Hi,
On 07.07.23 10:21, Christian König wrote:
Am 07.07.23 um 09:28 schrieb Friedrich Vock:
Hi Christian,
On 07.07.23 08:56, Christian König wrote:
Am 07.07.23 um 08:28 schrieb Friedrich Vock:
During gfxoff, the per-VMID GDS registers are reset and not restored
afterwards.
Hui? Since when? Those registers should be part of the saved ones.
Have you found that by observation?
yes. I tested this on my RX 6700 XT and the Steam Deck (Vangogh). In the
bug report I linked, a test program using GWS I developed hangs because
of this.
The hang occurs as soon as the kernel re-uses a VMID on which GWS was
already used once. In the hung state, inspecting the per-VMID GWS
registers shows that the values have been reset to 0.
The hang does not occur when gfxoff is disabled.
Even without causing hangs, you can confirm the behaviour by doing the
following:
1. Disable gfxoff.
2. Set some GWS registers.
3. Enable gfxoff and wait a bit.
4. Disable gfxoff and read the registers again. The GWS registers have
been reset.
I performed this test for the GDS_BASE/SIZE registers and it seems these
aren't affected, so it's only GWS that is buggy here.
That's most like a bug in the FW then. I'm going to ask around
internally.
Did the talks with the FW team result in anything yet? It's not that
high-priority, but it'd be nice to know if this is going to be fixed in
the firmware or if I should make a v2 (or if this isn't going to be
fixed at all).
Thanks,
Friedrich
I should probably make a v2 that combines the behaviour before this
patch for GDS and OA, and the patched behaviour for GWS.
Yeah, that sounds like a good idea to me. But let me ping the fw teams
first.
I'm not aware of userspace using GWS (yet, I had some ideas for using it
in RADV which is what I've been writing these tests for),
so perhaps the Cc to stable can also be omitted.
Depends on what the fw teams says. As far as I know GWS has never been
used widely on Linux.
Could be that they say there is a hw bug and we deprecated it for this
generation, or it's simply not handled by the fw and the driver needs
to take care of this (like this patch does) or whatever.
Thanks for the notice,
Christian.
Thanks,
Friedrich
Thanks,
Christian.
The kernel needs to emit a GDS switch to manually update the
GWS registers in this case. Since gfxoff can happen between any two
submissions and the kernel has no way of knowing, emit the GDS switch
before every submission.
Fixes: 56b0989e29 ("drm/amdgpu: fix GDS/GWS/OA switch handling")
Cc: stable@xxxxxxxxxxxxxxx
Link: https://gitlab.freedesktop.org/drm/amd/-/issues/2530
Signed-off-by: Friedrich Vock <friedrich.vock@xxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c | 22 +++++++---------------
drivers/gpu/drm/amd/amdgpu/amdgpu_job.h | 1 -
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 10 ++++++++--
3 files changed, 15 insertions(+), 18 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
index ff1ea99292fb..de73797e9279 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
@@ -165,24 +165,17 @@ bool amdgpu_vmid_had_gpu_reset(struct
amdgpu_device *adev,
atomic_read(&adev->gpu_reset_counter);
}
-/* Check if we need to switch to another set of resources */
-static bool amdgpu_vmid_gds_switch_needed(struct amdgpu_vmid *id,
- struct amdgpu_job *job)
-{
- return id->gds_base != job->gds_base ||
- id->gds_size != job->gds_size ||
- id->gws_base != job->gws_base ||
- id->gws_size != job->gws_size ||
- id->oa_base != job->oa_base ||
- id->oa_size != job->oa_size;
-}
-
/* Check if the id is compatible with the job */
static bool amdgpu_vmid_compatible(struct amdgpu_vmid *id,
struct amdgpu_job *job)
{
return id->pd_gpu_addr == job->vm_pd_addr &&
- !amdgpu_vmid_gds_switch_needed(id, job);
+ id->gds_base == job->gds_base &&
+ id->gds_size == job->gds_size &&
+ id->gws_base == job->gws_base &&
+ id->gws_size == job->gws_size &&
+ id->oa_base == job->oa_base &&
+ id->oa_size == job->oa_size;
}
/**
@@ -434,7 +427,6 @@ int amdgpu_vmid_grab(struct amdgpu_vm *vm, struct
amdgpu_ring *ring,
list_move_tail(&id->list, &id_mgr->ids_lru);
}
- job->gds_switch_needed = amdgpu_vmid_gds_switch_needed(id, job);
if (job->vm_needs_flush) {
id->flushed_updates = amdgpu_vm_tlb_seq(vm);
dma_fence_put(id->last_flush);
@@ -503,7 +495,7 @@ void amdgpu_vmid_free_reserved(struct
amdgpu_device *adev,
* @vmhub: vmhub type
* @vmid: vmid number to use
*
- * Reset saved GDW, GWS and OA to force switch on next flush.
+ * Reset saved GDS, GWS and OA data.
*/
void amdgpu_vmid_reset(struct amdgpu_device *adev, unsigned vmhub,
unsigned vmid)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
index a963a25ddd62..2898508b1ce4 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
@@ -53,7 +53,6 @@ struct amdgpu_job {
uint32_t preamble_status;
uint32_t preemption_status;
bool vm_needs_flush;
- bool gds_switch_needed;
bool spm_update_needed;
uint64_t vm_pd_addr;
unsigned vmid;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 291977b93b1d..61856040cae2 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -557,6 +557,12 @@ void amdgpu_vm_check_compute_bug(struct
amdgpu_device *adev)
}
}
+/* Check if the job needs a GDS switch */
+static bool amdgpu_vm_need_gds_switch(struct amdgpu_job *job)
+{
+ return job->gds_size || job->gws_size || job->oa_size;
+}
+
/**
* amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for
job.
*
@@ -579,7 +585,7 @@ bool amdgpu_vm_need_pipeline_sync(struct
amdgpu_ring *ring,
if (job->vm_needs_flush || ring->has_compute_vm_bug)
return true;
- if (ring->funcs->emit_gds_switch && job->gds_switch_needed)
+ if (ring->funcs->emit_gds_switch &&
amdgpu_vm_need_gds_switch(job))
return true;
if (amdgpu_vmid_had_gpu_reset(adev, &id_mgr->ids[job->vmid]))
@@ -609,7 +615,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring,
struct amdgpu_job *job,
struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
bool spm_update_needed = job->spm_update_needed;
bool gds_switch_needed = ring->funcs->emit_gds_switch &&
- job->gds_switch_needed;
+ amdgpu_vm_need_gds_switch(job);
bool vm_flush_needed = job->vm_needs_flush;
struct dma_fence *fence = NULL;
bool pasid_mapping_needed = false;
--
2.41.0
