Mir on vmwgfx
thellstrom at vmware.com
Tue Nov 5 07:22:28 UTC 2013
I'm new to this list and I'm trying to get Mir running in a VMware
virtual machine on top of the vmwgfx driver stack.
The idea is to first get "mir_demo_server_basic" running with demo
clients and then move on to Xmir, patching up our drivers as necessary.
So far, I've encountered a couple of issues that might need attention
from MIR developers:
1) function mggh::DRMHelper::is_appropriate_device() in
gbm_display_helpers.c checks whether a drm device has any children
except itself. This is not true for vmwgfx, and the server will fail to
start thinking that our drm device is not appropriate. Why the child
2) Once I get the basic server to start, the cursor disappears as soon
as I move the mouse. This boils down to Mir thinking that the cursor is
outside of the current mode bounding box. At Mir server startup, there
is no KMS setup configured, hence
DisplayConfigurationOutput::current_mode_index will be set to max (or
-1) in mgg::RealKMSDisplayConfiguration::add_or_update_output().
The value of DisplayConfigurationOutput::current_mode_index then doesn't
seem to change even when Mir sets a display configuration, and when the
mode bounding box is calculated, an out of bounds array access is performed.
3) Minor thing: The "Virtual" connector type is not recognized by Mir.
(actually it's not in xf86drmMode.h either, I'll see if I can fix that
up), but it's in the kernel user-space api file "drm_mode.h" and is
right after the "eDP" connector type. Should be added in
connector_type_name() in real_kms_output.cpp
4) vmwgfx does not yet implement the drm "Prime" mechanism for sharing
of dma buffers, which Mir relies on. I'm about to implement that.
However, it seems like Mir is using dma buffers in an illegal way:
1) Mir creates a GBM bufffer.
2) Mir uses Prime to export a dma_buf handle which it shares with its
3) The client imports the dma_buf handle and uses drm to turn it into a
drm buffer handle.
4) The buffer handle is typecast to a "dumb" buffer handle, and then
mmap'ed. in struct GBMMemoryRegion : mcl::MemoryRegion.
It's illegal to typecast a GBM buffer to a dumb buffer in this way. It
accidently happens to work on the major driver because deep inside, both
a GBM buffer and a dumb buffer is represented by a GEM buffer object.
With vmwgfx that's not the case either for a GBM buffer or a dumb
buffer, and they are different objects.
In fact, currently the only way to mmap() a GBM buffer (unless it's a
cursor buffer) is to export a dma_buf and use it's mmap() operation. But
that is not implemented by any of the major drivers since it's not
really desired. The reason is that a GBM buffer is completely opaque and
may not even reside in mappable memory. Hence any attempt to map it may
result in coherence issues and, in some cases, extremely costly
operations. This results in awkward driver code that attempts to guess
the usage-patterns of applications that mix mmap'ed cpu-access and
accelerated access to the same object.
The correct way to do this is to have the client import the buffer to
the appropriate API, and use texsubImage like operations (or perhaps
readPixels / writePixels) to access the buffer from the CPU.
IMHO, I think this is something that needs to be addressed as soon as
More information about the Mir-devel