invalid requirement from ethtool?

invalid requirement from ethtool?

[Eli Cohen]

Hi,
I see the following text in include/linux/ethtool.h and wonder what is
the reasoning for requiring that both params cannot be zero. I could
not track when and who inserted this text as it dates before git was
used to track kernel code, but my feeling is that is related to a
specific hardware limitation.

        /* How many packets to delay an RX interrupt after
         * a packet arrives.  If 0, only rx_coalesce_usecs is
         * used.  It is illegal to set both usecs and max frames
         * to zero as this would cause RX interrupts to never be
         * generated.
         */
        __u32   rx_max_coalesced_frames;

        /* How many packets to delay a TX interrupt after
         * a packet is sent.  If 0, only tx_coalesce_usecs is
         * used.  It is illegal to set both usecs and max frames
         * to zero as this would cause TX interrupts to never be
         * generated.
         */
        __u32   tx_max_coalesced_frames;

I found this in tg3 driver:
         /* No rx interrupts will be generated if both are zero */
        if ((ec->rx_coalesce_usecs == 0) &&
            (ec->rx_max_coalesced_frames == 0))
                return -EINVAL;

However, bnx2 for example allows setting both to zero.

I think both params zero should be allowed and mean coalescing is not
operational, thus we can remove these comments from ethtool.h

Re: invalid requirement from ethtool?

[David Miller]

From: Eli Cohen <eli@dev.mellanox.co.il>
Date: Tue, 26 Jul 2011 15:42:22 +0300

> I think both params zero should be allowed and mean coalescing is not
> operational, thus we can remove these comments from ethtool.h

The existing precendence has existed for more than 10 years.  You can't
just change it like this.

You'll need to find a new way to encode "disabled" coalescing.

Re: invalid requirement from ethtool?

[Eli Cohen]

On Wed, Jul 27, 2011 at 10:43:09PM -0700, David Miller wrote:
> From: Eli Cohen <eli@dev.mellanox.co.il>
> Date: Tue, 26 Jul 2011 15:42:22 +0300
> 
> > I think both params zero should be allowed and mean coalescing is not
> > operational, thus we can remove these comments from ethtool.h
> 
> The existing precendence has existed for more than 10 years.  You can't
> just change it like this.
> 
> You'll need to find a new way to encode "disabled" coalescing.

I can't see the text explicitly specifies how to "disable" coalescing.
If I ignore the specific comment that disallows both params 0, I could
interpret the text such that when they're both zero, the feature is
disabled.

BTW, I checked also Intel 10GBE Nic (ixgbe) driver and it allows
setting both 0.

Re: invalid requirement from ethtool?

[David Miller]

From: Eli Cohen <eli@dev.mellanox.co.il>
Date: Thu, 28 Jul 2011 10:23:26 +0300

> On Wed, Jul 27, 2011 at 10:43:09PM -0700, David Miller wrote:
>> From: Eli Cohen <eli@dev.mellanox.co.il>
>> Date: Tue, 26 Jul 2011 15:42:22 +0300
>> 
>> > I think both params zero should be allowed and mean coalescing is not
>> > operational, thus we can remove these comments from ethtool.h
>> 
>> The existing precendence has existed for more than 10 years.  You can't
>> just change it like this.
>> 
>> You'll need to find a new way to encode "disabled" coalescing.
> 
> I can't see the text explicitly specifies how to "disable" coalescing.
> If I ignore the specific comment that disallows both params 0, I could
> interpret the text such that when they're both zero, the feature is
> disabled.

The documentation does not determine what the rules are, the cpu
does not execute the documentation it executes the code, and that's
what determines the rules.

Re: invalid requirement from ethtool?

[Ben Hutchings]

On Thu, 2011-07-28 at 00:37 -0700, David Miller wrote:
> From: Eli Cohen <eli@dev.mellanox.co.il>
> Date: Thu, 28 Jul 2011 10:23:26 +0300
> 
> > On Wed, Jul 27, 2011 at 10:43:09PM -0700, David Miller wrote:
> >> From: Eli Cohen <eli@dev.mellanox.co.il>
> >> Date: Tue, 26 Jul 2011 15:42:22 +0300
> >> 
> >> > I think both params zero should be allowed and mean coalescing is not
> >> > operational, thus we can remove these comments from ethtool.h
> >> 
> >> The existing precendence has existed for more than 10 years.  You can't
> >> just change it like this.
> >> 
> >> You'll need to find a new way to encode "disabled" coalescing.
> > 
> > I can't see the text explicitly specifies how to "disable" coalescing.
> > If I ignore the specific comment that disallows both params 0, I could
> > interpret the text such that when they're both zero, the feature is
> > disabled.
> 
> The documentation does not determine what the rules are, the cpu
> does not execute the documentation it executes the code, and that's
> what determines the rules.

The ethtool core doesn't check the values in struct ethtool_coalesce, so
the rules are really driver-specific.

Ben.

-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: invalid requirement from ethtool?

[Eli Cohen]

On Thu, Jul 28, 2011 at 12:22:33PM +0200, Ben Hutchings wrote:
> 
> The ethtool core doesn't check the values in struct ethtool_coalesce, so
> the rules are really driver-specific.
> 

Exactly. That's why I think those comments are not in place.

Re: invalid requirement from ethtool?

[Ben Hutchings]

On Tue, 2011-07-26 at 15:42 +0300, Eli Cohen wrote:
> Hi,
> I see the following text in include/linux/ethtool.h and wonder what is
> the reasoning for requiring that both params cannot be zero. I could
> not track when and who inserted this text as it dates before git was
> used to track kernel code, but my feeling is that is related to a
> specific hardware limitation.
> 
>         /* How many packets to delay an RX interrupt after
>          * a packet arrives.  If 0, only rx_coalesce_usecs is
>          * used.  It is illegal to set both usecs and max frames
>          * to zero as this would cause RX interrupts to never be
>          * generated.
>          */
>         __u32   rx_max_coalesced_frames;
> 
>         /* How many packets to delay a TX interrupt after
>          * a packet is sent.  If 0, only tx_coalesce_usecs is
>          * used.  It is illegal to set both usecs and max frames
>          * to zero as this would cause TX interrupts to never be
>          * generated.
>          */
>         __u32   tx_max_coalesced_frames;
> 
> I found this in tg3 driver:
>          /* No rx interrupts will be generated if both are zero */
>         if ((ec->rx_coalesce_usecs == 0) &&
>             (ec->rx_max_coalesced_frames == 0))
>                 return -EINVAL;
> 
> However, bnx2 for example allows setting both to zero.
> 
> I think both params zero should be allowed and mean coalescing is not
> operational, thus we can remove these comments from ethtool.h

If coalescing is not operational, the maximum number of completions
before an interrupt is 1.  So logically {rx,tx}_max_coalesced_frames
should be 1, right?  Although the comment does say 'How many packets ...
after ...' which implies that the value of the field must be 1 less than
the wanted maximum, i.e. 0, which is supposedly invalid.

The first implementation of ethtool coalescing control was in tg3, so it
should be a useful reference.

David, I know you maintained tg3 for some time so I assume you have a
hardware reference.  Can you confirm whether a value of 1 in
HOSTCC_{RX,TX}MAX_FRAMES results in an interrupt after 1 completion or
after 2 completions?

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: invalid requirement from ethtool?

[David Miller]

From: Ben Hutchings <bhutchings@solarflare.com>
Date: Mon, 15 Aug 2011 18:24:00 +0100

> David, I know you maintained tg3 for some time so I assume you have a
> hardware reference.  Can you confirm whether a value of 1 in
> HOSTCC_{RX,TX}MAX_FRAMES results in an interrupt after 1 completion or
> after 2 completions?

A value of 1 results in an interrupt after 1 completion.

If you put a zero there, the packet count comparison never triggers.