Don't keep around the same pointer twice.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Signed-off-by: NChunming Zhou <David1.Zhou@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

When a BO is moved or destroyed it shouldn't be kmapped any more.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NFelix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

There is no need for page faults to force BOs into visible VRAM if it's
full, and the time it takes to do so is great enough to cause noticeable
stuttering. Add GTT as a possible placement so that if visible VRAM is
full, page faults move BOs to GTT instead of evicting other BOs from VRAM.
Suggested-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NJohn Brooks <john@fastquake.com>
Reviewed-by: NMichel Dänzer <michel.daenzer@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

When a BO is moved to VRAM, clear AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED.
This allows it to potentially later move to invisible VRAM if the CPU
does not access it again.

Setting the CPU_ACCESS flag in amdgpu_bo_fault_reserve_notify() also means
that we can remove the loop to restrict lpfn to the end of visible VRAM,
because amdgpu_ttm_placement_init() will do it for us.

v3 [Michel Dänzer]
* Use AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED instead of a new flag
  (Christian König)
* Clear flag in amdgpu_bo_move instead of amdgpu_move_ram_vram
  (Christian)
* Explicitly mention amdgpu_bo_fault_reserve_notify in amdgpu_bo_move
* Also clear flag in amdgpu_bo_create_restricted
Suggested-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NJohn Brooks <john@fastquake.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

The BO move throttling code is designed to allow VRAM to fill quickly if it
is relatively empty. However, this does not take into account situations
where the visible VRAM is smaller than total VRAM, and total VRAM may not
be close to full but the visible VRAM segment is under pressure. In such
situations, visible VRAM would experience unrestricted swapping and
performance would drop.

Add a separate counter specifically for moves involving visible VRAM, and
check it before moving BOs there.

v2: Only perform calculations for separate counter if visible VRAM is
    smaller than total VRAM. (Michel Dänzer)
v3: [Michel Dänzer]
* Use BO's location rather than the AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
  flag to determine whether to account a move for visible VRAM in most
  cases.
* Use a single

	if (adev->mc.visible_vram_size < adev->mc.real_vram_size) {

  block in amdgpu_cs_get_threshold_for_moves.

Fixes: 95844d20 (drm/amdgpu: throttle buffer migrations at CS using a fixed MBps limit (v2))
Signed-off-by: NJohn Brooks <john@fastquake.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

That is already fixed.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>

This allows us to flush the system VM here.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Acked-by: NFelix Kuehling <Felix.Kuehling@amd.com>

v2: bump the DRM version
Signed-off-by: NMarek Olšák <marek.olsak@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Reviewed-by: NChunming Zhou <david1.zhou@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NRoger.He <Hongbo.He@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

1. The signal interrupt can affect the expected behaviour.
2. There is no good mechanism to handle the corresponding error.
Signed-off-by: NAlex Xie <AlexBin.Xie@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Migration to VRAM will break the sharing, resulting in rendering on the exporting GPU never becoming
visible on the importing GPU.

v2: Don't pin BOs to GTT. Instead, refuse to migrate them out of GTT.
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NChristopher James Halse Rogers <christopher.halse.rogers@canonical.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

This avoids merging them together on page fault.

v2: squash in 64-bit division fix
Signed-off-by: NChristian König <christian.koenig@amd.com>
Acked-by: NMichel Dänzer <michel.daenzer@amd.com>
Reviewed-by: NNicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Implement AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS using TTM_PL_FLAG_CONTIGUOUS
instead of a placement limit. That allows us to better handle CPU
accessible placements.

v2: prevent virtual BO start address from overflowing
Signed-off-by: NChristian König <christian.koenig@amd.com>
Acked-by: NMichel Dänzer <michel.daenzer@amd.com>
Reviewed-by: NNicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

When VRAM pressue and trigger huge evictions there is performance drop,
this patch fix it.
Signed-off-by: NRoger.He <Hongbo.He@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Signed-off-by: NMarek Olšák <marek.olsak@amd.com>
Acked-by: NAlex Deucher <alexander.deucher@amd.com>
Acked-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

By using ttm_bo_init_reserved instead of the manual initialization of
the reservation object, the reservation lock will be properly unlocked
and destroyed when the TTM BO initialization fails.

Actual deadlocks caused by the missing unlock should have been fixed
by "drm/ttm: never add BO that failed to validate to the LRU list",
superseding the flawed fix in commit 38fc4856 ("drm/amdgpu: fix
a potential deadlock in amdgpu_bo_create_restricted()").

This change fixes remaining recursive locking errors that can be seen
with lock debugging enabled, and avoids the error of freeing a locked
mutex.

As an additional minor bonus, buffers created with resv == NULL and
the AMDGPU_GEM_CREATE_VRAM_CLEARED flag are now only added to the
global LRU list after the fill commands have been issued.

v2: use amdgpu_bo_unreserve instead of ttm_bo_unreserve

Fixes: 12a852219583 ("drm/amdgpu: improve AMDGPU_GEM_CREATE_VRAM_CLEARED handling (v2)")
Signed-off-by: NNicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

This reverts commit 38fc4856, which
introduces a use-after-free.

The underlying bug should be properly fixed with "drm/ttm: never add BO
that failed to validate to the LRU list".

Cc: zhoucm1 <david1.zhou@amd.com>
Signed-off-by: NNicolai Hähnle <nicolai.haehnle@amd.com>
Tested-by: NSamuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Like ttm_bo_validate(), ttm_bo_init() might need to move BO and
the number of bytes moved by TTM should be reported. This can help
the throttle buffer migration mechanism to make a better decision.

v2: fix computation
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NSamuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

When ttm_bo_init() fails, the reservation mutex should be unlocked.

In debug build, the kernel reported "possible recursive locking
detected" in this codepath. For debugging purposes, I also added
a "WARN_ON(ww_mutex_is_locked())" when ttm_bo_init() fails and the
mutex was locked as expected.

This should fix (random) GPU hangs. The easy way to reproduce the
issue is to change the "Super Sampling" option from 1.0 to 2.0 in
Hitman. It will create a huge buffer, evict a bunch of buffers
(around ~5k) and deadlock.

This regression has been introduced pretty recently.

v2: only release the mutex if resv is NULL

Fixes: 12a852219583 ("drm/amdgpu: improve AMDGPU_GEM_CREATE_VRAM_CLEARED handling (v2)")
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NSamuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

My randconfig tests on linux-next showed a newly introduced warning:

drivers/gpu/drm/amd/amdgpu/amdgpu_object.c: In function 'amdgpu_bo_create_restricted':
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c:377:2: error: #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance thanks to write-combining [-Werror=cpp]

Generally speaking, warnings about bad kernel configuration are not particularly
helpful. We could enforce the selection of X86_PAT through Kconfig, so the driver
cannot even be used unless it is enabled, or we could just rely on the runtime
warning that is also there.

In this version, I'm making the warning conditional on CONFIG_COMPILE_TEST, which
shuts it up for me, but not people that may actually want to run the kernel
as a compromize.

Fixes: a2e2f299 ("drm/amdgpu: Bring bo creation in line with radeon driver (v2)")
Reviewed-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Add the bo creation changes that have been done to the radeon driver in
recent times, e.g. disable GTT WC on 32 bit because it is broken there,
and also disable it generally (and print a warning message) when
CONFIG_X86_PAT is not set.

v2: agd: fix warning in defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
    case
Reviewed-by: NMichel Dänzer <michel.daenzer@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NNils Holland <nholland@tisys.org>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Give kernel allocations a higher priority cause it is often
more work to swap them back in.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NRoger.He <Hongbo.He@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Keeping groups of BOs on the LRU is to time consuming on command submission.

Instead use the newly added BO priority to give a certain eviction order.

v2: agd: trivial warning fix
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NRoger.He <Hongbo.He@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Unlock the resv lock only if we were the ones to lock it in the first
place.
Signed-off-by: NNicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: NEdward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

ttm_bo_init checks that the reservation object is locked. This is
the caller's responsibility when resv != NULL. Otherwise, the inline
reservation object of the newly allocated buffer is used and must
explicitly be locked.

Uninterruptible w/w locks without an acquire context are always
successful.

v2: use ww_mutex_lock
Signed-off-by: NNicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net> (v1)
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

This catches evictions of shadow page tables from the GART. Since shadow
page tables are always stored in system memory, amdgpu_bo_move is never
called for them.

This fixes a crash during command submission that occurs when only a shadow
page table and no other BOs were evicted since the last submission.

Fixes: 1baa439f ("drm/amdgpu: allocate shadow for pd/pt bo V2")
Signed-off-by: NNicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: NChunming Zhou <david1.zhou@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Ensure that the driver can listen to evictions even when they don't take the
path through ttm_bo_driver::move.

This is crucial for amdgpu, which relies on an eviction counter to skip
re-binding page tables when possible.
Signed-off-by: NNicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: NChunming Zhou <david1.zhou@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Drop this whole reserve/unreserve dance, just lock the
reservation object manually when creating the BO.

v2: rebase on dma_fence renaming
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Manually checking why a function could fail is not a good idea
if you can just check the functions return code.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

This only has the effect of scanning the invisible range twice
since the topdown flag is given anyway.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Reviewed-by: NJimQu <Jim.Qu@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

This fixes a regression in all these drivers since the cache
mode tracking was fixed for mixed mappings. It uses the new
arch API to add the VRAM range to the PAT mapping tracking
tables.

Fixes: 87744ab3 (mm: fix cache mode tracking in vm_insert_mixed())
Reviewed-by: Christian König <christian.koenig@amd.com>.
Signed-off-by: NDave Airlie <airlied@redhat.com>

It's completely pointless to have two pointers to the
device in the same structure.

v2: rename function to amdgpu_ttm_adev, fix typos
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Split VRAM won't have a valid offset, so just set an explicit limit
when the flag is given to trigger reallocation if necessary.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NEdward O'Callaghan <funfunctor@folklore1984.net>
Tested-by: NMike Lothian <mike@fireburn.co.uk>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Add a flag noting that a BO must be created using linear VRAM
and set this flag on all in kernel users where appropriate.

Hopefully I haven't missed anything.

v2: add it in a few more places, fix CPU mapping.
v3: rename to VRAM_CONTIGUOUS, fix typo in CS code.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NEdward O'Callaghan <funfunctor@folklore1984.net>
Tested-by: NMike Lothian <mike@fireburn.co.uk>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

I plan to usurp the short name of struct fence for a core kernel struct,
and so I need to rename the specialised fence/timeline for DMA
operations to make room.

A consensus was reached in
https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html
that making clear this fence applies to DMA operations was a good thing.
Since then the patch has grown a bit as usage increases, so hopefully it
remains a good thing!

(v2...: rebase, rerun spatch)
v3: Compile on msm, spotted a manual fixup that I broke.
v4: Try again for msm, sorry Daniel

coccinelle script:
@@

@@
- struct fence
+ struct dma_fence
@@

@@
- struct fence_ops
+ struct dma_fence_ops
@@

@@
- struct fence_cb
+ struct dma_fence_cb
@@

@@
- struct fence_array
+ struct dma_fence_array
@@

@@
- enum fence_flag_bits
+ enum dma_fence_flag_bits
@@

@@
(
- fence_init
+ dma_fence_init
|
- fence_release
+ dma_fence_release
|
- fence_free
+ dma_fence_free
|
- fence_get
+ dma_fence_get
|
- fence_get_rcu
+ dma_fence_get_rcu
|
- fence_put
+ dma_fence_put
|
- fence_signal
+ dma_fence_signal
|
- fence_signal_locked
+ dma_fence_signal_locked
|
- fence_default_wait
+ dma_fence_default_wait
|
- fence_add_callback
+ dma_fence_add_callback
|
- fence_remove_callback
+ dma_fence_remove_callback
|
- fence_enable_sw_signaling
+ dma_fence_enable_sw_signaling
|
- fence_is_signaled_locked
+ dma_fence_is_signaled_locked
|
- fence_is_signaled
+ dma_fence_is_signaled
|
- fence_is_later
+ dma_fence_is_later
|
- fence_later
+ dma_fence_later
|
- fence_wait_timeout
+ dma_fence_wait_timeout
|
- fence_wait_any_timeout
+ dma_fence_wait_any_timeout
|
- fence_wait
+ dma_fence_wait
|
- fence_context_alloc
+ dma_fence_context_alloc
|
- fence_array_create
+ dma_fence_array_create
|
- to_fence_array
+ to_dma_fence_array
|
- fence_is_array
+ dma_fence_is_array
|
- trace_fence_emit
+ trace_dma_fence_emit
|
- FENCE_TRACE
+ DMA_FENCE_TRACE
|
- FENCE_WARN
+ DMA_FENCE_WARN
|
- FENCE_ERR
+ DMA_FENCE_ERR
)
 (
 ...
 )
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NGustavo Padovan <gustavo.padovan@collabora.co.uk>
Acked-by: NSumit Semwal <sumit.semwal@linaro.org>
Acked-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk

Just to cleanup some radeon leftovers.

sed -i "s/rbo/abo/g" drivers/gpu/drm/amd/amdgpu/*.c
sed -i "s/rbo/abo/g" drivers/gpu/drm/amd/amdgpu/*.h

v2: rebased
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Not used in a while.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Only allocate address space when we really need it.

v2: fix a typo, add correct function description,
    stop leaking the node in the error case.
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Reviewed-by: NFelix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

We get 2 warnings when building kernel with W=1:

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c:985:5: warning: no previous prototype for 'amdgpu_ttm_init' [-Wmissing-prototypes]
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c:1092:6: warning: no previous prototype for 'amdgpu_ttm_fini' [-Wmissing-prototypes]

In fact, both functions are declared in
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c, but should be declared in
a header file, thus can be recognized in other file.

So this patch moves the declarations into
drivers/gpu/drm/amd/amdgpu/amdgpu.h.
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NBaoyou Xie <baoyou.xie@linaro.org>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>