[PATCH 1/3] cgroup: remove tasklist_lock from cgroup_attach_proc

[PATCH 1/3] cgroup: remove tasklist_lock from cgroup_attach_proc

[Mandeep Singh Baines]

Since cgroup_attach_proc is protected by a threadgroup_lock, we
no longer need a tasklist_lock to protect while_each_thread.
To keep the complexity of the double-check locking in one place,
I also moved the thread_group_leader check up into
attach_task_by_pid.

While at it, also converted a couple of returns to gotos.

The suggestion was made here:

https://lkml.org/lkml/2011/12/22/86

Suggested-by: Frederic Weisbecker <fweisbec-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
Signed-off-by: Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
Cc: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Cc: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>
Cc: containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
Cc: cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
Cc: Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
Cc: Paul Menage <paul-inf54ven1CmVyaH7bEyXVA@public.gmane.org>
---
 kernel/cgroup.c |   52 +++++++++++++++++++++-------------------------------
 1 files changed, 21 insertions(+), 31 deletions(-)

diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index 1042b3c..032139d 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c
@@ -2102,21 +2102,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 	if (retval)
 		goto out_free_group_list;
 
-	/* prevent changes to the threadgroup list while we take a snapshot. */
-	read_lock(&tasklist_lock);
-	if (!thread_group_leader(leader)) {
-		/*
-		 * a race with de_thread from another thread's exec() may strip
-		 * us of our leadership, making while_each_thread unsafe to use
-		 * on this task. if this happens, there is no choice but to
-		 * throw this task away and try again (from cgroup_procs_write);
-		 * this is "double-double-toil-and-trouble-check locking".
-		 */
-		read_unlock(&tasklist_lock);
-		retval = -EAGAIN;
-		goto out_free_group_list;
-	}
-
 	tsk = leader;
 	i = 0;
 	do {
@@ -2145,7 +2130,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 	group_size = i;
 	tset.tc_array = group;
 	tset.tc_array_len = group_size;
-	read_unlock(&tasklist_lock);
 
 	/* methods shouldn't be called if no task is actually migrating */
 	retval = 0;
@@ -2247,18 +2231,12 @@ static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
 		tsk = find_task_by_vpid(pid);
 		if (!tsk) {
 			rcu_read_unlock();
-			cgroup_unlock();
-			return -ESRCH;
+			ret= -ESRCH;
+			goto out_unlock_cgroup;
 		}
-		if (threadgroup) {
-			/*
-			 * RCU protects this access, since tsk was found in the
-			 * tid map. a race with de_thread may cause group_leader
-			 * to stop being the leader, but cgroup_attach_proc will
-			 * detect it later.
-			 */
+		/* we check later for a group_leader race with de_thread */
+		if (threadgroup)
 			tsk = tsk->group_leader;
-		}
 		/*
 		 * even if we're attaching all tasks in the thread group, we
 		 * only need to check permissions on one of them.
@@ -2268,8 +2246,8 @@ static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
 		    cred->euid != tcred->uid &&
 		    cred->euid != tcred->suid) {
 			rcu_read_unlock();
-			cgroup_unlock();
-			return -EACCES;
+			ret = -EACCES;
+			goto out_unlock_cgroup;
 		}
 		get_task_struct(tsk);
 		rcu_read_unlock();
@@ -2283,14 +2261,26 @@ static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
 
 	threadgroup_lock(tsk);
 
-	if (threadgroup)
+	if (threadgroup) {
+		if (!thread_group_leader(tsk)) {
+			/*
+			 * a race with de_thread from another thread's exec()
+			 * may strip us of our leadership, if this happens,
+			 * there is no choice but to throw this task away and
+			 * try again (from cgroup_procs_write); this is
+			 * "double-double-toil-and-trouble-check locking".
+			 */
+			ret = -EAGAIN;
+			goto out_unlock_threadgroup;
+		}
 		ret = cgroup_attach_proc(cgrp, tsk);
-	else
+	} else
 		ret = cgroup_attach_task(cgrp, tsk);
 
+out_unlock_threadgroup:
 	threadgroup_unlock(tsk);
-
 	put_task_struct(tsk);
+out_unlock_cgroup:
 	cgroup_unlock();
 	return ret;
 }
-- 
1.7.3.1

[PATCH 2/3] cgroup: remove double-checking locking from attach_task_by_pid

[Mandeep Singh Baines]

By reading group_leader after taking the threadgroup_lock, we can
avoid the double-check locking. This removes the return of
-EAGAIN so we can cleanup cgroup_procs_write at the same time.

The suggestion was made here:

https://lkml.org/lkml/2011/12/22/371

Suggested-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
Cc: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>
Cc: containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
Cc: cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
Cc: Frederic Weisbecker <fweisbec-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
Cc: Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
Cc: Paul Menage <paul-inf54ven1CmVyaH7bEyXVA@public.gmane.org>
---
 kernel/cgroup.c |   40 ++++++----------------------------------
 1 files changed, 6 insertions(+), 34 deletions(-)

diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index 032139d..a5f7d1b 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c
@@ -2234,9 +2234,6 @@ static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
 			ret= -ESRCH;
 			goto out_unlock_cgroup;
 		}
-		/* we check later for a group_leader race with de_thread */
-		if (threadgroup)
-			tsk = tsk->group_leader;
 		/*
 		 * even if we're attaching all tasks in the thread group, we
 		 * only need to check permissions on one of them.
@@ -2252,33 +2249,17 @@ static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
 		get_task_struct(tsk);
 		rcu_read_unlock();
 	} else {
-		if (threadgroup)
-			tsk = current->group_leader;
-		else
-			tsk = current;
+		tsk = current;
 		get_task_struct(tsk);
 	}
 
 	threadgroup_lock(tsk);
-
-	if (threadgroup) {
-		if (!thread_group_leader(tsk)) {
-			/*
-			 * a race with de_thread from another thread's exec()
-			 * may strip us of our leadership, if this happens,
-			 * there is no choice but to throw this task away and
-			 * try again (from cgroup_procs_write); this is
-			 * "double-double-toil-and-trouble-check locking".
-			 */
-			ret = -EAGAIN;
-			goto out_unlock_threadgroup;
-		}
-		ret = cgroup_attach_proc(cgrp, tsk);
-	} else
+	if (threadgroup)
+		ret = cgroup_attach_proc(cgrp, tsk->group_leader);
+	else
 		ret = cgroup_attach_task(cgrp, tsk);
-
-out_unlock_threadgroup:
 	threadgroup_unlock(tsk);
+
 	put_task_struct(tsk);
 out_unlock_cgroup:
 	cgroup_unlock();
@@ -2292,16 +2273,7 @@ static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
 
 static int cgroup_procs_write(struct cgroup *cgrp, struct cftype *cft, u64 tgid)
 {
-	int ret;
-	do {
-		/*
-		 * attach_proc fails with -EAGAIN if threadgroup leadership
-		 * changes in the middle of the operation, in which case we need
-		 * to find the task_struct for the new leader and start over.
-		 */
-		ret = attach_task_by_pid(cgrp, tgid, true);
-	} while (ret == -EAGAIN);
-	return ret;
+	return attach_task_by_pid(cgrp, tgid, true);
 }
 
 /**
-- 
1.7.3.1

[PATCH 3/3] cgroup: remove extra calls to find_existing_css_set

[Mandeep Singh Baines]

In cgroup_attach_proc, we indirectly call find_existing_css_set 3
times. It is an expensive call so we want to call it a minimum
of times. This patch only calls it once and stores the result so
that it can be used later on when we call cgroup_task_migrate.

This required modifying cgroup_task_migrate to take the new css_set
(which we obtained from find_css_set) as a parameter. The nice side
effect of this is that cgroup_task_migrate is now identical for
cgroup_attach_task and cgroup_attach_proc. It also now returns a
void since it can never fail.

Changes in V3:
* https://lkml.org/lkml/2011/12/22/13 (Li Zefan)
  * Fixed earlier bug by creating a seperate patch to remove tasklist_lock
Changes in V2:
* https://lkml.org/lkml/2011/12/20/372 (Tejun Heo)
  * Move find_css_set call into loop which creates the flex array
* Author
  * Kill css_set_refs and use group_size instead
  * Fix an off-by-one error in counting css_set refs
  * Add a retval check in out_list_teardown

Signed-off-by: Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
Cc: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Cc: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>
Cc: containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
Cc: cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
Cc: Frederic Weisbecker <fweisbec-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
Cc: Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
Cc: Paul Menage <paul-inf54ven1CmVyaH7bEyXVA@public.gmane.org>
---
 kernel/cgroup.c |  152 ++++++++++++-------------------------------------------
 1 files changed, 32 insertions(+), 120 deletions(-)

diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index a5f7d1b..0cb0489 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c
@@ -1763,6 +1763,7 @@ EXPORT_SYMBOL_GPL(cgroup_path);
 struct task_and_cgroup {
 	struct task_struct	*task;
 	struct cgroup		*cgrp;
+	struct css_set		*cg;
 };
 
 struct cgroup_taskset {
@@ -1843,11 +1844,10 @@ EXPORT_SYMBOL_GPL(cgroup_taskset_size);
  * will already exist. If not set, this function might sleep, and can fail with
  * -ENOMEM. Must be called with cgroup_mutex and threadgroup locked.
  */
-static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
-			       struct task_struct *tsk, bool guarantee)
+static void cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
+				struct task_struct *tsk, struct css_set *newcg)
 {
 	struct css_set *oldcg;
-	struct css_set *newcg;
 
 	/*
 	 * We are synchronized through threadgroup_lock() against PF_EXITING
@@ -1857,23 +1857,6 @@ static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
 	WARN_ON_ONCE(tsk->flags & PF_EXITING);
 	oldcg = tsk->cgroups;
 
-	/* locate or allocate a new css_set for this task. */
-	if (guarantee) {
-		/* we know the css_set we want already exists. */
-		struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
-		read_lock(&css_set_lock);
-		newcg = find_existing_css_set(oldcg, cgrp, template);
-		BUG_ON(!newcg);
-		get_css_set(newcg);
-		read_unlock(&css_set_lock);
-	} else {
-		might_sleep();
-		/* find_css_set will give us newcg already referenced. */
-		newcg = find_css_set(oldcg, cgrp);
-		if (!newcg)
-			return -ENOMEM;
-	}
-
 	task_lock(tsk);
 	rcu_assign_pointer(tsk->cgroups, newcg);
 	task_unlock(tsk);
@@ -1892,7 +1875,6 @@ static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
 	put_css_set(oldcg);
 
 	set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
-	return 0;
 }
 
 /**
@@ -1910,6 +1892,7 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
 	struct cgroup *oldcgrp;
 	struct cgroupfs_root *root = cgrp->root;
 	struct cgroup_taskset tset = { };
+	struct css_set *newcg;
 
 	/* @tsk either already exited or can't exit until the end */
 	if (tsk->flags & PF_EXITING)
@@ -1939,9 +1922,13 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
 		}
 	}
 
-	retval = cgroup_task_migrate(cgrp, oldcgrp, tsk, false);
-	if (retval)
+	newcg = find_css_set(tsk->cgroups, cgrp);
+	if (!newcg) {
+		retval = -ENOMEM;
 		goto out;
+	}
+
+	cgroup_task_migrate(cgrp, oldcgrp, tsk, newcg);
 
 	for_each_subsys(root, ss) {
 		if (ss->attach)
@@ -1997,66 +1984,6 @@ int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
 }
 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
 
-/*
- * cgroup_attach_proc works in two stages, the first of which prefetches all
- * new css_sets needed (to make sure we have enough memory before committing
- * to the move) and stores them in a list of entries of the following type.
- * TODO: possible optimization: use css_set->rcu_head for chaining instead
- */
-struct cg_list_entry {
-	struct css_set *cg;
-	struct list_head links;
-};
-
-static bool css_set_check_fetched(struct cgroup *cgrp,
-				  struct task_struct *tsk, struct css_set *cg,
-				  struct list_head *newcg_list)
-{
-	struct css_set *newcg;
-	struct cg_list_entry *cg_entry;
-	struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
-
-	read_lock(&css_set_lock);
-	newcg = find_existing_css_set(cg, cgrp, template);
-	read_unlock(&css_set_lock);
-
-	/* doesn't exist at all? */
-	if (!newcg)
-		return false;
-	/* see if it's already in the list */
-	list_for_each_entry(cg_entry, newcg_list, links)
-		if (cg_entry->cg == newcg)
-			return true;
-
-	/* not found */
-	return false;
-}
-
-/*
- * Find the new css_set and store it in the list in preparation for moving the
- * given task to the given cgroup. Returns 0 or -ENOMEM.
- */
-static int css_set_prefetch(struct cgroup *cgrp, struct css_set *cg,
-			    struct list_head *newcg_list)
-{
-	struct css_set *newcg;
-	struct cg_list_entry *cg_entry;
-
-	/* ensure a new css_set will exist for this thread */
-	newcg = find_css_set(cg, cgrp);
-	if (!newcg)
-		return -ENOMEM;
-	/* add it to the list */
-	cg_entry = kmalloc(sizeof(struct cg_list_entry), GFP_KERNEL);
-	if (!cg_entry) {
-		put_css_set(newcg);
-		return -ENOMEM;
-	}
-	cg_entry->cg = newcg;
-	list_add(&cg_entry->links, newcg_list);
-	return 0;
-}
-
 /**
  * cgroup_attach_proc - attach all threads in a threadgroup to a cgroup
  * @cgrp: the cgroup to attach to
@@ -2070,20 +1997,12 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 	int retval, i, group_size;
 	struct cgroup_subsys *ss, *failed_ss = NULL;
 	/* guaranteed to be initialized later, but the compiler needs this */
-	struct css_set *oldcg;
 	struct cgroupfs_root *root = cgrp->root;
 	/* threadgroup list cursor and array */
 	struct task_struct *tsk;
 	struct task_and_cgroup *tc;
 	struct flex_array *group;
 	struct cgroup_taskset tset = { };
-	/*
-	 * we need to make sure we have css_sets for all the tasks we're
-	 * going to move -before- we actually start moving them, so that in
-	 * case we get an ENOMEM we can bail out before making any changes.
-	 */
-	struct list_head newcg_list;
-	struct cg_list_entry *cg_entry, *temp_nobe;
 
 	/*
 	 * step 0: in order to do expensive, possibly blocking operations for
@@ -2091,6 +2010,10 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 	 * rcu or tasklist locked. instead, build an array of all threads in the
 	 * group - group_rwsem prevents new threads from appearing, and if
 	 * threads exit, this will just be an over-estimate.
+	 *
+	 * While creating the list, also make sure css_sets exist for all
+	 * threads to be migrated. we use find_css_set, which allocates a new
+	 * one if necessary.
 	 */
 	group_size = get_nr_threads(leader);
 	/* flex_array supports very large thread-groups better than kmalloc. */
@@ -2122,6 +2045,12 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 		/* nothing to do if this task is already in the cgroup */
 		if (ent.cgrp == cgrp)
 			continue;
+		ent.cg = find_css_set(tsk->cgroups, cgrp);
+		if (!ent.cg) {
+			retval = -ENOMEM;
+			group_size = i;
+			goto out_list_teardown;
+		}
 		retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
 		BUG_ON(retval != 0);
 		i++;
@@ -2134,7 +2063,7 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 	/* methods shouldn't be called if no task is actually migrating */
 	retval = 0;
 	if (!group_size)
-		goto out_free_group_list;
+		goto out_list_teardown;
 
 	/*
 	 * step 1: check that we can legitimately attach to the cgroup.
@@ -2150,34 +2079,18 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 	}
 
 	/*
-	 * step 2: make sure css_sets exist for all threads to be migrated.
-	 * we use find_css_set, which allocates a new one if necessary.
-	 */
-	INIT_LIST_HEAD(&newcg_list);
-	for (i = 0; i < group_size; i++) {
-		tc = flex_array_get(group, i);
-		oldcg = tc->task->cgroups;
-
-		/* if we don't already have it in the list get a new one */
-		if (!css_set_check_fetched(cgrp, tc->task, oldcg, &newcg_list))
-			if (retval = css_set_prefetch(cgrp, oldcg, &newcg_list))
-				goto out_list_teardown;
-	}
-
-	/*
-	 * step 3: now that we're guaranteed success wrt the css_sets,
+	 * step 2: now that we're guaranteed success wrt the css_sets,
 	 * proceed to move all tasks to the new cgroup.  There are no
 	 * failure cases after here, so this is the commit point.
 	 */
 	for (i = 0; i < group_size; i++) {
 		tc = flex_array_get(group, i);
-		retval = cgroup_task_migrate(cgrp, tc->cgrp, tc->task, true);
-		BUG_ON(retval);
+		cgroup_task_migrate(cgrp, tc->cgrp, tc->task, tc->cg);
 	}
 	/* nothing is sensitive to fork() after this point. */
 
 	/*
-	 * step 4: do subsystem attach callbacks.
+	 * step 3: do subsystem attach callbacks.
 	 */
 	for_each_subsys(root, ss) {
 		if (ss->attach)
@@ -2185,20 +2098,12 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 	}
 
 	/*
-	 * step 5: success! and cleanup
+	 * step 4: success! and cleanup
 	 */
 	synchronize_rcu();
 	cgroup_wakeup_rmdir_waiter(cgrp);
 	retval = 0;
-out_list_teardown:
-	/* clean up the list of prefetched css_sets. */
-	list_for_each_entry_safe(cg_entry, temp_nobe, &newcg_list, links) {
-		list_del(&cg_entry->links);
-		put_css_set(cg_entry->cg);
-		kfree(cg_entry);
-	}
 out_cancel_attach:
-	/* same deal as in cgroup_attach_task */
 	if (retval) {
 		for_each_subsys(root, ss) {
 			if (ss == failed_ss)
@@ -2207,6 +2112,13 @@ out_cancel_attach:
 				ss->cancel_attach(ss, cgrp, &tset);
 		}
 	}
+out_list_teardown:
+	if (retval) {
+		for (i = 0; i < group_size; i++) {
+			tc = flex_array_get(group, i);
+			put_css_set(tc->cg);
+		}
+	}
 out_free_group_list:
 	flex_array_free(group);
 	return retval;
-- 
1.7.3.1

Re: [PATCH 2/3] cgroup: remove double-checking locking from attach_task_by_pid

[Mandeep Singh Baines]

This is not going to work. Will fix up and resend. Need to goto
the top as suggested by Tejun.

Mandeep Singh Baines (msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org) wrote:
> By reading group_leader after taking the threadgroup_lock, we can
> avoid the double-check locking. This removes the return of
> -EAGAIN so we can cleanup cgroup_procs_write at the same time.
> 
> The suggestion was made here:
> 
> https://lkml.org/lkml/2011/12/22/371
> 
> Suggested-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
> Signed-off-by: Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
> Cc: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>
> Cc: containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
> Cc: cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
> Cc: Frederic Weisbecker <fweisbec-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
> Cc: Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
> Cc: Paul Menage <paul-inf54ven1CmVyaH7bEyXVA@public.gmane.org>
> ---
>  kernel/cgroup.c |   40 ++++++----------------------------------
>  1 files changed, 6 insertions(+), 34 deletions(-)
> 
> diff --git a/kernel/cgroup.c b/kernel/cgroup.c
> index 032139d..a5f7d1b 100644
> --- a/kernel/cgroup.c
> +++ b/kernel/cgroup.c
> @@ -2234,9 +2234,6 @@ static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
>  			ret= -ESRCH;
>  			goto out_unlock_cgroup;
>  		}
> -		/* we check later for a group_leader race with de_thread */
> -		if (threadgroup)
> -			tsk = tsk->group_leader;
>  		/*
>  		 * even if we're attaching all tasks in the thread group, we
>  		 * only need to check permissions on one of them.
> @@ -2252,33 +2249,17 @@ static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
>  		get_task_struct(tsk);
>  		rcu_read_unlock();
>  	} else {
> -		if (threadgroup)
> -			tsk = current->group_leader;
> -		else
> -			tsk = current;
> +		tsk = current;
>  		get_task_struct(tsk);
>  	}
>  
>  	threadgroup_lock(tsk);
> -
> -	if (threadgroup) {
> -		if (!thread_group_leader(tsk)) {
> -			/*
> -			 * a race with de_thread from another thread's exec()
> -			 * may strip us of our leadership, if this happens,
> -			 * there is no choice but to throw this task away and
> -			 * try again (from cgroup_procs_write); this is
> -			 * "double-double-toil-and-trouble-check locking".
> -			 */
> -			ret = -EAGAIN;
> -			goto out_unlock_threadgroup;
> -		}
> -		ret = cgroup_attach_proc(cgrp, tsk);
> -	} else
> +	if (threadgroup)
> +		ret = cgroup_attach_proc(cgrp, tsk->group_leader);
> +	else
>  		ret = cgroup_attach_task(cgrp, tsk);
> -
> -out_unlock_threadgroup:
>  	threadgroup_unlock(tsk);
> +
>  	put_task_struct(tsk);
>  out_unlock_cgroup:
>  	cgroup_unlock();
> @@ -2292,16 +2273,7 @@ static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
>  
>  static int cgroup_procs_write(struct cgroup *cgrp, struct cftype *cft, u64 tgid)
>  {
> -	int ret;
> -	do {
> -		/*
> -		 * attach_proc fails with -EAGAIN if threadgroup leadership
> -		 * changes in the middle of the operation, in which case we need
> -		 * to find the task_struct for the new leader and start over.
> -		 */
> -		ret = attach_task_by_pid(cgrp, tgid, true);
> -	} while (ret == -EAGAIN);
> -	return ret;
> +	return attach_task_by_pid(cgrp, tgid, true);
>  }
>  
>  /**
> -- 
> 1.7.3.1
> 

Re: [PATCH 1/3] cgroup: remove tasklist_lock from cgroup_attach_proc

[Frederic Weisbecker]

On Thu, Dec 22, 2011 at 04:57:51PM -0800, Mandeep Singh Baines wrote:
> Since cgroup_attach_proc is protected by a threadgroup_lock, we
> no longer need a tasklist_lock to protect while_each_thread.
> To keep the complexity of the double-check locking in one place,
> I also moved the thread_group_leader check up into
> attach_task_by_pid.
> 
> While at it, also converted a couple of returns to gotos.
> 
> The suggestion was made here:
> 
> https://lkml.org/lkml/2011/12/22/86
> 
> Suggested-by: Frederic Weisbecker <fweisbec-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
> Signed-off-by: Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
> Cc: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
> Cc: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>
> Cc: containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
> Cc: cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
> Cc: Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
> Cc: Paul Menage <paul-inf54ven1CmVyaH7bEyXVA@public.gmane.org>
> ---
>  kernel/cgroup.c |   52 +++++++++++++++++++++-------------------------------
>  1 files changed, 21 insertions(+), 31 deletions(-)
> 
> diff --git a/kernel/cgroup.c b/kernel/cgroup.c
> index 1042b3c..032139d 100644
> --- a/kernel/cgroup.c
> +++ b/kernel/cgroup.c
> @@ -2102,21 +2102,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
>  	if (retval)
>  		goto out_free_group_list;
>  
> -	/* prevent changes to the threadgroup list while we take a snapshot. */
> -	read_lock(&tasklist_lock);
> -	if (!thread_group_leader(leader)) {
> -		/*
> -		 * a race with de_thread from another thread's exec() may strip
> -		 * us of our leadership, making while_each_thread unsafe to use
> -		 * on this task. if this happens, there is no choice but to
> -		 * throw this task away and try again (from cgroup_procs_write);
> -		 * this is "double-double-toil-and-trouble-check locking".
> -		 */
> -		read_unlock(&tasklist_lock);
> -		retval = -EAGAIN;
> -		goto out_free_group_list;
> -	}
> -
>  	tsk = leader;
>  	i = 0;
>  	do {
> @@ -2145,7 +2130,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
>  	group_size = i;
>  	tset.tc_array = group;
>  	tset.tc_array_len = group_size;
> -	read_unlock(&tasklist_lock);

You still need rcu_read_lock()/rcu_read_unlock() around
	do {

	} while_each_thread()

because threadgroup_lock() doesn't lock the part that remove a thread from
its group on exit.

Re: [PATCH 1/3] cgroup: remove tasklist_lock from cgroup_attach_proc

[Frederic Weisbecker]

On Fri, Dec 23, 2011 at 03:27:45AM +0100, Frederic Weisbecker wrote:
> On Thu, Dec 22, 2011 at 04:57:51PM -0800, Mandeep Singh Baines wrote:
> > Since cgroup_attach_proc is protected by a threadgroup_lock, we
> > no longer need a tasklist_lock to protect while_each_thread.
> > To keep the complexity of the double-check locking in one place,
> > I also moved the thread_group_leader check up into
> > attach_task_by_pid.
> > 
> > While at it, also converted a couple of returns to gotos.
> > 
> > The suggestion was made here:
> > 
> > https://lkml.org/lkml/2011/12/22/86
> > 
> > Suggested-by: Frederic Weisbecker <fweisbec-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
> > Signed-off-by: Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
> > Cc: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
> > Cc: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>
> > Cc: containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
> > Cc: cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> > Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
> > Cc: Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> > Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
> > Cc: Paul Menage <paul-inf54ven1CmVyaH7bEyXVA@public.gmane.org>
> > ---
> >  kernel/cgroup.c |   52 +++++++++++++++++++++-------------------------------
> >  1 files changed, 21 insertions(+), 31 deletions(-)
> > 
> > diff --git a/kernel/cgroup.c b/kernel/cgroup.c
> > index 1042b3c..032139d 100644
> > --- a/kernel/cgroup.c
> > +++ b/kernel/cgroup.c
> > @@ -2102,21 +2102,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
> >  	if (retval)
> >  		goto out_free_group_list;
> >  
> > -	/* prevent changes to the threadgroup list while we take a snapshot. */
> > -	read_lock(&tasklist_lock);
> > -	if (!thread_group_leader(leader)) {
> > -		/*
> > -		 * a race with de_thread from another thread's exec() may strip
> > -		 * us of our leadership, making while_each_thread unsafe to use
> > -		 * on this task. if this happens, there is no choice but to
> > -		 * throw this task away and try again (from cgroup_procs_write);
> > -		 * this is "double-double-toil-and-trouble-check locking".
> > -		 */
> > -		read_unlock(&tasklist_lock);
> > -		retval = -EAGAIN;
> > -		goto out_free_group_list;
> > -	}
> > -
> >  	tsk = leader;
> >  	i = 0;
> >  	do {
> > @@ -2145,7 +2130,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
> >  	group_size = i;
> >  	tset.tc_array = group;
> >  	tset.tc_array_len = group_size;
> > -	read_unlock(&tasklist_lock);
> 
> You still need rcu_read_lock()/rcu_read_unlock() around
> 	do {
> 
> 	} while_each_thread()
> 
> because threadgroup_lock() doesn't lock the part that remove a thread from
> its group on exit.

Actually while_each_thread() takes care of the thread group list safe
walking. But we need RCU to ensure the task is not released in parallel.
threadgroup_lock() doesn't synchronize against that if the task has
already passed the setting of PF_EXITING.

Re: [PATCH 1/3] cgroup: remove tasklist_lock from cgroup_attach_proc

[Li Zefan]

>> diff --git a/kernel/cgroup.c b/kernel/cgroup.c
>> index 1042b3c..032139d 100644
>> --- a/kernel/cgroup.c
>> +++ b/kernel/cgroup.c
>> @@ -2102,21 +2102,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
>>  	if (retval)
>>  		goto out_free_group_list;
>>  
>> -	/* prevent changes to the threadgroup list while we take a snapshot. */
>> -	read_lock(&tasklist_lock);
>> -	if (!thread_group_leader(leader)) {
>> -		/*
>> -		 * a race with de_thread from another thread's exec() may strip
>> -		 * us of our leadership, making while_each_thread unsafe to use
>> -		 * on this task. if this happens, there is no choice but to
>> -		 * throw this task away and try again (from cgroup_procs_write);
>> -		 * this is "double-double-toil-and-trouble-check locking".
>> -		 */
>> -		read_unlock(&tasklist_lock);
>> -		retval = -EAGAIN;
>> -		goto out_free_group_list;
>> -	}
>> -
>>  	tsk = leader;
>>  	i = 0;
>>  	do {
>> @@ -2145,7 +2130,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
>>  	group_size = i;
>>  	tset.tc_array = group;
>>  	tset.tc_array_len = group_size;
>> -	read_unlock(&tasklist_lock);
> 
> You still need rcu_read_lock()/rcu_read_unlock() around
> 	do {
> 
> 	} while_each_thread()
> 
> because threadgroup_lock() doesn't lock the part that remove a thread from
> its group on exit.
> 

and inside rcu critical section, you can't call kmalloc(GFP_KERNEL)!!

Re: [PATCH 1/3] cgroup: remove tasklist_lock from cgroup_attach_proc

[Frederic Weisbecker]

On Fri, Dec 23, 2011 at 10:40:28AM +0800, Li Zefan wrote:
> >> diff --git a/kernel/cgroup.c b/kernel/cgroup.c
> >> index 1042b3c..032139d 100644
> >> --- a/kernel/cgroup.c
> >> +++ b/kernel/cgroup.c
> >> @@ -2102,21 +2102,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
> >>  	if (retval)
> >>  		goto out_free_group_list;
> >>  
> >> -	/* prevent changes to the threadgroup list while we take a snapshot. */
> >> -	read_lock(&tasklist_lock);
> >> -	if (!thread_group_leader(leader)) {
> >> -		/*
> >> -		 * a race with de_thread from another thread's exec() may strip
> >> -		 * us of our leadership, making while_each_thread unsafe to use
> >> -		 * on this task. if this happens, there is no choice but to
> >> -		 * throw this task away and try again (from cgroup_procs_write);
> >> -		 * this is "double-double-toil-and-trouble-check locking".
> >> -		 */
> >> -		read_unlock(&tasklist_lock);
> >> -		retval = -EAGAIN;
> >> -		goto out_free_group_list;
> >> -	}
> >> -
> >>  	tsk = leader;
> >>  	i = 0;
> >>  	do {
> >> @@ -2145,7 +2130,6 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
> >>  	group_size = i;
> >>  	tset.tc_array = group;
> >>  	tset.tc_array_len = group_size;
> >> -	read_unlock(&tasklist_lock);
> > 
> > You still need rcu_read_lock()/rcu_read_unlock() around
> > 	do {
> > 
> > 	} while_each_thread()
> > 
> > because threadgroup_lock() doesn't lock the part that remove a thread from
> > its group on exit.
> > 
> 
> and inside rcu critical section, you can't call kmalloc(GFP_KERNEL)!!

Good point. Well it's still worth replacing tasklist_lock by rcu though :)

[PATCH 3/3] cgroup: remove extra calls to find_existing_css_set

[Mandeep Singh Baines]

In cgroup_attach_proc, we indirectly call find_existing_css_set 3
times. It is an expensive call so we want to call it a minimum
of times. This patch only calls it once and stores the result so
that it can be used later on when we call cgroup_task_migrate.

This required modifying cgroup_task_migrate to take the new css_set
(which we obtained from find_css_set) as a parameter. The nice side
effect of this is that cgroup_task_migrate is now identical for
cgroup_attach_task and cgroup_attach_proc. It also now returns a
void since it can never fail.

Changes in V4:
* https://lkml.org/lkml/2011/12/22/421 (Li Zefan)
  * Avoid GFP_KERNEL (sleep) in rcu_read_lock by getting css_set in
    a separate loop not under an rcu_read_lock
Changes in V3:
* https://lkml.org/lkml/2011/12/22/13 (Li Zefan)
  * Fixed earlier bug by creating a seperate patch to remove tasklist_lock
Changes in V2:
* https://lkml.org/lkml/2011/12/20/372 (Tejun Heo)
  * Move find_css_set call into loop which creates the flex array
* Author
  * Kill css_set_refs and use group_size instead
  * Fix an off-by-one error in counting css_set refs
  * Add a retval check in out_list_teardown

Signed-off-by: Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
Cc: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Cc: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>
Cc: containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
Cc: cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
Cc: Frederic Weisbecker <fweisbec-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
Cc: Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
Cc: Paul Menage <paul-inf54ven1CmVyaH7bEyXVA@public.gmane.org>
---
 kernel/cgroup.c |  142 +++++++++++-------------------------------------------
 1 files changed, 29 insertions(+), 113 deletions(-)

diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index 354c539..16e8e70 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c
@@ -1763,6 +1763,7 @@ EXPORT_SYMBOL_GPL(cgroup_path);
 struct task_and_cgroup {
 	struct task_struct	*task;
 	struct cgroup		*cgrp;
+	struct css_set		*cg;
 };
 
 struct cgroup_taskset {
@@ -1843,11 +1844,10 @@ EXPORT_SYMBOL_GPL(cgroup_taskset_size);
  * will already exist. If not set, this function might sleep, and can fail with
  * -ENOMEM. Must be called with cgroup_mutex and threadgroup locked.
  */
-static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
-			       struct task_struct *tsk, bool guarantee)
+static void cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
+				struct task_struct *tsk, struct css_set *newcg)
 {
 	struct css_set *oldcg;
-	struct css_set *newcg;
 
 	/*
 	 * We are synchronized through threadgroup_lock() against PF_EXITING
@@ -1857,23 +1857,6 @@ static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
 	WARN_ON_ONCE(tsk->flags & PF_EXITING);
 	oldcg = tsk->cgroups;
 
-	/* locate or allocate a new css_set for this task. */
-	if (guarantee) {
-		/* we know the css_set we want already exists. */
-		struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
-		read_lock(&css_set_lock);
-		newcg = find_existing_css_set(oldcg, cgrp, template);
-		BUG_ON(!newcg);
-		get_css_set(newcg);
-		read_unlock(&css_set_lock);
-	} else {
-		might_sleep();
-		/* find_css_set will give us newcg already referenced. */
-		newcg = find_css_set(oldcg, cgrp);
-		if (!newcg)
-			return -ENOMEM;
-	}
-
 	task_lock(tsk);
 	rcu_assign_pointer(tsk->cgroups, newcg);
 	task_unlock(tsk);
@@ -1892,7 +1875,6 @@ static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
 	put_css_set(oldcg);
 
 	set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
-	return 0;
 }
 
 /**
@@ -1910,6 +1892,7 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
 	struct cgroup *oldcgrp;
 	struct cgroupfs_root *root = cgrp->root;
 	struct cgroup_taskset tset = { };
+	struct css_set *newcg;
 
 	/* @tsk either already exited or can't exit until the end */
 	if (tsk->flags & PF_EXITING)
@@ -1939,9 +1922,13 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
 		}
 	}
 
-	retval = cgroup_task_migrate(cgrp, oldcgrp, tsk, false);
-	if (retval)
+	newcg = find_css_set(tsk->cgroups, cgrp);
+	if (!newcg) {
+		retval = -ENOMEM;
 		goto out;
+	}
+
+	cgroup_task_migrate(cgrp, oldcgrp, tsk, newcg);
 
 	for_each_subsys(root, ss) {
 		if (ss->attach)
@@ -1997,66 +1984,6 @@ int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
 }
 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
 
-/*
- * cgroup_attach_proc works in two stages, the first of which prefetches all
- * new css_sets needed (to make sure we have enough memory before committing
- * to the move) and stores them in a list of entries of the following type.
- * TODO: possible optimization: use css_set->rcu_head for chaining instead
- */
-struct cg_list_entry {
-	struct css_set *cg;
-	struct list_head links;
-};
-
-static bool css_set_check_fetched(struct cgroup *cgrp,
-				  struct task_struct *tsk, struct css_set *cg,
-				  struct list_head *newcg_list)
-{
-	struct css_set *newcg;
-	struct cg_list_entry *cg_entry;
-	struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
-
-	read_lock(&css_set_lock);
-	newcg = find_existing_css_set(cg, cgrp, template);
-	read_unlock(&css_set_lock);
-
-	/* doesn't exist at all? */
-	if (!newcg)
-		return false;
-	/* see if it's already in the list */
-	list_for_each_entry(cg_entry, newcg_list, links)
-		if (cg_entry->cg == newcg)
-			return true;
-
-	/* not found */
-	return false;
-}
-
-/*
- * Find the new css_set and store it in the list in preparation for moving the
- * given task to the given cgroup. Returns 0 or -ENOMEM.
- */
-static int css_set_prefetch(struct cgroup *cgrp, struct css_set *cg,
-			    struct list_head *newcg_list)
-{
-	struct css_set *newcg;
-	struct cg_list_entry *cg_entry;
-
-	/* ensure a new css_set will exist for this thread */
-	newcg = find_css_set(cg, cgrp);
-	if (!newcg)
-		return -ENOMEM;
-	/* add it to the list */
-	cg_entry = kmalloc(sizeof(struct cg_list_entry), GFP_KERNEL);
-	if (!cg_entry) {
-		put_css_set(newcg);
-		return -ENOMEM;
-	}
-	cg_entry->cg = newcg;
-	list_add(&cg_entry->links, newcg_list);
-	return 0;
-}
-
 /**
  * cgroup_attach_proc - attach all threads in a threadgroup to a cgroup
  * @cgrp: the cgroup to attach to
@@ -2067,23 +1994,15 @@ static int css_set_prefetch(struct cgroup *cgrp, struct css_set *cg,
  */
 int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 {
-	int retval, i, group_size;
+	int retval, i, group_size, css_set_refs = 0;
 	struct cgroup_subsys *ss, *failed_ss = NULL;
 	/* guaranteed to be initialized later, but the compiler needs this */
-	struct css_set *oldcg;
 	struct cgroupfs_root *root = cgrp->root;
 	/* threadgroup list cursor and array */
 	struct task_struct *tsk;
 	struct task_and_cgroup *tc;
 	struct flex_array *group;
 	struct cgroup_taskset tset = { };
-	/*
-	 * we need to make sure we have css_sets for all the tasks we're
-	 * going to move -before- we actually start moving them, so that in
-	 * case we get an ENOMEM we can bail out before making any changes.
-	 */
-	struct list_head newcg_list;
-	struct cg_list_entry *cg_entry, *temp_nobe;
 
 	/*
 	 * step 0: in order to do expensive, possibly blocking operations for
@@ -2119,15 +2038,15 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 
 		/* as per above, nr_threads may decrease, but not increase. */
 		BUG_ON(i >= group_size);
-		/*
-		 * saying GFP_ATOMIC has no effect here because we did prealloc
-		 * earlier, but it's good form to communicate our expectations.
-		 */
 		ent.task = tsk;
 		ent.cgrp = task_cgroup_from_root(tsk, root);
 		/* nothing to do if this task is already in the cgroup */
 		if (ent.cgrp == cgrp)
 			continue;
+		/*
+		 * saying GFP_ATOMIC has no effect here because we did prealloc
+		 * earlier, but it's good form to communicate our expectations.
+		 */
 		retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
 		BUG_ON(retval != 0);
 		i++;
@@ -2156,19 +2075,18 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 		}
 	}
 
-	/*
+        /*
 	 * step 2: make sure css_sets exist for all threads to be migrated.
 	 * we use find_css_set, which allocates a new one if necessary.
 	 */
-	INIT_LIST_HEAD(&newcg_list);
 	for (i = 0; i < group_size; i++) {
 		tc = flex_array_get(group, i);
-		oldcg = tc->task->cgroups;
-
-		/* if we don't already have it in the list get a new one */
-		if (!css_set_check_fetched(cgrp, tc->task, oldcg, &newcg_list))
-			if (retval = css_set_prefetch(cgrp, oldcg, &newcg_list))
-				goto out_list_teardown;
+		tc->cg = find_css_set(tc->task->cgroups, cgrp);
+		if (!tc->cg) {
+			retval = -ENOMEM;
+			css_set_refs = i;
+			goto out_put_css_set_refs;
+		}
 	}
 
 	/*
@@ -2178,8 +2096,7 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 	 */
 	for (i = 0; i < group_size; i++) {
 		tc = flex_array_get(group, i);
-		retval = cgroup_task_migrate(cgrp, tc->cgrp, tc->task, true);
-		BUG_ON(retval);
+		cgroup_task_migrate(cgrp, tc->cgrp, tc->task, tc->cg);
 	}
 	/* nothing is sensitive to fork() after this point. */
 
@@ -2197,15 +2114,14 @@ int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
 	synchronize_rcu();
 	cgroup_wakeup_rmdir_waiter(cgrp);
 	retval = 0;
-out_list_teardown:
-	/* clean up the list of prefetched css_sets. */
-	list_for_each_entry_safe(cg_entry, temp_nobe, &newcg_list, links) {
-		list_del(&cg_entry->links);
-		put_css_set(cg_entry->cg);
-		kfree(cg_entry);
+out_put_css_set_refs:
+	if (retval) {
+		for (i = 0; i < css_set_refs; i++) {
+			tc = flex_array_get(group, i);
+			put_css_set(tc->cg);
+		}
 	}
 out_cancel_attach:
-	/* same deal as in cgroup_attach_task */
 	if (retval) {
 		for_each_subsys(root, ss) {
 			if (ss == failed_ss)
-- 
1.7.3.1

Re: [PATCH 3/3] cgroup: remove extra calls to find_existing_css_set

[Li Zefan]

Mandeep Singh Baines wrote:
> In cgroup_attach_proc, we indirectly call find_existing_css_set 3
> times. It is an expensive call so we want to call it a minimum
> of times. This patch only calls it once and stores the result so
> that it can be used later on when we call cgroup_task_migrate.
> 
> This required modifying cgroup_task_migrate to take the new css_set
> (which we obtained from find_css_set) as a parameter. The nice side
> effect of this is that cgroup_task_migrate is now identical for
> cgroup_attach_task and cgroup_attach_proc. It also now returns a
> void since it can never fail.
> 
> Changes in V4:
> * https://lkml.org/lkml/2011/12/22/421 (Li Zefan)
>   * Avoid GFP_KERNEL (sleep) in rcu_read_lock by getting css_set in
>     a separate loop not under an rcu_read_lock
> Changes in V3:
> * https://lkml.org/lkml/2011/12/22/13 (Li Zefan)
>   * Fixed earlier bug by creating a seperate patch to remove tasklist_lock
> Changes in V2:
> * https://lkml.org/lkml/2011/12/20/372 (Tejun Heo)
>   * Move find_css_set call into loop which creates the flex array
> * Author
>   * Kill css_set_refs and use group_size instead
>   * Fix an off-by-one error in counting css_set refs
>   * Add a retval check in out_list_teardown
> 
> Signed-off-by: Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>

Acked-by: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>

> Cc: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
> Cc: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>
> Cc: containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
> Cc: cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
> Cc: Frederic Weisbecker <fweisbec-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
> Cc: Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
> Cc: Paul Menage <paul-inf54ven1CmVyaH7bEyXVA@public.gmane.org>
> ---
>  kernel/cgroup.c |  142 +++++++++++-------------------------------------------
>  1 files changed, 29 insertions(+), 113 deletions(-)

Re: [PATCH 3/3] cgroup: remove extra calls to find_existing_css_set

[Tejun Heo]

Hello, guys.

On Thu, Jan 05, 2012 at 10:10:57AM +0800, Li Zefan wrote:
> Mandeep Singh Baines wrote:
> > In cgroup_attach_proc, we indirectly call find_existing_css_set 3
> > times. It is an expensive call so we want to call it a minimum
> > of times. This patch only calls it once and stores the result so
> > that it can be used later on when we call cgroup_task_migrate.
> > 
> > This required modifying cgroup_task_migrate to take the new css_set
> > (which we obtained from find_css_set) as a parameter. The nice side
> > effect of this is that cgroup_task_migrate is now identical for
> > cgroup_attach_task and cgroup_attach_proc. It also now returns a
> > void since it can never fail.
> > 
> > Changes in V4:
> > * https://lkml.org/lkml/2011/12/22/421 (Li Zefan)
> >   * Avoid GFP_KERNEL (sleep) in rcu_read_lock by getting css_set in
> >     a separate loop not under an rcu_read_lock
> > Changes in V3:
> > * https://lkml.org/lkml/2011/12/22/13 (Li Zefan)
> >   * Fixed earlier bug by creating a seperate patch to remove tasklist_lock
> > Changes in V2:
> > * https://lkml.org/lkml/2011/12/20/372 (Tejun Heo)
> >   * Move find_css_set call into loop which creates the flex array
> > * Author
> >   * Kill css_set_refs and use group_size instead
> >   * Fix an off-by-one error in counting css_set refs
> >   * Add a retval check in out_list_teardown
> > 
> > Signed-off-by: Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
> 
> Acked-by: Li Zefan <lizf-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>

I like this series but it's too late for the current merge window.
I'll apply them on for-3.4 once this merge window closes.  If I don't,
please ping me.

Thank you.

-- 
tejun

Re: [PATCH 3/3] cgroup: remove extra calls to find_existing_css_set

[Tejun Heo]

On Tue, Jan 03, 2012 at 09:18:32PM -0800, Mandeep Singh Baines wrote:
> +out_put_css_set_refs:
> +	if (retval) {
> +		for (i = 0; i < css_set_refs; i++) {
> +			tc = flex_array_get(group, i);
> +			put_css_set(tc->cg);

Why not the following

	for (i = 0; i < group_size; i++) {
		tc = flex_array_get(group, i);
		if (tc)
			put_css_set(tc->cg);
	}

and remove css_set_refs?

Thanks

-- 
tejun

Re: [PATCH 3/3] cgroup: remove extra calls to find_existing_css_set

[Mandeep Singh Baines]

Tejun Heo (tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org) wrote:
> On Tue, Jan 03, 2012 at 09:18:32PM -0800, Mandeep Singh Baines wrote:
> > +out_put_css_set_refs:
> > +	if (retval) {
> > +		for (i = 0; i < css_set_refs; i++) {
> > +			tc = flex_array_get(group, i);
> > +			put_css_set(tc->cg);
> 
> Why not the following
> 
> 	for (i = 0; i < group_size; i++) {
> 		tc = flex_array_get(group, i);
> 		if (tc)
> 			put_css_set(tc->cg);
> 	}
> 
> and remove css_set_refs?
> 

If I break out of step 2 via the goto (out_put_css_set_refs), I only
want to put back the refs I've gotten so far. I could reset group_size
in that case but wanted avoid overloading group_size: someone
could later add more error path code that isn't expecting that
group_size will get re-assigned and have its meaning change.

Regards,
Mandeep

> Thanks
> 
> -- 
> tejun

Re: [PATCH 3/3] cgroup: remove extra calls to find_existing_css_set

[Tejun Heo]

Hello,

On Fri, Jan 20, 2012 at 6:57 PM, Mandeep Singh Baines <msb-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org> wrote:
> If I break out of step 2 via the goto (out_put_css_set_refs), I only
> want to put back the refs I've gotten so far.

But isn't tc->cg NULL if that happens?  (ooh, the if statement should
be testing tc->cg, my bad).

Thanks.

-- 
tejun