cgroups: make procs file writable
Ben Blum [Thu, 26 May 2011 23:25:20 +0000 (16:25 -0700)]
Make procs file writable to move all threads by tgid at once.

Add functionality that enables users to move all threads in a threadgroup
at once to a cgroup by writing the tgid to the 'cgroup.procs' file.  This
current implementation makes use of a per-threadgroup rwsem that's taken
for reading in the fork() path to prevent newly forking threads within the
threadgroup from "escaping" while the move is in progress.

Signed-off-by: Ben Blum <bblum@andrew.cmu.edu>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Matt Helsley <matthltc@us.ibm.com>
Reviewed-by: Paul Menage <menage@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Documentation/cgroups/cgroups.txt
kernel/cgroup.c

index b3bd3bd..8c4f346 100644 (file)
@@ -236,7 +236,8 @@ containing the following files describing that cgroup:
  - cgroup.procs: list of tgids in the cgroup.  This list is not
    guaranteed to be sorted or free of duplicate tgids, and userspace
    should sort/uniquify the list if this property is required.
-   This is a read-only file, for now.
+   Writing a thread group id into this file moves all threads in that
+   group into this cgroup.
  - notify_on_release flag: run the release agent on exit?
  - release_agent: the path to use for release notifications (this file
    exists in the top cgroup only)
@@ -430,6 +431,12 @@ You can attach the current shell task by echoing 0:
 
 # echo 0 > tasks
 
+You can use the cgroup.procs file instead of the tasks file to move all
+threads in a threadgroup at once. Echoing the pid of any task in a
+threadgroup to cgroup.procs causes all tasks in that threadgroup to be
+be attached to the cgroup. Writing 0 to cgroup.procs moves all tasks
+in the writing task's threadgroup.
+
 Note: Since every task is always a member of exactly one cgroup in each
 mounted hierarchy, to remove a task from its current cgroup you must
 move it into a new cgroup (possibly the root cgroup) by writing to the
index 38fb0ad..5e6a974 100644 (file)
@@ -1735,6 +1735,76 @@ int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
 }
 EXPORT_SYMBOL_GPL(cgroup_path);
 
+/*
+ * cgroup_task_migrate - move a task from one cgroup to another.
+ *
+ * 'guarantee' is set if the caller promises that a new css_set for the task
+ * will already exist. If not set, this function might sleep, and can fail with
+ * -ENOMEM. Otherwise, it can only fail with -ESRCH.
+ */
+static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
+                              struct task_struct *tsk, bool guarantee)
+{
+       struct css_set *oldcg;
+       struct css_set *newcg;
+
+       /*
+        * get old css_set. we need to take task_lock and refcount it, because
+        * an exiting task can change its css_set to init_css_set and drop its
+        * old one without taking cgroup_mutex.
+        */
+       task_lock(tsk);
+       oldcg = tsk->cgroups;
+       get_css_set(oldcg);
+       task_unlock(tsk);
+
+       /* 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) {
+                       put_css_set(oldcg);
+                       return -ENOMEM;
+               }
+       }
+       put_css_set(oldcg);
+
+       /* if PF_EXITING is set, the tsk->cgroups pointer is no longer safe. */
+       task_lock(tsk);
+       if (tsk->flags & PF_EXITING) {
+               task_unlock(tsk);
+               put_css_set(newcg);
+               return -ESRCH;
+       }
+       rcu_assign_pointer(tsk->cgroups, newcg);
+       task_unlock(tsk);
+
+       /* Update the css_set linked lists if we're using them */
+       write_lock(&css_set_lock);
+       if (!list_empty(&tsk->cg_list))
+               list_move(&tsk->cg_list, &newcg->tasks);
+       write_unlock(&css_set_lock);
+
+       /*
+        * We just gained a reference on oldcg by taking it from the task. As
+        * trading it for newcg is protected by cgroup_mutex, we're safe to drop
+        * it here; it will be freed under RCU.
+        */
+       put_css_set(oldcg);
+
+       set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
+       return 0;
+}
+
 /**
  * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
  * @cgrp: the cgroup the task is attaching to
@@ -1745,11 +1815,9 @@ EXPORT_SYMBOL_GPL(cgroup_path);
  */
 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
 {
-       int retval = 0;
+       int retval;
        struct cgroup_subsys *ss, *failed_ss = NULL;
        struct cgroup *oldcgrp;
-       struct css_set *cg;
-       struct css_set *newcg;
        struct cgroupfs_root *root = cgrp->root;
 
        /* Nothing to do if the task is already in that cgroup */
@@ -1780,36 +1848,9 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
                }
        }
 
-       task_lock(tsk);
-       cg = tsk->cgroups;
-       get_css_set(cg);
-       task_unlock(tsk);
-       /*
-        * Locate or allocate a new css_set for this task,
-        * based on its final set of cgroups
-        */
-       newcg = find_css_set(cg, cgrp);
-       put_css_set(cg);
-       if (!newcg) {
-               retval = -ENOMEM;
-               goto out;
-       }
-
-       task_lock(tsk);
-       if (tsk->flags & PF_EXITING) {
-               task_unlock(tsk);
-               put_css_set(newcg);
-               retval = -ESRCH;
+       retval = cgroup_task_migrate(cgrp, oldcgrp, tsk, false);
+       if (retval)
                goto out;
-       }
-       rcu_assign_pointer(tsk->cgroups, newcg);
-       task_unlock(tsk);
-
-       /* Update the css_set linked lists if we're using them */
-       write_lock(&css_set_lock);
-       if (!list_empty(&tsk->cg_list))
-               list_move(&tsk->cg_list, &newcg->tasks);
-       write_unlock(&css_set_lock);
 
        for_each_subsys(root, ss) {
                if (ss->pre_attach)
@@ -1819,9 +1860,8 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
                if (ss->attach)
                        ss->attach(ss, cgrp, oldcgrp, tsk);
        }
-       set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
+
        synchronize_rcu();
-       put_css_set(cg);
 
        /*
         * wake up rmdir() waiter. the rmdir should fail since the cgroup
@@ -1871,49 +1911,356 @@ int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
 
 /*
- * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
- * held. May take task_lock of task
+ * 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);
+       if (newcg)
+               get_css_set(newcg);
+       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) {
+                       put_css_set(newcg);
+                       return true;
+               }
+       }
+
+       /* not found */
+       put_css_set(newcg);
+       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
+ * @leader: the threadgroup leader task_struct of the group to be attached
+ *
+ * Call holding cgroup_mutex and the threadgroup_fork_lock of the leader. Will
+ * take task_lock of each thread in leader's threadgroup individually in turn.
+ */
+int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
+{
+       int retval, i, group_size;
+       struct cgroup_subsys *ss, *failed_ss = NULL;
+       bool cancel_failed_ss = false;
+       /* guaranteed to be initialized later, but the compiler needs this */
+       struct cgroup *oldcgrp = NULL;
+       struct css_set *oldcg;
+       struct cgroupfs_root *root = cgrp->root;
+       /* threadgroup list cursor and array */
+       struct task_struct *tsk;
+       struct task_struct **group;
+       /*
+        * 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
+        * every thread, we cannot iterate the thread group list, since it needs
+        * rcu or tasklist locked. instead, build an array of all threads in the
+        * group - threadgroup_fork_lock prevents new threads from appearing,
+        * and if threads exit, this will just be an over-estimate.
+        */
+       group_size = get_nr_threads(leader);
+       group = kmalloc(group_size * sizeof(*group), GFP_KERNEL);
+       if (!group)
+               return -ENOMEM;
+
+       /* prevent changes to the threadgroup list while we take a snapshot. */
+       rcu_read_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".
+                */
+               rcu_read_unlock();
+               retval = -EAGAIN;
+               goto out_free_group_list;
+       }
+       /* take a reference on each task in the group to go in the array. */
+       tsk = leader;
+       i = 0;
+       do {
+               /* as per above, nr_threads may decrease, but not increase. */
+               BUG_ON(i >= group_size);
+               get_task_struct(tsk);
+               group[i] = tsk;
+               i++;
+       } while_each_thread(leader, tsk);
+       /* remember the number of threads in the array for later. */
+       group_size = i;
+       rcu_read_unlock();
+
+       /*
+        * step 1: check that we can legitimately attach to the cgroup.
+        */
+       for_each_subsys(root, ss) {
+               if (ss->can_attach) {
+                       retval = ss->can_attach(ss, cgrp, leader);
+                       if (retval) {
+                               failed_ss = ss;
+                               goto out_cancel_attach;
+                       }
+               }
+               /* a callback to be run on every thread in the threadgroup. */
+               if (ss->can_attach_task) {
+                       /* run on each task in the threadgroup. */
+                       for (i = 0; i < group_size; i++) {
+                               retval = ss->can_attach_task(cgrp, group[i]);
+                               if (retval) {
+                                       failed_ss = ss;
+                                       cancel_failed_ss = true;
+                                       goto out_cancel_attach;
+                               }
+                       }
+               }
+       }
+
+       /*
+        * 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++) {
+               tsk = group[i];
+               /* nothing to do if this task is already in the cgroup */
+               oldcgrp = task_cgroup_from_root(tsk, root);
+               if (cgrp == oldcgrp)
+                       continue;
+               /* get old css_set pointer */
+               task_lock(tsk);
+               if (tsk->flags & PF_EXITING) {
+                       /* ignore this task if it's going away */
+                       task_unlock(tsk);
+                       continue;
+               }
+               oldcg = tsk->cgroups;
+               get_css_set(oldcg);
+               task_unlock(tsk);
+               /* see if the new one for us is already in the list? */
+               if (css_set_check_fetched(cgrp, tsk, oldcg, &newcg_list)) {
+                       /* was already there, nothing to do. */
+                       put_css_set(oldcg);
+               } else {
+                       /* we don't already have it. get new one. */
+                       retval = css_set_prefetch(cgrp, oldcg, &newcg_list);
+                       put_css_set(oldcg);
+                       if (retval)
+                               goto out_list_teardown;
+               }
+       }
+
+       /*
+        * step 3: now that we're guaranteed success wrt the css_sets, proceed
+        * to move all tasks to the new cgroup, calling ss->attach_task for each
+        * one along the way. there are no failure cases after here, so this is
+        * the commit point.
+        */
+       for_each_subsys(root, ss) {
+               if (ss->pre_attach)
+                       ss->pre_attach(cgrp);
+       }
+       for (i = 0; i < group_size; i++) {
+               tsk = group[i];
+               /* leave current thread as it is if it's already there */
+               oldcgrp = task_cgroup_from_root(tsk, root);
+               if (cgrp == oldcgrp)
+                       continue;
+               /* attach each task to each subsystem */
+               for_each_subsys(root, ss) {
+                       if (ss->attach_task)
+                               ss->attach_task(cgrp, tsk);
+               }
+               /* if the thread is PF_EXITING, it can just get skipped. */
+               retval = cgroup_task_migrate(cgrp, oldcgrp, tsk, true);
+               BUG_ON(retval != 0 && retval != -ESRCH);
+       }
+       /* nothing is sensitive to fork() after this point. */
+
+       /*
+        * step 4: do expensive, non-thread-specific subsystem callbacks.
+        * TODO: if ever a subsystem needs to know the oldcgrp for each task
+        * being moved, this call will need to be reworked to communicate that.
+        */
+       for_each_subsys(root, ss) {
+               if (ss->attach)
+                       ss->attach(ss, cgrp, oldcgrp, leader);
+       }
+
+       /*
+        * step 5: 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) {
+                               if (cancel_failed_ss && ss->cancel_attach)
+                                       ss->cancel_attach(ss, cgrp, leader);
+                               break;
+                       }
+                       if (ss->cancel_attach)
+                               ss->cancel_attach(ss, cgrp, leader);
+               }
+       }
+       /* clean up the array of referenced threads in the group. */
+       for (i = 0; i < group_size; i++)
+               put_task_struct(group[i]);
+out_free_group_list:
+       kfree(group);
+       return retval;
+}
+
+/*
+ * Find the task_struct of the task to attach by vpid and pass it along to the
+ * function to attach either it or all tasks in its threadgroup. Will take
+ * cgroup_mutex; may take task_lock of task.
  */
-static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
+static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
 {
        struct task_struct *tsk;
        const struct cred *cred = current_cred(), *tcred;
        int ret;
 
+       if (!cgroup_lock_live_group(cgrp))
+               return -ENODEV;
+
        if (pid) {
                rcu_read_lock();
                tsk = find_task_by_vpid(pid);
-               if (!tsk || tsk->flags & PF_EXITING) {
+               if (!tsk) {
                        rcu_read_unlock();
+                       cgroup_unlock();
+                       return -ESRCH;
+               }
+               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.
+                        */
+                       tsk = tsk->group_leader;
+               } else if (tsk->flags & PF_EXITING) {
+                       /* optimization for the single-task-only case */
+                       rcu_read_unlock();
+                       cgroup_unlock();
                        return -ESRCH;
                }
 
+               /*
+                * even if we're attaching all tasks in the thread group, we
+                * only need to check permissions on one of them.
+                */
                tcred = __task_cred(tsk);
                if (cred->euid &&
                    cred->euid != tcred->uid &&
                    cred->euid != tcred->suid) {
                        rcu_read_unlock();
+                       cgroup_unlock();
                        return -EACCES;
                }
                get_task_struct(tsk);
                rcu_read_unlock();
        } else {
-               tsk = current;
+               if (threadgroup)
+                       tsk = current->group_leader;
+               else
+                       tsk = current;
                get_task_struct(tsk);
        }
 
-       ret = cgroup_attach_task(cgrp, tsk);
+       if (threadgroup) {
+               threadgroup_fork_write_lock(tsk);
+               ret = cgroup_attach_proc(cgrp, tsk);
+               threadgroup_fork_write_unlock(tsk);
+       } else {
+               ret = cgroup_attach_task(cgrp, tsk);
+       }
        put_task_struct(tsk);
+       cgroup_unlock();
        return ret;
 }
 
 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
 {
+       return attach_task_by_pid(cgrp, pid, false);
+}
+
+static int cgroup_procs_write(struct cgroup *cgrp, struct cftype *cft, u64 tgid)
+{
        int ret;
-       if (!cgroup_lock_live_group(cgrp))
-               return -ENODEV;
-       ret = attach_task_by_pid(cgrp, pid);
-       cgroup_unlock();
+       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;
 }
 
@@ -3270,9 +3617,9 @@ static struct cftype files[] = {
        {
                .name = CGROUP_FILE_GENERIC_PREFIX "procs",
                .open = cgroup_procs_open,
-               /* .write_u64 = cgroup_procs_write, TODO */
+               .write_u64 = cgroup_procs_write,
                .release = cgroup_pidlist_release,
-               .mode = S_IRUGO,
+               .mode = S_IRUGO | S_IWUSR,
        },
        {
                .name = "notify_on_release",