Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
drm/amdgpu: always backup and reemit fences
If when we backup the ring contents for reemit before a
ring reset, we skip jobs associated with the bad
context, however, we need to make sure the fences
are reemited as unprocessed submissions may depend on
them.
v2: clean up fence handling, make helpers static
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
+23
-6
+19
-5
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+19
-5
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
···
89
89
return seq;
90
90
}
91
91
92
+
static void amdgpu_fence_save_fence_wptr_start(struct amdgpu_fence *af)
93
+
{
94
+
af->fence_wptr_start = af->ring->wptr;
95
+
}
96
+
97
+
static void amdgpu_fence_save_fence_wptr_end(struct amdgpu_fence *af)
98
+
{
99
+
af->fence_wptr_end = af->ring->wptr;
100
+
}
101
+
92
102
/**
93
103
* amdgpu_fence_emit - emit a fence on the requested ring
94
104
*
···
126
116
&ring->fence_drv.lock,
127
117
adev->fence_context + ring->idx, seq);
128
118
119
+
amdgpu_fence_save_fence_wptr_start(af);
129
120
amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
130
121
seq, flags | AMDGPU_FENCE_FLAG_INT);
122
+
amdgpu_fence_save_fence_wptr_end(af);
131
123
amdgpu_fence_save_wptr(af);
132
124
pm_runtime_get_noresume(adev_to_drm(adev)->dev);
133
125
ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
···
754
742
/* if we've already reemitted once then just cancel everything */
755
743
amdgpu_fence_driver_force_completion(af->ring);
756
744
af->ring->ring_backup_entries_to_copy = 0;
757
-
} else {
758
-
/* signal the guilty fence */
759
-
amdgpu_fence_write(ring, (u32)af->base.seqno);
760
-
amdgpu_fence_process(ring);
761
745
}
762
746
}
763
747
···
803
795
* just save the content from other contexts.
804
796
*/
805
797
if (!fence->reemitted &&
806
-
(!guilty_fence || (fence->context != guilty_fence->context)))
798
+
(!guilty_fence || (fence->context != guilty_fence->context))) {
807
799
amdgpu_ring_backup_unprocessed_command(ring, wptr,
808
800
fence->wptr);
801
+
} else if (!fence->reemitted) {
802
+
/* always save the fence */
803
+
amdgpu_ring_backup_unprocessed_command(ring,
804
+
fence->fence_wptr_start,
805
+
fence->fence_wptr_end);
806
+
}
809
807
wptr = fence->wptr;
810
808
fence->reemitted++;
811
809
}
+4
-1
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
+4
-1
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
···
146
146
struct amdgpu_ring *ring;
147
147
ktime_t start_timestamp;
148
148
149
-
/* wptr for the fence for resets */
149
+
/* wptr for the total submission for resets */
150
150
u64 wptr;
151
151
/* fence context for resets */
152
152
u64 context;
153
153
/* has this fence been reemitted */
154
154
unsigned int reemitted;
155
+
/* wptr for the fence for the submission */
156
+
u64 fence_wptr_start;
157
+
u64 fence_wptr_end;
155
158
};
156
159
157
160
extern const struct drm_sched_backend_ops amdgpu_sched_ops;