Re: [PATCH] net: phy: fix PHY possibly unwork after MDIO bus resume back

From: Heiner Kallweit
Date: Tue Apr 06 2021 - 14:22:09 EST

On 06.04.2021 13:42, Heiner Kallweit wrote:
> On 06.04.2021 12:07, Joakim Zhang wrote:
>>
>>> -----Original Message-----
>>> From: Heiner Kallweit <hkallweit1@xxxxxxxxx>
>>> Sent: 2021年4月6日 14:29
>>> To: Joakim Zhang <qiangqing.zhang@xxxxxxx>; christian.melki@xxxxxxxxxx;
>>> andrew@xxxxxxx; linux@xxxxxxxxxxxxxxx; davem@xxxxxxxxxxxxx;
>>> kuba@xxxxxxxxxx
>>> Cc: netdev@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; dl-linux-imx
>>> <linux-imx@xxxxxxx>
>>> Subject: Re: [PATCH] net: phy: fix PHY possibly unwork after MDIO bus resume
>>> back
>>>
>>> On 06.04.2021 04:07, Joakim Zhang wrote:
>>>>
>>>> Hi Heiner,
>>>>
>>>>> -----Original Message-----
>>>>> From: Heiner Kallweit <hkallweit1@xxxxxxxxx>
>>>>> Sent: 2021年4月5日 20:10
>>>>> To: christian.melki@xxxxxxxxxx; Joakim Zhang
>>>>> <qiangqing.zhang@xxxxxxx>; andrew@xxxxxxx; linux@xxxxxxxxxxxxxxx;
>>>>> davem@xxxxxxxxxxxxx; kuba@xxxxxxxxxx
>>>>> Cc: netdev@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
>>>>> dl-linux-imx <linux-imx@xxxxxxx>
>>>>> Subject: Re: [PATCH] net: phy: fix PHY possibly unwork after MDIO bus
>>>>> resume back
>>>>>
>>>>> On 05.04.2021 10:43, Christian Melki wrote:
>>>>>> On 4/5/21 12:48 AM, Heiner Kallweit wrote:
>>>>>>> On 04.04.2021 16:09, Heiner Kallweit wrote:
>>>>>>>> On 04.04.2021 12:07, Joakim Zhang wrote:
>>>>>>>>> commit 4c0d2e96ba055 ("net: phy: consider that suspend2ram may
>>>>>>>>> cut off PHY power") invokes phy_init_hw() when MDIO bus resume,
>>>>>>>>> it will soft reset PHY if PHY driver implements soft_reset callback.
>>>>>>>>> commit 764d31cacfe4 ("net: phy: micrel: set soft_reset callback
>>>>>>>>> to genphy_soft_reset for KSZ8081") adds soft_reset for KSZ8081.
>>>>>>>>> After these two patches, I found i.MX6UL 14x14 EVK which
>>>>>>>>> connected to KSZ8081RNB doesn't work any more when system
>>> resume
>>>>>>>>> back, MAC
>>>>> driver is fec_main.c.
>>>>>>>>>
>>>>>>>>> It's obvious that initializing PHY hardware when MDIO bus resume
>>>>>>>>> back would introduce some regression when PHY implements
>>>>>>>>> soft_reset. When I
>>>>>>>>
>>>>>>>> Why is this obvious? Please elaborate on why a soft reset should
>>>>>>>> break something.
>>>>>>>>
>>>>>>>>> am debugging, I found PHY works fine if MAC doesn't support
>>>>>>>>> suspend/resume or phy_stop()/phy_start() doesn't been called
>>>>>>>>> during suspend/resume. This let me realize, PHY state machine
>>>>>>>>> phy_state_machine() could do something breaks the PHY.
>>>>>>>>>
>>>>>>>>> As we known, MAC resume first and then MDIO bus resume when
>>>>>>>>> system resume back from suspend. When MAC resume, usually it will
>>>>>>>>> invoke
>>>>>>>>> phy_start() where to change PHY state to PHY_UP, then trigger the
>>>>>>>>> stat> machine to run now. In phy_state_machine(), it will
>>>>>>>>> start/config auto-nego, then change PHY state to PHY_NOLINK, what
>>>>>>>>> to next is periodically check PHY link status. When MDIO bus
>>>>>>>>> resume, it will initialize PHY hardware, including soft_reset,
>>>>>>>>> what would soft_reset affect seems various from different PHYs.
>>>>>>>>> For KSZ8081RNB, when it in PHY_NOLINK state and then perform a
>>>>>>>>> soft reset,
>>>>> it will never complete auto-nego.
>>>>>>>>
>>>>>>>> Why? That would need to be checked in detail. Maybe chip errata
>>>>>>>> documentation provides a hint.
>>>>>>>>
>>>>>>>
>>>>>>> The KSZ8081 spec says the following about bit BMCR_PDOWN:
>>>>>>>
>>>>>>> If software reset (Register 0.15) is used to exit power-down mode
>>>>>>> (Register 0.11 = 1), two software reset writes (Register 0.15 = 1)
>>>>>>> are required. The first write clears power-down mode; the second
>>>>>>> write resets the chip and re-latches the pin strapping pin values.
>>>>>>>
>>>>>>> Maybe this causes the issue you see and genphy_soft_reset() isn't
>>>>>>> appropriate for this PHY. Please re-test with the KSZ8081 soft
>>>>>>> reset following the spec comment.
>>>>>>>
>>>>>>
>>>>>> Interesting. Never expected that behavior.
>>>>>> Thanks for catching it. Skimmed through the datasheets/erratas.
>>>>>> This is what I found (micrel.c):
>>>>>>
>>>>>> 10/100:
>>>>>> 8001 - Unaffected?
>>>>>> 8021/8031 - Double reset after PDOWN.
>>>>>> 8041 - Errata. PDOWN broken. Recommended do not use. Unclear if
>>>>>> reset solves the issue since errata says no error after reset but is
>>>>>> also claiming that only toggling PDOWN (may) or power will help.
>>>>>> 8051 - Double reset after PDOWN.
>>>>>> 8061 - Double reset after PDOWN.
>>>>>> 8081 - Double reset after PDOWN.
>>>>>> 8091 - Double reset after PDOWN.
>>>>>>
>>>>>> 10/100/1000:
>>>>>> Nothing in gigabit afaics.
>>>>>>
>>>>>> Switches:
>>>>>> 8862 - Not affected?
>>>>>> 8863 - Errata. PDOWN broken. Reset will not help. Workaround exists.
>>>>>> 8864 - Not affected?
>>>>>> 8873 - Errata. PDOWN broken. Reset will not help. Workaround exists.
>>>>>> 9477 - Errata. PDOWN broken. Will randomly cause link failure on
>>>>>> adjacent links. Do not use.
>>>>>>
>>>>>> This certainly explains a lot.
>>>>>>
>>>>>>>>>
>>>>>>>>> This patch changes PHY state to PHY_UP when MDIO bus resume back,
>>>>>>>>> it should be reasonable after PHY hardware re-initialized. Also
>>>>>>>>> give state machine a chance to start/config auto-nego again.
>>>>>>>>>
>>>>>>>>
>>>>>>>> If the MAC driver calls phy_stop() on suspend, then
>>>>>>>> phydev->suspended is true and mdio_bus_phy_may_suspend() returns
>>>>>>>> false. As a consequence
>>>>>>>> phydev->suspended_by_mdio_bus is false and mdio_bus_phy_resume()
>>>>>>>> skips the PHY hw initialization.
>>>>>>>> Please also note that mdio_bus_phy_suspend() calls
>>>>>>>> phy_stop_machine() that sets the state to PHY_UP.
>>>>>>>>
>>>>>>>
>>>>>>> Forgot that MDIO bus suspend is done before MAC driver suspend.
>>>>>>> Therefore disregard this part for now.
>>>>>>>
>>>>>>>> Having said that the current argumentation isn't convincing. I'm
>>>>>>>> not aware of such issues on other systems, therefore it's likely
>>>>>>>> that something is system-dependent.
>>>>>>>>
>>>>>>>> Please check the exact call sequence on your system, maybe it
>>>>>>>> provides a hint.
>>>>>>>>
>>>>>>>>> Signed-off-by: Joakim Zhang <qiangqing.zhang@xxxxxxx>
>>>>>>>>> ---
>>>>>>>>> drivers/net/phy/phy_device.c | 7 +++++++
>>>>>>>>> 1 file changed, 7 insertions(+)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/net/phy/phy_device.c
>>>>>>>>> b/drivers/net/phy/phy_device.c index cc38e326405a..312a6f662481
>>>>>>>>> 100644
>>>>>>>>> --- a/drivers/net/phy/phy_device.c
>>>>>>>>> +++ b/drivers/net/phy/phy_device.c
>>>>>>>>> @@ -306,6 +306,13 @@ static __maybe_unused int
>>>>> mdio_bus_phy_resume(struct device *dev)
>>>>>>>>> ret = phy_resume(phydev);
>>>>>>>>> if (ret < 0)
>>>>>>>>> return ret;
>>>>>>>>> +
>>>>>>>>> + /* PHY state could be changed to PHY_NOLINK from MAC controller
>>>>> resume
>>>>>>>>> + * rounte with phy_start(), here change to PHY_UP after
>>>>> re-initializing
>>>>>>>>> + * PHY hardware, let PHY state machine to start/config
>>>>>>>>> +auto-nego
>>>>> again.
>>>>>>>>> + */
>>>>>>>>> + phydev->state = PHY_UP;
>>>>>>>>> +
>>>>>>>>> no_resume:
>>>>>>>>> if (phydev->attached_dev && phydev->adjust_link)
>>>>>>>>> phy_start_machine(phydev);
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>> This is a quick draft of the modified soft reset for KSZ8081.
>>>>> Some tests would be appreciated.
>>>>>
>>>>
>>>> I test this patch at my side, unfortunately, it still can't work.
>>>>
>>>> I have not found what soft reset would affect in 8081 spec, is there
>>>> ang common Description for different PHYs?
>>>>
>>>
>>> You can check the clause 22 spec: 802.3 22.2.4.1.1
>>>
>>> Apart from that you can check BMCR value and which modes your PHY
>>> advertises after a soft reset. If the PHY indeed wouldn't restart aneg after a
>>> soft reset, then it would be the only one with this behavior I know. And I'd
>>> wonder why there aren't more such reports.
>>
>> Hi Heiner,
>>
>> We have reasons to believe it is a weird behavior of KSZ8081. I have another two PHYs at my side, AR8031 and RTL8211FD,
>> they can work fine if soft_reset implemented.
>>
>> As commit message described, phy_start() would change PHY state to PHY_UP finally to PHY_NOLINK when MAC resume.
>> After MDIO bus resume back, it will periodically check link status. I did below code change to dump all PHY registers.
>>
>> --- a/drivers/net/phy/phy.c
>> +++ b/drivers/net/phy/phy.c
>> @@ -35,7 +35,7 @@
>> #include <net/genetlink.h>
>> #include <net/sock.h>
>>
>> -#define PHY_STATE_TIME HZ
>> +#define PHY_STATE_TIME (10 * HZ)
>>
>> #define PHY_STATE_STR(_state) \
>> case PHY_##_state: \
>> @@ -738,6 +738,28 @@ static int phy_check_link_status(struct phy_device *phydev)
>> if (err)
>> return err;
>>
>> + printk("offset 0x00 data = %0x\n", phy_read(phydev, 0x00));
>> + printk("offset 0x01 data = %0x\n", phy_read(phydev, 0x01));
>> + printk("offset 0x02 data = %0x\n", phy_read(phydev, 0x02));
>> + printk("offset 0x03 data = %0x\n", phy_read(phydev, 0x03));
>> + printk("offset 0x04 data = %0x\n", phy_read(phydev, 0x04));
>> + printk("offset 0x05 data = %0x\n", phy_read(phydev, 0x05));
>> + printk("offset 0x06 data = %0x\n", phy_read(phydev, 0x06));
>> + printk("offset 0x07 data = %0x\n", phy_read(phydev, 0x07));
>> + printk("offset 0x08 data = %0x\n", phy_read(phydev, 0x08));
>> + printk("offset 0x09 data = %0x\n", phy_read(phydev, 0x09));
>> + printk("offset 0x10 data = %0x\n", phy_read(phydev, 0x10));
>> + printk("offset 0x11 data = %0x\n", phy_read(phydev, 0x11));
>> + printk("offset 0x15 data = %0x\n", phy_read(phydev, 0x15));
>> + printk("offset 0x16 data = %0x\n", phy_read(phydev, 0x16));
>> + printk("offset 0x17 data = %0x\n", phy_read(phydev, 0x17));
>> + printk("offset 0x18 data = %0x\n", phy_read(phydev, 0x18));
>> + printk("offset 0x1b data = %0x\n", phy_read(phydev, 0x1b));
>> + printk("offset 0x1c data = %0x\n", phy_read(phydev, 0x1c));
>> + printk("offset 0x1d data = %0x\n", phy_read(phydev, 0x1d));
>> + printk("offset 0x1e data = %0x\n", phy_read(phydev, 0x1e));
>> + printk("offset 0x1f data = %0x\n", phy_read(phydev, 0x1f));
>> + printk("\n\n");
>> if (phydev->link && phydev->state != PHY_RUNNING) {
>> phy_check_downshift(phydev);
>> phydev->state = PHY_RUNNING;
>>
>> After MDIO bus resume back, results as below:
>>
>> [ 119.545383] offset 0x00 data = 1100
>> [ 119.549237] offset 0x01 data = 7849
>> [ 119.563125] offset 0x02 data = 22
>> [ 119.566810] offset 0x03 data = 1560
>> [ 119.570597] offset 0x04 data = 85e1
>> [ 119.588016] offset 0x05 data = 0
>> [ 119.592891] offset 0x06 data = 4
>> [ 119.596452] offset 0x07 data = 2001
>> [ 119.600233] offset 0x08 data = 0
>> [ 119.617864] offset 0x09 data = 0
>> [ 119.625990] offset 0x10 data = 0
>> [ 119.629576] offset 0x11 data = 0
>> [ 119.642735] offset 0x15 data = 0
>> [ 119.646332] offset 0x16 data = 202
>> [ 119.650030] offset 0x17 data = 5c02
>> [ 119.668054] offset 0x18 data = 801
>> [ 119.672997] offset 0x1b data = 0
>> [ 119.676565] offset 0x1c data = 0
>> [ 119.680084] offset 0x1d data = 0
>> [ 119.698031] offset 0x1e data = 20
>> [ 119.706246] offset 0x1f data = 8190
>> [ 119.709984]
>> [ 119.709984]
>> [ 120.182120] offset 0x00 data = 100
>> [ 120.185894] offset 0x01 data = 784d
>> [ 120.189681] offset 0x02 data = 22
>> [ 120.206319] offset 0x03 data = 1560
>> [ 120.210171] offset 0x04 data = 8061
>> [ 120.225353] offset 0x05 data = 0
>> [ 120.228948] offset 0x06 data = 4
>> [ 120.242936] offset 0x07 data = 2001
>> [ 120.246792] offset 0x08 data = 0
>> [ 120.250313] offset 0x09 data = 0
>> [ 120.266748] offset 0x10 data = 0
>> [ 120.270335] offset 0x11 data = 0
>> [ 120.284167] offset 0x15 data = 0
>> [ 120.287760] offset 0x16 data = 202
>> [ 120.301031] offset 0x17 data = 3c02
>> [ 120.313209] offset 0x18 data = 801
>> [ 120.316983] offset 0x1b data = 0
>> [ 120.320513] offset 0x1c data = 1
>> [ 120.336589] offset 0x1d data = 0
>> [ 120.340184] offset 0x1e data = 115
>> [ 120.355357] offset 0x1f data = 8190
>> [ 120.359112]
>> [ 120.359112]
>> [ 129.785396] offset 0x00 data = 1100
>> [ 129.789252] offset 0x01 data = 7849
>> [ 129.809951] offset 0x02 data = 22
>> [ 129.815018] offset 0x03 data = 1560
>> [ 129.818845] offset 0x04 data = 85e1
>> [ 129.835808] offset 0x05 data = 0
>> [ 129.839398] offset 0x06 data = 4
>> [ 129.854514] offset 0x07 data = 2001
>> [ 129.858371] offset 0x08 data = 0
>> [ 129.873357] offset 0x09 data = 0
>> [ 129.876954] offset 0x10 data = 0
>> [ 129.880472] offset 0x11 data = 0
>> [ 129.896450] offset 0x15 data = 0
>> [ 129.900038] offset 0x16 data = 202
>> [ 129.915041] offset 0x17 data = 5c02
>> [ 129.918889] offset 0x18 data = 801
>> [ 129.932735] offset 0x1b data = 0
>> [ 129.946830] offset 0x1c data = 0
>> [ 129.950424] offset 0x1d data = 0
>> [ 129.964585] offset 0x1e data = 0
>> [ 129.968192] offset 0x1f data = 8190
>> [ 129.972938]
>> [ 129.972938]
>> [ 130.425125] offset 0x00 data = 100
>> [ 130.428889] offset 0x01 data = 784d
>> [ 130.442671] offset 0x02 data = 22
>> [ 130.446360] offset 0x03 data = 1560
>> [ 130.450142] offset 0x04 data = 8061
>> [ 130.467207] offset 0x05 data = 0
>> [ 130.470789] offset 0x06 data = 4
>> [ 130.485071] offset 0x07 data = 2001
>> [ 130.488934] offset 0x08 data = 0
>> [ 130.504320] offset 0x09 data = 0
>> [ 130.507911] offset 0x10 data = 0
>> [ 130.520950] offset 0x11 data = 0
>> [ 130.532865] offset 0x15 data = 0
>> [ 130.536461] offset 0x16 data = 202
>> [ 130.540156] offset 0x17 data = 3c02
>> [ 130.557218] offset 0x18 data = 801
>> [ 130.560987] offset 0x1b data = 0
>> [ 130.575158] offset 0x1c data = 1
>> [ 130.578754] offset 0x1d data = 0
>> [ 130.593543] offset 0x1e data = 115
>> [ 130.597312] offset 0x1f data = 8190
>>
>> We can see that BMCR and BMSR changes from 0x1100,0x7849 to 0x100,0x784d (BMCR[12] bit and BMSR[2]), and loop it.
>> Above process is auto-nego enabled, link is down, auto-nego is disabled, link is up. And auto-nego complete bit is always 0.
>>
>> PHY seems become unstable if soft reset after start/config auto-nego. I also want to fix it in micrel driver, but failed.
>>
>
> Waiting for ANEG_COMPLETE to be set wouldn't be a good option. Aneg may never
> complete for different reasons, e.g. no physical link. And even if we use a
> timeout this may add unwanted delays.
>
>> Do you have any other insights that can help me further locate the issue? Thanks.
>>
>
> I think current MAC/PHY PM handling isn't perfect. Often we have the following
> scenario:
>
> *suspend*
> 1. PHY is suspended (mdio_bus_phy_suspend)
> 2. MAC suspend callback (typically involving phy_stop())
>
> *resume*
> 1. MAC resume callback (typically involving phy_start())
> 2. PHY is resumed (mdio_bus_phy_resume), incl. calling phy_init_hw()
>
> Calling phy_init_hw() after phy_start() doesn't look right.
> It seems to work in most cases, but there's a certain risk
> that phy_init_hw() overwrites something, e.g. the advertised
> modes.
> I think we have two valid scenarios:
>
> 1. phylib PM callbacks are used, then the MAC driver shouldn't
> touch the PHY in its PM callbacks, especially not call
> phy_stop/phy_start.
>
> 2. MAC PM callbacks take care also of the PHY. Then I think we would
> need a flag at the phy_device telling it to make the PHY PM
> callbacks a no-op.
>
> Andrew, Russell: What do you think?
>
>> Best Regards,
>> Joakim Zhang
>>
>> [...]
>>
> Heiner
>

Based on scenario 1 you can also try the following.

diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
index 3db882322..cdf9160fc 100644
--- a/drivers/net/ethernet/freescale/fec_main.c
+++ b/drivers/net/ethernet/freescale/fec_main.c
@@ -3805,7 +3805,6 @@ static int __maybe_unused fec_suspend(struct device *dev)
if (netif_running(ndev)) {
if (fep->wol_flag & FEC_WOL_FLAG_ENABLE)
fep->wol_flag |= FEC_WOL_FLAG_SLEEP_ON;
- phy_stop(ndev->phydev);
napi_disable(&fep->napi);
netif_tx_lock_bh(ndev);
netif_device_detach(ndev);
@@ -3864,7 +3863,6 @@ static int __maybe_unused fec_resume(struct device *dev)
netif_device_attach(ndev);
netif_tx_unlock_bh(ndev);
napi_enable(&fep->napi);
- phy_start(ndev->phydev);
}
rtnl_unlock();

--
2.31.1