Extend map_queue and unmap_queue PM4 packets to support 8
SDMA engines. The new format is backward compatible.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If we shut down a process without having destroyed its GWS-using
queues, it is possible that GWS BO will still be in the process
BO list during the gpuvm destruction. This list should be empty
at that time, so we should remove the GWS allocation at the
process uninit point if it is still around.
Signed-off-by: Joseph Greathouse <Joseph.Greathouse@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Apply the same setting to SH_MEM_CONFIG and VM_CONTEXT1_CNTL. This
makes the noretry param no longer KFD-specific. On GFX10 I'm not
changing SH_MEM_CONFIG in this commit because GFX10 has different
retry behaviour in the SQ and I don't have a way to test it at the
moment.
Suggested-by: Christian König <Christian.Koenig@amd.com>
CC: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by : Shaoyun.liu < Shaoyun.liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The cp hang occurs in OCL conformance test only on supermicro
platform which has 40 cores and the test generates 40 threads.
The root cause is race condition in non-protected flags.
The fix is to add flags of is_evicted and is_active(init_mqd())
into protected area.
Signed-off-by: Eric Huang <JinhuiEric.Huang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
New flag to disable an idle runlist optimization that is causing soft
hangs with some diffult-to-reproduce customer workloads. This will
serve as a workaround until the problem can be reproduced and the
root-cause determined.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Oversubscription of queues or processes results in poor performance
mostly because HWS blinbly schedules busy and idle queues, resulting
in poor occupancy if many queues are idle.
Let users know with a warning message when transitioning from a
non-oversubscribed to an oversubscribed runlist.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Since amdgpu has always requested PCIE atomics, kfd don't
need duplicated PCIE atomics enablement. Referring to amdgpu
request result is enough.
Signed-off-by: Jack Xiao <Jack.Xiao@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In XGMI configuration, more than one asic can be reset at same time,
kfd is able to handle this and no need to trigger the warning
Signed-off-by: shaoyunl <shaoyun.liu@amd.com>
Reviewed-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Previous kfd doesn't use gws so this mask was set to 0.
Set it to 64 bit 1s because now kfd can use all 64 gws
resources.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add a folder structure to /sys/class/kfd/kfd/ called proc which contains
subfolders, each representing an active KFD process' PID, containing 1
file: pasid.
Signed-off-by: Kent Russell <kent.russell@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
SDMA queue allocation requires the dqm lock as it modify
the global dqm members. Enclose it in the dqm_lock.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Philip Yang <philip.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This reverts commit 06b89b38f3.
This fix is not proper. allocate_mqd can't be moved before
allocate_sdma_queue as it depends on q->properties->sdma_id
set in later.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Philip Yang <philip.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This reverts commit f77dac6cd6.
This fix is not proper. allocate_mqd can't be moved before
allocate_sdma_queue as it depends on q->properties->sdma_id
set in later.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Philip Yang <philip.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We can't have devices that are not completely initialized in kfd topology.
Otherwise it is a race condition when user access not completely
initialized device. This also addresses a kfd_topology_add_device accessing
NULL dqm pointer issue.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Move HSA_CAP_ATS_PRESENT initialization logic from kfd iommu codes to
kfd topology codes. This removes kfd_iommu_device_init's dependency
on kfd_topology_add_device. Also remove duplicate code setting the
same.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
SDMA queue allocation requires the dqm lock at it modify
the global dqm members. Move up the dqm_lock so sdma
queue allocation is enclosed in the critical section. Move
mqd allocation out of critical section to avoid circular
lock dependency.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Introduce a new mqd allocation interface and split the original
init_mqd function into two functions: allocate_mqd and init_mqd.
Also renamed uninit_mqd to free_mqd. This is preparation work to
fix a circular lock dependency.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Also calls load_mqd with current->mm struct. The mm
struct is used to read back user wptr of the queue.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Always mark evicted queues with q->properties.is_evicted = true, even
queues that are inactive for other reason. This simplifies maintaining
the eviction state as it doesn't require updating is_evicted when other
queue activation conditions change.
On the other hand, we now need to check those other queue activation
conditions whenever an evicted queues is restored. To minimize code
duplication, move the queue activation check into a macro so it can be
maintained in one central place.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Cox <Philip.Cox@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
New stuff for 5.3:
- Add new thermal sensors for vega asics
- Various RAS fixes
- Add sysfs interface for memory interface utilization
- Use HMM rather than mmu notifier for user pages
- Expose xgmi topology via kfd
- SR-IOV fixes
- Fixes for manual driver reload
- Add unique identifier for vega asics
- Clean up user fence handling with UVD/VCE/VCN blocks
- Convert DC to use core bpc attribute rather than a custom one
- Add GWS support for KFD
- Vega powerplay improvements
- Add CRC support for DCE 12
- SR-IOV support for new security policy
- Various cleanups
From: Alex Deucher <alexdeucher@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190529220944.14464-1-alexander.deucher@amd.com
The pointer dev is set to null yet it is being dereferenced when
checking dev->dqm->sched_policy. Fix this by performing the check
on dev->dqm->sched_policy after dev has been assigned and null
checked. Also remove the redundant null assignment to dev.
Addresses-Coverity: ("Explicit null dereference")
Fixes: 1a058c3376 ("drm/amdkfd: New IOCTL to allocate queue GWS")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add a field in map_queues packet to indicate whether
this is a gws control queue. Only one queue per process
can be gws control queue. Change num_gws field in
map_process packet to 7 bits
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add functions in process queue manager to
set/unset queue gws. Also set process's number
of gws used. Currently only one queue in
process can use and use all gws.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add amdgpu_amdkfd interface to get num_gws and add num_gws
to /sys/class/kfd/kfd/topology/nodes/x/properties. Only report
num_gws if MEC FW support GWS barriers. Currently it is
determined by a module parameter which will be replaced
with MEC FW version check when firmware is ready.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
A multi-socket server can have multiple PCIe segments so BFD is not enough
to distingush each GPU. Also add domain number into account when generating
gpu_id.
Signed-off-by: Amber Lin <Amber.Lin@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Fix a circular lock dependency exposed under userptr memory pressure.
The DQM lock is the only one taken inside the MMU notifier. We need
to make sure that no reclaim is done under this lock, and that
no other locks are taken under which reclaim is possible.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Existing QUEUE_TYPE_SDMA means PCIe optimized SDMA queues.
Introduce a new QUEUE_TYPE_SDMA_XGMI, which is optimized
for non-PCIe transfer such as XGMI.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Previous codes assumes there are two sdma engines.
This is not true e.g., Raven only has 1 SDMA engine.
Fix the issue by using sdma engine number info in
device_info.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
