| Message ID | 20230331200253.1435236-1-umesh.nerlige.ramappa@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | drm/i915/fdinfo: Enable fdinfo for GuC backends | expand |
On 31/03/2023 21:02, Umesh Nerlige Ramappa wrote: > The underlying mechanism used to provide the client specific utilization > in fdinfo is context busyness. When a context switches out, the HW > records the runtime in the specific context field. This information is > accumulated in fdinfo. This information is independent of the scheduling > backend as long as contexts are switching out in reasonable time. > > If a context is actively running on the engine, then the active busyness > is calculated by using the delta between current time and the context > switch in time. This part of the calculation depends on the scheduling > backend. For GuC backends, this data comes entirely from GuC FW. The FW > implementation is WIP. > > Enable fdinfo for GuC based platforms with the exception that long > running contexts will not provide reliable busyness data unless they > switch out at some reasonable point in time. > > Revert "drm/i915: Don't show engine information in fdinfo with GuC submission" > This reverts commit 8d0e3decf26bc423d8b1c42c05636ccc05ea6b17. > > Resolves: https://gitlab.freedesktop.org/drm/intel/issues/8303 > Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> > Test-with: 20230331195915.5839-1-umesh.nerlige.ramappa@intel.com > --- > drivers/gpu/drm/i915/i915_drm_client.c | 6 +----- > 1 file changed, 1 insertion(+), 5 deletions(-) > > diff --git a/drivers/gpu/drm/i915/i915_drm_client.c b/drivers/gpu/drm/i915/i915_drm_client.c > index e8fa172ebe5e..d18d0a3ed905 100644 > --- a/drivers/gpu/drm/i915/i915_drm_client.c > +++ b/drivers/gpu/drm/i915/i915_drm_client.c > @@ -147,11 +147,7 @@ void i915_drm_client_fdinfo(struct seq_file *m, struct file *f) > PCI_SLOT(pdev->devfn), PCI_FUNC(pdev->devfn)); > seq_printf(m, "drm-client-id:\t%u\n", client->id); > > - /* > - * Temporarily skip showing client engine information with GuC submission till > - * fetching engine busyness is implemented in the GuC submission backend > - */ > - if (GRAPHICS_VER(i915) < 8 || intel_uc_uses_guc_submission(&i915->gt0.uc)) > + if (GRAPHICS_VER(i915) < 8) > return; > > for (i = 0; i < ARRAY_SIZE(uabi_class_names); i++) Ack in principle but we have to figure out why Kurt reports this doesn't work - usage always zero with real workloads. Regards, Tvrtko
On Mon, Apr 03, 2023 at 11:21:51AM +0100, Tvrtko Ursulin wrote: > >On 31/03/2023 21:02, Umesh Nerlige Ramappa wrote: >>The underlying mechanism used to provide the client specific utilization >>in fdinfo is context busyness. When a context switches out, the HW >>records the runtime in the specific context field. This information is >>accumulated in fdinfo. This information is independent of the scheduling >>backend as long as contexts are switching out in reasonable time. >> >>If a context is actively running on the engine, then the active busyness >>is calculated by using the delta between current time and the context >>switch in time. This part of the calculation depends on the scheduling >>backend. For GuC backends, this data comes entirely from GuC FW. The FW >>implementation is WIP. >> >>Enable fdinfo for GuC based platforms with the exception that long >>running contexts will not provide reliable busyness data unless they >>switch out at some reasonable point in time. >> >>Revert "drm/i915: Don't show engine information in fdinfo with GuC submission" >>This reverts commit 8d0e3decf26bc423d8b1c42c05636ccc05ea6b17. >> >>Resolves: https://gitlab.freedesktop.org/drm/intel/issues/8303 >>Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> >>Test-with: 20230331195915.5839-1-umesh.nerlige.ramappa@intel.com >>--- >> drivers/gpu/drm/i915/i915_drm_client.c | 6 +----- >> 1 file changed, 1 insertion(+), 5 deletions(-) >> >>diff --git a/drivers/gpu/drm/i915/i915_drm_client.c b/drivers/gpu/drm/i915/i915_drm_client.c >>index e8fa172ebe5e..d18d0a3ed905 100644 >>--- a/drivers/gpu/drm/i915/i915_drm_client.c >>+++ b/drivers/gpu/drm/i915/i915_drm_client.c >>@@ -147,11 +147,7 @@ void i915_drm_client_fdinfo(struct seq_file *m, struct file *f) >> PCI_SLOT(pdev->devfn), PCI_FUNC(pdev->devfn)); >> seq_printf(m, "drm-client-id:\t%u\n", client->id); >>- /* >>- * Temporarily skip showing client engine information with GuC submission till >>- * fetching engine busyness is implemented in the GuC submission backend >>- */ >>- if (GRAPHICS_VER(i915) < 8 || intel_uc_uses_guc_submission(&i915->gt0.uc)) >>+ if (GRAPHICS_VER(i915) < 8) >> return; >> for (i = 0; i < ARRAY_SIZE(uabi_class_names); i++) > >Ack in principle but we have to figure out why Kurt reports this >doesn't work - usage always zero with real workloads. Hmm, looks like this might need some more changes that I initially thought. The changes may need to be extracted from this - https://patchwork.freedesktop.org/patch/496441/ lrc_update_runtime (which updates the stats by peeking into the hw context), is only called when the context is about to be unpinned. The issue is that the GuC context is not unpinned until the context scheduling is disabled. Disabling the context scheduling is asynchronous and does not align with context switch out. Also if requests are scheduled frequently to the context, then the context scheduling is never disabled. In the patch linked above, I am calling lrc_update_runtime whenever the update_stats hook is called (in case of intel_gpu_top, it's every second) and when context is inactive. This resolves the issue, but I need to check with Ashutosh/John if we can do that for this specific case because the way we check that the context is active is dependent on the WIP feature of GuC.. sigh... Regards, Umesh > >Regards, > >Tvrtko
diff --git a/drivers/gpu/drm/i915/i915_drm_client.c b/drivers/gpu/drm/i915/i915_drm_client.c index e8fa172ebe5e..d18d0a3ed905 100644 --- a/drivers/gpu/drm/i915/i915_drm_client.c +++ b/drivers/gpu/drm/i915/i915_drm_client.c @@ -147,11 +147,7 @@ void i915_drm_client_fdinfo(struct seq_file *m, struct file *f) PCI_SLOT(pdev->devfn), PCI_FUNC(pdev->devfn)); seq_printf(m, "drm-client-id:\t%u\n", client->id); - /* - * Temporarily skip showing client engine information with GuC submission till - * fetching engine busyness is implemented in the GuC submission backend - */ - if (GRAPHICS_VER(i915) < 8 || intel_uc_uses_guc_submission(&i915->gt0.uc)) + if (GRAPHICS_VER(i915) < 8) return; for (i = 0; i < ARRAY_SIZE(uabi_class_names); i++)
The underlying mechanism used to provide the client specific utilization in fdinfo is context busyness. When a context switches out, the HW records the runtime in the specific context field. This information is accumulated in fdinfo. This information is independent of the scheduling backend as long as contexts are switching out in reasonable time. If a context is actively running on the engine, then the active busyness is calculated by using the delta between current time and the context switch in time. This part of the calculation depends on the scheduling backend. For GuC backends, this data comes entirely from GuC FW. The FW implementation is WIP. Enable fdinfo for GuC based platforms with the exception that long running contexts will not provide reliable busyness data unless they switch out at some reasonable point in time. Revert "drm/i915: Don't show engine information in fdinfo with GuC submission" This reverts commit 8d0e3decf26bc423d8b1c42c05636ccc05ea6b17. Resolves: https://gitlab.freedesktop.org/drm/intel/issues/8303 Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> Test-with: 20230331195915.5839-1-umesh.nerlige.ramappa@intel.com --- drivers/gpu/drm/i915/i915_drm_client.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-)