![Davidlohr Bueso](/assets/img/avatar_default.png)
Both mutexes and rwsems took a performance hit when we switched over from the original mcs code to the cancelable variant (osq). The reason being the use of smp_load_acquire() when polling for node->locked. This is not needed as reordering is not an issue, as such, relax the barrier semantics. Paul describes the scenario nicely: https://lkml.org/lkml/2013/11/19/405 - If we start polling before the insertion is complete, all that happens is that the first few polls have no chance of seeing a lock grant. - Ordering the polling against the initialization -- the above xchg() is already doing that for us. The smp_load_acquire() when unqueuing make sense. In addition, we don't need to worry about leaking the critical region as osq is only used internally. This impacts both regular and large levels of concurrency, ie on a 40 core system with a disk intensive workload: disk-1 804.83 ( 0.00%) 828.16 ( 2.90%) disk-61 8063.45 ( 0.00%) 18181.82 (125.48%) disk-121 7187.41 ( 0.00%) 20119.17 (179.92%) disk-181 6933.32 ( 0.00%) 20509.91 (195.82%) disk-241 6850.81 ( 0.00%) 20397.80 (197.74%) disk-301 6815.22 ( 0.00%) 20287.58 (197.68%) disk-361 7080.40 ( 0.00%) 20205.22 (185.37%) disk-421 7076.13 ( 0.00%) 19957.33 (182.04%) disk-481 7083.25 ( 0.00%) 19784.06 (179.31%) disk-541 7038.39 ( 0.00%) 19610.92 (178.63%) disk-601 7072.04 ( 0.00%) 19464.53 (175.23%) disk-661 7010.97 ( 0.00%) 19348.23 (175.97%) disk-721 7069.44 ( 0.00%) 19255.33 (172.37%) disk-781 7007.58 ( 0.00%) 19103.14 (172.61%) disk-841 6981.18 ( 0.00%) 18964.22 (171.65%) disk-901 6968.47 ( 0.00%) 18826.72 (170.17%) disk-961 6964.61 ( 0.00%) 18708.02 (168.62%) Signed-off-by: Davidlohr Bueso <dbueso@suse.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1420573509-24774-7-git-send-email-dave@stgolabs.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
204 lines
4.9 KiB
C
204 lines
4.9 KiB
C
#include <linux/percpu.h>
|
|
#include <linux/sched.h>
|
|
#include <linux/osq_lock.h>
|
|
|
|
/*
|
|
* An MCS like lock especially tailored for optimistic spinning for sleeping
|
|
* lock implementations (mutex, rwsem, etc).
|
|
*
|
|
* Using a single mcs node per CPU is safe because sleeping locks should not be
|
|
* called from interrupt context and we have preemption disabled while
|
|
* spinning.
|
|
*/
|
|
static DEFINE_PER_CPU_SHARED_ALIGNED(struct optimistic_spin_node, osq_node);
|
|
|
|
/*
|
|
* We use the value 0 to represent "no CPU", thus the encoded value
|
|
* will be the CPU number incremented by 1.
|
|
*/
|
|
static inline int encode_cpu(int cpu_nr)
|
|
{
|
|
return cpu_nr + 1;
|
|
}
|
|
|
|
static inline struct optimistic_spin_node *decode_cpu(int encoded_cpu_val)
|
|
{
|
|
int cpu_nr = encoded_cpu_val - 1;
|
|
|
|
return per_cpu_ptr(&osq_node, cpu_nr);
|
|
}
|
|
|
|
/*
|
|
* Get a stable @node->next pointer, either for unlock() or unqueue() purposes.
|
|
* Can return NULL in case we were the last queued and we updated @lock instead.
|
|
*/
|
|
static inline struct optimistic_spin_node *
|
|
osq_wait_next(struct optimistic_spin_queue *lock,
|
|
struct optimistic_spin_node *node,
|
|
struct optimistic_spin_node *prev)
|
|
{
|
|
struct optimistic_spin_node *next = NULL;
|
|
int curr = encode_cpu(smp_processor_id());
|
|
int old;
|
|
|
|
/*
|
|
* If there is a prev node in queue, then the 'old' value will be
|
|
* the prev node's CPU #, else it's set to OSQ_UNLOCKED_VAL since if
|
|
* we're currently last in queue, then the queue will then become empty.
|
|
*/
|
|
old = prev ? prev->cpu : OSQ_UNLOCKED_VAL;
|
|
|
|
for (;;) {
|
|
if (atomic_read(&lock->tail) == curr &&
|
|
atomic_cmpxchg(&lock->tail, curr, old) == curr) {
|
|
/*
|
|
* We were the last queued, we moved @lock back. @prev
|
|
* will now observe @lock and will complete its
|
|
* unlock()/unqueue().
|
|
*/
|
|
break;
|
|
}
|
|
|
|
/*
|
|
* We must xchg() the @node->next value, because if we were to
|
|
* leave it in, a concurrent unlock()/unqueue() from
|
|
* @node->next might complete Step-A and think its @prev is
|
|
* still valid.
|
|
*
|
|
* If the concurrent unlock()/unqueue() wins the race, we'll
|
|
* wait for either @lock to point to us, through its Step-B, or
|
|
* wait for a new @node->next from its Step-C.
|
|
*/
|
|
if (node->next) {
|
|
next = xchg(&node->next, NULL);
|
|
if (next)
|
|
break;
|
|
}
|
|
|
|
cpu_relax_lowlatency();
|
|
}
|
|
|
|
return next;
|
|
}
|
|
|
|
bool osq_lock(struct optimistic_spin_queue *lock)
|
|
{
|
|
struct optimistic_spin_node *node = this_cpu_ptr(&osq_node);
|
|
struct optimistic_spin_node *prev, *next;
|
|
int curr = encode_cpu(smp_processor_id());
|
|
int old;
|
|
|
|
node->locked = 0;
|
|
node->next = NULL;
|
|
node->cpu = curr;
|
|
|
|
old = atomic_xchg(&lock->tail, curr);
|
|
if (old == OSQ_UNLOCKED_VAL)
|
|
return true;
|
|
|
|
prev = decode_cpu(old);
|
|
node->prev = prev;
|
|
ACCESS_ONCE(prev->next) = node;
|
|
|
|
/*
|
|
* Normally @prev is untouchable after the above store; because at that
|
|
* moment unlock can proceed and wipe the node element from stack.
|
|
*
|
|
* However, since our nodes are static per-cpu storage, we're
|
|
* guaranteed their existence -- this allows us to apply
|
|
* cmpxchg in an attempt to undo our queueing.
|
|
*/
|
|
|
|
while (!ACCESS_ONCE(node->locked)) {
|
|
/*
|
|
* If we need to reschedule bail... so we can block.
|
|
*/
|
|
if (need_resched())
|
|
goto unqueue;
|
|
|
|
cpu_relax_lowlatency();
|
|
}
|
|
return true;
|
|
|
|
unqueue:
|
|
/*
|
|
* Step - A -- stabilize @prev
|
|
*
|
|
* Undo our @prev->next assignment; this will make @prev's
|
|
* unlock()/unqueue() wait for a next pointer since @lock points to us
|
|
* (or later).
|
|
*/
|
|
|
|
for (;;) {
|
|
if (prev->next == node &&
|
|
cmpxchg(&prev->next, node, NULL) == node)
|
|
break;
|
|
|
|
/*
|
|
* We can only fail the cmpxchg() racing against an unlock(),
|
|
* in which case we should observe @node->locked becomming
|
|
* true.
|
|
*/
|
|
if (smp_load_acquire(&node->locked))
|
|
return true;
|
|
|
|
cpu_relax_lowlatency();
|
|
|
|
/*
|
|
* Or we race against a concurrent unqueue()'s step-B, in which
|
|
* case its step-C will write us a new @node->prev pointer.
|
|
*/
|
|
prev = ACCESS_ONCE(node->prev);
|
|
}
|
|
|
|
/*
|
|
* Step - B -- stabilize @next
|
|
*
|
|
* Similar to unlock(), wait for @node->next or move @lock from @node
|
|
* back to @prev.
|
|
*/
|
|
|
|
next = osq_wait_next(lock, node, prev);
|
|
if (!next)
|
|
return false;
|
|
|
|
/*
|
|
* Step - C -- unlink
|
|
*
|
|
* @prev is stable because its still waiting for a new @prev->next
|
|
* pointer, @next is stable because our @node->next pointer is NULL and
|
|
* it will wait in Step-A.
|
|
*/
|
|
|
|
ACCESS_ONCE(next->prev) = prev;
|
|
ACCESS_ONCE(prev->next) = next;
|
|
|
|
return false;
|
|
}
|
|
|
|
void osq_unlock(struct optimistic_spin_queue *lock)
|
|
{
|
|
struct optimistic_spin_node *node, *next;
|
|
int curr = encode_cpu(smp_processor_id());
|
|
|
|
/*
|
|
* Fast path for the uncontended case.
|
|
*/
|
|
if (likely(atomic_cmpxchg(&lock->tail, curr, OSQ_UNLOCKED_VAL) == curr))
|
|
return;
|
|
|
|
/*
|
|
* Second most likely case.
|
|
*/
|
|
node = this_cpu_ptr(&osq_node);
|
|
next = xchg(&node->next, NULL);
|
|
if (next) {
|
|
ACCESS_ONCE(next->locked) = 1;
|
|
return;
|
|
}
|
|
|
|
next = osq_wait_next(lock, node, NULL);
|
|
if (next)
|
|
ACCESS_ONCE(next->locked) = 1;
|
|
}
|