When a cpu belongs to a new group, there is no cpu has the same group id. This
means it can be assigned a new group id without checking with every others.
This patch does this optimiztion.
Signed-off-by: Wei Yang <[email protected]>
---
mm/percpu.c | 5 ++++-
1 files changed, 4 insertions(+), 1 deletions(-)
diff --git a/mm/percpu.c b/mm/percpu.c
index 8c8e08f..536ca4f 100644
--- a/mm/percpu.c
+++ b/mm/percpu.c
@@ -1488,7 +1488,10 @@ static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
(cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
group++;
- nr_groups = max(nr_groups, group + 1);
+ if (group == nr_groups) {
+ nr_groups++;
+ break;
+ }
goto next_group;
}
}
--
1.7.5.4
There are two loops setting up the group info of pcpu_alloc_info. They share
the same logic, so merge them could be time efficient when there are many
groups.
This patch merge these two loops into one.
Signed-off-by: Wei Yang <[email protected]>
---
mm/percpu.c | 8 +++-----
1 files changed, 3 insertions(+), 5 deletions(-)
diff --git a/mm/percpu.c b/mm/percpu.c
index 536ca4f..4f710a4f 100644
--- a/mm/percpu.c
+++ b/mm/percpu.c
@@ -1542,11 +1542,6 @@ static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
return ERR_PTR(-ENOMEM);
cpu_map = ai->groups[0].cpu_map;
- for (group = 0; group < nr_groups; group++) {
- ai->groups[group].cpu_map = cpu_map;
- cpu_map += roundup(group_cnt[group], upa);
- }
-
ai->static_size = static_size;
ai->reserved_size = reserved_size;
ai->dyn_size = dyn_size;
@@ -1557,6 +1552,8 @@ static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
for (group = 0, unit = 0; group_cnt[group]; group++) {
struct pcpu_group_info *gi = &ai->groups[group];
+ gi->cpu_map = cpu_map;
+
/*
* Initialize base_offset as if all groups are located
* back-to-back. The caller should update this to
@@ -1568,6 +1565,7 @@ static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
if (group_map[cpu] == group)
gi->cpu_map[gi->nr_units++] = cpu;
gi->nr_units = roundup(gi->nr_units, upa);
+ cpu_map += gi->nr_units;
unit += gi->nr_units;
}
BUG_ON(unit != nr_units);
--
1.7.5.4
pcpu_unit_size exactly equals to ai->unit_size.
This patch assign this value instead of calculating from pcpu_unit_pages. Also
it reorder them to make it looks more friendly to audience.
Signed-off-by: Wei Yang <[email protected]>
---
mm/percpu.c | 4 ++--
1 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/mm/percpu.c b/mm/percpu.c
index 4f710a4f..74677e0 100644
--- a/mm/percpu.c
+++ b/mm/percpu.c
@@ -1300,8 +1300,8 @@ int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
pcpu_unit_offsets = unit_off;
/* determine basic parameters */
- pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
- pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
+ pcpu_unit_size = ai->unit_size;
+ pcpu_unit_pages = pcpu_unit_size >> PAGE_SHIFT;
pcpu_atom_size = ai->atom_size;
pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
--
1.7.5.4
On Mon, Oct 21, 2013 at 04:58:11PM +0800, Wei Yang wrote:
> When a cpu belongs to a new group, there is no cpu has the same group id. This
> means it can be assigned a new group id without checking with every others.
>
> This patch does this optimiztion.
Does this actually matter? If so, it'd probably make a lot more sense
to start inner loop at @cpu + 1 so that it becomes O(N).
Thanks.
--
tejun
On Mon, Oct 21, 2013 at 04:58:12PM +0800, Wei Yang wrote:
> There are two loops setting up the group info of pcpu_alloc_info. They share
> the same logic, so merge them could be time efficient when there are many
> groups.
>
> This patch merge these two loops into one.
It *looks* correct to me but I'd rather not change this unless you can
show me this actually matters, which I find extremely doubtful given
nr_groups would be in the order of few thousands even on an extremely
large machine.
Thanks.
--
tejun
On Mon, Oct 21, 2013 at 04:58:13PM +0800, Wei Yang wrote:
> pcpu_unit_size exactly equals to ai->unit_size.
>
> This patch assign this value instead of calculating from pcpu_unit_pages. Also
> it reorder them to make it looks more friendly to audience.
Ditto. I'd rather not change unless this is clearly better.
Thanks.
--
tejun
On Sun, Oct 27, 2013 at 08:35:42AM -0400, Tejun Heo wrote:
>On Mon, Oct 21, 2013 at 04:58:12PM +0800, Wei Yang wrote:
>> There are two loops setting up the group info of pcpu_alloc_info. They share
>> the same logic, so merge them could be time efficient when there are many
>> groups.
>>
>> This patch merge these two loops into one.
>
>It *looks* correct to me but I'd rather not change this unless you can
>show me this actually matters, which I find extremely doubtful given
>nr_groups would be in the order of few thousands even on an extremely
>large machine.
Tejun, thanks for your review and comments.
I agree with you that the nr_groups won't be very large, which means it will
not bring many benefits.
This is just a small code refine. If you don't like it, just drop it :-)
>
>Thanks.
>
>--
>tejun
--
Richard Yang
Help you, Help me
On Sun, Oct 27, 2013 at 08:36:34AM -0400, Tejun Heo wrote:
>On Mon, Oct 21, 2013 at 04:58:13PM +0800, Wei Yang wrote:
>> pcpu_unit_size exactly equals to ai->unit_size.
>>
>> This patch assign this value instead of calculating from pcpu_unit_pages. Also
>> it reorder them to make it looks more friendly to audience.
>
>Ditto. I'd rather not change unless this is clearly better.
This one change an assignement to a shift, which in my mind is a little
faster.
Well, this is just executed once during the boot time, not a big deal.
>
>Thanks.
>
>--
>tejun
--
Richard Yang
Help you, Help me
On Sun, Oct 27, 2013 at 08:30:08AM -0400, Tejun Heo wrote:
>On Mon, Oct 21, 2013 at 04:58:11PM +0800, Wei Yang wrote:
>> When a cpu belongs to a new group, there is no cpu has the same group id. This
>> means it can be assigned a new group id without checking with every others.
>>
>> This patch does this optimiztion.
>
>Does this actually matter? If so, it'd probably make a lot more sense
>to start inner loop at @cpu + 1 so that it becomes O(N).
One of the worst case in my mind:
CPU: 0 1 2 3 4 ...
Group: 0 1 2 3 4 ...
(sounds it is impossible in the real world)
Every time, when we encounter a new CPU and try to assign it to a group, we
found it belongs to a new group. The original logic will iterate on all old
CPUs again, while the new logic could skip this and assign it to a new group.
Again, this is a tiny change, which doesn't matters a lot.
BTW, I don't get your point for "start inner loop at @cpu+1".
The original logic is:
loop 1: 0 - nr_cpus
loop 2: 0 - (cpu - 1)
If you found one better approach to improve the logic, I believe all the users
will appreciate your efforts :-)
Thanks for your review and comments again ~
>
>Thanks.
>
>--
>tejun
--
Richard Yang
Help you, Help me
Hello,
On Mon, Oct 28, 2013 at 11:00:55AM +0800, Wei Yang wrote:
> >Does this actually matter? If so, it'd probably make a lot more sense
> >to start inner loop at @cpu + 1 so that it becomes O(N).
>
> One of the worst case in my mind:
>
> CPU: 0 1 2 3 4 ...
> Group: 0 1 2 3 4 ...
> (sounds it is impossible in the real world)
I was wondering whether you had an actual case where this actually
matters or it's just something you thought of while reading the code.
> Every time, when we encounter a new CPU and try to assign it to a group, we
> found it belongs to a new group. The original logic will iterate on all old
> CPUs again, while the new logic could skip this and assign it to a new group.
>
> Again, this is a tiny change, which doesn't matters a lot.
I think it *could* matter because the current implementation is O(N^2)
where N is the number of CPUs. On machines, say, with 4k CPU, it's
gonna loop 16M times but then again even that takes only a few
millisecs on modern machines.
> BTW, I don't get your point for "start inner loop at @cpu+1".
>
> The original logic is:
> loop 1: 0 - nr_cpus
> loop 2: 0 - (cpu - 1)
>
> If you found one better approach to improve the logic, I believe all the users
> will appreciate your efforts :-)
Ooh, right, I forgot about the break and then I thought somehow that
would make it O(N). Sorry about that. I blame jetlag. :)
Yeah, I don't know. The function is quite hairy which makes me keep
things simpler and reluctant to make changes unless it actually makes
non-trivial difference. The change looks okay to me but it seems
neither necessary or substantially beneficial and if my experience is
anything to go by, *any* change involves some risk of brekage no
matter how innocent it may look, so given the circumstances, I'd like
to keep things the way they are.
Thanks a lot!
--
tejun
On Mon, Oct 28, 2013 at 07:31:20AM -0400, Tejun Heo wrote:
>Hello,
>
>On Mon, Oct 28, 2013 at 11:00:55AM +0800, Wei Yang wrote:
>> >Does this actually matter? If so, it'd probably make a lot more sense
>> >to start inner loop at @cpu + 1 so that it becomes O(N).
>>
>> One of the worst case in my mind:
>>
>> CPU: 0 1 2 3 4 ...
>> Group: 0 1 2 3 4 ...
>> (sounds it is impossible in the real world)
>
>I was wondering whether you had an actual case where this actually
>matters or it's just something you thought of while reading the code.
Tejun,
Thanks for your comments.
I found this just in code review. :-)
>
>> Every time, when we encounter a new CPU and try to assign it to a group, we
>> found it belongs to a new group. The original logic will iterate on all old
>> CPUs again, while the new logic could skip this and assign it to a new group.
>>
>> Again, this is a tiny change, which doesn't matters a lot.
>
>I think it *could* matter because the current implementation is O(N^2)
>where N is the number of CPUs. On machines, say, with 4k CPU, it's
>gonna loop 16M times but then again even that takes only a few
>millisecs on modern machines.
I am not familiar with the real cases of the CPU numbers. Thanks for leting me
know there could be 4K CPUs.
Yep, a few millisecs sounds not a big a mount.
>
>> BTW, I don't get your point for "start inner loop at @cpu+1".
>>
>> The original logic is:
>> loop 1: 0 - nr_cpus
>> loop 2: 0 - (cpu - 1)
>>
>> If you found one better approach to improve the logic, I believe all the users
>> will appreciate your efforts :-)
>
>Ooh, right, I forgot about the break and then I thought somehow that
>would make it O(N). Sorry about that. I blame jetlag. :)
>
>Yeah, I don't know. The function is quite hairy which makes me keep
>things simpler and reluctant to make changes unless it actually makes
>non-trivial difference. The change looks okay to me but it seems
>neither necessary or substantially beneficial and if my experience is
>anything to go by, *any* change involves some risk of brekage no
>matter how innocent it may look, so given the circumstances, I'd like
>to keep things the way they are.
Yep, I really agree with you. If no big improvement, it is really not
necessary to change the code, which will face some risk.
Here I have another one, which in my mind will improve it in one case. Looking
forward to your comments :-) If I am not correct, please let me know. :-)
>From bd70498b9df47b25ff20054e24bb510c5430c0c3 Mon Sep 17 00:00:00 2001
From: Wei Yang <[email protected]>
Date: Thu, 10 Oct 2013 09:42:14 +0800
Subject: [PATCH] percpu: optimize group assignment when cpu_distance_fn is
NULL
When cpu_distance_fn is NULL, all CPUs belongs to group 0. The original logic
will continue to go through each CPU and its predecessor. cpu_distance_fn is
always NULL when pcpu_build_alloc_info() is called from pcpu_page_first_chunk().
By applying this patch, the time complexity will drop to O(n) form O(n^2) in
case cpu_distance_fn is NULL.
Signed-off-by: Wei Yang <[email protected]>
---
mm/percpu.c | 23 ++++++++++++-----------
1 files changed, 12 insertions(+), 11 deletions(-)
diff --git a/mm/percpu.c b/mm/percpu.c
index f79c807..8e6034f 100644
--- a/mm/percpu.c
+++ b/mm/percpu.c
@@ -1481,20 +1481,21 @@ static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
for_each_possible_cpu(cpu) {
group = 0;
next_group:
- for_each_possible_cpu(tcpu) {
- if (cpu == tcpu)
- break;
- if (group_map[tcpu] == group && cpu_distance_fn &&
- (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
- cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
- group++;
- if (group == nr_groups) {
- nr_groups++;
+ if (cpu_distance_fn)
+ for_each_possible_cpu(tcpu) {
+ if (cpu == tcpu)
break;
+ if (group_map[tcpu] == group &&
+ (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
+ cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
+ group++;
+ if (group == nr_groups) {
+ nr_groups++;
+ break;
+ }
+ goto next_group;
}
- goto next_group;
}
- }
group_map[cpu] = group;
group_cnt[group]++;
}
--
1.7.5.4
BTW, this one is based on my previous patch.
>
>Thanks a lot!
>
>--
>tejun
--
Richard Yang
Help you, Help me