, Paul