gfx: drv: add pipe specific macros to access the dsp/palette regs

The registers are pipe specific, so add helpers to properly access them.
The actual register accesses will be fixed up by an upcoming patch.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: remove old panel timing register macros

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: use pipe specific macros to access panel timing regs

Use the new pipe specific macros to access these registers.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: add pipe specific macros to access panel timing regs

The registers are pipe specific, so add helpers to properly access them.
The actual register references will be fixed up by an upcoming patch.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: move macro to check pipe validity to psb_intel_reg.h

This will be needed by the next patch converting other pipe specific
register definitions in psb_intel_reg.h to use this macro.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: remove unused mdfld_wait_for_PIPEA_DISABLE

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: remove unused state save/restore functions

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Fix race between SGX and page flip

There is possible race condition where SGX might have parallel leads
happening while page flip is processing. There is small race condition
open where software managed sync object completion may block SGX from
completing rendering.

The deadlock situation would cause complete system deadlock which isn't
simple to recover from.

Issue: ANDROID-1661
Signed-off-by: Pauli Nieminen <pauli.nieminen@linux.intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: pvr: Expose helper macro to compare synchronization values

To make it simpler to compare synchronization values safetily expose the
helper macro in a header.

Signed-off-by: Pauli Nieminen <pauli.nieminen@linux.intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: pvr: Add an extended sync callback API

To allow advanced wait for specific pending value that isn't the lates
one caller needs more control for sync callback state. Giving full
control for caller what is in sync struct gives complete control to
caller what is waited for.

Signed-off-by: Pauli Nieminen <pauli.nieminen@linux.intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Fix unlikely race condition

Checking operation completeness before taking the lock may result to
state where IRQ work is run between checking for completion and before
the callback is added to the list. In practice that race will require
premption because time between microkernel modifying synchronization
object and kernel work check for status is many microseconds.

That could block the last page flip in the animation until next GPU
rendering is pushed from userspace. The missing frame in end of
animation could be extremy visible in some cases.

Signed-off-by: Pauli Nieminen <pauli.nieminen@linux.intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Fix invalid pointer reference in BUFER_UNDERRUN callback for HDMI Audio.

A Null pointer was being passed in the buffer underrun event, causing
tablet to crash as the hdmi audio data pointer is expected.

The scenario is reproducible when we plug HDMI during suspended mode
when audio is playing and try to wake up.
This patch will fix the above described issue.

ISSUE: HSD-206962

Signed-off-by: Mohan Kandra <mohan.b.kandra@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

enable parsing detailed timings for EDID rev1.3

Few Detailed Descriptor Timings were missing.
This fix parses through the EDID block 1 for detailed timing
descriptor for EDID revision 1.3. Earlier code was handling
for EDID revision 1.1

ISSUE: HSD-206749

Signed-off-by: Mohan Kandra <mohan.b.kandra@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Patch to mitigate HDMI Suspend/Resume crash/noise during Audio Playback.

We were sending HAD DRM_DPMS_ON event when resuming from sleep state
in gfx_late_resume. Which was causing the HAD HOT_PLUG sequence to
run again causing:
 1. Audio Noise
 2. Random crash.

With this patch, we handle the situation by checking if audio stream
is busy in resume sequence. If yes, we don't send DRM_DPMS_ON in
gfx_late_resume to both Local MIPI and HDMI as we never send
DRM_DPMS_OFF in case HDMI audio is busy becasue of a known JIRA
issue - ANDROID-1553. Otherwise we turn them ON.
After the Jira issue is resolved, we can turn the local MIPI
sequence independent of HDMI Audio state.

Refer JIRA: ANDROID-1553 for more details.

ISSUE: HSD-206943

Signed-off-by: Mohan Kandra <mohan.b.kandra@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

MUST REVERT: Fix the video slowness issue when unplug HDMI.

This is caused by the pending flips left in psb_intel_crtc.
After HDMI unplug, vblank interrupt will never be triggered,
where actually pending flips are handled. The solution is to
handle the HDMI pending flips in vblank handler of local display.

ISSUE: ANDROID-1463

Signed-off-by: Mohan Kandra <mohan.b.kandra@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Resume HDMI audio after suspend

HDMI Audio was not getting resumed after suspend. This was
because of changes in early_suspend which was not checking if HDMI
audio is busy and calling DPMS_OFF.

Also there was no check for HDMI busy status check in
psb_runtime_idle, but in psb_runtime_suspend itself.

There is a JIRA BUG: ANDROID-1553 for local MIPI fails to turn back
on from a DPMS off/on cycle if HDMI audio returns busy to disallow
system suspend. Once ANDROID-1553 is fixed, the expectation is to
turn off MIPI but keep display island on if there is active audio
playback over HDMI.
Refer Jira bug# Android-1553 for more details.

There is still HDMI Audio noise during suspend/resume, which
is a known issue.

Signed-off-by: Mohan Kandra <mohan.b.kandra@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Correct the VHDMI values inline to OTM HDMI values

Change the VHDMI values as these values in psb are not
inline to OTM HDMI. This will change MSIC VHDMI switch
off time to 0ms, and keeps VHDMI always on. This
is needed as HDMI HPD cannot be pulled up for some HDMI displays.

Signed-off-by: Mohan Kandra <mohan.b.kandra@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: display: initialize backlight PWM frequency

Some DV1s, and DV1.5s and later use I2C to adjust the backlight. We have
PWM based backlight control enabled for compatibility with older, non-CABC
devices. The PWM clock divider was not initialized, causing the PWM signal
to run at the base clock frequency of 19.2 MHz.

Initialize the clock divider, setting the PWM frequency to 9600 Hz. The
panel specification has no recommendation on the frequency, but the bug
ANDROID-1342 has this experimental value.

Issue: ANDROID-1342
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

drv: psb: print useful values in the PSB IRQ debug macros.

Signed-off-by: Oliver McFadden <oliver.mcfadden@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: avoid NULL deref when enabling PSB debug output.

Issue: GRA-43
Signed-off-by: Oliver McFadden <oliver.mcfadden@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

drv: gfx: do not use KERN_ALERT for debug messages.

Issue: GRA-39
Signed-off-by: Oliver McFadden <oliver.mcfadden@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

drv: tc35876x: add a named module parameter for register write debug.

Otherwise the output on the serial console is far too verbose.

Issue: GRA-19
Signed-off-by: Oliver McFadden <oliver.mcfadden@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

drv: tc35876x: set the brightness only when the panel is turned on.

Attempting to set the brightness when the panel is off would result in
several failed i2c writes; cache the value and write it when bringing
the panel turned on.

Issue: GRA-28
Signed-off-by: Oliver McFadden <oliver.mcfadden@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

drv: tc35876x: pass the DRM device to all tc35876x functions.

This patch is required for further patches in this series; no functional
changes.

Signed-off-by: Oliver McFadden <oliver.mcfadden@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Lack of MTX_CMDID_NULL causes topaz fence to timeout

this timeout results in reset the hardware for every
jpeg frame encode. vaSyncSurface will not complete until
the timeout occurs, since it is still waiting for
MTX_CMID_NULL. This shows up as slow performance
(we have to wait for timeout).

Issue: ANDROID-1633

Signed-off-by: naeem.m.afzal@intel.com
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Add DVI interoperability support for HDMI.

Kernel changes to add interoperability support for DVI. We do not want to send
NULL packets for DVI mode. This change depends on Android frameworks change to
actually make the decision of NOT routing audio to DVI.

ISSUE: ANDROID-1482
ISSUE: ANDROID-1519
Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

HDMI Audio PHY should be disabled if HDMI plane is disabled.

Disable audio Phy in DPMS when unplugging HDMI cable physically
and in cases like: deep sleep, etc., which was previously missing.

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Enable HDMI audio routing and signaling to user-space and HDMI audio driver.

We use the Android switch class driver to signal an HDMI state change to frameworks.
It is also important to notify the state change to HDMI audio driver.

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Enable HDMI Audio callbacks from HDMI driver.

HDMI audio driver needs to register itself with the HDMI display
driver in order to be notified of HDMI events.

This patch will register the audio driver during boot-up and callback
HDMI audio driver functions when an event like PLUG/UNPLUG occurs.

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

During page flip, update the fb_helper->fbdev to that in the swapped fb.

This fixes the bug of system crash if unplug the HDMI cable after play video
with overlay enabled.
It is because fb_helper->fbdev points to the old value in the last fb, which
is cleaned up after video playback. And in hotplug sequence, segmentation fault
happens when walking through fb_helper->fbdev content.

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

In encoder dpms power on display island before accessing HDMI registers

In mdfld encoder dpms function, before accessing the registers notify
Power Management module not to power down the display island and once done notify
the module about that access to registers is no longer needed.

Signed-off-by: Mohan Kandra <mohan.b.kandra@intel.com>
Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Change HDMI I2C adapter from 3 to 8.

I2C adapter 3 on Medfield has known issues with reliably reading EDIDs
and especially reading 4-block EDID. We change the i2c adapter for HDMI
to adapter 8 to enable EDID extension block read (4 blocks EDID).

OTM_HDMI: r2954

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

MUST_REVERT: Set overlay clip region for HDMI to be same as MIPI.

Since the same FB is used between MIPI and HDMI, in video clone mode,
the clip region for HDMI that overlay configures should be same as MIPI.

OTM_HDMI: r3079
ISSUE: ANDROID-1462

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Mark some unused functions in code to avoid warnings.

Some currently unused functions are annotated. These will get used in the
near future and will be enabled as required.

OTM_HDMI: r3068

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Cleanup of some hardcoded values in OTM HDMI driver.

Removed some hardcode values and replace with meaningful macros.

OTM_HDMI: r3001

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Port Panel fitting changes for HDMI from Gingerbread.

Panel fit hardware configuration dependent on required output HDMI mode.

OTM_HDMI: r2979

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Fix more build warnings found in OTM HDMI driver.

Some more build warning fixes for OTM HDMI driver.

OTM_HDMI: r2971

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Correct the compile condition for command line interface code

Command line interface code needs to be within OTM_HDMI_UNIT_TEST condition.

OTM_HDMI: r2967

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Fix compilation warnings in OTM HDMI driver.

Fixes multiple compilation warnings/sparse warnings in OTM HDMI driver.

OTM_HDMI: r2964

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Fix SPARSE issues found in OTM HDMI driver.

Patch fixes multiple issues found by SPARSE in OTM HDMI driver.

OTM_HDMI: r2947

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Cleanup of hooking OTM HDMI driver into mrst/Makefile

mrst/Makefile changes corresponding to the new OTM HDMI
Makefile where the OTM HDMI Makefile is now independent
of mrst/Makefile inclusions, definitions etc.

Signed-off-by: Mohan Kandra <mohan.b.kandra@intel.com>
Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Cleanup of OTM HDMI Makefile.

Removing uncessary CFLAGS and include dirs from Makefile and making is simpler.

OTM_HDMI: r2919

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Fix SMATCH issues found in OTM HDMI driver.

Fixes for multiple SMATCH issues inside OTM HDMI driver.

OTM_HDMI: r2913

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Static Analysis fixes for OTM HDMI driver

This patch comprises of multiple issues found by static analysis tools.

OTM_HDMI: r2897

Signed-off-by: Arun Kannan <arun.kannan@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Warn if mode_config.mutex is not locked in GTT code

The GTT ref counting relies on mode_config.mutex for protection. Print a
warning if the code is called without holding the mutex.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Kill psb_bo_offset()

.bo_offset() is the only function in struct psb_intel_mode_device
which takes psb_framebuffer as a parameter, whereas the others take a
pvr mem info. Also .bo_offset() is only called from
mdfld__intel_pipe_set_base() whereas the other functions are only called
from the hardware cursor code.

Eliminate this incosistency by removing the .bo_offset() function
pointer completely and just dereference the psb_framebuffer pointer
directly in mdfld__intel_pipe_set_base().

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Reference count cursor BOs

Refrence count cursor BOs the same way as FB BOs.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Increase mem info ref count safely

gPVRSRVLock must be held when looking up the mem info based on the
handle. Also increase the ref count on the mem info before releasing
the  lock. Otherwise some other thread could destroy the mem info
after we've looked it up, but before we've increased the ref count.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Add WARN_ON(1) to psb_gtt_{map, unmap}_meminfo_ioctl

psb_gtt_map_meminfo_ioctl() and psb_gtt_unmap_meminfo_ioctl() should not
be used by anyone. Add a WARN_ON(1) to both to catch anyone using them.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Fix locking in psb_gtt_{map, unmap}_meminfo_ioctl

mode_config.mutex is needed to protect the GTT reference counting, and
gPVRSRVLock is needed to protect the pvr data structures during these
ioctls. With locking properly handled in the code we can make both
ioctls DRM_UNLOCKED.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Kill psbfb_vdc_reg()

psbfb_vdc_reg() is not used anywhere, kill it.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Free the correct pointer in psb_framebuffer_destroy()

Pass the pointer to the psb_framebuffer structure to kfree(), not
the pointer to the embedded drm_framebuffer structure.

Luckily in this case the drm_framebuffer is embedded as the first
member inside psb_framebuffer, so the incorrect code didn't
manage to make a mess of things.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Use need_gtt() in psb_fb_gtt_ref()/unref()

Make sure the mem info needs a GTT mapping before trying to take a
reference to one.

Also swap the order of psb_fb_ref() and psb_gtt_map_meminfo() in
psb_gtt_fb_ref() to be consistent with the rest of the code.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Clean up error handling during fb creation

Try to handle errors during user framebuffer creation. Also use
psb_fb_ref()/unref() to take the reference to the mem info, in order
to get a reference to the per process data as well.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Keep a reference to pvr per process data

While we have a reference to the pvr mem info, we also must keep a
reference to the per process data for the owner. Freeing the mem
info will access data structures whose lifetime is tied to the per
process data, such as the context. So if we don't keep a reference
to the per process data, and the process exits before we unref the
mem info, the kernel could oops or some assertions could fail in
the pvr code.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

drm: disconnect plane from fb/crtc when disabled

Since plane->fb and plane->crtc are set in drm_mode_setplane()
after update_plane(), They should be cleared after disable().

Signed-off-by: Rob Clark <rob@ti.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: overlay: Optimize filter coefficient load with sychronized flips

Without synchronized flips the driver can instruct the hardware to stop
re-loading the filter coefficients from memory once they've been loaded
by the hardware at leas once.

Bring the same optimization over to synchronized flips. To do that the
dirty state tracking must be split into software and hardware variants
due to the asynchronous nature of the synchronized flip code.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Move flip trace commands into psb_fb

The overlay code also manipulates sync object counters these days.
Instead of duplicating the same tracing code in two places, move
the code to the place where the sync counters are actually
manipulated.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: overlay: Synchronize overlay updates with CRTC page flips

Tie the overlay updates to the CRTC page flips. Whenever the overlay is
updated, we simply "arm" the overlay, and when the next page flip happens
on the associated CRTC, we push the page flip and the pending overlay
updates to the hardware simultaneosly.

This will guarantee that the page flip and overlay updates happen on the
same vblank, and are thus perfectly synchronized with each other.

Issue: ANDROID-1574
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

HACK: gfx: display: Double the tc35876x pixel clock

The pixel clock and vertical refresh values the driver reports for the
internal display are wrong. The page flip logic needs accurate clock
information to convert time intervals to scanline intervals.

Double the pixel clock in the tc35876x driver so that the values are
at least a little closer to reality.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Avoid page flipping while too close to vblank start

The hardware doesn't have proper facilities to indicate when a specific
page flip has occured. We can deduce that information from the frame
counter and scanline counter, but only if we know on which side of the
vblank start the flip was programmed. So if we're too close to the
vblank start to be able make that determination, wait until the vblank
start has passed before programming the flip into the hardware.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Rewrite the CRTC page flipping code to use the new drm_flip mechanism

The drm_flip mechanism allows us to track page flip progess, and thus
maintain proper reference counts on the backing memory and GTT mappings
while the buffers are being scanned out.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Introduce drm_flip helper class

The drm_flip mechanism can be used to implement robust page flipping
support, and also to synchronize the flips on multiple hardware scanout
engines (eg. CRTCs and overlays).

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Add some helper function to manipulate fb refcounts

Upcoming page flip logic will need to maintain extra references
to the pvr mem_info, and the GTT mapping. Add some functions to
make that easier.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Move read ops sync counter functions into psb_fb.c

The overlay code will want to manipulate sync counters as well, so move
the code to a more appropriate place.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Check fb bpp before doing irreversible changes to state

When the fb uses an unsupported bpp layout, the current code messes with
the hardware state before checking whether it can actually use the
provided framebuffer. This could leave the hardware in some weird state.
So check the framebuffer bpp before touching the hardware state.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: gtt: Support unmapping from arbitrary context

Pass the 'tgid' down from the caller of psb_gtt_unmap_meminfo(). This
will allow unmapping from eg. workqueues. This will be used to keep
references to the GTT mapping possibly after the drm_framebuffer has
been destroyed.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: pvr: ossync: Indicate whether sync callback is called from MISR

The callee may want to do some extra locking when the sync callback is
called from MISR (as opposed to being called directly). Add a boolean
parameter to the sync callback which indicates this.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Add PIPE DSL regs

Add defines for display scanline counter registers.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: pvr: Lock gPVRSRVLock in PVRSRVMISR()

It's totally unclear how various data are protected in the driver.
Currently the MISR work is exectuted without holding the gPVRSRVLock
mutex, which is held across bridge ioctls. To play things safe,
take the same lock in MISR work.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Lock gPVRSRVLock mutex before touching the meminfo/bm ref counts

The pvr mem ref counts are not modified using atomic operations, so we
need to protect them against simultaneous RMW access. Additionally it's
not at all clear how the related data structures are protected in the
pvr code. Use gPVRSRVLock as that's always locked around bridge ioctls.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: pvr: Remove most of the display class code

The display class swap chain code calls into the psb_fb driver
which causes KMS mode_config mutex to be locked after gPVRSRVLock.
To avoid problems with non-atomic reference counts, a later patch
will take the gPVRSRVLock inside code sections where mode_config
mutex is already locked. That could lead to dead locks, so the simple
solution is to eliminate the DC swap chain code altogether. We don't
want to use it anyway.

Unfortunately the DC driver can't be removed completely, since the
current userspace code requires a DC device to be registered. So
leave a minimal DC driver in place.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: update meminfo reference count when creating/destroying fbs

The fbs may outlive the mem info if the client frees the memory before
removing the fb so make sure the memory backing the fb lives as long
as the fb itself.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: pvr: add an interface for inc/dec meminfo ref count

The drm part of the pvr driver stores pointers to mem infos but those
may be freed by the clients. If the reference counts aren't increased
properly, those mem infos may be freed while the kernel still
references them.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: support creating fb's from mapped memory

If a client passes a handle to a mapped memory to drmModeAddFB, the
ioctl will fail since the gtt mapping code will fail to get its page
list. Fix that by using a reference to the original meminfo instead
of the mapped one.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: pvr: add a way of getting the src meminfo backing a mapped one

This will be used to implement on the drm side support for creating
a drm fb from a mapped memory handle.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: pvr: unifdef RES_MAN_EXTEND

The PVRSRVGetResData function will be used in a follow up patches to
implement support for creating drm fb's from mapped exported memory.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: don't store meminfo handle on psbfb

The handle has no reference count and might be freed by the client
before the psbfb is destroyed.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: use psbfb->pvrBO instead of hKernelMemInfo on page flip code

The meminfo is already stored on psbfb. Remove the extra indirection
through psb_get_meminfo_by_handle.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: pass meminfo instead of handle to psb_gtt_{unmap, map}_meminfo

The handles are process specific and may be freed while the kernel is
still using the meminfo.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: gtt: refcount gtt mappings

Currently psb_gtt_map_* can handle multiple calls with the same
handle but will remove the pages from gtt as soon as one call to
psb_gtt_unmap_* is made. Since multiple drm_fb's can be added with
the same handle and the driver will cal psb_gtt_map_meminfo for each
one of them, this could lead to unmapping memory that is still being
used.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: gtt: refactor gtt mapping code

Make room for inserting reference counting in the psb_gtt_mem_mapping
objects by refactoring the map and unmap code so that lower level
objects are allocated in lower level functions.

There is a hash table per process that holds a psb_gtt_mem_mapping
object which in turn holds a drm_mm_node object. The current
implementation allocs a drm_mm_node object in the psb_gtt_map_*
functions and then uses helper functions to wrap this node inside
a mapping object. When freeing this object, the drm_mm_node has to
be passed to the psb_gtt_unmap_* functions so that it can be freed.

This patch changes this logic so that the psb_gtt_map_* functions
allocate the mapping object through a helper function
(psb_gtt_add_mapping) that in turns allocates the node object. When
freeing, the psb_gtt_unmap_* functions will call psb_gtt_remove_mapping
which in will in turn free the node object.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: fix resource leak in psb_gtt_map_meminfo

If psb_gtt_map_meminfo was called with an already mapped handle a node
would be allocated with drm_mm_get_block_atomic in psb_gtt_mm_alloc_mem.
Later, on psb_gtt_mm_alloc_insert_mem_mapping (called by
psb_gtt_add_node), the reference to the allocated node would be
replaced by the node previously allocated for the handle, so the former
was never free'd.

This commit changes psb_gtt_map_meminfo to check for an existing node
before allocating a new one.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv/psb_gtt.c: add spaces around comments

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv/psb_gtt.c: replace printk with DRM_DEBUG

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv/psb_gtt.c: improve readability by using ERR_PTR

A lot of functions in this file return a value through a parameter
and use the return value for returning an error value. Using ERR_PTR
makes the code a bit more readable.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv/psb_gtt.c: make function calls take as few lines as possible

Comestic fix. No code changes.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: change psb_gtt_insert_* to remove casting

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv/psb_gtt.c: remove unnecessary casting of psb_get_tgid() to u32

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: fix psb_gtt_{map, unmap}_pvr_memory interface

Change data types to avoid unnecessary casting.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv/psb_gtt.h: fix indentation

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: drv: Do not kfree() pvr buffer's page list

psb_get_pages_by_mem_handle() doesn't allocate a new page list. Instead
it simply returns a reference to the pvr buffer's page list. Do not
kfree() said list in the caller.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

drv: psb: check crtc pointer before calling in DPMS and mode set.

Avoid the possibility of oops-ing the kernel by running the
test-opengl-fillrate test (and others) which bypass DRM_MASTER
authentication.

Typically the test will execute succesfully, however when
unblanking/un-suspending we NULL deref in gfx_late_resume()

Signed-off-by: Oliver McFadden <oliver.mcfadden@intel.com>
Reviewed-by: Pauli Nieminen <pauli.nieminen@linux.intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: display: remove unused panel_desc field and defines

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Tested-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: display: cleanup ospm_power_suspend()

No functional changes, just a stylistic cleanup to make the function easier
to follow.

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Tested-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: display: remove useless gb{Suspend,Resume}InProgress flags

The gbSuspendInProgress and gbResumeInProgress flags are only ever true
when the g_ospm_mutex is held. The locking should be in order without
these.

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Tested-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: display: remove redundant code in ospm_power_using_hw_begin()

The display island should already be on at this point, and the power up
status for video enc/dec islands has already been checked some lines
earlier.

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Tested-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: display: drop redundant ospm_resume_pci() in ospm_power_using_hw_begin()

The runtime resume through pm_runtime_get_sync() should already have taken
care of ospm_resume_pci(). No need to repeat here.

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Tested-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: display: ospm_power_using_hw_begin() rewrite

ospm_power_using_hw_begin() had gotten overly complicated with a plethora
of variables and code paths. The code can be simplified considerably by
making the observation that the function is never called with force_on in
atomic context, i.e. if the power island is off, it will not be switched on
in atomic context.

Split the function into ospm_power_using_hw_begin() and
ospm_power_using_hw_begin_atomic(), of which only the latter may be called
in an atomic context. Clean up and simplify the code with the new
assumptions that can be made.

There is still plenty of strangeness left, but this is a big step forward
in making the code possible to understand.

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Tested-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

gfx: display: remove no-op assignments in ospm_power_using_hw_begin()

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Tested-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>