From: Bjorn Andersson <[email protected]>
Commit 'e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable
the domain")' aimed to make sure that a power-domain that is being
enabled without any particular performance-state requested will at least
turn the rail on, to avoid filling DeviceTree with otherwise unnecessary
required-opps properties.
But in the event that aggregation happens on a disabled power-domain, with
an enabled peer without performance-state, both the local and peer
corner are 0. The peer's enabled_corner is not considered, with the
result that the underlying (shared) resource is disabled.
One case where this can be observed is when the display stack keeps mmcx
enabled (but without a particular performance-state vote) in order to
access registers and sync_state happens in the rpmhpd driver. As mmcx_ao
is flushed the state of the peer (mmcx) is not considered and mmcx_ao
ends up turning off "mmcx.lvl" underneath mmcx. This has been observed
several times, but has been painted over in DeviceTree by adding an
explicit vote for the lowest non-disabled performance-state.
Fixes: e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable the domain")
Reported-by: Johan Hovold <[email protected]>
Closes: https://lore.kernel.org/linux-arm-msm/[email protected]/
Cc: <[email protected]>
Signed-off-by: Bjorn Andersson <[email protected]>
---
This issue is the root cause of a display regression on SC8280XP boards,
resulting in the system often resetting during boot. It was exposed by
the refactoring of the DisplayPort driver in v6.8-rc1.
---
drivers/pmdomain/qcom/rpmhpd.c | 7 +++++--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/drivers/pmdomain/qcom/rpmhpd.c b/drivers/pmdomain/qcom/rpmhpd.c
index 3078896b1300..47df910645f6 100644
--- a/drivers/pmdomain/qcom/rpmhpd.c
+++ b/drivers/pmdomain/qcom/rpmhpd.c
@@ -692,6 +692,7 @@ static int rpmhpd_aggregate_corner(struct rpmhpd *pd, unsigned int corner)
unsigned int active_corner, sleep_corner;
unsigned int this_active_corner = 0, this_sleep_corner = 0;
unsigned int peer_active_corner = 0, peer_sleep_corner = 0;
+ unsigned int peer_enabled_corner;
if (pd->state_synced) {
to_active_sleep(pd, corner, &this_active_corner, &this_sleep_corner);
@@ -701,9 +702,11 @@ static int rpmhpd_aggregate_corner(struct rpmhpd *pd, unsigned int corner)
this_sleep_corner = pd->level_count - 1;
}
- if (peer && peer->enabled)
- to_active_sleep(peer, peer->corner, &peer_active_corner,
+ if (peer && peer->enabled) {
+ peer_enabled_corner = max(peer->corner, peer->enable_corner);
+ to_active_sleep(peer, peer_enabled_corner, &peer_active_corner,
&peer_sleep_corner);
+ }
active_corner = max(this_active_corner, peer_active_corner);
---
base-commit: b401b621758e46812da61fa58a67c3fd8d91de0d
change-id: 20240226-rpmhpd-enable-corner-fix-c5e07fe7b986
Best regards,
--
Bjorn Andersson <[email protected]>
On Tue, 27 Feb 2024 at 03:45, Bjorn Andersson via B4 Relay
<[email protected]> wrote:
>
> From: Bjorn Andersson <[email protected]>
>
> Commit 'e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable
> the domain")' aimed to make sure that a power-domain that is being
> enabled without any particular performance-state requested will at least
> turn the rail on, to avoid filling DeviceTree with otherwise unnecessary
> required-opps properties.
>
> But in the event that aggregation happens on a disabled power-domain, with
> an enabled peer without performance-state, both the local and peer
> corner are 0. The peer's enabled_corner is not considered, with the
> result that the underlying (shared) resource is disabled.
>
> One case where this can be observed is when the display stack keeps mmcx
> enabled (but without a particular performance-state vote) in order to
> access registers and sync_state happens in the rpmhpd driver. As mmcx_ao
> is flushed the state of the peer (mmcx) is not considered and mmcx_ao
> ends up turning off "mmcx.lvl" underneath mmcx. This has been observed
> several times, but has been painted over in DeviceTree by adding an
> explicit vote for the lowest non-disabled performance-state.
>
> Fixes: e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable the domain")
> Reported-by: Johan Hovold <[email protected]>
> Closes: https://lore.kernel.org/linux-arm-msm/[email protected]/
> Cc: <[email protected]>
> Signed-off-by: Bjorn Andersson <[email protected]>
> ---
> This issue is the root cause of a display regression on SC8280XP boards,
> resulting in the system often resetting during boot. It was exposed by
> the refactoring of the DisplayPort driver in v6.8-rc1.
> ---
> drivers/pmdomain/qcom/rpmhpd.c | 7 +++++--
> 1 file changed, 5 insertions(+), 2 deletions(-)
Reviewed-by: Dmitry Baryshkov <[email protected]>
Tested-by: Dmitry Baryshkov <[email protected]>
Probably once this lands we can drop explicit required-opps properties.
--
With best wishes
Dmitry
On 2/27/24 02:49, Bjorn Andersson via B4 Relay wrote:
> From: Bjorn Andersson <[email protected]>
>
> Commit 'e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable
> the domain")' aimed to make sure that a power-domain that is being
> enabled without any particular performance-state requested will at least
> turn the rail on, to avoid filling DeviceTree with otherwise unnecessary
> required-opps properties.
>
> But in the event that aggregation happens on a disabled power-domain, with
> an enabled peer without performance-state, both the local and peer
> corner are 0. The peer's enabled_corner is not considered, with the
> result that the underlying (shared) resource is disabled.
>
> One case where this can be observed is when the display stack keeps mmcx
> enabled (but without a particular performance-state vote) in order to
> access registers and sync_state happens in the rpmhpd driver. As mmcx_ao
> is flushed the state of the peer (mmcx) is not considered and mmcx_ao
> ends up turning off "mmcx.lvl" underneath mmcx. This has been observed
> several times, but has been painted over in DeviceTree by adding an
> explicit vote for the lowest non-disabled performance-state.
>
> Fixes: e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable the domain")
> Reported-by: Johan Hovold <[email protected]>
> Closes: https://lore.kernel.org/linux-arm-msm/[email protected]/
> Cc: <[email protected]>
> Signed-off-by: Bjorn Andersson <[email protected]>
> ---
> This issue is the root cause of a display regression on SC8280XP boards,
> resulting in the system often resetting during boot. It was exposed by
> the refactoring of the DisplayPort driver in v6.8-rc1.
> ---
Very good find, thanks!
Reviewed-by: Konrad Dybcio <[email protected]>
Konrad
On 2/26/2024 5:49 PM, Bjorn Andersson via B4 Relay wrote:
> From: Bjorn Andersson <[email protected]>
>
> Commit 'e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable
> the domain")' aimed to make sure that a power-domain that is being
> enabled without any particular performance-state requested will at least
> turn the rail on, to avoid filling DeviceTree with otherwise unnecessary
> required-opps properties.
>
> But in the event that aggregation happens on a disabled power-domain, with
> an enabled peer without performance-state, both the local and peer
> corner are 0. The peer's enabled_corner is not considered, with the
> result that the underlying (shared) resource is disabled.
>
> One case where this can be observed is when the display stack keeps mmcx
> enabled (but without a particular performance-state vote) in order to
> access registers and sync_state happens in the rpmhpd driver. As mmcx_ao
> is flushed the state of the peer (mmcx) is not considered and mmcx_ao
> ends up turning off "mmcx.lvl" underneath mmcx. This has been observed
> several times, but has been painted over in DeviceTree by adding an
> explicit vote for the lowest non-disabled performance-state.
>
> Fixes: e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable the domain")
> Reported-by: Johan Hovold <[email protected]>
> Closes: https://lore.kernel.org/linux-arm-msm/[email protected]/
> Cc: <[email protected]>
> Signed-off-by: Bjorn Andersson <[email protected]>
> ---
> This issue is the root cause of a display regression on SC8280XP boards,
> resulting in the system often resetting during boot. It was exposed by
> the refactoring of the DisplayPort driver in v6.8-rc1.
> ---
> drivers/pmdomain/qcom/rpmhpd.c | 7 +++++--
> 1 file changed, 5 insertions(+), 2 deletions(-)
>
Reviewed-by: Abhinav Kumar <[email protected]>
Quoting Bjorn Andersson via B4 Relay (2024-02-26 17:49:57)
> From: Bjorn Andersson <[email protected]>
>
> Commit 'e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable
> the domain")' aimed to make sure that a power-domain that is being
> enabled without any particular performance-state requested will at least
> turn the rail on, to avoid filling DeviceTree with otherwise unnecessary
> required-opps properties.
>
> But in the event that aggregation happens on a disabled power-domain, with
> an enabled peer without performance-state, both the local and peer
> corner are 0. The peer's enabled_corner is not considered, with the
> result that the underlying (shared) resource is disabled.
>
> One case where this can be observed is when the display stack keeps mmcx
> enabled (but without a particular performance-state vote) in order to
> access registers and sync_state happens in the rpmhpd driver. As mmcx_ao
> is flushed the state of the peer (mmcx) is not considered and mmcx_ao
> ends up turning off "mmcx.lvl" underneath mmcx. This has been observed
> several times, but has been painted over in DeviceTree by adding an
> explicit vote for the lowest non-disabled performance-state.
>
> Fixes: e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable the domain")
> Reported-by: Johan Hovold <[email protected]>
> Closes: https://lore.kernel.org/linux-arm-msm/[email protected]/
> Cc: <[email protected]>
> Signed-off-by: Bjorn Andersson <[email protected]>
> ---
Reviewed-by: Stephen Boyd <[email protected]>
On Mon, Feb 26, 2024 at 05:49:57PM -0800, Bjorn Andersson via B4 Relay wrote:
> From: Bjorn Andersson <[email protected]>
>
> Commit 'e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable
> the domain")' aimed to make sure that a power-domain that is being
> enabled without any particular performance-state requested will at least
> turn the rail on, to avoid filling DeviceTree with otherwise unnecessary
> required-opps properties.
>
> But in the event that aggregation happens on a disabled power-domain, with
> an enabled peer without performance-state, both the local and peer
> corner are 0. The peer's enabled_corner is not considered, with the
> result that the underlying (shared) resource is disabled.
>
> One case where this can be observed is when the display stack keeps mmcx
> enabled (but without a particular performance-state vote) in order to
> access registers and sync_state happens in the rpmhpd driver. As mmcx_ao
> is flushed the state of the peer (mmcx) is not considered and mmcx_ao
> ends up turning off "mmcx.lvl" underneath mmcx. This has been observed
> several times, but has been painted over in DeviceTree by adding an
> explicit vote for the lowest non-disabled performance-state.
>
> Fixes: e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable the domain")
> Reported-by: Johan Hovold <[email protected]>
> Closes: https://lore.kernel.org/linux-arm-msm/[email protected]/
> Cc: <[email protected]>
> Signed-off-by: Bjorn Andersson <[email protected]>
> ---
> This issue is the root cause of a display regression on SC8280XP boards,
> resulting in the system often resetting during boot. It was exposed by
> the refactoring of the DisplayPort driver in v6.8-rc1.
This fixes the hard resets I've been seeing since rc1 when initialising
the display subsystem of the Lenovo ThinkPad X13s at boot. With some
instrumentation added I can see the resets coinciding with the call to
rpmhpd_aggregate_corner() for 'mx_ao':
Tested-by: Johan Hovold <[email protected]>
On Tue, 27 Feb 2024 at 02:45, Bjorn Andersson via B4 Relay
<[email protected]> wrote:
>
> From: Bjorn Andersson <[email protected]>
>
> Commit 'e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable
> the domain")' aimed to make sure that a power-domain that is being
> enabled without any particular performance-state requested will at least
> turn the rail on, to avoid filling DeviceTree with otherwise unnecessary
> required-opps properties.
>
> But in the event that aggregation happens on a disabled power-domain, with
> an enabled peer without performance-state, both the local and peer
> corner are 0. The peer's enabled_corner is not considered, with the
> result that the underlying (shared) resource is disabled.
>
> One case where this can be observed is when the display stack keeps mmcx
> enabled (but without a particular performance-state vote) in order to
> access registers and sync_state happens in the rpmhpd driver. As mmcx_ao
> is flushed the state of the peer (mmcx) is not considered and mmcx_ao
> ends up turning off "mmcx.lvl" underneath mmcx. This has been observed
> several times, but has been painted over in DeviceTree by adding an
> explicit vote for the lowest non-disabled performance-state.
>
> Fixes: e3e56c050ab6 ("soc: qcom: rpmhpd: Make power_on actually enable the domain")
> Reported-by: Johan Hovold <[email protected]>
> Closes: https://lore.kernel.org/linux-arm-msm/[email protected]/
> Cc: <[email protected]>
> Signed-off-by: Bjorn Andersson <[email protected]>
Applied for fixes, thanks!
Kind regards
Uffe
> ---
> This issue is the root cause of a display regression on SC8280XP boards,
> resulting in the system often resetting during boot. It was exposed by
> the refactoring of the DisplayPort driver in v6.8-rc1.
> ---
> drivers/pmdomain/qcom/rpmhpd.c | 7 +++++--
> 1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/pmdomain/qcom/rpmhpd.c b/drivers/pmdomain/qcom/rpmhpd.c
> index 3078896b1300..47df910645f6 100644
> --- a/drivers/pmdomain/qcom/rpmhpd.c
> +++ b/drivers/pmdomain/qcom/rpmhpd.c
> @@ -692,6 +692,7 @@ static int rpmhpd_aggregate_corner(struct rpmhpd *pd, unsigned int corner)
> unsigned int active_corner, sleep_corner;
> unsigned int this_active_corner = 0, this_sleep_corner = 0;
> unsigned int peer_active_corner = 0, peer_sleep_corner = 0;
> + unsigned int peer_enabled_corner;
>
> if (pd->state_synced) {
> to_active_sleep(pd, corner, &this_active_corner, &this_sleep_corner);
> @@ -701,9 +702,11 @@ static int rpmhpd_aggregate_corner(struct rpmhpd *pd, unsigned int corner)
> this_sleep_corner = pd->level_count - 1;
> }
>
> - if (peer && peer->enabled)
> - to_active_sleep(peer, peer->corner, &peer_active_corner,
> + if (peer && peer->enabled) {
> + peer_enabled_corner = max(peer->corner, peer->enable_corner);
> + to_active_sleep(peer, peer_enabled_corner, &peer_active_corner,
> &peer_sleep_corner);
> + }
>
> active_corner = max(this_active_corner, peer_active_corner);
>
>
> ---
> base-commit: b401b621758e46812da61fa58a67c3fd8d91de0d
> change-id: 20240226-rpmhpd-enable-corner-fix-c5e07fe7b986
>
> Best regards,
> --
> Bjorn Andersson <[email protected]>
>