[PATCH] delay accounting taskstats interface send tgid once

Send per-tgid data only once during exit of a thread group instead of once
with each member thread exit.

Currently, when a thread exits, besides its per-tid data, the per-tgid data
of its thread group is also sent out, if its thread group is non-empty.
The per-tgid data sent consists of the sum of per-tid stats for all
*remaining* threads of the thread group.

This patch modifies this sending in two ways:

- the per-tgid data is sent only when the last thread of a thread group
  exits.  This cuts down heavily on the overhead of sending/receiving
  per-tgid data, especially when other exploiters of the taskstats
  interface aren't interested in per-tgid stats

- the semantics of the per-tgid data sent are changed.  Instead of being
  the sum of per-tid data for remaining threads, the value now sent is the
  true total accumalated statistics for all threads that are/were part of
  the thread group.

The patch also addresses a minor issue where failure of one accounting
subsystem to fill in the taskstats structure was causing the send of
taskstats to not be sent at all.

The patch has been tested for stability and run cerberus for over 4 hours
on an SMP.

[akpm@osdl.org: bugfixes]
Signed-off-by: Shailabh Nagar <nagar@watson.ibm.com>
Signed-off-by: Balbir Singh <balbir@in.ibm.com>
Cc: Jay Lan <jlan@engr.sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

Documentation/accounting/delay-accounting.txt

@@ -48,9 +48,10 @@ counter (say cpu_delay_total) for a task will give the delay
 experienced by the task waiting for the corresponding resource
 in that interval.
 
-When a task exits, records containing the per-task and per-process statistics
+When a task exits, records containing the per-task statistics
-are sent to userspace without requiring a command. More details are given in
+are sent to userspace without requiring a command. If it is the last exiting
-the taskstats interface description.
+task of a thread group, the per-tgid statistics are also sent. More details
+are given in the taskstats interface description.
 
 The getdelays.c userspace utility in this directory allows simple commands to
 be run and the corresponding delay statistics to be displayed. It also serves
@@ -107,9 +108,3 @@ IO	count	delay total
 	0	0
 MEM	count	delay total
 	0	0
-
-
-
-
-
-

Documentation/accounting/taskstats.txt

@@ -32,12 +32,11 @@ The response contains statistics for a task (if pid is specified) or the sum of
 statistics for all tasks of the process (if tgid is specified).
 
 To obtain statistics for tasks which are exiting, userspace opens a multicast
-netlink socket. Each time a task exits, two records are sent by the kernel to
+netlink socket. Each time a task exits, its per-pid statistics is always sent
-each listener on the multicast socket. The first the per-pid task's statistics
+by the kernel to each listener on the multicast socket. In addition, if it is
-and the second is the sum for all tasks of the process to which the task
+the last thread exiting its thread group, an additional record containing the
-belongs (the task does not need to be the thread group leader). The need for
+per-tgid stats are also sent. The latter contains the sum of per-pid stats for
-per-tgid stats to be sent for each exiting task is explained in the per-tgid
+all threads in the thread group, both past and present.
-stats section below.
 
 getdelays.c is a simple utility demonstrating usage of the taskstats interface
 for reporting delay accounting statistics.
@@ -104,20 +103,14 @@ stats in userspace alone is inefficient and potentially inaccurate (due to lack
 of atomicity).
 
 However, maintaining per-process, in addition to per-task stats, within the
-kernel has space and time overheads. Hence the taskstats implementation
+kernel has space and time overheads. To address this, the taskstats code
-dynamically sums up the per-task stats for each task belonging to a process
+accumalates each exiting task's statistics into a process-wide data structure.
-whenever per-process stats are needed.
+When the last task of a process exits, the process level data accumalated also
+gets sent to userspace (along with the per-task data).
 
-Not maintaining per-tgid stats creates a problem when userspace is interested
+When a user queries to get per-tgid data, the sum of all other live threads in
-in getting these stats when the process dies i.e. the last thread of
+the group is added up and added to the accumalated total for previously exited
-a process exits. It isn't possible to simply return some aggregated per-process
+threads of the same thread group.
-statistic from the kernel.
-
-The approach taken by taskstats is to return the per-tgid stats *each* time
-a task exits, in addition to the per-pid stats for that task. Userspace can
-maintain task<->process mappings and use them to maintain the per-process stats
-in userspace, updating the aggregate appropriately as the tasks of a process
-exit.
 
 Extending taskstats
 -------------------

MAINTAINERS

@@ -2240,6 +2240,12 @@ M:	tsbogend@alpha.franken.de
 L:	netdev@vger.kernel.org
 S:	Maintained
 
+PER-TASK DELAY ACCOUNTING
+P:	Shailabh Nagar
+M:	nagar@watson.ibm.com
+L:	linux-kernel@vger.kernel.org
+S:	Maintained
+
 PERSONALITY HANDLING
 P:	Christoph Hellwig
 M:	hch@infradead.org
@@ -2767,6 +2773,12 @@ P:	Deepak Saxena
 M:	dsaxena@plexity.net
 S:	Maintained
 
+TASKSTATS STATISTICS INTERFACE
+P:	Shailabh Nagar
+M:	nagar@watson.ibm.com
+L:	linux-kernel@vger.kernel.org
+S:	Maintained
+
 TI PARALLEL LINK CABLE DRIVER
 P:     Romain Lievin
 M:     roms@lpg.ticalc.org

include/linux/sched.h

@@ -463,6 +463,10 @@ struct signal_struct {
 #ifdef CONFIG_BSD_PROCESS_ACCT
 	struct pacct_struct pacct;	/* per-process accounting information */
 #endif
+#ifdef CONFIG_TASKSTATS
+	spinlock_t stats_lock;
+	struct taskstats *stats;
+#endif
 };
 
 /* Context switch must be unlocked if interrupts are to be enabled */

include/linux/taskstats_kern.h

@@ -19,36 +19,75 @@ enum {
 extern kmem_cache_t *taskstats_cache;
 extern struct mutex taskstats_exit_mutex;
 
-static inline void taskstats_exit_alloc(struct taskstats **ptidstats,
+static inline void taskstats_exit_alloc(struct taskstats **ptidstats)
-					struct taskstats **ptgidstats)
 {
 	*ptidstats = kmem_cache_zalloc(taskstats_cache, SLAB_KERNEL);
-	*ptgidstats = kmem_cache_zalloc(taskstats_cache, SLAB_KERNEL);
 }
 
-static inline void taskstats_exit_free(struct taskstats *tidstats,
+static inline void taskstats_exit_free(struct taskstats *tidstats)
-					struct taskstats *tgidstats)
 {
 	if (tidstats)
 		kmem_cache_free(taskstats_cache, tidstats);
-	if (tgidstats)
-		kmem_cache_free(taskstats_cache, tgidstats);
 }
 
-extern void taskstats_exit_send(struct task_struct *, struct taskstats *,
+static inline void taskstats_tgid_init(struct signal_struct *sig)
-				struct taskstats *);
+{
-extern void taskstats_init_early(void);
+	spin_lock_init(&sig->stats_lock);
+	sig->stats = NULL;
+}
 
+static inline void taskstats_tgid_alloc(struct signal_struct *sig)
+{
+	struct taskstats *stats;
+	unsigned long flags;
+
+	stats = kmem_cache_zalloc(taskstats_cache, SLAB_KERNEL);
+	if (!stats)
+		return;
+
+	spin_lock_irqsave(&sig->stats_lock, flags);
+	if (!sig->stats) {
+		sig->stats = stats;
+		stats = NULL;
+	}
+	spin_unlock_irqrestore(&sig->stats_lock, flags);
+
+	if (stats)
+		kmem_cache_free(taskstats_cache, stats);
+}
+
+static inline void taskstats_tgid_free(struct signal_struct *sig)
+{
+	struct taskstats *stats = NULL;
+	unsigned long flags;
+
+	spin_lock_irqsave(&sig->stats_lock, flags);
+	if (sig->stats) {
+		stats = sig->stats;
+		sig->stats = NULL;
+	}
+	spin_unlock_irqrestore(&sig->stats_lock, flags);
+	if (stats)
+		kmem_cache_free(taskstats_cache, stats);
+}
+
+extern void taskstats_exit_send(struct task_struct *, struct taskstats *, int);
+extern void taskstats_init_early(void);
+extern void taskstats_tgid_alloc(struct signal_struct *);
 #else
-static inline void taskstats_exit_alloc(struct taskstats **ptidstats,
+static inline void taskstats_exit_alloc(struct taskstats **ptidstats)
-					struct taskstats **ptgidstats)
 {}
-static inline void taskstats_exit_free(struct taskstats *ptidstats,
+static inline void taskstats_exit_free(struct taskstats *ptidstats)
-					struct taskstats *ptgidstats)
 {}
 static inline void taskstats_exit_send(struct task_struct *tsk,
-					struct taskstats *tidstats,
+				       struct taskstats *tidstats,
-					struct taskstats *tgidstats)
+				       int group_dead)
+{}
+static inline void taskstats_tgid_init(struct signal_struct *sig)
+{}
+static inline void taskstats_tgid_alloc(struct signal_struct *sig)
+{}
+static inline void taskstats_tgid_free(struct signal_struct *sig)
 {}
 static inline void taskstats_init_early(void)
 {}

kernel/exit.c

@@ -845,7 +845,7 @@ static void exit_notify(struct task_struct *tsk)
 fastcall NORET_TYPE void do_exit(long code)
 {
 	struct task_struct *tsk = current;
-	struct taskstats *tidstats, *tgidstats;
+	struct taskstats *tidstats;
 	int group_dead;
 
 	profile_task_exit(tsk);
@@ -884,7 +884,7 @@ fastcall NORET_TYPE void do_exit(long code)
 				current->comm, current->pid,
 				preempt_count());
 
-	taskstats_exit_alloc(&tidstats, &tgidstats);
+	taskstats_exit_alloc(&tidstats);
 
 	acct_update_integrals(tsk);
 	if (tsk->mm) {
@@ -905,8 +905,8 @@ fastcall NORET_TYPE void do_exit(long code)
 #endif
 	if (unlikely(tsk->audit_context))
 		audit_free(tsk);
-	taskstats_exit_send(tsk, tidstats, tgidstats);
+	taskstats_exit_send(tsk, tidstats, group_dead);
-	taskstats_exit_free(tidstats, tgidstats);
+	taskstats_exit_free(tidstats);
 	delayacct_tsk_exit(tsk);
 
 	exit_mm(tsk);

kernel/fork.c

@@ -44,6 +44,7 @@
 #include <linux/acct.h>
 #include <linux/cn_proc.h>
 #include <linux/delayacct.h>
+#include <linux/taskstats_kern.h>
 
 #include <asm/pgtable.h>
 #include <asm/pgalloc.h>
@@ -819,6 +820,7 @@ static inline int copy_signal(unsigned long clone_flags, struct task_struct * ts
 	if (clone_flags & CLONE_THREAD) {
 		atomic_inc(&current->signal->count);
 		atomic_inc(&current->signal->live);
+		taskstats_tgid_alloc(current->signal);
 		return 0;
 	}
 	sig = kmem_cache_alloc(signal_cachep, GFP_KERNEL);
@@ -863,6 +865,7 @@ static inline int copy_signal(unsigned long clone_flags, struct task_struct * ts
 	INIT_LIST_HEAD(&sig->cpu_timers[0]);
 	INIT_LIST_HEAD(&sig->cpu_timers[1]);
 	INIT_LIST_HEAD(&sig->cpu_timers[2]);
+	taskstats_tgid_init(sig);
 
 	task_lock(current->group_leader);
 	memcpy(sig->rlim, current->signal->rlim, sizeof sig->rlim);
@@ -884,6 +887,7 @@ static inline int copy_signal(unsigned long clone_flags, struct task_struct * ts
 void __cleanup_signal(struct signal_struct *sig)
 {
 	exit_thread_group_keys(sig);
+	taskstats_tgid_free(sig);
 	kmem_cache_free(signal_cachep, sig);
 }
 

kernel/taskstats.c

@@ -132,46 +132,79 @@ static int fill_pid(pid_t pid, struct task_struct *pidtsk,
 static int fill_tgid(pid_t tgid, struct task_struct *tgidtsk,
 		struct taskstats *stats)
 {
-	int rc;
 	struct task_struct *tsk, *first;
+	unsigned long flags;
 
+	/*
+	 * Add additional stats from live tasks except zombie thread group
+	 * leaders who are already counted with the dead tasks
+	 */
 	first = tgidtsk;
-	read_lock(&tasklist_lock);
 	if (!first) {
+		read_lock(&tasklist_lock);
 		first = find_task_by_pid(tgid);
 		if (!first) {
 			read_unlock(&tasklist_lock);
 			return -ESRCH;
 		}
-	}
+		get_task_struct(first);
+		read_unlock(&tasklist_lock);
+	} else
+		get_task_struct(first);
+
+	/* Start with stats from dead tasks */
+	spin_lock_irqsave(&first->signal->stats_lock, flags);
+	if (first->signal->stats)
+		memcpy(stats, first->signal->stats, sizeof(*stats));
+	spin_unlock_irqrestore(&first->signal->stats_lock, flags);
+
 	tsk = first;
+	read_lock(&tasklist_lock);
 	do {
+		if (tsk->exit_state == EXIT_ZOMBIE && thread_group_leader(tsk))
+			continue;
 		/*
-		 * Each accounting subsystem adds calls its functions to
+		 * Accounting subsystem can call its functions here to
 		 * fill in relevant parts of struct taskstsats as follows
 		 *
-		 *	rc = per-task-foo(stats, tsk);
+		 *	per-task-foo(stats, tsk);
-		 *	if (rc)
-		 *		break;
 		 */
-
+		delayacct_add_tsk(stats, tsk);
-		rc = delayacct_add_tsk(stats, tsk);
-		if (rc)
-			break;
 
 	} while_each_thread(first, tsk);
 	read_unlock(&tasklist_lock);
 	stats->version = TASKSTATS_VERSION;
 
-
 	/*
-	 * Accounting subsytems can also add calls here if they don't
+	 * Accounting subsytems can also add calls here to modify
-	 * wish to aggregate statistics for per-tgid stats
+	 * fields of taskstats.
 	 */
 
-	return rc;
+	return 0;
 }
 
+
+static void fill_tgid_exit(struct task_struct *tsk)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&tsk->signal->stats_lock, flags);
+	if (!tsk->signal->stats)
+		goto ret;
+
+	/*
+	 * Each accounting subsystem calls its functions here to
+	 * accumalate its per-task stats for tsk, into the per-tgid structure
+	 *
+	 *	per-task-foo(tsk->signal->stats, tsk);
+	 */
+	delayacct_add_tsk(tsk->signal->stats, tsk);
+ret:
+	spin_unlock_irqrestore(&tsk->signal->stats_lock, flags);
+	return;
+}
+
+
 static int taskstats_send_stats(struct sk_buff *skb, struct genl_info *info)
 {
 	int rc = 0;
@@ -230,7 +263,7 @@ err:
 
 /* Send pid data out on exit */
 void taskstats_exit_send(struct task_struct *tsk, struct taskstats *tidstats,
-			struct taskstats *tgidstats)
+			int group_dead)
 {
 	int rc;
 	struct sk_buff *rep_skb;
@@ -238,13 +271,16 @@ void taskstats_exit_send(struct task_struct *tsk, struct taskstats *tidstats,
 	size_t size;
 	int is_thread_group;
 	struct nlattr *na;
+	unsigned long flags;
 
 	if (!family_registered || !tidstats)
 		return;
 
-	is_thread_group = !thread_group_empty(tsk);
+	spin_lock_irqsave(&tsk->signal->stats_lock, flags);
-	rc = 0;
+	is_thread_group = tsk->signal->stats ? 1 : 0;
+	spin_unlock_irqrestore(&tsk->signal->stats_lock, flags);
 
+	rc = 0;
 	/*
 	 * Size includes space for nested attributes
 	 */
@@ -268,30 +304,28 @@ void taskstats_exit_send(struct task_struct *tsk, struct taskstats *tidstats,
 			*tidstats);
 	nla_nest_end(rep_skb, na);
 
-	if (!is_thread_group || !tgidstats) {
+	if (!is_thread_group)
-		send_reply(rep_skb, 0, TASKSTATS_MSG_MULTICAST);
+		goto send;
-		goto ret;
-	}
 
-	rc = fill_tgid(tsk->pid, tsk, tgidstats);
 	/*
-	 * If fill_tgid() failed then one probable reason could be that the
+	 * tsk has/had a thread group so fill the tsk->signal->stats structure
-	 * thread group leader has exited. fill_tgid() will fail, send out
+	 * Doesn't matter if tsk is the leader or the last group member leaving
-	 * the pid statistics collected earlier.
 	 */
-	if (rc < 0) {
+
-		send_reply(rep_skb, 0, TASKSTATS_MSG_MULTICAST);
+	fill_tgid_exit(tsk);
-		goto ret;
+	if (!group_dead)
-	}
+		goto send;
 
 	na = nla_nest_start(rep_skb, TASKSTATS_TYPE_AGGR_TGID);
 	NLA_PUT_U32(rep_skb, TASKSTATS_TYPE_TGID, (u32)tsk->tgid);
+	/* No locking needed for tsk->signal->stats since group is dead */
 	NLA_PUT_TYPE(rep_skb, struct taskstats, TASKSTATS_TYPE_STATS,
-			*tgidstats);
+			*tsk->signal->stats);
 	nla_nest_end(rep_skb, na);
 
+send:
 	send_reply(rep_skb, 0, TASKSTATS_MSG_MULTICAST);
-	goto ret;
+	return;
 
 nla_put_failure:
 	genlmsg_cancel(rep_skb, reply);