Trying to resolve issues with missed vblanks and impossible
values inside delivered kms pageflip completion events showed
that radeon's irq handling sometimes doesn't handle valid irqs,
but silently skips them. This was observed for vblank interrupts.
Although those irqs have corresponding events queued in the gpu's
irq ring at time of interrupt, and therefore the corresponding
handling code gets triggered by these events, the handling code
sometimes silently skipped processing the irq. The reason for those
skips is that the handling code double-checks for each irq event if
the corresponding irq status bits in the irq status registers
are set. Sometimes those bits are not set at time of check
for valid irqs, maybe due to some hardware race on some setups?
The problem only seems to happen on some machine + card combos
sometimes, e.g., never happened during my testing of different PC
cards of the DCE-2/3/4 generation a year ago, but happens consistently
now on two different Apple Mac cards (RV730, DCE-3, Apple iMac and
Evergreen JUNIPER, DCE-4 in a Apple MacPro). It also doesn't happen
at each interrupt but only occassionally every couple of
hundred or thousand vblank interrupts.
This results in XOrg warning messages like
"[ 7084.472] (WW) RADEON(0): radeon_dri2_flip_event_handler:
Pageflip completion event has impossible msc 420120 < target_msc 420121"
as well as skipped frames and problems for applications that
use kms pageflip events or vblank events, e.g., users of DRI2 and
DRI3/Present, Waylands Weston compositor, etc. See also
https://bugs.freedesktop.org/show_bug.cgi?id=85203
After some talking to Alex and Michel, we decided to fix this
by turning the double-check for asserted irq status bits into a
warning. Whenever a irq event is queued in the IH ring, always
execute the corresponding interrupt handler. Still check the irq
status bits, but only to log a DRM_DEBUG message on a mismatch.
This fixed the problems reliably on both previously failing
cards, RV-730 dual-head tested on both crtcs (pipes D1 and D2)
and a triple-output Juniper HD-5770 card tested on all three
available crtcs (D1/D2/D3). The r600 and evergreen irq handling
is therefore tested, but the cik an si handling is only compile
tested due to lack of hw.
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
CC: Michel Dänzer <michel.daenzer@amd.com>
CC: Alex Deucher <alexander.deucher@amd.com>
CC: <stable@vger.kernel.org> # v3.16+
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In order for hibernation to reliably work we need to cleanup more
thoroughly the compute ring. Hibernation is different from suspend
resume as when we resume from hibernation the hardware is first
fully initialize by regular kernel then freeze callback happens
(which correspond to a suspend inside the radeon kernel driver)
and turn off each of the block. It turns out we were not cleanly
shutting down the compute ring. This patch fix that.
Hibernation and suspend to ram were tested (several times) on :
Bonaire
Hawaii
Mullins
Kaveri
Kabini
Changed since v1:
- Factor the ring stop logic into a function taking ring as arg.
Cc: stable@vger.kernel.org
Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1
iQEcBAABAgAGBQJVa7zvAAoJEHm+PkMAQRiGtfMIAILs3sxFtrC1hApgcfRLF/7z
K34bwTRqErzqUO/orTwakEr9kSIpIL0zIPSryTCOTPZLfMGkQjhHXO3KR/DSbbTV
MZ8y/BM/yelFA/Np+1LjbiYjTNRnTRvCoaQihkIH8Rn02g7ob9HyL4gIGKpuGFcZ
04GacL2cgChqsRSACdNef948jCoJXKgcuDpe39DXphDWZnBKNZ3HFuJ6bryGJf9A
1/eCI4is85BNwKPemQUYR0xx83UIzDfrghatZP2mOCDDSA2MNg8HNxLTd12LGoQD
tfgX4B7aftzW9Y7GSEDfZ0IKm2NRzgPmCVj6PjVR/iI0lIK4Aq0Z/lDJxxEq3XQ=
=AJM5
-----END PGP SIGNATURE-----
Merge tag 'v4.1-rc6' into drm-next
Linux 4.1-rc6
backmerge 4.1-rc6 as some of the later pull reqs are based on newer bases
and I'd prefer to do the fixup myself.
We have that bug for years and some users report side effects when fixing it on older hardware.
So revert it for VM_CONTEXT0_PAGE_TABLE_END_ADDR, but keep it for VM 1-15.
Signed-off-by: Christian König <christian.koenig@amd.com>
CC: stable@vger.kernel.org
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The mapping range is inclusive between starting and ending addresses.
Signed-off-by: Christian König <christian.koenig@amd.com>
CC: stable@vger.kernel.org
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Registers that can be fetched from the info ioctl.
Tested-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Enable at init and disable on fini. Workaround for hardware problems.
v2 (chk): extend commit message
v3: add new function
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com> (v2)
Cc: stable@vger.kernel.org
Emit the EOP twice to avoid cache flushing problems.
Signed-off-by: Christian König <christian.koenig@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Slava Grigorev <slava.grigorev@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Slava Grigorev <slava.grigorev@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Backmerge Linus tree after rc5 + drm-fixes went in.
There were a few amdkfd conflicts I wanted to avoid,
and Ben requested this for nouveau also.
Conflicts:
drivers/gpu/drm/amd/amdkfd/Makefile
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
drivers/gpu/drm/amd/amdkfd/kfd_priv.h
drivers/gpu/drm/amd/include/kgd_kfd_interface.h
drivers/gpu/drm/i915/intel_runtime_pm.c
drivers/gpu/drm/radeon/radeon_kfd.c
We need to wait for the GPUVM flush to complete. There
was some confusion as to how this mechanism was supposed
to work. The operation is not atomic. For GPU initiated
invalidations you need to read back a VM register to
introduce enough latency for the update to complete.
v2: drop gart changes
v3: just read back rather than polling
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Use multiple VMIDs for each VM, one for each ring. That allows
us to execute flushes separately on each ring, still not ideal
cause in a lot of cases rings can share IDs.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Previously we just allocated space for four hardware semaphores
in each software semaphore object. Make software semaphore objects
represent only one hardware semaphore address again by splitting
the sync code into it's own object.
v2: fix typo in comment
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Use ring structure instead of index and provide vm_id and pd_addr separately.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Always need to set bit 0 of RLC_CGTT_MGCG_OVERRIDE
to avoid unreliable doorbell updates in some cases.
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
The power management code calls into the display code for
certain things. If certain power management sysfs attributes
are called before the driver has finished initializing all of
the hardware we can run into problems with uninitialized
modesetting state. Add a check to make sure modesetting
init has completed to the bandwidth update callbacks to
fix this. Can be triggered by the tlp and laptop start
up scripts depending on the timing.
bugs:
https://bugzilla.kernel.org/show_bug.cgi?id=83611https://bugs.freedesktop.org/show_bug.cgi?id=85771
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
CE ram size is 32k/0k/0k for GFX/CS0/CS1 with CIK
Ported from amdgpu driver.
Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Pull drm updates from Dave Airlie:
"This is the main git pull for the drm,
I pretty much froze major pulls at -rc5/6 time, and haven't had much
fallout, so will probably continue doing that.
Lots of changes all over, big internal header cleanup to make it clear
drm features are legacy things and what are things that modern KMS
drivers should be using. Also big move to use the new generic fences
in all the TTM drivers.
core:
atomic prep work,
vblank rework changes, allows immediate vblank disables
major header reworking and cleanups to better delinate legacy
interfaces from what KMS drivers should be using.
cursor planes locking fixes
ttm:
move to generic fences (affects all TTM drivers)
ppc64 caching fixes
radeon:
userptr support,
uvd for old asics,
reset rework for fence changes
better buffer placement changes,
dpm feature enablement
hdmi audio support fixes
intel:
Cherryview work,
180 degree rotation,
skylake prep work,
execlist command submission
full ppgtt prep work
cursor improvements
edid caching,
vdd handling improvements
nouveau:
fence reworking
kepler memory clock work
gt21x clock work
fan control improvements
hdmi infoframe fixes
DP audio
ast:
ppc64 fixes
caching fix
rcar:
rcar-du DT support
ipuv3:
prep work for capture support
msm:
LVDS support for mdp4, new panel, gpu refactoring
exynos:
exynos3250 SoC support, drop bad mmap interface,
mipi dsi changes, and component match support"
* 'drm-next' of git://people.freedesktop.org/~airlied/linux: (640 commits)
drm/mst: rework payload table allocation to conform better.
drm/ast: Fix HW cursor image
drm/radeon/kv: add uvd/vce info to dpm debugfs output
drm/radeon/ci: add uvd/vce info to dpm debugfs output
drm/radeon: export reservation_object from dmabuf to ttm
drm/radeon: cope with foreign fences inside the reservation object
drm/radeon: cope with foreign fences inside display
drm/core: use helper to check driver features
drm/radeon/cik: write gfx ucode version to ucode addr reg
drm/radeon/si: print full CS when we hit a packet 0
drm/radeon: remove unecessary includes
drm/radeon/combios: declare legacy_connector_convert as static
drm/radeon/atombios: declare connector convert tables as static
drm/radeon: drop btc_get_max_clock_from_voltage_dependency_table
drm/radeon/dpm: drop clk/voltage dependency filters for BTC
drm/radeon/dpm: drop clk/voltage dependency filters for CI
drm/radeon/dpm: drop clk/voltage dependency filters for SI
drm/radeon/dpm: drop clk/voltage dependency filters for NI
drm/radeon: disable audio when we disable hdmi (v2)
drm/radeon: split audio enable between eg and r600 (v2)
...
Adds an extra argument to radeon_bo_create, which is only used in radeon_prime.c.
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Not the whole world is a radeon! :-)
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Helpful for debugging as the version shows up in a
register dump.
Cc: Jay Cornwall <jay.cornwall@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Otherwise we may fail to init the second compute ring.
Noticed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
This might decrease the chance of IH ring buffer overflows.
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Use the same format for all ring indices, and fix the calculation of the
post-overflow RPTR.
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Otherwise the bit remains set in rdev->ih.rptr, so the wptr can never
match that and we still have an infinite loop.
This fix allows me to successfully recover from an IH ring buffer
overflow.
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1
iQEcBAABAgAGBQJUFjfVAAoJEHm+PkMAQRiGANkIAIU3PNrAz9dIItq8a/rEAhnx
l2shHoOyEmyNR2apholM3BPUNX50cbsc/HGdi7lZKLkA/ifAj6B9nFD2NzVsIChD
1QWVcvdkKlVuxXCDd26qbijlfmbTOAWrLw9ntvM+J6ZtECM6zCAZF4MAV/FwogPq
ETGKD76AxJtVIhBMS99troAiC1YxmQ7DKgEr8CraTOR1qwXEonnPCmN/IZA6x2/G
EXiihOuQB5me1X7k4PI0V8CDscQOn+3B2CQHIrjRB+KiTF+iKIuI8n6ORC6bpFh+
U8UZP9wLlIG1BrUHG83pIndglIHotqPcjmtfl1WGrRr2hn7abzVSfV+g5Syo3Vg=
=Ep+s
-----END PGP SIGNATURE-----
drm: backmerge tag 'v3.17-rc5' into drm-next
This is requested to get the fixes for intel and radeon into the
same tree for future development work.
i915_display.c: fix missing dev_priv conflict.
This allows us to specify if we want to sync to
the shared fences of a reservation object or not.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If we assign a Radeon device to a virtual machine, we can no longer
assume a fixed hardware topology, like the GPU having a parent device.
This patch simply adds a few pci_is_root_bus() tests to avoid passing
a NULL pointer to PCI access functions, allowing the radeon driver to
work in a QEMU 440FX machine with an assigned HD8570 on the emulated
PCI root bus.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Blocking completely innocent processes with a GPU reset is
a pretty bad idea. Just set needs_reset and let the next
command submission or fence wait do the job.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This fixes a problem with GPU resets and TLB flushes on SI/CIK.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Need to initialize the mask to 0 on init, otherwise it
keeps increasing.
bug:
https://bugzilla.kernel.org/show_bug.cgi?id=82581
v2: also fix cu count
v3: split count fix into separate patch
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Cc: stable@vger.kernel.org
Fixes lockups due to CP read GPUVM faults when running piglit on Cape
Verde.
v2 (chk): apply the fix to R600+ as well, on CIK only the GFX CP has
a PFP, add more comments to R600 code, enable flushing again
v3: (agd5f): only apply to 7xx+. r6xx does not have the packet.
v4: (agd5f): split flush change into a separate patch, fix formatting
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
It isn't necessary for command streams generated by the kernel (at least
not while we aren't storing ring or indirect buffers in VRAM).
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch moves to radeon the initialization of compute vmid.
That initializations was done in kfd-->kgd interface, but doing it in radeon
as part of radeon's H/W initialization routines is more appropriate.
In addition, this simplifies the kfd-->kgd interface.
The patch removes the function from the interface file and from the interface
declaration file.
The function initializes memory apertures to fixed base/limit address and non
cached memory types.
Signed-off-by: Ben Goz <ben.goz@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Skip the "manual" pageflip completion checks via polling and
guessing in the vblank handler radeon_crtc_handle_vblank() on
asics which are known to reliably support hw pageflip completion
irqs. Those pflip irqs are a more reliable and race-free method
of handling pageflip completion detection, whereas the "classic"
polling method has some small races in combination with dpm on,
and with the reworked pageflip implementation since Linux 3.16.
On old asics without pflip irqs, the classic method is used.
On asics with known good pflip irqs, only pflip irqs are used
by default, but a new module parameter "use_pflipirqs" allows to
override this in case we encounter asics in the wild with
unreliable or faulty pflip irqs. A module parameter of 0 allows
to use the classic method only in such a case. A parameter of 1
allows to use both classic method and pflip irqs as additional
band-aid to avoid some small races which could happen with the
classic method alone. The setting 1 gives Linux 3.16 behaviour.
Hw pflip irqs are available since R600.
Tested on DCE-4, AMD Cedar - FirePro 2270.
v2: agd5f: only enable pflip interrupts on DCE4+ as they are not
reliable on older asics.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Older firmware didn't support the new nop packet.
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Andreas Boll <andreas.boll.dev@gmail.com>
Older firmware didn't support the new nop packet.
v2 (Andreas Boll):
- Drop usage of packet3 for new firmware
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com> (v1)
Signed-off-by: Andreas Boll <andreas.boll.dev@gmail.com>
Cc: stable@vger.kernel.org
This ensures the GPU sees all previous CPU writes to VRAM, which makes it
safe:
* For userspace to stream data from CPU to GPU via VRAM instead of GTT
* For IBs to be stored in VRAM instead of GTT
* For ring buffers to be stored in VRAM instead of GTT, if the HPD flush
is performed via MMIO
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Seems to make VM flushes more stable on SI and CIK.
v2: only use the PFP on the GFX ring on CIK
Signed-off-by: Christian König <christian.koenig@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
v2: fix rebase onto drm-fixes
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Doesn't seem necessary, the GART table memory should be persistent.
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
