mirrors/linux
1
0
Fork 0
mirror of https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git synced 2026-05-03 12:24:37 -04:00
Code Issues Packages Projects Releases Wiki Activity

drm/i915/gvt: Only pin/unpin intel_context along with workload

Browse Source 
One issue exposed after below commit with which the system will freeze
at suspend after vGPU is created (no need to activate the vGPU).
commit e6ba764802 ("drm/i915: Remove i915->kernel_context")

Old implementation pin the intel_context at setup_submission and
unpin it at clean_submission. So after some vGPU is created, the
intel_context is always pinned there although no workload using it.
It will then block i915 enter suspend state.

There is no need to pin it all the time. Pin/unpin it around workload
lifecycle is more reasonable. After GVT enabled suspend/resume, the
pinned intel_context will also get unpined when userspace put VM process
into suspend state since all workloads are retired, then it's safe to
unpin all intel_context for workloads created. So move the pin/unpin to
create_workload and destroy_workload, while still keep the
create/destroy in old place.

V2:
Rebase.

Fixes: e6ba764802 ("drm/i915: Remove i915->kernel_context")
Reviewed-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Colin Xu <colin.xu@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20201016054059.238371-1-colin.xu@intel.com
This commit is contained in:
Colin Xu
2020-10-16 13:40:59 +08:00
committed by Zhenyu Wang
parent 8fe1056797
commit baec997285
1 changed files with 8 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
drivers/gpu/drm/i915/gvt/scheduler.c
View File

@@ -1268,7 +1268,7 @@ void intel_vgpu_clean_submission(struct intel_vgpu *vgpu)
i915_context_ppgtt_root_restore(s, i915_vm_to_ppgtt(s->shadow[0]->vm));
for_each_engine(engine, vgpu->gvt->gt, id)
intel_context_unpin(s->shadow[id]);
intel_context_put(s->shadow[id]);
kmem_cache_destroy(s->workloads);
}
@@ -1360,11 +1360,6 @@ int intel_vgpu_setup_submission(struct intel_vgpu *vgpu)
ce->ring = __intel_context_ring_size(ring_size);
}
ret = intel_context_pin(ce);
intel_context_put(ce);
if (ret)
goto out_shadow_ctx;
s->shadow[i] = ce;
}
@@ -1396,7 +1391,6 @@ int intel_vgpu_setup_submission(struct intel_vgpu *vgpu)
if (IS_ERR(s->shadow[i]))
break;
intel_context_unpin(s->shadow[i]);
intel_context_put(s->shadow[i]);
}
i915_vm_put(&ppgtt->vm);
@@ -1470,6 +1464,7 @@ void intel_vgpu_destroy_workload(struct intel_vgpu_workload *workload)
{
struct intel_vgpu_submission *s = &workload->vgpu->submission;
intel_context_unpin(s->shadow[workload->engine->id]);
release_shadow_batch_buffer(workload);
release_shadow_wa_ctx(&workload->wa_ctx);
@@ -1715,6 +1710,12 @@ intel_vgpu_create_workload(struct intel_vgpu *vgpu,
return ERR_PTR(ret);
}
ret = intel_context_pin(s->shadow[engine->id]);
if (ret) {
intel_vgpu_destroy_workload(workload);
return ERR_PTR(ret);
}
return workload;
}