netdev - Re: [PATCH net-next 1/2] mlx4_en: Add PTP hardware clock

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <52B9AA7B.8090401@mellanox.com>
Date:	Tue, 24 Dec 2013 17:38:35 +0200
From:	Hadar Hen Zion <hadarh@...lanox.com>
To:	Shawn Bohrer <shawn.bohrer@...il.com>
CC:	"David S. Miller" <davem@...emloft.net>,
	Or Gerlitz <ogerlitz@...lanox.com>,
	Amir Vadai <amirv@...lanox.com>,
	Richard Cochran <richardcochran@...il.com>,
	<netdev@...r.kernel.org>, <tomk@...advisors.com>,
	Shawn Bohrer <sbohrer@...advisors.com>
Subject: Re: [PATCH net-next 1/2] mlx4_en: Add PTP hardware clock

On 12/23/2013 6:59 PM, Shawn Bohrer wrote:
> On Mon, Dec 23, 2013 at 10:29:58AM -0600, Shawn Bohrer wrote:
>> On Sun, Dec 22, 2013 at 03:13:12PM +0200, Hadar Hen Zion wrote:
>>> On 12/17/2013 10:32 PM, Shawn Bohrer wrote:
>>>> From: Shawn Bohrer <sbohrer@...advisors.com>
>>>>
>>>> This adds a PHC to the mlx4_en driver.  The code is largely based off of
>>>> the e1000e driver (drivers/net/ethernet/intel/e1000e/ptp.c) which seemed
>>>> very similar.
>>>>
>>>> This driver has been tested with both Documentation/ptp/testptp and the
>>>> linuxptp project (http://linuxptp.sourceforge.net/) and appears to work
>>>> on a Mellanox ConnectX-3 card.
>>>>
>>>> Signed-off-by: Shawn Bohrer <sbohrer@...advisors.com>
>>>> ---
>>>>   drivers/net/ethernet/mellanox/mlx4/en_clock.c   |  192 ++++++++++++++++++++++-
>>>>   drivers/net/ethernet/mellanox/mlx4/en_ethtool.c |    3 +
>>>>   drivers/net/ethernet/mellanox/mlx4/en_main.c    |    3 +
>>>>   drivers/net/ethernet/mellanox/mlx4/mlx4_en.h    |    6 +
>>>>   4 files changed, 196 insertions(+), 8 deletions(-)
>>>>
>>>> diff --git a/drivers/net/ethernet/mellanox/mlx4/en_clock.c b/drivers/net/ethernet/mellanox/mlx4/en_clock.c
>>>> index fd64410..9b0d515 100644
>>>> --- a/drivers/net/ethernet/mellanox/mlx4/en_clock.c
>>>> +++ b/drivers/net/ethernet/mellanox/mlx4/en_clock.c
>>>> @@ -103,17 +103,187 @@ void mlx4_en_fill_hwtstamps(struct mlx4_en_dev *mdev,
>>>>   			    struct skb_shared_hwtstamps *hwts,
>>>>   			    u64 timestamp)
>>>>   {
>>>> +	unsigned long flags;
>>>>   	u64 nsec;
>>>>
>>>> +	spin_lock_irqsave(&mdev->clock_lock, flags);
>>>
>>> 1. Missing initialization for clock_lock
>>> 2. Adding spin lock in the data path reduce performance by 15% when
>>> HW timestamping is enabled. I did some testing and replacing
>>> spin_lock_irqsave with read/write_lock_irqsave prevents the
>>> performance decrease.
>>
>> Thanks Hadar,
>>
>> I'm testing this change now, and will resend when I'm done.  However,
>> I noticed the following in Documentation/spinlocks.txt
>>
>>     NOTE! We are working hard to remove reader-writer spinlocks in most
>>     cases, so please don't add a new one without consensus.  (Instead, see
>>     Documentation/RCU/rcu.txt for complete information.)
>>
>> So is there consensus for a rwlock here?
>
> Also just to make sure I'm testing the correct thing.  These all need
> to be write_lock_irqsave() except for the one in
> mlx4_en_fill_hwtstamps() protecting timecounter_cyc2time() which can
> be a read_lock_irqsave().  All of the other timecounter* calls write
> to the timecounter including timecounter_read().  I'm assuming that is
> what you tested and that should still eliminate the performance loss
> since mlx4_en_fill_hwtstamps() should be the bottleneck.
>
> Thanks,
> Shawn
>
Yes, you were testing the correct thing.

But, after another check, I'm not sure we need any lock in 
mlx4_en_fill_hwtstamps() data path function.
Adding lock to mlx4_en_fill_hwtstamps() protects timecounter_cyc2time() 
which doesn't read hardware registers.
As you explained in your RFC mail:

 > In e1000e driver they protect the timecounter code with a spinlock
 > because the hardware reports the time in two 32bit registers.  The
 > Mellanox code looks similar.

The spin lock is needed when reading hardware registers.
My suggestion is to stay with spin locks in all the places protecting 
timecounter_read()/timecounter_init() and just remove the spin lock from 
timecounter_cyc2time()

Thanks,
Hadar

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html