Re: [RFC 04/11] net: phy: Handle c22 regs presence better

From: Jeremy Linton
Date: Mon May 25 2020 - 19:16:22 EST


On 5/25/20 5:01 PM, Russell King - ARM Linux admin wrote:
On Mon, May 25, 2020 at 04:51:16PM -0500, Jeremy Linton wrote:

On 5/25/20 5:06 AM, Russell King - ARM Linux admin wrote:
On Sun, May 24, 2020 at 10:34:13PM -0500, Jeremy Linton wrote:

On 5/23/20 1:37 PM, Russell King - ARM Linux admin wrote:
On Fri, May 22, 2020 at 04:30:52PM -0500, Jeremy Linton wrote:
Until this point, we have been sanitizing the c22
regs presence bit out of all the MMD device lists.
This is incorrect as it causes the 0xFFFFFFFF checks
to incorrectly fail. Further, it turns out that we
want to utilize this flag to make a determination that
there is actually a phy at this location and we should
be accessing it using c22.

Signed-off-by: Jeremy Linton <jeremy.linton@xxxxxxx>
drivers/net/phy/phy_device.c | 16 +++++++++++++---
1 file changed, 13 insertions(+), 3 deletions(-)

diff --git a/drivers/net/phy/phy_device.c b/drivers/net/phy/phy_device.c
index f0761fa5e40b..2d677490ecab 100644
--- a/drivers/net/phy/phy_device.c
+++ b/drivers/net/phy/phy_device.c
@@ -689,9 +689,6 @@ static int get_phy_c45_devs_in_pkg(struct mii_bus *bus, int addr, int dev_addr,
return -EIO;
*devices_in_package |= phy_reg;
- /* Bit 0 doesn't represent a device, it indicates c22 regs presence */
- *devices_in_package &= ~BIT(0);
return 0;
@@ -742,6 +739,8 @@ static int get_phy_c45_ids(struct mii_bus *bus, int addr, u32 *phy_id,
int i;
const int num_ids = ARRAY_SIZE(c45_ids->device_ids);
u32 *devs = &c45_ids->devices_in_package;
+ bool c22_present = false;
+ bool valid_id = false;
/* Find first non-zero Devices In package. Device zero is reserved
* for 802.3 c45 complied PHYs, so don't probe it at first.
@@ -770,6 +769,10 @@ static int get_phy_c45_ids(struct mii_bus *bus, int addr, u32 *phy_id,
return 0;
+ /* Bit 0 doesn't represent a device, it indicates c22 regs presence */
+ c22_present = *devs & BIT(0);
+ *devs &= ~BIT(0);
/* Now probe Device Identifiers for each device present. */
for (i = 1; i < num_ids; i++) {
if (!(c45_ids->devices_in_package & (1 << i)))
@@ -778,6 +781,13 @@ static int get_phy_c45_ids(struct mii_bus *bus, int addr, u32 *phy_id,
ret = _get_phy_id(bus, addr, i, &c45_ids->device_ids[i], true);
if (ret < 0)
return ret;
+ if (valid_phy_id(c45_ids->device_ids[i]))
+ valid_id = true;

Here you are using your "devices in package" validator to validate the
PHY ID value. One of the things it does is mask this value with
0x1fffffff. That means you lose some of the vendor OUI. To me, this
looks completely wrong.

I think in this case I was just using it like the comment in
get_phy_device() "if the phy_id is mostly F's, there is no device here".

My understanding is that the code is trying to avoid the 0xFFFFFFFF returns
that seem to indicate "bus ok, phy didn't respond".

I just checked the OUI registration, and while there are a couple OUI's
registered that have a number of FFF's in them, none of those cases seems to
overlap sufficiently to cause this to throw them out. Plus a phy would also
have to have model+revision set to 'F's. So while might be possible, if
unlikely, at the moment I think the OUI registration keeps this from being a
problem. Particularly, if i'm reading the mapping correctly, the OUI mapping
guarantees that the field cannot be all '1's due to the OUI having X & M
bits cleared. It sort of looks like the mapping is trying to lose those
bits, by tossing bit 1 & 2, but the X & M are in the wrong octet (AFAIK, I
just read it three times cause it didn't make any sense).

I should also note that we have at least one supported PHY where one
of the MMDs returns 0xfffe for even numbered registers and 0x0000 for
odd numbered registers in one of the vendor MMDs for addresses 0
through 0xefff - which has a bit set in the devices-in-package.

It also returns 0x0082 for almost every register in MMD 2, but MMD 2's
devices-in-package bit is clear in most of the valid MMDs, so we
shouldn't touch it.

These reveal the problem of randomly probing MMDs - they can return
unexpected values and not be as well behaved as we would like them to
be. Using register 8 to detect presence may be beneficial, but that
may also introduce problems as we haven't used that before (and we
don't know whether any PHY that wrong.) I know at least the 88x3310
gets it right for all except the vendor MMDs, where the low addresses
appear non-confromant to the 802.3 specs. Both vendor MMDs are
definitely implemented, just not with anything conforming to 802.3.

Yes, we know even for the NXP reference hardware, one of the phy's doesn't
probe out correctly because it doesn't respond to the ieee defined
registers. I think at this point, there really isn't anything we can do
about that unless we involve the (ACPI) firmware in currently nonstandard

So, my goals here have been to first, not break anything, and then do a
slightly better job finding phy's that are (mostly?) responding correctly to
the 802.3 spec. So we can say "if you hardware is ACPI conformant, and you
have IEEE conformant phy's you should be ok". So, for your example phy, I
guess the immediate answer is "use DT" or "find a conformant phy", or even
"abstract it in the firmware and use a mailbox interface".

You haven't understood. The PHY does conform for most of the MMDs,
but there are a number that do not conform.

Maybe I should clarify. This set is still terminating the search for a valid MMD device list on the first MMD that responds. It then probes the same ID registers of the flagged MMDs as before. What has changed is that we search higher into the MMD address space for a device list. So previously if a device didn't respond to MMD0-8 it was ignored. Now it needs to fail all of 0-31 to be ignored. Similarly for the ID's, if we find what appears to be a valid MMD device list, then we will probe not only the original 1-8 MMDs for IDs, but 1-31 MMDs for IDs.

So any device which presented a non zero, non "mostly ff's" devices list in 0-8 will get the same device list as before. Similarly we probe for ids at the same MMD addresses as before, with additional MMD detection >8. This change means we pick up additional phys, and we detect the correct MMD ids for more devices.

The possible negative differences are in 7/11 where we transition a device which responded with an ID=0 to 0xFFFFFFFF which will cause the code to not request a phy/mmd driver for 00000000 (or potentially some bogus OUI's due to the 1fffffff difference). Assuming a valid phy id gets applied somewhere, the same phy driver will load, and presumably it will know what to do with MMD's in the devices list.

So, I don't see anything in your example above which changes the detected MMD devices. Potentially though, not only will you get the previous MMD Id's, but you might get a few new ones. Whether MMD2 is probed is going to depend on whether MMD0/1's devices list has the MMD2 device bit set (you weren't clear in the description above).