Add DX includes and move all device includes to a separate directory.

Co-authored with Thomas Hellstrom, Charmaine Lee and above all,
the VMware device team.

Signed-off-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Charmaine Lee <charmainel@vmware.com>

This way drm_ioctl_permit() can be used by drivers

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

This patch fixes two issues.  One, when a surface is a proxy for a DMA
buffer, it holds an extra reference that needs to be cleared.

Two, when fbdev is enabled, we need to unpin the framebuffer before
unloading the driver.  This is done by a call to vmw_fb_off().

v2
Moved unreferencing surface to from vmw_framebuffer_surface_destroy()
to vmw_kms_new_framebuffer()

Added "struct vmw_framebuffer *vfb = NULL;" to silence a compiler
warning.

Removed error checking after calling vmw_surface/dmabuf_reference()

Signed-off-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>

On older hardware, texture max width and height is not available, so set
it to something reasonable, like 8192.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

For a Screen Target capable display device, the display topology is
limited by SVGA_REG_MAX_PRIMARY_BOUNDING_BOX_MEM.  Two values are
checked against this limit:
  1.  Size of the bounding box enclosing all the displays, and
  2.  Size of the total number of displays, e.g. framebuffers

The limitations above mean we do not have exact max width and
height for the topology.  The best current option is to set those to
the maximum texture width/height.

Signed-off-by: Sinclair Yeh <syeh@vmware.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>

Reported by Intel's kbuild robot.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>

When the size of dma_addr_t was 32 bits, the compiler warned
about the size of the 32 bit shift being larger than the size
of the data type.

Reported by Intel's kbuild robot.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>

We're giving up all attempts to keep cpu- and device byte ordering separate.

This silences sparse when compiled using
make C=2 CF="-D__CHECK_ENDIAN__"

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>

The preferred mode typically didn't end up first, since the function
drm_mode_connector_list_update() reordered the modes.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

It somehow got lost in a rewrite.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

With screen targets the old legacy display system fbdev doesn't work
satisfactory anymore. At best the resolution is severely restricted.
Therefore implement fbdev on top of the kms system. With this change, fbdev
will be using whatever KMS backend is chosen.

There are helpers available for this, so in the future we'd probably want
to implement the helper callbacks instead of calling into our KMS
implementation directly.

v2: Make sure we take the mode_config mutex around modesetting,
Also clear the initial framebuffer using vzalloc instead of vmalloc.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

The kernel interface is needed for fbdev, and needs to be free from
a file_priv member. To accomplish this, remove the fb surface mutex
and list which isn't used anymore, anyway.

Finally, make the pin() and unpin() pin the framebuffer for all display
system backends, so that fbdev can pin its framebuffer before mapping it.

v2: Address review comments:
- Fix vmw_framebuffer_unpin() to handle also the surface framebuffer case.
- Fix vmw_kms_new_framebuffer() to actually use the only_2d parameter.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

If the command buffer pool is out of space, the code waits until space is
available. However since the condition code tries to allocate a range manager
node while !TASK_RUNNING we get a kernel warning.

Avoid this by pre-allocating the mm node. This will also probably be more
efficient.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

Also implements the missing readback function and
fixes page flip in case of no event.

v2:
- Adapt to the work done for screen targets for 2d, in particular
Handle proxy surface updates.
- Remove execbuf quirks since we now use fifo reserve / commit.
- Revert the initial placement of vmw dma buffers.

v3: Address review comments.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

This makes it possible to use the same function for surface dirty and
present. Also fixes page flip without events.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

We need to make the dirty- and readback functions callable without a struct
drm_file pointer. We also need to unify the handling of dirty- and readback
cliprects that are now implemented in various places across the kms system,
som add helpers to facilitate this.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

v2: Fix dma buffer validation on resource pinning.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

This patch address the following underlying issues with SurfaceDMA

* SurfaceDMA command does not work in a 2D VM, but we can wrap a
  proxy surface around the same DMA buffer and use the SurfaceCopy
  command which does work in a 2D VM.

* Wrapping a DMA buffer with a proxy surface also gives us an
  added optimization path for the case when the DMA buf
  dimensions match the mode.  In this case, the DMA buf can
  be pinned as the display surface, saving an extra copy.
  This only works in a 2D VM because we won't be doing any
  rendering operations directly to the display surface.

v2
* Moved is_dmabuf_proxy field to vmw_framebuffer_surface
* Undone coding style changes
* Addressed other issues from review

Signed-off-by: Sinclair Yeh <syeh@vmware.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>

Add support for the screen target device interface.
Add a getparam parameter and bump minor to signal availability.

Signed-off-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>

For certain surface copies, we don't have a user space handle for
the destination surface.  In such cases, we are going to trust that
our caller is giving us the right surface ID.

To do this case, we created a quirk flag that may be useful
in the future for handling other cases.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Sinclair Yeh <syeh@vmware.com>

Signed-off-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>

Refactored vmw_gb_surface_define_ioctl() and made the surface
definition part a separate function.  This way other parts of vmwgfx
can use it to allocate kernel-visible GB surfaces.

Signed-off-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>

Update device definition headers to support screen targets.

Signed-off-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>

For screen targets it appears we need to pin surfaces while they are bound
as screen targets, so add a small interface to do that.

v2: Always increase pin_count on pin.
v3: Add missing reservation sem.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

Fix a circular locking dependency between
struct vmw_overlay::mutex and
struct vmw_private::reservation_sem

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>

Add command buffer support.
Currently we don't implement preemption or fancy error handling.
Tested with a couple of mesa-demos, compiz/unity and viewperf maya-03.

v2:
- Synchronize with pending work at command buffer manager takedown.
- Add an interface to flush the current command buffer for latency-critical
  command batches and apply it to framebuffer dirtying.

v3:
- Minor fixes of definitions and typos to address reviews.
- Removed new or moved branch predictor hints.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

Don't fence and free the BO if command submission fails.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>

This commit reworks device initialization so that we always enable the
FIFO at driver load, deferring SVGA enable until either first modeset
or fbdev enable.
This should always leave the fifo properly enabled for render- and
control nodes.
In addition,
*) We disable the use of VRAM when SVGA is not enabled.
*) We simplify PM support so that we only throw out resources on hibernate,
not on suspend, since the device keeps its state on suspend.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

A regression introduced when the master ttm lock was split into two.

Reported-and-tested-by: Brian Paul <brianp@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>

Two nice things here:
- drm_dev_register will truly register everything in the right order
  if the driver doesn't have a ->load callback. Before this we had to
  init the primary mode_group after the device nodes where already
  registered.

- Less things to keep track of when reworking the connector locking,
  yay!

Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

It's been dead code since forever since mode groups haven't ever been
implemented. On top of that it's also been non-functional since we
only ever filtered the getresources ioctl and not any of the others
nor the mode object lookup code.

Given overwhelming evidence it looks like this isn't a feature we
need, hence remove it.

Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

Remaining manual work in the drm core&helpers. Nothing special here,
no surprises.

Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

This function takes two locks, both of them the wrong ones. This
wasn't an oversight from my fb locking rework since both patches
landed in parallel. We really only need fb_lock when walking that
list, since everything we can reach from that is refcounted properly
already.

v2: Drop unused dev spotted by 0day.

Cc: Rob Clark <robdclark@gmail.com>
Cc: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

Now that we also grab the connection_mutex and so fixed the race with
atomic modeset we can use the iterator there too.

The other special case is drm_connector_unplug_all which would have a
locking inversion with the sysfs store/show functions if we'd grab the
mode_config.mutex around the unplug. We could just grab
connection_mutex instead, but that's a bit too much a dirty trick for
my taste. Also it's only used by udl, which doesn't do any other kind
of connector hotplugging, so should be race-free. Hence just stick
with a comment for now.

Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

Similar with the i915 take all modeset locks for mst hotplug. This is
needed to make sure radeon holds both mode_config.mutex and
mode_config.connection_mutex when updating the connector_list, which
is the new (interim) locking regime we want for that.

Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

While auditing various users of the connector/encoder lists I realized
that the atomic code is a very prolific user of them. And it only ever
grabs the mode_config->connection_mutex, but not the
mode_config->mutex like all the other code walking encoder/connector
lists.

The problem is that we can't grab the mode_config.mutex late in atomic
code since that would lead to locking inversions. And we don't want to
grab it unconditionally like the legacy set_config modeset path since
that would render all the fine-grained locking moot.

Instead just grab more locks in the dp mst hotplug code. Note that
drm_connector_init (which is the one adding the connector to these
lists) already uses drm_modeset_lock_all.

The other reason for grabbing all locks is that the dpms off in the
unplug function amounts to a modeset, so better to take all required
locks for that.

Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

Just so I have a user for this macro.

v2: Use the right macro - somehow I thought gcc should scream at me,
but list_for_each isn't really typesafe unfortunately. Spotted by
Ville.

Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

This is now truly only duct-tape to keep locking checks happy since
calling this function when hpd or polling are already enabled is a
bug. The fbdev helper can't cope with hotplug changes yet at this
point, only after that.

Otoh a bit more robustness in this function can't hurt, and with this
fbdev can actually cope with hotplug changes. And it's also more
consistent with the connector hotadd/remove dp mst needs to do.
Therefore document this as new official behavior.

Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

So on first looks this seems superflous since drivers should ensure
correct ordering to not make this a problem. Otoh ordering constraints
between hdp, fbdev load and enabling polling are already tricky on
some hardware and it helps to be more robust.

But the real goal is to just shut up a locking WARN_ON I'd like to
add, which means init code gets some additional locks just for
uniformity.

v2: Also grab the lock for the public poll_enable, not just poll_init
which is used for resume, with the same justification.

Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>