It's only likely to paper over bugs. Unlike the gpu, where we want to
keep things alive a bit longer in expectation of the next frame's
submit, when the display is shut down we can power off immediately.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Acked-by: Archit Taneja <architt@codeaurora.org>
When CONFIG_PM is disabled, we get harmless warnings about unused
functions:
drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.c:1025:12: error: 'mdp5_runtime_resume' defined but not used [-Werror=unused-function]
static int mdp5_runtime_resume(struct device *dev)
^~~~~~~~~~~~~~~~~~~
drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.c:1015:12: error: 'mdp5_runtime_suspend' defined but not used [-Werror=unused-function]
static int mdp5_runtime_suspend(struct device *dev)
^~~~~~~~~~~~~~~~~~~~
This marks both functions as __maybe_unused so the compiler
can drop them silently.
Fixes: d68fe15b18 ("drm/msm/mdp5: Use runtime PM get/put API instead of toggling clocks")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Accessing registers for an unclocked block is an insta-reboot on
snapdragon devices. So add a bit of logic to track the enable_count so
we can WARN_ON() unclocked register writes. This makes it much easier
to track down mistakes.
Signed-off-by: Rob Clark <robdclark@gmail.com>
mdp5_enable/disable calls are scattered all around in the MDP5 code.
Use the pm_runtime_get/put calls here instead, and populate the
runtime PM suspend/resume ops to manage the clocks.
About the overall design: MDP5 is a child of the top level MDSS
device. MDSS is also the parent to DSI, HDMI and other interfaces. When
we enable MDP5's power domain, we end up enabling MDSS's PD too. It is
only MDSS's PD that actually controlls the GDSC HW. Therefore, calling
runtime_get/put on the MDP5 device is like just requesting a vote to
enable/disable the GDSC.
Functionally, replacing the clock enable/disable calls with the RPM API
can result in the power domain (GDSC) state being toggled if no other
child isn't powered on. This can result in the register context being lost.
We make sure (in future commits) that code paths don't end up configuring
registers and then later lose state, resulting in a bad HW state.
For now, we've replaced each mdp5_enable/disable with runtime_get/put API.
We could optimize things later by removing runtime_get/put calls which
don't really need to be there. This could prevent unnecessary toggling of
the power domain and clocks.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
We have upstream bindings (msm8916) that have the "_clk" suffix in the
clock names. The downstream bindings also require it.
We want to drop the "_clk" suffix and at the same time support existing
bindings. Update the MDP5 code with the the msm_clk_get() helper to
support both old and new clock names.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Now that the msm_gem supports an arbitrary number of vma's, we no longer
need to assign an id (index) to each address space. So rip out the
associated code.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Before we can shift to passing the address-space object to _get_iova(),
we need to fix a few places (dsi+fbdev) that were hard-coding the adress
space id. That gets somewhat easier if we just move these to the kms
base class.
Prep work for next patch.
Signed-off-by: Rob Clark <robdclark@gmail.com>
If we restrict this helper to only kms drivers (which is the case) we
can look up the correct mode easily ourselves. But it's a bit tricky:
- All legacy drivers look at crtc->hwmode. But that is updated already
at the beginning of the modeset helper, which means when we disable
a pipe. Hence the final timestamps might be a bit off. But since
this is an existing bug I'm not going to change it, but just try to
be bug-for-bug compatible with the current code. This only applies
to radeon&amdgpu.
- i915 tries to get it perfect by updating crtc->hwmode when the pipe
is off (i.e. vblank->enabled = false).
- All other atomic drivers look at crtc->state->adjusted_mode. Those
that look at state->requested_mode simply don't adjust their mode,
so it's the same. That has two problems: Accessing crtc->state from
interrupt handling code is unsafe, and it's updated before we shut
down the pipe. For nonblocking modesets it's even worse.
For atomic drivers try to implement what i915 does. To do that we add
a new hwmode field to the vblank structure, and update it from
drm_calc_timestamping_constants(). For atomic drivers that's called
from the right spot by the helper library already, so all fine. But
for safety let's enforce that.
For legacy driver this function is only called at the end (oh the
fun), which is broken, so again let's not bother and just stay
bug-for-bug compatible.
The benefit is that we can use drm_calc_vbltimestamp_from_scanoutpos
directly to implement ->get_vblank_timestamp in every driver, deleting
a lot of code.
v2: Completely new approach, trying to mimick the i915 solution.
v3: Fixup kerneldoc.
v4: Drop the WARN_ON to check that the vblank is off, atomic helpers
currently unconditionally call this. Recomputing the same stuff should
be harmless.
v5: Fix typos and move misplaced hunks to the right patches (Neil).
v6: Undo hunk movement (kbuild).
Cc: Mario Kleiner <mario.kleiner@tuebingen.mpg.de>
Cc: Eric Anholt <eric@anholt.net>
Cc: Rob Clark <robdclark@gmail.com>
Cc: linux-arm-msm@vger.kernel.org
Cc: freedreno@lists.freedesktop.org
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Christian König <christian.koenig@amd.com>
Cc: Ben Skeggs <bskeggs@redhat.com>
Reviewed-by: Neil Armstrong <narmstrong@baylibre.com>
Acked-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170509140329.24114-4-daniel.vetter@ffwll.ch
There are reasons for a memory object to outlive the file descriptor
that created it and so the address space that a buffer object is
attached to must also outlive the file descriptor. Reference count
the address space so that it can remain viable until all the objects
have released their addresses.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
In the last few commits, we've been adding params to mdp5_crtc_state, and
assigning them in the atomic_check() funcs. Now it's time to actually
start using them.
Remove the duplicated params from the mdp5_crtc struct, and start using
them in the mdp5_crtc code. The majority of the references to these params
is in code that executes after the atomic swap has occurred, so it's okay
to use crtc->state in them. There are a couple of legacy LM cursor ops that
may not use the updated state, but (I think) it's okay to live with that.
Now that we dynamically allocate a mixer to the CRTC, we can also remove
the static assignment to it in mdp5_crtc_init, and also drop the code that
skipped init-ing WB bound mixers (those will now be rejected by
mdp5_mixer_assign()).
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add the stuff needed to allow dynamically assigning a mixer to a CRTC.
Since mixers are a resource that can be shared across multiple CRTCs, we
need to maintain a 'hwmixer_to_crtc' map in the global atomic state,
acquire the mdp5_kms.state_lock modeset lock and so on.
The mixer is assigned in the CRTC's atomic_check() func, a failure will
result in the new state being cleanly rolled back.
The mixer assignment itself is straightforward, and almost identical to
what we do for hwpipes. We don't need to grab the old hwmixer_to_crtc
state like we do in hwpipes since we don't need to compare anything
with the old state at the moment.
The only LM capability we care about at the moment is whether the mixer
instance can be used to display stuff (i.e, connect to an INTF
downstream).
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
mdp5_interface struct contains data corresponding to a INTF
instance in MDP5 hardware. This sturct is memcpy'd to the
mdp5_encoder struct, and then later to the mdp5_ctl struct.
Instead of copying around interface data, create mdp5_interface
instances in mdp5_init, like how it's done currently done for
pipes and layer mixers. Pass around the interface pointers to
mdp5_encoder and mdp5_ctl. This simplifies the code, and allows
us to decouple encoders from INTFs in the future if needed.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Use the mdp5_hw_mixer struct in the mdp5_crtc and mdp5_ctl instead of
using the LM index.
Like before, the Layer Mixers are assigned statically to the CRTCs.
The hwmixer(s) will later be dynamically assigned to CRTCs.
For now, ignore the hwmixers that can only do WB.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Create a struct to represent MDP5 Layer Mixer instances. This will
eventually allow us to detach CRTCs from the Layer Mixers, and
generally clean things up a bit.
This is very similar to how hwpipes were previously abstracted away
from drm planes.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Register cursor drm_planes. The loop in modeset_init that inits the
planes and crtcs has to be refactored a bit. We first iterate all the
hwpipes to find the cursor planes. Then, we loop again to create
crtcs.
In msm_atomic_wait_for_commit_done, remove the check which bypasses
waiting for vsyncs if state->legacy_cursor_updates is true.
We will later create a fast path for cursor position changes in the
cursor plane's update_plane func that doesn't go via the regular
atomic commit path. For rest of cursor related updates, we will have
to wait for vsyncs, so ignore the legacy_cursor_updates flag.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
These are various changes added in preparation for cursor planes:
- Add a pipe_cursor block for 8x96 in mdp5_cfg.
- Add a new pipe CAP called MDP_PIPE_CAP_CURSOR. Use this to ensure we
assign a cursor SSPP for a drm_plane with type DRM_PLANE_TYPE_CURSOR.
- Update mdp5_ctl_blend_mask/ext_blend_mask funcs to incorporate cursor
SSPPs.
- In mdp5_ctl_blend, iterate through MAX_STAGES instead of stage_cnt,
we need to do this because we can now have empty stages in between.
- In mdp5_crtc_atomic_check, make sure that the cursor plane has the
highest zorder, and stage the cursor plane to the maximum stage #
present on the HW.
- Create drm_crtc_funcs that doesn't try to implement cursors using the
older LM cursor HW.
- Pass drm_plane_type in mdp5_plane_init instead of a bool telling
whether plane is primary or not.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
We currently create CRTCs equaling to the # of Layer Mixer blocks we
have on the MDP5 HW. This number is generally more than the # of encoders
(INTFs) we have in the MDSS HW. The number of encoders connected to
displays on the platform (as described by DT) would be even lesser.
Create only N drm_crtcs, where N is the number of drm_encoders
successfully registered. To do this, we call modeset_init_intf() before
we init the drm_crtcs and drm_planes.
Because of this change, setting encoder->possible_crtcs needs to be moved
from construct_encoder() to a later point when we know how many CRTCs we
have.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
For the DSI interfaces, the mdp5_kms core creates 2 encoders for video
and command modes.
Create only a single encoder per interface. When creating the encoder, set
the interface type to MDP5_INTF_MODE_NONE. It's the bridge (DSI/HDMI/eDP)
driver's responsibility to set a different interface type. It can use the
the kms func op set_encoder_mode to change the mode of operation, which
in turn would configure the interface type for the INTF.
In mdp5_cmd_encoder.c, we remove the redundant code, and make the commmand
mode funcs as helpers that are used in mdp5_encoder.c
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Rename the mdp5_encoder_* ops for active displays to
mdp5_vid_encoder_* ops.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
The mdp5 kms driver currently sets up multiple encoders per interface
(INTF), one for each kind of mode of operation it supports.
We create 2 drm_encoders for DSI, one for Video Mode and the other
for Command Mode operation. The reason behind this approach could have
been that we aren't aware of the DSI device's mode of operation when
we create the encoders.
This makes things a bit complicated, since these encoders have to
be further attached to the same DSI bridge. The easier way out is
to create a single encoder, and make the DSI driver set its mode
of operation when we know what the DSI device's mode flags are.
Start with providing a way to set the mdp5_intf_mode using a kms
func that sets the encoder's mode of operation. When constructing
a DSI encoder, we set the mode of operation to Video Mode as
default. When the DSI device is attached to the host, we probe the
DSI mode flags and set the corresponding mode of operation.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
We currently create 2 encoders for DSI interfaces, one for command
mode and other for video mode operation. This isn't needed as we
can't really use both the encoders at the same time. It also makes
connecting bridges harder.
Switch to creating a single encoder. For now, we assume that the
encoder is configured only in video mode. Later, the same encoder
would be usable in both modes.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
It would race between userspace thread and commit worker. Ie. vblank
irq would trigger event and userspace could begin the next atomic
update, before the commit worker had a chance to clear the pending
flag.
If we do end up needing something to prevent userspace from trying
another pageflip before getting vblank event, it should probably be
implemented as a pending_planes bitmask, similar to pending_crtcs. See
start_atomic() and end_atomic().
Signed-off-by: Rob Clark <robdclark@gmail.com>
The mode_config->max_{width,height} is for the maximum size of a fb, not
the max scanout limits (of the layer-mixer). It is legal, and in fact
common, to create a larger fb, only only scan-out a smaller part of it.
For example multi-monitor configurations for x11, or android wallpaper
layer (which is created larger than the screen resolution for fast
scrolling by just changing the src x/y coordinates).
Signed-off-by: Rob Clark <robdclark@gmail.com>
Previously, SMP block allocation was not checked in the plane's
atomic_check() fxn, so we could fail allocation SMP block allocation at
atomic_update() time. Re-work the block allocation to request blocks
during atomic_check(), but not update the hw until committing the atomic
update.
Since SMP blocks allocated at atomic_check() time, we need to manage the
SMP state as part of mdp5_state (global atomic state). This actually
ends up significantly simplifying the SMP management, as the SMP module
does not need to manage the intermediate state between assigning new
blocks before setting flush bits and releasing old blocks after vblank.
(The SMP registers and SMP allocation is not double-buffered, so newly
allocated blocks need to be updated in kms->prepare_commit() released
blocks in kms->complete_commit().)
Signed-off-by: Rob Clark <robdclark@gmail.com>
(re)assign the hw pipes to planes based on required caps, and to handle
situations where we could not modify an in-use plane (ie. SMP block
reallocation).
This means all planes advertise the superset of formats and properties.
Userspace must (as always) use atomic TEST_ONLY step for atomic updates,
as not all planes may be available for use on every frame.
The mapping of hwpipe to plane is stored in mdp5_state, so that state
updates are atomically committed in the same way that plane/etc state
updates are managed. This is needed because the mdp5_plane_state keeps
a pointer to the hwpipe, and we don't want global state to become out
of sync with the plane state if an atomic update fails, we hit deadlock/
backoff scenario, etc. The use of state_lock keeps multiple parallel
updates which both re-assign hwpipes properly serialized.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add basic state duplication/apply mechanism. Following commits will
move actual global hw state into this.
The state_lock allows multiple concurrent updates to proceed as long as
they don't both try to alter global state. The ww_mutex mechanism will
trigger backoff in case of deadlock between multiple threads trying to
update state.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Archit Taneja <architt@codeaurora.org>
Split out the hardware pipe specifics from mdp5_plane. To start, the hw
pipes are statically assigned to planes, but next step is to assign the
hw pipes during plane->atomic_check() based on requested caps (scaling,
YUV, etc). And then hw pipe re-assignment if required if required SMP
blocks changes.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Archit Taneja <architt@codeaurora.org>
These are really plane-id's, not crtc-id's. Only connection to CRTCs is
that they are used as primary-planes.
Current name is just legacy from when we only supported RGB/primary
planes. Lets pick a better name now.
Signed-off-by: Rob Clark <robdclark@gmail.com>
We can have various combinations of 64b and 32b address space, ie. 64b
CPU but 32b display and gpu, or 64b CPU and GPU but 32b display. So
best to decouple the device iova's from mmap offset.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Introduce new compatible strings for the top level MDSS wrapper device,
and the MDP5 device.
Previously, the "qcom,mdp5" and "qcom,mdss_mdp" compatible strings
were used to match the top level platform_device (which was also tied
to the top level drm_device struct). Now, these strings are used
to match the MDP5 platform device.
Use "qcom,mdss" as the compatible string for top level MDSS device.
This is now used to match the top level platform_device (which is
tied to the drm_device struct).
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Since runtime PM isn't implemented yet, we need to call
mdp5_enable/disable in a few more places. These would later be
replaced by runtime PM get/put calls.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
With the new device hierarchy for MDP5, we need to enable runtime PM
for both the toplevel MDSS device and the MDP5 device itself. Enable
runtime PM for the new devices.
Since MDP4 and MDP5 now have different places where runtime PM is
enabled, remove the previous pm_runtime_enable/disable calls, and
squash them in the respective kms drivers.
The new device hierarchy (as expressed in the DT bindings) has the GDSC
tied only to the MDSS wrapper device. This GDSC needs to be enabled for
accessing any register in the MDSS sub-blocks. Once every driver is
runtime adapted, the GDSC will be enabled when any sub-block device
calls runtime_get because of the parent-child relationship with MDSS.
Until then, we call pm_runtime_get_sync() once for the MDSS device to
ensure the GDSC is never disabled. This will be removed once all the
drivers are runtime PM adapted.
The error handling paths become a bit tricky when we call these runtime
PM funcs. There doesn't seem to be any helper that checks if runtime PM
is enabled already. Add bool variables in mdp4_kms/mdp5_kms structs to
check if the driver had managed to call pm_runtime_enable before bailing
out.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Since MDSS registers were stuffed within the the MDP5 register
space, we had an __offset_MDP() macro to identify the offset
between the start of MDSS and MDP5 address spaces. This offset
macro expected a MDP index argument, which didn't make much
sense since we don't have multiple MDPs.
The offset is no longer needed now that we have devices for the 2
different register address spaces. Also, remove the "REG_MDP5_MDP_"
prefix to "REG_MDP5_".
Update the generated headers in mdp5.xml.h
We generally update headers as a separate patch, but we need to
do these together to prevent breaking build.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
With the new kms_init/destroy funcs in place for MDP5, we can get rid of
the old kms funcs. Some members of the mdp5_kms struct also become
redundant, so we remove those too.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Call msm_mdss_init in msm_drv to set up top level registers/irq line.
Start using the new kms_init2/destroy2 funcs to inititalize MDP5 KMS.
With the MDSS interrupt and irqdomain set up, the old MDP5 irq code
can be dropped.
The mdp5_hw_init kms func now uses the platform device tied to MDP5
instead of the one tied to the drm_device/MDSS.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
With MDP5 as a new device, we need to do less for MDP when initializing
modeset after all the components are bound.
Create mdp5_kms_init2/destroy2 funcs that inits modeset. These will
eventually replace the older kms_init/destroy funcs.
In the new kms_init2, the platform_device used is the one corresponding
to the new MDP5 platform_device. The new change here is that the irq is
now retrieved using irq_of_parse_and_map(), since MDP5 is a child interrupt
of the MDSS interrupt controller.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
In order to have a tree-like device hierarchy between MDSS and its
sub-blocks (MDP5, DSI, HDMI, eDP etc), we need to create a separate
device/driver for MDP5. Currently, MDP5 and MDSS are squashed
together are are tied to the top level platform_device, which is
also the one used to create drm_device.
The mdp5_kms_init code is split into two parts. The part where device
resources are allocated are associated with the MDP5 driver's probe,
the rest is executed later when we initialize modeset.
With this change, unlike MDP4, the MDP5 platform_device isn't tied to
the top level drm_device anymore. The top level drm_device is now
associated with a platform device that corresponds to MDSS wrapper
hardware.
Create mdp5_init/destroy funcs that will be used by the MDP5 driver
probe/remove. Use the HW_VERSION register in the MDP5 register address
space. Both the MDSS and MDP VERSION registers give out identical
version info.
The older mdp5_kms_init code is left as is for now, this would be removed
later when we have all the pieces to support the new device hierarchy.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
The driver gets the irq number using platform_get_irq on the main kms
platform device. This works fine since both MDP4 and MDP5 currently
have a flat device hierarchy. The platform device tied with the
drm_device points to the MDP DT node in both cases.
This won't work when MDP5 supports a tree-like hierarchy. In this
case, the platform device tied to the top level drm_device is the
MDSS DT node, and the irq we need for KMS is the one generated by
MDP5, not MDSS.
Get the irq number from the MDP4/5 kms driver itself. Each driver
can later provide the irq number based on what device hierarchy it
uses.
While we're at it, call drm_irq_install only when we have a valid KMS
driver.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
The driver expects DT to provide the parent to MDP core clock. The only
operation done to the parent clock is to set a rate. This can be
achieved by setting the rate on the core clock itsef. Don't try to
get the parent clock anymore.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Global symbols in the kernel should be prefixed by the name
of the subsystem and/or driver to avoid conflicts when all
code is built-in.
In this case, function names like 'hdmi_register' or 'hdmi_set_mode'
are way too generic for an MSM specific DRM driver, so I'm renaming
them all to msm_hdmi_* here.
I also rename a lot of the 'static' symbols along with the global
names for consistency, even though those are relatively harmless;
they might only be slightly confusing when they show up in
backtraces.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Rob Clark <robdclark@gmail.com>
MDP5 has line count and frame count registers for each interface. Enable
these counters and use them to implement the get_vblank_timestamp drm
driver op.
The line counter starts with the value 1 at the beginning of the VSYNC
pulse and ends with value VTOTAL at the end of VFP. This value is used
to determine whether we're in blanking period or not, and an adjusted
value of this counter is used to get vpos as expected by
get_scanout_position. Since there is no way to calculate hpos, we always
set it to 0.
Signed-off-by: Archit Taneja <architt@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
