LU-17251 osp: start OST object precreate earlier

If the OST object precreate count gets large (usually due to high
MDT file create workload, but sometimes also forced during testing)
then send an OST_CREATE RPC sooner when the number of precreated
objects gets low.

Currently the MDS will wait until 1/2 of the precreated OST objects
are consumed, but if create_count = 10000, then this can put bursty
create workloads on the OST.  Instead, send an OST_CREATE RPC when
the precreate pool is at most 1024 objects below target, so that the
MDS keeps its precreated pool more full and the OST doesn't have to
create so many objects at once (which also locks object directories
for a longer time).

Don't set opd_force_creation=true when osp.*.create_count is set
larger, and instead rely on the improved precreate check to force
OST object creation to start sooner, as opd_force_creation=true
can cause the OSP precreation to stop completely in some cases.

Test-Parameters: testlist=sanity env=ONLY=1-130,HONOR_EXCEPT=y
Test-Parameters: testlist=sanity env=ONLY=1-130,HONOR_EXCEPT=y
Test-Parameters: testlist=sanity env=ONLY=1-130,HONOR_EXCEPT=y
Test-Parameters: testlist=sanity env=ONLY=1-130,HONOR_EXCEPT=y
Test-Parameters: testlist=parallel-scale env=ONLY=rr_alloc,ONLY_REPEAT=10
Test-Parameters: testlist=parallel-scale env=ONLY=rr_alloc,ONLY_REPEAT=10
Test-Parameters: testlist=parallel-scale env=ONLY=rr_alloc,ONLY_REPEAT=10
Test-Parameters: testlist=parallel-scale env=ONLY=rr_alloc,ONLY_REPEAT=10
Fixes: df5b4c0a ("LU-17251 osp: force precreate if create_count grows")
Signed-off-by: Andreas Dilger <adilger@whamcloud.com>
Change-Id: Id2d12636d535485919ca5eec3adb18b1e6ce7057
Reviewed-on: https://review.whamcloud.com/c/fs/lustre-release/+/53245


Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Alex Deiter <alex.deiter@gmail.com>
Reviewed-by: Alex Zhuravlev <bzzz@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>

lustre/osp/lproc_osp.c

@@ -443,10 +443,8 @@ static ssize_t create_count_store(struct kobject *kobj, struct attribute *attr,
 		osp->opd_pre_create_count = rounddown_pow_of_two(val);
 
 	if (osp->opd_pre_create_count + osp->opd_pre_reserved >
-	    osp_objs_precreated(osp)) {
-		osp->opd_force_creation = true;
+	    osp_objs_precreated(osp))
 		wake_up(&osp->opd_pre_waitq);
-	}
 
 	return count;
 }

lustre/osp/osp_precreate.c

@@ -303,39 +303,43 @@ void osp_statfs_need_now(struct osp_device *d)
 }
 
 /**
- * Check pool of precreated objects is nearly empty
+ * Check pool of precreated objects is getting low.
  *
- * We should not wait till the pool of the precreated objects is exhausted,
+ * We should not wait till the pool of the precreated objects is too low,
  * because then there will be a long period of OSP being unavailable for the
- * new creations due to lenghty precreate RPC. Instead we ask for another
+ * new creations due to lengthy precreate RPC. Instead we ask for another
  * precreation ahead and hopefully have it ready before the current pool is
- * empty. Notice this function relies on an external locking.
+ * empty. Notice this function relies on external locking by opd_pre_lock.
  *
  * \param[in] d		OSP device
  *
  * \retval		0 - current pool is good enough, 1 - time to precreate
  */
-static inline int osp_precreate_near_empty_nolock(struct osp_device *d)
+static inline int osp_precreate_is_low_nolock(struct osp_device *d)
 {
-	int window = osp_objs_precreated(d);
+	int available = osp_objs_precreated(d) - d->opd_pre_reserved;
+	int precreate_needed = d->opd_pre_create_count > 1024 ?
+		d->opd_pre_create_count / 4 : d->opd_pre_create_count / 2;
 
-	/* don't consider new precreation till OST is healty and
-	 * has free space */
-	return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2 ||
+	if (precreate_needed > 1024)
+		precreate_needed = 1024;
+
+	/* no new precreation until OST is healthy and has free space */
+	return ((d->opd_pre_create_count - available > precreate_needed ||
 		 d->opd_force_creation) && (d->opd_pre_status == 0));
 }
 
 /**
  * Check pool of precreated objects
  *
- * This is protected version of osp_precreate_near_empty_nolock(), check that
+ * This is protected version of osp_precreate_is_low_nolock(), check that
  * for the details.
  *
  * \param[in] d		OSP device
  *
  * \retval		0 - current pool is good enough, 1 - time to precreate
  */
-static inline int osp_precreate_near_empty(struct osp_device *d)
+static inline int osp_precreate_is_low(struct osp_device *d)
 {
 	int rc;
 
@@ -344,7 +348,7 @@ static inline int osp_precreate_near_empty(struct osp_device *d)
 
 	/* XXX: do we really need locking here? */
 	spin_lock(&d->opd_pre_lock);
-	rc = osp_precreate_near_empty_nolock(d);
+	rc = osp_precreate_is_low_nolock(d);
 	spin_unlock(&d->opd_pre_lock);
 	return rc;
 }
@@ -1317,7 +1321,7 @@ static int osp_precreate_thread(void *_args)
 		while (!kthread_should_stop()) {
 			wait_event_idle(d->opd_pre_waitq,
 					kthread_should_stop() ||
-					(osp_precreate_near_empty(d) &&
+					(osp_precreate_is_low(d) &&
 					 !(osp_precreate_end_seq(d) &&
 					   osp_objs_precreated(d) != 0)) ||
 					osp_statfs_need_update(d) ||
@@ -1351,7 +1355,7 @@ static int osp_precreate_thread(void *_args)
 				}
 			}
 
-			if (osp_precreate_near_empty(d)) {
+			if (osp_precreate_is_low(d)) {
 				rc = osp_precreate_send(env, d);
 				/* osp_precreate_send() sets opd_pre_status
 				 * in case of error, that prevent the using of
@@ -1493,7 +1497,7 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
 				 * XXX: don't wake up if precreation
 				 * is in progress
 				 */
-				if (osp_precreate_near_empty_nolock(d) &&
+				if (osp_precreate_is_low_nolock(d) &&
 				   !osp_precreate_end_seq_nolock(d))
 					wake_up(&d->opd_pre_waitq);