GitLab

[ Upstream commit b6610101 ]

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>
Signed-off-by: Sasha Levin <sasha.levin@oracle.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>

Signed-off-by: Christian König <christian.koenig@amd.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>

Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-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>

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>

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>

Only increase the higher 32bits if we really detect a wrap around.

v2: instead of increasing the higher 32bits just use the higher
    32bits from the last emitted fence.
v3: also use last emitted fence value as upper limit.

The intention of this patch is to make fences as robust as
they where before introducing 64bit fences. This is
necessary because on older systems it looks like the fence
value gets corrupted on initialization.

Fixes:
https://bugs.freedesktop.org/show_bug.cgi?id=51344

Should also fix:
https://bugs.freedesktop.org/show_bug.cgi?id=54129
https://bugs.freedesktop.org/show_bug.cgi?id=54662
https://bugzilla.redhat.com/show_bug.cgi?id=846505
https://bugzilla.redhat.com/show_bug.cgi?id=845639



3.5 needs a separate patch due to changes in the
fence code.  Will send that out separately.
Signed-off-by: Christian König <deathsimple@vodafone.de>
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>

v2: BUG_ON not matching rings.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.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>