Message-ID: <575FFAC2.1070002@linaro.org>
Date:	Tue, 14 Jun 2016 15:38:26 +0300
From:	Ivan Khoronzhuk <ivan.khoronzhuk@...aro.org>
To:	"Andrew F. Davis" <afd@...com>,
	Mugunthan V N <mugunthanvnm@...com>,
	Schuyler Patton <spatton@...com>, linux-kernel@...r.kernel.org
Cc:	grygorii.strashko@...com, linux-omap@...r.kernel.org,
	netdev@...r.kernel.org, robh+dt@...nel.org, pawel.moll@....com,
	mark.rutland@....com, ijc+devicetree@...lion.org.uk,
	galak@...eaurora.org, bcousson@...libre.com, tony@...mide.com,
	devicetree@...r.kernel.org
Subject: Re: [PATCH v2 0/2] net: ethernet: ti: cpsw: delete rx_descs property



On 13.06.16 18:19, Andrew F. Davis wrote:
> On 06/13/2016 03:22 AM, Mugunthan V N wrote:
>> On Saturday 11 June 2016 04:34 AM, Schuyler Patton wrote:
>>>
>>>
>>> On 06/08/2016 07:03 PM, Ivan Khoronzhuk wrote:
>>>>
>>>>
>>>> On 09.06.16 02:11, Schuyler Patton wrote:
>>>>>
>>>>>
>>>>> On 06/08/2016 09:06 AM, Ivan Khoronzhuk wrote:
>>>>>>
>>>>>>
>>>>>> On 08.06.16 17:01, Ivan Khoronzhuk wrote:
>>>>>>> Hi Schuyer,
>>>>>>>
>>>>>>> On 07.06.16 18:26, Schuyler Patton wrote:
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> On 06/07/2016 08:59 AM, Ivan Khoronzhuk wrote:
>>>>>>>>> There is no reason in rx_descs property because davinici_cpdma
>>>>>>>>> driver splits pool of descriptors equally between tx and rx
>>>>>>>>> channels.
>>>>>>>>> So, this patch series makes driver to use available number of
>>>>>>>>> descriptors for rx channels.
>>>>>>>>
>>>>>>>> I agree with the idea of consolidating how the descriptors are
>>>>>>>> defined because of
>>>>>>>> the two variable components, number and size of the pool can be
>>>>>>>> confusing to
>>>>>>>> end users. I would like to request to change how it is being
>>>>>>>> proposed here.
>>>>>>>>
>>>>>>>> I think the number of descriptors should be left in the device
>>>>>>>> tree source file as
>>>>>>>> is and remove the BD size variable and have the driver calculate
>>>>>>>> the size of the
>>>>>>>> pool necessary to support the descriptor request. From an user
>>>>>>>> perspective it is
>>>>>>>> easier I think to be able to list the number of descriptors
>>>>>>>> necessary vs. the size
>>>>>>>> of the pool.
>>>>>>>>
>>>>>>>> Since the patch series points out how it is used so in the driver
>>>>>>>> so to make that
>>>>>>>> consistent is perhaps change rx_descs to total_descs.
>>>>>>>>
>>>>>>>> Regards,
>>>>>>>> Schuyler
>>>>>>>
>>>>>>> The DT entry for cpsw doesn't have property for size of the pool.
>>>>>>> It contains only BD ram size, if you mean this. The size of the
>>>>>>> pool is
>>>>>>> software decision. Current version of DT entry contain only rx desc
>>>>>>> number.
>>>>>>> That is not correct, as it depends on the size of the descriptor,
>>>>>>> which is also
>>>>>>> h/w parameter. The DT entry has to describe only h/w part and
>>>>>>> shouldn't contain
>>>>>>> driver implementation details, and I'm looking on it from this
>>>>>>> perspective.
>>>>>>>
>>>>>>> Besides, rx_descs describes only rx number of descriptors, that are
>>>>>>> taken from
>>>>>>> the same pool as tx descriptors, and setting rx desc to some new
>>>>>>> value doesn't
>>>>>>> mean that rest of them are freed for tx. Also, I'm going to send
>>>>>>> series that
>>>>>>> adds multi channel support to the driver, and in this case,
>>>>>>> splitting of the
>>>>>>> pool will be more sophisticated than now, after what setting those
>>>>>>> parameters
>>>>>>> for user (he should do this via device tree) can be even more
>>>>>>> confusing. But,
>>>>>> should -> shouldn't
>>>>>>
>>>>>>> as it's supposed, it's software decision that shouldn't leak to the
>>>>>>> DT.
>>>>>
>>>>> If this rx-desc field is removed how will the number of descriptors
>>>>> be set?
>>>>>
>>>>> This field has been used to increase the number of descriptors so high
>>>>> volume short packets are not dropped due to descriptor exhaustion.
>>>>> The current
>>>>> default number of 64 rx descriptors is too low for gigabit networks.
>>>>> Some users
>>>>> have a strong requirement for zero loss of UDP packets setting this
>>>>> field to a
>>>>> larger number and setting the descriptors off-chip was a means to solve
>>>>> the problem.
>>>> The current implementation of cpdma driver splits descs num on 2 parts
>>>> equally.
>>>> Total number = 256, then 128 reserved for rx and 128 for tx, but
>>>> setting this to
>>>> 64, simply limits usage of reserved rx descriptors to 64, so that:
>>>> 64 rx descs, 128 tx descs and 64 are always present in the pool but
>>>> cannot be used,
>>>> (as new rx descriptor is allocated only after previous was freed).
>>>> That means, 64 rx descs are unused. In case of rx descriptor
>>>> exhaustion, an user can
>>>> set rx_descs to 128, for instance, in this case all descriptors will
>>>> be in use, but then question,
>>>> why intentionally limit number of rx descs, anyway rest 64 descs
>>>> cannot be used for other
>>>> purposes. In case of this patch, all rx descs are in use, and no need
>>>> to correct number
>>>> of rx descs anymore, use all of them....and it doesn't have impact on
>>>> performance, as
>>>> anyway, bunch of rx descs were simply limited by DT and unused. So,
>>>> probably, there is no
>>>> reason to worry about that.
>>>
>>> When we see this issue we set the descriptors to DDR and put a large number
>>> in the desc count. unfortunately I wish I could provide a number,
>>> usually the issue
>>> is a volume burst of short UDP packets.
>>>
>>>>
>>>> PS:
>>>> It doesn't concern this patch, but, which PPS makes rx descs to be
>>>> exhausted?...
>>>> (In this case "desc_alloc_fail" counter contains some value for rx
>>>> channel,
>>>> and can be read with "ethtool -S eth0". Also, the user will be WARNed
>>>> ON by the driver)
>>>>
>>>> it's interesting to test it, I'm worrying about, because in case of
>>>> multichannel,
>>>> the pool is split between all channels... they are throughput limited,
>>>> but
>>>> anyway, it's good to correlate the number of descs with throughput
>>>> assigned to
>>>> a channel, if possible. That has to be possible, if setting to 128
>>>> helps, then
>>>> has to be value between 64 and 128 to make handling of rx packets fast
>>>> enough.
>>>> After what, can be calculated correlation between number of rx descs
>>>> and throughput
>>>> split between channels....
>>>
>>> With gigabit networks 64 or 128 rx descriptors is not going to enough to
>>> fix the
>>> DMA overrun problem. Usually we set this number to an arbitrarily large
>>> 2000
>>> descriptors in external DDR to demonstrate it is possible to not drop
>>> packets. All
>>> this does is move the problem higher up so that the drops occur in network
>>> stack if the ARM is overloaded. With the high speed networks I would like
>>> to propose that the descriptor pool or pools are moved to DDR by
>>> default. It would
>>> be nice to have some reconfigurability or set a pool size that reduces
>>> or eliminates
>>> the DMA issue that is seen in these types of applications.
>>>
>>> This test gets used a lot, which is to send very short UDP packets. If I
>>> have the math
>>> right, a 52 byte (64 byte with the inter-frame gap) UDP packet the
>>> default 64
>>> descriptors gets consumed in roughly 33uS. There are the switch fifos
>>> which will also
>>> allow some headroom, but a user was dropping packets at the switch when
>>> they
>>> were bursting 360 packets at the processor on a gigabit link
>>>
>>
>> I too agree that rx-descs can be derived from the pool size and
>> descriptor size in driver itself. The current driver uses bd_ram_size to
>> set the pool size when the descriptors are placed in DDR which is wrong.
>>
>> Here I propose an idea to solve Schuyler's concern to keep the
>> descriptors in DDR when a system need more rx descriptors for lossless
>> UDB performance.
>>
>> The DT property rx-descs can be removed and add a new DT property
>> *pool_size* to add support for descriptors memory size in DDR and define
>> a pool size which the system needs for a network to have lossless UDP
>> transfers.
>>
>
> I second the pool_size property, but being purely a driver configuration
> setting based on expected network environment and not a HW description
> it should probably be a module parameter, not DT property.
>
> Andrew

Agree, DT is not the best place for it.
And it be done with separate patch-set, as this patch solves rx_descs issue,
not pool_size.

>
>> So based on no_bd_ram DT entry, the driver can decide whether it can use
>> internal BD-ram or DDR to initialize the cpdma driver.
>>
>> Regards
>> Mugunthan V N

-- 
Regards,
Ivan Khoronzhuk

Powered by blists - more mailing lists