In function cgroup_migrate_add_src(), if dst_cgrp
equals to src_cgroup which the tasks link, dont
go on migrating tasks to another css_set.
This can save the cost of unnecessary migration.
Signed-off-by: shisiyuan <[email protected]>
---
kernel/cgroup/cgroup.c | 11 ++++++++++-
1 file changed, 10 insertions(+), 1 deletion(-)
diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
index 6139460..97d7f68 100644
--- a/kernel/cgroup/cgroup.c
+++ b/kernel/cgroup/cgroup.c
@@ -2634,6 +2634,9 @@ void cgroup_migrate_add_src(struct css_set *src_cset,
src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
+ if (src_cgrp == dst_cgrp)
+ return;
+
if (!list_empty(&src_cset->mg_preload_node))
return;
@@ -2780,6 +2783,9 @@ int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
rcu_read_unlock();
spin_unlock_irq(&css_set_lock);
+ if (list_empty(&mgctx.preloaded_src_csets))
+ return ret;
+
/* prepare dst csets and commit */
ret = cgroup_migrate_prepare_dst(&mgctx);
if (!ret)
@@ -2927,7 +2933,7 @@ static int cgroup_update_dfl_csses(struct cgroup *cgrp)
struct cgroup_subsys_state *d_css;
struct cgroup *dsct;
struct css_set *src_cset;
- int ret;
+ int ret = 0;
lockdep_assert_held(&cgroup_mutex);
@@ -2943,6 +2949,9 @@ static int cgroup_update_dfl_csses(struct cgroup *cgrp)
}
spin_unlock_irq(&css_set_lock);
+ if (list_empty(&mgctx.preloaded_src_csets))
+ goto out_finish;
+
/* NULL dst indicates self on default hierarchy */
ret = cgroup_migrate_prepare_dst(&mgctx);
if (ret)
--
2.7.4
Hello.
On Sat, May 21, 2022 at 01:02:00AM +0800, shisiyuan <[email protected]> wrote:
> In function cgroup_migrate_add_src(), if dst_cgrp
> equals to src_cgroup which the tasks link, dont
> go on migrating tasks to another css_set.
Do you have an example which operations lead to this situation?
> This can save the cost of unnecessary migration.
How much are the savings visible?
(Note, I suspect the shortcut on `src_cgrp == dst_cgrp` might not be
always correct when called from cgroup_update_dfl_csses() but not 100%
sure now.)
Thanks,
Michal
On Thu, May 26, 2022 at 12:40:17PM +0800, 史思远 <[email protected]> wrote:
> echo 1413 > /dev/cpuset/foreground/cgroup.procs
> echo 1413 > /dev/cpuset/foreground/cgroup.procs
> echo 1413 > /dev/cpuset/foreground/cgroup.procs
Thank you.
> In this situation, since the second echo, dst_cgrp equals to src_cgrp(both
> cgroups are the 'foreground' cpuset).
> cgroup1_procs_write（）-> __cgroup1_procs_write()
> -> cgrp = cgroup_kn_lock_live()
> -> task = cgroup_procs_write_start(buf...)
> -> cgroup_attach_task(cgrp, task, threadgroup);
> -> 1) cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx); //
> detect the (src_cgrp == dst_cgrp), don't need to continue the following
> migration flow.
> -> 2) cgroup_migrate_prepare_dst(&mgctx)
Here, I'd like to let you know about [1]
> -> 3) cgroup_migrate(leader, threadgroup, &mgctx)
> -> 4) cgroup_migrate_finish(&mgctx) // cleanup
> -> cgroup_procs_write_finish(); // cleanup
> >> The main purpose for this patch is to handle the situation above, and
> just noticed that cgroup_update_dfl_csses() also
> calls cgroup_migrate_add_src(). So I assume it's a similar situation. We
> can discuss more about it.
So I think the patch doesn't save that much and identical css_set is
better condition to check on (covers cgroup_update_dfl_csses too).
Michal
[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/cgroup/cgroup.c?h=v5.18#n2710