Skip to contentSkip to conversation

Optimization on intel HDMI detect and get_modes

PATCH NAK Ramalingam C 5 participants 18 messages Jan 13, 2014 – Jul 16, 2014 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover Optimization on intel HDMI detect and get_modes

On slow HDMI hotplug out, because of the physical interface design, DDC remains active for short duration even when HPD live status is indicating the disconnect state. Because of this on VLV and HSW, slow hotplug out events are not captured.

Hence this change uses the HPD pins live status to identify the HDMI connector status.

Multiple forced detect and read modes calls come from user space for different connectors, which can be handled with connector status and previous detect event's cached EDID. With this approach for hot pluggable displays, EDID retrieval is required only when there is a real hot-plug events.

This change also includes:
1. A logic to optimize those multiple calls, by re-using cached data from previous detect and get_mode calls.
2. Store HDMI EDID, and re-use it in read modes.
3. Read live status reg to suppress spurious interrupts

These two changes will optimize the access to the DDC and also the CPU cycles burned by intel_hdmi_detect and intel_hdmi_get_modes.

Ramalingam C (1):
drm/i915: HDMI detection based on HPD pin live status

Shashank Sharma (1):
drm/i915: Optimize EDID retrival on detect and get_modes

 drivers/gpu/drm/i915/intel_drv.h  |   12 +++
 drivers/gpu/drm/i915/intel_hdmi.c |  149 ++++++++++++++++++++++++++++++++-----
 2 files changed, 144 insertions(+), 17 deletions(-)
Signature · 1 line
1.7.9.5
#1 lore raw

PATCH 1/2 drm/i915: HDMI detection based on HPD pin live status

This change uses the HPD pins live status bit from South Display Engine(SDE) to identify the HDMI hotplug state.

On Soft HPD events (on automated test cases) only HPD pin will be toggled to notify the HDMI state change. But physical DDC will be alive. Similarly on slow HDMI hotplug out, because of the physical interface design, DDC remains active for short duration even when HPD live status is indicating the disconnect state. Because of this on VLV and HSW, slow hotplug out events and soft HPDs are not captured.

Hence this patch uses the HPD pins live status to identify the HDMI connector status and allows EDID retrival only when live status is up.

Change-Id I958b57fa139e52b45c8b349c861cb8eab7b67ae5
---
 drivers/gpu/drm/i915/intel_drv.h  |   10 +++++
 drivers/gpu/drm/i915/intel_hdmi.c |   87 ++++++++++++++++++++++++++++++++-----
 2 files changed, 85 insertions(+), 12 deletions(-)
drivers/gpu/drm/i915/intel_drv.h +10 -0
index 46aea6c..0f7d94b 100644
@@ -444,6 +444,16 @@ struct cxsr_latency {
 #define to_intel_framebuffer(x) container_of(x, struct intel_framebuffer, base)
 #define to_intel_plane(x) container_of(x, struct intel_plane, base)
 
+/* DisplayPort/HDMI Hotplug line status bit mask */
+#define VLV_HDMIB_HOTPLUG_LIVE_STATUS   (1 << 29)
+#define VLV_HDMIC_HOTPLUG_LIVE_STATUS   (1 << 28)
+#define VLV_HDMID_HOTPLUG_LIVE_STATUS   (1 << 27)
+
+/* DisplayPort/HDMI/DVI Hotplug line status bit mask */
+#define CORE_HDMIB_HOTPLUG_LIVE_STATUS  (1 << 21)
+#define CORE_HDMIC_HOTPLUG_LIVE_STATUS  (1 << 22)
+#define CORE_HDMID_HOTPLUG_LIVE_STATUS  (1 << 23)
+
 struct intel_hdmi {
 	u32 hdmi_reg;
 	int ddc_bus;
drivers/gpu/drm/i915/intel_hdmi.c +75 -12

105 lines not on this page · raw message

Signature · 1 line
1.7.9.5
#2 lore raw

PATCH 2/2 drm/i915: Optimize EDID retrival on detect and get_modes

In reply to Ramalingam C

From: Shashank Sharma <[email protected]>

Multiple forced detect and read modes calls come from user space for different connectors, which can be handled with connector status and previous detect event's cached EDID. With this approach for hot pluggable displays, EDID retrieval is required only when there is a real hot-plug events.

This approach optimizes access to the DDC interface and also the CPU cycles burned by intel_hdmi_detect and intel_hdmi_get_modes.

This patch contains:
1. A logic to optimize those multiple calls, by re-using cached data from previous detect and get_mode calls.
2. Store HDMI EDID, and re-use it in read modes.
3. Read live status reg to suppress spurious interrupts

Change-Id Ia46cbe346dcc18ef11381eabcb157c5bcfd51322
---
 drivers/gpu/drm/i915/intel_drv.h  |    2 ++
 drivers/gpu/drm/i915/intel_hdmi.c |   66 +++++++++++++++++++++++++++++++++----
 2 files changed, 61 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/intel_drv.h +2 -0
index 0f7d94b..4f7f81f 100644
@@ -463,6 +463,8 @@ struct intel_hdmi {
 	bool has_audio;
 	enum hdmi_force_audio force_audio;
 	bool rgb_quant_range_selectable;
+	struct edid *edid;
+	uint32_t edid_mode_count;
 	void (*write_infoframe)(struct drm_encoder *encoder,
 				enum hdmi_infoframe_type type,
 				const void *frame, ssize_t len);
drivers/gpu/drm/i915/intel_hdmi.c +59 -7

103 lines not on this page · raw message

Signature · 1 line
1.7.9.5
#3 lore raw
In reply to Ramalingam C
Quoted 26 lines · Ramalingam C

On slow HDMI hotplug out, because of the physical interface design, DDC remains active for short duration even when HPD live status is indicating the disconnect state. Because of this on VLV and HSW, slow hotplug out events are not captured.

Hence this change uses the HPD pins live status to identify the HDMI connector status.

Multiple forced detect and read modes calls come from user space for different connectors, which can be handled with connector status and previous detect event's cached EDID. With this approach for hot pluggable displays, EDID retrieval is required only when there is a real hot-plug events.

This change also includes:
1. A logic to optimize those multiple calls, by re-using cached data from previous detect and get_mode calls.
2. Store HDMI EDID, and re-use it in read modes.
3. Read live status reg to suppress spurious interrupts

These two changes will optimize the access to the DDC and also the CPU cycles burned by intel_hdmi_detect and intel_hdmi_get_modes.

Ramalingam C (1):
drm/i915: HDMI detection based on HPD pin live status

Quoted 6 lines

Shashank Sharma (1):
drm/i915: Optimize EDID retrival on detect and get_modes

 drivers/gpu/drm/i915/intel_drv.h  |   12 +++
 drivers/gpu/drm/i915/intel_hdmi.c |  149 ++++++++++++++++++++++++++++++++-----
 2 files changed, 144 insertions(+), 17 deletions(-)

Nak.

We've had code to use the hpd live status register on all platforms, but had to take it out again on all platforms due to broken hardware: The live status simply doesn't work sometimes, resulting in angry users reporting black screens. So as-is patch 1 can't go into upstream. I know that it'd be really useful if we could rely on the hpd live status but, but thus far I couldn't come up with a good idea to make it work. Git history should have all the details.

Also note that machines with noisy interrupt lines (see the interrupt storm handling code in i915_irq.c) complicate this further.

This means that patch 2 is also a no-go since we really can't rely upon the hpd stuff for hdmi detection. But we can save EDID caching if we automatically invalidate the cached edid after 1-2 seconds or the next hpd interrupt (whichever is first). 1-2 seconds should be long enough for userspace to read the EDID a few times and change the output configuration, but not long enough for users to get pissed. Note that the caching interval should be shorter than the polling interval, which we use as a fallback if the hpd lines are noise and atm is at 10 seconds.

Also this EDID caching and invalidation code should imo be a generic helper with data structures (for the invalidation work and cached edid) in struct intel_output. That way we can also wire it up for e.g. VGA or any other output that'll benefit from EDID caching.

Aside: On DP the hpd pins seem to be reliable thus far, but I'm not sure whether that's just lack of test coverage (since DP screens are less common) or because it actually works ...

Cheers, Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#4 lore raw

Hello Daniel

Thanks a lot for your time, for reviewing the changes, and giving us some pointers.
Both me and Ramalingam are designing this together, and we discussed about these changes and your suggestions. There are few things we would like to discuss about. Please correct us if some of our understanding is not proper.

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can still get the EDID).
2. Try to reduce the EDID reads over DDC channel for get connector and fill mode calls, by caching EDID, and using it until next HPD comes.

Patch 2: Reduce the EDID read over DDC channel
We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent HDMI disconnect calls.

The design philosophy here is, to maintain a state machine of HDMI connector status, and differentiate between IOCTL detect calls and HPD detect calls. If there is a detect() or get_modes() call due to any of the IOCTL, which makes sure that input variable force=1, we just use the cached EDID, to serve this calls. But if the detect call is coming from HPD work function, due to a HPD plug-out, we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 1 line

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the cached EDID after 1-2 seconds ?

Note: In this same patch, there is additional optimization, which you pointed out, where we check if the connector->status is same as live status. This can be removed independently, as you suggested.

About patch 1:
We have done some local experiments and we came to know that for VLV and HSW boards, we can rely on the live status, if we give it some time to settle (~300ms). Probably, we need to modify this patch, as you suggested, until it becomes handy to be used reliably. We are on it, and will send another patch soon.

But if somehow we are able to get some consistent results from live status, do you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing. Because I believe we will face this problem whenever we are trying to test something from automation, where the physical device is not removed, and DDC channel is up always.

Regards
Shashank

-----Original Message-----
From: Daniel Vetter [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 12:59 PM
To: C, Ramalingam
Cc: [email protected]; Sharma, Shashank
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 26 lines · Ramalingam C

On slow HDMI hotplug out, because of the physical interface design, DDC remains active for short duration even when HPD live status is indicating the disconnect state. Because of this on VLV and HSW, slow hotplug out events are not captured.

Hence this change uses the HPD pins live status to identify the HDMI connector status.

Multiple forced detect and read modes calls come from user space for different connectors, which can be handled with connector status and previous detect event's cached EDID.
With this approach for hot pluggable displays, EDID retrieval is required only when there is a real hot-plug events.

This change also includes:
1. A logic to optimize those multiple calls, by re-using cached data from previous detect and get_mode calls.
2. Store HDMI EDID, and re-use it in read modes.
3. Read live status reg to suppress spurious interrupts

These two changes will optimize the access to the DDC and also the CPU cycles burned by intel_hdmi_detect and intel_hdmi_get_modes.

Ramalingam C (1):
drm/i915: HDMI detection based on HPD pin live status

Quoted 7 lines

Shashank Sharma (1):
drm/i915: Optimize EDID retrival on detect and get_modes

 drivers/gpu/drm/i915/intel_drv.h  |   12 +++
 drivers/gpu/drm/i915/intel_hdmi.c |  149 
++++++++++++++++++++++++++++++++-----
 2 files changed, 144 insertions(+), 17 deletions(-)

Nak.

We've had code to use the hpd live status register on all platforms, but had to take it out again on all platforms due to broken hardware: The live status simply doesn't work sometimes, resulting in angry users reporting black screens. So as-is patch 1 can't go into upstream. I know that it'd be really useful if we could rely on the hpd live status but, but thus far I couldn't come up with a good idea to make it work. Git history should have all the details.

Also note that machines with noisy interrupt lines (see the interrupt storm handling code in i915_irq.c) complicate this further.

This means that patch 2 is also a no-go since we really can't rely upon the hpd stuff for hdmi detection. But we can save EDID caching if we automatically invalidate the cached edid after 1-2 seconds or the next hpd interrupt (whichever is first). 1-2 seconds should be long enough for userspace to read the EDID a few times and change the output configuration, but not long enough for users to get pissed. Note that the caching interval should be shorter than the polling interval, which we use as a fallback if the hpd lines are noise and atm is at 10 seconds.

Also this EDID caching and invalidation code should imo be a generic helper with data structures (for the invalidation work and cached edid) in struct intel_output. That way we can also wire it up for e.g. VGA or any other output that'll benefit from EDID caching.

Aside: On DP the hpd pins seem to be reliable thus far, but I'm not sure whether that's just lack of test coverage (since DP screens are less common) or because it actually works ...

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

#5 lore raw
Quoted 3 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some pointers.
Both me and Ramalingam are designing this together, and we discussed about these changes and your suggestions. There are few things we would like to discuss about. Please correct us if some of our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality just wont agree with the spec" :(

Quoted 7 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can still get the EDID).
2. Try to reduce the EDID reads over DDC channel for get connector and fill mode calls, by caching EDID, and using it until next HPD comes.

Patch 2: Reduce the EDID read over DDC channel
We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent HDMI disconnect calls.

Quoted 6 lines

The design philosophy here is, to maintain a state machine of HDMI connector status, and differentiate between IOCTL detect calls and HPD detect calls. If there is a detect() or get_modes() call due to any of the IOCTL, which makes sure that input variable force=1, we just use the cached EDID, to serve this calls. But if the detect call is coming from HPD work function, due to a HPD plug-out, we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in. From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out, where we check if the connector->status is same as live status. This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 6 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and HSW boards, we can rely on the live status, if we give it some time to settle (~300ms). Probably, we need to modify this patch, as you suggested, until it becomes handy to be used reliably. We are on it, and will send another patch soon.

But if somehow we are able to get some consistent results from live status, do you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing. Because I believe we will face this problem whenever we are trying to test something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Cheers, Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#6 lore raw

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 3 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some pointers.
Both me and Ramalingam are designing this together, and we discussed about these changes and your suggestions. There are few things we would like to discuss about. Please correct us if some of our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality just wont agree with the spec" :(

Quoted 7 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can still get the EDID).
2. Try to reduce the EDID reads over DDC channel for get connector and fill mode calls, by caching EDID, and using it until next HPD comes.

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID
at every HPD up, on HDMI detect calls, and we are freeing it on subsequent HDMI disconnect calls.

Quoted 6 lines

The design philosophy here is, to maintain a state machine of HDMI connector status, and differentiate between IOCTL detect calls and HPD detect calls. If there is a detect() or get_modes() call due to any of the IOCTL, which makes sure that input variable force=1, we just use the cached EDID, to serve this calls. But if the detect call is coming from HPD work function, due to a HPD plug-out, we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in. From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out, where we check if the connector->status is same as live status. This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 6 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and HSW boards, we can rely on the live status, if we give it some time to settle (~300ms). Probably, we need to modify this patch, as you suggested, until it becomes handy to be used reliably. We are on it, and will send another patch soon.

But if somehow we are able to get some consistent results from live status, do you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing. Because I believe we will face this problem whenever we are trying to test something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

#7 lore raw

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 126 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 1 line · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 1 line

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 1 line

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality just wont agree with the spec" :(

Quoted 2 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 1 line

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 3 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 2 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 1 line

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 1 line

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 3 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 1 line

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 2 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 1 line

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 2 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 1 line

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#8 lore raw

Hello Quanxian Wang

This patch is available and working on all MCG tree's (Main, R42B and R44B)
We were trying to opensource this patch, but due to the dependency on live_status reg, we had to change the design.

I was working on that, but couldn't finish the activity yet, Thanks for reminding me I will update soon. :)

Regards
Shashank
-----Original Message-----
From: Wang, Quanxian
Sent: Wednesday, April 09, 2014 11:49 AM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 142 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 2 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 2 lines

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 2 lines

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality just wont agree with the spec" :(

Quoted 3 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 2 lines

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 4 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 3 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 2 lines

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 2 lines

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 4 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 3 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 2 lines

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 3 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 2 lines

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#9 lore raw

Hi Daniel / Quanxian / All,

I have one question about the 'force' flag given to connector's detect() functions.
To design new EDID caching solution, I was trying to re-use this flag. As you all know, detect() gets called from few places in DRM and I915 layer, with this flag status:

drm_sysfs.c : status_show (sysfs status check inquire from USP) force = 1
drm_crtc_helper.c: drm_helper_probe_single_connector_modes (part of get_connector IOCTL) force = 1

drm_crtc_helper.c: output_poll_execute(scheduled from poll work) force = 0
drm_crtc_helper.c: drm_helper_hpd_irq_event (resume hotplug) force = 0
i915_irq.c: i915_hotplug_work_function (bottom half of hotplug IRQ) force = 0

What I am seeing is, the places where it's really required to probe the device, like in IRQ handlers or while resuming the device, the force flag is 0, whereas whenever there is a userspace interaction or query for status, the flag is 1. Please correct me if my understating is not proper, but I feel this should be the opposite way.

Please let me know your opinion about this.

Regards
Shashank
-----Original Message-----
From: Intel-gfx [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Wednesday, April 09, 2014 12:20 PM
To: Wang, Quanxian; Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hello Quanxian Wang

This patch is available and working on all MCG tree's (Main, R42B and R44B) We were trying to opensource this patch, but due to the dependency on live_status reg, we had to change the design.

I was working on that, but couldn't finish the activity yet, Thanks for reminding me I will update soon. :)

Regards
Shashank
-----Original Message-----
From: Wang, Quanxian
Sent: Wednesday, April 09, 2014 11:49 AM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 133 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 2 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 2 lines

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 2 lines

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality just wont agree with the spec" :(

Quoted 3 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 2 lines

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 4 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 3 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 2 lines

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 2 lines

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 4 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 3 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 2 lines

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 3 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 2 lines

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Quoted 9 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#10 lore raw
Quoted 14 lines · Shashank

Hi Daniel / Quanxian / All,

I have one question about the 'force' flag given to connector's detect() functions. To design new EDID caching solution, I was trying to re-use this flag. As you all know, detect() gets called from few places in DRM and I915 layer, with this flag status:

drm_sysfs.c : status_show (sysfs status check inquire from USP) force = 1
drm_crtc_helper.c: drm_helper_probe_single_connector_modes (part of get_connector IOCTL) force = 1

drm_crtc_helper.c: output_poll_execute(scheduled from poll work) force = 0
drm_crtc_helper.c: drm_helper_hpd_irq_event (resume hotplug) force = 0
i915_irq.c: i915_hotplug_work_function (bottom half of hotplug IRQ) force = 0

Quoted 8 lines

What I am seeing is, the places where it's really required to probe the device, like in IRQ handlers or while resuming the device, the force flag is 0, whereas whenever there is a userspace interaction or query for status, the flag is 1. Please correct me if my understating is not proper,
but I feel this should be the opposite way.

Please let me know your opinion about this.

Force is a bit misnomer, a better name might be non-invasive. Without force we don't do stuff like load-detect by default since at least when doing this manually some old crt screens switch on. But this is really only relevant for gen2/3 and tv-out, so not of any concern on modern platforms. Essentially force means "userspace asked for this and it's ok if the screen flickers a bit due to that".

Imo you can do the caching and use the cached version irrespective of force.
-Daniel

Quoted 181 lines

Regards
Shashank
-----Original Message-----
From: Intel-gfx [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Wednesday, April 09, 2014 12:20 PM
To: Wang, Quanxian; Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hello Quanxian Wang

This patch is available and working on all MCG tree's (Main, R42B and R44B) We were trying to opensource this patch, but due to the dependency on live_status reg, we had to change the design.

I was working on that, but couldn't finish the activity yet, Thanks for reminding me I will update soon. :)

Regards
Shashank
-----Original Message-----
From: Wang, Quanxian
Sent: Wednesday, April 09, 2014 11:49 AM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 133 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 2 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 2 lines

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 2 lines

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality just wont agree with the spec" :(

Quoted 3 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 2 lines

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 4 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 3 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 2 lines

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 2 lines

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 4 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 3 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 2 lines

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 3 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 2 lines

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Quoted 9 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#11 lore raw

Thanks for this clarification.

Regards
Shashank
-----Original Message-----
From: Daniel Vetter [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Thursday, April 10, 2014 1:39 PM
To: Sharma, Shashank
Cc: Wang, Quanxian; Daniel Vetter; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 14 lines · Shashank

Hi Daniel / Quanxian / All,

I have one question about the 'force' flag given to connector's detect() functions. To design new EDID caching solution, I was trying to re-use this flag. As you all know, detect() gets called from few places in DRM and I915 layer, with this flag status:

drm_sysfs.c : status_show (sysfs status check inquire from USP) force = 1
drm_crtc_helper.c: drm_helper_probe_single_connector_modes (part of get_connector IOCTL) force = 1

drm_crtc_helper.c: output_poll_execute(scheduled from poll work) force = 0
drm_crtc_helper.c: drm_helper_hpd_irq_event (resume hotplug) force = 0
i915_irq.c: i915_hotplug_work_function (bottom half of hotplug IRQ) force = 0

Quoted 7 lines

What I am seeing is, the places where it's really required to probe the device, like in IRQ handlers or while resuming the device, the force flag is 0, whereas whenever there is a userspace interaction or query for status, the flag is 1. Please correct me if my understating is not proper, but I feel this should be the opposite way.

Please let me know your opinion about this.

Force is a bit misnomer, a better name might be non-invasive. Without force we don't do stuff like load-detect by default since at least when doing this manually some old crt screens switch on. But this is really only relevant for gen2/3 and tv-out, so not of any concern on modern platforms. Essentially force means "userspace asked for this and it's ok if the screen flickers a bit due to that".

Imo you can do the caching and use the cached version irrespective of force. -Daniel

Quoted 181 lines

Regards
Shashank
-----Original Message-----
From: Intel-gfx [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Wednesday, April 09, 2014 12:20 PM
To: Wang, Quanxian; Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hello Quanxian Wang

This patch is available and working on all MCG tree's (Main, R42B and R44B) We were trying to opensource this patch, but due to the dependency on live_status reg, we had to change the design.

I was working on that, but couldn't finish the activity yet, Thanks for reminding me I will update soon. :)

Regards
Shashank
-----Original Message-----
From: Wang, Quanxian
Sent: Wednesday, April 09, 2014 11:49 AM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 133 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 2 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 2 lines

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 2 lines

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality
just wont agree with the spec" :(

Quoted 3 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 2 lines

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 4 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 3 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 2 lines

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 2 lines

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 4 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 3 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 2 lines

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 3 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 2 lines

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great,
but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This
way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change
happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Quoted 9 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Quoted 4 lines

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

#12 lore raw
In reply to Sharma, Shashank

Tizen-IVI has one feature will require hdmi edid cache. It will be happy to find your patch pushed into upstream tree.

Regards

Thanks

Quanxian Wang

-----Original Message-----
From: Intel-gfx [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Wednesday, April 09, 2014 12:20 PM
To: Wang, Quanxian; Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hello Quanxian Wang

This patch is available and working on all MCG tree's (Main, R42B and R44B) We were trying to opensource this patch, but due to the dependency on live_status reg, we had to change the design.

I was working on that, but couldn't finish the activity yet, Thanks for reminding me I will update soon. :)

Regards
Shashank
-----Original Message-----
From: Wang, Quanxian
Sent: Wednesday, April 09, 2014 11:49 AM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 133 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 2 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 2 lines

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 2 lines

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality just wont agree with the spec" :(

Quoted 3 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 2 lines

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 4 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 3 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 2 lines

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 2 lines

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 4 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 3 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 2 lines

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 3 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 2 lines

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Quoted 9 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

A message that is not in the archive (sent off-list or not yet received): FF3DDC77922A8A4BB08A3BC48A1EA8CB01692A7B@BGSMSX101.gar.corp.intel.com

#14 lore raw

Please don't drop the mailing list when discussing patches.

And nope, conditioning this on gen6+ won't help at all, since I have a gen6 and gen7 machine here which don't have reliable hdmi live status. Afaik that is _really_ broken across the board and going with the fancy invalidation scheme and delayed work items is the way forward.

Cheers, Daniel

Quoted 305 lines · Shashank

Hi Daniel,

First of all, sorry for the delay in coming up with the new patch set. I got stuck with some commercial projects :) .

There were few review comments what you gave for the previous optimization patch, those were:
1. This might break the old HWs. Few of them might not even be HPD capable, so its required to handle them.
2. Do not rely on the live_status check, as that HW is not yet proven.
And you recommended to have a WQ, which will invalidate the cached HDMI EDID after a timeout.

I have written another patch, which is addressing both of the above comments, but doesn't use a WQ.

Before sending this to ML, I wanted to have an opinion from you, if this qualifies the design you were expecting. I will be really glad, if you can spare some time, and just have a top level view on the patch, and give your opinion.

If you think that this is still not the way, and making a WQ is a better approach, I would gladly create a patch which applies that.
Or if you will prefer to have this reviewed in the mailing list only, please accept my apologies for direct contact, and just let me know so. I will publish this in ML.

The patch is (It's also attached):

From e7ec4e4827615c0ff3a4ddb020df35fdf1b0b82f Mon Sep 17 00:00:00 2001
From: Shashank Sharma <[email protected]>
Date: Fri, 11 Apr 2014 18:02:50 +0530
Subject: [PATCH] drm/i915: Cache HDMI EDID for future reference

This patch contains following changes:
1.Cache HDMI EDID and optimize detect() calls, which are
frequently called from userspace, causing un-necessary EDID reads.
2.This cached EDID will be used for detection of the status of
HDMI connector, for Non HPD detect() calls. HPD calls will update cached EDID.
3.This optimization is kept for new HW's (Gen6 and +), so that It
will not break old HWs who may not be even HPD capable.

---
 drivers/gpu/drm/i915/intel_drv.h  |    1 +
 drivers/gpu/drm/i915/intel_hdmi.c |   28 ++++++++++++++++++++++++++--
 2 files changed, 27 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_drv.h +1 -0
index 42762b7..b7911df 100644
@@ -478,6 +478,7 @@ struct intel_hdmi {
                                const void *frame, ssize_t len);
        void (*set_infoframes)(struct drm_encoder *encoder,
                               struct drm_display_mode *adjusted_mode);
+       struct edid *edid;
 };
 
 #define DP_MAX_DOWNSTREAM_PORTS                0x10
drivers/gpu/drm/i915/intel_hdmi.c +26 -3
index b0413e1..33550ca 100644
@@ -938,7 +938,7 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
                hdmi_to_dig_port(intel_hdmi);
        struct intel_encoder *intel_encoder = &intel_dig_port->base;
        struct drm_i915_private *dev_priv = dev->dev_private;
-       struct edid *edid;
+       struct edid *edid = NULL;
        enum intel_display_power_domain power_domain;
        enum drm_connector_status status = connector_status_disconnected;
 
@@ -948,6 +948,21 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
        power_domain = intel_display_port_power_domain(intel_encoder);
        intel_display_power_get(dev_priv, power_domain);
 
+       /*
+       * Support EDID caching only for new architectures
+       * Let old architectures probe and force read EDID
+       */
+       if (INTEL_INFO(dev)->gen >= 6) {
+               if (force && intel_hdmi->edid &&
+                       (connector->status != connector_status_unknown)) {
+                       /* Optimize userspace query, read EDID only
+                       in case of a real hot plug */
+                       DRM_DEBUG_KMS("HDMI %s", intel_hdmi->edid ?
+                               "Connected" : "Disconnected");
+                       return connector->status;
+               }
+       }
+
        intel_hdmi->has_hdmi_sink = false;
        intel_hdmi->has_audio = false;
        intel_hdmi->rgb_quant_range_selectable = false;
@@ -964,10 +979,19 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
                        intel_hdmi->has_audio = drm_detect_monitor_audio(edid);
                        intel_hdmi->rgb_quant_range_selectable =
                                drm_rgb_quant_range_selectable(edid);
+                       DRM_DEBUG_KMS("HDMI Connected\n");
                }
-               kfree(edid);
+       } else {
+               /*
+               * Connector disconnected, free cached EDID
+               * kfree is NULL protected, so will work for < gen 6 also */
+               kfree(intel_hdmi->edid);
+               DRM_DEBUG_KMS("HDMI Disconnected\n");
        }
 
+       /* Save latest EDID for further queries */
+       intel_hdmi->edid = edid;
+
        if (status == connector_status_connected) {
                if (intel_hdmi->force_audio != HDMI_AUDIO_AUTO)
                        intel_hdmi->has_audio =
--

1.7.10.4

-----Original Message-----
From: Wang, Quanxian
Sent: Thursday, April 10, 2014 4:12 PM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Tizen-IVI has one feature will require hdmi edid cache. It will be happy to find your patch pushed into upstream tree.

Regards

Thanks

Quanxian Wang

-----Original Message-----
From: Intel-gfx [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Wednesday, April 09, 2014 12:20 PM
To: Wang, Quanxian; Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hello Quanxian Wang

This patch is available and working on all MCG tree's (Main, R42B and R44B) We were trying to opensource this patch, but due to the dependency on live_status reg, we had to change the design.

I was working on that, but couldn't finish the activity yet, Thanks for reminding me I will update soon. :)

Regards
Shashank
-----Original Message-----
From: Wang, Quanxian
Sent: Wednesday, April 09, 2014 11:49 AM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 133 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 2 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 2 lines

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 2 lines

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality just wont agree with the spec" :(

Quoted 3 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 2 lines

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 4 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 3 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 2 lines

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 2 lines

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 4 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 3 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 2 lines

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 3 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 2 lines

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Quoted 9 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#15 lore raw

Thanks for the comments,
Actually, we are not using live_status at all.

The check for < gen6 is only for EDID caching. So if the HW is >= gen6 cache_edid. Else do not cache EDID, so that we will not block any of the old HW, which might not be HPD capable.

Does it sound ok now :) ?

Regards
Shashank
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Friday, April 11, 2014 6:28 PM
To: Sharma, Shashank
Cc: C, Ramalingam; Wang, Quanxian; intel-gfx
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Please don't drop the mailing list when discussing patches.

And nope, conditioning this on gen6+ won't help at all, since I have a
gen6 and gen7 machine here which don't have reliable hdmi live status.
Afaik that is _really_ broken across the board and going with the fancy invalidation scheme and delayed work items is the way forward.

Cheers, Daniel

Quoted 310 lines · Shashank

Hi Daniel,

First of all, sorry for the delay in coming up with the new patch set. I got stuck with some commercial projects :) .

There were few review comments what you gave for the previous optimization patch, those were:
1. This might break the old HWs. Few of them might not even be HPD capable, so its required to handle them.
2. Do not rely on the live_status check, as that HW is not yet proven.
And you recommended to have a WQ, which will invalidate the cached HDMI EDID after a timeout.

I have written another patch, which is addressing both of the above comments, but doesn't use a WQ.

Before sending this to ML, I wanted to have an opinion from you, if this qualifies the design you were expecting. I will be really glad, if you can spare some time, and just have a top level view on the patch, and give your opinion.

If you think that this is still not the way, and making a WQ is a better approach, I would gladly create a patch which applies that.
Or if you will prefer to have this reviewed in the mailing list only, please accept my apologies for direct contact, and just let me know so. I will publish this in ML.

The patch is (It's also attached):

From e7ec4e4827615c0ff3a4ddb020df35fdf1b0b82f Mon Sep 17 00:00:00 2001
From: Shashank Sharma <[email protected]>
Date: Fri, 11 Apr 2014 18:02:50 +0530
Subject: [PATCH] drm/i915: Cache HDMI EDID for future reference

This patch contains following changes:
1.Cache HDMI EDID and optimize detect() calls, which are
frequently called from userspace, causing un-necessary EDID reads.
2.This cached EDID will be used for detection of the status of
HDMI connector, for Non HPD detect() calls. HPD calls will update cached EDID.
3.This optimization is kept for new HW's (Gen6 and +), so that It
will not break old HWs who may not be even HPD capable.

---
 drivers/gpu/drm/i915/intel_drv.h  |    1 +
 drivers/gpu/drm/i915/intel_hdmi.c |   28 ++++++++++++++++++++++++++--
 2 files changed, 27 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_drv.h +1 -0
index 42762b7..b7911df 100644
@@ -478,6 +478,7 @@ struct intel_hdmi {
                                const void *frame, ssize_t len);
        void (*set_infoframes)(struct drm_encoder *encoder,
                               struct drm_display_mode 
 *adjusted_mode);
+       struct edid *edid;
 };
 

#define DP_MAX_DOWNSTREAM_PORTS 0x10

drivers/gpu/drm/i915/intel_hdmi.c +27 -3
index b0413e1..33550ca 100644
@@ -938,7 +938,7 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
                hdmi_to_dig_port(intel_hdmi);
        struct intel_encoder *intel_encoder = &intel_dig_port->base;
        struct drm_i915_private *dev_priv = dev->dev_private;
-       struct edid *edid;
+       struct edid *edid = NULL;
        enum intel_display_power_domain power_domain;
        enum drm_connector_status status = 
 connector_status_disconnected;
@@ -948,6 +948,21 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
        power_domain = intel_display_port_power_domain(intel_encoder);
        intel_display_power_get(dev_priv, power_domain);
 
+       /*
+       * Support EDID caching only for new architectures
+       * Let old architectures probe and force read EDID
+       */
+       if (INTEL_INFO(dev)->gen >= 6) {
+               if (force && intel_hdmi->edid &&
+                       (connector->status != connector_status_unknown)) {
+                       /* Optimize userspace query, read EDID only
+                       in case of a real hot plug */
+                       DRM_DEBUG_KMS("HDMI %s", intel_hdmi->edid ?
+                               "Connected" : "Disconnected");
+                       return connector->status;
+               }
+       }
+
        intel_hdmi->has_hdmi_sink = false;
        intel_hdmi->has_audio = false;
        intel_hdmi->rgb_quant_range_selectable = false; @@ -964,10 
+979,19 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
                        intel_hdmi->has_audio = drm_detect_monitor_audio(edid);
                        intel_hdmi->rgb_quant_range_selectable =
                                drm_rgb_quant_range_selectable(edid);
+                       DRM_DEBUG_KMS("HDMI Connected\n");
                }
-               kfree(edid);
+       } else {
+               /*
+               * Connector disconnected, free cached EDID
+               * kfree is NULL protected, so will work for < gen 6 also */
+               kfree(intel_hdmi->edid);
+               DRM_DEBUG_KMS("HDMI Disconnected\n");
        }
 
+       /* Save latest EDID for further queries */
+       intel_hdmi->edid = edid;
+
        if (status == connector_status_connected) {
                if (intel_hdmi->force_audio != HDMI_AUDIO_AUTO)
                        intel_hdmi->has_audio =
--

1.7.10.4

-----Original Message-----
From: Wang, Quanxian
Sent: Thursday, April 10, 2014 4:12 PM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Tizen-IVI has one feature will require hdmi edid cache. It will be happy to find your patch pushed into upstream tree.

Regards

Thanks

Quanxian Wang

-----Original Message-----
From: Intel-gfx [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Wednesday, April 09, 2014 12:20 PM
To: Wang, Quanxian; Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hello Quanxian Wang

This patch is available and working on all MCG tree's (Main, R42B and R44B) We were trying to opensource this patch, but due to the dependency on live_status reg, we had to change the design.

I was working on that, but couldn't finish the activity yet, Thanks for reminding me I will update soon. :)

Regards
Shashank
-----Original Message-----
From: Wang, Quanxian
Sent: Wednesday, April 09, 2014 11:49 AM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 133 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 2 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 2 lines

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 2 lines

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality
just wont agree with the spec" :(

Quoted 3 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 2 lines

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 4 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 3 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 2 lines

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 2 lines

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 4 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 3 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 2 lines

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 3 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 2 lines

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Quoted 9 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Quoted 4 lines

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

#16 lore raw
Quoted 5 lines · Shashank

Thanks for the comments,
Actually, we are not using live_status at all.

The check for < gen6 is only for EDID caching. So if the HW is >= gen6 cache_edid. Else do not cache EDID, so that we will not block any of the old HW, which might not be HPD capable.

Oh, I've thought that this is incremental on top of something you already have.

Quoted 2 lines

Does it sound ok now :) ?

No. HPD is _NOT_ I repeat _NOT_ reliably. Not even on gen6+. live status simply reflects the hpd pin, if either doesn't work, then neither does the other one.

Nacked-with-prejudice-by Daniel Vetter <[email protected]>

Cheers, Daniel

Quoted 340 lines

Regards
Shashank
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Friday, April 11, 2014 6:28 PM
To: Sharma, Shashank
Cc: C, Ramalingam; Wang, Quanxian; intel-gfx
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Please don't drop the mailing list when discussing patches.

And nope, conditioning this on gen6+ won't help at all, since I have a
gen6 and gen7 machine here which don't have reliable hdmi live status.
Afaik that is _really_ broken across the board and going with the fancy invalidation scheme and delayed work items is the way forward.

Cheers, Daniel

Quoted 310 lines · Shashank

Hi Daniel,

First of all, sorry for the delay in coming up with the new patch set. I got stuck with some commercial projects :) .

There were few review comments what you gave for the previous optimization patch, those were:
1. This might break the old HWs. Few of them might not even be HPD capable, so its required to handle them.
2. Do not rely on the live_status check, as that HW is not yet proven.
And you recommended to have a WQ, which will invalidate the cached HDMI EDID after a timeout.

I have written another patch, which is addressing both of the above comments, but doesn't use a WQ.

Before sending this to ML, I wanted to have an opinion from you, if this qualifies the design you were expecting. I will be really glad, if you can spare some time, and just have a top level view on the patch, and give your opinion.

If you think that this is still not the way, and making a WQ is a better approach, I would gladly create a patch which applies that.
Or if you will prefer to have this reviewed in the mailing list only, please accept my apologies for direct contact, and just let me know so. I will publish this in ML.

The patch is (It's also attached):

From e7ec4e4827615c0ff3a4ddb020df35fdf1b0b82f Mon Sep 17 00:00:00 2001
From: Shashank Sharma <[email protected]>
Date: Fri, 11 Apr 2014 18:02:50 +0530
Subject: [PATCH] drm/i915: Cache HDMI EDID for future reference

This patch contains following changes:
1.Cache HDMI EDID and optimize detect() calls, which are
frequently called from userspace, causing un-necessary EDID reads.
2.This cached EDID will be used for detection of the status of
HDMI connector, for Non HPD detect() calls. HPD calls will update cached EDID.
3.This optimization is kept for new HW's (Gen6 and +), so that It
will not break old HWs who may not be even HPD capable.

---
 drivers/gpu/drm/i915/intel_drv.h  |    1 +
 drivers/gpu/drm/i915/intel_hdmi.c |   28 ++++++++++++++++++++++++++--
 2 files changed, 27 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_drv.h +1 -0
index 42762b7..b7911df 100644
@@ -478,6 +478,7 @@ struct intel_hdmi {
                                const void *frame, ssize_t len);
        void (*set_infoframes)(struct drm_encoder *encoder,
                               struct drm_display_mode 
 *adjusted_mode);
+       struct edid *edid;
 };
 

#define DP_MAX_DOWNSTREAM_PORTS 0x10

drivers/gpu/drm/i915/intel_hdmi.c +27 -3
index b0413e1..33550ca 100644
@@ -938,7 +938,7 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
                hdmi_to_dig_port(intel_hdmi);
        struct intel_encoder *intel_encoder = &intel_dig_port->base;
        struct drm_i915_private *dev_priv = dev->dev_private;
-       struct edid *edid;
+       struct edid *edid = NULL;
        enum intel_display_power_domain power_domain;
        enum drm_connector_status status = 
 connector_status_disconnected;
@@ -948,6 +948,21 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
        power_domain = intel_display_port_power_domain(intel_encoder);
        intel_display_power_get(dev_priv, power_domain);
 
+       /*
+       * Support EDID caching only for new architectures
+       * Let old architectures probe and force read EDID
+       */
+       if (INTEL_INFO(dev)->gen >= 6) {
+               if (force && intel_hdmi->edid &&
+                       (connector->status != connector_status_unknown)) {
+                       /* Optimize userspace query, read EDID only
+                       in case of a real hot plug */
+                       DRM_DEBUG_KMS("HDMI %s", intel_hdmi->edid ?
+                               "Connected" : "Disconnected");
+                       return connector->status;
+               }
+       }
+
        intel_hdmi->has_hdmi_sink = false;
        intel_hdmi->has_audio = false;
        intel_hdmi->rgb_quant_range_selectable = false; @@ -964,10 
+979,19 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
                        intel_hdmi->has_audio = drm_detect_monitor_audio(edid);
                        intel_hdmi->rgb_quant_range_selectable =
                                drm_rgb_quant_range_selectable(edid);
+                       DRM_DEBUG_KMS("HDMI Connected\n");
                }
-               kfree(edid);
+       } else {
+               /*
+               * Connector disconnected, free cached EDID
+               * kfree is NULL protected, so will work for < gen 6 also */
+               kfree(intel_hdmi->edid);
+               DRM_DEBUG_KMS("HDMI Disconnected\n");
        }
 
+       /* Save latest EDID for further queries */
+       intel_hdmi->edid = edid;
+
        if (status == connector_status_connected) {
                if (intel_hdmi->force_audio != HDMI_AUDIO_AUTO)
                        intel_hdmi->has_audio =
--

1.7.10.4

-----Original Message-----
From: Wang, Quanxian
Sent: Thursday, April 10, 2014 4:12 PM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Tizen-IVI has one feature will require hdmi edid cache. It will be happy to find your patch pushed into upstream tree.

Regards

Thanks

Quanxian Wang

-----Original Message-----
From: Intel-gfx [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Wednesday, April 09, 2014 12:20 PM
To: Wang, Quanxian; Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hello Quanxian Wang

This patch is available and working on all MCG tree's (Main, R42B and R44B) We were trying to opensource this patch, but due to the dependency on live_status reg, we had to change the design.

I was working on that, but couldn't finish the activity yet, Thanks for reminding me I will update soon. :)

Regards
Shashank
-----Original Message-----
From: Wang, Quanxian
Sent: Wednesday, April 09, 2014 11:49 AM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 133 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 2 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 2 lines

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 2 lines

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality
just wont agree with the spec" :(

Quoted 3 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 2 lines

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 4 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 3 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 2 lines

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 2 lines

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 4 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 3 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 2 lines

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 3 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 2 lines

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great, but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change happens to be detected a uevent will be generate, irrespective of the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Quoted 9 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Quoted 4 lines

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#17 lore raw

Ok, I will change the implementation.

Regards
Shashank
-----Original Message-----
From: Daniel Vetter [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Friday, April 11, 2014 7:53 PM
To: Sharma, Shashank
Cc: Daniel Vetter; C, Ramalingam; Wang, Quanxian; intel-gfx
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 5 lines · Shashank

Thanks for the comments,
Actually, we are not using live_status at all.

The check for < gen6 is only for EDID caching. So if the HW is >= gen6 cache_edid. Else do not cache EDID, so that we will not block any of the old HW, which might not be HPD capable.

Oh, I've thought that this is incremental on top of something you already have.

Quoted 2 lines

Does it sound ok now :) ?

No. HPD is _NOT_ I repeat _NOT_ reliably. Not even on gen6+. live status simply reflects the hpd pin, if either doesn't work, then neither does the other one.

Nacked-with-prejudice-by Daniel Vetter <[email protected]>

Cheers, Daniel

Quoted 341 lines

Regards
Shashank
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Friday, April 11, 2014 6:28 PM
To: Sharma, Shashank
Cc: C, Ramalingam; Wang, Quanxian; intel-gfx
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Please don't drop the mailing list when discussing patches.

And nope, conditioning this on gen6+ won't help at all, since I have a
gen6 and gen7 machine here which don't have reliable hdmi live status.
Afaik that is _really_ broken across the board and going with the fancy invalidation scheme and delayed work items is the way forward.

Cheers, Daniel

Quoted 313 lines · Shashank

Hi Daniel,

First of all, sorry for the delay in coming up with the new patch set. I got stuck with some commercial projects :) .

There were few review comments what you gave for the previous optimization patch, those were:
1. This might break the old HWs. Few of them might not even be HPD capable, so its required to handle them.
2. Do not rely on the live_status check, as that HW is not yet proven.
And you recommended to have a WQ, which will invalidate the cached HDMI EDID after a timeout.

I have written another patch, which is addressing both of the above comments, but doesn't use a WQ.

Before sending this to ML, I wanted to have an opinion from you, if this qualifies the design you were expecting. I will be really glad, if you can spare some time, and just have a top level view on the patch, and give your opinion.

If you think that this is still not the way, and making a WQ is a better approach, I would gladly create a patch which applies that.
Or if you will prefer to have this reviewed in the mailing list only, please accept my apologies for direct contact, and just let me know so. I will publish this in ML.

The patch is (It's also attached):

From e7ec4e4827615c0ff3a4ddb020df35fdf1b0b82f Mon Sep 17 00:00:00 2001
From: Shashank Sharma <[email protected]>
Date: Fri, 11 Apr 2014 18:02:50 +0530
Subject: [PATCH] drm/i915: Cache HDMI EDID for future reference

This patch contains following changes:
1.Cache HDMI EDID and optimize detect() calls, which are
frequently called from userspace, causing un-necessary EDID reads.
2.This cached EDID will be used for detection of the status of
HDMI connector, for Non HPD detect() calls. HPD calls will update cached EDID.
3.This optimization is kept for new HW's (Gen6 and +), so that It
will not break old HWs who may not be even HPD capable.

---
 drivers/gpu/drm/i915/intel_drv.h  |    1 +
 drivers/gpu/drm/i915/intel_hdmi.c |   28 ++++++++++++++++++++++++++--
 2 files changed, 27 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_drv.h +1 -0
index 42762b7..b7911df 100644
@@ -478,6 +478,7 @@ struct intel_hdmi {
                                const void *frame, ssize_t len);
        void (*set_infoframes)(struct drm_encoder *encoder,
                               struct drm_display_mode 
 *adjusted_mode);
+       struct edid *edid;
 };
 

#define DP_MAX_DOWNSTREAM_PORTS 0x10

drivers/gpu/drm/i915/intel_hdmi.c +18 -1
index b0413e1..33550ca 100644
@@ -938,7 +938,7 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
                hdmi_to_dig_port(intel_hdmi);
        struct intel_encoder *intel_encoder = &intel_dig_port->base;
        struct drm_i915_private *dev_priv = dev->dev_private;
-       struct edid *edid;
+       struct edid *edid = NULL;
        enum intel_display_power_domain power_domain;
        enum drm_connector_status status = 
 connector_status_disconnected;
@@ -948,6 +948,21 @@ intel_hdmi_detect(struct drm_connector *connector, bool force)
        power_domain = intel_display_port_power_domain(intel_encoder);
        intel_display_power_get(dev_priv, power_domain);
 
+       /*
+       * Support EDID caching only for new architectures
+       * Let old architectures probe and force read EDID
+       */
+       if (INTEL_INFO(dev)->gen >= 6) {
+               if (force && intel_hdmi->edid &&
+                       (connector->status != connector_status_unknown)) {
+                       /* Optimize userspace query, read EDID only
+                       in case of a real hot plug */
+                       DRM_DEBUG_KMS("HDMI %s", intel_hdmi->edid ?
+                               "Connected" : "Disconnected");
+                       return connector->status;
+               }
+       }
+
        intel_hdmi->has_hdmi_sink = false;
        intel_hdmi->has_audio = false;
        intel_hdmi->rgb_quant_range_selectable = false; @@ -964,10
+979,19 @@ intel_hdmi_detect(struct drm_connector *connector, bool 
+force)
                        intel_hdmi->has_audio = drm_detect_monitor_audio(edid);
                        intel_hdmi->rgb_quant_range_selectable =

drm_rgb_quant_range_selectable(edid);

drivers/gpu/drm/i915/intel_hdmi.c +10 -2
+                       DRM_DEBUG_KMS("HDMI Connected\n");
                }
-               kfree(edid);
+       } else {
+               /*
+               * Connector disconnected, free cached EDID
+               * kfree is NULL protected, so will work for < gen 6 also */
+               kfree(intel_hdmi->edid);
+               DRM_DEBUG_KMS("HDMI Disconnected\n");
        }
 
+       /* Save latest EDID for further queries */
+       intel_hdmi->edid = edid;
+
        if (status == connector_status_connected) {
                if (intel_hdmi->force_audio != HDMI_AUDIO_AUTO)
                        intel_hdmi->has_audio =
--

1.7.10.4

-----Original Message-----
From: Wang, Quanxian
Sent: Thursday, April 10, 2014 4:12 PM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Tizen-IVI has one feature will require hdmi edid cache. It will be happy to find your patch pushed into upstream tree.

Regards

Thanks

Quanxian Wang

-----Original Message-----
From: Intel-gfx [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Wednesday, April 09, 2014 12:20 PM
To: Wang, Quanxian; Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hello Quanxian Wang

This patch is available and working on all MCG tree's (Main, R42B and R44B) We were trying to opensource this patch, but due to the dependency on live_status reg, we had to change the design.

I was working on that, but couldn't finish the activity yet, Thanks for reminding me I will update soon. :)

Regards
Shashank
-----Original Message-----
From: Wang, Quanxian
Sent: Wednesday, April 09, 2014 11:49 AM
To: Sharma, Shashank; Daniel Vetter
Cc: [email protected]
Subject: RE: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Hi, Sharma, Shashank

Is there any following patches to make it happen?

Thanks

Regards

Quanxian Wang

Quoted 133 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Sharma, Shashank
Sent: Tuesday, January 14, 2014 1:20 AM
To: Daniel Vetter
Cc: [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Thanks again for this explanation Daniel.
We will work on your suggestions and come up with a new patch.

Regards
Shashank / Ramalingam
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Monday, January 13, 2014 6:57 PM
To: Sharma, Shashank
Cc: C, Ramalingam; [email protected]
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 2 lines · Shashank

Thanks a lot for your time, for reviewing the changes, and giving us some

pointers.

Quoted 2 lines

Both me and Ramalingam are designing this together, and we discussed about

these changes and your suggestions.

Quoted 2 lines

There are few things we would like to discuss about. Please correct us if some of

our understanding is not proper.

First something I've forgotten in the original mail: Overall your patches look really nice and the commit messages and cover letter have been excellent. Unfortunately you've run into one of the nastier cases of "reality
just wont agree with the spec" :(

Quoted 3 lines

Those two patches provide two solution.
1. Support for soft HPD, and slow removal of HDMI (when the DDC channel can

still get the EDID).

Quoted 2 lines

2. Try to reduce the EDID reads over DDC channel for get connector and fill mode

calls, by caching EDID, and using it until next HPD comes.

Quoted 4 lines

Patch 2: Reduce the EDID read over DDC channel We are caching the EDID at every HPD up, on HDMI detect calls, and we are freeing it on subsequent

HDMI disconnect calls.

Quoted 3 lines

The design philosophy here is, to maintain a state machine of HDMI connector

status, and differentiate between IOCTL detect calls and HPD detect calls.

Quoted 2 lines

If there is a detect() or get_modes() call due to any of the IOCTL, which makes

sure that input variable force=1, we just use the cached EDID, to serve this calls.

Quoted 2 lines

But if the detect call is coming from HPD work function, due to a HPD plug-out,

we remove/invalidate the old cached EDID, and cache the new EDID, on subsequent HDMI plug-in.

Quoted 4 lines

From here, the same state machine follows.

Can you please let us know, why do you think that we should invalidate the

cached EDID after 1-2 seconds ?

Because there are machines out there where hpd never happens. So if you keep onto the cached value forever userspace will never notice a change in output configuration. Of course hotplug handling won't work, but at least users can still manually probe outputs. By unconditionally using the cached edid from ioctls you break this use case. Yes, such machines are broken, but we need to keep them working anyway.

Also in my experience all machines are affected, we have examples
covering gm45, ilk, snb & ivb. We haven't seen a case for hsw/byt
yet since we don't rely on the hpd bits any more (and so won't see bug reports any more).

Generally if you use the hpd stuff your code must be designed under the assumption that hpd is completely unreliably. We've seen
anything from random noise, flat-out not-working at all, stuck bits and unstable hpd values that occasionally flip-flop. So you can't rely on it at all.

Quoted 2 lines

Note: In this same patch, there is additional optimization, which you pointed out,

where we check if the connector->status is same as live status.

Quoted 1 line

This can be removed independently, as you suggested.

Hm, where have I pointed this out? Some other mail on internal discussions?

Quoted 3 lines

About patch 1:
We have done some local experiments and we came to know that for VLV and

HSW boards, we can rely on the live status, if we give it some time to settle (~300ms).

Quoted 2 lines

Probably, we need to modify this patch, as you suggested, until it becomes

handy to be used reliably. We are on it, and will send another patch soon.

Quoted 3 lines

But if somehow we are able to get some consistent results from live status, do

you think it would be worth accepting this change, so that it can handle soft HPDs and automation testing.

Quoted 2 lines

Because I believe we will face this problem whenever we are trying to test

something from automation, where the physical device is not removed, and DDC channel is up always.

It's very well possible that all the platforms you have, but experience says that some OEM will horrible screw this up. At least they've consistently botched this in the past on occasional machines.

Now the ghost hdmi detection on slow removal is obviously not great,
but we can't use the hpd bits to fix this. One approach would be.
1. Upon hpd interrupt do an immediate probe of the connector. This
way we'll have good userspace experience if the unplug happens quickly and the hw works.
2. Re-probe with a 1s delay to catch slow-uplugs. The current output probing helpers are clever enough already that if a state-change
happens to be detected a uevent will be generate, irrespective of
the source of the detect call (i.e. hpd, kernel poll or ioctl/sysfs).

Note that we already track the hpd interrupts on a per-source basis, so doing the re-poll shouldn't be costly. Maybe do the re-poll as part of the EDID invalidation to avoid stalling userspace.

But you can't rely upon the hpd pins unfortunately :(

This way we should be able to implement the 2 features you want, even on unreliable hw.

Quoted 9 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Quoted 4 lines

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Quoted 4 lines

--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

#18 lore raw
Quoted 18 lines · Shashank

Ok, I will change the implementation.

Regards
Shashank
-----Original Message-----
From: Daniel Vetter [mailto:[email protected]] On Behalf Of Daniel Vetter
Sent: Friday, April 11, 2014 7:53 PM
To: Sharma, Shashank
Cc: Daniel Vetter; C, Ramalingam; Wang, Quanxian; intel-gfx
Subject: Re: [Intel-gfx] [PATCH 0/2] Optimization on intel HDMI detect and get_modes

Quoted 5 lines · Shashank

Thanks for the comments,
Actually, we are not using live_status at all.

The check for < gen6 is only for EDID caching. So if the HW is >= gen6 cache_edid. Else do not cache EDID, so that we will not block any of the old HW, which might not be HPD capable.

Quoted 7 lines

Oh, I've thought that this is incremental on top of something you already have.

Quoted 2 lines

Does it sound ok now :) ?

No. HPD is _NOT_ I repeat _NOT_ reliably. Not even on gen6+. live status simply reflects the hpd pin, if either doesn't work, then neither does the other one.

Nacked-with-prejudice-by Daniel Vetter <[email protected]>

Again re-starting the thread on this. Discussed this with Daniel on IRC and want to put the design agreed here for any further comments if any -

On detect call if no edid cached, then read full edid and cache. Assumption is that usermode will do the processing within 1s and we can just maintain this cache for 1s and clear. This will work also for panels which do not reliably assert HPD/Live status and also we get benefit of caching.

This will still fail compliance where we should not read any thing on ddc if live status is not detected. So plan is to add a compliance mode by way of module param which will be by default off. This code path will rely on HPD and Live status.

Regards
Shobhit