BO are referenced once by reloc to make sure that they not destroyed
before we get a chance to flush the cmd stream, so we need to unreference
them once in cs submit or cs erase if cs i never submitted so bo can
be destructed.
Relocation now consist of the following informations (in this order) :
handle buffer object handle identifier
start_offset start offset of first data of the buffer object used by the cs
end_offset end offset of last data of the buffer object used by the cs
read_domain read domain (either VRAM, or GTT as GPU is invalid for CS)
write_domain write domain (either VRAM, or GTT as GPU is invalid for CS)
flags flags used for further optimization (like discard previous
buffer content or forget buffer content after cs which can
help in avoiding moving content in or out)
I wanted to avoid doing this, as it's a bunch of churn, but there was a
conflict between the dri_ symbols in libdrm and the symbols that were in
Mesa in 7.2, which broke Mesa 7.2 AIGLX when the 2D driver had loaded new
libdrm symbols. The new naming was recommended by cworth for giving the
code a unique prefix identifying where the code lives.
Additionally, take the opportunity to fix up two API mistakes: emit_reloc's
arguments were in a nonsensical order, and set_tiling lacked the stride
argument that the kernel will want to use soon. API compatibility with
released code is maintained using #defines.
This relies on a new kernel ioctl to get the available aperture size.
In order to provide reasonable performance from dri_bufmgr_check_aperture, we
now require that once a buffer has been used as the target of a relocation,
it gets no further relocations added to it. This cuts the cost of
check_aperture from 10% to 1% in the 3D driver with no code changes, but
slightly complicates our plans for the 2D driver.
Don't count on ioctl returning -errno; use errno directly.
Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
We want to be able to use the bufmgr from multiple threads for GL, and thus
we need to protect the internal structures.
The pthread-stubs package is used so that programs not linked against
pthreads get weak symbols to stubs and don't eat most of the cost.
When using bufmgr_fake without DRM, the X server idles the ring whenever it
wants to wait for something to complete (brutal, but effective). In this
case, bufmgr_fake must treat the pending fence as having passed. However, it
wasn't recording the fences as it emitted them, nor cleaning buffers as they
passed.
Signed-off-by: Keith Packard <keithp@keithp.com>
We want to be able to use the bufmgr from multiple threads for GL, and thus
we need to protect the internal structures.
The pthread-stubs package is used so that programs not linked against
pthreads get weak symbols to stubs and don't eat most of the cost.
We need a way of getting at the underlying handle for use with mode
setting. We can either export it in the dri_bo object or provide a new
callback to get it.
I'd swapped the operands, so if we weren't in lockstep with the hardware we
said the sequence was always passed. Additionally, a race was available that
we might have failed at recovering from. Instead, I've replaced the logic
with new stuff that should be more robust and not rely on all the parties in
userland following the same IRQ_EMIT() == 1 protocol. Also, in a radical
departure from past efforts, include a long comment describing the failure
modes and how we're working around them.
Thanks to haihao for catching the original issue.
Jesse Barnes
Dave Airlie
Eric Anholt
Jerome Glisse
Jakob Bornecrantz
Xiang, Haihao
Keith Packard
Dave Airlie
Julien Cristau