[LU-5397] optimize busy loop waiting in class_cleanup()

• Status: Resolved
• Project: Lustre
• Fix Version/s: Lustre 2.7.0
• Type: Bug
• Priority: Minor
• Reporter: Alexander Zarochentsev
• Assignee: WC Triage
• Resolution: Fixed
• Labels: patch
• Severity: 3

Description

the following code from class_cleanup():

	/* wait for already-arrived-connections to finish. */
	while (obd->obd_conn_inprogress > 0) {
		spin_unlock(&obd->obd_dev_lock);

		cond_resched();

		spin_lock(&obd->obd_dev_lock);
	}
	spin_unlock(&obd->obd_dev_lock);

contains at least two defects (if not going to substitute busy loop waiting by another more suitable synchronisation mechanism) :

1. spinlocks are not necessary for atomic access to obd_conn_inprogress field, reading is already atomic.
2. yield() should be used instead of cond_resched()

Comments

[Alexander Zarochentsev]

proposed fix http://review.whamcloud.com/#/c/11198/

[Alexander Zarochentsev]

patch updated to have memory barrier instead of spinlocks.

[Peter Jones]

Landed for 2.7