Re: [BUG] hwmon: (spd5118) resume fails -ENXIO on write protected hubs

[GG]

Sorry - I should have searched the archives before sending my last email. I've since
found this from February, which is the same problem:

  https://lore.kernel.org/linux-hwmon/20260205102942.28745-1-tinsaetadesse2015@xxxxxxxxx/

And I missed the obvious clue sitting in my own boot log:

  i801_smbus 0000:00:1f.4: SPD Write Disable is set

So my kernel already knew. Apologies for the noise.

One thing I'll mention in case it's useful: that series above has spd5118 refuse to
probe when writes are disabled. On my machine that would lose me two working
temperature sensors. Reads are completely fine here, I get correct temps,
and my temperature limits read back properly. It's only writing that fails.

Happy to test anything on my hardware if it would help.

Gabriel G

On Monday, August 17th, 2026 at 12:29 PM, GG <ggirouard@xxxxxxxxxxxxxx> wrote:

> Hi there,
> 
> I apologize in advance if this isn't the right place to send this. I ran into
> a quirk on my own machine, went digging, and ended up working out what causes
> it, so I figured it was probably worth sharing. Please feel free to ignore all
> of this if it's not useful or doesn't apply.
> 
> Before you read any further, three things you should probably know:
> 
>   - I'm not a programmer. I can tell you what my hardware does, and I ran
>     every test below myself on it, but if you ask me to defend the actual
>     implementation I'll be out of my depth pretty quickly.
> 
>   - I worked this out with an AI coding assistant (Claude, claude-opus-5) over
>     a long back-and-forth debugging session. It suggested things to try and
>     wrote the instrumented test modules, and I built and ran each one on this
>     machine and fed the results back. It drafted this write-up too, and I've
>     edited it.
> 
>   - So this isn't a formal patch submission and there's deliberately no
>     Signed-off-by on it. Treat it as a bug report that happens to come with a
>     suggestion attached, and please feel free to bin the patch entirely and do
>     something completely different.
> 
> The one thing I can offer that might actually be worth something is the
> hardware. This machine reproduces the problem every single time, so I'm happy
> to test whatever you'd like, in whatever form.
> 
> Anyway, here's what I found.
> 
> spd5118_resume() fails with -ENXIO on every single suspend/resume cycle here,
> and I think it's because the SPD hubs on this board are write protected by the
> platform firmware.
> 
> Hardware:
>   Gigabyte Z790 AORUS PRO X, BIOS F11a
>   Intel i9-14900KS
>   2x 32GB DDR5, hubs at i2c 12-0051 and 12-0053
>   Both report: DDR5 temperature sensor: vendor 0x06:0x32 revision 1.6
>   Kernel 7.1.8 (CachyOS)
> 
> This shows up on every resume:
> 
>   spd5118 12-0051: Failed to write b = 0: -6
>   spd5118 12-0051: PM: dpm_run_callback(): spd5118_resume returns -6
>   spd5118 12-0051: PM: spd5118_resume returned -6 after 15085 usecs
>   spd5118 12-0051: PM: failed to resume async: error -6
> 
> The sensor itself works fine either side of a suspend and the configured
> limits survive, so nothing is actually broken for me as a user. It's just the
> error, plus 5-18ms of pointless failing SMBus traffic per DIMM every time.
> 
> The thing that finally made it click is that this has nothing to do with
> suspend or resume at all. These hubs refuse writes all the time. Writing
> temp1_max back to the exact value it already holds, on a fully booted and idle
> system, fails the same way:
> 
>   $ cat /sys/class/hwmon/hwmon3/temp1_max
>   55000
>   $ sudo sh -c 'echo 55000 > /sys/class/hwmon/hwmon3/temp1_max'
>   sh: line 1: echo: write error: No such device or address
> 
> Reads are completely unaffected. During resume, raw i2c_smbus_read_byte_data()
> on MR11 and on every other register succeeds on the first try and returns
> sensible values. Every single i2c_smbus_write_byte_data() comes back -ENXIO.
> 
> Which means this, in spd5118_suspend():
> 
>   regmap_update_bits(regmap, SPD5118_REG_TEMP_CONFIG, SPD5118_TS_DISABLE,
>                      SPD5118_TS_DISABLE);
> 
> is failing silently, because its return value isn't checked. The sensor never
> actually gets disabled. But the cache still gets marked dirty, so on resume
> regcache_sync() goes off to restore state that was never changed in the first
> place.
> 
> The error message itself sent me down the wrong path for quite a while. It
> doesn't come from the sync, it comes from regcache_sync()'s epilogue, which
> force-writes every regmap range selector and then hands that write's error
> back as its own return value. For spd5118 in I2C legacy mode that selector is
> MR11, which is also a cached writeable register, and with window_start 0 and
> window_len 0x100 it sits inside its own paging window. That's why the message
> singles out register 0xb, which had me convinced for ages that there was
> something specifically wrong with MR11.
> 
> Dead ends, in case it saves anyone else the time:
> 
>   - Dropping MR11 from the regcache before the sync does silence the "Failed
>     to write b = 0" message, but resume still fails, because the next register
>     isn't writeable either.
>   - MR11 already reads back as 0 on resume, so there's no stale page.
>   - Reads succeed on the first attempt, so the hub isn't slow to wake up.
>   - Polling for a write to be accepted for up to 2 seconds never got one
>     through, and added 3 seconds to resume.
>   - Pinning i2c-i801 to power/control=on for the whole cycle changes nothing.
>     The adapter isn't marked suspended at that point either:
>     __i2c_check_suspended() (drivers/i2c/i2c-core.h) returns -ESHUTDOWN and
>     emits a "Transfer while suspended" dev_WARN, and neither of those shows up
>     anywhere in my logs.
> 
> The patch below is what ended up working for me here: check whether the device
> actually accepted that suspend-time write, and if it didn't, don't bother
> marking the cache dirty or syncing on resume. With it applied, resume returns
> 0 in 0us and the error is gone. I ran 5 suspend/resume cycles on a fresh boot
> with no debug kernel options, and /sys/kernel/debug/suspend_stats went from
> success=0 failed_resume=0 to success=5 failed_resume=0, with temperatures and
> limits unaffected throughout.
> 
> One thing I can't check myself: I don't have any hardware with a hub that does
> accept writes, so I can't confirm the normal path still behaves properly on a
> machine like that.
> 
> Candidate change against drivers/hwmon/spd5118.c:
> 
> --- a/drivers/hwmon/spd5118.c
> +++ b/drivers/hwmon/spd5118.c
> @@ -79,6 +79,7 @@ struct spd5118_data {
>  	struct regmap *regmap;
>  	struct mutex nvmem_lock;
>  	bool is_16bit;
> +	bool suspend_wp;	/* device NAKed the suspend write: nothing to restore */
>  };
> 
> @@ -498,10 +499,27 @@ static int spd5118_suspend(struct device *dev)
>  		return err;
> 
>  	regcache_cache_bypass(regmap, true);
> -	regmap_update_bits(regmap, SPD5118_REG_TEMP_CONFIG, SPD5118_TS_DISABLE,
> -			   SPD5118_TS_DISABLE);
> +	err = regmap_update_bits(regmap, SPD5118_REG_TEMP_CONFIG,
> +				 SPD5118_TS_DISABLE, SPD5118_TS_DISABLE);
>  	regcache_cache_bypass(regmap, false);
> 
> +	/*
> +	 * Platform firmware may write protect the SPD hub after memory
> +	 * training, in which case the device serves reads but NAKs every write
> +	 * with -ENXIO. The temperature sensor was then never disabled, so
> +	 * there is no device state to restore on resume.
> +	 */
> +	data->suspend_wp = err < 0;
> +	if (data->suspend_wp) {
> +		dev_dbg(dev, "write protected, skipping state save (%d)\n", err);
> +		/* Still avoid bus traffic while suspended. */
> +		regcache_cache_only(regmap, true);
> +		return 0;
> +	}
> +
>  	regcache_cache_only(regmap, true);
>  	regcache_mark_dirty(regmap);
> 
> @@ -514,6 +532,10 @@ static int spd5118_resume(struct device *dev)
>  	struct regmap *regmap = data->regmap;
> 
>  	regcache_cache_only(regmap, false);
> +
> +	if (data->suspend_wp)
> +		return 0;
> +
>  	return regcache_sync(regmap);
>  }
> 
> As far as I can tell this goes back to d1b4c755081a ("hwmon: (spd5118) Add
> suspend/resume support"), which first shipped in v6.11.
> 
> Thanks for taking a look, and no worries at all if this isn't useful.
> 
> Gabriel G