Commit fcf308e5 authored by Gerhard Engleder's avatar Gerhard Engleder Committed by Paolo Abeni
Browse files

ptp: Speed up vclock lookup 



ptp_convert_timestamp() is called in the RX path of network messages.
The current implementation takes ~5000ns on 1.2GHz A53. This is too much
for the hot path of packet processing.

Introduce hash table for fast vclock lookup in ptp_convert_timestamp().
The execution time of ptp_convert_timestamp() is reduced to ~700ns on
1.2GHz A53.

Signed-off-by: default avatarGerhard Engleder <gerhard@engleder-embedded.com>
Acked-by: default avatarRichard Cochran <richardcochran@gmail.com>
Signed-off-by: default avatarPaolo Abeni <pabeni@redhat.com>
parent 97dc7cd9

drivers/ptp/ptp_private.h

+1 −0
Original line number Diff line number Diff line
@@ -63,6 +63,7 @@ struct ptp_vclock { 
	struct ptp_clock *pclock;

	struct ptp_clock_info info;

	struct ptp_clock *clock;

	struct hlist_node vclock_hash_node;

	struct cyclecounter cc;

	struct timecounter tc;

	spinlock_t lock;	/* protects tc/cc */

drivers/ptp/ptp_vclock.c

+47 −19
Original line number Diff line number Diff line
@@ -5,6 +5,7 @@ 
 * Copyright 2021 NXP

 */

#include <linux/slab.h>

#include <linux/hashtable.h>

#include "ptp_private.h"



#define PTP_VCLOCK_CC_SHIFT		31
@@ -13,6 +14,32 @@ 
#define PTP_VCLOCK_FADJ_DENOMINATOR	15625ULL

#define PTP_VCLOCK_REFRESH_INTERVAL	(HZ * 2)



/* protects vclock_hash addition/deletion */

static DEFINE_SPINLOCK(vclock_hash_lock);



static DEFINE_READ_MOSTLY_HASHTABLE(vclock_hash, 8);



static void ptp_vclock_hash_add(struct ptp_vclock *vclock)

{

	spin_lock(&vclock_hash_lock);



	hlist_add_head_rcu(&vclock->vclock_hash_node,

			   &vclock_hash[vclock->clock->index % HASH_SIZE(vclock_hash)]);



	spin_unlock(&vclock_hash_lock);

}



static void ptp_vclock_hash_del(struct ptp_vclock *vclock)

{

	spin_lock(&vclock_hash_lock);



	hlist_del_init_rcu(&vclock->vclock_hash_node);



	spin_unlock(&vclock_hash_lock);



	synchronize_rcu();

}



static int ptp_vclock_adjfine(struct ptp_clock_info *ptp, long scaled_ppm)

{

	struct ptp_vclock *vclock = info_to_vclock(ptp);
@@ -176,6 +203,8 @@ struct ptp_vclock *ptp_vclock_register(struct ptp_clock *pclock) 
	snprintf(vclock->info.name, PTP_CLOCK_NAME_LEN, "ptp%d_virt",

		 pclock->index);



	INIT_HLIST_NODE(&vclock->vclock_hash_node);



	spin_lock_init(&vclock->lock);



	vclock->clock = ptp_clock_register(&vclock->info, &pclock->dev);
@@ -187,11 +216,15 @@ struct ptp_vclock *ptp_vclock_register(struct ptp_clock *pclock) 
	timecounter_init(&vclock->tc, &vclock->cc, 0);

	ptp_schedule_worker(vclock->clock, PTP_VCLOCK_REFRESH_INTERVAL);



	ptp_vclock_hash_add(vclock);



	return vclock;

}



void ptp_vclock_unregister(struct ptp_vclock *vclock)

{

	ptp_vclock_hash_del(vclock);



	ptp_clock_unregister(vclock->clock);

	kfree(vclock);

}
@@ -234,34 +267,29 @@ EXPORT_SYMBOL(ptp_get_vclocks_index); 


ktime_t ptp_convert_timestamp(const ktime_t *hwtstamp, int vclock_index)

{

	char name[PTP_CLOCK_NAME_LEN] = "";

	unsigned int hash = vclock_index % HASH_SIZE(vclock_hash);

	struct ptp_vclock *vclock;

	struct ptp_clock *ptp;

	unsigned long flags;

	struct device *dev;

	u64 ns;

	u64 vclock_ns = 0;



	snprintf(name, PTP_CLOCK_NAME_LEN, "ptp%d", vclock_index);

	dev = class_find_device_by_name(ptp_class, name);

	if (!dev)

		return 0;



	ptp = dev_get_drvdata(dev);

	if (!ptp->is_virtual_clock) {

		put_device(dev);

		return 0;

	}

	ns = ktime_to_ns(*hwtstamp);



	vclock = info_to_vclock(ptp->info);

	rcu_read_lock();



	ns = ktime_to_ns(*hwtstamp);

	hlist_for_each_entry_rcu(vclock, &vclock_hash[hash], vclock_hash_node) {

		if (vclock->clock->index != vclock_index)

			continue;



		spin_lock_irqsave(&vclock->lock, flags);

	ns = timecounter_cyc2time(&vclock->tc, ns);

		vclock_ns = timecounter_cyc2time(&vclock->tc, ns);

		spin_unlock_irqrestore(&vclock->lock, flags);

		break;

	}



	put_device(dev);

	return ns_to_ktime(ns);

	rcu_read_unlock();



	return ns_to_ktime(vclock_ns);

}

EXPORT_SYMBOL(ptp_convert_timestamp);

#endif