Re: [PATCH net-next v2] net: phy: dp83867: Add speed optimization feature

From: Dan Murphy
Date: Tue Feb 18 2020 - 12:17:33 EST


On 2/18/20 10:49 AM, Russell King - ARM Linux admin wrote:
On Tue, Feb 18, 2020 at 10:36:47AM -0600, Dan Murphy wrote:

On 2/18/20 10:25 AM, Russell King - ARM Linux admin wrote:
On Fri, Feb 14, 2020 at 12:31:52PM -0600, Dan Murphy wrote:

On 2/14/20 12:32 PM, Grygorii Strashko wrote:
I think it's good idea to have this message as just wrong cable might be

But this notifier make no sense in it current form - it will produce
noise in case of forced 100m/10M.

FYI. PHY sequence to update link:
ÂÂÂ |- .phy_link_change()->phy_link_change()
ÂÂÂÂ|-adjust_link() ----> netdev callback

So, log output has to be done or in .read_status() or
some info has to be saved in .read_status() and then re-used in

OK I will try to find a way to give some sort of message.
How do you know the speed that the PHY downshifted to?
The DP83867 has a register PHYSTS where BIT 15:14 indicate the speed that
the PHY negotiated.

In the same register BIT 13 indicates the duplex mode.

If the speed and duplex are available in some PHY specific status
register, then one way you can detect downshift is to decode the
negotiated speed/duplex from the advertisements (specifically the LPA
read from the registers and the advertisement that we should be
advertising - some PHYs modify their registers when downshifting) and
check whether it matches the negotiated parameters in the PHY
specific status register.

Alternatively, if the PHY modifies the advertisement register on
downshift, comparing the advertisement register with what it should
be will tell you if downshift has occurred.
The ISR register BIT 5 indicates if a downshift occurred or not. So we can
indicate that the PHY downshifted but there is no cause in the registers bit
field. My concern for this bit though is the register is clear on read so
all other interrupts are lost if we only read to check downshift. And the
link_change_notifier is called before the interrupt ACK call back. We could
call the interrupt function and get the downshift status but again it will
clear the interrupt register and any other statuses may be lost.
What's wrong with having an ack_interrupt() method that reads the
PHY ISR register, and records in a driver private flag that bit 5
has been set? The read_status() method can clear the flag if link
goes down, or check the flag if link is up and report that a
downshift event occurred.

If IRQs are not in use, then read_status() would have to read the
ISR itself.

It may be better to move ack_interrupt() to did_interrupt(), which
will ensure that it gets executed before the PHY state machine is
triggered by phy_interrupt().

Well now the read_status is becoming a lot more complex. It would be better to remove the ack_interrupt call back and just have read_status call the interrupt function regardless of interrupts or not. Because all the interrupt function would be doing is clearing all the interrupts in the ISR register on a link up/down event. And as you pointed out we can check and set a flag that indicates if a downshift has happened on link up status and clear it on link down. We would need to set the downshift interrupt mask to always report that bit. As opposed to not setting any interrupts if interrupts are not enabled. If the user wants to track WoL interrupt or any other feature interrupt we would have to add that flag to the read_status as well seems like it could get a bit out of control.

Again this is a lot of error prone complex changes and tracking just to populate a message in the kernel log. There is no guarantee that the LP did not force the downshift or advertise that it supports 1Gbps. So what condition is really being reported here? There seems like there are so many different scenarios why the PHY could not negotiate to its advertised 1Gbps.