The current TP futexes has no guarantee that the top waiter
(serialization mutex owner) can get the lock within a finite time.
As a result, lock starvation can happen.
A lock handoff mechanism is added to the TP futexes to prevent lock
starvation from happening. The idea is that the top waiter can set a
special handoff_pid variable with its own pid. The futex owner will
then check this variable at unlock time to see if it should free the
futex or change its value to the designated pid to transfer the lock
directly to the top waiter.
This change does have the effect of limiting the amount of unfairness
and hence may adversely affect performance depending on the workloads.
Using a futex locking microbenchmark on a 4-socket 72-core 144-thread
Haswell-EX system running 4.8-rc7 based kernel, the performance data
for the TP futexes before and after the patch were as follows:
CS length Before patch After patch % change
--------- ------------ ----------- --------
5 54,622,973 59,455,380 +8.8%
10 42,577,283 41,161,387 -3.3%
20 29,424,333 37,718,989 +28.2%
30 33,664,095 29,474,986 -12.4%
40 27,998,446 24,774,389 -11.5%
50 25,793,991 22,718,432 -11.9%
1us sleep 179,067 177,662 -0.8%
In fact, when the critical section is a 1us sleep, the wait-wake
futexes perform a bit better than the TP futexes as the former futexes
are more unfair in this case. So lock starvation can happen with
wait-wake futexes.
WW futex TP futex
-------- --------
Total locking ops 175,824 177,662
Per-thread avg/sec 122 123
Per-thread min/sec 2 106
Per-thread max/sec 498 199
% Stddev 6.26% 0.97%
Signed-off-by: Waiman Long <Waiman.Long@xxxxxxx>
---
kernel/futex.c | 55 ++++++++++++++++++++++++++++++++++++++++++++++++-------
1 files changed, 48 insertions(+), 7 deletions(-)
diff --git a/kernel/futex.c b/kernel/futex.c
index 5fa9a10..d260410 100644
--- a/kernel/futex.c
+++ b/kernel/futex.c
@@ -228,6 +228,8 @@ struct futex_state {
struct task_struct *owner;
atomic_t refcount;
+ u32 handoff_pid; /* For TP futexes only */
+
enum futex_type type;
union futex_key key;
};
@@ -898,6 +900,7 @@ static int refill_futex_state_cache(void)
/* pi_mutex gets initialized later */
state->owner = NULL;
+ state->handoff_pid = 0;
atomic_set(&state->refcount, 1);
state->key = FUTEX_KEY_INIT;
@@ -3303,8 +3306,9 @@ void exit_robust_list(struct task_struct *curr)
*
* Within the kernel, trylocks are done ignoring the FUTEX_WAITERS bit. The
* purpose of this FUTEX_WAITERS bit is to make the unlocker wake up the
- * serialization mutex owner. Not having the FUTEX_WAITERS bit set doesn't
- * mean there is no waiter in the kernel.
+ * serialization mutex owner or the hand off the lock directly to the top
+ * waiter. Not having the FUTEX_WAITERS bit set doesn't mean there is no
+ * waiter in the kernel.
*
* Like PI futexes, TP futexes are orthogonal to robust futexes can be
* used together.
@@ -3320,6 +3324,19 @@ void exit_robust_list(struct task_struct *curr)
* the ownership of the futex.
*/
+/*
+ * Spinning and sleeping thresholds for enabling lock handoff and prevent
+ * lock starvation.
+ *
+ * The current threshold values setting is kind of arbitrary. Setting them
+ * too low will impact the amount of lock stealing possible thus reducing
+ * performance. Setting them too high may make the top waiter wait too long
+ * before getting the lock. The current set of values are a bit on the
+ * high side to ensure adequate performance.
+ */
+#define TP_SPIN_THRESHOLD (1 << 14)
+#define TP_SLEEP_DECREMENT (TP_SPIN_THRESHOLD/64)
+
/**
* lookup_futex_state - Looking up the futex state structure.
* @hb: hash bucket
@@ -3398,6 +3415,7 @@ static inline int put_futex_state_unlocked(struct futex_state *state)
* If waiter is true
* then
* don't preserve the flag bits;
+ * check for handoff (futex word == own pid)
* else
* preserve the flag bits
* endif
@@ -3416,6 +3434,9 @@ static inline int futex_trylock(u32 __user *uaddr, u32 vpid, u32 *puval,
*puval = uval;
+ if (waiter && (uval & FUTEX_TID_MASK) == vpid)
+ return 1;
+
if (uval & FUTEX_TID_MASK)
return 0; /* Trylock fails */
@@ -3478,7 +3499,7 @@ static inline int futex_set_waiters_bit(u32 __user *uaddr, u32 *puval)
static int futex_spin_on_owner(u32 __user *uaddr, u32 vpid,
struct futex_state *state)
{
- int ret;
+ int ret, loop = TP_SPIN_THRESHOLD;
u32 uval;
u32 owner_pid = 0;
struct task_struct *owner_task = NULL;
@@ -3486,7 +3507,7 @@ static int futex_spin_on_owner(u32 __user *uaddr, u32 vpid,
WRITE_ONCE(state->owner, current);
preempt_disable();
- for (;;) {
+ for (;; loop--) {
ret = futex_trylock(uaddr, vpid, &uval, true);
if (ret)
break;
@@ -3544,6 +3565,18 @@ static int futex_spin_on_owner(u32 __user *uaddr, u32 vpid,
break;
}
+ /*
+ * Enable lock handoff if the threshold reaches 0.
+ * We also need to set the FUTEX_WAITERS bit to make sure
+ * that futex lock holder will initiate the handoff at
+ * unlock time.
+ */
+ if ((loop <= 0) && !READ_ONCE(state->handoff_pid)) {
+ WRITE_ONCE(state->handoff_pid, vpid);
+ if (futex_set_waiters_bit(uaddr, &uval) < 0)
+ goto efault;
+ }
+
if (owner_task->on_cpu) {
cpu_relax();
continue;
@@ -3587,8 +3620,10 @@ static int futex_spin_on_owner(u32 __user *uaddr, u32 vpid,
* lock stealing happen in between setting the FUTEX_WAITERS
* and setting state to TASK_INTERRUPTIBLE.
*/
- if (!(uval & FUTEX_OWNER_DIED) && (uval & FUTEX_WAITERS))
+ if (!(uval & FUTEX_OWNER_DIED) && (uval & FUTEX_WAITERS)) {
schedule_preempt_disabled();
+ loop -= TP_SLEEP_DECREMENT;
+ }
__set_current_state(TASK_RUNNING);
}
out:
@@ -3605,6 +3640,7 @@ out:
* Cleanup futex state.
*/
WRITE_ONCE(state->owner, NULL);
+ WRITE_ONCE(state->handoff_pid, 0);
return ret;
efault:
@@ -3728,6 +3764,7 @@ out:
static int futex_unlock(u32 __user *uaddr, unsigned int flags)
{
u32 uval, vpid = task_pid_vnr(current);
+ u32 newpid = 0;
union futex_key key = FUTEX_KEY_INIT;
struct futex_hash_bucket *hb;
struct futex_state *state = NULL;
@@ -3761,6 +3798,10 @@ static int futex_unlock(u32 __user *uaddr, unsigned int flags)
goto out_put_key;
}
+ newpid = READ_ONCE(state->handoff_pid);
+ if (newpid)
+ WRITE_ONCE(state->handoff_pid, 0);
+
owner = READ_ONCE(state->owner);
if (owner)
wake_q_add(&wake_q, owner);
@@ -3768,13 +3809,13 @@ static int futex_unlock(u32 __user *uaddr, unsigned int flags)
spin_unlock(&hb->fs_lock);
/*
- * Unlock the futex.
+ * Unlock the futex or handoff to the next owner.
* The flag bits are not preserved to encourage more lock stealing.
*/
for (;;) {
u32 old = uval;
- if (cmpxchg_futex_value(&uval, uaddr, old, 0)) {
+ if (cmpxchg_futex_value(&uval, uaddr, old, newpid)) {
ret = -EFAULT;
break;
}
--
1.7.1
--
To unsubscribe from this list: send the line "unsubscribe linux-doc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
