[v2,4/4] eventdev: relax smp barriers with c11 atomics

Message ID 1593667604-12029-4-git-send-email-phil.yang@arm.com (mailing list archive)
State Superseded, archived
Delegated to: Jerin Jacob
Headers
Series [v2,1/4] eventdev: fix race condition on timer list counter |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/travis-robot warning Travis build: failed
ci/Intel-compilation success Compilation OK

Commit Message

Phil Yang July 2, 2020, 5:26 a.m. UTC
The implementation-specific opaque data is shared between arm and cancel
operations. The state flag acts as a guard variable to make sure the
update of opaque data is synchronized. This patch uses c11 atomics with
explicit one way memory barrier instead of full barriers rte_smp_w/rmb()
to synchronize the opaque data between timer arm and cancel threads.

Signed-off-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
v2:
1. Removed implementation-specific opaque data cleanup code.
2. Replaced thread fence with atomic ACQURE/RELEASE ordering on state access.

 lib/librte_eventdev/rte_event_timer_adapter.c | 55 ++++++++++++++++++---------
 lib/librte_eventdev/rte_event_timer_adapter.h |  2 +-
 2 files changed, 38 insertions(+), 19 deletions(-)
  

Comments

Carrillo, Erik G July 2, 2020, 8:30 p.m. UTC | #1
> -----Original Message-----
> From: Phil Yang <phil.yang@arm.com>
> Sent: Thursday, July 2, 2020 12:27 AM
> To: Carrillo, Erik G <erik.g.carrillo@intel.com>; dev@dpdk.org
> Cc: jerinj@marvell.com; Honnappa.Nagarahalli@arm.com;
> drc@linux.vnet.ibm.com; Ruifeng.Wang@arm.com;
> Dharmik.Thakkar@arm.com; nd@arm.com
> Subject: [PATCH v2 4/4] eventdev: relax smp barriers with c11 atomics
> 
> The implementation-specific opaque data is shared between arm and cancel
> operations. The state flag acts as a guard variable to make sure the update of
> opaque data is synchronized. This patch uses c11 atomics with explicit one
> way memory barrier instead of full barriers rte_smp_w/rmb() to synchronize
> the opaque data between timer arm and cancel threads.
> 
> Signed-off-by: Phil Yang <phil.yang@arm.com>
> Reviewed-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
> Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Acked-by: Erik Gabriel Carrillo <erik.g.carrillo@intel.com>
  
Jerin Jacob July 3, 2020, 10:50 a.m. UTC | #2
On Fri, Jul 3, 2020 at 2:00 AM Carrillo, Erik G
<erik.g.carrillo@intel.com> wrote:
>
> > -----Original Message-----
> > From: Phil Yang <phil.yang@arm.com>
> > Sent: Thursday, July 2, 2020 12:27 AM
> > To: Carrillo, Erik G <erik.g.carrillo@intel.com>; dev@dpdk.org
> > Cc: jerinj@marvell.com; Honnappa.Nagarahalli@arm.com;
> > drc@linux.vnet.ibm.com; Ruifeng.Wang@arm.com;
> > Dharmik.Thakkar@arm.com; nd@arm.com
> > Subject: [PATCH v2 4/4] eventdev: relax smp barriers with c11 atomics
> >
> > The implementation-specific opaque data is shared between arm and cancel
> > operations. The state flag acts as a guard variable to make sure the update of
> > opaque data is synchronized. This patch uses c11 atomics with explicit one
> > way memory barrier instead of full barriers rte_smp_w/rmb() to synchronize
> > the opaque data between timer arm and cancel threads.
> >
> > Signed-off-by: Phil Yang <phil.yang@arm.com>
> > Reviewed-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
> > Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
> Acked-by: Erik Gabriel Carrillo <erik.g.carrillo@intel.com>

Series applied to dpdk-next-eventdev/master. Thanks.
  
Thomas Monjalon July 6, 2020, 10:04 a.m. UTC | #3
02/07/2020 07:26, Phil Yang:
> The implementation-specific opaque data is shared between arm and cancel
> operations. The state flag acts as a guard variable to make sure the
> update of opaque data is synchronized. This patch uses c11 atomics with
> explicit one way memory barrier instead of full barriers rte_smp_w/rmb()
> to synchronize the opaque data between timer arm and cancel threads.

I think we should write C11 (uppercase).

Please, in your explanations, try to be more specific.
Naming fields may help to make things clear.

[...]
> --- a/lib/librte_eventdev/rte_event_timer_adapter.h
> +++ b/lib/librte_eventdev/rte_event_timer_adapter.h
> @@ -467,7 +467,7 @@ struct rte_event_timer {
>  	 *  - op: RTE_EVENT_OP_NEW
>  	 *  - event_type: RTE_EVENT_TYPE_TIMER
>  	 */
> -	volatile enum rte_event_timer_state state;
> +	enum rte_event_timer_state state;
>  	/**< State of the event timer. */

Why do you remove the volatile keyword?
It is not explained in the commit log.

This change is triggering a warning in the ABI check:
http://mails.dpdk.org/archives/test-report/2020-July/140440.html
Moving from volatile to non-volatile is probably not an issue.
I expect the code generated for the volatile case to work the same
in non-volatile case. Do you confirm?

In any case, we need an explanation and an ABI check exception.
  
Phil Yang July 6, 2020, 3:32 p.m. UTC | #4
> -----Original Message-----
> From: Thomas Monjalon <thomas@monjalon.net>
> Sent: Monday, July 6, 2020 6:04 PM
> To: Phil Yang <Phil.Yang@arm.com>
> Cc: erik.g.carrillo@intel.com; dev@dpdk.org; jerinj@marvell.com; Honnappa
> Nagarahalli <Honnappa.Nagarahalli@arm.com>; drc@linux.vnet.ibm.com;
> Ruifeng Wang <Ruifeng.Wang@arm.com>; Dharmik Thakkar
> <Dharmik.Thakkar@arm.com>; nd <nd@arm.com>;
> david.marchand@redhat.com; mdr@ashroe.eu; Neil Horman
> <nhorman@tuxdriver.com>; Dodji Seketeli <dodji@redhat.com>
> Subject: Re: [dpdk-dev] [PATCH v2 4/4] eventdev: relax smp barriers with c11
> atomics
> 
> 02/07/2020 07:26, Phil Yang:
> > The implementation-specific opaque data is shared between arm and
> cancel
> > operations. The state flag acts as a guard variable to make sure the
> > update of opaque data is synchronized. This patch uses c11 atomics with
> > explicit one way memory barrier instead of full barriers rte_smp_w/rmb()
> > to synchronize the opaque data between timer arm and cancel threads.
> 
> I think we should write C11 (uppercase).
Agreed. 
I will change it in the next version.

> 
> Please, in your explanations, try to be more specific.
> Naming fields may help to make things clear.
OK. Thanks.

> 
> [...]
> > --- a/lib/librte_eventdev/rte_event_timer_adapter.h
> > +++ b/lib/librte_eventdev/rte_event_timer_adapter.h
> > @@ -467,7 +467,7 @@ struct rte_event_timer {
> >  	 *  - op: RTE_EVENT_OP_NEW
> >  	 *  - event_type: RTE_EVENT_TYPE_TIMER
> >  	 */
> > -	volatile enum rte_event_timer_state state;
> > +	enum rte_event_timer_state state;
> >  	/**< State of the event timer. */
> 
> Why do you remove the volatile keyword?
> It is not explained in the commit log.
By using the C11 atomic operations, it will generate the same instructions for non-volatile and volatile version.
Please check the sample code here: https://gcc.godbolt.org/z/8x5rWs

> 
> This change is triggering a warning in the ABI check:
> http://mails.dpdk.org/archives/test-report/2020-July/140440.html
> Moving from volatile to non-volatile is probably not an issue.
> I expect the code generated for the volatile case to work the same
> in non-volatile case. Do you confirm?
They generate the same instructions, so either way will work.
Do I need to revert it to the volatile version?


Thanks,
Phil
> 
> In any case, we need an explanation and an ABI check exception.
>
  
Thomas Monjalon July 6, 2020, 3:40 p.m. UTC | #5
06/07/2020 17:32, Phil Yang:
> From: Thomas Monjalon <thomas@monjalon.net>
> > 02/07/2020 07:26, Phil Yang:
> > > --- a/lib/librte_eventdev/rte_event_timer_adapter.h
> > > +++ b/lib/librte_eventdev/rte_event_timer_adapter.h
> > > @@ -467,7 +467,7 @@ struct rte_event_timer {
> > >  	 *  - op: RTE_EVENT_OP_NEW
> > >  	 *  - event_type: RTE_EVENT_TYPE_TIMER
> > >  	 */
> > > -	volatile enum rte_event_timer_state state;
> > > +	enum rte_event_timer_state state;
> > >  	/**< State of the event timer. */
> > 
> > Why do you remove the volatile keyword?
> > It is not explained in the commit log.
> By using the C11 atomic operations, it will generate the same instructions for non-volatile and volatile version.
> Please check the sample code here: https://gcc.godbolt.org/z/8x5rWs
> 
> > This change is triggering a warning in the ABI check:
> > http://mails.dpdk.org/archives/test-report/2020-July/140440.html
> > Moving from volatile to non-volatile is probably not an issue.
> > I expect the code generated for the volatile case to work the same
> > in non-volatile case. Do you confirm?
> They generate the same instructions, so either way will work.
> Do I need to revert it to the volatile version?

Either you revert, or you add explanation in the commit log
+ exception in libabigail.abignore
  

Patch

diff --git a/lib/librte_eventdev/rte_event_timer_adapter.c b/lib/librte_eventdev/rte_event_timer_adapter.c
index 8909a8c..ca00258 100644
--- a/lib/librte_eventdev/rte_event_timer_adapter.c
+++ b/lib/librte_eventdev/rte_event_timer_adapter.c
@@ -629,7 +629,8 @@  swtim_callback(struct rte_timer *tim)
 		sw->expired_timers[sw->n_expired_timers++] = tim;
 		sw->stats.evtim_exp_count++;
 
-		evtim->state = RTE_EVENT_TIMER_NOT_ARMED;
+		__atomic_store_n(&evtim->state, RTE_EVENT_TIMER_NOT_ARMED,
+				__ATOMIC_RELEASE);
 	}
 
 	if (event_buffer_batch_ready(&sw->buffer)) {
@@ -1020,6 +1021,7 @@  __swtim_arm_burst(const struct rte_event_timer_adapter *adapter,
 	int n_lcores;
 	/* Timer list for this lcore is not in use. */
 	uint16_t exp_state = 0;
+	enum rte_event_timer_state n_state;
 
 #ifdef RTE_LIBRTE_EVENTDEV_DEBUG
 	/* Check that the service is running. */
@@ -1060,30 +1062,36 @@  __swtim_arm_burst(const struct rte_event_timer_adapter *adapter,
 	}
 
 	for (i = 0; i < nb_evtims; i++) {
-		/* Don't modify the event timer state in these cases */
-		if (evtims[i]->state == RTE_EVENT_TIMER_ARMED) {
+		n_state = __atomic_load_n(&evtims[i]->state, __ATOMIC_ACQUIRE);
+		if (n_state == RTE_EVENT_TIMER_ARMED) {
 			rte_errno = EALREADY;
 			break;
-		} else if (!(evtims[i]->state == RTE_EVENT_TIMER_NOT_ARMED ||
-			     evtims[i]->state == RTE_EVENT_TIMER_CANCELED)) {
+		} else if (!(n_state == RTE_EVENT_TIMER_NOT_ARMED ||
+			     n_state == RTE_EVENT_TIMER_CANCELED)) {
 			rte_errno = EINVAL;
 			break;
 		}
 
 		ret = check_timeout(evtims[i], adapter);
 		if (unlikely(ret == -1)) {
-			evtims[i]->state = RTE_EVENT_TIMER_ERROR_TOOLATE;
+			__atomic_store_n(&evtims[i]->state,
+					RTE_EVENT_TIMER_ERROR_TOOLATE,
+					__ATOMIC_RELAXED);
 			rte_errno = EINVAL;
 			break;
 		} else if (unlikely(ret == -2)) {
-			evtims[i]->state = RTE_EVENT_TIMER_ERROR_TOOEARLY;
+			__atomic_store_n(&evtims[i]->state,
+					RTE_EVENT_TIMER_ERROR_TOOEARLY,
+					__ATOMIC_RELAXED);
 			rte_errno = EINVAL;
 			break;
 		}
 
 		if (unlikely(check_destination_event_queue(evtims[i],
 							   adapter) < 0)) {
-			evtims[i]->state = RTE_EVENT_TIMER_ERROR;
+			__atomic_store_n(&evtims[i]->state,
+					RTE_EVENT_TIMER_ERROR,
+					__ATOMIC_RELAXED);
 			rte_errno = EINVAL;
 			break;
 		}
@@ -1099,13 +1107,18 @@  __swtim_arm_burst(const struct rte_event_timer_adapter *adapter,
 					  SINGLE, lcore_id, NULL, evtims[i]);
 		if (ret < 0) {
 			/* tim was in RUNNING or CONFIG state */
-			evtims[i]->state = RTE_EVENT_TIMER_ERROR;
+			__atomic_store_n(&evtims[i]->state,
+					RTE_EVENT_TIMER_ERROR,
+					__ATOMIC_RELEASE);
 			break;
 		}
 
-		rte_smp_wmb();
 		EVTIM_LOG_DBG("armed an event timer");
-		evtims[i]->state = RTE_EVENT_TIMER_ARMED;
+		/* RELEASE ordering guarantees the adapter specific value
+		 * changes observed before the update of state.
+		 */
+		__atomic_store_n(&evtims[i]->state, RTE_EVENT_TIMER_ARMED,
+				__ATOMIC_RELEASE);
 	}
 
 	if (i < nb_evtims)
@@ -1132,6 +1145,7 @@  swtim_cancel_burst(const struct rte_event_timer_adapter *adapter,
 	struct rte_timer *timp;
 	uint64_t opaque;
 	struct swtim *sw = swtim_pmd_priv(adapter);
+	enum rte_event_timer_state n_state;
 
 #ifdef RTE_LIBRTE_EVENTDEV_DEBUG
 	/* Check that the service is running. */
@@ -1143,16 +1157,18 @@  swtim_cancel_burst(const struct rte_event_timer_adapter *adapter,
 
 	for (i = 0; i < nb_evtims; i++) {
 		/* Don't modify the event timer state in these cases */
-		if (evtims[i]->state == RTE_EVENT_TIMER_CANCELED) {
+		/* ACQUIRE ordering guarantees the access of implementation
+		 * specific opague data under the correct state.
+		 */
+		n_state = __atomic_load_n(&evtims[i]->state, __ATOMIC_ACQUIRE);
+		if (n_state == RTE_EVENT_TIMER_CANCELED) {
 			rte_errno = EALREADY;
 			break;
-		} else if (evtims[i]->state != RTE_EVENT_TIMER_ARMED) {
+		} else if (n_state != RTE_EVENT_TIMER_ARMED) {
 			rte_errno = EINVAL;
 			break;
 		}
 
-		rte_smp_rmb();
-
 		opaque = evtims[i]->impl_opaque[0];
 		timp = (struct rte_timer *)(uintptr_t)opaque;
 		RTE_ASSERT(timp != NULL);
@@ -1166,9 +1182,12 @@  swtim_cancel_burst(const struct rte_event_timer_adapter *adapter,
 
 		rte_mempool_put(sw->tim_pool, (void **)timp);
 
-		evtims[i]->state = RTE_EVENT_TIMER_CANCELED;
-
-		rte_smp_wmb();
+		/* The RELEASE ordering here pairs with atomic ordering
+		 * to make sure the state update data observed between
+		 * threads.
+		 */
+		__atomic_store_n(&evtims[i]->state, RTE_EVENT_TIMER_CANCELED,
+				__ATOMIC_RELEASE);
 	}
 
 	return i;
diff --git a/lib/librte_eventdev/rte_event_timer_adapter.h b/lib/librte_eventdev/rte_event_timer_adapter.h
index d2ebcb0..6f64b90 100644
--- a/lib/librte_eventdev/rte_event_timer_adapter.h
+++ b/lib/librte_eventdev/rte_event_timer_adapter.h
@@ -467,7 +467,7 @@  struct rte_event_timer {
 	 *  - op: RTE_EVENT_OP_NEW
 	 *  - event_type: RTE_EVENT_TYPE_TIMER
 	 */
-	volatile enum rte_event_timer_state state;
+	enum rte_event_timer_state state;
 	/**< State of the event timer. */
 	uint64_t timeout_ticks;
 	/**< Expiry timer ticks expressed in number of *timer_ticks_ns* from