We need to synchronize across rings when doing a bo move to make
sure we the buffer is idle if it's in use by a different ring than
the ring doing the move.
v2: fix fence setup for bo moves
v3: add missing ring lock/unlock
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Use semaphores to sync buffers across rings in the CS
ioctl. Add a reloc flag to allow userspace to skip
sync for buffers.
agd5f: port to latest CS ioctl changes.
v2: add ring lock/unlock to make sure changes hit the ring.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
drm_getclient, drm_getstats and drm_getmap (with a few minor
adjustments) do not need global mutex, so fix that and
make the said ioctls DRM_UNLOCKED. Details:
drm_getclient: the only thing that should be protected here
is dev->filelist and that is already protected everywhere with
dev->struct_mutex.
drm_getstats: there is no need for any mutex here because the
loop runs through quasi-static (set at load time only)
data, and the actual count access is done with atomic_read()
drm_getmap already uses dev->struct_mutex to protect
dev->maplist, which also used to protect the same structure
everywhere else except at three places:
* drm_getsarea, which doesn't grab *any* mutex before
touching dev->maplist (so no drm_global_mutex doesn't help
here either; different issue for a different patch).
However, drivers seem to call it only at
initialization time so it probably doesn't matter
* drm_master_destroy, which is called from drm_master_put,
which in turn is protected with dev->struct_mutex
everywhere else in drm module, so we are good here too.
* drm_getsareactx, which releases the dev->struct_mutex
too early, but this patch includes the fix for that.
v2: * incorporate comments received from Daniel Vetter
* include the (long) explanation above to make it clear what
we are doing (and why), also at Daniel Vetter's request
* tighten up mutex grab/release locations to only
encompass real critical sections, rather than some
random code around them
Signed-off-by: Ilija Hadzic <ihadzic@research.bell-labs.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
drm_getcap and drm_version ioctls only reads static data,
there is no need to protect them with drm_global_mutex,
so make them DRM_UNLOCKED
Signed-off-by: Ilija Hadzic <ihadzic@research.bell-labs.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
We often end up missing fences on older asics with
writeback enabled which leads to delays in the userspace
accel code, so just disable it by default on those asics.
Reported-by: Helge Deller <deller@gmx.de>
Reported-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
In cases where the scanout hw is sufficiently similar between "overlay"
and traditional crtc layers, it might be convenient to allow the driver
to create internal drm_plane helper objects used by the drm_crtc
implementation, rather than duplicate code between the plane and crtc.
A private plane is not exposed to userspace.
Signed-off-by: Rob Clark <rob@ti.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Since plane->fb and plane->crtc are set in drm_mode_setplane()
after update_plane(), They should be cleared after disable().
Signed-off-by: Rob Clark <rob@ti.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
If a PCH pipe PLL is being used by transcoder C, don't disable it.
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>
In the pre-gem days with non-existing hangcheck and gpu reset code,
this timeout of 3 seconds was pretty important to avoid stuck
processes.
But now we have the hangcheck code in gem that goes to great length
to ensure that the gpu is really dead before declaring it wedged.
So there's no need for this timeout anymore. Actually it's even harmful
because we can bail out too early (e.g. with xscreensaver slip)
when running giant batchbuffers. And our code isn't robust enough
to properly unroll any state-changes, we pretty much rely on the gpu
reset code cleaning up the mess (like cache tracking, fencing state,
active list/request tracking, ...).
With this change intel_begin_ring can only fail when the gpu is
wedged, and it will return -EAGAIN (like wait_request in case the
gpu reset is still outstanding).
v2: Chris Wilson noted that on resume timers aren't running and hence
we won't ever get kicked out of this loop by the hangcheck code. Use
an insanely large timeout instead for the HAS_GEM case to prevent
resume bugs from totally hanging the machine.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Keith Packard <keithp@keithp.com>
If our semaphore logic gets confused and we have a ring stuck waiting
for one, there's a decent chance it'll just execute garbage when being
kicked. Also, kicking the ring obscures the place where the error
first occured, making error_state decoding much harder.
So drop this an let gpu reset handle this mess in a clean fashion.
In contrast, kicking rings stuck on MI_WAIT is rather harmless, at
worst there'll be a bit of screen-flickering. There's also old
broken userspace out there which needs this as a work-around.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Chris Wilson <chris@hchris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Keith Packard <keithp@keithp.com>
These registers are automatically incremented by the hardware during
transform feedback to track where the next streamed vertex output
should go. Unlike the previous generation, which had a packet for
setting the corresponding registers to a defined value, gen7 only has
MI_LOAD_REGISTER_IMM to do so. That's a secure packet (since it loads
an arbitrary register), so we need to do it from the kernel, and it
needs to be settable atomically with the batchbuffer execution so that
two clients doing transform feedback don't stomp on each others'
state.
Instead of building a more complicated interface involcing setting the
registers to a specific value, just set them to 0 when asked and
userland can tweak its pointers accordingly.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Keith Packard <keithp@keithp.com>
The waits we do here are generally so short that sleeping is a bad
idea unless we have an IRQ to wake us up. Improves regression test
performance from 18 minutes to 3.5 minutes on gen7, which is now
consistent with the previous generation.
Signed-off-by: Eric Anholt <eric@anholt.net>
Tested-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Keith Packard <keithp@keithp.com>
Previous to this commit, testing easily reproduced a failure where the
seqno would apparently arrive after the IRQ associated with it, with test programs as simple as:
for (;;) {
glCopyPixels(0, 0, 1, 1);
glFinish();
}
Various workarounds we've seen for previous generations didn't work to
fix this issue, so until new information comes in, replace the IRQ
waits on the BLT ring with polling.
Signed-off-by: Eric Anholt <eric@anholt.net>
Tested-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Keith Packard <keithp@keithp.com>
They don't fix our problems alone, but we're told to set them.
Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Keith Packard <keithp@keithp.com>
Add new ioctls for getting and setting the current destination color
key. This allows for simple overlay display control by matching a color
key value in the primary plane before blending the overlay on top.
v2: remove unnecessary mutex acquire/release around reg accesses
v3: add support for full color key management
v4: fix copy & paste bug in snb_get_colorkey
don't bother checking min/max values against docs as the docs are likely
wrong (how could we handle 10bpc surface formats?)
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
To save power when the sprite is full screen, we can disable the primary
plane on the same pipe. Track the sprite status and enable/disable the
primary opportunistically.
v2: remove primary plane enable/disable hooks; they're identical
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>
The video sprites support various video surface formats natively and can
handle scaling as well. So add support for them using the new DRM core
sprite support functions.
v2: use drm specific fourcc header and defines
v3: address Daniel's comments:
- don't take struct mutex around register access (only needed for
regs in the GT power well)
- don't hold struct mutex across vblank waits
- fix up update_plane API (pass obj instead of GTT offset)
- add interlaced defines for sprite regs
- drop unnecessary 'reg' variables
- comment double buffered reg flushing
Also fix w/h confusion when writing the scaling reg.
v4: more fixes, address more comments from Daniel, and include Hai's fix
- prevent divide by zero in scaling calculation (Hai Lan)
- update to Ville's new DRM_FORMAT_* types
- fix sprite watermark handling (calc based on CRTC size, separate
from normal display wm)
- remove private refcounts now that the fb cleanups handles things
v5: add linear surface support
v6: remove color key clearing & setting from update_plane
For this version, I tested DPMS since it came up in the last review;
DPMS off/on works ok when a video player is working under X, but for
power saving we'll probably want to do something smarter. I'll leave
that for a separate patch on top. Likewise with the refcounting/fb
layer handling, which are really separate cleanups.
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>
Konrad Rzeszutek Wilk
Alex Deucher
Christian König
Jerome Glisse
Ilija Hadzic
Chen Jie
Rob Clark
Jesse Barnes
Daniel Vetter
Eric Anholt