Restoring multiple processes concurrently can lead to live-locks
where each process prevents the other from validating all its BOs.
v2: fix duplicate check of same variable
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Oded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
On large-BAR systems the VM page tables for compute are accessed by
the CPU. Always allow CPU access to the page directory so that it can
be used later by the CPU when a VM is converted to a compute VM.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Simulate large-BAR system by exporting only visible memory. This
limits the amount of available VRAM to the size of the BAR, but
enables CPU access to VRAM.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Oded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
The events page must be accessible in user mode by the GPU and CPU
as well as in kernel mode by the CPU. On dGPUs user mode virtual
addresses are managed by the Thunk's GPU memory allocation code.
Therefore we can't allocate the memory in kernel mode like we do
on APUs. But KFD still needs to map the memory for kernel access.
To facilitate this, the Thunk provides the buffer handle of the
events page to KFD when creating the first event.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
v2:
* Fix error handling after kfd_bind_process_to_device in
kfd_ioctl_map_memory_to_gpu
v3:
* Add ioctl to acquire VM from a DRM FD
v4:
* Return number of successful map/unmap operations in failure cases
* Facilitate partial retry after failed map/unmap
* Added comments with parameter descriptions to new APIs
* Defined AMDKFD_IOC_FREE_MEMORY_OF_GPU write-only
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
On GFX7 the CP does not perform a TC flush when queues are unmapped.
To avoid TC eviction from accessing an invalid VMID, flush it
explicitly before releasing a VMID.
v2: Fix unnecessary list_for_each_entry_safe
v3: Moved allocation to kfd_process_device_init_vm
Signed-off-by: Amber Lin <Amber.Lin@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Add helpers for allocating GPUVM memory in kernel mode and use them
to allocate memory for the CWSR trap handler.
v2: Use dev instead of pdd->dev in kfd_process_free_gpuvm
v3:
* Cleaned up and simplified kfd_process_alloc_gpuvm
* Moved allocation for dGPU to kfd_process_device_init_vm
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Also used for cleaning up on process termination.
v2: Refactored cleanup on process termination
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Set up the GPUVM aperture for SVM (shared virtual memory) that allows
sharing a part of virtual address space between GPUs and CPUs.
Report the size of the GPUVM aperture that is supported by KGD accurately.
The low part of the GPUVM aperture is reserved for kernel use. This is
for kernel-allocated buffers that are only accessed on the GPU:
- CWSR trap handler
- IB for submitting commands in user-mode context from kernel mode
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Currently the number of GPUs is limited by aperture placement options
available on GFX7 and GFX8 hardware. This limitation is not necessary.
Scratch and LDS represent per-work-item and per-work-group storage
respectively. Different work-items and work-groups use the same virtual
address to access their own data. Work running on different GPUs is by
definition in different work-groups (different dispatches, in fact).
That means the same virtual addresses can be used for these apertures
on different GPUs.
Add a new AMDKFD_IOC_GET_PROCESS_APERTURES_NEW ioctl that removes the
artificial limitation on the number of GPUs that can be supported. The
new ioctl allows user mode to query the number of GPUs to allocate
enough memory for all GPUs to be reported.
This deprecates AMDKFD_IOC_GET_PROCESS_APERTURES.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Populate DRM render device minor in kfd topology
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Instead of creating all VMs on process creation, create them when
a process is bound to a device. This will later allow registering
an existing VM from a DRM render node FD at runtime, before the
process is bound to the device. This way the render node VM can be
used for KFD instead of creating our own redundant VM.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
This allows acquiring an existing VM from a render node FD to use it
for a compute process.
Such VMs get destroyed when the original file descriptor is released.
Added a callback from amdgpu_vm_fini to handle KFD VM destruction
correctly in this case.
v2:
* Removed vm->vm_context check in amdgpu_amdkfd_gpuvm_destroy_cb,
check vm->process_info earlier instead
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
v2: Removed updating and checking of vm->vm_context
v3: Enable amdgpu_vm_clear_bo in amdgpu_vm_make_compute
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Make sure the PD BO is valid and attach the eviction fence during VM
creation. This ensures that the pd_phys_address is actually valid
and an eviction that would invalidate it triggers a KFD process
eviction like it should.
v2: Use uninterruptible waiting in initial PD validation
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Remove struct amdkfd_vm and move the fields into struct amdgpu_vm.
This will allow turning a VM created by a DRM render node into a
KFD VM.
v2: Removed vm_context field
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
When CONFIG_ACPI is disabled, we never initialize the acpi_table
structure in kfd_create_crat_image_virtual:
drivers/gpu/drm/amd/amdkfd/kfd_crat.c: In function 'kfd_create_crat_image_virtual':
drivers/gpu/drm/amd/amdkfd/kfd_crat.c:888:40: error: 'acpi_table' may be used uninitialized in this function [-Werror=maybe-uninitialized]
The undefined behavior also happens for any other acpi_get_table()
failure, but then the compiler can't warn about it.
This adds an error check that prevents the structure from
being used in error, avoiding both the undefined behavior and
the warning about it.
Fixes: 520b8fb755 ("drm/amdkfd: Add topology support for CPUs")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Last pull for 4.17. Highlights:
- Vega12 support
- A few more bug fixes and cleanups for powerplay
* 'drm-next-4.17' of git://people.freedesktop.org/~agd5f/linux: (77 commits)
drm/amd/pp: clean header file hwmgr.h
drm/amd/pp: use mlck_table.count for array loop index limit
drm/amdgpu: Add an ATPX quirk for hybrid laptop
drm/amdgpu: fix spelling mistake: "asssert" -> "assert"
drm/amd/pp: Add new asic support in pp_psm.c
drm/amd/pp: Clean up powerplay code on Vega12
drm/amd/pp: Add smu irq handlers for legacy asics
drm/amd/pp: Fix set wrong temperature range on smu7
drm/amdgpu: Don't change preferred domian when fallback GTT v5
drm/amdgpu: Fix NULL ptr on driver unload due to init failure.
drm/amdgpu: fix "mitigate workaround for i915"
drm/amd/pp: Add smu irq handlers in sw_init instand of hw_init
drm/amd/pp: Refine register_thermal_interrupt function
drm/amdgpu: Remove wrapper layer of cgs irq handling
drm/amd/powerplay: Return per DPM level clock
drm/amd/powerplay: Remove the SOC floor voltage setting
drm/amdgpu: no job timeout setting on compute queues
drm/amdgpu: add vega12 pci ids (v2)
drm/amd/powerplay: add the hw manager for vega12 (v4)
drm/amd/powerplay: add the smu manager for vega12 (v4)
...
v2: use temporaries to trivially reduces object size.
The for-loops process data in the mclk_table but use slck_table.count
as the loop index limit. I believe these are cut-n-paste errors from
the previous almost identical loops as indicated by static analysis.
Fix these.
Detected by CoverityScan, CID#1466001 ("Copy-paste error")
Fixes: 5d97cf39ff ("drm/amd/pp: Add and initialize OD_dpm_table for CI/VI.")
Fixes: 5e4d4fbea5 ("drm/amd/pp: Implement edit_dpm_table on smu7")
Reviewed-by: Rex Zhu <Rex.Zhu@amd.com>
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
GVT regression fix that caused guest VM GPU hang.
Fix for race conditions in declaring GPU wedged (hit in CI).
* tag 'drm-intel-next-fixes-2018-03-22' of git://anongit.freedesktop.org/drm/drm-intel:
drm/i915/gvt: force to set all context control bits from guest
drm/i915/gvt: Update PDPs after a vGPU mm object is pinned.
drm/i915/gvt: Invalidate vGPU PPGTT mm objects during a vGPU reset.
drm/i915/kvmgt: Handle kzalloc failure
drm/i915/gvt: fix spelling mistake: "destoried" -> "destroyed"
drm/i915/gvt: Remove reduntant printing of untracked mmio
drm/i915/pmu: Work around compiler warnings on some kernel configs
drm/i915: Only call tasklet_kill() on the first prepare_reset
drm/i915: Wrap engine->schedule in RCU locks for set-wedge protection
drm/i915/icl: do not save DDI A/E sharing bit for ICL
A relative large set of various improvements for vmwgfx. Some of them
have been around for a while, some are relatively new, but functionality
should have been tested in our standalone repo.
* 'vmwgfx-next' of git://people.freedesktop.org/~thomash/linux:
drm/vmwgfx: Bump version patchlevel and date
drm/vmwgfx: use monotonic event timestamps
drm/vmwgfx: Unpin the screen object backup buffer when not used
drm/vmwgfx: Stricter count of legacy surface device resources
drm/vmwgfx: Use kasprintf
drm/vmwgfx: Get rid of the device-private suspended member
drm/vmwgfx: Improve on hibernation
drm/vmwgfx: Avoid pinning fbdev framebuffers
drm/vmwgfx: Fix multiple command buffer context use
drm/vmwgfx: Use the cpu blit utility for framebuffer to screen target blits
drm/vmwgfx: Add a cpu blit utility that can be used for page-backed bos
drm/ttm: Export the ttm_k[un]map_atomic_prot API.
drm/ttm: Clean up kmap_atomic_prot selection code
drm/vmwgfx: Cursor update fixes
drm/vmwgfx: Send the correct nonblock option for atomic_commit
drm/vmwgfx: Move the stdu vblank event to atomic function
drm/vmwgfx: Move screen object page flip to atomic function
drm/vmwgfx: Remove drm_crtc_arm_vblank_event from atomic flush
drm/vmwgfx: Move surface copy cmd to atomic function
drm/vmwgfx: Avoid iterating over display unit if crtc is available
Changes this time mostly come down to:
- hook up the DRM GPU scheduler
- prep work for GC7000L support, to be completed in the next cycle
* 'etnaviv/next' of https://git.pengutronix.de/git/lst/linux: (22 commits)
drm/etnaviv: bump HW job limit to 4
drm/etnaviv: etnaviv_sched: Staticize functions when possible
drm/etnaviv: add PTA handling to MMUv2
drm/etnaviv: add function to load the initial PTA state
drm/etnaviv: handle security states
drm/etnaviv: add security handling mode enum
drm/etnaviv: add hardware database
drm/etnaviv: add more minor features fields
drm/etnaviv: update hardware headers from rnndb
drm/etnaviv: add support for slave interface clock
drm/etnaviv: split out and optimize MMU fault dumping
drm/etnaviv: remove the need for a gpu-subsystem DT node
dt-bindings: etnaviv: add slave interface clock
drm/etnaviv: use correct format specifier for size_t
drm/etnaviv: replace hangcheck with scheduler timeout
drm/etnaviv: lock BOs after all other submit work is done
drm/etnaviv: move dependency handling to scheduler
drm/etnaviv: hook up DRM GPU scheduler
drm/etnaviv: track fences by IDR instead of seqno
drm/etnaviv: add missing major features field to debugfs
...
Trivial fix to spelling mistake in pr_err error message text
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In new asics(vega12), no power state management in driver,
So no need to implement related callback functions.
and add some ps checks in pp_psm.c
Revert "drm/amd/powerplay: add new pp_psm infrastructure for vega12 (v2)"
This reverts commit 7d1a63f3aa331b853e41f92d0e7890ed31de8c13.
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Fix the issue thermal irq was always triggered
as GPU under temperature range detected
The low temp in default thermal policy
was set to -273. so need to use int type for the low temp.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
v2: add sanity checking
v3: make code open
v4: also handle visible to invisible fallback
v5: Since two fallback cases, re-use goto retry
Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
DRM_VMW_EVENT_FENCE_SIGNALED (struct drm_vmw_event_fence) and
DRM_EVENT_VBLANK (struct drm_event_vblank) pass timestamps in 32-bit
seconds/microseconds format.
As of commit c61eef726a ("drm: add support for monotonic vblank
timestamps"), other DRM drivers use monotonic times for drm_event_vblank,
but vmwgfx still uses CLOCK_REALTIME for both events, which suffers from
the y2038/y2106 overflow as well as time jumps.
For consistency, this changes vmwgfx to use ktime_get_ts64 as well,
which solves those problems and avoids the deprecated do_gettimeofday()
function.
This should be transparent to to user space, as long as it doesn't
compare the time against the result of gettimeofday().
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
We were relying on the pinned screen object backup buffer to be destroyed
when not used. But if we hold a copy of the atomic state, like when
hibernating, the backup buffer might not be destroyed since it's
refcounted by the atomic state. This causes us to hibernate with a
buffer pinned in VRAM.
Fix this by only having the buffer pinned when it is actually used by a
screen object.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
For legacy surfaces, they were previously registered as device resources
when the driver resources were created. Since they are evictable we instead
register them as device resources once they are created on the device,
just like for guest-backed surfaces. This has implications during
hibernation where we can't hibernate with device resources active.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Deepak Rawat <drawat@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
Use kasprintf instead of combination of kmalloc and sprintf. Also,
remove the local variables used for storing the string length as they
are not required now.
Signed-off-by: Himanshu Jha <himanshujha199640@gmail.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
It was used to early block fbdev dirty processing. Replace it with an
unprotected check of the par->dirty.active field. While this might
race with the vmw_fb_off() function, we do a protected check later so
the race will at worst lead to grabbing and releasing a couple of locks.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
Make it possible to hibernate also with masters that don't switch VT at
hibernation time. We save and restore modesetting state unless fbdev is
active and enabled at hibernation time.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
fbdev framebuffers were previously pinned to be able to keep them mapped
across updates.
This commit introduces a mechanism that instead revalidates the map on
each update, keeping the map cached across updates. The cached map is torn
down if the underlying pages change. Typically on buffer object moves and
swapouts.
This should be nicer to the system when we have resource contention.
Testing done: Basic fbdev functionality under Fedora 27.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Sinclair Yeh <syeh@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Deepak Rawat <drawat@vmware.com>
The start / stop and preempt commands don't honor the context argument
but rather acts on all available contexts.
Also add detection for context 1 availability.
Note that currently there's no driver interface for submitting buffers
using the high-priority command queue (context 1).
Testing done:
Change the default context for command submission to 1 instead of 0,
verify basic desktop functionality including faulty command injection and
recovery.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Deepak Rawat <drawat@vmware.com>
This blit was previously performed using two large vmaps, one of which
was teared down and remapped on each blit. Use the more resource-
conserving TTM cpu blit instead.
The blit is used in boundary-box computing mode which makes it possible
to minimize the bounding box used in host operations.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
The utility uses kmap_atomic() instead of vmapping the whole buffer
object. As a result there will be more book-keeping but on some
architectures this will help avoid exhausting vmalloc space and also
avoid expensive TLB flushes.
The blit utility also adds a provision to compute a bounding box of
changed content, which is very useful to optimize presentation speed
of ill-behaved applications that don't supply proper damage regions, and
for page-flips. The cost of computing the bounding box is not that
expensive when done in a cpu-blit utility like this.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
It will be used by vmwgfx cpu blit.
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Use helpers to perform the kmap_atomic_prot() functionality to
a) Avoid in-function ifdefs that violate the kernel coding policy,
b) Facilitate exporting the functionality.
This commit should not change any functionality.
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
The current limit of 2 leads to some GPU idle times, as the usual
IRQ latency leads to up to 3 jobs getting signaled at once with some
standard workloads.
A larger HW job limit might lead to slightly worse QoS, but we accept
that to not sacrifice GPU throughput in the common case.
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Use drm_plane_helper_check_update also for the cursor plane.
Some applications, like gdm on gnome shell still uses cursor front-buffer
like rendering without notifying the kernel. We do need some kind of
noficiation, but work around this for now by updating the cursor image on
every cursor move.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Page flip can be slow for vmwgfx in some cases, like need to do surface
copy to different surface or waiting for IN_FENCE_FD. Enabling
nonblocking commits for vmwgfx in case userspace request it.
Signed-off-by: Deepak Rawat <drawat@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Atomic ioctl can also send the same page flip flags as legacy ioctl.
In those cases also need to send the vblank event to userspace.
vmwgfx does not support flag DRM_MODE_PAGE_FLIP_ASYNC, so this flag is
never expected.
Signed-off-by: Deepak Rawat <drawat@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
The dmabuf_dirty/surface_dirty in case of screen object is moved to
plane atomic update, so that page flip in atomic ioctl also works.
vmwgfx does not support DRM_MODE_PAGE_FLIP_ASYNC, so this flag is never
expected.
Signed-off-by: Deepak Rawat <drawat@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
The function drm_crtc_arm_vblank_event should be used for the driver
which have vblank interrupt support. In case of vmwgfx we do not have
vblank interrupt.
Signed-off-by: Deepak Rawat <drawat@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
