When the radeon driver resets a gpu, it attempts to test whether all the
rings can successfully handle an IB. If these rings fail to respond, the
process will wait forever. Another gpu reset can't happen at this point,
as the current reset holds a lock required to do so. Instead, make all
the IB tests run with a timeout, so the system can attempt to recover
in this case.
While this doesn't fix the underlying issue with card resets failing, it
gives the system a higher chance of recovering. These timeouts have been
confirmed to help both a Tathi and Hawaii card recover after a gpu reset.
This also adds a new function, radeon_fence_wait_timeout, that behaves like
fence_wait_timeout. It is used instead of fence_wait_timeout as it continues
to work during a reset. radeon_fence_wait is changed to be implemented
using this function.
V2:
- Changed the timeout to 1s, as the default 10s from radeon_wait_timeout was
too long. A timeout of 100ms was tested and found to be too short.
- Changed radeon_fence_wait_timeout to behave more like fence_wait_timeout.
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Matthew Dawson <matthew@mjdsystems.ca>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In the rare situation where the kmalloc fails we're probably screwed anyway,
but let's try to be more robust about it.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Nicolai Hähnle <Nicolai.Haehnle@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
A normal wait adds to the front of the tail. By doing something
similar to fence_default_wait the fence code can run without racing.
This is a complete fix for "panic on suspend from KDE with radeon",
and a partial fix for "Radeon: System pauses on TAHITI". On tahiti
si_irq_set needs to be fixed too, to completely flush the writes
before radeon_fence_activity is called in radeon_fence_enable_signaling.
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=90741
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=90861
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@ubuntu.com>
Reported-by: Jon Arne Jørgensen <jonjon.arnearne@gmail.com>
Reported-and-tested-by: Gustaw Smolarczyk <wielkiegie@gmail.com>
Cc: stable@vger.kernel.org (v3.18+)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Note for each fence if it's a VM page table update or not. This allows
us to determine the last VM update in a sync object and so to figure
out if we need to flush the TLB or not.
Signed-off-by: Christian König <christian.koenig@amd.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>
Changes since v1:
- Kill the sw interrupt dance, add and use
radeon_irq_kms_sw_irq_get_delayed instead.
- Change custom wait function, lockdep complained about it.
Holding exclusive_lock in the wait function might cause deadlocks.
Instead do all the processing in .enable_signaling, and wait
on the global fence_queue to pick up gpu resets.
- Process all fences in radeon_gpu_reset after reset to close a race
with the trylock in enable_signaling.
Changes since v2:
- Small changes to work with the rewritten lockup recovery patches.
Changes since v3:
- Call radeon_fence_schedule_check when exclusive_lock cannot be
acquired to always cause a wake up.
- Reset irqs from hangup check.
- Drop reading seqno in the callback, use cached value.
- Fix indentation in radeon_fence_default_wait
- Add a radeon_test_signaled function, drop a few test_bit calls.
- Make to_radeon_fence global.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
It's causing issues with VMID handling and comparing the
fence value two times actually doesn't make handling faster.
v2: rebased on reset changes
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 makes it possible to wait for a specific amount of time,
rather than wait until infinity.
v2 (chk): rebased on other changes
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
v5 (chk): complete rework, start when the first fence is emitted,
stop when the last fence is signalled, make it work
correctly with GPU resets, cleanup radeon_fence_wait_seq
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Instead of resetting all fence numbers, only reset the
number of the problematic ring. Split out from a patch
from Maarten Lankhorst <maarten.lankhorst@canonical.com>
v2 (agd5f): rebase build fix
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>
Wake up all fences when we manually trigger a reset.
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>
We no longer need to take the ring lock while checking for
a gpu lockup, so just cleanup the code.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
To workaround bugs and/or certain limits it's sometimes
useful to fall back to waiting on fences.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Don't block forever if there is nothing to wait for.
Signed-off-by: Christian König <christian.koenig@amd.com>
Tested-by: Rafa? Mi?ecki <zajec5@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Stop fiddling with jiffies, always wait for RADEON_FENCE_JIFFIES_TIMEOUT.
Consolidate the two wait sequence implementations into just one function.
Activate all waiters and remember if the reset was already done instead of
trying to reset from only one thread.
v2: clear reset flag earlier to avoid timeout in IB test
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Otherwise just reinitialize from scratch on resume,
and so make it more likely to succeed.
Signed-off-by: Christian König <christian.koenig@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Changing the UVD BOs offset on suspend/resume doesn't work because the VCPU
internally keeps pointers to it. Just keep it always pinned and save the
content manually.
Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=66425
v2: fix compiler warning
v3: fix CIK support
Note: a version of this patch needs to go to stable.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
UVD ring can't use scratch thus it does need writeback buffer to keep
a valid address or radeon_ring_backup will trigger a kernel fault.
It's ok to not unpin the write back buffer on suspend as it leave in
gtt and thus does not need eviction.
v2: Fix the uvd case.
Reported and tracked by Wojtek <wojtask9@wp.pl>
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Also init the scratch reg to zero on the UVD ring.
This fixes UVD on AGP based cards.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Just everything needed to decode videos using UVD.
v6: just all the bugfixes and support for R7xx-SI merged in one patch
v7: UVD_CGC_GATE is a write only register, lockup detection fix
v8: split out VRAM fallback changes, remove support for RV770,
add support for HEMLOCK, add buffer sizes checks
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
radeon_fence_wait_empty_locked should not trigger GPU reset as no
place where it's call from would benefit from such thing and it
actually lead to a kernel deadlock in case the reset is triggered
from pm codepath. Instead force ring completion in place where it
makes sense or return early in others.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Force all fence to signal if GPU reset failed so no process get stuck
on waiting fence.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
The dma ring can't write to register thus have to write to memory
its fence value. This ensure that it doesn't try to use scratch
register for dma ring fence driver.
Should fix:
https://bugs.freedesktop.org/show_bug.cgi?id=58166
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Pull drm merge (part 1) from Dave Airlie:
"So first of all my tree and uapi stuff has a conflict mess, its my
fault as the nouveau stuff didn't hit -next as were trying to rebase
regressions out of it before we merged.
Highlights:
- SH mobile modesetting driver and associated helpers
- some DRM core documentation
- i915 modesetting rework, haswell hdmi, haswell and vlv fixes, write
combined pte writing, ilk rc6 support,
- nouveau: major driver rework into a hw core driver, makes features
like SLI a lot saner to implement,
- psb: add eDP/DP support for Cedarview
- radeon: 2 layer page tables, async VM pte updates, better PLL
selection for > 2 screens, better ACPI interactions
The rest is general grab bag of fixes.
So why part 1? well I have the exynos pull req which came in a bit
late but was waiting for me to do something they shouldn't have and it
looks fairly safe, and David Howells has some more header cleanups
he'd like me to pull, that seem like a good idea, but I'd like to get
this merge out of the way so -next dosen't get blocked."
Tons of conflicts mostly due to silly include line changes, but mostly
mindless. A few other small semantic conflicts too, noted from Dave's
pre-merged branch.
* 'drm-next' of git://people.freedesktop.org/~airlied/linux: (447 commits)
drm/nv98/crypt: fix fuc build with latest envyas
drm/nouveau/devinit: fixup various issues with subdev ctor/init ordering
drm/nv41/vm: fix and enable use of "real" pciegart
drm/nv44/vm: fix and enable use of "real" pciegart
drm/nv04/dmaobj: fixup vm target handling in preparation for nv4x pcie
drm/nouveau: store supported dma mask in vmmgr
drm/nvc0/ibus: initial implementation of subdev
drm/nouveau/therm: add support for fan-control modes
drm/nouveau/hwmon: rename pwm0* to pmw1* to follow hwmon's rules
drm/nouveau/therm: calculate the pwm divisor on nv50+
drm/nouveau/fan: rewrite the fan tachometer driver to get more precision, faster
drm/nouveau/therm: move thermal-related functions to the therm subdev
drm/nouveau/bios: parse the pwm divisor from the perf table
drm/nouveau/therm: use the EXTDEV table to detect i2c monitoring devices
drm/nouveau/therm: rework thermal table parsing
drm/nouveau/gpio: expose the PWM/TOGGLE parameter found in the gpio vbios table
drm/nouveau: fix pm initialization order
drm/nouveau/bios: check that fixed tvdac gpio data is valid before using it
drm/nouveau: log channel debug/error messages from client object rather than drm client
drm/nouveau: have drm debugging macros build on top of core macros
...
Convert #include "..." to #include <path/...> in drivers/gpu/.
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Dave Airlie <airlied@redhat.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Dave Jones <davej@redhat.com>
Remove redundant DRM UAPI header #inclusions from drivers/gpu/.
Remove redundant #inclusions of core DRM UAPI headers (drm.h, drm_mode.h and
drm_sarea.h). They are now #included via drmP.h and drm_crtc.h via a preceding
patch.
Without this patch and the patch to make include the UAPI headers from the core
headers, after the UAPI split, the DRM C sources cannot find these UAPI headers
because the DRM code relies on specific -I flags to make #include "..." work
on headers in include/drm/ - but that does not work after the UAPI split without
adding more -I flags.
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Dave Airlie <airlied@redhat.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Dave Jones <davej@redhat.com>
Let's allow GCC to optimize better.
This exposed some five unused functions, but this patch doesn't remove them.
Signed-off-by: Lauri Kasanen <cand@gmx.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Adds documentation to most of the functions in
radeon_fence.c
v2: address Christian's comments:
- split common concept description into it's own comment
- fix description of intr parameter
- Improve description of -EDEADLK error
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Otherwise the sa managers out of memory
handling doesn't work.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Start with last signaled fence number instead
of last emitted one.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
It is possible that radeon_fence_process is called
after writeback is disabled for suspend, leading
to an invalid read of register 0x0.
This fixes a problem for me where the fence value
is temporary incremented by 0x100000000 on
suspend/resume.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Instead of returning the error handle it directly
and while at it fix the comments about the ring lock.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
It is completely unnecessary to create fences
before they are emitted, so remove it and a bunch
of checks if fences are emitted or not.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
/ssd/git/drm-core-next/drivers/gpu/drm/radeon/radeon_fence.c: In function ‘radeon_debugfs_fence_info’:
/ssd/git/drm-core-next/drivers/gpu/drm/radeon/radeon_fence.c:606:7: warning: format ‘%lx’ expects argument of type ‘long unsigned int’, but argument 3 has type ‘long long int’ [-Wformat]
Signed-off-by: Dave Airlie <airlied@redhat.com>
It never really belonged there in the first place.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Directly use the suballocator to get small chunks of memory.
It's equally fast and doesn't crash when we encounter a GPU reset.
v2: rebased on new SA interface.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Use one wait queue for all rings. When one ring progress, other
likely does to and we are not expecting to have a lot of waiter
anyway.
Also add a fence_wait_any that will wait until the first fence
in the fence array (one fence per ring) is signaled. This allow
to wait on all rings.
v2: some minor cleanups and improvements.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Some callers illegal called fence_wait_next/empty
while holding the ring emission mutex. So don't
relock the mutex in that cases, and move the actual
locking into the fence code.
v2: Don't try to unlock the mutex if it isn't locked.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Using 64bits fence sequence we can directly compare sequence
number to know if a fence is signaled or not. Thus the fence
list became useless, so does the fence lock that mainly
protected the fence list.
Things like ring.ready are no longer behind a lock, this should
be ok as ring.ready is initialized once and will only change
when facing lockup. Worst case is that we return an -EBUSY just
after a successfull GPU reset, or we go into wait state instead
of returning -EBUSY (thus delaying reporting -EBUSY to fence
wait caller).
v2: Remove left over comment, force using writeback on cayman and
newer, thus not having to suffer from possibly scratch reg
exhaustion
v3: Rebase on top of change to uint64 fence patch
v4: Change DCE5 test to force write back on cayman and newer but
also any APU such as PALM or SUMO family
v5: Rebase on top of new uint64 fence patch
v6: Just break if seq doesn't change any more. Use radeon_fence
prefix for all function names. Even if it's now highly optimized,
try avoiding polling to often.
v7: We should never poll the last_seq from the hardware without
waking the sleeping threads, otherwise we might lose events.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This convert fence to use uint64_t sequence number intention is
to use the fact that uin64_t is big enough that we don't need to
care about wrap around.
Tested with and without writeback using 0xFFFFF000 as initial
fence sequence and thus allowing to test the wrap around from
32bits to 64bits.
v2: Add comment about possible race btw CPU & GPU, add comment
stressing that we need 2 dword aligned for R600_WB_EVENT_OFFSET
Read fence sequenc in reverse order of GPU write them so we
mitigate the race btw CPU and GPU.
v3: Drop the need for ring to emit the 64bits fence, and just have
each ring emit the lower 32bits of the fence sequence. We
handle the wrap over 32bits in fence_process.
v4: Just a small optimization: Don't reread the last_seq value
if loop restarts, since we already know its value anyway.
Also start at zero not one for seq value and use pre instead
of post increment in emmit, otherwise wait_empty will deadlock.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Instead of all this humpy pumpy with recursive
mutex (which also fixes only halve of the problem)
move the actual gpu reset out of the fence code,
return -EDEADLK and then reset the gpu in the
calling ioctl function.
v2: Split removal of radeon_mutex into separate patch.
Return -EAGAIN if reset is successful.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
It's never used and so practically superfluous.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
As discussed with Michel that name better
describes the behavior of this function.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
We should signal the caller that we haven't waited at all.
v2: only change fence_wait_next not fence_wait_last.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Matthew Dawson