Re: [PATCH 1/3] hwmon: spd5118: Do not fail resume on temporary I2C errors

[Guenter Roeck <linux@roeck-us.net>]

On 1/12/26 10:22, Armin Wolf wrote:
> Am 12.01.26 um 19:17 schrieb Guenter Roeck:
> 
>> On 1/12/26 09:46, Armin Wolf wrote:
>>> Am 12.01.26 um 17:36 schrieb Guenter Roeck:
>>>
>>>> On 1/10/26 14:27, Armin Wolf wrote:
>>>>> Am 10.01.26 um 18:19 schrieb Tinsae Tadesse:
>>>>>
>>>>>> SPD5118 DDR5 temperature sensors may be temporarily unavailable
>>>>>> during s2idle resume. Ignore temporary -ENXIO and -EIO errors during resume and allow
>>>>>> register synchronization to be retried later.
>>>>>
>>>>> Hi,
>>>>>
>>>>> do you know if the error is caused by the SPD5118 device itself or by the underlying
>>>>> i2c controller? Please also share the output of "acpidump" and the name of the i2c
>>>>> controller used to communicate with the SPD5118.
>>>>>
>>>>>> Signed-off-by: Tinsae Tadesse <tinsaetadesse2015@gmail.com>
>>>>>> ---
>>>>>>   drivers/hwmon/spd5118.c | 8 +++++++-
>>>>>>   1 file changed, 7 insertions(+), 1 deletion(-)
>>>>>>
>>>>>> diff --git a/drivers/hwmon/spd5118.c b/drivers/hwmon/spd5118.c
>>>>>> index 5da44571b6a0..ec9f14f6e0df 100644
>>>>>> --- a/drivers/hwmon/spd5118.c
>>>>>> +++ b/drivers/hwmon/spd5118.c
>>>>>> @@ -512,9 +512,15 @@ static int spd5118_resume(struct device *dev)
>>>>>>   {
>>>>>>       struct spd5118_data *data = dev_get_drvdata(dev);
>>>>>>       struct regmap *regmap = data->regmap;
>>>>>> +    int ret;
>>>>>>       regcache_cache_only(regmap, false);
>>>>>> -    return regcache_sync(regmap);
>>>>>> +    ret = regcache_sync(regmap);
>>>>>> +    if(ret == -ENXIO || ret == -EIO) {
>>>>>> +        dev_warn(dev, "SPD hub not responding on resume (%d), deferring init\n", ret);
>>>>>> +        return 0;
>>>>>> +    }
>>>>>
>>>>> The specification says that the SPD5118 might take up to 10ms to initialize its i2c interface
>>>>> after power on. Can you test if simply waiting for 10ms before syncing the regcache solves this
>>>>> issue?
>>>>>
>>>>
>>>> It seems to be highly unlikely that this code executes within 10ms of powering on the memory.
>>>>
>>>> Guenter
>>>>
>>> AFAIK the 10ms are associated with the VDDIO supply, the VDDSPD main supply is different from that.
>>> I just want to test if this device disables VDDIO during suspend-to-idle.
>>>
>>> I have another theory: if the SPD5118 somehow looses power, we might still need to manually put the
>>> device into 16-bit address mode using standard 8-bit i2c commands.
>>>
>>
>> Uh, no, we can not do that. I tried. Changing the access mode causes bad hiccups at least
>> with some boards. They interpret that as a memory configuration change, and the next warm
>> reboot will end up in the BIOS. Then, after the RAM configuration is updated, a cold reboot
>> will again detect a configuration change and BIOS will be entered again.
>>
>> That does make me wonder how the problem shows up in the first place, since the BIOS
>> usually does access the SPD5118 during resume, at least on my systems with DDR5. Granted,
>> those are with AMD CPUs, but I would assume that Intel BIOS versions are not different.
>>
>> Guenter
>>
> During suspend-to-idle the RAM stays active, so the firmware does not really need to access the SPD device.
> I meant that if the SPD device is configured during boot to operate in 16-bit mode and looses power during
> suspend-to-idle, the firmware might not reconfigure the SPD to continue operate in 16-bit mode after resume.
> 

This would be a severe BIOS/Firmware problem. I'd really want to see evidence that
restoring SPD5118_REG_I2C_LEGACY_MODE is necessary on resume before touching that
register. Even then I'd want to see evidence that touching it doesn't cause
problems on a variety of boards before actually doing it. After all, we would not
know if the register was reconfigured by the firmware for some reason, or if the
chip lost power and the firmware didn't handle it.

Having said that, even if there is evidence that the chip can lose power and end up
in the wrong mode on resume, I'd rather check for that condition, issue a WARN_ONCE(),
and disable the driver instead of touching its configuration.

Sorry, but from my experience SPD5118 access in general is quite fragile, and not
touching its configuration was the only way I could find to make it work reliably.

Guenter