Now that the vblank races are plugged, we can opt out of using
the vblank disable timer and just let vblank interrupts get
disabled immediately when the last reference is dropped.

Gen2 is the exception since it has no hardware frame counter.
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Add a flag to drm_device which will cause the vblank code to bypass the
disable timer and always disable the vblank interrupt immediately when
the last reference is dropped.

v2: Add some notes about the flag to the kernel doc
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Make drm_vblank_put() disable the vblank interrupt immediately when the
refcount drops to zero and drm_vblank_offdelay<0.

v2: Preserve the current drm_vblank_offdelay==0 'never disable' behaviur
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Currently it's possible that the following will happen:
1. drm_wait_vblank() calls drm_vblank_get()
2. drm_vblank_off() gets called
3. drm_wait_vblank() calls drm_queue_vblank_event() which
   adds the event to the queue event though vblank interrupts
   are currently disabled (and may not be re-enabled ever again).

To fix the problem, add another vblank->enabled check into
drm_queue_vblank_event().

drm_vblank_off() holds event_lock around the vblank disable,
so no further locking needs to be added to drm_queue_vblank_event().
vblank disable from another source is not possible since
drm_wait_vblank() already holds a vblank reference.
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Currently both drm_irq.c and several drivers call drm_vblank_put()
while holding event_lock. Now that drm_vblank_put() can disable the
vblank interrupt directly it may need to grab vbl_lock and
vblank_time_lock. That causes deadlocks since we take the locks
in the opposite order in two places in drm_irq.c. So let's make
sure the locking order is always event_lock->vbl_lock->vblank_time_lock.

In drm_vblank_off() pull up event_lock from underneath vbl_lock. Hold
the event_lock across the whole operation to make sure we only send
out the events that were on the queue when we disabled the interrupt,
and not ones that got added just after (assuming drm_vblank_on() already
managed to get called somewhere between).

To sort the other deadlock pull the event_lock out from
drm_handle_vblank_events() into drm_handle_vblank() to be taken outside
vblank_time_lock. Add the appropriate assert_spin_locked() to
drm_handle_vblank_events().
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Declare a local struct drm_vblank_crtc * and use that
instead of having to do dig it out via 'dev->vblank[crtc]'
everywhere.

Performed with the following coccinelle incantation,
and a few manual whitespace cleanups:

@@
identifier func,member;
expression num_crtcs;
struct drm_device *dev;
unsigned int crtc;
@@
func (...) {
+ struct drm_vblank_crtc *vblank;
...
if (crtc >= num_crtcs)
   return ...;
+ vblank = &dev->vblank[crtc];
<+...
(
- dev->vblank[crtc].member
+ vblank->member
|
- &(dev->vblank[crtc])
+ vblank
)
...+>
}

@@
struct drm_device *dev;
int crtc;
identifier member;
expression num_crtcs;
@@
for (crtc = 0; crtc < num_crtcs; crtc++) {
+ struct drm_vblank_crtc *vblank = &dev->vblank[crtc];
+
<+...
(
- dev->vblank[crtc].member
+ vblank->member
|
- &(dev->vblank[crtc])
+ vblank
)
...+>
}

@@
identifier func,member;
@@
func (struct drm_device *dev, int crtc, ...) {
+ struct drm_vblank_crtc *vblank = &dev->vblank[crtc];
<+...
(
- dev->vblank[crtc].member
+ vblank->member
|
- &(dev->vblank[crtc])
+ vblank
)
...+>
}

v2: Rebased
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

When drm_vblank_on() is called the hardware vblank counter may have
been reset, so we can't trust that the old values sampled prior to
drm_vblank_off() have anything to do with the new values.

So update the .last count in drm_vblank_on() to make the first
drm_vblank_enable() consider that as the reference point. This
will correct the user space visible counter to account for the
time between drm_vblank_on() and the first drm_vblank_enable()
calls.

For extra safety subtract one from the .last count in drm_vblank_on()
to make sure that user space will never see the same counter value
before and after modeset.
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

If the vblank irq has already been disabled (via the disable timer) when
we call drm_vblank_off() sample the counter and timestamp one last time.
This will make the sure that the user space visible counter will account
for time between vblank irq disable and drm_vblank_off().
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Move drm_update_vblank_count() to avoid forward a declaration.
No functional change.
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Clearing the timestamps causes us to send zeroed timestamps to userspace
if they get sent out in response to the drm_vblank_off(). It's better
to send the very latest timestamp and count instead.

Testcase: igt/kms_flip/modeset-vs-vblank-race
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

v2: Drop the drm_vblank_off() (Daniel)
    Use drm_crtc_vblank_{get,put}()
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Make sure drm_vblank_get() never succeeds when called between
drm_vblank_off() and drm_vblank_on(). Borrow a trick from the
old drm_vblank_{pre,post}_modeset() functions and just bump
the refcount in drm_vblank_off() and drop it in drm_vblank_on().

When drm_vblank_get() encounters a >0 refcount and the vblank
interrupt is already disabled it will simply return -EINVAL.

Hopefully the use of inmodeset won't conflict badly with
drm_vblank_{pre,post}_modeset().

For i915 there's a window between drm_vblank_off() and marking the
crtc as inactive where the current code still allows drm_vblank_get().

v2: Describe what drm_vblank_get() does to explain how
    a simple refcount bump manages to fix things (Daniel)
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Everyone agrees we should do this,
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: David Herrmann <dh.herrmann@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

bunch of cleanups
* 'drm-next' of git://people.freedesktop.org/~dvdhrm/linux:
  drm: mark drm_context support as legacy
  drm: make sysfs device always available for minors
  drm: make minor->index available early
  drm: merge drm_drv.c into drm_ioctl.c
  drm: move module initialization to drm_stub.c
  drm: don't de-authenticate clients on master-close
  drm: drop redundant drm_file->is_master
  drm: extract legacy ctxbitmap flushing

We need to take the connection mutex around the link status
check for non-MST case, but also around the MST link training
on short HPDs.

I suspect we actually should have a dpcd lock in the future as
well, that just lock the local copies of dpcd and flags stored
from that.
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>

- Additional Hawaii fixes
- Support for using the display scaler on non-fixed mode displays
- Support for new firmware format that makes it easier to update
- Enable dpm by default on additional asics
- GPUVM improvements
- Support for uncached and write combined gtt buffers
- Allow allocation of BOs larger than visible vram
- Various other small fixes and improvements

* 'drm-next-3.17' of git://people.freedesktop.org/~agd5f/linux: (57 commits)
  drm/radeon: Prevent hdmi deep color if max_tmds_clock is undefined.
  drm/radeon: Use pflip irqs for pageflip completion if possible. (v2)
  drm/radeon: tweak ACCEL_WORKING2 query for the new firmware for hawaii
  drm/radeon: use packet3 for nop on hawaii with new firmware
  drm/radeon: tweak ACCEL_WORKING2 query for hawaii
  drm/radeon: use packet2 for nop on hawaii with old firmware
  drm/radeon: update IB size estimation for VM
  drm/radeon: split PT setup in more functions
  drm/radeon: add VM GART copy optimization to NI as well
  drm/radeon: take a BO reference on VM cleanup
  drm/radeon: add radeon_bo_ref function
  drm/radeon: remove taking mclk_lock from radeon_bo_unref
  drm/radeon: adjust default radeon_vm_block_size v2
  drm/radeon: try to enable VM flushing once more
  drm/radeon: use an intervall tree to manage the VMA v2
  drm/radeon: remove radeon_bo_clear_va
  drm/radeon: invalidate moved BOs in the VM (v2)
  drm/radeon: re-enable dpm by default on BTC
  drm/radeon: re-enable dpm by default on cayman
  drm/radeon: Only flush HDP cache from idle ioctl if BO is in VRAM
  ...

This renames all drm-context helpers to drm_legacy_*() and moves the
internal definitions into the new drm_legacy.h header. This header is
local to DRM-core and drivers shouldn't access it.
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>

HDMI spec requires a valid max_tmds_clock from edid for hdmi
deep color modes. If a sink violates this, disable deep color.

Also add a hint to user about the deep_color module parameter if
deep color is disabled due to that.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

For each minor we allocate a sysfs device as minor->kdev. Currently, this
is allocated and registered in drm_minor_register(). This makes it
impossible to add sysfs-attributes to the device before it is registered.
Therefore, they are not added atomically, nor can we move device_add()
*after* ->load() is called.

This patch makes minor->kdev available early, but only adds the device
during minor-registration. Note that the registration is still called
before ->load() as debugfs needs to be split, too. This will be fixed in
follow-ups.
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>

Instead of allocating the minor-index during registration, we now do this
during allocation. This way, debug-messages between minor-allocation and
minor-registration will now use the correct minor instead of 0. Same is
done for unregistration vs. free, so debug-messages between
device-shutdown and device-destruction show proper indices.

Even though minor-indices are allocated early, we don't enable minor
lookup early. Instead, we keep the entry set to NULL and replace it during
registration / unregistration. This way, the index is allocated but lookup
only works if registered.
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>

All that is left in drm_drv.c is ioctl management. Merge it into
drm_ioctl.c so we have all ioctl management in one file (and the name is
much more fitting).

Maybe we should now rename drm_stub.c to drm_drv.c again?
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>

Most of the new DRM management functions are nowadays in drm_stub.c. By
moving the core module initialization to drm_stub.c we can make several
global variables static and keep the stub-open helper local.

The core files now look like this:
  drm_stub.c: Core management
   drm_drv.c: Ioctl dispatcher
 drm_ioctl.c: Actual ioctl backends
  drm_fops.c: Char-dev file-operations

A follow-up patch will move what is left from drm_drv.c into drm_ioctl.c.
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>

If an active DRM-Master closes its device, we deauthenticate all clients
on that master. However, if an inactive DRM-Master closes its device, we
do nothing. This is quite inconsistent and breaks several scenarios:

 1) If this was used as security mechanism, it fails horribly if a master
    closes a device while VT switched away. Furthermore, none of the few
    drivers using ->master_*() callbacks seems to require it, anyway.

 2) If you spawn weston (or any other non-UMS compositor) in background
    while another compositor is active, both will get assigned to the
    same "drm_master" object. If the foreground compositor now exits, all
    clients of both the foreground AND background compositor will be
    de-authenticated leading to unexpected behavior.

Stop this non-sense and keep clients authenticated. We don't do this when
dropping DRM-Master (i.e., switching VTs) so don't do it on active-close
either!
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>

The drm_file->is_master field is redundant as it's equivalent to:
    drm_file->master && drm_file->master == drm_file->minor->master

1) "=>"
  Whenever we set drm_file->is_master, we also set:
      drm_file->minor->master = drm_file->master;

  Whenever we clear drm_file->is_master, we also call:
      drm_master_put(&drm_file->minor->master);
  which implicitly clears it to NULL.

2) "<="
  minor->master cannot be set if it is non-NULL. Therefore, it stays as
  is unless a file drops it.

  If minor->master is NULL, it is only set by places that also adjust
  drm_file->is_master.

Therefore, we can safely drop is_master and replace it by an inline helper
that matches:
    drm_file->master && drm_file->master == drm_file->minor->master
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>

The ctxbitmap code is only used by legacy drivers so lets try to keep it
as separated as possible. Furthermore, the locking is non-obvious and
kinda weird with ctxlist_mutex *and* struct_mutex. Keeping all ctxbitmap
access in one file is much easier to review and makes drm_release() more
readable.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>

Skip the "manual" pageflip completion checks via polling and
guessing in the vblank handler radeon_crtc_handle_vblank() on
asics which are known to reliably support hw pageflip completion
irqs. Those pflip irqs are a more reliable and race-free method
of handling pageflip completion detection, whereas the "classic"
polling method has some small races in combination with dpm on,
and with the reworked pageflip implementation since Linux 3.16.

On old asics without pflip irqs, the classic method is used.

On asics with known good pflip irqs, only pflip irqs are used
by default, but a new module parameter "use_pflipirqs" allows to
override this in case we encounter asics in the wild with
unreliable or faulty pflip irqs. A module parameter of 0 allows
to use the classic method only in such a case. A parameter of 1
allows to use both classic method and pflip irqs as additional
band-aid to avoid some small races which could happen with the
classic method alone. The setting 1 gives Linux 3.16 behaviour.

Hw pflip irqs are available since R600.

Tested on DCE-4, AMD Cedar - FirePro 2270.

v2:  agd5f: only enable pflip interrupts on DCE4+ as they are not
reliable on older asics.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Adjust the previous tweak for hawaii to return 3 if the new firmware is used.
Signed-off-by: Andreas Boll <andreas.boll.dev@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Older firmware didn't support the new nop packet.
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Andreas Boll <andreas.boll.dev@gmail.com>

Return 2 so we can be sure the kernel has the necessary
changes for acceleration to work.

Note: This patch depends on these two commits:
 - drm/radeon: fix cut and paste issue for hawaii.
 - drm/radeon: use packet2 for nop on hawaii with old firmware
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Andreas Boll <andreas.boll.dev@gmail.com>
Cc: stable@vger.kernel.org

Older firmware didn't support the new nop packet.

v2 (Andreas Boll):
 - Drop usage of packet3 for new firmware
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com> (v1)
Signed-off-by: Andreas Boll <andreas.boll.dev@gmail.com>
Cc: stable@vger.kernel.org

That should allow us to allocate bigger BOs.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Move the decision what to use into the common VM code.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

This closes a small window where the GPU might have accessed freed up memory.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

To be consistent with radeon_bo_unref, needed in the following patch.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

It's causing lockdep warnings and why should
we access the memory that is freed up?
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

v2: rebase on vm_size scale change. Adjust vm_size default to 8,
    Better handle the default and smaller values.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Let's try to fix bugs related to this instead of just disabling it.
Signed-off-by: Christian König <christian.koenig@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Scales much better than scanning the address range linearly.

v2: store pfn instead of address
Signed-off-by: Christian König <christian.koenig@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Won't work anyway, instead WARN_ON if the VA list isn't
empty when we free the BO.
Signed-off-by: Christian König <christian.koenig@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>