Commit ed27f912 authored by Davidlohr Bueso's avatar Davidlohr Bueso Committed by Linus Torvalds

ipc/msg: avoid waking sender upon full queue

Blocked tasks queued in q_senders waiting for their message to fit in the
queue are blindly awoken every time we think there's a remote chance this
might happen.  This could cause numerous (and expensive -- thundering
herd-ish) bogus wakeups if the queue is still really full.  Adding to the
scheduling cost/overhead, there's also the fact that we need to take the
ipc object lock and requeue ourselves in the q_senders list.

By keeping track of the blocked sender's message size, we can know
previously if the wakeup ought to occur or not.  Otherwise, to maintain
the current wakeup order we just move it to the tail.  This is exactly
what occurs right now if the sender needs to go back to sleep.

The case of EIDRM is left completely untouched, as we need to wakeup all
the tasks, and shouldn't be playing games in the first place.

This patch was seen to save on the 'msgctl10' ltp testcase ~15% in context
switches (avg out of ten runs).  Although these tests are really about
functionality (as opposed to performance), is does show the direct
benefits of the optimization.

[ coding-style fixes]
Link: default avatarDavidlohr Bueso <>
Acked-by: default avatarPeter Zijlstra (Intel) <>
Cc: Manfred Spraul <>
Cc: Sebastian Andrzej Siewior <>
Signed-off-by: default avatarAndrew Morton <>
Signed-off-by: default avatarLinus Torvalds <>
parent d0d6a2a9
......@@ -58,6 +58,7 @@ struct msg_receiver {
struct msg_sender {
struct list_head list;
struct task_struct *tsk;
size_t msgsz;
#define SEARCH_ANY 1
......@@ -153,27 +154,60 @@ static int newque(struct ipc_namespace *ns, struct ipc_params *params)
return msq->;
static inline void ss_add(struct msg_queue *msq, struct msg_sender *mss)
static inline bool msg_fits_inqueue(struct msg_queue *msq, size_t msgsz)
return msgsz + msq->q_cbytes <= msq->q_qbytes &&
1 + msq->q_qnum <= msq->q_qbytes;
static inline void ss_add(struct msg_queue *msq,
struct msg_sender *mss, size_t msgsz)
mss->tsk = current;
mss->msgsz = msgsz;
list_add_tail(&mss->list, &msq->q_senders);
static inline void ss_del(struct msg_sender *mss)
if (mss-> != NULL)
if (mss->
static void ss_wakeup(struct list_head *h,
static void ss_wakeup(struct msg_queue *msq,
struct wake_q_head *wake_q, bool kill)
struct msg_sender *mss, *t;
struct task_struct *stop_tsk = NULL;
struct list_head *h = &msq->q_senders;
list_for_each_entry_safe(mss, t, h, list) {
if (kill)
mss-> = NULL;
* Stop at the first task we don't wakeup,
* we've already iterated the original
* sender queue.
else if (stop_tsk == mss->tsk)
* We are not in an EIDRM scenario here, therefore
* verify that we really need to wakeup the task.
* To maintain current semantics and wakeup order,
* move the sender to the tail on behalf of the
* blocked task.
else if (!msg_fits_inqueue(msq, mss->msgsz)) {
if (!stop_tsk)
stop_tsk = mss->tsk;
list_move_tail(&mss->list, &msq->q_senders);
wake_q_add(wake_q, mss->tsk);
......@@ -204,7 +238,7 @@ static void freeque(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)
expunge_all(msq, -EIDRM, &wake_q);
ss_wakeup(&msq->q_senders, &wake_q, true);
ss_wakeup(msq, &wake_q, true);
msg_rmid(ns, msq);
......@@ -388,7 +422,7 @@ static int msgctl_down(struct ipc_namespace *ns, int msqid, int cmd,
* Sleeping senders might be able to send
* due to a larger queue size.
ss_wakeup(&msq->q_senders, &wake_q, false);
ss_wakeup(msq, &wake_q, false);
......@@ -642,10 +676,8 @@ long do_msgsnd(int msqid, long mtype, void __user *mtext,
if (err)
goto out_unlock0;
if (msgsz + msq->q_cbytes <= msq->q_qbytes &&
1 + msq->q_qnum <= msq->q_qbytes) {
if (msg_fits_inqueue(msq, msgsz))
/* queue full, wait: */
if (msgflg & IPC_NOWAIT) {
......@@ -654,7 +686,7 @@ long do_msgsnd(int msqid, long mtype, void __user *mtext,
/* enqueue the sender and prepare to block */
ss_add(msq, &s);
ss_add(msq, &s, msgsz);
if (!ipc_rcu_getref(msq)) {
err = -EIDRM;
......@@ -682,6 +714,7 @@ long do_msgsnd(int msqid, long mtype, void __user *mtext,
msq->q_lspid = task_tgid_vnr(current);
msq->q_stime = get_seconds();
......@@ -882,7 +915,7 @@ long do_msgrcv(int msqid, void __user *buf, size_t bufsz, long msgtyp, int msgfl
msq->q_cbytes -= msg->m_ts;
atomic_sub(msg->m_ts, &ns->msg_bytes);
ss_wakeup(&msq->q_senders, &wake_q, false);
ss_wakeup(msq, &wake_q, false);
goto out_unlock0;
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment