Re: [PATCH] drm/amdgpu: Derive RLC SPM xcc from GFXHUB

[Lazar, Lijo]

On 25-Sep-26 4:04 PM, Tvrtko Ursulin wrote:

On 25/09/2026 10:37, Lijo Lazar wrote:
RLC SPM is updated for GFX subsystem related performance counters. Only
jobs on rings that use GFX subsystem need to be updated. Some rings like
SDMA could be inside/outside of GFX subsytem based on SOC. Identify the
rings using GFXHUB assignment and use hub id assigned to find the right
GC rather than relying on xcc id. xcc id may not be necessarily assigned
properly for all rings.

I don't know the hardware well enough to review or be certain of what I 
will ask, but if you remember the "kitten dies" meme it often comes to 
my mind that "every time amdgpu repeatedly walks a list of everything to 
find something at runtime a kitten dies". :)

On a serious note, question I have following the statement from the 
commit message that xcc id is not always correct for all rings is, would 
it be feasible to make it correct? At ring init time I mean. And so 
avoid having to repeatedly look for it.


Having to associate xcc_id with every ring (for ex: vcn/jpeg) is 
overloading it with too many things.

This is one of the changes in another set to remove xcc_id and keep a 
dedicated ip_inst. Then realized that a sdma job also could trigger a 
spm vmid update, then using ring->ip_inst doesn't work (even now, xcc id 
for sdma ring is not assigned properly).

That said, so far for gfxhub, xcc_id = ring->vmhub holds. The loop is 
only to appear it technically good. Also, max xcc = 8 (so far).

Thanks,
Lijo


Regards,

Tvrtko

Signed-off-by: Lijo Lazar <lijo.la...@amd.com>
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c  |  2 +-
  drivers/gpu/drm/amd/amdgpu/amdgpu_rlc.c | 31 +++++++++++++++++++++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_rlc.h |  2 ++
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c  |  2 +-
  4 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/ 
amd/amdgpu/amdgpu_ib.c
index 0f8b097001e3..6ea73ef6e672 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
@@ -212,7 +212,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, 
struct amdgpu_job *job,
          amdgpu_ring_emit_pipeline_sync(ring);
      if (emit_spm_needed)
-        adev->gfx.rlc.funcs->update_spm_vmid(adev, ring->xcc_id, 
ring, job->vmid);
+        amdgpu_gfx_rlc_update_spm(adev, ring, job->vmid);
      if (emit_gds_needed)
          amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_rlc.c b/drivers/gpu/ 
drm/amd/amdgpu/amdgpu_rlc.c
index 002fae3c380e..f791e1bcbf17 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_rlc.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_rlc.c
@@ -27,6 +27,37 @@
  #include "amdgpu_gfx.h"
  #include "amdgpu_rlc.h"
+/**
+ * amdgpu_gfx_rlc_update_spm - update RLC SPM VMID
+ *
+ * @adev: amdgpu_device pointer
+ * @ring: ring the VMID got assigned to
+ * @vmid: VMID to program
+ *
+ * RLC SPM counters are used only for GFX and GFX Cache subsystem. 
Only jobs submitted to
+ * rings in GFX subsystem need to update RLC SPM counters. Derive 
xcc_id from ring->vm_hub.
+ */
+void amdgpu_gfx_rlc_update_spm(struct amdgpu_device *adev,
+                   struct amdgpu_ring *ring, unsigned int vmid)
+{
+    int xcc_id, num_xcc;
+
+    if (!adev->gfx.rlc.funcs->update_spm_vmid || !ring)
+        return;
+
+    if (!AMDGPU_IS_GFXHUB(ring->vm_hub))
+        return;
+
+    num_xcc = NUM_XCC(adev->gfx.xcc_mask);
+    for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
+        if (AMDGPU_GFXHUB(xcc_id) == ring->vm_hub) {
+            adev->gfx.rlc.funcs->update_spm_vmid(adev, xcc_id, ring,
+                                 vmid);
+            return;
+        }
+    }
+}
+
  /**
   * amdgpu_gfx_rlc_enter_safe_mode - Set RLC into safe mode
   *
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_rlc.h b/drivers/gpu/ 
drm/amd/amdgpu/amdgpu_rlc.h
index 959d60c90dcd..1410954da5a6 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_rlc.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_rlc.h
@@ -370,6 +370,8 @@ struct amdgpu_rlc {
      struct amdgpu_rlcg_reg_access_ctrl 
reg_access_ctrl[AMDGPU_MAX_RLC_INSTANCES];
  };
+void amdgpu_gfx_rlc_update_spm(struct amdgpu_device *adev,
+                   struct amdgpu_ring *ring, unsigned int vmid);
  void amdgpu_gfx_rlc_enter_safe_mode(struct amdgpu_device *adev, int 
xcc_id);
  void amdgpu_gfx_rlc_exit_safe_mode(struct amdgpu_device *adev, int 
xcc_id);
  int amdgpu_gfx_rlc_init_sr(struct amdgpu_device *adev, u32 dws);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/ 
amd/amdgpu/amdgpu_vm.c
index 29a66e39f3d6..dc8f85176b8b 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -865,7 +865,7 @@ void amdgpu_vm_flush(struct amdgpu_ring *ring, 
struct amdgpu_job *job,
          amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
      if (spm_update_needed)
-        adev->gfx.rlc.funcs->update_spm_vmid(adev, ring->xcc_id, 
ring, job->vmid);
+        amdgpu_gfx_rlc_update_spm(adev, ring, job->vmid);
      if (gds_switch_needed)
          amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,