Skip to contentSkip to conversation

drm/amd/pm/smu7: Don't send VddC_Request on CI

PATCH In review Petr Belyaev 2 participants 4 messages Aug 24 – Aug 25

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amd/pm/smu7: Don't send VddC_Request on CI

smu7_upload_dpm_level_enable_mask() calls smu7_apply_minimum_dce_voltage_request() unconditionally, which sends PPSMC_MSG_VddC_Request to the SMC. CI SMC firmware does not implement that message: it is only defined in smu7_ppsmc.h, tonga_ppsmc.h and fiji_ppsmc.h, and there is no CI equivalent. The SMC therefore never writes SMC_RESP, PHM_WAIT_FIELD_UNEQUAL() spins until it gives up, and every call burns the full ~1.9s response timeout.

Nothing downstream notices, because ci_send_msg_to_smc() returns 0 regardless of the outcome. The only effects are the delay and

amdgpu: failed to send message 15f ret is 0

in dmesg. On a Radeon R7 260X (BONAIRE, 1002:6658) this is 40 messages per boot: 32 during the boot-time modeset, which becomes ~59s of blank screen, and 8 more on every DPMS wake, adding ~15s before the display comes back.

Commit 0138610c1413 ("drm/amd/pm/smu7: Fix SMU7 voltage dependency on display clock") extended this feature to CI, which the previous implementation had deliberately excluded -- the old call site was guarded by pp_table_version == PP_TABLE_V1 with an adjacent comment reading "TO DO for v0 iceland and Ci". That commit was verified on a Radeon R9 M380, which is VI, so the CI path went untested.

Restore the PP_TABLE_V1 condition. With it back the failed messages disappear entirely and both the boot modeset and the monitor wake are immediate.

Fixes 0138610c1413 ("drm/amd/pm/smu7: Fix SMU7 voltage dependency on display clock")
---
 drivers/gpu/drm/amd/pm/powerplay/hwmgr/smu7_hwmgr.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/pm/powerplay/hwmgr/smu7_hwmgr.c +5 -1
index 6529a91..aadf6d5 100644
@@ -3189,7 +3189,11 @@ static int smu7_upload_dpm_level_enable_mask(struct pp_hwmgr *hwmgr)
 {
 	struct smu7_hwmgr *data = (struct smu7_hwmgr *)(hwmgr->backend);
 
-	smu7_apply_minimum_dce_voltage_request(hwmgr);
+	/* CI SMC firmware does not implement PPSMC_MSG_VddC_Request, and each
+	 * attempt costs a full SMC response timeout.
+	 */
+	if (hwmgr->pp_table_version == PP_TABLE_V1)
+		smu7_apply_minimum_dce_voltage_request(hwmgr);
 
 	if (!data->sclk_dpm_key_disabled) {
 		if (data->dpm_level_enable_mask.sclk_dpm_enable_mask)
Signature · 2 lines
2.54.0
:wqa
#1 lore raw
Quoted 7 lines · Central European Summer Time Petr Belyaev

smu7_upload_dpm_level_enable_mask() calls smu7_apply_minimum_dce_voltage_request() unconditionally, which sends PPSMC_MSG_VddC_Request to the SMC. CI SMC firmware does not implement that message: it is only defined in smu7_ppsmc.h, tonga_ppsmc.h and fiji_ppsmc.h, and there is no CI equivalent. The SMC therefore never writes SMC_RESP, PHM_WAIT_FIELD_UNEQUAL() spins until it gives up, and every call burns the full ~1.9s response timeout.

Hi,

I'm sorry but this is wrong.
I added the VDDC request specifically to solve an issue with a Bonaire GPU. That system will boot into a black screen without requesting a high enough voltage for the display engine. This patch would regress that.

Quoted 10 lines

Nothing downstream notices, because ci_send_msg_to_smc() returns 0 regardless of the outcome. The only effects are the delay and

amdgpu: failed to send message 15f ret is 0

in dmesg. On a Radeon R7 260X (BONAIRE, 1002:6658) this is 40 messages per boot: 32 during the boot-time modeset, which becomes ~59s of blank screen, and 8 more on every DPMS wake, adding ~15s before the display comes back.

How do you reproduce this issue, and on which kernel version? I've tested this patch on two different Bonaire chips and a Hawaii chip, and haven't seen this error message. In fact I am running an R9 390X right now and I don't see this error message in my logs.

Quoted 7 lines

Commit 0138610c1413 ("drm/amd/pm/smu7: Fix SMU7 voltage dependency on display clock") extended this feature to CI, which the previous implementation had deliberately excluded -- the old call site was guarded by pp_table_version == PP_TABLE_V1 with an adjacent comment reading "TO DO for v0 iceland and Ci". That commit was verified on a Radeon R9 M380, which is VI, so the CI path went untested.

The CI path definitely did not go untested.
R9 M380 is Bonaire (it's in an old iMac).
You can see all the details here:
https://gitlab.freedesktop.org/drm/amd/-/work_items/3425#note_3338643

Best regards,
Timur

#2 lore raw

Hi,

You're right, and I withdraw the patch - hadn't considered that gating it off would regress the systems it was added for. Apologies also for the M380 claim; I thought it was VI from the commit message rather than checking.

To reproduce:
Stock 6.18.41, Bonaire XTX 1002:6658 rev 00 (XFX 1682:7260), VBIOS 113-260XT-HY12832-D5-2G-10, SMC firmware 11.35.0 from bonaire_smc.bin. DC is not enabled here (amdgpu.dc left at the default).
I'm getting roughly 40 "failed to send message 15f ret is 0" messages per boot, ~1.9s each.
gitlab: https://gitlab.freedesktop.org/drm/amd/-/work_items/5678

One more observation: Your MCLK DPM quirk does fire on this card - i see the msg 'disabling MCLK DPM on quirky ASIC' in dmesg.

Would be happy to test anything that might solve the problem.

Best regards,
Petr

Quoted 51 lines · Timur Kristóf
Quoted 7 lines · Central European Summer Time Petr Belyaev

smu7_upload_dpm_level_enable_mask() calls smu7_apply_minimum_dce_voltage_request() unconditionally, which sends PPSMC_MSG_VddC_Request to the SMC. CI SMC firmware does not implement that message: it is only defined in smu7_ppsmc.h, tonga_ppsmc.h and fiji_ppsmc.h, and there is no CI equivalent. The SMC therefore never writes SMC_RESP, PHM_WAIT_FIELD_UNEQUAL() spins until it gives up, and every call burns the full ~1.9s response timeout.

Hi,

I'm sorry but this is wrong.
I added the VDDC request specifically to solve an issue with a Bonaire GPU. That system will boot into a black screen without requesting a high enough voltage for the display engine. This patch would regress that.

Quoted 10 lines

Nothing downstream notices, because ci_send_msg_to_smc() returns 0 regardless of the outcome. The only effects are the delay and

amdgpu: failed to send message 15f ret is 0

in dmesg. On a Radeon R7 260X (BONAIRE, 1002:6658) this is 40 messages per boot: 32 during the boot-time modeset, which becomes ~59s of blank screen, and 8 more on every DPMS wake, adding ~15s before the display comes back.

How do you reproduce this issue, and on which kernel version? I've tested this patch on two different Bonaire chips and a Hawaii chip, and haven't seen this error message. In fact I am running an R9 390X right now and I don't see this error message in my logs.

Quoted 7 lines

Commit 0138610c1413 ("drm/amd/pm/smu7: Fix SMU7 voltage dependency on display clock") extended this feature to CI, which the previous implementation had deliberately excluded -- the old call site was guarded by pp_table_version == PP_TABLE_V1 with an adjacent comment reading "TO DO for v0 iceland and Ci". That commit was verified on a Radeon R9 M380, which is VI, so the CI path went untested.

The CI path definitely did not go untested.
R9 M380 is Bonaire (it's in an old iMac).
You can see all the details here:
https://gitlab.freedesktop.org/drm/amd/-/work_items/3425#note_3338643

Best regards,
Timur

#3 lore raw
Quoted 6 lines · Central European Summer Time Petr Belyaev

Hi,

You're right, and I withdraw the patch - hadn't considered that gating it off would regress the systems it was added for. Apologies also for the M380 claim; I thought it was VI from the commit message rather than checking.

Unfortunately, it is difficult to find information about which graphics card uses which GPU exactly, and the information online (Wikipedia, Techpowerup) is sometimes wrong.

Quoted 7 lines

To reproduce:
Stock 6.18.41, Bonaire XTX 1002:6658 rev 00 (XFX 1682:7260), VBIOS 113-260XT-HY12832-D5-2G-10, SMC firmware 11.35.0 from bonaire_smc.bin. DC is not enabled here (amdgpu.dc left at the default).
I'm getting roughly 40 "failed to send message 15f ret is 0" messages per boot, ~1.9s each.
gitlab: https://gitlab.freedesktop.org/drm/amd/-/work_items/5678

Thank you Petr. Let's continue the conversation on the GitLab.

Quoted 2 lines

On Mon, Aug 24, 2026 at 5:07 PM Timur Kristóf <[email protected]>

Quoted 48 lines · wrote:
Quoted 48 lines

On Monday, August 24, 2026 4:47:14 PM Central European Summer Time Petr

Quoted 7 lines · Belyaev

smu7_upload_dpm_level_enable_mask() calls smu7_apply_minimum_dce_voltage_request() unconditionally, which sends PPSMC_MSG_VddC_Request to the SMC. CI SMC firmware does not implement that message: it is only defined in smu7_ppsmc.h, tonga_ppsmc.h and fiji_ppsmc.h, and there is no CI equivalent. The SMC therefore never writes SMC_RESP, PHM_WAIT_FIELD_UNEQUAL() spins until it gives up, and every call burns the full ~1.9s response timeout.

Hi,

I'm sorry but this is wrong.
I added the VDDC request specifically to solve an issue with a Bonaire GPU.
That system will boot into a black screen without requesting a high enough voltage for the display engine. This patch would regress that.

Quoted 9 lines

Nothing downstream notices, because ci_send_msg_to_smc() returns 0 regardless of the outcome. The only effects are the delay and

amdgpu: failed to send message 15f ret is 0

in dmesg. On a Radeon R7 260X (BONAIRE, 1002:6658) this is 40 messages per boot: 32 during the boot-time modeset, which becomes ~59s of blank screen, and 8 more on every DPMS wake, adding ~15s before the display comes back.

How do you reproduce this issue, and on which kernel version? I've tested this patch on two different Bonaire chips and a Hawaii chip, and haven't seen this error message. In fact I am running an R9 390X right now and I don't see this error message in my logs.

Quoted 6 lines

Commit 0138610c1413 ("drm/amd/pm/smu7: Fix SMU7 voltage dependency on display clock") extended this feature to CI, which the previous implementation had deliberately excluded -- the old call site was guarded by pp_table_version == PP_TABLE_V1 with an adjacent comment reading "TO DO for v0 iceland and Ci". That commit was verified on a Radeon R9 M380, which is VI, so the CI path went untested.

The CI path definitely did not go untested.
R9 M380 is Bonaire (it's in an old iMac).
You can see all the details here:
https://gitlab.freedesktop.org/drm/amd/-/work_items/3425#note_3338643

Best regards,
Timur