[PATCH] drm/nouveau/therm: mask the update bit when reading back the fan PWM duty

[PATCH] drm/nouveau/therm: mask the update bit when reading back the fan PWM duty

[Alex]

nv50_fan_pwm_set() and gf119_fan_pwm_set() write the fan duty cycle to
0xe118 with bit 31 set.  Normally the bit reads back clear, but when
the register is read back shortly after such a write it can still be
set, and nv50_fan_pwm_get() and gf119_fan_pwm_get() return it as part
of the duty cycle.  nvkm_fanpwm_get() then takes divs = max(divs, duty),
so divs becomes 0x8000xxxx too, and duty * 100 wraps around in u32
(0x80000000 * 100 == 0 mod 2^32): the fan speed reads as 0%.
nvkm_fan_update() uses that as the current speed and steps up from it,
so it writes 3% to the fan and then ramps it back by 3% per bump
period.

The reads that hit this come from the two therm timers.
nvkm_therm_alarm() (the 1 s poll) and nvkm_fan_alarm() (the bump
alarm) both call nvkm_fan_update(), which reads the duty cycle before
it writes a new one.  When the two alarms fire within microseconds of
each other, the second one reads 0xe118 right after the first one
wrote it.  On the board below the bump period is 500 ms, so this
happens again and again while the temperature target keeps moving.

Seen on a GeForce GTX 560 Ti (GF114, nv50 PWM path) while the
temperature changed quickly under load: a bpftrace probe on
nv50_fan_pwm_get() saw 7 of 239 reads in two minutes come 7-25 us after
the other timer's write with bit 31 still set (e.g. 0x800001e1 with
divs 0x21c), and each of them was followed by a drop to 3%.

gm107_fan_pwm_get() and the PTHERM path of gf119_fan_pwm_get() already
mask their duty reads; do the same for 0xe118.

Testing: with this change the same workload and probe gave no such
reads in 155 and no fan drops.  The board was running out-of-tree
reclocking patches; the therm and fan code was unmodified.  The gf119
hunk changes the same register in the same way but was not tested on
GF119 or later hardware.  Both files build without warnings.

The problem was found and analysed, the bpftrace probes were written,
and this fix and changelog were drafted with an AI assistant (Claude
Opus 5.5 by Anthropic) during an interactive debugging session.  I
reviewed the change and ran the tests on my card.

Fixes: cb9fa62671ac ("drm/nv50/pm: add support for pwm fan control")
Fixes: bc79202f4c8f ("nvd0/therm: implement more appropriate pwm fan control functions")
Cc: stable@vger.kernel.org
---
 drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c | 6 +++++-
 drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c  | 6 +++++-
 2 files changed, 10 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c b/drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c
index 684aff7..f56e763 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c
@@ -74,7 +74,11 @@ gf119_fan_pwm_get(struct nvkm_therm *therm, int line, u32 *divs, u32 *duty)
 	else if (indx < 2) {
 		if (nvkm_rd32(device, 0x00d610 + (line * 0x04)) & 0x00000040) {
 			*divs = nvkm_rd32(device, 0x00e114 + (indx * 8));
-			*duty = nvkm_rd32(device, 0x00e118 + (indx * 8));
+			/*
+			 * Bit 31 is the update trigger set by pwm_set();
+			 * it can still read back as set after a write.
+			 */
+			*duty = nvkm_rd32(device, 0x00e118 + (indx * 8)) & 0x7fffffff;
 			return 0;
 		}
 	} else if (indx == 2) {
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c b/drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c
index 9cf16a7..9272256 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c
@@ -72,7 +72,11 @@ nv50_fan_pwm_get(struct nvkm_therm *therm, int line, u32 *divs, u32 *duty)
 
 	if (nvkm_rd32(device, ctrl) & (1 << line)) {
 		*divs = nvkm_rd32(device, 0x00e114 + (id * 8));
-		*duty = nvkm_rd32(device, 0x00e118 + (id * 8));
+		/*
+		 * Bit 31 is the update trigger set by pwm_set();
+		 * it can still read back as set after a write.
+		 */
+		*duty = nvkm_rd32(device, 0x00e118 + (id * 8)) & 0x7fffffff;
 		return 0;
 	}
 
-- 
2.53.0

Re: [PATCH] drm/nouveau/therm: mask the update bit when reading back the fan PWM duty

[Alex]

I forgot to add my sign-off to this patch, sorry about that:

Signed-off-by: Alex <xnderv@yandex.com>

Re: [PATCH] drm/nouveau/therm: mask the update bit when reading back the fan PWM duty

[Martin Roukala]

Hi Alex,

On 25/10/2026 12:16, Alex wrote:
> nv50_fan_pwm_set() and gf119_fan_pwm_set() write the fan duty cycle to
> 0xe118 with bit 31 set.  Normally the bit reads back clear, but when
> the register is read back shortly after such a write it can still be
> set, and nv50_fan_pwm_get() and gf119_fan_pwm_get() return it as part
> of the duty cycle.  nvkm_fanpwm_get() then takes divs = max(divs, duty),
> so divs becomes 0x8000xxxx too, and duty * 100 wraps around in u32
> (0x80000000 * 100 == 0 mod 2^32): the fan speed reads as 0%.
> nvkm_fan_update() uses that as the current speed and steps up from it,
> so it writes 3% to the fan and then ramps it back by 3% per bump
> period.

Ooops, I never saw the update bit being sticky! I thought this was a 
write-only bit :s Anyway, the code is:

    Acked-by: Martin Roukala <martin.roukala@mupuf.org>

Now, let's talk about the commit message: LLMs love to talk don't 
they?Let's trim the commit message to the just the above as it provides 
everything anyone bisecting to this change could ever want to know.

Whatever comes after (but not including the `Fixes:` and `Cc:` trailers) 
could be sent outside of the commit message, just as part of the email. 
It is valuable to the reviewer, as it allows us to see how you found the 
issue, figured out how to fix it, and how you tested it.

Anyway, thanks for your contribution. I am happy that you are taking 
over this old project of mine (back when I was contributing as Martin 
Peres, before I took my now-wife's name) :)

Thanks,
Martin Roukala
> 
> The reads that hit this come from the two therm timers.
> nvkm_therm_alarm() (the 1 s poll) and nvkm_fan_alarm() (the bump
> alarm) both call nvkm_fan_update(), which reads the duty cycle before
> it writes a new one.  When the two alarms fire within microseconds of
> each other, the second one reads 0xe118 right after the first one
> wrote it.  On the board below the bump period is 500 ms, so this
> happens again and again while the temperature target keeps moving.
> 
> Seen on a GeForce GTX 560 Ti (GF114, nv50 PWM path) while the
> temperature changed quickly under load: a bpftrace probe on
> nv50_fan_pwm_get() saw 7 of 239 reads in two minutes come 7-25 us after
> the other timer's write with bit 31 still set (e.g. 0x800001e1 with
> divs 0x21c), and each of them was followed by a drop to 3%.
> 
> gm107_fan_pwm_get() and the PTHERM path of gf119_fan_pwm_get() already
> mask their duty reads; do the same for 0xe118.
> 
> Testing: with this change the same workload and probe gave no such
> reads in 155 and no fan drops.  The board was running out-of-tree
> reclocking patches; the therm and fan code was unmodified.  The gf119
> hunk changes the same register in the same way but was not tested on
> GF119 or later hardware.  Both files build without warnings.
> 
> The problem was found and analysed, the bpftrace probes were written,
> and this fix and changelog were drafted with an AI assistant (Claude
> Opus 5.5 by Anthropic) during an interactive debugging session.  I
> reviewed the change and ran the tests on my card.
> 
> Fixes: cb9fa62671ac ("drm/nv50/pm: add support for pwm fan control")
> Fixes: bc79202f4c8f ("nvd0/therm: implement more appropriate pwm fan control functions")
> Cc: stable@vger.kernel.org
> ---
>   drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c | 6 +++++-
>   drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c  | 6 +++++-
>   2 files changed, 10 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c b/drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c
> index 684aff7..f56e763 100644
> --- a/drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c
> +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/therm/gf119.c
> @@ -74,7 +74,11 @@ gf119_fan_pwm_get(struct nvkm_therm *therm, int line, u32 *divs, u32 *duty)
>   	else if (indx < 2) {
>   		if (nvkm_rd32(device, 0x00d610 + (line * 0x04)) & 0x00000040) {
>   			*divs = nvkm_rd32(device, 0x00e114 + (indx * 8));
> -			*duty = nvkm_rd32(device, 0x00e118 + (indx * 8));
> +			/*
> +			 * Bit 31 is the update trigger set by pwm_set();
> +			 * it can still read back as set after a write.
> +			 */
> +			*duty = nvkm_rd32(device, 0x00e118 + (indx * 8)) & 0x7fffffff;
>   			return 0;
>   		}
>   	} else if (indx == 2) {
> diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c b/drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c
> index 9cf16a7..9272256 100644
> --- a/drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c
> +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/therm/nv50.c
> @@ -72,7 +72,11 @@ nv50_fan_pwm_get(struct nvkm_therm *therm, int line, u32 *divs, u32 *duty)
>   
>   	if (nvkm_rd32(device, ctrl) & (1 << line)) {
>   		*divs = nvkm_rd32(device, 0x00e114 + (id * 8));
> -		*duty = nvkm_rd32(device, 0x00e118 + (id * 8));
> +		/*
> +		 * Bit 31 is the update trigger set by pwm_set();
> +		 * it can still read back as set after a write.
> +		 */
> +		*duty = nvkm_rd32(device, 0x00e118 + (id * 8)) & 0x7fffffff;
>   		return 0;
>   	}
>