[Intel-gfx] [PATCH v6] drm/i915: Add tracepoints to track a vm during its lifetime

Ceraolo Spurio, Daniele daniele.ceraolospurio at intel.com
Mon Nov 10 12:43:49 CET 2014


On 11/10/2014 11:40 AM, Ceraolo Spurio, Daniele wrote:
> On 11/8/2014 8:44 AM, Chris Wilson wrote:
>> On Fri, Nov 07, 2014 at 05:45:01PM +0000,
>> daniele.ceraolospurio at intel.com wrote:
>>> +/**
>>> + * DOC: execlist_submit_context tracepoint
>>> + *
>>> + * These tracepoint are used to track the contexts that are
>>> submitted to the
>>> + * ring. An mm switch is automatically performed by the GPU during
>>> the context
>>> + * switch. Given the fact that the mm switch is an important point
>>> in the
>>> + * lifetime of a vm, the vm assigned to the context is also printed
>>> by the
>>> + * tracepoint when full ppgtt is enabled.
>>> + */
>>> +TRACE_EVENT(execlists_submit_context,
>>> +    TP_PROTO(struct intel_engine_cs *ring, u32 to_num, struct
>>> intel_context *to),
>>> +
>>> +    TP_ARGS(ring, to_num, to),
>>> +
>>> +    TP_STRUCT__entry(
>>> +            __field(u32, ring)
>>> +            __field(u32, to_num)
>>> +            __field(struct intel_context *, to)
>>> +            __field(struct i915_address_space *, vm)
>>> +            __field(u32, dev)
>>> +    ),
>>> +
>>> +    TP_fast_assign(
>>> +            __entry->ring = ring->id;
>>> +            __entry->to_num = to_num;
>>> +            __entry->to = to;
>>> +            __entry->vm = to->ppgtt ? &to->ppgtt->base : NULL;
>>> +            __entry->dev = ring->dev->primary->index;
>>> +    ),
>>> +
>>> +    TP_printk("dev=%u, ring=%u, ctx%d=%p, ctx_vm=%p",
>>> +          __entry->dev, __entry->ring,
>>> +          __entry->to_num, __entry->to, __entry->vm)
>>> +);
>>> +
>>>   #endif /* _I915_TRACE_H_ */
>>>
>>>   /* This part must be outside protection */
>>> diff --git a/drivers/gpu/drm/i915/intel_lrc.c
>>> b/drivers/gpu/drm/i915/intel_lrc.c
>>> index 6025ac7..e72759d 100644
>>> --- a/drivers/gpu/drm/i915/intel_lrc.c
>>> +++ b/drivers/gpu/drm/i915/intel_lrc.c
>>> @@ -135,6 +135,7 @@
>>>   #include <drm/drmP.h>
>>>   #include <drm/i915_drm.h>
>>>   #include "i915_drv.h"
>>> +#include "i915_trace.h"
>>>
>>>   #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE)
>>>   #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE)
>>> @@ -367,6 +368,7 @@ static void execlists_submit_contexts(struct
>>> intel_engine_cs *ring,
>>>       BUG_ON(!ctx_obj0);
>>>       WARN_ON(!i915_gem_obj_is_pinned(ctx_obj0));
>>>
>>> +    trace_execlists_submit_context(ring, 0, to0);
>>>       execlists_ctx_write_tail(ctx_obj0, tail0);
>>
>> This is very tenuous. This is not part of the context lifetime but of
>> the request.
>> -Chris
>>
>
> The aim wasn't to track the context or the request but to track the
> switch_mm. considering that in execlist mode that is automatically done
> by the GPU when the ctx is moved on the ring, this looked like a good
> place to put the trace. What were you exactly concerned about?
>
> Thanks,
> Daniele
>

+ intel-gfx



More information about the Intel-gfx mailing list